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: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2131: PetscFunctionBegin;
2132: PetscCall(MatConjugate_SeqAIJ(aij->A));
2133: PetscCall(MatConjugate_SeqAIJ(aij->B));
2134: PetscFunctionReturn(PETSC_SUCCESS);
2135: }
2137: static PetscErrorCode MatRealPart_MPIAIJ(Mat A)
2138: {
2139: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2141: PetscFunctionBegin;
2142: PetscCall(MatRealPart(a->A));
2143: PetscCall(MatRealPart(a->B));
2144: PetscFunctionReturn(PETSC_SUCCESS);
2145: }
2147: static PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A)
2148: {
2149: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2151: PetscFunctionBegin;
2152: PetscCall(MatImaginaryPart(a->A));
2153: PetscCall(MatImaginaryPart(a->B));
2154: PetscFunctionReturn(PETSC_SUCCESS);
2155: }
2157: static PetscErrorCode MatGetRowMaxAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2158: {
2159: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2160: PetscInt i, *idxb = NULL, m = A->rmap->n;
2161: PetscScalar *vv;
2162: Vec vB, vA;
2163: const PetscScalar *va, *vb;
2165: PetscFunctionBegin;
2166: PetscCall(MatCreateVecs(a->A, NULL, &vA));
2167: PetscCall(MatGetRowMaxAbs(a->A, vA, idx));
2169: PetscCall(VecGetArrayRead(vA, &va));
2170: if (idx) {
2171: for (i = 0; i < m; i++) {
2172: if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart;
2173: }
2174: }
2176: PetscCall(MatCreateVecs(a->B, NULL, &vB));
2177: PetscCall(PetscMalloc1(m, &idxb));
2178: PetscCall(MatGetRowMaxAbs(a->B, vB, idxb));
2180: PetscCall(VecGetArrayWrite(v, &vv));
2181: PetscCall(VecGetArrayRead(vB, &vb));
2182: for (i = 0; i < m; i++) {
2183: if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) {
2184: vv[i] = vb[i];
2185: if (idx) idx[i] = a->garray[idxb[i]];
2186: } else {
2187: vv[i] = va[i];
2188: if (idx && PetscAbsScalar(va[i]) == PetscAbsScalar(vb[i]) && idxb[i] != -1 && idx[i] > a->garray[idxb[i]]) idx[i] = a->garray[idxb[i]];
2189: }
2190: }
2191: PetscCall(VecRestoreArrayWrite(v, &vv));
2192: PetscCall(VecRestoreArrayRead(vA, &va));
2193: PetscCall(VecRestoreArrayRead(vB, &vb));
2194: PetscCall(PetscFree(idxb));
2195: PetscCall(VecDestroy(&vA));
2196: PetscCall(VecDestroy(&vB));
2197: PetscFunctionReturn(PETSC_SUCCESS);
2198: }
2200: static PetscErrorCode MatGetRowSumAbs_MPIAIJ(Mat A, Vec v)
2201: {
2202: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2203: Vec vB, vA;
2205: PetscFunctionBegin;
2206: PetscCall(MatCreateVecs(a->A, NULL, &vA));
2207: PetscCall(MatGetRowSumAbs(a->A, vA));
2208: PetscCall(MatCreateVecs(a->B, NULL, &vB));
2209: PetscCall(MatGetRowSumAbs(a->B, vB));
2210: PetscCall(VecAXPY(vA, 1.0, vB));
2211: PetscCall(VecDestroy(&vB));
2212: PetscCall(VecCopy(vA, v));
2213: PetscCall(VecDestroy(&vA));
2214: PetscFunctionReturn(PETSC_SUCCESS);
2215: }
2217: static PetscErrorCode MatGetRowMinAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2218: {
2219: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
2220: PetscInt m = A->rmap->n, n = A->cmap->n;
2221: PetscInt cstart = A->cmap->rstart, cend = A->cmap->rend;
2222: PetscInt *cmap = mat->garray;
2223: PetscInt *diagIdx, *offdiagIdx;
2224: Vec diagV, offdiagV;
2225: PetscScalar *a, *diagA, *offdiagA;
2226: const PetscScalar *ba, *bav;
2227: PetscInt r, j, col, ncols, *bi, *bj;
2228: Mat B = mat->B;
2229: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2231: PetscFunctionBegin;
2232: /* When a process holds entire A and other processes have no entry */
2233: if (A->cmap->N == n) {
2234: PetscCall(VecGetArrayWrite(v, &diagA));
2235: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2236: PetscCall(MatGetRowMinAbs(mat->A, diagV, idx));
2237: PetscCall(VecDestroy(&diagV));
2238: PetscCall(VecRestoreArrayWrite(v, &diagA));
2239: PetscFunctionReturn(PETSC_SUCCESS);
2240: } else if (n == 0) {
2241: if (m) {
2242: PetscCall(VecGetArrayWrite(v, &a));
2243: for (r = 0; r < m; r++) {
2244: a[r] = 0.0;
2245: if (idx) idx[r] = -1;
2246: }
2247: PetscCall(VecRestoreArrayWrite(v, &a));
2248: }
2249: PetscFunctionReturn(PETSC_SUCCESS);
2250: }
2252: PetscCall(PetscMalloc2(m, &diagIdx, m, &offdiagIdx));
2253: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2254: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2255: PetscCall(MatGetRowMinAbs(mat->A, diagV, diagIdx));
2257: /* Get offdiagIdx[] for implicit 0.0 */
2258: PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2259: ba = bav;
2260: bi = b->i;
2261: bj = b->j;
2262: PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2263: for (r = 0; r < m; r++) {
2264: ncols = bi[r + 1] - bi[r];
2265: if (ncols == A->cmap->N - n) { /* Brow is dense */
2266: offdiagA[r] = *ba;
2267: offdiagIdx[r] = cmap[0];
2268: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2269: offdiagA[r] = 0.0;
2271: /* Find first hole in the cmap */
2272: for (j = 0; j < ncols; j++) {
2273: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2274: if (col > j && j < cstart) {
2275: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2276: break;
2277: } else if (col > j + n && j >= cstart) {
2278: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2279: break;
2280: }
2281: }
2282: if (j == ncols && ncols < A->cmap->N - n) {
2283: /* a hole is outside compressed Bcols */
2284: if (ncols == 0) {
2285: if (cstart) {
2286: offdiagIdx[r] = 0;
2287: } else offdiagIdx[r] = cend;
2288: } else { /* ncols > 0 */
2289: offdiagIdx[r] = cmap[ncols - 1] + 1;
2290: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2291: }
2292: }
2293: }
2295: for (j = 0; j < ncols; j++) {
2296: if (PetscAbsScalar(offdiagA[r]) > PetscAbsScalar(*ba)) {
2297: offdiagA[r] = *ba;
2298: offdiagIdx[r] = cmap[*bj];
2299: }
2300: ba++;
2301: bj++;
2302: }
2303: }
2305: PetscCall(VecGetArrayWrite(v, &a));
2306: PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2307: for (r = 0; r < m; ++r) {
2308: if (PetscAbsScalar(diagA[r]) < PetscAbsScalar(offdiagA[r])) {
2309: a[r] = diagA[r];
2310: if (idx) idx[r] = cstart + diagIdx[r];
2311: } else if (PetscAbsScalar(diagA[r]) == PetscAbsScalar(offdiagA[r])) {
2312: a[r] = diagA[r];
2313: if (idx) {
2314: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2315: idx[r] = cstart + diagIdx[r];
2316: } else idx[r] = offdiagIdx[r];
2317: }
2318: } else {
2319: a[r] = offdiagA[r];
2320: if (idx) idx[r] = offdiagIdx[r];
2321: }
2322: }
2323: PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2324: PetscCall(VecRestoreArrayWrite(v, &a));
2325: PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2326: PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2327: PetscCall(VecDestroy(&diagV));
2328: PetscCall(VecDestroy(&offdiagV));
2329: PetscCall(PetscFree2(diagIdx, offdiagIdx));
2330: PetscFunctionReturn(PETSC_SUCCESS);
2331: }
2333: static PetscErrorCode MatGetRowMin_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2334: {
2335: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
2336: PetscInt m = A->rmap->n, n = A->cmap->n;
2337: PetscInt cstart = A->cmap->rstart, cend = A->cmap->rend;
2338: PetscInt *cmap = mat->garray;
2339: PetscInt *diagIdx, *offdiagIdx;
2340: Vec diagV, offdiagV;
2341: PetscScalar *a, *diagA, *offdiagA;
2342: const PetscScalar *ba, *bav;
2343: PetscInt r, j, col, ncols, *bi, *bj;
2344: Mat B = mat->B;
2345: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2347: PetscFunctionBegin;
2348: /* When a process holds entire A and other processes have no entry */
2349: if (A->cmap->N == n) {
2350: PetscCall(VecGetArrayWrite(v, &diagA));
2351: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2352: PetscCall(MatGetRowMin(mat->A, diagV, idx));
2353: PetscCall(VecDestroy(&diagV));
2354: PetscCall(VecRestoreArrayWrite(v, &diagA));
2355: PetscFunctionReturn(PETSC_SUCCESS);
2356: } else if (n == 0) {
2357: if (m) {
2358: PetscCall(VecGetArrayWrite(v, &a));
2359: for (r = 0; r < m; r++) {
2360: a[r] = PETSC_MAX_REAL;
2361: if (idx) idx[r] = -1;
2362: }
2363: PetscCall(VecRestoreArrayWrite(v, &a));
2364: }
2365: PetscFunctionReturn(PETSC_SUCCESS);
2366: }
2368: PetscCall(PetscCalloc2(m, &diagIdx, m, &offdiagIdx));
2369: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2370: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2371: PetscCall(MatGetRowMin(mat->A, diagV, diagIdx));
2373: /* Get offdiagIdx[] for implicit 0.0 */
2374: PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2375: ba = bav;
2376: bi = b->i;
2377: bj = b->j;
2378: PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2379: for (r = 0; r < m; r++) {
2380: ncols = bi[r + 1] - bi[r];
2381: if (ncols == A->cmap->N - n) { /* Brow is dense */
2382: offdiagA[r] = *ba;
2383: offdiagIdx[r] = cmap[0];
2384: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2385: offdiagA[r] = 0.0;
2387: /* Find first hole in the cmap */
2388: for (j = 0; j < ncols; j++) {
2389: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2390: if (col > j && j < cstart) {
2391: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2392: break;
2393: } else if (col > j + n && j >= cstart) {
2394: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2395: break;
2396: }
2397: }
2398: if (j == ncols && ncols < A->cmap->N - n) {
2399: /* a hole is outside compressed Bcols */
2400: if (ncols == 0) {
2401: if (cstart) {
2402: offdiagIdx[r] = 0;
2403: } else offdiagIdx[r] = cend;
2404: } else { /* ncols > 0 */
2405: offdiagIdx[r] = cmap[ncols - 1] + 1;
2406: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2407: }
2408: }
2409: }
2411: for (j = 0; j < ncols; j++) {
2412: if (PetscRealPart(offdiagA[r]) > PetscRealPart(*ba)) {
2413: offdiagA[r] = *ba;
2414: offdiagIdx[r] = cmap[*bj];
2415: }
2416: ba++;
2417: bj++;
2418: }
2419: }
2421: PetscCall(VecGetArrayWrite(v, &a));
2422: PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2423: for (r = 0; r < m; ++r) {
2424: if (PetscRealPart(diagA[r]) < PetscRealPart(offdiagA[r])) {
2425: a[r] = diagA[r];
2426: if (idx) idx[r] = cstart + diagIdx[r];
2427: } else if (PetscRealPart(diagA[r]) == PetscRealPart(offdiagA[r])) {
2428: a[r] = diagA[r];
2429: if (idx) {
2430: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2431: idx[r] = cstart + diagIdx[r];
2432: } else idx[r] = offdiagIdx[r];
2433: }
2434: } else {
2435: a[r] = offdiagA[r];
2436: if (idx) idx[r] = offdiagIdx[r];
2437: }
2438: }
2439: PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2440: PetscCall(VecRestoreArrayWrite(v, &a));
2441: PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2442: PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2443: PetscCall(VecDestroy(&diagV));
2444: PetscCall(VecDestroy(&offdiagV));
2445: PetscCall(PetscFree2(diagIdx, offdiagIdx));
2446: PetscFunctionReturn(PETSC_SUCCESS);
2447: }
2449: static PetscErrorCode MatGetRowMax_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2450: {
2451: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
2452: PetscInt m = A->rmap->n, n = A->cmap->n;
2453: PetscInt cstart = A->cmap->rstart, cend = A->cmap->rend;
2454: PetscInt *cmap = mat->garray;
2455: PetscInt *diagIdx, *offdiagIdx;
2456: Vec diagV, offdiagV;
2457: PetscScalar *a, *diagA, *offdiagA;
2458: const PetscScalar *ba, *bav;
2459: PetscInt r, j, col, ncols, *bi, *bj;
2460: Mat B = mat->B;
2461: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2463: PetscFunctionBegin;
2464: /* When a process holds entire A and other processes have no entry */
2465: if (A->cmap->N == n) {
2466: PetscCall(VecGetArrayWrite(v, &diagA));
2467: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2468: PetscCall(MatGetRowMax(mat->A, diagV, idx));
2469: PetscCall(VecDestroy(&diagV));
2470: PetscCall(VecRestoreArrayWrite(v, &diagA));
2471: PetscFunctionReturn(PETSC_SUCCESS);
2472: } else if (n == 0) {
2473: if (m) {
2474: PetscCall(VecGetArrayWrite(v, &a));
2475: for (r = 0; r < m; r++) {
2476: a[r] = PETSC_MIN_REAL;
2477: if (idx) idx[r] = -1;
2478: }
2479: PetscCall(VecRestoreArrayWrite(v, &a));
2480: }
2481: PetscFunctionReturn(PETSC_SUCCESS);
2482: }
2484: PetscCall(PetscMalloc2(m, &diagIdx, m, &offdiagIdx));
2485: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2486: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2487: PetscCall(MatGetRowMax(mat->A, diagV, diagIdx));
2489: /* Get offdiagIdx[] for implicit 0.0 */
2490: PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2491: ba = bav;
2492: bi = b->i;
2493: bj = b->j;
2494: PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2495: for (r = 0; r < m; r++) {
2496: ncols = bi[r + 1] - bi[r];
2497: if (ncols == A->cmap->N - n) { /* Brow is dense */
2498: offdiagA[r] = *ba;
2499: offdiagIdx[r] = cmap[0];
2500: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2501: offdiagA[r] = 0.0;
2503: /* Find first hole in the cmap */
2504: for (j = 0; j < ncols; j++) {
2505: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2506: if (col > j && j < cstart) {
2507: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2508: break;
2509: } else if (col > j + n && j >= cstart) {
2510: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2511: break;
2512: }
2513: }
2514: if (j == ncols && ncols < A->cmap->N - n) {
2515: /* a hole is outside compressed Bcols */
2516: if (ncols == 0) {
2517: if (cstart) {
2518: offdiagIdx[r] = 0;
2519: } else offdiagIdx[r] = cend;
2520: } else { /* ncols > 0 */
2521: offdiagIdx[r] = cmap[ncols - 1] + 1;
2522: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2523: }
2524: }
2525: }
2527: for (j = 0; j < ncols; j++) {
2528: if (PetscRealPart(offdiagA[r]) < PetscRealPart(*ba)) {
2529: offdiagA[r] = *ba;
2530: offdiagIdx[r] = cmap[*bj];
2531: }
2532: ba++;
2533: bj++;
2534: }
2535: }
2537: PetscCall(VecGetArrayWrite(v, &a));
2538: PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2539: for (r = 0; r < m; ++r) {
2540: if (PetscRealPart(diagA[r]) > PetscRealPart(offdiagA[r])) {
2541: a[r] = diagA[r];
2542: if (idx) idx[r] = cstart + diagIdx[r];
2543: } else if (PetscRealPart(diagA[r]) == PetscRealPart(offdiagA[r])) {
2544: a[r] = diagA[r];
2545: if (idx) {
2546: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2547: idx[r] = cstart + diagIdx[r];
2548: } else idx[r] = offdiagIdx[r];
2549: }
2550: } else {
2551: a[r] = offdiagA[r];
2552: if (idx) idx[r] = offdiagIdx[r];
2553: }
2554: }
2555: PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2556: PetscCall(VecRestoreArrayWrite(v, &a));
2557: PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2558: PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2559: PetscCall(VecDestroy(&diagV));
2560: PetscCall(VecDestroy(&offdiagV));
2561: PetscCall(PetscFree2(diagIdx, offdiagIdx));
2562: PetscFunctionReturn(PETSC_SUCCESS);
2563: }
2565: PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat mat, Mat *newmat)
2566: {
2567: Mat *dummy;
2569: PetscFunctionBegin;
2570: PetscCall(MatCreateSubMatrix_MPIAIJ_All(mat, MAT_DO_NOT_GET_VALUES, MAT_INITIAL_MATRIX, &dummy));
2571: *newmat = *dummy;
2572: PetscCall(PetscFree(dummy));
2573: PetscFunctionReturn(PETSC_SUCCESS);
2574: }
2576: static PetscErrorCode MatInvertBlockDiagonal_MPIAIJ(Mat A, const PetscScalar **values)
2577: {
2578: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2580: PetscFunctionBegin;
2581: PetscCall(MatInvertBlockDiagonal(a->A, values));
2582: A->factorerrortype = a->A->factorerrortype;
2583: PetscFunctionReturn(PETSC_SUCCESS);
2584: }
2586: static PetscErrorCode MatSetRandom_MPIAIJ(Mat x, PetscRandom rctx)
2587: {
2588: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)x->data;
2590: PetscFunctionBegin;
2591: PetscCheck(x->assembled || x->preallocated, PetscObjectComm((PetscObject)x), PETSC_ERR_ARG_WRONGSTATE, "MatSetRandom on an unassembled and unpreallocated MATMPIAIJ is not allowed");
2592: PetscCall(MatSetRandom(aij->A, rctx));
2593: if (x->assembled) {
2594: PetscCall(MatSetRandom(aij->B, rctx));
2595: } else {
2596: PetscCall(MatSetRandomSkipColumnRange_SeqAIJ_Private(aij->B, x->cmap->rstart, x->cmap->rend, rctx));
2597: }
2598: PetscCall(MatAssemblyBegin(x, MAT_FINAL_ASSEMBLY));
2599: PetscCall(MatAssemblyEnd(x, MAT_FINAL_ASSEMBLY));
2600: PetscFunctionReturn(PETSC_SUCCESS);
2601: }
2603: static PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ(Mat A, PetscBool sc)
2604: {
2605: PetscFunctionBegin;
2606: if (sc) A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ_Scalable;
2607: else A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ;
2608: PetscFunctionReturn(PETSC_SUCCESS);
2609: }
2611: /*@
2612: MatMPIAIJGetNumberNonzeros - gets the number of nonzeros in the matrix on this MPI rank
2614: Not Collective
2616: Input Parameter:
2617: . A - the matrix
2619: Output Parameter:
2620: . nz - the number of nonzeros
2622: Level: advanced
2624: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`
2625: @*/
2626: PetscErrorCode MatMPIAIJGetNumberNonzeros(Mat A, PetscCount *nz)
2627: {
2628: Mat_MPIAIJ *maij = (Mat_MPIAIJ *)A->data;
2629: Mat_SeqAIJ *aaij = (Mat_SeqAIJ *)maij->A->data, *baij = (Mat_SeqAIJ *)maij->B->data;
2630: PetscBool isaij;
2632: PetscFunctionBegin;
2633: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATMPIAIJ, &isaij));
2634: PetscCheck(isaij, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Not for type %s", ((PetscObject)A)->type_name);
2635: *nz = aaij->i[A->rmap->n] + baij->i[A->rmap->n];
2636: PetscFunctionReturn(PETSC_SUCCESS);
2637: }
2639: /*@
2640: MatMPIAIJSetUseScalableIncreaseOverlap - Determine if the matrix uses a scalable algorithm to compute the overlap
2642: Collective
2644: Input Parameters:
2645: + A - the matrix
2646: - sc - `PETSC_TRUE` indicates use the scalable algorithm (default is not to use the scalable algorithm)
2648: Level: advanced
2650: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`
2651: @*/
2652: PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap(Mat A, PetscBool sc)
2653: {
2654: PetscFunctionBegin;
2655: PetscTryMethod(A, "MatMPIAIJSetUseScalableIncreaseOverlap_C", (Mat, PetscBool), (A, sc));
2656: PetscFunctionReturn(PETSC_SUCCESS);
2657: }
2659: PetscErrorCode MatSetFromOptions_MPIAIJ(Mat A, PetscOptionItems PetscOptionsObject)
2660: {
2661: PetscBool sc = PETSC_FALSE, flg;
2663: PetscFunctionBegin;
2664: PetscOptionsHeadBegin(PetscOptionsObject, "MPIAIJ options");
2665: if (A->ops->increaseoverlap == MatIncreaseOverlap_MPIAIJ_Scalable) sc = PETSC_TRUE;
2666: PetscCall(PetscOptionsBool("-mat_increase_overlap_scalable", "Use a scalable algorithm to compute the overlap", "MatIncreaseOverlap", sc, &sc, &flg));
2667: if (flg) PetscCall(MatMPIAIJSetUseScalableIncreaseOverlap(A, sc));
2668: PetscOptionsHeadEnd();
2669: PetscFunctionReturn(PETSC_SUCCESS);
2670: }
2672: static PetscErrorCode MatShift_MPIAIJ(Mat Y, PetscScalar a)
2673: {
2674: Mat_MPIAIJ *maij = (Mat_MPIAIJ *)Y->data;
2675: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)maij->A->data;
2677: PetscFunctionBegin;
2678: if (!Y->preallocated) {
2679: PetscCall(MatMPIAIJSetPreallocation(Y, 1, NULL, 0, NULL));
2680: } else if (!aij->nz) { /* It does not matter if diagonals of Y only partially lie in maij->A. We just need an estimated preallocation. */
2681: PetscInt nonew = aij->nonew;
2682: PetscCall(MatSeqAIJSetPreallocation(maij->A, 1, NULL));
2683: aij->nonew = nonew;
2684: }
2685: PetscCall(MatShift_Basic(Y, a));
2686: PetscFunctionReturn(PETSC_SUCCESS);
2687: }
2689: static PetscErrorCode MatInvertVariableBlockDiagonal_MPIAIJ(Mat A, PetscInt nblocks, const PetscInt *bsizes, PetscScalar *diag)
2690: {
2691: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2693: PetscFunctionBegin;
2694: PetscCall(MatInvertVariableBlockDiagonal(a->A, nblocks, bsizes, diag));
2695: PetscFunctionReturn(PETSC_SUCCESS);
2696: }
2698: static PetscErrorCode MatEliminateZeros_MPIAIJ(Mat A, PetscBool keep)
2699: {
2700: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2702: PetscFunctionBegin;
2703: PetscCall(MatEliminateZeros_SeqAIJ(a->A, keep)); // possibly keep zero diagonal coefficients
2704: PetscCall(MatEliminateZeros_SeqAIJ(a->B, PETSC_FALSE)); // never keep zero diagonal coefficients
2705: PetscFunctionReturn(PETSC_SUCCESS);
2706: }
2708: static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ,
2709: MatGetRow_MPIAIJ,
2710: MatRestoreRow_MPIAIJ,
2711: MatMult_MPIAIJ,
2712: /* 4*/ MatMultAdd_MPIAIJ,
2713: MatMultTranspose_MPIAIJ,
2714: MatMultTransposeAdd_MPIAIJ,
2715: NULL,
2716: NULL,
2717: NULL,
2718: /*10*/ NULL,
2719: NULL,
2720: NULL,
2721: MatSOR_MPIAIJ,
2722: MatTranspose_MPIAIJ,
2723: /*15*/ MatGetInfo_MPIAIJ,
2724: MatEqual_MPIAIJ,
2725: MatGetDiagonal_MPIAIJ,
2726: MatDiagonalScale_MPIAIJ,
2727: MatNorm_MPIAIJ,
2728: /*20*/ MatAssemblyBegin_MPIAIJ,
2729: MatAssemblyEnd_MPIAIJ,
2730: MatSetOption_MPIAIJ,
2731: MatZeroEntries_MPIAIJ,
2732: /*24*/ MatZeroRows_MPIAIJ,
2733: NULL,
2734: NULL,
2735: NULL,
2736: NULL,
2737: /*29*/ MatSetUp_MPI_Hash,
2738: NULL,
2739: NULL,
2740: MatGetDiagonalBlock_MPIAIJ,
2741: NULL,
2742: /*34*/ MatDuplicate_MPIAIJ,
2743: NULL,
2744: NULL,
2745: NULL,
2746: NULL,
2747: /*39*/ MatAXPY_MPIAIJ,
2748: MatCreateSubMatrices_MPIAIJ,
2749: MatIncreaseOverlap_MPIAIJ,
2750: MatGetValues_MPIAIJ,
2751: MatCopy_MPIAIJ,
2752: /*44*/ MatGetRowMax_MPIAIJ,
2753: MatScale_MPIAIJ,
2754: MatShift_MPIAIJ,
2755: MatDiagonalSet_MPIAIJ,
2756: MatZeroRowsColumns_MPIAIJ,
2757: /*49*/ MatSetRandom_MPIAIJ,
2758: MatGetRowIJ_MPIAIJ,
2759: MatRestoreRowIJ_MPIAIJ,
2760: NULL,
2761: NULL,
2762: /*54*/ MatFDColoringCreate_MPIXAIJ,
2763: NULL,
2764: MatSetUnfactored_MPIAIJ,
2765: MatPermute_MPIAIJ,
2766: NULL,
2767: /*59*/ MatCreateSubMatrix_MPIAIJ,
2768: MatDestroy_MPIAIJ,
2769: MatView_MPIAIJ,
2770: NULL,
2771: NULL,
2772: /*64*/ MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ,
2773: NULL,
2774: NULL,
2775: NULL,
2776: MatGetRowMaxAbs_MPIAIJ,
2777: /*69*/ MatGetRowMinAbs_MPIAIJ,
2778: NULL,
2779: NULL,
2780: MatFDColoringApply_AIJ,
2781: MatSetFromOptions_MPIAIJ,
2782: MatFindZeroDiagonals_MPIAIJ,
2783: /*75*/ NULL,
2784: NULL,
2785: NULL,
2786: MatLoad_MPIAIJ,
2787: NULL,
2788: /*80*/ NULL,
2789: NULL,
2790: NULL,
2791: /*83*/ NULL,
2792: NULL,
2793: MatMatMultNumeric_MPIAIJ_MPIAIJ,
2794: MatPtAPNumeric_MPIAIJ_MPIAIJ,
2795: NULL,
2796: NULL,
2797: /*89*/ MatBindToCPU_MPIAIJ,
2798: MatProductSetFromOptions_MPIAIJ,
2799: NULL,
2800: NULL,
2801: MatConjugate_MPIAIJ,
2802: /*94*/ NULL,
2803: MatSetValuesRow_MPIAIJ,
2804: MatRealPart_MPIAIJ,
2805: MatImaginaryPart_MPIAIJ,
2806: NULL,
2807: /*99*/ NULL,
2808: NULL,
2809: NULL,
2810: MatGetRowMin_MPIAIJ,
2811: NULL,
2812: /*104*/ MatGetSeqNonzeroStructure_MPIAIJ,
2813: NULL,
2814: MatGetGhosts_MPIAIJ,
2815: NULL,
2816: NULL,
2817: /*109*/ MatMultDiagonalBlock_MPIAIJ,
2818: NULL,
2819: NULL,
2820: NULL,
2821: MatGetMultiProcBlock_MPIAIJ,
2822: /*114*/ MatFindNonzeroRows_MPIAIJ,
2823: MatGetColumnReductions_MPIAIJ,
2824: MatInvertBlockDiagonal_MPIAIJ,
2825: MatInvertVariableBlockDiagonal_MPIAIJ,
2826: MatCreateSubMatricesMPI_MPIAIJ,
2827: /*119*/ NULL,
2828: NULL,
2829: MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ,
2830: NULL,
2831: NULL,
2832: /*124*/ NULL,
2833: NULL,
2834: MatSetBlockSizes_MPIAIJ,
2835: NULL,
2836: MatFDColoringSetUp_MPIXAIJ,
2837: /*129*/ MatFindOffBlockDiagonalEntries_MPIAIJ,
2838: MatCreateMPIMatConcatenateSeqMat_MPIAIJ,
2839: NULL,
2840: NULL,
2841: NULL,
2842: /*134*/ MatCreateGraph_Simple_AIJ,
2843: NULL,
2844: MatEliminateZeros_MPIAIJ,
2845: MatGetRowSumAbs_MPIAIJ,
2846: NULL,
2847: /*139*/ NULL,
2848: NULL,
2849: MatCopyHashToXAIJ_MPI_Hash,
2850: MatGetCurrentMemType_MPIAIJ,
2851: NULL};
2853: static PetscErrorCode MatStoreValues_MPIAIJ(Mat mat)
2854: {
2855: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2857: PetscFunctionBegin;
2858: PetscCall(MatStoreValues(aij->A));
2859: PetscCall(MatStoreValues(aij->B));
2860: PetscFunctionReturn(PETSC_SUCCESS);
2861: }
2863: static PetscErrorCode MatRetrieveValues_MPIAIJ(Mat mat)
2864: {
2865: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2867: PetscFunctionBegin;
2868: PetscCall(MatRetrieveValues(aij->A));
2869: PetscCall(MatRetrieveValues(aij->B));
2870: PetscFunctionReturn(PETSC_SUCCESS);
2871: }
2873: PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat B, PetscInt d_nz, const PetscInt d_nnz[], PetscInt o_nz, const PetscInt o_nnz[])
2874: {
2875: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2876: PetscMPIInt size;
2878: PetscFunctionBegin;
2879: if (B->hash_active) {
2880: B->ops[0] = b->cops;
2881: B->hash_active = PETSC_FALSE;
2882: }
2883: PetscCall(PetscLayoutSetUp(B->rmap));
2884: PetscCall(PetscLayoutSetUp(B->cmap));
2886: #if defined(PETSC_USE_CTABLE)
2887: PetscCall(PetscHMapIDestroy(&b->colmap));
2888: #else
2889: PetscCall(PetscFree(b->colmap));
2890: #endif
2891: PetscCall(PetscFree(b->garray));
2892: PetscCall(VecDestroy(&b->lvec));
2893: PetscCall(VecScatterDestroy(&b->Mvctx));
2895: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
2897: MatSeqXAIJGetOptions_Private(b->B);
2898: PetscCall(MatDestroy(&b->B));
2899: PetscCall(MatCreate(PETSC_COMM_SELF, &b->B));
2900: PetscCall(MatSetSizes(b->B, B->rmap->n, size > 1 ? B->cmap->N : 0, B->rmap->n, size > 1 ? B->cmap->N : 0));
2901: PetscCall(MatSetBlockSizesFromMats(b->B, B, B));
2902: PetscCall(MatSetType(b->B, MATSEQAIJ));
2903: MatSeqXAIJRestoreOptions_Private(b->B);
2905: MatSeqXAIJGetOptions_Private(b->A);
2906: PetscCall(MatDestroy(&b->A));
2907: PetscCall(MatCreate(PETSC_COMM_SELF, &b->A));
2908: PetscCall(MatSetSizes(b->A, B->rmap->n, B->cmap->n, B->rmap->n, B->cmap->n));
2909: PetscCall(MatSetBlockSizesFromMats(b->A, B, B));
2910: PetscCall(MatSetType(b->A, MATSEQAIJ));
2911: MatSeqXAIJRestoreOptions_Private(b->A);
2913: PetscCall(MatSeqAIJSetPreallocation(b->A, d_nz, d_nnz));
2914: PetscCall(MatSeqAIJSetPreallocation(b->B, o_nz, o_nnz));
2915: B->preallocated = PETSC_TRUE;
2916: B->was_assembled = PETSC_FALSE;
2917: B->assembled = PETSC_FALSE;
2918: PetscFunctionReturn(PETSC_SUCCESS);
2919: }
2921: static PetscErrorCode MatResetPreallocation_MPIAIJ(Mat B)
2922: {
2923: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2924: PetscBool ondiagreset, offdiagreset, memoryreset;
2926: PetscFunctionBegin;
2928: 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()");
2929: if (B->num_ass == 0) PetscFunctionReturn(PETSC_SUCCESS);
2931: PetscCall(MatResetPreallocation_SeqAIJ_Private(b->A, &ondiagreset));
2932: PetscCall(MatResetPreallocation_SeqAIJ_Private(b->B, &offdiagreset));
2933: memoryreset = (PetscBool)(ondiagreset || offdiagreset);
2934: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &memoryreset, 1, MPI_C_BOOL, MPI_LOR, PetscObjectComm((PetscObject)B)));
2935: if (!memoryreset) PetscFunctionReturn(PETSC_SUCCESS);
2937: PetscCall(PetscLayoutSetUp(B->rmap));
2938: PetscCall(PetscLayoutSetUp(B->cmap));
2939: 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");
2940: PetscCall(MatDisAssemble_MPIAIJ(B, PETSC_TRUE));
2941: PetscCall(VecScatterDestroy(&b->Mvctx));
2943: B->preallocated = PETSC_TRUE;
2944: B->was_assembled = PETSC_FALSE;
2945: B->assembled = PETSC_FALSE;
2946: /* Log that the state of this object has changed; this will help guarantee that preconditioners get re-setup */
2947: PetscCall(PetscObjectStateIncrease((PetscObject)B));
2948: PetscFunctionReturn(PETSC_SUCCESS);
2949: }
2951: PetscErrorCode MatDuplicate_MPIAIJ(Mat matin, MatDuplicateOption cpvalues, Mat *newmat)
2952: {
2953: Mat mat;
2954: Mat_MPIAIJ *a, *oldmat = (Mat_MPIAIJ *)matin->data;
2956: PetscFunctionBegin;
2957: *newmat = NULL;
2958: PetscCall(MatCreate(PetscObjectComm((PetscObject)matin), &mat));
2959: PetscCall(MatSetSizes(mat, matin->rmap->n, matin->cmap->n, matin->rmap->N, matin->cmap->N));
2960: PetscCall(MatSetBlockSizesFromMats(mat, matin, matin));
2961: PetscCall(MatSetType(mat, ((PetscObject)matin)->type_name));
2962: a = (Mat_MPIAIJ *)mat->data;
2964: mat->factortype = matin->factortype;
2965: mat->assembled = matin->assembled;
2966: mat->insertmode = NOT_SET_VALUES;
2968: a->size = oldmat->size;
2969: a->rank = oldmat->rank;
2970: a->donotstash = oldmat->donotstash;
2971: a->roworiented = oldmat->roworiented;
2972: a->rowindices = NULL;
2973: a->rowvalues = NULL;
2974: a->getrowactive = PETSC_FALSE;
2976: PetscCall(PetscLayoutReference(matin->rmap, &mat->rmap));
2977: PetscCall(PetscLayoutReference(matin->cmap, &mat->cmap));
2978: if (matin->hash_active) {
2979: PetscCall(MatSetUp(mat));
2980: } else {
2981: mat->preallocated = matin->preallocated;
2982: if (oldmat->colmap) {
2983: #if defined(PETSC_USE_CTABLE)
2984: PetscCall(PetscHMapIDuplicate(oldmat->colmap, &a->colmap));
2985: #else
2986: PetscCall(PetscMalloc1(mat->cmap->N, &a->colmap));
2987: PetscCall(PetscArraycpy(a->colmap, oldmat->colmap, mat->cmap->N));
2988: #endif
2989: } else a->colmap = NULL;
2990: if (oldmat->garray) {
2991: PetscInt len;
2992: len = oldmat->B->cmap->n;
2993: PetscCall(PetscMalloc1(len, &a->garray));
2994: if (len) PetscCall(PetscArraycpy(a->garray, oldmat->garray, len));
2995: } else a->garray = NULL;
2997: /* It may happen MatDuplicate is called with a non-assembled matrix
2998: In fact, MatDuplicate only requires the matrix to be preallocated
2999: This may happen inside a DMCreateMatrix_Shell */
3000: if (oldmat->lvec) PetscCall(VecDuplicate(oldmat->lvec, &a->lvec));
3001: if (oldmat->Mvctx) {
3002: a->Mvctx = oldmat->Mvctx;
3003: PetscCall(PetscObjectReference((PetscObject)oldmat->Mvctx));
3004: }
3005: PetscCall(MatDuplicate(oldmat->A, cpvalues, &a->A));
3006: PetscCall(MatDuplicate(oldmat->B, cpvalues, &a->B));
3007: }
3008: PetscCall(PetscFunctionListDuplicate(((PetscObject)matin)->qlist, &((PetscObject)mat)->qlist));
3009: *newmat = mat;
3010: PetscFunctionReturn(PETSC_SUCCESS);
3011: }
3013: PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer)
3014: {
3015: PetscBool isbinary, ishdf5;
3017: PetscFunctionBegin;
3020: /* force binary viewer to load .info file if it has not yet done so */
3021: PetscCall(PetscViewerSetUp(viewer));
3022: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
3023: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERHDF5, &ishdf5));
3024: if (isbinary) {
3025: PetscCall(MatLoad_MPIAIJ_Binary(newMat, viewer));
3026: } else if (ishdf5) {
3027: #if defined(PETSC_HAVE_HDF5)
3028: PetscCall(MatLoad_AIJ_HDF5(newMat, viewer));
3029: #else
3030: SETERRQ(PetscObjectComm((PetscObject)newMat), PETSC_ERR_SUP, "HDF5 not supported in this build.\nPlease reconfigure using --download-hdf5");
3031: #endif
3032: } else {
3033: 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);
3034: }
3035: PetscFunctionReturn(PETSC_SUCCESS);
3036: }
3038: PetscErrorCode MatLoad_MPIAIJ_Binary(Mat mat, PetscViewer viewer)
3039: {
3040: PetscInt header[4], M, N, m, nz, rows, cols, sum, i;
3041: PetscInt *rowidxs, *colidxs;
3042: PetscScalar *matvals;
3044: PetscFunctionBegin;
3045: PetscCall(PetscViewerSetUp(viewer));
3047: /* read in matrix header */
3048: PetscCall(PetscViewerBinaryRead(viewer, header, 4, NULL, PETSC_INT));
3049: PetscCheck(header[0] == MAT_FILE_CLASSID, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Not a matrix object in file");
3050: M = header[1];
3051: N = header[2];
3052: nz = header[3];
3053: PetscCheck(M >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix row size (%" PetscInt_FMT ") in file is negative", M);
3054: PetscCheck(N >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix column size (%" PetscInt_FMT ") in file is negative", N);
3055: PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Matrix stored in special format on disk, cannot load as MPIAIJ");
3057: /* set block sizes from the viewer's .info file */
3058: PetscCall(MatLoad_Binary_BlockSizes(mat, viewer));
3059: /* set global sizes if not set already */
3060: if (mat->rmap->N < 0) mat->rmap->N = M;
3061: if (mat->cmap->N < 0) mat->cmap->N = N;
3062: PetscCall(PetscLayoutSetUp(mat->rmap));
3063: PetscCall(PetscLayoutSetUp(mat->cmap));
3065: /* check if the matrix sizes are correct */
3066: PetscCall(MatGetSize(mat, &rows, &cols));
3067: 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);
3069: /* read in row lengths and build row indices */
3070: PetscCall(MatGetLocalSize(mat, &m, NULL));
3071: PetscCall(PetscMalloc1(m + 1, &rowidxs));
3072: PetscCall(PetscViewerBinaryReadAll(viewer, rowidxs + 1, m, PETSC_DECIDE, M, PETSC_INT));
3073: rowidxs[0] = 0;
3074: for (i = 0; i < m; i++) rowidxs[i + 1] += rowidxs[i];
3075: if (nz != PETSC_INT_MAX) {
3076: PetscCallMPI(MPIU_Allreduce(&rowidxs[m], &sum, 1, MPIU_INT, MPI_SUM, PetscObjectComm((PetscObject)viewer)));
3077: 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);
3078: }
3080: /* read in column indices and matrix values */
3081: PetscCall(PetscMalloc2(rowidxs[m], &colidxs, rowidxs[m], &matvals));
3082: PetscCall(PetscViewerBinaryReadAll(viewer, colidxs, rowidxs[m], PETSC_DETERMINE, PETSC_DETERMINE, PETSC_INT));
3083: PetscCall(PetscViewerBinaryReadAll(viewer, matvals, rowidxs[m], PETSC_DETERMINE, PETSC_DETERMINE, PETSC_SCALAR));
3084: /* store matrix indices and values */
3085: PetscCall(MatMPIAIJSetPreallocationCSR(mat, rowidxs, colidxs, matvals));
3086: PetscCall(PetscFree(rowidxs));
3087: PetscCall(PetscFree2(colidxs, matvals));
3088: PetscFunctionReturn(PETSC_SUCCESS);
3089: }
3091: /* Not scalable because of ISAllGather() unless getting all columns. */
3092: static PetscErrorCode ISGetSeqIS_Private(Mat mat, IS iscol, IS *isseq)
3093: {
3094: IS iscol_local;
3095: PetscBool isstride;
3096: PetscMPIInt gisstride = 0;
3098: PetscFunctionBegin;
3099: /* check if we are grabbing all columns*/
3100: PetscCall(PetscObjectTypeCompare((PetscObject)iscol, ISSTRIDE, &isstride));
3102: if (isstride) {
3103: PetscInt start, len, mstart, mlen;
3104: PetscCall(ISStrideGetInfo(iscol, &start, NULL));
3105: PetscCall(ISGetLocalSize(iscol, &len));
3106: PetscCall(MatGetOwnershipRangeColumn(mat, &mstart, &mlen));
3107: if (mstart == start && mlen - mstart == len) gisstride = 1;
3108: }
3110: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &gisstride, 1, MPI_INT, MPI_MIN, PetscObjectComm((PetscObject)mat)));
3111: if (gisstride) {
3112: PetscInt N;
3113: PetscCall(MatGetSize(mat, NULL, &N));
3114: PetscCall(ISCreateStride(PETSC_COMM_SELF, N, 0, 1, &iscol_local));
3115: PetscCall(ISSetIdentity(iscol_local));
3116: PetscCall(PetscInfo(mat, "Optimizing for obtaining all columns of the matrix; skipping ISAllGather()\n"));
3117: } else {
3118: PetscInt cbs;
3119: PetscCall(ISGetBlockSize(iscol, &cbs));
3120: PetscCall(ISAllGather(iscol, &iscol_local));
3121: PetscCall(ISSetBlockSize(iscol_local, cbs));
3122: }
3124: *isseq = iscol_local;
3125: PetscFunctionReturn(PETSC_SUCCESS);
3126: }
3128: /*
3129: Used by MatCreateSubMatrix_MPIAIJ_SameRowColDist() to avoid ISAllGather() and global size of iscol_local
3130: (see MatCreateSubMatrix_MPIAIJ_nonscalable)
3132: Input Parameters:
3133: + mat - matrix
3134: . isrow - parallel row index set; its local indices are a subset of local columns of `mat`,
3135: i.e., mat->rstart <= isrow[i] < mat->rend
3136: - iscol - parallel column index set; its local indices are a subset of local columns of `mat`,
3137: i.e., mat->cstart <= iscol[i] < mat->cend
3139: Output Parameters:
3140: + isrow_d - sequential row index set for retrieving mat->A
3141: . iscol_d - sequential column index set for retrieving mat->A
3142: . iscol_o - sequential column index set for retrieving mat->B
3143: - garray - column map; garray[i] indicates global location of iscol_o[i] in `iscol`
3144: */
3145: static PetscErrorCode ISGetSeqIS_SameColDist_Private(Mat mat, IS isrow, IS iscol, IS *isrow_d, IS *iscol_d, IS *iscol_o, PetscInt *garray[])
3146: {
3147: Vec x, cmap;
3148: const PetscInt *is_idx;
3149: PetscScalar *xarray, *cmaparray;
3150: PetscInt ncols, isstart, *idx, m, rstart, *cmap1, count;
3151: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data;
3152: Mat B = a->B;
3153: Vec lvec = a->lvec, lcmap;
3154: PetscInt i, cstart, cend, Bn = B->cmap->N;
3155: MPI_Comm comm;
3156: VecScatter Mvctx = a->Mvctx;
3158: PetscFunctionBegin;
3159: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3160: PetscCall(ISGetLocalSize(iscol, &ncols));
3162: /* (1) iscol is a sub-column vector of mat, pad it with '-1.' to form a full vector x */
3163: PetscCall(MatCreateVecs(mat, &x, NULL));
3164: PetscCall(VecSet(x, -1.0));
3165: PetscCall(VecDuplicate(x, &cmap));
3166: PetscCall(VecSet(cmap, -1.0));
3168: /* Get start indices */
3169: PetscCallMPI(MPI_Scan(&ncols, &isstart, 1, MPIU_INT, MPI_SUM, comm));
3170: isstart -= ncols;
3171: PetscCall(MatGetOwnershipRangeColumn(mat, &cstart, &cend));
3173: PetscCall(ISGetIndices(iscol, &is_idx));
3174: PetscCall(VecGetArray(x, &xarray));
3175: PetscCall(VecGetArray(cmap, &cmaparray));
3176: PetscCall(PetscMalloc1(ncols, &idx));
3177: for (i = 0; i < ncols; i++) {
3178: xarray[is_idx[i] - cstart] = (PetscScalar)is_idx[i];
3179: cmaparray[is_idx[i] - cstart] = i + isstart; /* global index of iscol[i] */
3180: idx[i] = is_idx[i] - cstart; /* local index of iscol[i] */
3181: }
3182: PetscCall(VecRestoreArray(x, &xarray));
3183: PetscCall(VecRestoreArray(cmap, &cmaparray));
3184: PetscCall(ISRestoreIndices(iscol, &is_idx));
3186: /* Get iscol_d */
3187: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, iscol_d));
3188: PetscCall(ISGetBlockSize(iscol, &i));
3189: PetscCall(ISSetBlockSize(*iscol_d, i));
3191: /* Get isrow_d */
3192: PetscCall(ISGetLocalSize(isrow, &m));
3193: rstart = mat->rmap->rstart;
3194: PetscCall(PetscMalloc1(m, &idx));
3195: PetscCall(ISGetIndices(isrow, &is_idx));
3196: for (i = 0; i < m; i++) idx[i] = is_idx[i] - rstart;
3197: PetscCall(ISRestoreIndices(isrow, &is_idx));
3199: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, m, idx, PETSC_OWN_POINTER, isrow_d));
3200: PetscCall(ISGetBlockSize(isrow, &i));
3201: PetscCall(ISSetBlockSize(*isrow_d, i));
3203: /* (2) Scatter x and cmap using aij->Mvctx to get their off-process portions (see MatMult_MPIAIJ) */
3204: PetscCall(VecScatterBegin(Mvctx, x, lvec, INSERT_VALUES, SCATTER_FORWARD));
3205: PetscCall(VecScatterEnd(Mvctx, x, lvec, INSERT_VALUES, SCATTER_FORWARD));
3207: PetscCall(VecDuplicate(lvec, &lcmap));
3209: PetscCall(VecScatterBegin(Mvctx, cmap, lcmap, INSERT_VALUES, SCATTER_FORWARD));
3210: PetscCall(VecScatterEnd(Mvctx, cmap, lcmap, INSERT_VALUES, SCATTER_FORWARD));
3212: /* (3) create sequential iscol_o (a subset of iscol) and isgarray */
3213: /* off-process column indices */
3214: count = 0;
3215: PetscCall(PetscMalloc1(Bn, &idx));
3216: PetscCall(PetscMalloc1(Bn, &cmap1));
3218: PetscCall(VecGetArray(lvec, &xarray));
3219: PetscCall(VecGetArray(lcmap, &cmaparray));
3220: for (i = 0; i < Bn; i++) {
3221: if (PetscRealPart(xarray[i]) > -1.0) {
3222: idx[count] = i; /* local column index in off-diagonal part B */
3223: cmap1[count] = (PetscInt)PetscRealPart(cmaparray[i]); /* column index in submat */
3224: count++;
3225: }
3226: }
3227: PetscCall(VecRestoreArray(lvec, &xarray));
3228: PetscCall(VecRestoreArray(lcmap, &cmaparray));
3230: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, count, idx, PETSC_COPY_VALUES, iscol_o));
3231: /* cannot ensure iscol_o has same blocksize as iscol! */
3233: PetscCall(PetscFree(idx));
3234: *garray = cmap1;
3236: PetscCall(VecDestroy(&x));
3237: PetscCall(VecDestroy(&cmap));
3238: PetscCall(VecDestroy(&lcmap));
3239: PetscFunctionReturn(PETSC_SUCCESS);
3240: }
3242: /* isrow and iscol have same processor distribution as mat, output *submat is a submatrix of local mat */
3243: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowColDist(Mat mat, IS isrow, IS iscol, MatReuse call, Mat *submat)
3244: {
3245: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data, *asub;
3246: Mat M = NULL;
3247: MPI_Comm comm;
3248: IS iscol_d, isrow_d, iscol_o;
3249: Mat Asub = NULL, Bsub = NULL;
3250: PetscInt n, count, M_size, N_size;
3252: PetscFunctionBegin;
3253: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3255: if (call == MAT_REUSE_MATRIX) {
3256: /* Retrieve isrow_d, iscol_d and iscol_o from submat */
3257: PetscCall(PetscObjectQuery((PetscObject)*submat, "isrow_d", (PetscObject *)&isrow_d));
3258: PetscCheck(isrow_d, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "isrow_d passed in was not used before, cannot reuse");
3260: PetscCall(PetscObjectQuery((PetscObject)*submat, "iscol_d", (PetscObject *)&iscol_d));
3261: PetscCheck(iscol_d, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "iscol_d passed in was not used before, cannot reuse");
3263: PetscCall(PetscObjectQuery((PetscObject)*submat, "iscol_o", (PetscObject *)&iscol_o));
3264: PetscCheck(iscol_o, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "iscol_o passed in was not used before, cannot reuse");
3266: /* Update diagonal and off-diagonal portions of submat */
3267: asub = (Mat_MPIAIJ *)(*submat)->data;
3268: PetscCall(MatCreateSubMatrix_SeqAIJ(a->A, isrow_d, iscol_d, PETSC_DECIDE, MAT_REUSE_MATRIX, &asub->A));
3269: PetscCall(ISGetLocalSize(iscol_o, &n));
3270: if (n) PetscCall(MatCreateSubMatrix_SeqAIJ(a->B, isrow_d, iscol_o, PETSC_DECIDE, MAT_REUSE_MATRIX, &asub->B));
3271: PetscCall(MatAssemblyBegin(*submat, MAT_FINAL_ASSEMBLY));
3272: PetscCall(MatAssemblyEnd(*submat, MAT_FINAL_ASSEMBLY));
3274: } else { /* call == MAT_INITIAL_MATRIX) */
3275: PetscInt *garray, *garray_compact;
3276: PetscInt BsubN;
3278: /* Create isrow_d, iscol_d, iscol_o and isgarray (replace isgarray with array?) */
3279: PetscCall(ISGetSeqIS_SameColDist_Private(mat, isrow, iscol, &isrow_d, &iscol_d, &iscol_o, &garray));
3281: /* Create local submatrices Asub and Bsub */
3282: PetscCall(MatCreateSubMatrix_SeqAIJ(a->A, isrow_d, iscol_d, PETSC_DECIDE, MAT_INITIAL_MATRIX, &Asub));
3283: PetscCall(MatCreateSubMatrix_SeqAIJ(a->B, isrow_d, iscol_o, PETSC_DECIDE, MAT_INITIAL_MATRIX, &Bsub));
3285: // Compact garray so its not of size Bn
3286: PetscCall(ISGetSize(iscol_o, &count));
3287: PetscCall(PetscMalloc1(count, &garray_compact));
3288: PetscCall(PetscArraycpy(garray_compact, garray, count));
3290: /* Create submatrix M */
3291: PetscCall(ISGetSize(isrow, &M_size));
3292: PetscCall(ISGetSize(iscol, &N_size));
3293: PetscCall(MatCreateMPIAIJWithSeqAIJ(comm, M_size, N_size, Asub, Bsub, garray_compact, &M));
3295: /* If Bsub has empty columns, compress iscol_o such that it will retrieve condensed Bsub from a->B during reuse */
3296: asub = (Mat_MPIAIJ *)M->data;
3298: PetscCall(ISGetLocalSize(iscol_o, &BsubN));
3299: n = asub->B->cmap->N;
3300: if (BsubN > n) {
3301: /* This case can be tested using ~petsc/src/tao/bound/tutorials/runplate2_3 */
3302: const PetscInt *idx;
3303: PetscInt i, j, *idx_new, *subgarray = asub->garray;
3304: PetscCall(PetscInfo(M, "submatrix Bn %" PetscInt_FMT " != BsubN %" PetscInt_FMT ", update iscol_o\n", n, BsubN));
3306: PetscCall(PetscMalloc1(n, &idx_new));
3307: j = 0;
3308: PetscCall(ISGetIndices(iscol_o, &idx));
3309: for (i = 0; i < n; i++) {
3310: if (j >= BsubN) break;
3311: while (subgarray[i] > garray[j]) j++;
3313: 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]);
3314: idx_new[i] = idx[j++];
3315: }
3316: PetscCall(ISRestoreIndices(iscol_o, &idx));
3318: PetscCall(ISDestroy(&iscol_o));
3319: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, n, idx_new, PETSC_OWN_POINTER, &iscol_o));
3321: } 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);
3323: PetscCall(PetscFree(garray));
3324: *submat = M;
3326: /* Save isrow_d, iscol_d and iscol_o used in processor for next request */
3327: PetscCall(PetscObjectCompose((PetscObject)M, "isrow_d", (PetscObject)isrow_d));
3328: PetscCall(ISDestroy(&isrow_d));
3330: PetscCall(PetscObjectCompose((PetscObject)M, "iscol_d", (PetscObject)iscol_d));
3331: PetscCall(ISDestroy(&iscol_d));
3333: PetscCall(PetscObjectCompose((PetscObject)M, "iscol_o", (PetscObject)iscol_o));
3334: PetscCall(ISDestroy(&iscol_o));
3335: }
3336: PetscFunctionReturn(PETSC_SUCCESS);
3337: }
3339: PetscErrorCode MatCreateSubMatrix_MPIAIJ(Mat mat, IS isrow, IS iscol, MatReuse call, Mat *newmat)
3340: {
3341: IS iscol_local = NULL, isrow_d;
3342: PetscInt csize;
3343: PetscInt n, i, j, start, end;
3344: PetscBool sameRowDist = PETSC_FALSE, sameDist[2], tsameDist[2];
3345: MPI_Comm comm;
3347: PetscFunctionBegin;
3348: /* If isrow has same processor distribution as mat,
3349: call MatCreateSubMatrix_MPIAIJ_SameRowDist() to avoid using a hash table with global size of iscol */
3350: if (call == MAT_REUSE_MATRIX) {
3351: PetscCall(PetscObjectQuery((PetscObject)*newmat, "isrow_d", (PetscObject *)&isrow_d));
3352: if (isrow_d) {
3353: sameRowDist = PETSC_TRUE;
3354: tsameDist[1] = PETSC_TRUE; /* sameColDist */
3355: } else {
3356: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_local));
3357: if (iscol_local) {
3358: sameRowDist = PETSC_TRUE;
3359: tsameDist[1] = PETSC_FALSE; /* !sameColDist */
3360: }
3361: }
3362: } else {
3363: /* Check if isrow has same processor distribution as mat */
3364: sameDist[0] = PETSC_FALSE;
3365: PetscCall(ISGetLocalSize(isrow, &n));
3366: if (!n) {
3367: sameDist[0] = PETSC_TRUE;
3368: } else {
3369: PetscCall(ISGetMinMax(isrow, &i, &j));
3370: PetscCall(MatGetOwnershipRange(mat, &start, &end));
3371: if (i >= start && j < end) sameDist[0] = PETSC_TRUE;
3372: }
3374: /* Check if iscol has same processor distribution as mat */
3375: sameDist[1] = PETSC_FALSE;
3376: PetscCall(ISGetLocalSize(iscol, &n));
3377: if (!n) {
3378: sameDist[1] = PETSC_TRUE;
3379: } else {
3380: PetscCall(ISGetMinMax(iscol, &i, &j));
3381: PetscCall(MatGetOwnershipRangeColumn(mat, &start, &end));
3382: if (i >= start && j < end) sameDist[1] = PETSC_TRUE;
3383: }
3385: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3386: PetscCallMPI(MPIU_Allreduce(&sameDist, &tsameDist, 2, MPI_C_BOOL, MPI_LAND, comm));
3387: sameRowDist = tsameDist[0];
3388: }
3390: if (sameRowDist) {
3391: if (tsameDist[1]) { /* sameRowDist & sameColDist */
3392: /* isrow and iscol have same processor distribution as mat */
3393: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowColDist(mat, isrow, iscol, call, newmat));
3394: PetscFunctionReturn(PETSC_SUCCESS);
3395: } else { /* sameRowDist */
3396: /* isrow has same processor distribution as mat */
3397: if (call == MAT_INITIAL_MATRIX) {
3398: PetscBool sorted;
3399: PetscCall(ISGetSeqIS_Private(mat, iscol, &iscol_local));
3400: PetscCall(ISGetLocalSize(iscol_local, &n)); /* local size of iscol_local = global columns of newmat */
3401: PetscCall(ISGetSize(iscol, &i));
3402: PetscCheck(n == i, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "n %" PetscInt_FMT " != size of iscol %" PetscInt_FMT, n, i);
3404: PetscCall(ISSorted(iscol_local, &sorted));
3405: if (sorted) {
3406: /* MatCreateSubMatrix_MPIAIJ_SameRowDist() requires iscol_local be sorted; it can have duplicate indices */
3407: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowDist(mat, isrow, iscol, iscol_local, MAT_INITIAL_MATRIX, newmat));
3408: PetscFunctionReturn(PETSC_SUCCESS);
3409: }
3410: } else { /* call == MAT_REUSE_MATRIX */
3411: IS iscol_sub;
3412: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_sub));
3413: if (iscol_sub) {
3414: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowDist(mat, isrow, iscol, NULL, call, newmat));
3415: PetscFunctionReturn(PETSC_SUCCESS);
3416: }
3417: }
3418: }
3419: }
3421: /* General case: iscol -> iscol_local which has global size of iscol */
3422: if (call == MAT_REUSE_MATRIX) {
3423: PetscCall(PetscObjectQuery((PetscObject)*newmat, "ISAllGather", (PetscObject *)&iscol_local));
3424: PetscCheck(iscol_local, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3425: } else {
3426: if (!iscol_local) PetscCall(ISGetSeqIS_Private(mat, iscol, &iscol_local));
3427: }
3429: PetscCall(ISGetLocalSize(iscol, &csize));
3430: PetscCall(MatCreateSubMatrix_MPIAIJ_nonscalable(mat, isrow, iscol_local, csize, call, newmat));
3432: if (call == MAT_INITIAL_MATRIX) {
3433: PetscCall(PetscObjectCompose((PetscObject)*newmat, "ISAllGather", (PetscObject)iscol_local));
3434: PetscCall(ISDestroy(&iscol_local));
3435: }
3436: PetscFunctionReturn(PETSC_SUCCESS);
3437: }
3439: /*@C
3440: MatCreateMPIAIJWithSeqAIJ - creates a `MATMPIAIJ` matrix using `MATSEQAIJ` matrices that contain the "diagonal"
3441: and "off-diagonal" part of the matrix in CSR format.
3443: Collective
3445: Input Parameters:
3446: + comm - MPI communicator
3447: . M - the global row size
3448: . N - the global column size
3449: . A - "diagonal" portion of matrix
3450: . 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
3451: - 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.
3453: Output Parameter:
3454: . mat - the matrix, with input `A` as its local diagonal matrix
3456: Level: advanced
3458: Notes:
3459: See `MatCreateAIJ()` for the definition of "diagonal" and "off-diagonal" portion of the matrix.
3461: `A` and `B` becomes part of output mat. The user cannot use `A` and `B` anymore.
3463: 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
3464: `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
3465: `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`
3466: yourself, see algorithms in the private function `MatSetUpMultiply_MPIAIJ()`.
3468: 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.
3470: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MATSEQAIJ`, `MatCreateMPIAIJWithSplitArrays()`
3471: @*/
3472: PetscErrorCode MatCreateMPIAIJWithSeqAIJ(MPI_Comm comm, PetscInt M, PetscInt N, Mat A, Mat B, PetscInt *garray, Mat *mat)
3473: {
3474: PetscInt m, n;
3475: MatType mpi_mat_type;
3476: Mat_MPIAIJ *mpiaij;
3477: Mat C;
3479: PetscFunctionBegin;
3480: PetscCall(MatCreate(comm, &C));
3481: PetscCall(MatGetSize(A, &m, &n));
3482: PetscCheck(m == B->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Am %" PetscInt_FMT " != Bm %" PetscInt_FMT, m, B->rmap->N);
3483: 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);
3485: PetscCall(MatSetSizes(C, m, n, M, N));
3486: /* Determine the type of MPI matrix that should be created from the type of matrix A, which holds the "diagonal" portion. */
3487: PetscCall(MatGetMPIMatType_Private(A, &mpi_mat_type));
3488: PetscCall(MatSetType(C, mpi_mat_type));
3489: if (!garray) {
3490: const PetscScalar *ba;
3492: B->nonzerostate++;
3493: PetscCall(MatSeqAIJGetArrayRead(B, &ba)); /* Since we will destroy B's device copy, we need to make sure the host copy is up to date */
3494: PetscCall(MatSeqAIJRestoreArrayRead(B, &ba));
3495: }
3497: PetscCall(MatSetBlockSizes(C, A->rmap->bs, A->cmap->bs));
3498: PetscCall(PetscLayoutSetUp(C->rmap));
3499: PetscCall(PetscLayoutSetUp(C->cmap));
3501: mpiaij = (Mat_MPIAIJ *)C->data;
3502: mpiaij->A = A;
3503: mpiaij->B = B;
3504: mpiaij->garray = garray;
3505: C->preallocated = PETSC_TRUE;
3506: C->nooffprocentries = PETSC_TRUE; /* See MatAssemblyBegin_MPIAIJ. In effect, making MatAssemblyBegin a nop */
3508: PetscCall(MatSetOption(C, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
3509: PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY));
3510: /* MatAssemblyEnd is critical here. It sets mat->offloadmask according to A and B's, and
3511: also gets mpiaij->B compacted (if garray is NULL), with its col ids and size reduced
3512: */
3513: PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY));
3514: PetscCall(MatSetOption(C, MAT_NO_OFF_PROC_ENTRIES, PETSC_FALSE));
3515: PetscCall(MatSetOption(C, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
3516: *mat = C;
3517: PetscFunctionReturn(PETSC_SUCCESS);
3518: }
3520: extern PetscErrorCode MatCreateSubMatrices_MPIAIJ_SingleIS_Local(Mat, PetscInt, const IS[], const IS[], MatReuse, PetscBool, Mat *);
3522: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowDist(Mat mat, IS isrow, IS iscol, IS iscol_local, MatReuse call, Mat *newmat)
3523: {
3524: PetscInt i, m, n, rstart, row, rend, nz, j, bs, cbs;
3525: PetscInt *ii, *jj, nlocal, *dlens, *olens, dlen, olen, jend, mglobal;
3526: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data;
3527: Mat M, Msub, B = a->B;
3528: MatScalar *aa;
3529: Mat_SeqAIJ *aij;
3530: PetscInt *garray = a->garray, *colsub, Ncols;
3531: PetscInt count, Bn = B->cmap->N, cstart = mat->cmap->rstart, cend = mat->cmap->rend;
3532: IS iscol_sub, iscmap;
3533: const PetscInt *is_idx, *cmap;
3534: PetscBool allcolumns = PETSC_FALSE;
3535: MPI_Comm comm;
3537: PetscFunctionBegin;
3538: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3539: if (call == MAT_REUSE_MATRIX) {
3540: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_sub));
3541: PetscCheck(iscol_sub, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "SubIScol passed in was not used before, cannot reuse");
3542: PetscCall(ISGetLocalSize(iscol_sub, &count));
3544: PetscCall(PetscObjectQuery((PetscObject)*newmat, "Subcmap", (PetscObject *)&iscmap));
3545: PetscCheck(iscmap, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Subcmap passed in was not used before, cannot reuse");
3547: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubMatrix", (PetscObject *)&Msub));
3548: PetscCheck(Msub, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3550: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol_sub, MAT_REUSE_MATRIX, PETSC_FALSE, &Msub));
3552: } else { /* call == MAT_INITIAL_MATRIX) */
3553: PetscBool flg;
3555: PetscCall(ISGetLocalSize(iscol, &n));
3556: PetscCall(ISGetSize(iscol, &Ncols));
3558: /* (1) iscol -> nonscalable iscol_local */
3559: /* Check for special case: each processor gets entire matrix columns */
3560: PetscCall(ISIdentity(iscol_local, &flg));
3561: if (flg && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3562: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &allcolumns, 1, MPI_C_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
3563: if (allcolumns) {
3564: iscol_sub = iscol_local;
3565: PetscCall(PetscObjectReference((PetscObject)iscol_local));
3566: PetscCall(ISCreateStride(PETSC_COMM_SELF, n, 0, 1, &iscmap));
3568: } else {
3569: /* (2) iscol_local -> iscol_sub and iscmap. Implementation below requires iscol_local be sorted, it can have duplicate indices */
3570: PetscInt *idx, *cmap1, k;
3571: PetscCall(PetscMalloc1(Ncols, &idx));
3572: PetscCall(PetscMalloc1(Ncols, &cmap1));
3573: PetscCall(ISGetIndices(iscol_local, &is_idx));
3574: count = 0;
3575: k = 0;
3576: for (i = 0; i < Ncols; i++) {
3577: j = is_idx[i];
3578: if (j >= cstart && j < cend) {
3579: /* diagonal part of mat */
3580: idx[count] = j;
3581: cmap1[count++] = i; /* column index in submat */
3582: } else if (Bn) {
3583: /* off-diagonal part of mat */
3584: if (j == garray[k]) {
3585: idx[count] = j;
3586: cmap1[count++] = i; /* column index in submat */
3587: } else if (j > garray[k]) {
3588: while (j > garray[k] && k < Bn - 1) k++;
3589: if (j == garray[k]) {
3590: idx[count] = j;
3591: cmap1[count++] = i; /* column index in submat */
3592: }
3593: }
3594: }
3595: }
3596: PetscCall(ISRestoreIndices(iscol_local, &is_idx));
3598: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, count, idx, PETSC_OWN_POINTER, &iscol_sub));
3599: PetscCall(ISGetBlockSize(iscol, &cbs));
3600: PetscCall(ISSetBlockSize(iscol_sub, cbs));
3602: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)iscol_local), count, cmap1, PETSC_OWN_POINTER, &iscmap));
3603: }
3605: /* (3) Create sequential Msub */
3606: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol_sub, MAT_INITIAL_MATRIX, allcolumns, &Msub));
3607: }
3609: PetscCall(ISGetLocalSize(iscol_sub, &count));
3610: aij = (Mat_SeqAIJ *)Msub->data;
3611: ii = aij->i;
3612: PetscCall(ISGetIndices(iscmap, &cmap));
3614: /*
3615: m - number of local rows
3616: Ncols - number of columns (same on all processors)
3617: rstart - first row in new global matrix generated
3618: */
3619: PetscCall(MatGetSize(Msub, &m, NULL));
3621: if (call == MAT_INITIAL_MATRIX) {
3622: /* (4) Create parallel newmat */
3623: PetscMPIInt rank, size;
3624: PetscInt csize;
3626: PetscCallMPI(MPI_Comm_size(comm, &size));
3627: PetscCallMPI(MPI_Comm_rank(comm, &rank));
3629: /*
3630: Determine the number of non-zeros in the diagonal and off-diagonal
3631: portions of the matrix in order to do correct preallocation
3632: */
3634: /* first get start and end of "diagonal" columns */
3635: PetscCall(ISGetLocalSize(iscol, &csize));
3636: if (csize == PETSC_DECIDE) {
3637: PetscCall(ISGetSize(isrow, &mglobal));
3638: if (mglobal == Ncols) { /* square matrix */
3639: nlocal = m;
3640: } else {
3641: nlocal = Ncols / size + ((Ncols % size) > rank);
3642: }
3643: } else {
3644: nlocal = csize;
3645: }
3646: PetscCallMPI(MPI_Scan(&nlocal, &rend, 1, MPIU_INT, MPI_SUM, comm));
3647: rstart = rend - nlocal;
3648: 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);
3650: /* next, compute all the lengths */
3651: jj = aij->j;
3652: PetscCall(PetscMalloc1(2 * m + 1, &dlens));
3653: olens = dlens + m;
3654: for (i = 0; i < m; i++) {
3655: jend = ii[i + 1] - ii[i];
3656: olen = 0;
3657: dlen = 0;
3658: for (j = 0; j < jend; j++) {
3659: if (cmap[*jj] < rstart || cmap[*jj] >= rend) olen++;
3660: else dlen++;
3661: jj++;
3662: }
3663: olens[i] = olen;
3664: dlens[i] = dlen;
3665: }
3667: PetscCall(ISGetBlockSize(isrow, &bs));
3668: PetscCall(ISGetBlockSize(iscol, &cbs));
3670: PetscCall(MatCreate(comm, &M));
3671: PetscCall(MatSetSizes(M, m, nlocal, PETSC_DECIDE, Ncols));
3672: PetscCall(MatSetBlockSizes(M, bs, cbs));
3673: PetscCall(MatSetType(M, ((PetscObject)mat)->type_name));
3674: PetscCall(MatMPIAIJSetPreallocation(M, 0, dlens, 0, olens));
3675: PetscCall(PetscFree(dlens));
3677: } else { /* call == MAT_REUSE_MATRIX */
3678: M = *newmat;
3679: PetscCall(MatGetLocalSize(M, &i, NULL));
3680: PetscCheck(i == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Previous matrix must be same size/layout as request");
3681: PetscCall(MatZeroEntries(M));
3682: /*
3683: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3684: rather than the slower MatSetValues().
3685: */
3686: M->was_assembled = PETSC_TRUE;
3687: M->assembled = PETSC_FALSE;
3688: }
3690: /* (5) Set values of Msub to *newmat */
3691: PetscCall(PetscMalloc1(count, &colsub));
3692: PetscCall(MatGetOwnershipRange(M, &rstart, NULL));
3694: jj = aij->j;
3695: PetscCall(MatSeqAIJGetArrayRead(Msub, (const PetscScalar **)&aa));
3696: for (i = 0; i < m; i++) {
3697: row = rstart + i;
3698: nz = ii[i + 1] - ii[i];
3699: for (j = 0; j < nz; j++) colsub[j] = cmap[jj[j]];
3700: PetscCall(MatSetValues_MPIAIJ(M, 1, &row, nz, colsub, aa, INSERT_VALUES));
3701: jj += nz;
3702: aa += nz;
3703: }
3704: PetscCall(MatSeqAIJRestoreArrayRead(Msub, (const PetscScalar **)&aa));
3705: PetscCall(ISRestoreIndices(iscmap, &cmap));
3707: PetscCall(MatAssemblyBegin(M, MAT_FINAL_ASSEMBLY));
3708: PetscCall(MatAssemblyEnd(M, MAT_FINAL_ASSEMBLY));
3710: PetscCall(PetscFree(colsub));
3712: /* save Msub, iscol_sub and iscmap used in processor for next request */
3713: if (call == MAT_INITIAL_MATRIX) {
3714: *newmat = M;
3715: PetscCall(PetscObjectCompose((PetscObject)*newmat, "SubMatrix", (PetscObject)Msub));
3716: PetscCall(MatDestroy(&Msub));
3718: PetscCall(PetscObjectCompose((PetscObject)*newmat, "SubIScol", (PetscObject)iscol_sub));
3719: PetscCall(ISDestroy(&iscol_sub));
3721: PetscCall(PetscObjectCompose((PetscObject)*newmat, "Subcmap", (PetscObject)iscmap));
3722: PetscCall(ISDestroy(&iscmap));
3724: if (iscol_local) {
3725: PetscCall(PetscObjectCompose((PetscObject)*newmat, "ISAllGather", (PetscObject)iscol_local));
3726: PetscCall(ISDestroy(&iscol_local));
3727: }
3728: }
3729: PetscFunctionReturn(PETSC_SUCCESS);
3730: }
3732: /*
3733: Not great since it makes two copies of the submatrix, first an SeqAIJ
3734: in local and then by concatenating the local matrices the end result.
3735: Writing it directly would be much like MatCreateSubMatrices_MPIAIJ()
3737: This requires a sequential iscol with all indices.
3738: */
3739: PetscErrorCode MatCreateSubMatrix_MPIAIJ_nonscalable(Mat mat, IS isrow, IS iscol, PetscInt csize, MatReuse call, Mat *newmat)
3740: {
3741: PetscMPIInt rank, size;
3742: PetscInt i, m, n, rstart, row, rend, nz, *cwork, j, bs, cbs;
3743: PetscInt *ii, *jj, nlocal, *dlens, *olens, dlen, olen, jend, mglobal;
3744: Mat M, Mreuse;
3745: MatScalar *aa, *vwork;
3746: MPI_Comm comm;
3747: Mat_SeqAIJ *aij;
3748: PetscBool colflag, allcolumns = PETSC_FALSE;
3750: PetscFunctionBegin;
3751: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3752: PetscCallMPI(MPI_Comm_rank(comm, &rank));
3753: PetscCallMPI(MPI_Comm_size(comm, &size));
3755: /* Check for special case: each processor gets entire matrix columns */
3756: PetscCall(ISIdentity(iscol, &colflag));
3757: PetscCall(ISGetLocalSize(iscol, &n));
3758: if (colflag && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3759: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &allcolumns, 1, MPI_C_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
3761: if (call == MAT_REUSE_MATRIX) {
3762: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubMatrix", (PetscObject *)&Mreuse));
3763: PetscCheck(Mreuse, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3764: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol, MAT_REUSE_MATRIX, allcolumns, &Mreuse));
3765: } else {
3766: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol, MAT_INITIAL_MATRIX, allcolumns, &Mreuse));
3767: }
3769: /*
3770: m - number of local rows
3771: n - number of columns (same on all processors)
3772: rstart - first row in new global matrix generated
3773: */
3774: PetscCall(MatGetSize(Mreuse, &m, &n));
3775: PetscCall(MatGetBlockSizes(Mreuse, &bs, &cbs));
3776: if (call == MAT_INITIAL_MATRIX) {
3777: aij = (Mat_SeqAIJ *)Mreuse->data;
3778: ii = aij->i;
3779: jj = aij->j;
3781: /*
3782: Determine the number of non-zeros in the diagonal and off-diagonal
3783: portions of the matrix in order to do correct preallocation
3784: */
3786: /* first get start and end of "diagonal" columns */
3787: if (csize == PETSC_DECIDE) {
3788: PetscCall(ISGetSize(isrow, &mglobal));
3789: if (mglobal == n) { /* square matrix */
3790: nlocal = m;
3791: } else {
3792: nlocal = n / size + ((n % size) > rank);
3793: }
3794: } else {
3795: nlocal = csize;
3796: }
3797: PetscCallMPI(MPI_Scan(&nlocal, &rend, 1, MPIU_INT, MPI_SUM, comm));
3798: rstart = rend - nlocal;
3799: 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);
3801: /* next, compute all the lengths */
3802: PetscCall(PetscMalloc1(2 * m + 1, &dlens));
3803: olens = dlens + m;
3804: for (i = 0; i < m; i++) {
3805: jend = ii[i + 1] - ii[i];
3806: olen = 0;
3807: dlen = 0;
3808: for (j = 0; j < jend; j++) {
3809: if (*jj < rstart || *jj >= rend) olen++;
3810: else dlen++;
3811: jj++;
3812: }
3813: olens[i] = olen;
3814: dlens[i] = dlen;
3815: }
3816: PetscCall(MatCreate(comm, &M));
3817: PetscCall(MatSetSizes(M, m, nlocal, PETSC_DECIDE, n));
3818: PetscCall(MatSetBlockSizes(M, bs, cbs));
3819: PetscCall(MatSetType(M, ((PetscObject)mat)->type_name));
3820: PetscCall(MatMPIAIJSetPreallocation(M, 0, dlens, 0, olens));
3821: PetscCall(PetscFree(dlens));
3822: } else {
3823: PetscInt ml, nl;
3825: M = *newmat;
3826: PetscCall(MatGetLocalSize(M, &ml, &nl));
3827: PetscCheck(ml == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Previous matrix must be same size/layout as request");
3828: PetscCall(MatZeroEntries(M));
3829: /*
3830: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3831: rather than the slower MatSetValues().
3832: */
3833: M->was_assembled = PETSC_TRUE;
3834: M->assembled = PETSC_FALSE;
3835: }
3836: PetscCall(MatGetOwnershipRange(M, &rstart, &rend));
3837: aij = (Mat_SeqAIJ *)Mreuse->data;
3838: ii = aij->i;
3839: jj = aij->j;
3841: /* trigger copy to CPU if needed */
3842: PetscCall(MatSeqAIJGetArrayRead(Mreuse, (const PetscScalar **)&aa));
3843: for (i = 0; i < m; i++) {
3844: row = rstart + i;
3845: nz = ii[i + 1] - ii[i];
3846: cwork = jj;
3847: jj = PetscSafePointerPlusOffset(jj, nz);
3848: vwork = aa;
3849: aa = PetscSafePointerPlusOffset(aa, nz);
3850: PetscCall(MatSetValues_MPIAIJ(M, 1, &row, nz, cwork, vwork, INSERT_VALUES));
3851: }
3852: PetscCall(MatSeqAIJRestoreArrayRead(Mreuse, (const PetscScalar **)&aa));
3854: PetscCall(MatAssemblyBegin(M, MAT_FINAL_ASSEMBLY));
3855: PetscCall(MatAssemblyEnd(M, MAT_FINAL_ASSEMBLY));
3856: *newmat = M;
3858: /* save submatrix used in processor for next request */
3859: if (call == MAT_INITIAL_MATRIX) {
3860: PetscCall(PetscObjectCompose((PetscObject)M, "SubMatrix", (PetscObject)Mreuse));
3861: PetscCall(MatDestroy(&Mreuse));
3862: }
3863: PetscFunctionReturn(PETSC_SUCCESS);
3864: }
3866: static PetscErrorCode MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[])
3867: {
3868: PetscInt m, cstart, cend, j, nnz, i, d, *ld;
3869: PetscInt *d_nnz, *o_nnz, nnz_max = 0, rstart, ii, irstart;
3870: const PetscInt *JJ;
3871: PetscBool nooffprocentries;
3872: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)B->data;
3874: PetscFunctionBegin;
3875: PetscCall(PetscLayoutSetUp(B->rmap));
3876: PetscCall(PetscLayoutSetUp(B->cmap));
3877: m = B->rmap->n;
3878: cstart = B->cmap->rstart;
3879: cend = B->cmap->rend;
3880: rstart = B->rmap->rstart;
3881: irstart = Ii[0];
3883: PetscCall(PetscCalloc2(m, &d_nnz, m, &o_nnz));
3885: if (PetscDefined(USE_DEBUG)) {
3886: for (i = 0; i < m; i++) {
3887: nnz = Ii[i + 1] - Ii[i];
3888: JJ = PetscSafePointerPlusOffset(J, Ii[i] - irstart);
3889: PetscCheck(nnz >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Local row %" PetscInt_FMT " has a negative %" PetscInt_FMT " number of columns", i, nnz);
3890: 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]);
3891: 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);
3892: }
3893: }
3895: for (i = 0; i < m; i++) {
3896: nnz = Ii[i + 1] - Ii[i];
3897: JJ = PetscSafePointerPlusOffset(J, Ii[i] - irstart);
3898: nnz_max = PetscMax(nnz_max, nnz);
3899: d = 0;
3900: for (j = 0; j < nnz; j++) {
3901: if (cstart <= JJ[j] && JJ[j] < cend) d++;
3902: }
3903: d_nnz[i] = d;
3904: o_nnz[i] = nnz - d;
3905: }
3906: PetscCall(MatMPIAIJSetPreallocation(B, 0, d_nnz, 0, o_nnz));
3907: PetscCall(PetscFree2(d_nnz, o_nnz));
3909: for (i = 0; i < m; i++) {
3910: ii = i + rstart;
3911: PetscCall(MatSetValues_MPIAIJ(B, 1, &ii, Ii[i + 1] - Ii[i], PetscSafePointerPlusOffset(J, Ii[i] - irstart), PetscSafePointerPlusOffset(v, Ii[i] - irstart), INSERT_VALUES));
3912: }
3913: nooffprocentries = B->nooffprocentries;
3914: B->nooffprocentries = PETSC_TRUE;
3915: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
3916: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
3917: B->nooffprocentries = nooffprocentries;
3919: /* count number of entries below block diagonal */
3920: PetscCall(PetscFree(Aij->ld));
3921: PetscCall(PetscCalloc1(m, &ld));
3922: Aij->ld = ld;
3923: for (i = 0; i < m; i++) {
3924: nnz = Ii[i + 1] - Ii[i];
3925: j = 0;
3926: while (j < nnz && J[j] < cstart) j++;
3927: ld[i] = j;
3928: if (J) J += nnz;
3929: }
3931: PetscCall(MatSetOption(B, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
3932: PetscFunctionReturn(PETSC_SUCCESS);
3933: }
3935: /*@
3936: MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in `MATAIJ` format
3937: (the default parallel PETSc format).
3939: Collective
3941: Input Parameters:
3942: + B - the matrix
3943: . i - the indices into `j` for the start of each local row (indices start with zero)
3944: . j - the column indices for each local row (indices start with zero)
3945: - v - optional values in the matrix
3947: Level: developer
3949: Notes:
3950: The `i`, `j`, and `v` arrays ARE copied by this routine into the internal format used by PETSc;
3951: thus you CANNOT change the matrix entries by changing the values of `v` after you have
3952: called this routine. Use `MatCreateMPIAIJWithSplitArrays()` to avoid needing to copy the arrays.
3954: The `i` and `j` indices are 0 based, and `i` indices are indices corresponding to the local `j` array.
3956: A convenience routine for this functionality is `MatCreateMPIAIJWithArrays()`.
3958: You can update the matrix with new numerical values using `MatUpdateMPIAIJWithArrays()` after this call if the column indices in `j` are sorted.
3960: If you do **not** use `MatUpdateMPIAIJWithArrays()`, the column indices in `j` do not need to be sorted. If you will use
3961: `MatUpdateMPIAIJWithArrays()`, the column indices **must** be sorted.
3963: The format which is used for the sparse matrix input, is equivalent to a
3964: row-major ordering.. i.e for the following matrix, the input data expected is
3965: as shown
3966: .vb
3967: 1 0 0
3968: 2 0 3 P0
3969: -------
3970: 4 5 6 P1
3972: Process0 [P0] rows_owned=[0,1]
3973: i = {0,1,3} [size = nrow+1 = 2+1]
3974: j = {0,0,2} [size = 3]
3975: v = {1,2,3} [size = 3]
3977: Process1 [P1] rows_owned=[2]
3978: i = {0,3} [size = nrow+1 = 1+1]
3979: j = {0,1,2} [size = 3]
3980: v = {4,5,6} [size = 3]
3981: .ve
3983: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatCreateAIJ()`,
3984: `MatCreateSeqAIJWithArrays()`, `MatCreateMPIAIJWithSplitArrays()`, `MatCreateMPIAIJWithArrays()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
3985: @*/
3986: PetscErrorCode MatMPIAIJSetPreallocationCSR(Mat B, const PetscInt i[], const PetscInt j[], const PetscScalar v[])
3987: {
3988: PetscFunctionBegin;
3989: PetscTryMethod(B, "MatMPIAIJSetPreallocationCSR_C", (Mat, const PetscInt[], const PetscInt[], const PetscScalar[]), (B, i, j, v));
3990: PetscFunctionReturn(PETSC_SUCCESS);
3991: }
3993: /*@
3994: MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in `MATMPIAIJ` format
3995: (the default parallel PETSc format). For good matrix assembly performance
3996: the user should preallocate the matrix storage by setting the parameters
3997: `d_nz` (or `d_nnz`) and `o_nz` (or `o_nnz`).
3999: Collective
4001: Input Parameters:
4002: + B - the matrix
4003: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
4004: (same value is used for all local rows)
4005: . d_nnz - array containing the number of nonzeros in the various rows of the
4006: DIAGONAL portion of the local submatrix (possibly different for each row)
4007: or `NULL` (`PETSC_NULL_INTEGER` in Fortran), if `d_nz` is used to specify the nonzero structure.
4008: The size of this array is equal to the number of local rows, i.e 'm'.
4009: For matrices that will be factored, you must leave room for (and set)
4010: the diagonal entry even if it is zero.
4011: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
4012: submatrix (same value is used for all local rows).
4013: - o_nnz - array containing the number of nonzeros in the various rows of the
4014: OFF-DIAGONAL portion of the local submatrix (possibly different for
4015: each row) or `NULL` (`PETSC_NULL_INTEGER` in Fortran), if `o_nz` is used to specify the nonzero
4016: structure. The size of this array is equal to the number
4017: of local rows, i.e 'm'.
4019: Example Usage:
4020: Consider the following 8x8 matrix with 34 non-zero values, that is
4021: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4022: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4023: as follows
4025: .vb
4026: 1 2 0 | 0 3 0 | 0 4
4027: Proc0 0 5 6 | 7 0 0 | 8 0
4028: 9 0 10 | 11 0 0 | 12 0
4029: -------------------------------------
4030: 13 0 14 | 15 16 17 | 0 0
4031: Proc1 0 18 0 | 19 20 21 | 0 0
4032: 0 0 0 | 22 23 0 | 24 0
4033: -------------------------------------
4034: Proc2 25 26 27 | 0 0 28 | 29 0
4035: 30 0 0 | 31 32 33 | 0 34
4036: .ve
4038: This can be represented as a collection of submatrices as
4039: .vb
4040: A B C
4041: D E F
4042: G H I
4043: .ve
4045: Where the submatrices A,B,C are owned by proc0, D,E,F are
4046: owned by proc1, G,H,I are owned by proc2.
4048: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4049: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4050: The 'M','N' parameters are 8,8, and have the same values on all procs.
4052: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4053: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4054: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4055: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4056: part as `MATSEQAIJ` matrices. For example, proc1 will store [E] as a `MATSEQAIJ`
4057: matrix, and [DF] as another `MATSEQAIJ` matrix.
4059: When `d_nz`, `o_nz` parameters are specified, `d_nz` storage elements are
4060: allocated for every row of the local DIAGONAL submatrix, and `o_nz`
4061: storage locations are allocated for every row of the OFF-DIAGONAL submatrix.
4062: One way to choose `d_nz` and `o_nz` is to use the maximum number of nonzeros over
4063: the local rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4064: In this case, the values of `d_nz`, `o_nz` are
4065: .vb
4066: proc0 dnz = 2, o_nz = 2
4067: proc1 dnz = 3, o_nz = 2
4068: proc2 dnz = 1, o_nz = 4
4069: .ve
4070: We are allocating `m`*(`d_nz`+`o_nz`) storage locations for every proc. This
4071: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4072: for proc3. i.e we are using 12+15+10=37 storage locations to store
4073: 34 values.
4075: When `d_nnz`, `o_nnz` parameters are specified, the storage is specified
4076: for every row, corresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4077: In the above case the values for `d_nnz`, `o_nnz` are
4078: .vb
4079: proc0 d_nnz = [2,2,2] and o_nnz = [2,2,2]
4080: proc1 d_nnz = [3,3,2] and o_nnz = [2,1,1]
4081: proc2 d_nnz = [1,1] and o_nnz = [4,4]
4082: .ve
4083: Here the space allocated is sum of all the above values i.e 34, and
4084: hence pre-allocation is perfect.
4086: Level: intermediate
4088: Notes:
4089: If the *_nnz parameter is given then the *_nz parameter is ignored
4091: The `MATAIJ` format, also called compressed row storage (CSR), is compatible with standard Fortran
4092: storage. The stored row and column indices begin with zero.
4093: See [Sparse Matrices](sec_matsparse) for details.
4095: The parallel matrix is partitioned such that the first m0 rows belong to
4096: process 0, the next m1 rows belong to process 1, the next m2 rows belong
4097: to process 2 etc.. where m0,m1,m2... are the input parameter 'm'.
4099: The DIAGONAL portion of the local submatrix of a processor can be defined
4100: as the submatrix which is obtained by extraction the part corresponding to
4101: the rows r1-r2 and columns c1-c2 of the global matrix, where r1 is the
4102: first row that belongs to the processor, r2 is the last row belonging to
4103: the this processor, and c1-c2 is range of indices of the local part of a
4104: vector suitable for applying the matrix to. This is an mxn matrix. In the
4105: common case of a square matrix, the row and column ranges are the same and
4106: the DIAGONAL part is also square. The remaining portion of the local
4107: submatrix (mxN) constitute the OFF-DIAGONAL portion.
4109: If `o_nnz` and `d_nnz` are specified, then `o_nz` and `d_nz` are ignored.
4111: You can call `MatGetInfo()` to get information on how effective the preallocation was;
4112: for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
4113: You can also run with the option `-info` and look for messages with the string
4114: malloc in them to see if additional memory allocation was needed.
4116: .seealso: [](ch_matrices), `Mat`, [Sparse Matrices](sec_matsparse), `MATMPIAIJ`, `MATAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatCreateAIJ()`, `MatMPIAIJSetPreallocationCSR()`,
4117: `MatGetInfo()`, `PetscSplitOwnership()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4118: @*/
4119: PetscErrorCode MatMPIAIJSetPreallocation(Mat B, PetscInt d_nz, const PetscInt d_nnz[], PetscInt o_nz, const PetscInt o_nnz[])
4120: {
4121: PetscFunctionBegin;
4124: PetscTryMethod(B, "MatMPIAIJSetPreallocation_C", (Mat, PetscInt, const PetscInt[], PetscInt, const PetscInt[]), (B, d_nz, d_nnz, o_nz, o_nnz));
4125: PetscFunctionReturn(PETSC_SUCCESS);
4126: }
4128: /*@
4129: MatCreateMPIAIJWithArrays - creates a `MATMPIAIJ` matrix using arrays that contain in standard
4130: CSR format for the local rows.
4132: Collective
4134: Input Parameters:
4135: + comm - MPI communicator
4136: . m - number of local rows (Cannot be `PETSC_DECIDE`)
4137: . n - This value should be the same as the local size used in creating the
4138: x vector for the matrix-vector product $ y = Ax$. (or `PETSC_DECIDE` to have
4139: calculated if `N` is given) For square matrices n is almost always `m`.
4140: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if `m` is given)
4141: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if `n` is given)
4142: . 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
4143: . j - global column indices
4144: - a - optional matrix values
4146: Output Parameter:
4147: . mat - the matrix
4149: Level: intermediate
4151: Notes:
4152: The `i`, `j`, and `a` arrays ARE copied by this routine into the internal format used by PETSc;
4153: thus you CANNOT change the matrix entries by changing the values of `a[]` after you have
4154: called this routine. Use `MatCreateMPIAIJWithSplitArrays()` to avoid needing to copy the arrays.
4156: The `i` and `j` indices are 0 based, and `i` indices are indices corresponding to the local `j` array.
4158: Once you have created the matrix you can update it with new numerical values using `MatUpdateMPIAIJWithArray()`
4160: If you do **not** use `MatUpdateMPIAIJWithArray()`, the column indices in `j` do not need to be sorted. If you will use
4161: `MatUpdateMPIAIJWithArrays()`, the column indices **must** be sorted.
4163: The format which is used for the sparse matrix input, is equivalent to a
4164: row-major ordering, i.e., for the following matrix, the input data expected is
4165: as shown
4166: .vb
4167: 1 0 0
4168: 2 0 3 P0
4169: -------
4170: 4 5 6 P1
4172: Process0 [P0] rows_owned=[0,1]
4173: i = {0,1,3} [size = nrow+1 = 2+1]
4174: j = {0,0,2} [size = 3]
4175: v = {1,2,3} [size = 3]
4177: Process1 [P1] rows_owned=[2]
4178: i = {0,3} [size = nrow+1 = 1+1]
4179: j = {0,1,2} [size = 3]
4180: v = {4,5,6} [size = 3]
4181: .ve
4183: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4184: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArray()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4185: @*/
4186: PetscErrorCode MatCreateMPIAIJWithArrays(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt M, PetscInt N, const PetscInt i[], const PetscInt j[], const PetscScalar a[], Mat *mat)
4187: {
4188: PetscFunctionBegin;
4189: PetscCheck(!i || !i[0], PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
4190: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
4191: PetscCall(MatCreate(comm, mat));
4192: PetscCall(MatSetSizes(*mat, m, n, M, N));
4193: /* PetscCall(MatSetBlockSizes(M,bs,cbs)); */
4194: PetscCall(MatSetType(*mat, MATMPIAIJ));
4195: PetscCall(MatMPIAIJSetPreallocationCSR(*mat, i, j, a));
4196: PetscFunctionReturn(PETSC_SUCCESS);
4197: }
4199: /*@
4200: MatUpdateMPIAIJWithArrays - updates a `MATMPIAIJ` matrix using arrays that contain in standard
4201: CSR format for the local rows. Only the numerical values are updated the other arrays must be identical to what was passed
4202: from `MatCreateMPIAIJWithArrays()`
4204: Deprecated: Use `MatUpdateMPIAIJWithArray()`
4206: Collective
4208: Input Parameters:
4209: + mat - the matrix
4210: . m - number of local rows (Cannot be `PETSC_DECIDE`)
4211: . n - This value should be the same as the local size used in creating the
4212: x vector for the matrix-vector product y = Ax. (or `PETSC_DECIDE` to have
4213: calculated if N is given) For square matrices n is almost always m.
4214: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if m is given)
4215: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if n is given)
4216: . Ii - row indices; that is Ii[0] = 0, Ii[row] = Ii[row-1] + number of elements in that row of the matrix
4217: . J - column indices
4218: - v - matrix values
4220: Level: deprecated
4222: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4223: `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArray()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4224: @*/
4225: PetscErrorCode MatUpdateMPIAIJWithArrays(Mat mat, PetscInt m, PetscInt n, PetscInt M, PetscInt N, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[])
4226: {
4227: PetscInt nnz, i;
4228: PetscBool nooffprocentries;
4229: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)mat->data;
4230: Mat_SeqAIJ *Ad = (Mat_SeqAIJ *)Aij->A->data;
4231: PetscScalar *ad, *ao;
4232: PetscInt ldi, Iii, md;
4233: const PetscInt *Adi = Ad->i;
4234: PetscInt *ld = Aij->ld;
4236: PetscFunctionBegin;
4237: PetscCheck(Ii[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
4238: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
4239: PetscCheck(m == mat->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Local number of rows cannot change from call to MatUpdateMPIAIJWithArrays()");
4240: PetscCheck(n == mat->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Local number of columns cannot change from call to MatUpdateMPIAIJWithArrays()");
4242: PetscCall(MatSeqAIJGetArrayWrite(Aij->A, &ad));
4243: PetscCall(MatSeqAIJGetArrayWrite(Aij->B, &ao));
4245: for (i = 0; i < m; i++) {
4246: if (PetscDefined(USE_DEBUG)) {
4247: for (PetscInt j = Ii[i] + 1; j < Ii[i + 1]; ++j) {
4248: 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);
4249: 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);
4250: }
4251: }
4252: nnz = Ii[i + 1] - Ii[i];
4253: Iii = Ii[i];
4254: ldi = ld[i];
4255: md = Adi[i + 1] - Adi[i];
4256: PetscCall(PetscArraycpy(ao, v + Iii, ldi));
4257: PetscCall(PetscArraycpy(ad, v + Iii + ldi, md));
4258: PetscCall(PetscArraycpy(ao + ldi, v + Iii + ldi + md, nnz - ldi - md));
4259: ad += md;
4260: ao += nnz - md;
4261: }
4262: nooffprocentries = mat->nooffprocentries;
4263: mat->nooffprocentries = PETSC_TRUE;
4264: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->A, &ad));
4265: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->B, &ao));
4266: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->A));
4267: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->B));
4268: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
4269: PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
4270: PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
4271: mat->nooffprocentries = nooffprocentries;
4272: PetscFunctionReturn(PETSC_SUCCESS);
4273: }
4275: /*@
4276: MatUpdateMPIAIJWithArray - updates an `MATMPIAIJ` matrix using an array that contains the nonzero values
4278: Collective
4280: Input Parameters:
4281: + mat - the matrix
4282: - v - matrix values, stored by row
4284: Level: intermediate
4286: Notes:
4287: The matrix must have been obtained with `MatCreateMPIAIJWithArrays()` or `MatMPIAIJSetPreallocationCSR()`
4289: The column indices in the call to `MatCreateMPIAIJWithArrays()` or `MatMPIAIJSetPreallocationCSR()` must have been sorted for this call to work correctly
4291: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4292: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArrays()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4293: @*/
4294: PetscErrorCode MatUpdateMPIAIJWithArray(Mat mat, const PetscScalar v[])
4295: {
4296: PetscInt nnz, i, m;
4297: PetscBool nooffprocentries;
4298: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)mat->data;
4299: Mat_SeqAIJ *Ad = (Mat_SeqAIJ *)Aij->A->data;
4300: Mat_SeqAIJ *Ao = (Mat_SeqAIJ *)Aij->B->data;
4301: PetscScalar *ad, *ao;
4302: const PetscInt *Adi = Ad->i, *Adj = Ao->i;
4303: PetscInt ldi, Iii, md;
4304: PetscInt *ld = Aij->ld;
4306: PetscFunctionBegin;
4307: m = mat->rmap->n;
4309: PetscCall(MatSeqAIJGetArrayWrite(Aij->A, &ad));
4310: PetscCall(MatSeqAIJGetArrayWrite(Aij->B, &ao));
4311: Iii = 0;
4312: for (i = 0; i < m; i++) {
4313: nnz = Adi[i + 1] - Adi[i] + Adj[i + 1] - Adj[i];
4314: ldi = ld[i];
4315: md = Adi[i + 1] - Adi[i];
4316: PetscCall(PetscArraycpy(ad, v + Iii + ldi, md));
4317: ad += md;
4318: if (ao) {
4319: PetscCall(PetscArraycpy(ao, v + Iii, ldi));
4320: PetscCall(PetscArraycpy(ao + ldi, v + Iii + ldi + md, nnz - ldi - md));
4321: ao += nnz - md;
4322: }
4323: Iii += nnz;
4324: }
4325: nooffprocentries = mat->nooffprocentries;
4326: mat->nooffprocentries = PETSC_TRUE;
4327: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->A, &ad));
4328: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->B, &ao));
4329: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->A));
4330: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->B));
4331: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
4332: PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
4333: PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
4334: mat->nooffprocentries = nooffprocentries;
4335: PetscFunctionReturn(PETSC_SUCCESS);
4336: }
4338: /*@
4339: MatCreateAIJ - Creates a sparse parallel matrix in `MATAIJ` format
4340: (the default parallel PETSc format). For good matrix assembly performance
4341: the user should preallocate the matrix storage by setting the parameters
4342: `d_nz` (or `d_nnz`) and `o_nz` (or `o_nnz`).
4344: Collective
4346: Input Parameters:
4347: + comm - MPI communicator
4348: . m - number of local rows (or `PETSC_DECIDE` to have calculated if M is given)
4349: This value should be the same as the local size used in creating the
4350: y vector for the matrix-vector product y = Ax.
4351: . n - This value should be the same as the local size used in creating the
4352: x vector for the matrix-vector product y = Ax. (or `PETSC_DECIDE` to have
4353: calculated if N is given) For square matrices n is almost always m.
4354: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if m is given)
4355: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if n is given)
4356: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
4357: (same value is used for all local rows)
4358: . d_nnz - array containing the number of nonzeros in the various rows of the
4359: DIAGONAL portion of the local submatrix (possibly different for each row)
4360: or `NULL`, if `d_nz` is used to specify the nonzero structure.
4361: The size of this array is equal to the number of local rows, i.e 'm'.
4362: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
4363: submatrix (same value is used for all local rows).
4364: - o_nnz - array containing the number of nonzeros in the various rows of the
4365: OFF-DIAGONAL portion of the local submatrix (possibly different for
4366: each row) or `NULL`, if `o_nz` is used to specify the nonzero
4367: structure. The size of this array is equal to the number
4368: of local rows, i.e 'm'.
4370: Output Parameter:
4371: . A - the matrix
4373: Options Database Keys:
4374: + -mat_no_inode - Do not use inodes
4375: . -mat_inode_limit <limit> - Sets inode limit (max limit=5)
4376: - -matmult_vecscatter_view <viewer> - View the vecscatter (i.e., communication pattern) used in `MatMult()` of sparse parallel matrices.
4377: See viewer types in manual of `MatView()`. Of them, ascii_matlab, draw or binary cause the `VecScatter`
4378: 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.
4380: Level: intermediate
4382: Notes:
4383: It is recommended that one use `MatCreateFromOptions()` or the `MatCreate()`, `MatSetType()` and/or `MatSetFromOptions()`,
4384: MatXXXXSetPreallocation() paradigm instead of this routine directly.
4385: [MatXXXXSetPreallocation() is, for example, `MatSeqAIJSetPreallocation()`]
4387: If the *_nnz parameter is given then the *_nz parameter is ignored
4389: The `m`,`n`,`M`,`N` parameters specify the size of the matrix, and its partitioning across
4390: processors, while `d_nz`,`d_nnz`,`o_nz`,`o_nnz` parameters specify the approximate
4391: storage requirements for this matrix.
4393: If `PETSC_DECIDE` or `PETSC_DETERMINE` is used for a particular argument on one
4394: processor than it must be used on all processors that share the object for
4395: that argument.
4397: If `m` and `n` are not `PETSC_DECIDE`, then the values determine the `PetscLayout` of the matrix and the ranges returned by
4398: `MatGetOwnershipRange()`, `MatGetOwnershipRanges()`, `MatGetOwnershipRangeColumn()`, and `MatGetOwnershipRangesColumn()`.
4400: The user MUST specify either the local or global matrix dimensions
4401: (possibly both).
4403: The parallel matrix is partitioned across processors such that the
4404: first `m0` rows belong to process 0, the next `m1` rows belong to
4405: process 1, the next `m2` rows belong to process 2, etc., where
4406: `m0`, `m1`, `m2`... are the input parameter `m` on each MPI process. I.e., each MPI process stores
4407: values corresponding to [m x N] submatrix.
4409: The columns are logically partitioned with the n0 columns belonging
4410: to 0th partition, the next n1 columns belonging to the next
4411: partition etc.. where n0,n1,n2... are the input parameter 'n'.
4413: The DIAGONAL portion of the local submatrix on any given processor
4414: is the submatrix corresponding to the rows and columns m,n
4415: corresponding to the given processor. i.e diagonal matrix on
4416: process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1]
4417: etc. The remaining portion of the local submatrix [m x (N-n)]
4418: constitute the OFF-DIAGONAL portion. The example below better
4419: illustrates this concept. The two matrices, the DIAGONAL portion and
4420: the OFF-DIAGONAL portion are each stored as `MATSEQAIJ` matrices.
4422: For a square global matrix we define each processor's diagonal portion
4423: to be its local rows and the corresponding columns (a square submatrix);
4424: each processor's off-diagonal portion encompasses the remainder of the
4425: local matrix (a rectangular submatrix).
4427: If `o_nnz`, `d_nnz` are specified, then `o_nz`, and `d_nz` are ignored.
4429: When calling this routine with a single process communicator, a matrix of
4430: type `MATSEQAIJ` is returned. If a matrix of type `MATMPIAIJ` is desired for this
4431: type of communicator, use the construction mechanism
4432: .vb
4433: MatCreate(..., &A);
4434: MatSetType(A, MATMPIAIJ);
4435: MatSetSizes(A, m, n, M, N);
4436: MatMPIAIJSetPreallocation(A, ...);
4437: .ve
4439: By default, this format uses inodes (identical nodes) when possible.
4440: We search for consecutive rows with the same nonzero structure, thereby
4441: reusing matrix information to achieve increased efficiency.
4443: Example Usage:
4444: Consider the following 8x8 matrix with 34 non-zero values, that is
4445: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4446: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4447: as follows
4449: .vb
4450: 1 2 0 | 0 3 0 | 0 4
4451: Proc0 0 5 6 | 7 0 0 | 8 0
4452: 9 0 10 | 11 0 0 | 12 0
4453: -------------------------------------
4454: 13 0 14 | 15 16 17 | 0 0
4455: Proc1 0 18 0 | 19 20 21 | 0 0
4456: 0 0 0 | 22 23 0 | 24 0
4457: -------------------------------------
4458: Proc2 25 26 27 | 0 0 28 | 29 0
4459: 30 0 0 | 31 32 33 | 0 34
4460: .ve
4462: This can be represented as a collection of submatrices as
4464: .vb
4465: A B C
4466: D E F
4467: G H I
4468: .ve
4470: Where the submatrices A,B,C are owned by proc0, D,E,F are
4471: owned by proc1, G,H,I are owned by proc2.
4473: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4474: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4475: The 'M','N' parameters are 8,8, and have the same values on all procs.
4477: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4478: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4479: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4480: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4481: part as `MATSEQAIJ` matrices. For example, proc1 will store [E] as a `MATSEQAIJ`
4482: matrix, and [DF] as another SeqAIJ matrix.
4484: When `d_nz`, `o_nz` parameters are specified, `d_nz` storage elements are
4485: allocated for every row of the local DIAGONAL submatrix, and `o_nz`
4486: storage locations are allocated for every row of the OFF-DIAGONAL submatrix.
4487: One way to choose `d_nz` and `o_nz` is to use the maximum number of nonzeros over
4488: the local rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4489: In this case, the values of `d_nz`,`o_nz` are
4490: .vb
4491: proc0 dnz = 2, o_nz = 2
4492: proc1 dnz = 3, o_nz = 2
4493: proc2 dnz = 1, o_nz = 4
4494: .ve
4495: We are allocating m*(`d_nz`+`o_nz`) storage locations for every proc. This
4496: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4497: for proc3. i.e we are using 12+15+10=37 storage locations to store
4498: 34 values.
4500: When `d_nnz`, `o_nnz` parameters are specified, the storage is specified
4501: for every row, corresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4502: In the above case the values for d_nnz,o_nnz are
4503: .vb
4504: proc0 d_nnz = [2,2,2] and o_nnz = [2,2,2]
4505: proc1 d_nnz = [3,3,2] and o_nnz = [2,1,1]
4506: proc2 d_nnz = [1,1] and o_nnz = [4,4]
4507: .ve
4508: Here the space allocated is sum of all the above values i.e 34, and
4509: hence pre-allocation is perfect.
4511: .seealso: [](ch_matrices), `Mat`, [Sparse Matrix Creation](sec_matsparse), `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4512: `MATMPIAIJ`, `MatCreateMPIAIJWithArrays()`, `MatGetOwnershipRange()`, `MatGetOwnershipRanges()`, `MatGetOwnershipRangeColumn()`,
4513: `MatGetOwnershipRangesColumn()`, `PetscLayout`
4514: @*/
4515: 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)
4516: {
4517: PetscMPIInt size;
4519: PetscFunctionBegin;
4520: PetscCall(MatCreate(comm, A));
4521: PetscCall(MatSetSizes(*A, m, n, M, N));
4522: PetscCallMPI(MPI_Comm_size(comm, &size));
4523: if (size > 1) {
4524: PetscCall(MatSetType(*A, MATMPIAIJ));
4525: PetscCall(MatMPIAIJSetPreallocation(*A, d_nz, d_nnz, o_nz, o_nnz));
4526: } else {
4527: PetscCall(MatSetType(*A, MATSEQAIJ));
4528: PetscCall(MatSeqAIJSetPreallocation(*A, d_nz, d_nnz));
4529: }
4530: PetscFunctionReturn(PETSC_SUCCESS);
4531: }
4533: /*@C
4534: MatMPIAIJGetSeqAIJ - Returns the local pieces of this distributed matrix
4536: Not Collective
4538: Input Parameter:
4539: . A - The `MATMPIAIJ` matrix
4541: Output Parameters:
4542: + Ad - The local diagonal block as a `MATSEQAIJ` matrix
4543: . Ao - The local off-diagonal block as a `MATSEQAIJ` matrix
4544: - colmap - An array mapping local column numbers of `Ao` to global column numbers of the parallel matrix
4546: Level: intermediate
4548: Note:
4549: The rows in `Ad` and `Ao` are in [0, Nr), where Nr is the number of local rows on this process. The columns
4550: 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
4551: the number of nonzero columns in the local off-diagonal piece of the matrix `A`. The array colmap maps these
4552: local column numbers to global column numbers in the original matrix.
4554: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatMPIAIJGetLocalMat()`, `MatMPIAIJGetLocalMatCondensed()`, `MatCreateAIJ()`, `MATSEQAIJ`
4555: @*/
4556: PetscErrorCode MatMPIAIJGetSeqAIJ(Mat A, Mat *Ad, Mat *Ao, const PetscInt *colmap[])
4557: {
4558: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
4559: PetscBool flg;
4561: PetscFunctionBegin;
4562: PetscCall(PetscStrbeginswith(((PetscObject)A)->type_name, MATMPIAIJ, &flg));
4563: PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "This function requires a MATMPIAIJ matrix as input");
4564: if (Ad) *Ad = a->A;
4565: if (Ao) *Ao = a->B;
4566: if (colmap) *colmap = a->garray;
4567: PetscFunctionReturn(PETSC_SUCCESS);
4568: }
4570: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm comm, Mat inmat, PetscInt n, MatReuse scall, Mat *outmat)
4571: {
4572: PetscInt m, N, i, rstart, nnz, Ii;
4573: PetscInt *indx;
4574: PetscScalar *values;
4575: MatType rootType;
4577: PetscFunctionBegin;
4578: PetscCall(MatGetSize(inmat, &m, &N));
4579: if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */
4580: PetscInt *dnz, *onz, sum, bs, cbs;
4582: if (n == PETSC_DECIDE) PetscCall(PetscSplitOwnership(comm, &n, &N));
4583: /* Check sum(n) = N */
4584: PetscCallMPI(MPIU_Allreduce(&n, &sum, 1, MPIU_INT, MPI_SUM, comm));
4585: PetscCheck(sum == N, PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Sum of local columns %" PetscInt_FMT " != global columns %" PetscInt_FMT, sum, N);
4587: PetscCallMPI(MPI_Scan(&m, &rstart, 1, MPIU_INT, MPI_SUM, comm));
4588: rstart -= m;
4590: MatPreallocateBegin(comm, m, n, dnz, onz);
4591: for (i = 0; i < m; i++) {
4592: PetscCall(MatGetRow_SeqAIJ(inmat, i, &nnz, &indx, NULL));
4593: PetscCall(MatPreallocateSet(i + rstart, nnz, indx, dnz, onz));
4594: PetscCall(MatRestoreRow_SeqAIJ(inmat, i, &nnz, &indx, NULL));
4595: }
4597: PetscCall(MatCreate(comm, outmat));
4598: PetscCall(MatSetSizes(*outmat, m, n, PETSC_DETERMINE, PETSC_DETERMINE));
4599: PetscCall(MatGetBlockSizes(inmat, &bs, &cbs));
4600: PetscCall(MatSetBlockSizes(*outmat, bs, cbs));
4601: PetscCall(MatGetRootType_Private(inmat, &rootType));
4602: PetscCall(MatSetType(*outmat, rootType));
4603: PetscCall(MatSeqAIJSetPreallocation(*outmat, 0, dnz));
4604: PetscCall(MatMPIAIJSetPreallocation(*outmat, 0, dnz, 0, onz));
4605: MatPreallocateEnd(dnz, onz);
4606: PetscCall(MatSetOption(*outmat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
4607: }
4609: /* numeric phase */
4610: PetscCall(MatGetOwnershipRange(*outmat, &rstart, NULL));
4611: for (i = 0; i < m; i++) {
4612: PetscCall(MatGetRow_SeqAIJ(inmat, i, &nnz, &indx, &values));
4613: Ii = i + rstart;
4614: PetscCall(MatSetValues(*outmat, 1, &Ii, nnz, indx, values, INSERT_VALUES));
4615: PetscCall(MatRestoreRow_SeqAIJ(inmat, i, &nnz, &indx, &values));
4616: }
4617: PetscCall(MatAssemblyBegin(*outmat, MAT_FINAL_ASSEMBLY));
4618: PetscCall(MatAssemblyEnd(*outmat, MAT_FINAL_ASSEMBLY));
4619: PetscFunctionReturn(PETSC_SUCCESS);
4620: }
4622: static PetscErrorCode MatMergeSeqsToMPIDestroy(void **data)
4623: {
4624: MatMergeSeqsToMPI *merge = (MatMergeSeqsToMPI *)*data;
4626: PetscFunctionBegin;
4627: if (!merge) PetscFunctionReturn(PETSC_SUCCESS);
4628: PetscCall(PetscFree(merge->id_r));
4629: PetscCall(PetscFree(merge->len_s));
4630: PetscCall(PetscFree(merge->len_r));
4631: PetscCall(PetscFree(merge->bi));
4632: PetscCall(PetscFree(merge->bj));
4633: PetscCall(PetscFree(merge->buf_ri[0]));
4634: PetscCall(PetscFree(merge->buf_ri));
4635: PetscCall(PetscFree(merge->buf_rj[0]));
4636: PetscCall(PetscFree(merge->buf_rj));
4637: PetscCall(PetscFree(merge->coi));
4638: PetscCall(PetscFree(merge->coj));
4639: PetscCall(PetscFree(merge->owners_co));
4640: PetscCall(PetscLayoutDestroy(&merge->rowmap));
4641: PetscCall(PetscFree(merge));
4642: PetscFunctionReturn(PETSC_SUCCESS);
4643: }
4645: #include <../src/mat/utils/freespace.h>
4646: #include <petscbt.h>
4648: PetscErrorCode MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat, Mat mpimat)
4649: {
4650: MPI_Comm comm;
4651: Mat_SeqAIJ *a = (Mat_SeqAIJ *)seqmat->data;
4652: PetscMPIInt size, rank, taga, *len_s;
4653: PetscInt N = mpimat->cmap->N, i, j, *owners, *ai = a->i, *aj, m;
4654: PetscMPIInt proc, k;
4655: PetscInt **buf_ri, **buf_rj;
4656: PetscInt anzi, *bj_i, *bi, *bj, arow, bnzi, nextaj;
4657: PetscInt nrows, **buf_ri_k, **nextrow, **nextai;
4658: MPI_Request *s_waits, *r_waits;
4659: MPI_Status *status;
4660: const MatScalar *aa, *a_a;
4661: MatScalar **abuf_r, *ba_i;
4662: MatMergeSeqsToMPI *merge;
4663: PetscContainer container;
4665: PetscFunctionBegin;
4666: PetscCall(PetscObjectGetComm((PetscObject)mpimat, &comm));
4667: PetscCall(PetscLogEventBegin(MAT_Seqstompinum, seqmat, 0, 0, 0));
4669: PetscCallMPI(MPI_Comm_size(comm, &size));
4670: PetscCallMPI(MPI_Comm_rank(comm, &rank));
4672: PetscCall(PetscObjectQuery((PetscObject)mpimat, "MatMergeSeqsToMPI", (PetscObject *)&container));
4673: PetscCheck(container, PetscObjectComm((PetscObject)mpimat), PETSC_ERR_PLIB, "Mat not created from MatCreateMPIAIJSumSeqAIJSymbolic");
4674: PetscCall(PetscContainerGetPointer(container, (void **)&merge));
4675: PetscCall(MatSeqAIJGetArrayRead(seqmat, &a_a));
4676: aa = a_a;
4678: bi = merge->bi;
4679: bj = merge->bj;
4680: buf_ri = merge->buf_ri;
4681: buf_rj = merge->buf_rj;
4683: PetscCall(PetscMalloc1(size, &status));
4684: owners = merge->rowmap->range;
4685: len_s = merge->len_s;
4687: /* send and recv matrix values */
4688: PetscCall(PetscObjectGetNewTag((PetscObject)mpimat, &taga));
4689: PetscCall(PetscPostIrecvScalar(comm, taga, merge->nrecv, merge->id_r, merge->len_r, &abuf_r, &r_waits));
4691: PetscCall(PetscMalloc1(merge->nsend + 1, &s_waits));
4692: for (proc = 0, k = 0; proc < size; proc++) {
4693: if (!len_s[proc]) continue;
4694: i = owners[proc];
4695: PetscCallMPI(MPIU_Isend(aa + ai[i], len_s[proc], MPIU_MATSCALAR, proc, taga, comm, s_waits + k));
4696: k++;
4697: }
4699: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, r_waits, status));
4700: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, s_waits, status));
4701: PetscCall(PetscFree(status));
4703: PetscCall(PetscFree(s_waits));
4704: PetscCall(PetscFree(r_waits));
4706: /* insert mat values of mpimat */
4707: PetscCall(PetscMalloc1(N, &ba_i));
4708: PetscCall(PetscMalloc3(merge->nrecv, &buf_ri_k, merge->nrecv, &nextrow, merge->nrecv, &nextai));
4710: for (k = 0; k < merge->nrecv; k++) {
4711: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4712: nrows = *buf_ri_k[k];
4713: nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */
4714: nextai[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure */
4715: }
4717: /* set values of ba */
4718: m = merge->rowmap->n;
4719: for (i = 0; i < m; i++) {
4720: arow = owners[rank] + i;
4721: bj_i = bj + bi[i]; /* col indices of the i-th row of mpimat */
4722: bnzi = bi[i + 1] - bi[i];
4723: PetscCall(PetscArrayzero(ba_i, bnzi));
4725: /* add local non-zero vals of this proc's seqmat into ba */
4726: anzi = ai[arow + 1] - ai[arow];
4727: aj = a->j + ai[arow];
4728: aa = a_a + ai[arow];
4729: nextaj = 0;
4730: for (j = 0; nextaj < anzi; j++) {
4731: if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4732: ba_i[j] += aa[nextaj++];
4733: }
4734: }
4736: /* add received vals into ba */
4737: for (k = 0; k < merge->nrecv; k++) { /* k-th received message */
4738: /* i-th row */
4739: if (i == *nextrow[k]) {
4740: anzi = *(nextai[k] + 1) - *nextai[k];
4741: aj = buf_rj[k] + *nextai[k];
4742: aa = abuf_r[k] + *nextai[k];
4743: nextaj = 0;
4744: for (j = 0; nextaj < anzi; j++) {
4745: if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4746: ba_i[j] += aa[nextaj++];
4747: }
4748: }
4749: nextrow[k]++;
4750: nextai[k]++;
4751: }
4752: }
4753: PetscCall(MatSetValues(mpimat, 1, &arow, bnzi, bj_i, ba_i, INSERT_VALUES));
4754: }
4755: PetscCall(MatSeqAIJRestoreArrayRead(seqmat, &a_a));
4756: PetscCall(MatAssemblyBegin(mpimat, MAT_FINAL_ASSEMBLY));
4757: PetscCall(MatAssemblyEnd(mpimat, MAT_FINAL_ASSEMBLY));
4759: PetscCall(PetscFree(abuf_r[0]));
4760: PetscCall(PetscFree(abuf_r));
4761: PetscCall(PetscFree(ba_i));
4762: PetscCall(PetscFree3(buf_ri_k, nextrow, nextai));
4763: PetscCall(PetscLogEventEnd(MAT_Seqstompinum, seqmat, 0, 0, 0));
4764: PetscFunctionReturn(PETSC_SUCCESS);
4765: }
4767: PetscErrorCode MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm, Mat seqmat, PetscInt m, PetscInt n, Mat *mpimat)
4768: {
4769: Mat B_mpi;
4770: Mat_SeqAIJ *a = (Mat_SeqAIJ *)seqmat->data;
4771: PetscMPIInt size, rank, tagi, tagj, *len_s, *len_si, *len_ri;
4772: PetscInt **buf_rj, **buf_ri, **buf_ri_k;
4773: PetscInt M = seqmat->rmap->n, N = seqmat->cmap->n, i, *owners, *ai = a->i, *aj = a->j;
4774: PetscInt len, *dnz, *onz, bs, cbs;
4775: PetscInt k, anzi, *bi, *bj, *lnk, nlnk, arow, bnzi;
4776: PetscInt nrows, *buf_s, *buf_si, *buf_si_i, **nextrow, **nextai;
4777: MPI_Request *si_waits, *sj_waits, *ri_waits, *rj_waits;
4778: MPI_Status *status;
4779: PetscFreeSpaceList free_space = NULL, current_space = NULL;
4780: PetscBT lnkbt;
4781: MatMergeSeqsToMPI *merge;
4782: PetscContainer container;
4784: PetscFunctionBegin;
4785: PetscCall(PetscLogEventBegin(MAT_Seqstompisym, seqmat, 0, 0, 0));
4787: /* make sure it is a PETSc comm */
4788: PetscCall(PetscCommDuplicate(comm, &comm, NULL));
4789: PetscCallMPI(MPI_Comm_size(comm, &size));
4790: PetscCallMPI(MPI_Comm_rank(comm, &rank));
4792: PetscCall(PetscNew(&merge));
4793: PetscCall(PetscMalloc1(size, &status));
4795: /* determine row ownership */
4796: PetscCall(PetscLayoutCreate(comm, &merge->rowmap));
4797: PetscCall(PetscLayoutSetLocalSize(merge->rowmap, m));
4798: PetscCall(PetscLayoutSetSize(merge->rowmap, M));
4799: PetscCall(PetscLayoutSetBlockSize(merge->rowmap, 1));
4800: PetscCall(PetscLayoutSetUp(merge->rowmap));
4801: PetscCall(PetscMalloc1(size, &len_si));
4802: PetscCall(PetscMalloc1(size, &merge->len_s));
4804: m = merge->rowmap->n;
4805: owners = merge->rowmap->range;
4807: /* determine the number of messages to send, their lengths */
4808: len_s = merge->len_s;
4810: len = 0; /* length of buf_si[] */
4811: merge->nsend = 0;
4812: for (PetscMPIInt proc = 0; proc < size; proc++) {
4813: len_si[proc] = 0;
4814: if (proc == rank) {
4815: len_s[proc] = 0;
4816: } else {
4817: PetscCall(PetscMPIIntCast(owners[proc + 1] - owners[proc] + 1, &len_si[proc]));
4818: PetscCall(PetscMPIIntCast(ai[owners[proc + 1]] - ai[owners[proc]], &len_s[proc])); /* num of rows to be sent to [proc] */
4819: }
4820: if (len_s[proc]) {
4821: merge->nsend++;
4822: nrows = 0;
4823: for (i = owners[proc]; i < owners[proc + 1]; i++) {
4824: if (ai[i + 1] > ai[i]) nrows++;
4825: }
4826: PetscCall(PetscMPIIntCast(2 * (nrows + 1), &len_si[proc]));
4827: len += len_si[proc];
4828: }
4829: }
4831: /* determine the number and length of messages to receive for ij-structure */
4832: PetscCall(PetscGatherNumberOfMessages(comm, NULL, len_s, &merge->nrecv));
4833: PetscCall(PetscGatherMessageLengths2(comm, merge->nsend, merge->nrecv, len_s, len_si, &merge->id_r, &merge->len_r, &len_ri));
4835: /* post the Irecv of j-structure */
4836: PetscCall(PetscCommGetNewTag(comm, &tagj));
4837: PetscCall(PetscPostIrecvInt(comm, tagj, merge->nrecv, merge->id_r, merge->len_r, &buf_rj, &rj_waits));
4839: /* post the Isend of j-structure */
4840: PetscCall(PetscMalloc2(merge->nsend, &si_waits, merge->nsend, &sj_waits));
4842: for (PetscMPIInt proc = 0, k = 0; proc < size; proc++) {
4843: if (!len_s[proc]) continue;
4844: i = owners[proc];
4845: PetscCallMPI(MPIU_Isend(aj + ai[i], len_s[proc], MPIU_INT, proc, tagj, comm, sj_waits + k));
4846: k++;
4847: }
4849: /* receives and sends of j-structure are complete */
4850: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, rj_waits, status));
4851: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, sj_waits, status));
4853: /* send and recv i-structure */
4854: PetscCall(PetscCommGetNewTag(comm, &tagi));
4855: PetscCall(PetscPostIrecvInt(comm, tagi, merge->nrecv, merge->id_r, len_ri, &buf_ri, &ri_waits));
4857: PetscCall(PetscMalloc1(len + 1, &buf_s));
4858: buf_si = buf_s; /* points to the beginning of k-th msg to be sent */
4859: for (PetscMPIInt proc = 0, k = 0; proc < size; proc++) {
4860: if (!len_s[proc]) continue;
4861: /* form outgoing message for i-structure:
4862: buf_si[0]: nrows to be sent
4863: [1:nrows]: row index (global)
4864: [nrows+1:2*nrows+1]: i-structure index
4865: */
4866: nrows = len_si[proc] / 2 - 1;
4867: buf_si_i = buf_si + nrows + 1;
4868: buf_si[0] = nrows;
4869: buf_si_i[0] = 0;
4870: nrows = 0;
4871: for (i = owners[proc]; i < owners[proc + 1]; i++) {
4872: anzi = ai[i + 1] - ai[i];
4873: if (anzi) {
4874: buf_si_i[nrows + 1] = buf_si_i[nrows] + anzi; /* i-structure */
4875: buf_si[nrows + 1] = i - owners[proc]; /* local row index */
4876: nrows++;
4877: }
4878: }
4879: PetscCallMPI(MPIU_Isend(buf_si, len_si[proc], MPIU_INT, proc, tagi, comm, si_waits + k));
4880: k++;
4881: buf_si += len_si[proc];
4882: }
4884: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, ri_waits, status));
4885: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, si_waits, status));
4887: PetscCall(PetscInfo(seqmat, "nsend: %d, nrecv: %d\n", merge->nsend, merge->nrecv));
4888: 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]));
4890: PetscCall(PetscFree(len_si));
4891: PetscCall(PetscFree(len_ri));
4892: PetscCall(PetscFree(rj_waits));
4893: PetscCall(PetscFree2(si_waits, sj_waits));
4894: PetscCall(PetscFree(ri_waits));
4895: PetscCall(PetscFree(buf_s));
4896: PetscCall(PetscFree(status));
4898: /* compute a local seq matrix in each processor */
4899: /* allocate bi array and free space for accumulating nonzero column info */
4900: PetscCall(PetscMalloc1(m + 1, &bi));
4901: bi[0] = 0;
4903: /* create and initialize a linked list */
4904: nlnk = N + 1;
4905: PetscCall(PetscLLCreate(N, N, nlnk, lnk, lnkbt));
4907: /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */
4908: len = ai[owners[rank + 1]] - ai[owners[rank]];
4909: PetscCall(PetscFreeSpaceGet(PetscIntMultTruncate(2, len) + 1, &free_space));
4911: current_space = free_space;
4913: /* determine symbolic info for each local row */
4914: PetscCall(PetscMalloc3(merge->nrecv, &buf_ri_k, merge->nrecv, &nextrow, merge->nrecv, &nextai));
4916: for (k = 0; k < merge->nrecv; k++) {
4917: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4918: nrows = *buf_ri_k[k];
4919: nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */
4920: nextai[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure */
4921: }
4923: MatPreallocateBegin(comm, m, n, dnz, onz);
4924: len = 0;
4925: for (i = 0; i < m; i++) {
4926: bnzi = 0;
4927: /* add local non-zero cols of this proc's seqmat into lnk */
4928: arow = owners[rank] + i;
4929: anzi = ai[arow + 1] - ai[arow];
4930: aj = a->j + ai[arow];
4931: PetscCall(PetscLLAddSorted(anzi, aj, N, &nlnk, lnk, lnkbt));
4932: bnzi += nlnk;
4933: /* add received col data into lnk */
4934: for (k = 0; k < merge->nrecv; k++) { /* k-th received message */
4935: if (i == *nextrow[k]) { /* i-th row */
4936: anzi = *(nextai[k] + 1) - *nextai[k];
4937: aj = buf_rj[k] + *nextai[k];
4938: PetscCall(PetscLLAddSorted(anzi, aj, N, &nlnk, lnk, lnkbt));
4939: bnzi += nlnk;
4940: nextrow[k]++;
4941: nextai[k]++;
4942: }
4943: }
4944: if (len < bnzi) len = bnzi; /* =max(bnzi) */
4946: /* if free space is not available, make more free space */
4947: if (current_space->local_remaining < bnzi) PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(bnzi, current_space->total_array_size), ¤t_space));
4948: /* copy data into free space, then initialize lnk */
4949: PetscCall(PetscLLClean(N, N, bnzi, lnk, current_space->array, lnkbt));
4950: PetscCall(MatPreallocateSet(i + owners[rank], bnzi, current_space->array, dnz, onz));
4952: current_space->array += bnzi;
4953: current_space->local_used += bnzi;
4954: current_space->local_remaining -= bnzi;
4956: bi[i + 1] = bi[i] + bnzi;
4957: }
4959: PetscCall(PetscFree3(buf_ri_k, nextrow, nextai));
4961: PetscCall(PetscMalloc1(bi[m], &bj));
4962: PetscCall(PetscFreeSpaceContiguous(&free_space, bj));
4963: PetscCall(PetscLLDestroy(lnk, lnkbt));
4965: /* create symbolic parallel matrix B_mpi */
4966: PetscCall(MatGetBlockSizes(seqmat, &bs, &cbs));
4967: PetscCall(MatCreate(comm, &B_mpi));
4968: if (n == PETSC_DECIDE) {
4969: PetscCall(MatSetSizes(B_mpi, m, n, PETSC_DETERMINE, N));
4970: } else {
4971: PetscCall(MatSetSizes(B_mpi, m, n, PETSC_DETERMINE, PETSC_DETERMINE));
4972: }
4973: PetscCall(MatSetBlockSizes(B_mpi, bs, cbs));
4974: PetscCall(MatSetType(B_mpi, MATMPIAIJ));
4975: PetscCall(MatMPIAIJSetPreallocation(B_mpi, 0, dnz, 0, onz));
4976: MatPreallocateEnd(dnz, onz);
4977: PetscCall(MatSetOption(B_mpi, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_FALSE));
4979: /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */
4980: B_mpi->assembled = PETSC_FALSE;
4981: merge->bi = bi;
4982: merge->bj = bj;
4983: merge->buf_ri = buf_ri;
4984: merge->buf_rj = buf_rj;
4985: merge->coi = NULL;
4986: merge->coj = NULL;
4987: merge->owners_co = NULL;
4989: PetscCall(PetscCommDestroy(&comm));
4991: /* attach the supporting struct to B_mpi for reuse */
4992: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
4993: PetscCall(PetscContainerSetPointer(container, merge));
4994: PetscCall(PetscContainerSetCtxDestroy(container, MatMergeSeqsToMPIDestroy));
4995: PetscCall(PetscObjectCompose((PetscObject)B_mpi, "MatMergeSeqsToMPI", (PetscObject)container));
4996: PetscCall(PetscContainerDestroy(&container));
4997: *mpimat = B_mpi;
4999: PetscCall(PetscLogEventEnd(MAT_Seqstompisym, seqmat, 0, 0, 0));
5000: PetscFunctionReturn(PETSC_SUCCESS);
5001: }
5003: /*@
5004: MatCreateMPIAIJSumSeqAIJ - Creates a `MATMPIAIJ` matrix by adding sequential
5005: matrices from each processor
5007: Collective
5009: Input Parameters:
5010: + comm - the communicators the parallel matrix will live on
5011: . seqmat - the input sequential matrices
5012: . m - number of local rows (or `PETSC_DECIDE`)
5013: . n - number of local columns (or `PETSC_DECIDE`)
5014: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5016: Output Parameter:
5017: . mpimat - the parallel matrix generated
5019: Level: advanced
5021: Note:
5022: The dimensions of the sequential matrix in each processor MUST be the same.
5023: The input seqmat is included into the container `MatMergeSeqsToMPIDestroy`, and will be
5024: destroyed when `mpimat` is destroyed. Call `PetscObjectQuery()` to access `seqmat`.
5026: .seealso: [](ch_matrices), `Mat`, `MatCreateAIJ()`
5027: @*/
5028: PetscErrorCode MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm, Mat seqmat, PetscInt m, PetscInt n, MatReuse scall, Mat *mpimat)
5029: {
5030: PetscMPIInt size;
5032: PetscFunctionBegin;
5033: PetscCallMPI(MPI_Comm_size(comm, &size));
5034: if (size == 1) {
5035: PetscCall(PetscLogEventBegin(MAT_Seqstompi, seqmat, 0, 0, 0));
5036: if (scall == MAT_INITIAL_MATRIX) {
5037: PetscCall(MatDuplicate(seqmat, MAT_COPY_VALUES, mpimat));
5038: } else {
5039: PetscCall(MatCopy(seqmat, *mpimat, SAME_NONZERO_PATTERN));
5040: }
5041: PetscCall(PetscLogEventEnd(MAT_Seqstompi, seqmat, 0, 0, 0));
5042: PetscFunctionReturn(PETSC_SUCCESS);
5043: }
5044: PetscCall(PetscLogEventBegin(MAT_Seqstompi, seqmat, 0, 0, 0));
5045: if (scall == MAT_INITIAL_MATRIX) PetscCall(MatCreateMPIAIJSumSeqAIJSymbolic(comm, seqmat, m, n, mpimat));
5046: PetscCall(MatCreateMPIAIJSumSeqAIJNumeric(seqmat, *mpimat));
5047: PetscCall(PetscLogEventEnd(MAT_Seqstompi, seqmat, 0, 0, 0));
5048: PetscFunctionReturn(PETSC_SUCCESS);
5049: }
5051: /*@
5052: MatAIJGetLocalMat - Creates a `MATSEQAIJ` from a `MATAIJ` matrix.
5054: Not Collective
5056: Input Parameter:
5057: . A - the matrix
5059: Output Parameter:
5060: . A_loc - the local sequential matrix generated
5062: Level: developer
5064: Notes:
5065: The matrix is created by taking `A`'s local rows and putting them into a sequential matrix
5066: with `mlocal` rows and `n` columns. Where `mlocal` is obtained with `MatGetLocalSize()` and
5067: `n` is the global column count obtained with `MatGetSize()`
5069: In other words combines the two parts of a parallel `MATMPIAIJ` matrix on each process to a single matrix.
5071: For parallel matrices this creates an entirely new matrix. If the matrix is sequential it merely increases the reference count.
5073: Destroy the matrix with `MatDestroy()`
5075: .seealso: [](ch_matrices), `Mat`, `MatMPIAIJGetLocalMat()`
5076: @*/
5077: PetscErrorCode MatAIJGetLocalMat(Mat A, Mat *A_loc)
5078: {
5079: PetscBool mpi;
5081: PetscFunctionBegin;
5082: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIAIJ, &mpi));
5083: if (mpi) {
5084: PetscCall(MatMPIAIJGetLocalMat(A, MAT_INITIAL_MATRIX, A_loc));
5085: } else {
5086: *A_loc = A;
5087: PetscCall(PetscObjectReference((PetscObject)*A_loc));
5088: }
5089: PetscFunctionReturn(PETSC_SUCCESS);
5090: }
5092: /*@
5093: MatMPIAIJGetLocalMat - Creates a `MATSEQAIJ` from a `MATMPIAIJ` matrix.
5095: Not Collective
5097: Input Parameters:
5098: + A - the matrix
5099: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5101: Output Parameter:
5102: . A_loc - the local sequential matrix generated
5104: Level: developer
5106: Notes:
5107: The matrix is created by taking all `A`'s local rows and putting them into a sequential
5108: matrix with `mlocal` rows and `n` columns.`mlocal` is the row count obtained with
5109: `MatGetLocalSize()` and `n` is the global column count obtained with `MatGetSize()`.
5111: In other words combines the two parts of a parallel `MATMPIAIJ` matrix on each process to a single matrix.
5113: When `A` is sequential and `MAT_INITIAL_MATRIX` is requested, the matrix returned is the diagonal part of `A` (which contains the entire matrix),
5114: 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
5115: then `MatCopy`(Adiag,*`A_loc`,`SAME_NONZERO_PATTERN`) is called to fill `A_loc`. Thus one can preallocate the appropriate sequential matrix `A_loc`
5116: 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.
5118: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMatCondensed()`, `MatMPIAIJGetLocalMatMerge()`
5119: @*/
5120: PetscErrorCode MatMPIAIJGetLocalMat(Mat A, MatReuse scall, Mat *A_loc)
5121: {
5122: Mat_MPIAIJ *mpimat = (Mat_MPIAIJ *)A->data;
5123: Mat_SeqAIJ *mat, *a, *b;
5124: PetscInt *ai, *aj, *bi, *bj, *cmap = mpimat->garray;
5125: const PetscScalar *aa, *ba, *aav, *bav;
5126: PetscScalar *ca, *cam;
5127: PetscMPIInt size;
5128: PetscInt am = A->rmap->n, i, j, k, cstart = A->cmap->rstart;
5129: PetscInt *ci, *cj, col, ncols_d, ncols_o, jo;
5130: PetscBool match;
5132: PetscFunctionBegin;
5133: PetscCall(PetscStrbeginswith(((PetscObject)A)->type_name, MATMPIAIJ, &match));
5134: PetscCheck(match, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Requires MATMPIAIJ matrix as input");
5135: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
5136: if (size == 1) {
5137: if (scall == MAT_INITIAL_MATRIX) {
5138: PetscCall(PetscObjectReference((PetscObject)mpimat->A));
5139: *A_loc = mpimat->A;
5140: } else if (scall == MAT_REUSE_MATRIX) {
5141: PetscCall(MatCopy(mpimat->A, *A_loc, SAME_NONZERO_PATTERN));
5142: }
5143: PetscFunctionReturn(PETSC_SUCCESS);
5144: }
5146: PetscCall(PetscLogEventBegin(MAT_Getlocalmat, A, 0, 0, 0));
5147: a = (Mat_SeqAIJ *)mpimat->A->data;
5148: b = (Mat_SeqAIJ *)mpimat->B->data;
5149: ai = a->i;
5150: aj = a->j;
5151: bi = b->i;
5152: bj = b->j;
5153: PetscCall(MatSeqAIJGetArrayRead(mpimat->A, &aav));
5154: PetscCall(MatSeqAIJGetArrayRead(mpimat->B, &bav));
5155: aa = aav;
5156: ba = bav;
5157: if (scall == MAT_INITIAL_MATRIX) {
5158: PetscCall(PetscMalloc1(1 + am, &ci));
5159: ci[0] = 0;
5160: for (i = 0; i < am; i++) ci[i + 1] = ci[i] + (ai[i + 1] - ai[i]) + (bi[i + 1] - bi[i]);
5161: PetscCall(PetscMalloc1(1 + ci[am], &cj));
5162: PetscCall(PetscMalloc1(1 + ci[am], &ca));
5163: k = 0;
5164: for (i = 0; i < am; i++) {
5165: ncols_o = bi[i + 1] - bi[i];
5166: ncols_d = ai[i + 1] - ai[i];
5167: /* off-diagonal portion of A */
5168: for (jo = 0; jo < ncols_o; jo++) {
5169: col = cmap[*bj];
5170: if (col >= cstart) break;
5171: cj[k] = col;
5172: bj++;
5173: ca[k++] = *ba++;
5174: }
5175: /* diagonal portion of A */
5176: for (j = 0; j < ncols_d; j++) {
5177: cj[k] = cstart + *aj++;
5178: ca[k++] = *aa++;
5179: }
5180: /* off-diagonal portion of A */
5181: for (j = jo; j < ncols_o; j++) {
5182: cj[k] = cmap[*bj++];
5183: ca[k++] = *ba++;
5184: }
5185: }
5186: /* put together the new matrix */
5187: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, am, A->cmap->N, ci, cj, ca, A_loc));
5188: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5189: /* Since these are PETSc arrays, change flags to free them as necessary. */
5190: mat = (Mat_SeqAIJ *)(*A_loc)->data;
5191: mat->free_a = PETSC_TRUE;
5192: mat->free_ij = PETSC_TRUE;
5193: mat->nonew = 0;
5194: } else if (scall == MAT_REUSE_MATRIX) {
5195: mat = (Mat_SeqAIJ *)(*A_loc)->data;
5196: ci = mat->i;
5197: cj = mat->j;
5198: PetscCall(MatSeqAIJGetArrayWrite(*A_loc, &cam));
5199: for (i = 0; i < am; i++) {
5200: /* off-diagonal portion of A */
5201: ncols_o = bi[i + 1] - bi[i];
5202: for (jo = 0; jo < ncols_o; jo++) {
5203: col = cmap[*bj];
5204: if (col >= cstart) break;
5205: *cam++ = *ba++;
5206: bj++;
5207: }
5208: /* diagonal portion of A */
5209: ncols_d = ai[i + 1] - ai[i];
5210: for (j = 0; j < ncols_d; j++) *cam++ = *aa++;
5211: /* off-diagonal portion of A */
5212: for (j = jo; j < ncols_o; j++) {
5213: *cam++ = *ba++;
5214: bj++;
5215: }
5216: }
5217: PetscCall(MatSeqAIJRestoreArrayWrite(*A_loc, &cam));
5218: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Invalid MatReuse %d", (int)scall);
5219: PetscCall(MatSeqAIJRestoreArrayRead(mpimat->A, &aav));
5220: PetscCall(MatSeqAIJRestoreArrayRead(mpimat->B, &bav));
5221: PetscCall(PetscLogEventEnd(MAT_Getlocalmat, A, 0, 0, 0));
5222: PetscFunctionReturn(PETSC_SUCCESS);
5223: }
5225: /*@
5226: MatMPIAIJGetLocalMatMerge - Creates a `MATSEQAIJ` from a `MATMPIAIJ` matrix by taking all its local rows and putting them into a sequential matrix with
5227: mlocal rows and n columns. Where n is the sum of the number of columns of the diagonal and off-diagonal part
5229: Not Collective
5231: Input Parameters:
5232: + A - the matrix
5233: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5235: Output Parameters:
5236: + glob - sequential `IS` with global indices associated with the columns of the local sequential matrix generated (can be `NULL`)
5237: - A_loc - the local sequential matrix generated
5239: Level: developer
5241: Note:
5242: This is different from `MatMPIAIJGetLocalMat()` since the first columns in the returning matrix are those associated with the diagonal
5243: part, then those associated with the off-diagonal part (in its local ordering)
5245: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMat()`, `MatMPIAIJGetLocalMatCondensed()`
5246: @*/
5247: PetscErrorCode MatMPIAIJGetLocalMatMerge(Mat A, MatReuse scall, IS *glob, Mat *A_loc)
5248: {
5249: Mat Ao, Ad;
5250: const PetscInt *cmap;
5251: PetscMPIInt size;
5252: PetscErrorCode (*f)(Mat, MatReuse, IS *, Mat *);
5254: PetscFunctionBegin;
5255: PetscCall(MatMPIAIJGetSeqAIJ(A, &Ad, &Ao, &cmap));
5256: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
5257: if (size == 1) {
5258: if (scall == MAT_INITIAL_MATRIX) {
5259: PetscCall(PetscObjectReference((PetscObject)Ad));
5260: *A_loc = Ad;
5261: } else if (scall == MAT_REUSE_MATRIX) {
5262: PetscCall(MatCopy(Ad, *A_loc, SAME_NONZERO_PATTERN));
5263: }
5264: if (glob) PetscCall(ISCreateStride(PetscObjectComm((PetscObject)Ad), Ad->cmap->n, Ad->cmap->rstart, 1, glob));
5265: PetscFunctionReturn(PETSC_SUCCESS);
5266: }
5267: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatMPIAIJGetLocalMatMerge_C", &f));
5268: PetscCall(PetscLogEventBegin(MAT_Getlocalmat, A, 0, 0, 0));
5269: if (f) {
5270: PetscCall((*f)(A, scall, glob, A_loc));
5271: } else {
5272: Mat_SeqAIJ *a = (Mat_SeqAIJ *)Ad->data;
5273: Mat_SeqAIJ *b = (Mat_SeqAIJ *)Ao->data;
5274: Mat_SeqAIJ *c;
5275: PetscInt *ai = a->i, *aj = a->j;
5276: PetscInt *bi = b->i, *bj = b->j;
5277: PetscInt *ci, *cj;
5278: const PetscScalar *aa, *ba;
5279: PetscScalar *ca;
5280: PetscInt i, j, am, dn, on;
5282: PetscCall(MatGetLocalSize(Ad, &am, &dn));
5283: PetscCall(MatGetLocalSize(Ao, NULL, &on));
5284: PetscCall(MatSeqAIJGetArrayRead(Ad, &aa));
5285: PetscCall(MatSeqAIJGetArrayRead(Ao, &ba));
5286: if (scall == MAT_INITIAL_MATRIX) {
5287: PetscInt k;
5288: PetscCall(PetscMalloc1(1 + am, &ci));
5289: PetscCall(PetscMalloc1(ai[am] + bi[am], &cj));
5290: PetscCall(PetscMalloc1(ai[am] + bi[am], &ca));
5291: ci[0] = 0;
5292: for (i = 0, k = 0; i < am; i++) {
5293: const PetscInt ncols_o = bi[i + 1] - bi[i];
5294: const PetscInt ncols_d = ai[i + 1] - ai[i];
5295: ci[i + 1] = ci[i] + ncols_o + ncols_d;
5296: /* diagonal portion of A */
5297: for (j = 0; j < ncols_d; j++, k++) {
5298: cj[k] = *aj++;
5299: ca[k] = *aa++;
5300: }
5301: /* off-diagonal portion of A */
5302: for (j = 0; j < ncols_o; j++, k++) {
5303: cj[k] = dn + *bj++;
5304: ca[k] = *ba++;
5305: }
5306: }
5307: /* put together the new matrix */
5308: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, am, dn + on, ci, cj, ca, A_loc));
5309: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5310: /* Since these are PETSc arrays, change flags to free them as necessary. */
5311: c = (Mat_SeqAIJ *)(*A_loc)->data;
5312: c->free_a = PETSC_TRUE;
5313: c->free_ij = PETSC_TRUE;
5314: c->nonew = 0;
5315: PetscCall(MatSetType(*A_loc, ((PetscObject)Ad)->type_name));
5316: } else if (scall == MAT_REUSE_MATRIX) {
5317: PetscCall(MatSeqAIJGetArrayWrite(*A_loc, &ca));
5318: for (i = 0; i < am; i++) {
5319: const PetscInt ncols_d = ai[i + 1] - ai[i];
5320: const PetscInt ncols_o = bi[i + 1] - bi[i];
5321: /* diagonal portion of A */
5322: for (j = 0; j < ncols_d; j++) *ca++ = *aa++;
5323: /* off-diagonal portion of A */
5324: for (j = 0; j < ncols_o; j++) *ca++ = *ba++;
5325: }
5326: PetscCall(MatSeqAIJRestoreArrayWrite(*A_loc, &ca));
5327: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Invalid MatReuse %d", (int)scall);
5328: PetscCall(MatSeqAIJRestoreArrayRead(Ad, &aa));
5329: PetscCall(MatSeqAIJRestoreArrayRead(Ao, &aa));
5330: if (glob) {
5331: PetscInt cst, *gidx;
5333: PetscCall(MatGetOwnershipRangeColumn(A, &cst, NULL));
5334: PetscCall(PetscMalloc1(dn + on, &gidx));
5335: for (i = 0; i < dn; i++) gidx[i] = cst + i;
5336: for (i = 0; i < on; i++) gidx[i + dn] = cmap[i];
5337: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)Ad), dn + on, gidx, PETSC_OWN_POINTER, glob));
5338: }
5339: }
5340: PetscCall(PetscLogEventEnd(MAT_Getlocalmat, A, 0, 0, 0));
5341: PetscFunctionReturn(PETSC_SUCCESS);
5342: }
5344: /*@C
5345: MatMPIAIJGetLocalMatCondensed - Creates a `MATSEQAIJ` matrix from an `MATMPIAIJ` matrix by taking all its local rows and NON-ZERO columns
5347: Not Collective
5349: Input Parameters:
5350: + A - the matrix
5351: . scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5352: . row - index set of rows to extract (or `NULL`)
5353: - col - index set of columns to extract (or `NULL`)
5355: Output Parameter:
5356: . A_loc - the local sequential matrix generated
5358: Level: developer
5360: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMat()`
5361: @*/
5362: PetscErrorCode MatMPIAIJGetLocalMatCondensed(Mat A, MatReuse scall, IS *row, IS *col, Mat *A_loc)
5363: {
5364: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5365: PetscInt i, start, end, ncols, nzA, nzB, *cmap, imark, *idx;
5366: IS isrowa, iscola;
5367: Mat *aloc;
5368: PetscBool match;
5370: PetscFunctionBegin;
5371: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIAIJ, &match));
5372: PetscCheck(match, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Requires MATMPIAIJ matrix as input");
5373: PetscCall(PetscLogEventBegin(MAT_Getlocalmatcondensed, A, 0, 0, 0));
5374: if (!row) {
5375: start = A->rmap->rstart;
5376: end = A->rmap->rend;
5377: PetscCall(ISCreateStride(PETSC_COMM_SELF, end - start, start, 1, &isrowa));
5378: } else {
5379: isrowa = *row;
5380: }
5381: if (!col) {
5382: start = A->cmap->rstart;
5383: cmap = a->garray;
5384: nzA = a->A->cmap->n;
5385: nzB = a->B->cmap->n;
5386: PetscCall(PetscMalloc1(nzA + nzB, &idx));
5387: ncols = 0;
5388: for (i = 0; i < nzB; i++) {
5389: if (cmap[i] < start) idx[ncols++] = cmap[i];
5390: else break;
5391: }
5392: imark = i;
5393: for (i = 0; i < nzA; i++) idx[ncols++] = start + i;
5394: for (i = imark; i < nzB; i++) idx[ncols++] = cmap[i];
5395: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, &iscola));
5396: } else {
5397: iscola = *col;
5398: }
5399: if (scall != MAT_INITIAL_MATRIX) {
5400: PetscCall(PetscMalloc1(1, &aloc));
5401: aloc[0] = *A_loc;
5402: }
5403: PetscCall(MatCreateSubMatrices(A, 1, &isrowa, &iscola, scall, &aloc));
5404: if (!col) { /* attach global id of condensed columns */
5405: PetscCall(PetscObjectCompose((PetscObject)aloc[0], "_petsc_GetLocalMatCondensed_iscol", (PetscObject)iscola));
5406: }
5407: *A_loc = aloc[0];
5408: PetscCall(PetscFree(aloc));
5409: if (!row) PetscCall(ISDestroy(&isrowa));
5410: if (!col) PetscCall(ISDestroy(&iscola));
5411: PetscCall(PetscLogEventEnd(MAT_Getlocalmatcondensed, A, 0, 0, 0));
5412: PetscFunctionReturn(PETSC_SUCCESS);
5413: }
5415: /*
5416: * Create a sequential AIJ matrix based on row indices. a whole column is extracted once a row is matched.
5417: * Row could be local or remote.The routine is designed to be scalable in memory so that nothing is based
5418: * on a global size.
5419: * */
5420: static PetscErrorCode MatCreateSeqSubMatrixWithRows_Private(Mat P, IS rows, Mat *P_oth)
5421: {
5422: Mat_MPIAIJ *p = (Mat_MPIAIJ *)P->data;
5423: Mat_SeqAIJ *pd = (Mat_SeqAIJ *)p->A->data, *po = (Mat_SeqAIJ *)p->B->data, *p_oth;
5424: PetscInt plocalsize, nrows, *ilocal, *oilocal, i, lidx, *nrcols, *nlcols, ncol;
5425: PetscMPIInt owner;
5426: PetscSFNode *iremote, *oiremote;
5427: const PetscInt *lrowindices;
5428: PetscSF sf, osf;
5429: PetscInt pcstart, *roffsets, *loffsets, *pnnz, j;
5430: PetscInt ontotalcols, dntotalcols, ntotalcols, nout;
5431: MPI_Comm comm;
5432: ISLocalToGlobalMapping mapping;
5433: const PetscScalar *pd_a, *po_a;
5435: PetscFunctionBegin;
5436: PetscCall(PetscObjectGetComm((PetscObject)P, &comm));
5437: /* plocalsize is the number of roots
5438: * nrows is the number of leaves
5439: * */
5440: PetscCall(MatGetLocalSize(P, &plocalsize, NULL));
5441: PetscCall(ISGetLocalSize(rows, &nrows));
5442: PetscCall(PetscCalloc1(nrows, &iremote));
5443: PetscCall(ISGetIndices(rows, &lrowindices));
5444: for (i = 0; i < nrows; i++) {
5445: /* Find a remote index and an owner for a row
5446: * The row could be local or remote
5447: * */
5448: owner = 0;
5449: lidx = 0;
5450: PetscCall(PetscLayoutFindOwnerIndex(P->rmap, lrowindices[i], &owner, &lidx));
5451: iremote[i].index = lidx;
5452: iremote[i].rank = owner;
5453: }
5454: /* Create SF to communicate how many nonzero columns for each row */
5455: PetscCall(PetscSFCreate(comm, &sf));
5456: /* SF will figure out the number of nonzero columns for each row, and their
5457: * offsets
5458: * */
5459: PetscCall(PetscSFSetGraph(sf, plocalsize, nrows, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
5460: PetscCall(PetscSFSetFromOptions(sf));
5461: PetscCall(PetscSFSetUp(sf));
5463: PetscCall(PetscCalloc1(2 * (plocalsize + 1), &roffsets));
5464: PetscCall(PetscCalloc1(2 * plocalsize, &nrcols));
5465: PetscCall(PetscCalloc1(nrows, &pnnz));
5466: roffsets[0] = 0;
5467: roffsets[1] = 0;
5468: for (i = 0; i < plocalsize; i++) {
5469: /* diagonal */
5470: nrcols[i * 2 + 0] = pd->i[i + 1] - pd->i[i];
5471: /* off-diagonal */
5472: nrcols[i * 2 + 1] = po->i[i + 1] - po->i[i];
5473: /* compute offsets so that we relative location for each row */
5474: roffsets[(i + 1) * 2 + 0] = roffsets[i * 2 + 0] + nrcols[i * 2 + 0];
5475: roffsets[(i + 1) * 2 + 1] = roffsets[i * 2 + 1] + nrcols[i * 2 + 1];
5476: }
5477: PetscCall(PetscCalloc1(2 * nrows, &nlcols));
5478: PetscCall(PetscCalloc1(2 * nrows, &loffsets));
5479: /* 'r' means root, and 'l' means leaf */
5480: PetscCall(PetscSFBcastBegin(sf, MPIU_2INT, nrcols, nlcols, MPI_REPLACE));
5481: PetscCall(PetscSFBcastBegin(sf, MPIU_2INT, roffsets, loffsets, MPI_REPLACE));
5482: PetscCall(PetscSFBcastEnd(sf, MPIU_2INT, nrcols, nlcols, MPI_REPLACE));
5483: PetscCall(PetscSFBcastEnd(sf, MPIU_2INT, roffsets, loffsets, MPI_REPLACE));
5484: PetscCall(PetscSFDestroy(&sf));
5485: PetscCall(PetscFree(roffsets));
5486: PetscCall(PetscFree(nrcols));
5487: dntotalcols = 0;
5488: ontotalcols = 0;
5489: ncol = 0;
5490: for (i = 0; i < nrows; i++) {
5491: pnnz[i] = nlcols[i * 2 + 0] + nlcols[i * 2 + 1];
5492: ncol = PetscMax(pnnz[i], ncol);
5493: /* diagonal */
5494: dntotalcols += nlcols[i * 2 + 0];
5495: /* off-diagonal */
5496: ontotalcols += nlcols[i * 2 + 1];
5497: }
5498: /* We do not need to figure the right number of columns
5499: * since all the calculations will be done by going through the raw data
5500: * */
5501: PetscCall(MatCreateSeqAIJ(PETSC_COMM_SELF, nrows, ncol, 0, pnnz, P_oth));
5502: PetscCall(MatSetUp(*P_oth));
5503: PetscCall(PetscFree(pnnz));
5504: p_oth = (Mat_SeqAIJ *)(*P_oth)->data;
5505: /* diagonal */
5506: PetscCall(PetscCalloc1(dntotalcols, &iremote));
5507: /* off-diagonal */
5508: PetscCall(PetscCalloc1(ontotalcols, &oiremote));
5509: /* diagonal */
5510: PetscCall(PetscCalloc1(dntotalcols, &ilocal));
5511: /* off-diagonal */
5512: PetscCall(PetscCalloc1(ontotalcols, &oilocal));
5513: dntotalcols = 0;
5514: ontotalcols = 0;
5515: ntotalcols = 0;
5516: for (i = 0; i < nrows; i++) {
5517: owner = 0;
5518: PetscCall(PetscLayoutFindOwnerIndex(P->rmap, lrowindices[i], &owner, NULL));
5519: /* Set iremote for diag matrix */
5520: for (j = 0; j < nlcols[i * 2 + 0]; j++) {
5521: iremote[dntotalcols].index = loffsets[i * 2 + 0] + j;
5522: iremote[dntotalcols].rank = owner;
5523: /* P_oth is seqAIJ so that ilocal need to point to the first part of memory */
5524: ilocal[dntotalcols++] = ntotalcols++;
5525: }
5526: /* off-diagonal */
5527: for (j = 0; j < nlcols[i * 2 + 1]; j++) {
5528: oiremote[ontotalcols].index = loffsets[i * 2 + 1] + j;
5529: oiremote[ontotalcols].rank = owner;
5530: oilocal[ontotalcols++] = ntotalcols++;
5531: }
5532: }
5533: PetscCall(ISRestoreIndices(rows, &lrowindices));
5534: PetscCall(PetscFree(loffsets));
5535: PetscCall(PetscFree(nlcols));
5536: PetscCall(PetscSFCreate(comm, &sf));
5537: /* P serves as roots and P_oth is leaves
5538: * Diag matrix
5539: * */
5540: PetscCall(PetscSFSetGraph(sf, pd->i[plocalsize], dntotalcols, ilocal, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
5541: PetscCall(PetscSFSetFromOptions(sf));
5542: PetscCall(PetscSFSetUp(sf));
5544: PetscCall(PetscSFCreate(comm, &osf));
5545: /* off-diagonal */
5546: PetscCall(PetscSFSetGraph(osf, po->i[plocalsize], ontotalcols, oilocal, PETSC_OWN_POINTER, oiremote, PETSC_OWN_POINTER));
5547: PetscCall(PetscSFSetFromOptions(osf));
5548: PetscCall(PetscSFSetUp(osf));
5549: PetscCall(MatSeqAIJGetArrayRead(p->A, &pd_a));
5550: PetscCall(MatSeqAIJGetArrayRead(p->B, &po_a));
5551: /* operate on the matrix internal data to save memory */
5552: PetscCall(PetscSFBcastBegin(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5553: PetscCall(PetscSFBcastBegin(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5554: PetscCall(MatGetOwnershipRangeColumn(P, &pcstart, NULL));
5555: /* Convert to global indices for diag matrix */
5556: for (i = 0; i < pd->i[plocalsize]; i++) pd->j[i] += pcstart;
5557: PetscCall(PetscSFBcastBegin(sf, MPIU_INT, pd->j, p_oth->j, MPI_REPLACE));
5558: /* We want P_oth store global indices */
5559: PetscCall(ISLocalToGlobalMappingCreate(comm, 1, p->B->cmap->n, p->garray, PETSC_COPY_VALUES, &mapping));
5560: /* Use memory scalable approach */
5561: PetscCall(ISLocalToGlobalMappingSetType(mapping, ISLOCALTOGLOBALMAPPINGHASH));
5562: PetscCall(ISLocalToGlobalMappingApply(mapping, po->i[plocalsize], po->j, po->j));
5563: PetscCall(PetscSFBcastBegin(osf, MPIU_INT, po->j, p_oth->j, MPI_REPLACE));
5564: PetscCall(PetscSFBcastEnd(sf, MPIU_INT, pd->j, p_oth->j, MPI_REPLACE));
5565: /* Convert back to local indices */
5566: for (i = 0; i < pd->i[plocalsize]; i++) pd->j[i] -= pcstart;
5567: PetscCall(PetscSFBcastEnd(osf, MPIU_INT, po->j, p_oth->j, MPI_REPLACE));
5568: nout = 0;
5569: PetscCall(ISGlobalToLocalMappingApply(mapping, IS_GTOLM_DROP, po->i[plocalsize], po->j, &nout, po->j));
5570: PetscCheck(nout == po->i[plocalsize], comm, PETSC_ERR_ARG_INCOMP, "n %" PetscInt_FMT " does not equal to nout %" PetscInt_FMT " ", po->i[plocalsize], nout);
5571: PetscCall(ISLocalToGlobalMappingDestroy(&mapping));
5572: /* Exchange values */
5573: PetscCall(PetscSFBcastEnd(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5574: PetscCall(PetscSFBcastEnd(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5575: PetscCall(MatSeqAIJRestoreArrayRead(p->A, &pd_a));
5576: PetscCall(MatSeqAIJRestoreArrayRead(p->B, &po_a));
5577: /* Stop PETSc from shrinking memory */
5578: for (i = 0; i < nrows; i++) p_oth->ilen[i] = p_oth->imax[i];
5579: PetscCall(MatAssemblyBegin(*P_oth, MAT_FINAL_ASSEMBLY));
5580: PetscCall(MatAssemblyEnd(*P_oth, MAT_FINAL_ASSEMBLY));
5581: /* Attach PetscSF objects to P_oth so that we can reuse it later */
5582: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "diagsf", (PetscObject)sf));
5583: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "offdiagsf", (PetscObject)osf));
5584: PetscCall(PetscSFDestroy(&sf));
5585: PetscCall(PetscSFDestroy(&osf));
5586: PetscFunctionReturn(PETSC_SUCCESS);
5587: }
5589: /*
5590: * Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A
5591: * This supports MPIAIJ and MAIJ
5592: * */
5593: PetscErrorCode MatGetBrowsOfAcols_MPIXAIJ(Mat A, Mat P, PetscInt dof, MatReuse reuse, Mat *P_oth)
5594: {
5595: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data, *p = (Mat_MPIAIJ *)P->data;
5596: Mat_SeqAIJ *p_oth;
5597: IS rows, map;
5598: PetscHMapI hamp;
5599: PetscInt i, htsize, *rowindices, off, *mapping, key, count;
5600: MPI_Comm comm;
5601: PetscSF sf, osf;
5602: PetscBool has;
5604: PetscFunctionBegin;
5605: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
5606: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAocols, A, P, 0, 0));
5607: /* If it is the first time, create an index set of off-diag nonzero columns of A,
5608: * and then create a submatrix (that often is an overlapping matrix)
5609: * */
5610: if (reuse == MAT_INITIAL_MATRIX) {
5611: /* Use a hash table to figure out unique keys */
5612: PetscCall(PetscHMapICreateWithSize(a->B->cmap->n, &hamp));
5613: PetscCall(PetscCalloc1(a->B->cmap->n, &mapping));
5614: count = 0;
5615: /* Assume that a->g is sorted, otherwise the following does not make sense */
5616: for (i = 0; i < a->B->cmap->n; i++) {
5617: key = a->garray[i] / dof;
5618: PetscCall(PetscHMapIHas(hamp, key, &has));
5619: if (!has) {
5620: mapping[i] = count;
5621: PetscCall(PetscHMapISet(hamp, key, count++));
5622: } else {
5623: /* Current 'i' has the same value the previous step */
5624: mapping[i] = count - 1;
5625: }
5626: }
5627: PetscCall(ISCreateGeneral(comm, a->B->cmap->n, mapping, PETSC_OWN_POINTER, &map));
5628: PetscCall(PetscHMapIGetSize(hamp, &htsize));
5629: PetscCheck(htsize == count, comm, PETSC_ERR_ARG_INCOMP, " Size of hash map %" PetscInt_FMT " is inconsistent with count %" PetscInt_FMT, htsize, count);
5630: PetscCall(PetscCalloc1(htsize, &rowindices));
5631: off = 0;
5632: PetscCall(PetscHMapIGetKeys(hamp, &off, rowindices));
5633: PetscCall(PetscHMapIDestroy(&hamp));
5634: PetscCall(PetscSortInt(htsize, rowindices));
5635: PetscCall(ISCreateGeneral(comm, htsize, rowindices, PETSC_OWN_POINTER, &rows));
5636: /* In case, the matrix was already created but users want to recreate the matrix */
5637: PetscCall(MatDestroy(P_oth));
5638: PetscCall(MatCreateSeqSubMatrixWithRows_Private(P, rows, P_oth));
5639: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "aoffdiagtopothmapping", (PetscObject)map));
5640: PetscCall(ISDestroy(&map));
5641: PetscCall(ISDestroy(&rows));
5642: } else if (reuse == MAT_REUSE_MATRIX) {
5643: /* If matrix was already created, we simply update values using SF objects
5644: * that as attached to the matrix earlier.
5645: */
5646: const PetscScalar *pd_a, *po_a;
5648: PetscCall(PetscObjectQuery((PetscObject)*P_oth, "diagsf", (PetscObject *)&sf));
5649: PetscCall(PetscObjectQuery((PetscObject)*P_oth, "offdiagsf", (PetscObject *)&osf));
5650: PetscCheck(sf && osf, comm, PETSC_ERR_ARG_NULL, "Matrix is not initialized yet");
5651: p_oth = (Mat_SeqAIJ *)(*P_oth)->data;
5652: /* Update values in place */
5653: PetscCall(MatSeqAIJGetArrayRead(p->A, &pd_a));
5654: PetscCall(MatSeqAIJGetArrayRead(p->B, &po_a));
5655: PetscCall(PetscSFBcastBegin(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5656: PetscCall(PetscSFBcastBegin(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5657: PetscCall(PetscSFBcastEnd(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5658: PetscCall(PetscSFBcastEnd(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5659: PetscCall(MatSeqAIJRestoreArrayRead(p->A, &pd_a));
5660: PetscCall(MatSeqAIJRestoreArrayRead(p->B, &po_a));
5661: } else SETERRQ(comm, PETSC_ERR_ARG_UNKNOWN_TYPE, "Unknown reuse type");
5662: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAocols, A, P, 0, 0));
5663: PetscFunctionReturn(PETSC_SUCCESS);
5664: }
5666: /*@C
5667: MatGetBrowsOfAcols - Returns `IS` that contain rows of `B` that equal to nonzero columns of local `A`
5669: Collective
5671: Input Parameters:
5672: + A - the first matrix in `MATMPIAIJ` format
5673: . B - the second matrix in `MATMPIAIJ` format
5674: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5676: Output Parameters:
5677: + rowb - On input index sets of rows of B to extract (or `NULL`), modified on output
5678: . colb - On input index sets of columns of B to extract (or `NULL`), modified on output
5679: - B_seq - the sequential matrix generated
5681: Level: developer
5683: .seealso: `Mat`, `MATMPIAIJ`, `IS`, `MatReuse`
5684: @*/
5685: PetscErrorCode MatGetBrowsOfAcols(Mat A, Mat B, MatReuse scall, IS *rowb, IS *colb, Mat *B_seq)
5686: {
5687: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5688: PetscInt *idx, i, start, ncols, nzA, nzB, *cmap, imark;
5689: IS isrowb, iscolb;
5690: Mat *bseq = NULL;
5692: PetscFunctionBegin;
5693: 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 ")",
5694: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
5695: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAcols, A, B, 0, 0));
5697: if (scall == MAT_INITIAL_MATRIX) {
5698: start = A->cmap->rstart;
5699: cmap = a->garray;
5700: nzA = a->A->cmap->n;
5701: nzB = a->B->cmap->n;
5702: PetscCall(PetscMalloc1(nzA + nzB, &idx));
5703: ncols = 0;
5704: for (i = 0; i < nzB; i++) { /* row < local row index */
5705: if (cmap[i] < start) idx[ncols++] = cmap[i];
5706: else break;
5707: }
5708: imark = i;
5709: for (i = 0; i < nzA; i++) idx[ncols++] = start + i; /* local rows */
5710: for (i = imark; i < nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */
5711: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, &isrowb));
5712: PetscCall(ISCreateStride(PETSC_COMM_SELF, B->cmap->N, 0, 1, &iscolb));
5713: } else {
5714: PetscCheck(rowb && colb, PETSC_COMM_SELF, PETSC_ERR_SUP, "IS rowb and colb must be provided for MAT_REUSE_MATRIX");
5715: isrowb = *rowb;
5716: iscolb = *colb;
5717: PetscCall(PetscMalloc1(1, &bseq));
5718: bseq[0] = *B_seq;
5719: }
5720: PetscCall(MatCreateSubMatrices(B, 1, &isrowb, &iscolb, scall, &bseq));
5721: *B_seq = bseq[0];
5722: PetscCall(PetscFree(bseq));
5723: if (!rowb) {
5724: PetscCall(ISDestroy(&isrowb));
5725: } else {
5726: *rowb = isrowb;
5727: }
5728: if (!colb) {
5729: PetscCall(ISDestroy(&iscolb));
5730: } else {
5731: *colb = iscolb;
5732: }
5733: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAcols, A, B, 0, 0));
5734: PetscFunctionReturn(PETSC_SUCCESS);
5735: }
5737: /*
5738: MatGetBrowsOfAoCols_MPIAIJ - Creates a `MATSEQAIJ` matrix by taking rows of B that equal to nonzero columns
5739: of the OFF-DIAGONAL portion of local A
5741: Collective
5743: Input Parameters:
5744: + A,B - the matrices in `MATMPIAIJ` format
5745: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5747: Output Parameter:
5748: + startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or NULL)
5749: . startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or NULL)
5750: . bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or NULL)
5751: - B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N
5753: Developer Note:
5754: This directly accesses information inside the VecScatter associated with the matrix-vector product
5755: for this matrix. This is not desirable..
5757: Level: developer
5759: */
5761: PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat A, Mat B, MatReuse scall, PetscInt **startsj_s, PetscInt **startsj_r, MatScalar **bufa_ptr, Mat *B_oth)
5762: {
5763: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5764: VecScatter ctx;
5765: MPI_Comm comm;
5766: const PetscMPIInt *rprocs, *sprocs;
5767: PetscMPIInt nrecvs, nsends;
5768: const PetscInt *srow, *rstarts, *sstarts;
5769: PetscInt *rowlen, *bufj, *bufJ, ncols = 0, aBn = a->B->cmap->n, row, *b_othi, *b_othj, *rvalues = NULL, *svalues = NULL, *cols, sbs, rbs;
5770: PetscInt i, j, k = 0, l, ll, nrows, *rstartsj = NULL, *sstartsj, len;
5771: PetscScalar *b_otha, *bufa, *bufA, *vals = NULL;
5772: MPI_Request *reqs = NULL, *rwaits = NULL, *swaits = NULL;
5773: PetscMPIInt size, tag, rank, nreqs;
5775: PetscFunctionBegin;
5776: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
5777: PetscCallMPI(MPI_Comm_size(comm, &size));
5779: 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 ")",
5780: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
5781: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAocols, A, B, 0, 0));
5782: PetscCallMPI(MPI_Comm_rank(comm, &rank));
5784: if (size == 1) {
5785: startsj_s = NULL;
5786: bufa_ptr = NULL;
5787: *B_oth = NULL;
5788: PetscFunctionReturn(PETSC_SUCCESS);
5789: }
5791: ctx = a->Mvctx;
5792: tag = ((PetscObject)ctx)->tag;
5794: PetscCall(VecScatterGetRemote_Private(ctx, PETSC_TRUE /*send*/, &nsends, &sstarts, &srow, &sprocs, &sbs));
5795: /* rprocs[] must be ordered so that indices received from them are ordered in rvalues[], which is key to algorithms used in this subroutine */
5796: PetscCall(VecScatterGetRemoteOrdered_Private(ctx, PETSC_FALSE /*recv*/, &nrecvs, &rstarts, NULL /*indices not needed*/, &rprocs, &rbs));
5797: PetscCall(PetscMPIIntCast(nsends + nrecvs, &nreqs));
5798: PetscCall(PetscMalloc1(nreqs, &reqs));
5799: rwaits = reqs;
5800: swaits = PetscSafePointerPlusOffset(reqs, nrecvs);
5802: if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX;
5803: if (scall == MAT_INITIAL_MATRIX) {
5804: /* i-array */
5805: /* post receives */
5806: if (nrecvs) PetscCall(PetscMalloc1(rbs * (rstarts[nrecvs] - rstarts[0]), &rvalues)); /* rstarts can be NULL when nrecvs=0 */
5807: for (i = 0; i < nrecvs; i++) {
5808: rowlen = rvalues + rstarts[i] * rbs;
5809: nrows = (rstarts[i + 1] - rstarts[i]) * rbs; /* num of indices to be received */
5810: PetscCallMPI(MPIU_Irecv(rowlen, nrows, MPIU_INT, rprocs[i], tag, comm, rwaits + i));
5811: }
5813: /* pack the outgoing message */
5814: PetscCall(PetscMalloc2(nsends + 1, &sstartsj, nrecvs + 1, &rstartsj));
5816: sstartsj[0] = 0;
5817: rstartsj[0] = 0;
5818: len = 0; /* total length of j or a array to be sent */
5819: if (nsends) {
5820: k = sstarts[0]; /* ATTENTION: sstarts[0] and rstarts[0] are not necessarily zero */
5821: PetscCall(PetscMalloc1(sbs * (sstarts[nsends] - sstarts[0]), &svalues));
5822: }
5823: for (i = 0; i < nsends; i++) {
5824: rowlen = svalues + (sstarts[i] - sstarts[0]) * sbs;
5825: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5826: for (j = 0; j < nrows; j++) {
5827: row = srow[k] + B->rmap->range[rank]; /* global row idx */
5828: for (l = 0; l < sbs; l++) {
5829: PetscCall(MatGetRow_MPIAIJ(B, row + l, &ncols, NULL, NULL)); /* rowlength */
5831: rowlen[j * sbs + l] = ncols;
5833: len += ncols;
5834: PetscCall(MatRestoreRow_MPIAIJ(B, row + l, &ncols, NULL, NULL));
5835: }
5836: k++;
5837: }
5838: PetscCallMPI(MPIU_Isend(rowlen, nrows * sbs, MPIU_INT, sprocs[i], tag, comm, swaits + i));
5840: sstartsj[i + 1] = len; /* starting point of (i+1)-th outgoing msg in bufj and bufa */
5841: }
5842: /* recvs and sends of i-array are completed */
5843: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5844: PetscCall(PetscFree(svalues));
5846: /* allocate buffers for sending j and a arrays */
5847: PetscCall(PetscMalloc1(len, &bufj));
5848: PetscCall(PetscMalloc1(len, &bufa));
5850: /* create i-array of B_oth */
5851: PetscCall(PetscMalloc1(aBn + 1, &b_othi));
5853: b_othi[0] = 0;
5854: len = 0; /* total length of j or a array to be received */
5855: k = 0;
5856: for (i = 0; i < nrecvs; i++) {
5857: rowlen = rvalues + (rstarts[i] - rstarts[0]) * rbs;
5858: nrows = (rstarts[i + 1] - rstarts[i]) * rbs; /* num of rows to be received */
5859: for (j = 0; j < nrows; j++) {
5860: b_othi[k + 1] = b_othi[k] + rowlen[j];
5861: PetscCall(PetscIntSumError(rowlen[j], len, &len));
5862: k++;
5863: }
5864: rstartsj[i + 1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */
5865: }
5866: PetscCall(PetscFree(rvalues));
5868: /* allocate space for j and a arrays of B_oth */
5869: PetscCall(PetscMalloc1(b_othi[aBn], &b_othj));
5870: PetscCall(PetscMalloc1(b_othi[aBn], &b_otha));
5872: /* j-array */
5873: /* post receives of j-array */
5874: for (i = 0; i < nrecvs; i++) {
5875: nrows = rstartsj[i + 1] - rstartsj[i]; /* length of the msg received */
5876: PetscCallMPI(MPIU_Irecv(PetscSafePointerPlusOffset(b_othj, rstartsj[i]), nrows, MPIU_INT, rprocs[i], tag, comm, rwaits + i));
5877: }
5879: /* pack the outgoing message j-array */
5880: if (nsends) k = sstarts[0];
5881: for (i = 0; i < nsends; i++) {
5882: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5883: bufJ = PetscSafePointerPlusOffset(bufj, sstartsj[i]);
5884: for (j = 0; j < nrows; j++) {
5885: row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5886: for (ll = 0; ll < sbs; ll++) {
5887: PetscCall(MatGetRow_MPIAIJ(B, row + ll, &ncols, &cols, NULL));
5888: for (l = 0; l < ncols; l++) *bufJ++ = cols[l];
5889: PetscCall(MatRestoreRow_MPIAIJ(B, row + ll, &ncols, &cols, NULL));
5890: }
5891: }
5892: PetscCallMPI(MPIU_Isend(PetscSafePointerPlusOffset(bufj, sstartsj[i]), sstartsj[i + 1] - sstartsj[i], MPIU_INT, sprocs[i], tag, comm, swaits + i));
5893: }
5895: /* recvs and sends of j-array are completed */
5896: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5897: } else if (scall == MAT_REUSE_MATRIX) {
5898: sstartsj = *startsj_s;
5899: rstartsj = *startsj_r;
5900: bufa = *bufa_ptr;
5901: PetscCall(MatSeqAIJGetArrayWrite(*B_oth, &b_otha));
5902: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not possess an object container");
5904: /* a-array */
5905: /* post receives of a-array */
5906: for (i = 0; i < nrecvs; i++) {
5907: nrows = rstartsj[i + 1] - rstartsj[i]; /* length of the msg received */
5908: PetscCallMPI(MPIU_Irecv(PetscSafePointerPlusOffset(b_otha, rstartsj[i]), nrows, MPIU_SCALAR, rprocs[i], tag, comm, rwaits + i));
5909: }
5911: /* pack the outgoing message a-array */
5912: if (nsends) k = sstarts[0];
5913: for (i = 0; i < nsends; i++) {
5914: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5915: bufA = PetscSafePointerPlusOffset(bufa, sstartsj[i]);
5916: for (j = 0; j < nrows; j++) {
5917: row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5918: for (ll = 0; ll < sbs; ll++) {
5919: PetscCall(MatGetRow_MPIAIJ(B, row + ll, &ncols, NULL, &vals));
5920: for (l = 0; l < ncols; l++) *bufA++ = vals[l];
5921: PetscCall(MatRestoreRow_MPIAIJ(B, row + ll, &ncols, NULL, &vals));
5922: }
5923: }
5924: PetscCallMPI(MPIU_Isend(PetscSafePointerPlusOffset(bufa, sstartsj[i]), sstartsj[i + 1] - sstartsj[i], MPIU_SCALAR, sprocs[i], tag, comm, swaits + i));
5925: }
5926: /* recvs and sends of a-array are completed */
5927: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5928: PetscCall(PetscFree(reqs));
5930: if (scall == MAT_INITIAL_MATRIX) {
5931: Mat_SeqAIJ *b_oth;
5933: /* put together the new matrix */
5934: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, aBn, B->cmap->N, b_othi, b_othj, b_otha, B_oth));
5936: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5937: /* Since these are PETSc arrays, change flags to free them as necessary. */
5938: b_oth = (Mat_SeqAIJ *)(*B_oth)->data;
5939: b_oth->free_a = PETSC_TRUE;
5940: b_oth->free_ij = PETSC_TRUE;
5941: b_oth->nonew = 0;
5943: PetscCall(PetscFree(bufj));
5944: if (!startsj_s || !bufa_ptr) {
5945: PetscCall(PetscFree2(sstartsj, rstartsj));
5946: PetscCall(PetscFree(bufa_ptr));
5947: } else {
5948: *startsj_s = sstartsj;
5949: *startsj_r = rstartsj;
5950: *bufa_ptr = bufa;
5951: }
5952: } else if (scall == MAT_REUSE_MATRIX) {
5953: PetscCall(MatSeqAIJRestoreArrayWrite(*B_oth, &b_otha));
5954: }
5956: PetscCall(VecScatterRestoreRemote_Private(ctx, PETSC_TRUE, &nsends, &sstarts, &srow, &sprocs, &sbs));
5957: PetscCall(VecScatterRestoreRemoteOrdered_Private(ctx, PETSC_FALSE, &nrecvs, &rstarts, NULL, &rprocs, &rbs));
5958: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAocols, A, B, 0, 0));
5959: PetscFunctionReturn(PETSC_SUCCESS);
5960: }
5962: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat, MatType, MatReuse, Mat *);
5963: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJPERM(Mat, MatType, MatReuse, Mat *);
5964: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJSELL(Mat, MatType, MatReuse, Mat *);
5965: #if defined(PETSC_HAVE_MKL_SPARSE)
5966: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJMKL(Mat, MatType, MatReuse, Mat *);
5967: #endif
5968: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIBAIJ(Mat, MatType, MatReuse, Mat *);
5969: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISBAIJ(Mat, MatType, MatReuse, Mat *);
5970: #if defined(PETSC_HAVE_ELEMENTAL)
5971: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat, MatType, MatReuse, Mat *);
5972: #endif
5973: #if defined(PETSC_HAVE_SCALAPACK) && (defined(PETSC_USE_REAL_SINGLE) || defined(PETSC_USE_REAL_DOUBLE))
5974: PETSC_INTERN PetscErrorCode MatConvert_AIJ_ScaLAPACK(Mat, MatType, MatReuse, Mat *);
5975: #endif
5976: #if defined(PETSC_HAVE_HYPRE)
5977: PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat, MatType, MatReuse, Mat *);
5978: #endif
5979: #if defined(PETSC_HAVE_CUDA)
5980: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCUSPARSE(Mat, MatType, MatReuse, Mat *);
5981: #endif
5982: #if defined(PETSC_HAVE_HIP)
5983: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJHIPSPARSE(Mat, MatType, MatReuse, Mat *);
5984: #endif
5985: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
5986: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJKokkos(Mat, MatType, MatReuse, Mat *);
5987: #endif
5988: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISELL(Mat, MatType, MatReuse, Mat *);
5989: PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat, MatType, MatReuse, Mat *);
5990: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_IS_XAIJ(Mat);
5992: /*
5993: Computes (B'*A')' since computing B*A directly is untenable
5995: n p p
5996: [ ] [ ] [ ]
5997: m [ A ] * n [ B ] = m [ C ]
5998: [ ] [ ] [ ]
6000: */
6001: static PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A, Mat B, Mat C)
6002: {
6003: Mat At, Bt, Ct;
6005: PetscFunctionBegin;
6006: PetscCall(MatTranspose(A, MAT_INITIAL_MATRIX, &At));
6007: PetscCall(MatTranspose(B, MAT_INITIAL_MATRIX, &Bt));
6008: PetscCall(MatMatMult(Bt, At, MAT_INITIAL_MATRIX, PETSC_CURRENT, &Ct));
6009: PetscCall(MatDestroy(&At));
6010: PetscCall(MatDestroy(&Bt));
6011: PetscCall(MatTransposeSetPrecursor(Ct, C));
6012: PetscCall(MatTranspose(Ct, MAT_REUSE_MATRIX, &C));
6013: PetscCall(MatDestroy(&Ct));
6014: PetscFunctionReturn(PETSC_SUCCESS);
6015: }
6017: static PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A, Mat B, PetscReal fill, Mat C)
6018: {
6019: PetscBool cisdense;
6021: PetscFunctionBegin;
6022: 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);
6023: PetscCall(MatSetSizes(C, A->rmap->n, B->cmap->n, A->rmap->N, B->cmap->N));
6024: PetscCall(MatSetBlockSizesFromMats(C, A, B));
6025: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATMPIDENSE, MATMPIDENSECUDA, MATMPIDENSEHIP, ""));
6026: if (!cisdense) PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
6027: PetscCall(MatSetUp(C));
6029: C->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIAIJ;
6030: PetscFunctionReturn(PETSC_SUCCESS);
6031: }
6033: static PetscErrorCode MatProductSetFromOptions_MPIDense_MPIAIJ_AB(Mat C)
6034: {
6035: Mat_Product *product = C->product;
6036: Mat A = product->A, B = product->B;
6038: PetscFunctionBegin;
6039: 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 ")",
6040: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
6041: C->ops->matmultsymbolic = MatMatMultSymbolic_MPIDense_MPIAIJ;
6042: C->ops->productsymbolic = MatProductSymbolic_AB;
6043: PetscFunctionReturn(PETSC_SUCCESS);
6044: }
6046: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_MPIDense_MPIAIJ(Mat C)
6047: {
6048: Mat_Product *product = C->product;
6050: PetscFunctionBegin;
6051: if (product->type == MATPRODUCT_AB) PetscCall(MatProductSetFromOptions_MPIDense_MPIAIJ_AB(C));
6052: PetscFunctionReturn(PETSC_SUCCESS);
6053: }
6055: /*
6056: Merge two sets of sorted nonzeros and return a CSR for the merged (sequential) matrix
6058: Input Parameters:
6060: j1,rowBegin1,rowEnd1,jmap1: describe the first set of nonzeros (Set1)
6061: j2,rowBegin2,rowEnd2,jmap2: describe the second set of nonzeros (Set2)
6063: mat: both sets' nonzeros are on m rows, where m is the number of local rows of the matrix mat
6065: For Set1, j1[] contains column indices of the nonzeros.
6066: 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
6067: respectively (note rowEnd1[k] is not necessarily equal to rwoBegin1[k+1]). Indices in this range of j1[] are sorted,
6068: but might have repeats. jmap1[t+1] - jmap1[t] is the number of repeats for the t-th unique nonzero in Set1.
6070: Similar for Set2.
6072: This routine merges the two sets of nonzeros row by row and removes repeats.
6074: Output Parameters: (memory is allocated by the caller)
6076: i[],j[]: the CSR of the merged matrix, which has m rows.
6077: imap1[]: the k-th unique nonzero in Set1 (k=0,1,...) corresponds to imap1[k]-th unique nonzero in the merged matrix.
6078: imap2[]: similar to imap1[], but for Set2.
6079: Note we order nonzeros row-by-row and from left to right.
6080: */
6081: 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[])
6082: {
6083: PetscInt r, m; /* Row index of mat */
6084: PetscCount t, t1, t2, b1, e1, b2, e2;
6086: PetscFunctionBegin;
6087: PetscCall(MatGetLocalSize(mat, &m, NULL));
6088: t1 = t2 = t = 0; /* Count unique nonzeros of in Set1, Set1 and the merged respectively */
6089: i[0] = 0;
6090: for (r = 0; r < m; r++) { /* Do row by row merging */
6091: b1 = rowBegin1[r];
6092: e1 = rowEnd1[r];
6093: b2 = rowBegin2[r];
6094: e2 = rowEnd2[r];
6095: while (b1 < e1 && b2 < e2) {
6096: if (j1[b1] == j2[b2]) { /* Same column index and hence same nonzero */
6097: j[t] = j1[b1];
6098: imap1[t1] = t;
6099: imap2[t2] = t;
6100: b1 += jmap1[t1 + 1] - jmap1[t1]; /* Jump to next unique local nonzero */
6101: b2 += jmap2[t2 + 1] - jmap2[t2]; /* Jump to next unique remote nonzero */
6102: t1++;
6103: t2++;
6104: t++;
6105: } else if (j1[b1] < j2[b2]) {
6106: j[t] = j1[b1];
6107: imap1[t1] = t;
6108: b1 += jmap1[t1 + 1] - jmap1[t1];
6109: t1++;
6110: t++;
6111: } else {
6112: j[t] = j2[b2];
6113: imap2[t2] = t;
6114: b2 += jmap2[t2 + 1] - jmap2[t2];
6115: t2++;
6116: t++;
6117: }
6118: }
6119: /* Merge the remaining in either j1[] or j2[] */
6120: while (b1 < e1) {
6121: j[t] = j1[b1];
6122: imap1[t1] = t;
6123: b1 += jmap1[t1 + 1] - jmap1[t1];
6124: t1++;
6125: t++;
6126: }
6127: while (b2 < e2) {
6128: j[t] = j2[b2];
6129: imap2[t2] = t;
6130: b2 += jmap2[t2 + 1] - jmap2[t2];
6131: t2++;
6132: t++;
6133: }
6134: PetscCall(PetscIntCast(t, i + r + 1));
6135: }
6136: PetscFunctionReturn(PETSC_SUCCESS);
6137: }
6139: /*
6140: Split nonzeros in a block of local rows into two subsets: those in the diagonal block and those in the off-diagonal block
6142: Input Parameters:
6143: mat: an MPI matrix that provides row and column layout information for splitting. Let's say its number of local rows is m.
6144: 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[]
6145: respectively, along with a permutation array perm[]. Length of the i[],j[],perm[] arrays is n.
6147: i[] is already sorted, but within a row, j[] is not sorted and might have repeats.
6148: i[] might contain negative indices at the beginning, which means the corresponding entries should be ignored in the splitting.
6150: Output Parameters:
6151: j[],perm[]: the routine needs to sort j[] within each row along with perm[].
6152: rowBegin[],rowMid[],rowEnd[]: of length m, and the memory is preallocated and zeroed by the caller.
6153: 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,
6154: and [rowMid[r],rowEnd[r]) point to begin/end entries of row r of the off-diagonal block.
6156: Aperm[],Ajmap[],Atot,Annz: Arrays are allocated by this routine.
6157: Atot: number of entries belonging to the diagonal block.
6158: Annz: number of unique nonzeros belonging to the diagonal block.
6159: Aperm[Atot] stores values from perm[] for entries belonging to the diagonal block. Length of Aperm[] is Atot, though it may also count
6160: repeats (i.e., same 'i,j' pair).
6161: Ajmap[Annz+1] stores the number of repeats of each unique entry belonging to the diagonal block. More precisely, Ajmap[t+1] - Ajmap[t]
6162: is the number of repeats for the t-th unique entry in the diagonal block. Ajmap[0] is always 0.
6164: Atot: number of entries belonging to the diagonal block
6165: Annz: number of unique nonzeros belonging to the diagonal block.
6167: Bperm[], Bjmap[], Btot, Bnnz are similar but for the off-diagonal block.
6169: Aperm[],Bperm[],Ajmap[] and Bjmap[] are allocated separately by this routine with PetscMalloc1().
6170: */
6171: 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_)
6172: {
6173: PetscInt cstart, cend, rstart, rend, row, col;
6174: PetscCount Atot = 0, Btot = 0; /* Total number of nonzeros in the diagonal and off-diagonal blocks */
6175: PetscCount Annz = 0, Bnnz = 0; /* Number of unique nonzeros in the diagonal and off-diagonal blocks */
6176: PetscCount k, m, p, q, r, s, mid;
6177: PetscCount *Aperm, *Bperm, *Ajmap, *Bjmap;
6179: PetscFunctionBegin;
6180: PetscCall(PetscLayoutGetRange(mat->rmap, &rstart, &rend));
6181: PetscCall(PetscLayoutGetRange(mat->cmap, &cstart, &cend));
6182: m = rend - rstart;
6184: /* Skip negative rows */
6185: for (k = 0; k < n; k++)
6186: if (i[k] >= 0) break;
6188: /* Process [k,n): sort and partition each local row into diag and offdiag portions,
6189: fill rowBegin[], rowMid[], rowEnd[], and count Atot, Btot, Annz, Bnnz.
6190: */
6191: while (k < n) {
6192: row = i[k];
6193: /* Entries in [k,s) are in one row. Shift diagonal block col indices so that diag is ahead of offdiag after sorting the row */
6194: for (s = k; s < n; s++)
6195: if (i[s] != row) break;
6197: /* Shift diag columns to range of [-PETSC_INT_MAX, -1] */
6198: for (p = k; p < s; p++) {
6199: if (j[p] >= cstart && j[p] < cend) j[p] -= PETSC_INT_MAX;
6200: }
6201: PetscCall(PetscSortIntWithCountArray(s - k, j + k, perm + k));
6202: PetscCall(PetscSortedIntUpperBound(j, k, s, -1, &mid)); /* Separate [k,s) into [k,mid) for diag and [mid,s) for offdiag */
6203: rowBegin[row - rstart] = k;
6204: rowMid[row - rstart] = mid;
6205: rowEnd[row - rstart] = s;
6206: 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);
6208: /* Count nonzeros of this diag/offdiag row, which might have repeats */
6209: Atot += mid - k;
6210: Btot += s - mid;
6212: /* Count unique nonzeros of this diag row */
6213: for (p = k; p < mid;) {
6214: col = j[p];
6215: do {
6216: j[p] += PETSC_INT_MAX; /* Revert the modified diagonal indices */
6217: p++;
6218: } while (p < mid && j[p] == col);
6219: Annz++;
6220: }
6222: /* Count unique nonzeros of this offdiag row */
6223: for (p = mid; p < s;) {
6224: col = j[p];
6225: do {
6226: p++;
6227: } while (p < s && j[p] == col);
6228: Bnnz++;
6229: }
6230: k = s;
6231: }
6233: /* Allocation according to Atot, Btot, Annz, Bnnz */
6234: PetscCall(PetscMalloc1(Atot, &Aperm));
6235: PetscCall(PetscMalloc1(Btot, &Bperm));
6236: PetscCall(PetscMalloc1(Annz + 1, &Ajmap));
6237: PetscCall(PetscMalloc1(Bnnz + 1, &Bjmap));
6239: /* Re-scan indices and copy diag/offdiag permutation indices to Aperm, Bperm and also fill Ajmap and Bjmap */
6240: Ajmap[0] = Bjmap[0] = Atot = Btot = Annz = Bnnz = 0;
6241: for (r = 0; r < m; r++) {
6242: k = rowBegin[r];
6243: mid = rowMid[r];
6244: s = rowEnd[r];
6245: PetscCall(PetscArraycpy(PetscSafePointerPlusOffset(Aperm, Atot), PetscSafePointerPlusOffset(perm, k), mid - k));
6246: PetscCall(PetscArraycpy(PetscSafePointerPlusOffset(Bperm, Btot), PetscSafePointerPlusOffset(perm, mid), s - mid));
6247: Atot += mid - k;
6248: Btot += s - mid;
6250: /* Scan column indices in this row and find out how many repeats each unique nonzero has */
6251: for (p = k; p < mid;) {
6252: col = j[p];
6253: q = p;
6254: do {
6255: p++;
6256: } while (p < mid && j[p] == col);
6257: Ajmap[Annz + 1] = Ajmap[Annz] + (p - q);
6258: Annz++;
6259: }
6261: for (p = mid; p < s;) {
6262: col = j[p];
6263: q = p;
6264: do {
6265: p++;
6266: } while (p < s && j[p] == col);
6267: Bjmap[Bnnz + 1] = Bjmap[Bnnz] + (p - q);
6268: Bnnz++;
6269: }
6270: }
6271: /* Output */
6272: *Aperm_ = Aperm;
6273: *Annz_ = Annz;
6274: *Atot_ = Atot;
6275: *Ajmap_ = Ajmap;
6276: *Bperm_ = Bperm;
6277: *Bnnz_ = Bnnz;
6278: *Btot_ = Btot;
6279: *Bjmap_ = Bjmap;
6280: PetscFunctionReturn(PETSC_SUCCESS);
6281: }
6283: /*
6284: Expand the jmap[] array to make a new one in view of nonzeros in the merged matrix
6286: Input Parameters:
6287: nnz1: number of unique nonzeros in a set that was used to produce imap[], jmap[]
6288: nnz: number of unique nonzeros in the merged matrix
6289: imap[nnz1]: i-th nonzero in the set is the imap[i]-th nonzero in the merged matrix
6290: jmap[nnz1+1]: i-th nonzero in the set has jmap[i+1] - jmap[i] repeats in the set
6292: Output Parameter: (memory is allocated by the caller)
6293: jmap_new[nnz+1]: i-th nonzero in the merged matrix has jmap_new[i+1] - jmap_new[i] repeats in the set
6295: Example:
6296: nnz1 = 4
6297: nnz = 6
6298: imap = [1,3,4,5]
6299: jmap = [0,3,5,6,7]
6300: then,
6301: jmap_new = [0,0,3,3,5,6,7]
6302: */
6303: static PetscErrorCode ExpandJmap_Internal(PetscCount nnz1, PetscCount nnz, const PetscCount imap[], const PetscCount jmap[], PetscCount jmap_new[])
6304: {
6305: PetscCount k, p;
6307: PetscFunctionBegin;
6308: jmap_new[0] = 0;
6309: p = nnz; /* p loops over jmap_new[] backwards */
6310: for (k = nnz1 - 1; k >= 0; k--) { /* k loops over imap[] */
6311: for (; p > imap[k]; p--) jmap_new[p] = jmap[k + 1];
6312: }
6313: for (; p >= 0; p--) jmap_new[p] = jmap[0];
6314: PetscFunctionReturn(PETSC_SUCCESS);
6315: }
6317: static PetscErrorCode MatCOOStructDestroy_MPIAIJ(void **data)
6318: {
6319: MatCOOStruct_MPIAIJ *coo = (MatCOOStruct_MPIAIJ *)*data;
6321: PetscFunctionBegin;
6322: PetscCall(PetscSFDestroy(&coo->sf));
6323: PetscCall(PetscFree(coo->Aperm1));
6324: PetscCall(PetscFree(coo->Bperm1));
6325: PetscCall(PetscFree(coo->Ajmap1));
6326: PetscCall(PetscFree(coo->Bjmap1));
6327: PetscCall(PetscFree(coo->Aimap2));
6328: PetscCall(PetscFree(coo->Bimap2));
6329: PetscCall(PetscFree(coo->Aperm2));
6330: PetscCall(PetscFree(coo->Bperm2));
6331: PetscCall(PetscFree(coo->Ajmap2));
6332: PetscCall(PetscFree(coo->Bjmap2));
6333: PetscCall(PetscFree(coo->Cperm1));
6334: PetscCall(PetscFree2(coo->sendbuf, coo->recvbuf));
6335: PetscCall(PetscFree(coo));
6336: PetscFunctionReturn(PETSC_SUCCESS);
6337: }
6339: PetscErrorCode MatSetPreallocationCOO_MPIAIJ(Mat mat, PetscCount coo_n, PetscInt coo_i[], PetscInt coo_j[])
6340: {
6341: MPI_Comm comm;
6342: PetscMPIInt rank, size;
6343: PetscInt m, n, M, N, rstart, rend, cstart, cend; /* Sizes, indices of row/col, therefore with type PetscInt */
6344: PetscCount k, p, q, rem; /* Loop variables over coo arrays */
6345: Mat_MPIAIJ *mpiaij = (Mat_MPIAIJ *)mat->data;
6346: PetscContainer container;
6347: MatCOOStruct_MPIAIJ *coo;
6349: PetscFunctionBegin;
6350: PetscCall(PetscFree(mpiaij->garray));
6351: PetscCall(VecDestroy(&mpiaij->lvec));
6352: #if defined(PETSC_USE_CTABLE)
6353: PetscCall(PetscHMapIDestroy(&mpiaij->colmap));
6354: #else
6355: PetscCall(PetscFree(mpiaij->colmap));
6356: #endif
6357: PetscCall(VecScatterDestroy(&mpiaij->Mvctx));
6358: mat->assembled = PETSC_FALSE;
6359: mat->was_assembled = PETSC_FALSE;
6361: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
6362: PetscCallMPI(MPI_Comm_size(comm, &size));
6363: PetscCallMPI(MPI_Comm_rank(comm, &rank));
6364: PetscCall(PetscLayoutSetUp(mat->rmap));
6365: PetscCall(PetscLayoutSetUp(mat->cmap));
6366: PetscCall(PetscLayoutGetRange(mat->rmap, &rstart, &rend));
6367: PetscCall(PetscLayoutGetRange(mat->cmap, &cstart, &cend));
6368: PetscCall(MatGetLocalSize(mat, &m, &n));
6369: PetscCall(MatGetSize(mat, &M, &N));
6371: /* Sort (i,j) by row along with a permutation array, so that the to-be-ignored */
6372: /* entries come first, then local rows, then remote rows. */
6373: PetscCount n1 = coo_n, *perm1;
6374: PetscInt *i1 = coo_i, *j1 = coo_j;
6376: PetscCall(PetscMalloc1(n1, &perm1));
6377: for (k = 0; k < n1; k++) perm1[k] = k;
6379: /* Manipulate indices so that entries with negative row or col indices will have smallest
6380: row indices, local entries will have greater but negative row indices, and remote entries
6381: will have positive row indices.
6382: */
6383: for (k = 0; k < n1; k++) {
6384: if (i1[k] < 0 || j1[k] < 0) i1[k] = PETSC_INT_MIN; /* e.g., -2^31, minimal to move them ahead */
6385: 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] */
6386: else {
6387: PetscCheck(!mat->nooffprocentries, PETSC_COMM_SELF, PETSC_ERR_USER_INPUT, "MAT_NO_OFF_PROC_ENTRIES is set but insert to remote rows");
6388: if (mpiaij->donotstash) i1[k] = PETSC_INT_MIN; /* Ignore offproc entries as if they had negative indices */
6389: }
6390: }
6392: /* Sort by row; after that, [0,k) have ignored entries, [k,rem) have local rows and [rem,n1) have remote rows */
6393: PetscCall(PetscSortIntWithIntCountArrayPair(n1, i1, j1, perm1));
6395: /* Advance k to the first entry we need to take care of */
6396: for (k = 0; k < n1; k++)
6397: if (i1[k] > PETSC_INT_MIN) break;
6398: PetscCount i1start = k;
6400: PetscCall(PetscSortedIntUpperBound(i1, k, n1, rend - 1 - PETSC_INT_MAX, &rem)); /* rem is upper bound of the last local row */
6401: for (; k < rem; k++) i1[k] += PETSC_INT_MAX; /* Revert row indices of local rows*/
6403: 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);
6405: /* Send remote rows to their owner */
6406: /* Find which rows should be sent to which remote ranks*/
6407: PetscInt nsend = 0; /* Number of MPI ranks to send data to */
6408: PetscMPIInt *sendto; /* [nsend], storing remote ranks */
6409: PetscInt *nentries; /* [nsend], storing number of entries sent to remote ranks; Assume PetscInt is big enough for this count, and error if not */
6410: const PetscInt *ranges;
6411: PetscInt maxNsend = size >= 128 ? 128 : size; /* Assume max 128 neighbors; realloc when needed */
6413: PetscCall(PetscLayoutGetRanges(mat->rmap, &ranges));
6414: PetscCall(PetscMalloc2(maxNsend, &sendto, maxNsend, &nentries));
6415: for (k = rem; k < n1;) {
6416: PetscMPIInt owner;
6417: PetscInt firstRow, lastRow;
6419: /* Locate a row range */
6420: firstRow = i1[k]; /* first row of this owner */
6421: PetscCall(PetscLayoutFindOwner(mat->rmap, firstRow, &owner));
6422: lastRow = ranges[owner + 1] - 1; /* last row of this owner */
6424: /* Find the first index 'p' in [k,n) with i1[p] belonging to next owner */
6425: PetscCall(PetscSortedIntUpperBound(i1, k, n1, lastRow, &p));
6427: /* All entries in [k,p) belong to this remote owner */
6428: if (nsend >= maxNsend) { /* Double the remote ranks arrays if not long enough */
6429: PetscMPIInt *sendto2;
6430: PetscInt *nentries2;
6431: PetscInt maxNsend2 = (maxNsend <= size / 2) ? maxNsend * 2 : size;
6433: PetscCall(PetscMalloc2(maxNsend2, &sendto2, maxNsend2, &nentries2));
6434: PetscCall(PetscArraycpy(sendto2, sendto, maxNsend));
6435: PetscCall(PetscArraycpy(nentries2, nentries2, maxNsend + 1));
6436: PetscCall(PetscFree2(sendto, nentries2));
6437: sendto = sendto2;
6438: nentries = nentries2;
6439: maxNsend = maxNsend2;
6440: }
6441: sendto[nsend] = owner;
6442: PetscCall(PetscIntCast(p - k, &nentries[nsend]));
6443: nsend++;
6444: k = p;
6445: }
6447: /* Build 1st SF to know offsets on remote to send data */
6448: PetscSF sf1;
6449: PetscInt nroots = 1, nroots2 = 0;
6450: PetscInt nleaves = nsend, nleaves2 = 0;
6451: PetscInt *offsets;
6452: PetscSFNode *iremote;
6454: PetscCall(PetscSFCreate(comm, &sf1));
6455: PetscCall(PetscMalloc1(nsend, &iremote));
6456: PetscCall(PetscMalloc1(nsend, &offsets));
6457: for (k = 0; k < nsend; k++) {
6458: iremote[k].rank = sendto[k];
6459: iremote[k].index = 0;
6460: nleaves2 += nentries[k];
6461: PetscCheck(nleaves2 >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Number of SF leaves is too large for PetscInt");
6462: }
6463: PetscCall(PetscSFSetGraph(sf1, nroots, nleaves, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
6464: PetscCall(PetscSFFetchAndOpWithMemTypeBegin(sf1, MPIU_INT, PETSC_MEMTYPE_HOST, &nroots2 /*rootdata*/, PETSC_MEMTYPE_HOST, nentries /*leafdata*/, PETSC_MEMTYPE_HOST, offsets /*leafupdate*/, MPI_SUM));
6465: PetscCall(PetscSFFetchAndOpEnd(sf1, MPIU_INT, &nroots2, nentries, offsets, MPI_SUM)); /* Would nroots2 overflow, we check offsets[] below */
6466: PetscCall(PetscSFDestroy(&sf1));
6467: PetscAssert(nleaves2 == n1 - rem, PETSC_COMM_SELF, PETSC_ERR_PLIB, "nleaves2 %" PetscInt_FMT " != number of remote entries %" PetscCount_FMT, nleaves2, n1 - rem);
6469: /* Build 2nd SF to send remote COOs to their owner */
6470: PetscSF sf2;
6471: nroots = nroots2;
6472: nleaves = nleaves2;
6473: PetscCall(PetscSFCreate(comm, &sf2));
6474: PetscCall(PetscSFSetFromOptions(sf2));
6475: PetscCall(PetscMalloc1(nleaves, &iremote));
6476: p = 0;
6477: for (k = 0; k < nsend; k++) {
6478: PetscCheck(offsets[k] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Number of SF roots is too large for PetscInt");
6479: for (q = 0; q < nentries[k]; q++, p++) {
6480: iremote[p].rank = sendto[k];
6481: PetscCall(PetscIntCast(offsets[k] + q, &iremote[p].index));
6482: }
6483: }
6484: PetscCall(PetscSFSetGraph(sf2, nroots, nleaves, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
6486: /* Send the remote COOs to their owner */
6487: PetscInt n2 = nroots, *i2, *j2; /* Buffers for received COOs from other ranks, along with a permutation array */
6488: PetscCount *perm2; /* Though PetscInt is enough for remote entries, we use PetscCount here as we want to reuse MatSplitEntries_Internal() */
6489: PetscCall(PetscMalloc3(n2, &i2, n2, &j2, n2, &perm2));
6490: PetscAssert(rem == 0 || i1 != NULL, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot add nonzero offset to null");
6491: PetscAssert(rem == 0 || j1 != NULL, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot add nonzero offset to null");
6492: PetscInt *i1prem = PetscSafePointerPlusOffset(i1, rem);
6493: PetscInt *j1prem = PetscSafePointerPlusOffset(j1, rem);
6494: PetscCall(PetscSFReduceWithMemTypeBegin(sf2, MPIU_INT, PETSC_MEMTYPE_HOST, i1prem, PETSC_MEMTYPE_HOST, i2, MPI_REPLACE));
6495: PetscCall(PetscSFReduceEnd(sf2, MPIU_INT, i1prem, i2, MPI_REPLACE));
6496: PetscCall(PetscSFReduceWithMemTypeBegin(sf2, MPIU_INT, PETSC_MEMTYPE_HOST, j1prem, PETSC_MEMTYPE_HOST, j2, MPI_REPLACE));
6497: PetscCall(PetscSFReduceEnd(sf2, MPIU_INT, j1prem, j2, MPI_REPLACE));
6499: PetscCall(PetscFree(offsets));
6500: PetscCall(PetscFree2(sendto, nentries));
6502: /* Sort received COOs by row along with the permutation array */
6503: for (k = 0; k < n2; k++) perm2[k] = k;
6504: PetscCall(PetscSortIntWithIntCountArrayPair(n2, i2, j2, perm2));
6506: /* sf2 only sends contiguous leafdata to contiguous rootdata. We record the permutation which will be used to fill leafdata */
6507: PetscCount *Cperm1;
6508: PetscAssert(rem == 0 || perm1 != NULL, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot add nonzero offset to null");
6509: PetscCount *perm1prem = PetscSafePointerPlusOffset(perm1, rem);
6510: PetscCall(PetscMalloc1(nleaves, &Cperm1));
6511: PetscCall(PetscArraycpy(Cperm1, perm1prem, nleaves));
6513: /* Support for HYPRE matrices, kind of a hack.
6514: Swap min column with diagonal so that diagonal values will go first */
6515: PetscBool hypre;
6516: PetscCall(PetscStrcmp("_internal_COO_mat_for_hypre", ((PetscObject)mat)->name, &hypre));
6517: if (hypre) {
6518: PetscInt *minj;
6519: PetscBT hasdiag;
6521: PetscCall(PetscBTCreate(m, &hasdiag));
6522: PetscCall(PetscMalloc1(m, &minj));
6523: for (k = 0; k < m; k++) minj[k] = PETSC_INT_MAX;
6524: for (k = i1start; k < rem; k++) {
6525: if (j1[k] < cstart || j1[k] >= cend) continue;
6526: const PetscInt rindex = i1[k] - rstart;
6527: if ((j1[k] - cstart) == rindex) PetscCall(PetscBTSet(hasdiag, rindex));
6528: minj[rindex] = PetscMin(minj[rindex], j1[k]);
6529: }
6530: for (k = 0; k < n2; k++) {
6531: if (j2[k] < cstart || j2[k] >= cend) continue;
6532: const PetscInt rindex = i2[k] - rstart;
6533: if ((j2[k] - cstart) == rindex) PetscCall(PetscBTSet(hasdiag, rindex));
6534: minj[rindex] = PetscMin(minj[rindex], j2[k]);
6535: }
6536: for (k = i1start; k < rem; k++) {
6537: const PetscInt rindex = i1[k] - rstart;
6538: if (j1[k] < cstart || j1[k] >= cend || !PetscBTLookup(hasdiag, rindex)) continue;
6539: if (j1[k] == minj[rindex]) j1[k] = i1[k] + (cstart - rstart);
6540: else if ((j1[k] - cstart) == rindex) j1[k] = minj[rindex];
6541: }
6542: for (k = 0; k < n2; k++) {
6543: const PetscInt rindex = i2[k] - rstart;
6544: if (j2[k] < cstart || j2[k] >= cend || !PetscBTLookup(hasdiag, rindex)) continue;
6545: if (j2[k] == minj[rindex]) j2[k] = i2[k] + (cstart - rstart);
6546: else if ((j2[k] - cstart) == rindex) j2[k] = minj[rindex];
6547: }
6548: PetscCall(PetscBTDestroy(&hasdiag));
6549: PetscCall(PetscFree(minj));
6550: }
6552: /* Split local COOs and received COOs into diag/offdiag portions */
6553: PetscCount *rowBegin1, *rowMid1, *rowEnd1;
6554: PetscCount *Ajmap1, *Aperm1, *Bjmap1, *Bperm1;
6555: PetscCount Annz1, Bnnz1, Atot1, Btot1;
6556: PetscCount *rowBegin2, *rowMid2, *rowEnd2;
6557: PetscCount *Ajmap2, *Aperm2, *Bjmap2, *Bperm2;
6558: PetscCount Annz2, Bnnz2, Atot2, Btot2;
6560: PetscCall(PetscCalloc3(m, &rowBegin1, m, &rowMid1, m, &rowEnd1));
6561: PetscCall(PetscCalloc3(m, &rowBegin2, m, &rowMid2, m, &rowEnd2));
6562: PetscCall(MatSplitEntries_Internal(mat, rem, i1, j1, perm1, rowBegin1, rowMid1, rowEnd1, &Atot1, &Aperm1, &Annz1, &Ajmap1, &Btot1, &Bperm1, &Bnnz1, &Bjmap1));
6563: PetscCall(MatSplitEntries_Internal(mat, n2, i2, j2, perm2, rowBegin2, rowMid2, rowEnd2, &Atot2, &Aperm2, &Annz2, &Ajmap2, &Btot2, &Bperm2, &Bnnz2, &Bjmap2));
6565: /* Merge local COOs with received COOs: diag with diag, offdiag with offdiag */
6566: PetscInt *Ai, *Bi;
6567: PetscInt *Aj, *Bj;
6569: PetscCall(PetscMalloc1(m + 1, &Ai));
6570: PetscCall(PetscMalloc1(m + 1, &Bi));
6571: PetscCall(PetscMalloc1(Annz1 + Annz2, &Aj)); /* Since local and remote entries might have dups, we might allocate excess memory */
6572: PetscCall(PetscMalloc1(Bnnz1 + Bnnz2, &Bj));
6574: PetscCount *Aimap1, *Bimap1, *Aimap2, *Bimap2;
6575: PetscCall(PetscMalloc1(Annz1, &Aimap1));
6576: PetscCall(PetscMalloc1(Bnnz1, &Bimap1));
6577: PetscCall(PetscMalloc1(Annz2, &Aimap2));
6578: PetscCall(PetscMalloc1(Bnnz2, &Bimap2));
6580: PetscCall(MatMergeEntries_Internal(mat, j1, j2, rowBegin1, rowMid1, rowBegin2, rowMid2, Ajmap1, Ajmap2, Aimap1, Aimap2, Ai, Aj));
6581: PetscCall(MatMergeEntries_Internal(mat, j1, j2, rowMid1, rowEnd1, rowMid2, rowEnd2, Bjmap1, Bjmap2, Bimap1, Bimap2, Bi, Bj));
6583: /* Expand Ajmap1/Bjmap1 to make them based off nonzeros in A/B, since we */
6584: /* expect nonzeros in A/B most likely have local contributing entries */
6585: PetscInt Annz = Ai[m];
6586: PetscInt Bnnz = Bi[m];
6587: PetscCount *Ajmap1_new, *Bjmap1_new;
6589: PetscCall(PetscMalloc1(Annz + 1, &Ajmap1_new));
6590: PetscCall(PetscMalloc1(Bnnz + 1, &Bjmap1_new));
6592: PetscCall(ExpandJmap_Internal(Annz1, Annz, Aimap1, Ajmap1, Ajmap1_new));
6593: PetscCall(ExpandJmap_Internal(Bnnz1, Bnnz, Bimap1, Bjmap1, Bjmap1_new));
6595: PetscCall(PetscFree(Aimap1));
6596: PetscCall(PetscFree(Ajmap1));
6597: PetscCall(PetscFree(Bimap1));
6598: PetscCall(PetscFree(Bjmap1));
6599: PetscCall(PetscFree3(rowBegin1, rowMid1, rowEnd1));
6600: PetscCall(PetscFree3(rowBegin2, rowMid2, rowEnd2));
6601: PetscCall(PetscFree(perm1));
6602: PetscCall(PetscFree3(i2, j2, perm2));
6604: Ajmap1 = Ajmap1_new;
6605: Bjmap1 = Bjmap1_new;
6607: /* Reallocate Aj, Bj once we know actual numbers of unique nonzeros in A and B */
6608: if (Annz < Annz1 + Annz2) {
6609: PetscInt *Aj_new;
6610: PetscCall(PetscMalloc1(Annz, &Aj_new));
6611: PetscCall(PetscArraycpy(Aj_new, Aj, Annz));
6612: PetscCall(PetscFree(Aj));
6613: Aj = Aj_new;
6614: }
6616: if (Bnnz < Bnnz1 + Bnnz2) {
6617: PetscInt *Bj_new;
6618: PetscCall(PetscMalloc1(Bnnz, &Bj_new));
6619: PetscCall(PetscArraycpy(Bj_new, Bj, Bnnz));
6620: PetscCall(PetscFree(Bj));
6621: Bj = Bj_new;
6622: }
6624: /* Create new submatrices for on-process and off-process coupling */
6625: PetscScalar *Aa, *Ba;
6626: MatType rtype;
6627: Mat_SeqAIJ *a, *b;
6628: PetscObjectState state;
6629: PetscCall(PetscCalloc1(Annz, &Aa)); /* Zero matrix on device */
6630: PetscCall(PetscCalloc1(Bnnz, &Ba));
6631: /* make Aj[] local, i.e, based off the start column of the diagonal portion */
6632: if (cstart) {
6633: for (k = 0; k < Annz; k++) Aj[k] -= cstart;
6634: }
6636: PetscCall(MatGetRootType_Private(mat, &rtype));
6638: MatSeqXAIJGetOptions_Private(mpiaij->A);
6639: PetscCall(MatDestroy(&mpiaij->A));
6640: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, n, Ai, Aj, Aa, &mpiaij->A));
6641: PetscCall(MatSetBlockSizesFromMats(mpiaij->A, mat, mat));
6642: MatSeqXAIJRestoreOptions_Private(mpiaij->A);
6644: MatSeqXAIJGetOptions_Private(mpiaij->B);
6645: PetscCall(MatDestroy(&mpiaij->B));
6646: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, mat->cmap->N, Bi, Bj, Ba, &mpiaij->B));
6647: PetscCall(MatSetBlockSizesFromMats(mpiaij->B, mat, mat));
6648: MatSeqXAIJRestoreOptions_Private(mpiaij->B);
6650: PetscCall(MatSetUpMultiply_MPIAIJ(mat));
6651: mat->was_assembled = PETSC_TRUE; // was_assembled in effect means the Mvctx is built; doing so avoids redundant MatSetUpMultiply_MPIAIJ
6652: state = mpiaij->A->nonzerostate + mpiaij->B->nonzerostate;
6653: PetscCallMPI(MPIU_Allreduce(&state, &mat->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)mat)));
6655: a = (Mat_SeqAIJ *)mpiaij->A->data;
6656: b = (Mat_SeqAIJ *)mpiaij->B->data;
6657: a->free_a = PETSC_TRUE;
6658: a->free_ij = PETSC_TRUE;
6659: b->free_a = PETSC_TRUE;
6660: b->free_ij = PETSC_TRUE;
6661: a->maxnz = a->nz;
6662: b->maxnz = b->nz;
6664: /* conversion must happen AFTER multiply setup */
6665: PetscCall(MatConvert(mpiaij->A, rtype, MAT_INPLACE_MATRIX, &mpiaij->A));
6666: PetscCall(MatConvert(mpiaij->B, rtype, MAT_INPLACE_MATRIX, &mpiaij->B));
6667: PetscCall(VecDestroy(&mpiaij->lvec));
6668: PetscCall(MatCreateVecs(mpiaij->B, &mpiaij->lvec, NULL));
6670: // Put the COO struct in a container and then attach that to the matrix
6671: PetscCall(PetscMalloc1(1, &coo));
6672: coo->n = coo_n;
6673: coo->sf = sf2;
6674: coo->sendlen = nleaves;
6675: coo->recvlen = nroots;
6676: coo->Annz = Annz;
6677: coo->Bnnz = Bnnz;
6678: coo->Annz2 = Annz2;
6679: coo->Bnnz2 = Bnnz2;
6680: coo->Atot1 = Atot1;
6681: coo->Atot2 = Atot2;
6682: coo->Btot1 = Btot1;
6683: coo->Btot2 = Btot2;
6684: coo->Ajmap1 = Ajmap1;
6685: coo->Aperm1 = Aperm1;
6686: coo->Bjmap1 = Bjmap1;
6687: coo->Bperm1 = Bperm1;
6688: coo->Aimap2 = Aimap2;
6689: coo->Ajmap2 = Ajmap2;
6690: coo->Aperm2 = Aperm2;
6691: coo->Bimap2 = Bimap2;
6692: coo->Bjmap2 = Bjmap2;
6693: coo->Bperm2 = Bperm2;
6694: coo->Cperm1 = Cperm1;
6695: // Allocate in preallocation. If not used, it has zero cost on host
6696: PetscCall(PetscMalloc2(coo->sendlen, &coo->sendbuf, coo->recvlen, &coo->recvbuf));
6697: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
6698: PetscCall(PetscContainerSetPointer(container, coo));
6699: PetscCall(PetscContainerSetCtxDestroy(container, MatCOOStructDestroy_MPIAIJ));
6700: PetscCall(PetscObjectCompose((PetscObject)mat, "__PETSc_MatCOOStruct_Host", (PetscObject)container));
6701: PetscCall(PetscContainerDestroy(&container));
6702: PetscFunctionReturn(PETSC_SUCCESS);
6703: }
6705: static PetscErrorCode MatSetValuesCOO_MPIAIJ(Mat mat, const PetscScalar v[], InsertMode imode)
6706: {
6707: Mat_MPIAIJ *mpiaij = (Mat_MPIAIJ *)mat->data;
6708: Mat A = mpiaij->A, B = mpiaij->B;
6709: PetscScalar *Aa, *Ba;
6710: PetscScalar *sendbuf, *recvbuf;
6711: const PetscCount *Ajmap1, *Ajmap2, *Aimap2;
6712: const PetscCount *Bjmap1, *Bjmap2, *Bimap2;
6713: const PetscCount *Aperm1, *Aperm2, *Bperm1, *Bperm2;
6714: const PetscCount *Cperm1;
6715: PetscContainer container;
6716: MatCOOStruct_MPIAIJ *coo;
6718: PetscFunctionBegin;
6719: PetscCall(PetscObjectQuery((PetscObject)mat, "__PETSc_MatCOOStruct_Host", (PetscObject *)&container));
6720: PetscCheck(container, PetscObjectComm((PetscObject)mat), PETSC_ERR_PLIB, "Not found MatCOOStruct on this matrix");
6721: PetscCall(PetscContainerGetPointer(container, (void **)&coo));
6722: sendbuf = coo->sendbuf;
6723: recvbuf = coo->recvbuf;
6724: Ajmap1 = coo->Ajmap1;
6725: Ajmap2 = coo->Ajmap2;
6726: Aimap2 = coo->Aimap2;
6727: Bjmap1 = coo->Bjmap1;
6728: Bjmap2 = coo->Bjmap2;
6729: Bimap2 = coo->Bimap2;
6730: Aperm1 = coo->Aperm1;
6731: Aperm2 = coo->Aperm2;
6732: Bperm1 = coo->Bperm1;
6733: Bperm2 = coo->Bperm2;
6734: Cperm1 = coo->Cperm1;
6736: PetscCall(MatSeqAIJGetArray(A, &Aa)); /* Might read and write matrix values */
6737: PetscCall(MatSeqAIJGetArray(B, &Ba));
6739: /* Pack entries to be sent to remote */
6740: for (PetscCount i = 0; i < coo->sendlen; i++) sendbuf[i] = v[Cperm1[i]];
6742: /* Send remote entries to their owner and overlap the communication with local computation */
6743: PetscCall(PetscSFReduceWithMemTypeBegin(coo->sf, MPIU_SCALAR, PETSC_MEMTYPE_HOST, sendbuf, PETSC_MEMTYPE_HOST, recvbuf, MPI_REPLACE));
6744: /* Add local entries to A and B */
6745: for (PetscCount i = 0; i < coo->Annz; i++) { /* All nonzeros in A are either zero'ed or added with a value (i.e., initialized) */
6746: PetscScalar sum = 0.0; /* Do partial summation first to improve numerical stability */
6747: for (PetscCount k = Ajmap1[i]; k < Ajmap1[i + 1]; k++) sum += v[Aperm1[k]];
6748: Aa[i] = (imode == INSERT_VALUES ? 0.0 : Aa[i]) + sum;
6749: }
6750: for (PetscCount i = 0; i < coo->Bnnz; i++) {
6751: PetscScalar sum = 0.0;
6752: for (PetscCount k = Bjmap1[i]; k < Bjmap1[i + 1]; k++) sum += v[Bperm1[k]];
6753: Ba[i] = (imode == INSERT_VALUES ? 0.0 : Ba[i]) + sum;
6754: }
6755: PetscCall(PetscSFReduceEnd(coo->sf, MPIU_SCALAR, sendbuf, recvbuf, MPI_REPLACE));
6757: /* Add received remote entries to A and B */
6758: for (PetscCount i = 0; i < coo->Annz2; i++) {
6759: for (PetscCount k = Ajmap2[i]; k < Ajmap2[i + 1]; k++) Aa[Aimap2[i]] += recvbuf[Aperm2[k]];
6760: }
6761: for (PetscCount i = 0; i < coo->Bnnz2; i++) {
6762: for (PetscCount k = Bjmap2[i]; k < Bjmap2[i + 1]; k++) Ba[Bimap2[i]] += recvbuf[Bperm2[k]];
6763: }
6764: PetscCall(MatSeqAIJRestoreArray(A, &Aa));
6765: PetscCall(MatSeqAIJRestoreArray(B, &Ba));
6766: PetscFunctionReturn(PETSC_SUCCESS);
6767: }
6769: /*MC
6770: MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices.
6772: Options Database Keys:
6773: . -mat_type mpiaij - sets the matrix type to `MATMPIAIJ` during a call to `MatSetFromOptions()`
6775: Level: beginner
6777: Notes:
6778: `MatSetValues()` may be called for this matrix type with a `NULL` argument for the numerical values,
6779: in this case the values associated with the rows and columns one passes in are set to zero
6780: in the matrix
6782: `MatSetOptions`(,`MAT_STRUCTURE_ONLY`,`PETSC_TRUE`) may be called for this matrix type. In this no
6783: space is allocated for the nonzero entries and any entries passed with `MatSetValues()` are ignored
6785: .seealso: [](ch_matrices), `Mat`, `MATSEQAIJ`, `MATAIJ`, `MatCreateAIJ()`
6786: M*/
6787: PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat B)
6788: {
6789: Mat_MPIAIJ *b;
6790: PetscMPIInt size;
6792: PetscFunctionBegin;
6793: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
6795: PetscCall(PetscNew(&b));
6796: B->data = (void *)b;
6797: B->ops[0] = MatOps_Values;
6798: B->assembled = PETSC_FALSE;
6799: B->insertmode = NOT_SET_VALUES;
6800: b->size = size;
6802: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)B), &b->rank));
6804: /* build cache for off array entries formed */
6805: PetscCall(MatStashCreate_Private(PetscObjectComm((PetscObject)B), 1, &B->stash));
6807: b->donotstash = PETSC_FALSE;
6808: b->colmap = NULL;
6809: b->garray = NULL;
6810: b->roworiented = PETSC_TRUE;
6812: /* stuff used for matrix vector multiply */
6813: b->lvec = NULL;
6814: b->Mvctx = NULL;
6816: /* stuff for MatGetRow() */
6817: b->rowindices = NULL;
6818: b->rowvalues = NULL;
6819: b->getrowactive = PETSC_FALSE;
6821: /* flexible pointer used in CUSPARSE classes */
6822: b->spptr = NULL;
6824: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetUseScalableIncreaseOverlap_C", MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ));
6825: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatStoreValues_C", MatStoreValues_MPIAIJ));
6826: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatRetrieveValues_C", MatRetrieveValues_MPIAIJ));
6827: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatIsTranspose_C", MatIsTranspose_MPIAIJ));
6828: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetPreallocation_C", MatMPIAIJSetPreallocation_MPIAIJ));
6829: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatResetPreallocation_C", MatResetPreallocation_MPIAIJ));
6830: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatResetHash_C", MatResetHash_MPIAIJ));
6831: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetPreallocationCSR_C", MatMPIAIJSetPreallocationCSR_MPIAIJ));
6832: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDiagonalScaleLocal_C", MatDiagonalScaleLocal_MPIAIJ));
6833: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijperm_C", MatConvert_MPIAIJ_MPIAIJPERM));
6834: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijsell_C", MatConvert_MPIAIJ_MPIAIJSELL));
6835: #if defined(PETSC_HAVE_CUDA)
6836: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijcusparse_C", MatConvert_MPIAIJ_MPIAIJCUSPARSE));
6837: #endif
6838: #if defined(PETSC_HAVE_HIP)
6839: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijhipsparse_C", MatConvert_MPIAIJ_MPIAIJHIPSPARSE));
6840: #endif
6841: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
6842: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijkokkos_C", MatConvert_MPIAIJ_MPIAIJKokkos));
6843: #endif
6844: #if defined(PETSC_HAVE_MKL_SPARSE)
6845: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijmkl_C", MatConvert_MPIAIJ_MPIAIJMKL));
6846: #endif
6847: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijcrl_C", MatConvert_MPIAIJ_MPIAIJCRL));
6848: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpibaij_C", MatConvert_MPIAIJ_MPIBAIJ));
6849: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpisbaij_C", MatConvert_MPIAIJ_MPISBAIJ));
6850: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpidense_C", MatConvert_MPIAIJ_MPIDense));
6851: #if defined(PETSC_HAVE_ELEMENTAL)
6852: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_elemental_C", MatConvert_MPIAIJ_Elemental));
6853: #endif
6854: #if defined(PETSC_HAVE_SCALAPACK) && (defined(PETSC_USE_REAL_SINGLE) || defined(PETSC_USE_REAL_DOUBLE))
6855: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_scalapack_C", MatConvert_AIJ_ScaLAPACK));
6856: #endif
6857: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_is_C", MatConvert_XAIJ_IS));
6858: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpisell_C", MatConvert_MPIAIJ_MPISELL));
6859: #if defined(PETSC_HAVE_HYPRE)
6860: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_hypre_C", MatConvert_AIJ_HYPRE));
6861: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_transpose_mpiaij_mpiaij_C", MatProductSetFromOptions_Transpose_AIJ_AIJ));
6862: #endif
6863: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_is_mpiaij_C", MatProductSetFromOptions_IS_XAIJ));
6864: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_mpiaij_mpiaij_C", MatProductSetFromOptions_MPIAIJ));
6865: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetPreallocationCOO_C", MatSetPreallocationCOO_MPIAIJ));
6866: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetValuesCOO_C", MatSetValuesCOO_MPIAIJ));
6867: PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATMPIAIJ));
6868: PetscFunctionReturn(PETSC_SUCCESS);
6869: }
6871: /*@
6872: MatCreateMPIAIJWithSplitArrays - creates a `MATMPIAIJ` matrix using arrays that contain the "diagonal"
6873: and "off-diagonal" part of the matrix in CSR format.
6875: Collective
6877: Input Parameters:
6878: + comm - MPI communicator
6879: . m - number of local rows (Cannot be `PETSC_DECIDE`)
6880: . n - This value should be the same as the local size used in creating the
6881: x vector for the matrix-vector product $y = Ax$. (or `PETSC_DECIDE` to have
6882: calculated if `N` is given) For square matrices `n` is almost always `m`.
6883: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if `m` is given)
6884: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if `n` is given)
6885: . 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
6886: . j - column indices, which must be local, i.e., based off the start column of the diagonal portion
6887: . a - matrix values
6888: . 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
6889: . oj - column indices, which must be global, representing global columns in the `MATMPIAIJ` matrix
6890: - oa - matrix values
6892: Output Parameter:
6893: . mat - the matrix
6895: Level: advanced
6897: Notes:
6898: The `i`, `j`, and `a` arrays ARE NOT copied by this routine into the internal format used by PETSc (even in Fortran). The user
6899: must free the arrays once the matrix has been destroyed and not before.
6901: The `i` and `j` indices are 0 based
6903: See `MatCreateAIJ()` for the definition of "diagonal" and "off-diagonal" portion of the matrix
6905: This sets local rows and cannot be used to set off-processor values.
6907: Use of this routine is discouraged because it is inflexible and cumbersome to use. It is extremely rare that a
6908: legacy application natively assembles into exactly this split format. The code to do so is nontrivial and does
6909: not easily support in-place reassembly. It is recommended to use MatSetValues() (or a variant thereof) because
6910: the resulting assembly is easier to implement, will work with any matrix format, and the user does not have to
6911: keep track of the underlying array. Use `MatSetOption`(A,`MAT_NO_OFF_PROC_ENTRIES`,`PETSC_TRUE`) to disable all
6912: communication if it is known that only local entries will be set.
6914: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
6915: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithArrays()`
6916: @*/
6917: 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)
6918: {
6919: Mat_MPIAIJ *maij;
6921: PetscFunctionBegin;
6922: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
6923: PetscCheck(i[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
6924: PetscCheck(oi[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "oi (row indices) must start with 0");
6925: PetscCall(MatCreate(comm, mat));
6926: PetscCall(MatSetSizes(*mat, m, n, M, N));
6927: PetscCall(MatSetType(*mat, MATMPIAIJ));
6928: maij = (Mat_MPIAIJ *)(*mat)->data;
6930: (*mat)->preallocated = PETSC_TRUE;
6932: PetscCall(PetscLayoutSetUp((*mat)->rmap));
6933: PetscCall(PetscLayoutSetUp((*mat)->cmap));
6935: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, n, i, j, a, &maij->A));
6936: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, (*mat)->cmap->N, oi, oj, oa, &maij->B));
6938: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
6939: PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY));
6940: PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY));
6941: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_FALSE));
6942: PetscCall(MatSetOption(*mat, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
6943: PetscFunctionReturn(PETSC_SUCCESS);
6944: }
6946: typedef struct {
6947: Mat *mp; /* intermediate products */
6948: PetscBool *mptmp; /* is the intermediate product temporary ? */
6949: PetscInt cp; /* number of intermediate products */
6951: /* support for MatGetBrowsOfAoCols_MPIAIJ for P_oth */
6952: PetscInt *startsj_s, *startsj_r;
6953: PetscScalar *bufa;
6954: Mat P_oth;
6956: /* may take advantage of merging product->B */
6957: Mat Bloc; /* B-local by merging diag and off-diag */
6959: /* cusparse does not have support to split between symbolic and numeric phases.
6960: When api_user is true, we don't need to update the numerical values
6961: of the temporary storage */
6962: PetscBool reusesym;
6964: /* support for COO values insertion */
6965: PetscScalar *coo_v, *coo_w; /* store on-process and off-process COO scalars, and used as MPI recv/send buffers respectively */
6966: PetscInt **own; /* own[i] points to address of on-process COO indices for Mat mp[i] */
6967: PetscInt **off; /* off[i] points to address of off-process COO indices for Mat mp[i] */
6968: PetscBool hasoffproc; /* if true, have off-process values insertion (i.e. AtB or PtAP) */
6969: PetscSF sf; /* used for non-local values insertion and memory malloc */
6970: PetscMemType mtype;
6972: /* customization */
6973: PetscBool abmerge;
6974: PetscBool P_oth_bind;
6975: } MatMatMPIAIJBACKEND;
6977: static PetscErrorCode MatProductCtxDestroy_MatMatMPIAIJBACKEND(void **data)
6978: {
6979: MatMatMPIAIJBACKEND *mmdata = *(MatMatMPIAIJBACKEND **)data;
6980: PetscInt i;
6982: PetscFunctionBegin;
6983: PetscCall(PetscFree2(mmdata->startsj_s, mmdata->startsj_r));
6984: PetscCall(PetscFree(mmdata->bufa));
6985: PetscCall(PetscSFFree(mmdata->sf, mmdata->mtype, mmdata->coo_v));
6986: PetscCall(PetscSFFree(mmdata->sf, mmdata->mtype, mmdata->coo_w));
6987: PetscCall(MatDestroy(&mmdata->P_oth));
6988: PetscCall(MatDestroy(&mmdata->Bloc));
6989: PetscCall(PetscSFDestroy(&mmdata->sf));
6990: for (i = 0; i < mmdata->cp; i++) PetscCall(MatDestroy(&mmdata->mp[i]));
6991: PetscCall(PetscFree2(mmdata->mp, mmdata->mptmp));
6992: PetscCall(PetscFree(mmdata->own[0]));
6993: PetscCall(PetscFree(mmdata->own));
6994: PetscCall(PetscFree(mmdata->off[0]));
6995: PetscCall(PetscFree(mmdata->off));
6996: PetscCall(PetscFree(mmdata));
6997: PetscFunctionReturn(PETSC_SUCCESS);
6998: }
7000: /* Copy selected n entries with indices in idx[] of A to v[].
7001: If idx is NULL, copy the whole data array of A to v[]
7002: */
7003: static PetscErrorCode MatSeqAIJCopySubArray(Mat A, PetscInt n, const PetscInt idx[], PetscScalar v[])
7004: {
7005: PetscErrorCode (*f)(Mat, PetscInt, const PetscInt[], PetscScalar[]);
7007: PetscFunctionBegin;
7008: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatSeqAIJCopySubArray_C", &f));
7009: if (f) {
7010: PetscCall((*f)(A, n, idx, v));
7011: } else {
7012: const PetscScalar *vv;
7014: PetscCall(MatSeqAIJGetArrayRead(A, &vv));
7015: if (n && idx) {
7016: PetscScalar *w = v;
7017: const PetscInt *oi = idx;
7018: PetscInt j;
7020: for (j = 0; j < n; j++) *w++ = vv[*oi++];
7021: } else {
7022: PetscCall(PetscArraycpy(v, vv, n));
7023: }
7024: PetscCall(MatSeqAIJRestoreArrayRead(A, &vv));
7025: }
7026: PetscFunctionReturn(PETSC_SUCCESS);
7027: }
7029: static PetscErrorCode MatProductNumeric_MPIAIJBACKEND(Mat C)
7030: {
7031: MatMatMPIAIJBACKEND *mmdata;
7032: PetscInt i, n_d, n_o;
7034: PetscFunctionBegin;
7035: MatCheckProduct(C, 1);
7036: PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data empty");
7037: mmdata = (MatMatMPIAIJBACKEND *)C->product->data;
7038: if (!mmdata->reusesym) { /* update temporary matrices */
7039: 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));
7040: if (mmdata->Bloc) PetscCall(MatMPIAIJGetLocalMatMerge(C->product->B, MAT_REUSE_MATRIX, NULL, &mmdata->Bloc));
7041: }
7042: mmdata->reusesym = PETSC_FALSE;
7044: for (i = 0; i < mmdata->cp; i++) {
7045: 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]);
7046: PetscCall((*mmdata->mp[i]->ops->productnumeric)(mmdata->mp[i]));
7047: }
7048: for (i = 0, n_d = 0, n_o = 0; i < mmdata->cp; i++) {
7049: PetscInt noff;
7051: PetscCall(PetscIntCast(mmdata->off[i + 1] - mmdata->off[i], &noff));
7052: if (mmdata->mptmp[i]) continue;
7053: if (noff) {
7054: PetscInt nown;
7056: PetscCall(PetscIntCast(mmdata->own[i + 1] - mmdata->own[i], &nown));
7057: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], noff, mmdata->off[i], mmdata->coo_w + n_o));
7058: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], nown, mmdata->own[i], mmdata->coo_v + n_d));
7059: n_o += noff;
7060: n_d += nown;
7061: } else {
7062: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mmdata->mp[i]->data;
7064: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], mm->nz, NULL, mmdata->coo_v + n_d));
7065: n_d += mm->nz;
7066: }
7067: }
7068: if (mmdata->hasoffproc) { /* offprocess insertion */
7069: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_SCALAR, mmdata->coo_w, mmdata->coo_v + n_d));
7070: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_SCALAR, mmdata->coo_w, mmdata->coo_v + n_d));
7071: }
7072: PetscCall(MatSetValuesCOO(C, mmdata->coo_v, INSERT_VALUES));
7073: PetscFunctionReturn(PETSC_SUCCESS);
7074: }
7076: /* Support for Pt * A, A * P, or Pt * A * P */
7077: #define MAX_NUMBER_INTERMEDIATE 4
7078: PetscErrorCode MatProductSymbolic_MPIAIJBACKEND(Mat C)
7079: {
7080: Mat_Product *product = C->product;
7081: Mat A, P, mp[MAX_NUMBER_INTERMEDIATE]; /* A, P and a series of intermediate matrices */
7082: Mat_MPIAIJ *a, *p;
7083: MatMatMPIAIJBACKEND *mmdata;
7084: ISLocalToGlobalMapping P_oth_l2g = NULL;
7085: IS glob = NULL;
7086: const char *prefix;
7087: char pprefix[256];
7088: const PetscInt *globidx, *P_oth_idx;
7089: PetscInt i, j, cp, m, n, M, N, *coo_i, *coo_j;
7090: PetscCount ncoo, ncoo_d, ncoo_o, ncoo_oown;
7091: PetscInt cmapt[MAX_NUMBER_INTERMEDIATE], rmapt[MAX_NUMBER_INTERMEDIATE]; /* col/row map type for each Mat in mp[]. */
7092: /* type-0: consecutive, start from 0; type-1: consecutive with */
7093: /* a base offset; type-2: sparse with a local to global map table */
7094: const PetscInt *cmapa[MAX_NUMBER_INTERMEDIATE], *rmapa[MAX_NUMBER_INTERMEDIATE]; /* col/row local to global map array (table) for type-2 map type */
7096: MatProductType ptype;
7097: PetscBool mptmp[MAX_NUMBER_INTERMEDIATE], hasoffproc = PETSC_FALSE, iscuda, iship, iskokk;
7098: PetscMPIInt size;
7100: PetscFunctionBegin;
7101: MatCheckProduct(C, 1);
7102: PetscCheck(!product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data not empty");
7103: ptype = product->type;
7104: if (product->A->symmetric == PETSC_BOOL3_TRUE && ptype == MATPRODUCT_AtB) {
7105: ptype = MATPRODUCT_AB;
7106: product->symbolic_used_the_fact_A_is_symmetric = PETSC_TRUE;
7107: }
7108: switch (ptype) {
7109: case MATPRODUCT_AB:
7110: A = product->A;
7111: P = product->B;
7112: m = A->rmap->n;
7113: n = P->cmap->n;
7114: M = A->rmap->N;
7115: N = P->cmap->N;
7116: hasoffproc = PETSC_FALSE; /* will not scatter mat product values to other processes */
7117: break;
7118: case MATPRODUCT_AtB:
7119: P = product->A;
7120: A = product->B;
7121: m = P->cmap->n;
7122: n = A->cmap->n;
7123: M = P->cmap->N;
7124: N = A->cmap->N;
7125: hasoffproc = PETSC_TRUE;
7126: break;
7127: case MATPRODUCT_PtAP:
7128: A = product->A;
7129: P = product->B;
7130: m = P->cmap->n;
7131: n = P->cmap->n;
7132: M = P->cmap->N;
7133: N = P->cmap->N;
7134: hasoffproc = PETSC_TRUE;
7135: break;
7136: default:
7137: SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Not for product type %s", MatProductTypes[ptype]);
7138: }
7139: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)C), &size));
7140: if (size == 1) hasoffproc = PETSC_FALSE;
7142: /* defaults */
7143: for (i = 0; i < MAX_NUMBER_INTERMEDIATE; i++) {
7144: mp[i] = NULL;
7145: mptmp[i] = PETSC_FALSE;
7146: rmapt[i] = -1;
7147: cmapt[i] = -1;
7148: rmapa[i] = NULL;
7149: cmapa[i] = NULL;
7150: }
7152: /* customization */
7153: PetscCall(PetscNew(&mmdata));
7154: mmdata->reusesym = product->api_user;
7155: if (ptype == MATPRODUCT_AB) {
7156: if (product->api_user) {
7157: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatMatMult", "Mat");
7158: PetscCall(PetscOptionsBool("-matmatmult_backend_mergeB", "Merge product->B local matrices", "MatMatMult", mmdata->abmerge, &mmdata->abmerge, NULL));
7159: PetscCall(PetscOptionsBool("-matmatmult_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7160: PetscOptionsEnd();
7161: } else {
7162: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_AB", "Mat");
7163: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_mergeB", "Merge product->B local matrices", "MatMatMult", mmdata->abmerge, &mmdata->abmerge, NULL));
7164: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7165: PetscOptionsEnd();
7166: }
7167: } else if (ptype == MATPRODUCT_PtAP) {
7168: if (product->api_user) {
7169: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatPtAP", "Mat");
7170: PetscCall(PetscOptionsBool("-matptap_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7171: PetscOptionsEnd();
7172: } else {
7173: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_PtAP", "Mat");
7174: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7175: PetscOptionsEnd();
7176: }
7177: }
7178: a = (Mat_MPIAIJ *)A->data;
7179: p = (Mat_MPIAIJ *)P->data;
7180: PetscCall(MatSetSizes(C, m, n, M, N));
7181: PetscCall(PetscLayoutSetUp(C->rmap));
7182: PetscCall(PetscLayoutSetUp(C->cmap));
7183: PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
7184: PetscCall(MatGetOptionsPrefix(C, &prefix));
7186: cp = 0;
7187: switch (ptype) {
7188: case MATPRODUCT_AB: /* A * P */
7189: PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &mmdata->startsj_s, &mmdata->startsj_r, &mmdata->bufa, &mmdata->P_oth));
7191: /* A_diag * P_local (merged or not) */
7192: if (mmdata->abmerge) { /* P's diagonal and off-diag blocks are merged to one matrix, then multiplied by A_diag */
7193: /* P is product->B */
7194: PetscCall(MatMPIAIJGetLocalMatMerge(P, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7195: PetscCall(MatProductCreate(a->A, mmdata->Bloc, NULL, &mp[cp]));
7196: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7197: PetscCall(MatProductSetFill(mp[cp], product->fill));
7198: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7199: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7200: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7201: mp[cp]->product->api_user = product->api_user;
7202: PetscCall(MatProductSetFromOptions(mp[cp]));
7203: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7204: PetscCall(ISGetIndices(glob, &globidx));
7205: rmapt[cp] = 1;
7206: cmapt[cp] = 2;
7207: cmapa[cp] = globidx;
7208: mptmp[cp] = PETSC_FALSE;
7209: cp++;
7210: } else { /* A_diag * P_diag and A_diag * P_off */
7211: PetscCall(MatProductCreate(a->A, p->A, NULL, &mp[cp]));
7212: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7213: PetscCall(MatProductSetFill(mp[cp], product->fill));
7214: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7215: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7216: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7217: mp[cp]->product->api_user = product->api_user;
7218: PetscCall(MatProductSetFromOptions(mp[cp]));
7219: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7220: rmapt[cp] = 1;
7221: cmapt[cp] = 1;
7222: mptmp[cp] = PETSC_FALSE;
7223: cp++;
7224: PetscCall(MatProductCreate(a->A, p->B, NULL, &mp[cp]));
7225: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7226: PetscCall(MatProductSetFill(mp[cp], product->fill));
7227: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7228: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7229: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7230: mp[cp]->product->api_user = product->api_user;
7231: PetscCall(MatProductSetFromOptions(mp[cp]));
7232: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7233: rmapt[cp] = 1;
7234: cmapt[cp] = 2;
7235: cmapa[cp] = p->garray;
7236: mptmp[cp] = PETSC_FALSE;
7237: cp++;
7238: }
7240: /* A_off * P_other */
7241: if (mmdata->P_oth) {
7242: PetscCall(MatSeqAIJCompactOutExtraColumns_SeqAIJ(mmdata->P_oth, &P_oth_l2g)); /* make P_oth use local col ids */
7243: PetscCall(ISLocalToGlobalMappingGetIndices(P_oth_l2g, &P_oth_idx));
7244: PetscCall(MatSetType(mmdata->P_oth, ((PetscObject)a->B)->type_name));
7245: PetscCall(MatBindToCPU(mmdata->P_oth, mmdata->P_oth_bind));
7246: PetscCall(MatProductCreate(a->B, mmdata->P_oth, NULL, &mp[cp]));
7247: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7248: PetscCall(MatProductSetFill(mp[cp], product->fill));
7249: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7250: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7251: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7252: mp[cp]->product->api_user = product->api_user;
7253: PetscCall(MatProductSetFromOptions(mp[cp]));
7254: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7255: rmapt[cp] = 1;
7256: cmapt[cp] = 2;
7257: cmapa[cp] = P_oth_idx;
7258: mptmp[cp] = PETSC_FALSE;
7259: cp++;
7260: }
7261: break;
7263: case MATPRODUCT_AtB: /* (P^t * A): P_diag * A_loc + P_off * A_loc */
7264: /* A is product->B */
7265: PetscCall(MatMPIAIJGetLocalMatMerge(A, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7266: if (A == P) { /* when A==P, we can take advantage of the already merged mmdata->Bloc */
7267: PetscCall(MatProductCreate(mmdata->Bloc, mmdata->Bloc, NULL, &mp[cp]));
7268: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7269: PetscCall(MatProductSetFill(mp[cp], product->fill));
7270: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7271: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7272: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7273: mp[cp]->product->api_user = product->api_user;
7274: PetscCall(MatProductSetFromOptions(mp[cp]));
7275: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7276: PetscCall(ISGetIndices(glob, &globidx));
7277: rmapt[cp] = 2;
7278: rmapa[cp] = globidx;
7279: cmapt[cp] = 2;
7280: cmapa[cp] = globidx;
7281: mptmp[cp] = PETSC_FALSE;
7282: cp++;
7283: } else {
7284: PetscCall(MatProductCreate(p->A, mmdata->Bloc, NULL, &mp[cp]));
7285: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7286: PetscCall(MatProductSetFill(mp[cp], product->fill));
7287: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7288: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7289: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7290: mp[cp]->product->api_user = product->api_user;
7291: PetscCall(MatProductSetFromOptions(mp[cp]));
7292: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7293: PetscCall(ISGetIndices(glob, &globidx));
7294: rmapt[cp] = 1;
7295: cmapt[cp] = 2;
7296: cmapa[cp] = globidx;
7297: mptmp[cp] = PETSC_FALSE;
7298: cp++;
7299: PetscCall(MatProductCreate(p->B, mmdata->Bloc, NULL, &mp[cp]));
7300: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7301: PetscCall(MatProductSetFill(mp[cp], product->fill));
7302: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7303: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7304: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7305: mp[cp]->product->api_user = product->api_user;
7306: PetscCall(MatProductSetFromOptions(mp[cp]));
7307: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7308: rmapt[cp] = 2;
7309: rmapa[cp] = p->garray;
7310: cmapt[cp] = 2;
7311: cmapa[cp] = globidx;
7312: mptmp[cp] = PETSC_FALSE;
7313: cp++;
7314: }
7315: break;
7316: case MATPRODUCT_PtAP:
7317: PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &mmdata->startsj_s, &mmdata->startsj_r, &mmdata->bufa, &mmdata->P_oth));
7318: /* P is product->B */
7319: PetscCall(MatMPIAIJGetLocalMatMerge(P, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7320: PetscCall(MatProductCreate(a->A, mmdata->Bloc, NULL, &mp[cp]));
7321: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_PtAP));
7322: PetscCall(MatProductSetFill(mp[cp], product->fill));
7323: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7324: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7325: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7326: mp[cp]->product->api_user = product->api_user;
7327: PetscCall(MatProductSetFromOptions(mp[cp]));
7328: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7329: PetscCall(ISGetIndices(glob, &globidx));
7330: rmapt[cp] = 2;
7331: rmapa[cp] = globidx;
7332: cmapt[cp] = 2;
7333: cmapa[cp] = globidx;
7334: mptmp[cp] = PETSC_FALSE;
7335: cp++;
7336: if (mmdata->P_oth) {
7337: PetscCall(MatSeqAIJCompactOutExtraColumns_SeqAIJ(mmdata->P_oth, &P_oth_l2g));
7338: PetscCall(ISLocalToGlobalMappingGetIndices(P_oth_l2g, &P_oth_idx));
7339: PetscCall(MatSetType(mmdata->P_oth, ((PetscObject)a->B)->type_name));
7340: PetscCall(MatBindToCPU(mmdata->P_oth, mmdata->P_oth_bind));
7341: PetscCall(MatProductCreate(a->B, mmdata->P_oth, NULL, &mp[cp]));
7342: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7343: PetscCall(MatProductSetFill(mp[cp], product->fill));
7344: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7345: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7346: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7347: mp[cp]->product->api_user = product->api_user;
7348: PetscCall(MatProductSetFromOptions(mp[cp]));
7349: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7350: mptmp[cp] = PETSC_TRUE;
7351: cp++;
7352: PetscCall(MatProductCreate(mmdata->Bloc, mp[1], NULL, &mp[cp]));
7353: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7354: PetscCall(MatProductSetFill(mp[cp], product->fill));
7355: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7356: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7357: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7358: mp[cp]->product->api_user = product->api_user;
7359: PetscCall(MatProductSetFromOptions(mp[cp]));
7360: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7361: rmapt[cp] = 2;
7362: rmapa[cp] = globidx;
7363: cmapt[cp] = 2;
7364: cmapa[cp] = P_oth_idx;
7365: mptmp[cp] = PETSC_FALSE;
7366: cp++;
7367: }
7368: break;
7369: default:
7370: SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Not for product type %s", MatProductTypes[ptype]);
7371: }
7372: /* sanity check */
7373: if (size > 1)
7374: 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);
7376: PetscCall(PetscMalloc2(cp, &mmdata->mp, cp, &mmdata->mptmp));
7377: for (i = 0; i < cp; i++) {
7378: mmdata->mp[i] = mp[i];
7379: mmdata->mptmp[i] = mptmp[i];
7380: }
7381: mmdata->cp = cp;
7382: C->product->data = mmdata;
7383: C->product->destroy = MatProductCtxDestroy_MatMatMPIAIJBACKEND;
7384: C->ops->productnumeric = MatProductNumeric_MPIAIJBACKEND;
7386: /* memory type */
7387: mmdata->mtype = PETSC_MEMTYPE_HOST;
7388: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iscuda, MATSEQAIJCUSPARSE, MATMPIAIJCUSPARSE, ""));
7389: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iship, MATSEQAIJHIPSPARSE, MATMPIAIJHIPSPARSE, ""));
7390: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iskokk, MATSEQAIJKOKKOS, MATMPIAIJKOKKOS, ""));
7391: if (iscuda) mmdata->mtype = PETSC_MEMTYPE_CUDA;
7392: else if (iship) mmdata->mtype = PETSC_MEMTYPE_HIP;
7393: else if (iskokk) mmdata->mtype = PETSC_MEMTYPE_KOKKOS;
7395: /* prepare coo coordinates for values insertion */
7397: /* count total nonzeros of those intermediate seqaij Mats
7398: ncoo_d: # of nonzeros of matrices that do not have offproc entries
7399: ncoo_o: # of nonzeros (of matrices that might have offproc entries) that will be inserted to remote procs
7400: ncoo_oown: # of nonzeros (of matrices that might have offproc entries) that will be inserted locally
7401: */
7402: for (cp = 0, ncoo_d = 0, ncoo_o = 0, ncoo_oown = 0; cp < mmdata->cp; cp++) {
7403: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7404: if (mptmp[cp]) continue;
7405: if (rmapt[cp] == 2 && hasoffproc) { /* the rows need to be scatter to all processes (might include self) */
7406: const PetscInt *rmap = rmapa[cp];
7407: const PetscInt mr = mp[cp]->rmap->n;
7408: const PetscInt rs = C->rmap->rstart;
7409: const PetscInt re = C->rmap->rend;
7410: const PetscInt *ii = mm->i;
7411: for (i = 0; i < mr; i++) {
7412: const PetscInt gr = rmap[i];
7413: const PetscInt nz = ii[i + 1] - ii[i];
7414: if (gr < rs || gr >= re) ncoo_o += nz; /* this row is offproc */
7415: else ncoo_oown += nz; /* this row is local */
7416: }
7417: } else ncoo_d += mm->nz;
7418: }
7420: /*
7421: ncoo: total number of nonzeros (including those inserted by remote procs) belonging to this proc
7423: ncoo = ncoo_d + ncoo_oown + ncoo2, which ncoo2 is number of nonzeros inserted to me by other procs.
7425: off[0] points to a big index array, which is shared by off[1,2,...]. Similarly, for own[0].
7427: off[p]: points to the segment for matrix mp[p], storing location of nonzeros that mp[p] will insert to others
7428: own[p]: points to the segment for matrix mp[p], storing location of nonzeros that mp[p] will insert locally
7429: so, off[p+1]-off[p] is the number of nonzeros that mp[p] will send to others.
7431: coo_i/j/v[]: [ncoo] row/col/val of nonzeros belonging to this proc.
7432: 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.
7433: */
7434: PetscCall(PetscCalloc1(mmdata->cp + 1, &mmdata->off)); /* +1 to make a csr-like data structure */
7435: PetscCall(PetscCalloc1(mmdata->cp + 1, &mmdata->own));
7437: /* gather (i,j) of nonzeros inserted by remote procs */
7438: if (hasoffproc) {
7439: PetscSF msf;
7440: PetscInt ncoo2, *coo_i2, *coo_j2;
7442: PetscCall(PetscMalloc1(ncoo_o, &mmdata->off[0]));
7443: PetscCall(PetscMalloc1(ncoo_oown, &mmdata->own[0]));
7444: PetscCall(PetscMalloc2(ncoo_o, &coo_i, ncoo_o, &coo_j)); /* to collect (i,j) of entries to be sent to others */
7446: for (cp = 0, ncoo_o = 0; cp < mmdata->cp; cp++) {
7447: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7448: PetscInt *idxoff = mmdata->off[cp];
7449: PetscInt *idxown = mmdata->own[cp];
7450: if (!mptmp[cp] && rmapt[cp] == 2) { /* row map is sparse */
7451: const PetscInt *rmap = rmapa[cp];
7452: const PetscInt *cmap = cmapa[cp];
7453: const PetscInt *ii = mm->i;
7454: PetscInt *coi = coo_i + ncoo_o;
7455: PetscInt *coj = coo_j + ncoo_o;
7456: const PetscInt mr = mp[cp]->rmap->n;
7457: const PetscInt rs = C->rmap->rstart;
7458: const PetscInt re = C->rmap->rend;
7459: const PetscInt cs = C->cmap->rstart;
7460: for (i = 0; i < mr; i++) {
7461: const PetscInt *jj = mm->j + ii[i];
7462: const PetscInt gr = rmap[i];
7463: const PetscInt nz = ii[i + 1] - ii[i];
7464: if (gr < rs || gr >= re) { /* this is an offproc row */
7465: for (j = ii[i]; j < ii[i + 1]; j++) {
7466: *coi++ = gr;
7467: *idxoff++ = j;
7468: }
7469: if (!cmapt[cp]) { /* already global */
7470: for (j = 0; j < nz; j++) *coj++ = jj[j];
7471: } else if (cmapt[cp] == 1) { /* local to global for owned columns of C */
7472: for (j = 0; j < nz; j++) *coj++ = jj[j] + cs;
7473: } else { /* offdiag */
7474: for (j = 0; j < nz; j++) *coj++ = cmap[jj[j]];
7475: }
7476: ncoo_o += nz;
7477: } else { /* this is a local row */
7478: for (j = ii[i]; j < ii[i + 1]; j++) *idxown++ = j;
7479: }
7480: }
7481: }
7482: mmdata->off[cp + 1] = idxoff;
7483: mmdata->own[cp + 1] = idxown;
7484: }
7486: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)C), &mmdata->sf));
7487: PetscInt incoo_o;
7488: PetscCall(PetscIntCast(ncoo_o, &incoo_o));
7489: PetscCall(PetscSFSetGraphLayout(mmdata->sf, C->rmap, incoo_o /*nleaves*/, NULL /*ilocal*/, PETSC_OWN_POINTER, coo_i));
7490: PetscCall(PetscSFGetMultiSF(mmdata->sf, &msf));
7491: PetscCall(PetscSFGetGraph(msf, &ncoo2 /*nroots*/, NULL, NULL, NULL));
7492: ncoo = ncoo_d + ncoo_oown + ncoo2;
7493: PetscCall(PetscMalloc2(ncoo, &coo_i2, ncoo, &coo_j2));
7494: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_INT, coo_i, coo_i2 + ncoo_d + ncoo_oown)); /* put (i,j) of remote nonzeros at back */
7495: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_INT, coo_i, coo_i2 + ncoo_d + ncoo_oown));
7496: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_INT, coo_j, coo_j2 + ncoo_d + ncoo_oown));
7497: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_INT, coo_j, coo_j2 + ncoo_d + ncoo_oown));
7498: PetscCall(PetscFree2(coo_i, coo_j));
7499: /* allocate MPI send buffer to collect nonzero values to be sent to remote procs */
7500: PetscCall(PetscSFMalloc(mmdata->sf, mmdata->mtype, ncoo_o * sizeof(PetscScalar), (void **)&mmdata->coo_w));
7501: coo_i = coo_i2;
7502: coo_j = coo_j2;
7503: } else { /* no offproc values insertion */
7504: ncoo = ncoo_d;
7505: PetscCall(PetscMalloc2(ncoo, &coo_i, ncoo, &coo_j));
7507: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)C), &mmdata->sf));
7508: PetscCall(PetscSFSetGraph(mmdata->sf, 0, 0, NULL, PETSC_OWN_POINTER, NULL, PETSC_OWN_POINTER));
7509: PetscCall(PetscSFSetUp(mmdata->sf));
7510: }
7511: mmdata->hasoffproc = hasoffproc;
7513: /* gather (i,j) of nonzeros inserted locally */
7514: for (cp = 0, ncoo_d = 0; cp < mmdata->cp; cp++) {
7515: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7516: PetscInt *coi = coo_i + ncoo_d;
7517: PetscInt *coj = coo_j + ncoo_d;
7518: const PetscInt *jj = mm->j;
7519: const PetscInt *ii = mm->i;
7520: const PetscInt *cmap = cmapa[cp];
7521: const PetscInt *rmap = rmapa[cp];
7522: const PetscInt mr = mp[cp]->rmap->n;
7523: const PetscInt rs = C->rmap->rstart;
7524: const PetscInt re = C->rmap->rend;
7525: const PetscInt cs = C->cmap->rstart;
7527: if (mptmp[cp]) continue;
7528: if (rmapt[cp] == 1) { /* consecutive rows */
7529: /* fill coo_i */
7530: for (i = 0; i < mr; i++) {
7531: const PetscInt gr = i + rs;
7532: for (j = ii[i]; j < ii[i + 1]; j++) coi[j] = gr;
7533: }
7534: /* fill coo_j */
7535: if (!cmapt[cp]) { /* type-0, already global */
7536: PetscCall(PetscArraycpy(coj, jj, mm->nz));
7537: } else if (cmapt[cp] == 1) { /* type-1, local to global for consecutive columns of C */
7538: for (j = 0; j < mm->nz; j++) coj[j] = jj[j] + cs; /* lid + col start */
7539: } else { /* type-2, local to global for sparse columns */
7540: for (j = 0; j < mm->nz; j++) coj[j] = cmap[jj[j]];
7541: }
7542: ncoo_d += mm->nz;
7543: } else if (rmapt[cp] == 2) { /* sparse rows */
7544: for (i = 0; i < mr; i++) {
7545: const PetscInt *jj = mm->j + ii[i];
7546: const PetscInt gr = rmap[i];
7547: const PetscInt nz = ii[i + 1] - ii[i];
7548: if (gr >= rs && gr < re) { /* local rows */
7549: for (j = ii[i]; j < ii[i + 1]; j++) *coi++ = gr;
7550: if (!cmapt[cp]) { /* type-0, already global */
7551: for (j = 0; j < nz; j++) *coj++ = jj[j];
7552: } else if (cmapt[cp] == 1) { /* local to global for owned columns of C */
7553: for (j = 0; j < nz; j++) *coj++ = jj[j] + cs;
7554: } else { /* type-2, local to global for sparse columns */
7555: for (j = 0; j < nz; j++) *coj++ = cmap[jj[j]];
7556: }
7557: ncoo_d += nz;
7558: }
7559: }
7560: }
7561: }
7562: if (glob) PetscCall(ISRestoreIndices(glob, &globidx));
7563: PetscCall(ISDestroy(&glob));
7564: if (P_oth_l2g) PetscCall(ISLocalToGlobalMappingRestoreIndices(P_oth_l2g, &P_oth_idx));
7565: PetscCall(ISLocalToGlobalMappingDestroy(&P_oth_l2g));
7566: /* allocate an array to store all nonzeros (inserted locally or remotely) belonging to this proc */
7567: PetscCall(PetscSFMalloc(mmdata->sf, mmdata->mtype, ncoo * sizeof(PetscScalar), (void **)&mmdata->coo_v));
7569: /* set block sizes */
7570: A = product->A;
7571: P = product->B;
7572: switch (ptype) {
7573: case MATPRODUCT_PtAP:
7574: PetscCall(MatSetBlockSizes(C, P->cmap->bs, P->cmap->bs));
7575: break;
7576: case MATPRODUCT_RARt:
7577: PetscCall(MatSetBlockSizes(C, P->rmap->bs, P->rmap->bs));
7578: break;
7579: case MATPRODUCT_ABC:
7580: PetscCall(MatSetBlockSizesFromMats(C, A, product->C));
7581: break;
7582: case MATPRODUCT_AB:
7583: PetscCall(MatSetBlockSizesFromMats(C, A, P));
7584: break;
7585: case MATPRODUCT_AtB:
7586: PetscCall(MatSetBlockSizes(C, A->cmap->bs, P->cmap->bs));
7587: break;
7588: case MATPRODUCT_ABt:
7589: PetscCall(MatSetBlockSizes(C, A->rmap->bs, P->rmap->bs));
7590: break;
7591: default:
7592: SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Not for ProductType %s", MatProductTypes[ptype]);
7593: }
7595: /* preallocate with COO data */
7596: PetscCall(MatSetPreallocationCOO(C, ncoo, coo_i, coo_j));
7597: PetscCall(PetscFree2(coo_i, coo_j));
7598: PetscFunctionReturn(PETSC_SUCCESS);
7599: }
7601: PetscErrorCode MatProductSetFromOptions_MPIAIJBACKEND(Mat mat)
7602: {
7603: Mat_Product *product = mat->product;
7604: #if defined(PETSC_HAVE_DEVICE)
7605: PetscBool match = PETSC_FALSE;
7606: PetscBool usecpu = PETSC_FALSE;
7607: #else
7608: PetscBool match = PETSC_TRUE;
7609: #endif
7611: PetscFunctionBegin;
7612: MatCheckProduct(mat, 1);
7613: #if defined(PETSC_HAVE_DEVICE)
7614: if (!product->A->boundtocpu && !product->B->boundtocpu) PetscCall(PetscObjectTypeCompare((PetscObject)product->B, ((PetscObject)product->A)->type_name, &match));
7615: if (match) { /* we can always fallback to the CPU if requested */
7616: switch (product->type) {
7617: case MATPRODUCT_AB:
7618: if (product->api_user) {
7619: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatMatMult", "Mat");
7620: PetscCall(PetscOptionsBool("-matmatmult_backend_cpu", "Use CPU code", "MatMatMult", usecpu, &usecpu, NULL));
7621: PetscOptionsEnd();
7622: } else {
7623: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_AB", "Mat");
7624: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatMatMult", usecpu, &usecpu, NULL));
7625: PetscOptionsEnd();
7626: }
7627: break;
7628: case MATPRODUCT_AtB:
7629: if (product->api_user) {
7630: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatTransposeMatMult", "Mat");
7631: PetscCall(PetscOptionsBool("-mattransposematmult_backend_cpu", "Use CPU code", "MatTransposeMatMult", usecpu, &usecpu, NULL));
7632: PetscOptionsEnd();
7633: } else {
7634: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_AtB", "Mat");
7635: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatTransposeMatMult", usecpu, &usecpu, NULL));
7636: PetscOptionsEnd();
7637: }
7638: break;
7639: case MATPRODUCT_PtAP:
7640: if (product->api_user) {
7641: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatPtAP", "Mat");
7642: PetscCall(PetscOptionsBool("-matptap_backend_cpu", "Use CPU code", "MatPtAP", usecpu, &usecpu, NULL));
7643: PetscOptionsEnd();
7644: } else {
7645: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_PtAP", "Mat");
7646: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatPtAP", usecpu, &usecpu, NULL));
7647: PetscOptionsEnd();
7648: }
7649: break;
7650: default:
7651: break;
7652: }
7653: match = (PetscBool)!usecpu;
7654: }
7655: #endif
7656: if (match) {
7657: switch (product->type) {
7658: case MATPRODUCT_AB:
7659: case MATPRODUCT_AtB:
7660: case MATPRODUCT_PtAP:
7661: mat->ops->productsymbolic = MatProductSymbolic_MPIAIJBACKEND;
7662: break;
7663: default:
7664: break;
7665: }
7666: }
7667: /* fallback to MPIAIJ ops */
7668: if (!mat->ops->productsymbolic) PetscCall(MatProductSetFromOptions_MPIAIJ(mat));
7669: PetscFunctionReturn(PETSC_SUCCESS);
7670: }
7672: /*
7673: Produces a set of block column indices of the matrix row, one for each block represented in the original row
7675: n - the number of block indices in cc[]
7676: cc - the block indices (must be large enough to contain the indices)
7677: */
7678: static inline PetscErrorCode MatCollapseRow(Mat Amat, PetscInt row, PetscInt bs, PetscInt *n, PetscInt *cc)
7679: {
7680: PetscInt cnt = -1, nidx, j;
7681: const PetscInt *idx;
7683: PetscFunctionBegin;
7684: PetscCall(MatGetRow(Amat, row, &nidx, &idx, NULL));
7685: if (nidx) {
7686: cnt = 0;
7687: cc[cnt] = idx[0] / bs;
7688: for (j = 1; j < nidx; j++) {
7689: if (cc[cnt] < idx[j] / bs) cc[++cnt] = idx[j] / bs;
7690: }
7691: }
7692: PetscCall(MatRestoreRow(Amat, row, &nidx, &idx, NULL));
7693: *n = cnt + 1;
7694: PetscFunctionReturn(PETSC_SUCCESS);
7695: }
7697: /*
7698: Produces a set of block column indices of the matrix block row, one for each block represented in the original set of rows
7700: ncollapsed - the number of block indices
7701: collapsed - the block indices (must be large enough to contain the indices)
7702: */
7703: static inline PetscErrorCode MatCollapseRows(Mat Amat, PetscInt start, PetscInt bs, PetscInt *w0, PetscInt *w1, PetscInt *w2, PetscInt *ncollapsed, PetscInt **collapsed)
7704: {
7705: PetscInt i, nprev, *cprev = w0, ncur = 0, *ccur = w1, *merged = w2, *cprevtmp;
7707: PetscFunctionBegin;
7708: PetscCall(MatCollapseRow(Amat, start, bs, &nprev, cprev));
7709: for (i = start + 1; i < start + bs; i++) {
7710: PetscCall(MatCollapseRow(Amat, i, bs, &ncur, ccur));
7711: PetscCall(PetscMergeIntArray(nprev, cprev, ncur, ccur, &nprev, &merged));
7712: cprevtmp = cprev;
7713: cprev = merged;
7714: merged = cprevtmp;
7715: }
7716: *ncollapsed = nprev;
7717: if (collapsed) *collapsed = cprev;
7718: PetscFunctionReturn(PETSC_SUCCESS);
7719: }
7721: /*
7722: MatCreateGraph_Simple_AIJ - create simple scalar matrix (graph) from potentially blocked matrix
7724: Input Parameter:
7725: . Amat - matrix
7726: - symmetrize - make the result symmetric
7727: + scale - scale with diagonal
7729: Output Parameter:
7730: . a_Gmat - output scalar graph >= 0
7732: */
7733: PETSC_INTERN PetscErrorCode MatCreateGraph_Simple_AIJ(Mat Amat, PetscBool symmetrize, PetscBool scale, PetscReal filter, PetscInt index_size, PetscInt index[], Mat *a_Gmat)
7734: {
7735: PetscInt Istart, Iend, Ii, jj, kk, ncols, nloc, NN, MM, bs;
7736: MPI_Comm comm;
7737: Mat Gmat;
7738: PetscBool ismpiaij, isseqaij;
7739: Mat a, b, c;
7740: MatType jtype;
7742: PetscFunctionBegin;
7743: PetscCall(PetscObjectGetComm((PetscObject)Amat, &comm));
7744: PetscCall(MatGetOwnershipRange(Amat, &Istart, &Iend));
7745: PetscCall(MatGetSize(Amat, &MM, &NN));
7746: PetscCall(MatGetBlockSize(Amat, &bs));
7747: nloc = (Iend - Istart) / bs;
7749: PetscCall(PetscObjectBaseTypeCompare((PetscObject)Amat, MATSEQAIJ, &isseqaij));
7750: PetscCall(PetscObjectBaseTypeCompare((PetscObject)Amat, MATMPIAIJ, &ismpiaij));
7751: PetscCheck(isseqaij || ismpiaij, comm, PETSC_ERR_USER, "Require (MPI)AIJ matrix type");
7753: /* TODO GPU: these calls are potentially expensive if matrices are large and we want to use the GPU */
7754: /* A solution consists in providing a new API, MatAIJGetCollapsedAIJ, and each class can provide a fast
7755: implementation */
7756: if (bs > 1) {
7757: PetscCall(MatGetType(Amat, &jtype));
7758: PetscCall(MatCreate(comm, &Gmat));
7759: PetscCall(MatSetType(Gmat, jtype));
7760: PetscCall(MatSetSizes(Gmat, nloc, nloc, PETSC_DETERMINE, PETSC_DETERMINE));
7761: PetscCall(MatSetBlockSizes(Gmat, 1, 1));
7762: if (isseqaij || ((Mat_MPIAIJ *)Amat->data)->garray) {
7763: PetscInt *d_nnz, *o_nnz;
7764: MatScalar *aa, val, *AA;
7765: PetscInt *aj, *ai, *AJ, nc, nmax = 0;
7767: if (isseqaij) {
7768: a = Amat;
7769: b = NULL;
7770: } else {
7771: Mat_MPIAIJ *d = (Mat_MPIAIJ *)Amat->data;
7772: a = d->A;
7773: b = d->B;
7774: }
7775: PetscCall(PetscInfo(Amat, "New bs>1 Graph. nloc=%" PetscInt_FMT "\n", nloc));
7776: PetscCall(PetscMalloc2(nloc, &d_nnz, (isseqaij ? 0 : nloc), &o_nnz));
7777: for (c = a, kk = 0; c && kk < 2; c = b, kk++) {
7778: PetscInt *nnz = (c == a) ? d_nnz : o_nnz;
7779: const PetscInt *cols1, *cols2;
7781: for (PetscInt brow = 0, nc1, nc2, ok = 1; brow < nloc * bs; brow += bs) { // block rows
7782: PetscCall(MatGetRow(c, brow, &nc2, &cols2, NULL));
7783: nnz[brow / bs] = nc2 / bs;
7784: if (nc2 % bs) ok = 0;
7785: if (nnz[brow / bs] > nmax) nmax = nnz[brow / bs];
7786: for (PetscInt ii = 1; ii < bs; ii++) { // check for non-dense blocks
7787: PetscCall(MatGetRow(c, brow + ii, &nc1, &cols1, NULL));
7788: if (nc1 != nc2) ok = 0;
7789: else {
7790: for (PetscInt jj = 0; jj < nc1 && ok == 1; jj++) {
7791: if (cols1[jj] != cols2[jj]) ok = 0;
7792: if (cols1[jj] % bs != jj % bs) ok = 0;
7793: }
7794: }
7795: PetscCall(MatRestoreRow(c, brow + ii, &nc1, &cols1, NULL));
7796: }
7797: PetscCall(MatRestoreRow(c, brow, &nc2, &cols2, NULL));
7798: if (!ok) {
7799: PetscCall(PetscFree2(d_nnz, o_nnz));
7800: PetscCall(PetscInfo(Amat, "Found sparse blocks - revert to slow method\n"));
7801: goto old_bs;
7802: }
7803: }
7804: }
7805: PetscCall(MatSeqAIJSetPreallocation(Gmat, 0, d_nnz));
7806: PetscCall(MatMPIAIJSetPreallocation(Gmat, 0, d_nnz, 0, o_nnz));
7807: PetscCall(PetscFree2(d_nnz, o_nnz));
7808: PetscCall(PetscMalloc2(nmax, &AA, nmax, &AJ));
7809: // diag
7810: for (PetscInt brow = 0, n, grow; brow < nloc * bs; brow += bs) { // block rows
7811: Mat_SeqAIJ *aseq = (Mat_SeqAIJ *)a->data;
7813: ai = aseq->i;
7814: n = ai[brow + 1] - ai[brow];
7815: aj = aseq->j + ai[brow];
7816: for (PetscInt k = 0; k < n; k += bs) { // block columns
7817: AJ[k / bs] = aj[k] / bs + Istart / bs; // diag starts at (Istart,Istart)
7818: val = 0;
7819: if (index_size == 0) {
7820: for (PetscInt ii = 0; ii < bs; ii++) { // rows in block
7821: aa = aseq->a + ai[brow + ii] + k;
7822: for (PetscInt jj = 0; jj < bs; jj++) { // columns in block
7823: val += PetscAbs(PetscRealPart(aa[jj])); // a sort of norm
7824: }
7825: }
7826: } else { // use (index,index) value if provided
7827: for (PetscInt iii = 0; iii < index_size; iii++) { // rows in block
7828: PetscInt ii = index[iii];
7829: aa = aseq->a + ai[brow + ii] + k;
7830: for (PetscInt jjj = 0; jjj < index_size; jjj++) { // columns in block
7831: PetscInt jj = index[jjj];
7832: val += PetscAbs(PetscRealPart(aa[jj]));
7833: }
7834: }
7835: }
7836: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs (%" PetscInt_FMT ") >= nmax (%" PetscInt_FMT ")", k / bs, nmax);
7837: AA[k / bs] = val;
7838: }
7839: grow = Istart / bs + brow / bs;
7840: PetscCall(MatSetValues(Gmat, 1, &grow, n / bs, AJ, AA, ADD_VALUES));
7841: }
7842: // off-diag
7843: if (ismpiaij) {
7844: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)Amat->data;
7845: const PetscScalar *vals;
7846: const PetscInt *cols, *garray = aij->garray;
7848: PetscCheck(garray, PETSC_COMM_SELF, PETSC_ERR_USER, "No garray ?");
7849: for (PetscInt brow = 0, grow; brow < nloc * bs; brow += bs) { // block rows
7850: PetscCall(MatGetRow(b, brow, &ncols, &cols, NULL));
7851: for (PetscInt k = 0, cidx = 0; k < ncols; k += bs, cidx++) {
7852: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs >= nmax");
7853: AA[k / bs] = 0;
7854: AJ[cidx] = garray[cols[k]] / bs;
7855: }
7856: nc = ncols / bs;
7857: PetscCall(MatRestoreRow(b, brow, &ncols, &cols, NULL));
7858: if (index_size == 0) {
7859: for (PetscInt ii = 0; ii < bs; ii++) { // rows in block
7860: PetscCall(MatGetRow(b, brow + ii, &ncols, &cols, &vals));
7861: for (PetscInt k = 0; k < ncols; k += bs) {
7862: for (PetscInt jj = 0; jj < bs; jj++) { // cols in block
7863: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs (%" PetscInt_FMT ") >= nmax (%" PetscInt_FMT ")", k / bs, nmax);
7864: AA[k / bs] += PetscAbs(PetscRealPart(vals[k + jj]));
7865: }
7866: }
7867: PetscCall(MatRestoreRow(b, brow + ii, &ncols, &cols, &vals));
7868: }
7869: } else { // use (index,index) value if provided
7870: for (PetscInt iii = 0; iii < index_size; iii++) { // rows in block
7871: PetscInt ii = index[iii];
7872: PetscCall(MatGetRow(b, brow + ii, &ncols, &cols, &vals));
7873: for (PetscInt k = 0; k < ncols; k += bs) {
7874: for (PetscInt jjj = 0; jjj < index_size; jjj++) { // cols in block
7875: PetscInt jj = index[jjj];
7876: AA[k / bs] += PetscAbs(PetscRealPart(vals[k + jj]));
7877: }
7878: }
7879: PetscCall(MatRestoreRow(b, brow + ii, &ncols, &cols, &vals));
7880: }
7881: }
7882: grow = Istart / bs + brow / bs;
7883: PetscCall(MatSetValues(Gmat, 1, &grow, nc, AJ, AA, ADD_VALUES));
7884: }
7885: }
7886: PetscCall(MatAssemblyBegin(Gmat, MAT_FINAL_ASSEMBLY));
7887: PetscCall(MatAssemblyEnd(Gmat, MAT_FINAL_ASSEMBLY));
7888: PetscCall(PetscFree2(AA, AJ));
7889: } else {
7890: const PetscScalar *vals;
7891: const PetscInt *idx;
7892: PetscInt *d_nnz, *o_nnz, *w0, *w1, *w2;
7893: old_bs:
7894: /*
7895: Determine the preallocation needed for the scalar matrix derived from the vector matrix.
7896: */
7897: PetscCall(PetscInfo(Amat, "OLD bs>1 CreateGraph\n"));
7898: PetscCall(PetscMalloc2(nloc, &d_nnz, (isseqaij ? 0 : nloc), &o_nnz));
7899: if (isseqaij) {
7900: PetscInt max_d_nnz;
7902: /*
7903: Determine exact preallocation count for (sequential) scalar matrix
7904: */
7905: PetscCall(MatSeqAIJGetMaxRowNonzeros(Amat, &max_d_nnz));
7906: max_d_nnz = PetscMin(nloc, bs * max_d_nnz);
7907: PetscCall(PetscMalloc3(max_d_nnz, &w0, max_d_nnz, &w1, max_d_nnz, &w2));
7908: for (Ii = 0, jj = 0; Ii < Iend; Ii += bs, jj++) PetscCall(MatCollapseRows(Amat, Ii, bs, w0, w1, w2, &d_nnz[jj], NULL));
7909: PetscCall(PetscFree3(w0, w1, w2));
7910: } else if (ismpiaij) {
7911: Mat Daij, Oaij;
7912: const PetscInt *garray;
7913: PetscInt max_d_nnz;
7915: PetscCall(MatMPIAIJGetSeqAIJ(Amat, &Daij, &Oaij, &garray));
7916: /*
7917: Determine exact preallocation count for diagonal block portion of scalar matrix
7918: */
7919: PetscCall(MatSeqAIJGetMaxRowNonzeros(Daij, &max_d_nnz));
7920: max_d_nnz = PetscMin(nloc, bs * max_d_nnz);
7921: PetscCall(PetscMalloc3(max_d_nnz, &w0, max_d_nnz, &w1, max_d_nnz, &w2));
7922: for (Ii = 0, jj = 0; Ii < Iend - Istart; Ii += bs, jj++) PetscCall(MatCollapseRows(Daij, Ii, bs, w0, w1, w2, &d_nnz[jj], NULL));
7923: PetscCall(PetscFree3(w0, w1, w2));
7924: /*
7925: Over estimate (usually grossly over), preallocation count for off-diagonal portion of scalar matrix
7926: */
7927: for (Ii = 0, jj = 0; Ii < Iend - Istart; Ii += bs, jj++) {
7928: o_nnz[jj] = 0;
7929: for (kk = 0; kk < bs; kk++) { /* rows that get collapsed to a single row */
7930: PetscCall(MatGetRow(Oaij, Ii + kk, &ncols, NULL, NULL));
7931: o_nnz[jj] += ncols;
7932: PetscCall(MatRestoreRow(Oaij, Ii + kk, &ncols, NULL, NULL));
7933: }
7934: if (o_nnz[jj] > (NN / bs - nloc)) o_nnz[jj] = NN / bs - nloc;
7935: }
7936: } else SETERRQ(comm, PETSC_ERR_USER, "Require AIJ matrix type");
7937: /* get scalar copy (norms) of matrix */
7938: PetscCall(MatSeqAIJSetPreallocation(Gmat, 0, d_nnz));
7939: PetscCall(MatMPIAIJSetPreallocation(Gmat, 0, d_nnz, 0, o_nnz));
7940: PetscCall(PetscFree2(d_nnz, o_nnz));
7941: for (Ii = Istart; Ii < Iend; Ii++) {
7942: PetscInt dest_row = Ii / bs;
7944: PetscCall(MatGetRow(Amat, Ii, &ncols, &idx, &vals));
7945: for (jj = 0; jj < ncols; jj++) {
7946: PetscInt dest_col = idx[jj] / bs;
7947: PetscScalar sv = PetscAbs(PetscRealPart(vals[jj]));
7949: PetscCall(MatSetValues(Gmat, 1, &dest_row, 1, &dest_col, &sv, ADD_VALUES));
7950: }
7951: PetscCall(MatRestoreRow(Amat, Ii, &ncols, &idx, &vals));
7952: }
7953: PetscCall(MatAssemblyBegin(Gmat, MAT_FINAL_ASSEMBLY));
7954: PetscCall(MatAssemblyEnd(Gmat, MAT_FINAL_ASSEMBLY));
7955: }
7956: } else {
7957: if (symmetrize || filter >= 0 || scale) PetscCall(MatDuplicate(Amat, MAT_COPY_VALUES, &Gmat));
7958: else {
7959: Gmat = Amat;
7960: PetscCall(PetscObjectReference((PetscObject)Gmat));
7961: }
7962: if (isseqaij) {
7963: a = Gmat;
7964: b = NULL;
7965: } else {
7966: Mat_MPIAIJ *d = (Mat_MPIAIJ *)Gmat->data;
7967: a = d->A;
7968: b = d->B;
7969: }
7970: if (filter >= 0 || scale) {
7971: /* take absolute value of each entry */
7972: for (c = a, kk = 0; c && kk < 2; c = b, kk++) {
7973: MatInfo info;
7974: PetscScalar *avals;
7976: PetscCall(MatGetInfo(c, MAT_LOCAL, &info));
7977: PetscCall(MatSeqAIJGetArray(c, &avals));
7978: for (int jj = 0; jj < info.nz_used; jj++) avals[jj] = PetscAbsScalar(avals[jj]);
7979: PetscCall(MatSeqAIJRestoreArray(c, &avals));
7980: }
7981: }
7982: }
7983: if (symmetrize) {
7984: PetscBool isset, issym;
7986: PetscCall(MatIsSymmetricKnown(Amat, &isset, &issym));
7987: if (!isset || !issym) {
7988: Mat matTrans;
7990: PetscCall(MatTranspose(Gmat, MAT_INITIAL_MATRIX, &matTrans));
7991: PetscCall(MatAXPY(Gmat, 1.0, matTrans, Gmat->structurally_symmetric == PETSC_BOOL3_TRUE ? SAME_NONZERO_PATTERN : DIFFERENT_NONZERO_PATTERN));
7992: PetscCall(MatDestroy(&matTrans));
7993: }
7994: PetscCall(MatSetOption(Gmat, MAT_SYMMETRIC, PETSC_TRUE));
7995: } else if (Amat != Gmat) PetscCall(MatPropagateSymmetryOptions(Amat, Gmat));
7996: if (scale) {
7997: /* scale c for all diagonal values = 1 or -1 */
7998: Vec diag;
8000: PetscCall(MatCreateVecs(Gmat, &diag, NULL));
8001: PetscCall(MatGetDiagonal(Gmat, diag));
8002: PetscCall(VecReciprocal(diag));
8003: PetscCall(VecSqrtAbs(diag));
8004: PetscCall(MatDiagonalScale(Gmat, diag, diag));
8005: PetscCall(VecDestroy(&diag));
8006: }
8007: PetscCall(MatViewFromOptions(Gmat, NULL, "-mat_graph_view"));
8008: if (filter >= 0) {
8009: PetscCall(MatFilter(Gmat, filter, PETSC_TRUE, PETSC_TRUE));
8010: PetscCall(MatViewFromOptions(Gmat, NULL, "-mat_filter_graph_view"));
8011: }
8012: *a_Gmat = Gmat;
8013: PetscFunctionReturn(PETSC_SUCCESS);
8014: }
8016: PETSC_INTERN PetscErrorCode MatGetCurrentMemType_MPIAIJ(Mat A, PetscMemType *memtype)
8017: {
8018: Mat_MPIAIJ *mpiaij = (Mat_MPIAIJ *)A->data;
8019: PetscMemType mD = PETSC_MEMTYPE_HOST, mO = PETSC_MEMTYPE_HOST;
8021: PetscFunctionBegin;
8022: if (mpiaij->A) PetscCall(MatGetCurrentMemType(mpiaij->A, &mD));
8023: if (mpiaij->B) PetscCall(MatGetCurrentMemType(mpiaij->B, &mO));
8024: *memtype = (mD == mO) ? mD : PETSC_MEMTYPE_HOST;
8025: PetscFunctionReturn(PETSC_SUCCESS);
8026: }
8028: /*
8029: Special version for direct calls from Fortran
8030: */
8032: /* Change these macros so can be used in void function */
8033: /* Identical to PetscCallVoid, except it assigns to *_ierr */
8034: #undef PetscCall
8035: #define PetscCall(...) \
8036: do { \
8037: PetscErrorCode ierr_msv_mpiaij = __VA_ARGS__; \
8038: if (PetscUnlikely(ierr_msv_mpiaij)) { \
8039: *_ierr = PetscError(PETSC_COMM_SELF, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr_msv_mpiaij, PETSC_ERROR_REPEAT, " "); \
8040: return; \
8041: } \
8042: } while (0)
8044: #undef SETERRQ
8045: #define SETERRQ(comm, ierr, ...) \
8046: do { \
8047: *_ierr = PetscError(comm, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr, PETSC_ERROR_INITIAL, __VA_ARGS__); \
8048: return; \
8049: } while (0)
8051: #if defined(PETSC_HAVE_FORTRAN_CAPS)
8052: #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ
8053: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
8054: #define matsetvaluesmpiaij_ matsetvaluesmpiaij
8055: #else
8056: #endif
8057: PETSC_EXTERN void matsetvaluesmpiaij_(Mat *mmat, PetscInt *mm, const PetscInt im[], PetscInt *mn, const PetscInt in[], const PetscScalar v[], InsertMode *maddv, PetscErrorCode *_ierr)
8058: {
8059: Mat mat = *mmat;
8060: PetscInt m = *mm, n = *mn;
8061: InsertMode addv = *maddv;
8062: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
8063: PetscScalar value;
8065: MatCheckPreallocated(mat, 1);
8066: if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv;
8067: else PetscCheck(mat->insertmode == addv, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Cannot mix add values and insert values");
8068: {
8069: PetscInt i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend;
8070: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend, row, col;
8071: PetscBool roworiented = aij->roworiented;
8073: /* Some Variables required in the macro */
8074: Mat A = aij->A;
8075: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
8076: PetscInt *aimax = a->imax, *ai = a->i, *ailen = a->ilen, *aj = a->j;
8077: MatScalar *aa;
8078: PetscBool ignorezeroentries = ((a->ignorezeroentries && (addv == ADD_VALUES)) ? PETSC_TRUE : PETSC_FALSE);
8079: Mat B = aij->B;
8080: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
8081: PetscInt *bimax = b->imax, *bi = b->i, *bilen = b->ilen, *bj = b->j, bm = aij->B->rmap->n, am = aij->A->rmap->n;
8082: MatScalar *ba;
8083: /* This variable below is only for the PETSC_HAVE_VIENNACL or PETSC_HAVE_CUDA cases, but we define it in all cases because we
8084: * cannot use "#if defined" inside a macro. */
8085: PETSC_UNUSED PetscBool inserted = PETSC_FALSE;
8087: PetscInt *rp1, *rp2, ii, nrow1, nrow2, _i, rmax1, rmax2, N, low1, high1, low2, high2, t, lastcol1, lastcol2;
8088: PetscInt nonew = a->nonew;
8089: MatScalar *ap1, *ap2;
8091: PetscFunctionBegin;
8092: PetscCall(MatSeqAIJGetArray(A, &aa));
8093: PetscCall(MatSeqAIJGetArray(B, &ba));
8094: for (i = 0; i < m; i++) {
8095: if (im[i] < 0) continue;
8096: 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);
8097: if (im[i] >= rstart && im[i] < rend) {
8098: row = im[i] - rstart;
8099: lastcol1 = -1;
8100: rp1 = aj + ai[row];
8101: ap1 = aa + ai[row];
8102: rmax1 = aimax[row];
8103: nrow1 = ailen[row];
8104: low1 = 0;
8105: high1 = nrow1;
8106: lastcol2 = -1;
8107: rp2 = bj + bi[row];
8108: ap2 = ba + bi[row];
8109: rmax2 = bimax[row];
8110: nrow2 = bilen[row];
8111: low2 = 0;
8112: high2 = nrow2;
8114: for (j = 0; j < n; j++) {
8115: if (roworiented) value = v[i * n + j];
8116: else value = v[i + j * m];
8117: if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && im[i] != in[j]) continue;
8118: if (in[j] >= cstart && in[j] < cend) {
8119: col = in[j] - cstart;
8120: MatSetValues_SeqAIJ_A_Private(row, col, value, addv, im[i], in[j]);
8121: } else if (in[j] < 0) continue;
8122: else if (PetscUnlikelyDebug(in[j] >= mat->cmap->N)) {
8123: SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large: col %" PetscInt_FMT " max %" PetscInt_FMT, in[j], mat->cmap->N - 1);
8124: } else {
8125: if (mat->was_assembled) {
8126: if (!aij->colmap) PetscCall(MatCreateColmap_MPIAIJ_Private(mat));
8127: #if defined(PETSC_USE_CTABLE)
8128: PetscCall(PetscHMapIGetWithDefault(aij->colmap, in[j] + 1, 0, &col));
8129: col--;
8130: #else
8131: col = aij->colmap[in[j]] - 1;
8132: #endif
8133: if (col < 0 && !((Mat_SeqAIJ *)aij->A->data)->nonew) {
8134: PetscCall(MatDisAssemble_MPIAIJ(mat, PETSC_FALSE));
8135: col = in[j];
8136: /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
8137: B = aij->B;
8138: b = (Mat_SeqAIJ *)B->data;
8139: bimax = b->imax;
8140: bi = b->i;
8141: bilen = b->ilen;
8142: bj = b->j;
8143: rp2 = bj + bi[row];
8144: ap2 = ba + bi[row];
8145: rmax2 = bimax[row];
8146: nrow2 = bilen[row];
8147: low2 = 0;
8148: high2 = nrow2;
8149: bm = aij->B->rmap->n;
8150: ba = b->a;
8151: inserted = PETSC_FALSE;
8152: }
8153: } else col = in[j];
8154: MatSetValues_SeqAIJ_B_Private(row, col, value, addv, im[i], in[j]);
8155: }
8156: }
8157: } else if (!aij->donotstash) {
8158: if (roworiented) {
8159: PetscCall(MatStashValuesRow_Private(&mat->stash, im[i], n, in, v + i * n, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
8160: } else {
8161: PetscCall(MatStashValuesCol_Private(&mat->stash, im[i], n, in, v + i, m, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
8162: }
8163: }
8164: }
8165: PetscCall(MatSeqAIJRestoreArray(A, &aa));
8166: PetscCall(MatSeqAIJRestoreArray(B, &ba));
8167: }
8168: PetscFunctionReturnVoid();
8169: }
8171: /* Undefining these here since they were redefined from their original definition above! No
8172: * other PETSc functions should be defined past this point, as it is impossible to recover the
8173: * original definitions */
8174: #undef PetscCall
8175: #undef SETERRQ