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: if (a->A && A->rmap->n == A->cmap->n) PetscCall(MatSetOption(a->A, op, flg));
1690: break;
1691: case MAT_SUBMAT_SINGLEIS:
1692: A->submat_singleis = flg;
1693: break;
1694: case MAT_STRUCTURE_ONLY:
1695: /* The option is handled directly by MatSetOption() */
1696: break;
1697: default:
1698: SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "unknown option %d", op);
1699: }
1700: PetscFunctionReturn(PETSC_SUCCESS);
1701: }
1703: PetscErrorCode MatGetRow_MPIAIJ(Mat matin, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
1704: {
1705: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)matin->data;
1706: PetscScalar *vworkA, *vworkB, **pvA, **pvB, *v_p;
1707: PetscInt i, *cworkA, *cworkB, **pcA, **pcB, cstart = matin->cmap->rstart;
1708: PetscInt nztot, nzA, nzB, lrow, rstart = matin->rmap->rstart, rend = matin->rmap->rend;
1709: PetscInt *cmap, *idx_p;
1711: PetscFunctionBegin;
1712: PetscCheck(!mat->getrowactive, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Already active");
1713: mat->getrowactive = PETSC_TRUE;
1715: if (!mat->rowvalues && (idx || v)) {
1716: /*
1717: allocate enough space to hold information from the longest row.
1718: */
1719: Mat_SeqAIJ *Aa = (Mat_SeqAIJ *)mat->A->data, *Ba = (Mat_SeqAIJ *)mat->B->data;
1720: PetscInt max = 1, tmp;
1721: for (i = 0; i < matin->rmap->n; i++) {
1722: tmp = Aa->i[i + 1] - Aa->i[i] + Ba->i[i + 1] - Ba->i[i];
1723: if (max < tmp) max = tmp;
1724: }
1725: PetscCall(PetscMalloc2(max, &mat->rowvalues, max, &mat->rowindices));
1726: }
1728: PetscCheck(row >= rstart && row < rend, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Only local rows");
1729: lrow = row - rstart;
1731: pvA = &vworkA;
1732: pcA = &cworkA;
1733: pvB = &vworkB;
1734: pcB = &cworkB;
1735: if (!v) {
1736: pvA = NULL;
1737: pvB = NULL;
1738: }
1739: if (!idx) {
1740: pcA = NULL;
1741: if (!v) pcB = NULL;
1742: }
1743: PetscCall((*mat->A->ops->getrow)(mat->A, lrow, &nzA, pcA, pvA));
1744: PetscCall((*mat->B->ops->getrow)(mat->B, lrow, &nzB, pcB, pvB));
1745: nztot = nzA + nzB;
1747: cmap = mat->garray;
1748: if (v || idx) {
1749: if (nztot) {
1750: /* Sort by increasing column numbers, assuming A and B already sorted */
1751: PetscInt imark = -1;
1752: if (v) {
1753: *v = v_p = mat->rowvalues;
1754: for (i = 0; i < nzB; i++) {
1755: if (cmap[cworkB[i]] < cstart) v_p[i] = vworkB[i];
1756: else break;
1757: }
1758: imark = i;
1759: for (i = 0; i < nzA; i++) v_p[imark + i] = vworkA[i];
1760: for (i = imark; i < nzB; i++) v_p[nzA + i] = vworkB[i];
1761: }
1762: if (idx) {
1763: *idx = idx_p = mat->rowindices;
1764: if (imark > -1) {
1765: for (i = 0; i < imark; i++) idx_p[i] = cmap[cworkB[i]];
1766: } else {
1767: for (i = 0; i < nzB; i++) {
1768: if (cmap[cworkB[i]] < cstart) idx_p[i] = cmap[cworkB[i]];
1769: else break;
1770: }
1771: imark = i;
1772: }
1773: for (i = 0; i < nzA; i++) idx_p[imark + i] = cstart + cworkA[i];
1774: for (i = imark; i < nzB; i++) idx_p[nzA + i] = cmap[cworkB[i]];
1775: }
1776: } else {
1777: if (idx) *idx = NULL;
1778: if (v) *v = NULL;
1779: }
1780: }
1781: *nz = nztot;
1782: PetscCall((*mat->A->ops->restorerow)(mat->A, lrow, &nzA, pcA, pvA));
1783: PetscCall((*mat->B->ops->restorerow)(mat->B, lrow, &nzB, pcB, pvB));
1784: PetscFunctionReturn(PETSC_SUCCESS);
1785: }
1787: PetscErrorCode MatRestoreRow_MPIAIJ(Mat mat, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
1788: {
1789: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1791: PetscFunctionBegin;
1792: PetscCheck(aij->getrowactive, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "MatGetRow() must be called first");
1793: aij->getrowactive = PETSC_FALSE;
1794: PetscFunctionReturn(PETSC_SUCCESS);
1795: }
1797: static PetscErrorCode MatNorm_MPIAIJ(Mat mat, NormType type, PetscReal *norm)
1798: {
1799: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1800: Mat_SeqAIJ *amat = (Mat_SeqAIJ *)aij->A->data, *bmat = (Mat_SeqAIJ *)aij->B->data;
1801: PetscInt i, j, cstart = mat->cmap->rstart;
1802: PetscReal sum = 0.0;
1803: const MatScalar *v, *amata, *bmata;
1804: PetscMPIInt iN;
1806: PetscFunctionBegin;
1807: if (aij->size == 1) {
1808: PetscCall(MatNorm(aij->A, type, norm));
1809: } else {
1810: PetscCall(MatSeqAIJGetArrayRead(aij->A, &amata));
1811: PetscCall(MatSeqAIJGetArrayRead(aij->B, &bmata));
1812: if (type == NORM_FROBENIUS) {
1813: v = amata;
1814: for (i = 0; i < amat->nz; i++) {
1815: sum += PetscRealPart(PetscConj(*v) * (*v));
1816: v++;
1817: }
1818: v = bmata;
1819: for (i = 0; i < bmat->nz; i++) {
1820: sum += PetscRealPart(PetscConj(*v) * (*v));
1821: v++;
1822: }
1823: PetscCallMPI(MPIU_Allreduce(&sum, norm, 1, MPIU_REAL, MPIU_SUM, PetscObjectComm((PetscObject)mat)));
1824: *norm = PetscSqrtReal(*norm);
1825: PetscCall(PetscLogFlops(2.0 * amat->nz + 2.0 * bmat->nz));
1826: } else if (type == NORM_1) { /* max column norm */
1827: PetscReal *tmp, *tmp2;
1828: PetscInt *jj, *garray = aij->garray;
1829: PetscCall(PetscCalloc1(mat->cmap->N + 1, &tmp));
1830: PetscCall(PetscMalloc1(mat->cmap->N + 1, &tmp2));
1831: *norm = 0.0;
1832: v = amata;
1833: jj = amat->j;
1834: for (j = 0; j < amat->nz; j++) {
1835: tmp[cstart + *jj++] += PetscAbsScalar(*v);
1836: v++;
1837: }
1838: v = bmata;
1839: jj = bmat->j;
1840: for (j = 0; j < bmat->nz; j++) {
1841: tmp[garray[*jj++]] += PetscAbsScalar(*v);
1842: v++;
1843: }
1844: PetscCall(PetscMPIIntCast(mat->cmap->N, &iN));
1845: PetscCallMPI(MPIU_Allreduce(tmp, tmp2, iN, MPIU_REAL, MPIU_SUM, PetscObjectComm((PetscObject)mat)));
1846: for (j = 0; j < mat->cmap->N; j++) {
1847: if (tmp2[j] > *norm) *norm = tmp2[j];
1848: }
1849: PetscCall(PetscFree(tmp));
1850: PetscCall(PetscFree(tmp2));
1851: PetscCall(PetscLogFlops(PetscMax(amat->nz + bmat->nz - 1, 0)));
1852: } else if (type == NORM_INFINITY) { /* max row norm */
1853: PetscReal ntemp = 0.0;
1854: for (j = 0; j < aij->A->rmap->n; j++) {
1855: v = PetscSafePointerPlusOffset(amata, amat->i[j]);
1856: sum = 0.0;
1857: for (i = 0; i < amat->i[j + 1] - amat->i[j]; i++) {
1858: sum += PetscAbsScalar(*v);
1859: v++;
1860: }
1861: v = PetscSafePointerPlusOffset(bmata, bmat->i[j]);
1862: for (i = 0; i < bmat->i[j + 1] - bmat->i[j]; i++) {
1863: sum += PetscAbsScalar(*v);
1864: v++;
1865: }
1866: if (sum > ntemp) ntemp = sum;
1867: }
1868: PetscCallMPI(MPIU_Allreduce(&ntemp, norm, 1, MPIU_REAL, MPIU_MAX, PetscObjectComm((PetscObject)mat)));
1869: PetscCall(PetscLogFlops(PetscMax(amat->nz + bmat->nz - 1, 0)));
1870: } else SETERRQ(PetscObjectComm((PetscObject)mat), PETSC_ERR_SUP, "No support for two norm");
1871: PetscCall(MatSeqAIJRestoreArrayRead(aij->A, &amata));
1872: PetscCall(MatSeqAIJRestoreArrayRead(aij->B, &bmata));
1873: }
1874: PetscFunctionReturn(PETSC_SUCCESS);
1875: }
1877: static PetscErrorCode MatTranspose_MPIAIJ(Mat A, MatReuse reuse, Mat *matout)
1878: {
1879: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data, *b;
1880: Mat_SeqAIJ *Aloc = (Mat_SeqAIJ *)a->A->data, *Bloc = (Mat_SeqAIJ *)a->B->data, *sub_B_diag;
1881: 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;
1882: const PetscInt *ai, *aj, *bi, *bj, *B_diag_i;
1883: Mat B, A_diag, *B_diag;
1884: const MatScalar *pbv, *bv;
1886: PetscFunctionBegin;
1887: if (reuse == MAT_REUSE_MATRIX) PetscCall(MatTransposeCheckNonzeroState_Private(A, *matout));
1888: ma = A->rmap->n;
1889: na = A->cmap->n;
1890: mb = a->B->rmap->n;
1891: nb = a->B->cmap->n;
1892: ai = Aloc->i;
1893: aj = Aloc->j;
1894: bi = Bloc->i;
1895: bj = Bloc->j;
1896: if (reuse == MAT_INITIAL_MATRIX || *matout == A) {
1897: PetscInt *d_nnz, *g_nnz, *o_nnz;
1898: PetscSFNode *oloc;
1899: PETSC_UNUSED PetscSF sf;
1901: PetscCall(PetscMalloc4(na, &d_nnz, na, &o_nnz, nb, &g_nnz, nb, &oloc));
1902: /* compute d_nnz for preallocation */
1903: PetscCall(PetscArrayzero(d_nnz, na));
1904: for (i = 0; i < ai[ma]; i++) d_nnz[aj[i]]++;
1905: /* compute local off-diagonal contributions */
1906: PetscCall(PetscArrayzero(g_nnz, nb));
1907: for (i = 0; i < bi[ma]; i++) g_nnz[bj[i]]++;
1908: /* map those to global */
1909: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
1910: PetscCall(PetscSFSetGraphLayout(sf, A->cmap, nb, NULL, PETSC_USE_POINTER, a->garray));
1911: PetscCall(PetscSFSetFromOptions(sf));
1912: PetscCall(PetscArrayzero(o_nnz, na));
1913: PetscCall(PetscSFReduceBegin(sf, MPIU_INT, g_nnz, o_nnz, MPI_SUM));
1914: PetscCall(PetscSFReduceEnd(sf, MPIU_INT, g_nnz, o_nnz, MPI_SUM));
1915: PetscCall(PetscSFDestroy(&sf));
1917: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &B));
1918: PetscCall(MatSetSizes(B, A->cmap->n, A->rmap->n, N, M));
1919: PetscCall(MatSetBlockSizes(B, PetscAbs(A->cmap->bs), PetscAbs(A->rmap->bs)));
1920: PetscCall(MatSetType(B, ((PetscObject)A)->type_name));
1921: PetscCall(MatMPIAIJSetPreallocation(B, 0, d_nnz, 0, o_nnz));
1922: PetscCall(PetscFree4(d_nnz, o_nnz, g_nnz, oloc));
1923: } else {
1924: B = *matout;
1925: PetscCall(MatSetOption(B, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_TRUE));
1926: }
1928: b = (Mat_MPIAIJ *)B->data;
1929: A_diag = a->A;
1930: B_diag = &b->A;
1931: sub_B_diag = (Mat_SeqAIJ *)(*B_diag)->data;
1932: A_diag_ncol = A_diag->cmap->N;
1933: B_diag_ilen = sub_B_diag->ilen;
1934: B_diag_i = sub_B_diag->i;
1936: /* Set ilen for diagonal of B */
1937: for (i = 0; i < A_diag_ncol; i++) B_diag_ilen[i] = B_diag_i[i + 1] - B_diag_i[i];
1939: /* Transpose the diagonal part of the matrix. In contrast to the off-diagonal part, this can be done
1940: very quickly (=without using MatSetValues), because all writes are local. */
1941: PetscCall(MatTransposeSetPrecursor(A_diag, *B_diag));
1942: PetscCall(MatTranspose(A_diag, MAT_REUSE_MATRIX, B_diag));
1944: /* copy over the B part */
1945: PetscCall(PetscMalloc1(bi[mb], &cols));
1946: PetscCall(MatSeqAIJGetArrayRead(a->B, &bv));
1947: pbv = bv;
1948: row = A->rmap->rstart;
1949: for (i = 0; i < bi[mb]; i++) cols[i] = a->garray[bj[i]];
1950: cols_tmp = cols;
1951: for (i = 0; i < mb; i++) {
1952: ncol = bi[i + 1] - bi[i];
1953: PetscCall(MatSetValues(B, ncol, cols_tmp, 1, &row, pbv, INSERT_VALUES));
1954: row++;
1955: if (pbv) pbv += ncol;
1956: if (cols_tmp) cols_tmp += ncol;
1957: }
1958: PetscCall(PetscFree(cols));
1959: PetscCall(MatSeqAIJRestoreArrayRead(a->B, &bv));
1961: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
1962: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
1963: if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX) {
1964: *matout = B;
1965: } else {
1966: PetscCall(MatHeaderMerge(A, &B));
1967: }
1968: PetscFunctionReturn(PETSC_SUCCESS);
1969: }
1971: static PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat, Vec ll, Vec rr)
1972: {
1973: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1974: Mat a = aij->A, b = aij->B;
1975: PetscInt s1, s2, s3;
1977: PetscFunctionBegin;
1978: PetscCall(MatGetLocalSize(mat, &s2, &s3));
1979: if (rr) {
1980: PetscCall(VecGetLocalSize(rr, &s1));
1981: PetscCheck(s1 == s3, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "right vector non-conforming local size");
1982: /* Overlap communication with computation. */
1983: PetscCall(VecScatterBegin(aij->Mvctx, rr, aij->lvec, INSERT_VALUES, SCATTER_FORWARD));
1984: }
1985: if (ll) {
1986: PetscCall(VecGetLocalSize(ll, &s1));
1987: PetscCheck(s1 == s2, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "left vector non-conforming local size");
1988: PetscUseTypeMethod(b, diagonalscale, ll, NULL);
1989: }
1990: /* scale the diagonal block */
1991: PetscUseTypeMethod(a, diagonalscale, ll, rr);
1993: if (rr) {
1994: /* Do a scatter end and then right scale the off-diagonal block */
1995: PetscCall(VecScatterEnd(aij->Mvctx, rr, aij->lvec, INSERT_VALUES, SCATTER_FORWARD));
1996: PetscUseTypeMethod(b, diagonalscale, NULL, aij->lvec);
1997: }
1998: PetscFunctionReturn(PETSC_SUCCESS);
1999: }
2001: static PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A)
2002: {
2003: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2005: PetscFunctionBegin;
2006: PetscCall(MatSetUnfactored(a->A));
2007: PetscFunctionReturn(PETSC_SUCCESS);
2008: }
2010: static PetscErrorCode MatEqual_MPIAIJ(Mat A, Mat B, PetscBool *flag)
2011: {
2012: Mat_MPIAIJ *matB = (Mat_MPIAIJ *)B->data, *matA = (Mat_MPIAIJ *)A->data;
2013: Mat a, b, c, d;
2014: PetscBool flg;
2016: PetscFunctionBegin;
2017: a = matA->A;
2018: b = matA->B;
2019: c = matB->A;
2020: d = matB->B;
2022: PetscCall(MatEqual(a, c, &flg));
2023: if (flg) PetscCall(MatEqual(b, d, &flg));
2024: PetscCallMPI(MPIU_Allreduce(&flg, flag, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)A)));
2025: PetscFunctionReturn(PETSC_SUCCESS);
2026: }
2028: static PetscErrorCode MatCopy_MPIAIJ(Mat A, Mat B, MatStructure str)
2029: {
2030: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2031: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2033: PetscFunctionBegin;
2034: /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */
2035: if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) {
2036: /* because of the column compression in the off-processor part of the matrix a->B,
2037: the number of columns in a->B and b->B may be different, hence we cannot call
2038: the MatCopy() directly on the two parts. If need be, we can provide a more
2039: efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices
2040: then copying the submatrices */
2041: PetscCall(MatCopy_Basic(A, B, str));
2042: } else {
2043: PetscCall(MatCopy(a->A, b->A, str));
2044: PetscCall(MatCopy(a->B, b->B, str));
2045: }
2046: PetscCall(PetscObjectStateIncrease((PetscObject)B));
2047: PetscFunctionReturn(PETSC_SUCCESS);
2048: }
2050: /*
2051: Computes the number of nonzeros per row needed for preallocation when X and Y
2052: have different nonzero structure.
2053: */
2054: 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)
2055: {
2056: PetscInt i, j, k, nzx, nzy;
2058: PetscFunctionBegin;
2059: /* Set the number of nonzeros in the new matrix */
2060: for (i = 0; i < m; i++) {
2061: const PetscInt *xjj = PetscSafePointerPlusOffset(xj, xi[i]), *yjj = PetscSafePointerPlusOffset(yj, yi[i]);
2062: nzx = xi[i + 1] - xi[i];
2063: nzy = yi[i + 1] - yi[i];
2064: nnz[i] = 0;
2065: for (j = 0, k = 0; j < nzx; j++) { /* Point in X */
2066: for (; k < nzy && yltog[yjj[k]] < xltog[xjj[j]]; k++) nnz[i]++; /* Catch up to X */
2067: if (k < nzy && yltog[yjj[k]] == xltog[xjj[j]]) k++; /* Skip duplicate */
2068: nnz[i]++;
2069: }
2070: for (; k < nzy; k++) nnz[i]++;
2071: }
2072: PetscFunctionReturn(PETSC_SUCCESS);
2073: }
2075: /* This is the same as MatAXPYGetPreallocation_SeqAIJ, except that the local-to-global map is provided */
2076: static PetscErrorCode MatAXPYGetPreallocation_MPIAIJ(Mat Y, const PetscInt *yltog, Mat X, const PetscInt *xltog, PetscInt *nnz)
2077: {
2078: PetscInt m = Y->rmap->N;
2079: Mat_SeqAIJ *x = (Mat_SeqAIJ *)X->data;
2080: Mat_SeqAIJ *y = (Mat_SeqAIJ *)Y->data;
2082: PetscFunctionBegin;
2083: PetscCall(MatAXPYGetPreallocation_MPIX_private(m, x->i, x->j, xltog, y->i, y->j, yltog, nnz));
2084: PetscFunctionReturn(PETSC_SUCCESS);
2085: }
2087: static PetscErrorCode MatAXPY_MPIAIJ(Mat Y, PetscScalar a, Mat X, MatStructure str)
2088: {
2089: Mat_MPIAIJ *xx = (Mat_MPIAIJ *)X->data, *yy = (Mat_MPIAIJ *)Y->data;
2091: PetscFunctionBegin;
2092: if (str == SAME_NONZERO_PATTERN) {
2093: PetscCall(MatAXPY(yy->A, a, xx->A, str));
2094: PetscCall(MatAXPY(yy->B, a, xx->B, str));
2095: } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */
2096: PetscCall(MatAXPY_Basic(Y, a, X, str));
2097: } else {
2098: Mat B;
2099: PetscInt *nnz_d, *nnz_o;
2101: PetscCall(PetscMalloc1(yy->A->rmap->N, &nnz_d));
2102: PetscCall(PetscMalloc1(yy->B->rmap->N, &nnz_o));
2103: PetscCall(MatCreate(PetscObjectComm((PetscObject)Y), &B));
2104: PetscCall(PetscObjectSetName((PetscObject)B, ((PetscObject)Y)->name));
2105: PetscCall(MatSetLayouts(B, Y->rmap, Y->cmap));
2106: PetscCall(MatSetType(B, ((PetscObject)Y)->type_name));
2107: PetscCall(MatAXPYGetPreallocation_SeqAIJ(yy->A, xx->A, nnz_d));
2108: PetscCall(MatAXPYGetPreallocation_MPIAIJ(yy->B, yy->garray, xx->B, xx->garray, nnz_o));
2109: PetscCall(MatMPIAIJSetPreallocation(B, 0, nnz_d, 0, nnz_o));
2110: PetscCall(MatAXPY_BasicWithPreallocation(B, Y, a, X, str));
2111: PetscCall(MatHeaderMerge(Y, &B));
2112: PetscCall(PetscFree(nnz_d));
2113: PetscCall(PetscFree(nnz_o));
2114: }
2115: PetscFunctionReturn(PETSC_SUCCESS);
2116: }
2118: PETSC_INTERN PetscErrorCode MatConjugate_SeqAIJ(Mat);
2120: static PetscErrorCode MatConjugate_MPIAIJ(Mat mat)
2121: {
2122: PetscFunctionBegin;
2123: if (PetscDefined(USE_COMPLEX)) {
2124: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2126: PetscCall(MatConjugate_SeqAIJ(aij->A));
2127: PetscCall(MatConjugate_SeqAIJ(aij->B));
2128: }
2129: PetscFunctionReturn(PETSC_SUCCESS);
2130: }
2132: static PetscErrorCode MatRealPart_MPIAIJ(Mat A)
2133: {
2134: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2136: PetscFunctionBegin;
2137: PetscCall(MatRealPart(a->A));
2138: PetscCall(MatRealPart(a->B));
2139: PetscFunctionReturn(PETSC_SUCCESS);
2140: }
2142: static PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A)
2143: {
2144: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2146: PetscFunctionBegin;
2147: PetscCall(MatImaginaryPart(a->A));
2148: PetscCall(MatImaginaryPart(a->B));
2149: PetscFunctionReturn(PETSC_SUCCESS);
2150: }
2152: static PetscErrorCode MatGetRowMaxAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2153: {
2154: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2155: PetscInt i, *idxb = NULL, m = A->rmap->n;
2156: PetscScalar *vv;
2157: Vec vB, vA;
2158: const PetscScalar *va, *vb;
2160: PetscFunctionBegin;
2161: PetscCall(MatCreateVecs(a->A, NULL, &vA));
2162: PetscCall(MatGetRowMaxAbs(a->A, vA, idx));
2164: PetscCall(VecGetArrayRead(vA, &va));
2165: if (idx) {
2166: for (i = 0; i < m; i++) {
2167: if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart;
2168: }
2169: }
2171: PetscCall(MatCreateVecs(a->B, NULL, &vB));
2172: PetscCall(PetscMalloc1(m, &idxb));
2173: PetscCall(MatGetRowMaxAbs(a->B, vB, idxb));
2175: PetscCall(VecGetArrayWrite(v, &vv));
2176: PetscCall(VecGetArrayRead(vB, &vb));
2177: for (i = 0; i < m; i++) {
2178: if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) {
2179: vv[i] = vb[i];
2180: if (idx) idx[i] = a->garray[idxb[i]];
2181: } else {
2182: vv[i] = va[i];
2183: if (idx && PetscAbsScalar(va[i]) == PetscAbsScalar(vb[i]) && idxb[i] != -1 && idx[i] > a->garray[idxb[i]]) idx[i] = a->garray[idxb[i]];
2184: }
2185: }
2186: PetscCall(VecRestoreArrayWrite(v, &vv));
2187: PetscCall(VecRestoreArrayRead(vA, &va));
2188: PetscCall(VecRestoreArrayRead(vB, &vb));
2189: PetscCall(PetscFree(idxb));
2190: PetscCall(VecDestroy(&vA));
2191: PetscCall(VecDestroy(&vB));
2192: PetscFunctionReturn(PETSC_SUCCESS);
2193: }
2195: static PetscErrorCode MatGetRowSumAbs_MPIAIJ(Mat A, Vec v)
2196: {
2197: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2198: Vec vB, vA;
2200: PetscFunctionBegin;
2201: PetscCall(MatCreateVecs(a->A, NULL, &vA));
2202: PetscCall(MatGetRowSumAbs(a->A, vA));
2203: PetscCall(MatCreateVecs(a->B, NULL, &vB));
2204: PetscCall(MatGetRowSumAbs(a->B, vB));
2205: PetscCall(VecAXPY(vA, 1.0, vB));
2206: PetscCall(VecDestroy(&vB));
2207: PetscCall(VecCopy(vA, v));
2208: PetscCall(VecDestroy(&vA));
2209: PetscFunctionReturn(PETSC_SUCCESS);
2210: }
2212: static PetscErrorCode MatGetRowMinAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2213: {
2214: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
2215: PetscInt m = A->rmap->n, n = A->cmap->n;
2216: PetscInt cstart = A->cmap->rstart, cend = A->cmap->rend;
2217: PetscInt *cmap = mat->garray;
2218: PetscInt *diagIdx, *offdiagIdx;
2219: Vec diagV, offdiagV;
2220: PetscScalar *a, *diagA, *offdiagA;
2221: const PetscScalar *ba, *bav;
2222: PetscInt r, j, col, ncols, *bi, *bj;
2223: Mat B = mat->B;
2224: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2226: PetscFunctionBegin;
2227: /* When a process holds entire A and other processes have no entry */
2228: if (A->cmap->N == n) {
2229: PetscCall(VecGetArrayWrite(v, &diagA));
2230: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2231: PetscCall(MatGetRowMinAbs(mat->A, diagV, idx));
2232: PetscCall(VecDestroy(&diagV));
2233: PetscCall(VecRestoreArrayWrite(v, &diagA));
2234: PetscFunctionReturn(PETSC_SUCCESS);
2235: } else if (n == 0) {
2236: if (m) {
2237: PetscCall(VecGetArrayWrite(v, &a));
2238: for (r = 0; r < m; r++) {
2239: a[r] = 0.0;
2240: if (idx) idx[r] = -1;
2241: }
2242: PetscCall(VecRestoreArrayWrite(v, &a));
2243: }
2244: PetscFunctionReturn(PETSC_SUCCESS);
2245: }
2247: PetscCall(PetscMalloc2(m, &diagIdx, m, &offdiagIdx));
2248: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2249: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2250: PetscCall(MatGetRowMinAbs(mat->A, diagV, diagIdx));
2252: /* Get offdiagIdx[] for implicit 0.0 */
2253: PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2254: ba = bav;
2255: bi = b->i;
2256: bj = b->j;
2257: PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2258: for (r = 0; r < m; r++) {
2259: ncols = bi[r + 1] - bi[r];
2260: if (ncols == A->cmap->N - n) { /* Brow is dense */
2261: offdiagA[r] = *ba;
2262: offdiagIdx[r] = cmap[0];
2263: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2264: offdiagA[r] = 0.0;
2266: /* Find first hole in the cmap */
2267: for (j = 0; j < ncols; j++) {
2268: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2269: if (col > j && j < cstart) {
2270: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2271: break;
2272: } else if (col > j + n && j >= cstart) {
2273: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2274: break;
2275: }
2276: }
2277: if (j == ncols && ncols < A->cmap->N - n) {
2278: /* a hole is outside compressed Bcols */
2279: if (ncols == 0) {
2280: if (cstart) {
2281: offdiagIdx[r] = 0;
2282: } else offdiagIdx[r] = cend;
2283: } else { /* ncols > 0 */
2284: offdiagIdx[r] = cmap[ncols - 1] + 1;
2285: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2286: }
2287: }
2288: }
2290: for (j = 0; j < ncols; j++) {
2291: if (PetscAbsScalar(offdiagA[r]) > PetscAbsScalar(*ba)) {
2292: offdiagA[r] = *ba;
2293: offdiagIdx[r] = cmap[*bj];
2294: }
2295: ba++;
2296: bj++;
2297: }
2298: }
2300: PetscCall(VecGetArrayWrite(v, &a));
2301: PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2302: for (r = 0; r < m; ++r) {
2303: if (PetscAbsScalar(diagA[r]) < PetscAbsScalar(offdiagA[r])) {
2304: a[r] = diagA[r];
2305: if (idx) idx[r] = cstart + diagIdx[r];
2306: } else if (PetscAbsScalar(diagA[r]) == PetscAbsScalar(offdiagA[r])) {
2307: a[r] = diagA[r];
2308: if (idx) {
2309: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2310: idx[r] = cstart + diagIdx[r];
2311: } else idx[r] = offdiagIdx[r];
2312: }
2313: } else {
2314: a[r] = offdiagA[r];
2315: if (idx) idx[r] = offdiagIdx[r];
2316: }
2317: }
2318: PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2319: PetscCall(VecRestoreArrayWrite(v, &a));
2320: PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2321: PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2322: PetscCall(VecDestroy(&diagV));
2323: PetscCall(VecDestroy(&offdiagV));
2324: PetscCall(PetscFree2(diagIdx, offdiagIdx));
2325: PetscFunctionReturn(PETSC_SUCCESS);
2326: }
2328: static PetscErrorCode MatGetRowMin_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2329: {
2330: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
2331: PetscInt m = A->rmap->n, n = A->cmap->n;
2332: PetscInt cstart = A->cmap->rstart, cend = A->cmap->rend;
2333: PetscInt *cmap = mat->garray;
2334: PetscInt *diagIdx, *offdiagIdx;
2335: Vec diagV, offdiagV;
2336: PetscScalar *a, *diagA, *offdiagA;
2337: const PetscScalar *ba, *bav;
2338: PetscInt r, j, col, ncols, *bi, *bj;
2339: Mat B = mat->B;
2340: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2342: PetscFunctionBegin;
2343: /* When a process holds entire A and other processes have no entry */
2344: if (A->cmap->N == n) {
2345: PetscCall(VecGetArrayWrite(v, &diagA));
2346: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2347: PetscCall(MatGetRowMin(mat->A, diagV, idx));
2348: PetscCall(VecDestroy(&diagV));
2349: PetscCall(VecRestoreArrayWrite(v, &diagA));
2350: PetscFunctionReturn(PETSC_SUCCESS);
2351: } else if (n == 0) {
2352: if (m) {
2353: PetscCall(VecGetArrayWrite(v, &a));
2354: for (r = 0; r < m; r++) {
2355: a[r] = PETSC_MAX_REAL;
2356: if (idx) idx[r] = -1;
2357: }
2358: PetscCall(VecRestoreArrayWrite(v, &a));
2359: }
2360: PetscFunctionReturn(PETSC_SUCCESS);
2361: }
2363: PetscCall(PetscCalloc2(m, &diagIdx, m, &offdiagIdx));
2364: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2365: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2366: PetscCall(MatGetRowMin(mat->A, diagV, diagIdx));
2368: /* Get offdiagIdx[] for implicit 0.0 */
2369: PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2370: ba = bav;
2371: bi = b->i;
2372: bj = b->j;
2373: PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2374: for (r = 0; r < m; r++) {
2375: ncols = bi[r + 1] - bi[r];
2376: if (ncols == A->cmap->N - n) { /* Brow is dense */
2377: offdiagA[r] = *ba;
2378: offdiagIdx[r] = cmap[0];
2379: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2380: offdiagA[r] = 0.0;
2382: /* Find first hole in the cmap */
2383: for (j = 0; j < ncols; j++) {
2384: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2385: if (col > j && j < cstart) {
2386: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2387: break;
2388: } else if (col > j + n && j >= cstart) {
2389: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2390: break;
2391: }
2392: }
2393: if (j == ncols && ncols < A->cmap->N - n) {
2394: /* a hole is outside compressed Bcols */
2395: if (ncols == 0) {
2396: if (cstart) {
2397: offdiagIdx[r] = 0;
2398: } else offdiagIdx[r] = cend;
2399: } else { /* ncols > 0 */
2400: offdiagIdx[r] = cmap[ncols - 1] + 1;
2401: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2402: }
2403: }
2404: }
2406: for (j = 0; j < ncols; j++) {
2407: if (PetscRealPart(offdiagA[r]) > PetscRealPart(*ba)) {
2408: offdiagA[r] = *ba;
2409: offdiagIdx[r] = cmap[*bj];
2410: }
2411: ba++;
2412: bj++;
2413: }
2414: }
2416: PetscCall(VecGetArrayWrite(v, &a));
2417: PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2418: for (r = 0; r < m; ++r) {
2419: if (PetscRealPart(diagA[r]) < PetscRealPart(offdiagA[r])) {
2420: a[r] = diagA[r];
2421: if (idx) idx[r] = cstart + diagIdx[r];
2422: } else if (PetscRealPart(diagA[r]) == PetscRealPart(offdiagA[r])) {
2423: a[r] = diagA[r];
2424: if (idx) {
2425: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2426: idx[r] = cstart + diagIdx[r];
2427: } else idx[r] = offdiagIdx[r];
2428: }
2429: } else {
2430: a[r] = offdiagA[r];
2431: if (idx) idx[r] = offdiagIdx[r];
2432: }
2433: }
2434: PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2435: PetscCall(VecRestoreArrayWrite(v, &a));
2436: PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2437: PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2438: PetscCall(VecDestroy(&diagV));
2439: PetscCall(VecDestroy(&offdiagV));
2440: PetscCall(PetscFree2(diagIdx, offdiagIdx));
2441: PetscFunctionReturn(PETSC_SUCCESS);
2442: }
2444: static PetscErrorCode MatGetRowMax_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2445: {
2446: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
2447: PetscInt m = A->rmap->n, n = A->cmap->n;
2448: PetscInt cstart = A->cmap->rstart, cend = A->cmap->rend;
2449: PetscInt *cmap = mat->garray;
2450: PetscInt *diagIdx, *offdiagIdx;
2451: Vec diagV, offdiagV;
2452: PetscScalar *a, *diagA, *offdiagA;
2453: const PetscScalar *ba, *bav;
2454: PetscInt r, j, col, ncols, *bi, *bj;
2455: Mat B = mat->B;
2456: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2458: PetscFunctionBegin;
2459: /* When a process holds entire A and other processes have no entry */
2460: if (A->cmap->N == n) {
2461: PetscCall(VecGetArrayWrite(v, &diagA));
2462: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2463: PetscCall(MatGetRowMax(mat->A, diagV, idx));
2464: PetscCall(VecDestroy(&diagV));
2465: PetscCall(VecRestoreArrayWrite(v, &diagA));
2466: PetscFunctionReturn(PETSC_SUCCESS);
2467: } else if (n == 0) {
2468: if (m) {
2469: PetscCall(VecGetArrayWrite(v, &a));
2470: for (r = 0; r < m; r++) {
2471: a[r] = PETSC_MIN_REAL;
2472: if (idx) idx[r] = -1;
2473: }
2474: PetscCall(VecRestoreArrayWrite(v, &a));
2475: }
2476: PetscFunctionReturn(PETSC_SUCCESS);
2477: }
2479: PetscCall(PetscMalloc2(m, &diagIdx, m, &offdiagIdx));
2480: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2481: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2482: PetscCall(MatGetRowMax(mat->A, diagV, diagIdx));
2484: /* Get offdiagIdx[] for implicit 0.0 */
2485: PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2486: ba = bav;
2487: bi = b->i;
2488: bj = b->j;
2489: PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2490: for (r = 0; r < m; r++) {
2491: ncols = bi[r + 1] - bi[r];
2492: if (ncols == A->cmap->N - n) { /* Brow is dense */
2493: offdiagA[r] = *ba;
2494: offdiagIdx[r] = cmap[0];
2495: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2496: offdiagA[r] = 0.0;
2498: /* Find first hole in the cmap */
2499: for (j = 0; j < ncols; j++) {
2500: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2501: if (col > j && j < cstart) {
2502: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2503: break;
2504: } else if (col > j + n && j >= cstart) {
2505: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2506: break;
2507: }
2508: }
2509: if (j == ncols && ncols < A->cmap->N - n) {
2510: /* a hole is outside compressed Bcols */
2511: if (ncols == 0) {
2512: if (cstart) {
2513: offdiagIdx[r] = 0;
2514: } else offdiagIdx[r] = cend;
2515: } else { /* ncols > 0 */
2516: offdiagIdx[r] = cmap[ncols - 1] + 1;
2517: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2518: }
2519: }
2520: }
2522: for (j = 0; j < ncols; j++) {
2523: if (PetscRealPart(offdiagA[r]) < PetscRealPart(*ba)) {
2524: offdiagA[r] = *ba;
2525: offdiagIdx[r] = cmap[*bj];
2526: }
2527: ba++;
2528: bj++;
2529: }
2530: }
2532: PetscCall(VecGetArrayWrite(v, &a));
2533: PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2534: for (r = 0; r < m; ++r) {
2535: if (PetscRealPart(diagA[r]) > PetscRealPart(offdiagA[r])) {
2536: a[r] = diagA[r];
2537: if (idx) idx[r] = cstart + diagIdx[r];
2538: } else if (PetscRealPart(diagA[r]) == PetscRealPart(offdiagA[r])) {
2539: a[r] = diagA[r];
2540: if (idx) {
2541: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2542: idx[r] = cstart + diagIdx[r];
2543: } else idx[r] = offdiagIdx[r];
2544: }
2545: } else {
2546: a[r] = offdiagA[r];
2547: if (idx) idx[r] = offdiagIdx[r];
2548: }
2549: }
2550: PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2551: PetscCall(VecRestoreArrayWrite(v, &a));
2552: PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2553: PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2554: PetscCall(VecDestroy(&diagV));
2555: PetscCall(VecDestroy(&offdiagV));
2556: PetscCall(PetscFree2(diagIdx, offdiagIdx));
2557: PetscFunctionReturn(PETSC_SUCCESS);
2558: }
2560: PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat mat, Mat *newmat)
2561: {
2562: Mat *dummy;
2564: PetscFunctionBegin;
2565: PetscCall(MatCreateSubMatrix_MPIAIJ_All(mat, MAT_DO_NOT_GET_VALUES, MAT_INITIAL_MATRIX, &dummy));
2566: *newmat = *dummy;
2567: PetscCall(PetscFree(dummy));
2568: PetscFunctionReturn(PETSC_SUCCESS);
2569: }
2571: static PetscErrorCode MatInvertBlockDiagonal_MPIAIJ(Mat A, const PetscScalar **values)
2572: {
2573: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2575: PetscFunctionBegin;
2576: PetscCall(MatInvertBlockDiagonal(a->A, values));
2577: A->factorerrortype = a->A->factorerrortype;
2578: PetscFunctionReturn(PETSC_SUCCESS);
2579: }
2581: static PetscErrorCode MatSetRandom_MPIAIJ(Mat x, PetscRandom rctx)
2582: {
2583: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)x->data;
2585: PetscFunctionBegin;
2586: PetscCheck(x->assembled || x->preallocated, PetscObjectComm((PetscObject)x), PETSC_ERR_ARG_WRONGSTATE, "MatSetRandom on an unassembled and unpreallocated MATMPIAIJ is not allowed");
2587: PetscCall(MatSetRandom(aij->A, rctx));
2588: if (x->assembled) {
2589: PetscCall(MatSetRandom(aij->B, rctx));
2590: } else {
2591: PetscCall(MatSetRandomSkipColumnRange_SeqAIJ_Private(aij->B, x->cmap->rstart, x->cmap->rend, rctx));
2592: }
2593: PetscCall(MatAssemblyBegin(x, MAT_FINAL_ASSEMBLY));
2594: PetscCall(MatAssemblyEnd(x, MAT_FINAL_ASSEMBLY));
2595: PetscFunctionReturn(PETSC_SUCCESS);
2596: }
2598: static PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ(Mat A, PetscBool sc)
2599: {
2600: PetscFunctionBegin;
2601: if (sc) A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ_Scalable;
2602: else A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ;
2603: PetscFunctionReturn(PETSC_SUCCESS);
2604: }
2606: /*@
2607: MatMPIAIJGetNumberNonzeros - gets the number of nonzeros in the matrix on this MPI rank
2609: Not Collective
2611: Input Parameter:
2612: . A - the matrix
2614: Output Parameter:
2615: . nz - the number of nonzeros
2617: Level: advanced
2619: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`
2620: @*/
2621: PetscErrorCode MatMPIAIJGetNumberNonzeros(Mat A, PetscCount *nz)
2622: {
2623: Mat_MPIAIJ *maij = (Mat_MPIAIJ *)A->data;
2624: Mat_SeqAIJ *aaij = (Mat_SeqAIJ *)maij->A->data, *baij = (Mat_SeqAIJ *)maij->B->data;
2625: PetscBool isaij;
2627: PetscFunctionBegin;
2628: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATMPIAIJ, &isaij));
2629: PetscCheck(isaij, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Not for type %s", ((PetscObject)A)->type_name);
2630: *nz = aaij->i[A->rmap->n] + baij->i[A->rmap->n];
2631: PetscFunctionReturn(PETSC_SUCCESS);
2632: }
2634: /*@
2635: MatMPIAIJSetUseScalableIncreaseOverlap - Determine if the matrix uses a scalable algorithm to compute the overlap
2637: Collective
2639: Input Parameters:
2640: + A - the matrix
2641: - sc - `PETSC_TRUE` indicates use the scalable algorithm (default is not to use the scalable algorithm)
2643: Level: advanced
2645: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`
2646: @*/
2647: PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap(Mat A, PetscBool sc)
2648: {
2649: PetscFunctionBegin;
2650: PetscTryMethod(A, "MatMPIAIJSetUseScalableIncreaseOverlap_C", (Mat, PetscBool), (A, sc));
2651: PetscFunctionReturn(PETSC_SUCCESS);
2652: }
2654: PetscErrorCode MatSetFromOptions_MPIAIJ(Mat A, PetscOptionItems *PetscOptionsObject)
2655: {
2656: PetscBool sc = PETSC_FALSE, flg;
2658: PetscFunctionBegin;
2659: PetscOptionsHeadBegin(PetscOptionsObject, "MPIAIJ options");
2660: if (A->ops->increaseoverlap == MatIncreaseOverlap_MPIAIJ_Scalable) sc = PETSC_TRUE;
2661: PetscCall(PetscOptionsBool("-mat_increase_overlap_scalable", "Use a scalable algorithm to compute the overlap", "MatIncreaseOverlap", sc, &sc, &flg));
2662: if (flg) PetscCall(MatMPIAIJSetUseScalableIncreaseOverlap(A, sc));
2663: PetscOptionsHeadEnd();
2664: PetscFunctionReturn(PETSC_SUCCESS);
2665: }
2667: static PetscErrorCode MatShift_MPIAIJ(Mat Y, PetscScalar a)
2668: {
2669: Mat_MPIAIJ *maij = (Mat_MPIAIJ *)Y->data;
2670: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)maij->A->data;
2672: PetscFunctionBegin;
2673: if (!Y->preallocated) {
2674: PetscCall(MatMPIAIJSetPreallocation(Y, 1, NULL, 0, NULL));
2675: } else if (!aij->nz) { /* It does not matter if diagonals of Y only partially lie in maij->A. We just need an estimated preallocation. */
2676: PetscInt nonew = aij->nonew;
2677: PetscCall(MatSeqAIJSetPreallocation(maij->A, 1, NULL));
2678: aij->nonew = nonew;
2679: }
2680: PetscCall(MatShift_Basic(Y, a));
2681: PetscFunctionReturn(PETSC_SUCCESS);
2682: }
2684: static PetscErrorCode MatMissingDiagonal_MPIAIJ(Mat A, PetscBool *missing, PetscInt *d)
2685: {
2686: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2688: PetscFunctionBegin;
2689: PetscCheck(A->rmap->n == A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_SUP, "Only works for square matrices");
2690: PetscCall(MatMissingDiagonal(a->A, missing, d));
2691: if (d) {
2692: PetscInt rstart;
2693: PetscCall(MatGetOwnershipRange(A, &rstart, NULL));
2694: *d += rstart;
2695: }
2696: PetscFunctionReturn(PETSC_SUCCESS);
2697: }
2699: static PetscErrorCode MatInvertVariableBlockDiagonal_MPIAIJ(Mat A, PetscInt nblocks, const PetscInt *bsizes, PetscScalar *diag)
2700: {
2701: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2703: PetscFunctionBegin;
2704: PetscCall(MatInvertVariableBlockDiagonal(a->A, nblocks, bsizes, diag));
2705: PetscFunctionReturn(PETSC_SUCCESS);
2706: }
2708: static PetscErrorCode MatEliminateZeros_MPIAIJ(Mat A, PetscBool keep)
2709: {
2710: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2712: PetscFunctionBegin;
2713: PetscCall(MatEliminateZeros_SeqAIJ(a->A, keep)); // possibly keep zero diagonal coefficients
2714: PetscCall(MatEliminateZeros_SeqAIJ(a->B, PETSC_FALSE)); // never keep zero diagonal coefficients
2715: PetscFunctionReturn(PETSC_SUCCESS);
2716: }
2718: static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ,
2719: MatGetRow_MPIAIJ,
2720: MatRestoreRow_MPIAIJ,
2721: MatMult_MPIAIJ,
2722: /* 4*/ MatMultAdd_MPIAIJ,
2723: MatMultTranspose_MPIAIJ,
2724: MatMultTransposeAdd_MPIAIJ,
2725: NULL,
2726: NULL,
2727: NULL,
2728: /*10*/ NULL,
2729: NULL,
2730: NULL,
2731: MatSOR_MPIAIJ,
2732: MatTranspose_MPIAIJ,
2733: /*15*/ MatGetInfo_MPIAIJ,
2734: MatEqual_MPIAIJ,
2735: MatGetDiagonal_MPIAIJ,
2736: MatDiagonalScale_MPIAIJ,
2737: MatNorm_MPIAIJ,
2738: /*20*/ MatAssemblyBegin_MPIAIJ,
2739: MatAssemblyEnd_MPIAIJ,
2740: MatSetOption_MPIAIJ,
2741: MatZeroEntries_MPIAIJ,
2742: /*24*/ MatZeroRows_MPIAIJ,
2743: NULL,
2744: NULL,
2745: NULL,
2746: NULL,
2747: /*29*/ MatSetUp_MPI_Hash,
2748: NULL,
2749: NULL,
2750: MatGetDiagonalBlock_MPIAIJ,
2751: NULL,
2752: /*34*/ MatDuplicate_MPIAIJ,
2753: NULL,
2754: NULL,
2755: NULL,
2756: NULL,
2757: /*39*/ MatAXPY_MPIAIJ,
2758: MatCreateSubMatrices_MPIAIJ,
2759: MatIncreaseOverlap_MPIAIJ,
2760: MatGetValues_MPIAIJ,
2761: MatCopy_MPIAIJ,
2762: /*44*/ MatGetRowMax_MPIAIJ,
2763: MatScale_MPIAIJ,
2764: MatShift_MPIAIJ,
2765: MatDiagonalSet_MPIAIJ,
2766: MatZeroRowsColumns_MPIAIJ,
2767: /*49*/ MatSetRandom_MPIAIJ,
2768: MatGetRowIJ_MPIAIJ,
2769: MatRestoreRowIJ_MPIAIJ,
2770: NULL,
2771: NULL,
2772: /*54*/ MatFDColoringCreate_MPIXAIJ,
2773: NULL,
2774: MatSetUnfactored_MPIAIJ,
2775: MatPermute_MPIAIJ,
2776: NULL,
2777: /*59*/ MatCreateSubMatrix_MPIAIJ,
2778: MatDestroy_MPIAIJ,
2779: MatView_MPIAIJ,
2780: NULL,
2781: NULL,
2782: /*64*/ NULL,
2783: MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ,
2784: NULL,
2785: NULL,
2786: NULL,
2787: /*69*/ MatGetRowMaxAbs_MPIAIJ,
2788: MatGetRowMinAbs_MPIAIJ,
2789: NULL,
2790: NULL,
2791: NULL,
2792: NULL,
2793: /*75*/ MatFDColoringApply_AIJ,
2794: MatSetFromOptions_MPIAIJ,
2795: NULL,
2796: NULL,
2797: MatFindZeroDiagonals_MPIAIJ,
2798: /*80*/ NULL,
2799: NULL,
2800: NULL,
2801: /*83*/ MatLoad_MPIAIJ,
2802: NULL,
2803: NULL,
2804: NULL,
2805: NULL,
2806: NULL,
2807: /*89*/ NULL,
2808: NULL,
2809: MatMatMultNumeric_MPIAIJ_MPIAIJ,
2810: NULL,
2811: NULL,
2812: /*94*/ MatPtAPNumeric_MPIAIJ_MPIAIJ,
2813: NULL,
2814: NULL,
2815: NULL,
2816: MatBindToCPU_MPIAIJ,
2817: /*99*/ MatProductSetFromOptions_MPIAIJ,
2818: NULL,
2819: NULL,
2820: MatConjugate_MPIAIJ,
2821: NULL,
2822: /*104*/ MatSetValuesRow_MPIAIJ,
2823: MatRealPart_MPIAIJ,
2824: MatImaginaryPart_MPIAIJ,
2825: NULL,
2826: NULL,
2827: /*109*/ NULL,
2828: NULL,
2829: MatGetRowMin_MPIAIJ,
2830: NULL,
2831: MatMissingDiagonal_MPIAIJ,
2832: /*114*/ MatGetSeqNonzeroStructure_MPIAIJ,
2833: NULL,
2834: MatGetGhosts_MPIAIJ,
2835: NULL,
2836: NULL,
2837: /*119*/ MatMultDiagonalBlock_MPIAIJ,
2838: NULL,
2839: NULL,
2840: NULL,
2841: MatGetMultiProcBlock_MPIAIJ,
2842: /*124*/ MatFindNonzeroRows_MPIAIJ,
2843: MatGetColumnReductions_MPIAIJ,
2844: MatInvertBlockDiagonal_MPIAIJ,
2845: MatInvertVariableBlockDiagonal_MPIAIJ,
2846: MatCreateSubMatricesMPI_MPIAIJ,
2847: /*129*/ NULL,
2848: NULL,
2849: NULL,
2850: MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ,
2851: NULL,
2852: /*134*/ NULL,
2853: NULL,
2854: NULL,
2855: NULL,
2856: NULL,
2857: /*139*/ MatSetBlockSizes_MPIAIJ,
2858: NULL,
2859: NULL,
2860: MatFDColoringSetUp_MPIXAIJ,
2861: MatFindOffBlockDiagonalEntries_MPIAIJ,
2862: MatCreateMPIMatConcatenateSeqMat_MPIAIJ,
2863: /*145*/ NULL,
2864: NULL,
2865: NULL,
2866: MatCreateGraph_Simple_AIJ,
2867: NULL,
2868: /*150*/ NULL,
2869: MatEliminateZeros_MPIAIJ,
2870: MatGetRowSumAbs_MPIAIJ,
2871: NULL,
2872: NULL,
2873: NULL};
2875: static PetscErrorCode MatStoreValues_MPIAIJ(Mat mat)
2876: {
2877: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2879: PetscFunctionBegin;
2880: PetscCall(MatStoreValues(aij->A));
2881: PetscCall(MatStoreValues(aij->B));
2882: PetscFunctionReturn(PETSC_SUCCESS);
2883: }
2885: static PetscErrorCode MatRetrieveValues_MPIAIJ(Mat mat)
2886: {
2887: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2889: PetscFunctionBegin;
2890: PetscCall(MatRetrieveValues(aij->A));
2891: PetscCall(MatRetrieveValues(aij->B));
2892: PetscFunctionReturn(PETSC_SUCCESS);
2893: }
2895: PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat B, PetscInt d_nz, const PetscInt d_nnz[], PetscInt o_nz, const PetscInt o_nnz[])
2896: {
2897: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2898: PetscMPIInt size;
2900: PetscFunctionBegin;
2901: if (B->hash_active) {
2902: B->ops[0] = b->cops;
2903: B->hash_active = PETSC_FALSE;
2904: }
2905: PetscCall(PetscLayoutSetUp(B->rmap));
2906: PetscCall(PetscLayoutSetUp(B->cmap));
2908: #if defined(PETSC_USE_CTABLE)
2909: PetscCall(PetscHMapIDestroy(&b->colmap));
2910: #else
2911: PetscCall(PetscFree(b->colmap));
2912: #endif
2913: PetscCall(PetscFree(b->garray));
2914: PetscCall(VecDestroy(&b->lvec));
2915: PetscCall(VecScatterDestroy(&b->Mvctx));
2917: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
2919: MatSeqXAIJGetOptions_Private(b->B);
2920: PetscCall(MatDestroy(&b->B));
2921: PetscCall(MatCreate(PETSC_COMM_SELF, &b->B));
2922: PetscCall(MatSetSizes(b->B, B->rmap->n, size > 1 ? B->cmap->N : 0, B->rmap->n, size > 1 ? B->cmap->N : 0));
2923: PetscCall(MatSetBlockSizesFromMats(b->B, B, B));
2924: PetscCall(MatSetType(b->B, MATSEQAIJ));
2925: MatSeqXAIJRestoreOptions_Private(b->B);
2927: MatSeqXAIJGetOptions_Private(b->A);
2928: PetscCall(MatDestroy(&b->A));
2929: PetscCall(MatCreate(PETSC_COMM_SELF, &b->A));
2930: PetscCall(MatSetSizes(b->A, B->rmap->n, B->cmap->n, B->rmap->n, B->cmap->n));
2931: PetscCall(MatSetBlockSizesFromMats(b->A, B, B));
2932: PetscCall(MatSetType(b->A, MATSEQAIJ));
2933: MatSeqXAIJRestoreOptions_Private(b->A);
2935: PetscCall(MatSeqAIJSetPreallocation(b->A, d_nz, d_nnz));
2936: PetscCall(MatSeqAIJSetPreallocation(b->B, o_nz, o_nnz));
2937: B->preallocated = PETSC_TRUE;
2938: B->was_assembled = PETSC_FALSE;
2939: B->assembled = PETSC_FALSE;
2940: PetscFunctionReturn(PETSC_SUCCESS);
2941: }
2943: static PetscErrorCode MatResetPreallocation_MPIAIJ(Mat B)
2944: {
2945: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2946: PetscBool ondiagreset, offdiagreset, memoryreset;
2948: PetscFunctionBegin;
2950: PetscCheck(B->insertmode == NOT_SET_VALUES, PETSC_COMM_SELF, PETSC_ERR_SUP, "Cannot reset preallocation after setting some values but not yet calling MatAssemblyBegin()/MatAssemblyEnd()");
2951: if (B->num_ass == 0) PetscFunctionReturn(PETSC_SUCCESS);
2953: PetscCall(MatResetPreallocation_SeqAIJ_Private(b->A, &ondiagreset));
2954: PetscCall(MatResetPreallocation_SeqAIJ_Private(b->B, &offdiagreset));
2955: memoryreset = (PetscBool)(ondiagreset || offdiagreset);
2956: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &memoryreset, 1, MPIU_BOOL, MPI_LOR, PetscObjectComm((PetscObject)B)));
2957: if (!memoryreset) PetscFunctionReturn(PETSC_SUCCESS);
2959: PetscCall(PetscLayoutSetUp(B->rmap));
2960: PetscCall(PetscLayoutSetUp(B->cmap));
2961: PetscCheck(B->assembled || B->was_assembled, PetscObjectComm((PetscObject)B), PETSC_ERR_ARG_WRONGSTATE, "Should not need to reset preallocation if the matrix was never assembled");
2962: PetscCall(MatDisAssemble_MPIAIJ(B, PETSC_TRUE));
2963: PetscCall(VecScatterDestroy(&b->Mvctx));
2965: B->preallocated = PETSC_TRUE;
2966: B->was_assembled = PETSC_FALSE;
2967: B->assembled = PETSC_FALSE;
2968: /* Log that the state of this object has changed; this will help guarantee that preconditioners get re-setup */
2969: PetscCall(PetscObjectStateIncrease((PetscObject)B));
2970: PetscFunctionReturn(PETSC_SUCCESS);
2971: }
2973: PetscErrorCode MatDuplicate_MPIAIJ(Mat matin, MatDuplicateOption cpvalues, Mat *newmat)
2974: {
2975: Mat mat;
2976: Mat_MPIAIJ *a, *oldmat = (Mat_MPIAIJ *)matin->data;
2978: PetscFunctionBegin;
2979: *newmat = NULL;
2980: PetscCall(MatCreate(PetscObjectComm((PetscObject)matin), &mat));
2981: PetscCall(MatSetSizes(mat, matin->rmap->n, matin->cmap->n, matin->rmap->N, matin->cmap->N));
2982: PetscCall(MatSetBlockSizesFromMats(mat, matin, matin));
2983: PetscCall(MatSetType(mat, ((PetscObject)matin)->type_name));
2984: a = (Mat_MPIAIJ *)mat->data;
2986: mat->factortype = matin->factortype;
2987: mat->assembled = matin->assembled;
2988: mat->insertmode = NOT_SET_VALUES;
2990: a->size = oldmat->size;
2991: a->rank = oldmat->rank;
2992: a->donotstash = oldmat->donotstash;
2993: a->roworiented = oldmat->roworiented;
2994: a->rowindices = NULL;
2995: a->rowvalues = NULL;
2996: a->getrowactive = PETSC_FALSE;
2998: PetscCall(PetscLayoutReference(matin->rmap, &mat->rmap));
2999: PetscCall(PetscLayoutReference(matin->cmap, &mat->cmap));
3000: if (matin->hash_active) {
3001: PetscCall(MatSetUp(mat));
3002: } else {
3003: mat->preallocated = matin->preallocated;
3004: if (oldmat->colmap) {
3005: #if defined(PETSC_USE_CTABLE)
3006: PetscCall(PetscHMapIDuplicate(oldmat->colmap, &a->colmap));
3007: #else
3008: PetscCall(PetscMalloc1(mat->cmap->N, &a->colmap));
3009: PetscCall(PetscArraycpy(a->colmap, oldmat->colmap, mat->cmap->N));
3010: #endif
3011: } else a->colmap = NULL;
3012: if (oldmat->garray) {
3013: PetscInt len;
3014: len = oldmat->B->cmap->n;
3015: PetscCall(PetscMalloc1(len + 1, &a->garray));
3016: if (len) PetscCall(PetscArraycpy(a->garray, oldmat->garray, len));
3017: } else a->garray = NULL;
3019: /* It may happen MatDuplicate is called with a non-assembled matrix
3020: In fact, MatDuplicate only requires the matrix to be preallocated
3021: This may happen inside a DMCreateMatrix_Shell */
3022: if (oldmat->lvec) PetscCall(VecDuplicate(oldmat->lvec, &a->lvec));
3023: if (oldmat->Mvctx) {
3024: a->Mvctx = oldmat->Mvctx;
3025: PetscCall(PetscObjectReference((PetscObject)oldmat->Mvctx));
3026: }
3027: PetscCall(MatDuplicate(oldmat->A, cpvalues, &a->A));
3028: PetscCall(MatDuplicate(oldmat->B, cpvalues, &a->B));
3029: }
3030: PetscCall(PetscFunctionListDuplicate(((PetscObject)matin)->qlist, &((PetscObject)mat)->qlist));
3031: *newmat = mat;
3032: PetscFunctionReturn(PETSC_SUCCESS);
3033: }
3035: PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer)
3036: {
3037: PetscBool isbinary, ishdf5;
3039: PetscFunctionBegin;
3042: /* force binary viewer to load .info file if it has not yet done so */
3043: PetscCall(PetscViewerSetUp(viewer));
3044: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
3045: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERHDF5, &ishdf5));
3046: if (isbinary) {
3047: PetscCall(MatLoad_MPIAIJ_Binary(newMat, viewer));
3048: } else if (ishdf5) {
3049: #if defined(PETSC_HAVE_HDF5)
3050: PetscCall(MatLoad_AIJ_HDF5(newMat, viewer));
3051: #else
3052: SETERRQ(PetscObjectComm((PetscObject)newMat), PETSC_ERR_SUP, "HDF5 not supported in this build.\nPlease reconfigure using --download-hdf5");
3053: #endif
3054: } else {
3055: 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);
3056: }
3057: PetscFunctionReturn(PETSC_SUCCESS);
3058: }
3060: PetscErrorCode MatLoad_MPIAIJ_Binary(Mat mat, PetscViewer viewer)
3061: {
3062: PetscInt header[4], M, N, m, nz, rows, cols, sum, i;
3063: PetscInt *rowidxs, *colidxs;
3064: PetscScalar *matvals;
3066: PetscFunctionBegin;
3067: PetscCall(PetscViewerSetUp(viewer));
3069: /* read in matrix header */
3070: PetscCall(PetscViewerBinaryRead(viewer, header, 4, NULL, PETSC_INT));
3071: PetscCheck(header[0] == MAT_FILE_CLASSID, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Not a matrix object in file");
3072: M = header[1];
3073: N = header[2];
3074: nz = header[3];
3075: PetscCheck(M >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix row size (%" PetscInt_FMT ") in file is negative", M);
3076: PetscCheck(N >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix column size (%" PetscInt_FMT ") in file is negative", N);
3077: PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Matrix stored in special format on disk, cannot load as MPIAIJ");
3079: /* set block sizes from the viewer's .info file */
3080: PetscCall(MatLoad_Binary_BlockSizes(mat, viewer));
3081: /* set global sizes if not set already */
3082: if (mat->rmap->N < 0) mat->rmap->N = M;
3083: if (mat->cmap->N < 0) mat->cmap->N = N;
3084: PetscCall(PetscLayoutSetUp(mat->rmap));
3085: PetscCall(PetscLayoutSetUp(mat->cmap));
3087: /* check if the matrix sizes are correct */
3088: PetscCall(MatGetSize(mat, &rows, &cols));
3089: 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);
3091: /* read in row lengths and build row indices */
3092: PetscCall(MatGetLocalSize(mat, &m, NULL));
3093: PetscCall(PetscMalloc1(m + 1, &rowidxs));
3094: PetscCall(PetscViewerBinaryReadAll(viewer, rowidxs + 1, m, PETSC_DECIDE, M, PETSC_INT));
3095: rowidxs[0] = 0;
3096: for (i = 0; i < m; i++) rowidxs[i + 1] += rowidxs[i];
3097: if (nz != PETSC_INT_MAX) {
3098: PetscCallMPI(MPIU_Allreduce(&rowidxs[m], &sum, 1, MPIU_INT, MPI_SUM, PetscObjectComm((PetscObject)viewer)));
3099: 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);
3100: }
3102: /* read in column indices and matrix values */
3103: PetscCall(PetscMalloc2(rowidxs[m], &colidxs, rowidxs[m], &matvals));
3104: PetscCall(PetscViewerBinaryReadAll(viewer, colidxs, rowidxs[m], PETSC_DETERMINE, PETSC_DETERMINE, PETSC_INT));
3105: PetscCall(PetscViewerBinaryReadAll(viewer, matvals, rowidxs[m], PETSC_DETERMINE, PETSC_DETERMINE, PETSC_SCALAR));
3106: /* store matrix indices and values */
3107: PetscCall(MatMPIAIJSetPreallocationCSR(mat, rowidxs, colidxs, matvals));
3108: PetscCall(PetscFree(rowidxs));
3109: PetscCall(PetscFree2(colidxs, matvals));
3110: PetscFunctionReturn(PETSC_SUCCESS);
3111: }
3113: /* Not scalable because of ISAllGather() unless getting all columns. */
3114: static PetscErrorCode ISGetSeqIS_Private(Mat mat, IS iscol, IS *isseq)
3115: {
3116: IS iscol_local;
3117: PetscBool isstride;
3118: PetscMPIInt gisstride = 0;
3120: PetscFunctionBegin;
3121: /* check if we are grabbing all columns*/
3122: PetscCall(PetscObjectTypeCompare((PetscObject)iscol, ISSTRIDE, &isstride));
3124: if (isstride) {
3125: PetscInt start, len, mstart, mlen;
3126: PetscCall(ISStrideGetInfo(iscol, &start, NULL));
3127: PetscCall(ISGetLocalSize(iscol, &len));
3128: PetscCall(MatGetOwnershipRangeColumn(mat, &mstart, &mlen));
3129: if (mstart == start && mlen - mstart == len) gisstride = 1;
3130: }
3132: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &gisstride, 1, MPI_INT, MPI_MIN, PetscObjectComm((PetscObject)mat)));
3133: if (gisstride) {
3134: PetscInt N;
3135: PetscCall(MatGetSize(mat, NULL, &N));
3136: PetscCall(ISCreateStride(PETSC_COMM_SELF, N, 0, 1, &iscol_local));
3137: PetscCall(ISSetIdentity(iscol_local));
3138: PetscCall(PetscInfo(mat, "Optimizing for obtaining all columns of the matrix; skipping ISAllGather()\n"));
3139: } else {
3140: PetscInt cbs;
3141: PetscCall(ISGetBlockSize(iscol, &cbs));
3142: PetscCall(ISAllGather(iscol, &iscol_local));
3143: PetscCall(ISSetBlockSize(iscol_local, cbs));
3144: }
3146: *isseq = iscol_local;
3147: PetscFunctionReturn(PETSC_SUCCESS);
3148: }
3150: /*
3151: Used by MatCreateSubMatrix_MPIAIJ_SameRowColDist() to avoid ISAllGather() and global size of iscol_local
3152: (see MatCreateSubMatrix_MPIAIJ_nonscalable)
3154: Input Parameters:
3155: + mat - matrix
3156: . isrow - parallel row index set; its local indices are a subset of local columns of `mat`,
3157: i.e., mat->rstart <= isrow[i] < mat->rend
3158: - iscol - parallel column index set; its local indices are a subset of local columns of `mat`,
3159: i.e., mat->cstart <= iscol[i] < mat->cend
3161: Output Parameters:
3162: + isrow_d - sequential row index set for retrieving mat->A
3163: . iscol_d - sequential column index set for retrieving mat->A
3164: . iscol_o - sequential column index set for retrieving mat->B
3165: - garray - column map; garray[i] indicates global location of iscol_o[i] in `iscol`
3166: */
3167: static PetscErrorCode ISGetSeqIS_SameColDist_Private(Mat mat, IS isrow, IS iscol, IS *isrow_d, IS *iscol_d, IS *iscol_o, const PetscInt *garray[])
3168: {
3169: Vec x, cmap;
3170: const PetscInt *is_idx;
3171: PetscScalar *xarray, *cmaparray;
3172: PetscInt ncols, isstart, *idx, m, rstart, *cmap1, count;
3173: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data;
3174: Mat B = a->B;
3175: Vec lvec = a->lvec, lcmap;
3176: PetscInt i, cstart, cend, Bn = B->cmap->N;
3177: MPI_Comm comm;
3178: VecScatter Mvctx = a->Mvctx;
3180: PetscFunctionBegin;
3181: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3182: PetscCall(ISGetLocalSize(iscol, &ncols));
3184: /* (1) iscol is a sub-column vector of mat, pad it with '-1.' to form a full vector x */
3185: PetscCall(MatCreateVecs(mat, &x, NULL));
3186: PetscCall(VecSet(x, -1.0));
3187: PetscCall(VecDuplicate(x, &cmap));
3188: PetscCall(VecSet(cmap, -1.0));
3190: /* Get start indices */
3191: PetscCallMPI(MPI_Scan(&ncols, &isstart, 1, MPIU_INT, MPI_SUM, comm));
3192: isstart -= ncols;
3193: PetscCall(MatGetOwnershipRangeColumn(mat, &cstart, &cend));
3195: PetscCall(ISGetIndices(iscol, &is_idx));
3196: PetscCall(VecGetArray(x, &xarray));
3197: PetscCall(VecGetArray(cmap, &cmaparray));
3198: PetscCall(PetscMalloc1(ncols, &idx));
3199: for (i = 0; i < ncols; i++) {
3200: xarray[is_idx[i] - cstart] = (PetscScalar)is_idx[i];
3201: cmaparray[is_idx[i] - cstart] = i + isstart; /* global index of iscol[i] */
3202: idx[i] = is_idx[i] - cstart; /* local index of iscol[i] */
3203: }
3204: PetscCall(VecRestoreArray(x, &xarray));
3205: PetscCall(VecRestoreArray(cmap, &cmaparray));
3206: PetscCall(ISRestoreIndices(iscol, &is_idx));
3208: /* Get iscol_d */
3209: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, iscol_d));
3210: PetscCall(ISGetBlockSize(iscol, &i));
3211: PetscCall(ISSetBlockSize(*iscol_d, i));
3213: /* Get isrow_d */
3214: PetscCall(ISGetLocalSize(isrow, &m));
3215: rstart = mat->rmap->rstart;
3216: PetscCall(PetscMalloc1(m, &idx));
3217: PetscCall(ISGetIndices(isrow, &is_idx));
3218: for (i = 0; i < m; i++) idx[i] = is_idx[i] - rstart;
3219: PetscCall(ISRestoreIndices(isrow, &is_idx));
3221: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, m, idx, PETSC_OWN_POINTER, isrow_d));
3222: PetscCall(ISGetBlockSize(isrow, &i));
3223: PetscCall(ISSetBlockSize(*isrow_d, i));
3225: /* (2) Scatter x and cmap using aij->Mvctx to get their off-process portions (see MatMult_MPIAIJ) */
3226: PetscCall(VecScatterBegin(Mvctx, x, lvec, INSERT_VALUES, SCATTER_FORWARD));
3227: PetscCall(VecScatterEnd(Mvctx, x, lvec, INSERT_VALUES, SCATTER_FORWARD));
3229: PetscCall(VecDuplicate(lvec, &lcmap));
3231: PetscCall(VecScatterBegin(Mvctx, cmap, lcmap, INSERT_VALUES, SCATTER_FORWARD));
3232: PetscCall(VecScatterEnd(Mvctx, cmap, lcmap, INSERT_VALUES, SCATTER_FORWARD));
3234: /* (3) create sequential iscol_o (a subset of iscol) and isgarray */
3235: /* off-process column indices */
3236: count = 0;
3237: PetscCall(PetscMalloc1(Bn, &idx));
3238: PetscCall(PetscMalloc1(Bn, &cmap1));
3240: PetscCall(VecGetArray(lvec, &xarray));
3241: PetscCall(VecGetArray(lcmap, &cmaparray));
3242: for (i = 0; i < Bn; i++) {
3243: if (PetscRealPart(xarray[i]) > -1.0) {
3244: idx[count] = i; /* local column index in off-diagonal part B */
3245: cmap1[count] = (PetscInt)PetscRealPart(cmaparray[i]); /* column index in submat */
3246: count++;
3247: }
3248: }
3249: PetscCall(VecRestoreArray(lvec, &xarray));
3250: PetscCall(VecRestoreArray(lcmap, &cmaparray));
3252: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, count, idx, PETSC_COPY_VALUES, iscol_o));
3253: /* cannot ensure iscol_o has same blocksize as iscol! */
3255: PetscCall(PetscFree(idx));
3256: *garray = cmap1;
3258: PetscCall(VecDestroy(&x));
3259: PetscCall(VecDestroy(&cmap));
3260: PetscCall(VecDestroy(&lcmap));
3261: PetscFunctionReturn(PETSC_SUCCESS);
3262: }
3264: /* isrow and iscol have same processor distribution as mat, output *submat is a submatrix of local mat */
3265: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowColDist(Mat mat, IS isrow, IS iscol, MatReuse call, Mat *submat)
3266: {
3267: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data, *asub;
3268: Mat M = NULL;
3269: MPI_Comm comm;
3270: IS iscol_d, isrow_d, iscol_o;
3271: Mat Asub = NULL, Bsub = NULL;
3272: PetscInt n;
3274: PetscFunctionBegin;
3275: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3277: if (call == MAT_REUSE_MATRIX) {
3278: /* Retrieve isrow_d, iscol_d and iscol_o from submat */
3279: PetscCall(PetscObjectQuery((PetscObject)*submat, "isrow_d", (PetscObject *)&isrow_d));
3280: PetscCheck(isrow_d, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "isrow_d passed in was not used before, cannot reuse");
3282: PetscCall(PetscObjectQuery((PetscObject)*submat, "iscol_d", (PetscObject *)&iscol_d));
3283: PetscCheck(iscol_d, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "iscol_d passed in was not used before, cannot reuse");
3285: PetscCall(PetscObjectQuery((PetscObject)*submat, "iscol_o", (PetscObject *)&iscol_o));
3286: PetscCheck(iscol_o, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "iscol_o passed in was not used before, cannot reuse");
3288: /* Update diagonal and off-diagonal portions of submat */
3289: asub = (Mat_MPIAIJ *)(*submat)->data;
3290: PetscCall(MatCreateSubMatrix_SeqAIJ(a->A, isrow_d, iscol_d, PETSC_DECIDE, MAT_REUSE_MATRIX, &asub->A));
3291: PetscCall(ISGetLocalSize(iscol_o, &n));
3292: if (n) PetscCall(MatCreateSubMatrix_SeqAIJ(a->B, isrow_d, iscol_o, PETSC_DECIDE, MAT_REUSE_MATRIX, &asub->B));
3293: PetscCall(MatAssemblyBegin(*submat, MAT_FINAL_ASSEMBLY));
3294: PetscCall(MatAssemblyEnd(*submat, MAT_FINAL_ASSEMBLY));
3296: } else { /* call == MAT_INITIAL_MATRIX) */
3297: const PetscInt *garray;
3298: PetscInt BsubN;
3300: /* Create isrow_d, iscol_d, iscol_o and isgarray (replace isgarray with array?) */
3301: PetscCall(ISGetSeqIS_SameColDist_Private(mat, isrow, iscol, &isrow_d, &iscol_d, &iscol_o, &garray));
3303: /* Create local submatrices Asub and Bsub */
3304: PetscCall(MatCreateSubMatrix_SeqAIJ(a->A, isrow_d, iscol_d, PETSC_DECIDE, MAT_INITIAL_MATRIX, &Asub));
3305: PetscCall(MatCreateSubMatrix_SeqAIJ(a->B, isrow_d, iscol_o, PETSC_DECIDE, MAT_INITIAL_MATRIX, &Bsub));
3307: /* Create submatrix M */
3308: PetscCall(MatCreateMPIAIJWithSeqAIJ(comm, Asub, Bsub, garray, &M));
3310: /* If Bsub has empty columns, compress iscol_o such that it will retrieve condensed Bsub from a->B during reuse */
3311: asub = (Mat_MPIAIJ *)M->data;
3313: PetscCall(ISGetLocalSize(iscol_o, &BsubN));
3314: n = asub->B->cmap->N;
3315: if (BsubN > n) {
3316: /* This case can be tested using ~petsc/src/tao/bound/tutorials/runplate2_3 */
3317: const PetscInt *idx;
3318: PetscInt i, j, *idx_new, *subgarray = asub->garray;
3319: PetscCall(PetscInfo(M, "submatrix Bn %" PetscInt_FMT " != BsubN %" PetscInt_FMT ", update iscol_o\n", n, BsubN));
3321: PetscCall(PetscMalloc1(n, &idx_new));
3322: j = 0;
3323: PetscCall(ISGetIndices(iscol_o, &idx));
3324: for (i = 0; i < n; i++) {
3325: if (j >= BsubN) break;
3326: while (subgarray[i] > garray[j]) j++;
3328: if (subgarray[i] == garray[j]) {
3329: idx_new[i] = idx[j++];
3330: } 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]);
3331: }
3332: PetscCall(ISRestoreIndices(iscol_o, &idx));
3334: PetscCall(ISDestroy(&iscol_o));
3335: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, n, idx_new, PETSC_OWN_POINTER, &iscol_o));
3337: } else if (BsubN < n) {
3338: 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);
3339: }
3341: PetscCall(PetscFree(garray));
3342: *submat = M;
3344: /* Save isrow_d, iscol_d and iscol_o used in processor for next request */
3345: PetscCall(PetscObjectCompose((PetscObject)M, "isrow_d", (PetscObject)isrow_d));
3346: PetscCall(ISDestroy(&isrow_d));
3348: PetscCall(PetscObjectCompose((PetscObject)M, "iscol_d", (PetscObject)iscol_d));
3349: PetscCall(ISDestroy(&iscol_d));
3351: PetscCall(PetscObjectCompose((PetscObject)M, "iscol_o", (PetscObject)iscol_o));
3352: PetscCall(ISDestroy(&iscol_o));
3353: }
3354: PetscFunctionReturn(PETSC_SUCCESS);
3355: }
3357: PetscErrorCode MatCreateSubMatrix_MPIAIJ(Mat mat, IS isrow, IS iscol, MatReuse call, Mat *newmat)
3358: {
3359: IS iscol_local = NULL, isrow_d;
3360: PetscInt csize;
3361: PetscInt n, i, j, start, end;
3362: PetscBool sameRowDist = PETSC_FALSE, sameDist[2], tsameDist[2];
3363: MPI_Comm comm;
3365: PetscFunctionBegin;
3366: /* If isrow has same processor distribution as mat,
3367: call MatCreateSubMatrix_MPIAIJ_SameRowDist() to avoid using a hash table with global size of iscol */
3368: if (call == MAT_REUSE_MATRIX) {
3369: PetscCall(PetscObjectQuery((PetscObject)*newmat, "isrow_d", (PetscObject *)&isrow_d));
3370: if (isrow_d) {
3371: sameRowDist = PETSC_TRUE;
3372: tsameDist[1] = PETSC_TRUE; /* sameColDist */
3373: } else {
3374: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_local));
3375: if (iscol_local) {
3376: sameRowDist = PETSC_TRUE;
3377: tsameDist[1] = PETSC_FALSE; /* !sameColDist */
3378: }
3379: }
3380: } else {
3381: /* Check if isrow has same processor distribution as mat */
3382: sameDist[0] = PETSC_FALSE;
3383: PetscCall(ISGetLocalSize(isrow, &n));
3384: if (!n) {
3385: sameDist[0] = PETSC_TRUE;
3386: } else {
3387: PetscCall(ISGetMinMax(isrow, &i, &j));
3388: PetscCall(MatGetOwnershipRange(mat, &start, &end));
3389: if (i >= start && j < end) sameDist[0] = PETSC_TRUE;
3390: }
3392: /* Check if iscol has same processor distribution as mat */
3393: sameDist[1] = PETSC_FALSE;
3394: PetscCall(ISGetLocalSize(iscol, &n));
3395: if (!n) {
3396: sameDist[1] = PETSC_TRUE;
3397: } else {
3398: PetscCall(ISGetMinMax(iscol, &i, &j));
3399: PetscCall(MatGetOwnershipRangeColumn(mat, &start, &end));
3400: if (i >= start && j < end) sameDist[1] = PETSC_TRUE;
3401: }
3403: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3404: PetscCallMPI(MPIU_Allreduce(&sameDist, &tsameDist, 2, MPIU_BOOL, MPI_LAND, comm));
3405: sameRowDist = tsameDist[0];
3406: }
3408: if (sameRowDist) {
3409: if (tsameDist[1]) { /* sameRowDist & sameColDist */
3410: /* isrow and iscol have same processor distribution as mat */
3411: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowColDist(mat, isrow, iscol, call, newmat));
3412: PetscFunctionReturn(PETSC_SUCCESS);
3413: } else { /* sameRowDist */
3414: /* isrow has same processor distribution as mat */
3415: if (call == MAT_INITIAL_MATRIX) {
3416: PetscBool sorted;
3417: PetscCall(ISGetSeqIS_Private(mat, iscol, &iscol_local));
3418: PetscCall(ISGetLocalSize(iscol_local, &n)); /* local size of iscol_local = global columns of newmat */
3419: PetscCall(ISGetSize(iscol, &i));
3420: PetscCheck(n == i, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "n %" PetscInt_FMT " != size of iscol %" PetscInt_FMT, n, i);
3422: PetscCall(ISSorted(iscol_local, &sorted));
3423: if (sorted) {
3424: /* MatCreateSubMatrix_MPIAIJ_SameRowDist() requires iscol_local be sorted; it can have duplicate indices */
3425: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowDist(mat, isrow, iscol, iscol_local, MAT_INITIAL_MATRIX, newmat));
3426: PetscFunctionReturn(PETSC_SUCCESS);
3427: }
3428: } else { /* call == MAT_REUSE_MATRIX */
3429: IS iscol_sub;
3430: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_sub));
3431: if (iscol_sub) {
3432: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowDist(mat, isrow, iscol, NULL, call, newmat));
3433: PetscFunctionReturn(PETSC_SUCCESS);
3434: }
3435: }
3436: }
3437: }
3439: /* General case: iscol -> iscol_local which has global size of iscol */
3440: if (call == MAT_REUSE_MATRIX) {
3441: PetscCall(PetscObjectQuery((PetscObject)*newmat, "ISAllGather", (PetscObject *)&iscol_local));
3442: PetscCheck(iscol_local, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3443: } else {
3444: if (!iscol_local) PetscCall(ISGetSeqIS_Private(mat, iscol, &iscol_local));
3445: }
3447: PetscCall(ISGetLocalSize(iscol, &csize));
3448: PetscCall(MatCreateSubMatrix_MPIAIJ_nonscalable(mat, isrow, iscol_local, csize, call, newmat));
3450: if (call == MAT_INITIAL_MATRIX) {
3451: PetscCall(PetscObjectCompose((PetscObject)*newmat, "ISAllGather", (PetscObject)iscol_local));
3452: PetscCall(ISDestroy(&iscol_local));
3453: }
3454: PetscFunctionReturn(PETSC_SUCCESS);
3455: }
3457: /*@C
3458: MatCreateMPIAIJWithSeqAIJ - creates a `MATMPIAIJ` matrix using `MATSEQAIJ` matrices that contain the "diagonal"
3459: and "off-diagonal" part of the matrix in CSR format.
3461: Collective
3463: Input Parameters:
3464: + comm - MPI communicator
3465: . A - "diagonal" portion of matrix
3466: . B - "off-diagonal" portion of matrix, may have empty columns, will be destroyed by this routine
3467: - garray - global index of `B` columns
3469: Output Parameter:
3470: . mat - the matrix, with input `A` as its local diagonal matrix
3472: Level: advanced
3474: Notes:
3475: See `MatCreateAIJ()` for the definition of "diagonal" and "off-diagonal" portion of the matrix.
3477: `A` becomes part of output mat, `B` is destroyed by this routine. The user cannot use `A` and `B` anymore.
3479: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MATSEQAIJ`, `MatCreateMPIAIJWithSplitArrays()`
3480: @*/
3481: PetscErrorCode MatCreateMPIAIJWithSeqAIJ(MPI_Comm comm, Mat A, Mat B, const PetscInt garray[], Mat *mat)
3482: {
3483: Mat_MPIAIJ *maij;
3484: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data, *bnew;
3485: PetscInt *oi = b->i, *oj = b->j, i, nz, col;
3486: const PetscScalar *oa;
3487: Mat Bnew;
3488: PetscInt m, n, N;
3489: MatType mpi_mat_type;
3491: PetscFunctionBegin;
3492: PetscCall(MatCreate(comm, mat));
3493: PetscCall(MatGetSize(A, &m, &n));
3494: PetscCheck(m == B->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Am %" PetscInt_FMT " != Bm %" PetscInt_FMT, m, B->rmap->N);
3495: 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);
3496: /* remove check below; When B is created using iscol_o from ISGetSeqIS_SameColDist_Private(), its bs may not be same as A */
3497: /* 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); */
3499: /* Get global columns of mat */
3500: PetscCallMPI(MPIU_Allreduce(&n, &N, 1, MPIU_INT, MPI_SUM, comm));
3502: PetscCall(MatSetSizes(*mat, m, n, PETSC_DECIDE, N));
3503: /* Determine the type of MPI matrix that should be created from the type of matrix A, which holds the "diagonal" portion. */
3504: PetscCall(MatGetMPIMatType_Private(A, &mpi_mat_type));
3505: PetscCall(MatSetType(*mat, mpi_mat_type));
3507: if (A->rmap->bs > 1 || A->cmap->bs > 1) PetscCall(MatSetBlockSizes(*mat, A->rmap->bs, A->cmap->bs));
3508: maij = (Mat_MPIAIJ *)(*mat)->data;
3510: (*mat)->preallocated = PETSC_TRUE;
3512: PetscCall(PetscLayoutSetUp((*mat)->rmap));
3513: PetscCall(PetscLayoutSetUp((*mat)->cmap));
3515: /* Set A as diagonal portion of *mat */
3516: maij->A = A;
3518: nz = oi[m];
3519: for (i = 0; i < nz; i++) {
3520: col = oj[i];
3521: oj[i] = garray[col];
3522: }
3524: /* Set Bnew as off-diagonal portion of *mat */
3525: PetscCall(MatSeqAIJGetArrayRead(B, &oa));
3526: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, N, oi, oj, (PetscScalar *)oa, &Bnew));
3527: PetscCall(MatSeqAIJRestoreArrayRead(B, &oa));
3528: bnew = (Mat_SeqAIJ *)Bnew->data;
3529: bnew->maxnz = b->maxnz; /* allocated nonzeros of B */
3530: maij->B = Bnew;
3532: 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);
3534: b->free_a = PETSC_FALSE;
3535: b->free_ij = PETSC_FALSE;
3536: PetscCall(MatDestroy(&B));
3538: bnew->free_a = PETSC_TRUE;
3539: bnew->free_ij = PETSC_TRUE;
3541: /* condense columns of maij->B */
3542: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
3543: PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY));
3544: PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY));
3545: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_FALSE));
3546: PetscCall(MatSetOption(*mat, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
3547: PetscFunctionReturn(PETSC_SUCCESS);
3548: }
3550: extern PetscErrorCode MatCreateSubMatrices_MPIAIJ_SingleIS_Local(Mat, PetscInt, const IS[], const IS[], MatReuse, PetscBool, Mat *);
3552: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowDist(Mat mat, IS isrow, IS iscol, IS iscol_local, MatReuse call, Mat *newmat)
3553: {
3554: PetscInt i, m, n, rstart, row, rend, nz, j, bs, cbs;
3555: PetscInt *ii, *jj, nlocal, *dlens, *olens, dlen, olen, jend, mglobal;
3556: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data;
3557: Mat M, Msub, B = a->B;
3558: MatScalar *aa;
3559: Mat_SeqAIJ *aij;
3560: PetscInt *garray = a->garray, *colsub, Ncols;
3561: PetscInt count, Bn = B->cmap->N, cstart = mat->cmap->rstart, cend = mat->cmap->rend;
3562: IS iscol_sub, iscmap;
3563: const PetscInt *is_idx, *cmap;
3564: PetscBool allcolumns = PETSC_FALSE;
3565: MPI_Comm comm;
3567: PetscFunctionBegin;
3568: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3569: if (call == MAT_REUSE_MATRIX) {
3570: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_sub));
3571: PetscCheck(iscol_sub, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "SubIScol passed in was not used before, cannot reuse");
3572: PetscCall(ISGetLocalSize(iscol_sub, &count));
3574: PetscCall(PetscObjectQuery((PetscObject)*newmat, "Subcmap", (PetscObject *)&iscmap));
3575: PetscCheck(iscmap, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Subcmap passed in was not used before, cannot reuse");
3577: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubMatrix", (PetscObject *)&Msub));
3578: PetscCheck(Msub, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3580: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol_sub, MAT_REUSE_MATRIX, PETSC_FALSE, &Msub));
3582: } else { /* call == MAT_INITIAL_MATRIX) */
3583: PetscBool flg;
3585: PetscCall(ISGetLocalSize(iscol, &n));
3586: PetscCall(ISGetSize(iscol, &Ncols));
3588: /* (1) iscol -> nonscalable iscol_local */
3589: /* Check for special case: each processor gets entire matrix columns */
3590: PetscCall(ISIdentity(iscol_local, &flg));
3591: if (flg && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3592: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &allcolumns, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
3593: if (allcolumns) {
3594: iscol_sub = iscol_local;
3595: PetscCall(PetscObjectReference((PetscObject)iscol_local));
3596: PetscCall(ISCreateStride(PETSC_COMM_SELF, n, 0, 1, &iscmap));
3598: } else {
3599: /* (2) iscol_local -> iscol_sub and iscmap. Implementation below requires iscol_local be sorted, it can have duplicate indices */
3600: PetscInt *idx, *cmap1, k;
3601: PetscCall(PetscMalloc1(Ncols, &idx));
3602: PetscCall(PetscMalloc1(Ncols, &cmap1));
3603: PetscCall(ISGetIndices(iscol_local, &is_idx));
3604: count = 0;
3605: k = 0;
3606: for (i = 0; i < Ncols; i++) {
3607: j = is_idx[i];
3608: if (j >= cstart && j < cend) {
3609: /* diagonal part of mat */
3610: idx[count] = j;
3611: cmap1[count++] = i; /* column index in submat */
3612: } else if (Bn) {
3613: /* off-diagonal part of mat */
3614: if (j == garray[k]) {
3615: idx[count] = j;
3616: cmap1[count++] = i; /* column index in submat */
3617: } else if (j > garray[k]) {
3618: while (j > garray[k] && k < Bn - 1) k++;
3619: if (j == garray[k]) {
3620: idx[count] = j;
3621: cmap1[count++] = i; /* column index in submat */
3622: }
3623: }
3624: }
3625: }
3626: PetscCall(ISRestoreIndices(iscol_local, &is_idx));
3628: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, count, idx, PETSC_OWN_POINTER, &iscol_sub));
3629: PetscCall(ISGetBlockSize(iscol, &cbs));
3630: PetscCall(ISSetBlockSize(iscol_sub, cbs));
3632: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)iscol_local), count, cmap1, PETSC_OWN_POINTER, &iscmap));
3633: }
3635: /* (3) Create sequential Msub */
3636: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol_sub, MAT_INITIAL_MATRIX, allcolumns, &Msub));
3637: }
3639: PetscCall(ISGetLocalSize(iscol_sub, &count));
3640: aij = (Mat_SeqAIJ *)Msub->data;
3641: ii = aij->i;
3642: PetscCall(ISGetIndices(iscmap, &cmap));
3644: /*
3645: m - number of local rows
3646: Ncols - number of columns (same on all processors)
3647: rstart - first row in new global matrix generated
3648: */
3649: PetscCall(MatGetSize(Msub, &m, NULL));
3651: if (call == MAT_INITIAL_MATRIX) {
3652: /* (4) Create parallel newmat */
3653: PetscMPIInt rank, size;
3654: PetscInt csize;
3656: PetscCallMPI(MPI_Comm_size(comm, &size));
3657: PetscCallMPI(MPI_Comm_rank(comm, &rank));
3659: /*
3660: Determine the number of non-zeros in the diagonal and off-diagonal
3661: portions of the matrix in order to do correct preallocation
3662: */
3664: /* first get start and end of "diagonal" columns */
3665: PetscCall(ISGetLocalSize(iscol, &csize));
3666: if (csize == PETSC_DECIDE) {
3667: PetscCall(ISGetSize(isrow, &mglobal));
3668: if (mglobal == Ncols) { /* square matrix */
3669: nlocal = m;
3670: } else {
3671: nlocal = Ncols / size + ((Ncols % size) > rank);
3672: }
3673: } else {
3674: nlocal = csize;
3675: }
3676: PetscCallMPI(MPI_Scan(&nlocal, &rend, 1, MPIU_INT, MPI_SUM, comm));
3677: rstart = rend - nlocal;
3678: 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);
3680: /* next, compute all the lengths */
3681: jj = aij->j;
3682: PetscCall(PetscMalloc1(2 * m + 1, &dlens));
3683: olens = dlens + m;
3684: for (i = 0; i < m; i++) {
3685: jend = ii[i + 1] - ii[i];
3686: olen = 0;
3687: dlen = 0;
3688: for (j = 0; j < jend; j++) {
3689: if (cmap[*jj] < rstart || cmap[*jj] >= rend) olen++;
3690: else dlen++;
3691: jj++;
3692: }
3693: olens[i] = olen;
3694: dlens[i] = dlen;
3695: }
3697: PetscCall(ISGetBlockSize(isrow, &bs));
3698: PetscCall(ISGetBlockSize(iscol, &cbs));
3700: PetscCall(MatCreate(comm, &M));
3701: PetscCall(MatSetSizes(M, m, nlocal, PETSC_DECIDE, Ncols));
3702: PetscCall(MatSetBlockSizes(M, bs, cbs));
3703: PetscCall(MatSetType(M, ((PetscObject)mat)->type_name));
3704: PetscCall(MatMPIAIJSetPreallocation(M, 0, dlens, 0, olens));
3705: PetscCall(PetscFree(dlens));
3707: } else { /* call == MAT_REUSE_MATRIX */
3708: M = *newmat;
3709: PetscCall(MatGetLocalSize(M, &i, NULL));
3710: PetscCheck(i == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Previous matrix must be same size/layout as request");
3711: PetscCall(MatZeroEntries(M));
3712: /*
3713: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3714: rather than the slower MatSetValues().
3715: */
3716: M->was_assembled = PETSC_TRUE;
3717: M->assembled = PETSC_FALSE;
3718: }
3720: /* (5) Set values of Msub to *newmat */
3721: PetscCall(PetscMalloc1(count, &colsub));
3722: PetscCall(MatGetOwnershipRange(M, &rstart, NULL));
3724: jj = aij->j;
3725: PetscCall(MatSeqAIJGetArrayRead(Msub, (const PetscScalar **)&aa));
3726: for (i = 0; i < m; i++) {
3727: row = rstart + i;
3728: nz = ii[i + 1] - ii[i];
3729: for (j = 0; j < nz; j++) colsub[j] = cmap[jj[j]];
3730: PetscCall(MatSetValues_MPIAIJ(M, 1, &row, nz, colsub, aa, INSERT_VALUES));
3731: jj += nz;
3732: aa += nz;
3733: }
3734: PetscCall(MatSeqAIJRestoreArrayRead(Msub, (const PetscScalar **)&aa));
3735: PetscCall(ISRestoreIndices(iscmap, &cmap));
3737: PetscCall(MatAssemblyBegin(M, MAT_FINAL_ASSEMBLY));
3738: PetscCall(MatAssemblyEnd(M, MAT_FINAL_ASSEMBLY));
3740: PetscCall(PetscFree(colsub));
3742: /* save Msub, iscol_sub and iscmap used in processor for next request */
3743: if (call == MAT_INITIAL_MATRIX) {
3744: *newmat = M;
3745: PetscCall(PetscObjectCompose((PetscObject)*newmat, "SubMatrix", (PetscObject)Msub));
3746: PetscCall(MatDestroy(&Msub));
3748: PetscCall(PetscObjectCompose((PetscObject)*newmat, "SubIScol", (PetscObject)iscol_sub));
3749: PetscCall(ISDestroy(&iscol_sub));
3751: PetscCall(PetscObjectCompose((PetscObject)*newmat, "Subcmap", (PetscObject)iscmap));
3752: PetscCall(ISDestroy(&iscmap));
3754: if (iscol_local) {
3755: PetscCall(PetscObjectCompose((PetscObject)*newmat, "ISAllGather", (PetscObject)iscol_local));
3756: PetscCall(ISDestroy(&iscol_local));
3757: }
3758: }
3759: PetscFunctionReturn(PETSC_SUCCESS);
3760: }
3762: /*
3763: Not great since it makes two copies of the submatrix, first an SeqAIJ
3764: in local and then by concatenating the local matrices the end result.
3765: Writing it directly would be much like MatCreateSubMatrices_MPIAIJ()
3767: This requires a sequential iscol with all indices.
3768: */
3769: PetscErrorCode MatCreateSubMatrix_MPIAIJ_nonscalable(Mat mat, IS isrow, IS iscol, PetscInt csize, MatReuse call, Mat *newmat)
3770: {
3771: PetscMPIInt rank, size;
3772: PetscInt i, m, n, rstart, row, rend, nz, *cwork, j, bs, cbs;
3773: PetscInt *ii, *jj, nlocal, *dlens, *olens, dlen, olen, jend, mglobal;
3774: Mat M, Mreuse;
3775: MatScalar *aa, *vwork;
3776: MPI_Comm comm;
3777: Mat_SeqAIJ *aij;
3778: PetscBool colflag, allcolumns = PETSC_FALSE;
3780: PetscFunctionBegin;
3781: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3782: PetscCallMPI(MPI_Comm_rank(comm, &rank));
3783: PetscCallMPI(MPI_Comm_size(comm, &size));
3785: /* Check for special case: each processor gets entire matrix columns */
3786: PetscCall(ISIdentity(iscol, &colflag));
3787: PetscCall(ISGetLocalSize(iscol, &n));
3788: if (colflag && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3789: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &allcolumns, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
3791: if (call == MAT_REUSE_MATRIX) {
3792: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubMatrix", (PetscObject *)&Mreuse));
3793: PetscCheck(Mreuse, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3794: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol, MAT_REUSE_MATRIX, allcolumns, &Mreuse));
3795: } else {
3796: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol, MAT_INITIAL_MATRIX, allcolumns, &Mreuse));
3797: }
3799: /*
3800: m - number of local rows
3801: n - number of columns (same on all processors)
3802: rstart - first row in new global matrix generated
3803: */
3804: PetscCall(MatGetSize(Mreuse, &m, &n));
3805: PetscCall(MatGetBlockSizes(Mreuse, &bs, &cbs));
3806: if (call == MAT_INITIAL_MATRIX) {
3807: aij = (Mat_SeqAIJ *)Mreuse->data;
3808: ii = aij->i;
3809: jj = aij->j;
3811: /*
3812: Determine the number of non-zeros in the diagonal and off-diagonal
3813: portions of the matrix in order to do correct preallocation
3814: */
3816: /* first get start and end of "diagonal" columns */
3817: if (csize == PETSC_DECIDE) {
3818: PetscCall(ISGetSize(isrow, &mglobal));
3819: if (mglobal == n) { /* square matrix */
3820: nlocal = m;
3821: } else {
3822: nlocal = n / size + ((n % size) > rank);
3823: }
3824: } else {
3825: nlocal = csize;
3826: }
3827: PetscCallMPI(MPI_Scan(&nlocal, &rend, 1, MPIU_INT, MPI_SUM, comm));
3828: rstart = rend - nlocal;
3829: 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);
3831: /* next, compute all the lengths */
3832: PetscCall(PetscMalloc1(2 * m + 1, &dlens));
3833: olens = dlens + m;
3834: for (i = 0; i < m; i++) {
3835: jend = ii[i + 1] - ii[i];
3836: olen = 0;
3837: dlen = 0;
3838: for (j = 0; j < jend; j++) {
3839: if (*jj < rstart || *jj >= rend) olen++;
3840: else dlen++;
3841: jj++;
3842: }
3843: olens[i] = olen;
3844: dlens[i] = dlen;
3845: }
3846: PetscCall(MatCreate(comm, &M));
3847: PetscCall(MatSetSizes(M, m, nlocal, PETSC_DECIDE, n));
3848: PetscCall(MatSetBlockSizes(M, bs, cbs));
3849: PetscCall(MatSetType(M, ((PetscObject)mat)->type_name));
3850: PetscCall(MatMPIAIJSetPreallocation(M, 0, dlens, 0, olens));
3851: PetscCall(PetscFree(dlens));
3852: } else {
3853: PetscInt ml, nl;
3855: M = *newmat;
3856: PetscCall(MatGetLocalSize(M, &ml, &nl));
3857: PetscCheck(ml == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Previous matrix must be same size/layout as request");
3858: PetscCall(MatZeroEntries(M));
3859: /*
3860: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3861: rather than the slower MatSetValues().
3862: */
3863: M->was_assembled = PETSC_TRUE;
3864: M->assembled = PETSC_FALSE;
3865: }
3866: PetscCall(MatGetOwnershipRange(M, &rstart, &rend));
3867: aij = (Mat_SeqAIJ *)Mreuse->data;
3868: ii = aij->i;
3869: jj = aij->j;
3871: /* trigger copy to CPU if needed */
3872: PetscCall(MatSeqAIJGetArrayRead(Mreuse, (const PetscScalar **)&aa));
3873: for (i = 0; i < m; i++) {
3874: row = rstart + i;
3875: nz = ii[i + 1] - ii[i];
3876: cwork = jj;
3877: jj = PetscSafePointerPlusOffset(jj, nz);
3878: vwork = aa;
3879: aa = PetscSafePointerPlusOffset(aa, nz);
3880: PetscCall(MatSetValues_MPIAIJ(M, 1, &row, nz, cwork, vwork, INSERT_VALUES));
3881: }
3882: PetscCall(MatSeqAIJRestoreArrayRead(Mreuse, (const PetscScalar **)&aa));
3884: PetscCall(MatAssemblyBegin(M, MAT_FINAL_ASSEMBLY));
3885: PetscCall(MatAssemblyEnd(M, MAT_FINAL_ASSEMBLY));
3886: *newmat = M;
3888: /* save submatrix used in processor for next request */
3889: if (call == MAT_INITIAL_MATRIX) {
3890: PetscCall(PetscObjectCompose((PetscObject)M, "SubMatrix", (PetscObject)Mreuse));
3891: PetscCall(MatDestroy(&Mreuse));
3892: }
3893: PetscFunctionReturn(PETSC_SUCCESS);
3894: }
3896: static PetscErrorCode MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[])
3897: {
3898: PetscInt m, cstart, cend, j, nnz, i, d, *ld;
3899: PetscInt *d_nnz, *o_nnz, nnz_max = 0, rstart, ii, irstart;
3900: const PetscInt *JJ;
3901: PetscBool nooffprocentries;
3902: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)B->data;
3904: PetscFunctionBegin;
3905: PetscCall(PetscLayoutSetUp(B->rmap));
3906: PetscCall(PetscLayoutSetUp(B->cmap));
3907: m = B->rmap->n;
3908: cstart = B->cmap->rstart;
3909: cend = B->cmap->rend;
3910: rstart = B->rmap->rstart;
3911: irstart = Ii[0];
3913: PetscCall(PetscCalloc2(m, &d_nnz, m, &o_nnz));
3915: if (PetscDefined(USE_DEBUG)) {
3916: for (i = 0; i < m; i++) {
3917: nnz = Ii[i + 1] - Ii[i];
3918: JJ = PetscSafePointerPlusOffset(J, Ii[i] - irstart);
3919: PetscCheck(nnz >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Local row %" PetscInt_FMT " has a negative %" PetscInt_FMT " number of columns", i, nnz);
3920: 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]);
3921: 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);
3922: }
3923: }
3925: for (i = 0; i < m; i++) {
3926: nnz = Ii[i + 1] - Ii[i];
3927: JJ = PetscSafePointerPlusOffset(J, Ii[i] - irstart);
3928: nnz_max = PetscMax(nnz_max, nnz);
3929: d = 0;
3930: for (j = 0; j < nnz; j++) {
3931: if (cstart <= JJ[j] && JJ[j] < cend) d++;
3932: }
3933: d_nnz[i] = d;
3934: o_nnz[i] = nnz - d;
3935: }
3936: PetscCall(MatMPIAIJSetPreallocation(B, 0, d_nnz, 0, o_nnz));
3937: PetscCall(PetscFree2(d_nnz, o_nnz));
3939: for (i = 0; i < m; i++) {
3940: ii = i + rstart;
3941: PetscCall(MatSetValues_MPIAIJ(B, 1, &ii, Ii[i + 1] - Ii[i], PetscSafePointerPlusOffset(J, Ii[i] - irstart), PetscSafePointerPlusOffset(v, Ii[i] - irstart), INSERT_VALUES));
3942: }
3943: nooffprocentries = B->nooffprocentries;
3944: B->nooffprocentries = PETSC_TRUE;
3945: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
3946: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
3947: B->nooffprocentries = nooffprocentries;
3949: /* count number of entries below block diagonal */
3950: PetscCall(PetscFree(Aij->ld));
3951: PetscCall(PetscCalloc1(m, &ld));
3952: Aij->ld = ld;
3953: for (i = 0; i < m; i++) {
3954: nnz = Ii[i + 1] - Ii[i];
3955: j = 0;
3956: while (j < nnz && J[j] < cstart) j++;
3957: ld[i] = j;
3958: if (J) J += nnz;
3959: }
3961: PetscCall(MatSetOption(B, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
3962: PetscFunctionReturn(PETSC_SUCCESS);
3963: }
3965: /*@
3966: MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in `MATAIJ` format
3967: (the default parallel PETSc format).
3969: Collective
3971: Input Parameters:
3972: + B - the matrix
3973: . i - the indices into `j` for the start of each local row (indices start with zero)
3974: . j - the column indices for each local row (indices start with zero)
3975: - v - optional values in the matrix
3977: Level: developer
3979: Notes:
3980: The `i`, `j`, and `v` arrays ARE copied by this routine into the internal format used by PETSc;
3981: thus you CANNOT change the matrix entries by changing the values of `v` after you have
3982: called this routine. Use `MatCreateMPIAIJWithSplitArrays()` to avoid needing to copy the arrays.
3984: The `i` and `j` indices are 0 based, and `i` indices are indices corresponding to the local `j` array.
3986: A convenience routine for this functionality is `MatCreateMPIAIJWithArrays()`.
3988: You can update the matrix with new numerical values using `MatUpdateMPIAIJWithArrays()` after this call if the column indices in `j` are sorted.
3990: If you do **not** use `MatUpdateMPIAIJWithArrays()`, the column indices in `j` do not need to be sorted. If you will use
3991: `MatUpdateMPIAIJWithArrays()`, the column indices **must** be sorted.
3993: The format which is used for the sparse matrix input, is equivalent to a
3994: row-major ordering.. i.e for the following matrix, the input data expected is
3995: as shown
3996: .vb
3997: 1 0 0
3998: 2 0 3 P0
3999: -------
4000: 4 5 6 P1
4002: Process0 [P0] rows_owned=[0,1]
4003: i = {0,1,3} [size = nrow+1 = 2+1]
4004: j = {0,0,2} [size = 3]
4005: v = {1,2,3} [size = 3]
4007: Process1 [P1] rows_owned=[2]
4008: i = {0,3} [size = nrow+1 = 1+1]
4009: j = {0,1,2} [size = 3]
4010: v = {4,5,6} [size = 3]
4011: .ve
4013: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatCreateAIJ()`,
4014: `MatCreateSeqAIJWithArrays()`, `MatCreateMPIAIJWithSplitArrays()`, `MatCreateMPIAIJWithArrays()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4015: @*/
4016: PetscErrorCode MatMPIAIJSetPreallocationCSR(Mat B, const PetscInt i[], const PetscInt j[], const PetscScalar v[])
4017: {
4018: PetscFunctionBegin;
4019: PetscTryMethod(B, "MatMPIAIJSetPreallocationCSR_C", (Mat, const PetscInt[], const PetscInt[], const PetscScalar[]), (B, i, j, v));
4020: PetscFunctionReturn(PETSC_SUCCESS);
4021: }
4023: /*@
4024: MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in `MATMPIAIJ` format
4025: (the default parallel PETSc format). For good matrix assembly performance
4026: the user should preallocate the matrix storage by setting the parameters
4027: `d_nz` (or `d_nnz`) and `o_nz` (or `o_nnz`).
4029: Collective
4031: Input Parameters:
4032: + B - the matrix
4033: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
4034: (same value is used for all local rows)
4035: . d_nnz - array containing the number of nonzeros in the various rows of the
4036: DIAGONAL portion of the local submatrix (possibly different for each row)
4037: or `NULL` (`PETSC_NULL_INTEGER` in Fortran), if `d_nz` is used to specify the nonzero structure.
4038: The size of this array is equal to the number of local rows, i.e 'm'.
4039: For matrices that will be factored, you must leave room for (and set)
4040: the diagonal entry even if it is zero.
4041: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
4042: submatrix (same value is used for all local rows).
4043: - o_nnz - array containing the number of nonzeros in the various rows of the
4044: OFF-DIAGONAL portion of the local submatrix (possibly different for
4045: each row) or `NULL` (`PETSC_NULL_INTEGER` in Fortran), if `o_nz` is used to specify the nonzero
4046: structure. The size of this array is equal to the number
4047: of local rows, i.e 'm'.
4049: Example Usage:
4050: Consider the following 8x8 matrix with 34 non-zero values, that is
4051: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4052: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4053: as follows
4055: .vb
4056: 1 2 0 | 0 3 0 | 0 4
4057: Proc0 0 5 6 | 7 0 0 | 8 0
4058: 9 0 10 | 11 0 0 | 12 0
4059: -------------------------------------
4060: 13 0 14 | 15 16 17 | 0 0
4061: Proc1 0 18 0 | 19 20 21 | 0 0
4062: 0 0 0 | 22 23 0 | 24 0
4063: -------------------------------------
4064: Proc2 25 26 27 | 0 0 28 | 29 0
4065: 30 0 0 | 31 32 33 | 0 34
4066: .ve
4068: This can be represented as a collection of submatrices as
4069: .vb
4070: A B C
4071: D E F
4072: G H I
4073: .ve
4075: Where the submatrices A,B,C are owned by proc0, D,E,F are
4076: owned by proc1, G,H,I are owned by proc2.
4078: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4079: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4080: The 'M','N' parameters are 8,8, and have the same values on all procs.
4082: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4083: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4084: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4085: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4086: part as `MATSEQAIJ` matrices. For example, proc1 will store [E] as a `MATSEQAIJ`
4087: matrix, and [DF] as another `MATSEQAIJ` matrix.
4089: When `d_nz`, `o_nz` parameters are specified, `d_nz` storage elements are
4090: allocated for every row of the local DIAGONAL submatrix, and `o_nz`
4091: storage locations are allocated for every row of the OFF-DIAGONAL submatrix.
4092: One way to choose `d_nz` and `o_nz` is to use the maximum number of nonzeros over
4093: the local rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4094: In this case, the values of `d_nz`, `o_nz` are
4095: .vb
4096: proc0 dnz = 2, o_nz = 2
4097: proc1 dnz = 3, o_nz = 2
4098: proc2 dnz = 1, o_nz = 4
4099: .ve
4100: We are allocating `m`*(`d_nz`+`o_nz`) storage locations for every proc. This
4101: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4102: for proc3. i.e we are using 12+15+10=37 storage locations to store
4103: 34 values.
4105: When `d_nnz`, `o_nnz` parameters are specified, the storage is specified
4106: for every row, corresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4107: In the above case the values for `d_nnz`, `o_nnz` are
4108: .vb
4109: proc0 d_nnz = [2,2,2] and o_nnz = [2,2,2]
4110: proc1 d_nnz = [3,3,2] and o_nnz = [2,1,1]
4111: proc2 d_nnz = [1,1] and o_nnz = [4,4]
4112: .ve
4113: Here the space allocated is sum of all the above values i.e 34, and
4114: hence pre-allocation is perfect.
4116: Level: intermediate
4118: Notes:
4119: If the *_nnz parameter is given then the *_nz parameter is ignored
4121: The `MATAIJ` format, also called compressed row storage (CSR), is compatible with standard Fortran
4122: storage. The stored row and column indices begin with zero.
4123: See [Sparse Matrices](sec_matsparse) for details.
4125: The parallel matrix is partitioned such that the first m0 rows belong to
4126: process 0, the next m1 rows belong to process 1, the next m2 rows belong
4127: to process 2 etc.. where m0,m1,m2... are the input parameter 'm'.
4129: The DIAGONAL portion of the local submatrix of a processor can be defined
4130: as the submatrix which is obtained by extraction the part corresponding to
4131: the rows r1-r2 and columns c1-c2 of the global matrix, where r1 is the
4132: first row that belongs to the processor, r2 is the last row belonging to
4133: the this processor, and c1-c2 is range of indices of the local part of a
4134: vector suitable for applying the matrix to. This is an mxn matrix. In the
4135: common case of a square matrix, the row and column ranges are the same and
4136: the DIAGONAL part is also square. The remaining portion of the local
4137: submatrix (mxN) constitute the OFF-DIAGONAL portion.
4139: If `o_nnz` and `d_nnz` are specified, then `o_nz` and `d_nz` are ignored.
4141: You can call `MatGetInfo()` to get information on how effective the preallocation was;
4142: for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
4143: You can also run with the option `-info` and look for messages with the string
4144: malloc in them to see if additional memory allocation was needed.
4146: .seealso: [](ch_matrices), `Mat`, [Sparse Matrices](sec_matsparse), `MATMPIAIJ`, `MATAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatCreateAIJ()`, `MatMPIAIJSetPreallocationCSR()`,
4147: `MatGetInfo()`, `PetscSplitOwnership()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4148: @*/
4149: PetscErrorCode MatMPIAIJSetPreallocation(Mat B, PetscInt d_nz, const PetscInt d_nnz[], PetscInt o_nz, const PetscInt o_nnz[])
4150: {
4151: PetscFunctionBegin;
4154: PetscTryMethod(B, "MatMPIAIJSetPreallocation_C", (Mat, PetscInt, const PetscInt[], PetscInt, const PetscInt[]), (B, d_nz, d_nnz, o_nz, o_nnz));
4155: PetscFunctionReturn(PETSC_SUCCESS);
4156: }
4158: /*@
4159: MatCreateMPIAIJWithArrays - creates a `MATMPIAIJ` matrix using arrays that contain in standard
4160: CSR format for the local rows.
4162: Collective
4164: Input Parameters:
4165: + comm - MPI communicator
4166: . m - number of local rows (Cannot be `PETSC_DECIDE`)
4167: . n - This value should be the same as the local size used in creating the
4168: x vector for the matrix-vector product $ y = Ax$. (or `PETSC_DECIDE` to have
4169: calculated if `N` is given) For square matrices n is almost always `m`.
4170: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if `m` is given)
4171: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if `n` is given)
4172: . 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
4173: . j - global column indices
4174: - a - optional matrix values
4176: Output Parameter:
4177: . mat - the matrix
4179: Level: intermediate
4181: Notes:
4182: The `i`, `j`, and `a` arrays ARE copied by this routine into the internal format used by PETSc;
4183: thus you CANNOT change the matrix entries by changing the values of `a[]` after you have
4184: called this routine. Use `MatCreateMPIAIJWithSplitArrays()` to avoid needing to copy the arrays.
4186: The `i` and `j` indices are 0 based, and `i` indices are indices corresponding to the local `j` array.
4188: Once you have created the matrix you can update it with new numerical values using `MatUpdateMPIAIJWithArray()`
4190: If you do **not** use `MatUpdateMPIAIJWithArray()`, the column indices in `j` do not need to be sorted. If you will use
4191: `MatUpdateMPIAIJWithArrays()`, the column indices **must** be sorted.
4193: The format which is used for the sparse matrix input, is equivalent to a
4194: row-major ordering, i.e., for the following matrix, the input data expected is
4195: as shown
4196: .vb
4197: 1 0 0
4198: 2 0 3 P0
4199: -------
4200: 4 5 6 P1
4202: Process0 [P0] rows_owned=[0,1]
4203: i = {0,1,3} [size = nrow+1 = 2+1]
4204: j = {0,0,2} [size = 3]
4205: v = {1,2,3} [size = 3]
4207: Process1 [P1] rows_owned=[2]
4208: i = {0,3} [size = nrow+1 = 1+1]
4209: j = {0,1,2} [size = 3]
4210: v = {4,5,6} [size = 3]
4211: .ve
4213: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4214: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArray()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4215: @*/
4216: PetscErrorCode MatCreateMPIAIJWithArrays(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt M, PetscInt N, const PetscInt i[], const PetscInt j[], const PetscScalar a[], Mat *mat)
4217: {
4218: PetscFunctionBegin;
4219: PetscCheck(!i || !i[0], PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
4220: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
4221: PetscCall(MatCreate(comm, mat));
4222: PetscCall(MatSetSizes(*mat, m, n, M, N));
4223: /* PetscCall(MatSetBlockSizes(M,bs,cbs)); */
4224: PetscCall(MatSetType(*mat, MATMPIAIJ));
4225: PetscCall(MatMPIAIJSetPreallocationCSR(*mat, i, j, a));
4226: PetscFunctionReturn(PETSC_SUCCESS);
4227: }
4229: /*@
4230: MatUpdateMPIAIJWithArrays - updates a `MATMPIAIJ` matrix using arrays that contain in standard
4231: CSR format for the local rows. Only the numerical values are updated the other arrays must be identical to what was passed
4232: from `MatCreateMPIAIJWithArrays()`
4234: Deprecated: Use `MatUpdateMPIAIJWithArray()`
4236: Collective
4238: Input Parameters:
4239: + mat - the matrix
4240: . m - number of local rows (Cannot be `PETSC_DECIDE`)
4241: . n - This value should be the same as the local size used in creating the
4242: x vector for the matrix-vector product y = Ax. (or `PETSC_DECIDE` to have
4243: calculated if N is given) For square matrices n is almost always m.
4244: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if m is given)
4245: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if n is given)
4246: . Ii - row indices; that is Ii[0] = 0, Ii[row] = Ii[row-1] + number of elements in that row of the matrix
4247: . J - column indices
4248: - v - matrix values
4250: Level: deprecated
4252: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4253: `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArray()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4254: @*/
4255: PetscErrorCode MatUpdateMPIAIJWithArrays(Mat mat, PetscInt m, PetscInt n, PetscInt M, PetscInt N, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[])
4256: {
4257: PetscInt nnz, i;
4258: PetscBool nooffprocentries;
4259: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)mat->data;
4260: Mat_SeqAIJ *Ad = (Mat_SeqAIJ *)Aij->A->data;
4261: PetscScalar *ad, *ao;
4262: PetscInt ldi, Iii, md;
4263: const PetscInt *Adi = Ad->i;
4264: PetscInt *ld = Aij->ld;
4266: PetscFunctionBegin;
4267: PetscCheck(Ii[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
4268: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
4269: PetscCheck(m == mat->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Local number of rows cannot change from call to MatUpdateMPIAIJWithArrays()");
4270: PetscCheck(n == mat->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Local number of columns cannot change from call to MatUpdateMPIAIJWithArrays()");
4272: PetscCall(MatSeqAIJGetArrayWrite(Aij->A, &ad));
4273: PetscCall(MatSeqAIJGetArrayWrite(Aij->B, &ao));
4275: for (i = 0; i < m; i++) {
4276: if (PetscDefined(USE_DEBUG)) {
4277: for (PetscInt j = Ii[i] + 1; j < Ii[i + 1]; ++j) {
4278: 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);
4279: 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);
4280: }
4281: }
4282: nnz = Ii[i + 1] - Ii[i];
4283: Iii = Ii[i];
4284: ldi = ld[i];
4285: md = Adi[i + 1] - Adi[i];
4286: PetscCall(PetscArraycpy(ao, v + Iii, ldi));
4287: PetscCall(PetscArraycpy(ad, v + Iii + ldi, md));
4288: PetscCall(PetscArraycpy(ao + ldi, v + Iii + ldi + md, nnz - ldi - md));
4289: ad += md;
4290: ao += nnz - md;
4291: }
4292: nooffprocentries = mat->nooffprocentries;
4293: mat->nooffprocentries = PETSC_TRUE;
4294: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->A, &ad));
4295: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->B, &ao));
4296: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->A));
4297: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->B));
4298: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
4299: PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
4300: PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
4301: mat->nooffprocentries = nooffprocentries;
4302: PetscFunctionReturn(PETSC_SUCCESS);
4303: }
4305: /*@
4306: MatUpdateMPIAIJWithArray - updates an `MATMPIAIJ` matrix using an array that contains the nonzero values
4308: Collective
4310: Input Parameters:
4311: + mat - the matrix
4312: - v - matrix values, stored by row
4314: Level: intermediate
4316: Notes:
4317: The matrix must have been obtained with `MatCreateMPIAIJWithArrays()` or `MatMPIAIJSetPreallocationCSR()`
4319: The column indices in the call to `MatCreateMPIAIJWithArrays()` or `MatMPIAIJSetPreallocationCSR()` must have been sorted for this call to work correctly
4321: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4322: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArrays()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4323: @*/
4324: PetscErrorCode MatUpdateMPIAIJWithArray(Mat mat, const PetscScalar v[])
4325: {
4326: PetscInt nnz, i, m;
4327: PetscBool nooffprocentries;
4328: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)mat->data;
4329: Mat_SeqAIJ *Ad = (Mat_SeqAIJ *)Aij->A->data;
4330: Mat_SeqAIJ *Ao = (Mat_SeqAIJ *)Aij->B->data;
4331: PetscScalar *ad, *ao;
4332: const PetscInt *Adi = Ad->i, *Adj = Ao->i;
4333: PetscInt ldi, Iii, md;
4334: PetscInt *ld = Aij->ld;
4336: PetscFunctionBegin;
4337: m = mat->rmap->n;
4339: PetscCall(MatSeqAIJGetArrayWrite(Aij->A, &ad));
4340: PetscCall(MatSeqAIJGetArrayWrite(Aij->B, &ao));
4341: Iii = 0;
4342: for (i = 0; i < m; i++) {
4343: nnz = Adi[i + 1] - Adi[i] + Adj[i + 1] - Adj[i];
4344: ldi = ld[i];
4345: md = Adi[i + 1] - Adi[i];
4346: PetscCall(PetscArraycpy(ad, v + Iii + ldi, md));
4347: ad += md;
4348: if (ao) {
4349: PetscCall(PetscArraycpy(ao, v + Iii, ldi));
4350: PetscCall(PetscArraycpy(ao + ldi, v + Iii + ldi + md, nnz - ldi - md));
4351: ao += nnz - md;
4352: }
4353: Iii += nnz;
4354: }
4355: nooffprocentries = mat->nooffprocentries;
4356: mat->nooffprocentries = PETSC_TRUE;
4357: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->A, &ad));
4358: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->B, &ao));
4359: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->A));
4360: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->B));
4361: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
4362: PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
4363: PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
4364: mat->nooffprocentries = nooffprocentries;
4365: PetscFunctionReturn(PETSC_SUCCESS);
4366: }
4368: /*@
4369: MatCreateAIJ - Creates a sparse parallel matrix in `MATAIJ` format
4370: (the default parallel PETSc format). For good matrix assembly performance
4371: the user should preallocate the matrix storage by setting the parameters
4372: `d_nz` (or `d_nnz`) and `o_nz` (or `o_nnz`).
4374: Collective
4376: Input Parameters:
4377: + comm - MPI communicator
4378: . m - number of local rows (or `PETSC_DECIDE` to have calculated if M is given)
4379: This value should be the same as the local size used in creating the
4380: y vector for the matrix-vector product y = Ax.
4381: . n - This value should be the same as the local size used in creating the
4382: x vector for the matrix-vector product y = Ax. (or `PETSC_DECIDE` to have
4383: calculated if N is given) For square matrices n is almost always m.
4384: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if m is given)
4385: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if n is given)
4386: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
4387: (same value is used for all local rows)
4388: . d_nnz - array containing the number of nonzeros in the various rows of the
4389: DIAGONAL portion of the local submatrix (possibly different for each row)
4390: or `NULL`, if `d_nz` is used to specify the nonzero structure.
4391: The size of this array is equal to the number of local rows, i.e 'm'.
4392: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
4393: submatrix (same value is used for all local rows).
4394: - o_nnz - array containing the number of nonzeros in the various rows of the
4395: OFF-DIAGONAL portion of the local submatrix (possibly different for
4396: each row) or `NULL`, if `o_nz` is used to specify the nonzero
4397: structure. The size of this array is equal to the number
4398: of local rows, i.e 'm'.
4400: Output Parameter:
4401: . A - the matrix
4403: Options Database Keys:
4404: + -mat_no_inode - Do not use inodes
4405: . -mat_inode_limit <limit> - Sets inode limit (max limit=5)
4406: - -matmult_vecscatter_view <viewer> - View the vecscatter (i.e., communication pattern) used in `MatMult()` of sparse parallel matrices.
4407: See viewer types in manual of `MatView()`. Of them, ascii_matlab, draw or binary cause the `VecScatter`
4408: 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.
4410: Level: intermediate
4412: Notes:
4413: It is recommended that one use `MatCreateFromOptions()` or the `MatCreate()`, `MatSetType()` and/or `MatSetFromOptions()`,
4414: MatXXXXSetPreallocation() paradigm instead of this routine directly.
4415: [MatXXXXSetPreallocation() is, for example, `MatSeqAIJSetPreallocation()`]
4417: If the *_nnz parameter is given then the *_nz parameter is ignored
4419: The `m`,`n`,`M`,`N` parameters specify the size of the matrix, and its partitioning across
4420: processors, while `d_nz`,`d_nnz`,`o_nz`,`o_nnz` parameters specify the approximate
4421: storage requirements for this matrix.
4423: If `PETSC_DECIDE` or `PETSC_DETERMINE` is used for a particular argument on one
4424: processor than it must be used on all processors that share the object for
4425: that argument.
4427: If `m` and `n` are not `PETSC_DECIDE`, then the values determine the `PetscLayout` of the matrix and the ranges returned by
4428: `MatGetOwnershipRange()`, `MatGetOwnershipRanges()`, `MatGetOwnershipRangeColumn()`, and `MatGetOwnershipRangesColumn()`.
4430: The user MUST specify either the local or global matrix dimensions
4431: (possibly both).
4433: The parallel matrix is partitioned across processors such that the
4434: first `m0` rows belong to process 0, the next `m1` rows belong to
4435: process 1, the next `m2` rows belong to process 2, etc., where
4436: `m0`, `m1`, `m2`... are the input parameter `m` on each MPI process. I.e., each MPI process stores
4437: values corresponding to [m x N] submatrix.
4439: The columns are logically partitioned with the n0 columns belonging
4440: to 0th partition, the next n1 columns belonging to the next
4441: partition etc.. where n0,n1,n2... are the input parameter 'n'.
4443: The DIAGONAL portion of the local submatrix on any given processor
4444: is the submatrix corresponding to the rows and columns m,n
4445: corresponding to the given processor. i.e diagonal matrix on
4446: process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1]
4447: etc. The remaining portion of the local submatrix [m x (N-n)]
4448: constitute the OFF-DIAGONAL portion. The example below better
4449: illustrates this concept. The two matrices, the DIAGONAL portion and
4450: the OFF-DIAGONAL portion are each stored as `MATSEQAIJ` matrices.
4452: For a square global matrix we define each processor's diagonal portion
4453: to be its local rows and the corresponding columns (a square submatrix);
4454: each processor's off-diagonal portion encompasses the remainder of the
4455: local matrix (a rectangular submatrix).
4457: If `o_nnz`, `d_nnz` are specified, then `o_nz`, and `d_nz` are ignored.
4459: When calling this routine with a single process communicator, a matrix of
4460: type `MATSEQAIJ` is returned. If a matrix of type `MATMPIAIJ` is desired for this
4461: type of communicator, use the construction mechanism
4462: .vb
4463: MatCreate(..., &A);
4464: MatSetType(A, MATMPIAIJ);
4465: MatSetSizes(A, m, n, M, N);
4466: MatMPIAIJSetPreallocation(A, ...);
4467: .ve
4469: By default, this format uses inodes (identical nodes) when possible.
4470: We search for consecutive rows with the same nonzero structure, thereby
4471: reusing matrix information to achieve increased efficiency.
4473: Example Usage:
4474: Consider the following 8x8 matrix with 34 non-zero values, that is
4475: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4476: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4477: as follows
4479: .vb
4480: 1 2 0 | 0 3 0 | 0 4
4481: Proc0 0 5 6 | 7 0 0 | 8 0
4482: 9 0 10 | 11 0 0 | 12 0
4483: -------------------------------------
4484: 13 0 14 | 15 16 17 | 0 0
4485: Proc1 0 18 0 | 19 20 21 | 0 0
4486: 0 0 0 | 22 23 0 | 24 0
4487: -------------------------------------
4488: Proc2 25 26 27 | 0 0 28 | 29 0
4489: 30 0 0 | 31 32 33 | 0 34
4490: .ve
4492: This can be represented as a collection of submatrices as
4494: .vb
4495: A B C
4496: D E F
4497: G H I
4498: .ve
4500: Where the submatrices A,B,C are owned by proc0, D,E,F are
4501: owned by proc1, G,H,I are owned by proc2.
4503: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4504: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4505: The 'M','N' parameters are 8,8, and have the same values on all procs.
4507: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4508: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4509: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4510: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4511: part as `MATSEQAIJ` matrices. For example, proc1 will store [E] as a `MATSEQAIJ`
4512: matrix, and [DF] as another SeqAIJ matrix.
4514: When `d_nz`, `o_nz` parameters are specified, `d_nz` storage elements are
4515: allocated for every row of the local DIAGONAL submatrix, and `o_nz`
4516: storage locations are allocated for every row of the OFF-DIAGONAL submatrix.
4517: One way to choose `d_nz` and `o_nz` is to use the maximum number of nonzeros over
4518: the local rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4519: In this case, the values of `d_nz`,`o_nz` are
4520: .vb
4521: proc0 dnz = 2, o_nz = 2
4522: proc1 dnz = 3, o_nz = 2
4523: proc2 dnz = 1, o_nz = 4
4524: .ve
4525: We are allocating m*(`d_nz`+`o_nz`) storage locations for every proc. This
4526: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4527: for proc3. i.e we are using 12+15+10=37 storage locations to store
4528: 34 values.
4530: When `d_nnz`, `o_nnz` parameters are specified, the storage is specified
4531: for every row, corresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4532: In the above case the values for d_nnz,o_nnz are
4533: .vb
4534: proc0 d_nnz = [2,2,2] and o_nnz = [2,2,2]
4535: proc1 d_nnz = [3,3,2] and o_nnz = [2,1,1]
4536: proc2 d_nnz = [1,1] and o_nnz = [4,4]
4537: .ve
4538: Here the space allocated is sum of all the above values i.e 34, and
4539: hence pre-allocation is perfect.
4541: .seealso: [](ch_matrices), `Mat`, [Sparse Matrix Creation](sec_matsparse), `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4542: `MATMPIAIJ`, `MatCreateMPIAIJWithArrays()`, `MatGetOwnershipRange()`, `MatGetOwnershipRanges()`, `MatGetOwnershipRangeColumn()`,
4543: `MatGetOwnershipRangesColumn()`, `PetscLayout`
4544: @*/
4545: 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)
4546: {
4547: PetscMPIInt size;
4549: PetscFunctionBegin;
4550: PetscCall(MatCreate(comm, A));
4551: PetscCall(MatSetSizes(*A, m, n, M, N));
4552: PetscCallMPI(MPI_Comm_size(comm, &size));
4553: if (size > 1) {
4554: PetscCall(MatSetType(*A, MATMPIAIJ));
4555: PetscCall(MatMPIAIJSetPreallocation(*A, d_nz, d_nnz, o_nz, o_nnz));
4556: } else {
4557: PetscCall(MatSetType(*A, MATSEQAIJ));
4558: PetscCall(MatSeqAIJSetPreallocation(*A, d_nz, d_nnz));
4559: }
4560: PetscFunctionReturn(PETSC_SUCCESS);
4561: }
4563: /*MC
4564: MatMPIAIJGetSeqAIJF90 - Returns the local pieces of this distributed matrix
4566: Synopsis:
4567: MatMPIAIJGetSeqAIJF90(Mat A, Mat Ad, Mat Ao, {PetscInt, pointer :: colmap(:)},integer ierr)
4569: Not Collective
4571: Input Parameter:
4572: . A - the `MATMPIAIJ` matrix
4574: Output Parameters:
4575: + Ad - the diagonal portion of the matrix
4576: . Ao - the off-diagonal portion of the matrix
4577: . colmap - An array mapping local column numbers of `Ao` to global column numbers of the parallel matrix
4578: - ierr - error code
4580: Level: advanced
4582: Note:
4583: Use `MatMPIAIJRestoreSeqAIJF90()` when you no longer need access to the matrices and `colmap`
4585: .seealso: [](ch_matrices), `Mat`, [](sec_fortranarrays), `Mat`, `MATMPIAIJ`, `MatMPIAIJGetSeqAIJ()`, `MatMPIAIJRestoreSeqAIJF90()`
4586: M*/
4588: /*MC
4589: MatMPIAIJRestoreSeqAIJF90 - call after `MatMPIAIJGetSeqAIJF90()` when you no longer need access to the matrices and `colmap`
4591: Synopsis:
4592: MatMPIAIJRestoreSeqAIJF90(Mat A, Mat Ad, Mat Ao, {PetscInt, pointer :: colmap(:)},integer ierr)
4594: Not Collective
4596: Input Parameters:
4597: + A - the `MATMPIAIJ` matrix
4598: . Ad - the diagonal portion of the matrix
4599: . Ao - the off-diagonal portion of the matrix
4600: . colmap - An array mapping local column numbers of `Ao` to global column numbers of the parallel matrix
4601: - ierr - error code
4603: Level: advanced
4605: .seealso: [](ch_matrices), `Mat`, [](sec_fortranarrays), `Mat`, `MATMPIAIJ`, `MatMPIAIJGetSeqAIJ()`, `MatMPIAIJGetSeqAIJF90()`
4606: M*/
4608: /*@C
4609: MatMPIAIJGetSeqAIJ - Returns the local pieces of this distributed matrix
4611: Not Collective
4613: Input Parameter:
4614: . A - The `MATMPIAIJ` matrix
4616: Output Parameters:
4617: + Ad - The local diagonal block as a `MATSEQAIJ` matrix
4618: . Ao - The local off-diagonal block as a `MATSEQAIJ` matrix
4619: - colmap - An array mapping local column numbers of `Ao` to global column numbers of the parallel matrix
4621: Level: intermediate
4623: Note:
4624: The rows in `Ad` and `Ao` are in [0, Nr), where Nr is the number of local rows on this process. The columns
4625: 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
4626: the number of nonzero columns in the local off-diagonal piece of the matrix `A`. The array colmap maps these
4627: local column numbers to global column numbers in the original matrix.
4629: Fortran Notes:
4630: `MatMPIAIJGetSeqAIJ()` Fortran binding is deprecated (since PETSc 3.19), use `MatMPIAIJGetSeqAIJF90()`
4632: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatMPIAIJGetSeqAIJF90()`, `MatMPIAIJRestoreSeqAIJF90()`, `MatMPIAIJGetLocalMat()`, `MatMPIAIJGetLocalMatCondensed()`, `MatCreateAIJ()`, `MATSEQAIJ`
4633: @*/
4634: PetscErrorCode MatMPIAIJGetSeqAIJ(Mat A, Mat *Ad, Mat *Ao, const PetscInt *colmap[])
4635: {
4636: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
4637: PetscBool flg;
4639: PetscFunctionBegin;
4640: PetscCall(PetscStrbeginswith(((PetscObject)A)->type_name, MATMPIAIJ, &flg));
4641: PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "This function requires a MATMPIAIJ matrix as input");
4642: if (Ad) *Ad = a->A;
4643: if (Ao) *Ao = a->B;
4644: if (colmap) *colmap = a->garray;
4645: PetscFunctionReturn(PETSC_SUCCESS);
4646: }
4648: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm comm, Mat inmat, PetscInt n, MatReuse scall, Mat *outmat)
4649: {
4650: PetscInt m, N, i, rstart, nnz, Ii;
4651: PetscInt *indx;
4652: PetscScalar *values;
4653: MatType rootType;
4655: PetscFunctionBegin;
4656: PetscCall(MatGetSize(inmat, &m, &N));
4657: if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */
4658: PetscInt *dnz, *onz, sum, bs, cbs;
4660: if (n == PETSC_DECIDE) PetscCall(PetscSplitOwnership(comm, &n, &N));
4661: /* Check sum(n) = N */
4662: PetscCallMPI(MPIU_Allreduce(&n, &sum, 1, MPIU_INT, MPI_SUM, comm));
4663: PetscCheck(sum == N, PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Sum of local columns %" PetscInt_FMT " != global columns %" PetscInt_FMT, sum, N);
4665: PetscCallMPI(MPI_Scan(&m, &rstart, 1, MPIU_INT, MPI_SUM, comm));
4666: rstart -= m;
4668: MatPreallocateBegin(comm, m, n, dnz, onz);
4669: for (i = 0; i < m; i++) {
4670: PetscCall(MatGetRow_SeqAIJ(inmat, i, &nnz, &indx, NULL));
4671: PetscCall(MatPreallocateSet(i + rstart, nnz, indx, dnz, onz));
4672: PetscCall(MatRestoreRow_SeqAIJ(inmat, i, &nnz, &indx, NULL));
4673: }
4675: PetscCall(MatCreate(comm, outmat));
4676: PetscCall(MatSetSizes(*outmat, m, n, PETSC_DETERMINE, PETSC_DETERMINE));
4677: PetscCall(MatGetBlockSizes(inmat, &bs, &cbs));
4678: PetscCall(MatSetBlockSizes(*outmat, bs, cbs));
4679: PetscCall(MatGetRootType_Private(inmat, &rootType));
4680: PetscCall(MatSetType(*outmat, rootType));
4681: PetscCall(MatSeqAIJSetPreallocation(*outmat, 0, dnz));
4682: PetscCall(MatMPIAIJSetPreallocation(*outmat, 0, dnz, 0, onz));
4683: MatPreallocateEnd(dnz, onz);
4684: PetscCall(MatSetOption(*outmat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
4685: }
4687: /* numeric phase */
4688: PetscCall(MatGetOwnershipRange(*outmat, &rstart, NULL));
4689: for (i = 0; i < m; i++) {
4690: PetscCall(MatGetRow_SeqAIJ(inmat, i, &nnz, &indx, &values));
4691: Ii = i + rstart;
4692: PetscCall(MatSetValues(*outmat, 1, &Ii, nnz, indx, values, INSERT_VALUES));
4693: PetscCall(MatRestoreRow_SeqAIJ(inmat, i, &nnz, &indx, &values));
4694: }
4695: PetscCall(MatAssemblyBegin(*outmat, MAT_FINAL_ASSEMBLY));
4696: PetscCall(MatAssemblyEnd(*outmat, MAT_FINAL_ASSEMBLY));
4697: PetscFunctionReturn(PETSC_SUCCESS);
4698: }
4700: static PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(void *data)
4701: {
4702: Mat_Merge_SeqsToMPI *merge = (Mat_Merge_SeqsToMPI *)data;
4704: PetscFunctionBegin;
4705: if (!merge) PetscFunctionReturn(PETSC_SUCCESS);
4706: PetscCall(PetscFree(merge->id_r));
4707: PetscCall(PetscFree(merge->len_s));
4708: PetscCall(PetscFree(merge->len_r));
4709: PetscCall(PetscFree(merge->bi));
4710: PetscCall(PetscFree(merge->bj));
4711: PetscCall(PetscFree(merge->buf_ri[0]));
4712: PetscCall(PetscFree(merge->buf_ri));
4713: PetscCall(PetscFree(merge->buf_rj[0]));
4714: PetscCall(PetscFree(merge->buf_rj));
4715: PetscCall(PetscFree(merge->coi));
4716: PetscCall(PetscFree(merge->coj));
4717: PetscCall(PetscFree(merge->owners_co));
4718: PetscCall(PetscLayoutDestroy(&merge->rowmap));
4719: PetscCall(PetscFree(merge));
4720: PetscFunctionReturn(PETSC_SUCCESS);
4721: }
4723: #include <../src/mat/utils/freespace.h>
4724: #include <petscbt.h>
4726: PetscErrorCode MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat, Mat mpimat)
4727: {
4728: MPI_Comm comm;
4729: Mat_SeqAIJ *a = (Mat_SeqAIJ *)seqmat->data;
4730: PetscMPIInt size, rank, taga, *len_s;
4731: PetscInt N = mpimat->cmap->N, i, j, *owners, *ai = a->i, *aj, m;
4732: PetscMPIInt proc, k;
4733: PetscInt **buf_ri, **buf_rj;
4734: PetscInt anzi, *bj_i, *bi, *bj, arow, bnzi, nextaj;
4735: PetscInt nrows, **buf_ri_k, **nextrow, **nextai;
4736: MPI_Request *s_waits, *r_waits;
4737: MPI_Status *status;
4738: const MatScalar *aa, *a_a;
4739: MatScalar **abuf_r, *ba_i;
4740: Mat_Merge_SeqsToMPI *merge;
4741: PetscContainer container;
4743: PetscFunctionBegin;
4744: PetscCall(PetscObjectGetComm((PetscObject)mpimat, &comm));
4745: PetscCall(PetscLogEventBegin(MAT_Seqstompinum, seqmat, 0, 0, 0));
4747: PetscCallMPI(MPI_Comm_size(comm, &size));
4748: PetscCallMPI(MPI_Comm_rank(comm, &rank));
4750: PetscCall(PetscObjectQuery((PetscObject)mpimat, "MatMergeSeqsToMPI", (PetscObject *)&container));
4751: PetscCheck(container, PetscObjectComm((PetscObject)mpimat), PETSC_ERR_PLIB, "Mat not created from MatCreateMPIAIJSumSeqAIJSymbolic");
4752: PetscCall(PetscContainerGetPointer(container, (void **)&merge));
4753: PetscCall(MatSeqAIJGetArrayRead(seqmat, &a_a));
4754: aa = a_a;
4756: bi = merge->bi;
4757: bj = merge->bj;
4758: buf_ri = merge->buf_ri;
4759: buf_rj = merge->buf_rj;
4761: PetscCall(PetscMalloc1(size, &status));
4762: owners = merge->rowmap->range;
4763: len_s = merge->len_s;
4765: /* send and recv matrix values */
4766: PetscCall(PetscObjectGetNewTag((PetscObject)mpimat, &taga));
4767: PetscCall(PetscPostIrecvScalar(comm, taga, merge->nrecv, merge->id_r, merge->len_r, &abuf_r, &r_waits));
4769: PetscCall(PetscMalloc1(merge->nsend + 1, &s_waits));
4770: for (proc = 0, k = 0; proc < size; proc++) {
4771: if (!len_s[proc]) continue;
4772: i = owners[proc];
4773: PetscCallMPI(MPIU_Isend(aa + ai[i], len_s[proc], MPIU_MATSCALAR, proc, taga, comm, s_waits + k));
4774: k++;
4775: }
4777: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, r_waits, status));
4778: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, s_waits, status));
4779: PetscCall(PetscFree(status));
4781: PetscCall(PetscFree(s_waits));
4782: PetscCall(PetscFree(r_waits));
4784: /* insert mat values of mpimat */
4785: PetscCall(PetscMalloc1(N, &ba_i));
4786: PetscCall(PetscMalloc3(merge->nrecv, &buf_ri_k, merge->nrecv, &nextrow, merge->nrecv, &nextai));
4788: for (k = 0; k < merge->nrecv; k++) {
4789: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4790: nrows = *buf_ri_k[k];
4791: nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */
4792: nextai[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure */
4793: }
4795: /* set values of ba */
4796: m = merge->rowmap->n;
4797: for (i = 0; i < m; i++) {
4798: arow = owners[rank] + i;
4799: bj_i = bj + bi[i]; /* col indices of the i-th row of mpimat */
4800: bnzi = bi[i + 1] - bi[i];
4801: PetscCall(PetscArrayzero(ba_i, bnzi));
4803: /* add local non-zero vals of this proc's seqmat into ba */
4804: anzi = ai[arow + 1] - ai[arow];
4805: aj = a->j + ai[arow];
4806: aa = a_a + ai[arow];
4807: nextaj = 0;
4808: for (j = 0; nextaj < anzi; j++) {
4809: if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4810: ba_i[j] += aa[nextaj++];
4811: }
4812: }
4814: /* add received vals into ba */
4815: for (k = 0; k < merge->nrecv; k++) { /* k-th received message */
4816: /* i-th row */
4817: if (i == *nextrow[k]) {
4818: anzi = *(nextai[k] + 1) - *nextai[k];
4819: aj = buf_rj[k] + *nextai[k];
4820: aa = abuf_r[k] + *nextai[k];
4821: nextaj = 0;
4822: for (j = 0; nextaj < anzi; j++) {
4823: if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4824: ba_i[j] += aa[nextaj++];
4825: }
4826: }
4827: nextrow[k]++;
4828: nextai[k]++;
4829: }
4830: }
4831: PetscCall(MatSetValues(mpimat, 1, &arow, bnzi, bj_i, ba_i, INSERT_VALUES));
4832: }
4833: PetscCall(MatSeqAIJRestoreArrayRead(seqmat, &a_a));
4834: PetscCall(MatAssemblyBegin(mpimat, MAT_FINAL_ASSEMBLY));
4835: PetscCall(MatAssemblyEnd(mpimat, MAT_FINAL_ASSEMBLY));
4837: PetscCall(PetscFree(abuf_r[0]));
4838: PetscCall(PetscFree(abuf_r));
4839: PetscCall(PetscFree(ba_i));
4840: PetscCall(PetscFree3(buf_ri_k, nextrow, nextai));
4841: PetscCall(PetscLogEventEnd(MAT_Seqstompinum, seqmat, 0, 0, 0));
4842: PetscFunctionReturn(PETSC_SUCCESS);
4843: }
4845: PetscErrorCode MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm, Mat seqmat, PetscInt m, PetscInt n, Mat *mpimat)
4846: {
4847: Mat B_mpi;
4848: Mat_SeqAIJ *a = (Mat_SeqAIJ *)seqmat->data;
4849: PetscMPIInt size, rank, tagi, tagj, *len_s, *len_si, *len_ri;
4850: PetscInt **buf_rj, **buf_ri, **buf_ri_k;
4851: PetscInt M = seqmat->rmap->n, N = seqmat->cmap->n, i, *owners, *ai = a->i, *aj = a->j;
4852: PetscInt len, *dnz, *onz, bs, cbs;
4853: PetscInt k, anzi, *bi, *bj, *lnk, nlnk, arow, bnzi;
4854: PetscInt nrows, *buf_s, *buf_si, *buf_si_i, **nextrow, **nextai;
4855: MPI_Request *si_waits, *sj_waits, *ri_waits, *rj_waits;
4856: MPI_Status *status;
4857: PetscFreeSpaceList free_space = NULL, current_space = NULL;
4858: PetscBT lnkbt;
4859: Mat_Merge_SeqsToMPI *merge;
4860: PetscContainer container;
4862: PetscFunctionBegin;
4863: PetscCall(PetscLogEventBegin(MAT_Seqstompisym, seqmat, 0, 0, 0));
4865: /* make sure it is a PETSc comm */
4866: PetscCall(PetscCommDuplicate(comm, &comm, NULL));
4867: PetscCallMPI(MPI_Comm_size(comm, &size));
4868: PetscCallMPI(MPI_Comm_rank(comm, &rank));
4870: PetscCall(PetscNew(&merge));
4871: PetscCall(PetscMalloc1(size, &status));
4873: /* determine row ownership */
4874: PetscCall(PetscLayoutCreate(comm, &merge->rowmap));
4875: PetscCall(PetscLayoutSetLocalSize(merge->rowmap, m));
4876: PetscCall(PetscLayoutSetSize(merge->rowmap, M));
4877: PetscCall(PetscLayoutSetBlockSize(merge->rowmap, 1));
4878: PetscCall(PetscLayoutSetUp(merge->rowmap));
4879: PetscCall(PetscMalloc1(size, &len_si));
4880: PetscCall(PetscMalloc1(size, &merge->len_s));
4882: m = merge->rowmap->n;
4883: owners = merge->rowmap->range;
4885: /* determine the number of messages to send, their lengths */
4886: len_s = merge->len_s;
4888: len = 0; /* length of buf_si[] */
4889: merge->nsend = 0;
4890: for (PetscMPIInt proc = 0; proc < size; proc++) {
4891: len_si[proc] = 0;
4892: if (proc == rank) {
4893: len_s[proc] = 0;
4894: } else {
4895: PetscCall(PetscMPIIntCast(owners[proc + 1] - owners[proc] + 1, &len_si[proc]));
4896: PetscCall(PetscMPIIntCast(ai[owners[proc + 1]] - ai[owners[proc]], &len_s[proc])); /* num of rows to be sent to [proc] */
4897: }
4898: if (len_s[proc]) {
4899: merge->nsend++;
4900: nrows = 0;
4901: for (i = owners[proc]; i < owners[proc + 1]; i++) {
4902: if (ai[i + 1] > ai[i]) nrows++;
4903: }
4904: PetscCall(PetscMPIIntCast(2 * (nrows + 1), &len_si[proc]));
4905: len += len_si[proc];
4906: }
4907: }
4909: /* determine the number and length of messages to receive for ij-structure */
4910: PetscCall(PetscGatherNumberOfMessages(comm, NULL, len_s, &merge->nrecv));
4911: PetscCall(PetscGatherMessageLengths2(comm, merge->nsend, merge->nrecv, len_s, len_si, &merge->id_r, &merge->len_r, &len_ri));
4913: /* post the Irecv of j-structure */
4914: PetscCall(PetscCommGetNewTag(comm, &tagj));
4915: PetscCall(PetscPostIrecvInt(comm, tagj, merge->nrecv, merge->id_r, merge->len_r, &buf_rj, &rj_waits));
4917: /* post the Isend of j-structure */
4918: PetscCall(PetscMalloc2(merge->nsend, &si_waits, merge->nsend, &sj_waits));
4920: for (PetscMPIInt proc = 0, k = 0; proc < size; proc++) {
4921: if (!len_s[proc]) continue;
4922: i = owners[proc];
4923: PetscCallMPI(MPIU_Isend(aj + ai[i], len_s[proc], MPIU_INT, proc, tagj, comm, sj_waits + k));
4924: k++;
4925: }
4927: /* receives and sends of j-structure are complete */
4928: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, rj_waits, status));
4929: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, sj_waits, status));
4931: /* send and recv i-structure */
4932: PetscCall(PetscCommGetNewTag(comm, &tagi));
4933: PetscCall(PetscPostIrecvInt(comm, tagi, merge->nrecv, merge->id_r, len_ri, &buf_ri, &ri_waits));
4935: PetscCall(PetscMalloc1(len + 1, &buf_s));
4936: buf_si = buf_s; /* points to the beginning of k-th msg to be sent */
4937: for (PetscMPIInt proc = 0, k = 0; proc < size; proc++) {
4938: if (!len_s[proc]) continue;
4939: /* form outgoing message for i-structure:
4940: buf_si[0]: nrows to be sent
4941: [1:nrows]: row index (global)
4942: [nrows+1:2*nrows+1]: i-structure index
4943: */
4944: nrows = len_si[proc] / 2 - 1;
4945: buf_si_i = buf_si + nrows + 1;
4946: buf_si[0] = nrows;
4947: buf_si_i[0] = 0;
4948: nrows = 0;
4949: for (i = owners[proc]; i < owners[proc + 1]; i++) {
4950: anzi = ai[i + 1] - ai[i];
4951: if (anzi) {
4952: buf_si_i[nrows + 1] = buf_si_i[nrows] + anzi; /* i-structure */
4953: buf_si[nrows + 1] = i - owners[proc]; /* local row index */
4954: nrows++;
4955: }
4956: }
4957: PetscCallMPI(MPIU_Isend(buf_si, len_si[proc], MPIU_INT, proc, tagi, comm, si_waits + k));
4958: k++;
4959: buf_si += len_si[proc];
4960: }
4962: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, ri_waits, status));
4963: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, si_waits, status));
4965: PetscCall(PetscInfo(seqmat, "nsend: %d, nrecv: %d\n", merge->nsend, merge->nrecv));
4966: 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]));
4968: PetscCall(PetscFree(len_si));
4969: PetscCall(PetscFree(len_ri));
4970: PetscCall(PetscFree(rj_waits));
4971: PetscCall(PetscFree2(si_waits, sj_waits));
4972: PetscCall(PetscFree(ri_waits));
4973: PetscCall(PetscFree(buf_s));
4974: PetscCall(PetscFree(status));
4976: /* compute a local seq matrix in each processor */
4977: /* allocate bi array and free space for accumulating nonzero column info */
4978: PetscCall(PetscMalloc1(m + 1, &bi));
4979: bi[0] = 0;
4981: /* create and initialize a linked list */
4982: nlnk = N + 1;
4983: PetscCall(PetscLLCreate(N, N, nlnk, lnk, lnkbt));
4985: /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */
4986: len = ai[owners[rank + 1]] - ai[owners[rank]];
4987: PetscCall(PetscFreeSpaceGet(PetscIntMultTruncate(2, len) + 1, &free_space));
4989: current_space = free_space;
4991: /* determine symbolic info for each local row */
4992: PetscCall(PetscMalloc3(merge->nrecv, &buf_ri_k, merge->nrecv, &nextrow, merge->nrecv, &nextai));
4994: for (k = 0; k < merge->nrecv; k++) {
4995: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4996: nrows = *buf_ri_k[k];
4997: nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */
4998: nextai[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure */
4999: }
5001: MatPreallocateBegin(comm, m, n, dnz, onz);
5002: len = 0;
5003: for (i = 0; i < m; i++) {
5004: bnzi = 0;
5005: /* add local non-zero cols of this proc's seqmat into lnk */
5006: arow = owners[rank] + i;
5007: anzi = ai[arow + 1] - ai[arow];
5008: aj = a->j + ai[arow];
5009: PetscCall(PetscLLAddSorted(anzi, aj, N, &nlnk, lnk, lnkbt));
5010: bnzi += nlnk;
5011: /* add received col data into lnk */
5012: for (k = 0; k < merge->nrecv; k++) { /* k-th received message */
5013: if (i == *nextrow[k]) { /* i-th row */
5014: anzi = *(nextai[k] + 1) - *nextai[k];
5015: aj = buf_rj[k] + *nextai[k];
5016: PetscCall(PetscLLAddSorted(anzi, aj, N, &nlnk, lnk, lnkbt));
5017: bnzi += nlnk;
5018: nextrow[k]++;
5019: nextai[k]++;
5020: }
5021: }
5022: if (len < bnzi) len = bnzi; /* =max(bnzi) */
5024: /* if free space is not available, make more free space */
5025: if (current_space->local_remaining < bnzi) PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(bnzi, current_space->total_array_size), ¤t_space));
5026: /* copy data into free space, then initialize lnk */
5027: PetscCall(PetscLLClean(N, N, bnzi, lnk, current_space->array, lnkbt));
5028: PetscCall(MatPreallocateSet(i + owners[rank], bnzi, current_space->array, dnz, onz));
5030: current_space->array += bnzi;
5031: current_space->local_used += bnzi;
5032: current_space->local_remaining -= bnzi;
5034: bi[i + 1] = bi[i] + bnzi;
5035: }
5037: PetscCall(PetscFree3(buf_ri_k, nextrow, nextai));
5039: PetscCall(PetscMalloc1(bi[m] + 1, &bj));
5040: PetscCall(PetscFreeSpaceContiguous(&free_space, bj));
5041: PetscCall(PetscLLDestroy(lnk, lnkbt));
5043: /* create symbolic parallel matrix B_mpi */
5044: PetscCall(MatGetBlockSizes(seqmat, &bs, &cbs));
5045: PetscCall(MatCreate(comm, &B_mpi));
5046: if (n == PETSC_DECIDE) {
5047: PetscCall(MatSetSizes(B_mpi, m, n, PETSC_DETERMINE, N));
5048: } else {
5049: PetscCall(MatSetSizes(B_mpi, m, n, PETSC_DETERMINE, PETSC_DETERMINE));
5050: }
5051: PetscCall(MatSetBlockSizes(B_mpi, bs, cbs));
5052: PetscCall(MatSetType(B_mpi, MATMPIAIJ));
5053: PetscCall(MatMPIAIJSetPreallocation(B_mpi, 0, dnz, 0, onz));
5054: MatPreallocateEnd(dnz, onz);
5055: PetscCall(MatSetOption(B_mpi, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_FALSE));
5057: /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */
5058: B_mpi->assembled = PETSC_FALSE;
5059: merge->bi = bi;
5060: merge->bj = bj;
5061: merge->buf_ri = buf_ri;
5062: merge->buf_rj = buf_rj;
5063: merge->coi = NULL;
5064: merge->coj = NULL;
5065: merge->owners_co = NULL;
5067: PetscCall(PetscCommDestroy(&comm));
5069: /* attach the supporting struct to B_mpi for reuse */
5070: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
5071: PetscCall(PetscContainerSetPointer(container, merge));
5072: PetscCall(PetscContainerSetUserDestroy(container, MatDestroy_MPIAIJ_SeqsToMPI));
5073: PetscCall(PetscObjectCompose((PetscObject)B_mpi, "MatMergeSeqsToMPI", (PetscObject)container));
5074: PetscCall(PetscContainerDestroy(&container));
5075: *mpimat = B_mpi;
5077: PetscCall(PetscLogEventEnd(MAT_Seqstompisym, seqmat, 0, 0, 0));
5078: PetscFunctionReturn(PETSC_SUCCESS);
5079: }
5081: /*@
5082: MatCreateMPIAIJSumSeqAIJ - Creates a `MATMPIAIJ` matrix by adding sequential
5083: matrices from each processor
5085: Collective
5087: Input Parameters:
5088: + comm - the communicators the parallel matrix will live on
5089: . seqmat - the input sequential matrices
5090: . m - number of local rows (or `PETSC_DECIDE`)
5091: . n - number of local columns (or `PETSC_DECIDE`)
5092: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5094: Output Parameter:
5095: . mpimat - the parallel matrix generated
5097: Level: advanced
5099: Note:
5100: The dimensions of the sequential matrix in each processor MUST be the same.
5101: The input seqmat is included into the container "Mat_Merge_SeqsToMPI", and will be
5102: destroyed when `mpimat` is destroyed. Call `PetscObjectQuery()` to access `seqmat`.
5104: .seealso: [](ch_matrices), `Mat`, `MatCreateAIJ()`
5105: @*/
5106: PetscErrorCode MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm, Mat seqmat, PetscInt m, PetscInt n, MatReuse scall, Mat *mpimat)
5107: {
5108: PetscMPIInt size;
5110: PetscFunctionBegin;
5111: PetscCallMPI(MPI_Comm_size(comm, &size));
5112: if (size == 1) {
5113: PetscCall(PetscLogEventBegin(MAT_Seqstompi, seqmat, 0, 0, 0));
5114: if (scall == MAT_INITIAL_MATRIX) {
5115: PetscCall(MatDuplicate(seqmat, MAT_COPY_VALUES, mpimat));
5116: } else {
5117: PetscCall(MatCopy(seqmat, *mpimat, SAME_NONZERO_PATTERN));
5118: }
5119: PetscCall(PetscLogEventEnd(MAT_Seqstompi, seqmat, 0, 0, 0));
5120: PetscFunctionReturn(PETSC_SUCCESS);
5121: }
5122: PetscCall(PetscLogEventBegin(MAT_Seqstompi, seqmat, 0, 0, 0));
5123: if (scall == MAT_INITIAL_MATRIX) PetscCall(MatCreateMPIAIJSumSeqAIJSymbolic(comm, seqmat, m, n, mpimat));
5124: PetscCall(MatCreateMPIAIJSumSeqAIJNumeric(seqmat, *mpimat));
5125: PetscCall(PetscLogEventEnd(MAT_Seqstompi, seqmat, 0, 0, 0));
5126: PetscFunctionReturn(PETSC_SUCCESS);
5127: }
5129: /*@
5130: MatAIJGetLocalMat - Creates a `MATSEQAIJ` from a `MATAIJ` matrix.
5132: Not Collective
5134: Input Parameter:
5135: . A - the matrix
5137: Output Parameter:
5138: . A_loc - the local sequential matrix generated
5140: Level: developer
5142: Notes:
5143: The matrix is created by taking `A`'s local rows and putting them into a sequential matrix
5144: with `mlocal` rows and `n` columns. Where `mlocal` is obtained with `MatGetLocalSize()` and
5145: `n` is the global column count obtained with `MatGetSize()`
5147: In other words combines the two parts of a parallel `MATMPIAIJ` matrix on each process to a single matrix.
5149: For parallel matrices this creates an entirely new matrix. If the matrix is sequential it merely increases the reference count.
5151: Destroy the matrix with `MatDestroy()`
5153: .seealso: [](ch_matrices), `Mat`, `MatMPIAIJGetLocalMat()`
5154: @*/
5155: PetscErrorCode MatAIJGetLocalMat(Mat A, Mat *A_loc)
5156: {
5157: PetscBool mpi;
5159: PetscFunctionBegin;
5160: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIAIJ, &mpi));
5161: if (mpi) {
5162: PetscCall(MatMPIAIJGetLocalMat(A, MAT_INITIAL_MATRIX, A_loc));
5163: } else {
5164: *A_loc = A;
5165: PetscCall(PetscObjectReference((PetscObject)*A_loc));
5166: }
5167: PetscFunctionReturn(PETSC_SUCCESS);
5168: }
5170: /*@
5171: MatMPIAIJGetLocalMat - Creates a `MATSEQAIJ` from a `MATMPIAIJ` matrix.
5173: Not Collective
5175: Input Parameters:
5176: + A - the matrix
5177: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5179: Output Parameter:
5180: . A_loc - the local sequential matrix generated
5182: Level: developer
5184: Notes:
5185: The matrix is created by taking all `A`'s local rows and putting them into a sequential
5186: matrix with `mlocal` rows and `n` columns.`mlocal` is the row count obtained with
5187: `MatGetLocalSize()` and `n` is the global column count obtained with `MatGetSize()`.
5189: In other words combines the two parts of a parallel `MATMPIAIJ` matrix on each process to a single matrix.
5191: When `A` is sequential and `MAT_INITIAL_MATRIX` is requested, the matrix returned is the diagonal part of `A` (which contains the entire matrix),
5192: 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
5193: then `MatCopy`(Adiag,*`A_loc`,`SAME_NONZERO_PATTERN`) is called to fill `A_loc`. Thus one can preallocate the appropriate sequential matrix `A_loc`
5194: 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.
5196: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMatCondensed()`, `MatMPIAIJGetLocalMatMerge()`
5197: @*/
5198: PetscErrorCode MatMPIAIJGetLocalMat(Mat A, MatReuse scall, Mat *A_loc)
5199: {
5200: Mat_MPIAIJ *mpimat = (Mat_MPIAIJ *)A->data;
5201: Mat_SeqAIJ *mat, *a, *b;
5202: PetscInt *ai, *aj, *bi, *bj, *cmap = mpimat->garray;
5203: const PetscScalar *aa, *ba, *aav, *bav;
5204: PetscScalar *ca, *cam;
5205: PetscMPIInt size;
5206: PetscInt am = A->rmap->n, i, j, k, cstart = A->cmap->rstart;
5207: PetscInt *ci, *cj, col, ncols_d, ncols_o, jo;
5208: PetscBool match;
5210: PetscFunctionBegin;
5211: PetscCall(PetscStrbeginswith(((PetscObject)A)->type_name, MATMPIAIJ, &match));
5212: PetscCheck(match, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Requires MATMPIAIJ matrix as input");
5213: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
5214: if (size == 1) {
5215: if (scall == MAT_INITIAL_MATRIX) {
5216: PetscCall(PetscObjectReference((PetscObject)mpimat->A));
5217: *A_loc = mpimat->A;
5218: } else if (scall == MAT_REUSE_MATRIX) {
5219: PetscCall(MatCopy(mpimat->A, *A_loc, SAME_NONZERO_PATTERN));
5220: }
5221: PetscFunctionReturn(PETSC_SUCCESS);
5222: }
5224: PetscCall(PetscLogEventBegin(MAT_Getlocalmat, A, 0, 0, 0));
5225: a = (Mat_SeqAIJ *)mpimat->A->data;
5226: b = (Mat_SeqAIJ *)mpimat->B->data;
5227: ai = a->i;
5228: aj = a->j;
5229: bi = b->i;
5230: bj = b->j;
5231: PetscCall(MatSeqAIJGetArrayRead(mpimat->A, &aav));
5232: PetscCall(MatSeqAIJGetArrayRead(mpimat->B, &bav));
5233: aa = aav;
5234: ba = bav;
5235: if (scall == MAT_INITIAL_MATRIX) {
5236: PetscCall(PetscMalloc1(1 + am, &ci));
5237: ci[0] = 0;
5238: for (i = 0; i < am; i++) ci[i + 1] = ci[i] + (ai[i + 1] - ai[i]) + (bi[i + 1] - bi[i]);
5239: PetscCall(PetscMalloc1(1 + ci[am], &cj));
5240: PetscCall(PetscMalloc1(1 + ci[am], &ca));
5241: k = 0;
5242: for (i = 0; i < am; i++) {
5243: ncols_o = bi[i + 1] - bi[i];
5244: ncols_d = ai[i + 1] - ai[i];
5245: /* off-diagonal portion of A */
5246: for (jo = 0; jo < ncols_o; jo++) {
5247: col = cmap[*bj];
5248: if (col >= cstart) break;
5249: cj[k] = col;
5250: bj++;
5251: ca[k++] = *ba++;
5252: }
5253: /* diagonal portion of A */
5254: for (j = 0; j < ncols_d; j++) {
5255: cj[k] = cstart + *aj++;
5256: ca[k++] = *aa++;
5257: }
5258: /* off-diagonal portion of A */
5259: for (j = jo; j < ncols_o; j++) {
5260: cj[k] = cmap[*bj++];
5261: ca[k++] = *ba++;
5262: }
5263: }
5264: /* put together the new matrix */
5265: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, am, A->cmap->N, ci, cj, ca, A_loc));
5266: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5267: /* Since these are PETSc arrays, change flags to free them as necessary. */
5268: mat = (Mat_SeqAIJ *)(*A_loc)->data;
5269: mat->free_a = PETSC_TRUE;
5270: mat->free_ij = PETSC_TRUE;
5271: mat->nonew = 0;
5272: } else if (scall == MAT_REUSE_MATRIX) {
5273: mat = (Mat_SeqAIJ *)(*A_loc)->data;
5274: ci = mat->i;
5275: cj = mat->j;
5276: PetscCall(MatSeqAIJGetArrayWrite(*A_loc, &cam));
5277: for (i = 0; i < am; i++) {
5278: /* off-diagonal portion of A */
5279: ncols_o = bi[i + 1] - bi[i];
5280: for (jo = 0; jo < ncols_o; jo++) {
5281: col = cmap[*bj];
5282: if (col >= cstart) break;
5283: *cam++ = *ba++;
5284: bj++;
5285: }
5286: /* diagonal portion of A */
5287: ncols_d = ai[i + 1] - ai[i];
5288: for (j = 0; j < ncols_d; j++) *cam++ = *aa++;
5289: /* off-diagonal portion of A */
5290: for (j = jo; j < ncols_o; j++) {
5291: *cam++ = *ba++;
5292: bj++;
5293: }
5294: }
5295: PetscCall(MatSeqAIJRestoreArrayWrite(*A_loc, &cam));
5296: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Invalid MatReuse %d", (int)scall);
5297: PetscCall(MatSeqAIJRestoreArrayRead(mpimat->A, &aav));
5298: PetscCall(MatSeqAIJRestoreArrayRead(mpimat->B, &bav));
5299: PetscCall(PetscLogEventEnd(MAT_Getlocalmat, A, 0, 0, 0));
5300: PetscFunctionReturn(PETSC_SUCCESS);
5301: }
5303: /*@
5304: MatMPIAIJGetLocalMatMerge - Creates a `MATSEQAIJ` from a `MATMPIAIJ` matrix by taking all its local rows and putting them into a sequential matrix with
5305: mlocal rows and n columns. Where n is the sum of the number of columns of the diagonal and off-diagonal part
5307: Not Collective
5309: Input Parameters:
5310: + A - the matrix
5311: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5313: Output Parameters:
5314: + glob - sequential `IS` with global indices associated with the columns of the local sequential matrix generated (can be `NULL`)
5315: - A_loc - the local sequential matrix generated
5317: Level: developer
5319: Note:
5320: This is different from `MatMPIAIJGetLocalMat()` since the first columns in the returning matrix are those associated with the diagonal
5321: part, then those associated with the off-diagonal part (in its local ordering)
5323: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMat()`, `MatMPIAIJGetLocalMatCondensed()`
5324: @*/
5325: PetscErrorCode MatMPIAIJGetLocalMatMerge(Mat A, MatReuse scall, IS *glob, Mat *A_loc)
5326: {
5327: Mat Ao, Ad;
5328: const PetscInt *cmap;
5329: PetscMPIInt size;
5330: PetscErrorCode (*f)(Mat, MatReuse, IS *, Mat *);
5332: PetscFunctionBegin;
5333: PetscCall(MatMPIAIJGetSeqAIJ(A, &Ad, &Ao, &cmap));
5334: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
5335: if (size == 1) {
5336: if (scall == MAT_INITIAL_MATRIX) {
5337: PetscCall(PetscObjectReference((PetscObject)Ad));
5338: *A_loc = Ad;
5339: } else if (scall == MAT_REUSE_MATRIX) {
5340: PetscCall(MatCopy(Ad, *A_loc, SAME_NONZERO_PATTERN));
5341: }
5342: if (glob) PetscCall(ISCreateStride(PetscObjectComm((PetscObject)Ad), Ad->cmap->n, Ad->cmap->rstart, 1, glob));
5343: PetscFunctionReturn(PETSC_SUCCESS);
5344: }
5345: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatMPIAIJGetLocalMatMerge_C", &f));
5346: PetscCall(PetscLogEventBegin(MAT_Getlocalmat, A, 0, 0, 0));
5347: if (f) {
5348: PetscCall((*f)(A, scall, glob, A_loc));
5349: } else {
5350: Mat_SeqAIJ *a = (Mat_SeqAIJ *)Ad->data;
5351: Mat_SeqAIJ *b = (Mat_SeqAIJ *)Ao->data;
5352: Mat_SeqAIJ *c;
5353: PetscInt *ai = a->i, *aj = a->j;
5354: PetscInt *bi = b->i, *bj = b->j;
5355: PetscInt *ci, *cj;
5356: const PetscScalar *aa, *ba;
5357: PetscScalar *ca;
5358: PetscInt i, j, am, dn, on;
5360: PetscCall(MatGetLocalSize(Ad, &am, &dn));
5361: PetscCall(MatGetLocalSize(Ao, NULL, &on));
5362: PetscCall(MatSeqAIJGetArrayRead(Ad, &aa));
5363: PetscCall(MatSeqAIJGetArrayRead(Ao, &ba));
5364: if (scall == MAT_INITIAL_MATRIX) {
5365: PetscInt k;
5366: PetscCall(PetscMalloc1(1 + am, &ci));
5367: PetscCall(PetscMalloc1(ai[am] + bi[am], &cj));
5368: PetscCall(PetscMalloc1(ai[am] + bi[am], &ca));
5369: ci[0] = 0;
5370: for (i = 0, k = 0; i < am; i++) {
5371: const PetscInt ncols_o = bi[i + 1] - bi[i];
5372: const PetscInt ncols_d = ai[i + 1] - ai[i];
5373: ci[i + 1] = ci[i] + ncols_o + ncols_d;
5374: /* diagonal portion of A */
5375: for (j = 0; j < ncols_d; j++, k++) {
5376: cj[k] = *aj++;
5377: ca[k] = *aa++;
5378: }
5379: /* off-diagonal portion of A */
5380: for (j = 0; j < ncols_o; j++, k++) {
5381: cj[k] = dn + *bj++;
5382: ca[k] = *ba++;
5383: }
5384: }
5385: /* put together the new matrix */
5386: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, am, dn + on, ci, cj, ca, A_loc));
5387: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5388: /* Since these are PETSc arrays, change flags to free them as necessary. */
5389: c = (Mat_SeqAIJ *)(*A_loc)->data;
5390: c->free_a = PETSC_TRUE;
5391: c->free_ij = PETSC_TRUE;
5392: c->nonew = 0;
5393: PetscCall(MatSetType(*A_loc, ((PetscObject)Ad)->type_name));
5394: } else if (scall == MAT_REUSE_MATRIX) {
5395: PetscCall(MatSeqAIJGetArrayWrite(*A_loc, &ca));
5396: for (i = 0; i < am; i++) {
5397: const PetscInt ncols_d = ai[i + 1] - ai[i];
5398: const PetscInt ncols_o = bi[i + 1] - bi[i];
5399: /* diagonal portion of A */
5400: for (j = 0; j < ncols_d; j++) *ca++ = *aa++;
5401: /* off-diagonal portion of A */
5402: for (j = 0; j < ncols_o; j++) *ca++ = *ba++;
5403: }
5404: PetscCall(MatSeqAIJRestoreArrayWrite(*A_loc, &ca));
5405: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Invalid MatReuse %d", (int)scall);
5406: PetscCall(MatSeqAIJRestoreArrayRead(Ad, &aa));
5407: PetscCall(MatSeqAIJRestoreArrayRead(Ao, &aa));
5408: if (glob) {
5409: PetscInt cst, *gidx;
5411: PetscCall(MatGetOwnershipRangeColumn(A, &cst, NULL));
5412: PetscCall(PetscMalloc1(dn + on, &gidx));
5413: for (i = 0; i < dn; i++) gidx[i] = cst + i;
5414: for (i = 0; i < on; i++) gidx[i + dn] = cmap[i];
5415: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)Ad), dn + on, gidx, PETSC_OWN_POINTER, glob));
5416: }
5417: }
5418: PetscCall(PetscLogEventEnd(MAT_Getlocalmat, A, 0, 0, 0));
5419: PetscFunctionReturn(PETSC_SUCCESS);
5420: }
5422: /*@C
5423: MatMPIAIJGetLocalMatCondensed - Creates a `MATSEQAIJ` matrix from an `MATMPIAIJ` matrix by taking all its local rows and NON-ZERO columns
5425: Not Collective
5427: Input Parameters:
5428: + A - the matrix
5429: . scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5430: . row - index set of rows to extract (or `NULL`)
5431: - col - index set of columns to extract (or `NULL`)
5433: Output Parameter:
5434: . A_loc - the local sequential matrix generated
5436: Level: developer
5438: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMat()`
5439: @*/
5440: PetscErrorCode MatMPIAIJGetLocalMatCondensed(Mat A, MatReuse scall, IS *row, IS *col, Mat *A_loc)
5441: {
5442: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5443: PetscInt i, start, end, ncols, nzA, nzB, *cmap, imark, *idx;
5444: IS isrowa, iscola;
5445: Mat *aloc;
5446: PetscBool match;
5448: PetscFunctionBegin;
5449: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIAIJ, &match));
5450: PetscCheck(match, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Requires MATMPIAIJ matrix as input");
5451: PetscCall(PetscLogEventBegin(MAT_Getlocalmatcondensed, A, 0, 0, 0));
5452: if (!row) {
5453: start = A->rmap->rstart;
5454: end = A->rmap->rend;
5455: PetscCall(ISCreateStride(PETSC_COMM_SELF, end - start, start, 1, &isrowa));
5456: } else {
5457: isrowa = *row;
5458: }
5459: if (!col) {
5460: start = A->cmap->rstart;
5461: cmap = a->garray;
5462: nzA = a->A->cmap->n;
5463: nzB = a->B->cmap->n;
5464: PetscCall(PetscMalloc1(nzA + nzB, &idx));
5465: ncols = 0;
5466: for (i = 0; i < nzB; i++) {
5467: if (cmap[i] < start) idx[ncols++] = cmap[i];
5468: else break;
5469: }
5470: imark = i;
5471: for (i = 0; i < nzA; i++) idx[ncols++] = start + i;
5472: for (i = imark; i < nzB; i++) idx[ncols++] = cmap[i];
5473: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, &iscola));
5474: } else {
5475: iscola = *col;
5476: }
5477: if (scall != MAT_INITIAL_MATRIX) {
5478: PetscCall(PetscMalloc1(1, &aloc));
5479: aloc[0] = *A_loc;
5480: }
5481: PetscCall(MatCreateSubMatrices(A, 1, &isrowa, &iscola, scall, &aloc));
5482: if (!col) { /* attach global id of condensed columns */
5483: PetscCall(PetscObjectCompose((PetscObject)aloc[0], "_petsc_GetLocalMatCondensed_iscol", (PetscObject)iscola));
5484: }
5485: *A_loc = aloc[0];
5486: PetscCall(PetscFree(aloc));
5487: if (!row) PetscCall(ISDestroy(&isrowa));
5488: if (!col) PetscCall(ISDestroy(&iscola));
5489: PetscCall(PetscLogEventEnd(MAT_Getlocalmatcondensed, A, 0, 0, 0));
5490: PetscFunctionReturn(PETSC_SUCCESS);
5491: }
5493: /*
5494: * Create a sequential AIJ matrix based on row indices. a whole column is extracted once a row is matched.
5495: * Row could be local or remote.The routine is designed to be scalable in memory so that nothing is based
5496: * on a global size.
5497: * */
5498: static PetscErrorCode MatCreateSeqSubMatrixWithRows_Private(Mat P, IS rows, Mat *P_oth)
5499: {
5500: Mat_MPIAIJ *p = (Mat_MPIAIJ *)P->data;
5501: Mat_SeqAIJ *pd = (Mat_SeqAIJ *)p->A->data, *po = (Mat_SeqAIJ *)p->B->data, *p_oth;
5502: PetscInt plocalsize, nrows, *ilocal, *oilocal, i, lidx, *nrcols, *nlcols, ncol;
5503: PetscMPIInt owner;
5504: PetscSFNode *iremote, *oiremote;
5505: const PetscInt *lrowindices;
5506: PetscSF sf, osf;
5507: PetscInt pcstart, *roffsets, *loffsets, *pnnz, j;
5508: PetscInt ontotalcols, dntotalcols, ntotalcols, nout;
5509: MPI_Comm comm;
5510: ISLocalToGlobalMapping mapping;
5511: const PetscScalar *pd_a, *po_a;
5513: PetscFunctionBegin;
5514: PetscCall(PetscObjectGetComm((PetscObject)P, &comm));
5515: /* plocalsize is the number of roots
5516: * nrows is the number of leaves
5517: * */
5518: PetscCall(MatGetLocalSize(P, &plocalsize, NULL));
5519: PetscCall(ISGetLocalSize(rows, &nrows));
5520: PetscCall(PetscCalloc1(nrows, &iremote));
5521: PetscCall(ISGetIndices(rows, &lrowindices));
5522: for (i = 0; i < nrows; i++) {
5523: /* Find a remote index and an owner for a row
5524: * The row could be local or remote
5525: * */
5526: owner = 0;
5527: lidx = 0;
5528: PetscCall(PetscLayoutFindOwnerIndex(P->rmap, lrowindices[i], &owner, &lidx));
5529: iremote[i].index = lidx;
5530: iremote[i].rank = owner;
5531: }
5532: /* Create SF to communicate how many nonzero columns for each row */
5533: PetscCall(PetscSFCreate(comm, &sf));
5534: /* SF will figure out the number of nonzero columns for each row, and their
5535: * offsets
5536: * */
5537: PetscCall(PetscSFSetGraph(sf, plocalsize, nrows, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
5538: PetscCall(PetscSFSetFromOptions(sf));
5539: PetscCall(PetscSFSetUp(sf));
5541: PetscCall(PetscCalloc1(2 * (plocalsize + 1), &roffsets));
5542: PetscCall(PetscCalloc1(2 * plocalsize, &nrcols));
5543: PetscCall(PetscCalloc1(nrows, &pnnz));
5544: roffsets[0] = 0;
5545: roffsets[1] = 0;
5546: for (i = 0; i < plocalsize; i++) {
5547: /* diagonal */
5548: nrcols[i * 2 + 0] = pd->i[i + 1] - pd->i[i];
5549: /* off-diagonal */
5550: nrcols[i * 2 + 1] = po->i[i + 1] - po->i[i];
5551: /* compute offsets so that we relative location for each row */
5552: roffsets[(i + 1) * 2 + 0] = roffsets[i * 2 + 0] + nrcols[i * 2 + 0];
5553: roffsets[(i + 1) * 2 + 1] = roffsets[i * 2 + 1] + nrcols[i * 2 + 1];
5554: }
5555: PetscCall(PetscCalloc1(2 * nrows, &nlcols));
5556: PetscCall(PetscCalloc1(2 * nrows, &loffsets));
5557: /* 'r' means root, and 'l' means leaf */
5558: PetscCall(PetscSFBcastBegin(sf, MPIU_2INT, nrcols, nlcols, MPI_REPLACE));
5559: PetscCall(PetscSFBcastBegin(sf, MPIU_2INT, roffsets, loffsets, MPI_REPLACE));
5560: PetscCall(PetscSFBcastEnd(sf, MPIU_2INT, nrcols, nlcols, MPI_REPLACE));
5561: PetscCall(PetscSFBcastEnd(sf, MPIU_2INT, roffsets, loffsets, MPI_REPLACE));
5562: PetscCall(PetscSFDestroy(&sf));
5563: PetscCall(PetscFree(roffsets));
5564: PetscCall(PetscFree(nrcols));
5565: dntotalcols = 0;
5566: ontotalcols = 0;
5567: ncol = 0;
5568: for (i = 0; i < nrows; i++) {
5569: pnnz[i] = nlcols[i * 2 + 0] + nlcols[i * 2 + 1];
5570: ncol = PetscMax(pnnz[i], ncol);
5571: /* diagonal */
5572: dntotalcols += nlcols[i * 2 + 0];
5573: /* off-diagonal */
5574: ontotalcols += nlcols[i * 2 + 1];
5575: }
5576: /* We do not need to figure the right number of columns
5577: * since all the calculations will be done by going through the raw data
5578: * */
5579: PetscCall(MatCreateSeqAIJ(PETSC_COMM_SELF, nrows, ncol, 0, pnnz, P_oth));
5580: PetscCall(MatSetUp(*P_oth));
5581: PetscCall(PetscFree(pnnz));
5582: p_oth = (Mat_SeqAIJ *)(*P_oth)->data;
5583: /* diagonal */
5584: PetscCall(PetscCalloc1(dntotalcols, &iremote));
5585: /* off-diagonal */
5586: PetscCall(PetscCalloc1(ontotalcols, &oiremote));
5587: /* diagonal */
5588: PetscCall(PetscCalloc1(dntotalcols, &ilocal));
5589: /* off-diagonal */
5590: PetscCall(PetscCalloc1(ontotalcols, &oilocal));
5591: dntotalcols = 0;
5592: ontotalcols = 0;
5593: ntotalcols = 0;
5594: for (i = 0; i < nrows; i++) {
5595: owner = 0;
5596: PetscCall(PetscLayoutFindOwnerIndex(P->rmap, lrowindices[i], &owner, NULL));
5597: /* Set iremote for diag matrix */
5598: for (j = 0; j < nlcols[i * 2 + 0]; j++) {
5599: iremote[dntotalcols].index = loffsets[i * 2 + 0] + j;
5600: iremote[dntotalcols].rank = owner;
5601: /* P_oth is seqAIJ so that ilocal need to point to the first part of memory */
5602: ilocal[dntotalcols++] = ntotalcols++;
5603: }
5604: /* off-diagonal */
5605: for (j = 0; j < nlcols[i * 2 + 1]; j++) {
5606: oiremote[ontotalcols].index = loffsets[i * 2 + 1] + j;
5607: oiremote[ontotalcols].rank = owner;
5608: oilocal[ontotalcols++] = ntotalcols++;
5609: }
5610: }
5611: PetscCall(ISRestoreIndices(rows, &lrowindices));
5612: PetscCall(PetscFree(loffsets));
5613: PetscCall(PetscFree(nlcols));
5614: PetscCall(PetscSFCreate(comm, &sf));
5615: /* P serves as roots and P_oth is leaves
5616: * Diag matrix
5617: * */
5618: PetscCall(PetscSFSetGraph(sf, pd->i[plocalsize], dntotalcols, ilocal, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
5619: PetscCall(PetscSFSetFromOptions(sf));
5620: PetscCall(PetscSFSetUp(sf));
5622: PetscCall(PetscSFCreate(comm, &osf));
5623: /* off-diagonal */
5624: PetscCall(PetscSFSetGraph(osf, po->i[plocalsize], ontotalcols, oilocal, PETSC_OWN_POINTER, oiremote, PETSC_OWN_POINTER));
5625: PetscCall(PetscSFSetFromOptions(osf));
5626: PetscCall(PetscSFSetUp(osf));
5627: PetscCall(MatSeqAIJGetArrayRead(p->A, &pd_a));
5628: PetscCall(MatSeqAIJGetArrayRead(p->B, &po_a));
5629: /* operate on the matrix internal data to save memory */
5630: PetscCall(PetscSFBcastBegin(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5631: PetscCall(PetscSFBcastBegin(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5632: PetscCall(MatGetOwnershipRangeColumn(P, &pcstart, NULL));
5633: /* Convert to global indices for diag matrix */
5634: for (i = 0; i < pd->i[plocalsize]; i++) pd->j[i] += pcstart;
5635: PetscCall(PetscSFBcastBegin(sf, MPIU_INT, pd->j, p_oth->j, MPI_REPLACE));
5636: /* We want P_oth store global indices */
5637: PetscCall(ISLocalToGlobalMappingCreate(comm, 1, p->B->cmap->n, p->garray, PETSC_COPY_VALUES, &mapping));
5638: /* Use memory scalable approach */
5639: PetscCall(ISLocalToGlobalMappingSetType(mapping, ISLOCALTOGLOBALMAPPINGHASH));
5640: PetscCall(ISLocalToGlobalMappingApply(mapping, po->i[plocalsize], po->j, po->j));
5641: PetscCall(PetscSFBcastBegin(osf, MPIU_INT, po->j, p_oth->j, MPI_REPLACE));
5642: PetscCall(PetscSFBcastEnd(sf, MPIU_INT, pd->j, p_oth->j, MPI_REPLACE));
5643: /* Convert back to local indices */
5644: for (i = 0; i < pd->i[plocalsize]; i++) pd->j[i] -= pcstart;
5645: PetscCall(PetscSFBcastEnd(osf, MPIU_INT, po->j, p_oth->j, MPI_REPLACE));
5646: nout = 0;
5647: PetscCall(ISGlobalToLocalMappingApply(mapping, IS_GTOLM_DROP, po->i[plocalsize], po->j, &nout, po->j));
5648: PetscCheck(nout == po->i[plocalsize], comm, PETSC_ERR_ARG_INCOMP, "n %" PetscInt_FMT " does not equal to nout %" PetscInt_FMT " ", po->i[plocalsize], nout);
5649: PetscCall(ISLocalToGlobalMappingDestroy(&mapping));
5650: /* Exchange values */
5651: PetscCall(PetscSFBcastEnd(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5652: PetscCall(PetscSFBcastEnd(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5653: PetscCall(MatSeqAIJRestoreArrayRead(p->A, &pd_a));
5654: PetscCall(MatSeqAIJRestoreArrayRead(p->B, &po_a));
5655: /* Stop PETSc from shrinking memory */
5656: for (i = 0; i < nrows; i++) p_oth->ilen[i] = p_oth->imax[i];
5657: PetscCall(MatAssemblyBegin(*P_oth, MAT_FINAL_ASSEMBLY));
5658: PetscCall(MatAssemblyEnd(*P_oth, MAT_FINAL_ASSEMBLY));
5659: /* Attach PetscSF objects to P_oth so that we can reuse it later */
5660: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "diagsf", (PetscObject)sf));
5661: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "offdiagsf", (PetscObject)osf));
5662: PetscCall(PetscSFDestroy(&sf));
5663: PetscCall(PetscSFDestroy(&osf));
5664: PetscFunctionReturn(PETSC_SUCCESS);
5665: }
5667: /*
5668: * Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A
5669: * This supports MPIAIJ and MAIJ
5670: * */
5671: PetscErrorCode MatGetBrowsOfAcols_MPIXAIJ(Mat A, Mat P, PetscInt dof, MatReuse reuse, Mat *P_oth)
5672: {
5673: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data, *p = (Mat_MPIAIJ *)P->data;
5674: Mat_SeqAIJ *p_oth;
5675: IS rows, map;
5676: PetscHMapI hamp;
5677: PetscInt i, htsize, *rowindices, off, *mapping, key, count;
5678: MPI_Comm comm;
5679: PetscSF sf, osf;
5680: PetscBool has;
5682: PetscFunctionBegin;
5683: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
5684: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAocols, A, P, 0, 0));
5685: /* If it is the first time, create an index set of off-diag nonzero columns of A,
5686: * and then create a submatrix (that often is an overlapping matrix)
5687: * */
5688: if (reuse == MAT_INITIAL_MATRIX) {
5689: /* Use a hash table to figure out unique keys */
5690: PetscCall(PetscHMapICreateWithSize(a->B->cmap->n, &hamp));
5691: PetscCall(PetscCalloc1(a->B->cmap->n, &mapping));
5692: count = 0;
5693: /* Assume that a->g is sorted, otherwise the following does not make sense */
5694: for (i = 0; i < a->B->cmap->n; i++) {
5695: key = a->garray[i] / dof;
5696: PetscCall(PetscHMapIHas(hamp, key, &has));
5697: if (!has) {
5698: mapping[i] = count;
5699: PetscCall(PetscHMapISet(hamp, key, count++));
5700: } else {
5701: /* Current 'i' has the same value the previous step */
5702: mapping[i] = count - 1;
5703: }
5704: }
5705: PetscCall(ISCreateGeneral(comm, a->B->cmap->n, mapping, PETSC_OWN_POINTER, &map));
5706: PetscCall(PetscHMapIGetSize(hamp, &htsize));
5707: PetscCheck(htsize == count, comm, PETSC_ERR_ARG_INCOMP, " Size of hash map %" PetscInt_FMT " is inconsistent with count %" PetscInt_FMT, htsize, count);
5708: PetscCall(PetscCalloc1(htsize, &rowindices));
5709: off = 0;
5710: PetscCall(PetscHMapIGetKeys(hamp, &off, rowindices));
5711: PetscCall(PetscHMapIDestroy(&hamp));
5712: PetscCall(PetscSortInt(htsize, rowindices));
5713: PetscCall(ISCreateGeneral(comm, htsize, rowindices, PETSC_OWN_POINTER, &rows));
5714: /* In case, the matrix was already created but users want to recreate the matrix */
5715: PetscCall(MatDestroy(P_oth));
5716: PetscCall(MatCreateSeqSubMatrixWithRows_Private(P, rows, P_oth));
5717: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "aoffdiagtopothmapping", (PetscObject)map));
5718: PetscCall(ISDestroy(&map));
5719: PetscCall(ISDestroy(&rows));
5720: } else if (reuse == MAT_REUSE_MATRIX) {
5721: /* If matrix was already created, we simply update values using SF objects
5722: * that as attached to the matrix earlier.
5723: */
5724: const PetscScalar *pd_a, *po_a;
5726: PetscCall(PetscObjectQuery((PetscObject)*P_oth, "diagsf", (PetscObject *)&sf));
5727: PetscCall(PetscObjectQuery((PetscObject)*P_oth, "offdiagsf", (PetscObject *)&osf));
5728: PetscCheck(sf && osf, comm, PETSC_ERR_ARG_NULL, "Matrix is not initialized yet");
5729: p_oth = (Mat_SeqAIJ *)(*P_oth)->data;
5730: /* Update values in place */
5731: PetscCall(MatSeqAIJGetArrayRead(p->A, &pd_a));
5732: PetscCall(MatSeqAIJGetArrayRead(p->B, &po_a));
5733: PetscCall(PetscSFBcastBegin(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5734: PetscCall(PetscSFBcastBegin(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5735: PetscCall(PetscSFBcastEnd(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5736: PetscCall(PetscSFBcastEnd(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5737: PetscCall(MatSeqAIJRestoreArrayRead(p->A, &pd_a));
5738: PetscCall(MatSeqAIJRestoreArrayRead(p->B, &po_a));
5739: } else SETERRQ(comm, PETSC_ERR_ARG_UNKNOWN_TYPE, "Unknown reuse type");
5740: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAocols, A, P, 0, 0));
5741: PetscFunctionReturn(PETSC_SUCCESS);
5742: }
5744: /*@C
5745: MatGetBrowsOfAcols - Returns `IS` that contain rows of `B` that equal to nonzero columns of local `A`
5747: Collective
5749: Input Parameters:
5750: + A - the first matrix in `MATMPIAIJ` format
5751: . B - the second matrix in `MATMPIAIJ` format
5752: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5754: Output Parameters:
5755: + rowb - On input index sets of rows of B to extract (or `NULL`), modified on output
5756: . colb - On input index sets of columns of B to extract (or `NULL`), modified on output
5757: - B_seq - the sequential matrix generated
5759: Level: developer
5761: .seealso: `Mat`, `MATMPIAIJ`, `IS`, `MatReuse`
5762: @*/
5763: PetscErrorCode MatGetBrowsOfAcols(Mat A, Mat B, MatReuse scall, IS *rowb, IS *colb, Mat *B_seq)
5764: {
5765: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5766: PetscInt *idx, i, start, ncols, nzA, nzB, *cmap, imark;
5767: IS isrowb, iscolb;
5768: Mat *bseq = NULL;
5770: PetscFunctionBegin;
5771: 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 ")",
5772: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
5773: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAcols, A, B, 0, 0));
5775: if (scall == MAT_INITIAL_MATRIX) {
5776: start = A->cmap->rstart;
5777: cmap = a->garray;
5778: nzA = a->A->cmap->n;
5779: nzB = a->B->cmap->n;
5780: PetscCall(PetscMalloc1(nzA + nzB, &idx));
5781: ncols = 0;
5782: for (i = 0; i < nzB; i++) { /* row < local row index */
5783: if (cmap[i] < start) idx[ncols++] = cmap[i];
5784: else break;
5785: }
5786: imark = i;
5787: for (i = 0; i < nzA; i++) idx[ncols++] = start + i; /* local rows */
5788: for (i = imark; i < nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */
5789: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, &isrowb));
5790: PetscCall(ISCreateStride(PETSC_COMM_SELF, B->cmap->N, 0, 1, &iscolb));
5791: } else {
5792: PetscCheck(rowb && colb, PETSC_COMM_SELF, PETSC_ERR_SUP, "IS rowb and colb must be provided for MAT_REUSE_MATRIX");
5793: isrowb = *rowb;
5794: iscolb = *colb;
5795: PetscCall(PetscMalloc1(1, &bseq));
5796: bseq[0] = *B_seq;
5797: }
5798: PetscCall(MatCreateSubMatrices(B, 1, &isrowb, &iscolb, scall, &bseq));
5799: *B_seq = bseq[0];
5800: PetscCall(PetscFree(bseq));
5801: if (!rowb) {
5802: PetscCall(ISDestroy(&isrowb));
5803: } else {
5804: *rowb = isrowb;
5805: }
5806: if (!colb) {
5807: PetscCall(ISDestroy(&iscolb));
5808: } else {
5809: *colb = iscolb;
5810: }
5811: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAcols, A, B, 0, 0));
5812: PetscFunctionReturn(PETSC_SUCCESS);
5813: }
5815: /*
5816: MatGetBrowsOfAoCols_MPIAIJ - Creates a `MATSEQAIJ` matrix by taking rows of B that equal to nonzero columns
5817: of the OFF-DIAGONAL portion of local A
5819: Collective
5821: Input Parameters:
5822: + A,B - the matrices in `MATMPIAIJ` format
5823: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5825: Output Parameter:
5826: + startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or NULL)
5827: . startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or NULL)
5828: . bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or NULL)
5829: - B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N
5831: Developer Note:
5832: This directly accesses information inside the VecScatter associated with the matrix-vector product
5833: for this matrix. This is not desirable..
5835: Level: developer
5837: */
5839: PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat A, Mat B, MatReuse scall, PetscInt **startsj_s, PetscInt **startsj_r, MatScalar **bufa_ptr, Mat *B_oth)
5840: {
5841: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5842: VecScatter ctx;
5843: MPI_Comm comm;
5844: const PetscMPIInt *rprocs, *sprocs;
5845: PetscMPIInt nrecvs, nsends;
5846: const PetscInt *srow, *rstarts, *sstarts;
5847: PetscInt *rowlen, *bufj, *bufJ, ncols = 0, aBn = a->B->cmap->n, row, *b_othi, *b_othj, *rvalues = NULL, *svalues = NULL, *cols, sbs, rbs;
5848: PetscInt i, j, k = 0, l, ll, nrows, *rstartsj = NULL, *sstartsj, len;
5849: PetscScalar *b_otha, *bufa, *bufA, *vals = NULL;
5850: MPI_Request *reqs = NULL, *rwaits = NULL, *swaits = NULL;
5851: PetscMPIInt size, tag, rank, nreqs;
5853: PetscFunctionBegin;
5854: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
5855: PetscCallMPI(MPI_Comm_size(comm, &size));
5857: 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 ")",
5858: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
5859: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAocols, A, B, 0, 0));
5860: PetscCallMPI(MPI_Comm_rank(comm, &rank));
5862: if (size == 1) {
5863: startsj_s = NULL;
5864: bufa_ptr = NULL;
5865: *B_oth = NULL;
5866: PetscFunctionReturn(PETSC_SUCCESS);
5867: }
5869: ctx = a->Mvctx;
5870: tag = ((PetscObject)ctx)->tag;
5872: PetscCall(VecScatterGetRemote_Private(ctx, PETSC_TRUE /*send*/, &nsends, &sstarts, &srow, &sprocs, &sbs));
5873: /* rprocs[] must be ordered so that indices received from them are ordered in rvalues[], which is key to algorithms used in this subroutine */
5874: PetscCall(VecScatterGetRemoteOrdered_Private(ctx, PETSC_FALSE /*recv*/, &nrecvs, &rstarts, NULL /*indices not needed*/, &rprocs, &rbs));
5875: PetscCall(PetscMPIIntCast(nsends + nrecvs, &nreqs));
5876: PetscCall(PetscMalloc1(nreqs, &reqs));
5877: rwaits = reqs;
5878: swaits = PetscSafePointerPlusOffset(reqs, nrecvs);
5880: if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX;
5881: if (scall == MAT_INITIAL_MATRIX) {
5882: /* i-array */
5883: /* post receives */
5884: if (nrecvs) PetscCall(PetscMalloc1(rbs * (rstarts[nrecvs] - rstarts[0]), &rvalues)); /* rstarts can be NULL when nrecvs=0 */
5885: for (i = 0; i < nrecvs; i++) {
5886: rowlen = rvalues + rstarts[i] * rbs;
5887: nrows = (rstarts[i + 1] - rstarts[i]) * rbs; /* num of indices to be received */
5888: PetscCallMPI(MPIU_Irecv(rowlen, nrows, MPIU_INT, rprocs[i], tag, comm, rwaits + i));
5889: }
5891: /* pack the outgoing message */
5892: PetscCall(PetscMalloc2(nsends + 1, &sstartsj, nrecvs + 1, &rstartsj));
5894: sstartsj[0] = 0;
5895: rstartsj[0] = 0;
5896: len = 0; /* total length of j or a array to be sent */
5897: if (nsends) {
5898: k = sstarts[0]; /* ATTENTION: sstarts[0] and rstarts[0] are not necessarily zero */
5899: PetscCall(PetscMalloc1(sbs * (sstarts[nsends] - sstarts[0]), &svalues));
5900: }
5901: for (i = 0; i < nsends; i++) {
5902: rowlen = svalues + (sstarts[i] - sstarts[0]) * sbs;
5903: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5904: for (j = 0; j < nrows; j++) {
5905: row = srow[k] + B->rmap->range[rank]; /* global row idx */
5906: for (l = 0; l < sbs; l++) {
5907: PetscCall(MatGetRow_MPIAIJ(B, row + l, &ncols, NULL, NULL)); /* rowlength */
5909: rowlen[j * sbs + l] = ncols;
5911: len += ncols;
5912: PetscCall(MatRestoreRow_MPIAIJ(B, row + l, &ncols, NULL, NULL));
5913: }
5914: k++;
5915: }
5916: PetscCallMPI(MPIU_Isend(rowlen, nrows * sbs, MPIU_INT, sprocs[i], tag, comm, swaits + i));
5918: sstartsj[i + 1] = len; /* starting point of (i+1)-th outgoing msg in bufj and bufa */
5919: }
5920: /* recvs and sends of i-array are completed */
5921: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5922: PetscCall(PetscFree(svalues));
5924: /* allocate buffers for sending j and a arrays */
5925: PetscCall(PetscMalloc1(len + 1, &bufj));
5926: PetscCall(PetscMalloc1(len + 1, &bufa));
5928: /* create i-array of B_oth */
5929: PetscCall(PetscMalloc1(aBn + 2, &b_othi));
5931: b_othi[0] = 0;
5932: len = 0; /* total length of j or a array to be received */
5933: k = 0;
5934: for (i = 0; i < nrecvs; i++) {
5935: rowlen = rvalues + (rstarts[i] - rstarts[0]) * rbs;
5936: nrows = (rstarts[i + 1] - rstarts[i]) * rbs; /* num of rows to be received */
5937: for (j = 0; j < nrows; j++) {
5938: b_othi[k + 1] = b_othi[k] + rowlen[j];
5939: PetscCall(PetscIntSumError(rowlen[j], len, &len));
5940: k++;
5941: }
5942: rstartsj[i + 1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */
5943: }
5944: PetscCall(PetscFree(rvalues));
5946: /* allocate space for j and a arrays of B_oth */
5947: PetscCall(PetscMalloc1(b_othi[aBn] + 1, &b_othj));
5948: PetscCall(PetscMalloc1(b_othi[aBn] + 1, &b_otha));
5950: /* j-array */
5951: /* post receives of j-array */
5952: for (i = 0; i < nrecvs; i++) {
5953: nrows = rstartsj[i + 1] - rstartsj[i]; /* length of the msg received */
5954: PetscCallMPI(MPIU_Irecv(b_othj + rstartsj[i], nrows, MPIU_INT, rprocs[i], tag, comm, rwaits + i));
5955: }
5957: /* pack the outgoing message j-array */
5958: if (nsends) k = sstarts[0];
5959: for (i = 0; i < nsends; i++) {
5960: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5961: bufJ = bufj + sstartsj[i];
5962: for (j = 0; j < nrows; j++) {
5963: row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5964: for (ll = 0; ll < sbs; ll++) {
5965: PetscCall(MatGetRow_MPIAIJ(B, row + ll, &ncols, &cols, NULL));
5966: for (l = 0; l < ncols; l++) *bufJ++ = cols[l];
5967: PetscCall(MatRestoreRow_MPIAIJ(B, row + ll, &ncols, &cols, NULL));
5968: }
5969: }
5970: PetscCallMPI(MPIU_Isend(bufj + sstartsj[i], sstartsj[i + 1] - sstartsj[i], MPIU_INT, sprocs[i], tag, comm, swaits + i));
5971: }
5973: /* recvs and sends of j-array are completed */
5974: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5975: } else if (scall == MAT_REUSE_MATRIX) {
5976: sstartsj = *startsj_s;
5977: rstartsj = *startsj_r;
5978: bufa = *bufa_ptr;
5979: PetscCall(MatSeqAIJGetArrayWrite(*B_oth, &b_otha));
5980: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not possess an object container");
5982: /* a-array */
5983: /* post receives of a-array */
5984: for (i = 0; i < nrecvs; i++) {
5985: nrows = rstartsj[i + 1] - rstartsj[i]; /* length of the msg received */
5986: PetscCallMPI(MPIU_Irecv(b_otha + rstartsj[i], nrows, MPIU_SCALAR, rprocs[i], tag, comm, rwaits + i));
5987: }
5989: /* pack the outgoing message a-array */
5990: if (nsends) k = sstarts[0];
5991: for (i = 0; i < nsends; i++) {
5992: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5993: bufA = bufa + sstartsj[i];
5994: for (j = 0; j < nrows; j++) {
5995: row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5996: for (ll = 0; ll < sbs; ll++) {
5997: PetscCall(MatGetRow_MPIAIJ(B, row + ll, &ncols, NULL, &vals));
5998: for (l = 0; l < ncols; l++) *bufA++ = vals[l];
5999: PetscCall(MatRestoreRow_MPIAIJ(B, row + ll, &ncols, NULL, &vals));
6000: }
6001: }
6002: PetscCallMPI(MPIU_Isend(bufa + sstartsj[i], sstartsj[i + 1] - sstartsj[i], MPIU_SCALAR, sprocs[i], tag, comm, swaits + i));
6003: }
6004: /* recvs and sends of a-array are completed */
6005: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
6006: PetscCall(PetscFree(reqs));
6008: if (scall == MAT_INITIAL_MATRIX) {
6009: Mat_SeqAIJ *b_oth;
6011: /* put together the new matrix */
6012: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, aBn, B->cmap->N, b_othi, b_othj, b_otha, B_oth));
6014: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
6015: /* Since these are PETSc arrays, change flags to free them as necessary. */
6016: b_oth = (Mat_SeqAIJ *)(*B_oth)->data;
6017: b_oth->free_a = PETSC_TRUE;
6018: b_oth->free_ij = PETSC_TRUE;
6019: b_oth->nonew = 0;
6021: PetscCall(PetscFree(bufj));
6022: if (!startsj_s || !bufa_ptr) {
6023: PetscCall(PetscFree2(sstartsj, rstartsj));
6024: PetscCall(PetscFree(bufa_ptr));
6025: } else {
6026: *startsj_s = sstartsj;
6027: *startsj_r = rstartsj;
6028: *bufa_ptr = bufa;
6029: }
6030: } else if (scall == MAT_REUSE_MATRIX) {
6031: PetscCall(MatSeqAIJRestoreArrayWrite(*B_oth, &b_otha));
6032: }
6034: PetscCall(VecScatterRestoreRemote_Private(ctx, PETSC_TRUE, &nsends, &sstarts, &srow, &sprocs, &sbs));
6035: PetscCall(VecScatterRestoreRemoteOrdered_Private(ctx, PETSC_FALSE, &nrecvs, &rstarts, NULL, &rprocs, &rbs));
6036: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAocols, A, B, 0, 0));
6037: PetscFunctionReturn(PETSC_SUCCESS);
6038: }
6040: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat, MatType, MatReuse, Mat *);
6041: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJPERM(Mat, MatType, MatReuse, Mat *);
6042: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJSELL(Mat, MatType, MatReuse, Mat *);
6043: #if defined(PETSC_HAVE_MKL_SPARSE)
6044: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJMKL(Mat, MatType, MatReuse, Mat *);
6045: #endif
6046: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIBAIJ(Mat, MatType, MatReuse, Mat *);
6047: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISBAIJ(Mat, MatType, MatReuse, Mat *);
6048: #if defined(PETSC_HAVE_ELEMENTAL)
6049: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat, MatType, MatReuse, Mat *);
6050: #endif
6051: #if defined(PETSC_HAVE_SCALAPACK)
6052: PETSC_INTERN PetscErrorCode MatConvert_AIJ_ScaLAPACK(Mat, MatType, MatReuse, Mat *);
6053: #endif
6054: #if defined(PETSC_HAVE_HYPRE)
6055: PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat, MatType, MatReuse, Mat *);
6056: #endif
6057: #if defined(PETSC_HAVE_CUDA)
6058: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCUSPARSE(Mat, MatType, MatReuse, Mat *);
6059: #endif
6060: #if defined(PETSC_HAVE_HIP)
6061: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJHIPSPARSE(Mat, MatType, MatReuse, Mat *);
6062: #endif
6063: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
6064: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJKokkos(Mat, MatType, MatReuse, Mat *);
6065: #endif
6066: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISELL(Mat, MatType, MatReuse, Mat *);
6067: PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat, MatType, MatReuse, Mat *);
6068: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_IS_XAIJ(Mat);
6070: /*
6071: Computes (B'*A')' since computing B*A directly is untenable
6073: n p p
6074: [ ] [ ] [ ]
6075: m [ A ] * n [ B ] = m [ C ]
6076: [ ] [ ] [ ]
6078: */
6079: static PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A, Mat B, Mat C)
6080: {
6081: Mat At, Bt, Ct;
6083: PetscFunctionBegin;
6084: PetscCall(MatTranspose(A, MAT_INITIAL_MATRIX, &At));
6085: PetscCall(MatTranspose(B, MAT_INITIAL_MATRIX, &Bt));
6086: PetscCall(MatMatMult(Bt, At, MAT_INITIAL_MATRIX, PETSC_CURRENT, &Ct));
6087: PetscCall(MatDestroy(&At));
6088: PetscCall(MatDestroy(&Bt));
6089: PetscCall(MatTransposeSetPrecursor(Ct, C));
6090: PetscCall(MatTranspose(Ct, MAT_REUSE_MATRIX, &C));
6091: PetscCall(MatDestroy(&Ct));
6092: PetscFunctionReturn(PETSC_SUCCESS);
6093: }
6095: static PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A, Mat B, PetscReal fill, Mat C)
6096: {
6097: PetscBool cisdense;
6099: PetscFunctionBegin;
6100: 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);
6101: PetscCall(MatSetSizes(C, A->rmap->n, B->cmap->n, A->rmap->N, B->cmap->N));
6102: PetscCall(MatSetBlockSizesFromMats(C, A, B));
6103: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATMPIDENSE, MATMPIDENSECUDA, MATMPIDENSEHIP, ""));
6104: if (!cisdense) PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
6105: PetscCall(MatSetUp(C));
6107: C->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIAIJ;
6108: PetscFunctionReturn(PETSC_SUCCESS);
6109: }
6111: static PetscErrorCode MatProductSetFromOptions_MPIDense_MPIAIJ_AB(Mat C)
6112: {
6113: Mat_Product *product = C->product;
6114: Mat A = product->A, B = product->B;
6116: PetscFunctionBegin;
6117: 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 ")",
6118: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
6119: C->ops->matmultsymbolic = MatMatMultSymbolic_MPIDense_MPIAIJ;
6120: C->ops->productsymbolic = MatProductSymbolic_AB;
6121: PetscFunctionReturn(PETSC_SUCCESS);
6122: }
6124: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_MPIDense_MPIAIJ(Mat C)
6125: {
6126: Mat_Product *product = C->product;
6128: PetscFunctionBegin;
6129: if (product->type == MATPRODUCT_AB) PetscCall(MatProductSetFromOptions_MPIDense_MPIAIJ_AB(C));
6130: PetscFunctionReturn(PETSC_SUCCESS);
6131: }
6133: /*
6134: Merge two sets of sorted nonzeros and return a CSR for the merged (sequential) matrix
6136: Input Parameters:
6138: j1,rowBegin1,rowEnd1,jmap1: describe the first set of nonzeros (Set1)
6139: j2,rowBegin2,rowEnd2,jmap2: describe the second set of nonzeros (Set2)
6141: mat: both sets' nonzeros are on m rows, where m is the number of local rows of the matrix mat
6143: For Set1, j1[] contains column indices of the nonzeros.
6144: 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
6145: respectively (note rowEnd1[k] is not necessarily equal to rwoBegin1[k+1]). Indices in this range of j1[] are sorted,
6146: but might have repeats. jmap1[t+1] - jmap1[t] is the number of repeats for the t-th unique nonzero in Set1.
6148: Similar for Set2.
6150: This routine merges the two sets of nonzeros row by row and removes repeats.
6152: Output Parameters: (memory is allocated by the caller)
6154: i[],j[]: the CSR of the merged matrix, which has m rows.
6155: imap1[]: the k-th unique nonzero in Set1 (k=0,1,...) corresponds to imap1[k]-th unique nonzero in the merged matrix.
6156: imap2[]: similar to imap1[], but for Set2.
6157: Note we order nonzeros row-by-row and from left to right.
6158: */
6159: 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[])
6160: {
6161: PetscInt r, m; /* Row index of mat */
6162: PetscCount t, t1, t2, b1, e1, b2, e2;
6164: PetscFunctionBegin;
6165: PetscCall(MatGetLocalSize(mat, &m, NULL));
6166: t1 = t2 = t = 0; /* Count unique nonzeros of in Set1, Set1 and the merged respectively */
6167: i[0] = 0;
6168: for (r = 0; r < m; r++) { /* Do row by row merging */
6169: b1 = rowBegin1[r];
6170: e1 = rowEnd1[r];
6171: b2 = rowBegin2[r];
6172: e2 = rowEnd2[r];
6173: while (b1 < e1 && b2 < e2) {
6174: if (j1[b1] == j2[b2]) { /* Same column index and hence same nonzero */
6175: j[t] = j1[b1];
6176: imap1[t1] = t;
6177: imap2[t2] = t;
6178: b1 += jmap1[t1 + 1] - jmap1[t1]; /* Jump to next unique local nonzero */
6179: b2 += jmap2[t2 + 1] - jmap2[t2]; /* Jump to next unique remote nonzero */
6180: t1++;
6181: t2++;
6182: t++;
6183: } else if (j1[b1] < j2[b2]) {
6184: j[t] = j1[b1];
6185: imap1[t1] = t;
6186: b1 += jmap1[t1 + 1] - jmap1[t1];
6187: t1++;
6188: t++;
6189: } else {
6190: j[t] = j2[b2];
6191: imap2[t2] = t;
6192: b2 += jmap2[t2 + 1] - jmap2[t2];
6193: t2++;
6194: t++;
6195: }
6196: }
6197: /* Merge the remaining in either j1[] or j2[] */
6198: while (b1 < e1) {
6199: j[t] = j1[b1];
6200: imap1[t1] = t;
6201: b1 += jmap1[t1 + 1] - jmap1[t1];
6202: t1++;
6203: t++;
6204: }
6205: while (b2 < e2) {
6206: j[t] = j2[b2];
6207: imap2[t2] = t;
6208: b2 += jmap2[t2 + 1] - jmap2[t2];
6209: t2++;
6210: t++;
6211: }
6212: PetscCall(PetscIntCast(t, i + r + 1));
6213: }
6214: PetscFunctionReturn(PETSC_SUCCESS);
6215: }
6217: /*
6218: Split nonzeros in a block of local rows into two subsets: those in the diagonal block and those in the off-diagonal block
6220: Input Parameters:
6221: mat: an MPI matrix that provides row and column layout information for splitting. Let's say its number of local rows is m.
6222: 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[]
6223: respectively, along with a permutation array perm[]. Length of the i[],j[],perm[] arrays is n.
6225: i[] is already sorted, but within a row, j[] is not sorted and might have repeats.
6226: i[] might contain negative indices at the beginning, which means the corresponding entries should be ignored in the splitting.
6228: Output Parameters:
6229: j[],perm[]: the routine needs to sort j[] within each row along with perm[].
6230: rowBegin[],rowMid[],rowEnd[]: of length m, and the memory is preallocated and zeroed by the caller.
6231: 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,
6232: and [rowMid[r],rowEnd[r]) point to begin/end entries of row r of the off-diagonal block.
6234: Aperm[],Ajmap[],Atot,Annz: Arrays are allocated by this routine.
6235: Atot: number of entries belonging to the diagonal block.
6236: Annz: number of unique nonzeros belonging to the diagonal block.
6237: Aperm[Atot] stores values from perm[] for entries belonging to the diagonal block. Length of Aperm[] is Atot, though it may also count
6238: repeats (i.e., same 'i,j' pair).
6239: Ajmap[Annz+1] stores the number of repeats of each unique entry belonging to the diagonal block. More precisely, Ajmap[t+1] - Ajmap[t]
6240: is the number of repeats for the t-th unique entry in the diagonal block. Ajmap[0] is always 0.
6242: Atot: number of entries belonging to the diagonal block
6243: Annz: number of unique nonzeros belonging to the diagonal block.
6245: Bperm[], Bjmap[], Btot, Bnnz are similar but for the off-diagonal block.
6247: Aperm[],Bperm[],Ajmap[] and Bjmap[] are allocated separately by this routine with PetscMalloc1().
6248: */
6249: 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_)
6250: {
6251: PetscInt cstart, cend, rstart, rend, row, col;
6252: PetscCount Atot = 0, Btot = 0; /* Total number of nonzeros in the diagonal and off-diagonal blocks */
6253: PetscCount Annz = 0, Bnnz = 0; /* Number of unique nonzeros in the diagonal and off-diagonal blocks */
6254: PetscCount k, m, p, q, r, s, mid;
6255: PetscCount *Aperm, *Bperm, *Ajmap, *Bjmap;
6257: PetscFunctionBegin;
6258: PetscCall(PetscLayoutGetRange(mat->rmap, &rstart, &rend));
6259: PetscCall(PetscLayoutGetRange(mat->cmap, &cstart, &cend));
6260: m = rend - rstart;
6262: /* Skip negative rows */
6263: for (k = 0; k < n; k++)
6264: if (i[k] >= 0) break;
6266: /* Process [k,n): sort and partition each local row into diag and offdiag portions,
6267: fill rowBegin[], rowMid[], rowEnd[], and count Atot, Btot, Annz, Bnnz.
6268: */
6269: while (k < n) {
6270: row = i[k];
6271: /* Entries in [k,s) are in one row. Shift diagonal block col indices so that diag is ahead of offdiag after sorting the row */
6272: for (s = k; s < n; s++)
6273: if (i[s] != row) break;
6275: /* Shift diag columns to range of [-PETSC_INT_MAX, -1] */
6276: for (p = k; p < s; p++) {
6277: if (j[p] >= cstart && j[p] < cend) j[p] -= PETSC_INT_MAX;
6278: 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]);
6279: }
6280: PetscCall(PetscSortIntWithCountArray(s - k, j + k, perm + k));
6281: PetscCall(PetscSortedIntUpperBound(j, k, s, -1, &mid)); /* Separate [k,s) into [k,mid) for diag and [mid,s) for offdiag */
6282: rowBegin[row - rstart] = k;
6283: rowMid[row - rstart] = mid;
6284: rowEnd[row - rstart] = s;
6286: /* Count nonzeros of this diag/offdiag row, which might have repeats */
6287: Atot += mid - k;
6288: Btot += s - mid;
6290: /* Count unique nonzeros of this diag row */
6291: for (p = k; p < mid;) {
6292: col = j[p];
6293: do {
6294: j[p] += PETSC_INT_MAX; /* Revert the modified diagonal indices */
6295: p++;
6296: } while (p < mid && j[p] == col);
6297: Annz++;
6298: }
6300: /* Count unique nonzeros of this offdiag row */
6301: for (p = mid; p < s;) {
6302: col = j[p];
6303: do {
6304: p++;
6305: } while (p < s && j[p] == col);
6306: Bnnz++;
6307: }
6308: k = s;
6309: }
6311: /* Allocation according to Atot, Btot, Annz, Bnnz */
6312: PetscCall(PetscMalloc1(Atot, &Aperm));
6313: PetscCall(PetscMalloc1(Btot, &Bperm));
6314: PetscCall(PetscMalloc1(Annz + 1, &Ajmap));
6315: PetscCall(PetscMalloc1(Bnnz + 1, &Bjmap));
6317: /* Re-scan indices and copy diag/offdiag permutation indices to Aperm, Bperm and also fill Ajmap and Bjmap */
6318: Ajmap[0] = Bjmap[0] = Atot = Btot = Annz = Bnnz = 0;
6319: for (r = 0; r < m; r++) {
6320: k = rowBegin[r];
6321: mid = rowMid[r];
6322: s = rowEnd[r];
6323: PetscCall(PetscArraycpy(PetscSafePointerPlusOffset(Aperm, Atot), PetscSafePointerPlusOffset(perm, k), mid - k));
6324: PetscCall(PetscArraycpy(PetscSafePointerPlusOffset(Bperm, Btot), PetscSafePointerPlusOffset(perm, mid), s - mid));
6325: Atot += mid - k;
6326: Btot += s - mid;
6328: /* Scan column indices in this row and find out how many repeats each unique nonzero has */
6329: for (p = k; p < mid;) {
6330: col = j[p];
6331: q = p;
6332: do {
6333: p++;
6334: } while (p < mid && j[p] == col);
6335: Ajmap[Annz + 1] = Ajmap[Annz] + (p - q);
6336: Annz++;
6337: }
6339: for (p = mid; p < s;) {
6340: col = j[p];
6341: q = p;
6342: do {
6343: p++;
6344: } while (p < s && j[p] == col);
6345: Bjmap[Bnnz + 1] = Bjmap[Bnnz] + (p - q);
6346: Bnnz++;
6347: }
6348: }
6349: /* Output */
6350: *Aperm_ = Aperm;
6351: *Annz_ = Annz;
6352: *Atot_ = Atot;
6353: *Ajmap_ = Ajmap;
6354: *Bperm_ = Bperm;
6355: *Bnnz_ = Bnnz;
6356: *Btot_ = Btot;
6357: *Bjmap_ = Bjmap;
6358: PetscFunctionReturn(PETSC_SUCCESS);
6359: }
6361: /*
6362: Expand the jmap[] array to make a new one in view of nonzeros in the merged matrix
6364: Input Parameters:
6365: nnz1: number of unique nonzeros in a set that was used to produce imap[], jmap[]
6366: nnz: number of unique nonzeros in the merged matrix
6367: imap[nnz1]: i-th nonzero in the set is the imap[i]-th nonzero in the merged matrix
6368: jmap[nnz1+1]: i-th nonzero in the set has jmap[i+1] - jmap[i] repeats in the set
6370: Output Parameter: (memory is allocated by the caller)
6371: jmap_new[nnz+1]: i-th nonzero in the merged matrix has jmap_new[i+1] - jmap_new[i] repeats in the set
6373: Example:
6374: nnz1 = 4
6375: nnz = 6
6376: imap = [1,3,4,5]
6377: jmap = [0,3,5,6,7]
6378: then,
6379: jmap_new = [0,0,3,3,5,6,7]
6380: */
6381: static PetscErrorCode ExpandJmap_Internal(PetscCount nnz1, PetscCount nnz, const PetscCount imap[], const PetscCount jmap[], PetscCount jmap_new[])
6382: {
6383: PetscCount k, p;
6385: PetscFunctionBegin;
6386: jmap_new[0] = 0;
6387: p = nnz; /* p loops over jmap_new[] backwards */
6388: for (k = nnz1 - 1; k >= 0; k--) { /* k loops over imap[] */
6389: for (; p > imap[k]; p--) jmap_new[p] = jmap[k + 1];
6390: }
6391: for (; p >= 0; p--) jmap_new[p] = jmap[0];
6392: PetscFunctionReturn(PETSC_SUCCESS);
6393: }
6395: static PetscErrorCode MatCOOStructDestroy_MPIAIJ(void *data)
6396: {
6397: MatCOOStruct_MPIAIJ *coo = (MatCOOStruct_MPIAIJ *)data;
6399: PetscFunctionBegin;
6400: PetscCall(PetscSFDestroy(&coo->sf));
6401: PetscCall(PetscFree(coo->Aperm1));
6402: PetscCall(PetscFree(coo->Bperm1));
6403: PetscCall(PetscFree(coo->Ajmap1));
6404: PetscCall(PetscFree(coo->Bjmap1));
6405: PetscCall(PetscFree(coo->Aimap2));
6406: PetscCall(PetscFree(coo->Bimap2));
6407: PetscCall(PetscFree(coo->Aperm2));
6408: PetscCall(PetscFree(coo->Bperm2));
6409: PetscCall(PetscFree(coo->Ajmap2));
6410: PetscCall(PetscFree(coo->Bjmap2));
6411: PetscCall(PetscFree(coo->Cperm1));
6412: PetscCall(PetscFree2(coo->sendbuf, coo->recvbuf));
6413: PetscCall(PetscFree(coo));
6414: PetscFunctionReturn(PETSC_SUCCESS);
6415: }
6417: PetscErrorCode MatSetPreallocationCOO_MPIAIJ(Mat mat, PetscCount coo_n, PetscInt coo_i[], PetscInt coo_j[])
6418: {
6419: MPI_Comm comm;
6420: PetscMPIInt rank, size;
6421: PetscInt m, n, M, N, rstart, rend, cstart, cend; /* Sizes, indices of row/col, therefore with type PetscInt */
6422: PetscCount k, p, q, rem; /* Loop variables over coo arrays */
6423: Mat_MPIAIJ *mpiaij = (Mat_MPIAIJ *)mat->data;
6424: PetscContainer container;
6425: MatCOOStruct_MPIAIJ *coo;
6427: PetscFunctionBegin;
6428: PetscCall(PetscFree(mpiaij->garray));
6429: PetscCall(VecDestroy(&mpiaij->lvec));
6430: #if defined(PETSC_USE_CTABLE)
6431: PetscCall(PetscHMapIDestroy(&mpiaij->colmap));
6432: #else
6433: PetscCall(PetscFree(mpiaij->colmap));
6434: #endif
6435: PetscCall(VecScatterDestroy(&mpiaij->Mvctx));
6436: mat->assembled = PETSC_FALSE;
6437: mat->was_assembled = PETSC_FALSE;
6439: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
6440: PetscCallMPI(MPI_Comm_size(comm, &size));
6441: PetscCallMPI(MPI_Comm_rank(comm, &rank));
6442: PetscCall(PetscLayoutSetUp(mat->rmap));
6443: PetscCall(PetscLayoutSetUp(mat->cmap));
6444: PetscCall(PetscLayoutGetRange(mat->rmap, &rstart, &rend));
6445: PetscCall(PetscLayoutGetRange(mat->cmap, &cstart, &cend));
6446: PetscCall(MatGetLocalSize(mat, &m, &n));
6447: PetscCall(MatGetSize(mat, &M, &N));
6449: /* Sort (i,j) by row along with a permutation array, so that the to-be-ignored */
6450: /* entries come first, then local rows, then remote rows. */
6451: PetscCount n1 = coo_n, *perm1;
6452: PetscInt *i1 = coo_i, *j1 = coo_j;
6454: PetscCall(PetscMalloc1(n1, &perm1));
6455: for (k = 0; k < n1; k++) perm1[k] = k;
6457: /* Manipulate indices so that entries with negative row or col indices will have smallest
6458: row indices, local entries will have greater but negative row indices, and remote entries
6459: will have positive row indices.
6460: */
6461: for (k = 0; k < n1; k++) {
6462: if (i1[k] < 0 || j1[k] < 0) i1[k] = PETSC_INT_MIN; /* e.g., -2^31, minimal to move them ahead */
6463: 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] */
6464: else {
6465: PetscCheck(!mat->nooffprocentries, PETSC_COMM_SELF, PETSC_ERR_USER_INPUT, "MAT_NO_OFF_PROC_ENTRIES is set but insert to remote rows");
6466: if (mpiaij->donotstash) i1[k] = PETSC_INT_MIN; /* Ignore offproc entries as if they had negative indices */
6467: }
6468: }
6470: /* Sort by row; after that, [0,k) have ignored entries, [k,rem) have local rows and [rem,n1) have remote rows */
6471: PetscCall(PetscSortIntWithIntCountArrayPair(n1, i1, j1, perm1));
6473: /* Advance k to the first entry we need to take care of */
6474: for (k = 0; k < n1; k++)
6475: if (i1[k] > PETSC_INT_MIN) break;
6476: PetscCount i1start = k;
6478: PetscCall(PetscSortedIntUpperBound(i1, k, n1, rend - 1 - PETSC_INT_MAX, &rem)); /* rem is upper bound of the last local row */
6479: for (; k < rem; k++) i1[k] += PETSC_INT_MAX; /* Revert row indices of local rows*/
6481: /* Send remote rows to their owner */
6482: /* Find which rows should be sent to which remote ranks*/
6483: PetscInt nsend = 0; /* Number of MPI ranks to send data to */
6484: PetscMPIInt *sendto; /* [nsend], storing remote ranks */
6485: PetscInt *nentries; /* [nsend], storing number of entries sent to remote ranks; Assume PetscInt is big enough for this count, and error if not */
6486: const PetscInt *ranges;
6487: PetscInt maxNsend = size >= 128 ? 128 : size; /* Assume max 128 neighbors; realloc when needed */
6489: PetscCall(PetscLayoutGetRanges(mat->rmap, &ranges));
6490: PetscCall(PetscMalloc2(maxNsend, &sendto, maxNsend, &nentries));
6491: for (k = rem; k < n1;) {
6492: PetscMPIInt owner;
6493: PetscInt firstRow, lastRow;
6495: /* Locate a row range */
6496: firstRow = i1[k]; /* first row of this owner */
6497: PetscCall(PetscLayoutFindOwner(mat->rmap, firstRow, &owner));
6498: lastRow = ranges[owner + 1] - 1; /* last row of this owner */
6500: /* Find the first index 'p' in [k,n) with i[p] belonging to next owner */
6501: PetscCall(PetscSortedIntUpperBound(i1, k, n1, lastRow, &p));
6503: /* All entries in [k,p) belong to this remote owner */
6504: if (nsend >= maxNsend) { /* Double the remote ranks arrays if not long enough */
6505: PetscMPIInt *sendto2;
6506: PetscInt *nentries2;
6507: PetscInt maxNsend2 = (maxNsend <= size / 2) ? maxNsend * 2 : size;
6509: PetscCall(PetscMalloc2(maxNsend2, &sendto2, maxNsend2, &nentries2));
6510: PetscCall(PetscArraycpy(sendto2, sendto, maxNsend));
6511: PetscCall(PetscArraycpy(nentries2, nentries2, maxNsend + 1));
6512: PetscCall(PetscFree2(sendto, nentries2));
6513: sendto = sendto2;
6514: nentries = nentries2;
6515: maxNsend = maxNsend2;
6516: }
6517: sendto[nsend] = owner;
6518: PetscCall(PetscIntCast(p - k, &nentries[nsend]));
6519: nsend++;
6520: k = p;
6521: }
6523: /* Build 1st SF to know offsets on remote to send data */
6524: PetscSF sf1;
6525: PetscInt nroots = 1, nroots2 = 0;
6526: PetscInt nleaves = nsend, nleaves2 = 0;
6527: PetscInt *offsets;
6528: PetscSFNode *iremote;
6530: PetscCall(PetscSFCreate(comm, &sf1));
6531: PetscCall(PetscMalloc1(nsend, &iremote));
6532: PetscCall(PetscMalloc1(nsend, &offsets));
6533: for (k = 0; k < nsend; k++) {
6534: iremote[k].rank = sendto[k];
6535: iremote[k].index = 0;
6536: nleaves2 += nentries[k];
6537: PetscCheck(nleaves2 >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Number of SF leaves is too large for PetscInt");
6538: }
6539: PetscCall(PetscSFSetGraph(sf1, nroots, nleaves, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
6540: PetscCall(PetscSFFetchAndOpWithMemTypeBegin(sf1, MPIU_INT, PETSC_MEMTYPE_HOST, &nroots2 /*rootdata*/, PETSC_MEMTYPE_HOST, nentries /*leafdata*/, PETSC_MEMTYPE_HOST, offsets /*leafupdate*/, MPI_SUM));
6541: PetscCall(PetscSFFetchAndOpEnd(sf1, MPIU_INT, &nroots2, nentries, offsets, MPI_SUM)); /* Would nroots2 overflow, we check offsets[] below */
6542: PetscCall(PetscSFDestroy(&sf1));
6543: PetscAssert(nleaves2 == n1 - rem, PETSC_COMM_SELF, PETSC_ERR_PLIB, "nleaves2 %" PetscInt_FMT " != number of remote entries %" PetscCount_FMT, nleaves2, n1 - rem);
6545: /* Build 2nd SF to send remote COOs to their owner */
6546: PetscSF sf2;
6547: nroots = nroots2;
6548: nleaves = nleaves2;
6549: PetscCall(PetscSFCreate(comm, &sf2));
6550: PetscCall(PetscSFSetFromOptions(sf2));
6551: PetscCall(PetscMalloc1(nleaves, &iremote));
6552: p = 0;
6553: for (k = 0; k < nsend; k++) {
6554: PetscCheck(offsets[k] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Number of SF roots is too large for PetscInt");
6555: for (q = 0; q < nentries[k]; q++, p++) {
6556: iremote[p].rank = sendto[k];
6557: PetscCall(PetscIntCast(offsets[k] + q, &iremote[p].index));
6558: }
6559: }
6560: PetscCall(PetscSFSetGraph(sf2, nroots, nleaves, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
6562: /* Send the remote COOs to their owner */
6563: PetscInt n2 = nroots, *i2, *j2; /* Buffers for received COOs from other ranks, along with a permutation array */
6564: PetscCount *perm2; /* Though PetscInt is enough for remote entries, we use PetscCount here as we want to reuse MatSplitEntries_Internal() */
6565: PetscCall(PetscMalloc3(n2, &i2, n2, &j2, n2, &perm2));
6566: PetscAssert(rem == 0 || i1 != NULL, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot add nonzero offset to null");
6567: PetscAssert(rem == 0 || j1 != NULL, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot add nonzero offset to null");
6568: PetscInt *i1prem = PetscSafePointerPlusOffset(i1, rem);
6569: PetscInt *j1prem = PetscSafePointerPlusOffset(j1, rem);
6570: PetscCall(PetscSFReduceWithMemTypeBegin(sf2, MPIU_INT, PETSC_MEMTYPE_HOST, i1prem, PETSC_MEMTYPE_HOST, i2, MPI_REPLACE));
6571: PetscCall(PetscSFReduceEnd(sf2, MPIU_INT, i1prem, i2, MPI_REPLACE));
6572: PetscCall(PetscSFReduceWithMemTypeBegin(sf2, MPIU_INT, PETSC_MEMTYPE_HOST, j1prem, PETSC_MEMTYPE_HOST, j2, MPI_REPLACE));
6573: PetscCall(PetscSFReduceEnd(sf2, MPIU_INT, j1prem, j2, MPI_REPLACE));
6575: PetscCall(PetscFree(offsets));
6576: PetscCall(PetscFree2(sendto, nentries));
6578: /* Sort received COOs by row along with the permutation array */
6579: for (k = 0; k < n2; k++) perm2[k] = k;
6580: PetscCall(PetscSortIntWithIntCountArrayPair(n2, i2, j2, perm2));
6582: /* sf2 only sends contiguous leafdata to contiguous rootdata. We record the permutation which will be used to fill leafdata */
6583: PetscCount *Cperm1;
6584: PetscAssert(rem == 0 || perm1 != NULL, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot add nonzero offset to null");
6585: PetscCount *perm1prem = PetscSafePointerPlusOffset(perm1, rem);
6586: PetscCall(PetscMalloc1(nleaves, &Cperm1));
6587: PetscCall(PetscArraycpy(Cperm1, perm1prem, nleaves));
6589: /* Support for HYPRE matrices, kind of a hack.
6590: Swap min column with diagonal so that diagonal values will go first */
6591: PetscBool hypre;
6592: PetscCall(PetscStrcmp("_internal_COO_mat_for_hypre", ((PetscObject)mat)->name, &hypre));
6593: if (hypre) {
6594: PetscInt *minj;
6595: PetscBT hasdiag;
6597: PetscCall(PetscBTCreate(m, &hasdiag));
6598: PetscCall(PetscMalloc1(m, &minj));
6599: for (k = 0; k < m; k++) minj[k] = PETSC_INT_MAX;
6600: for (k = i1start; k < rem; k++) {
6601: if (j1[k] < cstart || j1[k] >= cend) continue;
6602: const PetscInt rindex = i1[k] - rstart;
6603: if ((j1[k] - cstart) == rindex) PetscCall(PetscBTSet(hasdiag, rindex));
6604: minj[rindex] = PetscMin(minj[rindex], j1[k]);
6605: }
6606: for (k = 0; k < n2; k++) {
6607: if (j2[k] < cstart || j2[k] >= cend) continue;
6608: const PetscInt rindex = i2[k] - rstart;
6609: if ((j2[k] - cstart) == rindex) PetscCall(PetscBTSet(hasdiag, rindex));
6610: minj[rindex] = PetscMin(minj[rindex], j2[k]);
6611: }
6612: for (k = i1start; k < rem; k++) {
6613: const PetscInt rindex = i1[k] - rstart;
6614: if (j1[k] < cstart || j1[k] >= cend || !PetscBTLookup(hasdiag, rindex)) continue;
6615: if (j1[k] == minj[rindex]) j1[k] = i1[k] + (cstart - rstart);
6616: else if ((j1[k] - cstart) == rindex) j1[k] = minj[rindex];
6617: }
6618: for (k = 0; k < n2; k++) {
6619: const PetscInt rindex = i2[k] - rstart;
6620: if (j2[k] < cstart || j2[k] >= cend || !PetscBTLookup(hasdiag, rindex)) continue;
6621: if (j2[k] == minj[rindex]) j2[k] = i2[k] + (cstart - rstart);
6622: else if ((j2[k] - cstart) == rindex) j2[k] = minj[rindex];
6623: }
6624: PetscCall(PetscBTDestroy(&hasdiag));
6625: PetscCall(PetscFree(minj));
6626: }
6628: /* Split local COOs and received COOs into diag/offdiag portions */
6629: PetscCount *rowBegin1, *rowMid1, *rowEnd1;
6630: PetscCount *Ajmap1, *Aperm1, *Bjmap1, *Bperm1;
6631: PetscCount Annz1, Bnnz1, Atot1, Btot1;
6632: PetscCount *rowBegin2, *rowMid2, *rowEnd2;
6633: PetscCount *Ajmap2, *Aperm2, *Bjmap2, *Bperm2;
6634: PetscCount Annz2, Bnnz2, Atot2, Btot2;
6636: PetscCall(PetscCalloc3(m, &rowBegin1, m, &rowMid1, m, &rowEnd1));
6637: PetscCall(PetscCalloc3(m, &rowBegin2, m, &rowMid2, m, &rowEnd2));
6638: PetscCall(MatSplitEntries_Internal(mat, rem, i1, j1, perm1, rowBegin1, rowMid1, rowEnd1, &Atot1, &Aperm1, &Annz1, &Ajmap1, &Btot1, &Bperm1, &Bnnz1, &Bjmap1));
6639: PetscCall(MatSplitEntries_Internal(mat, n2, i2, j2, perm2, rowBegin2, rowMid2, rowEnd2, &Atot2, &Aperm2, &Annz2, &Ajmap2, &Btot2, &Bperm2, &Bnnz2, &Bjmap2));
6641: /* Merge local COOs with received COOs: diag with diag, offdiag with offdiag */
6642: PetscInt *Ai, *Bi;
6643: PetscInt *Aj, *Bj;
6645: PetscCall(PetscMalloc1(m + 1, &Ai));
6646: PetscCall(PetscMalloc1(m + 1, &Bi));
6647: PetscCall(PetscMalloc1(Annz1 + Annz2, &Aj)); /* Since local and remote entries might have dups, we might allocate excess memory */
6648: PetscCall(PetscMalloc1(Bnnz1 + Bnnz2, &Bj));
6650: PetscCount *Aimap1, *Bimap1, *Aimap2, *Bimap2;
6651: PetscCall(PetscMalloc1(Annz1, &Aimap1));
6652: PetscCall(PetscMalloc1(Bnnz1, &Bimap1));
6653: PetscCall(PetscMalloc1(Annz2, &Aimap2));
6654: PetscCall(PetscMalloc1(Bnnz2, &Bimap2));
6656: PetscCall(MatMergeEntries_Internal(mat, j1, j2, rowBegin1, rowMid1, rowBegin2, rowMid2, Ajmap1, Ajmap2, Aimap1, Aimap2, Ai, Aj));
6657: PetscCall(MatMergeEntries_Internal(mat, j1, j2, rowMid1, rowEnd1, rowMid2, rowEnd2, Bjmap1, Bjmap2, Bimap1, Bimap2, Bi, Bj));
6659: /* Expand Ajmap1/Bjmap1 to make them based off nonzeros in A/B, since we */
6660: /* expect nonzeros in A/B most likely have local contributing entries */
6661: PetscInt Annz = Ai[m];
6662: PetscInt Bnnz = Bi[m];
6663: PetscCount *Ajmap1_new, *Bjmap1_new;
6665: PetscCall(PetscMalloc1(Annz + 1, &Ajmap1_new));
6666: PetscCall(PetscMalloc1(Bnnz + 1, &Bjmap1_new));
6668: PetscCall(ExpandJmap_Internal(Annz1, Annz, Aimap1, Ajmap1, Ajmap1_new));
6669: PetscCall(ExpandJmap_Internal(Bnnz1, Bnnz, Bimap1, Bjmap1, Bjmap1_new));
6671: PetscCall(PetscFree(Aimap1));
6672: PetscCall(PetscFree(Ajmap1));
6673: PetscCall(PetscFree(Bimap1));
6674: PetscCall(PetscFree(Bjmap1));
6675: PetscCall(PetscFree3(rowBegin1, rowMid1, rowEnd1));
6676: PetscCall(PetscFree3(rowBegin2, rowMid2, rowEnd2));
6677: PetscCall(PetscFree(perm1));
6678: PetscCall(PetscFree3(i2, j2, perm2));
6680: Ajmap1 = Ajmap1_new;
6681: Bjmap1 = Bjmap1_new;
6683: /* Reallocate Aj, Bj once we know actual numbers of unique nonzeros in A and B */
6684: if (Annz < Annz1 + Annz2) {
6685: PetscInt *Aj_new;
6686: PetscCall(PetscMalloc1(Annz, &Aj_new));
6687: PetscCall(PetscArraycpy(Aj_new, Aj, Annz));
6688: PetscCall(PetscFree(Aj));
6689: Aj = Aj_new;
6690: }
6692: if (Bnnz < Bnnz1 + Bnnz2) {
6693: PetscInt *Bj_new;
6694: PetscCall(PetscMalloc1(Bnnz, &Bj_new));
6695: PetscCall(PetscArraycpy(Bj_new, Bj, Bnnz));
6696: PetscCall(PetscFree(Bj));
6697: Bj = Bj_new;
6698: }
6700: /* Create new submatrices for on-process and off-process coupling */
6701: PetscScalar *Aa, *Ba;
6702: MatType rtype;
6703: Mat_SeqAIJ *a, *b;
6704: PetscObjectState state;
6705: PetscCall(PetscCalloc1(Annz, &Aa)); /* Zero matrix on device */
6706: PetscCall(PetscCalloc1(Bnnz, &Ba));
6707: /* make Aj[] local, i.e, based off the start column of the diagonal portion */
6708: if (cstart) {
6709: for (k = 0; k < Annz; k++) Aj[k] -= cstart;
6710: }
6712: PetscCall(MatGetRootType_Private(mat, &rtype));
6714: MatSeqXAIJGetOptions_Private(mpiaij->A);
6715: PetscCall(MatDestroy(&mpiaij->A));
6716: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, n, Ai, Aj, Aa, &mpiaij->A));
6717: PetscCall(MatSetBlockSizesFromMats(mpiaij->A, mat, mat));
6718: MatSeqXAIJRestoreOptions_Private(mpiaij->A);
6720: MatSeqXAIJGetOptions_Private(mpiaij->B);
6721: PetscCall(MatDestroy(&mpiaij->B));
6722: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, mat->cmap->N, Bi, Bj, Ba, &mpiaij->B));
6723: PetscCall(MatSetBlockSizesFromMats(mpiaij->B, mat, mat));
6724: MatSeqXAIJRestoreOptions_Private(mpiaij->B);
6726: PetscCall(MatSetUpMultiply_MPIAIJ(mat));
6727: mat->was_assembled = PETSC_TRUE; // was_assembled in effect means the Mvctx is built; doing so avoids redundant MatSetUpMultiply_MPIAIJ
6728: state = mpiaij->A->nonzerostate + mpiaij->B->nonzerostate;
6729: PetscCallMPI(MPIU_Allreduce(&state, &mat->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)mat)));
6731: a = (Mat_SeqAIJ *)mpiaij->A->data;
6732: b = (Mat_SeqAIJ *)mpiaij->B->data;
6733: a->free_a = PETSC_TRUE;
6734: a->free_ij = PETSC_TRUE;
6735: b->free_a = PETSC_TRUE;
6736: b->free_ij = PETSC_TRUE;
6737: a->maxnz = a->nz;
6738: b->maxnz = b->nz;
6740: /* conversion must happen AFTER multiply setup */
6741: PetscCall(MatConvert(mpiaij->A, rtype, MAT_INPLACE_MATRIX, &mpiaij->A));
6742: PetscCall(MatConvert(mpiaij->B, rtype, MAT_INPLACE_MATRIX, &mpiaij->B));
6743: PetscCall(VecDestroy(&mpiaij->lvec));
6744: PetscCall(MatCreateVecs(mpiaij->B, &mpiaij->lvec, NULL));
6746: // Put the COO struct in a container and then attach that to the matrix
6747: PetscCall(PetscMalloc1(1, &coo));
6748: coo->n = coo_n;
6749: coo->sf = sf2;
6750: coo->sendlen = nleaves;
6751: coo->recvlen = nroots;
6752: coo->Annz = Annz;
6753: coo->Bnnz = Bnnz;
6754: coo->Annz2 = Annz2;
6755: coo->Bnnz2 = Bnnz2;
6756: coo->Atot1 = Atot1;
6757: coo->Atot2 = Atot2;
6758: coo->Btot1 = Btot1;
6759: coo->Btot2 = Btot2;
6760: coo->Ajmap1 = Ajmap1;
6761: coo->Aperm1 = Aperm1;
6762: coo->Bjmap1 = Bjmap1;
6763: coo->Bperm1 = Bperm1;
6764: coo->Aimap2 = Aimap2;
6765: coo->Ajmap2 = Ajmap2;
6766: coo->Aperm2 = Aperm2;
6767: coo->Bimap2 = Bimap2;
6768: coo->Bjmap2 = Bjmap2;
6769: coo->Bperm2 = Bperm2;
6770: coo->Cperm1 = Cperm1;
6771: // Allocate in preallocation. If not used, it has zero cost on host
6772: PetscCall(PetscMalloc2(coo->sendlen, &coo->sendbuf, coo->recvlen, &coo->recvbuf));
6773: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
6774: PetscCall(PetscContainerSetPointer(container, coo));
6775: PetscCall(PetscContainerSetUserDestroy(container, MatCOOStructDestroy_MPIAIJ));
6776: PetscCall(PetscObjectCompose((PetscObject)mat, "__PETSc_MatCOOStruct_Host", (PetscObject)container));
6777: PetscCall(PetscContainerDestroy(&container));
6778: PetscFunctionReturn(PETSC_SUCCESS);
6779: }
6781: static PetscErrorCode MatSetValuesCOO_MPIAIJ(Mat mat, const PetscScalar v[], InsertMode imode)
6782: {
6783: Mat_MPIAIJ *mpiaij = (Mat_MPIAIJ *)mat->data;
6784: Mat A = mpiaij->A, B = mpiaij->B;
6785: PetscScalar *Aa, *Ba;
6786: PetscScalar *sendbuf, *recvbuf;
6787: const PetscCount *Ajmap1, *Ajmap2, *Aimap2;
6788: const PetscCount *Bjmap1, *Bjmap2, *Bimap2;
6789: const PetscCount *Aperm1, *Aperm2, *Bperm1, *Bperm2;
6790: const PetscCount *Cperm1;
6791: PetscContainer container;
6792: MatCOOStruct_MPIAIJ *coo;
6794: PetscFunctionBegin;
6795: PetscCall(PetscObjectQuery((PetscObject)mat, "__PETSc_MatCOOStruct_Host", (PetscObject *)&container));
6796: PetscCheck(container, PetscObjectComm((PetscObject)mat), PETSC_ERR_PLIB, "Not found MatCOOStruct on this matrix");
6797: PetscCall(PetscContainerGetPointer(container, (void **)&coo));
6798: sendbuf = coo->sendbuf;
6799: recvbuf = coo->recvbuf;
6800: Ajmap1 = coo->Ajmap1;
6801: Ajmap2 = coo->Ajmap2;
6802: Aimap2 = coo->Aimap2;
6803: Bjmap1 = coo->Bjmap1;
6804: Bjmap2 = coo->Bjmap2;
6805: Bimap2 = coo->Bimap2;
6806: Aperm1 = coo->Aperm1;
6807: Aperm2 = coo->Aperm2;
6808: Bperm1 = coo->Bperm1;
6809: Bperm2 = coo->Bperm2;
6810: Cperm1 = coo->Cperm1;
6812: PetscCall(MatSeqAIJGetArray(A, &Aa)); /* Might read and write matrix values */
6813: PetscCall(MatSeqAIJGetArray(B, &Ba));
6815: /* Pack entries to be sent to remote */
6816: for (PetscCount i = 0; i < coo->sendlen; i++) sendbuf[i] = v[Cperm1[i]];
6818: /* Send remote entries to their owner and overlap the communication with local computation */
6819: PetscCall(PetscSFReduceWithMemTypeBegin(coo->sf, MPIU_SCALAR, PETSC_MEMTYPE_HOST, sendbuf, PETSC_MEMTYPE_HOST, recvbuf, MPI_REPLACE));
6820: /* Add local entries to A and B */
6821: for (PetscCount i = 0; i < coo->Annz; i++) { /* All nonzeros in A are either zero'ed or added with a value (i.e., initialized) */
6822: PetscScalar sum = 0.0; /* Do partial summation first to improve numerical stability */
6823: for (PetscCount k = Ajmap1[i]; k < Ajmap1[i + 1]; k++) sum += v[Aperm1[k]];
6824: Aa[i] = (imode == INSERT_VALUES ? 0.0 : Aa[i]) + sum;
6825: }
6826: for (PetscCount i = 0; i < coo->Bnnz; i++) {
6827: PetscScalar sum = 0.0;
6828: for (PetscCount k = Bjmap1[i]; k < Bjmap1[i + 1]; k++) sum += v[Bperm1[k]];
6829: Ba[i] = (imode == INSERT_VALUES ? 0.0 : Ba[i]) + sum;
6830: }
6831: PetscCall(PetscSFReduceEnd(coo->sf, MPIU_SCALAR, sendbuf, recvbuf, MPI_REPLACE));
6833: /* Add received remote entries to A and B */
6834: for (PetscCount i = 0; i < coo->Annz2; i++) {
6835: for (PetscCount k = Ajmap2[i]; k < Ajmap2[i + 1]; k++) Aa[Aimap2[i]] += recvbuf[Aperm2[k]];
6836: }
6837: for (PetscCount i = 0; i < coo->Bnnz2; i++) {
6838: for (PetscCount k = Bjmap2[i]; k < Bjmap2[i + 1]; k++) Ba[Bimap2[i]] += recvbuf[Bperm2[k]];
6839: }
6840: PetscCall(MatSeqAIJRestoreArray(A, &Aa));
6841: PetscCall(MatSeqAIJRestoreArray(B, &Ba));
6842: PetscFunctionReturn(PETSC_SUCCESS);
6843: }
6845: /*MC
6846: MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices.
6848: Options Database Keys:
6849: . -mat_type mpiaij - sets the matrix type to `MATMPIAIJ` during a call to `MatSetFromOptions()`
6851: Level: beginner
6853: Notes:
6854: `MatSetValues()` may be called for this matrix type with a `NULL` argument for the numerical values,
6855: in this case the values associated with the rows and columns one passes in are set to zero
6856: in the matrix
6858: `MatSetOptions`(,`MAT_STRUCTURE_ONLY`,`PETSC_TRUE`) may be called for this matrix type. In this no
6859: space is allocated for the nonzero entries and any entries passed with `MatSetValues()` are ignored
6861: .seealso: [](ch_matrices), `Mat`, `MATSEQAIJ`, `MATAIJ`, `MatCreateAIJ()`
6862: M*/
6863: PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat B)
6864: {
6865: Mat_MPIAIJ *b;
6866: PetscMPIInt size;
6868: PetscFunctionBegin;
6869: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
6871: PetscCall(PetscNew(&b));
6872: B->data = (void *)b;
6873: B->ops[0] = MatOps_Values;
6874: B->assembled = PETSC_FALSE;
6875: B->insertmode = NOT_SET_VALUES;
6876: b->size = size;
6878: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)B), &b->rank));
6880: /* build cache for off array entries formed */
6881: PetscCall(MatStashCreate_Private(PetscObjectComm((PetscObject)B), 1, &B->stash));
6883: b->donotstash = PETSC_FALSE;
6884: b->colmap = NULL;
6885: b->garray = NULL;
6886: b->roworiented = PETSC_TRUE;
6888: /* stuff used for matrix vector multiply */
6889: b->lvec = NULL;
6890: b->Mvctx = NULL;
6892: /* stuff for MatGetRow() */
6893: b->rowindices = NULL;
6894: b->rowvalues = NULL;
6895: b->getrowactive = PETSC_FALSE;
6897: /* flexible pointer used in CUSPARSE classes */
6898: b->spptr = NULL;
6900: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetUseScalableIncreaseOverlap_C", MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ));
6901: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatStoreValues_C", MatStoreValues_MPIAIJ));
6902: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatRetrieveValues_C", MatRetrieveValues_MPIAIJ));
6903: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatIsTranspose_C", MatIsTranspose_MPIAIJ));
6904: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetPreallocation_C", MatMPIAIJSetPreallocation_MPIAIJ));
6905: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatResetPreallocation_C", MatResetPreallocation_MPIAIJ));
6906: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetPreallocationCSR_C", MatMPIAIJSetPreallocationCSR_MPIAIJ));
6907: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDiagonalScaleLocal_C", MatDiagonalScaleLocal_MPIAIJ));
6908: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijperm_C", MatConvert_MPIAIJ_MPIAIJPERM));
6909: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijsell_C", MatConvert_MPIAIJ_MPIAIJSELL));
6910: #if defined(PETSC_HAVE_CUDA)
6911: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijcusparse_C", MatConvert_MPIAIJ_MPIAIJCUSPARSE));
6912: #endif
6913: #if defined(PETSC_HAVE_HIP)
6914: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijhipsparse_C", MatConvert_MPIAIJ_MPIAIJHIPSPARSE));
6915: #endif
6916: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
6917: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijkokkos_C", MatConvert_MPIAIJ_MPIAIJKokkos));
6918: #endif
6919: #if defined(PETSC_HAVE_MKL_SPARSE)
6920: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijmkl_C", MatConvert_MPIAIJ_MPIAIJMKL));
6921: #endif
6922: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijcrl_C", MatConvert_MPIAIJ_MPIAIJCRL));
6923: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpibaij_C", MatConvert_MPIAIJ_MPIBAIJ));
6924: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpisbaij_C", MatConvert_MPIAIJ_MPISBAIJ));
6925: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpidense_C", MatConvert_MPIAIJ_MPIDense));
6926: #if defined(PETSC_HAVE_ELEMENTAL)
6927: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_elemental_C", MatConvert_MPIAIJ_Elemental));
6928: #endif
6929: #if defined(PETSC_HAVE_SCALAPACK)
6930: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_scalapack_C", MatConvert_AIJ_ScaLAPACK));
6931: #endif
6932: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_is_C", MatConvert_XAIJ_IS));
6933: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpisell_C", MatConvert_MPIAIJ_MPISELL));
6934: #if defined(PETSC_HAVE_HYPRE)
6935: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_hypre_C", MatConvert_AIJ_HYPRE));
6936: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_transpose_mpiaij_mpiaij_C", MatProductSetFromOptions_Transpose_AIJ_AIJ));
6937: #endif
6938: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_is_mpiaij_C", MatProductSetFromOptions_IS_XAIJ));
6939: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_mpiaij_mpiaij_C", MatProductSetFromOptions_MPIAIJ));
6940: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetPreallocationCOO_C", MatSetPreallocationCOO_MPIAIJ));
6941: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetValuesCOO_C", MatSetValuesCOO_MPIAIJ));
6942: PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATMPIAIJ));
6943: PetscFunctionReturn(PETSC_SUCCESS);
6944: }
6946: /*@
6947: MatCreateMPIAIJWithSplitArrays - creates a `MATMPIAIJ` matrix using arrays that contain the "diagonal"
6948: and "off-diagonal" part of the matrix in CSR format.
6950: Collective
6952: Input Parameters:
6953: + comm - MPI communicator
6954: . m - number of local rows (Cannot be `PETSC_DECIDE`)
6955: . n - This value should be the same as the local size used in creating the
6956: x vector for the matrix-vector product $y = Ax$. (or `PETSC_DECIDE` to have
6957: calculated if `N` is given) For square matrices `n` is almost always `m`.
6958: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if `m` is given)
6959: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if `n` is given)
6960: . 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
6961: . j - column indices, which must be local, i.e., based off the start column of the diagonal portion
6962: . a - matrix values
6963: . 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
6964: . oj - column indices, which must be global, representing global columns in the `MATMPIAIJ` matrix
6965: - oa - matrix values
6967: Output Parameter:
6968: . mat - the matrix
6970: Level: advanced
6972: Notes:
6973: The `i`, `j`, and `a` arrays ARE NOT copied by this routine into the internal format used by PETSc (even in Fortran). The user
6974: must free the arrays once the matrix has been destroyed and not before.
6976: The `i` and `j` indices are 0 based
6978: See `MatCreateAIJ()` for the definition of "diagonal" and "off-diagonal" portion of the matrix
6980: This sets local rows and cannot be used to set off-processor values.
6982: Use of this routine is discouraged because it is inflexible and cumbersome to use. It is extremely rare that a
6983: legacy application natively assembles into exactly this split format. The code to do so is nontrivial and does
6984: not easily support in-place reassembly. It is recommended to use MatSetValues() (or a variant thereof) because
6985: the resulting assembly is easier to implement, will work with any matrix format, and the user does not have to
6986: keep track of the underlying array. Use `MatSetOption`(A,`MAT_NO_OFF_PROC_ENTRIES`,`PETSC_TRUE`) to disable all
6987: communication if it is known that only local entries will be set.
6989: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
6990: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithArrays()`
6991: @*/
6992: 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)
6993: {
6994: Mat_MPIAIJ *maij;
6996: PetscFunctionBegin;
6997: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
6998: PetscCheck(i[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
6999: PetscCheck(oi[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "oi (row indices) must start with 0");
7000: PetscCall(MatCreate(comm, mat));
7001: PetscCall(MatSetSizes(*mat, m, n, M, N));
7002: PetscCall(MatSetType(*mat, MATMPIAIJ));
7003: maij = (Mat_MPIAIJ *)(*mat)->data;
7005: (*mat)->preallocated = PETSC_TRUE;
7007: PetscCall(PetscLayoutSetUp((*mat)->rmap));
7008: PetscCall(PetscLayoutSetUp((*mat)->cmap));
7010: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, n, i, j, a, &maij->A));
7011: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, (*mat)->cmap->N, oi, oj, oa, &maij->B));
7013: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
7014: PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY));
7015: PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY));
7016: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_FALSE));
7017: PetscCall(MatSetOption(*mat, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
7018: PetscFunctionReturn(PETSC_SUCCESS);
7019: }
7021: typedef struct {
7022: Mat *mp; /* intermediate products */
7023: PetscBool *mptmp; /* is the intermediate product temporary ? */
7024: PetscInt cp; /* number of intermediate products */
7026: /* support for MatGetBrowsOfAoCols_MPIAIJ for P_oth */
7027: PetscInt *startsj_s, *startsj_r;
7028: PetscScalar *bufa;
7029: Mat P_oth;
7031: /* may take advantage of merging product->B */
7032: Mat Bloc; /* B-local by merging diag and off-diag */
7034: /* cusparse does not have support to split between symbolic and numeric phases.
7035: When api_user is true, we don't need to update the numerical values
7036: of the temporary storage */
7037: PetscBool reusesym;
7039: /* support for COO values insertion */
7040: PetscScalar *coo_v, *coo_w; /* store on-process and off-process COO scalars, and used as MPI recv/send buffers respectively */
7041: PetscInt **own; /* own[i] points to address of on-process COO indices for Mat mp[i] */
7042: PetscInt **off; /* off[i] points to address of off-process COO indices for Mat mp[i] */
7043: PetscBool hasoffproc; /* if true, have off-process values insertion (i.e. AtB or PtAP) */
7044: PetscSF sf; /* used for non-local values insertion and memory malloc */
7045: PetscMemType mtype;
7047: /* customization */
7048: PetscBool abmerge;
7049: PetscBool P_oth_bind;
7050: } MatMatMPIAIJBACKEND;
7052: static PetscErrorCode MatDestroy_MatMatMPIAIJBACKEND(void *data)
7053: {
7054: MatMatMPIAIJBACKEND *mmdata = (MatMatMPIAIJBACKEND *)data;
7055: PetscInt i;
7057: PetscFunctionBegin;
7058: PetscCall(PetscFree2(mmdata->startsj_s, mmdata->startsj_r));
7059: PetscCall(PetscFree(mmdata->bufa));
7060: PetscCall(PetscSFFree(mmdata->sf, mmdata->mtype, mmdata->coo_v));
7061: PetscCall(PetscSFFree(mmdata->sf, mmdata->mtype, mmdata->coo_w));
7062: PetscCall(MatDestroy(&mmdata->P_oth));
7063: PetscCall(MatDestroy(&mmdata->Bloc));
7064: PetscCall(PetscSFDestroy(&mmdata->sf));
7065: for (i = 0; i < mmdata->cp; i++) PetscCall(MatDestroy(&mmdata->mp[i]));
7066: PetscCall(PetscFree2(mmdata->mp, mmdata->mptmp));
7067: PetscCall(PetscFree(mmdata->own[0]));
7068: PetscCall(PetscFree(mmdata->own));
7069: PetscCall(PetscFree(mmdata->off[0]));
7070: PetscCall(PetscFree(mmdata->off));
7071: PetscCall(PetscFree(mmdata));
7072: PetscFunctionReturn(PETSC_SUCCESS);
7073: }
7075: /* Copy selected n entries with indices in idx[] of A to v[].
7076: If idx is NULL, copy the whole data array of A to v[]
7077: */
7078: static PetscErrorCode MatSeqAIJCopySubArray(Mat A, PetscInt n, const PetscInt idx[], PetscScalar v[])
7079: {
7080: PetscErrorCode (*f)(Mat, PetscInt, const PetscInt[], PetscScalar[]);
7082: PetscFunctionBegin;
7083: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatSeqAIJCopySubArray_C", &f));
7084: if (f) {
7085: PetscCall((*f)(A, n, idx, v));
7086: } else {
7087: const PetscScalar *vv;
7089: PetscCall(MatSeqAIJGetArrayRead(A, &vv));
7090: if (n && idx) {
7091: PetscScalar *w = v;
7092: const PetscInt *oi = idx;
7093: PetscInt j;
7095: for (j = 0; j < n; j++) *w++ = vv[*oi++];
7096: } else {
7097: PetscCall(PetscArraycpy(v, vv, n));
7098: }
7099: PetscCall(MatSeqAIJRestoreArrayRead(A, &vv));
7100: }
7101: PetscFunctionReturn(PETSC_SUCCESS);
7102: }
7104: static PetscErrorCode MatProductNumeric_MPIAIJBACKEND(Mat C)
7105: {
7106: MatMatMPIAIJBACKEND *mmdata;
7107: PetscInt i, n_d, n_o;
7109: PetscFunctionBegin;
7110: MatCheckProduct(C, 1);
7111: PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data empty");
7112: mmdata = (MatMatMPIAIJBACKEND *)C->product->data;
7113: if (!mmdata->reusesym) { /* update temporary matrices */
7114: 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));
7115: if (mmdata->Bloc) PetscCall(MatMPIAIJGetLocalMatMerge(C->product->B, MAT_REUSE_MATRIX, NULL, &mmdata->Bloc));
7116: }
7117: mmdata->reusesym = PETSC_FALSE;
7119: for (i = 0; i < mmdata->cp; i++) {
7120: 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]);
7121: PetscCall((*mmdata->mp[i]->ops->productnumeric)(mmdata->mp[i]));
7122: }
7123: for (i = 0, n_d = 0, n_o = 0; i < mmdata->cp; i++) {
7124: PetscInt noff;
7126: PetscCall(PetscIntCast(mmdata->off[i + 1] - mmdata->off[i], &noff));
7127: if (mmdata->mptmp[i]) continue;
7128: if (noff) {
7129: PetscInt nown;
7131: PetscCall(PetscIntCast(mmdata->own[i + 1] - mmdata->own[i], &nown));
7132: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], noff, mmdata->off[i], mmdata->coo_w + n_o));
7133: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], nown, mmdata->own[i], mmdata->coo_v + n_d));
7134: n_o += noff;
7135: n_d += nown;
7136: } else {
7137: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mmdata->mp[i]->data;
7139: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], mm->nz, NULL, mmdata->coo_v + n_d));
7140: n_d += mm->nz;
7141: }
7142: }
7143: if (mmdata->hasoffproc) { /* offprocess insertion */
7144: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_SCALAR, mmdata->coo_w, mmdata->coo_v + n_d));
7145: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_SCALAR, mmdata->coo_w, mmdata->coo_v + n_d));
7146: }
7147: PetscCall(MatSetValuesCOO(C, mmdata->coo_v, INSERT_VALUES));
7148: PetscFunctionReturn(PETSC_SUCCESS);
7149: }
7151: /* Support for Pt * A, A * P, or Pt * A * P */
7152: #define MAX_NUMBER_INTERMEDIATE 4
7153: PetscErrorCode MatProductSymbolic_MPIAIJBACKEND(Mat C)
7154: {
7155: Mat_Product *product = C->product;
7156: Mat A, P, mp[MAX_NUMBER_INTERMEDIATE]; /* A, P and a series of intermediate matrices */
7157: Mat_MPIAIJ *a, *p;
7158: MatMatMPIAIJBACKEND *mmdata;
7159: ISLocalToGlobalMapping P_oth_l2g = NULL;
7160: IS glob = NULL;
7161: const char *prefix;
7162: char pprefix[256];
7163: const PetscInt *globidx, *P_oth_idx;
7164: PetscInt i, j, cp, m, n, M, N, *coo_i, *coo_j;
7165: PetscCount ncoo, ncoo_d, ncoo_o, ncoo_oown;
7166: PetscInt cmapt[MAX_NUMBER_INTERMEDIATE], rmapt[MAX_NUMBER_INTERMEDIATE]; /* col/row map type for each Mat in mp[]. */
7167: /* type-0: consecutive, start from 0; type-1: consecutive with */
7168: /* a base offset; type-2: sparse with a local to global map table */
7169: const PetscInt *cmapa[MAX_NUMBER_INTERMEDIATE], *rmapa[MAX_NUMBER_INTERMEDIATE]; /* col/row local to global map array (table) for type-2 map type */
7171: MatProductType ptype;
7172: PetscBool mptmp[MAX_NUMBER_INTERMEDIATE], hasoffproc = PETSC_FALSE, iscuda, iship, iskokk;
7173: PetscMPIInt size;
7175: PetscFunctionBegin;
7176: MatCheckProduct(C, 1);
7177: PetscCheck(!product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data not empty");
7178: ptype = product->type;
7179: if (product->A->symmetric == PETSC_BOOL3_TRUE && ptype == MATPRODUCT_AtB) {
7180: ptype = MATPRODUCT_AB;
7181: product->symbolic_used_the_fact_A_is_symmetric = PETSC_TRUE;
7182: }
7183: switch (ptype) {
7184: case MATPRODUCT_AB:
7185: A = product->A;
7186: P = product->B;
7187: m = A->rmap->n;
7188: n = P->cmap->n;
7189: M = A->rmap->N;
7190: N = P->cmap->N;
7191: hasoffproc = PETSC_FALSE; /* will not scatter mat product values to other processes */
7192: break;
7193: case MATPRODUCT_AtB:
7194: P = product->A;
7195: A = product->B;
7196: m = P->cmap->n;
7197: n = A->cmap->n;
7198: M = P->cmap->N;
7199: N = A->cmap->N;
7200: hasoffproc = PETSC_TRUE;
7201: break;
7202: case MATPRODUCT_PtAP:
7203: A = product->A;
7204: P = product->B;
7205: m = P->cmap->n;
7206: n = P->cmap->n;
7207: M = P->cmap->N;
7208: N = P->cmap->N;
7209: hasoffproc = PETSC_TRUE;
7210: break;
7211: default:
7212: SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Not for product type %s", MatProductTypes[ptype]);
7213: }
7214: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)C), &size));
7215: if (size == 1) hasoffproc = PETSC_FALSE;
7217: /* defaults */
7218: for (i = 0; i < MAX_NUMBER_INTERMEDIATE; i++) {
7219: mp[i] = NULL;
7220: mptmp[i] = PETSC_FALSE;
7221: rmapt[i] = -1;
7222: cmapt[i] = -1;
7223: rmapa[i] = NULL;
7224: cmapa[i] = NULL;
7225: }
7227: /* customization */
7228: PetscCall(PetscNew(&mmdata));
7229: mmdata->reusesym = product->api_user;
7230: if (ptype == MATPRODUCT_AB) {
7231: if (product->api_user) {
7232: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatMatMult", "Mat");
7233: PetscCall(PetscOptionsBool("-matmatmult_backend_mergeB", "Merge product->B local matrices", "MatMatMult", mmdata->abmerge, &mmdata->abmerge, NULL));
7234: PetscCall(PetscOptionsBool("-matmatmult_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7235: PetscOptionsEnd();
7236: } else {
7237: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_AB", "Mat");
7238: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_mergeB", "Merge product->B local matrices", "MatMatMult", mmdata->abmerge, &mmdata->abmerge, NULL));
7239: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7240: PetscOptionsEnd();
7241: }
7242: } else if (ptype == MATPRODUCT_PtAP) {
7243: if (product->api_user) {
7244: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatPtAP", "Mat");
7245: PetscCall(PetscOptionsBool("-matptap_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7246: PetscOptionsEnd();
7247: } else {
7248: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_PtAP", "Mat");
7249: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7250: PetscOptionsEnd();
7251: }
7252: }
7253: a = (Mat_MPIAIJ *)A->data;
7254: p = (Mat_MPIAIJ *)P->data;
7255: PetscCall(MatSetSizes(C, m, n, M, N));
7256: PetscCall(PetscLayoutSetUp(C->rmap));
7257: PetscCall(PetscLayoutSetUp(C->cmap));
7258: PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
7259: PetscCall(MatGetOptionsPrefix(C, &prefix));
7261: cp = 0;
7262: switch (ptype) {
7263: case MATPRODUCT_AB: /* A * P */
7264: PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &mmdata->startsj_s, &mmdata->startsj_r, &mmdata->bufa, &mmdata->P_oth));
7266: /* A_diag * P_local (merged or not) */
7267: if (mmdata->abmerge) { /* P's diagonal and off-diag blocks are merged to one matrix, then multiplied by A_diag */
7268: /* P is product->B */
7269: PetscCall(MatMPIAIJGetLocalMatMerge(P, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7270: PetscCall(MatProductCreate(a->A, mmdata->Bloc, NULL, &mp[cp]));
7271: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7272: PetscCall(MatProductSetFill(mp[cp], product->fill));
7273: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7274: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7275: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7276: mp[cp]->product->api_user = product->api_user;
7277: PetscCall(MatProductSetFromOptions(mp[cp]));
7278: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7279: PetscCall(ISGetIndices(glob, &globidx));
7280: rmapt[cp] = 1;
7281: cmapt[cp] = 2;
7282: cmapa[cp] = globidx;
7283: mptmp[cp] = PETSC_FALSE;
7284: cp++;
7285: } else { /* A_diag * P_diag and A_diag * P_off */
7286: PetscCall(MatProductCreate(a->A, p->A, NULL, &mp[cp]));
7287: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7288: PetscCall(MatProductSetFill(mp[cp], product->fill));
7289: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7290: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7291: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7292: mp[cp]->product->api_user = product->api_user;
7293: PetscCall(MatProductSetFromOptions(mp[cp]));
7294: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7295: rmapt[cp] = 1;
7296: cmapt[cp] = 1;
7297: mptmp[cp] = PETSC_FALSE;
7298: cp++;
7299: PetscCall(MatProductCreate(a->A, p->B, NULL, &mp[cp]));
7300: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7301: PetscCall(MatProductSetFill(mp[cp], product->fill));
7302: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7303: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7304: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7305: mp[cp]->product->api_user = product->api_user;
7306: PetscCall(MatProductSetFromOptions(mp[cp]));
7307: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7308: rmapt[cp] = 1;
7309: cmapt[cp] = 2;
7310: cmapa[cp] = p->garray;
7311: mptmp[cp] = PETSC_FALSE;
7312: cp++;
7313: }
7315: /* A_off * P_other */
7316: if (mmdata->P_oth) {
7317: PetscCall(MatSeqAIJCompactOutExtraColumns_SeqAIJ(mmdata->P_oth, &P_oth_l2g)); /* make P_oth use local col ids */
7318: PetscCall(ISLocalToGlobalMappingGetIndices(P_oth_l2g, &P_oth_idx));
7319: PetscCall(MatSetType(mmdata->P_oth, ((PetscObject)a->B)->type_name));
7320: PetscCall(MatBindToCPU(mmdata->P_oth, mmdata->P_oth_bind));
7321: PetscCall(MatProductCreate(a->B, mmdata->P_oth, NULL, &mp[cp]));
7322: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7323: PetscCall(MatProductSetFill(mp[cp], product->fill));
7324: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7325: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7326: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7327: mp[cp]->product->api_user = product->api_user;
7328: PetscCall(MatProductSetFromOptions(mp[cp]));
7329: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7330: rmapt[cp] = 1;
7331: cmapt[cp] = 2;
7332: cmapa[cp] = P_oth_idx;
7333: mptmp[cp] = PETSC_FALSE;
7334: cp++;
7335: }
7336: break;
7338: case MATPRODUCT_AtB: /* (P^t * A): P_diag * A_loc + P_off * A_loc */
7339: /* A is product->B */
7340: PetscCall(MatMPIAIJGetLocalMatMerge(A, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7341: if (A == P) { /* when A==P, we can take advantage of the already merged mmdata->Bloc */
7342: PetscCall(MatProductCreate(mmdata->Bloc, mmdata->Bloc, NULL, &mp[cp]));
7343: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7344: PetscCall(MatProductSetFill(mp[cp], product->fill));
7345: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7346: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7347: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7348: mp[cp]->product->api_user = product->api_user;
7349: PetscCall(MatProductSetFromOptions(mp[cp]));
7350: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7351: PetscCall(ISGetIndices(glob, &globidx));
7352: rmapt[cp] = 2;
7353: rmapa[cp] = globidx;
7354: cmapt[cp] = 2;
7355: cmapa[cp] = globidx;
7356: mptmp[cp] = PETSC_FALSE;
7357: cp++;
7358: } else {
7359: PetscCall(MatProductCreate(p->A, mmdata->Bloc, NULL, &mp[cp]));
7360: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7361: PetscCall(MatProductSetFill(mp[cp], product->fill));
7362: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7363: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7364: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7365: mp[cp]->product->api_user = product->api_user;
7366: PetscCall(MatProductSetFromOptions(mp[cp]));
7367: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7368: PetscCall(ISGetIndices(glob, &globidx));
7369: rmapt[cp] = 1;
7370: cmapt[cp] = 2;
7371: cmapa[cp] = globidx;
7372: mptmp[cp] = PETSC_FALSE;
7373: cp++;
7374: PetscCall(MatProductCreate(p->B, mmdata->Bloc, NULL, &mp[cp]));
7375: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7376: PetscCall(MatProductSetFill(mp[cp], product->fill));
7377: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7378: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7379: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7380: mp[cp]->product->api_user = product->api_user;
7381: PetscCall(MatProductSetFromOptions(mp[cp]));
7382: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7383: rmapt[cp] = 2;
7384: rmapa[cp] = p->garray;
7385: cmapt[cp] = 2;
7386: cmapa[cp] = globidx;
7387: mptmp[cp] = PETSC_FALSE;
7388: cp++;
7389: }
7390: break;
7391: case MATPRODUCT_PtAP:
7392: PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &mmdata->startsj_s, &mmdata->startsj_r, &mmdata->bufa, &mmdata->P_oth));
7393: /* P is product->B */
7394: PetscCall(MatMPIAIJGetLocalMatMerge(P, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7395: PetscCall(MatProductCreate(a->A, mmdata->Bloc, NULL, &mp[cp]));
7396: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_PtAP));
7397: PetscCall(MatProductSetFill(mp[cp], product->fill));
7398: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7399: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7400: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7401: mp[cp]->product->api_user = product->api_user;
7402: PetscCall(MatProductSetFromOptions(mp[cp]));
7403: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7404: PetscCall(ISGetIndices(glob, &globidx));
7405: rmapt[cp] = 2;
7406: rmapa[cp] = globidx;
7407: cmapt[cp] = 2;
7408: cmapa[cp] = globidx;
7409: mptmp[cp] = PETSC_FALSE;
7410: cp++;
7411: if (mmdata->P_oth) {
7412: PetscCall(MatSeqAIJCompactOutExtraColumns_SeqAIJ(mmdata->P_oth, &P_oth_l2g));
7413: PetscCall(ISLocalToGlobalMappingGetIndices(P_oth_l2g, &P_oth_idx));
7414: PetscCall(MatSetType(mmdata->P_oth, ((PetscObject)a->B)->type_name));
7415: PetscCall(MatBindToCPU(mmdata->P_oth, mmdata->P_oth_bind));
7416: PetscCall(MatProductCreate(a->B, mmdata->P_oth, NULL, &mp[cp]));
7417: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7418: PetscCall(MatProductSetFill(mp[cp], product->fill));
7419: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7420: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7421: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7422: mp[cp]->product->api_user = product->api_user;
7423: PetscCall(MatProductSetFromOptions(mp[cp]));
7424: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7425: mptmp[cp] = PETSC_TRUE;
7426: cp++;
7427: PetscCall(MatProductCreate(mmdata->Bloc, mp[1], NULL, &mp[cp]));
7428: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7429: PetscCall(MatProductSetFill(mp[cp], product->fill));
7430: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7431: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7432: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7433: mp[cp]->product->api_user = product->api_user;
7434: PetscCall(MatProductSetFromOptions(mp[cp]));
7435: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7436: rmapt[cp] = 2;
7437: rmapa[cp] = globidx;
7438: cmapt[cp] = 2;
7439: cmapa[cp] = P_oth_idx;
7440: mptmp[cp] = PETSC_FALSE;
7441: cp++;
7442: }
7443: break;
7444: default:
7445: SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Not for product type %s", MatProductTypes[ptype]);
7446: }
7447: /* sanity check */
7448: if (size > 1)
7449: 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);
7451: PetscCall(PetscMalloc2(cp, &mmdata->mp, cp, &mmdata->mptmp));
7452: for (i = 0; i < cp; i++) {
7453: mmdata->mp[i] = mp[i];
7454: mmdata->mptmp[i] = mptmp[i];
7455: }
7456: mmdata->cp = cp;
7457: C->product->data = mmdata;
7458: C->product->destroy = MatDestroy_MatMatMPIAIJBACKEND;
7459: C->ops->productnumeric = MatProductNumeric_MPIAIJBACKEND;
7461: /* memory type */
7462: mmdata->mtype = PETSC_MEMTYPE_HOST;
7463: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iscuda, MATSEQAIJCUSPARSE, MATMPIAIJCUSPARSE, ""));
7464: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iship, MATSEQAIJHIPSPARSE, MATMPIAIJHIPSPARSE, ""));
7465: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iskokk, MATSEQAIJKOKKOS, MATMPIAIJKOKKOS, ""));
7466: if (iscuda) mmdata->mtype = PETSC_MEMTYPE_CUDA;
7467: else if (iship) mmdata->mtype = PETSC_MEMTYPE_HIP;
7468: else if (iskokk) mmdata->mtype = PETSC_MEMTYPE_KOKKOS;
7470: /* prepare coo coordinates for values insertion */
7472: /* count total nonzeros of those intermediate seqaij Mats
7473: ncoo_d: # of nonzeros of matrices that do not have offproc entries
7474: ncoo_o: # of nonzeros (of matrices that might have offproc entries) that will be inserted to remote procs
7475: ncoo_oown: # of nonzeros (of matrices that might have offproc entries) that will be inserted locally
7476: */
7477: for (cp = 0, ncoo_d = 0, ncoo_o = 0, ncoo_oown = 0; cp < mmdata->cp; cp++) {
7478: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7479: if (mptmp[cp]) continue;
7480: if (rmapt[cp] == 2 && hasoffproc) { /* the rows need to be scatter to all processes (might include self) */
7481: const PetscInt *rmap = rmapa[cp];
7482: const PetscInt mr = mp[cp]->rmap->n;
7483: const PetscInt rs = C->rmap->rstart;
7484: const PetscInt re = C->rmap->rend;
7485: const PetscInt *ii = mm->i;
7486: for (i = 0; i < mr; i++) {
7487: const PetscInt gr = rmap[i];
7488: const PetscInt nz = ii[i + 1] - ii[i];
7489: if (gr < rs || gr >= re) ncoo_o += nz; /* this row is offproc */
7490: else ncoo_oown += nz; /* this row is local */
7491: }
7492: } else ncoo_d += mm->nz;
7493: }
7495: /*
7496: ncoo: total number of nonzeros (including those inserted by remote procs) belonging to this proc
7498: ncoo = ncoo_d + ncoo_oown + ncoo2, which ncoo2 is number of nonzeros inserted to me by other procs.
7500: off[0] points to a big index array, which is shared by off[1,2,...]. Similarly, for own[0].
7502: off[p]: points to the segment for matrix mp[p], storing location of nonzeros that mp[p] will insert to others
7503: own[p]: points to the segment for matrix mp[p], storing location of nonzeros that mp[p] will insert locally
7504: so, off[p+1]-off[p] is the number of nonzeros that mp[p] will send to others.
7506: coo_i/j/v[]: [ncoo] row/col/val of nonzeros belonging to this proc.
7507: 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.
7508: */
7509: PetscCall(PetscCalloc1(mmdata->cp + 1, &mmdata->off)); /* +1 to make a csr-like data structure */
7510: PetscCall(PetscCalloc1(mmdata->cp + 1, &mmdata->own));
7512: /* gather (i,j) of nonzeros inserted by remote procs */
7513: if (hasoffproc) {
7514: PetscSF msf;
7515: PetscInt ncoo2, *coo_i2, *coo_j2;
7517: PetscCall(PetscMalloc1(ncoo_o, &mmdata->off[0]));
7518: PetscCall(PetscMalloc1(ncoo_oown, &mmdata->own[0]));
7519: PetscCall(PetscMalloc2(ncoo_o, &coo_i, ncoo_o, &coo_j)); /* to collect (i,j) of entries to be sent to others */
7521: for (cp = 0, ncoo_o = 0; cp < mmdata->cp; cp++) {
7522: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7523: PetscInt *idxoff = mmdata->off[cp];
7524: PetscInt *idxown = mmdata->own[cp];
7525: if (!mptmp[cp] && rmapt[cp] == 2) { /* row map is sparse */
7526: const PetscInt *rmap = rmapa[cp];
7527: const PetscInt *cmap = cmapa[cp];
7528: const PetscInt *ii = mm->i;
7529: PetscInt *coi = coo_i + ncoo_o;
7530: PetscInt *coj = coo_j + ncoo_o;
7531: const PetscInt mr = mp[cp]->rmap->n;
7532: const PetscInt rs = C->rmap->rstart;
7533: const PetscInt re = C->rmap->rend;
7534: const PetscInt cs = C->cmap->rstart;
7535: for (i = 0; i < mr; i++) {
7536: const PetscInt *jj = mm->j + ii[i];
7537: const PetscInt gr = rmap[i];
7538: const PetscInt nz = ii[i + 1] - ii[i];
7539: if (gr < rs || gr >= re) { /* this is an offproc row */
7540: for (j = ii[i]; j < ii[i + 1]; j++) {
7541: *coi++ = gr;
7542: *idxoff++ = j;
7543: }
7544: if (!cmapt[cp]) { /* already global */
7545: for (j = 0; j < nz; j++) *coj++ = jj[j];
7546: } else if (cmapt[cp] == 1) { /* local to global for owned columns of C */
7547: for (j = 0; j < nz; j++) *coj++ = jj[j] + cs;
7548: } else { /* offdiag */
7549: for (j = 0; j < nz; j++) *coj++ = cmap[jj[j]];
7550: }
7551: ncoo_o += nz;
7552: } else { /* this is a local row */
7553: for (j = ii[i]; j < ii[i + 1]; j++) *idxown++ = j;
7554: }
7555: }
7556: }
7557: mmdata->off[cp + 1] = idxoff;
7558: mmdata->own[cp + 1] = idxown;
7559: }
7561: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)C), &mmdata->sf));
7562: PetscInt incoo_o;
7563: PetscCall(PetscIntCast(ncoo_o, &incoo_o));
7564: PetscCall(PetscSFSetGraphLayout(mmdata->sf, C->rmap, incoo_o /*nleaves*/, NULL /*ilocal*/, PETSC_OWN_POINTER, coo_i));
7565: PetscCall(PetscSFGetMultiSF(mmdata->sf, &msf));
7566: PetscCall(PetscSFGetGraph(msf, &ncoo2 /*nroots*/, NULL, NULL, NULL));
7567: ncoo = ncoo_d + ncoo_oown + ncoo2;
7568: PetscCall(PetscMalloc2(ncoo, &coo_i2, ncoo, &coo_j2));
7569: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_INT, coo_i, coo_i2 + ncoo_d + ncoo_oown)); /* put (i,j) of remote nonzeros at back */
7570: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_INT, coo_i, coo_i2 + ncoo_d + ncoo_oown));
7571: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_INT, coo_j, coo_j2 + ncoo_d + ncoo_oown));
7572: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_INT, coo_j, coo_j2 + ncoo_d + ncoo_oown));
7573: PetscCall(PetscFree2(coo_i, coo_j));
7574: /* allocate MPI send buffer to collect nonzero values to be sent to remote procs */
7575: PetscCall(PetscSFMalloc(mmdata->sf, mmdata->mtype, ncoo_o * sizeof(PetscScalar), (void **)&mmdata->coo_w));
7576: coo_i = coo_i2;
7577: coo_j = coo_j2;
7578: } else { /* no offproc values insertion */
7579: ncoo = ncoo_d;
7580: PetscCall(PetscMalloc2(ncoo, &coo_i, ncoo, &coo_j));
7582: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)C), &mmdata->sf));
7583: PetscCall(PetscSFSetGraph(mmdata->sf, 0, 0, NULL, PETSC_OWN_POINTER, NULL, PETSC_OWN_POINTER));
7584: PetscCall(PetscSFSetUp(mmdata->sf));
7585: }
7586: mmdata->hasoffproc = hasoffproc;
7588: /* gather (i,j) of nonzeros inserted locally */
7589: for (cp = 0, ncoo_d = 0; cp < mmdata->cp; cp++) {
7590: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7591: PetscInt *coi = coo_i + ncoo_d;
7592: PetscInt *coj = coo_j + ncoo_d;
7593: const PetscInt *jj = mm->j;
7594: const PetscInt *ii = mm->i;
7595: const PetscInt *cmap = cmapa[cp];
7596: const PetscInt *rmap = rmapa[cp];
7597: const PetscInt mr = mp[cp]->rmap->n;
7598: const PetscInt rs = C->rmap->rstart;
7599: const PetscInt re = C->rmap->rend;
7600: const PetscInt cs = C->cmap->rstart;
7602: if (mptmp[cp]) continue;
7603: if (rmapt[cp] == 1) { /* consecutive rows */
7604: /* fill coo_i */
7605: for (i = 0; i < mr; i++) {
7606: const PetscInt gr = i + rs;
7607: for (j = ii[i]; j < ii[i + 1]; j++) coi[j] = gr;
7608: }
7609: /* fill coo_j */
7610: if (!cmapt[cp]) { /* type-0, already global */
7611: PetscCall(PetscArraycpy(coj, jj, mm->nz));
7612: } else if (cmapt[cp] == 1) { /* type-1, local to global for consecutive columns of C */
7613: for (j = 0; j < mm->nz; j++) coj[j] = jj[j] + cs; /* lid + col start */
7614: } else { /* type-2, local to global for sparse columns */
7615: for (j = 0; j < mm->nz; j++) coj[j] = cmap[jj[j]];
7616: }
7617: ncoo_d += mm->nz;
7618: } else if (rmapt[cp] == 2) { /* sparse rows */
7619: for (i = 0; i < mr; i++) {
7620: const PetscInt *jj = mm->j + ii[i];
7621: const PetscInt gr = rmap[i];
7622: const PetscInt nz = ii[i + 1] - ii[i];
7623: if (gr >= rs && gr < re) { /* local rows */
7624: for (j = ii[i]; j < ii[i + 1]; j++) *coi++ = gr;
7625: if (!cmapt[cp]) { /* type-0, already global */
7626: for (j = 0; j < nz; j++) *coj++ = jj[j];
7627: } else if (cmapt[cp] == 1) { /* local to global for owned columns of C */
7628: for (j = 0; j < nz; j++) *coj++ = jj[j] + cs;
7629: } else { /* type-2, local to global for sparse columns */
7630: for (j = 0; j < nz; j++) *coj++ = cmap[jj[j]];
7631: }
7632: ncoo_d += nz;
7633: }
7634: }
7635: }
7636: }
7637: if (glob) PetscCall(ISRestoreIndices(glob, &globidx));
7638: PetscCall(ISDestroy(&glob));
7639: if (P_oth_l2g) PetscCall(ISLocalToGlobalMappingRestoreIndices(P_oth_l2g, &P_oth_idx));
7640: PetscCall(ISLocalToGlobalMappingDestroy(&P_oth_l2g));
7641: /* allocate an array to store all nonzeros (inserted locally or remotely) belonging to this proc */
7642: PetscCall(PetscSFMalloc(mmdata->sf, mmdata->mtype, ncoo * sizeof(PetscScalar), (void **)&mmdata->coo_v));
7644: /* preallocate with COO data */
7645: PetscCall(MatSetPreallocationCOO(C, ncoo, coo_i, coo_j));
7646: PetscCall(PetscFree2(coo_i, coo_j));
7647: PetscFunctionReturn(PETSC_SUCCESS);
7648: }
7650: PetscErrorCode MatProductSetFromOptions_MPIAIJBACKEND(Mat mat)
7651: {
7652: Mat_Product *product = mat->product;
7653: #if defined(PETSC_HAVE_DEVICE)
7654: PetscBool match = PETSC_FALSE;
7655: PetscBool usecpu = PETSC_FALSE;
7656: #else
7657: PetscBool match = PETSC_TRUE;
7658: #endif
7660: PetscFunctionBegin;
7661: MatCheckProduct(mat, 1);
7662: #if defined(PETSC_HAVE_DEVICE)
7663: if (!product->A->boundtocpu && !product->B->boundtocpu) PetscCall(PetscObjectTypeCompare((PetscObject)product->B, ((PetscObject)product->A)->type_name, &match));
7664: if (match) { /* we can always fallback to the CPU if requested */
7665: switch (product->type) {
7666: case MATPRODUCT_AB:
7667: if (product->api_user) {
7668: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatMatMult", "Mat");
7669: PetscCall(PetscOptionsBool("-matmatmult_backend_cpu", "Use CPU code", "MatMatMult", usecpu, &usecpu, NULL));
7670: PetscOptionsEnd();
7671: } else {
7672: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_AB", "Mat");
7673: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatMatMult", usecpu, &usecpu, NULL));
7674: PetscOptionsEnd();
7675: }
7676: break;
7677: case MATPRODUCT_AtB:
7678: if (product->api_user) {
7679: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatTransposeMatMult", "Mat");
7680: PetscCall(PetscOptionsBool("-mattransposematmult_backend_cpu", "Use CPU code", "MatTransposeMatMult", usecpu, &usecpu, NULL));
7681: PetscOptionsEnd();
7682: } else {
7683: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_AtB", "Mat");
7684: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatTransposeMatMult", usecpu, &usecpu, NULL));
7685: PetscOptionsEnd();
7686: }
7687: break;
7688: case MATPRODUCT_PtAP:
7689: if (product->api_user) {
7690: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatPtAP", "Mat");
7691: PetscCall(PetscOptionsBool("-matptap_backend_cpu", "Use CPU code", "MatPtAP", usecpu, &usecpu, NULL));
7692: PetscOptionsEnd();
7693: } else {
7694: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_PtAP", "Mat");
7695: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatPtAP", usecpu, &usecpu, NULL));
7696: PetscOptionsEnd();
7697: }
7698: break;
7699: default:
7700: break;
7701: }
7702: match = (PetscBool)!usecpu;
7703: }
7704: #endif
7705: if (match) {
7706: switch (product->type) {
7707: case MATPRODUCT_AB:
7708: case MATPRODUCT_AtB:
7709: case MATPRODUCT_PtAP:
7710: mat->ops->productsymbolic = MatProductSymbolic_MPIAIJBACKEND;
7711: break;
7712: default:
7713: break;
7714: }
7715: }
7716: /* fallback to MPIAIJ ops */
7717: if (!mat->ops->productsymbolic) PetscCall(MatProductSetFromOptions_MPIAIJ(mat));
7718: PetscFunctionReturn(PETSC_SUCCESS);
7719: }
7721: /*
7722: Produces a set of block column indices of the matrix row, one for each block represented in the original row
7724: n - the number of block indices in cc[]
7725: cc - the block indices (must be large enough to contain the indices)
7726: */
7727: static inline PetscErrorCode MatCollapseRow(Mat Amat, PetscInt row, PetscInt bs, PetscInt *n, PetscInt *cc)
7728: {
7729: PetscInt cnt = -1, nidx, j;
7730: const PetscInt *idx;
7732: PetscFunctionBegin;
7733: PetscCall(MatGetRow(Amat, row, &nidx, &idx, NULL));
7734: if (nidx) {
7735: cnt = 0;
7736: cc[cnt] = idx[0] / bs;
7737: for (j = 1; j < nidx; j++) {
7738: if (cc[cnt] < idx[j] / bs) cc[++cnt] = idx[j] / bs;
7739: }
7740: }
7741: PetscCall(MatRestoreRow(Amat, row, &nidx, &idx, NULL));
7742: *n = cnt + 1;
7743: PetscFunctionReturn(PETSC_SUCCESS);
7744: }
7746: /*
7747: Produces a set of block column indices of the matrix block row, one for each block represented in the original set of rows
7749: ncollapsed - the number of block indices
7750: collapsed - the block indices (must be large enough to contain the indices)
7751: */
7752: static inline PetscErrorCode MatCollapseRows(Mat Amat, PetscInt start, PetscInt bs, PetscInt *w0, PetscInt *w1, PetscInt *w2, PetscInt *ncollapsed, PetscInt **collapsed)
7753: {
7754: PetscInt i, nprev, *cprev = w0, ncur = 0, *ccur = w1, *merged = w2, *cprevtmp;
7756: PetscFunctionBegin;
7757: PetscCall(MatCollapseRow(Amat, start, bs, &nprev, cprev));
7758: for (i = start + 1; i < start + bs; i++) {
7759: PetscCall(MatCollapseRow(Amat, i, bs, &ncur, ccur));
7760: PetscCall(PetscMergeIntArray(nprev, cprev, ncur, ccur, &nprev, &merged));
7761: cprevtmp = cprev;
7762: cprev = merged;
7763: merged = cprevtmp;
7764: }
7765: *ncollapsed = nprev;
7766: if (collapsed) *collapsed = cprev;
7767: PetscFunctionReturn(PETSC_SUCCESS);
7768: }
7770: /*
7771: MatCreateGraph_Simple_AIJ - create simple scalar matrix (graph) from potentially blocked matrix
7773: Input Parameter:
7774: . Amat - matrix
7775: - symmetrize - make the result symmetric
7776: + scale - scale with diagonal
7778: Output Parameter:
7779: . a_Gmat - output scalar graph >= 0
7781: */
7782: PETSC_INTERN PetscErrorCode MatCreateGraph_Simple_AIJ(Mat Amat, PetscBool symmetrize, PetscBool scale, PetscReal filter, PetscInt index_size, PetscInt index[], Mat *a_Gmat)
7783: {
7784: PetscInt Istart, Iend, Ii, jj, kk, ncols, nloc, NN, MM, bs;
7785: MPI_Comm comm;
7786: Mat Gmat;
7787: PetscBool ismpiaij, isseqaij;
7788: Mat a, b, c;
7789: MatType jtype;
7791: PetscFunctionBegin;
7792: PetscCall(PetscObjectGetComm((PetscObject)Amat, &comm));
7793: PetscCall(MatGetOwnershipRange(Amat, &Istart, &Iend));
7794: PetscCall(MatGetSize(Amat, &MM, &NN));
7795: PetscCall(MatGetBlockSize(Amat, &bs));
7796: nloc = (Iend - Istart) / bs;
7798: PetscCall(PetscObjectBaseTypeCompare((PetscObject)Amat, MATSEQAIJ, &isseqaij));
7799: PetscCall(PetscObjectBaseTypeCompare((PetscObject)Amat, MATMPIAIJ, &ismpiaij));
7800: PetscCheck(isseqaij || ismpiaij, comm, PETSC_ERR_USER, "Require (MPI)AIJ matrix type");
7802: /* TODO GPU: these calls are potentially expensive if matrices are large and we want to use the GPU */
7803: /* A solution consists in providing a new API, MatAIJGetCollapsedAIJ, and each class can provide a fast
7804: implementation */
7805: if (bs > 1) {
7806: PetscCall(MatGetType(Amat, &jtype));
7807: PetscCall(MatCreate(comm, &Gmat));
7808: PetscCall(MatSetType(Gmat, jtype));
7809: PetscCall(MatSetSizes(Gmat, nloc, nloc, PETSC_DETERMINE, PETSC_DETERMINE));
7810: PetscCall(MatSetBlockSizes(Gmat, 1, 1));
7811: if (isseqaij || ((Mat_MPIAIJ *)Amat->data)->garray) {
7812: PetscInt *d_nnz, *o_nnz;
7813: MatScalar *aa, val, *AA;
7814: PetscInt *aj, *ai, *AJ, nc, nmax = 0;
7816: if (isseqaij) {
7817: a = Amat;
7818: b = NULL;
7819: } else {
7820: Mat_MPIAIJ *d = (Mat_MPIAIJ *)Amat->data;
7821: a = d->A;
7822: b = d->B;
7823: }
7824: PetscCall(PetscInfo(Amat, "New bs>1 Graph. nloc=%" PetscInt_FMT "\n", nloc));
7825: PetscCall(PetscMalloc2(nloc, &d_nnz, (isseqaij ? 0 : nloc), &o_nnz));
7826: for (c = a, kk = 0; c && kk < 2; c = b, kk++) {
7827: PetscInt *nnz = (c == a) ? d_nnz : o_nnz;
7828: const PetscInt *cols1, *cols2;
7830: for (PetscInt brow = 0, nc1, nc2, ok = 1; brow < nloc * bs; brow += bs) { // block rows
7831: PetscCall(MatGetRow(c, brow, &nc2, &cols2, NULL));
7832: nnz[brow / bs] = nc2 / bs;
7833: if (nc2 % bs) ok = 0;
7834: if (nnz[brow / bs] > nmax) nmax = nnz[brow / bs];
7835: for (PetscInt ii = 1; ii < bs; ii++) { // check for non-dense blocks
7836: PetscCall(MatGetRow(c, brow + ii, &nc1, &cols1, NULL));
7837: if (nc1 != nc2) ok = 0;
7838: else {
7839: for (PetscInt jj = 0; jj < nc1 && ok == 1; jj++) {
7840: if (cols1[jj] != cols2[jj]) ok = 0;
7841: if (cols1[jj] % bs != jj % bs) ok = 0;
7842: }
7843: }
7844: PetscCall(MatRestoreRow(c, brow + ii, &nc1, &cols1, NULL));
7845: }
7846: PetscCall(MatRestoreRow(c, brow, &nc2, &cols2, NULL));
7847: if (!ok) {
7848: PetscCall(PetscFree2(d_nnz, o_nnz));
7849: PetscCall(PetscInfo(Amat, "Found sparse blocks - revert to slow method\n"));
7850: goto old_bs;
7851: }
7852: }
7853: }
7854: PetscCall(MatSeqAIJSetPreallocation(Gmat, 0, d_nnz));
7855: PetscCall(MatMPIAIJSetPreallocation(Gmat, 0, d_nnz, 0, o_nnz));
7856: PetscCall(PetscFree2(d_nnz, o_nnz));
7857: PetscCall(PetscMalloc2(nmax, &AA, nmax, &AJ));
7858: // diag
7859: for (PetscInt brow = 0, n, grow; brow < nloc * bs; brow += bs) { // block rows
7860: Mat_SeqAIJ *aseq = (Mat_SeqAIJ *)a->data;
7862: ai = aseq->i;
7863: n = ai[brow + 1] - ai[brow];
7864: aj = aseq->j + ai[brow];
7865: for (PetscInt k = 0; k < n; k += bs) { // block columns
7866: AJ[k / bs] = aj[k] / bs + Istart / bs; // diag starts at (Istart,Istart)
7867: val = 0;
7868: if (index_size == 0) {
7869: for (PetscInt ii = 0; ii < bs; ii++) { // rows in block
7870: aa = aseq->a + ai[brow + ii] + k;
7871: for (PetscInt jj = 0; jj < bs; jj++) { // columns in block
7872: val += PetscAbs(PetscRealPart(aa[jj])); // a sort of norm
7873: }
7874: }
7875: } else { // use (index,index) value if provided
7876: for (PetscInt iii = 0; iii < index_size; iii++) { // rows in block
7877: PetscInt ii = index[iii];
7878: aa = aseq->a + ai[brow + ii] + k;
7879: for (PetscInt jjj = 0; jjj < index_size; jjj++) { // columns in block
7880: PetscInt jj = index[jjj];
7881: val += PetscAbs(PetscRealPart(aa[jj]));
7882: }
7883: }
7884: }
7885: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs (%d) >= nmax (%d)", (int)(k / bs), (int)nmax);
7886: AA[k / bs] = val;
7887: }
7888: grow = Istart / bs + brow / bs;
7889: PetscCall(MatSetValues(Gmat, 1, &grow, n / bs, AJ, AA, ADD_VALUES));
7890: }
7891: // off-diag
7892: if (ismpiaij) {
7893: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)Amat->data;
7894: const PetscScalar *vals;
7895: const PetscInt *cols, *garray = aij->garray;
7897: PetscCheck(garray, PETSC_COMM_SELF, PETSC_ERR_USER, "No garray ?");
7898: for (PetscInt brow = 0, grow; brow < nloc * bs; brow += bs) { // block rows
7899: PetscCall(MatGetRow(b, brow, &ncols, &cols, NULL));
7900: for (PetscInt k = 0, cidx = 0; k < ncols; k += bs, cidx++) {
7901: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs >= nmax");
7902: AA[k / bs] = 0;
7903: AJ[cidx] = garray[cols[k]] / bs;
7904: }
7905: nc = ncols / bs;
7906: PetscCall(MatRestoreRow(b, brow, &ncols, &cols, NULL));
7907: if (index_size == 0) {
7908: for (PetscInt ii = 0; ii < bs; ii++) { // rows in block
7909: PetscCall(MatGetRow(b, brow + ii, &ncols, &cols, &vals));
7910: for (PetscInt k = 0; k < ncols; k += bs) {
7911: for (PetscInt jj = 0; jj < bs; jj++) { // cols in block
7912: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs (%d) >= nmax (%d)", (int)(k / bs), (int)nmax);
7913: AA[k / bs] += PetscAbs(PetscRealPart(vals[k + jj]));
7914: }
7915: }
7916: PetscCall(MatRestoreRow(b, brow + ii, &ncols, &cols, &vals));
7917: }
7918: } else { // use (index,index) value if provided
7919: for (PetscInt iii = 0; iii < index_size; iii++) { // rows in block
7920: PetscInt ii = index[iii];
7921: PetscCall(MatGetRow(b, brow + ii, &ncols, &cols, &vals));
7922: for (PetscInt k = 0; k < ncols; k += bs) {
7923: for (PetscInt jjj = 0; jjj < index_size; jjj++) { // cols in block
7924: PetscInt jj = index[jjj];
7925: AA[k / bs] += PetscAbs(PetscRealPart(vals[k + jj]));
7926: }
7927: }
7928: PetscCall(MatRestoreRow(b, brow + ii, &ncols, &cols, &vals));
7929: }
7930: }
7931: grow = Istart / bs + brow / bs;
7932: PetscCall(MatSetValues(Gmat, 1, &grow, nc, AJ, AA, ADD_VALUES));
7933: }
7934: }
7935: PetscCall(MatAssemblyBegin(Gmat, MAT_FINAL_ASSEMBLY));
7936: PetscCall(MatAssemblyEnd(Gmat, MAT_FINAL_ASSEMBLY));
7937: PetscCall(PetscFree2(AA, AJ));
7938: } else {
7939: const PetscScalar *vals;
7940: const PetscInt *idx;
7941: PetscInt *d_nnz, *o_nnz, *w0, *w1, *w2;
7942: old_bs:
7943: /*
7944: Determine the preallocation needed for the scalar matrix derived from the vector matrix.
7945: */
7946: PetscCall(PetscInfo(Amat, "OLD bs>1 CreateGraph\n"));
7947: PetscCall(PetscMalloc2(nloc, &d_nnz, (isseqaij ? 0 : nloc), &o_nnz));
7948: if (isseqaij) {
7949: PetscInt max_d_nnz;
7951: /*
7952: Determine exact preallocation count for (sequential) scalar matrix
7953: */
7954: PetscCall(MatSeqAIJGetMaxRowNonzeros(Amat, &max_d_nnz));
7955: max_d_nnz = PetscMin(nloc, bs * max_d_nnz);
7956: PetscCall(PetscMalloc3(max_d_nnz, &w0, max_d_nnz, &w1, max_d_nnz, &w2));
7957: for (Ii = 0, jj = 0; Ii < Iend; Ii += bs, jj++) PetscCall(MatCollapseRows(Amat, Ii, bs, w0, w1, w2, &d_nnz[jj], NULL));
7958: PetscCall(PetscFree3(w0, w1, w2));
7959: } else if (ismpiaij) {
7960: Mat Daij, Oaij;
7961: const PetscInt *garray;
7962: PetscInt max_d_nnz;
7964: PetscCall(MatMPIAIJGetSeqAIJ(Amat, &Daij, &Oaij, &garray));
7965: /*
7966: Determine exact preallocation count for diagonal block portion of scalar matrix
7967: */
7968: PetscCall(MatSeqAIJGetMaxRowNonzeros(Daij, &max_d_nnz));
7969: max_d_nnz = PetscMin(nloc, bs * max_d_nnz);
7970: PetscCall(PetscMalloc3(max_d_nnz, &w0, max_d_nnz, &w1, max_d_nnz, &w2));
7971: for (Ii = 0, jj = 0; Ii < Iend - Istart; Ii += bs, jj++) PetscCall(MatCollapseRows(Daij, Ii, bs, w0, w1, w2, &d_nnz[jj], NULL));
7972: PetscCall(PetscFree3(w0, w1, w2));
7973: /*
7974: Over estimate (usually grossly over), preallocation count for off-diagonal portion of scalar matrix
7975: */
7976: for (Ii = 0, jj = 0; Ii < Iend - Istart; Ii += bs, jj++) {
7977: o_nnz[jj] = 0;
7978: for (kk = 0; kk < bs; kk++) { /* rows that get collapsed to a single row */
7979: PetscCall(MatGetRow(Oaij, Ii + kk, &ncols, NULL, NULL));
7980: o_nnz[jj] += ncols;
7981: PetscCall(MatRestoreRow(Oaij, Ii + kk, &ncols, NULL, NULL));
7982: }
7983: if (o_nnz[jj] > (NN / bs - nloc)) o_nnz[jj] = NN / bs - nloc;
7984: }
7985: } else SETERRQ(comm, PETSC_ERR_USER, "Require AIJ matrix type");
7986: /* get scalar copy (norms) of matrix */
7987: PetscCall(MatSeqAIJSetPreallocation(Gmat, 0, d_nnz));
7988: PetscCall(MatMPIAIJSetPreallocation(Gmat, 0, d_nnz, 0, o_nnz));
7989: PetscCall(PetscFree2(d_nnz, o_nnz));
7990: for (Ii = Istart; Ii < Iend; Ii++) {
7991: PetscInt dest_row = Ii / bs;
7993: PetscCall(MatGetRow(Amat, Ii, &ncols, &idx, &vals));
7994: for (jj = 0; jj < ncols; jj++) {
7995: PetscInt dest_col = idx[jj] / bs;
7996: PetscScalar sv = PetscAbs(PetscRealPart(vals[jj]));
7998: PetscCall(MatSetValues(Gmat, 1, &dest_row, 1, &dest_col, &sv, ADD_VALUES));
7999: }
8000: PetscCall(MatRestoreRow(Amat, Ii, &ncols, &idx, &vals));
8001: }
8002: PetscCall(MatAssemblyBegin(Gmat, MAT_FINAL_ASSEMBLY));
8003: PetscCall(MatAssemblyEnd(Gmat, MAT_FINAL_ASSEMBLY));
8004: }
8005: } else {
8006: if (symmetrize || filter >= 0 || scale) PetscCall(MatDuplicate(Amat, MAT_COPY_VALUES, &Gmat));
8007: else {
8008: Gmat = Amat;
8009: PetscCall(PetscObjectReference((PetscObject)Gmat));
8010: }
8011: if (isseqaij) {
8012: a = Gmat;
8013: b = NULL;
8014: } else {
8015: Mat_MPIAIJ *d = (Mat_MPIAIJ *)Gmat->data;
8016: a = d->A;
8017: b = d->B;
8018: }
8019: if (filter >= 0 || scale) {
8020: /* take absolute value of each entry */
8021: for (c = a, kk = 0; c && kk < 2; c = b, kk++) {
8022: MatInfo info;
8023: PetscScalar *avals;
8025: PetscCall(MatGetInfo(c, MAT_LOCAL, &info));
8026: PetscCall(MatSeqAIJGetArray(c, &avals));
8027: for (int jj = 0; jj < info.nz_used; jj++) avals[jj] = PetscAbsScalar(avals[jj]);
8028: PetscCall(MatSeqAIJRestoreArray(c, &avals));
8029: }
8030: }
8031: }
8032: if (symmetrize) {
8033: PetscBool isset, issym;
8035: PetscCall(MatIsSymmetricKnown(Amat, &isset, &issym));
8036: if (!isset || !issym) {
8037: Mat matTrans;
8039: PetscCall(MatTranspose(Gmat, MAT_INITIAL_MATRIX, &matTrans));
8040: PetscCall(MatAXPY(Gmat, 1.0, matTrans, Gmat->structurally_symmetric == PETSC_BOOL3_TRUE ? SAME_NONZERO_PATTERN : DIFFERENT_NONZERO_PATTERN));
8041: PetscCall(MatDestroy(&matTrans));
8042: }
8043: PetscCall(MatSetOption(Gmat, MAT_SYMMETRIC, PETSC_TRUE));
8044: } else if (Amat != Gmat) PetscCall(MatPropagateSymmetryOptions(Amat, Gmat));
8045: if (scale) {
8046: /* scale c for all diagonal values = 1 or -1 */
8047: Vec diag;
8049: PetscCall(MatCreateVecs(Gmat, &diag, NULL));
8050: PetscCall(MatGetDiagonal(Gmat, diag));
8051: PetscCall(VecReciprocal(diag));
8052: PetscCall(VecSqrtAbs(diag));
8053: PetscCall(MatDiagonalScale(Gmat, diag, diag));
8054: PetscCall(VecDestroy(&diag));
8055: }
8056: PetscCall(MatViewFromOptions(Gmat, NULL, "-mat_graph_view"));
8057: if (filter >= 0) {
8058: PetscCall(MatFilter(Gmat, filter, PETSC_TRUE, PETSC_TRUE));
8059: PetscCall(MatViewFromOptions(Gmat, NULL, "-mat_filter_graph_view"));
8060: }
8061: *a_Gmat = Gmat;
8062: PetscFunctionReturn(PETSC_SUCCESS);
8063: }
8065: /*
8066: Special version for direct calls from Fortran
8067: */
8069: /* Change these macros so can be used in void function */
8070: /* Identical to PetscCallVoid, except it assigns to *_ierr */
8071: #undef PetscCall
8072: #define PetscCall(...) \
8073: do { \
8074: PetscErrorCode ierr_msv_mpiaij = __VA_ARGS__; \
8075: if (PetscUnlikely(ierr_msv_mpiaij)) { \
8076: *_ierr = PetscError(PETSC_COMM_SELF, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr_msv_mpiaij, PETSC_ERROR_REPEAT, " "); \
8077: return; \
8078: } \
8079: } while (0)
8081: #undef SETERRQ
8082: #define SETERRQ(comm, ierr, ...) \
8083: do { \
8084: *_ierr = PetscError(comm, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr, PETSC_ERROR_INITIAL, __VA_ARGS__); \
8085: return; \
8086: } while (0)
8088: #if defined(PETSC_HAVE_FORTRAN_CAPS)
8089: #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ
8090: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
8091: #define matsetvaluesmpiaij_ matsetvaluesmpiaij
8092: #else
8093: #endif
8094: PETSC_EXTERN void matsetvaluesmpiaij_(Mat *mmat, PetscInt *mm, const PetscInt im[], PetscInt *mn, const PetscInt in[], const PetscScalar v[], InsertMode *maddv, PetscErrorCode *_ierr)
8095: {
8096: Mat mat = *mmat;
8097: PetscInt m = *mm, n = *mn;
8098: InsertMode addv = *maddv;
8099: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
8100: PetscScalar value;
8102: MatCheckPreallocated(mat, 1);
8103: if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv;
8104: else PetscCheck(mat->insertmode == addv, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Cannot mix add values and insert values");
8105: {
8106: PetscInt i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend;
8107: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend, row, col;
8108: PetscBool roworiented = aij->roworiented;
8110: /* Some Variables required in the macro */
8111: Mat A = aij->A;
8112: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
8113: PetscInt *aimax = a->imax, *ai = a->i, *ailen = a->ilen, *aj = a->j;
8114: MatScalar *aa;
8115: PetscBool ignorezeroentries = ((a->ignorezeroentries && (addv == ADD_VALUES)) ? PETSC_TRUE : PETSC_FALSE);
8116: Mat B = aij->B;
8117: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
8118: PetscInt *bimax = b->imax, *bi = b->i, *bilen = b->ilen, *bj = b->j, bm = aij->B->rmap->n, am = aij->A->rmap->n;
8119: MatScalar *ba;
8120: /* This variable below is only for the PETSC_HAVE_VIENNACL or PETSC_HAVE_CUDA cases, but we define it in all cases because we
8121: * cannot use "#if defined" inside a macro. */
8122: PETSC_UNUSED PetscBool inserted = PETSC_FALSE;
8124: PetscInt *rp1, *rp2, ii, nrow1, nrow2, _i, rmax1, rmax2, N, low1, high1, low2, high2, t, lastcol1, lastcol2;
8125: PetscInt nonew = a->nonew;
8126: MatScalar *ap1, *ap2;
8128: PetscFunctionBegin;
8129: PetscCall(MatSeqAIJGetArray(A, &aa));
8130: PetscCall(MatSeqAIJGetArray(B, &ba));
8131: for (i = 0; i < m; i++) {
8132: if (im[i] < 0) continue;
8133: 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);
8134: if (im[i] >= rstart && im[i] < rend) {
8135: row = im[i] - rstart;
8136: lastcol1 = -1;
8137: rp1 = aj + ai[row];
8138: ap1 = aa + ai[row];
8139: rmax1 = aimax[row];
8140: nrow1 = ailen[row];
8141: low1 = 0;
8142: high1 = nrow1;
8143: lastcol2 = -1;
8144: rp2 = bj + bi[row];
8145: ap2 = ba + bi[row];
8146: rmax2 = bimax[row];
8147: nrow2 = bilen[row];
8148: low2 = 0;
8149: high2 = nrow2;
8151: for (j = 0; j < n; j++) {
8152: if (roworiented) value = v[i * n + j];
8153: else value = v[i + j * m];
8154: if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && im[i] != in[j]) continue;
8155: if (in[j] >= cstart && in[j] < cend) {
8156: col = in[j] - cstart;
8157: MatSetValues_SeqAIJ_A_Private(row, col, value, addv, im[i], in[j]);
8158: } else if (in[j] < 0) continue;
8159: else if (PetscUnlikelyDebug(in[j] >= mat->cmap->N)) {
8160: SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large: col %" PetscInt_FMT " max %" PetscInt_FMT, in[j], mat->cmap->N - 1);
8161: } else {
8162: if (mat->was_assembled) {
8163: if (!aij->colmap) PetscCall(MatCreateColmap_MPIAIJ_Private(mat));
8164: #if defined(PETSC_USE_CTABLE)
8165: PetscCall(PetscHMapIGetWithDefault(aij->colmap, in[j] + 1, 0, &col));
8166: col--;
8167: #else
8168: col = aij->colmap[in[j]] - 1;
8169: #endif
8170: if (col < 0 && !((Mat_SeqAIJ *)aij->A->data)->nonew) {
8171: PetscCall(MatDisAssemble_MPIAIJ(mat, PETSC_FALSE));
8172: col = in[j];
8173: /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
8174: B = aij->B;
8175: b = (Mat_SeqAIJ *)B->data;
8176: bimax = b->imax;
8177: bi = b->i;
8178: bilen = b->ilen;
8179: bj = b->j;
8180: rp2 = bj + bi[row];
8181: ap2 = ba + bi[row];
8182: rmax2 = bimax[row];
8183: nrow2 = bilen[row];
8184: low2 = 0;
8185: high2 = nrow2;
8186: bm = aij->B->rmap->n;
8187: ba = b->a;
8188: inserted = PETSC_FALSE;
8189: }
8190: } else col = in[j];
8191: MatSetValues_SeqAIJ_B_Private(row, col, value, addv, im[i], in[j]);
8192: }
8193: }
8194: } else if (!aij->donotstash) {
8195: if (roworiented) {
8196: PetscCall(MatStashValuesRow_Private(&mat->stash, im[i], n, in, v + i * n, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
8197: } else {
8198: PetscCall(MatStashValuesCol_Private(&mat->stash, im[i], n, in, v + i, m, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
8199: }
8200: }
8201: }
8202: PetscCall(MatSeqAIJRestoreArray(A, &aa));
8203: PetscCall(MatSeqAIJRestoreArray(B, &ba));
8204: }
8205: PetscFunctionReturnVoid();
8206: }
8208: /* Undefining these here since they were redefined from their original definition above! No
8209: * other PETSc functions should be defined past this point, as it is impossible to recover the
8210: * original definitions */
8211: #undef PetscCall
8212: #undef SETERRQ