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: /* defines MatSetValues_MPI_Hash(), MatAssemblyBegin_MPI_Hash(), and MatAssemblyEnd_MPI_Hash() */
10: #define TYPE AIJ
11: #define TYPE_AIJ
12: #include "../src/mat/impls/aij/mpi/mpihashmat.h"
13: #undef TYPE
14: #undef TYPE_AIJ
16: static PetscErrorCode MatReset_MPIAIJ(Mat mat)
17: {
18: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
20: PetscFunctionBegin;
21: PetscCall(PetscLogObjectState((PetscObject)mat, "Rows=%" PetscInt_FMT ", Cols=%" PetscInt_FMT, mat->rmap->N, mat->cmap->N));
22: PetscCall(MatStashDestroy_Private(&mat->stash));
23: PetscCall(VecDestroy(&aij->diag));
24: PetscCall(MatDestroy(&aij->A));
25: PetscCall(MatDestroy(&aij->B));
26: #if PetscDefined(USE_CTABLE)
27: PetscCall(PetscHMapIDestroy(&aij->colmap));
28: #else
29: PetscCall(PetscFree(aij->colmap));
30: #endif
31: PetscCall(PetscFree(aij->garray));
32: PetscCall(VecDestroy(&aij->lvec));
33: PetscCall(VecScatterDestroy(&aij->Mvctx));
34: PetscCall(PetscFree2(aij->rowvalues, aij->rowindices));
35: PetscCall(PetscFree(aij->ld));
36: PetscFunctionReturn(PETSC_SUCCESS);
37: }
39: static PetscErrorCode MatResetHash_MPIAIJ(Mat mat)
40: {
41: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
42: /* Save the nonzero states of the component matrices because those are what are used to determine
43: the nonzero state of mat */
44: PetscObjectState Astate = aij->A->nonzerostate, Bstate = aij->B->nonzerostate;
46: PetscFunctionBegin;
47: PetscCall(MatReset_MPIAIJ(mat));
48: PetscCall(MatSetUp_MPI_Hash(mat));
49: aij->A->nonzerostate = ++Astate, aij->B->nonzerostate = ++Bstate;
50: PetscFunctionReturn(PETSC_SUCCESS);
51: }
53: PetscErrorCode MatDestroy_MPIAIJ(Mat mat)
54: {
55: PetscFunctionBegin;
56: PetscCall(MatReset_MPIAIJ(mat));
58: PetscCall(PetscFree(mat->data));
60: /* may be created by MatCreateMPIAIJSumSeqAIJSymbolic */
61: PetscCall(PetscObjectCompose((PetscObject)mat, "MatMergeSeqsToMPI", NULL));
63: PetscCall(PetscObjectChangeTypeName((PetscObject)mat, NULL));
64: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatStoreValues_C", NULL));
65: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatRetrieveValues_C", NULL));
66: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatGetMultPetscSF_C", NULL));
67: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatIsTranspose_C", NULL));
68: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMPIAIJSetPreallocation_C", NULL));
69: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatResetPreallocation_C", NULL));
70: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatResetHash_C", NULL));
71: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMPIAIJSetPreallocationCSR_C", NULL));
72: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDiagonalScaleLocal_C", NULL));
73: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpibaij_C", NULL));
74: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpisbaij_C", NULL));
75: #if PetscDefined(HAVE_CUDA)
76: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijcusparse_C", NULL));
77: #endif
78: #if PetscDefined(HAVE_HIP)
79: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijhipsparse_C", NULL));
80: #endif
81: #if PetscDefined(HAVE_KOKKOS_KERNELS)
82: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijkokkos_C", NULL));
83: #endif
84: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpidense_C", NULL));
85: #if PetscDefined(HAVE_ELEMENTAL)
86: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_elemental_C", NULL));
87: #endif
88: #if PetscDefined(HAVE_SCALAPACK) && (PetscDefined(USE_REAL_SINGLE) || PetscDefined(USE_REAL_DOUBLE))
89: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_scalapack_C", NULL));
90: #endif
91: #if PetscDefined(HAVE_HYPRE)
92: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_hypre_C", NULL));
93: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_transpose_mpiaij_mpiaij_C", NULL));
94: #endif
95: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_is_C", NULL));
96: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_is_mpiaij_C", NULL));
97: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpiaij_mpiaij_C", NULL));
98: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMPIAIJSetUseScalableIncreaseOverlap_C", NULL));
99: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijperm_C", NULL));
100: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijsell_C", NULL));
101: #if PetscDefined(HAVE_MKL_SPARSE)
102: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijmkl_C", NULL));
103: #endif
104: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijcrl_C", NULL));
105: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_is_C", NULL));
106: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpisell_C", NULL));
107: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatSetPreallocationCOO_C", NULL));
108: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatSetValuesCOO_C", NULL));
109: PetscFunctionReturn(PETSC_SUCCESS);
110: }
112: static PetscErrorCode MatGetRowIJ_MPIAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *m, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done)
113: {
114: Mat B;
116: PetscFunctionBegin;
117: PetscCall(MatMPIAIJGetLocalMat(A, MAT_INITIAL_MATRIX, &B));
118: PetscCall(PetscObjectCompose((PetscObject)A, "MatGetRowIJ_MPIAIJ", (PetscObject)B));
119: PetscCall(MatGetRowIJ(B, oshift, symmetric, inodecompressed, m, ia, ja, done));
120: PetscCall(MatDestroy(&B));
121: PetscFunctionReturn(PETSC_SUCCESS);
122: }
124: static PetscErrorCode MatRestoreRowIJ_MPIAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *m, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done)
125: {
126: Mat B;
128: PetscFunctionBegin;
129: PetscCall(PetscObjectQuery((PetscObject)A, "MatGetRowIJ_MPIAIJ", (PetscObject *)&B));
130: PetscCall(MatRestoreRowIJ(B, oshift, symmetric, inodecompressed, m, ia, ja, done));
131: PetscCall(PetscObjectCompose((PetscObject)A, "MatGetRowIJ_MPIAIJ", NULL));
132: PetscFunctionReturn(PETSC_SUCCESS);
133: }
135: /*MC
136: MATAIJCRL - MATAIJCRL = "aijcrl" - A matrix type to be used for sparse matrices.
138: This matrix type is identical to `MATSEQAIJCRL` when constructed with a single process communicator,
139: and `MATMPIAIJCRL` otherwise. As a result, for single process communicators,
140: `MatSeqAIJSetPreallocation()` is supported, and similarly `MatMPIAIJSetPreallocation()` is supported
141: for communicators controlling multiple processes. It is recommended that you call both of
142: the above preallocation routines for simplicity.
144: Options Database Key:
145: . -mat_type aijcrl - sets the matrix type to `MATMPIAIJCRL` during a call to `MatSetFromOptions()`
147: Level: beginner
149: .seealso: [](ch_matrices), `Mat`, `MatCreateMPIAIJCRL`, `MATSEQAIJCRL`, `MATMPIAIJCRL`, `MATSEQAIJ`, `MATMPIAIJ`, `MATAIJ`
150: M*/
152: static PetscErrorCode MatBindToCPU_MPIAIJ(Mat A, PetscBool flg)
153: {
154: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
156: PetscFunctionBegin;
157: #if PetscDefined(HAVE_CUDA) || PetscDefined(HAVE_HIP) || PetscDefined(HAVE_VIENNACL)
158: A->boundtocpu = flg;
159: #endif
160: if (a->A) PetscCall(MatBindToCPU(a->A, flg));
161: if (a->B) PetscCall(MatBindToCPU(a->B, flg));
163: /* In addition to binding the diagonal and off-diagonal matrices, bind the local vectors used for matrix-vector products.
164: * This maybe seems a little odd for a MatBindToCPU() call to do, but it makes no sense for the binding of these vectors
165: * to differ from the parent matrix. */
166: if (a->lvec) PetscCall(VecBindToCPU(a->lvec, flg));
167: if (a->diag) PetscCall(VecBindToCPU(a->diag, flg));
168: PetscFunctionReturn(PETSC_SUCCESS);
169: }
171: static PetscErrorCode MatSetBlockSizes_MPIAIJ(Mat M, PetscInt rbs, PetscInt cbs)
172: {
173: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)M->data;
175: PetscFunctionBegin;
176: if (mat->A) {
177: PetscCall(MatSetBlockSizes(mat->A, rbs, cbs));
178: PetscCall(MatSetBlockSizes(mat->B, rbs, 1));
179: }
180: PetscFunctionReturn(PETSC_SUCCESS);
181: }
183: static PetscErrorCode MatFindNonzeroRows_MPIAIJ(Mat M, IS *keptrows)
184: {
185: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)M->data;
186: Mat_SeqAIJ *a = (Mat_SeqAIJ *)mat->A->data;
187: Mat_SeqAIJ *b = (Mat_SeqAIJ *)mat->B->data;
188: const PetscInt *ia, *ib;
189: const MatScalar *aa, *bb, *aav, *bav;
190: PetscInt na, nb, i, j, *rows, cnt = 0, n0rows;
191: PetscInt m = M->rmap->n, rstart = M->rmap->rstart;
193: PetscFunctionBegin;
194: *keptrows = NULL;
196: ia = a->i;
197: ib = b->i;
198: PetscCall(MatSeqAIJGetArrayRead(mat->A, &aav));
199: PetscCall(MatSeqAIJGetArrayRead(mat->B, &bav));
200: for (i = 0; i < m; i++) {
201: na = ia[i + 1] - ia[i];
202: nb = ib[i + 1] - ib[i];
203: if (!na && !nb) {
204: cnt++;
205: goto ok1;
206: }
207: aa = aav + ia[i];
208: for (j = 0; j < na; j++) {
209: if (aa[j] != 0.0) goto ok1;
210: }
211: bb = PetscSafePointerPlusOffset(bav, ib[i]);
212: for (j = 0; j < nb; j++) {
213: if (bb[j] != 0.0) goto ok1;
214: }
215: cnt++;
216: ok1:;
217: }
218: PetscCallMPI(MPIU_Allreduce(&cnt, &n0rows, 1, MPIU_INT, MPI_SUM, PetscObjectComm((PetscObject)M)));
219: if (!n0rows) {
220: PetscCall(MatSeqAIJRestoreArrayRead(mat->A, &aav));
221: PetscCall(MatSeqAIJRestoreArrayRead(mat->B, &bav));
222: PetscFunctionReturn(PETSC_SUCCESS);
223: }
224: PetscCall(PetscMalloc1(M->rmap->n - cnt, &rows));
225: cnt = 0;
226: for (i = 0; i < m; i++) {
227: na = ia[i + 1] - ia[i];
228: nb = ib[i + 1] - ib[i];
229: if (!na && !nb) continue;
230: aa = aav + ia[i];
231: for (j = 0; j < na; j++) {
232: if (aa[j] != 0.0) {
233: rows[cnt++] = rstart + i;
234: goto ok2;
235: }
236: }
237: bb = PetscSafePointerPlusOffset(bav, ib[i]);
238: for (j = 0; j < nb; j++) {
239: if (bb[j] != 0.0) {
240: rows[cnt++] = rstart + i;
241: goto ok2;
242: }
243: }
244: ok2:;
245: }
246: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)M), cnt, rows, PETSC_OWN_POINTER, keptrows));
247: PetscCall(MatSeqAIJRestoreArrayRead(mat->A, &aav));
248: PetscCall(MatSeqAIJRestoreArrayRead(mat->B, &bav));
249: PetscFunctionReturn(PETSC_SUCCESS);
250: }
252: static PetscErrorCode MatDiagonalSet_MPIAIJ(Mat Y, Vec D, InsertMode is)
253: {
254: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)Y->data;
255: PetscBool cong;
257: PetscFunctionBegin;
258: PetscCall(MatHasCongruentLayouts(Y, &cong));
259: if (Y->assembled && cong) PetscCall(MatDiagonalSet(aij->A, D, is));
260: else PetscCall(MatDiagonalSet_Default(Y, D, is));
261: PetscFunctionReturn(PETSC_SUCCESS);
262: }
264: static PetscErrorCode MatFindZeroDiagonals_MPIAIJ(Mat M, IS *zrows)
265: {
266: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)M->data;
267: PetscInt i, rstart, nrows, *rows;
269: PetscFunctionBegin;
270: *zrows = NULL;
271: PetscCall(MatFindZeroDiagonals_SeqAIJ_Private(aij->A, &nrows, &rows));
272: PetscCall(MatGetOwnershipRange(M, &rstart, NULL));
273: for (i = 0; i < nrows; i++) rows[i] += rstart;
274: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)M), nrows, rows, PETSC_OWN_POINTER, zrows));
275: PetscFunctionReturn(PETSC_SUCCESS);
276: }
278: static PetscErrorCode MatGetColumnReductions_MPIAIJ(Mat A, PetscInt type, PetscReal *reductions)
279: {
280: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)A->data;
281: PetscInt i, m, n, *garray = aij->garray;
282: Mat_SeqAIJ *a_aij = (Mat_SeqAIJ *)aij->A->data;
283: Mat_SeqAIJ *b_aij = (Mat_SeqAIJ *)aij->B->data;
284: const PetscScalar *dummy;
286: PetscFunctionBegin;
287: PetscCall(MatGetSize(A, &m, &n));
288: PetscCall(PetscArrayzero(reductions, n));
289: PetscCall(MatSeqAIJGetArrayRead(aij->A, &dummy));
290: PetscCall(MatSeqAIJRestoreArrayRead(aij->A, &dummy));
291: PetscCall(MatSeqAIJGetArrayRead(aij->B, &dummy));
292: PetscCall(MatSeqAIJRestoreArrayRead(aij->B, &dummy));
293: if (type == NORM_2) {
294: for (i = 0; i < a_aij->i[aij->A->rmap->n]; i++) reductions[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i] * a_aij->a[i]);
295: for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) reductions[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i] * b_aij->a[i]);
296: } else if (type == NORM_1) {
297: for (i = 0; i < a_aij->i[aij->A->rmap->n]; i++) reductions[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]);
298: for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) reductions[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]);
299: } else if (type == NORM_INFINITY) {
300: for (i = 0; i < a_aij->i[aij->A->rmap->n]; i++) reductions[A->cmap->rstart + a_aij->j[i]] = PetscMax(PetscAbsScalar(a_aij->a[i]), reductions[A->cmap->rstart + a_aij->j[i]]);
301: for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) reductions[garray[b_aij->j[i]]] = PetscMax(PetscAbsScalar(b_aij->a[i]), reductions[garray[b_aij->j[i]]]);
302: } else if (type == REDUCTION_SUM_REALPART || type == REDUCTION_MEAN_REALPART) {
303: for (i = 0; i < a_aij->i[aij->A->rmap->n]; i++) reductions[A->cmap->rstart + a_aij->j[i]] += PetscRealPart(a_aij->a[i]);
304: for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) reductions[garray[b_aij->j[i]]] += PetscRealPart(b_aij->a[i]);
305: } else {
306: PetscCheck(type == REDUCTION_SUM_IMAGINARYPART || type == REDUCTION_MEAN_IMAGINARYPART, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Unknown reduction type");
307: for (i = 0; i < a_aij->i[aij->A->rmap->n]; i++) reductions[A->cmap->rstart + a_aij->j[i]] += PetscImaginaryPart(a_aij->a[i]);
308: for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) reductions[garray[b_aij->j[i]]] += PetscImaginaryPart(b_aij->a[i]);
309: }
310: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, reductions, n, MPIU_REAL, type == NORM_INFINITY ? MPIU_MAX : MPIU_SUM, PetscObjectComm((PetscObject)A)));
311: if (type == NORM_2) {
312: for (i = 0; i < n; i++) reductions[i] = PetscSqrtReal(reductions[i]);
313: } else if (type == REDUCTION_MEAN_REALPART || type == REDUCTION_MEAN_IMAGINARYPART) {
314: for (i = 0; i < n; i++) reductions[i] /= m;
315: }
316: PetscFunctionReturn(PETSC_SUCCESS);
317: }
319: static PetscErrorCode MatFindOffBlockDiagonalEntries_MPIAIJ(Mat A, IS *is)
320: {
321: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
322: IS sis, gis;
323: const PetscInt *isis, *igis;
324: PetscInt n, *iis, nsis, ngis, rstart, i;
326: PetscFunctionBegin;
327: PetscCall(MatFindOffBlockDiagonalEntries(a->A, &sis));
328: PetscCall(MatFindNonzeroRows(a->B, &gis));
329: PetscCall(ISGetSize(gis, &ngis));
330: PetscCall(ISGetSize(sis, &nsis));
331: PetscCall(ISGetIndices(sis, &isis));
332: PetscCall(ISGetIndices(gis, &igis));
334: PetscCall(PetscMalloc1(ngis + nsis, &iis));
335: PetscCall(PetscArraycpy(iis, igis, ngis));
336: PetscCall(PetscArraycpy(iis + ngis, isis, nsis));
337: n = ngis + nsis;
338: PetscCall(PetscSortRemoveDupsInt(&n, iis));
339: PetscCall(MatGetOwnershipRange(A, &rstart, NULL));
340: for (i = 0; i < n; i++) iis[i] += rstart;
341: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)A), n, iis, PETSC_OWN_POINTER, is));
343: PetscCall(ISRestoreIndices(sis, &isis));
344: PetscCall(ISRestoreIndices(gis, &igis));
345: PetscCall(ISDestroy(&sis));
346: PetscCall(ISDestroy(&gis));
347: PetscFunctionReturn(PETSC_SUCCESS);
348: }
350: /*
351: Local utility routine that creates a mapping from the global column
352: number to the local number in the off-diagonal part of the local
353: storage of the matrix. When PETSC_USE_CTABLE is used this is scalable at
354: a slightly higher hash table cost; without it it is not scalable (each processor
355: has an order N integer array but is fast to access.
356: */
357: PetscErrorCode MatCreateColmap_MPIAIJ_Private(Mat mat)
358: {
359: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
360: PetscInt n = aij->B->cmap->n, i;
362: PetscFunctionBegin;
363: PetscCheck(!n || aij->garray, PETSC_COMM_SELF, PETSC_ERR_PLIB, "MPIAIJ Matrix was assembled but is missing garray");
364: #if PetscDefined(USE_CTABLE)
365: PetscCall(PetscHMapICreateWithSize(n, &aij->colmap));
366: for (i = 0; i < n; i++) PetscCall(PetscHMapISet(aij->colmap, aij->garray[i] + 1, i + 1));
367: #else
368: PetscCall(PetscCalloc1(mat->cmap->N + 1, &aij->colmap));
369: for (i = 0; i < n; i++) aij->colmap[aij->garray[i]] = i + 1;
370: #endif
371: PetscFunctionReturn(PETSC_SUCCESS);
372: }
374: #define MatSetValues_SeqAIJ_A_Private(row, col, value, addv, orow, ocol) \
375: do { \
376: if ((col) <= lastcol1) low1 = 0; \
377: else high1 = nrow1; \
378: lastcol1 = col; \
379: while (high1 - low1 > 5) { \
380: t = (low1 + high1) / 2; \
381: if (rp1[t] > (col)) high1 = t; \
382: else low1 = t; \
383: } \
384: for (_i = low1; _i < high1; _i++) { \
385: if (rp1[_i] > (col)) break; \
386: if (rp1[_i] == (col)) { \
387: if (A->structure_only) goto a_noinsert; \
388: if ((addv) == ADD_VALUES) { \
389: ap1[_i] += value; \
390: /* Not sure LogFlops will slow down the code or not */ \
391: (void)PetscLogFlops(1.0); \
392: } else ap1[_i] = value; \
393: goto a_noinsert; \
394: } \
395: } \
396: if (!A->structure_only && (value) == 0.0 && ignorezeroentries && (orow) != (ocol)) { \
397: low1 = 0; \
398: high1 = nrow1; \
399: goto a_noinsert; \
400: } \
401: if (nonew == 1) { \
402: low1 = 0; \
403: high1 = nrow1; \
404: goto a_noinsert; \
405: } \
406: 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); \
407: if (A->structure_only) MatSeqXAIJReallocateAIJ_structure_only(A, am, 1, nrow1, row, col, rmax1, ai, aj, rp1, aimax, nonew, MatScalar); \
408: else MatSeqXAIJReallocateAIJ(A, am, 1, nrow1, row, col, rmax1, aa, ai, aj, rp1, ap1, aimax, nonew, MatScalar); \
409: N = nrow1++ - 1; \
410: a->nz++; \
411: high1++; \
412: /* shift up all the later entries in this row */ \
413: PetscCall(PetscArraymove(rp1 + _i + 1, rp1 + _i, N - _i + 1)); \
414: rp1[_i] = col; \
415: if (!A->structure_only) { \
416: PetscCall(PetscArraymove(ap1 + _i + 1, ap1 + _i, N - _i + 1)); \
417: ap1[_i] = value; \
418: } \
419: a_noinsert:; \
420: ailen[row] = nrow1; \
421: } while (0)
423: #define MatSetValues_SeqAIJ_B_Private(row, col, value, addv, orow, ocol) \
424: do { \
425: if ((col) <= lastcol2) low2 = 0; \
426: else high2 = nrow2; \
427: lastcol2 = col; \
428: while (high2 - low2 > 5) { \
429: t = (low2 + high2) / 2; \
430: if (rp2[t] > (col)) high2 = t; \
431: else low2 = t; \
432: } \
433: for (_i = low2; _i < high2; _i++) { \
434: if (rp2[_i] > (col)) break; \
435: if (rp2[_i] == (col)) { \
436: if (B->structure_only) goto b_noinsert; \
437: if ((addv) == ADD_VALUES) { \
438: ap2[_i] += value; \
439: (void)PetscLogFlops(1.0); \
440: } else ap2[_i] = value; \
441: goto b_noinsert; \
442: } \
443: } \
444: if (!B->structure_only && (value) == 0.0 && ignorezeroentries) { \
445: low2 = 0; \
446: high2 = nrow2; \
447: goto b_noinsert; \
448: } \
449: if (nonew == 1) { \
450: low2 = 0; \
451: high2 = nrow2; \
452: goto b_noinsert; \
453: } \
454: 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); \
455: if (B->structure_only) MatSeqXAIJReallocateAIJ_structure_only(B, bm, 1, nrow2, row, col, rmax2, bi, bj, rp2, bimax, nonew, MatScalar); \
456: else MatSeqXAIJReallocateAIJ(B, bm, 1, nrow2, row, col, rmax2, ba, bi, bj, rp2, ap2, bimax, nonew, MatScalar); \
457: N = nrow2++ - 1; \
458: b->nz++; \
459: high2++; \
460: /* shift up all the later entries in this row */ \
461: PetscCall(PetscArraymove(rp2 + _i + 1, rp2 + _i, N - _i + 1)); \
462: rp2[_i] = col; \
463: if (!B->structure_only) { \
464: PetscCall(PetscArraymove(ap2 + _i + 1, ap2 + _i, N - _i + 1)); \
465: ap2[_i] = value; \
466: } \
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: const PetscBool ignorezeroentries = (PetscBool)(!mat->structure_only && 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 && !mat->structure_only) value = roworiented ? v[i * n + j] : v[i + j * m];
555: if (!mat->structure_only && 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 PetscDefined(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: PetscCheck(1 == ((Mat_SeqAIJ *)aij->B->data)->nonew, 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]);
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: }
595: } else col = in[j];
596: nonew = b->nonew;
597: MatSetValues_SeqAIJ_B_Private(row, col, value, addv, im[i], in[j]);
598: }
599: }
600: } else {
601: 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]);
602: if (!aij->donotstash) {
603: mat->assembled = PETSC_FALSE;
604: if (roworiented) {
605: PetscCall(MatStashValuesRow_Private(&mat->stash, im[i], n, in, PetscSafePointerPlusOffset(v, i * n), (PetscBool)(ignorezeroentries && addv == ADD_VALUES)));
606: } else {
607: PetscCall(MatStashValuesCol_Private(&mat->stash, im[i], n, in, PetscSafePointerPlusOffset(v, i), m, (PetscBool)(ignorezeroentries && addv == ADD_VALUES)));
608: }
609: }
610: }
611: }
612: PetscCall(MatSeqAIJRestoreArray(A, &aa)); /* aa, bb might have been free'd due to reallocation above. But we don't access them here */
613: PetscCall(MatSeqAIJRestoreArray(B, &ba));
614: PetscFunctionReturn(PETSC_SUCCESS);
615: }
617: /*
618: This function sets the j and ilen arrays (of the diagonal and off-diagonal part) of an MPIAIJ-matrix.
619: The values in mat_i have to be sorted and the values in mat_j have to be sorted for each row (CSR-like).
620: No off-processor parts off the matrix are allowed here and mat->was_assembled has to be PETSC_FALSE.
621: */
622: PetscErrorCode MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(Mat mat, const PetscInt mat_j[], const PetscInt mat_i[])
623: {
624: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
625: Mat A = aij->A; /* diagonal part of the matrix */
626: Mat B = aij->B; /* off-diagonal part of the matrix */
627: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
628: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
629: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend, col;
630: PetscInt *ailen = a->ilen, *aj = a->j;
631: PetscInt *bilen = b->ilen, *bj = b->j;
632: PetscInt am = aij->A->rmap->n, j;
633: PetscInt diag_so_far = 0, dnz;
634: PetscInt offd_so_far = 0, onz;
636: PetscFunctionBegin;
637: /* Iterate over all rows of the matrix */
638: for (j = 0; j < am; j++) {
639: dnz = onz = 0;
640: /* Iterate over all non-zero columns of the current row */
641: for (col = mat_i[j]; col < mat_i[j + 1]; col++) {
642: /* If column is in the diagonal */
643: if (mat_j[col] >= cstart && mat_j[col] < cend) {
644: aj[diag_so_far++] = mat_j[col] - cstart;
645: dnz++;
646: } else { /* off-diagonal entries */
647: bj[offd_so_far++] = mat_j[col];
648: onz++;
649: }
650: }
651: ailen[j] = dnz;
652: bilen[j] = onz;
653: }
654: PetscFunctionReturn(PETSC_SUCCESS);
655: }
657: /*
658: This function sets the local j, a and ilen arrays (of the diagonal and off-diagonal part) of an MPIAIJ-matrix.
659: The values in mat_i have to be sorted and the values in mat_j have to be sorted for each row (CSR-like).
660: No off-processor parts off the matrix are allowed here, they are set at a later point by MatSetValues_MPIAIJ.
661: Also, mat->was_assembled has to be false, otherwise the statement aj[rowstart_diag+dnz_row] = mat_j[col] - cstart;
662: would not be true and the more complex MatSetValues_MPIAIJ has to be used.
663: */
664: PetscErrorCode MatSetValues_MPIAIJ_CopyFromCSRFormat(Mat mat, const PetscInt mat_j[], const PetscInt mat_i[], const PetscScalar mat_a[])
665: {
666: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
667: Mat A = aij->A; /* diagonal part of the matrix */
668: Mat B = aij->B; /* off-diagonal part of the matrix */
669: Mat_SeqAIJ *aijd = (Mat_SeqAIJ *)aij->A->data, *aijo = (Mat_SeqAIJ *)aij->B->data;
670: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
671: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
672: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend;
673: PetscInt *ailen = a->ilen, *aj = a->j;
674: PetscInt *bilen = b->ilen, *bj = b->j;
675: PetscInt am = aij->A->rmap->n, j;
676: 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. */
677: PetscInt col, dnz_row, onz_row, rowstart_diag, rowstart_offd;
678: PetscScalar *aa = a->a, *ba = b->a;
680: PetscFunctionBegin;
681: /* Iterate over all rows of the matrix */
682: for (j = 0; j < am; j++) {
683: dnz_row = onz_row = 0;
684: rowstart_offd = full_offd_i[j];
685: rowstart_diag = full_diag_i[j];
686: /* Iterate over all non-zero columns of the current row */
687: for (col = mat_i[j]; col < mat_i[j + 1]; col++) {
688: /* If column is in the diagonal */
689: if (mat_j[col] >= cstart && mat_j[col] < cend) {
690: aj[rowstart_diag + dnz_row] = mat_j[col] - cstart;
691: aa[rowstart_diag + dnz_row] = mat_a[col];
692: dnz_row++;
693: } else { /* off-diagonal entries */
694: bj[rowstart_offd + onz_row] = mat_j[col];
695: ba[rowstart_offd + onz_row] = mat_a[col];
696: onz_row++;
697: }
698: }
699: ailen[j] = dnz_row;
700: bilen[j] = onz_row;
701: }
702: PetscFunctionReturn(PETSC_SUCCESS);
703: }
705: static PetscErrorCode MatGetValues_MPIAIJ(Mat mat, PetscInt m, const PetscInt idxm[], PetscInt n, const PetscInt idxn[], PetscScalar v[])
706: {
707: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
708: PetscInt i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend;
709: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend, row, col;
710: PetscBool roworiented = aij->roworiented;
711: PetscScalar *value;
713: PetscFunctionBegin;
714: for (i = 0; i < m; i++) {
715: if (idxm[i] < 0) continue; /* negative row */
716: 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);
717: 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);
718: row = idxm[i] - rstart;
719: for (j = 0; j < n; j++) {
720: if (idxn[j] < 0) continue; /* negative column */
721: 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);
722: value = roworiented ? &v[j + i * n] : &v[i + j * m];
723: if (idxn[j] >= cstart && idxn[j] < cend) {
724: col = idxn[j] - cstart;
725: PetscCall(MatGetValues(aij->A, 1, &row, 1, &col, value));
726: } else {
727: if (!aij->colmap) PetscCall(MatCreateColmap_MPIAIJ_Private(mat));
728: #if PetscDefined(USE_CTABLE)
729: PetscCall(PetscHMapIGetWithDefault(aij->colmap, idxn[j] + 1, 0, &col));
730: col--;
731: #else
732: col = aij->colmap[idxn[j]] - 1;
733: #endif
734: if (col < 0 || aij->garray[col] != idxn[j]) *value = 0.0;
735: else PetscCall(MatGetValues(aij->B, 1, &row, 1, &col, value));
736: }
737: }
738: }
739: PetscFunctionReturn(PETSC_SUCCESS);
740: }
742: static PetscErrorCode MatAssemblyBegin_MPIAIJ(Mat mat, MatAssemblyType mode)
743: {
744: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
745: PetscInt nstash, reallocs;
747: PetscFunctionBegin;
748: if (aij->donotstash || mat->nooffprocentries) PetscFunctionReturn(PETSC_SUCCESS);
750: PetscCall(MatStashScatterBegin_Private(mat, &mat->stash, mat->rmap->range));
751: PetscCall(MatStashGetInfo_Private(&mat->stash, &nstash, &reallocs));
752: PetscCall(PetscInfo(mat, "Stash has %" PetscInt_FMT " entries, uses %" PetscInt_FMT " mallocs.\n", nstash, reallocs));
753: PetscFunctionReturn(PETSC_SUCCESS);
754: }
756: PetscErrorCode MatAssemblyEnd_MPIAIJ(Mat mat, MatAssemblyType mode)
757: {
758: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
759: PetscMPIInt n;
760: PetscInt i, j, rstart, ncols, flg;
761: PetscInt *row, *col;
762: PetscBool all_assembled;
763: PetscScalar *val;
765: /* do not use 'b = (Mat_SeqAIJ*)aij->B->data' as B can be reset in disassembly */
767: PetscFunctionBegin;
768: if (!aij->donotstash && !mat->nooffprocentries) {
769: while (1) {
770: PetscCall(MatStashScatterGetMesg_Private(&mat->stash, &n, &row, &col, &val, &flg));
771: if (!flg) break;
773: for (i = 0; i < n;) {
774: /* Now identify the consecutive vals belonging to the same row */
775: for (j = i, rstart = row[j]; j < n; j++) {
776: if (row[j] != rstart) break;
777: }
778: if (j < n) ncols = j - i;
779: else ncols = n - i;
780: /* Now assemble all these values with a single function call */
781: PetscCall(MatSetValues_MPIAIJ(mat, 1, row + i, ncols, col + i, val + i, mat->insertmode));
782: i = j;
783: }
784: }
785: PetscCall(MatStashScatterEnd_Private(&mat->stash));
786: }
787: #if PetscDefined(HAVE_DEVICE)
788: if (mat->offloadmask == PETSC_OFFLOAD_CPU) aij->A->offloadmask = PETSC_OFFLOAD_CPU;
789: /* We call MatBindToCPU() on aij->A and aij->B here, because if MatBindToCPU_MPIAIJ() is called before assembly, it cannot bind these. */
790: if (mat->boundtocpu) {
791: PetscCall(MatBindToCPU(aij->A, PETSC_TRUE));
792: PetscCall(MatBindToCPU(aij->B, PETSC_TRUE));
793: }
794: #endif
795: PetscCall(MatAssemblyBegin(aij->A, mode));
796: PetscCall(MatAssemblyEnd(aij->A, mode));
798: /* determine if any process has disassembled, if so we must
799: also disassemble ourself, in order that we may reassemble. */
800: /*
801: if nonzero structure of submatrix B cannot change then we know that
802: no process disassembled thus we can skip this stuff
803: */
804: if (!((Mat_SeqAIJ *)aij->B->data)->nonew) {
805: PetscCallMPI(MPIU_Allreduce(&mat->was_assembled, &all_assembled, 1, MPI_C_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
806: if (mat->was_assembled && !all_assembled) { /* mat on this rank has reduced off-diag B with local col ids, but globally it does not */
807: PetscCall(MatDisAssemble_MPIAIJ(mat, PETSC_FALSE));
808: }
809: }
810: if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) PetscCall(MatSetUpMultiply_MPIAIJ(mat));
811: PetscCall(MatSetOption(aij->B, MAT_USE_INODES, PETSC_FALSE));
812: #if PetscDefined(HAVE_DEVICE)
813: if (mat->offloadmask == PETSC_OFFLOAD_CPU && aij->B->offloadmask != PETSC_OFFLOAD_UNALLOCATED) aij->B->offloadmask = PETSC_OFFLOAD_CPU;
814: #endif
815: PetscCall(MatAssemblyBegin(aij->B, mode));
816: PetscCall(MatAssemblyEnd(aij->B, mode));
818: PetscCall(PetscFree2(aij->rowvalues, aij->rowindices));
820: aij->rowvalues = NULL;
822: PetscCall(VecDestroy(&aij->diag));
824: /* if no new nonzero locations are allowed in matrix then only set the matrix state the first time through */
825: if ((!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) || !((Mat_SeqAIJ *)aij->A->data)->nonew) {
826: mat->nonzerostate = aij->A->nonzerostate + aij->B->nonzerostate;
827: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &mat->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)mat)));
828: }
829: #if PetscDefined(HAVE_DEVICE)
830: mat->offloadmask = PETSC_OFFLOAD_BOTH;
831: #endif
832: PetscFunctionReturn(PETSC_SUCCESS);
833: }
835: static PetscErrorCode MatZeroEntries_MPIAIJ(Mat A)
836: {
837: Mat_MPIAIJ *l = (Mat_MPIAIJ *)A->data;
839: PetscFunctionBegin;
840: PetscCall(MatZeroEntries(l->A));
841: PetscCall(MatZeroEntries(l->B));
842: PetscFunctionReturn(PETSC_SUCCESS);
843: }
845: static PetscErrorCode MatZeroRows_MPIAIJ(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diag, Vec x, Vec b)
846: {
847: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
848: PetscInt *lrows;
849: PetscInt r, len;
850: PetscBool cong;
852: PetscFunctionBegin;
853: /* get locally owned rows */
854: PetscCall(MatZeroRowsMapLocal_Private(A, N, rows, &len, &lrows));
855: PetscCall(MatHasCongruentLayouts(A, &cong));
856: /* fix right-hand side if needed */
857: if (x && b) {
858: const PetscScalar *xx;
859: PetscScalar *bb;
861: PetscCheck(cong, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Need matching row/col layout");
862: PetscCall(VecGetArrayRead(x, &xx));
863: PetscCall(VecGetArray(b, &bb));
864: for (r = 0; r < len; ++r) bb[lrows[r]] = diag * xx[lrows[r]];
865: PetscCall(VecRestoreArrayRead(x, &xx));
866: PetscCall(VecRestoreArray(b, &bb));
867: }
869: if (diag != 0.0 && cong) {
870: PetscCall(MatZeroRows(mat->A, len, lrows, diag, NULL, NULL));
871: PetscCall(MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL));
872: } else if (diag != 0.0) { /* non-square or non congruent layouts -> if keepnonzeropattern is false, we allow for new insertion */
873: Mat_SeqAIJ *aijA = (Mat_SeqAIJ *)mat->A->data;
874: Mat_SeqAIJ *aijB = (Mat_SeqAIJ *)mat->B->data;
875: PetscInt nnwA, nnwB;
876: PetscBool nnzA, nnzB;
878: nnwA = aijA->nonew;
879: nnwB = aijB->nonew;
880: nnzA = aijA->keepnonzeropattern;
881: nnzB = aijB->keepnonzeropattern;
882: if (!nnzA) {
883: PetscCall(PetscInfo(mat->A, "Requested to not keep the pattern and add a nonzero diagonal; may encounter reallocations on diagonal block.\n"));
884: aijA->nonew = 0;
885: }
886: if (!nnzB) {
887: PetscCall(PetscInfo(mat->B, "Requested to not keep the pattern and add a nonzero diagonal; may encounter reallocations on off-diagonal block.\n"));
888: aijB->nonew = 0;
889: }
890: /* Must zero here before the next loop */
891: PetscCall(MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL));
892: PetscCall(MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL));
893: for (r = 0; r < len; ++r) {
894: const PetscInt row = lrows[r] + A->rmap->rstart;
895: if (row >= A->cmap->N) continue;
896: PetscCall(MatSetValues(A, 1, &row, 1, &row, &diag, INSERT_VALUES));
897: }
898: aijA->nonew = nnwA;
899: aijB->nonew = nnwB;
900: } else {
901: PetscCall(MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL));
902: PetscCall(MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL));
903: }
904: PetscCall(PetscFree(lrows));
905: PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY));
906: PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY));
908: /* only change matrix nonzero state if pattern was allowed to be changed */
909: if (!((Mat_SeqAIJ *)mat->A->data)->keepnonzeropattern || !((Mat_SeqAIJ *)mat->A->data)->nonew) {
910: A->nonzerostate = mat->A->nonzerostate + mat->B->nonzerostate;
911: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &A->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)A)));
912: }
913: PetscFunctionReturn(PETSC_SUCCESS);
914: }
916: static PetscErrorCode MatZeroRowsColumns_MPIAIJ(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diag, Vec x, Vec b)
917: {
918: Mat_MPIAIJ *l = (Mat_MPIAIJ *)A->data;
919: PetscInt n = A->rmap->n;
920: PetscInt i, j, r, m, len = 0;
921: PetscInt *lrows, *owners = A->rmap->range;
922: PetscMPIInt p = 0;
923: PetscSFNode *rrows;
924: PetscSF sf;
925: const PetscScalar *xx;
926: PetscScalar *bb, *mask, *aij_a;
927: Vec xmask, lmask;
928: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)l->B->data;
929: const PetscInt *aj, *ii, *ridx;
930: PetscScalar *aa;
932: PetscFunctionBegin;
933: /* Create SF where leaves are input rows and roots are owned rows */
934: PetscCall(PetscMalloc1(n, &lrows));
935: for (r = 0; r < n; ++r) lrows[r] = -1;
936: PetscCall(PetscMalloc1(N, &rrows));
937: for (r = 0; r < N; ++r) {
938: const PetscInt idx = rows[r];
939: 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);
940: if (idx < owners[p] || owners[p + 1] <= idx) { /* short-circuit the search if the last p owns this row too */
941: PetscCall(PetscLayoutFindOwner(A->rmap, idx, &p));
942: }
943: rrows[r].rank = p;
944: rrows[r].index = rows[r] - owners[p];
945: }
946: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
947: PetscCall(PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER));
948: /* Collect flags for rows to be zeroed */
949: PetscCall(PetscSFReduceBegin(sf, MPIU_INT, (PetscInt *)rows, lrows, MPI_LOR));
950: PetscCall(PetscSFReduceEnd(sf, MPIU_INT, (PetscInt *)rows, lrows, MPI_LOR));
951: PetscCall(PetscSFDestroy(&sf));
952: /* Compress and put in row numbers */
953: for (r = 0; r < n; ++r)
954: if (lrows[r] >= 0) lrows[len++] = r;
955: /* zero diagonal part of matrix */
956: PetscCall(MatZeroRowsColumns(l->A, len, lrows, diag, x, b));
957: /* handle off-diagonal part of matrix */
958: PetscCall(MatCreateVecs(A, &xmask, NULL));
959: PetscCall(VecDuplicate(l->lvec, &lmask));
960: PetscCall(VecGetArray(xmask, &bb));
961: for (i = 0; i < len; i++) bb[lrows[i]] = 1;
962: PetscCall(VecRestoreArray(xmask, &bb));
963: PetscCall(VecScatterBegin(l->Mvctx, xmask, lmask, ADD_VALUES, SCATTER_FORWARD));
964: PetscCall(VecScatterEnd(l->Mvctx, xmask, lmask, ADD_VALUES, SCATTER_FORWARD));
965: PetscCall(VecDestroy(&xmask));
966: if (x && b) { /* this code is buggy when the row and column layout don't match */
967: PetscBool cong;
969: PetscCall(MatHasCongruentLayouts(A, &cong));
970: PetscCheck(cong, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Need matching row/col layout");
971: PetscCall(VecScatterBegin(l->Mvctx, x, l->lvec, INSERT_VALUES, SCATTER_FORWARD));
972: PetscCall(VecScatterEnd(l->Mvctx, x, l->lvec, INSERT_VALUES, SCATTER_FORWARD));
973: PetscCall(VecGetArrayRead(l->lvec, &xx));
974: PetscCall(VecGetArray(b, &bb));
975: }
976: PetscCall(VecGetArray(lmask, &mask));
977: /* remove zeroed rows of off-diagonal matrix */
978: PetscCall(MatSeqAIJGetArray(l->B, &aij_a));
979: ii = aij->i;
980: for (i = 0; i < len; i++) PetscCall(PetscArrayzero(PetscSafePointerPlusOffset(aij_a, ii[lrows[i]]), ii[lrows[i] + 1] - ii[lrows[i]]));
981: /* loop over all elements of off process part of matrix zeroing removed columns*/
982: if (aij->compressedrow.use) {
983: m = aij->compressedrow.nrows;
984: ii = aij->compressedrow.i;
985: ridx = aij->compressedrow.rindex;
986: for (i = 0; i < m; i++) {
987: n = ii[i + 1] - ii[i];
988: aj = aij->j + ii[i];
989: aa = aij_a + ii[i];
991: for (j = 0; j < n; j++) {
992: if (PetscAbsScalar(mask[*aj])) {
993: if (b) bb[*ridx] -= *aa * xx[*aj];
994: *aa = 0.0;
995: }
996: aa++;
997: aj++;
998: }
999: ridx++;
1000: }
1001: } else { /* do not use compressed row format */
1002: m = l->B->rmap->n;
1003: for (i = 0; i < m; i++) {
1004: n = ii[i + 1] - ii[i];
1005: aj = aij->j + ii[i];
1006: aa = aij_a + ii[i];
1007: for (j = 0; j < n; j++) {
1008: if (PetscAbsScalar(mask[*aj])) {
1009: if (b) bb[i] -= *aa * xx[*aj];
1010: *aa = 0.0;
1011: }
1012: aa++;
1013: aj++;
1014: }
1015: }
1016: }
1017: if (x && b) {
1018: PetscCall(VecRestoreArray(b, &bb));
1019: PetscCall(VecRestoreArrayRead(l->lvec, &xx));
1020: }
1021: PetscCall(MatSeqAIJRestoreArray(l->B, &aij_a));
1022: PetscCall(VecRestoreArray(lmask, &mask));
1023: PetscCall(VecDestroy(&lmask));
1024: PetscCall(PetscFree(lrows));
1026: /* only change matrix nonzero state if pattern was allowed to be changed */
1027: if (!((Mat_SeqAIJ *)l->A->data)->nonew) {
1028: A->nonzerostate = l->A->nonzerostate + l->B->nonzerostate;
1029: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &A->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)A)));
1030: }
1031: PetscFunctionReturn(PETSC_SUCCESS);
1032: }
1034: static PetscErrorCode MatMult_MPIAIJ(Mat A, Vec xx, Vec yy)
1035: {
1036: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1037: PetscInt nt;
1038: VecScatter Mvctx = a->Mvctx;
1040: PetscFunctionBegin;
1041: PetscCall(VecGetLocalSize(xx, &nt));
1042: 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);
1043: PetscCall(VecScatterBegin(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1044: PetscUseTypeMethod(a->A, mult, xx, yy);
1045: PetscCall(VecScatterEnd(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1046: PetscUseTypeMethod(a->B, multadd, a->lvec, yy, yy);
1047: PetscFunctionReturn(PETSC_SUCCESS);
1048: }
1050: static PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat A, Vec bb, Vec xx)
1051: {
1052: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1054: PetscFunctionBegin;
1055: PetscCall(MatMultDiagonalBlock(a->A, bb, xx));
1056: PetscFunctionReturn(PETSC_SUCCESS);
1057: }
1059: static PetscErrorCode MatMultAdd_MPIAIJ(Mat A, Vec xx, Vec yy, Vec zz)
1060: {
1061: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1062: VecScatter Mvctx = a->Mvctx;
1064: PetscFunctionBegin;
1065: PetscCall(VecScatterBegin(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1066: PetscUseTypeMethod(a->A, multadd, xx, yy, zz);
1067: PetscCall(VecScatterEnd(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1068: PetscUseTypeMethod(a->B, multadd, a->lvec, zz, zz);
1069: PetscFunctionReturn(PETSC_SUCCESS);
1070: }
1072: static PetscErrorCode MatMultTranspose_MPIAIJ(Mat A, Vec xx, Vec yy)
1073: {
1074: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1076: PetscFunctionBegin;
1077: /* do nondiagonal part */
1078: PetscUseTypeMethod(a->B, multtranspose, xx, a->lvec);
1079: /* do local part */
1080: PetscUseTypeMethod(a->A, multtranspose, xx, yy);
1081: /* add partial results together */
1082: PetscCall(VecScatterBegin(a->Mvctx, a->lvec, yy, ADD_VALUES, SCATTER_REVERSE));
1083: PetscCall(VecScatterEnd(a->Mvctx, a->lvec, yy, ADD_VALUES, SCATTER_REVERSE));
1084: PetscFunctionReturn(PETSC_SUCCESS);
1085: }
1087: static PetscErrorCode MatIsTranspose_MPIAIJ(Mat Amat, Mat Bmat, PetscReal tol, PetscBool *f)
1088: {
1089: MPI_Comm comm;
1090: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)Amat->data, *Bij = (Mat_MPIAIJ *)Bmat->data;
1091: Mat Adia = Aij->A, Bdia = Bij->A, Aoff, Boff, *Aoffs, *Boffs;
1092: IS Me, Notme;
1093: PetscInt M, N, first, last, *notme, i;
1094: PetscMPIInt size;
1096: PetscFunctionBegin;
1097: /* Easy test: symmetric diagonal block */
1098: PetscCall(MatIsTranspose(Adia, Bdia, tol, f));
1099: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, f, 1, MPI_C_BOOL, MPI_LAND, PetscObjectComm((PetscObject)Amat)));
1100: if (!*f) PetscFunctionReturn(PETSC_SUCCESS);
1101: PetscCall(PetscObjectGetComm((PetscObject)Amat, &comm));
1102: PetscCallMPI(MPI_Comm_size(comm, &size));
1103: if (size == 1) PetscFunctionReturn(PETSC_SUCCESS);
1105: /* Hard test: off-diagonal block. This takes a MatCreateSubMatrix. */
1106: PetscCall(MatGetSize(Amat, &M, &N));
1107: PetscCall(MatGetOwnershipRange(Amat, &first, &last));
1108: PetscCall(PetscMalloc1(N - last + first, ¬me));
1109: for (i = 0; i < first; i++) notme[i] = i;
1110: for (i = last; i < M; i++) notme[i - last + first] = i;
1111: PetscCall(ISCreateGeneral(MPI_COMM_SELF, N - last + first, notme, PETSC_COPY_VALUES, &Notme));
1112: PetscCall(ISCreateStride(MPI_COMM_SELF, last - first, first, 1, &Me));
1113: PetscCall(MatCreateSubMatrices(Amat, 1, &Me, &Notme, MAT_INITIAL_MATRIX, &Aoffs));
1114: Aoff = Aoffs[0];
1115: PetscCall(MatCreateSubMatrices(Bmat, 1, &Notme, &Me, MAT_INITIAL_MATRIX, &Boffs));
1116: Boff = Boffs[0];
1117: PetscCall(MatIsTranspose(Aoff, Boff, tol, f));
1118: PetscCall(MatDestroyMatrices(1, &Aoffs));
1119: PetscCall(MatDestroyMatrices(1, &Boffs));
1120: PetscCall(ISDestroy(&Me));
1121: PetscCall(ISDestroy(&Notme));
1122: PetscCall(PetscFree(notme));
1123: PetscFunctionReturn(PETSC_SUCCESS);
1124: }
1126: static PetscErrorCode MatMultTransposeAdd_MPIAIJ(Mat A, Vec xx, Vec yy, Vec zz)
1127: {
1128: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1130: PetscFunctionBegin;
1131: /* do nondiagonal part */
1132: PetscUseTypeMethod(a->B, multtranspose, xx, a->lvec);
1133: /* do local part */
1134: PetscUseTypeMethod(a->A, multtransposeadd, xx, yy, zz);
1135: /* add partial results together */
1136: PetscCall(VecScatterBegin(a->Mvctx, a->lvec, zz, ADD_VALUES, SCATTER_REVERSE));
1137: PetscCall(VecScatterEnd(a->Mvctx, a->lvec, zz, ADD_VALUES, SCATTER_REVERSE));
1138: PetscFunctionReturn(PETSC_SUCCESS);
1139: }
1141: /*
1142: This only works correctly for square matrices where the subblock A->A is the
1143: diagonal block
1144: */
1145: static PetscErrorCode MatGetDiagonal_MPIAIJ(Mat A, Vec v)
1146: {
1147: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1149: PetscFunctionBegin;
1150: PetscCheck(A->rmap->N == A->cmap->N, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Supports only square matrix where A->A is diag block");
1151: 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");
1152: PetscCall(MatGetDiagonal(a->A, v));
1153: PetscFunctionReturn(PETSC_SUCCESS);
1154: }
1156: static PetscErrorCode MatScale_MPIAIJ(Mat A, PetscScalar aa)
1157: {
1158: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1160: PetscFunctionBegin;
1161: PetscCall(MatScale(a->A, aa));
1162: PetscCall(MatScale(a->B, aa));
1163: PetscFunctionReturn(PETSC_SUCCESS);
1164: }
1166: static PetscErrorCode MatView_MPIAIJ_Binary(Mat mat, PetscViewer viewer)
1167: {
1168: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1169: Mat_SeqAIJ *A = (Mat_SeqAIJ *)aij->A->data;
1170: Mat_SeqAIJ *B = (Mat_SeqAIJ *)aij->B->data;
1171: const PetscInt *garray = aij->garray;
1172: const PetscScalar *aa, *ba;
1173: PetscInt header[4], M, N, m, rs, cs, cnt, i, ja, jb;
1174: PetscInt64 nz, hnz;
1175: PetscInt *rowlens;
1176: PetscInt *colidxs;
1177: PetscScalar *matvals;
1178: PetscMPIInt rank;
1180: PetscFunctionBegin;
1181: PetscCall(PetscViewerSetUp(viewer));
1183: M = mat->rmap->N;
1184: N = mat->cmap->N;
1185: m = mat->rmap->n;
1186: rs = mat->rmap->rstart;
1187: cs = mat->cmap->rstart;
1188: nz = A->nz + B->nz;
1190: /* write matrix header */
1191: header[0] = MAT_FILE_CLASSID;
1192: header[1] = M;
1193: header[2] = N;
1194: PetscCallMPI(MPI_Reduce(&nz, &hnz, 1, MPIU_INT64, MPI_SUM, 0, PetscObjectComm((PetscObject)mat)));
1195: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)mat), &rank));
1196: if (rank == 0) PetscCall(PetscIntCast(hnz, &header[3]));
1197: PetscCall(PetscViewerBinaryWrite(viewer, header, 4, PETSC_INT));
1199: /* fill in and store row lengths */
1200: PetscCall(PetscMalloc1(m, &rowlens));
1201: for (i = 0; i < m; i++) rowlens[i] = A->i[i + 1] - A->i[i] + B->i[i + 1] - B->i[i];
1202: PetscCall(PetscViewerBinaryWriteAll(viewer, rowlens, m, rs, M, PETSC_INT));
1203: PetscCall(PetscFree(rowlens));
1205: /* fill in and store column indices */
1206: PetscCall(PetscMalloc1(nz, &colidxs));
1207: for (cnt = 0, i = 0; i < m; i++) {
1208: for (jb = B->i[i]; jb < B->i[i + 1]; jb++) {
1209: if (garray[B->j[jb]] > cs) break;
1210: colidxs[cnt++] = garray[B->j[jb]];
1211: }
1212: for (ja = A->i[i]; ja < A->i[i + 1]; ja++) colidxs[cnt++] = A->j[ja] + cs;
1213: for (; jb < B->i[i + 1]; jb++) colidxs[cnt++] = garray[B->j[jb]];
1214: }
1215: PetscCheck(cnt == nz, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Internal PETSc error: cnt = %" PetscInt_FMT " nz = %" PetscInt64_FMT, cnt, nz);
1216: PetscCall(PetscViewerBinaryWriteAll(viewer, colidxs, nz, PETSC_DETERMINE, PETSC_DETERMINE, PETSC_INT));
1217: PetscCall(PetscFree(colidxs));
1219: /* fill in and store nonzero values */
1220: PetscCall(MatSeqAIJGetArrayRead(aij->A, &aa));
1221: PetscCall(MatSeqAIJGetArrayRead(aij->B, &ba));
1222: PetscCall(PetscMalloc1(nz, &matvals));
1223: for (cnt = 0, i = 0; i < m; i++) {
1224: for (jb = B->i[i]; jb < B->i[i + 1]; jb++) {
1225: if (garray[B->j[jb]] > cs) break;
1226: matvals[cnt++] = ba[jb];
1227: }
1228: for (ja = A->i[i]; ja < A->i[i + 1]; ja++) matvals[cnt++] = aa[ja];
1229: for (; jb < B->i[i + 1]; jb++) matvals[cnt++] = ba[jb];
1230: }
1231: PetscCall(MatSeqAIJRestoreArrayRead(aij->A, &aa));
1232: PetscCall(MatSeqAIJRestoreArrayRead(aij->B, &ba));
1233: PetscCheck(cnt == nz, PETSC_COMM_SELF, PETSC_ERR_LIB, "Internal PETSc error: cnt = %" PetscInt_FMT " nz = %" PetscInt64_FMT, cnt, nz);
1234: PetscCall(PetscViewerBinaryWriteAll(viewer, matvals, nz, PETSC_DETERMINE, PETSC_DETERMINE, PETSC_SCALAR));
1235: PetscCall(PetscFree(matvals));
1237: /* write block size option to the viewer's .info file */
1238: PetscCall(MatView_Binary_BlockSizes(mat, viewer));
1239: PetscFunctionReturn(PETSC_SUCCESS);
1240: }
1242: #include <petscdraw.h>
1243: static PetscErrorCode MatView_MPIAIJ_ASCIIorDraworSocket(Mat mat, PetscViewer viewer)
1244: {
1245: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1246: PetscMPIInt rank = aij->rank, size = aij->size;
1247: PetscBool isdraw, isascii, isbinary;
1248: PetscViewer sviewer;
1249: PetscViewerFormat format;
1251: PetscFunctionBegin;
1252: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw));
1253: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii));
1254: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
1255: if (isascii) {
1256: PetscCall(PetscViewerGetFormat(viewer, &format));
1257: if (format == PETSC_VIEWER_LOAD_BALANCE) {
1258: PetscInt i, nmax = 0, nmin = PETSC_INT_MAX, navg = 0, *nz, nzlocal = ((Mat_SeqAIJ *)aij->A->data)->nz + ((Mat_SeqAIJ *)aij->B->data)->nz;
1259: PetscCall(PetscMalloc1(size, &nz));
1260: PetscCallMPI(MPI_Allgather(&nzlocal, 1, MPIU_INT, nz, 1, MPIU_INT, PetscObjectComm((PetscObject)mat)));
1261: for (i = 0; i < size; i++) {
1262: nmax = PetscMax(nmax, nz[i]);
1263: nmin = PetscMin(nmin, nz[i]);
1264: navg += nz[i];
1265: }
1266: PetscCall(PetscFree(nz));
1267: navg = navg / size;
1268: PetscCall(PetscViewerASCIIPrintf(viewer, "Load Balance - Nonzeros: Min %" PetscInt_FMT " avg %" PetscInt_FMT " max %" PetscInt_FMT "\n", nmin, navg, nmax));
1269: PetscFunctionReturn(PETSC_SUCCESS);
1270: }
1271: PetscCall(PetscViewerGetFormat(viewer, &format));
1272: if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
1273: MatInfo info;
1274: PetscInt *inodes = NULL;
1276: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)mat), &rank));
1277: PetscCall(MatGetInfo(mat, MAT_LOCAL, &info));
1278: PetscCall(MatInodeGetInodeSizes(aij->A, NULL, &inodes, NULL));
1279: PetscCall(PetscViewerASCIIPushSynchronized(viewer));
1280: if (!inodes) {
1281: 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,
1282: info.memory));
1283: } else {
1284: PetscCall(
1285: 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, info.memory));
1286: }
1287: PetscCall(MatGetInfo(aij->A, MAT_LOCAL, &info));
1288: PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] on-diagonal part: nz %" PetscInt_FMT " \n", rank, (PetscInt)info.nz_used));
1289: PetscCall(MatGetInfo(aij->B, MAT_LOCAL, &info));
1290: PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] off-diagonal part: nz %" PetscInt_FMT " \n", rank, (PetscInt)info.nz_used));
1291: PetscCall(PetscViewerFlush(viewer));
1292: PetscCall(PetscViewerASCIIPopSynchronized(viewer));
1293: PetscCall(PetscViewerASCIIPrintf(viewer, "Information on VecScatter used in matrix-vector product: \n"));
1294: PetscCall(VecScatterView(aij->Mvctx, viewer));
1295: PetscFunctionReturn(PETSC_SUCCESS);
1296: } else if (format == PETSC_VIEWER_ASCII_INFO) {
1297: PetscInt inodecount, inodelimit, *inodes;
1298: PetscCall(MatInodeGetInodeSizes(aij->A, &inodecount, &inodes, &inodelimit));
1299: if (inodes) {
1300: PetscCall(PetscViewerASCIIPrintf(viewer, "using I-node (on process 0) routines: found %" PetscInt_FMT " nodes, limit used is %" PetscInt_FMT "\n", inodecount, inodelimit));
1301: } else {
1302: PetscCall(PetscViewerASCIIPrintf(viewer, "not using I-node (on process 0) routines\n"));
1303: }
1304: PetscFunctionReturn(PETSC_SUCCESS);
1305: } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) {
1306: PetscFunctionReturn(PETSC_SUCCESS);
1307: }
1308: } else if (isbinary) {
1309: if (size == 1) {
1310: PetscCall(PetscObjectSetName((PetscObject)aij->A, ((PetscObject)mat)->name));
1311: PetscCall(MatView(aij->A, viewer));
1312: } else {
1313: PetscCall(MatView_MPIAIJ_Binary(mat, viewer));
1314: }
1315: PetscFunctionReturn(PETSC_SUCCESS);
1316: } else if (isascii && size == 1) {
1317: PetscCall(PetscObjectSetName((PetscObject)aij->A, ((PetscObject)mat)->name));
1318: PetscCall(MatView(aij->A, viewer));
1319: PetscFunctionReturn(PETSC_SUCCESS);
1320: } else if (isdraw) {
1321: PetscDraw draw;
1322: PetscBool isnull;
1323: PetscCall(PetscViewerDrawGetDraw(viewer, 0, &draw));
1324: PetscCall(PetscDrawIsNull(draw, &isnull));
1325: if (isnull) PetscFunctionReturn(PETSC_SUCCESS);
1326: }
1328: { /* assemble the entire matrix onto first process */
1329: Mat A, Av;
1330: IS isrow, iscol;
1332: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat), rank == 0 ? mat->rmap->N : 0, 0, 1, &isrow));
1333: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat), rank == 0 ? mat->cmap->N : 0, 0, 1, &iscol));
1334: PetscCall(MatCreateSubMatrix(mat, isrow, iscol, MAT_INITIAL_MATRIX, &A));
1335: PetscCall(MatMPIAIJGetSeqAIJ(A, &Av, NULL, NULL));
1336: PetscCall(ISDestroy(&iscol));
1337: PetscCall(ISDestroy(&isrow));
1338: /*
1339: Everyone has to call to draw the matrix since the graphics waits are
1340: synchronized across all processors that share the PetscDraw object
1341: */
1342: PetscCall(PetscViewerGetSubViewer(viewer, PETSC_COMM_SELF, &sviewer));
1343: if (rank == 0) {
1344: if (((PetscObject)mat)->name) PetscCall(PetscObjectSetName((PetscObject)Av, ((PetscObject)mat)->name));
1345: PetscCall(MatView_SeqAIJ(Av, sviewer));
1346: }
1347: PetscCall(PetscViewerRestoreSubViewer(viewer, PETSC_COMM_SELF, &sviewer));
1348: PetscCall(MatDestroy(&A));
1349: }
1350: PetscFunctionReturn(PETSC_SUCCESS);
1351: }
1353: PetscErrorCode MatView_MPIAIJ(Mat mat, PetscViewer viewer)
1354: {
1355: PetscBool isascii, isdraw, issocket, isbinary;
1357: PetscFunctionBegin;
1358: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii));
1359: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw));
1360: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
1361: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERSOCKET, &issocket));
1362: if (isascii || isdraw || isbinary || issocket) PetscCall(MatView_MPIAIJ_ASCIIorDraworSocket(mat, viewer));
1363: PetscFunctionReturn(PETSC_SUCCESS);
1364: }
1366: static PetscErrorCode MatSOR_MPIAIJ(Mat matin, Vec bb, PetscReal omega, MatSORType flag, PetscReal fshift, PetscInt its, PetscInt lits, Vec xx)
1367: {
1368: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)matin->data;
1369: Vec bb1 = NULL;
1370: PetscBool hasop;
1372: PetscFunctionBegin;
1373: if (flag == SOR_APPLY_UPPER) {
1374: PetscUseTypeMethod(mat->A, sor, bb, omega, flag, fshift, lits, 1, xx);
1375: PetscFunctionReturn(PETSC_SUCCESS);
1376: }
1378: if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS || flag & SOR_EISENSTAT) PetscCall(VecDuplicate(bb, &bb1));
1380: if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) {
1381: if (flag & SOR_ZERO_INITIAL_GUESS) {
1382: PetscUseTypeMethod(mat->A, sor, bb, omega, flag, fshift, lits, 1, xx);
1383: its--;
1384: }
1386: while (its--) {
1387: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1388: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1390: /* update rhs: bb1 = bb - B*x */
1391: PetscCall(VecScale(mat->lvec, -1.0));
1392: PetscUseTypeMethod(mat->B, multadd, mat->lvec, bb, bb1);
1394: /* local sweep */
1395: PetscUseTypeMethod(mat->A, sor, bb1, omega, SOR_SYMMETRIC_SWEEP, fshift, lits, 1, xx);
1396: }
1397: } else if (flag & SOR_LOCAL_FORWARD_SWEEP) {
1398: if (flag & SOR_ZERO_INITIAL_GUESS) {
1399: PetscUseTypeMethod(mat->A, sor, bb, omega, flag, fshift, lits, 1, xx);
1400: its--;
1401: }
1402: while (its--) {
1403: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1404: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1406: /* update rhs: bb1 = bb - B*x */
1407: PetscCall(VecScale(mat->lvec, -1.0));
1408: PetscUseTypeMethod(mat->B, multadd, mat->lvec, bb, bb1);
1410: /* local sweep */
1411: PetscUseTypeMethod(mat->A, sor, bb1, omega, SOR_FORWARD_SWEEP, fshift, lits, 1, xx);
1412: }
1413: } else if (flag & SOR_LOCAL_BACKWARD_SWEEP) {
1414: if (flag & SOR_ZERO_INITIAL_GUESS) {
1415: PetscUseTypeMethod(mat->A, sor, bb, omega, flag, fshift, lits, 1, xx);
1416: its--;
1417: }
1418: while (its--) {
1419: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1420: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1422: /* update rhs: bb1 = bb - B*x */
1423: PetscCall(VecScale(mat->lvec, -1.0));
1424: PetscUseTypeMethod(mat->B, multadd, mat->lvec, bb, bb1);
1426: /* local sweep */
1427: PetscUseTypeMethod(mat->A, sor, bb1, omega, SOR_BACKWARD_SWEEP, fshift, lits, 1, xx);
1428: }
1429: } else if (flag & SOR_EISENSTAT) {
1430: Vec xx1;
1432: PetscCall(VecDuplicate(bb, &xx1));
1433: PetscUseTypeMethod(mat->A, sor, bb, omega, (MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_BACKWARD_SWEEP), fshift, lits, 1, xx);
1435: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1436: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1437: if (!mat->diag) {
1438: PetscCall(MatCreateVecs(matin, &mat->diag, NULL));
1439: PetscCall(MatGetDiagonal(matin, mat->diag));
1440: }
1441: PetscCall(MatHasOperation(matin, MATOP_MULT_DIAGONAL_BLOCK, &hasop));
1442: if (hasop) PetscCall(MatMultDiagonalBlock(matin, xx, bb1));
1443: else PetscCall(VecPointwiseMult(bb1, mat->diag, xx));
1444: PetscCall(VecAYPX(bb1, (omega - 2.0) / omega, bb));
1446: PetscCall(MatMultAdd(mat->B, mat->lvec, bb1, bb1));
1448: /* local sweep */
1449: PetscUseTypeMethod(mat->A, sor, bb1, omega, (MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_FORWARD_SWEEP), fshift, lits, 1, xx1);
1450: PetscCall(VecAXPY(xx, 1.0, xx1));
1451: PetscCall(VecDestroy(&xx1));
1452: } else SETERRQ(PetscObjectComm((PetscObject)matin), PETSC_ERR_SUP, "Parallel SOR not supported");
1454: PetscCall(VecDestroy(&bb1));
1456: matin->factorerrortype = mat->A->factorerrortype;
1457: PetscFunctionReturn(PETSC_SUCCESS);
1458: }
1460: static PetscErrorCode MatPermute_MPIAIJ(Mat A, IS rowp, IS colp, Mat *B)
1461: {
1462: Mat aA, aB, Aperm;
1463: const PetscInt *rwant, *cwant, *gcols, *ai, *bi, *aj, *bj;
1464: PetscScalar *aa, *ba;
1465: PetscInt i, j, m, n, ng, anz, bnz, *dnnz, *onnz, *tdnnz, *tonnz, *rdest, *cdest, *work, *gcdest;
1466: PetscSF rowsf, sf;
1467: IS parcolp = NULL;
1468: PetscBool done;
1470: PetscFunctionBegin;
1471: PetscCall(MatGetLocalSize(A, &m, &n));
1472: PetscCall(ISGetIndices(rowp, &rwant));
1473: PetscCall(ISGetIndices(colp, &cwant));
1474: PetscCall(PetscMalloc3(PetscMax(m, n), &work, m, &rdest, n, &cdest));
1476: /* Invert row permutation to find out where my rows should go */
1477: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &rowsf));
1478: PetscCall(PetscSFSetGraphLayout(rowsf, A->rmap, A->rmap->n, NULL, PETSC_OWN_POINTER, rwant));
1479: PetscCall(PetscSFSetFromOptions(rowsf));
1480: for (i = 0; i < m; i++) work[i] = A->rmap->rstart + i;
1481: PetscCall(PetscSFReduceBegin(rowsf, MPIU_INT, work, rdest, MPI_REPLACE));
1482: PetscCall(PetscSFReduceEnd(rowsf, MPIU_INT, work, rdest, MPI_REPLACE));
1484: /* Invert column permutation to find out where my columns should go */
1485: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
1486: PetscCall(PetscSFSetGraphLayout(sf, A->cmap, A->cmap->n, NULL, PETSC_OWN_POINTER, cwant));
1487: PetscCall(PetscSFSetFromOptions(sf));
1488: for (i = 0; i < n; i++) work[i] = A->cmap->rstart + i;
1489: PetscCall(PetscSFReduceBegin(sf, MPIU_INT, work, cdest, MPI_REPLACE));
1490: PetscCall(PetscSFReduceEnd(sf, MPIU_INT, work, cdest, MPI_REPLACE));
1491: PetscCall(PetscSFDestroy(&sf));
1493: PetscCall(ISRestoreIndices(rowp, &rwant));
1494: PetscCall(ISRestoreIndices(colp, &cwant));
1495: PetscCall(MatMPIAIJGetSeqAIJ(A, &aA, &aB, &gcols));
1497: /* Find out where my gcols should go */
1498: PetscCall(MatGetSize(aB, NULL, &ng));
1499: PetscCall(PetscMalloc1(ng, &gcdest));
1500: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
1501: PetscCall(PetscSFSetGraphLayout(sf, A->cmap, ng, NULL, PETSC_OWN_POINTER, gcols));
1502: PetscCall(PetscSFSetFromOptions(sf));
1503: PetscCall(PetscSFBcastBegin(sf, MPIU_INT, cdest, gcdest, MPI_REPLACE));
1504: PetscCall(PetscSFBcastEnd(sf, MPIU_INT, cdest, gcdest, MPI_REPLACE));
1505: PetscCall(PetscSFDestroy(&sf));
1507: PetscCall(PetscCalloc4(m, &dnnz, m, &onnz, m, &tdnnz, m, &tonnz));
1508: PetscCall(MatGetRowIJ(aA, 0, PETSC_FALSE, PETSC_FALSE, &anz, &ai, &aj, &done));
1509: PetscCall(MatGetRowIJ(aB, 0, PETSC_FALSE, PETSC_FALSE, &bnz, &bi, &bj, &done));
1510: for (i = 0; i < m; i++) {
1511: PetscInt row = rdest[i];
1512: PetscMPIInt rowner;
1513: PetscCall(PetscLayoutFindOwner(A->rmap, row, &rowner));
1514: for (j = ai[i]; j < ai[i + 1]; j++) {
1515: PetscInt col = cdest[aj[j]];
1516: PetscMPIInt cowner;
1517: PetscCall(PetscLayoutFindOwner(A->cmap, col, &cowner)); /* Could build an index for the columns to eliminate this search */
1518: if (rowner == cowner) dnnz[i]++;
1519: else onnz[i]++;
1520: }
1521: for (j = bi[i]; j < bi[i + 1]; j++) {
1522: PetscInt col = gcdest[bj[j]];
1523: PetscMPIInt cowner;
1524: PetscCall(PetscLayoutFindOwner(A->cmap, col, &cowner));
1525: if (rowner == cowner) dnnz[i]++;
1526: else onnz[i]++;
1527: }
1528: }
1529: PetscCall(PetscSFBcastBegin(rowsf, MPIU_INT, dnnz, tdnnz, MPI_REPLACE));
1530: PetscCall(PetscSFBcastEnd(rowsf, MPIU_INT, dnnz, tdnnz, MPI_REPLACE));
1531: PetscCall(PetscSFBcastBegin(rowsf, MPIU_INT, onnz, tonnz, MPI_REPLACE));
1532: PetscCall(PetscSFBcastEnd(rowsf, MPIU_INT, onnz, tonnz, MPI_REPLACE));
1533: PetscCall(PetscSFDestroy(&rowsf));
1535: PetscCall(MatCreateAIJ(PetscObjectComm((PetscObject)A), A->rmap->n, A->cmap->n, A->rmap->N, A->cmap->N, 0, tdnnz, 0, tonnz, &Aperm));
1536: PetscCall(MatSeqAIJGetArray(aA, &aa));
1537: PetscCall(MatSeqAIJGetArray(aB, &ba));
1538: for (i = 0; i < m; i++) {
1539: PetscInt *acols = dnnz, *bcols = onnz; /* Repurpose now-unneeded arrays */
1540: PetscInt rowlen;
1541: rowlen = ai[i + 1] - ai[i];
1542: for (PetscInt j0 = j = 0; j < rowlen; j0 = j) { /* rowlen could be larger than number of rows m, so sum in batches */
1543: for (; j < PetscMin(rowlen, j0 + m); j++) acols[j - j0] = cdest[aj[ai[i] + j]];
1544: PetscCall(MatSetValues(Aperm, 1, &rdest[i], j - j0, acols, aa + ai[i] + j0, INSERT_VALUES));
1545: }
1546: rowlen = bi[i + 1] - bi[i];
1547: for (PetscInt j0 = j = 0; j < rowlen; j0 = j) {
1548: for (; j < PetscMin(rowlen, j0 + m); j++) bcols[j - j0] = gcdest[bj[bi[i] + j]];
1549: PetscCall(MatSetValues(Aperm, 1, &rdest[i], j - j0, bcols, ba + bi[i] + j0, INSERT_VALUES));
1550: }
1551: }
1552: PetscCall(MatAssemblyBegin(Aperm, MAT_FINAL_ASSEMBLY));
1553: PetscCall(MatAssemblyEnd(Aperm, MAT_FINAL_ASSEMBLY));
1554: PetscCall(MatRestoreRowIJ(aA, 0, PETSC_FALSE, PETSC_FALSE, &anz, &ai, &aj, &done));
1555: PetscCall(MatRestoreRowIJ(aB, 0, PETSC_FALSE, PETSC_FALSE, &bnz, &bi, &bj, &done));
1556: PetscCall(MatSeqAIJRestoreArray(aA, &aa));
1557: PetscCall(MatSeqAIJRestoreArray(aB, &ba));
1558: PetscCall(PetscFree4(dnnz, onnz, tdnnz, tonnz));
1559: PetscCall(PetscFree3(work, rdest, cdest));
1560: PetscCall(PetscFree(gcdest));
1561: if (parcolp) PetscCall(ISDestroy(&colp));
1562: *B = Aperm;
1563: PetscFunctionReturn(PETSC_SUCCESS);
1564: }
1566: static PetscErrorCode MatGetGhosts_MPIAIJ(Mat mat, PetscInt *nghosts, const PetscInt *ghosts[])
1567: {
1568: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1570: PetscFunctionBegin;
1571: PetscCall(MatGetSize(aij->B, NULL, nghosts));
1572: if (ghosts) *ghosts = aij->garray;
1573: PetscFunctionReturn(PETSC_SUCCESS);
1574: }
1576: static PetscErrorCode MatGetInfo_MPIAIJ(Mat matin, MatInfoType flag, MatInfo *info)
1577: {
1578: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)matin->data;
1579: Mat A = mat->A, B = mat->B;
1580: PetscLogDouble irecv[5];
1582: PetscFunctionBegin;
1583: info->block_size = 1.0;
1584: PetscCall(MatGetInfo(A, MAT_LOCAL, info));
1586: irecv[0] = info->nz_used;
1587: irecv[1] = info->nz_allocated;
1588: irecv[2] = info->nz_unneeded;
1589: irecv[3] = info->memory;
1590: irecv[4] = info->mallocs;
1592: PetscCall(MatGetInfo(B, MAT_LOCAL, info));
1594: irecv[0] += info->nz_used;
1595: irecv[1] += info->nz_allocated;
1596: irecv[2] += info->nz_unneeded;
1597: irecv[3] += info->memory;
1598: irecv[4] += info->mallocs;
1599: if (flag == MAT_LOCAL) {
1600: info->nz_used = irecv[0];
1601: info->nz_allocated = irecv[1];
1602: info->nz_unneeded = irecv[2];
1603: info->memory = irecv[3];
1604: info->mallocs = irecv[4];
1605: } else if (flag == MAT_GLOBAL_MAX) {
1606: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, irecv, 5, MPIU_PETSCLOGDOUBLE, MPI_MAX, PetscObjectComm((PetscObject)matin)));
1608: info->nz_used = irecv[0];
1609: info->nz_allocated = irecv[1];
1610: info->nz_unneeded = irecv[2];
1611: info->memory = irecv[3];
1612: info->mallocs = irecv[4];
1613: } else if (flag == MAT_GLOBAL_SUM) {
1614: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, irecv, 5, MPIU_PETSCLOGDOUBLE, MPI_SUM, PetscObjectComm((PetscObject)matin)));
1616: info->nz_used = irecv[0];
1617: info->nz_allocated = irecv[1];
1618: info->nz_unneeded = irecv[2];
1619: info->memory = irecv[3];
1620: info->mallocs = irecv[4];
1621: }
1622: info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */
1623: info->fill_ratio_needed = 0;
1624: info->factor_mallocs = 0;
1625: PetscFunctionReturn(PETSC_SUCCESS);
1626: }
1628: PetscErrorCode MatSetOption_MPIAIJ(Mat A, MatOption op, PetscBool flg)
1629: {
1630: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1632: PetscFunctionBegin;
1633: switch (op) {
1634: case MAT_NEW_NONZERO_LOCATIONS:
1635: case MAT_NEW_NONZERO_ALLOCATION_ERR:
1636: case MAT_UNUSED_NONZERO_LOCATION_ERR:
1637: case MAT_KEEP_NONZERO_PATTERN:
1638: case MAT_NEW_NONZERO_LOCATION_ERR:
1639: case MAT_USE_INODES:
1640: case MAT_IGNORE_ZERO_ENTRIES:
1641: case MAT_FORM_EXPLICIT_TRANSPOSE:
1642: case MAT_ROW_ORIENTED:
1643: MatCheckPreallocated(A, 1);
1644: if (op == MAT_ROW_ORIENTED) a->roworiented = flg;
1645: PetscCall(MatSetOption(a->A, op, flg));
1646: PetscCall(MatSetOption(a->B, op, flg));
1647: break;
1648: case MAT_STRUCTURE_ONLY:
1649: if (a->A) PetscCall(MatSetOption(a->A, op, flg));
1650: if (a->B) PetscCall(MatSetOption(a->B, op, flg));
1651: break;
1652: case MAT_IGNORE_OFF_PROC_ENTRIES:
1653: a->donotstash = flg;
1654: break;
1655: /* Symmetry flags are handled directly by MatSetOption() and they don't affect preallocation */
1656: case MAT_SPD:
1657: case MAT_SYMMETRIC:
1658: case MAT_STRUCTURALLY_SYMMETRIC:
1659: case MAT_HERMITIAN:
1660: case MAT_SYMMETRY_ETERNAL:
1661: case MAT_STRUCTURAL_SYMMETRY_ETERNAL:
1662: case MAT_SPD_ETERNAL:
1663: /* if the diagonal matrix is square it inherits some of the properties above */
1664: if (a->A && A->rmap->n == A->cmap->n) PetscCall(MatSetOption(a->A, op, flg));
1665: break;
1666: case MAT_SUBMAT_SINGLEIS:
1667: A->submat_singleis = flg;
1668: break;
1669: default:
1670: break;
1671: }
1672: PetscFunctionReturn(PETSC_SUCCESS);
1673: }
1675: PetscErrorCode MatGetRow_MPIAIJ(Mat matin, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
1676: {
1677: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)matin->data;
1678: PetscScalar *vworkA, *vworkB, **pvA, **pvB, *v_p;
1679: PetscInt i, *cworkA, *cworkB, **pcA, **pcB, cstart = matin->cmap->rstart;
1680: PetscInt nztot, nzA, nzB, lrow, rstart = matin->rmap->rstart, rend = matin->rmap->rend;
1681: PetscInt *cmap, *idx_p;
1683: PetscFunctionBegin;
1684: PetscCheck(!mat->getrowactive, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Already active");
1685: mat->getrowactive = PETSC_TRUE;
1687: if ((!mat->rowindices && (idx || v)) || (!mat->rowvalues && v)) {
1688: /*
1689: allocate enough space to hold information from the longest row.
1690: */
1691: Mat_SeqAIJ *Aa = (Mat_SeqAIJ *)mat->A->data, *Ba = (Mat_SeqAIJ *)mat->B->data;
1692: PetscInt max = 1, tmp;
1693: for (i = 0; i < matin->rmap->n; i++) {
1694: tmp = Aa->i[i + 1] - Aa->i[i] + Ba->i[i + 1] - Ba->i[i];
1695: if (max < tmp) max = tmp;
1696: }
1697: PetscCall(PetscFree2(mat->rowvalues, mat->rowindices));
1698: PetscCall(PetscMalloc2(v ? max : 0, &mat->rowvalues, max, &mat->rowindices));
1699: }
1701: PetscCheck(row >= rstart && row < rend, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Only local rows");
1702: lrow = row - rstart;
1704: pvA = &vworkA;
1705: pcA = &cworkA;
1706: pvB = &vworkB;
1707: pcB = &cworkB;
1708: if (!v) {
1709: pvA = NULL;
1710: pvB = NULL;
1711: }
1712: if (!idx) {
1713: pcA = NULL;
1714: if (!v) pcB = NULL;
1715: }
1716: PetscUseTypeMethod(mat->A, getrow, lrow, &nzA, pcA, pvA);
1717: PetscUseTypeMethod(mat->B, getrow, lrow, &nzB, pcB, pvB);
1718: nztot = nzA + nzB;
1720: cmap = mat->garray;
1721: if (v || idx) {
1722: if (nztot) {
1723: /* Sort by increasing column numbers, assuming A and B already sorted */
1724: PetscInt imark = -1;
1725: if (v) {
1726: *v = v_p = mat->rowvalues;
1727: for (i = 0; i < nzB; i++) {
1728: if (cmap[cworkB[i]] < cstart) v_p[i] = vworkB[i];
1729: else break;
1730: }
1731: imark = i;
1732: for (i = 0; i < nzA; i++) v_p[imark + i] = vworkA[i];
1733: for (i = imark; i < nzB; i++) v_p[nzA + i] = vworkB[i];
1734: }
1735: if (idx) {
1736: *idx = idx_p = mat->rowindices;
1737: if (imark > -1) {
1738: for (i = 0; i < imark; i++) idx_p[i] = cmap[cworkB[i]];
1739: } else {
1740: for (i = 0; i < nzB; i++) {
1741: if (cmap[cworkB[i]] < cstart) idx_p[i] = cmap[cworkB[i]];
1742: else break;
1743: }
1744: imark = i;
1745: }
1746: for (i = 0; i < nzA; i++) idx_p[imark + i] = cstart + cworkA[i];
1747: for (i = imark; i < nzB; i++) idx_p[nzA + i] = cmap[cworkB[i]];
1748: }
1749: } else {
1750: if (idx) *idx = NULL;
1751: if (v) *v = NULL;
1752: }
1753: }
1754: *nz = nztot;
1755: PetscUseTypeMethod(mat->A, restorerow, lrow, &nzA, pcA, pvA);
1756: PetscUseTypeMethod(mat->B, restorerow, lrow, &nzB, pcB, pvB);
1757: PetscFunctionReturn(PETSC_SUCCESS);
1758: }
1760: PetscErrorCode MatRestoreRow_MPIAIJ(Mat mat, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
1761: {
1762: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1764: PetscFunctionBegin;
1765: PetscCheck(aij->getrowactive, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "MatGetRow() must be called first");
1766: aij->getrowactive = PETSC_FALSE;
1767: PetscFunctionReturn(PETSC_SUCCESS);
1768: }
1770: static PetscErrorCode MatNorm_MPIAIJ(Mat mat, NormType type, PetscReal *norm)
1771: {
1772: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1773: Mat_SeqAIJ *amat = (Mat_SeqAIJ *)aij->A->data, *bmat = (Mat_SeqAIJ *)aij->B->data;
1774: PetscInt i, j;
1775: PetscReal sum = 0.0;
1776: const MatScalar *v, *amata, *bmata;
1778: PetscFunctionBegin;
1779: if (aij->size == 1) {
1780: PetscCall(MatNorm(aij->A, type, norm));
1781: } else {
1782: PetscCall(MatSeqAIJGetArrayRead(aij->A, &amata));
1783: PetscCall(MatSeqAIJGetArrayRead(aij->B, &bmata));
1784: if (type == NORM_FROBENIUS) {
1785: v = amata;
1786: for (i = 0; i < amat->nz; i++) {
1787: sum += PetscRealPart(PetscConj(*v) * (*v));
1788: v++;
1789: }
1790: v = bmata;
1791: for (i = 0; i < bmat->nz; i++) {
1792: sum += PetscRealPart(PetscConj(*v) * (*v));
1793: v++;
1794: }
1795: PetscCallMPI(MPIU_Allreduce(&sum, norm, 1, MPIU_REAL, MPIU_SUM, PetscObjectComm((PetscObject)mat)));
1796: *norm = PetscSqrtReal(*norm);
1797: PetscCall(PetscLogFlops(2.0 * amat->nz + 2.0 * bmat->nz));
1798: } else if (type == NORM_1) { /* max column norm */
1799: Vec col, bcol;
1800: PetscScalar *array;
1801: PetscInt *jj, *garray = aij->garray;
1803: PetscCall(MatCreateVecs(mat, &col, NULL));
1804: PetscCall(VecGetArrayWrite(col, &array));
1805: v = amata;
1806: jj = amat->j;
1807: for (j = 0; j < amat->nz; j++) array[*jj++] += PetscAbsScalar(*v++);
1808: PetscCall(VecRestoreArrayWrite(col, &array));
1809: PetscCall(MatCreateVecs(aij->B, &bcol, NULL));
1810: PetscCall(VecGetArrayWrite(bcol, &array));
1811: v = bmata;
1812: jj = bmat->j;
1813: for (j = 0; j < bmat->nz; j++) array[*jj++] += PetscAbsScalar(*v++);
1814: PetscCall(VecSetValues(col, aij->B->cmap->n, garray, array, ADD_VALUES));
1815: PetscCall(VecRestoreArrayWrite(bcol, &array));
1816: PetscCall(VecDestroy(&bcol));
1817: PetscCall(VecAssemblyBegin(col));
1818: PetscCall(VecAssemblyEnd(col));
1819: PetscCall(VecNorm(col, NORM_INFINITY, norm));
1820: PetscCall(VecDestroy(&col));
1821: } else if (type == NORM_INFINITY) { /* max row norm */
1822: *norm = 0.0;
1823: for (j = 0; j < aij->A->rmap->n; j++) {
1824: v = PetscSafePointerPlusOffset(amata, amat->i[j]);
1825: sum = 0.0;
1826: for (i = 0; i < amat->i[j + 1] - amat->i[j]; i++) {
1827: sum += PetscAbsScalar(*v);
1828: v++;
1829: }
1830: v = PetscSafePointerPlusOffset(bmata, bmat->i[j]);
1831: for (i = 0; i < bmat->i[j + 1] - bmat->i[j]; i++) {
1832: sum += PetscAbsScalar(*v);
1833: v++;
1834: }
1835: if (sum > *norm) *norm = sum;
1836: }
1837: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, norm, 1, MPIU_REAL, MPIU_MAX, PetscObjectComm((PetscObject)mat)));
1838: PetscCall(PetscLogFlops(PetscMax(amat->nz + bmat->nz - 1, 0)));
1839: } else SETERRQ(PetscObjectComm((PetscObject)mat), PETSC_ERR_SUP, "No support for two norm");
1840: PetscCall(MatSeqAIJRestoreArrayRead(aij->A, &amata));
1841: PetscCall(MatSeqAIJRestoreArrayRead(aij->B, &bmata));
1842: }
1843: PetscFunctionReturn(PETSC_SUCCESS);
1844: }
1846: static PetscErrorCode MatTranspose_MPIAIJ(Mat A, MatReuse reuse, Mat *matout)
1847: {
1848: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data, *b;
1849: Mat_SeqAIJ *Aloc = (Mat_SeqAIJ *)a->A->data, *Bloc = (Mat_SeqAIJ *)a->B->data, *sub_B_diag;
1850: 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;
1851: const PetscInt *ai, *aj, *bi, *bj, *B_diag_i;
1852: Mat B, A_diag, *B_diag;
1853: const MatScalar *pbv, *bv;
1855: PetscFunctionBegin;
1856: if (reuse == MAT_REUSE_MATRIX) PetscCall(MatTransposeCheckNonzeroState_Private(A, *matout));
1857: ma = A->rmap->n;
1858: na = A->cmap->n;
1859: mb = a->B->rmap->n;
1860: nb = a->B->cmap->n;
1861: ai = Aloc->i;
1862: aj = Aloc->j;
1863: bi = Bloc->i;
1864: bj = Bloc->j;
1865: if (reuse == MAT_INITIAL_MATRIX || *matout == A) {
1866: PetscInt *d_nnz, *g_nnz, *o_nnz;
1867: PetscSFNode *oloc;
1868: PETSC_UNUSED PetscSF sf;
1870: PetscCall(PetscMalloc4(na, &d_nnz, na, &o_nnz, nb, &g_nnz, nb, &oloc));
1871: /* compute d_nnz for preallocation */
1872: PetscCall(PetscArrayzero(d_nnz, na));
1873: for (i = 0; i < ai[ma]; i++) d_nnz[aj[i]]++;
1874: /* compute local off-diagonal contributions */
1875: PetscCall(PetscArrayzero(g_nnz, nb));
1876: for (i = 0; i < bi[ma]; i++) g_nnz[bj[i]]++;
1877: /* map those to global */
1878: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
1879: PetscCall(PetscSFSetGraphLayout(sf, A->cmap, nb, NULL, PETSC_USE_POINTER, a->garray));
1880: PetscCall(PetscSFSetFromOptions(sf));
1881: PetscCall(PetscArrayzero(o_nnz, na));
1882: PetscCall(PetscSFReduceBegin(sf, MPIU_INT, g_nnz, o_nnz, MPI_SUM));
1883: PetscCall(PetscSFReduceEnd(sf, MPIU_INT, g_nnz, o_nnz, MPI_SUM));
1884: PetscCall(PetscSFDestroy(&sf));
1886: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &B));
1887: PetscCall(MatSetSizes(B, A->cmap->n, A->rmap->n, N, M));
1888: PetscCall(MatSetBlockSizes(B, A->cmap->bs, A->rmap->bs));
1889: PetscCall(MatSetType(B, ((PetscObject)A)->type_name));
1890: PetscCall(MatMPIAIJSetPreallocation(B, 0, d_nnz, 0, o_nnz));
1891: PetscCall(PetscFree4(d_nnz, o_nnz, g_nnz, oloc));
1892: } else {
1893: B = *matout;
1894: PetscCall(MatSetOption(B, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_TRUE));
1895: }
1897: b = (Mat_MPIAIJ *)B->data;
1898: A_diag = a->A;
1899: B_diag = &b->A;
1900: sub_B_diag = (Mat_SeqAIJ *)(*B_diag)->data;
1901: A_diag_ncol = A_diag->cmap->N;
1902: B_diag_ilen = sub_B_diag->ilen;
1903: B_diag_i = sub_B_diag->i;
1905: /* Set ilen for diagonal of B */
1906: for (i = 0; i < A_diag_ncol; i++) B_diag_ilen[i] = B_diag_i[i + 1] - B_diag_i[i];
1908: /* Transpose the diagonal part of the matrix. In contrast to the off-diagonal part, this can be done
1909: very quickly (=without using MatSetValues), because all writes are local. */
1910: PetscCall(MatTransposeSetPrecursor(A_diag, *B_diag));
1911: PetscCall(MatTranspose(A_diag, MAT_REUSE_MATRIX, B_diag));
1913: /* copy over the B part */
1914: PetscCall(PetscMalloc1(bi[mb], &cols));
1915: PetscCall(MatSeqAIJGetArrayRead(a->B, &bv));
1916: pbv = bv;
1917: row = A->rmap->rstart;
1918: for (i = 0; i < bi[mb]; i++) cols[i] = a->garray[bj[i]];
1919: cols_tmp = cols;
1920: for (i = 0; i < mb; i++) {
1921: ncol = bi[i + 1] - bi[i];
1922: PetscCall(MatSetValues(B, ncol, cols_tmp, 1, &row, pbv, INSERT_VALUES));
1923: row++;
1924: if (pbv) pbv += ncol;
1925: if (cols_tmp) cols_tmp += ncol;
1926: }
1927: PetscCall(PetscFree(cols));
1928: PetscCall(MatSeqAIJRestoreArrayRead(a->B, &bv));
1930: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
1931: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
1932: if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX) {
1933: *matout = B;
1934: } else {
1935: PetscCall(MatHeaderMerge(A, &B));
1936: }
1937: PetscFunctionReturn(PETSC_SUCCESS);
1938: }
1940: static PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat, Vec ll, Vec rr)
1941: {
1942: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1943: Mat a = aij->A, b = aij->B;
1944: PetscInt s1, s2, s3;
1946: PetscFunctionBegin;
1947: PetscCall(MatGetLocalSize(mat, &s2, &s3));
1948: if (rr) {
1949: PetscCall(VecGetLocalSize(rr, &s1));
1950: PetscCheck(s1 == s3, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "right vector non-conforming local size");
1951: /* Overlap communication with computation. */
1952: PetscCall(VecScatterBegin(aij->Mvctx, rr, aij->lvec, INSERT_VALUES, SCATTER_FORWARD));
1953: }
1954: if (ll) {
1955: PetscCall(VecGetLocalSize(ll, &s1));
1956: PetscCheck(s1 == s2, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "left vector non-conforming local size");
1957: PetscUseTypeMethod(b, diagonalscale, ll, NULL);
1958: }
1959: /* scale the diagonal block */
1960: PetscUseTypeMethod(a, diagonalscale, ll, rr);
1962: if (rr) {
1963: /* Do a scatter end and then right scale the off-diagonal block */
1964: PetscCall(VecScatterEnd(aij->Mvctx, rr, aij->lvec, INSERT_VALUES, SCATTER_FORWARD));
1965: PetscUseTypeMethod(b, diagonalscale, NULL, aij->lvec);
1966: }
1967: PetscFunctionReturn(PETSC_SUCCESS);
1968: }
1970: static PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A)
1971: {
1972: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1974: PetscFunctionBegin;
1975: PetscCall(MatSetUnfactored(a->A));
1976: PetscFunctionReturn(PETSC_SUCCESS);
1977: }
1979: static PetscErrorCode MatEqual_MPIAIJ(Mat A, Mat B, PetscBool *flag)
1980: {
1981: Mat_MPIAIJ *matB = (Mat_MPIAIJ *)B->data, *matA = (Mat_MPIAIJ *)A->data;
1982: Mat a, b, c, d;
1984: PetscFunctionBegin;
1985: a = matA->A;
1986: b = matA->B;
1987: c = matB->A;
1988: d = matB->B;
1990: PetscCall(MatEqual(a, c, flag));
1991: if (*flag) PetscCall(MatEqual(b, d, flag));
1992: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, flag, 1, MPI_C_BOOL, MPI_LAND, PetscObjectComm((PetscObject)A)));
1993: PetscFunctionReturn(PETSC_SUCCESS);
1994: }
1996: static PetscErrorCode MatCopy_MPIAIJ(Mat A, Mat B, MatStructure str)
1997: {
1998: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1999: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2001: PetscFunctionBegin;
2002: /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */
2003: if (str != SAME_NONZERO_PATTERN || A->ops->copy != B->ops->copy) {
2004: /* because of the column compression in the off-processor part of the matrix a->B,
2005: the number of columns in a->B and b->B may be different, hence we cannot call
2006: the MatCopy() directly on the two parts. If need be, we can provide a more
2007: efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices
2008: then copying the submatrices */
2009: PetscCall(MatCopy_Basic(A, B, str));
2010: } else {
2011: PetscCall(MatCopy(a->A, b->A, str));
2012: PetscCall(MatCopy(a->B, b->B, str));
2013: }
2014: PetscCall(PetscObjectStateIncrease((PetscObject)B));
2015: PetscFunctionReturn(PETSC_SUCCESS);
2016: }
2018: /*
2019: Computes the number of nonzeros per row needed for preallocation when X and Y
2020: have different nonzero structure.
2021: */
2022: 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)
2023: {
2024: PetscInt i, j, k, nzx, nzy;
2026: PetscFunctionBegin;
2027: /* Set the number of nonzeros in the new matrix */
2028: for (i = 0; i < m; i++) {
2029: const PetscInt *xjj = PetscSafePointerPlusOffset(xj, xi[i]), *yjj = PetscSafePointerPlusOffset(yj, yi[i]);
2030: nzx = xi[i + 1] - xi[i];
2031: nzy = yi[i + 1] - yi[i];
2032: nnz[i] = 0;
2033: for (j = 0, k = 0; j < nzx; j++) { /* Point in X */
2034: for (; k < nzy && yltog[yjj[k]] < xltog[xjj[j]]; k++) nnz[i]++; /* Catch up to X */
2035: if (k < nzy && yltog[yjj[k]] == xltog[xjj[j]]) k++; /* Skip duplicate */
2036: nnz[i]++;
2037: }
2038: for (; k < nzy; k++) nnz[i]++;
2039: }
2040: PetscFunctionReturn(PETSC_SUCCESS);
2041: }
2043: /* This is the same as MatAXPYGetPreallocation_SeqAIJ, except that the local-to-global map is provided */
2044: static PetscErrorCode MatAXPYGetPreallocation_MPIAIJ(Mat Y, const PetscInt *yltog, Mat X, const PetscInt *xltog, PetscInt *nnz)
2045: {
2046: PetscInt m = Y->rmap->N;
2047: Mat_SeqAIJ *x = (Mat_SeqAIJ *)X->data;
2048: Mat_SeqAIJ *y = (Mat_SeqAIJ *)Y->data;
2050: PetscFunctionBegin;
2051: PetscCall(MatAXPYGetPreallocation_MPIX_private(m, x->i, x->j, xltog, y->i, y->j, yltog, nnz));
2052: PetscFunctionReturn(PETSC_SUCCESS);
2053: }
2055: static PetscErrorCode MatAXPY_MPIAIJ(Mat Y, PetscScalar a, Mat X, MatStructure str)
2056: {
2057: Mat_MPIAIJ *xx = (Mat_MPIAIJ *)X->data, *yy = (Mat_MPIAIJ *)Y->data;
2059: PetscFunctionBegin;
2060: if (str == SAME_NONZERO_PATTERN) {
2061: PetscCall(MatAXPY(yy->A, a, xx->A, str));
2062: PetscCall(MatAXPY(yy->B, a, xx->B, str));
2063: } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */
2064: PetscCall(MatAXPY_Basic(Y, a, X, str));
2065: } else {
2066: Mat B;
2067: PetscInt *nnz_d, *nnz_o;
2068: PetscBool subset = PETSC_TRUE;
2069: Mat_SeqAIJ *ya = (Mat_SeqAIJ *)yy->A->data, *yb = (Mat_SeqAIJ *)yy->B->data;
2071: PetscCall(PetscMalloc1(yy->A->rmap->N, &nnz_d));
2072: PetscCall(PetscMalloc1(yy->B->rmap->N, &nnz_o));
2073: PetscCall(MatAXPYGetPreallocation_SeqAIJ(yy->A, xx->A, nnz_d));
2074: PetscCall(MatAXPYGetPreallocation_MPIAIJ(yy->B, yy->garray, xx->B, xx->garray, nnz_o));
2075: /* nnz_d[i] and nnz_o[i] are the sizes of the unions of row i of the blocks of X and Y, so the pattern of X is contained in
2076: that of Y when, on every process, they equal the lengths of row i of the blocks of Y. Adding in place then keeps the nonzero
2077: pattern of Y, and its nonzero state, unchanged. */
2078: for (PetscInt i = 0; i < yy->A->rmap->N && subset; i++) subset = (PetscBool)(nnz_d[i] == ya->i[i + 1] - ya->i[i] && nnz_o[i] == yb->i[i + 1] - yb->i[i]);
2079: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &subset, 1, MPI_C_BOOL, MPI_LAND, PetscObjectComm((PetscObject)Y)));
2080: if (subset) PetscCall(MatAXPY_Basic(Y, a, X, SUBSET_NONZERO_PATTERN));
2081: else {
2082: PetscCall(MatCreate(PetscObjectComm((PetscObject)Y), &B));
2083: PetscCall(PetscObjectSetName((PetscObject)B, ((PetscObject)Y)->name));
2084: PetscCall(MatSetLayouts(B, Y->rmap, Y->cmap));
2085: PetscCall(MatSetType(B, ((PetscObject)Y)->type_name));
2086: PetscCall(MatMPIAIJSetPreallocation(B, 0, nnz_d, 0, nnz_o));
2087: PetscCall(MatAXPY_BasicWithPreallocation(B, Y, a, X, str));
2088: PetscCall(MatHeaderMerge(Y, &B));
2089: }
2090: PetscCall(PetscFree(nnz_d));
2091: PetscCall(PetscFree(nnz_o));
2092: }
2093: PetscFunctionReturn(PETSC_SUCCESS);
2094: }
2096: PETSC_INTERN PetscErrorCode MatConjugate_SeqAIJ(Mat);
2098: static PetscErrorCode MatConjugate_MPIAIJ(Mat mat)
2099: {
2100: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2102: PetscFunctionBegin;
2103: PetscCall(MatConjugate_SeqAIJ(aij->A));
2104: PetscCall(MatConjugate_SeqAIJ(aij->B));
2105: PetscFunctionReturn(PETSC_SUCCESS);
2106: }
2108: static PetscErrorCode MatRealPart_MPIAIJ(Mat A)
2109: {
2110: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2112: PetscFunctionBegin;
2113: PetscCall(MatRealPart(a->A));
2114: PetscCall(MatRealPart(a->B));
2115: PetscFunctionReturn(PETSC_SUCCESS);
2116: }
2118: static PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A)
2119: {
2120: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2122: PetscFunctionBegin;
2123: PetscCall(MatImaginaryPart(a->A));
2124: PetscCall(MatImaginaryPart(a->B));
2125: PetscFunctionReturn(PETSC_SUCCESS);
2126: }
2128: static PetscErrorCode MatGetRowMaxAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2129: {
2130: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2131: PetscInt i, *idxb = NULL, m = A->rmap->n;
2132: PetscScalar *vv;
2133: Vec vB, vA;
2134: const PetscScalar *va, *vb;
2136: PetscFunctionBegin;
2137: PetscCall(MatCreateVecs(a->A, NULL, &vA));
2138: PetscCall(MatGetRowMaxAbs(a->A, vA, idx));
2140: PetscCall(VecGetArrayRead(vA, &va));
2141: if (idx) {
2142: for (i = 0; i < m; i++) {
2143: if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart;
2144: }
2145: }
2147: PetscCall(MatCreateVecs(a->B, NULL, &vB));
2148: PetscCall(PetscMalloc1(m, &idxb));
2149: PetscCall(MatGetRowMaxAbs(a->B, vB, idxb));
2151: PetscCall(VecGetArrayWrite(v, &vv));
2152: PetscCall(VecGetArrayRead(vB, &vb));
2153: for (i = 0; i < m; i++) {
2154: if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) {
2155: vv[i] = vb[i];
2156: if (idx) idx[i] = a->garray[idxb[i]];
2157: } else {
2158: vv[i] = va[i];
2159: if (idx && PetscAbsScalar(va[i]) == PetscAbsScalar(vb[i]) && idxb[i] != -1 && idx[i] > a->garray[idxb[i]]) idx[i] = a->garray[idxb[i]];
2160: }
2161: }
2162: PetscCall(VecRestoreArrayWrite(v, &vv));
2163: PetscCall(VecRestoreArrayRead(vA, &va));
2164: PetscCall(VecRestoreArrayRead(vB, &vb));
2165: PetscCall(PetscFree(idxb));
2166: PetscCall(VecDestroy(&vA));
2167: PetscCall(VecDestroy(&vB));
2168: PetscFunctionReturn(PETSC_SUCCESS);
2169: }
2171: static PetscErrorCode MatGetRowSumAbs_MPIAIJ(Mat A, Vec v)
2172: {
2173: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2174: Vec vB, vA;
2176: PetscFunctionBegin;
2177: PetscCall(MatCreateVecs(a->A, NULL, &vA));
2178: PetscCall(MatGetRowSumAbs(a->A, vA));
2179: PetscCall(MatCreateVecs(a->B, NULL, &vB));
2180: PetscCall(MatGetRowSumAbs(a->B, vB));
2181: PetscCall(VecAXPY(vA, 1.0, vB));
2182: PetscCall(VecDestroy(&vB));
2183: PetscCall(VecCopy(vA, v));
2184: PetscCall(VecDestroy(&vA));
2185: PetscFunctionReturn(PETSC_SUCCESS);
2186: }
2188: static PetscErrorCode MatGetRowMinAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2189: {
2190: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
2191: PetscInt m = A->rmap->n, n = A->cmap->n;
2192: PetscInt cstart = A->cmap->rstart, cend = A->cmap->rend;
2193: PetscInt *cmap = mat->garray;
2194: PetscInt *diagIdx, *offdiagIdx;
2195: Vec diagV, offdiagV;
2196: PetscScalar *a, *diagA, *offdiagA;
2197: const PetscScalar *ba, *bav;
2198: PetscInt r, j, col, ncols, *bi, *bj;
2199: Mat B = mat->B;
2200: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2202: PetscFunctionBegin;
2203: /* When a process holds entire A and other processes have no entry */
2204: if (A->cmap->N == n) {
2205: PetscCall(VecGetArrayWrite(v, &diagA));
2206: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2207: PetscCall(MatGetRowMinAbs(mat->A, diagV, idx));
2208: PetscCall(VecDestroy(&diagV));
2209: PetscCall(VecRestoreArrayWrite(v, &diagA));
2210: PetscFunctionReturn(PETSC_SUCCESS);
2211: } else if (n == 0) {
2212: if (m) {
2213: PetscCall(VecGetArrayWrite(v, &a));
2214: for (r = 0; r < m; r++) {
2215: a[r] = 0.0;
2216: if (idx) idx[r] = -1;
2217: }
2218: PetscCall(VecRestoreArrayWrite(v, &a));
2219: }
2220: PetscFunctionReturn(PETSC_SUCCESS);
2221: }
2223: PetscCall(PetscMalloc2(m, &diagIdx, m, &offdiagIdx));
2224: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2225: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2226: PetscCall(MatGetRowMinAbs(mat->A, diagV, diagIdx));
2228: /* Get offdiagIdx[] for implicit 0.0 */
2229: PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2230: ba = bav;
2231: bi = b->i;
2232: bj = b->j;
2233: PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2234: for (r = 0; r < m; r++) {
2235: ncols = bi[r + 1] - bi[r];
2236: if (ncols == A->cmap->N - n) { /* Brow is dense */
2237: offdiagA[r] = *ba;
2238: offdiagIdx[r] = cmap[0];
2239: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2240: offdiagA[r] = 0.0;
2242: /* Find first hole in the cmap */
2243: for (j = 0; j < ncols; j++) {
2244: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2245: if (col > j && j < cstart) {
2246: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2247: break;
2248: } else if (col > j + n && j >= cstart) {
2249: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2250: break;
2251: }
2252: }
2253: if (j == ncols && ncols < A->cmap->N - n) {
2254: /* a hole is outside compressed Bcols */
2255: if (ncols == 0) {
2256: if (cstart) {
2257: offdiagIdx[r] = 0;
2258: } else offdiagIdx[r] = cend;
2259: } else { /* ncols > 0 */
2260: offdiagIdx[r] = cmap[ncols - 1] + 1;
2261: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2262: }
2263: }
2264: }
2266: for (j = 0; j < ncols; j++) {
2267: if (PetscAbsScalar(offdiagA[r]) > PetscAbsScalar(*ba)) {
2268: offdiagA[r] = *ba;
2269: offdiagIdx[r] = cmap[*bj];
2270: }
2271: ba++;
2272: bj++;
2273: }
2274: }
2276: PetscCall(VecGetArrayWrite(v, &a));
2277: PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2278: for (r = 0; r < m; ++r) {
2279: if (PetscAbsScalar(diagA[r]) < PetscAbsScalar(offdiagA[r])) {
2280: a[r] = diagA[r];
2281: if (idx) idx[r] = cstart + diagIdx[r];
2282: } else if (PetscAbsScalar(diagA[r]) == PetscAbsScalar(offdiagA[r])) {
2283: a[r] = diagA[r];
2284: if (idx) {
2285: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2286: idx[r] = cstart + diagIdx[r];
2287: } else idx[r] = offdiagIdx[r];
2288: }
2289: } else {
2290: a[r] = offdiagA[r];
2291: if (idx) idx[r] = offdiagIdx[r];
2292: }
2293: }
2294: PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2295: PetscCall(VecRestoreArrayWrite(v, &a));
2296: PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2297: PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2298: PetscCall(VecDestroy(&diagV));
2299: PetscCall(VecDestroy(&offdiagV));
2300: PetscCall(PetscFree2(diagIdx, offdiagIdx));
2301: PetscFunctionReturn(PETSC_SUCCESS);
2302: }
2304: static PetscErrorCode MatGetRowMin_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2305: {
2306: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
2307: PetscInt m = A->rmap->n, n = A->cmap->n;
2308: PetscInt cstart = A->cmap->rstart, cend = A->cmap->rend;
2309: PetscInt *cmap = mat->garray;
2310: PetscInt *diagIdx, *offdiagIdx;
2311: Vec diagV, offdiagV;
2312: PetscScalar *a, *diagA, *offdiagA;
2313: const PetscScalar *ba, *bav;
2314: PetscInt r, j, col, ncols, *bi, *bj;
2315: Mat B = mat->B;
2316: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2318: PetscFunctionBegin;
2319: /* When a process holds entire A and other processes have no entry */
2320: if (A->cmap->N == n) {
2321: PetscCall(VecGetArrayWrite(v, &diagA));
2322: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2323: PetscCall(MatGetRowMin(mat->A, diagV, idx));
2324: PetscCall(VecDestroy(&diagV));
2325: PetscCall(VecRestoreArrayWrite(v, &diagA));
2326: PetscFunctionReturn(PETSC_SUCCESS);
2327: } else if (n == 0) {
2328: if (m) {
2329: PetscCall(VecGetArrayWrite(v, &a));
2330: for (r = 0; r < m; r++) {
2331: a[r] = PETSC_MAX_REAL;
2332: if (idx) idx[r] = -1;
2333: }
2334: PetscCall(VecRestoreArrayWrite(v, &a));
2335: }
2336: PetscFunctionReturn(PETSC_SUCCESS);
2337: }
2339: PetscCall(PetscCalloc2(m, &diagIdx, m, &offdiagIdx));
2340: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2341: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2342: PetscCall(MatGetRowMin(mat->A, diagV, diagIdx));
2344: /* Get offdiagIdx[] for implicit 0.0 */
2345: PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2346: ba = bav;
2347: bi = b->i;
2348: bj = b->j;
2349: PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2350: for (r = 0; r < m; r++) {
2351: ncols = bi[r + 1] - bi[r];
2352: if (ncols == A->cmap->N - n) { /* Brow is dense */
2353: offdiagA[r] = *ba;
2354: offdiagIdx[r] = cmap[0];
2355: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2356: offdiagA[r] = 0.0;
2358: /* Find first hole in the cmap */
2359: for (j = 0; j < ncols; j++) {
2360: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2361: if (col > j && j < cstart) {
2362: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2363: break;
2364: } else if (col > j + n && j >= cstart) {
2365: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2366: break;
2367: }
2368: }
2369: if (j == ncols && ncols < A->cmap->N - n) {
2370: /* a hole is outside compressed Bcols */
2371: if (ncols == 0) {
2372: if (cstart) {
2373: offdiagIdx[r] = 0;
2374: } else offdiagIdx[r] = cend;
2375: } else { /* ncols > 0 */
2376: offdiagIdx[r] = cmap[ncols - 1] + 1;
2377: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2378: }
2379: }
2380: }
2382: for (j = 0; j < ncols; j++) {
2383: if (PetscRealPart(offdiagA[r]) > PetscRealPart(*ba)) {
2384: offdiagA[r] = *ba;
2385: offdiagIdx[r] = cmap[*bj];
2386: }
2387: ba++;
2388: bj++;
2389: }
2390: }
2392: PetscCall(VecGetArrayWrite(v, &a));
2393: PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2394: for (r = 0; r < m; ++r) {
2395: if (PetscRealPart(diagA[r]) < PetscRealPart(offdiagA[r])) {
2396: a[r] = diagA[r];
2397: if (idx) idx[r] = cstart + diagIdx[r];
2398: } else if (PetscRealPart(diagA[r]) == PetscRealPart(offdiagA[r])) {
2399: a[r] = diagA[r];
2400: if (idx) {
2401: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2402: idx[r] = cstart + diagIdx[r];
2403: } else idx[r] = offdiagIdx[r];
2404: }
2405: } else {
2406: a[r] = offdiagA[r];
2407: if (idx) idx[r] = offdiagIdx[r];
2408: }
2409: }
2410: PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2411: PetscCall(VecRestoreArrayWrite(v, &a));
2412: PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2413: PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2414: PetscCall(VecDestroy(&diagV));
2415: PetscCall(VecDestroy(&offdiagV));
2416: PetscCall(PetscFree2(diagIdx, offdiagIdx));
2417: PetscFunctionReturn(PETSC_SUCCESS);
2418: }
2420: static PetscErrorCode MatGetRowMax_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2421: {
2422: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
2423: PetscInt m = A->rmap->n, n = A->cmap->n;
2424: PetscInt cstart = A->cmap->rstart, cend = A->cmap->rend;
2425: PetscInt *cmap = mat->garray;
2426: PetscInt *diagIdx, *offdiagIdx;
2427: Vec diagV, offdiagV;
2428: PetscScalar *a, *diagA, *offdiagA;
2429: const PetscScalar *ba, *bav;
2430: PetscInt r, j, col, ncols, *bi, *bj;
2431: Mat B = mat->B;
2432: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2434: PetscFunctionBegin;
2435: /* When a process holds entire A and other processes have no entry */
2436: if (A->cmap->N == n) {
2437: PetscCall(VecGetArrayWrite(v, &diagA));
2438: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2439: PetscCall(MatGetRowMax(mat->A, diagV, idx));
2440: PetscCall(VecDestroy(&diagV));
2441: PetscCall(VecRestoreArrayWrite(v, &diagA));
2442: PetscFunctionReturn(PETSC_SUCCESS);
2443: } else if (n == 0) {
2444: if (m) {
2445: PetscCall(VecGetArrayWrite(v, &a));
2446: for (r = 0; r < m; r++) {
2447: a[r] = PETSC_MIN_REAL;
2448: if (idx) idx[r] = -1;
2449: }
2450: PetscCall(VecRestoreArrayWrite(v, &a));
2451: }
2452: PetscFunctionReturn(PETSC_SUCCESS);
2453: }
2455: PetscCall(PetscMalloc2(m, &diagIdx, m, &offdiagIdx));
2456: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2457: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2458: PetscCall(MatGetRowMax(mat->A, diagV, diagIdx));
2460: /* Get offdiagIdx[] for implicit 0.0 */
2461: PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2462: ba = bav;
2463: bi = b->i;
2464: bj = b->j;
2465: PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2466: for (r = 0; r < m; r++) {
2467: ncols = bi[r + 1] - bi[r];
2468: if (ncols == A->cmap->N - n) { /* Brow is dense */
2469: offdiagA[r] = *ba;
2470: offdiagIdx[r] = cmap[0];
2471: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2472: offdiagA[r] = 0.0;
2474: /* Find first hole in the cmap */
2475: for (j = 0; j < ncols; j++) {
2476: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2477: if (col > j && j < cstart) {
2478: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2479: break;
2480: } else if (col > j + n && j >= cstart) {
2481: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2482: break;
2483: }
2484: }
2485: if (j == ncols && ncols < A->cmap->N - n) {
2486: /* a hole is outside compressed Bcols */
2487: if (ncols == 0) {
2488: if (cstart) {
2489: offdiagIdx[r] = 0;
2490: } else offdiagIdx[r] = cend;
2491: } else { /* ncols > 0 */
2492: offdiagIdx[r] = cmap[ncols - 1] + 1;
2493: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2494: }
2495: }
2496: }
2498: for (j = 0; j < ncols; j++) {
2499: if (PetscRealPart(offdiagA[r]) < PetscRealPart(*ba)) {
2500: offdiagA[r] = *ba;
2501: offdiagIdx[r] = cmap[*bj];
2502: }
2503: ba++;
2504: bj++;
2505: }
2506: }
2508: PetscCall(VecGetArrayWrite(v, &a));
2509: PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2510: for (r = 0; r < m; ++r) {
2511: if (PetscRealPart(diagA[r]) > PetscRealPart(offdiagA[r])) {
2512: a[r] = diagA[r];
2513: if (idx) idx[r] = cstart + diagIdx[r];
2514: } else if (PetscRealPart(diagA[r]) == PetscRealPart(offdiagA[r])) {
2515: a[r] = diagA[r];
2516: if (idx) {
2517: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2518: idx[r] = cstart + diagIdx[r];
2519: } else idx[r] = offdiagIdx[r];
2520: }
2521: } else {
2522: a[r] = offdiagA[r];
2523: if (idx) idx[r] = offdiagIdx[r];
2524: }
2525: }
2526: PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2527: PetscCall(VecRestoreArrayWrite(v, &a));
2528: PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2529: PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2530: PetscCall(VecDestroy(&diagV));
2531: PetscCall(VecDestroy(&offdiagV));
2532: PetscCall(PetscFree2(diagIdx, offdiagIdx));
2533: PetscFunctionReturn(PETSC_SUCCESS);
2534: }
2536: PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat mat, Mat *newmat)
2537: {
2538: Mat *dummy;
2540: PetscFunctionBegin;
2541: PetscCall(MatCreateSubMatrix_MPIAIJ_All(mat, MAT_DO_NOT_GET_VALUES, MAT_INITIAL_MATRIX, &dummy));
2542: *newmat = *dummy;
2543: PetscCall(PetscFree(dummy));
2544: PetscFunctionReturn(PETSC_SUCCESS);
2545: }
2547: static PetscErrorCode MatInvertBlockDiagonal_MPIAIJ(Mat A, const PetscScalar **values)
2548: {
2549: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2551: PetscFunctionBegin;
2552: PetscCall(MatInvertBlockDiagonal(a->A, values));
2553: A->factorerrortype = a->A->factorerrortype;
2554: PetscFunctionReturn(PETSC_SUCCESS);
2555: }
2557: static PetscErrorCode MatSetRandom_MPIAIJ(Mat x, PetscRandom rctx)
2558: {
2559: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)x->data;
2560: PetscBool nooffprocentries = x->nooffprocentries;
2562: PetscFunctionBegin;
2563: PetscCheck(x->assembled || x->preallocated, PetscObjectComm((PetscObject)x), PETSC_ERR_ARG_WRONGSTATE, "MatSetRandom() on an unassembled and unpreallocated MATMPIAIJ is not allowed");
2564: PetscCall(MatSetRandom(aij->A, rctx));
2565: if (x->assembled) PetscCall(MatSetRandom(aij->B, rctx));
2566: else PetscCall(MatSetRandomSkipColumnRange_SeqAIJ_Private(aij->B, x->cmap->rstart, x->cmap->rend, rctx));
2567: PetscCall(MatSetOption(x, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
2568: PetscCall(MatAssemblyBegin(x, MAT_FINAL_ASSEMBLY));
2569: PetscCall(MatAssemblyEnd(x, MAT_FINAL_ASSEMBLY));
2570: PetscCall(MatSetOption(x, MAT_NO_OFF_PROC_ENTRIES, nooffprocentries));
2571: PetscFunctionReturn(PETSC_SUCCESS);
2572: }
2574: static PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ(Mat A, PetscBool sc)
2575: {
2576: PetscFunctionBegin;
2577: if (sc) A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ_Scalable;
2578: else A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ;
2579: PetscFunctionReturn(PETSC_SUCCESS);
2580: }
2582: /*@
2583: MatMPIAIJGetNumberNonzeros - gets the number of nonzeros in the matrix on this MPI rank
2585: Not Collective
2587: Input Parameter:
2588: . A - the matrix
2590: Output Parameter:
2591: . nz - the number of nonzeros
2593: Level: advanced
2595: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`
2596: @*/
2597: PetscErrorCode MatMPIAIJGetNumberNonzeros(Mat A, PetscCount *nz)
2598: {
2599: Mat_MPIAIJ *maij = (Mat_MPIAIJ *)A->data;
2600: Mat_SeqAIJ *aaij = (Mat_SeqAIJ *)maij->A->data, *baij = (Mat_SeqAIJ *)maij->B->data;
2601: PetscBool isaij;
2603: PetscFunctionBegin;
2604: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATMPIAIJ, &isaij));
2605: PetscCheck(isaij, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Not for type %s", ((PetscObject)A)->type_name);
2606: *nz = aaij->i[A->rmap->n] + baij->i[A->rmap->n];
2607: PetscFunctionReturn(PETSC_SUCCESS);
2608: }
2610: /*@
2611: MatMPIAIJSetUseScalableIncreaseOverlap - Determine if the matrix uses a scalable algorithm to compute the overlap
2613: Collective
2615: Input Parameters:
2616: + A - the matrix
2617: - sc - `PETSC_TRUE` indicates use the scalable algorithm (default is not to use the scalable algorithm)
2619: Level: advanced
2621: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`
2622: @*/
2623: PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap(Mat A, PetscBool sc)
2624: {
2625: PetscFunctionBegin;
2626: PetscTryMethod(A, "MatMPIAIJSetUseScalableIncreaseOverlap_C", (Mat, PetscBool), (A, sc));
2627: PetscFunctionReturn(PETSC_SUCCESS);
2628: }
2630: PetscErrorCode MatSetFromOptions_MPIAIJ(Mat A, PetscOptionItems PetscOptionsObject)
2631: {
2632: PetscBool sc = PETSC_FALSE, flg;
2634: PetscFunctionBegin;
2635: PetscOptionsHeadBegin(PetscOptionsObject, "MPIAIJ options");
2636: if (A->ops->increaseoverlap == MatIncreaseOverlap_MPIAIJ_Scalable) sc = PETSC_TRUE;
2637: PetscCall(PetscOptionsBool("-mat_increase_overlap_scalable", "Use a scalable algorithm to compute the overlap", "MatIncreaseOverlap", sc, &sc, &flg));
2638: if (flg) PetscCall(MatMPIAIJSetUseScalableIncreaseOverlap(A, sc));
2639: PetscOptionsHeadEnd();
2640: PetscFunctionReturn(PETSC_SUCCESS);
2641: }
2643: static PetscErrorCode MatShift_MPIAIJ(Mat Y, PetscScalar a)
2644: {
2645: Mat_MPIAIJ *maij = (Mat_MPIAIJ *)Y->data;
2646: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)maij->A->data;
2648: PetscFunctionBegin;
2649: if (!Y->preallocated) {
2650: PetscCall(MatMPIAIJSetPreallocation(Y, 1, NULL, 0, NULL));
2651: } else if (!aij->nz && PetscMax(Y->rmap->rstart, Y->cmap->rstart) < PetscMin(Y->rmap->rend, Y->cmap->rend)) {
2652: /* If maij->A has no nonzeros, preallocate it again only if one of the local rows has its diagonal entry in maij->A, otherwise
2653: its nonzero state would increase even though the shift creates no new nonzeros in it */
2654: PetscInt nonew = aij->nonew;
2655: PetscCall(MatSeqAIJSetPreallocation(maij->A, 1, NULL));
2656: aij->nonew = nonew;
2657: }
2658: PetscCall(MatShift_Basic(Y, a));
2659: PetscFunctionReturn(PETSC_SUCCESS);
2660: }
2662: static PetscErrorCode MatInvertVariableBlockDiagonal_MPIAIJ(Mat A, PetscInt nblocks, const PetscInt *bsizes, PetscScalar *diag)
2663: {
2664: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2666: PetscFunctionBegin;
2667: PetscCall(MatInvertVariableBlockDiagonal(a->A, nblocks, bsizes, diag));
2668: PetscFunctionReturn(PETSC_SUCCESS);
2669: }
2671: static PetscErrorCode MatEliminateZeros_MPIAIJ(Mat A, PetscBool keep)
2672: {
2673: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2675: PetscFunctionBegin;
2676: PetscCall(MatEliminateZeros_SeqAIJ(a->A, keep)); // possibly keep zero diagonal coefficients
2677: PetscCall(MatEliminateZeros_SeqAIJ(a->B, PETSC_FALSE)); // never keep zero diagonal coefficients
2678: PetscFunctionReturn(PETSC_SUCCESS);
2679: }
2681: static PetscErrorCode MatGetOrdering_MPIAIJ(Mat A, MatOrderingType type, IS *rperm, IS *cperm)
2682: {
2683: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2684: IS lrowperm, lcolperm;
2685: PetscInt i, rstart, rend, *idx;
2686: const PetscInt *lidx;
2688: PetscFunctionBegin;
2689: PetscCall(MatGetOrdering(a->A, type, &lrowperm, &lcolperm));
2690: PetscCall(MatGetOwnershipRange(A, &rstart, &rend));
2691: /* Remap row index set to global space */
2692: PetscCall(ISGetIndices(lrowperm, &lidx));
2693: PetscCall(PetscMalloc1(rend - rstart, &idx));
2694: for (i = 0; i + rstart < rend; i++) idx[i] = rstart + lidx[i];
2695: PetscCall(ISRestoreIndices(lrowperm, &lidx));
2696: PetscCall(ISDestroy(&lrowperm));
2697: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)A), rend - rstart, idx, PETSC_OWN_POINTER, rperm));
2698: PetscCall(ISSetPermutation(*rperm));
2699: /* Remap column index set to global space */
2700: PetscCall(ISGetIndices(lcolperm, &lidx));
2701: PetscCall(PetscMalloc1(rend - rstart, &idx));
2702: for (i = 0; i + rstart < rend; i++) idx[i] = rstart + lidx[i];
2703: PetscCall(ISRestoreIndices(lcolperm, &lidx));
2704: PetscCall(ISDestroy(&lcolperm));
2705: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)A), rend - rstart, idx, PETSC_OWN_POINTER, cperm));
2706: PetscCall(ISSetPermutation(*cperm));
2707: PetscFunctionReturn(PETSC_SUCCESS);
2708: }
2710: static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ,
2711: MatGetRow_MPIAIJ,
2712: MatRestoreRow_MPIAIJ,
2713: MatMult_MPIAIJ,
2714: /* 4*/ MatMultAdd_MPIAIJ,
2715: MatMultTranspose_MPIAIJ,
2716: MatMultTransposeAdd_MPIAIJ,
2717: NULL,
2718: NULL,
2719: NULL,
2720: /*10*/ NULL,
2721: NULL,
2722: NULL,
2723: MatSOR_MPIAIJ,
2724: MatTranspose_MPIAIJ,
2725: /*15*/ MatGetInfo_MPIAIJ,
2726: MatEqual_MPIAIJ,
2727: MatGetDiagonal_MPIAIJ,
2728: MatDiagonalScale_MPIAIJ,
2729: MatNorm_MPIAIJ,
2730: /*20*/ MatAssemblyBegin_MPIAIJ,
2731: MatAssemblyEnd_MPIAIJ,
2732: MatSetOption_MPIAIJ,
2733: MatZeroEntries_MPIAIJ,
2734: /*24*/ MatZeroRows_MPIAIJ,
2735: NULL,
2736: NULL,
2737: NULL,
2738: NULL,
2739: /*29*/ MatSetUp_MPI_Hash,
2740: NULL,
2741: NULL,
2742: MatGetDiagonalBlock_MPIAIJ,
2743: NULL,
2744: /*34*/ MatDuplicate_MPIAIJ,
2745: NULL,
2746: NULL,
2747: NULL,
2748: NULL,
2749: /*39*/ MatAXPY_MPIAIJ,
2750: MatCreateSubMatrices_MPIAIJ,
2751: MatIncreaseOverlap_MPIAIJ,
2752: MatGetValues_MPIAIJ,
2753: MatCopy_MPIAIJ,
2754: /*44*/ MatGetRowMax_MPIAIJ,
2755: MatScale_MPIAIJ,
2756: MatShift_MPIAIJ,
2757: MatDiagonalSet_MPIAIJ,
2758: MatZeroRowsColumns_MPIAIJ,
2759: /*49*/ MatSetRandom_MPIAIJ,
2760: MatGetRowIJ_MPIAIJ,
2761: MatRestoreRowIJ_MPIAIJ,
2762: NULL,
2763: NULL,
2764: /*54*/ MatFDColoringCreate_MPIXAIJ,
2765: NULL,
2766: MatSetUnfactored_MPIAIJ,
2767: MatPermute_MPIAIJ,
2768: NULL,
2769: /*59*/ MatCreateSubMatrix_MPIAIJ,
2770: MatDestroy_MPIAIJ,
2771: MatView_MPIAIJ,
2772: NULL,
2773: NULL,
2774: /*64*/ MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ,
2775: NULL,
2776: NULL,
2777: NULL,
2778: MatGetRowMaxAbs_MPIAIJ,
2779: /*69*/ MatGetRowMinAbs_MPIAIJ,
2780: NULL,
2781: NULL,
2782: MatFDColoringApply_AIJ,
2783: MatSetFromOptions_MPIAIJ,
2784: MatFindZeroDiagonals_MPIAIJ,
2785: /*75*/ NULL,
2786: NULL,
2787: NULL,
2788: MatLoad_MPIAIJ,
2789: NULL,
2790: /*80*/ NULL,
2791: NULL,
2792: NULL,
2793: /*83*/ NULL,
2794: NULL,
2795: MatMatMultNumeric_MPIAIJ_MPIAIJ,
2796: MatPtAPNumeric_MPIAIJ_MPIAIJ,
2797: NULL,
2798: NULL,
2799: /*89*/ MatBindToCPU_MPIAIJ,
2800: MatProductSetFromOptions_MPIAIJ,
2801: NULL,
2802: NULL,
2803: MatConjugate_MPIAIJ,
2804: /*94*/ NULL,
2805: MatSetValuesRow_MPIAIJ,
2806: MatRealPart_MPIAIJ,
2807: MatImaginaryPart_MPIAIJ,
2808: NULL,
2809: /*99*/ NULL,
2810: NULL,
2811: NULL,
2812: MatGetRowMin_MPIAIJ,
2813: NULL,
2814: /*104*/ MatGetSeqNonzeroStructure_MPIAIJ,
2815: NULL,
2816: MatGetGhosts_MPIAIJ,
2817: NULL,
2818: NULL,
2819: /*109*/ MatMultDiagonalBlock_MPIAIJ,
2820: NULL,
2821: NULL,
2822: NULL,
2823: MatGetMultiProcBlock_MPIAIJ,
2824: /*114*/ MatFindNonzeroRows_MPIAIJ,
2825: MatGetColumnReductions_MPIAIJ,
2826: MatInvertBlockDiagonal_MPIAIJ,
2827: MatInvertVariableBlockDiagonal_MPIAIJ,
2828: MatCreateSubMatricesMPI_MPIAIJ,
2829: /*119*/ NULL,
2830: MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ,
2831: NULL,
2832: NULL,
2833: NULL,
2834: /*124*/ NULL,
2835: MatSetBlockSizes_MPIAIJ,
2836: NULL,
2837: MatFDColoringSetUp_MPIXAIJ,
2838: MatFindOffBlockDiagonalEntries_MPIAIJ,
2839: /*129*/ MatCreateMPIMatConcatenateSeqMat_MPIAIJ,
2840: NULL,
2841: NULL,
2842: NULL,
2843: MatCreateGraph_Simple_AIJ,
2844: /*134*/ NULL,
2845: MatEliminateZeros_MPIAIJ,
2846: MatGetRowSumAbs_MPIAIJ,
2847: NULL,
2848: NULL,
2849: /*139*/ NULL,
2850: MatCopyHashToXAIJ_MPI_Hash,
2851: MatGetCurrentMemType_MPIAIJ,
2852: NULL,
2853: NULL,
2854: /*144*/ NULL,
2855: NULL,
2856: NULL,
2857: MatGetOrdering_MPIAIJ};
2859: static PetscErrorCode MatStoreValues_MPIAIJ(Mat mat)
2860: {
2861: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2863: PetscFunctionBegin;
2864: PetscCall(MatStoreValues(aij->A));
2865: PetscCall(MatStoreValues(aij->B));
2866: PetscFunctionReturn(PETSC_SUCCESS);
2867: }
2869: static PetscErrorCode MatRetrieveValues_MPIAIJ(Mat mat)
2870: {
2871: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2873: PetscFunctionBegin;
2874: PetscCall(MatRetrieveValues(aij->A));
2875: PetscCall(MatRetrieveValues(aij->B));
2876: PetscFunctionReturn(PETSC_SUCCESS);
2877: }
2879: PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat B, PetscInt d_nz, const PetscInt d_nnz[], PetscInt o_nz, const PetscInt o_nnz[])
2880: {
2881: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2882: PetscMPIInt size;
2884: PetscFunctionBegin;
2885: if (B->hash_active) {
2886: B->ops[0] = b->cops;
2887: B->hash_active = PETSC_FALSE;
2888: }
2889: PetscCall(PetscLayoutSetUp(B->rmap));
2890: PetscCall(PetscLayoutSetUp(B->cmap));
2892: #if PetscDefined(USE_CTABLE)
2893: PetscCall(PetscHMapIDestroy(&b->colmap));
2894: #else
2895: PetscCall(PetscFree(b->colmap));
2896: #endif
2897: PetscCall(PetscFree(b->garray));
2898: PetscCall(VecDestroy(&b->lvec));
2899: PetscCall(VecScatterDestroy(&b->Mvctx));
2901: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
2903: MatSeqXAIJGetOptions_Private(b->B);
2904: PetscCall(MatDestroy(&b->B));
2905: PetscCall(MatCreate(PETSC_COMM_SELF, &b->B));
2906: PetscCall(MatSetSizes(b->B, B->rmap->n, size > 1 ? B->cmap->N : 0, B->rmap->n, size > 1 ? B->cmap->N : 0));
2907: PetscCall(MatSetBlockSizesFromMats(b->B, B, B));
2908: PetscCall(MatSetType(b->B, MATSEQAIJ));
2909: MatSeqXAIJRestoreOptions_Private(b->B);
2910: PetscCall(MatSetOption(b->B, MAT_STRUCTURE_ONLY, B->structure_only));
2912: MatSeqXAIJGetOptions_Private(b->A);
2913: PetscCall(MatDestroy(&b->A));
2914: PetscCall(MatCreate(PETSC_COMM_SELF, &b->A));
2915: PetscCall(MatSetSizes(b->A, B->rmap->n, B->cmap->n, B->rmap->n, B->cmap->n));
2916: PetscCall(MatSetBlockSizesFromMats(b->A, B, B));
2917: PetscCall(MatSetType(b->A, MATSEQAIJ));
2918: MatSeqXAIJRestoreOptions_Private(b->A);
2919: PetscCall(MatSetOption(b->A, MAT_STRUCTURE_ONLY, B->structure_only));
2921: PetscCall(MatSeqAIJSetPreallocation(b->A, d_nz, d_nnz));
2922: PetscCall(MatSeqAIJSetPreallocation(b->B, o_nz, o_nnz));
2923: B->preallocated = PETSC_TRUE;
2924: B->was_assembled = PETSC_FALSE;
2925: B->assembled = PETSC_FALSE;
2926: PetscFunctionReturn(PETSC_SUCCESS);
2927: }
2929: static PetscErrorCode MatResetPreallocation_MPIAIJ(Mat B)
2930: {
2931: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2932: PetscBool ondiagreset, offdiagreset, memoryreset;
2934: PetscFunctionBegin;
2936: 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()");
2937: if (B->num_ass == 0) PetscFunctionReturn(PETSC_SUCCESS);
2939: PetscCall(MatResetPreallocation_SeqAIJ_Private(b->A, &ondiagreset));
2940: PetscCall(MatResetPreallocation_SeqAIJ_Private(b->B, &offdiagreset));
2941: memoryreset = (PetscBool)(ondiagreset || offdiagreset);
2942: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &memoryreset, 1, MPI_C_BOOL, MPI_LOR, PetscObjectComm((PetscObject)B)));
2943: if (!memoryreset) PetscFunctionReturn(PETSC_SUCCESS);
2945: PetscCall(PetscLayoutSetUp(B->rmap));
2946: PetscCall(PetscLayoutSetUp(B->cmap));
2947: 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");
2948: PetscCall(MatDisAssemble_MPIAIJ(B, PETSC_TRUE));
2949: PetscCall(VecScatterDestroy(&b->Mvctx));
2951: B->preallocated = PETSC_TRUE;
2952: B->was_assembled = PETSC_FALSE;
2953: B->assembled = PETSC_FALSE;
2954: /* Log that the state of this object has changed; this will help guarantee that preconditioners get re-setup */
2955: PetscCall(PetscObjectStateIncrease((PetscObject)B));
2956: PetscFunctionReturn(PETSC_SUCCESS);
2957: }
2959: PetscErrorCode MatDuplicate_MPIAIJ(Mat matin, MatDuplicateOption cpvalues, Mat *newmat)
2960: {
2961: Mat mat;
2962: Mat_MPIAIJ *a, *oldmat = (Mat_MPIAIJ *)matin->data;
2964: PetscFunctionBegin;
2965: *newmat = NULL;
2966: PetscCall(MatCreate(PetscObjectComm((PetscObject)matin), &mat));
2967: PetscCall(MatSetSizes(mat, matin->rmap->n, matin->cmap->n, matin->rmap->N, matin->cmap->N));
2968: PetscCall(MatSetBlockSizesFromMats(mat, matin, matin));
2969: PetscCall(MatSetType(mat, ((PetscObject)matin)->type_name));
2970: PetscCall(MatSetOption(mat, MAT_STRUCTURE_ONLY, matin->structure_only));
2971: a = (Mat_MPIAIJ *)mat->data;
2973: mat->factortype = matin->factortype;
2974: mat->assembled = matin->assembled;
2975: mat->insertmode = NOT_SET_VALUES;
2977: a->size = oldmat->size;
2978: a->rank = oldmat->rank;
2979: a->donotstash = oldmat->donotstash;
2980: a->roworiented = oldmat->roworiented;
2981: a->rowindices = NULL;
2982: a->rowvalues = NULL;
2983: a->getrowactive = PETSC_FALSE;
2985: PetscCall(PetscLayoutReference(matin->rmap, &mat->rmap));
2986: PetscCall(PetscLayoutReference(matin->cmap, &mat->cmap));
2987: if (matin->hash_active) PetscCall(MatSetUp(mat));
2988: else {
2989: mat->preallocated = matin->preallocated;
2990: if (oldmat->colmap) {
2991: #if PetscDefined(USE_CTABLE)
2992: PetscCall(PetscHMapIDuplicate(oldmat->colmap, &a->colmap));
2993: #else
2994: PetscCall(PetscMalloc1(mat->cmap->N, &a->colmap));
2995: PetscCall(PetscArraycpy(a->colmap, oldmat->colmap, mat->cmap->N));
2996: #endif
2997: } else a->colmap = NULL;
2998: if (oldmat->garray) {
2999: PetscInt len;
3000: len = oldmat->B->cmap->n;
3001: PetscCall(PetscMalloc1(len, &a->garray));
3002: if (len) PetscCall(PetscArraycpy(a->garray, oldmat->garray, len));
3003: } else a->garray = NULL;
3005: /* It may happen MatDuplicate is called with a non-assembled matrix
3006: In fact, MatDuplicate only requires the matrix to be preallocated
3007: This may happen inside a DMCreateMatrix_Shell */
3008: if (oldmat->lvec) PetscCall(VecDuplicate(oldmat->lvec, &a->lvec));
3009: if (oldmat->Mvctx) {
3010: a->Mvctx = oldmat->Mvctx;
3011: PetscCall(PetscObjectReference((PetscObject)oldmat->Mvctx));
3012: }
3013: PetscCall(MatDuplicate(oldmat->A, cpvalues, &a->A));
3014: PetscCall(MatDuplicate(oldmat->B, cpvalues, &a->B));
3015: }
3016: PetscCall(PetscFunctionListDuplicate(((PetscObject)matin)->qlist, &((PetscObject)mat)->qlist));
3017: *newmat = mat;
3018: PetscFunctionReturn(PETSC_SUCCESS);
3019: }
3021: PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer)
3022: {
3023: PetscBool isbinary, ishdf5;
3025: PetscFunctionBegin;
3028: /* force binary viewer to load .info file if it has not yet done so */
3029: PetscCall(PetscViewerSetUp(viewer));
3030: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
3031: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERHDF5, &ishdf5));
3032: if (isbinary) {
3033: PetscCall(MatLoad_MPIAIJ_Binary(newMat, viewer));
3034: } else if (ishdf5) {
3035: #if PetscDefined(HAVE_HDF5)
3036: PetscCall(MatLoad_AIJ_HDF5(newMat, viewer));
3037: #else
3038: SETERRQ(PetscObjectComm((PetscObject)newMat), PETSC_ERR_SUP, "HDF5 not supported in this build.\nPlease reconfigure using --download-hdf5");
3039: #endif
3040: } else {
3041: 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);
3042: }
3043: PetscFunctionReturn(PETSC_SUCCESS);
3044: }
3046: PetscErrorCode MatLoad_MPIAIJ_Binary(Mat mat, PetscViewer viewer)
3047: {
3048: PetscInt header[4], M, N, m, nz, rows, cols, sum, i;
3049: PetscInt *rowidxs, *colidxs;
3050: PetscScalar *matvals;
3052: PetscFunctionBegin;
3053: PetscCall(PetscViewerSetUp(viewer));
3055: /* read in matrix header */
3056: PetscCall(PetscViewerBinaryRead(viewer, header, 4, NULL, PETSC_INT));
3057: PetscCheck(header[0] == MAT_FILE_CLASSID, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Not a matrix object in file");
3058: M = header[1];
3059: N = header[2];
3060: nz = header[3];
3061: PetscCheck(M >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix row size (%" PetscInt_FMT ") in file is negative", M);
3062: PetscCheck(N >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix column size (%" PetscInt_FMT ") in file is negative", N);
3063: PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Matrix stored in special format on disk, cannot load as MPIAIJ");
3065: /* set block sizes from the viewer's .info file */
3066: PetscCall(MatLoad_Binary_BlockSizes(mat, viewer));
3067: /* set global sizes if not set already */
3068: if (mat->rmap->N < 0) mat->rmap->N = M;
3069: if (mat->cmap->N < 0) mat->cmap->N = N;
3070: PetscCall(PetscLayoutSetUp(mat->rmap));
3071: PetscCall(PetscLayoutSetUp(mat->cmap));
3073: /* check if the matrix sizes are correct */
3074: PetscCall(MatGetSize(mat, &rows, &cols));
3075: 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);
3077: /* read in row lengths and build row indices */
3078: PetscCall(MatGetLocalSize(mat, &m, NULL));
3079: PetscCall(PetscMalloc1(m + 1, &rowidxs));
3080: PetscCall(PetscViewerBinaryReadAll(viewer, rowidxs + 1, m, PETSC_DECIDE, M, PETSC_INT));
3081: rowidxs[0] = 0;
3082: for (i = 0; i < m; i++) rowidxs[i + 1] += rowidxs[i];
3083: if (nz != PETSC_INT_MAX) {
3084: PetscCallMPI(MPIU_Allreduce(&rowidxs[m], &sum, 1, MPIU_INT, MPI_SUM, PetscObjectComm((PetscObject)viewer)));
3085: 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);
3086: }
3088: /* read in column indices and matrix values */
3089: PetscCall(PetscMalloc2(rowidxs[m], &colidxs, rowidxs[m], &matvals));
3090: PetscCall(PetscViewerBinaryReadAll(viewer, colidxs, rowidxs[m], PETSC_DETERMINE, PETSC_DETERMINE, PETSC_INT));
3091: PetscCall(PetscViewerBinaryReadAll(viewer, matvals, rowidxs[m], PETSC_DETERMINE, PETSC_DETERMINE, PETSC_SCALAR));
3092: /* store matrix indices and values */
3093: PetscCall(MatMPIAIJSetPreallocationCSR(mat, rowidxs, colidxs, matvals));
3094: PetscCall(PetscFree(rowidxs));
3095: PetscCall(PetscFree2(colidxs, matvals));
3096: PetscFunctionReturn(PETSC_SUCCESS);
3097: }
3099: /* Not scalable because of ISAllGather() unless getting all columns. */
3100: static PetscErrorCode ISGetSeqIS_Private(Mat mat, IS iscol, IS *isseq)
3101: {
3102: IS iscol_local;
3103: PetscBool isstride;
3104: PetscMPIInt gisstride = 0;
3106: PetscFunctionBegin;
3107: /* check if we are grabbing all columns*/
3108: PetscCall(PetscObjectTypeCompare((PetscObject)iscol, ISSTRIDE, &isstride));
3110: if (isstride) {
3111: PetscInt start, len, mstart, mlen;
3112: PetscCall(ISStrideGetInfo(iscol, &start, NULL));
3113: PetscCall(ISGetLocalSize(iscol, &len));
3114: PetscCall(MatGetOwnershipRangeColumn(mat, &mstart, &mlen));
3115: if (mstart == start && mlen - mstart == len) gisstride = 1;
3116: }
3118: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &gisstride, 1, MPI_INT, MPI_MIN, PetscObjectComm((PetscObject)mat)));
3119: if (gisstride) {
3120: PetscInt N;
3121: PetscCall(MatGetSize(mat, NULL, &N));
3122: PetscCall(ISCreateStride(PETSC_COMM_SELF, N, 0, 1, &iscol_local));
3123: PetscCall(ISSetIdentity(iscol_local));
3124: PetscCall(PetscInfo(mat, "Optimizing for obtaining all columns of the matrix; skipping ISAllGather()\n"));
3125: } else {
3126: PetscInt cbs;
3127: PetscCall(ISGetBlockSize(iscol, &cbs));
3128: PetscCall(ISAllGather(iscol, &iscol_local));
3129: PetscCall(ISSetBlockSize(iscol_local, cbs));
3130: }
3132: *isseq = iscol_local;
3133: PetscFunctionReturn(PETSC_SUCCESS);
3134: }
3136: /*
3137: Used by MatCreateSubMatrix_MPIAIJ_SameRowColDist() to avoid ISAllGather() and global size of iscol_local
3138: (see MatCreateSubMatrix_MPIAIJ_nonscalable)
3140: Input Parameters:
3141: + mat - matrix
3142: . isrow - parallel row index set; its local indices are a subset of local columns of `mat`,
3143: i.e., mat->rstart <= isrow[i] < mat->rend
3144: - iscol - parallel column index set; its local indices are a subset of local columns of `mat`,
3145: i.e., mat->cstart <= iscol[i] < mat->cend
3147: Output Parameters:
3148: + isrow_d - sequential row index set for retrieving mat->A
3149: . iscol_d - sequential column index set for retrieving mat->A
3150: . iscol_o - sequential column index set for retrieving mat->B
3151: - garray - column map; garray[i] indicates global location of iscol_o[i] in `iscol`
3152: */
3153: static PetscErrorCode ISGetSeqIS_SameColDist_Private(Mat mat, IS isrow, IS iscol, IS *isrow_d, IS *iscol_d, IS *iscol_o, PetscInt *garray[])
3154: {
3155: Vec x, cmap;
3156: const PetscInt *is_idx;
3157: PetscScalar *xarray, *cmaparray;
3158: PetscInt ncols, isstart, *idx, m, rstart, *cmap1, count;
3159: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data;
3160: Mat B = a->B;
3161: Vec lvec = a->lvec, lcmap;
3162: PetscInt i, cstart, cend, Bn = B->cmap->N;
3163: MPI_Comm comm;
3164: VecScatter Mvctx = a->Mvctx;
3166: PetscFunctionBegin;
3167: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3168: PetscCall(ISGetLocalSize(iscol, &ncols));
3170: /* (1) iscol is a sub-column vector of mat, pad it with '-1.' to form a full vector x */
3171: PetscCall(MatCreateVecs(mat, &x, NULL));
3172: PetscCall(VecSet(x, -1.0));
3173: PetscCall(VecDuplicate(x, &cmap));
3174: PetscCall(VecSet(cmap, -1.0));
3176: /* Get start indices */
3177: PetscCallMPI(MPI_Scan(&ncols, &isstart, 1, MPIU_INT, MPI_SUM, comm));
3178: isstart -= ncols;
3179: PetscCall(MatGetOwnershipRangeColumn(mat, &cstart, &cend));
3181: PetscCall(ISGetIndices(iscol, &is_idx));
3182: PetscCall(VecGetArray(x, &xarray));
3183: PetscCall(VecGetArray(cmap, &cmaparray));
3184: PetscCall(PetscMalloc1(ncols, &idx));
3185: for (i = 0; i < ncols; i++) {
3186: xarray[is_idx[i] - cstart] = (PetscScalar)is_idx[i];
3187: cmaparray[is_idx[i] - cstart] = i + isstart; /* global index of iscol[i] */
3188: idx[i] = is_idx[i] - cstart; /* local index of iscol[i] */
3189: }
3190: PetscCall(VecRestoreArray(x, &xarray));
3191: PetscCall(VecRestoreArray(cmap, &cmaparray));
3192: PetscCall(ISRestoreIndices(iscol, &is_idx));
3194: /* Get iscol_d */
3195: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, iscol_d));
3196: PetscCall(ISGetBlockSize(iscol, &i));
3197: PetscCall(ISSetBlockSize(*iscol_d, i));
3199: /* Get isrow_d */
3200: PetscCall(ISGetLocalSize(isrow, &m));
3201: rstart = mat->rmap->rstart;
3202: PetscCall(PetscMalloc1(m, &idx));
3203: PetscCall(ISGetIndices(isrow, &is_idx));
3204: for (i = 0; i < m; i++) idx[i] = is_idx[i] - rstart;
3205: PetscCall(ISRestoreIndices(isrow, &is_idx));
3207: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, m, idx, PETSC_OWN_POINTER, isrow_d));
3208: PetscCall(ISGetBlockSize(isrow, &i));
3209: PetscCall(ISSetBlockSize(*isrow_d, i));
3211: /* (2) Scatter x and cmap using aij->Mvctx to get their off-process portions (see MatMult_MPIAIJ) */
3212: PetscCall(VecScatterBegin(Mvctx, x, lvec, INSERT_VALUES, SCATTER_FORWARD));
3213: PetscCall(VecScatterEnd(Mvctx, x, lvec, INSERT_VALUES, SCATTER_FORWARD));
3215: PetscCall(VecDuplicate(lvec, &lcmap));
3217: PetscCall(VecScatterBegin(Mvctx, cmap, lcmap, INSERT_VALUES, SCATTER_FORWARD));
3218: PetscCall(VecScatterEnd(Mvctx, cmap, lcmap, INSERT_VALUES, SCATTER_FORWARD));
3220: /* (3) create sequential iscol_o (a subset of iscol) and isgarray */
3221: /* off-process column indices */
3222: count = 0;
3223: PetscCall(PetscMalloc1(Bn, &idx));
3224: PetscCall(PetscMalloc1(Bn, &cmap1));
3226: PetscCall(VecGetArray(lvec, &xarray));
3227: PetscCall(VecGetArray(lcmap, &cmaparray));
3228: for (i = 0; i < Bn; i++) {
3229: if (PetscRealPart(xarray[i]) > -1.0) {
3230: idx[count] = i; /* local column index in off-diagonal part B */
3231: cmap1[count] = (PetscInt)PetscRealPart(cmaparray[i]); /* column index in submat */
3232: count++;
3233: }
3234: }
3235: PetscCall(VecRestoreArray(lvec, &xarray));
3236: PetscCall(VecRestoreArray(lcmap, &cmaparray));
3238: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, count, idx, PETSC_COPY_VALUES, iscol_o));
3239: /* cannot ensure iscol_o has same blocksize as iscol! */
3241: PetscCall(PetscFree(idx));
3242: *garray = cmap1;
3244: PetscCall(VecDestroy(&x));
3245: PetscCall(VecDestroy(&cmap));
3246: PetscCall(VecDestroy(&lcmap));
3247: PetscFunctionReturn(PETSC_SUCCESS);
3248: }
3250: /* isrow and iscol have same processor distribution as mat, output *submat is a submatrix of local mat */
3251: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowColDist(Mat mat, IS isrow, IS iscol, MatReuse call, Mat *submat)
3252: {
3253: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data, *asub;
3254: Mat M = NULL;
3255: MPI_Comm comm;
3256: IS iscol_d, isrow_d, iscol_o;
3257: Mat Asub = NULL, Bsub = NULL;
3258: PetscInt n, count, M_size, N_size;
3260: PetscFunctionBegin;
3261: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3263: if (call == MAT_REUSE_MATRIX) {
3264: /* Retrieve isrow_d, iscol_d and iscol_o from submat */
3265: PetscCall(PetscObjectQuery((PetscObject)*submat, "isrow_d", (PetscObject *)&isrow_d));
3266: PetscCheck(isrow_d, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "isrow_d passed in was not used before, cannot reuse");
3268: PetscCall(PetscObjectQuery((PetscObject)*submat, "iscol_d", (PetscObject *)&iscol_d));
3269: PetscCheck(iscol_d, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "iscol_d passed in was not used before, cannot reuse");
3271: PetscCall(PetscObjectQuery((PetscObject)*submat, "iscol_o", (PetscObject *)&iscol_o));
3272: PetscCheck(iscol_o, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "iscol_o passed in was not used before, cannot reuse");
3274: /* Update diagonal and off-diagonal portions of submat */
3275: asub = (Mat_MPIAIJ *)(*submat)->data;
3276: PetscCall(MatCreateSubMatrix_SeqAIJ(a->A, isrow_d, iscol_d, PETSC_DECIDE, MAT_REUSE_MATRIX, &asub->A));
3277: PetscCall(ISGetLocalSize(iscol_o, &n));
3278: if (n) PetscCall(MatCreateSubMatrix_SeqAIJ(a->B, isrow_d, iscol_o, PETSC_DECIDE, MAT_REUSE_MATRIX, &asub->B));
3279: PetscCall(MatAssemblyBegin(*submat, MAT_FINAL_ASSEMBLY));
3280: PetscCall(MatAssemblyEnd(*submat, MAT_FINAL_ASSEMBLY));
3282: } else { /* call == MAT_INITIAL_MATRIX) */
3283: PetscInt *garray, *garray_compact;
3284: PetscInt BsubN;
3286: /* Create isrow_d, iscol_d, iscol_o and isgarray (replace isgarray with array?) */
3287: PetscCall(ISGetSeqIS_SameColDist_Private(mat, isrow, iscol, &isrow_d, &iscol_d, &iscol_o, &garray));
3289: /* Create local submatrices Asub and Bsub */
3290: PetscCall(MatCreateSubMatrix_SeqAIJ(a->A, isrow_d, iscol_d, PETSC_DECIDE, MAT_INITIAL_MATRIX, &Asub));
3291: PetscCall(MatCreateSubMatrix_SeqAIJ(a->B, isrow_d, iscol_o, PETSC_DECIDE, MAT_INITIAL_MATRIX, &Bsub));
3293: // Compact garray so its not of size Bn
3294: PetscCall(ISGetSize(iscol_o, &count));
3295: PetscCall(PetscMalloc1(count, &garray_compact));
3296: PetscCall(PetscArraycpy(garray_compact, garray, count));
3298: /* Create submatrix M */
3299: PetscCall(ISGetSize(isrow, &M_size));
3300: PetscCall(ISGetSize(iscol, &N_size));
3301: PetscCall(MatCreateMPIAIJWithSeqAIJ(comm, M_size, N_size, Asub, Bsub, garray_compact, &M));
3303: /* If Bsub has empty columns, compress iscol_o such that it will retrieve condensed Bsub from a->B during reuse */
3304: asub = (Mat_MPIAIJ *)M->data;
3306: PetscCall(ISGetLocalSize(iscol_o, &BsubN));
3307: n = asub->B->cmap->N;
3308: if (BsubN > n) {
3309: /* This case can be tested using ~petsc/src/tao/bound/tutorials/runplate2_3 */
3310: const PetscInt *idx;
3311: PetscInt i, j, *idx_new, *subgarray = asub->garray;
3312: PetscCall(PetscInfo(M, "submatrix Bn %" PetscInt_FMT " != BsubN %" PetscInt_FMT ", update iscol_o\n", n, BsubN));
3314: PetscCall(PetscMalloc1(n, &idx_new));
3315: j = 0;
3316: PetscCall(ISGetIndices(iscol_o, &idx));
3317: for (i = 0; i < n; i++) {
3318: if (j >= BsubN) break;
3319: while (subgarray[i] > garray[j]) j++;
3321: PetscCheck(subgarray[i] == garray[j], PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "subgarray[%" PetscInt_FMT "]=%" PetscInt_FMT " cannot < garray[%" PetscInt_FMT "]=%" PetscInt_FMT, i, subgarray[i], j, garray[j]);
3322: idx_new[i] = idx[j++];
3323: }
3324: PetscCall(ISRestoreIndices(iscol_o, &idx));
3326: PetscCall(ISDestroy(&iscol_o));
3327: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, n, idx_new, PETSC_OWN_POINTER, &iscol_o));
3329: } else PetscCheck(BsubN >= n, 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);
3331: PetscCall(PetscFree(garray));
3332: *submat = M;
3334: /* Save isrow_d, iscol_d and iscol_o used in processor for next request */
3335: PetscCall(PetscObjectCompose((PetscObject)M, "isrow_d", (PetscObject)isrow_d));
3336: PetscCall(ISDestroy(&isrow_d));
3338: PetscCall(PetscObjectCompose((PetscObject)M, "iscol_d", (PetscObject)iscol_d));
3339: PetscCall(ISDestroy(&iscol_d));
3341: PetscCall(PetscObjectCompose((PetscObject)M, "iscol_o", (PetscObject)iscol_o));
3342: PetscCall(ISDestroy(&iscol_o));
3343: }
3344: PetscFunctionReturn(PETSC_SUCCESS);
3345: }
3347: PetscErrorCode MatCreateSubMatrix_MPIAIJ(Mat mat, IS isrow, IS iscol, MatReuse call, Mat *newmat)
3348: {
3349: IS iscol_local = NULL, isrow_d;
3350: PetscInt csize;
3351: PetscInt n, i, j, start, end;
3352: PetscBool sameRowDist = PETSC_FALSE, tsameDist[2];
3353: MPI_Comm comm;
3355: PetscFunctionBegin;
3356: /* If isrow has same processor distribution as mat,
3357: call MatCreateSubMatrix_MPIAIJ_SameRowDist() to avoid using a hash table with global size of iscol */
3358: if (call == MAT_REUSE_MATRIX) {
3359: PetscCall(PetscObjectQuery((PetscObject)*newmat, "isrow_d", (PetscObject *)&isrow_d));
3360: if (isrow_d) {
3361: sameRowDist = PETSC_TRUE;
3362: tsameDist[1] = PETSC_TRUE; /* sameColDist */
3363: } else {
3364: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_local));
3365: if (iscol_local) {
3366: sameRowDist = PETSC_TRUE;
3367: tsameDist[1] = PETSC_FALSE; /* !sameColDist */
3368: }
3369: }
3370: } else {
3371: /* Check if isrow has same processor distribution as mat */
3372: tsameDist[0] = PETSC_FALSE;
3373: PetscCall(ISGetLocalSize(isrow, &n));
3374: if (!n) {
3375: tsameDist[0] = PETSC_TRUE;
3376: } else {
3377: PetscCall(ISGetMinMax(isrow, &i, &j));
3378: PetscCall(MatGetOwnershipRange(mat, &start, &end));
3379: if (i >= start && j < end) tsameDist[0] = PETSC_TRUE;
3380: }
3382: /* Check if iscol has same processor distribution as mat */
3383: tsameDist[1] = PETSC_FALSE;
3384: PetscCall(ISGetLocalSize(iscol, &n));
3385: if (!n) {
3386: tsameDist[1] = PETSC_TRUE;
3387: } else {
3388: PetscCall(ISGetMinMax(iscol, &i, &j));
3389: PetscCall(MatGetOwnershipRangeColumn(mat, &start, &end));
3390: if (i >= start && j < end) tsameDist[1] = PETSC_TRUE;
3391: }
3393: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3394: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, tsameDist, 2, MPI_C_BOOL, MPI_LAND, comm));
3395: sameRowDist = tsameDist[0];
3396: }
3398: if (sameRowDist) {
3399: if (tsameDist[1]) { /* sameRowDist & sameColDist */
3400: /* isrow and iscol have same processor distribution as mat */
3401: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowColDist(mat, isrow, iscol, call, newmat));
3402: PetscFunctionReturn(PETSC_SUCCESS);
3403: } else { /* sameRowDist */
3404: /* isrow has same processor distribution as mat */
3405: if (call == MAT_INITIAL_MATRIX) {
3406: PetscBool sorted;
3407: PetscCall(ISGetSeqIS_Private(mat, iscol, &iscol_local));
3408: PetscCall(ISGetLocalSize(iscol_local, &n)); /* local size of iscol_local = global columns of newmat */
3409: PetscCall(ISGetSize(iscol, &i));
3410: PetscCheck(n == i, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "n %" PetscInt_FMT " != size of iscol %" PetscInt_FMT, n, i);
3412: PetscCall(ISSorted(iscol_local, &sorted));
3413: if (sorted) {
3414: /* MatCreateSubMatrix_MPIAIJ_SameRowDist() requires iscol_local be sorted; it can have duplicate indices */
3415: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowDist(mat, isrow, iscol, iscol_local, MAT_INITIAL_MATRIX, newmat));
3416: PetscFunctionReturn(PETSC_SUCCESS);
3417: }
3418: } else { /* call == MAT_REUSE_MATRIX */
3419: IS iscol_sub;
3420: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_sub));
3421: if (iscol_sub) {
3422: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowDist(mat, isrow, iscol, NULL, call, newmat));
3423: PetscFunctionReturn(PETSC_SUCCESS);
3424: }
3425: }
3426: }
3427: }
3429: /* General case: iscol -> iscol_local which has global size of iscol */
3430: if (call == MAT_REUSE_MATRIX) {
3431: PetscCall(PetscObjectQuery((PetscObject)*newmat, "ISAllGather", (PetscObject *)&iscol_local));
3432: PetscCheck(iscol_local, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3433: } else {
3434: if (!iscol_local) PetscCall(ISGetSeqIS_Private(mat, iscol, &iscol_local));
3435: }
3437: PetscCall(ISGetLocalSize(iscol, &csize));
3438: PetscCall(MatCreateSubMatrix_MPIAIJ_nonscalable(mat, isrow, iscol_local, csize, call, newmat));
3440: if (call == MAT_INITIAL_MATRIX) {
3441: PetscCall(PetscObjectCompose((PetscObject)*newmat, "ISAllGather", (PetscObject)iscol_local));
3442: PetscCall(ISDestroy(&iscol_local));
3443: }
3444: PetscFunctionReturn(PETSC_SUCCESS);
3445: }
3447: /*@
3448: MatCreateMPIAIJWithSeqAIJ - creates a `MATMPIAIJ` matrix using `MATSEQAIJ` matrices that contain the "diagonal"
3449: and "off-diagonal" part of the matrix in CSR format.
3451: Collective
3453: Input Parameters:
3454: + comm - MPI communicator
3455: . M - the global row size
3456: . N - the global column size
3457: . A - "diagonal" portion of matrix
3458: . B - if garray is `NULL`, B should be the offdiag matrix using global col ids and of size N - if garray is not `NULL`, B should be the offdiag matrix using local col ids and of size garray
3459: - garray - either `NULL` or the global index of `B` columns. If not `NULL`, it should be allocated by `PetscMalloc1()` and will be owned by `mat` thereafter.
3461: Output Parameter:
3462: . mat - the matrix, with input `A` as its local diagonal matrix
3464: Level: advanced
3466: Notes:
3467: See `MatCreateAIJ()` for the definition of "diagonal" and "off-diagonal" portion of the matrix.
3469: `A` and `B` becomes part of output mat. The user cannot use `A` and `B` anymore.
3471: If `garray` is `NULL`, `B` will be compacted to use local indices. In this sense, `B`'s sparsity pattern (nonzerostate) will be changed. If `B` is a device matrix, we need to somehow also update
3472: `B`'s copy on device. We do so by increasing `B`'s nonzerostate. In use of `B` on device, device matrix types should detect this change (ref. internal routines `MatSeqAIJCUSPARSECopyToGPU()` or
3473: `MatAssemblyEnd_SeqAIJKokkos()`) and will just destroy and then recreate the device copy of `B`. It is not optimal, but is easy to implement and less hacky. To avoid this overhead, try to compute `garray`
3474: yourself, see algorithms in the private function `MatSetUpMultiply_MPIAIJ()`.
3476: The `NULL`-ness of `garray` doesn't need to be collective, in other words, `garray` can be `NULL` on some processes while not on others.
3478: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MATSEQAIJ`, `MatCreateMPIAIJWithSplitArrays()`
3479: @*/
3480: PetscErrorCode MatCreateMPIAIJWithSeqAIJ(MPI_Comm comm, PetscInt M, PetscInt N, Mat A, Mat B, PetscInt *garray, Mat *mat)
3481: {
3482: PetscInt m, n;
3483: MatType mpi_mat_type;
3484: Mat_MPIAIJ *mpiaij;
3485: Mat C;
3487: PetscFunctionBegin;
3488: PetscCall(MatCreate(comm, &C));
3489: PetscCall(MatGetSize(A, &m, &n));
3490: PetscCheck(m == B->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Am %" PetscInt_FMT " != Bm %" PetscInt_FMT, m, B->rmap->N);
3491: PetscCheck(A->rmap->bs == B->rmap->bs, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "A row bs %" PetscInt_FMT " != B row bs %" PetscInt_FMT, A->rmap->bs, B->rmap->bs);
3493: PetscCheck(A->structure_only == B->structure_only, PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Diagonal and off-diagonal matrices must have the same MAT_STRUCTURE_ONLY option value");
3494: PetscCall(MatSetSizes(C, m, n, M, N));
3495: /* Determine the type of MPI matrix that should be created from the type of matrix A, which holds the "diagonal" portion. */
3496: PetscCall(MatGetMPIMatType_Private(A, &mpi_mat_type));
3497: PetscCall(MatSetType(C, mpi_mat_type));
3498: if (!garray) {
3499: const PetscScalar *ba;
3501: B->nonzerostate++;
3502: PetscCall(MatSeqAIJGetArrayRead(B, &ba)); /* Since we will destroy B's device copy, we need to make sure the host copy is up to date */
3503: PetscCall(MatSeqAIJRestoreArrayRead(B, &ba));
3504: }
3506: PetscCall(MatSetBlockSizes(C, A->rmap->bs, A->cmap->bs));
3507: PetscCall(PetscLayoutSetUp(C->rmap));
3508: PetscCall(PetscLayoutSetUp(C->cmap));
3510: mpiaij = (Mat_MPIAIJ *)C->data;
3511: mpiaij->A = A;
3512: mpiaij->B = B;
3513: mpiaij->garray = garray;
3514: C->preallocated = PETSC_TRUE;
3515: C->nooffprocentries = PETSC_TRUE; /* See MatAssemblyBegin_MPIAIJ. In effect, making MatAssemblyBegin a nop */
3517: PetscCall(MatSetOption(C, MAT_STRUCTURE_ONLY, A->structure_only));
3518: PetscCall(MatSetOption(C, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
3519: PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY));
3520: /* MatAssemblyEnd is critical here. It sets mat->offloadmask according to A and B's, and
3521: also gets mpiaij->B compacted (if garray is NULL), with its col ids and size reduced
3522: */
3523: PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY));
3524: PetscCall(MatSetOption(C, MAT_NO_OFF_PROC_ENTRIES, PETSC_FALSE));
3525: PetscCall(MatSetOption(C, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
3526: *mat = C;
3527: PetscFunctionReturn(PETSC_SUCCESS);
3528: }
3530: extern PetscErrorCode MatCreateSubMatrices_MPIAIJ_SingleIS_Local(Mat, PetscInt, const IS[], const IS[], MatReuse, PetscBool, Mat *);
3532: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowDist(Mat mat, IS isrow, IS iscol, IS iscol_local, MatReuse call, Mat *newmat)
3533: {
3534: PetscInt i, m, n, rstart, row, rend, nz, j, bs, cbs;
3535: PetscInt *ii, *jj, nlocal, *dlens, *olens, dlen, olen, jend, mglobal;
3536: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data;
3537: Mat M, Msub, B = a->B;
3538: MatScalar *aa;
3539: Mat_SeqAIJ *aij;
3540: PetscInt *garray = a->garray, *colsub, Ncols;
3541: PetscInt count, Bn = B->cmap->N, cstart = mat->cmap->rstart, cend = mat->cmap->rend;
3542: IS iscol_sub, iscmap;
3543: const PetscInt *is_idx, *cmap;
3544: PetscBool allcolumns = PETSC_FALSE;
3545: MPI_Comm comm;
3547: PetscFunctionBegin;
3548: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3549: if (call == MAT_REUSE_MATRIX) {
3550: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_sub));
3551: PetscCheck(iscol_sub, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "SubIScol passed in was not used before, cannot reuse");
3552: PetscCall(ISGetLocalSize(iscol_sub, &count));
3554: PetscCall(PetscObjectQuery((PetscObject)*newmat, "Subcmap", (PetscObject *)&iscmap));
3555: PetscCheck(iscmap, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Subcmap passed in was not used before, cannot reuse");
3557: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubMatrix", (PetscObject *)&Msub));
3558: PetscCheck(Msub, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3560: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol_sub, MAT_REUSE_MATRIX, PETSC_FALSE, &Msub));
3562: } else { /* call == MAT_INITIAL_MATRIX) */
3563: PetscBool flg;
3565: PetscCall(ISGetLocalSize(iscol, &n));
3566: PetscCall(ISGetSize(iscol, &Ncols));
3568: /* (1) iscol -> nonscalable iscol_local */
3569: /* Check for special case: each processor gets entire matrix columns */
3570: PetscCall(ISIdentity(iscol_local, &flg));
3571: if (flg && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3572: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &allcolumns, 1, MPI_C_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
3573: if (allcolumns) {
3574: iscol_sub = iscol_local;
3575: PetscCall(PetscObjectReference((PetscObject)iscol_local));
3576: PetscCall(ISCreateStride(PETSC_COMM_SELF, n, 0, 1, &iscmap));
3578: } else {
3579: /* (2) iscol_local -> iscol_sub and iscmap. Implementation below requires iscol_local be sorted, it can have duplicate indices */
3580: PetscInt *idx, *cmap1, k;
3581: PetscCall(PetscMalloc1(Ncols, &idx));
3582: PetscCall(PetscMalloc1(Ncols, &cmap1));
3583: PetscCall(ISGetIndices(iscol_local, &is_idx));
3584: count = 0;
3585: k = 0;
3586: for (i = 0; i < Ncols; i++) {
3587: j = is_idx[i];
3588: if (j >= cstart && j < cend) {
3589: /* diagonal part of mat */
3590: idx[count] = j;
3591: cmap1[count++] = i; /* column index in submat */
3592: } else if (Bn) {
3593: /* off-diagonal part of mat */
3594: if (j == garray[k]) {
3595: idx[count] = j;
3596: cmap1[count++] = i; /* column index in submat */
3597: } else if (j > garray[k]) {
3598: while (j > garray[k] && k < Bn - 1) k++;
3599: if (j == garray[k]) {
3600: idx[count] = j;
3601: cmap1[count++] = i; /* column index in submat */
3602: }
3603: }
3604: }
3605: }
3606: PetscCall(ISRestoreIndices(iscol_local, &is_idx));
3608: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, count, idx, PETSC_OWN_POINTER, &iscol_sub));
3609: PetscCall(ISGetBlockSize(iscol, &cbs));
3610: PetscCall(ISSetBlockSize(iscol_sub, cbs));
3612: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)iscol_local), count, cmap1, PETSC_OWN_POINTER, &iscmap));
3613: }
3615: /* (3) Create sequential Msub */
3616: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol_sub, MAT_INITIAL_MATRIX, allcolumns, &Msub));
3617: }
3619: PetscCall(ISGetLocalSize(iscol_sub, &count));
3620: aij = (Mat_SeqAIJ *)Msub->data;
3621: ii = aij->i;
3622: PetscCall(ISGetIndices(iscmap, &cmap));
3624: /*
3625: m - number of local rows
3626: Ncols - number of columns (same on all processors)
3627: rstart - first row in new global matrix generated
3628: */
3629: PetscCall(MatGetSize(Msub, &m, NULL));
3631: if (call == MAT_INITIAL_MATRIX) {
3632: /* (4) Create parallel newmat */
3633: PetscMPIInt rank, size;
3634: PetscInt csize;
3636: PetscCallMPI(MPI_Comm_size(comm, &size));
3637: PetscCallMPI(MPI_Comm_rank(comm, &rank));
3639: /*
3640: Determine the number of non-zeros in the diagonal and off-diagonal
3641: portions of the matrix in order to do correct preallocation
3642: */
3644: /* first get start and end of "diagonal" columns */
3645: PetscCall(ISGetLocalSize(iscol, &csize));
3646: if (csize == PETSC_DECIDE) {
3647: PetscCall(ISGetSize(isrow, &mglobal));
3648: if (mglobal == Ncols) { /* square matrix */
3649: nlocal = m;
3650: } else {
3651: nlocal = Ncols / size + ((Ncols % size) > rank);
3652: }
3653: } else {
3654: nlocal = csize;
3655: }
3656: PetscCallMPI(MPI_Scan(&nlocal, &rend, 1, MPIU_INT, MPI_SUM, comm));
3657: rstart = rend - nlocal;
3658: 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);
3660: /* next, compute all the lengths */
3661: jj = aij->j;
3662: PetscCall(PetscMalloc1(2 * m + 1, &dlens));
3663: olens = dlens + m;
3664: for (i = 0; i < m; i++) {
3665: jend = ii[i + 1] - ii[i];
3666: olen = 0;
3667: dlen = 0;
3668: for (j = 0; j < jend; j++) {
3669: if (cmap[*jj] < rstart || cmap[*jj] >= rend) olen++;
3670: else dlen++;
3671: jj++;
3672: }
3673: olens[i] = olen;
3674: dlens[i] = dlen;
3675: }
3677: PetscCall(ISGetBlockSize(isrow, &bs));
3678: PetscCall(ISGetBlockSize(iscol, &cbs));
3680: PetscCall(MatCreate(comm, &M));
3681: PetscCall(MatSetSizes(M, m, nlocal, PETSC_DECIDE, Ncols));
3682: PetscCall(MatSetBlockSizes(M, bs, cbs));
3683: PetscCall(MatSetType(M, ((PetscObject)mat)->type_name));
3684: PetscCall(MatSetOption(M, MAT_STRUCTURE_ONLY, mat->structure_only));
3685: PetscCall(MatMPIAIJSetPreallocation(M, 0, dlens, 0, olens));
3686: PetscCall(PetscFree(dlens));
3688: } else { /* call == MAT_REUSE_MATRIX */
3689: M = *newmat;
3690: PetscCall(MatGetLocalSize(M, &i, NULL));
3691: PetscCheck(i == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Previous matrix must be same size/layout as request");
3692: PetscCall(MatZeroEntries(M));
3693: /*
3694: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3695: rather than the slower MatSetValues().
3696: */
3697: M->was_assembled = PETSC_TRUE;
3698: M->assembled = PETSC_FALSE;
3699: }
3701: /* (5) Set values of Msub to *newmat */
3702: PetscCall(PetscMalloc1(count, &colsub));
3703: PetscCall(MatGetOwnershipRange(M, &rstart, NULL));
3705: jj = aij->j;
3706: PetscCall(MatSeqAIJGetArrayRead(Msub, (const PetscScalar **)&aa));
3707: for (i = 0; i < m; i++) {
3708: row = rstart + i;
3709: nz = ii[i + 1] - ii[i];
3710: for (j = 0; j < nz; j++) colsub[j] = cmap[jj[j]];
3711: PetscCall(MatSetValues_MPIAIJ(M, 1, &row, nz, colsub, aa, INSERT_VALUES));
3712: jj += nz;
3713: if (!mat->structure_only) aa += nz;
3714: }
3715: PetscCall(MatSeqAIJRestoreArrayRead(Msub, (const PetscScalar **)&aa));
3716: PetscCall(ISRestoreIndices(iscmap, &cmap));
3718: PetscCall(MatAssemblyBegin(M, MAT_FINAL_ASSEMBLY));
3719: PetscCall(MatAssemblyEnd(M, MAT_FINAL_ASSEMBLY));
3721: PetscCall(PetscFree(colsub));
3723: /* save Msub, iscol_sub and iscmap used in processor for next request */
3724: if (call == MAT_INITIAL_MATRIX) {
3725: *newmat = M;
3726: PetscCall(PetscObjectCompose((PetscObject)*newmat, "SubMatrix", (PetscObject)Msub));
3727: PetscCall(MatDestroy(&Msub));
3729: PetscCall(PetscObjectCompose((PetscObject)*newmat, "SubIScol", (PetscObject)iscol_sub));
3730: PetscCall(ISDestroy(&iscol_sub));
3732: PetscCall(PetscObjectCompose((PetscObject)*newmat, "Subcmap", (PetscObject)iscmap));
3733: PetscCall(ISDestroy(&iscmap));
3735: if (iscol_local) {
3736: PetscCall(PetscObjectCompose((PetscObject)*newmat, "ISAllGather", (PetscObject)iscol_local));
3737: PetscCall(ISDestroy(&iscol_local));
3738: }
3739: }
3740: PetscFunctionReturn(PETSC_SUCCESS);
3741: }
3743: /*
3744: Not great since it makes two copies of the submatrix, first an SeqAIJ
3745: in local and then by concatenating the local matrices the end result.
3746: Writing it directly would be much like MatCreateSubMatrices_MPIAIJ()
3748: This requires a sequential iscol with all indices.
3749: */
3750: PetscErrorCode MatCreateSubMatrix_MPIAIJ_nonscalable(Mat mat, IS isrow, IS iscol, PetscInt csize, MatReuse call, Mat *newmat)
3751: {
3752: PetscMPIInt rank, size;
3753: PetscInt i, m, n, rstart, row, rend, nz, *cwork, j, bs, cbs;
3754: PetscInt *ii, *jj, nlocal, *dlens, *olens, dlen, olen, jend, mglobal;
3755: Mat M, Mreuse;
3756: MatScalar *aa, *vwork;
3757: MPI_Comm comm;
3758: Mat_SeqAIJ *aij;
3759: PetscBool colflag, allcolumns = PETSC_FALSE;
3761: PetscFunctionBegin;
3762: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3763: PetscCallMPI(MPI_Comm_rank(comm, &rank));
3764: PetscCallMPI(MPI_Comm_size(comm, &size));
3766: /* Check for special case: each processor gets entire matrix columns */
3767: PetscCall(ISIdentity(iscol, &colflag));
3768: PetscCall(ISGetLocalSize(iscol, &n));
3769: if (colflag && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3770: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &allcolumns, 1, MPI_C_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
3772: if (call == MAT_REUSE_MATRIX) {
3773: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubMatrix", (PetscObject *)&Mreuse));
3774: PetscCheck(Mreuse, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3775: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol, MAT_REUSE_MATRIX, allcolumns, &Mreuse));
3776: } else {
3777: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol, MAT_INITIAL_MATRIX, allcolumns, &Mreuse));
3778: }
3780: /*
3781: m - number of local rows
3782: n - number of columns (same on all processors)
3783: rstart - first row in new global matrix generated
3784: */
3785: PetscCall(MatGetSize(Mreuse, &m, &n));
3786: PetscCall(MatGetBlockSizes(Mreuse, &bs, &cbs));
3787: if (call == MAT_INITIAL_MATRIX) {
3788: aij = (Mat_SeqAIJ *)Mreuse->data;
3789: ii = aij->i;
3790: jj = aij->j;
3792: /*
3793: Determine the number of non-zeros in the diagonal and off-diagonal
3794: portions of the matrix in order to do correct preallocation
3795: */
3797: /* first get start and end of "diagonal" columns */
3798: if (csize == PETSC_DECIDE) {
3799: PetscCall(ISGetSize(isrow, &mglobal));
3800: if (mglobal == n) { /* square matrix */
3801: nlocal = m;
3802: } else {
3803: nlocal = n / size + ((n % size) > rank);
3804: }
3805: } else {
3806: nlocal = csize;
3807: }
3808: PetscCallMPI(MPI_Scan(&nlocal, &rend, 1, MPIU_INT, MPI_SUM, comm));
3809: rstart = rend - nlocal;
3810: 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);
3812: /* next, compute all the lengths */
3813: PetscCall(PetscMalloc1(2 * m + 1, &dlens));
3814: olens = dlens + m;
3815: for (i = 0; i < m; i++) {
3816: jend = ii[i + 1] - ii[i];
3817: olen = 0;
3818: dlen = 0;
3819: for (j = 0; j < jend; j++) {
3820: if (*jj < rstart || *jj >= rend) olen++;
3821: else dlen++;
3822: jj++;
3823: }
3824: olens[i] = olen;
3825: dlens[i] = dlen;
3826: }
3827: PetscCall(MatCreate(comm, &M));
3828: PetscCall(MatSetSizes(M, m, nlocal, PETSC_DECIDE, n));
3829: PetscCall(MatSetBlockSizes(M, bs, cbs));
3830: PetscCall(MatSetType(M, ((PetscObject)mat)->type_name));
3831: PetscCall(MatSetOption(M, MAT_STRUCTURE_ONLY, mat->structure_only));
3832: PetscCall(MatMPIAIJSetPreallocation(M, 0, dlens, 0, olens));
3833: PetscCall(PetscFree(dlens));
3834: } else {
3835: PetscInt ml, nl;
3837: M = *newmat;
3838: PetscCall(MatGetLocalSize(M, &ml, &nl));
3839: PetscCheck(ml == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Previous matrix must be same size/layout as request");
3840: PetscCall(MatZeroEntries(M));
3841: /*
3842: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3843: rather than the slower MatSetValues().
3844: */
3845: M->was_assembled = PETSC_TRUE;
3846: M->assembled = PETSC_FALSE;
3847: }
3848: PetscCall(MatGetOwnershipRange(M, &rstart, &rend));
3849: aij = (Mat_SeqAIJ *)Mreuse->data;
3850: ii = aij->i;
3851: jj = aij->j;
3853: /* trigger copy to CPU if needed */
3854: PetscCall(MatSeqAIJGetArrayRead(Mreuse, (const PetscScalar **)&aa));
3855: for (i = 0; i < m; i++) {
3856: row = rstart + i;
3857: nz = ii[i + 1] - ii[i];
3858: cwork = jj;
3859: jj = PetscSafePointerPlusOffset(jj, nz);
3860: vwork = aa;
3861: aa = PetscSafePointerPlusOffset(aa, nz);
3862: PetscCall(MatSetValues_MPIAIJ(M, 1, &row, nz, cwork, vwork, INSERT_VALUES));
3863: }
3864: PetscCall(MatSeqAIJRestoreArrayRead(Mreuse, (const PetscScalar **)&aa));
3866: PetscCall(MatAssemblyBegin(M, MAT_FINAL_ASSEMBLY));
3867: PetscCall(MatAssemblyEnd(M, MAT_FINAL_ASSEMBLY));
3868: *newmat = M;
3870: /* save submatrix used in processor for next request */
3871: if (call == MAT_INITIAL_MATRIX) {
3872: PetscCall(PetscObjectCompose((PetscObject)M, "SubMatrix", (PetscObject)Mreuse));
3873: PetscCall(MatDestroy(&Mreuse));
3874: }
3875: PetscFunctionReturn(PETSC_SUCCESS);
3876: }
3878: static PetscErrorCode MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[])
3879: {
3880: PetscInt m, cstart, cend, j, nnz, i, d, *ld;
3881: PetscInt *d_nnz, *o_nnz, nnz_max = 0, rstart, ii, irstart;
3882: const PetscInt *JJ;
3883: PetscBool nooffprocentries;
3884: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)B->data;
3886: PetscFunctionBegin;
3887: PetscCall(PetscLayoutSetUp(B->rmap));
3888: PetscCall(PetscLayoutSetUp(B->cmap));
3889: m = B->rmap->n;
3890: cstart = B->cmap->rstart;
3891: cend = B->cmap->rend;
3892: rstart = B->rmap->rstart;
3893: irstart = Ii[0];
3895: PetscCall(PetscCalloc2(m, &d_nnz, m, &o_nnz));
3897: if (PetscDefined(USE_DEBUG)) {
3898: for (i = 0; i < m; i++) {
3899: nnz = Ii[i + 1] - Ii[i];
3900: JJ = PetscSafePointerPlusOffset(J, Ii[i] - irstart);
3901: PetscCheck(nnz >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Local row %" PetscInt_FMT " has a negative %" PetscInt_FMT " number of columns", i, nnz);
3902: 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]);
3903: 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);
3904: }
3905: }
3907: for (i = 0; i < m; i++) {
3908: nnz = Ii[i + 1] - Ii[i];
3909: JJ = PetscSafePointerPlusOffset(J, Ii[i] - irstart);
3910: nnz_max = PetscMax(nnz_max, nnz);
3911: d = 0;
3912: for (j = 0; j < nnz; j++) {
3913: if (cstart <= JJ[j] && JJ[j] < cend) d++;
3914: }
3915: d_nnz[i] = d;
3916: o_nnz[i] = nnz - d;
3917: }
3918: PetscCall(MatMPIAIJSetPreallocation(B, 0, d_nnz, 0, o_nnz));
3919: PetscCall(PetscFree2(d_nnz, o_nnz));
3921: for (i = 0; i < m; i++) {
3922: ii = i + rstart;
3923: PetscCall(MatSetValues_MPIAIJ(B, 1, &ii, Ii[i + 1] - Ii[i], PetscSafePointerPlusOffset(J, Ii[i] - irstart), PetscSafePointerPlusOffset(v, Ii[i] - irstart), INSERT_VALUES));
3924: }
3925: nooffprocentries = B->nooffprocentries;
3926: B->nooffprocentries = PETSC_TRUE;
3927: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
3928: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
3929: B->nooffprocentries = nooffprocentries;
3931: /* count number of entries below block diagonal */
3932: PetscCall(PetscFree(Aij->ld));
3933: PetscCall(PetscCalloc1(m, &ld));
3934: Aij->ld = ld;
3935: for (i = 0; i < m; i++) {
3936: nnz = Ii[i + 1] - Ii[i];
3937: j = 0;
3938: while (j < nnz && J[j] < cstart) j++;
3939: ld[i] = j;
3940: if (J) J += nnz;
3941: }
3943: PetscCall(MatSetOption(B, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
3944: PetscFunctionReturn(PETSC_SUCCESS);
3945: }
3947: /*@
3948: MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in `MATAIJ` format
3949: (the default parallel PETSc format).
3951: Collective
3953: Input Parameters:
3954: + B - the matrix
3955: . i - the indices into `j` for the start of each local row (indices start with zero)
3956: . j - the column indices for each local row (indices start with zero)
3957: - v - optional values in the matrix
3959: Level: developer
3961: Notes:
3962: The `i`, `j`, and `v` arrays ARE copied by this routine into the internal format used by PETSc;
3963: thus you CANNOT change the matrix entries by changing the values of `v` after you have
3964: called this routine. Use `MatCreateMPIAIJWithSplitArrays()` to avoid needing to copy the arrays.
3966: The `i` and `j` indices are 0 based, and `i` indices are indices corresponding to the local `j` array.
3968: A convenience routine for this functionality is `MatCreateMPIAIJWithArrays()`.
3970: You can update the matrix with new numerical values using `MatUpdateMPIAIJWithArrays()` after this call if the column indices in `j` are sorted.
3972: If you do **not** use `MatUpdateMPIAIJWithArrays()`, the column indices in `j` do not need to be sorted. If you will use
3973: `MatUpdateMPIAIJWithArrays()`, the column indices **must** be sorted.
3975: The format which is used for the sparse matrix input, is equivalent to a
3976: row-major ordering.. i.e for the following matrix, the input data expected is
3977: as shown
3978: .vb
3979: 1 0 0
3980: 2 0 3 P0
3981: -------
3982: 4 5 6 P1
3984: Process0 [P0] rows_owned=[0,1]
3985: i = {0,1,3} [size = nrow+1 = 2+1]
3986: j = {0,0,2} [size = 3]
3987: v = {1,2,3} [size = 3]
3989: Process1 [P1] rows_owned=[2]
3990: i = {0,3} [size = nrow+1 = 1+1]
3991: j = {0,1,2} [size = 3]
3992: v = {4,5,6} [size = 3]
3993: .ve
3995: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatCreateAIJ()`,
3996: `MatCreateSeqAIJWithArrays()`, `MatCreateMPIAIJWithSplitArrays()`, `MatCreateMPIAIJWithArrays()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
3997: @*/
3998: PetscErrorCode MatMPIAIJSetPreallocationCSR(Mat B, const PetscInt i[], const PetscInt j[], const PetscScalar v[])
3999: {
4000: PetscFunctionBegin;
4001: PetscTryMethod(B, "MatMPIAIJSetPreallocationCSR_C", (Mat, const PetscInt[], const PetscInt[], const PetscScalar[]), (B, i, j, v));
4002: PetscFunctionReturn(PETSC_SUCCESS);
4003: }
4005: /*@
4006: MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in `MATMPIAIJ` format
4007: (the default parallel PETSc format). For good matrix assembly performance
4008: the user should preallocate the matrix storage by setting the parameters
4009: `d_nz` (or `d_nnz`) and `o_nz` (or `o_nnz`).
4011: Collective
4013: Input Parameters:
4014: + B - the matrix
4015: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
4016: (same value is used for all local rows)
4017: . d_nnz - array containing the number of nonzeros in the various rows of the
4018: DIAGONAL portion of the local submatrix (possibly different for each row)
4019: or `NULL` (`PETSC_NULL_INTEGER` in Fortran), if `d_nz` is used to specify the nonzero structure.
4020: The size of this array is equal to the number of local rows, i.e 'm'.
4021: For matrices that will be factored, you must leave room for (and set)
4022: the diagonal entry even if it is zero.
4023: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
4024: submatrix (same value is used for all local rows).
4025: - o_nnz - array containing the number of nonzeros in the various rows of the
4026: OFF-DIAGONAL portion of the local submatrix (possibly different for
4027: each row) or `NULL` (`PETSC_NULL_INTEGER` in Fortran), if `o_nz` is used to specify the nonzero
4028: structure. The size of this array is equal to the number
4029: of local rows, i.e 'm'.
4031: Example Usage:
4032: Consider the following 8x8 matrix with 34 non-zero values, that is
4033: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4034: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4035: as follows
4037: .vb
4038: 1 2 0 | 0 3 0 | 0 4
4039: Proc0 0 5 6 | 7 0 0 | 8 0
4040: 9 0 10 | 11 0 0 | 12 0
4041: -------------------------------------
4042: 13 0 14 | 15 16 17 | 0 0
4043: Proc1 0 18 0 | 19 20 21 | 0 0
4044: 0 0 0 | 22 23 0 | 24 0
4045: -------------------------------------
4046: Proc2 25 26 27 | 0 0 28 | 29 0
4047: 30 0 0 | 31 32 33 | 0 34
4048: .ve
4050: This can be represented as a collection of submatrices as
4051: .vb
4052: A B C
4053: D E F
4054: G H I
4055: .ve
4057: Where the submatrices A,B,C are owned by proc0, D,E,F are
4058: owned by proc1, G,H,I are owned by proc2.
4060: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4061: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4062: The 'M','N' parameters are 8,8, and have the same values on all procs.
4064: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4065: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4066: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4067: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4068: part as `MATSEQAIJ` matrices. For example, proc1 will store [E] as a `MATSEQAIJ`
4069: matrix, and [DF] as another `MATSEQAIJ` matrix.
4071: When `d_nz`, `o_nz` parameters are specified, `d_nz` storage elements are
4072: allocated for every row of the local DIAGONAL submatrix, and `o_nz`
4073: storage locations are allocated for every row of the OFF-DIAGONAL submatrix.
4074: One way to choose `d_nz` and `o_nz` is to use the maximum number of nonzeros over
4075: the local rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4076: In this case, the values of `d_nz`, `o_nz` are
4077: .vb
4078: proc0 dnz = 2, o_nz = 2
4079: proc1 dnz = 3, o_nz = 2
4080: proc2 dnz = 1, o_nz = 4
4081: .ve
4082: We are allocating `m`*(`d_nz`+`o_nz`) storage locations for every proc. This
4083: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4084: for proc3. i.e we are using 12+15+10=37 storage locations to store
4085: 34 values.
4087: When `d_nnz`, `o_nnz` parameters are specified, the storage is specified
4088: for every row, corresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4089: In the above case the values for `d_nnz`, `o_nnz` are
4090: .vb
4091: proc0 d_nnz = [2,2,2] and o_nnz = [2,2,2]
4092: proc1 d_nnz = [3,3,2] and o_nnz = [2,1,1]
4093: proc2 d_nnz = [1,1] and o_nnz = [4,4]
4094: .ve
4095: Here the space allocated is sum of all the above values i.e 34, and
4096: hence pre-allocation is perfect.
4098: Level: intermediate
4100: Notes:
4101: If the *_nnz parameter is given then the *_nz parameter is ignored
4103: The `MATAIJ` format, also called compressed row storage (CSR), is compatible with standard Fortran
4104: storage. The stored row and column indices begin with zero.
4105: See [Sparse Matrices](sec_matsparse) for details.
4107: The parallel matrix is partitioned such that the first m0 rows belong to
4108: process 0, the next m1 rows belong to process 1, the next m2 rows belong
4109: to process 2 etc.. where m0,m1,m2... are the input parameter 'm'.
4111: The DIAGONAL portion of the local submatrix of a processor can be defined
4112: as the submatrix which is obtained by extraction the part corresponding to
4113: the rows r1-r2 and columns c1-c2 of the global matrix, where r1 is the
4114: first row that belongs to the processor, r2 is the last row belonging to
4115: the this processor, and c1-c2 is range of indices of the local part of a
4116: vector suitable for applying the matrix to. This is an mxn matrix. In the
4117: common case of a square matrix, the row and column ranges are the same and
4118: the DIAGONAL part is also square. The remaining portion of the local
4119: submatrix (mxN) constitute the OFF-DIAGONAL portion.
4121: If `o_nnz` and `d_nnz` are specified, then `o_nz` and `d_nz` are ignored.
4123: You can call `MatGetInfo()` to get information on how effective the preallocation was;
4124: for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
4125: You can also run with the option `-info` and look for messages with the string
4126: malloc in them to see if additional memory allocation was needed.
4128: .seealso: [](ch_matrices), `Mat`, [Sparse Matrices](sec_matsparse), `MATMPIAIJ`, `MATAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatCreateAIJ()`, `MatMPIAIJSetPreallocationCSR()`,
4129: `MatGetInfo()`, `PetscSplitOwnership()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4130: @*/
4131: PetscErrorCode MatMPIAIJSetPreallocation(Mat B, PetscInt d_nz, const PetscInt d_nnz[], PetscInt o_nz, const PetscInt o_nnz[])
4132: {
4133: PetscFunctionBegin;
4136: PetscTryMethod(B, "MatMPIAIJSetPreallocation_C", (Mat, PetscInt, const PetscInt[], PetscInt, const PetscInt[]), (B, d_nz, d_nnz, o_nz, o_nnz));
4137: PetscFunctionReturn(PETSC_SUCCESS);
4138: }
4140: /*@
4141: MatCreateMPIAIJWithArrays - creates a `MATMPIAIJ` matrix using arrays that contain in standard
4142: CSR format for the local rows.
4144: Collective
4146: Input Parameters:
4147: + comm - MPI communicator
4148: . m - number of local rows (Cannot be `PETSC_DECIDE`)
4149: . n - This value should be the same as the local size used in creating the
4150: x vector for the matrix-vector product $ y = Ax$. (or `PETSC_DECIDE` to have
4151: calculated if `N` is given) For square matrices n is almost always `m`.
4152: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if `m` is given)
4153: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if `n` is given)
4154: . 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
4155: . j - global column indices
4156: - a - optional matrix values
4158: Output Parameter:
4159: . mat - the matrix
4161: Level: intermediate
4163: Notes:
4164: The `i`, `j`, and `a` arrays ARE copied by this routine into the internal format used by PETSc;
4165: thus you CANNOT change the matrix entries by changing the values of `a[]` after you have
4166: called this routine. Use `MatCreateMPIAIJWithSplitArrays()` to avoid needing to copy the arrays.
4168: The `i` and `j` indices are 0 based, and `i` indices are indices corresponding to the local `j` array.
4170: Once you have created the matrix you can update it with new numerical values using `MatUpdateMPIAIJWithArray()`
4172: If you do **not** use `MatUpdateMPIAIJWithArray()`, the column indices in `j` do not need to be sorted. If you will use
4173: `MatUpdateMPIAIJWithArrays()`, the column indices **must** be sorted.
4175: The format which is used for the sparse matrix input, is equivalent to a
4176: row-major ordering, i.e., for the following matrix, the input data expected is
4177: as shown
4178: .vb
4179: 1 0 0
4180: 2 0 3 P0
4181: -------
4182: 4 5 6 P1
4184: Process0 [P0] rows_owned=[0,1]
4185: i = {0,1,3} [size = nrow+1 = 2+1]
4186: j = {0,0,2} [size = 3]
4187: v = {1,2,3} [size = 3]
4189: Process1 [P1] rows_owned=[2]
4190: i = {0,3} [size = nrow+1 = 1+1]
4191: j = {0,1,2} [size = 3]
4192: v = {4,5,6} [size = 3]
4193: .ve
4195: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4196: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArray()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4197: @*/
4198: PetscErrorCode MatCreateMPIAIJWithArrays(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt M, PetscInt N, const PetscInt i[], const PetscInt j[], const PetscScalar a[], Mat *mat)
4199: {
4200: PetscFunctionBegin;
4201: PetscCheck(!i || !i[0], PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
4202: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
4203: PetscCall(MatCreate(comm, mat));
4204: PetscCall(MatSetSizes(*mat, m, n, M, N));
4205: /* PetscCall(MatSetBlockSizes(M,bs,cbs)); */
4206: PetscCall(MatSetType(*mat, MATMPIAIJ));
4207: PetscCall(MatMPIAIJSetPreallocationCSR(*mat, i, j, a));
4208: PetscFunctionReturn(PETSC_SUCCESS);
4209: }
4211: /*@
4212: MatUpdateMPIAIJWithArrays - updates a `MATMPIAIJ` matrix using arrays that contain in standard
4213: CSR format for the local rows. Only the numerical values are updated the other arrays must be identical to what was passed
4214: from `MatCreateMPIAIJWithArrays()`
4216: Deprecated: Use `MatUpdateMPIAIJWithArray()`
4218: Collective
4220: Input Parameters:
4221: + mat - the matrix
4222: . m - number of local rows (Cannot be `PETSC_DECIDE`)
4223: . n - This value should be the same as the local size used in creating the
4224: x vector for the matrix-vector product y = Ax. (or `PETSC_DECIDE` to have
4225: calculated if N is given) For square matrices n is almost always m.
4226: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if m is given)
4227: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if n is given)
4228: . Ii - row indices; that is Ii[0] = 0, Ii[row] = Ii[row-1] + number of elements in that row of the matrix
4229: . J - column indices
4230: - v - matrix values
4232: Level: deprecated
4234: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4235: `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArray()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4236: @*/
4237: PetscErrorCode MatUpdateMPIAIJWithArrays(Mat mat, PetscInt m, PetscInt n, PetscInt M, PetscInt N, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[])
4238: {
4239: PetscInt nnz, i;
4240: PetscBool nooffprocentries;
4241: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)mat->data;
4242: Mat_SeqAIJ *Ad = (Mat_SeqAIJ *)Aij->A->data;
4243: PetscScalar *ad, *ao;
4244: PetscInt ldi, Iii, md;
4245: const PetscInt *Adi = Ad->i;
4246: PetscInt *ld = Aij->ld;
4248: PetscFunctionBegin;
4249: PetscCheck(Ii[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
4250: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
4251: PetscCheck(m == mat->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Local number of rows cannot change from call to MatUpdateMPIAIJWithArrays()");
4252: PetscCheck(n == mat->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Local number of columns cannot change from call to MatUpdateMPIAIJWithArrays()");
4254: PetscCall(MatSeqAIJGetArrayWrite(Aij->A, &ad));
4255: PetscCall(MatSeqAIJGetArrayWrite(Aij->B, &ao));
4257: for (i = 0; i < m; i++) {
4258: if (PetscDefined(USE_DEBUG)) {
4259: for (PetscInt j = Ii[i] + 1; j < Ii[i + 1]; ++j) {
4260: 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);
4261: 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);
4262: }
4263: }
4264: nnz = Ii[i + 1] - Ii[i];
4265: Iii = Ii[i];
4266: ldi = ld[i];
4267: md = Adi[i + 1] - Adi[i];
4268: PetscCall(PetscArraycpy(ao, v + Iii, ldi));
4269: PetscCall(PetscArraycpy(ad, v + Iii + ldi, md));
4270: PetscCall(PetscArraycpy(ao + ldi, v + Iii + ldi + md, nnz - ldi - md));
4271: ad += md;
4272: ao += nnz - md;
4273: }
4274: nooffprocentries = mat->nooffprocentries;
4275: mat->nooffprocentries = PETSC_TRUE;
4276: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->A, &ad));
4277: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->B, &ao));
4278: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->A));
4279: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->B));
4280: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
4281: PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
4282: PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
4283: mat->nooffprocentries = nooffprocentries;
4284: PetscFunctionReturn(PETSC_SUCCESS);
4285: }
4287: /*@
4288: MatUpdateMPIAIJWithArray - updates an `MATMPIAIJ` matrix using an array that contains the nonzero values
4290: Collective
4292: Input Parameters:
4293: + mat - the matrix
4294: - v - matrix values, stored by row
4296: Level: intermediate
4298: Notes:
4299: The matrix must have been obtained with `MatCreateMPIAIJWithArrays()` or `MatMPIAIJSetPreallocationCSR()`
4301: The column indices in the call to `MatCreateMPIAIJWithArrays()` or `MatMPIAIJSetPreallocationCSR()` must have been sorted for this call to work correctly
4303: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4304: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArrays()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4305: @*/
4306: PetscErrorCode MatUpdateMPIAIJWithArray(Mat mat, const PetscScalar v[])
4307: {
4308: PetscInt nnz, i, m;
4309: PetscBool nooffprocentries;
4310: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)mat->data;
4311: Mat_SeqAIJ *Ad = (Mat_SeqAIJ *)Aij->A->data;
4312: Mat_SeqAIJ *Ao = (Mat_SeqAIJ *)Aij->B->data;
4313: PetscScalar *ad, *ao;
4314: const PetscInt *Adi = Ad->i, *Adj = Ao->i;
4315: PetscInt ldi, Iii, md;
4316: PetscInt *ld = Aij->ld;
4318: PetscFunctionBegin;
4319: m = mat->rmap->n;
4321: PetscCall(MatSeqAIJGetArrayWrite(Aij->A, &ad));
4322: PetscCall(MatSeqAIJGetArrayWrite(Aij->B, &ao));
4323: Iii = 0;
4324: for (i = 0; i < m; i++) {
4325: nnz = Adi[i + 1] - Adi[i] + Adj[i + 1] - Adj[i];
4326: ldi = ld[i];
4327: md = Adi[i + 1] - Adi[i];
4328: PetscCall(PetscArraycpy(ad, v + Iii + ldi, md));
4329: ad += md;
4330: if (ao) {
4331: PetscCall(PetscArraycpy(ao, v + Iii, ldi));
4332: PetscCall(PetscArraycpy(ao + ldi, v + Iii + ldi + md, nnz - ldi - md));
4333: ao += nnz - md;
4334: }
4335: Iii += nnz;
4336: }
4337: nooffprocentries = mat->nooffprocentries;
4338: mat->nooffprocentries = PETSC_TRUE;
4339: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->A, &ad));
4340: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->B, &ao));
4341: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->A));
4342: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->B));
4343: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
4344: PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
4345: PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
4346: mat->nooffprocentries = nooffprocentries;
4347: PetscFunctionReturn(PETSC_SUCCESS);
4348: }
4350: /*@
4351: MatCreateAIJ - Creates a sparse parallel matrix in `MATAIJ` format
4352: (the default parallel PETSc format). For good matrix assembly performance
4353: the user should preallocate the matrix storage by setting the parameters
4354: `d_nz` (or `d_nnz`) and `o_nz` (or `o_nnz`).
4356: Collective
4358: Input Parameters:
4359: + comm - MPI communicator
4360: . m - number of local rows (or `PETSC_DECIDE` to have calculated if M is given)
4361: This value should be the same as the local size used in creating the
4362: y vector for the matrix-vector product y = Ax.
4363: . n - This value should be the same as the local size used in creating the
4364: x vector for the matrix-vector product y = Ax. (or `PETSC_DECIDE` to have
4365: calculated if N is given) For square matrices n is almost always m.
4366: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if m is given)
4367: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if n is given)
4368: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
4369: (same value is used for all local rows)
4370: . d_nnz - array containing the number of nonzeros in the various rows of the
4371: DIAGONAL portion of the local submatrix (possibly different for each row)
4372: or `NULL`, if `d_nz` is used to specify the nonzero structure.
4373: The size of this array is equal to the number of local rows, i.e 'm'.
4374: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
4375: submatrix (same value is used for all local rows).
4376: - o_nnz - array containing the number of nonzeros in the various rows of the
4377: OFF-DIAGONAL portion of the local submatrix (possibly different for
4378: each row) or `NULL`, if `o_nz` is used to specify the nonzero
4379: structure. The size of this array is equal to the number
4380: of local rows, i.e 'm'.
4382: Output Parameter:
4383: . A - the matrix
4385: Options Database Keys:
4386: + -mat_no_inode - Do not use inodes
4387: . -mat_inode_limit limit - Sets inode limit (max limit=5)
4388: - -matmult_vecscatter_view viewer - View the vecscatter (i.e., communication pattern) used in `MatMult()` of sparse parallel matrices.
4389: See viewer types in manual of `MatView()`. Of them, ascii_matlab, draw or binary cause the `VecScatter`
4390: 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.
4392: Level: intermediate
4394: Notes:
4395: It is recommended that one use `MatCreateFromOptions()` or the `MatCreate()`, `MatSetType()` and/or `MatSetFromOptions()`,
4396: MatXXXXSetPreallocation() paradigm instead of this routine directly.
4397: [MatXXXXSetPreallocation() is, for example, `MatSeqAIJSetPreallocation()`]
4399: If the *_nnz parameter is given then the *_nz parameter is ignored
4401: The `m`,`n`,`M`,`N` parameters specify the size of the matrix, and its partitioning across
4402: processors, while `d_nz`,`d_nnz`,`o_nz`,`o_nnz` parameters specify the approximate
4403: storage requirements for this matrix.
4405: If `PETSC_DECIDE` or `PETSC_DETERMINE` is used for a particular argument on one
4406: processor than it must be used on all processors that share the object for
4407: that argument.
4409: If `m` and `n` are not `PETSC_DECIDE`, then the values determine the `PetscLayout` of the matrix and the ranges returned by
4410: `MatGetOwnershipRange()`, `MatGetOwnershipRanges()`, `MatGetOwnershipRangeColumn()`, and `MatGetOwnershipRangesColumn()`.
4412: The user MUST specify either the local or global matrix dimensions
4413: (possibly both).
4415: The parallel matrix is partitioned across processors such that the
4416: first `m0` rows belong to process 0, the next `m1` rows belong to
4417: process 1, the next `m2` rows belong to process 2, etc., where
4418: `m0`, `m1`, `m2`... are the input parameter `m` on each MPI process. I.e., each MPI process stores
4419: values corresponding to [m x N] submatrix.
4421: The columns are logically partitioned with the n0 columns belonging
4422: to 0th partition, the next n1 columns belonging to the next
4423: partition etc.. where n0,n1,n2... are the input parameter 'n'.
4425: The DIAGONAL portion of the local submatrix on any given processor
4426: is the submatrix corresponding to the rows and columns m,n
4427: corresponding to the given processor. i.e diagonal matrix on
4428: process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1]
4429: etc. The remaining portion of the local submatrix [m x (N-n)]
4430: constitute the OFF-DIAGONAL portion. The example below better
4431: illustrates this concept. The two matrices, the DIAGONAL portion and
4432: the OFF-DIAGONAL portion are each stored as `MATSEQAIJ` matrices.
4434: For a square global matrix we define each processor's diagonal portion
4435: to be its local rows and the corresponding columns (a square submatrix);
4436: each processor's off-diagonal portion encompasses the remainder of the
4437: local matrix (a rectangular submatrix).
4439: If `o_nnz`, `d_nnz` are specified, then `o_nz`, and `d_nz` are ignored.
4441: When calling this routine with a single process communicator, a matrix of
4442: type `MATSEQAIJ` is returned. If a matrix of type `MATMPIAIJ` is desired for this
4443: type of communicator, use the construction mechanism
4444: .vb
4445: MatCreate(..., &A);
4446: MatSetType(A, MATMPIAIJ);
4447: MatSetSizes(A, m, n, M, N);
4448: MatMPIAIJSetPreallocation(A, ...);
4449: .ve
4451: By default, this format uses inodes (identical nodes) when possible.
4452: We search for consecutive rows with the same nonzero structure, thereby
4453: reusing matrix information to achieve increased efficiency.
4455: Example Usage:
4456: Consider the following 8x8 matrix with 34 non-zero values, that is
4457: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4458: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4459: as follows
4461: .vb
4462: 1 2 0 | 0 3 0 | 0 4
4463: Proc0 0 5 6 | 7 0 0 | 8 0
4464: 9 0 10 | 11 0 0 | 12 0
4465: -------------------------------------
4466: 13 0 14 | 15 16 17 | 0 0
4467: Proc1 0 18 0 | 19 20 21 | 0 0
4468: 0 0 0 | 22 23 0 | 24 0
4469: -------------------------------------
4470: Proc2 25 26 27 | 0 0 28 | 29 0
4471: 30 0 0 | 31 32 33 | 0 34
4472: .ve
4474: This can be represented as a collection of submatrices as
4476: .vb
4477: A B C
4478: D E F
4479: G H I
4480: .ve
4482: Where the submatrices A,B,C are owned by proc0, D,E,F are
4483: owned by proc1, G,H,I are owned by proc2.
4485: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4486: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4487: The 'M','N' parameters are 8,8, and have the same values on all procs.
4489: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4490: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4491: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4492: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4493: part as `MATSEQAIJ` matrices. For example, proc1 will store [E] as a `MATSEQAIJ`
4494: matrix, and [DF] as another SeqAIJ matrix.
4496: When `d_nz`, `o_nz` parameters are specified, `d_nz` storage elements are
4497: allocated for every row of the local DIAGONAL submatrix, and `o_nz`
4498: storage locations are allocated for every row of the OFF-DIAGONAL submatrix.
4499: One way to choose `d_nz` and `o_nz` is to use the maximum number of nonzeros over
4500: the local rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4501: In this case, the values of `d_nz`,`o_nz` are
4502: .vb
4503: proc0 dnz = 2, o_nz = 2
4504: proc1 dnz = 3, o_nz = 2
4505: proc2 dnz = 1, o_nz = 4
4506: .ve
4507: We are allocating m*(`d_nz`+`o_nz`) storage locations for every proc. This
4508: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4509: for proc3. i.e we are using 12+15+10=37 storage locations to store
4510: 34 values.
4512: When `d_nnz`, `o_nnz` parameters are specified, the storage is specified
4513: for every row, corresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4514: In the above case the values for d_nnz,o_nnz are
4515: .vb
4516: proc0 d_nnz = [2,2,2] and o_nnz = [2,2,2]
4517: proc1 d_nnz = [3,3,2] and o_nnz = [2,1,1]
4518: proc2 d_nnz = [1,1] and o_nnz = [4,4]
4519: .ve
4520: Here the space allocated is sum of all the above values i.e 34, and
4521: hence pre-allocation is perfect.
4523: .seealso: [](ch_matrices), `Mat`, [Sparse Matrix Creation](sec_matsparse), `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4524: `MATMPIAIJ`, `MatCreateMPIAIJWithArrays()`, `MatGetOwnershipRange()`, `MatGetOwnershipRanges()`, `MatGetOwnershipRangeColumn()`,
4525: `MatGetOwnershipRangesColumn()`, `PetscLayout`
4526: @*/
4527: 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)
4528: {
4529: PetscMPIInt size;
4531: PetscFunctionBegin;
4532: PetscCall(MatCreate(comm, A));
4533: PetscCall(MatSetSizes(*A, m, n, M, N));
4534: PetscCallMPI(MPI_Comm_size(comm, &size));
4535: if (size > 1) {
4536: PetscCall(MatSetType(*A, MATMPIAIJ));
4537: PetscCall(MatMPIAIJSetPreallocation(*A, d_nz, d_nnz, o_nz, o_nnz));
4538: } else {
4539: PetscCall(MatSetType(*A, MATSEQAIJ));
4540: PetscCall(MatSeqAIJSetPreallocation(*A, d_nz, d_nnz));
4541: }
4542: PetscFunctionReturn(PETSC_SUCCESS);
4543: }
4545: /*@
4546: MatMPIAIJGetSeqAIJ - Returns the local pieces of this distributed matrix
4548: Not Collective
4550: Input Parameter:
4551: . A - The `MATMPIAIJ` matrix
4553: Output Parameters:
4554: + Ad - The local diagonal block as a `MATSEQAIJ` matrix
4555: . Ao - The local off-diagonal block as a `MATSEQAIJ` matrix
4556: - colmap - An array mapping local column numbers of `Ao` to global column numbers of the parallel matrix
4558: Level: intermediate
4560: Note:
4561: The rows in `Ad` and `Ao` are in [0, Nr), where Nr is the number of local rows on this process. The columns
4562: 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
4563: the number of nonzero columns in the local off-diagonal piece of the matrix `A`. The array colmap maps these
4564: local column numbers to global column numbers in the original matrix.
4566: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatMPIAIJGetLocalMat()`, `MatMPIAIJGetLocalMatCondensed()`, `MatCreateAIJ()`, `MATSEQAIJ`
4567: @*/
4568: PetscErrorCode MatMPIAIJGetSeqAIJ(Mat A, Mat *Ad, Mat *Ao, const PetscInt *colmap[])
4569: {
4570: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
4571: PetscBool flg;
4573: PetscFunctionBegin;
4574: PetscCall(PetscStrbeginswith(((PetscObject)A)->type_name, MATMPIAIJ, &flg));
4575: PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "This function requires a MATMPIAIJ matrix as input");
4576: if (Ad) *Ad = a->A;
4577: if (Ao) *Ao = a->B;
4578: if (colmap) *colmap = a->garray;
4579: PetscFunctionReturn(PETSC_SUCCESS);
4580: }
4582: static PetscErrorCode MatGetMultPetscSF_MPIAIJ(Mat A, PetscSF *sf)
4583: {
4584: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
4586: PetscFunctionBegin;
4587: *sf = a->Mvctx;
4588: PetscFunctionReturn(PETSC_SUCCESS);
4589: }
4591: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm comm, Mat inmat, PetscInt n, MatReuse scall, Mat *outmat)
4592: {
4593: PetscInt m, N, i, rstart, nnz, Ii;
4594: PetscInt *indx;
4595: PetscScalar *values;
4596: MatType rootType;
4598: PetscFunctionBegin;
4599: PetscCall(MatGetSize(inmat, &m, &N));
4600: if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */
4601: PetscInt *dnz, *onz, sum, bs, cbs;
4603: if (n == PETSC_DECIDE) PetscCall(PetscSplitOwnership(comm, &n, &N));
4604: /* Check sum(n) = N */
4605: PetscCallMPI(MPIU_Allreduce(&n, &sum, 1, MPIU_INT, MPI_SUM, comm));
4606: PetscCheck(sum == N, PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Sum of local columns %" PetscInt_FMT " != global columns %" PetscInt_FMT, sum, N);
4608: PetscCallMPI(MPI_Scan(&m, &rstart, 1, MPIU_INT, MPI_SUM, comm));
4609: rstart -= m;
4611: MatPreallocateBegin(comm, m, n, dnz, onz);
4612: for (i = 0; i < m; i++) {
4613: PetscCall(MatGetRow_SeqAIJ(inmat, i, &nnz, &indx, NULL));
4614: PetscCall(MatPreallocateSet(i + rstart, nnz, indx, dnz, onz));
4615: PetscCall(MatRestoreRow_SeqAIJ(inmat, i, &nnz, &indx, NULL));
4616: }
4618: PetscCall(MatCreate(comm, outmat));
4619: PetscCall(MatSetSizes(*outmat, m, n, PETSC_DETERMINE, PETSC_DETERMINE));
4620: PetscCall(MatGetBlockSizes(inmat, &bs, &cbs));
4621: PetscCall(MatSetBlockSizes(*outmat, bs, cbs));
4622: PetscCall(MatGetRootType_Private(inmat, &rootType));
4623: PetscCall(MatSetType(*outmat, rootType));
4624: PetscCall(MatSeqAIJSetPreallocation(*outmat, 0, dnz));
4625: PetscCall(MatMPIAIJSetPreallocation(*outmat, 0, dnz, 0, onz));
4626: MatPreallocateEnd(dnz, onz);
4627: PetscCall(MatSetOption(*outmat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
4628: }
4630: /* numeric phase */
4631: PetscCall(MatGetOwnershipRange(*outmat, &rstart, NULL));
4632: for (i = 0; i < m; i++) {
4633: PetscCall(MatGetRow_SeqAIJ(inmat, i, &nnz, &indx, &values));
4634: Ii = i + rstart;
4635: PetscCall(MatSetValues(*outmat, 1, &Ii, nnz, indx, values, INSERT_VALUES));
4636: PetscCall(MatRestoreRow_SeqAIJ(inmat, i, &nnz, &indx, &values));
4637: }
4638: PetscCall(MatAssemblyBegin(*outmat, MAT_FINAL_ASSEMBLY));
4639: PetscCall(MatAssemblyEnd(*outmat, MAT_FINAL_ASSEMBLY));
4640: PetscFunctionReturn(PETSC_SUCCESS);
4641: }
4643: static PetscErrorCode MatMergeSeqsToMPIDestroy(PetscCtxRt data)
4644: {
4645: MatMergeSeqsToMPI *merge = *(MatMergeSeqsToMPI **)data;
4647: PetscFunctionBegin;
4648: if (!merge) PetscFunctionReturn(PETSC_SUCCESS);
4649: PetscCall(PetscFree(merge->id_r));
4650: PetscCall(PetscFree(merge->len_s));
4651: PetscCall(PetscFree(merge->len_r));
4652: PetscCall(PetscFree(merge->bi));
4653: PetscCall(PetscFree(merge->bj));
4654: PetscCall(PetscFree(merge->buf_ri[0]));
4655: PetscCall(PetscFree(merge->buf_ri));
4656: PetscCall(PetscFree(merge->buf_rj[0]));
4657: PetscCall(PetscFree(merge->buf_rj));
4658: PetscCall(PetscFree(merge->coi));
4659: PetscCall(PetscFree(merge->coj));
4660: PetscCall(PetscFree(merge->owners_co));
4661: PetscCall(PetscLayoutDestroy(&merge->rowmap));
4662: PetscCall(PetscFree(merge));
4663: PetscFunctionReturn(PETSC_SUCCESS);
4664: }
4666: #include <../src/mat/utils/freespace.h>
4667: #include <petscbt.h>
4669: /*@
4670: MatCreateMPIAIJSumSeqAIJNumeric - Fill the numerical values of an `MATMPIAIJ` matrix previously created by
4671: `MatCreateMPIAIJSumSeqAIJSymbolic()` by summing the local `MATSEQAIJ` contributions from each process.
4673: Collective
4675: Input Parameters:
4676: + seqmat - the local `MATSEQAIJ` contribution from this process
4677: - mpimat - the target `MATMPIAIJ` matrix created by `MatCreateMPIAIJSumSeqAIJSymbolic()`
4679: Level: developer
4681: .seealso: `Mat`, `MATMPIAIJ`, `MATSEQAIJ`, `MatCreateMPIAIJSumSeqAIJSymbolic()`, `MatCreateMPIAIJSumSeqAIJ()`
4682: @*/
4683: PetscErrorCode MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat, Mat mpimat)
4684: {
4685: MPI_Comm comm;
4686: Mat_SeqAIJ *a = (Mat_SeqAIJ *)seqmat->data;
4687: PetscMPIInt size, rank, taga, *len_s;
4688: PetscInt N = mpimat->cmap->N, i, j, *owners, *ai = a->i, *aj, m;
4689: PetscMPIInt proc, k;
4690: PetscInt **buf_ri, **buf_rj;
4691: PetscInt anzi, *bj_i, *bi, *bj, arow, bnzi, nextaj;
4692: PetscInt nrows, **buf_ri_k, **nextrow, **nextai;
4693: MPI_Request *s_waits, *r_waits;
4694: MPI_Status *status;
4695: const MatScalar *aa, *a_a;
4696: MatScalar **abuf_r, *ba_i;
4697: MatMergeSeqsToMPI *merge;
4698: PetscContainer container;
4700: PetscFunctionBegin;
4701: PetscCall(PetscObjectGetComm((PetscObject)mpimat, &comm));
4702: PetscCall(PetscLogEventBegin(MAT_Seqstompinum, seqmat, 0, 0, 0));
4704: PetscCallMPI(MPI_Comm_size(comm, &size));
4705: PetscCallMPI(MPI_Comm_rank(comm, &rank));
4707: PetscCall(PetscObjectQuery((PetscObject)mpimat, "MatMergeSeqsToMPI", (PetscObject *)&container));
4708: PetscCheck(container, PetscObjectComm((PetscObject)mpimat), PETSC_ERR_PLIB, "Mat not created from MatCreateMPIAIJSumSeqAIJSymbolic");
4709: PetscCall(PetscContainerGetPointer(container, &merge));
4710: PetscCall(MatSeqAIJGetArrayRead(seqmat, &a_a));
4711: aa = a_a;
4713: bi = merge->bi;
4714: bj = merge->bj;
4715: buf_ri = merge->buf_ri;
4716: buf_rj = merge->buf_rj;
4718: PetscCall(PetscMalloc1(size, &status));
4719: owners = merge->rowmap->range;
4720: len_s = merge->len_s;
4722: /* send and recv matrix values */
4723: PetscCall(PetscObjectGetNewTag((PetscObject)mpimat, &taga));
4724: PetscCall(PetscPostIrecvScalar(comm, taga, merge->nrecv, merge->id_r, merge->len_r, &abuf_r, &r_waits));
4726: PetscCall(PetscMalloc1(merge->nsend + 1, &s_waits));
4727: for (proc = 0, k = 0; proc < size; proc++) {
4728: if (!len_s[proc]) continue;
4729: i = owners[proc];
4730: PetscCallMPI(MPIU_Isend(aa + ai[i], len_s[proc], MPIU_MATSCALAR, proc, taga, comm, s_waits + k));
4731: k++;
4732: }
4734: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, r_waits, status));
4735: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, s_waits, status));
4736: PetscCall(PetscFree(status));
4738: PetscCall(PetscFree(s_waits));
4739: PetscCall(PetscFree(r_waits));
4741: /* insert mat values of mpimat */
4742: PetscCall(PetscMalloc1(N, &ba_i));
4743: PetscCall(PetscMalloc3(merge->nrecv, &buf_ri_k, merge->nrecv, &nextrow, merge->nrecv, &nextai));
4745: for (k = 0; k < merge->nrecv; k++) {
4746: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4747: nrows = *buf_ri_k[k];
4748: nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */
4749: nextai[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure */
4750: }
4752: /* set values of ba */
4753: m = merge->rowmap->n;
4754: for (i = 0; i < m; i++) {
4755: arow = owners[rank] + i;
4756: bj_i = bj + bi[i]; /* col indices of the i-th row of mpimat */
4757: bnzi = bi[i + 1] - bi[i];
4758: PetscCall(PetscArrayzero(ba_i, bnzi));
4760: /* add local non-zero vals of this proc's seqmat into ba */
4761: anzi = ai[arow + 1] - ai[arow];
4762: aj = a->j + ai[arow];
4763: aa = a_a + ai[arow];
4764: nextaj = 0;
4765: for (j = 0; nextaj < anzi; j++) {
4766: if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4767: ba_i[j] += aa[nextaj++];
4768: }
4769: }
4771: /* add received vals into ba */
4772: for (k = 0; k < merge->nrecv; k++) { /* k-th received message */
4773: /* i-th row */
4774: if (i == *nextrow[k]) {
4775: anzi = *(nextai[k] + 1) - *nextai[k];
4776: aj = buf_rj[k] + *nextai[k];
4777: aa = abuf_r[k] + *nextai[k];
4778: nextaj = 0;
4779: for (j = 0; nextaj < anzi; j++) {
4780: if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4781: ba_i[j] += aa[nextaj++];
4782: }
4783: }
4784: nextrow[k]++;
4785: nextai[k]++;
4786: }
4787: }
4788: PetscCall(MatSetValues(mpimat, 1, &arow, bnzi, bj_i, ba_i, INSERT_VALUES));
4789: }
4790: PetscCall(MatSeqAIJRestoreArrayRead(seqmat, &a_a));
4791: PetscCall(MatAssemblyBegin(mpimat, MAT_FINAL_ASSEMBLY));
4792: PetscCall(MatAssemblyEnd(mpimat, MAT_FINAL_ASSEMBLY));
4794: PetscCall(PetscFree(abuf_r[0]));
4795: PetscCall(PetscFree(abuf_r));
4796: PetscCall(PetscFree(ba_i));
4797: PetscCall(PetscFree3(buf_ri_k, nextrow, nextai));
4798: PetscCall(PetscLogEventEnd(MAT_Seqstompinum, seqmat, 0, 0, 0));
4799: PetscFunctionReturn(PETSC_SUCCESS);
4800: }
4802: /*@
4803: MatCreateMPIAIJSumSeqAIJSymbolic - Create the symbolic (nonzero-pattern) portion of an `MATMPIAIJ` matrix
4804: obtained by summing local `MATSEQAIJ` contributions from each process.
4806: Collective
4808: Input Parameters:
4809: + comm - the communicator
4810: . seqmat - the local `MATSEQAIJ` contribution from this process
4811: . m - the number of local rows for the resulting `MATMPIAIJ` matrix, or `PETSC_DECIDE`
4812: - n - the number of local columns for the resulting `MATMPIAIJ` matrix, or `PETSC_DECIDE`
4814: Output Parameter:
4815: . mpimat - the newly created `MATMPIAIJ` matrix
4817: Level: developer
4819: Note:
4820: The numerical values are filled in by a subsequent call to `MatCreateMPIAIJSumSeqAIJNumeric()`.
4822: .seealso: `Mat`, `MATMPIAIJ`, `MATSEQAIJ`, `MatCreateMPIAIJSumSeqAIJNumeric()`, `MatCreateMPIAIJSumSeqAIJ()`
4823: @*/
4824: PetscErrorCode MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm, Mat seqmat, PetscInt m, PetscInt n, Mat *mpimat)
4825: {
4826: Mat B_mpi;
4827: Mat_SeqAIJ *a = (Mat_SeqAIJ *)seqmat->data;
4828: PetscMPIInt size, rank, tagi, tagj, *len_s, *len_si, *len_ri;
4829: PetscInt **buf_rj, **buf_ri, **buf_ri_k;
4830: PetscInt M = seqmat->rmap->n, N = seqmat->cmap->n, i, *owners, *ai = a->i, *aj = a->j;
4831: PetscInt len, *dnz, *onz, bs, cbs;
4832: PetscInt k, anzi, *bi, *bj, *lnk, nlnk, arow, bnzi;
4833: PetscInt nrows, *buf_s, *buf_si, *buf_si_i, **nextrow, **nextai;
4834: MPI_Request *si_waits, *sj_waits, *ri_waits, *rj_waits;
4835: MPI_Status *status;
4836: PetscFreeSpaceList free_space = NULL, current_space = NULL;
4837: PetscBT lnkbt;
4838: MatMergeSeqsToMPI *merge;
4839: PetscContainer container;
4841: PetscFunctionBegin;
4842: PetscCall(PetscLogEventBegin(MAT_Seqstompisym, seqmat, 0, 0, 0));
4844: /* make sure it is a PETSc comm */
4845: PetscCall(PetscCommDuplicate(comm, &comm, NULL));
4846: PetscCallMPI(MPI_Comm_size(comm, &size));
4847: PetscCallMPI(MPI_Comm_rank(comm, &rank));
4849: PetscCall(PetscNew(&merge));
4850: PetscCall(PetscMalloc1(size, &status));
4852: /* determine row ownership */
4853: PetscCall(PetscLayoutCreate(comm, &merge->rowmap));
4854: PetscCall(PetscLayoutSetLocalSize(merge->rowmap, m));
4855: PetscCall(PetscLayoutSetSize(merge->rowmap, M));
4856: PetscCall(PetscLayoutSetBlockSize(merge->rowmap, 1));
4857: PetscCall(PetscLayoutSetUp(merge->rowmap));
4858: PetscCall(PetscMalloc1(size, &len_si));
4859: PetscCall(PetscMalloc1(size, &merge->len_s));
4861: m = merge->rowmap->n;
4862: owners = merge->rowmap->range;
4864: /* determine the number of messages to send, their lengths */
4865: len_s = merge->len_s;
4867: len = 0; /* length of buf_si[] */
4868: merge->nsend = 0;
4869: for (PetscMPIInt proc = 0; proc < size; proc++) {
4870: len_si[proc] = 0;
4871: if (proc == rank) {
4872: len_s[proc] = 0;
4873: } else {
4874: PetscCall(PetscMPIIntCast(owners[proc + 1] - owners[proc] + 1, &len_si[proc]));
4875: PetscCall(PetscMPIIntCast(ai[owners[proc + 1]] - ai[owners[proc]], &len_s[proc])); /* num of rows to be sent to [proc] */
4876: }
4877: if (len_s[proc]) {
4878: merge->nsend++;
4879: nrows = 0;
4880: for (i = owners[proc]; i < owners[proc + 1]; i++) {
4881: if (ai[i + 1] > ai[i]) nrows++;
4882: }
4883: PetscCall(PetscMPIIntCast(2 * (nrows + 1), &len_si[proc]));
4884: len += len_si[proc];
4885: }
4886: }
4888: /* determine the number and length of messages to receive for ij-structure */
4889: PetscCall(PetscGatherNumberOfMessages(comm, NULL, len_s, &merge->nrecv));
4890: PetscCall(PetscGatherMessageLengths2(comm, merge->nsend, merge->nrecv, len_s, len_si, &merge->id_r, &merge->len_r, &len_ri));
4892: /* post the Irecv of j-structure */
4893: PetscCall(PetscCommGetNewTag(comm, &tagj));
4894: PetscCall(PetscPostIrecvInt(comm, tagj, merge->nrecv, merge->id_r, merge->len_r, &buf_rj, &rj_waits));
4896: /* post the Isend of j-structure */
4897: PetscCall(PetscMalloc2(merge->nsend, &si_waits, merge->nsend, &sj_waits));
4899: for (PetscMPIInt proc = 0, k = 0; proc < size; proc++) {
4900: if (!len_s[proc]) continue;
4901: i = owners[proc];
4902: PetscCallMPI(MPIU_Isend(aj + ai[i], len_s[proc], MPIU_INT, proc, tagj, comm, sj_waits + k));
4903: k++;
4904: }
4906: /* receives and sends of j-structure are complete */
4907: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, rj_waits, status));
4908: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, sj_waits, status));
4910: /* send and recv i-structure */
4911: PetscCall(PetscCommGetNewTag(comm, &tagi));
4912: PetscCall(PetscPostIrecvInt(comm, tagi, merge->nrecv, merge->id_r, len_ri, &buf_ri, &ri_waits));
4914: PetscCall(PetscMalloc1(len + 1, &buf_s));
4915: buf_si = buf_s; /* points to the beginning of k-th msg to be sent */
4916: for (PetscMPIInt proc = 0, k = 0; proc < size; proc++) {
4917: if (!len_s[proc]) continue;
4918: /* form outgoing message for i-structure:
4919: buf_si[0]: nrows to be sent
4920: [1:nrows]: row index (global)
4921: [nrows+1:2*nrows+1]: i-structure index
4922: */
4923: nrows = len_si[proc] / 2 - 1;
4924: buf_si_i = buf_si + nrows + 1;
4925: buf_si[0] = nrows;
4926: buf_si_i[0] = 0;
4927: nrows = 0;
4928: for (i = owners[proc]; i < owners[proc + 1]; i++) {
4929: anzi = ai[i + 1] - ai[i];
4930: if (anzi) {
4931: buf_si_i[nrows + 1] = buf_si_i[nrows] + anzi; /* i-structure */
4932: buf_si[nrows + 1] = i - owners[proc]; /* local row index */
4933: nrows++;
4934: }
4935: }
4936: PetscCallMPI(MPIU_Isend(buf_si, len_si[proc], MPIU_INT, proc, tagi, comm, si_waits + k));
4937: k++;
4938: buf_si += len_si[proc];
4939: }
4941: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, ri_waits, status));
4942: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, si_waits, status));
4944: PetscCall(PetscInfo(seqmat, "nsend: %d, nrecv: %d\n", merge->nsend, merge->nrecv));
4945: 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]));
4947: PetscCall(PetscFree(len_si));
4948: PetscCall(PetscFree(len_ri));
4949: PetscCall(PetscFree(rj_waits));
4950: PetscCall(PetscFree2(si_waits, sj_waits));
4951: PetscCall(PetscFree(ri_waits));
4952: PetscCall(PetscFree(buf_s));
4953: PetscCall(PetscFree(status));
4955: /* compute a local seq matrix in each processor */
4956: /* allocate bi array and free space for accumulating nonzero column info */
4957: PetscCall(PetscMalloc1(m + 1, &bi));
4958: bi[0] = 0;
4960: /* create and initialize a linked list */
4961: nlnk = N + 1;
4962: PetscCall(PetscLLCreate(N, N, nlnk, lnk, lnkbt));
4964: /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */
4965: len = ai[owners[rank + 1]] - ai[owners[rank]];
4966: PetscCall(PetscFreeSpaceGet(PetscIntMultTruncate(2, len) + 1, &free_space));
4968: current_space = free_space;
4970: /* determine symbolic info for each local row */
4971: PetscCall(PetscMalloc3(merge->nrecv, &buf_ri_k, merge->nrecv, &nextrow, merge->nrecv, &nextai));
4973: for (k = 0; k < merge->nrecv; k++) {
4974: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4975: nrows = *buf_ri_k[k];
4976: nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */
4977: nextai[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure */
4978: }
4980: MatPreallocateBegin(comm, m, n, dnz, onz);
4981: len = 0;
4982: for (i = 0; i < m; i++) {
4983: bnzi = 0;
4984: /* add local non-zero cols of this proc's seqmat into lnk */
4985: arow = owners[rank] + i;
4986: anzi = ai[arow + 1] - ai[arow];
4987: aj = a->j + ai[arow];
4988: PetscCall(PetscLLAddSorted(anzi, aj, N, &nlnk, lnk, lnkbt));
4989: bnzi += nlnk;
4990: /* add received col data into lnk */
4991: for (k = 0; k < merge->nrecv; k++) { /* k-th received message */
4992: if (i == *nextrow[k]) { /* i-th row */
4993: anzi = *(nextai[k] + 1) - *nextai[k];
4994: aj = buf_rj[k] + *nextai[k];
4995: PetscCall(PetscLLAddSorted(anzi, aj, N, &nlnk, lnk, lnkbt));
4996: bnzi += nlnk;
4997: nextrow[k]++;
4998: nextai[k]++;
4999: }
5000: }
5001: if (len < bnzi) len = bnzi; /* =max(bnzi) */
5003: /* if free space is not available, make more free space */
5004: if (current_space->local_remaining < bnzi) PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(bnzi, current_space->total_array_size), ¤t_space));
5005: /* copy data into free space, then initialize lnk */
5006: PetscCall(PetscLLClean(N, N, bnzi, lnk, current_space->array, lnkbt));
5007: PetscCall(MatPreallocateSet(i + owners[rank], bnzi, current_space->array, dnz, onz));
5009: current_space->array += bnzi;
5010: current_space->local_used += bnzi;
5011: current_space->local_remaining -= bnzi;
5013: bi[i + 1] = bi[i] + bnzi;
5014: }
5016: PetscCall(PetscFree3(buf_ri_k, nextrow, nextai));
5018: PetscCall(PetscMalloc1(bi[m], &bj));
5019: PetscCall(PetscFreeSpaceContiguous(&free_space, bj));
5020: PetscCall(PetscLLDestroy(lnk, lnkbt));
5022: /* create symbolic parallel matrix B_mpi */
5023: PetscCall(MatGetBlockSizes(seqmat, &bs, &cbs));
5024: PetscCall(MatCreate(comm, &B_mpi));
5025: if (n == PETSC_DECIDE) PetscCall(MatSetSizes(B_mpi, m, n, PETSC_DETERMINE, N));
5026: else PetscCall(MatSetSizes(B_mpi, m, n, PETSC_DETERMINE, PETSC_DETERMINE));
5027: PetscCall(MatSetBlockSizes(B_mpi, bs, cbs));
5028: PetscCall(MatSetType(B_mpi, MATMPIAIJ));
5029: PetscCall(MatMPIAIJSetPreallocation(B_mpi, 0, dnz, 0, onz));
5030: MatPreallocateEnd(dnz, onz);
5031: PetscCall(MatSetOption(B_mpi, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_FALSE));
5033: /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */
5034: B_mpi->assembled = PETSC_FALSE;
5035: merge->bi = bi;
5036: merge->bj = bj;
5037: merge->buf_ri = buf_ri;
5038: merge->buf_rj = buf_rj;
5039: merge->coi = NULL;
5040: merge->coj = NULL;
5041: merge->owners_co = NULL;
5043: PetscCall(PetscCommDestroy(&comm));
5045: /* attach the supporting struct to B_mpi for reuse */
5046: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
5047: PetscCall(PetscContainerSetPointer(container, merge));
5048: PetscCall(PetscContainerSetCtxDestroy(container, MatMergeSeqsToMPIDestroy));
5049: PetscCall(PetscObjectCompose((PetscObject)B_mpi, "MatMergeSeqsToMPI", (PetscObject)container));
5050: PetscCall(PetscContainerDestroy(&container));
5051: *mpimat = B_mpi;
5053: PetscCall(PetscLogEventEnd(MAT_Seqstompisym, seqmat, 0, 0, 0));
5054: PetscFunctionReturn(PETSC_SUCCESS);
5055: }
5057: /*@
5058: MatCreateMPIAIJSumSeqAIJ - Creates a `MATMPIAIJ` matrix by adding sequential
5059: matrices from each processor
5061: Collective
5063: Input Parameters:
5064: + comm - the communicators the parallel matrix will live on
5065: . seqmat - the input sequential matrices
5066: . m - number of local rows (or `PETSC_DECIDE`)
5067: . n - number of local columns (or `PETSC_DECIDE`)
5068: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5070: Output Parameter:
5071: . mpimat - the parallel matrix generated
5073: Level: advanced
5075: Note:
5076: The dimensions of the sequential matrix in each processor MUST be the same.
5077: The input seqmat is included into the container `MatMergeSeqsToMPIDestroy`, and will be
5078: destroyed when `mpimat` is destroyed. Call `PetscObjectQuery()` to access `seqmat`.
5080: .seealso: [](ch_matrices), `Mat`, `MatCreateAIJ()`
5081: @*/
5082: PetscErrorCode MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm, Mat seqmat, PetscInt m, PetscInt n, MatReuse scall, Mat *mpimat)
5083: {
5084: PetscMPIInt size;
5086: PetscFunctionBegin;
5087: PetscCallMPI(MPI_Comm_size(comm, &size));
5088: if (size == 1) {
5089: PetscCall(PetscLogEventBegin(MAT_Seqstompi, seqmat, 0, 0, 0));
5090: if (scall == MAT_INITIAL_MATRIX) PetscCall(MatDuplicate(seqmat, MAT_COPY_VALUES, mpimat));
5091: else PetscCall(MatCopy(seqmat, *mpimat, SAME_NONZERO_PATTERN));
5092: PetscCall(PetscLogEventEnd(MAT_Seqstompi, seqmat, 0, 0, 0));
5093: PetscFunctionReturn(PETSC_SUCCESS);
5094: }
5095: PetscCall(PetscLogEventBegin(MAT_Seqstompi, seqmat, 0, 0, 0));
5096: if (scall == MAT_INITIAL_MATRIX) PetscCall(MatCreateMPIAIJSumSeqAIJSymbolic(comm, seqmat, m, n, mpimat));
5097: PetscCall(MatCreateMPIAIJSumSeqAIJNumeric(seqmat, *mpimat));
5098: PetscCall(PetscLogEventEnd(MAT_Seqstompi, seqmat, 0, 0, 0));
5099: PetscFunctionReturn(PETSC_SUCCESS);
5100: }
5102: /*@
5103: MatAIJGetLocalMat - Creates a `MATSEQAIJ` from a `MATAIJ` matrix.
5105: Not Collective
5107: Input Parameter:
5108: . A - the matrix
5110: Output Parameter:
5111: . A_loc - the local sequential matrix generated
5113: Level: developer
5115: Notes:
5116: The matrix is created by taking `A`'s local rows and putting them into a sequential matrix
5117: with `mlocal` rows and `n` columns. Where `mlocal` is obtained with `MatGetLocalSize()` and
5118: `n` is the global column count obtained with `MatGetSize()`
5120: In other words combines the two parts of a parallel `MATMPIAIJ` matrix on each process to a single matrix.
5122: For parallel matrices this creates an entirely new matrix. If the matrix is sequential it merely increases the reference count.
5124: Destroy the matrix with `MatDestroy()`
5126: .seealso: [](ch_matrices), `Mat`, `MatMPIAIJGetLocalMat()`
5127: @*/
5128: PetscErrorCode MatAIJGetLocalMat(Mat A, Mat *A_loc)
5129: {
5130: PetscBool mpi;
5132: PetscFunctionBegin;
5133: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIAIJ, &mpi));
5134: if (mpi) PetscCall(MatMPIAIJGetLocalMat(A, MAT_INITIAL_MATRIX, A_loc));
5135: else {
5136: *A_loc = A;
5137: PetscCall(PetscObjectReference((PetscObject)*A_loc));
5138: }
5139: PetscFunctionReturn(PETSC_SUCCESS);
5140: }
5142: /*@
5143: MatMPIAIJGetLocalMat - Creates a `MATSEQAIJ` from a `MATMPIAIJ` matrix.
5145: Not Collective
5147: Input Parameters:
5148: + A - the matrix
5149: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5151: Output Parameter:
5152: . A_loc - the local sequential matrix generated
5154: Level: developer
5156: Notes:
5157: The matrix is created by taking all `A`'s local rows and putting them into a sequential
5158: matrix with `mlocal` rows and `n` columns.`mlocal` is the row count obtained with
5159: `MatGetLocalSize()` and `n` is the global column count obtained with `MatGetSize()`.
5161: In other words combines the two parts of a parallel `MATMPIAIJ` matrix on each process to a single matrix.
5163: When `A` is sequential and `MAT_INITIAL_MATRIX` is requested, the matrix returned is the diagonal part of `A` (which contains the entire matrix),
5164: 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
5165: then `MatCopy`(Adiag,*`A_loc`,`SAME_NONZERO_PATTERN`) is called to fill `A_loc`. Thus one can preallocate the appropriate sequential matrix `A_loc`
5166: 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.
5168: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMatCondensed()`, `MatMPIAIJGetLocalMatMerge()`
5169: @*/
5170: PetscErrorCode MatMPIAIJGetLocalMat(Mat A, MatReuse scall, Mat *A_loc)
5171: {
5172: PetscFunctionBegin;
5173: PetscCall(MatMPIAIJGetLocalMat_Private(A, scall, PETSC_FALSE, A_loc));
5174: PetscFunctionReturn(PETSC_SUCCESS);
5175: }
5177: PetscErrorCode MatMPIAIJGetLocalMat_Private(Mat A, MatReuse scall, PetscBool structure_only, Mat *A_loc)
5178: {
5179: Mat_MPIAIJ *mpimat = (Mat_MPIAIJ *)A->data;
5180: Mat_SeqAIJ *mat, *a, *b;
5181: PetscInt *ai, *aj, *bi, *bj, *cmap = mpimat->garray;
5182: const PetscScalar *aa, *ba, *aav = NULL, *bav = NULL;
5183: PetscScalar *ca = NULL, *cam;
5184: PetscMPIInt size;
5185: PetscInt am = A->rmap->n, i, j, k, cstart = A->cmap->rstart;
5186: PetscInt *ci, *cj, col, ncols_d, ncols_o, jo;
5187: PetscBool match;
5189: PetscFunctionBegin;
5190: PetscCheck(!structure_only || scall == MAT_INITIAL_MATRIX, PETSC_COMM_SELF, PETSC_ERR_SUP, "Structure-only extraction requires MAT_INITIAL_MATRIX");
5191: PetscCall(PetscStrbeginswith(((PetscObject)A)->type_name, MATMPIAIJ, &match));
5192: PetscCheck(match, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Requires MATMPIAIJ matrix as input");
5193: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
5194: if (size == 1 && !structure_only) {
5195: if (scall == MAT_INITIAL_MATRIX) {
5196: PetscCall(PetscObjectReference((PetscObject)mpimat->A));
5197: *A_loc = mpimat->A;
5198: } else if (scall == MAT_REUSE_MATRIX) PetscCall(MatCopy(mpimat->A, *A_loc, SAME_NONZERO_PATTERN));
5199: PetscFunctionReturn(PETSC_SUCCESS);
5200: }
5202: PetscCall(PetscLogEventBegin(MAT_Getlocalmat, A, 0, 0, 0));
5203: a = (Mat_SeqAIJ *)mpimat->A->data;
5204: b = (Mat_SeqAIJ *)mpimat->B->data;
5205: ai = a->i;
5206: aj = a->j;
5207: bi = b->i;
5208: bj = b->j;
5209: if (!structure_only) {
5210: PetscCall(MatSeqAIJGetArrayRead(mpimat->A, &aav));
5211: PetscCall(MatSeqAIJGetArrayRead(mpimat->B, &bav));
5212: }
5213: aa = aav;
5214: ba = bav;
5215: if (scall == MAT_INITIAL_MATRIX) {
5216: PetscCall(PetscMalloc1(1 + am, &ci));
5217: ci[0] = 0;
5218: for (i = 0; i < am; i++) ci[i + 1] = ci[i] + (ai[i + 1] - ai[i]) + (bi[i + 1] - bi[i]);
5219: PetscCall(PetscMalloc1(ci[am], &cj));
5220: if (!structure_only) PetscCall(PetscMalloc1(ci[am], &ca));
5221: k = 0;
5222: for (i = 0; i < am; i++) {
5223: ncols_o = bi[i + 1] - bi[i];
5224: ncols_d = ai[i + 1] - ai[i];
5225: /* off-diagonal portion of A */
5226: for (jo = 0; jo < ncols_o; jo++, bj++, k++) {
5227: col = cmap[*bj];
5228: if (col >= cstart) break;
5229: cj[k] = col;
5230: if (!structure_only) ca[k] = *ba++;
5231: }
5232: /* diagonal portion of A */
5233: for (j = 0; j < ncols_d; j++, k++) {
5234: cj[k] = cstart + *aj++;
5235: if (!structure_only) ca[k] = *aa++;
5236: }
5237: /* off-diagonal portion of A */
5238: for (j = jo; j < ncols_o; j++, k++) {
5239: cj[k] = cmap[*bj++];
5240: if (!structure_only) ca[k] = *ba++;
5241: }
5242: }
5243: /* put together the new matrix */
5244: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, am, A->cmap->N, ci, cj, ca, A_loc));
5245: PetscCall(MatSetOption(*A_loc, MAT_STRUCTURE_ONLY, structure_only));
5246: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5247: /* Since these are PETSc arrays, change flags to free them as necessary. */
5248: mat = (Mat_SeqAIJ *)(*A_loc)->data;
5249: mat->free_a = PETSC_TRUE;
5250: mat->free_ij = PETSC_TRUE;
5251: mat->nonew = 0;
5252: } else if (scall == MAT_REUSE_MATRIX) {
5253: mat = (Mat_SeqAIJ *)(*A_loc)->data;
5254: ci = mat->i;
5255: cj = mat->j;
5256: PetscCall(MatSeqAIJGetArrayWrite(*A_loc, &cam));
5257: for (i = 0; i < am; i++) {
5258: /* off-diagonal portion of A */
5259: ncols_o = bi[i + 1] - bi[i];
5260: for (jo = 0; jo < ncols_o; jo++, bj++) {
5261: col = cmap[*bj];
5262: if (col >= cstart) break;
5263: *cam++ = *ba++;
5264: }
5265: /* diagonal portion of A */
5266: ncols_d = ai[i + 1] - ai[i];
5267: for (j = 0; j < ncols_d; j++) *cam++ = *aa++;
5268: /* off-diagonal portion of A */
5269: for (j = jo; j < ncols_o; j++, bj++) *cam++ = *ba++;
5270: }
5271: PetscCall(MatSeqAIJRestoreArrayWrite(*A_loc, &cam));
5272: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Invalid MatReuse %d", (int)scall);
5273: if (!structure_only) {
5274: PetscCall(MatSeqAIJRestoreArrayRead(mpimat->A, &aav));
5275: PetscCall(MatSeqAIJRestoreArrayRead(mpimat->B, &bav));
5276: }
5277: PetscCall(PetscLogEventEnd(MAT_Getlocalmat, A, 0, 0, 0));
5278: PetscFunctionReturn(PETSC_SUCCESS);
5279: }
5281: /*@
5282: MatMPIAIJGetLocalMatMerge - Creates a `MATSEQAIJ` from a `MATMPIAIJ` matrix by taking all its local rows and putting them into a sequential matrix with
5283: mlocal rows and n columns. Where n is the sum of the number of columns of the diagonal and off-diagonal part
5285: Not Collective
5287: Input Parameters:
5288: + A - the matrix
5289: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5291: Output Parameters:
5292: + glob - sequential `IS` with global indices associated with the columns of the local sequential matrix generated (can be `NULL`)
5293: - A_loc - the local sequential matrix generated
5295: Level: developer
5297: Note:
5298: This is different from `MatMPIAIJGetLocalMat()` since the first columns in the returning matrix are those associated with the diagonal
5299: part, then those associated with the off-diagonal part (in its local ordering)
5301: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMat()`, `MatMPIAIJGetLocalMatCondensed()`
5302: @*/
5303: PetscErrorCode MatMPIAIJGetLocalMatMerge(Mat A, MatReuse scall, IS *glob, Mat *A_loc)
5304: {
5305: Mat Ao, Ad;
5306: const PetscInt *cmap;
5307: PetscMPIInt size;
5308: PetscErrorCode (*f)(Mat, MatReuse, IS *, Mat *);
5310: PetscFunctionBegin;
5311: PetscCall(MatMPIAIJGetSeqAIJ(A, &Ad, &Ao, &cmap));
5312: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
5313: if (size == 1) {
5314: if (scall == MAT_INITIAL_MATRIX) {
5315: PetscCall(PetscObjectReference((PetscObject)Ad));
5316: *A_loc = Ad;
5317: } else if (scall == MAT_REUSE_MATRIX) {
5318: PetscCall(MatCopy(Ad, *A_loc, SAME_NONZERO_PATTERN));
5319: }
5320: if (glob) PetscCall(ISCreateStride(PetscObjectComm((PetscObject)Ad), Ad->cmap->n, Ad->cmap->rstart, 1, glob));
5321: PetscFunctionReturn(PETSC_SUCCESS);
5322: }
5323: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatMPIAIJGetLocalMatMerge_C", &f));
5324: PetscCall(PetscLogEventBegin(MAT_Getlocalmat, A, 0, 0, 0));
5325: if (f) PetscCall((*f)(A, scall, glob, A_loc));
5326: else {
5327: Mat_SeqAIJ *a = (Mat_SeqAIJ *)Ad->data;
5328: Mat_SeqAIJ *b = (Mat_SeqAIJ *)Ao->data;
5329: Mat_SeqAIJ *c;
5330: PetscInt *ai = a->i, *aj = a->j;
5331: PetscInt *bi = b->i, *bj = b->j;
5332: PetscInt *ci, *cj;
5333: const PetscScalar *aa, *ba;
5334: PetscScalar *ca;
5335: PetscInt i, j, am, dn, on;
5337: PetscCall(MatGetLocalSize(Ad, &am, &dn));
5338: PetscCall(MatGetLocalSize(Ao, NULL, &on));
5339: PetscCall(MatSeqAIJGetArrayRead(Ad, &aa));
5340: PetscCall(MatSeqAIJGetArrayRead(Ao, &ba));
5341: if (scall == MAT_INITIAL_MATRIX) {
5342: PetscInt k;
5343: PetscCall(PetscMalloc1(1 + am, &ci));
5344: PetscCall(PetscMalloc1(ai[am] + bi[am], &cj));
5345: PetscCall(PetscMalloc1(ai[am] + bi[am], &ca));
5346: ci[0] = 0;
5347: for (i = 0, k = 0; i < am; i++) {
5348: const PetscInt ncols_o = bi[i + 1] - bi[i];
5349: const PetscInt ncols_d = ai[i + 1] - ai[i];
5350: ci[i + 1] = ci[i] + ncols_o + ncols_d;
5351: /* diagonal portion of A */
5352: for (j = 0; j < ncols_d; j++, k++) {
5353: cj[k] = *aj++;
5354: ca[k] = *aa++;
5355: }
5356: /* off-diagonal portion of A */
5357: for (j = 0; j < ncols_o; j++, k++) {
5358: cj[k] = dn + *bj++;
5359: ca[k] = *ba++;
5360: }
5361: }
5362: /* put together the new matrix */
5363: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, am, dn + on, ci, cj, ca, A_loc));
5364: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5365: /* Since these are PETSc arrays, change flags to free them as necessary. */
5366: c = (Mat_SeqAIJ *)(*A_loc)->data;
5367: c->free_a = PETSC_TRUE;
5368: c->free_ij = PETSC_TRUE;
5369: c->nonew = 0;
5370: PetscCall(MatSetType(*A_loc, ((PetscObject)Ad)->type_name));
5371: } else if (scall == MAT_REUSE_MATRIX) {
5372: PetscCall(MatSeqAIJGetArrayWrite(*A_loc, &ca));
5373: for (i = 0; i < am; i++) {
5374: const PetscInt ncols_d = ai[i + 1] - ai[i];
5375: const PetscInt ncols_o = bi[i + 1] - bi[i];
5376: /* diagonal portion of A */
5377: for (j = 0; j < ncols_d; j++) *ca++ = *aa++;
5378: /* off-diagonal portion of A */
5379: for (j = 0; j < ncols_o; j++) *ca++ = *ba++;
5380: }
5381: PetscCall(MatSeqAIJRestoreArrayWrite(*A_loc, &ca));
5382: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Invalid MatReuse %d", (int)scall);
5383: PetscCall(MatSeqAIJRestoreArrayRead(Ad, &aa));
5384: PetscCall(MatSeqAIJRestoreArrayRead(Ao, &aa));
5385: if (glob) {
5386: PetscInt cst, *gidx;
5388: PetscCall(MatGetOwnershipRangeColumn(A, &cst, NULL));
5389: PetscCall(PetscMalloc1(dn + on, &gidx));
5390: for (i = 0; i < dn; i++) gidx[i] = cst + i;
5391: for (i = 0; i < on; i++) gidx[i + dn] = cmap[i];
5392: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)Ad), dn + on, gidx, PETSC_OWN_POINTER, glob));
5393: }
5394: }
5395: PetscCall(PetscLogEventEnd(MAT_Getlocalmat, A, 0, 0, 0));
5396: PetscFunctionReturn(PETSC_SUCCESS);
5397: }
5399: /*@
5400: MatMPIAIJGetLocalMatCondensed - Creates a `MATSEQAIJ` matrix from an `MATMPIAIJ` matrix by taking all its local rows and NON-ZERO columns
5402: Not Collective
5404: Input Parameters:
5405: + A - the matrix
5406: . scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5407: . row - index set of rows to extract (or `NULL`)
5408: - col - index set of columns to extract (or `NULL`)
5410: Output Parameter:
5411: . A_loc - the local sequential matrix generated
5413: Level: developer
5415: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMat()`
5416: @*/
5417: PetscErrorCode MatMPIAIJGetLocalMatCondensed(Mat A, MatReuse scall, IS *row, IS *col, Mat *A_loc)
5418: {
5419: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5420: PetscInt i, start, end, ncols, nzA, nzB, *cmap, imark, *idx;
5421: IS isrowa, iscola;
5422: Mat *aloc;
5423: PetscBool match;
5425: PetscFunctionBegin;
5426: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIAIJ, &match));
5427: PetscCheck(match, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Requires MATMPIAIJ matrix as input");
5428: PetscCall(PetscLogEventBegin(MAT_Getlocalmatcondensed, A, 0, 0, 0));
5429: if (!row) {
5430: start = A->rmap->rstart;
5431: end = A->rmap->rend;
5432: PetscCall(ISCreateStride(PETSC_COMM_SELF, end - start, start, 1, &isrowa));
5433: } else {
5434: isrowa = *row;
5435: }
5436: if (!col) {
5437: start = A->cmap->rstart;
5438: cmap = a->garray;
5439: nzA = a->A->cmap->n;
5440: nzB = a->B->cmap->n;
5441: PetscCall(PetscMalloc1(nzA + nzB, &idx));
5442: ncols = 0;
5443: for (i = 0; i < nzB; i++) {
5444: if (cmap[i] < start) idx[ncols++] = cmap[i];
5445: else break;
5446: }
5447: imark = i;
5448: for (i = 0; i < nzA; i++) idx[ncols++] = start + i;
5449: for (i = imark; i < nzB; i++) idx[ncols++] = cmap[i];
5450: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, &iscola));
5451: } else iscola = *col;
5452: if (scall != MAT_INITIAL_MATRIX) {
5453: PetscCall(PetscMalloc1(1, &aloc));
5454: aloc[0] = *A_loc;
5455: }
5456: PetscCall(MatCreateSubMatrices(A, 1, &isrowa, &iscola, scall, &aloc));
5457: if (!col) { /* attach global id of condensed columns */
5458: PetscCall(PetscObjectCompose((PetscObject)aloc[0], "_petsc_GetLocalMatCondensed_iscol", (PetscObject)iscola));
5459: }
5460: *A_loc = aloc[0];
5461: PetscCall(PetscFree(aloc));
5462: if (!row) PetscCall(ISDestroy(&isrowa));
5463: if (!col) PetscCall(ISDestroy(&iscola));
5464: PetscCall(PetscLogEventEnd(MAT_Getlocalmatcondensed, A, 0, 0, 0));
5465: PetscFunctionReturn(PETSC_SUCCESS);
5466: }
5468: /*
5469: * Create a sequential AIJ matrix based on row indices. a whole column is extracted once a row is matched.
5470: * Row could be local or remote.The routine is designed to be scalable in memory so that nothing is based
5471: * on a global size.
5472: * */
5473: static PetscErrorCode MatCreateSeqSubMatrixWithRows_Private(Mat P, IS rows, Mat *P_oth)
5474: {
5475: Mat_MPIAIJ *p = (Mat_MPIAIJ *)P->data;
5476: Mat_SeqAIJ *pd = (Mat_SeqAIJ *)p->A->data, *po = (Mat_SeqAIJ *)p->B->data, *p_oth;
5477: PetscInt plocalsize, nrows, *ilocal, *oilocal, i, lidx, *nrcols, *nlcols, ncol;
5478: PetscMPIInt owner;
5479: PetscSFNode *iremote, *oiremote;
5480: const PetscInt *lrowindices;
5481: PetscSF sf, osf;
5482: PetscInt pcstart, *roffsets, *loffsets, *pnnz, j;
5483: PetscInt ontotalcols, dntotalcols, ntotalcols, nout;
5484: MPI_Comm comm;
5485: ISLocalToGlobalMapping mapping;
5486: const PetscScalar *pd_a, *po_a;
5488: PetscFunctionBegin;
5489: PetscCall(PetscObjectGetComm((PetscObject)P, &comm));
5490: /* plocalsize is the number of roots
5491: * nrows is the number of leaves
5492: * */
5493: PetscCall(MatGetLocalSize(P, &plocalsize, NULL));
5494: PetscCall(ISGetLocalSize(rows, &nrows));
5495: PetscCall(PetscCalloc1(nrows, &iremote));
5496: PetscCall(ISGetIndices(rows, &lrowindices));
5497: for (i = 0; i < nrows; i++) {
5498: /* Find a remote index and an owner for a row
5499: * The row could be local or remote
5500: * */
5501: owner = 0;
5502: lidx = 0;
5503: PetscCall(PetscLayoutFindOwnerIndex(P->rmap, lrowindices[i], &owner, &lidx));
5504: iremote[i].index = lidx;
5505: iremote[i].rank = owner;
5506: }
5507: /* Create SF to communicate how many nonzero columns for each row */
5508: PetscCall(PetscSFCreate(comm, &sf));
5509: /* SF will figure out the number of nonzero columns for each row, and their
5510: * offsets
5511: * */
5512: PetscCall(PetscSFSetGraph(sf, plocalsize, nrows, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
5513: PetscCall(PetscSFSetFromOptions(sf));
5514: PetscCall(PetscSFSetUp(sf));
5516: PetscCall(PetscCalloc1(2 * (plocalsize + 1), &roffsets));
5517: PetscCall(PetscCalloc1(2 * plocalsize, &nrcols));
5518: PetscCall(PetscCalloc1(nrows, &pnnz));
5519: roffsets[0] = 0;
5520: roffsets[1] = 0;
5521: for (i = 0; i < plocalsize; i++) {
5522: /* diagonal */
5523: nrcols[i * 2 + 0] = pd->i[i + 1] - pd->i[i];
5524: /* off-diagonal */
5525: nrcols[i * 2 + 1] = po->i[i + 1] - po->i[i];
5526: /* compute offsets so that we relative location for each row */
5527: roffsets[(i + 1) * 2 + 0] = roffsets[i * 2 + 0] + nrcols[i * 2 + 0];
5528: roffsets[(i + 1) * 2 + 1] = roffsets[i * 2 + 1] + nrcols[i * 2 + 1];
5529: }
5530: PetscCall(PetscCalloc1(2 * nrows, &nlcols));
5531: PetscCall(PetscCalloc1(2 * nrows, &loffsets));
5532: /* 'r' means root, and 'l' means leaf */
5533: PetscCall(PetscSFBcastBegin(sf, MPIU_2INT, nrcols, nlcols, MPI_REPLACE));
5534: PetscCall(PetscSFBcastBegin(sf, MPIU_2INT, roffsets, loffsets, MPI_REPLACE));
5535: PetscCall(PetscSFBcastEnd(sf, MPIU_2INT, nrcols, nlcols, MPI_REPLACE));
5536: PetscCall(PetscSFBcastEnd(sf, MPIU_2INT, roffsets, loffsets, MPI_REPLACE));
5537: PetscCall(PetscSFDestroy(&sf));
5538: PetscCall(PetscFree(roffsets));
5539: PetscCall(PetscFree(nrcols));
5540: dntotalcols = 0;
5541: ontotalcols = 0;
5542: ncol = 0;
5543: for (i = 0; i < nrows; i++) {
5544: pnnz[i] = nlcols[i * 2 + 0] + nlcols[i * 2 + 1];
5545: ncol = PetscMax(pnnz[i], ncol);
5546: /* diagonal */
5547: dntotalcols += nlcols[i * 2 + 0];
5548: /* off-diagonal */
5549: ontotalcols += nlcols[i * 2 + 1];
5550: }
5551: /* We do not need to figure the right number of columns
5552: * since all the calculations will be done by going through the raw data
5553: * */
5554: PetscCall(MatCreateSeqAIJ(PETSC_COMM_SELF, nrows, ncol, 0, pnnz, P_oth));
5555: PetscCall(MatSetUp(*P_oth));
5556: PetscCall(PetscFree(pnnz));
5557: p_oth = (Mat_SeqAIJ *)(*P_oth)->data;
5558: /* diagonal */
5559: PetscCall(PetscCalloc1(dntotalcols, &iremote));
5560: /* off-diagonal */
5561: PetscCall(PetscCalloc1(ontotalcols, &oiremote));
5562: /* diagonal */
5563: PetscCall(PetscCalloc1(dntotalcols, &ilocal));
5564: /* off-diagonal */
5565: PetscCall(PetscCalloc1(ontotalcols, &oilocal));
5566: dntotalcols = 0;
5567: ontotalcols = 0;
5568: ntotalcols = 0;
5569: for (i = 0; i < nrows; i++) {
5570: owner = 0;
5571: PetscCall(PetscLayoutFindOwnerIndex(P->rmap, lrowindices[i], &owner, NULL));
5572: /* Set iremote for diag matrix */
5573: for (j = 0; j < nlcols[i * 2 + 0]; j++) {
5574: iremote[dntotalcols].index = loffsets[i * 2 + 0] + j;
5575: iremote[dntotalcols].rank = owner;
5576: /* P_oth is seqAIJ so that ilocal need to point to the first part of memory */
5577: ilocal[dntotalcols++] = ntotalcols++;
5578: }
5579: /* off-diagonal */
5580: for (j = 0; j < nlcols[i * 2 + 1]; j++) {
5581: oiremote[ontotalcols].index = loffsets[i * 2 + 1] + j;
5582: oiremote[ontotalcols].rank = owner;
5583: oilocal[ontotalcols++] = ntotalcols++;
5584: }
5585: }
5586: PetscCall(ISRestoreIndices(rows, &lrowindices));
5587: PetscCall(PetscFree(loffsets));
5588: PetscCall(PetscFree(nlcols));
5589: PetscCall(PetscSFCreate(comm, &sf));
5590: /* P serves as roots and P_oth is leaves
5591: * Diag matrix
5592: * */
5593: PetscCall(PetscSFSetGraph(sf, pd->i[plocalsize], dntotalcols, ilocal, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
5594: PetscCall(PetscSFSetFromOptions(sf));
5595: PetscCall(PetscSFSetUp(sf));
5597: PetscCall(PetscSFCreate(comm, &osf));
5598: /* off-diagonal */
5599: PetscCall(PetscSFSetGraph(osf, po->i[plocalsize], ontotalcols, oilocal, PETSC_OWN_POINTER, oiremote, PETSC_OWN_POINTER));
5600: PetscCall(PetscSFSetFromOptions(osf));
5601: PetscCall(PetscSFSetUp(osf));
5602: PetscCall(MatSeqAIJGetArrayRead(p->A, &pd_a));
5603: PetscCall(MatSeqAIJGetArrayRead(p->B, &po_a));
5604: /* operate on the matrix internal data to save memory */
5605: PetscCall(PetscSFBcastBegin(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5606: PetscCall(PetscSFBcastBegin(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5607: PetscCall(MatGetOwnershipRangeColumn(P, &pcstart, NULL));
5608: /* Convert to global indices for diag matrix */
5609: for (i = 0; i < pd->i[plocalsize]; i++) pd->j[i] += pcstart;
5610: PetscCall(PetscSFBcastBegin(sf, MPIU_INT, pd->j, p_oth->j, MPI_REPLACE));
5611: /* We want P_oth store global indices */
5612: PetscCall(ISLocalToGlobalMappingCreate(comm, 1, p->B->cmap->n, p->garray, PETSC_COPY_VALUES, &mapping));
5613: /* Use memory scalable approach */
5614: PetscCall(ISLocalToGlobalMappingSetType(mapping, ISLOCALTOGLOBALMAPPINGHASH));
5615: PetscCall(ISLocalToGlobalMappingApply(mapping, po->i[plocalsize], po->j, po->j));
5616: PetscCall(PetscSFBcastBegin(osf, MPIU_INT, po->j, p_oth->j, MPI_REPLACE));
5617: PetscCall(PetscSFBcastEnd(sf, MPIU_INT, pd->j, p_oth->j, MPI_REPLACE));
5618: /* Convert back to local indices */
5619: for (i = 0; i < pd->i[plocalsize]; i++) pd->j[i] -= pcstart;
5620: PetscCall(PetscSFBcastEnd(osf, MPIU_INT, po->j, p_oth->j, MPI_REPLACE));
5621: nout = 0;
5622: PetscCall(ISGlobalToLocalMappingApply(mapping, IS_GTOLM_DROP, po->i[plocalsize], po->j, &nout, po->j));
5623: PetscCheck(nout == po->i[plocalsize], comm, PETSC_ERR_ARG_INCOMP, "n %" PetscInt_FMT " does not equal to nout %" PetscInt_FMT " ", po->i[plocalsize], nout);
5624: PetscCall(ISLocalToGlobalMappingDestroy(&mapping));
5625: /* Exchange values */
5626: PetscCall(PetscSFBcastEnd(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5627: PetscCall(PetscSFBcastEnd(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5628: PetscCall(MatSeqAIJRestoreArrayRead(p->A, &pd_a));
5629: PetscCall(MatSeqAIJRestoreArrayRead(p->B, &po_a));
5630: /* Stop PETSc from shrinking memory */
5631: for (i = 0; i < nrows; i++) p_oth->ilen[i] = p_oth->imax[i];
5632: PetscCall(MatAssemblyBegin(*P_oth, MAT_FINAL_ASSEMBLY));
5633: PetscCall(MatAssemblyEnd(*P_oth, MAT_FINAL_ASSEMBLY));
5634: /* Attach PetscSF objects to P_oth so that we can reuse it later */
5635: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "diagsf", (PetscObject)sf));
5636: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "offdiagsf", (PetscObject)osf));
5637: PetscCall(PetscSFDestroy(&sf));
5638: PetscCall(PetscSFDestroy(&osf));
5639: PetscFunctionReturn(PETSC_SUCCESS);
5640: }
5642: /*
5643: * Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A
5644: * This supports MPIAIJ and MAIJ
5645: * */
5646: PetscErrorCode MatGetBrowsOfAcols_MPIXAIJ(Mat A, Mat P, PetscInt dof, MatReuse reuse, Mat *P_oth)
5647: {
5648: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data, *p = (Mat_MPIAIJ *)P->data;
5649: Mat_SeqAIJ *p_oth;
5650: IS rows, map;
5651: PetscHMapI hamp;
5652: PetscInt i, htsize, *rowindices, off, *mapping, key, count;
5653: MPI_Comm comm;
5654: PetscSF sf, osf;
5655: PetscBool has;
5657: PetscFunctionBegin;
5658: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
5659: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAocols, A, P, 0, 0));
5660: /* If it is the first time, create an index set of off-diag nonzero columns of A,
5661: * and then create a submatrix (that often is an overlapping matrix)
5662: * */
5663: if (reuse == MAT_INITIAL_MATRIX) {
5664: /* Use a hash table to figure out unique keys */
5665: PetscCall(PetscHMapICreateWithSize(a->B->cmap->n, &hamp));
5666: PetscCall(PetscCalloc1(a->B->cmap->n, &mapping));
5667: count = 0;
5668: /* Assume that a->g is sorted, otherwise the following does not make sense */
5669: for (i = 0; i < a->B->cmap->n; i++) {
5670: key = a->garray[i] / dof;
5671: PetscCall(PetscHMapIHas(hamp, key, &has));
5672: if (!has) {
5673: mapping[i] = count;
5674: PetscCall(PetscHMapISet(hamp, key, count++));
5675: } else {
5676: /* Current 'i' has the same value the previous step */
5677: mapping[i] = count - 1;
5678: }
5679: }
5680: PetscCall(ISCreateGeneral(comm, a->B->cmap->n, mapping, PETSC_OWN_POINTER, &map));
5681: PetscCall(PetscHMapIGetSize(hamp, &htsize));
5682: PetscCheck(htsize == count, comm, PETSC_ERR_ARG_INCOMP, " Size of hash map %" PetscInt_FMT " is inconsistent with count %" PetscInt_FMT, htsize, count);
5683: PetscCall(PetscCalloc1(htsize, &rowindices));
5684: off = 0;
5685: PetscCall(PetscHMapIGetKeys(hamp, &off, rowindices));
5686: PetscCall(PetscHMapIDestroy(&hamp));
5687: PetscCall(PetscSortInt(htsize, rowindices));
5688: PetscCall(ISCreateGeneral(comm, htsize, rowindices, PETSC_OWN_POINTER, &rows));
5689: /* In case, the matrix was already created but users want to recreate the matrix */
5690: PetscCall(MatDestroy(P_oth));
5691: PetscCall(MatCreateSeqSubMatrixWithRows_Private(P, rows, P_oth));
5692: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "aoffdiagtopothmapping", (PetscObject)map));
5693: PetscCall(ISDestroy(&map));
5694: PetscCall(ISDestroy(&rows));
5695: } else if (reuse == MAT_REUSE_MATRIX) {
5696: /* If matrix was already created, we simply update values using SF objects
5697: * that as attached to the matrix earlier.
5698: */
5699: const PetscScalar *pd_a, *po_a;
5701: PetscCall(PetscObjectQuery((PetscObject)*P_oth, "diagsf", (PetscObject *)&sf));
5702: PetscCall(PetscObjectQuery((PetscObject)*P_oth, "offdiagsf", (PetscObject *)&osf));
5703: PetscCheck(sf && osf, comm, PETSC_ERR_ARG_NULL, "Matrix is not initialized yet");
5704: p_oth = (Mat_SeqAIJ *)(*P_oth)->data;
5705: /* Update values in place */
5706: PetscCall(MatSeqAIJGetArrayRead(p->A, &pd_a));
5707: PetscCall(MatSeqAIJGetArrayRead(p->B, &po_a));
5708: PetscCall(PetscSFBcastBegin(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5709: PetscCall(PetscSFBcastBegin(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5710: PetscCall(PetscSFBcastEnd(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5711: PetscCall(PetscSFBcastEnd(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5712: PetscCall(MatSeqAIJRestoreArrayRead(p->A, &pd_a));
5713: PetscCall(MatSeqAIJRestoreArrayRead(p->B, &po_a));
5714: } else SETERRQ(comm, PETSC_ERR_ARG_UNKNOWN_TYPE, "Unknown reuse type");
5715: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAocols, A, P, 0, 0));
5716: PetscFunctionReturn(PETSC_SUCCESS);
5717: }
5719: /*@
5720: MatGetBrowsOfAcols - Returns `IS` that contain rows of `B` that equal to nonzero columns of local `A`
5722: Collective
5724: Input Parameters:
5725: + A - the first matrix in `MATMPIAIJ` format
5726: . B - the second matrix in `MATMPIAIJ` format
5727: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5729: Output Parameters:
5730: + rowb - On input index sets of rows of B to extract (or `NULL`), modified on output
5731: . colb - On input index sets of columns of B to extract (or `NULL`), modified on output
5732: - B_seq - the sequential matrix generated
5734: Level: developer
5736: .seealso: `Mat`, `MATMPIAIJ`, `IS`, `MatReuse`
5737: @*/
5738: PetscErrorCode MatGetBrowsOfAcols(Mat A, Mat B, MatReuse scall, IS *rowb, IS *colb, Mat *B_seq)
5739: {
5740: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5741: PetscInt *idx, i, start, ncols, nzA, nzB, *cmap, imark;
5742: IS isrowb, iscolb;
5743: Mat *bseq = NULL;
5745: PetscFunctionBegin;
5746: 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 ")",
5747: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
5748: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAcols, A, B, 0, 0));
5750: if (scall == MAT_INITIAL_MATRIX) {
5751: start = A->cmap->rstart;
5752: cmap = a->garray;
5753: nzA = a->A->cmap->n;
5754: nzB = a->B->cmap->n;
5755: PetscCall(PetscMalloc1(nzA + nzB, &idx));
5756: ncols = 0;
5757: for (i = 0; i < nzB; i++) { /* row < local row index */
5758: if (cmap[i] < start) idx[ncols++] = cmap[i];
5759: else break;
5760: }
5761: imark = i;
5762: for (i = 0; i < nzA; i++) idx[ncols++] = start + i; /* local rows */
5763: for (i = imark; i < nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */
5764: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, &isrowb));
5765: PetscCall(ISCreateStride(PETSC_COMM_SELF, B->cmap->N, 0, 1, &iscolb));
5766: } else {
5767: PetscCheck(rowb && colb, PETSC_COMM_SELF, PETSC_ERR_SUP, "IS rowb and colb must be provided for MAT_REUSE_MATRIX");
5768: isrowb = *rowb;
5769: iscolb = *colb;
5770: PetscCall(PetscMalloc1(1, &bseq));
5771: bseq[0] = *B_seq;
5772: }
5773: PetscCall(MatCreateSubMatrices(B, 1, &isrowb, &iscolb, scall, &bseq));
5774: *B_seq = bseq[0];
5775: PetscCall(PetscFree(bseq));
5776: if (!rowb) {
5777: PetscCall(ISDestroy(&isrowb));
5778: } else {
5779: *rowb = isrowb;
5780: }
5781: if (!colb) {
5782: PetscCall(ISDestroy(&iscolb));
5783: } else {
5784: *colb = iscolb;
5785: }
5786: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAcols, A, B, 0, 0));
5787: PetscFunctionReturn(PETSC_SUCCESS);
5788: }
5790: // PetscClangLinter pragma disable: -fdoc-sowing-chars
5791: /*
5792: MatGetBrowsOfAoCols_MPIAIJ - Creates a `MATSEQAIJ` matrix by taking rows of B that equal to nonzero columns
5793: of the OFF-DIAGONAL portion of local A
5795: Collective
5797: Input Parameters:
5798: + A - the first matrix in `MATMPIAIJ` format
5799: . B - the second matrix in `MATMPIAIJ` format
5800: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5802: Output Parameters:
5803: + startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or NULL)
5804: . startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or NULL)
5805: . bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or NULL)
5806: - B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N
5808: Level: developer
5810: Developer Note:
5811: This directly accesses information inside the VecScatter associated with the matrix-vector product
5812: for this matrix. This is not desirable.
5814: .seealso: [](ch_mat), `Mat`, `MATMPIAIJ`
5815: */
5816: PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat A, Mat B, MatReuse scall, PetscInt **startsj_s, PetscInt **startsj_r, MatScalar **bufa_ptr, Mat *B_oth)
5817: {
5818: PetscFunctionBegin;
5819: PetscCall(MatGetBrowsOfAoCols_MPIAIJ_Private(A, B, scall, PETSC_FALSE, startsj_s, startsj_r, bufa_ptr, B_oth));
5820: PetscFunctionReturn(PETSC_SUCCESS);
5821: }
5823: PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ_Private(Mat A, Mat B, MatReuse scall, PetscBool structure_only, PetscInt **startsj_s, PetscInt **startsj_r, MatScalar **bufa_ptr, Mat *B_oth)
5824: {
5825: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5826: VecScatter ctx;
5827: MPI_Comm comm;
5828: const PetscMPIInt *rprocs, *sprocs;
5829: PetscMPIInt nrecvs, nsends;
5830: const PetscInt *srow, *rstarts, *sstarts;
5831: PetscInt *rowlen, *bufj, *bufJ, ncols = 0, aBn = a->B->cmap->n, row, *b_othi, *b_othj, *rvalues = NULL, *svalues = NULL, *cols, sbs, rbs;
5832: PetscInt i, j, k = 0, l, ll, nrows, *rstartsj = NULL, *sstartsj, len;
5833: PetscScalar *b_otha = NULL, *bufa = NULL, *bufA, *vals = NULL;
5834: MPI_Request *reqs = NULL, *rwaits = NULL, *swaits = NULL;
5835: PetscMPIInt size, tag, rank, nreqs;
5837: PetscFunctionBegin;
5838: PetscCheck(!structure_only || scall == MAT_INITIAL_MATRIX, PETSC_COMM_SELF, PETSC_ERR_SUP, "Structure-only extraction requires MAT_INITIAL_MATRIX");
5839: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
5840: PetscCallMPI(MPI_Comm_size(comm, &size));
5842: 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 ")",
5843: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
5844: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAocols, A, B, 0, 0));
5845: PetscCallMPI(MPI_Comm_rank(comm, &rank));
5847: if (size == 1) {
5848: startsj_s = NULL;
5849: bufa_ptr = NULL;
5850: *B_oth = NULL;
5851: PetscFunctionReturn(PETSC_SUCCESS);
5852: }
5854: ctx = a->Mvctx;
5855: tag = ((PetscObject)ctx)->tag;
5857: PetscCall(VecScatterGetRemote_Private(ctx, PETSC_TRUE /*send*/, &nsends, &sstarts, &srow, &sprocs, &sbs));
5858: /* rprocs[] must be ordered so that indices received from them are ordered in rvalues[], which is key to algorithms used in this subroutine */
5859: PetscCall(VecScatterGetRemoteOrdered_Private(ctx, PETSC_FALSE /*recv*/, &nrecvs, &rstarts, NULL /*indices not needed*/, &rprocs, &rbs));
5860: PetscCall(PetscMPIIntCast(nsends + nrecvs, &nreqs));
5861: PetscCall(PetscMalloc1(nreqs, &reqs));
5862: rwaits = reqs;
5863: swaits = PetscSafePointerPlusOffset(reqs, nrecvs);
5865: if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX;
5866: if (scall == MAT_INITIAL_MATRIX) {
5867: /* i-array */
5868: /* post receives */
5869: if (nrecvs) PetscCall(PetscMalloc1(rbs * (rstarts[nrecvs] - rstarts[0]), &rvalues)); /* rstarts can be NULL when nrecvs=0 */
5870: for (i = 0; i < nrecvs; i++) {
5871: rowlen = rvalues + rstarts[i] * rbs;
5872: nrows = (rstarts[i + 1] - rstarts[i]) * rbs; /* num of indices to be received */
5873: PetscCallMPI(MPIU_Irecv(rowlen, nrows, MPIU_INT, rprocs[i], tag, comm, rwaits + i));
5874: }
5876: /* pack the outgoing message */
5877: PetscCall(PetscMalloc2(nsends + 1, &sstartsj, nrecvs + 1, &rstartsj));
5879: sstartsj[0] = 0;
5880: rstartsj[0] = 0;
5881: len = 0; /* total length of j or a array to be sent */
5882: if (nsends) {
5883: k = sstarts[0]; /* ATTENTION: sstarts[0] and rstarts[0] are not necessarily zero */
5884: PetscCall(PetscMalloc1(sbs * (sstarts[nsends] - sstarts[0]), &svalues));
5885: }
5886: for (i = 0; i < nsends; i++) {
5887: rowlen = svalues + (sstarts[i] - sstarts[0]) * sbs;
5888: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5889: for (j = 0; j < nrows; j++) {
5890: row = srow[k] + B->rmap->range[rank]; /* global row idx */
5891: for (l = 0; l < sbs; l++) {
5892: PetscCall(MatGetRow_MPIAIJ(B, row + l, &ncols, NULL, NULL)); /* rowlength */
5894: rowlen[j * sbs + l] = ncols;
5896: len += ncols;
5897: PetscCall(MatRestoreRow_MPIAIJ(B, row + l, &ncols, NULL, NULL));
5898: }
5899: k++;
5900: }
5901: PetscCallMPI(MPIU_Isend(rowlen, nrows * sbs, MPIU_INT, sprocs[i], tag, comm, swaits + i));
5903: sstartsj[i + 1] = len; /* starting point of (i+1)-th outgoing msg in bufj and bufa */
5904: }
5905: /* recvs and sends of i-array are completed */
5906: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5907: PetscCall(PetscFree(svalues));
5909: /* allocate buffers for sending j and a arrays */
5910: PetscCall(PetscMalloc1(len, &bufj));
5911: if (!structure_only) PetscCall(PetscMalloc1(len, &bufa));
5913: /* create i-array of B_oth */
5914: PetscCall(PetscMalloc1(aBn + 1, &b_othi));
5916: b_othi[0] = 0;
5917: len = 0; /* total length of j or a array to be received */
5918: k = 0;
5919: for (i = 0; i < nrecvs; i++) {
5920: rowlen = rvalues + (rstarts[i] - rstarts[0]) * rbs;
5921: nrows = (rstarts[i + 1] - rstarts[i]) * rbs; /* num of rows to be received */
5922: for (j = 0; j < nrows; j++) {
5923: b_othi[k + 1] = b_othi[k] + rowlen[j];
5924: PetscCall(PetscIntSumError(rowlen[j], len, &len));
5925: k++;
5926: }
5927: rstartsj[i + 1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */
5928: }
5929: PetscCall(PetscFree(rvalues));
5931: /* allocate space for j and a arrays of B_oth */
5932: PetscCall(PetscMalloc1(b_othi[aBn], &b_othj));
5933: if (!structure_only) PetscCall(PetscMalloc1(b_othi[aBn], &b_otha));
5935: /* j-array */
5936: /* post receives of j-array */
5937: for (i = 0; i < nrecvs; i++) {
5938: nrows = rstartsj[i + 1] - rstartsj[i]; /* length of the msg received */
5939: PetscCallMPI(MPIU_Irecv(PetscSafePointerPlusOffset(b_othj, rstartsj[i]), nrows, MPIU_INT, rprocs[i], tag, comm, rwaits + i));
5940: }
5942: /* pack the outgoing message j-array */
5943: if (nsends) k = sstarts[0];
5944: for (i = 0; i < nsends; i++) {
5945: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5946: bufJ = PetscSafePointerPlusOffset(bufj, sstartsj[i]);
5947: for (j = 0; j < nrows; j++) {
5948: row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5949: for (ll = 0; ll < sbs; ll++) {
5950: PetscCall(MatGetRow_MPIAIJ(B, row + ll, &ncols, &cols, NULL));
5951: for (l = 0; l < ncols; l++) *bufJ++ = cols[l];
5952: PetscCall(MatRestoreRow_MPIAIJ(B, row + ll, &ncols, &cols, NULL));
5953: }
5954: }
5955: PetscCallMPI(MPIU_Isend(PetscSafePointerPlusOffset(bufj, sstartsj[i]), sstartsj[i + 1] - sstartsj[i], MPIU_INT, sprocs[i], tag, comm, swaits + i));
5956: }
5958: /* recvs and sends of j-array are completed */
5959: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5960: } else if (scall == MAT_REUSE_MATRIX) {
5961: sstartsj = *startsj_s;
5962: rstartsj = *startsj_r;
5963: bufa = *bufa_ptr;
5964: PetscCall(MatSeqAIJGetArrayWrite(*B_oth, &b_otha));
5965: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not possess an object container");
5967: if (!structure_only) {
5968: /* a-array */
5969: /* post receives of a-array */
5970: for (i = 0; i < nrecvs; i++) {
5971: nrows = rstartsj[i + 1] - rstartsj[i]; /* length of the msg received */
5972: PetscCallMPI(MPIU_Irecv(PetscSafePointerPlusOffset(b_otha, rstartsj[i]), nrows, MPIU_SCALAR, rprocs[i], tag, comm, rwaits + i));
5973: }
5975: /* pack the outgoing message a-array */
5976: if (nsends) k = sstarts[0];
5977: for (i = 0; i < nsends; i++) {
5978: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5979: bufA = PetscSafePointerPlusOffset(bufa, sstartsj[i]);
5980: for (j = 0; j < nrows; j++) {
5981: row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5982: for (ll = 0; ll < sbs; ll++) {
5983: PetscCall(MatGetRow_MPIAIJ(B, row + ll, &ncols, NULL, &vals));
5984: for (l = 0; l < ncols; l++) *bufA++ = vals[l];
5985: PetscCall(MatRestoreRow_MPIAIJ(B, row + ll, &ncols, NULL, &vals));
5986: }
5987: }
5988: PetscCallMPI(MPIU_Isend(PetscSafePointerPlusOffset(bufa, sstartsj[i]), sstartsj[i + 1] - sstartsj[i], MPIU_SCALAR, sprocs[i], tag, comm, swaits + i));
5989: }
5990: /* recvs and sends of a-array are completed */
5991: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5992: }
5993: PetscCall(PetscFree(reqs));
5995: if (scall == MAT_INITIAL_MATRIX) {
5996: Mat_SeqAIJ *b_oth;
5998: /* put together the new matrix */
5999: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, aBn, B->cmap->N, b_othi, b_othj, b_otha, B_oth));
6000: PetscCall(MatSetOption(*B_oth, MAT_STRUCTURE_ONLY, structure_only));
6002: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
6003: /* Since these are PETSc arrays, change flags to free them as necessary. */
6004: b_oth = (Mat_SeqAIJ *)(*B_oth)->data;
6005: b_oth->free_a = PETSC_TRUE;
6006: b_oth->free_ij = PETSC_TRUE;
6007: b_oth->nonew = 0;
6009: PetscCall(PetscFree(bufj));
6010: if (!startsj_s || !bufa_ptr) {
6011: PetscCall(PetscFree2(sstartsj, rstartsj));
6012: PetscCall(PetscFree(bufa_ptr));
6013: } else {
6014: *startsj_s = sstartsj;
6015: *startsj_r = rstartsj;
6016: *bufa_ptr = bufa;
6017: }
6018: } else if (scall == MAT_REUSE_MATRIX) {
6019: PetscCall(MatSeqAIJRestoreArrayWrite(*B_oth, &b_otha));
6020: }
6022: PetscCall(VecScatterRestoreRemote_Private(ctx, PETSC_TRUE, &nsends, &sstarts, &srow, &sprocs, &sbs));
6023: PetscCall(VecScatterRestoreRemoteOrdered_Private(ctx, PETSC_FALSE, &nrecvs, &rstarts, NULL, &rprocs, &rbs));
6024: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAocols, A, B, 0, 0));
6025: PetscFunctionReturn(PETSC_SUCCESS);
6026: }
6028: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat, MatType, MatReuse, Mat *);
6029: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJPERM(Mat, MatType, MatReuse, Mat *);
6030: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJSELL(Mat, MatType, MatReuse, Mat *);
6031: #if PetscDefined(HAVE_MKL_SPARSE)
6032: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJMKL(Mat, MatType, MatReuse, Mat *);
6033: #endif
6034: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIBAIJ(Mat, MatType, MatReuse, Mat *);
6035: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISBAIJ(Mat, MatType, MatReuse, Mat *);
6036: #if PetscDefined(HAVE_ELEMENTAL)
6037: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat, MatType, MatReuse, Mat *);
6038: #endif
6039: #if PetscDefined(HAVE_SCALAPACK) && (PetscDefined(USE_REAL_SINGLE) || PetscDefined(USE_REAL_DOUBLE))
6040: PETSC_INTERN PetscErrorCode MatConvert_AIJ_ScaLAPACK(Mat, MatType, MatReuse, Mat *);
6041: #endif
6042: #if PetscDefined(HAVE_HYPRE)
6043: PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat, MatType, MatReuse, Mat *);
6044: #endif
6045: #if PetscDefined(HAVE_CUDA)
6046: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCUSPARSE(Mat, MatType, MatReuse, Mat *);
6047: #endif
6048: #if PetscDefined(HAVE_HIP)
6049: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJHIPSPARSE(Mat, MatType, MatReuse, Mat *);
6050: #endif
6051: #if PetscDefined(HAVE_KOKKOS_KERNELS)
6052: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJKokkos(Mat, MatType, MatReuse, Mat *);
6053: #endif
6054: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISELL(Mat, MatType, MatReuse, Mat *);
6055: PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat, MatType, MatReuse, Mat *);
6056: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_IS_XAIJ(Mat);
6058: /*
6059: Computes (B'*A')' since computing B*A directly is untenable
6061: n p p
6062: [ ] [ ] [ ]
6063: m [ A ] * n [ B ] = m [ C ]
6064: [ ] [ ] [ ]
6066: */
6067: static PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A, Mat B, Mat C)
6068: {
6069: Mat At, Bt, Ct;
6071: PetscFunctionBegin;
6072: PetscCall(MatTranspose(A, MAT_INITIAL_MATRIX, &At));
6073: PetscCall(MatTranspose(B, MAT_INITIAL_MATRIX, &Bt));
6074: PetscCall(MatMatMult(Bt, At, MAT_INITIAL_MATRIX, PETSC_CURRENT, &Ct));
6075: PetscCall(MatDestroy(&At));
6076: PetscCall(MatDestroy(&Bt));
6077: PetscCall(MatTransposeSetPrecursor(Ct, C));
6078: PetscCall(MatTranspose(Ct, MAT_REUSE_MATRIX, &C));
6079: PetscCall(MatDestroy(&Ct));
6080: PetscFunctionReturn(PETSC_SUCCESS);
6081: }
6083: static PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A, Mat B, PetscReal fill, Mat C)
6084: {
6085: PetscBool cisdense;
6087: PetscFunctionBegin;
6088: 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);
6089: PetscCall(MatSetSizes(C, A->rmap->n, B->cmap->n, A->rmap->N, B->cmap->N));
6090: PetscCall(MatSetBlockSizesFromMats(C, A, B));
6091: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATMPIDENSE, MATMPIDENSECUDA, MATMPIDENSEHIP, ""));
6092: if (!cisdense) {
6093: PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
6094: PetscCall(MatSetVecType(C, A->defaultvectype));
6095: }
6096: PetscCall(MatSetUp(C));
6098: C->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIAIJ;
6099: PetscFunctionReturn(PETSC_SUCCESS);
6100: }
6102: static PetscErrorCode MatProductSetFromOptions_MPIDense_MPIAIJ_AB(Mat C)
6103: {
6104: Mat_Product *product = C->product;
6105: Mat A = product->A, B = product->B;
6107: PetscFunctionBegin;
6108: PetscCheck(A->cmap->rstart == B->rmap->rstart && A->cmap->rend == B->rmap->rend, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Matrix local dimensions are incompatible, (%" PetscInt_FMT ", %" PetscInt_FMT ") != (%" PetscInt_FMT ",%" PetscInt_FMT ")",
6109: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
6110: C->ops->matmultsymbolic = MatMatMultSymbolic_MPIDense_MPIAIJ;
6111: C->ops->productsymbolic = MatProductSymbolic_AB;
6112: PetscFunctionReturn(PETSC_SUCCESS);
6113: }
6115: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_MPIDense_MPIAIJ(Mat C)
6116: {
6117: Mat_Product *product = C->product;
6119: PetscFunctionBegin;
6120: if (product->type == MATPRODUCT_AB) PetscCall(MatProductSetFromOptions_MPIDense_MPIAIJ_AB(C));
6121: PetscFunctionReturn(PETSC_SUCCESS);
6122: }
6124: /*
6125: Merge two sets of sorted nonzeros and return a CSR for the merged (sequential) matrix
6127: Input Parameters:
6129: j1,rowBegin1,rowEnd1,jmap1: describe the first set of nonzeros (Set1)
6130: j2,rowBegin2,rowEnd2,jmap2: describe the second set of nonzeros (Set2)
6132: mat: both sets' nonzeros are on m rows, where m is the number of local rows of the matrix mat
6134: For Set1, j1[] contains column indices of the nonzeros.
6135: For the k-th row (0<=k<m), [rowBegin1[k],rowEnd1[k]) index into j1[] and point to the begin/end nonzero in row k
6136: respectively (note rowEnd1[k] is not necessarily equal to rwoBegin1[k+1]). Indices in this range of j1[] are sorted,
6137: but might have repeats. jmap1[t+1] - jmap1[t] is the number of repeats for the t-th unique nonzero in Set1.
6139: Similar for Set2.
6141: This routine merges the two sets of nonzeros row by row and removes repeats.
6143: Output Parameters: (memory is allocated by the caller)
6145: i[],j[]: the CSR of the merged matrix, which has m rows.
6146: imap1[]: the k-th unique nonzero in Set1 (k=0,1,...) corresponds to imap1[k]-th unique nonzero in the merged matrix.
6147: imap2[]: similar to imap1[], but for Set2.
6148: Note we order nonzeros row-by-row and from left to right.
6149: */
6150: static PetscErrorCode MatMergeEntries_Internal(Mat mat, const PetscInt j1[], const PetscInt j2[], const PetscCount rowBegin1[], const PetscCount rowEnd1[], const PetscCount rowBegin2[], const PetscCount rowEnd2[], const PetscCount jmap1[], const PetscCount jmap2[], PetscCount imap1[], PetscCount imap2[], PetscInt i[], PetscInt j[])
6151: {
6152: PetscInt r, m; /* Row index of mat */
6153: PetscCount t, t1, t2, b1, e1, b2, e2;
6155: PetscFunctionBegin;
6156: PetscCall(MatGetLocalSize(mat, &m, NULL));
6157: t1 = t2 = t = 0; /* Count unique nonzeros of in Set1, Set1 and the merged respectively */
6158: i[0] = 0;
6159: for (r = 0; r < m; r++) { /* Do row by row merging */
6160: b1 = rowBegin1[r];
6161: e1 = rowEnd1[r];
6162: b2 = rowBegin2[r];
6163: e2 = rowEnd2[r];
6164: while (b1 < e1 && b2 < e2) {
6165: if (j1[b1] == j2[b2]) { /* Same column index and hence same nonzero */
6166: j[t] = j1[b1];
6167: imap1[t1] = t;
6168: imap2[t2] = t;
6169: b1 += jmap1[t1 + 1] - jmap1[t1]; /* Jump to next unique local nonzero */
6170: b2 += jmap2[t2 + 1] - jmap2[t2]; /* Jump to next unique remote nonzero */
6171: t1++;
6172: t2++;
6173: t++;
6174: } else if (j1[b1] < j2[b2]) {
6175: j[t] = j1[b1];
6176: imap1[t1] = t;
6177: b1 += jmap1[t1 + 1] - jmap1[t1];
6178: t1++;
6179: t++;
6180: } else {
6181: j[t] = j2[b2];
6182: imap2[t2] = t;
6183: b2 += jmap2[t2 + 1] - jmap2[t2];
6184: t2++;
6185: t++;
6186: }
6187: }
6188: /* Merge the remaining in either j1[] or j2[] */
6189: while (b1 < e1) {
6190: j[t] = j1[b1];
6191: imap1[t1] = t;
6192: b1 += jmap1[t1 + 1] - jmap1[t1];
6193: t1++;
6194: t++;
6195: }
6196: while (b2 < e2) {
6197: j[t] = j2[b2];
6198: imap2[t2] = t;
6199: b2 += jmap2[t2 + 1] - jmap2[t2];
6200: t2++;
6201: t++;
6202: }
6203: PetscCall(PetscIntCast(t, i + r + 1));
6204: }
6205: PetscFunctionReturn(PETSC_SUCCESS);
6206: }
6208: /*
6209: Split nonzeros in a block of local rows into two subsets: those in the diagonal block and those in the off-diagonal block
6211: Input Parameters:
6212: mat: an MPI matrix that provides row and column layout information for splitting. Let's say its number of local rows is m.
6213: n,i[],j[],perm[]: there are n input entries, belonging to m rows. Row/col indices of the entries are stored in i[] and j[]
6214: respectively, along with a permutation array perm[]. Length of the i[],j[],perm[] arrays is n.
6216: i[] is already sorted, but within a row, j[] is not sorted and might have repeats.
6217: i[] might contain negative indices at the beginning, which means the corresponding entries should be ignored in the splitting.
6219: Output Parameters:
6220: j[],perm[]: the routine needs to sort j[] within each row along with perm[].
6221: rowBegin[],rowMid[],rowEnd[]: of length m, and the memory is preallocated and zeroed by the caller.
6222: They contain indices pointing to j[]. For 0<=r<m, [rowBegin[r],rowMid[r]) point to begin/end entries of row r of the diagonal block,
6223: and [rowMid[r],rowEnd[r]) point to begin/end entries of row r of the off-diagonal block.
6225: Aperm[],Ajmap[],Atot,Annz: Arrays are allocated by this routine.
6226: Atot: number of entries belonging to the diagonal block.
6227: Annz: number of unique nonzeros belonging to the diagonal block.
6228: Aperm[Atot] stores values from perm[] for entries belonging to the diagonal block. Length of Aperm[] is Atot, though it may also count
6229: repeats (i.e., same 'i,j' pair).
6230: Ajmap[Annz+1] stores the number of repeats of each unique entry belonging to the diagonal block. More precisely, Ajmap[t+1] - Ajmap[t]
6231: is the number of repeats for the t-th unique entry in the diagonal block. Ajmap[0] is always 0.
6233: Atot: number of entries belonging to the diagonal block
6234: Annz: number of unique nonzeros belonging to the diagonal block.
6236: Bperm[], Bjmap[], Btot, Bnnz are similar but for the off-diagonal block.
6238: Aperm[],Bperm[],Ajmap[] and Bjmap[] are allocated separately by this routine with PetscMalloc1().
6239: */
6240: static PetscErrorCode MatSplitEntries_Internal(Mat mat, PetscCount n, const PetscInt i[], PetscInt j[], PetscCount perm[], PetscCount rowBegin[], PetscCount rowMid[], PetscCount rowEnd[], PetscCount *Atot_, PetscCount **Aperm_, PetscCount *Annz_, PetscCount **Ajmap_, PetscCount *Btot_, PetscCount **Bperm_, PetscCount *Bnnz_, PetscCount **Bjmap_)
6241: {
6242: PetscInt cstart, cend, rstart, rend, row, col;
6243: PetscCount Atot = 0, Btot = 0; /* Total number of nonzeros in the diagonal and off-diagonal blocks */
6244: PetscCount Annz = 0, Bnnz = 0; /* Number of unique nonzeros in the diagonal and off-diagonal blocks */
6245: PetscCount k, m, p, q, r, s, mid;
6246: PetscCount *Aperm, *Bperm, *Ajmap, *Bjmap;
6248: PetscFunctionBegin;
6249: PetscCall(PetscLayoutGetRange(mat->rmap, &rstart, &rend));
6250: PetscCall(PetscLayoutGetRange(mat->cmap, &cstart, &cend));
6251: m = rend - rstart;
6253: /* Skip negative rows */
6254: for (k = 0; k < n; k++)
6255: if (i[k] >= 0) break;
6257: /* Process [k,n): sort and partition each local row into diag and offdiag portions,
6258: fill rowBegin[], rowMid[], rowEnd[], and count Atot, Btot, Annz, Bnnz.
6259: */
6260: while (k < n) {
6261: row = i[k];
6262: /* Entries in [k,s) are in one row. Shift diagonal block col indices so that diag is ahead of offdiag after sorting the row */
6263: for (s = k; s < n; s++)
6264: if (i[s] != row) break;
6266: /* Shift diag columns to range of [-PETSC_INT_MAX, -1] */
6267: for (p = k; p < s; p++) {
6268: if (j[p] >= cstart && j[p] < cend) j[p] -= PETSC_INT_MAX;
6269: }
6270: PetscCall(PetscSortIntWithCountArray(s - k, j + k, perm + k));
6271: PetscCall(PetscSortedIntUpperBound(j, k, s, -1, &mid)); /* Separate [k,s) into [k,mid) for diag and [mid,s) for offdiag */
6272: rowBegin[row - rstart] = k;
6273: rowMid[row - rstart] = mid;
6274: rowEnd[row - rstart] = s;
6275: PetscCheck(k == s || j[s - 1] < mat->cmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column index %" PetscInt_FMT " is >= matrix column size %" PetscInt_FMT, j[s - 1], mat->cmap->N);
6277: /* Count nonzeros of this diag/offdiag row, which might have repeats */
6278: Atot += mid - k;
6279: Btot += s - mid;
6281: /* Count unique nonzeros of this diag row */
6282: for (p = k; p < mid;) {
6283: col = j[p];
6284: do {
6285: j[p] += PETSC_INT_MAX; /* Revert the modified diagonal indices */
6286: p++;
6287: } while (p < mid && j[p] == col);
6288: Annz++;
6289: }
6291: /* Count unique nonzeros of this offdiag row */
6292: for (p = mid; p < s;) {
6293: col = j[p];
6294: do {
6295: p++;
6296: } while (p < s && j[p] == col);
6297: Bnnz++;
6298: }
6299: k = s;
6300: }
6302: /* Allocation according to Atot, Btot, Annz, Bnnz */
6303: PetscCall(PetscMalloc1(Atot, &Aperm));
6304: PetscCall(PetscMalloc1(Btot, &Bperm));
6305: PetscCall(PetscMalloc1(Annz + 1, &Ajmap));
6306: PetscCall(PetscMalloc1(Bnnz + 1, &Bjmap));
6308: /* Re-scan indices and copy diag/offdiag permutation indices to Aperm, Bperm and also fill Ajmap and Bjmap */
6309: Ajmap[0] = Bjmap[0] = Atot = Btot = Annz = Bnnz = 0;
6310: for (r = 0; r < m; r++) {
6311: k = rowBegin[r];
6312: mid = rowMid[r];
6313: s = rowEnd[r];
6314: PetscCall(PetscArraycpy(PetscSafePointerPlusOffset(Aperm, Atot), PetscSafePointerPlusOffset(perm, k), mid - k));
6315: PetscCall(PetscArraycpy(PetscSafePointerPlusOffset(Bperm, Btot), PetscSafePointerPlusOffset(perm, mid), s - mid));
6316: Atot += mid - k;
6317: Btot += s - mid;
6319: /* Scan column indices in this row and find out how many repeats each unique nonzero has */
6320: for (p = k; p < mid;) {
6321: col = j[p];
6322: q = p;
6323: do {
6324: p++;
6325: } while (p < mid && j[p] == col);
6326: Ajmap[Annz + 1] = Ajmap[Annz] + (p - q);
6327: Annz++;
6328: }
6330: for (p = mid; p < s;) {
6331: col = j[p];
6332: q = p;
6333: do {
6334: p++;
6335: } while (p < s && j[p] == col);
6336: Bjmap[Bnnz + 1] = Bjmap[Bnnz] + (p - q);
6337: Bnnz++;
6338: }
6339: }
6340: /* Output */
6341: *Aperm_ = Aperm;
6342: *Annz_ = Annz;
6343: *Atot_ = Atot;
6344: *Ajmap_ = Ajmap;
6345: *Bperm_ = Bperm;
6346: *Bnnz_ = Bnnz;
6347: *Btot_ = Btot;
6348: *Bjmap_ = Bjmap;
6349: PetscFunctionReturn(PETSC_SUCCESS);
6350: }
6352: /*
6353: Expand the jmap[] array to make a new one in view of nonzeros in the merged matrix
6355: Input Parameters:
6356: nnz1: number of unique nonzeros in a set that was used to produce imap[], jmap[]
6357: nnz: number of unique nonzeros in the merged matrix
6358: imap[nnz1]: i-th nonzero in the set is the imap[i]-th nonzero in the merged matrix
6359: jmap[nnz1+1]: i-th nonzero in the set has jmap[i+1] - jmap[i] repeats in the set
6361: Output Parameter: (memory is allocated by the caller)
6362: jmap_new[nnz+1]: i-th nonzero in the merged matrix has jmap_new[i+1] - jmap_new[i] repeats in the set
6364: Example:
6365: nnz1 = 4
6366: nnz = 6
6367: imap = [1,3,4,5]
6368: jmap = [0,3,5,6,7]
6369: then,
6370: jmap_new = [0,0,3,3,5,6,7]
6371: */
6372: static PetscErrorCode ExpandJmap_Internal(PetscCount nnz1, PetscCount nnz, const PetscCount imap[], const PetscCount jmap[], PetscCount jmap_new[])
6373: {
6374: PetscCount k, p;
6376: PetscFunctionBegin;
6377: jmap_new[0] = 0;
6378: p = nnz; /* p loops over jmap_new[] backwards */
6379: for (k = nnz1 - 1; k >= 0; k--) { /* k loops over imap[] */
6380: for (; p > imap[k]; p--) jmap_new[p] = jmap[k + 1];
6381: }
6382: for (; p >= 0; p--) jmap_new[p] = jmap[0];
6383: PetscFunctionReturn(PETSC_SUCCESS);
6384: }
6386: static PetscErrorCode MatCOOStructDestroy_MPIAIJ(PetscCtxRt data)
6387: {
6388: MatCOOStruct_MPIAIJ *coo = *(MatCOOStruct_MPIAIJ **)data;
6390: PetscFunctionBegin;
6391: PetscCall(PetscSFDestroy(&coo->sf));
6392: PetscCall(PetscFree(coo->Aperm1));
6393: PetscCall(PetscFree(coo->Bperm1));
6394: PetscCall(PetscFree(coo->Ajmap1));
6395: PetscCall(PetscFree(coo->Bjmap1));
6396: PetscCall(PetscFree(coo->Aimap2));
6397: PetscCall(PetscFree(coo->Bimap2));
6398: PetscCall(PetscFree(coo->Aperm2));
6399: PetscCall(PetscFree(coo->Bperm2));
6400: PetscCall(PetscFree(coo->Ajmap2));
6401: PetscCall(PetscFree(coo->Bjmap2));
6402: PetscCall(PetscFree(coo->Cperm1));
6403: PetscCall(PetscFree2(coo->sendbuf, coo->recvbuf));
6404: PetscCall(PetscFree(coo));
6405: PetscFunctionReturn(PETSC_SUCCESS);
6406: }
6408: PetscErrorCode MatSetPreallocationCOO_MPIAIJ(Mat mat, PetscCount coo_n, PetscInt coo_i[], PetscInt coo_j[])
6409: {
6410: MPI_Comm comm;
6411: PetscMPIInt rank, size;
6412: PetscInt m, n, M, N, rstart, rend, cstart, cend; /* Sizes, indices of row/col, therefore with type PetscInt */
6413: PetscCount k, p, q, rem; /* Loop variables over coo arrays */
6414: Mat_MPIAIJ *mpiaij = (Mat_MPIAIJ *)mat->data;
6415: PetscContainer container;
6416: MatCOOStruct_MPIAIJ *coo;
6418: PetscFunctionBegin;
6419: // The A and B blocks built below leave B with compacted column indices, which MatSetValues_MPI_Hash() cannot fill
6420: if (mat->hash_active) {
6421: mat->ops[0] = mpiaij->cops;
6422: mat->hash_active = PETSC_FALSE;
6423: }
6424: PetscCall(PetscFree(mpiaij->garray));
6425: PetscCall(VecDestroy(&mpiaij->lvec));
6426: #if PetscDefined(USE_CTABLE)
6427: PetscCall(PetscHMapIDestroy(&mpiaij->colmap));
6428: #else
6429: PetscCall(PetscFree(mpiaij->colmap));
6430: #endif
6431: PetscCall(VecScatterDestroy(&mpiaij->Mvctx));
6432: mat->assembled = PETSC_FALSE;
6433: mat->was_assembled = PETSC_FALSE;
6435: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
6436: PetscCallMPI(MPI_Comm_size(comm, &size));
6437: PetscCallMPI(MPI_Comm_rank(comm, &rank));
6438: PetscCall(PetscLayoutSetUp(mat->rmap));
6439: PetscCall(PetscLayoutSetUp(mat->cmap));
6440: PetscCall(PetscLayoutGetRange(mat->rmap, &rstart, &rend));
6441: PetscCall(PetscLayoutGetRange(mat->cmap, &cstart, &cend));
6442: PetscCall(MatGetLocalSize(mat, &m, &n));
6443: PetscCall(MatGetSize(mat, &M, &N));
6445: /* Sort (i,j) by row along with a permutation array, so that the to-be-ignored */
6446: /* entries come first, then local rows, then remote rows. */
6447: PetscCount n1 = coo_n, *perm1;
6448: PetscInt *i1 = coo_i, *j1 = coo_j;
6450: PetscCall(PetscMalloc1(n1, &perm1));
6451: for (k = 0; k < n1; k++) perm1[k] = k;
6453: /* Manipulate indices so that entries with negative row or col indices will have smallest
6454: row indices, local entries will have greater but negative row indices, and remote entries
6455: will have positive row indices.
6456: */
6457: for (k = 0; k < n1; k++) {
6458: if (i1[k] < 0 || j1[k] < 0) i1[k] = PETSC_INT_MIN; /* e.g., -2^31, minimal to move them ahead */
6459: 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] */
6460: else {
6461: PetscCheck(!mat->nooffprocentries, PETSC_COMM_SELF, PETSC_ERR_USER_INPUT, "MAT_NO_OFF_PROC_ENTRIES is set but insert to remote rows");
6462: if (mpiaij->donotstash) i1[k] = PETSC_INT_MIN; /* Ignore offproc entries as if they had negative indices */
6463: }
6464: }
6466: /* Sort by row; after that, [0,k) have ignored entries, [k,rem) have local rows and [rem,n1) have remote rows */
6467: PetscCall(PetscSortIntWithIntCountArrayPair(n1, i1, j1, perm1));
6469: /* Advance k to the first entry we need to take care of */
6470: for (k = 0; k < n1; k++)
6471: if (i1[k] > PETSC_INT_MIN) break;
6472: PetscCount i1start = k;
6474: PetscCall(PetscSortedIntUpperBound(i1, k, n1, rend - 1 - PETSC_INT_MAX, &rem)); /* rem is upper bound of the last local row */
6475: for (; k < rem; k++) i1[k] += PETSC_INT_MAX; /* Revert row indices of local rows*/
6477: PetscCheck(n1 == 0 || i1[n1 - 1] < M, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "COO row index %" PetscInt_FMT " is >= the matrix row size %" PetscInt_FMT, i1[n1 - 1], M);
6479: /* Send remote rows to their owner */
6480: /* Find which rows should be sent to which remote ranks*/
6481: PetscInt nsend = 0; /* Number of MPI ranks to send data to */
6482: PetscMPIInt *sendto; /* [nsend], storing remote ranks */
6483: PetscInt *nentries; /* [nsend], storing number of entries sent to remote ranks; Assume PetscInt is big enough for this count, and error if not */
6484: const PetscInt *ranges;
6485: PetscInt maxNsend = size >= 128 ? 128 : size; /* Assume max 128 neighbors; realloc when needed */
6487: PetscCall(PetscLayoutGetRanges(mat->rmap, &ranges));
6488: PetscCall(PetscMalloc2(maxNsend, &sendto, maxNsend, &nentries));
6489: for (k = rem; k < n1;) {
6490: PetscMPIInt owner;
6491: PetscInt firstRow, lastRow;
6493: /* Locate a row range */
6494: firstRow = i1[k]; /* first row of this owner */
6495: PetscCall(PetscLayoutFindOwner(mat->rmap, firstRow, &owner));
6496: lastRow = ranges[owner + 1] - 1; /* last row of this owner */
6498: /* Find the first index 'p' in [k,n) with i1[p] belonging to next owner */
6499: PetscCall(PetscSortedIntUpperBound(i1, k, n1, lastRow, &p));
6501: /* All entries in [k,p) belong to this remote owner */
6502: if (nsend >= maxNsend) { /* Double the remote ranks arrays if not long enough */
6503: PetscMPIInt *sendto2;
6504: PetscInt *nentries2;
6505: PetscInt maxNsend2 = (maxNsend <= size / 2) ? maxNsend * 2 : size;
6507: PetscCall(PetscMalloc2(maxNsend2, &sendto2, maxNsend2, &nentries2));
6508: PetscCall(PetscArraycpy(sendto2, sendto, maxNsend));
6509: PetscCall(PetscArraycpy(nentries2, nentries, maxNsend));
6510: PetscCall(PetscFree2(sendto, nentries));
6511: sendto = sendto2;
6512: nentries = nentries2;
6513: maxNsend = maxNsend2;
6514: }
6515: sendto[nsend] = owner;
6516: PetscCall(PetscIntCast(p - k, &nentries[nsend]));
6517: nsend++;
6518: k = p;
6519: }
6521: /* Build 1st SF to know offsets on remote to send data */
6522: PetscSF sf1;
6523: PetscInt nroots = 1, nroots2 = 0;
6524: PetscInt nleaves = nsend, nleaves2 = 0;
6525: PetscInt *offsets;
6526: PetscSFNode *iremote;
6528: PetscCall(PetscSFCreate(comm, &sf1));
6529: PetscCall(PetscMalloc1(nsend, &iremote));
6530: PetscCall(PetscMalloc1(nsend, &offsets));
6531: for (k = 0; k < nsend; k++) {
6532: iremote[k].rank = sendto[k];
6533: iremote[k].index = 0;
6534: nleaves2 += nentries[k];
6535: PetscCheck(nleaves2 >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Number of SF leaves is too large for PetscInt");
6536: }
6537: PetscCall(PetscSFSetGraph(sf1, nroots, nleaves, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
6538: PetscCall(PetscSFFetchAndOpWithMemTypeBegin(sf1, MPIU_INT, PETSC_MEMTYPE_HOST, &nroots2 /*rootdata*/, PETSC_MEMTYPE_HOST, nentries /*leafdata*/, PETSC_MEMTYPE_HOST, offsets /*leafupdate*/, MPI_SUM));
6539: PetscCall(PetscSFFetchAndOpEnd(sf1, MPIU_INT, &nroots2, nentries, offsets, MPI_SUM)); /* Would nroots2 overflow, we check offsets[] below */
6540: PetscCall(PetscSFDestroy(&sf1));
6541: PetscAssert(nleaves2 == n1 - rem, PETSC_COMM_SELF, PETSC_ERR_PLIB, "nleaves2 %" PetscInt_FMT " != number of remote entries %" PetscCount_FMT, nleaves2, n1 - rem);
6543: /* Build 2nd SF to send remote COOs to their owner */
6544: PetscSF sf2;
6545: nroots = nroots2;
6546: nleaves = nleaves2;
6547: PetscCall(PetscSFCreate(comm, &sf2));
6548: PetscCall(PetscSFSetFromOptions(sf2));
6549: PetscCall(PetscMalloc1(nleaves, &iremote));
6550: p = 0;
6551: for (k = 0; k < nsend; k++) {
6552: PetscCheck(offsets[k] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Number of SF roots is too large for PetscInt");
6553: for (q = 0; q < nentries[k]; q++, p++) {
6554: iremote[p].rank = sendto[k];
6555: PetscCall(PetscIntCast(offsets[k] + q, &iremote[p].index));
6556: }
6557: }
6558: PetscCall(PetscSFSetGraph(sf2, nroots, nleaves, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
6560: /* Send the remote COOs to their owner */
6561: PetscInt n2 = nroots, *i2, *j2; /* Buffers for received COOs from other ranks, along with a permutation array */
6562: PetscCount *perm2; /* Though PetscInt is enough for remote entries, we use PetscCount here as we want to reuse MatSplitEntries_Internal() */
6563: PetscCall(PetscMalloc3(n2, &i2, n2, &j2, n2, &perm2));
6564: PetscAssert(rem == 0 || i1 != NULL, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot add nonzero offset to null");
6565: PetscAssert(rem == 0 || j1 != NULL, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot add nonzero offset to null");
6566: PetscInt *i1prem = PetscSafePointerPlusOffset(i1, rem);
6567: PetscInt *j1prem = PetscSafePointerPlusOffset(j1, rem);
6568: PetscCall(PetscSFReduceWithMemTypeBegin(sf2, MPIU_INT, PETSC_MEMTYPE_HOST, i1prem, PETSC_MEMTYPE_HOST, i2, MPI_REPLACE));
6569: PetscCall(PetscSFReduceEnd(sf2, MPIU_INT, i1prem, i2, MPI_REPLACE));
6570: PetscCall(PetscSFReduceWithMemTypeBegin(sf2, MPIU_INT, PETSC_MEMTYPE_HOST, j1prem, PETSC_MEMTYPE_HOST, j2, MPI_REPLACE));
6571: PetscCall(PetscSFReduceEnd(sf2, MPIU_INT, j1prem, j2, MPI_REPLACE));
6573: PetscCall(PetscFree(offsets));
6574: PetscCall(PetscFree2(sendto, nentries));
6576: /* Sort received COOs by row along with the permutation array */
6577: for (k = 0; k < n2; k++) perm2[k] = k;
6578: PetscCall(PetscSortIntWithIntCountArrayPair(n2, i2, j2, perm2));
6580: /* sf2 only sends contiguous leafdata to contiguous rootdata. We record the permutation which will be used to fill leafdata */
6581: PetscCount *Cperm1;
6582: PetscAssert(rem == 0 || perm1 != NULL, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot add nonzero offset to null");
6583: PetscCount *perm1prem = PetscSafePointerPlusOffset(perm1, rem);
6584: PetscCall(PetscMalloc1(nleaves, &Cperm1));
6585: PetscCall(PetscArraycpy(Cperm1, perm1prem, nleaves));
6587: /* Support for HYPRE matrices, kind of a hack.
6588: Swap min column with diagonal so that diagonal values will go first */
6589: PetscBool hypre;
6590: PetscCall(PetscStrcmp("_internal_COO_mat_for_hypre", ((PetscObject)mat)->name, &hypre));
6591: if (hypre) {
6592: PetscInt *minj;
6593: PetscBT hasdiag;
6595: PetscCall(PetscBTCreate(m, &hasdiag));
6596: PetscCall(PetscMalloc1(m, &minj));
6597: for (k = 0; k < m; k++) minj[k] = PETSC_INT_MAX;
6598: for (k = i1start; k < rem; k++) {
6599: if (j1[k] < cstart || j1[k] >= cend) continue;
6600: const PetscInt rindex = i1[k] - rstart;
6601: if ((j1[k] - cstart) == rindex) PetscCall(PetscBTSet(hasdiag, rindex));
6602: minj[rindex] = PetscMin(minj[rindex], j1[k]);
6603: }
6604: for (k = 0; k < n2; k++) {
6605: if (j2[k] < cstart || j2[k] >= cend) continue;
6606: const PetscInt rindex = i2[k] - rstart;
6607: if ((j2[k] - cstart) == rindex) PetscCall(PetscBTSet(hasdiag, rindex));
6608: minj[rindex] = PetscMin(minj[rindex], j2[k]);
6609: }
6610: for (k = i1start; k < rem; k++) {
6611: const PetscInt rindex = i1[k] - rstart;
6612: if (j1[k] < cstart || j1[k] >= cend || !PetscBTLookup(hasdiag, rindex)) continue;
6613: if (j1[k] == minj[rindex]) j1[k] = i1[k] + (cstart - rstart);
6614: else if ((j1[k] - cstart) == rindex) j1[k] = minj[rindex];
6615: }
6616: for (k = 0; k < n2; k++) {
6617: const PetscInt rindex = i2[k] - rstart;
6618: if (j2[k] < cstart || j2[k] >= cend || !PetscBTLookup(hasdiag, rindex)) continue;
6619: if (j2[k] == minj[rindex]) j2[k] = i2[k] + (cstart - rstart);
6620: else if ((j2[k] - cstart) == rindex) j2[k] = minj[rindex];
6621: }
6622: PetscCall(PetscBTDestroy(&hasdiag));
6623: PetscCall(PetscFree(minj));
6624: }
6626: /* Split local COOs and received COOs into diag/offdiag portions */
6627: PetscCount *rowBegin1, *rowMid1, *rowEnd1;
6628: PetscCount *Ajmap1, *Aperm1, *Bjmap1, *Bperm1;
6629: PetscCount Annz1, Bnnz1, Atot1, Btot1;
6630: PetscCount *rowBegin2, *rowMid2, *rowEnd2;
6631: PetscCount *Ajmap2, *Aperm2, *Bjmap2, *Bperm2;
6632: PetscCount Annz2, Bnnz2, Atot2, Btot2;
6634: PetscCall(PetscCalloc3(m, &rowBegin1, m, &rowMid1, m, &rowEnd1));
6635: PetscCall(PetscCalloc3(m, &rowBegin2, m, &rowMid2, m, &rowEnd2));
6636: PetscCall(MatSplitEntries_Internal(mat, rem, i1, j1, perm1, rowBegin1, rowMid1, rowEnd1, &Atot1, &Aperm1, &Annz1, &Ajmap1, &Btot1, &Bperm1, &Bnnz1, &Bjmap1));
6637: PetscCall(MatSplitEntries_Internal(mat, n2, i2, j2, perm2, rowBegin2, rowMid2, rowEnd2, &Atot2, &Aperm2, &Annz2, &Ajmap2, &Btot2, &Bperm2, &Bnnz2, &Bjmap2));
6639: /* Merge local COOs with received COOs: diag with diag, offdiag with offdiag */
6640: PetscInt *Ai, *Bi;
6641: PetscInt *Aj, *Bj;
6643: PetscCall(PetscMalloc1(m + 1, &Ai));
6644: PetscCall(PetscMalloc1(m + 1, &Bi));
6645: PetscCall(PetscMalloc1(Annz1 + Annz2, &Aj)); /* Since local and remote entries might have dups, we might allocate excess memory */
6646: PetscCall(PetscMalloc1(Bnnz1 + Bnnz2, &Bj));
6648: PetscCount *Aimap1, *Bimap1, *Aimap2, *Bimap2;
6649: PetscCall(PetscMalloc1(Annz1, &Aimap1));
6650: PetscCall(PetscMalloc1(Bnnz1, &Bimap1));
6651: PetscCall(PetscMalloc1(Annz2, &Aimap2));
6652: PetscCall(PetscMalloc1(Bnnz2, &Bimap2));
6654: PetscCall(MatMergeEntries_Internal(mat, j1, j2, rowBegin1, rowMid1, rowBegin2, rowMid2, Ajmap1, Ajmap2, Aimap1, Aimap2, Ai, Aj));
6655: PetscCall(MatMergeEntries_Internal(mat, j1, j2, rowMid1, rowEnd1, rowMid2, rowEnd2, Bjmap1, Bjmap2, Bimap1, Bimap2, Bi, Bj));
6657: /* Expand Ajmap1/Bjmap1 to make them based off nonzeros in A/B, since we */
6658: /* expect nonzeros in A/B most likely have local contributing entries */
6659: PetscInt Annz = Ai[m];
6660: PetscInt Bnnz = Bi[m];
6661: PetscCount *Ajmap1_new, *Bjmap1_new;
6663: PetscCall(PetscMalloc1(Annz + 1, &Ajmap1_new));
6664: PetscCall(PetscMalloc1(Bnnz + 1, &Bjmap1_new));
6666: PetscCall(ExpandJmap_Internal(Annz1, Annz, Aimap1, Ajmap1, Ajmap1_new));
6667: PetscCall(ExpandJmap_Internal(Bnnz1, Bnnz, Bimap1, Bjmap1, Bjmap1_new));
6669: PetscCall(PetscFree(Aimap1));
6670: PetscCall(PetscFree(Ajmap1));
6671: PetscCall(PetscFree(Bimap1));
6672: PetscCall(PetscFree(Bjmap1));
6673: PetscCall(PetscFree3(rowBegin1, rowMid1, rowEnd1));
6674: PetscCall(PetscFree3(rowBegin2, rowMid2, rowEnd2));
6675: PetscCall(PetscFree(perm1));
6676: PetscCall(PetscFree3(i2, j2, perm2));
6678: Ajmap1 = Ajmap1_new;
6679: Bjmap1 = Bjmap1_new;
6681: /* Reallocate Aj, Bj once we know actual numbers of unique nonzeros in A and B */
6682: if (Annz < Annz1 + Annz2) {
6683: PetscInt *Aj_new;
6684: PetscCall(PetscMalloc1(Annz, &Aj_new));
6685: PetscCall(PetscArraycpy(Aj_new, Aj, Annz));
6686: PetscCall(PetscFree(Aj));
6687: Aj = Aj_new;
6688: }
6690: if (Bnnz < Bnnz1 + Bnnz2) {
6691: PetscInt *Bj_new;
6692: PetscCall(PetscMalloc1(Bnnz, &Bj_new));
6693: PetscCall(PetscArraycpy(Bj_new, Bj, Bnnz));
6694: PetscCall(PetscFree(Bj));
6695: Bj = Bj_new;
6696: }
6698: /* Create new submatrices for on-process and off-process coupling */
6699: PetscScalar *Aa = NULL, *Ba = NULL;
6700: MatType rtype;
6701: Mat_SeqAIJ *a, *b;
6702: if (!mat->structure_only) {
6703: PetscCall(PetscCalloc1(Annz, &Aa)); /* Zero matrix on device */
6704: PetscCall(PetscCalloc1(Bnnz, &Ba));
6705: }
6706: /* make Aj[] local, i.e, based off the start column of the diagonal portion */
6707: if (cstart) {
6708: for (k = 0; k < Annz; k++) Aj[k] -= cstart;
6709: }
6711: PetscCall(MatGetRootType_Private(mat, &rtype));
6713: MatSeqXAIJGetOptions_Private(mpiaij->A);
6714: PetscCall(MatDestroy(&mpiaij->A));
6715: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, n, Ai, Aj, Aa, &mpiaij->A));
6716: PetscCall(MatSetBlockSizesFromMats(mpiaij->A, mat, mat));
6717: MatSeqXAIJRestoreOptions_Private(mpiaij->A);
6718: PetscCall(MatSetOption(mpiaij->A, MAT_STRUCTURE_ONLY, mat->structure_only));
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);
6725: PetscCall(MatSetOption(mpiaij->B, MAT_STRUCTURE_ONLY, mat->structure_only));
6727: PetscCall(MatSetUpMultiply_MPIAIJ(mat));
6728: mat->was_assembled = PETSC_TRUE; // was_assembled in effect means the Mvctx is built; doing so avoids redundant MatSetUpMultiply_MPIAIJ
6729: mat->nonzerostate = mpiaij->A->nonzerostate + mpiaij->B->nonzerostate;
6730: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &mat->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)mat)));
6732: a = (Mat_SeqAIJ *)mpiaij->A->data;
6733: b = (Mat_SeqAIJ *)mpiaij->B->data;
6734: a->free_a = (PetscBool)!mat->structure_only;
6735: a->free_ij = PETSC_TRUE;
6736: b->free_a = (PetscBool)!mat->structure_only;
6737: b->free_ij = PETSC_TRUE;
6738: a->maxnz = a->nz;
6739: b->maxnz = b->nz;
6741: /* conversion must happen AFTER multiply setup */
6742: PetscCall(MatConvert(mpiaij->A, rtype, MAT_INPLACE_MATRIX, &mpiaij->A));
6743: PetscCall(MatConvert(mpiaij->B, rtype, MAT_INPLACE_MATRIX, &mpiaij->B));
6744: PetscCall(VecDestroy(&mpiaij->lvec));
6745: PetscCall(MatCreateVecs(mpiaij->B, &mpiaij->lvec, NULL));
6747: // Put the COO struct in a container and then attach that to the matrix
6748: PetscCall(PetscMalloc1(1, &coo));
6749: coo->n = coo_n;
6750: coo->sf = sf2;
6751: coo->sendlen = nleaves;
6752: coo->recvlen = nroots;
6753: coo->Annz = Annz;
6754: coo->Bnnz = Bnnz;
6755: coo->Annz2 = Annz2;
6756: coo->Bnnz2 = Bnnz2;
6757: coo->Atot1 = Atot1;
6758: coo->Atot2 = Atot2;
6759: coo->Btot1 = Btot1;
6760: coo->Btot2 = Btot2;
6761: coo->Ajmap1 = Ajmap1;
6762: coo->Aperm1 = Aperm1;
6763: coo->Bjmap1 = Bjmap1;
6764: coo->Bperm1 = Bperm1;
6765: coo->Aimap2 = Aimap2;
6766: coo->Ajmap2 = Ajmap2;
6767: coo->Aperm2 = Aperm2;
6768: coo->Bimap2 = Bimap2;
6769: coo->Bjmap2 = Bjmap2;
6770: coo->Bperm2 = Bperm2;
6771: coo->Cperm1 = Cperm1;
6772: // Allocate in preallocation. If not used, it has zero cost on host
6773: if (!mat->structure_only) PetscCall(PetscMalloc2(coo->sendlen, &coo->sendbuf, coo->recvlen, &coo->recvbuf));
6774: else coo->sendbuf = coo->recvbuf = NULL;
6775: coo->Anonzerostate = mpiaij->A->nonzerostate;
6776: coo->Bnonzerostate = mpiaij->B->nonzerostate;
6777: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
6778: PetscCall(PetscContainerSetPointer(container, coo));
6779: PetscCall(PetscContainerSetCtxDestroy(container, MatCOOStructDestroy_MPIAIJ));
6780: PetscCall(PetscObjectCompose((PetscObject)mat, "__PETSc_MatCOOStruct_Host", (PetscObject)container));
6781: PetscCall(PetscContainerDestroy(&container));
6782: PetscFunctionReturn(PETSC_SUCCESS);
6783: }
6785: // PetscClangLinter pragma disable: -fdoc-internal-linkage
6786: /*@
6787: MatMPIAIJCheckCOONonzeroState - Checks that the COO maps built by `MatSetPreallocationCOO()` can still be used by
6788: `MatSetValuesCOO()` on a `MATMPIAIJ` matrix
6790: Not Collective
6792: Input Parameter:
6793: . mat - the `MATMPIAIJ` matrix
6795: Level: developer
6797: Notes:
6798: The maps address the nonzeros of the diagonal and off-diagonal blocks by position, so they are only valid while the blocks keep
6799: the nonzero patterns built by `MatSetPreallocationCOO()`. Any change to them, for example a new nonzero inserted with
6800: `MatSetValues()`, would make `MatSetValuesCOO()` write values into the wrong locations, so this routine generates an error
6801: instead. It also generates an error when `MatSetPreallocationCOO()` has not been called.
6803: The check is local to each MPI process, so only the processes whose blocks changed generate the error.
6805: The `MatSetValuesCOO()` implementations of `MATMPIAIJ` and its device subtypes call this routine before using the maps.
6807: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`, `MatSeqAIJCheckCOONonzeroState()`
6808: @*/
6809: PetscErrorCode MatMPIAIJCheckCOONonzeroState(Mat mat)
6810: {
6811: Mat_MPIAIJ *mpiaij = (Mat_MPIAIJ *)mat->data;
6812: PetscContainer container;
6813: MatCOOStruct_MPIAIJ *coo;
6815: PetscFunctionBegin;
6816: PetscCall(PetscObjectQuery((PetscObject)mat, "__PETSc_MatCOOStruct_Host", (PetscObject *)&container));
6817: PetscCheck(container, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "MatSetPreallocationCOO() must be called before MatSetValuesCOO()");
6818: PetscCall(PetscContainerGetPointer(container, &coo));
6819: PetscCheck(coo->Anonzerostate == mpiaij->A->nonzerostate && coo->Bnonzerostate == mpiaij->B->nonzerostate, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "The nonzero pattern of the matrix has changed since MatSetPreallocationCOO(), for example because MatSetValues() inserted a new nonzero; call MatSetPreallocationCOO() again before MatSetValuesCOO(), or stop inserting new nonzeros with MatSetValues()");
6820: PetscFunctionReturn(PETSC_SUCCESS);
6821: }
6823: static PetscErrorCode MatSetValuesCOO_MPIAIJ(Mat mat, const PetscScalar v[], InsertMode imode)
6824: {
6825: Mat_MPIAIJ *mpiaij = (Mat_MPIAIJ *)mat->data;
6826: Mat A = mpiaij->A, B = mpiaij->B;
6827: PetscScalar *Aa, *Ba;
6828: PetscScalar *sendbuf, *recvbuf;
6829: const PetscCount *Ajmap1, *Ajmap2, *Aimap2;
6830: const PetscCount *Bjmap1, *Bjmap2, *Bimap2;
6831: const PetscCount *Aperm1, *Aperm2, *Bperm1, *Bperm2;
6832: const PetscCount *Cperm1;
6833: PetscContainer container;
6834: MatCOOStruct_MPIAIJ *coo;
6836: PetscFunctionBegin;
6837: PetscCall(MatMPIAIJCheckCOONonzeroState(mat));
6838: PetscCall(PetscObjectQuery((PetscObject)mat, "__PETSc_MatCOOStruct_Host", (PetscObject *)&container));
6839: PetscCall(PetscContainerGetPointer(container, &coo));
6840: sendbuf = coo->sendbuf;
6841: recvbuf = coo->recvbuf;
6842: Ajmap1 = coo->Ajmap1;
6843: Ajmap2 = coo->Ajmap2;
6844: Aimap2 = coo->Aimap2;
6845: Bjmap1 = coo->Bjmap1;
6846: Bjmap2 = coo->Bjmap2;
6847: Bimap2 = coo->Bimap2;
6848: Aperm1 = coo->Aperm1;
6849: Aperm2 = coo->Aperm2;
6850: Bperm1 = coo->Bperm1;
6851: Bperm2 = coo->Bperm2;
6852: Cperm1 = coo->Cperm1;
6854: PetscCall(MatSeqAIJGetArray(A, &Aa)); /* Might read and write matrix values */
6855: PetscCall(MatSeqAIJGetArray(B, &Ba));
6857: /* Pack entries to be sent to remote */
6858: for (PetscCount i = 0; i < coo->sendlen; i++) sendbuf[i] = v[Cperm1[i]];
6860: /* Send remote entries to their owner and overlap the communication with local computation */
6861: PetscCall(PetscSFReduceWithMemTypeBegin(coo->sf, MPIU_SCALAR, PETSC_MEMTYPE_HOST, sendbuf, PETSC_MEMTYPE_HOST, recvbuf, MPI_REPLACE));
6862: /* Add local entries to A and B */
6863: for (PetscCount i = 0; i < coo->Annz; i++) { /* All nonzeros in A are either zero'ed or added with a value (i.e., initialized) */
6864: PetscScalar sum = 0.0; /* Do partial summation first to improve numerical stability */
6865: for (PetscCount k = Ajmap1[i]; k < Ajmap1[i + 1]; k++) sum += v[Aperm1[k]];
6866: Aa[i] = (imode == INSERT_VALUES ? 0.0 : Aa[i]) + sum;
6867: }
6868: for (PetscCount i = 0; i < coo->Bnnz; i++) {
6869: PetscScalar sum = 0.0;
6870: for (PetscCount k = Bjmap1[i]; k < Bjmap1[i + 1]; k++) sum += v[Bperm1[k]];
6871: Ba[i] = (imode == INSERT_VALUES ? 0.0 : Ba[i]) + sum;
6872: }
6873: PetscCall(PetscSFReduceEnd(coo->sf, MPIU_SCALAR, sendbuf, recvbuf, MPI_REPLACE));
6875: /* Add received remote entries to A and B */
6876: for (PetscCount i = 0; i < coo->Annz2; i++) {
6877: for (PetscCount k = Ajmap2[i]; k < Ajmap2[i + 1]; k++) Aa[Aimap2[i]] += recvbuf[Aperm2[k]];
6878: }
6879: for (PetscCount i = 0; i < coo->Bnnz2; i++) {
6880: for (PetscCount k = Bjmap2[i]; k < Bjmap2[i + 1]; k++) Ba[Bimap2[i]] += recvbuf[Bperm2[k]];
6881: }
6882: PetscCall(MatSeqAIJRestoreArray(A, &Aa));
6883: PetscCall(MatSeqAIJRestoreArray(B, &Ba));
6884: PetscFunctionReturn(PETSC_SUCCESS);
6885: }
6887: /*MC
6888: MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices.
6890: Options Database Keys:
6891: . -mat_type mpiaij - sets the matrix type to `MATMPIAIJ` during a call to `MatSetFromOptions()`
6893: Level: beginner
6895: Notes:
6896: `MatSetValues()` may be called with a `NULL` argument for the numerical values to insert zeros at the supplied row and column indices.
6898: Call `MatSetOption(A, MAT_STRUCTURE_ONLY, PETSC_TRUE)` before preallocation or `MatSetUp()` to store only the nonzero pattern.
6899: The assembled matrix has no numerical value array. Row and column indices supplied during insertion are retained, while numerical values are ignored.
6900: Such matrices can be used for structural operations, but not for numerical operations.
6902: .seealso: [](ch_matrices), `Mat`, `MATSEQAIJ`, `MATAIJ`, `MatCreateAIJ()`
6903: M*/
6904: PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat B)
6905: {
6906: Mat_MPIAIJ *b;
6907: PetscMPIInt size;
6909: PetscFunctionBegin;
6910: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
6912: PetscCall(PetscNew(&b));
6913: B->data = (void *)b;
6914: B->ops[0] = MatOps_Values;
6915: B->assembled = PETSC_FALSE;
6916: B->insertmode = NOT_SET_VALUES;
6917: b->size = size;
6919: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)B), &b->rank));
6921: /* build cache for off array entries formed */
6922: PetscCall(MatStashCreate_Private(PetscObjectComm((PetscObject)B), 1, &B->stash));
6924: b->donotstash = PETSC_FALSE;
6925: b->colmap = NULL;
6926: b->garray = NULL;
6927: b->roworiented = PETSC_TRUE;
6929: /* stuff used for matrix vector multiply */
6930: b->lvec = NULL;
6931: b->Mvctx = NULL;
6933: /* stuff for MatGetRow() */
6934: b->rowindices = NULL;
6935: b->rowvalues = NULL;
6936: b->getrowactive = PETSC_FALSE;
6938: /* flexible pointer used in CUSPARSE classes */
6939: b->spptr = NULL;
6941: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetUseScalableIncreaseOverlap_C", MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ));
6942: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatStoreValues_C", MatStoreValues_MPIAIJ));
6943: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatRetrieveValues_C", MatRetrieveValues_MPIAIJ));
6944: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatIsTranspose_C", MatIsTranspose_MPIAIJ));
6945: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetPreallocation_C", MatMPIAIJSetPreallocation_MPIAIJ));
6946: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatResetPreallocation_C", MatResetPreallocation_MPIAIJ));
6947: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatResetHash_C", MatResetHash_MPIAIJ));
6948: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetPreallocationCSR_C", MatMPIAIJSetPreallocationCSR_MPIAIJ));
6949: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDiagonalScaleLocal_C", MatDiagonalScaleLocal_MPIAIJ));
6950: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijperm_C", MatConvert_MPIAIJ_MPIAIJPERM));
6951: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijsell_C", MatConvert_MPIAIJ_MPIAIJSELL));
6952: #if PetscDefined(HAVE_CUDA)
6953: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijcusparse_C", MatConvert_MPIAIJ_MPIAIJCUSPARSE));
6954: #endif
6955: #if PetscDefined(HAVE_HIP)
6956: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijhipsparse_C", MatConvert_MPIAIJ_MPIAIJHIPSPARSE));
6957: #endif
6958: #if PetscDefined(HAVE_KOKKOS_KERNELS)
6959: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijkokkos_C", MatConvert_MPIAIJ_MPIAIJKokkos));
6960: #endif
6961: #if PetscDefined(HAVE_MKL_SPARSE)
6962: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijmkl_C", MatConvert_MPIAIJ_MPIAIJMKL));
6963: #endif
6964: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijcrl_C", MatConvert_MPIAIJ_MPIAIJCRL));
6965: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpibaij_C", MatConvert_MPIAIJ_MPIBAIJ));
6966: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpisbaij_C", MatConvert_MPIAIJ_MPISBAIJ));
6967: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpidense_C", MatConvert_MPIAIJ_MPIDense));
6968: #if PetscDefined(HAVE_ELEMENTAL)
6969: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_elemental_C", MatConvert_MPIAIJ_Elemental));
6970: #endif
6971: #if PetscDefined(HAVE_SCALAPACK) && (PetscDefined(USE_REAL_SINGLE) || PetscDefined(USE_REAL_DOUBLE))
6972: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_scalapack_C", MatConvert_AIJ_ScaLAPACK));
6973: #endif
6974: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_is_C", MatConvert_XAIJ_IS));
6975: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpisell_C", MatConvert_MPIAIJ_MPISELL));
6976: #if PetscDefined(HAVE_HYPRE)
6977: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_hypre_C", MatConvert_AIJ_HYPRE));
6978: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_transpose_mpiaij_mpiaij_C", MatProductSetFromOptions_Transpose_AIJ_AIJ));
6979: #endif
6980: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_is_mpiaij_C", MatProductSetFromOptions_IS_XAIJ));
6981: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_mpiaij_mpiaij_C", MatProductSetFromOptions_MPIAIJ));
6982: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetPreallocationCOO_C", MatSetPreallocationCOO_MPIAIJ));
6983: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetValuesCOO_C", MatSetValuesCOO_MPIAIJ));
6984: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatGetMultPetscSF_C", MatGetMultPetscSF_MPIAIJ));
6985: PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATMPIAIJ));
6986: PetscFunctionReturn(PETSC_SUCCESS);
6987: }
6989: /*@
6990: MatCreateMPIAIJWithSplitArrays - creates a `MATMPIAIJ` matrix using arrays that contain the "diagonal"
6991: and "off-diagonal" part of the matrix in CSR format.
6993: Collective
6995: Input Parameters:
6996: + comm - MPI communicator
6997: . m - number of local rows (Cannot be `PETSC_DECIDE`)
6998: . n - This value should be the same as the local size used in creating the
6999: x vector for the matrix-vector product $y = Ax$. (or `PETSC_DECIDE` to have
7000: calculated if `N` is given) For square matrices `n` is almost always `m`.
7001: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if `m` is given)
7002: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if `n` is given)
7003: . 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
7004: . j - column indices, which must be local, i.e., based off the start column of the diagonal portion
7005: . a - matrix values
7006: . 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
7007: . oj - column indices, which must be global, representing global columns in the `MATMPIAIJ` matrix
7008: - oa - matrix values
7010: Output Parameter:
7011: . mat - the matrix
7013: Level: advanced
7015: Notes:
7016: The `i`, `j`, and `a` arrays ARE NOT copied by this routine into the internal format used by PETSc (even in Fortran). The user
7017: must free the arrays once the matrix has been destroyed and not before.
7019: The `i` and `j` indices are 0 based
7021: See `MatCreateAIJ()` for the definition of "diagonal" and "off-diagonal" portion of the matrix
7023: This sets local rows and cannot be used to set off-processor values.
7025: Use of this routine is discouraged because it is inflexible and cumbersome to use. It is extremely rare that a
7026: legacy application natively assembles into exactly this split format. The code to do so is nontrivial and does
7027: not easily support in-place reassembly. It is recommended to use MatSetValues() (or a variant thereof) because
7028: the resulting assembly is easier to implement, will work with any matrix format, and the user does not have to
7029: keep track of the underlying array. Use `MatSetOption`(A,`MAT_NO_OFF_PROC_ENTRIES`,`PETSC_TRUE`) to disable all
7030: communication if it is known that only local entries will be set.
7032: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
7033: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithArrays()`
7034: @*/
7035: 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)
7036: {
7037: Mat_MPIAIJ *maij;
7039: PetscFunctionBegin;
7040: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
7041: PetscCheck(i[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
7042: PetscCheck(oi[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "oi (row indices) must start with 0");
7043: PetscCall(MatCreate(comm, mat));
7044: PetscCall(MatSetSizes(*mat, m, n, M, N));
7045: PetscCall(MatSetType(*mat, MATMPIAIJ));
7046: maij = (Mat_MPIAIJ *)(*mat)->data;
7048: (*mat)->preallocated = PETSC_TRUE;
7050: PetscCall(PetscLayoutSetUp((*mat)->rmap));
7051: PetscCall(PetscLayoutSetUp((*mat)->cmap));
7053: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, n, i, j, a, &maij->A));
7054: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, (*mat)->cmap->N, oi, oj, oa, &maij->B));
7056: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
7057: PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY));
7058: PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY));
7059: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_FALSE));
7060: PetscCall(MatSetOption(*mat, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
7061: PetscFunctionReturn(PETSC_SUCCESS);
7062: }
7064: typedef struct {
7065: Mat *mp; /* intermediate products */
7066: PetscBool *mptmp; /* is the intermediate product temporary ? */
7067: PetscInt cp; /* number of intermediate products */
7069: /* support for MatGetBrowsOfAoCols_MPIAIJ for P_oth */
7070: PetscInt *startsj_s, *startsj_r;
7071: PetscScalar *bufa;
7072: Mat P_oth;
7074: /* may take advantage of merging product->B */
7075: Mat Bloc; /* B-local by merging diag and off-diag */
7077: /* cusparse does not have support to split between symbolic and numeric phases.
7078: When api_user is true, we don't need to update the numerical values
7079: of the temporary storage */
7080: PetscBool reusesym;
7082: /* support for COO values insertion */
7083: PetscScalar *coo_v, *coo_w; /* store on-process and off-process COO scalars, and used as MPI recv/send buffers respectively */
7084: PetscInt **own; /* own[i] points to address of on-process COO indices for Mat mp[i] */
7085: PetscInt **off; /* off[i] points to address of off-process COO indices for Mat mp[i] */
7086: PetscBool hasoffproc; /* if true, have off-process values insertion (i.e. AtB or PtAP) */
7087: PetscSF sf; /* used for non-local values insertion and memory malloc */
7088: PetscMemType mtype;
7090: /* customization */
7091: PetscBool abmerge;
7092: PetscBool P_oth_bind;
7093: } MatMatMPIAIJBACKEND;
7095: static PetscErrorCode MatProductCtxDestroy_MatMatMPIAIJBACKEND(PetscCtxRt data)
7096: {
7097: MatMatMPIAIJBACKEND *mmdata = *(MatMatMPIAIJBACKEND **)data;
7098: PetscInt i;
7100: PetscFunctionBegin;
7101: PetscCall(PetscFree2(mmdata->startsj_s, mmdata->startsj_r));
7102: PetscCall(PetscFree(mmdata->bufa));
7103: PetscCall(PetscSFFree(mmdata->sf, mmdata->mtype, mmdata->coo_v));
7104: PetscCall(PetscSFFree(mmdata->sf, mmdata->mtype, mmdata->coo_w));
7105: PetscCall(MatDestroy(&mmdata->P_oth));
7106: PetscCall(MatDestroy(&mmdata->Bloc));
7107: PetscCall(PetscSFDestroy(&mmdata->sf));
7108: for (i = 0; i < mmdata->cp; i++) PetscCall(MatDestroy(&mmdata->mp[i]));
7109: PetscCall(PetscFree2(mmdata->mp, mmdata->mptmp));
7110: PetscCall(PetscFree(mmdata->own[0]));
7111: PetscCall(PetscFree(mmdata->own));
7112: PetscCall(PetscFree(mmdata->off[0]));
7113: PetscCall(PetscFree(mmdata->off));
7114: PetscCall(PetscFree(mmdata));
7115: PetscFunctionReturn(PETSC_SUCCESS);
7116: }
7118: /* Copy selected n entries with indices in idx[] of A to v[].
7119: If idx is NULL, copy the whole data array of A to v[]
7120: */
7121: static PetscErrorCode MatSeqAIJCopySubArray(Mat A, PetscInt n, const PetscInt idx[], PetscScalar v[])
7122: {
7123: PetscErrorCode (*f)(Mat, PetscInt, const PetscInt[], PetscScalar[]);
7125: PetscFunctionBegin;
7126: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatSeqAIJCopySubArray_C", &f));
7127: if (f) PetscCall((*f)(A, n, idx, v));
7128: else {
7129: const PetscScalar *vv;
7131: PetscCall(MatSeqAIJGetArrayRead(A, &vv));
7132: if (n && idx) {
7133: PetscScalar *w = v;
7134: const PetscInt *oi = idx;
7136: for (PetscInt j = 0; j < n; j++) *w++ = vv[*oi++];
7137: } else {
7138: PetscCall(PetscArraycpy(v, vv, n));
7139: }
7140: PetscCall(MatSeqAIJRestoreArrayRead(A, &vv));
7141: }
7142: PetscFunctionReturn(PETSC_SUCCESS);
7143: }
7145: static PetscErrorCode MatProductNumeric_MPIAIJBACKEND(Mat C)
7146: {
7147: MatMatMPIAIJBACKEND *mmdata;
7148: PetscInt i, n_d, n_o;
7150: PetscFunctionBegin;
7151: MatCheckProduct(C, 1);
7152: PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data empty");
7153: mmdata = (MatMatMPIAIJBACKEND *)C->product->data;
7154: if (!mmdata->reusesym) { /* update temporary matrices */
7155: 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));
7156: if (mmdata->Bloc) PetscCall(MatMPIAIJGetLocalMatMerge(C->product->B, MAT_REUSE_MATRIX, NULL, &mmdata->Bloc));
7157: }
7158: mmdata->reusesym = PETSC_FALSE;
7160: for (i = 0; i < mmdata->cp; i++) {
7161: 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]);
7162: PetscCall((*mmdata->mp[i]->ops->productnumeric)(mmdata->mp[i]));
7163: }
7164: for (i = 0, n_d = 0, n_o = 0; i < mmdata->cp; i++) {
7165: PetscInt noff;
7167: PetscCall(PetscIntCast(mmdata->off[i + 1] - mmdata->off[i], &noff));
7168: if (mmdata->mptmp[i]) continue;
7169: if (noff) {
7170: PetscInt nown;
7172: PetscCall(PetscIntCast(mmdata->own[i + 1] - mmdata->own[i], &nown));
7173: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], noff, mmdata->off[i], mmdata->coo_w + n_o));
7174: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], nown, mmdata->own[i], mmdata->coo_v + n_d));
7175: n_o += noff;
7176: n_d += nown;
7177: } else {
7178: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mmdata->mp[i]->data;
7180: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], mm->nz, NULL, mmdata->coo_v + n_d));
7181: n_d += mm->nz;
7182: }
7183: }
7184: if (mmdata->hasoffproc) { /* offprocess insertion */
7185: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_SCALAR, mmdata->coo_w, mmdata->coo_v + n_d));
7186: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_SCALAR, mmdata->coo_w, mmdata->coo_v + n_d));
7187: }
7188: PetscCall(MatSetValuesCOO(C, mmdata->coo_v, INSERT_VALUES));
7189: PetscFunctionReturn(PETSC_SUCCESS);
7190: }
7192: /* Support for Pt * A, A * P, or Pt * A * P */
7193: #define MAX_NUMBER_INTERMEDIATE 4
7194: PetscErrorCode MatProductSymbolic_MPIAIJBACKEND(Mat C)
7195: {
7196: Mat_Product *product = C->product;
7197: Mat A, P, mp[MAX_NUMBER_INTERMEDIATE]; /* A, P and a series of intermediate matrices */
7198: Mat_MPIAIJ *a, *p;
7199: MatMatMPIAIJBACKEND *mmdata;
7200: ISLocalToGlobalMapping P_oth_l2g = NULL;
7201: IS glob = NULL;
7202: const char *prefix;
7203: char pprefix[256];
7204: const PetscInt *globidx, *P_oth_idx;
7205: PetscInt i, j, cp, m, n, M, N, *coo_i, *coo_j;
7206: PetscCount ncoo, ncoo_d, ncoo_o, ncoo_oown;
7207: PetscInt cmapt[MAX_NUMBER_INTERMEDIATE], rmapt[MAX_NUMBER_INTERMEDIATE]; /* col/row map type for each Mat in mp[]. */
7208: /* type-0: consecutive, start from 0; type-1: consecutive with */
7209: /* a base offset; type-2: sparse with a local to global map table */
7210: const PetscInt *cmapa[MAX_NUMBER_INTERMEDIATE], *rmapa[MAX_NUMBER_INTERMEDIATE]; /* col/row local to global map array (table) for type-2 map type */
7212: MatProductType ptype;
7213: PetscBool mptmp[MAX_NUMBER_INTERMEDIATE], hasoffproc = PETSC_FALSE, iscuda, iship, iskokk;
7214: PetscMPIInt size;
7216: PetscFunctionBegin;
7217: MatCheckProduct(C, 1);
7218: PetscCheck(!product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data not empty");
7219: ptype = product->type;
7220: if (product->A->symmetric == PETSC_BOOL3_TRUE && ptype == MATPRODUCT_AtB) {
7221: ptype = MATPRODUCT_AB;
7222: product->symbolic_used_the_fact_A_is_symmetric = PETSC_TRUE;
7223: }
7224: switch (ptype) {
7225: case MATPRODUCT_AB:
7226: A = product->A;
7227: P = product->B;
7228: m = A->rmap->n;
7229: n = P->cmap->n;
7230: M = A->rmap->N;
7231: N = P->cmap->N;
7232: hasoffproc = PETSC_FALSE; /* will not scatter mat product values to other processes */
7233: break;
7234: case MATPRODUCT_AtB:
7235: P = product->A;
7236: A = product->B;
7237: m = P->cmap->n;
7238: n = A->cmap->n;
7239: M = P->cmap->N;
7240: N = A->cmap->N;
7241: hasoffproc = PETSC_TRUE;
7242: break;
7243: case MATPRODUCT_PtAP:
7244: A = product->A;
7245: P = product->B;
7246: m = P->cmap->n;
7247: n = P->cmap->n;
7248: M = P->cmap->N;
7249: N = P->cmap->N;
7250: hasoffproc = PETSC_TRUE;
7251: break;
7252: default:
7253: SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Not for product type %s", MatProductTypes[ptype]);
7254: }
7255: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)C), &size));
7256: if (size == 1) hasoffproc = PETSC_FALSE;
7258: /* defaults */
7259: for (i = 0; i < MAX_NUMBER_INTERMEDIATE; i++) {
7260: mp[i] = NULL;
7261: mptmp[i] = PETSC_FALSE;
7262: rmapt[i] = -1;
7263: cmapt[i] = -1;
7264: rmapa[i] = NULL;
7265: cmapa[i] = NULL;
7266: }
7268: /* customization */
7269: PetscCall(PetscNew(&mmdata));
7270: mmdata->reusesym = product->api_user;
7271: if (ptype == MATPRODUCT_AB) {
7272: if (product->api_user) {
7273: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatMatMult", "Mat");
7274: PetscCall(PetscOptionsBool("-matmatmult_backend_mergeB", "Merge product->B local matrices", "MatMatMult", mmdata->abmerge, &mmdata->abmerge, NULL));
7275: PetscCall(PetscOptionsBool("-matmatmult_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7276: PetscOptionsEnd();
7277: } else {
7278: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_AB", "Mat");
7279: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_mergeB", "Merge product->B local matrices", "MatMatMult", mmdata->abmerge, &mmdata->abmerge, NULL));
7280: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7281: PetscOptionsEnd();
7282: }
7283: } else if (ptype == MATPRODUCT_PtAP) {
7284: if (product->api_user) {
7285: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatPtAP", "Mat");
7286: PetscCall(PetscOptionsBool("-matptap_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7287: PetscOptionsEnd();
7288: } else {
7289: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_PtAP", "Mat");
7290: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7291: PetscOptionsEnd();
7292: }
7293: }
7294: a = (Mat_MPIAIJ *)A->data;
7295: p = (Mat_MPIAIJ *)P->data;
7296: PetscCall(MatSetSizes(C, m, n, M, N));
7297: PetscCall(PetscLayoutSetUp(C->rmap));
7298: PetscCall(PetscLayoutSetUp(C->cmap));
7299: PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
7300: PetscCall(MatGetOptionsPrefix(C, &prefix));
7302: cp = 0;
7303: switch (ptype) {
7304: case MATPRODUCT_AB: /* A * P */
7305: PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &mmdata->startsj_s, &mmdata->startsj_r, &mmdata->bufa, &mmdata->P_oth));
7307: /* A_diag * P_local (merged or not) */
7308: if (mmdata->abmerge) { /* P's diagonal and off-diag blocks are merged to one matrix, then multiplied by A_diag */
7309: /* P is product->B */
7310: PetscCall(MatMPIAIJGetLocalMatMerge(P, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7311: PetscCall(MatProductCreate(a->A, mmdata->Bloc, NULL, &mp[cp]));
7312: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7313: PetscCall(MatProductSetFill(mp[cp], product->fill));
7314: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7315: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7316: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7317: mp[cp]->product->api_user = product->api_user;
7318: PetscCall(MatProductSetFromOptions(mp[cp]));
7319: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7320: PetscCall(ISGetIndices(glob, &globidx));
7321: rmapt[cp] = 1;
7322: cmapt[cp] = 2;
7323: cmapa[cp] = globidx;
7324: mptmp[cp] = PETSC_FALSE;
7325: cp++;
7326: } else { /* A_diag * P_diag and A_diag * P_off */
7327: PetscCall(MatProductCreate(a->A, p->A, NULL, &mp[cp]));
7328: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7329: PetscCall(MatProductSetFill(mp[cp], product->fill));
7330: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7331: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7332: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7333: mp[cp]->product->api_user = product->api_user;
7334: PetscCall(MatProductSetFromOptions(mp[cp]));
7335: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7336: rmapt[cp] = 1;
7337: cmapt[cp] = 1;
7338: mptmp[cp] = PETSC_FALSE;
7339: cp++;
7340: PetscCall(MatProductCreate(a->A, p->B, NULL, &mp[cp]));
7341: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7342: PetscCall(MatProductSetFill(mp[cp], product->fill));
7343: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7344: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7345: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7346: mp[cp]->product->api_user = product->api_user;
7347: PetscCall(MatProductSetFromOptions(mp[cp]));
7348: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7349: rmapt[cp] = 1;
7350: cmapt[cp] = 2;
7351: cmapa[cp] = p->garray;
7352: mptmp[cp] = PETSC_FALSE;
7353: cp++;
7354: }
7356: /* A_off * P_other */
7357: if (mmdata->P_oth) {
7358: PetscCall(MatSeqAIJCompactOutExtraColumns_SeqAIJ(mmdata->P_oth, &P_oth_l2g)); /* make P_oth use local col ids */
7359: PetscCall(ISLocalToGlobalMappingGetIndices(P_oth_l2g, &P_oth_idx));
7360: PetscCall(MatSetType(mmdata->P_oth, ((PetscObject)a->B)->type_name));
7361: PetscCall(MatBindToCPU(mmdata->P_oth, mmdata->P_oth_bind));
7362: PetscCall(MatProductCreate(a->B, mmdata->P_oth, NULL, &mp[cp]));
7363: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7364: PetscCall(MatProductSetFill(mp[cp], product->fill));
7365: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7366: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7367: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7368: mp[cp]->product->api_user = product->api_user;
7369: PetscCall(MatProductSetFromOptions(mp[cp]));
7370: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7371: rmapt[cp] = 1;
7372: cmapt[cp] = 2;
7373: cmapa[cp] = P_oth_idx;
7374: mptmp[cp] = PETSC_FALSE;
7375: cp++;
7376: }
7377: break;
7379: case MATPRODUCT_AtB: /* (P^t * A): P_diag * A_loc + P_off * A_loc */
7380: /* A is product->B */
7381: PetscCall(MatMPIAIJGetLocalMatMerge(A, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7382: if (A == P) { /* when A==P, we can take advantage of the already merged mmdata->Bloc */
7383: PetscCall(MatProductCreate(mmdata->Bloc, mmdata->Bloc, NULL, &mp[cp]));
7384: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7385: PetscCall(MatProductSetFill(mp[cp], product->fill));
7386: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7387: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7388: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7389: mp[cp]->product->api_user = product->api_user;
7390: PetscCall(MatProductSetFromOptions(mp[cp]));
7391: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7392: PetscCall(ISGetIndices(glob, &globidx));
7393: rmapt[cp] = 2;
7394: rmapa[cp] = globidx;
7395: cmapt[cp] = 2;
7396: cmapa[cp] = globidx;
7397: mptmp[cp] = PETSC_FALSE;
7398: cp++;
7399: } else {
7400: PetscCall(MatProductCreate(p->A, mmdata->Bloc, NULL, &mp[cp]));
7401: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7402: PetscCall(MatProductSetFill(mp[cp], product->fill));
7403: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7404: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7405: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7406: mp[cp]->product->api_user = product->api_user;
7407: PetscCall(MatProductSetFromOptions(mp[cp]));
7408: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7409: PetscCall(ISGetIndices(glob, &globidx));
7410: rmapt[cp] = 1;
7411: cmapt[cp] = 2;
7412: cmapa[cp] = globidx;
7413: mptmp[cp] = PETSC_FALSE;
7414: cp++;
7415: PetscCall(MatProductCreate(p->B, mmdata->Bloc, NULL, &mp[cp]));
7416: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7417: PetscCall(MatProductSetFill(mp[cp], product->fill));
7418: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7419: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7420: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7421: mp[cp]->product->api_user = product->api_user;
7422: PetscCall(MatProductSetFromOptions(mp[cp]));
7423: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7424: rmapt[cp] = 2;
7425: rmapa[cp] = p->garray;
7426: cmapt[cp] = 2;
7427: cmapa[cp] = globidx;
7428: mptmp[cp] = PETSC_FALSE;
7429: cp++;
7430: }
7431: break;
7432: case MATPRODUCT_PtAP:
7433: PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &mmdata->startsj_s, &mmdata->startsj_r, &mmdata->bufa, &mmdata->P_oth));
7434: /* P is product->B */
7435: PetscCall(MatMPIAIJGetLocalMatMerge(P, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7436: PetscCall(MatProductCreate(a->A, mmdata->Bloc, NULL, &mp[cp]));
7437: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_PtAP));
7438: PetscCall(MatProductSetFill(mp[cp], product->fill));
7439: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7440: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7441: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7442: mp[cp]->product->api_user = product->api_user;
7443: PetscCall(MatProductSetFromOptions(mp[cp]));
7444: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7445: PetscCall(ISGetIndices(glob, &globidx));
7446: rmapt[cp] = 2;
7447: rmapa[cp] = globidx;
7448: cmapt[cp] = 2;
7449: cmapa[cp] = globidx;
7450: mptmp[cp] = PETSC_FALSE;
7451: cp++;
7452: if (mmdata->P_oth) {
7453: PetscCall(MatSeqAIJCompactOutExtraColumns_SeqAIJ(mmdata->P_oth, &P_oth_l2g));
7454: PetscCall(ISLocalToGlobalMappingGetIndices(P_oth_l2g, &P_oth_idx));
7455: PetscCall(MatSetType(mmdata->P_oth, ((PetscObject)a->B)->type_name));
7456: PetscCall(MatBindToCPU(mmdata->P_oth, mmdata->P_oth_bind));
7457: PetscCall(MatProductCreate(a->B, mmdata->P_oth, NULL, &mp[cp]));
7458: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7459: PetscCall(MatProductSetFill(mp[cp], product->fill));
7460: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7461: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7462: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7463: mp[cp]->product->api_user = product->api_user;
7464: PetscCall(MatProductSetFromOptions(mp[cp]));
7465: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7466: mptmp[cp] = PETSC_TRUE;
7467: cp++;
7468: PetscCall(MatProductCreate(mmdata->Bloc, mp[1], NULL, &mp[cp]));
7469: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7470: PetscCall(MatProductSetFill(mp[cp], product->fill));
7471: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7472: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7473: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7474: mp[cp]->product->api_user = product->api_user;
7475: PetscCall(MatProductSetFromOptions(mp[cp]));
7476: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7477: rmapt[cp] = 2;
7478: rmapa[cp] = globidx;
7479: cmapt[cp] = 2;
7480: cmapa[cp] = P_oth_idx;
7481: mptmp[cp] = PETSC_FALSE;
7482: cp++;
7483: }
7484: break;
7485: default:
7486: SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Not for product type %s", MatProductTypes[ptype]);
7487: }
7488: /* sanity check */
7489: if (size > 1)
7490: 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);
7492: PetscCall(PetscMalloc2(cp, &mmdata->mp, cp, &mmdata->mptmp));
7493: for (i = 0; i < cp; i++) {
7494: mmdata->mp[i] = mp[i];
7495: mmdata->mptmp[i] = mptmp[i];
7496: }
7497: mmdata->cp = cp;
7498: C->product->data = mmdata;
7499: C->product->destroy = MatProductCtxDestroy_MatMatMPIAIJBACKEND;
7500: C->ops->productnumeric = MatProductNumeric_MPIAIJBACKEND;
7502: /* memory type */
7503: mmdata->mtype = PETSC_MEMTYPE_HOST;
7504: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iscuda, MATSEQAIJCUSPARSE, MATMPIAIJCUSPARSE, ""));
7505: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iship, MATSEQAIJHIPSPARSE, MATMPIAIJHIPSPARSE, ""));
7506: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iskokk, MATSEQAIJKOKKOS, MATMPIAIJKOKKOS, ""));
7507: if (iscuda) mmdata->mtype = PETSC_MEMTYPE_CUDA;
7508: else if (iship) mmdata->mtype = PETSC_MEMTYPE_HIP;
7509: else if (iskokk) mmdata->mtype = PETSC_MEMTYPE_KOKKOS;
7511: /* prepare coo coordinates for values insertion */
7513: /* count total nonzeros of those intermediate seqaij Mats
7514: ncoo_d: # of nonzeros of matrices that do not have offproc entries
7515: ncoo_o: # of nonzeros (of matrices that might have offproc entries) that will be inserted to remote procs
7516: ncoo_oown: # of nonzeros (of matrices that might have offproc entries) that will be inserted locally
7517: */
7518: for (cp = 0, ncoo_d = 0, ncoo_o = 0, ncoo_oown = 0; cp < mmdata->cp; cp++) {
7519: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7520: if (mptmp[cp]) continue;
7521: if (rmapt[cp] == 2 && hasoffproc) { /* the rows need to be scatter to all processes (might include self) */
7522: const PetscInt *rmap = rmapa[cp];
7523: const PetscInt mr = mp[cp]->rmap->n;
7524: const PetscInt rs = C->rmap->rstart;
7525: const PetscInt re = C->rmap->rend;
7526: const PetscInt *ii = mm->i;
7527: for (i = 0; i < mr; i++) {
7528: const PetscInt gr = rmap[i];
7529: const PetscInt nz = ii[i + 1] - ii[i];
7530: if (gr < rs || gr >= re) ncoo_o += nz; /* this row is offproc */
7531: else ncoo_oown += nz; /* this row is local */
7532: }
7533: } else ncoo_d += mm->nz;
7534: }
7536: /*
7537: ncoo: total number of nonzeros (including those inserted by remote procs) belonging to this proc
7539: ncoo = ncoo_d + ncoo_oown + ncoo2, which ncoo2 is number of nonzeros inserted to me by other procs.
7541: off[0] points to a big index array, which is shared by off[1,2,...]. Similarly, for own[0].
7543: off[p]: points to the segment for matrix mp[p], storing location of nonzeros that mp[p] will insert to others
7544: own[p]: points to the segment for matrix mp[p], storing location of nonzeros that mp[p] will insert locally
7545: so, off[p+1]-off[p] is the number of nonzeros that mp[p] will send to others.
7547: coo_i/j/v[]: [ncoo] row/col/val of nonzeros belonging to this proc.
7548: 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.
7549: */
7550: PetscCall(PetscCalloc1(mmdata->cp + 1, &mmdata->off)); /* +1 to make a csr-like data structure */
7551: PetscCall(PetscCalloc1(mmdata->cp + 1, &mmdata->own));
7553: /* gather (i,j) of nonzeros inserted by remote procs */
7554: if (hasoffproc) {
7555: PetscSF msf;
7556: PetscInt ncoo2, *coo_i2, *coo_j2;
7558: PetscCall(PetscMalloc1(ncoo_o, &mmdata->off[0]));
7559: PetscCall(PetscMalloc1(ncoo_oown, &mmdata->own[0]));
7560: PetscCall(PetscMalloc2(ncoo_o, &coo_i, ncoo_o, &coo_j)); /* to collect (i,j) of entries to be sent to others */
7562: for (cp = 0, ncoo_o = 0; cp < mmdata->cp; cp++) {
7563: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7564: PetscInt *idxoff = mmdata->off[cp];
7565: PetscInt *idxown = mmdata->own[cp];
7566: if (!mptmp[cp] && rmapt[cp] == 2) { /* row map is sparse */
7567: const PetscInt *rmap = rmapa[cp];
7568: const PetscInt *cmap = cmapa[cp];
7569: const PetscInt *ii = mm->i;
7570: PetscInt *coi = coo_i + ncoo_o;
7571: PetscInt *coj = coo_j + ncoo_o;
7572: const PetscInt mr = mp[cp]->rmap->n;
7573: const PetscInt rs = C->rmap->rstart;
7574: const PetscInt re = C->rmap->rend;
7575: const PetscInt cs = C->cmap->rstart;
7576: for (i = 0; i < mr; i++) {
7577: const PetscInt *jj = mm->j + ii[i];
7578: const PetscInt gr = rmap[i];
7579: const PetscInt nz = ii[i + 1] - ii[i];
7580: if (gr < rs || gr >= re) { /* this is an offproc row */
7581: for (j = ii[i]; j < ii[i + 1]; j++) {
7582: *coi++ = gr;
7583: *idxoff++ = j;
7584: }
7585: if (!cmapt[cp]) { /* already global */
7586: for (j = 0; j < nz; j++) *coj++ = jj[j];
7587: } else if (cmapt[cp] == 1) { /* local to global for owned columns of C */
7588: for (j = 0; j < nz; j++) *coj++ = jj[j] + cs;
7589: } else { /* offdiag */
7590: for (j = 0; j < nz; j++) *coj++ = cmap[jj[j]];
7591: }
7592: ncoo_o += nz;
7593: } else { /* this is a local row */
7594: for (j = ii[i]; j < ii[i + 1]; j++) *idxown++ = j;
7595: }
7596: }
7597: }
7598: mmdata->off[cp + 1] = idxoff;
7599: mmdata->own[cp + 1] = idxown;
7600: }
7602: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)C), &mmdata->sf));
7603: PetscInt incoo_o;
7604: PetscCall(PetscIntCast(ncoo_o, &incoo_o));
7605: PetscCall(PetscSFSetGraphLayout(mmdata->sf, C->rmap, incoo_o /*nleaves*/, NULL /*ilocal*/, PETSC_OWN_POINTER, coo_i));
7606: PetscCall(PetscSFGetMultiSF(mmdata->sf, &msf));
7607: PetscCall(PetscSFGetGraph(msf, &ncoo2 /*nroots*/, NULL, NULL, NULL));
7608: ncoo = ncoo_d + ncoo_oown + ncoo2;
7609: PetscCall(PetscMalloc2(ncoo, &coo_i2, ncoo, &coo_j2));
7610: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_INT, coo_i, coo_i2 + ncoo_d + ncoo_oown)); /* put (i,j) of remote nonzeros at back */
7611: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_INT, coo_i, coo_i2 + ncoo_d + ncoo_oown));
7612: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_INT, coo_j, coo_j2 + ncoo_d + ncoo_oown));
7613: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_INT, coo_j, coo_j2 + ncoo_d + ncoo_oown));
7614: PetscCall(PetscFree2(coo_i, coo_j));
7615: /* allocate MPI send buffer to collect nonzero values to be sent to remote procs */
7616: PetscCall(PetscSFMalloc(mmdata->sf, mmdata->mtype, ncoo_o * sizeof(PetscScalar), (void **)&mmdata->coo_w));
7617: coo_i = coo_i2;
7618: coo_j = coo_j2;
7619: } else { /* no offproc values insertion */
7620: ncoo = ncoo_d;
7621: PetscCall(PetscMalloc2(ncoo, &coo_i, ncoo, &coo_j));
7623: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)C), &mmdata->sf));
7624: PetscCall(PetscSFSetGraph(mmdata->sf, 0, 0, NULL, PETSC_OWN_POINTER, NULL, PETSC_OWN_POINTER));
7625: PetscCall(PetscSFSetUp(mmdata->sf));
7626: }
7627: mmdata->hasoffproc = hasoffproc;
7629: /* gather (i,j) of nonzeros inserted locally */
7630: for (cp = 0, ncoo_d = 0; cp < mmdata->cp; cp++) {
7631: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7632: PetscInt *coi = coo_i + ncoo_d;
7633: PetscInt *coj = coo_j + ncoo_d;
7634: const PetscInt *jj = mm->j;
7635: const PetscInt *ii = mm->i;
7636: const PetscInt *cmap = cmapa[cp];
7637: const PetscInt *rmap = rmapa[cp];
7638: const PetscInt mr = mp[cp]->rmap->n;
7639: const PetscInt rs = C->rmap->rstart;
7640: const PetscInt re = C->rmap->rend;
7641: const PetscInt cs = C->cmap->rstart;
7643: if (mptmp[cp]) continue;
7644: if (rmapt[cp] == 1) { /* consecutive rows */
7645: /* fill coo_i */
7646: for (i = 0; i < mr; i++) {
7647: const PetscInt gr = i + rs;
7648: for (j = ii[i]; j < ii[i + 1]; j++) coi[j] = gr;
7649: }
7650: /* fill coo_j */
7651: if (!cmapt[cp]) { /* type-0, already global */
7652: PetscCall(PetscArraycpy(coj, jj, mm->nz));
7653: } else if (cmapt[cp] == 1) { /* type-1, local to global for consecutive columns of C */
7654: for (j = 0; j < mm->nz; j++) coj[j] = jj[j] + cs; /* lid + col start */
7655: } else { /* type-2, local to global for sparse columns */
7656: for (j = 0; j < mm->nz; j++) coj[j] = cmap[jj[j]];
7657: }
7658: ncoo_d += mm->nz;
7659: } else if (rmapt[cp] == 2) { /* sparse rows */
7660: for (i = 0; i < mr; i++) {
7661: const PetscInt *jj = mm->j + ii[i];
7662: const PetscInt gr = rmap[i];
7663: const PetscInt nz = ii[i + 1] - ii[i];
7664: if (gr >= rs && gr < re) { /* local rows */
7665: for (j = ii[i]; j < ii[i + 1]; j++) *coi++ = gr;
7666: if (!cmapt[cp]) { /* type-0, already global */
7667: for (j = 0; j < nz; j++) *coj++ = jj[j];
7668: } else if (cmapt[cp] == 1) { /* local to global for owned columns of C */
7669: for (j = 0; j < nz; j++) *coj++ = jj[j] + cs;
7670: } else { /* type-2, local to global for sparse columns */
7671: for (j = 0; j < nz; j++) *coj++ = cmap[jj[j]];
7672: }
7673: ncoo_d += nz;
7674: }
7675: }
7676: }
7677: }
7678: if (glob) PetscCall(ISRestoreIndices(glob, &globidx));
7679: PetscCall(ISDestroy(&glob));
7680: if (P_oth_l2g) PetscCall(ISLocalToGlobalMappingRestoreIndices(P_oth_l2g, &P_oth_idx));
7681: PetscCall(ISLocalToGlobalMappingDestroy(&P_oth_l2g));
7682: /* allocate an array to store all nonzeros (inserted locally or remotely) belonging to this proc */
7683: PetscCall(PetscSFMalloc(mmdata->sf, mmdata->mtype, ncoo * sizeof(PetscScalar), (void **)&mmdata->coo_v));
7685: /* set block sizes */
7686: A = product->A;
7687: P = product->B;
7688: switch (ptype) {
7689: case MATPRODUCT_PtAP:
7690: PetscCall(MatSetBlockSizes(C, P->cmap->bs, P->cmap->bs));
7691: break;
7692: case MATPRODUCT_RARt:
7693: PetscCall(MatSetBlockSizes(C, P->rmap->bs, P->rmap->bs));
7694: break;
7695: case MATPRODUCT_ABC:
7696: PetscCall(MatSetBlockSizesFromMats(C, A, product->C));
7697: break;
7698: case MATPRODUCT_AB:
7699: PetscCall(MatSetBlockSizesFromMats(C, A, P));
7700: break;
7701: case MATPRODUCT_AtB:
7702: PetscCall(MatSetBlockSizes(C, A->cmap->bs, P->cmap->bs));
7703: break;
7704: case MATPRODUCT_ABt:
7705: PetscCall(MatSetBlockSizes(C, A->rmap->bs, P->rmap->bs));
7706: break;
7707: default:
7708: SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Not for ProductType %s", MatProductTypes[ptype]);
7709: }
7711: /* preallocate with COO data */
7712: PetscCall(MatSetPreallocationCOO(C, ncoo, coo_i, coo_j));
7713: PetscCall(PetscFree2(coo_i, coo_j));
7714: PetscFunctionReturn(PETSC_SUCCESS);
7715: }
7717: PetscErrorCode MatProductSetFromOptions_MPIAIJBACKEND(Mat mat)
7718: {
7719: Mat_Product *product = mat->product;
7720: #if PetscDefined(HAVE_DEVICE)
7721: PetscBool match = PETSC_FALSE;
7722: PetscBool usecpu = PETSC_FALSE;
7723: #else
7724: PetscBool match = PETSC_TRUE;
7725: #endif
7727: PetscFunctionBegin;
7728: MatCheckProduct(mat, 1);
7729: #if PetscDefined(HAVE_DEVICE)
7730: if (!product->A->boundtocpu && !product->B->boundtocpu) PetscCall(PetscObjectTypeCompare((PetscObject)product->B, ((PetscObject)product->A)->type_name, &match));
7731: if (match) { /* we can always fallback to the CPU if requested */
7732: switch (product->type) {
7733: case MATPRODUCT_AB:
7734: if (product->api_user) {
7735: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatMatMult", "Mat");
7736: PetscCall(PetscOptionsBool("-matmatmult_backend_cpu", "Use CPU code", "MatMatMult", usecpu, &usecpu, NULL));
7737: PetscOptionsEnd();
7738: } else {
7739: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_AB", "Mat");
7740: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatMatMult", usecpu, &usecpu, NULL));
7741: PetscOptionsEnd();
7742: }
7743: break;
7744: case MATPRODUCT_AtB:
7745: if (product->api_user) {
7746: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatTransposeMatMult", "Mat");
7747: PetscCall(PetscOptionsBool("-mattransposematmult_backend_cpu", "Use CPU code", "MatTransposeMatMult", usecpu, &usecpu, NULL));
7748: PetscOptionsEnd();
7749: } else {
7750: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_AtB", "Mat");
7751: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatTransposeMatMult", usecpu, &usecpu, NULL));
7752: PetscOptionsEnd();
7753: }
7754: break;
7755: case MATPRODUCT_PtAP:
7756: if (product->api_user) {
7757: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatPtAP", "Mat");
7758: PetscCall(PetscOptionsBool("-matptap_backend_cpu", "Use CPU code", "MatPtAP", usecpu, &usecpu, NULL));
7759: PetscOptionsEnd();
7760: } else {
7761: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_PtAP", "Mat");
7762: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatPtAP", usecpu, &usecpu, NULL));
7763: PetscOptionsEnd();
7764: }
7765: break;
7766: default:
7767: break;
7768: }
7769: match = (PetscBool)!usecpu;
7770: }
7771: #endif
7772: if (match) {
7773: switch (product->type) {
7774: case MATPRODUCT_AB:
7775: case MATPRODUCT_AtB:
7776: case MATPRODUCT_PtAP:
7777: mat->ops->productsymbolic = MatProductSymbolic_MPIAIJBACKEND;
7778: break;
7779: default:
7780: break;
7781: }
7782: }
7783: /* fallback to MPIAIJ ops */
7784: if (!mat->ops->productsymbolic) PetscCall(MatProductSetFromOptions_MPIAIJ(mat));
7785: PetscFunctionReturn(PETSC_SUCCESS);
7786: }
7788: /*
7789: Produces a set of block column indices of the matrix row, one for each block represented in the original row
7791: n - the number of block indices in cc[]
7792: cc - the block indices (must be large enough to contain the indices)
7793: */
7794: static inline PetscErrorCode MatCollapseRow(Mat Amat, PetscInt row, PetscInt bs, PetscInt *n, PetscInt *cc)
7795: {
7796: PetscInt cnt = -1, nidx, j;
7797: const PetscInt *idx;
7799: PetscFunctionBegin;
7800: PetscCall(MatGetRow(Amat, row, &nidx, &idx, NULL));
7801: if (nidx) {
7802: cnt = 0;
7803: cc[cnt] = idx[0] / bs;
7804: for (j = 1; j < nidx; j++) {
7805: if (cc[cnt] < idx[j] / bs) cc[++cnt] = idx[j] / bs;
7806: }
7807: }
7808: PetscCall(MatRestoreRow(Amat, row, &nidx, &idx, NULL));
7809: *n = cnt + 1;
7810: PetscFunctionReturn(PETSC_SUCCESS);
7811: }
7813: /*
7814: Produces a set of block column indices of the matrix block row, one for each block represented in the original set of rows
7816: ncollapsed - the number of block indices
7817: collapsed - the block indices (must be large enough to contain the indices)
7818: */
7819: static inline PetscErrorCode MatCollapseRows(Mat Amat, PetscInt start, PetscInt bs, PetscInt *w0, PetscInt *w1, PetscInt *w2, PetscInt *ncollapsed, PetscInt **collapsed)
7820: {
7821: PetscInt i, nprev, *cprev = w0, ncur = 0, *ccur = w1, *merged = w2, *cprevtmp;
7823: PetscFunctionBegin;
7824: PetscCall(MatCollapseRow(Amat, start, bs, &nprev, cprev));
7825: for (i = start + 1; i < start + bs; i++) {
7826: PetscCall(MatCollapseRow(Amat, i, bs, &ncur, ccur));
7827: PetscCall(PetscMergeIntArray(nprev, cprev, ncur, ccur, &nprev, &merged));
7828: cprevtmp = cprev;
7829: cprev = merged;
7830: merged = cprevtmp;
7831: }
7832: *ncollapsed = nprev;
7833: if (collapsed) *collapsed = cprev;
7834: PetscFunctionReturn(PETSC_SUCCESS);
7835: }
7837: PETSC_INTERN PetscErrorCode MatCreateGraph_Simple_AIJ(Mat Amat, PetscBool symmetrize, PetscBool scale, PetscReal filter, PetscInt index_size, PetscInt index[], Mat *a_Gmat)
7838: {
7839: PetscInt Istart, Iend, Ii, jj, kk, ncols, nloc, NN, MM, bs;
7840: MPI_Comm comm;
7841: Mat Gmat;
7842: PetscBool ismpiaij, isseqaij;
7843: Mat a, b, c;
7844: MatType jtype;
7846: PetscFunctionBegin;
7847: PetscCall(PetscObjectGetComm((PetscObject)Amat, &comm));
7848: PetscCall(MatGetOwnershipRange(Amat, &Istart, &Iend));
7849: PetscCall(MatGetSize(Amat, &MM, &NN));
7850: PetscCall(MatGetBlockSize(Amat, &bs));
7851: nloc = (Iend - Istart) / bs;
7853: PetscCall(PetscObjectBaseTypeCompare((PetscObject)Amat, MATSEQAIJ, &isseqaij));
7854: PetscCall(PetscObjectBaseTypeCompare((PetscObject)Amat, MATMPIAIJ, &ismpiaij));
7855: PetscCheck(isseqaij || ismpiaij, comm, PETSC_ERR_USER, "Require (MPI)AIJ matrix type");
7857: /* TODO GPU: these calls are potentially expensive if matrices are large and we want to use the GPU */
7858: /* A solution consists in providing a new API, MatAIJGetCollapsedAIJ, and each class can provide a fast
7859: implementation */
7860: if (bs > 1) {
7861: PetscCall(MatGetType(Amat, &jtype));
7862: PetscCall(MatCreate(comm, &Gmat));
7863: PetscCall(MatSetType(Gmat, jtype));
7864: PetscCall(MatSetSizes(Gmat, nloc, nloc, PETSC_DETERMINE, PETSC_DETERMINE));
7865: PetscCall(MatSetBlockSizes(Gmat, 1, 1));
7866: if (isseqaij || ((Mat_MPIAIJ *)Amat->data)->garray) {
7867: PetscInt *d_nnz, *o_nnz;
7868: MatScalar *aa, val, *AA;
7869: PetscInt *aj, *ai, *AJ, nc, nmax = 0;
7871: if (isseqaij) {
7872: a = Amat;
7873: b = NULL;
7874: } else {
7875: Mat_MPIAIJ *d = (Mat_MPIAIJ *)Amat->data;
7876: a = d->A;
7877: b = d->B;
7878: }
7879: PetscCall(PetscInfo(Amat, "New bs>1 Graph. nloc=%" PetscInt_FMT "\n", nloc));
7880: PetscCall(PetscMalloc2(nloc, &d_nnz, (isseqaij ? 0 : nloc), &o_nnz));
7881: for (c = a, kk = 0; c && kk < 2; c = b, kk++) {
7882: PetscInt *nnz = (c == a) ? d_nnz : o_nnz;
7883: const PetscInt *cols1, *cols2;
7885: for (PetscInt brow = 0, nc1, nc2, ok = 1; brow < nloc * bs; brow += bs) { // block rows
7886: PetscCall(MatGetRow(c, brow, &nc2, &cols2, NULL));
7887: nnz[brow / bs] = nc2 / bs;
7888: if (nc2 % bs) ok = 0;
7889: if (nnz[brow / bs] > nmax) nmax = nnz[brow / bs];
7890: for (PetscInt ii = 1; ii < bs; ii++) { // check for non-dense blocks
7891: PetscCall(MatGetRow(c, brow + ii, &nc1, &cols1, NULL));
7892: if (nc1 != nc2) ok = 0;
7893: else {
7894: for (PetscInt jj = 0; jj < nc1 && ok == 1; jj++) {
7895: if (cols1[jj] != cols2[jj]) ok = 0;
7896: if (cols1[jj] % bs != jj % bs) ok = 0;
7897: }
7898: }
7899: PetscCall(MatRestoreRow(c, brow + ii, &nc1, &cols1, NULL));
7900: }
7901: PetscCall(MatRestoreRow(c, brow, &nc2, &cols2, NULL));
7902: if (!ok) {
7903: PetscCall(PetscFree2(d_nnz, o_nnz));
7904: PetscCall(PetscInfo(Amat, "Found sparse blocks - revert to slow method\n"));
7905: goto old_bs;
7906: }
7907: }
7908: }
7909: PetscCall(MatSeqAIJSetPreallocation(Gmat, 0, d_nnz));
7910: PetscCall(MatMPIAIJSetPreallocation(Gmat, 0, d_nnz, 0, o_nnz));
7911: PetscCall(PetscFree2(d_nnz, o_nnz));
7912: PetscCall(PetscMalloc2(nmax, &AA, nmax, &AJ));
7913: // diag
7914: for (PetscInt brow = 0, n, grow; brow < nloc * bs; brow += bs) { // block rows
7915: Mat_SeqAIJ *aseq = (Mat_SeqAIJ *)a->data;
7917: ai = aseq->i;
7918: n = ai[brow + 1] - ai[brow];
7919: aj = aseq->j + ai[brow];
7920: for (PetscInt k = 0; k < n; k += bs) { // block columns
7921: AJ[k / bs] = aj[k] / bs + Istart / bs; // diag starts at (Istart,Istart)
7922: val = 0;
7923: if (index_size == 0) {
7924: for (PetscInt ii = 0; ii < bs; ii++) { // rows in block
7925: aa = aseq->a + ai[brow + ii] + k;
7926: for (PetscInt jj = 0; jj < bs; jj++) { // columns in block
7927: val += PetscAbs(PetscRealPart(aa[jj])); // a sort of norm
7928: }
7929: }
7930: } else { // use (index,index) value if provided
7931: for (PetscInt iii = 0; iii < index_size; iii++) { // rows in block
7932: PetscInt ii = index[iii];
7933: aa = aseq->a + ai[brow + ii] + k;
7934: for (PetscInt jjj = 0; jjj < index_size; jjj++) { // columns in block
7935: PetscInt jj = index[jjj];
7936: val += PetscAbs(PetscRealPart(aa[jj]));
7937: }
7938: }
7939: }
7940: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs (%" PetscInt_FMT ") >= nmax (%" PetscInt_FMT ")", k / bs, nmax);
7941: AA[k / bs] = val;
7942: }
7943: grow = Istart / bs + brow / bs;
7944: PetscCall(MatSetValues(Gmat, 1, &grow, n / bs, AJ, AA, ADD_VALUES));
7945: }
7946: // off-diag
7947: if (ismpiaij) {
7948: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)Amat->data;
7949: const PetscScalar *vals;
7950: const PetscInt *cols, *garray = aij->garray;
7952: PetscCheck(garray, PETSC_COMM_SELF, PETSC_ERR_USER, "No garray ?");
7953: for (PetscInt brow = 0, grow; brow < nloc * bs; brow += bs) { // block rows
7954: PetscCall(MatGetRow(b, brow, &ncols, &cols, NULL));
7955: for (PetscInt k = 0, cidx = 0; k < ncols; k += bs, cidx++) {
7956: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs >= nmax");
7957: AA[k / bs] = 0;
7958: AJ[cidx] = garray[cols[k]] / bs;
7959: }
7960: nc = ncols / bs;
7961: PetscCall(MatRestoreRow(b, brow, &ncols, &cols, NULL));
7962: if (index_size == 0) {
7963: for (PetscInt ii = 0; ii < bs; ii++) { // rows in block
7964: PetscCall(MatGetRow(b, brow + ii, &ncols, &cols, &vals));
7965: for (PetscInt k = 0; k < ncols; k += bs) {
7966: for (PetscInt jj = 0; jj < bs; jj++) { // cols in block
7967: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs (%" PetscInt_FMT ") >= nmax (%" PetscInt_FMT ")", k / bs, nmax);
7968: AA[k / bs] += PetscAbs(PetscRealPart(vals[k + jj]));
7969: }
7970: }
7971: PetscCall(MatRestoreRow(b, brow + ii, &ncols, &cols, &vals));
7972: }
7973: } else { // use (index,index) value if provided
7974: for (PetscInt iii = 0; iii < index_size; iii++) { // rows in block
7975: PetscInt ii = index[iii];
7976: PetscCall(MatGetRow(b, brow + ii, &ncols, &cols, &vals));
7977: for (PetscInt k = 0; k < ncols; k += bs) {
7978: for (PetscInt jjj = 0; jjj < index_size; jjj++) { // cols in block
7979: PetscInt jj = index[jjj];
7980: AA[k / bs] += PetscAbs(PetscRealPart(vals[k + jj]));
7981: }
7982: }
7983: PetscCall(MatRestoreRow(b, brow + ii, &ncols, &cols, &vals));
7984: }
7985: }
7986: grow = Istart / bs + brow / bs;
7987: PetscCall(MatSetValues(Gmat, 1, &grow, nc, AJ, AA, ADD_VALUES));
7988: }
7989: }
7990: PetscCall(MatAssemblyBegin(Gmat, MAT_FINAL_ASSEMBLY));
7991: PetscCall(MatAssemblyEnd(Gmat, MAT_FINAL_ASSEMBLY));
7992: PetscCall(PetscFree2(AA, AJ));
7993: } else {
7994: const PetscScalar *vals;
7995: const PetscInt *idx;
7996: PetscInt *d_nnz, *o_nnz, *w0, *w1, *w2;
7997: old_bs:
7998: /*
7999: Determine the preallocation needed for the scalar matrix derived from the vector matrix.
8000: */
8001: PetscCall(PetscInfo(Amat, "OLD bs>1 CreateGraph\n"));
8002: PetscCall(PetscMalloc2(nloc, &d_nnz, (isseqaij ? 0 : nloc), &o_nnz));
8003: if (isseqaij) {
8004: PetscInt max_d_nnz;
8006: /*
8007: Determine exact preallocation count for (sequential) scalar matrix
8008: */
8009: PetscCall(MatSeqAIJGetMaxRowNonzeros(Amat, &max_d_nnz));
8010: max_d_nnz = PetscMin(nloc, bs * max_d_nnz);
8011: PetscCall(PetscMalloc3(max_d_nnz, &w0, max_d_nnz, &w1, max_d_nnz, &w2));
8012: for (Ii = 0, jj = 0; Ii < Iend; Ii += bs, jj++) PetscCall(MatCollapseRows(Amat, Ii, bs, w0, w1, w2, &d_nnz[jj], NULL));
8013: PetscCall(PetscFree3(w0, w1, w2));
8014: } else if (ismpiaij) {
8015: Mat Daij, Oaij;
8016: const PetscInt *garray;
8017: PetscInt max_d_nnz;
8019: PetscCall(MatMPIAIJGetSeqAIJ(Amat, &Daij, &Oaij, &garray));
8020: /*
8021: Determine exact preallocation count for diagonal block portion of scalar matrix
8022: */
8023: PetscCall(MatSeqAIJGetMaxRowNonzeros(Daij, &max_d_nnz));
8024: max_d_nnz = PetscMin(nloc, bs * max_d_nnz);
8025: PetscCall(PetscMalloc3(max_d_nnz, &w0, max_d_nnz, &w1, max_d_nnz, &w2));
8026: for (Ii = 0, jj = 0; Ii < Iend - Istart; Ii += bs, jj++) PetscCall(MatCollapseRows(Daij, Ii, bs, w0, w1, w2, &d_nnz[jj], NULL));
8027: PetscCall(PetscFree3(w0, w1, w2));
8028: /*
8029: Over estimate (usually grossly over), preallocation count for off-diagonal portion of scalar matrix
8030: */
8031: for (Ii = 0, jj = 0; Ii < Iend - Istart; Ii += bs, jj++) {
8032: o_nnz[jj] = 0;
8033: for (kk = 0; kk < bs; kk++) { /* rows that get collapsed to a single row */
8034: PetscCall(MatGetRow(Oaij, Ii + kk, &ncols, NULL, NULL));
8035: o_nnz[jj] += ncols;
8036: PetscCall(MatRestoreRow(Oaij, Ii + kk, &ncols, NULL, NULL));
8037: }
8038: if (o_nnz[jj] > (NN / bs - nloc)) o_nnz[jj] = NN / bs - nloc;
8039: }
8040: } else SETERRQ(comm, PETSC_ERR_USER, "Require AIJ matrix type");
8041: /* get scalar copy (norms) of matrix */
8042: PetscCall(MatSeqAIJSetPreallocation(Gmat, 0, d_nnz));
8043: PetscCall(MatMPIAIJSetPreallocation(Gmat, 0, d_nnz, 0, o_nnz));
8044: PetscCall(PetscFree2(d_nnz, o_nnz));
8045: for (Ii = Istart; Ii < Iend; Ii++) {
8046: PetscInt dest_row = Ii / bs;
8048: PetscCall(MatGetRow(Amat, Ii, &ncols, &idx, &vals));
8049: for (jj = 0; jj < ncols; jj++) {
8050: PetscInt dest_col = idx[jj] / bs;
8051: PetscScalar sv = PetscAbs(PetscRealPart(vals[jj]));
8053: PetscCall(MatSetValues(Gmat, 1, &dest_row, 1, &dest_col, &sv, ADD_VALUES));
8054: }
8055: PetscCall(MatRestoreRow(Amat, Ii, &ncols, &idx, &vals));
8056: }
8057: PetscCall(MatAssemblyBegin(Gmat, MAT_FINAL_ASSEMBLY));
8058: PetscCall(MatAssemblyEnd(Gmat, MAT_FINAL_ASSEMBLY));
8059: }
8060: } else {
8061: if (symmetrize || filter >= 0 || scale) PetscCall(MatDuplicate(Amat, MAT_COPY_VALUES, &Gmat));
8062: else {
8063: Gmat = Amat;
8064: PetscCall(PetscObjectReference((PetscObject)Gmat));
8065: }
8066: if (isseqaij) {
8067: a = Gmat;
8068: b = NULL;
8069: } else {
8070: Mat_MPIAIJ *d = (Mat_MPIAIJ *)Gmat->data;
8071: a = d->A;
8072: b = d->B;
8073: }
8074: if (filter >= 0 || scale) {
8075: /* take absolute value of each entry */
8076: for (c = a, kk = 0; c && kk < 2; c = b, kk++) {
8077: MatInfo info;
8078: PetscScalar *avals;
8080: PetscCall(MatGetInfo(c, MAT_LOCAL, &info));
8081: PetscCall(MatSeqAIJGetArray(c, &avals));
8082: for (int jj = 0; jj < info.nz_used; jj++) avals[jj] = PetscAbsScalar(avals[jj]);
8083: PetscCall(MatSeqAIJRestoreArray(c, &avals));
8084: }
8085: }
8086: }
8087: if (symmetrize) {
8088: PetscBool isset, issym;
8090: PetscCall(MatIsSymmetricKnown(Amat, &isset, &issym));
8091: if (!isset || !issym) {
8092: Mat matTrans;
8094: PetscCall(MatTranspose(Gmat, MAT_INITIAL_MATRIX, &matTrans));
8095: PetscCall(MatAXPY(Gmat, 1.0, matTrans, Gmat->structurally_symmetric == PETSC_BOOL3_TRUE ? SAME_NONZERO_PATTERN : DIFFERENT_NONZERO_PATTERN));
8096: PetscCall(MatDestroy(&matTrans));
8097: }
8098: PetscCall(MatSetOption(Gmat, MAT_SYMMETRIC, PETSC_TRUE));
8099: } else if (Amat != Gmat) PetscCall(MatPropagateSymmetryOptions(Amat, Gmat));
8100: if (scale) {
8101: /* scale c for all diagonal values = 1 or -1 */
8102: Vec diag;
8104: PetscCall(MatCreateVecs(Gmat, &diag, NULL));
8105: PetscCall(MatGetDiagonal(Gmat, diag));
8106: PetscCall(VecReciprocal(diag));
8107: PetscCall(VecSqrtAbs(diag));
8108: PetscCall(MatDiagonalScale(Gmat, diag, diag));
8109: PetscCall(VecDestroy(&diag));
8110: }
8111: PetscCall(MatViewFromOptions(Gmat, NULL, "-mat_graph_view"));
8112: if (filter >= 0) {
8113: PetscCall(MatFilter(Gmat, filter, PETSC_TRUE, PETSC_TRUE));
8114: PetscCall(MatViewFromOptions(Gmat, NULL, "-mat_filter_graph_view"));
8115: }
8116: *a_Gmat = Gmat;
8117: PetscFunctionReturn(PETSC_SUCCESS);
8118: }
8120: PETSC_INTERN PetscErrorCode MatGetCurrentMemType_MPIAIJ(Mat A, PetscMemType *memtype)
8121: {
8122: Mat_MPIAIJ *mpiaij = (Mat_MPIAIJ *)A->data;
8123: PetscMemType mD = PETSC_MEMTYPE_HOST, mO = PETSC_MEMTYPE_HOST;
8125: PetscFunctionBegin;
8126: if (mpiaij->A) PetscCall(MatGetCurrentMemType(mpiaij->A, &mD));
8127: if (mpiaij->B) PetscCall(MatGetCurrentMemType(mpiaij->B, &mO));
8128: *memtype = (mD == mO) ? mD : PETSC_MEMTYPE_HOST;
8129: PetscFunctionReturn(PETSC_SUCCESS);
8130: }
8132: /*
8133: Special version for direct calls from Fortran
8134: */
8136: /* Change these macros so can be used in void function */
8137: /* Identical to PetscCallVoid, except it assigns to *_ierr */
8138: #undef PetscCall
8139: #define PetscCall(...) \
8140: do { \
8141: PetscErrorCode ierr_msv_mpiaij = __VA_ARGS__; \
8142: if (PetscUnlikely(ierr_msv_mpiaij)) { \
8143: *_ierr = PetscError(PETSC_COMM_SELF, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr_msv_mpiaij, PETSC_ERROR_REPEAT, " "); \
8144: return; \
8145: } \
8146: } while (0)
8148: #undef SETERRQ
8149: #define SETERRQ(comm, ierr, ...) \
8150: do { \
8151: *_ierr = PetscError(comm, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr, PETSC_ERROR_INITIAL, __VA_ARGS__); \
8152: return; \
8153: } while (0)
8155: #if PetscDefined(HAVE_FORTRAN_CAPS)
8156: #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ
8157: #elif !PetscDefined(HAVE_FORTRAN_UNDERSCORE)
8158: #define matsetvaluesmpiaij_ matsetvaluesmpiaij
8159: #else
8160: #endif
8161: PETSC_EXTERN void matsetvaluesmpiaij_(Mat *mmat, PetscInt *mm, const PetscInt im[], PetscInt *mn, const PetscInt in[], const PetscScalar v[], InsertMode *maddv, PetscErrorCode *_ierr)
8162: {
8163: Mat mat = *mmat;
8164: PetscInt m = *mm, n = *mn;
8165: InsertMode addv = *maddv;
8166: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
8167: PetscScalar value;
8169: MatCheckPreallocated(mat, 1);
8170: if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv;
8171: else PetscCheck(mat->insertmode == addv, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Cannot mix add values and insert values");
8172: {
8173: PetscInt i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend;
8174: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend, row, col;
8175: PetscBool roworiented = aij->roworiented;
8177: /* Some Variables required in the macro */
8178: Mat A = aij->A;
8179: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
8180: PetscInt *aimax = a->imax, *ai = a->i, *ailen = a->ilen, *aj = a->j;
8181: MatScalar *aa;
8182: PetscBool ignorezeroentries = (a->ignorezeroentries && addv == ADD_VALUES) ? PETSC_TRUE : PETSC_FALSE;
8183: Mat B = aij->B;
8184: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
8185: PetscInt *bimax = b->imax, *bi = b->i, *bilen = b->ilen, *bj = b->j, bm = aij->B->rmap->n, am = aij->A->rmap->n;
8186: MatScalar *ba;
8187: /* This variable below is only for the PETSC_HAVE_VIENNACL or PETSC_HAVE_CUDA cases, but we define it in all cases because we
8188: * cannot use "#if defined" inside a macro. */
8189: PETSC_UNUSED PetscBool inserted = PETSC_FALSE;
8191: PetscInt *rp1, *rp2, ii, nrow1, nrow2, _i, rmax1, rmax2, N, low1, high1, low2, high2, t, lastcol1, lastcol2;
8192: PetscInt nonew = a->nonew;
8193: MatScalar *ap1, *ap2;
8195: PetscFunctionBegin;
8196: PetscCall(MatSeqAIJGetArray(A, &aa));
8197: PetscCall(MatSeqAIJGetArray(B, &ba));
8198: for (i = 0; i < m; i++) {
8199: if (im[i] < 0) continue;
8200: 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);
8201: if (im[i] >= rstart && im[i] < rend) {
8202: row = im[i] - rstart;
8203: lastcol1 = -1;
8204: rp1 = aj + ai[row];
8205: ap1 = aa + ai[row];
8206: rmax1 = aimax[row];
8207: nrow1 = ailen[row];
8208: low1 = 0;
8209: high1 = nrow1;
8210: lastcol2 = -1;
8211: rp2 = bj + bi[row];
8212: ap2 = ba + bi[row];
8213: rmax2 = bimax[row];
8214: nrow2 = bilen[row];
8215: low2 = 0;
8216: high2 = nrow2;
8218: for (j = 0; j < n; j++) {
8219: if (roworiented) value = v[i * n + j];
8220: else value = v[i + j * m];
8221: if (ignorezeroentries && value == 0.0 && addv == ADD_VALUES && im[i] != in[j]) continue;
8222: if (in[j] >= cstart && in[j] < cend) {
8223: col = in[j] - cstart;
8224: MatSetValues_SeqAIJ_A_Private(row, col, value, addv, im[i], in[j]);
8225: } else if (in[j] < 0) continue;
8226: else if (PetscUnlikelyDebug(in[j] >= mat->cmap->N)) {
8227: SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large: col %" PetscInt_FMT " max %" PetscInt_FMT, in[j], mat->cmap->N - 1);
8228: } else {
8229: if (mat->was_assembled) {
8230: if (!aij->colmap) PetscCall(MatCreateColmap_MPIAIJ_Private(mat));
8231: #if PetscDefined(USE_CTABLE)
8232: PetscCall(PetscHMapIGetWithDefault(aij->colmap, in[j] + 1, 0, &col));
8233: col--;
8234: #else
8235: col = aij->colmap[in[j]] - 1;
8236: #endif
8237: if (col < 0 && !((Mat_SeqAIJ *)aij->A->data)->nonew) {
8238: PetscCall(MatDisAssemble_MPIAIJ(mat, PETSC_FALSE));
8239: col = in[j];
8240: /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
8241: B = aij->B;
8242: b = (Mat_SeqAIJ *)B->data;
8243: bimax = b->imax;
8244: bi = b->i;
8245: bilen = b->ilen;
8246: bj = b->j;
8247: rp2 = bj + bi[row];
8248: ap2 = ba + bi[row];
8249: rmax2 = bimax[row];
8250: nrow2 = bilen[row];
8251: low2 = 0;
8252: high2 = nrow2;
8253: bm = aij->B->rmap->n;
8254: ba = b->a;
8255: inserted = PETSC_FALSE;
8256: }
8257: } else col = in[j];
8258: MatSetValues_SeqAIJ_B_Private(row, col, value, addv, im[i], in[j]);
8259: }
8260: }
8261: } else if (!aij->donotstash) {
8262: if (roworiented) {
8263: PetscCall(MatStashValuesRow_Private(&mat->stash, im[i], n, in, v + i * n, (PetscBool)(ignorezeroentries && addv == ADD_VALUES)));
8264: } else {
8265: PetscCall(MatStashValuesCol_Private(&mat->stash, im[i], n, in, v + i, m, (PetscBool)(ignorezeroentries && addv == ADD_VALUES)));
8266: }
8267: }
8268: }
8269: PetscCall(MatSeqAIJRestoreArray(A, &aa));
8270: PetscCall(MatSeqAIJRestoreArray(B, &ba));
8271: }
8272: PetscFunctionReturnVoid();
8273: }
8275: /* Undefining these here since they were redefined from their original definition above! No
8276: * other PETSc functions should be defined past this point, as it is impossible to recover the
8277: * original definitions */
8278: #undef PetscCall
8279: #undef SETERRQ