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)
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 dow 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: if (1 == ((Mat_SeqAIJ *)aij->B->data)->nonew) {
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: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Inserting a new nonzero at global row/column (%" PetscInt_FMT ", %" PetscInt_FMT ") into matrix", im[i], in[j]);
615: }
616: } else col = in[j];
617: nonew = b->nonew;
618: MatSetValues_SeqAIJ_B_Private(row, col, value, addv, im[i], in[j]);
619: }
620: }
621: } else {
622: 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]);
623: if (!aij->donotstash) {
624: mat->assembled = PETSC_FALSE;
625: if (roworiented) {
626: PetscCall(MatStashValuesRow_Private(&mat->stash, im[i], n, in, PetscSafePointerPlusOffset(v, i * n), (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
627: } else {
628: PetscCall(MatStashValuesCol_Private(&mat->stash, im[i], n, in, PetscSafePointerPlusOffset(v, i), m, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
629: }
630: }
631: }
632: }
633: PetscCall(MatSeqAIJRestoreArray(A, &aa)); /* aa, bb might have been free'd due to reallocation above. But we don't access them here */
634: PetscCall(MatSeqAIJRestoreArray(B, &ba));
635: PetscFunctionReturn(PETSC_SUCCESS);
636: }
638: /*
639: This function sets the j and ilen arrays (of the diagonal and off-diagonal part) of an MPIAIJ-matrix.
640: The values in mat_i have to be sorted and the values in mat_j have to be sorted for each row (CSR-like).
641: No off-processor parts off the matrix are allowed here and mat->was_assembled has to be PETSC_FALSE.
642: */
643: PetscErrorCode MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(Mat mat, const PetscInt mat_j[], const PetscInt mat_i[])
644: {
645: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
646: Mat A = aij->A; /* diagonal part of the matrix */
647: Mat B = aij->B; /* off-diagonal part of the matrix */
648: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
649: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
650: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend, col;
651: PetscInt *ailen = a->ilen, *aj = a->j;
652: PetscInt *bilen = b->ilen, *bj = b->j;
653: PetscInt am = aij->A->rmap->n, j;
654: PetscInt diag_so_far = 0, dnz;
655: PetscInt offd_so_far = 0, onz;
657: PetscFunctionBegin;
658: /* Iterate over all rows of the matrix */
659: for (j = 0; j < am; j++) {
660: dnz = onz = 0;
661: /* Iterate over all non-zero columns of the current row */
662: for (col = mat_i[j]; col < mat_i[j + 1]; col++) {
663: /* If column is in the diagonal */
664: if (mat_j[col] >= cstart && mat_j[col] < cend) {
665: aj[diag_so_far++] = mat_j[col] - cstart;
666: dnz++;
667: } else { /* off-diagonal entries */
668: bj[offd_so_far++] = mat_j[col];
669: onz++;
670: }
671: }
672: ailen[j] = dnz;
673: bilen[j] = onz;
674: }
675: PetscFunctionReturn(PETSC_SUCCESS);
676: }
678: /*
679: This function sets the local j, a and ilen arrays (of the diagonal and off-diagonal part) of an MPIAIJ-matrix.
680: The values in mat_i have to be sorted and the values in mat_j have to be sorted for each row (CSR-like).
681: No off-processor parts off the matrix are allowed here, they are set at a later point by MatSetValues_MPIAIJ.
682: Also, mat->was_assembled has to be false, otherwise the statement aj[rowstart_diag+dnz_row] = mat_j[col] - cstart;
683: would not be true and the more complex MatSetValues_MPIAIJ has to be used.
684: */
685: PetscErrorCode MatSetValues_MPIAIJ_CopyFromCSRFormat(Mat mat, const PetscInt mat_j[], const PetscInt mat_i[], const PetscScalar mat_a[])
686: {
687: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
688: Mat A = aij->A; /* diagonal part of the matrix */
689: Mat B = aij->B; /* off-diagonal part of the matrix */
690: Mat_SeqAIJ *aijd = (Mat_SeqAIJ *)aij->A->data, *aijo = (Mat_SeqAIJ *)aij->B->data;
691: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
692: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
693: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend;
694: PetscInt *ailen = a->ilen, *aj = a->j;
695: PetscInt *bilen = b->ilen, *bj = b->j;
696: PetscInt am = aij->A->rmap->n, j;
697: 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. */
698: PetscInt col, dnz_row, onz_row, rowstart_diag, rowstart_offd;
699: PetscScalar *aa = a->a, *ba = b->a;
701: PetscFunctionBegin;
702: /* Iterate over all rows of the matrix */
703: for (j = 0; j < am; j++) {
704: dnz_row = onz_row = 0;
705: rowstart_offd = full_offd_i[j];
706: rowstart_diag = full_diag_i[j];
707: /* Iterate over all non-zero columns of the current row */
708: for (col = mat_i[j]; col < mat_i[j + 1]; col++) {
709: /* If column is in the diagonal */
710: if (mat_j[col] >= cstart && mat_j[col] < cend) {
711: aj[rowstart_diag + dnz_row] = mat_j[col] - cstart;
712: aa[rowstart_diag + dnz_row] = mat_a[col];
713: dnz_row++;
714: } else { /* off-diagonal entries */
715: bj[rowstart_offd + onz_row] = mat_j[col];
716: ba[rowstart_offd + onz_row] = mat_a[col];
717: onz_row++;
718: }
719: }
720: ailen[j] = dnz_row;
721: bilen[j] = onz_row;
722: }
723: PetscFunctionReturn(PETSC_SUCCESS);
724: }
726: static PetscErrorCode MatGetValues_MPIAIJ(Mat mat, PetscInt m, const PetscInt idxm[], PetscInt n, const PetscInt idxn[], PetscScalar v[])
727: {
728: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
729: PetscInt i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend;
730: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend, row, col;
732: PetscFunctionBegin;
733: for (i = 0; i < m; i++) {
734: if (idxm[i] < 0) continue; /* negative row */
735: 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);
736: 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);
737: row = idxm[i] - rstart;
738: for (j = 0; j < n; j++) {
739: if (idxn[j] < 0) continue; /* negative column */
740: 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);
741: if (idxn[j] >= cstart && idxn[j] < cend) {
742: col = idxn[j] - cstart;
743: PetscCall(MatGetValues(aij->A, 1, &row, 1, &col, v + i * n + j));
744: } else {
745: if (!aij->colmap) PetscCall(MatCreateColmap_MPIAIJ_Private(mat));
746: #if defined(PETSC_USE_CTABLE)
747: PetscCall(PetscHMapIGetWithDefault(aij->colmap, idxn[j] + 1, 0, &col));
748: col--;
749: #else
750: col = aij->colmap[idxn[j]] - 1;
751: #endif
752: if ((col < 0) || (aij->garray[col] != idxn[j])) *(v + i * n + j) = 0.0;
753: else PetscCall(MatGetValues(aij->B, 1, &row, 1, &col, v + i * n + j));
754: }
755: }
756: }
757: PetscFunctionReturn(PETSC_SUCCESS);
758: }
760: static PetscErrorCode MatAssemblyBegin_MPIAIJ(Mat mat, MatAssemblyType mode)
761: {
762: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
763: PetscInt nstash, reallocs;
765: PetscFunctionBegin;
766: if (aij->donotstash || mat->nooffprocentries) PetscFunctionReturn(PETSC_SUCCESS);
768: PetscCall(MatStashScatterBegin_Private(mat, &mat->stash, mat->rmap->range));
769: PetscCall(MatStashGetInfo_Private(&mat->stash, &nstash, &reallocs));
770: PetscCall(PetscInfo(aij->A, "Stash has %" PetscInt_FMT " entries, uses %" PetscInt_FMT " mallocs.\n", nstash, reallocs));
771: PetscFunctionReturn(PETSC_SUCCESS);
772: }
774: PetscErrorCode MatAssemblyEnd_MPIAIJ(Mat mat, MatAssemblyType mode)
775: {
776: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
777: PetscMPIInt n;
778: PetscInt i, j, rstart, ncols, flg;
779: PetscInt *row, *col;
780: PetscBool other_disassembled;
781: PetscScalar *val;
783: /* do not use 'b = (Mat_SeqAIJ*)aij->B->data' as B can be reset in disassembly */
785: PetscFunctionBegin;
786: if (!aij->donotstash && !mat->nooffprocentries) {
787: while (1) {
788: PetscCall(MatStashScatterGetMesg_Private(&mat->stash, &n, &row, &col, &val, &flg));
789: if (!flg) break;
791: for (i = 0; i < n;) {
792: /* Now identify the consecutive vals belonging to the same row */
793: for (j = i, rstart = row[j]; j < n; j++) {
794: if (row[j] != rstart) break;
795: }
796: if (j < n) ncols = j - i;
797: else ncols = n - i;
798: /* Now assemble all these values with a single function call */
799: PetscCall(MatSetValues_MPIAIJ(mat, 1, row + i, ncols, col + i, val + i, mat->insertmode));
800: i = j;
801: }
802: }
803: PetscCall(MatStashScatterEnd_Private(&mat->stash));
804: }
805: #if defined(PETSC_HAVE_DEVICE)
806: if (mat->offloadmask == PETSC_OFFLOAD_CPU) aij->A->offloadmask = PETSC_OFFLOAD_CPU;
807: /* We call MatBindToCPU() on aij->A and aij->B here, because if MatBindToCPU_MPIAIJ() is called before assembly, it cannot bind these. */
808: if (mat->boundtocpu) {
809: PetscCall(MatBindToCPU(aij->A, PETSC_TRUE));
810: PetscCall(MatBindToCPU(aij->B, PETSC_TRUE));
811: }
812: #endif
813: PetscCall(MatAssemblyBegin(aij->A, mode));
814: PetscCall(MatAssemblyEnd(aij->A, mode));
816: /* determine if any processor has disassembled, if so we must
817: also disassemble ourself, in order that we may reassemble. */
818: /*
819: if nonzero structure of submatrix B cannot change then we know that
820: no processor disassembled thus we can skip this stuff
821: */
822: if (!((Mat_SeqAIJ *)aij->B->data)->nonew) {
823: PetscCallMPI(MPIU_Allreduce(&mat->was_assembled, &other_disassembled, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
824: if (mat->was_assembled && !other_disassembled) { /* mat on this rank has reduced off-diag B with local col ids, but globally it does not */
825: PetscCall(MatDisAssemble_MPIAIJ(mat, PETSC_FALSE));
826: }
827: }
828: if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) PetscCall(MatSetUpMultiply_MPIAIJ(mat));
829: PetscCall(MatSetOption(aij->B, MAT_USE_INODES, PETSC_FALSE));
830: #if defined(PETSC_HAVE_DEVICE)
831: if (mat->offloadmask == PETSC_OFFLOAD_CPU && aij->B->offloadmask != PETSC_OFFLOAD_UNALLOCATED) aij->B->offloadmask = PETSC_OFFLOAD_CPU;
832: #endif
833: PetscCall(MatAssemblyBegin(aij->B, mode));
834: PetscCall(MatAssemblyEnd(aij->B, mode));
836: PetscCall(PetscFree2(aij->rowvalues, aij->rowindices));
838: aij->rowvalues = NULL;
840: PetscCall(VecDestroy(&aij->diag));
842: /* if no new nonzero locations are allowed in matrix then only set the matrix state the first time through */
843: if ((!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) || !((Mat_SeqAIJ *)aij->A->data)->nonew) {
844: PetscObjectState state = aij->A->nonzerostate + aij->B->nonzerostate;
845: PetscCallMPI(MPIU_Allreduce(&state, &mat->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)mat)));
846: }
847: #if defined(PETSC_HAVE_DEVICE)
848: mat->offloadmask = PETSC_OFFLOAD_BOTH;
849: #endif
850: PetscFunctionReturn(PETSC_SUCCESS);
851: }
853: static PetscErrorCode MatZeroEntries_MPIAIJ(Mat A)
854: {
855: Mat_MPIAIJ *l = (Mat_MPIAIJ *)A->data;
857: PetscFunctionBegin;
858: PetscCall(MatZeroEntries(l->A));
859: PetscCall(MatZeroEntries(l->B));
860: PetscFunctionReturn(PETSC_SUCCESS);
861: }
863: static PetscErrorCode MatZeroRows_MPIAIJ(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diag, Vec x, Vec b)
864: {
865: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
866: PetscInt *lrows;
867: PetscInt r, len;
868: PetscBool cong;
870: PetscFunctionBegin;
871: /* get locally owned rows */
872: PetscCall(MatZeroRowsMapLocal_Private(A, N, rows, &len, &lrows));
873: PetscCall(MatHasCongruentLayouts(A, &cong));
874: /* fix right-hand side if needed */
875: if (x && b) {
876: const PetscScalar *xx;
877: PetscScalar *bb;
879: PetscCheck(cong, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Need matching row/col layout");
880: PetscCall(VecGetArrayRead(x, &xx));
881: PetscCall(VecGetArray(b, &bb));
882: for (r = 0; r < len; ++r) bb[lrows[r]] = diag * xx[lrows[r]];
883: PetscCall(VecRestoreArrayRead(x, &xx));
884: PetscCall(VecRestoreArray(b, &bb));
885: }
887: if (diag != 0.0 && cong) {
888: PetscCall(MatZeroRows(mat->A, len, lrows, diag, NULL, NULL));
889: PetscCall(MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL));
890: } else if (diag != 0.0) { /* non-square or non congruent layouts -> if keepnonzeropattern is false, we allow for new insertion */
891: Mat_SeqAIJ *aijA = (Mat_SeqAIJ *)mat->A->data;
892: Mat_SeqAIJ *aijB = (Mat_SeqAIJ *)mat->B->data;
893: PetscInt nnwA, nnwB;
894: PetscBool nnzA, nnzB;
896: nnwA = aijA->nonew;
897: nnwB = aijB->nonew;
898: nnzA = aijA->keepnonzeropattern;
899: nnzB = aijB->keepnonzeropattern;
900: if (!nnzA) {
901: PetscCall(PetscInfo(mat->A, "Requested to not keep the pattern and add a nonzero diagonal; may encounter reallocations on diagonal block.\n"));
902: aijA->nonew = 0;
903: }
904: if (!nnzB) {
905: PetscCall(PetscInfo(mat->B, "Requested to not keep the pattern and add a nonzero diagonal; may encounter reallocations on off-diagonal block.\n"));
906: aijB->nonew = 0;
907: }
908: /* Must zero here before the next loop */
909: PetscCall(MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL));
910: PetscCall(MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL));
911: for (r = 0; r < len; ++r) {
912: const PetscInt row = lrows[r] + A->rmap->rstart;
913: if (row >= A->cmap->N) continue;
914: PetscCall(MatSetValues(A, 1, &row, 1, &row, &diag, INSERT_VALUES));
915: }
916: aijA->nonew = nnwA;
917: aijB->nonew = nnwB;
918: } else {
919: PetscCall(MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL));
920: PetscCall(MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL));
921: }
922: PetscCall(PetscFree(lrows));
923: PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY));
924: PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY));
926: /* only change matrix nonzero state if pattern was allowed to be changed */
927: if (!((Mat_SeqAIJ *)mat->A->data)->keepnonzeropattern || !((Mat_SeqAIJ *)mat->A->data)->nonew) {
928: PetscObjectState state = mat->A->nonzerostate + mat->B->nonzerostate;
929: PetscCallMPI(MPIU_Allreduce(&state, &A->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)A)));
930: }
931: PetscFunctionReturn(PETSC_SUCCESS);
932: }
934: static PetscErrorCode MatZeroRowsColumns_MPIAIJ(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diag, Vec x, Vec b)
935: {
936: Mat_MPIAIJ *l = (Mat_MPIAIJ *)A->data;
937: PetscInt n = A->rmap->n;
938: PetscInt i, j, r, m, len = 0;
939: PetscInt *lrows, *owners = A->rmap->range;
940: PetscMPIInt p = 0;
941: PetscSFNode *rrows;
942: PetscSF sf;
943: const PetscScalar *xx;
944: PetscScalar *bb, *mask, *aij_a;
945: Vec xmask, lmask;
946: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)l->B->data;
947: const PetscInt *aj, *ii, *ridx;
948: PetscScalar *aa;
950: PetscFunctionBegin;
951: /* Create SF where leaves are input rows and roots are owned rows */
952: PetscCall(PetscMalloc1(n, &lrows));
953: for (r = 0; r < n; ++r) lrows[r] = -1;
954: PetscCall(PetscMalloc1(N, &rrows));
955: for (r = 0; r < N; ++r) {
956: const PetscInt idx = rows[r];
957: 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);
958: if (idx < owners[p] || owners[p + 1] <= idx) { /* short-circuit the search if the last p owns this row too */
959: PetscCall(PetscLayoutFindOwner(A->rmap, idx, &p));
960: }
961: rrows[r].rank = p;
962: rrows[r].index = rows[r] - owners[p];
963: }
964: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
965: PetscCall(PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER));
966: /* Collect flags for rows to be zeroed */
967: PetscCall(PetscSFReduceBegin(sf, MPIU_INT, (PetscInt *)rows, lrows, MPI_LOR));
968: PetscCall(PetscSFReduceEnd(sf, MPIU_INT, (PetscInt *)rows, lrows, MPI_LOR));
969: PetscCall(PetscSFDestroy(&sf));
970: /* Compress and put in row numbers */
971: for (r = 0; r < n; ++r)
972: if (lrows[r] >= 0) lrows[len++] = r;
973: /* zero diagonal part of matrix */
974: PetscCall(MatZeroRowsColumns(l->A, len, lrows, diag, x, b));
975: /* handle off-diagonal part of matrix */
976: PetscCall(MatCreateVecs(A, &xmask, NULL));
977: PetscCall(VecDuplicate(l->lvec, &lmask));
978: PetscCall(VecGetArray(xmask, &bb));
979: for (i = 0; i < len; i++) bb[lrows[i]] = 1;
980: PetscCall(VecRestoreArray(xmask, &bb));
981: PetscCall(VecScatterBegin(l->Mvctx, xmask, lmask, ADD_VALUES, SCATTER_FORWARD));
982: PetscCall(VecScatterEnd(l->Mvctx, xmask, lmask, ADD_VALUES, SCATTER_FORWARD));
983: PetscCall(VecDestroy(&xmask));
984: if (x && b) { /* this code is buggy when the row and column layout don't match */
985: PetscBool cong;
987: PetscCall(MatHasCongruentLayouts(A, &cong));
988: PetscCheck(cong, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Need matching row/col layout");
989: PetscCall(VecScatterBegin(l->Mvctx, x, l->lvec, INSERT_VALUES, SCATTER_FORWARD));
990: PetscCall(VecScatterEnd(l->Mvctx, x, l->lvec, INSERT_VALUES, SCATTER_FORWARD));
991: PetscCall(VecGetArrayRead(l->lvec, &xx));
992: PetscCall(VecGetArray(b, &bb));
993: }
994: PetscCall(VecGetArray(lmask, &mask));
995: /* remove zeroed rows of off-diagonal matrix */
996: PetscCall(MatSeqAIJGetArray(l->B, &aij_a));
997: ii = aij->i;
998: for (i = 0; i < len; i++) PetscCall(PetscArrayzero(PetscSafePointerPlusOffset(aij_a, ii[lrows[i]]), ii[lrows[i] + 1] - ii[lrows[i]]));
999: /* loop over all elements of off process part of matrix zeroing removed columns*/
1000: if (aij->compressedrow.use) {
1001: m = aij->compressedrow.nrows;
1002: ii = aij->compressedrow.i;
1003: ridx = aij->compressedrow.rindex;
1004: for (i = 0; i < m; i++) {
1005: n = ii[i + 1] - ii[i];
1006: aj = aij->j + ii[i];
1007: aa = aij_a + ii[i];
1009: for (j = 0; j < n; j++) {
1010: if (PetscAbsScalar(mask[*aj])) {
1011: if (b) bb[*ridx] -= *aa * xx[*aj];
1012: *aa = 0.0;
1013: }
1014: aa++;
1015: aj++;
1016: }
1017: ridx++;
1018: }
1019: } else { /* do not use compressed row format */
1020: m = l->B->rmap->n;
1021: for (i = 0; i < m; i++) {
1022: n = ii[i + 1] - ii[i];
1023: aj = aij->j + ii[i];
1024: aa = aij_a + ii[i];
1025: for (j = 0; j < n; j++) {
1026: if (PetscAbsScalar(mask[*aj])) {
1027: if (b) bb[i] -= *aa * xx[*aj];
1028: *aa = 0.0;
1029: }
1030: aa++;
1031: aj++;
1032: }
1033: }
1034: }
1035: if (x && b) {
1036: PetscCall(VecRestoreArray(b, &bb));
1037: PetscCall(VecRestoreArrayRead(l->lvec, &xx));
1038: }
1039: PetscCall(MatSeqAIJRestoreArray(l->B, &aij_a));
1040: PetscCall(VecRestoreArray(lmask, &mask));
1041: PetscCall(VecDestroy(&lmask));
1042: PetscCall(PetscFree(lrows));
1044: /* only change matrix nonzero state if pattern was allowed to be changed */
1045: if (!((Mat_SeqAIJ *)l->A->data)->nonew) {
1046: PetscObjectState state = l->A->nonzerostate + l->B->nonzerostate;
1047: PetscCallMPI(MPIU_Allreduce(&state, &A->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)A)));
1048: }
1049: PetscFunctionReturn(PETSC_SUCCESS);
1050: }
1052: static PetscErrorCode MatMult_MPIAIJ(Mat A, Vec xx, Vec yy)
1053: {
1054: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1055: PetscInt nt;
1056: VecScatter Mvctx = a->Mvctx;
1058: PetscFunctionBegin;
1059: PetscCall(VecGetLocalSize(xx, &nt));
1060: 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);
1061: PetscCall(VecScatterBegin(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1062: PetscUseTypeMethod(a->A, mult, xx, yy);
1063: PetscCall(VecScatterEnd(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1064: PetscUseTypeMethod(a->B, multadd, a->lvec, yy, yy);
1065: PetscFunctionReturn(PETSC_SUCCESS);
1066: }
1068: static PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat A, Vec bb, Vec xx)
1069: {
1070: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1072: PetscFunctionBegin;
1073: PetscCall(MatMultDiagonalBlock(a->A, bb, xx));
1074: PetscFunctionReturn(PETSC_SUCCESS);
1075: }
1077: static PetscErrorCode MatMultAdd_MPIAIJ(Mat A, Vec xx, Vec yy, Vec zz)
1078: {
1079: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1080: VecScatter Mvctx = a->Mvctx;
1082: PetscFunctionBegin;
1083: PetscCall(VecScatterBegin(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1084: PetscCall((*a->A->ops->multadd)(a->A, xx, yy, zz));
1085: PetscCall(VecScatterEnd(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1086: PetscCall((*a->B->ops->multadd)(a->B, a->lvec, zz, zz));
1087: PetscFunctionReturn(PETSC_SUCCESS);
1088: }
1090: static PetscErrorCode MatMultTranspose_MPIAIJ(Mat A, Vec xx, Vec yy)
1091: {
1092: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1094: PetscFunctionBegin;
1095: /* do nondiagonal part */
1096: PetscCall((*a->B->ops->multtranspose)(a->B, xx, a->lvec));
1097: /* do local part */
1098: PetscCall((*a->A->ops->multtranspose)(a->A, xx, yy));
1099: /* add partial results together */
1100: PetscCall(VecScatterBegin(a->Mvctx, a->lvec, yy, ADD_VALUES, SCATTER_REVERSE));
1101: PetscCall(VecScatterEnd(a->Mvctx, a->lvec, yy, ADD_VALUES, SCATTER_REVERSE));
1102: PetscFunctionReturn(PETSC_SUCCESS);
1103: }
1105: static PetscErrorCode MatIsTranspose_MPIAIJ(Mat Amat, Mat Bmat, PetscReal tol, PetscBool *f)
1106: {
1107: MPI_Comm comm;
1108: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)Amat->data, *Bij = (Mat_MPIAIJ *)Bmat->data;
1109: Mat Adia = Aij->A, Bdia = Bij->A, Aoff, Boff, *Aoffs, *Boffs;
1110: IS Me, Notme;
1111: PetscInt M, N, first, last, *notme, i;
1112: PetscBool lf;
1113: PetscMPIInt size;
1115: PetscFunctionBegin;
1116: /* Easy test: symmetric diagonal block */
1117: PetscCall(MatIsTranspose(Adia, Bdia, tol, &lf));
1118: PetscCallMPI(MPIU_Allreduce(&lf, f, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)Amat)));
1119: if (!*f) PetscFunctionReturn(PETSC_SUCCESS);
1120: PetscCall(PetscObjectGetComm((PetscObject)Amat, &comm));
1121: PetscCallMPI(MPI_Comm_size(comm, &size));
1122: if (size == 1) PetscFunctionReturn(PETSC_SUCCESS);
1124: /* Hard test: off-diagonal block. This takes a MatCreateSubMatrix. */
1125: PetscCall(MatGetSize(Amat, &M, &N));
1126: PetscCall(MatGetOwnershipRange(Amat, &first, &last));
1127: PetscCall(PetscMalloc1(N - last + first, ¬me));
1128: for (i = 0; i < first; i++) notme[i] = i;
1129: for (i = last; i < M; i++) notme[i - last + first] = i;
1130: PetscCall(ISCreateGeneral(MPI_COMM_SELF, N - last + first, notme, PETSC_COPY_VALUES, &Notme));
1131: PetscCall(ISCreateStride(MPI_COMM_SELF, last - first, first, 1, &Me));
1132: PetscCall(MatCreateSubMatrices(Amat, 1, &Me, &Notme, MAT_INITIAL_MATRIX, &Aoffs));
1133: Aoff = Aoffs[0];
1134: PetscCall(MatCreateSubMatrices(Bmat, 1, &Notme, &Me, MAT_INITIAL_MATRIX, &Boffs));
1135: Boff = Boffs[0];
1136: PetscCall(MatIsTranspose(Aoff, Boff, tol, f));
1137: PetscCall(MatDestroyMatrices(1, &Aoffs));
1138: PetscCall(MatDestroyMatrices(1, &Boffs));
1139: PetscCall(ISDestroy(&Me));
1140: PetscCall(ISDestroy(&Notme));
1141: PetscCall(PetscFree(notme));
1142: PetscFunctionReturn(PETSC_SUCCESS);
1143: }
1145: static PetscErrorCode MatMultTransposeAdd_MPIAIJ(Mat A, Vec xx, Vec yy, Vec zz)
1146: {
1147: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1149: PetscFunctionBegin;
1150: /* do nondiagonal part */
1151: PetscCall((*a->B->ops->multtranspose)(a->B, xx, a->lvec));
1152: /* do local part */
1153: PetscCall((*a->A->ops->multtransposeadd)(a->A, xx, yy, zz));
1154: /* add partial results together */
1155: PetscCall(VecScatterBegin(a->Mvctx, a->lvec, zz, ADD_VALUES, SCATTER_REVERSE));
1156: PetscCall(VecScatterEnd(a->Mvctx, a->lvec, zz, ADD_VALUES, SCATTER_REVERSE));
1157: PetscFunctionReturn(PETSC_SUCCESS);
1158: }
1160: /*
1161: This only works correctly for square matrices where the subblock A->A is the
1162: diagonal block
1163: */
1164: static PetscErrorCode MatGetDiagonal_MPIAIJ(Mat A, Vec v)
1165: {
1166: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1168: PetscFunctionBegin;
1169: PetscCheck(A->rmap->N == A->cmap->N, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Supports only square matrix where A->A is diag block");
1170: 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");
1171: PetscCall(MatGetDiagonal(a->A, v));
1172: PetscFunctionReturn(PETSC_SUCCESS);
1173: }
1175: static PetscErrorCode MatScale_MPIAIJ(Mat A, PetscScalar aa)
1176: {
1177: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1179: PetscFunctionBegin;
1180: PetscCall(MatScale(a->A, aa));
1181: PetscCall(MatScale(a->B, aa));
1182: PetscFunctionReturn(PETSC_SUCCESS);
1183: }
1185: static PetscErrorCode MatView_MPIAIJ_Binary(Mat mat, PetscViewer viewer)
1186: {
1187: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1188: Mat_SeqAIJ *A = (Mat_SeqAIJ *)aij->A->data;
1189: Mat_SeqAIJ *B = (Mat_SeqAIJ *)aij->B->data;
1190: const PetscInt *garray = aij->garray;
1191: const PetscScalar *aa, *ba;
1192: PetscInt header[4], M, N, m, rs, cs, cnt, i, ja, jb;
1193: PetscInt64 nz, hnz;
1194: PetscInt *rowlens;
1195: PetscInt *colidxs;
1196: PetscScalar *matvals;
1197: PetscMPIInt rank;
1199: PetscFunctionBegin;
1200: PetscCall(PetscViewerSetUp(viewer));
1202: M = mat->rmap->N;
1203: N = mat->cmap->N;
1204: m = mat->rmap->n;
1205: rs = mat->rmap->rstart;
1206: cs = mat->cmap->rstart;
1207: nz = A->nz + B->nz;
1209: /* write matrix header */
1210: header[0] = MAT_FILE_CLASSID;
1211: header[1] = M;
1212: header[2] = N;
1213: PetscCallMPI(MPI_Reduce(&nz, &hnz, 1, MPIU_INT64, MPI_SUM, 0, PetscObjectComm((PetscObject)mat)));
1214: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)mat), &rank));
1215: if (rank == 0) PetscCall(PetscIntCast(hnz, &header[3]));
1216: PetscCall(PetscViewerBinaryWrite(viewer, header, 4, PETSC_INT));
1218: /* fill in and store row lengths */
1219: PetscCall(PetscMalloc1(m, &rowlens));
1220: for (i = 0; i < m; i++) rowlens[i] = A->i[i + 1] - A->i[i] + B->i[i + 1] - B->i[i];
1221: PetscCall(PetscViewerBinaryWriteAll(viewer, rowlens, m, rs, M, PETSC_INT));
1222: PetscCall(PetscFree(rowlens));
1224: /* fill in and store column indices */
1225: PetscCall(PetscMalloc1(nz, &colidxs));
1226: for (cnt = 0, i = 0; i < m; i++) {
1227: for (jb = B->i[i]; jb < B->i[i + 1]; jb++) {
1228: if (garray[B->j[jb]] > cs) break;
1229: colidxs[cnt++] = garray[B->j[jb]];
1230: }
1231: for (ja = A->i[i]; ja < A->i[i + 1]; ja++) colidxs[cnt++] = A->j[ja] + cs;
1232: for (; jb < B->i[i + 1]; jb++) colidxs[cnt++] = garray[B->j[jb]];
1233: }
1234: PetscCheck(cnt == nz, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Internal PETSc error: cnt = %" PetscInt_FMT " nz = %" PetscInt64_FMT, cnt, nz);
1235: PetscCall(PetscViewerBinaryWriteAll(viewer, colidxs, nz, PETSC_DETERMINE, PETSC_DETERMINE, PETSC_INT));
1236: PetscCall(PetscFree(colidxs));
1238: /* fill in and store nonzero values */
1239: PetscCall(MatSeqAIJGetArrayRead(aij->A, &aa));
1240: PetscCall(MatSeqAIJGetArrayRead(aij->B, &ba));
1241: PetscCall(PetscMalloc1(nz, &matvals));
1242: for (cnt = 0, i = 0; i < m; i++) {
1243: for (jb = B->i[i]; jb < B->i[i + 1]; jb++) {
1244: if (garray[B->j[jb]] > cs) break;
1245: matvals[cnt++] = ba[jb];
1246: }
1247: for (ja = A->i[i]; ja < A->i[i + 1]; ja++) matvals[cnt++] = aa[ja];
1248: for (; jb < B->i[i + 1]; jb++) matvals[cnt++] = ba[jb];
1249: }
1250: PetscCall(MatSeqAIJRestoreArrayRead(aij->A, &aa));
1251: PetscCall(MatSeqAIJRestoreArrayRead(aij->B, &ba));
1252: PetscCheck(cnt == nz, PETSC_COMM_SELF, PETSC_ERR_LIB, "Internal PETSc error: cnt = %" PetscInt_FMT " nz = %" PetscInt64_FMT, cnt, nz);
1253: PetscCall(PetscViewerBinaryWriteAll(viewer, matvals, nz, PETSC_DETERMINE, PETSC_DETERMINE, PETSC_SCALAR));
1254: PetscCall(PetscFree(matvals));
1256: /* write block size option to the viewer's .info file */
1257: PetscCall(MatView_Binary_BlockSizes(mat, viewer));
1258: PetscFunctionReturn(PETSC_SUCCESS);
1259: }
1261: #include <petscdraw.h>
1262: static PetscErrorCode MatView_MPIAIJ_ASCIIorDraworSocket(Mat mat, PetscViewer viewer)
1263: {
1264: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1265: PetscMPIInt rank = aij->rank, size = aij->size;
1266: PetscBool isdraw, iascii, isbinary;
1267: PetscViewer sviewer;
1268: PetscViewerFormat format;
1270: PetscFunctionBegin;
1271: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw));
1272: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &iascii));
1273: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
1274: if (iascii) {
1275: PetscCall(PetscViewerGetFormat(viewer, &format));
1276: if (format == PETSC_VIEWER_LOAD_BALANCE) {
1277: PetscInt i, nmax = 0, nmin = PETSC_INT_MAX, navg = 0, *nz, nzlocal = ((Mat_SeqAIJ *)aij->A->data)->nz + ((Mat_SeqAIJ *)aij->B->data)->nz;
1278: PetscCall(PetscMalloc1(size, &nz));
1279: PetscCallMPI(MPI_Allgather(&nzlocal, 1, MPIU_INT, nz, 1, MPIU_INT, PetscObjectComm((PetscObject)mat)));
1280: for (i = 0; i < size; i++) {
1281: nmax = PetscMax(nmax, nz[i]);
1282: nmin = PetscMin(nmin, nz[i]);
1283: navg += nz[i];
1284: }
1285: PetscCall(PetscFree(nz));
1286: navg = navg / size;
1287: PetscCall(PetscViewerASCIIPrintf(viewer, "Load Balance - Nonzeros: Min %" PetscInt_FMT " avg %" PetscInt_FMT " max %" PetscInt_FMT "\n", nmin, navg, nmax));
1288: PetscFunctionReturn(PETSC_SUCCESS);
1289: }
1290: PetscCall(PetscViewerGetFormat(viewer, &format));
1291: if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
1292: MatInfo info;
1293: PetscInt *inodes = NULL;
1295: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)mat), &rank));
1296: PetscCall(MatGetInfo(mat, MAT_LOCAL, &info));
1297: PetscCall(MatInodeGetInodeSizes(aij->A, NULL, &inodes, NULL));
1298: PetscCall(PetscViewerASCIIPushSynchronized(viewer));
1299: if (!inodes) {
1300: 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,
1301: info.memory));
1302: } else {
1303: PetscCall(
1304: 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));
1305: }
1306: PetscCall(MatGetInfo(aij->A, MAT_LOCAL, &info));
1307: PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] on-diagonal part: nz %" PetscInt_FMT " \n", rank, (PetscInt)info.nz_used));
1308: PetscCall(MatGetInfo(aij->B, MAT_LOCAL, &info));
1309: PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] off-diagonal part: nz %" PetscInt_FMT " \n", rank, (PetscInt)info.nz_used));
1310: PetscCall(PetscViewerFlush(viewer));
1311: PetscCall(PetscViewerASCIIPopSynchronized(viewer));
1312: PetscCall(PetscViewerASCIIPrintf(viewer, "Information on VecScatter used in matrix-vector product: \n"));
1313: PetscCall(VecScatterView(aij->Mvctx, viewer));
1314: PetscFunctionReturn(PETSC_SUCCESS);
1315: } else if (format == PETSC_VIEWER_ASCII_INFO) {
1316: PetscInt inodecount, inodelimit, *inodes;
1317: PetscCall(MatInodeGetInodeSizes(aij->A, &inodecount, &inodes, &inodelimit));
1318: if (inodes) {
1319: PetscCall(PetscViewerASCIIPrintf(viewer, "using I-node (on process 0) routines: found %" PetscInt_FMT " nodes, limit used is %" PetscInt_FMT "\n", inodecount, inodelimit));
1320: } else {
1321: PetscCall(PetscViewerASCIIPrintf(viewer, "not using I-node (on process 0) routines\n"));
1322: }
1323: PetscFunctionReturn(PETSC_SUCCESS);
1324: } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) {
1325: PetscFunctionReturn(PETSC_SUCCESS);
1326: }
1327: } else if (isbinary) {
1328: if (size == 1) {
1329: PetscCall(PetscObjectSetName((PetscObject)aij->A, ((PetscObject)mat)->name));
1330: PetscCall(MatView(aij->A, viewer));
1331: } else {
1332: PetscCall(MatView_MPIAIJ_Binary(mat, viewer));
1333: }
1334: PetscFunctionReturn(PETSC_SUCCESS);
1335: } else if (iascii && size == 1) {
1336: PetscCall(PetscObjectSetName((PetscObject)aij->A, ((PetscObject)mat)->name));
1337: PetscCall(MatView(aij->A, viewer));
1338: PetscFunctionReturn(PETSC_SUCCESS);
1339: } else if (isdraw) {
1340: PetscDraw draw;
1341: PetscBool isnull;
1342: PetscCall(PetscViewerDrawGetDraw(viewer, 0, &draw));
1343: PetscCall(PetscDrawIsNull(draw, &isnull));
1344: if (isnull) PetscFunctionReturn(PETSC_SUCCESS);
1345: }
1347: { /* assemble the entire matrix onto first processor */
1348: Mat A = NULL, Av;
1349: IS isrow, iscol;
1351: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat), rank == 0 ? mat->rmap->N : 0, 0, 1, &isrow));
1352: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat), rank == 0 ? mat->cmap->N : 0, 0, 1, &iscol));
1353: PetscCall(MatCreateSubMatrix(mat, isrow, iscol, MAT_INITIAL_MATRIX, &A));
1354: PetscCall(MatMPIAIJGetSeqAIJ(A, &Av, NULL, NULL));
1355: /* The commented code uses MatCreateSubMatrices instead */
1356: /*
1357: Mat *AA, A = NULL, Av;
1358: IS isrow,iscol;
1360: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat),rank == 0 ? mat->rmap->N : 0,0,1,&isrow));
1361: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat),rank == 0 ? mat->cmap->N : 0,0,1,&iscol));
1362: PetscCall(MatCreateSubMatrices(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,&AA));
1363: if (rank == 0) {
1364: PetscCall(PetscObjectReference((PetscObject)AA[0]));
1365: A = AA[0];
1366: Av = AA[0];
1367: }
1368: PetscCall(MatDestroySubMatrices(1,&AA));
1369: */
1370: PetscCall(ISDestroy(&iscol));
1371: PetscCall(ISDestroy(&isrow));
1372: /*
1373: Everyone has to call to draw the matrix since the graphics waits are
1374: synchronized across all processors that share the PetscDraw object
1375: */
1376: PetscCall(PetscViewerGetSubViewer(viewer, PETSC_COMM_SELF, &sviewer));
1377: if (rank == 0) {
1378: if (((PetscObject)mat)->name) PetscCall(PetscObjectSetName((PetscObject)Av, ((PetscObject)mat)->name));
1379: PetscCall(MatView_SeqAIJ(Av, sviewer));
1380: }
1381: PetscCall(PetscViewerRestoreSubViewer(viewer, PETSC_COMM_SELF, &sviewer));
1382: PetscCall(MatDestroy(&A));
1383: }
1384: PetscFunctionReturn(PETSC_SUCCESS);
1385: }
1387: PetscErrorCode MatView_MPIAIJ(Mat mat, PetscViewer viewer)
1388: {
1389: PetscBool iascii, isdraw, issocket, isbinary;
1391: PetscFunctionBegin;
1392: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &iascii));
1393: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw));
1394: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
1395: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERSOCKET, &issocket));
1396: if (iascii || isdraw || isbinary || issocket) PetscCall(MatView_MPIAIJ_ASCIIorDraworSocket(mat, viewer));
1397: PetscFunctionReturn(PETSC_SUCCESS);
1398: }
1400: static PetscErrorCode MatSOR_MPIAIJ(Mat matin, Vec bb, PetscReal omega, MatSORType flag, PetscReal fshift, PetscInt its, PetscInt lits, Vec xx)
1401: {
1402: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)matin->data;
1403: Vec bb1 = NULL;
1404: PetscBool hasop;
1406: PetscFunctionBegin;
1407: if (flag == SOR_APPLY_UPPER) {
1408: PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
1409: PetscFunctionReturn(PETSC_SUCCESS);
1410: }
1412: if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS || flag & SOR_EISENSTAT) PetscCall(VecDuplicate(bb, &bb1));
1414: if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) {
1415: if (flag & SOR_ZERO_INITIAL_GUESS) {
1416: PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
1417: its--;
1418: }
1420: while (its--) {
1421: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1422: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1424: /* update rhs: bb1 = bb - B*x */
1425: PetscCall(VecScale(mat->lvec, -1.0));
1426: PetscCall((*mat->B->ops->multadd)(mat->B, mat->lvec, bb, bb1));
1428: /* local sweep */
1429: PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, SOR_SYMMETRIC_SWEEP, fshift, lits, 1, xx));
1430: }
1431: } else if (flag & SOR_LOCAL_FORWARD_SWEEP) {
1432: if (flag & SOR_ZERO_INITIAL_GUESS) {
1433: PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
1434: its--;
1435: }
1436: while (its--) {
1437: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1438: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1440: /* update rhs: bb1 = bb - B*x */
1441: PetscCall(VecScale(mat->lvec, -1.0));
1442: PetscCall((*mat->B->ops->multadd)(mat->B, mat->lvec, bb, bb1));
1444: /* local sweep */
1445: PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, SOR_FORWARD_SWEEP, fshift, lits, 1, xx));
1446: }
1447: } else if (flag & SOR_LOCAL_BACKWARD_SWEEP) {
1448: if (flag & SOR_ZERO_INITIAL_GUESS) {
1449: PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
1450: its--;
1451: }
1452: while (its--) {
1453: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1454: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1456: /* update rhs: bb1 = bb - B*x */
1457: PetscCall(VecScale(mat->lvec, -1.0));
1458: PetscCall((*mat->B->ops->multadd)(mat->B, mat->lvec, bb, bb1));
1460: /* local sweep */
1461: PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, SOR_BACKWARD_SWEEP, fshift, lits, 1, xx));
1462: }
1463: } else if (flag & SOR_EISENSTAT) {
1464: Vec xx1;
1466: PetscCall(VecDuplicate(bb, &xx1));
1467: PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, (MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_BACKWARD_SWEEP), fshift, lits, 1, xx));
1469: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1470: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1471: if (!mat->diag) {
1472: PetscCall(MatCreateVecs(matin, &mat->diag, NULL));
1473: PetscCall(MatGetDiagonal(matin, mat->diag));
1474: }
1475: PetscCall(MatHasOperation(matin, MATOP_MULT_DIAGONAL_BLOCK, &hasop));
1476: if (hasop) {
1477: PetscCall(MatMultDiagonalBlock(matin, xx, bb1));
1478: } else {
1479: PetscCall(VecPointwiseMult(bb1, mat->diag, xx));
1480: }
1481: PetscCall(VecAYPX(bb1, (omega - 2.0) / omega, bb));
1483: PetscCall(MatMultAdd(mat->B, mat->lvec, bb1, bb1));
1485: /* local sweep */
1486: PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, (MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_FORWARD_SWEEP), fshift, lits, 1, xx1));
1487: PetscCall(VecAXPY(xx, 1.0, xx1));
1488: PetscCall(VecDestroy(&xx1));
1489: } else SETERRQ(PetscObjectComm((PetscObject)matin), PETSC_ERR_SUP, "Parallel SOR not supported");
1491: PetscCall(VecDestroy(&bb1));
1493: matin->factorerrortype = mat->A->factorerrortype;
1494: PetscFunctionReturn(PETSC_SUCCESS);
1495: }
1497: static PetscErrorCode MatPermute_MPIAIJ(Mat A, IS rowp, IS colp, Mat *B)
1498: {
1499: Mat aA, aB, Aperm;
1500: const PetscInt *rwant, *cwant, *gcols, *ai, *bi, *aj, *bj;
1501: PetscScalar *aa, *ba;
1502: PetscInt i, j, m, n, ng, anz, bnz, *dnnz, *onnz, *tdnnz, *tonnz, *rdest, *cdest, *work, *gcdest;
1503: PetscSF rowsf, sf;
1504: IS parcolp = NULL;
1505: PetscBool done;
1507: PetscFunctionBegin;
1508: PetscCall(MatGetLocalSize(A, &m, &n));
1509: PetscCall(ISGetIndices(rowp, &rwant));
1510: PetscCall(ISGetIndices(colp, &cwant));
1511: PetscCall(PetscMalloc3(PetscMax(m, n), &work, m, &rdest, n, &cdest));
1513: /* Invert row permutation to find out where my rows should go */
1514: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &rowsf));
1515: PetscCall(PetscSFSetGraphLayout(rowsf, A->rmap, A->rmap->n, NULL, PETSC_OWN_POINTER, rwant));
1516: PetscCall(PetscSFSetFromOptions(rowsf));
1517: for (i = 0; i < m; i++) work[i] = A->rmap->rstart + i;
1518: PetscCall(PetscSFReduceBegin(rowsf, MPIU_INT, work, rdest, MPI_REPLACE));
1519: PetscCall(PetscSFReduceEnd(rowsf, MPIU_INT, work, rdest, MPI_REPLACE));
1521: /* Invert column permutation to find out where my columns should go */
1522: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
1523: PetscCall(PetscSFSetGraphLayout(sf, A->cmap, A->cmap->n, NULL, PETSC_OWN_POINTER, cwant));
1524: PetscCall(PetscSFSetFromOptions(sf));
1525: for (i = 0; i < n; i++) work[i] = A->cmap->rstart + i;
1526: PetscCall(PetscSFReduceBegin(sf, MPIU_INT, work, cdest, MPI_REPLACE));
1527: PetscCall(PetscSFReduceEnd(sf, MPIU_INT, work, cdest, MPI_REPLACE));
1528: PetscCall(PetscSFDestroy(&sf));
1530: PetscCall(ISRestoreIndices(rowp, &rwant));
1531: PetscCall(ISRestoreIndices(colp, &cwant));
1532: PetscCall(MatMPIAIJGetSeqAIJ(A, &aA, &aB, &gcols));
1534: /* Find out where my gcols should go */
1535: PetscCall(MatGetSize(aB, NULL, &ng));
1536: PetscCall(PetscMalloc1(ng, &gcdest));
1537: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
1538: PetscCall(PetscSFSetGraphLayout(sf, A->cmap, ng, NULL, PETSC_OWN_POINTER, gcols));
1539: PetscCall(PetscSFSetFromOptions(sf));
1540: PetscCall(PetscSFBcastBegin(sf, MPIU_INT, cdest, gcdest, MPI_REPLACE));
1541: PetscCall(PetscSFBcastEnd(sf, MPIU_INT, cdest, gcdest, MPI_REPLACE));
1542: PetscCall(PetscSFDestroy(&sf));
1544: PetscCall(PetscCalloc4(m, &dnnz, m, &onnz, m, &tdnnz, m, &tonnz));
1545: PetscCall(MatGetRowIJ(aA, 0, PETSC_FALSE, PETSC_FALSE, &anz, &ai, &aj, &done));
1546: PetscCall(MatGetRowIJ(aB, 0, PETSC_FALSE, PETSC_FALSE, &bnz, &bi, &bj, &done));
1547: for (i = 0; i < m; i++) {
1548: PetscInt row = rdest[i];
1549: PetscMPIInt rowner;
1550: PetscCall(PetscLayoutFindOwner(A->rmap, row, &rowner));
1551: for (j = ai[i]; j < ai[i + 1]; j++) {
1552: PetscInt col = cdest[aj[j]];
1553: PetscMPIInt cowner;
1554: PetscCall(PetscLayoutFindOwner(A->cmap, col, &cowner)); /* Could build an index for the columns to eliminate this search */
1555: if (rowner == cowner) dnnz[i]++;
1556: else onnz[i]++;
1557: }
1558: for (j = bi[i]; j < bi[i + 1]; j++) {
1559: PetscInt col = gcdest[bj[j]];
1560: PetscMPIInt cowner;
1561: PetscCall(PetscLayoutFindOwner(A->cmap, col, &cowner));
1562: if (rowner == cowner) dnnz[i]++;
1563: else onnz[i]++;
1564: }
1565: }
1566: PetscCall(PetscSFBcastBegin(rowsf, MPIU_INT, dnnz, tdnnz, MPI_REPLACE));
1567: PetscCall(PetscSFBcastEnd(rowsf, MPIU_INT, dnnz, tdnnz, MPI_REPLACE));
1568: PetscCall(PetscSFBcastBegin(rowsf, MPIU_INT, onnz, tonnz, MPI_REPLACE));
1569: PetscCall(PetscSFBcastEnd(rowsf, MPIU_INT, onnz, tonnz, MPI_REPLACE));
1570: PetscCall(PetscSFDestroy(&rowsf));
1572: PetscCall(MatCreateAIJ(PetscObjectComm((PetscObject)A), A->rmap->n, A->cmap->n, A->rmap->N, A->cmap->N, 0, tdnnz, 0, tonnz, &Aperm));
1573: PetscCall(MatSeqAIJGetArray(aA, &aa));
1574: PetscCall(MatSeqAIJGetArray(aB, &ba));
1575: for (i = 0; i < m; i++) {
1576: PetscInt *acols = dnnz, *bcols = onnz; /* Repurpose now-unneeded arrays */
1577: PetscInt j0, rowlen;
1578: rowlen = ai[i + 1] - ai[i];
1579: for (j0 = j = 0; j < rowlen; j0 = j) { /* rowlen could be larger than number of rows m, so sum in batches */
1580: for (; j < PetscMin(rowlen, j0 + m); j++) acols[j - j0] = cdest[aj[ai[i] + j]];
1581: PetscCall(MatSetValues(Aperm, 1, &rdest[i], j - j0, acols, aa + ai[i] + j0, INSERT_VALUES));
1582: }
1583: rowlen = bi[i + 1] - bi[i];
1584: for (j0 = j = 0; j < rowlen; j0 = j) {
1585: for (; j < PetscMin(rowlen, j0 + m); j++) bcols[j - j0] = gcdest[bj[bi[i] + j]];
1586: PetscCall(MatSetValues(Aperm, 1, &rdest[i], j - j0, bcols, ba + bi[i] + j0, INSERT_VALUES));
1587: }
1588: }
1589: PetscCall(MatAssemblyBegin(Aperm, MAT_FINAL_ASSEMBLY));
1590: PetscCall(MatAssemblyEnd(Aperm, MAT_FINAL_ASSEMBLY));
1591: PetscCall(MatRestoreRowIJ(aA, 0, PETSC_FALSE, PETSC_FALSE, &anz, &ai, &aj, &done));
1592: PetscCall(MatRestoreRowIJ(aB, 0, PETSC_FALSE, PETSC_FALSE, &bnz, &bi, &bj, &done));
1593: PetscCall(MatSeqAIJRestoreArray(aA, &aa));
1594: PetscCall(MatSeqAIJRestoreArray(aB, &ba));
1595: PetscCall(PetscFree4(dnnz, onnz, tdnnz, tonnz));
1596: PetscCall(PetscFree3(work, rdest, cdest));
1597: PetscCall(PetscFree(gcdest));
1598: if (parcolp) PetscCall(ISDestroy(&colp));
1599: *B = Aperm;
1600: PetscFunctionReturn(PETSC_SUCCESS);
1601: }
1603: static PetscErrorCode MatGetGhosts_MPIAIJ(Mat mat, PetscInt *nghosts, const PetscInt *ghosts[])
1604: {
1605: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1607: PetscFunctionBegin;
1608: PetscCall(MatGetSize(aij->B, NULL, nghosts));
1609: if (ghosts) *ghosts = aij->garray;
1610: PetscFunctionReturn(PETSC_SUCCESS);
1611: }
1613: static PetscErrorCode MatGetInfo_MPIAIJ(Mat matin, MatInfoType flag, MatInfo *info)
1614: {
1615: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)matin->data;
1616: Mat A = mat->A, B = mat->B;
1617: PetscLogDouble isend[5], irecv[5];
1619: PetscFunctionBegin;
1620: info->block_size = 1.0;
1621: PetscCall(MatGetInfo(A, MAT_LOCAL, info));
1623: isend[0] = info->nz_used;
1624: isend[1] = info->nz_allocated;
1625: isend[2] = info->nz_unneeded;
1626: isend[3] = info->memory;
1627: isend[4] = info->mallocs;
1629: PetscCall(MatGetInfo(B, MAT_LOCAL, info));
1631: isend[0] += info->nz_used;
1632: isend[1] += info->nz_allocated;
1633: isend[2] += info->nz_unneeded;
1634: isend[3] += info->memory;
1635: isend[4] += info->mallocs;
1636: if (flag == MAT_LOCAL) {
1637: info->nz_used = isend[0];
1638: info->nz_allocated = isend[1];
1639: info->nz_unneeded = isend[2];
1640: info->memory = isend[3];
1641: info->mallocs = isend[4];
1642: } else if (flag == MAT_GLOBAL_MAX) {
1643: PetscCallMPI(MPIU_Allreduce(isend, irecv, 5, MPIU_PETSCLOGDOUBLE, MPI_MAX, PetscObjectComm((PetscObject)matin)));
1645: info->nz_used = irecv[0];
1646: info->nz_allocated = irecv[1];
1647: info->nz_unneeded = irecv[2];
1648: info->memory = irecv[3];
1649: info->mallocs = irecv[4];
1650: } else if (flag == MAT_GLOBAL_SUM) {
1651: PetscCallMPI(MPIU_Allreduce(isend, irecv, 5, MPIU_PETSCLOGDOUBLE, MPI_SUM, PetscObjectComm((PetscObject)matin)));
1653: info->nz_used = irecv[0];
1654: info->nz_allocated = irecv[1];
1655: info->nz_unneeded = irecv[2];
1656: info->memory = irecv[3];
1657: info->mallocs = irecv[4];
1658: }
1659: info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */
1660: info->fill_ratio_needed = 0;
1661: info->factor_mallocs = 0;
1662: PetscFunctionReturn(PETSC_SUCCESS);
1663: }
1665: PetscErrorCode MatSetOption_MPIAIJ(Mat A, MatOption op, PetscBool flg)
1666: {
1667: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1669: PetscFunctionBegin;
1670: switch (op) {
1671: case MAT_NEW_NONZERO_LOCATIONS:
1672: case MAT_NEW_NONZERO_ALLOCATION_ERR:
1673: case MAT_UNUSED_NONZERO_LOCATION_ERR:
1674: case MAT_KEEP_NONZERO_PATTERN:
1675: case MAT_NEW_NONZERO_LOCATION_ERR:
1676: case MAT_USE_INODES:
1677: case MAT_IGNORE_ZERO_ENTRIES:
1678: case MAT_FORM_EXPLICIT_TRANSPOSE:
1679: MatCheckPreallocated(A, 1);
1680: PetscCall(MatSetOption(a->A, op, flg));
1681: PetscCall(MatSetOption(a->B, op, flg));
1682: break;
1683: case MAT_ROW_ORIENTED:
1684: MatCheckPreallocated(A, 1);
1685: a->roworiented = flg;
1687: PetscCall(MatSetOption(a->A, op, flg));
1688: PetscCall(MatSetOption(a->B, op, flg));
1689: break;
1690: case MAT_FORCE_DIAGONAL_ENTRIES:
1691: case MAT_SORTED_FULL:
1692: PetscCall(PetscInfo(A, "Option %s ignored\n", MatOptions[op]));
1693: break;
1694: case MAT_IGNORE_OFF_PROC_ENTRIES:
1695: a->donotstash = flg;
1696: break;
1697: /* Symmetry flags are handled directly by MatSetOption() and they don't affect preallocation */
1698: case MAT_SPD:
1699: case MAT_SYMMETRIC:
1700: case MAT_STRUCTURALLY_SYMMETRIC:
1701: case MAT_HERMITIAN:
1702: case MAT_SYMMETRY_ETERNAL:
1703: case MAT_STRUCTURAL_SYMMETRY_ETERNAL:
1704: case MAT_SPD_ETERNAL:
1705: /* if the diagonal matrix is square it inherits some of the properties above */
1706: break;
1707: case MAT_SUBMAT_SINGLEIS:
1708: A->submat_singleis = flg;
1709: break;
1710: case MAT_STRUCTURE_ONLY:
1711: /* The option is handled directly by MatSetOption() */
1712: break;
1713: default:
1714: SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "unknown option %d", op);
1715: }
1716: PetscFunctionReturn(PETSC_SUCCESS);
1717: }
1719: PetscErrorCode MatGetRow_MPIAIJ(Mat matin, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
1720: {
1721: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)matin->data;
1722: PetscScalar *vworkA, *vworkB, **pvA, **pvB, *v_p;
1723: PetscInt i, *cworkA, *cworkB, **pcA, **pcB, cstart = matin->cmap->rstart;
1724: PetscInt nztot, nzA, nzB, lrow, rstart = matin->rmap->rstart, rend = matin->rmap->rend;
1725: PetscInt *cmap, *idx_p;
1727: PetscFunctionBegin;
1728: PetscCheck(!mat->getrowactive, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Already active");
1729: mat->getrowactive = PETSC_TRUE;
1731: if (!mat->rowvalues && (idx || v)) {
1732: /*
1733: allocate enough space to hold information from the longest row.
1734: */
1735: Mat_SeqAIJ *Aa = (Mat_SeqAIJ *)mat->A->data, *Ba = (Mat_SeqAIJ *)mat->B->data;
1736: PetscInt max = 1, tmp;
1737: for (i = 0; i < matin->rmap->n; i++) {
1738: tmp = Aa->i[i + 1] - Aa->i[i] + Ba->i[i + 1] - Ba->i[i];
1739: if (max < tmp) max = tmp;
1740: }
1741: PetscCall(PetscMalloc2(max, &mat->rowvalues, max, &mat->rowindices));
1742: }
1744: PetscCheck(row >= rstart && row < rend, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Only local rows");
1745: lrow = row - rstart;
1747: pvA = &vworkA;
1748: pcA = &cworkA;
1749: pvB = &vworkB;
1750: pcB = &cworkB;
1751: if (!v) {
1752: pvA = NULL;
1753: pvB = NULL;
1754: }
1755: if (!idx) {
1756: pcA = NULL;
1757: if (!v) pcB = NULL;
1758: }
1759: PetscCall((*mat->A->ops->getrow)(mat->A, lrow, &nzA, pcA, pvA));
1760: PetscCall((*mat->B->ops->getrow)(mat->B, lrow, &nzB, pcB, pvB));
1761: nztot = nzA + nzB;
1763: cmap = mat->garray;
1764: if (v || idx) {
1765: if (nztot) {
1766: /* Sort by increasing column numbers, assuming A and B already sorted */
1767: PetscInt imark = -1;
1768: if (v) {
1769: *v = v_p = mat->rowvalues;
1770: for (i = 0; i < nzB; i++) {
1771: if (cmap[cworkB[i]] < cstart) v_p[i] = vworkB[i];
1772: else break;
1773: }
1774: imark = i;
1775: for (i = 0; i < nzA; i++) v_p[imark + i] = vworkA[i];
1776: for (i = imark; i < nzB; i++) v_p[nzA + i] = vworkB[i];
1777: }
1778: if (idx) {
1779: *idx = idx_p = mat->rowindices;
1780: if (imark > -1) {
1781: for (i = 0; i < imark; i++) idx_p[i] = cmap[cworkB[i]];
1782: } else {
1783: for (i = 0; i < nzB; i++) {
1784: if (cmap[cworkB[i]] < cstart) idx_p[i] = cmap[cworkB[i]];
1785: else break;
1786: }
1787: imark = i;
1788: }
1789: for (i = 0; i < nzA; i++) idx_p[imark + i] = cstart + cworkA[i];
1790: for (i = imark; i < nzB; i++) idx_p[nzA + i] = cmap[cworkB[i]];
1791: }
1792: } else {
1793: if (idx) *idx = NULL;
1794: if (v) *v = NULL;
1795: }
1796: }
1797: *nz = nztot;
1798: PetscCall((*mat->A->ops->restorerow)(mat->A, lrow, &nzA, pcA, pvA));
1799: PetscCall((*mat->B->ops->restorerow)(mat->B, lrow, &nzB, pcB, pvB));
1800: PetscFunctionReturn(PETSC_SUCCESS);
1801: }
1803: PetscErrorCode MatRestoreRow_MPIAIJ(Mat mat, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
1804: {
1805: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1807: PetscFunctionBegin;
1808: PetscCheck(aij->getrowactive, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "MatGetRow() must be called first");
1809: aij->getrowactive = PETSC_FALSE;
1810: PetscFunctionReturn(PETSC_SUCCESS);
1811: }
1813: static PetscErrorCode MatNorm_MPIAIJ(Mat mat, NormType type, PetscReal *norm)
1814: {
1815: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1816: Mat_SeqAIJ *amat = (Mat_SeqAIJ *)aij->A->data, *bmat = (Mat_SeqAIJ *)aij->B->data;
1817: PetscInt i, j, cstart = mat->cmap->rstart;
1818: PetscReal sum = 0.0;
1819: const MatScalar *v, *amata, *bmata;
1821: PetscFunctionBegin;
1822: if (aij->size == 1) {
1823: PetscCall(MatNorm(aij->A, type, norm));
1824: } else {
1825: PetscCall(MatSeqAIJGetArrayRead(aij->A, &amata));
1826: PetscCall(MatSeqAIJGetArrayRead(aij->B, &bmata));
1827: if (type == NORM_FROBENIUS) {
1828: v = amata;
1829: for (i = 0; i < amat->nz; i++) {
1830: sum += PetscRealPart(PetscConj(*v) * (*v));
1831: v++;
1832: }
1833: v = bmata;
1834: for (i = 0; i < bmat->nz; i++) {
1835: sum += PetscRealPart(PetscConj(*v) * (*v));
1836: v++;
1837: }
1838: PetscCallMPI(MPIU_Allreduce(&sum, norm, 1, MPIU_REAL, MPIU_SUM, PetscObjectComm((PetscObject)mat)));
1839: *norm = PetscSqrtReal(*norm);
1840: PetscCall(PetscLogFlops(2.0 * amat->nz + 2.0 * bmat->nz));
1841: } else if (type == NORM_1) { /* max column norm */
1842: PetscReal *tmp;
1843: PetscInt *jj, *garray = aij->garray;
1844: PetscCall(PetscCalloc1(mat->cmap->N + 1, &tmp));
1845: *norm = 0.0;
1846: v = amata;
1847: jj = amat->j;
1848: for (j = 0; j < amat->nz; j++) {
1849: tmp[cstart + *jj++] += PetscAbsScalar(*v);
1850: v++;
1851: }
1852: v = bmata;
1853: jj = bmat->j;
1854: for (j = 0; j < bmat->nz; j++) {
1855: tmp[garray[*jj++]] += PetscAbsScalar(*v);
1856: v++;
1857: }
1858: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, tmp, mat->cmap->N, MPIU_REAL, MPIU_SUM, PetscObjectComm((PetscObject)mat)));
1859: for (j = 0; j < mat->cmap->N; j++) {
1860: if (tmp[j] > *norm) *norm = tmp[j];
1861: }
1862: PetscCall(PetscFree(tmp));
1863: PetscCall(PetscLogFlops(PetscMax(amat->nz + bmat->nz - 1, 0)));
1864: } else if (type == NORM_INFINITY) { /* max row norm */
1865: PetscReal ntemp = 0.0;
1866: for (j = 0; j < aij->A->rmap->n; j++) {
1867: v = PetscSafePointerPlusOffset(amata, amat->i[j]);
1868: sum = 0.0;
1869: for (i = 0; i < amat->i[j + 1] - amat->i[j]; i++) {
1870: sum += PetscAbsScalar(*v);
1871: v++;
1872: }
1873: v = PetscSafePointerPlusOffset(bmata, bmat->i[j]);
1874: for (i = 0; i < bmat->i[j + 1] - bmat->i[j]; i++) {
1875: sum += PetscAbsScalar(*v);
1876: v++;
1877: }
1878: if (sum > ntemp) ntemp = sum;
1879: }
1880: PetscCallMPI(MPIU_Allreduce(&ntemp, norm, 1, MPIU_REAL, MPIU_MAX, PetscObjectComm((PetscObject)mat)));
1881: PetscCall(PetscLogFlops(PetscMax(amat->nz + bmat->nz - 1, 0)));
1882: } else SETERRQ(PetscObjectComm((PetscObject)mat), PETSC_ERR_SUP, "No support for two norm");
1883: PetscCall(MatSeqAIJRestoreArrayRead(aij->A, &amata));
1884: PetscCall(MatSeqAIJRestoreArrayRead(aij->B, &bmata));
1885: }
1886: PetscFunctionReturn(PETSC_SUCCESS);
1887: }
1889: static PetscErrorCode MatTranspose_MPIAIJ(Mat A, MatReuse reuse, Mat *matout)
1890: {
1891: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data, *b;
1892: Mat_SeqAIJ *Aloc = (Mat_SeqAIJ *)a->A->data, *Bloc = (Mat_SeqAIJ *)a->B->data, *sub_B_diag;
1893: 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;
1894: const PetscInt *ai, *aj, *bi, *bj, *B_diag_i;
1895: Mat B, A_diag, *B_diag;
1896: const MatScalar *pbv, *bv;
1898: PetscFunctionBegin;
1899: if (reuse == MAT_REUSE_MATRIX) PetscCall(MatTransposeCheckNonzeroState_Private(A, *matout));
1900: ma = A->rmap->n;
1901: na = A->cmap->n;
1902: mb = a->B->rmap->n;
1903: nb = a->B->cmap->n;
1904: ai = Aloc->i;
1905: aj = Aloc->j;
1906: bi = Bloc->i;
1907: bj = Bloc->j;
1908: if (reuse == MAT_INITIAL_MATRIX || *matout == A) {
1909: PetscInt *d_nnz, *g_nnz, *o_nnz;
1910: PetscSFNode *oloc;
1911: PETSC_UNUSED PetscSF sf;
1913: PetscCall(PetscMalloc4(na, &d_nnz, na, &o_nnz, nb, &g_nnz, nb, &oloc));
1914: /* compute d_nnz for preallocation */
1915: PetscCall(PetscArrayzero(d_nnz, na));
1916: for (i = 0; i < ai[ma]; i++) d_nnz[aj[i]]++;
1917: /* compute local off-diagonal contributions */
1918: PetscCall(PetscArrayzero(g_nnz, nb));
1919: for (i = 0; i < bi[ma]; i++) g_nnz[bj[i]]++;
1920: /* map those to global */
1921: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
1922: PetscCall(PetscSFSetGraphLayout(sf, A->cmap, nb, NULL, PETSC_USE_POINTER, a->garray));
1923: PetscCall(PetscSFSetFromOptions(sf));
1924: PetscCall(PetscArrayzero(o_nnz, na));
1925: PetscCall(PetscSFReduceBegin(sf, MPIU_INT, g_nnz, o_nnz, MPI_SUM));
1926: PetscCall(PetscSFReduceEnd(sf, MPIU_INT, g_nnz, o_nnz, MPI_SUM));
1927: PetscCall(PetscSFDestroy(&sf));
1929: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &B));
1930: PetscCall(MatSetSizes(B, A->cmap->n, A->rmap->n, N, M));
1931: PetscCall(MatSetBlockSizes(B, PetscAbs(A->cmap->bs), PetscAbs(A->rmap->bs)));
1932: PetscCall(MatSetType(B, ((PetscObject)A)->type_name));
1933: PetscCall(MatMPIAIJSetPreallocation(B, 0, d_nnz, 0, o_nnz));
1934: PetscCall(PetscFree4(d_nnz, o_nnz, g_nnz, oloc));
1935: } else {
1936: B = *matout;
1937: PetscCall(MatSetOption(B, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_TRUE));
1938: }
1940: b = (Mat_MPIAIJ *)B->data;
1941: A_diag = a->A;
1942: B_diag = &b->A;
1943: sub_B_diag = (Mat_SeqAIJ *)(*B_diag)->data;
1944: A_diag_ncol = A_diag->cmap->N;
1945: B_diag_ilen = sub_B_diag->ilen;
1946: B_diag_i = sub_B_diag->i;
1948: /* Set ilen for diagonal of B */
1949: for (i = 0; i < A_diag_ncol; i++) B_diag_ilen[i] = B_diag_i[i + 1] - B_diag_i[i];
1951: /* Transpose the diagonal part of the matrix. In contrast to the off-diagonal part, this can be done
1952: very quickly (=without using MatSetValues), because all writes are local. */
1953: PetscCall(MatTransposeSetPrecursor(A_diag, *B_diag));
1954: PetscCall(MatTranspose(A_diag, MAT_REUSE_MATRIX, B_diag));
1956: /* copy over the B part */
1957: PetscCall(PetscMalloc1(bi[mb], &cols));
1958: PetscCall(MatSeqAIJGetArrayRead(a->B, &bv));
1959: pbv = bv;
1960: row = A->rmap->rstart;
1961: for (i = 0; i < bi[mb]; i++) cols[i] = a->garray[bj[i]];
1962: cols_tmp = cols;
1963: for (i = 0; i < mb; i++) {
1964: ncol = bi[i + 1] - bi[i];
1965: PetscCall(MatSetValues(B, ncol, cols_tmp, 1, &row, pbv, INSERT_VALUES));
1966: row++;
1967: if (pbv) pbv += ncol;
1968: if (cols_tmp) cols_tmp += ncol;
1969: }
1970: PetscCall(PetscFree(cols));
1971: PetscCall(MatSeqAIJRestoreArrayRead(a->B, &bv));
1973: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
1974: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
1975: if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX) {
1976: *matout = B;
1977: } else {
1978: PetscCall(MatHeaderMerge(A, &B));
1979: }
1980: PetscFunctionReturn(PETSC_SUCCESS);
1981: }
1983: static PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat, Vec ll, Vec rr)
1984: {
1985: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1986: Mat a = aij->A, b = aij->B;
1987: PetscInt s1, s2, s3;
1989: PetscFunctionBegin;
1990: PetscCall(MatGetLocalSize(mat, &s2, &s3));
1991: if (rr) {
1992: PetscCall(VecGetLocalSize(rr, &s1));
1993: PetscCheck(s1 == s3, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "right vector non-conforming local size");
1994: /* Overlap communication with computation. */
1995: PetscCall(VecScatterBegin(aij->Mvctx, rr, aij->lvec, INSERT_VALUES, SCATTER_FORWARD));
1996: }
1997: if (ll) {
1998: PetscCall(VecGetLocalSize(ll, &s1));
1999: PetscCheck(s1 == s2, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "left vector non-conforming local size");
2000: PetscUseTypeMethod(b, diagonalscale, ll, NULL);
2001: }
2002: /* scale the diagonal block */
2003: PetscUseTypeMethod(a, diagonalscale, ll, rr);
2005: if (rr) {
2006: /* Do a scatter end and then right scale the off-diagonal block */
2007: PetscCall(VecScatterEnd(aij->Mvctx, rr, aij->lvec, INSERT_VALUES, SCATTER_FORWARD));
2008: PetscUseTypeMethod(b, diagonalscale, NULL, aij->lvec);
2009: }
2010: PetscFunctionReturn(PETSC_SUCCESS);
2011: }
2013: static PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A)
2014: {
2015: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2017: PetscFunctionBegin;
2018: PetscCall(MatSetUnfactored(a->A));
2019: PetscFunctionReturn(PETSC_SUCCESS);
2020: }
2022: static PetscErrorCode MatEqual_MPIAIJ(Mat A, Mat B, PetscBool *flag)
2023: {
2024: Mat_MPIAIJ *matB = (Mat_MPIAIJ *)B->data, *matA = (Mat_MPIAIJ *)A->data;
2025: Mat a, b, c, d;
2026: PetscBool flg;
2028: PetscFunctionBegin;
2029: a = matA->A;
2030: b = matA->B;
2031: c = matB->A;
2032: d = matB->B;
2034: PetscCall(MatEqual(a, c, &flg));
2035: if (flg) PetscCall(MatEqual(b, d, &flg));
2036: PetscCallMPI(MPIU_Allreduce(&flg, flag, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)A)));
2037: PetscFunctionReturn(PETSC_SUCCESS);
2038: }
2040: static PetscErrorCode MatCopy_MPIAIJ(Mat A, Mat B, MatStructure str)
2041: {
2042: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2043: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2045: PetscFunctionBegin;
2046: /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */
2047: if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) {
2048: /* because of the column compression in the off-processor part of the matrix a->B,
2049: the number of columns in a->B and b->B may be different, hence we cannot call
2050: the MatCopy() directly on the two parts. If need be, we can provide a more
2051: efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices
2052: then copying the submatrices */
2053: PetscCall(MatCopy_Basic(A, B, str));
2054: } else {
2055: PetscCall(MatCopy(a->A, b->A, str));
2056: PetscCall(MatCopy(a->B, b->B, str));
2057: }
2058: PetscCall(PetscObjectStateIncrease((PetscObject)B));
2059: PetscFunctionReturn(PETSC_SUCCESS);
2060: }
2062: /*
2063: Computes the number of nonzeros per row needed for preallocation when X and Y
2064: have different nonzero structure.
2065: */
2066: 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)
2067: {
2068: PetscInt i, j, k, nzx, nzy;
2070: PetscFunctionBegin;
2071: /* Set the number of nonzeros in the new matrix */
2072: for (i = 0; i < m; i++) {
2073: const PetscInt *xjj = PetscSafePointerPlusOffset(xj, xi[i]), *yjj = PetscSafePointerPlusOffset(yj, yi[i]);
2074: nzx = xi[i + 1] - xi[i];
2075: nzy = yi[i + 1] - yi[i];
2076: nnz[i] = 0;
2077: for (j = 0, k = 0; j < nzx; j++) { /* Point in X */
2078: for (; k < nzy && yltog[yjj[k]] < xltog[xjj[j]]; k++) nnz[i]++; /* Catch up to X */
2079: if (k < nzy && yltog[yjj[k]] == xltog[xjj[j]]) k++; /* Skip duplicate */
2080: nnz[i]++;
2081: }
2082: for (; k < nzy; k++) nnz[i]++;
2083: }
2084: PetscFunctionReturn(PETSC_SUCCESS);
2085: }
2087: /* This is the same as MatAXPYGetPreallocation_SeqAIJ, except that the local-to-global map is provided */
2088: static PetscErrorCode MatAXPYGetPreallocation_MPIAIJ(Mat Y, const PetscInt *yltog, Mat X, const PetscInt *xltog, PetscInt *nnz)
2089: {
2090: PetscInt m = Y->rmap->N;
2091: Mat_SeqAIJ *x = (Mat_SeqAIJ *)X->data;
2092: Mat_SeqAIJ *y = (Mat_SeqAIJ *)Y->data;
2094: PetscFunctionBegin;
2095: PetscCall(MatAXPYGetPreallocation_MPIX_private(m, x->i, x->j, xltog, y->i, y->j, yltog, nnz));
2096: PetscFunctionReturn(PETSC_SUCCESS);
2097: }
2099: static PetscErrorCode MatAXPY_MPIAIJ(Mat Y, PetscScalar a, Mat X, MatStructure str)
2100: {
2101: Mat_MPIAIJ *xx = (Mat_MPIAIJ *)X->data, *yy = (Mat_MPIAIJ *)Y->data;
2103: PetscFunctionBegin;
2104: if (str == SAME_NONZERO_PATTERN) {
2105: PetscCall(MatAXPY(yy->A, a, xx->A, str));
2106: PetscCall(MatAXPY(yy->B, a, xx->B, str));
2107: } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */
2108: PetscCall(MatAXPY_Basic(Y, a, X, str));
2109: } else {
2110: Mat B;
2111: PetscInt *nnz_d, *nnz_o;
2113: PetscCall(PetscMalloc1(yy->A->rmap->N, &nnz_d));
2114: PetscCall(PetscMalloc1(yy->B->rmap->N, &nnz_o));
2115: PetscCall(MatCreate(PetscObjectComm((PetscObject)Y), &B));
2116: PetscCall(PetscObjectSetName((PetscObject)B, ((PetscObject)Y)->name));
2117: PetscCall(MatSetLayouts(B, Y->rmap, Y->cmap));
2118: PetscCall(MatSetType(B, ((PetscObject)Y)->type_name));
2119: PetscCall(MatAXPYGetPreallocation_SeqAIJ(yy->A, xx->A, nnz_d));
2120: PetscCall(MatAXPYGetPreallocation_MPIAIJ(yy->B, yy->garray, xx->B, xx->garray, nnz_o));
2121: PetscCall(MatMPIAIJSetPreallocation(B, 0, nnz_d, 0, nnz_o));
2122: PetscCall(MatAXPY_BasicWithPreallocation(B, Y, a, X, str));
2123: PetscCall(MatHeaderMerge(Y, &B));
2124: PetscCall(PetscFree(nnz_d));
2125: PetscCall(PetscFree(nnz_o));
2126: }
2127: PetscFunctionReturn(PETSC_SUCCESS);
2128: }
2130: PETSC_INTERN PetscErrorCode MatConjugate_SeqAIJ(Mat);
2132: static PetscErrorCode MatConjugate_MPIAIJ(Mat mat)
2133: {
2134: PetscFunctionBegin;
2135: if (PetscDefined(USE_COMPLEX)) {
2136: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2138: PetscCall(MatConjugate_SeqAIJ(aij->A));
2139: PetscCall(MatConjugate_SeqAIJ(aij->B));
2140: }
2141: PetscFunctionReturn(PETSC_SUCCESS);
2142: }
2144: static PetscErrorCode MatRealPart_MPIAIJ(Mat A)
2145: {
2146: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2148: PetscFunctionBegin;
2149: PetscCall(MatRealPart(a->A));
2150: PetscCall(MatRealPart(a->B));
2151: PetscFunctionReturn(PETSC_SUCCESS);
2152: }
2154: static PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A)
2155: {
2156: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2158: PetscFunctionBegin;
2159: PetscCall(MatImaginaryPart(a->A));
2160: PetscCall(MatImaginaryPart(a->B));
2161: PetscFunctionReturn(PETSC_SUCCESS);
2162: }
2164: static PetscErrorCode MatGetRowMaxAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2165: {
2166: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2167: PetscInt i, *idxb = NULL, m = A->rmap->n;
2168: PetscScalar *va, *vv;
2169: Vec vB, vA;
2170: const PetscScalar *vb;
2172: PetscFunctionBegin;
2173: PetscCall(MatCreateVecs(a->A, NULL, &vA));
2174: PetscCall(MatGetRowMaxAbs(a->A, vA, idx));
2176: PetscCall(VecGetArrayWrite(vA, &va));
2177: if (idx) {
2178: for (i = 0; i < m; i++) {
2179: if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart;
2180: }
2181: }
2183: PetscCall(MatCreateVecs(a->B, NULL, &vB));
2184: PetscCall(PetscMalloc1(m, &idxb));
2185: PetscCall(MatGetRowMaxAbs(a->B, vB, idxb));
2187: PetscCall(VecGetArrayWrite(v, &vv));
2188: PetscCall(VecGetArrayRead(vB, &vb));
2189: for (i = 0; i < m; i++) {
2190: if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) {
2191: vv[i] = vb[i];
2192: if (idx) idx[i] = a->garray[idxb[i]];
2193: } else {
2194: vv[i] = va[i];
2195: if (idx && PetscAbsScalar(va[i]) == PetscAbsScalar(vb[i]) && idxb[i] != -1 && idx[i] > a->garray[idxb[i]]) idx[i] = a->garray[idxb[i]];
2196: }
2197: }
2198: PetscCall(VecRestoreArrayWrite(vA, &vv));
2199: PetscCall(VecRestoreArrayWrite(vA, &va));
2200: PetscCall(VecRestoreArrayRead(vB, &vb));
2201: PetscCall(PetscFree(idxb));
2202: PetscCall(VecDestroy(&vA));
2203: PetscCall(VecDestroy(&vB));
2204: PetscFunctionReturn(PETSC_SUCCESS);
2205: }
2207: static PetscErrorCode MatGetRowSumAbs_MPIAIJ(Mat A, Vec v)
2208: {
2209: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2210: Vec vB, vA;
2212: PetscFunctionBegin;
2213: PetscCall(MatCreateVecs(a->A, NULL, &vA));
2214: PetscCall(MatGetRowSumAbs(a->A, vA));
2215: PetscCall(MatCreateVecs(a->B, NULL, &vB));
2216: PetscCall(MatGetRowSumAbs(a->B, vB));
2217: PetscCall(VecAXPY(vA, 1.0, vB));
2218: PetscCall(VecDestroy(&vB));
2219: PetscCall(VecCopy(vA, v));
2220: PetscCall(VecDestroy(&vA));
2221: PetscFunctionReturn(PETSC_SUCCESS);
2222: }
2224: static PetscErrorCode MatGetRowMinAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2225: {
2226: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
2227: PetscInt m = A->rmap->n, n = A->cmap->n;
2228: PetscInt cstart = A->cmap->rstart, cend = A->cmap->rend;
2229: PetscInt *cmap = mat->garray;
2230: PetscInt *diagIdx, *offdiagIdx;
2231: Vec diagV, offdiagV;
2232: PetscScalar *a, *diagA, *offdiagA;
2233: const PetscScalar *ba, *bav;
2234: PetscInt r, j, col, ncols, *bi, *bj;
2235: Mat B = mat->B;
2236: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2238: PetscFunctionBegin;
2239: /* When a process holds entire A and other processes have no entry */
2240: if (A->cmap->N == n) {
2241: PetscCall(VecGetArrayWrite(v, &diagA));
2242: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2243: PetscCall(MatGetRowMinAbs(mat->A, diagV, idx));
2244: PetscCall(VecDestroy(&diagV));
2245: PetscCall(VecRestoreArrayWrite(v, &diagA));
2246: PetscFunctionReturn(PETSC_SUCCESS);
2247: } else if (n == 0) {
2248: if (m) {
2249: PetscCall(VecGetArrayWrite(v, &a));
2250: for (r = 0; r < m; r++) {
2251: a[r] = 0.0;
2252: if (idx) idx[r] = -1;
2253: }
2254: PetscCall(VecRestoreArrayWrite(v, &a));
2255: }
2256: PetscFunctionReturn(PETSC_SUCCESS);
2257: }
2259: PetscCall(PetscMalloc2(m, &diagIdx, m, &offdiagIdx));
2260: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2261: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2262: PetscCall(MatGetRowMinAbs(mat->A, diagV, diagIdx));
2264: /* Get offdiagIdx[] for implicit 0.0 */
2265: PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2266: ba = bav;
2267: bi = b->i;
2268: bj = b->j;
2269: PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2270: for (r = 0; r < m; r++) {
2271: ncols = bi[r + 1] - bi[r];
2272: if (ncols == A->cmap->N - n) { /* Brow is dense */
2273: offdiagA[r] = *ba;
2274: offdiagIdx[r] = cmap[0];
2275: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2276: offdiagA[r] = 0.0;
2278: /* Find first hole in the cmap */
2279: for (j = 0; j < ncols; j++) {
2280: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2281: if (col > j && j < cstart) {
2282: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2283: break;
2284: } else if (col > j + n && j >= cstart) {
2285: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2286: break;
2287: }
2288: }
2289: if (j == ncols && ncols < A->cmap->N - n) {
2290: /* a hole is outside compressed Bcols */
2291: if (ncols == 0) {
2292: if (cstart) {
2293: offdiagIdx[r] = 0;
2294: } else offdiagIdx[r] = cend;
2295: } else { /* ncols > 0 */
2296: offdiagIdx[r] = cmap[ncols - 1] + 1;
2297: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2298: }
2299: }
2300: }
2302: for (j = 0; j < ncols; j++) {
2303: if (PetscAbsScalar(offdiagA[r]) > PetscAbsScalar(*ba)) {
2304: offdiagA[r] = *ba;
2305: offdiagIdx[r] = cmap[*bj];
2306: }
2307: ba++;
2308: bj++;
2309: }
2310: }
2312: PetscCall(VecGetArrayWrite(v, &a));
2313: PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2314: for (r = 0; r < m; ++r) {
2315: if (PetscAbsScalar(diagA[r]) < PetscAbsScalar(offdiagA[r])) {
2316: a[r] = diagA[r];
2317: if (idx) idx[r] = cstart + diagIdx[r];
2318: } else if (PetscAbsScalar(diagA[r]) == PetscAbsScalar(offdiagA[r])) {
2319: a[r] = diagA[r];
2320: if (idx) {
2321: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2322: idx[r] = cstart + diagIdx[r];
2323: } else idx[r] = offdiagIdx[r];
2324: }
2325: } else {
2326: a[r] = offdiagA[r];
2327: if (idx) idx[r] = offdiagIdx[r];
2328: }
2329: }
2330: PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2331: PetscCall(VecRestoreArrayWrite(v, &a));
2332: PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2333: PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2334: PetscCall(VecDestroy(&diagV));
2335: PetscCall(VecDestroy(&offdiagV));
2336: PetscCall(PetscFree2(diagIdx, offdiagIdx));
2337: PetscFunctionReturn(PETSC_SUCCESS);
2338: }
2340: static PetscErrorCode MatGetRowMin_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2341: {
2342: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
2343: PetscInt m = A->rmap->n, n = A->cmap->n;
2344: PetscInt cstart = A->cmap->rstart, cend = A->cmap->rend;
2345: PetscInt *cmap = mat->garray;
2346: PetscInt *diagIdx, *offdiagIdx;
2347: Vec diagV, offdiagV;
2348: PetscScalar *a, *diagA, *offdiagA;
2349: const PetscScalar *ba, *bav;
2350: PetscInt r, j, col, ncols, *bi, *bj;
2351: Mat B = mat->B;
2352: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2354: PetscFunctionBegin;
2355: /* When a process holds entire A and other processes have no entry */
2356: if (A->cmap->N == n) {
2357: PetscCall(VecGetArrayWrite(v, &diagA));
2358: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2359: PetscCall(MatGetRowMin(mat->A, diagV, idx));
2360: PetscCall(VecDestroy(&diagV));
2361: PetscCall(VecRestoreArrayWrite(v, &diagA));
2362: PetscFunctionReturn(PETSC_SUCCESS);
2363: } else if (n == 0) {
2364: if (m) {
2365: PetscCall(VecGetArrayWrite(v, &a));
2366: for (r = 0; r < m; r++) {
2367: a[r] = PETSC_MAX_REAL;
2368: if (idx) idx[r] = -1;
2369: }
2370: PetscCall(VecRestoreArrayWrite(v, &a));
2371: }
2372: PetscFunctionReturn(PETSC_SUCCESS);
2373: }
2375: PetscCall(PetscCalloc2(m, &diagIdx, m, &offdiagIdx));
2376: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2377: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2378: PetscCall(MatGetRowMin(mat->A, diagV, diagIdx));
2380: /* Get offdiagIdx[] for implicit 0.0 */
2381: PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2382: ba = bav;
2383: bi = b->i;
2384: bj = b->j;
2385: PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2386: for (r = 0; r < m; r++) {
2387: ncols = bi[r + 1] - bi[r];
2388: if (ncols == A->cmap->N - n) { /* Brow is dense */
2389: offdiagA[r] = *ba;
2390: offdiagIdx[r] = cmap[0];
2391: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2392: offdiagA[r] = 0.0;
2394: /* Find first hole in the cmap */
2395: for (j = 0; j < ncols; j++) {
2396: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2397: if (col > j && j < cstart) {
2398: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2399: break;
2400: } else if (col > j + n && j >= cstart) {
2401: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2402: break;
2403: }
2404: }
2405: if (j == ncols && ncols < A->cmap->N - n) {
2406: /* a hole is outside compressed Bcols */
2407: if (ncols == 0) {
2408: if (cstart) {
2409: offdiagIdx[r] = 0;
2410: } else offdiagIdx[r] = cend;
2411: } else { /* ncols > 0 */
2412: offdiagIdx[r] = cmap[ncols - 1] + 1;
2413: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2414: }
2415: }
2416: }
2418: for (j = 0; j < ncols; j++) {
2419: if (PetscRealPart(offdiagA[r]) > PetscRealPart(*ba)) {
2420: offdiagA[r] = *ba;
2421: offdiagIdx[r] = cmap[*bj];
2422: }
2423: ba++;
2424: bj++;
2425: }
2426: }
2428: PetscCall(VecGetArrayWrite(v, &a));
2429: PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2430: for (r = 0; r < m; ++r) {
2431: if (PetscRealPart(diagA[r]) < PetscRealPart(offdiagA[r])) {
2432: a[r] = diagA[r];
2433: if (idx) idx[r] = cstart + diagIdx[r];
2434: } else if (PetscRealPart(diagA[r]) == PetscRealPart(offdiagA[r])) {
2435: a[r] = diagA[r];
2436: if (idx) {
2437: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2438: idx[r] = cstart + diagIdx[r];
2439: } else idx[r] = offdiagIdx[r];
2440: }
2441: } else {
2442: a[r] = offdiagA[r];
2443: if (idx) idx[r] = offdiagIdx[r];
2444: }
2445: }
2446: PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2447: PetscCall(VecRestoreArrayWrite(v, &a));
2448: PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2449: PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2450: PetscCall(VecDestroy(&diagV));
2451: PetscCall(VecDestroy(&offdiagV));
2452: PetscCall(PetscFree2(diagIdx, offdiagIdx));
2453: PetscFunctionReturn(PETSC_SUCCESS);
2454: }
2456: static PetscErrorCode MatGetRowMax_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2457: {
2458: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
2459: PetscInt m = A->rmap->n, n = A->cmap->n;
2460: PetscInt cstart = A->cmap->rstart, cend = A->cmap->rend;
2461: PetscInt *cmap = mat->garray;
2462: PetscInt *diagIdx, *offdiagIdx;
2463: Vec diagV, offdiagV;
2464: PetscScalar *a, *diagA, *offdiagA;
2465: const PetscScalar *ba, *bav;
2466: PetscInt r, j, col, ncols, *bi, *bj;
2467: Mat B = mat->B;
2468: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2470: PetscFunctionBegin;
2471: /* When a process holds entire A and other processes have no entry */
2472: if (A->cmap->N == n) {
2473: PetscCall(VecGetArrayWrite(v, &diagA));
2474: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2475: PetscCall(MatGetRowMax(mat->A, diagV, idx));
2476: PetscCall(VecDestroy(&diagV));
2477: PetscCall(VecRestoreArrayWrite(v, &diagA));
2478: PetscFunctionReturn(PETSC_SUCCESS);
2479: } else if (n == 0) {
2480: if (m) {
2481: PetscCall(VecGetArrayWrite(v, &a));
2482: for (r = 0; r < m; r++) {
2483: a[r] = PETSC_MIN_REAL;
2484: if (idx) idx[r] = -1;
2485: }
2486: PetscCall(VecRestoreArrayWrite(v, &a));
2487: }
2488: PetscFunctionReturn(PETSC_SUCCESS);
2489: }
2491: PetscCall(PetscMalloc2(m, &diagIdx, m, &offdiagIdx));
2492: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2493: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2494: PetscCall(MatGetRowMax(mat->A, diagV, diagIdx));
2496: /* Get offdiagIdx[] for implicit 0.0 */
2497: PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2498: ba = bav;
2499: bi = b->i;
2500: bj = b->j;
2501: PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2502: for (r = 0; r < m; r++) {
2503: ncols = bi[r + 1] - bi[r];
2504: if (ncols == A->cmap->N - n) { /* Brow is dense */
2505: offdiagA[r] = *ba;
2506: offdiagIdx[r] = cmap[0];
2507: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2508: offdiagA[r] = 0.0;
2510: /* Find first hole in the cmap */
2511: for (j = 0; j < ncols; j++) {
2512: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2513: if (col > j && j < cstart) {
2514: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2515: break;
2516: } else if (col > j + n && j >= cstart) {
2517: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2518: break;
2519: }
2520: }
2521: if (j == ncols && ncols < A->cmap->N - n) {
2522: /* a hole is outside compressed Bcols */
2523: if (ncols == 0) {
2524: if (cstart) {
2525: offdiagIdx[r] = 0;
2526: } else offdiagIdx[r] = cend;
2527: } else { /* ncols > 0 */
2528: offdiagIdx[r] = cmap[ncols - 1] + 1;
2529: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2530: }
2531: }
2532: }
2534: for (j = 0; j < ncols; j++) {
2535: if (PetscRealPart(offdiagA[r]) < PetscRealPart(*ba)) {
2536: offdiagA[r] = *ba;
2537: offdiagIdx[r] = cmap[*bj];
2538: }
2539: ba++;
2540: bj++;
2541: }
2542: }
2544: PetscCall(VecGetArrayWrite(v, &a));
2545: PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2546: for (r = 0; r < m; ++r) {
2547: if (PetscRealPart(diagA[r]) > PetscRealPart(offdiagA[r])) {
2548: a[r] = diagA[r];
2549: if (idx) idx[r] = cstart + diagIdx[r];
2550: } else if (PetscRealPart(diagA[r]) == PetscRealPart(offdiagA[r])) {
2551: a[r] = diagA[r];
2552: if (idx) {
2553: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2554: idx[r] = cstart + diagIdx[r];
2555: } else idx[r] = offdiagIdx[r];
2556: }
2557: } else {
2558: a[r] = offdiagA[r];
2559: if (idx) idx[r] = offdiagIdx[r];
2560: }
2561: }
2562: PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2563: PetscCall(VecRestoreArrayWrite(v, &a));
2564: PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2565: PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2566: PetscCall(VecDestroy(&diagV));
2567: PetscCall(VecDestroy(&offdiagV));
2568: PetscCall(PetscFree2(diagIdx, offdiagIdx));
2569: PetscFunctionReturn(PETSC_SUCCESS);
2570: }
2572: PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat mat, Mat *newmat)
2573: {
2574: Mat *dummy;
2576: PetscFunctionBegin;
2577: PetscCall(MatCreateSubMatrix_MPIAIJ_All(mat, MAT_DO_NOT_GET_VALUES, MAT_INITIAL_MATRIX, &dummy));
2578: *newmat = *dummy;
2579: PetscCall(PetscFree(dummy));
2580: PetscFunctionReturn(PETSC_SUCCESS);
2581: }
2583: static PetscErrorCode MatInvertBlockDiagonal_MPIAIJ(Mat A, const PetscScalar **values)
2584: {
2585: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2587: PetscFunctionBegin;
2588: PetscCall(MatInvertBlockDiagonal(a->A, values));
2589: A->factorerrortype = a->A->factorerrortype;
2590: PetscFunctionReturn(PETSC_SUCCESS);
2591: }
2593: static PetscErrorCode MatSetRandom_MPIAIJ(Mat x, PetscRandom rctx)
2594: {
2595: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)x->data;
2597: PetscFunctionBegin;
2598: PetscCheck(x->assembled || x->preallocated, PetscObjectComm((PetscObject)x), PETSC_ERR_ARG_WRONGSTATE, "MatSetRandom on an unassembled and unpreallocated MATMPIAIJ is not allowed");
2599: PetscCall(MatSetRandom(aij->A, rctx));
2600: if (x->assembled) {
2601: PetscCall(MatSetRandom(aij->B, rctx));
2602: } else {
2603: PetscCall(MatSetRandomSkipColumnRange_SeqAIJ_Private(aij->B, x->cmap->rstart, x->cmap->rend, rctx));
2604: }
2605: PetscCall(MatAssemblyBegin(x, MAT_FINAL_ASSEMBLY));
2606: PetscCall(MatAssemblyEnd(x, MAT_FINAL_ASSEMBLY));
2607: PetscFunctionReturn(PETSC_SUCCESS);
2608: }
2610: static PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ(Mat A, PetscBool sc)
2611: {
2612: PetscFunctionBegin;
2613: if (sc) A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ_Scalable;
2614: else A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ;
2615: PetscFunctionReturn(PETSC_SUCCESS);
2616: }
2618: /*@
2619: MatMPIAIJGetNumberNonzeros - gets the number of nonzeros in the matrix on this MPI rank
2621: Not Collective
2623: Input Parameter:
2624: . A - the matrix
2626: Output Parameter:
2627: . nz - the number of nonzeros
2629: Level: advanced
2631: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`
2632: @*/
2633: PetscErrorCode MatMPIAIJGetNumberNonzeros(Mat A, PetscCount *nz)
2634: {
2635: Mat_MPIAIJ *maij = (Mat_MPIAIJ *)A->data;
2636: Mat_SeqAIJ *aaij = (Mat_SeqAIJ *)maij->A->data, *baij = (Mat_SeqAIJ *)maij->B->data;
2637: PetscBool isaij;
2639: PetscFunctionBegin;
2640: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATMPIAIJ, &isaij));
2641: PetscCheck(isaij, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Not for type %s", ((PetscObject)A)->type_name);
2642: *nz = aaij->i[A->rmap->n] + baij->i[A->rmap->n];
2643: PetscFunctionReturn(PETSC_SUCCESS);
2644: }
2646: /*@
2647: MatMPIAIJSetUseScalableIncreaseOverlap - Determine if the matrix uses a scalable algorithm to compute the overlap
2649: Collective
2651: Input Parameters:
2652: + A - the matrix
2653: - sc - `PETSC_TRUE` indicates use the scalable algorithm (default is not to use the scalable algorithm)
2655: Level: advanced
2657: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`
2658: @*/
2659: PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap(Mat A, PetscBool sc)
2660: {
2661: PetscFunctionBegin;
2662: PetscTryMethod(A, "MatMPIAIJSetUseScalableIncreaseOverlap_C", (Mat, PetscBool), (A, sc));
2663: PetscFunctionReturn(PETSC_SUCCESS);
2664: }
2666: PetscErrorCode MatSetFromOptions_MPIAIJ(Mat A, PetscOptionItems *PetscOptionsObject)
2667: {
2668: PetscBool sc = PETSC_FALSE, flg;
2670: PetscFunctionBegin;
2671: PetscOptionsHeadBegin(PetscOptionsObject, "MPIAIJ options");
2672: if (A->ops->increaseoverlap == MatIncreaseOverlap_MPIAIJ_Scalable) sc = PETSC_TRUE;
2673: PetscCall(PetscOptionsBool("-mat_increase_overlap_scalable", "Use a scalable algorithm to compute the overlap", "MatIncreaseOverlap", sc, &sc, &flg));
2674: if (flg) PetscCall(MatMPIAIJSetUseScalableIncreaseOverlap(A, sc));
2675: PetscOptionsHeadEnd();
2676: PetscFunctionReturn(PETSC_SUCCESS);
2677: }
2679: static PetscErrorCode MatShift_MPIAIJ(Mat Y, PetscScalar a)
2680: {
2681: Mat_MPIAIJ *maij = (Mat_MPIAIJ *)Y->data;
2682: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)maij->A->data;
2684: PetscFunctionBegin;
2685: if (!Y->preallocated) {
2686: PetscCall(MatMPIAIJSetPreallocation(Y, 1, NULL, 0, NULL));
2687: } else if (!aij->nz) { /* It does not matter if diagonals of Y only partially lie in maij->A. We just need an estimated preallocation. */
2688: PetscInt nonew = aij->nonew;
2689: PetscCall(MatSeqAIJSetPreallocation(maij->A, 1, NULL));
2690: aij->nonew = nonew;
2691: }
2692: PetscCall(MatShift_Basic(Y, a));
2693: PetscFunctionReturn(PETSC_SUCCESS);
2694: }
2696: static PetscErrorCode MatMissingDiagonal_MPIAIJ(Mat A, PetscBool *missing, PetscInt *d)
2697: {
2698: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2700: PetscFunctionBegin;
2701: PetscCheck(A->rmap->n == A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_SUP, "Only works for square matrices");
2702: PetscCall(MatMissingDiagonal(a->A, missing, d));
2703: if (d) {
2704: PetscInt rstart;
2705: PetscCall(MatGetOwnershipRange(A, &rstart, NULL));
2706: *d += rstart;
2707: }
2708: PetscFunctionReturn(PETSC_SUCCESS);
2709: }
2711: static PetscErrorCode MatInvertVariableBlockDiagonal_MPIAIJ(Mat A, PetscInt nblocks, const PetscInt *bsizes, PetscScalar *diag)
2712: {
2713: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2715: PetscFunctionBegin;
2716: PetscCall(MatInvertVariableBlockDiagonal(a->A, nblocks, bsizes, diag));
2717: PetscFunctionReturn(PETSC_SUCCESS);
2718: }
2720: static PetscErrorCode MatEliminateZeros_MPIAIJ(Mat A, PetscBool keep)
2721: {
2722: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2724: PetscFunctionBegin;
2725: PetscCall(MatEliminateZeros_SeqAIJ(a->A, keep)); // possibly keep zero diagonal coefficients
2726: PetscCall(MatEliminateZeros_SeqAIJ(a->B, PETSC_FALSE)); // never keep zero diagonal coefficients
2727: PetscFunctionReturn(PETSC_SUCCESS);
2728: }
2730: static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ,
2731: MatGetRow_MPIAIJ,
2732: MatRestoreRow_MPIAIJ,
2733: MatMult_MPIAIJ,
2734: /* 4*/ MatMultAdd_MPIAIJ,
2735: MatMultTranspose_MPIAIJ,
2736: MatMultTransposeAdd_MPIAIJ,
2737: NULL,
2738: NULL,
2739: NULL,
2740: /*10*/ NULL,
2741: NULL,
2742: NULL,
2743: MatSOR_MPIAIJ,
2744: MatTranspose_MPIAIJ,
2745: /*15*/ MatGetInfo_MPIAIJ,
2746: MatEqual_MPIAIJ,
2747: MatGetDiagonal_MPIAIJ,
2748: MatDiagonalScale_MPIAIJ,
2749: MatNorm_MPIAIJ,
2750: /*20*/ MatAssemblyBegin_MPIAIJ,
2751: MatAssemblyEnd_MPIAIJ,
2752: MatSetOption_MPIAIJ,
2753: MatZeroEntries_MPIAIJ,
2754: /*24*/ MatZeroRows_MPIAIJ,
2755: NULL,
2756: NULL,
2757: NULL,
2758: NULL,
2759: /*29*/ MatSetUp_MPI_Hash,
2760: NULL,
2761: NULL,
2762: MatGetDiagonalBlock_MPIAIJ,
2763: NULL,
2764: /*34*/ MatDuplicate_MPIAIJ,
2765: NULL,
2766: NULL,
2767: NULL,
2768: NULL,
2769: /*39*/ MatAXPY_MPIAIJ,
2770: MatCreateSubMatrices_MPIAIJ,
2771: MatIncreaseOverlap_MPIAIJ,
2772: MatGetValues_MPIAIJ,
2773: MatCopy_MPIAIJ,
2774: /*44*/ MatGetRowMax_MPIAIJ,
2775: MatScale_MPIAIJ,
2776: MatShift_MPIAIJ,
2777: MatDiagonalSet_MPIAIJ,
2778: MatZeroRowsColumns_MPIAIJ,
2779: /*49*/ MatSetRandom_MPIAIJ,
2780: MatGetRowIJ_MPIAIJ,
2781: MatRestoreRowIJ_MPIAIJ,
2782: NULL,
2783: NULL,
2784: /*54*/ MatFDColoringCreate_MPIXAIJ,
2785: NULL,
2786: MatSetUnfactored_MPIAIJ,
2787: MatPermute_MPIAIJ,
2788: NULL,
2789: /*59*/ MatCreateSubMatrix_MPIAIJ,
2790: MatDestroy_MPIAIJ,
2791: MatView_MPIAIJ,
2792: NULL,
2793: NULL,
2794: /*64*/ NULL,
2795: MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ,
2796: NULL,
2797: NULL,
2798: NULL,
2799: /*69*/ MatGetRowMaxAbs_MPIAIJ,
2800: MatGetRowMinAbs_MPIAIJ,
2801: NULL,
2802: NULL,
2803: NULL,
2804: NULL,
2805: /*75*/ MatFDColoringApply_AIJ,
2806: MatSetFromOptions_MPIAIJ,
2807: NULL,
2808: NULL,
2809: MatFindZeroDiagonals_MPIAIJ,
2810: /*80*/ NULL,
2811: NULL,
2812: NULL,
2813: /*83*/ MatLoad_MPIAIJ,
2814: NULL,
2815: NULL,
2816: NULL,
2817: NULL,
2818: NULL,
2819: /*89*/ NULL,
2820: NULL,
2821: MatMatMultNumeric_MPIAIJ_MPIAIJ,
2822: NULL,
2823: NULL,
2824: /*94*/ MatPtAPNumeric_MPIAIJ_MPIAIJ,
2825: NULL,
2826: NULL,
2827: NULL,
2828: MatBindToCPU_MPIAIJ,
2829: /*99*/ MatProductSetFromOptions_MPIAIJ,
2830: NULL,
2831: NULL,
2832: MatConjugate_MPIAIJ,
2833: NULL,
2834: /*104*/ MatSetValuesRow_MPIAIJ,
2835: MatRealPart_MPIAIJ,
2836: MatImaginaryPart_MPIAIJ,
2837: NULL,
2838: NULL,
2839: /*109*/ NULL,
2840: NULL,
2841: MatGetRowMin_MPIAIJ,
2842: NULL,
2843: MatMissingDiagonal_MPIAIJ,
2844: /*114*/ MatGetSeqNonzeroStructure_MPIAIJ,
2845: NULL,
2846: MatGetGhosts_MPIAIJ,
2847: NULL,
2848: NULL,
2849: /*119*/ MatMultDiagonalBlock_MPIAIJ,
2850: NULL,
2851: NULL,
2852: NULL,
2853: MatGetMultiProcBlock_MPIAIJ,
2854: /*124*/ MatFindNonzeroRows_MPIAIJ,
2855: MatGetColumnReductions_MPIAIJ,
2856: MatInvertBlockDiagonal_MPIAIJ,
2857: MatInvertVariableBlockDiagonal_MPIAIJ,
2858: MatCreateSubMatricesMPI_MPIAIJ,
2859: /*129*/ NULL,
2860: NULL,
2861: NULL,
2862: MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ,
2863: NULL,
2864: /*134*/ NULL,
2865: NULL,
2866: NULL,
2867: NULL,
2868: NULL,
2869: /*139*/ MatSetBlockSizes_MPIAIJ,
2870: NULL,
2871: NULL,
2872: MatFDColoringSetUp_MPIXAIJ,
2873: MatFindOffBlockDiagonalEntries_MPIAIJ,
2874: MatCreateMPIMatConcatenateSeqMat_MPIAIJ,
2875: /*145*/ NULL,
2876: NULL,
2877: NULL,
2878: MatCreateGraph_Simple_AIJ,
2879: NULL,
2880: /*150*/ NULL,
2881: MatEliminateZeros_MPIAIJ,
2882: MatGetRowSumAbs_MPIAIJ,
2883: NULL,
2884: NULL,
2885: /*155*/ NULL,
2886: MatCopyHashToXAIJ_MPI_Hash};
2888: static PetscErrorCode MatStoreValues_MPIAIJ(Mat mat)
2889: {
2890: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2892: PetscFunctionBegin;
2893: PetscCall(MatStoreValues(aij->A));
2894: PetscCall(MatStoreValues(aij->B));
2895: PetscFunctionReturn(PETSC_SUCCESS);
2896: }
2898: static PetscErrorCode MatRetrieveValues_MPIAIJ(Mat mat)
2899: {
2900: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2902: PetscFunctionBegin;
2903: PetscCall(MatRetrieveValues(aij->A));
2904: PetscCall(MatRetrieveValues(aij->B));
2905: PetscFunctionReturn(PETSC_SUCCESS);
2906: }
2908: PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat B, PetscInt d_nz, const PetscInt d_nnz[], PetscInt o_nz, const PetscInt o_nnz[])
2909: {
2910: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2911: PetscMPIInt size;
2913: PetscFunctionBegin;
2914: if (B->hash_active) {
2915: B->ops[0] = b->cops;
2916: B->hash_active = PETSC_FALSE;
2917: }
2918: PetscCall(PetscLayoutSetUp(B->rmap));
2919: PetscCall(PetscLayoutSetUp(B->cmap));
2921: #if defined(PETSC_USE_CTABLE)
2922: PetscCall(PetscHMapIDestroy(&b->colmap));
2923: #else
2924: PetscCall(PetscFree(b->colmap));
2925: #endif
2926: PetscCall(PetscFree(b->garray));
2927: PetscCall(VecDestroy(&b->lvec));
2928: PetscCall(VecScatterDestroy(&b->Mvctx));
2930: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
2932: MatSeqXAIJGetOptions_Private(b->B);
2933: PetscCall(MatDestroy(&b->B));
2934: PetscCall(MatCreate(PETSC_COMM_SELF, &b->B));
2935: PetscCall(MatSetSizes(b->B, B->rmap->n, size > 1 ? B->cmap->N : 0, B->rmap->n, size > 1 ? B->cmap->N : 0));
2936: PetscCall(MatSetBlockSizesFromMats(b->B, B, B));
2937: PetscCall(MatSetType(b->B, MATSEQAIJ));
2938: MatSeqXAIJRestoreOptions_Private(b->B);
2940: MatSeqXAIJGetOptions_Private(b->A);
2941: PetscCall(MatDestroy(&b->A));
2942: PetscCall(MatCreate(PETSC_COMM_SELF, &b->A));
2943: PetscCall(MatSetSizes(b->A, B->rmap->n, B->cmap->n, B->rmap->n, B->cmap->n));
2944: PetscCall(MatSetBlockSizesFromMats(b->A, B, B));
2945: PetscCall(MatSetType(b->A, MATSEQAIJ));
2946: MatSeqXAIJRestoreOptions_Private(b->A);
2948: PetscCall(MatSeqAIJSetPreallocation(b->A, d_nz, d_nnz));
2949: PetscCall(MatSeqAIJSetPreallocation(b->B, o_nz, o_nnz));
2950: B->preallocated = PETSC_TRUE;
2951: B->was_assembled = PETSC_FALSE;
2952: B->assembled = PETSC_FALSE;
2953: PetscFunctionReturn(PETSC_SUCCESS);
2954: }
2956: static PetscErrorCode MatResetPreallocation_MPIAIJ(Mat B)
2957: {
2958: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2960: PetscFunctionBegin;
2962: PetscCall(PetscLayoutSetUp(B->rmap));
2963: PetscCall(PetscLayoutSetUp(B->cmap));
2964: if (B->assembled || B->was_assembled) PetscCall(MatDisAssemble_MPIAIJ(B, PETSC_TRUE));
2965: else {
2966: #if defined(PETSC_USE_CTABLE)
2967: PetscCall(PetscHMapIDestroy(&b->colmap));
2968: #else
2969: PetscCall(PetscFree(b->colmap));
2970: #endif
2971: PetscCall(PetscFree(b->garray));
2972: PetscCall(VecDestroy(&b->lvec));
2973: }
2974: PetscCall(VecScatterDestroy(&b->Mvctx));
2976: PetscCall(MatResetPreallocation(b->A));
2977: PetscCall(MatResetPreallocation(b->B));
2978: B->preallocated = PETSC_TRUE;
2979: B->was_assembled = PETSC_FALSE;
2980: B->assembled = PETSC_FALSE;
2981: PetscFunctionReturn(PETSC_SUCCESS);
2982: }
2984: PetscErrorCode MatDuplicate_MPIAIJ(Mat matin, MatDuplicateOption cpvalues, Mat *newmat)
2985: {
2986: Mat mat;
2987: Mat_MPIAIJ *a, *oldmat = (Mat_MPIAIJ *)matin->data;
2989: PetscFunctionBegin;
2990: *newmat = NULL;
2991: PetscCall(MatCreate(PetscObjectComm((PetscObject)matin), &mat));
2992: PetscCall(MatSetSizes(mat, matin->rmap->n, matin->cmap->n, matin->rmap->N, matin->cmap->N));
2993: PetscCall(MatSetBlockSizesFromMats(mat, matin, matin));
2994: PetscCall(MatSetType(mat, ((PetscObject)matin)->type_name));
2995: a = (Mat_MPIAIJ *)mat->data;
2997: mat->factortype = matin->factortype;
2998: mat->assembled = matin->assembled;
2999: mat->insertmode = NOT_SET_VALUES;
3001: a->size = oldmat->size;
3002: a->rank = oldmat->rank;
3003: a->donotstash = oldmat->donotstash;
3004: a->roworiented = oldmat->roworiented;
3005: a->rowindices = NULL;
3006: a->rowvalues = NULL;
3007: a->getrowactive = PETSC_FALSE;
3009: PetscCall(PetscLayoutReference(matin->rmap, &mat->rmap));
3010: PetscCall(PetscLayoutReference(matin->cmap, &mat->cmap));
3011: if (matin->hash_active) {
3012: PetscCall(MatSetUp(mat));
3013: } else {
3014: mat->preallocated = matin->preallocated;
3015: if (oldmat->colmap) {
3016: #if defined(PETSC_USE_CTABLE)
3017: PetscCall(PetscHMapIDuplicate(oldmat->colmap, &a->colmap));
3018: #else
3019: PetscCall(PetscMalloc1(mat->cmap->N, &a->colmap));
3020: PetscCall(PetscArraycpy(a->colmap, oldmat->colmap, mat->cmap->N));
3021: #endif
3022: } else a->colmap = NULL;
3023: if (oldmat->garray) {
3024: PetscInt len;
3025: len = oldmat->B->cmap->n;
3026: PetscCall(PetscMalloc1(len + 1, &a->garray));
3027: if (len) PetscCall(PetscArraycpy(a->garray, oldmat->garray, len));
3028: } else a->garray = NULL;
3030: /* It may happen MatDuplicate is called with a non-assembled matrix
3031: In fact, MatDuplicate only requires the matrix to be preallocated
3032: This may happen inside a DMCreateMatrix_Shell */
3033: if (oldmat->lvec) PetscCall(VecDuplicate(oldmat->lvec, &a->lvec));
3034: if (oldmat->Mvctx) {
3035: a->Mvctx = oldmat->Mvctx;
3036: PetscCall(PetscObjectReference((PetscObject)oldmat->Mvctx));
3037: }
3038: PetscCall(MatDuplicate(oldmat->A, cpvalues, &a->A));
3039: PetscCall(MatDuplicate(oldmat->B, cpvalues, &a->B));
3040: }
3041: PetscCall(PetscFunctionListDuplicate(((PetscObject)matin)->qlist, &((PetscObject)mat)->qlist));
3042: *newmat = mat;
3043: PetscFunctionReturn(PETSC_SUCCESS);
3044: }
3046: PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer)
3047: {
3048: PetscBool isbinary, ishdf5;
3050: PetscFunctionBegin;
3053: /* force binary viewer to load .info file if it has not yet done so */
3054: PetscCall(PetscViewerSetUp(viewer));
3055: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
3056: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERHDF5, &ishdf5));
3057: if (isbinary) {
3058: PetscCall(MatLoad_MPIAIJ_Binary(newMat, viewer));
3059: } else if (ishdf5) {
3060: #if defined(PETSC_HAVE_HDF5)
3061: PetscCall(MatLoad_AIJ_HDF5(newMat, viewer));
3062: #else
3063: SETERRQ(PetscObjectComm((PetscObject)newMat), PETSC_ERR_SUP, "HDF5 not supported in this build.\nPlease reconfigure using --download-hdf5");
3064: #endif
3065: } else {
3066: 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);
3067: }
3068: PetscFunctionReturn(PETSC_SUCCESS);
3069: }
3071: PetscErrorCode MatLoad_MPIAIJ_Binary(Mat mat, PetscViewer viewer)
3072: {
3073: PetscInt header[4], M, N, m, nz, rows, cols, sum, i;
3074: PetscInt *rowidxs, *colidxs;
3075: PetscScalar *matvals;
3077: PetscFunctionBegin;
3078: PetscCall(PetscViewerSetUp(viewer));
3080: /* read in matrix header */
3081: PetscCall(PetscViewerBinaryRead(viewer, header, 4, NULL, PETSC_INT));
3082: PetscCheck(header[0] == MAT_FILE_CLASSID, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Not a matrix object in file");
3083: M = header[1];
3084: N = header[2];
3085: nz = header[3];
3086: PetscCheck(M >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix row size (%" PetscInt_FMT ") in file is negative", M);
3087: PetscCheck(N >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix column size (%" PetscInt_FMT ") in file is negative", N);
3088: PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Matrix stored in special format on disk, cannot load as MPIAIJ");
3090: /* set block sizes from the viewer's .info file */
3091: PetscCall(MatLoad_Binary_BlockSizes(mat, viewer));
3092: /* set global sizes if not set already */
3093: if (mat->rmap->N < 0) mat->rmap->N = M;
3094: if (mat->cmap->N < 0) mat->cmap->N = N;
3095: PetscCall(PetscLayoutSetUp(mat->rmap));
3096: PetscCall(PetscLayoutSetUp(mat->cmap));
3098: /* check if the matrix sizes are correct */
3099: PetscCall(MatGetSize(mat, &rows, &cols));
3100: 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);
3102: /* read in row lengths and build row indices */
3103: PetscCall(MatGetLocalSize(mat, &m, NULL));
3104: PetscCall(PetscMalloc1(m + 1, &rowidxs));
3105: PetscCall(PetscViewerBinaryReadAll(viewer, rowidxs + 1, m, PETSC_DECIDE, M, PETSC_INT));
3106: rowidxs[0] = 0;
3107: for (i = 0; i < m; i++) rowidxs[i + 1] += rowidxs[i];
3108: if (nz != PETSC_INT_MAX) {
3109: PetscCallMPI(MPIU_Allreduce(&rowidxs[m], &sum, 1, MPIU_INT, MPI_SUM, PetscObjectComm((PetscObject)viewer)));
3110: 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);
3111: }
3113: /* read in column indices and matrix values */
3114: PetscCall(PetscMalloc2(rowidxs[m], &colidxs, rowidxs[m], &matvals));
3115: PetscCall(PetscViewerBinaryReadAll(viewer, colidxs, rowidxs[m], PETSC_DETERMINE, PETSC_DETERMINE, PETSC_INT));
3116: PetscCall(PetscViewerBinaryReadAll(viewer, matvals, rowidxs[m], PETSC_DETERMINE, PETSC_DETERMINE, PETSC_SCALAR));
3117: /* store matrix indices and values */
3118: PetscCall(MatMPIAIJSetPreallocationCSR(mat, rowidxs, colidxs, matvals));
3119: PetscCall(PetscFree(rowidxs));
3120: PetscCall(PetscFree2(colidxs, matvals));
3121: PetscFunctionReturn(PETSC_SUCCESS);
3122: }
3124: /* Not scalable because of ISAllGather() unless getting all columns. */
3125: static PetscErrorCode ISGetSeqIS_Private(Mat mat, IS iscol, IS *isseq)
3126: {
3127: IS iscol_local;
3128: PetscBool isstride;
3129: PetscMPIInt lisstride = 0, gisstride;
3131: PetscFunctionBegin;
3132: /* check if we are grabbing all columns*/
3133: PetscCall(PetscObjectTypeCompare((PetscObject)iscol, ISSTRIDE, &isstride));
3135: if (isstride) {
3136: PetscInt start, len, mstart, mlen;
3137: PetscCall(ISStrideGetInfo(iscol, &start, NULL));
3138: PetscCall(ISGetLocalSize(iscol, &len));
3139: PetscCall(MatGetOwnershipRangeColumn(mat, &mstart, &mlen));
3140: if (mstart == start && mlen - mstart == len) lisstride = 1;
3141: }
3143: PetscCallMPI(MPIU_Allreduce(&lisstride, &gisstride, 1, MPI_INT, MPI_MIN, PetscObjectComm((PetscObject)mat)));
3144: if (gisstride) {
3145: PetscInt N;
3146: PetscCall(MatGetSize(mat, NULL, &N));
3147: PetscCall(ISCreateStride(PETSC_COMM_SELF, N, 0, 1, &iscol_local));
3148: PetscCall(ISSetIdentity(iscol_local));
3149: PetscCall(PetscInfo(mat, "Optimizing for obtaining all columns of the matrix; skipping ISAllGather()\n"));
3150: } else {
3151: PetscInt cbs;
3152: PetscCall(ISGetBlockSize(iscol, &cbs));
3153: PetscCall(ISAllGather(iscol, &iscol_local));
3154: PetscCall(ISSetBlockSize(iscol_local, cbs));
3155: }
3157: *isseq = iscol_local;
3158: PetscFunctionReturn(PETSC_SUCCESS);
3159: }
3161: /*
3162: Used by MatCreateSubMatrix_MPIAIJ_SameRowColDist() to avoid ISAllGather() and global size of iscol_local
3163: (see MatCreateSubMatrix_MPIAIJ_nonscalable)
3165: Input Parameters:
3166: + mat - matrix
3167: . isrow - parallel row index set; its local indices are a subset of local columns of `mat`,
3168: i.e., mat->rstart <= isrow[i] < mat->rend
3169: - iscol - parallel column index set; its local indices are a subset of local columns of `mat`,
3170: i.e., mat->cstart <= iscol[i] < mat->cend
3172: Output Parameters:
3173: + isrow_d - sequential row index set for retrieving mat->A
3174: . iscol_d - sequential column index set for retrieving mat->A
3175: . iscol_o - sequential column index set for retrieving mat->B
3176: - garray - column map; garray[i] indicates global location of iscol_o[i] in `iscol`
3177: */
3178: static PetscErrorCode ISGetSeqIS_SameColDist_Private(Mat mat, IS isrow, IS iscol, IS *isrow_d, IS *iscol_d, IS *iscol_o, PetscInt *garray[])
3179: {
3180: Vec x, cmap;
3181: const PetscInt *is_idx;
3182: PetscScalar *xarray, *cmaparray;
3183: PetscInt ncols, isstart, *idx, m, rstart, *cmap1, count;
3184: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data;
3185: Mat B = a->B;
3186: Vec lvec = a->lvec, lcmap;
3187: PetscInt i, cstart, cend, Bn = B->cmap->N;
3188: MPI_Comm comm;
3189: VecScatter Mvctx = a->Mvctx;
3191: PetscFunctionBegin;
3192: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3193: PetscCall(ISGetLocalSize(iscol, &ncols));
3195: /* (1) iscol is a sub-column vector of mat, pad it with '-1.' to form a full vector x */
3196: PetscCall(MatCreateVecs(mat, &x, NULL));
3197: PetscCall(VecSet(x, -1.0));
3198: PetscCall(VecDuplicate(x, &cmap));
3199: PetscCall(VecSet(cmap, -1.0));
3201: /* Get start indices */
3202: PetscCallMPI(MPI_Scan(&ncols, &isstart, 1, MPIU_INT, MPI_SUM, comm));
3203: isstart -= ncols;
3204: PetscCall(MatGetOwnershipRangeColumn(mat, &cstart, &cend));
3206: PetscCall(ISGetIndices(iscol, &is_idx));
3207: PetscCall(VecGetArray(x, &xarray));
3208: PetscCall(VecGetArray(cmap, &cmaparray));
3209: PetscCall(PetscMalloc1(ncols, &idx));
3210: for (i = 0; i < ncols; i++) {
3211: xarray[is_idx[i] - cstart] = (PetscScalar)is_idx[i];
3212: cmaparray[is_idx[i] - cstart] = i + isstart; /* global index of iscol[i] */
3213: idx[i] = is_idx[i] - cstart; /* local index of iscol[i] */
3214: }
3215: PetscCall(VecRestoreArray(x, &xarray));
3216: PetscCall(VecRestoreArray(cmap, &cmaparray));
3217: PetscCall(ISRestoreIndices(iscol, &is_idx));
3219: /* Get iscol_d */
3220: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, iscol_d));
3221: PetscCall(ISGetBlockSize(iscol, &i));
3222: PetscCall(ISSetBlockSize(*iscol_d, i));
3224: /* Get isrow_d */
3225: PetscCall(ISGetLocalSize(isrow, &m));
3226: rstart = mat->rmap->rstart;
3227: PetscCall(PetscMalloc1(m, &idx));
3228: PetscCall(ISGetIndices(isrow, &is_idx));
3229: for (i = 0; i < m; i++) idx[i] = is_idx[i] - rstart;
3230: PetscCall(ISRestoreIndices(isrow, &is_idx));
3232: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, m, idx, PETSC_OWN_POINTER, isrow_d));
3233: PetscCall(ISGetBlockSize(isrow, &i));
3234: PetscCall(ISSetBlockSize(*isrow_d, i));
3236: /* (2) Scatter x and cmap using aij->Mvctx to get their off-process portions (see MatMult_MPIAIJ) */
3237: PetscCall(VecScatterBegin(Mvctx, x, lvec, INSERT_VALUES, SCATTER_FORWARD));
3238: PetscCall(VecScatterEnd(Mvctx, x, lvec, INSERT_VALUES, SCATTER_FORWARD));
3240: PetscCall(VecDuplicate(lvec, &lcmap));
3242: PetscCall(VecScatterBegin(Mvctx, cmap, lcmap, INSERT_VALUES, SCATTER_FORWARD));
3243: PetscCall(VecScatterEnd(Mvctx, cmap, lcmap, INSERT_VALUES, SCATTER_FORWARD));
3245: /* (3) create sequential iscol_o (a subset of iscol) and isgarray */
3246: /* off-process column indices */
3247: count = 0;
3248: PetscCall(PetscMalloc1(Bn, &idx));
3249: PetscCall(PetscMalloc1(Bn, &cmap1));
3251: PetscCall(VecGetArray(lvec, &xarray));
3252: PetscCall(VecGetArray(lcmap, &cmaparray));
3253: for (i = 0; i < Bn; i++) {
3254: if (PetscRealPart(xarray[i]) > -1.0) {
3255: idx[count] = i; /* local column index in off-diagonal part B */
3256: cmap1[count] = (PetscInt)PetscRealPart(cmaparray[i]); /* column index in submat */
3257: count++;
3258: }
3259: }
3260: PetscCall(VecRestoreArray(lvec, &xarray));
3261: PetscCall(VecRestoreArray(lcmap, &cmaparray));
3263: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, count, idx, PETSC_COPY_VALUES, iscol_o));
3264: /* cannot ensure iscol_o has same blocksize as iscol! */
3266: PetscCall(PetscFree(idx));
3267: *garray = cmap1;
3269: PetscCall(VecDestroy(&x));
3270: PetscCall(VecDestroy(&cmap));
3271: PetscCall(VecDestroy(&lcmap));
3272: PetscFunctionReturn(PETSC_SUCCESS);
3273: }
3275: /* isrow and iscol have same processor distribution as mat, output *submat is a submatrix of local mat */
3276: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowColDist(Mat mat, IS isrow, IS iscol, MatReuse call, Mat *submat)
3277: {
3278: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data, *asub;
3279: Mat M = NULL;
3280: MPI_Comm comm;
3281: IS iscol_d, isrow_d, iscol_o;
3282: Mat Asub = NULL, Bsub = NULL;
3283: PetscInt n;
3285: PetscFunctionBegin;
3286: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3288: if (call == MAT_REUSE_MATRIX) {
3289: /* Retrieve isrow_d, iscol_d and iscol_o from submat */
3290: PetscCall(PetscObjectQuery((PetscObject)*submat, "isrow_d", (PetscObject *)&isrow_d));
3291: PetscCheck(isrow_d, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "isrow_d passed in was not used before, cannot reuse");
3293: PetscCall(PetscObjectQuery((PetscObject)*submat, "iscol_d", (PetscObject *)&iscol_d));
3294: PetscCheck(iscol_d, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "iscol_d passed in was not used before, cannot reuse");
3296: PetscCall(PetscObjectQuery((PetscObject)*submat, "iscol_o", (PetscObject *)&iscol_o));
3297: PetscCheck(iscol_o, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "iscol_o passed in was not used before, cannot reuse");
3299: /* Update diagonal and off-diagonal portions of submat */
3300: asub = (Mat_MPIAIJ *)(*submat)->data;
3301: PetscCall(MatCreateSubMatrix_SeqAIJ(a->A, isrow_d, iscol_d, PETSC_DECIDE, MAT_REUSE_MATRIX, &asub->A));
3302: PetscCall(ISGetLocalSize(iscol_o, &n));
3303: if (n) PetscCall(MatCreateSubMatrix_SeqAIJ(a->B, isrow_d, iscol_o, PETSC_DECIDE, MAT_REUSE_MATRIX, &asub->B));
3304: PetscCall(MatAssemblyBegin(*submat, MAT_FINAL_ASSEMBLY));
3305: PetscCall(MatAssemblyEnd(*submat, MAT_FINAL_ASSEMBLY));
3307: } else { /* call == MAT_INITIAL_MATRIX) */
3308: PetscInt *garray;
3309: PetscInt BsubN;
3311: /* Create isrow_d, iscol_d, iscol_o and isgarray (replace isgarray with array?) */
3312: PetscCall(ISGetSeqIS_SameColDist_Private(mat, isrow, iscol, &isrow_d, &iscol_d, &iscol_o, &garray));
3314: /* Create local submatrices Asub and Bsub */
3315: PetscCall(MatCreateSubMatrix_SeqAIJ(a->A, isrow_d, iscol_d, PETSC_DECIDE, MAT_INITIAL_MATRIX, &Asub));
3316: PetscCall(MatCreateSubMatrix_SeqAIJ(a->B, isrow_d, iscol_o, PETSC_DECIDE, MAT_INITIAL_MATRIX, &Bsub));
3318: /* Create submatrix M */
3319: PetscCall(MatCreateMPIAIJWithSeqAIJ(comm, Asub, Bsub, garray, &M));
3321: /* If Bsub has empty columns, compress iscol_o such that it will retrieve condensed Bsub from a->B during reuse */
3322: asub = (Mat_MPIAIJ *)M->data;
3324: PetscCall(ISGetLocalSize(iscol_o, &BsubN));
3325: n = asub->B->cmap->N;
3326: if (BsubN > n) {
3327: /* This case can be tested using ~petsc/src/tao/bound/tutorials/runplate2_3 */
3328: const PetscInt *idx;
3329: PetscInt i, j, *idx_new, *subgarray = asub->garray;
3330: PetscCall(PetscInfo(M, "submatrix Bn %" PetscInt_FMT " != BsubN %" PetscInt_FMT ", update iscol_o\n", n, BsubN));
3332: PetscCall(PetscMalloc1(n, &idx_new));
3333: j = 0;
3334: PetscCall(ISGetIndices(iscol_o, &idx));
3335: for (i = 0; i < n; i++) {
3336: if (j >= BsubN) break;
3337: while (subgarray[i] > garray[j]) j++;
3339: if (subgarray[i] == garray[j]) {
3340: idx_new[i] = idx[j++];
3341: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "subgarray[%" PetscInt_FMT "]=%" PetscInt_FMT " cannot < garray[%" PetscInt_FMT "]=%" PetscInt_FMT, i, subgarray[i], j, garray[j]);
3342: }
3343: PetscCall(ISRestoreIndices(iscol_o, &idx));
3345: PetscCall(ISDestroy(&iscol_o));
3346: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, n, idx_new, PETSC_OWN_POINTER, &iscol_o));
3348: } else if (BsubN < n) {
3349: SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Columns of Bsub (%" PetscInt_FMT ") cannot be smaller than B's (%" PetscInt_FMT ")", BsubN, asub->B->cmap->N);
3350: }
3352: PetscCall(PetscFree(garray));
3353: *submat = M;
3355: /* Save isrow_d, iscol_d and iscol_o used in processor for next request */
3356: PetscCall(PetscObjectCompose((PetscObject)M, "isrow_d", (PetscObject)isrow_d));
3357: PetscCall(ISDestroy(&isrow_d));
3359: PetscCall(PetscObjectCompose((PetscObject)M, "iscol_d", (PetscObject)iscol_d));
3360: PetscCall(ISDestroy(&iscol_d));
3362: PetscCall(PetscObjectCompose((PetscObject)M, "iscol_o", (PetscObject)iscol_o));
3363: PetscCall(ISDestroy(&iscol_o));
3364: }
3365: PetscFunctionReturn(PETSC_SUCCESS);
3366: }
3368: PetscErrorCode MatCreateSubMatrix_MPIAIJ(Mat mat, IS isrow, IS iscol, MatReuse call, Mat *newmat)
3369: {
3370: IS iscol_local = NULL, isrow_d;
3371: PetscInt csize;
3372: PetscInt n, i, j, start, end;
3373: PetscBool sameRowDist = PETSC_FALSE, sameDist[2], tsameDist[2];
3374: MPI_Comm comm;
3376: PetscFunctionBegin;
3377: /* If isrow has same processor distribution as mat,
3378: call MatCreateSubMatrix_MPIAIJ_SameRowDist() to avoid using a hash table with global size of iscol */
3379: if (call == MAT_REUSE_MATRIX) {
3380: PetscCall(PetscObjectQuery((PetscObject)*newmat, "isrow_d", (PetscObject *)&isrow_d));
3381: if (isrow_d) {
3382: sameRowDist = PETSC_TRUE;
3383: tsameDist[1] = PETSC_TRUE; /* sameColDist */
3384: } else {
3385: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_local));
3386: if (iscol_local) {
3387: sameRowDist = PETSC_TRUE;
3388: tsameDist[1] = PETSC_FALSE; /* !sameColDist */
3389: }
3390: }
3391: } else {
3392: /* Check if isrow has same processor distribution as mat */
3393: sameDist[0] = PETSC_FALSE;
3394: PetscCall(ISGetLocalSize(isrow, &n));
3395: if (!n) {
3396: sameDist[0] = PETSC_TRUE;
3397: } else {
3398: PetscCall(ISGetMinMax(isrow, &i, &j));
3399: PetscCall(MatGetOwnershipRange(mat, &start, &end));
3400: if (i >= start && j < end) sameDist[0] = PETSC_TRUE;
3401: }
3403: /* Check if iscol has same processor distribution as mat */
3404: sameDist[1] = PETSC_FALSE;
3405: PetscCall(ISGetLocalSize(iscol, &n));
3406: if (!n) {
3407: sameDist[1] = PETSC_TRUE;
3408: } else {
3409: PetscCall(ISGetMinMax(iscol, &i, &j));
3410: PetscCall(MatGetOwnershipRangeColumn(mat, &start, &end));
3411: if (i >= start && j < end) sameDist[1] = PETSC_TRUE;
3412: }
3414: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3415: PetscCallMPI(MPIU_Allreduce(&sameDist, &tsameDist, 2, MPIU_BOOL, MPI_LAND, comm));
3416: sameRowDist = tsameDist[0];
3417: }
3419: if (sameRowDist) {
3420: if (tsameDist[1]) { /* sameRowDist & sameColDist */
3421: /* isrow and iscol have same processor distribution as mat */
3422: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowColDist(mat, isrow, iscol, call, newmat));
3423: PetscFunctionReturn(PETSC_SUCCESS);
3424: } else { /* sameRowDist */
3425: /* isrow has same processor distribution as mat */
3426: if (call == MAT_INITIAL_MATRIX) {
3427: PetscBool sorted;
3428: PetscCall(ISGetSeqIS_Private(mat, iscol, &iscol_local));
3429: PetscCall(ISGetLocalSize(iscol_local, &n)); /* local size of iscol_local = global columns of newmat */
3430: PetscCall(ISGetSize(iscol, &i));
3431: PetscCheck(n == i, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "n %" PetscInt_FMT " != size of iscol %" PetscInt_FMT, n, i);
3433: PetscCall(ISSorted(iscol_local, &sorted));
3434: if (sorted) {
3435: /* MatCreateSubMatrix_MPIAIJ_SameRowDist() requires iscol_local be sorted; it can have duplicate indices */
3436: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowDist(mat, isrow, iscol, iscol_local, MAT_INITIAL_MATRIX, newmat));
3437: PetscFunctionReturn(PETSC_SUCCESS);
3438: }
3439: } else { /* call == MAT_REUSE_MATRIX */
3440: IS iscol_sub;
3441: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_sub));
3442: if (iscol_sub) {
3443: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowDist(mat, isrow, iscol, NULL, call, newmat));
3444: PetscFunctionReturn(PETSC_SUCCESS);
3445: }
3446: }
3447: }
3448: }
3450: /* General case: iscol -> iscol_local which has global size of iscol */
3451: if (call == MAT_REUSE_MATRIX) {
3452: PetscCall(PetscObjectQuery((PetscObject)*newmat, "ISAllGather", (PetscObject *)&iscol_local));
3453: PetscCheck(iscol_local, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3454: } else {
3455: if (!iscol_local) PetscCall(ISGetSeqIS_Private(mat, iscol, &iscol_local));
3456: }
3458: PetscCall(ISGetLocalSize(iscol, &csize));
3459: PetscCall(MatCreateSubMatrix_MPIAIJ_nonscalable(mat, isrow, iscol_local, csize, call, newmat));
3461: if (call == MAT_INITIAL_MATRIX) {
3462: PetscCall(PetscObjectCompose((PetscObject)*newmat, "ISAllGather", (PetscObject)iscol_local));
3463: PetscCall(ISDestroy(&iscol_local));
3464: }
3465: PetscFunctionReturn(PETSC_SUCCESS);
3466: }
3468: /*@C
3469: MatCreateMPIAIJWithSeqAIJ - creates a `MATMPIAIJ` matrix using `MATSEQAIJ` matrices that contain the "diagonal"
3470: and "off-diagonal" part of the matrix in CSR format.
3472: Collective
3474: Input Parameters:
3475: + comm - MPI communicator
3476: . A - "diagonal" portion of matrix
3477: . B - "off-diagonal" portion of matrix, may have empty columns, will be destroyed by this routine
3478: - garray - global index of `B` columns
3480: Output Parameter:
3481: . mat - the matrix, with input `A` as its local diagonal matrix
3483: Level: advanced
3485: Notes:
3486: See `MatCreateAIJ()` for the definition of "diagonal" and "off-diagonal" portion of the matrix.
3488: `A` becomes part of output mat, `B` is destroyed by this routine. The user cannot use `A` and `B` anymore.
3490: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MATSEQAIJ`, `MatCreateMPIAIJWithSplitArrays()`
3491: @*/
3492: PetscErrorCode MatCreateMPIAIJWithSeqAIJ(MPI_Comm comm, Mat A, Mat B, const PetscInt garray[], Mat *mat)
3493: {
3494: Mat_MPIAIJ *maij;
3495: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data, *bnew;
3496: PetscInt *oi = b->i, *oj = b->j, i, nz, col;
3497: const PetscScalar *oa;
3498: Mat Bnew;
3499: PetscInt m, n, N;
3500: MatType mpi_mat_type;
3502: PetscFunctionBegin;
3503: PetscCall(MatCreate(comm, mat));
3504: PetscCall(MatGetSize(A, &m, &n));
3505: PetscCheck(m == B->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Am %" PetscInt_FMT " != Bm %" PetscInt_FMT, m, B->rmap->N);
3506: PetscCheck(PetscAbs(A->rmap->bs) == PetscAbs(B->rmap->bs), PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "A row bs %" PetscInt_FMT " != B row bs %" PetscInt_FMT, A->rmap->bs, B->rmap->bs);
3507: /* remove check below; When B is created using iscol_o from ISGetSeqIS_SameColDist_Private(), its bs may not be same as A */
3508: /* PetscCheck(A->cmap->bs == B->cmap->bs,PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"A column bs %" PetscInt_FMT " != B column bs %" PetscInt_FMT,A->cmap->bs,B->cmap->bs); */
3510: /* Get global columns of mat */
3511: PetscCallMPI(MPIU_Allreduce(&n, &N, 1, MPIU_INT, MPI_SUM, comm));
3513: PetscCall(MatSetSizes(*mat, m, n, PETSC_DECIDE, N));
3514: /* Determine the type of MPI matrix that should be created from the type of matrix A, which holds the "diagonal" portion. */
3515: PetscCall(MatGetMPIMatType_Private(A, &mpi_mat_type));
3516: PetscCall(MatSetType(*mat, mpi_mat_type));
3518: if (A->rmap->bs > 1 || A->cmap->bs > 1) PetscCall(MatSetBlockSizes(*mat, A->rmap->bs, A->cmap->bs));
3519: maij = (Mat_MPIAIJ *)(*mat)->data;
3521: (*mat)->preallocated = PETSC_TRUE;
3523: PetscCall(PetscLayoutSetUp((*mat)->rmap));
3524: PetscCall(PetscLayoutSetUp((*mat)->cmap));
3526: /* Set A as diagonal portion of *mat */
3527: maij->A = A;
3529: nz = oi[m];
3530: for (i = 0; i < nz; i++) {
3531: col = oj[i];
3532: oj[i] = garray[col];
3533: }
3535: /* Set Bnew as off-diagonal portion of *mat */
3536: PetscCall(MatSeqAIJGetArrayRead(B, &oa));
3537: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, N, oi, oj, (PetscScalar *)oa, &Bnew));
3538: PetscCall(MatSeqAIJRestoreArrayRead(B, &oa));
3539: bnew = (Mat_SeqAIJ *)Bnew->data;
3540: bnew->maxnz = b->maxnz; /* allocated nonzeros of B */
3541: maij->B = Bnew;
3543: PetscCheck(B->rmap->N == Bnew->rmap->N, PETSC_COMM_SELF, PETSC_ERR_PLIB, "BN %" PetscInt_FMT " != BnewN %" PetscInt_FMT, B->rmap->N, Bnew->rmap->N);
3545: b->free_a = PETSC_FALSE;
3546: b->free_ij = PETSC_FALSE;
3547: PetscCall(MatDestroy(&B));
3549: bnew->free_a = PETSC_TRUE;
3550: bnew->free_ij = PETSC_TRUE;
3552: /* condense columns of maij->B */
3553: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
3554: PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY));
3555: PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY));
3556: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_FALSE));
3557: PetscCall(MatSetOption(*mat, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
3558: PetscFunctionReturn(PETSC_SUCCESS);
3559: }
3561: extern PetscErrorCode MatCreateSubMatrices_MPIAIJ_SingleIS_Local(Mat, PetscInt, const IS[], const IS[], MatReuse, PetscBool, Mat *);
3563: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowDist(Mat mat, IS isrow, IS iscol, IS iscol_local, MatReuse call, Mat *newmat)
3564: {
3565: PetscInt i, m, n, rstart, row, rend, nz, j, bs, cbs;
3566: PetscInt *ii, *jj, nlocal, *dlens, *olens, dlen, olen, jend, mglobal;
3567: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data;
3568: Mat M, Msub, B = a->B;
3569: MatScalar *aa;
3570: Mat_SeqAIJ *aij;
3571: PetscInt *garray = a->garray, *colsub, Ncols;
3572: PetscInt count, Bn = B->cmap->N, cstart = mat->cmap->rstart, cend = mat->cmap->rend;
3573: IS iscol_sub, iscmap;
3574: const PetscInt *is_idx, *cmap;
3575: PetscBool allcolumns = PETSC_FALSE;
3576: MPI_Comm comm;
3578: PetscFunctionBegin;
3579: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3580: if (call == MAT_REUSE_MATRIX) {
3581: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_sub));
3582: PetscCheck(iscol_sub, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "SubIScol passed in was not used before, cannot reuse");
3583: PetscCall(ISGetLocalSize(iscol_sub, &count));
3585: PetscCall(PetscObjectQuery((PetscObject)*newmat, "Subcmap", (PetscObject *)&iscmap));
3586: PetscCheck(iscmap, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Subcmap passed in was not used before, cannot reuse");
3588: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubMatrix", (PetscObject *)&Msub));
3589: PetscCheck(Msub, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3591: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol_sub, MAT_REUSE_MATRIX, PETSC_FALSE, &Msub));
3593: } else { /* call == MAT_INITIAL_MATRIX) */
3594: PetscBool flg;
3596: PetscCall(ISGetLocalSize(iscol, &n));
3597: PetscCall(ISGetSize(iscol, &Ncols));
3599: /* (1) iscol -> nonscalable iscol_local */
3600: /* Check for special case: each processor gets entire matrix columns */
3601: PetscCall(ISIdentity(iscol_local, &flg));
3602: if (flg && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3603: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &allcolumns, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
3604: if (allcolumns) {
3605: iscol_sub = iscol_local;
3606: PetscCall(PetscObjectReference((PetscObject)iscol_local));
3607: PetscCall(ISCreateStride(PETSC_COMM_SELF, n, 0, 1, &iscmap));
3609: } else {
3610: /* (2) iscol_local -> iscol_sub and iscmap. Implementation below requires iscol_local be sorted, it can have duplicate indices */
3611: PetscInt *idx, *cmap1, k;
3612: PetscCall(PetscMalloc1(Ncols, &idx));
3613: PetscCall(PetscMalloc1(Ncols, &cmap1));
3614: PetscCall(ISGetIndices(iscol_local, &is_idx));
3615: count = 0;
3616: k = 0;
3617: for (i = 0; i < Ncols; i++) {
3618: j = is_idx[i];
3619: if (j >= cstart && j < cend) {
3620: /* diagonal part of mat */
3621: idx[count] = j;
3622: cmap1[count++] = i; /* column index in submat */
3623: } else if (Bn) {
3624: /* off-diagonal part of mat */
3625: if (j == garray[k]) {
3626: idx[count] = j;
3627: cmap1[count++] = i; /* column index in submat */
3628: } else if (j > garray[k]) {
3629: while (j > garray[k] && k < Bn - 1) k++;
3630: if (j == garray[k]) {
3631: idx[count] = j;
3632: cmap1[count++] = i; /* column index in submat */
3633: }
3634: }
3635: }
3636: }
3637: PetscCall(ISRestoreIndices(iscol_local, &is_idx));
3639: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, count, idx, PETSC_OWN_POINTER, &iscol_sub));
3640: PetscCall(ISGetBlockSize(iscol, &cbs));
3641: PetscCall(ISSetBlockSize(iscol_sub, cbs));
3643: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)iscol_local), count, cmap1, PETSC_OWN_POINTER, &iscmap));
3644: }
3646: /* (3) Create sequential Msub */
3647: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol_sub, MAT_INITIAL_MATRIX, allcolumns, &Msub));
3648: }
3650: PetscCall(ISGetLocalSize(iscol_sub, &count));
3651: aij = (Mat_SeqAIJ *)Msub->data;
3652: ii = aij->i;
3653: PetscCall(ISGetIndices(iscmap, &cmap));
3655: /*
3656: m - number of local rows
3657: Ncols - number of columns (same on all processors)
3658: rstart - first row in new global matrix generated
3659: */
3660: PetscCall(MatGetSize(Msub, &m, NULL));
3662: if (call == MAT_INITIAL_MATRIX) {
3663: /* (4) Create parallel newmat */
3664: PetscMPIInt rank, size;
3665: PetscInt csize;
3667: PetscCallMPI(MPI_Comm_size(comm, &size));
3668: PetscCallMPI(MPI_Comm_rank(comm, &rank));
3670: /*
3671: Determine the number of non-zeros in the diagonal and off-diagonal
3672: portions of the matrix in order to do correct preallocation
3673: */
3675: /* first get start and end of "diagonal" columns */
3676: PetscCall(ISGetLocalSize(iscol, &csize));
3677: if (csize == PETSC_DECIDE) {
3678: PetscCall(ISGetSize(isrow, &mglobal));
3679: if (mglobal == Ncols) { /* square matrix */
3680: nlocal = m;
3681: } else {
3682: nlocal = Ncols / size + ((Ncols % size) > rank);
3683: }
3684: } else {
3685: nlocal = csize;
3686: }
3687: PetscCallMPI(MPI_Scan(&nlocal, &rend, 1, MPIU_INT, MPI_SUM, comm));
3688: rstart = rend - nlocal;
3689: 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);
3691: /* next, compute all the lengths */
3692: jj = aij->j;
3693: PetscCall(PetscMalloc1(2 * m + 1, &dlens));
3694: olens = dlens + m;
3695: for (i = 0; i < m; i++) {
3696: jend = ii[i + 1] - ii[i];
3697: olen = 0;
3698: dlen = 0;
3699: for (j = 0; j < jend; j++) {
3700: if (cmap[*jj] < rstart || cmap[*jj] >= rend) olen++;
3701: else dlen++;
3702: jj++;
3703: }
3704: olens[i] = olen;
3705: dlens[i] = dlen;
3706: }
3708: PetscCall(ISGetBlockSize(isrow, &bs));
3709: PetscCall(ISGetBlockSize(iscol, &cbs));
3711: PetscCall(MatCreate(comm, &M));
3712: PetscCall(MatSetSizes(M, m, nlocal, PETSC_DECIDE, Ncols));
3713: PetscCall(MatSetBlockSizes(M, bs, cbs));
3714: PetscCall(MatSetType(M, ((PetscObject)mat)->type_name));
3715: PetscCall(MatMPIAIJSetPreallocation(M, 0, dlens, 0, olens));
3716: PetscCall(PetscFree(dlens));
3718: } else { /* call == MAT_REUSE_MATRIX */
3719: M = *newmat;
3720: PetscCall(MatGetLocalSize(M, &i, NULL));
3721: PetscCheck(i == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Previous matrix must be same size/layout as request");
3722: PetscCall(MatZeroEntries(M));
3723: /*
3724: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3725: rather than the slower MatSetValues().
3726: */
3727: M->was_assembled = PETSC_TRUE;
3728: M->assembled = PETSC_FALSE;
3729: }
3731: /* (5) Set values of Msub to *newmat */
3732: PetscCall(PetscMalloc1(count, &colsub));
3733: PetscCall(MatGetOwnershipRange(M, &rstart, NULL));
3735: jj = aij->j;
3736: PetscCall(MatSeqAIJGetArrayRead(Msub, (const PetscScalar **)&aa));
3737: for (i = 0; i < m; i++) {
3738: row = rstart + i;
3739: nz = ii[i + 1] - ii[i];
3740: for (j = 0; j < nz; j++) colsub[j] = cmap[jj[j]];
3741: PetscCall(MatSetValues_MPIAIJ(M, 1, &row, nz, colsub, aa, INSERT_VALUES));
3742: jj += nz;
3743: aa += nz;
3744: }
3745: PetscCall(MatSeqAIJRestoreArrayRead(Msub, (const PetscScalar **)&aa));
3746: PetscCall(ISRestoreIndices(iscmap, &cmap));
3748: PetscCall(MatAssemblyBegin(M, MAT_FINAL_ASSEMBLY));
3749: PetscCall(MatAssemblyEnd(M, MAT_FINAL_ASSEMBLY));
3751: PetscCall(PetscFree(colsub));
3753: /* save Msub, iscol_sub and iscmap used in processor for next request */
3754: if (call == MAT_INITIAL_MATRIX) {
3755: *newmat = M;
3756: PetscCall(PetscObjectCompose((PetscObject)*newmat, "SubMatrix", (PetscObject)Msub));
3757: PetscCall(MatDestroy(&Msub));
3759: PetscCall(PetscObjectCompose((PetscObject)*newmat, "SubIScol", (PetscObject)iscol_sub));
3760: PetscCall(ISDestroy(&iscol_sub));
3762: PetscCall(PetscObjectCompose((PetscObject)*newmat, "Subcmap", (PetscObject)iscmap));
3763: PetscCall(ISDestroy(&iscmap));
3765: if (iscol_local) {
3766: PetscCall(PetscObjectCompose((PetscObject)*newmat, "ISAllGather", (PetscObject)iscol_local));
3767: PetscCall(ISDestroy(&iscol_local));
3768: }
3769: }
3770: PetscFunctionReturn(PETSC_SUCCESS);
3771: }
3773: /*
3774: Not great since it makes two copies of the submatrix, first an SeqAIJ
3775: in local and then by concatenating the local matrices the end result.
3776: Writing it directly would be much like MatCreateSubMatrices_MPIAIJ()
3778: This requires a sequential iscol with all indices.
3779: */
3780: PetscErrorCode MatCreateSubMatrix_MPIAIJ_nonscalable(Mat mat, IS isrow, IS iscol, PetscInt csize, MatReuse call, Mat *newmat)
3781: {
3782: PetscMPIInt rank, size;
3783: PetscInt i, m, n, rstart, row, rend, nz, *cwork, j, bs, cbs;
3784: PetscInt *ii, *jj, nlocal, *dlens, *olens, dlen, olen, jend, mglobal;
3785: Mat M, Mreuse;
3786: MatScalar *aa, *vwork;
3787: MPI_Comm comm;
3788: Mat_SeqAIJ *aij;
3789: PetscBool colflag, allcolumns = PETSC_FALSE;
3791: PetscFunctionBegin;
3792: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3793: PetscCallMPI(MPI_Comm_rank(comm, &rank));
3794: PetscCallMPI(MPI_Comm_size(comm, &size));
3796: /* Check for special case: each processor gets entire matrix columns */
3797: PetscCall(ISIdentity(iscol, &colflag));
3798: PetscCall(ISGetLocalSize(iscol, &n));
3799: if (colflag && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3800: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &allcolumns, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
3802: if (call == MAT_REUSE_MATRIX) {
3803: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubMatrix", (PetscObject *)&Mreuse));
3804: PetscCheck(Mreuse, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3805: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol, MAT_REUSE_MATRIX, allcolumns, &Mreuse));
3806: } else {
3807: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol, MAT_INITIAL_MATRIX, allcolumns, &Mreuse));
3808: }
3810: /*
3811: m - number of local rows
3812: n - number of columns (same on all processors)
3813: rstart - first row in new global matrix generated
3814: */
3815: PetscCall(MatGetSize(Mreuse, &m, &n));
3816: PetscCall(MatGetBlockSizes(Mreuse, &bs, &cbs));
3817: if (call == MAT_INITIAL_MATRIX) {
3818: aij = (Mat_SeqAIJ *)Mreuse->data;
3819: ii = aij->i;
3820: jj = aij->j;
3822: /*
3823: Determine the number of non-zeros in the diagonal and off-diagonal
3824: portions of the matrix in order to do correct preallocation
3825: */
3827: /* first get start and end of "diagonal" columns */
3828: if (csize == PETSC_DECIDE) {
3829: PetscCall(ISGetSize(isrow, &mglobal));
3830: if (mglobal == n) { /* square matrix */
3831: nlocal = m;
3832: } else {
3833: nlocal = n / size + ((n % size) > rank);
3834: }
3835: } else {
3836: nlocal = csize;
3837: }
3838: PetscCallMPI(MPI_Scan(&nlocal, &rend, 1, MPIU_INT, MPI_SUM, comm));
3839: rstart = rend - nlocal;
3840: 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);
3842: /* next, compute all the lengths */
3843: PetscCall(PetscMalloc1(2 * m + 1, &dlens));
3844: olens = dlens + m;
3845: for (i = 0; i < m; i++) {
3846: jend = ii[i + 1] - ii[i];
3847: olen = 0;
3848: dlen = 0;
3849: for (j = 0; j < jend; j++) {
3850: if (*jj < rstart || *jj >= rend) olen++;
3851: else dlen++;
3852: jj++;
3853: }
3854: olens[i] = olen;
3855: dlens[i] = dlen;
3856: }
3857: PetscCall(MatCreate(comm, &M));
3858: PetscCall(MatSetSizes(M, m, nlocal, PETSC_DECIDE, n));
3859: PetscCall(MatSetBlockSizes(M, bs, cbs));
3860: PetscCall(MatSetType(M, ((PetscObject)mat)->type_name));
3861: PetscCall(MatMPIAIJSetPreallocation(M, 0, dlens, 0, olens));
3862: PetscCall(PetscFree(dlens));
3863: } else {
3864: PetscInt ml, nl;
3866: M = *newmat;
3867: PetscCall(MatGetLocalSize(M, &ml, &nl));
3868: PetscCheck(ml == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Previous matrix must be same size/layout as request");
3869: PetscCall(MatZeroEntries(M));
3870: /*
3871: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3872: rather than the slower MatSetValues().
3873: */
3874: M->was_assembled = PETSC_TRUE;
3875: M->assembled = PETSC_FALSE;
3876: }
3877: PetscCall(MatGetOwnershipRange(M, &rstart, &rend));
3878: aij = (Mat_SeqAIJ *)Mreuse->data;
3879: ii = aij->i;
3880: jj = aij->j;
3882: /* trigger copy to CPU if needed */
3883: PetscCall(MatSeqAIJGetArrayRead(Mreuse, (const PetscScalar **)&aa));
3884: for (i = 0; i < m; i++) {
3885: row = rstart + i;
3886: nz = ii[i + 1] - ii[i];
3887: cwork = jj;
3888: jj = PetscSafePointerPlusOffset(jj, nz);
3889: vwork = aa;
3890: aa = PetscSafePointerPlusOffset(aa, nz);
3891: PetscCall(MatSetValues_MPIAIJ(M, 1, &row, nz, cwork, vwork, INSERT_VALUES));
3892: }
3893: PetscCall(MatSeqAIJRestoreArrayRead(Mreuse, (const PetscScalar **)&aa));
3895: PetscCall(MatAssemblyBegin(M, MAT_FINAL_ASSEMBLY));
3896: PetscCall(MatAssemblyEnd(M, MAT_FINAL_ASSEMBLY));
3897: *newmat = M;
3899: /* save submatrix used in processor for next request */
3900: if (call == MAT_INITIAL_MATRIX) {
3901: PetscCall(PetscObjectCompose((PetscObject)M, "SubMatrix", (PetscObject)Mreuse));
3902: PetscCall(MatDestroy(&Mreuse));
3903: }
3904: PetscFunctionReturn(PETSC_SUCCESS);
3905: }
3907: static PetscErrorCode MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[])
3908: {
3909: PetscInt m, cstart, cend, j, nnz, i, d, *ld;
3910: PetscInt *d_nnz, *o_nnz, nnz_max = 0, rstart, ii, irstart;
3911: const PetscInt *JJ;
3912: PetscBool nooffprocentries;
3913: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)B->data;
3915: PetscFunctionBegin;
3916: PetscCall(PetscLayoutSetUp(B->rmap));
3917: PetscCall(PetscLayoutSetUp(B->cmap));
3918: m = B->rmap->n;
3919: cstart = B->cmap->rstart;
3920: cend = B->cmap->rend;
3921: rstart = B->rmap->rstart;
3922: irstart = Ii[0];
3924: PetscCall(PetscCalloc2(m, &d_nnz, m, &o_nnz));
3926: if (PetscDefined(USE_DEBUG)) {
3927: for (i = 0; i < m; i++) {
3928: nnz = Ii[i + 1] - Ii[i];
3929: JJ = PetscSafePointerPlusOffset(J, Ii[i] - irstart);
3930: PetscCheck(nnz >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Local row %" PetscInt_FMT " has a negative %" PetscInt_FMT " number of columns", i, nnz);
3931: 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]);
3932: 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);
3933: }
3934: }
3936: for (i = 0; i < m; i++) {
3937: nnz = Ii[i + 1] - Ii[i];
3938: JJ = PetscSafePointerPlusOffset(J, Ii[i] - irstart);
3939: nnz_max = PetscMax(nnz_max, nnz);
3940: d = 0;
3941: for (j = 0; j < nnz; j++) {
3942: if (cstart <= JJ[j] && JJ[j] < cend) d++;
3943: }
3944: d_nnz[i] = d;
3945: o_nnz[i] = nnz - d;
3946: }
3947: PetscCall(MatMPIAIJSetPreallocation(B, 0, d_nnz, 0, o_nnz));
3948: PetscCall(PetscFree2(d_nnz, o_nnz));
3950: for (i = 0; i < m; i++) {
3951: ii = i + rstart;
3952: PetscCall(MatSetValues_MPIAIJ(B, 1, &ii, Ii[i + 1] - Ii[i], PetscSafePointerPlusOffset(J, Ii[i] - irstart), PetscSafePointerPlusOffset(v, Ii[i] - irstart), INSERT_VALUES));
3953: }
3954: nooffprocentries = B->nooffprocentries;
3955: B->nooffprocentries = PETSC_TRUE;
3956: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
3957: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
3958: B->nooffprocentries = nooffprocentries;
3960: /* count number of entries below block diagonal */
3961: PetscCall(PetscFree(Aij->ld));
3962: PetscCall(PetscCalloc1(m, &ld));
3963: Aij->ld = ld;
3964: for (i = 0; i < m; i++) {
3965: nnz = Ii[i + 1] - Ii[i];
3966: j = 0;
3967: while (j < nnz && J[j] < cstart) j++;
3968: ld[i] = j;
3969: if (J) J += nnz;
3970: }
3972: PetscCall(MatSetOption(B, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
3973: PetscFunctionReturn(PETSC_SUCCESS);
3974: }
3976: /*@
3977: MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in `MATAIJ` format
3978: (the default parallel PETSc format).
3980: Collective
3982: Input Parameters:
3983: + B - the matrix
3984: . i - the indices into `j` for the start of each local row (indices start with zero)
3985: . j - the column indices for each local row (indices start with zero)
3986: - v - optional values in the matrix
3988: Level: developer
3990: Notes:
3991: The `i`, `j`, and `v` arrays ARE copied by this routine into the internal format used by PETSc;
3992: thus you CANNOT change the matrix entries by changing the values of `v` after you have
3993: called this routine. Use `MatCreateMPIAIJWithSplitArrays()` to avoid needing to copy the arrays.
3995: The `i` and `j` indices are 0 based, and `i` indices are indices corresponding to the local `j` array.
3997: A convenience routine for this functionality is `MatCreateMPIAIJWithArrays()`.
3999: You can update the matrix with new numerical values using `MatUpdateMPIAIJWithArrays()` after this call if the column indices in `j` are sorted.
4001: If you do **not** use `MatUpdateMPIAIJWithArrays()`, the column indices in `j` do not need to be sorted. If you will use
4002: `MatUpdateMPIAIJWithArrays()`, the column indices **must** be sorted.
4004: The format which is used for the sparse matrix input, is equivalent to a
4005: row-major ordering.. i.e for the following matrix, the input data expected is
4006: as shown
4007: .vb
4008: 1 0 0
4009: 2 0 3 P0
4010: -------
4011: 4 5 6 P1
4013: Process0 [P0] rows_owned=[0,1]
4014: i = {0,1,3} [size = nrow+1 = 2+1]
4015: j = {0,0,2} [size = 3]
4016: v = {1,2,3} [size = 3]
4018: Process1 [P1] rows_owned=[2]
4019: i = {0,3} [size = nrow+1 = 1+1]
4020: j = {0,1,2} [size = 3]
4021: v = {4,5,6} [size = 3]
4022: .ve
4024: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatCreateAIJ()`,
4025: `MatCreateSeqAIJWithArrays()`, `MatCreateMPIAIJWithSplitArrays()`, `MatCreateMPIAIJWithArrays()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4026: @*/
4027: PetscErrorCode MatMPIAIJSetPreallocationCSR(Mat B, const PetscInt i[], const PetscInt j[], const PetscScalar v[])
4028: {
4029: PetscFunctionBegin;
4030: PetscTryMethod(B, "MatMPIAIJSetPreallocationCSR_C", (Mat, const PetscInt[], const PetscInt[], const PetscScalar[]), (B, i, j, v));
4031: PetscFunctionReturn(PETSC_SUCCESS);
4032: }
4034: /*@
4035: MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in `MATMPIAIJ` format
4036: (the default parallel PETSc format). For good matrix assembly performance
4037: the user should preallocate the matrix storage by setting the parameters
4038: `d_nz` (or `d_nnz`) and `o_nz` (or `o_nnz`).
4040: Collective
4042: Input Parameters:
4043: + B - the matrix
4044: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
4045: (same value is used for all local rows)
4046: . d_nnz - array containing the number of nonzeros in the various rows of the
4047: DIAGONAL portion of the local submatrix (possibly different for each row)
4048: or `NULL` (`PETSC_NULL_INTEGER` in Fortran), if `d_nz` is used to specify the nonzero structure.
4049: The size of this array is equal to the number of local rows, i.e 'm'.
4050: For matrices that will be factored, you must leave room for (and set)
4051: the diagonal entry even if it is zero.
4052: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
4053: submatrix (same value is used for all local rows).
4054: - o_nnz - array containing the number of nonzeros in the various rows of the
4055: OFF-DIAGONAL portion of the local submatrix (possibly different for
4056: each row) or `NULL` (`PETSC_NULL_INTEGER` in Fortran), if `o_nz` is used to specify the nonzero
4057: structure. The size of this array is equal to the number
4058: of local rows, i.e 'm'.
4060: Example Usage:
4061: Consider the following 8x8 matrix with 34 non-zero values, that is
4062: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4063: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4064: as follows
4066: .vb
4067: 1 2 0 | 0 3 0 | 0 4
4068: Proc0 0 5 6 | 7 0 0 | 8 0
4069: 9 0 10 | 11 0 0 | 12 0
4070: -------------------------------------
4071: 13 0 14 | 15 16 17 | 0 0
4072: Proc1 0 18 0 | 19 20 21 | 0 0
4073: 0 0 0 | 22 23 0 | 24 0
4074: -------------------------------------
4075: Proc2 25 26 27 | 0 0 28 | 29 0
4076: 30 0 0 | 31 32 33 | 0 34
4077: .ve
4079: This can be represented as a collection of submatrices as
4080: .vb
4081: A B C
4082: D E F
4083: G H I
4084: .ve
4086: Where the submatrices A,B,C are owned by proc0, D,E,F are
4087: owned by proc1, G,H,I are owned by proc2.
4089: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4090: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4091: The 'M','N' parameters are 8,8, and have the same values on all procs.
4093: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4094: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4095: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4096: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4097: part as `MATSEQAIJ` matrices. For example, proc1 will store [E] as a `MATSEQAIJ`
4098: matrix, and [DF] as another `MATSEQAIJ` matrix.
4100: When `d_nz`, `o_nz` parameters are specified, `d_nz` storage elements are
4101: allocated for every row of the local DIAGONAL submatrix, and `o_nz`
4102: storage locations are allocated for every row of the OFF-DIAGONAL submatrix.
4103: One way to choose `d_nz` and `o_nz` is to use the maximum number of nonzeros over
4104: the local rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4105: In this case, the values of `d_nz`, `o_nz` are
4106: .vb
4107: proc0 dnz = 2, o_nz = 2
4108: proc1 dnz = 3, o_nz = 2
4109: proc2 dnz = 1, o_nz = 4
4110: .ve
4111: We are allocating `m`*(`d_nz`+`o_nz`) storage locations for every proc. This
4112: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4113: for proc3. i.e we are using 12+15+10=37 storage locations to store
4114: 34 values.
4116: When `d_nnz`, `o_nnz` parameters are specified, the storage is specified
4117: for every row, corresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4118: In the above case the values for `d_nnz`, `o_nnz` are
4119: .vb
4120: proc0 d_nnz = [2,2,2] and o_nnz = [2,2,2]
4121: proc1 d_nnz = [3,3,2] and o_nnz = [2,1,1]
4122: proc2 d_nnz = [1,1] and o_nnz = [4,4]
4123: .ve
4124: Here the space allocated is sum of all the above values i.e 34, and
4125: hence pre-allocation is perfect.
4127: Level: intermediate
4129: Notes:
4130: If the *_nnz parameter is given then the *_nz parameter is ignored
4132: The `MATAIJ` format, also called compressed row storage (CSR), is compatible with standard Fortran
4133: storage. The stored row and column indices begin with zero.
4134: See [Sparse Matrices](sec_matsparse) for details.
4136: The parallel matrix is partitioned such that the first m0 rows belong to
4137: process 0, the next m1 rows belong to process 1, the next m2 rows belong
4138: to process 2 etc.. where m0,m1,m2... are the input parameter 'm'.
4140: The DIAGONAL portion of the local submatrix of a processor can be defined
4141: as the submatrix which is obtained by extraction the part corresponding to
4142: the rows r1-r2 and columns c1-c2 of the global matrix, where r1 is the
4143: first row that belongs to the processor, r2 is the last row belonging to
4144: the this processor, and c1-c2 is range of indices of the local part of a
4145: vector suitable for applying the matrix to. This is an mxn matrix. In the
4146: common case of a square matrix, the row and column ranges are the same and
4147: the DIAGONAL part is also square. The remaining portion of the local
4148: submatrix (mxN) constitute the OFF-DIAGONAL portion.
4150: If `o_nnz` and `d_nnz` are specified, then `o_nz` and `d_nz` are ignored.
4152: You can call `MatGetInfo()` to get information on how effective the preallocation was;
4153: for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
4154: You can also run with the option `-info` and look for messages with the string
4155: malloc in them to see if additional memory allocation was needed.
4157: .seealso: [](ch_matrices), `Mat`, [Sparse Matrices](sec_matsparse), `MATMPIAIJ`, `MATAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatCreateAIJ()`, `MatMPIAIJSetPreallocationCSR()`,
4158: `MatGetInfo()`, `PetscSplitOwnership()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4159: @*/
4160: PetscErrorCode MatMPIAIJSetPreallocation(Mat B, PetscInt d_nz, const PetscInt d_nnz[], PetscInt o_nz, const PetscInt o_nnz[])
4161: {
4162: PetscFunctionBegin;
4165: PetscTryMethod(B, "MatMPIAIJSetPreallocation_C", (Mat, PetscInt, const PetscInt[], PetscInt, const PetscInt[]), (B, d_nz, d_nnz, o_nz, o_nnz));
4166: PetscFunctionReturn(PETSC_SUCCESS);
4167: }
4169: /*@
4170: MatCreateMPIAIJWithArrays - creates a `MATMPIAIJ` matrix using arrays that contain in standard
4171: CSR format for the local rows.
4173: Collective
4175: Input Parameters:
4176: + comm - MPI communicator
4177: . m - number of local rows (Cannot be `PETSC_DECIDE`)
4178: . n - This value should be the same as the local size used in creating the
4179: x vector for the matrix-vector product $ y = Ax$. (or `PETSC_DECIDE` to have
4180: calculated if `N` is given) For square matrices n is almost always `m`.
4181: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if `m` is given)
4182: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if `n` is given)
4183: . 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
4184: . j - global column indices
4185: - a - optional matrix values
4187: Output Parameter:
4188: . mat - the matrix
4190: Level: intermediate
4192: Notes:
4193: The `i`, `j`, and `a` arrays ARE copied by this routine into the internal format used by PETSc;
4194: thus you CANNOT change the matrix entries by changing the values of `a[]` after you have
4195: called this routine. Use `MatCreateMPIAIJWithSplitArrays()` to avoid needing to copy the arrays.
4197: The `i` and `j` indices are 0 based, and `i` indices are indices corresponding to the local `j` array.
4199: Once you have created the matrix you can update it with new numerical values using `MatUpdateMPIAIJWithArray()`
4201: If you do **not** use `MatUpdateMPIAIJWithArray()`, the column indices in `j` do not need to be sorted. If you will use
4202: `MatUpdateMPIAIJWithArrays()`, the column indices **must** be sorted.
4204: The format which is used for the sparse matrix input, is equivalent to a
4205: row-major ordering, i.e., for the following matrix, the input data expected is
4206: as shown
4207: .vb
4208: 1 0 0
4209: 2 0 3 P0
4210: -------
4211: 4 5 6 P1
4213: Process0 [P0] rows_owned=[0,1]
4214: i = {0,1,3} [size = nrow+1 = 2+1]
4215: j = {0,0,2} [size = 3]
4216: v = {1,2,3} [size = 3]
4218: Process1 [P1] rows_owned=[2]
4219: i = {0,3} [size = nrow+1 = 1+1]
4220: j = {0,1,2} [size = 3]
4221: v = {4,5,6} [size = 3]
4222: .ve
4224: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4225: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArray()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4226: @*/
4227: PetscErrorCode MatCreateMPIAIJWithArrays(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt M, PetscInt N, const PetscInt i[], const PetscInt j[], const PetscScalar a[], Mat *mat)
4228: {
4229: PetscFunctionBegin;
4230: PetscCheck(!i || !i[0], PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
4231: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
4232: PetscCall(MatCreate(comm, mat));
4233: PetscCall(MatSetSizes(*mat, m, n, M, N));
4234: /* PetscCall(MatSetBlockSizes(M,bs,cbs)); */
4235: PetscCall(MatSetType(*mat, MATMPIAIJ));
4236: PetscCall(MatMPIAIJSetPreallocationCSR(*mat, i, j, a));
4237: PetscFunctionReturn(PETSC_SUCCESS);
4238: }
4240: /*@
4241: MatUpdateMPIAIJWithArrays - updates a `MATMPIAIJ` matrix using arrays that contain in standard
4242: CSR format for the local rows. Only the numerical values are updated the other arrays must be identical to what was passed
4243: from `MatCreateMPIAIJWithArrays()`
4245: Deprecated: Use `MatUpdateMPIAIJWithArray()`
4247: Collective
4249: Input Parameters:
4250: + mat - the matrix
4251: . m - number of local rows (Cannot be `PETSC_DECIDE`)
4252: . n - This value should be the same as the local size used in creating the
4253: x vector for the matrix-vector product y = Ax. (or `PETSC_DECIDE` to have
4254: calculated if N is given) For square matrices n is almost always m.
4255: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if m is given)
4256: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if n is given)
4257: . Ii - row indices; that is Ii[0] = 0, Ii[row] = Ii[row-1] + number of elements in that row of the matrix
4258: . J - column indices
4259: - v - matrix values
4261: Level: deprecated
4263: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4264: `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArray()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4265: @*/
4266: PetscErrorCode MatUpdateMPIAIJWithArrays(Mat mat, PetscInt m, PetscInt n, PetscInt M, PetscInt N, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[])
4267: {
4268: PetscInt nnz, i;
4269: PetscBool nooffprocentries;
4270: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)mat->data;
4271: Mat_SeqAIJ *Ad = (Mat_SeqAIJ *)Aij->A->data;
4272: PetscScalar *ad, *ao;
4273: PetscInt ldi, Iii, md;
4274: const PetscInt *Adi = Ad->i;
4275: PetscInt *ld = Aij->ld;
4277: PetscFunctionBegin;
4278: PetscCheck(Ii[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
4279: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
4280: PetscCheck(m == mat->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Local number of rows cannot change from call to MatUpdateMPIAIJWithArrays()");
4281: PetscCheck(n == mat->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Local number of columns cannot change from call to MatUpdateMPIAIJWithArrays()");
4283: PetscCall(MatSeqAIJGetArrayWrite(Aij->A, &ad));
4284: PetscCall(MatSeqAIJGetArrayWrite(Aij->B, &ao));
4286: for (i = 0; i < m; i++) {
4287: if (PetscDefined(USE_DEBUG)) {
4288: for (PetscInt j = Ii[i] + 1; j < Ii[i + 1]; ++j) {
4289: 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);
4290: 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);
4291: }
4292: }
4293: nnz = Ii[i + 1] - Ii[i];
4294: Iii = Ii[i];
4295: ldi = ld[i];
4296: md = Adi[i + 1] - Adi[i];
4297: PetscCall(PetscArraycpy(ao, v + Iii, ldi));
4298: PetscCall(PetscArraycpy(ad, v + Iii + ldi, md));
4299: PetscCall(PetscArraycpy(ao + ldi, v + Iii + ldi + md, nnz - ldi - md));
4300: ad += md;
4301: ao += nnz - md;
4302: }
4303: nooffprocentries = mat->nooffprocentries;
4304: mat->nooffprocentries = PETSC_TRUE;
4305: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->A, &ad));
4306: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->B, &ao));
4307: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->A));
4308: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->B));
4309: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
4310: PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
4311: PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
4312: mat->nooffprocentries = nooffprocentries;
4313: PetscFunctionReturn(PETSC_SUCCESS);
4314: }
4316: /*@
4317: MatUpdateMPIAIJWithArray - updates an `MATMPIAIJ` matrix using an array that contains the nonzero values
4319: Collective
4321: Input Parameters:
4322: + mat - the matrix
4323: - v - matrix values, stored by row
4325: Level: intermediate
4327: Notes:
4328: The matrix must have been obtained with `MatCreateMPIAIJWithArrays()` or `MatMPIAIJSetPreallocationCSR()`
4330: The column indices in the call to `MatCreateMPIAIJWithArrays()` or `MatMPIAIJSetPreallocationCSR()` must have been sorted for this call to work correctly
4332: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4333: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArrays()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4334: @*/
4335: PetscErrorCode MatUpdateMPIAIJWithArray(Mat mat, const PetscScalar v[])
4336: {
4337: PetscInt nnz, i, m;
4338: PetscBool nooffprocentries;
4339: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)mat->data;
4340: Mat_SeqAIJ *Ad = (Mat_SeqAIJ *)Aij->A->data;
4341: Mat_SeqAIJ *Ao = (Mat_SeqAIJ *)Aij->B->data;
4342: PetscScalar *ad, *ao;
4343: const PetscInt *Adi = Ad->i, *Adj = Ao->i;
4344: PetscInt ldi, Iii, md;
4345: PetscInt *ld = Aij->ld;
4347: PetscFunctionBegin;
4348: m = mat->rmap->n;
4350: PetscCall(MatSeqAIJGetArrayWrite(Aij->A, &ad));
4351: PetscCall(MatSeqAIJGetArrayWrite(Aij->B, &ao));
4352: Iii = 0;
4353: for (i = 0; i < m; i++) {
4354: nnz = Adi[i + 1] - Adi[i] + Adj[i + 1] - Adj[i];
4355: ldi = ld[i];
4356: md = Adi[i + 1] - Adi[i];
4357: PetscCall(PetscArraycpy(ad, v + Iii + ldi, md));
4358: ad += md;
4359: if (ao) {
4360: PetscCall(PetscArraycpy(ao, v + Iii, ldi));
4361: PetscCall(PetscArraycpy(ao + ldi, v + Iii + ldi + md, nnz - ldi - md));
4362: ao += nnz - md;
4363: }
4364: Iii += nnz;
4365: }
4366: nooffprocentries = mat->nooffprocentries;
4367: mat->nooffprocentries = PETSC_TRUE;
4368: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->A, &ad));
4369: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->B, &ao));
4370: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->A));
4371: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->B));
4372: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
4373: PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
4374: PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
4375: mat->nooffprocentries = nooffprocentries;
4376: PetscFunctionReturn(PETSC_SUCCESS);
4377: }
4379: /*@
4380: MatCreateAIJ - Creates a sparse parallel matrix in `MATAIJ` format
4381: (the default parallel PETSc format). For good matrix assembly performance
4382: the user should preallocate the matrix storage by setting the parameters
4383: `d_nz` (or `d_nnz`) and `o_nz` (or `o_nnz`).
4385: Collective
4387: Input Parameters:
4388: + comm - MPI communicator
4389: . m - number of local rows (or `PETSC_DECIDE` to have calculated if M is given)
4390: This value should be the same as the local size used in creating the
4391: y vector for the matrix-vector product y = Ax.
4392: . n - This value should be the same as the local size used in creating the
4393: x vector for the matrix-vector product y = Ax. (or `PETSC_DECIDE` to have
4394: calculated if N is given) For square matrices n is almost always m.
4395: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if m is given)
4396: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if n is given)
4397: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
4398: (same value is used for all local rows)
4399: . d_nnz - array containing the number of nonzeros in the various rows of the
4400: DIAGONAL portion of the local submatrix (possibly different for each row)
4401: or `NULL`, if `d_nz` is used to specify the nonzero structure.
4402: The size of this array is equal to the number of local rows, i.e 'm'.
4403: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
4404: submatrix (same value is used for all local rows).
4405: - o_nnz - array containing the number of nonzeros in the various rows of the
4406: OFF-DIAGONAL portion of the local submatrix (possibly different for
4407: each row) or `NULL`, if `o_nz` is used to specify the nonzero
4408: structure. The size of this array is equal to the number
4409: of local rows, i.e 'm'.
4411: Output Parameter:
4412: . A - the matrix
4414: Options Database Keys:
4415: + -mat_no_inode - Do not use inodes
4416: . -mat_inode_limit <limit> - Sets inode limit (max limit=5)
4417: - -matmult_vecscatter_view <viewer> - View the vecscatter (i.e., communication pattern) used in `MatMult()` of sparse parallel matrices.
4418: See viewer types in manual of `MatView()`. Of them, ascii_matlab, draw or binary cause the `VecScatter`
4419: 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.
4421: Level: intermediate
4423: Notes:
4424: It is recommended that one use `MatCreateFromOptions()` or the `MatCreate()`, `MatSetType()` and/or `MatSetFromOptions()`,
4425: MatXXXXSetPreallocation() paradigm instead of this routine directly.
4426: [MatXXXXSetPreallocation() is, for example, `MatSeqAIJSetPreallocation()`]
4428: If the *_nnz parameter is given then the *_nz parameter is ignored
4430: The `m`,`n`,`M`,`N` parameters specify the size of the matrix, and its partitioning across
4431: processors, while `d_nz`,`d_nnz`,`o_nz`,`o_nnz` parameters specify the approximate
4432: storage requirements for this matrix.
4434: If `PETSC_DECIDE` or `PETSC_DETERMINE` is used for a particular argument on one
4435: processor than it must be used on all processors that share the object for
4436: that argument.
4438: If `m` and `n` are not `PETSC_DECIDE`, then the values determine the `PetscLayout` of the matrix and the ranges returned by
4439: `MatGetOwnershipRange()`, `MatGetOwnershipRanges()`, `MatGetOwnershipRangeColumn()`, and `MatGetOwnershipRangesColumn()`.
4441: The user MUST specify either the local or global matrix dimensions
4442: (possibly both).
4444: The parallel matrix is partitioned across processors such that the
4445: first `m0` rows belong to process 0, the next `m1` rows belong to
4446: process 1, the next `m2` rows belong to process 2, etc., where
4447: `m0`, `m1`, `m2`... are the input parameter `m` on each MPI process. I.e., each MPI process stores
4448: values corresponding to [m x N] submatrix.
4450: The columns are logically partitioned with the n0 columns belonging
4451: to 0th partition, the next n1 columns belonging to the next
4452: partition etc.. where n0,n1,n2... are the input parameter 'n'.
4454: The DIAGONAL portion of the local submatrix on any given processor
4455: is the submatrix corresponding to the rows and columns m,n
4456: corresponding to the given processor. i.e diagonal matrix on
4457: process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1]
4458: etc. The remaining portion of the local submatrix [m x (N-n)]
4459: constitute the OFF-DIAGONAL portion. The example below better
4460: illustrates this concept. The two matrices, the DIAGONAL portion and
4461: the OFF-DIAGONAL portion are each stored as `MATSEQAIJ` matrices.
4463: For a square global matrix we define each processor's diagonal portion
4464: to be its local rows and the corresponding columns (a square submatrix);
4465: each processor's off-diagonal portion encompasses the remainder of the
4466: local matrix (a rectangular submatrix).
4468: If `o_nnz`, `d_nnz` are specified, then `o_nz`, and `d_nz` are ignored.
4470: When calling this routine with a single process communicator, a matrix of
4471: type `MATSEQAIJ` is returned. If a matrix of type `MATMPIAIJ` is desired for this
4472: type of communicator, use the construction mechanism
4473: .vb
4474: MatCreate(..., &A);
4475: MatSetType(A, MATMPIAIJ);
4476: MatSetSizes(A, m, n, M, N);
4477: MatMPIAIJSetPreallocation(A, ...);
4478: .ve
4480: By default, this format uses inodes (identical nodes) when possible.
4481: We search for consecutive rows with the same nonzero structure, thereby
4482: reusing matrix information to achieve increased efficiency.
4484: Example Usage:
4485: Consider the following 8x8 matrix with 34 non-zero values, that is
4486: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4487: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4488: as follows
4490: .vb
4491: 1 2 0 | 0 3 0 | 0 4
4492: Proc0 0 5 6 | 7 0 0 | 8 0
4493: 9 0 10 | 11 0 0 | 12 0
4494: -------------------------------------
4495: 13 0 14 | 15 16 17 | 0 0
4496: Proc1 0 18 0 | 19 20 21 | 0 0
4497: 0 0 0 | 22 23 0 | 24 0
4498: -------------------------------------
4499: Proc2 25 26 27 | 0 0 28 | 29 0
4500: 30 0 0 | 31 32 33 | 0 34
4501: .ve
4503: This can be represented as a collection of submatrices as
4505: .vb
4506: A B C
4507: D E F
4508: G H I
4509: .ve
4511: Where the submatrices A,B,C are owned by proc0, D,E,F are
4512: owned by proc1, G,H,I are owned by proc2.
4514: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4515: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4516: The 'M','N' parameters are 8,8, and have the same values on all procs.
4518: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4519: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4520: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4521: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4522: part as `MATSEQAIJ` matrices. For example, proc1 will store [E] as a `MATSEQAIJ`
4523: matrix, and [DF] as another SeqAIJ matrix.
4525: When `d_nz`, `o_nz` parameters are specified, `d_nz` storage elements are
4526: allocated for every row of the local DIAGONAL submatrix, and `o_nz`
4527: storage locations are allocated for every row of the OFF-DIAGONAL submatrix.
4528: One way to choose `d_nz` and `o_nz` is to use the maximum number of nonzeros over
4529: the local rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4530: In this case, the values of `d_nz`,`o_nz` are
4531: .vb
4532: proc0 dnz = 2, o_nz = 2
4533: proc1 dnz = 3, o_nz = 2
4534: proc2 dnz = 1, o_nz = 4
4535: .ve
4536: We are allocating m*(`d_nz`+`o_nz`) storage locations for every proc. This
4537: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4538: for proc3. i.e we are using 12+15+10=37 storage locations to store
4539: 34 values.
4541: When `d_nnz`, `o_nnz` parameters are specified, the storage is specified
4542: for every row, corresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4543: In the above case the values for d_nnz,o_nnz are
4544: .vb
4545: proc0 d_nnz = [2,2,2] and o_nnz = [2,2,2]
4546: proc1 d_nnz = [3,3,2] and o_nnz = [2,1,1]
4547: proc2 d_nnz = [1,1] and o_nnz = [4,4]
4548: .ve
4549: Here the space allocated is sum of all the above values i.e 34, and
4550: hence pre-allocation is perfect.
4552: .seealso: [](ch_matrices), `Mat`, [Sparse Matrix Creation](sec_matsparse), `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4553: `MATMPIAIJ`, `MatCreateMPIAIJWithArrays()`, `MatGetOwnershipRange()`, `MatGetOwnershipRanges()`, `MatGetOwnershipRangeColumn()`,
4554: `MatGetOwnershipRangesColumn()`, `PetscLayout`
4555: @*/
4556: 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)
4557: {
4558: PetscMPIInt size;
4560: PetscFunctionBegin;
4561: PetscCall(MatCreate(comm, A));
4562: PetscCall(MatSetSizes(*A, m, n, M, N));
4563: PetscCallMPI(MPI_Comm_size(comm, &size));
4564: if (size > 1) {
4565: PetscCall(MatSetType(*A, MATMPIAIJ));
4566: PetscCall(MatMPIAIJSetPreallocation(*A, d_nz, d_nnz, o_nz, o_nnz));
4567: } else {
4568: PetscCall(MatSetType(*A, MATSEQAIJ));
4569: PetscCall(MatSeqAIJSetPreallocation(*A, d_nz, d_nnz));
4570: }
4571: PetscFunctionReturn(PETSC_SUCCESS);
4572: }
4574: /*MC
4575: MatMPIAIJGetSeqAIJF90 - Returns the local pieces of this distributed matrix
4577: Synopsis:
4578: MatMPIAIJGetSeqAIJF90(Mat A, Mat Ad, Mat Ao, {PetscInt, pointer :: colmap(:)},integer ierr)
4580: Not Collective
4582: Input Parameter:
4583: . A - the `MATMPIAIJ` matrix
4585: Output Parameters:
4586: + Ad - the diagonal portion of the matrix
4587: . Ao - the off-diagonal portion of the matrix
4588: . colmap - An array mapping local column numbers of `Ao` to global column numbers of the parallel matrix
4589: - ierr - error code
4591: Level: advanced
4593: Note:
4594: Use `MatMPIAIJRestoreSeqAIJF90()` when you no longer need access to the matrices and `colmap`
4596: .seealso: [](ch_matrices), `Mat`, [](sec_fortranarrays), `Mat`, `MATMPIAIJ`, `MatMPIAIJGetSeqAIJ()`, `MatMPIAIJRestoreSeqAIJF90()`
4597: M*/
4599: /*MC
4600: MatMPIAIJRestoreSeqAIJF90 - call after `MatMPIAIJGetSeqAIJF90()` when you no longer need access to the matrices and `colmap`
4602: Synopsis:
4603: MatMPIAIJRestoreSeqAIJF90(Mat A, Mat Ad, Mat Ao, {PetscInt, pointer :: colmap(:)},integer ierr)
4605: Not Collective
4607: Input Parameters:
4608: + A - the `MATMPIAIJ` matrix
4609: . Ad - the diagonal portion of the matrix
4610: . Ao - the off-diagonal portion of the matrix
4611: . colmap - An array mapping local column numbers of `Ao` to global column numbers of the parallel matrix
4612: - ierr - error code
4614: Level: advanced
4616: .seealso: [](ch_matrices), `Mat`, [](sec_fortranarrays), `Mat`, `MATMPIAIJ`, `MatMPIAIJGetSeqAIJ()`, `MatMPIAIJGetSeqAIJF90()`
4617: M*/
4619: /*@C
4620: MatMPIAIJGetSeqAIJ - Returns the local pieces of this distributed matrix
4622: Not Collective
4624: Input Parameter:
4625: . A - The `MATMPIAIJ` matrix
4627: Output Parameters:
4628: + Ad - The local diagonal block as a `MATSEQAIJ` matrix
4629: . Ao - The local off-diagonal block as a `MATSEQAIJ` matrix
4630: - colmap - An array mapping local column numbers of `Ao` to global column numbers of the parallel matrix
4632: Level: intermediate
4634: Note:
4635: The rows in `Ad` and `Ao` are in [0, Nr), where Nr is the number of local rows on this process. The columns
4636: 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
4637: the number of nonzero columns in the local off-diagonal piece of the matrix `A`. The array colmap maps these
4638: local column numbers to global column numbers in the original matrix.
4640: Fortran Notes:
4641: `MatMPIAIJGetSeqAIJ()` Fortran binding is deprecated (since PETSc 3.19), use `MatMPIAIJGetSeqAIJF90()`
4643: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatMPIAIJGetSeqAIJF90()`, `MatMPIAIJRestoreSeqAIJF90()`, `MatMPIAIJGetLocalMat()`, `MatMPIAIJGetLocalMatCondensed()`, `MatCreateAIJ()`, `MATSEQAIJ`
4644: @*/
4645: PetscErrorCode MatMPIAIJGetSeqAIJ(Mat A, Mat *Ad, Mat *Ao, const PetscInt *colmap[])
4646: {
4647: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
4648: PetscBool flg;
4650: PetscFunctionBegin;
4651: PetscCall(PetscStrbeginswith(((PetscObject)A)->type_name, MATMPIAIJ, &flg));
4652: PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "This function requires a MATMPIAIJ matrix as input");
4653: if (Ad) *Ad = a->A;
4654: if (Ao) *Ao = a->B;
4655: if (colmap) *colmap = a->garray;
4656: PetscFunctionReturn(PETSC_SUCCESS);
4657: }
4659: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm comm, Mat inmat, PetscInt n, MatReuse scall, Mat *outmat)
4660: {
4661: PetscInt m, N, i, rstart, nnz, Ii;
4662: PetscInt *indx;
4663: PetscScalar *values;
4664: MatType rootType;
4666: PetscFunctionBegin;
4667: PetscCall(MatGetSize(inmat, &m, &N));
4668: if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */
4669: PetscInt *dnz, *onz, sum, bs, cbs;
4671: if (n == PETSC_DECIDE) PetscCall(PetscSplitOwnership(comm, &n, &N));
4672: /* Check sum(n) = N */
4673: PetscCallMPI(MPIU_Allreduce(&n, &sum, 1, MPIU_INT, MPI_SUM, comm));
4674: PetscCheck(sum == N, PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Sum of local columns %" PetscInt_FMT " != global columns %" PetscInt_FMT, sum, N);
4676: PetscCallMPI(MPI_Scan(&m, &rstart, 1, MPIU_INT, MPI_SUM, comm));
4677: rstart -= m;
4679: MatPreallocateBegin(comm, m, n, dnz, onz);
4680: for (i = 0; i < m; i++) {
4681: PetscCall(MatGetRow_SeqAIJ(inmat, i, &nnz, &indx, NULL));
4682: PetscCall(MatPreallocateSet(i + rstart, nnz, indx, dnz, onz));
4683: PetscCall(MatRestoreRow_SeqAIJ(inmat, i, &nnz, &indx, NULL));
4684: }
4686: PetscCall(MatCreate(comm, outmat));
4687: PetscCall(MatSetSizes(*outmat, m, n, PETSC_DETERMINE, PETSC_DETERMINE));
4688: PetscCall(MatGetBlockSizes(inmat, &bs, &cbs));
4689: PetscCall(MatSetBlockSizes(*outmat, bs, cbs));
4690: PetscCall(MatGetRootType_Private(inmat, &rootType));
4691: PetscCall(MatSetType(*outmat, rootType));
4692: PetscCall(MatSeqAIJSetPreallocation(*outmat, 0, dnz));
4693: PetscCall(MatMPIAIJSetPreallocation(*outmat, 0, dnz, 0, onz));
4694: MatPreallocateEnd(dnz, onz);
4695: PetscCall(MatSetOption(*outmat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
4696: }
4698: /* numeric phase */
4699: PetscCall(MatGetOwnershipRange(*outmat, &rstart, NULL));
4700: for (i = 0; i < m; i++) {
4701: PetscCall(MatGetRow_SeqAIJ(inmat, i, &nnz, &indx, &values));
4702: Ii = i + rstart;
4703: PetscCall(MatSetValues(*outmat, 1, &Ii, nnz, indx, values, INSERT_VALUES));
4704: PetscCall(MatRestoreRow_SeqAIJ(inmat, i, &nnz, &indx, &values));
4705: }
4706: PetscCall(MatAssemblyBegin(*outmat, MAT_FINAL_ASSEMBLY));
4707: PetscCall(MatAssemblyEnd(*outmat, MAT_FINAL_ASSEMBLY));
4708: PetscFunctionReturn(PETSC_SUCCESS);
4709: }
4711: static PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(void **data)
4712: {
4713: Mat_Merge_SeqsToMPI *merge = (Mat_Merge_SeqsToMPI *)*data;
4715: PetscFunctionBegin;
4716: if (!merge) PetscFunctionReturn(PETSC_SUCCESS);
4717: PetscCall(PetscFree(merge->id_r));
4718: PetscCall(PetscFree(merge->len_s));
4719: PetscCall(PetscFree(merge->len_r));
4720: PetscCall(PetscFree(merge->bi));
4721: PetscCall(PetscFree(merge->bj));
4722: PetscCall(PetscFree(merge->buf_ri[0]));
4723: PetscCall(PetscFree(merge->buf_ri));
4724: PetscCall(PetscFree(merge->buf_rj[0]));
4725: PetscCall(PetscFree(merge->buf_rj));
4726: PetscCall(PetscFree(merge->coi));
4727: PetscCall(PetscFree(merge->coj));
4728: PetscCall(PetscFree(merge->owners_co));
4729: PetscCall(PetscLayoutDestroy(&merge->rowmap));
4730: PetscCall(PetscFree(merge));
4731: PetscFunctionReturn(PETSC_SUCCESS);
4732: }
4734: #include <../src/mat/utils/freespace.h>
4735: #include <petscbt.h>
4737: PetscErrorCode MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat, Mat mpimat)
4738: {
4739: MPI_Comm comm;
4740: Mat_SeqAIJ *a = (Mat_SeqAIJ *)seqmat->data;
4741: PetscMPIInt size, rank, taga, *len_s;
4742: PetscInt N = mpimat->cmap->N, i, j, *owners, *ai = a->i, *aj, m;
4743: PetscMPIInt proc, k;
4744: PetscInt **buf_ri, **buf_rj;
4745: PetscInt anzi, *bj_i, *bi, *bj, arow, bnzi, nextaj;
4746: PetscInt nrows, **buf_ri_k, **nextrow, **nextai;
4747: MPI_Request *s_waits, *r_waits;
4748: MPI_Status *status;
4749: const MatScalar *aa, *a_a;
4750: MatScalar **abuf_r, *ba_i;
4751: Mat_Merge_SeqsToMPI *merge;
4752: PetscContainer container;
4754: PetscFunctionBegin;
4755: PetscCall(PetscObjectGetComm((PetscObject)mpimat, &comm));
4756: PetscCall(PetscLogEventBegin(MAT_Seqstompinum, seqmat, 0, 0, 0));
4758: PetscCallMPI(MPI_Comm_size(comm, &size));
4759: PetscCallMPI(MPI_Comm_rank(comm, &rank));
4761: PetscCall(PetscObjectQuery((PetscObject)mpimat, "MatMergeSeqsToMPI", (PetscObject *)&container));
4762: PetscCheck(container, PetscObjectComm((PetscObject)mpimat), PETSC_ERR_PLIB, "Mat not created from MatCreateMPIAIJSumSeqAIJSymbolic");
4763: PetscCall(PetscContainerGetPointer(container, (void **)&merge));
4764: PetscCall(MatSeqAIJGetArrayRead(seqmat, &a_a));
4765: aa = a_a;
4767: bi = merge->bi;
4768: bj = merge->bj;
4769: buf_ri = merge->buf_ri;
4770: buf_rj = merge->buf_rj;
4772: PetscCall(PetscMalloc1(size, &status));
4773: owners = merge->rowmap->range;
4774: len_s = merge->len_s;
4776: /* send and recv matrix values */
4777: PetscCall(PetscObjectGetNewTag((PetscObject)mpimat, &taga));
4778: PetscCall(PetscPostIrecvScalar(comm, taga, merge->nrecv, merge->id_r, merge->len_r, &abuf_r, &r_waits));
4780: PetscCall(PetscMalloc1(merge->nsend + 1, &s_waits));
4781: for (proc = 0, k = 0; proc < size; proc++) {
4782: if (!len_s[proc]) continue;
4783: i = owners[proc];
4784: PetscCallMPI(MPIU_Isend(aa + ai[i], len_s[proc], MPIU_MATSCALAR, proc, taga, comm, s_waits + k));
4785: k++;
4786: }
4788: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, r_waits, status));
4789: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, s_waits, status));
4790: PetscCall(PetscFree(status));
4792: PetscCall(PetscFree(s_waits));
4793: PetscCall(PetscFree(r_waits));
4795: /* insert mat values of mpimat */
4796: PetscCall(PetscMalloc1(N, &ba_i));
4797: PetscCall(PetscMalloc3(merge->nrecv, &buf_ri_k, merge->nrecv, &nextrow, merge->nrecv, &nextai));
4799: for (k = 0; k < merge->nrecv; k++) {
4800: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4801: nrows = *buf_ri_k[k];
4802: nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */
4803: nextai[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure */
4804: }
4806: /* set values of ba */
4807: m = merge->rowmap->n;
4808: for (i = 0; i < m; i++) {
4809: arow = owners[rank] + i;
4810: bj_i = bj + bi[i]; /* col indices of the i-th row of mpimat */
4811: bnzi = bi[i + 1] - bi[i];
4812: PetscCall(PetscArrayzero(ba_i, bnzi));
4814: /* add local non-zero vals of this proc's seqmat into ba */
4815: anzi = ai[arow + 1] - ai[arow];
4816: aj = a->j + ai[arow];
4817: aa = a_a + ai[arow];
4818: nextaj = 0;
4819: for (j = 0; nextaj < anzi; j++) {
4820: if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4821: ba_i[j] += aa[nextaj++];
4822: }
4823: }
4825: /* add received vals into ba */
4826: for (k = 0; k < merge->nrecv; k++) { /* k-th received message */
4827: /* i-th row */
4828: if (i == *nextrow[k]) {
4829: anzi = *(nextai[k] + 1) - *nextai[k];
4830: aj = buf_rj[k] + *nextai[k];
4831: aa = abuf_r[k] + *nextai[k];
4832: nextaj = 0;
4833: for (j = 0; nextaj < anzi; j++) {
4834: if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4835: ba_i[j] += aa[nextaj++];
4836: }
4837: }
4838: nextrow[k]++;
4839: nextai[k]++;
4840: }
4841: }
4842: PetscCall(MatSetValues(mpimat, 1, &arow, bnzi, bj_i, ba_i, INSERT_VALUES));
4843: }
4844: PetscCall(MatSeqAIJRestoreArrayRead(seqmat, &a_a));
4845: PetscCall(MatAssemblyBegin(mpimat, MAT_FINAL_ASSEMBLY));
4846: PetscCall(MatAssemblyEnd(mpimat, MAT_FINAL_ASSEMBLY));
4848: PetscCall(PetscFree(abuf_r[0]));
4849: PetscCall(PetscFree(abuf_r));
4850: PetscCall(PetscFree(ba_i));
4851: PetscCall(PetscFree3(buf_ri_k, nextrow, nextai));
4852: PetscCall(PetscLogEventEnd(MAT_Seqstompinum, seqmat, 0, 0, 0));
4853: PetscFunctionReturn(PETSC_SUCCESS);
4854: }
4856: PetscErrorCode MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm, Mat seqmat, PetscInt m, PetscInt n, Mat *mpimat)
4857: {
4858: Mat B_mpi;
4859: Mat_SeqAIJ *a = (Mat_SeqAIJ *)seqmat->data;
4860: PetscMPIInt size, rank, tagi, tagj, *len_s, *len_si, *len_ri;
4861: PetscInt **buf_rj, **buf_ri, **buf_ri_k;
4862: PetscInt M = seqmat->rmap->n, N = seqmat->cmap->n, i, *owners, *ai = a->i, *aj = a->j;
4863: PetscInt len, *dnz, *onz, bs, cbs;
4864: PetscInt k, anzi, *bi, *bj, *lnk, nlnk, arow, bnzi;
4865: PetscInt nrows, *buf_s, *buf_si, *buf_si_i, **nextrow, **nextai;
4866: MPI_Request *si_waits, *sj_waits, *ri_waits, *rj_waits;
4867: MPI_Status *status;
4868: PetscFreeSpaceList free_space = NULL, current_space = NULL;
4869: PetscBT lnkbt;
4870: Mat_Merge_SeqsToMPI *merge;
4871: PetscContainer container;
4873: PetscFunctionBegin;
4874: PetscCall(PetscLogEventBegin(MAT_Seqstompisym, seqmat, 0, 0, 0));
4876: /* make sure it is a PETSc comm */
4877: PetscCall(PetscCommDuplicate(comm, &comm, NULL));
4878: PetscCallMPI(MPI_Comm_size(comm, &size));
4879: PetscCallMPI(MPI_Comm_rank(comm, &rank));
4881: PetscCall(PetscNew(&merge));
4882: PetscCall(PetscMalloc1(size, &status));
4884: /* determine row ownership */
4885: PetscCall(PetscLayoutCreate(comm, &merge->rowmap));
4886: PetscCall(PetscLayoutSetLocalSize(merge->rowmap, m));
4887: PetscCall(PetscLayoutSetSize(merge->rowmap, M));
4888: PetscCall(PetscLayoutSetBlockSize(merge->rowmap, 1));
4889: PetscCall(PetscLayoutSetUp(merge->rowmap));
4890: PetscCall(PetscMalloc1(size, &len_si));
4891: PetscCall(PetscMalloc1(size, &merge->len_s));
4893: m = merge->rowmap->n;
4894: owners = merge->rowmap->range;
4896: /* determine the number of messages to send, their lengths */
4897: len_s = merge->len_s;
4899: len = 0; /* length of buf_si[] */
4900: merge->nsend = 0;
4901: for (PetscMPIInt proc = 0; proc < size; proc++) {
4902: len_si[proc] = 0;
4903: if (proc == rank) {
4904: len_s[proc] = 0;
4905: } else {
4906: PetscCall(PetscMPIIntCast(owners[proc + 1] - owners[proc] + 1, &len_si[proc]));
4907: PetscCall(PetscMPIIntCast(ai[owners[proc + 1]] - ai[owners[proc]], &len_s[proc])); /* num of rows to be sent to [proc] */
4908: }
4909: if (len_s[proc]) {
4910: merge->nsend++;
4911: nrows = 0;
4912: for (i = owners[proc]; i < owners[proc + 1]; i++) {
4913: if (ai[i + 1] > ai[i]) nrows++;
4914: }
4915: PetscCall(PetscMPIIntCast(2 * (nrows + 1), &len_si[proc]));
4916: len += len_si[proc];
4917: }
4918: }
4920: /* determine the number and length of messages to receive for ij-structure */
4921: PetscCall(PetscGatherNumberOfMessages(comm, NULL, len_s, &merge->nrecv));
4922: PetscCall(PetscGatherMessageLengths2(comm, merge->nsend, merge->nrecv, len_s, len_si, &merge->id_r, &merge->len_r, &len_ri));
4924: /* post the Irecv of j-structure */
4925: PetscCall(PetscCommGetNewTag(comm, &tagj));
4926: PetscCall(PetscPostIrecvInt(comm, tagj, merge->nrecv, merge->id_r, merge->len_r, &buf_rj, &rj_waits));
4928: /* post the Isend of j-structure */
4929: PetscCall(PetscMalloc2(merge->nsend, &si_waits, merge->nsend, &sj_waits));
4931: for (PetscMPIInt proc = 0, k = 0; proc < size; proc++) {
4932: if (!len_s[proc]) continue;
4933: i = owners[proc];
4934: PetscCallMPI(MPIU_Isend(aj + ai[i], len_s[proc], MPIU_INT, proc, tagj, comm, sj_waits + k));
4935: k++;
4936: }
4938: /* receives and sends of j-structure are complete */
4939: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, rj_waits, status));
4940: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, sj_waits, status));
4942: /* send and recv i-structure */
4943: PetscCall(PetscCommGetNewTag(comm, &tagi));
4944: PetscCall(PetscPostIrecvInt(comm, tagi, merge->nrecv, merge->id_r, len_ri, &buf_ri, &ri_waits));
4946: PetscCall(PetscMalloc1(len + 1, &buf_s));
4947: buf_si = buf_s; /* points to the beginning of k-th msg to be sent */
4948: for (PetscMPIInt proc = 0, k = 0; proc < size; proc++) {
4949: if (!len_s[proc]) continue;
4950: /* form outgoing message for i-structure:
4951: buf_si[0]: nrows to be sent
4952: [1:nrows]: row index (global)
4953: [nrows+1:2*nrows+1]: i-structure index
4954: */
4955: nrows = len_si[proc] / 2 - 1;
4956: buf_si_i = buf_si + nrows + 1;
4957: buf_si[0] = nrows;
4958: buf_si_i[0] = 0;
4959: nrows = 0;
4960: for (i = owners[proc]; i < owners[proc + 1]; i++) {
4961: anzi = ai[i + 1] - ai[i];
4962: if (anzi) {
4963: buf_si_i[nrows + 1] = buf_si_i[nrows] + anzi; /* i-structure */
4964: buf_si[nrows + 1] = i - owners[proc]; /* local row index */
4965: nrows++;
4966: }
4967: }
4968: PetscCallMPI(MPIU_Isend(buf_si, len_si[proc], MPIU_INT, proc, tagi, comm, si_waits + k));
4969: k++;
4970: buf_si += len_si[proc];
4971: }
4973: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, ri_waits, status));
4974: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, si_waits, status));
4976: PetscCall(PetscInfo(seqmat, "nsend: %d, nrecv: %d\n", merge->nsend, merge->nrecv));
4977: 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]));
4979: PetscCall(PetscFree(len_si));
4980: PetscCall(PetscFree(len_ri));
4981: PetscCall(PetscFree(rj_waits));
4982: PetscCall(PetscFree2(si_waits, sj_waits));
4983: PetscCall(PetscFree(ri_waits));
4984: PetscCall(PetscFree(buf_s));
4985: PetscCall(PetscFree(status));
4987: /* compute a local seq matrix in each processor */
4988: /* allocate bi array and free space for accumulating nonzero column info */
4989: PetscCall(PetscMalloc1(m + 1, &bi));
4990: bi[0] = 0;
4992: /* create and initialize a linked list */
4993: nlnk = N + 1;
4994: PetscCall(PetscLLCreate(N, N, nlnk, lnk, lnkbt));
4996: /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */
4997: len = ai[owners[rank + 1]] - ai[owners[rank]];
4998: PetscCall(PetscFreeSpaceGet(PetscIntMultTruncate(2, len) + 1, &free_space));
5000: current_space = free_space;
5002: /* determine symbolic info for each local row */
5003: PetscCall(PetscMalloc3(merge->nrecv, &buf_ri_k, merge->nrecv, &nextrow, merge->nrecv, &nextai));
5005: for (k = 0; k < merge->nrecv; k++) {
5006: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
5007: nrows = *buf_ri_k[k];
5008: nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */
5009: nextai[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure */
5010: }
5012: MatPreallocateBegin(comm, m, n, dnz, onz);
5013: len = 0;
5014: for (i = 0; i < m; i++) {
5015: bnzi = 0;
5016: /* add local non-zero cols of this proc's seqmat into lnk */
5017: arow = owners[rank] + i;
5018: anzi = ai[arow + 1] - ai[arow];
5019: aj = a->j + ai[arow];
5020: PetscCall(PetscLLAddSorted(anzi, aj, N, &nlnk, lnk, lnkbt));
5021: bnzi += nlnk;
5022: /* add received col data into lnk */
5023: for (k = 0; k < merge->nrecv; k++) { /* k-th received message */
5024: if (i == *nextrow[k]) { /* i-th row */
5025: anzi = *(nextai[k] + 1) - *nextai[k];
5026: aj = buf_rj[k] + *nextai[k];
5027: PetscCall(PetscLLAddSorted(anzi, aj, N, &nlnk, lnk, lnkbt));
5028: bnzi += nlnk;
5029: nextrow[k]++;
5030: nextai[k]++;
5031: }
5032: }
5033: if (len < bnzi) len = bnzi; /* =max(bnzi) */
5035: /* if free space is not available, make more free space */
5036: if (current_space->local_remaining < bnzi) PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(bnzi, current_space->total_array_size), ¤t_space));
5037: /* copy data into free space, then initialize lnk */
5038: PetscCall(PetscLLClean(N, N, bnzi, lnk, current_space->array, lnkbt));
5039: PetscCall(MatPreallocateSet(i + owners[rank], bnzi, current_space->array, dnz, onz));
5041: current_space->array += bnzi;
5042: current_space->local_used += bnzi;
5043: current_space->local_remaining -= bnzi;
5045: bi[i + 1] = bi[i] + bnzi;
5046: }
5048: PetscCall(PetscFree3(buf_ri_k, nextrow, nextai));
5050: PetscCall(PetscMalloc1(bi[m] + 1, &bj));
5051: PetscCall(PetscFreeSpaceContiguous(&free_space, bj));
5052: PetscCall(PetscLLDestroy(lnk, lnkbt));
5054: /* create symbolic parallel matrix B_mpi */
5055: PetscCall(MatGetBlockSizes(seqmat, &bs, &cbs));
5056: PetscCall(MatCreate(comm, &B_mpi));
5057: if (n == PETSC_DECIDE) {
5058: PetscCall(MatSetSizes(B_mpi, m, n, PETSC_DETERMINE, N));
5059: } else {
5060: PetscCall(MatSetSizes(B_mpi, m, n, PETSC_DETERMINE, PETSC_DETERMINE));
5061: }
5062: PetscCall(MatSetBlockSizes(B_mpi, bs, cbs));
5063: PetscCall(MatSetType(B_mpi, MATMPIAIJ));
5064: PetscCall(MatMPIAIJSetPreallocation(B_mpi, 0, dnz, 0, onz));
5065: MatPreallocateEnd(dnz, onz);
5066: PetscCall(MatSetOption(B_mpi, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_FALSE));
5068: /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */
5069: B_mpi->assembled = PETSC_FALSE;
5070: merge->bi = bi;
5071: merge->bj = bj;
5072: merge->buf_ri = buf_ri;
5073: merge->buf_rj = buf_rj;
5074: merge->coi = NULL;
5075: merge->coj = NULL;
5076: merge->owners_co = NULL;
5078: PetscCall(PetscCommDestroy(&comm));
5080: /* attach the supporting struct to B_mpi for reuse */
5081: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
5082: PetscCall(PetscContainerSetPointer(container, merge));
5083: PetscCall(PetscContainerSetCtxDestroy(container, MatDestroy_MPIAIJ_SeqsToMPI));
5084: PetscCall(PetscObjectCompose((PetscObject)B_mpi, "MatMergeSeqsToMPI", (PetscObject)container));
5085: PetscCall(PetscContainerDestroy(&container));
5086: *mpimat = B_mpi;
5088: PetscCall(PetscLogEventEnd(MAT_Seqstompisym, seqmat, 0, 0, 0));
5089: PetscFunctionReturn(PETSC_SUCCESS);
5090: }
5092: /*@
5093: MatCreateMPIAIJSumSeqAIJ - Creates a `MATMPIAIJ` matrix by adding sequential
5094: matrices from each processor
5096: Collective
5098: Input Parameters:
5099: + comm - the communicators the parallel matrix will live on
5100: . seqmat - the input sequential matrices
5101: . m - number of local rows (or `PETSC_DECIDE`)
5102: . n - number of local columns (or `PETSC_DECIDE`)
5103: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5105: Output Parameter:
5106: . mpimat - the parallel matrix generated
5108: Level: advanced
5110: Note:
5111: The dimensions of the sequential matrix in each processor MUST be the same.
5112: The input seqmat is included into the container "Mat_Merge_SeqsToMPI", and will be
5113: destroyed when `mpimat` is destroyed. Call `PetscObjectQuery()` to access `seqmat`.
5115: .seealso: [](ch_matrices), `Mat`, `MatCreateAIJ()`
5116: @*/
5117: PetscErrorCode MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm, Mat seqmat, PetscInt m, PetscInt n, MatReuse scall, Mat *mpimat)
5118: {
5119: PetscMPIInt size;
5121: PetscFunctionBegin;
5122: PetscCallMPI(MPI_Comm_size(comm, &size));
5123: if (size == 1) {
5124: PetscCall(PetscLogEventBegin(MAT_Seqstompi, seqmat, 0, 0, 0));
5125: if (scall == MAT_INITIAL_MATRIX) {
5126: PetscCall(MatDuplicate(seqmat, MAT_COPY_VALUES, mpimat));
5127: } else {
5128: PetscCall(MatCopy(seqmat, *mpimat, SAME_NONZERO_PATTERN));
5129: }
5130: PetscCall(PetscLogEventEnd(MAT_Seqstompi, seqmat, 0, 0, 0));
5131: PetscFunctionReturn(PETSC_SUCCESS);
5132: }
5133: PetscCall(PetscLogEventBegin(MAT_Seqstompi, seqmat, 0, 0, 0));
5134: if (scall == MAT_INITIAL_MATRIX) PetscCall(MatCreateMPIAIJSumSeqAIJSymbolic(comm, seqmat, m, n, mpimat));
5135: PetscCall(MatCreateMPIAIJSumSeqAIJNumeric(seqmat, *mpimat));
5136: PetscCall(PetscLogEventEnd(MAT_Seqstompi, seqmat, 0, 0, 0));
5137: PetscFunctionReturn(PETSC_SUCCESS);
5138: }
5140: /*@
5141: MatAIJGetLocalMat - Creates a `MATSEQAIJ` from a `MATAIJ` matrix.
5143: Not Collective
5145: Input Parameter:
5146: . A - the matrix
5148: Output Parameter:
5149: . A_loc - the local sequential matrix generated
5151: Level: developer
5153: Notes:
5154: The matrix is created by taking `A`'s local rows and putting them into a sequential matrix
5155: with `mlocal` rows and `n` columns. Where `mlocal` is obtained with `MatGetLocalSize()` and
5156: `n` is the global column count obtained with `MatGetSize()`
5158: In other words combines the two parts of a parallel `MATMPIAIJ` matrix on each process to a single matrix.
5160: For parallel matrices this creates an entirely new matrix. If the matrix is sequential it merely increases the reference count.
5162: Destroy the matrix with `MatDestroy()`
5164: .seealso: [](ch_matrices), `Mat`, `MatMPIAIJGetLocalMat()`
5165: @*/
5166: PetscErrorCode MatAIJGetLocalMat(Mat A, Mat *A_loc)
5167: {
5168: PetscBool mpi;
5170: PetscFunctionBegin;
5171: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIAIJ, &mpi));
5172: if (mpi) {
5173: PetscCall(MatMPIAIJGetLocalMat(A, MAT_INITIAL_MATRIX, A_loc));
5174: } else {
5175: *A_loc = A;
5176: PetscCall(PetscObjectReference((PetscObject)*A_loc));
5177: }
5178: PetscFunctionReturn(PETSC_SUCCESS);
5179: }
5181: /*@
5182: MatMPIAIJGetLocalMat - Creates a `MATSEQAIJ` from a `MATMPIAIJ` matrix.
5184: Not Collective
5186: Input Parameters:
5187: + A - the matrix
5188: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5190: Output Parameter:
5191: . A_loc - the local sequential matrix generated
5193: Level: developer
5195: Notes:
5196: The matrix is created by taking all `A`'s local rows and putting them into a sequential
5197: matrix with `mlocal` rows and `n` columns.`mlocal` is the row count obtained with
5198: `MatGetLocalSize()` and `n` is the global column count obtained with `MatGetSize()`.
5200: In other words combines the two parts of a parallel `MATMPIAIJ` matrix on each process to a single matrix.
5202: When `A` is sequential and `MAT_INITIAL_MATRIX` is requested, the matrix returned is the diagonal part of `A` (which contains the entire matrix),
5203: 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
5204: then `MatCopy`(Adiag,*`A_loc`,`SAME_NONZERO_PATTERN`) is called to fill `A_loc`. Thus one can preallocate the appropriate sequential matrix `A_loc`
5205: 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.
5207: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMatCondensed()`, `MatMPIAIJGetLocalMatMerge()`
5208: @*/
5209: PetscErrorCode MatMPIAIJGetLocalMat(Mat A, MatReuse scall, Mat *A_loc)
5210: {
5211: Mat_MPIAIJ *mpimat = (Mat_MPIAIJ *)A->data;
5212: Mat_SeqAIJ *mat, *a, *b;
5213: PetscInt *ai, *aj, *bi, *bj, *cmap = mpimat->garray;
5214: const PetscScalar *aa, *ba, *aav, *bav;
5215: PetscScalar *ca, *cam;
5216: PetscMPIInt size;
5217: PetscInt am = A->rmap->n, i, j, k, cstart = A->cmap->rstart;
5218: PetscInt *ci, *cj, col, ncols_d, ncols_o, jo;
5219: PetscBool match;
5221: PetscFunctionBegin;
5222: PetscCall(PetscStrbeginswith(((PetscObject)A)->type_name, MATMPIAIJ, &match));
5223: PetscCheck(match, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Requires MATMPIAIJ matrix as input");
5224: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
5225: if (size == 1) {
5226: if (scall == MAT_INITIAL_MATRIX) {
5227: PetscCall(PetscObjectReference((PetscObject)mpimat->A));
5228: *A_loc = mpimat->A;
5229: } else if (scall == MAT_REUSE_MATRIX) {
5230: PetscCall(MatCopy(mpimat->A, *A_loc, SAME_NONZERO_PATTERN));
5231: }
5232: PetscFunctionReturn(PETSC_SUCCESS);
5233: }
5235: PetscCall(PetscLogEventBegin(MAT_Getlocalmat, A, 0, 0, 0));
5236: a = (Mat_SeqAIJ *)mpimat->A->data;
5237: b = (Mat_SeqAIJ *)mpimat->B->data;
5238: ai = a->i;
5239: aj = a->j;
5240: bi = b->i;
5241: bj = b->j;
5242: PetscCall(MatSeqAIJGetArrayRead(mpimat->A, &aav));
5243: PetscCall(MatSeqAIJGetArrayRead(mpimat->B, &bav));
5244: aa = aav;
5245: ba = bav;
5246: if (scall == MAT_INITIAL_MATRIX) {
5247: PetscCall(PetscMalloc1(1 + am, &ci));
5248: ci[0] = 0;
5249: for (i = 0; i < am; i++) ci[i + 1] = ci[i] + (ai[i + 1] - ai[i]) + (bi[i + 1] - bi[i]);
5250: PetscCall(PetscMalloc1(1 + ci[am], &cj));
5251: PetscCall(PetscMalloc1(1 + ci[am], &ca));
5252: k = 0;
5253: for (i = 0; i < am; i++) {
5254: ncols_o = bi[i + 1] - bi[i];
5255: ncols_d = ai[i + 1] - ai[i];
5256: /* off-diagonal portion of A */
5257: for (jo = 0; jo < ncols_o; jo++) {
5258: col = cmap[*bj];
5259: if (col >= cstart) break;
5260: cj[k] = col;
5261: bj++;
5262: ca[k++] = *ba++;
5263: }
5264: /* diagonal portion of A */
5265: for (j = 0; j < ncols_d; j++) {
5266: cj[k] = cstart + *aj++;
5267: ca[k++] = *aa++;
5268: }
5269: /* off-diagonal portion of A */
5270: for (j = jo; j < ncols_o; j++) {
5271: cj[k] = cmap[*bj++];
5272: ca[k++] = *ba++;
5273: }
5274: }
5275: /* put together the new matrix */
5276: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, am, A->cmap->N, ci, cj, ca, A_loc));
5277: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5278: /* Since these are PETSc arrays, change flags to free them as necessary. */
5279: mat = (Mat_SeqAIJ *)(*A_loc)->data;
5280: mat->free_a = PETSC_TRUE;
5281: mat->free_ij = PETSC_TRUE;
5282: mat->nonew = 0;
5283: } else if (scall == MAT_REUSE_MATRIX) {
5284: mat = (Mat_SeqAIJ *)(*A_loc)->data;
5285: ci = mat->i;
5286: cj = mat->j;
5287: PetscCall(MatSeqAIJGetArrayWrite(*A_loc, &cam));
5288: for (i = 0; i < am; i++) {
5289: /* off-diagonal portion of A */
5290: ncols_o = bi[i + 1] - bi[i];
5291: for (jo = 0; jo < ncols_o; jo++) {
5292: col = cmap[*bj];
5293: if (col >= cstart) break;
5294: *cam++ = *ba++;
5295: bj++;
5296: }
5297: /* diagonal portion of A */
5298: ncols_d = ai[i + 1] - ai[i];
5299: for (j = 0; j < ncols_d; j++) *cam++ = *aa++;
5300: /* off-diagonal portion of A */
5301: for (j = jo; j < ncols_o; j++) {
5302: *cam++ = *ba++;
5303: bj++;
5304: }
5305: }
5306: PetscCall(MatSeqAIJRestoreArrayWrite(*A_loc, &cam));
5307: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Invalid MatReuse %d", (int)scall);
5308: PetscCall(MatSeqAIJRestoreArrayRead(mpimat->A, &aav));
5309: PetscCall(MatSeqAIJRestoreArrayRead(mpimat->B, &bav));
5310: PetscCall(PetscLogEventEnd(MAT_Getlocalmat, A, 0, 0, 0));
5311: PetscFunctionReturn(PETSC_SUCCESS);
5312: }
5314: /*@
5315: MatMPIAIJGetLocalMatMerge - Creates a `MATSEQAIJ` from a `MATMPIAIJ` matrix by taking all its local rows and putting them into a sequential matrix with
5316: mlocal rows and n columns. Where n is the sum of the number of columns of the diagonal and off-diagonal part
5318: Not Collective
5320: Input Parameters:
5321: + A - the matrix
5322: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5324: Output Parameters:
5325: + glob - sequential `IS` with global indices associated with the columns of the local sequential matrix generated (can be `NULL`)
5326: - A_loc - the local sequential matrix generated
5328: Level: developer
5330: Note:
5331: This is different from `MatMPIAIJGetLocalMat()` since the first columns in the returning matrix are those associated with the diagonal
5332: part, then those associated with the off-diagonal part (in its local ordering)
5334: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMat()`, `MatMPIAIJGetLocalMatCondensed()`
5335: @*/
5336: PetscErrorCode MatMPIAIJGetLocalMatMerge(Mat A, MatReuse scall, IS *glob, Mat *A_loc)
5337: {
5338: Mat Ao, Ad;
5339: const PetscInt *cmap;
5340: PetscMPIInt size;
5341: PetscErrorCode (*f)(Mat, MatReuse, IS *, Mat *);
5343: PetscFunctionBegin;
5344: PetscCall(MatMPIAIJGetSeqAIJ(A, &Ad, &Ao, &cmap));
5345: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
5346: if (size == 1) {
5347: if (scall == MAT_INITIAL_MATRIX) {
5348: PetscCall(PetscObjectReference((PetscObject)Ad));
5349: *A_loc = Ad;
5350: } else if (scall == MAT_REUSE_MATRIX) {
5351: PetscCall(MatCopy(Ad, *A_loc, SAME_NONZERO_PATTERN));
5352: }
5353: if (glob) PetscCall(ISCreateStride(PetscObjectComm((PetscObject)Ad), Ad->cmap->n, Ad->cmap->rstart, 1, glob));
5354: PetscFunctionReturn(PETSC_SUCCESS);
5355: }
5356: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatMPIAIJGetLocalMatMerge_C", &f));
5357: PetscCall(PetscLogEventBegin(MAT_Getlocalmat, A, 0, 0, 0));
5358: if (f) {
5359: PetscCall((*f)(A, scall, glob, A_loc));
5360: } else {
5361: Mat_SeqAIJ *a = (Mat_SeqAIJ *)Ad->data;
5362: Mat_SeqAIJ *b = (Mat_SeqAIJ *)Ao->data;
5363: Mat_SeqAIJ *c;
5364: PetscInt *ai = a->i, *aj = a->j;
5365: PetscInt *bi = b->i, *bj = b->j;
5366: PetscInt *ci, *cj;
5367: const PetscScalar *aa, *ba;
5368: PetscScalar *ca;
5369: PetscInt i, j, am, dn, on;
5371: PetscCall(MatGetLocalSize(Ad, &am, &dn));
5372: PetscCall(MatGetLocalSize(Ao, NULL, &on));
5373: PetscCall(MatSeqAIJGetArrayRead(Ad, &aa));
5374: PetscCall(MatSeqAIJGetArrayRead(Ao, &ba));
5375: if (scall == MAT_INITIAL_MATRIX) {
5376: PetscInt k;
5377: PetscCall(PetscMalloc1(1 + am, &ci));
5378: PetscCall(PetscMalloc1(ai[am] + bi[am], &cj));
5379: PetscCall(PetscMalloc1(ai[am] + bi[am], &ca));
5380: ci[0] = 0;
5381: for (i = 0, k = 0; i < am; i++) {
5382: const PetscInt ncols_o = bi[i + 1] - bi[i];
5383: const PetscInt ncols_d = ai[i + 1] - ai[i];
5384: ci[i + 1] = ci[i] + ncols_o + ncols_d;
5385: /* diagonal portion of A */
5386: for (j = 0; j < ncols_d; j++, k++) {
5387: cj[k] = *aj++;
5388: ca[k] = *aa++;
5389: }
5390: /* off-diagonal portion of A */
5391: for (j = 0; j < ncols_o; j++, k++) {
5392: cj[k] = dn + *bj++;
5393: ca[k] = *ba++;
5394: }
5395: }
5396: /* put together the new matrix */
5397: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, am, dn + on, ci, cj, ca, A_loc));
5398: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5399: /* Since these are PETSc arrays, change flags to free them as necessary. */
5400: c = (Mat_SeqAIJ *)(*A_loc)->data;
5401: c->free_a = PETSC_TRUE;
5402: c->free_ij = PETSC_TRUE;
5403: c->nonew = 0;
5404: PetscCall(MatSetType(*A_loc, ((PetscObject)Ad)->type_name));
5405: } else if (scall == MAT_REUSE_MATRIX) {
5406: PetscCall(MatSeqAIJGetArrayWrite(*A_loc, &ca));
5407: for (i = 0; i < am; i++) {
5408: const PetscInt ncols_d = ai[i + 1] - ai[i];
5409: const PetscInt ncols_o = bi[i + 1] - bi[i];
5410: /* diagonal portion of A */
5411: for (j = 0; j < ncols_d; j++) *ca++ = *aa++;
5412: /* off-diagonal portion of A */
5413: for (j = 0; j < ncols_o; j++) *ca++ = *ba++;
5414: }
5415: PetscCall(MatSeqAIJRestoreArrayWrite(*A_loc, &ca));
5416: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Invalid MatReuse %d", (int)scall);
5417: PetscCall(MatSeqAIJRestoreArrayRead(Ad, &aa));
5418: PetscCall(MatSeqAIJRestoreArrayRead(Ao, &aa));
5419: if (glob) {
5420: PetscInt cst, *gidx;
5422: PetscCall(MatGetOwnershipRangeColumn(A, &cst, NULL));
5423: PetscCall(PetscMalloc1(dn + on, &gidx));
5424: for (i = 0; i < dn; i++) gidx[i] = cst + i;
5425: for (i = 0; i < on; i++) gidx[i + dn] = cmap[i];
5426: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)Ad), dn + on, gidx, PETSC_OWN_POINTER, glob));
5427: }
5428: }
5429: PetscCall(PetscLogEventEnd(MAT_Getlocalmat, A, 0, 0, 0));
5430: PetscFunctionReturn(PETSC_SUCCESS);
5431: }
5433: /*@C
5434: MatMPIAIJGetLocalMatCondensed - Creates a `MATSEQAIJ` matrix from an `MATMPIAIJ` matrix by taking all its local rows and NON-ZERO columns
5436: Not Collective
5438: Input Parameters:
5439: + A - the matrix
5440: . scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5441: . row - index set of rows to extract (or `NULL`)
5442: - col - index set of columns to extract (or `NULL`)
5444: Output Parameter:
5445: . A_loc - the local sequential matrix generated
5447: Level: developer
5449: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMat()`
5450: @*/
5451: PetscErrorCode MatMPIAIJGetLocalMatCondensed(Mat A, MatReuse scall, IS *row, IS *col, Mat *A_loc)
5452: {
5453: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5454: PetscInt i, start, end, ncols, nzA, nzB, *cmap, imark, *idx;
5455: IS isrowa, iscola;
5456: Mat *aloc;
5457: PetscBool match;
5459: PetscFunctionBegin;
5460: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIAIJ, &match));
5461: PetscCheck(match, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Requires MATMPIAIJ matrix as input");
5462: PetscCall(PetscLogEventBegin(MAT_Getlocalmatcondensed, A, 0, 0, 0));
5463: if (!row) {
5464: start = A->rmap->rstart;
5465: end = A->rmap->rend;
5466: PetscCall(ISCreateStride(PETSC_COMM_SELF, end - start, start, 1, &isrowa));
5467: } else {
5468: isrowa = *row;
5469: }
5470: if (!col) {
5471: start = A->cmap->rstart;
5472: cmap = a->garray;
5473: nzA = a->A->cmap->n;
5474: nzB = a->B->cmap->n;
5475: PetscCall(PetscMalloc1(nzA + nzB, &idx));
5476: ncols = 0;
5477: for (i = 0; i < nzB; i++) {
5478: if (cmap[i] < start) idx[ncols++] = cmap[i];
5479: else break;
5480: }
5481: imark = i;
5482: for (i = 0; i < nzA; i++) idx[ncols++] = start + i;
5483: for (i = imark; i < nzB; i++) idx[ncols++] = cmap[i];
5484: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, &iscola));
5485: } else {
5486: iscola = *col;
5487: }
5488: if (scall != MAT_INITIAL_MATRIX) {
5489: PetscCall(PetscMalloc1(1, &aloc));
5490: aloc[0] = *A_loc;
5491: }
5492: PetscCall(MatCreateSubMatrices(A, 1, &isrowa, &iscola, scall, &aloc));
5493: if (!col) { /* attach global id of condensed columns */
5494: PetscCall(PetscObjectCompose((PetscObject)aloc[0], "_petsc_GetLocalMatCondensed_iscol", (PetscObject)iscola));
5495: }
5496: *A_loc = aloc[0];
5497: PetscCall(PetscFree(aloc));
5498: if (!row) PetscCall(ISDestroy(&isrowa));
5499: if (!col) PetscCall(ISDestroy(&iscola));
5500: PetscCall(PetscLogEventEnd(MAT_Getlocalmatcondensed, A, 0, 0, 0));
5501: PetscFunctionReturn(PETSC_SUCCESS);
5502: }
5504: /*
5505: * Create a sequential AIJ matrix based on row indices. a whole column is extracted once a row is matched.
5506: * Row could be local or remote.The routine is designed to be scalable in memory so that nothing is based
5507: * on a global size.
5508: * */
5509: static PetscErrorCode MatCreateSeqSubMatrixWithRows_Private(Mat P, IS rows, Mat *P_oth)
5510: {
5511: Mat_MPIAIJ *p = (Mat_MPIAIJ *)P->data;
5512: Mat_SeqAIJ *pd = (Mat_SeqAIJ *)p->A->data, *po = (Mat_SeqAIJ *)p->B->data, *p_oth;
5513: PetscInt plocalsize, nrows, *ilocal, *oilocal, i, lidx, *nrcols, *nlcols, ncol;
5514: PetscMPIInt owner;
5515: PetscSFNode *iremote, *oiremote;
5516: const PetscInt *lrowindices;
5517: PetscSF sf, osf;
5518: PetscInt pcstart, *roffsets, *loffsets, *pnnz, j;
5519: PetscInt ontotalcols, dntotalcols, ntotalcols, nout;
5520: MPI_Comm comm;
5521: ISLocalToGlobalMapping mapping;
5522: const PetscScalar *pd_a, *po_a;
5524: PetscFunctionBegin;
5525: PetscCall(PetscObjectGetComm((PetscObject)P, &comm));
5526: /* plocalsize is the number of roots
5527: * nrows is the number of leaves
5528: * */
5529: PetscCall(MatGetLocalSize(P, &plocalsize, NULL));
5530: PetscCall(ISGetLocalSize(rows, &nrows));
5531: PetscCall(PetscCalloc1(nrows, &iremote));
5532: PetscCall(ISGetIndices(rows, &lrowindices));
5533: for (i = 0; i < nrows; i++) {
5534: /* Find a remote index and an owner for a row
5535: * The row could be local or remote
5536: * */
5537: owner = 0;
5538: lidx = 0;
5539: PetscCall(PetscLayoutFindOwnerIndex(P->rmap, lrowindices[i], &owner, &lidx));
5540: iremote[i].index = lidx;
5541: iremote[i].rank = owner;
5542: }
5543: /* Create SF to communicate how many nonzero columns for each row */
5544: PetscCall(PetscSFCreate(comm, &sf));
5545: /* SF will figure out the number of nonzero columns for each row, and their
5546: * offsets
5547: * */
5548: PetscCall(PetscSFSetGraph(sf, plocalsize, nrows, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
5549: PetscCall(PetscSFSetFromOptions(sf));
5550: PetscCall(PetscSFSetUp(sf));
5552: PetscCall(PetscCalloc1(2 * (plocalsize + 1), &roffsets));
5553: PetscCall(PetscCalloc1(2 * plocalsize, &nrcols));
5554: PetscCall(PetscCalloc1(nrows, &pnnz));
5555: roffsets[0] = 0;
5556: roffsets[1] = 0;
5557: for (i = 0; i < plocalsize; i++) {
5558: /* diagonal */
5559: nrcols[i * 2 + 0] = pd->i[i + 1] - pd->i[i];
5560: /* off-diagonal */
5561: nrcols[i * 2 + 1] = po->i[i + 1] - po->i[i];
5562: /* compute offsets so that we relative location for each row */
5563: roffsets[(i + 1) * 2 + 0] = roffsets[i * 2 + 0] + nrcols[i * 2 + 0];
5564: roffsets[(i + 1) * 2 + 1] = roffsets[i * 2 + 1] + nrcols[i * 2 + 1];
5565: }
5566: PetscCall(PetscCalloc1(2 * nrows, &nlcols));
5567: PetscCall(PetscCalloc1(2 * nrows, &loffsets));
5568: /* 'r' means root, and 'l' means leaf */
5569: PetscCall(PetscSFBcastBegin(sf, MPIU_2INT, nrcols, nlcols, MPI_REPLACE));
5570: PetscCall(PetscSFBcastBegin(sf, MPIU_2INT, roffsets, loffsets, MPI_REPLACE));
5571: PetscCall(PetscSFBcastEnd(sf, MPIU_2INT, nrcols, nlcols, MPI_REPLACE));
5572: PetscCall(PetscSFBcastEnd(sf, MPIU_2INT, roffsets, loffsets, MPI_REPLACE));
5573: PetscCall(PetscSFDestroy(&sf));
5574: PetscCall(PetscFree(roffsets));
5575: PetscCall(PetscFree(nrcols));
5576: dntotalcols = 0;
5577: ontotalcols = 0;
5578: ncol = 0;
5579: for (i = 0; i < nrows; i++) {
5580: pnnz[i] = nlcols[i * 2 + 0] + nlcols[i * 2 + 1];
5581: ncol = PetscMax(pnnz[i], ncol);
5582: /* diagonal */
5583: dntotalcols += nlcols[i * 2 + 0];
5584: /* off-diagonal */
5585: ontotalcols += nlcols[i * 2 + 1];
5586: }
5587: /* We do not need to figure the right number of columns
5588: * since all the calculations will be done by going through the raw data
5589: * */
5590: PetscCall(MatCreateSeqAIJ(PETSC_COMM_SELF, nrows, ncol, 0, pnnz, P_oth));
5591: PetscCall(MatSetUp(*P_oth));
5592: PetscCall(PetscFree(pnnz));
5593: p_oth = (Mat_SeqAIJ *)(*P_oth)->data;
5594: /* diagonal */
5595: PetscCall(PetscCalloc1(dntotalcols, &iremote));
5596: /* off-diagonal */
5597: PetscCall(PetscCalloc1(ontotalcols, &oiremote));
5598: /* diagonal */
5599: PetscCall(PetscCalloc1(dntotalcols, &ilocal));
5600: /* off-diagonal */
5601: PetscCall(PetscCalloc1(ontotalcols, &oilocal));
5602: dntotalcols = 0;
5603: ontotalcols = 0;
5604: ntotalcols = 0;
5605: for (i = 0; i < nrows; i++) {
5606: owner = 0;
5607: PetscCall(PetscLayoutFindOwnerIndex(P->rmap, lrowindices[i], &owner, NULL));
5608: /* Set iremote for diag matrix */
5609: for (j = 0; j < nlcols[i * 2 + 0]; j++) {
5610: iremote[dntotalcols].index = loffsets[i * 2 + 0] + j;
5611: iremote[dntotalcols].rank = owner;
5612: /* P_oth is seqAIJ so that ilocal need to point to the first part of memory */
5613: ilocal[dntotalcols++] = ntotalcols++;
5614: }
5615: /* off-diagonal */
5616: for (j = 0; j < nlcols[i * 2 + 1]; j++) {
5617: oiremote[ontotalcols].index = loffsets[i * 2 + 1] + j;
5618: oiremote[ontotalcols].rank = owner;
5619: oilocal[ontotalcols++] = ntotalcols++;
5620: }
5621: }
5622: PetscCall(ISRestoreIndices(rows, &lrowindices));
5623: PetscCall(PetscFree(loffsets));
5624: PetscCall(PetscFree(nlcols));
5625: PetscCall(PetscSFCreate(comm, &sf));
5626: /* P serves as roots and P_oth is leaves
5627: * Diag matrix
5628: * */
5629: PetscCall(PetscSFSetGraph(sf, pd->i[plocalsize], dntotalcols, ilocal, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
5630: PetscCall(PetscSFSetFromOptions(sf));
5631: PetscCall(PetscSFSetUp(sf));
5633: PetscCall(PetscSFCreate(comm, &osf));
5634: /* off-diagonal */
5635: PetscCall(PetscSFSetGraph(osf, po->i[plocalsize], ontotalcols, oilocal, PETSC_OWN_POINTER, oiremote, PETSC_OWN_POINTER));
5636: PetscCall(PetscSFSetFromOptions(osf));
5637: PetscCall(PetscSFSetUp(osf));
5638: PetscCall(MatSeqAIJGetArrayRead(p->A, &pd_a));
5639: PetscCall(MatSeqAIJGetArrayRead(p->B, &po_a));
5640: /* operate on the matrix internal data to save memory */
5641: PetscCall(PetscSFBcastBegin(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5642: PetscCall(PetscSFBcastBegin(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5643: PetscCall(MatGetOwnershipRangeColumn(P, &pcstart, NULL));
5644: /* Convert to global indices for diag matrix */
5645: for (i = 0; i < pd->i[plocalsize]; i++) pd->j[i] += pcstart;
5646: PetscCall(PetscSFBcastBegin(sf, MPIU_INT, pd->j, p_oth->j, MPI_REPLACE));
5647: /* We want P_oth store global indices */
5648: PetscCall(ISLocalToGlobalMappingCreate(comm, 1, p->B->cmap->n, p->garray, PETSC_COPY_VALUES, &mapping));
5649: /* Use memory scalable approach */
5650: PetscCall(ISLocalToGlobalMappingSetType(mapping, ISLOCALTOGLOBALMAPPINGHASH));
5651: PetscCall(ISLocalToGlobalMappingApply(mapping, po->i[plocalsize], po->j, po->j));
5652: PetscCall(PetscSFBcastBegin(osf, MPIU_INT, po->j, p_oth->j, MPI_REPLACE));
5653: PetscCall(PetscSFBcastEnd(sf, MPIU_INT, pd->j, p_oth->j, MPI_REPLACE));
5654: /* Convert back to local indices */
5655: for (i = 0; i < pd->i[plocalsize]; i++) pd->j[i] -= pcstart;
5656: PetscCall(PetscSFBcastEnd(osf, MPIU_INT, po->j, p_oth->j, MPI_REPLACE));
5657: nout = 0;
5658: PetscCall(ISGlobalToLocalMappingApply(mapping, IS_GTOLM_DROP, po->i[plocalsize], po->j, &nout, po->j));
5659: PetscCheck(nout == po->i[plocalsize], comm, PETSC_ERR_ARG_INCOMP, "n %" PetscInt_FMT " does not equal to nout %" PetscInt_FMT " ", po->i[plocalsize], nout);
5660: PetscCall(ISLocalToGlobalMappingDestroy(&mapping));
5661: /* Exchange values */
5662: PetscCall(PetscSFBcastEnd(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5663: PetscCall(PetscSFBcastEnd(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5664: PetscCall(MatSeqAIJRestoreArrayRead(p->A, &pd_a));
5665: PetscCall(MatSeqAIJRestoreArrayRead(p->B, &po_a));
5666: /* Stop PETSc from shrinking memory */
5667: for (i = 0; i < nrows; i++) p_oth->ilen[i] = p_oth->imax[i];
5668: PetscCall(MatAssemblyBegin(*P_oth, MAT_FINAL_ASSEMBLY));
5669: PetscCall(MatAssemblyEnd(*P_oth, MAT_FINAL_ASSEMBLY));
5670: /* Attach PetscSF objects to P_oth so that we can reuse it later */
5671: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "diagsf", (PetscObject)sf));
5672: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "offdiagsf", (PetscObject)osf));
5673: PetscCall(PetscSFDestroy(&sf));
5674: PetscCall(PetscSFDestroy(&osf));
5675: PetscFunctionReturn(PETSC_SUCCESS);
5676: }
5678: /*
5679: * Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A
5680: * This supports MPIAIJ and MAIJ
5681: * */
5682: PetscErrorCode MatGetBrowsOfAcols_MPIXAIJ(Mat A, Mat P, PetscInt dof, MatReuse reuse, Mat *P_oth)
5683: {
5684: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data, *p = (Mat_MPIAIJ *)P->data;
5685: Mat_SeqAIJ *p_oth;
5686: IS rows, map;
5687: PetscHMapI hamp;
5688: PetscInt i, htsize, *rowindices, off, *mapping, key, count;
5689: MPI_Comm comm;
5690: PetscSF sf, osf;
5691: PetscBool has;
5693: PetscFunctionBegin;
5694: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
5695: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAocols, A, P, 0, 0));
5696: /* If it is the first time, create an index set of off-diag nonzero columns of A,
5697: * and then create a submatrix (that often is an overlapping matrix)
5698: * */
5699: if (reuse == MAT_INITIAL_MATRIX) {
5700: /* Use a hash table to figure out unique keys */
5701: PetscCall(PetscHMapICreateWithSize(a->B->cmap->n, &hamp));
5702: PetscCall(PetscCalloc1(a->B->cmap->n, &mapping));
5703: count = 0;
5704: /* Assume that a->g is sorted, otherwise the following does not make sense */
5705: for (i = 0; i < a->B->cmap->n; i++) {
5706: key = a->garray[i] / dof;
5707: PetscCall(PetscHMapIHas(hamp, key, &has));
5708: if (!has) {
5709: mapping[i] = count;
5710: PetscCall(PetscHMapISet(hamp, key, count++));
5711: } else {
5712: /* Current 'i' has the same value the previous step */
5713: mapping[i] = count - 1;
5714: }
5715: }
5716: PetscCall(ISCreateGeneral(comm, a->B->cmap->n, mapping, PETSC_OWN_POINTER, &map));
5717: PetscCall(PetscHMapIGetSize(hamp, &htsize));
5718: PetscCheck(htsize == count, comm, PETSC_ERR_ARG_INCOMP, " Size of hash map %" PetscInt_FMT " is inconsistent with count %" PetscInt_FMT, htsize, count);
5719: PetscCall(PetscCalloc1(htsize, &rowindices));
5720: off = 0;
5721: PetscCall(PetscHMapIGetKeys(hamp, &off, rowindices));
5722: PetscCall(PetscHMapIDestroy(&hamp));
5723: PetscCall(PetscSortInt(htsize, rowindices));
5724: PetscCall(ISCreateGeneral(comm, htsize, rowindices, PETSC_OWN_POINTER, &rows));
5725: /* In case, the matrix was already created but users want to recreate the matrix */
5726: PetscCall(MatDestroy(P_oth));
5727: PetscCall(MatCreateSeqSubMatrixWithRows_Private(P, rows, P_oth));
5728: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "aoffdiagtopothmapping", (PetscObject)map));
5729: PetscCall(ISDestroy(&map));
5730: PetscCall(ISDestroy(&rows));
5731: } else if (reuse == MAT_REUSE_MATRIX) {
5732: /* If matrix was already created, we simply update values using SF objects
5733: * that as attached to the matrix earlier.
5734: */
5735: const PetscScalar *pd_a, *po_a;
5737: PetscCall(PetscObjectQuery((PetscObject)*P_oth, "diagsf", (PetscObject *)&sf));
5738: PetscCall(PetscObjectQuery((PetscObject)*P_oth, "offdiagsf", (PetscObject *)&osf));
5739: PetscCheck(sf && osf, comm, PETSC_ERR_ARG_NULL, "Matrix is not initialized yet");
5740: p_oth = (Mat_SeqAIJ *)(*P_oth)->data;
5741: /* Update values in place */
5742: PetscCall(MatSeqAIJGetArrayRead(p->A, &pd_a));
5743: PetscCall(MatSeqAIJGetArrayRead(p->B, &po_a));
5744: PetscCall(PetscSFBcastBegin(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5745: PetscCall(PetscSFBcastBegin(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5746: PetscCall(PetscSFBcastEnd(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5747: PetscCall(PetscSFBcastEnd(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5748: PetscCall(MatSeqAIJRestoreArrayRead(p->A, &pd_a));
5749: PetscCall(MatSeqAIJRestoreArrayRead(p->B, &po_a));
5750: } else SETERRQ(comm, PETSC_ERR_ARG_UNKNOWN_TYPE, "Unknown reuse type");
5751: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAocols, A, P, 0, 0));
5752: PetscFunctionReturn(PETSC_SUCCESS);
5753: }
5755: /*@C
5756: MatGetBrowsOfAcols - Returns `IS` that contain rows of `B` that equal to nonzero columns of local `A`
5758: Collective
5760: Input Parameters:
5761: + A - the first matrix in `MATMPIAIJ` format
5762: . B - the second matrix in `MATMPIAIJ` format
5763: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5765: Output Parameters:
5766: + rowb - On input index sets of rows of B to extract (or `NULL`), modified on output
5767: . colb - On input index sets of columns of B to extract (or `NULL`), modified on output
5768: - B_seq - the sequential matrix generated
5770: Level: developer
5772: .seealso: `Mat`, `MATMPIAIJ`, `IS`, `MatReuse`
5773: @*/
5774: PetscErrorCode MatGetBrowsOfAcols(Mat A, Mat B, MatReuse scall, IS *rowb, IS *colb, Mat *B_seq)
5775: {
5776: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5777: PetscInt *idx, i, start, ncols, nzA, nzB, *cmap, imark;
5778: IS isrowb, iscolb;
5779: Mat *bseq = NULL;
5781: PetscFunctionBegin;
5782: 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 ")",
5783: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
5784: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAcols, A, B, 0, 0));
5786: if (scall == MAT_INITIAL_MATRIX) {
5787: start = A->cmap->rstart;
5788: cmap = a->garray;
5789: nzA = a->A->cmap->n;
5790: nzB = a->B->cmap->n;
5791: PetscCall(PetscMalloc1(nzA + nzB, &idx));
5792: ncols = 0;
5793: for (i = 0; i < nzB; i++) { /* row < local row index */
5794: if (cmap[i] < start) idx[ncols++] = cmap[i];
5795: else break;
5796: }
5797: imark = i;
5798: for (i = 0; i < nzA; i++) idx[ncols++] = start + i; /* local rows */
5799: for (i = imark; i < nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */
5800: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, &isrowb));
5801: PetscCall(ISCreateStride(PETSC_COMM_SELF, B->cmap->N, 0, 1, &iscolb));
5802: } else {
5803: PetscCheck(rowb && colb, PETSC_COMM_SELF, PETSC_ERR_SUP, "IS rowb and colb must be provided for MAT_REUSE_MATRIX");
5804: isrowb = *rowb;
5805: iscolb = *colb;
5806: PetscCall(PetscMalloc1(1, &bseq));
5807: bseq[0] = *B_seq;
5808: }
5809: PetscCall(MatCreateSubMatrices(B, 1, &isrowb, &iscolb, scall, &bseq));
5810: *B_seq = bseq[0];
5811: PetscCall(PetscFree(bseq));
5812: if (!rowb) {
5813: PetscCall(ISDestroy(&isrowb));
5814: } else {
5815: *rowb = isrowb;
5816: }
5817: if (!colb) {
5818: PetscCall(ISDestroy(&iscolb));
5819: } else {
5820: *colb = iscolb;
5821: }
5822: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAcols, A, B, 0, 0));
5823: PetscFunctionReturn(PETSC_SUCCESS);
5824: }
5826: /*
5827: MatGetBrowsOfAoCols_MPIAIJ - Creates a `MATSEQAIJ` matrix by taking rows of B that equal to nonzero columns
5828: of the OFF-DIAGONAL portion of local A
5830: Collective
5832: Input Parameters:
5833: + A,B - the matrices in `MATMPIAIJ` format
5834: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5836: Output Parameter:
5837: + startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or NULL)
5838: . startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or NULL)
5839: . bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or NULL)
5840: - B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N
5842: Developer Note:
5843: This directly accesses information inside the VecScatter associated with the matrix-vector product
5844: for this matrix. This is not desirable..
5846: Level: developer
5848: */
5850: PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat A, Mat B, MatReuse scall, PetscInt **startsj_s, PetscInt **startsj_r, MatScalar **bufa_ptr, Mat *B_oth)
5851: {
5852: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5853: VecScatter ctx;
5854: MPI_Comm comm;
5855: const PetscMPIInt *rprocs, *sprocs;
5856: PetscMPIInt nrecvs, nsends;
5857: const PetscInt *srow, *rstarts, *sstarts;
5858: PetscInt *rowlen, *bufj, *bufJ, ncols = 0, aBn = a->B->cmap->n, row, *b_othi, *b_othj, *rvalues = NULL, *svalues = NULL, *cols, sbs, rbs;
5859: PetscInt i, j, k = 0, l, ll, nrows, *rstartsj = NULL, *sstartsj, len;
5860: PetscScalar *b_otha, *bufa, *bufA, *vals = NULL;
5861: MPI_Request *reqs = NULL, *rwaits = NULL, *swaits = NULL;
5862: PetscMPIInt size, tag, rank, nreqs;
5864: PetscFunctionBegin;
5865: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
5866: PetscCallMPI(MPI_Comm_size(comm, &size));
5868: 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 ")",
5869: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
5870: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAocols, A, B, 0, 0));
5871: PetscCallMPI(MPI_Comm_rank(comm, &rank));
5873: if (size == 1) {
5874: startsj_s = NULL;
5875: bufa_ptr = NULL;
5876: *B_oth = NULL;
5877: PetscFunctionReturn(PETSC_SUCCESS);
5878: }
5880: ctx = a->Mvctx;
5881: tag = ((PetscObject)ctx)->tag;
5883: PetscCall(VecScatterGetRemote_Private(ctx, PETSC_TRUE /*send*/, &nsends, &sstarts, &srow, &sprocs, &sbs));
5884: /* rprocs[] must be ordered so that indices received from them are ordered in rvalues[], which is key to algorithms used in this subroutine */
5885: PetscCall(VecScatterGetRemoteOrdered_Private(ctx, PETSC_FALSE /*recv*/, &nrecvs, &rstarts, NULL /*indices not needed*/, &rprocs, &rbs));
5886: PetscCall(PetscMPIIntCast(nsends + nrecvs, &nreqs));
5887: PetscCall(PetscMalloc1(nreqs, &reqs));
5888: rwaits = reqs;
5889: swaits = PetscSafePointerPlusOffset(reqs, nrecvs);
5891: if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX;
5892: if (scall == MAT_INITIAL_MATRIX) {
5893: /* i-array */
5894: /* post receives */
5895: if (nrecvs) PetscCall(PetscMalloc1(rbs * (rstarts[nrecvs] - rstarts[0]), &rvalues)); /* rstarts can be NULL when nrecvs=0 */
5896: for (i = 0; i < nrecvs; i++) {
5897: rowlen = rvalues + rstarts[i] * rbs;
5898: nrows = (rstarts[i + 1] - rstarts[i]) * rbs; /* num of indices to be received */
5899: PetscCallMPI(MPIU_Irecv(rowlen, nrows, MPIU_INT, rprocs[i], tag, comm, rwaits + i));
5900: }
5902: /* pack the outgoing message */
5903: PetscCall(PetscMalloc2(nsends + 1, &sstartsj, nrecvs + 1, &rstartsj));
5905: sstartsj[0] = 0;
5906: rstartsj[0] = 0;
5907: len = 0; /* total length of j or a array to be sent */
5908: if (nsends) {
5909: k = sstarts[0]; /* ATTENTION: sstarts[0] and rstarts[0] are not necessarily zero */
5910: PetscCall(PetscMalloc1(sbs * (sstarts[nsends] - sstarts[0]), &svalues));
5911: }
5912: for (i = 0; i < nsends; i++) {
5913: rowlen = svalues + (sstarts[i] - sstarts[0]) * sbs;
5914: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5915: for (j = 0; j < nrows; j++) {
5916: row = srow[k] + B->rmap->range[rank]; /* global row idx */
5917: for (l = 0; l < sbs; l++) {
5918: PetscCall(MatGetRow_MPIAIJ(B, row + l, &ncols, NULL, NULL)); /* rowlength */
5920: rowlen[j * sbs + l] = ncols;
5922: len += ncols;
5923: PetscCall(MatRestoreRow_MPIAIJ(B, row + l, &ncols, NULL, NULL));
5924: }
5925: k++;
5926: }
5927: PetscCallMPI(MPIU_Isend(rowlen, nrows * sbs, MPIU_INT, sprocs[i], tag, comm, swaits + i));
5929: sstartsj[i + 1] = len; /* starting point of (i+1)-th outgoing msg in bufj and bufa */
5930: }
5931: /* recvs and sends of i-array are completed */
5932: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5933: PetscCall(PetscFree(svalues));
5935: /* allocate buffers for sending j and a arrays */
5936: PetscCall(PetscMalloc1(len + 1, &bufj));
5937: PetscCall(PetscMalloc1(len + 1, &bufa));
5939: /* create i-array of B_oth */
5940: PetscCall(PetscMalloc1(aBn + 2, &b_othi));
5942: b_othi[0] = 0;
5943: len = 0; /* total length of j or a array to be received */
5944: k = 0;
5945: for (i = 0; i < nrecvs; i++) {
5946: rowlen = rvalues + (rstarts[i] - rstarts[0]) * rbs;
5947: nrows = (rstarts[i + 1] - rstarts[i]) * rbs; /* num of rows to be received */
5948: for (j = 0; j < nrows; j++) {
5949: b_othi[k + 1] = b_othi[k] + rowlen[j];
5950: PetscCall(PetscIntSumError(rowlen[j], len, &len));
5951: k++;
5952: }
5953: rstartsj[i + 1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */
5954: }
5955: PetscCall(PetscFree(rvalues));
5957: /* allocate space for j and a arrays of B_oth */
5958: PetscCall(PetscMalloc1(b_othi[aBn] + 1, &b_othj));
5959: PetscCall(PetscMalloc1(b_othi[aBn] + 1, &b_otha));
5961: /* j-array */
5962: /* post receives of j-array */
5963: for (i = 0; i < nrecvs; i++) {
5964: nrows = rstartsj[i + 1] - rstartsj[i]; /* length of the msg received */
5965: PetscCallMPI(MPIU_Irecv(b_othj + rstartsj[i], nrows, MPIU_INT, rprocs[i], tag, comm, rwaits + i));
5966: }
5968: /* pack the outgoing message j-array */
5969: if (nsends) k = sstarts[0];
5970: for (i = 0; i < nsends; i++) {
5971: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5972: bufJ = bufj + sstartsj[i];
5973: for (j = 0; j < nrows; j++) {
5974: row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5975: for (ll = 0; ll < sbs; ll++) {
5976: PetscCall(MatGetRow_MPIAIJ(B, row + ll, &ncols, &cols, NULL));
5977: for (l = 0; l < ncols; l++) *bufJ++ = cols[l];
5978: PetscCall(MatRestoreRow_MPIAIJ(B, row + ll, &ncols, &cols, NULL));
5979: }
5980: }
5981: PetscCallMPI(MPIU_Isend(bufj + sstartsj[i], sstartsj[i + 1] - sstartsj[i], MPIU_INT, sprocs[i], tag, comm, swaits + i));
5982: }
5984: /* recvs and sends of j-array are completed */
5985: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5986: } else if (scall == MAT_REUSE_MATRIX) {
5987: sstartsj = *startsj_s;
5988: rstartsj = *startsj_r;
5989: bufa = *bufa_ptr;
5990: PetscCall(MatSeqAIJGetArrayWrite(*B_oth, &b_otha));
5991: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not possess an object container");
5993: /* a-array */
5994: /* post receives of a-array */
5995: for (i = 0; i < nrecvs; i++) {
5996: nrows = rstartsj[i + 1] - rstartsj[i]; /* length of the msg received */
5997: PetscCallMPI(MPIU_Irecv(b_otha + rstartsj[i], nrows, MPIU_SCALAR, rprocs[i], tag, comm, rwaits + i));
5998: }
6000: /* pack the outgoing message a-array */
6001: if (nsends) k = sstarts[0];
6002: for (i = 0; i < nsends; i++) {
6003: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
6004: bufA = bufa + sstartsj[i];
6005: for (j = 0; j < nrows; j++) {
6006: row = srow[k++] + B->rmap->range[rank]; /* global row idx */
6007: for (ll = 0; ll < sbs; ll++) {
6008: PetscCall(MatGetRow_MPIAIJ(B, row + ll, &ncols, NULL, &vals));
6009: for (l = 0; l < ncols; l++) *bufA++ = vals[l];
6010: PetscCall(MatRestoreRow_MPIAIJ(B, row + ll, &ncols, NULL, &vals));
6011: }
6012: }
6013: PetscCallMPI(MPIU_Isend(bufa + sstartsj[i], sstartsj[i + 1] - sstartsj[i], MPIU_SCALAR, sprocs[i], tag, comm, swaits + i));
6014: }
6015: /* recvs and sends of a-array are completed */
6016: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
6017: PetscCall(PetscFree(reqs));
6019: if (scall == MAT_INITIAL_MATRIX) {
6020: Mat_SeqAIJ *b_oth;
6022: /* put together the new matrix */
6023: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, aBn, B->cmap->N, b_othi, b_othj, b_otha, B_oth));
6025: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
6026: /* Since these are PETSc arrays, change flags to free them as necessary. */
6027: b_oth = (Mat_SeqAIJ *)(*B_oth)->data;
6028: b_oth->free_a = PETSC_TRUE;
6029: b_oth->free_ij = PETSC_TRUE;
6030: b_oth->nonew = 0;
6032: PetscCall(PetscFree(bufj));
6033: if (!startsj_s || !bufa_ptr) {
6034: PetscCall(PetscFree2(sstartsj, rstartsj));
6035: PetscCall(PetscFree(bufa_ptr));
6036: } else {
6037: *startsj_s = sstartsj;
6038: *startsj_r = rstartsj;
6039: *bufa_ptr = bufa;
6040: }
6041: } else if (scall == MAT_REUSE_MATRIX) {
6042: PetscCall(MatSeqAIJRestoreArrayWrite(*B_oth, &b_otha));
6043: }
6045: PetscCall(VecScatterRestoreRemote_Private(ctx, PETSC_TRUE, &nsends, &sstarts, &srow, &sprocs, &sbs));
6046: PetscCall(VecScatterRestoreRemoteOrdered_Private(ctx, PETSC_FALSE, &nrecvs, &rstarts, NULL, &rprocs, &rbs));
6047: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAocols, A, B, 0, 0));
6048: PetscFunctionReturn(PETSC_SUCCESS);
6049: }
6051: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat, MatType, MatReuse, Mat *);
6052: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJPERM(Mat, MatType, MatReuse, Mat *);
6053: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJSELL(Mat, MatType, MatReuse, Mat *);
6054: #if defined(PETSC_HAVE_MKL_SPARSE)
6055: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJMKL(Mat, MatType, MatReuse, Mat *);
6056: #endif
6057: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIBAIJ(Mat, MatType, MatReuse, Mat *);
6058: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISBAIJ(Mat, MatType, MatReuse, Mat *);
6059: #if defined(PETSC_HAVE_ELEMENTAL)
6060: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat, MatType, MatReuse, Mat *);
6061: #endif
6062: #if defined(PETSC_HAVE_SCALAPACK)
6063: PETSC_INTERN PetscErrorCode MatConvert_AIJ_ScaLAPACK(Mat, MatType, MatReuse, Mat *);
6064: #endif
6065: #if defined(PETSC_HAVE_HYPRE)
6066: PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat, MatType, MatReuse, Mat *);
6067: #endif
6068: #if defined(PETSC_HAVE_CUDA)
6069: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCUSPARSE(Mat, MatType, MatReuse, Mat *);
6070: #endif
6071: #if defined(PETSC_HAVE_HIP)
6072: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJHIPSPARSE(Mat, MatType, MatReuse, Mat *);
6073: #endif
6074: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
6075: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJKokkos(Mat, MatType, MatReuse, Mat *);
6076: #endif
6077: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISELL(Mat, MatType, MatReuse, Mat *);
6078: PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat, MatType, MatReuse, Mat *);
6079: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_IS_XAIJ(Mat);
6081: /*
6082: Computes (B'*A')' since computing B*A directly is untenable
6084: n p p
6085: [ ] [ ] [ ]
6086: m [ A ] * n [ B ] = m [ C ]
6087: [ ] [ ] [ ]
6089: */
6090: static PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A, Mat B, Mat C)
6091: {
6092: Mat At, Bt, Ct;
6094: PetscFunctionBegin;
6095: PetscCall(MatTranspose(A, MAT_INITIAL_MATRIX, &At));
6096: PetscCall(MatTranspose(B, MAT_INITIAL_MATRIX, &Bt));
6097: PetscCall(MatMatMult(Bt, At, MAT_INITIAL_MATRIX, PETSC_CURRENT, &Ct));
6098: PetscCall(MatDestroy(&At));
6099: PetscCall(MatDestroy(&Bt));
6100: PetscCall(MatTransposeSetPrecursor(Ct, C));
6101: PetscCall(MatTranspose(Ct, MAT_REUSE_MATRIX, &C));
6102: PetscCall(MatDestroy(&Ct));
6103: PetscFunctionReturn(PETSC_SUCCESS);
6104: }
6106: static PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A, Mat B, PetscReal fill, Mat C)
6107: {
6108: PetscBool cisdense;
6110: PetscFunctionBegin;
6111: 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);
6112: PetscCall(MatSetSizes(C, A->rmap->n, B->cmap->n, A->rmap->N, B->cmap->N));
6113: PetscCall(MatSetBlockSizesFromMats(C, A, B));
6114: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATMPIDENSE, MATMPIDENSECUDA, MATMPIDENSEHIP, ""));
6115: if (!cisdense) PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
6116: PetscCall(MatSetUp(C));
6118: C->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIAIJ;
6119: PetscFunctionReturn(PETSC_SUCCESS);
6120: }
6122: static PetscErrorCode MatProductSetFromOptions_MPIDense_MPIAIJ_AB(Mat C)
6123: {
6124: Mat_Product *product = C->product;
6125: Mat A = product->A, B = product->B;
6127: PetscFunctionBegin;
6128: 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 ")",
6129: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
6130: C->ops->matmultsymbolic = MatMatMultSymbolic_MPIDense_MPIAIJ;
6131: C->ops->productsymbolic = MatProductSymbolic_AB;
6132: PetscFunctionReturn(PETSC_SUCCESS);
6133: }
6135: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_MPIDense_MPIAIJ(Mat C)
6136: {
6137: Mat_Product *product = C->product;
6139: PetscFunctionBegin;
6140: if (product->type == MATPRODUCT_AB) PetscCall(MatProductSetFromOptions_MPIDense_MPIAIJ_AB(C));
6141: PetscFunctionReturn(PETSC_SUCCESS);
6142: }
6144: /*
6145: Merge two sets of sorted nonzeros and return a CSR for the merged (sequential) matrix
6147: Input Parameters:
6149: j1,rowBegin1,rowEnd1,jmap1: describe the first set of nonzeros (Set1)
6150: j2,rowBegin2,rowEnd2,jmap2: describe the second set of nonzeros (Set2)
6152: mat: both sets' nonzeros are on m rows, where m is the number of local rows of the matrix mat
6154: For Set1, j1[] contains column indices of the nonzeros.
6155: 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
6156: respectively (note rowEnd1[k] is not necessarily equal to rwoBegin1[k+1]). Indices in this range of j1[] are sorted,
6157: but might have repeats. jmap1[t+1] - jmap1[t] is the number of repeats for the t-th unique nonzero in Set1.
6159: Similar for Set2.
6161: This routine merges the two sets of nonzeros row by row and removes repeats.
6163: Output Parameters: (memory is allocated by the caller)
6165: i[],j[]: the CSR of the merged matrix, which has m rows.
6166: imap1[]: the k-th unique nonzero in Set1 (k=0,1,...) corresponds to imap1[k]-th unique nonzero in the merged matrix.
6167: imap2[]: similar to imap1[], but for Set2.
6168: Note we order nonzeros row-by-row and from left to right.
6169: */
6170: 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[])
6171: {
6172: PetscInt r, m; /* Row index of mat */
6173: PetscCount t, t1, t2, b1, e1, b2, e2;
6175: PetscFunctionBegin;
6176: PetscCall(MatGetLocalSize(mat, &m, NULL));
6177: t1 = t2 = t = 0; /* Count unique nonzeros of in Set1, Set1 and the merged respectively */
6178: i[0] = 0;
6179: for (r = 0; r < m; r++) { /* Do row by row merging */
6180: b1 = rowBegin1[r];
6181: e1 = rowEnd1[r];
6182: b2 = rowBegin2[r];
6183: e2 = rowEnd2[r];
6184: while (b1 < e1 && b2 < e2) {
6185: if (j1[b1] == j2[b2]) { /* Same column index and hence same nonzero */
6186: j[t] = j1[b1];
6187: imap1[t1] = t;
6188: imap2[t2] = t;
6189: b1 += jmap1[t1 + 1] - jmap1[t1]; /* Jump to next unique local nonzero */
6190: b2 += jmap2[t2 + 1] - jmap2[t2]; /* Jump to next unique remote nonzero */
6191: t1++;
6192: t2++;
6193: t++;
6194: } else if (j1[b1] < j2[b2]) {
6195: j[t] = j1[b1];
6196: imap1[t1] = t;
6197: b1 += jmap1[t1 + 1] - jmap1[t1];
6198: t1++;
6199: t++;
6200: } else {
6201: j[t] = j2[b2];
6202: imap2[t2] = t;
6203: b2 += jmap2[t2 + 1] - jmap2[t2];
6204: t2++;
6205: t++;
6206: }
6207: }
6208: /* Merge the remaining in either j1[] or j2[] */
6209: while (b1 < e1) {
6210: j[t] = j1[b1];
6211: imap1[t1] = t;
6212: b1 += jmap1[t1 + 1] - jmap1[t1];
6213: t1++;
6214: t++;
6215: }
6216: while (b2 < e2) {
6217: j[t] = j2[b2];
6218: imap2[t2] = t;
6219: b2 += jmap2[t2 + 1] - jmap2[t2];
6220: t2++;
6221: t++;
6222: }
6223: PetscCall(PetscIntCast(t, i + r + 1));
6224: }
6225: PetscFunctionReturn(PETSC_SUCCESS);
6226: }
6228: /*
6229: Split nonzeros in a block of local rows into two subsets: those in the diagonal block and those in the off-diagonal block
6231: Input Parameters:
6232: mat: an MPI matrix that provides row and column layout information for splitting. Let's say its number of local rows is m.
6233: 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[]
6234: respectively, along with a permutation array perm[]. Length of the i[],j[],perm[] arrays is n.
6236: i[] is already sorted, but within a row, j[] is not sorted and might have repeats.
6237: i[] might contain negative indices at the beginning, which means the corresponding entries should be ignored in the splitting.
6239: Output Parameters:
6240: j[],perm[]: the routine needs to sort j[] within each row along with perm[].
6241: rowBegin[],rowMid[],rowEnd[]: of length m, and the memory is preallocated and zeroed by the caller.
6242: 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,
6243: and [rowMid[r],rowEnd[r]) point to begin/end entries of row r of the off-diagonal block.
6245: Aperm[],Ajmap[],Atot,Annz: Arrays are allocated by this routine.
6246: Atot: number of entries belonging to the diagonal block.
6247: Annz: number of unique nonzeros belonging to the diagonal block.
6248: Aperm[Atot] stores values from perm[] for entries belonging to the diagonal block. Length of Aperm[] is Atot, though it may also count
6249: repeats (i.e., same 'i,j' pair).
6250: Ajmap[Annz+1] stores the number of repeats of each unique entry belonging to the diagonal block. More precisely, Ajmap[t+1] - Ajmap[t]
6251: is the number of repeats for the t-th unique entry in the diagonal block. Ajmap[0] is always 0.
6253: Atot: number of entries belonging to the diagonal block
6254: Annz: number of unique nonzeros belonging to the diagonal block.
6256: Bperm[], Bjmap[], Btot, Bnnz are similar but for the off-diagonal block.
6258: Aperm[],Bperm[],Ajmap[] and Bjmap[] are allocated separately by this routine with PetscMalloc1().
6259: */
6260: 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_)
6261: {
6262: PetscInt cstart, cend, rstart, rend, row, col;
6263: PetscCount Atot = 0, Btot = 0; /* Total number of nonzeros in the diagonal and off-diagonal blocks */
6264: PetscCount Annz = 0, Bnnz = 0; /* Number of unique nonzeros in the diagonal and off-diagonal blocks */
6265: PetscCount k, m, p, q, r, s, mid;
6266: PetscCount *Aperm, *Bperm, *Ajmap, *Bjmap;
6268: PetscFunctionBegin;
6269: PetscCall(PetscLayoutGetRange(mat->rmap, &rstart, &rend));
6270: PetscCall(PetscLayoutGetRange(mat->cmap, &cstart, &cend));
6271: m = rend - rstart;
6273: /* Skip negative rows */
6274: for (k = 0; k < n; k++)
6275: if (i[k] >= 0) break;
6277: /* Process [k,n): sort and partition each local row into diag and offdiag portions,
6278: fill rowBegin[], rowMid[], rowEnd[], and count Atot, Btot, Annz, Bnnz.
6279: */
6280: while (k < n) {
6281: row = i[k];
6282: /* Entries in [k,s) are in one row. Shift diagonal block col indices so that diag is ahead of offdiag after sorting the row */
6283: for (s = k; s < n; s++)
6284: if (i[s] != row) break;
6286: /* Shift diag columns to range of [-PETSC_INT_MAX, -1] */
6287: for (p = k; p < s; p++) {
6288: if (j[p] >= cstart && j[p] < cend) j[p] -= PETSC_INT_MAX;
6289: else PetscAssert((j[p] >= 0) && (j[p] <= mat->cmap->N), PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column index %" PetscInt_FMT " is out of range", j[p]);
6290: }
6291: PetscCall(PetscSortIntWithCountArray(s - k, j + k, perm + k));
6292: PetscCall(PetscSortedIntUpperBound(j, k, s, -1, &mid)); /* Separate [k,s) into [k,mid) for diag and [mid,s) for offdiag */
6293: rowBegin[row - rstart] = k;
6294: rowMid[row - rstart] = mid;
6295: rowEnd[row - rstart] = s;
6297: /* Count nonzeros of this diag/offdiag row, which might have repeats */
6298: Atot += mid - k;
6299: Btot += s - mid;
6301: /* Count unique nonzeros of this diag row */
6302: for (p = k; p < mid;) {
6303: col = j[p];
6304: do {
6305: j[p] += PETSC_INT_MAX; /* Revert the modified diagonal indices */
6306: p++;
6307: } while (p < mid && j[p] == col);
6308: Annz++;
6309: }
6311: /* Count unique nonzeros of this offdiag row */
6312: for (p = mid; p < s;) {
6313: col = j[p];
6314: do {
6315: p++;
6316: } while (p < s && j[p] == col);
6317: Bnnz++;
6318: }
6319: k = s;
6320: }
6322: /* Allocation according to Atot, Btot, Annz, Bnnz */
6323: PetscCall(PetscMalloc1(Atot, &Aperm));
6324: PetscCall(PetscMalloc1(Btot, &Bperm));
6325: PetscCall(PetscMalloc1(Annz + 1, &Ajmap));
6326: PetscCall(PetscMalloc1(Bnnz + 1, &Bjmap));
6328: /* Re-scan indices and copy diag/offdiag permutation indices to Aperm, Bperm and also fill Ajmap and Bjmap */
6329: Ajmap[0] = Bjmap[0] = Atot = Btot = Annz = Bnnz = 0;
6330: for (r = 0; r < m; r++) {
6331: k = rowBegin[r];
6332: mid = rowMid[r];
6333: s = rowEnd[r];
6334: PetscCall(PetscArraycpy(PetscSafePointerPlusOffset(Aperm, Atot), PetscSafePointerPlusOffset(perm, k), mid - k));
6335: PetscCall(PetscArraycpy(PetscSafePointerPlusOffset(Bperm, Btot), PetscSafePointerPlusOffset(perm, mid), s - mid));
6336: Atot += mid - k;
6337: Btot += s - mid;
6339: /* Scan column indices in this row and find out how many repeats each unique nonzero has */
6340: for (p = k; p < mid;) {
6341: col = j[p];
6342: q = p;
6343: do {
6344: p++;
6345: } while (p < mid && j[p] == col);
6346: Ajmap[Annz + 1] = Ajmap[Annz] + (p - q);
6347: Annz++;
6348: }
6350: for (p = mid; p < s;) {
6351: col = j[p];
6352: q = p;
6353: do {
6354: p++;
6355: } while (p < s && j[p] == col);
6356: Bjmap[Bnnz + 1] = Bjmap[Bnnz] + (p - q);
6357: Bnnz++;
6358: }
6359: }
6360: /* Output */
6361: *Aperm_ = Aperm;
6362: *Annz_ = Annz;
6363: *Atot_ = Atot;
6364: *Ajmap_ = Ajmap;
6365: *Bperm_ = Bperm;
6366: *Bnnz_ = Bnnz;
6367: *Btot_ = Btot;
6368: *Bjmap_ = Bjmap;
6369: PetscFunctionReturn(PETSC_SUCCESS);
6370: }
6372: /*
6373: Expand the jmap[] array to make a new one in view of nonzeros in the merged matrix
6375: Input Parameters:
6376: nnz1: number of unique nonzeros in a set that was used to produce imap[], jmap[]
6377: nnz: number of unique nonzeros in the merged matrix
6378: imap[nnz1]: i-th nonzero in the set is the imap[i]-th nonzero in the merged matrix
6379: jmap[nnz1+1]: i-th nonzero in the set has jmap[i+1] - jmap[i] repeats in the set
6381: Output Parameter: (memory is allocated by the caller)
6382: jmap_new[nnz+1]: i-th nonzero in the merged matrix has jmap_new[i+1] - jmap_new[i] repeats in the set
6384: Example:
6385: nnz1 = 4
6386: nnz = 6
6387: imap = [1,3,4,5]
6388: jmap = [0,3,5,6,7]
6389: then,
6390: jmap_new = [0,0,3,3,5,6,7]
6391: */
6392: static PetscErrorCode ExpandJmap_Internal(PetscCount nnz1, PetscCount nnz, const PetscCount imap[], const PetscCount jmap[], PetscCount jmap_new[])
6393: {
6394: PetscCount k, p;
6396: PetscFunctionBegin;
6397: jmap_new[0] = 0;
6398: p = nnz; /* p loops over jmap_new[] backwards */
6399: for (k = nnz1 - 1; k >= 0; k--) { /* k loops over imap[] */
6400: for (; p > imap[k]; p--) jmap_new[p] = jmap[k + 1];
6401: }
6402: for (; p >= 0; p--) jmap_new[p] = jmap[0];
6403: PetscFunctionReturn(PETSC_SUCCESS);
6404: }
6406: static PetscErrorCode MatCOOStructDestroy_MPIAIJ(void **data)
6407: {
6408: MatCOOStruct_MPIAIJ *coo = (MatCOOStruct_MPIAIJ *)*data;
6410: PetscFunctionBegin;
6411: PetscCall(PetscSFDestroy(&coo->sf));
6412: PetscCall(PetscFree(coo->Aperm1));
6413: PetscCall(PetscFree(coo->Bperm1));
6414: PetscCall(PetscFree(coo->Ajmap1));
6415: PetscCall(PetscFree(coo->Bjmap1));
6416: PetscCall(PetscFree(coo->Aimap2));
6417: PetscCall(PetscFree(coo->Bimap2));
6418: PetscCall(PetscFree(coo->Aperm2));
6419: PetscCall(PetscFree(coo->Bperm2));
6420: PetscCall(PetscFree(coo->Ajmap2));
6421: PetscCall(PetscFree(coo->Bjmap2));
6422: PetscCall(PetscFree(coo->Cperm1));
6423: PetscCall(PetscFree2(coo->sendbuf, coo->recvbuf));
6424: PetscCall(PetscFree(coo));
6425: PetscFunctionReturn(PETSC_SUCCESS);
6426: }
6428: PetscErrorCode MatSetPreallocationCOO_MPIAIJ(Mat mat, PetscCount coo_n, PetscInt coo_i[], PetscInt coo_j[])
6429: {
6430: MPI_Comm comm;
6431: PetscMPIInt rank, size;
6432: PetscInt m, n, M, N, rstart, rend, cstart, cend; /* Sizes, indices of row/col, therefore with type PetscInt */
6433: PetscCount k, p, q, rem; /* Loop variables over coo arrays */
6434: Mat_MPIAIJ *mpiaij = (Mat_MPIAIJ *)mat->data;
6435: PetscContainer container;
6436: MatCOOStruct_MPIAIJ *coo;
6438: PetscFunctionBegin;
6439: PetscCall(PetscFree(mpiaij->garray));
6440: PetscCall(VecDestroy(&mpiaij->lvec));
6441: #if defined(PETSC_USE_CTABLE)
6442: PetscCall(PetscHMapIDestroy(&mpiaij->colmap));
6443: #else
6444: PetscCall(PetscFree(mpiaij->colmap));
6445: #endif
6446: PetscCall(VecScatterDestroy(&mpiaij->Mvctx));
6447: mat->assembled = PETSC_FALSE;
6448: mat->was_assembled = PETSC_FALSE;
6450: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
6451: PetscCallMPI(MPI_Comm_size(comm, &size));
6452: PetscCallMPI(MPI_Comm_rank(comm, &rank));
6453: PetscCall(PetscLayoutSetUp(mat->rmap));
6454: PetscCall(PetscLayoutSetUp(mat->cmap));
6455: PetscCall(PetscLayoutGetRange(mat->rmap, &rstart, &rend));
6456: PetscCall(PetscLayoutGetRange(mat->cmap, &cstart, &cend));
6457: PetscCall(MatGetLocalSize(mat, &m, &n));
6458: PetscCall(MatGetSize(mat, &M, &N));
6460: /* Sort (i,j) by row along with a permutation array, so that the to-be-ignored */
6461: /* entries come first, then local rows, then remote rows. */
6462: PetscCount n1 = coo_n, *perm1;
6463: PetscInt *i1 = coo_i, *j1 = coo_j;
6465: PetscCall(PetscMalloc1(n1, &perm1));
6466: for (k = 0; k < n1; k++) perm1[k] = k;
6468: /* Manipulate indices so that entries with negative row or col indices will have smallest
6469: row indices, local entries will have greater but negative row indices, and remote entries
6470: will have positive row indices.
6471: */
6472: for (k = 0; k < n1; k++) {
6473: if (i1[k] < 0 || j1[k] < 0) i1[k] = PETSC_INT_MIN; /* e.g., -2^31, minimal to move them ahead */
6474: 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] */
6475: else {
6476: PetscCheck(!mat->nooffprocentries, PETSC_COMM_SELF, PETSC_ERR_USER_INPUT, "MAT_NO_OFF_PROC_ENTRIES is set but insert to remote rows");
6477: if (mpiaij->donotstash) i1[k] = PETSC_INT_MIN; /* Ignore offproc entries as if they had negative indices */
6478: }
6479: }
6481: /* Sort by row; after that, [0,k) have ignored entries, [k,rem) have local rows and [rem,n1) have remote rows */
6482: PetscCall(PetscSortIntWithIntCountArrayPair(n1, i1, j1, perm1));
6484: /* Advance k to the first entry we need to take care of */
6485: for (k = 0; k < n1; k++)
6486: if (i1[k] > PETSC_INT_MIN) break;
6487: PetscCount i1start = k;
6489: PetscCall(PetscSortedIntUpperBound(i1, k, n1, rend - 1 - PETSC_INT_MAX, &rem)); /* rem is upper bound of the last local row */
6490: for (; k < rem; k++) i1[k] += PETSC_INT_MAX; /* Revert row indices of local rows*/
6492: /* Send remote rows to their owner */
6493: /* Find which rows should be sent to which remote ranks*/
6494: PetscInt nsend = 0; /* Number of MPI ranks to send data to */
6495: PetscMPIInt *sendto; /* [nsend], storing remote ranks */
6496: PetscInt *nentries; /* [nsend], storing number of entries sent to remote ranks; Assume PetscInt is big enough for this count, and error if not */
6497: const PetscInt *ranges;
6498: PetscInt maxNsend = size >= 128 ? 128 : size; /* Assume max 128 neighbors; realloc when needed */
6500: PetscCall(PetscLayoutGetRanges(mat->rmap, &ranges));
6501: PetscCall(PetscMalloc2(maxNsend, &sendto, maxNsend, &nentries));
6502: for (k = rem; k < n1;) {
6503: PetscMPIInt owner;
6504: PetscInt firstRow, lastRow;
6506: /* Locate a row range */
6507: firstRow = i1[k]; /* first row of this owner */
6508: PetscCall(PetscLayoutFindOwner(mat->rmap, firstRow, &owner));
6509: lastRow = ranges[owner + 1] - 1; /* last row of this owner */
6511: /* Find the first index 'p' in [k,n) with i[p] belonging to next owner */
6512: PetscCall(PetscSortedIntUpperBound(i1, k, n1, lastRow, &p));
6514: /* All entries in [k,p) belong to this remote owner */
6515: if (nsend >= maxNsend) { /* Double the remote ranks arrays if not long enough */
6516: PetscMPIInt *sendto2;
6517: PetscInt *nentries2;
6518: PetscInt maxNsend2 = (maxNsend <= size / 2) ? maxNsend * 2 : size;
6520: PetscCall(PetscMalloc2(maxNsend2, &sendto2, maxNsend2, &nentries2));
6521: PetscCall(PetscArraycpy(sendto2, sendto, maxNsend));
6522: PetscCall(PetscArraycpy(nentries2, nentries2, maxNsend + 1));
6523: PetscCall(PetscFree2(sendto, nentries2));
6524: sendto = sendto2;
6525: nentries = nentries2;
6526: maxNsend = maxNsend2;
6527: }
6528: sendto[nsend] = owner;
6529: PetscCall(PetscIntCast(p - k, &nentries[nsend]));
6530: nsend++;
6531: k = p;
6532: }
6534: /* Build 1st SF to know offsets on remote to send data */
6535: PetscSF sf1;
6536: PetscInt nroots = 1, nroots2 = 0;
6537: PetscInt nleaves = nsend, nleaves2 = 0;
6538: PetscInt *offsets;
6539: PetscSFNode *iremote;
6541: PetscCall(PetscSFCreate(comm, &sf1));
6542: PetscCall(PetscMalloc1(nsend, &iremote));
6543: PetscCall(PetscMalloc1(nsend, &offsets));
6544: for (k = 0; k < nsend; k++) {
6545: iremote[k].rank = sendto[k];
6546: iremote[k].index = 0;
6547: nleaves2 += nentries[k];
6548: PetscCheck(nleaves2 >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Number of SF leaves is too large for PetscInt");
6549: }
6550: PetscCall(PetscSFSetGraph(sf1, nroots, nleaves, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
6551: PetscCall(PetscSFFetchAndOpWithMemTypeBegin(sf1, MPIU_INT, PETSC_MEMTYPE_HOST, &nroots2 /*rootdata*/, PETSC_MEMTYPE_HOST, nentries /*leafdata*/, PETSC_MEMTYPE_HOST, offsets /*leafupdate*/, MPI_SUM));
6552: PetscCall(PetscSFFetchAndOpEnd(sf1, MPIU_INT, &nroots2, nentries, offsets, MPI_SUM)); /* Would nroots2 overflow, we check offsets[] below */
6553: PetscCall(PetscSFDestroy(&sf1));
6554: PetscAssert(nleaves2 == n1 - rem, PETSC_COMM_SELF, PETSC_ERR_PLIB, "nleaves2 %" PetscInt_FMT " != number of remote entries %" PetscCount_FMT, nleaves2, n1 - rem);
6556: /* Build 2nd SF to send remote COOs to their owner */
6557: PetscSF sf2;
6558: nroots = nroots2;
6559: nleaves = nleaves2;
6560: PetscCall(PetscSFCreate(comm, &sf2));
6561: PetscCall(PetscSFSetFromOptions(sf2));
6562: PetscCall(PetscMalloc1(nleaves, &iremote));
6563: p = 0;
6564: for (k = 0; k < nsend; k++) {
6565: PetscCheck(offsets[k] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Number of SF roots is too large for PetscInt");
6566: for (q = 0; q < nentries[k]; q++, p++) {
6567: iremote[p].rank = sendto[k];
6568: PetscCall(PetscIntCast(offsets[k] + q, &iremote[p].index));
6569: }
6570: }
6571: PetscCall(PetscSFSetGraph(sf2, nroots, nleaves, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
6573: /* Send the remote COOs to their owner */
6574: PetscInt n2 = nroots, *i2, *j2; /* Buffers for received COOs from other ranks, along with a permutation array */
6575: PetscCount *perm2; /* Though PetscInt is enough for remote entries, we use PetscCount here as we want to reuse MatSplitEntries_Internal() */
6576: PetscCall(PetscMalloc3(n2, &i2, n2, &j2, n2, &perm2));
6577: PetscAssert(rem == 0 || i1 != NULL, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot add nonzero offset to null");
6578: PetscAssert(rem == 0 || j1 != NULL, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot add nonzero offset to null");
6579: PetscInt *i1prem = PetscSafePointerPlusOffset(i1, rem);
6580: PetscInt *j1prem = PetscSafePointerPlusOffset(j1, rem);
6581: PetscCall(PetscSFReduceWithMemTypeBegin(sf2, MPIU_INT, PETSC_MEMTYPE_HOST, i1prem, PETSC_MEMTYPE_HOST, i2, MPI_REPLACE));
6582: PetscCall(PetscSFReduceEnd(sf2, MPIU_INT, i1prem, i2, MPI_REPLACE));
6583: PetscCall(PetscSFReduceWithMemTypeBegin(sf2, MPIU_INT, PETSC_MEMTYPE_HOST, j1prem, PETSC_MEMTYPE_HOST, j2, MPI_REPLACE));
6584: PetscCall(PetscSFReduceEnd(sf2, MPIU_INT, j1prem, j2, MPI_REPLACE));
6586: PetscCall(PetscFree(offsets));
6587: PetscCall(PetscFree2(sendto, nentries));
6589: /* Sort received COOs by row along with the permutation array */
6590: for (k = 0; k < n2; k++) perm2[k] = k;
6591: PetscCall(PetscSortIntWithIntCountArrayPair(n2, i2, j2, perm2));
6593: /* sf2 only sends contiguous leafdata to contiguous rootdata. We record the permutation which will be used to fill leafdata */
6594: PetscCount *Cperm1;
6595: PetscAssert(rem == 0 || perm1 != NULL, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot add nonzero offset to null");
6596: PetscCount *perm1prem = PetscSafePointerPlusOffset(perm1, rem);
6597: PetscCall(PetscMalloc1(nleaves, &Cperm1));
6598: PetscCall(PetscArraycpy(Cperm1, perm1prem, nleaves));
6600: /* Support for HYPRE matrices, kind of a hack.
6601: Swap min column with diagonal so that diagonal values will go first */
6602: PetscBool hypre;
6603: PetscCall(PetscStrcmp("_internal_COO_mat_for_hypre", ((PetscObject)mat)->name, &hypre));
6604: if (hypre) {
6605: PetscInt *minj;
6606: PetscBT hasdiag;
6608: PetscCall(PetscBTCreate(m, &hasdiag));
6609: PetscCall(PetscMalloc1(m, &minj));
6610: for (k = 0; k < m; k++) minj[k] = PETSC_INT_MAX;
6611: for (k = i1start; k < rem; k++) {
6612: if (j1[k] < cstart || j1[k] >= cend) continue;
6613: const PetscInt rindex = i1[k] - rstart;
6614: if ((j1[k] - cstart) == rindex) PetscCall(PetscBTSet(hasdiag, rindex));
6615: minj[rindex] = PetscMin(minj[rindex], j1[k]);
6616: }
6617: for (k = 0; k < n2; k++) {
6618: if (j2[k] < cstart || j2[k] >= cend) continue;
6619: const PetscInt rindex = i2[k] - rstart;
6620: if ((j2[k] - cstart) == rindex) PetscCall(PetscBTSet(hasdiag, rindex));
6621: minj[rindex] = PetscMin(minj[rindex], j2[k]);
6622: }
6623: for (k = i1start; k < rem; k++) {
6624: const PetscInt rindex = i1[k] - rstart;
6625: if (j1[k] < cstart || j1[k] >= cend || !PetscBTLookup(hasdiag, rindex)) continue;
6626: if (j1[k] == minj[rindex]) j1[k] = i1[k] + (cstart - rstart);
6627: else if ((j1[k] - cstart) == rindex) j1[k] = minj[rindex];
6628: }
6629: for (k = 0; k < n2; k++) {
6630: const PetscInt rindex = i2[k] - rstart;
6631: if (j2[k] < cstart || j2[k] >= cend || !PetscBTLookup(hasdiag, rindex)) continue;
6632: if (j2[k] == minj[rindex]) j2[k] = i2[k] + (cstart - rstart);
6633: else if ((j2[k] - cstart) == rindex) j2[k] = minj[rindex];
6634: }
6635: PetscCall(PetscBTDestroy(&hasdiag));
6636: PetscCall(PetscFree(minj));
6637: }
6639: /* Split local COOs and received COOs into diag/offdiag portions */
6640: PetscCount *rowBegin1, *rowMid1, *rowEnd1;
6641: PetscCount *Ajmap1, *Aperm1, *Bjmap1, *Bperm1;
6642: PetscCount Annz1, Bnnz1, Atot1, Btot1;
6643: PetscCount *rowBegin2, *rowMid2, *rowEnd2;
6644: PetscCount *Ajmap2, *Aperm2, *Bjmap2, *Bperm2;
6645: PetscCount Annz2, Bnnz2, Atot2, Btot2;
6647: PetscCall(PetscCalloc3(m, &rowBegin1, m, &rowMid1, m, &rowEnd1));
6648: PetscCall(PetscCalloc3(m, &rowBegin2, m, &rowMid2, m, &rowEnd2));
6649: PetscCall(MatSplitEntries_Internal(mat, rem, i1, j1, perm1, rowBegin1, rowMid1, rowEnd1, &Atot1, &Aperm1, &Annz1, &Ajmap1, &Btot1, &Bperm1, &Bnnz1, &Bjmap1));
6650: PetscCall(MatSplitEntries_Internal(mat, n2, i2, j2, perm2, rowBegin2, rowMid2, rowEnd2, &Atot2, &Aperm2, &Annz2, &Ajmap2, &Btot2, &Bperm2, &Bnnz2, &Bjmap2));
6652: /* Merge local COOs with received COOs: diag with diag, offdiag with offdiag */
6653: PetscInt *Ai, *Bi;
6654: PetscInt *Aj, *Bj;
6656: PetscCall(PetscMalloc1(m + 1, &Ai));
6657: PetscCall(PetscMalloc1(m + 1, &Bi));
6658: PetscCall(PetscMalloc1(Annz1 + Annz2, &Aj)); /* Since local and remote entries might have dups, we might allocate excess memory */
6659: PetscCall(PetscMalloc1(Bnnz1 + Bnnz2, &Bj));
6661: PetscCount *Aimap1, *Bimap1, *Aimap2, *Bimap2;
6662: PetscCall(PetscMalloc1(Annz1, &Aimap1));
6663: PetscCall(PetscMalloc1(Bnnz1, &Bimap1));
6664: PetscCall(PetscMalloc1(Annz2, &Aimap2));
6665: PetscCall(PetscMalloc1(Bnnz2, &Bimap2));
6667: PetscCall(MatMergeEntries_Internal(mat, j1, j2, rowBegin1, rowMid1, rowBegin2, rowMid2, Ajmap1, Ajmap2, Aimap1, Aimap2, Ai, Aj));
6668: PetscCall(MatMergeEntries_Internal(mat, j1, j2, rowMid1, rowEnd1, rowMid2, rowEnd2, Bjmap1, Bjmap2, Bimap1, Bimap2, Bi, Bj));
6670: /* Expand Ajmap1/Bjmap1 to make them based off nonzeros in A/B, since we */
6671: /* expect nonzeros in A/B most likely have local contributing entries */
6672: PetscInt Annz = Ai[m];
6673: PetscInt Bnnz = Bi[m];
6674: PetscCount *Ajmap1_new, *Bjmap1_new;
6676: PetscCall(PetscMalloc1(Annz + 1, &Ajmap1_new));
6677: PetscCall(PetscMalloc1(Bnnz + 1, &Bjmap1_new));
6679: PetscCall(ExpandJmap_Internal(Annz1, Annz, Aimap1, Ajmap1, Ajmap1_new));
6680: PetscCall(ExpandJmap_Internal(Bnnz1, Bnnz, Bimap1, Bjmap1, Bjmap1_new));
6682: PetscCall(PetscFree(Aimap1));
6683: PetscCall(PetscFree(Ajmap1));
6684: PetscCall(PetscFree(Bimap1));
6685: PetscCall(PetscFree(Bjmap1));
6686: PetscCall(PetscFree3(rowBegin1, rowMid1, rowEnd1));
6687: PetscCall(PetscFree3(rowBegin2, rowMid2, rowEnd2));
6688: PetscCall(PetscFree(perm1));
6689: PetscCall(PetscFree3(i2, j2, perm2));
6691: Ajmap1 = Ajmap1_new;
6692: Bjmap1 = Bjmap1_new;
6694: /* Reallocate Aj, Bj once we know actual numbers of unique nonzeros in A and B */
6695: if (Annz < Annz1 + Annz2) {
6696: PetscInt *Aj_new;
6697: PetscCall(PetscMalloc1(Annz, &Aj_new));
6698: PetscCall(PetscArraycpy(Aj_new, Aj, Annz));
6699: PetscCall(PetscFree(Aj));
6700: Aj = Aj_new;
6701: }
6703: if (Bnnz < Bnnz1 + Bnnz2) {
6704: PetscInt *Bj_new;
6705: PetscCall(PetscMalloc1(Bnnz, &Bj_new));
6706: PetscCall(PetscArraycpy(Bj_new, Bj, Bnnz));
6707: PetscCall(PetscFree(Bj));
6708: Bj = Bj_new;
6709: }
6711: /* Create new submatrices for on-process and off-process coupling */
6712: PetscScalar *Aa, *Ba;
6713: MatType rtype;
6714: Mat_SeqAIJ *a, *b;
6715: PetscObjectState state;
6716: PetscCall(PetscCalloc1(Annz, &Aa)); /* Zero matrix on device */
6717: PetscCall(PetscCalloc1(Bnnz, &Ba));
6718: /* make Aj[] local, i.e, based off the start column of the diagonal portion */
6719: if (cstart) {
6720: for (k = 0; k < Annz; k++) Aj[k] -= cstart;
6721: }
6723: PetscCall(MatGetRootType_Private(mat, &rtype));
6725: MatSeqXAIJGetOptions_Private(mpiaij->A);
6726: PetscCall(MatDestroy(&mpiaij->A));
6727: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, n, Ai, Aj, Aa, &mpiaij->A));
6728: PetscCall(MatSetBlockSizesFromMats(mpiaij->A, mat, mat));
6729: MatSeqXAIJRestoreOptions_Private(mpiaij->A);
6731: MatSeqXAIJGetOptions_Private(mpiaij->B);
6732: PetscCall(MatDestroy(&mpiaij->B));
6733: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, mat->cmap->N, Bi, Bj, Ba, &mpiaij->B));
6734: PetscCall(MatSetBlockSizesFromMats(mpiaij->B, mat, mat));
6735: MatSeqXAIJRestoreOptions_Private(mpiaij->B);
6737: PetscCall(MatSetUpMultiply_MPIAIJ(mat));
6738: mat->was_assembled = PETSC_TRUE; // was_assembled in effect means the Mvctx is built; doing so avoids redundant MatSetUpMultiply_MPIAIJ
6739: state = mpiaij->A->nonzerostate + mpiaij->B->nonzerostate;
6740: PetscCallMPI(MPIU_Allreduce(&state, &mat->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)mat)));
6742: a = (Mat_SeqAIJ *)mpiaij->A->data;
6743: b = (Mat_SeqAIJ *)mpiaij->B->data;
6744: a->free_a = PETSC_TRUE;
6745: a->free_ij = PETSC_TRUE;
6746: b->free_a = PETSC_TRUE;
6747: b->free_ij = PETSC_TRUE;
6748: a->maxnz = a->nz;
6749: b->maxnz = b->nz;
6751: /* conversion must happen AFTER multiply setup */
6752: PetscCall(MatConvert(mpiaij->A, rtype, MAT_INPLACE_MATRIX, &mpiaij->A));
6753: PetscCall(MatConvert(mpiaij->B, rtype, MAT_INPLACE_MATRIX, &mpiaij->B));
6754: PetscCall(VecDestroy(&mpiaij->lvec));
6755: PetscCall(MatCreateVecs(mpiaij->B, &mpiaij->lvec, NULL));
6757: // Put the COO struct in a container and then attach that to the matrix
6758: PetscCall(PetscMalloc1(1, &coo));
6759: coo->n = coo_n;
6760: coo->sf = sf2;
6761: coo->sendlen = nleaves;
6762: coo->recvlen = nroots;
6763: coo->Annz = Annz;
6764: coo->Bnnz = Bnnz;
6765: coo->Annz2 = Annz2;
6766: coo->Bnnz2 = Bnnz2;
6767: coo->Atot1 = Atot1;
6768: coo->Atot2 = Atot2;
6769: coo->Btot1 = Btot1;
6770: coo->Btot2 = Btot2;
6771: coo->Ajmap1 = Ajmap1;
6772: coo->Aperm1 = Aperm1;
6773: coo->Bjmap1 = Bjmap1;
6774: coo->Bperm1 = Bperm1;
6775: coo->Aimap2 = Aimap2;
6776: coo->Ajmap2 = Ajmap2;
6777: coo->Aperm2 = Aperm2;
6778: coo->Bimap2 = Bimap2;
6779: coo->Bjmap2 = Bjmap2;
6780: coo->Bperm2 = Bperm2;
6781: coo->Cperm1 = Cperm1;
6782: // Allocate in preallocation. If not used, it has zero cost on host
6783: PetscCall(PetscMalloc2(coo->sendlen, &coo->sendbuf, coo->recvlen, &coo->recvbuf));
6784: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
6785: PetscCall(PetscContainerSetPointer(container, coo));
6786: PetscCall(PetscContainerSetCtxDestroy(container, MatCOOStructDestroy_MPIAIJ));
6787: PetscCall(PetscObjectCompose((PetscObject)mat, "__PETSc_MatCOOStruct_Host", (PetscObject)container));
6788: PetscCall(PetscContainerDestroy(&container));
6789: PetscFunctionReturn(PETSC_SUCCESS);
6790: }
6792: static PetscErrorCode MatSetValuesCOO_MPIAIJ(Mat mat, const PetscScalar v[], InsertMode imode)
6793: {
6794: Mat_MPIAIJ *mpiaij = (Mat_MPIAIJ *)mat->data;
6795: Mat A = mpiaij->A, B = mpiaij->B;
6796: PetscScalar *Aa, *Ba;
6797: PetscScalar *sendbuf, *recvbuf;
6798: const PetscCount *Ajmap1, *Ajmap2, *Aimap2;
6799: const PetscCount *Bjmap1, *Bjmap2, *Bimap2;
6800: const PetscCount *Aperm1, *Aperm2, *Bperm1, *Bperm2;
6801: const PetscCount *Cperm1;
6802: PetscContainer container;
6803: MatCOOStruct_MPIAIJ *coo;
6805: PetscFunctionBegin;
6806: PetscCall(PetscObjectQuery((PetscObject)mat, "__PETSc_MatCOOStruct_Host", (PetscObject *)&container));
6807: PetscCheck(container, PetscObjectComm((PetscObject)mat), PETSC_ERR_PLIB, "Not found MatCOOStruct on this matrix");
6808: PetscCall(PetscContainerGetPointer(container, (void **)&coo));
6809: sendbuf = coo->sendbuf;
6810: recvbuf = coo->recvbuf;
6811: Ajmap1 = coo->Ajmap1;
6812: Ajmap2 = coo->Ajmap2;
6813: Aimap2 = coo->Aimap2;
6814: Bjmap1 = coo->Bjmap1;
6815: Bjmap2 = coo->Bjmap2;
6816: Bimap2 = coo->Bimap2;
6817: Aperm1 = coo->Aperm1;
6818: Aperm2 = coo->Aperm2;
6819: Bperm1 = coo->Bperm1;
6820: Bperm2 = coo->Bperm2;
6821: Cperm1 = coo->Cperm1;
6823: PetscCall(MatSeqAIJGetArray(A, &Aa)); /* Might read and write matrix values */
6824: PetscCall(MatSeqAIJGetArray(B, &Ba));
6826: /* Pack entries to be sent to remote */
6827: for (PetscCount i = 0; i < coo->sendlen; i++) sendbuf[i] = v[Cperm1[i]];
6829: /* Send remote entries to their owner and overlap the communication with local computation */
6830: PetscCall(PetscSFReduceWithMemTypeBegin(coo->sf, MPIU_SCALAR, PETSC_MEMTYPE_HOST, sendbuf, PETSC_MEMTYPE_HOST, recvbuf, MPI_REPLACE));
6831: /* Add local entries to A and B */
6832: for (PetscCount i = 0; i < coo->Annz; i++) { /* All nonzeros in A are either zero'ed or added with a value (i.e., initialized) */
6833: PetscScalar sum = 0.0; /* Do partial summation first to improve numerical stability */
6834: for (PetscCount k = Ajmap1[i]; k < Ajmap1[i + 1]; k++) sum += v[Aperm1[k]];
6835: Aa[i] = (imode == INSERT_VALUES ? 0.0 : Aa[i]) + sum;
6836: }
6837: for (PetscCount i = 0; i < coo->Bnnz; i++) {
6838: PetscScalar sum = 0.0;
6839: for (PetscCount k = Bjmap1[i]; k < Bjmap1[i + 1]; k++) sum += v[Bperm1[k]];
6840: Ba[i] = (imode == INSERT_VALUES ? 0.0 : Ba[i]) + sum;
6841: }
6842: PetscCall(PetscSFReduceEnd(coo->sf, MPIU_SCALAR, sendbuf, recvbuf, MPI_REPLACE));
6844: /* Add received remote entries to A and B */
6845: for (PetscCount i = 0; i < coo->Annz2; i++) {
6846: for (PetscCount k = Ajmap2[i]; k < Ajmap2[i + 1]; k++) Aa[Aimap2[i]] += recvbuf[Aperm2[k]];
6847: }
6848: for (PetscCount i = 0; i < coo->Bnnz2; i++) {
6849: for (PetscCount k = Bjmap2[i]; k < Bjmap2[i + 1]; k++) Ba[Bimap2[i]] += recvbuf[Bperm2[k]];
6850: }
6851: PetscCall(MatSeqAIJRestoreArray(A, &Aa));
6852: PetscCall(MatSeqAIJRestoreArray(B, &Ba));
6853: PetscFunctionReturn(PETSC_SUCCESS);
6854: }
6856: /*MC
6857: MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices.
6859: Options Database Keys:
6860: . -mat_type mpiaij - sets the matrix type to `MATMPIAIJ` during a call to `MatSetFromOptions()`
6862: Level: beginner
6864: Notes:
6865: `MatSetValues()` may be called for this matrix type with a `NULL` argument for the numerical values,
6866: in this case the values associated with the rows and columns one passes in are set to zero
6867: in the matrix
6869: `MatSetOptions`(,`MAT_STRUCTURE_ONLY`,`PETSC_TRUE`) may be called for this matrix type. In this no
6870: space is allocated for the nonzero entries and any entries passed with `MatSetValues()` are ignored
6872: .seealso: [](ch_matrices), `Mat`, `MATSEQAIJ`, `MATAIJ`, `MatCreateAIJ()`
6873: M*/
6874: PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat B)
6875: {
6876: Mat_MPIAIJ *b;
6877: PetscMPIInt size;
6879: PetscFunctionBegin;
6880: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
6882: PetscCall(PetscNew(&b));
6883: B->data = (void *)b;
6884: B->ops[0] = MatOps_Values;
6885: B->assembled = PETSC_FALSE;
6886: B->insertmode = NOT_SET_VALUES;
6887: b->size = size;
6889: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)B), &b->rank));
6891: /* build cache for off array entries formed */
6892: PetscCall(MatStashCreate_Private(PetscObjectComm((PetscObject)B), 1, &B->stash));
6894: b->donotstash = PETSC_FALSE;
6895: b->colmap = NULL;
6896: b->garray = NULL;
6897: b->roworiented = PETSC_TRUE;
6899: /* stuff used for matrix vector multiply */
6900: b->lvec = NULL;
6901: b->Mvctx = NULL;
6903: /* stuff for MatGetRow() */
6904: b->rowindices = NULL;
6905: b->rowvalues = NULL;
6906: b->getrowactive = PETSC_FALSE;
6908: /* flexible pointer used in CUSPARSE classes */
6909: b->spptr = NULL;
6911: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetUseScalableIncreaseOverlap_C", MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ));
6912: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatStoreValues_C", MatStoreValues_MPIAIJ));
6913: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatRetrieveValues_C", MatRetrieveValues_MPIAIJ));
6914: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatIsTranspose_C", MatIsTranspose_MPIAIJ));
6915: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetPreallocation_C", MatMPIAIJSetPreallocation_MPIAIJ));
6916: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatResetPreallocation_C", MatResetPreallocation_MPIAIJ));
6917: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatResetHash_C", MatResetHash_MPIAIJ));
6918: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetPreallocationCSR_C", MatMPIAIJSetPreallocationCSR_MPIAIJ));
6919: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDiagonalScaleLocal_C", MatDiagonalScaleLocal_MPIAIJ));
6920: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijperm_C", MatConvert_MPIAIJ_MPIAIJPERM));
6921: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijsell_C", MatConvert_MPIAIJ_MPIAIJSELL));
6922: #if defined(PETSC_HAVE_CUDA)
6923: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijcusparse_C", MatConvert_MPIAIJ_MPIAIJCUSPARSE));
6924: #endif
6925: #if defined(PETSC_HAVE_HIP)
6926: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijhipsparse_C", MatConvert_MPIAIJ_MPIAIJHIPSPARSE));
6927: #endif
6928: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
6929: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijkokkos_C", MatConvert_MPIAIJ_MPIAIJKokkos));
6930: #endif
6931: #if defined(PETSC_HAVE_MKL_SPARSE)
6932: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijmkl_C", MatConvert_MPIAIJ_MPIAIJMKL));
6933: #endif
6934: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijcrl_C", MatConvert_MPIAIJ_MPIAIJCRL));
6935: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpibaij_C", MatConvert_MPIAIJ_MPIBAIJ));
6936: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpisbaij_C", MatConvert_MPIAIJ_MPISBAIJ));
6937: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpidense_C", MatConvert_MPIAIJ_MPIDense));
6938: #if defined(PETSC_HAVE_ELEMENTAL)
6939: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_elemental_C", MatConvert_MPIAIJ_Elemental));
6940: #endif
6941: #if defined(PETSC_HAVE_SCALAPACK)
6942: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_scalapack_C", MatConvert_AIJ_ScaLAPACK));
6943: #endif
6944: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_is_C", MatConvert_XAIJ_IS));
6945: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpisell_C", MatConvert_MPIAIJ_MPISELL));
6946: #if defined(PETSC_HAVE_HYPRE)
6947: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_hypre_C", MatConvert_AIJ_HYPRE));
6948: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_transpose_mpiaij_mpiaij_C", MatProductSetFromOptions_Transpose_AIJ_AIJ));
6949: #endif
6950: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_is_mpiaij_C", MatProductSetFromOptions_IS_XAIJ));
6951: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_mpiaij_mpiaij_C", MatProductSetFromOptions_MPIAIJ));
6952: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetPreallocationCOO_C", MatSetPreallocationCOO_MPIAIJ));
6953: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetValuesCOO_C", MatSetValuesCOO_MPIAIJ));
6954: PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATMPIAIJ));
6955: PetscFunctionReturn(PETSC_SUCCESS);
6956: }
6958: /*@
6959: MatCreateMPIAIJWithSplitArrays - creates a `MATMPIAIJ` matrix using arrays that contain the "diagonal"
6960: and "off-diagonal" part of the matrix in CSR format.
6962: Collective
6964: Input Parameters:
6965: + comm - MPI communicator
6966: . m - number of local rows (Cannot be `PETSC_DECIDE`)
6967: . n - This value should be the same as the local size used in creating the
6968: x vector for the matrix-vector product $y = Ax$. (or `PETSC_DECIDE` to have
6969: calculated if `N` is given) For square matrices `n` is almost always `m`.
6970: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if `m` is given)
6971: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if `n` is given)
6972: . 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
6973: . j - column indices, which must be local, i.e., based off the start column of the diagonal portion
6974: . a - matrix values
6975: . 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
6976: . oj - column indices, which must be global, representing global columns in the `MATMPIAIJ` matrix
6977: - oa - matrix values
6979: Output Parameter:
6980: . mat - the matrix
6982: Level: advanced
6984: Notes:
6985: The `i`, `j`, and `a` arrays ARE NOT copied by this routine into the internal format used by PETSc (even in Fortran). The user
6986: must free the arrays once the matrix has been destroyed and not before.
6988: The `i` and `j` indices are 0 based
6990: See `MatCreateAIJ()` for the definition of "diagonal" and "off-diagonal" portion of the matrix
6992: This sets local rows and cannot be used to set off-processor values.
6994: Use of this routine is discouraged because it is inflexible and cumbersome to use. It is extremely rare that a
6995: legacy application natively assembles into exactly this split format. The code to do so is nontrivial and does
6996: not easily support in-place reassembly. It is recommended to use MatSetValues() (or a variant thereof) because
6997: the resulting assembly is easier to implement, will work with any matrix format, and the user does not have to
6998: keep track of the underlying array. Use `MatSetOption`(A,`MAT_NO_OFF_PROC_ENTRIES`,`PETSC_TRUE`) to disable all
6999: communication if it is known that only local entries will be set.
7001: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
7002: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithArrays()`
7003: @*/
7004: 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)
7005: {
7006: Mat_MPIAIJ *maij;
7008: PetscFunctionBegin;
7009: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
7010: PetscCheck(i[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
7011: PetscCheck(oi[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "oi (row indices) must start with 0");
7012: PetscCall(MatCreate(comm, mat));
7013: PetscCall(MatSetSizes(*mat, m, n, M, N));
7014: PetscCall(MatSetType(*mat, MATMPIAIJ));
7015: maij = (Mat_MPIAIJ *)(*mat)->data;
7017: (*mat)->preallocated = PETSC_TRUE;
7019: PetscCall(PetscLayoutSetUp((*mat)->rmap));
7020: PetscCall(PetscLayoutSetUp((*mat)->cmap));
7022: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, n, i, j, a, &maij->A));
7023: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, (*mat)->cmap->N, oi, oj, oa, &maij->B));
7025: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
7026: PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY));
7027: PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY));
7028: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_FALSE));
7029: PetscCall(MatSetOption(*mat, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
7030: PetscFunctionReturn(PETSC_SUCCESS);
7031: }
7033: typedef struct {
7034: Mat *mp; /* intermediate products */
7035: PetscBool *mptmp; /* is the intermediate product temporary ? */
7036: PetscInt cp; /* number of intermediate products */
7038: /* support for MatGetBrowsOfAoCols_MPIAIJ for P_oth */
7039: PetscInt *startsj_s, *startsj_r;
7040: PetscScalar *bufa;
7041: Mat P_oth;
7043: /* may take advantage of merging product->B */
7044: Mat Bloc; /* B-local by merging diag and off-diag */
7046: /* cusparse does not have support to split between symbolic and numeric phases.
7047: When api_user is true, we don't need to update the numerical values
7048: of the temporary storage */
7049: PetscBool reusesym;
7051: /* support for COO values insertion */
7052: PetscScalar *coo_v, *coo_w; /* store on-process and off-process COO scalars, and used as MPI recv/send buffers respectively */
7053: PetscInt **own; /* own[i] points to address of on-process COO indices for Mat mp[i] */
7054: PetscInt **off; /* off[i] points to address of off-process COO indices for Mat mp[i] */
7055: PetscBool hasoffproc; /* if true, have off-process values insertion (i.e. AtB or PtAP) */
7056: PetscSF sf; /* used for non-local values insertion and memory malloc */
7057: PetscMemType mtype;
7059: /* customization */
7060: PetscBool abmerge;
7061: PetscBool P_oth_bind;
7062: } MatMatMPIAIJBACKEND;
7064: static PetscErrorCode MatDestroy_MatMatMPIAIJBACKEND(void *data)
7065: {
7066: MatMatMPIAIJBACKEND *mmdata = (MatMatMPIAIJBACKEND *)data;
7067: PetscInt i;
7069: PetscFunctionBegin;
7070: PetscCall(PetscFree2(mmdata->startsj_s, mmdata->startsj_r));
7071: PetscCall(PetscFree(mmdata->bufa));
7072: PetscCall(PetscSFFree(mmdata->sf, mmdata->mtype, mmdata->coo_v));
7073: PetscCall(PetscSFFree(mmdata->sf, mmdata->mtype, mmdata->coo_w));
7074: PetscCall(MatDestroy(&mmdata->P_oth));
7075: PetscCall(MatDestroy(&mmdata->Bloc));
7076: PetscCall(PetscSFDestroy(&mmdata->sf));
7077: for (i = 0; i < mmdata->cp; i++) PetscCall(MatDestroy(&mmdata->mp[i]));
7078: PetscCall(PetscFree2(mmdata->mp, mmdata->mptmp));
7079: PetscCall(PetscFree(mmdata->own[0]));
7080: PetscCall(PetscFree(mmdata->own));
7081: PetscCall(PetscFree(mmdata->off[0]));
7082: PetscCall(PetscFree(mmdata->off));
7083: PetscCall(PetscFree(mmdata));
7084: PetscFunctionReturn(PETSC_SUCCESS);
7085: }
7087: /* Copy selected n entries with indices in idx[] of A to v[].
7088: If idx is NULL, copy the whole data array of A to v[]
7089: */
7090: static PetscErrorCode MatSeqAIJCopySubArray(Mat A, PetscInt n, const PetscInt idx[], PetscScalar v[])
7091: {
7092: PetscErrorCode (*f)(Mat, PetscInt, const PetscInt[], PetscScalar[]);
7094: PetscFunctionBegin;
7095: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatSeqAIJCopySubArray_C", &f));
7096: if (f) {
7097: PetscCall((*f)(A, n, idx, v));
7098: } else {
7099: const PetscScalar *vv;
7101: PetscCall(MatSeqAIJGetArrayRead(A, &vv));
7102: if (n && idx) {
7103: PetscScalar *w = v;
7104: const PetscInt *oi = idx;
7105: PetscInt j;
7107: for (j = 0; j < n; j++) *w++ = vv[*oi++];
7108: } else {
7109: PetscCall(PetscArraycpy(v, vv, n));
7110: }
7111: PetscCall(MatSeqAIJRestoreArrayRead(A, &vv));
7112: }
7113: PetscFunctionReturn(PETSC_SUCCESS);
7114: }
7116: static PetscErrorCode MatProductNumeric_MPIAIJBACKEND(Mat C)
7117: {
7118: MatMatMPIAIJBACKEND *mmdata;
7119: PetscInt i, n_d, n_o;
7121: PetscFunctionBegin;
7122: MatCheckProduct(C, 1);
7123: PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data empty");
7124: mmdata = (MatMatMPIAIJBACKEND *)C->product->data;
7125: if (!mmdata->reusesym) { /* update temporary matrices */
7126: 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));
7127: if (mmdata->Bloc) PetscCall(MatMPIAIJGetLocalMatMerge(C->product->B, MAT_REUSE_MATRIX, NULL, &mmdata->Bloc));
7128: }
7129: mmdata->reusesym = PETSC_FALSE;
7131: for (i = 0; i < mmdata->cp; i++) {
7132: 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]);
7133: PetscCall((*mmdata->mp[i]->ops->productnumeric)(mmdata->mp[i]));
7134: }
7135: for (i = 0, n_d = 0, n_o = 0; i < mmdata->cp; i++) {
7136: PetscInt noff;
7138: PetscCall(PetscIntCast(mmdata->off[i + 1] - mmdata->off[i], &noff));
7139: if (mmdata->mptmp[i]) continue;
7140: if (noff) {
7141: PetscInt nown;
7143: PetscCall(PetscIntCast(mmdata->own[i + 1] - mmdata->own[i], &nown));
7144: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], noff, mmdata->off[i], mmdata->coo_w + n_o));
7145: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], nown, mmdata->own[i], mmdata->coo_v + n_d));
7146: n_o += noff;
7147: n_d += nown;
7148: } else {
7149: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mmdata->mp[i]->data;
7151: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], mm->nz, NULL, mmdata->coo_v + n_d));
7152: n_d += mm->nz;
7153: }
7154: }
7155: if (mmdata->hasoffproc) { /* offprocess insertion */
7156: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_SCALAR, mmdata->coo_w, mmdata->coo_v + n_d));
7157: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_SCALAR, mmdata->coo_w, mmdata->coo_v + n_d));
7158: }
7159: PetscCall(MatSetValuesCOO(C, mmdata->coo_v, INSERT_VALUES));
7160: PetscFunctionReturn(PETSC_SUCCESS);
7161: }
7163: /* Support for Pt * A, A * P, or Pt * A * P */
7164: #define MAX_NUMBER_INTERMEDIATE 4
7165: PetscErrorCode MatProductSymbolic_MPIAIJBACKEND(Mat C)
7166: {
7167: Mat_Product *product = C->product;
7168: Mat A, P, mp[MAX_NUMBER_INTERMEDIATE]; /* A, P and a series of intermediate matrices */
7169: Mat_MPIAIJ *a, *p;
7170: MatMatMPIAIJBACKEND *mmdata;
7171: ISLocalToGlobalMapping P_oth_l2g = NULL;
7172: IS glob = NULL;
7173: const char *prefix;
7174: char pprefix[256];
7175: const PetscInt *globidx, *P_oth_idx;
7176: PetscInt i, j, cp, m, n, M, N, *coo_i, *coo_j;
7177: PetscCount ncoo, ncoo_d, ncoo_o, ncoo_oown;
7178: PetscInt cmapt[MAX_NUMBER_INTERMEDIATE], rmapt[MAX_NUMBER_INTERMEDIATE]; /* col/row map type for each Mat in mp[]. */
7179: /* type-0: consecutive, start from 0; type-1: consecutive with */
7180: /* a base offset; type-2: sparse with a local to global map table */
7181: const PetscInt *cmapa[MAX_NUMBER_INTERMEDIATE], *rmapa[MAX_NUMBER_INTERMEDIATE]; /* col/row local to global map array (table) for type-2 map type */
7183: MatProductType ptype;
7184: PetscBool mptmp[MAX_NUMBER_INTERMEDIATE], hasoffproc = PETSC_FALSE, iscuda, iship, iskokk;
7185: PetscMPIInt size;
7187: PetscFunctionBegin;
7188: MatCheckProduct(C, 1);
7189: PetscCheck(!product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data not empty");
7190: ptype = product->type;
7191: if (product->A->symmetric == PETSC_BOOL3_TRUE && ptype == MATPRODUCT_AtB) {
7192: ptype = MATPRODUCT_AB;
7193: product->symbolic_used_the_fact_A_is_symmetric = PETSC_TRUE;
7194: }
7195: switch (ptype) {
7196: case MATPRODUCT_AB:
7197: A = product->A;
7198: P = product->B;
7199: m = A->rmap->n;
7200: n = P->cmap->n;
7201: M = A->rmap->N;
7202: N = P->cmap->N;
7203: hasoffproc = PETSC_FALSE; /* will not scatter mat product values to other processes */
7204: break;
7205: case MATPRODUCT_AtB:
7206: P = product->A;
7207: A = product->B;
7208: m = P->cmap->n;
7209: n = A->cmap->n;
7210: M = P->cmap->N;
7211: N = A->cmap->N;
7212: hasoffproc = PETSC_TRUE;
7213: break;
7214: case MATPRODUCT_PtAP:
7215: A = product->A;
7216: P = product->B;
7217: m = P->cmap->n;
7218: n = P->cmap->n;
7219: M = P->cmap->N;
7220: N = P->cmap->N;
7221: hasoffproc = PETSC_TRUE;
7222: break;
7223: default:
7224: SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Not for product type %s", MatProductTypes[ptype]);
7225: }
7226: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)C), &size));
7227: if (size == 1) hasoffproc = PETSC_FALSE;
7229: /* defaults */
7230: for (i = 0; i < MAX_NUMBER_INTERMEDIATE; i++) {
7231: mp[i] = NULL;
7232: mptmp[i] = PETSC_FALSE;
7233: rmapt[i] = -1;
7234: cmapt[i] = -1;
7235: rmapa[i] = NULL;
7236: cmapa[i] = NULL;
7237: }
7239: /* customization */
7240: PetscCall(PetscNew(&mmdata));
7241: mmdata->reusesym = product->api_user;
7242: if (ptype == MATPRODUCT_AB) {
7243: if (product->api_user) {
7244: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatMatMult", "Mat");
7245: PetscCall(PetscOptionsBool("-matmatmult_backend_mergeB", "Merge product->B local matrices", "MatMatMult", mmdata->abmerge, &mmdata->abmerge, NULL));
7246: PetscCall(PetscOptionsBool("-matmatmult_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7247: PetscOptionsEnd();
7248: } else {
7249: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_AB", "Mat");
7250: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_mergeB", "Merge product->B local matrices", "MatMatMult", mmdata->abmerge, &mmdata->abmerge, NULL));
7251: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7252: PetscOptionsEnd();
7253: }
7254: } else if (ptype == MATPRODUCT_PtAP) {
7255: if (product->api_user) {
7256: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatPtAP", "Mat");
7257: PetscCall(PetscOptionsBool("-matptap_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7258: PetscOptionsEnd();
7259: } else {
7260: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_PtAP", "Mat");
7261: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7262: PetscOptionsEnd();
7263: }
7264: }
7265: a = (Mat_MPIAIJ *)A->data;
7266: p = (Mat_MPIAIJ *)P->data;
7267: PetscCall(MatSetSizes(C, m, n, M, N));
7268: PetscCall(PetscLayoutSetUp(C->rmap));
7269: PetscCall(PetscLayoutSetUp(C->cmap));
7270: PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
7271: PetscCall(MatGetOptionsPrefix(C, &prefix));
7273: cp = 0;
7274: switch (ptype) {
7275: case MATPRODUCT_AB: /* A * P */
7276: PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &mmdata->startsj_s, &mmdata->startsj_r, &mmdata->bufa, &mmdata->P_oth));
7278: /* A_diag * P_local (merged or not) */
7279: if (mmdata->abmerge) { /* P's diagonal and off-diag blocks are merged to one matrix, then multiplied by A_diag */
7280: /* P is product->B */
7281: PetscCall(MatMPIAIJGetLocalMatMerge(P, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7282: PetscCall(MatProductCreate(a->A, mmdata->Bloc, NULL, &mp[cp]));
7283: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7284: PetscCall(MatProductSetFill(mp[cp], product->fill));
7285: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7286: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7287: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7288: mp[cp]->product->api_user = product->api_user;
7289: PetscCall(MatProductSetFromOptions(mp[cp]));
7290: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7291: PetscCall(ISGetIndices(glob, &globidx));
7292: rmapt[cp] = 1;
7293: cmapt[cp] = 2;
7294: cmapa[cp] = globidx;
7295: mptmp[cp] = PETSC_FALSE;
7296: cp++;
7297: } else { /* A_diag * P_diag and A_diag * P_off */
7298: PetscCall(MatProductCreate(a->A, p->A, NULL, &mp[cp]));
7299: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7300: PetscCall(MatProductSetFill(mp[cp], product->fill));
7301: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7302: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7303: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7304: mp[cp]->product->api_user = product->api_user;
7305: PetscCall(MatProductSetFromOptions(mp[cp]));
7306: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7307: rmapt[cp] = 1;
7308: cmapt[cp] = 1;
7309: mptmp[cp] = PETSC_FALSE;
7310: cp++;
7311: PetscCall(MatProductCreate(a->A, p->B, NULL, &mp[cp]));
7312: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7313: PetscCall(MatProductSetFill(mp[cp], product->fill));
7314: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7315: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7316: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7317: mp[cp]->product->api_user = product->api_user;
7318: PetscCall(MatProductSetFromOptions(mp[cp]));
7319: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7320: rmapt[cp] = 1;
7321: cmapt[cp] = 2;
7322: cmapa[cp] = p->garray;
7323: mptmp[cp] = PETSC_FALSE;
7324: cp++;
7325: }
7327: /* A_off * P_other */
7328: if (mmdata->P_oth) {
7329: PetscCall(MatSeqAIJCompactOutExtraColumns_SeqAIJ(mmdata->P_oth, &P_oth_l2g)); /* make P_oth use local col ids */
7330: PetscCall(ISLocalToGlobalMappingGetIndices(P_oth_l2g, &P_oth_idx));
7331: PetscCall(MatSetType(mmdata->P_oth, ((PetscObject)a->B)->type_name));
7332: PetscCall(MatBindToCPU(mmdata->P_oth, mmdata->P_oth_bind));
7333: PetscCall(MatProductCreate(a->B, mmdata->P_oth, NULL, &mp[cp]));
7334: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7335: PetscCall(MatProductSetFill(mp[cp], product->fill));
7336: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7337: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7338: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7339: mp[cp]->product->api_user = product->api_user;
7340: PetscCall(MatProductSetFromOptions(mp[cp]));
7341: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7342: rmapt[cp] = 1;
7343: cmapt[cp] = 2;
7344: cmapa[cp] = P_oth_idx;
7345: mptmp[cp] = PETSC_FALSE;
7346: cp++;
7347: }
7348: break;
7350: case MATPRODUCT_AtB: /* (P^t * A): P_diag * A_loc + P_off * A_loc */
7351: /* A is product->B */
7352: PetscCall(MatMPIAIJGetLocalMatMerge(A, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7353: if (A == P) { /* when A==P, we can take advantage of the already merged mmdata->Bloc */
7354: PetscCall(MatProductCreate(mmdata->Bloc, mmdata->Bloc, NULL, &mp[cp]));
7355: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7356: PetscCall(MatProductSetFill(mp[cp], product->fill));
7357: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7358: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7359: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7360: mp[cp]->product->api_user = product->api_user;
7361: PetscCall(MatProductSetFromOptions(mp[cp]));
7362: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7363: PetscCall(ISGetIndices(glob, &globidx));
7364: rmapt[cp] = 2;
7365: rmapa[cp] = globidx;
7366: cmapt[cp] = 2;
7367: cmapa[cp] = globidx;
7368: mptmp[cp] = PETSC_FALSE;
7369: cp++;
7370: } else {
7371: PetscCall(MatProductCreate(p->A, mmdata->Bloc, NULL, &mp[cp]));
7372: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7373: PetscCall(MatProductSetFill(mp[cp], product->fill));
7374: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7375: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7376: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7377: mp[cp]->product->api_user = product->api_user;
7378: PetscCall(MatProductSetFromOptions(mp[cp]));
7379: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7380: PetscCall(ISGetIndices(glob, &globidx));
7381: rmapt[cp] = 1;
7382: cmapt[cp] = 2;
7383: cmapa[cp] = globidx;
7384: mptmp[cp] = PETSC_FALSE;
7385: cp++;
7386: PetscCall(MatProductCreate(p->B, mmdata->Bloc, NULL, &mp[cp]));
7387: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7388: PetscCall(MatProductSetFill(mp[cp], product->fill));
7389: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7390: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7391: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7392: mp[cp]->product->api_user = product->api_user;
7393: PetscCall(MatProductSetFromOptions(mp[cp]));
7394: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7395: rmapt[cp] = 2;
7396: rmapa[cp] = p->garray;
7397: cmapt[cp] = 2;
7398: cmapa[cp] = globidx;
7399: mptmp[cp] = PETSC_FALSE;
7400: cp++;
7401: }
7402: break;
7403: case MATPRODUCT_PtAP:
7404: PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &mmdata->startsj_s, &mmdata->startsj_r, &mmdata->bufa, &mmdata->P_oth));
7405: /* P is product->B */
7406: PetscCall(MatMPIAIJGetLocalMatMerge(P, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7407: PetscCall(MatProductCreate(a->A, mmdata->Bloc, NULL, &mp[cp]));
7408: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_PtAP));
7409: PetscCall(MatProductSetFill(mp[cp], product->fill));
7410: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7411: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7412: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7413: mp[cp]->product->api_user = product->api_user;
7414: PetscCall(MatProductSetFromOptions(mp[cp]));
7415: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7416: PetscCall(ISGetIndices(glob, &globidx));
7417: rmapt[cp] = 2;
7418: rmapa[cp] = globidx;
7419: cmapt[cp] = 2;
7420: cmapa[cp] = globidx;
7421: mptmp[cp] = PETSC_FALSE;
7422: cp++;
7423: if (mmdata->P_oth) {
7424: PetscCall(MatSeqAIJCompactOutExtraColumns_SeqAIJ(mmdata->P_oth, &P_oth_l2g));
7425: PetscCall(ISLocalToGlobalMappingGetIndices(P_oth_l2g, &P_oth_idx));
7426: PetscCall(MatSetType(mmdata->P_oth, ((PetscObject)a->B)->type_name));
7427: PetscCall(MatBindToCPU(mmdata->P_oth, mmdata->P_oth_bind));
7428: PetscCall(MatProductCreate(a->B, mmdata->P_oth, NULL, &mp[cp]));
7429: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7430: PetscCall(MatProductSetFill(mp[cp], product->fill));
7431: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7432: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7433: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7434: mp[cp]->product->api_user = product->api_user;
7435: PetscCall(MatProductSetFromOptions(mp[cp]));
7436: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7437: mptmp[cp] = PETSC_TRUE;
7438: cp++;
7439: PetscCall(MatProductCreate(mmdata->Bloc, mp[1], NULL, &mp[cp]));
7440: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7441: PetscCall(MatProductSetFill(mp[cp], product->fill));
7442: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7443: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7444: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7445: mp[cp]->product->api_user = product->api_user;
7446: PetscCall(MatProductSetFromOptions(mp[cp]));
7447: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7448: rmapt[cp] = 2;
7449: rmapa[cp] = globidx;
7450: cmapt[cp] = 2;
7451: cmapa[cp] = P_oth_idx;
7452: mptmp[cp] = PETSC_FALSE;
7453: cp++;
7454: }
7455: break;
7456: default:
7457: SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Not for product type %s", MatProductTypes[ptype]);
7458: }
7459: /* sanity check */
7460: if (size > 1)
7461: 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);
7463: PetscCall(PetscMalloc2(cp, &mmdata->mp, cp, &mmdata->mptmp));
7464: for (i = 0; i < cp; i++) {
7465: mmdata->mp[i] = mp[i];
7466: mmdata->mptmp[i] = mptmp[i];
7467: }
7468: mmdata->cp = cp;
7469: C->product->data = mmdata;
7470: C->product->destroy = MatDestroy_MatMatMPIAIJBACKEND;
7471: C->ops->productnumeric = MatProductNumeric_MPIAIJBACKEND;
7473: /* memory type */
7474: mmdata->mtype = PETSC_MEMTYPE_HOST;
7475: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iscuda, MATSEQAIJCUSPARSE, MATMPIAIJCUSPARSE, ""));
7476: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iship, MATSEQAIJHIPSPARSE, MATMPIAIJHIPSPARSE, ""));
7477: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iskokk, MATSEQAIJKOKKOS, MATMPIAIJKOKKOS, ""));
7478: if (iscuda) mmdata->mtype = PETSC_MEMTYPE_CUDA;
7479: else if (iship) mmdata->mtype = PETSC_MEMTYPE_HIP;
7480: else if (iskokk) mmdata->mtype = PETSC_MEMTYPE_KOKKOS;
7482: /* prepare coo coordinates for values insertion */
7484: /* count total nonzeros of those intermediate seqaij Mats
7485: ncoo_d: # of nonzeros of matrices that do not have offproc entries
7486: ncoo_o: # of nonzeros (of matrices that might have offproc entries) that will be inserted to remote procs
7487: ncoo_oown: # of nonzeros (of matrices that might have offproc entries) that will be inserted locally
7488: */
7489: for (cp = 0, ncoo_d = 0, ncoo_o = 0, ncoo_oown = 0; cp < mmdata->cp; cp++) {
7490: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7491: if (mptmp[cp]) continue;
7492: if (rmapt[cp] == 2 && hasoffproc) { /* the rows need to be scatter to all processes (might include self) */
7493: const PetscInt *rmap = rmapa[cp];
7494: const PetscInt mr = mp[cp]->rmap->n;
7495: const PetscInt rs = C->rmap->rstart;
7496: const PetscInt re = C->rmap->rend;
7497: const PetscInt *ii = mm->i;
7498: for (i = 0; i < mr; i++) {
7499: const PetscInt gr = rmap[i];
7500: const PetscInt nz = ii[i + 1] - ii[i];
7501: if (gr < rs || gr >= re) ncoo_o += nz; /* this row is offproc */
7502: else ncoo_oown += nz; /* this row is local */
7503: }
7504: } else ncoo_d += mm->nz;
7505: }
7507: /*
7508: ncoo: total number of nonzeros (including those inserted by remote procs) belonging to this proc
7510: ncoo = ncoo_d + ncoo_oown + ncoo2, which ncoo2 is number of nonzeros inserted to me by other procs.
7512: off[0] points to a big index array, which is shared by off[1,2,...]. Similarly, for own[0].
7514: off[p]: points to the segment for matrix mp[p], storing location of nonzeros that mp[p] will insert to others
7515: own[p]: points to the segment for matrix mp[p], storing location of nonzeros that mp[p] will insert locally
7516: so, off[p+1]-off[p] is the number of nonzeros that mp[p] will send to others.
7518: coo_i/j/v[]: [ncoo] row/col/val of nonzeros belonging to this proc.
7519: 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.
7520: */
7521: PetscCall(PetscCalloc1(mmdata->cp + 1, &mmdata->off)); /* +1 to make a csr-like data structure */
7522: PetscCall(PetscCalloc1(mmdata->cp + 1, &mmdata->own));
7524: /* gather (i,j) of nonzeros inserted by remote procs */
7525: if (hasoffproc) {
7526: PetscSF msf;
7527: PetscInt ncoo2, *coo_i2, *coo_j2;
7529: PetscCall(PetscMalloc1(ncoo_o, &mmdata->off[0]));
7530: PetscCall(PetscMalloc1(ncoo_oown, &mmdata->own[0]));
7531: PetscCall(PetscMalloc2(ncoo_o, &coo_i, ncoo_o, &coo_j)); /* to collect (i,j) of entries to be sent to others */
7533: for (cp = 0, ncoo_o = 0; cp < mmdata->cp; cp++) {
7534: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7535: PetscInt *idxoff = mmdata->off[cp];
7536: PetscInt *idxown = mmdata->own[cp];
7537: if (!mptmp[cp] && rmapt[cp] == 2) { /* row map is sparse */
7538: const PetscInt *rmap = rmapa[cp];
7539: const PetscInt *cmap = cmapa[cp];
7540: const PetscInt *ii = mm->i;
7541: PetscInt *coi = coo_i + ncoo_o;
7542: PetscInt *coj = coo_j + ncoo_o;
7543: const PetscInt mr = mp[cp]->rmap->n;
7544: const PetscInt rs = C->rmap->rstart;
7545: const PetscInt re = C->rmap->rend;
7546: const PetscInt cs = C->cmap->rstart;
7547: for (i = 0; i < mr; i++) {
7548: const PetscInt *jj = mm->j + ii[i];
7549: const PetscInt gr = rmap[i];
7550: const PetscInt nz = ii[i + 1] - ii[i];
7551: if (gr < rs || gr >= re) { /* this is an offproc row */
7552: for (j = ii[i]; j < ii[i + 1]; j++) {
7553: *coi++ = gr;
7554: *idxoff++ = j;
7555: }
7556: if (!cmapt[cp]) { /* already global */
7557: for (j = 0; j < nz; j++) *coj++ = jj[j];
7558: } else if (cmapt[cp] == 1) { /* local to global for owned columns of C */
7559: for (j = 0; j < nz; j++) *coj++ = jj[j] + cs;
7560: } else { /* offdiag */
7561: for (j = 0; j < nz; j++) *coj++ = cmap[jj[j]];
7562: }
7563: ncoo_o += nz;
7564: } else { /* this is a local row */
7565: for (j = ii[i]; j < ii[i + 1]; j++) *idxown++ = j;
7566: }
7567: }
7568: }
7569: mmdata->off[cp + 1] = idxoff;
7570: mmdata->own[cp + 1] = idxown;
7571: }
7573: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)C), &mmdata->sf));
7574: PetscInt incoo_o;
7575: PetscCall(PetscIntCast(ncoo_o, &incoo_o));
7576: PetscCall(PetscSFSetGraphLayout(mmdata->sf, C->rmap, incoo_o /*nleaves*/, NULL /*ilocal*/, PETSC_OWN_POINTER, coo_i));
7577: PetscCall(PetscSFGetMultiSF(mmdata->sf, &msf));
7578: PetscCall(PetscSFGetGraph(msf, &ncoo2 /*nroots*/, NULL, NULL, NULL));
7579: ncoo = ncoo_d + ncoo_oown + ncoo2;
7580: PetscCall(PetscMalloc2(ncoo, &coo_i2, ncoo, &coo_j2));
7581: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_INT, coo_i, coo_i2 + ncoo_d + ncoo_oown)); /* put (i,j) of remote nonzeros at back */
7582: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_INT, coo_i, coo_i2 + ncoo_d + ncoo_oown));
7583: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_INT, coo_j, coo_j2 + ncoo_d + ncoo_oown));
7584: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_INT, coo_j, coo_j2 + ncoo_d + ncoo_oown));
7585: PetscCall(PetscFree2(coo_i, coo_j));
7586: /* allocate MPI send buffer to collect nonzero values to be sent to remote procs */
7587: PetscCall(PetscSFMalloc(mmdata->sf, mmdata->mtype, ncoo_o * sizeof(PetscScalar), (void **)&mmdata->coo_w));
7588: coo_i = coo_i2;
7589: coo_j = coo_j2;
7590: } else { /* no offproc values insertion */
7591: ncoo = ncoo_d;
7592: PetscCall(PetscMalloc2(ncoo, &coo_i, ncoo, &coo_j));
7594: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)C), &mmdata->sf));
7595: PetscCall(PetscSFSetGraph(mmdata->sf, 0, 0, NULL, PETSC_OWN_POINTER, NULL, PETSC_OWN_POINTER));
7596: PetscCall(PetscSFSetUp(mmdata->sf));
7597: }
7598: mmdata->hasoffproc = hasoffproc;
7600: /* gather (i,j) of nonzeros inserted locally */
7601: for (cp = 0, ncoo_d = 0; cp < mmdata->cp; cp++) {
7602: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7603: PetscInt *coi = coo_i + ncoo_d;
7604: PetscInt *coj = coo_j + ncoo_d;
7605: const PetscInt *jj = mm->j;
7606: const PetscInt *ii = mm->i;
7607: const PetscInt *cmap = cmapa[cp];
7608: const PetscInt *rmap = rmapa[cp];
7609: const PetscInt mr = mp[cp]->rmap->n;
7610: const PetscInt rs = C->rmap->rstart;
7611: const PetscInt re = C->rmap->rend;
7612: const PetscInt cs = C->cmap->rstart;
7614: if (mptmp[cp]) continue;
7615: if (rmapt[cp] == 1) { /* consecutive rows */
7616: /* fill coo_i */
7617: for (i = 0; i < mr; i++) {
7618: const PetscInt gr = i + rs;
7619: for (j = ii[i]; j < ii[i + 1]; j++) coi[j] = gr;
7620: }
7621: /* fill coo_j */
7622: if (!cmapt[cp]) { /* type-0, already global */
7623: PetscCall(PetscArraycpy(coj, jj, mm->nz));
7624: } else if (cmapt[cp] == 1) { /* type-1, local to global for consecutive columns of C */
7625: for (j = 0; j < mm->nz; j++) coj[j] = jj[j] + cs; /* lid + col start */
7626: } else { /* type-2, local to global for sparse columns */
7627: for (j = 0; j < mm->nz; j++) coj[j] = cmap[jj[j]];
7628: }
7629: ncoo_d += mm->nz;
7630: } else if (rmapt[cp] == 2) { /* sparse rows */
7631: for (i = 0; i < mr; i++) {
7632: const PetscInt *jj = mm->j + ii[i];
7633: const PetscInt gr = rmap[i];
7634: const PetscInt nz = ii[i + 1] - ii[i];
7635: if (gr >= rs && gr < re) { /* local rows */
7636: for (j = ii[i]; j < ii[i + 1]; j++) *coi++ = gr;
7637: if (!cmapt[cp]) { /* type-0, already global */
7638: for (j = 0; j < nz; j++) *coj++ = jj[j];
7639: } else if (cmapt[cp] == 1) { /* local to global for owned columns of C */
7640: for (j = 0; j < nz; j++) *coj++ = jj[j] + cs;
7641: } else { /* type-2, local to global for sparse columns */
7642: for (j = 0; j < nz; j++) *coj++ = cmap[jj[j]];
7643: }
7644: ncoo_d += nz;
7645: }
7646: }
7647: }
7648: }
7649: if (glob) PetscCall(ISRestoreIndices(glob, &globidx));
7650: PetscCall(ISDestroy(&glob));
7651: if (P_oth_l2g) PetscCall(ISLocalToGlobalMappingRestoreIndices(P_oth_l2g, &P_oth_idx));
7652: PetscCall(ISLocalToGlobalMappingDestroy(&P_oth_l2g));
7653: /* allocate an array to store all nonzeros (inserted locally or remotely) belonging to this proc */
7654: PetscCall(PetscSFMalloc(mmdata->sf, mmdata->mtype, ncoo * sizeof(PetscScalar), (void **)&mmdata->coo_v));
7656: /* preallocate with COO data */
7657: PetscCall(MatSetPreallocationCOO(C, ncoo, coo_i, coo_j));
7658: PetscCall(PetscFree2(coo_i, coo_j));
7659: PetscFunctionReturn(PETSC_SUCCESS);
7660: }
7662: PetscErrorCode MatProductSetFromOptions_MPIAIJBACKEND(Mat mat)
7663: {
7664: Mat_Product *product = mat->product;
7665: #if defined(PETSC_HAVE_DEVICE)
7666: PetscBool match = PETSC_FALSE;
7667: PetscBool usecpu = PETSC_FALSE;
7668: #else
7669: PetscBool match = PETSC_TRUE;
7670: #endif
7672: PetscFunctionBegin;
7673: MatCheckProduct(mat, 1);
7674: #if defined(PETSC_HAVE_DEVICE)
7675: if (!product->A->boundtocpu && !product->B->boundtocpu) PetscCall(PetscObjectTypeCompare((PetscObject)product->B, ((PetscObject)product->A)->type_name, &match));
7676: if (match) { /* we can always fallback to the CPU if requested */
7677: switch (product->type) {
7678: case MATPRODUCT_AB:
7679: if (product->api_user) {
7680: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatMatMult", "Mat");
7681: PetscCall(PetscOptionsBool("-matmatmult_backend_cpu", "Use CPU code", "MatMatMult", usecpu, &usecpu, NULL));
7682: PetscOptionsEnd();
7683: } else {
7684: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_AB", "Mat");
7685: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatMatMult", usecpu, &usecpu, NULL));
7686: PetscOptionsEnd();
7687: }
7688: break;
7689: case MATPRODUCT_AtB:
7690: if (product->api_user) {
7691: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatTransposeMatMult", "Mat");
7692: PetscCall(PetscOptionsBool("-mattransposematmult_backend_cpu", "Use CPU code", "MatTransposeMatMult", usecpu, &usecpu, NULL));
7693: PetscOptionsEnd();
7694: } else {
7695: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_AtB", "Mat");
7696: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatTransposeMatMult", usecpu, &usecpu, NULL));
7697: PetscOptionsEnd();
7698: }
7699: break;
7700: case MATPRODUCT_PtAP:
7701: if (product->api_user) {
7702: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatPtAP", "Mat");
7703: PetscCall(PetscOptionsBool("-matptap_backend_cpu", "Use CPU code", "MatPtAP", usecpu, &usecpu, NULL));
7704: PetscOptionsEnd();
7705: } else {
7706: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_PtAP", "Mat");
7707: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatPtAP", usecpu, &usecpu, NULL));
7708: PetscOptionsEnd();
7709: }
7710: break;
7711: default:
7712: break;
7713: }
7714: match = (PetscBool)!usecpu;
7715: }
7716: #endif
7717: if (match) {
7718: switch (product->type) {
7719: case MATPRODUCT_AB:
7720: case MATPRODUCT_AtB:
7721: case MATPRODUCT_PtAP:
7722: mat->ops->productsymbolic = MatProductSymbolic_MPIAIJBACKEND;
7723: break;
7724: default:
7725: break;
7726: }
7727: }
7728: /* fallback to MPIAIJ ops */
7729: if (!mat->ops->productsymbolic) PetscCall(MatProductSetFromOptions_MPIAIJ(mat));
7730: PetscFunctionReturn(PETSC_SUCCESS);
7731: }
7733: /*
7734: Produces a set of block column indices of the matrix row, one for each block represented in the original row
7736: n - the number of block indices in cc[]
7737: cc - the block indices (must be large enough to contain the indices)
7738: */
7739: static inline PetscErrorCode MatCollapseRow(Mat Amat, PetscInt row, PetscInt bs, PetscInt *n, PetscInt *cc)
7740: {
7741: PetscInt cnt = -1, nidx, j;
7742: const PetscInt *idx;
7744: PetscFunctionBegin;
7745: PetscCall(MatGetRow(Amat, row, &nidx, &idx, NULL));
7746: if (nidx) {
7747: cnt = 0;
7748: cc[cnt] = idx[0] / bs;
7749: for (j = 1; j < nidx; j++) {
7750: if (cc[cnt] < idx[j] / bs) cc[++cnt] = idx[j] / bs;
7751: }
7752: }
7753: PetscCall(MatRestoreRow(Amat, row, &nidx, &idx, NULL));
7754: *n = cnt + 1;
7755: PetscFunctionReturn(PETSC_SUCCESS);
7756: }
7758: /*
7759: Produces a set of block column indices of the matrix block row, one for each block represented in the original set of rows
7761: ncollapsed - the number of block indices
7762: collapsed - the block indices (must be large enough to contain the indices)
7763: */
7764: static inline PetscErrorCode MatCollapseRows(Mat Amat, PetscInt start, PetscInt bs, PetscInt *w0, PetscInt *w1, PetscInt *w2, PetscInt *ncollapsed, PetscInt **collapsed)
7765: {
7766: PetscInt i, nprev, *cprev = w0, ncur = 0, *ccur = w1, *merged = w2, *cprevtmp;
7768: PetscFunctionBegin;
7769: PetscCall(MatCollapseRow(Amat, start, bs, &nprev, cprev));
7770: for (i = start + 1; i < start + bs; i++) {
7771: PetscCall(MatCollapseRow(Amat, i, bs, &ncur, ccur));
7772: PetscCall(PetscMergeIntArray(nprev, cprev, ncur, ccur, &nprev, &merged));
7773: cprevtmp = cprev;
7774: cprev = merged;
7775: merged = cprevtmp;
7776: }
7777: *ncollapsed = nprev;
7778: if (collapsed) *collapsed = cprev;
7779: PetscFunctionReturn(PETSC_SUCCESS);
7780: }
7782: /*
7783: MatCreateGraph_Simple_AIJ - create simple scalar matrix (graph) from potentially blocked matrix
7785: Input Parameter:
7786: . Amat - matrix
7787: - symmetrize - make the result symmetric
7788: + scale - scale with diagonal
7790: Output Parameter:
7791: . a_Gmat - output scalar graph >= 0
7793: */
7794: PETSC_INTERN PetscErrorCode MatCreateGraph_Simple_AIJ(Mat Amat, PetscBool symmetrize, PetscBool scale, PetscReal filter, PetscInt index_size, PetscInt index[], Mat *a_Gmat)
7795: {
7796: PetscInt Istart, Iend, Ii, jj, kk, ncols, nloc, NN, MM, bs;
7797: MPI_Comm comm;
7798: Mat Gmat;
7799: PetscBool ismpiaij, isseqaij;
7800: Mat a, b, c;
7801: MatType jtype;
7803: PetscFunctionBegin;
7804: PetscCall(PetscObjectGetComm((PetscObject)Amat, &comm));
7805: PetscCall(MatGetOwnershipRange(Amat, &Istart, &Iend));
7806: PetscCall(MatGetSize(Amat, &MM, &NN));
7807: PetscCall(MatGetBlockSize(Amat, &bs));
7808: nloc = (Iend - Istart) / bs;
7810: PetscCall(PetscObjectBaseTypeCompare((PetscObject)Amat, MATSEQAIJ, &isseqaij));
7811: PetscCall(PetscObjectBaseTypeCompare((PetscObject)Amat, MATMPIAIJ, &ismpiaij));
7812: PetscCheck(isseqaij || ismpiaij, comm, PETSC_ERR_USER, "Require (MPI)AIJ matrix type");
7814: /* TODO GPU: these calls are potentially expensive if matrices are large and we want to use the GPU */
7815: /* A solution consists in providing a new API, MatAIJGetCollapsedAIJ, and each class can provide a fast
7816: implementation */
7817: if (bs > 1) {
7818: PetscCall(MatGetType(Amat, &jtype));
7819: PetscCall(MatCreate(comm, &Gmat));
7820: PetscCall(MatSetType(Gmat, jtype));
7821: PetscCall(MatSetSizes(Gmat, nloc, nloc, PETSC_DETERMINE, PETSC_DETERMINE));
7822: PetscCall(MatSetBlockSizes(Gmat, 1, 1));
7823: if (isseqaij || ((Mat_MPIAIJ *)Amat->data)->garray) {
7824: PetscInt *d_nnz, *o_nnz;
7825: MatScalar *aa, val, *AA;
7826: PetscInt *aj, *ai, *AJ, nc, nmax = 0;
7828: if (isseqaij) {
7829: a = Amat;
7830: b = NULL;
7831: } else {
7832: Mat_MPIAIJ *d = (Mat_MPIAIJ *)Amat->data;
7833: a = d->A;
7834: b = d->B;
7835: }
7836: PetscCall(PetscInfo(Amat, "New bs>1 Graph. nloc=%" PetscInt_FMT "\n", nloc));
7837: PetscCall(PetscMalloc2(nloc, &d_nnz, (isseqaij ? 0 : nloc), &o_nnz));
7838: for (c = a, kk = 0; c && kk < 2; c = b, kk++) {
7839: PetscInt *nnz = (c == a) ? d_nnz : o_nnz;
7840: const PetscInt *cols1, *cols2;
7842: for (PetscInt brow = 0, nc1, nc2, ok = 1; brow < nloc * bs; brow += bs) { // block rows
7843: PetscCall(MatGetRow(c, brow, &nc2, &cols2, NULL));
7844: nnz[brow / bs] = nc2 / bs;
7845: if (nc2 % bs) ok = 0;
7846: if (nnz[brow / bs] > nmax) nmax = nnz[brow / bs];
7847: for (PetscInt ii = 1; ii < bs; ii++) { // check for non-dense blocks
7848: PetscCall(MatGetRow(c, brow + ii, &nc1, &cols1, NULL));
7849: if (nc1 != nc2) ok = 0;
7850: else {
7851: for (PetscInt jj = 0; jj < nc1 && ok == 1; jj++) {
7852: if (cols1[jj] != cols2[jj]) ok = 0;
7853: if (cols1[jj] % bs != jj % bs) ok = 0;
7854: }
7855: }
7856: PetscCall(MatRestoreRow(c, brow + ii, &nc1, &cols1, NULL));
7857: }
7858: PetscCall(MatRestoreRow(c, brow, &nc2, &cols2, NULL));
7859: if (!ok) {
7860: PetscCall(PetscFree2(d_nnz, o_nnz));
7861: PetscCall(PetscInfo(Amat, "Found sparse blocks - revert to slow method\n"));
7862: goto old_bs;
7863: }
7864: }
7865: }
7866: PetscCall(MatSeqAIJSetPreallocation(Gmat, 0, d_nnz));
7867: PetscCall(MatMPIAIJSetPreallocation(Gmat, 0, d_nnz, 0, o_nnz));
7868: PetscCall(PetscFree2(d_nnz, o_nnz));
7869: PetscCall(PetscMalloc2(nmax, &AA, nmax, &AJ));
7870: // diag
7871: for (PetscInt brow = 0, n, grow; brow < nloc * bs; brow += bs) { // block rows
7872: Mat_SeqAIJ *aseq = (Mat_SeqAIJ *)a->data;
7874: ai = aseq->i;
7875: n = ai[brow + 1] - ai[brow];
7876: aj = aseq->j + ai[brow];
7877: for (PetscInt k = 0; k < n; k += bs) { // block columns
7878: AJ[k / bs] = aj[k] / bs + Istart / bs; // diag starts at (Istart,Istart)
7879: val = 0;
7880: if (index_size == 0) {
7881: for (PetscInt ii = 0; ii < bs; ii++) { // rows in block
7882: aa = aseq->a + ai[brow + ii] + k;
7883: for (PetscInt jj = 0; jj < bs; jj++) { // columns in block
7884: val += PetscAbs(PetscRealPart(aa[jj])); // a sort of norm
7885: }
7886: }
7887: } else { // use (index,index) value if provided
7888: for (PetscInt iii = 0; iii < index_size; iii++) { // rows in block
7889: PetscInt ii = index[iii];
7890: aa = aseq->a + ai[brow + ii] + k;
7891: for (PetscInt jjj = 0; jjj < index_size; jjj++) { // columns in block
7892: PetscInt jj = index[jjj];
7893: val += PetscAbs(PetscRealPart(aa[jj]));
7894: }
7895: }
7896: }
7897: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs (%" PetscInt_FMT ") >= nmax (%" PetscInt_FMT ")", k / bs, nmax);
7898: AA[k / bs] = val;
7899: }
7900: grow = Istart / bs + brow / bs;
7901: PetscCall(MatSetValues(Gmat, 1, &grow, n / bs, AJ, AA, ADD_VALUES));
7902: }
7903: // off-diag
7904: if (ismpiaij) {
7905: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)Amat->data;
7906: const PetscScalar *vals;
7907: const PetscInt *cols, *garray = aij->garray;
7909: PetscCheck(garray, PETSC_COMM_SELF, PETSC_ERR_USER, "No garray ?");
7910: for (PetscInt brow = 0, grow; brow < nloc * bs; brow += bs) { // block rows
7911: PetscCall(MatGetRow(b, brow, &ncols, &cols, NULL));
7912: for (PetscInt k = 0, cidx = 0; k < ncols; k += bs, cidx++) {
7913: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs >= nmax");
7914: AA[k / bs] = 0;
7915: AJ[cidx] = garray[cols[k]] / bs;
7916: }
7917: nc = ncols / bs;
7918: PetscCall(MatRestoreRow(b, brow, &ncols, &cols, NULL));
7919: if (index_size == 0) {
7920: for (PetscInt ii = 0; ii < bs; ii++) { // rows in block
7921: PetscCall(MatGetRow(b, brow + ii, &ncols, &cols, &vals));
7922: for (PetscInt k = 0; k < ncols; k += bs) {
7923: for (PetscInt jj = 0; jj < bs; jj++) { // cols in block
7924: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs (%" PetscInt_FMT ") >= nmax (%" PetscInt_FMT ")", k / bs, nmax);
7925: AA[k / bs] += PetscAbs(PetscRealPart(vals[k + jj]));
7926: }
7927: }
7928: PetscCall(MatRestoreRow(b, brow + ii, &ncols, &cols, &vals));
7929: }
7930: } else { // use (index,index) value if provided
7931: for (PetscInt iii = 0; iii < index_size; iii++) { // rows in block
7932: PetscInt ii = index[iii];
7933: PetscCall(MatGetRow(b, brow + ii, &ncols, &cols, &vals));
7934: for (PetscInt k = 0; k < ncols; k += bs) {
7935: for (PetscInt jjj = 0; jjj < index_size; jjj++) { // cols in block
7936: PetscInt jj = index[jjj];
7937: AA[k / bs] += PetscAbs(PetscRealPart(vals[k + jj]));
7938: }
7939: }
7940: PetscCall(MatRestoreRow(b, brow + ii, &ncols, &cols, &vals));
7941: }
7942: }
7943: grow = Istart / bs + brow / bs;
7944: PetscCall(MatSetValues(Gmat, 1, &grow, nc, AJ, AA, ADD_VALUES));
7945: }
7946: }
7947: PetscCall(MatAssemblyBegin(Gmat, MAT_FINAL_ASSEMBLY));
7948: PetscCall(MatAssemblyEnd(Gmat, MAT_FINAL_ASSEMBLY));
7949: PetscCall(PetscFree2(AA, AJ));
7950: } else {
7951: const PetscScalar *vals;
7952: const PetscInt *idx;
7953: PetscInt *d_nnz, *o_nnz, *w0, *w1, *w2;
7954: old_bs:
7955: /*
7956: Determine the preallocation needed for the scalar matrix derived from the vector matrix.
7957: */
7958: PetscCall(PetscInfo(Amat, "OLD bs>1 CreateGraph\n"));
7959: PetscCall(PetscMalloc2(nloc, &d_nnz, (isseqaij ? 0 : nloc), &o_nnz));
7960: if (isseqaij) {
7961: PetscInt max_d_nnz;
7963: /*
7964: Determine exact preallocation count for (sequential) scalar matrix
7965: */
7966: PetscCall(MatSeqAIJGetMaxRowNonzeros(Amat, &max_d_nnz));
7967: max_d_nnz = PetscMin(nloc, bs * max_d_nnz);
7968: PetscCall(PetscMalloc3(max_d_nnz, &w0, max_d_nnz, &w1, max_d_nnz, &w2));
7969: for (Ii = 0, jj = 0; Ii < Iend; Ii += bs, jj++) PetscCall(MatCollapseRows(Amat, Ii, bs, w0, w1, w2, &d_nnz[jj], NULL));
7970: PetscCall(PetscFree3(w0, w1, w2));
7971: } else if (ismpiaij) {
7972: Mat Daij, Oaij;
7973: const PetscInt *garray;
7974: PetscInt max_d_nnz;
7976: PetscCall(MatMPIAIJGetSeqAIJ(Amat, &Daij, &Oaij, &garray));
7977: /*
7978: Determine exact preallocation count for diagonal block portion of scalar matrix
7979: */
7980: PetscCall(MatSeqAIJGetMaxRowNonzeros(Daij, &max_d_nnz));
7981: max_d_nnz = PetscMin(nloc, bs * max_d_nnz);
7982: PetscCall(PetscMalloc3(max_d_nnz, &w0, max_d_nnz, &w1, max_d_nnz, &w2));
7983: for (Ii = 0, jj = 0; Ii < Iend - Istart; Ii += bs, jj++) PetscCall(MatCollapseRows(Daij, Ii, bs, w0, w1, w2, &d_nnz[jj], NULL));
7984: PetscCall(PetscFree3(w0, w1, w2));
7985: /*
7986: Over estimate (usually grossly over), preallocation count for off-diagonal portion of scalar matrix
7987: */
7988: for (Ii = 0, jj = 0; Ii < Iend - Istart; Ii += bs, jj++) {
7989: o_nnz[jj] = 0;
7990: for (kk = 0; kk < bs; kk++) { /* rows that get collapsed to a single row */
7991: PetscCall(MatGetRow(Oaij, Ii + kk, &ncols, NULL, NULL));
7992: o_nnz[jj] += ncols;
7993: PetscCall(MatRestoreRow(Oaij, Ii + kk, &ncols, NULL, NULL));
7994: }
7995: if (o_nnz[jj] > (NN / bs - nloc)) o_nnz[jj] = NN / bs - nloc;
7996: }
7997: } else SETERRQ(comm, PETSC_ERR_USER, "Require AIJ matrix type");
7998: /* get scalar copy (norms) of matrix */
7999: PetscCall(MatSeqAIJSetPreallocation(Gmat, 0, d_nnz));
8000: PetscCall(MatMPIAIJSetPreallocation(Gmat, 0, d_nnz, 0, o_nnz));
8001: PetscCall(PetscFree2(d_nnz, o_nnz));
8002: for (Ii = Istart; Ii < Iend; Ii++) {
8003: PetscInt dest_row = Ii / bs;
8005: PetscCall(MatGetRow(Amat, Ii, &ncols, &idx, &vals));
8006: for (jj = 0; jj < ncols; jj++) {
8007: PetscInt dest_col = idx[jj] / bs;
8008: PetscScalar sv = PetscAbs(PetscRealPart(vals[jj]));
8010: PetscCall(MatSetValues(Gmat, 1, &dest_row, 1, &dest_col, &sv, ADD_VALUES));
8011: }
8012: PetscCall(MatRestoreRow(Amat, Ii, &ncols, &idx, &vals));
8013: }
8014: PetscCall(MatAssemblyBegin(Gmat, MAT_FINAL_ASSEMBLY));
8015: PetscCall(MatAssemblyEnd(Gmat, MAT_FINAL_ASSEMBLY));
8016: }
8017: } else {
8018: if (symmetrize || filter >= 0 || scale) PetscCall(MatDuplicate(Amat, MAT_COPY_VALUES, &Gmat));
8019: else {
8020: Gmat = Amat;
8021: PetscCall(PetscObjectReference((PetscObject)Gmat));
8022: }
8023: if (isseqaij) {
8024: a = Gmat;
8025: b = NULL;
8026: } else {
8027: Mat_MPIAIJ *d = (Mat_MPIAIJ *)Gmat->data;
8028: a = d->A;
8029: b = d->B;
8030: }
8031: if (filter >= 0 || scale) {
8032: /* take absolute value of each entry */
8033: for (c = a, kk = 0; c && kk < 2; c = b, kk++) {
8034: MatInfo info;
8035: PetscScalar *avals;
8037: PetscCall(MatGetInfo(c, MAT_LOCAL, &info));
8038: PetscCall(MatSeqAIJGetArray(c, &avals));
8039: for (int jj = 0; jj < info.nz_used; jj++) avals[jj] = PetscAbsScalar(avals[jj]);
8040: PetscCall(MatSeqAIJRestoreArray(c, &avals));
8041: }
8042: }
8043: }
8044: if (symmetrize) {
8045: PetscBool isset, issym;
8047: PetscCall(MatIsSymmetricKnown(Amat, &isset, &issym));
8048: if (!isset || !issym) {
8049: Mat matTrans;
8051: PetscCall(MatTranspose(Gmat, MAT_INITIAL_MATRIX, &matTrans));
8052: PetscCall(MatAXPY(Gmat, 1.0, matTrans, Gmat->structurally_symmetric == PETSC_BOOL3_TRUE ? SAME_NONZERO_PATTERN : DIFFERENT_NONZERO_PATTERN));
8053: PetscCall(MatDestroy(&matTrans));
8054: }
8055: PetscCall(MatSetOption(Gmat, MAT_SYMMETRIC, PETSC_TRUE));
8056: } else if (Amat != Gmat) PetscCall(MatPropagateSymmetryOptions(Amat, Gmat));
8057: if (scale) {
8058: /* scale c for all diagonal values = 1 or -1 */
8059: Vec diag;
8061: PetscCall(MatCreateVecs(Gmat, &diag, NULL));
8062: PetscCall(MatGetDiagonal(Gmat, diag));
8063: PetscCall(VecReciprocal(diag));
8064: PetscCall(VecSqrtAbs(diag));
8065: PetscCall(MatDiagonalScale(Gmat, diag, diag));
8066: PetscCall(VecDestroy(&diag));
8067: }
8068: PetscCall(MatViewFromOptions(Gmat, NULL, "-mat_graph_view"));
8069: if (filter >= 0) {
8070: PetscCall(MatFilter(Gmat, filter, PETSC_TRUE, PETSC_TRUE));
8071: PetscCall(MatViewFromOptions(Gmat, NULL, "-mat_filter_graph_view"));
8072: }
8073: *a_Gmat = Gmat;
8074: PetscFunctionReturn(PETSC_SUCCESS);
8075: }
8077: /*
8078: Special version for direct calls from Fortran
8079: */
8081: /* Change these macros so can be used in void function */
8082: /* Identical to PetscCallVoid, except it assigns to *_ierr */
8083: #undef PetscCall
8084: #define PetscCall(...) \
8085: do { \
8086: PetscErrorCode ierr_msv_mpiaij = __VA_ARGS__; \
8087: if (PetscUnlikely(ierr_msv_mpiaij)) { \
8088: *_ierr = PetscError(PETSC_COMM_SELF, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr_msv_mpiaij, PETSC_ERROR_REPEAT, " "); \
8089: return; \
8090: } \
8091: } while (0)
8093: #undef SETERRQ
8094: #define SETERRQ(comm, ierr, ...) \
8095: do { \
8096: *_ierr = PetscError(comm, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr, PETSC_ERROR_INITIAL, __VA_ARGS__); \
8097: return; \
8098: } while (0)
8100: #if defined(PETSC_HAVE_FORTRAN_CAPS)
8101: #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ
8102: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
8103: #define matsetvaluesmpiaij_ matsetvaluesmpiaij
8104: #else
8105: #endif
8106: PETSC_EXTERN void matsetvaluesmpiaij_(Mat *mmat, PetscInt *mm, const PetscInt im[], PetscInt *mn, const PetscInt in[], const PetscScalar v[], InsertMode *maddv, PetscErrorCode *_ierr)
8107: {
8108: Mat mat = *mmat;
8109: PetscInt m = *mm, n = *mn;
8110: InsertMode addv = *maddv;
8111: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
8112: PetscScalar value;
8114: MatCheckPreallocated(mat, 1);
8115: if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv;
8116: else PetscCheck(mat->insertmode == addv, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Cannot mix add values and insert values");
8117: {
8118: PetscInt i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend;
8119: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend, row, col;
8120: PetscBool roworiented = aij->roworiented;
8122: /* Some Variables required in the macro */
8123: Mat A = aij->A;
8124: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
8125: PetscInt *aimax = a->imax, *ai = a->i, *ailen = a->ilen, *aj = a->j;
8126: MatScalar *aa;
8127: PetscBool ignorezeroentries = ((a->ignorezeroentries && (addv == ADD_VALUES)) ? PETSC_TRUE : PETSC_FALSE);
8128: Mat B = aij->B;
8129: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
8130: PetscInt *bimax = b->imax, *bi = b->i, *bilen = b->ilen, *bj = b->j, bm = aij->B->rmap->n, am = aij->A->rmap->n;
8131: MatScalar *ba;
8132: /* This variable below is only for the PETSC_HAVE_VIENNACL or PETSC_HAVE_CUDA cases, but we define it in all cases because we
8133: * cannot use "#if defined" inside a macro. */
8134: PETSC_UNUSED PetscBool inserted = PETSC_FALSE;
8136: PetscInt *rp1, *rp2, ii, nrow1, nrow2, _i, rmax1, rmax2, N, low1, high1, low2, high2, t, lastcol1, lastcol2;
8137: PetscInt nonew = a->nonew;
8138: MatScalar *ap1, *ap2;
8140: PetscFunctionBegin;
8141: PetscCall(MatSeqAIJGetArray(A, &aa));
8142: PetscCall(MatSeqAIJGetArray(B, &ba));
8143: for (i = 0; i < m; i++) {
8144: if (im[i] < 0) continue;
8145: 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);
8146: if (im[i] >= rstart && im[i] < rend) {
8147: row = im[i] - rstart;
8148: lastcol1 = -1;
8149: rp1 = aj + ai[row];
8150: ap1 = aa + ai[row];
8151: rmax1 = aimax[row];
8152: nrow1 = ailen[row];
8153: low1 = 0;
8154: high1 = nrow1;
8155: lastcol2 = -1;
8156: rp2 = bj + bi[row];
8157: ap2 = ba + bi[row];
8158: rmax2 = bimax[row];
8159: nrow2 = bilen[row];
8160: low2 = 0;
8161: high2 = nrow2;
8163: for (j = 0; j < n; j++) {
8164: if (roworiented) value = v[i * n + j];
8165: else value = v[i + j * m];
8166: if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && im[i] != in[j]) continue;
8167: if (in[j] >= cstart && in[j] < cend) {
8168: col = in[j] - cstart;
8169: MatSetValues_SeqAIJ_A_Private(row, col, value, addv, im[i], in[j]);
8170: } else if (in[j] < 0) continue;
8171: else if (PetscUnlikelyDebug(in[j] >= mat->cmap->N)) {
8172: SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large: col %" PetscInt_FMT " max %" PetscInt_FMT, in[j], mat->cmap->N - 1);
8173: } else {
8174: if (mat->was_assembled) {
8175: if (!aij->colmap) PetscCall(MatCreateColmap_MPIAIJ_Private(mat));
8176: #if defined(PETSC_USE_CTABLE)
8177: PetscCall(PetscHMapIGetWithDefault(aij->colmap, in[j] + 1, 0, &col));
8178: col--;
8179: #else
8180: col = aij->colmap[in[j]] - 1;
8181: #endif
8182: if (col < 0 && !((Mat_SeqAIJ *)aij->A->data)->nonew) {
8183: PetscCall(MatDisAssemble_MPIAIJ(mat, PETSC_FALSE));
8184: col = in[j];
8185: /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
8186: B = aij->B;
8187: b = (Mat_SeqAIJ *)B->data;
8188: bimax = b->imax;
8189: bi = b->i;
8190: bilen = b->ilen;
8191: bj = b->j;
8192: rp2 = bj + bi[row];
8193: ap2 = ba + bi[row];
8194: rmax2 = bimax[row];
8195: nrow2 = bilen[row];
8196: low2 = 0;
8197: high2 = nrow2;
8198: bm = aij->B->rmap->n;
8199: ba = b->a;
8200: inserted = PETSC_FALSE;
8201: }
8202: } else col = in[j];
8203: MatSetValues_SeqAIJ_B_Private(row, col, value, addv, im[i], in[j]);
8204: }
8205: }
8206: } else if (!aij->donotstash) {
8207: if (roworiented) {
8208: PetscCall(MatStashValuesRow_Private(&mat->stash, im[i], n, in, v + i * n, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
8209: } else {
8210: PetscCall(MatStashValuesCol_Private(&mat->stash, im[i], n, in, v + i, m, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
8211: }
8212: }
8213: }
8214: PetscCall(MatSeqAIJRestoreArray(A, &aa));
8215: PetscCall(MatSeqAIJRestoreArray(B, &ba));
8216: }
8217: PetscFunctionReturnVoid();
8218: }
8220: /* Undefining these here since they were redefined from their original definition above! No
8221: * other PETSc functions should be defined past this point, as it is impossible to recover the
8222: * original definitions */
8223: #undef PetscCall
8224: #undef SETERRQ