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: MATAIJ - MATAIJ = "aij" - A matrix type to be used for sparse matrices.
138: This matrix type is identical to` MATSEQAIJ` when constructed with a single process communicator,
139: and `MATMPIAIJ` 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 aij - sets the matrix type to `MATAIJ` during a call to `MatSetFromOptions()`
147: Developer Note:
148: Level: beginner
150: Subclasses include `MATAIJCUSPARSE`, `MATAIJPERM`, `MATAIJSELL`, `MATAIJMKL`, `MATAIJCRL`, `MATAIJKOKKOS`,and also automatically switches over to use inodes when
151: enough exist.
153: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MATSEQAIJ`, `MatCreateAIJ()`, `MatCreateSeqAIJ()`, `MATBAIJ`
154: M*/
156: /*MC
157: MATAIJCRL - MATAIJCRL = "aijcrl" - A matrix type to be used for sparse matrices.
159: This matrix type is identical to `MATSEQAIJCRL` when constructed with a single process communicator,
160: and `MATMPIAIJCRL` otherwise. As a result, for single process communicators,
161: `MatSeqAIJSetPreallocation()` is supported, and similarly `MatMPIAIJSetPreallocation()` is supported
162: for communicators controlling multiple processes. It is recommended that you call both of
163: the above preallocation routines for simplicity.
165: Options Database Key:
166: . -mat_type aijcrl - sets the matrix type to `MATMPIAIJCRL` during a call to `MatSetFromOptions()`
168: Level: beginner
170: .seealso: [](ch_matrices), `Mat`, `MatCreateMPIAIJCRL`, `MATSEQAIJCRL`, `MATMPIAIJCRL`, `MATSEQAIJ`, `MATMPIAIJ`, `MATAIJ`
171: M*/
173: static PetscErrorCode MatBindToCPU_MPIAIJ(Mat A, PetscBool flg)
174: {
175: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
177: PetscFunctionBegin;
178: #if PetscDefined(HAVE_CUDA) || PetscDefined(HAVE_HIP) || PetscDefined(HAVE_VIENNACL)
179: A->boundtocpu = flg;
180: #endif
181: if (a->A) PetscCall(MatBindToCPU(a->A, flg));
182: if (a->B) PetscCall(MatBindToCPU(a->B, flg));
184: /* In addition to binding the diagonal and off-diagonal matrices, bind the local vectors used for matrix-vector products.
185: * This maybe seems a little odd for a MatBindToCPU() call to do, but it makes no sense for the binding of these vectors
186: * to differ from the parent matrix. */
187: if (a->lvec) PetscCall(VecBindToCPU(a->lvec, flg));
188: if (a->diag) PetscCall(VecBindToCPU(a->diag, flg));
189: PetscFunctionReturn(PETSC_SUCCESS);
190: }
192: static PetscErrorCode MatSetBlockSizes_MPIAIJ(Mat M, PetscInt rbs, PetscInt cbs)
193: {
194: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)M->data;
196: PetscFunctionBegin;
197: if (mat->A) {
198: PetscCall(MatSetBlockSizes(mat->A, rbs, cbs));
199: PetscCall(MatSetBlockSizes(mat->B, rbs, 1));
200: }
201: PetscFunctionReturn(PETSC_SUCCESS);
202: }
204: static PetscErrorCode MatFindNonzeroRows_MPIAIJ(Mat M, IS *keptrows)
205: {
206: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)M->data;
207: Mat_SeqAIJ *a = (Mat_SeqAIJ *)mat->A->data;
208: Mat_SeqAIJ *b = (Mat_SeqAIJ *)mat->B->data;
209: const PetscInt *ia, *ib;
210: const MatScalar *aa, *bb, *aav, *bav;
211: PetscInt na, nb, i, j, *rows, cnt = 0, n0rows;
212: PetscInt m = M->rmap->n, rstart = M->rmap->rstart;
214: PetscFunctionBegin;
215: *keptrows = NULL;
217: ia = a->i;
218: ib = b->i;
219: PetscCall(MatSeqAIJGetArrayRead(mat->A, &aav));
220: PetscCall(MatSeqAIJGetArrayRead(mat->B, &bav));
221: for (i = 0; i < m; i++) {
222: na = ia[i + 1] - ia[i];
223: nb = ib[i + 1] - ib[i];
224: if (!na && !nb) {
225: cnt++;
226: goto ok1;
227: }
228: aa = aav + ia[i];
229: for (j = 0; j < na; j++) {
230: if (aa[j] != 0.0) goto ok1;
231: }
232: bb = PetscSafePointerPlusOffset(bav, ib[i]);
233: for (j = 0; j < nb; j++) {
234: if (bb[j] != 0.0) goto ok1;
235: }
236: cnt++;
237: ok1:;
238: }
239: PetscCallMPI(MPIU_Allreduce(&cnt, &n0rows, 1, MPIU_INT, MPI_SUM, PetscObjectComm((PetscObject)M)));
240: if (!n0rows) {
241: PetscCall(MatSeqAIJRestoreArrayRead(mat->A, &aav));
242: PetscCall(MatSeqAIJRestoreArrayRead(mat->B, &bav));
243: PetscFunctionReturn(PETSC_SUCCESS);
244: }
245: PetscCall(PetscMalloc1(M->rmap->n - cnt, &rows));
246: cnt = 0;
247: for (i = 0; i < m; i++) {
248: na = ia[i + 1] - ia[i];
249: nb = ib[i + 1] - ib[i];
250: if (!na && !nb) continue;
251: aa = aav + ia[i];
252: for (j = 0; j < na; j++) {
253: if (aa[j] != 0.0) {
254: rows[cnt++] = rstart + i;
255: goto ok2;
256: }
257: }
258: bb = PetscSafePointerPlusOffset(bav, ib[i]);
259: for (j = 0; j < nb; j++) {
260: if (bb[j] != 0.0) {
261: rows[cnt++] = rstart + i;
262: goto ok2;
263: }
264: }
265: ok2:;
266: }
267: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)M), cnt, rows, PETSC_OWN_POINTER, keptrows));
268: PetscCall(MatSeqAIJRestoreArrayRead(mat->A, &aav));
269: PetscCall(MatSeqAIJRestoreArrayRead(mat->B, &bav));
270: PetscFunctionReturn(PETSC_SUCCESS);
271: }
273: static PetscErrorCode MatDiagonalSet_MPIAIJ(Mat Y, Vec D, InsertMode is)
274: {
275: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)Y->data;
276: PetscBool cong;
278: PetscFunctionBegin;
279: PetscCall(MatHasCongruentLayouts(Y, &cong));
280: if (Y->assembled && cong) PetscCall(MatDiagonalSet(aij->A, D, is));
281: else PetscCall(MatDiagonalSet_Default(Y, D, is));
282: PetscFunctionReturn(PETSC_SUCCESS);
283: }
285: static PetscErrorCode MatFindZeroDiagonals_MPIAIJ(Mat M, IS *zrows)
286: {
287: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)M->data;
288: PetscInt i, rstart, nrows, *rows;
290: PetscFunctionBegin;
291: *zrows = NULL;
292: PetscCall(MatFindZeroDiagonals_SeqAIJ_Private(aij->A, &nrows, &rows));
293: PetscCall(MatGetOwnershipRange(M, &rstart, NULL));
294: for (i = 0; i < nrows; i++) rows[i] += rstart;
295: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)M), nrows, rows, PETSC_OWN_POINTER, zrows));
296: PetscFunctionReturn(PETSC_SUCCESS);
297: }
299: static PetscErrorCode MatGetColumnReductions_MPIAIJ(Mat A, PetscInt type, PetscReal *reductions)
300: {
301: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)A->data;
302: PetscInt i, m, n, *garray = aij->garray;
303: Mat_SeqAIJ *a_aij = (Mat_SeqAIJ *)aij->A->data;
304: Mat_SeqAIJ *b_aij = (Mat_SeqAIJ *)aij->B->data;
305: const PetscScalar *dummy;
307: PetscFunctionBegin;
308: PetscCall(MatGetSize(A, &m, &n));
309: PetscCall(PetscArrayzero(reductions, n));
310: PetscCall(MatSeqAIJGetArrayRead(aij->A, &dummy));
311: PetscCall(MatSeqAIJRestoreArrayRead(aij->A, &dummy));
312: PetscCall(MatSeqAIJGetArrayRead(aij->B, &dummy));
313: PetscCall(MatSeqAIJRestoreArrayRead(aij->B, &dummy));
314: if (type == NORM_2) {
315: 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]);
316: 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]);
317: } else if (type == NORM_1) {
318: 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]);
319: for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) reductions[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]);
320: } else if (type == NORM_INFINITY) {
321: 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]]);
322: 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]]]);
323: } else if (type == REDUCTION_SUM_REALPART || type == REDUCTION_MEAN_REALPART) {
324: 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]);
325: for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) reductions[garray[b_aij->j[i]]] += PetscRealPart(b_aij->a[i]);
326: } else {
327: PetscCheck(type == REDUCTION_SUM_IMAGINARYPART || type == REDUCTION_MEAN_IMAGINARYPART, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Unknown reduction type");
328: 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]);
329: for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) reductions[garray[b_aij->j[i]]] += PetscImaginaryPart(b_aij->a[i]);
330: }
331: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, reductions, n, MPIU_REAL, type == NORM_INFINITY ? MPIU_MAX : MPIU_SUM, PetscObjectComm((PetscObject)A)));
332: if (type == NORM_2) {
333: for (i = 0; i < n; i++) reductions[i] = PetscSqrtReal(reductions[i]);
334: } else if (type == REDUCTION_MEAN_REALPART || type == REDUCTION_MEAN_IMAGINARYPART) {
335: for (i = 0; i < n; i++) reductions[i] /= m;
336: }
337: PetscFunctionReturn(PETSC_SUCCESS);
338: }
340: static PetscErrorCode MatFindOffBlockDiagonalEntries_MPIAIJ(Mat A, IS *is)
341: {
342: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
343: IS sis, gis;
344: const PetscInt *isis, *igis;
345: PetscInt n, *iis, nsis, ngis, rstart, i;
347: PetscFunctionBegin;
348: PetscCall(MatFindOffBlockDiagonalEntries(a->A, &sis));
349: PetscCall(MatFindNonzeroRows(a->B, &gis));
350: PetscCall(ISGetSize(gis, &ngis));
351: PetscCall(ISGetSize(sis, &nsis));
352: PetscCall(ISGetIndices(sis, &isis));
353: PetscCall(ISGetIndices(gis, &igis));
355: PetscCall(PetscMalloc1(ngis + nsis, &iis));
356: PetscCall(PetscArraycpy(iis, igis, ngis));
357: PetscCall(PetscArraycpy(iis + ngis, isis, nsis));
358: n = ngis + nsis;
359: PetscCall(PetscSortRemoveDupsInt(&n, iis));
360: PetscCall(MatGetOwnershipRange(A, &rstart, NULL));
361: for (i = 0; i < n; i++) iis[i] += rstart;
362: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)A), n, iis, PETSC_OWN_POINTER, is));
364: PetscCall(ISRestoreIndices(sis, &isis));
365: PetscCall(ISRestoreIndices(gis, &igis));
366: PetscCall(ISDestroy(&sis));
367: PetscCall(ISDestroy(&gis));
368: PetscFunctionReturn(PETSC_SUCCESS);
369: }
371: /*
372: Local utility routine that creates a mapping from the global column
373: number to the local number in the off-diagonal part of the local
374: storage of the matrix. When PETSC_USE_CTABLE is used this is scalable at
375: a slightly higher hash table cost; without it it is not scalable (each processor
376: has an order N integer array but is fast to access.
377: */
378: PetscErrorCode MatCreateColmap_MPIAIJ_Private(Mat mat)
379: {
380: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
381: PetscInt n = aij->B->cmap->n, i;
383: PetscFunctionBegin;
384: PetscCheck(!n || aij->garray, PETSC_COMM_SELF, PETSC_ERR_PLIB, "MPIAIJ Matrix was assembled but is missing garray");
385: #if PetscDefined(USE_CTABLE)
386: PetscCall(PetscHMapICreateWithSize(n, &aij->colmap));
387: for (i = 0; i < n; i++) PetscCall(PetscHMapISet(aij->colmap, aij->garray[i] + 1, i + 1));
388: #else
389: PetscCall(PetscCalloc1(mat->cmap->N + 1, &aij->colmap));
390: for (i = 0; i < n; i++) aij->colmap[aij->garray[i]] = i + 1;
391: #endif
392: PetscFunctionReturn(PETSC_SUCCESS);
393: }
395: #define MatSetValues_SeqAIJ_A_Private(row, col, value, addv, orow, ocol) \
396: do { \
397: if (col <= lastcol1) low1 = 0; \
398: else high1 = nrow1; \
399: lastcol1 = col; \
400: while (high1 - low1 > 5) { \
401: t = (low1 + high1) / 2; \
402: if (rp1[t] > col) high1 = t; \
403: else low1 = t; \
404: } \
405: for (_i = low1; _i < high1; _i++) { \
406: if (rp1[_i] > col) break; \
407: if (rp1[_i] == col) { \
408: if (addv == ADD_VALUES) { \
409: ap1[_i] += value; \
410: /* Not sure LogFlops will slow down the code or not */ \
411: (void)PetscLogFlops(1.0); \
412: } else ap1[_i] = value; \
413: goto a_noinsert; \
414: } \
415: } \
416: if (value == 0.0 && ignorezeroentries && row != col) { \
417: low1 = 0; \
418: high1 = nrow1; \
419: goto a_noinsert; \
420: } \
421: if (nonew == 1) { \
422: low1 = 0; \
423: high1 = nrow1; \
424: goto a_noinsert; \
425: } \
426: 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); \
427: MatSeqXAIJReallocateAIJ(A, am, 1, nrow1, row, col, rmax1, aa, ai, aj, rp1, ap1, aimax, nonew, MatScalar); \
428: N = nrow1++ - 1; \
429: a->nz++; \
430: high1++; \
431: /* shift up all the later entries in this row */ \
432: PetscCall(PetscArraymove(rp1 + _i + 1, rp1 + _i, N - _i + 1)); \
433: PetscCall(PetscArraymove(ap1 + _i + 1, ap1 + _i, N - _i + 1)); \
434: rp1[_i] = col; \
435: ap1[_i] = value; \
436: a_noinsert:; \
437: ailen[row] = nrow1; \
438: } while (0)
440: #define MatSetValues_SeqAIJ_B_Private(row, col, value, addv, orow, ocol) \
441: do { \
442: if (col <= lastcol2) low2 = 0; \
443: else high2 = nrow2; \
444: lastcol2 = col; \
445: while (high2 - low2 > 5) { \
446: t = (low2 + high2) / 2; \
447: if (rp2[t] > col) high2 = t; \
448: else low2 = t; \
449: } \
450: for (_i = low2; _i < high2; _i++) { \
451: if (rp2[_i] > col) break; \
452: if (rp2[_i] == col) { \
453: if (addv == ADD_VALUES) { \
454: ap2[_i] += value; \
455: (void)PetscLogFlops(1.0); \
456: } else ap2[_i] = value; \
457: goto b_noinsert; \
458: } \
459: } \
460: if (value == 0.0 && ignorezeroentries) { \
461: low2 = 0; \
462: high2 = nrow2; \
463: goto b_noinsert; \
464: } \
465: if (nonew == 1) { \
466: low2 = 0; \
467: high2 = nrow2; \
468: goto b_noinsert; \
469: } \
470: 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); \
471: MatSeqXAIJReallocateAIJ(B, bm, 1, nrow2, row, col, rmax2, ba, bi, bj, rp2, ap2, bimax, nonew, MatScalar); \
472: N = nrow2++ - 1; \
473: b->nz++; \
474: high2++; \
475: /* shift up all the later entries in this row */ \
476: PetscCall(PetscArraymove(rp2 + _i + 1, rp2 + _i, N - _i + 1)); \
477: PetscCall(PetscArraymove(ap2 + _i + 1, ap2 + _i, N - _i + 1)); \
478: rp2[_i] = col; \
479: ap2[_i] = value; \
480: b_noinsert:; \
481: bilen[row] = nrow2; \
482: } while (0)
484: static PetscErrorCode MatSetValuesRow_MPIAIJ(Mat A, PetscInt row, const PetscScalar v[])
485: {
486: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
487: Mat_SeqAIJ *a = (Mat_SeqAIJ *)mat->A->data, *b = (Mat_SeqAIJ *)mat->B->data;
488: PetscInt l, *garray = mat->garray, diag;
489: PetscScalar *aa, *ba;
491: PetscFunctionBegin;
492: /* code only works for square matrices A */
494: /* find size of row to the left of the diagonal part */
495: PetscCall(MatGetOwnershipRange(A, &diag, NULL));
496: row = row - diag;
497: for (l = 0; l < b->i[row + 1] - b->i[row]; l++) {
498: if (garray[b->j[b->i[row] + l]] > diag) break;
499: }
500: if (l) {
501: PetscCall(MatSeqAIJGetArray(mat->B, &ba));
502: PetscCall(PetscArraycpy(ba + b->i[row], v, l));
503: PetscCall(MatSeqAIJRestoreArray(mat->B, &ba));
504: }
506: /* diagonal part */
507: if (a->i[row + 1] - a->i[row]) {
508: PetscCall(MatSeqAIJGetArray(mat->A, &aa));
509: PetscCall(PetscArraycpy(aa + a->i[row], v + l, a->i[row + 1] - a->i[row]));
510: PetscCall(MatSeqAIJRestoreArray(mat->A, &aa));
511: }
513: /* right of diagonal part */
514: if (b->i[row + 1] - b->i[row] - l) {
515: PetscCall(MatSeqAIJGetArray(mat->B, &ba));
516: PetscCall(PetscArraycpy(ba + b->i[row] + l, v + l + a->i[row + 1] - a->i[row], b->i[row + 1] - b->i[row] - l));
517: PetscCall(MatSeqAIJRestoreArray(mat->B, &ba));
518: }
519: PetscFunctionReturn(PETSC_SUCCESS);
520: }
522: PetscErrorCode MatSetValues_MPIAIJ(Mat mat, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode addv)
523: {
524: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
525: PetscScalar value = 0.0;
526: PetscInt i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend;
527: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend, row, col;
528: PetscBool roworiented = aij->roworiented;
530: /* Some Variables required in the macro */
531: Mat A = aij->A;
532: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
533: PetscInt *aimax = a->imax, *ai = a->i, *ailen = a->ilen, *aj = a->j;
534: PetscBool ignorezeroentries = a->ignorezeroentries;
535: Mat B = aij->B;
536: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
537: PetscInt *bimax = b->imax, *bi = b->i, *bilen = b->ilen, *bj = b->j, bm = aij->B->rmap->n, am = aij->A->rmap->n;
538: MatScalar *aa, *ba;
539: PetscInt *rp1, *rp2, ii, nrow1, nrow2, _i, rmax1, rmax2, N, low1, high1, low2, high2, t, lastcol1, lastcol2;
540: PetscInt nonew;
541: MatScalar *ap1, *ap2;
543: PetscFunctionBegin;
544: PetscCall(MatSeqAIJGetArray(A, &aa));
545: PetscCall(MatSeqAIJGetArray(B, &ba));
546: for (i = 0; i < m; i++) {
547: if (im[i] < 0) continue;
548: 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);
549: if (im[i] >= rstart && im[i] < rend) {
550: row = im[i] - rstart;
551: lastcol1 = -1;
552: rp1 = PetscSafePointerPlusOffset(aj, ai[row]);
553: ap1 = PetscSafePointerPlusOffset(aa, ai[row]);
554: rmax1 = aimax[row];
555: nrow1 = ailen[row];
556: low1 = 0;
557: high1 = nrow1;
558: lastcol2 = -1;
559: rp2 = PetscSafePointerPlusOffset(bj, bi[row]);
560: ap2 = PetscSafePointerPlusOffset(ba, bi[row]);
561: rmax2 = bimax[row];
562: nrow2 = bilen[row];
563: low2 = 0;
564: high2 = nrow2;
566: for (j = 0; j < n; j++) {
567: if (v) value = roworiented ? v[i * n + j] : v[i + j * m];
568: if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && im[i] != in[j]) continue;
569: if (in[j] >= cstart && in[j] < cend) {
570: col = in[j] - cstart;
571: nonew = a->nonew;
572: MatSetValues_SeqAIJ_A_Private(row, col, value, addv, im[i], in[j]);
573: } else if (in[j] < 0) {
574: continue;
575: } else {
576: 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);
577: if (mat->was_assembled) {
578: if (!aij->colmap) PetscCall(MatCreateColmap_MPIAIJ_Private(mat));
579: #if PetscDefined(USE_CTABLE)
580: PetscCall(PetscHMapIGetWithDefault(aij->colmap, in[j] + 1, 0, &col)); /* map global col ids to local ones */
581: col--;
582: #else
583: col = aij->colmap[in[j]] - 1;
584: #endif
585: if (col < 0 && !((Mat_SeqAIJ *)aij->B->data)->nonew) { /* col < 0 means in[j] is a new col for B */
586: PetscCall(MatDisAssemble_MPIAIJ(mat, PETSC_FALSE)); /* Change aij->B from reduced/local format to expanded/global format */
587: col = in[j];
588: /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
589: B = aij->B;
590: b = (Mat_SeqAIJ *)B->data;
591: bimax = b->imax;
592: bi = b->i;
593: bilen = b->ilen;
594: bj = b->j;
595: ba = b->a;
596: rp2 = PetscSafePointerPlusOffset(bj, bi[row]);
597: ap2 = PetscSafePointerPlusOffset(ba, bi[row]);
598: rmax2 = bimax[row];
599: nrow2 = bilen[row];
600: low2 = 0;
601: high2 = nrow2;
602: bm = aij->B->rmap->n;
603: ba = b->a;
604: } else if (col < 0 && !(ignorezeroentries && value == 0.0)) {
605: 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]);
606: 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]));
607: }
608: } else col = in[j];
609: nonew = b->nonew;
610: MatSetValues_SeqAIJ_B_Private(row, col, value, addv, im[i], in[j]);
611: }
612: }
613: } else {
614: 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]);
615: if (!aij->donotstash) {
616: mat->assembled = PETSC_FALSE;
617: if (roworiented) {
618: PetscCall(MatStashValuesRow_Private(&mat->stash, im[i], n, in, PetscSafePointerPlusOffset(v, i * n), (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
619: } else {
620: PetscCall(MatStashValuesCol_Private(&mat->stash, im[i], n, in, PetscSafePointerPlusOffset(v, i), m, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
621: }
622: }
623: }
624: }
625: PetscCall(MatSeqAIJRestoreArray(A, &aa)); /* aa, bb might have been free'd due to reallocation above. But we don't access them here */
626: PetscCall(MatSeqAIJRestoreArray(B, &ba));
627: PetscFunctionReturn(PETSC_SUCCESS);
628: }
630: /*
631: This function sets the j and ilen arrays (of the diagonal and off-diagonal part) of an MPIAIJ-matrix.
632: The values in mat_i have to be sorted and the values in mat_j have to be sorted for each row (CSR-like).
633: No off-processor parts off the matrix are allowed here and mat->was_assembled has to be PETSC_FALSE.
634: */
635: PetscErrorCode MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(Mat mat, const PetscInt mat_j[], const PetscInt mat_i[])
636: {
637: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
638: Mat A = aij->A; /* diagonal part of the matrix */
639: Mat B = aij->B; /* off-diagonal part of the matrix */
640: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
641: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
642: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend, col;
643: PetscInt *ailen = a->ilen, *aj = a->j;
644: PetscInt *bilen = b->ilen, *bj = b->j;
645: PetscInt am = aij->A->rmap->n, j;
646: PetscInt diag_so_far = 0, dnz;
647: PetscInt offd_so_far = 0, onz;
649: PetscFunctionBegin;
650: /* Iterate over all rows of the matrix */
651: for (j = 0; j < am; j++) {
652: dnz = onz = 0;
653: /* Iterate over all non-zero columns of the current row */
654: for (col = mat_i[j]; col < mat_i[j + 1]; col++) {
655: /* If column is in the diagonal */
656: if (mat_j[col] >= cstart && mat_j[col] < cend) {
657: aj[diag_so_far++] = mat_j[col] - cstart;
658: dnz++;
659: } else { /* off-diagonal entries */
660: bj[offd_so_far++] = mat_j[col];
661: onz++;
662: }
663: }
664: ailen[j] = dnz;
665: bilen[j] = onz;
666: }
667: PetscFunctionReturn(PETSC_SUCCESS);
668: }
670: /*
671: This function sets the local j, a and ilen arrays (of the diagonal and off-diagonal part) of an MPIAIJ-matrix.
672: The values in mat_i have to be sorted and the values in mat_j have to be sorted for each row (CSR-like).
673: No off-processor parts off the matrix are allowed here, they are set at a later point by MatSetValues_MPIAIJ.
674: Also, mat->was_assembled has to be false, otherwise the statement aj[rowstart_diag+dnz_row] = mat_j[col] - cstart;
675: would not be true and the more complex MatSetValues_MPIAIJ has to be used.
676: */
677: PetscErrorCode MatSetValues_MPIAIJ_CopyFromCSRFormat(Mat mat, const PetscInt mat_j[], const PetscInt mat_i[], const PetscScalar mat_a[])
678: {
679: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
680: Mat A = aij->A; /* diagonal part of the matrix */
681: Mat B = aij->B; /* off-diagonal part of the matrix */
682: Mat_SeqAIJ *aijd = (Mat_SeqAIJ *)aij->A->data, *aijo = (Mat_SeqAIJ *)aij->B->data;
683: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
684: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
685: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend;
686: PetscInt *ailen = a->ilen, *aj = a->j;
687: PetscInt *bilen = b->ilen, *bj = b->j;
688: PetscInt am = aij->A->rmap->n, j;
689: 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. */
690: PetscInt col, dnz_row, onz_row, rowstart_diag, rowstart_offd;
691: PetscScalar *aa = a->a, *ba = b->a;
693: PetscFunctionBegin;
694: /* Iterate over all rows of the matrix */
695: for (j = 0; j < am; j++) {
696: dnz_row = onz_row = 0;
697: rowstart_offd = full_offd_i[j];
698: rowstart_diag = full_diag_i[j];
699: /* Iterate over all non-zero columns of the current row */
700: for (col = mat_i[j]; col < mat_i[j + 1]; col++) {
701: /* If column is in the diagonal */
702: if (mat_j[col] >= cstart && mat_j[col] < cend) {
703: aj[rowstart_diag + dnz_row] = mat_j[col] - cstart;
704: aa[rowstart_diag + dnz_row] = mat_a[col];
705: dnz_row++;
706: } else { /* off-diagonal entries */
707: bj[rowstart_offd + onz_row] = mat_j[col];
708: ba[rowstart_offd + onz_row] = mat_a[col];
709: onz_row++;
710: }
711: }
712: ailen[j] = dnz_row;
713: bilen[j] = onz_row;
714: }
715: PetscFunctionReturn(PETSC_SUCCESS);
716: }
718: static PetscErrorCode MatGetValues_MPIAIJ(Mat mat, PetscInt m, const PetscInt idxm[], PetscInt n, const PetscInt idxn[], PetscScalar v[])
719: {
720: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
721: PetscInt i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend;
722: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend, row, col;
723: PetscBool roworiented = aij->roworiented;
724: PetscScalar *value;
726: PetscFunctionBegin;
727: for (i = 0; i < m; i++) {
728: if (idxm[i] < 0) continue; /* negative row */
729: 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);
730: 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);
731: row = idxm[i] - rstart;
732: for (j = 0; j < n; j++) {
733: if (idxn[j] < 0) continue; /* negative column */
734: 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);
735: value = roworiented ? &v[j + i * n] : &v[i + j * m];
736: if (idxn[j] >= cstart && idxn[j] < cend) {
737: col = idxn[j] - cstart;
738: PetscCall(MatGetValues(aij->A, 1, &row, 1, &col, value));
739: } else {
740: if (!aij->colmap) PetscCall(MatCreateColmap_MPIAIJ_Private(mat));
741: #if PetscDefined(USE_CTABLE)
742: PetscCall(PetscHMapIGetWithDefault(aij->colmap, idxn[j] + 1, 0, &col));
743: col--;
744: #else
745: col = aij->colmap[idxn[j]] - 1;
746: #endif
747: if ((col < 0) || (aij->garray[col] != idxn[j])) *value = 0.0;
748: else PetscCall(MatGetValues(aij->B, 1, &row, 1, &col, value));
749: }
750: }
751: }
752: PetscFunctionReturn(PETSC_SUCCESS);
753: }
755: static PetscErrorCode MatAssemblyBegin_MPIAIJ(Mat mat, MatAssemblyType mode)
756: {
757: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
758: PetscInt nstash, reallocs;
760: PetscFunctionBegin;
761: if (aij->donotstash || mat->nooffprocentries) PetscFunctionReturn(PETSC_SUCCESS);
763: PetscCall(MatStashScatterBegin_Private(mat, &mat->stash, mat->rmap->range));
764: PetscCall(MatStashGetInfo_Private(&mat->stash, &nstash, &reallocs));
765: PetscCall(PetscInfo(mat, "Stash has %" PetscInt_FMT " entries, uses %" PetscInt_FMT " mallocs.\n", nstash, reallocs));
766: PetscFunctionReturn(PETSC_SUCCESS);
767: }
769: PetscErrorCode MatAssemblyEnd_MPIAIJ(Mat mat, MatAssemblyType mode)
770: {
771: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
772: PetscMPIInt n;
773: PetscInt i, j, rstart, ncols, flg;
774: PetscInt *row, *col;
775: PetscBool all_assembled;
776: PetscScalar *val;
778: /* do not use 'b = (Mat_SeqAIJ*)aij->B->data' as B can be reset in disassembly */
780: PetscFunctionBegin;
781: if (!aij->donotstash && !mat->nooffprocentries) {
782: while (1) {
783: PetscCall(MatStashScatterGetMesg_Private(&mat->stash, &n, &row, &col, &val, &flg));
784: if (!flg) break;
786: for (i = 0; i < n;) {
787: /* Now identify the consecutive vals belonging to the same row */
788: for (j = i, rstart = row[j]; j < n; j++) {
789: if (row[j] != rstart) break;
790: }
791: if (j < n) ncols = j - i;
792: else ncols = n - i;
793: /* Now assemble all these values with a single function call */
794: PetscCall(MatSetValues_MPIAIJ(mat, 1, row + i, ncols, col + i, val + i, mat->insertmode));
795: i = j;
796: }
797: }
798: PetscCall(MatStashScatterEnd_Private(&mat->stash));
799: }
800: #if PetscDefined(HAVE_DEVICE)
801: if (mat->offloadmask == PETSC_OFFLOAD_CPU) aij->A->offloadmask = PETSC_OFFLOAD_CPU;
802: /* We call MatBindToCPU() on aij->A and aij->B here, because if MatBindToCPU_MPIAIJ() is called before assembly, it cannot bind these. */
803: if (mat->boundtocpu) {
804: PetscCall(MatBindToCPU(aij->A, PETSC_TRUE));
805: PetscCall(MatBindToCPU(aij->B, PETSC_TRUE));
806: }
807: #endif
808: PetscCall(MatAssemblyBegin(aij->A, mode));
809: PetscCall(MatAssemblyEnd(aij->A, mode));
811: /* determine if any process has disassembled, if so we must
812: also disassemble ourself, in order that we may reassemble. */
813: /*
814: if nonzero structure of submatrix B cannot change then we know that
815: no process disassembled thus we can skip this stuff
816: */
817: if (!((Mat_SeqAIJ *)aij->B->data)->nonew) {
818: PetscCallMPI(MPIU_Allreduce(&mat->was_assembled, &all_assembled, 1, MPI_C_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
819: if (mat->was_assembled && !all_assembled) { /* mat on this rank has reduced off-diag B with local col ids, but globally it does not */
820: PetscCall(MatDisAssemble_MPIAIJ(mat, PETSC_FALSE));
821: }
822: }
823: if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) PetscCall(MatSetUpMultiply_MPIAIJ(mat));
824: PetscCall(MatSetOption(aij->B, MAT_USE_INODES, PETSC_FALSE));
825: #if PetscDefined(HAVE_DEVICE)
826: if (mat->offloadmask == PETSC_OFFLOAD_CPU && aij->B->offloadmask != PETSC_OFFLOAD_UNALLOCATED) aij->B->offloadmask = PETSC_OFFLOAD_CPU;
827: #endif
828: PetscCall(MatAssemblyBegin(aij->B, mode));
829: PetscCall(MatAssemblyEnd(aij->B, mode));
831: PetscCall(PetscFree2(aij->rowvalues, aij->rowindices));
833: aij->rowvalues = NULL;
835: PetscCall(VecDestroy(&aij->diag));
837: /* if no new nonzero locations are allowed in matrix then only set the matrix state the first time through */
838: if ((!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) || !((Mat_SeqAIJ *)aij->A->data)->nonew) {
839: mat->nonzerostate = aij->A->nonzerostate + aij->B->nonzerostate;
840: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &mat->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)mat)));
841: }
842: #if PetscDefined(HAVE_DEVICE)
843: mat->offloadmask = PETSC_OFFLOAD_BOTH;
844: #endif
845: PetscFunctionReturn(PETSC_SUCCESS);
846: }
848: static PetscErrorCode MatZeroEntries_MPIAIJ(Mat A)
849: {
850: Mat_MPIAIJ *l = (Mat_MPIAIJ *)A->data;
852: PetscFunctionBegin;
853: PetscCall(MatZeroEntries(l->A));
854: PetscCall(MatZeroEntries(l->B));
855: PetscFunctionReturn(PETSC_SUCCESS);
856: }
858: static PetscErrorCode MatZeroRows_MPIAIJ(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diag, Vec x, Vec b)
859: {
860: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
861: PetscInt *lrows;
862: PetscInt r, len;
863: PetscBool cong;
865: PetscFunctionBegin;
866: /* get locally owned rows */
867: PetscCall(MatZeroRowsMapLocal_Private(A, N, rows, &len, &lrows));
868: PetscCall(MatHasCongruentLayouts(A, &cong));
869: /* fix right-hand side if needed */
870: if (x && b) {
871: const PetscScalar *xx;
872: PetscScalar *bb;
874: PetscCheck(cong, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Need matching row/col layout");
875: PetscCall(VecGetArrayRead(x, &xx));
876: PetscCall(VecGetArray(b, &bb));
877: for (r = 0; r < len; ++r) bb[lrows[r]] = diag * xx[lrows[r]];
878: PetscCall(VecRestoreArrayRead(x, &xx));
879: PetscCall(VecRestoreArray(b, &bb));
880: }
882: if (diag != 0.0 && cong) {
883: PetscCall(MatZeroRows(mat->A, len, lrows, diag, NULL, NULL));
884: PetscCall(MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL));
885: } else if (diag != 0.0) { /* non-square or non congruent layouts -> if keepnonzeropattern is false, we allow for new insertion */
886: Mat_SeqAIJ *aijA = (Mat_SeqAIJ *)mat->A->data;
887: Mat_SeqAIJ *aijB = (Mat_SeqAIJ *)mat->B->data;
888: PetscInt nnwA, nnwB;
889: PetscBool nnzA, nnzB;
891: nnwA = aijA->nonew;
892: nnwB = aijB->nonew;
893: nnzA = aijA->keepnonzeropattern;
894: nnzB = aijB->keepnonzeropattern;
895: if (!nnzA) {
896: PetscCall(PetscInfo(mat->A, "Requested to not keep the pattern and add a nonzero diagonal; may encounter reallocations on diagonal block.\n"));
897: aijA->nonew = 0;
898: }
899: if (!nnzB) {
900: PetscCall(PetscInfo(mat->B, "Requested to not keep the pattern and add a nonzero diagonal; may encounter reallocations on off-diagonal block.\n"));
901: aijB->nonew = 0;
902: }
903: /* Must zero here before the next loop */
904: PetscCall(MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL));
905: PetscCall(MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL));
906: for (r = 0; r < len; ++r) {
907: const PetscInt row = lrows[r] + A->rmap->rstart;
908: if (row >= A->cmap->N) continue;
909: PetscCall(MatSetValues(A, 1, &row, 1, &row, &diag, INSERT_VALUES));
910: }
911: aijA->nonew = nnwA;
912: aijB->nonew = nnwB;
913: } else {
914: PetscCall(MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL));
915: PetscCall(MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL));
916: }
917: PetscCall(PetscFree(lrows));
918: PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY));
919: PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY));
921: /* only change matrix nonzero state if pattern was allowed to be changed */
922: if (!((Mat_SeqAIJ *)mat->A->data)->keepnonzeropattern || !((Mat_SeqAIJ *)mat->A->data)->nonew) {
923: A->nonzerostate = mat->A->nonzerostate + mat->B->nonzerostate;
924: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &A->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)A)));
925: }
926: PetscFunctionReturn(PETSC_SUCCESS);
927: }
929: static PetscErrorCode MatZeroRowsColumns_MPIAIJ(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diag, Vec x, Vec b)
930: {
931: Mat_MPIAIJ *l = (Mat_MPIAIJ *)A->data;
932: PetscInt n = A->rmap->n;
933: PetscInt i, j, r, m, len = 0;
934: PetscInt *lrows, *owners = A->rmap->range;
935: PetscMPIInt p = 0;
936: PetscSFNode *rrows;
937: PetscSF sf;
938: const PetscScalar *xx;
939: PetscScalar *bb, *mask, *aij_a;
940: Vec xmask, lmask;
941: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)l->B->data;
942: const PetscInt *aj, *ii, *ridx;
943: PetscScalar *aa;
945: PetscFunctionBegin;
946: /* Create SF where leaves are input rows and roots are owned rows */
947: PetscCall(PetscMalloc1(n, &lrows));
948: for (r = 0; r < n; ++r) lrows[r] = -1;
949: PetscCall(PetscMalloc1(N, &rrows));
950: for (r = 0; r < N; ++r) {
951: const PetscInt idx = rows[r];
952: 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);
953: if (idx < owners[p] || owners[p + 1] <= idx) { /* short-circuit the search if the last p owns this row too */
954: PetscCall(PetscLayoutFindOwner(A->rmap, idx, &p));
955: }
956: rrows[r].rank = p;
957: rrows[r].index = rows[r] - owners[p];
958: }
959: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
960: PetscCall(PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER));
961: /* Collect flags for rows to be zeroed */
962: PetscCall(PetscSFReduceBegin(sf, MPIU_INT, (PetscInt *)rows, lrows, MPI_LOR));
963: PetscCall(PetscSFReduceEnd(sf, MPIU_INT, (PetscInt *)rows, lrows, MPI_LOR));
964: PetscCall(PetscSFDestroy(&sf));
965: /* Compress and put in row numbers */
966: for (r = 0; r < n; ++r)
967: if (lrows[r] >= 0) lrows[len++] = r;
968: /* zero diagonal part of matrix */
969: PetscCall(MatZeroRowsColumns(l->A, len, lrows, diag, x, b));
970: /* handle off-diagonal part of matrix */
971: PetscCall(MatCreateVecs(A, &xmask, NULL));
972: PetscCall(VecDuplicate(l->lvec, &lmask));
973: PetscCall(VecGetArray(xmask, &bb));
974: for (i = 0; i < len; i++) bb[lrows[i]] = 1;
975: PetscCall(VecRestoreArray(xmask, &bb));
976: PetscCall(VecScatterBegin(l->Mvctx, xmask, lmask, ADD_VALUES, SCATTER_FORWARD));
977: PetscCall(VecScatterEnd(l->Mvctx, xmask, lmask, ADD_VALUES, SCATTER_FORWARD));
978: PetscCall(VecDestroy(&xmask));
979: if (x && b) { /* this code is buggy when the row and column layout don't match */
980: PetscBool cong;
982: PetscCall(MatHasCongruentLayouts(A, &cong));
983: PetscCheck(cong, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Need matching row/col layout");
984: PetscCall(VecScatterBegin(l->Mvctx, x, l->lvec, INSERT_VALUES, SCATTER_FORWARD));
985: PetscCall(VecScatterEnd(l->Mvctx, x, l->lvec, INSERT_VALUES, SCATTER_FORWARD));
986: PetscCall(VecGetArrayRead(l->lvec, &xx));
987: PetscCall(VecGetArray(b, &bb));
988: }
989: PetscCall(VecGetArray(lmask, &mask));
990: /* remove zeroed rows of off-diagonal matrix */
991: PetscCall(MatSeqAIJGetArray(l->B, &aij_a));
992: ii = aij->i;
993: for (i = 0; i < len; i++) PetscCall(PetscArrayzero(PetscSafePointerPlusOffset(aij_a, ii[lrows[i]]), ii[lrows[i] + 1] - ii[lrows[i]]));
994: /* loop over all elements of off process part of matrix zeroing removed columns*/
995: if (aij->compressedrow.use) {
996: m = aij->compressedrow.nrows;
997: ii = aij->compressedrow.i;
998: ridx = aij->compressedrow.rindex;
999: for (i = 0; i < m; i++) {
1000: n = ii[i + 1] - ii[i];
1001: aj = aij->j + ii[i];
1002: aa = aij_a + ii[i];
1004: for (j = 0; j < n; j++) {
1005: if (PetscAbsScalar(mask[*aj])) {
1006: if (b) bb[*ridx] -= *aa * xx[*aj];
1007: *aa = 0.0;
1008: }
1009: aa++;
1010: aj++;
1011: }
1012: ridx++;
1013: }
1014: } else { /* do not use compressed row format */
1015: m = l->B->rmap->n;
1016: for (i = 0; i < m; i++) {
1017: n = ii[i + 1] - ii[i];
1018: aj = aij->j + ii[i];
1019: aa = aij_a + ii[i];
1020: for (j = 0; j < n; j++) {
1021: if (PetscAbsScalar(mask[*aj])) {
1022: if (b) bb[i] -= *aa * xx[*aj];
1023: *aa = 0.0;
1024: }
1025: aa++;
1026: aj++;
1027: }
1028: }
1029: }
1030: if (x && b) {
1031: PetscCall(VecRestoreArray(b, &bb));
1032: PetscCall(VecRestoreArrayRead(l->lvec, &xx));
1033: }
1034: PetscCall(MatSeqAIJRestoreArray(l->B, &aij_a));
1035: PetscCall(VecRestoreArray(lmask, &mask));
1036: PetscCall(VecDestroy(&lmask));
1037: PetscCall(PetscFree(lrows));
1039: /* only change matrix nonzero state if pattern was allowed to be changed */
1040: if (!((Mat_SeqAIJ *)l->A->data)->nonew) {
1041: A->nonzerostate = l->A->nonzerostate + l->B->nonzerostate;
1042: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &A->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)A)));
1043: }
1044: PetscFunctionReturn(PETSC_SUCCESS);
1045: }
1047: static PetscErrorCode MatMult_MPIAIJ(Mat A, Vec xx, Vec yy)
1048: {
1049: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1050: PetscInt nt;
1051: VecScatter Mvctx = a->Mvctx;
1053: PetscFunctionBegin;
1054: PetscCall(VecGetLocalSize(xx, &nt));
1055: 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);
1056: PetscCall(VecScatterBegin(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1057: PetscUseTypeMethod(a->A, mult, xx, yy);
1058: PetscCall(VecScatterEnd(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1059: PetscUseTypeMethod(a->B, multadd, a->lvec, yy, yy);
1060: PetscFunctionReturn(PETSC_SUCCESS);
1061: }
1063: static PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat A, Vec bb, Vec xx)
1064: {
1065: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1067: PetscFunctionBegin;
1068: PetscCall(MatMultDiagonalBlock(a->A, bb, xx));
1069: PetscFunctionReturn(PETSC_SUCCESS);
1070: }
1072: static PetscErrorCode MatMultAdd_MPIAIJ(Mat A, Vec xx, Vec yy, Vec zz)
1073: {
1074: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1075: VecScatter Mvctx = a->Mvctx;
1077: PetscFunctionBegin;
1078: PetscCall(VecScatterBegin(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1079: PetscUseTypeMethod(a->A, multadd, xx, yy, zz);
1080: PetscCall(VecScatterEnd(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1081: PetscUseTypeMethod(a->B, multadd, a->lvec, zz, zz);
1082: PetscFunctionReturn(PETSC_SUCCESS);
1083: }
1085: static PetscErrorCode MatMultTranspose_MPIAIJ(Mat A, Vec xx, Vec yy)
1086: {
1087: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1089: PetscFunctionBegin;
1090: /* do nondiagonal part */
1091: PetscUseTypeMethod(a->B, multtranspose, xx, a->lvec);
1092: /* do local part */
1093: PetscUseTypeMethod(a->A, multtranspose, xx, yy);
1094: /* add partial results together */
1095: PetscCall(VecScatterBegin(a->Mvctx, a->lvec, yy, ADD_VALUES, SCATTER_REVERSE));
1096: PetscCall(VecScatterEnd(a->Mvctx, a->lvec, yy, ADD_VALUES, SCATTER_REVERSE));
1097: PetscFunctionReturn(PETSC_SUCCESS);
1098: }
1100: static PetscErrorCode MatIsTranspose_MPIAIJ(Mat Amat, Mat Bmat, PetscReal tol, PetscBool *f)
1101: {
1102: MPI_Comm comm;
1103: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)Amat->data, *Bij = (Mat_MPIAIJ *)Bmat->data;
1104: Mat Adia = Aij->A, Bdia = Bij->A, Aoff, Boff, *Aoffs, *Boffs;
1105: IS Me, Notme;
1106: PetscInt M, N, first, last, *notme, i;
1107: PetscMPIInt size;
1109: PetscFunctionBegin;
1110: /* Easy test: symmetric diagonal block */
1111: PetscCall(MatIsTranspose(Adia, Bdia, tol, f));
1112: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, f, 1, MPI_C_BOOL, MPI_LAND, PetscObjectComm((PetscObject)Amat)));
1113: if (!*f) PetscFunctionReturn(PETSC_SUCCESS);
1114: PetscCall(PetscObjectGetComm((PetscObject)Amat, &comm));
1115: PetscCallMPI(MPI_Comm_size(comm, &size));
1116: if (size == 1) PetscFunctionReturn(PETSC_SUCCESS);
1118: /* Hard test: off-diagonal block. This takes a MatCreateSubMatrix. */
1119: PetscCall(MatGetSize(Amat, &M, &N));
1120: PetscCall(MatGetOwnershipRange(Amat, &first, &last));
1121: PetscCall(PetscMalloc1(N - last + first, ¬me));
1122: for (i = 0; i < first; i++) notme[i] = i;
1123: for (i = last; i < M; i++) notme[i - last + first] = i;
1124: PetscCall(ISCreateGeneral(MPI_COMM_SELF, N - last + first, notme, PETSC_COPY_VALUES, &Notme));
1125: PetscCall(ISCreateStride(MPI_COMM_SELF, last - first, first, 1, &Me));
1126: PetscCall(MatCreateSubMatrices(Amat, 1, &Me, &Notme, MAT_INITIAL_MATRIX, &Aoffs));
1127: Aoff = Aoffs[0];
1128: PetscCall(MatCreateSubMatrices(Bmat, 1, &Notme, &Me, MAT_INITIAL_MATRIX, &Boffs));
1129: Boff = Boffs[0];
1130: PetscCall(MatIsTranspose(Aoff, Boff, tol, f));
1131: PetscCall(MatDestroyMatrices(1, &Aoffs));
1132: PetscCall(MatDestroyMatrices(1, &Boffs));
1133: PetscCall(ISDestroy(&Me));
1134: PetscCall(ISDestroy(&Notme));
1135: PetscCall(PetscFree(notme));
1136: PetscFunctionReturn(PETSC_SUCCESS);
1137: }
1139: static PetscErrorCode MatMultTransposeAdd_MPIAIJ(Mat A, Vec xx, Vec yy, Vec zz)
1140: {
1141: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1143: PetscFunctionBegin;
1144: /* do nondiagonal part */
1145: PetscUseTypeMethod(a->B, multtranspose, xx, a->lvec);
1146: /* do local part */
1147: PetscUseTypeMethod(a->A, multtransposeadd, xx, yy, zz);
1148: /* add partial results together */
1149: PetscCall(VecScatterBegin(a->Mvctx, a->lvec, zz, ADD_VALUES, SCATTER_REVERSE));
1150: PetscCall(VecScatterEnd(a->Mvctx, a->lvec, zz, ADD_VALUES, SCATTER_REVERSE));
1151: PetscFunctionReturn(PETSC_SUCCESS);
1152: }
1154: /*
1155: This only works correctly for square matrices where the subblock A->A is the
1156: diagonal block
1157: */
1158: static PetscErrorCode MatGetDiagonal_MPIAIJ(Mat A, Vec v)
1159: {
1160: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1162: PetscFunctionBegin;
1163: PetscCheck(A->rmap->N == A->cmap->N, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Supports only square matrix where A->A is diag block");
1164: 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");
1165: PetscCall(MatGetDiagonal(a->A, v));
1166: PetscFunctionReturn(PETSC_SUCCESS);
1167: }
1169: static PetscErrorCode MatScale_MPIAIJ(Mat A, PetscScalar aa)
1170: {
1171: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1173: PetscFunctionBegin;
1174: PetscCall(MatScale(a->A, aa));
1175: PetscCall(MatScale(a->B, aa));
1176: PetscFunctionReturn(PETSC_SUCCESS);
1177: }
1179: static PetscErrorCode MatView_MPIAIJ_Binary(Mat mat, PetscViewer viewer)
1180: {
1181: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1182: Mat_SeqAIJ *A = (Mat_SeqAIJ *)aij->A->data;
1183: Mat_SeqAIJ *B = (Mat_SeqAIJ *)aij->B->data;
1184: const PetscInt *garray = aij->garray;
1185: const PetscScalar *aa, *ba;
1186: PetscInt header[4], M, N, m, rs, cs, cnt, i, ja, jb;
1187: PetscInt64 nz, hnz;
1188: PetscInt *rowlens;
1189: PetscInt *colidxs;
1190: PetscScalar *matvals;
1191: PetscMPIInt rank;
1193: PetscFunctionBegin;
1194: PetscCall(PetscViewerSetUp(viewer));
1196: M = mat->rmap->N;
1197: N = mat->cmap->N;
1198: m = mat->rmap->n;
1199: rs = mat->rmap->rstart;
1200: cs = mat->cmap->rstart;
1201: nz = A->nz + B->nz;
1203: /* write matrix header */
1204: header[0] = MAT_FILE_CLASSID;
1205: header[1] = M;
1206: header[2] = N;
1207: PetscCallMPI(MPI_Reduce(&nz, &hnz, 1, MPIU_INT64, MPI_SUM, 0, PetscObjectComm((PetscObject)mat)));
1208: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)mat), &rank));
1209: if (rank == 0) PetscCall(PetscIntCast(hnz, &header[3]));
1210: PetscCall(PetscViewerBinaryWrite(viewer, header, 4, PETSC_INT));
1212: /* fill in and store row lengths */
1213: PetscCall(PetscMalloc1(m, &rowlens));
1214: for (i = 0; i < m; i++) rowlens[i] = A->i[i + 1] - A->i[i] + B->i[i + 1] - B->i[i];
1215: PetscCall(PetscViewerBinaryWriteAll(viewer, rowlens, m, rs, M, PETSC_INT));
1216: PetscCall(PetscFree(rowlens));
1218: /* fill in and store column indices */
1219: PetscCall(PetscMalloc1(nz, &colidxs));
1220: for (cnt = 0, i = 0; i < m; i++) {
1221: for (jb = B->i[i]; jb < B->i[i + 1]; jb++) {
1222: if (garray[B->j[jb]] > cs) break;
1223: colidxs[cnt++] = garray[B->j[jb]];
1224: }
1225: for (ja = A->i[i]; ja < A->i[i + 1]; ja++) colidxs[cnt++] = A->j[ja] + cs;
1226: for (; jb < B->i[i + 1]; jb++) colidxs[cnt++] = garray[B->j[jb]];
1227: }
1228: PetscCheck(cnt == nz, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Internal PETSc error: cnt = %" PetscInt_FMT " nz = %" PetscInt64_FMT, cnt, nz);
1229: PetscCall(PetscViewerBinaryWriteAll(viewer, colidxs, nz, PETSC_DETERMINE, PETSC_DETERMINE, PETSC_INT));
1230: PetscCall(PetscFree(colidxs));
1232: /* fill in and store nonzero values */
1233: PetscCall(MatSeqAIJGetArrayRead(aij->A, &aa));
1234: PetscCall(MatSeqAIJGetArrayRead(aij->B, &ba));
1235: PetscCall(PetscMalloc1(nz, &matvals));
1236: for (cnt = 0, i = 0; i < m; i++) {
1237: for (jb = B->i[i]; jb < B->i[i + 1]; jb++) {
1238: if (garray[B->j[jb]] > cs) break;
1239: matvals[cnt++] = ba[jb];
1240: }
1241: for (ja = A->i[i]; ja < A->i[i + 1]; ja++) matvals[cnt++] = aa[ja];
1242: for (; jb < B->i[i + 1]; jb++) matvals[cnt++] = ba[jb];
1243: }
1244: PetscCall(MatSeqAIJRestoreArrayRead(aij->A, &aa));
1245: PetscCall(MatSeqAIJRestoreArrayRead(aij->B, &ba));
1246: PetscCheck(cnt == nz, PETSC_COMM_SELF, PETSC_ERR_LIB, "Internal PETSc error: cnt = %" PetscInt_FMT " nz = %" PetscInt64_FMT, cnt, nz);
1247: PetscCall(PetscViewerBinaryWriteAll(viewer, matvals, nz, PETSC_DETERMINE, PETSC_DETERMINE, PETSC_SCALAR));
1248: PetscCall(PetscFree(matvals));
1250: /* write block size option to the viewer's .info file */
1251: PetscCall(MatView_Binary_BlockSizes(mat, viewer));
1252: PetscFunctionReturn(PETSC_SUCCESS);
1253: }
1255: #include <petscdraw.h>
1256: static PetscErrorCode MatView_MPIAIJ_ASCIIorDraworSocket(Mat mat, PetscViewer viewer)
1257: {
1258: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1259: PetscMPIInt rank = aij->rank, size = aij->size;
1260: PetscBool isdraw, isascii, isbinary;
1261: PetscViewer sviewer;
1262: PetscViewerFormat format;
1264: PetscFunctionBegin;
1265: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw));
1266: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii));
1267: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
1268: if (isascii) {
1269: PetscCall(PetscViewerGetFormat(viewer, &format));
1270: if (format == PETSC_VIEWER_LOAD_BALANCE) {
1271: PetscInt i, nmax = 0, nmin = PETSC_INT_MAX, navg = 0, *nz, nzlocal = ((Mat_SeqAIJ *)aij->A->data)->nz + ((Mat_SeqAIJ *)aij->B->data)->nz;
1272: PetscCall(PetscMalloc1(size, &nz));
1273: PetscCallMPI(MPI_Allgather(&nzlocal, 1, MPIU_INT, nz, 1, MPIU_INT, PetscObjectComm((PetscObject)mat)));
1274: for (i = 0; i < size; i++) {
1275: nmax = PetscMax(nmax, nz[i]);
1276: nmin = PetscMin(nmin, nz[i]);
1277: navg += nz[i];
1278: }
1279: PetscCall(PetscFree(nz));
1280: navg = navg / size;
1281: PetscCall(PetscViewerASCIIPrintf(viewer, "Load Balance - Nonzeros: Min %" PetscInt_FMT " avg %" PetscInt_FMT " max %" PetscInt_FMT "\n", nmin, navg, nmax));
1282: PetscFunctionReturn(PETSC_SUCCESS);
1283: }
1284: PetscCall(PetscViewerGetFormat(viewer, &format));
1285: if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
1286: MatInfo info;
1287: PetscInt *inodes = NULL;
1289: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)mat), &rank));
1290: PetscCall(MatGetInfo(mat, MAT_LOCAL, &info));
1291: PetscCall(MatInodeGetInodeSizes(aij->A, NULL, &inodes, NULL));
1292: PetscCall(PetscViewerASCIIPushSynchronized(viewer));
1293: if (!inodes) {
1294: 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,
1295: info.memory));
1296: } else {
1297: PetscCall(
1298: 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));
1299: }
1300: PetscCall(MatGetInfo(aij->A, MAT_LOCAL, &info));
1301: PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] on-diagonal part: nz %" PetscInt_FMT " \n", rank, (PetscInt)info.nz_used));
1302: PetscCall(MatGetInfo(aij->B, MAT_LOCAL, &info));
1303: PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] off-diagonal part: nz %" PetscInt_FMT " \n", rank, (PetscInt)info.nz_used));
1304: PetscCall(PetscViewerFlush(viewer));
1305: PetscCall(PetscViewerASCIIPopSynchronized(viewer));
1306: PetscCall(PetscViewerASCIIPrintf(viewer, "Information on VecScatter used in matrix-vector product: \n"));
1307: PetscCall(VecScatterView(aij->Mvctx, viewer));
1308: PetscFunctionReturn(PETSC_SUCCESS);
1309: } else if (format == PETSC_VIEWER_ASCII_INFO) {
1310: PetscInt inodecount, inodelimit, *inodes;
1311: PetscCall(MatInodeGetInodeSizes(aij->A, &inodecount, &inodes, &inodelimit));
1312: if (inodes) {
1313: PetscCall(PetscViewerASCIIPrintf(viewer, "using I-node (on process 0) routines: found %" PetscInt_FMT " nodes, limit used is %" PetscInt_FMT "\n", inodecount, inodelimit));
1314: } else {
1315: PetscCall(PetscViewerASCIIPrintf(viewer, "not using I-node (on process 0) routines\n"));
1316: }
1317: PetscFunctionReturn(PETSC_SUCCESS);
1318: } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) {
1319: PetscFunctionReturn(PETSC_SUCCESS);
1320: }
1321: } else if (isbinary) {
1322: if (size == 1) {
1323: PetscCall(PetscObjectSetName((PetscObject)aij->A, ((PetscObject)mat)->name));
1324: PetscCall(MatView(aij->A, viewer));
1325: } else {
1326: PetscCall(MatView_MPIAIJ_Binary(mat, viewer));
1327: }
1328: PetscFunctionReturn(PETSC_SUCCESS);
1329: } else if (isascii && size == 1) {
1330: PetscCall(PetscObjectSetName((PetscObject)aij->A, ((PetscObject)mat)->name));
1331: PetscCall(MatView(aij->A, viewer));
1332: PetscFunctionReturn(PETSC_SUCCESS);
1333: } else if (isdraw) {
1334: PetscDraw draw;
1335: PetscBool isnull;
1336: PetscCall(PetscViewerDrawGetDraw(viewer, 0, &draw));
1337: PetscCall(PetscDrawIsNull(draw, &isnull));
1338: if (isnull) PetscFunctionReturn(PETSC_SUCCESS);
1339: }
1341: { /* assemble the entire matrix onto first processor */
1342: Mat A = NULL, Av;
1343: IS isrow, iscol;
1345: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat), rank == 0 ? mat->rmap->N : 0, 0, 1, &isrow));
1346: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat), rank == 0 ? mat->cmap->N : 0, 0, 1, &iscol));
1347: PetscCall(MatCreateSubMatrix(mat, isrow, iscol, MAT_INITIAL_MATRIX, &A));
1348: PetscCall(MatMPIAIJGetSeqAIJ(A, &Av, NULL, NULL));
1349: /* The commented code uses MatCreateSubMatrices instead */
1350: /*
1351: Mat *AA, A = NULL, Av;
1352: IS isrow,iscol;
1354: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat),rank == 0 ? mat->rmap->N : 0,0,1,&isrow));
1355: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat),rank == 0 ? mat->cmap->N : 0,0,1,&iscol));
1356: PetscCall(MatCreateSubMatrices(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,&AA));
1357: if (rank == 0) {
1358: PetscCall(PetscObjectReference((PetscObject)AA[0]));
1359: A = AA[0];
1360: Av = AA[0];
1361: }
1362: PetscCall(MatDestroySubMatrices(1,&AA));
1363: */
1364: PetscCall(ISDestroy(&iscol));
1365: PetscCall(ISDestroy(&isrow));
1366: /*
1367: Everyone has to call to draw the matrix since the graphics waits are
1368: synchronized across all processors that share the PetscDraw object
1369: */
1370: PetscCall(PetscViewerGetSubViewer(viewer, PETSC_COMM_SELF, &sviewer));
1371: if (rank == 0) {
1372: if (((PetscObject)mat)->name) PetscCall(PetscObjectSetName((PetscObject)Av, ((PetscObject)mat)->name));
1373: PetscCall(MatView_SeqAIJ(Av, sviewer));
1374: }
1375: PetscCall(PetscViewerRestoreSubViewer(viewer, PETSC_COMM_SELF, &sviewer));
1376: PetscCall(MatDestroy(&A));
1377: }
1378: PetscFunctionReturn(PETSC_SUCCESS);
1379: }
1381: PetscErrorCode MatView_MPIAIJ(Mat mat, PetscViewer viewer)
1382: {
1383: PetscBool isascii, isdraw, issocket, isbinary;
1385: PetscFunctionBegin;
1386: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii));
1387: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw));
1388: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
1389: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERSOCKET, &issocket));
1390: if (isascii || isdraw || isbinary || issocket) PetscCall(MatView_MPIAIJ_ASCIIorDraworSocket(mat, viewer));
1391: PetscFunctionReturn(PETSC_SUCCESS);
1392: }
1394: static PetscErrorCode MatSOR_MPIAIJ(Mat matin, Vec bb, PetscReal omega, MatSORType flag, PetscReal fshift, PetscInt its, PetscInt lits, Vec xx)
1395: {
1396: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)matin->data;
1397: Vec bb1 = NULL;
1398: PetscBool hasop;
1400: PetscFunctionBegin;
1401: if (flag == SOR_APPLY_UPPER) {
1402: PetscUseTypeMethod(mat->A, sor, bb, omega, flag, fshift, lits, 1, xx);
1403: PetscFunctionReturn(PETSC_SUCCESS);
1404: }
1406: if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS || flag & SOR_EISENSTAT) PetscCall(VecDuplicate(bb, &bb1));
1408: if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) {
1409: if (flag & SOR_ZERO_INITIAL_GUESS) {
1410: PetscUseTypeMethod(mat->A, sor, bb, omega, flag, fshift, lits, 1, xx);
1411: its--;
1412: }
1414: while (its--) {
1415: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1416: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1418: /* update rhs: bb1 = bb - B*x */
1419: PetscCall(VecScale(mat->lvec, -1.0));
1420: PetscUseTypeMethod(mat->B, multadd, mat->lvec, bb, bb1);
1422: /* local sweep */
1423: PetscUseTypeMethod(mat->A, sor, bb1, omega, SOR_SYMMETRIC_SWEEP, fshift, lits, 1, xx);
1424: }
1425: } else if (flag & SOR_LOCAL_FORWARD_SWEEP) {
1426: if (flag & SOR_ZERO_INITIAL_GUESS) {
1427: PetscUseTypeMethod(mat->A, sor, bb, omega, flag, fshift, lits, 1, xx);
1428: its--;
1429: }
1430: while (its--) {
1431: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1432: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1434: /* update rhs: bb1 = bb - B*x */
1435: PetscCall(VecScale(mat->lvec, -1.0));
1436: PetscUseTypeMethod(mat->B, multadd, mat->lvec, bb, bb1);
1438: /* local sweep */
1439: PetscUseTypeMethod(mat->A, sor, bb1, omega, SOR_FORWARD_SWEEP, fshift, lits, 1, xx);
1440: }
1441: } else if (flag & SOR_LOCAL_BACKWARD_SWEEP) {
1442: if (flag & SOR_ZERO_INITIAL_GUESS) {
1443: PetscUseTypeMethod(mat->A, sor, bb, omega, flag, fshift, lits, 1, xx);
1444: its--;
1445: }
1446: while (its--) {
1447: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1448: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1450: /* update rhs: bb1 = bb - B*x */
1451: PetscCall(VecScale(mat->lvec, -1.0));
1452: PetscUseTypeMethod(mat->B, multadd, mat->lvec, bb, bb1);
1454: /* local sweep */
1455: PetscUseTypeMethod(mat->A, sor, bb1, omega, SOR_BACKWARD_SWEEP, fshift, lits, 1, xx);
1456: }
1457: } else if (flag & SOR_EISENSTAT) {
1458: Vec xx1;
1460: PetscCall(VecDuplicate(bb, &xx1));
1461: PetscUseTypeMethod(mat->A, sor, bb, omega, (MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_BACKWARD_SWEEP), fshift, lits, 1, xx);
1463: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1464: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1465: if (!mat->diag) {
1466: PetscCall(MatCreateVecs(matin, &mat->diag, NULL));
1467: PetscCall(MatGetDiagonal(matin, mat->diag));
1468: }
1469: PetscCall(MatHasOperation(matin, MATOP_MULT_DIAGONAL_BLOCK, &hasop));
1470: if (hasop) PetscCall(MatMultDiagonalBlock(matin, xx, bb1));
1471: else PetscCall(VecPointwiseMult(bb1, mat->diag, xx));
1472: PetscCall(VecAYPX(bb1, (omega - 2.0) / omega, bb));
1474: PetscCall(MatMultAdd(mat->B, mat->lvec, bb1, bb1));
1476: /* local sweep */
1477: PetscUseTypeMethod(mat->A, sor, bb1, omega, (MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_FORWARD_SWEEP), fshift, lits, 1, xx1);
1478: PetscCall(VecAXPY(xx, 1.0, xx1));
1479: PetscCall(VecDestroy(&xx1));
1480: } else SETERRQ(PetscObjectComm((PetscObject)matin), PETSC_ERR_SUP, "Parallel SOR not supported");
1482: PetscCall(VecDestroy(&bb1));
1484: matin->factorerrortype = mat->A->factorerrortype;
1485: PetscFunctionReturn(PETSC_SUCCESS);
1486: }
1488: static PetscErrorCode MatPermute_MPIAIJ(Mat A, IS rowp, IS colp, Mat *B)
1489: {
1490: Mat aA, aB, Aperm;
1491: const PetscInt *rwant, *cwant, *gcols, *ai, *bi, *aj, *bj;
1492: PetscScalar *aa, *ba;
1493: PetscInt i, j, m, n, ng, anz, bnz, *dnnz, *onnz, *tdnnz, *tonnz, *rdest, *cdest, *work, *gcdest;
1494: PetscSF rowsf, sf;
1495: IS parcolp = NULL;
1496: PetscBool done;
1498: PetscFunctionBegin;
1499: PetscCall(MatGetLocalSize(A, &m, &n));
1500: PetscCall(ISGetIndices(rowp, &rwant));
1501: PetscCall(ISGetIndices(colp, &cwant));
1502: PetscCall(PetscMalloc3(PetscMax(m, n), &work, m, &rdest, n, &cdest));
1504: /* Invert row permutation to find out where my rows should go */
1505: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &rowsf));
1506: PetscCall(PetscSFSetGraphLayout(rowsf, A->rmap, A->rmap->n, NULL, PETSC_OWN_POINTER, rwant));
1507: PetscCall(PetscSFSetFromOptions(rowsf));
1508: for (i = 0; i < m; i++) work[i] = A->rmap->rstart + i;
1509: PetscCall(PetscSFReduceBegin(rowsf, MPIU_INT, work, rdest, MPI_REPLACE));
1510: PetscCall(PetscSFReduceEnd(rowsf, MPIU_INT, work, rdest, MPI_REPLACE));
1512: /* Invert column permutation to find out where my columns should go */
1513: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
1514: PetscCall(PetscSFSetGraphLayout(sf, A->cmap, A->cmap->n, NULL, PETSC_OWN_POINTER, cwant));
1515: PetscCall(PetscSFSetFromOptions(sf));
1516: for (i = 0; i < n; i++) work[i] = A->cmap->rstart + i;
1517: PetscCall(PetscSFReduceBegin(sf, MPIU_INT, work, cdest, MPI_REPLACE));
1518: PetscCall(PetscSFReduceEnd(sf, MPIU_INT, work, cdest, MPI_REPLACE));
1519: PetscCall(PetscSFDestroy(&sf));
1521: PetscCall(ISRestoreIndices(rowp, &rwant));
1522: PetscCall(ISRestoreIndices(colp, &cwant));
1523: PetscCall(MatMPIAIJGetSeqAIJ(A, &aA, &aB, &gcols));
1525: /* Find out where my gcols should go */
1526: PetscCall(MatGetSize(aB, NULL, &ng));
1527: PetscCall(PetscMalloc1(ng, &gcdest));
1528: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
1529: PetscCall(PetscSFSetGraphLayout(sf, A->cmap, ng, NULL, PETSC_OWN_POINTER, gcols));
1530: PetscCall(PetscSFSetFromOptions(sf));
1531: PetscCall(PetscSFBcastBegin(sf, MPIU_INT, cdest, gcdest, MPI_REPLACE));
1532: PetscCall(PetscSFBcastEnd(sf, MPIU_INT, cdest, gcdest, MPI_REPLACE));
1533: PetscCall(PetscSFDestroy(&sf));
1535: PetscCall(PetscCalloc4(m, &dnnz, m, &onnz, m, &tdnnz, m, &tonnz));
1536: PetscCall(MatGetRowIJ(aA, 0, PETSC_FALSE, PETSC_FALSE, &anz, &ai, &aj, &done));
1537: PetscCall(MatGetRowIJ(aB, 0, PETSC_FALSE, PETSC_FALSE, &bnz, &bi, &bj, &done));
1538: for (i = 0; i < m; i++) {
1539: PetscInt row = rdest[i];
1540: PetscMPIInt rowner;
1541: PetscCall(PetscLayoutFindOwner(A->rmap, row, &rowner));
1542: for (j = ai[i]; j < ai[i + 1]; j++) {
1543: PetscInt col = cdest[aj[j]];
1544: PetscMPIInt cowner;
1545: PetscCall(PetscLayoutFindOwner(A->cmap, col, &cowner)); /* Could build an index for the columns to eliminate this search */
1546: if (rowner == cowner) dnnz[i]++;
1547: else onnz[i]++;
1548: }
1549: for (j = bi[i]; j < bi[i + 1]; j++) {
1550: PetscInt col = gcdest[bj[j]];
1551: PetscMPIInt cowner;
1552: PetscCall(PetscLayoutFindOwner(A->cmap, col, &cowner));
1553: if (rowner == cowner) dnnz[i]++;
1554: else onnz[i]++;
1555: }
1556: }
1557: PetscCall(PetscSFBcastBegin(rowsf, MPIU_INT, dnnz, tdnnz, MPI_REPLACE));
1558: PetscCall(PetscSFBcastEnd(rowsf, MPIU_INT, dnnz, tdnnz, MPI_REPLACE));
1559: PetscCall(PetscSFBcastBegin(rowsf, MPIU_INT, onnz, tonnz, MPI_REPLACE));
1560: PetscCall(PetscSFBcastEnd(rowsf, MPIU_INT, onnz, tonnz, MPI_REPLACE));
1561: PetscCall(PetscSFDestroy(&rowsf));
1563: PetscCall(MatCreateAIJ(PetscObjectComm((PetscObject)A), A->rmap->n, A->cmap->n, A->rmap->N, A->cmap->N, 0, tdnnz, 0, tonnz, &Aperm));
1564: PetscCall(MatSeqAIJGetArray(aA, &aa));
1565: PetscCall(MatSeqAIJGetArray(aB, &ba));
1566: for (i = 0; i < m; i++) {
1567: PetscInt *acols = dnnz, *bcols = onnz; /* Repurpose now-unneeded arrays */
1568: PetscInt rowlen;
1569: rowlen = ai[i + 1] - ai[i];
1570: for (PetscInt j0 = j = 0; j < rowlen; j0 = j) { /* rowlen could be larger than number of rows m, so sum in batches */
1571: for (; j < PetscMin(rowlen, j0 + m); j++) acols[j - j0] = cdest[aj[ai[i] + j]];
1572: PetscCall(MatSetValues(Aperm, 1, &rdest[i], j - j0, acols, aa + ai[i] + j0, INSERT_VALUES));
1573: }
1574: rowlen = bi[i + 1] - bi[i];
1575: for (PetscInt j0 = j = 0; j < rowlen; j0 = j) {
1576: for (; j < PetscMin(rowlen, j0 + m); j++) bcols[j - j0] = gcdest[bj[bi[i] + j]];
1577: PetscCall(MatSetValues(Aperm, 1, &rdest[i], j - j0, bcols, ba + bi[i] + j0, INSERT_VALUES));
1578: }
1579: }
1580: PetscCall(MatAssemblyBegin(Aperm, MAT_FINAL_ASSEMBLY));
1581: PetscCall(MatAssemblyEnd(Aperm, MAT_FINAL_ASSEMBLY));
1582: PetscCall(MatRestoreRowIJ(aA, 0, PETSC_FALSE, PETSC_FALSE, &anz, &ai, &aj, &done));
1583: PetscCall(MatRestoreRowIJ(aB, 0, PETSC_FALSE, PETSC_FALSE, &bnz, &bi, &bj, &done));
1584: PetscCall(MatSeqAIJRestoreArray(aA, &aa));
1585: PetscCall(MatSeqAIJRestoreArray(aB, &ba));
1586: PetscCall(PetscFree4(dnnz, onnz, tdnnz, tonnz));
1587: PetscCall(PetscFree3(work, rdest, cdest));
1588: PetscCall(PetscFree(gcdest));
1589: if (parcolp) PetscCall(ISDestroy(&colp));
1590: *B = Aperm;
1591: PetscFunctionReturn(PETSC_SUCCESS);
1592: }
1594: static PetscErrorCode MatGetGhosts_MPIAIJ(Mat mat, PetscInt *nghosts, const PetscInt *ghosts[])
1595: {
1596: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1598: PetscFunctionBegin;
1599: PetscCall(MatGetSize(aij->B, NULL, nghosts));
1600: if (ghosts) *ghosts = aij->garray;
1601: PetscFunctionReturn(PETSC_SUCCESS);
1602: }
1604: static PetscErrorCode MatGetInfo_MPIAIJ(Mat matin, MatInfoType flag, MatInfo *info)
1605: {
1606: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)matin->data;
1607: Mat A = mat->A, B = mat->B;
1608: PetscLogDouble irecv[5];
1610: PetscFunctionBegin;
1611: info->block_size = 1.0;
1612: PetscCall(MatGetInfo(A, MAT_LOCAL, info));
1614: irecv[0] = info->nz_used;
1615: irecv[1] = info->nz_allocated;
1616: irecv[2] = info->nz_unneeded;
1617: irecv[3] = info->memory;
1618: irecv[4] = info->mallocs;
1620: PetscCall(MatGetInfo(B, MAT_LOCAL, info));
1622: irecv[0] += info->nz_used;
1623: irecv[1] += info->nz_allocated;
1624: irecv[2] += info->nz_unneeded;
1625: irecv[3] += info->memory;
1626: irecv[4] += info->mallocs;
1627: if (flag == MAT_LOCAL) {
1628: info->nz_used = irecv[0];
1629: info->nz_allocated = irecv[1];
1630: info->nz_unneeded = irecv[2];
1631: info->memory = irecv[3];
1632: info->mallocs = irecv[4];
1633: } else if (flag == MAT_GLOBAL_MAX) {
1634: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, irecv, 5, MPIU_PETSCLOGDOUBLE, MPI_MAX, PetscObjectComm((PetscObject)matin)));
1636: info->nz_used = irecv[0];
1637: info->nz_allocated = irecv[1];
1638: info->nz_unneeded = irecv[2];
1639: info->memory = irecv[3];
1640: info->mallocs = irecv[4];
1641: } else if (flag == MAT_GLOBAL_SUM) {
1642: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, irecv, 5, MPIU_PETSCLOGDOUBLE, MPI_SUM, PetscObjectComm((PetscObject)matin)));
1644: info->nz_used = irecv[0];
1645: info->nz_allocated = irecv[1];
1646: info->nz_unneeded = irecv[2];
1647: info->memory = irecv[3];
1648: info->mallocs = irecv[4];
1649: }
1650: info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */
1651: info->fill_ratio_needed = 0;
1652: info->factor_mallocs = 0;
1653: PetscFunctionReturn(PETSC_SUCCESS);
1654: }
1656: PetscErrorCode MatSetOption_MPIAIJ(Mat A, MatOption op, PetscBool flg)
1657: {
1658: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1660: PetscFunctionBegin;
1661: switch (op) {
1662: case MAT_NEW_NONZERO_LOCATIONS:
1663: case MAT_NEW_NONZERO_ALLOCATION_ERR:
1664: case MAT_UNUSED_NONZERO_LOCATION_ERR:
1665: case MAT_KEEP_NONZERO_PATTERN:
1666: case MAT_NEW_NONZERO_LOCATION_ERR:
1667: case MAT_USE_INODES:
1668: case MAT_IGNORE_ZERO_ENTRIES:
1669: case MAT_FORM_EXPLICIT_TRANSPOSE:
1670: MatCheckPreallocated(A, 1);
1671: PetscCall(MatSetOption(a->A, op, flg));
1672: PetscCall(MatSetOption(a->B, op, flg));
1673: break;
1674: case MAT_ROW_ORIENTED:
1675: MatCheckPreallocated(A, 1);
1676: a->roworiented = flg;
1678: PetscCall(MatSetOption(a->A, op, flg));
1679: PetscCall(MatSetOption(a->B, op, flg));
1680: break;
1681: case MAT_IGNORE_OFF_PROC_ENTRIES:
1682: a->donotstash = flg;
1683: break;
1684: /* Symmetry flags are handled directly by MatSetOption() and they don't affect preallocation */
1685: case MAT_SPD:
1686: case MAT_SYMMETRIC:
1687: case MAT_STRUCTURALLY_SYMMETRIC:
1688: case MAT_HERMITIAN:
1689: case MAT_SYMMETRY_ETERNAL:
1690: case MAT_STRUCTURAL_SYMMETRY_ETERNAL:
1691: case MAT_SPD_ETERNAL:
1692: /* if the diagonal matrix is square it inherits some of the properties above */
1693: if (a->A && A->rmap->n == A->cmap->n) PetscCall(MatSetOption(a->A, op, flg));
1694: break;
1695: case MAT_SUBMAT_SINGLEIS:
1696: A->submat_singleis = flg;
1697: break;
1698: default:
1699: break;
1700: }
1701: PetscFunctionReturn(PETSC_SUCCESS);
1702: }
1704: PetscErrorCode MatGetRow_MPIAIJ(Mat matin, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
1705: {
1706: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)matin->data;
1707: PetscScalar *vworkA, *vworkB, **pvA, **pvB, *v_p;
1708: PetscInt i, *cworkA, *cworkB, **pcA, **pcB, cstart = matin->cmap->rstart;
1709: PetscInt nztot, nzA, nzB, lrow, rstart = matin->rmap->rstart, rend = matin->rmap->rend;
1710: PetscInt *cmap, *idx_p;
1712: PetscFunctionBegin;
1713: PetscCheck(!mat->getrowactive, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Already active");
1714: mat->getrowactive = PETSC_TRUE;
1716: if (!mat->rowvalues && (idx || v)) {
1717: /*
1718: allocate enough space to hold information from the longest row.
1719: */
1720: Mat_SeqAIJ *Aa = (Mat_SeqAIJ *)mat->A->data, *Ba = (Mat_SeqAIJ *)mat->B->data;
1721: PetscInt max = 1, tmp;
1722: for (i = 0; i < matin->rmap->n; i++) {
1723: tmp = Aa->i[i + 1] - Aa->i[i] + Ba->i[i + 1] - Ba->i[i];
1724: if (max < tmp) max = tmp;
1725: }
1726: PetscCall(PetscMalloc2(max, &mat->rowvalues, max, &mat->rowindices));
1727: }
1729: PetscCheck(row >= rstart && row < rend, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Only local rows");
1730: lrow = row - rstart;
1732: pvA = &vworkA;
1733: pcA = &cworkA;
1734: pvB = &vworkB;
1735: pcB = &cworkB;
1736: if (!v) {
1737: pvA = NULL;
1738: pvB = NULL;
1739: }
1740: if (!idx) {
1741: pcA = NULL;
1742: if (!v) pcB = NULL;
1743: }
1744: PetscUseTypeMethod(mat->A, getrow, lrow, &nzA, pcA, pvA);
1745: PetscUseTypeMethod(mat->B, getrow, lrow, &nzB, pcB, pvB);
1746: nztot = nzA + nzB;
1748: cmap = mat->garray;
1749: if (v || idx) {
1750: if (nztot) {
1751: /* Sort by increasing column numbers, assuming A and B already sorted */
1752: PetscInt imark = -1;
1753: if (v) {
1754: *v = v_p = mat->rowvalues;
1755: for (i = 0; i < nzB; i++) {
1756: if (cmap[cworkB[i]] < cstart) v_p[i] = vworkB[i];
1757: else break;
1758: }
1759: imark = i;
1760: for (i = 0; i < nzA; i++) v_p[imark + i] = vworkA[i];
1761: for (i = imark; i < nzB; i++) v_p[nzA + i] = vworkB[i];
1762: }
1763: if (idx) {
1764: *idx = idx_p = mat->rowindices;
1765: if (imark > -1) {
1766: for (i = 0; i < imark; i++) idx_p[i] = cmap[cworkB[i]];
1767: } else {
1768: for (i = 0; i < nzB; i++) {
1769: if (cmap[cworkB[i]] < cstart) idx_p[i] = cmap[cworkB[i]];
1770: else break;
1771: }
1772: imark = i;
1773: }
1774: for (i = 0; i < nzA; i++) idx_p[imark + i] = cstart + cworkA[i];
1775: for (i = imark; i < nzB; i++) idx_p[nzA + i] = cmap[cworkB[i]];
1776: }
1777: } else {
1778: if (idx) *idx = NULL;
1779: if (v) *v = NULL;
1780: }
1781: }
1782: *nz = nztot;
1783: PetscUseTypeMethod(mat->A, restorerow, lrow, &nzA, pcA, pvA);
1784: PetscUseTypeMethod(mat->B, restorerow, lrow, &nzB, pcB, pvB);
1785: PetscFunctionReturn(PETSC_SUCCESS);
1786: }
1788: PetscErrorCode MatRestoreRow_MPIAIJ(Mat mat, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
1789: {
1790: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1792: PetscFunctionBegin;
1793: PetscCheck(aij->getrowactive, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "MatGetRow() must be called first");
1794: aij->getrowactive = PETSC_FALSE;
1795: PetscFunctionReturn(PETSC_SUCCESS);
1796: }
1798: static PetscErrorCode MatNorm_MPIAIJ(Mat mat, NormType type, PetscReal *norm)
1799: {
1800: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1801: Mat_SeqAIJ *amat = (Mat_SeqAIJ *)aij->A->data, *bmat = (Mat_SeqAIJ *)aij->B->data;
1802: PetscInt i, j;
1803: PetscReal sum = 0.0;
1804: const MatScalar *v, *amata, *bmata;
1806: PetscFunctionBegin;
1807: if (aij->size == 1) {
1808: PetscCall(MatNorm(aij->A, type, norm));
1809: } else {
1810: PetscCall(MatSeqAIJGetArrayRead(aij->A, &amata));
1811: PetscCall(MatSeqAIJGetArrayRead(aij->B, &bmata));
1812: if (type == NORM_FROBENIUS) {
1813: v = amata;
1814: for (i = 0; i < amat->nz; i++) {
1815: sum += PetscRealPart(PetscConj(*v) * (*v));
1816: v++;
1817: }
1818: v = bmata;
1819: for (i = 0; i < bmat->nz; i++) {
1820: sum += PetscRealPart(PetscConj(*v) * (*v));
1821: v++;
1822: }
1823: PetscCallMPI(MPIU_Allreduce(&sum, norm, 1, MPIU_REAL, MPIU_SUM, PetscObjectComm((PetscObject)mat)));
1824: *norm = PetscSqrtReal(*norm);
1825: PetscCall(PetscLogFlops(2.0 * amat->nz + 2.0 * bmat->nz));
1826: } else if (type == NORM_1) { /* max column norm */
1827: Vec col, bcol;
1828: PetscScalar *array;
1829: PetscInt *jj, *garray = aij->garray;
1831: PetscCall(MatCreateVecs(mat, &col, NULL));
1832: PetscCall(VecGetArrayWrite(col, &array));
1833: v = amata;
1834: jj = amat->j;
1835: for (j = 0; j < amat->nz; j++) array[*jj++] += PetscAbsScalar(*v++);
1836: PetscCall(VecRestoreArrayWrite(col, &array));
1837: PetscCall(MatCreateVecs(aij->B, &bcol, NULL));
1838: PetscCall(VecGetArrayWrite(bcol, &array));
1839: v = bmata;
1840: jj = bmat->j;
1841: for (j = 0; j < bmat->nz; j++) array[*jj++] += PetscAbsScalar(*v++);
1842: PetscCall(VecSetValues(col, aij->B->cmap->n, garray, array, ADD_VALUES));
1843: PetscCall(VecRestoreArrayWrite(bcol, &array));
1844: PetscCall(VecDestroy(&bcol));
1845: PetscCall(VecAssemblyBegin(col));
1846: PetscCall(VecAssemblyEnd(col));
1847: PetscCall(VecNorm(col, NORM_INFINITY, norm));
1848: PetscCall(VecDestroy(&col));
1849: } else if (type == NORM_INFINITY) { /* max row norm */
1850: *norm = 0.0;
1851: for (j = 0; j < aij->A->rmap->n; j++) {
1852: v = PetscSafePointerPlusOffset(amata, amat->i[j]);
1853: sum = 0.0;
1854: for (i = 0; i < amat->i[j + 1] - amat->i[j]; i++) {
1855: sum += PetscAbsScalar(*v);
1856: v++;
1857: }
1858: v = PetscSafePointerPlusOffset(bmata, bmat->i[j]);
1859: for (i = 0; i < bmat->i[j + 1] - bmat->i[j]; i++) {
1860: sum += PetscAbsScalar(*v);
1861: v++;
1862: }
1863: if (sum > *norm) *norm = sum;
1864: }
1865: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, norm, 1, MPIU_REAL, MPIU_MAX, PetscObjectComm((PetscObject)mat)));
1866: PetscCall(PetscLogFlops(PetscMax(amat->nz + bmat->nz - 1, 0)));
1867: } else SETERRQ(PetscObjectComm((PetscObject)mat), PETSC_ERR_SUP, "No support for two norm");
1868: PetscCall(MatSeqAIJRestoreArrayRead(aij->A, &amata));
1869: PetscCall(MatSeqAIJRestoreArrayRead(aij->B, &bmata));
1870: }
1871: PetscFunctionReturn(PETSC_SUCCESS);
1872: }
1874: static PetscErrorCode MatTranspose_MPIAIJ(Mat A, MatReuse reuse, Mat *matout)
1875: {
1876: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data, *b;
1877: Mat_SeqAIJ *Aloc = (Mat_SeqAIJ *)a->A->data, *Bloc = (Mat_SeqAIJ *)a->B->data, *sub_B_diag;
1878: PetscInt M = A->rmap->N, N = A->cmap->N, ma, na, mb, nb, row, *cols, *cols_tmp, *B_diag_ilen, i, ncol, A_diag_ncol;
1879: const PetscInt *ai, *aj, *bi, *bj, *B_diag_i;
1880: Mat B, A_diag, *B_diag;
1881: const MatScalar *pbv, *bv;
1883: PetscFunctionBegin;
1884: if (reuse == MAT_REUSE_MATRIX) PetscCall(MatTransposeCheckNonzeroState_Private(A, *matout));
1885: ma = A->rmap->n;
1886: na = A->cmap->n;
1887: mb = a->B->rmap->n;
1888: nb = a->B->cmap->n;
1889: ai = Aloc->i;
1890: aj = Aloc->j;
1891: bi = Bloc->i;
1892: bj = Bloc->j;
1893: if (reuse == MAT_INITIAL_MATRIX || *matout == A) {
1894: PetscInt *d_nnz, *g_nnz, *o_nnz;
1895: PetscSFNode *oloc;
1896: PETSC_UNUSED PetscSF sf;
1898: PetscCall(PetscMalloc4(na, &d_nnz, na, &o_nnz, nb, &g_nnz, nb, &oloc));
1899: /* compute d_nnz for preallocation */
1900: PetscCall(PetscArrayzero(d_nnz, na));
1901: for (i = 0; i < ai[ma]; i++) d_nnz[aj[i]]++;
1902: /* compute local off-diagonal contributions */
1903: PetscCall(PetscArrayzero(g_nnz, nb));
1904: for (i = 0; i < bi[ma]; i++) g_nnz[bj[i]]++;
1905: /* map those to global */
1906: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
1907: PetscCall(PetscSFSetGraphLayout(sf, A->cmap, nb, NULL, PETSC_USE_POINTER, a->garray));
1908: PetscCall(PetscSFSetFromOptions(sf));
1909: PetscCall(PetscArrayzero(o_nnz, na));
1910: PetscCall(PetscSFReduceBegin(sf, MPIU_INT, g_nnz, o_nnz, MPI_SUM));
1911: PetscCall(PetscSFReduceEnd(sf, MPIU_INT, g_nnz, o_nnz, MPI_SUM));
1912: PetscCall(PetscSFDestroy(&sf));
1914: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &B));
1915: PetscCall(MatSetSizes(B, A->cmap->n, A->rmap->n, N, M));
1916: PetscCall(MatSetBlockSizes(B, A->cmap->bs, A->rmap->bs));
1917: PetscCall(MatSetType(B, ((PetscObject)A)->type_name));
1918: PetscCall(MatMPIAIJSetPreallocation(B, 0, d_nnz, 0, o_nnz));
1919: PetscCall(PetscFree4(d_nnz, o_nnz, g_nnz, oloc));
1920: } else {
1921: B = *matout;
1922: PetscCall(MatSetOption(B, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_TRUE));
1923: }
1925: b = (Mat_MPIAIJ *)B->data;
1926: A_diag = a->A;
1927: B_diag = &b->A;
1928: sub_B_diag = (Mat_SeqAIJ *)(*B_diag)->data;
1929: A_diag_ncol = A_diag->cmap->N;
1930: B_diag_ilen = sub_B_diag->ilen;
1931: B_diag_i = sub_B_diag->i;
1933: /* Set ilen for diagonal of B */
1934: for (i = 0; i < A_diag_ncol; i++) B_diag_ilen[i] = B_diag_i[i + 1] - B_diag_i[i];
1936: /* Transpose the diagonal part of the matrix. In contrast to the off-diagonal part, this can be done
1937: very quickly (=without using MatSetValues), because all writes are local. */
1938: PetscCall(MatTransposeSetPrecursor(A_diag, *B_diag));
1939: PetscCall(MatTranspose(A_diag, MAT_REUSE_MATRIX, B_diag));
1941: /* copy over the B part */
1942: PetscCall(PetscMalloc1(bi[mb], &cols));
1943: PetscCall(MatSeqAIJGetArrayRead(a->B, &bv));
1944: pbv = bv;
1945: row = A->rmap->rstart;
1946: for (i = 0; i < bi[mb]; i++) cols[i] = a->garray[bj[i]];
1947: cols_tmp = cols;
1948: for (i = 0; i < mb; i++) {
1949: ncol = bi[i + 1] - bi[i];
1950: PetscCall(MatSetValues(B, ncol, cols_tmp, 1, &row, pbv, INSERT_VALUES));
1951: row++;
1952: if (pbv) pbv += ncol;
1953: if (cols_tmp) cols_tmp += ncol;
1954: }
1955: PetscCall(PetscFree(cols));
1956: PetscCall(MatSeqAIJRestoreArrayRead(a->B, &bv));
1958: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
1959: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
1960: if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX) {
1961: *matout = B;
1962: } else {
1963: PetscCall(MatHeaderMerge(A, &B));
1964: }
1965: PetscFunctionReturn(PETSC_SUCCESS);
1966: }
1968: static PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat, Vec ll, Vec rr)
1969: {
1970: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1971: Mat a = aij->A, b = aij->B;
1972: PetscInt s1, s2, s3;
1974: PetscFunctionBegin;
1975: PetscCall(MatGetLocalSize(mat, &s2, &s3));
1976: if (rr) {
1977: PetscCall(VecGetLocalSize(rr, &s1));
1978: PetscCheck(s1 == s3, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "right vector non-conforming local size");
1979: /* Overlap communication with computation. */
1980: PetscCall(VecScatterBegin(aij->Mvctx, rr, aij->lvec, INSERT_VALUES, SCATTER_FORWARD));
1981: }
1982: if (ll) {
1983: PetscCall(VecGetLocalSize(ll, &s1));
1984: PetscCheck(s1 == s2, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "left vector non-conforming local size");
1985: PetscUseTypeMethod(b, diagonalscale, ll, NULL);
1986: }
1987: /* scale the diagonal block */
1988: PetscUseTypeMethod(a, diagonalscale, ll, rr);
1990: if (rr) {
1991: /* Do a scatter end and then right scale the off-diagonal block */
1992: PetscCall(VecScatterEnd(aij->Mvctx, rr, aij->lvec, INSERT_VALUES, SCATTER_FORWARD));
1993: PetscUseTypeMethod(b, diagonalscale, NULL, aij->lvec);
1994: }
1995: PetscFunctionReturn(PETSC_SUCCESS);
1996: }
1998: static PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A)
1999: {
2000: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2002: PetscFunctionBegin;
2003: PetscCall(MatSetUnfactored(a->A));
2004: PetscFunctionReturn(PETSC_SUCCESS);
2005: }
2007: static PetscErrorCode MatEqual_MPIAIJ(Mat A, Mat B, PetscBool *flag)
2008: {
2009: Mat_MPIAIJ *matB = (Mat_MPIAIJ *)B->data, *matA = (Mat_MPIAIJ *)A->data;
2010: Mat a, b, c, d;
2012: PetscFunctionBegin;
2013: a = matA->A;
2014: b = matA->B;
2015: c = matB->A;
2016: d = matB->B;
2018: PetscCall(MatEqual(a, c, flag));
2019: if (*flag) PetscCall(MatEqual(b, d, flag));
2020: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, flag, 1, MPI_C_BOOL, MPI_LAND, PetscObjectComm((PetscObject)A)));
2021: PetscFunctionReturn(PETSC_SUCCESS);
2022: }
2024: static PetscErrorCode MatCopy_MPIAIJ(Mat A, Mat B, MatStructure str)
2025: {
2026: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2027: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2029: PetscFunctionBegin;
2030: /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */
2031: if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) {
2032: /* because of the column compression in the off-processor part of the matrix a->B,
2033: the number of columns in a->B and b->B may be different, hence we cannot call
2034: the MatCopy() directly on the two parts. If need be, we can provide a more
2035: efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices
2036: then copying the submatrices */
2037: PetscCall(MatCopy_Basic(A, B, str));
2038: } else {
2039: PetscCall(MatCopy(a->A, b->A, str));
2040: PetscCall(MatCopy(a->B, b->B, str));
2041: }
2042: PetscCall(PetscObjectStateIncrease((PetscObject)B));
2043: PetscFunctionReturn(PETSC_SUCCESS);
2044: }
2046: /*
2047: Computes the number of nonzeros per row needed for preallocation when X and Y
2048: have different nonzero structure.
2049: */
2050: 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)
2051: {
2052: PetscInt i, j, k, nzx, nzy;
2054: PetscFunctionBegin;
2055: /* Set the number of nonzeros in the new matrix */
2056: for (i = 0; i < m; i++) {
2057: const PetscInt *xjj = PetscSafePointerPlusOffset(xj, xi[i]), *yjj = PetscSafePointerPlusOffset(yj, yi[i]);
2058: nzx = xi[i + 1] - xi[i];
2059: nzy = yi[i + 1] - yi[i];
2060: nnz[i] = 0;
2061: for (j = 0, k = 0; j < nzx; j++) { /* Point in X */
2062: for (; k < nzy && yltog[yjj[k]] < xltog[xjj[j]]; k++) nnz[i]++; /* Catch up to X */
2063: if (k < nzy && yltog[yjj[k]] == xltog[xjj[j]]) k++; /* Skip duplicate */
2064: nnz[i]++;
2065: }
2066: for (; k < nzy; k++) nnz[i]++;
2067: }
2068: PetscFunctionReturn(PETSC_SUCCESS);
2069: }
2071: /* This is the same as MatAXPYGetPreallocation_SeqAIJ, except that the local-to-global map is provided */
2072: static PetscErrorCode MatAXPYGetPreallocation_MPIAIJ(Mat Y, const PetscInt *yltog, Mat X, const PetscInt *xltog, PetscInt *nnz)
2073: {
2074: PetscInt m = Y->rmap->N;
2075: Mat_SeqAIJ *x = (Mat_SeqAIJ *)X->data;
2076: Mat_SeqAIJ *y = (Mat_SeqAIJ *)Y->data;
2078: PetscFunctionBegin;
2079: PetscCall(MatAXPYGetPreallocation_MPIX_private(m, x->i, x->j, xltog, y->i, y->j, yltog, nnz));
2080: PetscFunctionReturn(PETSC_SUCCESS);
2081: }
2083: static PetscErrorCode MatAXPY_MPIAIJ(Mat Y, PetscScalar a, Mat X, MatStructure str)
2084: {
2085: Mat_MPIAIJ *xx = (Mat_MPIAIJ *)X->data, *yy = (Mat_MPIAIJ *)Y->data;
2087: PetscFunctionBegin;
2088: if (str == SAME_NONZERO_PATTERN) {
2089: PetscCall(MatAXPY(yy->A, a, xx->A, str));
2090: PetscCall(MatAXPY(yy->B, a, xx->B, str));
2091: } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */
2092: PetscCall(MatAXPY_Basic(Y, a, X, str));
2093: } else {
2094: Mat B;
2095: PetscInt *nnz_d, *nnz_o;
2097: PetscCall(PetscMalloc1(yy->A->rmap->N, &nnz_d));
2098: PetscCall(PetscMalloc1(yy->B->rmap->N, &nnz_o));
2099: PetscCall(MatCreate(PetscObjectComm((PetscObject)Y), &B));
2100: PetscCall(PetscObjectSetName((PetscObject)B, ((PetscObject)Y)->name));
2101: PetscCall(MatSetLayouts(B, Y->rmap, Y->cmap));
2102: PetscCall(MatSetType(B, ((PetscObject)Y)->type_name));
2103: PetscCall(MatAXPYGetPreallocation_SeqAIJ(yy->A, xx->A, nnz_d));
2104: PetscCall(MatAXPYGetPreallocation_MPIAIJ(yy->B, yy->garray, xx->B, xx->garray, nnz_o));
2105: PetscCall(MatMPIAIJSetPreallocation(B, 0, nnz_d, 0, nnz_o));
2106: PetscCall(MatAXPY_BasicWithPreallocation(B, Y, a, X, str));
2107: PetscCall(MatHeaderMerge(Y, &B));
2108: PetscCall(PetscFree(nnz_d));
2109: PetscCall(PetscFree(nnz_o));
2110: }
2111: PetscFunctionReturn(PETSC_SUCCESS);
2112: }
2114: PETSC_INTERN PetscErrorCode MatConjugate_SeqAIJ(Mat);
2116: static PetscErrorCode MatConjugate_MPIAIJ(Mat mat)
2117: {
2118: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2120: PetscFunctionBegin;
2121: PetscCall(MatConjugate_SeqAIJ(aij->A));
2122: PetscCall(MatConjugate_SeqAIJ(aij->B));
2123: PetscFunctionReturn(PETSC_SUCCESS);
2124: }
2126: static PetscErrorCode MatRealPart_MPIAIJ(Mat A)
2127: {
2128: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2130: PetscFunctionBegin;
2131: PetscCall(MatRealPart(a->A));
2132: PetscCall(MatRealPart(a->B));
2133: PetscFunctionReturn(PETSC_SUCCESS);
2134: }
2136: static PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A)
2137: {
2138: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2140: PetscFunctionBegin;
2141: PetscCall(MatImaginaryPart(a->A));
2142: PetscCall(MatImaginaryPart(a->B));
2143: PetscFunctionReturn(PETSC_SUCCESS);
2144: }
2146: static PetscErrorCode MatGetRowMaxAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2147: {
2148: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2149: PetscInt i, *idxb = NULL, m = A->rmap->n;
2150: PetscScalar *vv;
2151: Vec vB, vA;
2152: const PetscScalar *va, *vb;
2154: PetscFunctionBegin;
2155: PetscCall(MatCreateVecs(a->A, NULL, &vA));
2156: PetscCall(MatGetRowMaxAbs(a->A, vA, idx));
2158: PetscCall(VecGetArrayRead(vA, &va));
2159: if (idx) {
2160: for (i = 0; i < m; i++) {
2161: if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart;
2162: }
2163: }
2165: PetscCall(MatCreateVecs(a->B, NULL, &vB));
2166: PetscCall(PetscMalloc1(m, &idxb));
2167: PetscCall(MatGetRowMaxAbs(a->B, vB, idxb));
2169: PetscCall(VecGetArrayWrite(v, &vv));
2170: PetscCall(VecGetArrayRead(vB, &vb));
2171: for (i = 0; i < m; i++) {
2172: if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) {
2173: vv[i] = vb[i];
2174: if (idx) idx[i] = a->garray[idxb[i]];
2175: } else {
2176: vv[i] = va[i];
2177: if (idx && PetscAbsScalar(va[i]) == PetscAbsScalar(vb[i]) && idxb[i] != -1 && idx[i] > a->garray[idxb[i]]) idx[i] = a->garray[idxb[i]];
2178: }
2179: }
2180: PetscCall(VecRestoreArrayWrite(v, &vv));
2181: PetscCall(VecRestoreArrayRead(vA, &va));
2182: PetscCall(VecRestoreArrayRead(vB, &vb));
2183: PetscCall(PetscFree(idxb));
2184: PetscCall(VecDestroy(&vA));
2185: PetscCall(VecDestroy(&vB));
2186: PetscFunctionReturn(PETSC_SUCCESS);
2187: }
2189: static PetscErrorCode MatGetRowSumAbs_MPIAIJ(Mat A, Vec v)
2190: {
2191: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2192: Vec vB, vA;
2194: PetscFunctionBegin;
2195: PetscCall(MatCreateVecs(a->A, NULL, &vA));
2196: PetscCall(MatGetRowSumAbs(a->A, vA));
2197: PetscCall(MatCreateVecs(a->B, NULL, &vB));
2198: PetscCall(MatGetRowSumAbs(a->B, vB));
2199: PetscCall(VecAXPY(vA, 1.0, vB));
2200: PetscCall(VecDestroy(&vB));
2201: PetscCall(VecCopy(vA, v));
2202: PetscCall(VecDestroy(&vA));
2203: PetscFunctionReturn(PETSC_SUCCESS);
2204: }
2206: static PetscErrorCode MatGetRowMinAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2207: {
2208: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
2209: PetscInt m = A->rmap->n, n = A->cmap->n;
2210: PetscInt cstart = A->cmap->rstart, cend = A->cmap->rend;
2211: PetscInt *cmap = mat->garray;
2212: PetscInt *diagIdx, *offdiagIdx;
2213: Vec diagV, offdiagV;
2214: PetscScalar *a, *diagA, *offdiagA;
2215: const PetscScalar *ba, *bav;
2216: PetscInt r, j, col, ncols, *bi, *bj;
2217: Mat B = mat->B;
2218: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2220: PetscFunctionBegin;
2221: /* When a process holds entire A and other processes have no entry */
2222: if (A->cmap->N == n) {
2223: PetscCall(VecGetArrayWrite(v, &diagA));
2224: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2225: PetscCall(MatGetRowMinAbs(mat->A, diagV, idx));
2226: PetscCall(VecDestroy(&diagV));
2227: PetscCall(VecRestoreArrayWrite(v, &diagA));
2228: PetscFunctionReturn(PETSC_SUCCESS);
2229: } else if (n == 0) {
2230: if (m) {
2231: PetscCall(VecGetArrayWrite(v, &a));
2232: for (r = 0; r < m; r++) {
2233: a[r] = 0.0;
2234: if (idx) idx[r] = -1;
2235: }
2236: PetscCall(VecRestoreArrayWrite(v, &a));
2237: }
2238: PetscFunctionReturn(PETSC_SUCCESS);
2239: }
2241: PetscCall(PetscMalloc2(m, &diagIdx, m, &offdiagIdx));
2242: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2243: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2244: PetscCall(MatGetRowMinAbs(mat->A, diagV, diagIdx));
2246: /* Get offdiagIdx[] for implicit 0.0 */
2247: PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2248: ba = bav;
2249: bi = b->i;
2250: bj = b->j;
2251: PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2252: for (r = 0; r < m; r++) {
2253: ncols = bi[r + 1] - bi[r];
2254: if (ncols == A->cmap->N - n) { /* Brow is dense */
2255: offdiagA[r] = *ba;
2256: offdiagIdx[r] = cmap[0];
2257: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2258: offdiagA[r] = 0.0;
2260: /* Find first hole in the cmap */
2261: for (j = 0; j < ncols; j++) {
2262: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2263: if (col > j && j < cstart) {
2264: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2265: break;
2266: } else if (col > j + n && j >= cstart) {
2267: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2268: break;
2269: }
2270: }
2271: if (j == ncols && ncols < A->cmap->N - n) {
2272: /* a hole is outside compressed Bcols */
2273: if (ncols == 0) {
2274: if (cstart) {
2275: offdiagIdx[r] = 0;
2276: } else offdiagIdx[r] = cend;
2277: } else { /* ncols > 0 */
2278: offdiagIdx[r] = cmap[ncols - 1] + 1;
2279: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2280: }
2281: }
2282: }
2284: for (j = 0; j < ncols; j++) {
2285: if (PetscAbsScalar(offdiagA[r]) > PetscAbsScalar(*ba)) {
2286: offdiagA[r] = *ba;
2287: offdiagIdx[r] = cmap[*bj];
2288: }
2289: ba++;
2290: bj++;
2291: }
2292: }
2294: PetscCall(VecGetArrayWrite(v, &a));
2295: PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2296: for (r = 0; r < m; ++r) {
2297: if (PetscAbsScalar(diagA[r]) < PetscAbsScalar(offdiagA[r])) {
2298: a[r] = diagA[r];
2299: if (idx) idx[r] = cstart + diagIdx[r];
2300: } else if (PetscAbsScalar(diagA[r]) == PetscAbsScalar(offdiagA[r])) {
2301: a[r] = diagA[r];
2302: if (idx) {
2303: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2304: idx[r] = cstart + diagIdx[r];
2305: } else idx[r] = offdiagIdx[r];
2306: }
2307: } else {
2308: a[r] = offdiagA[r];
2309: if (idx) idx[r] = offdiagIdx[r];
2310: }
2311: }
2312: PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2313: PetscCall(VecRestoreArrayWrite(v, &a));
2314: PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2315: PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2316: PetscCall(VecDestroy(&diagV));
2317: PetscCall(VecDestroy(&offdiagV));
2318: PetscCall(PetscFree2(diagIdx, offdiagIdx));
2319: PetscFunctionReturn(PETSC_SUCCESS);
2320: }
2322: static PetscErrorCode MatGetRowMin_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2323: {
2324: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
2325: PetscInt m = A->rmap->n, n = A->cmap->n;
2326: PetscInt cstart = A->cmap->rstart, cend = A->cmap->rend;
2327: PetscInt *cmap = mat->garray;
2328: PetscInt *diagIdx, *offdiagIdx;
2329: Vec diagV, offdiagV;
2330: PetscScalar *a, *diagA, *offdiagA;
2331: const PetscScalar *ba, *bav;
2332: PetscInt r, j, col, ncols, *bi, *bj;
2333: Mat B = mat->B;
2334: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2336: PetscFunctionBegin;
2337: /* When a process holds entire A and other processes have no entry */
2338: if (A->cmap->N == n) {
2339: PetscCall(VecGetArrayWrite(v, &diagA));
2340: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2341: PetscCall(MatGetRowMin(mat->A, diagV, idx));
2342: PetscCall(VecDestroy(&diagV));
2343: PetscCall(VecRestoreArrayWrite(v, &diagA));
2344: PetscFunctionReturn(PETSC_SUCCESS);
2345: } else if (n == 0) {
2346: if (m) {
2347: PetscCall(VecGetArrayWrite(v, &a));
2348: for (r = 0; r < m; r++) {
2349: a[r] = PETSC_MAX_REAL;
2350: if (idx) idx[r] = -1;
2351: }
2352: PetscCall(VecRestoreArrayWrite(v, &a));
2353: }
2354: PetscFunctionReturn(PETSC_SUCCESS);
2355: }
2357: PetscCall(PetscCalloc2(m, &diagIdx, m, &offdiagIdx));
2358: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2359: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2360: PetscCall(MatGetRowMin(mat->A, diagV, diagIdx));
2362: /* Get offdiagIdx[] for implicit 0.0 */
2363: PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2364: ba = bav;
2365: bi = b->i;
2366: bj = b->j;
2367: PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2368: for (r = 0; r < m; r++) {
2369: ncols = bi[r + 1] - bi[r];
2370: if (ncols == A->cmap->N - n) { /* Brow is dense */
2371: offdiagA[r] = *ba;
2372: offdiagIdx[r] = cmap[0];
2373: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2374: offdiagA[r] = 0.0;
2376: /* Find first hole in the cmap */
2377: for (j = 0; j < ncols; j++) {
2378: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2379: if (col > j && j < cstart) {
2380: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2381: break;
2382: } else if (col > j + n && j >= cstart) {
2383: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2384: break;
2385: }
2386: }
2387: if (j == ncols && ncols < A->cmap->N - n) {
2388: /* a hole is outside compressed Bcols */
2389: if (ncols == 0) {
2390: if (cstart) {
2391: offdiagIdx[r] = 0;
2392: } else offdiagIdx[r] = cend;
2393: } else { /* ncols > 0 */
2394: offdiagIdx[r] = cmap[ncols - 1] + 1;
2395: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2396: }
2397: }
2398: }
2400: for (j = 0; j < ncols; j++) {
2401: if (PetscRealPart(offdiagA[r]) > PetscRealPart(*ba)) {
2402: offdiagA[r] = *ba;
2403: offdiagIdx[r] = cmap[*bj];
2404: }
2405: ba++;
2406: bj++;
2407: }
2408: }
2410: PetscCall(VecGetArrayWrite(v, &a));
2411: PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2412: for (r = 0; r < m; ++r) {
2413: if (PetscRealPart(diagA[r]) < PetscRealPart(offdiagA[r])) {
2414: a[r] = diagA[r];
2415: if (idx) idx[r] = cstart + diagIdx[r];
2416: } else if (PetscRealPart(diagA[r]) == PetscRealPart(offdiagA[r])) {
2417: a[r] = diagA[r];
2418: if (idx) {
2419: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2420: idx[r] = cstart + diagIdx[r];
2421: } else idx[r] = offdiagIdx[r];
2422: }
2423: } else {
2424: a[r] = offdiagA[r];
2425: if (idx) idx[r] = offdiagIdx[r];
2426: }
2427: }
2428: PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2429: PetscCall(VecRestoreArrayWrite(v, &a));
2430: PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2431: PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2432: PetscCall(VecDestroy(&diagV));
2433: PetscCall(VecDestroy(&offdiagV));
2434: PetscCall(PetscFree2(diagIdx, offdiagIdx));
2435: PetscFunctionReturn(PETSC_SUCCESS);
2436: }
2438: static PetscErrorCode MatGetRowMax_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2439: {
2440: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
2441: PetscInt m = A->rmap->n, n = A->cmap->n;
2442: PetscInt cstart = A->cmap->rstart, cend = A->cmap->rend;
2443: PetscInt *cmap = mat->garray;
2444: PetscInt *diagIdx, *offdiagIdx;
2445: Vec diagV, offdiagV;
2446: PetscScalar *a, *diagA, *offdiagA;
2447: const PetscScalar *ba, *bav;
2448: PetscInt r, j, col, ncols, *bi, *bj;
2449: Mat B = mat->B;
2450: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2452: PetscFunctionBegin;
2453: /* When a process holds entire A and other processes have no entry */
2454: if (A->cmap->N == n) {
2455: PetscCall(VecGetArrayWrite(v, &diagA));
2456: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2457: PetscCall(MatGetRowMax(mat->A, diagV, idx));
2458: PetscCall(VecDestroy(&diagV));
2459: PetscCall(VecRestoreArrayWrite(v, &diagA));
2460: PetscFunctionReturn(PETSC_SUCCESS);
2461: } else if (n == 0) {
2462: if (m) {
2463: PetscCall(VecGetArrayWrite(v, &a));
2464: for (r = 0; r < m; r++) {
2465: a[r] = PETSC_MIN_REAL;
2466: if (idx) idx[r] = -1;
2467: }
2468: PetscCall(VecRestoreArrayWrite(v, &a));
2469: }
2470: PetscFunctionReturn(PETSC_SUCCESS);
2471: }
2473: PetscCall(PetscMalloc2(m, &diagIdx, m, &offdiagIdx));
2474: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2475: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2476: PetscCall(MatGetRowMax(mat->A, diagV, diagIdx));
2478: /* Get offdiagIdx[] for implicit 0.0 */
2479: PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2480: ba = bav;
2481: bi = b->i;
2482: bj = b->j;
2483: PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2484: for (r = 0; r < m; r++) {
2485: ncols = bi[r + 1] - bi[r];
2486: if (ncols == A->cmap->N - n) { /* Brow is dense */
2487: offdiagA[r] = *ba;
2488: offdiagIdx[r] = cmap[0];
2489: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2490: offdiagA[r] = 0.0;
2492: /* Find first hole in the cmap */
2493: for (j = 0; j < ncols; j++) {
2494: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2495: if (col > j && j < cstart) {
2496: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2497: break;
2498: } else if (col > j + n && j >= cstart) {
2499: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2500: break;
2501: }
2502: }
2503: if (j == ncols && ncols < A->cmap->N - n) {
2504: /* a hole is outside compressed Bcols */
2505: if (ncols == 0) {
2506: if (cstart) {
2507: offdiagIdx[r] = 0;
2508: } else offdiagIdx[r] = cend;
2509: } else { /* ncols > 0 */
2510: offdiagIdx[r] = cmap[ncols - 1] + 1;
2511: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2512: }
2513: }
2514: }
2516: for (j = 0; j < ncols; j++) {
2517: if (PetscRealPart(offdiagA[r]) < PetscRealPart(*ba)) {
2518: offdiagA[r] = *ba;
2519: offdiagIdx[r] = cmap[*bj];
2520: }
2521: ba++;
2522: bj++;
2523: }
2524: }
2526: PetscCall(VecGetArrayWrite(v, &a));
2527: PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2528: for (r = 0; r < m; ++r) {
2529: if (PetscRealPart(diagA[r]) > PetscRealPart(offdiagA[r])) {
2530: a[r] = diagA[r];
2531: if (idx) idx[r] = cstart + diagIdx[r];
2532: } else if (PetscRealPart(diagA[r]) == PetscRealPart(offdiagA[r])) {
2533: a[r] = diagA[r];
2534: if (idx) {
2535: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2536: idx[r] = cstart + diagIdx[r];
2537: } else idx[r] = offdiagIdx[r];
2538: }
2539: } else {
2540: a[r] = offdiagA[r];
2541: if (idx) idx[r] = offdiagIdx[r];
2542: }
2543: }
2544: PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2545: PetscCall(VecRestoreArrayWrite(v, &a));
2546: PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2547: PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2548: PetscCall(VecDestroy(&diagV));
2549: PetscCall(VecDestroy(&offdiagV));
2550: PetscCall(PetscFree2(diagIdx, offdiagIdx));
2551: PetscFunctionReturn(PETSC_SUCCESS);
2552: }
2554: PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat mat, Mat *newmat)
2555: {
2556: Mat *dummy;
2558: PetscFunctionBegin;
2559: PetscCall(MatCreateSubMatrix_MPIAIJ_All(mat, MAT_DO_NOT_GET_VALUES, MAT_INITIAL_MATRIX, &dummy));
2560: *newmat = *dummy;
2561: PetscCall(PetscFree(dummy));
2562: PetscFunctionReturn(PETSC_SUCCESS);
2563: }
2565: static PetscErrorCode MatInvertBlockDiagonal_MPIAIJ(Mat A, const PetscScalar **values)
2566: {
2567: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2569: PetscFunctionBegin;
2570: PetscCall(MatInvertBlockDiagonal(a->A, values));
2571: A->factorerrortype = a->A->factorerrortype;
2572: PetscFunctionReturn(PETSC_SUCCESS);
2573: }
2575: static PetscErrorCode MatSetRandom_MPIAIJ(Mat x, PetscRandom rctx)
2576: {
2577: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)x->data;
2579: PetscFunctionBegin;
2580: PetscCheck(x->assembled || x->preallocated, PetscObjectComm((PetscObject)x), PETSC_ERR_ARG_WRONGSTATE, "MatSetRandom on an unassembled and unpreallocated MATMPIAIJ is not allowed");
2581: PetscCall(MatSetRandom(aij->A, rctx));
2582: if (x->assembled) {
2583: PetscCall(MatSetRandom(aij->B, rctx));
2584: } else {
2585: PetscCall(MatSetRandomSkipColumnRange_SeqAIJ_Private(aij->B, x->cmap->rstart, x->cmap->rend, rctx));
2586: }
2587: PetscCall(MatAssemblyBegin(x, MAT_FINAL_ASSEMBLY));
2588: PetscCall(MatAssemblyEnd(x, MAT_FINAL_ASSEMBLY));
2589: PetscFunctionReturn(PETSC_SUCCESS);
2590: }
2592: static PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ(Mat A, PetscBool sc)
2593: {
2594: PetscFunctionBegin;
2595: if (sc) A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ_Scalable;
2596: else A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ;
2597: PetscFunctionReturn(PETSC_SUCCESS);
2598: }
2600: /*@
2601: MatMPIAIJGetNumberNonzeros - gets the number of nonzeros in the matrix on this MPI rank
2603: Not Collective
2605: Input Parameter:
2606: . A - the matrix
2608: Output Parameter:
2609: . nz - the number of nonzeros
2611: Level: advanced
2613: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`
2614: @*/
2615: PetscErrorCode MatMPIAIJGetNumberNonzeros(Mat A, PetscCount *nz)
2616: {
2617: Mat_MPIAIJ *maij = (Mat_MPIAIJ *)A->data;
2618: Mat_SeqAIJ *aaij = (Mat_SeqAIJ *)maij->A->data, *baij = (Mat_SeqAIJ *)maij->B->data;
2619: PetscBool isaij;
2621: PetscFunctionBegin;
2622: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATMPIAIJ, &isaij));
2623: PetscCheck(isaij, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Not for type %s", ((PetscObject)A)->type_name);
2624: *nz = aaij->i[A->rmap->n] + baij->i[A->rmap->n];
2625: PetscFunctionReturn(PETSC_SUCCESS);
2626: }
2628: /*@
2629: MatMPIAIJSetUseScalableIncreaseOverlap - Determine if the matrix uses a scalable algorithm to compute the overlap
2631: Collective
2633: Input Parameters:
2634: + A - the matrix
2635: - sc - `PETSC_TRUE` indicates use the scalable algorithm (default is not to use the scalable algorithm)
2637: Level: advanced
2639: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`
2640: @*/
2641: PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap(Mat A, PetscBool sc)
2642: {
2643: PetscFunctionBegin;
2644: PetscTryMethod(A, "MatMPIAIJSetUseScalableIncreaseOverlap_C", (Mat, PetscBool), (A, sc));
2645: PetscFunctionReturn(PETSC_SUCCESS);
2646: }
2648: PetscErrorCode MatSetFromOptions_MPIAIJ(Mat A, PetscOptionItems PetscOptionsObject)
2649: {
2650: PetscBool sc = PETSC_FALSE, flg;
2652: PetscFunctionBegin;
2653: PetscOptionsHeadBegin(PetscOptionsObject, "MPIAIJ options");
2654: if (A->ops->increaseoverlap == MatIncreaseOverlap_MPIAIJ_Scalable) sc = PETSC_TRUE;
2655: PetscCall(PetscOptionsBool("-mat_increase_overlap_scalable", "Use a scalable algorithm to compute the overlap", "MatIncreaseOverlap", sc, &sc, &flg));
2656: if (flg) PetscCall(MatMPIAIJSetUseScalableIncreaseOverlap(A, sc));
2657: PetscOptionsHeadEnd();
2658: PetscFunctionReturn(PETSC_SUCCESS);
2659: }
2661: static PetscErrorCode MatShift_MPIAIJ(Mat Y, PetscScalar a)
2662: {
2663: Mat_MPIAIJ *maij = (Mat_MPIAIJ *)Y->data;
2664: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)maij->A->data;
2666: PetscFunctionBegin;
2667: if (!Y->preallocated) {
2668: PetscCall(MatMPIAIJSetPreallocation(Y, 1, NULL, 0, NULL));
2669: } else if (!aij->nz) { /* It does not matter if diagonals of Y only partially lie in maij->A. We just need an estimated preallocation. */
2670: PetscInt nonew = aij->nonew;
2671: PetscCall(MatSeqAIJSetPreallocation(maij->A, 1, NULL));
2672: aij->nonew = nonew;
2673: }
2674: PetscCall(MatShift_Basic(Y, a));
2675: PetscFunctionReturn(PETSC_SUCCESS);
2676: }
2678: static PetscErrorCode MatInvertVariableBlockDiagonal_MPIAIJ(Mat A, PetscInt nblocks, const PetscInt *bsizes, PetscScalar *diag)
2679: {
2680: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2682: PetscFunctionBegin;
2683: PetscCall(MatInvertVariableBlockDiagonal(a->A, nblocks, bsizes, diag));
2684: PetscFunctionReturn(PETSC_SUCCESS);
2685: }
2687: static PetscErrorCode MatEliminateZeros_MPIAIJ(Mat A, PetscBool keep)
2688: {
2689: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2691: PetscFunctionBegin;
2692: PetscCall(MatEliminateZeros_SeqAIJ(a->A, keep)); // possibly keep zero diagonal coefficients
2693: PetscCall(MatEliminateZeros_SeqAIJ(a->B, PETSC_FALSE)); // never keep zero diagonal coefficients
2694: PetscFunctionReturn(PETSC_SUCCESS);
2695: }
2697: static PetscErrorCode MatGetOrdering_MPIAIJ(Mat A, MatOrderingType type, IS *rperm, IS *cperm)
2698: {
2699: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2700: IS lrowperm, lcolperm;
2701: PetscInt i, rstart, rend, *idx;
2702: const PetscInt *lidx;
2704: PetscFunctionBegin;
2705: PetscCall(MatGetOrdering(a->A, type, &lrowperm, &lcolperm));
2706: PetscCall(MatGetOwnershipRange(A, &rstart, &rend));
2707: /* Remap row index set to global space */
2708: PetscCall(ISGetIndices(lrowperm, &lidx));
2709: PetscCall(PetscMalloc1(rend - rstart, &idx));
2710: for (i = 0; i + rstart < rend; i++) idx[i] = rstart + lidx[i];
2711: PetscCall(ISRestoreIndices(lrowperm, &lidx));
2712: PetscCall(ISDestroy(&lrowperm));
2713: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)A), rend - rstart, idx, PETSC_OWN_POINTER, rperm));
2714: PetscCall(ISSetPermutation(*rperm));
2715: /* Remap column index set to global space */
2716: PetscCall(ISGetIndices(lcolperm, &lidx));
2717: PetscCall(PetscMalloc1(rend - rstart, &idx));
2718: for (i = 0; i + rstart < rend; i++) idx[i] = rstart + lidx[i];
2719: PetscCall(ISRestoreIndices(lcolperm, &lidx));
2720: PetscCall(ISDestroy(&lcolperm));
2721: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)A), rend - rstart, idx, PETSC_OWN_POINTER, cperm));
2722: PetscCall(ISSetPermutation(*cperm));
2723: PetscFunctionReturn(PETSC_SUCCESS);
2724: }
2726: static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ,
2727: MatGetRow_MPIAIJ,
2728: MatRestoreRow_MPIAIJ,
2729: MatMult_MPIAIJ,
2730: /* 4*/ MatMultAdd_MPIAIJ,
2731: MatMultTranspose_MPIAIJ,
2732: MatMultTransposeAdd_MPIAIJ,
2733: NULL,
2734: NULL,
2735: NULL,
2736: /*10*/ NULL,
2737: NULL,
2738: NULL,
2739: MatSOR_MPIAIJ,
2740: MatTranspose_MPIAIJ,
2741: /*15*/ MatGetInfo_MPIAIJ,
2742: MatEqual_MPIAIJ,
2743: MatGetDiagonal_MPIAIJ,
2744: MatDiagonalScale_MPIAIJ,
2745: MatNorm_MPIAIJ,
2746: /*20*/ MatAssemblyBegin_MPIAIJ,
2747: MatAssemblyEnd_MPIAIJ,
2748: MatSetOption_MPIAIJ,
2749: MatZeroEntries_MPIAIJ,
2750: /*24*/ MatZeroRows_MPIAIJ,
2751: NULL,
2752: NULL,
2753: NULL,
2754: NULL,
2755: /*29*/ MatSetUp_MPI_Hash,
2756: NULL,
2757: NULL,
2758: MatGetDiagonalBlock_MPIAIJ,
2759: NULL,
2760: /*34*/ MatDuplicate_MPIAIJ,
2761: NULL,
2762: NULL,
2763: NULL,
2764: NULL,
2765: /*39*/ MatAXPY_MPIAIJ,
2766: MatCreateSubMatrices_MPIAIJ,
2767: MatIncreaseOverlap_MPIAIJ,
2768: MatGetValues_MPIAIJ,
2769: MatCopy_MPIAIJ,
2770: /*44*/ MatGetRowMax_MPIAIJ,
2771: MatScale_MPIAIJ,
2772: MatShift_MPIAIJ,
2773: MatDiagonalSet_MPIAIJ,
2774: MatZeroRowsColumns_MPIAIJ,
2775: /*49*/ MatSetRandom_MPIAIJ,
2776: MatGetRowIJ_MPIAIJ,
2777: MatRestoreRowIJ_MPIAIJ,
2778: NULL,
2779: NULL,
2780: /*54*/ MatFDColoringCreate_MPIXAIJ,
2781: NULL,
2782: MatSetUnfactored_MPIAIJ,
2783: MatPermute_MPIAIJ,
2784: NULL,
2785: /*59*/ MatCreateSubMatrix_MPIAIJ,
2786: MatDestroy_MPIAIJ,
2787: MatView_MPIAIJ,
2788: NULL,
2789: NULL,
2790: /*64*/ MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ,
2791: NULL,
2792: NULL,
2793: NULL,
2794: MatGetRowMaxAbs_MPIAIJ,
2795: /*69*/ MatGetRowMinAbs_MPIAIJ,
2796: NULL,
2797: NULL,
2798: MatFDColoringApply_AIJ,
2799: MatSetFromOptions_MPIAIJ,
2800: MatFindZeroDiagonals_MPIAIJ,
2801: /*75*/ NULL,
2802: NULL,
2803: NULL,
2804: MatLoad_MPIAIJ,
2805: NULL,
2806: /*80*/ NULL,
2807: NULL,
2808: NULL,
2809: /*83*/ NULL,
2810: NULL,
2811: MatMatMultNumeric_MPIAIJ_MPIAIJ,
2812: MatPtAPNumeric_MPIAIJ_MPIAIJ,
2813: NULL,
2814: NULL,
2815: /*89*/ MatBindToCPU_MPIAIJ,
2816: MatProductSetFromOptions_MPIAIJ,
2817: NULL,
2818: NULL,
2819: MatConjugate_MPIAIJ,
2820: /*94*/ NULL,
2821: MatSetValuesRow_MPIAIJ,
2822: MatRealPart_MPIAIJ,
2823: MatImaginaryPart_MPIAIJ,
2824: NULL,
2825: /*99*/ NULL,
2826: NULL,
2827: NULL,
2828: MatGetRowMin_MPIAIJ,
2829: NULL,
2830: /*104*/ MatGetSeqNonzeroStructure_MPIAIJ,
2831: NULL,
2832: MatGetGhosts_MPIAIJ,
2833: NULL,
2834: NULL,
2835: /*109*/ MatMultDiagonalBlock_MPIAIJ,
2836: NULL,
2837: NULL,
2838: NULL,
2839: MatGetMultiProcBlock_MPIAIJ,
2840: /*114*/ MatFindNonzeroRows_MPIAIJ,
2841: MatGetColumnReductions_MPIAIJ,
2842: MatInvertBlockDiagonal_MPIAIJ,
2843: MatInvertVariableBlockDiagonal_MPIAIJ,
2844: MatCreateSubMatricesMPI_MPIAIJ,
2845: /*119*/ NULL,
2846: MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ,
2847: NULL,
2848: NULL,
2849: NULL,
2850: /*124*/ NULL,
2851: MatSetBlockSizes_MPIAIJ,
2852: NULL,
2853: MatFDColoringSetUp_MPIXAIJ,
2854: MatFindOffBlockDiagonalEntries_MPIAIJ,
2855: /*129*/ MatCreateMPIMatConcatenateSeqMat_MPIAIJ,
2856: NULL,
2857: NULL,
2858: NULL,
2859: MatCreateGraph_Simple_AIJ,
2860: /*134*/ NULL,
2861: MatEliminateZeros_MPIAIJ,
2862: MatGetRowSumAbs_MPIAIJ,
2863: NULL,
2864: NULL,
2865: /*139*/ NULL,
2866: MatCopyHashToXAIJ_MPI_Hash,
2867: MatGetCurrentMemType_MPIAIJ,
2868: NULL,
2869: MatADot_Default,
2870: /*144*/ MatANorm_Default,
2871: NULL,
2872: NULL,
2873: MatGetOrdering_MPIAIJ};
2875: static PetscErrorCode MatStoreValues_MPIAIJ(Mat mat)
2876: {
2877: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2879: PetscFunctionBegin;
2880: PetscCall(MatStoreValues(aij->A));
2881: PetscCall(MatStoreValues(aij->B));
2882: PetscFunctionReturn(PETSC_SUCCESS);
2883: }
2885: static PetscErrorCode MatRetrieveValues_MPIAIJ(Mat mat)
2886: {
2887: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2889: PetscFunctionBegin;
2890: PetscCall(MatRetrieveValues(aij->A));
2891: PetscCall(MatRetrieveValues(aij->B));
2892: PetscFunctionReturn(PETSC_SUCCESS);
2893: }
2895: PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat B, PetscInt d_nz, const PetscInt d_nnz[], PetscInt o_nz, const PetscInt o_nnz[])
2896: {
2897: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2898: PetscMPIInt size;
2900: PetscFunctionBegin;
2901: if (B->hash_active) {
2902: B->ops[0] = b->cops;
2903: B->hash_active = PETSC_FALSE;
2904: }
2905: PetscCall(PetscLayoutSetUp(B->rmap));
2906: PetscCall(PetscLayoutSetUp(B->cmap));
2908: #if PetscDefined(USE_CTABLE)
2909: PetscCall(PetscHMapIDestroy(&b->colmap));
2910: #else
2911: PetscCall(PetscFree(b->colmap));
2912: #endif
2913: PetscCall(PetscFree(b->garray));
2914: PetscCall(VecDestroy(&b->lvec));
2915: PetscCall(VecScatterDestroy(&b->Mvctx));
2917: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
2919: MatSeqXAIJGetOptions_Private(b->B);
2920: PetscCall(MatDestroy(&b->B));
2921: PetscCall(MatCreate(PETSC_COMM_SELF, &b->B));
2922: PetscCall(MatSetSizes(b->B, B->rmap->n, size > 1 ? B->cmap->N : 0, B->rmap->n, size > 1 ? B->cmap->N : 0));
2923: PetscCall(MatSetBlockSizesFromMats(b->B, B, B));
2924: PetscCall(MatSetType(b->B, MATSEQAIJ));
2925: MatSeqXAIJRestoreOptions_Private(b->B);
2927: MatSeqXAIJGetOptions_Private(b->A);
2928: PetscCall(MatDestroy(&b->A));
2929: PetscCall(MatCreate(PETSC_COMM_SELF, &b->A));
2930: PetscCall(MatSetSizes(b->A, B->rmap->n, B->cmap->n, B->rmap->n, B->cmap->n));
2931: PetscCall(MatSetBlockSizesFromMats(b->A, B, B));
2932: PetscCall(MatSetType(b->A, MATSEQAIJ));
2933: MatSeqXAIJRestoreOptions_Private(b->A);
2935: PetscCall(MatSeqAIJSetPreallocation(b->A, d_nz, d_nnz));
2936: PetscCall(MatSeqAIJSetPreallocation(b->B, o_nz, o_nnz));
2937: B->preallocated = PETSC_TRUE;
2938: B->was_assembled = PETSC_FALSE;
2939: B->assembled = PETSC_FALSE;
2940: PetscFunctionReturn(PETSC_SUCCESS);
2941: }
2943: static PetscErrorCode MatResetPreallocation_MPIAIJ(Mat B)
2944: {
2945: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2946: PetscBool ondiagreset, offdiagreset, memoryreset;
2948: PetscFunctionBegin;
2950: PetscCheck(B->insertmode == NOT_SET_VALUES, PETSC_COMM_SELF, PETSC_ERR_SUP, "Cannot reset preallocation after setting some values but not yet calling MatAssemblyBegin()/MatAssemblyEnd()");
2951: if (B->num_ass == 0) PetscFunctionReturn(PETSC_SUCCESS);
2953: PetscCall(MatResetPreallocation_SeqAIJ_Private(b->A, &ondiagreset));
2954: PetscCall(MatResetPreallocation_SeqAIJ_Private(b->B, &offdiagreset));
2955: memoryreset = (PetscBool)(ondiagreset || offdiagreset);
2956: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &memoryreset, 1, MPI_C_BOOL, MPI_LOR, PetscObjectComm((PetscObject)B)));
2957: if (!memoryreset) PetscFunctionReturn(PETSC_SUCCESS);
2959: PetscCall(PetscLayoutSetUp(B->rmap));
2960: PetscCall(PetscLayoutSetUp(B->cmap));
2961: PetscCheck(B->assembled || B->was_assembled, PetscObjectComm((PetscObject)B), PETSC_ERR_ARG_WRONGSTATE, "Should not need to reset preallocation if the matrix was never assembled");
2962: PetscCall(MatDisAssemble_MPIAIJ(B, PETSC_TRUE));
2963: PetscCall(VecScatterDestroy(&b->Mvctx));
2965: B->preallocated = PETSC_TRUE;
2966: B->was_assembled = PETSC_FALSE;
2967: B->assembled = PETSC_FALSE;
2968: /* Log that the state of this object has changed; this will help guarantee that preconditioners get re-setup */
2969: PetscCall(PetscObjectStateIncrease((PetscObject)B));
2970: PetscFunctionReturn(PETSC_SUCCESS);
2971: }
2973: PetscErrorCode MatDuplicate_MPIAIJ(Mat matin, MatDuplicateOption cpvalues, Mat *newmat)
2974: {
2975: Mat mat;
2976: Mat_MPIAIJ *a, *oldmat = (Mat_MPIAIJ *)matin->data;
2978: PetscFunctionBegin;
2979: *newmat = NULL;
2980: PetscCall(MatCreate(PetscObjectComm((PetscObject)matin), &mat));
2981: PetscCall(MatSetSizes(mat, matin->rmap->n, matin->cmap->n, matin->rmap->N, matin->cmap->N));
2982: PetscCall(MatSetBlockSizesFromMats(mat, matin, matin));
2983: PetscCall(MatSetType(mat, ((PetscObject)matin)->type_name));
2984: a = (Mat_MPIAIJ *)mat->data;
2986: mat->factortype = matin->factortype;
2987: mat->assembled = matin->assembled;
2988: mat->insertmode = NOT_SET_VALUES;
2990: a->size = oldmat->size;
2991: a->rank = oldmat->rank;
2992: a->donotstash = oldmat->donotstash;
2993: a->roworiented = oldmat->roworiented;
2994: a->rowindices = NULL;
2995: a->rowvalues = NULL;
2996: a->getrowactive = PETSC_FALSE;
2998: PetscCall(PetscLayoutReference(matin->rmap, &mat->rmap));
2999: PetscCall(PetscLayoutReference(matin->cmap, &mat->cmap));
3000: if (matin->hash_active) PetscCall(MatSetUp(mat));
3001: else {
3002: mat->preallocated = matin->preallocated;
3003: if (oldmat->colmap) {
3004: #if PetscDefined(USE_CTABLE)
3005: PetscCall(PetscHMapIDuplicate(oldmat->colmap, &a->colmap));
3006: #else
3007: PetscCall(PetscMalloc1(mat->cmap->N, &a->colmap));
3008: PetscCall(PetscArraycpy(a->colmap, oldmat->colmap, mat->cmap->N));
3009: #endif
3010: } else a->colmap = NULL;
3011: if (oldmat->garray) {
3012: PetscInt len;
3013: len = oldmat->B->cmap->n;
3014: PetscCall(PetscMalloc1(len, &a->garray));
3015: if (len) PetscCall(PetscArraycpy(a->garray, oldmat->garray, len));
3016: } else a->garray = NULL;
3018: /* It may happen MatDuplicate is called with a non-assembled matrix
3019: In fact, MatDuplicate only requires the matrix to be preallocated
3020: This may happen inside a DMCreateMatrix_Shell */
3021: if (oldmat->lvec) PetscCall(VecDuplicate(oldmat->lvec, &a->lvec));
3022: if (oldmat->Mvctx) {
3023: a->Mvctx = oldmat->Mvctx;
3024: PetscCall(PetscObjectReference((PetscObject)oldmat->Mvctx));
3025: }
3026: PetscCall(MatDuplicate(oldmat->A, cpvalues, &a->A));
3027: PetscCall(MatDuplicate(oldmat->B, cpvalues, &a->B));
3028: }
3029: PetscCall(PetscFunctionListDuplicate(((PetscObject)matin)->qlist, &((PetscObject)mat)->qlist));
3030: *newmat = mat;
3031: PetscFunctionReturn(PETSC_SUCCESS);
3032: }
3034: PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer)
3035: {
3036: PetscBool isbinary, ishdf5;
3038: PetscFunctionBegin;
3041: /* force binary viewer to load .info file if it has not yet done so */
3042: PetscCall(PetscViewerSetUp(viewer));
3043: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
3044: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERHDF5, &ishdf5));
3045: if (isbinary) {
3046: PetscCall(MatLoad_MPIAIJ_Binary(newMat, viewer));
3047: } else if (ishdf5) {
3048: #if PetscDefined(HAVE_HDF5)
3049: PetscCall(MatLoad_AIJ_HDF5(newMat, viewer));
3050: #else
3051: SETERRQ(PetscObjectComm((PetscObject)newMat), PETSC_ERR_SUP, "HDF5 not supported in this build.\nPlease reconfigure using --download-hdf5");
3052: #endif
3053: } else {
3054: 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);
3055: }
3056: PetscFunctionReturn(PETSC_SUCCESS);
3057: }
3059: PetscErrorCode MatLoad_MPIAIJ_Binary(Mat mat, PetscViewer viewer)
3060: {
3061: PetscInt header[4], M, N, m, nz, rows, cols, sum, i;
3062: PetscInt *rowidxs, *colidxs;
3063: PetscScalar *matvals;
3065: PetscFunctionBegin;
3066: PetscCall(PetscViewerSetUp(viewer));
3068: /* read in matrix header */
3069: PetscCall(PetscViewerBinaryRead(viewer, header, 4, NULL, PETSC_INT));
3070: PetscCheck(header[0] == MAT_FILE_CLASSID, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Not a matrix object in file");
3071: M = header[1];
3072: N = header[2];
3073: nz = header[3];
3074: PetscCheck(M >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix row size (%" PetscInt_FMT ") in file is negative", M);
3075: PetscCheck(N >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix column size (%" PetscInt_FMT ") in file is negative", N);
3076: PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Matrix stored in special format on disk, cannot load as MPIAIJ");
3078: /* set block sizes from the viewer's .info file */
3079: PetscCall(MatLoad_Binary_BlockSizes(mat, viewer));
3080: /* set global sizes if not set already */
3081: if (mat->rmap->N < 0) mat->rmap->N = M;
3082: if (mat->cmap->N < 0) mat->cmap->N = N;
3083: PetscCall(PetscLayoutSetUp(mat->rmap));
3084: PetscCall(PetscLayoutSetUp(mat->cmap));
3086: /* check if the matrix sizes are correct */
3087: PetscCall(MatGetSize(mat, &rows, &cols));
3088: 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);
3090: /* read in row lengths and build row indices */
3091: PetscCall(MatGetLocalSize(mat, &m, NULL));
3092: PetscCall(PetscMalloc1(m + 1, &rowidxs));
3093: PetscCall(PetscViewerBinaryReadAll(viewer, rowidxs + 1, m, PETSC_DECIDE, M, PETSC_INT));
3094: rowidxs[0] = 0;
3095: for (i = 0; i < m; i++) rowidxs[i + 1] += rowidxs[i];
3096: if (nz != PETSC_INT_MAX) {
3097: PetscCallMPI(MPIU_Allreduce(&rowidxs[m], &sum, 1, MPIU_INT, MPI_SUM, PetscObjectComm((PetscObject)viewer)));
3098: 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);
3099: }
3101: /* read in column indices and matrix values */
3102: PetscCall(PetscMalloc2(rowidxs[m], &colidxs, rowidxs[m], &matvals));
3103: PetscCall(PetscViewerBinaryReadAll(viewer, colidxs, rowidxs[m], PETSC_DETERMINE, PETSC_DETERMINE, PETSC_INT));
3104: PetscCall(PetscViewerBinaryReadAll(viewer, matvals, rowidxs[m], PETSC_DETERMINE, PETSC_DETERMINE, PETSC_SCALAR));
3105: /* store matrix indices and values */
3106: PetscCall(MatMPIAIJSetPreallocationCSR(mat, rowidxs, colidxs, matvals));
3107: PetscCall(PetscFree(rowidxs));
3108: PetscCall(PetscFree2(colidxs, matvals));
3109: PetscFunctionReturn(PETSC_SUCCESS);
3110: }
3112: /* Not scalable because of ISAllGather() unless getting all columns. */
3113: static PetscErrorCode ISGetSeqIS_Private(Mat mat, IS iscol, IS *isseq)
3114: {
3115: IS iscol_local;
3116: PetscBool isstride;
3117: PetscMPIInt gisstride = 0;
3119: PetscFunctionBegin;
3120: /* check if we are grabbing all columns*/
3121: PetscCall(PetscObjectTypeCompare((PetscObject)iscol, ISSTRIDE, &isstride));
3123: if (isstride) {
3124: PetscInt start, len, mstart, mlen;
3125: PetscCall(ISStrideGetInfo(iscol, &start, NULL));
3126: PetscCall(ISGetLocalSize(iscol, &len));
3127: PetscCall(MatGetOwnershipRangeColumn(mat, &mstart, &mlen));
3128: if (mstart == start && mlen - mstart == len) gisstride = 1;
3129: }
3131: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &gisstride, 1, MPI_INT, MPI_MIN, PetscObjectComm((PetscObject)mat)));
3132: if (gisstride) {
3133: PetscInt N;
3134: PetscCall(MatGetSize(mat, NULL, &N));
3135: PetscCall(ISCreateStride(PETSC_COMM_SELF, N, 0, 1, &iscol_local));
3136: PetscCall(ISSetIdentity(iscol_local));
3137: PetscCall(PetscInfo(mat, "Optimizing for obtaining all columns of the matrix; skipping ISAllGather()\n"));
3138: } else {
3139: PetscInt cbs;
3140: PetscCall(ISGetBlockSize(iscol, &cbs));
3141: PetscCall(ISAllGather(iscol, &iscol_local));
3142: PetscCall(ISSetBlockSize(iscol_local, cbs));
3143: }
3145: *isseq = iscol_local;
3146: PetscFunctionReturn(PETSC_SUCCESS);
3147: }
3149: /*
3150: Used by MatCreateSubMatrix_MPIAIJ_SameRowColDist() to avoid ISAllGather() and global size of iscol_local
3151: (see MatCreateSubMatrix_MPIAIJ_nonscalable)
3153: Input Parameters:
3154: + mat - matrix
3155: . isrow - parallel row index set; its local indices are a subset of local columns of `mat`,
3156: i.e., mat->rstart <= isrow[i] < mat->rend
3157: - iscol - parallel column index set; its local indices are a subset of local columns of `mat`,
3158: i.e., mat->cstart <= iscol[i] < mat->cend
3160: Output Parameters:
3161: + isrow_d - sequential row index set for retrieving mat->A
3162: . iscol_d - sequential column index set for retrieving mat->A
3163: . iscol_o - sequential column index set for retrieving mat->B
3164: - garray - column map; garray[i] indicates global location of iscol_o[i] in `iscol`
3165: */
3166: static PetscErrorCode ISGetSeqIS_SameColDist_Private(Mat mat, IS isrow, IS iscol, IS *isrow_d, IS *iscol_d, IS *iscol_o, PetscInt *garray[])
3167: {
3168: Vec x, cmap;
3169: const PetscInt *is_idx;
3170: PetscScalar *xarray, *cmaparray;
3171: PetscInt ncols, isstart, *idx, m, rstart, *cmap1, count;
3172: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data;
3173: Mat B = a->B;
3174: Vec lvec = a->lvec, lcmap;
3175: PetscInt i, cstart, cend, Bn = B->cmap->N;
3176: MPI_Comm comm;
3177: VecScatter Mvctx = a->Mvctx;
3179: PetscFunctionBegin;
3180: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3181: PetscCall(ISGetLocalSize(iscol, &ncols));
3183: /* (1) iscol is a sub-column vector of mat, pad it with '-1.' to form a full vector x */
3184: PetscCall(MatCreateVecs(mat, &x, NULL));
3185: PetscCall(VecSet(x, -1.0));
3186: PetscCall(VecDuplicate(x, &cmap));
3187: PetscCall(VecSet(cmap, -1.0));
3189: /* Get start indices */
3190: PetscCallMPI(MPI_Scan(&ncols, &isstart, 1, MPIU_INT, MPI_SUM, comm));
3191: isstart -= ncols;
3192: PetscCall(MatGetOwnershipRangeColumn(mat, &cstart, &cend));
3194: PetscCall(ISGetIndices(iscol, &is_idx));
3195: PetscCall(VecGetArray(x, &xarray));
3196: PetscCall(VecGetArray(cmap, &cmaparray));
3197: PetscCall(PetscMalloc1(ncols, &idx));
3198: for (i = 0; i < ncols; i++) {
3199: xarray[is_idx[i] - cstart] = (PetscScalar)is_idx[i];
3200: cmaparray[is_idx[i] - cstart] = i + isstart; /* global index of iscol[i] */
3201: idx[i] = is_idx[i] - cstart; /* local index of iscol[i] */
3202: }
3203: PetscCall(VecRestoreArray(x, &xarray));
3204: PetscCall(VecRestoreArray(cmap, &cmaparray));
3205: PetscCall(ISRestoreIndices(iscol, &is_idx));
3207: /* Get iscol_d */
3208: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, iscol_d));
3209: PetscCall(ISGetBlockSize(iscol, &i));
3210: PetscCall(ISSetBlockSize(*iscol_d, i));
3212: /* Get isrow_d */
3213: PetscCall(ISGetLocalSize(isrow, &m));
3214: rstart = mat->rmap->rstart;
3215: PetscCall(PetscMalloc1(m, &idx));
3216: PetscCall(ISGetIndices(isrow, &is_idx));
3217: for (i = 0; i < m; i++) idx[i] = is_idx[i] - rstart;
3218: PetscCall(ISRestoreIndices(isrow, &is_idx));
3220: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, m, idx, PETSC_OWN_POINTER, isrow_d));
3221: PetscCall(ISGetBlockSize(isrow, &i));
3222: PetscCall(ISSetBlockSize(*isrow_d, i));
3224: /* (2) Scatter x and cmap using aij->Mvctx to get their off-process portions (see MatMult_MPIAIJ) */
3225: PetscCall(VecScatterBegin(Mvctx, x, lvec, INSERT_VALUES, SCATTER_FORWARD));
3226: PetscCall(VecScatterEnd(Mvctx, x, lvec, INSERT_VALUES, SCATTER_FORWARD));
3228: PetscCall(VecDuplicate(lvec, &lcmap));
3230: PetscCall(VecScatterBegin(Mvctx, cmap, lcmap, INSERT_VALUES, SCATTER_FORWARD));
3231: PetscCall(VecScatterEnd(Mvctx, cmap, lcmap, INSERT_VALUES, SCATTER_FORWARD));
3233: /* (3) create sequential iscol_o (a subset of iscol) and isgarray */
3234: /* off-process column indices */
3235: count = 0;
3236: PetscCall(PetscMalloc1(Bn, &idx));
3237: PetscCall(PetscMalloc1(Bn, &cmap1));
3239: PetscCall(VecGetArray(lvec, &xarray));
3240: PetscCall(VecGetArray(lcmap, &cmaparray));
3241: for (i = 0; i < Bn; i++) {
3242: if (PetscRealPart(xarray[i]) > -1.0) {
3243: idx[count] = i; /* local column index in off-diagonal part B */
3244: cmap1[count] = (PetscInt)PetscRealPart(cmaparray[i]); /* column index in submat */
3245: count++;
3246: }
3247: }
3248: PetscCall(VecRestoreArray(lvec, &xarray));
3249: PetscCall(VecRestoreArray(lcmap, &cmaparray));
3251: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, count, idx, PETSC_COPY_VALUES, iscol_o));
3252: /* cannot ensure iscol_o has same blocksize as iscol! */
3254: PetscCall(PetscFree(idx));
3255: *garray = cmap1;
3257: PetscCall(VecDestroy(&x));
3258: PetscCall(VecDestroy(&cmap));
3259: PetscCall(VecDestroy(&lcmap));
3260: PetscFunctionReturn(PETSC_SUCCESS);
3261: }
3263: /* isrow and iscol have same processor distribution as mat, output *submat is a submatrix of local mat */
3264: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowColDist(Mat mat, IS isrow, IS iscol, MatReuse call, Mat *submat)
3265: {
3266: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data, *asub;
3267: Mat M = NULL;
3268: MPI_Comm comm;
3269: IS iscol_d, isrow_d, iscol_o;
3270: Mat Asub = NULL, Bsub = NULL;
3271: PetscInt n, count, M_size, N_size;
3273: PetscFunctionBegin;
3274: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3276: if (call == MAT_REUSE_MATRIX) {
3277: /* Retrieve isrow_d, iscol_d and iscol_o from submat */
3278: PetscCall(PetscObjectQuery((PetscObject)*submat, "isrow_d", (PetscObject *)&isrow_d));
3279: PetscCheck(isrow_d, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "isrow_d passed in was not used before, cannot reuse");
3281: PetscCall(PetscObjectQuery((PetscObject)*submat, "iscol_d", (PetscObject *)&iscol_d));
3282: PetscCheck(iscol_d, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "iscol_d passed in was not used before, cannot reuse");
3284: PetscCall(PetscObjectQuery((PetscObject)*submat, "iscol_o", (PetscObject *)&iscol_o));
3285: PetscCheck(iscol_o, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "iscol_o passed in was not used before, cannot reuse");
3287: /* Update diagonal and off-diagonal portions of submat */
3288: asub = (Mat_MPIAIJ *)(*submat)->data;
3289: PetscCall(MatCreateSubMatrix_SeqAIJ(a->A, isrow_d, iscol_d, PETSC_DECIDE, MAT_REUSE_MATRIX, &asub->A));
3290: PetscCall(ISGetLocalSize(iscol_o, &n));
3291: if (n) PetscCall(MatCreateSubMatrix_SeqAIJ(a->B, isrow_d, iscol_o, PETSC_DECIDE, MAT_REUSE_MATRIX, &asub->B));
3292: PetscCall(MatAssemblyBegin(*submat, MAT_FINAL_ASSEMBLY));
3293: PetscCall(MatAssemblyEnd(*submat, MAT_FINAL_ASSEMBLY));
3295: } else { /* call == MAT_INITIAL_MATRIX) */
3296: PetscInt *garray, *garray_compact;
3297: PetscInt BsubN;
3299: /* Create isrow_d, iscol_d, iscol_o and isgarray (replace isgarray with array?) */
3300: PetscCall(ISGetSeqIS_SameColDist_Private(mat, isrow, iscol, &isrow_d, &iscol_d, &iscol_o, &garray));
3302: /* Create local submatrices Asub and Bsub */
3303: PetscCall(MatCreateSubMatrix_SeqAIJ(a->A, isrow_d, iscol_d, PETSC_DECIDE, MAT_INITIAL_MATRIX, &Asub));
3304: PetscCall(MatCreateSubMatrix_SeqAIJ(a->B, isrow_d, iscol_o, PETSC_DECIDE, MAT_INITIAL_MATRIX, &Bsub));
3306: // Compact garray so its not of size Bn
3307: PetscCall(ISGetSize(iscol_o, &count));
3308: PetscCall(PetscMalloc1(count, &garray_compact));
3309: PetscCall(PetscArraycpy(garray_compact, garray, count));
3311: /* Create submatrix M */
3312: PetscCall(ISGetSize(isrow, &M_size));
3313: PetscCall(ISGetSize(iscol, &N_size));
3314: PetscCall(MatCreateMPIAIJWithSeqAIJ(comm, M_size, N_size, Asub, Bsub, garray_compact, &M));
3316: /* If Bsub has empty columns, compress iscol_o such that it will retrieve condensed Bsub from a->B during reuse */
3317: asub = (Mat_MPIAIJ *)M->data;
3319: PetscCall(ISGetLocalSize(iscol_o, &BsubN));
3320: n = asub->B->cmap->N;
3321: if (BsubN > n) {
3322: /* This case can be tested using ~petsc/src/tao/bound/tutorials/runplate2_3 */
3323: const PetscInt *idx;
3324: PetscInt i, j, *idx_new, *subgarray = asub->garray;
3325: PetscCall(PetscInfo(M, "submatrix Bn %" PetscInt_FMT " != BsubN %" PetscInt_FMT ", update iscol_o\n", n, BsubN));
3327: PetscCall(PetscMalloc1(n, &idx_new));
3328: j = 0;
3329: PetscCall(ISGetIndices(iscol_o, &idx));
3330: for (i = 0; i < n; i++) {
3331: if (j >= BsubN) break;
3332: while (subgarray[i] > garray[j]) j++;
3334: 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]);
3335: idx_new[i] = idx[j++];
3336: }
3337: PetscCall(ISRestoreIndices(iscol_o, &idx));
3339: PetscCall(ISDestroy(&iscol_o));
3340: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, n, idx_new, PETSC_OWN_POINTER, &iscol_o));
3342: } 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);
3344: PetscCall(PetscFree(garray));
3345: *submat = M;
3347: /* Save isrow_d, iscol_d and iscol_o used in processor for next request */
3348: PetscCall(PetscObjectCompose((PetscObject)M, "isrow_d", (PetscObject)isrow_d));
3349: PetscCall(ISDestroy(&isrow_d));
3351: PetscCall(PetscObjectCompose((PetscObject)M, "iscol_d", (PetscObject)iscol_d));
3352: PetscCall(ISDestroy(&iscol_d));
3354: PetscCall(PetscObjectCompose((PetscObject)M, "iscol_o", (PetscObject)iscol_o));
3355: PetscCall(ISDestroy(&iscol_o));
3356: }
3357: PetscFunctionReturn(PETSC_SUCCESS);
3358: }
3360: PetscErrorCode MatCreateSubMatrix_MPIAIJ(Mat mat, IS isrow, IS iscol, MatReuse call, Mat *newmat)
3361: {
3362: IS iscol_local = NULL, isrow_d;
3363: PetscInt csize;
3364: PetscInt n, i, j, start, end;
3365: PetscBool sameRowDist = PETSC_FALSE, tsameDist[2];
3366: MPI_Comm comm;
3368: PetscFunctionBegin;
3369: /* If isrow has same processor distribution as mat,
3370: call MatCreateSubMatrix_MPIAIJ_SameRowDist() to avoid using a hash table with global size of iscol */
3371: if (call == MAT_REUSE_MATRIX) {
3372: PetscCall(PetscObjectQuery((PetscObject)*newmat, "isrow_d", (PetscObject *)&isrow_d));
3373: if (isrow_d) {
3374: sameRowDist = PETSC_TRUE;
3375: tsameDist[1] = PETSC_TRUE; /* sameColDist */
3376: } else {
3377: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_local));
3378: if (iscol_local) {
3379: sameRowDist = PETSC_TRUE;
3380: tsameDist[1] = PETSC_FALSE; /* !sameColDist */
3381: }
3382: }
3383: } else {
3384: /* Check if isrow has same processor distribution as mat */
3385: tsameDist[0] = PETSC_FALSE;
3386: PetscCall(ISGetLocalSize(isrow, &n));
3387: if (!n) {
3388: tsameDist[0] = PETSC_TRUE;
3389: } else {
3390: PetscCall(ISGetMinMax(isrow, &i, &j));
3391: PetscCall(MatGetOwnershipRange(mat, &start, &end));
3392: if (i >= start && j < end) tsameDist[0] = PETSC_TRUE;
3393: }
3395: /* Check if iscol has same processor distribution as mat */
3396: tsameDist[1] = PETSC_FALSE;
3397: PetscCall(ISGetLocalSize(iscol, &n));
3398: if (!n) {
3399: tsameDist[1] = PETSC_TRUE;
3400: } else {
3401: PetscCall(ISGetMinMax(iscol, &i, &j));
3402: PetscCall(MatGetOwnershipRangeColumn(mat, &start, &end));
3403: if (i >= start && j < end) tsameDist[1] = PETSC_TRUE;
3404: }
3406: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3407: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, tsameDist, 2, MPI_C_BOOL, MPI_LAND, comm));
3408: sameRowDist = tsameDist[0];
3409: }
3411: if (sameRowDist) {
3412: if (tsameDist[1]) { /* sameRowDist & sameColDist */
3413: /* isrow and iscol have same processor distribution as mat */
3414: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowColDist(mat, isrow, iscol, call, newmat));
3415: PetscFunctionReturn(PETSC_SUCCESS);
3416: } else { /* sameRowDist */
3417: /* isrow has same processor distribution as mat */
3418: if (call == MAT_INITIAL_MATRIX) {
3419: PetscBool sorted;
3420: PetscCall(ISGetSeqIS_Private(mat, iscol, &iscol_local));
3421: PetscCall(ISGetLocalSize(iscol_local, &n)); /* local size of iscol_local = global columns of newmat */
3422: PetscCall(ISGetSize(iscol, &i));
3423: PetscCheck(n == i, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "n %" PetscInt_FMT " != size of iscol %" PetscInt_FMT, n, i);
3425: PetscCall(ISSorted(iscol_local, &sorted));
3426: if (sorted) {
3427: /* MatCreateSubMatrix_MPIAIJ_SameRowDist() requires iscol_local be sorted; it can have duplicate indices */
3428: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowDist(mat, isrow, iscol, iscol_local, MAT_INITIAL_MATRIX, newmat));
3429: PetscFunctionReturn(PETSC_SUCCESS);
3430: }
3431: } else { /* call == MAT_REUSE_MATRIX */
3432: IS iscol_sub;
3433: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_sub));
3434: if (iscol_sub) {
3435: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowDist(mat, isrow, iscol, NULL, call, newmat));
3436: PetscFunctionReturn(PETSC_SUCCESS);
3437: }
3438: }
3439: }
3440: }
3442: /* General case: iscol -> iscol_local which has global size of iscol */
3443: if (call == MAT_REUSE_MATRIX) {
3444: PetscCall(PetscObjectQuery((PetscObject)*newmat, "ISAllGather", (PetscObject *)&iscol_local));
3445: PetscCheck(iscol_local, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3446: } else {
3447: if (!iscol_local) PetscCall(ISGetSeqIS_Private(mat, iscol, &iscol_local));
3448: }
3450: PetscCall(ISGetLocalSize(iscol, &csize));
3451: PetscCall(MatCreateSubMatrix_MPIAIJ_nonscalable(mat, isrow, iscol_local, csize, call, newmat));
3453: if (call == MAT_INITIAL_MATRIX) {
3454: PetscCall(PetscObjectCompose((PetscObject)*newmat, "ISAllGather", (PetscObject)iscol_local));
3455: PetscCall(ISDestroy(&iscol_local));
3456: }
3457: PetscFunctionReturn(PETSC_SUCCESS);
3458: }
3460: /*@C
3461: MatCreateMPIAIJWithSeqAIJ - creates a `MATMPIAIJ` matrix using `MATSEQAIJ` matrices that contain the "diagonal"
3462: and "off-diagonal" part of the matrix in CSR format.
3464: Collective
3466: Input Parameters:
3467: + comm - MPI communicator
3468: . M - the global row size
3469: . N - the global column size
3470: . A - "diagonal" portion of matrix
3471: . 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
3472: - 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.
3474: Output Parameter:
3475: . mat - the matrix, with input `A` as its local diagonal matrix
3477: Level: advanced
3479: Notes:
3480: See `MatCreateAIJ()` for the definition of "diagonal" and "off-diagonal" portion of the matrix.
3482: `A` and `B` becomes part of output mat. The user cannot use `A` and `B` anymore.
3484: 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
3485: `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
3486: `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`
3487: yourself, see algorithms in the private function `MatSetUpMultiply_MPIAIJ()`.
3489: 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.
3491: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MATSEQAIJ`, `MatCreateMPIAIJWithSplitArrays()`
3492: @*/
3493: PetscErrorCode MatCreateMPIAIJWithSeqAIJ(MPI_Comm comm, PetscInt M, PetscInt N, Mat A, Mat B, PetscInt *garray, Mat *mat)
3494: {
3495: PetscInt m, n;
3496: MatType mpi_mat_type;
3497: Mat_MPIAIJ *mpiaij;
3498: Mat C;
3500: PetscFunctionBegin;
3501: PetscCall(MatCreate(comm, &C));
3502: PetscCall(MatGetSize(A, &m, &n));
3503: PetscCheck(m == B->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Am %" PetscInt_FMT " != Bm %" PetscInt_FMT, m, B->rmap->N);
3504: 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);
3506: PetscCall(MatSetSizes(C, m, n, M, N));
3507: /* Determine the type of MPI matrix that should be created from the type of matrix A, which holds the "diagonal" portion. */
3508: PetscCall(MatGetMPIMatType_Private(A, &mpi_mat_type));
3509: PetscCall(MatSetType(C, mpi_mat_type));
3510: if (!garray) {
3511: const PetscScalar *ba;
3513: B->nonzerostate++;
3514: PetscCall(MatSeqAIJGetArrayRead(B, &ba)); /* Since we will destroy B's device copy, we need to make sure the host copy is up to date */
3515: PetscCall(MatSeqAIJRestoreArrayRead(B, &ba));
3516: }
3518: PetscCall(MatSetBlockSizes(C, A->rmap->bs, A->cmap->bs));
3519: PetscCall(PetscLayoutSetUp(C->rmap));
3520: PetscCall(PetscLayoutSetUp(C->cmap));
3522: mpiaij = (Mat_MPIAIJ *)C->data;
3523: mpiaij->A = A;
3524: mpiaij->B = B;
3525: mpiaij->garray = garray;
3526: C->preallocated = PETSC_TRUE;
3527: C->nooffprocentries = PETSC_TRUE; /* See MatAssemblyBegin_MPIAIJ. In effect, making MatAssemblyBegin a nop */
3529: PetscCall(MatSetOption(C, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
3530: PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY));
3531: /* MatAssemblyEnd is critical here. It sets mat->offloadmask according to A and B's, and
3532: also gets mpiaij->B compacted (if garray is NULL), with its col ids and size reduced
3533: */
3534: PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY));
3535: PetscCall(MatSetOption(C, MAT_NO_OFF_PROC_ENTRIES, PETSC_FALSE));
3536: PetscCall(MatSetOption(C, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
3537: *mat = C;
3538: PetscFunctionReturn(PETSC_SUCCESS);
3539: }
3541: extern PetscErrorCode MatCreateSubMatrices_MPIAIJ_SingleIS_Local(Mat, PetscInt, const IS[], const IS[], MatReuse, PetscBool, Mat *);
3543: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowDist(Mat mat, IS isrow, IS iscol, IS iscol_local, MatReuse call, Mat *newmat)
3544: {
3545: PetscInt i, m, n, rstart, row, rend, nz, j, bs, cbs;
3546: PetscInt *ii, *jj, nlocal, *dlens, *olens, dlen, olen, jend, mglobal;
3547: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data;
3548: Mat M, Msub, B = a->B;
3549: MatScalar *aa;
3550: Mat_SeqAIJ *aij;
3551: PetscInt *garray = a->garray, *colsub, Ncols;
3552: PetscInt count, Bn = B->cmap->N, cstart = mat->cmap->rstart, cend = mat->cmap->rend;
3553: IS iscol_sub, iscmap;
3554: const PetscInt *is_idx, *cmap;
3555: PetscBool allcolumns = PETSC_FALSE;
3556: MPI_Comm comm;
3558: PetscFunctionBegin;
3559: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3560: if (call == MAT_REUSE_MATRIX) {
3561: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_sub));
3562: PetscCheck(iscol_sub, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "SubIScol passed in was not used before, cannot reuse");
3563: PetscCall(ISGetLocalSize(iscol_sub, &count));
3565: PetscCall(PetscObjectQuery((PetscObject)*newmat, "Subcmap", (PetscObject *)&iscmap));
3566: PetscCheck(iscmap, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Subcmap passed in was not used before, cannot reuse");
3568: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubMatrix", (PetscObject *)&Msub));
3569: PetscCheck(Msub, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3571: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol_sub, MAT_REUSE_MATRIX, PETSC_FALSE, &Msub));
3573: } else { /* call == MAT_INITIAL_MATRIX) */
3574: PetscBool flg;
3576: PetscCall(ISGetLocalSize(iscol, &n));
3577: PetscCall(ISGetSize(iscol, &Ncols));
3579: /* (1) iscol -> nonscalable iscol_local */
3580: /* Check for special case: each processor gets entire matrix columns */
3581: PetscCall(ISIdentity(iscol_local, &flg));
3582: if (flg && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3583: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &allcolumns, 1, MPI_C_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
3584: if (allcolumns) {
3585: iscol_sub = iscol_local;
3586: PetscCall(PetscObjectReference((PetscObject)iscol_local));
3587: PetscCall(ISCreateStride(PETSC_COMM_SELF, n, 0, 1, &iscmap));
3589: } else {
3590: /* (2) iscol_local -> iscol_sub and iscmap. Implementation below requires iscol_local be sorted, it can have duplicate indices */
3591: PetscInt *idx, *cmap1, k;
3592: PetscCall(PetscMalloc1(Ncols, &idx));
3593: PetscCall(PetscMalloc1(Ncols, &cmap1));
3594: PetscCall(ISGetIndices(iscol_local, &is_idx));
3595: count = 0;
3596: k = 0;
3597: for (i = 0; i < Ncols; i++) {
3598: j = is_idx[i];
3599: if (j >= cstart && j < cend) {
3600: /* diagonal part of mat */
3601: idx[count] = j;
3602: cmap1[count++] = i; /* column index in submat */
3603: } else if (Bn) {
3604: /* off-diagonal part of mat */
3605: if (j == garray[k]) {
3606: idx[count] = j;
3607: cmap1[count++] = i; /* column index in submat */
3608: } else if (j > garray[k]) {
3609: while (j > garray[k] && k < Bn - 1) k++;
3610: if (j == garray[k]) {
3611: idx[count] = j;
3612: cmap1[count++] = i; /* column index in submat */
3613: }
3614: }
3615: }
3616: }
3617: PetscCall(ISRestoreIndices(iscol_local, &is_idx));
3619: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, count, idx, PETSC_OWN_POINTER, &iscol_sub));
3620: PetscCall(ISGetBlockSize(iscol, &cbs));
3621: PetscCall(ISSetBlockSize(iscol_sub, cbs));
3623: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)iscol_local), count, cmap1, PETSC_OWN_POINTER, &iscmap));
3624: }
3626: /* (3) Create sequential Msub */
3627: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol_sub, MAT_INITIAL_MATRIX, allcolumns, &Msub));
3628: }
3630: PetscCall(ISGetLocalSize(iscol_sub, &count));
3631: aij = (Mat_SeqAIJ *)Msub->data;
3632: ii = aij->i;
3633: PetscCall(ISGetIndices(iscmap, &cmap));
3635: /*
3636: m - number of local rows
3637: Ncols - number of columns (same on all processors)
3638: rstart - first row in new global matrix generated
3639: */
3640: PetscCall(MatGetSize(Msub, &m, NULL));
3642: if (call == MAT_INITIAL_MATRIX) {
3643: /* (4) Create parallel newmat */
3644: PetscMPIInt rank, size;
3645: PetscInt csize;
3647: PetscCallMPI(MPI_Comm_size(comm, &size));
3648: PetscCallMPI(MPI_Comm_rank(comm, &rank));
3650: /*
3651: Determine the number of non-zeros in the diagonal and off-diagonal
3652: portions of the matrix in order to do correct preallocation
3653: */
3655: /* first get start and end of "diagonal" columns */
3656: PetscCall(ISGetLocalSize(iscol, &csize));
3657: if (csize == PETSC_DECIDE) {
3658: PetscCall(ISGetSize(isrow, &mglobal));
3659: if (mglobal == Ncols) { /* square matrix */
3660: nlocal = m;
3661: } else {
3662: nlocal = Ncols / size + ((Ncols % size) > rank);
3663: }
3664: } else {
3665: nlocal = csize;
3666: }
3667: PetscCallMPI(MPI_Scan(&nlocal, &rend, 1, MPIU_INT, MPI_SUM, comm));
3668: rstart = rend - nlocal;
3669: 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);
3671: /* next, compute all the lengths */
3672: jj = aij->j;
3673: PetscCall(PetscMalloc1(2 * m + 1, &dlens));
3674: olens = dlens + m;
3675: for (i = 0; i < m; i++) {
3676: jend = ii[i + 1] - ii[i];
3677: olen = 0;
3678: dlen = 0;
3679: for (j = 0; j < jend; j++) {
3680: if (cmap[*jj] < rstart || cmap[*jj] >= rend) olen++;
3681: else dlen++;
3682: jj++;
3683: }
3684: olens[i] = olen;
3685: dlens[i] = dlen;
3686: }
3688: PetscCall(ISGetBlockSize(isrow, &bs));
3689: PetscCall(ISGetBlockSize(iscol, &cbs));
3691: PetscCall(MatCreate(comm, &M));
3692: PetscCall(MatSetSizes(M, m, nlocal, PETSC_DECIDE, Ncols));
3693: PetscCall(MatSetBlockSizes(M, bs, cbs));
3694: PetscCall(MatSetType(M, ((PetscObject)mat)->type_name));
3695: PetscCall(MatMPIAIJSetPreallocation(M, 0, dlens, 0, olens));
3696: PetscCall(PetscFree(dlens));
3698: } else { /* call == MAT_REUSE_MATRIX */
3699: M = *newmat;
3700: PetscCall(MatGetLocalSize(M, &i, NULL));
3701: PetscCheck(i == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Previous matrix must be same size/layout as request");
3702: PetscCall(MatZeroEntries(M));
3703: /*
3704: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3705: rather than the slower MatSetValues().
3706: */
3707: M->was_assembled = PETSC_TRUE;
3708: M->assembled = PETSC_FALSE;
3709: }
3711: /* (5) Set values of Msub to *newmat */
3712: PetscCall(PetscMalloc1(count, &colsub));
3713: PetscCall(MatGetOwnershipRange(M, &rstart, NULL));
3715: jj = aij->j;
3716: PetscCall(MatSeqAIJGetArrayRead(Msub, (const PetscScalar **)&aa));
3717: for (i = 0; i < m; i++) {
3718: row = rstart + i;
3719: nz = ii[i + 1] - ii[i];
3720: for (j = 0; j < nz; j++) colsub[j] = cmap[jj[j]];
3721: PetscCall(MatSetValues_MPIAIJ(M, 1, &row, nz, colsub, aa, INSERT_VALUES));
3722: jj += nz;
3723: aa += nz;
3724: }
3725: PetscCall(MatSeqAIJRestoreArrayRead(Msub, (const PetscScalar **)&aa));
3726: PetscCall(ISRestoreIndices(iscmap, &cmap));
3728: PetscCall(MatAssemblyBegin(M, MAT_FINAL_ASSEMBLY));
3729: PetscCall(MatAssemblyEnd(M, MAT_FINAL_ASSEMBLY));
3731: PetscCall(PetscFree(colsub));
3733: /* save Msub, iscol_sub and iscmap used in processor for next request */
3734: if (call == MAT_INITIAL_MATRIX) {
3735: *newmat = M;
3736: PetscCall(PetscObjectCompose((PetscObject)*newmat, "SubMatrix", (PetscObject)Msub));
3737: PetscCall(MatDestroy(&Msub));
3739: PetscCall(PetscObjectCompose((PetscObject)*newmat, "SubIScol", (PetscObject)iscol_sub));
3740: PetscCall(ISDestroy(&iscol_sub));
3742: PetscCall(PetscObjectCompose((PetscObject)*newmat, "Subcmap", (PetscObject)iscmap));
3743: PetscCall(ISDestroy(&iscmap));
3745: if (iscol_local) {
3746: PetscCall(PetscObjectCompose((PetscObject)*newmat, "ISAllGather", (PetscObject)iscol_local));
3747: PetscCall(ISDestroy(&iscol_local));
3748: }
3749: }
3750: PetscFunctionReturn(PETSC_SUCCESS);
3751: }
3753: /*
3754: Not great since it makes two copies of the submatrix, first an SeqAIJ
3755: in local and then by concatenating the local matrices the end result.
3756: Writing it directly would be much like MatCreateSubMatrices_MPIAIJ()
3758: This requires a sequential iscol with all indices.
3759: */
3760: PetscErrorCode MatCreateSubMatrix_MPIAIJ_nonscalable(Mat mat, IS isrow, IS iscol, PetscInt csize, MatReuse call, Mat *newmat)
3761: {
3762: PetscMPIInt rank, size;
3763: PetscInt i, m, n, rstart, row, rend, nz, *cwork, j, bs, cbs;
3764: PetscInt *ii, *jj, nlocal, *dlens, *olens, dlen, olen, jend, mglobal;
3765: Mat M, Mreuse;
3766: MatScalar *aa, *vwork;
3767: MPI_Comm comm;
3768: Mat_SeqAIJ *aij;
3769: PetscBool colflag, allcolumns = PETSC_FALSE;
3771: PetscFunctionBegin;
3772: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3773: PetscCallMPI(MPI_Comm_rank(comm, &rank));
3774: PetscCallMPI(MPI_Comm_size(comm, &size));
3776: /* Check for special case: each processor gets entire matrix columns */
3777: PetscCall(ISIdentity(iscol, &colflag));
3778: PetscCall(ISGetLocalSize(iscol, &n));
3779: if (colflag && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3780: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &allcolumns, 1, MPI_C_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
3782: if (call == MAT_REUSE_MATRIX) {
3783: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubMatrix", (PetscObject *)&Mreuse));
3784: PetscCheck(Mreuse, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3785: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol, MAT_REUSE_MATRIX, allcolumns, &Mreuse));
3786: } else {
3787: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol, MAT_INITIAL_MATRIX, allcolumns, &Mreuse));
3788: }
3790: /*
3791: m - number of local rows
3792: n - number of columns (same on all processors)
3793: rstart - first row in new global matrix generated
3794: */
3795: PetscCall(MatGetSize(Mreuse, &m, &n));
3796: PetscCall(MatGetBlockSizes(Mreuse, &bs, &cbs));
3797: if (call == MAT_INITIAL_MATRIX) {
3798: aij = (Mat_SeqAIJ *)Mreuse->data;
3799: ii = aij->i;
3800: jj = aij->j;
3802: /*
3803: Determine the number of non-zeros in the diagonal and off-diagonal
3804: portions of the matrix in order to do correct preallocation
3805: */
3807: /* first get start and end of "diagonal" columns */
3808: if (csize == PETSC_DECIDE) {
3809: PetscCall(ISGetSize(isrow, &mglobal));
3810: if (mglobal == n) { /* square matrix */
3811: nlocal = m;
3812: } else {
3813: nlocal = n / size + ((n % size) > rank);
3814: }
3815: } else {
3816: nlocal = csize;
3817: }
3818: PetscCallMPI(MPI_Scan(&nlocal, &rend, 1, MPIU_INT, MPI_SUM, comm));
3819: rstart = rend - nlocal;
3820: 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);
3822: /* next, compute all the lengths */
3823: PetscCall(PetscMalloc1(2 * m + 1, &dlens));
3824: olens = dlens + m;
3825: for (i = 0; i < m; i++) {
3826: jend = ii[i + 1] - ii[i];
3827: olen = 0;
3828: dlen = 0;
3829: for (j = 0; j < jend; j++) {
3830: if (*jj < rstart || *jj >= rend) olen++;
3831: else dlen++;
3832: jj++;
3833: }
3834: olens[i] = olen;
3835: dlens[i] = dlen;
3836: }
3837: PetscCall(MatCreate(comm, &M));
3838: PetscCall(MatSetSizes(M, m, nlocal, PETSC_DECIDE, n));
3839: PetscCall(MatSetBlockSizes(M, bs, cbs));
3840: PetscCall(MatSetType(M, ((PetscObject)mat)->type_name));
3841: PetscCall(MatMPIAIJSetPreallocation(M, 0, dlens, 0, olens));
3842: PetscCall(PetscFree(dlens));
3843: } else {
3844: PetscInt ml, nl;
3846: M = *newmat;
3847: PetscCall(MatGetLocalSize(M, &ml, &nl));
3848: PetscCheck(ml == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Previous matrix must be same size/layout as request");
3849: PetscCall(MatZeroEntries(M));
3850: /*
3851: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3852: rather than the slower MatSetValues().
3853: */
3854: M->was_assembled = PETSC_TRUE;
3855: M->assembled = PETSC_FALSE;
3856: }
3857: PetscCall(MatGetOwnershipRange(M, &rstart, &rend));
3858: aij = (Mat_SeqAIJ *)Mreuse->data;
3859: ii = aij->i;
3860: jj = aij->j;
3862: /* trigger copy to CPU if needed */
3863: PetscCall(MatSeqAIJGetArrayRead(Mreuse, (const PetscScalar **)&aa));
3864: for (i = 0; i < m; i++) {
3865: row = rstart + i;
3866: nz = ii[i + 1] - ii[i];
3867: cwork = jj;
3868: jj = PetscSafePointerPlusOffset(jj, nz);
3869: vwork = aa;
3870: aa = PetscSafePointerPlusOffset(aa, nz);
3871: PetscCall(MatSetValues_MPIAIJ(M, 1, &row, nz, cwork, vwork, INSERT_VALUES));
3872: }
3873: PetscCall(MatSeqAIJRestoreArrayRead(Mreuse, (const PetscScalar **)&aa));
3875: PetscCall(MatAssemblyBegin(M, MAT_FINAL_ASSEMBLY));
3876: PetscCall(MatAssemblyEnd(M, MAT_FINAL_ASSEMBLY));
3877: *newmat = M;
3879: /* save submatrix used in processor for next request */
3880: if (call == MAT_INITIAL_MATRIX) {
3881: PetscCall(PetscObjectCompose((PetscObject)M, "SubMatrix", (PetscObject)Mreuse));
3882: PetscCall(MatDestroy(&Mreuse));
3883: }
3884: PetscFunctionReturn(PETSC_SUCCESS);
3885: }
3887: static PetscErrorCode MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[])
3888: {
3889: PetscInt m, cstart, cend, j, nnz, i, d, *ld;
3890: PetscInt *d_nnz, *o_nnz, nnz_max = 0, rstart, ii, irstart;
3891: const PetscInt *JJ;
3892: PetscBool nooffprocentries;
3893: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)B->data;
3895: PetscFunctionBegin;
3896: PetscCall(PetscLayoutSetUp(B->rmap));
3897: PetscCall(PetscLayoutSetUp(B->cmap));
3898: m = B->rmap->n;
3899: cstart = B->cmap->rstart;
3900: cend = B->cmap->rend;
3901: rstart = B->rmap->rstart;
3902: irstart = Ii[0];
3904: PetscCall(PetscCalloc2(m, &d_nnz, m, &o_nnz));
3906: if (PetscDefined(USE_DEBUG)) {
3907: for (i = 0; i < m; i++) {
3908: nnz = Ii[i + 1] - Ii[i];
3909: JJ = PetscSafePointerPlusOffset(J, Ii[i] - irstart);
3910: PetscCheck(nnz >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Local row %" PetscInt_FMT " has a negative %" PetscInt_FMT " number of columns", i, nnz);
3911: 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]);
3912: 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);
3913: }
3914: }
3916: for (i = 0; i < m; i++) {
3917: nnz = Ii[i + 1] - Ii[i];
3918: JJ = PetscSafePointerPlusOffset(J, Ii[i] - irstart);
3919: nnz_max = PetscMax(nnz_max, nnz);
3920: d = 0;
3921: for (j = 0; j < nnz; j++) {
3922: if (cstart <= JJ[j] && JJ[j] < cend) d++;
3923: }
3924: d_nnz[i] = d;
3925: o_nnz[i] = nnz - d;
3926: }
3927: PetscCall(MatMPIAIJSetPreallocation(B, 0, d_nnz, 0, o_nnz));
3928: PetscCall(PetscFree2(d_nnz, o_nnz));
3930: for (i = 0; i < m; i++) {
3931: ii = i + rstart;
3932: PetscCall(MatSetValues_MPIAIJ(B, 1, &ii, Ii[i + 1] - Ii[i], PetscSafePointerPlusOffset(J, Ii[i] - irstart), PetscSafePointerPlusOffset(v, Ii[i] - irstart), INSERT_VALUES));
3933: }
3934: nooffprocentries = B->nooffprocentries;
3935: B->nooffprocentries = PETSC_TRUE;
3936: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
3937: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
3938: B->nooffprocentries = nooffprocentries;
3940: /* count number of entries below block diagonal */
3941: PetscCall(PetscFree(Aij->ld));
3942: PetscCall(PetscCalloc1(m, &ld));
3943: Aij->ld = ld;
3944: for (i = 0; i < m; i++) {
3945: nnz = Ii[i + 1] - Ii[i];
3946: j = 0;
3947: while (j < nnz && J[j] < cstart) j++;
3948: ld[i] = j;
3949: if (J) J += nnz;
3950: }
3952: PetscCall(MatSetOption(B, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
3953: PetscFunctionReturn(PETSC_SUCCESS);
3954: }
3956: /*@
3957: MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in `MATAIJ` format
3958: (the default parallel PETSc format).
3960: Collective
3962: Input Parameters:
3963: + B - the matrix
3964: . i - the indices into `j` for the start of each local row (indices start with zero)
3965: . j - the column indices for each local row (indices start with zero)
3966: - v - optional values in the matrix
3968: Level: developer
3970: Notes:
3971: The `i`, `j`, and `v` arrays ARE copied by this routine into the internal format used by PETSc;
3972: thus you CANNOT change the matrix entries by changing the values of `v` after you have
3973: called this routine. Use `MatCreateMPIAIJWithSplitArrays()` to avoid needing to copy the arrays.
3975: The `i` and `j` indices are 0 based, and `i` indices are indices corresponding to the local `j` array.
3977: A convenience routine for this functionality is `MatCreateMPIAIJWithArrays()`.
3979: You can update the matrix with new numerical values using `MatUpdateMPIAIJWithArrays()` after this call if the column indices in `j` are sorted.
3981: If you do **not** use `MatUpdateMPIAIJWithArrays()`, the column indices in `j` do not need to be sorted. If you will use
3982: `MatUpdateMPIAIJWithArrays()`, the column indices **must** be sorted.
3984: The format which is used for the sparse matrix input, is equivalent to a
3985: row-major ordering.. i.e for the following matrix, the input data expected is
3986: as shown
3987: .vb
3988: 1 0 0
3989: 2 0 3 P0
3990: -------
3991: 4 5 6 P1
3993: Process0 [P0] rows_owned=[0,1]
3994: i = {0,1,3} [size = nrow+1 = 2+1]
3995: j = {0,0,2} [size = 3]
3996: v = {1,2,3} [size = 3]
3998: Process1 [P1] rows_owned=[2]
3999: i = {0,3} [size = nrow+1 = 1+1]
4000: j = {0,1,2} [size = 3]
4001: v = {4,5,6} [size = 3]
4002: .ve
4004: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatCreateAIJ()`,
4005: `MatCreateSeqAIJWithArrays()`, `MatCreateMPIAIJWithSplitArrays()`, `MatCreateMPIAIJWithArrays()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4006: @*/
4007: PetscErrorCode MatMPIAIJSetPreallocationCSR(Mat B, const PetscInt i[], const PetscInt j[], const PetscScalar v[])
4008: {
4009: PetscFunctionBegin;
4010: PetscTryMethod(B, "MatMPIAIJSetPreallocationCSR_C", (Mat, const PetscInt[], const PetscInt[], const PetscScalar[]), (B, i, j, v));
4011: PetscFunctionReturn(PETSC_SUCCESS);
4012: }
4014: /*@
4015: MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in `MATMPIAIJ` format
4016: (the default parallel PETSc format). For good matrix assembly performance
4017: the user should preallocate the matrix storage by setting the parameters
4018: `d_nz` (or `d_nnz`) and `o_nz` (or `o_nnz`).
4020: Collective
4022: Input Parameters:
4023: + B - the matrix
4024: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
4025: (same value is used for all local rows)
4026: . d_nnz - array containing the number of nonzeros in the various rows of the
4027: DIAGONAL portion of the local submatrix (possibly different for each row)
4028: or `NULL` (`PETSC_NULL_INTEGER` in Fortran), if `d_nz` is used to specify the nonzero structure.
4029: The size of this array is equal to the number of local rows, i.e 'm'.
4030: For matrices that will be factored, you must leave room for (and set)
4031: the diagonal entry even if it is zero.
4032: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
4033: submatrix (same value is used for all local rows).
4034: - o_nnz - array containing the number of nonzeros in the various rows of the
4035: OFF-DIAGONAL portion of the local submatrix (possibly different for
4036: each row) or `NULL` (`PETSC_NULL_INTEGER` in Fortran), if `o_nz` is used to specify the nonzero
4037: structure. The size of this array is equal to the number
4038: of local rows, i.e 'm'.
4040: Example Usage:
4041: Consider the following 8x8 matrix with 34 non-zero values, that is
4042: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4043: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4044: as follows
4046: .vb
4047: 1 2 0 | 0 3 0 | 0 4
4048: Proc0 0 5 6 | 7 0 0 | 8 0
4049: 9 0 10 | 11 0 0 | 12 0
4050: -------------------------------------
4051: 13 0 14 | 15 16 17 | 0 0
4052: Proc1 0 18 0 | 19 20 21 | 0 0
4053: 0 0 0 | 22 23 0 | 24 0
4054: -------------------------------------
4055: Proc2 25 26 27 | 0 0 28 | 29 0
4056: 30 0 0 | 31 32 33 | 0 34
4057: .ve
4059: This can be represented as a collection of submatrices as
4060: .vb
4061: A B C
4062: D E F
4063: G H I
4064: .ve
4066: Where the submatrices A,B,C are owned by proc0, D,E,F are
4067: owned by proc1, G,H,I are owned by proc2.
4069: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4070: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4071: The 'M','N' parameters are 8,8, and have the same values on all procs.
4073: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4074: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4075: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4076: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4077: part as `MATSEQAIJ` matrices. For example, proc1 will store [E] as a `MATSEQAIJ`
4078: matrix, and [DF] as another `MATSEQAIJ` matrix.
4080: When `d_nz`, `o_nz` parameters are specified, `d_nz` storage elements are
4081: allocated for every row of the local DIAGONAL submatrix, and `o_nz`
4082: storage locations are allocated for every row of the OFF-DIAGONAL submatrix.
4083: One way to choose `d_nz` and `o_nz` is to use the maximum number of nonzeros over
4084: the local rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4085: In this case, the values of `d_nz`, `o_nz` are
4086: .vb
4087: proc0 dnz = 2, o_nz = 2
4088: proc1 dnz = 3, o_nz = 2
4089: proc2 dnz = 1, o_nz = 4
4090: .ve
4091: We are allocating `m`*(`d_nz`+`o_nz`) storage locations for every proc. This
4092: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4093: for proc3. i.e we are using 12+15+10=37 storage locations to store
4094: 34 values.
4096: When `d_nnz`, `o_nnz` parameters are specified, the storage is specified
4097: for every row, corresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4098: In the above case the values for `d_nnz`, `o_nnz` are
4099: .vb
4100: proc0 d_nnz = [2,2,2] and o_nnz = [2,2,2]
4101: proc1 d_nnz = [3,3,2] and o_nnz = [2,1,1]
4102: proc2 d_nnz = [1,1] and o_nnz = [4,4]
4103: .ve
4104: Here the space allocated is sum of all the above values i.e 34, and
4105: hence pre-allocation is perfect.
4107: Level: intermediate
4109: Notes:
4110: If the *_nnz parameter is given then the *_nz parameter is ignored
4112: The `MATAIJ` format, also called compressed row storage (CSR), is compatible with standard Fortran
4113: storage. The stored row and column indices begin with zero.
4114: See [Sparse Matrices](sec_matsparse) for details.
4116: The parallel matrix is partitioned such that the first m0 rows belong to
4117: process 0, the next m1 rows belong to process 1, the next m2 rows belong
4118: to process 2 etc.. where m0,m1,m2... are the input parameter 'm'.
4120: The DIAGONAL portion of the local submatrix of a processor can be defined
4121: as the submatrix which is obtained by extraction the part corresponding to
4122: the rows r1-r2 and columns c1-c2 of the global matrix, where r1 is the
4123: first row that belongs to the processor, r2 is the last row belonging to
4124: the this processor, and c1-c2 is range of indices of the local part of a
4125: vector suitable for applying the matrix to. This is an mxn matrix. In the
4126: common case of a square matrix, the row and column ranges are the same and
4127: the DIAGONAL part is also square. The remaining portion of the local
4128: submatrix (mxN) constitute the OFF-DIAGONAL portion.
4130: If `o_nnz` and `d_nnz` are specified, then `o_nz` and `d_nz` are ignored.
4132: You can call `MatGetInfo()` to get information on how effective the preallocation was;
4133: for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
4134: You can also run with the option `-info` and look for messages with the string
4135: malloc in them to see if additional memory allocation was needed.
4137: .seealso: [](ch_matrices), `Mat`, [Sparse Matrices](sec_matsparse), `MATMPIAIJ`, `MATAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatCreateAIJ()`, `MatMPIAIJSetPreallocationCSR()`,
4138: `MatGetInfo()`, `PetscSplitOwnership()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4139: @*/
4140: PetscErrorCode MatMPIAIJSetPreallocation(Mat B, PetscInt d_nz, const PetscInt d_nnz[], PetscInt o_nz, const PetscInt o_nnz[])
4141: {
4142: PetscFunctionBegin;
4145: PetscTryMethod(B, "MatMPIAIJSetPreallocation_C", (Mat, PetscInt, const PetscInt[], PetscInt, const PetscInt[]), (B, d_nz, d_nnz, o_nz, o_nnz));
4146: PetscFunctionReturn(PETSC_SUCCESS);
4147: }
4149: /*@
4150: MatCreateMPIAIJWithArrays - creates a `MATMPIAIJ` matrix using arrays that contain in standard
4151: CSR format for the local rows.
4153: Collective
4155: Input Parameters:
4156: + comm - MPI communicator
4157: . m - number of local rows (Cannot be `PETSC_DECIDE`)
4158: . n - This value should be the same as the local size used in creating the
4159: x vector for the matrix-vector product $ y = Ax$. (or `PETSC_DECIDE` to have
4160: calculated if `N` is given) For square matrices n is almost always `m`.
4161: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if `m` is given)
4162: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if `n` is given)
4163: . 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
4164: . j - global column indices
4165: - a - optional matrix values
4167: Output Parameter:
4168: . mat - the matrix
4170: Level: intermediate
4172: Notes:
4173: The `i`, `j`, and `a` arrays ARE copied by this routine into the internal format used by PETSc;
4174: thus you CANNOT change the matrix entries by changing the values of `a[]` after you have
4175: called this routine. Use `MatCreateMPIAIJWithSplitArrays()` to avoid needing to copy the arrays.
4177: The `i` and `j` indices are 0 based, and `i` indices are indices corresponding to the local `j` array.
4179: Once you have created the matrix you can update it with new numerical values using `MatUpdateMPIAIJWithArray()`
4181: If you do **not** use `MatUpdateMPIAIJWithArray()`, the column indices in `j` do not need to be sorted. If you will use
4182: `MatUpdateMPIAIJWithArrays()`, the column indices **must** be sorted.
4184: The format which is used for the sparse matrix input, is equivalent to a
4185: row-major ordering, i.e., for the following matrix, the input data expected is
4186: as shown
4187: .vb
4188: 1 0 0
4189: 2 0 3 P0
4190: -------
4191: 4 5 6 P1
4193: Process0 [P0] rows_owned=[0,1]
4194: i = {0,1,3} [size = nrow+1 = 2+1]
4195: j = {0,0,2} [size = 3]
4196: v = {1,2,3} [size = 3]
4198: Process1 [P1] rows_owned=[2]
4199: i = {0,3} [size = nrow+1 = 1+1]
4200: j = {0,1,2} [size = 3]
4201: v = {4,5,6} [size = 3]
4202: .ve
4204: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4205: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArray()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4206: @*/
4207: PetscErrorCode MatCreateMPIAIJWithArrays(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt M, PetscInt N, const PetscInt i[], const PetscInt j[], const PetscScalar a[], Mat *mat)
4208: {
4209: PetscFunctionBegin;
4210: PetscCheck(!i || !i[0], PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
4211: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
4212: PetscCall(MatCreate(comm, mat));
4213: PetscCall(MatSetSizes(*mat, m, n, M, N));
4214: /* PetscCall(MatSetBlockSizes(M,bs,cbs)); */
4215: PetscCall(MatSetType(*mat, MATMPIAIJ));
4216: PetscCall(MatMPIAIJSetPreallocationCSR(*mat, i, j, a));
4217: PetscFunctionReturn(PETSC_SUCCESS);
4218: }
4220: /*@
4221: MatUpdateMPIAIJWithArrays - updates a `MATMPIAIJ` matrix using arrays that contain in standard
4222: CSR format for the local rows. Only the numerical values are updated the other arrays must be identical to what was passed
4223: from `MatCreateMPIAIJWithArrays()`
4225: Deprecated: Use `MatUpdateMPIAIJWithArray()`
4227: Collective
4229: Input Parameters:
4230: + mat - the matrix
4231: . m - number of local rows (Cannot be `PETSC_DECIDE`)
4232: . n - This value should be the same as the local size used in creating the
4233: x vector for the matrix-vector product y = Ax. (or `PETSC_DECIDE` to have
4234: calculated if N is given) For square matrices n is almost always m.
4235: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if m is given)
4236: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if n is given)
4237: . Ii - row indices; that is Ii[0] = 0, Ii[row] = Ii[row-1] + number of elements in that row of the matrix
4238: . J - column indices
4239: - v - matrix values
4241: Level: deprecated
4243: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4244: `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArray()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4245: @*/
4246: PetscErrorCode MatUpdateMPIAIJWithArrays(Mat mat, PetscInt m, PetscInt n, PetscInt M, PetscInt N, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[])
4247: {
4248: PetscInt nnz, i;
4249: PetscBool nooffprocentries;
4250: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)mat->data;
4251: Mat_SeqAIJ *Ad = (Mat_SeqAIJ *)Aij->A->data;
4252: PetscScalar *ad, *ao;
4253: PetscInt ldi, Iii, md;
4254: const PetscInt *Adi = Ad->i;
4255: PetscInt *ld = Aij->ld;
4257: PetscFunctionBegin;
4258: PetscCheck(Ii[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
4259: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
4260: PetscCheck(m == mat->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Local number of rows cannot change from call to MatUpdateMPIAIJWithArrays()");
4261: PetscCheck(n == mat->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Local number of columns cannot change from call to MatUpdateMPIAIJWithArrays()");
4263: PetscCall(MatSeqAIJGetArrayWrite(Aij->A, &ad));
4264: PetscCall(MatSeqAIJGetArrayWrite(Aij->B, &ao));
4266: for (i = 0; i < m; i++) {
4267: if (PetscDefined(USE_DEBUG)) {
4268: for (PetscInt j = Ii[i] + 1; j < Ii[i + 1]; ++j) {
4269: PetscCheck(J[j] >= J[j - 1], PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column entry number %" PetscInt_FMT " (actual column %" PetscInt_FMT ") in row %" PetscInt_FMT " is not sorted", j - Ii[i], J[j], i);
4270: 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);
4271: }
4272: }
4273: nnz = Ii[i + 1] - Ii[i];
4274: Iii = Ii[i];
4275: ldi = ld[i];
4276: md = Adi[i + 1] - Adi[i];
4277: PetscCall(PetscArraycpy(ao, v + Iii, ldi));
4278: PetscCall(PetscArraycpy(ad, v + Iii + ldi, md));
4279: PetscCall(PetscArraycpy(ao + ldi, v + Iii + ldi + md, nnz - ldi - md));
4280: ad += md;
4281: ao += nnz - md;
4282: }
4283: nooffprocentries = mat->nooffprocentries;
4284: mat->nooffprocentries = PETSC_TRUE;
4285: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->A, &ad));
4286: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->B, &ao));
4287: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->A));
4288: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->B));
4289: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
4290: PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
4291: PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
4292: mat->nooffprocentries = nooffprocentries;
4293: PetscFunctionReturn(PETSC_SUCCESS);
4294: }
4296: /*@
4297: MatUpdateMPIAIJWithArray - updates an `MATMPIAIJ` matrix using an array that contains the nonzero values
4299: Collective
4301: Input Parameters:
4302: + mat - the matrix
4303: - v - matrix values, stored by row
4305: Level: intermediate
4307: Notes:
4308: The matrix must have been obtained with `MatCreateMPIAIJWithArrays()` or `MatMPIAIJSetPreallocationCSR()`
4310: The column indices in the call to `MatCreateMPIAIJWithArrays()` or `MatMPIAIJSetPreallocationCSR()` must have been sorted for this call to work correctly
4312: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4313: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArrays()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4314: @*/
4315: PetscErrorCode MatUpdateMPIAIJWithArray(Mat mat, const PetscScalar v[])
4316: {
4317: PetscInt nnz, i, m;
4318: PetscBool nooffprocentries;
4319: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)mat->data;
4320: Mat_SeqAIJ *Ad = (Mat_SeqAIJ *)Aij->A->data;
4321: Mat_SeqAIJ *Ao = (Mat_SeqAIJ *)Aij->B->data;
4322: PetscScalar *ad, *ao;
4323: const PetscInt *Adi = Ad->i, *Adj = Ao->i;
4324: PetscInt ldi, Iii, md;
4325: PetscInt *ld = Aij->ld;
4327: PetscFunctionBegin;
4328: m = mat->rmap->n;
4330: PetscCall(MatSeqAIJGetArrayWrite(Aij->A, &ad));
4331: PetscCall(MatSeqAIJGetArrayWrite(Aij->B, &ao));
4332: Iii = 0;
4333: for (i = 0; i < m; i++) {
4334: nnz = Adi[i + 1] - Adi[i] + Adj[i + 1] - Adj[i];
4335: ldi = ld[i];
4336: md = Adi[i + 1] - Adi[i];
4337: PetscCall(PetscArraycpy(ad, v + Iii + ldi, md));
4338: ad += md;
4339: if (ao) {
4340: PetscCall(PetscArraycpy(ao, v + Iii, ldi));
4341: PetscCall(PetscArraycpy(ao + ldi, v + Iii + ldi + md, nnz - ldi - md));
4342: ao += nnz - md;
4343: }
4344: Iii += nnz;
4345: }
4346: nooffprocentries = mat->nooffprocentries;
4347: mat->nooffprocentries = PETSC_TRUE;
4348: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->A, &ad));
4349: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->B, &ao));
4350: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->A));
4351: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->B));
4352: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
4353: PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
4354: PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
4355: mat->nooffprocentries = nooffprocentries;
4356: PetscFunctionReturn(PETSC_SUCCESS);
4357: }
4359: /*@
4360: MatCreateAIJ - Creates a sparse parallel matrix in `MATAIJ` format
4361: (the default parallel PETSc format). For good matrix assembly performance
4362: the user should preallocate the matrix storage by setting the parameters
4363: `d_nz` (or `d_nnz`) and `o_nz` (or `o_nnz`).
4365: Collective
4367: Input Parameters:
4368: + comm - MPI communicator
4369: . m - number of local rows (or `PETSC_DECIDE` to have calculated if M is given)
4370: This value should be the same as the local size used in creating the
4371: y vector for the matrix-vector product y = Ax.
4372: . n - This value should be the same as the local size used in creating the
4373: x vector for the matrix-vector product y = Ax. (or `PETSC_DECIDE` to have
4374: calculated if N is given) For square matrices n is almost always m.
4375: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if m is given)
4376: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if n is given)
4377: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
4378: (same value is used for all local rows)
4379: . d_nnz - array containing the number of nonzeros in the various rows of the
4380: DIAGONAL portion of the local submatrix (possibly different for each row)
4381: or `NULL`, if `d_nz` is used to specify the nonzero structure.
4382: The size of this array is equal to the number of local rows, i.e 'm'.
4383: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
4384: submatrix (same value is used for all local rows).
4385: - o_nnz - array containing the number of nonzeros in the various rows of the
4386: OFF-DIAGONAL portion of the local submatrix (possibly different for
4387: each row) or `NULL`, if `o_nz` is used to specify the nonzero
4388: structure. The size of this array is equal to the number
4389: of local rows, i.e 'm'.
4391: Output Parameter:
4392: . A - the matrix
4394: Options Database Keys:
4395: + -mat_no_inode - Do not use inodes
4396: . -mat_inode_limit limit - Sets inode limit (max limit=5)
4397: - -matmult_vecscatter_view viewer - View the vecscatter (i.e., communication pattern) used in `MatMult()` of sparse parallel matrices.
4398: See viewer types in manual of `MatView()`. Of them, ascii_matlab, draw or binary cause the `VecScatter`
4399: 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.
4401: Level: intermediate
4403: Notes:
4404: It is recommended that one use `MatCreateFromOptions()` or the `MatCreate()`, `MatSetType()` and/or `MatSetFromOptions()`,
4405: MatXXXXSetPreallocation() paradigm instead of this routine directly.
4406: [MatXXXXSetPreallocation() is, for example, `MatSeqAIJSetPreallocation()`]
4408: If the *_nnz parameter is given then the *_nz parameter is ignored
4410: The `m`,`n`,`M`,`N` parameters specify the size of the matrix, and its partitioning across
4411: processors, while `d_nz`,`d_nnz`,`o_nz`,`o_nnz` parameters specify the approximate
4412: storage requirements for this matrix.
4414: If `PETSC_DECIDE` or `PETSC_DETERMINE` is used for a particular argument on one
4415: processor than it must be used on all processors that share the object for
4416: that argument.
4418: If `m` and `n` are not `PETSC_DECIDE`, then the values determine the `PetscLayout` of the matrix and the ranges returned by
4419: `MatGetOwnershipRange()`, `MatGetOwnershipRanges()`, `MatGetOwnershipRangeColumn()`, and `MatGetOwnershipRangesColumn()`.
4421: The user MUST specify either the local or global matrix dimensions
4422: (possibly both).
4424: The parallel matrix is partitioned across processors such that the
4425: first `m0` rows belong to process 0, the next `m1` rows belong to
4426: process 1, the next `m2` rows belong to process 2, etc., where
4427: `m0`, `m1`, `m2`... are the input parameter `m` on each MPI process. I.e., each MPI process stores
4428: values corresponding to [m x N] submatrix.
4430: The columns are logically partitioned with the n0 columns belonging
4431: to 0th partition, the next n1 columns belonging to the next
4432: partition etc.. where n0,n1,n2... are the input parameter 'n'.
4434: The DIAGONAL portion of the local submatrix on any given processor
4435: is the submatrix corresponding to the rows and columns m,n
4436: corresponding to the given processor. i.e diagonal matrix on
4437: process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1]
4438: etc. The remaining portion of the local submatrix [m x (N-n)]
4439: constitute the OFF-DIAGONAL portion. The example below better
4440: illustrates this concept. The two matrices, the DIAGONAL portion and
4441: the OFF-DIAGONAL portion are each stored as `MATSEQAIJ` matrices.
4443: For a square global matrix we define each processor's diagonal portion
4444: to be its local rows and the corresponding columns (a square submatrix);
4445: each processor's off-diagonal portion encompasses the remainder of the
4446: local matrix (a rectangular submatrix).
4448: If `o_nnz`, `d_nnz` are specified, then `o_nz`, and `d_nz` are ignored.
4450: When calling this routine with a single process communicator, a matrix of
4451: type `MATSEQAIJ` is returned. If a matrix of type `MATMPIAIJ` is desired for this
4452: type of communicator, use the construction mechanism
4453: .vb
4454: MatCreate(..., &A);
4455: MatSetType(A, MATMPIAIJ);
4456: MatSetSizes(A, m, n, M, N);
4457: MatMPIAIJSetPreallocation(A, ...);
4458: .ve
4460: By default, this format uses inodes (identical nodes) when possible.
4461: We search for consecutive rows with the same nonzero structure, thereby
4462: reusing matrix information to achieve increased efficiency.
4464: Example Usage:
4465: Consider the following 8x8 matrix with 34 non-zero values, that is
4466: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4467: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4468: as follows
4470: .vb
4471: 1 2 0 | 0 3 0 | 0 4
4472: Proc0 0 5 6 | 7 0 0 | 8 0
4473: 9 0 10 | 11 0 0 | 12 0
4474: -------------------------------------
4475: 13 0 14 | 15 16 17 | 0 0
4476: Proc1 0 18 0 | 19 20 21 | 0 0
4477: 0 0 0 | 22 23 0 | 24 0
4478: -------------------------------------
4479: Proc2 25 26 27 | 0 0 28 | 29 0
4480: 30 0 0 | 31 32 33 | 0 34
4481: .ve
4483: This can be represented as a collection of submatrices as
4485: .vb
4486: A B C
4487: D E F
4488: G H I
4489: .ve
4491: Where the submatrices A,B,C are owned by proc0, D,E,F are
4492: owned by proc1, G,H,I are owned by proc2.
4494: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4495: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4496: The 'M','N' parameters are 8,8, and have the same values on all procs.
4498: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4499: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4500: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4501: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4502: part as `MATSEQAIJ` matrices. For example, proc1 will store [E] as a `MATSEQAIJ`
4503: matrix, and [DF] as another SeqAIJ matrix.
4505: When `d_nz`, `o_nz` parameters are specified, `d_nz` storage elements are
4506: allocated for every row of the local DIAGONAL submatrix, and `o_nz`
4507: storage locations are allocated for every row of the OFF-DIAGONAL submatrix.
4508: One way to choose `d_nz` and `o_nz` is to use the maximum number of nonzeros over
4509: the local rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4510: In this case, the values of `d_nz`,`o_nz` are
4511: .vb
4512: proc0 dnz = 2, o_nz = 2
4513: proc1 dnz = 3, o_nz = 2
4514: proc2 dnz = 1, o_nz = 4
4515: .ve
4516: We are allocating m*(`d_nz`+`o_nz`) storage locations for every proc. This
4517: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4518: for proc3. i.e we are using 12+15+10=37 storage locations to store
4519: 34 values.
4521: When `d_nnz`, `o_nnz` parameters are specified, the storage is specified
4522: for every row, corresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4523: In the above case the values for d_nnz,o_nnz are
4524: .vb
4525: proc0 d_nnz = [2,2,2] and o_nnz = [2,2,2]
4526: proc1 d_nnz = [3,3,2] and o_nnz = [2,1,1]
4527: proc2 d_nnz = [1,1] and o_nnz = [4,4]
4528: .ve
4529: Here the space allocated is sum of all the above values i.e 34, and
4530: hence pre-allocation is perfect.
4532: .seealso: [](ch_matrices), `Mat`, [Sparse Matrix Creation](sec_matsparse), `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4533: `MATMPIAIJ`, `MatCreateMPIAIJWithArrays()`, `MatGetOwnershipRange()`, `MatGetOwnershipRanges()`, `MatGetOwnershipRangeColumn()`,
4534: `MatGetOwnershipRangesColumn()`, `PetscLayout`
4535: @*/
4536: 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)
4537: {
4538: PetscMPIInt size;
4540: PetscFunctionBegin;
4541: PetscCall(MatCreate(comm, A));
4542: PetscCall(MatSetSizes(*A, m, n, M, N));
4543: PetscCallMPI(MPI_Comm_size(comm, &size));
4544: if (size > 1) {
4545: PetscCall(MatSetType(*A, MATMPIAIJ));
4546: PetscCall(MatMPIAIJSetPreallocation(*A, d_nz, d_nnz, o_nz, o_nnz));
4547: } else {
4548: PetscCall(MatSetType(*A, MATSEQAIJ));
4549: PetscCall(MatSeqAIJSetPreallocation(*A, d_nz, d_nnz));
4550: }
4551: PetscFunctionReturn(PETSC_SUCCESS);
4552: }
4554: /*@C
4555: MatMPIAIJGetSeqAIJ - Returns the local pieces of this distributed matrix
4557: Not Collective
4559: Input Parameter:
4560: . A - The `MATMPIAIJ` matrix
4562: Output Parameters:
4563: + Ad - The local diagonal block as a `MATSEQAIJ` matrix
4564: . Ao - The local off-diagonal block as a `MATSEQAIJ` matrix
4565: - colmap - An array mapping local column numbers of `Ao` to global column numbers of the parallel matrix
4567: Level: intermediate
4569: Note:
4570: The rows in `Ad` and `Ao` are in [0, Nr), where Nr is the number of local rows on this process. The columns
4571: 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
4572: the number of nonzero columns in the local off-diagonal piece of the matrix `A`. The array colmap maps these
4573: local column numbers to global column numbers in the original matrix.
4575: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatMPIAIJGetLocalMat()`, `MatMPIAIJGetLocalMatCondensed()`, `MatCreateAIJ()`, `MATSEQAIJ`
4576: @*/
4577: PetscErrorCode MatMPIAIJGetSeqAIJ(Mat A, Mat *Ad, Mat *Ao, const PetscInt *colmap[])
4578: {
4579: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
4580: PetscBool flg;
4582: PetscFunctionBegin;
4583: PetscCall(PetscStrbeginswith(((PetscObject)A)->type_name, MATMPIAIJ, &flg));
4584: PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "This function requires a MATMPIAIJ matrix as input");
4585: if (Ad) *Ad = a->A;
4586: if (Ao) *Ao = a->B;
4587: if (colmap) *colmap = a->garray;
4588: PetscFunctionReturn(PETSC_SUCCESS);
4589: }
4591: static PetscErrorCode MatGetMultPetscSF_MPIAIJ(Mat A, PetscSF *sf)
4592: {
4593: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
4595: PetscFunctionBegin;
4596: *sf = a->Mvctx;
4597: PetscFunctionReturn(PETSC_SUCCESS);
4598: }
4600: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm comm, Mat inmat, PetscInt n, MatReuse scall, Mat *outmat)
4601: {
4602: PetscInt m, N, i, rstart, nnz, Ii;
4603: PetscInt *indx;
4604: PetscScalar *values;
4605: MatType rootType;
4607: PetscFunctionBegin;
4608: PetscCall(MatGetSize(inmat, &m, &N));
4609: if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */
4610: PetscInt *dnz, *onz, sum, bs, cbs;
4612: if (n == PETSC_DECIDE) PetscCall(PetscSplitOwnership(comm, &n, &N));
4613: /* Check sum(n) = N */
4614: PetscCallMPI(MPIU_Allreduce(&n, &sum, 1, MPIU_INT, MPI_SUM, comm));
4615: PetscCheck(sum == N, PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Sum of local columns %" PetscInt_FMT " != global columns %" PetscInt_FMT, sum, N);
4617: PetscCallMPI(MPI_Scan(&m, &rstart, 1, MPIU_INT, MPI_SUM, comm));
4618: rstart -= m;
4620: MatPreallocateBegin(comm, m, n, dnz, onz);
4621: for (i = 0; i < m; i++) {
4622: PetscCall(MatGetRow_SeqAIJ(inmat, i, &nnz, &indx, NULL));
4623: PetscCall(MatPreallocateSet(i + rstart, nnz, indx, dnz, onz));
4624: PetscCall(MatRestoreRow_SeqAIJ(inmat, i, &nnz, &indx, NULL));
4625: }
4627: PetscCall(MatCreate(comm, outmat));
4628: PetscCall(MatSetSizes(*outmat, m, n, PETSC_DETERMINE, PETSC_DETERMINE));
4629: PetscCall(MatGetBlockSizes(inmat, &bs, &cbs));
4630: PetscCall(MatSetBlockSizes(*outmat, bs, cbs));
4631: PetscCall(MatGetRootType_Private(inmat, &rootType));
4632: PetscCall(MatSetType(*outmat, rootType));
4633: PetscCall(MatSeqAIJSetPreallocation(*outmat, 0, dnz));
4634: PetscCall(MatMPIAIJSetPreallocation(*outmat, 0, dnz, 0, onz));
4635: MatPreallocateEnd(dnz, onz);
4636: PetscCall(MatSetOption(*outmat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
4637: }
4639: /* numeric phase */
4640: PetscCall(MatGetOwnershipRange(*outmat, &rstart, NULL));
4641: for (i = 0; i < m; i++) {
4642: PetscCall(MatGetRow_SeqAIJ(inmat, i, &nnz, &indx, &values));
4643: Ii = i + rstart;
4644: PetscCall(MatSetValues(*outmat, 1, &Ii, nnz, indx, values, INSERT_VALUES));
4645: PetscCall(MatRestoreRow_SeqAIJ(inmat, i, &nnz, &indx, &values));
4646: }
4647: PetscCall(MatAssemblyBegin(*outmat, MAT_FINAL_ASSEMBLY));
4648: PetscCall(MatAssemblyEnd(*outmat, MAT_FINAL_ASSEMBLY));
4649: PetscFunctionReturn(PETSC_SUCCESS);
4650: }
4652: static PetscErrorCode MatMergeSeqsToMPIDestroy(PetscCtxRt data)
4653: {
4654: MatMergeSeqsToMPI *merge = *(MatMergeSeqsToMPI **)data;
4656: PetscFunctionBegin;
4657: if (!merge) PetscFunctionReturn(PETSC_SUCCESS);
4658: PetscCall(PetscFree(merge->id_r));
4659: PetscCall(PetscFree(merge->len_s));
4660: PetscCall(PetscFree(merge->len_r));
4661: PetscCall(PetscFree(merge->bi));
4662: PetscCall(PetscFree(merge->bj));
4663: PetscCall(PetscFree(merge->buf_ri[0]));
4664: PetscCall(PetscFree(merge->buf_ri));
4665: PetscCall(PetscFree(merge->buf_rj[0]));
4666: PetscCall(PetscFree(merge->buf_rj));
4667: PetscCall(PetscFree(merge->coi));
4668: PetscCall(PetscFree(merge->coj));
4669: PetscCall(PetscFree(merge->owners_co));
4670: PetscCall(PetscLayoutDestroy(&merge->rowmap));
4671: PetscCall(PetscFree(merge));
4672: PetscFunctionReturn(PETSC_SUCCESS);
4673: }
4675: #include <../src/mat/utils/freespace.h>
4676: #include <petscbt.h>
4678: /*@
4679: MatCreateMPIAIJSumSeqAIJNumeric - Fill the numerical values of an `MATMPIAIJ` matrix previously created by
4680: `MatCreateMPIAIJSumSeqAIJSymbolic()` by summing the local `MATSEQAIJ` contributions from each process.
4682: Collective
4684: Input Parameters:
4685: + seqmat - the local `MATSEQAIJ` contribution from this process
4686: - mpimat - the target `MATMPIAIJ` matrix created by `MatCreateMPIAIJSumSeqAIJSymbolic()`
4688: Level: developer
4690: .seealso: `Mat`, `MATMPIAIJ`, `MATSEQAIJ`, `MatCreateMPIAIJSumSeqAIJSymbolic()`, `MatCreateMPIAIJSumSeqAIJ()`
4691: @*/
4692: PetscErrorCode MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat, Mat mpimat)
4693: {
4694: MPI_Comm comm;
4695: Mat_SeqAIJ *a = (Mat_SeqAIJ *)seqmat->data;
4696: PetscMPIInt size, rank, taga, *len_s;
4697: PetscInt N = mpimat->cmap->N, i, j, *owners, *ai = a->i, *aj, m;
4698: PetscMPIInt proc, k;
4699: PetscInt **buf_ri, **buf_rj;
4700: PetscInt anzi, *bj_i, *bi, *bj, arow, bnzi, nextaj;
4701: PetscInt nrows, **buf_ri_k, **nextrow, **nextai;
4702: MPI_Request *s_waits, *r_waits;
4703: MPI_Status *status;
4704: const MatScalar *aa, *a_a;
4705: MatScalar **abuf_r, *ba_i;
4706: MatMergeSeqsToMPI *merge;
4707: PetscContainer container;
4709: PetscFunctionBegin;
4710: PetscCall(PetscObjectGetComm((PetscObject)mpimat, &comm));
4711: PetscCall(PetscLogEventBegin(MAT_Seqstompinum, seqmat, 0, 0, 0));
4713: PetscCallMPI(MPI_Comm_size(comm, &size));
4714: PetscCallMPI(MPI_Comm_rank(comm, &rank));
4716: PetscCall(PetscObjectQuery((PetscObject)mpimat, "MatMergeSeqsToMPI", (PetscObject *)&container));
4717: PetscCheck(container, PetscObjectComm((PetscObject)mpimat), PETSC_ERR_PLIB, "Mat not created from MatCreateMPIAIJSumSeqAIJSymbolic");
4718: PetscCall(PetscContainerGetPointer(container, &merge));
4719: PetscCall(MatSeqAIJGetArrayRead(seqmat, &a_a));
4720: aa = a_a;
4722: bi = merge->bi;
4723: bj = merge->bj;
4724: buf_ri = merge->buf_ri;
4725: buf_rj = merge->buf_rj;
4727: PetscCall(PetscMalloc1(size, &status));
4728: owners = merge->rowmap->range;
4729: len_s = merge->len_s;
4731: /* send and recv matrix values */
4732: PetscCall(PetscObjectGetNewTag((PetscObject)mpimat, &taga));
4733: PetscCall(PetscPostIrecvScalar(comm, taga, merge->nrecv, merge->id_r, merge->len_r, &abuf_r, &r_waits));
4735: PetscCall(PetscMalloc1(merge->nsend + 1, &s_waits));
4736: for (proc = 0, k = 0; proc < size; proc++) {
4737: if (!len_s[proc]) continue;
4738: i = owners[proc];
4739: PetscCallMPI(MPIU_Isend(aa + ai[i], len_s[proc], MPIU_MATSCALAR, proc, taga, comm, s_waits + k));
4740: k++;
4741: }
4743: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, r_waits, status));
4744: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, s_waits, status));
4745: PetscCall(PetscFree(status));
4747: PetscCall(PetscFree(s_waits));
4748: PetscCall(PetscFree(r_waits));
4750: /* insert mat values of mpimat */
4751: PetscCall(PetscMalloc1(N, &ba_i));
4752: PetscCall(PetscMalloc3(merge->nrecv, &buf_ri_k, merge->nrecv, &nextrow, merge->nrecv, &nextai));
4754: for (k = 0; k < merge->nrecv; k++) {
4755: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4756: nrows = *buf_ri_k[k];
4757: nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */
4758: nextai[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure */
4759: }
4761: /* set values of ba */
4762: m = merge->rowmap->n;
4763: for (i = 0; i < m; i++) {
4764: arow = owners[rank] + i;
4765: bj_i = bj + bi[i]; /* col indices of the i-th row of mpimat */
4766: bnzi = bi[i + 1] - bi[i];
4767: PetscCall(PetscArrayzero(ba_i, bnzi));
4769: /* add local non-zero vals of this proc's seqmat into ba */
4770: anzi = ai[arow + 1] - ai[arow];
4771: aj = a->j + ai[arow];
4772: aa = a_a + ai[arow];
4773: nextaj = 0;
4774: for (j = 0; nextaj < anzi; j++) {
4775: if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4776: ba_i[j] += aa[nextaj++];
4777: }
4778: }
4780: /* add received vals into ba */
4781: for (k = 0; k < merge->nrecv; k++) { /* k-th received message */
4782: /* i-th row */
4783: if (i == *nextrow[k]) {
4784: anzi = *(nextai[k] + 1) - *nextai[k];
4785: aj = buf_rj[k] + *nextai[k];
4786: aa = abuf_r[k] + *nextai[k];
4787: nextaj = 0;
4788: for (j = 0; nextaj < anzi; j++) {
4789: if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4790: ba_i[j] += aa[nextaj++];
4791: }
4792: }
4793: nextrow[k]++;
4794: nextai[k]++;
4795: }
4796: }
4797: PetscCall(MatSetValues(mpimat, 1, &arow, bnzi, bj_i, ba_i, INSERT_VALUES));
4798: }
4799: PetscCall(MatSeqAIJRestoreArrayRead(seqmat, &a_a));
4800: PetscCall(MatAssemblyBegin(mpimat, MAT_FINAL_ASSEMBLY));
4801: PetscCall(MatAssemblyEnd(mpimat, MAT_FINAL_ASSEMBLY));
4803: PetscCall(PetscFree(abuf_r[0]));
4804: PetscCall(PetscFree(abuf_r));
4805: PetscCall(PetscFree(ba_i));
4806: PetscCall(PetscFree3(buf_ri_k, nextrow, nextai));
4807: PetscCall(PetscLogEventEnd(MAT_Seqstompinum, seqmat, 0, 0, 0));
4808: PetscFunctionReturn(PETSC_SUCCESS);
4809: }
4811: /*@
4812: MatCreateMPIAIJSumSeqAIJSymbolic - Create the symbolic (nonzero-pattern) portion of an `MATMPIAIJ` matrix
4813: obtained by summing local `MATSEQAIJ` contributions from each process.
4815: Collective
4817: Input Parameters:
4818: + comm - the communicator
4819: . seqmat - the local `MATSEQAIJ` contribution from this process
4820: . m - the number of local rows for the resulting `MATMPIAIJ` matrix, or `PETSC_DECIDE`
4821: - n - the number of local columns for the resulting `MATMPIAIJ` matrix, or `PETSC_DECIDE`
4823: Output Parameter:
4824: . mpimat - the newly created `MATMPIAIJ` matrix
4826: Level: developer
4828: Note:
4829: The numerical values are filled in by a subsequent call to `MatCreateMPIAIJSumSeqAIJNumeric()`.
4831: .seealso: `Mat`, `MATMPIAIJ`, `MATSEQAIJ`, `MatCreateMPIAIJSumSeqAIJNumeric()`, `MatCreateMPIAIJSumSeqAIJ()`
4832: @*/
4833: PetscErrorCode MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm, Mat seqmat, PetscInt m, PetscInt n, Mat *mpimat)
4834: {
4835: Mat B_mpi;
4836: Mat_SeqAIJ *a = (Mat_SeqAIJ *)seqmat->data;
4837: PetscMPIInt size, rank, tagi, tagj, *len_s, *len_si, *len_ri;
4838: PetscInt **buf_rj, **buf_ri, **buf_ri_k;
4839: PetscInt M = seqmat->rmap->n, N = seqmat->cmap->n, i, *owners, *ai = a->i, *aj = a->j;
4840: PetscInt len, *dnz, *onz, bs, cbs;
4841: PetscInt k, anzi, *bi, *bj, *lnk, nlnk, arow, bnzi;
4842: PetscInt nrows, *buf_s, *buf_si, *buf_si_i, **nextrow, **nextai;
4843: MPI_Request *si_waits, *sj_waits, *ri_waits, *rj_waits;
4844: MPI_Status *status;
4845: PetscFreeSpaceList free_space = NULL, current_space = NULL;
4846: PetscBT lnkbt;
4847: MatMergeSeqsToMPI *merge;
4848: PetscContainer container;
4850: PetscFunctionBegin;
4851: PetscCall(PetscLogEventBegin(MAT_Seqstompisym, seqmat, 0, 0, 0));
4853: /* make sure it is a PETSc comm */
4854: PetscCall(PetscCommDuplicate(comm, &comm, NULL));
4855: PetscCallMPI(MPI_Comm_size(comm, &size));
4856: PetscCallMPI(MPI_Comm_rank(comm, &rank));
4858: PetscCall(PetscNew(&merge));
4859: PetscCall(PetscMalloc1(size, &status));
4861: /* determine row ownership */
4862: PetscCall(PetscLayoutCreate(comm, &merge->rowmap));
4863: PetscCall(PetscLayoutSetLocalSize(merge->rowmap, m));
4864: PetscCall(PetscLayoutSetSize(merge->rowmap, M));
4865: PetscCall(PetscLayoutSetBlockSize(merge->rowmap, 1));
4866: PetscCall(PetscLayoutSetUp(merge->rowmap));
4867: PetscCall(PetscMalloc1(size, &len_si));
4868: PetscCall(PetscMalloc1(size, &merge->len_s));
4870: m = merge->rowmap->n;
4871: owners = merge->rowmap->range;
4873: /* determine the number of messages to send, their lengths */
4874: len_s = merge->len_s;
4876: len = 0; /* length of buf_si[] */
4877: merge->nsend = 0;
4878: for (PetscMPIInt proc = 0; proc < size; proc++) {
4879: len_si[proc] = 0;
4880: if (proc == rank) {
4881: len_s[proc] = 0;
4882: } else {
4883: PetscCall(PetscMPIIntCast(owners[proc + 1] - owners[proc] + 1, &len_si[proc]));
4884: PetscCall(PetscMPIIntCast(ai[owners[proc + 1]] - ai[owners[proc]], &len_s[proc])); /* num of rows to be sent to [proc] */
4885: }
4886: if (len_s[proc]) {
4887: merge->nsend++;
4888: nrows = 0;
4889: for (i = owners[proc]; i < owners[proc + 1]; i++) {
4890: if (ai[i + 1] > ai[i]) nrows++;
4891: }
4892: PetscCall(PetscMPIIntCast(2 * (nrows + 1), &len_si[proc]));
4893: len += len_si[proc];
4894: }
4895: }
4897: /* determine the number and length of messages to receive for ij-structure */
4898: PetscCall(PetscGatherNumberOfMessages(comm, NULL, len_s, &merge->nrecv));
4899: PetscCall(PetscGatherMessageLengths2(comm, merge->nsend, merge->nrecv, len_s, len_si, &merge->id_r, &merge->len_r, &len_ri));
4901: /* post the Irecv of j-structure */
4902: PetscCall(PetscCommGetNewTag(comm, &tagj));
4903: PetscCall(PetscPostIrecvInt(comm, tagj, merge->nrecv, merge->id_r, merge->len_r, &buf_rj, &rj_waits));
4905: /* post the Isend of j-structure */
4906: PetscCall(PetscMalloc2(merge->nsend, &si_waits, merge->nsend, &sj_waits));
4908: for (PetscMPIInt proc = 0, k = 0; proc < size; proc++) {
4909: if (!len_s[proc]) continue;
4910: i = owners[proc];
4911: PetscCallMPI(MPIU_Isend(aj + ai[i], len_s[proc], MPIU_INT, proc, tagj, comm, sj_waits + k));
4912: k++;
4913: }
4915: /* receives and sends of j-structure are complete */
4916: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, rj_waits, status));
4917: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, sj_waits, status));
4919: /* send and recv i-structure */
4920: PetscCall(PetscCommGetNewTag(comm, &tagi));
4921: PetscCall(PetscPostIrecvInt(comm, tagi, merge->nrecv, merge->id_r, len_ri, &buf_ri, &ri_waits));
4923: PetscCall(PetscMalloc1(len + 1, &buf_s));
4924: buf_si = buf_s; /* points to the beginning of k-th msg to be sent */
4925: for (PetscMPIInt proc = 0, k = 0; proc < size; proc++) {
4926: if (!len_s[proc]) continue;
4927: /* form outgoing message for i-structure:
4928: buf_si[0]: nrows to be sent
4929: [1:nrows]: row index (global)
4930: [nrows+1:2*nrows+1]: i-structure index
4931: */
4932: nrows = len_si[proc] / 2 - 1;
4933: buf_si_i = buf_si + nrows + 1;
4934: buf_si[0] = nrows;
4935: buf_si_i[0] = 0;
4936: nrows = 0;
4937: for (i = owners[proc]; i < owners[proc + 1]; i++) {
4938: anzi = ai[i + 1] - ai[i];
4939: if (anzi) {
4940: buf_si_i[nrows + 1] = buf_si_i[nrows] + anzi; /* i-structure */
4941: buf_si[nrows + 1] = i - owners[proc]; /* local row index */
4942: nrows++;
4943: }
4944: }
4945: PetscCallMPI(MPIU_Isend(buf_si, len_si[proc], MPIU_INT, proc, tagi, comm, si_waits + k));
4946: k++;
4947: buf_si += len_si[proc];
4948: }
4950: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, ri_waits, status));
4951: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, si_waits, status));
4953: PetscCall(PetscInfo(seqmat, "nsend: %d, nrecv: %d\n", merge->nsend, merge->nrecv));
4954: 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]));
4956: PetscCall(PetscFree(len_si));
4957: PetscCall(PetscFree(len_ri));
4958: PetscCall(PetscFree(rj_waits));
4959: PetscCall(PetscFree2(si_waits, sj_waits));
4960: PetscCall(PetscFree(ri_waits));
4961: PetscCall(PetscFree(buf_s));
4962: PetscCall(PetscFree(status));
4964: /* compute a local seq matrix in each processor */
4965: /* allocate bi array and free space for accumulating nonzero column info */
4966: PetscCall(PetscMalloc1(m + 1, &bi));
4967: bi[0] = 0;
4969: /* create and initialize a linked list */
4970: nlnk = N + 1;
4971: PetscCall(PetscLLCreate(N, N, nlnk, lnk, lnkbt));
4973: /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */
4974: len = ai[owners[rank + 1]] - ai[owners[rank]];
4975: PetscCall(PetscFreeSpaceGet(PetscIntMultTruncate(2, len) + 1, &free_space));
4977: current_space = free_space;
4979: /* determine symbolic info for each local row */
4980: PetscCall(PetscMalloc3(merge->nrecv, &buf_ri_k, merge->nrecv, &nextrow, merge->nrecv, &nextai));
4982: for (k = 0; k < merge->nrecv; k++) {
4983: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4984: nrows = *buf_ri_k[k];
4985: nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */
4986: nextai[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure */
4987: }
4989: MatPreallocateBegin(comm, m, n, dnz, onz);
4990: len = 0;
4991: for (i = 0; i < m; i++) {
4992: bnzi = 0;
4993: /* add local non-zero cols of this proc's seqmat into lnk */
4994: arow = owners[rank] + i;
4995: anzi = ai[arow + 1] - ai[arow];
4996: aj = a->j + ai[arow];
4997: PetscCall(PetscLLAddSorted(anzi, aj, N, &nlnk, lnk, lnkbt));
4998: bnzi += nlnk;
4999: /* add received col data into lnk */
5000: for (k = 0; k < merge->nrecv; k++) { /* k-th received message */
5001: if (i == *nextrow[k]) { /* i-th row */
5002: anzi = *(nextai[k] + 1) - *nextai[k];
5003: aj = buf_rj[k] + *nextai[k];
5004: PetscCall(PetscLLAddSorted(anzi, aj, N, &nlnk, lnk, lnkbt));
5005: bnzi += nlnk;
5006: nextrow[k]++;
5007: nextai[k]++;
5008: }
5009: }
5010: if (len < bnzi) len = bnzi; /* =max(bnzi) */
5012: /* if free space is not available, make more free space */
5013: if (current_space->local_remaining < bnzi) PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(bnzi, current_space->total_array_size), ¤t_space));
5014: /* copy data into free space, then initialize lnk */
5015: PetscCall(PetscLLClean(N, N, bnzi, lnk, current_space->array, lnkbt));
5016: PetscCall(MatPreallocateSet(i + owners[rank], bnzi, current_space->array, dnz, onz));
5018: current_space->array += bnzi;
5019: current_space->local_used += bnzi;
5020: current_space->local_remaining -= bnzi;
5022: bi[i + 1] = bi[i] + bnzi;
5023: }
5025: PetscCall(PetscFree3(buf_ri_k, nextrow, nextai));
5027: PetscCall(PetscMalloc1(bi[m], &bj));
5028: PetscCall(PetscFreeSpaceContiguous(&free_space, bj));
5029: PetscCall(PetscLLDestroy(lnk, lnkbt));
5031: /* create symbolic parallel matrix B_mpi */
5032: PetscCall(MatGetBlockSizes(seqmat, &bs, &cbs));
5033: PetscCall(MatCreate(comm, &B_mpi));
5034: if (n == PETSC_DECIDE) PetscCall(MatSetSizes(B_mpi, m, n, PETSC_DETERMINE, N));
5035: else PetscCall(MatSetSizes(B_mpi, m, n, PETSC_DETERMINE, PETSC_DETERMINE));
5036: PetscCall(MatSetBlockSizes(B_mpi, bs, cbs));
5037: PetscCall(MatSetType(B_mpi, MATMPIAIJ));
5038: PetscCall(MatMPIAIJSetPreallocation(B_mpi, 0, dnz, 0, onz));
5039: MatPreallocateEnd(dnz, onz);
5040: PetscCall(MatSetOption(B_mpi, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_FALSE));
5042: /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */
5043: B_mpi->assembled = PETSC_FALSE;
5044: merge->bi = bi;
5045: merge->bj = bj;
5046: merge->buf_ri = buf_ri;
5047: merge->buf_rj = buf_rj;
5048: merge->coi = NULL;
5049: merge->coj = NULL;
5050: merge->owners_co = NULL;
5052: PetscCall(PetscCommDestroy(&comm));
5054: /* attach the supporting struct to B_mpi for reuse */
5055: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
5056: PetscCall(PetscContainerSetPointer(container, merge));
5057: PetscCall(PetscContainerSetCtxDestroy(container, MatMergeSeqsToMPIDestroy));
5058: PetscCall(PetscObjectCompose((PetscObject)B_mpi, "MatMergeSeqsToMPI", (PetscObject)container));
5059: PetscCall(PetscContainerDestroy(&container));
5060: *mpimat = B_mpi;
5062: PetscCall(PetscLogEventEnd(MAT_Seqstompisym, seqmat, 0, 0, 0));
5063: PetscFunctionReturn(PETSC_SUCCESS);
5064: }
5066: /*@
5067: MatCreateMPIAIJSumSeqAIJ - Creates a `MATMPIAIJ` matrix by adding sequential
5068: matrices from each processor
5070: Collective
5072: Input Parameters:
5073: + comm - the communicators the parallel matrix will live on
5074: . seqmat - the input sequential matrices
5075: . m - number of local rows (or `PETSC_DECIDE`)
5076: . n - number of local columns (or `PETSC_DECIDE`)
5077: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5079: Output Parameter:
5080: . mpimat - the parallel matrix generated
5082: Level: advanced
5084: Note:
5085: The dimensions of the sequential matrix in each processor MUST be the same.
5086: The input seqmat is included into the container `MatMergeSeqsToMPIDestroy`, and will be
5087: destroyed when `mpimat` is destroyed. Call `PetscObjectQuery()` to access `seqmat`.
5089: .seealso: [](ch_matrices), `Mat`, `MatCreateAIJ()`
5090: @*/
5091: PetscErrorCode MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm, Mat seqmat, PetscInt m, PetscInt n, MatReuse scall, Mat *mpimat)
5092: {
5093: PetscMPIInt size;
5095: PetscFunctionBegin;
5096: PetscCallMPI(MPI_Comm_size(comm, &size));
5097: if (size == 1) {
5098: PetscCall(PetscLogEventBegin(MAT_Seqstompi, seqmat, 0, 0, 0));
5099: if (scall == MAT_INITIAL_MATRIX) PetscCall(MatDuplicate(seqmat, MAT_COPY_VALUES, mpimat));
5100: else PetscCall(MatCopy(seqmat, *mpimat, SAME_NONZERO_PATTERN));
5101: PetscCall(PetscLogEventEnd(MAT_Seqstompi, seqmat, 0, 0, 0));
5102: PetscFunctionReturn(PETSC_SUCCESS);
5103: }
5104: PetscCall(PetscLogEventBegin(MAT_Seqstompi, seqmat, 0, 0, 0));
5105: if (scall == MAT_INITIAL_MATRIX) PetscCall(MatCreateMPIAIJSumSeqAIJSymbolic(comm, seqmat, m, n, mpimat));
5106: PetscCall(MatCreateMPIAIJSumSeqAIJNumeric(seqmat, *mpimat));
5107: PetscCall(PetscLogEventEnd(MAT_Seqstompi, seqmat, 0, 0, 0));
5108: PetscFunctionReturn(PETSC_SUCCESS);
5109: }
5111: /*@
5112: MatAIJGetLocalMat - Creates a `MATSEQAIJ` from a `MATAIJ` matrix.
5114: Not Collective
5116: Input Parameter:
5117: . A - the matrix
5119: Output Parameter:
5120: . A_loc - the local sequential matrix generated
5122: Level: developer
5124: Notes:
5125: The matrix is created by taking `A`'s local rows and putting them into a sequential matrix
5126: with `mlocal` rows and `n` columns. Where `mlocal` is obtained with `MatGetLocalSize()` and
5127: `n` is the global column count obtained with `MatGetSize()`
5129: In other words combines the two parts of a parallel `MATMPIAIJ` matrix on each process to a single matrix.
5131: For parallel matrices this creates an entirely new matrix. If the matrix is sequential it merely increases the reference count.
5133: Destroy the matrix with `MatDestroy()`
5135: .seealso: [](ch_matrices), `Mat`, `MatMPIAIJGetLocalMat()`
5136: @*/
5137: PetscErrorCode MatAIJGetLocalMat(Mat A, Mat *A_loc)
5138: {
5139: PetscBool mpi;
5141: PetscFunctionBegin;
5142: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIAIJ, &mpi));
5143: if (mpi) PetscCall(MatMPIAIJGetLocalMat(A, MAT_INITIAL_MATRIX, A_loc));
5144: else {
5145: *A_loc = A;
5146: PetscCall(PetscObjectReference((PetscObject)*A_loc));
5147: }
5148: PetscFunctionReturn(PETSC_SUCCESS);
5149: }
5151: /*@
5152: MatMPIAIJGetLocalMat - Creates a `MATSEQAIJ` from a `MATMPIAIJ` matrix.
5154: Not Collective
5156: Input Parameters:
5157: + A - the matrix
5158: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5160: Output Parameter:
5161: . A_loc - the local sequential matrix generated
5163: Level: developer
5165: Notes:
5166: The matrix is created by taking all `A`'s local rows and putting them into a sequential
5167: matrix with `mlocal` rows and `n` columns.`mlocal` is the row count obtained with
5168: `MatGetLocalSize()` and `n` is the global column count obtained with `MatGetSize()`.
5170: In other words combines the two parts of a parallel `MATMPIAIJ` matrix on each process to a single matrix.
5172: When `A` is sequential and `MAT_INITIAL_MATRIX` is requested, the matrix returned is the diagonal part of `A` (which contains the entire matrix),
5173: 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
5174: then `MatCopy`(Adiag,*`A_loc`,`SAME_NONZERO_PATTERN`) is called to fill `A_loc`. Thus one can preallocate the appropriate sequential matrix `A_loc`
5175: 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.
5177: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMatCondensed()`, `MatMPIAIJGetLocalMatMerge()`
5178: @*/
5179: PetscErrorCode MatMPIAIJGetLocalMat(Mat A, MatReuse scall, Mat *A_loc)
5180: {
5181: Mat_MPIAIJ *mpimat = (Mat_MPIAIJ *)A->data;
5182: Mat_SeqAIJ *mat, *a, *b;
5183: PetscInt *ai, *aj, *bi, *bj, *cmap = mpimat->garray;
5184: const PetscScalar *aa, *ba, *aav, *bav;
5185: PetscScalar *ca, *cam;
5186: PetscMPIInt size;
5187: PetscInt am = A->rmap->n, i, j, k, cstart = A->cmap->rstart;
5188: PetscInt *ci, *cj, col, ncols_d, ncols_o, jo;
5189: PetscBool match;
5191: PetscFunctionBegin;
5192: PetscCall(PetscStrbeginswith(((PetscObject)A)->type_name, MATMPIAIJ, &match));
5193: PetscCheck(match, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Requires MATMPIAIJ matrix as input");
5194: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
5195: if (size == 1) {
5196: if (scall == MAT_INITIAL_MATRIX) {
5197: PetscCall(PetscObjectReference((PetscObject)mpimat->A));
5198: *A_loc = mpimat->A;
5199: } else if (scall == MAT_REUSE_MATRIX) {
5200: PetscCall(MatCopy(mpimat->A, *A_loc, SAME_NONZERO_PATTERN));
5201: }
5202: PetscFunctionReturn(PETSC_SUCCESS);
5203: }
5205: PetscCall(PetscLogEventBegin(MAT_Getlocalmat, A, 0, 0, 0));
5206: a = (Mat_SeqAIJ *)mpimat->A->data;
5207: b = (Mat_SeqAIJ *)mpimat->B->data;
5208: ai = a->i;
5209: aj = a->j;
5210: bi = b->i;
5211: bj = b->j;
5212: PetscCall(MatSeqAIJGetArrayRead(mpimat->A, &aav));
5213: PetscCall(MatSeqAIJGetArrayRead(mpimat->B, &bav));
5214: aa = aav;
5215: ba = bav;
5216: if (scall == MAT_INITIAL_MATRIX) {
5217: PetscCall(PetscMalloc1(1 + am, &ci));
5218: ci[0] = 0;
5219: for (i = 0; i < am; i++) ci[i + 1] = ci[i] + (ai[i + 1] - ai[i]) + (bi[i + 1] - bi[i]);
5220: PetscCall(PetscMalloc1(1 + ci[am], &cj));
5221: PetscCall(PetscMalloc1(1 + ci[am], &ca));
5222: k = 0;
5223: for (i = 0; i < am; i++) {
5224: ncols_o = bi[i + 1] - bi[i];
5225: ncols_d = ai[i + 1] - ai[i];
5226: /* off-diagonal portion of A */
5227: for (jo = 0; jo < ncols_o; jo++) {
5228: col = cmap[*bj];
5229: if (col >= cstart) break;
5230: cj[k] = col;
5231: bj++;
5232: ca[k++] = *ba++;
5233: }
5234: /* diagonal portion of A */
5235: for (j = 0; j < ncols_d; j++) {
5236: cj[k] = cstart + *aj++;
5237: ca[k++] = *aa++;
5238: }
5239: /* off-diagonal portion of A */
5240: for (j = jo; j < ncols_o; j++) {
5241: cj[k] = cmap[*bj++];
5242: ca[k++] = *ba++;
5243: }
5244: }
5245: /* put together the new matrix */
5246: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, am, A->cmap->N, ci, cj, ca, A_loc));
5247: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5248: /* Since these are PETSc arrays, change flags to free them as necessary. */
5249: mat = (Mat_SeqAIJ *)(*A_loc)->data;
5250: mat->free_a = PETSC_TRUE;
5251: mat->free_ij = PETSC_TRUE;
5252: mat->nonew = 0;
5253: } else if (scall == MAT_REUSE_MATRIX) {
5254: mat = (Mat_SeqAIJ *)(*A_loc)->data;
5255: ci = mat->i;
5256: cj = mat->j;
5257: PetscCall(MatSeqAIJGetArrayWrite(*A_loc, &cam));
5258: for (i = 0; i < am; i++) {
5259: /* off-diagonal portion of A */
5260: ncols_o = bi[i + 1] - bi[i];
5261: for (jo = 0; jo < ncols_o; jo++) {
5262: col = cmap[*bj];
5263: if (col >= cstart) break;
5264: *cam++ = *ba++;
5265: bj++;
5266: }
5267: /* diagonal portion of A */
5268: ncols_d = ai[i + 1] - ai[i];
5269: for (j = 0; j < ncols_d; j++) *cam++ = *aa++;
5270: /* off-diagonal portion of A */
5271: for (j = jo; j < ncols_o; j++) {
5272: *cam++ = *ba++;
5273: bj++;
5274: }
5275: }
5276: PetscCall(MatSeqAIJRestoreArrayWrite(*A_loc, &cam));
5277: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Invalid MatReuse %d", (int)scall);
5278: PetscCall(MatSeqAIJRestoreArrayRead(mpimat->A, &aav));
5279: PetscCall(MatSeqAIJRestoreArrayRead(mpimat->B, &bav));
5280: PetscCall(PetscLogEventEnd(MAT_Getlocalmat, A, 0, 0, 0));
5281: PetscFunctionReturn(PETSC_SUCCESS);
5282: }
5284: /*@
5285: MatMPIAIJGetLocalMatMerge - Creates a `MATSEQAIJ` from a `MATMPIAIJ` matrix by taking all its local rows and putting them into a sequential matrix with
5286: mlocal rows and n columns. Where n is the sum of the number of columns of the diagonal and off-diagonal part
5288: Not Collective
5290: Input Parameters:
5291: + A - the matrix
5292: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5294: Output Parameters:
5295: + glob - sequential `IS` with global indices associated with the columns of the local sequential matrix generated (can be `NULL`)
5296: - A_loc - the local sequential matrix generated
5298: Level: developer
5300: Note:
5301: This is different from `MatMPIAIJGetLocalMat()` since the first columns in the returning matrix are those associated with the diagonal
5302: part, then those associated with the off-diagonal part (in its local ordering)
5304: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMat()`, `MatMPIAIJGetLocalMatCondensed()`
5305: @*/
5306: PetscErrorCode MatMPIAIJGetLocalMatMerge(Mat A, MatReuse scall, IS *glob, Mat *A_loc)
5307: {
5308: Mat Ao, Ad;
5309: const PetscInt *cmap;
5310: PetscMPIInt size;
5311: PetscErrorCode (*f)(Mat, MatReuse, IS *, Mat *);
5313: PetscFunctionBegin;
5314: PetscCall(MatMPIAIJGetSeqAIJ(A, &Ad, &Ao, &cmap));
5315: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
5316: if (size == 1) {
5317: if (scall == MAT_INITIAL_MATRIX) {
5318: PetscCall(PetscObjectReference((PetscObject)Ad));
5319: *A_loc = Ad;
5320: } else if (scall == MAT_REUSE_MATRIX) {
5321: PetscCall(MatCopy(Ad, *A_loc, SAME_NONZERO_PATTERN));
5322: }
5323: if (glob) PetscCall(ISCreateStride(PetscObjectComm((PetscObject)Ad), Ad->cmap->n, Ad->cmap->rstart, 1, glob));
5324: PetscFunctionReturn(PETSC_SUCCESS);
5325: }
5326: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatMPIAIJGetLocalMatMerge_C", &f));
5327: PetscCall(PetscLogEventBegin(MAT_Getlocalmat, A, 0, 0, 0));
5328: if (f) PetscCall((*f)(A, scall, glob, A_loc));
5329: else {
5330: Mat_SeqAIJ *a = (Mat_SeqAIJ *)Ad->data;
5331: Mat_SeqAIJ *b = (Mat_SeqAIJ *)Ao->data;
5332: Mat_SeqAIJ *c;
5333: PetscInt *ai = a->i, *aj = a->j;
5334: PetscInt *bi = b->i, *bj = b->j;
5335: PetscInt *ci, *cj;
5336: const PetscScalar *aa, *ba;
5337: PetscScalar *ca;
5338: PetscInt i, j, am, dn, on;
5340: PetscCall(MatGetLocalSize(Ad, &am, &dn));
5341: PetscCall(MatGetLocalSize(Ao, NULL, &on));
5342: PetscCall(MatSeqAIJGetArrayRead(Ad, &aa));
5343: PetscCall(MatSeqAIJGetArrayRead(Ao, &ba));
5344: if (scall == MAT_INITIAL_MATRIX) {
5345: PetscInt k;
5346: PetscCall(PetscMalloc1(1 + am, &ci));
5347: PetscCall(PetscMalloc1(ai[am] + bi[am], &cj));
5348: PetscCall(PetscMalloc1(ai[am] + bi[am], &ca));
5349: ci[0] = 0;
5350: for (i = 0, k = 0; i < am; i++) {
5351: const PetscInt ncols_o = bi[i + 1] - bi[i];
5352: const PetscInt ncols_d = ai[i + 1] - ai[i];
5353: ci[i + 1] = ci[i] + ncols_o + ncols_d;
5354: /* diagonal portion of A */
5355: for (j = 0; j < ncols_d; j++, k++) {
5356: cj[k] = *aj++;
5357: ca[k] = *aa++;
5358: }
5359: /* off-diagonal portion of A */
5360: for (j = 0; j < ncols_o; j++, k++) {
5361: cj[k] = dn + *bj++;
5362: ca[k] = *ba++;
5363: }
5364: }
5365: /* put together the new matrix */
5366: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, am, dn + on, ci, cj, ca, A_loc));
5367: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5368: /* Since these are PETSc arrays, change flags to free them as necessary. */
5369: c = (Mat_SeqAIJ *)(*A_loc)->data;
5370: c->free_a = PETSC_TRUE;
5371: c->free_ij = PETSC_TRUE;
5372: c->nonew = 0;
5373: PetscCall(MatSetType(*A_loc, ((PetscObject)Ad)->type_name));
5374: } else if (scall == MAT_REUSE_MATRIX) {
5375: PetscCall(MatSeqAIJGetArrayWrite(*A_loc, &ca));
5376: for (i = 0; i < am; i++) {
5377: const PetscInt ncols_d = ai[i + 1] - ai[i];
5378: const PetscInt ncols_o = bi[i + 1] - bi[i];
5379: /* diagonal portion of A */
5380: for (j = 0; j < ncols_d; j++) *ca++ = *aa++;
5381: /* off-diagonal portion of A */
5382: for (j = 0; j < ncols_o; j++) *ca++ = *ba++;
5383: }
5384: PetscCall(MatSeqAIJRestoreArrayWrite(*A_loc, &ca));
5385: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Invalid MatReuse %d", (int)scall);
5386: PetscCall(MatSeqAIJRestoreArrayRead(Ad, &aa));
5387: PetscCall(MatSeqAIJRestoreArrayRead(Ao, &aa));
5388: if (glob) {
5389: PetscInt cst, *gidx;
5391: PetscCall(MatGetOwnershipRangeColumn(A, &cst, NULL));
5392: PetscCall(PetscMalloc1(dn + on, &gidx));
5393: for (i = 0; i < dn; i++) gidx[i] = cst + i;
5394: for (i = 0; i < on; i++) gidx[i + dn] = cmap[i];
5395: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)Ad), dn + on, gidx, PETSC_OWN_POINTER, glob));
5396: }
5397: }
5398: PetscCall(PetscLogEventEnd(MAT_Getlocalmat, A, 0, 0, 0));
5399: PetscFunctionReturn(PETSC_SUCCESS);
5400: }
5402: /*@C
5403: MatMPIAIJGetLocalMatCondensed - Creates a `MATSEQAIJ` matrix from an `MATMPIAIJ` matrix by taking all its local rows and NON-ZERO columns
5405: Not Collective
5407: Input Parameters:
5408: + A - the matrix
5409: . scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5410: . row - index set of rows to extract (or `NULL`)
5411: - col - index set of columns to extract (or `NULL`)
5413: Output Parameter:
5414: . A_loc - the local sequential matrix generated
5416: Level: developer
5418: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMat()`
5419: @*/
5420: PetscErrorCode MatMPIAIJGetLocalMatCondensed(Mat A, MatReuse scall, IS *row, IS *col, Mat *A_loc)
5421: {
5422: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5423: PetscInt i, start, end, ncols, nzA, nzB, *cmap, imark, *idx;
5424: IS isrowa, iscola;
5425: Mat *aloc;
5426: PetscBool match;
5428: PetscFunctionBegin;
5429: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIAIJ, &match));
5430: PetscCheck(match, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Requires MATMPIAIJ matrix as input");
5431: PetscCall(PetscLogEventBegin(MAT_Getlocalmatcondensed, A, 0, 0, 0));
5432: if (!row) {
5433: start = A->rmap->rstart;
5434: end = A->rmap->rend;
5435: PetscCall(ISCreateStride(PETSC_COMM_SELF, end - start, start, 1, &isrowa));
5436: } else {
5437: isrowa = *row;
5438: }
5439: if (!col) {
5440: start = A->cmap->rstart;
5441: cmap = a->garray;
5442: nzA = a->A->cmap->n;
5443: nzB = a->B->cmap->n;
5444: PetscCall(PetscMalloc1(nzA + nzB, &idx));
5445: ncols = 0;
5446: for (i = 0; i < nzB; i++) {
5447: if (cmap[i] < start) idx[ncols++] = cmap[i];
5448: else break;
5449: }
5450: imark = i;
5451: for (i = 0; i < nzA; i++) idx[ncols++] = start + i;
5452: for (i = imark; i < nzB; i++) idx[ncols++] = cmap[i];
5453: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, &iscola));
5454: } else {
5455: iscola = *col;
5456: }
5457: if (scall != MAT_INITIAL_MATRIX) {
5458: PetscCall(PetscMalloc1(1, &aloc));
5459: aloc[0] = *A_loc;
5460: }
5461: PetscCall(MatCreateSubMatrices(A, 1, &isrowa, &iscola, scall, &aloc));
5462: if (!col) { /* attach global id of condensed columns */
5463: PetscCall(PetscObjectCompose((PetscObject)aloc[0], "_petsc_GetLocalMatCondensed_iscol", (PetscObject)iscola));
5464: }
5465: *A_loc = aloc[0];
5466: PetscCall(PetscFree(aloc));
5467: if (!row) PetscCall(ISDestroy(&isrowa));
5468: if (!col) PetscCall(ISDestroy(&iscola));
5469: PetscCall(PetscLogEventEnd(MAT_Getlocalmatcondensed, A, 0, 0, 0));
5470: PetscFunctionReturn(PETSC_SUCCESS);
5471: }
5473: /*
5474: * Create a sequential AIJ matrix based on row indices. a whole column is extracted once a row is matched.
5475: * Row could be local or remote.The routine is designed to be scalable in memory so that nothing is based
5476: * on a global size.
5477: * */
5478: static PetscErrorCode MatCreateSeqSubMatrixWithRows_Private(Mat P, IS rows, Mat *P_oth)
5479: {
5480: Mat_MPIAIJ *p = (Mat_MPIAIJ *)P->data;
5481: Mat_SeqAIJ *pd = (Mat_SeqAIJ *)p->A->data, *po = (Mat_SeqAIJ *)p->B->data, *p_oth;
5482: PetscInt plocalsize, nrows, *ilocal, *oilocal, i, lidx, *nrcols, *nlcols, ncol;
5483: PetscMPIInt owner;
5484: PetscSFNode *iremote, *oiremote;
5485: const PetscInt *lrowindices;
5486: PetscSF sf, osf;
5487: PetscInt pcstart, *roffsets, *loffsets, *pnnz, j;
5488: PetscInt ontotalcols, dntotalcols, ntotalcols, nout;
5489: MPI_Comm comm;
5490: ISLocalToGlobalMapping mapping;
5491: const PetscScalar *pd_a, *po_a;
5493: PetscFunctionBegin;
5494: PetscCall(PetscObjectGetComm((PetscObject)P, &comm));
5495: /* plocalsize is the number of roots
5496: * nrows is the number of leaves
5497: * */
5498: PetscCall(MatGetLocalSize(P, &plocalsize, NULL));
5499: PetscCall(ISGetLocalSize(rows, &nrows));
5500: PetscCall(PetscCalloc1(nrows, &iremote));
5501: PetscCall(ISGetIndices(rows, &lrowindices));
5502: for (i = 0; i < nrows; i++) {
5503: /* Find a remote index and an owner for a row
5504: * The row could be local or remote
5505: * */
5506: owner = 0;
5507: lidx = 0;
5508: PetscCall(PetscLayoutFindOwnerIndex(P->rmap, lrowindices[i], &owner, &lidx));
5509: iremote[i].index = lidx;
5510: iremote[i].rank = owner;
5511: }
5512: /* Create SF to communicate how many nonzero columns for each row */
5513: PetscCall(PetscSFCreate(comm, &sf));
5514: /* SF will figure out the number of nonzero columns for each row, and their
5515: * offsets
5516: * */
5517: PetscCall(PetscSFSetGraph(sf, plocalsize, nrows, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
5518: PetscCall(PetscSFSetFromOptions(sf));
5519: PetscCall(PetscSFSetUp(sf));
5521: PetscCall(PetscCalloc1(2 * (plocalsize + 1), &roffsets));
5522: PetscCall(PetscCalloc1(2 * plocalsize, &nrcols));
5523: PetscCall(PetscCalloc1(nrows, &pnnz));
5524: roffsets[0] = 0;
5525: roffsets[1] = 0;
5526: for (i = 0; i < plocalsize; i++) {
5527: /* diagonal */
5528: nrcols[i * 2 + 0] = pd->i[i + 1] - pd->i[i];
5529: /* off-diagonal */
5530: nrcols[i * 2 + 1] = po->i[i + 1] - po->i[i];
5531: /* compute offsets so that we relative location for each row */
5532: roffsets[(i + 1) * 2 + 0] = roffsets[i * 2 + 0] + nrcols[i * 2 + 0];
5533: roffsets[(i + 1) * 2 + 1] = roffsets[i * 2 + 1] + nrcols[i * 2 + 1];
5534: }
5535: PetscCall(PetscCalloc1(2 * nrows, &nlcols));
5536: PetscCall(PetscCalloc1(2 * nrows, &loffsets));
5537: /* 'r' means root, and 'l' means leaf */
5538: PetscCall(PetscSFBcastBegin(sf, MPIU_2INT, nrcols, nlcols, MPI_REPLACE));
5539: PetscCall(PetscSFBcastBegin(sf, MPIU_2INT, roffsets, loffsets, MPI_REPLACE));
5540: PetscCall(PetscSFBcastEnd(sf, MPIU_2INT, nrcols, nlcols, MPI_REPLACE));
5541: PetscCall(PetscSFBcastEnd(sf, MPIU_2INT, roffsets, loffsets, MPI_REPLACE));
5542: PetscCall(PetscSFDestroy(&sf));
5543: PetscCall(PetscFree(roffsets));
5544: PetscCall(PetscFree(nrcols));
5545: dntotalcols = 0;
5546: ontotalcols = 0;
5547: ncol = 0;
5548: for (i = 0; i < nrows; i++) {
5549: pnnz[i] = nlcols[i * 2 + 0] + nlcols[i * 2 + 1];
5550: ncol = PetscMax(pnnz[i], ncol);
5551: /* diagonal */
5552: dntotalcols += nlcols[i * 2 + 0];
5553: /* off-diagonal */
5554: ontotalcols += nlcols[i * 2 + 1];
5555: }
5556: /* We do not need to figure the right number of columns
5557: * since all the calculations will be done by going through the raw data
5558: * */
5559: PetscCall(MatCreateSeqAIJ(PETSC_COMM_SELF, nrows, ncol, 0, pnnz, P_oth));
5560: PetscCall(MatSetUp(*P_oth));
5561: PetscCall(PetscFree(pnnz));
5562: p_oth = (Mat_SeqAIJ *)(*P_oth)->data;
5563: /* diagonal */
5564: PetscCall(PetscCalloc1(dntotalcols, &iremote));
5565: /* off-diagonal */
5566: PetscCall(PetscCalloc1(ontotalcols, &oiremote));
5567: /* diagonal */
5568: PetscCall(PetscCalloc1(dntotalcols, &ilocal));
5569: /* off-diagonal */
5570: PetscCall(PetscCalloc1(ontotalcols, &oilocal));
5571: dntotalcols = 0;
5572: ontotalcols = 0;
5573: ntotalcols = 0;
5574: for (i = 0; i < nrows; i++) {
5575: owner = 0;
5576: PetscCall(PetscLayoutFindOwnerIndex(P->rmap, lrowindices[i], &owner, NULL));
5577: /* Set iremote for diag matrix */
5578: for (j = 0; j < nlcols[i * 2 + 0]; j++) {
5579: iremote[dntotalcols].index = loffsets[i * 2 + 0] + j;
5580: iremote[dntotalcols].rank = owner;
5581: /* P_oth is seqAIJ so that ilocal need to point to the first part of memory */
5582: ilocal[dntotalcols++] = ntotalcols++;
5583: }
5584: /* off-diagonal */
5585: for (j = 0; j < nlcols[i * 2 + 1]; j++) {
5586: oiremote[ontotalcols].index = loffsets[i * 2 + 1] + j;
5587: oiremote[ontotalcols].rank = owner;
5588: oilocal[ontotalcols++] = ntotalcols++;
5589: }
5590: }
5591: PetscCall(ISRestoreIndices(rows, &lrowindices));
5592: PetscCall(PetscFree(loffsets));
5593: PetscCall(PetscFree(nlcols));
5594: PetscCall(PetscSFCreate(comm, &sf));
5595: /* P serves as roots and P_oth is leaves
5596: * Diag matrix
5597: * */
5598: PetscCall(PetscSFSetGraph(sf, pd->i[plocalsize], dntotalcols, ilocal, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
5599: PetscCall(PetscSFSetFromOptions(sf));
5600: PetscCall(PetscSFSetUp(sf));
5602: PetscCall(PetscSFCreate(comm, &osf));
5603: /* off-diagonal */
5604: PetscCall(PetscSFSetGraph(osf, po->i[plocalsize], ontotalcols, oilocal, PETSC_OWN_POINTER, oiremote, PETSC_OWN_POINTER));
5605: PetscCall(PetscSFSetFromOptions(osf));
5606: PetscCall(PetscSFSetUp(osf));
5607: PetscCall(MatSeqAIJGetArrayRead(p->A, &pd_a));
5608: PetscCall(MatSeqAIJGetArrayRead(p->B, &po_a));
5609: /* operate on the matrix internal data to save memory */
5610: PetscCall(PetscSFBcastBegin(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5611: PetscCall(PetscSFBcastBegin(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5612: PetscCall(MatGetOwnershipRangeColumn(P, &pcstart, NULL));
5613: /* Convert to global indices for diag matrix */
5614: for (i = 0; i < pd->i[plocalsize]; i++) pd->j[i] += pcstart;
5615: PetscCall(PetscSFBcastBegin(sf, MPIU_INT, pd->j, p_oth->j, MPI_REPLACE));
5616: /* We want P_oth store global indices */
5617: PetscCall(ISLocalToGlobalMappingCreate(comm, 1, p->B->cmap->n, p->garray, PETSC_COPY_VALUES, &mapping));
5618: /* Use memory scalable approach */
5619: PetscCall(ISLocalToGlobalMappingSetType(mapping, ISLOCALTOGLOBALMAPPINGHASH));
5620: PetscCall(ISLocalToGlobalMappingApply(mapping, po->i[plocalsize], po->j, po->j));
5621: PetscCall(PetscSFBcastBegin(osf, MPIU_INT, po->j, p_oth->j, MPI_REPLACE));
5622: PetscCall(PetscSFBcastEnd(sf, MPIU_INT, pd->j, p_oth->j, MPI_REPLACE));
5623: /* Convert back to local indices */
5624: for (i = 0; i < pd->i[plocalsize]; i++) pd->j[i] -= pcstart;
5625: PetscCall(PetscSFBcastEnd(osf, MPIU_INT, po->j, p_oth->j, MPI_REPLACE));
5626: nout = 0;
5627: PetscCall(ISGlobalToLocalMappingApply(mapping, IS_GTOLM_DROP, po->i[plocalsize], po->j, &nout, po->j));
5628: PetscCheck(nout == po->i[plocalsize], comm, PETSC_ERR_ARG_INCOMP, "n %" PetscInt_FMT " does not equal to nout %" PetscInt_FMT " ", po->i[plocalsize], nout);
5629: PetscCall(ISLocalToGlobalMappingDestroy(&mapping));
5630: /* Exchange values */
5631: PetscCall(PetscSFBcastEnd(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5632: PetscCall(PetscSFBcastEnd(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5633: PetscCall(MatSeqAIJRestoreArrayRead(p->A, &pd_a));
5634: PetscCall(MatSeqAIJRestoreArrayRead(p->B, &po_a));
5635: /* Stop PETSc from shrinking memory */
5636: for (i = 0; i < nrows; i++) p_oth->ilen[i] = p_oth->imax[i];
5637: PetscCall(MatAssemblyBegin(*P_oth, MAT_FINAL_ASSEMBLY));
5638: PetscCall(MatAssemblyEnd(*P_oth, MAT_FINAL_ASSEMBLY));
5639: /* Attach PetscSF objects to P_oth so that we can reuse it later */
5640: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "diagsf", (PetscObject)sf));
5641: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "offdiagsf", (PetscObject)osf));
5642: PetscCall(PetscSFDestroy(&sf));
5643: PetscCall(PetscSFDestroy(&osf));
5644: PetscFunctionReturn(PETSC_SUCCESS);
5645: }
5647: /*
5648: * Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A
5649: * This supports MPIAIJ and MAIJ
5650: * */
5651: PetscErrorCode MatGetBrowsOfAcols_MPIXAIJ(Mat A, Mat P, PetscInt dof, MatReuse reuse, Mat *P_oth)
5652: {
5653: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data, *p = (Mat_MPIAIJ *)P->data;
5654: Mat_SeqAIJ *p_oth;
5655: IS rows, map;
5656: PetscHMapI hamp;
5657: PetscInt i, htsize, *rowindices, off, *mapping, key, count;
5658: MPI_Comm comm;
5659: PetscSF sf, osf;
5660: PetscBool has;
5662: PetscFunctionBegin;
5663: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
5664: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAocols, A, P, 0, 0));
5665: /* If it is the first time, create an index set of off-diag nonzero columns of A,
5666: * and then create a submatrix (that often is an overlapping matrix)
5667: * */
5668: if (reuse == MAT_INITIAL_MATRIX) {
5669: /* Use a hash table to figure out unique keys */
5670: PetscCall(PetscHMapICreateWithSize(a->B->cmap->n, &hamp));
5671: PetscCall(PetscCalloc1(a->B->cmap->n, &mapping));
5672: count = 0;
5673: /* Assume that a->g is sorted, otherwise the following does not make sense */
5674: for (i = 0; i < a->B->cmap->n; i++) {
5675: key = a->garray[i] / dof;
5676: PetscCall(PetscHMapIHas(hamp, key, &has));
5677: if (!has) {
5678: mapping[i] = count;
5679: PetscCall(PetscHMapISet(hamp, key, count++));
5680: } else {
5681: /* Current 'i' has the same value the previous step */
5682: mapping[i] = count - 1;
5683: }
5684: }
5685: PetscCall(ISCreateGeneral(comm, a->B->cmap->n, mapping, PETSC_OWN_POINTER, &map));
5686: PetscCall(PetscHMapIGetSize(hamp, &htsize));
5687: PetscCheck(htsize == count, comm, PETSC_ERR_ARG_INCOMP, " Size of hash map %" PetscInt_FMT " is inconsistent with count %" PetscInt_FMT, htsize, count);
5688: PetscCall(PetscCalloc1(htsize, &rowindices));
5689: off = 0;
5690: PetscCall(PetscHMapIGetKeys(hamp, &off, rowindices));
5691: PetscCall(PetscHMapIDestroy(&hamp));
5692: PetscCall(PetscSortInt(htsize, rowindices));
5693: PetscCall(ISCreateGeneral(comm, htsize, rowindices, PETSC_OWN_POINTER, &rows));
5694: /* In case, the matrix was already created but users want to recreate the matrix */
5695: PetscCall(MatDestroy(P_oth));
5696: PetscCall(MatCreateSeqSubMatrixWithRows_Private(P, rows, P_oth));
5697: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "aoffdiagtopothmapping", (PetscObject)map));
5698: PetscCall(ISDestroy(&map));
5699: PetscCall(ISDestroy(&rows));
5700: } else if (reuse == MAT_REUSE_MATRIX) {
5701: /* If matrix was already created, we simply update values using SF objects
5702: * that as attached to the matrix earlier.
5703: */
5704: const PetscScalar *pd_a, *po_a;
5706: PetscCall(PetscObjectQuery((PetscObject)*P_oth, "diagsf", (PetscObject *)&sf));
5707: PetscCall(PetscObjectQuery((PetscObject)*P_oth, "offdiagsf", (PetscObject *)&osf));
5708: PetscCheck(sf && osf, comm, PETSC_ERR_ARG_NULL, "Matrix is not initialized yet");
5709: p_oth = (Mat_SeqAIJ *)(*P_oth)->data;
5710: /* Update values in place */
5711: PetscCall(MatSeqAIJGetArrayRead(p->A, &pd_a));
5712: PetscCall(MatSeqAIJGetArrayRead(p->B, &po_a));
5713: PetscCall(PetscSFBcastBegin(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5714: PetscCall(PetscSFBcastBegin(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5715: PetscCall(PetscSFBcastEnd(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5716: PetscCall(PetscSFBcastEnd(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5717: PetscCall(MatSeqAIJRestoreArrayRead(p->A, &pd_a));
5718: PetscCall(MatSeqAIJRestoreArrayRead(p->B, &po_a));
5719: } else SETERRQ(comm, PETSC_ERR_ARG_UNKNOWN_TYPE, "Unknown reuse type");
5720: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAocols, A, P, 0, 0));
5721: PetscFunctionReturn(PETSC_SUCCESS);
5722: }
5724: /*@C
5725: MatGetBrowsOfAcols - Returns `IS` that contain rows of `B` that equal to nonzero columns of local `A`
5727: Collective
5729: Input Parameters:
5730: + A - the first matrix in `MATMPIAIJ` format
5731: . B - the second matrix in `MATMPIAIJ` format
5732: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5734: Output Parameters:
5735: + rowb - On input index sets of rows of B to extract (or `NULL`), modified on output
5736: . colb - On input index sets of columns of B to extract (or `NULL`), modified on output
5737: - B_seq - the sequential matrix generated
5739: Level: developer
5741: .seealso: `Mat`, `MATMPIAIJ`, `IS`, `MatReuse`
5742: @*/
5743: PetscErrorCode MatGetBrowsOfAcols(Mat A, Mat B, MatReuse scall, IS *rowb, IS *colb, Mat *B_seq)
5744: {
5745: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5746: PetscInt *idx, i, start, ncols, nzA, nzB, *cmap, imark;
5747: IS isrowb, iscolb;
5748: Mat *bseq = NULL;
5750: PetscFunctionBegin;
5751: 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 ")",
5752: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
5753: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAcols, A, B, 0, 0));
5755: if (scall == MAT_INITIAL_MATRIX) {
5756: start = A->cmap->rstart;
5757: cmap = a->garray;
5758: nzA = a->A->cmap->n;
5759: nzB = a->B->cmap->n;
5760: PetscCall(PetscMalloc1(nzA + nzB, &idx));
5761: ncols = 0;
5762: for (i = 0; i < nzB; i++) { /* row < local row index */
5763: if (cmap[i] < start) idx[ncols++] = cmap[i];
5764: else break;
5765: }
5766: imark = i;
5767: for (i = 0; i < nzA; i++) idx[ncols++] = start + i; /* local rows */
5768: for (i = imark; i < nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */
5769: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, &isrowb));
5770: PetscCall(ISCreateStride(PETSC_COMM_SELF, B->cmap->N, 0, 1, &iscolb));
5771: } else {
5772: PetscCheck(rowb && colb, PETSC_COMM_SELF, PETSC_ERR_SUP, "IS rowb and colb must be provided for MAT_REUSE_MATRIX");
5773: isrowb = *rowb;
5774: iscolb = *colb;
5775: PetscCall(PetscMalloc1(1, &bseq));
5776: bseq[0] = *B_seq;
5777: }
5778: PetscCall(MatCreateSubMatrices(B, 1, &isrowb, &iscolb, scall, &bseq));
5779: *B_seq = bseq[0];
5780: PetscCall(PetscFree(bseq));
5781: if (!rowb) {
5782: PetscCall(ISDestroy(&isrowb));
5783: } else {
5784: *rowb = isrowb;
5785: }
5786: if (!colb) {
5787: PetscCall(ISDestroy(&iscolb));
5788: } else {
5789: *colb = iscolb;
5790: }
5791: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAcols, A, B, 0, 0));
5792: PetscFunctionReturn(PETSC_SUCCESS);
5793: }
5795: // PetscClangLinter pragma disable: -fdoc-sowing-chars
5796: /*
5797: MatGetBrowsOfAoCols_MPIAIJ - Creates a `MATSEQAIJ` matrix by taking rows of B that equal to nonzero columns
5798: of the OFF-DIAGONAL portion of local A
5800: Collective
5802: Input Parameters:
5803: + A - the first matrix in `MATMPIAIJ` format
5804: . B - the second matrix in `MATMPIAIJ` format
5805: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5807: Output Parameters:
5808: + startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or NULL)
5809: . startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or NULL)
5810: . bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or NULL)
5811: - B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N
5813: Level: developer
5815: Developer Note:
5816: This directly accesses information inside the VecScatter associated with the matrix-vector product
5817: for this matrix. This is not desirable.
5819: .seealso: [](ch_mat), `Mat`, `MATMPIAIJ`
5820: */
5821: PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat A, Mat B, MatReuse scall, PetscInt **startsj_s, PetscInt **startsj_r, MatScalar **bufa_ptr, Mat *B_oth)
5822: {
5823: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5824: VecScatter ctx;
5825: MPI_Comm comm;
5826: const PetscMPIInt *rprocs, *sprocs;
5827: PetscMPIInt nrecvs, nsends;
5828: const PetscInt *srow, *rstarts, *sstarts;
5829: PetscInt *rowlen, *bufj, *bufJ, ncols = 0, aBn = a->B->cmap->n, row, *b_othi, *b_othj, *rvalues = NULL, *svalues = NULL, *cols, sbs, rbs;
5830: PetscInt i, j, k = 0, l, ll, nrows, *rstartsj = NULL, *sstartsj, len;
5831: PetscScalar *b_otha, *bufa, *bufA, *vals = NULL;
5832: MPI_Request *reqs = NULL, *rwaits = NULL, *swaits = NULL;
5833: PetscMPIInt size, tag, rank, nreqs;
5835: PetscFunctionBegin;
5836: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
5837: PetscCallMPI(MPI_Comm_size(comm, &size));
5839: 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 ")",
5840: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
5841: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAocols, A, B, 0, 0));
5842: PetscCallMPI(MPI_Comm_rank(comm, &rank));
5844: if (size == 1) {
5845: startsj_s = NULL;
5846: bufa_ptr = NULL;
5847: *B_oth = NULL;
5848: PetscFunctionReturn(PETSC_SUCCESS);
5849: }
5851: ctx = a->Mvctx;
5852: tag = ((PetscObject)ctx)->tag;
5854: PetscCall(VecScatterGetRemote_Private(ctx, PETSC_TRUE /*send*/, &nsends, &sstarts, &srow, &sprocs, &sbs));
5855: /* rprocs[] must be ordered so that indices received from them are ordered in rvalues[], which is key to algorithms used in this subroutine */
5856: PetscCall(VecScatterGetRemoteOrdered_Private(ctx, PETSC_FALSE /*recv*/, &nrecvs, &rstarts, NULL /*indices not needed*/, &rprocs, &rbs));
5857: PetscCall(PetscMPIIntCast(nsends + nrecvs, &nreqs));
5858: PetscCall(PetscMalloc1(nreqs, &reqs));
5859: rwaits = reqs;
5860: swaits = PetscSafePointerPlusOffset(reqs, nrecvs);
5862: if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX;
5863: if (scall == MAT_INITIAL_MATRIX) {
5864: /* i-array */
5865: /* post receives */
5866: if (nrecvs) PetscCall(PetscMalloc1(rbs * (rstarts[nrecvs] - rstarts[0]), &rvalues)); /* rstarts can be NULL when nrecvs=0 */
5867: for (i = 0; i < nrecvs; i++) {
5868: rowlen = rvalues + rstarts[i] * rbs;
5869: nrows = (rstarts[i + 1] - rstarts[i]) * rbs; /* num of indices to be received */
5870: PetscCallMPI(MPIU_Irecv(rowlen, nrows, MPIU_INT, rprocs[i], tag, comm, rwaits + i));
5871: }
5873: /* pack the outgoing message */
5874: PetscCall(PetscMalloc2(nsends + 1, &sstartsj, nrecvs + 1, &rstartsj));
5876: sstartsj[0] = 0;
5877: rstartsj[0] = 0;
5878: len = 0; /* total length of j or a array to be sent */
5879: if (nsends) {
5880: k = sstarts[0]; /* ATTENTION: sstarts[0] and rstarts[0] are not necessarily zero */
5881: PetscCall(PetscMalloc1(sbs * (sstarts[nsends] - sstarts[0]), &svalues));
5882: }
5883: for (i = 0; i < nsends; i++) {
5884: rowlen = svalues + (sstarts[i] - sstarts[0]) * sbs;
5885: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5886: for (j = 0; j < nrows; j++) {
5887: row = srow[k] + B->rmap->range[rank]; /* global row idx */
5888: for (l = 0; l < sbs; l++) {
5889: PetscCall(MatGetRow_MPIAIJ(B, row + l, &ncols, NULL, NULL)); /* rowlength */
5891: rowlen[j * sbs + l] = ncols;
5893: len += ncols;
5894: PetscCall(MatRestoreRow_MPIAIJ(B, row + l, &ncols, NULL, NULL));
5895: }
5896: k++;
5897: }
5898: PetscCallMPI(MPIU_Isend(rowlen, nrows * sbs, MPIU_INT, sprocs[i], tag, comm, swaits + i));
5900: sstartsj[i + 1] = len; /* starting point of (i+1)-th outgoing msg in bufj and bufa */
5901: }
5902: /* recvs and sends of i-array are completed */
5903: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5904: PetscCall(PetscFree(svalues));
5906: /* allocate buffers for sending j and a arrays */
5907: PetscCall(PetscMalloc1(len, &bufj));
5908: PetscCall(PetscMalloc1(len, &bufa));
5910: /* create i-array of B_oth */
5911: PetscCall(PetscMalloc1(aBn + 1, &b_othi));
5913: b_othi[0] = 0;
5914: len = 0; /* total length of j or a array to be received */
5915: k = 0;
5916: for (i = 0; i < nrecvs; i++) {
5917: rowlen = rvalues + (rstarts[i] - rstarts[0]) * rbs;
5918: nrows = (rstarts[i + 1] - rstarts[i]) * rbs; /* num of rows to be received */
5919: for (j = 0; j < nrows; j++) {
5920: b_othi[k + 1] = b_othi[k] + rowlen[j];
5921: PetscCall(PetscIntSumError(rowlen[j], len, &len));
5922: k++;
5923: }
5924: rstartsj[i + 1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */
5925: }
5926: PetscCall(PetscFree(rvalues));
5928: /* allocate space for j and a arrays of B_oth */
5929: PetscCall(PetscMalloc1(b_othi[aBn], &b_othj));
5930: PetscCall(PetscMalloc1(b_othi[aBn], &b_otha));
5932: /* j-array */
5933: /* post receives of j-array */
5934: for (i = 0; i < nrecvs; i++) {
5935: nrows = rstartsj[i + 1] - rstartsj[i]; /* length of the msg received */
5936: PetscCallMPI(MPIU_Irecv(PetscSafePointerPlusOffset(b_othj, rstartsj[i]), nrows, MPIU_INT, rprocs[i], tag, comm, rwaits + i));
5937: }
5939: /* pack the outgoing message j-array */
5940: if (nsends) k = sstarts[0];
5941: for (i = 0; i < nsends; i++) {
5942: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5943: bufJ = PetscSafePointerPlusOffset(bufj, sstartsj[i]);
5944: for (j = 0; j < nrows; j++) {
5945: row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5946: for (ll = 0; ll < sbs; ll++) {
5947: PetscCall(MatGetRow_MPIAIJ(B, row + ll, &ncols, &cols, NULL));
5948: for (l = 0; l < ncols; l++) *bufJ++ = cols[l];
5949: PetscCall(MatRestoreRow_MPIAIJ(B, row + ll, &ncols, &cols, NULL));
5950: }
5951: }
5952: PetscCallMPI(MPIU_Isend(PetscSafePointerPlusOffset(bufj, sstartsj[i]), sstartsj[i + 1] - sstartsj[i], MPIU_INT, sprocs[i], tag, comm, swaits + i));
5953: }
5955: /* recvs and sends of j-array are completed */
5956: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5957: } else if (scall == MAT_REUSE_MATRIX) {
5958: sstartsj = *startsj_s;
5959: rstartsj = *startsj_r;
5960: bufa = *bufa_ptr;
5961: PetscCall(MatSeqAIJGetArrayWrite(*B_oth, &b_otha));
5962: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not possess an object container");
5964: /* a-array */
5965: /* post receives of a-array */
5966: for (i = 0; i < nrecvs; i++) {
5967: nrows = rstartsj[i + 1] - rstartsj[i]; /* length of the msg received */
5968: PetscCallMPI(MPIU_Irecv(PetscSafePointerPlusOffset(b_otha, rstartsj[i]), nrows, MPIU_SCALAR, rprocs[i], tag, comm, rwaits + i));
5969: }
5971: /* pack the outgoing message a-array */
5972: if (nsends) k = sstarts[0];
5973: for (i = 0; i < nsends; i++) {
5974: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5975: bufA = PetscSafePointerPlusOffset(bufa, sstartsj[i]);
5976: for (j = 0; j < nrows; j++) {
5977: row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5978: for (ll = 0; ll < sbs; ll++) {
5979: PetscCall(MatGetRow_MPIAIJ(B, row + ll, &ncols, NULL, &vals));
5980: for (l = 0; l < ncols; l++) *bufA++ = vals[l];
5981: PetscCall(MatRestoreRow_MPIAIJ(B, row + ll, &ncols, NULL, &vals));
5982: }
5983: }
5984: PetscCallMPI(MPIU_Isend(PetscSafePointerPlusOffset(bufa, sstartsj[i]), sstartsj[i + 1] - sstartsj[i], MPIU_SCALAR, sprocs[i], tag, comm, swaits + i));
5985: }
5986: /* recvs and sends of a-array are completed */
5987: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5988: PetscCall(PetscFree(reqs));
5990: if (scall == MAT_INITIAL_MATRIX) {
5991: Mat_SeqAIJ *b_oth;
5993: /* put together the new matrix */
5994: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, aBn, B->cmap->N, b_othi, b_othj, b_otha, B_oth));
5996: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5997: /* Since these are PETSc arrays, change flags to free them as necessary. */
5998: b_oth = (Mat_SeqAIJ *)(*B_oth)->data;
5999: b_oth->free_a = PETSC_TRUE;
6000: b_oth->free_ij = PETSC_TRUE;
6001: b_oth->nonew = 0;
6003: PetscCall(PetscFree(bufj));
6004: if (!startsj_s || !bufa_ptr) {
6005: PetscCall(PetscFree2(sstartsj, rstartsj));
6006: PetscCall(PetscFree(bufa_ptr));
6007: } else {
6008: *startsj_s = sstartsj;
6009: *startsj_r = rstartsj;
6010: *bufa_ptr = bufa;
6011: }
6012: } else if (scall == MAT_REUSE_MATRIX) {
6013: PetscCall(MatSeqAIJRestoreArrayWrite(*B_oth, &b_otha));
6014: }
6016: PetscCall(VecScatterRestoreRemote_Private(ctx, PETSC_TRUE, &nsends, &sstarts, &srow, &sprocs, &sbs));
6017: PetscCall(VecScatterRestoreRemoteOrdered_Private(ctx, PETSC_FALSE, &nrecvs, &rstarts, NULL, &rprocs, &rbs));
6018: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAocols, A, B, 0, 0));
6019: PetscFunctionReturn(PETSC_SUCCESS);
6020: }
6022: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat, MatType, MatReuse, Mat *);
6023: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJPERM(Mat, MatType, MatReuse, Mat *);
6024: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJSELL(Mat, MatType, MatReuse, Mat *);
6025: #if PetscDefined(HAVE_MKL_SPARSE)
6026: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJMKL(Mat, MatType, MatReuse, Mat *);
6027: #endif
6028: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIBAIJ(Mat, MatType, MatReuse, Mat *);
6029: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISBAIJ(Mat, MatType, MatReuse, Mat *);
6030: #if PetscDefined(HAVE_ELEMENTAL)
6031: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat, MatType, MatReuse, Mat *);
6032: #endif
6033: #if PetscDefined(HAVE_SCALAPACK) && (PetscDefined(USE_REAL_SINGLE) || PetscDefined(USE_REAL_DOUBLE))
6034: PETSC_INTERN PetscErrorCode MatConvert_AIJ_ScaLAPACK(Mat, MatType, MatReuse, Mat *);
6035: #endif
6036: #if PetscDefined(HAVE_HYPRE)
6037: PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat, MatType, MatReuse, Mat *);
6038: #endif
6039: #if PetscDefined(HAVE_CUDA)
6040: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCUSPARSE(Mat, MatType, MatReuse, Mat *);
6041: #endif
6042: #if PetscDefined(HAVE_HIP)
6043: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJHIPSPARSE(Mat, MatType, MatReuse, Mat *);
6044: #endif
6045: #if PetscDefined(HAVE_KOKKOS_KERNELS)
6046: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJKokkos(Mat, MatType, MatReuse, Mat *);
6047: #endif
6048: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISELL(Mat, MatType, MatReuse, Mat *);
6049: PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat, MatType, MatReuse, Mat *);
6050: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_IS_XAIJ(Mat);
6052: /*
6053: Computes (B'*A')' since computing B*A directly is untenable
6055: n p p
6056: [ ] [ ] [ ]
6057: m [ A ] * n [ B ] = m [ C ]
6058: [ ] [ ] [ ]
6060: */
6061: static PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A, Mat B, Mat C)
6062: {
6063: Mat At, Bt, Ct;
6065: PetscFunctionBegin;
6066: PetscCall(MatTranspose(A, MAT_INITIAL_MATRIX, &At));
6067: PetscCall(MatTranspose(B, MAT_INITIAL_MATRIX, &Bt));
6068: PetscCall(MatMatMult(Bt, At, MAT_INITIAL_MATRIX, PETSC_CURRENT, &Ct));
6069: PetscCall(MatDestroy(&At));
6070: PetscCall(MatDestroy(&Bt));
6071: PetscCall(MatTransposeSetPrecursor(Ct, C));
6072: PetscCall(MatTranspose(Ct, MAT_REUSE_MATRIX, &C));
6073: PetscCall(MatDestroy(&Ct));
6074: PetscFunctionReturn(PETSC_SUCCESS);
6075: }
6077: static PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A, Mat B, PetscReal fill, Mat C)
6078: {
6079: PetscBool cisdense;
6081: PetscFunctionBegin;
6082: 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);
6083: PetscCall(MatSetSizes(C, A->rmap->n, B->cmap->n, A->rmap->N, B->cmap->N));
6084: PetscCall(MatSetBlockSizesFromMats(C, A, B));
6085: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATMPIDENSE, MATMPIDENSECUDA, MATMPIDENSEHIP, ""));
6086: if (!cisdense) PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
6087: PetscCall(MatSetUp(C));
6089: C->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIAIJ;
6090: PetscFunctionReturn(PETSC_SUCCESS);
6091: }
6093: static PetscErrorCode MatProductSetFromOptions_MPIDense_MPIAIJ_AB(Mat C)
6094: {
6095: Mat_Product *product = C->product;
6096: Mat A = product->A, B = product->B;
6098: PetscFunctionBegin;
6099: 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 ")",
6100: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
6101: C->ops->matmultsymbolic = MatMatMultSymbolic_MPIDense_MPIAIJ;
6102: C->ops->productsymbolic = MatProductSymbolic_AB;
6103: PetscFunctionReturn(PETSC_SUCCESS);
6104: }
6106: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_MPIDense_MPIAIJ(Mat C)
6107: {
6108: Mat_Product *product = C->product;
6110: PetscFunctionBegin;
6111: if (product->type == MATPRODUCT_AB) PetscCall(MatProductSetFromOptions_MPIDense_MPIAIJ_AB(C));
6112: PetscFunctionReturn(PETSC_SUCCESS);
6113: }
6115: /*
6116: Merge two sets of sorted nonzeros and return a CSR for the merged (sequential) matrix
6118: Input Parameters:
6120: j1,rowBegin1,rowEnd1,jmap1: describe the first set of nonzeros (Set1)
6121: j2,rowBegin2,rowEnd2,jmap2: describe the second set of nonzeros (Set2)
6123: mat: both sets' nonzeros are on m rows, where m is the number of local rows of the matrix mat
6125: For Set1, j1[] contains column indices of the nonzeros.
6126: 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
6127: respectively (note rowEnd1[k] is not necessarily equal to rwoBegin1[k+1]). Indices in this range of j1[] are sorted,
6128: but might have repeats. jmap1[t+1] - jmap1[t] is the number of repeats for the t-th unique nonzero in Set1.
6130: Similar for Set2.
6132: This routine merges the two sets of nonzeros row by row and removes repeats.
6134: Output Parameters: (memory is allocated by the caller)
6136: i[],j[]: the CSR of the merged matrix, which has m rows.
6137: imap1[]: the k-th unique nonzero in Set1 (k=0,1,...) corresponds to imap1[k]-th unique nonzero in the merged matrix.
6138: imap2[]: similar to imap1[], but for Set2.
6139: Note we order nonzeros row-by-row and from left to right.
6140: */
6141: 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[])
6142: {
6143: PetscInt r, m; /* Row index of mat */
6144: PetscCount t, t1, t2, b1, e1, b2, e2;
6146: PetscFunctionBegin;
6147: PetscCall(MatGetLocalSize(mat, &m, NULL));
6148: t1 = t2 = t = 0; /* Count unique nonzeros of in Set1, Set1 and the merged respectively */
6149: i[0] = 0;
6150: for (r = 0; r < m; r++) { /* Do row by row merging */
6151: b1 = rowBegin1[r];
6152: e1 = rowEnd1[r];
6153: b2 = rowBegin2[r];
6154: e2 = rowEnd2[r];
6155: while (b1 < e1 && b2 < e2) {
6156: if (j1[b1] == j2[b2]) { /* Same column index and hence same nonzero */
6157: j[t] = j1[b1];
6158: imap1[t1] = t;
6159: imap2[t2] = t;
6160: b1 += jmap1[t1 + 1] - jmap1[t1]; /* Jump to next unique local nonzero */
6161: b2 += jmap2[t2 + 1] - jmap2[t2]; /* Jump to next unique remote nonzero */
6162: t1++;
6163: t2++;
6164: t++;
6165: } else if (j1[b1] < j2[b2]) {
6166: j[t] = j1[b1];
6167: imap1[t1] = t;
6168: b1 += jmap1[t1 + 1] - jmap1[t1];
6169: t1++;
6170: t++;
6171: } else {
6172: j[t] = j2[b2];
6173: imap2[t2] = t;
6174: b2 += jmap2[t2 + 1] - jmap2[t2];
6175: t2++;
6176: t++;
6177: }
6178: }
6179: /* Merge the remaining in either j1[] or j2[] */
6180: while (b1 < e1) {
6181: j[t] = j1[b1];
6182: imap1[t1] = t;
6183: b1 += jmap1[t1 + 1] - jmap1[t1];
6184: t1++;
6185: t++;
6186: }
6187: while (b2 < e2) {
6188: j[t] = j2[b2];
6189: imap2[t2] = t;
6190: b2 += jmap2[t2 + 1] - jmap2[t2];
6191: t2++;
6192: t++;
6193: }
6194: PetscCall(PetscIntCast(t, i + r + 1));
6195: }
6196: PetscFunctionReturn(PETSC_SUCCESS);
6197: }
6199: /*
6200: Split nonzeros in a block of local rows into two subsets: those in the diagonal block and those in the off-diagonal block
6202: Input Parameters:
6203: mat: an MPI matrix that provides row and column layout information for splitting. Let's say its number of local rows is m.
6204: 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[]
6205: respectively, along with a permutation array perm[]. Length of the i[],j[],perm[] arrays is n.
6207: i[] is already sorted, but within a row, j[] is not sorted and might have repeats.
6208: i[] might contain negative indices at the beginning, which means the corresponding entries should be ignored in the splitting.
6210: Output Parameters:
6211: j[],perm[]: the routine needs to sort j[] within each row along with perm[].
6212: rowBegin[],rowMid[],rowEnd[]: of length m, and the memory is preallocated and zeroed by the caller.
6213: 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,
6214: and [rowMid[r],rowEnd[r]) point to begin/end entries of row r of the off-diagonal block.
6216: Aperm[],Ajmap[],Atot,Annz: Arrays are allocated by this routine.
6217: Atot: number of entries belonging to the diagonal block.
6218: Annz: number of unique nonzeros belonging to the diagonal block.
6219: Aperm[Atot] stores values from perm[] for entries belonging to the diagonal block. Length of Aperm[] is Atot, though it may also count
6220: repeats (i.e., same 'i,j' pair).
6221: Ajmap[Annz+1] stores the number of repeats of each unique entry belonging to the diagonal block. More precisely, Ajmap[t+1] - Ajmap[t]
6222: is the number of repeats for the t-th unique entry in the diagonal block. Ajmap[0] is always 0.
6224: Atot: number of entries belonging to the diagonal block
6225: Annz: number of unique nonzeros belonging to the diagonal block.
6227: Bperm[], Bjmap[], Btot, Bnnz are similar but for the off-diagonal block.
6229: Aperm[],Bperm[],Ajmap[] and Bjmap[] are allocated separately by this routine with PetscMalloc1().
6230: */
6231: 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_)
6232: {
6233: PetscInt cstart, cend, rstart, rend, row, col;
6234: PetscCount Atot = 0, Btot = 0; /* Total number of nonzeros in the diagonal and off-diagonal blocks */
6235: PetscCount Annz = 0, Bnnz = 0; /* Number of unique nonzeros in the diagonal and off-diagonal blocks */
6236: PetscCount k, m, p, q, r, s, mid;
6237: PetscCount *Aperm, *Bperm, *Ajmap, *Bjmap;
6239: PetscFunctionBegin;
6240: PetscCall(PetscLayoutGetRange(mat->rmap, &rstart, &rend));
6241: PetscCall(PetscLayoutGetRange(mat->cmap, &cstart, &cend));
6242: m = rend - rstart;
6244: /* Skip negative rows */
6245: for (k = 0; k < n; k++)
6246: if (i[k] >= 0) break;
6248: /* Process [k,n): sort and partition each local row into diag and offdiag portions,
6249: fill rowBegin[], rowMid[], rowEnd[], and count Atot, Btot, Annz, Bnnz.
6250: */
6251: while (k < n) {
6252: row = i[k];
6253: /* Entries in [k,s) are in one row. Shift diagonal block col indices so that diag is ahead of offdiag after sorting the row */
6254: for (s = k; s < n; s++)
6255: if (i[s] != row) break;
6257: /* Shift diag columns to range of [-PETSC_INT_MAX, -1] */
6258: for (p = k; p < s; p++) {
6259: if (j[p] >= cstart && j[p] < cend) j[p] -= PETSC_INT_MAX;
6260: }
6261: PetscCall(PetscSortIntWithCountArray(s - k, j + k, perm + k));
6262: PetscCall(PetscSortedIntUpperBound(j, k, s, -1, &mid)); /* Separate [k,s) into [k,mid) for diag and [mid,s) for offdiag */
6263: rowBegin[row - rstart] = k;
6264: rowMid[row - rstart] = mid;
6265: rowEnd[row - rstart] = s;
6266: 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);
6268: /* Count nonzeros of this diag/offdiag row, which might have repeats */
6269: Atot += mid - k;
6270: Btot += s - mid;
6272: /* Count unique nonzeros of this diag row */
6273: for (p = k; p < mid;) {
6274: col = j[p];
6275: do {
6276: j[p] += PETSC_INT_MAX; /* Revert the modified diagonal indices */
6277: p++;
6278: } while (p < mid && j[p] == col);
6279: Annz++;
6280: }
6282: /* Count unique nonzeros of this offdiag row */
6283: for (p = mid; p < s;) {
6284: col = j[p];
6285: do {
6286: p++;
6287: } while (p < s && j[p] == col);
6288: Bnnz++;
6289: }
6290: k = s;
6291: }
6293: /* Allocation according to Atot, Btot, Annz, Bnnz */
6294: PetscCall(PetscMalloc1(Atot, &Aperm));
6295: PetscCall(PetscMalloc1(Btot, &Bperm));
6296: PetscCall(PetscMalloc1(Annz + 1, &Ajmap));
6297: PetscCall(PetscMalloc1(Bnnz + 1, &Bjmap));
6299: /* Re-scan indices and copy diag/offdiag permutation indices to Aperm, Bperm and also fill Ajmap and Bjmap */
6300: Ajmap[0] = Bjmap[0] = Atot = Btot = Annz = Bnnz = 0;
6301: for (r = 0; r < m; r++) {
6302: k = rowBegin[r];
6303: mid = rowMid[r];
6304: s = rowEnd[r];
6305: PetscCall(PetscArraycpy(PetscSafePointerPlusOffset(Aperm, Atot), PetscSafePointerPlusOffset(perm, k), mid - k));
6306: PetscCall(PetscArraycpy(PetscSafePointerPlusOffset(Bperm, Btot), PetscSafePointerPlusOffset(perm, mid), s - mid));
6307: Atot += mid - k;
6308: Btot += s - mid;
6310: /* Scan column indices in this row and find out how many repeats each unique nonzero has */
6311: for (p = k; p < mid;) {
6312: col = j[p];
6313: q = p;
6314: do {
6315: p++;
6316: } while (p < mid && j[p] == col);
6317: Ajmap[Annz + 1] = Ajmap[Annz] + (p - q);
6318: Annz++;
6319: }
6321: for (p = mid; p < s;) {
6322: col = j[p];
6323: q = p;
6324: do {
6325: p++;
6326: } while (p < s && j[p] == col);
6327: Bjmap[Bnnz + 1] = Bjmap[Bnnz] + (p - q);
6328: Bnnz++;
6329: }
6330: }
6331: /* Output */
6332: *Aperm_ = Aperm;
6333: *Annz_ = Annz;
6334: *Atot_ = Atot;
6335: *Ajmap_ = Ajmap;
6336: *Bperm_ = Bperm;
6337: *Bnnz_ = Bnnz;
6338: *Btot_ = Btot;
6339: *Bjmap_ = Bjmap;
6340: PetscFunctionReturn(PETSC_SUCCESS);
6341: }
6343: /*
6344: Expand the jmap[] array to make a new one in view of nonzeros in the merged matrix
6346: Input Parameters:
6347: nnz1: number of unique nonzeros in a set that was used to produce imap[], jmap[]
6348: nnz: number of unique nonzeros in the merged matrix
6349: imap[nnz1]: i-th nonzero in the set is the imap[i]-th nonzero in the merged matrix
6350: jmap[nnz1+1]: i-th nonzero in the set has jmap[i+1] - jmap[i] repeats in the set
6352: Output Parameter: (memory is allocated by the caller)
6353: jmap_new[nnz+1]: i-th nonzero in the merged matrix has jmap_new[i+1] - jmap_new[i] repeats in the set
6355: Example:
6356: nnz1 = 4
6357: nnz = 6
6358: imap = [1,3,4,5]
6359: jmap = [0,3,5,6,7]
6360: then,
6361: jmap_new = [0,0,3,3,5,6,7]
6362: */
6363: static PetscErrorCode ExpandJmap_Internal(PetscCount nnz1, PetscCount nnz, const PetscCount imap[], const PetscCount jmap[], PetscCount jmap_new[])
6364: {
6365: PetscCount k, p;
6367: PetscFunctionBegin;
6368: jmap_new[0] = 0;
6369: p = nnz; /* p loops over jmap_new[] backwards */
6370: for (k = nnz1 - 1; k >= 0; k--) { /* k loops over imap[] */
6371: for (; p > imap[k]; p--) jmap_new[p] = jmap[k + 1];
6372: }
6373: for (; p >= 0; p--) jmap_new[p] = jmap[0];
6374: PetscFunctionReturn(PETSC_SUCCESS);
6375: }
6377: static PetscErrorCode MatCOOStructDestroy_MPIAIJ(PetscCtxRt data)
6378: {
6379: MatCOOStruct_MPIAIJ *coo = *(MatCOOStruct_MPIAIJ **)data;
6381: PetscFunctionBegin;
6382: PetscCall(PetscSFDestroy(&coo->sf));
6383: PetscCall(PetscFree(coo->Aperm1));
6384: PetscCall(PetscFree(coo->Bperm1));
6385: PetscCall(PetscFree(coo->Ajmap1));
6386: PetscCall(PetscFree(coo->Bjmap1));
6387: PetscCall(PetscFree(coo->Aimap2));
6388: PetscCall(PetscFree(coo->Bimap2));
6389: PetscCall(PetscFree(coo->Aperm2));
6390: PetscCall(PetscFree(coo->Bperm2));
6391: PetscCall(PetscFree(coo->Ajmap2));
6392: PetscCall(PetscFree(coo->Bjmap2));
6393: PetscCall(PetscFree(coo->Cperm1));
6394: PetscCall(PetscFree2(coo->sendbuf, coo->recvbuf));
6395: PetscCall(PetscFree(coo));
6396: PetscFunctionReturn(PETSC_SUCCESS);
6397: }
6399: PetscErrorCode MatSetPreallocationCOO_MPIAIJ(Mat mat, PetscCount coo_n, PetscInt coo_i[], PetscInt coo_j[])
6400: {
6401: MPI_Comm comm;
6402: PetscMPIInt rank, size;
6403: PetscInt m, n, M, N, rstart, rend, cstart, cend; /* Sizes, indices of row/col, therefore with type PetscInt */
6404: PetscCount k, p, q, rem; /* Loop variables over coo arrays */
6405: Mat_MPIAIJ *mpiaij = (Mat_MPIAIJ *)mat->data;
6406: PetscContainer container;
6407: MatCOOStruct_MPIAIJ *coo;
6409: PetscFunctionBegin;
6410: PetscCall(PetscFree(mpiaij->garray));
6411: PetscCall(VecDestroy(&mpiaij->lvec));
6412: #if PetscDefined(USE_CTABLE)
6413: PetscCall(PetscHMapIDestroy(&mpiaij->colmap));
6414: #else
6415: PetscCall(PetscFree(mpiaij->colmap));
6416: #endif
6417: PetscCall(VecScatterDestroy(&mpiaij->Mvctx));
6418: mat->assembled = PETSC_FALSE;
6419: mat->was_assembled = PETSC_FALSE;
6421: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
6422: PetscCallMPI(MPI_Comm_size(comm, &size));
6423: PetscCallMPI(MPI_Comm_rank(comm, &rank));
6424: PetscCall(PetscLayoutSetUp(mat->rmap));
6425: PetscCall(PetscLayoutSetUp(mat->cmap));
6426: PetscCall(PetscLayoutGetRange(mat->rmap, &rstart, &rend));
6427: PetscCall(PetscLayoutGetRange(mat->cmap, &cstart, &cend));
6428: PetscCall(MatGetLocalSize(mat, &m, &n));
6429: PetscCall(MatGetSize(mat, &M, &N));
6431: /* Sort (i,j) by row along with a permutation array, so that the to-be-ignored */
6432: /* entries come first, then local rows, then remote rows. */
6433: PetscCount n1 = coo_n, *perm1;
6434: PetscInt *i1 = coo_i, *j1 = coo_j;
6436: PetscCall(PetscMalloc1(n1, &perm1));
6437: for (k = 0; k < n1; k++) perm1[k] = k;
6439: /* Manipulate indices so that entries with negative row or col indices will have smallest
6440: row indices, local entries will have greater but negative row indices, and remote entries
6441: will have positive row indices.
6442: */
6443: for (k = 0; k < n1; k++) {
6444: if (i1[k] < 0 || j1[k] < 0) i1[k] = PETSC_INT_MIN; /* e.g., -2^31, minimal to move them ahead */
6445: 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] */
6446: else {
6447: PetscCheck(!mat->nooffprocentries, PETSC_COMM_SELF, PETSC_ERR_USER_INPUT, "MAT_NO_OFF_PROC_ENTRIES is set but insert to remote rows");
6448: if (mpiaij->donotstash) i1[k] = PETSC_INT_MIN; /* Ignore offproc entries as if they had negative indices */
6449: }
6450: }
6452: /* Sort by row; after that, [0,k) have ignored entries, [k,rem) have local rows and [rem,n1) have remote rows */
6453: PetscCall(PetscSortIntWithIntCountArrayPair(n1, i1, j1, perm1));
6455: /* Advance k to the first entry we need to take care of */
6456: for (k = 0; k < n1; k++)
6457: if (i1[k] > PETSC_INT_MIN) break;
6458: PetscCount i1start = k;
6460: PetscCall(PetscSortedIntUpperBound(i1, k, n1, rend - 1 - PETSC_INT_MAX, &rem)); /* rem is upper bound of the last local row */
6461: for (; k < rem; k++) i1[k] += PETSC_INT_MAX; /* Revert row indices of local rows*/
6463: 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);
6465: /* Send remote rows to their owner */
6466: /* Find which rows should be sent to which remote ranks*/
6467: PetscInt nsend = 0; /* Number of MPI ranks to send data to */
6468: PetscMPIInt *sendto; /* [nsend], storing remote ranks */
6469: PetscInt *nentries; /* [nsend], storing number of entries sent to remote ranks; Assume PetscInt is big enough for this count, and error if not */
6470: const PetscInt *ranges;
6471: PetscInt maxNsend = size >= 128 ? 128 : size; /* Assume max 128 neighbors; realloc when needed */
6473: PetscCall(PetscLayoutGetRanges(mat->rmap, &ranges));
6474: PetscCall(PetscMalloc2(maxNsend, &sendto, maxNsend, &nentries));
6475: for (k = rem; k < n1;) {
6476: PetscMPIInt owner;
6477: PetscInt firstRow, lastRow;
6479: /* Locate a row range */
6480: firstRow = i1[k]; /* first row of this owner */
6481: PetscCall(PetscLayoutFindOwner(mat->rmap, firstRow, &owner));
6482: lastRow = ranges[owner + 1] - 1; /* last row of this owner */
6484: /* Find the first index 'p' in [k,n) with i1[p] belonging to next owner */
6485: PetscCall(PetscSortedIntUpperBound(i1, k, n1, lastRow, &p));
6487: /* All entries in [k,p) belong to this remote owner */
6488: if (nsend >= maxNsend) { /* Double the remote ranks arrays if not long enough */
6489: PetscMPIInt *sendto2;
6490: PetscInt *nentries2;
6491: PetscInt maxNsend2 = (maxNsend <= size / 2) ? maxNsend * 2 : size;
6493: PetscCall(PetscMalloc2(maxNsend2, &sendto2, maxNsend2, &nentries2));
6494: PetscCall(PetscArraycpy(sendto2, sendto, maxNsend));
6495: PetscCall(PetscArraycpy(nentries2, nentries, maxNsend));
6496: PetscCall(PetscFree2(sendto, nentries));
6497: sendto = sendto2;
6498: nentries = nentries2;
6499: maxNsend = maxNsend2;
6500: }
6501: sendto[nsend] = owner;
6502: PetscCall(PetscIntCast(p - k, &nentries[nsend]));
6503: nsend++;
6504: k = p;
6505: }
6507: /* Build 1st SF to know offsets on remote to send data */
6508: PetscSF sf1;
6509: PetscInt nroots = 1, nroots2 = 0;
6510: PetscInt nleaves = nsend, nleaves2 = 0;
6511: PetscInt *offsets;
6512: PetscSFNode *iremote;
6514: PetscCall(PetscSFCreate(comm, &sf1));
6515: PetscCall(PetscMalloc1(nsend, &iremote));
6516: PetscCall(PetscMalloc1(nsend, &offsets));
6517: for (k = 0; k < nsend; k++) {
6518: iremote[k].rank = sendto[k];
6519: iremote[k].index = 0;
6520: nleaves2 += nentries[k];
6521: PetscCheck(nleaves2 >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Number of SF leaves is too large for PetscInt");
6522: }
6523: PetscCall(PetscSFSetGraph(sf1, nroots, nleaves, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
6524: PetscCall(PetscSFFetchAndOpWithMemTypeBegin(sf1, MPIU_INT, PETSC_MEMTYPE_HOST, &nroots2 /*rootdata*/, PETSC_MEMTYPE_HOST, nentries /*leafdata*/, PETSC_MEMTYPE_HOST, offsets /*leafupdate*/, MPI_SUM));
6525: PetscCall(PetscSFFetchAndOpEnd(sf1, MPIU_INT, &nroots2, nentries, offsets, MPI_SUM)); /* Would nroots2 overflow, we check offsets[] below */
6526: PetscCall(PetscSFDestroy(&sf1));
6527: PetscAssert(nleaves2 == n1 - rem, PETSC_COMM_SELF, PETSC_ERR_PLIB, "nleaves2 %" PetscInt_FMT " != number of remote entries %" PetscCount_FMT, nleaves2, n1 - rem);
6529: /* Build 2nd SF to send remote COOs to their owner */
6530: PetscSF sf2;
6531: nroots = nroots2;
6532: nleaves = nleaves2;
6533: PetscCall(PetscSFCreate(comm, &sf2));
6534: PetscCall(PetscSFSetFromOptions(sf2));
6535: PetscCall(PetscMalloc1(nleaves, &iremote));
6536: p = 0;
6537: for (k = 0; k < nsend; k++) {
6538: PetscCheck(offsets[k] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Number of SF roots is too large for PetscInt");
6539: for (q = 0; q < nentries[k]; q++, p++) {
6540: iremote[p].rank = sendto[k];
6541: PetscCall(PetscIntCast(offsets[k] + q, &iremote[p].index));
6542: }
6543: }
6544: PetscCall(PetscSFSetGraph(sf2, nroots, nleaves, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
6546: /* Send the remote COOs to their owner */
6547: PetscInt n2 = nroots, *i2, *j2; /* Buffers for received COOs from other ranks, along with a permutation array */
6548: PetscCount *perm2; /* Though PetscInt is enough for remote entries, we use PetscCount here as we want to reuse MatSplitEntries_Internal() */
6549: PetscCall(PetscMalloc3(n2, &i2, n2, &j2, n2, &perm2));
6550: PetscAssert(rem == 0 || i1 != NULL, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot add nonzero offset to null");
6551: PetscAssert(rem == 0 || j1 != NULL, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot add nonzero offset to null");
6552: PetscInt *i1prem = PetscSafePointerPlusOffset(i1, rem);
6553: PetscInt *j1prem = PetscSafePointerPlusOffset(j1, rem);
6554: PetscCall(PetscSFReduceWithMemTypeBegin(sf2, MPIU_INT, PETSC_MEMTYPE_HOST, i1prem, PETSC_MEMTYPE_HOST, i2, MPI_REPLACE));
6555: PetscCall(PetscSFReduceEnd(sf2, MPIU_INT, i1prem, i2, MPI_REPLACE));
6556: PetscCall(PetscSFReduceWithMemTypeBegin(sf2, MPIU_INT, PETSC_MEMTYPE_HOST, j1prem, PETSC_MEMTYPE_HOST, j2, MPI_REPLACE));
6557: PetscCall(PetscSFReduceEnd(sf2, MPIU_INT, j1prem, j2, MPI_REPLACE));
6559: PetscCall(PetscFree(offsets));
6560: PetscCall(PetscFree2(sendto, nentries));
6562: /* Sort received COOs by row along with the permutation array */
6563: for (k = 0; k < n2; k++) perm2[k] = k;
6564: PetscCall(PetscSortIntWithIntCountArrayPair(n2, i2, j2, perm2));
6566: /* sf2 only sends contiguous leafdata to contiguous rootdata. We record the permutation which will be used to fill leafdata */
6567: PetscCount *Cperm1;
6568: PetscAssert(rem == 0 || perm1 != NULL, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot add nonzero offset to null");
6569: PetscCount *perm1prem = PetscSafePointerPlusOffset(perm1, rem);
6570: PetscCall(PetscMalloc1(nleaves, &Cperm1));
6571: PetscCall(PetscArraycpy(Cperm1, perm1prem, nleaves));
6573: /* Support for HYPRE matrices, kind of a hack.
6574: Swap min column with diagonal so that diagonal values will go first */
6575: PetscBool hypre;
6576: PetscCall(PetscStrcmp("_internal_COO_mat_for_hypre", ((PetscObject)mat)->name, &hypre));
6577: if (hypre) {
6578: PetscInt *minj;
6579: PetscBT hasdiag;
6581: PetscCall(PetscBTCreate(m, &hasdiag));
6582: PetscCall(PetscMalloc1(m, &minj));
6583: for (k = 0; k < m; k++) minj[k] = PETSC_INT_MAX;
6584: for (k = i1start; k < rem; k++) {
6585: if (j1[k] < cstart || j1[k] >= cend) continue;
6586: const PetscInt rindex = i1[k] - rstart;
6587: if ((j1[k] - cstart) == rindex) PetscCall(PetscBTSet(hasdiag, rindex));
6588: minj[rindex] = PetscMin(minj[rindex], j1[k]);
6589: }
6590: for (k = 0; k < n2; k++) {
6591: if (j2[k] < cstart || j2[k] >= cend) continue;
6592: const PetscInt rindex = i2[k] - rstart;
6593: if ((j2[k] - cstart) == rindex) PetscCall(PetscBTSet(hasdiag, rindex));
6594: minj[rindex] = PetscMin(minj[rindex], j2[k]);
6595: }
6596: for (k = i1start; k < rem; k++) {
6597: const PetscInt rindex = i1[k] - rstart;
6598: if (j1[k] < cstart || j1[k] >= cend || !PetscBTLookup(hasdiag, rindex)) continue;
6599: if (j1[k] == minj[rindex]) j1[k] = i1[k] + (cstart - rstart);
6600: else if ((j1[k] - cstart) == rindex) j1[k] = minj[rindex];
6601: }
6602: for (k = 0; k < n2; k++) {
6603: const PetscInt rindex = i2[k] - rstart;
6604: if (j2[k] < cstart || j2[k] >= cend || !PetscBTLookup(hasdiag, rindex)) continue;
6605: if (j2[k] == minj[rindex]) j2[k] = i2[k] + (cstart - rstart);
6606: else if ((j2[k] - cstart) == rindex) j2[k] = minj[rindex];
6607: }
6608: PetscCall(PetscBTDestroy(&hasdiag));
6609: PetscCall(PetscFree(minj));
6610: }
6612: /* Split local COOs and received COOs into diag/offdiag portions */
6613: PetscCount *rowBegin1, *rowMid1, *rowEnd1;
6614: PetscCount *Ajmap1, *Aperm1, *Bjmap1, *Bperm1;
6615: PetscCount Annz1, Bnnz1, Atot1, Btot1;
6616: PetscCount *rowBegin2, *rowMid2, *rowEnd2;
6617: PetscCount *Ajmap2, *Aperm2, *Bjmap2, *Bperm2;
6618: PetscCount Annz2, Bnnz2, Atot2, Btot2;
6620: PetscCall(PetscCalloc3(m, &rowBegin1, m, &rowMid1, m, &rowEnd1));
6621: PetscCall(PetscCalloc3(m, &rowBegin2, m, &rowMid2, m, &rowEnd2));
6622: PetscCall(MatSplitEntries_Internal(mat, rem, i1, j1, perm1, rowBegin1, rowMid1, rowEnd1, &Atot1, &Aperm1, &Annz1, &Ajmap1, &Btot1, &Bperm1, &Bnnz1, &Bjmap1));
6623: PetscCall(MatSplitEntries_Internal(mat, n2, i2, j2, perm2, rowBegin2, rowMid2, rowEnd2, &Atot2, &Aperm2, &Annz2, &Ajmap2, &Btot2, &Bperm2, &Bnnz2, &Bjmap2));
6625: /* Merge local COOs with received COOs: diag with diag, offdiag with offdiag */
6626: PetscInt *Ai, *Bi;
6627: PetscInt *Aj, *Bj;
6629: PetscCall(PetscMalloc1(m + 1, &Ai));
6630: PetscCall(PetscMalloc1(m + 1, &Bi));
6631: PetscCall(PetscMalloc1(Annz1 + Annz2, &Aj)); /* Since local and remote entries might have dups, we might allocate excess memory */
6632: PetscCall(PetscMalloc1(Bnnz1 + Bnnz2, &Bj));
6634: PetscCount *Aimap1, *Bimap1, *Aimap2, *Bimap2;
6635: PetscCall(PetscMalloc1(Annz1, &Aimap1));
6636: PetscCall(PetscMalloc1(Bnnz1, &Bimap1));
6637: PetscCall(PetscMalloc1(Annz2, &Aimap2));
6638: PetscCall(PetscMalloc1(Bnnz2, &Bimap2));
6640: PetscCall(MatMergeEntries_Internal(mat, j1, j2, rowBegin1, rowMid1, rowBegin2, rowMid2, Ajmap1, Ajmap2, Aimap1, Aimap2, Ai, Aj));
6641: PetscCall(MatMergeEntries_Internal(mat, j1, j2, rowMid1, rowEnd1, rowMid2, rowEnd2, Bjmap1, Bjmap2, Bimap1, Bimap2, Bi, Bj));
6643: /* Expand Ajmap1/Bjmap1 to make them based off nonzeros in A/B, since we */
6644: /* expect nonzeros in A/B most likely have local contributing entries */
6645: PetscInt Annz = Ai[m];
6646: PetscInt Bnnz = Bi[m];
6647: PetscCount *Ajmap1_new, *Bjmap1_new;
6649: PetscCall(PetscMalloc1(Annz + 1, &Ajmap1_new));
6650: PetscCall(PetscMalloc1(Bnnz + 1, &Bjmap1_new));
6652: PetscCall(ExpandJmap_Internal(Annz1, Annz, Aimap1, Ajmap1, Ajmap1_new));
6653: PetscCall(ExpandJmap_Internal(Bnnz1, Bnnz, Bimap1, Bjmap1, Bjmap1_new));
6655: PetscCall(PetscFree(Aimap1));
6656: PetscCall(PetscFree(Ajmap1));
6657: PetscCall(PetscFree(Bimap1));
6658: PetscCall(PetscFree(Bjmap1));
6659: PetscCall(PetscFree3(rowBegin1, rowMid1, rowEnd1));
6660: PetscCall(PetscFree3(rowBegin2, rowMid2, rowEnd2));
6661: PetscCall(PetscFree(perm1));
6662: PetscCall(PetscFree3(i2, j2, perm2));
6664: Ajmap1 = Ajmap1_new;
6665: Bjmap1 = Bjmap1_new;
6667: /* Reallocate Aj, Bj once we know actual numbers of unique nonzeros in A and B */
6668: if (Annz < Annz1 + Annz2) {
6669: PetscInt *Aj_new;
6670: PetscCall(PetscMalloc1(Annz, &Aj_new));
6671: PetscCall(PetscArraycpy(Aj_new, Aj, Annz));
6672: PetscCall(PetscFree(Aj));
6673: Aj = Aj_new;
6674: }
6676: if (Bnnz < Bnnz1 + Bnnz2) {
6677: PetscInt *Bj_new;
6678: PetscCall(PetscMalloc1(Bnnz, &Bj_new));
6679: PetscCall(PetscArraycpy(Bj_new, Bj, Bnnz));
6680: PetscCall(PetscFree(Bj));
6681: Bj = Bj_new;
6682: }
6684: /* Create new submatrices for on-process and off-process coupling */
6685: PetscScalar *Aa, *Ba;
6686: MatType rtype;
6687: Mat_SeqAIJ *a, *b;
6688: PetscCall(PetscCalloc1(Annz, &Aa)); /* Zero matrix on device */
6689: PetscCall(PetscCalloc1(Bnnz, &Ba));
6690: /* make Aj[] local, i.e, based off the start column of the diagonal portion */
6691: if (cstart) {
6692: for (k = 0; k < Annz; k++) Aj[k] -= cstart;
6693: }
6695: PetscCall(MatGetRootType_Private(mat, &rtype));
6697: MatSeqXAIJGetOptions_Private(mpiaij->A);
6698: PetscCall(MatDestroy(&mpiaij->A));
6699: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, n, Ai, Aj, Aa, &mpiaij->A));
6700: PetscCall(MatSetBlockSizesFromMats(mpiaij->A, mat, mat));
6701: MatSeqXAIJRestoreOptions_Private(mpiaij->A);
6703: MatSeqXAIJGetOptions_Private(mpiaij->B);
6704: PetscCall(MatDestroy(&mpiaij->B));
6705: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, mat->cmap->N, Bi, Bj, Ba, &mpiaij->B));
6706: PetscCall(MatSetBlockSizesFromMats(mpiaij->B, mat, mat));
6707: MatSeqXAIJRestoreOptions_Private(mpiaij->B);
6709: PetscCall(MatSetUpMultiply_MPIAIJ(mat));
6710: mat->was_assembled = PETSC_TRUE; // was_assembled in effect means the Mvctx is built; doing so avoids redundant MatSetUpMultiply_MPIAIJ
6711: mat->nonzerostate = mpiaij->A->nonzerostate + mpiaij->B->nonzerostate;
6712: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &mat->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)mat)));
6714: a = (Mat_SeqAIJ *)mpiaij->A->data;
6715: b = (Mat_SeqAIJ *)mpiaij->B->data;
6716: a->free_a = PETSC_TRUE;
6717: a->free_ij = PETSC_TRUE;
6718: b->free_a = PETSC_TRUE;
6719: b->free_ij = PETSC_TRUE;
6720: a->maxnz = a->nz;
6721: b->maxnz = b->nz;
6723: /* conversion must happen AFTER multiply setup */
6724: PetscCall(MatConvert(mpiaij->A, rtype, MAT_INPLACE_MATRIX, &mpiaij->A));
6725: PetscCall(MatConvert(mpiaij->B, rtype, MAT_INPLACE_MATRIX, &mpiaij->B));
6726: PetscCall(VecDestroy(&mpiaij->lvec));
6727: PetscCall(MatCreateVecs(mpiaij->B, &mpiaij->lvec, NULL));
6729: // Put the COO struct in a container and then attach that to the matrix
6730: PetscCall(PetscMalloc1(1, &coo));
6731: coo->n = coo_n;
6732: coo->sf = sf2;
6733: coo->sendlen = nleaves;
6734: coo->recvlen = nroots;
6735: coo->Annz = Annz;
6736: coo->Bnnz = Bnnz;
6737: coo->Annz2 = Annz2;
6738: coo->Bnnz2 = Bnnz2;
6739: coo->Atot1 = Atot1;
6740: coo->Atot2 = Atot2;
6741: coo->Btot1 = Btot1;
6742: coo->Btot2 = Btot2;
6743: coo->Ajmap1 = Ajmap1;
6744: coo->Aperm1 = Aperm1;
6745: coo->Bjmap1 = Bjmap1;
6746: coo->Bperm1 = Bperm1;
6747: coo->Aimap2 = Aimap2;
6748: coo->Ajmap2 = Ajmap2;
6749: coo->Aperm2 = Aperm2;
6750: coo->Bimap2 = Bimap2;
6751: coo->Bjmap2 = Bjmap2;
6752: coo->Bperm2 = Bperm2;
6753: coo->Cperm1 = Cperm1;
6754: // Allocate in preallocation. If not used, it has zero cost on host
6755: PetscCall(PetscMalloc2(coo->sendlen, &coo->sendbuf, coo->recvlen, &coo->recvbuf));
6756: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
6757: PetscCall(PetscContainerSetPointer(container, coo));
6758: PetscCall(PetscContainerSetCtxDestroy(container, MatCOOStructDestroy_MPIAIJ));
6759: PetscCall(PetscObjectCompose((PetscObject)mat, "__PETSc_MatCOOStruct_Host", (PetscObject)container));
6760: PetscCall(PetscContainerDestroy(&container));
6761: PetscFunctionReturn(PETSC_SUCCESS);
6762: }
6764: static PetscErrorCode MatSetValuesCOO_MPIAIJ(Mat mat, const PetscScalar v[], InsertMode imode)
6765: {
6766: Mat_MPIAIJ *mpiaij = (Mat_MPIAIJ *)mat->data;
6767: Mat A = mpiaij->A, B = mpiaij->B;
6768: PetscScalar *Aa, *Ba;
6769: PetscScalar *sendbuf, *recvbuf;
6770: const PetscCount *Ajmap1, *Ajmap2, *Aimap2;
6771: const PetscCount *Bjmap1, *Bjmap2, *Bimap2;
6772: const PetscCount *Aperm1, *Aperm2, *Bperm1, *Bperm2;
6773: const PetscCount *Cperm1;
6774: PetscContainer container;
6775: MatCOOStruct_MPIAIJ *coo;
6777: PetscFunctionBegin;
6778: PetscCall(PetscObjectQuery((PetscObject)mat, "__PETSc_MatCOOStruct_Host", (PetscObject *)&container));
6779: PetscCheck(container, PetscObjectComm((PetscObject)mat), PETSC_ERR_PLIB, "Not found MatCOOStruct on this matrix");
6780: PetscCall(PetscContainerGetPointer(container, &coo));
6781: sendbuf = coo->sendbuf;
6782: recvbuf = coo->recvbuf;
6783: Ajmap1 = coo->Ajmap1;
6784: Ajmap2 = coo->Ajmap2;
6785: Aimap2 = coo->Aimap2;
6786: Bjmap1 = coo->Bjmap1;
6787: Bjmap2 = coo->Bjmap2;
6788: Bimap2 = coo->Bimap2;
6789: Aperm1 = coo->Aperm1;
6790: Aperm2 = coo->Aperm2;
6791: Bperm1 = coo->Bperm1;
6792: Bperm2 = coo->Bperm2;
6793: Cperm1 = coo->Cperm1;
6795: PetscCall(MatSeqAIJGetArray(A, &Aa)); /* Might read and write matrix values */
6796: PetscCall(MatSeqAIJGetArray(B, &Ba));
6798: /* Pack entries to be sent to remote */
6799: for (PetscCount i = 0; i < coo->sendlen; i++) sendbuf[i] = v[Cperm1[i]];
6801: /* Send remote entries to their owner and overlap the communication with local computation */
6802: PetscCall(PetscSFReduceWithMemTypeBegin(coo->sf, MPIU_SCALAR, PETSC_MEMTYPE_HOST, sendbuf, PETSC_MEMTYPE_HOST, recvbuf, MPI_REPLACE));
6803: /* Add local entries to A and B */
6804: for (PetscCount i = 0; i < coo->Annz; i++) { /* All nonzeros in A are either zero'ed or added with a value (i.e., initialized) */
6805: PetscScalar sum = 0.0; /* Do partial summation first to improve numerical stability */
6806: for (PetscCount k = Ajmap1[i]; k < Ajmap1[i + 1]; k++) sum += v[Aperm1[k]];
6807: Aa[i] = (imode == INSERT_VALUES ? 0.0 : Aa[i]) + sum;
6808: }
6809: for (PetscCount i = 0; i < coo->Bnnz; i++) {
6810: PetscScalar sum = 0.0;
6811: for (PetscCount k = Bjmap1[i]; k < Bjmap1[i + 1]; k++) sum += v[Bperm1[k]];
6812: Ba[i] = (imode == INSERT_VALUES ? 0.0 : Ba[i]) + sum;
6813: }
6814: PetscCall(PetscSFReduceEnd(coo->sf, MPIU_SCALAR, sendbuf, recvbuf, MPI_REPLACE));
6816: /* Add received remote entries to A and B */
6817: for (PetscCount i = 0; i < coo->Annz2; i++) {
6818: for (PetscCount k = Ajmap2[i]; k < Ajmap2[i + 1]; k++) Aa[Aimap2[i]] += recvbuf[Aperm2[k]];
6819: }
6820: for (PetscCount i = 0; i < coo->Bnnz2; i++) {
6821: for (PetscCount k = Bjmap2[i]; k < Bjmap2[i + 1]; k++) Ba[Bimap2[i]] += recvbuf[Bperm2[k]];
6822: }
6823: PetscCall(MatSeqAIJRestoreArray(A, &Aa));
6824: PetscCall(MatSeqAIJRestoreArray(B, &Ba));
6825: PetscFunctionReturn(PETSC_SUCCESS);
6826: }
6828: /*MC
6829: MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices.
6831: Options Database Keys:
6832: . -mat_type mpiaij - sets the matrix type to `MATMPIAIJ` during a call to `MatSetFromOptions()`
6834: Level: beginner
6836: Notes:
6837: `MatSetValues()` may be called for this matrix type with a `NULL` argument for the numerical values,
6838: in this case the values associated with the rows and columns one passes in are set to zero
6839: in the matrix
6841: `MatSetOptions`(,`MAT_STRUCTURE_ONLY`,`PETSC_TRUE`) may be called for this matrix type. In this no
6842: space is allocated for the nonzero entries and any entries passed with `MatSetValues()` are ignored
6844: .seealso: [](ch_matrices), `Mat`, `MATSEQAIJ`, `MATAIJ`, `MatCreateAIJ()`
6845: M*/
6846: PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat B)
6847: {
6848: Mat_MPIAIJ *b;
6849: PetscMPIInt size;
6851: PetscFunctionBegin;
6852: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
6854: PetscCall(PetscNew(&b));
6855: B->data = (void *)b;
6856: B->ops[0] = MatOps_Values;
6857: B->assembled = PETSC_FALSE;
6858: B->insertmode = NOT_SET_VALUES;
6859: b->size = size;
6861: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)B), &b->rank));
6863: /* build cache for off array entries formed */
6864: PetscCall(MatStashCreate_Private(PetscObjectComm((PetscObject)B), 1, &B->stash));
6866: b->donotstash = PETSC_FALSE;
6867: b->colmap = NULL;
6868: b->garray = NULL;
6869: b->roworiented = PETSC_TRUE;
6871: /* stuff used for matrix vector multiply */
6872: b->lvec = NULL;
6873: b->Mvctx = NULL;
6875: /* stuff for MatGetRow() */
6876: b->rowindices = NULL;
6877: b->rowvalues = NULL;
6878: b->getrowactive = PETSC_FALSE;
6880: /* flexible pointer used in CUSPARSE classes */
6881: b->spptr = NULL;
6883: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetUseScalableIncreaseOverlap_C", MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ));
6884: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatStoreValues_C", MatStoreValues_MPIAIJ));
6885: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatRetrieveValues_C", MatRetrieveValues_MPIAIJ));
6886: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatIsTranspose_C", MatIsTranspose_MPIAIJ));
6887: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetPreallocation_C", MatMPIAIJSetPreallocation_MPIAIJ));
6888: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatResetPreallocation_C", MatResetPreallocation_MPIAIJ));
6889: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatResetHash_C", MatResetHash_MPIAIJ));
6890: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetPreallocationCSR_C", MatMPIAIJSetPreallocationCSR_MPIAIJ));
6891: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDiagonalScaleLocal_C", MatDiagonalScaleLocal_MPIAIJ));
6892: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijperm_C", MatConvert_MPIAIJ_MPIAIJPERM));
6893: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijsell_C", MatConvert_MPIAIJ_MPIAIJSELL));
6894: #if PetscDefined(HAVE_CUDA)
6895: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijcusparse_C", MatConvert_MPIAIJ_MPIAIJCUSPARSE));
6896: #endif
6897: #if PetscDefined(HAVE_HIP)
6898: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijhipsparse_C", MatConvert_MPIAIJ_MPIAIJHIPSPARSE));
6899: #endif
6900: #if PetscDefined(HAVE_KOKKOS_KERNELS)
6901: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijkokkos_C", MatConvert_MPIAIJ_MPIAIJKokkos));
6902: #endif
6903: #if PetscDefined(HAVE_MKL_SPARSE)
6904: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijmkl_C", MatConvert_MPIAIJ_MPIAIJMKL));
6905: #endif
6906: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijcrl_C", MatConvert_MPIAIJ_MPIAIJCRL));
6907: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpibaij_C", MatConvert_MPIAIJ_MPIBAIJ));
6908: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpisbaij_C", MatConvert_MPIAIJ_MPISBAIJ));
6909: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpidense_C", MatConvert_MPIAIJ_MPIDense));
6910: #if PetscDefined(HAVE_ELEMENTAL)
6911: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_elemental_C", MatConvert_MPIAIJ_Elemental));
6912: #endif
6913: #if PetscDefined(HAVE_SCALAPACK) && (PetscDefined(USE_REAL_SINGLE) || PetscDefined(USE_REAL_DOUBLE))
6914: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_scalapack_C", MatConvert_AIJ_ScaLAPACK));
6915: #endif
6916: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_is_C", MatConvert_XAIJ_IS));
6917: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpisell_C", MatConvert_MPIAIJ_MPISELL));
6918: #if PetscDefined(HAVE_HYPRE)
6919: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_hypre_C", MatConvert_AIJ_HYPRE));
6920: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_transpose_mpiaij_mpiaij_C", MatProductSetFromOptions_Transpose_AIJ_AIJ));
6921: #endif
6922: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_is_mpiaij_C", MatProductSetFromOptions_IS_XAIJ));
6923: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_mpiaij_mpiaij_C", MatProductSetFromOptions_MPIAIJ));
6924: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetPreallocationCOO_C", MatSetPreallocationCOO_MPIAIJ));
6925: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetValuesCOO_C", MatSetValuesCOO_MPIAIJ));
6926: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatGetMultPetscSF_C", MatGetMultPetscSF_MPIAIJ));
6927: PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATMPIAIJ));
6928: PetscFunctionReturn(PETSC_SUCCESS);
6929: }
6931: /*@
6932: MatCreateMPIAIJWithSplitArrays - creates a `MATMPIAIJ` matrix using arrays that contain the "diagonal"
6933: and "off-diagonal" part of the matrix in CSR format.
6935: Collective
6937: Input Parameters:
6938: + comm - MPI communicator
6939: . m - number of local rows (Cannot be `PETSC_DECIDE`)
6940: . n - This value should be the same as the local size used in creating the
6941: x vector for the matrix-vector product $y = Ax$. (or `PETSC_DECIDE` to have
6942: calculated if `N` is given) For square matrices `n` is almost always `m`.
6943: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if `m` is given)
6944: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if `n` is given)
6945: . 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
6946: . j - column indices, which must be local, i.e., based off the start column of the diagonal portion
6947: . a - matrix values
6948: . 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
6949: . oj - column indices, which must be global, representing global columns in the `MATMPIAIJ` matrix
6950: - oa - matrix values
6952: Output Parameter:
6953: . mat - the matrix
6955: Level: advanced
6957: Notes:
6958: The `i`, `j`, and `a` arrays ARE NOT copied by this routine into the internal format used by PETSc (even in Fortran). The user
6959: must free the arrays once the matrix has been destroyed and not before.
6961: The `i` and `j` indices are 0 based
6963: See `MatCreateAIJ()` for the definition of "diagonal" and "off-diagonal" portion of the matrix
6965: This sets local rows and cannot be used to set off-processor values.
6967: Use of this routine is discouraged because it is inflexible and cumbersome to use. It is extremely rare that a
6968: legacy application natively assembles into exactly this split format. The code to do so is nontrivial and does
6969: not easily support in-place reassembly. It is recommended to use MatSetValues() (or a variant thereof) because
6970: the resulting assembly is easier to implement, will work with any matrix format, and the user does not have to
6971: keep track of the underlying array. Use `MatSetOption`(A,`MAT_NO_OFF_PROC_ENTRIES`,`PETSC_TRUE`) to disable all
6972: communication if it is known that only local entries will be set.
6974: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
6975: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithArrays()`
6976: @*/
6977: 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)
6978: {
6979: Mat_MPIAIJ *maij;
6981: PetscFunctionBegin;
6982: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
6983: PetscCheck(i[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
6984: PetscCheck(oi[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "oi (row indices) must start with 0");
6985: PetscCall(MatCreate(comm, mat));
6986: PetscCall(MatSetSizes(*mat, m, n, M, N));
6987: PetscCall(MatSetType(*mat, MATMPIAIJ));
6988: maij = (Mat_MPIAIJ *)(*mat)->data;
6990: (*mat)->preallocated = PETSC_TRUE;
6992: PetscCall(PetscLayoutSetUp((*mat)->rmap));
6993: PetscCall(PetscLayoutSetUp((*mat)->cmap));
6995: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, n, i, j, a, &maij->A));
6996: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, (*mat)->cmap->N, oi, oj, oa, &maij->B));
6998: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
6999: PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY));
7000: PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY));
7001: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_FALSE));
7002: PetscCall(MatSetOption(*mat, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
7003: PetscFunctionReturn(PETSC_SUCCESS);
7004: }
7006: typedef struct {
7007: Mat *mp; /* intermediate products */
7008: PetscBool *mptmp; /* is the intermediate product temporary ? */
7009: PetscInt cp; /* number of intermediate products */
7011: /* support for MatGetBrowsOfAoCols_MPIAIJ for P_oth */
7012: PetscInt *startsj_s, *startsj_r;
7013: PetscScalar *bufa;
7014: Mat P_oth;
7016: /* may take advantage of merging product->B */
7017: Mat Bloc; /* B-local by merging diag and off-diag */
7019: /* cusparse does not have support to split between symbolic and numeric phases.
7020: When api_user is true, we don't need to update the numerical values
7021: of the temporary storage */
7022: PetscBool reusesym;
7024: /* support for COO values insertion */
7025: PetscScalar *coo_v, *coo_w; /* store on-process and off-process COO scalars, and used as MPI recv/send buffers respectively */
7026: PetscInt **own; /* own[i] points to address of on-process COO indices for Mat mp[i] */
7027: PetscInt **off; /* off[i] points to address of off-process COO indices for Mat mp[i] */
7028: PetscBool hasoffproc; /* if true, have off-process values insertion (i.e. AtB or PtAP) */
7029: PetscSF sf; /* used for non-local values insertion and memory malloc */
7030: PetscMemType mtype;
7032: /* customization */
7033: PetscBool abmerge;
7034: PetscBool P_oth_bind;
7035: } MatMatMPIAIJBACKEND;
7037: static PetscErrorCode MatProductCtxDestroy_MatMatMPIAIJBACKEND(PetscCtxRt data)
7038: {
7039: MatMatMPIAIJBACKEND *mmdata = *(MatMatMPIAIJBACKEND **)data;
7040: PetscInt i;
7042: PetscFunctionBegin;
7043: PetscCall(PetscFree2(mmdata->startsj_s, mmdata->startsj_r));
7044: PetscCall(PetscFree(mmdata->bufa));
7045: PetscCall(PetscSFFree(mmdata->sf, mmdata->mtype, mmdata->coo_v));
7046: PetscCall(PetscSFFree(mmdata->sf, mmdata->mtype, mmdata->coo_w));
7047: PetscCall(MatDestroy(&mmdata->P_oth));
7048: PetscCall(MatDestroy(&mmdata->Bloc));
7049: PetscCall(PetscSFDestroy(&mmdata->sf));
7050: for (i = 0; i < mmdata->cp; i++) PetscCall(MatDestroy(&mmdata->mp[i]));
7051: PetscCall(PetscFree2(mmdata->mp, mmdata->mptmp));
7052: PetscCall(PetscFree(mmdata->own[0]));
7053: PetscCall(PetscFree(mmdata->own));
7054: PetscCall(PetscFree(mmdata->off[0]));
7055: PetscCall(PetscFree(mmdata->off));
7056: PetscCall(PetscFree(mmdata));
7057: PetscFunctionReturn(PETSC_SUCCESS);
7058: }
7060: /* Copy selected n entries with indices in idx[] of A to v[].
7061: If idx is NULL, copy the whole data array of A to v[]
7062: */
7063: static PetscErrorCode MatSeqAIJCopySubArray(Mat A, PetscInt n, const PetscInt idx[], PetscScalar v[])
7064: {
7065: PetscErrorCode (*f)(Mat, PetscInt, const PetscInt[], PetscScalar[]);
7067: PetscFunctionBegin;
7068: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatSeqAIJCopySubArray_C", &f));
7069: if (f) PetscCall((*f)(A, n, idx, v));
7070: else {
7071: const PetscScalar *vv;
7073: PetscCall(MatSeqAIJGetArrayRead(A, &vv));
7074: if (n && idx) {
7075: PetscScalar *w = v;
7076: const PetscInt *oi = idx;
7078: for (PetscInt j = 0; j < n; j++) *w++ = vv[*oi++];
7079: } else {
7080: PetscCall(PetscArraycpy(v, vv, n));
7081: }
7082: PetscCall(MatSeqAIJRestoreArrayRead(A, &vv));
7083: }
7084: PetscFunctionReturn(PETSC_SUCCESS);
7085: }
7087: static PetscErrorCode MatProductNumeric_MPIAIJBACKEND(Mat C)
7088: {
7089: MatMatMPIAIJBACKEND *mmdata;
7090: PetscInt i, n_d, n_o;
7092: PetscFunctionBegin;
7093: MatCheckProduct(C, 1);
7094: PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data empty");
7095: mmdata = (MatMatMPIAIJBACKEND *)C->product->data;
7096: if (!mmdata->reusesym) { /* update temporary matrices */
7097: 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));
7098: if (mmdata->Bloc) PetscCall(MatMPIAIJGetLocalMatMerge(C->product->B, MAT_REUSE_MATRIX, NULL, &mmdata->Bloc));
7099: }
7100: mmdata->reusesym = PETSC_FALSE;
7102: for (i = 0; i < mmdata->cp; i++) {
7103: 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]);
7104: PetscCall((*mmdata->mp[i]->ops->productnumeric)(mmdata->mp[i]));
7105: }
7106: for (i = 0, n_d = 0, n_o = 0; i < mmdata->cp; i++) {
7107: PetscInt noff;
7109: PetscCall(PetscIntCast(mmdata->off[i + 1] - mmdata->off[i], &noff));
7110: if (mmdata->mptmp[i]) continue;
7111: if (noff) {
7112: PetscInt nown;
7114: PetscCall(PetscIntCast(mmdata->own[i + 1] - mmdata->own[i], &nown));
7115: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], noff, mmdata->off[i], mmdata->coo_w + n_o));
7116: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], nown, mmdata->own[i], mmdata->coo_v + n_d));
7117: n_o += noff;
7118: n_d += nown;
7119: } else {
7120: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mmdata->mp[i]->data;
7122: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], mm->nz, NULL, mmdata->coo_v + n_d));
7123: n_d += mm->nz;
7124: }
7125: }
7126: if (mmdata->hasoffproc) { /* offprocess insertion */
7127: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_SCALAR, mmdata->coo_w, mmdata->coo_v + n_d));
7128: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_SCALAR, mmdata->coo_w, mmdata->coo_v + n_d));
7129: }
7130: PetscCall(MatSetValuesCOO(C, mmdata->coo_v, INSERT_VALUES));
7131: PetscFunctionReturn(PETSC_SUCCESS);
7132: }
7134: /* Support for Pt * A, A * P, or Pt * A * P */
7135: #define MAX_NUMBER_INTERMEDIATE 4
7136: PetscErrorCode MatProductSymbolic_MPIAIJBACKEND(Mat C)
7137: {
7138: Mat_Product *product = C->product;
7139: Mat A, P, mp[MAX_NUMBER_INTERMEDIATE]; /* A, P and a series of intermediate matrices */
7140: Mat_MPIAIJ *a, *p;
7141: MatMatMPIAIJBACKEND *mmdata;
7142: ISLocalToGlobalMapping P_oth_l2g = NULL;
7143: IS glob = NULL;
7144: const char *prefix;
7145: char pprefix[256];
7146: const PetscInt *globidx, *P_oth_idx;
7147: PetscInt i, j, cp, m, n, M, N, *coo_i, *coo_j;
7148: PetscCount ncoo, ncoo_d, ncoo_o, ncoo_oown;
7149: PetscInt cmapt[MAX_NUMBER_INTERMEDIATE], rmapt[MAX_NUMBER_INTERMEDIATE]; /* col/row map type for each Mat in mp[]. */
7150: /* type-0: consecutive, start from 0; type-1: consecutive with */
7151: /* a base offset; type-2: sparse with a local to global map table */
7152: const PetscInt *cmapa[MAX_NUMBER_INTERMEDIATE], *rmapa[MAX_NUMBER_INTERMEDIATE]; /* col/row local to global map array (table) for type-2 map type */
7154: MatProductType ptype;
7155: PetscBool mptmp[MAX_NUMBER_INTERMEDIATE], hasoffproc = PETSC_FALSE, iscuda, iship, iskokk;
7156: PetscMPIInt size;
7158: PetscFunctionBegin;
7159: MatCheckProduct(C, 1);
7160: PetscCheck(!product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data not empty");
7161: ptype = product->type;
7162: if (product->A->symmetric == PETSC_BOOL3_TRUE && ptype == MATPRODUCT_AtB) {
7163: ptype = MATPRODUCT_AB;
7164: product->symbolic_used_the_fact_A_is_symmetric = PETSC_TRUE;
7165: }
7166: switch (ptype) {
7167: case MATPRODUCT_AB:
7168: A = product->A;
7169: P = product->B;
7170: m = A->rmap->n;
7171: n = P->cmap->n;
7172: M = A->rmap->N;
7173: N = P->cmap->N;
7174: hasoffproc = PETSC_FALSE; /* will not scatter mat product values to other processes */
7175: break;
7176: case MATPRODUCT_AtB:
7177: P = product->A;
7178: A = product->B;
7179: m = P->cmap->n;
7180: n = A->cmap->n;
7181: M = P->cmap->N;
7182: N = A->cmap->N;
7183: hasoffproc = PETSC_TRUE;
7184: break;
7185: case MATPRODUCT_PtAP:
7186: A = product->A;
7187: P = product->B;
7188: m = P->cmap->n;
7189: n = P->cmap->n;
7190: M = P->cmap->N;
7191: N = P->cmap->N;
7192: hasoffproc = PETSC_TRUE;
7193: break;
7194: default:
7195: SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Not for product type %s", MatProductTypes[ptype]);
7196: }
7197: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)C), &size));
7198: if (size == 1) hasoffproc = PETSC_FALSE;
7200: /* defaults */
7201: for (i = 0; i < MAX_NUMBER_INTERMEDIATE; i++) {
7202: mp[i] = NULL;
7203: mptmp[i] = PETSC_FALSE;
7204: rmapt[i] = -1;
7205: cmapt[i] = -1;
7206: rmapa[i] = NULL;
7207: cmapa[i] = NULL;
7208: }
7210: /* customization */
7211: PetscCall(PetscNew(&mmdata));
7212: mmdata->reusesym = product->api_user;
7213: if (ptype == MATPRODUCT_AB) {
7214: if (product->api_user) {
7215: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatMatMult", "Mat");
7216: PetscCall(PetscOptionsBool("-matmatmult_backend_mergeB", "Merge product->B local matrices", "MatMatMult", mmdata->abmerge, &mmdata->abmerge, NULL));
7217: PetscCall(PetscOptionsBool("-matmatmult_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7218: PetscOptionsEnd();
7219: } else {
7220: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_AB", "Mat");
7221: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_mergeB", "Merge product->B local matrices", "MatMatMult", mmdata->abmerge, &mmdata->abmerge, NULL));
7222: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7223: PetscOptionsEnd();
7224: }
7225: } else if (ptype == MATPRODUCT_PtAP) {
7226: if (product->api_user) {
7227: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatPtAP", "Mat");
7228: PetscCall(PetscOptionsBool("-matptap_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7229: PetscOptionsEnd();
7230: } else {
7231: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_PtAP", "Mat");
7232: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7233: PetscOptionsEnd();
7234: }
7235: }
7236: a = (Mat_MPIAIJ *)A->data;
7237: p = (Mat_MPIAIJ *)P->data;
7238: PetscCall(MatSetSizes(C, m, n, M, N));
7239: PetscCall(PetscLayoutSetUp(C->rmap));
7240: PetscCall(PetscLayoutSetUp(C->cmap));
7241: PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
7242: PetscCall(MatGetOptionsPrefix(C, &prefix));
7244: cp = 0;
7245: switch (ptype) {
7246: case MATPRODUCT_AB: /* A * P */
7247: PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &mmdata->startsj_s, &mmdata->startsj_r, &mmdata->bufa, &mmdata->P_oth));
7249: /* A_diag * P_local (merged or not) */
7250: if (mmdata->abmerge) { /* P's diagonal and off-diag blocks are merged to one matrix, then multiplied by A_diag */
7251: /* P is product->B */
7252: PetscCall(MatMPIAIJGetLocalMatMerge(P, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7253: PetscCall(MatProductCreate(a->A, mmdata->Bloc, NULL, &mp[cp]));
7254: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7255: PetscCall(MatProductSetFill(mp[cp], product->fill));
7256: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7257: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7258: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7259: mp[cp]->product->api_user = product->api_user;
7260: PetscCall(MatProductSetFromOptions(mp[cp]));
7261: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7262: PetscCall(ISGetIndices(glob, &globidx));
7263: rmapt[cp] = 1;
7264: cmapt[cp] = 2;
7265: cmapa[cp] = globidx;
7266: mptmp[cp] = PETSC_FALSE;
7267: cp++;
7268: } else { /* A_diag * P_diag and A_diag * P_off */
7269: PetscCall(MatProductCreate(a->A, p->A, NULL, &mp[cp]));
7270: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7271: PetscCall(MatProductSetFill(mp[cp], product->fill));
7272: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7273: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7274: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7275: mp[cp]->product->api_user = product->api_user;
7276: PetscCall(MatProductSetFromOptions(mp[cp]));
7277: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7278: rmapt[cp] = 1;
7279: cmapt[cp] = 1;
7280: mptmp[cp] = PETSC_FALSE;
7281: cp++;
7282: PetscCall(MatProductCreate(a->A, p->B, NULL, &mp[cp]));
7283: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7284: PetscCall(MatProductSetFill(mp[cp], product->fill));
7285: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7286: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7287: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7288: mp[cp]->product->api_user = product->api_user;
7289: PetscCall(MatProductSetFromOptions(mp[cp]));
7290: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7291: rmapt[cp] = 1;
7292: cmapt[cp] = 2;
7293: cmapa[cp] = p->garray;
7294: mptmp[cp] = PETSC_FALSE;
7295: cp++;
7296: }
7298: /* A_off * P_other */
7299: if (mmdata->P_oth) {
7300: PetscCall(MatSeqAIJCompactOutExtraColumns_SeqAIJ(mmdata->P_oth, &P_oth_l2g)); /* make P_oth use local col ids */
7301: PetscCall(ISLocalToGlobalMappingGetIndices(P_oth_l2g, &P_oth_idx));
7302: PetscCall(MatSetType(mmdata->P_oth, ((PetscObject)a->B)->type_name));
7303: PetscCall(MatBindToCPU(mmdata->P_oth, mmdata->P_oth_bind));
7304: PetscCall(MatProductCreate(a->B, mmdata->P_oth, NULL, &mp[cp]));
7305: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7306: PetscCall(MatProductSetFill(mp[cp], product->fill));
7307: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7308: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7309: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7310: mp[cp]->product->api_user = product->api_user;
7311: PetscCall(MatProductSetFromOptions(mp[cp]));
7312: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7313: rmapt[cp] = 1;
7314: cmapt[cp] = 2;
7315: cmapa[cp] = P_oth_idx;
7316: mptmp[cp] = PETSC_FALSE;
7317: cp++;
7318: }
7319: break;
7321: case MATPRODUCT_AtB: /* (P^t * A): P_diag * A_loc + P_off * A_loc */
7322: /* A is product->B */
7323: PetscCall(MatMPIAIJGetLocalMatMerge(A, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7324: if (A == P) { /* when A==P, we can take advantage of the already merged mmdata->Bloc */
7325: PetscCall(MatProductCreate(mmdata->Bloc, mmdata->Bloc, NULL, &mp[cp]));
7326: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7327: PetscCall(MatProductSetFill(mp[cp], product->fill));
7328: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7329: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7330: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7331: mp[cp]->product->api_user = product->api_user;
7332: PetscCall(MatProductSetFromOptions(mp[cp]));
7333: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7334: PetscCall(ISGetIndices(glob, &globidx));
7335: rmapt[cp] = 2;
7336: rmapa[cp] = globidx;
7337: cmapt[cp] = 2;
7338: cmapa[cp] = globidx;
7339: mptmp[cp] = PETSC_FALSE;
7340: cp++;
7341: } else {
7342: PetscCall(MatProductCreate(p->A, mmdata->Bloc, NULL, &mp[cp]));
7343: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7344: PetscCall(MatProductSetFill(mp[cp], product->fill));
7345: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7346: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7347: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7348: mp[cp]->product->api_user = product->api_user;
7349: PetscCall(MatProductSetFromOptions(mp[cp]));
7350: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7351: PetscCall(ISGetIndices(glob, &globidx));
7352: rmapt[cp] = 1;
7353: cmapt[cp] = 2;
7354: cmapa[cp] = globidx;
7355: mptmp[cp] = PETSC_FALSE;
7356: cp++;
7357: PetscCall(MatProductCreate(p->B, mmdata->Bloc, NULL, &mp[cp]));
7358: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7359: PetscCall(MatProductSetFill(mp[cp], product->fill));
7360: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7361: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7362: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7363: mp[cp]->product->api_user = product->api_user;
7364: PetscCall(MatProductSetFromOptions(mp[cp]));
7365: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7366: rmapt[cp] = 2;
7367: rmapa[cp] = p->garray;
7368: cmapt[cp] = 2;
7369: cmapa[cp] = globidx;
7370: mptmp[cp] = PETSC_FALSE;
7371: cp++;
7372: }
7373: break;
7374: case MATPRODUCT_PtAP:
7375: PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &mmdata->startsj_s, &mmdata->startsj_r, &mmdata->bufa, &mmdata->P_oth));
7376: /* P is product->B */
7377: PetscCall(MatMPIAIJGetLocalMatMerge(P, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7378: PetscCall(MatProductCreate(a->A, mmdata->Bloc, NULL, &mp[cp]));
7379: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_PtAP));
7380: PetscCall(MatProductSetFill(mp[cp], product->fill));
7381: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7382: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7383: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7384: mp[cp]->product->api_user = product->api_user;
7385: PetscCall(MatProductSetFromOptions(mp[cp]));
7386: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7387: PetscCall(ISGetIndices(glob, &globidx));
7388: rmapt[cp] = 2;
7389: rmapa[cp] = globidx;
7390: cmapt[cp] = 2;
7391: cmapa[cp] = globidx;
7392: mptmp[cp] = PETSC_FALSE;
7393: cp++;
7394: if (mmdata->P_oth) {
7395: PetscCall(MatSeqAIJCompactOutExtraColumns_SeqAIJ(mmdata->P_oth, &P_oth_l2g));
7396: PetscCall(ISLocalToGlobalMappingGetIndices(P_oth_l2g, &P_oth_idx));
7397: PetscCall(MatSetType(mmdata->P_oth, ((PetscObject)a->B)->type_name));
7398: PetscCall(MatBindToCPU(mmdata->P_oth, mmdata->P_oth_bind));
7399: PetscCall(MatProductCreate(a->B, mmdata->P_oth, NULL, &mp[cp]));
7400: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7401: PetscCall(MatProductSetFill(mp[cp], product->fill));
7402: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7403: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7404: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7405: mp[cp]->product->api_user = product->api_user;
7406: PetscCall(MatProductSetFromOptions(mp[cp]));
7407: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7408: mptmp[cp] = PETSC_TRUE;
7409: cp++;
7410: PetscCall(MatProductCreate(mmdata->Bloc, mp[1], NULL, &mp[cp]));
7411: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7412: PetscCall(MatProductSetFill(mp[cp], product->fill));
7413: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7414: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7415: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7416: mp[cp]->product->api_user = product->api_user;
7417: PetscCall(MatProductSetFromOptions(mp[cp]));
7418: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7419: rmapt[cp] = 2;
7420: rmapa[cp] = globidx;
7421: cmapt[cp] = 2;
7422: cmapa[cp] = P_oth_idx;
7423: mptmp[cp] = PETSC_FALSE;
7424: cp++;
7425: }
7426: break;
7427: default:
7428: SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Not for product type %s", MatProductTypes[ptype]);
7429: }
7430: /* sanity check */
7431: if (size > 1)
7432: 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);
7434: PetscCall(PetscMalloc2(cp, &mmdata->mp, cp, &mmdata->mptmp));
7435: for (i = 0; i < cp; i++) {
7436: mmdata->mp[i] = mp[i];
7437: mmdata->mptmp[i] = mptmp[i];
7438: }
7439: mmdata->cp = cp;
7440: C->product->data = mmdata;
7441: C->product->destroy = MatProductCtxDestroy_MatMatMPIAIJBACKEND;
7442: C->ops->productnumeric = MatProductNumeric_MPIAIJBACKEND;
7444: /* memory type */
7445: mmdata->mtype = PETSC_MEMTYPE_HOST;
7446: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iscuda, MATSEQAIJCUSPARSE, MATMPIAIJCUSPARSE, ""));
7447: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iship, MATSEQAIJHIPSPARSE, MATMPIAIJHIPSPARSE, ""));
7448: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iskokk, MATSEQAIJKOKKOS, MATMPIAIJKOKKOS, ""));
7449: if (iscuda) mmdata->mtype = PETSC_MEMTYPE_CUDA;
7450: else if (iship) mmdata->mtype = PETSC_MEMTYPE_HIP;
7451: else if (iskokk) mmdata->mtype = PETSC_MEMTYPE_KOKKOS;
7453: /* prepare coo coordinates for values insertion */
7455: /* count total nonzeros of those intermediate seqaij Mats
7456: ncoo_d: # of nonzeros of matrices that do not have offproc entries
7457: ncoo_o: # of nonzeros (of matrices that might have offproc entries) that will be inserted to remote procs
7458: ncoo_oown: # of nonzeros (of matrices that might have offproc entries) that will be inserted locally
7459: */
7460: for (cp = 0, ncoo_d = 0, ncoo_o = 0, ncoo_oown = 0; cp < mmdata->cp; cp++) {
7461: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7462: if (mptmp[cp]) continue;
7463: if (rmapt[cp] == 2 && hasoffproc) { /* the rows need to be scatter to all processes (might include self) */
7464: const PetscInt *rmap = rmapa[cp];
7465: const PetscInt mr = mp[cp]->rmap->n;
7466: const PetscInt rs = C->rmap->rstart;
7467: const PetscInt re = C->rmap->rend;
7468: const PetscInt *ii = mm->i;
7469: for (i = 0; i < mr; i++) {
7470: const PetscInt gr = rmap[i];
7471: const PetscInt nz = ii[i + 1] - ii[i];
7472: if (gr < rs || gr >= re) ncoo_o += nz; /* this row is offproc */
7473: else ncoo_oown += nz; /* this row is local */
7474: }
7475: } else ncoo_d += mm->nz;
7476: }
7478: /*
7479: ncoo: total number of nonzeros (including those inserted by remote procs) belonging to this proc
7481: ncoo = ncoo_d + ncoo_oown + ncoo2, which ncoo2 is number of nonzeros inserted to me by other procs.
7483: off[0] points to a big index array, which is shared by off[1,2,...]. Similarly, for own[0].
7485: off[p]: points to the segment for matrix mp[p], storing location of nonzeros that mp[p] will insert to others
7486: own[p]: points to the segment for matrix mp[p], storing location of nonzeros that mp[p] will insert locally
7487: so, off[p+1]-off[p] is the number of nonzeros that mp[p] will send to others.
7489: coo_i/j/v[]: [ncoo] row/col/val of nonzeros belonging to this proc.
7490: 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.
7491: */
7492: PetscCall(PetscCalloc1(mmdata->cp + 1, &mmdata->off)); /* +1 to make a csr-like data structure */
7493: PetscCall(PetscCalloc1(mmdata->cp + 1, &mmdata->own));
7495: /* gather (i,j) of nonzeros inserted by remote procs */
7496: if (hasoffproc) {
7497: PetscSF msf;
7498: PetscInt ncoo2, *coo_i2, *coo_j2;
7500: PetscCall(PetscMalloc1(ncoo_o, &mmdata->off[0]));
7501: PetscCall(PetscMalloc1(ncoo_oown, &mmdata->own[0]));
7502: PetscCall(PetscMalloc2(ncoo_o, &coo_i, ncoo_o, &coo_j)); /* to collect (i,j) of entries to be sent to others */
7504: for (cp = 0, ncoo_o = 0; cp < mmdata->cp; cp++) {
7505: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7506: PetscInt *idxoff = mmdata->off[cp];
7507: PetscInt *idxown = mmdata->own[cp];
7508: if (!mptmp[cp] && rmapt[cp] == 2) { /* row map is sparse */
7509: const PetscInt *rmap = rmapa[cp];
7510: const PetscInt *cmap = cmapa[cp];
7511: const PetscInt *ii = mm->i;
7512: PetscInt *coi = coo_i + ncoo_o;
7513: PetscInt *coj = coo_j + ncoo_o;
7514: const PetscInt mr = mp[cp]->rmap->n;
7515: const PetscInt rs = C->rmap->rstart;
7516: const PetscInt re = C->rmap->rend;
7517: const PetscInt cs = C->cmap->rstart;
7518: for (i = 0; i < mr; i++) {
7519: const PetscInt *jj = mm->j + ii[i];
7520: const PetscInt gr = rmap[i];
7521: const PetscInt nz = ii[i + 1] - ii[i];
7522: if (gr < rs || gr >= re) { /* this is an offproc row */
7523: for (j = ii[i]; j < ii[i + 1]; j++) {
7524: *coi++ = gr;
7525: *idxoff++ = j;
7526: }
7527: if (!cmapt[cp]) { /* already global */
7528: for (j = 0; j < nz; j++) *coj++ = jj[j];
7529: } else if (cmapt[cp] == 1) { /* local to global for owned columns of C */
7530: for (j = 0; j < nz; j++) *coj++ = jj[j] + cs;
7531: } else { /* offdiag */
7532: for (j = 0; j < nz; j++) *coj++ = cmap[jj[j]];
7533: }
7534: ncoo_o += nz;
7535: } else { /* this is a local row */
7536: for (j = ii[i]; j < ii[i + 1]; j++) *idxown++ = j;
7537: }
7538: }
7539: }
7540: mmdata->off[cp + 1] = idxoff;
7541: mmdata->own[cp + 1] = idxown;
7542: }
7544: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)C), &mmdata->sf));
7545: PetscInt incoo_o;
7546: PetscCall(PetscIntCast(ncoo_o, &incoo_o));
7547: PetscCall(PetscSFSetGraphLayout(mmdata->sf, C->rmap, incoo_o /*nleaves*/, NULL /*ilocal*/, PETSC_OWN_POINTER, coo_i));
7548: PetscCall(PetscSFGetMultiSF(mmdata->sf, &msf));
7549: PetscCall(PetscSFGetGraph(msf, &ncoo2 /*nroots*/, NULL, NULL, NULL));
7550: ncoo = ncoo_d + ncoo_oown + ncoo2;
7551: PetscCall(PetscMalloc2(ncoo, &coo_i2, ncoo, &coo_j2));
7552: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_INT, coo_i, coo_i2 + ncoo_d + ncoo_oown)); /* put (i,j) of remote nonzeros at back */
7553: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_INT, coo_i, coo_i2 + ncoo_d + ncoo_oown));
7554: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_INT, coo_j, coo_j2 + ncoo_d + ncoo_oown));
7555: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_INT, coo_j, coo_j2 + ncoo_d + ncoo_oown));
7556: PetscCall(PetscFree2(coo_i, coo_j));
7557: /* allocate MPI send buffer to collect nonzero values to be sent to remote procs */
7558: PetscCall(PetscSFMalloc(mmdata->sf, mmdata->mtype, ncoo_o * sizeof(PetscScalar), (void **)&mmdata->coo_w));
7559: coo_i = coo_i2;
7560: coo_j = coo_j2;
7561: } else { /* no offproc values insertion */
7562: ncoo = ncoo_d;
7563: PetscCall(PetscMalloc2(ncoo, &coo_i, ncoo, &coo_j));
7565: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)C), &mmdata->sf));
7566: PetscCall(PetscSFSetGraph(mmdata->sf, 0, 0, NULL, PETSC_OWN_POINTER, NULL, PETSC_OWN_POINTER));
7567: PetscCall(PetscSFSetUp(mmdata->sf));
7568: }
7569: mmdata->hasoffproc = hasoffproc;
7571: /* gather (i,j) of nonzeros inserted locally */
7572: for (cp = 0, ncoo_d = 0; cp < mmdata->cp; cp++) {
7573: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7574: PetscInt *coi = coo_i + ncoo_d;
7575: PetscInt *coj = coo_j + ncoo_d;
7576: const PetscInt *jj = mm->j;
7577: const PetscInt *ii = mm->i;
7578: const PetscInt *cmap = cmapa[cp];
7579: const PetscInt *rmap = rmapa[cp];
7580: const PetscInt mr = mp[cp]->rmap->n;
7581: const PetscInt rs = C->rmap->rstart;
7582: const PetscInt re = C->rmap->rend;
7583: const PetscInt cs = C->cmap->rstart;
7585: if (mptmp[cp]) continue;
7586: if (rmapt[cp] == 1) { /* consecutive rows */
7587: /* fill coo_i */
7588: for (i = 0; i < mr; i++) {
7589: const PetscInt gr = i + rs;
7590: for (j = ii[i]; j < ii[i + 1]; j++) coi[j] = gr;
7591: }
7592: /* fill coo_j */
7593: if (!cmapt[cp]) { /* type-0, already global */
7594: PetscCall(PetscArraycpy(coj, jj, mm->nz));
7595: } else if (cmapt[cp] == 1) { /* type-1, local to global for consecutive columns of C */
7596: for (j = 0; j < mm->nz; j++) coj[j] = jj[j] + cs; /* lid + col start */
7597: } else { /* type-2, local to global for sparse columns */
7598: for (j = 0; j < mm->nz; j++) coj[j] = cmap[jj[j]];
7599: }
7600: ncoo_d += mm->nz;
7601: } else if (rmapt[cp] == 2) { /* sparse rows */
7602: for (i = 0; i < mr; i++) {
7603: const PetscInt *jj = mm->j + ii[i];
7604: const PetscInt gr = rmap[i];
7605: const PetscInt nz = ii[i + 1] - ii[i];
7606: if (gr >= rs && gr < re) { /* local rows */
7607: for (j = ii[i]; j < ii[i + 1]; j++) *coi++ = gr;
7608: if (!cmapt[cp]) { /* type-0, already global */
7609: for (j = 0; j < nz; j++) *coj++ = jj[j];
7610: } else if (cmapt[cp] == 1) { /* local to global for owned columns of C */
7611: for (j = 0; j < nz; j++) *coj++ = jj[j] + cs;
7612: } else { /* type-2, local to global for sparse columns */
7613: for (j = 0; j < nz; j++) *coj++ = cmap[jj[j]];
7614: }
7615: ncoo_d += nz;
7616: }
7617: }
7618: }
7619: }
7620: if (glob) PetscCall(ISRestoreIndices(glob, &globidx));
7621: PetscCall(ISDestroy(&glob));
7622: if (P_oth_l2g) PetscCall(ISLocalToGlobalMappingRestoreIndices(P_oth_l2g, &P_oth_idx));
7623: PetscCall(ISLocalToGlobalMappingDestroy(&P_oth_l2g));
7624: /* allocate an array to store all nonzeros (inserted locally or remotely) belonging to this proc */
7625: PetscCall(PetscSFMalloc(mmdata->sf, mmdata->mtype, ncoo * sizeof(PetscScalar), (void **)&mmdata->coo_v));
7627: /* set block sizes */
7628: A = product->A;
7629: P = product->B;
7630: switch (ptype) {
7631: case MATPRODUCT_PtAP:
7632: PetscCall(MatSetBlockSizes(C, P->cmap->bs, P->cmap->bs));
7633: break;
7634: case MATPRODUCT_RARt:
7635: PetscCall(MatSetBlockSizes(C, P->rmap->bs, P->rmap->bs));
7636: break;
7637: case MATPRODUCT_ABC:
7638: PetscCall(MatSetBlockSizesFromMats(C, A, product->C));
7639: break;
7640: case MATPRODUCT_AB:
7641: PetscCall(MatSetBlockSizesFromMats(C, A, P));
7642: break;
7643: case MATPRODUCT_AtB:
7644: PetscCall(MatSetBlockSizes(C, A->cmap->bs, P->cmap->bs));
7645: break;
7646: case MATPRODUCT_ABt:
7647: PetscCall(MatSetBlockSizes(C, A->rmap->bs, P->rmap->bs));
7648: break;
7649: default:
7650: SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Not for ProductType %s", MatProductTypes[ptype]);
7651: }
7653: /* preallocate with COO data */
7654: PetscCall(MatSetPreallocationCOO(C, ncoo, coo_i, coo_j));
7655: PetscCall(PetscFree2(coo_i, coo_j));
7656: PetscFunctionReturn(PETSC_SUCCESS);
7657: }
7659: PetscErrorCode MatProductSetFromOptions_MPIAIJBACKEND(Mat mat)
7660: {
7661: Mat_Product *product = mat->product;
7662: #if PetscDefined(HAVE_DEVICE)
7663: PetscBool match = PETSC_FALSE;
7664: PetscBool usecpu = PETSC_FALSE;
7665: #else
7666: PetscBool match = PETSC_TRUE;
7667: #endif
7669: PetscFunctionBegin;
7670: MatCheckProduct(mat, 1);
7671: #if PetscDefined(HAVE_DEVICE)
7672: if (!product->A->boundtocpu && !product->B->boundtocpu) PetscCall(PetscObjectTypeCompare((PetscObject)product->B, ((PetscObject)product->A)->type_name, &match));
7673: if (match) { /* we can always fallback to the CPU if requested */
7674: switch (product->type) {
7675: case MATPRODUCT_AB:
7676: if (product->api_user) {
7677: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatMatMult", "Mat");
7678: PetscCall(PetscOptionsBool("-matmatmult_backend_cpu", "Use CPU code", "MatMatMult", usecpu, &usecpu, NULL));
7679: PetscOptionsEnd();
7680: } else {
7681: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_AB", "Mat");
7682: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatMatMult", usecpu, &usecpu, NULL));
7683: PetscOptionsEnd();
7684: }
7685: break;
7686: case MATPRODUCT_AtB:
7687: if (product->api_user) {
7688: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatTransposeMatMult", "Mat");
7689: PetscCall(PetscOptionsBool("-mattransposematmult_backend_cpu", "Use CPU code", "MatTransposeMatMult", usecpu, &usecpu, NULL));
7690: PetscOptionsEnd();
7691: } else {
7692: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_AtB", "Mat");
7693: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatTransposeMatMult", usecpu, &usecpu, NULL));
7694: PetscOptionsEnd();
7695: }
7696: break;
7697: case MATPRODUCT_PtAP:
7698: if (product->api_user) {
7699: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatPtAP", "Mat");
7700: PetscCall(PetscOptionsBool("-matptap_backend_cpu", "Use CPU code", "MatPtAP", usecpu, &usecpu, NULL));
7701: PetscOptionsEnd();
7702: } else {
7703: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_PtAP", "Mat");
7704: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatPtAP", usecpu, &usecpu, NULL));
7705: PetscOptionsEnd();
7706: }
7707: break;
7708: default:
7709: break;
7710: }
7711: match = (PetscBool)!usecpu;
7712: }
7713: #endif
7714: if (match) {
7715: switch (product->type) {
7716: case MATPRODUCT_AB:
7717: case MATPRODUCT_AtB:
7718: case MATPRODUCT_PtAP:
7719: mat->ops->productsymbolic = MatProductSymbolic_MPIAIJBACKEND;
7720: break;
7721: default:
7722: break;
7723: }
7724: }
7725: /* fallback to MPIAIJ ops */
7726: if (!mat->ops->productsymbolic) PetscCall(MatProductSetFromOptions_MPIAIJ(mat));
7727: PetscFunctionReturn(PETSC_SUCCESS);
7728: }
7730: /*
7731: Produces a set of block column indices of the matrix row, one for each block represented in the original row
7733: n - the number of block indices in cc[]
7734: cc - the block indices (must be large enough to contain the indices)
7735: */
7736: static inline PetscErrorCode MatCollapseRow(Mat Amat, PetscInt row, PetscInt bs, PetscInt *n, PetscInt *cc)
7737: {
7738: PetscInt cnt = -1, nidx, j;
7739: const PetscInt *idx;
7741: PetscFunctionBegin;
7742: PetscCall(MatGetRow(Amat, row, &nidx, &idx, NULL));
7743: if (nidx) {
7744: cnt = 0;
7745: cc[cnt] = idx[0] / bs;
7746: for (j = 1; j < nidx; j++) {
7747: if (cc[cnt] < idx[j] / bs) cc[++cnt] = idx[j] / bs;
7748: }
7749: }
7750: PetscCall(MatRestoreRow(Amat, row, &nidx, &idx, NULL));
7751: *n = cnt + 1;
7752: PetscFunctionReturn(PETSC_SUCCESS);
7753: }
7755: /*
7756: Produces a set of block column indices of the matrix block row, one for each block represented in the original set of rows
7758: ncollapsed - the number of block indices
7759: collapsed - the block indices (must be large enough to contain the indices)
7760: */
7761: static inline PetscErrorCode MatCollapseRows(Mat Amat, PetscInt start, PetscInt bs, PetscInt *w0, PetscInt *w1, PetscInt *w2, PetscInt *ncollapsed, PetscInt **collapsed)
7762: {
7763: PetscInt i, nprev, *cprev = w0, ncur = 0, *ccur = w1, *merged = w2, *cprevtmp;
7765: PetscFunctionBegin;
7766: PetscCall(MatCollapseRow(Amat, start, bs, &nprev, cprev));
7767: for (i = start + 1; i < start + bs; i++) {
7768: PetscCall(MatCollapseRow(Amat, i, bs, &ncur, ccur));
7769: PetscCall(PetscMergeIntArray(nprev, cprev, ncur, ccur, &nprev, &merged));
7770: cprevtmp = cprev;
7771: cprev = merged;
7772: merged = cprevtmp;
7773: }
7774: *ncollapsed = nprev;
7775: if (collapsed) *collapsed = cprev;
7776: PetscFunctionReturn(PETSC_SUCCESS);
7777: }
7779: // PetscClangLinter pragma disable: -fdoc-sowing-chars
7780: /*
7781: MatCreateGraph_Simple_AIJ - create simple scalar matrix (graph) from potentially blocked matrix
7783: Input Parameters:
7784: + Amat - matrix
7785: . symmetrize - make the result symmetric
7786: . scale - scale with diagonal
7787: . filter - threshold for filter
7788: . index_size - length of `index`
7789: - index - indices of unknown purpose
7791: Output Parameter:
7792: . a_Gmat - output scalar graph >= 0
7794: Level: developer
7796: .seealso: `MATAIJ`
7797: */
7798: PETSC_INTERN PetscErrorCode MatCreateGraph_Simple_AIJ(Mat Amat, PetscBool symmetrize, PetscBool scale, PetscReal filter, PetscInt index_size, PetscInt index[], Mat *a_Gmat)
7799: {
7800: PetscInt Istart, Iend, Ii, jj, kk, ncols, nloc, NN, MM, bs;
7801: MPI_Comm comm;
7802: Mat Gmat;
7803: PetscBool ismpiaij, isseqaij;
7804: Mat a, b, c;
7805: MatType jtype;
7807: PetscFunctionBegin;
7808: PetscCall(PetscObjectGetComm((PetscObject)Amat, &comm));
7809: PetscCall(MatGetOwnershipRange(Amat, &Istart, &Iend));
7810: PetscCall(MatGetSize(Amat, &MM, &NN));
7811: PetscCall(MatGetBlockSize(Amat, &bs));
7812: nloc = (Iend - Istart) / bs;
7814: PetscCall(PetscObjectBaseTypeCompare((PetscObject)Amat, MATSEQAIJ, &isseqaij));
7815: PetscCall(PetscObjectBaseTypeCompare((PetscObject)Amat, MATMPIAIJ, &ismpiaij));
7816: PetscCheck(isseqaij || ismpiaij, comm, PETSC_ERR_USER, "Require (MPI)AIJ matrix type");
7818: /* TODO GPU: these calls are potentially expensive if matrices are large and we want to use the GPU */
7819: /* A solution consists in providing a new API, MatAIJGetCollapsedAIJ, and each class can provide a fast
7820: implementation */
7821: if (bs > 1) {
7822: PetscCall(MatGetType(Amat, &jtype));
7823: PetscCall(MatCreate(comm, &Gmat));
7824: PetscCall(MatSetType(Gmat, jtype));
7825: PetscCall(MatSetSizes(Gmat, nloc, nloc, PETSC_DETERMINE, PETSC_DETERMINE));
7826: PetscCall(MatSetBlockSizes(Gmat, 1, 1));
7827: if (isseqaij || ((Mat_MPIAIJ *)Amat->data)->garray) {
7828: PetscInt *d_nnz, *o_nnz;
7829: MatScalar *aa, val, *AA;
7830: PetscInt *aj, *ai, *AJ, nc, nmax = 0;
7832: if (isseqaij) {
7833: a = Amat;
7834: b = NULL;
7835: } else {
7836: Mat_MPIAIJ *d = (Mat_MPIAIJ *)Amat->data;
7837: a = d->A;
7838: b = d->B;
7839: }
7840: PetscCall(PetscInfo(Amat, "New bs>1 Graph. nloc=%" PetscInt_FMT "\n", nloc));
7841: PetscCall(PetscMalloc2(nloc, &d_nnz, (isseqaij ? 0 : nloc), &o_nnz));
7842: for (c = a, kk = 0; c && kk < 2; c = b, kk++) {
7843: PetscInt *nnz = (c == a) ? d_nnz : o_nnz;
7844: const PetscInt *cols1, *cols2;
7846: for (PetscInt brow = 0, nc1, nc2, ok = 1; brow < nloc * bs; brow += bs) { // block rows
7847: PetscCall(MatGetRow(c, brow, &nc2, &cols2, NULL));
7848: nnz[brow / bs] = nc2 / bs;
7849: if (nc2 % bs) ok = 0;
7850: if (nnz[brow / bs] > nmax) nmax = nnz[brow / bs];
7851: for (PetscInt ii = 1; ii < bs; ii++) { // check for non-dense blocks
7852: PetscCall(MatGetRow(c, brow + ii, &nc1, &cols1, NULL));
7853: if (nc1 != nc2) ok = 0;
7854: else {
7855: for (PetscInt jj = 0; jj < nc1 && ok == 1; jj++) {
7856: if (cols1[jj] != cols2[jj]) ok = 0;
7857: if (cols1[jj] % bs != jj % bs) ok = 0;
7858: }
7859: }
7860: PetscCall(MatRestoreRow(c, brow + ii, &nc1, &cols1, NULL));
7861: }
7862: PetscCall(MatRestoreRow(c, brow, &nc2, &cols2, NULL));
7863: if (!ok) {
7864: PetscCall(PetscFree2(d_nnz, o_nnz));
7865: PetscCall(PetscInfo(Amat, "Found sparse blocks - revert to slow method\n"));
7866: goto old_bs;
7867: }
7868: }
7869: }
7870: PetscCall(MatSeqAIJSetPreallocation(Gmat, 0, d_nnz));
7871: PetscCall(MatMPIAIJSetPreallocation(Gmat, 0, d_nnz, 0, o_nnz));
7872: PetscCall(PetscFree2(d_nnz, o_nnz));
7873: PetscCall(PetscMalloc2(nmax, &AA, nmax, &AJ));
7874: // diag
7875: for (PetscInt brow = 0, n, grow; brow < nloc * bs; brow += bs) { // block rows
7876: Mat_SeqAIJ *aseq = (Mat_SeqAIJ *)a->data;
7878: ai = aseq->i;
7879: n = ai[brow + 1] - ai[brow];
7880: aj = aseq->j + ai[brow];
7881: for (PetscInt k = 0; k < n; k += bs) { // block columns
7882: AJ[k / bs] = aj[k] / bs + Istart / bs; // diag starts at (Istart,Istart)
7883: val = 0;
7884: if (index_size == 0) {
7885: for (PetscInt ii = 0; ii < bs; ii++) { // rows in block
7886: aa = aseq->a + ai[brow + ii] + k;
7887: for (PetscInt jj = 0; jj < bs; jj++) { // columns in block
7888: val += PetscAbs(PetscRealPart(aa[jj])); // a sort of norm
7889: }
7890: }
7891: } else { // use (index,index) value if provided
7892: for (PetscInt iii = 0; iii < index_size; iii++) { // rows in block
7893: PetscInt ii = index[iii];
7894: aa = aseq->a + ai[brow + ii] + k;
7895: for (PetscInt jjj = 0; jjj < index_size; jjj++) { // columns in block
7896: PetscInt jj = index[jjj];
7897: val += PetscAbs(PetscRealPart(aa[jj]));
7898: }
7899: }
7900: }
7901: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs (%" PetscInt_FMT ") >= nmax (%" PetscInt_FMT ")", k / bs, nmax);
7902: AA[k / bs] = val;
7903: }
7904: grow = Istart / bs + brow / bs;
7905: PetscCall(MatSetValues(Gmat, 1, &grow, n / bs, AJ, AA, ADD_VALUES));
7906: }
7907: // off-diag
7908: if (ismpiaij) {
7909: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)Amat->data;
7910: const PetscScalar *vals;
7911: const PetscInt *cols, *garray = aij->garray;
7913: PetscCheck(garray, PETSC_COMM_SELF, PETSC_ERR_USER, "No garray ?");
7914: for (PetscInt brow = 0, grow; brow < nloc * bs; brow += bs) { // block rows
7915: PetscCall(MatGetRow(b, brow, &ncols, &cols, NULL));
7916: for (PetscInt k = 0, cidx = 0; k < ncols; k += bs, cidx++) {
7917: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs >= nmax");
7918: AA[k / bs] = 0;
7919: AJ[cidx] = garray[cols[k]] / bs;
7920: }
7921: nc = ncols / bs;
7922: PetscCall(MatRestoreRow(b, brow, &ncols, &cols, NULL));
7923: if (index_size == 0) {
7924: for (PetscInt ii = 0; ii < bs; ii++) { // rows in block
7925: PetscCall(MatGetRow(b, brow + ii, &ncols, &cols, &vals));
7926: for (PetscInt k = 0; k < ncols; k += bs) {
7927: for (PetscInt jj = 0; jj < bs; jj++) { // cols in block
7928: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs (%" PetscInt_FMT ") >= nmax (%" PetscInt_FMT ")", k / bs, nmax);
7929: AA[k / bs] += PetscAbs(PetscRealPart(vals[k + jj]));
7930: }
7931: }
7932: PetscCall(MatRestoreRow(b, brow + ii, &ncols, &cols, &vals));
7933: }
7934: } else { // use (index,index) value if provided
7935: for (PetscInt iii = 0; iii < index_size; iii++) { // rows in block
7936: PetscInt ii = index[iii];
7937: PetscCall(MatGetRow(b, brow + ii, &ncols, &cols, &vals));
7938: for (PetscInt k = 0; k < ncols; k += bs) {
7939: for (PetscInt jjj = 0; jjj < index_size; jjj++) { // cols in block
7940: PetscInt jj = index[jjj];
7941: AA[k / bs] += PetscAbs(PetscRealPart(vals[k + jj]));
7942: }
7943: }
7944: PetscCall(MatRestoreRow(b, brow + ii, &ncols, &cols, &vals));
7945: }
7946: }
7947: grow = Istart / bs + brow / bs;
7948: PetscCall(MatSetValues(Gmat, 1, &grow, nc, AJ, AA, ADD_VALUES));
7949: }
7950: }
7951: PetscCall(MatAssemblyBegin(Gmat, MAT_FINAL_ASSEMBLY));
7952: PetscCall(MatAssemblyEnd(Gmat, MAT_FINAL_ASSEMBLY));
7953: PetscCall(PetscFree2(AA, AJ));
7954: } else {
7955: const PetscScalar *vals;
7956: const PetscInt *idx;
7957: PetscInt *d_nnz, *o_nnz, *w0, *w1, *w2;
7958: old_bs:
7959: /*
7960: Determine the preallocation needed for the scalar matrix derived from the vector matrix.
7961: */
7962: PetscCall(PetscInfo(Amat, "OLD bs>1 CreateGraph\n"));
7963: PetscCall(PetscMalloc2(nloc, &d_nnz, (isseqaij ? 0 : nloc), &o_nnz));
7964: if (isseqaij) {
7965: PetscInt max_d_nnz;
7967: /*
7968: Determine exact preallocation count for (sequential) scalar matrix
7969: */
7970: PetscCall(MatSeqAIJGetMaxRowNonzeros(Amat, &max_d_nnz));
7971: max_d_nnz = PetscMin(nloc, bs * max_d_nnz);
7972: PetscCall(PetscMalloc3(max_d_nnz, &w0, max_d_nnz, &w1, max_d_nnz, &w2));
7973: for (Ii = 0, jj = 0; Ii < Iend; Ii += bs, jj++) PetscCall(MatCollapseRows(Amat, Ii, bs, w0, w1, w2, &d_nnz[jj], NULL));
7974: PetscCall(PetscFree3(w0, w1, w2));
7975: } else if (ismpiaij) {
7976: Mat Daij, Oaij;
7977: const PetscInt *garray;
7978: PetscInt max_d_nnz;
7980: PetscCall(MatMPIAIJGetSeqAIJ(Amat, &Daij, &Oaij, &garray));
7981: /*
7982: Determine exact preallocation count for diagonal block portion of scalar matrix
7983: */
7984: PetscCall(MatSeqAIJGetMaxRowNonzeros(Daij, &max_d_nnz));
7985: max_d_nnz = PetscMin(nloc, bs * max_d_nnz);
7986: PetscCall(PetscMalloc3(max_d_nnz, &w0, max_d_nnz, &w1, max_d_nnz, &w2));
7987: for (Ii = 0, jj = 0; Ii < Iend - Istart; Ii += bs, jj++) PetscCall(MatCollapseRows(Daij, Ii, bs, w0, w1, w2, &d_nnz[jj], NULL));
7988: PetscCall(PetscFree3(w0, w1, w2));
7989: /*
7990: Over estimate (usually grossly over), preallocation count for off-diagonal portion of scalar matrix
7991: */
7992: for (Ii = 0, jj = 0; Ii < Iend - Istart; Ii += bs, jj++) {
7993: o_nnz[jj] = 0;
7994: for (kk = 0; kk < bs; kk++) { /* rows that get collapsed to a single row */
7995: PetscCall(MatGetRow(Oaij, Ii + kk, &ncols, NULL, NULL));
7996: o_nnz[jj] += ncols;
7997: PetscCall(MatRestoreRow(Oaij, Ii + kk, &ncols, NULL, NULL));
7998: }
7999: if (o_nnz[jj] > (NN / bs - nloc)) o_nnz[jj] = NN / bs - nloc;
8000: }
8001: } else SETERRQ(comm, PETSC_ERR_USER, "Require AIJ matrix type");
8002: /* get scalar copy (norms) of matrix */
8003: PetscCall(MatSeqAIJSetPreallocation(Gmat, 0, d_nnz));
8004: PetscCall(MatMPIAIJSetPreallocation(Gmat, 0, d_nnz, 0, o_nnz));
8005: PetscCall(PetscFree2(d_nnz, o_nnz));
8006: for (Ii = Istart; Ii < Iend; Ii++) {
8007: PetscInt dest_row = Ii / bs;
8009: PetscCall(MatGetRow(Amat, Ii, &ncols, &idx, &vals));
8010: for (jj = 0; jj < ncols; jj++) {
8011: PetscInt dest_col = idx[jj] / bs;
8012: PetscScalar sv = PetscAbs(PetscRealPart(vals[jj]));
8014: PetscCall(MatSetValues(Gmat, 1, &dest_row, 1, &dest_col, &sv, ADD_VALUES));
8015: }
8016: PetscCall(MatRestoreRow(Amat, Ii, &ncols, &idx, &vals));
8017: }
8018: PetscCall(MatAssemblyBegin(Gmat, MAT_FINAL_ASSEMBLY));
8019: PetscCall(MatAssemblyEnd(Gmat, MAT_FINAL_ASSEMBLY));
8020: }
8021: } else {
8022: if (symmetrize || filter >= 0 || scale) PetscCall(MatDuplicate(Amat, MAT_COPY_VALUES, &Gmat));
8023: else {
8024: Gmat = Amat;
8025: PetscCall(PetscObjectReference((PetscObject)Gmat));
8026: }
8027: if (isseqaij) {
8028: a = Gmat;
8029: b = NULL;
8030: } else {
8031: Mat_MPIAIJ *d = (Mat_MPIAIJ *)Gmat->data;
8032: a = d->A;
8033: b = d->B;
8034: }
8035: if (filter >= 0 || scale) {
8036: /* take absolute value of each entry */
8037: for (c = a, kk = 0; c && kk < 2; c = b, kk++) {
8038: MatInfo info;
8039: PetscScalar *avals;
8041: PetscCall(MatGetInfo(c, MAT_LOCAL, &info));
8042: PetscCall(MatSeqAIJGetArray(c, &avals));
8043: for (int jj = 0; jj < info.nz_used; jj++) avals[jj] = PetscAbsScalar(avals[jj]);
8044: PetscCall(MatSeqAIJRestoreArray(c, &avals));
8045: }
8046: }
8047: }
8048: if (symmetrize) {
8049: PetscBool isset, issym;
8051: PetscCall(MatIsSymmetricKnown(Amat, &isset, &issym));
8052: if (!isset || !issym) {
8053: Mat matTrans;
8055: PetscCall(MatTranspose(Gmat, MAT_INITIAL_MATRIX, &matTrans));
8056: PetscCall(MatAXPY(Gmat, 1.0, matTrans, Gmat->structurally_symmetric == PETSC_BOOL3_TRUE ? SAME_NONZERO_PATTERN : DIFFERENT_NONZERO_PATTERN));
8057: PetscCall(MatDestroy(&matTrans));
8058: }
8059: PetscCall(MatSetOption(Gmat, MAT_SYMMETRIC, PETSC_TRUE));
8060: } else if (Amat != Gmat) PetscCall(MatPropagateSymmetryOptions(Amat, Gmat));
8061: if (scale) {
8062: /* scale c for all diagonal values = 1 or -1 */
8063: Vec diag;
8065: PetscCall(MatCreateVecs(Gmat, &diag, NULL));
8066: PetscCall(MatGetDiagonal(Gmat, diag));
8067: PetscCall(VecReciprocal(diag));
8068: PetscCall(VecSqrtAbs(diag));
8069: PetscCall(MatDiagonalScale(Gmat, diag, diag));
8070: PetscCall(VecDestroy(&diag));
8071: }
8072: PetscCall(MatViewFromOptions(Gmat, NULL, "-mat_graph_view"));
8073: if (filter >= 0) {
8074: PetscCall(MatFilter(Gmat, filter, PETSC_TRUE, PETSC_TRUE));
8075: PetscCall(MatViewFromOptions(Gmat, NULL, "-mat_filter_graph_view"));
8076: }
8077: *a_Gmat = Gmat;
8078: PetscFunctionReturn(PETSC_SUCCESS);
8079: }
8081: PETSC_INTERN PetscErrorCode MatGetCurrentMemType_MPIAIJ(Mat A, PetscMemType *memtype)
8082: {
8083: Mat_MPIAIJ *mpiaij = (Mat_MPIAIJ *)A->data;
8084: PetscMemType mD = PETSC_MEMTYPE_HOST, mO = PETSC_MEMTYPE_HOST;
8086: PetscFunctionBegin;
8087: if (mpiaij->A) PetscCall(MatGetCurrentMemType(mpiaij->A, &mD));
8088: if (mpiaij->B) PetscCall(MatGetCurrentMemType(mpiaij->B, &mO));
8089: *memtype = (mD == mO) ? mD : PETSC_MEMTYPE_HOST;
8090: PetscFunctionReturn(PETSC_SUCCESS);
8091: }
8093: /*
8094: Special version for direct calls from Fortran
8095: */
8097: /* Change these macros so can be used in void function */
8098: /* Identical to PetscCallVoid, except it assigns to *_ierr */
8099: #undef PetscCall
8100: #define PetscCall(...) \
8101: do { \
8102: PetscErrorCode ierr_msv_mpiaij = __VA_ARGS__; \
8103: if (PetscUnlikely(ierr_msv_mpiaij)) { \
8104: *_ierr = PetscError(PETSC_COMM_SELF, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr_msv_mpiaij, PETSC_ERROR_REPEAT, " "); \
8105: return; \
8106: } \
8107: } while (0)
8109: #undef SETERRQ
8110: #define SETERRQ(comm, ierr, ...) \
8111: do { \
8112: *_ierr = PetscError(comm, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr, PETSC_ERROR_INITIAL, __VA_ARGS__); \
8113: return; \
8114: } while (0)
8116: #if PetscDefined(HAVE_FORTRAN_CAPS)
8117: #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ
8118: #elif !PetscDefined(HAVE_FORTRAN_UNDERSCORE)
8119: #define matsetvaluesmpiaij_ matsetvaluesmpiaij
8120: #else
8121: #endif
8122: PETSC_EXTERN void matsetvaluesmpiaij_(Mat *mmat, PetscInt *mm, const PetscInt im[], PetscInt *mn, const PetscInt in[], const PetscScalar v[], InsertMode *maddv, PetscErrorCode *_ierr)
8123: {
8124: Mat mat = *mmat;
8125: PetscInt m = *mm, n = *mn;
8126: InsertMode addv = *maddv;
8127: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
8128: PetscScalar value;
8130: MatCheckPreallocated(mat, 1);
8131: if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv;
8132: else PetscCheck(mat->insertmode == addv, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Cannot mix add values and insert values");
8133: {
8134: PetscInt i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend;
8135: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend, row, col;
8136: PetscBool roworiented = aij->roworiented;
8138: /* Some Variables required in the macro */
8139: Mat A = aij->A;
8140: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
8141: PetscInt *aimax = a->imax, *ai = a->i, *ailen = a->ilen, *aj = a->j;
8142: MatScalar *aa;
8143: PetscBool ignorezeroentries = ((a->ignorezeroentries && (addv == ADD_VALUES)) ? PETSC_TRUE : PETSC_FALSE);
8144: Mat B = aij->B;
8145: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
8146: PetscInt *bimax = b->imax, *bi = b->i, *bilen = b->ilen, *bj = b->j, bm = aij->B->rmap->n, am = aij->A->rmap->n;
8147: MatScalar *ba;
8148: /* This variable below is only for the PETSC_HAVE_VIENNACL or PETSC_HAVE_CUDA cases, but we define it in all cases because we
8149: * cannot use "#if defined" inside a macro. */
8150: PETSC_UNUSED PetscBool inserted = PETSC_FALSE;
8152: PetscInt *rp1, *rp2, ii, nrow1, nrow2, _i, rmax1, rmax2, N, low1, high1, low2, high2, t, lastcol1, lastcol2;
8153: PetscInt nonew = a->nonew;
8154: MatScalar *ap1, *ap2;
8156: PetscFunctionBegin;
8157: PetscCall(MatSeqAIJGetArray(A, &aa));
8158: PetscCall(MatSeqAIJGetArray(B, &ba));
8159: for (i = 0; i < m; i++) {
8160: if (im[i] < 0) continue;
8161: 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);
8162: if (im[i] >= rstart && im[i] < rend) {
8163: row = im[i] - rstart;
8164: lastcol1 = -1;
8165: rp1 = aj + ai[row];
8166: ap1 = aa + ai[row];
8167: rmax1 = aimax[row];
8168: nrow1 = ailen[row];
8169: low1 = 0;
8170: high1 = nrow1;
8171: lastcol2 = -1;
8172: rp2 = bj + bi[row];
8173: ap2 = ba + bi[row];
8174: rmax2 = bimax[row];
8175: nrow2 = bilen[row];
8176: low2 = 0;
8177: high2 = nrow2;
8179: for (j = 0; j < n; j++) {
8180: if (roworiented) value = v[i * n + j];
8181: else value = v[i + j * m];
8182: if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && im[i] != in[j]) continue;
8183: if (in[j] >= cstart && in[j] < cend) {
8184: col = in[j] - cstart;
8185: MatSetValues_SeqAIJ_A_Private(row, col, value, addv, im[i], in[j]);
8186: } else if (in[j] < 0) continue;
8187: else if (PetscUnlikelyDebug(in[j] >= mat->cmap->N)) {
8188: SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large: col %" PetscInt_FMT " max %" PetscInt_FMT, in[j], mat->cmap->N - 1);
8189: } else {
8190: if (mat->was_assembled) {
8191: if (!aij->colmap) PetscCall(MatCreateColmap_MPIAIJ_Private(mat));
8192: #if PetscDefined(USE_CTABLE)
8193: PetscCall(PetscHMapIGetWithDefault(aij->colmap, in[j] + 1, 0, &col));
8194: col--;
8195: #else
8196: col = aij->colmap[in[j]] - 1;
8197: #endif
8198: if (col < 0 && !((Mat_SeqAIJ *)aij->A->data)->nonew) {
8199: PetscCall(MatDisAssemble_MPIAIJ(mat, PETSC_FALSE));
8200: col = in[j];
8201: /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
8202: B = aij->B;
8203: b = (Mat_SeqAIJ *)B->data;
8204: bimax = b->imax;
8205: bi = b->i;
8206: bilen = b->ilen;
8207: bj = b->j;
8208: rp2 = bj + bi[row];
8209: ap2 = ba + bi[row];
8210: rmax2 = bimax[row];
8211: nrow2 = bilen[row];
8212: low2 = 0;
8213: high2 = nrow2;
8214: bm = aij->B->rmap->n;
8215: ba = b->a;
8216: inserted = PETSC_FALSE;
8217: }
8218: } else col = in[j];
8219: MatSetValues_SeqAIJ_B_Private(row, col, value, addv, im[i], in[j]);
8220: }
8221: }
8222: } else if (!aij->donotstash) {
8223: if (roworiented) {
8224: PetscCall(MatStashValuesRow_Private(&mat->stash, im[i], n, in, v + i * n, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
8225: } else {
8226: PetscCall(MatStashValuesCol_Private(&mat->stash, im[i], n, in, v + i, m, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
8227: }
8228: }
8229: }
8230: PetscCall(MatSeqAIJRestoreArray(A, &aa));
8231: PetscCall(MatSeqAIJRestoreArray(B, &ba));
8232: }
8233: PetscFunctionReturnVoid();
8234: }
8236: /* Undefining these here since they were redefined from their original definition above! No
8237: * other PETSc functions should be defined past this point, as it is impossible to recover the
8238: * original definitions */
8239: #undef PetscCall
8240: #undef SETERRQ