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 defined(PETSC_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, "MatIsTranspose_C", NULL));
67: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMPIAIJSetPreallocation_C", NULL));
68: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatResetPreallocation_C", NULL));
69: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatResetHash_C", NULL));
70: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMPIAIJSetPreallocationCSR_C", NULL));
71: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDiagonalScaleLocal_C", NULL));
72: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpibaij_C", NULL));
73: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpisbaij_C", NULL));
74: #if defined(PETSC_HAVE_CUDA)
75: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijcusparse_C", NULL));
76: #endif
77: #if defined(PETSC_HAVE_HIP)
78: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijhipsparse_C", NULL));
79: #endif
80: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
81: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijkokkos_C", NULL));
82: #endif
83: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpidense_C", NULL));
84: #if defined(PETSC_HAVE_ELEMENTAL)
85: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_elemental_C", NULL));
86: #endif
87: #if defined(PETSC_HAVE_SCALAPACK) && (defined(PETSC_USE_REAL_SINGLE) || defined(PETSC_USE_REAL_DOUBLE))
88: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_scalapack_C", NULL));
89: #endif
90: #if defined(PETSC_HAVE_HYPRE)
91: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_hypre_C", NULL));
92: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_transpose_mpiaij_mpiaij_C", NULL));
93: #endif
94: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_is_C", NULL));
95: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_is_mpiaij_C", NULL));
96: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpiaij_mpiaij_C", NULL));
97: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMPIAIJSetUseScalableIncreaseOverlap_C", NULL));
98: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijperm_C", NULL));
99: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijsell_C", NULL));
100: #if defined(PETSC_HAVE_MKL_SPARSE)
101: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijmkl_C", NULL));
102: #endif
103: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijcrl_C", NULL));
104: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_is_C", NULL));
105: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpisell_C", NULL));
106: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatSetPreallocationCOO_C", NULL));
107: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatSetValuesCOO_C", NULL));
108: PetscFunctionReturn(PETSC_SUCCESS);
109: }
111: static PetscErrorCode MatGetRowIJ_MPIAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *m, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done)
112: {
113: Mat B;
115: PetscFunctionBegin;
116: PetscCall(MatMPIAIJGetLocalMat(A, MAT_INITIAL_MATRIX, &B));
117: PetscCall(PetscObjectCompose((PetscObject)A, "MatGetRowIJ_MPIAIJ", (PetscObject)B));
118: PetscCall(MatGetRowIJ(B, oshift, symmetric, inodecompressed, m, ia, ja, done));
119: PetscCall(MatDestroy(&B));
120: PetscFunctionReturn(PETSC_SUCCESS);
121: }
123: static PetscErrorCode MatRestoreRowIJ_MPIAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *m, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done)
124: {
125: Mat B;
127: PetscFunctionBegin;
128: PetscCall(PetscObjectQuery((PetscObject)A, "MatGetRowIJ_MPIAIJ", (PetscObject *)&B));
129: PetscCall(MatRestoreRowIJ(B, oshift, symmetric, inodecompressed, m, ia, ja, done));
130: PetscCall(PetscObjectCompose((PetscObject)A, "MatGetRowIJ_MPIAIJ", NULL));
131: PetscFunctionReturn(PETSC_SUCCESS);
132: }
134: /*MC
135: MATAIJ - MATAIJ = "aij" - A matrix type to be used for sparse matrices.
137: This matrix type is identical to` MATSEQAIJ` when constructed with a single process communicator,
138: and `MATMPIAIJ` otherwise. As a result, for single process communicators,
139: `MatSeqAIJSetPreallocation()` is supported, and similarly `MatMPIAIJSetPreallocation()` is supported
140: for communicators controlling multiple processes. It is recommended that you call both of
141: the above preallocation routines for simplicity.
143: Options Database Key:
144: . -mat_type aij - sets the matrix type to `MATAIJ` during a call to `MatSetFromOptions()`
146: Developer Note:
147: Level: beginner
149: Subclasses include `MATAIJCUSPARSE`, `MATAIJPERM`, `MATAIJSELL`, `MATAIJMKL`, `MATAIJCRL`, `MATAIJKOKKOS`,and also automatically switches over to use inodes when
150: enough exist.
152: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MATSEQAIJ`, `MatCreateAIJ()`, `MatCreateSeqAIJ()`, `MATSEQAIJ`, `MATMPIAIJ`
153: M*/
155: /*MC
156: MATAIJCRL - MATAIJCRL = "aijcrl" - A matrix type to be used for sparse matrices.
158: This matrix type is identical to `MATSEQAIJCRL` when constructed with a single process communicator,
159: and `MATMPIAIJCRL` otherwise. As a result, for single process communicators,
160: `MatSeqAIJSetPreallocation()` is supported, and similarly `MatMPIAIJSetPreallocation()` is supported
161: for communicators controlling multiple processes. It is recommended that you call both of
162: the above preallocation routines for simplicity.
164: Options Database Key:
165: . -mat_type aijcrl - sets the matrix type to `MATMPIAIJCRL` during a call to `MatSetFromOptions()`
167: Level: beginner
169: .seealso: [](ch_matrices), `Mat`, `MatCreateMPIAIJCRL`, `MATSEQAIJCRL`, `MATMPIAIJCRL`, `MATSEQAIJCRL`, `MATMPIAIJCRL`
170: M*/
172: static PetscErrorCode MatBindToCPU_MPIAIJ(Mat A, PetscBool flg)
173: {
174: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
176: PetscFunctionBegin;
177: #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP) || defined(PETSC_HAVE_VIENNACL)
178: A->boundtocpu = flg;
179: #endif
180: if (a->A) PetscCall(MatBindToCPU(a->A, flg));
181: if (a->B) PetscCall(MatBindToCPU(a->B, flg));
183: /* In addition to binding the diagonal and off-diagonal matrices, bind the local vectors used for matrix-vector products.
184: * This maybe seems a little odd for a MatBindToCPU() call to do, but it makes no sense for the binding of these vectors
185: * to differ from the parent matrix. */
186: if (a->lvec) PetscCall(VecBindToCPU(a->lvec, flg));
187: if (a->diag) PetscCall(VecBindToCPU(a->diag, flg));
188: PetscFunctionReturn(PETSC_SUCCESS);
189: }
191: static PetscErrorCode MatSetBlockSizes_MPIAIJ(Mat M, PetscInt rbs, PetscInt cbs)
192: {
193: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)M->data;
195: PetscFunctionBegin;
196: if (mat->A) {
197: PetscCall(MatSetBlockSizes(mat->A, rbs, cbs));
198: PetscCall(MatSetBlockSizes(mat->B, rbs, 1));
199: }
200: PetscFunctionReturn(PETSC_SUCCESS);
201: }
203: static PetscErrorCode MatFindNonzeroRows_MPIAIJ(Mat M, IS *keptrows)
204: {
205: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)M->data;
206: Mat_SeqAIJ *a = (Mat_SeqAIJ *)mat->A->data;
207: Mat_SeqAIJ *b = (Mat_SeqAIJ *)mat->B->data;
208: const PetscInt *ia, *ib;
209: const MatScalar *aa, *bb, *aav, *bav;
210: PetscInt na, nb, i, j, *rows, cnt = 0, n0rows;
211: PetscInt m = M->rmap->n, rstart = M->rmap->rstart;
213: PetscFunctionBegin;
214: *keptrows = NULL;
216: ia = a->i;
217: ib = b->i;
218: PetscCall(MatSeqAIJGetArrayRead(mat->A, &aav));
219: PetscCall(MatSeqAIJGetArrayRead(mat->B, &bav));
220: for (i = 0; i < m; i++) {
221: na = ia[i + 1] - ia[i];
222: nb = ib[i + 1] - ib[i];
223: if (!na && !nb) {
224: cnt++;
225: goto ok1;
226: }
227: aa = aav + ia[i];
228: for (j = 0; j < na; j++) {
229: if (aa[j] != 0.0) goto ok1;
230: }
231: bb = PetscSafePointerPlusOffset(bav, ib[i]);
232: for (j = 0; j < nb; j++) {
233: if (bb[j] != 0.0) goto ok1;
234: }
235: cnt++;
236: ok1:;
237: }
238: PetscCallMPI(MPIU_Allreduce(&cnt, &n0rows, 1, MPIU_INT, MPI_SUM, PetscObjectComm((PetscObject)M)));
239: if (!n0rows) {
240: PetscCall(MatSeqAIJRestoreArrayRead(mat->A, &aav));
241: PetscCall(MatSeqAIJRestoreArrayRead(mat->B, &bav));
242: PetscFunctionReturn(PETSC_SUCCESS);
243: }
244: PetscCall(PetscMalloc1(M->rmap->n - cnt, &rows));
245: cnt = 0;
246: for (i = 0; i < m; i++) {
247: na = ia[i + 1] - ia[i];
248: nb = ib[i + 1] - ib[i];
249: if (!na && !nb) continue;
250: aa = aav + ia[i];
251: for (j = 0; j < na; j++) {
252: if (aa[j] != 0.0) {
253: rows[cnt++] = rstart + i;
254: goto ok2;
255: }
256: }
257: bb = PetscSafePointerPlusOffset(bav, ib[i]);
258: for (j = 0; j < nb; j++) {
259: if (bb[j] != 0.0) {
260: rows[cnt++] = rstart + i;
261: goto ok2;
262: }
263: }
264: ok2:;
265: }
266: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)M), cnt, rows, PETSC_OWN_POINTER, keptrows));
267: PetscCall(MatSeqAIJRestoreArrayRead(mat->A, &aav));
268: PetscCall(MatSeqAIJRestoreArrayRead(mat->B, &bav));
269: PetscFunctionReturn(PETSC_SUCCESS);
270: }
272: static PetscErrorCode MatDiagonalSet_MPIAIJ(Mat Y, Vec D, InsertMode is)
273: {
274: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)Y->data;
275: PetscBool cong;
277: PetscFunctionBegin;
278: PetscCall(MatHasCongruentLayouts(Y, &cong));
279: if (Y->assembled && cong) {
280: PetscCall(MatDiagonalSet(aij->A, D, is));
281: } else {
282: PetscCall(MatDiagonalSet_Default(Y, D, is));
283: }
284: PetscFunctionReturn(PETSC_SUCCESS);
285: }
287: static PetscErrorCode MatFindZeroDiagonals_MPIAIJ(Mat M, IS *zrows)
288: {
289: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)M->data;
290: PetscInt i, rstart, nrows, *rows;
292: PetscFunctionBegin;
293: *zrows = NULL;
294: PetscCall(MatFindZeroDiagonals_SeqAIJ_Private(aij->A, &nrows, &rows));
295: PetscCall(MatGetOwnershipRange(M, &rstart, NULL));
296: for (i = 0; i < nrows; i++) rows[i] += rstart;
297: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)M), nrows, rows, PETSC_OWN_POINTER, zrows));
298: PetscFunctionReturn(PETSC_SUCCESS);
299: }
301: static PetscErrorCode MatGetColumnReductions_MPIAIJ(Mat A, PetscInt type, PetscReal *reductions)
302: {
303: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)A->data;
304: PetscInt i, m, n, *garray = aij->garray;
305: Mat_SeqAIJ *a_aij = (Mat_SeqAIJ *)aij->A->data;
306: Mat_SeqAIJ *b_aij = (Mat_SeqAIJ *)aij->B->data;
307: PetscReal *work;
308: const PetscScalar *dummy;
310: PetscFunctionBegin;
311: PetscCall(MatGetSize(A, &m, &n));
312: PetscCall(PetscCalloc1(n, &work));
313: PetscCall(MatSeqAIJGetArrayRead(aij->A, &dummy));
314: PetscCall(MatSeqAIJRestoreArrayRead(aij->A, &dummy));
315: PetscCall(MatSeqAIJGetArrayRead(aij->B, &dummy));
316: PetscCall(MatSeqAIJRestoreArrayRead(aij->B, &dummy));
317: if (type == NORM_2) {
318: for (i = 0; i < a_aij->i[aij->A->rmap->n]; i++) work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i] * a_aij->a[i]);
319: for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i] * b_aij->a[i]);
320: } else if (type == NORM_1) {
321: for (i = 0; i < a_aij->i[aij->A->rmap->n]; i++) work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]);
322: for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]);
323: } else if (type == NORM_INFINITY) {
324: for (i = 0; i < a_aij->i[aij->A->rmap->n]; i++) work[A->cmap->rstart + a_aij->j[i]] = PetscMax(PetscAbsScalar(a_aij->a[i]), work[A->cmap->rstart + a_aij->j[i]]);
325: for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) work[garray[b_aij->j[i]]] = PetscMax(PetscAbsScalar(b_aij->a[i]), work[garray[b_aij->j[i]]]);
326: } else if (type == REDUCTION_SUM_REALPART || type == REDUCTION_MEAN_REALPART) {
327: for (i = 0; i < a_aij->i[aij->A->rmap->n]; i++) work[A->cmap->rstart + a_aij->j[i]] += PetscRealPart(a_aij->a[i]);
328: for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) work[garray[b_aij->j[i]]] += PetscRealPart(b_aij->a[i]);
329: } else if (type == REDUCTION_SUM_IMAGINARYPART || type == REDUCTION_MEAN_IMAGINARYPART) {
330: for (i = 0; i < a_aij->i[aij->A->rmap->n]; i++) work[A->cmap->rstart + a_aij->j[i]] += PetscImaginaryPart(a_aij->a[i]);
331: for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) work[garray[b_aij->j[i]]] += PetscImaginaryPart(b_aij->a[i]);
332: } else SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Unknown reduction type");
333: if (type == NORM_INFINITY) {
334: PetscCallMPI(MPIU_Allreduce(work, reductions, n, MPIU_REAL, MPIU_MAX, PetscObjectComm((PetscObject)A)));
335: } else {
336: PetscCallMPI(MPIU_Allreduce(work, reductions, n, MPIU_REAL, MPIU_SUM, PetscObjectComm((PetscObject)A)));
337: }
338: PetscCall(PetscFree(work));
339: if (type == NORM_2) {
340: for (i = 0; i < n; i++) reductions[i] = PetscSqrtReal(reductions[i]);
341: } else if (type == REDUCTION_MEAN_REALPART || type == REDUCTION_MEAN_IMAGINARYPART) {
342: for (i = 0; i < n; i++) reductions[i] /= m;
343: }
344: PetscFunctionReturn(PETSC_SUCCESS);
345: }
347: static PetscErrorCode MatFindOffBlockDiagonalEntries_MPIAIJ(Mat A, IS *is)
348: {
349: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
350: IS sis, gis;
351: const PetscInt *isis, *igis;
352: PetscInt n, *iis, nsis, ngis, rstart, i;
354: PetscFunctionBegin;
355: PetscCall(MatFindOffBlockDiagonalEntries(a->A, &sis));
356: PetscCall(MatFindNonzeroRows(a->B, &gis));
357: PetscCall(ISGetSize(gis, &ngis));
358: PetscCall(ISGetSize(sis, &nsis));
359: PetscCall(ISGetIndices(sis, &isis));
360: PetscCall(ISGetIndices(gis, &igis));
362: PetscCall(PetscMalloc1(ngis + nsis, &iis));
363: PetscCall(PetscArraycpy(iis, igis, ngis));
364: PetscCall(PetscArraycpy(iis + ngis, isis, nsis));
365: n = ngis + nsis;
366: PetscCall(PetscSortRemoveDupsInt(&n, iis));
367: PetscCall(MatGetOwnershipRange(A, &rstart, NULL));
368: for (i = 0; i < n; i++) iis[i] += rstart;
369: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)A), n, iis, PETSC_OWN_POINTER, is));
371: PetscCall(ISRestoreIndices(sis, &isis));
372: PetscCall(ISRestoreIndices(gis, &igis));
373: PetscCall(ISDestroy(&sis));
374: PetscCall(ISDestroy(&gis));
375: PetscFunctionReturn(PETSC_SUCCESS);
376: }
378: /*
379: Local utility routine that creates a mapping from the global column
380: number to the local number in the off-diagonal part of the local
381: storage of the matrix. When PETSC_USE_CTABLE is used this is scalable at
382: a slightly higher hash table cost; without it it is not scalable (each processor
383: has an order N integer array but is fast to access.
384: */
385: PetscErrorCode MatCreateColmap_MPIAIJ_Private(Mat mat)
386: {
387: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
388: PetscInt n = aij->B->cmap->n, i;
390: PetscFunctionBegin;
391: PetscCheck(!n || aij->garray, PETSC_COMM_SELF, PETSC_ERR_PLIB, "MPIAIJ Matrix was assembled but is missing garray");
392: #if defined(PETSC_USE_CTABLE)
393: PetscCall(PetscHMapICreateWithSize(n, &aij->colmap));
394: for (i = 0; i < n; i++) PetscCall(PetscHMapISet(aij->colmap, aij->garray[i] + 1, i + 1));
395: #else
396: PetscCall(PetscCalloc1(mat->cmap->N + 1, &aij->colmap));
397: for (i = 0; i < n; i++) aij->colmap[aij->garray[i]] = i + 1;
398: #endif
399: PetscFunctionReturn(PETSC_SUCCESS);
400: }
402: #define MatSetValues_SeqAIJ_A_Private(row, col, value, addv, orow, ocol) \
403: do { \
404: if (col <= lastcol1) low1 = 0; \
405: else high1 = nrow1; \
406: lastcol1 = col; \
407: while (high1 - low1 > 5) { \
408: t = (low1 + high1) / 2; \
409: if (rp1[t] > col) high1 = t; \
410: else low1 = t; \
411: } \
412: for (_i = low1; _i < high1; _i++) { \
413: if (rp1[_i] > col) break; \
414: if (rp1[_i] == col) { \
415: if (addv == ADD_VALUES) { \
416: ap1[_i] += value; \
417: /* Not sure LogFlops will slow down the code or not */ \
418: (void)PetscLogFlops(1.0); \
419: } else ap1[_i] = value; \
420: goto a_noinsert; \
421: } \
422: } \
423: if (value == 0.0 && ignorezeroentries && row != col) { \
424: low1 = 0; \
425: high1 = nrow1; \
426: goto a_noinsert; \
427: } \
428: if (nonew == 1) { \
429: low1 = 0; \
430: high1 = nrow1; \
431: goto a_noinsert; \
432: } \
433: 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); \
434: MatSeqXAIJReallocateAIJ(A, am, 1, nrow1, row, col, rmax1, aa, ai, aj, rp1, ap1, aimax, nonew, MatScalar); \
435: N = nrow1++ - 1; \
436: a->nz++; \
437: high1++; \
438: /* shift up all the later entries in this row */ \
439: PetscCall(PetscArraymove(rp1 + _i + 1, rp1 + _i, N - _i + 1)); \
440: PetscCall(PetscArraymove(ap1 + _i + 1, ap1 + _i, N - _i + 1)); \
441: rp1[_i] = col; \
442: ap1[_i] = value; \
443: a_noinsert:; \
444: ailen[row] = nrow1; \
445: } while (0)
447: #define MatSetValues_SeqAIJ_B_Private(row, col, value, addv, orow, ocol) \
448: do { \
449: if (col <= lastcol2) low2 = 0; \
450: else high2 = nrow2; \
451: lastcol2 = col; \
452: while (high2 - low2 > 5) { \
453: t = (low2 + high2) / 2; \
454: if (rp2[t] > col) high2 = t; \
455: else low2 = t; \
456: } \
457: for (_i = low2; _i < high2; _i++) { \
458: if (rp2[_i] > col) break; \
459: if (rp2[_i] == col) { \
460: if (addv == ADD_VALUES) { \
461: ap2[_i] += value; \
462: (void)PetscLogFlops(1.0); \
463: } else ap2[_i] = value; \
464: goto b_noinsert; \
465: } \
466: } \
467: if (value == 0.0 && ignorezeroentries) { \
468: low2 = 0; \
469: high2 = nrow2; \
470: goto b_noinsert; \
471: } \
472: if (nonew == 1) { \
473: low2 = 0; \
474: high2 = nrow2; \
475: goto b_noinsert; \
476: } \
477: 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); \
478: MatSeqXAIJReallocateAIJ(B, bm, 1, nrow2, row, col, rmax2, ba, bi, bj, rp2, ap2, bimax, nonew, MatScalar); \
479: N = nrow2++ - 1; \
480: b->nz++; \
481: high2++; \
482: /* shift up all the later entries in this row */ \
483: PetscCall(PetscArraymove(rp2 + _i + 1, rp2 + _i, N - _i + 1)); \
484: PetscCall(PetscArraymove(ap2 + _i + 1, ap2 + _i, N - _i + 1)); \
485: rp2[_i] = col; \
486: ap2[_i] = value; \
487: b_noinsert:; \
488: bilen[row] = nrow2; \
489: } while (0)
491: static PetscErrorCode MatSetValuesRow_MPIAIJ(Mat A, PetscInt row, const PetscScalar v[])
492: {
493: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
494: Mat_SeqAIJ *a = (Mat_SeqAIJ *)mat->A->data, *b = (Mat_SeqAIJ *)mat->B->data;
495: PetscInt l, *garray = mat->garray, diag;
496: PetscScalar *aa, *ba;
498: PetscFunctionBegin;
499: /* code only works for square matrices A */
501: /* find size of row to the left of the diagonal part */
502: PetscCall(MatGetOwnershipRange(A, &diag, NULL));
503: row = row - diag;
504: for (l = 0; l < b->i[row + 1] - b->i[row]; l++) {
505: if (garray[b->j[b->i[row] + l]] > diag) break;
506: }
507: if (l) {
508: PetscCall(MatSeqAIJGetArray(mat->B, &ba));
509: PetscCall(PetscArraycpy(ba + b->i[row], v, l));
510: PetscCall(MatSeqAIJRestoreArray(mat->B, &ba));
511: }
513: /* diagonal part */
514: if (a->i[row + 1] - a->i[row]) {
515: PetscCall(MatSeqAIJGetArray(mat->A, &aa));
516: PetscCall(PetscArraycpy(aa + a->i[row], v + l, a->i[row + 1] - a->i[row]));
517: PetscCall(MatSeqAIJRestoreArray(mat->A, &aa));
518: }
520: /* right of diagonal part */
521: if (b->i[row + 1] - b->i[row] - l) {
522: PetscCall(MatSeqAIJGetArray(mat->B, &ba));
523: PetscCall(PetscArraycpy(ba + b->i[row] + l, v + l + a->i[row + 1] - a->i[row], b->i[row + 1] - b->i[row] - l));
524: PetscCall(MatSeqAIJRestoreArray(mat->B, &ba));
525: }
526: PetscFunctionReturn(PETSC_SUCCESS);
527: }
529: PetscErrorCode MatSetValues_MPIAIJ(Mat mat, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode addv)
530: {
531: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
532: PetscScalar value = 0.0;
533: PetscInt i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend;
534: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend, row, col;
535: PetscBool roworiented = aij->roworiented;
537: /* Some Variables required in the macro */
538: Mat A = aij->A;
539: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
540: PetscInt *aimax = a->imax, *ai = a->i, *ailen = a->ilen, *aj = a->j;
541: PetscBool ignorezeroentries = a->ignorezeroentries;
542: Mat B = aij->B;
543: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
544: PetscInt *bimax = b->imax, *bi = b->i, *bilen = b->ilen, *bj = b->j, bm = aij->B->rmap->n, am = aij->A->rmap->n;
545: MatScalar *aa, *ba;
546: PetscInt *rp1, *rp2, ii, nrow1, nrow2, _i, rmax1, rmax2, N, low1, high1, low2, high2, t, lastcol1, lastcol2;
547: PetscInt nonew;
548: MatScalar *ap1, *ap2;
550: PetscFunctionBegin;
551: PetscCall(MatSeqAIJGetArray(A, &aa));
552: PetscCall(MatSeqAIJGetArray(B, &ba));
553: for (i = 0; i < m; i++) {
554: if (im[i] < 0) continue;
555: 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);
556: if (im[i] >= rstart && im[i] < rend) {
557: row = im[i] - rstart;
558: lastcol1 = -1;
559: rp1 = PetscSafePointerPlusOffset(aj, ai[row]);
560: ap1 = PetscSafePointerPlusOffset(aa, ai[row]);
561: rmax1 = aimax[row];
562: nrow1 = ailen[row];
563: low1 = 0;
564: high1 = nrow1;
565: lastcol2 = -1;
566: rp2 = PetscSafePointerPlusOffset(bj, bi[row]);
567: ap2 = PetscSafePointerPlusOffset(ba, bi[row]);
568: rmax2 = bimax[row];
569: nrow2 = bilen[row];
570: low2 = 0;
571: high2 = nrow2;
573: for (j = 0; j < n; j++) {
574: if (v) value = roworiented ? v[i * n + j] : v[i + j * m];
575: if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && im[i] != in[j]) continue;
576: if (in[j] >= cstart && in[j] < cend) {
577: col = in[j] - cstart;
578: nonew = a->nonew;
579: MatSetValues_SeqAIJ_A_Private(row, col, value, addv, im[i], in[j]);
580: } else if (in[j] < 0) {
581: continue;
582: } else {
583: 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);
584: if (mat->was_assembled) {
585: if (!aij->colmap) PetscCall(MatCreateColmap_MPIAIJ_Private(mat));
586: #if defined(PETSC_USE_CTABLE)
587: PetscCall(PetscHMapIGetWithDefault(aij->colmap, in[j] + 1, 0, &col)); /* map global col ids to local ones */
588: col--;
589: #else
590: col = aij->colmap[in[j]] - 1;
591: #endif
592: if (col < 0 && !((Mat_SeqAIJ *)aij->B->data)->nonew) { /* col < 0 means in[j] is a new col for B */
593: PetscCall(MatDisAssemble_MPIAIJ(mat, PETSC_FALSE)); /* Change aij->B from reduced/local format to expanded/global format */
594: col = in[j];
595: /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
596: B = aij->B;
597: b = (Mat_SeqAIJ *)B->data;
598: bimax = b->imax;
599: bi = b->i;
600: bilen = b->ilen;
601: bj = b->j;
602: ba = b->a;
603: rp2 = PetscSafePointerPlusOffset(bj, bi[row]);
604: ap2 = PetscSafePointerPlusOffset(ba, bi[row]);
605: rmax2 = bimax[row];
606: nrow2 = bilen[row];
607: low2 = 0;
608: high2 = nrow2;
609: bm = aij->B->rmap->n;
610: ba = b->a;
611: } else if (col < 0 && !(ignorezeroentries && value == 0.0)) {
612: 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]);
613: 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]));
614: }
615: } else col = in[j];
616: nonew = b->nonew;
617: MatSetValues_SeqAIJ_B_Private(row, col, value, addv, im[i], in[j]);
618: }
619: }
620: } else {
621: 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]);
622: if (!aij->donotstash) {
623: mat->assembled = PETSC_FALSE;
624: if (roworiented) {
625: PetscCall(MatStashValuesRow_Private(&mat->stash, im[i], n, in, PetscSafePointerPlusOffset(v, i * n), (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
626: } else {
627: PetscCall(MatStashValuesCol_Private(&mat->stash, im[i], n, in, PetscSafePointerPlusOffset(v, i), m, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
628: }
629: }
630: }
631: }
632: PetscCall(MatSeqAIJRestoreArray(A, &aa)); /* aa, bb might have been free'd due to reallocation above. But we don't access them here */
633: PetscCall(MatSeqAIJRestoreArray(B, &ba));
634: PetscFunctionReturn(PETSC_SUCCESS);
635: }
637: /*
638: This function sets the j and ilen arrays (of the diagonal and off-diagonal part) of an MPIAIJ-matrix.
639: The values in mat_i have to be sorted and the values in mat_j have to be sorted for each row (CSR-like).
640: No off-processor parts off the matrix are allowed here and mat->was_assembled has to be PETSC_FALSE.
641: */
642: PetscErrorCode MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(Mat mat, const PetscInt mat_j[], const PetscInt mat_i[])
643: {
644: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
645: Mat A = aij->A; /* diagonal part of the matrix */
646: Mat B = aij->B; /* off-diagonal part of the matrix */
647: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
648: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
649: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend, col;
650: PetscInt *ailen = a->ilen, *aj = a->j;
651: PetscInt *bilen = b->ilen, *bj = b->j;
652: PetscInt am = aij->A->rmap->n, j;
653: PetscInt diag_so_far = 0, dnz;
654: PetscInt offd_so_far = 0, onz;
656: PetscFunctionBegin;
657: /* Iterate over all rows of the matrix */
658: for (j = 0; j < am; j++) {
659: dnz = onz = 0;
660: /* Iterate over all non-zero columns of the current row */
661: for (col = mat_i[j]; col < mat_i[j + 1]; col++) {
662: /* If column is in the diagonal */
663: if (mat_j[col] >= cstart && mat_j[col] < cend) {
664: aj[diag_so_far++] = mat_j[col] - cstart;
665: dnz++;
666: } else { /* off-diagonal entries */
667: bj[offd_so_far++] = mat_j[col];
668: onz++;
669: }
670: }
671: ailen[j] = dnz;
672: bilen[j] = onz;
673: }
674: PetscFunctionReturn(PETSC_SUCCESS);
675: }
677: /*
678: This function sets the local j, a and ilen arrays (of the diagonal and off-diagonal part) of an MPIAIJ-matrix.
679: The values in mat_i have to be sorted and the values in mat_j have to be sorted for each row (CSR-like).
680: No off-processor parts off the matrix are allowed here, they are set at a later point by MatSetValues_MPIAIJ.
681: Also, mat->was_assembled has to be false, otherwise the statement aj[rowstart_diag+dnz_row] = mat_j[col] - cstart;
682: would not be true and the more complex MatSetValues_MPIAIJ has to be used.
683: */
684: PetscErrorCode MatSetValues_MPIAIJ_CopyFromCSRFormat(Mat mat, const PetscInt mat_j[], const PetscInt mat_i[], const PetscScalar mat_a[])
685: {
686: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
687: Mat A = aij->A; /* diagonal part of the matrix */
688: Mat B = aij->B; /* off-diagonal part of the matrix */
689: Mat_SeqAIJ *aijd = (Mat_SeqAIJ *)aij->A->data, *aijo = (Mat_SeqAIJ *)aij->B->data;
690: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
691: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
692: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend;
693: PetscInt *ailen = a->ilen, *aj = a->j;
694: PetscInt *bilen = b->ilen, *bj = b->j;
695: PetscInt am = aij->A->rmap->n, j;
696: 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. */
697: PetscInt col, dnz_row, onz_row, rowstart_diag, rowstart_offd;
698: PetscScalar *aa = a->a, *ba = b->a;
700: PetscFunctionBegin;
701: /* Iterate over all rows of the matrix */
702: for (j = 0; j < am; j++) {
703: dnz_row = onz_row = 0;
704: rowstart_offd = full_offd_i[j];
705: rowstart_diag = full_diag_i[j];
706: /* Iterate over all non-zero columns of the current row */
707: for (col = mat_i[j]; col < mat_i[j + 1]; col++) {
708: /* If column is in the diagonal */
709: if (mat_j[col] >= cstart && mat_j[col] < cend) {
710: aj[rowstart_diag + dnz_row] = mat_j[col] - cstart;
711: aa[rowstart_diag + dnz_row] = mat_a[col];
712: dnz_row++;
713: } else { /* off-diagonal entries */
714: bj[rowstart_offd + onz_row] = mat_j[col];
715: ba[rowstart_offd + onz_row] = mat_a[col];
716: onz_row++;
717: }
718: }
719: ailen[j] = dnz_row;
720: bilen[j] = onz_row;
721: }
722: PetscFunctionReturn(PETSC_SUCCESS);
723: }
725: static PetscErrorCode MatGetValues_MPIAIJ(Mat mat, PetscInt m, const PetscInt idxm[], PetscInt n, const PetscInt idxn[], PetscScalar v[])
726: {
727: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
728: PetscInt i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend;
729: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend, row, col;
731: PetscFunctionBegin;
732: for (i = 0; i < m; i++) {
733: if (idxm[i] < 0) continue; /* negative row */
734: 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);
735: 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);
736: row = idxm[i] - rstart;
737: for (j = 0; j < n; j++) {
738: if (idxn[j] < 0) continue; /* negative column */
739: 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);
740: if (idxn[j] >= cstart && idxn[j] < cend) {
741: col = idxn[j] - cstart;
742: PetscCall(MatGetValues(aij->A, 1, &row, 1, &col, v + i * n + j));
743: } else {
744: if (!aij->colmap) PetscCall(MatCreateColmap_MPIAIJ_Private(mat));
745: #if defined(PETSC_USE_CTABLE)
746: PetscCall(PetscHMapIGetWithDefault(aij->colmap, idxn[j] + 1, 0, &col));
747: col--;
748: #else
749: col = aij->colmap[idxn[j]] - 1;
750: #endif
751: if ((col < 0) || (aij->garray[col] != idxn[j])) *(v + i * n + j) = 0.0;
752: else PetscCall(MatGetValues(aij->B, 1, &row, 1, &col, v + i * n + j));
753: }
754: }
755: }
756: PetscFunctionReturn(PETSC_SUCCESS);
757: }
759: static PetscErrorCode MatAssemblyBegin_MPIAIJ(Mat mat, MatAssemblyType mode)
760: {
761: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
762: PetscInt nstash, reallocs;
764: PetscFunctionBegin;
765: if (aij->donotstash || mat->nooffprocentries) PetscFunctionReturn(PETSC_SUCCESS);
767: PetscCall(MatStashScatterBegin_Private(mat, &mat->stash, mat->rmap->range));
768: PetscCall(MatStashGetInfo_Private(&mat->stash, &nstash, &reallocs));
769: PetscCall(PetscInfo(mat, "Stash has %" PetscInt_FMT " entries, uses %" PetscInt_FMT " mallocs.\n", nstash, reallocs));
770: PetscFunctionReturn(PETSC_SUCCESS);
771: }
773: PetscErrorCode MatAssemblyEnd_MPIAIJ(Mat mat, MatAssemblyType mode)
774: {
775: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
776: PetscMPIInt n;
777: PetscInt i, j, rstart, ncols, flg;
778: PetscInt *row, *col;
779: PetscBool all_assembled;
780: PetscScalar *val;
782: /* do not use 'b = (Mat_SeqAIJ*)aij->B->data' as B can be reset in disassembly */
784: PetscFunctionBegin;
785: if (!aij->donotstash && !mat->nooffprocentries) {
786: while (1) {
787: PetscCall(MatStashScatterGetMesg_Private(&mat->stash, &n, &row, &col, &val, &flg));
788: if (!flg) break;
790: for (i = 0; i < n;) {
791: /* Now identify the consecutive vals belonging to the same row */
792: for (j = i, rstart = row[j]; j < n; j++) {
793: if (row[j] != rstart) break;
794: }
795: if (j < n) ncols = j - i;
796: else ncols = n - i;
797: /* Now assemble all these values with a single function call */
798: PetscCall(MatSetValues_MPIAIJ(mat, 1, row + i, ncols, col + i, val + i, mat->insertmode));
799: i = j;
800: }
801: }
802: PetscCall(MatStashScatterEnd_Private(&mat->stash));
803: }
804: #if defined(PETSC_HAVE_DEVICE)
805: if (mat->offloadmask == PETSC_OFFLOAD_CPU) aij->A->offloadmask = PETSC_OFFLOAD_CPU;
806: /* We call MatBindToCPU() on aij->A and aij->B here, because if MatBindToCPU_MPIAIJ() is called before assembly, it cannot bind these. */
807: if (mat->boundtocpu) {
808: PetscCall(MatBindToCPU(aij->A, PETSC_TRUE));
809: PetscCall(MatBindToCPU(aij->B, PETSC_TRUE));
810: }
811: #endif
812: PetscCall(MatAssemblyBegin(aij->A, mode));
813: PetscCall(MatAssemblyEnd(aij->A, mode));
815: /* determine if any process has disassembled, if so we must
816: also disassemble ourself, in order that we may reassemble. */
817: /*
818: if nonzero structure of submatrix B cannot change then we know that
819: no process disassembled thus we can skip this stuff
820: */
821: if (!((Mat_SeqAIJ *)aij->B->data)->nonew) {
822: PetscCallMPI(MPIU_Allreduce(&mat->was_assembled, &all_assembled, 1, MPI_C_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
823: if (mat->was_assembled && !all_assembled) { /* mat on this rank has reduced off-diag B with local col ids, but globally it does not */
824: PetscCall(MatDisAssemble_MPIAIJ(mat, PETSC_FALSE));
825: }
826: }
827: if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) PetscCall(MatSetUpMultiply_MPIAIJ(mat));
828: PetscCall(MatSetOption(aij->B, MAT_USE_INODES, PETSC_FALSE));
829: #if defined(PETSC_HAVE_DEVICE)
830: if (mat->offloadmask == PETSC_OFFLOAD_CPU && aij->B->offloadmask != PETSC_OFFLOAD_UNALLOCATED) aij->B->offloadmask = PETSC_OFFLOAD_CPU;
831: #endif
832: PetscCall(MatAssemblyBegin(aij->B, mode));
833: PetscCall(MatAssemblyEnd(aij->B, mode));
835: PetscCall(PetscFree2(aij->rowvalues, aij->rowindices));
837: aij->rowvalues = NULL;
839: PetscCall(VecDestroy(&aij->diag));
841: /* if no new nonzero locations are allowed in matrix then only set the matrix state the first time through */
842: if ((!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) || !((Mat_SeqAIJ *)aij->A->data)->nonew) {
843: PetscObjectState state = aij->A->nonzerostate + aij->B->nonzerostate;
844: PetscCallMPI(MPIU_Allreduce(&state, &mat->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)mat)));
845: }
846: #if defined(PETSC_HAVE_DEVICE)
847: mat->offloadmask = PETSC_OFFLOAD_BOTH;
848: #endif
849: PetscFunctionReturn(PETSC_SUCCESS);
850: }
852: static PetscErrorCode MatZeroEntries_MPIAIJ(Mat A)
853: {
854: Mat_MPIAIJ *l = (Mat_MPIAIJ *)A->data;
856: PetscFunctionBegin;
857: PetscCall(MatZeroEntries(l->A));
858: PetscCall(MatZeroEntries(l->B));
859: PetscFunctionReturn(PETSC_SUCCESS);
860: }
862: static PetscErrorCode MatZeroRows_MPIAIJ(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diag, Vec x, Vec b)
863: {
864: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
865: PetscInt *lrows;
866: PetscInt r, len;
867: PetscBool cong;
869: PetscFunctionBegin;
870: /* get locally owned rows */
871: PetscCall(MatZeroRowsMapLocal_Private(A, N, rows, &len, &lrows));
872: PetscCall(MatHasCongruentLayouts(A, &cong));
873: /* fix right-hand side if needed */
874: if (x && b) {
875: const PetscScalar *xx;
876: PetscScalar *bb;
878: PetscCheck(cong, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Need matching row/col layout");
879: PetscCall(VecGetArrayRead(x, &xx));
880: PetscCall(VecGetArray(b, &bb));
881: for (r = 0; r < len; ++r) bb[lrows[r]] = diag * xx[lrows[r]];
882: PetscCall(VecRestoreArrayRead(x, &xx));
883: PetscCall(VecRestoreArray(b, &bb));
884: }
886: if (diag != 0.0 && cong) {
887: PetscCall(MatZeroRows(mat->A, len, lrows, diag, NULL, NULL));
888: PetscCall(MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL));
889: } else if (diag != 0.0) { /* non-square or non congruent layouts -> if keepnonzeropattern is false, we allow for new insertion */
890: Mat_SeqAIJ *aijA = (Mat_SeqAIJ *)mat->A->data;
891: Mat_SeqAIJ *aijB = (Mat_SeqAIJ *)mat->B->data;
892: PetscInt nnwA, nnwB;
893: PetscBool nnzA, nnzB;
895: nnwA = aijA->nonew;
896: nnwB = aijB->nonew;
897: nnzA = aijA->keepnonzeropattern;
898: nnzB = aijB->keepnonzeropattern;
899: if (!nnzA) {
900: PetscCall(PetscInfo(mat->A, "Requested to not keep the pattern and add a nonzero diagonal; may encounter reallocations on diagonal block.\n"));
901: aijA->nonew = 0;
902: }
903: if (!nnzB) {
904: PetscCall(PetscInfo(mat->B, "Requested to not keep the pattern and add a nonzero diagonal; may encounter reallocations on off-diagonal block.\n"));
905: aijB->nonew = 0;
906: }
907: /* Must zero here before the next loop */
908: PetscCall(MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL));
909: PetscCall(MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL));
910: for (r = 0; r < len; ++r) {
911: const PetscInt row = lrows[r] + A->rmap->rstart;
912: if (row >= A->cmap->N) continue;
913: PetscCall(MatSetValues(A, 1, &row, 1, &row, &diag, INSERT_VALUES));
914: }
915: aijA->nonew = nnwA;
916: aijB->nonew = nnwB;
917: } else {
918: PetscCall(MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL));
919: PetscCall(MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL));
920: }
921: PetscCall(PetscFree(lrows));
922: PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY));
923: PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY));
925: /* only change matrix nonzero state if pattern was allowed to be changed */
926: if (!((Mat_SeqAIJ *)mat->A->data)->keepnonzeropattern || !((Mat_SeqAIJ *)mat->A->data)->nonew) {
927: PetscObjectState state = mat->A->nonzerostate + mat->B->nonzerostate;
928: PetscCallMPI(MPIU_Allreduce(&state, &A->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)A)));
929: }
930: PetscFunctionReturn(PETSC_SUCCESS);
931: }
933: static PetscErrorCode MatZeroRowsColumns_MPIAIJ(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diag, Vec x, Vec b)
934: {
935: Mat_MPIAIJ *l = (Mat_MPIAIJ *)A->data;
936: PetscInt n = A->rmap->n;
937: PetscInt i, j, r, m, len = 0;
938: PetscInt *lrows, *owners = A->rmap->range;
939: PetscMPIInt p = 0;
940: PetscSFNode *rrows;
941: PetscSF sf;
942: const PetscScalar *xx;
943: PetscScalar *bb, *mask, *aij_a;
944: Vec xmask, lmask;
945: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)l->B->data;
946: const PetscInt *aj, *ii, *ridx;
947: PetscScalar *aa;
949: PetscFunctionBegin;
950: /* Create SF where leaves are input rows and roots are owned rows */
951: PetscCall(PetscMalloc1(n, &lrows));
952: for (r = 0; r < n; ++r) lrows[r] = -1;
953: PetscCall(PetscMalloc1(N, &rrows));
954: for (r = 0; r < N; ++r) {
955: const PetscInt idx = rows[r];
956: 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);
957: if (idx < owners[p] || owners[p + 1] <= idx) { /* short-circuit the search if the last p owns this row too */
958: PetscCall(PetscLayoutFindOwner(A->rmap, idx, &p));
959: }
960: rrows[r].rank = p;
961: rrows[r].index = rows[r] - owners[p];
962: }
963: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
964: PetscCall(PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER));
965: /* Collect flags for rows to be zeroed */
966: PetscCall(PetscSFReduceBegin(sf, MPIU_INT, (PetscInt *)rows, lrows, MPI_LOR));
967: PetscCall(PetscSFReduceEnd(sf, MPIU_INT, (PetscInt *)rows, lrows, MPI_LOR));
968: PetscCall(PetscSFDestroy(&sf));
969: /* Compress and put in row numbers */
970: for (r = 0; r < n; ++r)
971: if (lrows[r] >= 0) lrows[len++] = r;
972: /* zero diagonal part of matrix */
973: PetscCall(MatZeroRowsColumns(l->A, len, lrows, diag, x, b));
974: /* handle off-diagonal part of matrix */
975: PetscCall(MatCreateVecs(A, &xmask, NULL));
976: PetscCall(VecDuplicate(l->lvec, &lmask));
977: PetscCall(VecGetArray(xmask, &bb));
978: for (i = 0; i < len; i++) bb[lrows[i]] = 1;
979: PetscCall(VecRestoreArray(xmask, &bb));
980: PetscCall(VecScatterBegin(l->Mvctx, xmask, lmask, ADD_VALUES, SCATTER_FORWARD));
981: PetscCall(VecScatterEnd(l->Mvctx, xmask, lmask, ADD_VALUES, SCATTER_FORWARD));
982: PetscCall(VecDestroy(&xmask));
983: if (x && b) { /* this code is buggy when the row and column layout don't match */
984: PetscBool cong;
986: PetscCall(MatHasCongruentLayouts(A, &cong));
987: PetscCheck(cong, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Need matching row/col layout");
988: PetscCall(VecScatterBegin(l->Mvctx, x, l->lvec, INSERT_VALUES, SCATTER_FORWARD));
989: PetscCall(VecScatterEnd(l->Mvctx, x, l->lvec, INSERT_VALUES, SCATTER_FORWARD));
990: PetscCall(VecGetArrayRead(l->lvec, &xx));
991: PetscCall(VecGetArray(b, &bb));
992: }
993: PetscCall(VecGetArray(lmask, &mask));
994: /* remove zeroed rows of off-diagonal matrix */
995: PetscCall(MatSeqAIJGetArray(l->B, &aij_a));
996: ii = aij->i;
997: for (i = 0; i < len; i++) PetscCall(PetscArrayzero(PetscSafePointerPlusOffset(aij_a, ii[lrows[i]]), ii[lrows[i] + 1] - ii[lrows[i]]));
998: /* loop over all elements of off process part of matrix zeroing removed columns*/
999: if (aij->compressedrow.use) {
1000: m = aij->compressedrow.nrows;
1001: ii = aij->compressedrow.i;
1002: ridx = aij->compressedrow.rindex;
1003: for (i = 0; i < m; i++) {
1004: n = ii[i + 1] - ii[i];
1005: aj = aij->j + ii[i];
1006: aa = aij_a + ii[i];
1008: for (j = 0; j < n; j++) {
1009: if (PetscAbsScalar(mask[*aj])) {
1010: if (b) bb[*ridx] -= *aa * xx[*aj];
1011: *aa = 0.0;
1012: }
1013: aa++;
1014: aj++;
1015: }
1016: ridx++;
1017: }
1018: } else { /* do not use compressed row format */
1019: m = l->B->rmap->n;
1020: for (i = 0; i < m; i++) {
1021: n = ii[i + 1] - ii[i];
1022: aj = aij->j + ii[i];
1023: aa = aij_a + ii[i];
1024: for (j = 0; j < n; j++) {
1025: if (PetscAbsScalar(mask[*aj])) {
1026: if (b) bb[i] -= *aa * xx[*aj];
1027: *aa = 0.0;
1028: }
1029: aa++;
1030: aj++;
1031: }
1032: }
1033: }
1034: if (x && b) {
1035: PetscCall(VecRestoreArray(b, &bb));
1036: PetscCall(VecRestoreArrayRead(l->lvec, &xx));
1037: }
1038: PetscCall(MatSeqAIJRestoreArray(l->B, &aij_a));
1039: PetscCall(VecRestoreArray(lmask, &mask));
1040: PetscCall(VecDestroy(&lmask));
1041: PetscCall(PetscFree(lrows));
1043: /* only change matrix nonzero state if pattern was allowed to be changed */
1044: if (!((Mat_SeqAIJ *)l->A->data)->nonew) {
1045: PetscObjectState state = l->A->nonzerostate + l->B->nonzerostate;
1046: PetscCallMPI(MPIU_Allreduce(&state, &A->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)A)));
1047: }
1048: PetscFunctionReturn(PETSC_SUCCESS);
1049: }
1051: static PetscErrorCode MatMult_MPIAIJ(Mat A, Vec xx, Vec yy)
1052: {
1053: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1054: PetscInt nt;
1055: VecScatter Mvctx = a->Mvctx;
1057: PetscFunctionBegin;
1058: PetscCall(VecGetLocalSize(xx, &nt));
1059: 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);
1060: PetscCall(VecScatterBegin(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1061: PetscUseTypeMethod(a->A, mult, xx, yy);
1062: PetscCall(VecScatterEnd(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1063: PetscUseTypeMethod(a->B, multadd, a->lvec, yy, yy);
1064: PetscFunctionReturn(PETSC_SUCCESS);
1065: }
1067: static PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat A, Vec bb, Vec xx)
1068: {
1069: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1071: PetscFunctionBegin;
1072: PetscCall(MatMultDiagonalBlock(a->A, bb, xx));
1073: PetscFunctionReturn(PETSC_SUCCESS);
1074: }
1076: static PetscErrorCode MatMultAdd_MPIAIJ(Mat A, Vec xx, Vec yy, Vec zz)
1077: {
1078: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1079: VecScatter Mvctx = a->Mvctx;
1081: PetscFunctionBegin;
1082: PetscCall(VecScatterBegin(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1083: PetscCall((*a->A->ops->multadd)(a->A, xx, yy, zz));
1084: PetscCall(VecScatterEnd(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1085: PetscCall((*a->B->ops->multadd)(a->B, a->lvec, zz, zz));
1086: PetscFunctionReturn(PETSC_SUCCESS);
1087: }
1089: static PetscErrorCode MatMultTranspose_MPIAIJ(Mat A, Vec xx, Vec yy)
1090: {
1091: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1093: PetscFunctionBegin;
1094: /* do nondiagonal part */
1095: PetscCall((*a->B->ops->multtranspose)(a->B, xx, a->lvec));
1096: /* do local part */
1097: PetscCall((*a->A->ops->multtranspose)(a->A, xx, yy));
1098: /* add partial results together */
1099: PetscCall(VecScatterBegin(a->Mvctx, a->lvec, yy, ADD_VALUES, SCATTER_REVERSE));
1100: PetscCall(VecScatterEnd(a->Mvctx, a->lvec, yy, ADD_VALUES, SCATTER_REVERSE));
1101: PetscFunctionReturn(PETSC_SUCCESS);
1102: }
1104: static PetscErrorCode MatIsTranspose_MPIAIJ(Mat Amat, Mat Bmat, PetscReal tol, PetscBool *f)
1105: {
1106: MPI_Comm comm;
1107: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)Amat->data, *Bij = (Mat_MPIAIJ *)Bmat->data;
1108: Mat Adia = Aij->A, Bdia = Bij->A, Aoff, Boff, *Aoffs, *Boffs;
1109: IS Me, Notme;
1110: PetscInt M, N, first, last, *notme, i;
1111: PetscBool lf;
1112: PetscMPIInt size;
1114: PetscFunctionBegin;
1115: /* Easy test: symmetric diagonal block */
1116: PetscCall(MatIsTranspose(Adia, Bdia, tol, &lf));
1117: PetscCallMPI(MPIU_Allreduce(&lf, f, 1, MPI_C_BOOL, MPI_LAND, PetscObjectComm((PetscObject)Amat)));
1118: if (!*f) PetscFunctionReturn(PETSC_SUCCESS);
1119: PetscCall(PetscObjectGetComm((PetscObject)Amat, &comm));
1120: PetscCallMPI(MPI_Comm_size(comm, &size));
1121: if (size == 1) PetscFunctionReturn(PETSC_SUCCESS);
1123: /* Hard test: off-diagonal block. This takes a MatCreateSubMatrix. */
1124: PetscCall(MatGetSize(Amat, &M, &N));
1125: PetscCall(MatGetOwnershipRange(Amat, &first, &last));
1126: PetscCall(PetscMalloc1(N - last + first, ¬me));
1127: for (i = 0; i < first; i++) notme[i] = i;
1128: for (i = last; i < M; i++) notme[i - last + first] = i;
1129: PetscCall(ISCreateGeneral(MPI_COMM_SELF, N - last + first, notme, PETSC_COPY_VALUES, &Notme));
1130: PetscCall(ISCreateStride(MPI_COMM_SELF, last - first, first, 1, &Me));
1131: PetscCall(MatCreateSubMatrices(Amat, 1, &Me, &Notme, MAT_INITIAL_MATRIX, &Aoffs));
1132: Aoff = Aoffs[0];
1133: PetscCall(MatCreateSubMatrices(Bmat, 1, &Notme, &Me, MAT_INITIAL_MATRIX, &Boffs));
1134: Boff = Boffs[0];
1135: PetscCall(MatIsTranspose(Aoff, Boff, tol, f));
1136: PetscCall(MatDestroyMatrices(1, &Aoffs));
1137: PetscCall(MatDestroyMatrices(1, &Boffs));
1138: PetscCall(ISDestroy(&Me));
1139: PetscCall(ISDestroy(&Notme));
1140: PetscCall(PetscFree(notme));
1141: PetscFunctionReturn(PETSC_SUCCESS);
1142: }
1144: static PetscErrorCode MatMultTransposeAdd_MPIAIJ(Mat A, Vec xx, Vec yy, Vec zz)
1145: {
1146: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1148: PetscFunctionBegin;
1149: /* do nondiagonal part */
1150: PetscCall((*a->B->ops->multtranspose)(a->B, xx, a->lvec));
1151: /* do local part */
1152: PetscCall((*a->A->ops->multtransposeadd)(a->A, xx, yy, zz));
1153: /* add partial results together */
1154: PetscCall(VecScatterBegin(a->Mvctx, a->lvec, zz, ADD_VALUES, SCATTER_REVERSE));
1155: PetscCall(VecScatterEnd(a->Mvctx, a->lvec, zz, ADD_VALUES, SCATTER_REVERSE));
1156: PetscFunctionReturn(PETSC_SUCCESS);
1157: }
1159: /*
1160: This only works correctly for square matrices where the subblock A->A is the
1161: diagonal block
1162: */
1163: static PetscErrorCode MatGetDiagonal_MPIAIJ(Mat A, Vec v)
1164: {
1165: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1167: PetscFunctionBegin;
1168: PetscCheck(A->rmap->N == A->cmap->N, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Supports only square matrix where A->A is diag block");
1169: 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");
1170: PetscCall(MatGetDiagonal(a->A, v));
1171: PetscFunctionReturn(PETSC_SUCCESS);
1172: }
1174: static PetscErrorCode MatScale_MPIAIJ(Mat A, PetscScalar aa)
1175: {
1176: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1178: PetscFunctionBegin;
1179: PetscCall(MatScale(a->A, aa));
1180: PetscCall(MatScale(a->B, aa));
1181: PetscFunctionReturn(PETSC_SUCCESS);
1182: }
1184: static PetscErrorCode MatView_MPIAIJ_Binary(Mat mat, PetscViewer viewer)
1185: {
1186: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1187: Mat_SeqAIJ *A = (Mat_SeqAIJ *)aij->A->data;
1188: Mat_SeqAIJ *B = (Mat_SeqAIJ *)aij->B->data;
1189: const PetscInt *garray = aij->garray;
1190: const PetscScalar *aa, *ba;
1191: PetscInt header[4], M, N, m, rs, cs, cnt, i, ja, jb;
1192: PetscInt64 nz, hnz;
1193: PetscInt *rowlens;
1194: PetscInt *colidxs;
1195: PetscScalar *matvals;
1196: PetscMPIInt rank;
1198: PetscFunctionBegin;
1199: PetscCall(PetscViewerSetUp(viewer));
1201: M = mat->rmap->N;
1202: N = mat->cmap->N;
1203: m = mat->rmap->n;
1204: rs = mat->rmap->rstart;
1205: cs = mat->cmap->rstart;
1206: nz = A->nz + B->nz;
1208: /* write matrix header */
1209: header[0] = MAT_FILE_CLASSID;
1210: header[1] = M;
1211: header[2] = N;
1212: PetscCallMPI(MPI_Reduce(&nz, &hnz, 1, MPIU_INT64, MPI_SUM, 0, PetscObjectComm((PetscObject)mat)));
1213: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)mat), &rank));
1214: if (rank == 0) PetscCall(PetscIntCast(hnz, &header[3]));
1215: PetscCall(PetscViewerBinaryWrite(viewer, header, 4, PETSC_INT));
1217: /* fill in and store row lengths */
1218: PetscCall(PetscMalloc1(m, &rowlens));
1219: for (i = 0; i < m; i++) rowlens[i] = A->i[i + 1] - A->i[i] + B->i[i + 1] - B->i[i];
1220: PetscCall(PetscViewerBinaryWriteAll(viewer, rowlens, m, rs, M, PETSC_INT));
1221: PetscCall(PetscFree(rowlens));
1223: /* fill in and store column indices */
1224: PetscCall(PetscMalloc1(nz, &colidxs));
1225: for (cnt = 0, i = 0; i < m; i++) {
1226: for (jb = B->i[i]; jb < B->i[i + 1]; jb++) {
1227: if (garray[B->j[jb]] > cs) break;
1228: colidxs[cnt++] = garray[B->j[jb]];
1229: }
1230: for (ja = A->i[i]; ja < A->i[i + 1]; ja++) colidxs[cnt++] = A->j[ja] + cs;
1231: for (; jb < B->i[i + 1]; jb++) colidxs[cnt++] = garray[B->j[jb]];
1232: }
1233: PetscCheck(cnt == nz, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Internal PETSc error: cnt = %" PetscInt_FMT " nz = %" PetscInt64_FMT, cnt, nz);
1234: PetscCall(PetscViewerBinaryWriteAll(viewer, colidxs, nz, PETSC_DETERMINE, PETSC_DETERMINE, PETSC_INT));
1235: PetscCall(PetscFree(colidxs));
1237: /* fill in and store nonzero values */
1238: PetscCall(MatSeqAIJGetArrayRead(aij->A, &aa));
1239: PetscCall(MatSeqAIJGetArrayRead(aij->B, &ba));
1240: PetscCall(PetscMalloc1(nz, &matvals));
1241: for (cnt = 0, i = 0; i < m; i++) {
1242: for (jb = B->i[i]; jb < B->i[i + 1]; jb++) {
1243: if (garray[B->j[jb]] > cs) break;
1244: matvals[cnt++] = ba[jb];
1245: }
1246: for (ja = A->i[i]; ja < A->i[i + 1]; ja++) matvals[cnt++] = aa[ja];
1247: for (; jb < B->i[i + 1]; jb++) matvals[cnt++] = ba[jb];
1248: }
1249: PetscCall(MatSeqAIJRestoreArrayRead(aij->A, &aa));
1250: PetscCall(MatSeqAIJRestoreArrayRead(aij->B, &ba));
1251: PetscCheck(cnt == nz, PETSC_COMM_SELF, PETSC_ERR_LIB, "Internal PETSc error: cnt = %" PetscInt_FMT " nz = %" PetscInt64_FMT, cnt, nz);
1252: PetscCall(PetscViewerBinaryWriteAll(viewer, matvals, nz, PETSC_DETERMINE, PETSC_DETERMINE, PETSC_SCALAR));
1253: PetscCall(PetscFree(matvals));
1255: /* write block size option to the viewer's .info file */
1256: PetscCall(MatView_Binary_BlockSizes(mat, viewer));
1257: PetscFunctionReturn(PETSC_SUCCESS);
1258: }
1260: #include <petscdraw.h>
1261: static PetscErrorCode MatView_MPIAIJ_ASCIIorDraworSocket(Mat mat, PetscViewer viewer)
1262: {
1263: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1264: PetscMPIInt rank = aij->rank, size = aij->size;
1265: PetscBool isdraw, isascii, isbinary;
1266: PetscViewer sviewer;
1267: PetscViewerFormat format;
1269: PetscFunctionBegin;
1270: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw));
1271: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii));
1272: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
1273: if (isascii) {
1274: PetscCall(PetscViewerGetFormat(viewer, &format));
1275: if (format == PETSC_VIEWER_LOAD_BALANCE) {
1276: PetscInt i, nmax = 0, nmin = PETSC_INT_MAX, navg = 0, *nz, nzlocal = ((Mat_SeqAIJ *)aij->A->data)->nz + ((Mat_SeqAIJ *)aij->B->data)->nz;
1277: PetscCall(PetscMalloc1(size, &nz));
1278: PetscCallMPI(MPI_Allgather(&nzlocal, 1, MPIU_INT, nz, 1, MPIU_INT, PetscObjectComm((PetscObject)mat)));
1279: for (i = 0; i < size; i++) {
1280: nmax = PetscMax(nmax, nz[i]);
1281: nmin = PetscMin(nmin, nz[i]);
1282: navg += nz[i];
1283: }
1284: PetscCall(PetscFree(nz));
1285: navg = navg / size;
1286: PetscCall(PetscViewerASCIIPrintf(viewer, "Load Balance - Nonzeros: Min %" PetscInt_FMT " avg %" PetscInt_FMT " max %" PetscInt_FMT "\n", nmin, navg, nmax));
1287: PetscFunctionReturn(PETSC_SUCCESS);
1288: }
1289: PetscCall(PetscViewerGetFormat(viewer, &format));
1290: if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
1291: MatInfo info;
1292: PetscInt *inodes = NULL;
1294: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)mat), &rank));
1295: PetscCall(MatGetInfo(mat, MAT_LOCAL, &info));
1296: PetscCall(MatInodeGetInodeSizes(aij->A, NULL, &inodes, NULL));
1297: PetscCall(PetscViewerASCIIPushSynchronized(viewer));
1298: if (!inodes) {
1299: 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,
1300: info.memory));
1301: } else {
1302: PetscCall(
1303: 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));
1304: }
1305: PetscCall(MatGetInfo(aij->A, MAT_LOCAL, &info));
1306: PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] on-diagonal part: nz %" PetscInt_FMT " \n", rank, (PetscInt)info.nz_used));
1307: PetscCall(MatGetInfo(aij->B, MAT_LOCAL, &info));
1308: PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] off-diagonal part: nz %" PetscInt_FMT " \n", rank, (PetscInt)info.nz_used));
1309: PetscCall(PetscViewerFlush(viewer));
1310: PetscCall(PetscViewerASCIIPopSynchronized(viewer));
1311: PetscCall(PetscViewerASCIIPrintf(viewer, "Information on VecScatter used in matrix-vector product: \n"));
1312: PetscCall(VecScatterView(aij->Mvctx, viewer));
1313: PetscFunctionReturn(PETSC_SUCCESS);
1314: } else if (format == PETSC_VIEWER_ASCII_INFO) {
1315: PetscInt inodecount, inodelimit, *inodes;
1316: PetscCall(MatInodeGetInodeSizes(aij->A, &inodecount, &inodes, &inodelimit));
1317: if (inodes) {
1318: PetscCall(PetscViewerASCIIPrintf(viewer, "using I-node (on process 0) routines: found %" PetscInt_FMT " nodes, limit used is %" PetscInt_FMT "\n", inodecount, inodelimit));
1319: } else {
1320: PetscCall(PetscViewerASCIIPrintf(viewer, "not using I-node (on process 0) routines\n"));
1321: }
1322: PetscFunctionReturn(PETSC_SUCCESS);
1323: } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) {
1324: PetscFunctionReturn(PETSC_SUCCESS);
1325: }
1326: } else if (isbinary) {
1327: if (size == 1) {
1328: PetscCall(PetscObjectSetName((PetscObject)aij->A, ((PetscObject)mat)->name));
1329: PetscCall(MatView(aij->A, viewer));
1330: } else {
1331: PetscCall(MatView_MPIAIJ_Binary(mat, viewer));
1332: }
1333: PetscFunctionReturn(PETSC_SUCCESS);
1334: } else if (isascii && size == 1) {
1335: PetscCall(PetscObjectSetName((PetscObject)aij->A, ((PetscObject)mat)->name));
1336: PetscCall(MatView(aij->A, viewer));
1337: PetscFunctionReturn(PETSC_SUCCESS);
1338: } else if (isdraw) {
1339: PetscDraw draw;
1340: PetscBool isnull;
1341: PetscCall(PetscViewerDrawGetDraw(viewer, 0, &draw));
1342: PetscCall(PetscDrawIsNull(draw, &isnull));
1343: if (isnull) PetscFunctionReturn(PETSC_SUCCESS);
1344: }
1346: { /* assemble the entire matrix onto first processor */
1347: Mat A = NULL, Av;
1348: IS isrow, iscol;
1350: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat), rank == 0 ? mat->rmap->N : 0, 0, 1, &isrow));
1351: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat), rank == 0 ? mat->cmap->N : 0, 0, 1, &iscol));
1352: PetscCall(MatCreateSubMatrix(mat, isrow, iscol, MAT_INITIAL_MATRIX, &A));
1353: PetscCall(MatMPIAIJGetSeqAIJ(A, &Av, NULL, NULL));
1354: /* The commented code uses MatCreateSubMatrices instead */
1355: /*
1356: Mat *AA, A = NULL, Av;
1357: IS isrow,iscol;
1359: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat),rank == 0 ? mat->rmap->N : 0,0,1,&isrow));
1360: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat),rank == 0 ? mat->cmap->N : 0,0,1,&iscol));
1361: PetscCall(MatCreateSubMatrices(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,&AA));
1362: if (rank == 0) {
1363: PetscCall(PetscObjectReference((PetscObject)AA[0]));
1364: A = AA[0];
1365: Av = AA[0];
1366: }
1367: PetscCall(MatDestroySubMatrices(1,&AA));
1368: */
1369: PetscCall(ISDestroy(&iscol));
1370: PetscCall(ISDestroy(&isrow));
1371: /*
1372: Everyone has to call to draw the matrix since the graphics waits are
1373: synchronized across all processors that share the PetscDraw object
1374: */
1375: PetscCall(PetscViewerGetSubViewer(viewer, PETSC_COMM_SELF, &sviewer));
1376: if (rank == 0) {
1377: if (((PetscObject)mat)->name) PetscCall(PetscObjectSetName((PetscObject)Av, ((PetscObject)mat)->name));
1378: PetscCall(MatView_SeqAIJ(Av, sviewer));
1379: }
1380: PetscCall(PetscViewerRestoreSubViewer(viewer, PETSC_COMM_SELF, &sviewer));
1381: PetscCall(MatDestroy(&A));
1382: }
1383: PetscFunctionReturn(PETSC_SUCCESS);
1384: }
1386: PetscErrorCode MatView_MPIAIJ(Mat mat, PetscViewer viewer)
1387: {
1388: PetscBool isascii, isdraw, issocket, isbinary;
1390: PetscFunctionBegin;
1391: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii));
1392: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw));
1393: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
1394: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERSOCKET, &issocket));
1395: if (isascii || isdraw || isbinary || issocket) PetscCall(MatView_MPIAIJ_ASCIIorDraworSocket(mat, viewer));
1396: PetscFunctionReturn(PETSC_SUCCESS);
1397: }
1399: static PetscErrorCode MatSOR_MPIAIJ(Mat matin, Vec bb, PetscReal omega, MatSORType flag, PetscReal fshift, PetscInt its, PetscInt lits, Vec xx)
1400: {
1401: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)matin->data;
1402: Vec bb1 = NULL;
1403: PetscBool hasop;
1405: PetscFunctionBegin;
1406: if (flag == SOR_APPLY_UPPER) {
1407: PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
1408: PetscFunctionReturn(PETSC_SUCCESS);
1409: }
1411: if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS || flag & SOR_EISENSTAT) PetscCall(VecDuplicate(bb, &bb1));
1413: if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) {
1414: if (flag & SOR_ZERO_INITIAL_GUESS) {
1415: PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
1416: its--;
1417: }
1419: while (its--) {
1420: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1421: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1423: /* update rhs: bb1 = bb - B*x */
1424: PetscCall(VecScale(mat->lvec, -1.0));
1425: PetscCall((*mat->B->ops->multadd)(mat->B, mat->lvec, bb, bb1));
1427: /* local sweep */
1428: PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, SOR_SYMMETRIC_SWEEP, fshift, lits, 1, xx));
1429: }
1430: } else if (flag & SOR_LOCAL_FORWARD_SWEEP) {
1431: if (flag & SOR_ZERO_INITIAL_GUESS) {
1432: PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
1433: its--;
1434: }
1435: while (its--) {
1436: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1437: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1439: /* update rhs: bb1 = bb - B*x */
1440: PetscCall(VecScale(mat->lvec, -1.0));
1441: PetscCall((*mat->B->ops->multadd)(mat->B, mat->lvec, bb, bb1));
1443: /* local sweep */
1444: PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, SOR_FORWARD_SWEEP, fshift, lits, 1, xx));
1445: }
1446: } else if (flag & SOR_LOCAL_BACKWARD_SWEEP) {
1447: if (flag & SOR_ZERO_INITIAL_GUESS) {
1448: PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
1449: its--;
1450: }
1451: while (its--) {
1452: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1453: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1455: /* update rhs: bb1 = bb - B*x */
1456: PetscCall(VecScale(mat->lvec, -1.0));
1457: PetscCall((*mat->B->ops->multadd)(mat->B, mat->lvec, bb, bb1));
1459: /* local sweep */
1460: PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, SOR_BACKWARD_SWEEP, fshift, lits, 1, xx));
1461: }
1462: } else if (flag & SOR_EISENSTAT) {
1463: Vec xx1;
1465: PetscCall(VecDuplicate(bb, &xx1));
1466: PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, (MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_BACKWARD_SWEEP), fshift, lits, 1, xx));
1468: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1469: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1470: if (!mat->diag) {
1471: PetscCall(MatCreateVecs(matin, &mat->diag, NULL));
1472: PetscCall(MatGetDiagonal(matin, mat->diag));
1473: }
1474: PetscCall(MatHasOperation(matin, MATOP_MULT_DIAGONAL_BLOCK, &hasop));
1475: if (hasop) {
1476: PetscCall(MatMultDiagonalBlock(matin, xx, bb1));
1477: } else {
1478: PetscCall(VecPointwiseMult(bb1, mat->diag, xx));
1479: }
1480: PetscCall(VecAYPX(bb1, (omega - 2.0) / omega, bb));
1482: PetscCall(MatMultAdd(mat->B, mat->lvec, bb1, bb1));
1484: /* local sweep */
1485: PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, (MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_FORWARD_SWEEP), fshift, lits, 1, xx1));
1486: PetscCall(VecAXPY(xx, 1.0, xx1));
1487: PetscCall(VecDestroy(&xx1));
1488: } else SETERRQ(PetscObjectComm((PetscObject)matin), PETSC_ERR_SUP, "Parallel SOR not supported");
1490: PetscCall(VecDestroy(&bb1));
1492: matin->factorerrortype = mat->A->factorerrortype;
1493: PetscFunctionReturn(PETSC_SUCCESS);
1494: }
1496: static PetscErrorCode MatPermute_MPIAIJ(Mat A, IS rowp, IS colp, Mat *B)
1497: {
1498: Mat aA, aB, Aperm;
1499: const PetscInt *rwant, *cwant, *gcols, *ai, *bi, *aj, *bj;
1500: PetscScalar *aa, *ba;
1501: PetscInt i, j, m, n, ng, anz, bnz, *dnnz, *onnz, *tdnnz, *tonnz, *rdest, *cdest, *work, *gcdest;
1502: PetscSF rowsf, sf;
1503: IS parcolp = NULL;
1504: PetscBool done;
1506: PetscFunctionBegin;
1507: PetscCall(MatGetLocalSize(A, &m, &n));
1508: PetscCall(ISGetIndices(rowp, &rwant));
1509: PetscCall(ISGetIndices(colp, &cwant));
1510: PetscCall(PetscMalloc3(PetscMax(m, n), &work, m, &rdest, n, &cdest));
1512: /* Invert row permutation to find out where my rows should go */
1513: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &rowsf));
1514: PetscCall(PetscSFSetGraphLayout(rowsf, A->rmap, A->rmap->n, NULL, PETSC_OWN_POINTER, rwant));
1515: PetscCall(PetscSFSetFromOptions(rowsf));
1516: for (i = 0; i < m; i++) work[i] = A->rmap->rstart + i;
1517: PetscCall(PetscSFReduceBegin(rowsf, MPIU_INT, work, rdest, MPI_REPLACE));
1518: PetscCall(PetscSFReduceEnd(rowsf, MPIU_INT, work, rdest, MPI_REPLACE));
1520: /* Invert column permutation to find out where my columns should go */
1521: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
1522: PetscCall(PetscSFSetGraphLayout(sf, A->cmap, A->cmap->n, NULL, PETSC_OWN_POINTER, cwant));
1523: PetscCall(PetscSFSetFromOptions(sf));
1524: for (i = 0; i < n; i++) work[i] = A->cmap->rstart + i;
1525: PetscCall(PetscSFReduceBegin(sf, MPIU_INT, work, cdest, MPI_REPLACE));
1526: PetscCall(PetscSFReduceEnd(sf, MPIU_INT, work, cdest, MPI_REPLACE));
1527: PetscCall(PetscSFDestroy(&sf));
1529: PetscCall(ISRestoreIndices(rowp, &rwant));
1530: PetscCall(ISRestoreIndices(colp, &cwant));
1531: PetscCall(MatMPIAIJGetSeqAIJ(A, &aA, &aB, &gcols));
1533: /* Find out where my gcols should go */
1534: PetscCall(MatGetSize(aB, NULL, &ng));
1535: PetscCall(PetscMalloc1(ng, &gcdest));
1536: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
1537: PetscCall(PetscSFSetGraphLayout(sf, A->cmap, ng, NULL, PETSC_OWN_POINTER, gcols));
1538: PetscCall(PetscSFSetFromOptions(sf));
1539: PetscCall(PetscSFBcastBegin(sf, MPIU_INT, cdest, gcdest, MPI_REPLACE));
1540: PetscCall(PetscSFBcastEnd(sf, MPIU_INT, cdest, gcdest, MPI_REPLACE));
1541: PetscCall(PetscSFDestroy(&sf));
1543: PetscCall(PetscCalloc4(m, &dnnz, m, &onnz, m, &tdnnz, m, &tonnz));
1544: PetscCall(MatGetRowIJ(aA, 0, PETSC_FALSE, PETSC_FALSE, &anz, &ai, &aj, &done));
1545: PetscCall(MatGetRowIJ(aB, 0, PETSC_FALSE, PETSC_FALSE, &bnz, &bi, &bj, &done));
1546: for (i = 0; i < m; i++) {
1547: PetscInt row = rdest[i];
1548: PetscMPIInt rowner;
1549: PetscCall(PetscLayoutFindOwner(A->rmap, row, &rowner));
1550: for (j = ai[i]; j < ai[i + 1]; j++) {
1551: PetscInt col = cdest[aj[j]];
1552: PetscMPIInt cowner;
1553: PetscCall(PetscLayoutFindOwner(A->cmap, col, &cowner)); /* Could build an index for the columns to eliminate this search */
1554: if (rowner == cowner) dnnz[i]++;
1555: else onnz[i]++;
1556: }
1557: for (j = bi[i]; j < bi[i + 1]; j++) {
1558: PetscInt col = gcdest[bj[j]];
1559: PetscMPIInt cowner;
1560: PetscCall(PetscLayoutFindOwner(A->cmap, col, &cowner));
1561: if (rowner == cowner) dnnz[i]++;
1562: else onnz[i]++;
1563: }
1564: }
1565: PetscCall(PetscSFBcastBegin(rowsf, MPIU_INT, dnnz, tdnnz, MPI_REPLACE));
1566: PetscCall(PetscSFBcastEnd(rowsf, MPIU_INT, dnnz, tdnnz, MPI_REPLACE));
1567: PetscCall(PetscSFBcastBegin(rowsf, MPIU_INT, onnz, tonnz, MPI_REPLACE));
1568: PetscCall(PetscSFBcastEnd(rowsf, MPIU_INT, onnz, tonnz, MPI_REPLACE));
1569: PetscCall(PetscSFDestroy(&rowsf));
1571: PetscCall(MatCreateAIJ(PetscObjectComm((PetscObject)A), A->rmap->n, A->cmap->n, A->rmap->N, A->cmap->N, 0, tdnnz, 0, tonnz, &Aperm));
1572: PetscCall(MatSeqAIJGetArray(aA, &aa));
1573: PetscCall(MatSeqAIJGetArray(aB, &ba));
1574: for (i = 0; i < m; i++) {
1575: PetscInt *acols = dnnz, *bcols = onnz; /* Repurpose now-unneeded arrays */
1576: PetscInt j0, rowlen;
1577: rowlen = ai[i + 1] - ai[i];
1578: for (j0 = j = 0; j < rowlen; j0 = j) { /* rowlen could be larger than number of rows m, so sum in batches */
1579: for (; j < PetscMin(rowlen, j0 + m); j++) acols[j - j0] = cdest[aj[ai[i] + j]];
1580: PetscCall(MatSetValues(Aperm, 1, &rdest[i], j - j0, acols, aa + ai[i] + j0, INSERT_VALUES));
1581: }
1582: rowlen = bi[i + 1] - bi[i];
1583: for (j0 = j = 0; j < rowlen; j0 = j) {
1584: for (; j < PetscMin(rowlen, j0 + m); j++) bcols[j - j0] = gcdest[bj[bi[i] + j]];
1585: PetscCall(MatSetValues(Aperm, 1, &rdest[i], j - j0, bcols, ba + bi[i] + j0, INSERT_VALUES));
1586: }
1587: }
1588: PetscCall(MatAssemblyBegin(Aperm, MAT_FINAL_ASSEMBLY));
1589: PetscCall(MatAssemblyEnd(Aperm, MAT_FINAL_ASSEMBLY));
1590: PetscCall(MatRestoreRowIJ(aA, 0, PETSC_FALSE, PETSC_FALSE, &anz, &ai, &aj, &done));
1591: PetscCall(MatRestoreRowIJ(aB, 0, PETSC_FALSE, PETSC_FALSE, &bnz, &bi, &bj, &done));
1592: PetscCall(MatSeqAIJRestoreArray(aA, &aa));
1593: PetscCall(MatSeqAIJRestoreArray(aB, &ba));
1594: PetscCall(PetscFree4(dnnz, onnz, tdnnz, tonnz));
1595: PetscCall(PetscFree3(work, rdest, cdest));
1596: PetscCall(PetscFree(gcdest));
1597: if (parcolp) PetscCall(ISDestroy(&colp));
1598: *B = Aperm;
1599: PetscFunctionReturn(PETSC_SUCCESS);
1600: }
1602: static PetscErrorCode MatGetGhosts_MPIAIJ(Mat mat, PetscInt *nghosts, const PetscInt *ghosts[])
1603: {
1604: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1606: PetscFunctionBegin;
1607: PetscCall(MatGetSize(aij->B, NULL, nghosts));
1608: if (ghosts) *ghosts = aij->garray;
1609: PetscFunctionReturn(PETSC_SUCCESS);
1610: }
1612: static PetscErrorCode MatGetInfo_MPIAIJ(Mat matin, MatInfoType flag, MatInfo *info)
1613: {
1614: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)matin->data;
1615: Mat A = mat->A, B = mat->B;
1616: PetscLogDouble isend[5], irecv[5];
1618: PetscFunctionBegin;
1619: info->block_size = 1.0;
1620: PetscCall(MatGetInfo(A, MAT_LOCAL, info));
1622: isend[0] = info->nz_used;
1623: isend[1] = info->nz_allocated;
1624: isend[2] = info->nz_unneeded;
1625: isend[3] = info->memory;
1626: isend[4] = info->mallocs;
1628: PetscCall(MatGetInfo(B, MAT_LOCAL, info));
1630: isend[0] += info->nz_used;
1631: isend[1] += info->nz_allocated;
1632: isend[2] += info->nz_unneeded;
1633: isend[3] += info->memory;
1634: isend[4] += info->mallocs;
1635: if (flag == MAT_LOCAL) {
1636: info->nz_used = isend[0];
1637: info->nz_allocated = isend[1];
1638: info->nz_unneeded = isend[2];
1639: info->memory = isend[3];
1640: info->mallocs = isend[4];
1641: } else if (flag == MAT_GLOBAL_MAX) {
1642: PetscCallMPI(MPIU_Allreduce(isend, irecv, 5, MPIU_PETSCLOGDOUBLE, MPI_MAX, 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: } else if (flag == MAT_GLOBAL_SUM) {
1650: PetscCallMPI(MPIU_Allreduce(isend, irecv, 5, MPIU_PETSCLOGDOUBLE, MPI_SUM, PetscObjectComm((PetscObject)matin)));
1652: info->nz_used = irecv[0];
1653: info->nz_allocated = irecv[1];
1654: info->nz_unneeded = irecv[2];
1655: info->memory = irecv[3];
1656: info->mallocs = irecv[4];
1657: }
1658: info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */
1659: info->fill_ratio_needed = 0;
1660: info->factor_mallocs = 0;
1661: PetscFunctionReturn(PETSC_SUCCESS);
1662: }
1664: PetscErrorCode MatSetOption_MPIAIJ(Mat A, MatOption op, PetscBool flg)
1665: {
1666: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1668: PetscFunctionBegin;
1669: switch (op) {
1670: case MAT_NEW_NONZERO_LOCATIONS:
1671: case MAT_NEW_NONZERO_ALLOCATION_ERR:
1672: case MAT_UNUSED_NONZERO_LOCATION_ERR:
1673: case MAT_KEEP_NONZERO_PATTERN:
1674: case MAT_NEW_NONZERO_LOCATION_ERR:
1675: case MAT_USE_INODES:
1676: case MAT_IGNORE_ZERO_ENTRIES:
1677: case MAT_FORM_EXPLICIT_TRANSPOSE:
1678: MatCheckPreallocated(A, 1);
1679: PetscCall(MatSetOption(a->A, op, flg));
1680: PetscCall(MatSetOption(a->B, op, flg));
1681: break;
1682: case MAT_ROW_ORIENTED:
1683: MatCheckPreallocated(A, 1);
1684: a->roworiented = flg;
1686: PetscCall(MatSetOption(a->A, op, flg));
1687: PetscCall(MatSetOption(a->B, op, flg));
1688: break;
1689: case MAT_IGNORE_OFF_PROC_ENTRIES:
1690: a->donotstash = flg;
1691: break;
1692: /* Symmetry flags are handled directly by MatSetOption() and they don't affect preallocation */
1693: case MAT_SPD:
1694: case MAT_SYMMETRIC:
1695: case MAT_STRUCTURALLY_SYMMETRIC:
1696: case MAT_HERMITIAN:
1697: case MAT_SYMMETRY_ETERNAL:
1698: case MAT_STRUCTURAL_SYMMETRY_ETERNAL:
1699: case MAT_SPD_ETERNAL:
1700: /* if the diagonal matrix is square it inherits some of the properties above */
1701: if (a->A && A->rmap->n == A->cmap->n) PetscCall(MatSetOption(a->A, op, flg));
1702: break;
1703: case MAT_SUBMAT_SINGLEIS:
1704: A->submat_singleis = flg;
1705: break;
1706: default:
1707: break;
1708: }
1709: PetscFunctionReturn(PETSC_SUCCESS);
1710: }
1712: PetscErrorCode MatGetRow_MPIAIJ(Mat matin, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
1713: {
1714: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)matin->data;
1715: PetscScalar *vworkA, *vworkB, **pvA, **pvB, *v_p;
1716: PetscInt i, *cworkA, *cworkB, **pcA, **pcB, cstart = matin->cmap->rstart;
1717: PetscInt nztot, nzA, nzB, lrow, rstart = matin->rmap->rstart, rend = matin->rmap->rend;
1718: PetscInt *cmap, *idx_p;
1720: PetscFunctionBegin;
1721: PetscCheck(!mat->getrowactive, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Already active");
1722: mat->getrowactive = PETSC_TRUE;
1724: if (!mat->rowvalues && (idx || v)) {
1725: /*
1726: allocate enough space to hold information from the longest row.
1727: */
1728: Mat_SeqAIJ *Aa = (Mat_SeqAIJ *)mat->A->data, *Ba = (Mat_SeqAIJ *)mat->B->data;
1729: PetscInt max = 1, tmp;
1730: for (i = 0; i < matin->rmap->n; i++) {
1731: tmp = Aa->i[i + 1] - Aa->i[i] + Ba->i[i + 1] - Ba->i[i];
1732: if (max < tmp) max = tmp;
1733: }
1734: PetscCall(PetscMalloc2(max, &mat->rowvalues, max, &mat->rowindices));
1735: }
1737: PetscCheck(row >= rstart && row < rend, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Only local rows");
1738: lrow = row - rstart;
1740: pvA = &vworkA;
1741: pcA = &cworkA;
1742: pvB = &vworkB;
1743: pcB = &cworkB;
1744: if (!v) {
1745: pvA = NULL;
1746: pvB = NULL;
1747: }
1748: if (!idx) {
1749: pcA = NULL;
1750: if (!v) pcB = NULL;
1751: }
1752: PetscCall((*mat->A->ops->getrow)(mat->A, lrow, &nzA, pcA, pvA));
1753: PetscCall((*mat->B->ops->getrow)(mat->B, lrow, &nzB, pcB, pvB));
1754: nztot = nzA + nzB;
1756: cmap = mat->garray;
1757: if (v || idx) {
1758: if (nztot) {
1759: /* Sort by increasing column numbers, assuming A and B already sorted */
1760: PetscInt imark = -1;
1761: if (v) {
1762: *v = v_p = mat->rowvalues;
1763: for (i = 0; i < nzB; i++) {
1764: if (cmap[cworkB[i]] < cstart) v_p[i] = vworkB[i];
1765: else break;
1766: }
1767: imark = i;
1768: for (i = 0; i < nzA; i++) v_p[imark + i] = vworkA[i];
1769: for (i = imark; i < nzB; i++) v_p[nzA + i] = vworkB[i];
1770: }
1771: if (idx) {
1772: *idx = idx_p = mat->rowindices;
1773: if (imark > -1) {
1774: for (i = 0; i < imark; i++) idx_p[i] = cmap[cworkB[i]];
1775: } else {
1776: for (i = 0; i < nzB; i++) {
1777: if (cmap[cworkB[i]] < cstart) idx_p[i] = cmap[cworkB[i]];
1778: else break;
1779: }
1780: imark = i;
1781: }
1782: for (i = 0; i < nzA; i++) idx_p[imark + i] = cstart + cworkA[i];
1783: for (i = imark; i < nzB; i++) idx_p[nzA + i] = cmap[cworkB[i]];
1784: }
1785: } else {
1786: if (idx) *idx = NULL;
1787: if (v) *v = NULL;
1788: }
1789: }
1790: *nz = nztot;
1791: PetscCall((*mat->A->ops->restorerow)(mat->A, lrow, &nzA, pcA, pvA));
1792: PetscCall((*mat->B->ops->restorerow)(mat->B, lrow, &nzB, pcB, pvB));
1793: PetscFunctionReturn(PETSC_SUCCESS);
1794: }
1796: PetscErrorCode MatRestoreRow_MPIAIJ(Mat mat, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
1797: {
1798: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1800: PetscFunctionBegin;
1801: PetscCheck(aij->getrowactive, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "MatGetRow() must be called first");
1802: aij->getrowactive = PETSC_FALSE;
1803: PetscFunctionReturn(PETSC_SUCCESS);
1804: }
1806: static PetscErrorCode MatNorm_MPIAIJ(Mat mat, NormType type, PetscReal *norm)
1807: {
1808: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1809: Mat_SeqAIJ *amat = (Mat_SeqAIJ *)aij->A->data, *bmat = (Mat_SeqAIJ *)aij->B->data;
1810: PetscInt i, j;
1811: PetscReal sum = 0.0;
1812: const MatScalar *v, *amata, *bmata;
1814: PetscFunctionBegin;
1815: if (aij->size == 1) {
1816: PetscCall(MatNorm(aij->A, type, norm));
1817: } else {
1818: PetscCall(MatSeqAIJGetArrayRead(aij->A, &amata));
1819: PetscCall(MatSeqAIJGetArrayRead(aij->B, &bmata));
1820: if (type == NORM_FROBENIUS) {
1821: v = amata;
1822: for (i = 0; i < amat->nz; i++) {
1823: sum += PetscRealPart(PetscConj(*v) * (*v));
1824: v++;
1825: }
1826: v = bmata;
1827: for (i = 0; i < bmat->nz; i++) {
1828: sum += PetscRealPart(PetscConj(*v) * (*v));
1829: v++;
1830: }
1831: PetscCallMPI(MPIU_Allreduce(&sum, norm, 1, MPIU_REAL, MPIU_SUM, PetscObjectComm((PetscObject)mat)));
1832: *norm = PetscSqrtReal(*norm);
1833: PetscCall(PetscLogFlops(2.0 * amat->nz + 2.0 * bmat->nz));
1834: } else if (type == NORM_1) { /* max column norm */
1835: Vec col, bcol;
1836: PetscScalar *array;
1837: PetscInt *jj, *garray = aij->garray;
1839: PetscCall(MatCreateVecs(mat, &col, NULL));
1840: PetscCall(VecSet(col, 0.0));
1841: PetscCall(VecGetArrayWrite(col, &array));
1842: v = amata;
1843: jj = amat->j;
1844: for (j = 0; j < amat->nz; j++) array[*jj++] += PetscAbsScalar(*v++);
1845: PetscCall(VecRestoreArrayWrite(col, &array));
1846: PetscCall(MatCreateVecs(aij->B, &bcol, NULL));
1847: PetscCall(VecSet(bcol, 0.0));
1848: PetscCall(VecGetArrayWrite(bcol, &array));
1849: v = bmata;
1850: jj = bmat->j;
1851: for (j = 0; j < bmat->nz; j++) array[*jj++] += PetscAbsScalar(*v++);
1852: PetscCall(VecSetValues(col, aij->B->cmap->n, garray, array, ADD_VALUES));
1853: PetscCall(VecRestoreArrayWrite(bcol, &array));
1854: PetscCall(VecDestroy(&bcol));
1855: PetscCall(VecAssemblyBegin(col));
1856: PetscCall(VecAssemblyEnd(col));
1857: PetscCall(VecNorm(col, NORM_INFINITY, norm));
1858: PetscCall(VecDestroy(&col));
1859: } else if (type == NORM_INFINITY) { /* max row norm */
1860: PetscReal ntemp = 0.0;
1861: for (j = 0; j < aij->A->rmap->n; j++) {
1862: v = PetscSafePointerPlusOffset(amata, amat->i[j]);
1863: sum = 0.0;
1864: for (i = 0; i < amat->i[j + 1] - amat->i[j]; i++) {
1865: sum += PetscAbsScalar(*v);
1866: v++;
1867: }
1868: v = PetscSafePointerPlusOffset(bmata, bmat->i[j]);
1869: for (i = 0; i < bmat->i[j + 1] - bmat->i[j]; i++) {
1870: sum += PetscAbsScalar(*v);
1871: v++;
1872: }
1873: if (sum > ntemp) ntemp = sum;
1874: }
1875: PetscCallMPI(MPIU_Allreduce(&ntemp, norm, 1, MPIU_REAL, MPIU_MAX, PetscObjectComm((PetscObject)mat)));
1876: PetscCall(PetscLogFlops(PetscMax(amat->nz + bmat->nz - 1, 0)));
1877: } else SETERRQ(PetscObjectComm((PetscObject)mat), PETSC_ERR_SUP, "No support for two norm");
1878: PetscCall(MatSeqAIJRestoreArrayRead(aij->A, &amata));
1879: PetscCall(MatSeqAIJRestoreArrayRead(aij->B, &bmata));
1880: }
1881: PetscFunctionReturn(PETSC_SUCCESS);
1882: }
1884: static PetscErrorCode MatTranspose_MPIAIJ(Mat A, MatReuse reuse, Mat *matout)
1885: {
1886: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data, *b;
1887: Mat_SeqAIJ *Aloc = (Mat_SeqAIJ *)a->A->data, *Bloc = (Mat_SeqAIJ *)a->B->data, *sub_B_diag;
1888: 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;
1889: const PetscInt *ai, *aj, *bi, *bj, *B_diag_i;
1890: Mat B, A_diag, *B_diag;
1891: const MatScalar *pbv, *bv;
1893: PetscFunctionBegin;
1894: if (reuse == MAT_REUSE_MATRIX) PetscCall(MatTransposeCheckNonzeroState_Private(A, *matout));
1895: ma = A->rmap->n;
1896: na = A->cmap->n;
1897: mb = a->B->rmap->n;
1898: nb = a->B->cmap->n;
1899: ai = Aloc->i;
1900: aj = Aloc->j;
1901: bi = Bloc->i;
1902: bj = Bloc->j;
1903: if (reuse == MAT_INITIAL_MATRIX || *matout == A) {
1904: PetscInt *d_nnz, *g_nnz, *o_nnz;
1905: PetscSFNode *oloc;
1906: PETSC_UNUSED PetscSF sf;
1908: PetscCall(PetscMalloc4(na, &d_nnz, na, &o_nnz, nb, &g_nnz, nb, &oloc));
1909: /* compute d_nnz for preallocation */
1910: PetscCall(PetscArrayzero(d_nnz, na));
1911: for (i = 0; i < ai[ma]; i++) d_nnz[aj[i]]++;
1912: /* compute local off-diagonal contributions */
1913: PetscCall(PetscArrayzero(g_nnz, nb));
1914: for (i = 0; i < bi[ma]; i++) g_nnz[bj[i]]++;
1915: /* map those to global */
1916: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
1917: PetscCall(PetscSFSetGraphLayout(sf, A->cmap, nb, NULL, PETSC_USE_POINTER, a->garray));
1918: PetscCall(PetscSFSetFromOptions(sf));
1919: PetscCall(PetscArrayzero(o_nnz, na));
1920: PetscCall(PetscSFReduceBegin(sf, MPIU_INT, g_nnz, o_nnz, MPI_SUM));
1921: PetscCall(PetscSFReduceEnd(sf, MPIU_INT, g_nnz, o_nnz, MPI_SUM));
1922: PetscCall(PetscSFDestroy(&sf));
1924: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &B));
1925: PetscCall(MatSetSizes(B, A->cmap->n, A->rmap->n, N, M));
1926: PetscCall(MatSetBlockSizes(B, A->cmap->bs, A->rmap->bs));
1927: PetscCall(MatSetType(B, ((PetscObject)A)->type_name));
1928: PetscCall(MatMPIAIJSetPreallocation(B, 0, d_nnz, 0, o_nnz));
1929: PetscCall(PetscFree4(d_nnz, o_nnz, g_nnz, oloc));
1930: } else {
1931: B = *matout;
1932: PetscCall(MatSetOption(B, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_TRUE));
1933: }
1935: b = (Mat_MPIAIJ *)B->data;
1936: A_diag = a->A;
1937: B_diag = &b->A;
1938: sub_B_diag = (Mat_SeqAIJ *)(*B_diag)->data;
1939: A_diag_ncol = A_diag->cmap->N;
1940: B_diag_ilen = sub_B_diag->ilen;
1941: B_diag_i = sub_B_diag->i;
1943: /* Set ilen for diagonal of B */
1944: for (i = 0; i < A_diag_ncol; i++) B_diag_ilen[i] = B_diag_i[i + 1] - B_diag_i[i];
1946: /* Transpose the diagonal part of the matrix. In contrast to the off-diagonal part, this can be done
1947: very quickly (=without using MatSetValues), because all writes are local. */
1948: PetscCall(MatTransposeSetPrecursor(A_diag, *B_diag));
1949: PetscCall(MatTranspose(A_diag, MAT_REUSE_MATRIX, B_diag));
1951: /* copy over the B part */
1952: PetscCall(PetscMalloc1(bi[mb], &cols));
1953: PetscCall(MatSeqAIJGetArrayRead(a->B, &bv));
1954: pbv = bv;
1955: row = A->rmap->rstart;
1956: for (i = 0; i < bi[mb]; i++) cols[i] = a->garray[bj[i]];
1957: cols_tmp = cols;
1958: for (i = 0; i < mb; i++) {
1959: ncol = bi[i + 1] - bi[i];
1960: PetscCall(MatSetValues(B, ncol, cols_tmp, 1, &row, pbv, INSERT_VALUES));
1961: row++;
1962: if (pbv) pbv += ncol;
1963: if (cols_tmp) cols_tmp += ncol;
1964: }
1965: PetscCall(PetscFree(cols));
1966: PetscCall(MatSeqAIJRestoreArrayRead(a->B, &bv));
1968: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
1969: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
1970: if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX) {
1971: *matout = B;
1972: } else {
1973: PetscCall(MatHeaderMerge(A, &B));
1974: }
1975: PetscFunctionReturn(PETSC_SUCCESS);
1976: }
1978: static PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat, Vec ll, Vec rr)
1979: {
1980: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1981: Mat a = aij->A, b = aij->B;
1982: PetscInt s1, s2, s3;
1984: PetscFunctionBegin;
1985: PetscCall(MatGetLocalSize(mat, &s2, &s3));
1986: if (rr) {
1987: PetscCall(VecGetLocalSize(rr, &s1));
1988: PetscCheck(s1 == s3, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "right vector non-conforming local size");
1989: /* Overlap communication with computation. */
1990: PetscCall(VecScatterBegin(aij->Mvctx, rr, aij->lvec, INSERT_VALUES, SCATTER_FORWARD));
1991: }
1992: if (ll) {
1993: PetscCall(VecGetLocalSize(ll, &s1));
1994: PetscCheck(s1 == s2, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "left vector non-conforming local size");
1995: PetscUseTypeMethod(b, diagonalscale, ll, NULL);
1996: }
1997: /* scale the diagonal block */
1998: PetscUseTypeMethod(a, diagonalscale, ll, rr);
2000: if (rr) {
2001: /* Do a scatter end and then right scale the off-diagonal block */
2002: PetscCall(VecScatterEnd(aij->Mvctx, rr, aij->lvec, INSERT_VALUES, SCATTER_FORWARD));
2003: PetscUseTypeMethod(b, diagonalscale, NULL, aij->lvec);
2004: }
2005: PetscFunctionReturn(PETSC_SUCCESS);
2006: }
2008: static PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A)
2009: {
2010: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2012: PetscFunctionBegin;
2013: PetscCall(MatSetUnfactored(a->A));
2014: PetscFunctionReturn(PETSC_SUCCESS);
2015: }
2017: static PetscErrorCode MatEqual_MPIAIJ(Mat A, Mat B, PetscBool *flag)
2018: {
2019: Mat_MPIAIJ *matB = (Mat_MPIAIJ *)B->data, *matA = (Mat_MPIAIJ *)A->data;
2020: Mat a, b, c, d;
2021: PetscBool flg;
2023: PetscFunctionBegin;
2024: a = matA->A;
2025: b = matA->B;
2026: c = matB->A;
2027: d = matB->B;
2029: PetscCall(MatEqual(a, c, &flg));
2030: if (flg) PetscCall(MatEqual(b, d, &flg));
2031: PetscCallMPI(MPIU_Allreduce(&flg, flag, 1, MPI_C_BOOL, MPI_LAND, PetscObjectComm((PetscObject)A)));
2032: PetscFunctionReturn(PETSC_SUCCESS);
2033: }
2035: static PetscErrorCode MatCopy_MPIAIJ(Mat A, Mat B, MatStructure str)
2036: {
2037: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2038: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2040: PetscFunctionBegin;
2041: /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */
2042: if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) {
2043: /* because of the column compression in the off-processor part of the matrix a->B,
2044: the number of columns in a->B and b->B may be different, hence we cannot call
2045: the MatCopy() directly on the two parts. If need be, we can provide a more
2046: efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices
2047: then copying the submatrices */
2048: PetscCall(MatCopy_Basic(A, B, str));
2049: } else {
2050: PetscCall(MatCopy(a->A, b->A, str));
2051: PetscCall(MatCopy(a->B, b->B, str));
2052: }
2053: PetscCall(PetscObjectStateIncrease((PetscObject)B));
2054: PetscFunctionReturn(PETSC_SUCCESS);
2055: }
2057: /*
2058: Computes the number of nonzeros per row needed for preallocation when X and Y
2059: have different nonzero structure.
2060: */
2061: 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)
2062: {
2063: PetscInt i, j, k, nzx, nzy;
2065: PetscFunctionBegin;
2066: /* Set the number of nonzeros in the new matrix */
2067: for (i = 0; i < m; i++) {
2068: const PetscInt *xjj = PetscSafePointerPlusOffset(xj, xi[i]), *yjj = PetscSafePointerPlusOffset(yj, yi[i]);
2069: nzx = xi[i + 1] - xi[i];
2070: nzy = yi[i + 1] - yi[i];
2071: nnz[i] = 0;
2072: for (j = 0, k = 0; j < nzx; j++) { /* Point in X */
2073: for (; k < nzy && yltog[yjj[k]] < xltog[xjj[j]]; k++) nnz[i]++; /* Catch up to X */
2074: if (k < nzy && yltog[yjj[k]] == xltog[xjj[j]]) k++; /* Skip duplicate */
2075: nnz[i]++;
2076: }
2077: for (; k < nzy; k++) nnz[i]++;
2078: }
2079: PetscFunctionReturn(PETSC_SUCCESS);
2080: }
2082: /* This is the same as MatAXPYGetPreallocation_SeqAIJ, except that the local-to-global map is provided */
2083: static PetscErrorCode MatAXPYGetPreallocation_MPIAIJ(Mat Y, const PetscInt *yltog, Mat X, const PetscInt *xltog, PetscInt *nnz)
2084: {
2085: PetscInt m = Y->rmap->N;
2086: Mat_SeqAIJ *x = (Mat_SeqAIJ *)X->data;
2087: Mat_SeqAIJ *y = (Mat_SeqAIJ *)Y->data;
2089: PetscFunctionBegin;
2090: PetscCall(MatAXPYGetPreallocation_MPIX_private(m, x->i, x->j, xltog, y->i, y->j, yltog, nnz));
2091: PetscFunctionReturn(PETSC_SUCCESS);
2092: }
2094: static PetscErrorCode MatAXPY_MPIAIJ(Mat Y, PetscScalar a, Mat X, MatStructure str)
2095: {
2096: Mat_MPIAIJ *xx = (Mat_MPIAIJ *)X->data, *yy = (Mat_MPIAIJ *)Y->data;
2098: PetscFunctionBegin;
2099: if (str == SAME_NONZERO_PATTERN) {
2100: PetscCall(MatAXPY(yy->A, a, xx->A, str));
2101: PetscCall(MatAXPY(yy->B, a, xx->B, str));
2102: } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */
2103: PetscCall(MatAXPY_Basic(Y, a, X, str));
2104: } else {
2105: Mat B;
2106: PetscInt *nnz_d, *nnz_o;
2108: PetscCall(PetscMalloc1(yy->A->rmap->N, &nnz_d));
2109: PetscCall(PetscMalloc1(yy->B->rmap->N, &nnz_o));
2110: PetscCall(MatCreate(PetscObjectComm((PetscObject)Y), &B));
2111: PetscCall(PetscObjectSetName((PetscObject)B, ((PetscObject)Y)->name));
2112: PetscCall(MatSetLayouts(B, Y->rmap, Y->cmap));
2113: PetscCall(MatSetType(B, ((PetscObject)Y)->type_name));
2114: PetscCall(MatAXPYGetPreallocation_SeqAIJ(yy->A, xx->A, nnz_d));
2115: PetscCall(MatAXPYGetPreallocation_MPIAIJ(yy->B, yy->garray, xx->B, xx->garray, nnz_o));
2116: PetscCall(MatMPIAIJSetPreallocation(B, 0, nnz_d, 0, nnz_o));
2117: PetscCall(MatAXPY_BasicWithPreallocation(B, Y, a, X, str));
2118: PetscCall(MatHeaderMerge(Y, &B));
2119: PetscCall(PetscFree(nnz_d));
2120: PetscCall(PetscFree(nnz_o));
2121: }
2122: PetscFunctionReturn(PETSC_SUCCESS);
2123: }
2125: PETSC_INTERN PetscErrorCode MatConjugate_SeqAIJ(Mat);
2127: static PetscErrorCode MatConjugate_MPIAIJ(Mat mat)
2128: {
2129: PetscFunctionBegin;
2130: if (PetscDefined(USE_COMPLEX)) {
2131: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2133: PetscCall(MatConjugate_SeqAIJ(aij->A));
2134: PetscCall(MatConjugate_SeqAIJ(aij->B));
2135: }
2136: PetscFunctionReturn(PETSC_SUCCESS);
2137: }
2139: static PetscErrorCode MatRealPart_MPIAIJ(Mat A)
2140: {
2141: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2143: PetscFunctionBegin;
2144: PetscCall(MatRealPart(a->A));
2145: PetscCall(MatRealPart(a->B));
2146: PetscFunctionReturn(PETSC_SUCCESS);
2147: }
2149: static PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A)
2150: {
2151: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2153: PetscFunctionBegin;
2154: PetscCall(MatImaginaryPart(a->A));
2155: PetscCall(MatImaginaryPart(a->B));
2156: PetscFunctionReturn(PETSC_SUCCESS);
2157: }
2159: static PetscErrorCode MatGetRowMaxAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2160: {
2161: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2162: PetscInt i, *idxb = NULL, m = A->rmap->n;
2163: PetscScalar *vv;
2164: Vec vB, vA;
2165: const PetscScalar *va, *vb;
2167: PetscFunctionBegin;
2168: PetscCall(MatCreateVecs(a->A, NULL, &vA));
2169: PetscCall(MatGetRowMaxAbs(a->A, vA, idx));
2171: PetscCall(VecGetArrayRead(vA, &va));
2172: if (idx) {
2173: for (i = 0; i < m; i++) {
2174: if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart;
2175: }
2176: }
2178: PetscCall(MatCreateVecs(a->B, NULL, &vB));
2179: PetscCall(PetscMalloc1(m, &idxb));
2180: PetscCall(MatGetRowMaxAbs(a->B, vB, idxb));
2182: PetscCall(VecGetArrayWrite(v, &vv));
2183: PetscCall(VecGetArrayRead(vB, &vb));
2184: for (i = 0; i < m; i++) {
2185: if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) {
2186: vv[i] = vb[i];
2187: if (idx) idx[i] = a->garray[idxb[i]];
2188: } else {
2189: vv[i] = va[i];
2190: if (idx && PetscAbsScalar(va[i]) == PetscAbsScalar(vb[i]) && idxb[i] != -1 && idx[i] > a->garray[idxb[i]]) idx[i] = a->garray[idxb[i]];
2191: }
2192: }
2193: PetscCall(VecRestoreArrayWrite(v, &vv));
2194: PetscCall(VecRestoreArrayRead(vA, &va));
2195: PetscCall(VecRestoreArrayRead(vB, &vb));
2196: PetscCall(PetscFree(idxb));
2197: PetscCall(VecDestroy(&vA));
2198: PetscCall(VecDestroy(&vB));
2199: PetscFunctionReturn(PETSC_SUCCESS);
2200: }
2202: static PetscErrorCode MatGetRowSumAbs_MPIAIJ(Mat A, Vec v)
2203: {
2204: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2205: Vec vB, vA;
2207: PetscFunctionBegin;
2208: PetscCall(MatCreateVecs(a->A, NULL, &vA));
2209: PetscCall(MatGetRowSumAbs(a->A, vA));
2210: PetscCall(MatCreateVecs(a->B, NULL, &vB));
2211: PetscCall(MatGetRowSumAbs(a->B, vB));
2212: PetscCall(VecAXPY(vA, 1.0, vB));
2213: PetscCall(VecDestroy(&vB));
2214: PetscCall(VecCopy(vA, v));
2215: PetscCall(VecDestroy(&vA));
2216: PetscFunctionReturn(PETSC_SUCCESS);
2217: }
2219: static PetscErrorCode MatGetRowMinAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2220: {
2221: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
2222: PetscInt m = A->rmap->n, n = A->cmap->n;
2223: PetscInt cstart = A->cmap->rstart, cend = A->cmap->rend;
2224: PetscInt *cmap = mat->garray;
2225: PetscInt *diagIdx, *offdiagIdx;
2226: Vec diagV, offdiagV;
2227: PetscScalar *a, *diagA, *offdiagA;
2228: const PetscScalar *ba, *bav;
2229: PetscInt r, j, col, ncols, *bi, *bj;
2230: Mat B = mat->B;
2231: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2233: PetscFunctionBegin;
2234: /* When a process holds entire A and other processes have no entry */
2235: if (A->cmap->N == n) {
2236: PetscCall(VecGetArrayWrite(v, &diagA));
2237: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2238: PetscCall(MatGetRowMinAbs(mat->A, diagV, idx));
2239: PetscCall(VecDestroy(&diagV));
2240: PetscCall(VecRestoreArrayWrite(v, &diagA));
2241: PetscFunctionReturn(PETSC_SUCCESS);
2242: } else if (n == 0) {
2243: if (m) {
2244: PetscCall(VecGetArrayWrite(v, &a));
2245: for (r = 0; r < m; r++) {
2246: a[r] = 0.0;
2247: if (idx) idx[r] = -1;
2248: }
2249: PetscCall(VecRestoreArrayWrite(v, &a));
2250: }
2251: PetscFunctionReturn(PETSC_SUCCESS);
2252: }
2254: PetscCall(PetscMalloc2(m, &diagIdx, m, &offdiagIdx));
2255: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2256: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2257: PetscCall(MatGetRowMinAbs(mat->A, diagV, diagIdx));
2259: /* Get offdiagIdx[] for implicit 0.0 */
2260: PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2261: ba = bav;
2262: bi = b->i;
2263: bj = b->j;
2264: PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2265: for (r = 0; r < m; r++) {
2266: ncols = bi[r + 1] - bi[r];
2267: if (ncols == A->cmap->N - n) { /* Brow is dense */
2268: offdiagA[r] = *ba;
2269: offdiagIdx[r] = cmap[0];
2270: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2271: offdiagA[r] = 0.0;
2273: /* Find first hole in the cmap */
2274: for (j = 0; j < ncols; j++) {
2275: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2276: if (col > j && j < cstart) {
2277: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2278: break;
2279: } else if (col > j + n && j >= cstart) {
2280: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2281: break;
2282: }
2283: }
2284: if (j == ncols && ncols < A->cmap->N - n) {
2285: /* a hole is outside compressed Bcols */
2286: if (ncols == 0) {
2287: if (cstart) {
2288: offdiagIdx[r] = 0;
2289: } else offdiagIdx[r] = cend;
2290: } else { /* ncols > 0 */
2291: offdiagIdx[r] = cmap[ncols - 1] + 1;
2292: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2293: }
2294: }
2295: }
2297: for (j = 0; j < ncols; j++) {
2298: if (PetscAbsScalar(offdiagA[r]) > PetscAbsScalar(*ba)) {
2299: offdiagA[r] = *ba;
2300: offdiagIdx[r] = cmap[*bj];
2301: }
2302: ba++;
2303: bj++;
2304: }
2305: }
2307: PetscCall(VecGetArrayWrite(v, &a));
2308: PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2309: for (r = 0; r < m; ++r) {
2310: if (PetscAbsScalar(diagA[r]) < PetscAbsScalar(offdiagA[r])) {
2311: a[r] = diagA[r];
2312: if (idx) idx[r] = cstart + diagIdx[r];
2313: } else if (PetscAbsScalar(diagA[r]) == PetscAbsScalar(offdiagA[r])) {
2314: a[r] = diagA[r];
2315: if (idx) {
2316: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2317: idx[r] = cstart + diagIdx[r];
2318: } else idx[r] = offdiagIdx[r];
2319: }
2320: } else {
2321: a[r] = offdiagA[r];
2322: if (idx) idx[r] = offdiagIdx[r];
2323: }
2324: }
2325: PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2326: PetscCall(VecRestoreArrayWrite(v, &a));
2327: PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2328: PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2329: PetscCall(VecDestroy(&diagV));
2330: PetscCall(VecDestroy(&offdiagV));
2331: PetscCall(PetscFree2(diagIdx, offdiagIdx));
2332: PetscFunctionReturn(PETSC_SUCCESS);
2333: }
2335: static PetscErrorCode MatGetRowMin_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2336: {
2337: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
2338: PetscInt m = A->rmap->n, n = A->cmap->n;
2339: PetscInt cstart = A->cmap->rstart, cend = A->cmap->rend;
2340: PetscInt *cmap = mat->garray;
2341: PetscInt *diagIdx, *offdiagIdx;
2342: Vec diagV, offdiagV;
2343: PetscScalar *a, *diagA, *offdiagA;
2344: const PetscScalar *ba, *bav;
2345: PetscInt r, j, col, ncols, *bi, *bj;
2346: Mat B = mat->B;
2347: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2349: PetscFunctionBegin;
2350: /* When a process holds entire A and other processes have no entry */
2351: if (A->cmap->N == n) {
2352: PetscCall(VecGetArrayWrite(v, &diagA));
2353: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2354: PetscCall(MatGetRowMin(mat->A, diagV, idx));
2355: PetscCall(VecDestroy(&diagV));
2356: PetscCall(VecRestoreArrayWrite(v, &diagA));
2357: PetscFunctionReturn(PETSC_SUCCESS);
2358: } else if (n == 0) {
2359: if (m) {
2360: PetscCall(VecGetArrayWrite(v, &a));
2361: for (r = 0; r < m; r++) {
2362: a[r] = PETSC_MAX_REAL;
2363: if (idx) idx[r] = -1;
2364: }
2365: PetscCall(VecRestoreArrayWrite(v, &a));
2366: }
2367: PetscFunctionReturn(PETSC_SUCCESS);
2368: }
2370: PetscCall(PetscCalloc2(m, &diagIdx, m, &offdiagIdx));
2371: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2372: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2373: PetscCall(MatGetRowMin(mat->A, diagV, diagIdx));
2375: /* Get offdiagIdx[] for implicit 0.0 */
2376: PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2377: ba = bav;
2378: bi = b->i;
2379: bj = b->j;
2380: PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2381: for (r = 0; r < m; r++) {
2382: ncols = bi[r + 1] - bi[r];
2383: if (ncols == A->cmap->N - n) { /* Brow is dense */
2384: offdiagA[r] = *ba;
2385: offdiagIdx[r] = cmap[0];
2386: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2387: offdiagA[r] = 0.0;
2389: /* Find first hole in the cmap */
2390: for (j = 0; j < ncols; j++) {
2391: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2392: if (col > j && j < cstart) {
2393: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2394: break;
2395: } else if (col > j + n && j >= cstart) {
2396: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2397: break;
2398: }
2399: }
2400: if (j == ncols && ncols < A->cmap->N - n) {
2401: /* a hole is outside compressed Bcols */
2402: if (ncols == 0) {
2403: if (cstart) {
2404: offdiagIdx[r] = 0;
2405: } else offdiagIdx[r] = cend;
2406: } else { /* ncols > 0 */
2407: offdiagIdx[r] = cmap[ncols - 1] + 1;
2408: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2409: }
2410: }
2411: }
2413: for (j = 0; j < ncols; j++) {
2414: if (PetscRealPart(offdiagA[r]) > PetscRealPart(*ba)) {
2415: offdiagA[r] = *ba;
2416: offdiagIdx[r] = cmap[*bj];
2417: }
2418: ba++;
2419: bj++;
2420: }
2421: }
2423: PetscCall(VecGetArrayWrite(v, &a));
2424: PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2425: for (r = 0; r < m; ++r) {
2426: if (PetscRealPart(diagA[r]) < PetscRealPart(offdiagA[r])) {
2427: a[r] = diagA[r];
2428: if (idx) idx[r] = cstart + diagIdx[r];
2429: } else if (PetscRealPart(diagA[r]) == PetscRealPart(offdiagA[r])) {
2430: a[r] = diagA[r];
2431: if (idx) {
2432: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2433: idx[r] = cstart + diagIdx[r];
2434: } else idx[r] = offdiagIdx[r];
2435: }
2436: } else {
2437: a[r] = offdiagA[r];
2438: if (idx) idx[r] = offdiagIdx[r];
2439: }
2440: }
2441: PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2442: PetscCall(VecRestoreArrayWrite(v, &a));
2443: PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2444: PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2445: PetscCall(VecDestroy(&diagV));
2446: PetscCall(VecDestroy(&offdiagV));
2447: PetscCall(PetscFree2(diagIdx, offdiagIdx));
2448: PetscFunctionReturn(PETSC_SUCCESS);
2449: }
2451: static PetscErrorCode MatGetRowMax_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2452: {
2453: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
2454: PetscInt m = A->rmap->n, n = A->cmap->n;
2455: PetscInt cstart = A->cmap->rstart, cend = A->cmap->rend;
2456: PetscInt *cmap = mat->garray;
2457: PetscInt *diagIdx, *offdiagIdx;
2458: Vec diagV, offdiagV;
2459: PetscScalar *a, *diagA, *offdiagA;
2460: const PetscScalar *ba, *bav;
2461: PetscInt r, j, col, ncols, *bi, *bj;
2462: Mat B = mat->B;
2463: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2465: PetscFunctionBegin;
2466: /* When a process holds entire A and other processes have no entry */
2467: if (A->cmap->N == n) {
2468: PetscCall(VecGetArrayWrite(v, &diagA));
2469: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2470: PetscCall(MatGetRowMax(mat->A, diagV, idx));
2471: PetscCall(VecDestroy(&diagV));
2472: PetscCall(VecRestoreArrayWrite(v, &diagA));
2473: PetscFunctionReturn(PETSC_SUCCESS);
2474: } else if (n == 0) {
2475: if (m) {
2476: PetscCall(VecGetArrayWrite(v, &a));
2477: for (r = 0; r < m; r++) {
2478: a[r] = PETSC_MIN_REAL;
2479: if (idx) idx[r] = -1;
2480: }
2481: PetscCall(VecRestoreArrayWrite(v, &a));
2482: }
2483: PetscFunctionReturn(PETSC_SUCCESS);
2484: }
2486: PetscCall(PetscMalloc2(m, &diagIdx, m, &offdiagIdx));
2487: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2488: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2489: PetscCall(MatGetRowMax(mat->A, diagV, diagIdx));
2491: /* Get offdiagIdx[] for implicit 0.0 */
2492: PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2493: ba = bav;
2494: bi = b->i;
2495: bj = b->j;
2496: PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2497: for (r = 0; r < m; r++) {
2498: ncols = bi[r + 1] - bi[r];
2499: if (ncols == A->cmap->N - n) { /* Brow is dense */
2500: offdiagA[r] = *ba;
2501: offdiagIdx[r] = cmap[0];
2502: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2503: offdiagA[r] = 0.0;
2505: /* Find first hole in the cmap */
2506: for (j = 0; j < ncols; j++) {
2507: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2508: if (col > j && j < cstart) {
2509: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2510: break;
2511: } else if (col > j + n && j >= cstart) {
2512: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2513: break;
2514: }
2515: }
2516: if (j == ncols && ncols < A->cmap->N - n) {
2517: /* a hole is outside compressed Bcols */
2518: if (ncols == 0) {
2519: if (cstart) {
2520: offdiagIdx[r] = 0;
2521: } else offdiagIdx[r] = cend;
2522: } else { /* ncols > 0 */
2523: offdiagIdx[r] = cmap[ncols - 1] + 1;
2524: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2525: }
2526: }
2527: }
2529: for (j = 0; j < ncols; j++) {
2530: if (PetscRealPart(offdiagA[r]) < PetscRealPart(*ba)) {
2531: offdiagA[r] = *ba;
2532: offdiagIdx[r] = cmap[*bj];
2533: }
2534: ba++;
2535: bj++;
2536: }
2537: }
2539: PetscCall(VecGetArrayWrite(v, &a));
2540: PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2541: for (r = 0; r < m; ++r) {
2542: if (PetscRealPart(diagA[r]) > PetscRealPart(offdiagA[r])) {
2543: a[r] = diagA[r];
2544: if (idx) idx[r] = cstart + diagIdx[r];
2545: } else if (PetscRealPart(diagA[r]) == PetscRealPart(offdiagA[r])) {
2546: a[r] = diagA[r];
2547: if (idx) {
2548: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2549: idx[r] = cstart + diagIdx[r];
2550: } else idx[r] = offdiagIdx[r];
2551: }
2552: } else {
2553: a[r] = offdiagA[r];
2554: if (idx) idx[r] = offdiagIdx[r];
2555: }
2556: }
2557: PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2558: PetscCall(VecRestoreArrayWrite(v, &a));
2559: PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2560: PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2561: PetscCall(VecDestroy(&diagV));
2562: PetscCall(VecDestroy(&offdiagV));
2563: PetscCall(PetscFree2(diagIdx, offdiagIdx));
2564: PetscFunctionReturn(PETSC_SUCCESS);
2565: }
2567: PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat mat, Mat *newmat)
2568: {
2569: Mat *dummy;
2571: PetscFunctionBegin;
2572: PetscCall(MatCreateSubMatrix_MPIAIJ_All(mat, MAT_DO_NOT_GET_VALUES, MAT_INITIAL_MATRIX, &dummy));
2573: *newmat = *dummy;
2574: PetscCall(PetscFree(dummy));
2575: PetscFunctionReturn(PETSC_SUCCESS);
2576: }
2578: static PetscErrorCode MatInvertBlockDiagonal_MPIAIJ(Mat A, const PetscScalar **values)
2579: {
2580: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2582: PetscFunctionBegin;
2583: PetscCall(MatInvertBlockDiagonal(a->A, values));
2584: A->factorerrortype = a->A->factorerrortype;
2585: PetscFunctionReturn(PETSC_SUCCESS);
2586: }
2588: static PetscErrorCode MatSetRandom_MPIAIJ(Mat x, PetscRandom rctx)
2589: {
2590: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)x->data;
2592: PetscFunctionBegin;
2593: PetscCheck(x->assembled || x->preallocated, PetscObjectComm((PetscObject)x), PETSC_ERR_ARG_WRONGSTATE, "MatSetRandom on an unassembled and unpreallocated MATMPIAIJ is not allowed");
2594: PetscCall(MatSetRandom(aij->A, rctx));
2595: if (x->assembled) {
2596: PetscCall(MatSetRandom(aij->B, rctx));
2597: } else {
2598: PetscCall(MatSetRandomSkipColumnRange_SeqAIJ_Private(aij->B, x->cmap->rstart, x->cmap->rend, rctx));
2599: }
2600: PetscCall(MatAssemblyBegin(x, MAT_FINAL_ASSEMBLY));
2601: PetscCall(MatAssemblyEnd(x, MAT_FINAL_ASSEMBLY));
2602: PetscFunctionReturn(PETSC_SUCCESS);
2603: }
2605: static PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ(Mat A, PetscBool sc)
2606: {
2607: PetscFunctionBegin;
2608: if (sc) A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ_Scalable;
2609: else A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ;
2610: PetscFunctionReturn(PETSC_SUCCESS);
2611: }
2613: /*@
2614: MatMPIAIJGetNumberNonzeros - gets the number of nonzeros in the matrix on this MPI rank
2616: Not Collective
2618: Input Parameter:
2619: . A - the matrix
2621: Output Parameter:
2622: . nz - the number of nonzeros
2624: Level: advanced
2626: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`
2627: @*/
2628: PetscErrorCode MatMPIAIJGetNumberNonzeros(Mat A, PetscCount *nz)
2629: {
2630: Mat_MPIAIJ *maij = (Mat_MPIAIJ *)A->data;
2631: Mat_SeqAIJ *aaij = (Mat_SeqAIJ *)maij->A->data, *baij = (Mat_SeqAIJ *)maij->B->data;
2632: PetscBool isaij;
2634: PetscFunctionBegin;
2635: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATMPIAIJ, &isaij));
2636: PetscCheck(isaij, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Not for type %s", ((PetscObject)A)->type_name);
2637: *nz = aaij->i[A->rmap->n] + baij->i[A->rmap->n];
2638: PetscFunctionReturn(PETSC_SUCCESS);
2639: }
2641: /*@
2642: MatMPIAIJSetUseScalableIncreaseOverlap - Determine if the matrix uses a scalable algorithm to compute the overlap
2644: Collective
2646: Input Parameters:
2647: + A - the matrix
2648: - sc - `PETSC_TRUE` indicates use the scalable algorithm (default is not to use the scalable algorithm)
2650: Level: advanced
2652: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`
2653: @*/
2654: PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap(Mat A, PetscBool sc)
2655: {
2656: PetscFunctionBegin;
2657: PetscTryMethod(A, "MatMPIAIJSetUseScalableIncreaseOverlap_C", (Mat, PetscBool), (A, sc));
2658: PetscFunctionReturn(PETSC_SUCCESS);
2659: }
2661: PetscErrorCode MatSetFromOptions_MPIAIJ(Mat A, PetscOptionItems PetscOptionsObject)
2662: {
2663: PetscBool sc = PETSC_FALSE, flg;
2665: PetscFunctionBegin;
2666: PetscOptionsHeadBegin(PetscOptionsObject, "MPIAIJ options");
2667: if (A->ops->increaseoverlap == MatIncreaseOverlap_MPIAIJ_Scalable) sc = PETSC_TRUE;
2668: PetscCall(PetscOptionsBool("-mat_increase_overlap_scalable", "Use a scalable algorithm to compute the overlap", "MatIncreaseOverlap", sc, &sc, &flg));
2669: if (flg) PetscCall(MatMPIAIJSetUseScalableIncreaseOverlap(A, sc));
2670: PetscOptionsHeadEnd();
2671: PetscFunctionReturn(PETSC_SUCCESS);
2672: }
2674: static PetscErrorCode MatShift_MPIAIJ(Mat Y, PetscScalar a)
2675: {
2676: Mat_MPIAIJ *maij = (Mat_MPIAIJ *)Y->data;
2677: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)maij->A->data;
2679: PetscFunctionBegin;
2680: if (!Y->preallocated) {
2681: PetscCall(MatMPIAIJSetPreallocation(Y, 1, NULL, 0, NULL));
2682: } else if (!aij->nz) { /* It does not matter if diagonals of Y only partially lie in maij->A. We just need an estimated preallocation. */
2683: PetscInt nonew = aij->nonew;
2684: PetscCall(MatSeqAIJSetPreallocation(maij->A, 1, NULL));
2685: aij->nonew = nonew;
2686: }
2687: PetscCall(MatShift_Basic(Y, a));
2688: PetscFunctionReturn(PETSC_SUCCESS);
2689: }
2691: static PetscErrorCode MatInvertVariableBlockDiagonal_MPIAIJ(Mat A, PetscInt nblocks, const PetscInt *bsizes, PetscScalar *diag)
2692: {
2693: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2695: PetscFunctionBegin;
2696: PetscCall(MatInvertVariableBlockDiagonal(a->A, nblocks, bsizes, diag));
2697: PetscFunctionReturn(PETSC_SUCCESS);
2698: }
2700: static PetscErrorCode MatEliminateZeros_MPIAIJ(Mat A, PetscBool keep)
2701: {
2702: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2704: PetscFunctionBegin;
2705: PetscCall(MatEliminateZeros_SeqAIJ(a->A, keep)); // possibly keep zero diagonal coefficients
2706: PetscCall(MatEliminateZeros_SeqAIJ(a->B, PETSC_FALSE)); // never keep zero diagonal coefficients
2707: PetscFunctionReturn(PETSC_SUCCESS);
2708: }
2710: static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ,
2711: MatGetRow_MPIAIJ,
2712: MatRestoreRow_MPIAIJ,
2713: MatMult_MPIAIJ,
2714: /* 4*/ MatMultAdd_MPIAIJ,
2715: MatMultTranspose_MPIAIJ,
2716: MatMultTransposeAdd_MPIAIJ,
2717: NULL,
2718: NULL,
2719: NULL,
2720: /*10*/ NULL,
2721: NULL,
2722: NULL,
2723: MatSOR_MPIAIJ,
2724: MatTranspose_MPIAIJ,
2725: /*15*/ MatGetInfo_MPIAIJ,
2726: MatEqual_MPIAIJ,
2727: MatGetDiagonal_MPIAIJ,
2728: MatDiagonalScale_MPIAIJ,
2729: MatNorm_MPIAIJ,
2730: /*20*/ MatAssemblyBegin_MPIAIJ,
2731: MatAssemblyEnd_MPIAIJ,
2732: MatSetOption_MPIAIJ,
2733: MatZeroEntries_MPIAIJ,
2734: /*24*/ MatZeroRows_MPIAIJ,
2735: NULL,
2736: NULL,
2737: NULL,
2738: NULL,
2739: /*29*/ MatSetUp_MPI_Hash,
2740: NULL,
2741: NULL,
2742: MatGetDiagonalBlock_MPIAIJ,
2743: NULL,
2744: /*34*/ MatDuplicate_MPIAIJ,
2745: NULL,
2746: NULL,
2747: NULL,
2748: NULL,
2749: /*39*/ MatAXPY_MPIAIJ,
2750: MatCreateSubMatrices_MPIAIJ,
2751: MatIncreaseOverlap_MPIAIJ,
2752: MatGetValues_MPIAIJ,
2753: MatCopy_MPIAIJ,
2754: /*44*/ MatGetRowMax_MPIAIJ,
2755: MatScale_MPIAIJ,
2756: MatShift_MPIAIJ,
2757: MatDiagonalSet_MPIAIJ,
2758: MatZeroRowsColumns_MPIAIJ,
2759: /*49*/ MatSetRandom_MPIAIJ,
2760: MatGetRowIJ_MPIAIJ,
2761: MatRestoreRowIJ_MPIAIJ,
2762: NULL,
2763: NULL,
2764: /*54*/ MatFDColoringCreate_MPIXAIJ,
2765: NULL,
2766: MatSetUnfactored_MPIAIJ,
2767: MatPermute_MPIAIJ,
2768: NULL,
2769: /*59*/ MatCreateSubMatrix_MPIAIJ,
2770: MatDestroy_MPIAIJ,
2771: MatView_MPIAIJ,
2772: NULL,
2773: NULL,
2774: /*64*/ MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ,
2775: NULL,
2776: NULL,
2777: NULL,
2778: MatGetRowMaxAbs_MPIAIJ,
2779: /*69*/ MatGetRowMinAbs_MPIAIJ,
2780: NULL,
2781: NULL,
2782: MatFDColoringApply_AIJ,
2783: MatSetFromOptions_MPIAIJ,
2784: MatFindZeroDiagonals_MPIAIJ,
2785: /*75*/ NULL,
2786: NULL,
2787: NULL,
2788: MatLoad_MPIAIJ,
2789: NULL,
2790: /*80*/ NULL,
2791: NULL,
2792: NULL,
2793: /*83*/ NULL,
2794: NULL,
2795: MatMatMultNumeric_MPIAIJ_MPIAIJ,
2796: MatPtAPNumeric_MPIAIJ_MPIAIJ,
2797: NULL,
2798: NULL,
2799: /*89*/ MatBindToCPU_MPIAIJ,
2800: MatProductSetFromOptions_MPIAIJ,
2801: NULL,
2802: NULL,
2803: MatConjugate_MPIAIJ,
2804: /*94*/ NULL,
2805: MatSetValuesRow_MPIAIJ,
2806: MatRealPart_MPIAIJ,
2807: MatImaginaryPart_MPIAIJ,
2808: NULL,
2809: /*99*/ NULL,
2810: NULL,
2811: NULL,
2812: MatGetRowMin_MPIAIJ,
2813: NULL,
2814: /*104*/ MatGetSeqNonzeroStructure_MPIAIJ,
2815: NULL,
2816: MatGetGhosts_MPIAIJ,
2817: NULL,
2818: NULL,
2819: /*109*/ MatMultDiagonalBlock_MPIAIJ,
2820: NULL,
2821: NULL,
2822: NULL,
2823: MatGetMultiProcBlock_MPIAIJ,
2824: /*114*/ MatFindNonzeroRows_MPIAIJ,
2825: MatGetColumnReductions_MPIAIJ,
2826: MatInvertBlockDiagonal_MPIAIJ,
2827: MatInvertVariableBlockDiagonal_MPIAIJ,
2828: MatCreateSubMatricesMPI_MPIAIJ,
2829: /*119*/ NULL,
2830: NULL,
2831: MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ,
2832: NULL,
2833: NULL,
2834: /*124*/ NULL,
2835: NULL,
2836: MatSetBlockSizes_MPIAIJ,
2837: NULL,
2838: MatFDColoringSetUp_MPIXAIJ,
2839: /*129*/ MatFindOffBlockDiagonalEntries_MPIAIJ,
2840: MatCreateMPIMatConcatenateSeqMat_MPIAIJ,
2841: NULL,
2842: NULL,
2843: NULL,
2844: /*134*/ MatCreateGraph_Simple_AIJ,
2845: NULL,
2846: MatEliminateZeros_MPIAIJ,
2847: MatGetRowSumAbs_MPIAIJ,
2848: NULL,
2849: /*139*/ NULL,
2850: NULL,
2851: MatCopyHashToXAIJ_MPI_Hash,
2852: MatGetCurrentMemType_MPIAIJ,
2853: NULL};
2855: static PetscErrorCode MatStoreValues_MPIAIJ(Mat mat)
2856: {
2857: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2859: PetscFunctionBegin;
2860: PetscCall(MatStoreValues(aij->A));
2861: PetscCall(MatStoreValues(aij->B));
2862: PetscFunctionReturn(PETSC_SUCCESS);
2863: }
2865: static PetscErrorCode MatRetrieveValues_MPIAIJ(Mat mat)
2866: {
2867: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2869: PetscFunctionBegin;
2870: PetscCall(MatRetrieveValues(aij->A));
2871: PetscCall(MatRetrieveValues(aij->B));
2872: PetscFunctionReturn(PETSC_SUCCESS);
2873: }
2875: PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat B, PetscInt d_nz, const PetscInt d_nnz[], PetscInt o_nz, const PetscInt o_nnz[])
2876: {
2877: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2878: PetscMPIInt size;
2880: PetscFunctionBegin;
2881: if (B->hash_active) {
2882: B->ops[0] = b->cops;
2883: B->hash_active = PETSC_FALSE;
2884: }
2885: PetscCall(PetscLayoutSetUp(B->rmap));
2886: PetscCall(PetscLayoutSetUp(B->cmap));
2888: #if defined(PETSC_USE_CTABLE)
2889: PetscCall(PetscHMapIDestroy(&b->colmap));
2890: #else
2891: PetscCall(PetscFree(b->colmap));
2892: #endif
2893: PetscCall(PetscFree(b->garray));
2894: PetscCall(VecDestroy(&b->lvec));
2895: PetscCall(VecScatterDestroy(&b->Mvctx));
2897: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
2899: MatSeqXAIJGetOptions_Private(b->B);
2900: PetscCall(MatDestroy(&b->B));
2901: PetscCall(MatCreate(PETSC_COMM_SELF, &b->B));
2902: PetscCall(MatSetSizes(b->B, B->rmap->n, size > 1 ? B->cmap->N : 0, B->rmap->n, size > 1 ? B->cmap->N : 0));
2903: PetscCall(MatSetBlockSizesFromMats(b->B, B, B));
2904: PetscCall(MatSetType(b->B, MATSEQAIJ));
2905: MatSeqXAIJRestoreOptions_Private(b->B);
2907: MatSeqXAIJGetOptions_Private(b->A);
2908: PetscCall(MatDestroy(&b->A));
2909: PetscCall(MatCreate(PETSC_COMM_SELF, &b->A));
2910: PetscCall(MatSetSizes(b->A, B->rmap->n, B->cmap->n, B->rmap->n, B->cmap->n));
2911: PetscCall(MatSetBlockSizesFromMats(b->A, B, B));
2912: PetscCall(MatSetType(b->A, MATSEQAIJ));
2913: MatSeqXAIJRestoreOptions_Private(b->A);
2915: PetscCall(MatSeqAIJSetPreallocation(b->A, d_nz, d_nnz));
2916: PetscCall(MatSeqAIJSetPreallocation(b->B, o_nz, o_nnz));
2917: B->preallocated = PETSC_TRUE;
2918: B->was_assembled = PETSC_FALSE;
2919: B->assembled = PETSC_FALSE;
2920: PetscFunctionReturn(PETSC_SUCCESS);
2921: }
2923: static PetscErrorCode MatResetPreallocation_MPIAIJ(Mat B)
2924: {
2925: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2926: PetscBool ondiagreset, offdiagreset, memoryreset;
2928: PetscFunctionBegin;
2930: 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()");
2931: if (B->num_ass == 0) PetscFunctionReturn(PETSC_SUCCESS);
2933: PetscCall(MatResetPreallocation_SeqAIJ_Private(b->A, &ondiagreset));
2934: PetscCall(MatResetPreallocation_SeqAIJ_Private(b->B, &offdiagreset));
2935: memoryreset = (PetscBool)(ondiagreset || offdiagreset);
2936: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &memoryreset, 1, MPI_C_BOOL, MPI_LOR, PetscObjectComm((PetscObject)B)));
2937: if (!memoryreset) PetscFunctionReturn(PETSC_SUCCESS);
2939: PetscCall(PetscLayoutSetUp(B->rmap));
2940: PetscCall(PetscLayoutSetUp(B->cmap));
2941: 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");
2942: PetscCall(MatDisAssemble_MPIAIJ(B, PETSC_TRUE));
2943: PetscCall(VecScatterDestroy(&b->Mvctx));
2945: B->preallocated = PETSC_TRUE;
2946: B->was_assembled = PETSC_FALSE;
2947: B->assembled = PETSC_FALSE;
2948: /* Log that the state of this object has changed; this will help guarantee that preconditioners get re-setup */
2949: PetscCall(PetscObjectStateIncrease((PetscObject)B));
2950: PetscFunctionReturn(PETSC_SUCCESS);
2951: }
2953: PetscErrorCode MatDuplicate_MPIAIJ(Mat matin, MatDuplicateOption cpvalues, Mat *newmat)
2954: {
2955: Mat mat;
2956: Mat_MPIAIJ *a, *oldmat = (Mat_MPIAIJ *)matin->data;
2958: PetscFunctionBegin;
2959: *newmat = NULL;
2960: PetscCall(MatCreate(PetscObjectComm((PetscObject)matin), &mat));
2961: PetscCall(MatSetSizes(mat, matin->rmap->n, matin->cmap->n, matin->rmap->N, matin->cmap->N));
2962: PetscCall(MatSetBlockSizesFromMats(mat, matin, matin));
2963: PetscCall(MatSetType(mat, ((PetscObject)matin)->type_name));
2964: a = (Mat_MPIAIJ *)mat->data;
2966: mat->factortype = matin->factortype;
2967: mat->assembled = matin->assembled;
2968: mat->insertmode = NOT_SET_VALUES;
2970: a->size = oldmat->size;
2971: a->rank = oldmat->rank;
2972: a->donotstash = oldmat->donotstash;
2973: a->roworiented = oldmat->roworiented;
2974: a->rowindices = NULL;
2975: a->rowvalues = NULL;
2976: a->getrowactive = PETSC_FALSE;
2978: PetscCall(PetscLayoutReference(matin->rmap, &mat->rmap));
2979: PetscCall(PetscLayoutReference(matin->cmap, &mat->cmap));
2980: if (matin->hash_active) {
2981: PetscCall(MatSetUp(mat));
2982: } else {
2983: mat->preallocated = matin->preallocated;
2984: if (oldmat->colmap) {
2985: #if defined(PETSC_USE_CTABLE)
2986: PetscCall(PetscHMapIDuplicate(oldmat->colmap, &a->colmap));
2987: #else
2988: PetscCall(PetscMalloc1(mat->cmap->N, &a->colmap));
2989: PetscCall(PetscArraycpy(a->colmap, oldmat->colmap, mat->cmap->N));
2990: #endif
2991: } else a->colmap = NULL;
2992: if (oldmat->garray) {
2993: PetscInt len;
2994: len = oldmat->B->cmap->n;
2995: PetscCall(PetscMalloc1(len, &a->garray));
2996: if (len) PetscCall(PetscArraycpy(a->garray, oldmat->garray, len));
2997: } else a->garray = NULL;
2999: /* It may happen MatDuplicate is called with a non-assembled matrix
3000: In fact, MatDuplicate only requires the matrix to be preallocated
3001: This may happen inside a DMCreateMatrix_Shell */
3002: if (oldmat->lvec) PetscCall(VecDuplicate(oldmat->lvec, &a->lvec));
3003: if (oldmat->Mvctx) {
3004: a->Mvctx = oldmat->Mvctx;
3005: PetscCall(PetscObjectReference((PetscObject)oldmat->Mvctx));
3006: }
3007: PetscCall(MatDuplicate(oldmat->A, cpvalues, &a->A));
3008: PetscCall(MatDuplicate(oldmat->B, cpvalues, &a->B));
3009: }
3010: PetscCall(PetscFunctionListDuplicate(((PetscObject)matin)->qlist, &((PetscObject)mat)->qlist));
3011: *newmat = mat;
3012: PetscFunctionReturn(PETSC_SUCCESS);
3013: }
3015: PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer)
3016: {
3017: PetscBool isbinary, ishdf5;
3019: PetscFunctionBegin;
3022: /* force binary viewer to load .info file if it has not yet done so */
3023: PetscCall(PetscViewerSetUp(viewer));
3024: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
3025: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERHDF5, &ishdf5));
3026: if (isbinary) {
3027: PetscCall(MatLoad_MPIAIJ_Binary(newMat, viewer));
3028: } else if (ishdf5) {
3029: #if defined(PETSC_HAVE_HDF5)
3030: PetscCall(MatLoad_AIJ_HDF5(newMat, viewer));
3031: #else
3032: SETERRQ(PetscObjectComm((PetscObject)newMat), PETSC_ERR_SUP, "HDF5 not supported in this build.\nPlease reconfigure using --download-hdf5");
3033: #endif
3034: } else {
3035: 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);
3036: }
3037: PetscFunctionReturn(PETSC_SUCCESS);
3038: }
3040: PetscErrorCode MatLoad_MPIAIJ_Binary(Mat mat, PetscViewer viewer)
3041: {
3042: PetscInt header[4], M, N, m, nz, rows, cols, sum, i;
3043: PetscInt *rowidxs, *colidxs;
3044: PetscScalar *matvals;
3046: PetscFunctionBegin;
3047: PetscCall(PetscViewerSetUp(viewer));
3049: /* read in matrix header */
3050: PetscCall(PetscViewerBinaryRead(viewer, header, 4, NULL, PETSC_INT));
3051: PetscCheck(header[0] == MAT_FILE_CLASSID, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Not a matrix object in file");
3052: M = header[1];
3053: N = header[2];
3054: nz = header[3];
3055: PetscCheck(M >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix row size (%" PetscInt_FMT ") in file is negative", M);
3056: PetscCheck(N >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix column size (%" PetscInt_FMT ") in file is negative", N);
3057: PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Matrix stored in special format on disk, cannot load as MPIAIJ");
3059: /* set block sizes from the viewer's .info file */
3060: PetscCall(MatLoad_Binary_BlockSizes(mat, viewer));
3061: /* set global sizes if not set already */
3062: if (mat->rmap->N < 0) mat->rmap->N = M;
3063: if (mat->cmap->N < 0) mat->cmap->N = N;
3064: PetscCall(PetscLayoutSetUp(mat->rmap));
3065: PetscCall(PetscLayoutSetUp(mat->cmap));
3067: /* check if the matrix sizes are correct */
3068: PetscCall(MatGetSize(mat, &rows, &cols));
3069: 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);
3071: /* read in row lengths and build row indices */
3072: PetscCall(MatGetLocalSize(mat, &m, NULL));
3073: PetscCall(PetscMalloc1(m + 1, &rowidxs));
3074: PetscCall(PetscViewerBinaryReadAll(viewer, rowidxs + 1, m, PETSC_DECIDE, M, PETSC_INT));
3075: rowidxs[0] = 0;
3076: for (i = 0; i < m; i++) rowidxs[i + 1] += rowidxs[i];
3077: if (nz != PETSC_INT_MAX) {
3078: PetscCallMPI(MPIU_Allreduce(&rowidxs[m], &sum, 1, MPIU_INT, MPI_SUM, PetscObjectComm((PetscObject)viewer)));
3079: 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);
3080: }
3082: /* read in column indices and matrix values */
3083: PetscCall(PetscMalloc2(rowidxs[m], &colidxs, rowidxs[m], &matvals));
3084: PetscCall(PetscViewerBinaryReadAll(viewer, colidxs, rowidxs[m], PETSC_DETERMINE, PETSC_DETERMINE, PETSC_INT));
3085: PetscCall(PetscViewerBinaryReadAll(viewer, matvals, rowidxs[m], PETSC_DETERMINE, PETSC_DETERMINE, PETSC_SCALAR));
3086: /* store matrix indices and values */
3087: PetscCall(MatMPIAIJSetPreallocationCSR(mat, rowidxs, colidxs, matvals));
3088: PetscCall(PetscFree(rowidxs));
3089: PetscCall(PetscFree2(colidxs, matvals));
3090: PetscFunctionReturn(PETSC_SUCCESS);
3091: }
3093: /* Not scalable because of ISAllGather() unless getting all columns. */
3094: static PetscErrorCode ISGetSeqIS_Private(Mat mat, IS iscol, IS *isseq)
3095: {
3096: IS iscol_local;
3097: PetscBool isstride;
3098: PetscMPIInt gisstride = 0;
3100: PetscFunctionBegin;
3101: /* check if we are grabbing all columns*/
3102: PetscCall(PetscObjectTypeCompare((PetscObject)iscol, ISSTRIDE, &isstride));
3104: if (isstride) {
3105: PetscInt start, len, mstart, mlen;
3106: PetscCall(ISStrideGetInfo(iscol, &start, NULL));
3107: PetscCall(ISGetLocalSize(iscol, &len));
3108: PetscCall(MatGetOwnershipRangeColumn(mat, &mstart, &mlen));
3109: if (mstart == start && mlen - mstart == len) gisstride = 1;
3110: }
3112: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &gisstride, 1, MPI_INT, MPI_MIN, PetscObjectComm((PetscObject)mat)));
3113: if (gisstride) {
3114: PetscInt N;
3115: PetscCall(MatGetSize(mat, NULL, &N));
3116: PetscCall(ISCreateStride(PETSC_COMM_SELF, N, 0, 1, &iscol_local));
3117: PetscCall(ISSetIdentity(iscol_local));
3118: PetscCall(PetscInfo(mat, "Optimizing for obtaining all columns of the matrix; skipping ISAllGather()\n"));
3119: } else {
3120: PetscInt cbs;
3121: PetscCall(ISGetBlockSize(iscol, &cbs));
3122: PetscCall(ISAllGather(iscol, &iscol_local));
3123: PetscCall(ISSetBlockSize(iscol_local, cbs));
3124: }
3126: *isseq = iscol_local;
3127: PetscFunctionReturn(PETSC_SUCCESS);
3128: }
3130: /*
3131: Used by MatCreateSubMatrix_MPIAIJ_SameRowColDist() to avoid ISAllGather() and global size of iscol_local
3132: (see MatCreateSubMatrix_MPIAIJ_nonscalable)
3134: Input Parameters:
3135: + mat - matrix
3136: . isrow - parallel row index set; its local indices are a subset of local columns of `mat`,
3137: i.e., mat->rstart <= isrow[i] < mat->rend
3138: - iscol - parallel column index set; its local indices are a subset of local columns of `mat`,
3139: i.e., mat->cstart <= iscol[i] < mat->cend
3141: Output Parameters:
3142: + isrow_d - sequential row index set for retrieving mat->A
3143: . iscol_d - sequential column index set for retrieving mat->A
3144: . iscol_o - sequential column index set for retrieving mat->B
3145: - garray - column map; garray[i] indicates global location of iscol_o[i] in `iscol`
3146: */
3147: static PetscErrorCode ISGetSeqIS_SameColDist_Private(Mat mat, IS isrow, IS iscol, IS *isrow_d, IS *iscol_d, IS *iscol_o, PetscInt *garray[])
3148: {
3149: Vec x, cmap;
3150: const PetscInt *is_idx;
3151: PetscScalar *xarray, *cmaparray;
3152: PetscInt ncols, isstart, *idx, m, rstart, *cmap1, count;
3153: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data;
3154: Mat B = a->B;
3155: Vec lvec = a->lvec, lcmap;
3156: PetscInt i, cstart, cend, Bn = B->cmap->N;
3157: MPI_Comm comm;
3158: VecScatter Mvctx = a->Mvctx;
3160: PetscFunctionBegin;
3161: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3162: PetscCall(ISGetLocalSize(iscol, &ncols));
3164: /* (1) iscol is a sub-column vector of mat, pad it with '-1.' to form a full vector x */
3165: PetscCall(MatCreateVecs(mat, &x, NULL));
3166: PetscCall(VecSet(x, -1.0));
3167: PetscCall(VecDuplicate(x, &cmap));
3168: PetscCall(VecSet(cmap, -1.0));
3170: /* Get start indices */
3171: PetscCallMPI(MPI_Scan(&ncols, &isstart, 1, MPIU_INT, MPI_SUM, comm));
3172: isstart -= ncols;
3173: PetscCall(MatGetOwnershipRangeColumn(mat, &cstart, &cend));
3175: PetscCall(ISGetIndices(iscol, &is_idx));
3176: PetscCall(VecGetArray(x, &xarray));
3177: PetscCall(VecGetArray(cmap, &cmaparray));
3178: PetscCall(PetscMalloc1(ncols, &idx));
3179: for (i = 0; i < ncols; i++) {
3180: xarray[is_idx[i] - cstart] = (PetscScalar)is_idx[i];
3181: cmaparray[is_idx[i] - cstart] = i + isstart; /* global index of iscol[i] */
3182: idx[i] = is_idx[i] - cstart; /* local index of iscol[i] */
3183: }
3184: PetscCall(VecRestoreArray(x, &xarray));
3185: PetscCall(VecRestoreArray(cmap, &cmaparray));
3186: PetscCall(ISRestoreIndices(iscol, &is_idx));
3188: /* Get iscol_d */
3189: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, iscol_d));
3190: PetscCall(ISGetBlockSize(iscol, &i));
3191: PetscCall(ISSetBlockSize(*iscol_d, i));
3193: /* Get isrow_d */
3194: PetscCall(ISGetLocalSize(isrow, &m));
3195: rstart = mat->rmap->rstart;
3196: PetscCall(PetscMalloc1(m, &idx));
3197: PetscCall(ISGetIndices(isrow, &is_idx));
3198: for (i = 0; i < m; i++) idx[i] = is_idx[i] - rstart;
3199: PetscCall(ISRestoreIndices(isrow, &is_idx));
3201: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, m, idx, PETSC_OWN_POINTER, isrow_d));
3202: PetscCall(ISGetBlockSize(isrow, &i));
3203: PetscCall(ISSetBlockSize(*isrow_d, i));
3205: /* (2) Scatter x and cmap using aij->Mvctx to get their off-process portions (see MatMult_MPIAIJ) */
3206: PetscCall(VecScatterBegin(Mvctx, x, lvec, INSERT_VALUES, SCATTER_FORWARD));
3207: PetscCall(VecScatterEnd(Mvctx, x, lvec, INSERT_VALUES, SCATTER_FORWARD));
3209: PetscCall(VecDuplicate(lvec, &lcmap));
3211: PetscCall(VecScatterBegin(Mvctx, cmap, lcmap, INSERT_VALUES, SCATTER_FORWARD));
3212: PetscCall(VecScatterEnd(Mvctx, cmap, lcmap, INSERT_VALUES, SCATTER_FORWARD));
3214: /* (3) create sequential iscol_o (a subset of iscol) and isgarray */
3215: /* off-process column indices */
3216: count = 0;
3217: PetscCall(PetscMalloc1(Bn, &idx));
3218: PetscCall(PetscMalloc1(Bn, &cmap1));
3220: PetscCall(VecGetArray(lvec, &xarray));
3221: PetscCall(VecGetArray(lcmap, &cmaparray));
3222: for (i = 0; i < Bn; i++) {
3223: if (PetscRealPart(xarray[i]) > -1.0) {
3224: idx[count] = i; /* local column index in off-diagonal part B */
3225: cmap1[count] = (PetscInt)PetscRealPart(cmaparray[i]); /* column index in submat */
3226: count++;
3227: }
3228: }
3229: PetscCall(VecRestoreArray(lvec, &xarray));
3230: PetscCall(VecRestoreArray(lcmap, &cmaparray));
3232: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, count, idx, PETSC_COPY_VALUES, iscol_o));
3233: /* cannot ensure iscol_o has same blocksize as iscol! */
3235: PetscCall(PetscFree(idx));
3236: *garray = cmap1;
3238: PetscCall(VecDestroy(&x));
3239: PetscCall(VecDestroy(&cmap));
3240: PetscCall(VecDestroy(&lcmap));
3241: PetscFunctionReturn(PETSC_SUCCESS);
3242: }
3244: /* isrow and iscol have same processor distribution as mat, output *submat is a submatrix of local mat */
3245: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowColDist(Mat mat, IS isrow, IS iscol, MatReuse call, Mat *submat)
3246: {
3247: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data, *asub;
3248: Mat M = NULL;
3249: MPI_Comm comm;
3250: IS iscol_d, isrow_d, iscol_o;
3251: Mat Asub = NULL, Bsub = NULL;
3252: PetscInt n, count, M_size, N_size;
3254: PetscFunctionBegin;
3255: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3257: if (call == MAT_REUSE_MATRIX) {
3258: /* Retrieve isrow_d, iscol_d and iscol_o from submat */
3259: PetscCall(PetscObjectQuery((PetscObject)*submat, "isrow_d", (PetscObject *)&isrow_d));
3260: PetscCheck(isrow_d, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "isrow_d passed in was not used before, cannot reuse");
3262: PetscCall(PetscObjectQuery((PetscObject)*submat, "iscol_d", (PetscObject *)&iscol_d));
3263: PetscCheck(iscol_d, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "iscol_d passed in was not used before, cannot reuse");
3265: PetscCall(PetscObjectQuery((PetscObject)*submat, "iscol_o", (PetscObject *)&iscol_o));
3266: PetscCheck(iscol_o, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "iscol_o passed in was not used before, cannot reuse");
3268: /* Update diagonal and off-diagonal portions of submat */
3269: asub = (Mat_MPIAIJ *)(*submat)->data;
3270: PetscCall(MatCreateSubMatrix_SeqAIJ(a->A, isrow_d, iscol_d, PETSC_DECIDE, MAT_REUSE_MATRIX, &asub->A));
3271: PetscCall(ISGetLocalSize(iscol_o, &n));
3272: if (n) PetscCall(MatCreateSubMatrix_SeqAIJ(a->B, isrow_d, iscol_o, PETSC_DECIDE, MAT_REUSE_MATRIX, &asub->B));
3273: PetscCall(MatAssemblyBegin(*submat, MAT_FINAL_ASSEMBLY));
3274: PetscCall(MatAssemblyEnd(*submat, MAT_FINAL_ASSEMBLY));
3276: } else { /* call == MAT_INITIAL_MATRIX) */
3277: PetscInt *garray, *garray_compact;
3278: PetscInt BsubN;
3280: /* Create isrow_d, iscol_d, iscol_o and isgarray (replace isgarray with array?) */
3281: PetscCall(ISGetSeqIS_SameColDist_Private(mat, isrow, iscol, &isrow_d, &iscol_d, &iscol_o, &garray));
3283: /* Create local submatrices Asub and Bsub */
3284: PetscCall(MatCreateSubMatrix_SeqAIJ(a->A, isrow_d, iscol_d, PETSC_DECIDE, MAT_INITIAL_MATRIX, &Asub));
3285: PetscCall(MatCreateSubMatrix_SeqAIJ(a->B, isrow_d, iscol_o, PETSC_DECIDE, MAT_INITIAL_MATRIX, &Bsub));
3287: // Compact garray so its not of size Bn
3288: PetscCall(ISGetSize(iscol_o, &count));
3289: PetscCall(PetscMalloc1(count, &garray_compact));
3290: PetscCall(PetscArraycpy(garray_compact, garray, count));
3292: /* Create submatrix M */
3293: PetscCall(ISGetSize(isrow, &M_size));
3294: PetscCall(ISGetSize(iscol, &N_size));
3295: PetscCall(MatCreateMPIAIJWithSeqAIJ(comm, M_size, N_size, Asub, Bsub, garray_compact, &M));
3297: /* If Bsub has empty columns, compress iscol_o such that it will retrieve condensed Bsub from a->B during reuse */
3298: asub = (Mat_MPIAIJ *)M->data;
3300: PetscCall(ISGetLocalSize(iscol_o, &BsubN));
3301: n = asub->B->cmap->N;
3302: if (BsubN > n) {
3303: /* This case can be tested using ~petsc/src/tao/bound/tutorials/runplate2_3 */
3304: const PetscInt *idx;
3305: PetscInt i, j, *idx_new, *subgarray = asub->garray;
3306: PetscCall(PetscInfo(M, "submatrix Bn %" PetscInt_FMT " != BsubN %" PetscInt_FMT ", update iscol_o\n", n, BsubN));
3308: PetscCall(PetscMalloc1(n, &idx_new));
3309: j = 0;
3310: PetscCall(ISGetIndices(iscol_o, &idx));
3311: for (i = 0; i < n; i++) {
3312: if (j >= BsubN) break;
3313: while (subgarray[i] > garray[j]) j++;
3315: 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]);
3316: idx_new[i] = idx[j++];
3317: }
3318: PetscCall(ISRestoreIndices(iscol_o, &idx));
3320: PetscCall(ISDestroy(&iscol_o));
3321: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, n, idx_new, PETSC_OWN_POINTER, &iscol_o));
3323: } 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);
3325: PetscCall(PetscFree(garray));
3326: *submat = M;
3328: /* Save isrow_d, iscol_d and iscol_o used in processor for next request */
3329: PetscCall(PetscObjectCompose((PetscObject)M, "isrow_d", (PetscObject)isrow_d));
3330: PetscCall(ISDestroy(&isrow_d));
3332: PetscCall(PetscObjectCompose((PetscObject)M, "iscol_d", (PetscObject)iscol_d));
3333: PetscCall(ISDestroy(&iscol_d));
3335: PetscCall(PetscObjectCompose((PetscObject)M, "iscol_o", (PetscObject)iscol_o));
3336: PetscCall(ISDestroy(&iscol_o));
3337: }
3338: PetscFunctionReturn(PETSC_SUCCESS);
3339: }
3341: PetscErrorCode MatCreateSubMatrix_MPIAIJ(Mat mat, IS isrow, IS iscol, MatReuse call, Mat *newmat)
3342: {
3343: IS iscol_local = NULL, isrow_d;
3344: PetscInt csize;
3345: PetscInt n, i, j, start, end;
3346: PetscBool sameRowDist = PETSC_FALSE, sameDist[2], tsameDist[2];
3347: MPI_Comm comm;
3349: PetscFunctionBegin;
3350: /* If isrow has same processor distribution as mat,
3351: call MatCreateSubMatrix_MPIAIJ_SameRowDist() to avoid using a hash table with global size of iscol */
3352: if (call == MAT_REUSE_MATRIX) {
3353: PetscCall(PetscObjectQuery((PetscObject)*newmat, "isrow_d", (PetscObject *)&isrow_d));
3354: if (isrow_d) {
3355: sameRowDist = PETSC_TRUE;
3356: tsameDist[1] = PETSC_TRUE; /* sameColDist */
3357: } else {
3358: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_local));
3359: if (iscol_local) {
3360: sameRowDist = PETSC_TRUE;
3361: tsameDist[1] = PETSC_FALSE; /* !sameColDist */
3362: }
3363: }
3364: } else {
3365: /* Check if isrow has same processor distribution as mat */
3366: sameDist[0] = PETSC_FALSE;
3367: PetscCall(ISGetLocalSize(isrow, &n));
3368: if (!n) {
3369: sameDist[0] = PETSC_TRUE;
3370: } else {
3371: PetscCall(ISGetMinMax(isrow, &i, &j));
3372: PetscCall(MatGetOwnershipRange(mat, &start, &end));
3373: if (i >= start && j < end) sameDist[0] = PETSC_TRUE;
3374: }
3376: /* Check if iscol has same processor distribution as mat */
3377: sameDist[1] = PETSC_FALSE;
3378: PetscCall(ISGetLocalSize(iscol, &n));
3379: if (!n) {
3380: sameDist[1] = PETSC_TRUE;
3381: } else {
3382: PetscCall(ISGetMinMax(iscol, &i, &j));
3383: PetscCall(MatGetOwnershipRangeColumn(mat, &start, &end));
3384: if (i >= start && j < end) sameDist[1] = PETSC_TRUE;
3385: }
3387: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3388: PetscCallMPI(MPIU_Allreduce(&sameDist, &tsameDist, 2, MPI_C_BOOL, MPI_LAND, comm));
3389: sameRowDist = tsameDist[0];
3390: }
3392: if (sameRowDist) {
3393: if (tsameDist[1]) { /* sameRowDist & sameColDist */
3394: /* isrow and iscol have same processor distribution as mat */
3395: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowColDist(mat, isrow, iscol, call, newmat));
3396: PetscFunctionReturn(PETSC_SUCCESS);
3397: } else { /* sameRowDist */
3398: /* isrow has same processor distribution as mat */
3399: if (call == MAT_INITIAL_MATRIX) {
3400: PetscBool sorted;
3401: PetscCall(ISGetSeqIS_Private(mat, iscol, &iscol_local));
3402: PetscCall(ISGetLocalSize(iscol_local, &n)); /* local size of iscol_local = global columns of newmat */
3403: PetscCall(ISGetSize(iscol, &i));
3404: PetscCheck(n == i, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "n %" PetscInt_FMT " != size of iscol %" PetscInt_FMT, n, i);
3406: PetscCall(ISSorted(iscol_local, &sorted));
3407: if (sorted) {
3408: /* MatCreateSubMatrix_MPIAIJ_SameRowDist() requires iscol_local be sorted; it can have duplicate indices */
3409: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowDist(mat, isrow, iscol, iscol_local, MAT_INITIAL_MATRIX, newmat));
3410: PetscFunctionReturn(PETSC_SUCCESS);
3411: }
3412: } else { /* call == MAT_REUSE_MATRIX */
3413: IS iscol_sub;
3414: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_sub));
3415: if (iscol_sub) {
3416: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowDist(mat, isrow, iscol, NULL, call, newmat));
3417: PetscFunctionReturn(PETSC_SUCCESS);
3418: }
3419: }
3420: }
3421: }
3423: /* General case: iscol -> iscol_local which has global size of iscol */
3424: if (call == MAT_REUSE_MATRIX) {
3425: PetscCall(PetscObjectQuery((PetscObject)*newmat, "ISAllGather", (PetscObject *)&iscol_local));
3426: PetscCheck(iscol_local, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3427: } else {
3428: if (!iscol_local) PetscCall(ISGetSeqIS_Private(mat, iscol, &iscol_local));
3429: }
3431: PetscCall(ISGetLocalSize(iscol, &csize));
3432: PetscCall(MatCreateSubMatrix_MPIAIJ_nonscalable(mat, isrow, iscol_local, csize, call, newmat));
3434: if (call == MAT_INITIAL_MATRIX) {
3435: PetscCall(PetscObjectCompose((PetscObject)*newmat, "ISAllGather", (PetscObject)iscol_local));
3436: PetscCall(ISDestroy(&iscol_local));
3437: }
3438: PetscFunctionReturn(PETSC_SUCCESS);
3439: }
3441: /*@C
3442: MatCreateMPIAIJWithSeqAIJ - creates a `MATMPIAIJ` matrix using `MATSEQAIJ` matrices that contain the "diagonal"
3443: and "off-diagonal" part of the matrix in CSR format.
3445: Collective
3447: Input Parameters:
3448: + comm - MPI communicator
3449: . M - the global row size
3450: . N - the global column size
3451: . A - "diagonal" portion of matrix
3452: . 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
3453: - 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.
3455: Output Parameter:
3456: . mat - the matrix, with input `A` as its local diagonal matrix
3458: Level: advanced
3460: Notes:
3461: See `MatCreateAIJ()` for the definition of "diagonal" and "off-diagonal" portion of the matrix.
3463: `A` and `B` becomes part of output mat. The user cannot use `A` and `B` anymore.
3465: 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
3466: `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
3467: `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`
3468: yourself, see algorithms in the private function `MatSetUpMultiply_MPIAIJ()`.
3470: 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.
3472: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MATSEQAIJ`, `MatCreateMPIAIJWithSplitArrays()`
3473: @*/
3474: PetscErrorCode MatCreateMPIAIJWithSeqAIJ(MPI_Comm comm, PetscInt M, PetscInt N, Mat A, Mat B, PetscInt *garray, Mat *mat)
3475: {
3476: PetscInt m, n;
3477: MatType mpi_mat_type;
3478: Mat_MPIAIJ *mpiaij;
3479: Mat C;
3481: PetscFunctionBegin;
3482: PetscCall(MatCreate(comm, &C));
3483: PetscCall(MatGetSize(A, &m, &n));
3484: PetscCheck(m == B->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Am %" PetscInt_FMT " != Bm %" PetscInt_FMT, m, B->rmap->N);
3485: 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);
3487: PetscCall(MatSetSizes(C, m, n, M, N));
3488: /* Determine the type of MPI matrix that should be created from the type of matrix A, which holds the "diagonal" portion. */
3489: PetscCall(MatGetMPIMatType_Private(A, &mpi_mat_type));
3490: PetscCall(MatSetType(C, mpi_mat_type));
3491: if (!garray) {
3492: const PetscScalar *ba;
3494: B->nonzerostate++;
3495: PetscCall(MatSeqAIJGetArrayRead(B, &ba)); /* Since we will destroy B's device copy, we need to make sure the host copy is up to date */
3496: PetscCall(MatSeqAIJRestoreArrayRead(B, &ba));
3497: }
3499: PetscCall(MatSetBlockSizes(C, A->rmap->bs, A->cmap->bs));
3500: PetscCall(PetscLayoutSetUp(C->rmap));
3501: PetscCall(PetscLayoutSetUp(C->cmap));
3503: mpiaij = (Mat_MPIAIJ *)C->data;
3504: mpiaij->A = A;
3505: mpiaij->B = B;
3506: mpiaij->garray = garray;
3507: C->preallocated = PETSC_TRUE;
3508: C->nooffprocentries = PETSC_TRUE; /* See MatAssemblyBegin_MPIAIJ. In effect, making MatAssemblyBegin a nop */
3510: PetscCall(MatSetOption(C, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
3511: PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY));
3512: /* MatAssemblyEnd is critical here. It sets mat->offloadmask according to A and B's, and
3513: also gets mpiaij->B compacted (if garray is NULL), with its col ids and size reduced
3514: */
3515: PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY));
3516: PetscCall(MatSetOption(C, MAT_NO_OFF_PROC_ENTRIES, PETSC_FALSE));
3517: PetscCall(MatSetOption(C, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
3518: *mat = C;
3519: PetscFunctionReturn(PETSC_SUCCESS);
3520: }
3522: extern PetscErrorCode MatCreateSubMatrices_MPIAIJ_SingleIS_Local(Mat, PetscInt, const IS[], const IS[], MatReuse, PetscBool, Mat *);
3524: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowDist(Mat mat, IS isrow, IS iscol, IS iscol_local, MatReuse call, Mat *newmat)
3525: {
3526: PetscInt i, m, n, rstart, row, rend, nz, j, bs, cbs;
3527: PetscInt *ii, *jj, nlocal, *dlens, *olens, dlen, olen, jend, mglobal;
3528: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data;
3529: Mat M, Msub, B = a->B;
3530: MatScalar *aa;
3531: Mat_SeqAIJ *aij;
3532: PetscInt *garray = a->garray, *colsub, Ncols;
3533: PetscInt count, Bn = B->cmap->N, cstart = mat->cmap->rstart, cend = mat->cmap->rend;
3534: IS iscol_sub, iscmap;
3535: const PetscInt *is_idx, *cmap;
3536: PetscBool allcolumns = PETSC_FALSE;
3537: MPI_Comm comm;
3539: PetscFunctionBegin;
3540: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3541: if (call == MAT_REUSE_MATRIX) {
3542: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_sub));
3543: PetscCheck(iscol_sub, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "SubIScol passed in was not used before, cannot reuse");
3544: PetscCall(ISGetLocalSize(iscol_sub, &count));
3546: PetscCall(PetscObjectQuery((PetscObject)*newmat, "Subcmap", (PetscObject *)&iscmap));
3547: PetscCheck(iscmap, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Subcmap passed in was not used before, cannot reuse");
3549: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubMatrix", (PetscObject *)&Msub));
3550: PetscCheck(Msub, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3552: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol_sub, MAT_REUSE_MATRIX, PETSC_FALSE, &Msub));
3554: } else { /* call == MAT_INITIAL_MATRIX) */
3555: PetscBool flg;
3557: PetscCall(ISGetLocalSize(iscol, &n));
3558: PetscCall(ISGetSize(iscol, &Ncols));
3560: /* (1) iscol -> nonscalable iscol_local */
3561: /* Check for special case: each processor gets entire matrix columns */
3562: PetscCall(ISIdentity(iscol_local, &flg));
3563: if (flg && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3564: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &allcolumns, 1, MPI_C_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
3565: if (allcolumns) {
3566: iscol_sub = iscol_local;
3567: PetscCall(PetscObjectReference((PetscObject)iscol_local));
3568: PetscCall(ISCreateStride(PETSC_COMM_SELF, n, 0, 1, &iscmap));
3570: } else {
3571: /* (2) iscol_local -> iscol_sub and iscmap. Implementation below requires iscol_local be sorted, it can have duplicate indices */
3572: PetscInt *idx, *cmap1, k;
3573: PetscCall(PetscMalloc1(Ncols, &idx));
3574: PetscCall(PetscMalloc1(Ncols, &cmap1));
3575: PetscCall(ISGetIndices(iscol_local, &is_idx));
3576: count = 0;
3577: k = 0;
3578: for (i = 0; i < Ncols; i++) {
3579: j = is_idx[i];
3580: if (j >= cstart && j < cend) {
3581: /* diagonal part of mat */
3582: idx[count] = j;
3583: cmap1[count++] = i; /* column index in submat */
3584: } else if (Bn) {
3585: /* off-diagonal part of mat */
3586: if (j == garray[k]) {
3587: idx[count] = j;
3588: cmap1[count++] = i; /* column index in submat */
3589: } else if (j > garray[k]) {
3590: while (j > garray[k] && k < Bn - 1) k++;
3591: if (j == garray[k]) {
3592: idx[count] = j;
3593: cmap1[count++] = i; /* column index in submat */
3594: }
3595: }
3596: }
3597: }
3598: PetscCall(ISRestoreIndices(iscol_local, &is_idx));
3600: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, count, idx, PETSC_OWN_POINTER, &iscol_sub));
3601: PetscCall(ISGetBlockSize(iscol, &cbs));
3602: PetscCall(ISSetBlockSize(iscol_sub, cbs));
3604: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)iscol_local), count, cmap1, PETSC_OWN_POINTER, &iscmap));
3605: }
3607: /* (3) Create sequential Msub */
3608: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol_sub, MAT_INITIAL_MATRIX, allcolumns, &Msub));
3609: }
3611: PetscCall(ISGetLocalSize(iscol_sub, &count));
3612: aij = (Mat_SeqAIJ *)Msub->data;
3613: ii = aij->i;
3614: PetscCall(ISGetIndices(iscmap, &cmap));
3616: /*
3617: m - number of local rows
3618: Ncols - number of columns (same on all processors)
3619: rstart - first row in new global matrix generated
3620: */
3621: PetscCall(MatGetSize(Msub, &m, NULL));
3623: if (call == MAT_INITIAL_MATRIX) {
3624: /* (4) Create parallel newmat */
3625: PetscMPIInt rank, size;
3626: PetscInt csize;
3628: PetscCallMPI(MPI_Comm_size(comm, &size));
3629: PetscCallMPI(MPI_Comm_rank(comm, &rank));
3631: /*
3632: Determine the number of non-zeros in the diagonal and off-diagonal
3633: portions of the matrix in order to do correct preallocation
3634: */
3636: /* first get start and end of "diagonal" columns */
3637: PetscCall(ISGetLocalSize(iscol, &csize));
3638: if (csize == PETSC_DECIDE) {
3639: PetscCall(ISGetSize(isrow, &mglobal));
3640: if (mglobal == Ncols) { /* square matrix */
3641: nlocal = m;
3642: } else {
3643: nlocal = Ncols / size + ((Ncols % size) > rank);
3644: }
3645: } else {
3646: nlocal = csize;
3647: }
3648: PetscCallMPI(MPI_Scan(&nlocal, &rend, 1, MPIU_INT, MPI_SUM, comm));
3649: rstart = rend - nlocal;
3650: 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);
3652: /* next, compute all the lengths */
3653: jj = aij->j;
3654: PetscCall(PetscMalloc1(2 * m + 1, &dlens));
3655: olens = dlens + m;
3656: for (i = 0; i < m; i++) {
3657: jend = ii[i + 1] - ii[i];
3658: olen = 0;
3659: dlen = 0;
3660: for (j = 0; j < jend; j++) {
3661: if (cmap[*jj] < rstart || cmap[*jj] >= rend) olen++;
3662: else dlen++;
3663: jj++;
3664: }
3665: olens[i] = olen;
3666: dlens[i] = dlen;
3667: }
3669: PetscCall(ISGetBlockSize(isrow, &bs));
3670: PetscCall(ISGetBlockSize(iscol, &cbs));
3672: PetscCall(MatCreate(comm, &M));
3673: PetscCall(MatSetSizes(M, m, nlocal, PETSC_DECIDE, Ncols));
3674: PetscCall(MatSetBlockSizes(M, bs, cbs));
3675: PetscCall(MatSetType(M, ((PetscObject)mat)->type_name));
3676: PetscCall(MatMPIAIJSetPreallocation(M, 0, dlens, 0, olens));
3677: PetscCall(PetscFree(dlens));
3679: } else { /* call == MAT_REUSE_MATRIX */
3680: M = *newmat;
3681: PetscCall(MatGetLocalSize(M, &i, NULL));
3682: PetscCheck(i == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Previous matrix must be same size/layout as request");
3683: PetscCall(MatZeroEntries(M));
3684: /*
3685: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3686: rather than the slower MatSetValues().
3687: */
3688: M->was_assembled = PETSC_TRUE;
3689: M->assembled = PETSC_FALSE;
3690: }
3692: /* (5) Set values of Msub to *newmat */
3693: PetscCall(PetscMalloc1(count, &colsub));
3694: PetscCall(MatGetOwnershipRange(M, &rstart, NULL));
3696: jj = aij->j;
3697: PetscCall(MatSeqAIJGetArrayRead(Msub, (const PetscScalar **)&aa));
3698: for (i = 0; i < m; i++) {
3699: row = rstart + i;
3700: nz = ii[i + 1] - ii[i];
3701: for (j = 0; j < nz; j++) colsub[j] = cmap[jj[j]];
3702: PetscCall(MatSetValues_MPIAIJ(M, 1, &row, nz, colsub, aa, INSERT_VALUES));
3703: jj += nz;
3704: aa += nz;
3705: }
3706: PetscCall(MatSeqAIJRestoreArrayRead(Msub, (const PetscScalar **)&aa));
3707: PetscCall(ISRestoreIndices(iscmap, &cmap));
3709: PetscCall(MatAssemblyBegin(M, MAT_FINAL_ASSEMBLY));
3710: PetscCall(MatAssemblyEnd(M, MAT_FINAL_ASSEMBLY));
3712: PetscCall(PetscFree(colsub));
3714: /* save Msub, iscol_sub and iscmap used in processor for next request */
3715: if (call == MAT_INITIAL_MATRIX) {
3716: *newmat = M;
3717: PetscCall(PetscObjectCompose((PetscObject)*newmat, "SubMatrix", (PetscObject)Msub));
3718: PetscCall(MatDestroy(&Msub));
3720: PetscCall(PetscObjectCompose((PetscObject)*newmat, "SubIScol", (PetscObject)iscol_sub));
3721: PetscCall(ISDestroy(&iscol_sub));
3723: PetscCall(PetscObjectCompose((PetscObject)*newmat, "Subcmap", (PetscObject)iscmap));
3724: PetscCall(ISDestroy(&iscmap));
3726: if (iscol_local) {
3727: PetscCall(PetscObjectCompose((PetscObject)*newmat, "ISAllGather", (PetscObject)iscol_local));
3728: PetscCall(ISDestroy(&iscol_local));
3729: }
3730: }
3731: PetscFunctionReturn(PETSC_SUCCESS);
3732: }
3734: /*
3735: Not great since it makes two copies of the submatrix, first an SeqAIJ
3736: in local and then by concatenating the local matrices the end result.
3737: Writing it directly would be much like MatCreateSubMatrices_MPIAIJ()
3739: This requires a sequential iscol with all indices.
3740: */
3741: PetscErrorCode MatCreateSubMatrix_MPIAIJ_nonscalable(Mat mat, IS isrow, IS iscol, PetscInt csize, MatReuse call, Mat *newmat)
3742: {
3743: PetscMPIInt rank, size;
3744: PetscInt i, m, n, rstart, row, rend, nz, *cwork, j, bs, cbs;
3745: PetscInt *ii, *jj, nlocal, *dlens, *olens, dlen, olen, jend, mglobal;
3746: Mat M, Mreuse;
3747: MatScalar *aa, *vwork;
3748: MPI_Comm comm;
3749: Mat_SeqAIJ *aij;
3750: PetscBool colflag, allcolumns = PETSC_FALSE;
3752: PetscFunctionBegin;
3753: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3754: PetscCallMPI(MPI_Comm_rank(comm, &rank));
3755: PetscCallMPI(MPI_Comm_size(comm, &size));
3757: /* Check for special case: each processor gets entire matrix columns */
3758: PetscCall(ISIdentity(iscol, &colflag));
3759: PetscCall(ISGetLocalSize(iscol, &n));
3760: if (colflag && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3761: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &allcolumns, 1, MPI_C_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
3763: if (call == MAT_REUSE_MATRIX) {
3764: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubMatrix", (PetscObject *)&Mreuse));
3765: PetscCheck(Mreuse, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3766: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol, MAT_REUSE_MATRIX, allcolumns, &Mreuse));
3767: } else {
3768: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol, MAT_INITIAL_MATRIX, allcolumns, &Mreuse));
3769: }
3771: /*
3772: m - number of local rows
3773: n - number of columns (same on all processors)
3774: rstart - first row in new global matrix generated
3775: */
3776: PetscCall(MatGetSize(Mreuse, &m, &n));
3777: PetscCall(MatGetBlockSizes(Mreuse, &bs, &cbs));
3778: if (call == MAT_INITIAL_MATRIX) {
3779: aij = (Mat_SeqAIJ *)Mreuse->data;
3780: ii = aij->i;
3781: jj = aij->j;
3783: /*
3784: Determine the number of non-zeros in the diagonal and off-diagonal
3785: portions of the matrix in order to do correct preallocation
3786: */
3788: /* first get start and end of "diagonal" columns */
3789: if (csize == PETSC_DECIDE) {
3790: PetscCall(ISGetSize(isrow, &mglobal));
3791: if (mglobal == n) { /* square matrix */
3792: nlocal = m;
3793: } else {
3794: nlocal = n / size + ((n % size) > rank);
3795: }
3796: } else {
3797: nlocal = csize;
3798: }
3799: PetscCallMPI(MPI_Scan(&nlocal, &rend, 1, MPIU_INT, MPI_SUM, comm));
3800: rstart = rend - nlocal;
3801: 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);
3803: /* next, compute all the lengths */
3804: PetscCall(PetscMalloc1(2 * m + 1, &dlens));
3805: olens = dlens + m;
3806: for (i = 0; i < m; i++) {
3807: jend = ii[i + 1] - ii[i];
3808: olen = 0;
3809: dlen = 0;
3810: for (j = 0; j < jend; j++) {
3811: if (*jj < rstart || *jj >= rend) olen++;
3812: else dlen++;
3813: jj++;
3814: }
3815: olens[i] = olen;
3816: dlens[i] = dlen;
3817: }
3818: PetscCall(MatCreate(comm, &M));
3819: PetscCall(MatSetSizes(M, m, nlocal, PETSC_DECIDE, n));
3820: PetscCall(MatSetBlockSizes(M, bs, cbs));
3821: PetscCall(MatSetType(M, ((PetscObject)mat)->type_name));
3822: PetscCall(MatMPIAIJSetPreallocation(M, 0, dlens, 0, olens));
3823: PetscCall(PetscFree(dlens));
3824: } else {
3825: PetscInt ml, nl;
3827: M = *newmat;
3828: PetscCall(MatGetLocalSize(M, &ml, &nl));
3829: PetscCheck(ml == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Previous matrix must be same size/layout as request");
3830: PetscCall(MatZeroEntries(M));
3831: /*
3832: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3833: rather than the slower MatSetValues().
3834: */
3835: M->was_assembled = PETSC_TRUE;
3836: M->assembled = PETSC_FALSE;
3837: }
3838: PetscCall(MatGetOwnershipRange(M, &rstart, &rend));
3839: aij = (Mat_SeqAIJ *)Mreuse->data;
3840: ii = aij->i;
3841: jj = aij->j;
3843: /* trigger copy to CPU if needed */
3844: PetscCall(MatSeqAIJGetArrayRead(Mreuse, (const PetscScalar **)&aa));
3845: for (i = 0; i < m; i++) {
3846: row = rstart + i;
3847: nz = ii[i + 1] - ii[i];
3848: cwork = jj;
3849: jj = PetscSafePointerPlusOffset(jj, nz);
3850: vwork = aa;
3851: aa = PetscSafePointerPlusOffset(aa, nz);
3852: PetscCall(MatSetValues_MPIAIJ(M, 1, &row, nz, cwork, vwork, INSERT_VALUES));
3853: }
3854: PetscCall(MatSeqAIJRestoreArrayRead(Mreuse, (const PetscScalar **)&aa));
3856: PetscCall(MatAssemblyBegin(M, MAT_FINAL_ASSEMBLY));
3857: PetscCall(MatAssemblyEnd(M, MAT_FINAL_ASSEMBLY));
3858: *newmat = M;
3860: /* save submatrix used in processor for next request */
3861: if (call == MAT_INITIAL_MATRIX) {
3862: PetscCall(PetscObjectCompose((PetscObject)M, "SubMatrix", (PetscObject)Mreuse));
3863: PetscCall(MatDestroy(&Mreuse));
3864: }
3865: PetscFunctionReturn(PETSC_SUCCESS);
3866: }
3868: static PetscErrorCode MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[])
3869: {
3870: PetscInt m, cstart, cend, j, nnz, i, d, *ld;
3871: PetscInt *d_nnz, *o_nnz, nnz_max = 0, rstart, ii, irstart;
3872: const PetscInt *JJ;
3873: PetscBool nooffprocentries;
3874: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)B->data;
3876: PetscFunctionBegin;
3877: PetscCall(PetscLayoutSetUp(B->rmap));
3878: PetscCall(PetscLayoutSetUp(B->cmap));
3879: m = B->rmap->n;
3880: cstart = B->cmap->rstart;
3881: cend = B->cmap->rend;
3882: rstart = B->rmap->rstart;
3883: irstart = Ii[0];
3885: PetscCall(PetscCalloc2(m, &d_nnz, m, &o_nnz));
3887: if (PetscDefined(USE_DEBUG)) {
3888: for (i = 0; i < m; i++) {
3889: nnz = Ii[i + 1] - Ii[i];
3890: JJ = PetscSafePointerPlusOffset(J, Ii[i] - irstart);
3891: PetscCheck(nnz >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Local row %" PetscInt_FMT " has a negative %" PetscInt_FMT " number of columns", i, nnz);
3892: 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]);
3893: 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);
3894: }
3895: }
3897: for (i = 0; i < m; i++) {
3898: nnz = Ii[i + 1] - Ii[i];
3899: JJ = PetscSafePointerPlusOffset(J, Ii[i] - irstart);
3900: nnz_max = PetscMax(nnz_max, nnz);
3901: d = 0;
3902: for (j = 0; j < nnz; j++) {
3903: if (cstart <= JJ[j] && JJ[j] < cend) d++;
3904: }
3905: d_nnz[i] = d;
3906: o_nnz[i] = nnz - d;
3907: }
3908: PetscCall(MatMPIAIJSetPreallocation(B, 0, d_nnz, 0, o_nnz));
3909: PetscCall(PetscFree2(d_nnz, o_nnz));
3911: for (i = 0; i < m; i++) {
3912: ii = i + rstart;
3913: PetscCall(MatSetValues_MPIAIJ(B, 1, &ii, Ii[i + 1] - Ii[i], PetscSafePointerPlusOffset(J, Ii[i] - irstart), PetscSafePointerPlusOffset(v, Ii[i] - irstart), INSERT_VALUES));
3914: }
3915: nooffprocentries = B->nooffprocentries;
3916: B->nooffprocentries = PETSC_TRUE;
3917: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
3918: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
3919: B->nooffprocentries = nooffprocentries;
3921: /* count number of entries below block diagonal */
3922: PetscCall(PetscFree(Aij->ld));
3923: PetscCall(PetscCalloc1(m, &ld));
3924: Aij->ld = ld;
3925: for (i = 0; i < m; i++) {
3926: nnz = Ii[i + 1] - Ii[i];
3927: j = 0;
3928: while (j < nnz && J[j] < cstart) j++;
3929: ld[i] = j;
3930: if (J) J += nnz;
3931: }
3933: PetscCall(MatSetOption(B, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
3934: PetscFunctionReturn(PETSC_SUCCESS);
3935: }
3937: /*@
3938: MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in `MATAIJ` format
3939: (the default parallel PETSc format).
3941: Collective
3943: Input Parameters:
3944: + B - the matrix
3945: . i - the indices into `j` for the start of each local row (indices start with zero)
3946: . j - the column indices for each local row (indices start with zero)
3947: - v - optional values in the matrix
3949: Level: developer
3951: Notes:
3952: The `i`, `j`, and `v` arrays ARE copied by this routine into the internal format used by PETSc;
3953: thus you CANNOT change the matrix entries by changing the values of `v` after you have
3954: called this routine. Use `MatCreateMPIAIJWithSplitArrays()` to avoid needing to copy the arrays.
3956: The `i` and `j` indices are 0 based, and `i` indices are indices corresponding to the local `j` array.
3958: A convenience routine for this functionality is `MatCreateMPIAIJWithArrays()`.
3960: You can update the matrix with new numerical values using `MatUpdateMPIAIJWithArrays()` after this call if the column indices in `j` are sorted.
3962: If you do **not** use `MatUpdateMPIAIJWithArrays()`, the column indices in `j` do not need to be sorted. If you will use
3963: `MatUpdateMPIAIJWithArrays()`, the column indices **must** be sorted.
3965: The format which is used for the sparse matrix input, is equivalent to a
3966: row-major ordering.. i.e for the following matrix, the input data expected is
3967: as shown
3968: .vb
3969: 1 0 0
3970: 2 0 3 P0
3971: -------
3972: 4 5 6 P1
3974: Process0 [P0] rows_owned=[0,1]
3975: i = {0,1,3} [size = nrow+1 = 2+1]
3976: j = {0,0,2} [size = 3]
3977: v = {1,2,3} [size = 3]
3979: Process1 [P1] rows_owned=[2]
3980: i = {0,3} [size = nrow+1 = 1+1]
3981: j = {0,1,2} [size = 3]
3982: v = {4,5,6} [size = 3]
3983: .ve
3985: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatCreateAIJ()`,
3986: `MatCreateSeqAIJWithArrays()`, `MatCreateMPIAIJWithSplitArrays()`, `MatCreateMPIAIJWithArrays()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
3987: @*/
3988: PetscErrorCode MatMPIAIJSetPreallocationCSR(Mat B, const PetscInt i[], const PetscInt j[], const PetscScalar v[])
3989: {
3990: PetscFunctionBegin;
3991: PetscTryMethod(B, "MatMPIAIJSetPreallocationCSR_C", (Mat, const PetscInt[], const PetscInt[], const PetscScalar[]), (B, i, j, v));
3992: PetscFunctionReturn(PETSC_SUCCESS);
3993: }
3995: /*@
3996: MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in `MATMPIAIJ` format
3997: (the default parallel PETSc format). For good matrix assembly performance
3998: the user should preallocate the matrix storage by setting the parameters
3999: `d_nz` (or `d_nnz`) and `o_nz` (or `o_nnz`).
4001: Collective
4003: Input Parameters:
4004: + B - the matrix
4005: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
4006: (same value is used for all local rows)
4007: . d_nnz - array containing the number of nonzeros in the various rows of the
4008: DIAGONAL portion of the local submatrix (possibly different for each row)
4009: or `NULL` (`PETSC_NULL_INTEGER` in Fortran), if `d_nz` is used to specify the nonzero structure.
4010: The size of this array is equal to the number of local rows, i.e 'm'.
4011: For matrices that will be factored, you must leave room for (and set)
4012: the diagonal entry even if it is zero.
4013: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
4014: submatrix (same value is used for all local rows).
4015: - o_nnz - array containing the number of nonzeros in the various rows of the
4016: OFF-DIAGONAL portion of the local submatrix (possibly different for
4017: each row) or `NULL` (`PETSC_NULL_INTEGER` in Fortran), if `o_nz` is used to specify the nonzero
4018: structure. The size of this array is equal to the number
4019: of local rows, i.e 'm'.
4021: Example Usage:
4022: Consider the following 8x8 matrix with 34 non-zero values, that is
4023: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4024: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4025: as follows
4027: .vb
4028: 1 2 0 | 0 3 0 | 0 4
4029: Proc0 0 5 6 | 7 0 0 | 8 0
4030: 9 0 10 | 11 0 0 | 12 0
4031: -------------------------------------
4032: 13 0 14 | 15 16 17 | 0 0
4033: Proc1 0 18 0 | 19 20 21 | 0 0
4034: 0 0 0 | 22 23 0 | 24 0
4035: -------------------------------------
4036: Proc2 25 26 27 | 0 0 28 | 29 0
4037: 30 0 0 | 31 32 33 | 0 34
4038: .ve
4040: This can be represented as a collection of submatrices as
4041: .vb
4042: A B C
4043: D E F
4044: G H I
4045: .ve
4047: Where the submatrices A,B,C are owned by proc0, D,E,F are
4048: owned by proc1, G,H,I are owned by proc2.
4050: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4051: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4052: The 'M','N' parameters are 8,8, and have the same values on all procs.
4054: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4055: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4056: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4057: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4058: part as `MATSEQAIJ` matrices. For example, proc1 will store [E] as a `MATSEQAIJ`
4059: matrix, and [DF] as another `MATSEQAIJ` matrix.
4061: When `d_nz`, `o_nz` parameters are specified, `d_nz` storage elements are
4062: allocated for every row of the local DIAGONAL submatrix, and `o_nz`
4063: storage locations are allocated for every row of the OFF-DIAGONAL submatrix.
4064: One way to choose `d_nz` and `o_nz` is to use the maximum number of nonzeros over
4065: the local rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4066: In this case, the values of `d_nz`, `o_nz` are
4067: .vb
4068: proc0 dnz = 2, o_nz = 2
4069: proc1 dnz = 3, o_nz = 2
4070: proc2 dnz = 1, o_nz = 4
4071: .ve
4072: We are allocating `m`*(`d_nz`+`o_nz`) storage locations for every proc. This
4073: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4074: for proc3. i.e we are using 12+15+10=37 storage locations to store
4075: 34 values.
4077: When `d_nnz`, `o_nnz` parameters are specified, the storage is specified
4078: for every row, corresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4079: In the above case the values for `d_nnz`, `o_nnz` are
4080: .vb
4081: proc0 d_nnz = [2,2,2] and o_nnz = [2,2,2]
4082: proc1 d_nnz = [3,3,2] and o_nnz = [2,1,1]
4083: proc2 d_nnz = [1,1] and o_nnz = [4,4]
4084: .ve
4085: Here the space allocated is sum of all the above values i.e 34, and
4086: hence pre-allocation is perfect.
4088: Level: intermediate
4090: Notes:
4091: If the *_nnz parameter is given then the *_nz parameter is ignored
4093: The `MATAIJ` format, also called compressed row storage (CSR), is compatible with standard Fortran
4094: storage. The stored row and column indices begin with zero.
4095: See [Sparse Matrices](sec_matsparse) for details.
4097: The parallel matrix is partitioned such that the first m0 rows belong to
4098: process 0, the next m1 rows belong to process 1, the next m2 rows belong
4099: to process 2 etc.. where m0,m1,m2... are the input parameter 'm'.
4101: The DIAGONAL portion of the local submatrix of a processor can be defined
4102: as the submatrix which is obtained by extraction the part corresponding to
4103: the rows r1-r2 and columns c1-c2 of the global matrix, where r1 is the
4104: first row that belongs to the processor, r2 is the last row belonging to
4105: the this processor, and c1-c2 is range of indices of the local part of a
4106: vector suitable for applying the matrix to. This is an mxn matrix. In the
4107: common case of a square matrix, the row and column ranges are the same and
4108: the DIAGONAL part is also square. The remaining portion of the local
4109: submatrix (mxN) constitute the OFF-DIAGONAL portion.
4111: If `o_nnz` and `d_nnz` are specified, then `o_nz` and `d_nz` are ignored.
4113: You can call `MatGetInfo()` to get information on how effective the preallocation was;
4114: for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
4115: You can also run with the option `-info` and look for messages with the string
4116: malloc in them to see if additional memory allocation was needed.
4118: .seealso: [](ch_matrices), `Mat`, [Sparse Matrices](sec_matsparse), `MATMPIAIJ`, `MATAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatCreateAIJ()`, `MatMPIAIJSetPreallocationCSR()`,
4119: `MatGetInfo()`, `PetscSplitOwnership()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4120: @*/
4121: PetscErrorCode MatMPIAIJSetPreallocation(Mat B, PetscInt d_nz, const PetscInt d_nnz[], PetscInt o_nz, const PetscInt o_nnz[])
4122: {
4123: PetscFunctionBegin;
4126: PetscTryMethod(B, "MatMPIAIJSetPreallocation_C", (Mat, PetscInt, const PetscInt[], PetscInt, const PetscInt[]), (B, d_nz, d_nnz, o_nz, o_nnz));
4127: PetscFunctionReturn(PETSC_SUCCESS);
4128: }
4130: /*@
4131: MatCreateMPIAIJWithArrays - creates a `MATMPIAIJ` matrix using arrays that contain in standard
4132: CSR format for the local rows.
4134: Collective
4136: Input Parameters:
4137: + comm - MPI communicator
4138: . m - number of local rows (Cannot be `PETSC_DECIDE`)
4139: . n - This value should be the same as the local size used in creating the
4140: x vector for the matrix-vector product $ y = Ax$. (or `PETSC_DECIDE` to have
4141: calculated if `N` is given) For square matrices n is almost always `m`.
4142: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if `m` is given)
4143: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if `n` is given)
4144: . 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
4145: . j - global column indices
4146: - a - optional matrix values
4148: Output Parameter:
4149: . mat - the matrix
4151: Level: intermediate
4153: Notes:
4154: The `i`, `j`, and `a` arrays ARE copied by this routine into the internal format used by PETSc;
4155: thus you CANNOT change the matrix entries by changing the values of `a[]` after you have
4156: called this routine. Use `MatCreateMPIAIJWithSplitArrays()` to avoid needing to copy the arrays.
4158: The `i` and `j` indices are 0 based, and `i` indices are indices corresponding to the local `j` array.
4160: Once you have created the matrix you can update it with new numerical values using `MatUpdateMPIAIJWithArray()`
4162: If you do **not** use `MatUpdateMPIAIJWithArray()`, the column indices in `j` do not need to be sorted. If you will use
4163: `MatUpdateMPIAIJWithArrays()`, the column indices **must** be sorted.
4165: The format which is used for the sparse matrix input, is equivalent to a
4166: row-major ordering, i.e., for the following matrix, the input data expected is
4167: as shown
4168: .vb
4169: 1 0 0
4170: 2 0 3 P0
4171: -------
4172: 4 5 6 P1
4174: Process0 [P0] rows_owned=[0,1]
4175: i = {0,1,3} [size = nrow+1 = 2+1]
4176: j = {0,0,2} [size = 3]
4177: v = {1,2,3} [size = 3]
4179: Process1 [P1] rows_owned=[2]
4180: i = {0,3} [size = nrow+1 = 1+1]
4181: j = {0,1,2} [size = 3]
4182: v = {4,5,6} [size = 3]
4183: .ve
4185: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4186: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArray()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4187: @*/
4188: PetscErrorCode MatCreateMPIAIJWithArrays(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt M, PetscInt N, const PetscInt i[], const PetscInt j[], const PetscScalar a[], Mat *mat)
4189: {
4190: PetscFunctionBegin;
4191: PetscCheck(!i || !i[0], PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
4192: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
4193: PetscCall(MatCreate(comm, mat));
4194: PetscCall(MatSetSizes(*mat, m, n, M, N));
4195: /* PetscCall(MatSetBlockSizes(M,bs,cbs)); */
4196: PetscCall(MatSetType(*mat, MATMPIAIJ));
4197: PetscCall(MatMPIAIJSetPreallocationCSR(*mat, i, j, a));
4198: PetscFunctionReturn(PETSC_SUCCESS);
4199: }
4201: /*@
4202: MatUpdateMPIAIJWithArrays - updates a `MATMPIAIJ` matrix using arrays that contain in standard
4203: CSR format for the local rows. Only the numerical values are updated the other arrays must be identical to what was passed
4204: from `MatCreateMPIAIJWithArrays()`
4206: Deprecated: Use `MatUpdateMPIAIJWithArray()`
4208: Collective
4210: Input Parameters:
4211: + mat - the matrix
4212: . m - number of local rows (Cannot be `PETSC_DECIDE`)
4213: . n - This value should be the same as the local size used in creating the
4214: x vector for the matrix-vector product y = Ax. (or `PETSC_DECIDE` to have
4215: calculated if N is given) For square matrices n is almost always m.
4216: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if m is given)
4217: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if n is given)
4218: . Ii - row indices; that is Ii[0] = 0, Ii[row] = Ii[row-1] + number of elements in that row of the matrix
4219: . J - column indices
4220: - v - matrix values
4222: Level: deprecated
4224: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4225: `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArray()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4226: @*/
4227: PetscErrorCode MatUpdateMPIAIJWithArrays(Mat mat, PetscInt m, PetscInt n, PetscInt M, PetscInt N, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[])
4228: {
4229: PetscInt nnz, i;
4230: PetscBool nooffprocentries;
4231: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)mat->data;
4232: Mat_SeqAIJ *Ad = (Mat_SeqAIJ *)Aij->A->data;
4233: PetscScalar *ad, *ao;
4234: PetscInt ldi, Iii, md;
4235: const PetscInt *Adi = Ad->i;
4236: PetscInt *ld = Aij->ld;
4238: PetscFunctionBegin;
4239: PetscCheck(Ii[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
4240: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
4241: PetscCheck(m == mat->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Local number of rows cannot change from call to MatUpdateMPIAIJWithArrays()");
4242: PetscCheck(n == mat->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Local number of columns cannot change from call to MatUpdateMPIAIJWithArrays()");
4244: PetscCall(MatSeqAIJGetArrayWrite(Aij->A, &ad));
4245: PetscCall(MatSeqAIJGetArrayWrite(Aij->B, &ao));
4247: for (i = 0; i < m; i++) {
4248: if (PetscDefined(USE_DEBUG)) {
4249: for (PetscInt j = Ii[i] + 1; j < Ii[i + 1]; ++j) {
4250: 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);
4251: 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);
4252: }
4253: }
4254: nnz = Ii[i + 1] - Ii[i];
4255: Iii = Ii[i];
4256: ldi = ld[i];
4257: md = Adi[i + 1] - Adi[i];
4258: PetscCall(PetscArraycpy(ao, v + Iii, ldi));
4259: PetscCall(PetscArraycpy(ad, v + Iii + ldi, md));
4260: PetscCall(PetscArraycpy(ao + ldi, v + Iii + ldi + md, nnz - ldi - md));
4261: ad += md;
4262: ao += nnz - md;
4263: }
4264: nooffprocentries = mat->nooffprocentries;
4265: mat->nooffprocentries = PETSC_TRUE;
4266: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->A, &ad));
4267: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->B, &ao));
4268: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->A));
4269: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->B));
4270: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
4271: PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
4272: PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
4273: mat->nooffprocentries = nooffprocentries;
4274: PetscFunctionReturn(PETSC_SUCCESS);
4275: }
4277: /*@
4278: MatUpdateMPIAIJWithArray - updates an `MATMPIAIJ` matrix using an array that contains the nonzero values
4280: Collective
4282: Input Parameters:
4283: + mat - the matrix
4284: - v - matrix values, stored by row
4286: Level: intermediate
4288: Notes:
4289: The matrix must have been obtained with `MatCreateMPIAIJWithArrays()` or `MatMPIAIJSetPreallocationCSR()`
4291: The column indices in the call to `MatCreateMPIAIJWithArrays()` or `MatMPIAIJSetPreallocationCSR()` must have been sorted for this call to work correctly
4293: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4294: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArrays()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4295: @*/
4296: PetscErrorCode MatUpdateMPIAIJWithArray(Mat mat, const PetscScalar v[])
4297: {
4298: PetscInt nnz, i, m;
4299: PetscBool nooffprocentries;
4300: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)mat->data;
4301: Mat_SeqAIJ *Ad = (Mat_SeqAIJ *)Aij->A->data;
4302: Mat_SeqAIJ *Ao = (Mat_SeqAIJ *)Aij->B->data;
4303: PetscScalar *ad, *ao;
4304: const PetscInt *Adi = Ad->i, *Adj = Ao->i;
4305: PetscInt ldi, Iii, md;
4306: PetscInt *ld = Aij->ld;
4308: PetscFunctionBegin;
4309: m = mat->rmap->n;
4311: PetscCall(MatSeqAIJGetArrayWrite(Aij->A, &ad));
4312: PetscCall(MatSeqAIJGetArrayWrite(Aij->B, &ao));
4313: Iii = 0;
4314: for (i = 0; i < m; i++) {
4315: nnz = Adi[i + 1] - Adi[i] + Adj[i + 1] - Adj[i];
4316: ldi = ld[i];
4317: md = Adi[i + 1] - Adi[i];
4318: PetscCall(PetscArraycpy(ad, v + Iii + ldi, md));
4319: ad += md;
4320: if (ao) {
4321: PetscCall(PetscArraycpy(ao, v + Iii, ldi));
4322: PetscCall(PetscArraycpy(ao + ldi, v + Iii + ldi + md, nnz - ldi - md));
4323: ao += nnz - md;
4324: }
4325: Iii += nnz;
4326: }
4327: nooffprocentries = mat->nooffprocentries;
4328: mat->nooffprocentries = PETSC_TRUE;
4329: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->A, &ad));
4330: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->B, &ao));
4331: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->A));
4332: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->B));
4333: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
4334: PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
4335: PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
4336: mat->nooffprocentries = nooffprocentries;
4337: PetscFunctionReturn(PETSC_SUCCESS);
4338: }
4340: /*@
4341: MatCreateAIJ - Creates a sparse parallel matrix in `MATAIJ` format
4342: (the default parallel PETSc format). For good matrix assembly performance
4343: the user should preallocate the matrix storage by setting the parameters
4344: `d_nz` (or `d_nnz`) and `o_nz` (or `o_nnz`).
4346: Collective
4348: Input Parameters:
4349: + comm - MPI communicator
4350: . m - number of local rows (or `PETSC_DECIDE` to have calculated if M is given)
4351: This value should be the same as the local size used in creating the
4352: y vector for the matrix-vector product y = Ax.
4353: . n - This value should be the same as the local size used in creating the
4354: x vector for the matrix-vector product y = Ax. (or `PETSC_DECIDE` to have
4355: calculated if N is given) For square matrices n is almost always m.
4356: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if m is given)
4357: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if n is given)
4358: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
4359: (same value is used for all local rows)
4360: . d_nnz - array containing the number of nonzeros in the various rows of the
4361: DIAGONAL portion of the local submatrix (possibly different for each row)
4362: or `NULL`, if `d_nz` is used to specify the nonzero structure.
4363: The size of this array is equal to the number of local rows, i.e 'm'.
4364: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
4365: submatrix (same value is used for all local rows).
4366: - o_nnz - array containing the number of nonzeros in the various rows of the
4367: OFF-DIAGONAL portion of the local submatrix (possibly different for
4368: each row) or `NULL`, if `o_nz` is used to specify the nonzero
4369: structure. The size of this array is equal to the number
4370: of local rows, i.e 'm'.
4372: Output Parameter:
4373: . A - the matrix
4375: Options Database Keys:
4376: + -mat_no_inode - Do not use inodes
4377: . -mat_inode_limit <limit> - Sets inode limit (max limit=5)
4378: - -matmult_vecscatter_view <viewer> - View the vecscatter (i.e., communication pattern) used in `MatMult()` of sparse parallel matrices.
4379: See viewer types in manual of `MatView()`. Of them, ascii_matlab, draw or binary cause the `VecScatter`
4380: 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.
4382: Level: intermediate
4384: Notes:
4385: It is recommended that one use `MatCreateFromOptions()` or the `MatCreate()`, `MatSetType()` and/or `MatSetFromOptions()`,
4386: MatXXXXSetPreallocation() paradigm instead of this routine directly.
4387: [MatXXXXSetPreallocation() is, for example, `MatSeqAIJSetPreallocation()`]
4389: If the *_nnz parameter is given then the *_nz parameter is ignored
4391: The `m`,`n`,`M`,`N` parameters specify the size of the matrix, and its partitioning across
4392: processors, while `d_nz`,`d_nnz`,`o_nz`,`o_nnz` parameters specify the approximate
4393: storage requirements for this matrix.
4395: If `PETSC_DECIDE` or `PETSC_DETERMINE` is used for a particular argument on one
4396: processor than it must be used on all processors that share the object for
4397: that argument.
4399: If `m` and `n` are not `PETSC_DECIDE`, then the values determine the `PetscLayout` of the matrix and the ranges returned by
4400: `MatGetOwnershipRange()`, `MatGetOwnershipRanges()`, `MatGetOwnershipRangeColumn()`, and `MatGetOwnershipRangesColumn()`.
4402: The user MUST specify either the local or global matrix dimensions
4403: (possibly both).
4405: The parallel matrix is partitioned across processors such that the
4406: first `m0` rows belong to process 0, the next `m1` rows belong to
4407: process 1, the next `m2` rows belong to process 2, etc., where
4408: `m0`, `m1`, `m2`... are the input parameter `m` on each MPI process. I.e., each MPI process stores
4409: values corresponding to [m x N] submatrix.
4411: The columns are logically partitioned with the n0 columns belonging
4412: to 0th partition, the next n1 columns belonging to the next
4413: partition etc.. where n0,n1,n2... are the input parameter 'n'.
4415: The DIAGONAL portion of the local submatrix on any given processor
4416: is the submatrix corresponding to the rows and columns m,n
4417: corresponding to the given processor. i.e diagonal matrix on
4418: process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1]
4419: etc. The remaining portion of the local submatrix [m x (N-n)]
4420: constitute the OFF-DIAGONAL portion. The example below better
4421: illustrates this concept. The two matrices, the DIAGONAL portion and
4422: the OFF-DIAGONAL portion are each stored as `MATSEQAIJ` matrices.
4424: For a square global matrix we define each processor's diagonal portion
4425: to be its local rows and the corresponding columns (a square submatrix);
4426: each processor's off-diagonal portion encompasses the remainder of the
4427: local matrix (a rectangular submatrix).
4429: If `o_nnz`, `d_nnz` are specified, then `o_nz`, and `d_nz` are ignored.
4431: When calling this routine with a single process communicator, a matrix of
4432: type `MATSEQAIJ` is returned. If a matrix of type `MATMPIAIJ` is desired for this
4433: type of communicator, use the construction mechanism
4434: .vb
4435: MatCreate(..., &A);
4436: MatSetType(A, MATMPIAIJ);
4437: MatSetSizes(A, m, n, M, N);
4438: MatMPIAIJSetPreallocation(A, ...);
4439: .ve
4441: By default, this format uses inodes (identical nodes) when possible.
4442: We search for consecutive rows with the same nonzero structure, thereby
4443: reusing matrix information to achieve increased efficiency.
4445: Example Usage:
4446: Consider the following 8x8 matrix with 34 non-zero values, that is
4447: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4448: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4449: as follows
4451: .vb
4452: 1 2 0 | 0 3 0 | 0 4
4453: Proc0 0 5 6 | 7 0 0 | 8 0
4454: 9 0 10 | 11 0 0 | 12 0
4455: -------------------------------------
4456: 13 0 14 | 15 16 17 | 0 0
4457: Proc1 0 18 0 | 19 20 21 | 0 0
4458: 0 0 0 | 22 23 0 | 24 0
4459: -------------------------------------
4460: Proc2 25 26 27 | 0 0 28 | 29 0
4461: 30 0 0 | 31 32 33 | 0 34
4462: .ve
4464: This can be represented as a collection of submatrices as
4466: .vb
4467: A B C
4468: D E F
4469: G H I
4470: .ve
4472: Where the submatrices A,B,C are owned by proc0, D,E,F are
4473: owned by proc1, G,H,I are owned by proc2.
4475: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4476: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4477: The 'M','N' parameters are 8,8, and have the same values on all procs.
4479: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4480: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4481: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4482: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4483: part as `MATSEQAIJ` matrices. For example, proc1 will store [E] as a `MATSEQAIJ`
4484: matrix, and [DF] as another SeqAIJ matrix.
4486: When `d_nz`, `o_nz` parameters are specified, `d_nz` storage elements are
4487: allocated for every row of the local DIAGONAL submatrix, and `o_nz`
4488: storage locations are allocated for every row of the OFF-DIAGONAL submatrix.
4489: One way to choose `d_nz` and `o_nz` is to use the maximum number of nonzeros over
4490: the local rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4491: In this case, the values of `d_nz`,`o_nz` are
4492: .vb
4493: proc0 dnz = 2, o_nz = 2
4494: proc1 dnz = 3, o_nz = 2
4495: proc2 dnz = 1, o_nz = 4
4496: .ve
4497: We are allocating m*(`d_nz`+`o_nz`) storage locations for every proc. This
4498: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4499: for proc3. i.e we are using 12+15+10=37 storage locations to store
4500: 34 values.
4502: When `d_nnz`, `o_nnz` parameters are specified, the storage is specified
4503: for every row, corresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4504: In the above case the values for d_nnz,o_nnz are
4505: .vb
4506: proc0 d_nnz = [2,2,2] and o_nnz = [2,2,2]
4507: proc1 d_nnz = [3,3,2] and o_nnz = [2,1,1]
4508: proc2 d_nnz = [1,1] and o_nnz = [4,4]
4509: .ve
4510: Here the space allocated is sum of all the above values i.e 34, and
4511: hence pre-allocation is perfect.
4513: .seealso: [](ch_matrices), `Mat`, [Sparse Matrix Creation](sec_matsparse), `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4514: `MATMPIAIJ`, `MatCreateMPIAIJWithArrays()`, `MatGetOwnershipRange()`, `MatGetOwnershipRanges()`, `MatGetOwnershipRangeColumn()`,
4515: `MatGetOwnershipRangesColumn()`, `PetscLayout`
4516: @*/
4517: 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)
4518: {
4519: PetscMPIInt size;
4521: PetscFunctionBegin;
4522: PetscCall(MatCreate(comm, A));
4523: PetscCall(MatSetSizes(*A, m, n, M, N));
4524: PetscCallMPI(MPI_Comm_size(comm, &size));
4525: if (size > 1) {
4526: PetscCall(MatSetType(*A, MATMPIAIJ));
4527: PetscCall(MatMPIAIJSetPreallocation(*A, d_nz, d_nnz, o_nz, o_nnz));
4528: } else {
4529: PetscCall(MatSetType(*A, MATSEQAIJ));
4530: PetscCall(MatSeqAIJSetPreallocation(*A, d_nz, d_nnz));
4531: }
4532: PetscFunctionReturn(PETSC_SUCCESS);
4533: }
4535: /*@C
4536: MatMPIAIJGetSeqAIJ - Returns the local pieces of this distributed matrix
4538: Not Collective
4540: Input Parameter:
4541: . A - The `MATMPIAIJ` matrix
4543: Output Parameters:
4544: + Ad - The local diagonal block as a `MATSEQAIJ` matrix
4545: . Ao - The local off-diagonal block as a `MATSEQAIJ` matrix
4546: - colmap - An array mapping local column numbers of `Ao` to global column numbers of the parallel matrix
4548: Level: intermediate
4550: Note:
4551: The rows in `Ad` and `Ao` are in [0, Nr), where Nr is the number of local rows on this process. The columns
4552: 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
4553: the number of nonzero columns in the local off-diagonal piece of the matrix `A`. The array colmap maps these
4554: local column numbers to global column numbers in the original matrix.
4556: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatMPIAIJGetLocalMat()`, `MatMPIAIJGetLocalMatCondensed()`, `MatCreateAIJ()`, `MATSEQAIJ`
4557: @*/
4558: PetscErrorCode MatMPIAIJGetSeqAIJ(Mat A, Mat *Ad, Mat *Ao, const PetscInt *colmap[])
4559: {
4560: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
4561: PetscBool flg;
4563: PetscFunctionBegin;
4564: PetscCall(PetscStrbeginswith(((PetscObject)A)->type_name, MATMPIAIJ, &flg));
4565: PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "This function requires a MATMPIAIJ matrix as input");
4566: if (Ad) *Ad = a->A;
4567: if (Ao) *Ao = a->B;
4568: if (colmap) *colmap = a->garray;
4569: PetscFunctionReturn(PETSC_SUCCESS);
4570: }
4572: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm comm, Mat inmat, PetscInt n, MatReuse scall, Mat *outmat)
4573: {
4574: PetscInt m, N, i, rstart, nnz, Ii;
4575: PetscInt *indx;
4576: PetscScalar *values;
4577: MatType rootType;
4579: PetscFunctionBegin;
4580: PetscCall(MatGetSize(inmat, &m, &N));
4581: if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */
4582: PetscInt *dnz, *onz, sum, bs, cbs;
4584: if (n == PETSC_DECIDE) PetscCall(PetscSplitOwnership(comm, &n, &N));
4585: /* Check sum(n) = N */
4586: PetscCallMPI(MPIU_Allreduce(&n, &sum, 1, MPIU_INT, MPI_SUM, comm));
4587: PetscCheck(sum == N, PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Sum of local columns %" PetscInt_FMT " != global columns %" PetscInt_FMT, sum, N);
4589: PetscCallMPI(MPI_Scan(&m, &rstart, 1, MPIU_INT, MPI_SUM, comm));
4590: rstart -= m;
4592: MatPreallocateBegin(comm, m, n, dnz, onz);
4593: for (i = 0; i < m; i++) {
4594: PetscCall(MatGetRow_SeqAIJ(inmat, i, &nnz, &indx, NULL));
4595: PetscCall(MatPreallocateSet(i + rstart, nnz, indx, dnz, onz));
4596: PetscCall(MatRestoreRow_SeqAIJ(inmat, i, &nnz, &indx, NULL));
4597: }
4599: PetscCall(MatCreate(comm, outmat));
4600: PetscCall(MatSetSizes(*outmat, m, n, PETSC_DETERMINE, PETSC_DETERMINE));
4601: PetscCall(MatGetBlockSizes(inmat, &bs, &cbs));
4602: PetscCall(MatSetBlockSizes(*outmat, bs, cbs));
4603: PetscCall(MatGetRootType_Private(inmat, &rootType));
4604: PetscCall(MatSetType(*outmat, rootType));
4605: PetscCall(MatSeqAIJSetPreallocation(*outmat, 0, dnz));
4606: PetscCall(MatMPIAIJSetPreallocation(*outmat, 0, dnz, 0, onz));
4607: MatPreallocateEnd(dnz, onz);
4608: PetscCall(MatSetOption(*outmat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
4609: }
4611: /* numeric phase */
4612: PetscCall(MatGetOwnershipRange(*outmat, &rstart, NULL));
4613: for (i = 0; i < m; i++) {
4614: PetscCall(MatGetRow_SeqAIJ(inmat, i, &nnz, &indx, &values));
4615: Ii = i + rstart;
4616: PetscCall(MatSetValues(*outmat, 1, &Ii, nnz, indx, values, INSERT_VALUES));
4617: PetscCall(MatRestoreRow_SeqAIJ(inmat, i, &nnz, &indx, &values));
4618: }
4619: PetscCall(MatAssemblyBegin(*outmat, MAT_FINAL_ASSEMBLY));
4620: PetscCall(MatAssemblyEnd(*outmat, MAT_FINAL_ASSEMBLY));
4621: PetscFunctionReturn(PETSC_SUCCESS);
4622: }
4624: static PetscErrorCode MatMergeSeqsToMPIDestroy(void **data)
4625: {
4626: MatMergeSeqsToMPI *merge = (MatMergeSeqsToMPI *)*data;
4628: PetscFunctionBegin;
4629: if (!merge) PetscFunctionReturn(PETSC_SUCCESS);
4630: PetscCall(PetscFree(merge->id_r));
4631: PetscCall(PetscFree(merge->len_s));
4632: PetscCall(PetscFree(merge->len_r));
4633: PetscCall(PetscFree(merge->bi));
4634: PetscCall(PetscFree(merge->bj));
4635: PetscCall(PetscFree(merge->buf_ri[0]));
4636: PetscCall(PetscFree(merge->buf_ri));
4637: PetscCall(PetscFree(merge->buf_rj[0]));
4638: PetscCall(PetscFree(merge->buf_rj));
4639: PetscCall(PetscFree(merge->coi));
4640: PetscCall(PetscFree(merge->coj));
4641: PetscCall(PetscFree(merge->owners_co));
4642: PetscCall(PetscLayoutDestroy(&merge->rowmap));
4643: PetscCall(PetscFree(merge));
4644: PetscFunctionReturn(PETSC_SUCCESS);
4645: }
4647: #include <../src/mat/utils/freespace.h>
4648: #include <petscbt.h>
4650: PetscErrorCode MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat, Mat mpimat)
4651: {
4652: MPI_Comm comm;
4653: Mat_SeqAIJ *a = (Mat_SeqAIJ *)seqmat->data;
4654: PetscMPIInt size, rank, taga, *len_s;
4655: PetscInt N = mpimat->cmap->N, i, j, *owners, *ai = a->i, *aj, m;
4656: PetscMPIInt proc, k;
4657: PetscInt **buf_ri, **buf_rj;
4658: PetscInt anzi, *bj_i, *bi, *bj, arow, bnzi, nextaj;
4659: PetscInt nrows, **buf_ri_k, **nextrow, **nextai;
4660: MPI_Request *s_waits, *r_waits;
4661: MPI_Status *status;
4662: const MatScalar *aa, *a_a;
4663: MatScalar **abuf_r, *ba_i;
4664: MatMergeSeqsToMPI *merge;
4665: PetscContainer container;
4667: PetscFunctionBegin;
4668: PetscCall(PetscObjectGetComm((PetscObject)mpimat, &comm));
4669: PetscCall(PetscLogEventBegin(MAT_Seqstompinum, seqmat, 0, 0, 0));
4671: PetscCallMPI(MPI_Comm_size(comm, &size));
4672: PetscCallMPI(MPI_Comm_rank(comm, &rank));
4674: PetscCall(PetscObjectQuery((PetscObject)mpimat, "MatMergeSeqsToMPI", (PetscObject *)&container));
4675: PetscCheck(container, PetscObjectComm((PetscObject)mpimat), PETSC_ERR_PLIB, "Mat not created from MatCreateMPIAIJSumSeqAIJSymbolic");
4676: PetscCall(PetscContainerGetPointer(container, (void **)&merge));
4677: PetscCall(MatSeqAIJGetArrayRead(seqmat, &a_a));
4678: aa = a_a;
4680: bi = merge->bi;
4681: bj = merge->bj;
4682: buf_ri = merge->buf_ri;
4683: buf_rj = merge->buf_rj;
4685: PetscCall(PetscMalloc1(size, &status));
4686: owners = merge->rowmap->range;
4687: len_s = merge->len_s;
4689: /* send and recv matrix values */
4690: PetscCall(PetscObjectGetNewTag((PetscObject)mpimat, &taga));
4691: PetscCall(PetscPostIrecvScalar(comm, taga, merge->nrecv, merge->id_r, merge->len_r, &abuf_r, &r_waits));
4693: PetscCall(PetscMalloc1(merge->nsend + 1, &s_waits));
4694: for (proc = 0, k = 0; proc < size; proc++) {
4695: if (!len_s[proc]) continue;
4696: i = owners[proc];
4697: PetscCallMPI(MPIU_Isend(aa + ai[i], len_s[proc], MPIU_MATSCALAR, proc, taga, comm, s_waits + k));
4698: k++;
4699: }
4701: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, r_waits, status));
4702: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, s_waits, status));
4703: PetscCall(PetscFree(status));
4705: PetscCall(PetscFree(s_waits));
4706: PetscCall(PetscFree(r_waits));
4708: /* insert mat values of mpimat */
4709: PetscCall(PetscMalloc1(N, &ba_i));
4710: PetscCall(PetscMalloc3(merge->nrecv, &buf_ri_k, merge->nrecv, &nextrow, merge->nrecv, &nextai));
4712: for (k = 0; k < merge->nrecv; k++) {
4713: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4714: nrows = *buf_ri_k[k];
4715: nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */
4716: nextai[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure */
4717: }
4719: /* set values of ba */
4720: m = merge->rowmap->n;
4721: for (i = 0; i < m; i++) {
4722: arow = owners[rank] + i;
4723: bj_i = bj + bi[i]; /* col indices of the i-th row of mpimat */
4724: bnzi = bi[i + 1] - bi[i];
4725: PetscCall(PetscArrayzero(ba_i, bnzi));
4727: /* add local non-zero vals of this proc's seqmat into ba */
4728: anzi = ai[arow + 1] - ai[arow];
4729: aj = a->j + ai[arow];
4730: aa = a_a + ai[arow];
4731: nextaj = 0;
4732: for (j = 0; nextaj < anzi; j++) {
4733: if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4734: ba_i[j] += aa[nextaj++];
4735: }
4736: }
4738: /* add received vals into ba */
4739: for (k = 0; k < merge->nrecv; k++) { /* k-th received message */
4740: /* i-th row */
4741: if (i == *nextrow[k]) {
4742: anzi = *(nextai[k] + 1) - *nextai[k];
4743: aj = buf_rj[k] + *nextai[k];
4744: aa = abuf_r[k] + *nextai[k];
4745: nextaj = 0;
4746: for (j = 0; nextaj < anzi; j++) {
4747: if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4748: ba_i[j] += aa[nextaj++];
4749: }
4750: }
4751: nextrow[k]++;
4752: nextai[k]++;
4753: }
4754: }
4755: PetscCall(MatSetValues(mpimat, 1, &arow, bnzi, bj_i, ba_i, INSERT_VALUES));
4756: }
4757: PetscCall(MatSeqAIJRestoreArrayRead(seqmat, &a_a));
4758: PetscCall(MatAssemblyBegin(mpimat, MAT_FINAL_ASSEMBLY));
4759: PetscCall(MatAssemblyEnd(mpimat, MAT_FINAL_ASSEMBLY));
4761: PetscCall(PetscFree(abuf_r[0]));
4762: PetscCall(PetscFree(abuf_r));
4763: PetscCall(PetscFree(ba_i));
4764: PetscCall(PetscFree3(buf_ri_k, nextrow, nextai));
4765: PetscCall(PetscLogEventEnd(MAT_Seqstompinum, seqmat, 0, 0, 0));
4766: PetscFunctionReturn(PETSC_SUCCESS);
4767: }
4769: PetscErrorCode MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm, Mat seqmat, PetscInt m, PetscInt n, Mat *mpimat)
4770: {
4771: Mat B_mpi;
4772: Mat_SeqAIJ *a = (Mat_SeqAIJ *)seqmat->data;
4773: PetscMPIInt size, rank, tagi, tagj, *len_s, *len_si, *len_ri;
4774: PetscInt **buf_rj, **buf_ri, **buf_ri_k;
4775: PetscInt M = seqmat->rmap->n, N = seqmat->cmap->n, i, *owners, *ai = a->i, *aj = a->j;
4776: PetscInt len, *dnz, *onz, bs, cbs;
4777: PetscInt k, anzi, *bi, *bj, *lnk, nlnk, arow, bnzi;
4778: PetscInt nrows, *buf_s, *buf_si, *buf_si_i, **nextrow, **nextai;
4779: MPI_Request *si_waits, *sj_waits, *ri_waits, *rj_waits;
4780: MPI_Status *status;
4781: PetscFreeSpaceList free_space = NULL, current_space = NULL;
4782: PetscBT lnkbt;
4783: MatMergeSeqsToMPI *merge;
4784: PetscContainer container;
4786: PetscFunctionBegin;
4787: PetscCall(PetscLogEventBegin(MAT_Seqstompisym, seqmat, 0, 0, 0));
4789: /* make sure it is a PETSc comm */
4790: PetscCall(PetscCommDuplicate(comm, &comm, NULL));
4791: PetscCallMPI(MPI_Comm_size(comm, &size));
4792: PetscCallMPI(MPI_Comm_rank(comm, &rank));
4794: PetscCall(PetscNew(&merge));
4795: PetscCall(PetscMalloc1(size, &status));
4797: /* determine row ownership */
4798: PetscCall(PetscLayoutCreate(comm, &merge->rowmap));
4799: PetscCall(PetscLayoutSetLocalSize(merge->rowmap, m));
4800: PetscCall(PetscLayoutSetSize(merge->rowmap, M));
4801: PetscCall(PetscLayoutSetBlockSize(merge->rowmap, 1));
4802: PetscCall(PetscLayoutSetUp(merge->rowmap));
4803: PetscCall(PetscMalloc1(size, &len_si));
4804: PetscCall(PetscMalloc1(size, &merge->len_s));
4806: m = merge->rowmap->n;
4807: owners = merge->rowmap->range;
4809: /* determine the number of messages to send, their lengths */
4810: len_s = merge->len_s;
4812: len = 0; /* length of buf_si[] */
4813: merge->nsend = 0;
4814: for (PetscMPIInt proc = 0; proc < size; proc++) {
4815: len_si[proc] = 0;
4816: if (proc == rank) {
4817: len_s[proc] = 0;
4818: } else {
4819: PetscCall(PetscMPIIntCast(owners[proc + 1] - owners[proc] + 1, &len_si[proc]));
4820: PetscCall(PetscMPIIntCast(ai[owners[proc + 1]] - ai[owners[proc]], &len_s[proc])); /* num of rows to be sent to [proc] */
4821: }
4822: if (len_s[proc]) {
4823: merge->nsend++;
4824: nrows = 0;
4825: for (i = owners[proc]; i < owners[proc + 1]; i++) {
4826: if (ai[i + 1] > ai[i]) nrows++;
4827: }
4828: PetscCall(PetscMPIIntCast(2 * (nrows + 1), &len_si[proc]));
4829: len += len_si[proc];
4830: }
4831: }
4833: /* determine the number and length of messages to receive for ij-structure */
4834: PetscCall(PetscGatherNumberOfMessages(comm, NULL, len_s, &merge->nrecv));
4835: PetscCall(PetscGatherMessageLengths2(comm, merge->nsend, merge->nrecv, len_s, len_si, &merge->id_r, &merge->len_r, &len_ri));
4837: /* post the Irecv of j-structure */
4838: PetscCall(PetscCommGetNewTag(comm, &tagj));
4839: PetscCall(PetscPostIrecvInt(comm, tagj, merge->nrecv, merge->id_r, merge->len_r, &buf_rj, &rj_waits));
4841: /* post the Isend of j-structure */
4842: PetscCall(PetscMalloc2(merge->nsend, &si_waits, merge->nsend, &sj_waits));
4844: for (PetscMPIInt proc = 0, k = 0; proc < size; proc++) {
4845: if (!len_s[proc]) continue;
4846: i = owners[proc];
4847: PetscCallMPI(MPIU_Isend(aj + ai[i], len_s[proc], MPIU_INT, proc, tagj, comm, sj_waits + k));
4848: k++;
4849: }
4851: /* receives and sends of j-structure are complete */
4852: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, rj_waits, status));
4853: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, sj_waits, status));
4855: /* send and recv i-structure */
4856: PetscCall(PetscCommGetNewTag(comm, &tagi));
4857: PetscCall(PetscPostIrecvInt(comm, tagi, merge->nrecv, merge->id_r, len_ri, &buf_ri, &ri_waits));
4859: PetscCall(PetscMalloc1(len + 1, &buf_s));
4860: buf_si = buf_s; /* points to the beginning of k-th msg to be sent */
4861: for (PetscMPIInt proc = 0, k = 0; proc < size; proc++) {
4862: if (!len_s[proc]) continue;
4863: /* form outgoing message for i-structure:
4864: buf_si[0]: nrows to be sent
4865: [1:nrows]: row index (global)
4866: [nrows+1:2*nrows+1]: i-structure index
4867: */
4868: nrows = len_si[proc] / 2 - 1;
4869: buf_si_i = buf_si + nrows + 1;
4870: buf_si[0] = nrows;
4871: buf_si_i[0] = 0;
4872: nrows = 0;
4873: for (i = owners[proc]; i < owners[proc + 1]; i++) {
4874: anzi = ai[i + 1] - ai[i];
4875: if (anzi) {
4876: buf_si_i[nrows + 1] = buf_si_i[nrows] + anzi; /* i-structure */
4877: buf_si[nrows + 1] = i - owners[proc]; /* local row index */
4878: nrows++;
4879: }
4880: }
4881: PetscCallMPI(MPIU_Isend(buf_si, len_si[proc], MPIU_INT, proc, tagi, comm, si_waits + k));
4882: k++;
4883: buf_si += len_si[proc];
4884: }
4886: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, ri_waits, status));
4887: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, si_waits, status));
4889: PetscCall(PetscInfo(seqmat, "nsend: %d, nrecv: %d\n", merge->nsend, merge->nrecv));
4890: 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]));
4892: PetscCall(PetscFree(len_si));
4893: PetscCall(PetscFree(len_ri));
4894: PetscCall(PetscFree(rj_waits));
4895: PetscCall(PetscFree2(si_waits, sj_waits));
4896: PetscCall(PetscFree(ri_waits));
4897: PetscCall(PetscFree(buf_s));
4898: PetscCall(PetscFree(status));
4900: /* compute a local seq matrix in each processor */
4901: /* allocate bi array and free space for accumulating nonzero column info */
4902: PetscCall(PetscMalloc1(m + 1, &bi));
4903: bi[0] = 0;
4905: /* create and initialize a linked list */
4906: nlnk = N + 1;
4907: PetscCall(PetscLLCreate(N, N, nlnk, lnk, lnkbt));
4909: /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */
4910: len = ai[owners[rank + 1]] - ai[owners[rank]];
4911: PetscCall(PetscFreeSpaceGet(PetscIntMultTruncate(2, len) + 1, &free_space));
4913: current_space = free_space;
4915: /* determine symbolic info for each local row */
4916: PetscCall(PetscMalloc3(merge->nrecv, &buf_ri_k, merge->nrecv, &nextrow, merge->nrecv, &nextai));
4918: for (k = 0; k < merge->nrecv; k++) {
4919: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4920: nrows = *buf_ri_k[k];
4921: nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */
4922: nextai[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure */
4923: }
4925: MatPreallocateBegin(comm, m, n, dnz, onz);
4926: len = 0;
4927: for (i = 0; i < m; i++) {
4928: bnzi = 0;
4929: /* add local non-zero cols of this proc's seqmat into lnk */
4930: arow = owners[rank] + i;
4931: anzi = ai[arow + 1] - ai[arow];
4932: aj = a->j + ai[arow];
4933: PetscCall(PetscLLAddSorted(anzi, aj, N, &nlnk, lnk, lnkbt));
4934: bnzi += nlnk;
4935: /* add received col data into lnk */
4936: for (k = 0; k < merge->nrecv; k++) { /* k-th received message */
4937: if (i == *nextrow[k]) { /* i-th row */
4938: anzi = *(nextai[k] + 1) - *nextai[k];
4939: aj = buf_rj[k] + *nextai[k];
4940: PetscCall(PetscLLAddSorted(anzi, aj, N, &nlnk, lnk, lnkbt));
4941: bnzi += nlnk;
4942: nextrow[k]++;
4943: nextai[k]++;
4944: }
4945: }
4946: if (len < bnzi) len = bnzi; /* =max(bnzi) */
4948: /* if free space is not available, make more free space */
4949: if (current_space->local_remaining < bnzi) PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(bnzi, current_space->total_array_size), ¤t_space));
4950: /* copy data into free space, then initialize lnk */
4951: PetscCall(PetscLLClean(N, N, bnzi, lnk, current_space->array, lnkbt));
4952: PetscCall(MatPreallocateSet(i + owners[rank], bnzi, current_space->array, dnz, onz));
4954: current_space->array += bnzi;
4955: current_space->local_used += bnzi;
4956: current_space->local_remaining -= bnzi;
4958: bi[i + 1] = bi[i] + bnzi;
4959: }
4961: PetscCall(PetscFree3(buf_ri_k, nextrow, nextai));
4963: PetscCall(PetscMalloc1(bi[m], &bj));
4964: PetscCall(PetscFreeSpaceContiguous(&free_space, bj));
4965: PetscCall(PetscLLDestroy(lnk, lnkbt));
4967: /* create symbolic parallel matrix B_mpi */
4968: PetscCall(MatGetBlockSizes(seqmat, &bs, &cbs));
4969: PetscCall(MatCreate(comm, &B_mpi));
4970: if (n == PETSC_DECIDE) {
4971: PetscCall(MatSetSizes(B_mpi, m, n, PETSC_DETERMINE, N));
4972: } else {
4973: PetscCall(MatSetSizes(B_mpi, m, n, PETSC_DETERMINE, PETSC_DETERMINE));
4974: }
4975: PetscCall(MatSetBlockSizes(B_mpi, bs, cbs));
4976: PetscCall(MatSetType(B_mpi, MATMPIAIJ));
4977: PetscCall(MatMPIAIJSetPreallocation(B_mpi, 0, dnz, 0, onz));
4978: MatPreallocateEnd(dnz, onz);
4979: PetscCall(MatSetOption(B_mpi, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_FALSE));
4981: /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */
4982: B_mpi->assembled = PETSC_FALSE;
4983: merge->bi = bi;
4984: merge->bj = bj;
4985: merge->buf_ri = buf_ri;
4986: merge->buf_rj = buf_rj;
4987: merge->coi = NULL;
4988: merge->coj = NULL;
4989: merge->owners_co = NULL;
4991: PetscCall(PetscCommDestroy(&comm));
4993: /* attach the supporting struct to B_mpi for reuse */
4994: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
4995: PetscCall(PetscContainerSetPointer(container, merge));
4996: PetscCall(PetscContainerSetCtxDestroy(container, MatMergeSeqsToMPIDestroy));
4997: PetscCall(PetscObjectCompose((PetscObject)B_mpi, "MatMergeSeqsToMPI", (PetscObject)container));
4998: PetscCall(PetscContainerDestroy(&container));
4999: *mpimat = B_mpi;
5001: PetscCall(PetscLogEventEnd(MAT_Seqstompisym, seqmat, 0, 0, 0));
5002: PetscFunctionReturn(PETSC_SUCCESS);
5003: }
5005: /*@
5006: MatCreateMPIAIJSumSeqAIJ - Creates a `MATMPIAIJ` matrix by adding sequential
5007: matrices from each processor
5009: Collective
5011: Input Parameters:
5012: + comm - the communicators the parallel matrix will live on
5013: . seqmat - the input sequential matrices
5014: . m - number of local rows (or `PETSC_DECIDE`)
5015: . n - number of local columns (or `PETSC_DECIDE`)
5016: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5018: Output Parameter:
5019: . mpimat - the parallel matrix generated
5021: Level: advanced
5023: Note:
5024: The dimensions of the sequential matrix in each processor MUST be the same.
5025: The input seqmat is included into the container `MatMergeSeqsToMPIDestroy`, and will be
5026: destroyed when `mpimat` is destroyed. Call `PetscObjectQuery()` to access `seqmat`.
5028: .seealso: [](ch_matrices), `Mat`, `MatCreateAIJ()`
5029: @*/
5030: PetscErrorCode MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm, Mat seqmat, PetscInt m, PetscInt n, MatReuse scall, Mat *mpimat)
5031: {
5032: PetscMPIInt size;
5034: PetscFunctionBegin;
5035: PetscCallMPI(MPI_Comm_size(comm, &size));
5036: if (size == 1) {
5037: PetscCall(PetscLogEventBegin(MAT_Seqstompi, seqmat, 0, 0, 0));
5038: if (scall == MAT_INITIAL_MATRIX) {
5039: PetscCall(MatDuplicate(seqmat, MAT_COPY_VALUES, mpimat));
5040: } else {
5041: PetscCall(MatCopy(seqmat, *mpimat, SAME_NONZERO_PATTERN));
5042: }
5043: PetscCall(PetscLogEventEnd(MAT_Seqstompi, seqmat, 0, 0, 0));
5044: PetscFunctionReturn(PETSC_SUCCESS);
5045: }
5046: PetscCall(PetscLogEventBegin(MAT_Seqstompi, seqmat, 0, 0, 0));
5047: if (scall == MAT_INITIAL_MATRIX) PetscCall(MatCreateMPIAIJSumSeqAIJSymbolic(comm, seqmat, m, n, mpimat));
5048: PetscCall(MatCreateMPIAIJSumSeqAIJNumeric(seqmat, *mpimat));
5049: PetscCall(PetscLogEventEnd(MAT_Seqstompi, seqmat, 0, 0, 0));
5050: PetscFunctionReturn(PETSC_SUCCESS);
5051: }
5053: /*@
5054: MatAIJGetLocalMat - Creates a `MATSEQAIJ` from a `MATAIJ` matrix.
5056: Not Collective
5058: Input Parameter:
5059: . A - the matrix
5061: Output Parameter:
5062: . A_loc - the local sequential matrix generated
5064: Level: developer
5066: Notes:
5067: The matrix is created by taking `A`'s local rows and putting them into a sequential matrix
5068: with `mlocal` rows and `n` columns. Where `mlocal` is obtained with `MatGetLocalSize()` and
5069: `n` is the global column count obtained with `MatGetSize()`
5071: In other words combines the two parts of a parallel `MATMPIAIJ` matrix on each process to a single matrix.
5073: For parallel matrices this creates an entirely new matrix. If the matrix is sequential it merely increases the reference count.
5075: Destroy the matrix with `MatDestroy()`
5077: .seealso: [](ch_matrices), `Mat`, `MatMPIAIJGetLocalMat()`
5078: @*/
5079: PetscErrorCode MatAIJGetLocalMat(Mat A, Mat *A_loc)
5080: {
5081: PetscBool mpi;
5083: PetscFunctionBegin;
5084: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIAIJ, &mpi));
5085: if (mpi) {
5086: PetscCall(MatMPIAIJGetLocalMat(A, MAT_INITIAL_MATRIX, A_loc));
5087: } else {
5088: *A_loc = A;
5089: PetscCall(PetscObjectReference((PetscObject)*A_loc));
5090: }
5091: PetscFunctionReturn(PETSC_SUCCESS);
5092: }
5094: /*@
5095: MatMPIAIJGetLocalMat - Creates a `MATSEQAIJ` from a `MATMPIAIJ` matrix.
5097: Not Collective
5099: Input Parameters:
5100: + A - the matrix
5101: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5103: Output Parameter:
5104: . A_loc - the local sequential matrix generated
5106: Level: developer
5108: Notes:
5109: The matrix is created by taking all `A`'s local rows and putting them into a sequential
5110: matrix with `mlocal` rows and `n` columns.`mlocal` is the row count obtained with
5111: `MatGetLocalSize()` and `n` is the global column count obtained with `MatGetSize()`.
5113: In other words combines the two parts of a parallel `MATMPIAIJ` matrix on each process to a single matrix.
5115: When `A` is sequential and `MAT_INITIAL_MATRIX` is requested, the matrix returned is the diagonal part of `A` (which contains the entire matrix),
5116: 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
5117: then `MatCopy`(Adiag,*`A_loc`,`SAME_NONZERO_PATTERN`) is called to fill `A_loc`. Thus one can preallocate the appropriate sequential matrix `A_loc`
5118: 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.
5120: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMatCondensed()`, `MatMPIAIJGetLocalMatMerge()`
5121: @*/
5122: PetscErrorCode MatMPIAIJGetLocalMat(Mat A, MatReuse scall, Mat *A_loc)
5123: {
5124: Mat_MPIAIJ *mpimat = (Mat_MPIAIJ *)A->data;
5125: Mat_SeqAIJ *mat, *a, *b;
5126: PetscInt *ai, *aj, *bi, *bj, *cmap = mpimat->garray;
5127: const PetscScalar *aa, *ba, *aav, *bav;
5128: PetscScalar *ca, *cam;
5129: PetscMPIInt size;
5130: PetscInt am = A->rmap->n, i, j, k, cstart = A->cmap->rstart;
5131: PetscInt *ci, *cj, col, ncols_d, ncols_o, jo;
5132: PetscBool match;
5134: PetscFunctionBegin;
5135: PetscCall(PetscStrbeginswith(((PetscObject)A)->type_name, MATMPIAIJ, &match));
5136: PetscCheck(match, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Requires MATMPIAIJ matrix as input");
5137: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
5138: if (size == 1) {
5139: if (scall == MAT_INITIAL_MATRIX) {
5140: PetscCall(PetscObjectReference((PetscObject)mpimat->A));
5141: *A_loc = mpimat->A;
5142: } else if (scall == MAT_REUSE_MATRIX) {
5143: PetscCall(MatCopy(mpimat->A, *A_loc, SAME_NONZERO_PATTERN));
5144: }
5145: PetscFunctionReturn(PETSC_SUCCESS);
5146: }
5148: PetscCall(PetscLogEventBegin(MAT_Getlocalmat, A, 0, 0, 0));
5149: a = (Mat_SeqAIJ *)mpimat->A->data;
5150: b = (Mat_SeqAIJ *)mpimat->B->data;
5151: ai = a->i;
5152: aj = a->j;
5153: bi = b->i;
5154: bj = b->j;
5155: PetscCall(MatSeqAIJGetArrayRead(mpimat->A, &aav));
5156: PetscCall(MatSeqAIJGetArrayRead(mpimat->B, &bav));
5157: aa = aav;
5158: ba = bav;
5159: if (scall == MAT_INITIAL_MATRIX) {
5160: PetscCall(PetscMalloc1(1 + am, &ci));
5161: ci[0] = 0;
5162: for (i = 0; i < am; i++) ci[i + 1] = ci[i] + (ai[i + 1] - ai[i]) + (bi[i + 1] - bi[i]);
5163: PetscCall(PetscMalloc1(1 + ci[am], &cj));
5164: PetscCall(PetscMalloc1(1 + ci[am], &ca));
5165: k = 0;
5166: for (i = 0; i < am; i++) {
5167: ncols_o = bi[i + 1] - bi[i];
5168: ncols_d = ai[i + 1] - ai[i];
5169: /* off-diagonal portion of A */
5170: for (jo = 0; jo < ncols_o; jo++) {
5171: col = cmap[*bj];
5172: if (col >= cstart) break;
5173: cj[k] = col;
5174: bj++;
5175: ca[k++] = *ba++;
5176: }
5177: /* diagonal portion of A */
5178: for (j = 0; j < ncols_d; j++) {
5179: cj[k] = cstart + *aj++;
5180: ca[k++] = *aa++;
5181: }
5182: /* off-diagonal portion of A */
5183: for (j = jo; j < ncols_o; j++) {
5184: cj[k] = cmap[*bj++];
5185: ca[k++] = *ba++;
5186: }
5187: }
5188: /* put together the new matrix */
5189: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, am, A->cmap->N, ci, cj, ca, A_loc));
5190: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5191: /* Since these are PETSc arrays, change flags to free them as necessary. */
5192: mat = (Mat_SeqAIJ *)(*A_loc)->data;
5193: mat->free_a = PETSC_TRUE;
5194: mat->free_ij = PETSC_TRUE;
5195: mat->nonew = 0;
5196: } else if (scall == MAT_REUSE_MATRIX) {
5197: mat = (Mat_SeqAIJ *)(*A_loc)->data;
5198: ci = mat->i;
5199: cj = mat->j;
5200: PetscCall(MatSeqAIJGetArrayWrite(*A_loc, &cam));
5201: for (i = 0; i < am; i++) {
5202: /* off-diagonal portion of A */
5203: ncols_o = bi[i + 1] - bi[i];
5204: for (jo = 0; jo < ncols_o; jo++) {
5205: col = cmap[*bj];
5206: if (col >= cstart) break;
5207: *cam++ = *ba++;
5208: bj++;
5209: }
5210: /* diagonal portion of A */
5211: ncols_d = ai[i + 1] - ai[i];
5212: for (j = 0; j < ncols_d; j++) *cam++ = *aa++;
5213: /* off-diagonal portion of A */
5214: for (j = jo; j < ncols_o; j++) {
5215: *cam++ = *ba++;
5216: bj++;
5217: }
5218: }
5219: PetscCall(MatSeqAIJRestoreArrayWrite(*A_loc, &cam));
5220: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Invalid MatReuse %d", (int)scall);
5221: PetscCall(MatSeqAIJRestoreArrayRead(mpimat->A, &aav));
5222: PetscCall(MatSeqAIJRestoreArrayRead(mpimat->B, &bav));
5223: PetscCall(PetscLogEventEnd(MAT_Getlocalmat, A, 0, 0, 0));
5224: PetscFunctionReturn(PETSC_SUCCESS);
5225: }
5227: /*@
5228: MatMPIAIJGetLocalMatMerge - Creates a `MATSEQAIJ` from a `MATMPIAIJ` matrix by taking all its local rows and putting them into a sequential matrix with
5229: mlocal rows and n columns. Where n is the sum of the number of columns of the diagonal and off-diagonal part
5231: Not Collective
5233: Input Parameters:
5234: + A - the matrix
5235: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5237: Output Parameters:
5238: + glob - sequential `IS` with global indices associated with the columns of the local sequential matrix generated (can be `NULL`)
5239: - A_loc - the local sequential matrix generated
5241: Level: developer
5243: Note:
5244: This is different from `MatMPIAIJGetLocalMat()` since the first columns in the returning matrix are those associated with the diagonal
5245: part, then those associated with the off-diagonal part (in its local ordering)
5247: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMat()`, `MatMPIAIJGetLocalMatCondensed()`
5248: @*/
5249: PetscErrorCode MatMPIAIJGetLocalMatMerge(Mat A, MatReuse scall, IS *glob, Mat *A_loc)
5250: {
5251: Mat Ao, Ad;
5252: const PetscInt *cmap;
5253: PetscMPIInt size;
5254: PetscErrorCode (*f)(Mat, MatReuse, IS *, Mat *);
5256: PetscFunctionBegin;
5257: PetscCall(MatMPIAIJGetSeqAIJ(A, &Ad, &Ao, &cmap));
5258: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
5259: if (size == 1) {
5260: if (scall == MAT_INITIAL_MATRIX) {
5261: PetscCall(PetscObjectReference((PetscObject)Ad));
5262: *A_loc = Ad;
5263: } else if (scall == MAT_REUSE_MATRIX) {
5264: PetscCall(MatCopy(Ad, *A_loc, SAME_NONZERO_PATTERN));
5265: }
5266: if (glob) PetscCall(ISCreateStride(PetscObjectComm((PetscObject)Ad), Ad->cmap->n, Ad->cmap->rstart, 1, glob));
5267: PetscFunctionReturn(PETSC_SUCCESS);
5268: }
5269: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatMPIAIJGetLocalMatMerge_C", &f));
5270: PetscCall(PetscLogEventBegin(MAT_Getlocalmat, A, 0, 0, 0));
5271: if (f) {
5272: PetscCall((*f)(A, scall, glob, A_loc));
5273: } else {
5274: Mat_SeqAIJ *a = (Mat_SeqAIJ *)Ad->data;
5275: Mat_SeqAIJ *b = (Mat_SeqAIJ *)Ao->data;
5276: Mat_SeqAIJ *c;
5277: PetscInt *ai = a->i, *aj = a->j;
5278: PetscInt *bi = b->i, *bj = b->j;
5279: PetscInt *ci, *cj;
5280: const PetscScalar *aa, *ba;
5281: PetscScalar *ca;
5282: PetscInt i, j, am, dn, on;
5284: PetscCall(MatGetLocalSize(Ad, &am, &dn));
5285: PetscCall(MatGetLocalSize(Ao, NULL, &on));
5286: PetscCall(MatSeqAIJGetArrayRead(Ad, &aa));
5287: PetscCall(MatSeqAIJGetArrayRead(Ao, &ba));
5288: if (scall == MAT_INITIAL_MATRIX) {
5289: PetscInt k;
5290: PetscCall(PetscMalloc1(1 + am, &ci));
5291: PetscCall(PetscMalloc1(ai[am] + bi[am], &cj));
5292: PetscCall(PetscMalloc1(ai[am] + bi[am], &ca));
5293: ci[0] = 0;
5294: for (i = 0, k = 0; i < am; i++) {
5295: const PetscInt ncols_o = bi[i + 1] - bi[i];
5296: const PetscInt ncols_d = ai[i + 1] - ai[i];
5297: ci[i + 1] = ci[i] + ncols_o + ncols_d;
5298: /* diagonal portion of A */
5299: for (j = 0; j < ncols_d; j++, k++) {
5300: cj[k] = *aj++;
5301: ca[k] = *aa++;
5302: }
5303: /* off-diagonal portion of A */
5304: for (j = 0; j < ncols_o; j++, k++) {
5305: cj[k] = dn + *bj++;
5306: ca[k] = *ba++;
5307: }
5308: }
5309: /* put together the new matrix */
5310: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, am, dn + on, ci, cj, ca, A_loc));
5311: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5312: /* Since these are PETSc arrays, change flags to free them as necessary. */
5313: c = (Mat_SeqAIJ *)(*A_loc)->data;
5314: c->free_a = PETSC_TRUE;
5315: c->free_ij = PETSC_TRUE;
5316: c->nonew = 0;
5317: PetscCall(MatSetType(*A_loc, ((PetscObject)Ad)->type_name));
5318: } else if (scall == MAT_REUSE_MATRIX) {
5319: PetscCall(MatSeqAIJGetArrayWrite(*A_loc, &ca));
5320: for (i = 0; i < am; i++) {
5321: const PetscInt ncols_d = ai[i + 1] - ai[i];
5322: const PetscInt ncols_o = bi[i + 1] - bi[i];
5323: /* diagonal portion of A */
5324: for (j = 0; j < ncols_d; j++) *ca++ = *aa++;
5325: /* off-diagonal portion of A */
5326: for (j = 0; j < ncols_o; j++) *ca++ = *ba++;
5327: }
5328: PetscCall(MatSeqAIJRestoreArrayWrite(*A_loc, &ca));
5329: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Invalid MatReuse %d", (int)scall);
5330: PetscCall(MatSeqAIJRestoreArrayRead(Ad, &aa));
5331: PetscCall(MatSeqAIJRestoreArrayRead(Ao, &aa));
5332: if (glob) {
5333: PetscInt cst, *gidx;
5335: PetscCall(MatGetOwnershipRangeColumn(A, &cst, NULL));
5336: PetscCall(PetscMalloc1(dn + on, &gidx));
5337: for (i = 0; i < dn; i++) gidx[i] = cst + i;
5338: for (i = 0; i < on; i++) gidx[i + dn] = cmap[i];
5339: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)Ad), dn + on, gidx, PETSC_OWN_POINTER, glob));
5340: }
5341: }
5342: PetscCall(PetscLogEventEnd(MAT_Getlocalmat, A, 0, 0, 0));
5343: PetscFunctionReturn(PETSC_SUCCESS);
5344: }
5346: /*@C
5347: MatMPIAIJGetLocalMatCondensed - Creates a `MATSEQAIJ` matrix from an `MATMPIAIJ` matrix by taking all its local rows and NON-ZERO columns
5349: Not Collective
5351: Input Parameters:
5352: + A - the matrix
5353: . scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5354: . row - index set of rows to extract (or `NULL`)
5355: - col - index set of columns to extract (or `NULL`)
5357: Output Parameter:
5358: . A_loc - the local sequential matrix generated
5360: Level: developer
5362: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMat()`
5363: @*/
5364: PetscErrorCode MatMPIAIJGetLocalMatCondensed(Mat A, MatReuse scall, IS *row, IS *col, Mat *A_loc)
5365: {
5366: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5367: PetscInt i, start, end, ncols, nzA, nzB, *cmap, imark, *idx;
5368: IS isrowa, iscola;
5369: Mat *aloc;
5370: PetscBool match;
5372: PetscFunctionBegin;
5373: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIAIJ, &match));
5374: PetscCheck(match, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Requires MATMPIAIJ matrix as input");
5375: PetscCall(PetscLogEventBegin(MAT_Getlocalmatcondensed, A, 0, 0, 0));
5376: if (!row) {
5377: start = A->rmap->rstart;
5378: end = A->rmap->rend;
5379: PetscCall(ISCreateStride(PETSC_COMM_SELF, end - start, start, 1, &isrowa));
5380: } else {
5381: isrowa = *row;
5382: }
5383: if (!col) {
5384: start = A->cmap->rstart;
5385: cmap = a->garray;
5386: nzA = a->A->cmap->n;
5387: nzB = a->B->cmap->n;
5388: PetscCall(PetscMalloc1(nzA + nzB, &idx));
5389: ncols = 0;
5390: for (i = 0; i < nzB; i++) {
5391: if (cmap[i] < start) idx[ncols++] = cmap[i];
5392: else break;
5393: }
5394: imark = i;
5395: for (i = 0; i < nzA; i++) idx[ncols++] = start + i;
5396: for (i = imark; i < nzB; i++) idx[ncols++] = cmap[i];
5397: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, &iscola));
5398: } else {
5399: iscola = *col;
5400: }
5401: if (scall != MAT_INITIAL_MATRIX) {
5402: PetscCall(PetscMalloc1(1, &aloc));
5403: aloc[0] = *A_loc;
5404: }
5405: PetscCall(MatCreateSubMatrices(A, 1, &isrowa, &iscola, scall, &aloc));
5406: if (!col) { /* attach global id of condensed columns */
5407: PetscCall(PetscObjectCompose((PetscObject)aloc[0], "_petsc_GetLocalMatCondensed_iscol", (PetscObject)iscola));
5408: }
5409: *A_loc = aloc[0];
5410: PetscCall(PetscFree(aloc));
5411: if (!row) PetscCall(ISDestroy(&isrowa));
5412: if (!col) PetscCall(ISDestroy(&iscola));
5413: PetscCall(PetscLogEventEnd(MAT_Getlocalmatcondensed, A, 0, 0, 0));
5414: PetscFunctionReturn(PETSC_SUCCESS);
5415: }
5417: /*
5418: * Create a sequential AIJ matrix based on row indices. a whole column is extracted once a row is matched.
5419: * Row could be local or remote.The routine is designed to be scalable in memory so that nothing is based
5420: * on a global size.
5421: * */
5422: static PetscErrorCode MatCreateSeqSubMatrixWithRows_Private(Mat P, IS rows, Mat *P_oth)
5423: {
5424: Mat_MPIAIJ *p = (Mat_MPIAIJ *)P->data;
5425: Mat_SeqAIJ *pd = (Mat_SeqAIJ *)p->A->data, *po = (Mat_SeqAIJ *)p->B->data, *p_oth;
5426: PetscInt plocalsize, nrows, *ilocal, *oilocal, i, lidx, *nrcols, *nlcols, ncol;
5427: PetscMPIInt owner;
5428: PetscSFNode *iremote, *oiremote;
5429: const PetscInt *lrowindices;
5430: PetscSF sf, osf;
5431: PetscInt pcstart, *roffsets, *loffsets, *pnnz, j;
5432: PetscInt ontotalcols, dntotalcols, ntotalcols, nout;
5433: MPI_Comm comm;
5434: ISLocalToGlobalMapping mapping;
5435: const PetscScalar *pd_a, *po_a;
5437: PetscFunctionBegin;
5438: PetscCall(PetscObjectGetComm((PetscObject)P, &comm));
5439: /* plocalsize is the number of roots
5440: * nrows is the number of leaves
5441: * */
5442: PetscCall(MatGetLocalSize(P, &plocalsize, NULL));
5443: PetscCall(ISGetLocalSize(rows, &nrows));
5444: PetscCall(PetscCalloc1(nrows, &iremote));
5445: PetscCall(ISGetIndices(rows, &lrowindices));
5446: for (i = 0; i < nrows; i++) {
5447: /* Find a remote index and an owner for a row
5448: * The row could be local or remote
5449: * */
5450: owner = 0;
5451: lidx = 0;
5452: PetscCall(PetscLayoutFindOwnerIndex(P->rmap, lrowindices[i], &owner, &lidx));
5453: iremote[i].index = lidx;
5454: iremote[i].rank = owner;
5455: }
5456: /* Create SF to communicate how many nonzero columns for each row */
5457: PetscCall(PetscSFCreate(comm, &sf));
5458: /* SF will figure out the number of nonzero columns for each row, and their
5459: * offsets
5460: * */
5461: PetscCall(PetscSFSetGraph(sf, plocalsize, nrows, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
5462: PetscCall(PetscSFSetFromOptions(sf));
5463: PetscCall(PetscSFSetUp(sf));
5465: PetscCall(PetscCalloc1(2 * (plocalsize + 1), &roffsets));
5466: PetscCall(PetscCalloc1(2 * plocalsize, &nrcols));
5467: PetscCall(PetscCalloc1(nrows, &pnnz));
5468: roffsets[0] = 0;
5469: roffsets[1] = 0;
5470: for (i = 0; i < plocalsize; i++) {
5471: /* diagonal */
5472: nrcols[i * 2 + 0] = pd->i[i + 1] - pd->i[i];
5473: /* off-diagonal */
5474: nrcols[i * 2 + 1] = po->i[i + 1] - po->i[i];
5475: /* compute offsets so that we relative location for each row */
5476: roffsets[(i + 1) * 2 + 0] = roffsets[i * 2 + 0] + nrcols[i * 2 + 0];
5477: roffsets[(i + 1) * 2 + 1] = roffsets[i * 2 + 1] + nrcols[i * 2 + 1];
5478: }
5479: PetscCall(PetscCalloc1(2 * nrows, &nlcols));
5480: PetscCall(PetscCalloc1(2 * nrows, &loffsets));
5481: /* 'r' means root, and 'l' means leaf */
5482: PetscCall(PetscSFBcastBegin(sf, MPIU_2INT, nrcols, nlcols, MPI_REPLACE));
5483: PetscCall(PetscSFBcastBegin(sf, MPIU_2INT, roffsets, loffsets, MPI_REPLACE));
5484: PetscCall(PetscSFBcastEnd(sf, MPIU_2INT, nrcols, nlcols, MPI_REPLACE));
5485: PetscCall(PetscSFBcastEnd(sf, MPIU_2INT, roffsets, loffsets, MPI_REPLACE));
5486: PetscCall(PetscSFDestroy(&sf));
5487: PetscCall(PetscFree(roffsets));
5488: PetscCall(PetscFree(nrcols));
5489: dntotalcols = 0;
5490: ontotalcols = 0;
5491: ncol = 0;
5492: for (i = 0; i < nrows; i++) {
5493: pnnz[i] = nlcols[i * 2 + 0] + nlcols[i * 2 + 1];
5494: ncol = PetscMax(pnnz[i], ncol);
5495: /* diagonal */
5496: dntotalcols += nlcols[i * 2 + 0];
5497: /* off-diagonal */
5498: ontotalcols += nlcols[i * 2 + 1];
5499: }
5500: /* We do not need to figure the right number of columns
5501: * since all the calculations will be done by going through the raw data
5502: * */
5503: PetscCall(MatCreateSeqAIJ(PETSC_COMM_SELF, nrows, ncol, 0, pnnz, P_oth));
5504: PetscCall(MatSetUp(*P_oth));
5505: PetscCall(PetscFree(pnnz));
5506: p_oth = (Mat_SeqAIJ *)(*P_oth)->data;
5507: /* diagonal */
5508: PetscCall(PetscCalloc1(dntotalcols, &iremote));
5509: /* off-diagonal */
5510: PetscCall(PetscCalloc1(ontotalcols, &oiremote));
5511: /* diagonal */
5512: PetscCall(PetscCalloc1(dntotalcols, &ilocal));
5513: /* off-diagonal */
5514: PetscCall(PetscCalloc1(ontotalcols, &oilocal));
5515: dntotalcols = 0;
5516: ontotalcols = 0;
5517: ntotalcols = 0;
5518: for (i = 0; i < nrows; i++) {
5519: owner = 0;
5520: PetscCall(PetscLayoutFindOwnerIndex(P->rmap, lrowindices[i], &owner, NULL));
5521: /* Set iremote for diag matrix */
5522: for (j = 0; j < nlcols[i * 2 + 0]; j++) {
5523: iremote[dntotalcols].index = loffsets[i * 2 + 0] + j;
5524: iremote[dntotalcols].rank = owner;
5525: /* P_oth is seqAIJ so that ilocal need to point to the first part of memory */
5526: ilocal[dntotalcols++] = ntotalcols++;
5527: }
5528: /* off-diagonal */
5529: for (j = 0; j < nlcols[i * 2 + 1]; j++) {
5530: oiremote[ontotalcols].index = loffsets[i * 2 + 1] + j;
5531: oiremote[ontotalcols].rank = owner;
5532: oilocal[ontotalcols++] = ntotalcols++;
5533: }
5534: }
5535: PetscCall(ISRestoreIndices(rows, &lrowindices));
5536: PetscCall(PetscFree(loffsets));
5537: PetscCall(PetscFree(nlcols));
5538: PetscCall(PetscSFCreate(comm, &sf));
5539: /* P serves as roots and P_oth is leaves
5540: * Diag matrix
5541: * */
5542: PetscCall(PetscSFSetGraph(sf, pd->i[plocalsize], dntotalcols, ilocal, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
5543: PetscCall(PetscSFSetFromOptions(sf));
5544: PetscCall(PetscSFSetUp(sf));
5546: PetscCall(PetscSFCreate(comm, &osf));
5547: /* off-diagonal */
5548: PetscCall(PetscSFSetGraph(osf, po->i[plocalsize], ontotalcols, oilocal, PETSC_OWN_POINTER, oiremote, PETSC_OWN_POINTER));
5549: PetscCall(PetscSFSetFromOptions(osf));
5550: PetscCall(PetscSFSetUp(osf));
5551: PetscCall(MatSeqAIJGetArrayRead(p->A, &pd_a));
5552: PetscCall(MatSeqAIJGetArrayRead(p->B, &po_a));
5553: /* operate on the matrix internal data to save memory */
5554: PetscCall(PetscSFBcastBegin(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5555: PetscCall(PetscSFBcastBegin(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5556: PetscCall(MatGetOwnershipRangeColumn(P, &pcstart, NULL));
5557: /* Convert to global indices for diag matrix */
5558: for (i = 0; i < pd->i[plocalsize]; i++) pd->j[i] += pcstart;
5559: PetscCall(PetscSFBcastBegin(sf, MPIU_INT, pd->j, p_oth->j, MPI_REPLACE));
5560: /* We want P_oth store global indices */
5561: PetscCall(ISLocalToGlobalMappingCreate(comm, 1, p->B->cmap->n, p->garray, PETSC_COPY_VALUES, &mapping));
5562: /* Use memory scalable approach */
5563: PetscCall(ISLocalToGlobalMappingSetType(mapping, ISLOCALTOGLOBALMAPPINGHASH));
5564: PetscCall(ISLocalToGlobalMappingApply(mapping, po->i[plocalsize], po->j, po->j));
5565: PetscCall(PetscSFBcastBegin(osf, MPIU_INT, po->j, p_oth->j, MPI_REPLACE));
5566: PetscCall(PetscSFBcastEnd(sf, MPIU_INT, pd->j, p_oth->j, MPI_REPLACE));
5567: /* Convert back to local indices */
5568: for (i = 0; i < pd->i[plocalsize]; i++) pd->j[i] -= pcstart;
5569: PetscCall(PetscSFBcastEnd(osf, MPIU_INT, po->j, p_oth->j, MPI_REPLACE));
5570: nout = 0;
5571: PetscCall(ISGlobalToLocalMappingApply(mapping, IS_GTOLM_DROP, po->i[plocalsize], po->j, &nout, po->j));
5572: PetscCheck(nout == po->i[plocalsize], comm, PETSC_ERR_ARG_INCOMP, "n %" PetscInt_FMT " does not equal to nout %" PetscInt_FMT " ", po->i[plocalsize], nout);
5573: PetscCall(ISLocalToGlobalMappingDestroy(&mapping));
5574: /* Exchange values */
5575: PetscCall(PetscSFBcastEnd(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5576: PetscCall(PetscSFBcastEnd(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5577: PetscCall(MatSeqAIJRestoreArrayRead(p->A, &pd_a));
5578: PetscCall(MatSeqAIJRestoreArrayRead(p->B, &po_a));
5579: /* Stop PETSc from shrinking memory */
5580: for (i = 0; i < nrows; i++) p_oth->ilen[i] = p_oth->imax[i];
5581: PetscCall(MatAssemblyBegin(*P_oth, MAT_FINAL_ASSEMBLY));
5582: PetscCall(MatAssemblyEnd(*P_oth, MAT_FINAL_ASSEMBLY));
5583: /* Attach PetscSF objects to P_oth so that we can reuse it later */
5584: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "diagsf", (PetscObject)sf));
5585: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "offdiagsf", (PetscObject)osf));
5586: PetscCall(PetscSFDestroy(&sf));
5587: PetscCall(PetscSFDestroy(&osf));
5588: PetscFunctionReturn(PETSC_SUCCESS);
5589: }
5591: /*
5592: * Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A
5593: * This supports MPIAIJ and MAIJ
5594: * */
5595: PetscErrorCode MatGetBrowsOfAcols_MPIXAIJ(Mat A, Mat P, PetscInt dof, MatReuse reuse, Mat *P_oth)
5596: {
5597: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data, *p = (Mat_MPIAIJ *)P->data;
5598: Mat_SeqAIJ *p_oth;
5599: IS rows, map;
5600: PetscHMapI hamp;
5601: PetscInt i, htsize, *rowindices, off, *mapping, key, count;
5602: MPI_Comm comm;
5603: PetscSF sf, osf;
5604: PetscBool has;
5606: PetscFunctionBegin;
5607: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
5608: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAocols, A, P, 0, 0));
5609: /* If it is the first time, create an index set of off-diag nonzero columns of A,
5610: * and then create a submatrix (that often is an overlapping matrix)
5611: * */
5612: if (reuse == MAT_INITIAL_MATRIX) {
5613: /* Use a hash table to figure out unique keys */
5614: PetscCall(PetscHMapICreateWithSize(a->B->cmap->n, &hamp));
5615: PetscCall(PetscCalloc1(a->B->cmap->n, &mapping));
5616: count = 0;
5617: /* Assume that a->g is sorted, otherwise the following does not make sense */
5618: for (i = 0; i < a->B->cmap->n; i++) {
5619: key = a->garray[i] / dof;
5620: PetscCall(PetscHMapIHas(hamp, key, &has));
5621: if (!has) {
5622: mapping[i] = count;
5623: PetscCall(PetscHMapISet(hamp, key, count++));
5624: } else {
5625: /* Current 'i' has the same value the previous step */
5626: mapping[i] = count - 1;
5627: }
5628: }
5629: PetscCall(ISCreateGeneral(comm, a->B->cmap->n, mapping, PETSC_OWN_POINTER, &map));
5630: PetscCall(PetscHMapIGetSize(hamp, &htsize));
5631: PetscCheck(htsize == count, comm, PETSC_ERR_ARG_INCOMP, " Size of hash map %" PetscInt_FMT " is inconsistent with count %" PetscInt_FMT, htsize, count);
5632: PetscCall(PetscCalloc1(htsize, &rowindices));
5633: off = 0;
5634: PetscCall(PetscHMapIGetKeys(hamp, &off, rowindices));
5635: PetscCall(PetscHMapIDestroy(&hamp));
5636: PetscCall(PetscSortInt(htsize, rowindices));
5637: PetscCall(ISCreateGeneral(comm, htsize, rowindices, PETSC_OWN_POINTER, &rows));
5638: /* In case, the matrix was already created but users want to recreate the matrix */
5639: PetscCall(MatDestroy(P_oth));
5640: PetscCall(MatCreateSeqSubMatrixWithRows_Private(P, rows, P_oth));
5641: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "aoffdiagtopothmapping", (PetscObject)map));
5642: PetscCall(ISDestroy(&map));
5643: PetscCall(ISDestroy(&rows));
5644: } else if (reuse == MAT_REUSE_MATRIX) {
5645: /* If matrix was already created, we simply update values using SF objects
5646: * that as attached to the matrix earlier.
5647: */
5648: const PetscScalar *pd_a, *po_a;
5650: PetscCall(PetscObjectQuery((PetscObject)*P_oth, "diagsf", (PetscObject *)&sf));
5651: PetscCall(PetscObjectQuery((PetscObject)*P_oth, "offdiagsf", (PetscObject *)&osf));
5652: PetscCheck(sf && osf, comm, PETSC_ERR_ARG_NULL, "Matrix is not initialized yet");
5653: p_oth = (Mat_SeqAIJ *)(*P_oth)->data;
5654: /* Update values in place */
5655: PetscCall(MatSeqAIJGetArrayRead(p->A, &pd_a));
5656: PetscCall(MatSeqAIJGetArrayRead(p->B, &po_a));
5657: PetscCall(PetscSFBcastBegin(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5658: PetscCall(PetscSFBcastBegin(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5659: PetscCall(PetscSFBcastEnd(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5660: PetscCall(PetscSFBcastEnd(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5661: PetscCall(MatSeqAIJRestoreArrayRead(p->A, &pd_a));
5662: PetscCall(MatSeqAIJRestoreArrayRead(p->B, &po_a));
5663: } else SETERRQ(comm, PETSC_ERR_ARG_UNKNOWN_TYPE, "Unknown reuse type");
5664: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAocols, A, P, 0, 0));
5665: PetscFunctionReturn(PETSC_SUCCESS);
5666: }
5668: /*@C
5669: MatGetBrowsOfAcols - Returns `IS` that contain rows of `B` that equal to nonzero columns of local `A`
5671: Collective
5673: Input Parameters:
5674: + A - the first matrix in `MATMPIAIJ` format
5675: . B - the second matrix in `MATMPIAIJ` format
5676: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5678: Output Parameters:
5679: + rowb - On input index sets of rows of B to extract (or `NULL`), modified on output
5680: . colb - On input index sets of columns of B to extract (or `NULL`), modified on output
5681: - B_seq - the sequential matrix generated
5683: Level: developer
5685: .seealso: `Mat`, `MATMPIAIJ`, `IS`, `MatReuse`
5686: @*/
5687: PetscErrorCode MatGetBrowsOfAcols(Mat A, Mat B, MatReuse scall, IS *rowb, IS *colb, Mat *B_seq)
5688: {
5689: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5690: PetscInt *idx, i, start, ncols, nzA, nzB, *cmap, imark;
5691: IS isrowb, iscolb;
5692: Mat *bseq = NULL;
5694: PetscFunctionBegin;
5695: 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 ")",
5696: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
5697: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAcols, A, B, 0, 0));
5699: if (scall == MAT_INITIAL_MATRIX) {
5700: start = A->cmap->rstart;
5701: cmap = a->garray;
5702: nzA = a->A->cmap->n;
5703: nzB = a->B->cmap->n;
5704: PetscCall(PetscMalloc1(nzA + nzB, &idx));
5705: ncols = 0;
5706: for (i = 0; i < nzB; i++) { /* row < local row index */
5707: if (cmap[i] < start) idx[ncols++] = cmap[i];
5708: else break;
5709: }
5710: imark = i;
5711: for (i = 0; i < nzA; i++) idx[ncols++] = start + i; /* local rows */
5712: for (i = imark; i < nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */
5713: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, &isrowb));
5714: PetscCall(ISCreateStride(PETSC_COMM_SELF, B->cmap->N, 0, 1, &iscolb));
5715: } else {
5716: PetscCheck(rowb && colb, PETSC_COMM_SELF, PETSC_ERR_SUP, "IS rowb and colb must be provided for MAT_REUSE_MATRIX");
5717: isrowb = *rowb;
5718: iscolb = *colb;
5719: PetscCall(PetscMalloc1(1, &bseq));
5720: bseq[0] = *B_seq;
5721: }
5722: PetscCall(MatCreateSubMatrices(B, 1, &isrowb, &iscolb, scall, &bseq));
5723: *B_seq = bseq[0];
5724: PetscCall(PetscFree(bseq));
5725: if (!rowb) {
5726: PetscCall(ISDestroy(&isrowb));
5727: } else {
5728: *rowb = isrowb;
5729: }
5730: if (!colb) {
5731: PetscCall(ISDestroy(&iscolb));
5732: } else {
5733: *colb = iscolb;
5734: }
5735: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAcols, A, B, 0, 0));
5736: PetscFunctionReturn(PETSC_SUCCESS);
5737: }
5739: /*
5740: MatGetBrowsOfAoCols_MPIAIJ - Creates a `MATSEQAIJ` matrix by taking rows of B that equal to nonzero columns
5741: of the OFF-DIAGONAL portion of local A
5743: Collective
5745: Input Parameters:
5746: + A,B - the matrices in `MATMPIAIJ` format
5747: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5749: Output Parameter:
5750: + startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or NULL)
5751: . startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or NULL)
5752: . bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or NULL)
5753: - B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N
5755: Developer Note:
5756: This directly accesses information inside the VecScatter associated with the matrix-vector product
5757: for this matrix. This is not desirable..
5759: Level: developer
5761: */
5763: PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat A, Mat B, MatReuse scall, PetscInt **startsj_s, PetscInt **startsj_r, MatScalar **bufa_ptr, Mat *B_oth)
5764: {
5765: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5766: VecScatter ctx;
5767: MPI_Comm comm;
5768: const PetscMPIInt *rprocs, *sprocs;
5769: PetscMPIInt nrecvs, nsends;
5770: const PetscInt *srow, *rstarts, *sstarts;
5771: PetscInt *rowlen, *bufj, *bufJ, ncols = 0, aBn = a->B->cmap->n, row, *b_othi, *b_othj, *rvalues = NULL, *svalues = NULL, *cols, sbs, rbs;
5772: PetscInt i, j, k = 0, l, ll, nrows, *rstartsj = NULL, *sstartsj, len;
5773: PetscScalar *b_otha, *bufa, *bufA, *vals = NULL;
5774: MPI_Request *reqs = NULL, *rwaits = NULL, *swaits = NULL;
5775: PetscMPIInt size, tag, rank, nreqs;
5777: PetscFunctionBegin;
5778: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
5779: PetscCallMPI(MPI_Comm_size(comm, &size));
5781: 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 ")",
5782: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
5783: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAocols, A, B, 0, 0));
5784: PetscCallMPI(MPI_Comm_rank(comm, &rank));
5786: if (size == 1) {
5787: startsj_s = NULL;
5788: bufa_ptr = NULL;
5789: *B_oth = NULL;
5790: PetscFunctionReturn(PETSC_SUCCESS);
5791: }
5793: ctx = a->Mvctx;
5794: tag = ((PetscObject)ctx)->tag;
5796: PetscCall(VecScatterGetRemote_Private(ctx, PETSC_TRUE /*send*/, &nsends, &sstarts, &srow, &sprocs, &sbs));
5797: /* rprocs[] must be ordered so that indices received from them are ordered in rvalues[], which is key to algorithms used in this subroutine */
5798: PetscCall(VecScatterGetRemoteOrdered_Private(ctx, PETSC_FALSE /*recv*/, &nrecvs, &rstarts, NULL /*indices not needed*/, &rprocs, &rbs));
5799: PetscCall(PetscMPIIntCast(nsends + nrecvs, &nreqs));
5800: PetscCall(PetscMalloc1(nreqs, &reqs));
5801: rwaits = reqs;
5802: swaits = PetscSafePointerPlusOffset(reqs, nrecvs);
5804: if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX;
5805: if (scall == MAT_INITIAL_MATRIX) {
5806: /* i-array */
5807: /* post receives */
5808: if (nrecvs) PetscCall(PetscMalloc1(rbs * (rstarts[nrecvs] - rstarts[0]), &rvalues)); /* rstarts can be NULL when nrecvs=0 */
5809: for (i = 0; i < nrecvs; i++) {
5810: rowlen = rvalues + rstarts[i] * rbs;
5811: nrows = (rstarts[i + 1] - rstarts[i]) * rbs; /* num of indices to be received */
5812: PetscCallMPI(MPIU_Irecv(rowlen, nrows, MPIU_INT, rprocs[i], tag, comm, rwaits + i));
5813: }
5815: /* pack the outgoing message */
5816: PetscCall(PetscMalloc2(nsends + 1, &sstartsj, nrecvs + 1, &rstartsj));
5818: sstartsj[0] = 0;
5819: rstartsj[0] = 0;
5820: len = 0; /* total length of j or a array to be sent */
5821: if (nsends) {
5822: k = sstarts[0]; /* ATTENTION: sstarts[0] and rstarts[0] are not necessarily zero */
5823: PetscCall(PetscMalloc1(sbs * (sstarts[nsends] - sstarts[0]), &svalues));
5824: }
5825: for (i = 0; i < nsends; i++) {
5826: rowlen = svalues + (sstarts[i] - sstarts[0]) * sbs;
5827: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5828: for (j = 0; j < nrows; j++) {
5829: row = srow[k] + B->rmap->range[rank]; /* global row idx */
5830: for (l = 0; l < sbs; l++) {
5831: PetscCall(MatGetRow_MPIAIJ(B, row + l, &ncols, NULL, NULL)); /* rowlength */
5833: rowlen[j * sbs + l] = ncols;
5835: len += ncols;
5836: PetscCall(MatRestoreRow_MPIAIJ(B, row + l, &ncols, NULL, NULL));
5837: }
5838: k++;
5839: }
5840: PetscCallMPI(MPIU_Isend(rowlen, nrows * sbs, MPIU_INT, sprocs[i], tag, comm, swaits + i));
5842: sstartsj[i + 1] = len; /* starting point of (i+1)-th outgoing msg in bufj and bufa */
5843: }
5844: /* recvs and sends of i-array are completed */
5845: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5846: PetscCall(PetscFree(svalues));
5848: /* allocate buffers for sending j and a arrays */
5849: PetscCall(PetscMalloc1(len, &bufj));
5850: PetscCall(PetscMalloc1(len, &bufa));
5852: /* create i-array of B_oth */
5853: PetscCall(PetscMalloc1(aBn + 1, &b_othi));
5855: b_othi[0] = 0;
5856: len = 0; /* total length of j or a array to be received */
5857: k = 0;
5858: for (i = 0; i < nrecvs; i++) {
5859: rowlen = rvalues + (rstarts[i] - rstarts[0]) * rbs;
5860: nrows = (rstarts[i + 1] - rstarts[i]) * rbs; /* num of rows to be received */
5861: for (j = 0; j < nrows; j++) {
5862: b_othi[k + 1] = b_othi[k] + rowlen[j];
5863: PetscCall(PetscIntSumError(rowlen[j], len, &len));
5864: k++;
5865: }
5866: rstartsj[i + 1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */
5867: }
5868: PetscCall(PetscFree(rvalues));
5870: /* allocate space for j and a arrays of B_oth */
5871: PetscCall(PetscMalloc1(b_othi[aBn], &b_othj));
5872: PetscCall(PetscMalloc1(b_othi[aBn], &b_otha));
5874: /* j-array */
5875: /* post receives of j-array */
5876: for (i = 0; i < nrecvs; i++) {
5877: nrows = rstartsj[i + 1] - rstartsj[i]; /* length of the msg received */
5878: PetscCallMPI(MPIU_Irecv(PetscSafePointerPlusOffset(b_othj, rstartsj[i]), nrows, MPIU_INT, rprocs[i], tag, comm, rwaits + i));
5879: }
5881: /* pack the outgoing message j-array */
5882: if (nsends) k = sstarts[0];
5883: for (i = 0; i < nsends; i++) {
5884: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5885: bufJ = PetscSafePointerPlusOffset(bufj, sstartsj[i]);
5886: for (j = 0; j < nrows; j++) {
5887: row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5888: for (ll = 0; ll < sbs; ll++) {
5889: PetscCall(MatGetRow_MPIAIJ(B, row + ll, &ncols, &cols, NULL));
5890: for (l = 0; l < ncols; l++) *bufJ++ = cols[l];
5891: PetscCall(MatRestoreRow_MPIAIJ(B, row + ll, &ncols, &cols, NULL));
5892: }
5893: }
5894: PetscCallMPI(MPIU_Isend(PetscSafePointerPlusOffset(bufj, sstartsj[i]), sstartsj[i + 1] - sstartsj[i], MPIU_INT, sprocs[i], tag, comm, swaits + i));
5895: }
5897: /* recvs and sends of j-array are completed */
5898: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5899: } else if (scall == MAT_REUSE_MATRIX) {
5900: sstartsj = *startsj_s;
5901: rstartsj = *startsj_r;
5902: bufa = *bufa_ptr;
5903: PetscCall(MatSeqAIJGetArrayWrite(*B_oth, &b_otha));
5904: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not possess an object container");
5906: /* a-array */
5907: /* post receives of a-array */
5908: for (i = 0; i < nrecvs; i++) {
5909: nrows = rstartsj[i + 1] - rstartsj[i]; /* length of the msg received */
5910: PetscCallMPI(MPIU_Irecv(PetscSafePointerPlusOffset(b_otha, rstartsj[i]), nrows, MPIU_SCALAR, rprocs[i], tag, comm, rwaits + i));
5911: }
5913: /* pack the outgoing message a-array */
5914: if (nsends) k = sstarts[0];
5915: for (i = 0; i < nsends; i++) {
5916: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5917: bufA = PetscSafePointerPlusOffset(bufa, sstartsj[i]);
5918: for (j = 0; j < nrows; j++) {
5919: row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5920: for (ll = 0; ll < sbs; ll++) {
5921: PetscCall(MatGetRow_MPIAIJ(B, row + ll, &ncols, NULL, &vals));
5922: for (l = 0; l < ncols; l++) *bufA++ = vals[l];
5923: PetscCall(MatRestoreRow_MPIAIJ(B, row + ll, &ncols, NULL, &vals));
5924: }
5925: }
5926: PetscCallMPI(MPIU_Isend(PetscSafePointerPlusOffset(bufa, sstartsj[i]), sstartsj[i + 1] - sstartsj[i], MPIU_SCALAR, sprocs[i], tag, comm, swaits + i));
5927: }
5928: /* recvs and sends of a-array are completed */
5929: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5930: PetscCall(PetscFree(reqs));
5932: if (scall == MAT_INITIAL_MATRIX) {
5933: Mat_SeqAIJ *b_oth;
5935: /* put together the new matrix */
5936: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, aBn, B->cmap->N, b_othi, b_othj, b_otha, B_oth));
5938: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5939: /* Since these are PETSc arrays, change flags to free them as necessary. */
5940: b_oth = (Mat_SeqAIJ *)(*B_oth)->data;
5941: b_oth->free_a = PETSC_TRUE;
5942: b_oth->free_ij = PETSC_TRUE;
5943: b_oth->nonew = 0;
5945: PetscCall(PetscFree(bufj));
5946: if (!startsj_s || !bufa_ptr) {
5947: PetscCall(PetscFree2(sstartsj, rstartsj));
5948: PetscCall(PetscFree(bufa_ptr));
5949: } else {
5950: *startsj_s = sstartsj;
5951: *startsj_r = rstartsj;
5952: *bufa_ptr = bufa;
5953: }
5954: } else if (scall == MAT_REUSE_MATRIX) {
5955: PetscCall(MatSeqAIJRestoreArrayWrite(*B_oth, &b_otha));
5956: }
5958: PetscCall(VecScatterRestoreRemote_Private(ctx, PETSC_TRUE, &nsends, &sstarts, &srow, &sprocs, &sbs));
5959: PetscCall(VecScatterRestoreRemoteOrdered_Private(ctx, PETSC_FALSE, &nrecvs, &rstarts, NULL, &rprocs, &rbs));
5960: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAocols, A, B, 0, 0));
5961: PetscFunctionReturn(PETSC_SUCCESS);
5962: }
5964: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat, MatType, MatReuse, Mat *);
5965: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJPERM(Mat, MatType, MatReuse, Mat *);
5966: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJSELL(Mat, MatType, MatReuse, Mat *);
5967: #if defined(PETSC_HAVE_MKL_SPARSE)
5968: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJMKL(Mat, MatType, MatReuse, Mat *);
5969: #endif
5970: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIBAIJ(Mat, MatType, MatReuse, Mat *);
5971: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISBAIJ(Mat, MatType, MatReuse, Mat *);
5972: #if defined(PETSC_HAVE_ELEMENTAL)
5973: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat, MatType, MatReuse, Mat *);
5974: #endif
5975: #if defined(PETSC_HAVE_SCALAPACK) && (defined(PETSC_USE_REAL_SINGLE) || defined(PETSC_USE_REAL_DOUBLE))
5976: PETSC_INTERN PetscErrorCode MatConvert_AIJ_ScaLAPACK(Mat, MatType, MatReuse, Mat *);
5977: #endif
5978: #if defined(PETSC_HAVE_HYPRE)
5979: PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat, MatType, MatReuse, Mat *);
5980: #endif
5981: #if defined(PETSC_HAVE_CUDA)
5982: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCUSPARSE(Mat, MatType, MatReuse, Mat *);
5983: #endif
5984: #if defined(PETSC_HAVE_HIP)
5985: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJHIPSPARSE(Mat, MatType, MatReuse, Mat *);
5986: #endif
5987: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
5988: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJKokkos(Mat, MatType, MatReuse, Mat *);
5989: #endif
5990: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISELL(Mat, MatType, MatReuse, Mat *);
5991: PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat, MatType, MatReuse, Mat *);
5992: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_IS_XAIJ(Mat);
5994: /*
5995: Computes (B'*A')' since computing B*A directly is untenable
5997: n p p
5998: [ ] [ ] [ ]
5999: m [ A ] * n [ B ] = m [ C ]
6000: [ ] [ ] [ ]
6002: */
6003: static PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A, Mat B, Mat C)
6004: {
6005: Mat At, Bt, Ct;
6007: PetscFunctionBegin;
6008: PetscCall(MatTranspose(A, MAT_INITIAL_MATRIX, &At));
6009: PetscCall(MatTranspose(B, MAT_INITIAL_MATRIX, &Bt));
6010: PetscCall(MatMatMult(Bt, At, MAT_INITIAL_MATRIX, PETSC_CURRENT, &Ct));
6011: PetscCall(MatDestroy(&At));
6012: PetscCall(MatDestroy(&Bt));
6013: PetscCall(MatTransposeSetPrecursor(Ct, C));
6014: PetscCall(MatTranspose(Ct, MAT_REUSE_MATRIX, &C));
6015: PetscCall(MatDestroy(&Ct));
6016: PetscFunctionReturn(PETSC_SUCCESS);
6017: }
6019: static PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A, Mat B, PetscReal fill, Mat C)
6020: {
6021: PetscBool cisdense;
6023: PetscFunctionBegin;
6024: 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);
6025: PetscCall(MatSetSizes(C, A->rmap->n, B->cmap->n, A->rmap->N, B->cmap->N));
6026: PetscCall(MatSetBlockSizesFromMats(C, A, B));
6027: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATMPIDENSE, MATMPIDENSECUDA, MATMPIDENSEHIP, ""));
6028: if (!cisdense) PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
6029: PetscCall(MatSetUp(C));
6031: C->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIAIJ;
6032: PetscFunctionReturn(PETSC_SUCCESS);
6033: }
6035: static PetscErrorCode MatProductSetFromOptions_MPIDense_MPIAIJ_AB(Mat C)
6036: {
6037: Mat_Product *product = C->product;
6038: Mat A = product->A, B = product->B;
6040: PetscFunctionBegin;
6041: 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 ")",
6042: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
6043: C->ops->matmultsymbolic = MatMatMultSymbolic_MPIDense_MPIAIJ;
6044: C->ops->productsymbolic = MatProductSymbolic_AB;
6045: PetscFunctionReturn(PETSC_SUCCESS);
6046: }
6048: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_MPIDense_MPIAIJ(Mat C)
6049: {
6050: Mat_Product *product = C->product;
6052: PetscFunctionBegin;
6053: if (product->type == MATPRODUCT_AB) PetscCall(MatProductSetFromOptions_MPIDense_MPIAIJ_AB(C));
6054: PetscFunctionReturn(PETSC_SUCCESS);
6055: }
6057: /*
6058: Merge two sets of sorted nonzeros and return a CSR for the merged (sequential) matrix
6060: Input Parameters:
6062: j1,rowBegin1,rowEnd1,jmap1: describe the first set of nonzeros (Set1)
6063: j2,rowBegin2,rowEnd2,jmap2: describe the second set of nonzeros (Set2)
6065: mat: both sets' nonzeros are on m rows, where m is the number of local rows of the matrix mat
6067: For Set1, j1[] contains column indices of the nonzeros.
6068: 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
6069: respectively (note rowEnd1[k] is not necessarily equal to rwoBegin1[k+1]). Indices in this range of j1[] are sorted,
6070: but might have repeats. jmap1[t+1] - jmap1[t] is the number of repeats for the t-th unique nonzero in Set1.
6072: Similar for Set2.
6074: This routine merges the two sets of nonzeros row by row and removes repeats.
6076: Output Parameters: (memory is allocated by the caller)
6078: i[],j[]: the CSR of the merged matrix, which has m rows.
6079: imap1[]: the k-th unique nonzero in Set1 (k=0,1,...) corresponds to imap1[k]-th unique nonzero in the merged matrix.
6080: imap2[]: similar to imap1[], but for Set2.
6081: Note we order nonzeros row-by-row and from left to right.
6082: */
6083: 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[])
6084: {
6085: PetscInt r, m; /* Row index of mat */
6086: PetscCount t, t1, t2, b1, e1, b2, e2;
6088: PetscFunctionBegin;
6089: PetscCall(MatGetLocalSize(mat, &m, NULL));
6090: t1 = t2 = t = 0; /* Count unique nonzeros of in Set1, Set1 and the merged respectively */
6091: i[0] = 0;
6092: for (r = 0; r < m; r++) { /* Do row by row merging */
6093: b1 = rowBegin1[r];
6094: e1 = rowEnd1[r];
6095: b2 = rowBegin2[r];
6096: e2 = rowEnd2[r];
6097: while (b1 < e1 && b2 < e2) {
6098: if (j1[b1] == j2[b2]) { /* Same column index and hence same nonzero */
6099: j[t] = j1[b1];
6100: imap1[t1] = t;
6101: imap2[t2] = t;
6102: b1 += jmap1[t1 + 1] - jmap1[t1]; /* Jump to next unique local nonzero */
6103: b2 += jmap2[t2 + 1] - jmap2[t2]; /* Jump to next unique remote nonzero */
6104: t1++;
6105: t2++;
6106: t++;
6107: } else if (j1[b1] < j2[b2]) {
6108: j[t] = j1[b1];
6109: imap1[t1] = t;
6110: b1 += jmap1[t1 + 1] - jmap1[t1];
6111: t1++;
6112: t++;
6113: } else {
6114: j[t] = j2[b2];
6115: imap2[t2] = t;
6116: b2 += jmap2[t2 + 1] - jmap2[t2];
6117: t2++;
6118: t++;
6119: }
6120: }
6121: /* Merge the remaining in either j1[] or j2[] */
6122: while (b1 < e1) {
6123: j[t] = j1[b1];
6124: imap1[t1] = t;
6125: b1 += jmap1[t1 + 1] - jmap1[t1];
6126: t1++;
6127: t++;
6128: }
6129: while (b2 < e2) {
6130: j[t] = j2[b2];
6131: imap2[t2] = t;
6132: b2 += jmap2[t2 + 1] - jmap2[t2];
6133: t2++;
6134: t++;
6135: }
6136: PetscCall(PetscIntCast(t, i + r + 1));
6137: }
6138: PetscFunctionReturn(PETSC_SUCCESS);
6139: }
6141: /*
6142: Split nonzeros in a block of local rows into two subsets: those in the diagonal block and those in the off-diagonal block
6144: Input Parameters:
6145: mat: an MPI matrix that provides row and column layout information for splitting. Let's say its number of local rows is m.
6146: 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[]
6147: respectively, along with a permutation array perm[]. Length of the i[],j[],perm[] arrays is n.
6149: i[] is already sorted, but within a row, j[] is not sorted and might have repeats.
6150: i[] might contain negative indices at the beginning, which means the corresponding entries should be ignored in the splitting.
6152: Output Parameters:
6153: j[],perm[]: the routine needs to sort j[] within each row along with perm[].
6154: rowBegin[],rowMid[],rowEnd[]: of length m, and the memory is preallocated and zeroed by the caller.
6155: 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,
6156: and [rowMid[r],rowEnd[r]) point to begin/end entries of row r of the off-diagonal block.
6158: Aperm[],Ajmap[],Atot,Annz: Arrays are allocated by this routine.
6159: Atot: number of entries belonging to the diagonal block.
6160: Annz: number of unique nonzeros belonging to the diagonal block.
6161: Aperm[Atot] stores values from perm[] for entries belonging to the diagonal block. Length of Aperm[] is Atot, though it may also count
6162: repeats (i.e., same 'i,j' pair).
6163: Ajmap[Annz+1] stores the number of repeats of each unique entry belonging to the diagonal block. More precisely, Ajmap[t+1] - Ajmap[t]
6164: is the number of repeats for the t-th unique entry in the diagonal block. Ajmap[0] is always 0.
6166: Atot: number of entries belonging to the diagonal block
6167: Annz: number of unique nonzeros belonging to the diagonal block.
6169: Bperm[], Bjmap[], Btot, Bnnz are similar but for the off-diagonal block.
6171: Aperm[],Bperm[],Ajmap[] and Bjmap[] are allocated separately by this routine with PetscMalloc1().
6172: */
6173: 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_)
6174: {
6175: PetscInt cstart, cend, rstart, rend, row, col;
6176: PetscCount Atot = 0, Btot = 0; /* Total number of nonzeros in the diagonal and off-diagonal blocks */
6177: PetscCount Annz = 0, Bnnz = 0; /* Number of unique nonzeros in the diagonal and off-diagonal blocks */
6178: PetscCount k, m, p, q, r, s, mid;
6179: PetscCount *Aperm, *Bperm, *Ajmap, *Bjmap;
6181: PetscFunctionBegin;
6182: PetscCall(PetscLayoutGetRange(mat->rmap, &rstart, &rend));
6183: PetscCall(PetscLayoutGetRange(mat->cmap, &cstart, &cend));
6184: m = rend - rstart;
6186: /* Skip negative rows */
6187: for (k = 0; k < n; k++)
6188: if (i[k] >= 0) break;
6190: /* Process [k,n): sort and partition each local row into diag and offdiag portions,
6191: fill rowBegin[], rowMid[], rowEnd[], and count Atot, Btot, Annz, Bnnz.
6192: */
6193: while (k < n) {
6194: row = i[k];
6195: /* Entries in [k,s) are in one row. Shift diagonal block col indices so that diag is ahead of offdiag after sorting the row */
6196: for (s = k; s < n; s++)
6197: if (i[s] != row) break;
6199: /* Shift diag columns to range of [-PETSC_INT_MAX, -1] */
6200: for (p = k; p < s; p++) {
6201: if (j[p] >= cstart && j[p] < cend) j[p] -= PETSC_INT_MAX;
6202: }
6203: PetscCall(PetscSortIntWithCountArray(s - k, j + k, perm + k));
6204: PetscCall(PetscSortedIntUpperBound(j, k, s, -1, &mid)); /* Separate [k,s) into [k,mid) for diag and [mid,s) for offdiag */
6205: rowBegin[row - rstart] = k;
6206: rowMid[row - rstart] = mid;
6207: rowEnd[row - rstart] = s;
6208: 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);
6210: /* Count nonzeros of this diag/offdiag row, which might have repeats */
6211: Atot += mid - k;
6212: Btot += s - mid;
6214: /* Count unique nonzeros of this diag row */
6215: for (p = k; p < mid;) {
6216: col = j[p];
6217: do {
6218: j[p] += PETSC_INT_MAX; /* Revert the modified diagonal indices */
6219: p++;
6220: } while (p < mid && j[p] == col);
6221: Annz++;
6222: }
6224: /* Count unique nonzeros of this offdiag row */
6225: for (p = mid; p < s;) {
6226: col = j[p];
6227: do {
6228: p++;
6229: } while (p < s && j[p] == col);
6230: Bnnz++;
6231: }
6232: k = s;
6233: }
6235: /* Allocation according to Atot, Btot, Annz, Bnnz */
6236: PetscCall(PetscMalloc1(Atot, &Aperm));
6237: PetscCall(PetscMalloc1(Btot, &Bperm));
6238: PetscCall(PetscMalloc1(Annz + 1, &Ajmap));
6239: PetscCall(PetscMalloc1(Bnnz + 1, &Bjmap));
6241: /* Re-scan indices and copy diag/offdiag permutation indices to Aperm, Bperm and also fill Ajmap and Bjmap */
6242: Ajmap[0] = Bjmap[0] = Atot = Btot = Annz = Bnnz = 0;
6243: for (r = 0; r < m; r++) {
6244: k = rowBegin[r];
6245: mid = rowMid[r];
6246: s = rowEnd[r];
6247: PetscCall(PetscArraycpy(PetscSafePointerPlusOffset(Aperm, Atot), PetscSafePointerPlusOffset(perm, k), mid - k));
6248: PetscCall(PetscArraycpy(PetscSafePointerPlusOffset(Bperm, Btot), PetscSafePointerPlusOffset(perm, mid), s - mid));
6249: Atot += mid - k;
6250: Btot += s - mid;
6252: /* Scan column indices in this row and find out how many repeats each unique nonzero has */
6253: for (p = k; p < mid;) {
6254: col = j[p];
6255: q = p;
6256: do {
6257: p++;
6258: } while (p < mid && j[p] == col);
6259: Ajmap[Annz + 1] = Ajmap[Annz] + (p - q);
6260: Annz++;
6261: }
6263: for (p = mid; p < s;) {
6264: col = j[p];
6265: q = p;
6266: do {
6267: p++;
6268: } while (p < s && j[p] == col);
6269: Bjmap[Bnnz + 1] = Bjmap[Bnnz] + (p - q);
6270: Bnnz++;
6271: }
6272: }
6273: /* Output */
6274: *Aperm_ = Aperm;
6275: *Annz_ = Annz;
6276: *Atot_ = Atot;
6277: *Ajmap_ = Ajmap;
6278: *Bperm_ = Bperm;
6279: *Bnnz_ = Bnnz;
6280: *Btot_ = Btot;
6281: *Bjmap_ = Bjmap;
6282: PetscFunctionReturn(PETSC_SUCCESS);
6283: }
6285: /*
6286: Expand the jmap[] array to make a new one in view of nonzeros in the merged matrix
6288: Input Parameters:
6289: nnz1: number of unique nonzeros in a set that was used to produce imap[], jmap[]
6290: nnz: number of unique nonzeros in the merged matrix
6291: imap[nnz1]: i-th nonzero in the set is the imap[i]-th nonzero in the merged matrix
6292: jmap[nnz1+1]: i-th nonzero in the set has jmap[i+1] - jmap[i] repeats in the set
6294: Output Parameter: (memory is allocated by the caller)
6295: jmap_new[nnz+1]: i-th nonzero in the merged matrix has jmap_new[i+1] - jmap_new[i] repeats in the set
6297: Example:
6298: nnz1 = 4
6299: nnz = 6
6300: imap = [1,3,4,5]
6301: jmap = [0,3,5,6,7]
6302: then,
6303: jmap_new = [0,0,3,3,5,6,7]
6304: */
6305: static PetscErrorCode ExpandJmap_Internal(PetscCount nnz1, PetscCount nnz, const PetscCount imap[], const PetscCount jmap[], PetscCount jmap_new[])
6306: {
6307: PetscCount k, p;
6309: PetscFunctionBegin;
6310: jmap_new[0] = 0;
6311: p = nnz; /* p loops over jmap_new[] backwards */
6312: for (k = nnz1 - 1; k >= 0; k--) { /* k loops over imap[] */
6313: for (; p > imap[k]; p--) jmap_new[p] = jmap[k + 1];
6314: }
6315: for (; p >= 0; p--) jmap_new[p] = jmap[0];
6316: PetscFunctionReturn(PETSC_SUCCESS);
6317: }
6319: static PetscErrorCode MatCOOStructDestroy_MPIAIJ(void **data)
6320: {
6321: MatCOOStruct_MPIAIJ *coo = (MatCOOStruct_MPIAIJ *)*data;
6323: PetscFunctionBegin;
6324: PetscCall(PetscSFDestroy(&coo->sf));
6325: PetscCall(PetscFree(coo->Aperm1));
6326: PetscCall(PetscFree(coo->Bperm1));
6327: PetscCall(PetscFree(coo->Ajmap1));
6328: PetscCall(PetscFree(coo->Bjmap1));
6329: PetscCall(PetscFree(coo->Aimap2));
6330: PetscCall(PetscFree(coo->Bimap2));
6331: PetscCall(PetscFree(coo->Aperm2));
6332: PetscCall(PetscFree(coo->Bperm2));
6333: PetscCall(PetscFree(coo->Ajmap2));
6334: PetscCall(PetscFree(coo->Bjmap2));
6335: PetscCall(PetscFree(coo->Cperm1));
6336: PetscCall(PetscFree2(coo->sendbuf, coo->recvbuf));
6337: PetscCall(PetscFree(coo));
6338: PetscFunctionReturn(PETSC_SUCCESS);
6339: }
6341: PetscErrorCode MatSetPreallocationCOO_MPIAIJ(Mat mat, PetscCount coo_n, PetscInt coo_i[], PetscInt coo_j[])
6342: {
6343: MPI_Comm comm;
6344: PetscMPIInt rank, size;
6345: PetscInt m, n, M, N, rstart, rend, cstart, cend; /* Sizes, indices of row/col, therefore with type PetscInt */
6346: PetscCount k, p, q, rem; /* Loop variables over coo arrays */
6347: Mat_MPIAIJ *mpiaij = (Mat_MPIAIJ *)mat->data;
6348: PetscContainer container;
6349: MatCOOStruct_MPIAIJ *coo;
6351: PetscFunctionBegin;
6352: PetscCall(PetscFree(mpiaij->garray));
6353: PetscCall(VecDestroy(&mpiaij->lvec));
6354: #if defined(PETSC_USE_CTABLE)
6355: PetscCall(PetscHMapIDestroy(&mpiaij->colmap));
6356: #else
6357: PetscCall(PetscFree(mpiaij->colmap));
6358: #endif
6359: PetscCall(VecScatterDestroy(&mpiaij->Mvctx));
6360: mat->assembled = PETSC_FALSE;
6361: mat->was_assembled = PETSC_FALSE;
6363: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
6364: PetscCallMPI(MPI_Comm_size(comm, &size));
6365: PetscCallMPI(MPI_Comm_rank(comm, &rank));
6366: PetscCall(PetscLayoutSetUp(mat->rmap));
6367: PetscCall(PetscLayoutSetUp(mat->cmap));
6368: PetscCall(PetscLayoutGetRange(mat->rmap, &rstart, &rend));
6369: PetscCall(PetscLayoutGetRange(mat->cmap, &cstart, &cend));
6370: PetscCall(MatGetLocalSize(mat, &m, &n));
6371: PetscCall(MatGetSize(mat, &M, &N));
6373: /* Sort (i,j) by row along with a permutation array, so that the to-be-ignored */
6374: /* entries come first, then local rows, then remote rows. */
6375: PetscCount n1 = coo_n, *perm1;
6376: PetscInt *i1 = coo_i, *j1 = coo_j;
6378: PetscCall(PetscMalloc1(n1, &perm1));
6379: for (k = 0; k < n1; k++) perm1[k] = k;
6381: /* Manipulate indices so that entries with negative row or col indices will have smallest
6382: row indices, local entries will have greater but negative row indices, and remote entries
6383: will have positive row indices.
6384: */
6385: for (k = 0; k < n1; k++) {
6386: if (i1[k] < 0 || j1[k] < 0) i1[k] = PETSC_INT_MIN; /* e.g., -2^31, minimal to move them ahead */
6387: 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] */
6388: else {
6389: PetscCheck(!mat->nooffprocentries, PETSC_COMM_SELF, PETSC_ERR_USER_INPUT, "MAT_NO_OFF_PROC_ENTRIES is set but insert to remote rows");
6390: if (mpiaij->donotstash) i1[k] = PETSC_INT_MIN; /* Ignore offproc entries as if they had negative indices */
6391: }
6392: }
6394: /* Sort by row; after that, [0,k) have ignored entries, [k,rem) have local rows and [rem,n1) have remote rows */
6395: PetscCall(PetscSortIntWithIntCountArrayPair(n1, i1, j1, perm1));
6397: /* Advance k to the first entry we need to take care of */
6398: for (k = 0; k < n1; k++)
6399: if (i1[k] > PETSC_INT_MIN) break;
6400: PetscCount i1start = k;
6402: PetscCall(PetscSortedIntUpperBound(i1, k, n1, rend - 1 - PETSC_INT_MAX, &rem)); /* rem is upper bound of the last local row */
6403: for (; k < rem; k++) i1[k] += PETSC_INT_MAX; /* Revert row indices of local rows*/
6405: 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);
6407: /* Send remote rows to their owner */
6408: /* Find which rows should be sent to which remote ranks*/
6409: PetscInt nsend = 0; /* Number of MPI ranks to send data to */
6410: PetscMPIInt *sendto; /* [nsend], storing remote ranks */
6411: PetscInt *nentries; /* [nsend], storing number of entries sent to remote ranks; Assume PetscInt is big enough for this count, and error if not */
6412: const PetscInt *ranges;
6413: PetscInt maxNsend = size >= 128 ? 128 : size; /* Assume max 128 neighbors; realloc when needed */
6415: PetscCall(PetscLayoutGetRanges(mat->rmap, &ranges));
6416: PetscCall(PetscMalloc2(maxNsend, &sendto, maxNsend, &nentries));
6417: for (k = rem; k < n1;) {
6418: PetscMPIInt owner;
6419: PetscInt firstRow, lastRow;
6421: /* Locate a row range */
6422: firstRow = i1[k]; /* first row of this owner */
6423: PetscCall(PetscLayoutFindOwner(mat->rmap, firstRow, &owner));
6424: lastRow = ranges[owner + 1] - 1; /* last row of this owner */
6426: /* Find the first index 'p' in [k,n) with i1[p] belonging to next owner */
6427: PetscCall(PetscSortedIntUpperBound(i1, k, n1, lastRow, &p));
6429: /* All entries in [k,p) belong to this remote owner */
6430: if (nsend >= maxNsend) { /* Double the remote ranks arrays if not long enough */
6431: PetscMPIInt *sendto2;
6432: PetscInt *nentries2;
6433: PetscInt maxNsend2 = (maxNsend <= size / 2) ? maxNsend * 2 : size;
6435: PetscCall(PetscMalloc2(maxNsend2, &sendto2, maxNsend2, &nentries2));
6436: PetscCall(PetscArraycpy(sendto2, sendto, maxNsend));
6437: PetscCall(PetscArraycpy(nentries2, nentries2, maxNsend + 1));
6438: PetscCall(PetscFree2(sendto, nentries2));
6439: sendto = sendto2;
6440: nentries = nentries2;
6441: maxNsend = maxNsend2;
6442: }
6443: sendto[nsend] = owner;
6444: PetscCall(PetscIntCast(p - k, &nentries[nsend]));
6445: nsend++;
6446: k = p;
6447: }
6449: /* Build 1st SF to know offsets on remote to send data */
6450: PetscSF sf1;
6451: PetscInt nroots = 1, nroots2 = 0;
6452: PetscInt nleaves = nsend, nleaves2 = 0;
6453: PetscInt *offsets;
6454: PetscSFNode *iremote;
6456: PetscCall(PetscSFCreate(comm, &sf1));
6457: PetscCall(PetscMalloc1(nsend, &iremote));
6458: PetscCall(PetscMalloc1(nsend, &offsets));
6459: for (k = 0; k < nsend; k++) {
6460: iremote[k].rank = sendto[k];
6461: iremote[k].index = 0;
6462: nleaves2 += nentries[k];
6463: PetscCheck(nleaves2 >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Number of SF leaves is too large for PetscInt");
6464: }
6465: PetscCall(PetscSFSetGraph(sf1, nroots, nleaves, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
6466: PetscCall(PetscSFFetchAndOpWithMemTypeBegin(sf1, MPIU_INT, PETSC_MEMTYPE_HOST, &nroots2 /*rootdata*/, PETSC_MEMTYPE_HOST, nentries /*leafdata*/, PETSC_MEMTYPE_HOST, offsets /*leafupdate*/, MPI_SUM));
6467: PetscCall(PetscSFFetchAndOpEnd(sf1, MPIU_INT, &nroots2, nentries, offsets, MPI_SUM)); /* Would nroots2 overflow, we check offsets[] below */
6468: PetscCall(PetscSFDestroy(&sf1));
6469: PetscAssert(nleaves2 == n1 - rem, PETSC_COMM_SELF, PETSC_ERR_PLIB, "nleaves2 %" PetscInt_FMT " != number of remote entries %" PetscCount_FMT, nleaves2, n1 - rem);
6471: /* Build 2nd SF to send remote COOs to their owner */
6472: PetscSF sf2;
6473: nroots = nroots2;
6474: nleaves = nleaves2;
6475: PetscCall(PetscSFCreate(comm, &sf2));
6476: PetscCall(PetscSFSetFromOptions(sf2));
6477: PetscCall(PetscMalloc1(nleaves, &iremote));
6478: p = 0;
6479: for (k = 0; k < nsend; k++) {
6480: PetscCheck(offsets[k] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Number of SF roots is too large for PetscInt");
6481: for (q = 0; q < nentries[k]; q++, p++) {
6482: iremote[p].rank = sendto[k];
6483: PetscCall(PetscIntCast(offsets[k] + q, &iremote[p].index));
6484: }
6485: }
6486: PetscCall(PetscSFSetGraph(sf2, nroots, nleaves, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
6488: /* Send the remote COOs to their owner */
6489: PetscInt n2 = nroots, *i2, *j2; /* Buffers for received COOs from other ranks, along with a permutation array */
6490: PetscCount *perm2; /* Though PetscInt is enough for remote entries, we use PetscCount here as we want to reuse MatSplitEntries_Internal() */
6491: PetscCall(PetscMalloc3(n2, &i2, n2, &j2, n2, &perm2));
6492: PetscAssert(rem == 0 || i1 != NULL, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot add nonzero offset to null");
6493: PetscAssert(rem == 0 || j1 != NULL, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot add nonzero offset to null");
6494: PetscInt *i1prem = PetscSafePointerPlusOffset(i1, rem);
6495: PetscInt *j1prem = PetscSafePointerPlusOffset(j1, rem);
6496: PetscCall(PetscSFReduceWithMemTypeBegin(sf2, MPIU_INT, PETSC_MEMTYPE_HOST, i1prem, PETSC_MEMTYPE_HOST, i2, MPI_REPLACE));
6497: PetscCall(PetscSFReduceEnd(sf2, MPIU_INT, i1prem, i2, MPI_REPLACE));
6498: PetscCall(PetscSFReduceWithMemTypeBegin(sf2, MPIU_INT, PETSC_MEMTYPE_HOST, j1prem, PETSC_MEMTYPE_HOST, j2, MPI_REPLACE));
6499: PetscCall(PetscSFReduceEnd(sf2, MPIU_INT, j1prem, j2, MPI_REPLACE));
6501: PetscCall(PetscFree(offsets));
6502: PetscCall(PetscFree2(sendto, nentries));
6504: /* Sort received COOs by row along with the permutation array */
6505: for (k = 0; k < n2; k++) perm2[k] = k;
6506: PetscCall(PetscSortIntWithIntCountArrayPair(n2, i2, j2, perm2));
6508: /* sf2 only sends contiguous leafdata to contiguous rootdata. We record the permutation which will be used to fill leafdata */
6509: PetscCount *Cperm1;
6510: PetscAssert(rem == 0 || perm1 != NULL, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot add nonzero offset to null");
6511: PetscCount *perm1prem = PetscSafePointerPlusOffset(perm1, rem);
6512: PetscCall(PetscMalloc1(nleaves, &Cperm1));
6513: PetscCall(PetscArraycpy(Cperm1, perm1prem, nleaves));
6515: /* Support for HYPRE matrices, kind of a hack.
6516: Swap min column with diagonal so that diagonal values will go first */
6517: PetscBool hypre;
6518: PetscCall(PetscStrcmp("_internal_COO_mat_for_hypre", ((PetscObject)mat)->name, &hypre));
6519: if (hypre) {
6520: PetscInt *minj;
6521: PetscBT hasdiag;
6523: PetscCall(PetscBTCreate(m, &hasdiag));
6524: PetscCall(PetscMalloc1(m, &minj));
6525: for (k = 0; k < m; k++) minj[k] = PETSC_INT_MAX;
6526: for (k = i1start; k < rem; k++) {
6527: if (j1[k] < cstart || j1[k] >= cend) continue;
6528: const PetscInt rindex = i1[k] - rstart;
6529: if ((j1[k] - cstart) == rindex) PetscCall(PetscBTSet(hasdiag, rindex));
6530: minj[rindex] = PetscMin(minj[rindex], j1[k]);
6531: }
6532: for (k = 0; k < n2; k++) {
6533: if (j2[k] < cstart || j2[k] >= cend) continue;
6534: const PetscInt rindex = i2[k] - rstart;
6535: if ((j2[k] - cstart) == rindex) PetscCall(PetscBTSet(hasdiag, rindex));
6536: minj[rindex] = PetscMin(minj[rindex], j2[k]);
6537: }
6538: for (k = i1start; k < rem; k++) {
6539: const PetscInt rindex = i1[k] - rstart;
6540: if (j1[k] < cstart || j1[k] >= cend || !PetscBTLookup(hasdiag, rindex)) continue;
6541: if (j1[k] == minj[rindex]) j1[k] = i1[k] + (cstart - rstart);
6542: else if ((j1[k] - cstart) == rindex) j1[k] = minj[rindex];
6543: }
6544: for (k = 0; k < n2; k++) {
6545: const PetscInt rindex = i2[k] - rstart;
6546: if (j2[k] < cstart || j2[k] >= cend || !PetscBTLookup(hasdiag, rindex)) continue;
6547: if (j2[k] == minj[rindex]) j2[k] = i2[k] + (cstart - rstart);
6548: else if ((j2[k] - cstart) == rindex) j2[k] = minj[rindex];
6549: }
6550: PetscCall(PetscBTDestroy(&hasdiag));
6551: PetscCall(PetscFree(minj));
6552: }
6554: /* Split local COOs and received COOs into diag/offdiag portions */
6555: PetscCount *rowBegin1, *rowMid1, *rowEnd1;
6556: PetscCount *Ajmap1, *Aperm1, *Bjmap1, *Bperm1;
6557: PetscCount Annz1, Bnnz1, Atot1, Btot1;
6558: PetscCount *rowBegin2, *rowMid2, *rowEnd2;
6559: PetscCount *Ajmap2, *Aperm2, *Bjmap2, *Bperm2;
6560: PetscCount Annz2, Bnnz2, Atot2, Btot2;
6562: PetscCall(PetscCalloc3(m, &rowBegin1, m, &rowMid1, m, &rowEnd1));
6563: PetscCall(PetscCalloc3(m, &rowBegin2, m, &rowMid2, m, &rowEnd2));
6564: PetscCall(MatSplitEntries_Internal(mat, rem, i1, j1, perm1, rowBegin1, rowMid1, rowEnd1, &Atot1, &Aperm1, &Annz1, &Ajmap1, &Btot1, &Bperm1, &Bnnz1, &Bjmap1));
6565: PetscCall(MatSplitEntries_Internal(mat, n2, i2, j2, perm2, rowBegin2, rowMid2, rowEnd2, &Atot2, &Aperm2, &Annz2, &Ajmap2, &Btot2, &Bperm2, &Bnnz2, &Bjmap2));
6567: /* Merge local COOs with received COOs: diag with diag, offdiag with offdiag */
6568: PetscInt *Ai, *Bi;
6569: PetscInt *Aj, *Bj;
6571: PetscCall(PetscMalloc1(m + 1, &Ai));
6572: PetscCall(PetscMalloc1(m + 1, &Bi));
6573: PetscCall(PetscMalloc1(Annz1 + Annz2, &Aj)); /* Since local and remote entries might have dups, we might allocate excess memory */
6574: PetscCall(PetscMalloc1(Bnnz1 + Bnnz2, &Bj));
6576: PetscCount *Aimap1, *Bimap1, *Aimap2, *Bimap2;
6577: PetscCall(PetscMalloc1(Annz1, &Aimap1));
6578: PetscCall(PetscMalloc1(Bnnz1, &Bimap1));
6579: PetscCall(PetscMalloc1(Annz2, &Aimap2));
6580: PetscCall(PetscMalloc1(Bnnz2, &Bimap2));
6582: PetscCall(MatMergeEntries_Internal(mat, j1, j2, rowBegin1, rowMid1, rowBegin2, rowMid2, Ajmap1, Ajmap2, Aimap1, Aimap2, Ai, Aj));
6583: PetscCall(MatMergeEntries_Internal(mat, j1, j2, rowMid1, rowEnd1, rowMid2, rowEnd2, Bjmap1, Bjmap2, Bimap1, Bimap2, Bi, Bj));
6585: /* Expand Ajmap1/Bjmap1 to make them based off nonzeros in A/B, since we */
6586: /* expect nonzeros in A/B most likely have local contributing entries */
6587: PetscInt Annz = Ai[m];
6588: PetscInt Bnnz = Bi[m];
6589: PetscCount *Ajmap1_new, *Bjmap1_new;
6591: PetscCall(PetscMalloc1(Annz + 1, &Ajmap1_new));
6592: PetscCall(PetscMalloc1(Bnnz + 1, &Bjmap1_new));
6594: PetscCall(ExpandJmap_Internal(Annz1, Annz, Aimap1, Ajmap1, Ajmap1_new));
6595: PetscCall(ExpandJmap_Internal(Bnnz1, Bnnz, Bimap1, Bjmap1, Bjmap1_new));
6597: PetscCall(PetscFree(Aimap1));
6598: PetscCall(PetscFree(Ajmap1));
6599: PetscCall(PetscFree(Bimap1));
6600: PetscCall(PetscFree(Bjmap1));
6601: PetscCall(PetscFree3(rowBegin1, rowMid1, rowEnd1));
6602: PetscCall(PetscFree3(rowBegin2, rowMid2, rowEnd2));
6603: PetscCall(PetscFree(perm1));
6604: PetscCall(PetscFree3(i2, j2, perm2));
6606: Ajmap1 = Ajmap1_new;
6607: Bjmap1 = Bjmap1_new;
6609: /* Reallocate Aj, Bj once we know actual numbers of unique nonzeros in A and B */
6610: if (Annz < Annz1 + Annz2) {
6611: PetscInt *Aj_new;
6612: PetscCall(PetscMalloc1(Annz, &Aj_new));
6613: PetscCall(PetscArraycpy(Aj_new, Aj, Annz));
6614: PetscCall(PetscFree(Aj));
6615: Aj = Aj_new;
6616: }
6618: if (Bnnz < Bnnz1 + Bnnz2) {
6619: PetscInt *Bj_new;
6620: PetscCall(PetscMalloc1(Bnnz, &Bj_new));
6621: PetscCall(PetscArraycpy(Bj_new, Bj, Bnnz));
6622: PetscCall(PetscFree(Bj));
6623: Bj = Bj_new;
6624: }
6626: /* Create new submatrices for on-process and off-process coupling */
6627: PetscScalar *Aa, *Ba;
6628: MatType rtype;
6629: Mat_SeqAIJ *a, *b;
6630: PetscObjectState state;
6631: PetscCall(PetscCalloc1(Annz, &Aa)); /* Zero matrix on device */
6632: PetscCall(PetscCalloc1(Bnnz, &Ba));
6633: /* make Aj[] local, i.e, based off the start column of the diagonal portion */
6634: if (cstart) {
6635: for (k = 0; k < Annz; k++) Aj[k] -= cstart;
6636: }
6638: PetscCall(MatGetRootType_Private(mat, &rtype));
6640: MatSeqXAIJGetOptions_Private(mpiaij->A);
6641: PetscCall(MatDestroy(&mpiaij->A));
6642: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, n, Ai, Aj, Aa, &mpiaij->A));
6643: PetscCall(MatSetBlockSizesFromMats(mpiaij->A, mat, mat));
6644: MatSeqXAIJRestoreOptions_Private(mpiaij->A);
6646: MatSeqXAIJGetOptions_Private(mpiaij->B);
6647: PetscCall(MatDestroy(&mpiaij->B));
6648: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, mat->cmap->N, Bi, Bj, Ba, &mpiaij->B));
6649: PetscCall(MatSetBlockSizesFromMats(mpiaij->B, mat, mat));
6650: MatSeqXAIJRestoreOptions_Private(mpiaij->B);
6652: PetscCall(MatSetUpMultiply_MPIAIJ(mat));
6653: mat->was_assembled = PETSC_TRUE; // was_assembled in effect means the Mvctx is built; doing so avoids redundant MatSetUpMultiply_MPIAIJ
6654: state = mpiaij->A->nonzerostate + mpiaij->B->nonzerostate;
6655: PetscCallMPI(MPIU_Allreduce(&state, &mat->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)mat)));
6657: a = (Mat_SeqAIJ *)mpiaij->A->data;
6658: b = (Mat_SeqAIJ *)mpiaij->B->data;
6659: a->free_a = PETSC_TRUE;
6660: a->free_ij = PETSC_TRUE;
6661: b->free_a = PETSC_TRUE;
6662: b->free_ij = PETSC_TRUE;
6663: a->maxnz = a->nz;
6664: b->maxnz = b->nz;
6666: /* conversion must happen AFTER multiply setup */
6667: PetscCall(MatConvert(mpiaij->A, rtype, MAT_INPLACE_MATRIX, &mpiaij->A));
6668: PetscCall(MatConvert(mpiaij->B, rtype, MAT_INPLACE_MATRIX, &mpiaij->B));
6669: PetscCall(VecDestroy(&mpiaij->lvec));
6670: PetscCall(MatCreateVecs(mpiaij->B, &mpiaij->lvec, NULL));
6672: // Put the COO struct in a container and then attach that to the matrix
6673: PetscCall(PetscMalloc1(1, &coo));
6674: coo->n = coo_n;
6675: coo->sf = sf2;
6676: coo->sendlen = nleaves;
6677: coo->recvlen = nroots;
6678: coo->Annz = Annz;
6679: coo->Bnnz = Bnnz;
6680: coo->Annz2 = Annz2;
6681: coo->Bnnz2 = Bnnz2;
6682: coo->Atot1 = Atot1;
6683: coo->Atot2 = Atot2;
6684: coo->Btot1 = Btot1;
6685: coo->Btot2 = Btot2;
6686: coo->Ajmap1 = Ajmap1;
6687: coo->Aperm1 = Aperm1;
6688: coo->Bjmap1 = Bjmap1;
6689: coo->Bperm1 = Bperm1;
6690: coo->Aimap2 = Aimap2;
6691: coo->Ajmap2 = Ajmap2;
6692: coo->Aperm2 = Aperm2;
6693: coo->Bimap2 = Bimap2;
6694: coo->Bjmap2 = Bjmap2;
6695: coo->Bperm2 = Bperm2;
6696: coo->Cperm1 = Cperm1;
6697: // Allocate in preallocation. If not used, it has zero cost on host
6698: PetscCall(PetscMalloc2(coo->sendlen, &coo->sendbuf, coo->recvlen, &coo->recvbuf));
6699: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
6700: PetscCall(PetscContainerSetPointer(container, coo));
6701: PetscCall(PetscContainerSetCtxDestroy(container, MatCOOStructDestroy_MPIAIJ));
6702: PetscCall(PetscObjectCompose((PetscObject)mat, "__PETSc_MatCOOStruct_Host", (PetscObject)container));
6703: PetscCall(PetscContainerDestroy(&container));
6704: PetscFunctionReturn(PETSC_SUCCESS);
6705: }
6707: static PetscErrorCode MatSetValuesCOO_MPIAIJ(Mat mat, const PetscScalar v[], InsertMode imode)
6708: {
6709: Mat_MPIAIJ *mpiaij = (Mat_MPIAIJ *)mat->data;
6710: Mat A = mpiaij->A, B = mpiaij->B;
6711: PetscScalar *Aa, *Ba;
6712: PetscScalar *sendbuf, *recvbuf;
6713: const PetscCount *Ajmap1, *Ajmap2, *Aimap2;
6714: const PetscCount *Bjmap1, *Bjmap2, *Bimap2;
6715: const PetscCount *Aperm1, *Aperm2, *Bperm1, *Bperm2;
6716: const PetscCount *Cperm1;
6717: PetscContainer container;
6718: MatCOOStruct_MPIAIJ *coo;
6720: PetscFunctionBegin;
6721: PetscCall(PetscObjectQuery((PetscObject)mat, "__PETSc_MatCOOStruct_Host", (PetscObject *)&container));
6722: PetscCheck(container, PetscObjectComm((PetscObject)mat), PETSC_ERR_PLIB, "Not found MatCOOStruct on this matrix");
6723: PetscCall(PetscContainerGetPointer(container, (void **)&coo));
6724: sendbuf = coo->sendbuf;
6725: recvbuf = coo->recvbuf;
6726: Ajmap1 = coo->Ajmap1;
6727: Ajmap2 = coo->Ajmap2;
6728: Aimap2 = coo->Aimap2;
6729: Bjmap1 = coo->Bjmap1;
6730: Bjmap2 = coo->Bjmap2;
6731: Bimap2 = coo->Bimap2;
6732: Aperm1 = coo->Aperm1;
6733: Aperm2 = coo->Aperm2;
6734: Bperm1 = coo->Bperm1;
6735: Bperm2 = coo->Bperm2;
6736: Cperm1 = coo->Cperm1;
6738: PetscCall(MatSeqAIJGetArray(A, &Aa)); /* Might read and write matrix values */
6739: PetscCall(MatSeqAIJGetArray(B, &Ba));
6741: /* Pack entries to be sent to remote */
6742: for (PetscCount i = 0; i < coo->sendlen; i++) sendbuf[i] = v[Cperm1[i]];
6744: /* Send remote entries to their owner and overlap the communication with local computation */
6745: PetscCall(PetscSFReduceWithMemTypeBegin(coo->sf, MPIU_SCALAR, PETSC_MEMTYPE_HOST, sendbuf, PETSC_MEMTYPE_HOST, recvbuf, MPI_REPLACE));
6746: /* Add local entries to A and B */
6747: for (PetscCount i = 0; i < coo->Annz; i++) { /* All nonzeros in A are either zero'ed or added with a value (i.e., initialized) */
6748: PetscScalar sum = 0.0; /* Do partial summation first to improve numerical stability */
6749: for (PetscCount k = Ajmap1[i]; k < Ajmap1[i + 1]; k++) sum += v[Aperm1[k]];
6750: Aa[i] = (imode == INSERT_VALUES ? 0.0 : Aa[i]) + sum;
6751: }
6752: for (PetscCount i = 0; i < coo->Bnnz; i++) {
6753: PetscScalar sum = 0.0;
6754: for (PetscCount k = Bjmap1[i]; k < Bjmap1[i + 1]; k++) sum += v[Bperm1[k]];
6755: Ba[i] = (imode == INSERT_VALUES ? 0.0 : Ba[i]) + sum;
6756: }
6757: PetscCall(PetscSFReduceEnd(coo->sf, MPIU_SCALAR, sendbuf, recvbuf, MPI_REPLACE));
6759: /* Add received remote entries to A and B */
6760: for (PetscCount i = 0; i < coo->Annz2; i++) {
6761: for (PetscCount k = Ajmap2[i]; k < Ajmap2[i + 1]; k++) Aa[Aimap2[i]] += recvbuf[Aperm2[k]];
6762: }
6763: for (PetscCount i = 0; i < coo->Bnnz2; i++) {
6764: for (PetscCount k = Bjmap2[i]; k < Bjmap2[i + 1]; k++) Ba[Bimap2[i]] += recvbuf[Bperm2[k]];
6765: }
6766: PetscCall(MatSeqAIJRestoreArray(A, &Aa));
6767: PetscCall(MatSeqAIJRestoreArray(B, &Ba));
6768: PetscFunctionReturn(PETSC_SUCCESS);
6769: }
6771: /*MC
6772: MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices.
6774: Options Database Keys:
6775: . -mat_type mpiaij - sets the matrix type to `MATMPIAIJ` during a call to `MatSetFromOptions()`
6777: Level: beginner
6779: Notes:
6780: `MatSetValues()` may be called for this matrix type with a `NULL` argument for the numerical values,
6781: in this case the values associated with the rows and columns one passes in are set to zero
6782: in the matrix
6784: `MatSetOptions`(,`MAT_STRUCTURE_ONLY`,`PETSC_TRUE`) may be called for this matrix type. In this no
6785: space is allocated for the nonzero entries and any entries passed with `MatSetValues()` are ignored
6787: .seealso: [](ch_matrices), `Mat`, `MATSEQAIJ`, `MATAIJ`, `MatCreateAIJ()`
6788: M*/
6789: PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat B)
6790: {
6791: Mat_MPIAIJ *b;
6792: PetscMPIInt size;
6794: PetscFunctionBegin;
6795: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
6797: PetscCall(PetscNew(&b));
6798: B->data = (void *)b;
6799: B->ops[0] = MatOps_Values;
6800: B->assembled = PETSC_FALSE;
6801: B->insertmode = NOT_SET_VALUES;
6802: b->size = size;
6804: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)B), &b->rank));
6806: /* build cache for off array entries formed */
6807: PetscCall(MatStashCreate_Private(PetscObjectComm((PetscObject)B), 1, &B->stash));
6809: b->donotstash = PETSC_FALSE;
6810: b->colmap = NULL;
6811: b->garray = NULL;
6812: b->roworiented = PETSC_TRUE;
6814: /* stuff used for matrix vector multiply */
6815: b->lvec = NULL;
6816: b->Mvctx = NULL;
6818: /* stuff for MatGetRow() */
6819: b->rowindices = NULL;
6820: b->rowvalues = NULL;
6821: b->getrowactive = PETSC_FALSE;
6823: /* flexible pointer used in CUSPARSE classes */
6824: b->spptr = NULL;
6826: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetUseScalableIncreaseOverlap_C", MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ));
6827: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatStoreValues_C", MatStoreValues_MPIAIJ));
6828: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatRetrieveValues_C", MatRetrieveValues_MPIAIJ));
6829: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatIsTranspose_C", MatIsTranspose_MPIAIJ));
6830: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetPreallocation_C", MatMPIAIJSetPreallocation_MPIAIJ));
6831: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatResetPreallocation_C", MatResetPreallocation_MPIAIJ));
6832: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatResetHash_C", MatResetHash_MPIAIJ));
6833: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetPreallocationCSR_C", MatMPIAIJSetPreallocationCSR_MPIAIJ));
6834: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDiagonalScaleLocal_C", MatDiagonalScaleLocal_MPIAIJ));
6835: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijperm_C", MatConvert_MPIAIJ_MPIAIJPERM));
6836: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijsell_C", MatConvert_MPIAIJ_MPIAIJSELL));
6837: #if defined(PETSC_HAVE_CUDA)
6838: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijcusparse_C", MatConvert_MPIAIJ_MPIAIJCUSPARSE));
6839: #endif
6840: #if defined(PETSC_HAVE_HIP)
6841: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijhipsparse_C", MatConvert_MPIAIJ_MPIAIJHIPSPARSE));
6842: #endif
6843: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
6844: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijkokkos_C", MatConvert_MPIAIJ_MPIAIJKokkos));
6845: #endif
6846: #if defined(PETSC_HAVE_MKL_SPARSE)
6847: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijmkl_C", MatConvert_MPIAIJ_MPIAIJMKL));
6848: #endif
6849: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijcrl_C", MatConvert_MPIAIJ_MPIAIJCRL));
6850: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpibaij_C", MatConvert_MPIAIJ_MPIBAIJ));
6851: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpisbaij_C", MatConvert_MPIAIJ_MPISBAIJ));
6852: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpidense_C", MatConvert_MPIAIJ_MPIDense));
6853: #if defined(PETSC_HAVE_ELEMENTAL)
6854: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_elemental_C", MatConvert_MPIAIJ_Elemental));
6855: #endif
6856: #if defined(PETSC_HAVE_SCALAPACK) && (defined(PETSC_USE_REAL_SINGLE) || defined(PETSC_USE_REAL_DOUBLE))
6857: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_scalapack_C", MatConvert_AIJ_ScaLAPACK));
6858: #endif
6859: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_is_C", MatConvert_XAIJ_IS));
6860: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpisell_C", MatConvert_MPIAIJ_MPISELL));
6861: #if defined(PETSC_HAVE_HYPRE)
6862: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_hypre_C", MatConvert_AIJ_HYPRE));
6863: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_transpose_mpiaij_mpiaij_C", MatProductSetFromOptions_Transpose_AIJ_AIJ));
6864: #endif
6865: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_is_mpiaij_C", MatProductSetFromOptions_IS_XAIJ));
6866: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_mpiaij_mpiaij_C", MatProductSetFromOptions_MPIAIJ));
6867: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetPreallocationCOO_C", MatSetPreallocationCOO_MPIAIJ));
6868: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetValuesCOO_C", MatSetValuesCOO_MPIAIJ));
6869: PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATMPIAIJ));
6870: PetscFunctionReturn(PETSC_SUCCESS);
6871: }
6873: /*@
6874: MatCreateMPIAIJWithSplitArrays - creates a `MATMPIAIJ` matrix using arrays that contain the "diagonal"
6875: and "off-diagonal" part of the matrix in CSR format.
6877: Collective
6879: Input Parameters:
6880: + comm - MPI communicator
6881: . m - number of local rows (Cannot be `PETSC_DECIDE`)
6882: . n - This value should be the same as the local size used in creating the
6883: x vector for the matrix-vector product $y = Ax$. (or `PETSC_DECIDE` to have
6884: calculated if `N` is given) For square matrices `n` is almost always `m`.
6885: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if `m` is given)
6886: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if `n` is given)
6887: . 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
6888: . j - column indices, which must be local, i.e., based off the start column of the diagonal portion
6889: . a - matrix values
6890: . 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
6891: . oj - column indices, which must be global, representing global columns in the `MATMPIAIJ` matrix
6892: - oa - matrix values
6894: Output Parameter:
6895: . mat - the matrix
6897: Level: advanced
6899: Notes:
6900: The `i`, `j`, and `a` arrays ARE NOT copied by this routine into the internal format used by PETSc (even in Fortran). The user
6901: must free the arrays once the matrix has been destroyed and not before.
6903: The `i` and `j` indices are 0 based
6905: See `MatCreateAIJ()` for the definition of "diagonal" and "off-diagonal" portion of the matrix
6907: This sets local rows and cannot be used to set off-processor values.
6909: Use of this routine is discouraged because it is inflexible and cumbersome to use. It is extremely rare that a
6910: legacy application natively assembles into exactly this split format. The code to do so is nontrivial and does
6911: not easily support in-place reassembly. It is recommended to use MatSetValues() (or a variant thereof) because
6912: the resulting assembly is easier to implement, will work with any matrix format, and the user does not have to
6913: keep track of the underlying array. Use `MatSetOption`(A,`MAT_NO_OFF_PROC_ENTRIES`,`PETSC_TRUE`) to disable all
6914: communication if it is known that only local entries will be set.
6916: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
6917: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithArrays()`
6918: @*/
6919: 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)
6920: {
6921: Mat_MPIAIJ *maij;
6923: PetscFunctionBegin;
6924: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
6925: PetscCheck(i[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
6926: PetscCheck(oi[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "oi (row indices) must start with 0");
6927: PetscCall(MatCreate(comm, mat));
6928: PetscCall(MatSetSizes(*mat, m, n, M, N));
6929: PetscCall(MatSetType(*mat, MATMPIAIJ));
6930: maij = (Mat_MPIAIJ *)(*mat)->data;
6932: (*mat)->preallocated = PETSC_TRUE;
6934: PetscCall(PetscLayoutSetUp((*mat)->rmap));
6935: PetscCall(PetscLayoutSetUp((*mat)->cmap));
6937: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, n, i, j, a, &maij->A));
6938: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, (*mat)->cmap->N, oi, oj, oa, &maij->B));
6940: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
6941: PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY));
6942: PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY));
6943: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_FALSE));
6944: PetscCall(MatSetOption(*mat, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
6945: PetscFunctionReturn(PETSC_SUCCESS);
6946: }
6948: typedef struct {
6949: Mat *mp; /* intermediate products */
6950: PetscBool *mptmp; /* is the intermediate product temporary ? */
6951: PetscInt cp; /* number of intermediate products */
6953: /* support for MatGetBrowsOfAoCols_MPIAIJ for P_oth */
6954: PetscInt *startsj_s, *startsj_r;
6955: PetscScalar *bufa;
6956: Mat P_oth;
6958: /* may take advantage of merging product->B */
6959: Mat Bloc; /* B-local by merging diag and off-diag */
6961: /* cusparse does not have support to split between symbolic and numeric phases.
6962: When api_user is true, we don't need to update the numerical values
6963: of the temporary storage */
6964: PetscBool reusesym;
6966: /* support for COO values insertion */
6967: PetscScalar *coo_v, *coo_w; /* store on-process and off-process COO scalars, and used as MPI recv/send buffers respectively */
6968: PetscInt **own; /* own[i] points to address of on-process COO indices for Mat mp[i] */
6969: PetscInt **off; /* off[i] points to address of off-process COO indices for Mat mp[i] */
6970: PetscBool hasoffproc; /* if true, have off-process values insertion (i.e. AtB or PtAP) */
6971: PetscSF sf; /* used for non-local values insertion and memory malloc */
6972: PetscMemType mtype;
6974: /* customization */
6975: PetscBool abmerge;
6976: PetscBool P_oth_bind;
6977: } MatMatMPIAIJBACKEND;
6979: static PetscErrorCode MatProductCtxDestroy_MatMatMPIAIJBACKEND(void **data)
6980: {
6981: MatMatMPIAIJBACKEND *mmdata = *(MatMatMPIAIJBACKEND **)data;
6982: PetscInt i;
6984: PetscFunctionBegin;
6985: PetscCall(PetscFree2(mmdata->startsj_s, mmdata->startsj_r));
6986: PetscCall(PetscFree(mmdata->bufa));
6987: PetscCall(PetscSFFree(mmdata->sf, mmdata->mtype, mmdata->coo_v));
6988: PetscCall(PetscSFFree(mmdata->sf, mmdata->mtype, mmdata->coo_w));
6989: PetscCall(MatDestroy(&mmdata->P_oth));
6990: PetscCall(MatDestroy(&mmdata->Bloc));
6991: PetscCall(PetscSFDestroy(&mmdata->sf));
6992: for (i = 0; i < mmdata->cp; i++) PetscCall(MatDestroy(&mmdata->mp[i]));
6993: PetscCall(PetscFree2(mmdata->mp, mmdata->mptmp));
6994: PetscCall(PetscFree(mmdata->own[0]));
6995: PetscCall(PetscFree(mmdata->own));
6996: PetscCall(PetscFree(mmdata->off[0]));
6997: PetscCall(PetscFree(mmdata->off));
6998: PetscCall(PetscFree(mmdata));
6999: PetscFunctionReturn(PETSC_SUCCESS);
7000: }
7002: /* Copy selected n entries with indices in idx[] of A to v[].
7003: If idx is NULL, copy the whole data array of A to v[]
7004: */
7005: static PetscErrorCode MatSeqAIJCopySubArray(Mat A, PetscInt n, const PetscInt idx[], PetscScalar v[])
7006: {
7007: PetscErrorCode (*f)(Mat, PetscInt, const PetscInt[], PetscScalar[]);
7009: PetscFunctionBegin;
7010: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatSeqAIJCopySubArray_C", &f));
7011: if (f) {
7012: PetscCall((*f)(A, n, idx, v));
7013: } else {
7014: const PetscScalar *vv;
7016: PetscCall(MatSeqAIJGetArrayRead(A, &vv));
7017: if (n && idx) {
7018: PetscScalar *w = v;
7019: const PetscInt *oi = idx;
7020: PetscInt j;
7022: for (j = 0; j < n; j++) *w++ = vv[*oi++];
7023: } else {
7024: PetscCall(PetscArraycpy(v, vv, n));
7025: }
7026: PetscCall(MatSeqAIJRestoreArrayRead(A, &vv));
7027: }
7028: PetscFunctionReturn(PETSC_SUCCESS);
7029: }
7031: static PetscErrorCode MatProductNumeric_MPIAIJBACKEND(Mat C)
7032: {
7033: MatMatMPIAIJBACKEND *mmdata;
7034: PetscInt i, n_d, n_o;
7036: PetscFunctionBegin;
7037: MatCheckProduct(C, 1);
7038: PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data empty");
7039: mmdata = (MatMatMPIAIJBACKEND *)C->product->data;
7040: if (!mmdata->reusesym) { /* update temporary matrices */
7041: 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));
7042: if (mmdata->Bloc) PetscCall(MatMPIAIJGetLocalMatMerge(C->product->B, MAT_REUSE_MATRIX, NULL, &mmdata->Bloc));
7043: }
7044: mmdata->reusesym = PETSC_FALSE;
7046: for (i = 0; i < mmdata->cp; i++) {
7047: 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]);
7048: PetscCall((*mmdata->mp[i]->ops->productnumeric)(mmdata->mp[i]));
7049: }
7050: for (i = 0, n_d = 0, n_o = 0; i < mmdata->cp; i++) {
7051: PetscInt noff;
7053: PetscCall(PetscIntCast(mmdata->off[i + 1] - mmdata->off[i], &noff));
7054: if (mmdata->mptmp[i]) continue;
7055: if (noff) {
7056: PetscInt nown;
7058: PetscCall(PetscIntCast(mmdata->own[i + 1] - mmdata->own[i], &nown));
7059: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], noff, mmdata->off[i], mmdata->coo_w + n_o));
7060: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], nown, mmdata->own[i], mmdata->coo_v + n_d));
7061: n_o += noff;
7062: n_d += nown;
7063: } else {
7064: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mmdata->mp[i]->data;
7066: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], mm->nz, NULL, mmdata->coo_v + n_d));
7067: n_d += mm->nz;
7068: }
7069: }
7070: if (mmdata->hasoffproc) { /* offprocess insertion */
7071: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_SCALAR, mmdata->coo_w, mmdata->coo_v + n_d));
7072: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_SCALAR, mmdata->coo_w, mmdata->coo_v + n_d));
7073: }
7074: PetscCall(MatSetValuesCOO(C, mmdata->coo_v, INSERT_VALUES));
7075: PetscFunctionReturn(PETSC_SUCCESS);
7076: }
7078: /* Support for Pt * A, A * P, or Pt * A * P */
7079: #define MAX_NUMBER_INTERMEDIATE 4
7080: PetscErrorCode MatProductSymbolic_MPIAIJBACKEND(Mat C)
7081: {
7082: Mat_Product *product = C->product;
7083: Mat A, P, mp[MAX_NUMBER_INTERMEDIATE]; /* A, P and a series of intermediate matrices */
7084: Mat_MPIAIJ *a, *p;
7085: MatMatMPIAIJBACKEND *mmdata;
7086: ISLocalToGlobalMapping P_oth_l2g = NULL;
7087: IS glob = NULL;
7088: const char *prefix;
7089: char pprefix[256];
7090: const PetscInt *globidx, *P_oth_idx;
7091: PetscInt i, j, cp, m, n, M, N, *coo_i, *coo_j;
7092: PetscCount ncoo, ncoo_d, ncoo_o, ncoo_oown;
7093: PetscInt cmapt[MAX_NUMBER_INTERMEDIATE], rmapt[MAX_NUMBER_INTERMEDIATE]; /* col/row map type for each Mat in mp[]. */
7094: /* type-0: consecutive, start from 0; type-1: consecutive with */
7095: /* a base offset; type-2: sparse with a local to global map table */
7096: const PetscInt *cmapa[MAX_NUMBER_INTERMEDIATE], *rmapa[MAX_NUMBER_INTERMEDIATE]; /* col/row local to global map array (table) for type-2 map type */
7098: MatProductType ptype;
7099: PetscBool mptmp[MAX_NUMBER_INTERMEDIATE], hasoffproc = PETSC_FALSE, iscuda, iship, iskokk;
7100: PetscMPIInt size;
7102: PetscFunctionBegin;
7103: MatCheckProduct(C, 1);
7104: PetscCheck(!product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data not empty");
7105: ptype = product->type;
7106: if (product->A->symmetric == PETSC_BOOL3_TRUE && ptype == MATPRODUCT_AtB) {
7107: ptype = MATPRODUCT_AB;
7108: product->symbolic_used_the_fact_A_is_symmetric = PETSC_TRUE;
7109: }
7110: switch (ptype) {
7111: case MATPRODUCT_AB:
7112: A = product->A;
7113: P = product->B;
7114: m = A->rmap->n;
7115: n = P->cmap->n;
7116: M = A->rmap->N;
7117: N = P->cmap->N;
7118: hasoffproc = PETSC_FALSE; /* will not scatter mat product values to other processes */
7119: break;
7120: case MATPRODUCT_AtB:
7121: P = product->A;
7122: A = product->B;
7123: m = P->cmap->n;
7124: n = A->cmap->n;
7125: M = P->cmap->N;
7126: N = A->cmap->N;
7127: hasoffproc = PETSC_TRUE;
7128: break;
7129: case MATPRODUCT_PtAP:
7130: A = product->A;
7131: P = product->B;
7132: m = P->cmap->n;
7133: n = P->cmap->n;
7134: M = P->cmap->N;
7135: N = P->cmap->N;
7136: hasoffproc = PETSC_TRUE;
7137: break;
7138: default:
7139: SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Not for product type %s", MatProductTypes[ptype]);
7140: }
7141: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)C), &size));
7142: if (size == 1) hasoffproc = PETSC_FALSE;
7144: /* defaults */
7145: for (i = 0; i < MAX_NUMBER_INTERMEDIATE; i++) {
7146: mp[i] = NULL;
7147: mptmp[i] = PETSC_FALSE;
7148: rmapt[i] = -1;
7149: cmapt[i] = -1;
7150: rmapa[i] = NULL;
7151: cmapa[i] = NULL;
7152: }
7154: /* customization */
7155: PetscCall(PetscNew(&mmdata));
7156: mmdata->reusesym = product->api_user;
7157: if (ptype == MATPRODUCT_AB) {
7158: if (product->api_user) {
7159: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatMatMult", "Mat");
7160: PetscCall(PetscOptionsBool("-matmatmult_backend_mergeB", "Merge product->B local matrices", "MatMatMult", mmdata->abmerge, &mmdata->abmerge, NULL));
7161: PetscCall(PetscOptionsBool("-matmatmult_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7162: PetscOptionsEnd();
7163: } else {
7164: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_AB", "Mat");
7165: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_mergeB", "Merge product->B local matrices", "MatMatMult", mmdata->abmerge, &mmdata->abmerge, NULL));
7166: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7167: PetscOptionsEnd();
7168: }
7169: } else if (ptype == MATPRODUCT_PtAP) {
7170: if (product->api_user) {
7171: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatPtAP", "Mat");
7172: PetscCall(PetscOptionsBool("-matptap_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7173: PetscOptionsEnd();
7174: } else {
7175: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_PtAP", "Mat");
7176: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7177: PetscOptionsEnd();
7178: }
7179: }
7180: a = (Mat_MPIAIJ *)A->data;
7181: p = (Mat_MPIAIJ *)P->data;
7182: PetscCall(MatSetSizes(C, m, n, M, N));
7183: PetscCall(PetscLayoutSetUp(C->rmap));
7184: PetscCall(PetscLayoutSetUp(C->cmap));
7185: PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
7186: PetscCall(MatGetOptionsPrefix(C, &prefix));
7188: cp = 0;
7189: switch (ptype) {
7190: case MATPRODUCT_AB: /* A * P */
7191: PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &mmdata->startsj_s, &mmdata->startsj_r, &mmdata->bufa, &mmdata->P_oth));
7193: /* A_diag * P_local (merged or not) */
7194: if (mmdata->abmerge) { /* P's diagonal and off-diag blocks are merged to one matrix, then multiplied by A_diag */
7195: /* P is product->B */
7196: PetscCall(MatMPIAIJGetLocalMatMerge(P, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7197: PetscCall(MatProductCreate(a->A, mmdata->Bloc, NULL, &mp[cp]));
7198: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7199: PetscCall(MatProductSetFill(mp[cp], product->fill));
7200: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7201: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7202: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7203: mp[cp]->product->api_user = product->api_user;
7204: PetscCall(MatProductSetFromOptions(mp[cp]));
7205: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7206: PetscCall(ISGetIndices(glob, &globidx));
7207: rmapt[cp] = 1;
7208: cmapt[cp] = 2;
7209: cmapa[cp] = globidx;
7210: mptmp[cp] = PETSC_FALSE;
7211: cp++;
7212: } else { /* A_diag * P_diag and A_diag * P_off */
7213: PetscCall(MatProductCreate(a->A, p->A, NULL, &mp[cp]));
7214: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7215: PetscCall(MatProductSetFill(mp[cp], product->fill));
7216: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7217: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7218: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7219: mp[cp]->product->api_user = product->api_user;
7220: PetscCall(MatProductSetFromOptions(mp[cp]));
7221: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7222: rmapt[cp] = 1;
7223: cmapt[cp] = 1;
7224: mptmp[cp] = PETSC_FALSE;
7225: cp++;
7226: PetscCall(MatProductCreate(a->A, p->B, NULL, &mp[cp]));
7227: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7228: PetscCall(MatProductSetFill(mp[cp], product->fill));
7229: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7230: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7231: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7232: mp[cp]->product->api_user = product->api_user;
7233: PetscCall(MatProductSetFromOptions(mp[cp]));
7234: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7235: rmapt[cp] = 1;
7236: cmapt[cp] = 2;
7237: cmapa[cp] = p->garray;
7238: mptmp[cp] = PETSC_FALSE;
7239: cp++;
7240: }
7242: /* A_off * P_other */
7243: if (mmdata->P_oth) {
7244: PetscCall(MatSeqAIJCompactOutExtraColumns_SeqAIJ(mmdata->P_oth, &P_oth_l2g)); /* make P_oth use local col ids */
7245: PetscCall(ISLocalToGlobalMappingGetIndices(P_oth_l2g, &P_oth_idx));
7246: PetscCall(MatSetType(mmdata->P_oth, ((PetscObject)a->B)->type_name));
7247: PetscCall(MatBindToCPU(mmdata->P_oth, mmdata->P_oth_bind));
7248: PetscCall(MatProductCreate(a->B, mmdata->P_oth, NULL, &mp[cp]));
7249: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7250: PetscCall(MatProductSetFill(mp[cp], product->fill));
7251: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7252: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7253: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7254: mp[cp]->product->api_user = product->api_user;
7255: PetscCall(MatProductSetFromOptions(mp[cp]));
7256: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7257: rmapt[cp] = 1;
7258: cmapt[cp] = 2;
7259: cmapa[cp] = P_oth_idx;
7260: mptmp[cp] = PETSC_FALSE;
7261: cp++;
7262: }
7263: break;
7265: case MATPRODUCT_AtB: /* (P^t * A): P_diag * A_loc + P_off * A_loc */
7266: /* A is product->B */
7267: PetscCall(MatMPIAIJGetLocalMatMerge(A, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7268: if (A == P) { /* when A==P, we can take advantage of the already merged mmdata->Bloc */
7269: PetscCall(MatProductCreate(mmdata->Bloc, mmdata->Bloc, NULL, &mp[cp]));
7270: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
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: PetscCall(ISGetIndices(glob, &globidx));
7279: rmapt[cp] = 2;
7280: rmapa[cp] = globidx;
7281: cmapt[cp] = 2;
7282: cmapa[cp] = globidx;
7283: mptmp[cp] = PETSC_FALSE;
7284: cp++;
7285: } else {
7286: PetscCall(MatProductCreate(p->A, mmdata->Bloc, NULL, &mp[cp]));
7287: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7288: PetscCall(MatProductSetFill(mp[cp], product->fill));
7289: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7290: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7291: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7292: mp[cp]->product->api_user = product->api_user;
7293: PetscCall(MatProductSetFromOptions(mp[cp]));
7294: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7295: PetscCall(ISGetIndices(glob, &globidx));
7296: rmapt[cp] = 1;
7297: cmapt[cp] = 2;
7298: cmapa[cp] = globidx;
7299: mptmp[cp] = PETSC_FALSE;
7300: cp++;
7301: PetscCall(MatProductCreate(p->B, mmdata->Bloc, NULL, &mp[cp]));
7302: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7303: PetscCall(MatProductSetFill(mp[cp], product->fill));
7304: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7305: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7306: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7307: mp[cp]->product->api_user = product->api_user;
7308: PetscCall(MatProductSetFromOptions(mp[cp]));
7309: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7310: rmapt[cp] = 2;
7311: rmapa[cp] = p->garray;
7312: cmapt[cp] = 2;
7313: cmapa[cp] = globidx;
7314: mptmp[cp] = PETSC_FALSE;
7315: cp++;
7316: }
7317: break;
7318: case MATPRODUCT_PtAP:
7319: PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &mmdata->startsj_s, &mmdata->startsj_r, &mmdata->bufa, &mmdata->P_oth));
7320: /* P is product->B */
7321: PetscCall(MatMPIAIJGetLocalMatMerge(P, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7322: PetscCall(MatProductCreate(a->A, mmdata->Bloc, NULL, &mp[cp]));
7323: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_PtAP));
7324: PetscCall(MatProductSetFill(mp[cp], product->fill));
7325: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7326: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7327: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7328: mp[cp]->product->api_user = product->api_user;
7329: PetscCall(MatProductSetFromOptions(mp[cp]));
7330: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7331: PetscCall(ISGetIndices(glob, &globidx));
7332: rmapt[cp] = 2;
7333: rmapa[cp] = globidx;
7334: cmapt[cp] = 2;
7335: cmapa[cp] = globidx;
7336: mptmp[cp] = PETSC_FALSE;
7337: cp++;
7338: if (mmdata->P_oth) {
7339: PetscCall(MatSeqAIJCompactOutExtraColumns_SeqAIJ(mmdata->P_oth, &P_oth_l2g));
7340: PetscCall(ISLocalToGlobalMappingGetIndices(P_oth_l2g, &P_oth_idx));
7341: PetscCall(MatSetType(mmdata->P_oth, ((PetscObject)a->B)->type_name));
7342: PetscCall(MatBindToCPU(mmdata->P_oth, mmdata->P_oth_bind));
7343: PetscCall(MatProductCreate(a->B, mmdata->P_oth, NULL, &mp[cp]));
7344: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7345: PetscCall(MatProductSetFill(mp[cp], product->fill));
7346: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7347: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7348: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7349: mp[cp]->product->api_user = product->api_user;
7350: PetscCall(MatProductSetFromOptions(mp[cp]));
7351: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7352: mptmp[cp] = PETSC_TRUE;
7353: cp++;
7354: PetscCall(MatProductCreate(mmdata->Bloc, mp[1], NULL, &mp[cp]));
7355: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7356: PetscCall(MatProductSetFill(mp[cp], product->fill));
7357: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7358: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7359: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7360: mp[cp]->product->api_user = product->api_user;
7361: PetscCall(MatProductSetFromOptions(mp[cp]));
7362: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7363: rmapt[cp] = 2;
7364: rmapa[cp] = globidx;
7365: cmapt[cp] = 2;
7366: cmapa[cp] = P_oth_idx;
7367: mptmp[cp] = PETSC_FALSE;
7368: cp++;
7369: }
7370: break;
7371: default:
7372: SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Not for product type %s", MatProductTypes[ptype]);
7373: }
7374: /* sanity check */
7375: if (size > 1)
7376: 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);
7378: PetscCall(PetscMalloc2(cp, &mmdata->mp, cp, &mmdata->mptmp));
7379: for (i = 0; i < cp; i++) {
7380: mmdata->mp[i] = mp[i];
7381: mmdata->mptmp[i] = mptmp[i];
7382: }
7383: mmdata->cp = cp;
7384: C->product->data = mmdata;
7385: C->product->destroy = MatProductCtxDestroy_MatMatMPIAIJBACKEND;
7386: C->ops->productnumeric = MatProductNumeric_MPIAIJBACKEND;
7388: /* memory type */
7389: mmdata->mtype = PETSC_MEMTYPE_HOST;
7390: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iscuda, MATSEQAIJCUSPARSE, MATMPIAIJCUSPARSE, ""));
7391: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iship, MATSEQAIJHIPSPARSE, MATMPIAIJHIPSPARSE, ""));
7392: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iskokk, MATSEQAIJKOKKOS, MATMPIAIJKOKKOS, ""));
7393: if (iscuda) mmdata->mtype = PETSC_MEMTYPE_CUDA;
7394: else if (iship) mmdata->mtype = PETSC_MEMTYPE_HIP;
7395: else if (iskokk) mmdata->mtype = PETSC_MEMTYPE_KOKKOS;
7397: /* prepare coo coordinates for values insertion */
7399: /* count total nonzeros of those intermediate seqaij Mats
7400: ncoo_d: # of nonzeros of matrices that do not have offproc entries
7401: ncoo_o: # of nonzeros (of matrices that might have offproc entries) that will be inserted to remote procs
7402: ncoo_oown: # of nonzeros (of matrices that might have offproc entries) that will be inserted locally
7403: */
7404: for (cp = 0, ncoo_d = 0, ncoo_o = 0, ncoo_oown = 0; cp < mmdata->cp; cp++) {
7405: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7406: if (mptmp[cp]) continue;
7407: if (rmapt[cp] == 2 && hasoffproc) { /* the rows need to be scatter to all processes (might include self) */
7408: const PetscInt *rmap = rmapa[cp];
7409: const PetscInt mr = mp[cp]->rmap->n;
7410: const PetscInt rs = C->rmap->rstart;
7411: const PetscInt re = C->rmap->rend;
7412: const PetscInt *ii = mm->i;
7413: for (i = 0; i < mr; i++) {
7414: const PetscInt gr = rmap[i];
7415: const PetscInt nz = ii[i + 1] - ii[i];
7416: if (gr < rs || gr >= re) ncoo_o += nz; /* this row is offproc */
7417: else ncoo_oown += nz; /* this row is local */
7418: }
7419: } else ncoo_d += mm->nz;
7420: }
7422: /*
7423: ncoo: total number of nonzeros (including those inserted by remote procs) belonging to this proc
7425: ncoo = ncoo_d + ncoo_oown + ncoo2, which ncoo2 is number of nonzeros inserted to me by other procs.
7427: off[0] points to a big index array, which is shared by off[1,2,...]. Similarly, for own[0].
7429: off[p]: points to the segment for matrix mp[p], storing location of nonzeros that mp[p] will insert to others
7430: own[p]: points to the segment for matrix mp[p], storing location of nonzeros that mp[p] will insert locally
7431: so, off[p+1]-off[p] is the number of nonzeros that mp[p] will send to others.
7433: coo_i/j/v[]: [ncoo] row/col/val of nonzeros belonging to this proc.
7434: 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.
7435: */
7436: PetscCall(PetscCalloc1(mmdata->cp + 1, &mmdata->off)); /* +1 to make a csr-like data structure */
7437: PetscCall(PetscCalloc1(mmdata->cp + 1, &mmdata->own));
7439: /* gather (i,j) of nonzeros inserted by remote procs */
7440: if (hasoffproc) {
7441: PetscSF msf;
7442: PetscInt ncoo2, *coo_i2, *coo_j2;
7444: PetscCall(PetscMalloc1(ncoo_o, &mmdata->off[0]));
7445: PetscCall(PetscMalloc1(ncoo_oown, &mmdata->own[0]));
7446: PetscCall(PetscMalloc2(ncoo_o, &coo_i, ncoo_o, &coo_j)); /* to collect (i,j) of entries to be sent to others */
7448: for (cp = 0, ncoo_o = 0; cp < mmdata->cp; cp++) {
7449: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7450: PetscInt *idxoff = mmdata->off[cp];
7451: PetscInt *idxown = mmdata->own[cp];
7452: if (!mptmp[cp] && rmapt[cp] == 2) { /* row map is sparse */
7453: const PetscInt *rmap = rmapa[cp];
7454: const PetscInt *cmap = cmapa[cp];
7455: const PetscInt *ii = mm->i;
7456: PetscInt *coi = coo_i + ncoo_o;
7457: PetscInt *coj = coo_j + ncoo_o;
7458: const PetscInt mr = mp[cp]->rmap->n;
7459: const PetscInt rs = C->rmap->rstart;
7460: const PetscInt re = C->rmap->rend;
7461: const PetscInt cs = C->cmap->rstart;
7462: for (i = 0; i < mr; i++) {
7463: const PetscInt *jj = mm->j + ii[i];
7464: const PetscInt gr = rmap[i];
7465: const PetscInt nz = ii[i + 1] - ii[i];
7466: if (gr < rs || gr >= re) { /* this is an offproc row */
7467: for (j = ii[i]; j < ii[i + 1]; j++) {
7468: *coi++ = gr;
7469: *idxoff++ = j;
7470: }
7471: if (!cmapt[cp]) { /* already global */
7472: for (j = 0; j < nz; j++) *coj++ = jj[j];
7473: } else if (cmapt[cp] == 1) { /* local to global for owned columns of C */
7474: for (j = 0; j < nz; j++) *coj++ = jj[j] + cs;
7475: } else { /* offdiag */
7476: for (j = 0; j < nz; j++) *coj++ = cmap[jj[j]];
7477: }
7478: ncoo_o += nz;
7479: } else { /* this is a local row */
7480: for (j = ii[i]; j < ii[i + 1]; j++) *idxown++ = j;
7481: }
7482: }
7483: }
7484: mmdata->off[cp + 1] = idxoff;
7485: mmdata->own[cp + 1] = idxown;
7486: }
7488: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)C), &mmdata->sf));
7489: PetscInt incoo_o;
7490: PetscCall(PetscIntCast(ncoo_o, &incoo_o));
7491: PetscCall(PetscSFSetGraphLayout(mmdata->sf, C->rmap, incoo_o /*nleaves*/, NULL /*ilocal*/, PETSC_OWN_POINTER, coo_i));
7492: PetscCall(PetscSFGetMultiSF(mmdata->sf, &msf));
7493: PetscCall(PetscSFGetGraph(msf, &ncoo2 /*nroots*/, NULL, NULL, NULL));
7494: ncoo = ncoo_d + ncoo_oown + ncoo2;
7495: PetscCall(PetscMalloc2(ncoo, &coo_i2, ncoo, &coo_j2));
7496: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_INT, coo_i, coo_i2 + ncoo_d + ncoo_oown)); /* put (i,j) of remote nonzeros at back */
7497: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_INT, coo_i, coo_i2 + ncoo_d + ncoo_oown));
7498: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_INT, coo_j, coo_j2 + ncoo_d + ncoo_oown));
7499: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_INT, coo_j, coo_j2 + ncoo_d + ncoo_oown));
7500: PetscCall(PetscFree2(coo_i, coo_j));
7501: /* allocate MPI send buffer to collect nonzero values to be sent to remote procs */
7502: PetscCall(PetscSFMalloc(mmdata->sf, mmdata->mtype, ncoo_o * sizeof(PetscScalar), (void **)&mmdata->coo_w));
7503: coo_i = coo_i2;
7504: coo_j = coo_j2;
7505: } else { /* no offproc values insertion */
7506: ncoo = ncoo_d;
7507: PetscCall(PetscMalloc2(ncoo, &coo_i, ncoo, &coo_j));
7509: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)C), &mmdata->sf));
7510: PetscCall(PetscSFSetGraph(mmdata->sf, 0, 0, NULL, PETSC_OWN_POINTER, NULL, PETSC_OWN_POINTER));
7511: PetscCall(PetscSFSetUp(mmdata->sf));
7512: }
7513: mmdata->hasoffproc = hasoffproc;
7515: /* gather (i,j) of nonzeros inserted locally */
7516: for (cp = 0, ncoo_d = 0; cp < mmdata->cp; cp++) {
7517: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7518: PetscInt *coi = coo_i + ncoo_d;
7519: PetscInt *coj = coo_j + ncoo_d;
7520: const PetscInt *jj = mm->j;
7521: const PetscInt *ii = mm->i;
7522: const PetscInt *cmap = cmapa[cp];
7523: const PetscInt *rmap = rmapa[cp];
7524: const PetscInt mr = mp[cp]->rmap->n;
7525: const PetscInt rs = C->rmap->rstart;
7526: const PetscInt re = C->rmap->rend;
7527: const PetscInt cs = C->cmap->rstart;
7529: if (mptmp[cp]) continue;
7530: if (rmapt[cp] == 1) { /* consecutive rows */
7531: /* fill coo_i */
7532: for (i = 0; i < mr; i++) {
7533: const PetscInt gr = i + rs;
7534: for (j = ii[i]; j < ii[i + 1]; j++) coi[j] = gr;
7535: }
7536: /* fill coo_j */
7537: if (!cmapt[cp]) { /* type-0, already global */
7538: PetscCall(PetscArraycpy(coj, jj, mm->nz));
7539: } else if (cmapt[cp] == 1) { /* type-1, local to global for consecutive columns of C */
7540: for (j = 0; j < mm->nz; j++) coj[j] = jj[j] + cs; /* lid + col start */
7541: } else { /* type-2, local to global for sparse columns */
7542: for (j = 0; j < mm->nz; j++) coj[j] = cmap[jj[j]];
7543: }
7544: ncoo_d += mm->nz;
7545: } else if (rmapt[cp] == 2) { /* sparse rows */
7546: for (i = 0; i < mr; i++) {
7547: const PetscInt *jj = mm->j + ii[i];
7548: const PetscInt gr = rmap[i];
7549: const PetscInt nz = ii[i + 1] - ii[i];
7550: if (gr >= rs && gr < re) { /* local rows */
7551: for (j = ii[i]; j < ii[i + 1]; j++) *coi++ = gr;
7552: if (!cmapt[cp]) { /* type-0, already global */
7553: for (j = 0; j < nz; j++) *coj++ = jj[j];
7554: } else if (cmapt[cp] == 1) { /* local to global for owned columns of C */
7555: for (j = 0; j < nz; j++) *coj++ = jj[j] + cs;
7556: } else { /* type-2, local to global for sparse columns */
7557: for (j = 0; j < nz; j++) *coj++ = cmap[jj[j]];
7558: }
7559: ncoo_d += nz;
7560: }
7561: }
7562: }
7563: }
7564: if (glob) PetscCall(ISRestoreIndices(glob, &globidx));
7565: PetscCall(ISDestroy(&glob));
7566: if (P_oth_l2g) PetscCall(ISLocalToGlobalMappingRestoreIndices(P_oth_l2g, &P_oth_idx));
7567: PetscCall(ISLocalToGlobalMappingDestroy(&P_oth_l2g));
7568: /* allocate an array to store all nonzeros (inserted locally or remotely) belonging to this proc */
7569: PetscCall(PetscSFMalloc(mmdata->sf, mmdata->mtype, ncoo * sizeof(PetscScalar), (void **)&mmdata->coo_v));
7571: /* set block sizes */
7572: A = product->A;
7573: P = product->B;
7574: switch (ptype) {
7575: case MATPRODUCT_PtAP:
7576: PetscCall(MatSetBlockSizes(C, P->cmap->bs, P->cmap->bs));
7577: break;
7578: case MATPRODUCT_RARt:
7579: PetscCall(MatSetBlockSizes(C, P->rmap->bs, P->rmap->bs));
7580: break;
7581: case MATPRODUCT_ABC:
7582: PetscCall(MatSetBlockSizesFromMats(C, A, product->C));
7583: break;
7584: case MATPRODUCT_AB:
7585: PetscCall(MatSetBlockSizesFromMats(C, A, P));
7586: break;
7587: case MATPRODUCT_AtB:
7588: PetscCall(MatSetBlockSizes(C, A->cmap->bs, P->cmap->bs));
7589: break;
7590: case MATPRODUCT_ABt:
7591: PetscCall(MatSetBlockSizes(C, A->rmap->bs, P->rmap->bs));
7592: break;
7593: default:
7594: SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Not for ProductType %s", MatProductTypes[ptype]);
7595: }
7597: /* preallocate with COO data */
7598: PetscCall(MatSetPreallocationCOO(C, ncoo, coo_i, coo_j));
7599: PetscCall(PetscFree2(coo_i, coo_j));
7600: PetscFunctionReturn(PETSC_SUCCESS);
7601: }
7603: PetscErrorCode MatProductSetFromOptions_MPIAIJBACKEND(Mat mat)
7604: {
7605: Mat_Product *product = mat->product;
7606: #if defined(PETSC_HAVE_DEVICE)
7607: PetscBool match = PETSC_FALSE;
7608: PetscBool usecpu = PETSC_FALSE;
7609: #else
7610: PetscBool match = PETSC_TRUE;
7611: #endif
7613: PetscFunctionBegin;
7614: MatCheckProduct(mat, 1);
7615: #if defined(PETSC_HAVE_DEVICE)
7616: if (!product->A->boundtocpu && !product->B->boundtocpu) PetscCall(PetscObjectTypeCompare((PetscObject)product->B, ((PetscObject)product->A)->type_name, &match));
7617: if (match) { /* we can always fallback to the CPU if requested */
7618: switch (product->type) {
7619: case MATPRODUCT_AB:
7620: if (product->api_user) {
7621: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatMatMult", "Mat");
7622: PetscCall(PetscOptionsBool("-matmatmult_backend_cpu", "Use CPU code", "MatMatMult", usecpu, &usecpu, NULL));
7623: PetscOptionsEnd();
7624: } else {
7625: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_AB", "Mat");
7626: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatMatMult", usecpu, &usecpu, NULL));
7627: PetscOptionsEnd();
7628: }
7629: break;
7630: case MATPRODUCT_AtB:
7631: if (product->api_user) {
7632: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatTransposeMatMult", "Mat");
7633: PetscCall(PetscOptionsBool("-mattransposematmult_backend_cpu", "Use CPU code", "MatTransposeMatMult", usecpu, &usecpu, NULL));
7634: PetscOptionsEnd();
7635: } else {
7636: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_AtB", "Mat");
7637: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatTransposeMatMult", usecpu, &usecpu, NULL));
7638: PetscOptionsEnd();
7639: }
7640: break;
7641: case MATPRODUCT_PtAP:
7642: if (product->api_user) {
7643: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatPtAP", "Mat");
7644: PetscCall(PetscOptionsBool("-matptap_backend_cpu", "Use CPU code", "MatPtAP", usecpu, &usecpu, NULL));
7645: PetscOptionsEnd();
7646: } else {
7647: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_PtAP", "Mat");
7648: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatPtAP", usecpu, &usecpu, NULL));
7649: PetscOptionsEnd();
7650: }
7651: break;
7652: default:
7653: break;
7654: }
7655: match = (PetscBool)!usecpu;
7656: }
7657: #endif
7658: if (match) {
7659: switch (product->type) {
7660: case MATPRODUCT_AB:
7661: case MATPRODUCT_AtB:
7662: case MATPRODUCT_PtAP:
7663: mat->ops->productsymbolic = MatProductSymbolic_MPIAIJBACKEND;
7664: break;
7665: default:
7666: break;
7667: }
7668: }
7669: /* fallback to MPIAIJ ops */
7670: if (!mat->ops->productsymbolic) PetscCall(MatProductSetFromOptions_MPIAIJ(mat));
7671: PetscFunctionReturn(PETSC_SUCCESS);
7672: }
7674: /*
7675: Produces a set of block column indices of the matrix row, one for each block represented in the original row
7677: n - the number of block indices in cc[]
7678: cc - the block indices (must be large enough to contain the indices)
7679: */
7680: static inline PetscErrorCode MatCollapseRow(Mat Amat, PetscInt row, PetscInt bs, PetscInt *n, PetscInt *cc)
7681: {
7682: PetscInt cnt = -1, nidx, j;
7683: const PetscInt *idx;
7685: PetscFunctionBegin;
7686: PetscCall(MatGetRow(Amat, row, &nidx, &idx, NULL));
7687: if (nidx) {
7688: cnt = 0;
7689: cc[cnt] = idx[0] / bs;
7690: for (j = 1; j < nidx; j++) {
7691: if (cc[cnt] < idx[j] / bs) cc[++cnt] = idx[j] / bs;
7692: }
7693: }
7694: PetscCall(MatRestoreRow(Amat, row, &nidx, &idx, NULL));
7695: *n = cnt + 1;
7696: PetscFunctionReturn(PETSC_SUCCESS);
7697: }
7699: /*
7700: Produces a set of block column indices of the matrix block row, one for each block represented in the original set of rows
7702: ncollapsed - the number of block indices
7703: collapsed - the block indices (must be large enough to contain the indices)
7704: */
7705: static inline PetscErrorCode MatCollapseRows(Mat Amat, PetscInt start, PetscInt bs, PetscInt *w0, PetscInt *w1, PetscInt *w2, PetscInt *ncollapsed, PetscInt **collapsed)
7706: {
7707: PetscInt i, nprev, *cprev = w0, ncur = 0, *ccur = w1, *merged = w2, *cprevtmp;
7709: PetscFunctionBegin;
7710: PetscCall(MatCollapseRow(Amat, start, bs, &nprev, cprev));
7711: for (i = start + 1; i < start + bs; i++) {
7712: PetscCall(MatCollapseRow(Amat, i, bs, &ncur, ccur));
7713: PetscCall(PetscMergeIntArray(nprev, cprev, ncur, ccur, &nprev, &merged));
7714: cprevtmp = cprev;
7715: cprev = merged;
7716: merged = cprevtmp;
7717: }
7718: *ncollapsed = nprev;
7719: if (collapsed) *collapsed = cprev;
7720: PetscFunctionReturn(PETSC_SUCCESS);
7721: }
7723: /*
7724: MatCreateGraph_Simple_AIJ - create simple scalar matrix (graph) from potentially blocked matrix
7726: Input Parameter:
7727: . Amat - matrix
7728: - symmetrize - make the result symmetric
7729: + scale - scale with diagonal
7731: Output Parameter:
7732: . a_Gmat - output scalar graph >= 0
7734: */
7735: PETSC_INTERN PetscErrorCode MatCreateGraph_Simple_AIJ(Mat Amat, PetscBool symmetrize, PetscBool scale, PetscReal filter, PetscInt index_size, PetscInt index[], Mat *a_Gmat)
7736: {
7737: PetscInt Istart, Iend, Ii, jj, kk, ncols, nloc, NN, MM, bs;
7738: MPI_Comm comm;
7739: Mat Gmat;
7740: PetscBool ismpiaij, isseqaij;
7741: Mat a, b, c;
7742: MatType jtype;
7744: PetscFunctionBegin;
7745: PetscCall(PetscObjectGetComm((PetscObject)Amat, &comm));
7746: PetscCall(MatGetOwnershipRange(Amat, &Istart, &Iend));
7747: PetscCall(MatGetSize(Amat, &MM, &NN));
7748: PetscCall(MatGetBlockSize(Amat, &bs));
7749: nloc = (Iend - Istart) / bs;
7751: PetscCall(PetscObjectBaseTypeCompare((PetscObject)Amat, MATSEQAIJ, &isseqaij));
7752: PetscCall(PetscObjectBaseTypeCompare((PetscObject)Amat, MATMPIAIJ, &ismpiaij));
7753: PetscCheck(isseqaij || ismpiaij, comm, PETSC_ERR_USER, "Require (MPI)AIJ matrix type");
7755: /* TODO GPU: these calls are potentially expensive if matrices are large and we want to use the GPU */
7756: /* A solution consists in providing a new API, MatAIJGetCollapsedAIJ, and each class can provide a fast
7757: implementation */
7758: if (bs > 1) {
7759: PetscCall(MatGetType(Amat, &jtype));
7760: PetscCall(MatCreate(comm, &Gmat));
7761: PetscCall(MatSetType(Gmat, jtype));
7762: PetscCall(MatSetSizes(Gmat, nloc, nloc, PETSC_DETERMINE, PETSC_DETERMINE));
7763: PetscCall(MatSetBlockSizes(Gmat, 1, 1));
7764: if (isseqaij || ((Mat_MPIAIJ *)Amat->data)->garray) {
7765: PetscInt *d_nnz, *o_nnz;
7766: MatScalar *aa, val, *AA;
7767: PetscInt *aj, *ai, *AJ, nc, nmax = 0;
7769: if (isseqaij) {
7770: a = Amat;
7771: b = NULL;
7772: } else {
7773: Mat_MPIAIJ *d = (Mat_MPIAIJ *)Amat->data;
7774: a = d->A;
7775: b = d->B;
7776: }
7777: PetscCall(PetscInfo(Amat, "New bs>1 Graph. nloc=%" PetscInt_FMT "\n", nloc));
7778: PetscCall(PetscMalloc2(nloc, &d_nnz, (isseqaij ? 0 : nloc), &o_nnz));
7779: for (c = a, kk = 0; c && kk < 2; c = b, kk++) {
7780: PetscInt *nnz = (c == a) ? d_nnz : o_nnz;
7781: const PetscInt *cols1, *cols2;
7783: for (PetscInt brow = 0, nc1, nc2, ok = 1; brow < nloc * bs; brow += bs) { // block rows
7784: PetscCall(MatGetRow(c, brow, &nc2, &cols2, NULL));
7785: nnz[brow / bs] = nc2 / bs;
7786: if (nc2 % bs) ok = 0;
7787: if (nnz[brow / bs] > nmax) nmax = nnz[brow / bs];
7788: for (PetscInt ii = 1; ii < bs; ii++) { // check for non-dense blocks
7789: PetscCall(MatGetRow(c, brow + ii, &nc1, &cols1, NULL));
7790: if (nc1 != nc2) ok = 0;
7791: else {
7792: for (PetscInt jj = 0; jj < nc1 && ok == 1; jj++) {
7793: if (cols1[jj] != cols2[jj]) ok = 0;
7794: if (cols1[jj] % bs != jj % bs) ok = 0;
7795: }
7796: }
7797: PetscCall(MatRestoreRow(c, brow + ii, &nc1, &cols1, NULL));
7798: }
7799: PetscCall(MatRestoreRow(c, brow, &nc2, &cols2, NULL));
7800: if (!ok) {
7801: PetscCall(PetscFree2(d_nnz, o_nnz));
7802: PetscCall(PetscInfo(Amat, "Found sparse blocks - revert to slow method\n"));
7803: goto old_bs;
7804: }
7805: }
7806: }
7807: PetscCall(MatSeqAIJSetPreallocation(Gmat, 0, d_nnz));
7808: PetscCall(MatMPIAIJSetPreallocation(Gmat, 0, d_nnz, 0, o_nnz));
7809: PetscCall(PetscFree2(d_nnz, o_nnz));
7810: PetscCall(PetscMalloc2(nmax, &AA, nmax, &AJ));
7811: // diag
7812: for (PetscInt brow = 0, n, grow; brow < nloc * bs; brow += bs) { // block rows
7813: Mat_SeqAIJ *aseq = (Mat_SeqAIJ *)a->data;
7815: ai = aseq->i;
7816: n = ai[brow + 1] - ai[brow];
7817: aj = aseq->j + ai[brow];
7818: for (PetscInt k = 0; k < n; k += bs) { // block columns
7819: AJ[k / bs] = aj[k] / bs + Istart / bs; // diag starts at (Istart,Istart)
7820: val = 0;
7821: if (index_size == 0) {
7822: for (PetscInt ii = 0; ii < bs; ii++) { // rows in block
7823: aa = aseq->a + ai[brow + ii] + k;
7824: for (PetscInt jj = 0; jj < bs; jj++) { // columns in block
7825: val += PetscAbs(PetscRealPart(aa[jj])); // a sort of norm
7826: }
7827: }
7828: } else { // use (index,index) value if provided
7829: for (PetscInt iii = 0; iii < index_size; iii++) { // rows in block
7830: PetscInt ii = index[iii];
7831: aa = aseq->a + ai[brow + ii] + k;
7832: for (PetscInt jjj = 0; jjj < index_size; jjj++) { // columns in block
7833: PetscInt jj = index[jjj];
7834: val += PetscAbs(PetscRealPart(aa[jj]));
7835: }
7836: }
7837: }
7838: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs (%" PetscInt_FMT ") >= nmax (%" PetscInt_FMT ")", k / bs, nmax);
7839: AA[k / bs] = val;
7840: }
7841: grow = Istart / bs + brow / bs;
7842: PetscCall(MatSetValues(Gmat, 1, &grow, n / bs, AJ, AA, ADD_VALUES));
7843: }
7844: // off-diag
7845: if (ismpiaij) {
7846: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)Amat->data;
7847: const PetscScalar *vals;
7848: const PetscInt *cols, *garray = aij->garray;
7850: PetscCheck(garray, PETSC_COMM_SELF, PETSC_ERR_USER, "No garray ?");
7851: for (PetscInt brow = 0, grow; brow < nloc * bs; brow += bs) { // block rows
7852: PetscCall(MatGetRow(b, brow, &ncols, &cols, NULL));
7853: for (PetscInt k = 0, cidx = 0; k < ncols; k += bs, cidx++) {
7854: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs >= nmax");
7855: AA[k / bs] = 0;
7856: AJ[cidx] = garray[cols[k]] / bs;
7857: }
7858: nc = ncols / bs;
7859: PetscCall(MatRestoreRow(b, brow, &ncols, &cols, NULL));
7860: if (index_size == 0) {
7861: for (PetscInt ii = 0; ii < bs; ii++) { // rows in block
7862: PetscCall(MatGetRow(b, brow + ii, &ncols, &cols, &vals));
7863: for (PetscInt k = 0; k < ncols; k += bs) {
7864: for (PetscInt jj = 0; jj < bs; jj++) { // cols in block
7865: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs (%" PetscInt_FMT ") >= nmax (%" PetscInt_FMT ")", k / bs, nmax);
7866: AA[k / bs] += PetscAbs(PetscRealPart(vals[k + jj]));
7867: }
7868: }
7869: PetscCall(MatRestoreRow(b, brow + ii, &ncols, &cols, &vals));
7870: }
7871: } else { // use (index,index) value if provided
7872: for (PetscInt iii = 0; iii < index_size; iii++) { // rows in block
7873: PetscInt ii = index[iii];
7874: PetscCall(MatGetRow(b, brow + ii, &ncols, &cols, &vals));
7875: for (PetscInt k = 0; k < ncols; k += bs) {
7876: for (PetscInt jjj = 0; jjj < index_size; jjj++) { // cols in block
7877: PetscInt jj = index[jjj];
7878: AA[k / bs] += PetscAbs(PetscRealPart(vals[k + jj]));
7879: }
7880: }
7881: PetscCall(MatRestoreRow(b, brow + ii, &ncols, &cols, &vals));
7882: }
7883: }
7884: grow = Istart / bs + brow / bs;
7885: PetscCall(MatSetValues(Gmat, 1, &grow, nc, AJ, AA, ADD_VALUES));
7886: }
7887: }
7888: PetscCall(MatAssemblyBegin(Gmat, MAT_FINAL_ASSEMBLY));
7889: PetscCall(MatAssemblyEnd(Gmat, MAT_FINAL_ASSEMBLY));
7890: PetscCall(PetscFree2(AA, AJ));
7891: } else {
7892: const PetscScalar *vals;
7893: const PetscInt *idx;
7894: PetscInt *d_nnz, *o_nnz, *w0, *w1, *w2;
7895: old_bs:
7896: /*
7897: Determine the preallocation needed for the scalar matrix derived from the vector matrix.
7898: */
7899: PetscCall(PetscInfo(Amat, "OLD bs>1 CreateGraph\n"));
7900: PetscCall(PetscMalloc2(nloc, &d_nnz, (isseqaij ? 0 : nloc), &o_nnz));
7901: if (isseqaij) {
7902: PetscInt max_d_nnz;
7904: /*
7905: Determine exact preallocation count for (sequential) scalar matrix
7906: */
7907: PetscCall(MatSeqAIJGetMaxRowNonzeros(Amat, &max_d_nnz));
7908: max_d_nnz = PetscMin(nloc, bs * max_d_nnz);
7909: PetscCall(PetscMalloc3(max_d_nnz, &w0, max_d_nnz, &w1, max_d_nnz, &w2));
7910: for (Ii = 0, jj = 0; Ii < Iend; Ii += bs, jj++) PetscCall(MatCollapseRows(Amat, Ii, bs, w0, w1, w2, &d_nnz[jj], NULL));
7911: PetscCall(PetscFree3(w0, w1, w2));
7912: } else if (ismpiaij) {
7913: Mat Daij, Oaij;
7914: const PetscInt *garray;
7915: PetscInt max_d_nnz;
7917: PetscCall(MatMPIAIJGetSeqAIJ(Amat, &Daij, &Oaij, &garray));
7918: /*
7919: Determine exact preallocation count for diagonal block portion of scalar matrix
7920: */
7921: PetscCall(MatSeqAIJGetMaxRowNonzeros(Daij, &max_d_nnz));
7922: max_d_nnz = PetscMin(nloc, bs * max_d_nnz);
7923: PetscCall(PetscMalloc3(max_d_nnz, &w0, max_d_nnz, &w1, max_d_nnz, &w2));
7924: for (Ii = 0, jj = 0; Ii < Iend - Istart; Ii += bs, jj++) PetscCall(MatCollapseRows(Daij, Ii, bs, w0, w1, w2, &d_nnz[jj], NULL));
7925: PetscCall(PetscFree3(w0, w1, w2));
7926: /*
7927: Over estimate (usually grossly over), preallocation count for off-diagonal portion of scalar matrix
7928: */
7929: for (Ii = 0, jj = 0; Ii < Iend - Istart; Ii += bs, jj++) {
7930: o_nnz[jj] = 0;
7931: for (kk = 0; kk < bs; kk++) { /* rows that get collapsed to a single row */
7932: PetscCall(MatGetRow(Oaij, Ii + kk, &ncols, NULL, NULL));
7933: o_nnz[jj] += ncols;
7934: PetscCall(MatRestoreRow(Oaij, Ii + kk, &ncols, NULL, NULL));
7935: }
7936: if (o_nnz[jj] > (NN / bs - nloc)) o_nnz[jj] = NN / bs - nloc;
7937: }
7938: } else SETERRQ(comm, PETSC_ERR_USER, "Require AIJ matrix type");
7939: /* get scalar copy (norms) of matrix */
7940: PetscCall(MatSeqAIJSetPreallocation(Gmat, 0, d_nnz));
7941: PetscCall(MatMPIAIJSetPreallocation(Gmat, 0, d_nnz, 0, o_nnz));
7942: PetscCall(PetscFree2(d_nnz, o_nnz));
7943: for (Ii = Istart; Ii < Iend; Ii++) {
7944: PetscInt dest_row = Ii / bs;
7946: PetscCall(MatGetRow(Amat, Ii, &ncols, &idx, &vals));
7947: for (jj = 0; jj < ncols; jj++) {
7948: PetscInt dest_col = idx[jj] / bs;
7949: PetscScalar sv = PetscAbs(PetscRealPart(vals[jj]));
7951: PetscCall(MatSetValues(Gmat, 1, &dest_row, 1, &dest_col, &sv, ADD_VALUES));
7952: }
7953: PetscCall(MatRestoreRow(Amat, Ii, &ncols, &idx, &vals));
7954: }
7955: PetscCall(MatAssemblyBegin(Gmat, MAT_FINAL_ASSEMBLY));
7956: PetscCall(MatAssemblyEnd(Gmat, MAT_FINAL_ASSEMBLY));
7957: }
7958: } else {
7959: if (symmetrize || filter >= 0 || scale) PetscCall(MatDuplicate(Amat, MAT_COPY_VALUES, &Gmat));
7960: else {
7961: Gmat = Amat;
7962: PetscCall(PetscObjectReference((PetscObject)Gmat));
7963: }
7964: if (isseqaij) {
7965: a = Gmat;
7966: b = NULL;
7967: } else {
7968: Mat_MPIAIJ *d = (Mat_MPIAIJ *)Gmat->data;
7969: a = d->A;
7970: b = d->B;
7971: }
7972: if (filter >= 0 || scale) {
7973: /* take absolute value of each entry */
7974: for (c = a, kk = 0; c && kk < 2; c = b, kk++) {
7975: MatInfo info;
7976: PetscScalar *avals;
7978: PetscCall(MatGetInfo(c, MAT_LOCAL, &info));
7979: PetscCall(MatSeqAIJGetArray(c, &avals));
7980: for (int jj = 0; jj < info.nz_used; jj++) avals[jj] = PetscAbsScalar(avals[jj]);
7981: PetscCall(MatSeqAIJRestoreArray(c, &avals));
7982: }
7983: }
7984: }
7985: if (symmetrize) {
7986: PetscBool isset, issym;
7988: PetscCall(MatIsSymmetricKnown(Amat, &isset, &issym));
7989: if (!isset || !issym) {
7990: Mat matTrans;
7992: PetscCall(MatTranspose(Gmat, MAT_INITIAL_MATRIX, &matTrans));
7993: PetscCall(MatAXPY(Gmat, 1.0, matTrans, Gmat->structurally_symmetric == PETSC_BOOL3_TRUE ? SAME_NONZERO_PATTERN : DIFFERENT_NONZERO_PATTERN));
7994: PetscCall(MatDestroy(&matTrans));
7995: }
7996: PetscCall(MatSetOption(Gmat, MAT_SYMMETRIC, PETSC_TRUE));
7997: } else if (Amat != Gmat) PetscCall(MatPropagateSymmetryOptions(Amat, Gmat));
7998: if (scale) {
7999: /* scale c for all diagonal values = 1 or -1 */
8000: Vec diag;
8002: PetscCall(MatCreateVecs(Gmat, &diag, NULL));
8003: PetscCall(MatGetDiagonal(Gmat, diag));
8004: PetscCall(VecReciprocal(diag));
8005: PetscCall(VecSqrtAbs(diag));
8006: PetscCall(MatDiagonalScale(Gmat, diag, diag));
8007: PetscCall(VecDestroy(&diag));
8008: }
8009: PetscCall(MatViewFromOptions(Gmat, NULL, "-mat_graph_view"));
8010: if (filter >= 0) {
8011: PetscCall(MatFilter(Gmat, filter, PETSC_TRUE, PETSC_TRUE));
8012: PetscCall(MatViewFromOptions(Gmat, NULL, "-mat_filter_graph_view"));
8013: }
8014: *a_Gmat = Gmat;
8015: PetscFunctionReturn(PETSC_SUCCESS);
8016: }
8018: PETSC_INTERN PetscErrorCode MatGetCurrentMemType_MPIAIJ(Mat A, PetscMemType *memtype)
8019: {
8020: Mat_MPIAIJ *mpiaij = (Mat_MPIAIJ *)A->data;
8021: PetscMemType mD = PETSC_MEMTYPE_HOST, mO = PETSC_MEMTYPE_HOST;
8023: PetscFunctionBegin;
8024: if (mpiaij->A) PetscCall(MatGetCurrentMemType(mpiaij->A, &mD));
8025: if (mpiaij->B) PetscCall(MatGetCurrentMemType(mpiaij->B, &mO));
8026: *memtype = (mD == mO) ? mD : PETSC_MEMTYPE_HOST;
8027: PetscFunctionReturn(PETSC_SUCCESS);
8028: }
8030: /*
8031: Special version for direct calls from Fortran
8032: */
8034: /* Change these macros so can be used in void function */
8035: /* Identical to PetscCallVoid, except it assigns to *_ierr */
8036: #undef PetscCall
8037: #define PetscCall(...) \
8038: do { \
8039: PetscErrorCode ierr_msv_mpiaij = __VA_ARGS__; \
8040: if (PetscUnlikely(ierr_msv_mpiaij)) { \
8041: *_ierr = PetscError(PETSC_COMM_SELF, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr_msv_mpiaij, PETSC_ERROR_REPEAT, " "); \
8042: return; \
8043: } \
8044: } while (0)
8046: #undef SETERRQ
8047: #define SETERRQ(comm, ierr, ...) \
8048: do { \
8049: *_ierr = PetscError(comm, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr, PETSC_ERROR_INITIAL, __VA_ARGS__); \
8050: return; \
8051: } while (0)
8053: #if defined(PETSC_HAVE_FORTRAN_CAPS)
8054: #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ
8055: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
8056: #define matsetvaluesmpiaij_ matsetvaluesmpiaij
8057: #else
8058: #endif
8059: PETSC_EXTERN void matsetvaluesmpiaij_(Mat *mmat, PetscInt *mm, const PetscInt im[], PetscInt *mn, const PetscInt in[], const PetscScalar v[], InsertMode *maddv, PetscErrorCode *_ierr)
8060: {
8061: Mat mat = *mmat;
8062: PetscInt m = *mm, n = *mn;
8063: InsertMode addv = *maddv;
8064: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
8065: PetscScalar value;
8067: MatCheckPreallocated(mat, 1);
8068: if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv;
8069: else PetscCheck(mat->insertmode == addv, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Cannot mix add values and insert values");
8070: {
8071: PetscInt i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend;
8072: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend, row, col;
8073: PetscBool roworiented = aij->roworiented;
8075: /* Some Variables required in the macro */
8076: Mat A = aij->A;
8077: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
8078: PetscInt *aimax = a->imax, *ai = a->i, *ailen = a->ilen, *aj = a->j;
8079: MatScalar *aa;
8080: PetscBool ignorezeroentries = ((a->ignorezeroentries && (addv == ADD_VALUES)) ? PETSC_TRUE : PETSC_FALSE);
8081: Mat B = aij->B;
8082: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
8083: PetscInt *bimax = b->imax, *bi = b->i, *bilen = b->ilen, *bj = b->j, bm = aij->B->rmap->n, am = aij->A->rmap->n;
8084: MatScalar *ba;
8085: /* This variable below is only for the PETSC_HAVE_VIENNACL or PETSC_HAVE_CUDA cases, but we define it in all cases because we
8086: * cannot use "#if defined" inside a macro. */
8087: PETSC_UNUSED PetscBool inserted = PETSC_FALSE;
8089: PetscInt *rp1, *rp2, ii, nrow1, nrow2, _i, rmax1, rmax2, N, low1, high1, low2, high2, t, lastcol1, lastcol2;
8090: PetscInt nonew = a->nonew;
8091: MatScalar *ap1, *ap2;
8093: PetscFunctionBegin;
8094: PetscCall(MatSeqAIJGetArray(A, &aa));
8095: PetscCall(MatSeqAIJGetArray(B, &ba));
8096: for (i = 0; i < m; i++) {
8097: if (im[i] < 0) continue;
8098: 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);
8099: if (im[i] >= rstart && im[i] < rend) {
8100: row = im[i] - rstart;
8101: lastcol1 = -1;
8102: rp1 = aj + ai[row];
8103: ap1 = aa + ai[row];
8104: rmax1 = aimax[row];
8105: nrow1 = ailen[row];
8106: low1 = 0;
8107: high1 = nrow1;
8108: lastcol2 = -1;
8109: rp2 = bj + bi[row];
8110: ap2 = ba + bi[row];
8111: rmax2 = bimax[row];
8112: nrow2 = bilen[row];
8113: low2 = 0;
8114: high2 = nrow2;
8116: for (j = 0; j < n; j++) {
8117: if (roworiented) value = v[i * n + j];
8118: else value = v[i + j * m];
8119: if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && im[i] != in[j]) continue;
8120: if (in[j] >= cstart && in[j] < cend) {
8121: col = in[j] - cstart;
8122: MatSetValues_SeqAIJ_A_Private(row, col, value, addv, im[i], in[j]);
8123: } else if (in[j] < 0) continue;
8124: else if (PetscUnlikelyDebug(in[j] >= mat->cmap->N)) {
8125: SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large: col %" PetscInt_FMT " max %" PetscInt_FMT, in[j], mat->cmap->N - 1);
8126: } else {
8127: if (mat->was_assembled) {
8128: if (!aij->colmap) PetscCall(MatCreateColmap_MPIAIJ_Private(mat));
8129: #if defined(PETSC_USE_CTABLE)
8130: PetscCall(PetscHMapIGetWithDefault(aij->colmap, in[j] + 1, 0, &col));
8131: col--;
8132: #else
8133: col = aij->colmap[in[j]] - 1;
8134: #endif
8135: if (col < 0 && !((Mat_SeqAIJ *)aij->A->data)->nonew) {
8136: PetscCall(MatDisAssemble_MPIAIJ(mat, PETSC_FALSE));
8137: col = in[j];
8138: /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
8139: B = aij->B;
8140: b = (Mat_SeqAIJ *)B->data;
8141: bimax = b->imax;
8142: bi = b->i;
8143: bilen = b->ilen;
8144: bj = b->j;
8145: rp2 = bj + bi[row];
8146: ap2 = ba + bi[row];
8147: rmax2 = bimax[row];
8148: nrow2 = bilen[row];
8149: low2 = 0;
8150: high2 = nrow2;
8151: bm = aij->B->rmap->n;
8152: ba = b->a;
8153: inserted = PETSC_FALSE;
8154: }
8155: } else col = in[j];
8156: MatSetValues_SeqAIJ_B_Private(row, col, value, addv, im[i], in[j]);
8157: }
8158: }
8159: } else if (!aij->donotstash) {
8160: if (roworiented) {
8161: PetscCall(MatStashValuesRow_Private(&mat->stash, im[i], n, in, v + i * n, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
8162: } else {
8163: PetscCall(MatStashValuesCol_Private(&mat->stash, im[i], n, in, v + i, m, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
8164: }
8165: }
8166: }
8167: PetscCall(MatSeqAIJRestoreArray(A, &aa));
8168: PetscCall(MatSeqAIJRestoreArray(B, &ba));
8169: }
8170: PetscFunctionReturnVoid();
8171: }
8173: /* Undefining these here since they were redefined from their original definition above! No
8174: * other PETSc functions should be defined past this point, as it is impossible to recover the
8175: * original definitions */
8176: #undef PetscCall
8177: #undef SETERRQ