Actual source code: bddcprivate.c
1: #include <../src/mat/impls/aij/seq/aij.h>
2: #include <petsc/private/pcbddcimpl.h>
3: #include <petsc/private/pcbddcprivateimpl.h>
4: #include <petsc/private/kernels/blockinvert.h>
5: #include <../src/mat/impls/dense/seq/dense.h>
6: #include <petscdmplex.h>
7: #include <petscblaslapack.h>
8: #include <petsc/private/sfimpl.h>
9: #include <petsc/private/dmpleximpl.h>
10: #include <petscdmda.h>
12: static PetscErrorCode MatMPIAIJRestrict(Mat, MPI_Comm, Mat *);
14: /* if range is true, it returns B s.t. span{B} = range(A)
15: if range is false, it returns B s.t. range(B) _|_ range(A) */
16: static PetscErrorCode MatDenseOrthogonalRangeOrComplement(Mat A, PetscBool range, PetscInt lw, PetscScalar *work, PetscReal *rwork, Mat *B)
17: {
18: PetscScalar *uwork, *data, *U, ds = 0.;
19: PetscReal *sing;
20: PetscBLASInt bM, bN, lwork, lierr, di = 1;
21: PetscInt ulw, i, nr, nc, n;
22: #if defined(PETSC_USE_COMPLEX)
23: PetscReal *rwork2;
24: #endif
26: PetscFunctionBegin;
27: PetscCall(MatGetSize(A, &nr, &nc));
28: if (!nr || !nc) PetscFunctionReturn(PETSC_SUCCESS);
30: /* workspace */
31: if (!work) {
32: ulw = PetscMax(PetscMax(1, 5 * PetscMin(nr, nc)), 3 * PetscMin(nr, nc) + PetscMax(nr, nc));
33: PetscCall(PetscMalloc1(ulw, &uwork));
34: } else {
35: ulw = lw;
36: uwork = work;
37: }
38: n = PetscMin(nr, nc);
39: if (!rwork) {
40: PetscCall(PetscMalloc1(n, &sing));
41: } else {
42: sing = rwork;
43: }
45: /* SVD */
46: PetscCall(PetscMalloc1(nr * nr, &U));
47: PetscCall(PetscBLASIntCast(nr, &bM));
48: PetscCall(PetscBLASIntCast(nc, &bN));
49: PetscCall(PetscBLASIntCast(ulw, &lwork));
50: PetscCall(MatDenseGetArray(A, &data));
51: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
52: #if !defined(PETSC_USE_COMPLEX)
53: PetscCallBLAS("LAPACKgesvd", LAPACKgesvd_("A", "N", &bM, &bN, data, &bM, sing, U, &bM, &ds, &di, uwork, &lwork, &lierr));
54: #else
55: PetscCall(PetscMalloc1(5 * n, &rwork2));
56: PetscCallBLAS("LAPACKgesvd", LAPACKgesvd_("A", "N", &bM, &bN, data, &bM, sing, U, &bM, &ds, &di, uwork, &lwork, rwork2, &lierr));
57: PetscCall(PetscFree(rwork2));
58: #endif
59: PetscCall(PetscFPTrapPop());
60: PetscCheck(!lierr, PETSC_COMM_SELF, PETSC_ERR_LIB, "Error in GESVD Lapack routine %" PetscBLASInt_FMT, lierr);
61: PetscCall(MatDenseRestoreArray(A, &data));
62: for (i = 0; i < n; i++)
63: if (sing[i] < PETSC_SMALL) break;
64: if (!rwork) PetscCall(PetscFree(sing));
65: if (!work) PetscCall(PetscFree(uwork));
66: /* create B */
67: if (!range) {
68: PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, nr, nr - i, NULL, B));
69: PetscCall(MatDenseGetArray(*B, &data));
70: PetscCall(PetscArraycpy(data, U + nr * i, (nr - i) * nr));
71: } else {
72: PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, nr, i, NULL, B));
73: PetscCall(MatDenseGetArray(*B, &data));
74: PetscCall(PetscArraycpy(data, U, i * nr));
75: }
76: PetscCall(MatDenseRestoreArray(*B, &data));
77: PetscCall(PetscFree(U));
78: PetscFunctionReturn(PETSC_SUCCESS);
79: }
81: /* TODO REMOVE */
82: #if defined(PRINT_GDET)
83: static int inc = 0;
84: static int lev = 0;
85: #endif
87: static PetscErrorCode PCBDDCComputeNedelecChangeEdge(Mat lG, IS edge, IS extrow, IS extcol, IS corners, Mat *Gins, Mat *GKins, PetscScalar cvals[2], PetscScalar *work, PetscReal *rwork)
88: {
89: Mat GE, GEd;
90: PetscInt rsize, csize, esize;
91: PetscScalar *ptr;
93: PetscFunctionBegin;
94: PetscCall(ISGetSize(edge, &esize));
95: if (!esize) PetscFunctionReturn(PETSC_SUCCESS);
96: PetscCall(ISGetSize(extrow, &rsize));
97: PetscCall(ISGetSize(extcol, &csize));
99: /* gradients */
100: ptr = work + 5 * esize;
101: PetscCall(MatCreateSubMatrix(lG, extrow, extcol, MAT_INITIAL_MATRIX, &GE));
102: PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, rsize, csize, ptr, Gins));
103: PetscCall(MatConvert(GE, MATSEQDENSE, MAT_REUSE_MATRIX, Gins));
104: PetscCall(MatDestroy(&GE));
106: /* constants */
107: ptr += rsize * csize;
108: PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, esize, csize, ptr, &GEd));
109: PetscCall(MatCreateSubMatrix(lG, edge, extcol, MAT_INITIAL_MATRIX, &GE));
110: PetscCall(MatConvert(GE, MATSEQDENSE, MAT_REUSE_MATRIX, &GEd));
111: PetscCall(MatDestroy(&GE));
112: PetscCall(MatDenseOrthogonalRangeOrComplement(GEd, PETSC_FALSE, 5 * esize, work, rwork, GKins));
113: PetscCall(MatDestroy(&GEd));
115: if (corners) {
116: Mat GEc;
117: const PetscScalar *vals;
118: PetscScalar v;
120: PetscCall(MatCreateSubMatrix(lG, edge, corners, MAT_INITIAL_MATRIX, &GEc));
121: PetscCall(MatTransposeMatMult(GEc, *GKins, MAT_INITIAL_MATRIX, 1.0, &GEd));
122: PetscCall(MatDenseGetArrayRead(GEd, &vals));
123: /* v = PetscAbsScalar(vals[0]); */
124: v = 1.;
125: cvals[0] = vals[0] / v;
126: cvals[1] = vals[1] / v;
127: PetscCall(MatDenseRestoreArrayRead(GEd, &vals));
128: PetscCall(MatScale(*GKins, 1. / v));
129: #if defined(PRINT_GDET)
130: {
131: PetscViewer viewer;
132: char filename[256];
133: PetscCall(PetscSNPrintf(filename, PETSC_STATIC_ARRAY_LENGTH(filename), "Gdet_l%d_r%d_cc%d.m", lev, PetscGlobalRank, inc++));
134: PetscCall(PetscViewerASCIIOpen(PETSC_COMM_SELF, filename, &viewer));
135: PetscCall(PetscViewerPushFormat(viewer, PETSC_VIEWER_ASCII_MATLAB));
136: PetscCall(PetscObjectSetName((PetscObject)GEc, "GEc"));
137: PetscCall(MatView(GEc, viewer));
138: PetscCall(PetscObjectSetName((PetscObject)*GKins, "GK"));
139: PetscCall(MatView(*GKins, viewer));
140: PetscCall(PetscObjectSetName((PetscObject)GEd, "Gproj"));
141: PetscCall(MatView(GEd, viewer));
142: PetscCall(PetscViewerDestroy(&viewer));
143: }
144: #endif
145: PetscCall(MatDestroy(&GEd));
146: PetscCall(MatDestroy(&GEc));
147: }
148: PetscFunctionReturn(PETSC_SUCCESS);
149: }
151: static PetscErrorCode MatAIJExtractRows(Mat, IS, Mat *);
153: PetscErrorCode PCBDDCNedelecSupport(PC pc)
154: {
155: PC_BDDC *pcbddc = (PC_BDDC *)pc->data;
156: Mat_IS *matis = (Mat_IS *)pc->pmat->data;
157: Mat G, T, conn, lG, lGt, lGis, lGall, lGe, lGinit;
158: PetscSF sfv;
159: ISLocalToGlobalMapping el2g, vl2g, fl2g, al2g;
160: MPI_Comm comm;
161: IS lned, primals, allprimals, nedfieldlocal, elements_corners = NULL;
162: IS *eedges, *extrows, *extcols, *alleedges;
163: PetscBT btv, bte, btvc, btb, btbd, btvcand, btvi, btee, bter;
164: PetscScalar *vals, *work;
165: PetscReal *rwork;
166: const PetscInt *idxs, *ii, *jj, *iit, *jjt;
167: PetscInt ne, nv, Lv, order, n, field;
168: PetscInt i, j, extmem, cum, maxsize, nee;
169: PetscInt *extrow, *extrowcum, *marks, *vmarks, *gidxs;
170: PetscInt *sfvleaves, *sfvroots;
171: PetscInt *corners, *cedges;
172: PetscInt *ecount, **eneighs, *vcount, **vneighs;
173: PetscInt *emarks;
174: PetscBool print, eerr, done, lrc[2], conforming, global, setprimal;
176: PetscFunctionBegin;
177: /* If the discrete gradient is defined for a subset of dofs and global is true,
178: it assumes G is given in global ordering for all the dofs.
179: Otherwise, the ordering is global for the Nedelec field */
180: order = pcbddc->nedorder;
181: conforming = pcbddc->conforming;
182: field = pcbddc->nedfield;
183: global = pcbddc->nedglobal;
184: setprimal = PETSC_FALSE;
185: print = PETSC_FALSE;
187: /* Command line customization */
188: PetscOptionsBegin(PetscObjectComm((PetscObject)pc), ((PetscObject)pc)->prefix, "BDDC Nedelec options", "PC");
189: PetscCall(PetscOptionsBool("-pc_bddc_nedelec_field_primal", "All edge dofs set as primals: Toselli's algorithm C", NULL, setprimal, &setprimal, NULL));
190: /* print debug info and adaptive order TODO: to be removed */
191: PetscCall(PetscOptionsInt("-pc_bddc_nedelec_order", "Test variable order code (to be removed)", NULL, order, &order, NULL));
192: PetscCall(PetscOptionsBool("-pc_bddc_nedelec_print", "Print debug info", NULL, print, &print, NULL));
193: PetscOptionsEnd();
195: /* Return if there are no edges in the decomposition */
196: PetscCall(MatISGetLocalToGlobalMapping(pc->pmat, &al2g, NULL));
197: PetscCall(ISLocalToGlobalMappingGetSize(al2g, &n));
198: PetscCall(PetscObjectGetComm((PetscObject)pc, &comm));
199: PetscCall(VecGetArrayRead(matis->counter, (const PetscScalar **)&vals));
200: lrc[0] = PETSC_FALSE;
201: for (i = 0; i < n; i++) {
202: if (PetscRealPart(vals[i]) > 2.) {
203: lrc[0] = PETSC_TRUE;
204: break;
205: }
206: }
207: PetscCall(VecRestoreArrayRead(matis->counter, (const PetscScalar **)&vals));
208: PetscCallMPI(MPIU_Allreduce(&lrc[0], &lrc[1], 1, MPI_C_BOOL, MPI_LOR, comm));
209: if (!lrc[1]) PetscFunctionReturn(PETSC_SUCCESS);
211: /* Get Nedelec field */
212: PetscCheck(!pcbddc->n_ISForDofsLocal || field < pcbddc->n_ISForDofsLocal, comm, PETSC_ERR_USER, "Invalid field for Nedelec %" PetscInt_FMT ": number of fields is %" PetscInt_FMT, field, pcbddc->n_ISForDofsLocal);
213: if (pcbddc->n_ISForDofsLocal && field >= 0) {
214: PetscCall(PetscObjectReference((PetscObject)pcbddc->ISForDofsLocal[field]));
215: nedfieldlocal = pcbddc->ISForDofsLocal[field];
216: PetscCall(ISGetLocalSize(nedfieldlocal, &ne));
217: } else if (!pcbddc->n_ISForDofsLocal && field != PETSC_DECIDE) {
218: ne = n;
219: nedfieldlocal = NULL;
220: global = PETSC_TRUE;
221: } else if (field == PETSC_DECIDE) {
222: PetscInt rst, ren, *idx;
224: PetscCall(PetscArrayzero(matis->sf_leafdata, n));
225: PetscCall(PetscArrayzero(matis->sf_rootdata, pc->pmat->rmap->n));
226: PetscCall(MatGetOwnershipRange(pcbddc->discretegradient, &rst, &ren));
227: for (i = rst; i < ren; i++) {
228: PetscInt nc;
230: PetscCall(MatGetRow(pcbddc->discretegradient, i, &nc, NULL, NULL));
231: if (nc > 1) matis->sf_rootdata[i - rst] = 1;
232: PetscCall(MatRestoreRow(pcbddc->discretegradient, i, &nc, NULL, NULL));
233: }
234: PetscCall(PetscSFBcastBegin(matis->sf, MPIU_INT, matis->sf_rootdata, matis->sf_leafdata, MPI_REPLACE));
235: PetscCall(PetscSFBcastEnd(matis->sf, MPIU_INT, matis->sf_rootdata, matis->sf_leafdata, MPI_REPLACE));
236: PetscCall(PetscMalloc1(n, &idx));
237: for (i = 0, ne = 0; i < n; i++)
238: if (matis->sf_leafdata[i]) idx[ne++] = i;
239: PetscCall(ISCreateGeneral(comm, ne, idx, PETSC_OWN_POINTER, &nedfieldlocal));
240: } else {
241: SETERRQ(comm, PETSC_ERR_USER, "When multiple fields are present, the Nedelec field has to be specified");
242: }
244: /* Sanity checks */
245: PetscCheck(order || conforming, comm, PETSC_ERR_SUP, "Variable order and non-conforming spaces are not supported at the same time");
246: PetscCheck(!pcbddc->user_ChangeOfBasisMatrix, comm, PETSC_ERR_SUP, "Cannot generate Nedelec support with user defined change of basis");
247: PetscCheck(!order || (ne % order == 0), PETSC_COMM_SELF, PETSC_ERR_USER, "The number of local edge dofs %" PetscInt_FMT " is not a multiple of the order %" PetscInt_FMT, ne, order);
249: /* Just set primal dofs and return */
250: if (setprimal) {
251: IS enedfieldlocal;
252: PetscInt *eidxs;
254: PetscCall(PetscMalloc1(ne, &eidxs));
255: PetscCall(VecGetArrayRead(matis->counter, (const PetscScalar **)&vals));
256: if (nedfieldlocal) {
257: PetscCall(ISGetIndices(nedfieldlocal, &idxs));
258: for (i = 0, cum = 0; i < ne; i++) {
259: if (PetscRealPart(vals[idxs[i]]) > 2.) eidxs[cum++] = idxs[i];
260: }
261: PetscCall(ISRestoreIndices(nedfieldlocal, &idxs));
262: } else {
263: for (i = 0, cum = 0; i < ne; i++) {
264: if (PetscRealPart(vals[i]) > 2.) eidxs[cum++] = i;
265: }
266: }
267: PetscCall(VecRestoreArrayRead(matis->counter, (const PetscScalar **)&vals));
268: PetscCall(ISCreateGeneral(comm, cum, eidxs, PETSC_COPY_VALUES, &enedfieldlocal));
269: PetscCall(PCBDDCSetPrimalVerticesLocalIS(pc, enedfieldlocal));
270: PetscCall(PetscFree(eidxs));
271: PetscCall(ISDestroy(&nedfieldlocal));
272: PetscCall(ISDestroy(&enedfieldlocal));
273: PetscFunctionReturn(PETSC_SUCCESS);
274: }
276: /* Compute some l2g maps */
277: if (nedfieldlocal) {
278: IS is;
280: /* need to map from the local Nedelec field to local numbering */
281: PetscCall(ISLocalToGlobalMappingCreateIS(nedfieldlocal, &fl2g));
282: /* need to map from the local Nedelec field to global numbering for the whole dofs*/
283: PetscCall(ISLocalToGlobalMappingApplyIS(al2g, nedfieldlocal, &is));
284: PetscCall(ISLocalToGlobalMappingCreateIS(is, &al2g));
285: /* need to map from the local Nedelec field to global numbering (for Nedelec only) */
286: if (global) {
287: PetscCall(PetscObjectReference((PetscObject)al2g));
288: el2g = al2g;
289: } else {
290: IS gis;
292: PetscCall(ISRenumber(is, NULL, NULL, &gis));
293: PetscCall(ISLocalToGlobalMappingCreateIS(gis, &el2g));
294: PetscCall(ISDestroy(&gis));
295: }
296: PetscCall(ISDestroy(&is));
297: } else {
298: /* one ref for the destruction of al2g, one for el2g */
299: PetscCall(PetscObjectReference((PetscObject)al2g));
300: PetscCall(PetscObjectReference((PetscObject)al2g));
301: el2g = al2g;
302: fl2g = NULL;
303: }
305: /* Start communication to drop connections for interior edges (for cc analysis only) */
306: PetscCall(PetscArrayzero(matis->sf_leafdata, n));
307: PetscCall(PetscArrayzero(matis->sf_rootdata, pc->pmat->rmap->n));
308: if (nedfieldlocal) {
309: PetscCall(ISGetIndices(nedfieldlocal, &idxs));
310: for (i = 0; i < ne; i++) matis->sf_leafdata[idxs[i]] = 1;
311: PetscCall(ISRestoreIndices(nedfieldlocal, &idxs));
312: } else {
313: for (i = 0; i < ne; i++) matis->sf_leafdata[i] = 1;
314: }
315: PetscCall(PetscSFReduceBegin(matis->sf, MPIU_INT, matis->sf_leafdata, matis->sf_rootdata, MPI_SUM));
316: PetscCall(PetscSFReduceEnd(matis->sf, MPIU_INT, matis->sf_leafdata, matis->sf_rootdata, MPI_SUM));
318: /* There's no way to detect all possible corner candidates in a element-by-element case in a pure algebraic setting
319: Firedrake attaches a index set to identify them upfront. If it is present, we assume we are in such a case */
320: if (matis->allow_repeated) PetscCall(PetscObjectQuery((PetscObject)pcbddc->discretegradient, "_elements_corners", (PetscObject *)&elements_corners));
322: /* drop connections with interior edges to avoid unneeded communications and memory movements */
323: PetscCall(MatViewFromOptions(pcbddc->discretegradient, (PetscObject)pc, "-pc_bddc_discrete_gradient_view"));
324: PetscCall(MatDuplicate(pcbddc->discretegradient, MAT_COPY_VALUES, &G));
325: PetscCall(MatSetOption(G, MAT_KEEP_NONZERO_PATTERN, PETSC_FALSE));
326: if (global) {
327: PetscInt rst;
329: PetscCall(MatGetOwnershipRange(G, &rst, NULL));
330: for (i = 0, cum = 0; i < pc->pmat->rmap->n; i++) {
331: if (matis->sf_rootdata[i] < 2) matis->sf_rootdata[cum++] = i + rst;
332: }
333: PetscCall(MatSetOption(G, MAT_NO_OFF_PROC_ZERO_ROWS, PETSC_TRUE));
334: PetscCall(MatZeroRows(G, cum, matis->sf_rootdata, 0., NULL, NULL));
335: } else {
336: PetscInt *tbz;
338: PetscCall(PetscMalloc1(ne, &tbz));
339: PetscCall(PetscSFBcastBegin(matis->sf, MPIU_INT, matis->sf_rootdata, matis->sf_leafdata, MPI_REPLACE));
340: PetscCall(PetscSFBcastEnd(matis->sf, MPIU_INT, matis->sf_rootdata, matis->sf_leafdata, MPI_REPLACE));
341: PetscCall(ISGetIndices(nedfieldlocal, &idxs));
342: for (i = 0, cum = 0; i < ne; i++)
343: if (matis->sf_leafdata[idxs[i]] == 1) tbz[cum++] = i;
344: PetscCall(ISRestoreIndices(nedfieldlocal, &idxs));
345: PetscCall(ISLocalToGlobalMappingApply(el2g, cum, tbz, tbz));
346: PetscCall(MatZeroRows(G, cum, tbz, 0., NULL, NULL));
347: PetscCall(PetscFree(tbz));
348: }
350: /* Extract subdomain relevant rows of G */
351: PetscCall(ISLocalToGlobalMappingGetIndices(el2g, &idxs));
352: PetscCall(ISCreateGeneral(comm, ne, idxs, PETSC_USE_POINTER, &lned));
353: PetscCall(MatAIJExtractRows(G, lned, &lGall));
354: /* PetscCall(MatCreateSubMatrix(G, lned, NULL, MAT_INITIAL_MATRIX, &lGall)); */
355: PetscCall(ISLocalToGlobalMappingRestoreIndices(el2g, &idxs));
356: PetscCall(ISDestroy(&lned));
357: PetscCall(MatConvert(lGall, MATIS, MAT_INITIAL_MATRIX, &lGis));
358: PetscCall(MatDestroy(&lGall));
359: PetscCall(MatISGetLocalMat(lGis, &lG));
360: if (matis->allow_repeated) { /* multi-element support */
361: Mat *lGn, B;
362: IS *is_rows, *tcols, tmap, nmap;
363: PetscInt subnv;
364: const PetscInt *subvidxs;
365: ISLocalToGlobalMapping mapn;
367: PetscCall(PetscCalloc1(pcbddc->n_local_subs * pcbddc->n_local_subs, &lGn));
368: PetscCall(PetscMalloc1(pcbddc->n_local_subs, &is_rows));
369: PetscCall(PetscMalloc1(pcbddc->n_local_subs, &tcols));
370: for (PetscInt i = 0; i < pcbddc->n_local_subs; i++) {
371: if (fl2g) {
372: PetscCall(ISGlobalToLocalMappingApplyIS(fl2g, IS_GTOLM_MASK, pcbddc->local_subs[i], &is_rows[i]));
373: } else {
374: PetscCall(PetscObjectReference((PetscObject)pcbddc->local_subs[i]));
375: is_rows[i] = pcbddc->local_subs[i];
376: }
377: PetscCall(MatCreateSubMatrix(lG, is_rows[i], NULL, MAT_INITIAL_MATRIX, &lGn[i * (1 + pcbddc->n_local_subs)]));
378: PetscCall(MatSeqAIJCompactOutExtraColumns_SeqAIJ(lGn[i * (1 + pcbddc->n_local_subs)], &mapn));
379: PetscCall(ISLocalToGlobalMappingGetSize(mapn, &subnv));
380: PetscCall(ISLocalToGlobalMappingGetIndices(mapn, &subvidxs));
381: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, subnv, subvidxs, PETSC_COPY_VALUES, &tcols[i]));
382: PetscCall(ISLocalToGlobalMappingRestoreIndices(mapn, &subvidxs));
383: PetscCall(ISLocalToGlobalMappingDestroy(&mapn));
384: }
386: /* Create new MATIS with repeated vertices */
387: PetscCall(MatCreate(comm, &B));
388: PetscCall(MatSetSizes(B, lGis->rmap->n, lGis->cmap->n, lGis->rmap->N, lGis->cmap->N));
389: PetscCall(MatSetType(B, MATIS));
390: PetscCall(MatISSetAllowRepeated(B, PETSC_TRUE));
391: PetscCall(ISConcatenate(PETSC_COMM_SELF, pcbddc->n_local_subs, tcols, &tmap));
392: PetscCall(ISLocalToGlobalMappingApplyIS(lGis->cmap->mapping, tmap, &nmap));
393: PetscCall(ISDestroy(&tmap));
394: PetscCall(ISGetLocalSize(nmap, &subnv));
395: PetscCall(ISGetIndices(nmap, &subvidxs));
396: PetscCall(ISCreateGeneral(comm, subnv, subvidxs, PETSC_USE_POINTER, &tmap));
397: PetscCall(ISRestoreIndices(nmap, &subvidxs));
398: PetscCall(ISLocalToGlobalMappingCreateIS(tmap, &mapn));
399: PetscCall(ISDestroy(&tmap));
400: PetscCall(ISDestroy(&nmap));
401: PetscCall(MatSetLocalToGlobalMapping(B, lGis->rmap->mapping, mapn));
402: PetscCall(ISLocalToGlobalMappingDestroy(&mapn));
403: PetscCall(MatCreateNest(PETSC_COMM_SELF, pcbddc->n_local_subs, is_rows, pcbddc->n_local_subs, NULL, lGn, &lG));
404: for (PetscInt i = 0; i < pcbddc->n_local_subs; i++) {
405: PetscCall(MatDestroy(&lGn[i * (1 + pcbddc->n_local_subs)]));
406: PetscCall(ISDestroy(&is_rows[i]));
407: PetscCall(ISDestroy(&tcols[i]));
408: }
409: PetscCall(MatConvert(lG, MATSEQAIJ, MAT_INPLACE_MATRIX, &lG));
410: PetscCall(PetscFree(lGn));
411: PetscCall(PetscFree(is_rows));
412: PetscCall(PetscFree(tcols));
413: PetscCall(MatISSetLocalMat(B, lG));
414: PetscCall(MatDestroy(&lG));
416: PetscCall(MatDestroy(&lGis));
417: lGis = B;
419: lGis->assembled = PETSC_TRUE;
420: }
421: PetscCall(MatViewFromOptions(lGis, (PetscObject)pc, "-pc_bddc_nedelec_init_G_view"));
423: /* SF for nodal dofs communications */
424: PetscCall(MatGetLocalSize(G, NULL, &Lv));
425: PetscCall(MatISGetLocalToGlobalMapping(lGis, NULL, &vl2g));
426: PetscCall(PetscObjectReference((PetscObject)vl2g));
427: PetscCall(ISLocalToGlobalMappingGetSize(vl2g, &nv));
428: PetscCall(PetscSFCreate(comm, &sfv));
429: PetscCall(ISLocalToGlobalMappingGetIndices(vl2g, &idxs));
430: PetscCall(PetscSFSetGraphLayout(sfv, lGis->cmap, nv, NULL, PETSC_OWN_POINTER, idxs));
431: PetscCall(ISLocalToGlobalMappingRestoreIndices(vl2g, &idxs));
433: if (elements_corners) {
434: IS tmp;
435: Vec global, local;
436: Mat_IS *tGis = (Mat_IS *)lGis->data;
438: PetscCall(MatCreateVecs(lGis, &global, NULL));
439: PetscCall(MatCreateVecs(tGis->A, &local, NULL));
440: PetscCall(PCBDDCGlobalToLocal(tGis->cctx, global, local, elements_corners, &tmp));
441: PetscCall(VecDestroy(&global));
442: PetscCall(VecDestroy(&local));
443: elements_corners = tmp;
444: }
446: /* Destroy temporary G */
447: PetscCall(MatISGetLocalMat(lGis, &lG));
448: PetscCall(PetscObjectReference((PetscObject)lG));
449: PetscCall(MatDestroy(&G));
450: PetscCall(MatDestroy(&lGis));
452: if (print) {
453: PetscCall(PetscObjectSetName((PetscObject)lG, "initial_lG"));
454: PetscCall(MatView(lG, NULL));
455: }
457: /* Save lG for values insertion in change of basis */
458: PetscCall(MatDuplicate(lG, MAT_COPY_VALUES, &lGinit));
460: /* Analyze the edge-nodes connections (duplicate lG) */
461: PetscCall(MatDuplicate(lG, MAT_COPY_VALUES, &lGe));
462: PetscCall(MatSetOption(lGe, MAT_KEEP_NONZERO_PATTERN, PETSC_FALSE));
463: PetscCall(PetscBTCreate(nv, &btv));
464: PetscCall(PetscBTCreate(ne, &bte));
465: PetscCall(PetscBTCreate(ne, &btb));
466: PetscCall(PetscBTCreate(ne, &btbd));
467: /* need to import the boundary specification to ensure the
468: proper detection of coarse edges' endpoints */
469: if (pcbddc->DirichletBoundariesLocal) {
470: IS is;
472: if (fl2g) {
473: PetscCall(ISGlobalToLocalMappingApplyIS(fl2g, IS_GTOLM_MASK, pcbddc->DirichletBoundariesLocal, &is));
474: } else {
475: is = pcbddc->DirichletBoundariesLocal;
476: }
477: PetscCall(ISGetLocalSize(is, &cum));
478: PetscCall(ISGetIndices(is, &idxs));
479: for (i = 0; i < cum; i++) {
480: if (idxs[i] >= 0 && idxs[i] < ne) {
481: PetscCall(PetscBTSet(btb, idxs[i]));
482: PetscCall(PetscBTSet(btbd, idxs[i]));
483: }
484: }
485: PetscCall(ISRestoreIndices(is, &idxs));
486: if (fl2g) PetscCall(ISDestroy(&is));
487: }
488: if (pcbddc->NeumannBoundariesLocal) {
489: IS is;
491: if (fl2g) {
492: PetscCall(ISGlobalToLocalMappingApplyIS(fl2g, IS_GTOLM_MASK, pcbddc->NeumannBoundariesLocal, &is));
493: } else {
494: is = pcbddc->NeumannBoundariesLocal;
495: }
496: PetscCall(ISGetLocalSize(is, &cum));
497: PetscCall(ISGetIndices(is, &idxs));
498: for (i = 0; i < cum; i++) {
499: if (idxs[i] >= 0 && idxs[i] < ne) PetscCall(PetscBTSet(btb, idxs[i]));
500: }
501: PetscCall(ISRestoreIndices(is, &idxs));
502: if (fl2g) PetscCall(ISDestroy(&is));
503: }
505: /* Count neighs per dof */
506: PetscCall(ISLocalToGlobalMappingGetNodeInfo(el2g, NULL, &ecount, NULL));
507: PetscCall(ISLocalToGlobalMappingGetNodeInfo(vl2g, NULL, &vcount, NULL));
509: /* need to remove coarse faces' dofs and coarse edges' dirichlet dofs
510: for proper detection of coarse edges' endpoints */
511: PetscCall(PetscBTCreate(ne, &btee));
512: for (i = 0; i < ne; i++) {
513: if ((ecount[i] > 2 && !PetscBTLookup(btbd, i)) || (ecount[i] == 2 && PetscBTLookup(btb, i))) PetscCall(PetscBTSet(btee, i));
514: }
515: PetscCall(PetscMalloc1(ne, &marks));
516: if (!conforming) {
517: PetscCall(MatTranspose(lGe, MAT_INITIAL_MATRIX, &lGt));
518: PetscCall(MatGetRowIJ(lGt, 0, PETSC_FALSE, PETSC_FALSE, &i, &iit, &jjt, &done));
519: }
520: PetscCall(MatGetRowIJ(lGe, 0, PETSC_FALSE, PETSC_FALSE, &i, &ii, &jj, &done));
521: PetscCall(MatSeqAIJGetArray(lGe, &vals));
522: cum = 0;
523: for (i = 0; i < ne; i++) {
524: /* eliminate rows corresponding to edge dofs belonging to coarse faces */
525: if (!PetscBTLookup(btee, i)) {
526: marks[cum++] = i;
527: continue;
528: }
529: /* set badly connected edge dofs as primal */
530: if (!conforming) {
531: if (ii[i + 1] - ii[i] != order + 1) { /* every row of G on the coarse edge should list order+1 nodal dofs */
532: marks[cum++] = i;
533: PetscCall(PetscBTSet(bte, i));
534: for (j = ii[i]; j < ii[i + 1]; j++) PetscCall(PetscBTSet(btv, jj[j]));
535: } else {
536: /* every edge dofs should be connected through a certain number of nodal dofs
537: to other edge dofs belonging to coarse edges
538: - at most 2 endpoints
539: - order-1 interior nodal dofs
540: - no undefined nodal dofs (nconn < order)
541: */
542: PetscInt ends = 0, ints = 0, undef = 0;
543: for (j = ii[i]; j < ii[i + 1]; j++) {
544: PetscInt v = jj[j], k;
545: PetscInt nconn = iit[v + 1] - iit[v];
546: for (k = iit[v]; k < iit[v + 1]; k++)
547: if (!PetscBTLookup(btee, jjt[k])) nconn--;
548: if (nconn > order) ends++;
549: else if (nconn == order) ints++;
550: else undef++;
551: }
552: if (undef || ends > 2 || ints != order - 1) {
553: marks[cum++] = i;
554: PetscCall(PetscBTSet(bte, i));
555: for (j = ii[i]; j < ii[i + 1]; j++) PetscCall(PetscBTSet(btv, jj[j]));
556: }
557: }
558: }
559: /* We assume the order on the element edge is ii[i+1]-ii[i]-1 */
560: if (!order && ii[i + 1] != ii[i]) {
561: PetscScalar val = 1. / (ii[i + 1] - ii[i] - 1);
562: for (j = ii[i]; j < ii[i + 1]; j++) vals[j] = val;
563: }
564: }
565: PetscCall(PetscBTDestroy(&btee));
566: PetscCall(MatSeqAIJRestoreArray(lGe, &vals));
567: PetscCall(MatRestoreRowIJ(lGe, 0, PETSC_FALSE, PETSC_FALSE, &i, &ii, &jj, &done));
568: if (!conforming) {
569: PetscCall(MatRestoreRowIJ(lGt, 0, PETSC_FALSE, PETSC_FALSE, &i, &iit, &jjt, &done));
570: PetscCall(MatDestroy(&lGt));
571: }
572: PetscCall(MatZeroRows(lGe, cum, marks, 0., NULL, NULL));
574: /* identify splitpoints and corner candidates */
575: PetscCall(PetscMalloc2(nv, &sfvleaves, Lv, &sfvroots));
576: PetscCall(PetscBTCreate(nv, &btvcand));
577: if (elements_corners) {
578: PetscCall(ISGetLocalSize(elements_corners, &cum));
579: PetscCall(ISGetIndices(elements_corners, &idxs));
580: for (i = 0; i < cum; i++) PetscCall(PetscBTSet(btvcand, idxs[i]));
581: PetscCall(ISRestoreIndices(elements_corners, &idxs));
582: }
584: if (matis->allow_repeated) { /* assign a uniq global id to edge local subsets and communicate it with nodal space */
585: PetscSF emlsf, vmlsf;
586: PetscInt *eleaves, *vleaves, *meleaves, *mvleaves;
587: PetscInt cum_subs = 0, n_subs = pcbddc->n_local_subs, bs, emnr, emnl, vmnr, vmnl;
589: PetscCall(ISLocalToGlobalMappingGetBlockSize(el2g, &bs));
590: PetscCheck(bs == 1, comm, PETSC_ERR_SUP, "Not coded");
591: PetscCall(ISLocalToGlobalMappingGetBlockSize(vl2g, &bs));
592: PetscCheck(bs == 1, comm, PETSC_ERR_SUP, "Not coded");
594: PetscCall(ISLocalToGlobalMappingGetBlockMultiLeavesSF(el2g, &emlsf));
595: PetscCall(ISLocalToGlobalMappingGetBlockMultiLeavesSF(vl2g, &vmlsf));
597: PetscCall(PetscSFGetGraph(emlsf, &emnr, &emnl, NULL, NULL));
598: for (i = 0, j = 0; i < ne; i++) j += ecount[i];
599: PetscCheck(emnr == ne, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Invalid number of roots in edge multi-leaves SF %" PetscInt_FMT " != %" PetscInt_FMT, emnr, ne);
600: PetscCheck(emnl == j, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Invalid number of leaves in edge multi-leaves SF %" PetscInt_FMT " != %" PetscInt_FMT, emnl, j);
602: PetscCall(PetscSFGetGraph(vmlsf, &vmnr, &vmnl, NULL, NULL));
603: for (i = 0, j = 0; i < nv; i++) j += vcount[i];
604: PetscCheck(vmnr == nv, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Invalid number of roots in nodal multi-leaves SF %" PetscInt_FMT " != %" PetscInt_FMT, vmnr, nv);
605: PetscCheck(vmnl == j, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Invalid number of leaves in nodal multi-leaves SF %" PetscInt_FMT " != %" PetscInt_FMT, vmnl, j);
607: PetscCall(PetscMalloc1(ne, &eleaves));
608: PetscCall(PetscMalloc1(nv, &vleaves));
609: for (i = 0; i < ne; i++) eleaves[i] = PETSC_INT_MAX;
610: for (i = 0; i < nv; i++) vleaves[i] = PETSC_INT_MAX;
611: PetscCall(PetscMalloc1(emnl, &meleaves));
612: PetscCall(PetscMalloc1(vmnl, &mvleaves));
614: PetscCallMPI(MPI_Exscan(&n_subs, &cum_subs, 1, MPIU_INT, MPI_SUM, comm));
615: PetscCall(MatGetRowIJ(lGinit, 0, PETSC_FALSE, PETSC_FALSE, &i, &ii, &jj, &done));
616: for (i = 0; i < n_subs; i++) {
617: const PetscInt *idxs;
618: const PetscInt subid = cum_subs + i;
619: PetscInt ns;
621: PetscCall(ISGetLocalSize(pcbddc->local_subs[i], &ns));
622: PetscCall(ISGetIndices(pcbddc->local_subs[i], &idxs));
623: for (j = 0; j < ns; j++) {
624: const PetscInt e = idxs[j];
626: eleaves[e] = subid;
627: for (PetscInt k = ii[e]; k < ii[e + 1]; k++) vleaves[jj[k]] = subid;
628: }
629: PetscCall(ISRestoreIndices(pcbddc->local_subs[i], &idxs));
630: }
631: PetscCall(MatRestoreRowIJ(lGinit, 0, PETSC_FALSE, PETSC_FALSE, &i, &ii, &jj, &done));
632: PetscCall(PetscSFBcastBegin(emlsf, MPIU_INT, eleaves, meleaves, MPI_REPLACE));
633: PetscCall(PetscSFBcastEnd(emlsf, MPIU_INT, eleaves, meleaves, MPI_REPLACE));
634: PetscCall(PetscSFBcastBegin(vmlsf, MPIU_INT, vleaves, mvleaves, MPI_REPLACE));
635: PetscCall(PetscSFBcastEnd(vmlsf, MPIU_INT, vleaves, mvleaves, MPI_REPLACE));
636: PetscCall(PetscFree(eleaves));
637: PetscCall(PetscFree(vleaves));
639: PetscCall(PetscMalloc1(ne + 1, &eneighs));
640: eneighs[0] = meleaves;
641: for (i = 0; i < ne; i++) {
642: PetscCall(PetscSortInt(ecount[i], eneighs[i]));
643: eneighs[i + 1] = eneighs[i] + ecount[i];
644: }
645: PetscCall(PetscMalloc1(nv + 1, &vneighs));
646: vneighs[0] = mvleaves;
647: for (i = 0; i < nv; i++) {
648: PetscCall(PetscSortInt(vcount[i], vneighs[i]));
649: vneighs[i + 1] = vneighs[i] + vcount[i];
650: }
651: } else {
652: PetscCall(ISLocalToGlobalMappingGetNodeInfo(el2g, NULL, NULL, &eneighs));
653: PetscCall(ISLocalToGlobalMappingGetNodeInfo(vl2g, NULL, NULL, &vneighs));
654: }
656: PetscCall(MatTranspose(lGe, MAT_INITIAL_MATRIX, &lGt));
657: if (print) {
658: PetscCall(PetscObjectSetName((PetscObject)lGe, "edgerestr_lG"));
659: PetscCall(MatView(lGe, NULL));
660: PetscCall(PetscObjectSetName((PetscObject)lGt, "edgerestr_lGt"));
661: PetscCall(MatView(lGt, NULL));
662: }
663: PetscCall(MatGetRowIJ(lGt, 0, PETSC_FALSE, PETSC_FALSE, &i, &ii, &jj, &done));
664: PetscCall(MatSeqAIJGetArray(lGt, &vals));
665: for (i = 0; i < nv; i++) {
666: PetscInt ord = order, test = ii[i + 1] - ii[i], vc = vcount[i];
667: PetscBool sneighs = PETSC_TRUE, bdir = PETSC_FALSE;
668: if (!order) { /* variable order */
669: PetscReal vorder = 0.;
671: for (j = ii[i]; j < ii[i + 1]; j++) vorder += PetscRealPart(vals[j]);
672: test = PetscFloorReal(vorder + 10. * PETSC_SQRT_MACHINE_EPSILON);
673: PetscCheck(vorder - test <= PETSC_SQRT_MACHINE_EPSILON, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Unexpected value for vorder: %g (%" PetscInt_FMT ")", (double)vorder, test);
674: ord = 1;
675: }
676: for (j = ii[i]; j < ii[i + 1] && sneighs; j++) {
677: const PetscInt e = jj[j];
679: if (PetscBTLookup(btbd, e)) {
680: bdir = PETSC_TRUE;
681: break;
682: }
683: if (vc != ecount[e]) {
684: sneighs = PETSC_FALSE;
685: } else {
686: const PetscInt *vn = vneighs[i], *en = eneighs[e];
688: for (PetscInt k = 0; k < vc; k++) {
689: if (vn[k] != en[k]) {
690: sneighs = PETSC_FALSE;
691: break;
692: }
693: }
694: }
695: }
696: if (elements_corners) test = 0;
697: if (!sneighs || test >= 3 * ord || bdir) { /* splitpoints */
698: if (print) PetscCall(PetscPrintf(PETSC_COMM_SELF, "SPLITPOINT %" PetscInt_FMT " (%s %s %s)\n", i, PetscBools[!sneighs], PetscBools[test >= 3 * ord], PetscBools[bdir]));
699: PetscCall(PetscBTSet(btv, i));
700: } else if (test == ord) {
701: if (order == 1 || (!order && ii[i + 1] - ii[i] == 1)) {
702: if (print) PetscCall(PetscPrintf(PETSC_COMM_SELF, "ENDPOINT %" PetscInt_FMT "\n", i));
703: PetscCall(PetscBTSet(btv, i));
704: } else if (!elements_corners) {
705: if (print) PetscCall(PetscPrintf(PETSC_COMM_SELF, "CORNER CANDIDATE %" PetscInt_FMT "\n", i));
706: PetscCall(PetscBTSet(btvcand, i));
707: }
708: }
709: }
710: PetscCall(PetscBTDestroy(&btbd));
712: /* a candidate is valid if it is connected to another candidate via a non-primal edge dof */
713: if (order != 1) {
714: if (print) PetscCall(PetscPrintf(PETSC_COMM_SELF, "INSPECTING CANDIDATES\n"));
715: PetscCall(MatGetRowIJ(lGe, 0, PETSC_FALSE, PETSC_FALSE, &i, &iit, &jjt, &done));
716: for (i = 0; i < nv; i++) {
717: if (PetscBTLookup(btvcand, i)) {
718: PetscBool found = PETSC_FALSE;
719: for (j = ii[i]; j < ii[i + 1] && !found; j++) {
720: PetscInt k, e = jj[j];
721: if (PetscBTLookup(bte, e)) continue;
722: for (k = iit[e]; k < iit[e + 1]; k++) {
723: PetscInt v = jjt[k];
724: if (v != i && PetscBTLookup(btvcand, v)) {
725: found = PETSC_TRUE;
726: break;
727: }
728: }
729: }
730: if (!found) {
731: if (print) PetscCall(PetscPrintf(PETSC_COMM_SELF, " CANDIDATE %" PetscInt_FMT " CLEARED\n", i));
732: PetscCall(PetscBTClear(btvcand, i));
733: } else {
734: if (print) PetscCall(PetscPrintf(PETSC_COMM_SELF, " CANDIDATE %" PetscInt_FMT " ACCEPTED\n", i));
735: }
736: }
737: }
738: PetscCall(MatRestoreRowIJ(lGe, 0, PETSC_FALSE, PETSC_FALSE, &i, &iit, &jjt, &done));
739: }
740: PetscCall(MatSeqAIJRestoreArray(lGt, &vals));
741: PetscCall(MatRestoreRowIJ(lGt, 0, PETSC_FALSE, PETSC_FALSE, &i, &ii, &jj, &done));
742: PetscCall(MatDestroy(&lGe));
744: /* Get the local G^T explicitly */
745: PetscCall(MatDestroy(&lGt));
746: PetscCall(MatTranspose(lG, MAT_INITIAL_MATRIX, &lGt));
747: PetscCall(MatSetOption(lGt, MAT_KEEP_NONZERO_PATTERN, PETSC_FALSE));
749: /* Mark shared nodal dofs */
750: PetscCall(PetscBTCreate(nv, &btvi));
751: for (i = 0; i < nv; i++) {
752: if (vcount[i] > 1) PetscCall(PetscBTSet(btvi, i));
753: }
755: if (matis->allow_repeated) {
756: PetscCall(PetscFree(eneighs[0]));
757: PetscCall(PetscFree(vneighs[0]));
758: PetscCall(PetscFree(eneighs));
759: PetscCall(PetscFree(vneighs));
760: }
761: PetscCall(ISLocalToGlobalMappingRestoreNodeInfo(el2g, NULL, &ecount, &eneighs));
762: PetscCall(ISLocalToGlobalMappingRestoreNodeInfo(vl2g, NULL, &vcount, &vneighs));
764: /* communicate corners and splitpoints */
765: PetscCall(PetscMalloc1(nv, &vmarks));
766: PetscCall(PetscArrayzero(sfvleaves, nv));
767: PetscCall(PetscArrayzero(sfvroots, Lv));
768: for (i = 0; i < nv; i++)
769: if (PetscUnlikely(PetscBTLookup(btv, i))) sfvleaves[i] = 1;
771: if (print) {
772: IS tbz;
774: cum = 0;
775: for (i = 0; i < nv; i++)
776: if (sfvleaves[i]) vmarks[cum++] = i;
778: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, cum, vmarks, PETSC_COPY_VALUES, &tbz));
779: PetscCall(PetscObjectSetName((PetscObject)tbz, "corners_to_be_zeroed_local"));
780: PetscCall(ISView(tbz, NULL));
781: PetscCall(ISDestroy(&tbz));
782: }
784: PetscCall(PetscSFReduceBegin(sfv, MPIU_INT, sfvleaves, sfvroots, MPI_SUM));
785: PetscCall(PetscSFReduceEnd(sfv, MPIU_INT, sfvleaves, sfvroots, MPI_SUM));
786: PetscCall(PetscSFBcastBegin(sfv, MPIU_INT, sfvroots, sfvleaves, MPI_REPLACE));
787: PetscCall(PetscSFBcastEnd(sfv, MPIU_INT, sfvroots, sfvleaves, MPI_REPLACE));
789: /* Zero rows of lGt corresponding to identified corners
790: and interior nodal dofs */
791: cum = 0;
792: for (i = 0; i < nv; i++) {
793: if (sfvleaves[i]) {
794: vmarks[cum++] = i;
795: PetscCall(PetscBTSet(btv, i));
796: } else if (!PetscBTLookup(btvi, i)) vmarks[cum++] = i;
797: }
798: PetscCall(PetscBTDestroy(&btvi));
799: if (print) {
800: IS tbz;
802: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, cum, vmarks, PETSC_COPY_VALUES, &tbz));
803: PetscCall(PetscObjectSetName((PetscObject)tbz, "corners_to_be_zeroed_with_interior"));
804: PetscCall(ISView(tbz, NULL));
805: PetscCall(ISDestroy(&tbz));
806: }
807: PetscCall(MatZeroRows(lGt, cum, vmarks, 0., NULL, NULL));
808: PetscCall(PetscFree(vmarks));
809: PetscCall(PetscSFDestroy(&sfv));
810: PetscCall(PetscFree2(sfvleaves, sfvroots));
812: /* Recompute G */
813: PetscCall(MatDestroy(&lG));
814: PetscCall(MatTranspose(lGt, MAT_INITIAL_MATRIX, &lG));
815: if (print) {
816: PetscCall(PetscObjectSetName((PetscObject)lG, "used_lG"));
817: PetscCall(MatView(lG, NULL));
818: PetscCall(PetscObjectSetName((PetscObject)lGt, "used_lGt"));
819: PetscCall(MatView(lGt, NULL));
820: }
822: /* Get primal dofs (if any) */
823: cum = 0;
824: for (i = 0; i < ne; i++) {
825: if (PetscUnlikely(PetscBTLookup(bte, i))) marks[cum++] = i;
826: }
827: if (fl2g) PetscCall(ISLocalToGlobalMappingApply(fl2g, cum, marks, marks));
828: PetscCall(ISCreateGeneral(comm, cum, marks, PETSC_COPY_VALUES, &primals));
829: if (print) {
830: PetscCall(PetscObjectSetName((PetscObject)primals, "prescribed_primal_dofs"));
831: PetscCall(ISView(primals, NULL));
832: }
833: PetscCall(PetscBTDestroy(&bte));
834: /* TODO: what if the user passed in some of them ? */
835: PetscCall(PCBDDCSetPrimalVerticesLocalIS(pc, primals));
836: PetscCall(ISDestroy(&primals));
838: /* Compute edge connectivity */
839: PetscCall(PetscObjectSetOptionsPrefix((PetscObject)lG, "econn_"));
841: /* Symbolic conn = lG*lGt */
842: if (!elements_corners) { /* if present, we assume we are in the element-by-element case and the CSR graph is not needed */
843: PetscCall(MatProductCreate(lG, lGt, NULL, &conn));
844: PetscCall(MatProductSetType(conn, MATPRODUCT_AB));
845: PetscCall(MatProductSetAlgorithm(conn, "default"));
846: PetscCall(MatProductSetFill(conn, PETSC_DEFAULT));
847: PetscCall(PetscObjectSetOptionsPrefix((PetscObject)conn, "econn_"));
848: PetscCall(MatProductSetFromOptions(conn));
849: PetscCall(MatProductSymbolic(conn));
850: PetscCall(MatGetRowIJ(conn, 0, PETSC_FALSE, PETSC_FALSE, &i, &ii, &jj, &done));
851: if (fl2g) {
852: PetscBT btf;
853: PetscInt *iia, *jja, *iiu, *jju;
854: PetscBool rest = PETSC_FALSE, free = PETSC_FALSE;
856: /* create CSR for all local dofs */
857: PetscCall(PetscMalloc1(n + 1, &iia));
858: if (pcbddc->mat_graph->nvtxs_csr) { /* the user has passed in a CSR graph */
859: PetscCheck(pcbddc->mat_graph->nvtxs_csr == n, PETSC_COMM_SELF, PETSC_ERR_USER, "Invalid size of CSR graph %" PetscInt_FMT ". Should be %" PetscInt_FMT, pcbddc->mat_graph->nvtxs_csr, n);
860: iiu = pcbddc->mat_graph->xadj;
861: jju = pcbddc->mat_graph->adjncy;
862: } else if (pcbddc->use_local_adj) {
863: rest = PETSC_TRUE;
864: PetscCall(MatGetRowIJ(matis->A, 0, PETSC_TRUE, PETSC_FALSE, &i, (const PetscInt **)&iiu, (const PetscInt **)&jju, &done));
865: } else {
866: free = PETSC_TRUE;
867: PetscCall(PetscMalloc2(n + 1, &iiu, n, &jju));
868: iiu[0] = 0;
869: for (i = 0; i < n; i++) {
870: iiu[i + 1] = i + 1;
871: jju[i] = -1;
872: }
873: }
875: /* import sizes of CSR */
876: iia[0] = 0;
877: for (i = 0; i < n; i++) iia[i + 1] = iiu[i + 1] - iiu[i];
879: /* overwrite entries corresponding to the Nedelec field */
880: PetscCall(PetscBTCreate(n, &btf));
881: PetscCall(ISGetIndices(nedfieldlocal, &idxs));
882: for (i = 0; i < ne; i++) {
883: PetscCall(PetscBTSet(btf, idxs[i]));
884: iia[idxs[i] + 1] = ii[i + 1] - ii[i];
885: }
887: /* iia in CSR */
888: for (i = 0; i < n; i++) iia[i + 1] += iia[i];
890: /* jja in CSR */
891: PetscCall(PetscMalloc1(iia[n], &jja));
892: for (i = 0; i < n; i++)
893: if (!PetscBTLookup(btf, i))
894: for (j = 0; j < iiu[i + 1] - iiu[i]; j++) jja[iia[i] + j] = jju[iiu[i] + j];
896: /* map edge dofs connectivity */
897: if (jj) {
898: PetscCall(ISLocalToGlobalMappingApply(fl2g, ii[ne], jj, (PetscInt *)jj));
899: for (i = 0; i < ne; i++) {
900: PetscInt e = idxs[i];
901: for (j = 0; j < ii[i + 1] - ii[i]; j++) jja[iia[e] + j] = jj[ii[i] + j];
902: }
903: }
904: PetscCall(ISRestoreIndices(nedfieldlocal, &idxs));
905: PetscCall(PCBDDCSetLocalAdjacencyGraph(pc, n, iia, jja, PETSC_COPY_VALUES));
906: if (rest) PetscCall(MatRestoreRowIJ(matis->A, 0, PETSC_TRUE, PETSC_FALSE, &i, (const PetscInt **)&iiu, (const PetscInt **)&jju, &done));
907: if (free) PetscCall(PetscFree2(iiu, jju));
908: PetscCall(PetscBTDestroy(&btf));
909: } else {
910: PetscCall(PCBDDCSetLocalAdjacencyGraph(pc, n, ii, jj, PETSC_COPY_VALUES));
911: }
912: PetscCall(MatRestoreRowIJ(conn, 0, PETSC_FALSE, PETSC_FALSE, &i, &ii, &jj, &done));
913: PetscCall(MatDestroy(&conn));
914: }
916: /* Analyze interface for edge dofs */
917: PetscCall(PCBDDCAnalyzeInterface(pc));
918: pcbddc->mat_graph->twodim = PETSC_FALSE;
920: /* Get coarse edges in the edge space */
921: PetscCall(PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph, NULL, NULL, &nee, &alleedges, &allprimals));
923: if (fl2g) {
924: PetscCall(ISGlobalToLocalMappingApplyIS(fl2g, IS_GTOLM_DROP, allprimals, &primals));
925: PetscCall(PetscMalloc1(nee, &eedges));
926: for (i = 0; i < nee; i++) PetscCall(ISGlobalToLocalMappingApplyIS(fl2g, IS_GTOLM_DROP, alleedges[i], &eedges[i]));
927: } else {
928: eedges = alleedges;
929: primals = allprimals;
930: }
932: /* Mark fine edge dofs with their coarse edge id */
933: PetscCall(PetscArrayzero(marks, ne));
934: PetscCall(ISGetLocalSize(primals, &cum));
935: PetscCall(ISGetIndices(primals, &idxs));
936: for (i = 0; i < cum; i++) marks[idxs[i]] = nee + 1;
937: PetscCall(ISRestoreIndices(primals, &idxs));
938: if (print) {
939: PetscCall(PetscObjectSetName((PetscObject)primals, "obtained_primal_dofs"));
940: PetscCall(ISView(primals, NULL));
941: }
943: maxsize = 0;
944: for (i = 0; i < nee; i++) {
945: PetscInt size, mark = i + 1;
947: PetscCall(ISGetLocalSize(eedges[i], &size));
948: PetscCall(ISGetIndices(eedges[i], &idxs));
949: for (j = 0; j < size; j++) marks[idxs[j]] = mark;
950: PetscCall(ISRestoreIndices(eedges[i], &idxs));
951: maxsize = PetscMax(maxsize, size);
952: }
954: /* Find coarse edge endpoints */
955: PetscCall(MatGetRowIJ(lG, 0, PETSC_FALSE, PETSC_FALSE, &i, &ii, &jj, &done));
956: PetscCall(MatGetRowIJ(lGt, 0, PETSC_FALSE, PETSC_FALSE, &i, &iit, &jjt, &done));
957: for (i = 0; i < nee; i++) {
958: PetscInt mark = i + 1, size;
960: PetscCall(ISGetLocalSize(eedges[i], &size));
961: if (!size && nedfieldlocal) continue;
962: PetscCheck(size, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Unexpected zero sized edge %" PetscInt_FMT, i);
963: PetscCall(ISGetIndices(eedges[i], &idxs));
964: if (print) {
965: PetscCall(PetscPrintf(PETSC_COMM_SELF, "ENDPOINTS ANALYSIS EDGE %" PetscInt_FMT "\n", i));
966: PetscCall(ISView(eedges[i], NULL));
967: }
968: for (j = 0; j < size; j++) {
969: PetscInt k, ee = idxs[j];
970: if (print) PetscCall(PetscPrintf(PETSC_COMM_SELF, " idx %" PetscInt_FMT "\n", ee));
971: for (k = ii[ee]; k < ii[ee + 1]; k++) {
972: if (print) PetscCall(PetscPrintf(PETSC_COMM_SELF, " inspect %" PetscInt_FMT "\n", jj[k]));
973: if (PetscBTLookup(btv, jj[k])) {
974: if (print) PetscCall(PetscPrintf(PETSC_COMM_SELF, " corner found (already set) %" PetscInt_FMT "\n", jj[k]));
975: } else if (PetscBTLookup(btvcand, jj[k])) { /* is it ok? */
976: PetscInt k2;
977: PetscBool corner = PETSC_FALSE;
978: for (k2 = iit[jj[k]]; k2 < iit[jj[k] + 1]; k2++) {
979: if (print) PetscCall(PetscPrintf(PETSC_COMM_SELF, " INSPECTING %" PetscInt_FMT ": mark %" PetscInt_FMT " (ref mark %" PetscInt_FMT "), boundary %d\n", jjt[k2], marks[jjt[k2]], mark, (int)!!PetscBTLookup(btb, jjt[k2])));
980: /* it's a corner if either is connected with an edge dof belonging to a different cc or
981: if the edge dof lie on the natural part of the boundary */
982: if ((marks[jjt[k2]] && marks[jjt[k2]] != mark) || (!marks[jjt[k2]] && PetscBTLookup(btb, jjt[k2]))) {
983: corner = PETSC_TRUE;
984: break;
985: }
986: }
987: if (corner) { /* found the nodal dof corresponding to the endpoint of the edge */
988: if (print) PetscCall(PetscPrintf(PETSC_COMM_SELF, " corner found %" PetscInt_FMT "\n", jj[k]));
989: PetscCall(PetscBTSet(btv, jj[k]));
990: } else {
991: if (print) PetscCall(PetscPrintf(PETSC_COMM_SELF, " no corners found\n"));
992: }
993: }
994: }
995: }
996: PetscCall(ISRestoreIndices(eedges[i], &idxs));
997: }
998: PetscCall(MatRestoreRowIJ(lGt, 0, PETSC_FALSE, PETSC_FALSE, &i, &iit, &jjt, &done));
999: PetscCall(MatRestoreRowIJ(lG, 0, PETSC_FALSE, PETSC_FALSE, &i, &ii, &jj, &done));
1000: PetscCall(PetscBTDestroy(&btb));
1002: /* Reset marked primal dofs */
1003: PetscCall(ISGetLocalSize(primals, &cum));
1004: PetscCall(ISGetIndices(primals, &idxs));
1005: for (i = 0; i < cum; i++) marks[idxs[i]] = 0;
1006: PetscCall(ISRestoreIndices(primals, &idxs));
1008: /* Now use the initial lG */
1009: PetscCall(MatDestroy(&lG));
1010: PetscCall(MatDestroy(&lGt));
1011: lG = lGinit;
1012: PetscCall(MatTranspose(lG, MAT_INITIAL_MATRIX, &lGt));
1014: /* Compute extended cols indices */
1015: PetscCall(PetscBTCreate(nv, &btvc));
1016: PetscCall(PetscBTCreate(nee, &bter));
1017: PetscCall(MatGetRowIJ(lG, 0, PETSC_FALSE, PETSC_FALSE, &i, &ii, &jj, &done));
1018: PetscCall(MatSeqAIJGetMaxRowNonzeros(lG, &i));
1019: i *= maxsize;
1020: PetscCall(PetscCalloc1(nee, &extcols));
1021: PetscCall(PetscMalloc2(i, &extrow, i, &gidxs));
1022: eerr = PETSC_FALSE;
1023: for (i = 0; i < nee; i++) {
1024: PetscInt size, found = 0;
1026: cum = 0;
1027: PetscCall(ISGetLocalSize(eedges[i], &size));
1028: if (!size && nedfieldlocal) continue;
1029: PetscCheck(size, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Unexpected zero sized edge %" PetscInt_FMT, i);
1030: PetscCall(ISGetIndices(eedges[i], &idxs));
1031: PetscCall(PetscBTMemzero(nv, btvc));
1032: for (j = 0; j < size; j++) {
1033: PetscInt k, ee = idxs[j];
1034: for (k = ii[ee]; k < ii[ee + 1]; k++) {
1035: PetscInt vv = jj[k];
1036: if (!PetscBTLookup(btv, vv)) extrow[cum++] = vv;
1037: else if (!PetscBTLookupSet(btvc, vv)) found++;
1038: }
1039: }
1040: PetscCall(ISRestoreIndices(eedges[i], &idxs));
1041: PetscCall(PetscSortRemoveDupsInt(&cum, extrow));
1042: PetscCall(ISLocalToGlobalMappingApply(vl2g, cum, extrow, gidxs));
1043: PetscCall(PetscSortIntWithArray(cum, gidxs, extrow));
1044: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, cum, extrow, PETSC_COPY_VALUES, &extcols[i]));
1045: /* it may happen that endpoints are not defined at this point
1046: if it is the case, mark this edge for a second pass */
1047: if (cum != size - 1 || found != 2) {
1048: PetscCall(PetscBTSet(bter, i));
1049: if (print) {
1050: PetscCall(PetscObjectSetName((PetscObject)eedges[i], "error_edge"));
1051: PetscCall(ISView(eedges[i], NULL));
1052: PetscCall(PetscObjectSetName((PetscObject)extcols[i], "error_extcol"));
1053: PetscCall(ISView(extcols[i], NULL));
1054: }
1055: eerr = PETSC_TRUE;
1056: }
1057: }
1058: /* PetscCheck(!eerr,PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected SIZE OF EDGE > EXTCOL FIRST PASS"); */
1059: PetscCallMPI(MPIU_Allreduce(&eerr, &done, 1, MPI_C_BOOL, MPI_LOR, comm));
1060: if (done) {
1061: PetscInt *newprimals;
1063: PetscCall(PetscMalloc1(ne, &newprimals));
1064: PetscCall(ISGetLocalSize(primals, &cum));
1065: PetscCall(ISGetIndices(primals, &idxs));
1066: PetscCall(PetscArraycpy(newprimals, idxs, cum));
1067: PetscCall(ISRestoreIndices(primals, &idxs));
1068: PetscCall(MatGetRowIJ(lGt, 0, PETSC_FALSE, PETSC_FALSE, &i, &iit, &jjt, &done));
1069: if (print) PetscCall(PetscPrintf(PETSC_COMM_SELF, "DOING SECOND PASS (eerr %s)\n", PetscBools[eerr]));
1070: for (i = 0; i < nee; i++) {
1071: PetscBool has_candidates = PETSC_FALSE;
1072: if (PetscBTLookup(bter, i)) {
1073: PetscInt size, mark = i + 1;
1075: PetscCall(ISGetLocalSize(eedges[i], &size));
1076: PetscCall(ISGetIndices(eedges[i], &idxs));
1077: /* for (j=0;j<size;j++) newprimals[cum++] = idxs[j]; */
1078: for (j = 0; j < size; j++) {
1079: PetscInt k, ee = idxs[j];
1080: if (print) PetscCall(PetscPrintf(PETSC_COMM_SELF, "Inspecting edge dof %" PetscInt_FMT " [%" PetscInt_FMT " %" PetscInt_FMT ")\n", ee, ii[ee], ii[ee + 1]));
1081: for (k = ii[ee]; k < ii[ee + 1]; k++) {
1082: /* set all candidates located on the edge as corners */
1083: if (PetscBTLookup(btvcand, jj[k])) {
1084: PetscInt k2, vv = jj[k];
1085: has_candidates = PETSC_TRUE;
1086: if (print) PetscCall(PetscPrintf(PETSC_COMM_SELF, " Candidate set to vertex %" PetscInt_FMT "\n", vv));
1087: PetscCall(PetscBTSet(btv, vv));
1088: /* set all edge dofs connected to candidate as primals */
1089: for (k2 = iit[vv]; k2 < iit[vv + 1]; k2++) {
1090: if (marks[jjt[k2]] == mark) {
1091: PetscInt k3, ee2 = jjt[k2];
1092: if (print) PetscCall(PetscPrintf(PETSC_COMM_SELF, " Connected edge dof set to primal %" PetscInt_FMT "\n", ee2));
1093: newprimals[cum++] = ee2;
1094: /* finally set the new corners */
1095: for (k3 = ii[ee2]; k3 < ii[ee2 + 1]; k3++) {
1096: if (print) PetscCall(PetscPrintf(PETSC_COMM_SELF, " Connected nodal dof set to vertex %" PetscInt_FMT "\n", jj[k3]));
1097: PetscCall(PetscBTSet(btv, jj[k3]));
1098: }
1099: }
1100: }
1101: } else {
1102: if (print) PetscCall(PetscPrintf(PETSC_COMM_SELF, " Not a candidate vertex %" PetscInt_FMT "\n", jj[k]));
1103: }
1104: }
1105: }
1106: if (!has_candidates) { /* circular edge */
1107: PetscInt k, ee = idxs[0], *tmarks;
1109: PetscCall(PetscCalloc1(ne, &tmarks));
1110: if (print) PetscCall(PetscPrintf(PETSC_COMM_SELF, " Circular edge %" PetscInt_FMT "\n", i));
1111: for (k = ii[ee]; k < ii[ee + 1]; k++) {
1112: PetscInt k2;
1113: if (print) PetscCall(PetscPrintf(PETSC_COMM_SELF, " Set to corner %" PetscInt_FMT "\n", jj[k]));
1114: PetscCall(PetscBTSet(btv, jj[k]));
1115: for (k2 = iit[jj[k]]; k2 < iit[jj[k] + 1]; k2++) tmarks[jjt[k2]]++;
1116: }
1117: for (j = 0; j < size; j++) {
1118: if (tmarks[idxs[j]] > 1) {
1119: if (print) PetscCall(PetscPrintf(PETSC_COMM_SELF, " Edge dof set to primal %" PetscInt_FMT "\n", idxs[j]));
1120: newprimals[cum++] = idxs[j];
1121: }
1122: }
1123: PetscCall(PetscFree(tmarks));
1124: }
1125: PetscCall(ISRestoreIndices(eedges[i], &idxs));
1126: }
1127: PetscCall(ISDestroy(&extcols[i]));
1128: }
1129: PetscCall(PetscFree(extcols));
1130: PetscCall(MatRestoreRowIJ(lGt, 0, PETSC_FALSE, PETSC_FALSE, &i, &iit, &jjt, &done));
1131: PetscCall(PetscSortRemoveDupsInt(&cum, newprimals));
1132: if (fl2g) {
1133: PetscCall(ISLocalToGlobalMappingApply(fl2g, cum, newprimals, newprimals));
1134: PetscCall(ISDestroy(&primals));
1135: for (i = 0; i < nee; i++) PetscCall(ISDestroy(&eedges[i]));
1136: PetscCall(PetscFree(eedges));
1137: }
1138: PetscCall(PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph, NULL, NULL, &nee, &alleedges, &allprimals));
1139: PetscCall(ISCreateGeneral(comm, cum, newprimals, PETSC_COPY_VALUES, &primals));
1140: PetscCall(PetscFree(newprimals));
1141: PetscCall(PCBDDCSetPrimalVerticesLocalIS(pc, primals));
1142: PetscCall(ISDestroy(&primals));
1143: PetscCall(PCBDDCAnalyzeInterface(pc));
1144: pcbddc->mat_graph->twodim = PETSC_FALSE;
1145: PetscCall(PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph, NULL, NULL, &nee, &alleedges, &allprimals));
1146: if (fl2g) {
1147: PetscCall(ISGlobalToLocalMappingApplyIS(fl2g, IS_GTOLM_DROP, allprimals, &primals));
1148: PetscCall(PetscMalloc1(nee, &eedges));
1149: for (i = 0; i < nee; i++) PetscCall(ISGlobalToLocalMappingApplyIS(fl2g, IS_GTOLM_DROP, alleedges[i], &eedges[i]));
1150: } else {
1151: eedges = alleedges;
1152: primals = allprimals;
1153: }
1154: PetscCall(PetscCalloc1(nee, &extcols));
1156: /* Mark again */
1157: PetscCall(PetscArrayzero(marks, ne));
1158: for (i = 0; i < nee; i++) {
1159: PetscInt size, mark = i + 1;
1161: PetscCall(ISGetLocalSize(eedges[i], &size));
1162: PetscCall(ISGetIndices(eedges[i], &idxs));
1163: for (j = 0; j < size; j++) marks[idxs[j]] = mark;
1164: PetscCall(ISRestoreIndices(eedges[i], &idxs));
1165: }
1166: if (print) {
1167: PetscCall(PetscObjectSetName((PetscObject)primals, "obtained_primal_dofs_secondpass"));
1168: PetscCall(ISView(primals, NULL));
1169: }
1171: /* Recompute extended cols */
1172: eerr = PETSC_FALSE;
1173: for (i = 0; i < nee; i++) {
1174: PetscInt size;
1176: cum = 0;
1177: PetscCall(ISGetLocalSize(eedges[i], &size));
1178: if (!size && nedfieldlocal) continue;
1179: PetscCheck(size, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Unexpected zero sized edge %" PetscInt_FMT, i);
1180: PetscCall(ISGetIndices(eedges[i], &idxs));
1181: for (j = 0; j < size; j++) {
1182: PetscInt k, ee = idxs[j];
1183: for (k = ii[ee]; k < ii[ee + 1]; k++)
1184: if (!PetscBTLookup(btv, jj[k])) extrow[cum++] = jj[k];
1185: }
1186: PetscCall(ISRestoreIndices(eedges[i], &idxs));
1187: PetscCall(PetscSortRemoveDupsInt(&cum, extrow));
1188: PetscCall(ISLocalToGlobalMappingApply(vl2g, cum, extrow, gidxs));
1189: PetscCall(PetscSortIntWithArray(cum, gidxs, extrow));
1190: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, cum, extrow, PETSC_COPY_VALUES, &extcols[i]));
1191: if (cum != size - 1) {
1192: if (print) {
1193: PetscCall(PetscObjectSetName((PetscObject)eedges[i], "error_edge_secondpass"));
1194: PetscCall(ISView(eedges[i], NULL));
1195: PetscCall(PetscObjectSetName((PetscObject)extcols[i], "error_extcol_secondpass"));
1196: PetscCall(ISView(extcols[i], NULL));
1197: }
1198: eerr = PETSC_TRUE;
1199: }
1200: }
1201: }
1202: PetscCall(MatRestoreRowIJ(lG, 0, PETSC_FALSE, PETSC_FALSE, &i, &ii, &jj, &done));
1203: PetscCall(PetscFree2(extrow, gidxs));
1204: PetscCall(PetscBTDestroy(&bter));
1205: if (print) PetscCall(PCBDDCGraphASCIIView(pcbddc->mat_graph, 5, PETSC_VIEWER_STDOUT_SELF));
1206: /* an error should not occur at this point */
1207: PetscCheck(!eerr, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Unexpected SIZE OF EDGE > EXTCOL SECOND PASS");
1209: /* Check the number of endpoints */
1210: PetscCall(MatGetRowIJ(lG, 0, PETSC_FALSE, PETSC_FALSE, &i, &ii, &jj, &done));
1211: PetscCall(PetscMalloc1(2 * nee, &corners));
1212: PetscCall(PetscMalloc1(nee, &cedges));
1213: for (i = 0; i < nee; i++) {
1214: PetscInt size, found = 0, gc[2];
1216: /* init with defaults */
1217: cedges[i] = corners[i * 2] = corners[i * 2 + 1] = -1;
1218: PetscCall(ISGetLocalSize(eedges[i], &size));
1219: if (!size && nedfieldlocal) continue;
1220: PetscCheck(size, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Unexpected zero sized edge %" PetscInt_FMT, i);
1221: PetscCall(ISGetIndices(eedges[i], &idxs));
1222: PetscCall(PetscBTMemzero(nv, btvc));
1223: for (j = 0; j < size; j++) {
1224: PetscInt k, ee = idxs[j];
1225: for (k = ii[ee]; k < ii[ee + 1]; k++) {
1226: PetscInt vv = jj[k];
1227: if (PetscBTLookup(btv, vv) && !PetscBTLookupSet(btvc, vv)) {
1228: PetscCheck(found != 2, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Found more than two corners for edge %" PetscInt_FMT, i);
1229: corners[i * 2 + found++] = vv;
1230: }
1231: }
1232: }
1233: if (found != 2) {
1234: PetscInt e;
1235: if (fl2g) PetscCall(ISLocalToGlobalMappingApply(fl2g, 1, idxs, &e));
1236: else e = idxs[0];
1237: SETERRQ(PETSC_COMM_SELF, PETSC_ERR_PLIB, "Found %" PetscInt_FMT " corners for edge %" PetscInt_FMT " (astart %" PetscInt_FMT ", estart %" PetscInt_FMT ")", found, i, e, idxs[0]);
1238: }
1240: /* get primal dof index on this coarse edge */
1241: PetscCall(ISLocalToGlobalMappingApply(vl2g, 2, corners + 2 * i, gc));
1242: if (gc[0] > gc[1]) {
1243: PetscInt swap = corners[2 * i];
1244: corners[2 * i] = corners[2 * i + 1];
1245: corners[2 * i + 1] = swap;
1246: }
1247: cedges[i] = idxs[size - 1];
1248: PetscCall(ISRestoreIndices(eedges[i], &idxs));
1249: if (print) PetscCall(PetscPrintf(PETSC_COMM_SELF, "EDGE %" PetscInt_FMT ": ce %" PetscInt_FMT ", corners (%" PetscInt_FMT ",%" PetscInt_FMT ")\n", i, cedges[i], corners[2 * i], corners[2 * i + 1]));
1250: }
1251: PetscCall(MatRestoreRowIJ(lG, 0, PETSC_FALSE, PETSC_FALSE, &i, &ii, &jj, &done));
1252: PetscCall(PetscBTDestroy(&btvc));
1254: if (PetscDefined(USE_DEBUG)) {
1255: /* Inspects columns of lG (rows of lGt) and make sure the change of basis will
1256: not interfere with neighbouring coarse edges */
1257: PetscCall(PetscMalloc1(nee + 1, &emarks));
1258: PetscCall(MatGetRowIJ(lGt, 0, PETSC_FALSE, PETSC_FALSE, &i, &ii, &jj, &done));
1259: for (i = 0; i < nv; i++) {
1260: PetscInt emax = 0, eemax = 0;
1262: if (ii[i + 1] == ii[i] || PetscBTLookup(btv, i)) continue;
1263: PetscCall(PetscArrayzero(emarks, nee + 1));
1264: for (j = ii[i]; j < ii[i + 1]; j++) emarks[marks[jj[j]]]++;
1265: for (j = 1; j < nee + 1; j++) {
1266: if (emax < emarks[j]) {
1267: emax = emarks[j];
1268: eemax = j;
1269: }
1270: }
1271: /* not relevant for edges */
1272: if (!eemax) continue;
1274: for (j = ii[i]; j < ii[i + 1]; j++) {
1275: PetscCheck(!marks[jj[j]] || marks[jj[j]] == eemax, PETSC_COMM_SELF, PETSC_ERR_SUP, "Found 2 coarse edges (id %" PetscInt_FMT " and %" PetscInt_FMT ") connected through the %" PetscInt_FMT " nodal dof at edge dof %" PetscInt_FMT, marks[jj[j]] - 1, eemax, i, jj[j]);
1276: }
1277: }
1278: PetscCall(PetscFree(emarks));
1279: PetscCall(MatRestoreRowIJ(lGt, 0, PETSC_FALSE, PETSC_FALSE, &i, &ii, &jj, &done));
1280: }
1282: /* Compute extended rows indices for edge blocks of the change of basis */
1283: PetscCall(MatGetRowIJ(lGt, 0, PETSC_FALSE, PETSC_FALSE, &i, &ii, &jj, &done));
1284: PetscCall(MatSeqAIJGetMaxRowNonzeros(lGt, &extmem));
1285: extmem *= maxsize;
1286: PetscCall(PetscMalloc1(extmem * nee, &extrow));
1287: PetscCall(PetscMalloc1(nee, &extrows));
1288: PetscCall(PetscCalloc1(nee, &extrowcum));
1289: for (i = 0; i < nv; i++) {
1290: PetscInt mark = 0, size, start;
1292: if (ii[i + 1] == ii[i] || PetscBTLookup(btv, i)) continue;
1293: for (j = ii[i]; j < ii[i + 1]; j++)
1294: if (marks[jj[j]] && !mark) mark = marks[jj[j]];
1296: /* not relevant */
1297: if (!mark) continue;
1299: /* import extended row */
1300: mark--;
1301: start = mark * extmem + extrowcum[mark];
1302: size = ii[i + 1] - ii[i];
1303: PetscCheck(extrowcum[mark] + size <= extmem, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Not enough memory allocated %" PetscInt_FMT " > %" PetscInt_FMT, extrowcum[mark] + size, extmem);
1304: PetscCall(PetscArraycpy(extrow + start, jj + ii[i], size));
1305: extrowcum[mark] += size;
1306: }
1307: PetscCall(MatRestoreRowIJ(lGt, 0, PETSC_FALSE, PETSC_FALSE, &i, &ii, &jj, &done));
1308: PetscCall(MatDestroy(&lGt));
1309: PetscCall(PetscFree(marks));
1311: /* Compress extrows */
1312: cum = 0;
1313: for (i = 0; i < nee; i++) {
1314: PetscInt size = extrowcum[i], *start = extrow + i * extmem;
1315: PetscCall(PetscSortRemoveDupsInt(&size, start));
1316: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, size, start, PETSC_USE_POINTER, &extrows[i]));
1317: cum = PetscMax(cum, size);
1318: }
1319: PetscCall(PetscFree(extrowcum));
1320: PetscCall(PetscBTDestroy(&btv));
1321: PetscCall(PetscBTDestroy(&btvcand));
1323: /* Workspace for lapack inner calls and VecSetValues */
1324: PetscCall(PetscMalloc2((5 + cum + maxsize) * maxsize, &work, maxsize, &rwork));
1326: /* Create change of basis matrix (no preallocation) */
1327: PetscCall(MatCreate(comm, &T));
1328: PetscCall(MatSetLayouts(T, pc->mat->rmap, pc->mat->cmap));
1329: PetscCall(MatSetType(T, MATAIJ));
1330: PetscCall(MatSetLocalToGlobalMapping(T, al2g, al2g));
1331: PetscCall(MatSetOption(T, MAT_ROW_ORIENTED, PETSC_FALSE));
1332: PetscCall(MatSetOption(T, MAT_IGNORE_ZERO_ENTRIES, PETSC_TRUE));
1333: //PetscCall(MatSeqAIJSetPreallocation(T, maxsize, NULL));
1334: //PetscCall(MatMPIAIJSetPreallocation(T, maxsize, NULL, maxsize, NULL));
1335: //PetscCall(MatSetOption(T, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_FALSE));
1337: /* Defaults to identity */
1338: {
1339: Vec w;
1340: const PetscScalar *wa;
1342: PetscCall(MatCreateVecs(T, &w, NULL));
1343: PetscCall(VecSetLocalToGlobalMapping(w, al2g));
1344: PetscCall(VecSet(w, 1.0));
1345: for (i = 0; i < nee; i++) {
1346: const PetscInt *idxs;
1347: PetscInt nl;
1349: PetscCall(ISGetLocalSize(eedges[i], &nl));
1350: PetscCall(ISGetIndices(eedges[i], &idxs));
1351: PetscCall(VecSetValuesLocal(w, nl, idxs, NULL, INSERT_VALUES));
1352: PetscCall(ISRestoreIndices(eedges[i], &idxs));
1353: }
1354: PetscCall(VecAssemblyBegin(w));
1355: PetscCall(VecAssemblyEnd(w));
1356: PetscCall(VecGetArrayRead(w, &wa));
1357: for (i = T->rmap->rstart; i < T->rmap->rend; i++)
1358: if (PetscAbsScalar(wa[i - T->rmap->rstart])) PetscCall(MatSetValue(T, i, i, 1.0, INSERT_VALUES));
1359: PetscCall(VecRestoreArrayRead(w, &wa));
1360: PetscCall(VecDestroy(&w));
1361: }
1363: /* Create discrete gradient for the coarser level if needed */
1364: PetscCall(MatDestroy(&pcbddc->nedcG));
1365: PetscCall(ISDestroy(&pcbddc->nedclocal));
1366: if (pcbddc->current_level < pcbddc->max_levels) {
1367: ISLocalToGlobalMapping cel2g, cvl2g;
1368: IS wis, gwis;
1369: PetscInt cnv, cne;
1371: PetscCall(ISCreateGeneral(comm, nee, cedges, PETSC_COPY_VALUES, &wis));
1372: if (fl2g) PetscCall(ISLocalToGlobalMappingApplyIS(fl2g, wis, &pcbddc->nedclocal));
1373: else {
1374: PetscCall(PetscObjectReference((PetscObject)wis));
1375: pcbddc->nedclocal = wis;
1376: }
1377: PetscCall(ISLocalToGlobalMappingApplyIS(el2g, wis, &gwis));
1378: PetscCall(ISDestroy(&wis));
1379: PetscCall(ISRenumber(gwis, NULL, &cne, &wis));
1380: PetscCall(ISLocalToGlobalMappingCreateIS(wis, &cel2g));
1381: PetscCall(ISDestroy(&wis));
1382: PetscCall(ISDestroy(&gwis));
1384: PetscCall(ISCreateGeneral(comm, 2 * nee, corners, PETSC_USE_POINTER, &wis));
1385: PetscCall(ISLocalToGlobalMappingApplyIS(vl2g, wis, &gwis));
1386: PetscCall(ISDestroy(&wis));
1387: PetscCall(ISRenumber(gwis, NULL, &cnv, &wis));
1388: PetscCall(ISLocalToGlobalMappingCreateIS(wis, &cvl2g));
1389: PetscCall(ISDestroy(&wis));
1390: PetscCall(ISDestroy(&gwis));
1392: PetscCall(MatCreate(comm, &pcbddc->nedcG));
1393: PetscCall(MatSetSizes(pcbddc->nedcG, PETSC_DECIDE, PETSC_DECIDE, cne, cnv));
1394: PetscCall(MatSetType(pcbddc->nedcG, MATAIJ));
1395: PetscCall(MatSeqAIJSetPreallocation(pcbddc->nedcG, 2, NULL));
1396: PetscCall(MatMPIAIJSetPreallocation(pcbddc->nedcG, 2, NULL, 2, NULL));
1397: PetscCall(MatSetLocalToGlobalMapping(pcbddc->nedcG, cel2g, cvl2g));
1398: PetscCall(ISLocalToGlobalMappingDestroy(&cel2g));
1399: PetscCall(ISLocalToGlobalMappingDestroy(&cvl2g));
1400: }
1402: MatNullSpace nnsp;
1403: PetscBool nnsp_has_const = PETSC_FALSE;
1404: const Vec *nnsp_vecs = NULL;
1405: PetscInt nnsp_nvecs = 0;
1406: VecScatter nnsp_vscat = NULL;
1407: PetscCall(MatGetNullSpace(pcbddc->discretegradient, &nnsp));
1408: if (nnsp) PetscCall(MatNullSpaceGetVecs(nnsp, &nnsp_has_const, &nnsp_nvecs, &nnsp_vecs));
1409: if (nnsp_has_const || nnsp_nvecs) { /* create scatter to import edge constraints */
1410: IS allextcols, gallextcols, galleedges, is_E_to_zero;
1411: Vec E, V;
1412: PetscInt *eedgesidxs;
1413: const PetscScalar *evals;
1415: PetscCall(MatCreateVecs(pc->pmat, &E, NULL));
1416: PetscCall(MatCreateVecs(pcbddc->discretegradient, &V, NULL));
1417: PetscCall(ISConcatenate(PETSC_COMM_SELF, nee, extcols, &allextcols));
1418: cum = 0;
1419: for (i = 0; i < nee; i++) {
1420: PetscInt j;
1422: PetscCall(ISGetLocalSize(eedges[i], &j));
1423: PetscCheck(j, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Zero sized edge %" PetscInt_FMT, i);
1424: cum += j - 1;
1425: }
1426: PetscCall(PetscMalloc1(PetscMax(cum, pc->pmat->rmap->n), &eedgesidxs));
1427: cum = 0;
1428: for (i = 0; i < nee; i++) {
1429: const PetscInt *idxs;
1430: PetscInt j;
1432: PetscCall(ISGetLocalSize(eedges[i], &j));
1433: PetscCall(ISGetIndices(eedges[i], &idxs));
1434: PetscCall(PetscArraycpy(eedgesidxs + cum, idxs, j - 1)); /* last on the edge is primal */
1435: PetscCall(ISRestoreIndices(eedges[i], &idxs));
1436: cum += j - 1;
1437: }
1438: PetscCall(ISLocalToGlobalMappingApply(al2g, cum, eedgesidxs, eedgesidxs));
1439: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, cum, eedgesidxs, PETSC_USE_POINTER, &galleedges));
1440: PetscCall(ISLocalToGlobalMappingApplyIS(vl2g, allextcols, &gallextcols));
1441: PetscCall(VecScatterCreate(V, gallextcols, E, galleedges, &nnsp_vscat));
1442: PetscCall(ISDestroy(&allextcols));
1443: PetscCall(ISDestroy(&gallextcols));
1444: PetscCall(ISDestroy(&galleedges));
1446: /* identify dofs we must zero if importing user-defined near nullspace from pmat */
1447: PetscCall(VecSet(E, 1.0));
1448: PetscCall(VecSetValues(E, cum, eedgesidxs, NULL, INSERT_VALUES));
1449: PetscCall(VecAssemblyBegin(E));
1450: PetscCall(VecAssemblyEnd(E));
1451: PetscCall(VecGetArrayRead(E, &evals));
1452: for (i = 0, cum = 0; i < pc->pmat->rmap->n; i++)
1453: if (evals[i] == 0.0) eedgesidxs[cum++] = i + pc->pmat->rmap->rstart;
1454: PetscCall(VecRestoreArrayRead(E, &evals));
1455: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, cum, eedgesidxs, PETSC_COPY_VALUES, &is_E_to_zero));
1456: PetscCall(PetscFree(eedgesidxs));
1458: PetscCall(PetscObjectCompose((PetscObject)nnsp_vscat, "__V_Vec", (PetscObject)V));
1459: PetscCall(PetscObjectCompose((PetscObject)nnsp_vscat, "__E_Vec", (PetscObject)E));
1460: PetscCall(PetscObjectCompose((PetscObject)nnsp_vscat, "__E_zero", (PetscObject)is_E_to_zero));
1461: PetscCall(ISDestroy(&is_E_to_zero));
1462: PetscCall(VecDestroy(&V));
1463: PetscCall(VecDestroy(&E));
1464: }
1465: #if defined(PRINT_GDET)
1466: inc = 0;
1467: lev = pcbddc->current_level;
1468: #endif
1470: /* Insert values in the change of basis matrix */
1471: for (i = 0; i < nee; i++) {
1472: Mat Gins = NULL, GKins = NULL;
1473: IS cornersis = NULL;
1474: PetscScalar cvals[2];
1476: if (pcbddc->nedcG) PetscCall(ISCreateGeneral(PETSC_COMM_SELF, 2, corners + 2 * i, PETSC_USE_POINTER, &cornersis));
1477: PetscCall(PCBDDCComputeNedelecChangeEdge(lG, eedges[i], extrows[i], extcols[i], cornersis, &Gins, &GKins, cvals, work, rwork));
1478: if (Gins && GKins) {
1479: const PetscScalar *data;
1480: const PetscInt *rows, *cols;
1481: PetscInt nrh, nch, nrc, ncc;
1483: PetscCall(ISGetIndices(eedges[i], &cols));
1484: /* H1 */
1485: PetscCall(ISGetIndices(extrows[i], &rows));
1486: PetscCall(MatGetSize(Gins, &nrh, &nch));
1487: PetscCall(MatDenseGetArrayRead(Gins, &data));
1488: PetscCall(MatSetValuesLocal(T, nrh, rows, nch, cols, data, INSERT_VALUES));
1489: PetscCall(MatDenseRestoreArrayRead(Gins, &data));
1490: PetscCall(ISRestoreIndices(extrows[i], &rows));
1491: /* complement */
1492: PetscCall(MatGetSize(GKins, &nrc, &ncc));
1493: PetscCheck(ncc, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Constant function has not been generated for coarse edge %" PetscInt_FMT, i);
1494: PetscCheck(ncc + nch == nrc, PETSC_COMM_SELF, PETSC_ERR_PLIB, "The sum of the number of columns of GKins %" PetscInt_FMT " and Gins %" PetscInt_FMT " does not match %" PetscInt_FMT " for coarse edge %" PetscInt_FMT, ncc, nch, nrc, i);
1495: PetscCheck(ncc == 1 || !pcbddc->nedcG, PETSC_COMM_SELF, PETSC_ERR_SUP, "Cannot generate the coarse discrete gradient for coarse edge %" PetscInt_FMT " with ncc %" PetscInt_FMT, i, ncc);
1496: PetscCall(MatDenseGetArrayRead(GKins, &data));
1497: PetscCall(MatSetValuesLocal(T, nrc, cols, ncc, cols + nch, data, INSERT_VALUES));
1498: PetscCall(MatDenseRestoreArrayRead(GKins, &data));
1500: /* coarse discrete gradient */
1501: if (pcbddc->nedcG) {
1502: PetscInt cols[2];
1504: cols[0] = 2 * i;
1505: cols[1] = 2 * i + 1;
1506: PetscCall(MatSetValuesLocal(pcbddc->nedcG, 1, &i, 2, cols, cvals, INSERT_VALUES));
1507: }
1508: PetscCall(ISRestoreIndices(eedges[i], &cols));
1509: }
1510: PetscCall(ISDestroy(&extrows[i]));
1511: PetscCall(ISDestroy(&extcols[i]));
1512: PetscCall(ISDestroy(&cornersis));
1513: PetscCall(MatDestroy(&Gins));
1514: PetscCall(MatDestroy(&GKins));
1515: }
1517: /* import edge constraints */
1518: if (nnsp_vscat) {
1519: Vec V, E, *quadvecs;
1520: PetscInt nvecs, nvecs_orth;
1521: MatNullSpace onnsp = NULL;
1522: PetscBool onnsp_has_const = PETSC_FALSE;
1523: const Vec *onnsp_vecs = NULL;
1524: PetscInt onnsp_nvecs = 0, new_nnsp_nvecs, old_nnsp_nvecs;
1525: IS is_E_to_zero;
1527: /* import nearnullspace from preconditioning matrix if user-defined */
1528: PetscCall(MatGetNearNullSpace(pc->pmat, &onnsp));
1529: if (onnsp) {
1530: PetscBool isinternal;
1532: PetscCall(PetscStrcmp("_internal_BDDC_nedelec_nnsp", ((PetscObject)onnsp)->name, &isinternal));
1533: if (!isinternal) PetscCall(MatNullSpaceGetVecs(onnsp, &onnsp_has_const, &onnsp_nvecs, &onnsp_vecs));
1534: }
1535: new_nnsp_nvecs = nnsp_nvecs + (nnsp_has_const ? 1 : 0);
1536: old_nnsp_nvecs = onnsp_nvecs + (onnsp_has_const ? 1 : 0);
1537: nvecs = old_nnsp_nvecs + new_nnsp_nvecs;
1538: PetscCall(PetscMalloc1(nvecs, &quadvecs));
1540: PetscCall(PetscObjectQuery((PetscObject)nnsp_vscat, "__V_Vec", (PetscObject *)&V));
1541: PetscCall(PetscObjectQuery((PetscObject)nnsp_vscat, "__E_Vec", (PetscObject *)&E));
1542: PetscCall(PetscObjectQuery((PetscObject)nnsp_vscat, "__E_zero", (PetscObject *)&is_E_to_zero));
1543: for (i = 0; i < nvecs; i++) PetscCall(VecDuplicate(E, &quadvecs[i]));
1544: cum = 0;
1545: if (nnsp_has_const) {
1546: PetscCall(VecSet(V, 1.0));
1547: PetscCall(VecScatterBegin(nnsp_vscat, V, quadvecs[0], INSERT_VALUES, SCATTER_FORWARD));
1548: PetscCall(VecScatterEnd(nnsp_vscat, V, quadvecs[0], INSERT_VALUES, SCATTER_FORWARD));
1549: cum = 1;
1550: }
1551: for (i = 0; i < nnsp_nvecs; i++) {
1552: PetscCall(VecScatterBegin(nnsp_vscat, nnsp_vecs[i], quadvecs[i + cum], INSERT_VALUES, SCATTER_FORWARD));
1553: PetscCall(VecScatterEnd(nnsp_vscat, nnsp_vecs[i], quadvecs[i + cum], INSERT_VALUES, SCATTER_FORWARD));
1554: }
1556: /* Now add old nnsp if present */
1557: cum = 0;
1558: if (onnsp_has_const) {
1559: PetscCall(VecSet(quadvecs[new_nnsp_nvecs], 1.0));
1560: PetscCall(VecISSet(quadvecs[new_nnsp_nvecs], is_E_to_zero, 0));
1561: cum = 1;
1562: }
1563: for (i = 0; i < onnsp_nvecs; i++) {
1564: PetscCall(VecCopy(onnsp_vecs[i], quadvecs[i + cum + new_nnsp_nvecs]));
1565: PetscCall(VecISSet(quadvecs[i + cum + new_nnsp_nvecs], is_E_to_zero, 0));
1566: }
1567: nvecs_orth = nvecs;
1568: PetscCall(PCBDDCOrthonormalizeVecs(&nvecs_orth, quadvecs));
1569: PetscCall(MatNullSpaceCreate(PetscObjectComm((PetscObject)pc), PETSC_FALSE, nvecs_orth, quadvecs, &nnsp));
1570: for (i = 0; i < nvecs; i++) PetscCall(VecDestroy(&quadvecs[i]));
1571: PetscCall(PetscFree(quadvecs));
1572: PetscCall(PetscObjectSetName((PetscObject)nnsp, "_internal_BDDC_nedelec_nnsp"));
1573: PetscCall(MatSetNearNullSpace(pc->pmat, nnsp));
1574: PetscCall(MatNullSpaceDestroy(&nnsp));
1575: }
1576: PetscCall(VecScatterDestroy(&nnsp_vscat));
1577: PetscCall(ISLocalToGlobalMappingDestroy(&vl2g));
1578: PetscCall(ISLocalToGlobalMappingDestroy(&el2g));
1579: PetscCall(ISLocalToGlobalMappingDestroy(&al2g));
1581: /* Start assembling */
1582: PetscCall(MatAssemblyBegin(T, MAT_FINAL_ASSEMBLY));
1583: if (pcbddc->nedcG) PetscCall(MatAssemblyBegin(pcbddc->nedcG, MAT_FINAL_ASSEMBLY));
1585: /* Free */
1586: if (fl2g) {
1587: PetscCall(ISDestroy(&primals));
1588: for (i = 0; i < nee; i++) PetscCall(ISDestroy(&eedges[i]));
1589: PetscCall(PetscFree(eedges));
1590: }
1592: /* hack mat_graph with primal dofs on the coarse edges */
1593: {
1594: PCBDDCGraph graph = pcbddc->mat_graph;
1595: PetscInt *oqueue = graph->queue;
1596: PetscInt *ocptr = graph->cptr;
1597: PetscInt ncc, *idxs;
1599: /* find first primal edge */
1600: if (pcbddc->nedclocal) PetscCall(ISGetIndices(pcbddc->nedclocal, (const PetscInt **)&idxs));
1601: else {
1602: if (fl2g) PetscCall(ISLocalToGlobalMappingApply(fl2g, nee, cedges, cedges));
1603: idxs = cedges;
1604: }
1605: cum = 0;
1606: while (cum < nee && cedges[cum] < 0) cum++;
1608: /* adapt connected components */
1609: PetscCall(PetscMalloc2(graph->nvtxs + 1, &graph->cptr, ocptr[graph->ncc], &graph->queue));
1610: graph->cptr[0] = 0;
1611: for (i = 0, ncc = 0; i < graph->ncc; i++) {
1612: PetscInt lc = ocptr[i + 1] - ocptr[i];
1613: if (cum != nee && oqueue[ocptr[i + 1] - 1] == cedges[cum]) { /* this cc has a primal dof */
1614: graph->cptr[ncc + 1] = graph->cptr[ncc] + 1;
1615: graph->queue[graph->cptr[ncc]] = cedges[cum];
1616: ncc++;
1617: lc--;
1618: cum++;
1619: while (cum < nee && cedges[cum] < 0) cum++;
1620: }
1621: graph->cptr[ncc + 1] = graph->cptr[ncc] + lc;
1622: for (j = 0; j < lc; j++) graph->queue[graph->cptr[ncc] + j] = oqueue[ocptr[i] + j];
1623: ncc++;
1624: }
1625: graph->ncc = ncc;
1626: if (pcbddc->nedclocal) PetscCall(ISRestoreIndices(pcbddc->nedclocal, (const PetscInt **)&idxs));
1627: PetscCall(PetscFree2(ocptr, oqueue));
1628: }
1629: PetscCall(ISLocalToGlobalMappingDestroy(&fl2g));
1630: PetscCall(PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph, NULL, NULL, &nee, &alleedges, &allprimals));
1631: PetscCall(PCBDDCGraphResetCSR(pcbddc->mat_graph));
1633: PetscCall(ISDestroy(&nedfieldlocal));
1634: PetscCall(PetscFree(extrow));
1635: PetscCall(PetscFree2(work, rwork));
1636: PetscCall(PetscFree(corners));
1637: PetscCall(PetscFree(cedges));
1638: PetscCall(PetscFree(extrows));
1639: PetscCall(PetscFree(extcols));
1640: PetscCall(MatDestroy(&lG));
1642: /* Complete assembling */
1643: PetscCall(MatAssemblyEnd(T, MAT_FINAL_ASSEMBLY));
1644: PetscCall(MatViewFromOptions(T, (PetscObject)pc, "-pc_bddc_nedelec_change_view"));
1645: if (pcbddc->nedcG) {
1646: PetscCall(MatAssemblyEnd(pcbddc->nedcG, MAT_FINAL_ASSEMBLY));
1647: PetscCall(MatViewFromOptions(pcbddc->nedcG, (PetscObject)pc, "-pc_bddc_nedelec_coarse_change_view"));
1648: }
1650: PetscCall(ISDestroy(&elements_corners));
1652: /* set change of basis */
1653: PetscCall(PCBDDCSetChangeOfBasisMat(pc, T, PETSC_FALSE));
1654: PetscCall(MatDestroy(&T));
1655: PetscFunctionReturn(PETSC_SUCCESS);
1656: }
1658: /* the near-null space of BDDC carries information on quadrature weights,
1659: and these can be collinear -> so cheat with MatNullSpaceCreate
1660: and create a suitable set of basis vectors first */
1661: PetscErrorCode PCBDDCNullSpaceCreate(MPI_Comm comm, PetscBool has_const, PetscInt nvecs, Vec quad_vecs[], MatNullSpace *nnsp)
1662: {
1663: PetscInt i;
1665: PetscFunctionBegin;
1666: for (i = 0; i < nvecs; i++) {
1667: PetscInt first, last;
1669: PetscCall(VecGetOwnershipRange(quad_vecs[i], &first, &last));
1670: PetscCheck(last - first >= 2 * nvecs || !has_const, PETSC_COMM_SELF, PETSC_ERR_SUP, "Not implemented");
1671: if (i >= first && i < last) {
1672: PetscScalar *data;
1673: PetscCall(VecGetArray(quad_vecs[i], &data));
1674: if (!has_const) {
1675: data[i - first] = 1.;
1676: } else {
1677: data[2 * i - first] = 1. / PetscSqrtReal(2.);
1678: data[2 * i - first + 1] = -1. / PetscSqrtReal(2.);
1679: }
1680: PetscCall(VecRestoreArray(quad_vecs[i], &data));
1681: }
1682: PetscCall(PetscObjectStateIncrease((PetscObject)quad_vecs[i]));
1683: }
1684: PetscCall(MatNullSpaceCreate(comm, has_const, nvecs, quad_vecs, nnsp));
1685: for (i = 0; i < nvecs; i++) { /* reset vectors */
1686: PetscInt first, last;
1687: PetscCall(VecLockReadPop(quad_vecs[i]));
1688: PetscCall(VecGetOwnershipRange(quad_vecs[i], &first, &last));
1689: if (i >= first && i < last) {
1690: PetscScalar *data;
1691: PetscCall(VecGetArray(quad_vecs[i], &data));
1692: if (!has_const) {
1693: data[i - first] = 0.;
1694: } else {
1695: data[2 * i - first] = 0.;
1696: data[2 * i - first + 1] = 0.;
1697: }
1698: PetscCall(VecRestoreArray(quad_vecs[i], &data));
1699: }
1700: PetscCall(PetscObjectStateIncrease((PetscObject)quad_vecs[i]));
1701: PetscCall(VecLockReadPush(quad_vecs[i]));
1702: }
1703: PetscFunctionReturn(PETSC_SUCCESS);
1704: }
1706: PetscErrorCode PCBDDCComputeNoNetFlux(Mat A, Mat divudotp, PetscBool transpose, IS vl2l, PCBDDCGraph graph, MatNullSpace *nnsp)
1707: {
1708: Mat loc_divudotp;
1709: Vec p, v, quad_vec;
1710: ISLocalToGlobalMapping map;
1711: PetscScalar *array;
1713: PetscFunctionBegin;
1714: PetscCall(MatCreateVecs(A, &quad_vec, NULL));
1715: if (!transpose) {
1716: PetscCall(MatISGetLocalToGlobalMapping(A, &map, NULL));
1717: } else {
1718: PetscCall(MatISGetLocalToGlobalMapping(A, NULL, &map));
1719: }
1720: PetscCall(PCBDDCNullSpaceCreate(PetscObjectComm((PetscObject)A), PETSC_FALSE, 1, &quad_vec, nnsp));
1721: PetscCall(VecLockReadPop(quad_vec));
1722: PetscCall(VecSetLocalToGlobalMapping(quad_vec, map));
1724: /* compute local quad vec */
1725: PetscCall(MatISGetLocalMat(divudotp, &loc_divudotp));
1726: if (!transpose) {
1727: PetscCall(MatCreateVecs(loc_divudotp, &v, &p));
1728: } else {
1729: PetscCall(MatCreateVecs(loc_divudotp, &p, &v));
1730: }
1731: /* the assumption here is that the constant vector interpolates the constant on the L2 conforming space */
1732: PetscCall(VecSet(p, 1.));
1733: if (!transpose) {
1734: PetscCall(MatMultTranspose(loc_divudotp, p, v));
1735: } else {
1736: PetscCall(MatMult(loc_divudotp, p, v));
1737: }
1738: PetscCall(VecDestroy(&p));
1739: if (vl2l) {
1740: Mat lA;
1741: VecScatter sc;
1742: Vec vins;
1744: PetscCall(MatISGetLocalMat(A, &lA));
1745: PetscCall(MatCreateVecs(lA, &vins, NULL));
1746: PetscCall(VecScatterCreate(v, NULL, vins, vl2l, &sc));
1747: PetscCall(VecScatterBegin(sc, v, vins, INSERT_VALUES, SCATTER_FORWARD));
1748: PetscCall(VecScatterEnd(sc, v, vins, INSERT_VALUES, SCATTER_FORWARD));
1749: PetscCall(VecScatterDestroy(&sc));
1750: PetscCall(VecDestroy(&v));
1751: v = vins;
1752: }
1754: /* mask summation of interface values */
1755: PetscInt n, *mmask, *mask, *idxs, nmr, nr;
1756: const PetscInt *degree;
1757: PetscSF msf;
1759: PetscCall(VecGetLocalSize(v, &n));
1760: PetscCall(PetscSFGetGraph(graph->interface_subset_sf, &nr, NULL, NULL, NULL));
1761: PetscCall(PetscSFGetMultiSF(graph->interface_subset_sf, &msf));
1762: PetscCall(PetscSFGetGraph(msf, &nmr, NULL, NULL, NULL));
1763: PetscCall(PetscCalloc3(nmr, &mmask, n, &mask, n, &idxs));
1764: PetscCall(PetscSFComputeDegreeBegin(graph->interface_subset_sf, °ree));
1765: PetscCall(PetscSFComputeDegreeEnd(graph->interface_subset_sf, °ree));
1766: for (PetscInt i = 0, c = 0; i < nr; i++) {
1767: mmask[c] = 1;
1768: c += degree[i];
1769: }
1770: PetscCall(PetscSFScatterBegin(graph->interface_subset_sf, MPIU_INT, mmask, mask));
1771: PetscCall(PetscSFScatterEnd(graph->interface_subset_sf, MPIU_INT, mmask, mask));
1772: PetscCall(VecGetArray(v, &array));
1773: for (PetscInt i = 0; i < n; i++) {
1774: array[i] *= mask[i];
1775: idxs[i] = i;
1776: }
1777: PetscCall(VecSetValuesLocal(quad_vec, n, idxs, array, ADD_VALUES));
1778: PetscCall(VecRestoreArray(v, &array));
1779: PetscCall(PetscFree3(mmask, mask, idxs));
1780: PetscCall(VecDestroy(&v));
1781: PetscCall(VecAssemblyBegin(quad_vec));
1782: PetscCall(VecAssemblyEnd(quad_vec));
1783: PetscCall(VecViewFromOptions(quad_vec, NULL, "-pc_bddc_quad_vec_view"));
1784: PetscCall(VecLockReadPush(quad_vec));
1785: PetscCall(VecDestroy(&quad_vec));
1786: PetscFunctionReturn(PETSC_SUCCESS);
1787: }
1789: PetscErrorCode PCBDDCAddPrimalVerticesLocalIS(PC pc, IS primalv)
1790: {
1791: PC_BDDC *pcbddc = (PC_BDDC *)pc->data;
1793: PetscFunctionBegin;
1794: if (primalv) {
1795: if (pcbddc->user_primal_vertices_local) {
1796: IS list[2], newp;
1798: list[0] = primalv;
1799: list[1] = pcbddc->user_primal_vertices_local;
1800: PetscCall(ISConcatenate(PetscObjectComm((PetscObject)pc), 2, list, &newp));
1801: PetscCall(ISSortRemoveDups(newp));
1802: PetscCall(ISDestroy(&list[1]));
1803: pcbddc->user_primal_vertices_local = newp;
1804: } else {
1805: PetscCall(PCBDDCSetPrimalVerticesLocalIS(pc, primalv));
1806: }
1807: }
1808: PetscFunctionReturn(PETSC_SUCCESS);
1809: }
1811: static PetscErrorCode func_coords_private(PetscInt dim, PetscReal t, const PetscReal X[], PetscInt Nf, PetscScalar *out, PetscCtx ctx)
1812: {
1813: PetscInt f, *comp = (PetscInt *)ctx;
1815: PetscFunctionBegin;
1816: for (f = 0; f < Nf; f++) out[f] = X[*comp];
1817: PetscFunctionReturn(PETSC_SUCCESS);
1818: }
1820: PetscErrorCode PCBDDCComputeLocalTopologyInfo(PC pc)
1821: {
1822: Vec local, global;
1823: PC_BDDC *pcbddc = (PC_BDDC *)pc->data;
1824: Mat_IS *matis = (Mat_IS *)pc->pmat->data;
1825: PetscBool monolithic = PETSC_FALSE;
1827: PetscFunctionBegin;
1828: PetscOptionsBegin(PetscObjectComm((PetscObject)pc), ((PetscObject)pc)->prefix, "BDDC topology options", "PC");
1829: PetscCall(PetscOptionsBool("-pc_bddc_monolithic", "Discard any information on dofs splitting", NULL, monolithic, &monolithic, NULL));
1830: PetscOptionsEnd();
1831: /* need to convert from global to local topology information and remove references to information in global ordering */
1832: PetscCall(MatCreateVecs(pc->pmat, &global, NULL));
1833: PetscCall(MatCreateVecs(matis->A, &local, NULL));
1834: PetscCall(VecBindToCPU(global, PETSC_TRUE));
1835: PetscCall(VecBindToCPU(local, PETSC_TRUE));
1836: if (monolithic) { /* just get block size to properly compute vertices */
1837: if (pcbddc->vertex_size == 1) PetscCall(MatGetBlockSize(pc->pmat, &pcbddc->vertex_size));
1838: goto boundary;
1839: }
1841: if (pcbddc->user_provided_isfordofs) {
1842: if (pcbddc->n_ISForDofs) {
1843: PetscInt i;
1845: PetscCall(PetscMalloc1(pcbddc->n_ISForDofs, &pcbddc->ISForDofsLocal));
1846: for (i = 0; i < pcbddc->n_ISForDofs; i++) {
1847: PetscInt bs;
1849: PetscCall(PCBDDCGlobalToLocal(matis->rctx, global, local, pcbddc->ISForDofs[i], &pcbddc->ISForDofsLocal[i]));
1850: PetscCall(ISGetBlockSize(pcbddc->ISForDofs[i], &bs));
1851: PetscCall(ISSetBlockSize(pcbddc->ISForDofsLocal[i], bs));
1852: PetscCall(ISDestroy(&pcbddc->ISForDofs[i]));
1853: }
1854: pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs;
1855: pcbddc->n_ISForDofs = 0;
1856: PetscCall(PetscFree(pcbddc->ISForDofs));
1857: }
1858: } else {
1859: if (!pcbddc->n_ISForDofsLocal) { /* field split not present */
1860: DM dm;
1862: PetscCall(MatGetDM(pc->pmat, &dm));
1863: if (!dm) PetscCall(PCGetDM(pc, &dm));
1864: if (dm) {
1865: IS *fields;
1866: PetscInt nf, i;
1868: PetscCall(DMCreateFieldDecomposition(dm, &nf, NULL, &fields, NULL));
1869: PetscCall(PetscMalloc1(nf, &pcbddc->ISForDofsLocal));
1870: for (i = 0; i < nf; i++) {
1871: PetscInt bs;
1873: PetscCall(PCBDDCGlobalToLocal(matis->rctx, global, local, fields[i], &pcbddc->ISForDofsLocal[i]));
1874: PetscCall(ISGetBlockSize(fields[i], &bs));
1875: PetscCall(ISSetBlockSize(pcbddc->ISForDofsLocal[i], bs));
1876: PetscCall(ISDestroy(&fields[i]));
1877: }
1878: PetscCall(PetscFree(fields));
1879: pcbddc->n_ISForDofsLocal = nf;
1880: } else { /* See if MATIS has fields attached by the conversion from MatNest */
1881: PetscContainer c;
1883: PetscCall(PetscObjectQuery((PetscObject)pc->pmat, "_convert_nest_lfields", (PetscObject *)&c));
1884: if (c) {
1885: MatISLocalFields lf;
1886: PetscCall(PetscContainerGetPointer(c, &lf));
1887: PetscCall(PCBDDCSetDofsSplittingLocal(pc, lf->nr, lf->rf));
1888: } else { /* fallback, create the default fields if bs > 1 */
1889: PetscInt i, n = matis->A->rmap->n;
1890: PetscCall(MatGetBlockSize(pc->pmat, &i));
1891: if (i > 1) {
1892: pcbddc->n_ISForDofsLocal = i;
1893: PetscCall(PetscMalloc1(pcbddc->n_ISForDofsLocal, &pcbddc->ISForDofsLocal));
1894: for (i = 0; i < pcbddc->n_ISForDofsLocal; i++) PetscCall(ISCreateStride(PetscObjectComm((PetscObject)pc), n / pcbddc->n_ISForDofsLocal, i, pcbddc->n_ISForDofsLocal, &pcbddc->ISForDofsLocal[i]));
1895: }
1896: }
1897: }
1898: } else {
1899: PetscInt i;
1900: for (i = 0; i < pcbddc->n_ISForDofsLocal; i++) PetscCall(PCBDDCConsistencyCheckIS(pc, MPI_LAND, &pcbddc->ISForDofsLocal[i]));
1901: }
1902: }
1904: boundary:
1905: if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) {
1906: PetscCall(PCBDDCGlobalToLocal(matis->rctx, global, local, pcbddc->DirichletBoundaries, &pcbddc->DirichletBoundariesLocal));
1907: } else if (pcbddc->DirichletBoundariesLocal) {
1908: PetscCall(PCBDDCConsistencyCheckIS(pc, MPI_LAND, &pcbddc->DirichletBoundariesLocal));
1909: }
1910: if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) {
1911: PetscCall(PCBDDCGlobalToLocal(matis->rctx, global, local, pcbddc->NeumannBoundaries, &pcbddc->NeumannBoundariesLocal));
1912: } else if (pcbddc->NeumannBoundariesLocal) {
1913: PetscCall(PCBDDCConsistencyCheckIS(pc, MPI_LOR, &pcbddc->NeumannBoundariesLocal));
1914: }
1915: if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) PetscCall(PCBDDCGlobalToLocal(matis->rctx, global, local, pcbddc->user_primal_vertices, &pcbddc->user_primal_vertices_local));
1916: PetscCall(VecDestroy(&global));
1917: PetscCall(VecDestroy(&local));
1918: /* detect local disconnected subdomains if requested or needed */
1919: if (pcbddc->detect_disconnected || matis->allow_repeated) {
1920: IS primalv = NULL;
1921: PetscInt nel;
1922: PetscBool filter = pcbddc->detect_disconnected_filter;
1924: for (PetscInt i = 0; i < pcbddc->n_local_subs; i++) PetscCall(ISDestroy(&pcbddc->local_subs[i]));
1925: PetscCall(PetscFree(pcbddc->local_subs));
1926: PetscCall(MatGetVariableBlockSizes(matis->A, &nel, NULL));
1927: if (matis->allow_repeated && nel) {
1928: const PetscInt *elsizes;
1930: pcbddc->n_local_subs = nel;
1931: PetscCall(MatGetVariableBlockSizes(matis->A, NULL, &elsizes));
1932: PetscCall(PetscMalloc1(nel, &pcbddc->local_subs));
1933: for (PetscInt i = 0, c = 0; i < nel; i++) {
1934: PetscCall(ISCreateStride(PETSC_COMM_SELF, elsizes[i], c, 1, &pcbddc->local_subs[i]));
1935: c += elsizes[i];
1936: }
1937: } else {
1938: PetscCall(PCBDDCDetectDisconnectedComponents(pc, filter, &pcbddc->n_local_subs, &pcbddc->local_subs, &primalv));
1939: }
1940: PetscCall(PCBDDCAddPrimalVerticesLocalIS(pc, primalv));
1941: PetscCall(ISDestroy(&primalv));
1942: }
1943: /* early stage corner detection */
1944: {
1945: DM dm;
1947: PetscCall(MatGetDM(pc->pmat, &dm));
1948: if (!dm) PetscCall(PCGetDM(pc, &dm));
1949: if (dm) {
1950: PetscBool isda;
1952: PetscCall(PetscObjectTypeCompare((PetscObject)dm, DMDA, &isda));
1953: if (isda) {
1954: ISLocalToGlobalMapping l2l;
1955: IS corners;
1956: Mat lA;
1957: PetscBool gl, lo;
1959: {
1960: Vec cvec;
1961: const PetscScalar *coords;
1962: PetscInt dof, n, cdim;
1963: PetscBool memc = PetscDefined(USE_COMPLEX) ? PETSC_FALSE : PETSC_TRUE;
1965: PetscCall(DMDAGetInfo(dm, NULL, NULL, NULL, NULL, NULL, NULL, NULL, &dof, NULL, NULL, NULL, NULL, NULL));
1966: PetscCall(DMGetCoordinates(dm, &cvec));
1967: PetscCall(VecGetLocalSize(cvec, &n));
1968: PetscCall(VecGetBlockSize(cvec, &cdim));
1969: n /= cdim;
1970: PetscCall(PetscFree(pcbddc->mat_graph->coords));
1971: PetscCall(PetscMalloc1(dof * n * cdim, &pcbddc->mat_graph->coords));
1972: PetscCall(VecGetArrayRead(cvec, &coords));
1973: if (dof != 1) memc = PETSC_FALSE;
1974: if (memc) {
1975: PetscCall(PetscArraycpy(pcbddc->mat_graph->coords, coords, cdim * n * dof));
1976: } else { /* BDDC graph does not use any blocked information, we need to replicate the data */
1977: PetscReal *bcoords = pcbddc->mat_graph->coords;
1978: PetscInt i, b, d;
1980: for (i = 0; i < n; i++) {
1981: for (b = 0; b < dof; b++) {
1982: for (d = 0; d < cdim; d++) bcoords[i * dof * cdim + b * cdim + d] = PetscRealPart(coords[i * cdim + d]);
1983: }
1984: }
1985: }
1986: PetscCall(VecRestoreArrayRead(cvec, &coords));
1987: pcbddc->mat_graph->cdim = cdim;
1988: pcbddc->mat_graph->cnloc = dof * n;
1989: pcbddc->mat_graph->cloc = PETSC_FALSE;
1990: }
1991: PetscCall(DMDAGetSubdomainCornersIS(dm, &corners));
1992: PetscCall(MatISGetLocalMat(pc->pmat, &lA));
1993: PetscCall(MatGetLocalToGlobalMapping(lA, &l2l, NULL));
1994: PetscCall(MatISRestoreLocalMat(pc->pmat, &lA));
1995: lo = (PetscBool)(l2l && corners);
1996: PetscCallMPI(MPIU_Allreduce(&lo, &gl, 1, MPI_C_BOOL, MPI_LAND, PetscObjectComm((PetscObject)pc)));
1997: if (gl) { /* From PETSc's DMDA */
1998: const PetscInt *idx;
1999: PetscInt dof, bs, *idxout, n;
2001: PetscCall(DMDAGetInfo(dm, NULL, NULL, NULL, NULL, NULL, NULL, NULL, &dof, NULL, NULL, NULL, NULL, NULL));
2002: PetscCall(ISLocalToGlobalMappingGetBlockSize(l2l, &bs));
2003: PetscCall(ISGetLocalSize(corners, &n));
2004: PetscCall(ISGetIndices(corners, &idx));
2005: if (bs == dof) {
2006: PetscCall(PetscMalloc1(n, &idxout));
2007: PetscCall(ISLocalToGlobalMappingApplyBlock(l2l, n, idx, idxout));
2008: } else { /* the original DMDA local-to-local map have been modified */
2009: PetscInt i, d;
2011: PetscCall(PetscMalloc1(dof * n, &idxout));
2012: for (i = 0; i < n; i++)
2013: for (d = 0; d < dof; d++) idxout[dof * i + d] = dof * idx[i] + d;
2014: PetscCall(ISLocalToGlobalMappingApply(l2l, dof * n, idxout, idxout));
2016: bs = 1;
2017: n *= dof;
2018: }
2019: PetscCall(ISRestoreIndices(corners, &idx));
2020: PetscCall(DMDARestoreSubdomainCornersIS(dm, &corners));
2021: PetscCall(ISCreateBlock(PetscObjectComm((PetscObject)pc), bs, n, idxout, PETSC_OWN_POINTER, &corners));
2022: PetscCall(PCBDDCAddPrimalVerticesLocalIS(pc, corners));
2023: PetscCall(ISDestroy(&corners));
2024: pcbddc->corner_selected = PETSC_TRUE;
2025: pcbddc->corner_selection = PETSC_TRUE;
2026: }
2027: if (corners) PetscCall(DMDARestoreSubdomainCornersIS(dm, &corners));
2028: }
2029: }
2030: }
2031: if (pcbddc->corner_selection && !pcbddc->mat_graph->cdim) {
2032: DM dm;
2034: PetscCall(MatGetDM(pc->pmat, &dm));
2035: if (!dm) PetscCall(PCGetDM(pc, &dm));
2036: if (dm) { /* this can get very expensive, I need to find a faster alternative */
2037: Vec vcoords;
2038: PetscSection section;
2039: PetscReal *coords;
2040: PetscInt d, cdim, nl, nf, **ctxs;
2041: PetscErrorCode (**funcs)(PetscInt, PetscReal, const PetscReal *, PetscInt, PetscScalar *, void *);
2042: /* debug coordinates */
2043: PetscViewer viewer;
2044: PetscBool flg;
2045: PetscViewerFormat format;
2046: const char *prefix;
2048: PetscCall(DMGetCoordinateDim(dm, &cdim));
2049: PetscCall(DMGetLocalSection(dm, §ion));
2050: PetscCall(PetscSectionGetNumFields(section, &nf));
2051: PetscCall(DMCreateGlobalVector(dm, &vcoords));
2052: PetscCall(VecGetLocalSize(vcoords, &nl));
2053: PetscCall(PetscMalloc1(nl * cdim, &coords));
2054: PetscCall(PetscMalloc2(nf, &funcs, nf, &ctxs));
2055: PetscCall(PetscMalloc1(nf, &ctxs[0]));
2056: for (d = 0; d < nf; d++) funcs[d] = func_coords_private;
2057: for (d = 1; d < nf; d++) ctxs[d] = ctxs[d - 1] + 1;
2059: /* debug coordinates */
2060: PetscCall(PCGetOptionsPrefix(pc, &prefix));
2061: PetscCall(PetscOptionsCreateViewer(PetscObjectComm((PetscObject)vcoords), ((PetscObject)vcoords)->options, prefix, "-pc_bddc_coords_vec_view", &viewer, &format, &flg));
2062: if (flg) PetscCall(PetscViewerPushFormat(viewer, format));
2063: for (d = 0; d < cdim; d++) {
2064: PetscInt i;
2065: const PetscScalar *v;
2066: char name[16];
2068: for (i = 0; i < nf; i++) ctxs[i][0] = d;
2069: PetscCall(PetscSNPrintf(name, sizeof(name), "bddc_coords_%" PetscInt_FMT, d));
2070: PetscCall(PetscObjectSetName((PetscObject)vcoords, name));
2071: PetscCall(DMProjectFunction(dm, 0.0, funcs, (void **)ctxs, INSERT_VALUES, vcoords));
2072: if (flg) PetscCall(VecView(vcoords, viewer));
2073: PetscCall(VecGetArrayRead(vcoords, &v));
2074: for (i = 0; i < nl; i++) coords[i * cdim + d] = PetscRealPart(v[i]);
2075: PetscCall(VecRestoreArrayRead(vcoords, &v));
2076: }
2077: PetscCall(VecDestroy(&vcoords));
2078: PetscCall(PCSetCoordinates(pc, cdim, nl, coords));
2079: PetscCall(PetscFree(coords));
2080: PetscCall(PetscFree(ctxs[0]));
2081: PetscCall(PetscFree2(funcs, ctxs));
2082: if (flg) {
2083: PetscCall(PetscViewerPopFormat(viewer));
2084: PetscCall(PetscViewerDestroy(&viewer));
2085: }
2086: }
2087: }
2088: PetscFunctionReturn(PETSC_SUCCESS);
2089: }
2091: PetscErrorCode PCBDDCConsistencyCheckIS(PC pc, MPI_Op mop, IS *is)
2092: {
2093: Mat_IS *matis = (Mat_IS *)pc->pmat->data;
2094: IS nis;
2095: const PetscInt *idxs;
2096: PetscInt i, nd, n = matis->A->rmap->n, *nidxs, nnd;
2098: PetscFunctionBegin;
2099: PetscCheck(mop == MPI_LAND || mop == MPI_LOR, PetscObjectComm((PetscObject)pc), PETSC_ERR_SUP, "Supported are MPI_LAND and MPI_LOR");
2100: if (mop == MPI_LAND) {
2101: /* init rootdata with true */
2102: for (i = 0; i < pc->pmat->rmap->n; i++) matis->sf_rootdata[i] = 1;
2103: } else {
2104: PetscCall(PetscArrayzero(matis->sf_rootdata, pc->pmat->rmap->n));
2105: }
2106: PetscCall(PetscArrayzero(matis->sf_leafdata, n));
2107: PetscCall(ISGetLocalSize(*is, &nd));
2108: PetscCall(ISGetIndices(*is, &idxs));
2109: for (i = 0; i < nd; i++)
2110: if (-1 < idxs[i] && idxs[i] < n) matis->sf_leafdata[idxs[i]] = 1;
2111: PetscCall(ISRestoreIndices(*is, &idxs));
2112: PetscCall(PetscSFReduceBegin(matis->sf, MPIU_INT, matis->sf_leafdata, matis->sf_rootdata, mop));
2113: PetscCall(PetscSFReduceEnd(matis->sf, MPIU_INT, matis->sf_leafdata, matis->sf_rootdata, mop));
2114: PetscCall(PetscSFBcastBegin(matis->sf, MPIU_INT, matis->sf_rootdata, matis->sf_leafdata, MPI_REPLACE));
2115: PetscCall(PetscSFBcastEnd(matis->sf, MPIU_INT, matis->sf_rootdata, matis->sf_leafdata, MPI_REPLACE));
2116: if (mop == MPI_LAND) PetscCall(PetscMalloc1(nd, &nidxs));
2117: else PetscCall(PetscMalloc1(n, &nidxs));
2118: for (i = 0, nnd = 0; i < n; i++)
2119: if (matis->sf_leafdata[i]) nidxs[nnd++] = i;
2120: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)*is), nnd, nidxs, PETSC_OWN_POINTER, &nis));
2121: PetscCall(ISDestroy(is));
2122: *is = nis;
2123: PetscFunctionReturn(PETSC_SUCCESS);
2124: }
2126: PetscErrorCode PCBDDCBenignRemoveInterior(PC pc, Vec r, Vec z)
2127: {
2128: PC_IS *pcis = (PC_IS *)pc->data;
2129: PC_BDDC *pcbddc = (PC_BDDC *)pc->data;
2131: PetscFunctionBegin;
2132: if (!pcbddc->benign_have_null) PetscFunctionReturn(PETSC_SUCCESS);
2133: if (pcbddc->ChangeOfBasisMatrix) {
2134: Vec swap;
2136: PetscCall(MatMultTranspose(pcbddc->ChangeOfBasisMatrix, r, pcbddc->work_change));
2137: swap = pcbddc->work_change;
2138: pcbddc->work_change = r;
2139: r = swap;
2140: }
2141: PetscCall(VecScatterBegin(pcis->global_to_D, r, pcis->vec1_D, INSERT_VALUES, SCATTER_FORWARD));
2142: PetscCall(VecScatterEnd(pcis->global_to_D, r, pcis->vec1_D, INSERT_VALUES, SCATTER_FORWARD));
2143: PetscCall(PetscLogEventBegin(PC_BDDC_Solves[pcbddc->current_level][0], pc, 0, 0, 0));
2144: PetscCall(KSPSolve(pcbddc->ksp_D, pcis->vec1_D, pcis->vec2_D));
2145: PetscCall(PetscLogEventEnd(PC_BDDC_Solves[pcbddc->current_level][0], pc, 0, 0, 0));
2146: PetscCall(KSPCheckSolve(pcbddc->ksp_D, pc, pcis->vec2_D));
2147: PetscCall(VecSet(z, 0.));
2148: PetscCall(VecScatterBegin(pcis->global_to_D, pcis->vec2_D, z, INSERT_VALUES, SCATTER_REVERSE));
2149: PetscCall(VecScatterEnd(pcis->global_to_D, pcis->vec2_D, z, INSERT_VALUES, SCATTER_REVERSE));
2150: if (pcbddc->ChangeOfBasisMatrix) {
2151: pcbddc->work_change = r;
2152: PetscCall(VecCopy(z, pcbddc->work_change));
2153: PetscCall(MatMult(pcbddc->ChangeOfBasisMatrix, pcbddc->work_change, z));
2154: }
2155: PetscFunctionReturn(PETSC_SUCCESS);
2156: }
2158: static PetscErrorCode PCBDDCBenignMatMult_Private_Private(Mat A, Vec x, Vec y, PetscBool transpose)
2159: {
2160: PCBDDCBenignMatMult_ctx ctx;
2161: PetscBool apply_right, apply_left, reset_x;
2163: PetscFunctionBegin;
2164: PetscCall(MatShellGetContext(A, &ctx));
2165: if (transpose) {
2166: apply_right = ctx->apply_left;
2167: apply_left = ctx->apply_right;
2168: } else {
2169: apply_right = ctx->apply_right;
2170: apply_left = ctx->apply_left;
2171: }
2172: reset_x = PETSC_FALSE;
2173: if (apply_right) {
2174: const PetscScalar *ax;
2175: PetscInt nl, i;
2177: PetscCall(VecGetLocalSize(x, &nl));
2178: PetscCall(VecGetArrayRead(x, &ax));
2179: PetscCall(PetscArraycpy(ctx->work, ax, nl));
2180: PetscCall(VecRestoreArrayRead(x, &ax));
2181: for (i = 0; i < ctx->benign_n; i++) {
2182: PetscScalar sum, val;
2183: const PetscInt *idxs;
2184: PetscInt nz, j;
2185: PetscCall(ISGetLocalSize(ctx->benign_zerodiag_subs[i], &nz));
2186: PetscCall(ISGetIndices(ctx->benign_zerodiag_subs[i], &idxs));
2187: sum = 0.;
2188: if (ctx->apply_p0) {
2189: val = ctx->work[idxs[nz - 1]];
2190: for (j = 0; j < nz - 1; j++) {
2191: sum += ctx->work[idxs[j]];
2192: ctx->work[idxs[j]] += val;
2193: }
2194: } else {
2195: for (j = 0; j < nz - 1; j++) sum += ctx->work[idxs[j]];
2196: }
2197: ctx->work[idxs[nz - 1]] -= sum;
2198: PetscCall(ISRestoreIndices(ctx->benign_zerodiag_subs[i], &idxs));
2199: }
2200: PetscCall(VecPlaceArray(x, ctx->work));
2201: reset_x = PETSC_TRUE;
2202: }
2203: if (transpose) {
2204: PetscCall(MatMultTranspose(ctx->A, x, y));
2205: } else {
2206: PetscCall(MatMult(ctx->A, x, y));
2207: }
2208: if (reset_x) PetscCall(VecResetArray(x));
2209: if (apply_left) {
2210: PetscScalar *ay;
2211: PetscInt i;
2213: PetscCall(VecGetArray(y, &ay));
2214: for (i = 0; i < ctx->benign_n; i++) {
2215: PetscScalar sum, val;
2216: const PetscInt *idxs;
2217: PetscInt nz, j;
2218: PetscCall(ISGetLocalSize(ctx->benign_zerodiag_subs[i], &nz));
2219: PetscCall(ISGetIndices(ctx->benign_zerodiag_subs[i], &idxs));
2220: val = -ay[idxs[nz - 1]];
2221: if (ctx->apply_p0) {
2222: sum = 0.;
2223: for (j = 0; j < nz - 1; j++) {
2224: sum += ay[idxs[j]];
2225: ay[idxs[j]] += val;
2226: }
2227: ay[idxs[nz - 1]] += sum;
2228: } else {
2229: for (j = 0; j < nz - 1; j++) ay[idxs[j]] += val;
2230: ay[idxs[nz - 1]] = 0.;
2231: }
2232: PetscCall(ISRestoreIndices(ctx->benign_zerodiag_subs[i], &idxs));
2233: }
2234: PetscCall(VecRestoreArray(y, &ay));
2235: }
2236: PetscFunctionReturn(PETSC_SUCCESS);
2237: }
2239: static PetscErrorCode PCBDDCBenignMatMultTranspose_Private(Mat A, Vec x, Vec y)
2240: {
2241: PetscFunctionBegin;
2242: PetscCall(PCBDDCBenignMatMult_Private_Private(A, x, y, PETSC_TRUE));
2243: PetscFunctionReturn(PETSC_SUCCESS);
2244: }
2246: static PetscErrorCode PCBDDCBenignMatMult_Private(Mat A, Vec x, Vec y)
2247: {
2248: PetscFunctionBegin;
2249: PetscCall(PCBDDCBenignMatMult_Private_Private(A, x, y, PETSC_FALSE));
2250: PetscFunctionReturn(PETSC_SUCCESS);
2251: }
2253: PetscErrorCode PCBDDCBenignShellMat(PC pc, PetscBool restore)
2254: {
2255: PC_IS *pcis = (PC_IS *)pc->data;
2256: PC_BDDC *pcbddc = (PC_BDDC *)pc->data;
2257: PCBDDCBenignMatMult_ctx ctx;
2259: PetscFunctionBegin;
2260: if (!restore) {
2261: Mat A_IB, A_BI;
2262: PetscScalar *work;
2263: PCBDDCReuseSolvers reuse = pcbddc->sub_schurs ? pcbddc->sub_schurs->reuse_solver : NULL;
2265: PetscCheck(!pcbddc->benign_original_mat, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Benign original mat has not been restored");
2266: if (!pcbddc->benign_change || !pcbddc->benign_n || pcbddc->benign_change_explicit) PetscFunctionReturn(PETSC_SUCCESS);
2267: PetscCall(PetscMalloc1(pcis->n, &work));
2268: PetscCall(MatCreate(PETSC_COMM_SELF, &A_IB));
2269: PetscCall(MatSetSizes(A_IB, pcis->n - pcis->n_B, pcis->n_B, PETSC_DECIDE, PETSC_DECIDE));
2270: PetscCall(MatSetType(A_IB, MATSHELL));
2271: PetscCall(MatShellSetOperation(A_IB, MATOP_MULT, (PetscErrorCodeFn *)PCBDDCBenignMatMult_Private));
2272: PetscCall(MatShellSetOperation(A_IB, MATOP_MULT_TRANSPOSE, (PetscErrorCodeFn *)PCBDDCBenignMatMultTranspose_Private));
2273: PetscCall(PetscNew(&ctx));
2274: PetscCall(MatShellSetContext(A_IB, ctx));
2275: ctx->apply_left = PETSC_TRUE;
2276: ctx->apply_right = PETSC_FALSE;
2277: ctx->apply_p0 = PETSC_FALSE;
2278: ctx->benign_n = pcbddc->benign_n;
2279: if (reuse) {
2280: ctx->benign_zerodiag_subs = reuse->benign_zerodiag_subs;
2281: ctx->free = PETSC_FALSE;
2282: } else { /* TODO: could be optimized for successive solves */
2283: ISLocalToGlobalMapping N_to_D;
2284: PetscInt i;
2286: PetscCall(ISLocalToGlobalMappingCreateIS(pcis->is_I_local, &N_to_D));
2287: PetscCall(PetscMalloc1(pcbddc->benign_n, &ctx->benign_zerodiag_subs));
2288: for (i = 0; i < pcbddc->benign_n; i++) PetscCall(ISGlobalToLocalMappingApplyIS(N_to_D, IS_GTOLM_DROP, pcbddc->benign_zerodiag_subs[i], &ctx->benign_zerodiag_subs[i]));
2289: PetscCall(ISLocalToGlobalMappingDestroy(&N_to_D));
2290: ctx->free = PETSC_TRUE;
2291: }
2292: ctx->A = pcis->A_IB;
2293: ctx->work = work;
2294: PetscCall(MatSetUp(A_IB));
2295: PetscCall(MatAssemblyBegin(A_IB, MAT_FINAL_ASSEMBLY));
2296: PetscCall(MatAssemblyEnd(A_IB, MAT_FINAL_ASSEMBLY));
2297: pcis->A_IB = A_IB;
2299: /* A_BI as A_IB^T */
2300: PetscCall(MatCreateTranspose(A_IB, &A_BI));
2301: pcbddc->benign_original_mat = pcis->A_BI;
2302: pcis->A_BI = A_BI;
2303: } else {
2304: if (!pcbddc->benign_original_mat) PetscFunctionReturn(PETSC_SUCCESS);
2305: PetscCall(MatShellGetContext(pcis->A_IB, &ctx));
2306: PetscCall(MatDestroy(&pcis->A_IB));
2307: pcis->A_IB = ctx->A;
2308: ctx->A = NULL;
2309: PetscCall(MatDestroy(&pcis->A_BI));
2310: pcis->A_BI = pcbddc->benign_original_mat;
2311: pcbddc->benign_original_mat = NULL;
2312: if (ctx->free) {
2313: PetscInt i;
2314: for (i = 0; i < ctx->benign_n; i++) PetscCall(ISDestroy(&ctx->benign_zerodiag_subs[i]));
2315: PetscCall(PetscFree(ctx->benign_zerodiag_subs));
2316: }
2317: PetscCall(PetscFree(ctx->work));
2318: PetscCall(PetscFree(ctx));
2319: }
2320: PetscFunctionReturn(PETSC_SUCCESS);
2321: }
2323: /* used just in bddc debug mode */
2324: static PetscErrorCode PCBDDCBenignProject(PC pc, IS is1, IS is2, Mat *B)
2325: {
2326: PC_BDDC *pcbddc = (PC_BDDC *)pc->data;
2327: Mat_IS *matis = (Mat_IS *)pc->pmat->data;
2328: Mat An;
2330: PetscFunctionBegin;
2331: PetscCall(MatPtAP(matis->A, pcbddc->benign_change, MAT_INITIAL_MATRIX, 2.0, &An));
2332: PetscCall(MatZeroRowsColumns(An, pcbddc->benign_n, pcbddc->benign_p0_lidx, 1.0, NULL, NULL));
2333: if (is1) {
2334: PetscCall(MatCreateSubMatrix(An, is1, is2, MAT_INITIAL_MATRIX, B));
2335: PetscCall(MatDestroy(&An));
2336: } else {
2337: *B = An;
2338: }
2339: PetscFunctionReturn(PETSC_SUCCESS);
2340: }
2342: /* TODO: add reuse flag */
2343: PetscErrorCode MatSeqAIJCompress(Mat A, Mat *B)
2344: {
2345: Mat Bt;
2346: PetscScalar *a, *bdata;
2347: const PetscInt *ii, *ij;
2348: PetscInt m, n, i, nnz, *bii, *bij;
2349: PetscBool flg_row;
2351: PetscFunctionBegin;
2352: PetscCall(MatGetSize(A, &n, &m));
2353: PetscCall(MatGetRowIJ(A, 0, PETSC_FALSE, PETSC_FALSE, &n, &ii, &ij, &flg_row));
2354: PetscCall(MatSeqAIJGetArray(A, &a));
2355: nnz = n;
2356: for (i = 0; i < ii[n]; i++) {
2357: if (PetscLikely(PetscAbsScalar(a[i]) > PETSC_SMALL)) nnz++;
2358: }
2359: PetscCall(PetscMalloc1(n + 1, &bii));
2360: PetscCall(PetscMalloc1(nnz, &bij));
2361: PetscCall(PetscMalloc1(nnz, &bdata));
2362: nnz = 0;
2363: bii[0] = 0;
2364: for (i = 0; i < n; i++) {
2365: PetscInt j;
2366: for (j = ii[i]; j < ii[i + 1]; j++) {
2367: PetscScalar entry = a[j];
2368: if (PetscLikely(PetscAbsScalar(entry) > PETSC_SMALL) || (n == m && ij[j] == i)) {
2369: bij[nnz] = ij[j];
2370: bdata[nnz] = entry;
2371: nnz++;
2372: }
2373: }
2374: bii[i + 1] = nnz;
2375: }
2376: PetscCall(MatSeqAIJRestoreArray(A, &a));
2377: PetscCall(MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A), n, m, bii, bij, bdata, &Bt));
2378: PetscCall(MatRestoreRowIJ(A, 0, PETSC_FALSE, PETSC_FALSE, &n, &ii, &ij, &flg_row));
2379: {
2380: Mat_SeqAIJ *b = (Mat_SeqAIJ *)Bt->data;
2381: b->free_a = PETSC_TRUE;
2382: b->free_ij = PETSC_TRUE;
2383: }
2384: if (*B == A) PetscCall(MatDestroy(&A));
2385: *B = Bt;
2386: PetscFunctionReturn(PETSC_SUCCESS);
2387: }
2389: PetscErrorCode PCBDDCDetectDisconnectedComponents(PC pc, PetscBool filter, PetscInt *ncc, IS *cc[], IS *primalv)
2390: {
2391: Mat B = NULL;
2392: DM dm;
2393: IS is_dummy, *cc_n;
2394: ISLocalToGlobalMapping l2gmap_dummy;
2395: PCBDDCGraph graph;
2396: PetscInt *xadj_filtered = NULL, *adjncy_filtered = NULL;
2397: PetscInt i, n;
2398: PetscInt *xadj, *adjncy;
2399: PetscBool isplex = PETSC_FALSE;
2401: PetscFunctionBegin;
2402: if (ncc) *ncc = 0;
2403: if (cc) *cc = NULL;
2404: if (primalv) *primalv = NULL;
2405: PetscCall(PCBDDCGraphCreate(&graph));
2406: PetscCall(MatGetDM(pc->pmat, &dm));
2407: if (!dm) PetscCall(PCGetDM(pc, &dm));
2408: if (dm) PetscCall(PetscObjectTypeCompareAny((PetscObject)dm, &isplex, DMPLEX, DMP4EST, DMP8EST, ""));
2409: if (filter) isplex = PETSC_FALSE;
2411: if (isplex) { /* this code has been modified from plexpartition.c */
2412: PetscInt p, pStart, pEnd, a, adjSize, idx, size, nroots;
2413: PetscInt *adj = NULL;
2414: IS cellNumbering;
2415: const PetscInt *cellNum;
2416: PetscBool useCone, useClosure;
2417: PetscSection section;
2418: PetscSegBuffer adjBuffer;
2419: PetscSF sfPoint;
2421: PetscCall(DMConvert(dm, DMPLEX, &dm));
2422: PetscCall(DMPlexGetHeightStratum(dm, 0, &pStart, &pEnd));
2423: PetscCall(DMGetPointSF(dm, &sfPoint));
2424: PetscCall(PetscSFGetGraph(sfPoint, &nroots, NULL, NULL, NULL));
2425: /* Build adjacency graph via a section/segbuffer */
2426: PetscCall(PetscSectionCreate(PetscObjectComm((PetscObject)dm), §ion));
2427: PetscCall(PetscSectionSetChart(section, pStart, pEnd));
2428: PetscCall(PetscSegBufferCreate(sizeof(PetscInt), 1000, &adjBuffer));
2429: /* Always use FVM adjacency to create partitioner graph */
2430: PetscCall(DMGetBasicAdjacency(dm, &useCone, &useClosure));
2431: PetscCall(DMSetBasicAdjacency(dm, PETSC_TRUE, PETSC_FALSE));
2432: PetscCall(DMPlexGetCellNumbering(dm, &cellNumbering));
2433: PetscCall(ISGetIndices(cellNumbering, &cellNum));
2434: for (n = 0, p = pStart; p < pEnd; p++) {
2435: /* Skip non-owned cells in parallel (ParMetis expects no overlap) */
2436: if (nroots > 0) {
2437: if (cellNum[p] < 0) continue;
2438: }
2439: adjSize = PETSC_DETERMINE;
2440: PetscCall(DMPlexGetAdjacency(dm, p, &adjSize, &adj));
2441: for (a = 0; a < adjSize; ++a) {
2442: const PetscInt point = adj[a];
2443: if (pStart <= point && point < pEnd) {
2444: PetscInt *PETSC_RESTRICT pBuf;
2445: PetscCall(PetscSectionAddDof(section, p, 1));
2446: PetscCall(PetscSegBufferGetInts(adjBuffer, 1, &pBuf));
2447: *pBuf = point;
2448: }
2449: }
2450: n++;
2451: }
2452: PetscCall(DMSetBasicAdjacency(dm, useCone, useClosure));
2453: /* Derive CSR graph from section/segbuffer */
2454: PetscCall(PetscSectionSetUp(section));
2455: PetscCall(PetscSectionGetStorageSize(section, &size));
2456: PetscCall(PetscMalloc1(n + 1, &xadj));
2457: for (idx = 0, p = pStart; p < pEnd; p++) {
2458: if (nroots > 0) {
2459: if (cellNum[p] < 0) continue;
2460: }
2461: PetscCall(PetscSectionGetOffset(section, p, &xadj[idx++]));
2462: }
2463: xadj[n] = size;
2464: PetscCall(PetscSegBufferExtractAlloc(adjBuffer, &adjncy));
2465: /* Clean up */
2466: PetscCall(PetscSegBufferDestroy(&adjBuffer));
2467: PetscCall(PetscSectionDestroy(§ion));
2468: PetscCall(PetscFree(adj));
2469: graph->xadj = xadj;
2470: graph->adjncy = adjncy;
2471: } else {
2472: Mat A;
2473: PetscBool isseqaij, flg_row;
2475: PetscCall(MatISGetLocalMat(pc->pmat, &A));
2476: if (!A->rmap->N || !A->cmap->N) {
2477: PetscCall(PCBDDCGraphDestroy(&graph));
2478: PetscFunctionReturn(PETSC_SUCCESS);
2479: }
2480: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATSEQAIJ, &isseqaij));
2481: if (!isseqaij && filter) {
2482: PetscBool isseqdense;
2484: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATSEQDENSE, &isseqdense));
2485: if (!isseqdense) {
2486: PetscCall(MatConvert(A, MATSEQAIJ, MAT_INITIAL_MATRIX, &B));
2487: } else { /* TODO: rectangular case and LDA */
2488: PetscScalar *array;
2489: PetscReal chop = 1.e-6;
2491: PetscCall(MatDuplicate(A, MAT_COPY_VALUES, &B));
2492: PetscCall(MatDenseGetArray(B, &array));
2493: PetscCall(MatGetSize(B, &n, NULL));
2494: for (i = 0; i < n; i++) {
2495: PetscInt j;
2496: for (j = i + 1; j < n; j++) {
2497: PetscReal thresh = chop * (PetscAbsScalar(array[i * (n + 1)]) + PetscAbsScalar(array[j * (n + 1)]));
2498: if (PetscAbsScalar(array[i * n + j]) < thresh) array[i * n + j] = 0.;
2499: if (PetscAbsScalar(array[j * n + i]) < thresh) array[j * n + i] = 0.;
2500: }
2501: }
2502: PetscCall(MatDenseRestoreArray(B, &array));
2503: PetscCall(MatConvert(B, MATSEQAIJ, MAT_INPLACE_MATRIX, &B));
2504: }
2505: } else {
2506: PetscCall(PetscObjectReference((PetscObject)A));
2507: B = A;
2508: }
2509: PetscCall(MatGetRowIJ(B, 0, PETSC_TRUE, PETSC_FALSE, &n, (const PetscInt **)&xadj, (const PetscInt **)&adjncy, &flg_row));
2511: /* if filter is true, then removes entries lower than PETSC_SMALL in magnitude */
2512: if (filter) {
2513: PetscScalar *data;
2514: PetscInt j, cum;
2516: PetscCall(PetscCalloc2(n + 1, &xadj_filtered, xadj[n], &adjncy_filtered));
2517: PetscCall(MatSeqAIJGetArray(B, &data));
2518: cum = 0;
2519: for (i = 0; i < n; i++) {
2520: PetscInt t;
2522: for (j = xadj[i]; j < xadj[i + 1]; j++) {
2523: if (PetscUnlikely(PetscAbsScalar(data[j]) < PETSC_SMALL)) continue;
2524: adjncy_filtered[cum + xadj_filtered[i]++] = adjncy[j];
2525: }
2526: t = xadj_filtered[i];
2527: xadj_filtered[i] = cum;
2528: cum += t;
2529: }
2530: PetscCall(MatSeqAIJRestoreArray(B, &data));
2531: graph->xadj = xadj_filtered;
2532: graph->adjncy = adjncy_filtered;
2533: } else {
2534: graph->xadj = xadj;
2535: graph->adjncy = adjncy;
2536: }
2537: }
2538: /* compute local connected components using PCBDDCGraph */
2539: graph->seq_graph = PETSC_TRUE; /* analyze local connected components (i.e. disconnected subdomains) irrespective of dofs count */
2540: PetscCall(ISCreateStride(PETSC_COMM_SELF, n, 0, 1, &is_dummy));
2541: PetscCall(ISLocalToGlobalMappingCreateIS(is_dummy, &l2gmap_dummy));
2542: PetscCall(ISDestroy(&is_dummy));
2543: PetscCall(PCBDDCGraphInit(graph, l2gmap_dummy, n, PETSC_INT_MAX));
2544: PetscCall(ISLocalToGlobalMappingDestroy(&l2gmap_dummy));
2545: PetscCall(PCBDDCGraphSetUp(graph, 1, NULL, NULL, 0, NULL, NULL));
2546: PetscCall(PCBDDCGraphComputeConnectedComponents(graph));
2548: /* partial clean up */
2549: PetscCall(PetscFree2(xadj_filtered, adjncy_filtered));
2550: if (B) {
2551: PetscBool flg_row;
2552: PetscCall(MatRestoreRowIJ(B, 0, PETSC_TRUE, PETSC_FALSE, &n, (const PetscInt **)&xadj, (const PetscInt **)&adjncy, &flg_row));
2553: PetscCall(MatDestroy(&B));
2554: }
2555: if (isplex) {
2556: PetscCall(PetscFree(xadj));
2557: PetscCall(PetscFree(adjncy));
2558: }
2560: /* get back data */
2561: if (isplex) {
2562: if (ncc) *ncc = graph->ncc;
2563: if (cc || primalv) {
2564: Mat A;
2565: PetscBT btv, btvt, btvc;
2566: PetscSection subSection;
2567: PetscInt *ids, cum, cump, *cids, *pids;
2568: PetscInt dim, cStart, cEnd, fStart, fEnd, vStart, vEnd, pStart, pEnd;
2570: PetscCall(DMGetDimension(dm, &dim));
2571: PetscCall(DMPlexGetSubdomainSection(dm, &subSection));
2572: PetscCall(DMPlexGetHeightStratum(dm, 1, &fStart, &fEnd));
2573: PetscCall(DMPlexGetHeightStratum(dm, 0, &cStart, &cEnd));
2574: PetscCall(DMPlexGetDepthStratum(dm, 0, &vStart, &vEnd));
2575: PetscCall(DMPlexGetChart(dm, &pStart, &pEnd));
2576: PetscCall(MatISGetLocalMat(pc->pmat, &A));
2577: PetscCall(PetscMalloc3(A->rmap->n, &ids, graph->ncc + 1, &cids, A->rmap->n, &pids));
2578: PetscCall(PetscBTCreate(A->rmap->n, &btv));
2579: PetscCall(PetscBTCreate(A->rmap->n, &btvt));
2580: PetscCall(PetscBTCreate(pEnd - pStart, &btvc));
2582: /* First see if we find corners for the subdomains, i.e. a vertex
2583: shared by at least dim subdomain boundary faces. This does not
2584: cover all the possible cases with simplices but it is enough
2585: for tensor cells */
2586: if (vStart != fStart && dim <= 3) {
2587: for (PetscInt c = cStart; c < cEnd; c++) {
2588: PetscInt nf, cnt = 0, mcnt = dim, *cfaces;
2589: const PetscInt *faces;
2591: PetscCall(DMPlexGetConeSize(dm, c, &nf));
2592: PetscCall(DMGetWorkArray(dm, nf, MPIU_INT, &cfaces));
2593: PetscCall(DMPlexGetCone(dm, c, &faces));
2594: for (PetscInt f = 0; f < nf; f++) {
2595: PetscInt nc, ff;
2597: PetscCall(DMPlexGetSupportSize(dm, faces[f], &nc));
2598: PetscCall(DMPlexGetTreeParent(dm, faces[f], &ff, NULL));
2599: if (nc == 1 && faces[f] == ff) cfaces[cnt++] = faces[f];
2600: }
2601: if (cnt >= mcnt) {
2602: PetscInt size, *closure = NULL;
2604: PetscCall(DMPlexGetTransitiveClosure(dm, c, PETSC_TRUE, &size, &closure));
2605: for (PetscInt k = 0; k < 2 * size; k += 2) {
2606: PetscInt v = closure[k];
2607: if (v >= vStart && v < vEnd) {
2608: PetscInt vsize, *vclosure = NULL;
2610: cnt = 0;
2611: PetscCall(DMPlexGetTransitiveClosure(dm, v, PETSC_FALSE, &vsize, &vclosure));
2612: for (PetscInt vk = 0; vk < 2 * vsize; vk += 2) {
2613: PetscInt f = vclosure[vk];
2614: if (f >= fStart && f < fEnd) {
2615: PetscInt nc, ff;
2616: PetscBool valid = PETSC_FALSE;
2618: for (PetscInt fk = 0; fk < nf; fk++)
2619: if (f == cfaces[fk]) valid = PETSC_TRUE;
2620: if (!valid) continue;
2621: PetscCall(DMPlexGetSupportSize(dm, f, &nc));
2622: PetscCall(DMPlexGetTreeParent(dm, f, &ff, NULL));
2623: if (nc == 1 && f == ff) cnt++;
2624: }
2625: }
2626: if (cnt >= mcnt) PetscCall(PetscBTSet(btvc, v - pStart));
2627: PetscCall(DMPlexRestoreTransitiveClosure(dm, v, PETSC_FALSE, &vsize, &vclosure));
2628: }
2629: }
2630: PetscCall(DMPlexRestoreTransitiveClosure(dm, c, PETSC_TRUE, &size, &closure));
2631: }
2632: PetscCall(DMRestoreWorkArray(dm, nf, MPIU_INT, &cfaces));
2633: }
2634: }
2636: cids[0] = 0;
2637: for (i = 0, cump = 0, cum = 0; i < graph->ncc; i++) {
2638: PetscInt j;
2640: PetscCall(PetscBTMemzero(A->rmap->n, btvt));
2641: for (j = graph->cptr[i]; j < graph->cptr[i + 1]; j++) {
2642: PetscInt k, size, *closure = NULL, cell = graph->queue[j];
2644: PetscCall(DMPlexGetTransitiveClosure(dm, cell, PETSC_TRUE, &size, &closure));
2645: for (k = 0; k < 2 * size; k += 2) {
2646: PetscInt s, pp, p = closure[k], off, dof, cdof;
2648: PetscCall(PetscSectionGetConstraintDof(subSection, p, &cdof));
2649: PetscCall(PetscSectionGetOffset(subSection, p, &off));
2650: PetscCall(PetscSectionGetDof(subSection, p, &dof));
2651: for (s = 0; s < dof - cdof; s++) {
2652: if (PetscBTLookupSet(btvt, off + s)) continue;
2653: if (PetscBTLookup(btvc, p - pStart)) pids[cump++] = off + s; /* subdomain corner */
2654: else if (!PetscBTLookup(btv, off + s)) ids[cum++] = off + s;
2655: else pids[cump++] = off + s; /* cross-vertex */
2656: }
2657: PetscCall(DMPlexGetTreeParent(dm, p, &pp, NULL));
2658: if (pp != p) {
2659: PetscCall(PetscSectionGetConstraintDof(subSection, pp, &cdof));
2660: PetscCall(PetscSectionGetOffset(subSection, pp, &off));
2661: PetscCall(PetscSectionGetDof(subSection, pp, &dof));
2662: for (s = 0; s < dof - cdof; s++) {
2663: if (PetscBTLookupSet(btvt, off + s)) continue;
2664: if (PetscBTLookup(btvc, pp - pStart)) pids[cump++] = off + s; /* subdomain corner */
2665: else if (!PetscBTLookup(btv, off + s)) ids[cum++] = off + s;
2666: else pids[cump++] = off + s; /* cross-vertex */
2667: }
2668: }
2669: }
2670: PetscCall(DMPlexRestoreTransitiveClosure(dm, cell, PETSC_TRUE, &size, &closure));
2671: }
2672: cids[i + 1] = cum;
2673: /* mark dofs as already assigned */
2674: for (j = cids[i]; j < cids[i + 1]; j++) PetscCall(PetscBTSet(btv, ids[j]));
2675: }
2676: if (cc) {
2677: PetscCall(PetscMalloc1(graph->ncc, &cc_n));
2678: for (i = 0; i < graph->ncc; i++) PetscCall(ISCreateGeneral(PETSC_COMM_SELF, cids[i + 1] - cids[i], ids + cids[i], PETSC_COPY_VALUES, &cc_n[i]));
2679: *cc = cc_n;
2680: }
2681: if (primalv) PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)pc), cump, pids, PETSC_COPY_VALUES, primalv));
2682: PetscCall(PetscFree3(ids, cids, pids));
2683: PetscCall(PetscBTDestroy(&btv));
2684: PetscCall(PetscBTDestroy(&btvt));
2685: PetscCall(PetscBTDestroy(&btvc));
2686: PetscCall(DMDestroy(&dm));
2687: }
2688: } else {
2689: if (ncc) *ncc = graph->ncc;
2690: if (cc) {
2691: PetscCall(PetscMalloc1(graph->ncc, &cc_n));
2692: for (i = 0; i < graph->ncc; i++) PetscCall(ISCreateGeneral(PETSC_COMM_SELF, graph->cptr[i + 1] - graph->cptr[i], graph->queue + graph->cptr[i], PETSC_COPY_VALUES, &cc_n[i]));
2693: *cc = cc_n;
2694: }
2695: }
2696: /* clean up graph */
2697: graph->xadj = NULL;
2698: graph->adjncy = NULL;
2699: PetscCall(PCBDDCGraphDestroy(&graph));
2700: PetscFunctionReturn(PETSC_SUCCESS);
2701: }
2703: PetscErrorCode PCBDDCBenignCheck(PC pc, IS zerodiag)
2704: {
2705: PC_BDDC *pcbddc = (PC_BDDC *)pc->data;
2706: PC_IS *pcis = (PC_IS *)pc->data;
2707: IS dirIS = NULL;
2708: PetscInt i;
2710: PetscFunctionBegin;
2711: PetscCall(PCBDDCGraphGetDirichletDofs(pcbddc->mat_graph, &dirIS));
2712: if (zerodiag) {
2713: Mat A;
2714: Vec vec3_N;
2715: PetscScalar *vals;
2716: const PetscInt *idxs;
2717: PetscInt nz, *count;
2719: /* p0 */
2720: PetscCall(VecSet(pcis->vec1_N, 0.));
2721: PetscCall(PetscMalloc1(pcis->n, &vals));
2722: PetscCall(ISGetLocalSize(zerodiag, &nz));
2723: PetscCall(ISGetIndices(zerodiag, &idxs));
2724: for (i = 0; i < nz; i++) vals[i] = 1.;
2725: PetscCall(VecSetValues(pcis->vec1_N, nz, idxs, vals, INSERT_VALUES));
2726: PetscCall(VecAssemblyBegin(pcis->vec1_N));
2727: PetscCall(VecAssemblyEnd(pcis->vec1_N));
2728: /* v_I */
2729: PetscCall(VecSetRandom(pcis->vec2_N, NULL));
2730: for (i = 0; i < nz; i++) vals[i] = 0.;
2731: PetscCall(VecSetValues(pcis->vec2_N, nz, idxs, vals, INSERT_VALUES));
2732: PetscCall(ISRestoreIndices(zerodiag, &idxs));
2733: PetscCall(ISGetIndices(pcis->is_B_local, &idxs));
2734: for (i = 0; i < pcis->n_B; i++) vals[i] = 0.;
2735: PetscCall(VecSetValues(pcis->vec2_N, pcis->n_B, idxs, vals, INSERT_VALUES));
2736: PetscCall(ISRestoreIndices(pcis->is_B_local, &idxs));
2737: if (dirIS) {
2738: PetscInt n;
2740: PetscCall(ISGetLocalSize(dirIS, &n));
2741: PetscCall(ISGetIndices(dirIS, &idxs));
2742: for (i = 0; i < n; i++) vals[i] = 0.;
2743: PetscCall(VecSetValues(pcis->vec2_N, n, idxs, vals, INSERT_VALUES));
2744: PetscCall(ISRestoreIndices(dirIS, &idxs));
2745: }
2746: PetscCall(VecAssemblyBegin(pcis->vec2_N));
2747: PetscCall(VecAssemblyEnd(pcis->vec2_N));
2748: PetscCall(VecDuplicate(pcis->vec1_N, &vec3_N));
2749: PetscCall(VecSet(vec3_N, 0.));
2750: PetscCall(MatISGetLocalMat(pc->pmat, &A));
2751: PetscCall(MatMult(A, pcis->vec1_N, vec3_N));
2752: PetscCall(VecDot(vec3_N, pcis->vec2_N, &vals[0]));
2753: PetscCheck(PetscAbsScalar(vals[0]) <= 1.e-1, PETSC_COMM_SELF, PETSC_ERR_SUP, "Benign trick can not be applied! b(v_I,p_0) = %1.6e (should be numerically 0.)", (double)PetscAbsScalar(vals[0]));
2754: PetscCall(PetscFree(vals));
2755: PetscCall(VecDestroy(&vec3_N));
2757: /* there should not be any pressure dofs lying on the interface */
2758: PetscCall(PetscCalloc1(pcis->n, &count));
2759: PetscCall(ISGetIndices(pcis->is_B_local, &idxs));
2760: for (i = 0; i < pcis->n_B; i++) count[idxs[i]]++;
2761: PetscCall(ISRestoreIndices(pcis->is_B_local, &idxs));
2762: PetscCall(ISGetIndices(zerodiag, &idxs));
2763: for (i = 0; i < nz; i++) PetscCheck(!count[idxs[i]], PETSC_COMM_SELF, PETSC_ERR_SUP, "Benign trick can not be applied! pressure dof %" PetscInt_FMT " is an interface dof", idxs[i]);
2764: PetscCall(ISRestoreIndices(zerodiag, &idxs));
2765: PetscCall(PetscFree(count));
2766: }
2767: PetscCall(ISDestroy(&dirIS));
2769: /* check PCBDDCBenignGetOrSetP0 */
2770: PetscCall(VecSetRandom(pcis->vec1_global, NULL));
2771: for (i = 0; i < pcbddc->benign_n; i++) pcbddc->benign_p0[i] = -PetscGlobalRank - i;
2772: PetscCall(PCBDDCBenignGetOrSetP0(pc, pcis->vec1_global, PETSC_FALSE));
2773: for (i = 0; i < pcbddc->benign_n; i++) pcbddc->benign_p0[i] = 1;
2774: PetscCall(PCBDDCBenignGetOrSetP0(pc, pcis->vec1_global, PETSC_TRUE));
2775: for (i = 0; i < pcbddc->benign_n; i++) {
2776: PetscInt val = PetscRealPart(pcbddc->benign_p0[i]);
2777: PetscCheck(val == -PetscGlobalRank - i, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Error testing PCBDDCBenignGetOrSetP0! Found %g at %" PetscInt_FMT " instead of %g", (double)PetscRealPart(pcbddc->benign_p0[i]), i, (double)(-PetscGlobalRank - i));
2778: }
2779: PetscFunctionReturn(PETSC_SUCCESS);
2780: }
2782: PetscErrorCode PCBDDCBenignDetectSaddlePoint(PC pc, PetscBool reuse, IS *zerodiaglocal)
2783: {
2784: PC_BDDC *pcbddc = (PC_BDDC *)pc->data;
2785: Mat_IS *matis = (Mat_IS *)pc->pmat->data;
2786: IS pressures = NULL, zerodiag = NULL, *bzerodiag = NULL, zerodiag_save, *zerodiag_subs;
2787: PetscInt nz, n, benign_n, bsp = 1;
2788: PetscInt *interior_dofs, n_interior_dofs, nneu;
2789: PetscBool sorted, have_null, has_null_pressures, recompute_zerodiag, checkb;
2791: PetscFunctionBegin;
2792: if (reuse) goto project_b0;
2793: PetscCall(PetscSFDestroy(&pcbddc->benign_sf));
2794: PetscCall(MatDestroy(&pcbddc->benign_B0));
2795: for (n = 0; n < pcbddc->benign_n; n++) PetscCall(ISDestroy(&pcbddc->benign_zerodiag_subs[n]));
2796: PetscCall(PetscFree(pcbddc->benign_zerodiag_subs));
2797: has_null_pressures = PETSC_TRUE;
2798: have_null = PETSC_TRUE;
2799: /* if a local information on dofs is present, gets pressure dofs from command line (uses the last field is not provided)
2800: Without local information, it uses only the zerodiagonal dofs (ok if the pressure block is all zero and it is a scalar field)
2801: Checks if all the pressure dofs in each subdomain have a zero diagonal
2802: If not, a change of basis on pressures is not needed
2803: since the local Schur complements are already SPD
2804: */
2805: if (pcbddc->n_ISForDofsLocal) {
2806: IS iP = NULL;
2807: PetscInt p, *pp;
2808: PetscBool flg, blocked = PETSC_FALSE;
2810: PetscCall(PetscMalloc1(pcbddc->n_ISForDofsLocal, &pp));
2811: n = pcbddc->n_ISForDofsLocal;
2812: PetscOptionsBegin(PetscObjectComm((PetscObject)pc), ((PetscObject)pc)->prefix, "BDDC benign options", "PC");
2813: PetscCall(PetscOptionsIntArray("-pc_bddc_pressure_field", "Field id for pressures", NULL, pp, &n, &flg));
2814: PetscCall(PetscOptionsBool("-pc_bddc_pressure_blocked", "Use blocked pressure fields", NULL, blocked, &blocked, NULL));
2815: PetscOptionsEnd();
2816: if (!flg) {
2817: n = 1;
2818: pp[0] = pcbddc->n_ISForDofsLocal - 1;
2819: }
2821: bsp = 0;
2822: for (p = 0; p < n; p++) {
2823: PetscInt bs = 1;
2825: PetscCheck(pp[p] >= 0 && pp[p] < pcbddc->n_ISForDofsLocal, PetscObjectComm((PetscObject)pc), PETSC_ERR_USER, "Invalid field id for pressures %" PetscInt_FMT, pp[p]);
2826: if (blocked) PetscCall(ISGetBlockSize(pcbddc->ISForDofsLocal[pp[p]], &bs));
2827: bsp += bs;
2828: }
2829: PetscCall(PetscMalloc1(bsp, &bzerodiag));
2830: bsp = 0;
2831: for (p = 0; p < n; p++) {
2832: const PetscInt *idxs;
2833: PetscInt b, bs = 1, npl, *bidxs;
2835: if (blocked) PetscCall(ISGetBlockSize(pcbddc->ISForDofsLocal[pp[p]], &bs));
2836: PetscCall(ISGetLocalSize(pcbddc->ISForDofsLocal[pp[p]], &npl));
2837: PetscCall(ISGetIndices(pcbddc->ISForDofsLocal[pp[p]], &idxs));
2838: PetscCall(PetscMalloc1(npl / bs, &bidxs));
2839: for (b = 0; b < bs; b++) {
2840: PetscInt i;
2842: for (i = 0; i < npl / bs; i++) bidxs[i] = idxs[bs * i + b];
2843: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, npl / bs, bidxs, PETSC_COPY_VALUES, &bzerodiag[bsp]));
2844: bsp++;
2845: }
2846: PetscCall(PetscFree(bidxs));
2847: PetscCall(ISRestoreIndices(pcbddc->ISForDofsLocal[pp[p]], &idxs));
2848: }
2849: PetscCall(ISConcatenate(PETSC_COMM_SELF, bsp, bzerodiag, &pressures));
2851: /* remove zeroed out pressures if we are setting up a BDDC solver for a saddle-point FETI-DP */
2852: PetscCall(PetscObjectQuery((PetscObject)pc, "__KSPFETIDP_lP", (PetscObject *)&iP));
2853: if (iP) {
2854: IS newpressures;
2856: PetscCall(ISDifference(pressures, iP, &newpressures));
2857: PetscCall(ISDestroy(&pressures));
2858: pressures = newpressures;
2859: }
2860: PetscCall(ISSorted(pressures, &sorted));
2861: if (!sorted) PetscCall(ISSort(pressures));
2862: PetscCall(PetscFree(pp));
2863: }
2865: /* pcis has not been setup yet, so get the local size from the subdomain matrix */
2866: PetscCall(MatGetLocalSize(pcbddc->local_mat, &n, NULL));
2867: if (!n) pcbddc->benign_change_explicit = PETSC_TRUE;
2868: PetscCall(MatFindZeroDiagonals(pcbddc->local_mat, &zerodiag));
2869: PetscCall(ISSorted(zerodiag, &sorted));
2870: if (!sorted) PetscCall(ISSort(zerodiag));
2871: PetscCall(PetscObjectReference((PetscObject)zerodiag));
2872: zerodiag_save = zerodiag;
2873: PetscCall(ISGetLocalSize(zerodiag, &nz));
2874: if (!nz) {
2875: if (n) have_null = PETSC_FALSE;
2876: has_null_pressures = PETSC_FALSE;
2877: PetscCall(ISDestroy(&zerodiag));
2878: }
2879: recompute_zerodiag = PETSC_FALSE;
2881: /* in case disconnected subdomains info is present, split the pressures accordingly (otherwise the benign trick could fail) */
2882: zerodiag_subs = NULL;
2883: benign_n = 0;
2884: n_interior_dofs = 0;
2885: interior_dofs = NULL;
2886: nneu = 0;
2887: if (pcbddc->NeumannBoundariesLocal) PetscCall(ISGetLocalSize(pcbddc->NeumannBoundariesLocal, &nneu));
2888: checkb = (PetscBool)(!pcbddc->NeumannBoundariesLocal || pcbddc->current_level);
2889: if (checkb) { /* need to compute interior nodes */
2890: PetscInt n, i;
2891: PetscInt *count;
2892: ISLocalToGlobalMapping mapping;
2894: PetscCall(MatISGetLocalToGlobalMapping(pc->pmat, &mapping, NULL));
2895: PetscCall(ISLocalToGlobalMappingGetNodeInfo(mapping, &n, &count, NULL));
2896: PetscCall(PetscMalloc1(n, &interior_dofs));
2897: for (i = 0; i < n; i++)
2898: if (count[i] < 2) interior_dofs[n_interior_dofs++] = i;
2899: PetscCall(ISLocalToGlobalMappingRestoreNodeInfo(mapping, &n, &count, NULL));
2900: }
2901: if (has_null_pressures) {
2902: IS *subs;
2903: PetscInt nsubs, i, j, nl;
2904: const PetscInt *idxs;
2905: PetscScalar *array;
2906: Vec *work;
2908: subs = pcbddc->local_subs;
2909: nsubs = pcbddc->n_local_subs;
2910: /* these vectors are needed to check if the constant on pressures is in the kernel of the local operator B (i.e. B(v_I,p0) should be zero) */
2911: if (checkb) {
2912: PetscCall(VecDuplicateVecs(matis->y, 2, &work));
2913: PetscCall(ISGetLocalSize(zerodiag, &nl));
2914: PetscCall(ISGetIndices(zerodiag, &idxs));
2915: /* work[0] = 1_p */
2916: PetscCall(VecSet(work[0], 0.));
2917: PetscCall(VecGetArray(work[0], &array));
2918: for (j = 0; j < nl; j++) array[idxs[j]] = 1.;
2919: PetscCall(VecRestoreArray(work[0], &array));
2920: /* work[0] = 1_v */
2921: PetscCall(VecSet(work[1], 1.));
2922: PetscCall(VecGetArray(work[1], &array));
2923: for (j = 0; j < nl; j++) array[idxs[j]] = 0.;
2924: PetscCall(VecRestoreArray(work[1], &array));
2925: PetscCall(ISRestoreIndices(zerodiag, &idxs));
2926: }
2928: if (nsubs > 1 || bsp > 1) {
2929: IS *is;
2930: PetscInt b, totb;
2932: totb = bsp;
2933: is = bsp > 1 ? bzerodiag : &zerodiag;
2934: nsubs = PetscMax(nsubs, 1);
2935: PetscCall(PetscCalloc1(nsubs * totb, &zerodiag_subs));
2936: for (b = 0; b < totb; b++) {
2937: for (i = 0; i < nsubs; i++) {
2938: ISLocalToGlobalMapping l2g;
2939: IS t_zerodiag_subs;
2940: PetscInt nl;
2942: if (subs) {
2943: PetscCall(ISLocalToGlobalMappingCreateIS(subs[i], &l2g));
2944: } else {
2945: IS tis;
2947: PetscCall(MatGetLocalSize(pcbddc->local_mat, &nl, NULL));
2948: PetscCall(ISCreateStride(PETSC_COMM_SELF, nl, 0, 1, &tis));
2949: PetscCall(ISLocalToGlobalMappingCreateIS(tis, &l2g));
2950: PetscCall(ISDestroy(&tis));
2951: }
2952: PetscCall(ISGlobalToLocalMappingApplyIS(l2g, IS_GTOLM_DROP, is[b], &t_zerodiag_subs));
2953: PetscCall(ISGetLocalSize(t_zerodiag_subs, &nl));
2954: if (nl) {
2955: PetscBool valid = PETSC_TRUE;
2957: if (checkb) {
2958: PetscCall(VecSet(matis->x, 0));
2959: PetscCall(ISGetLocalSize(subs[i], &nl));
2960: PetscCall(ISGetIndices(subs[i], &idxs));
2961: PetscCall(VecGetArray(matis->x, &array));
2962: for (j = 0; j < nl; j++) array[idxs[j]] = 1.;
2963: PetscCall(VecRestoreArray(matis->x, &array));
2964: PetscCall(ISRestoreIndices(subs[i], &idxs));
2965: PetscCall(VecPointwiseMult(matis->x, work[0], matis->x));
2966: PetscCall(MatMult(matis->A, matis->x, matis->y));
2967: PetscCall(VecPointwiseMult(matis->y, work[1], matis->y));
2968: PetscCall(VecGetArray(matis->y, &array));
2969: for (j = 0; j < n_interior_dofs; j++) {
2970: if (PetscAbsScalar(array[interior_dofs[j]]) > PETSC_SMALL) {
2971: valid = PETSC_FALSE;
2972: break;
2973: }
2974: }
2975: PetscCall(VecRestoreArray(matis->y, &array));
2976: }
2977: if (valid && nneu) {
2978: const PetscInt *idxs;
2979: PetscInt nzb;
2981: PetscCall(ISGetIndices(pcbddc->NeumannBoundariesLocal, &idxs));
2982: PetscCall(ISGlobalToLocalMappingApply(l2g, IS_GTOLM_DROP, nneu, idxs, &nzb, NULL));
2983: PetscCall(ISRestoreIndices(pcbddc->NeumannBoundariesLocal, &idxs));
2984: if (nzb) valid = PETSC_FALSE;
2985: }
2986: if (valid && pressures) {
2987: IS t_pressure_subs, tmp;
2988: PetscInt i1, i2;
2990: PetscCall(ISGlobalToLocalMappingApplyIS(l2g, IS_GTOLM_DROP, pressures, &t_pressure_subs));
2991: PetscCall(ISEmbed(t_zerodiag_subs, t_pressure_subs, PETSC_TRUE, &tmp));
2992: PetscCall(ISGetLocalSize(tmp, &i1));
2993: PetscCall(ISGetLocalSize(t_zerodiag_subs, &i2));
2994: if (i2 != i1) valid = PETSC_FALSE;
2995: PetscCall(ISDestroy(&t_pressure_subs));
2996: PetscCall(ISDestroy(&tmp));
2997: }
2998: if (valid) {
2999: PetscCall(ISLocalToGlobalMappingApplyIS(l2g, t_zerodiag_subs, &zerodiag_subs[benign_n]));
3000: benign_n++;
3001: } else recompute_zerodiag = PETSC_TRUE;
3002: }
3003: PetscCall(ISDestroy(&t_zerodiag_subs));
3004: PetscCall(ISLocalToGlobalMappingDestroy(&l2g));
3005: }
3006: }
3007: } else { /* there's just one subdomain (or zero if they have not been detected */
3008: PetscBool valid = PETSC_TRUE;
3010: if (nneu) valid = PETSC_FALSE;
3011: if (valid && pressures) PetscCall(ISEqual(pressures, zerodiag, &valid));
3012: if (valid && checkb) {
3013: PetscCall(MatMult(matis->A, work[0], matis->x));
3014: PetscCall(VecPointwiseMult(matis->x, work[1], matis->x));
3015: PetscCall(VecGetArray(matis->x, &array));
3016: for (j = 0; j < n_interior_dofs; j++) {
3017: if (PetscAbsScalar(array[interior_dofs[j]]) > PETSC_SMALL) {
3018: valid = PETSC_FALSE;
3019: break;
3020: }
3021: }
3022: PetscCall(VecRestoreArray(matis->x, &array));
3023: }
3024: if (valid) {
3025: benign_n = 1;
3026: PetscCall(PetscMalloc1(benign_n, &zerodiag_subs));
3027: PetscCall(PetscObjectReference((PetscObject)zerodiag));
3028: zerodiag_subs[0] = zerodiag;
3029: }
3030: }
3031: if (checkb) PetscCall(VecDestroyVecs(2, &work));
3032: }
3033: PetscCall(PetscFree(interior_dofs));
3035: if (!benign_n) {
3036: PetscInt n;
3038: PetscCall(ISDestroy(&zerodiag));
3039: recompute_zerodiag = PETSC_FALSE;
3040: PetscCall(MatGetLocalSize(pcbddc->local_mat, &n, NULL));
3041: if (n) have_null = PETSC_FALSE;
3042: }
3044: /* final check for null pressures */
3045: if (zerodiag && pressures) PetscCall(ISEqual(pressures, zerodiag, &have_null));
3047: if (recompute_zerodiag) {
3048: PetscCall(ISDestroy(&zerodiag));
3049: if (benign_n == 1) {
3050: PetscCall(PetscObjectReference((PetscObject)zerodiag_subs[0]));
3051: zerodiag = zerodiag_subs[0];
3052: } else {
3053: PetscInt i, nzn, *new_idxs;
3055: nzn = 0;
3056: for (i = 0; i < benign_n; i++) {
3057: PetscInt ns;
3058: PetscCall(ISGetLocalSize(zerodiag_subs[i], &ns));
3059: nzn += ns;
3060: }
3061: PetscCall(PetscMalloc1(nzn, &new_idxs));
3062: nzn = 0;
3063: for (i = 0; i < benign_n; i++) {
3064: PetscInt ns, *idxs;
3065: PetscCall(ISGetLocalSize(zerodiag_subs[i], &ns));
3066: PetscCall(ISGetIndices(zerodiag_subs[i], (const PetscInt **)&idxs));
3067: PetscCall(PetscArraycpy(new_idxs + nzn, idxs, ns));
3068: PetscCall(ISRestoreIndices(zerodiag_subs[i], (const PetscInt **)&idxs));
3069: nzn += ns;
3070: }
3071: PetscCall(PetscSortInt(nzn, new_idxs));
3072: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, nzn, new_idxs, PETSC_OWN_POINTER, &zerodiag));
3073: }
3074: have_null = PETSC_FALSE;
3075: }
3077: /* determines if the coarse solver will be singular or not */
3078: PetscCallMPI(MPIU_Allreduce(&have_null, &pcbddc->benign_null, 1, MPI_C_BOOL, MPI_LAND, PetscObjectComm((PetscObject)pc)));
3080: /* Prepare matrix to compute no-net-flux */
3081: if (pcbddc->compute_nonetflux && !pcbddc->divudotp) {
3082: Mat A, loc_divudotp;
3083: ISLocalToGlobalMapping rl2g, cl2g, l2gmap;
3084: IS row, col, isused = NULL;
3085: PetscInt M, N, n, st, n_isused;
3087: if (pressures) {
3088: isused = pressures;
3089: } else {
3090: isused = zerodiag_save;
3091: }
3092: PetscCall(MatISGetLocalToGlobalMapping(pc->pmat, &l2gmap, NULL));
3093: PetscCall(MatISGetLocalMat(pc->pmat, &A));
3094: PetscCall(MatGetLocalSize(A, &n, NULL));
3095: PetscCheck(isused || (n == 0), PETSC_COMM_SELF, PETSC_ERR_USER, "Don't know how to extract div u dot p! Please provide the pressure field");
3096: n_isused = 0;
3097: if (isused) PetscCall(ISGetLocalSize(isused, &n_isused));
3098: PetscCallMPI(MPI_Scan(&n_isused, &st, 1, MPIU_INT, MPI_SUM, PetscObjectComm((PetscObject)pc)));
3099: st = st - n_isused;
3100: if (n) {
3101: const PetscInt *gidxs;
3103: PetscCall(MatCreateSubMatrix(A, isused, NULL, MAT_INITIAL_MATRIX, &loc_divudotp));
3104: PetscCall(ISLocalToGlobalMappingGetIndices(l2gmap, &gidxs));
3105: /* TODO: extend ISCreateStride with st = PETSC_DECIDE */
3106: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)pc), n_isused, st, 1, &row));
3107: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)pc), n, gidxs, PETSC_COPY_VALUES, &col));
3108: PetscCall(ISLocalToGlobalMappingRestoreIndices(l2gmap, &gidxs));
3109: } else {
3110: PetscCall(MatCreateSeqAIJ(PETSC_COMM_SELF, 0, 0, 1, NULL, &loc_divudotp));
3111: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)pc), n_isused, st, 1, &row));
3112: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)pc), 0, NULL, PETSC_COPY_VALUES, &col));
3113: }
3114: PetscCall(MatGetSize(pc->pmat, NULL, &N));
3115: PetscCall(ISGetSize(row, &M));
3116: PetscCall(ISLocalToGlobalMappingCreateIS(row, &rl2g));
3117: PetscCall(ISLocalToGlobalMappingCreateIS(col, &cl2g));
3118: PetscCall(ISDestroy(&row));
3119: PetscCall(ISDestroy(&col));
3120: PetscCall(MatCreate(PetscObjectComm((PetscObject)pc), &pcbddc->divudotp));
3121: PetscCall(MatSetType(pcbddc->divudotp, MATIS));
3122: PetscCall(MatSetSizes(pcbddc->divudotp, PETSC_DECIDE, PETSC_DECIDE, M, N));
3123: PetscCall(MatSetLocalToGlobalMapping(pcbddc->divudotp, rl2g, cl2g));
3124: PetscCall(ISLocalToGlobalMappingDestroy(&rl2g));
3125: PetscCall(ISLocalToGlobalMappingDestroy(&cl2g));
3126: PetscCall(MatISSetLocalMat(pcbddc->divudotp, loc_divudotp));
3127: PetscCall(MatDestroy(&loc_divudotp));
3128: PetscCall(MatAssemblyBegin(pcbddc->divudotp, MAT_FINAL_ASSEMBLY));
3129: PetscCall(MatAssemblyEnd(pcbddc->divudotp, MAT_FINAL_ASSEMBLY));
3130: }
3131: PetscCall(ISDestroy(&zerodiag_save));
3132: PetscCall(ISDestroy(&pressures));
3133: if (bzerodiag) {
3134: PetscInt i;
3136: for (i = 0; i < bsp; i++) PetscCall(ISDestroy(&bzerodiag[i]));
3137: PetscCall(PetscFree(bzerodiag));
3138: }
3139: pcbddc->benign_n = benign_n;
3140: pcbddc->benign_zerodiag_subs = zerodiag_subs;
3142: /* determines if the problem has subdomains with 0 pressure block */
3143: have_null = (PetscBool)(!!pcbddc->benign_n);
3144: PetscCallMPI(MPIU_Allreduce(&have_null, &pcbddc->benign_have_null, 1, MPI_C_BOOL, MPI_LOR, PetscObjectComm((PetscObject)pc)));
3146: project_b0:
3147: PetscCall(MatGetLocalSize(pcbddc->local_mat, &n, NULL));
3148: /* change of basis and p0 dofs */
3149: if (pcbddc->benign_n) {
3150: PetscInt i, s, *nnz;
3152: /* local change of basis for pressures */
3153: PetscCall(MatDestroy(&pcbddc->benign_change));
3154: PetscCall(MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat), &pcbddc->benign_change));
3155: PetscCall(MatSetType(pcbddc->benign_change, MATAIJ));
3156: PetscCall(MatSetSizes(pcbddc->benign_change, n, n, PETSC_DECIDE, PETSC_DECIDE));
3157: PetscCall(PetscMalloc1(n, &nnz));
3158: for (i = 0; i < n; i++) nnz[i] = 1; /* defaults to identity */
3159: for (i = 0; i < pcbddc->benign_n; i++) {
3160: const PetscInt *idxs;
3161: PetscInt nzs, j;
3163: PetscCall(ISGetLocalSize(pcbddc->benign_zerodiag_subs[i], &nzs));
3164: PetscCall(ISGetIndices(pcbddc->benign_zerodiag_subs[i], &idxs));
3165: for (j = 0; j < nzs - 1; j++) nnz[idxs[j]] = 2; /* change on pressures */
3166: nnz[idxs[nzs - 1]] = nzs; /* last local pressure dof in subdomain */
3167: PetscCall(ISRestoreIndices(pcbddc->benign_zerodiag_subs[i], &idxs));
3168: }
3169: PetscCall(MatSeqAIJSetPreallocation(pcbddc->benign_change, 0, nnz));
3170: PetscCall(MatSetOption(pcbddc->benign_change, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_TRUE));
3171: PetscCall(PetscFree(nnz));
3172: /* set identity by default */
3173: for (i = 0; i < n; i++) PetscCall(MatSetValue(pcbddc->benign_change, i, i, 1., INSERT_VALUES));
3174: PetscCall(PetscFree3(pcbddc->benign_p0_lidx, pcbddc->benign_p0_gidx, pcbddc->benign_p0));
3175: PetscCall(PetscMalloc3(pcbddc->benign_n, &pcbddc->benign_p0_lidx, pcbddc->benign_n, &pcbddc->benign_p0_gidx, pcbddc->benign_n, &pcbddc->benign_p0));
3176: /* set change on pressures */
3177: for (s = 0; s < pcbddc->benign_n; s++) {
3178: PetscScalar *array;
3179: const PetscInt *idxs;
3180: PetscInt nzs;
3182: PetscCall(ISGetLocalSize(pcbddc->benign_zerodiag_subs[s], &nzs));
3183: PetscCall(ISGetIndices(pcbddc->benign_zerodiag_subs[s], &idxs));
3184: for (i = 0; i < nzs - 1; i++) {
3185: PetscScalar vals[2];
3186: PetscInt cols[2];
3188: cols[0] = idxs[i];
3189: cols[1] = idxs[nzs - 1];
3190: vals[0] = 1.;
3191: vals[1] = 1.;
3192: PetscCall(MatSetValues(pcbddc->benign_change, 1, cols, 2, cols, vals, INSERT_VALUES));
3193: }
3194: PetscCall(PetscMalloc1(nzs, &array));
3195: for (i = 0; i < nzs - 1; i++) array[i] = -1.;
3196: array[nzs - 1] = 1.;
3197: PetscCall(MatSetValues(pcbddc->benign_change, 1, idxs + nzs - 1, nzs, idxs, array, INSERT_VALUES));
3198: /* store local idxs for p0 */
3199: pcbddc->benign_p0_lidx[s] = idxs[nzs - 1];
3200: PetscCall(ISRestoreIndices(pcbddc->benign_zerodiag_subs[s], &idxs));
3201: PetscCall(PetscFree(array));
3202: }
3203: PetscCall(MatAssemblyBegin(pcbddc->benign_change, MAT_FINAL_ASSEMBLY));
3204: PetscCall(MatAssemblyEnd(pcbddc->benign_change, MAT_FINAL_ASSEMBLY));
3206: /* project if needed */
3207: if (pcbddc->benign_change_explicit) {
3208: Mat M;
3210: PetscCall(MatPtAP(pcbddc->local_mat, pcbddc->benign_change, MAT_INITIAL_MATRIX, 2.0, &M));
3211: PetscCall(MatDestroy(&pcbddc->local_mat));
3212: PetscCall(MatSeqAIJCompress(M, &pcbddc->local_mat));
3213: PetscCall(MatDestroy(&M));
3214: }
3215: /* store global idxs for p0 */
3216: PetscCall(ISLocalToGlobalMappingApply(matis->rmapping, pcbddc->benign_n, pcbddc->benign_p0_lidx, pcbddc->benign_p0_gidx));
3217: }
3218: *zerodiaglocal = zerodiag;
3219: PetscFunctionReturn(PETSC_SUCCESS);
3220: }
3222: PetscErrorCode PCBDDCBenignGetOrSetP0(PC pc, Vec v, PetscBool get)
3223: {
3224: PC_BDDC *pcbddc = (PC_BDDC *)pc->data;
3225: PetscScalar *array;
3227: PetscFunctionBegin;
3228: if (!pcbddc->benign_sf) {
3229: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)pc), &pcbddc->benign_sf));
3230: PetscCall(PetscSFSetGraphLayout(pcbddc->benign_sf, pc->pmat->rmap, pcbddc->benign_n, NULL, PETSC_OWN_POINTER, pcbddc->benign_p0_gidx));
3231: }
3232: if (get) {
3233: PetscCall(VecGetArrayRead(v, (const PetscScalar **)&array));
3234: PetscCall(PetscSFBcastBegin(pcbddc->benign_sf, MPIU_SCALAR, array, pcbddc->benign_p0, MPI_REPLACE));
3235: PetscCall(PetscSFBcastEnd(pcbddc->benign_sf, MPIU_SCALAR, array, pcbddc->benign_p0, MPI_REPLACE));
3236: PetscCall(VecRestoreArrayRead(v, (const PetscScalar **)&array));
3237: } else {
3238: PetscCall(VecGetArray(v, &array));
3239: PetscCall(PetscSFReduceBegin(pcbddc->benign_sf, MPIU_SCALAR, pcbddc->benign_p0, array, MPI_REPLACE));
3240: PetscCall(PetscSFReduceEnd(pcbddc->benign_sf, MPIU_SCALAR, pcbddc->benign_p0, array, MPI_REPLACE));
3241: PetscCall(VecRestoreArray(v, &array));
3242: }
3243: PetscFunctionReturn(PETSC_SUCCESS);
3244: }
3246: PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop)
3247: {
3248: PC_BDDC *pcbddc = (PC_BDDC *)pc->data;
3250: PetscFunctionBegin;
3251: /* TODO: add error checking
3252: - avoid nested pop (or push) calls.
3253: - cannot push before pop.
3254: - cannot call this if pcbddc->local_mat is NULL
3255: */
3256: if (!pcbddc->benign_n) PetscFunctionReturn(PETSC_SUCCESS);
3257: if (pop) {
3258: if (pcbddc->benign_change_explicit) {
3259: IS is_p0;
3260: MatReuse reuse;
3262: /* extract B_0 */
3263: reuse = MAT_INITIAL_MATRIX;
3264: if (pcbddc->benign_B0) reuse = MAT_REUSE_MATRIX;
3265: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, pcbddc->benign_n, pcbddc->benign_p0_lidx, PETSC_COPY_VALUES, &is_p0));
3266: PetscCall(MatCreateSubMatrix(pcbddc->local_mat, is_p0, NULL, reuse, &pcbddc->benign_B0));
3267: /* remove rows and cols from local problem */
3268: PetscCall(MatSetOption(pcbddc->local_mat, MAT_KEEP_NONZERO_PATTERN, PETSC_TRUE));
3269: PetscCall(MatSetOption(pcbddc->local_mat, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_FALSE));
3270: PetscCall(MatZeroRowsColumnsIS(pcbddc->local_mat, is_p0, 1.0, NULL, NULL));
3271: PetscCall(ISDestroy(&is_p0));
3272: } else {
3273: Mat_IS *matis = (Mat_IS *)pc->pmat->data;
3274: PetscScalar *vals;
3275: PetscInt i, n, *idxs_ins;
3277: PetscCall(VecGetLocalSize(matis->y, &n));
3278: PetscCall(PetscMalloc2(n, &idxs_ins, n, &vals));
3279: if (!pcbddc->benign_B0) {
3280: PetscInt *nnz;
3281: PetscCall(MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat), &pcbddc->benign_B0));
3282: PetscCall(MatSetType(pcbddc->benign_B0, MATAIJ));
3283: PetscCall(MatSetSizes(pcbddc->benign_B0, pcbddc->benign_n, n, PETSC_DECIDE, PETSC_DECIDE));
3284: PetscCall(PetscMalloc1(pcbddc->benign_n, &nnz));
3285: for (i = 0; i < pcbddc->benign_n; i++) {
3286: PetscCall(ISGetLocalSize(pcbddc->benign_zerodiag_subs[i], &nnz[i]));
3287: nnz[i] = n - nnz[i];
3288: }
3289: PetscCall(MatSeqAIJSetPreallocation(pcbddc->benign_B0, 0, nnz));
3290: PetscCall(MatSetOption(pcbddc->benign_B0, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_TRUE));
3291: PetscCall(PetscFree(nnz));
3292: }
3294: for (i = 0; i < pcbddc->benign_n; i++) {
3295: PetscScalar *array;
3296: PetscInt *idxs, j, nz, cum;
3298: PetscCall(VecSet(matis->x, 0.));
3299: PetscCall(ISGetLocalSize(pcbddc->benign_zerodiag_subs[i], &nz));
3300: PetscCall(ISGetIndices(pcbddc->benign_zerodiag_subs[i], (const PetscInt **)&idxs));
3301: for (j = 0; j < nz; j++) vals[j] = 1.;
3302: PetscCall(VecSetValues(matis->x, nz, idxs, vals, INSERT_VALUES));
3303: PetscCall(VecAssemblyBegin(matis->x));
3304: PetscCall(VecAssemblyEnd(matis->x));
3305: PetscCall(VecSet(matis->y, 0.));
3306: PetscCall(MatMult(matis->A, matis->x, matis->y));
3307: PetscCall(VecGetArray(matis->y, &array));
3308: cum = 0;
3309: for (j = 0; j < n; j++) {
3310: if (PetscUnlikely(PetscAbsScalar(array[j]) > PETSC_SMALL)) {
3311: vals[cum] = array[j];
3312: idxs_ins[cum] = j;
3313: cum++;
3314: }
3315: }
3316: PetscCall(MatSetValues(pcbddc->benign_B0, 1, &i, cum, idxs_ins, vals, INSERT_VALUES));
3317: PetscCall(VecRestoreArray(matis->y, &array));
3318: PetscCall(ISRestoreIndices(pcbddc->benign_zerodiag_subs[i], (const PetscInt **)&idxs));
3319: }
3320: PetscCall(MatAssemblyBegin(pcbddc->benign_B0, MAT_FINAL_ASSEMBLY));
3321: PetscCall(MatAssemblyEnd(pcbddc->benign_B0, MAT_FINAL_ASSEMBLY));
3322: PetscCall(PetscFree2(idxs_ins, vals));
3323: }
3324: } else { /* push */
3326: PetscCheck(pcbddc->benign_change_explicit, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot push B0!");
3327: for (PetscInt i = 0; i < pcbddc->benign_n; i++) {
3328: PetscScalar *B0_vals;
3329: PetscInt *B0_cols, B0_ncol;
3331: PetscCall(MatGetRow(pcbddc->benign_B0, i, &B0_ncol, (const PetscInt **)&B0_cols, (const PetscScalar **)&B0_vals));
3332: PetscCall(MatSetValues(pcbddc->local_mat, 1, pcbddc->benign_p0_lidx + i, B0_ncol, B0_cols, B0_vals, INSERT_VALUES));
3333: PetscCall(MatSetValues(pcbddc->local_mat, B0_ncol, B0_cols, 1, pcbddc->benign_p0_lidx + i, B0_vals, INSERT_VALUES));
3334: PetscCall(MatSetValue(pcbddc->local_mat, pcbddc->benign_p0_lidx[i], pcbddc->benign_p0_lidx[i], 0.0, INSERT_VALUES));
3335: PetscCall(MatRestoreRow(pcbddc->benign_B0, i, &B0_ncol, (const PetscInt **)&B0_cols, (const PetscScalar **)&B0_vals));
3336: }
3337: PetscCall(MatAssemblyBegin(pcbddc->local_mat, MAT_FINAL_ASSEMBLY));
3338: PetscCall(MatAssemblyEnd(pcbddc->local_mat, MAT_FINAL_ASSEMBLY));
3339: }
3340: PetscFunctionReturn(PETSC_SUCCESS);
3341: }
3343: PetscErrorCode PCBDDCAdaptiveSelection(PC pc)
3344: {
3345: PC_BDDC *pcbddc = (PC_BDDC *)pc->data;
3346: PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs;
3347: PetscBLASInt B_neigs, B_ierr, B_lwork;
3348: PetscBLASInt *B_iwork, *B_ifail;
3349: PetscScalar *work, lwork;
3350: PetscScalar *St, *S, *eigv;
3351: PetscScalar *Sarray, *Starray;
3352: PetscReal *eigs, thresh, lthresh, uthresh;
3353: PetscInt i, nmax, nmin, nv, cum, mss, cum2, cumarray, maxneigs;
3354: PetscBool allocated_S_St, upart;
3355: #if defined(PETSC_USE_COMPLEX)
3356: PetscReal *rwork;
3357: #endif
3359: PetscFunctionBegin;
3360: if (!pcbddc->adaptive_selection) PetscFunctionReturn(PETSC_SUCCESS);
3361: PetscCheck(sub_schurs, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Adaptive selection of constraints requires SubSchurs data");
3362: PetscCheck(sub_schurs->schur_explicit || !sub_schurs->n_subs, PetscObjectComm((PetscObject)pc), PETSC_ERR_SUP, "Adaptive selection of constraints requires MUMPS and/or MKL_CPARDISO");
3363: PetscCheck(!sub_schurs->n_subs || sub_schurs->is_symmetric, PETSC_COMM_SELF, PETSC_ERR_SUP, "Adaptive selection not yet implemented for this matrix pencil (herm %d, symm %d, posdef %d)", sub_schurs->is_hermitian, sub_schurs->is_symmetric,
3364: sub_schurs->is_posdef);
3365: PetscCall(PetscLogEventBegin(PC_BDDC_AdaptiveSetUp[pcbddc->current_level], pc, 0, 0, 0));
3367: if (pcbddc->dbg_flag) {
3368: if (!pcbddc->dbg_viewer) pcbddc->dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pc));
3369: PetscCall(PetscViewerFlush(pcbddc->dbg_viewer));
3370: PetscCall(PetscViewerASCIIPrintf(pcbddc->dbg_viewer, "--------------------------------------------------\n"));
3371: PetscCall(PetscViewerASCIIPrintf(pcbddc->dbg_viewer, "Check adaptive selection of constraints\n"));
3372: PetscCall(PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer));
3373: }
3375: if (pcbddc->dbg_flag) PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, "Subdomain %04d cc %" PetscInt_FMT " (%d,%d).\n", PetscGlobalRank, sub_schurs->n_subs, sub_schurs->is_hermitian, sub_schurs->is_posdef));
3377: /* max size of subsets */
3378: mss = 0;
3379: for (i = 0; i < sub_schurs->n_subs; i++) {
3380: PetscInt subset_size;
3382: PetscCall(ISGetLocalSize(sub_schurs->is_subs[i], &subset_size));
3383: mss = PetscMax(mss, subset_size);
3384: }
3386: /* min/max and threshold */
3387: nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss;
3388: nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0;
3389: nmax = PetscMax(nmin, nmax);
3390: allocated_S_St = PETSC_FALSE;
3391: if (nmin || !sub_schurs->is_posdef) { /* XXX */
3392: allocated_S_St = PETSC_TRUE;
3393: }
3395: /* allocate lapack workspace */
3396: cum = cum2 = 0;
3397: maxneigs = 0;
3398: for (i = 0; i < sub_schurs->n_subs; i++) {
3399: PetscInt n, subset_size;
3401: PetscCall(ISGetLocalSize(sub_schurs->is_subs[i], &subset_size));
3402: n = PetscMin(subset_size, nmax);
3403: cum += subset_size;
3404: cum2 += subset_size * n;
3405: maxneigs = PetscMax(maxneigs, n);
3406: }
3407: lwork = 0;
3408: if (mss) {
3409: PetscScalar sdummy = 0.;
3410: PetscBLASInt B_itype = 1;
3411: PetscBLASInt B_N, idummy = 0;
3412: PetscReal rdummy = 0., zero = 0.0;
3413: PetscReal eps = 0.0; /* dlamch? */
3415: PetscCheck(sub_schurs->is_symmetric, PETSC_COMM_SELF, PETSC_ERR_SUP, "Not yet implemented");
3416: PetscCall(PetscBLASIntCast(mss, &B_N));
3417: B_lwork = -1;
3418: /* some implementations may complain about NULL pointers, even if we are querying */
3419: S = &sdummy;
3420: St = &sdummy;
3421: eigs = &rdummy;
3422: eigv = &sdummy;
3423: B_iwork = &idummy;
3424: B_ifail = &idummy;
3425: #if defined(PETSC_USE_COMPLEX)
3426: rwork = &rdummy;
3427: #endif
3428: thresh = 1.0;
3429: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
3430: #if defined(PETSC_USE_COMPLEX)
3431: PetscCallBLAS("LAPACKsygvx", LAPACKsygvx_(&B_itype, "V", "V", "L", &B_N, St, &B_N, S, &B_N, &zero, &thresh, B_iwork, B_iwork, &eps, &B_neigs, eigs, eigv, &B_N, &lwork, &B_lwork, rwork, B_iwork, B_ifail, &B_ierr));
3432: #else
3433: PetscCallBLAS("LAPACKsygvx", LAPACKsygvx_(&B_itype, "V", "V", "L", &B_N, St, &B_N, S, &B_N, &zero, &thresh, B_iwork, B_iwork, &eps, &B_neigs, eigs, eigv, &B_N, &lwork, &B_lwork, B_iwork, B_ifail, &B_ierr));
3434: #endif
3435: PetscCheck(B_ierr == 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "Error in query to SYGVX Lapack routine %" PetscBLASInt_FMT, B_ierr);
3436: PetscCall(PetscFPTrapPop());
3437: }
3439: nv = 0;
3440: if (sub_schurs->is_vertices && pcbddc->use_vertices) { /* complement set of active subsets, each entry is a vertex (boundary made by active subsets, vertices and dirichlet dofs) */
3441: PetscCall(ISGetLocalSize(sub_schurs->is_vertices, &nv));
3442: }
3443: PetscCall(PetscBLASIntCast((PetscInt)PetscRealPart(lwork), &B_lwork));
3444: if (allocated_S_St) PetscCall(PetscMalloc2(mss * mss, &S, mss * mss, &St));
3445: PetscCall(PetscMalloc5(mss * mss, &eigv, mss, &eigs, B_lwork, &work, 5 * mss, &B_iwork, mss, &B_ifail));
3446: #if defined(PETSC_USE_COMPLEX)
3447: PetscCall(PetscMalloc1(7 * mss, &rwork));
3448: #endif
3449: PetscCall(PetscMalloc5(nv + sub_schurs->n_subs, &pcbddc->adaptive_constraints_n, nv + sub_schurs->n_subs + 1, &pcbddc->adaptive_constraints_idxs_ptr, nv + sub_schurs->n_subs + 1, &pcbddc->adaptive_constraints_data_ptr, nv + cum, &pcbddc->adaptive_constraints_idxs, nv + cum2,
3450: &pcbddc->adaptive_constraints_data));
3451: PetscCall(PetscArrayzero(pcbddc->adaptive_constraints_n, nv + sub_schurs->n_subs));
3453: maxneigs = 0;
3454: cum = cumarray = 0;
3455: pcbddc->adaptive_constraints_idxs_ptr[0] = 0;
3456: pcbddc->adaptive_constraints_data_ptr[0] = 0;
3457: if (sub_schurs->is_vertices && pcbddc->use_vertices) {
3458: const PetscInt *idxs;
3460: PetscCall(ISGetIndices(sub_schurs->is_vertices, &idxs));
3461: for (cum = 0; cum < nv; cum++) {
3462: pcbddc->adaptive_constraints_n[cum] = 1;
3463: pcbddc->adaptive_constraints_idxs[cum] = idxs[cum];
3464: pcbddc->adaptive_constraints_data[cum] = 1.0;
3465: pcbddc->adaptive_constraints_idxs_ptr[cum + 1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + 1;
3466: pcbddc->adaptive_constraints_data_ptr[cum + 1] = pcbddc->adaptive_constraints_data_ptr[cum] + 1;
3467: }
3468: PetscCall(ISRestoreIndices(sub_schurs->is_vertices, &idxs));
3469: }
3471: if (mss) { /* multilevel */
3472: if (sub_schurs->gdsw) {
3473: PetscCall(MatSeqAIJGetArray(sub_schurs->sum_S_Ej_all, &Sarray));
3474: PetscCall(MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all, &Starray));
3475: } else {
3476: PetscCall(MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all, &Sarray));
3477: PetscCall(MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all, &Starray));
3478: }
3479: }
3481: lthresh = pcbddc->adaptive_threshold[0];
3482: uthresh = pcbddc->adaptive_threshold[1];
3483: upart = pcbddc->use_deluxe_scaling;
3484: for (i = 0; i < sub_schurs->n_subs; i++) {
3485: const PetscInt *idxs;
3486: PetscReal upper, lower;
3487: PetscInt j, subset_size, eigs_start = 0;
3488: PetscBLASInt B_N;
3489: PetscBool same_data = PETSC_FALSE;
3490: PetscBool scal = PETSC_FALSE;
3492: if (upart) {
3493: upper = PETSC_MAX_REAL;
3494: lower = uthresh;
3495: } else {
3496: if (sub_schurs->gdsw) {
3497: upper = uthresh;
3498: lower = PETSC_MIN_REAL;
3499: } else {
3500: PetscCheck(sub_schurs->is_posdef, PETSC_COMM_SELF, PETSC_ERR_SUP, "Not yet implemented without deluxe scaling");
3501: upper = 1. / uthresh;
3502: lower = 0.;
3503: }
3504: }
3505: PetscCall(ISGetLocalSize(sub_schurs->is_subs[i], &subset_size));
3506: PetscCall(ISGetIndices(sub_schurs->is_subs[i], &idxs));
3507: PetscCall(PetscBLASIntCast(subset_size, &B_N));
3508: /* this is experimental: we assume the dofs have been properly grouped to have
3509: the diagonal blocks Schur complements either positive or negative definite (true for Stokes) */
3510: if (!sub_schurs->is_posdef) {
3511: Mat T;
3513: for (j = 0; j < subset_size; j++) {
3514: if (PetscRealPart(*(Sarray + cumarray + j * (subset_size + 1))) < 0.0) {
3515: PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, subset_size, subset_size, Sarray + cumarray, &T));
3516: PetscCall(MatScale(T, -1.0));
3517: PetscCall(MatDestroy(&T));
3518: PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, subset_size, subset_size, Starray + cumarray, &T));
3519: PetscCall(MatScale(T, -1.0));
3520: PetscCall(MatDestroy(&T));
3521: if (sub_schurs->change_primal_sub) {
3522: PetscInt nz, k;
3523: const PetscInt *idxs;
3525: PetscCall(ISGetLocalSize(sub_schurs->change_primal_sub[i], &nz));
3526: PetscCall(ISGetIndices(sub_schurs->change_primal_sub[i], &idxs));
3527: for (k = 0; k < nz; k++) {
3528: *(Sarray + cumarray + idxs[k] * (subset_size + 1)) *= -1.0;
3529: *(Starray + cumarray + idxs[k] * (subset_size + 1)) = 0.0;
3530: }
3531: PetscCall(ISRestoreIndices(sub_schurs->change_primal_sub[i], &idxs));
3532: }
3533: scal = PETSC_TRUE;
3534: break;
3535: }
3536: }
3537: }
3539: if (allocated_S_St) { /* S and S_t should be copied since we could need them later */
3540: if (sub_schurs->is_symmetric) {
3541: PetscInt j, k;
3542: if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscArraycmp() later */
3543: PetscCall(PetscArrayzero(S, subset_size * subset_size));
3544: PetscCall(PetscArrayzero(St, subset_size * subset_size));
3545: }
3546: for (j = 0; j < subset_size; j++) {
3547: for (k = j; k < subset_size; k++) {
3548: S[j * subset_size + k] = Sarray[cumarray + j * subset_size + k];
3549: St[j * subset_size + k] = Starray[cumarray + j * subset_size + k];
3550: }
3551: }
3552: } else {
3553: PetscCall(PetscArraycpy(S, Sarray + cumarray, subset_size * subset_size));
3554: PetscCall(PetscArraycpy(St, Starray + cumarray, subset_size * subset_size));
3555: }
3556: } else {
3557: S = Sarray + cumarray;
3558: St = Starray + cumarray;
3559: }
3560: /* see if we can save some work */
3561: if (sub_schurs->n_subs == 1 && pcbddc->use_deluxe_scaling) PetscCall(PetscArraycmp(S, St, subset_size * subset_size, &same_data));
3563: if (same_data && !sub_schurs->change) { /* there's no need of constraints here */
3564: B_neigs = 0;
3565: } else {
3566: PetscBLASInt B_itype = 1, B_IL = 1, B_IU = 0;
3567: PetscReal eps = -1.0; /* dlamch? */
3568: PetscInt nmin_s;
3569: PetscBool compute_range;
3571: PetscCheck(sub_schurs->is_symmetric, PETSC_COMM_SELF, PETSC_ERR_SUP, "Not yet implemented");
3572: B_neigs = 0;
3573: compute_range = (PetscBool)!same_data;
3574: if (nmin >= subset_size) compute_range = PETSC_FALSE;
3576: if (pcbddc->dbg_flag) {
3577: PetscInt nc = 0, c = pcbddc->mat_graph->nodes[idxs[0]].count, w = pcbddc->mat_graph->nodes[idxs[0]].which_dof;
3579: if (sub_schurs->change_primal_sub) PetscCall(ISGetLocalSize(sub_schurs->change_primal_sub[i], &nc));
3580: PetscCall(
3581: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, "Computing for sub %" PetscInt_FMT "/%" PetscInt_FMT " size %" PetscInt_FMT " count %" PetscInt_FMT " fid %" PetscInt_FMT " (range %d) (change %" PetscInt_FMT ").\n", i, sub_schurs->n_subs, subset_size, c, w, compute_range, nc));
3582: }
3584: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
3585: if (compute_range) {
3586: /* ask for eigenvalues larger than thresh */
3587: if (sub_schurs->is_posdef) {
3588: #if defined(PETSC_USE_COMPLEX)
3589: PetscCallBLAS("LAPACKsygvx", LAPACKsygvx_(&B_itype, "V", "V", "L", &B_N, St, &B_N, S, &B_N, &lower, &upper, &B_IL, &B_IU, &eps, &B_neigs, eigs, eigv, &B_N, work, &B_lwork, rwork, B_iwork, B_ifail, &B_ierr));
3590: #else
3591: PetscCallBLAS("LAPACKsygvx", LAPACKsygvx_(&B_itype, "V", "V", "L", &B_N, St, &B_N, S, &B_N, &lower, &upper, &B_IL, &B_IU, &eps, &B_neigs, eigs, eigv, &B_N, work, &B_lwork, B_iwork, B_ifail, &B_ierr));
3592: #endif
3593: PetscCall(PetscLogFlops((4.0 * subset_size * subset_size * subset_size) / 3.0));
3594: } else { /* no theory so far, but it works nicely */
3595: PetscInt recipe = 0, recipe_m = 1;
3596: PetscReal bb[2];
3598: PetscCall(PetscOptionsGetInt(NULL, ((PetscObject)pc)->prefix, "-pc_bddc_adaptive_recipe", &recipe, NULL));
3599: switch (recipe) {
3600: case 0:
3601: if (scal) {
3602: bb[0] = PETSC_MIN_REAL;
3603: bb[1] = lthresh;
3604: } else {
3605: bb[0] = uthresh;
3606: bb[1] = PETSC_MAX_REAL;
3607: }
3608: #if defined(PETSC_USE_COMPLEX)
3609: PetscCallBLAS("LAPACKsygvx", LAPACKsygvx_(&B_itype, "V", "V", "L", &B_N, St, &B_N, S, &B_N, &bb[0], &bb[1], &B_IL, &B_IU, &eps, &B_neigs, eigs, eigv, &B_N, work, &B_lwork, rwork, B_iwork, B_ifail, &B_ierr));
3610: #else
3611: PetscCallBLAS("LAPACKsygvx", LAPACKsygvx_(&B_itype, "V", "V", "L", &B_N, St, &B_N, S, &B_N, &bb[0], &bb[1], &B_IL, &B_IU, &eps, &B_neigs, eigs, eigv, &B_N, work, &B_lwork, B_iwork, B_ifail, &B_ierr));
3612: #endif
3613: PetscCall(PetscLogFlops((4.0 * subset_size * subset_size * subset_size) / 3.0));
3614: break;
3615: case 1:
3616: bb[0] = PETSC_MIN_REAL;
3617: bb[1] = lthresh * lthresh;
3618: #if defined(PETSC_USE_COMPLEX)
3619: PetscCallBLAS("LAPACKsygvx", LAPACKsygvx_(&B_itype, "V", "V", "L", &B_N, St, &B_N, S, &B_N, &bb[0], &bb[1], &B_IL, &B_IU, &eps, &B_neigs, eigs, eigv, &B_N, work, &B_lwork, rwork, B_iwork, B_ifail, &B_ierr));
3620: #else
3621: PetscCallBLAS("LAPACKsygvx", LAPACKsygvx_(&B_itype, "V", "V", "L", &B_N, St, &B_N, S, &B_N, &bb[0], &bb[1], &B_IL, &B_IU, &eps, &B_neigs, eigs, eigv, &B_N, work, &B_lwork, B_iwork, B_ifail, &B_ierr));
3622: #endif
3623: PetscCall(PetscLogFlops((4.0 * subset_size * subset_size * subset_size) / 3.0));
3624: if (!scal) {
3625: PetscBLASInt B_neigs2 = 0;
3627: bb[0] = PetscMax(lthresh * lthresh, uthresh);
3628: bb[1] = PETSC_MAX_REAL;
3629: PetscCall(PetscArraycpy(S, Sarray + cumarray, subset_size * subset_size));
3630: PetscCall(PetscArraycpy(St, Starray + cumarray, subset_size * subset_size));
3631: #if defined(PETSC_USE_COMPLEX)
3632: PetscCallBLAS("LAPACKsygvx", LAPACKsygvx_(&B_itype, "V", "V", "L", &B_N, St, &B_N, S, &B_N, &bb[0], &bb[1], &B_IL, &B_IU, &eps, &B_neigs2, eigs + B_neigs, eigv + B_neigs * B_N, &B_N, work, &B_lwork, rwork, B_iwork, B_ifail, &B_ierr));
3633: #else
3634: PetscCallBLAS("LAPACKsygvx", LAPACKsygvx_(&B_itype, "V", "V", "L", &B_N, St, &B_N, S, &B_N, &bb[0], &bb[1], &B_IL, &B_IU, &eps, &B_neigs2, eigs + B_neigs, eigv + B_neigs * B_N, &B_N, work, &B_lwork, B_iwork, B_ifail, &B_ierr));
3635: #endif
3636: PetscCall(PetscLogFlops((4.0 * subset_size * subset_size * subset_size) / 3.0));
3637: B_neigs += B_neigs2;
3638: }
3639: break;
3640: case 2:
3641: if (scal) {
3642: bb[0] = PETSC_MIN_REAL;
3643: bb[1] = 0;
3644: #if defined(PETSC_USE_COMPLEX)
3645: PetscCallBLAS("LAPACKsygvx", LAPACKsygvx_(&B_itype, "V", "V", "L", &B_N, St, &B_N, S, &B_N, &bb[0], &bb[1], &B_IL, &B_IU, &eps, &B_neigs, eigs, eigv, &B_N, work, &B_lwork, rwork, B_iwork, B_ifail, &B_ierr));
3646: #else
3647: PetscCallBLAS("LAPACKsygvx", LAPACKsygvx_(&B_itype, "V", "V", "L", &B_N, St, &B_N, S, &B_N, &bb[0], &bb[1], &B_IL, &B_IU, &eps, &B_neigs, eigs, eigv, &B_N, work, &B_lwork, B_iwork, B_ifail, &B_ierr));
3648: #endif
3649: PetscCall(PetscLogFlops((4.0 * subset_size * subset_size * subset_size) / 3.0));
3650: } else {
3651: PetscBLASInt B_neigs2 = 0;
3652: PetscBool do_copy = PETSC_FALSE;
3654: lthresh = PetscMax(lthresh, 0.0);
3655: if (lthresh > 0.0) {
3656: bb[0] = PETSC_MIN_REAL;
3657: bb[1] = lthresh * lthresh;
3659: do_copy = PETSC_TRUE;
3660: #if defined(PETSC_USE_COMPLEX)
3661: PetscCallBLAS("LAPACKsygvx", LAPACKsygvx_(&B_itype, "V", "V", "L", &B_N, St, &B_N, S, &B_N, &bb[0], &bb[1], &B_IL, &B_IU, &eps, &B_neigs, eigs, eigv, &B_N, work, &B_lwork, rwork, B_iwork, B_ifail, &B_ierr));
3662: #else
3663: PetscCallBLAS("LAPACKsygvx", LAPACKsygvx_(&B_itype, "V", "V", "L", &B_N, St, &B_N, S, &B_N, &bb[0], &bb[1], &B_IL, &B_IU, &eps, &B_neigs, eigs, eigv, &B_N, work, &B_lwork, B_iwork, B_ifail, &B_ierr));
3664: #endif
3665: PetscCall(PetscLogFlops((4.0 * subset_size * subset_size * subset_size) / 3.0));
3666: }
3667: bb[0] = PetscMax(lthresh * lthresh, uthresh);
3668: bb[1] = PETSC_MAX_REAL;
3669: if (do_copy) {
3670: PetscCall(PetscArraycpy(S, Sarray + cumarray, subset_size * subset_size));
3671: PetscCall(PetscArraycpy(St, Starray + cumarray, subset_size * subset_size));
3672: }
3673: #if defined(PETSC_USE_COMPLEX)
3674: PetscCallBLAS("LAPACKsygvx", LAPACKsygvx_(&B_itype, "V", "V", "L", &B_N, St, &B_N, S, &B_N, &bb[0], &bb[1], &B_IL, &B_IU, &eps, &B_neigs2, eigs + B_neigs, eigv + B_neigs * B_N, &B_N, work, &B_lwork, rwork, B_iwork, B_ifail, &B_ierr));
3675: #else
3676: PetscCallBLAS("LAPACKsygvx", LAPACKsygvx_(&B_itype, "V", "V", "L", &B_N, St, &B_N, S, &B_N, &bb[0], &bb[1], &B_IL, &B_IU, &eps, &B_neigs2, eigs + B_neigs, eigv + B_neigs * B_N, &B_N, work, &B_lwork, B_iwork, B_ifail, &B_ierr));
3677: #endif
3678: PetscCall(PetscLogFlops((4.0 * subset_size * subset_size * subset_size) / 3.0));
3679: B_neigs += B_neigs2;
3680: }
3681: break;
3682: case 3:
3683: if (scal) {
3684: PetscCall(PetscOptionsGetInt(NULL, ((PetscObject)pc)->prefix, "-pc_bddc_adaptive_recipe3_min_scal", &recipe_m, NULL));
3685: } else {
3686: PetscCall(PetscOptionsGetInt(NULL, ((PetscObject)pc)->prefix, "-pc_bddc_adaptive_recipe3_min", &recipe_m, NULL));
3687: }
3688: if (!scal) {
3689: bb[0] = uthresh;
3690: bb[1] = PETSC_MAX_REAL;
3691: #if defined(PETSC_USE_COMPLEX)
3692: PetscCallBLAS("LAPACKsygvx", LAPACKsygvx_(&B_itype, "V", "V", "L", &B_N, St, &B_N, S, &B_N, &bb[0], &bb[1], &B_IL, &B_IU, &eps, &B_neigs, eigs, eigv, &B_N, work, &B_lwork, rwork, B_iwork, B_ifail, &B_ierr));
3693: #else
3694: PetscCallBLAS("LAPACKsygvx", LAPACKsygvx_(&B_itype, "V", "V", "L", &B_N, St, &B_N, S, &B_N, &bb[0], &bb[1], &B_IL, &B_IU, &eps, &B_neigs, eigs, eigv, &B_N, work, &B_lwork, B_iwork, B_ifail, &B_ierr));
3695: #endif
3696: PetscCall(PetscLogFlops((4.0 * subset_size * subset_size * subset_size) / 3.0));
3697: }
3698: if (recipe_m > 0 && B_N - B_neigs > 0) {
3699: PetscBLASInt B_neigs2 = 0;
3701: PetscCall(PetscBLASIntCast(PetscMin(recipe_m, B_N - B_neigs), &B_IU));
3702: PetscCall(PetscArraycpy(S, Sarray + cumarray, subset_size * subset_size));
3703: PetscCall(PetscArraycpy(St, Starray + cumarray, subset_size * subset_size));
3704: #if defined(PETSC_USE_COMPLEX)
3705: PetscCallBLAS("LAPACKsygvx", LAPACKsygvx_(&B_itype, "V", "I", "L", &B_N, St, &B_N, S, &B_N, &lower, &upper, &B_IL, &B_IU, &eps, &B_neigs2, eigs + B_neigs, eigv + B_neigs * B_N, &B_N, work, &B_lwork, rwork, B_iwork, B_ifail, &B_ierr));
3706: #else
3707: PetscCallBLAS("LAPACKsygvx", LAPACKsygvx_(&B_itype, "V", "I", "L", &B_N, St, &B_N, S, &B_N, &lower, &upper, &B_IL, &B_IU, &eps, &B_neigs2, eigs + B_neigs, eigv + B_neigs * B_N, &B_N, work, &B_lwork, B_iwork, B_ifail, &B_ierr));
3708: #endif
3709: PetscCall(PetscLogFlops((4.0 * subset_size * subset_size * subset_size) / 3.0));
3710: B_neigs += B_neigs2;
3711: }
3712: break;
3713: case 4:
3714: bb[0] = PETSC_MIN_REAL;
3715: bb[1] = lthresh;
3716: #if defined(PETSC_USE_COMPLEX)
3717: PetscCallBLAS("LAPACKsygvx", LAPACKsygvx_(&B_itype, "V", "V", "L", &B_N, St, &B_N, S, &B_N, &bb[0], &bb[1], &B_IL, &B_IU, &eps, &B_neigs, eigs, eigv, &B_N, work, &B_lwork, rwork, B_iwork, B_ifail, &B_ierr));
3718: #else
3719: PetscCallBLAS("LAPACKsygvx", LAPACKsygvx_(&B_itype, "V", "V", "L", &B_N, St, &B_N, S, &B_N, &bb[0], &bb[1], &B_IL, &B_IU, &eps, &B_neigs, eigs, eigv, &B_N, work, &B_lwork, B_iwork, B_ifail, &B_ierr));
3720: #endif
3721: PetscCall(PetscLogFlops((4.0 * subset_size * subset_size * subset_size) / 3.0));
3722: {
3723: PetscBLASInt B_neigs2 = 0;
3725: bb[0] = PetscMax(lthresh + PETSC_SMALL, uthresh);
3726: bb[1] = PETSC_MAX_REAL;
3727: PetscCall(PetscArraycpy(S, Sarray + cumarray, subset_size * subset_size));
3728: PetscCall(PetscArraycpy(St, Starray + cumarray, subset_size * subset_size));
3729: #if defined(PETSC_USE_COMPLEX)
3730: PetscCallBLAS("LAPACKsygvx", LAPACKsygvx_(&B_itype, "V", "V", "L", &B_N, St, &B_N, S, &B_N, &bb[0], &bb[1], &B_IL, &B_IU, &eps, &B_neigs2, eigs + B_neigs, eigv + B_neigs * B_N, &B_N, work, &B_lwork, rwork, B_iwork, B_ifail, &B_ierr));
3731: #else
3732: PetscCallBLAS("LAPACKsygvx", LAPACKsygvx_(&B_itype, "V", "V", "L", &B_N, St, &B_N, S, &B_N, &bb[0], &bb[1], &B_IL, &B_IU, &eps, &B_neigs2, eigs + B_neigs, eigv + B_neigs * B_N, &B_N, work, &B_lwork, B_iwork, B_ifail, &B_ierr));
3733: #endif
3734: PetscCall(PetscLogFlops((4.0 * subset_size * subset_size * subset_size) / 3.0));
3735: B_neigs += B_neigs2;
3736: }
3737: break;
3738: case 5: /* same as before: first compute all eigenvalues, then filter */
3739: #if defined(PETSC_USE_COMPLEX)
3740: PetscCallBLAS("LAPACKsygvx", LAPACKsygvx_(&B_itype, "V", "A", "L", &B_N, St, &B_N, S, &B_N, &bb[0], &bb[1], &B_IL, &B_IU, &eps, &B_neigs, eigs, eigv, &B_N, work, &B_lwork, rwork, B_iwork, B_ifail, &B_ierr));
3741: #else
3742: PetscCallBLAS("LAPACKsygvx", LAPACKsygvx_(&B_itype, "V", "A", "L", &B_N, St, &B_N, S, &B_N, &bb[0], &bb[1], &B_IL, &B_IU, &eps, &B_neigs, eigs, eigv, &B_N, work, &B_lwork, B_iwork, B_ifail, &B_ierr));
3743: #endif
3744: PetscCall(PetscLogFlops((4.0 * subset_size * subset_size * subset_size) / 3.0));
3745: {
3746: PetscInt e, k, ne;
3747: for (e = 0, ne = 0; e < B_neigs; e++) {
3748: if (eigs[e] < lthresh || eigs[e] > uthresh) {
3749: for (k = 0; k < B_N; k++) S[ne * B_N + k] = eigv[e * B_N + k];
3750: eigs[ne] = eigs[e];
3751: ne++;
3752: }
3753: }
3754: PetscCall(PetscArraycpy(eigv, S, B_N * ne));
3755: PetscCall(PetscBLASIntCast(ne, &B_neigs));
3756: }
3757: break;
3758: default:
3759: SETERRQ(PetscObjectComm((PetscObject)pc), PETSC_ERR_SUP, "Unknown recipe %" PetscInt_FMT, recipe);
3760: }
3761: }
3762: } else if (!same_data) { /* this is just to see all the eigenvalues */
3763: PetscCall(PetscBLASIntCast(PetscMax(1, PetscMin(B_N, nmax)), &B_IU));
3764: #if defined(PETSC_USE_COMPLEX)
3765: PetscCallBLAS("LAPACKsygvx", LAPACKsygvx_(&B_itype, "V", "I", "L", &B_N, St, &B_N, S, &B_N, &lower, &upper, &B_IL, &B_IU, &eps, &B_neigs, eigs, eigv, &B_N, work, &B_lwork, rwork, B_iwork, B_ifail, &B_ierr));
3766: #else
3767: PetscCallBLAS("LAPACKsygvx", LAPACKsygvx_(&B_itype, "V", "I", "L", &B_N, St, &B_N, S, &B_N, &lower, &upper, &B_IL, &B_IU, &eps, &B_neigs, eigs, eigv, &B_N, work, &B_lwork, B_iwork, B_ifail, &B_ierr));
3768: #endif
3769: PetscCall(PetscLogFlops((4.0 * subset_size * subset_size * subset_size) / 3.0));
3770: } else { /* same_data is true, so just get the adaptive functional requested by the user */
3771: PetscInt k;
3772: PetscCheck(sub_schurs->change_primal_sub, PETSC_COMM_SELF, PETSC_ERR_PLIB, "This should not happen");
3773: PetscCall(ISGetLocalSize(sub_schurs->change_primal_sub[i], &nmax));
3774: PetscCall(PetscBLASIntCast(nmax, &B_neigs));
3775: nmin = nmax;
3776: PetscCall(PetscArrayzero(eigv, subset_size * nmax));
3777: for (k = 0; k < nmax; k++) {
3778: eigs[k] = 1. / PETSC_SMALL;
3779: eigv[k * (subset_size + 1)] = 1.0;
3780: }
3781: }
3782: PetscCall(PetscFPTrapPop());
3783: if (B_ierr) {
3784: PetscCheck(B_ierr >= 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "Error in SYGVX Lapack routine: illegal value for argument %" PetscBLASInt_FMT, -B_ierr);
3785: PetscCheck(B_ierr > B_N, PETSC_COMM_SELF, PETSC_ERR_LIB, "Error in SYGVX Lapack routine: %" PetscBLASInt_FMT " eigenvalues failed to converge", B_ierr);
3786: SETERRQ(PETSC_COMM_SELF, PETSC_ERR_LIB, "Error in SYGVX Lapack routine: leading minor of order %" PetscBLASInt_FMT " is not positive definite", B_ierr - B_N - 1);
3787: }
3789: if (B_neigs > nmax) {
3790: if (pcbddc->dbg_flag) PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, " found %" PetscBLASInt_FMT " eigs, more than maximum required %" PetscInt_FMT ".\n", B_neigs, nmax));
3791: if (upart) eigs_start = scal ? 0 : B_neigs - nmax;
3792: PetscCall(PetscBLASIntCast(nmax, &B_neigs));
3793: }
3795: nmin_s = PetscMin(nmin, B_N);
3796: if (B_neigs < nmin_s) {
3797: PetscBLASInt B_neigs2 = 0;
3799: if (upart) {
3800: if (scal) {
3801: PetscCall(PetscBLASIntCast(nmin_s, &B_IU));
3802: B_IL = B_neigs + 1;
3803: } else {
3804: PetscCall(PetscBLASIntCast(B_N - nmin_s + 1, &B_IL));
3805: B_IU = B_N - B_neigs;
3806: }
3807: } else {
3808: B_IL = B_neigs + 1;
3809: PetscCall(PetscBLASIntCast(nmin_s, &B_IU));
3810: }
3811: if (pcbddc->dbg_flag) {
3812: PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, " found %" PetscBLASInt_FMT " eigs, less than minimum required %" PetscInt_FMT ". Asking for %" PetscBLASInt_FMT " to %" PetscBLASInt_FMT " incl (fortran like)\n", B_neigs, nmin, B_IL, B_IU));
3813: }
3814: if (sub_schurs->is_symmetric) {
3815: PetscInt j, k;
3816: for (j = 0; j < subset_size; j++) {
3817: for (k = j; k < subset_size; k++) {
3818: S[j * subset_size + k] = Sarray[cumarray + j * subset_size + k];
3819: St[j * subset_size + k] = Starray[cumarray + j * subset_size + k];
3820: }
3821: }
3822: } else {
3823: PetscCall(PetscArraycpy(S, Sarray + cumarray, subset_size * subset_size));
3824: PetscCall(PetscArraycpy(St, Starray + cumarray, subset_size * subset_size));
3825: }
3826: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
3827: #if defined(PETSC_USE_COMPLEX)
3828: PetscCallBLAS("LAPACKsygvx", LAPACKsygvx_(&B_itype, "V", "I", "L", &B_N, St, &B_N, S, &B_N, &lower, &upper, &B_IL, &B_IU, &eps, &B_neigs2, eigs + B_neigs, eigv + B_neigs * subset_size, &B_N, work, &B_lwork, rwork, B_iwork, B_ifail, &B_ierr));
3829: #else
3830: PetscCallBLAS("LAPACKsygvx", LAPACKsygvx_(&B_itype, "V", "I", "L", &B_N, St, &B_N, S, &B_N, &lower, &upper, &B_IL, &B_IU, &eps, &B_neigs2, eigs + B_neigs, eigv + B_neigs * subset_size, &B_N, work, &B_lwork, B_iwork, B_ifail, &B_ierr));
3831: #endif
3832: PetscCall(PetscLogFlops((4.0 * subset_size * subset_size * subset_size) / 3.0));
3833: PetscCall(PetscFPTrapPop());
3834: B_neigs += B_neigs2;
3835: }
3836: if (B_ierr) {
3837: PetscCheck(B_ierr >= 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "Error in SYGVX Lapack routine: illegal value for argument %" PetscBLASInt_FMT, -B_ierr);
3838: PetscCheck(B_ierr > B_N, PETSC_COMM_SELF, PETSC_ERR_LIB, "Error in SYGVX Lapack routine: %" PetscBLASInt_FMT " eigenvalues failed to converge", B_ierr);
3839: SETERRQ(PETSC_COMM_SELF, PETSC_ERR_LIB, "Error in SYGVX Lapack routine: leading minor of order %" PetscBLASInt_FMT " is not positive definite", B_ierr - B_N - 1);
3840: }
3841: if (pcbddc->dbg_flag) {
3842: PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, " -> Got %" PetscBLASInt_FMT " eigs\n", B_neigs));
3843: for (j = 0; j < B_neigs; j++) {
3844: if (!sub_schurs->gdsw) {
3845: if (eigs[j] == 0.0) {
3846: PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, " Inf\n"));
3847: } else {
3848: if (upart) {
3849: PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, " %1.6e\n", (double)eigs[j + eigs_start]));
3850: } else {
3851: PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, " %1.6e\n", (double)(1 / eigs[j + eigs_start])));
3852: }
3853: }
3854: } else {
3855: double pg = (double)eigs[j + eigs_start];
3856: if (pg < 2 * PETSC_SMALL) pg = 0.0;
3857: PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, " %1.6e\n", pg));
3858: }
3859: }
3860: }
3861: }
3862: /* change the basis back to the original one */
3863: if (sub_schurs->change) {
3864: Mat change, phi, phit;
3866: if (pcbddc->dbg_flag > 2) {
3867: PetscInt ii;
3868: for (ii = 0; ii < B_neigs; ii++) {
3869: PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, " -> Eigenvector (old basis) %" PetscInt_FMT "/%" PetscBLASInt_FMT " (%" PetscBLASInt_FMT ")\n", ii, B_neigs, B_N));
3870: for (j = 0; j < B_N; j++) {
3871: #if defined(PETSC_USE_COMPLEX)
3872: PetscReal r = PetscRealPart(eigv[(ii + eigs_start) * subset_size + j]);
3873: PetscReal c = PetscImaginaryPart(eigv[(ii + eigs_start) * subset_size + j]);
3874: PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, " %1.4e + %1.4e i\n", (double)r, (double)c));
3875: #else
3876: PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, " %1.4e\n", (double)(eigv[(ii + eigs_start) * subset_size + j])));
3877: #endif
3878: }
3879: }
3880: }
3881: PetscCall(KSPGetOperators(sub_schurs->change[i], &change, NULL));
3882: PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, subset_size, B_neigs, eigv + eigs_start * subset_size, &phit));
3883: PetscCall(MatMatMult(change, phit, MAT_INITIAL_MATRIX, PETSC_DETERMINE, &phi));
3884: PetscCall(MatCopy(phi, phit, SAME_NONZERO_PATTERN));
3885: PetscCall(MatDestroy(&phit));
3886: PetscCall(MatDestroy(&phi));
3887: }
3888: maxneigs = PetscMax(B_neigs, maxneigs);
3889: pcbddc->adaptive_constraints_n[i + nv] = B_neigs;
3890: if (B_neigs) {
3891: PetscCall(PetscArraycpy(pcbddc->adaptive_constraints_data + pcbddc->adaptive_constraints_data_ptr[cum], eigv + eigs_start * subset_size, B_neigs * subset_size));
3893: if (pcbddc->dbg_flag > 1) {
3894: PetscInt ii;
3895: for (ii = 0; ii < B_neigs; ii++) {
3896: PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, " -> Eigenvector %" PetscInt_FMT "/%" PetscBLASInt_FMT " (%" PetscBLASInt_FMT ")\n", ii, B_neigs, B_N));
3897: for (j = 0; j < B_N; j++) {
3898: #if defined(PETSC_USE_COMPLEX)
3899: PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii * subset_size + j + pcbddc->adaptive_constraints_data_ptr[cum]]);
3900: PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii * subset_size + j + pcbddc->adaptive_constraints_data_ptr[cum]]);
3901: PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, " %1.4e + %1.4e i\n", (double)r, (double)c));
3902: #else
3903: PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, " %1.4e\n", (double)PetscRealPart(pcbddc->adaptive_constraints_data[ii * subset_size + j + pcbddc->adaptive_constraints_data_ptr[cum]])));
3904: #endif
3905: }
3906: }
3907: }
3908: PetscCall(PetscArraycpy(pcbddc->adaptive_constraints_idxs + pcbddc->adaptive_constraints_idxs_ptr[cum], idxs, subset_size));
3909: pcbddc->adaptive_constraints_idxs_ptr[cum + 1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size;
3910: pcbddc->adaptive_constraints_data_ptr[cum + 1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size * B_neigs;
3911: cum++;
3912: }
3913: PetscCall(ISRestoreIndices(sub_schurs->is_subs[i], &idxs));
3914: /* shift for next computation */
3915: cumarray += subset_size * subset_size;
3916: }
3917: if (pcbddc->dbg_flag) PetscCall(PetscViewerFlush(pcbddc->dbg_viewer));
3919: if (mss) {
3920: if (sub_schurs->gdsw) {
3921: PetscCall(MatSeqAIJGetArray(sub_schurs->sum_S_Ej_all, &Sarray));
3922: PetscCall(MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all, &Starray));
3923: } else {
3924: PetscCall(MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all, &Sarray));
3925: PetscCall(MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all, &Starray));
3926: /* destroy matrices (junk) */
3927: PetscCall(MatDestroy(&sub_schurs->sum_S_Ej_inv_all));
3928: PetscCall(MatDestroy(&sub_schurs->sum_S_Ej_tilda_all));
3929: }
3930: }
3931: if (allocated_S_St) PetscCall(PetscFree2(S, St));
3932: PetscCall(PetscFree5(eigv, eigs, work, B_iwork, B_ifail));
3933: #if defined(PETSC_USE_COMPLEX)
3934: PetscCall(PetscFree(rwork));
3935: #endif
3936: if (pcbddc->dbg_flag) {
3937: PetscInt maxneigs_r;
3938: PetscCallMPI(MPIU_Allreduce(&maxneigs, &maxneigs_r, 1, MPIU_INT, MPI_MAX, PetscObjectComm((PetscObject)pc)));
3939: PetscCall(PetscViewerASCIIPrintf(pcbddc->dbg_viewer, "Maximum number of constraints per cc %" PetscInt_FMT "\n", maxneigs_r));
3940: }
3941: PetscCall(PetscLogEventEnd(PC_BDDC_AdaptiveSetUp[pcbddc->current_level], pc, 0, 0, 0));
3942: PetscFunctionReturn(PETSC_SUCCESS);
3943: }
3945: PetscErrorCode PCBDDCSetUpSolvers(PC pc)
3946: {
3947: Mat coarse_submat;
3949: PetscFunctionBegin;
3950: /* Setup local scatters R_to_B and (optionally) R_to_D */
3951: /* PCBDDCSetUpLocalWorkVectors should be called first! */
3952: PetscCall(PCBDDCSetUpLocalScatters(pc));
3954: /* Setup local neumann solver ksp_R */
3955: /* PCBDDCSetUpLocalScatters should be called first! */
3956: PetscCall(PCBDDCSetUpLocalSolvers(pc, PETSC_FALSE, PETSC_TRUE));
3958: /*
3959: Setup local correction and local part of coarse basis.
3960: Gives back the dense local part of the coarse matrix in column major ordering
3961: */
3962: PetscCall(PCBDDCSetUpCorrection(pc, &coarse_submat));
3964: /* Compute total number of coarse nodes and setup coarse solver */
3965: PetscCall(PCBDDCSetUpCoarseSolver(pc, coarse_submat));
3966: PetscCall(MatDestroy(&coarse_submat));
3967: PetscFunctionReturn(PETSC_SUCCESS);
3968: }
3970: PetscErrorCode PCBDDCResetCustomization(PC pc)
3971: {
3972: PC_BDDC *pcbddc = (PC_BDDC *)pc->data;
3974: PetscFunctionBegin;
3975: PetscCall(ISDestroy(&pcbddc->user_primal_vertices));
3976: PetscCall(ISDestroy(&pcbddc->user_primal_vertices_local));
3977: PetscCall(ISDestroy(&pcbddc->NeumannBoundaries));
3978: PetscCall(ISDestroy(&pcbddc->NeumannBoundariesLocal));
3979: PetscCall(ISDestroy(&pcbddc->DirichletBoundaries));
3980: PetscCall(MatNullSpaceDestroy(&pcbddc->onearnullspace));
3981: PetscCall(PetscFree(pcbddc->onearnullvecs_state));
3982: PetscCall(ISDestroy(&pcbddc->DirichletBoundariesLocal));
3983: PetscCall(PCBDDCSetDofsSplitting(pc, 0, NULL));
3984: PetscCall(PCBDDCSetDofsSplittingLocal(pc, 0, NULL));
3985: PetscFunctionReturn(PETSC_SUCCESS);
3986: }
3988: PetscErrorCode PCBDDCResetTopography(PC pc)
3989: {
3990: PC_BDDC *pcbddc = (PC_BDDC *)pc->data;
3991: PetscInt i;
3993: PetscFunctionBegin;
3994: PetscCall(MatDestroy(&pcbddc->nedcG));
3995: PetscCall(ISDestroy(&pcbddc->nedclocal));
3996: PetscCall(MatDestroy(&pcbddc->discretegradient));
3997: PetscCall(MatDestroy(&pcbddc->user_ChangeOfBasisMatrix));
3998: PetscCall(MatDestroy(&pcbddc->ChangeOfBasisMatrix));
3999: PetscCall(MatDestroy(&pcbddc->switch_static_change));
4000: PetscCall(VecDestroy(&pcbddc->work_change));
4001: PetscCall(MatDestroy(&pcbddc->ConstraintMatrix));
4002: PetscCall(MatDestroy(&pcbddc->divudotp));
4003: PetscCall(ISDestroy(&pcbddc->divudotp_vl2l));
4004: PetscCall(PCBDDCGraphDestroy(&pcbddc->mat_graph));
4005: for (i = 0; i < pcbddc->n_local_subs; i++) PetscCall(ISDestroy(&pcbddc->local_subs[i]));
4006: pcbddc->n_local_subs = 0;
4007: PetscCall(PetscFree(pcbddc->local_subs));
4008: PetscCall(PCBDDCSubSchursDestroy(&pcbddc->sub_schurs));
4009: pcbddc->graphanalyzed = PETSC_FALSE;
4010: pcbddc->recompute_topography = PETSC_TRUE;
4011: pcbddc->corner_selected = PETSC_FALSE;
4012: PetscFunctionReturn(PETSC_SUCCESS);
4013: }
4015: PetscErrorCode PCBDDCResetSolvers(PC pc)
4016: {
4017: PC_BDDC *pcbddc = (PC_BDDC *)pc->data;
4019: PetscFunctionBegin;
4020: PetscCall(VecDestroy(&pcbddc->coarse_vec));
4021: PetscCall(MatDestroy(&pcbddc->coarse_phi_B));
4022: PetscCall(MatDestroy(&pcbddc->coarse_phi_D));
4023: PetscCall(MatDestroy(&pcbddc->coarse_psi_B));
4024: PetscCall(MatDestroy(&pcbddc->coarse_psi_D));
4025: PetscCall(VecDestroy(&pcbddc->vec1_P));
4026: PetscCall(VecDestroy(&pcbddc->vec1_C));
4027: PetscCall(MatDestroy(&pcbddc->local_auxmat2));
4028: PetscCall(MatDestroy(&pcbddc->local_auxmat1));
4029: PetscCall(VecDestroy(&pcbddc->vec1_R));
4030: PetscCall(VecDestroy(&pcbddc->vec2_R));
4031: PetscCall(ISDestroy(&pcbddc->is_R_local));
4032: PetscCall(VecScatterDestroy(&pcbddc->R_to_B));
4033: PetscCall(VecScatterDestroy(&pcbddc->R_to_D));
4034: PetscCall(VecScatterDestroy(&pcbddc->coarse_loc_to_glob));
4035: PetscCall(KSPReset(pcbddc->ksp_D));
4036: PetscCall(KSPReset(pcbddc->ksp_R));
4037: PetscCall(KSPReset(pcbddc->coarse_ksp));
4038: PetscCall(MatDestroy(&pcbddc->local_mat));
4039: PetscCall(PetscFree(pcbddc->primal_indices_local_idxs));
4040: PetscCall(PetscFree2(pcbddc->local_primal_ref_node, pcbddc->local_primal_ref_mult));
4041: PetscCall(PetscFree(pcbddc->global_primal_indices));
4042: PetscCall(ISDestroy(&pcbddc->coarse_subassembling));
4043: PetscCall(MatDestroy(&pcbddc->benign_change));
4044: PetscCall(VecDestroy(&pcbddc->benign_vec));
4045: PetscCall(PCBDDCBenignShellMat(pc, PETSC_TRUE));
4046: PetscCall(MatDestroy(&pcbddc->benign_B0));
4047: PetscCall(PetscSFDestroy(&pcbddc->benign_sf));
4048: if (pcbddc->benign_zerodiag_subs) {
4049: PetscInt i;
4050: for (i = 0; i < pcbddc->benign_n; i++) PetscCall(ISDestroy(&pcbddc->benign_zerodiag_subs[i]));
4051: PetscCall(PetscFree(pcbddc->benign_zerodiag_subs));
4052: }
4053: PetscCall(PetscFree3(pcbddc->benign_p0_lidx, pcbddc->benign_p0_gidx, pcbddc->benign_p0));
4054: PetscFunctionReturn(PETSC_SUCCESS);
4055: }
4057: PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc)
4058: {
4059: PC_BDDC *pcbddc = (PC_BDDC *)pc->data;
4060: PC_IS *pcis = (PC_IS *)pc->data;
4061: VecType impVecType;
4062: PetscInt n_constraints, n_R, old_size;
4064: PetscFunctionBegin;
4065: n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices;
4066: n_R = pcis->n - pcbddc->n_vertices;
4067: PetscCall(VecGetType(pcis->vec1_N, &impVecType));
4068: /* local work vectors (try to avoid unneeded work)*/
4069: /* R nodes */
4070: old_size = -1;
4071: if (pcbddc->vec1_R) PetscCall(VecGetSize(pcbddc->vec1_R, &old_size));
4072: if (n_R != old_size) {
4073: PetscCall(VecDestroy(&pcbddc->vec1_R));
4074: PetscCall(VecDestroy(&pcbddc->vec2_R));
4075: PetscCall(VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N), &pcbddc->vec1_R));
4076: PetscCall(VecSetSizes(pcbddc->vec1_R, PETSC_DECIDE, n_R));
4077: PetscCall(VecSetType(pcbddc->vec1_R, impVecType));
4078: PetscCall(VecDuplicate(pcbddc->vec1_R, &pcbddc->vec2_R));
4079: }
4080: /* local primal dofs */
4081: old_size = -1;
4082: if (pcbddc->vec1_P) PetscCall(VecGetSize(pcbddc->vec1_P, &old_size));
4083: if (pcbddc->local_primal_size != old_size) {
4084: PetscCall(VecDestroy(&pcbddc->vec1_P));
4085: PetscCall(VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N), &pcbddc->vec1_P));
4086: PetscCall(VecSetSizes(pcbddc->vec1_P, PETSC_DECIDE, pcbddc->local_primal_size));
4087: PetscCall(VecSetType(pcbddc->vec1_P, impVecType));
4088: }
4089: /* local explicit constraints */
4090: old_size = -1;
4091: if (pcbddc->vec1_C) PetscCall(VecGetSize(pcbddc->vec1_C, &old_size));
4092: if (n_constraints && n_constraints != old_size) {
4093: PetscCall(VecDestroy(&pcbddc->vec1_C));
4094: PetscCall(VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N), &pcbddc->vec1_C));
4095: PetscCall(VecSetSizes(pcbddc->vec1_C, PETSC_DECIDE, n_constraints));
4096: PetscCall(VecSetType(pcbddc->vec1_C, impVecType));
4097: }
4098: PetscFunctionReturn(PETSC_SUCCESS);
4099: }
4101: static PetscErrorCode MatSetValuesSubMat(Mat A, Mat S, PetscInt nr, const PetscInt rows[], PetscInt nc, const PetscInt cols[], InsertMode imode)
4102: {
4103: PetscBool flg;
4104: const PetscScalar *a;
4106: PetscFunctionBegin;
4107: PetscCall(PetscObjectBaseTypeCompare((PetscObject)S, MATSEQDENSE, &flg));
4108: if (flg) {
4109: PetscCall(MatDenseGetArrayRead(S, &a));
4110: PetscCall(MatSetOption(A, MAT_ROW_ORIENTED, PETSC_FALSE));
4111: PetscCall(MatSetValues(A, nr, rows, nc, cols, a, imode));
4112: PetscCall(MatSetOption(A, MAT_ROW_ORIENTED, PETSC_TRUE));
4113: PetscCall(MatDenseRestoreArrayRead(S, &a));
4114: } else {
4115: const PetscInt *ii, *jj;
4116: PetscInt n;
4117: PetscInt buf[8192], *bufc = NULL;
4118: PetscBool freeb = PETSC_FALSE;
4119: Mat Sm = S;
4121: PetscCall(PetscObjectBaseTypeCompare((PetscObject)S, MATSEQAIJ, &flg));
4122: if (!flg) PetscCall(MatConvert(S, MATSEQAIJ, MAT_INITIAL_MATRIX, &Sm));
4123: else PetscCall(PetscObjectReference((PetscObject)S));
4124: PetscCall(MatSeqAIJGetArrayRead(Sm, &a));
4125: PetscCall(MatGetRowIJ(Sm, 0, PETSC_FALSE, PETSC_FALSE, &n, &ii, &jj, &flg));
4126: PetscCheck(flg, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot get IJ structure");
4127: if (nc <= (PetscInt)PETSC_STATIC_ARRAY_LENGTH(buf)) {
4128: bufc = buf;
4129: } else {
4130: PetscCall(PetscMalloc1(nc, &bufc));
4131: freeb = PETSC_TRUE;
4132: }
4134: for (PetscInt i = 0; i < n; i++) {
4135: const PetscInt nci = ii[i + 1] - ii[i];
4137: for (PetscInt j = 0; j < nci; j++) bufc[j] = cols[jj[ii[i] + j]];
4138: PetscCall(MatSetValues(A, 1, rows + i, nci, bufc, a + ii[i], imode));
4139: }
4140: PetscCall(MatRestoreRowIJ(Sm, 0, PETSC_FALSE, PETSC_FALSE, &n, &ii, &jj, &flg));
4141: PetscCall(MatSeqAIJRestoreArrayRead(Sm, &a));
4142: PetscCall(MatDestroy(&Sm));
4143: if (freeb) PetscCall(PetscFree(bufc));
4144: }
4145: PetscCall(MatAssemblyBegin(A, MAT_FLUSH_ASSEMBLY));
4146: PetscCall(MatAssemblyEnd(A, MAT_FLUSH_ASSEMBLY));
4147: PetscFunctionReturn(PETSC_SUCCESS);
4148: }
4150: static PetscErrorCode MatCreateSeqAIJFromDenseExpand(Mat D, PetscInt n, const PetscInt j[], Mat *mat)
4151: {
4152: Mat_SeqAIJ *aij;
4153: PetscInt *ii, *jj;
4154: PetscScalar *aa;
4155: PetscInt nnz = 0, m, nc;
4156: const PetscScalar *a;
4157: const PetscScalar zero = 0.0;
4159: PetscFunctionBegin;
4160: PetscCall(MatGetLocalSize(D, &m, &nc));
4161: PetscCall(MatDenseGetArrayRead(D, &a));
4162: PetscCall(PetscMalloc1(m + 1, &ii));
4163: PetscCall(PetscMalloc1(m * nc, &jj));
4164: PetscCall(PetscMalloc1(m * nc, &aa));
4165: ii[0] = 0;
4166: for (PetscInt k = 0; k < m; k++) {
4167: for (PetscInt s = 0; s < nc; s++) {
4168: const PetscInt c = s + k * nc;
4169: const PetscScalar v = a[k + s * m];
4171: if (PetscUnlikely(j[c] < 0 || v == zero)) continue;
4172: jj[nnz] = j[c];
4173: aa[nnz] = a[k + s * m];
4174: nnz++;
4175: }
4176: ii[k + 1] = nnz;
4177: }
4179: PetscCall(MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)D), m, n, ii, jj, aa, mat));
4180: PetscCall(MatDenseRestoreArrayRead(D, &a));
4182: aij = (Mat_SeqAIJ *)(*mat)->data;
4183: aij->free_a = PETSC_TRUE;
4184: aij->free_ij = PETSC_TRUE;
4185: PetscFunctionReturn(PETSC_SUCCESS);
4186: }
4188: /* adapted from MatInvertVariableBlockDiagonal_SeqAIJ */
4189: static PetscErrorCode MatSeqAIJInvertVariableBlockDiagonalMat(Mat A, PetscInt nblocks, const PetscInt *bsizes, Mat *B)
4190: {
4191: PetscInt n = A->rmap->n, ncnt = 0, ncnt2 = 0, bsizemax = 0, *v_pivots = NULL;
4192: const PetscBool allowzeropivot = PETSC_FALSE;
4193: PetscBool zeropivotdetected = PETSC_FALSE;
4194: const PetscReal shift = 0.0;
4195: PetscInt ipvt[5], *ii, *jj, *indi, *indj;
4196: PetscScalar work[25], *v_work = NULL, *aa, *diag;
4197: PetscLogDouble flops = 0.0;
4199: PetscFunctionBegin;
4200: PetscCheck(A->rmap->n == A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Not for rectangular matrices");
4201: for (PetscInt i = 0; i < nblocks; i++) {
4202: ncnt += bsizes[i];
4203: ncnt2 += PetscSqr(bsizes[i]);
4204: }
4205: PetscCheck(ncnt == n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Total blocksizes %" PetscInt_FMT " doesn't match number matrix rows %" PetscInt_FMT, ncnt, n);
4206: for (PetscInt i = 0; i < nblocks; i++) bsizemax = PetscMax(bsizemax, bsizes[i]);
4207: if (bsizemax > 7) PetscCall(PetscMalloc2(bsizemax, &v_work, bsizemax, &v_pivots));
4209: PetscCall(PetscMalloc1(n + 1, &ii));
4210: PetscCall(PetscMalloc1(ncnt2, &jj));
4211: PetscCall(PetscCalloc1(ncnt2, &aa));
4213: ncnt = 0;
4214: ii[0] = 0;
4215: indi = ii;
4216: indj = jj;
4217: diag = aa;
4218: for (PetscInt i = 0; i < nblocks; i++) {
4219: const PetscInt bs = bsizes[i];
4221: for (PetscInt k = 0; k < bs; k++) {
4222: indi[k + 1] = indi[k] + bs;
4223: for (PetscInt j = 0; j < bs; j++) indj[k * bs + j] = ncnt + j;
4224: }
4225: PetscCall(MatGetValues(A, bs, indj, bs, indj, diag));
4226: switch (bs) {
4227: case 1:
4228: *diag = 1.0 / (*diag);
4229: break;
4230: case 2:
4231: PetscCall(PetscKernel_A_gets_inverse_A_2(diag, shift, allowzeropivot, &zeropivotdetected));
4232: break;
4233: case 3:
4234: PetscCall(PetscKernel_A_gets_inverse_A_3(diag, shift, allowzeropivot, &zeropivotdetected));
4235: break;
4236: case 4:
4237: PetscCall(PetscKernel_A_gets_inverse_A_4(diag, shift, allowzeropivot, &zeropivotdetected));
4238: break;
4239: case 5:
4240: PetscCall(PetscKernel_A_gets_inverse_A_5(diag, ipvt, work, shift, allowzeropivot, &zeropivotdetected));
4241: break;
4242: case 6:
4243: PetscCall(PetscKernel_A_gets_inverse_A_6(diag, shift, allowzeropivot, &zeropivotdetected));
4244: break;
4245: case 7:
4246: PetscCall(PetscKernel_A_gets_inverse_A_7(diag, shift, allowzeropivot, &zeropivotdetected));
4247: break;
4248: default:
4249: PetscCall(PetscKernel_A_gets_inverse_A(bs, diag, v_pivots, v_work, allowzeropivot, &zeropivotdetected));
4250: }
4251: ncnt += bs;
4252: flops += 2.0 * PetscPowInt(bs, 3) / 3.0;
4253: diag += bs * bs;
4254: indj += bs * bs;
4255: indi += bs;
4256: }
4257: PetscCall(PetscLogFlops(flops));
4258: PetscCall(PetscFree2(v_work, v_pivots));
4259: PetscCall(MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A), n, n, ii, jj, aa, B));
4260: {
4261: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)(*B)->data;
4262: aij->free_a = PETSC_TRUE;
4263: aij->free_ij = PETSC_TRUE;
4264: }
4265: PetscFunctionReturn(PETSC_SUCCESS);
4266: }
4268: PetscErrorCode PCBDDCSetUpCorrection(PC pc, Mat *coarse_submat)
4269: {
4270: PC_IS *pcis = (PC_IS *)pc->data;
4271: PC_BDDC *pcbddc = (PC_BDDC *)pc->data;
4272: PCBDDCGraph graph = pcbddc->mat_graph;
4273: PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs;
4274: /* submatrices of local problem */
4275: Mat A_RV = NULL, A_VR, A_VV, local_auxmat2_R = NULL;
4276: /* submatrices of local coarse problem */
4277: Mat S_CV = NULL, S_VC = NULL, S_CC = NULL;
4278: /* working matrices */
4279: Mat C_CR;
4281: /* additional working stuff */
4282: PC pc_R;
4283: IS is_R, is_V, is_C;
4284: const PetscInt *idx_V, *idx_C;
4285: Mat F, Brhs = NULL;
4286: Vec dummy_vec;
4287: PetscBool isPreonly, isLU, isCHOL, need_benign_correction, sparserhs;
4288: PetscInt *idx_V_B;
4289: PetscInt lda_rhs, n_vertices, n_constraints, *p0_lidx_I;
4290: PetscInt n_eff_vertices, n_eff_constraints;
4291: PetscInt i, n_R, n_D, n_B;
4292: PetscScalar one = 1.0, m_one = -1.0;
4294: /* Multi-element support */
4295: PetscBool multi_element = graph->multi_element;
4296: PetscInt *V_to_eff_V = NULL, *C_to_eff_C = NULL;
4297: PetscInt *B_eff_V_J = NULL, *R_eff_V_J = NULL, *B_eff_C_J = NULL, *R_eff_C_J = NULL;
4298: IS is_C_perm = NULL;
4299: PetscInt n_C_bss = 0, *C_bss = NULL;
4300: Mat coarse_phi_multi;
4302: PetscFunctionBegin;
4303: PetscCheck(pcbddc->symmetric_primal || !pcbddc->benign_n, PETSC_COMM_SELF, PETSC_ERR_SUP, "Non-symmetric primal basis computation with benign trick not yet implemented");
4304: PetscCall(PetscLogEventBegin(PC_BDDC_CorrectionSetUp[pcbddc->current_level], pc, 0, 0, 0));
4306: /* Set Non-overlapping dimensions */
4307: n_vertices = pcbddc->n_vertices;
4308: n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices;
4309: n_B = pcis->n_B;
4310: n_D = pcis->n - n_B;
4311: n_R = pcis->n - n_vertices;
4313: /* vertices in boundary numbering */
4314: PetscCall(PetscMalloc1(n_vertices, &idx_V_B));
4315: PetscCall(ISGlobalToLocalMappingApply(pcis->BtoNmap, IS_GTOLM_DROP, n_vertices, pcbddc->local_primal_ref_node, &i, idx_V_B));
4316: PetscCheck(i == n_vertices, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Error in boundary numbering for BDDC vertices! %" PetscInt_FMT " != %" PetscInt_FMT, n_vertices, i);
4318: /* these two cases still need to be optimized */
4319: if (pcbddc->benign_saddle_point || !pcbddc->symmetric_primal) multi_element = PETSC_FALSE;
4321: /* Subdomain contribution (Non-overlapping) to coarse matrix */
4322: if (multi_element) {
4323: PetscCheck(!pcbddc->benign_n, PETSC_COMM_SELF, PETSC_ERR_SUP, "Not yet implemented");
4325: PetscCall(MatCreate(PETSC_COMM_SELF, coarse_submat));
4326: PetscCall(MatSetSizes(*coarse_submat, pcbddc->local_primal_size, pcbddc->local_primal_size, pcbddc->local_primal_size, pcbddc->local_primal_size));
4327: PetscCall(MatSetType(*coarse_submat, MATSEQAIJ));
4328: PetscCall(MatSetOption(*coarse_submat, MAT_IGNORE_ZERO_ENTRIES, PETSC_TRUE));
4329: PetscCall(MatSetOption(*coarse_submat, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
4331: /* group vertices and constraints by subdomain id */
4332: const PetscInt *vidxs = pcbddc->primal_indices_local_idxs;
4333: const PetscInt *cidxs = pcbddc->primal_indices_local_idxs + n_vertices;
4334: PetscInt *count_eff, *V_eff_to_V, *C_eff_to_C, *nnz;
4335: PetscInt n_el = PetscMax(graph->n_local_subs, 1);
4337: PetscCall(PetscCalloc1(2 * n_el, &count_eff));
4338: PetscCall(PetscMalloc1(n_vertices, &V_to_eff_V));
4339: PetscCall(PetscMalloc1(n_constraints, &C_to_eff_C));
4340: for (PetscInt i = 0; i < n_vertices; i++) {
4341: PetscInt s = 2 * graph->nodes[vidxs[i]].local_sub;
4343: V_to_eff_V[i] = count_eff[s];
4344: count_eff[s] += 1;
4345: }
4346: for (PetscInt i = 0; i < n_constraints; i++) {
4347: PetscInt s = 2 * graph->nodes[cidxs[i]].local_sub + 1;
4349: C_to_eff_C[i] = count_eff[s];
4350: count_eff[s] += 1;
4351: }
4353: /* preallocation */
4354: PetscCall(PetscMalloc1(n_vertices + n_constraints, &nnz));
4355: for (PetscInt i = 0; i < n_vertices; i++) {
4356: PetscInt s = 2 * graph->nodes[vidxs[i]].local_sub;
4358: nnz[i] = count_eff[s] + count_eff[s + 1];
4359: }
4360: for (PetscInt i = 0; i < n_constraints; i++) {
4361: PetscInt s = 2 * graph->nodes[cidxs[i]].local_sub;
4363: nnz[i + n_vertices] = count_eff[s] + count_eff[s + 1];
4364: }
4365: PetscCall(MatSeqAIJSetPreallocation(*coarse_submat, 0, nnz));
4366: PetscCall(PetscFree(nnz));
4368: n_eff_vertices = 0;
4369: n_eff_constraints = 0;
4370: for (PetscInt i = 0; i < n_el; i++) {
4371: n_eff_vertices = PetscMax(n_eff_vertices, count_eff[2 * i]);
4372: n_eff_constraints = PetscMax(n_eff_constraints, count_eff[2 * i + 1]);
4373: count_eff[2 * i] = 0;
4374: count_eff[2 * i + 1] = 0;
4375: }
4377: const PetscInt *idx;
4378: PetscCall(PetscMalloc2(n_el * n_eff_vertices, &V_eff_to_V, n_el * n_eff_constraints, &C_eff_to_C));
4380: for (PetscInt i = 0; i < n_vertices; i++) {
4381: const PetscInt e = graph->nodes[vidxs[i]].local_sub;
4382: const PetscInt s = 2 * e;
4384: V_eff_to_V[e * n_eff_vertices + count_eff[s]] = i;
4385: count_eff[s] += 1;
4386: }
4387: for (PetscInt i = 0; i < n_constraints; i++) {
4388: const PetscInt e = graph->nodes[cidxs[i]].local_sub;
4389: const PetscInt s = 2 * e + 1;
4391: C_eff_to_C[e * n_eff_constraints + count_eff[s]] = i;
4392: count_eff[s] += 1;
4393: }
4395: PetscCall(PetscMalloc1(n_R * n_eff_vertices, &R_eff_V_J));
4396: PetscCall(PetscMalloc1(n_R * n_eff_constraints, &R_eff_C_J));
4397: PetscCall(PetscMalloc1(n_B * n_eff_vertices, &B_eff_V_J));
4398: PetscCall(PetscMalloc1(n_B * n_eff_constraints, &B_eff_C_J));
4399: for (PetscInt i = 0; i < n_R * n_eff_vertices; i++) R_eff_V_J[i] = -1;
4400: for (PetscInt i = 0; i < n_R * n_eff_constraints; i++) R_eff_C_J[i] = -1;
4401: for (PetscInt i = 0; i < n_B * n_eff_vertices; i++) B_eff_V_J[i] = -1;
4402: for (PetscInt i = 0; i < n_B * n_eff_constraints; i++) B_eff_C_J[i] = -1;
4404: PetscCall(ISGetIndices(pcbddc->is_R_local, &idx));
4405: for (PetscInt i = 0; i < n_R; i++) {
4406: const PetscInt e = graph->nodes[idx[i]].local_sub;
4407: const PetscInt s = 2 * e;
4408: PetscInt j;
4410: for (j = 0; j < count_eff[s]; j++) R_eff_V_J[i * n_eff_vertices + j] = V_eff_to_V[e * n_eff_vertices + j];
4411: for (j = 0; j < count_eff[s + 1]; j++) R_eff_C_J[i * n_eff_constraints + j] = C_eff_to_C[e * n_eff_constraints + j];
4412: }
4413: PetscCall(ISRestoreIndices(pcbddc->is_R_local, &idx));
4414: PetscCall(ISGetIndices(pcis->is_B_local, &idx));
4415: for (PetscInt i = 0; i < n_B; i++) {
4416: const PetscInt e = graph->nodes[idx[i]].local_sub;
4417: const PetscInt s = 2 * e;
4418: PetscInt j;
4420: for (j = 0; j < count_eff[s]; j++) B_eff_V_J[i * n_eff_vertices + j] = V_eff_to_V[e * n_eff_vertices + j];
4421: for (j = 0; j < count_eff[s + 1]; j++) B_eff_C_J[i * n_eff_constraints + j] = C_eff_to_C[e * n_eff_constraints + j];
4422: }
4423: PetscCall(ISRestoreIndices(pcis->is_B_local, &idx));
4425: /* permutation and blocksizes for block invert of S_CC */
4426: PetscInt *idxp;
4428: PetscCall(PetscMalloc1(n_constraints, &idxp));
4429: PetscCall(PetscMalloc1(n_el, &C_bss));
4430: n_C_bss = 0;
4431: for (PetscInt e = 0, cnt = 0; e < n_el; e++) {
4432: const PetscInt nc = count_eff[2 * e + 1];
4434: if (nc) C_bss[n_C_bss++] = nc;
4435: for (PetscInt c = 0; c < nc; c++) idxp[cnt + c] = C_eff_to_C[e * n_eff_constraints + c];
4436: cnt += nc;
4437: }
4439: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, n_constraints, idxp, PETSC_OWN_POINTER, &is_C_perm));
4441: PetscCall(PetscFree2(V_eff_to_V, C_eff_to_C));
4442: PetscCall(PetscFree(count_eff));
4443: } else {
4444: PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, pcbddc->local_primal_size, pcbddc->local_primal_size, NULL, coarse_submat));
4445: n_eff_constraints = n_constraints;
4446: n_eff_vertices = n_vertices;
4447: }
4449: /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */
4450: PetscCall(KSPGetPC(pcbddc->ksp_R, &pc_R));
4451: PetscCall(PCSetUp(pc_R));
4452: PetscCall(PetscObjectTypeCompare((PetscObject)pcbddc->ksp_R, KSPPREONLY, &isPreonly));
4453: PetscCall(PetscObjectTypeCompare((PetscObject)pc_R, PCLU, &isLU));
4454: PetscCall(PetscObjectTypeCompare((PetscObject)pc_R, PCCHOLESKY, &isCHOL));
4455: lda_rhs = n_R;
4456: need_benign_correction = PETSC_FALSE;
4457: F = NULL;
4458: if (isPreonly && (isLU || isCHOL)) {
4459: PetscCall(PCFactorGetMatrix(pc_R, &F));
4460: } else if (sub_schurs && sub_schurs->reuse_solver) {
4461: PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
4462: MatFactorType type;
4464: F = reuse_solver->F;
4465: PetscCall(MatGetFactorType(F, &type));
4466: if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE;
4467: if (type == MAT_FACTOR_LU) isLU = PETSC_TRUE;
4468: PetscCall(MatGetSize(F, &lda_rhs, NULL));
4469: need_benign_correction = (PetscBool)(!!reuse_solver->benign_n);
4470: }
4472: /* determine if we can use a sparse right-hand side */
4473: sparserhs = PETSC_FALSE;
4474: if (F && !multi_element) {
4475: MatSolverType solver;
4477: PetscCall(MatFactorGetSolverType(F, &solver));
4478: PetscCall(PetscStrcmp(solver, MATSOLVERMUMPS, &sparserhs));
4479: }
4481: /* create dummy vector to modify rhs and sol of MatMatSolve (work array will never be used) */
4482: dummy_vec = NULL;
4483: if (need_benign_correction && lda_rhs != n_R && F) {
4484: PetscCall(VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N), &dummy_vec));
4485: PetscCall(VecSetSizes(dummy_vec, lda_rhs, PETSC_DECIDE));
4486: PetscCall(VecSetType(dummy_vec, ((PetscObject)pcis->vec1_N)->type_name));
4487: }
4489: PetscCall(MatDestroy(&pcbddc->local_auxmat1));
4490: PetscCall(MatDestroy(&pcbddc->local_auxmat2));
4492: PetscCall(ISCreateStride(PETSC_COMM_SELF, n_R, 0, 1, &is_R));
4493: PetscCall(ISCreateStride(PETSC_COMM_SELF, n_vertices, 0, 1, &is_V));
4494: PetscCall(ISCreateStride(PETSC_COMM_SELF, n_constraints, n_vertices, 1, &is_C));
4495: PetscCall(ISGetIndices(is_V, &idx_V));
4496: PetscCall(ISGetIndices(is_C, &idx_C));
4498: /* Precompute stuffs needed for preprocessing and application of BDDC*/
4499: if (n_constraints) {
4500: Mat C_B;
4502: /* Extract constraints on R nodes: C_{CR} */
4503: PetscCall(MatCreateSubMatrix(pcbddc->ConstraintMatrix, is_C, pcbddc->is_R_local, MAT_INITIAL_MATRIX, &C_CR));
4504: PetscCall(MatCreateSubMatrix(pcbddc->ConstraintMatrix, is_C, pcis->is_B_local, MAT_INITIAL_MATRIX, &C_B));
4506: /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */
4507: /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */
4508: if (!sparserhs) {
4509: PetscScalar *marr;
4511: PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, lda_rhs, n_eff_constraints, NULL, &Brhs));
4512: PetscCall(MatDenseGetArrayWrite(Brhs, &marr));
4513: for (i = 0; i < n_constraints; i++) {
4514: const PetscScalar *row_cmat_values;
4515: const PetscInt *row_cmat_indices;
4516: PetscInt size_of_constraint, j, col = C_to_eff_C ? C_to_eff_C[i] : i;
4518: PetscCall(MatGetRow(C_CR, i, &size_of_constraint, &row_cmat_indices, &row_cmat_values));
4519: for (j = 0; j < size_of_constraint; j++) marr[row_cmat_indices[j] + col * lda_rhs] = -row_cmat_values[j];
4520: PetscCall(MatRestoreRow(C_CR, i, &size_of_constraint, &row_cmat_indices, &row_cmat_values));
4521: }
4522: PetscCall(MatDenseRestoreArrayWrite(Brhs, &marr));
4523: } else {
4524: Mat tC_CR;
4526: PetscCall(MatScale(C_CR, -1.0));
4527: if (lda_rhs != n_R) {
4528: PetscScalar *aa;
4529: PetscInt r, *ii, *jj;
4530: PetscBool done;
4532: PetscCall(MatGetRowIJ(C_CR, 0, PETSC_FALSE, PETSC_FALSE, &r, (const PetscInt **)&ii, (const PetscInt **)&jj, &done));
4533: PetscCheck(done, PETSC_COMM_SELF, PETSC_ERR_PLIB, "GetRowIJ failed");
4534: PetscCall(MatSeqAIJGetArray(C_CR, &aa));
4535: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, n_constraints, lda_rhs, ii, jj, aa, &tC_CR));
4536: PetscCall(MatRestoreRowIJ(C_CR, 0, PETSC_FALSE, PETSC_FALSE, &r, (const PetscInt **)&ii, (const PetscInt **)&jj, &done));
4537: PetscCheck(done, PETSC_COMM_SELF, PETSC_ERR_PLIB, "RestoreRowIJ failed");
4538: } else {
4539: PetscCall(PetscObjectReference((PetscObject)C_CR));
4540: tC_CR = C_CR;
4541: }
4542: PetscCall(MatCreateTranspose(tC_CR, &Brhs));
4543: PetscCall(MatDestroy(&tC_CR));
4544: }
4545: PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, lda_rhs, n_eff_constraints, NULL, &local_auxmat2_R));
4546: if (F) {
4547: if (need_benign_correction) {
4548: PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
4550: /* rhs is already zero on interior dofs, no need to change the rhs */
4551: PetscCall(PetscArrayzero(reuse_solver->benign_save_vals, pcbddc->benign_n));
4552: }
4553: PetscCall(MatMatSolve(F, Brhs, local_auxmat2_R));
4554: if (need_benign_correction) {
4555: PetscScalar *marr;
4556: PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
4558: /* XXX multi_element? */
4559: PetscCall(MatDenseGetArray(local_auxmat2_R, &marr));
4560: if (lda_rhs != n_R) {
4561: for (i = 0; i < n_eff_constraints; i++) {
4562: PetscCall(VecPlaceArray(dummy_vec, marr + i * lda_rhs));
4563: PetscCall(PCBDDCReuseSolversBenignAdapt(reuse_solver, dummy_vec, NULL, PETSC_TRUE, PETSC_TRUE));
4564: PetscCall(VecResetArray(dummy_vec));
4565: }
4566: } else {
4567: for (i = 0; i < n_eff_constraints; i++) {
4568: PetscCall(VecPlaceArray(pcbddc->vec1_R, marr + i * lda_rhs));
4569: PetscCall(PCBDDCReuseSolversBenignAdapt(reuse_solver, pcbddc->vec1_R, NULL, PETSC_TRUE, PETSC_TRUE));
4570: PetscCall(VecResetArray(pcbddc->vec1_R));
4571: }
4572: }
4573: PetscCall(MatDenseRestoreArray(local_auxmat2_R, &marr));
4574: }
4575: } else {
4576: const PetscScalar *barr;
4577: PetscScalar *marr;
4579: PetscCall(MatDenseGetArrayRead(Brhs, &barr));
4580: PetscCall(MatDenseGetArray(local_auxmat2_R, &marr));
4581: for (i = 0; i < n_eff_constraints; i++) {
4582: PetscCall(VecPlaceArray(pcbddc->vec1_R, barr + i * lda_rhs));
4583: PetscCall(VecPlaceArray(pcbddc->vec2_R, marr + i * lda_rhs));
4584: PetscCall(KSPSolve(pcbddc->ksp_R, pcbddc->vec1_R, pcbddc->vec2_R));
4585: PetscCall(KSPCheckSolve(pcbddc->ksp_R, pc, pcbddc->vec2_R));
4586: PetscCall(VecResetArray(pcbddc->vec1_R));
4587: PetscCall(VecResetArray(pcbddc->vec2_R));
4588: }
4589: PetscCall(MatDenseRestoreArrayRead(Brhs, &barr));
4590: PetscCall(MatDenseRestoreArray(local_auxmat2_R, &marr));
4591: }
4592: if (sparserhs) PetscCall(MatScale(C_CR, -1.0));
4593: PetscCall(MatDestroy(&Brhs));
4594: /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR})^{-1} */
4595: if (!pcbddc->switch_static) {
4596: PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, n_B, n_eff_constraints, NULL, &pcbddc->local_auxmat2));
4597: PetscCall(MatDenseScatter_Private(pcbddc->R_to_B, local_auxmat2_R, pcbddc->local_auxmat2, INSERT_VALUES, SCATTER_FORWARD));
4598: if (multi_element) {
4599: Mat T;
4601: PetscCall(MatCreateSeqAIJFromDenseExpand(local_auxmat2_R, n_constraints, R_eff_C_J, &T));
4602: PetscCall(MatDestroy(&local_auxmat2_R));
4603: local_auxmat2_R = T;
4604: PetscCall(MatCreateSeqAIJFromDenseExpand(pcbddc->local_auxmat2, n_constraints, B_eff_C_J, &T));
4605: PetscCall(MatDestroy(&pcbddc->local_auxmat2));
4606: pcbddc->local_auxmat2 = T;
4607: }
4608: PetscCall(MatMatMult(C_B, pcbddc->local_auxmat2, MAT_INITIAL_MATRIX, PETSC_DETERMINE, &S_CC));
4609: } else {
4610: if (multi_element) {
4611: Mat T;
4613: PetscCall(MatCreateSeqAIJFromDenseExpand(local_auxmat2_R, n_constraints, R_eff_C_J, &T));
4614: PetscCall(MatDestroy(&local_auxmat2_R));
4615: local_auxmat2_R = T;
4616: }
4617: if (lda_rhs != n_R) {
4618: PetscCall(MatCreateSubMatrix(local_auxmat2_R, is_R, NULL, MAT_INITIAL_MATRIX, &pcbddc->local_auxmat2));
4619: } else {
4620: PetscCall(PetscObjectReference((PetscObject)local_auxmat2_R));
4621: pcbddc->local_auxmat2 = local_auxmat2_R;
4622: }
4623: PetscCall(MatMatMult(C_CR, pcbddc->local_auxmat2, MAT_INITIAL_MATRIX, PETSC_DETERMINE, &S_CC));
4624: }
4625: PetscCall(MatScale(S_CC, m_one));
4626: if (multi_element) {
4627: Mat T, T2;
4628: IS isp, ispi;
4630: isp = is_C_perm;
4632: PetscCall(ISInvertPermutation(isp, PETSC_DECIDE, &ispi));
4633: PetscCall(MatPermute(S_CC, isp, isp, &T));
4634: PetscCall(MatSeqAIJInvertVariableBlockDiagonalMat(T, n_C_bss, C_bss, &T2));
4635: PetscCall(MatDestroy(&T));
4636: PetscCall(MatDestroy(&S_CC));
4637: PetscCall(MatPermute(T2, ispi, ispi, &S_CC));
4638: PetscCall(MatDestroy(&T2));
4639: PetscCall(ISDestroy(&ispi));
4640: } else {
4641: if (isCHOL) {
4642: PetscCall(MatCholeskyFactor(S_CC, NULL, NULL));
4643: } else {
4644: PetscCall(MatLUFactor(S_CC, NULL, NULL, NULL));
4645: }
4646: PetscCall(MatSeqDenseInvertFactors_Private(S_CC));
4647: }
4648: /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */
4649: PetscCall(MatMatMult(S_CC, C_B, MAT_INITIAL_MATRIX, PETSC_DETERMINE, &pcbddc->local_auxmat1));
4650: PetscCall(MatDestroy(&C_B));
4651: PetscCall(MatSetValuesSubMat(*coarse_submat, S_CC, n_constraints, idx_C, n_constraints, idx_C, INSERT_VALUES));
4652: }
4654: /* Get submatrices from subdomain matrix */
4655: if (n_vertices) {
4656: #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
4657: PetscBool oldpin;
4658: #endif
4659: IS is_aux;
4661: if (sub_schurs && sub_schurs->reuse_solver) { /* is_R_local is not sorted, ISComplement doesn't like it */
4662: IS tis;
4664: PetscCall(ISDuplicate(pcbddc->is_R_local, &tis));
4665: PetscCall(ISSort(tis));
4666: PetscCall(ISComplement(tis, 0, pcis->n, &is_aux));
4667: PetscCall(ISDestroy(&tis));
4668: } else {
4669: PetscCall(ISComplement(pcbddc->is_R_local, 0, pcis->n, &is_aux));
4670: }
4671: #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
4672: oldpin = pcbddc->local_mat->boundtocpu;
4673: #endif
4674: PetscCall(MatBindToCPU(pcbddc->local_mat, PETSC_TRUE));
4675: PetscCall(MatCreateSubMatrix(pcbddc->local_mat, pcbddc->is_R_local, is_aux, MAT_INITIAL_MATRIX, &A_RV));
4676: PetscCall(MatCreateSubMatrix(pcbddc->local_mat, is_aux, pcbddc->is_R_local, MAT_INITIAL_MATRIX, &A_VR));
4677: /* TODO REMOVE: MatMatMult(A_VR,A_RRmA_RV) below may raise an error */
4678: PetscCall(MatConvert(A_VR, MATSEQAIJ, MAT_INPLACE_MATRIX, &A_VR));
4679: PetscCall(MatCreateSubMatrix(pcbddc->local_mat, is_aux, is_aux, MAT_INITIAL_MATRIX, &A_VV));
4680: #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
4681: PetscCall(MatBindToCPU(pcbddc->local_mat, oldpin));
4682: #endif
4683: PetscCall(ISDestroy(&is_aux));
4684: }
4685: PetscCall(ISDestroy(&is_C_perm));
4686: PetscCall(PetscFree(C_bss));
4688: p0_lidx_I = NULL;
4689: if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) {
4690: const PetscInt *idxs;
4692: PetscCall(ISGetIndices(pcis->is_I_local, &idxs));
4693: PetscCall(PetscMalloc1(pcbddc->benign_n, &p0_lidx_I));
4694: for (i = 0; i < pcbddc->benign_n; i++) PetscCall(PetscFindInt(pcbddc->benign_p0_lidx[i], pcis->n - pcis->n_B, idxs, &p0_lidx_I[i]));
4695: PetscCall(ISRestoreIndices(pcis->is_I_local, &idxs));
4696: }
4698: /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */
4700: /* Matrices of coarse basis functions (local) */
4701: PetscCall(MatDestroy(&pcbddc->coarse_phi_B));
4702: PetscCall(MatDestroy(&pcbddc->coarse_psi_B));
4703: PetscCall(MatDestroy(&pcbddc->coarse_phi_D));
4704: PetscCall(MatDestroy(&pcbddc->coarse_psi_D));
4705: if (!multi_element) {
4706: PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, n_B, pcbddc->local_primal_size, NULL, &pcbddc->coarse_phi_B));
4707: if (pcbddc->switch_static || pcbddc->dbg_flag) PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, n_D, pcbddc->local_primal_size, NULL, &pcbddc->coarse_phi_D));
4708: coarse_phi_multi = NULL;
4709: } else { /* Create temporary NEST matrix to hold coarse basis functions blocks */
4710: IS is_rows[2] = {pcbddc->is_R_local, NULL};
4711: IS is_cols[2] = {is_V, is_C};
4713: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, n_vertices, pcbddc->local_primal_ref_node, PETSC_USE_POINTER, &is_rows[1]));
4714: PetscCall(MatCreateNest(PETSC_COMM_SELF, 2, is_rows, 2, is_cols, NULL, &coarse_phi_multi));
4715: PetscCall(ISDestroy(&is_rows[1]));
4716: }
4718: /* vertices */
4719: if (n_vertices) {
4720: PetscBool restoreavr = PETSC_FALSE;
4721: Mat A_RRmA_RV = NULL;
4723: PetscCall(MatSetValuesSubMat(*coarse_submat, A_VV, n_vertices, idx_V, n_vertices, idx_V, ADD_VALUES));
4724: PetscCall(MatDestroy(&A_VV));
4726: if (n_R) {
4727: Mat A_RV_bcorr = NULL, S_VV;
4729: PetscCall(MatScale(A_RV, m_one));
4730: if (need_benign_correction) {
4731: ISLocalToGlobalMapping RtoN;
4732: IS is_p0;
4733: PetscInt *idxs_p0, n;
4735: PetscCall(PetscMalloc1(pcbddc->benign_n, &idxs_p0));
4736: PetscCall(ISLocalToGlobalMappingCreateIS(pcbddc->is_R_local, &RtoN));
4737: PetscCall(ISGlobalToLocalMappingApply(RtoN, IS_GTOLM_DROP, pcbddc->benign_n, pcbddc->benign_p0_lidx, &n, idxs_p0));
4738: PetscCheck(n == pcbddc->benign_n, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Error in R numbering for benign p0! %" PetscInt_FMT " != %" PetscInt_FMT, n, pcbddc->benign_n);
4739: PetscCall(ISLocalToGlobalMappingDestroy(&RtoN));
4740: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, n, idxs_p0, PETSC_OWN_POINTER, &is_p0));
4741: PetscCall(MatCreateSubMatrix(A_RV, is_p0, NULL, MAT_INITIAL_MATRIX, &A_RV_bcorr));
4742: PetscCall(ISDestroy(&is_p0));
4743: }
4745: PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, lda_rhs, n_eff_vertices, NULL, &A_RRmA_RV));
4746: if (!sparserhs || need_benign_correction) {
4747: if (lda_rhs == n_R && !multi_element) {
4748: PetscCall(MatConvert(A_RV, MATDENSE, MAT_INPLACE_MATRIX, &A_RV));
4749: } else {
4750: Mat T;
4751: PetscScalar *av, *array;
4752: const PetscInt *xadj, *adjncy;
4753: PetscInt n;
4754: PetscBool flg_row;
4756: PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, lda_rhs, n_eff_vertices, NULL, &T));
4757: PetscCall(MatDenseGetArrayWrite(T, &array));
4758: PetscCall(MatConvert(A_RV, MATSEQAIJ, MAT_INPLACE_MATRIX, &A_RV));
4759: PetscCall(MatGetRowIJ(A_RV, 0, PETSC_FALSE, PETSC_FALSE, &n, &xadj, &adjncy, &flg_row));
4760: PetscCall(MatSeqAIJGetArray(A_RV, &av));
4761: for (i = 0; i < n; i++) {
4762: PetscInt j;
4763: for (j = xadj[i]; j < xadj[i + 1]; j++) array[lda_rhs * (V_to_eff_V ? V_to_eff_V[adjncy[j]] : adjncy[j]) + i] = av[j];
4764: }
4765: PetscCall(MatRestoreRowIJ(A_RV, 0, PETSC_FALSE, PETSC_FALSE, &n, &xadj, &adjncy, &flg_row));
4766: PetscCall(MatDenseRestoreArrayWrite(T, &array));
4767: PetscCall(MatDestroy(&A_RV));
4768: A_RV = T;
4769: }
4770: if (need_benign_correction) {
4771: PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
4772: PetscScalar *marr;
4774: /* XXX multi_element */
4775: PetscCall(MatDenseGetArray(A_RV, &marr));
4776: /* need \Phi^T A_RV = (I+L)A_RV, L given by
4778: | 0 0 0 | (V)
4779: L = | 0 0 -1 | (P-p0)
4780: | 0 0 -1 | (p0)
4782: */
4783: for (i = 0; i < reuse_solver->benign_n; i++) {
4784: const PetscScalar *vals;
4785: const PetscInt *idxs, *idxs_zero;
4786: PetscInt n, j, nz;
4788: PetscCall(ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i], &nz));
4789: PetscCall(ISGetIndices(reuse_solver->benign_zerodiag_subs[i], &idxs_zero));
4790: PetscCall(MatGetRow(A_RV_bcorr, i, &n, &idxs, &vals));
4791: for (j = 0; j < n; j++) {
4792: PetscScalar val = vals[j];
4793: PetscInt k, col = idxs[j];
4794: for (k = 0; k < nz; k++) marr[idxs_zero[k] + lda_rhs * col] -= val;
4795: }
4796: PetscCall(MatRestoreRow(A_RV_bcorr, i, &n, &idxs, &vals));
4797: PetscCall(ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i], &idxs_zero));
4798: }
4799: PetscCall(MatDenseRestoreArray(A_RV, &marr));
4800: }
4801: PetscCall(PetscObjectReference((PetscObject)A_RV));
4802: Brhs = A_RV;
4803: } else {
4804: Mat tA_RVT, A_RVT;
4806: if (!pcbddc->symmetric_primal) {
4807: /* A_RV already scaled by -1 */
4808: PetscCall(MatTranspose(A_RV, MAT_INITIAL_MATRIX, &A_RVT));
4809: } else {
4810: restoreavr = PETSC_TRUE;
4811: PetscCall(MatScale(A_VR, -1.0));
4812: PetscCall(PetscObjectReference((PetscObject)A_VR));
4813: A_RVT = A_VR;
4814: }
4815: if (lda_rhs != n_R) {
4816: PetscScalar *aa;
4817: PetscInt r, *ii, *jj;
4818: PetscBool done;
4820: PetscCall(MatGetRowIJ(A_RVT, 0, PETSC_FALSE, PETSC_FALSE, &r, (const PetscInt **)&ii, (const PetscInt **)&jj, &done));
4821: PetscCheck(done, PETSC_COMM_SELF, PETSC_ERR_PLIB, "GetRowIJ failed");
4822: PetscCall(MatSeqAIJGetArray(A_RVT, &aa));
4823: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, n_vertices, lda_rhs, ii, jj, aa, &tA_RVT));
4824: PetscCall(MatRestoreRowIJ(A_RVT, 0, PETSC_FALSE, PETSC_FALSE, &r, (const PetscInt **)&ii, (const PetscInt **)&jj, &done));
4825: PetscCheck(done, PETSC_COMM_SELF, PETSC_ERR_PLIB, "RestoreRowIJ failed");
4826: } else {
4827: PetscCall(PetscObjectReference((PetscObject)A_RVT));
4828: tA_RVT = A_RVT;
4829: }
4830: PetscCall(MatCreateTranspose(tA_RVT, &Brhs));
4831: PetscCall(MatDestroy(&tA_RVT));
4832: PetscCall(MatDestroy(&A_RVT));
4833: }
4834: if (F) {
4835: /* need to correct the rhs */
4836: if (need_benign_correction) {
4837: PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
4838: PetscScalar *marr;
4840: PetscCall(MatDenseGetArray(Brhs, &marr));
4841: if (lda_rhs != n_R) {
4842: for (i = 0; i < n_eff_vertices; i++) {
4843: PetscCall(VecPlaceArray(dummy_vec, marr + i * lda_rhs));
4844: PetscCall(PCBDDCReuseSolversBenignAdapt(reuse_solver, dummy_vec, NULL, PETSC_FALSE, PETSC_TRUE));
4845: PetscCall(VecResetArray(dummy_vec));
4846: }
4847: } else {
4848: for (i = 0; i < n_eff_vertices; i++) {
4849: PetscCall(VecPlaceArray(pcbddc->vec1_R, marr + i * lda_rhs));
4850: PetscCall(PCBDDCReuseSolversBenignAdapt(reuse_solver, pcbddc->vec1_R, NULL, PETSC_FALSE, PETSC_TRUE));
4851: PetscCall(VecResetArray(pcbddc->vec1_R));
4852: }
4853: }
4854: PetscCall(MatDenseRestoreArray(Brhs, &marr));
4855: }
4856: PetscCall(MatMatSolve(F, Brhs, A_RRmA_RV));
4857: if (restoreavr) PetscCall(MatScale(A_VR, -1.0));
4858: /* need to correct the solution */
4859: if (need_benign_correction) {
4860: PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
4861: PetscScalar *marr;
4863: PetscCall(MatDenseGetArray(A_RRmA_RV, &marr));
4864: if (lda_rhs != n_R) {
4865: for (i = 0; i < n_eff_vertices; i++) {
4866: PetscCall(VecPlaceArray(dummy_vec, marr + i * lda_rhs));
4867: PetscCall(PCBDDCReuseSolversBenignAdapt(reuse_solver, dummy_vec, NULL, PETSC_TRUE, PETSC_TRUE));
4868: PetscCall(VecResetArray(dummy_vec));
4869: }
4870: } else {
4871: for (i = 0; i < n_eff_vertices; i++) {
4872: PetscCall(VecPlaceArray(pcbddc->vec1_R, marr + i * lda_rhs));
4873: PetscCall(PCBDDCReuseSolversBenignAdapt(reuse_solver, pcbddc->vec1_R, NULL, PETSC_TRUE, PETSC_TRUE));
4874: PetscCall(VecResetArray(pcbddc->vec1_R));
4875: }
4876: }
4877: PetscCall(MatDenseRestoreArray(A_RRmA_RV, &marr));
4878: }
4879: } else {
4880: const PetscScalar *barr;
4881: PetscScalar *marr;
4883: PetscCall(MatDenseGetArrayRead(Brhs, &barr));
4884: PetscCall(MatDenseGetArray(A_RRmA_RV, &marr));
4885: for (i = 0; i < n_eff_vertices; i++) {
4886: PetscCall(VecPlaceArray(pcbddc->vec1_R, barr + i * lda_rhs));
4887: PetscCall(VecPlaceArray(pcbddc->vec2_R, marr + i * lda_rhs));
4888: PetscCall(KSPSolve(pcbddc->ksp_R, pcbddc->vec1_R, pcbddc->vec2_R));
4889: PetscCall(KSPCheckSolve(pcbddc->ksp_R, pc, pcbddc->vec2_R));
4890: PetscCall(VecResetArray(pcbddc->vec1_R));
4891: PetscCall(VecResetArray(pcbddc->vec2_R));
4892: }
4893: PetscCall(MatDenseRestoreArrayRead(Brhs, &barr));
4894: PetscCall(MatDenseRestoreArray(A_RRmA_RV, &marr));
4895: }
4896: PetscCall(MatDestroy(&A_RV));
4897: PetscCall(MatDestroy(&Brhs));
4898: /* S_VV and S_CV */
4899: if (n_constraints) {
4900: Mat B;
4902: PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, n_B, n_eff_vertices, NULL, &B));
4903: PetscCall(MatDenseScatter_Private(pcbddc->R_to_B, A_RRmA_RV, B, INSERT_VALUES, SCATTER_FORWARD));
4905: /* S_CV = pcbddc->local_auxmat1 * B */
4906: if (multi_element) {
4907: Mat T;
4909: PetscCall(MatCreateSeqAIJFromDenseExpand(B, n_vertices, B_eff_V_J, &T));
4910: PetscCall(MatDestroy(&B));
4911: B = T;
4912: }
4913: PetscCall(MatProductCreate(pcbddc->local_auxmat1, B, NULL, &S_CV));
4914: PetscCall(MatProductSetType(S_CV, MATPRODUCT_AB));
4915: PetscCall(MatProductSetFromOptions(S_CV));
4916: PetscCall(MatProductSymbolic(S_CV));
4917: PetscCall(MatProductNumeric(S_CV));
4918: PetscCall(MatProductClear(S_CV));
4919: PetscCall(MatDestroy(&B));
4921: /* B = local_auxmat2_R * S_CV */
4922: PetscCall(MatProductCreate(local_auxmat2_R, S_CV, NULL, &B));
4923: PetscCall(MatProductSetType(B, MATPRODUCT_AB));
4924: PetscCall(MatProductSetFromOptions(B));
4925: PetscCall(MatProductSymbolic(B));
4926: PetscCall(MatProductNumeric(B));
4928: PetscCall(MatScale(S_CV, m_one));
4929: PetscCall(MatSetValuesSubMat(*coarse_submat, S_CV, n_constraints, idx_C, n_vertices, idx_V, INSERT_VALUES));
4931: if (multi_element) {
4932: Mat T;
4934: PetscCall(MatCreateSeqAIJFromDenseExpand(A_RRmA_RV, n_vertices, R_eff_V_J, &T));
4935: PetscCall(MatDestroy(&A_RRmA_RV));
4936: A_RRmA_RV = T;
4937: }
4938: PetscCall(MatAXPY(A_RRmA_RV, 1.0, B, UNKNOWN_NONZERO_PATTERN)); /* XXX ? */
4939: PetscCall(MatDestroy(&B));
4940: } else if (multi_element) {
4941: Mat T;
4943: PetscCall(MatCreateSeqAIJFromDenseExpand(A_RRmA_RV, n_vertices, R_eff_V_J, &T));
4944: PetscCall(MatDestroy(&A_RRmA_RV));
4945: A_RRmA_RV = T;
4946: }
4948: if (lda_rhs != n_R) {
4949: Mat T;
4951: PetscCall(MatCreateSubMatrix(A_RRmA_RV, is_R, NULL, MAT_INITIAL_MATRIX, &T));
4952: PetscCall(MatDestroy(&A_RRmA_RV));
4953: A_RRmA_RV = T;
4954: }
4956: /* need A_VR * \Phi * A_RRmA_RV = A_VR * (I+L)^T * A_RRmA_RV, L given as before */
4957: if (need_benign_correction) { /* XXX SPARSE */
4958: PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
4959: PetscScalar *sums;
4960: const PetscScalar *marr;
4962: PetscCall(MatDenseGetArrayRead(A_RRmA_RV, &marr));
4963: PetscCall(PetscMalloc1(n_vertices, &sums));
4964: for (i = 0; i < reuse_solver->benign_n; i++) {
4965: const PetscScalar *vals;
4966: const PetscInt *idxs, *idxs_zero;
4967: PetscInt n, j, nz;
4969: PetscCall(ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i], &nz));
4970: PetscCall(ISGetIndices(reuse_solver->benign_zerodiag_subs[i], &idxs_zero));
4971: for (j = 0; j < n_vertices; j++) {
4972: sums[j] = 0.;
4973: for (PetscInt k = 0; k < nz; k++) sums[j] += marr[idxs_zero[k] + j * n_R];
4974: }
4975: PetscCall(MatGetRow(A_RV_bcorr, i, &n, &idxs, &vals));
4976: for (j = 0; j < n; j++) {
4977: PetscScalar val = vals[j];
4978: for (PetscInt k = 0; k < n_vertices; k++) PetscCall(MatSetValue(*coarse_submat, idx_V[idxs[j]], idx_V[k], val * sums[k], ADD_VALUES));
4979: }
4980: PetscCall(MatRestoreRow(A_RV_bcorr, i, &n, &idxs, &vals));
4981: PetscCall(ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i], &idxs_zero));
4982: }
4983: PetscCall(PetscFree(sums));
4984: PetscCall(MatDestroy(&A_RV_bcorr));
4985: PetscCall(MatDenseRestoreArrayRead(A_RRmA_RV, &marr));
4986: }
4988: PetscCall(MatMatMult(A_VR, A_RRmA_RV, MAT_INITIAL_MATRIX, PETSC_DETERMINE, &S_VV));
4989: PetscCall(MatSetValuesSubMat(*coarse_submat, S_VV, n_vertices, idx_V, n_vertices, idx_V, ADD_VALUES));
4990: PetscCall(MatDestroy(&S_VV));
4991: }
4993: /* coarse basis functions */
4994: if (coarse_phi_multi) {
4995: Mat Vid;
4997: PetscCall(MatCreateSeqAIJ(PETSC_COMM_SELF, n_vertices, n_vertices, 1, NULL, &Vid));
4998: PetscCall(MatShift_Basic(Vid, 1.0));
4999: PetscCall(MatNestSetSubMat(coarse_phi_multi, 0, 0, A_RRmA_RV));
5000: PetscCall(MatNestSetSubMat(coarse_phi_multi, 1, 0, Vid));
5001: PetscCall(MatDestroy(&Vid));
5002: } else {
5003: if (A_RRmA_RV) {
5004: Mat B;
5006: PetscCall(MatDenseGetSubMatrix(pcbddc->coarse_phi_B, PETSC_DECIDE, PETSC_DECIDE, 0, n_vertices, &B));
5007: PetscCall(MatDenseScatter_Private(pcbddc->R_to_B, A_RRmA_RV, B, INSERT_VALUES, SCATTER_FORWARD));
5008: PetscCall(MatDenseRestoreSubMatrix(pcbddc->coarse_phi_B, &B));
5009: if (pcbddc->switch_static || pcbddc->dbg_flag) {
5010: PetscCall(MatDenseGetSubMatrix(pcbddc->coarse_phi_D, PETSC_DECIDE, PETSC_DECIDE, 0, n_vertices, &B));
5011: PetscCall(MatDenseScatter_Private(pcbddc->R_to_D, A_RRmA_RV, B, INSERT_VALUES, SCATTER_FORWARD));
5012: PetscCall(MatDenseRestoreSubMatrix(pcbddc->coarse_phi_D, &B));
5013: if (pcbddc->benign_n) {
5014: for (i = 0; i < n_vertices; i++) PetscCall(MatSetValues(pcbddc->coarse_phi_D, pcbddc->benign_n, p0_lidx_I, 1, &i, NULL, INSERT_VALUES));
5015: PetscCall(MatAssemblyBegin(pcbddc->coarse_phi_D, MAT_FINAL_ASSEMBLY));
5016: PetscCall(MatAssemblyEnd(pcbddc->coarse_phi_D, MAT_FINAL_ASSEMBLY));
5017: }
5018: }
5019: }
5020: for (i = 0; i < n_vertices; i++) PetscCall(MatSetValues(pcbddc->coarse_phi_B, 1, &idx_V_B[i], 1, &i, &one, INSERT_VALUES));
5021: PetscCall(MatAssemblyBegin(pcbddc->coarse_phi_B, MAT_FINAL_ASSEMBLY));
5022: PetscCall(MatAssemblyEnd(pcbddc->coarse_phi_B, MAT_FINAL_ASSEMBLY));
5023: }
5024: PetscCall(MatDestroy(&A_RRmA_RV));
5025: }
5026: PetscCall(MatDestroy(&A_RV));
5027: PetscCall(VecDestroy(&dummy_vec));
5029: if (n_constraints) {
5030: Mat B, B2;
5032: PetscCall(MatScale(S_CC, m_one));
5033: PetscCall(MatProductCreate(local_auxmat2_R, S_CC, NULL, &B));
5034: PetscCall(MatProductSetType(B, MATPRODUCT_AB));
5035: PetscCall(MatProductSetFromOptions(B));
5036: PetscCall(MatProductSymbolic(B));
5037: PetscCall(MatProductNumeric(B));
5039: if (n_vertices) {
5040: if (isCHOL || need_benign_correction) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */
5041: PetscCall(MatTranspose(S_CV, MAT_INITIAL_MATRIX, &S_VC));
5042: } else {
5043: if (lda_rhs != n_R) {
5044: Mat tB;
5046: PetscCall(MatCreateSubMatrix(B, is_R, NULL, MAT_INITIAL_MATRIX, &tB));
5047: PetscCall(MatDestroy(&B));
5048: B = tB;
5049: }
5050: PetscCall(MatMatMult(A_VR, B, MAT_INITIAL_MATRIX, PETSC_DETERMINE, &S_VC));
5051: }
5052: PetscCall(MatSetValuesSubMat(*coarse_submat, S_VC, n_vertices, idx_V, n_constraints, idx_C, INSERT_VALUES));
5053: }
5055: /* coarse basis functions */
5056: if (coarse_phi_multi) {
5057: PetscCall(MatNestSetSubMat(coarse_phi_multi, 0, 1, B));
5058: } else {
5059: PetscCall(MatDenseGetSubMatrix(pcbddc->coarse_phi_B, PETSC_DECIDE, PETSC_DECIDE, n_vertices, n_vertices + n_constraints, &B2));
5060: PetscCall(MatDenseScatter_Private(pcbddc->R_to_B, B, B2, INSERT_VALUES, SCATTER_FORWARD));
5061: PetscCall(MatDenseRestoreSubMatrix(pcbddc->coarse_phi_B, &B2));
5062: if (pcbddc->switch_static || pcbddc->dbg_flag) {
5063: PetscCall(MatDenseGetSubMatrix(pcbddc->coarse_phi_D, PETSC_DECIDE, PETSC_DECIDE, n_vertices, n_vertices + n_constraints, &B2));
5064: PetscCall(MatDenseScatter_Private(pcbddc->R_to_D, B, B2, INSERT_VALUES, SCATTER_FORWARD));
5065: if (pcbddc->benign_n) {
5066: for (i = 0; i < n_constraints; i++) PetscCall(MatSetValues(B2, pcbddc->benign_n, p0_lidx_I, 1, &i, NULL, INSERT_VALUES));
5067: }
5068: PetscCall(MatDenseRestoreSubMatrix(pcbddc->coarse_phi_D, &B2));
5069: }
5070: }
5071: PetscCall(MatDestroy(&B));
5072: }
5074: /* assemble sparse coarse basis functions */
5075: if (coarse_phi_multi) {
5076: Mat T;
5078: PetscCall(MatConvert(coarse_phi_multi, MATSEQAIJ, MAT_INITIAL_MATRIX, &T));
5079: PetscCall(MatDestroy(&coarse_phi_multi));
5080: PetscCall(MatCreateSubMatrix(T, pcis->is_B_local, NULL, MAT_INITIAL_MATRIX, &pcbddc->coarse_phi_B));
5081: if (pcbddc->switch_static || pcbddc->dbg_flag) PetscCall(MatCreateSubMatrix(T, pcis->is_I_local, NULL, MAT_INITIAL_MATRIX, &pcbddc->coarse_phi_D));
5082: PetscCall(MatDestroy(&T));
5083: }
5084: PetscCall(MatDestroy(&local_auxmat2_R));
5085: PetscCall(PetscFree(p0_lidx_I));
5087: /* coarse matrix entries relative to B_0 */
5088: if (pcbddc->benign_n) {
5089: Mat B0_B, B0_BPHI;
5090: IS is_dummy;
5091: const PetscScalar *data;
5092: PetscInt j;
5094: PetscCall(ISCreateStride(PETSC_COMM_SELF, pcbddc->benign_n, 0, 1, &is_dummy));
5095: PetscCall(MatCreateSubMatrix(pcbddc->benign_B0, is_dummy, pcis->is_B_local, MAT_INITIAL_MATRIX, &B0_B));
5096: PetscCall(ISDestroy(&is_dummy));
5097: PetscCall(MatMatMult(B0_B, pcbddc->coarse_phi_B, MAT_INITIAL_MATRIX, 1.0, &B0_BPHI));
5098: PetscCall(MatConvert(B0_BPHI, MATSEQDENSE, MAT_INPLACE_MATRIX, &B0_BPHI));
5099: PetscCall(MatDenseGetArrayRead(B0_BPHI, &data));
5100: for (j = 0; j < pcbddc->benign_n; j++) {
5101: PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j;
5102: for (i = 0; i < pcbddc->local_primal_size; i++) {
5103: PetscCall(MatSetValue(*coarse_submat, primal_idx, i, data[i * pcbddc->benign_n + j], INSERT_VALUES));
5104: PetscCall(MatSetValue(*coarse_submat, i, primal_idx, data[i * pcbddc->benign_n + j], INSERT_VALUES));
5105: }
5106: }
5107: PetscCall(MatDenseRestoreArrayRead(B0_BPHI, &data));
5108: PetscCall(MatDestroy(&B0_B));
5109: PetscCall(MatDestroy(&B0_BPHI));
5110: }
5112: /* compute other basis functions for non-symmetric problems */
5113: if (!pcbddc->symmetric_primal) {
5114: Mat B_V = NULL, B_C = NULL;
5115: PetscScalar *marray, *work;
5117: /* TODO multi_element MatDenseScatter */
5118: if (n_constraints) {
5119: Mat S_CCT, C_CRT;
5121: PetscCall(MatScale(S_CC, m_one));
5122: PetscCall(MatTranspose(C_CR, MAT_INITIAL_MATRIX, &C_CRT));
5123: PetscCall(MatTranspose(S_CC, MAT_INITIAL_MATRIX, &S_CCT));
5124: PetscCall(MatMatMult(C_CRT, S_CCT, MAT_INITIAL_MATRIX, PETSC_DETERMINE, &B_C));
5125: PetscCall(MatConvert(B_C, MATDENSE, MAT_INPLACE_MATRIX, &B_C));
5126: PetscCall(MatDestroy(&S_CCT));
5127: if (n_vertices) {
5128: Mat S_VCT;
5130: PetscCall(MatTranspose(S_VC, MAT_INITIAL_MATRIX, &S_VCT));
5131: PetscCall(MatMatMult(C_CRT, S_VCT, MAT_INITIAL_MATRIX, PETSC_DETERMINE, &B_V));
5132: PetscCall(MatDestroy(&S_VCT));
5133: PetscCall(MatConvert(B_V, MATDENSE, MAT_INPLACE_MATRIX, &B_V));
5134: }
5135: PetscCall(MatDestroy(&C_CRT));
5136: } else {
5137: PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, n_R, n_vertices, NULL, &B_V));
5138: }
5139: if (n_vertices && n_R) {
5140: PetscScalar *av, *marray;
5141: const PetscInt *xadj, *adjncy;
5142: PetscInt n;
5143: PetscBool flg_row;
5145: /* B_V = B_V - A_VR^T */
5146: PetscCall(MatConvert(A_VR, MATSEQAIJ, MAT_INPLACE_MATRIX, &A_VR));
5147: PetscCall(MatGetRowIJ(A_VR, 0, PETSC_FALSE, PETSC_FALSE, &n, &xadj, &adjncy, &flg_row));
5148: PetscCall(MatSeqAIJGetArray(A_VR, &av));
5149: PetscCall(MatDenseGetArray(B_V, &marray));
5150: for (i = 0; i < n; i++) {
5151: PetscInt j;
5152: for (j = xadj[i]; j < xadj[i + 1]; j++) marray[i * n_R + adjncy[j]] -= av[j];
5153: }
5154: PetscCall(MatDenseRestoreArray(B_V, &marray));
5155: PetscCall(MatRestoreRowIJ(A_VR, 0, PETSC_FALSE, PETSC_FALSE, &n, &xadj, &adjncy, &flg_row));
5156: PetscCall(MatDestroy(&A_VR));
5157: }
5159: /* currently there's no support for MatTransposeMatSolve(F,B,X) */
5160: PetscCall(PetscMalloc1(n_R * pcbddc->local_primal_size, &work));
5161: if (n_vertices) {
5162: PetscCall(MatDenseGetArray(B_V, &marray));
5163: for (i = 0; i < n_vertices; i++) {
5164: PetscCall(VecPlaceArray(pcbddc->vec1_R, marray + i * n_R));
5165: PetscCall(VecPlaceArray(pcbddc->vec2_R, work + i * n_R));
5166: PetscCall(KSPSolveTranspose(pcbddc->ksp_R, pcbddc->vec1_R, pcbddc->vec2_R));
5167: PetscCall(KSPCheckSolve(pcbddc->ksp_R, pc, pcbddc->vec2_R));
5168: PetscCall(VecResetArray(pcbddc->vec1_R));
5169: PetscCall(VecResetArray(pcbddc->vec2_R));
5170: }
5171: PetscCall(MatDenseRestoreArray(B_V, &marray));
5172: }
5173: if (B_C) {
5174: PetscCall(MatDenseGetArray(B_C, &marray));
5175: for (i = n_vertices; i < n_constraints + n_vertices; i++) {
5176: PetscCall(VecPlaceArray(pcbddc->vec1_R, marray + (i - n_vertices) * n_R));
5177: PetscCall(VecPlaceArray(pcbddc->vec2_R, work + i * n_R));
5178: PetscCall(KSPSolveTranspose(pcbddc->ksp_R, pcbddc->vec1_R, pcbddc->vec2_R));
5179: PetscCall(KSPCheckSolve(pcbddc->ksp_R, pc, pcbddc->vec2_R));
5180: PetscCall(VecResetArray(pcbddc->vec1_R));
5181: PetscCall(VecResetArray(pcbddc->vec2_R));
5182: }
5183: PetscCall(MatDenseRestoreArray(B_C, &marray));
5184: }
5185: /* coarse basis functions */
5186: PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, n_B, pcbddc->local_primal_size, NULL, &pcbddc->coarse_psi_B));
5187: if (pcbddc->switch_static || pcbddc->dbg_flag) PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, n_D, pcbddc->local_primal_size, NULL, &pcbddc->coarse_psi_D));
5188: for (i = 0; i < pcbddc->local_primal_size; i++) {
5189: Vec v;
5191: PetscCall(VecPlaceArray(pcbddc->vec1_R, work + i * n_R));
5192: PetscCall(MatDenseGetColumnVec(pcbddc->coarse_psi_B, i, &v));
5193: PetscCall(VecScatterBegin(pcbddc->R_to_B, pcbddc->vec1_R, v, INSERT_VALUES, SCATTER_FORWARD));
5194: PetscCall(VecScatterEnd(pcbddc->R_to_B, pcbddc->vec1_R, v, INSERT_VALUES, SCATTER_FORWARD));
5195: if (i < n_vertices) {
5196: PetscScalar one = 1.0;
5197: PetscCall(VecSetValues(v, 1, &idx_V_B[i], &one, INSERT_VALUES));
5198: PetscCall(VecAssemblyBegin(v));
5199: PetscCall(VecAssemblyEnd(v));
5200: }
5201: PetscCall(MatDenseRestoreColumnVec(pcbddc->coarse_psi_B, i, &v));
5203: if (pcbddc->switch_static || pcbddc->dbg_flag) {
5204: PetscCall(MatDenseGetColumnVec(pcbddc->coarse_psi_D, i, &v));
5205: PetscCall(VecScatterBegin(pcbddc->R_to_D, pcbddc->vec1_R, v, INSERT_VALUES, SCATTER_FORWARD));
5206: PetscCall(VecScatterEnd(pcbddc->R_to_D, pcbddc->vec1_R, v, INSERT_VALUES, SCATTER_FORWARD));
5207: PetscCall(MatDenseRestoreColumnVec(pcbddc->coarse_psi_D, i, &v));
5208: }
5209: PetscCall(VecResetArray(pcbddc->vec1_R));
5210: }
5211: PetscCall(MatDestroy(&B_V));
5212: PetscCall(MatDestroy(&B_C));
5213: PetscCall(PetscFree(work));
5214: } else {
5215: PetscCall(PetscObjectReference((PetscObject)pcbddc->coarse_phi_B));
5216: pcbddc->coarse_psi_B = pcbddc->coarse_phi_B;
5217: PetscCall(PetscObjectReference((PetscObject)pcbddc->coarse_phi_D));
5218: pcbddc->coarse_psi_D = pcbddc->coarse_phi_D;
5219: }
5220: PetscCall(MatAssemblyBegin(*coarse_submat, MAT_FINAL_ASSEMBLY));
5221: PetscCall(MatAssemblyEnd(*coarse_submat, MAT_FINAL_ASSEMBLY));
5223: /* free memory */
5224: PetscCall(PetscFree(V_to_eff_V));
5225: PetscCall(PetscFree(C_to_eff_C));
5226: PetscCall(PetscFree(R_eff_V_J));
5227: PetscCall(PetscFree(R_eff_C_J));
5228: PetscCall(PetscFree(B_eff_V_J));
5229: PetscCall(PetscFree(B_eff_C_J));
5230: PetscCall(ISDestroy(&is_R));
5231: PetscCall(ISRestoreIndices(is_V, &idx_V));
5232: PetscCall(ISRestoreIndices(is_C, &idx_C));
5233: PetscCall(ISDestroy(&is_V));
5234: PetscCall(ISDestroy(&is_C));
5235: PetscCall(PetscFree(idx_V_B));
5236: PetscCall(MatDestroy(&S_CV));
5237: PetscCall(MatDestroy(&S_VC));
5238: PetscCall(MatDestroy(&S_CC));
5239: if (n_vertices) PetscCall(MatDestroy(&A_VR));
5240: if (n_constraints) PetscCall(MatDestroy(&C_CR));
5241: PetscCall(PetscLogEventEnd(PC_BDDC_CorrectionSetUp[pcbddc->current_level], pc, 0, 0, 0));
5243: /* Checking coarse_sub_mat and coarse basis functions */
5244: /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */
5245: /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */
5246: if (pcbddc->dbg_flag) {
5247: Mat AUXMAT, TM1, TM2, TM3, TM4;
5248: Mat coarse_phi_D, coarse_phi_B;
5249: Mat coarse_psi_D, coarse_psi_B;
5250: Mat A_II, A_BB, A_IB, A_BI;
5251: Mat C_B, CPHI;
5252: IS is_dummy;
5253: Vec mones;
5254: MatType checkmattype = MATSEQAIJ;
5255: PetscReal real_value;
5257: if (pcbddc->benign_n && !pcbddc->benign_change_explicit) {
5258: Mat A;
5259: PetscCall(PCBDDCBenignProject(pc, NULL, NULL, &A));
5260: PetscCall(MatCreateSubMatrix(A, pcis->is_I_local, pcis->is_I_local, MAT_INITIAL_MATRIX, &A_II));
5261: PetscCall(MatCreateSubMatrix(A, pcis->is_I_local, pcis->is_B_local, MAT_INITIAL_MATRIX, &A_IB));
5262: PetscCall(MatCreateSubMatrix(A, pcis->is_B_local, pcis->is_I_local, MAT_INITIAL_MATRIX, &A_BI));
5263: PetscCall(MatCreateSubMatrix(A, pcis->is_B_local, pcis->is_B_local, MAT_INITIAL_MATRIX, &A_BB));
5264: PetscCall(MatDestroy(&A));
5265: } else {
5266: PetscCall(MatConvert(pcis->A_II, checkmattype, MAT_INITIAL_MATRIX, &A_II));
5267: PetscCall(MatConvert(pcis->A_IB, checkmattype, MAT_INITIAL_MATRIX, &A_IB));
5268: PetscCall(MatConvert(pcis->A_BI, checkmattype, MAT_INITIAL_MATRIX, &A_BI));
5269: PetscCall(MatConvert(pcis->A_BB, checkmattype, MAT_INITIAL_MATRIX, &A_BB));
5270: }
5271: PetscCall(MatConvert(pcbddc->coarse_phi_D, checkmattype, MAT_INITIAL_MATRIX, &coarse_phi_D));
5272: PetscCall(MatConvert(pcbddc->coarse_phi_B, checkmattype, MAT_INITIAL_MATRIX, &coarse_phi_B));
5273: if (!pcbddc->symmetric_primal) {
5274: PetscCall(MatConvert(pcbddc->coarse_psi_D, checkmattype, MAT_INITIAL_MATRIX, &coarse_psi_D));
5275: PetscCall(MatConvert(pcbddc->coarse_psi_B, checkmattype, MAT_INITIAL_MATRIX, &coarse_psi_B));
5276: }
5277: PetscCall(PetscViewerASCIIPrintf(pcbddc->dbg_viewer, "--------------------------------------------------\n"));
5278: PetscCall(PetscViewerASCIIPrintf(pcbddc->dbg_viewer, "Check coarse sub mat computation (symmetric %d)\n", pcbddc->symmetric_primal));
5279: PetscCall(PetscViewerFlush(pcbddc->dbg_viewer));
5280: if (!pcbddc->symmetric_primal) {
5281: PetscCall(MatMatMult(A_II, coarse_phi_D, MAT_INITIAL_MATRIX, 1.0, &AUXMAT));
5282: PetscCall(MatTransposeMatMult(coarse_psi_D, AUXMAT, MAT_INITIAL_MATRIX, 1.0, &TM1));
5283: PetscCall(MatDestroy(&AUXMAT));
5284: PetscCall(MatMatMult(A_BB, coarse_phi_B, MAT_INITIAL_MATRIX, 1.0, &AUXMAT));
5285: PetscCall(MatTransposeMatMult(coarse_psi_B, AUXMAT, MAT_INITIAL_MATRIX, 1.0, &TM2));
5286: PetscCall(MatDestroy(&AUXMAT));
5287: PetscCall(MatMatMult(A_IB, coarse_phi_B, MAT_INITIAL_MATRIX, 1.0, &AUXMAT));
5288: PetscCall(MatTransposeMatMult(coarse_psi_D, AUXMAT, MAT_INITIAL_MATRIX, 1.0, &TM3));
5289: PetscCall(MatDestroy(&AUXMAT));
5290: PetscCall(MatMatMult(A_BI, coarse_phi_D, MAT_INITIAL_MATRIX, 1.0, &AUXMAT));
5291: PetscCall(MatTransposeMatMult(coarse_psi_B, AUXMAT, MAT_INITIAL_MATRIX, 1.0, &TM4));
5292: PetscCall(MatDestroy(&AUXMAT));
5293: } else {
5294: PetscCall(MatPtAP(A_II, coarse_phi_D, MAT_INITIAL_MATRIX, 1.0, &TM1));
5295: PetscCall(MatPtAP(A_BB, coarse_phi_B, MAT_INITIAL_MATRIX, 1.0, &TM2));
5296: PetscCall(MatMatMult(A_IB, coarse_phi_B, MAT_INITIAL_MATRIX, 1.0, &AUXMAT));
5297: PetscCall(MatTransposeMatMult(coarse_phi_D, AUXMAT, MAT_INITIAL_MATRIX, 1.0, &TM3));
5298: PetscCall(MatDestroy(&AUXMAT));
5299: PetscCall(MatMatMult(A_BI, coarse_phi_D, MAT_INITIAL_MATRIX, 1.0, &AUXMAT));
5300: PetscCall(MatTransposeMatMult(coarse_phi_B, AUXMAT, MAT_INITIAL_MATRIX, 1.0, &TM4));
5301: PetscCall(MatDestroy(&AUXMAT));
5302: }
5303: PetscCall(MatAXPY(TM1, one, TM2, DIFFERENT_NONZERO_PATTERN));
5304: PetscCall(MatAXPY(TM1, one, TM3, DIFFERENT_NONZERO_PATTERN));
5305: PetscCall(MatAXPY(TM1, one, TM4, DIFFERENT_NONZERO_PATTERN));
5306: PetscCall(MatConvert(TM1, MATSEQDENSE, MAT_INPLACE_MATRIX, &TM1));
5307: if (pcbddc->benign_n) {
5308: Mat B0_B, B0_BPHI;
5309: const PetscScalar *data2;
5310: PetscScalar *data;
5311: PetscInt j;
5313: PetscCall(ISCreateStride(PETSC_COMM_SELF, pcbddc->benign_n, 0, 1, &is_dummy));
5314: PetscCall(MatCreateSubMatrix(pcbddc->benign_B0, is_dummy, pcis->is_B_local, MAT_INITIAL_MATRIX, &B0_B));
5315: PetscCall(MatMatMult(B0_B, coarse_phi_B, MAT_INITIAL_MATRIX, 1.0, &B0_BPHI));
5316: PetscCall(MatConvert(B0_BPHI, MATSEQDENSE, MAT_INPLACE_MATRIX, &B0_BPHI));
5317: PetscCall(MatDenseGetArray(TM1, &data));
5318: PetscCall(MatDenseGetArrayRead(B0_BPHI, &data2));
5319: for (j = 0; j < pcbddc->benign_n; j++) {
5320: PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j;
5321: for (i = 0; i < pcbddc->local_primal_size; i++) {
5322: data[primal_idx * pcbddc->local_primal_size + i] += data2[i * pcbddc->benign_n + j];
5323: data[i * pcbddc->local_primal_size + primal_idx] += data2[i * pcbddc->benign_n + j];
5324: }
5325: }
5326: PetscCall(MatDenseRestoreArray(TM1, &data));
5327: PetscCall(MatDenseRestoreArrayRead(B0_BPHI, &data2));
5328: PetscCall(MatDestroy(&B0_B));
5329: PetscCall(ISDestroy(&is_dummy));
5330: PetscCall(MatDestroy(&B0_BPHI));
5331: }
5332: PetscCall(MatAXPY(TM1, m_one, *coarse_submat, DIFFERENT_NONZERO_PATTERN));
5333: PetscCall(MatNorm(TM1, NORM_FROBENIUS, &real_value));
5334: PetscCall(PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer));
5335: PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, "Subdomain %04d matrix error % 1.14e\n", PetscGlobalRank, (double)real_value));
5337: /* check constraints */
5338: PetscCall(ISCreateStride(PETSC_COMM_SELF, pcbddc->local_primal_size - pcbddc->benign_n, 0, 1, &is_dummy));
5339: PetscCall(MatCreateSubMatrix(pcbddc->ConstraintMatrix, is_dummy, pcis->is_B_local, MAT_INITIAL_MATRIX, &C_B));
5340: if (!pcbddc->benign_n) { /* TODO: add benign case */
5341: PetscCall(MatMatMult(C_B, coarse_phi_B, MAT_INITIAL_MATRIX, 1.0, &CPHI));
5342: } else {
5343: PetscScalar *data;
5344: Mat tmat;
5345: PetscCall(MatDenseGetArray(pcbddc->coarse_phi_B, &data));
5346: PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, pcis->n_B, pcbddc->local_primal_size - pcbddc->benign_n, data, &tmat));
5347: PetscCall(MatDenseRestoreArray(pcbddc->coarse_phi_B, &data));
5348: PetscCall(MatMatMult(C_B, tmat, MAT_INITIAL_MATRIX, 1.0, &CPHI));
5349: PetscCall(MatDestroy(&tmat));
5350: }
5351: PetscCall(MatCreateVecs(CPHI, &mones, NULL));
5352: PetscCall(VecSet(mones, -1.0));
5353: PetscCall(MatDiagonalSet(CPHI, mones, ADD_VALUES));
5354: PetscCall(MatNorm(CPHI, NORM_FROBENIUS, &real_value));
5355: PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, "Subdomain %04d phi constraints error % 1.14e\n", PetscGlobalRank, (double)real_value));
5356: if (!pcbddc->symmetric_primal) {
5357: PetscCall(MatMatMult(C_B, coarse_psi_B, MAT_REUSE_MATRIX, 1.0, &CPHI));
5358: PetscCall(VecSet(mones, -1.0));
5359: PetscCall(MatDiagonalSet(CPHI, mones, ADD_VALUES));
5360: PetscCall(MatNorm(CPHI, NORM_FROBENIUS, &real_value));
5361: PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, "Subdomain %04d psi constraints error % 1.14e\n", PetscGlobalRank, (double)real_value));
5362: }
5363: PetscCall(MatDestroy(&C_B));
5364: PetscCall(MatDestroy(&CPHI));
5365: PetscCall(ISDestroy(&is_dummy));
5366: PetscCall(VecDestroy(&mones));
5367: PetscCall(PetscViewerFlush(pcbddc->dbg_viewer));
5368: PetscCall(MatDestroy(&A_II));
5369: PetscCall(MatDestroy(&A_BB));
5370: PetscCall(MatDestroy(&A_IB));
5371: PetscCall(MatDestroy(&A_BI));
5372: PetscCall(MatDestroy(&TM1));
5373: PetscCall(MatDestroy(&TM2));
5374: PetscCall(MatDestroy(&TM3));
5375: PetscCall(MatDestroy(&TM4));
5376: PetscCall(MatDestroy(&coarse_phi_D));
5377: PetscCall(MatDestroy(&coarse_phi_B));
5378: if (!pcbddc->symmetric_primal) {
5379: PetscCall(MatDestroy(&coarse_psi_D));
5380: PetscCall(MatDestroy(&coarse_psi_B));
5381: }
5382: }
5384: #if 0
5385: {
5386: PetscViewer viewer;
5387: char filename[256];
5389: PetscCall(PetscSNPrintf(filename, PETSC_STATIC_ARRAY_LENGTH(filename), "details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level));
5390: PetscCall(PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer));
5391: PetscCall(PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB));
5392: PetscCall(PetscObjectSetName((PetscObject)*coarse_submat,"coarse submat"));
5393: PetscCall(MatView(*coarse_submat,viewer));
5394: if (pcbddc->coarse_phi_B) {
5395: PetscCall(PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B"));
5396: PetscCall(MatView(pcbddc->coarse_phi_B,viewer));
5397: }
5398: if (pcbddc->coarse_phi_D) {
5399: PetscCall(PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D"));
5400: PetscCall(MatView(pcbddc->coarse_phi_D,viewer));
5401: }
5402: if (pcbddc->coarse_psi_B) {
5403: PetscCall(PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B"));
5404: PetscCall(MatView(pcbddc->coarse_psi_B,viewer));
5405: }
5406: if (pcbddc->coarse_psi_D) {
5407: PetscCall(PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D"));
5408: PetscCall(MatView(pcbddc->coarse_psi_D,viewer));
5409: }
5410: PetscCall(PetscObjectSetName((PetscObject)pcbddc->local_mat,"A"));
5411: PetscCall(MatView(pcbddc->local_mat,viewer));
5412: PetscCall(PetscObjectSetName((PetscObject)pcbddc->ConstraintMatrix,"C"));
5413: PetscCall(MatView(pcbddc->ConstraintMatrix,viewer));
5414: PetscCall(PetscObjectSetName((PetscObject)pcis->is_I_local,"I"));
5415: PetscCall(ISView(pcis->is_I_local,viewer));
5416: PetscCall(PetscObjectSetName((PetscObject)pcis->is_B_local,"B"));
5417: PetscCall(ISView(pcis->is_B_local,viewer));
5418: PetscCall(PetscObjectSetName((PetscObject)pcbddc->is_R_local,"R"));
5419: PetscCall(ISView(pcbddc->is_R_local,viewer));
5420: PetscCall(PetscViewerDestroy(&viewer));
5421: }
5422: #endif
5424: /* device support */
5425: {
5426: PetscBool iscuda, iship, iskokkos;
5427: MatType mtype = NULL;
5429: PetscCall(PetscObjectTypeCompareAny((PetscObject)pcis->vec1_N, &iscuda, VECCUDA, VECMPICUDA, VECSEQCUDA, ""));
5430: PetscCall(PetscObjectTypeCompareAny((PetscObject)pcis->vec1_N, &iship, VECHIP, VECMPIHIP, VECSEQHIP, ""));
5431: PetscCall(PetscObjectTypeCompareAny((PetscObject)pcis->vec1_N, &iskokkos, VECKOKKOS, VECMPIKOKKOS, VECSEQKOKKOS, ""));
5432: if (iskokkos) {
5433: if (PetscDefined(HAVE_MACRO_KOKKOS_ENABLE_CUDA)) iscuda = PETSC_TRUE;
5434: else if (PetscDefined(HAVE_MACRO_KOKKOS_ENABLE_HIP)) iship = PETSC_TRUE;
5435: }
5436: if (iskokkos) mtype = multi_element ? MATSEQAIJKOKKOS : (iscuda ? MATSEQDENSECUDA : MATSEQDENSEHIP);
5437: else if (iship) mtype = multi_element ? MATSEQAIJHIPSPARSE : MATSEQDENSEHIP;
5438: else if (iscuda) mtype = multi_element ? MATSEQAIJCUSPARSE : MATSEQDENSECUDA;
5439: if (mtype) {
5440: if (pcbddc->local_auxmat1) PetscCall(MatConvert(pcbddc->local_auxmat1, mtype, MAT_INPLACE_MATRIX, &pcbddc->local_auxmat1));
5441: if (pcbddc->local_auxmat2) PetscCall(MatConvert(pcbddc->local_auxmat2, mtype, MAT_INPLACE_MATRIX, &pcbddc->local_auxmat2));
5442: if (pcbddc->coarse_phi_B) PetscCall(MatConvert(pcbddc->coarse_phi_B, mtype, MAT_INPLACE_MATRIX, &pcbddc->coarse_phi_B));
5443: if (pcbddc->coarse_phi_D) PetscCall(MatConvert(pcbddc->coarse_phi_D, mtype, MAT_INPLACE_MATRIX, &pcbddc->coarse_phi_D));
5444: if (pcbddc->coarse_psi_B) PetscCall(MatConvert(pcbddc->coarse_psi_B, mtype, MAT_INPLACE_MATRIX, &pcbddc->coarse_psi_B));
5445: if (pcbddc->coarse_psi_D) PetscCall(MatConvert(pcbddc->coarse_psi_D, mtype, MAT_INPLACE_MATRIX, &pcbddc->coarse_psi_D));
5446: }
5447: }
5448: PetscFunctionReturn(PETSC_SUCCESS);
5449: }
5451: PetscErrorCode MatCreateSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat *B)
5452: {
5453: Mat *work_mat;
5454: IS isrow_s, iscol_s;
5455: PetscBool rsorted, csorted;
5456: PetscInt rsize, *idxs_perm_r = NULL, csize, *idxs_perm_c = NULL;
5458: PetscFunctionBegin;
5459: PetscCall(ISSorted(isrow, &rsorted));
5460: PetscCall(ISSorted(iscol, &csorted));
5461: PetscCall(ISGetLocalSize(isrow, &rsize));
5462: PetscCall(ISGetLocalSize(iscol, &csize));
5464: if (!rsorted) {
5465: const PetscInt *idxs;
5466: PetscInt *idxs_sorted, i;
5468: PetscCall(PetscMalloc1(rsize, &idxs_perm_r));
5469: PetscCall(PetscMalloc1(rsize, &idxs_sorted));
5470: for (i = 0; i < rsize; i++) idxs_perm_r[i] = i;
5471: PetscCall(ISGetIndices(isrow, &idxs));
5472: PetscCall(PetscSortIntWithPermutation(rsize, idxs, idxs_perm_r));
5473: for (i = 0; i < rsize; i++) idxs_sorted[i] = idxs[idxs_perm_r[i]];
5474: PetscCall(ISRestoreIndices(isrow, &idxs));
5475: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, rsize, idxs_sorted, PETSC_OWN_POINTER, &isrow_s));
5476: } else {
5477: PetscCall(PetscObjectReference((PetscObject)isrow));
5478: isrow_s = isrow;
5479: }
5481: if (!csorted) {
5482: if (isrow == iscol) {
5483: PetscCall(PetscObjectReference((PetscObject)isrow_s));
5484: iscol_s = isrow_s;
5485: } else {
5486: const PetscInt *idxs;
5487: PetscInt *idxs_sorted, i;
5489: PetscCall(PetscMalloc1(csize, &idxs_perm_c));
5490: PetscCall(PetscMalloc1(csize, &idxs_sorted));
5491: for (i = 0; i < csize; i++) idxs_perm_c[i] = i;
5492: PetscCall(ISGetIndices(iscol, &idxs));
5493: PetscCall(PetscSortIntWithPermutation(csize, idxs, idxs_perm_c));
5494: for (i = 0; i < csize; i++) idxs_sorted[i] = idxs[idxs_perm_c[i]];
5495: PetscCall(ISRestoreIndices(iscol, &idxs));
5496: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, csize, idxs_sorted, PETSC_OWN_POINTER, &iscol_s));
5497: }
5498: } else {
5499: PetscCall(PetscObjectReference((PetscObject)iscol));
5500: iscol_s = iscol;
5501: }
5503: PetscCall(MatCreateSubMatrices(A, 1, &isrow_s, &iscol_s, MAT_INITIAL_MATRIX, &work_mat));
5505: if (!rsorted || !csorted) {
5506: Mat new_mat;
5507: IS is_perm_r, is_perm_c;
5509: if (!rsorted) {
5510: PetscInt *idxs_r, i;
5511: PetscCall(PetscMalloc1(rsize, &idxs_r));
5512: for (i = 0; i < rsize; i++) idxs_r[idxs_perm_r[i]] = i;
5513: PetscCall(PetscFree(idxs_perm_r));
5514: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, rsize, idxs_r, PETSC_OWN_POINTER, &is_perm_r));
5515: } else {
5516: PetscCall(ISCreateStride(PETSC_COMM_SELF, rsize, 0, 1, &is_perm_r));
5517: }
5518: PetscCall(ISSetPermutation(is_perm_r));
5520: if (!csorted) {
5521: if (isrow_s == iscol_s) {
5522: PetscCall(PetscObjectReference((PetscObject)is_perm_r));
5523: is_perm_c = is_perm_r;
5524: } else {
5525: PetscInt *idxs_c, i;
5526: PetscCheck(idxs_perm_c, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Permutation array not present");
5527: PetscCall(PetscMalloc1(csize, &idxs_c));
5528: for (i = 0; i < csize; i++) idxs_c[idxs_perm_c[i]] = i;
5529: PetscCall(PetscFree(idxs_perm_c));
5530: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, csize, idxs_c, PETSC_OWN_POINTER, &is_perm_c));
5531: }
5532: } else {
5533: PetscCall(ISCreateStride(PETSC_COMM_SELF, csize, 0, 1, &is_perm_c));
5534: }
5535: PetscCall(ISSetPermutation(is_perm_c));
5537: PetscCall(MatPermute(work_mat[0], is_perm_r, is_perm_c, &new_mat));
5538: PetscCall(MatDestroy(&work_mat[0]));
5539: work_mat[0] = new_mat;
5540: PetscCall(ISDestroy(&is_perm_r));
5541: PetscCall(ISDestroy(&is_perm_c));
5542: }
5544: PetscCall(PetscObjectReference((PetscObject)work_mat[0]));
5545: *B = work_mat[0];
5546: PetscCall(MatDestroyMatrices(1, &work_mat));
5547: PetscCall(ISDestroy(&isrow_s));
5548: PetscCall(ISDestroy(&iscol_s));
5549: PetscFunctionReturn(PETSC_SUCCESS);
5550: }
5552: static PetscErrorCode MatPtAPWithPrefix_Private(Mat A, Mat P, PetscReal fill, const char *prefix, Mat *C)
5553: {
5554: PetscFunctionBegin;
5555: PetscCall(MatProductCreate(A, P, NULL, C));
5556: PetscCall(MatProductSetType(*C, MATPRODUCT_PtAP));
5557: PetscCall(MatProductSetAlgorithm(*C, "default"));
5558: PetscCall(MatProductSetFill(*C, fill));
5559: PetscCall(MatSetOptionsPrefix(*C, prefix));
5560: PetscCall(MatProductSetFromOptions(*C));
5561: PetscCall(MatProductSymbolic(*C));
5562: PetscCall(MatProductNumeric(*C));
5563: (*C)->symmetric = A->symmetric;
5564: (*C)->spd = A->spd;
5565: PetscFunctionReturn(PETSC_SUCCESS);
5566: }
5568: PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix)
5569: {
5570: Mat_IS *matis = (Mat_IS *)pc->pmat->data;
5571: PC_BDDC *pcbddc = (PC_BDDC *)pc->data;
5572: Mat new_mat, lA;
5573: IS is_local, is_global;
5574: PetscInt local_size;
5575: PetscBool isseqaij, issym, isset;
5576: char ptapprefix[256];
5578: PetscFunctionBegin;
5579: PetscCall(MatDestroy(&pcbddc->local_mat));
5580: PetscCall(MatGetSize(matis->A, &local_size, NULL));
5581: if (pcbddc->mat_graph->multi_element) {
5582: Mat *mats, *bdiags;
5583: IS *gsubs;
5584: PetscInt nsubs = pcbddc->n_local_subs;
5586: PetscCall(PetscCalloc1(nsubs * nsubs, &mats));
5587: #if 1
5588: PetscCall(PetscMalloc1(nsubs, &gsubs));
5589: for (PetscInt i = 0; i < nsubs; i++) PetscCall(ISLocalToGlobalMappingApplyIS(matis->rmapping, pcbddc->local_subs[i], &gsubs[i]));
5590: PetscCall(MatCreateSubMatrices(ChangeOfBasisMatrix, nsubs, gsubs, gsubs, MAT_INITIAL_MATRIX, &bdiags));
5591: for (PetscInt i = 0; i < nsubs; i++) PetscCall(ISDestroy(&gsubs[i]));
5592: PetscCall(PetscFree(gsubs));
5593: #else /* this does not work since MatCreateSubMatrices does not support repeated indices */
5594: Mat *tmats;
5595: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)matis->A), local_size, 0, 1, &is_local));
5596: PetscCall(ISLocalToGlobalMappingApplyIS(matis->rmapping, is_local, &is_global));
5597: PetscCall(ISDestroy(&is_local));
5598: PetscCall(MatSetOption(ChangeOfBasisMatrix, MAT_SUBMAT_SINGLEIS, PETSC_TRUE));
5599: PetscCall(MatCreateSubMatrices(ChangeOfBasisMatrix, 1, &is_global, &is_global, MAT_INITIAL_MATRIX, &tmats));
5600: PetscCall(ISDestroy(&is_global));
5601: PetscCall(MatCreateSubMatrices(tmats[0], nsubs, pcbddc->local_subs, pcbddc->local_subs, MAT_INITIAL_MATRIX, &bdiags));
5602: PetscCall(MatDestroySubMatrices(1, &tmats));
5603: #endif
5604: for (PetscInt i = 0; i < nsubs; i++) mats[i * (1 + nsubs)] = bdiags[i];
5605: PetscCall(MatCreateNest(PETSC_COMM_SELF, nsubs, pcbddc->local_subs, nsubs, pcbddc->local_subs, mats, &new_mat));
5606: PetscCall(MatConvert(new_mat, MATSEQAIJ, MAT_INPLACE_MATRIX, &new_mat));
5607: PetscCall(MatDestroySubMatrices(nsubs, &bdiags));
5608: PetscCall(PetscFree(mats));
5609: } else {
5610: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)matis->A), local_size, 0, 1, &is_local));
5611: PetscCall(ISLocalToGlobalMappingApplyIS(matis->rmapping, is_local, &is_global));
5612: PetscCall(ISDestroy(&is_local));
5613: PetscCall(MatCreateSubMatrixUnsorted(ChangeOfBasisMatrix, is_global, is_global, &new_mat));
5614: PetscCall(ISDestroy(&is_global));
5615: }
5616: if (pcbddc->dbg_flag) {
5617: Vec x, x_change;
5618: PetscReal error;
5620: PetscCall(MatCreateVecs(ChangeOfBasisMatrix, &x, &x_change));
5621: PetscCall(VecSetRandom(x, NULL));
5622: PetscCall(MatMult(ChangeOfBasisMatrix, x, x_change));
5623: PetscCall(VecScatterBegin(matis->cctx, x, matis->x, INSERT_VALUES, SCATTER_FORWARD));
5624: PetscCall(VecScatterEnd(matis->cctx, x, matis->x, INSERT_VALUES, SCATTER_FORWARD));
5625: PetscCall(MatMult(new_mat, matis->x, matis->y));
5626: if (!pcbddc->change_interior) {
5627: const PetscScalar *x, *y, *v;
5628: PetscReal lerror = 0.;
5629: PetscInt i;
5631: PetscCall(VecGetArrayRead(matis->x, &x));
5632: PetscCall(VecGetArrayRead(matis->y, &y));
5633: PetscCall(VecGetArrayRead(matis->counter, &v));
5634: for (i = 0; i < local_size; i++)
5635: if (PetscRealPart(v[i]) < 1.5 && PetscAbsScalar(x[i] - y[i]) > lerror) lerror = PetscAbsScalar(x[i] - y[i]);
5636: PetscCall(VecRestoreArrayRead(matis->x, &x));
5637: PetscCall(VecRestoreArrayRead(matis->y, &y));
5638: PetscCall(VecRestoreArrayRead(matis->counter, &v));
5639: PetscCallMPI(MPIU_Allreduce(&lerror, &error, 1, MPIU_REAL, MPIU_MAX, PetscObjectComm((PetscObject)pc)));
5640: if (error > PETSC_SMALL) {
5641: if (!pcbddc->user_ChangeOfBasisMatrix || pcbddc->current_level) {
5642: SETERRQ(PetscObjectComm((PetscObject)pc), PETSC_ERR_PLIB, "Error global vs local change on I: %1.6e", (double)error);
5643: } else {
5644: SETERRQ(PetscObjectComm((PetscObject)pc), PETSC_ERR_USER, "Error global vs local change on I: %1.6e", (double)error);
5645: }
5646: }
5647: }
5648: PetscCall(VecScatterBegin(matis->rctx, matis->y, x, INSERT_VALUES, SCATTER_REVERSE));
5649: PetscCall(VecScatterEnd(matis->rctx, matis->y, x, INSERT_VALUES, SCATTER_REVERSE));
5650: PetscCall(VecAXPY(x, -1.0, x_change));
5651: PetscCall(VecNorm(x, NORM_INFINITY, &error));
5652: if (error > PETSC_SMALL) {
5653: if (!pcbddc->user_ChangeOfBasisMatrix || pcbddc->current_level) {
5654: SETERRQ(PetscObjectComm((PetscObject)pc), PETSC_ERR_PLIB, "Error global vs local change on N: %1.6e", (double)error);
5655: } else {
5656: SETERRQ(PetscObjectComm((PetscObject)pc), PETSC_ERR_USER, "Error global vs local change on N: %1.6e", (double)error);
5657: }
5658: }
5659: PetscCall(VecDestroy(&x));
5660: PetscCall(VecDestroy(&x_change));
5661: }
5663: /* lA is present if we are setting up an inner BDDC for a saddle point FETI-DP */
5664: PetscCall(PetscObjectQuery((PetscObject)pc, "__KSPFETIDP_lA", (PetscObject *)&lA));
5666: /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */
5667: if (((PetscObject)pc)->prefix) PetscCall(PetscSNPrintf(ptapprefix, sizeof(ptapprefix), "%spc_bddc_change_", ((PetscObject)pc)->prefix));
5668: else PetscCall(PetscSNPrintf(ptapprefix, sizeof(ptapprefix), "pc_bddc_change_"));
5669: PetscCall(PetscObjectBaseTypeCompare((PetscObject)matis->A, MATSEQAIJ, &isseqaij));
5670: if (isseqaij) {
5671: PetscCall(MatDestroy(&pcbddc->local_mat));
5672: PetscCall(MatPtAPWithPrefix_Private(matis->A, new_mat, PETSC_DEFAULT, ptapprefix, &pcbddc->local_mat));
5673: if (lA) {
5674: Mat work;
5675: PetscCall(MatPtAPWithPrefix_Private(lA, new_mat, PETSC_DEFAULT, ptapprefix, &work));
5676: PetscCall(PetscObjectCompose((PetscObject)pc, "__KSPFETIDP_lA", (PetscObject)work));
5677: PetscCall(MatDestroy(&work));
5678: }
5679: } else {
5680: Mat work_mat;
5682: PetscCall(MatDestroy(&pcbddc->local_mat));
5683: PetscCall(MatConvert(matis->A, MATSEQAIJ, MAT_INITIAL_MATRIX, &work_mat));
5684: PetscCall(MatPtAPWithPrefix_Private(work_mat, new_mat, PETSC_DEFAULT, ptapprefix, &pcbddc->local_mat));
5685: PetscCall(MatDestroy(&work_mat));
5686: if (lA) {
5687: Mat work;
5688: PetscCall(MatConvert(lA, MATSEQAIJ, MAT_INITIAL_MATRIX, &work_mat));
5689: PetscCall(MatPtAPWithPrefix_Private(work_mat, new_mat, PETSC_DEFAULT, ptapprefix, &work));
5690: PetscCall(PetscObjectCompose((PetscObject)pc, "__KSPFETIDP_lA", (PetscObject)work));
5691: PetscCall(MatDestroy(&work));
5692: }
5693: }
5694: PetscCall(MatIsSymmetricKnown(matis->A, &isset, &issym));
5695: if (isset) PetscCall(MatSetOption(pcbddc->local_mat, MAT_SYMMETRIC, issym));
5696: PetscCall(MatDestroy(&new_mat));
5697: PetscFunctionReturn(PETSC_SUCCESS);
5698: }
5700: PetscErrorCode PCBDDCSetUpLocalScatters(PC pc)
5701: {
5702: PC_IS *pcis = (PC_IS *)pc->data;
5703: PC_BDDC *pcbddc = (PC_BDDC *)pc->data;
5704: PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs;
5705: PetscInt *idx_R_local = NULL;
5706: PetscInt n_vertices, i, j, n_R, n_D, n_B;
5707: PetscInt vbs, bs;
5708: PetscBT bitmask = NULL;
5710: PetscFunctionBegin;
5711: /*
5712: No need to setup local scatters if
5713: - primal space is unchanged
5714: AND
5715: - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains)
5716: AND
5717: - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine
5718: */
5719: if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) PetscFunctionReturn(PETSC_SUCCESS);
5720: /* destroy old objects */
5721: PetscCall(ISDestroy(&pcbddc->is_R_local));
5722: PetscCall(VecScatterDestroy(&pcbddc->R_to_B));
5723: PetscCall(VecScatterDestroy(&pcbddc->R_to_D));
5724: /* Set Non-overlapping dimensions */
5725: n_B = pcis->n_B;
5726: n_D = pcis->n - n_B;
5727: n_vertices = pcbddc->n_vertices;
5729: /* Dohrmann's notation: dofs split in R (Remaining: all dofs but the vertices) and V (Vertices) */
5731: /* create auxiliary bitmask and allocate workspace */
5732: if (!sub_schurs || !sub_schurs->reuse_solver) {
5733: PetscCall(PetscMalloc1(pcis->n - n_vertices, &idx_R_local));
5734: PetscCall(PetscBTCreate(pcis->n, &bitmask));
5735: for (i = 0; i < n_vertices; i++) PetscCall(PetscBTSet(bitmask, pcbddc->local_primal_ref_node[i]));
5737: for (i = 0, n_R = 0; i < pcis->n; i++) {
5738: if (!PetscBTLookup(bitmask, i)) idx_R_local[n_R++] = i;
5739: }
5740: } else { /* A different ordering (already computed) is present if we are reusing the Schur solver */
5741: PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
5743: PetscCall(ISGetIndices(reuse_solver->is_R, (const PetscInt **)&idx_R_local));
5744: PetscCall(ISGetLocalSize(reuse_solver->is_R, &n_R));
5745: }
5747: /* Block code */
5748: vbs = 1;
5749: PetscCall(MatGetBlockSize(pcbddc->local_mat, &bs));
5750: if (bs > 1 && !(n_vertices % bs)) {
5751: PetscBool is_blocked = PETSC_TRUE;
5752: PetscInt *vary;
5753: if (!sub_schurs || !sub_schurs->reuse_solver) {
5754: PetscCall(PetscMalloc1(pcis->n / bs, &vary));
5755: PetscCall(PetscArrayzero(vary, pcis->n / bs));
5756: /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */
5757: /* it is ok to check this way since local_primal_ref_node are always sorted by local numbering and idx_R_local is obtained as a complement */
5758: for (i = 0; i < n_vertices; i++) vary[pcbddc->local_primal_ref_node[i] / bs]++;
5759: for (i = 0; i < pcis->n / bs; i++) {
5760: if (vary[i] != 0 && vary[i] != bs) {
5761: is_blocked = PETSC_FALSE;
5762: break;
5763: }
5764: }
5765: PetscCall(PetscFree(vary));
5766: } else {
5767: /* Verify directly the R set */
5768: for (i = 0; i < n_R / bs; i++) {
5769: PetscInt j, node = idx_R_local[bs * i];
5770: for (j = 1; j < bs; j++) {
5771: if (node != idx_R_local[bs * i + j] - j) {
5772: is_blocked = PETSC_FALSE;
5773: break;
5774: }
5775: }
5776: }
5777: }
5778: if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */
5779: vbs = bs;
5780: for (i = 0; i < n_R / vbs; i++) idx_R_local[i] = idx_R_local[vbs * i] / vbs;
5781: }
5782: }
5783: PetscCall(ISCreateBlock(PETSC_COMM_SELF, vbs, n_R / vbs, idx_R_local, PETSC_COPY_VALUES, &pcbddc->is_R_local));
5784: if (sub_schurs && sub_schurs->reuse_solver) {
5785: PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
5787: PetscCall(ISRestoreIndices(reuse_solver->is_R, (const PetscInt **)&idx_R_local));
5788: PetscCall(ISDestroy(&reuse_solver->is_R));
5789: PetscCall(PetscObjectReference((PetscObject)pcbddc->is_R_local));
5790: reuse_solver->is_R = pcbddc->is_R_local;
5791: } else {
5792: PetscCall(PetscFree(idx_R_local));
5793: }
5795: /* print some info if requested */
5796: if (pcbddc->dbg_flag) {
5797: PetscCall(PetscViewerASCIIPrintf(pcbddc->dbg_viewer, "--------------------------------------------------\n"));
5798: PetscCall(PetscViewerFlush(pcbddc->dbg_viewer));
5799: PetscCall(PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer));
5800: PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, "Subdomain %04d local dimensions\n", PetscGlobalRank));
5801: PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, "local_size = %" PetscInt_FMT ", dirichlet_size = %" PetscInt_FMT ", boundary_size = %" PetscInt_FMT "\n", pcis->n, n_D, n_B));
5802: PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, "r_size = %" PetscInt_FMT ", v_size = %" PetscInt_FMT ", constraints = %" PetscInt_FMT ", local_primal_size = %" PetscInt_FMT "\n", n_R, n_vertices,
5803: pcbddc->local_primal_size - n_vertices - pcbddc->benign_n, pcbddc->local_primal_size));
5804: PetscCall(PetscViewerFlush(pcbddc->dbg_viewer));
5805: }
5807: /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */
5808: if (!sub_schurs || !sub_schurs->reuse_solver) {
5809: IS is_aux1, is_aux2;
5810: PetscInt *aux_array1, *aux_array2, *is_indices, *idx_R_local;
5812: PetscCall(ISGetIndices(pcbddc->is_R_local, (const PetscInt **)&idx_R_local));
5813: PetscCall(PetscMalloc1(pcis->n_B - n_vertices, &aux_array1));
5814: PetscCall(PetscMalloc1(pcis->n_B - n_vertices, &aux_array2));
5815: PetscCall(ISGetIndices(pcis->is_I_local, (const PetscInt **)&is_indices));
5816: for (i = 0; i < n_D; i++) PetscCall(PetscBTSet(bitmask, is_indices[i]));
5817: PetscCall(ISRestoreIndices(pcis->is_I_local, (const PetscInt **)&is_indices));
5818: for (i = 0, j = 0; i < n_R; i++) {
5819: if (!PetscBTLookup(bitmask, idx_R_local[i])) aux_array1[j++] = i;
5820: }
5821: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, j, aux_array1, PETSC_OWN_POINTER, &is_aux1));
5822: PetscCall(ISGetIndices(pcis->is_B_local, (const PetscInt **)&is_indices));
5823: for (i = 0, j = 0; i < n_B; i++) {
5824: if (!PetscBTLookup(bitmask, is_indices[i])) aux_array2[j++] = i;
5825: }
5826: PetscCall(ISRestoreIndices(pcis->is_B_local, (const PetscInt **)&is_indices));
5827: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, j, aux_array2, PETSC_OWN_POINTER, &is_aux2));
5828: PetscCall(VecScatterCreate(pcbddc->vec1_R, is_aux1, pcis->vec1_B, is_aux2, &pcbddc->R_to_B));
5829: PetscCall(ISDestroy(&is_aux1));
5830: PetscCall(ISDestroy(&is_aux2));
5832: if (pcbddc->switch_static || pcbddc->dbg_flag) {
5833: PetscCall(PetscMalloc1(n_D, &aux_array1));
5834: for (i = 0, j = 0; i < n_R; i++) {
5835: if (PetscBTLookup(bitmask, idx_R_local[i])) aux_array1[j++] = i;
5836: }
5837: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, j, aux_array1, PETSC_OWN_POINTER, &is_aux1));
5838: PetscCall(VecScatterCreate(pcbddc->vec1_R, is_aux1, pcis->vec1_D, (IS)0, &pcbddc->R_to_D));
5839: PetscCall(ISDestroy(&is_aux1));
5840: }
5841: PetscCall(PetscBTDestroy(&bitmask));
5842: PetscCall(ISRestoreIndices(pcbddc->is_R_local, (const PetscInt **)&idx_R_local));
5843: } else {
5844: PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
5845: IS tis;
5846: PetscInt schur_size;
5848: PetscCall(ISGetLocalSize(reuse_solver->is_B, &schur_size));
5849: PetscCall(ISCreateStride(PETSC_COMM_SELF, schur_size, n_D, 1, &tis));
5850: PetscCall(VecScatterCreate(pcbddc->vec1_R, tis, pcis->vec1_B, reuse_solver->is_B, &pcbddc->R_to_B));
5851: PetscCall(ISDestroy(&tis));
5852: if (pcbddc->switch_static || pcbddc->dbg_flag) {
5853: PetscCall(ISCreateStride(PETSC_COMM_SELF, n_D, 0, 1, &tis));
5854: PetscCall(VecScatterCreate(pcbddc->vec1_R, tis, pcis->vec1_D, (IS)0, &pcbddc->R_to_D));
5855: PetscCall(ISDestroy(&tis));
5856: }
5857: }
5858: PetscFunctionReturn(PETSC_SUCCESS);
5859: }
5861: PetscErrorCode MatNullSpacePropagateAny_Private(Mat A, IS is, Mat B)
5862: {
5863: MatNullSpace NullSpace;
5864: Mat dmat;
5865: const Vec *nullvecs;
5866: Vec v, v2, *nullvecs2;
5867: VecScatter sct = NULL;
5868: PetscScalar *ddata;
5869: PetscInt k, nnsp_size, bsiz, bsiz2, n, N, bs;
5870: PetscBool nnsp_has_cnst;
5872: PetscFunctionBegin;
5873: if (!is && !B) { /* MATIS */
5874: Mat_IS *matis = (Mat_IS *)A->data;
5876: if (!B) PetscCall(MatISGetLocalMat(A, &B));
5877: sct = matis->cctx;
5878: PetscCall(PetscObjectReference((PetscObject)sct));
5879: } else {
5880: PetscCall(MatGetNullSpace(B, &NullSpace));
5881: if (!NullSpace) PetscCall(MatGetNearNullSpace(B, &NullSpace));
5882: if (NullSpace) PetscFunctionReturn(PETSC_SUCCESS);
5883: }
5884: PetscCall(MatGetNullSpace(A, &NullSpace));
5885: if (!NullSpace) PetscCall(MatGetNearNullSpace(A, &NullSpace));
5886: if (!NullSpace) PetscFunctionReturn(PETSC_SUCCESS);
5888: PetscCall(MatCreateVecs(A, &v, NULL));
5889: PetscCall(MatCreateVecs(B, &v2, NULL));
5890: if (!sct) PetscCall(VecScatterCreate(v, is, v2, NULL, &sct));
5891: PetscCall(MatNullSpaceGetVecs(NullSpace, &nnsp_has_cnst, &nnsp_size, &nullvecs));
5892: bsiz = bsiz2 = nnsp_size + !!nnsp_has_cnst;
5893: PetscCall(PetscMalloc1(bsiz, &nullvecs2));
5894: PetscCall(VecGetBlockSize(v2, &bs));
5895: PetscCall(VecGetSize(v2, &N));
5896: PetscCall(VecGetLocalSize(v2, &n));
5897: PetscCall(PetscMalloc1(n * bsiz, &ddata));
5898: for (k = 0; k < nnsp_size; k++) {
5899: PetscCall(VecCreateMPIWithArray(PetscObjectComm((PetscObject)B), bs, n, N, ddata + n * k, &nullvecs2[k]));
5900: PetscCall(VecScatterBegin(sct, nullvecs[k], nullvecs2[k], INSERT_VALUES, SCATTER_FORWARD));
5901: PetscCall(VecScatterEnd(sct, nullvecs[k], nullvecs2[k], INSERT_VALUES, SCATTER_FORWARD));
5902: }
5903: if (nnsp_has_cnst) {
5904: PetscCall(VecCreateMPIWithArray(PetscObjectComm((PetscObject)B), bs, n, N, ddata + n * nnsp_size, &nullvecs2[nnsp_size]));
5905: PetscCall(VecSet(nullvecs2[nnsp_size], 1.0));
5906: }
5907: PetscCall(PCBDDCOrthonormalizeVecs(&bsiz2, nullvecs2));
5908: PetscCall(MatNullSpaceCreate(PetscObjectComm((PetscObject)B), PETSC_FALSE, bsiz2, nullvecs2, &NullSpace));
5910: PetscCall(MatCreateDense(PetscObjectComm((PetscObject)B), n, PETSC_DECIDE, N, bsiz2, ddata, &dmat));
5911: PetscCall(PetscObjectContainerCompose((PetscObject)dmat, "_PBDDC_Null_dmat_arr", ddata, PetscCtxDestroyDefault));
5912: PetscCall(PetscObjectCompose((PetscObject)NullSpace, "_PBDDC_Null_dmat", (PetscObject)dmat));
5913: PetscCall(MatDestroy(&dmat));
5915: for (k = 0; k < bsiz; k++) PetscCall(VecDestroy(&nullvecs2[k]));
5916: PetscCall(PetscFree(nullvecs2));
5917: PetscCall(MatSetNearNullSpace(B, NullSpace));
5918: PetscCall(MatNullSpaceDestroy(&NullSpace));
5919: PetscCall(VecDestroy(&v));
5920: PetscCall(VecDestroy(&v2));
5921: PetscCall(VecScatterDestroy(&sct));
5922: PetscFunctionReturn(PETSC_SUCCESS);
5923: }
5925: PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann)
5926: {
5927: PC_BDDC *pcbddc = (PC_BDDC *)pc->data;
5928: PC_IS *pcis = (PC_IS *)pc->data;
5929: PC pc_temp;
5930: Mat A_RR;
5931: MatNullSpace nnsp;
5932: MatReuse reuse;
5933: PetscScalar m_one = -1.0;
5934: PetscReal value;
5935: PetscInt n_D, n_R;
5936: PetscBool issbaij, opts, isset, issym;
5937: PetscBool f = PETSC_FALSE;
5938: char dir_prefix[256], neu_prefix[256], str_level[16];
5939: size_t len;
5941: PetscFunctionBegin;
5942: PetscCall(PetscLogEventBegin(PC_BDDC_LocalSolvers[pcbddc->current_level], pc, 0, 0, 0));
5943: /* approximate solver, propagate NearNullSpace if needed */
5944: if (!pc->setupcalled && (pcbddc->NullSpace_corr[0] || pcbddc->NullSpace_corr[2])) {
5945: MatNullSpace gnnsp1, gnnsp2;
5946: PetscBool lhas, ghas;
5948: PetscCall(MatGetNearNullSpace(pcbddc->local_mat, &nnsp));
5949: PetscCall(MatGetNearNullSpace(pc->pmat, &gnnsp1));
5950: PetscCall(MatGetNullSpace(pc->pmat, &gnnsp2));
5951: lhas = nnsp ? PETSC_TRUE : PETSC_FALSE;
5952: PetscCallMPI(MPIU_Allreduce(&lhas, &ghas, 1, MPI_C_BOOL, MPI_LOR, PetscObjectComm((PetscObject)pc)));
5953: if (!ghas && (gnnsp1 || gnnsp2)) PetscCall(MatNullSpacePropagateAny_Private(pc->pmat, NULL, NULL));
5954: }
5956: /* compute prefixes */
5957: PetscCall(PetscStrncpy(dir_prefix, "", sizeof(dir_prefix)));
5958: PetscCall(PetscStrncpy(neu_prefix, "", sizeof(neu_prefix)));
5959: if (!pcbddc->current_level) {
5960: PetscCall(PetscStrncpy(dir_prefix, ((PetscObject)pc)->prefix, sizeof(dir_prefix)));
5961: PetscCall(PetscStrncpy(neu_prefix, ((PetscObject)pc)->prefix, sizeof(neu_prefix)));
5962: PetscCall(PetscStrlcat(dir_prefix, "pc_bddc_dirichlet_", sizeof(dir_prefix)));
5963: PetscCall(PetscStrlcat(neu_prefix, "pc_bddc_neumann_", sizeof(neu_prefix)));
5964: } else {
5965: PetscCall(PetscSNPrintf(str_level, sizeof(str_level), "l%" PetscInt_FMT "_", pcbddc->current_level));
5966: PetscCall(PetscStrlen(((PetscObject)pc)->prefix, &len));
5967: len -= 15; /* remove "pc_bddc_coarse_" */
5968: if (pcbddc->current_level > 1) len -= 3; /* remove "lX_" with X level number */
5969: if (pcbddc->current_level > 10) len -= 1; /* remove another char from level number */
5970: /* Nonstandard use of PetscStrncpy() to only copy a portion of the input string */
5971: PetscCall(PetscStrncpy(dir_prefix, ((PetscObject)pc)->prefix, len + 1));
5972: PetscCall(PetscStrncpy(neu_prefix, ((PetscObject)pc)->prefix, len + 1));
5973: PetscCall(PetscStrlcat(dir_prefix, "pc_bddc_dirichlet_", sizeof(dir_prefix)));
5974: PetscCall(PetscStrlcat(neu_prefix, "pc_bddc_neumann_", sizeof(neu_prefix)));
5975: PetscCall(PetscStrlcat(dir_prefix, str_level, sizeof(dir_prefix)));
5976: PetscCall(PetscStrlcat(neu_prefix, str_level, sizeof(neu_prefix)));
5977: }
5979: /* DIRICHLET PROBLEM */
5980: if (dirichlet) {
5981: PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs;
5982: if (pcbddc->benign_n && !pcbddc->benign_change_explicit) {
5983: PetscCheck(sub_schurs && sub_schurs->reuse_solver, PETSC_COMM_SELF, PETSC_ERR_SUP, "Not yet implemented");
5984: if (pcbddc->dbg_flag) {
5985: Mat A_IIn;
5987: PetscCall(PCBDDCBenignProject(pc, pcis->is_I_local, pcis->is_I_local, &A_IIn));
5988: PetscCall(MatDestroy(&pcis->A_II));
5989: pcis->A_II = A_IIn;
5990: }
5991: }
5992: PetscCall(MatIsSymmetricKnown(pcbddc->local_mat, &isset, &issym));
5993: if (isset) PetscCall(MatSetOption(pcis->A_II, MAT_SYMMETRIC, issym));
5995: /* Matrix for Dirichlet problem is pcis->A_II */
5996: n_D = pcis->n - pcis->n_B;
5997: opts = PETSC_FALSE;
5998: if (!pcbddc->ksp_D) { /* create object if not yet build */
5999: opts = PETSC_TRUE;
6000: PetscCall(KSPCreate(PETSC_COMM_SELF, &pcbddc->ksp_D));
6001: PetscCall(KSPSetNestLevel(pcbddc->ksp_D, pc->kspnestlevel));
6002: PetscCall(PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D, (PetscObject)pc, 1));
6003: /* default */
6004: PetscCall(KSPSetType(pcbddc->ksp_D, KSPPREONLY));
6005: PetscCall(KSPSetOptionsPrefix(pcbddc->ksp_D, dir_prefix));
6006: PetscCall(PetscObjectTypeCompare((PetscObject)pcis->pA_II, MATSEQSBAIJ, &issbaij));
6007: PetscCall(KSPGetPC(pcbddc->ksp_D, &pc_temp));
6008: if (issbaij) PetscCall(PCSetType(pc_temp, PCCHOLESKY));
6009: else PetscCall(PCSetType(pc_temp, PCLU));
6010: PetscCall(KSPSetErrorIfNotConverged(pcbddc->ksp_D, pc->erroriffailure));
6011: }
6012: PetscCall(MatSetOptionsPrefix(pcis->pA_II, ((PetscObject)pcbddc->ksp_D)->prefix));
6013: PetscCall(MatViewFromOptions(pcis->pA_II, NULL, "-mat_view"));
6014: PetscCall(KSPSetOperators(pcbddc->ksp_D, pcis->A_II, pcis->pA_II));
6015: /* Allow user's customization */
6016: if (opts) PetscCall(KSPSetFromOptions(pcbddc->ksp_D));
6017: PetscCall(MatGetNearNullSpace(pcis->pA_II, &nnsp));
6018: if (pcbddc->NullSpace_corr[0] && !nnsp) { /* approximate solver, propagate NearNullSpace */
6019: PetscCall(MatNullSpacePropagateAny_Private(pcbddc->local_mat, pcis->is_I_local, pcis->pA_II));
6020: }
6021: PetscCall(MatGetNearNullSpace(pcis->pA_II, &nnsp));
6022: PetscCall(KSPGetPC(pcbddc->ksp_D, &pc_temp));
6023: PetscCall(PetscObjectHasFunction((PetscObject)pc_temp, "PCSetCoordinates_C", &f));
6024: if (f && pcbddc->mat_graph->cloc && !nnsp) {
6025: PetscReal *coords = pcbddc->mat_graph->coords, *scoords;
6026: const PetscInt *idxs;
6027: PetscInt cdim = pcbddc->mat_graph->cdim, nl, i, d;
6029: PetscCall(ISGetLocalSize(pcis->is_I_local, &nl));
6030: PetscCall(ISGetIndices(pcis->is_I_local, &idxs));
6031: PetscCall(PetscMalloc1(nl * cdim, &scoords));
6032: for (i = 0; i < nl; i++) {
6033: for (d = 0; d < cdim; d++) scoords[i * cdim + d] = coords[idxs[i] * cdim + d];
6034: }
6035: PetscCall(ISRestoreIndices(pcis->is_I_local, &idxs));
6036: PetscCall(PCSetCoordinates(pc_temp, cdim, nl, scoords));
6037: PetscCall(PetscFree(scoords));
6038: }
6039: if (sub_schurs && sub_schurs->reuse_solver) {
6040: PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
6042: PetscCall(KSPSetPC(pcbddc->ksp_D, reuse_solver->interior_solver));
6043: }
6045: /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */
6046: if (!n_D) {
6047: PetscCall(KSPGetPC(pcbddc->ksp_D, &pc_temp));
6048: PetscCall(PCSetType(pc_temp, PCNONE));
6049: }
6050: PetscCall(KSPSetUp(pcbddc->ksp_D));
6051: /* set ksp_D into pcis data */
6052: PetscCall(PetscObjectReference((PetscObject)pcbddc->ksp_D));
6053: PetscCall(KSPDestroy(&pcis->ksp_D));
6054: pcis->ksp_D = pcbddc->ksp_D;
6055: }
6057: /* NEUMANN PROBLEM */
6058: A_RR = NULL;
6059: if (neumann) {
6060: PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs;
6061: PetscInt ibs, mbs;
6062: PetscBool issbaij, reuse_neumann_solver, isset, issym;
6063: Mat_IS *matis = (Mat_IS *)pc->pmat->data;
6065: reuse_neumann_solver = PETSC_FALSE;
6066: if (sub_schurs && sub_schurs->reuse_solver) {
6067: IS iP;
6069: reuse_neumann_solver = PETSC_TRUE;
6070: PetscCall(PetscObjectQuery((PetscObject)sub_schurs->A, "__KSPFETIDP_iP", (PetscObject *)&iP));
6071: if (iP) reuse_neumann_solver = PETSC_FALSE;
6072: }
6073: /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */
6074: PetscCall(ISGetSize(pcbddc->is_R_local, &n_R));
6075: if (pcbddc->ksp_R) { /* already created ksp */
6076: PetscInt nn_R;
6077: PetscCall(KSPGetOperators(pcbddc->ksp_R, NULL, &A_RR));
6078: PetscCall(PetscObjectReference((PetscObject)A_RR));
6079: PetscCall(MatGetSize(A_RR, &nn_R, NULL));
6080: if (nn_R != n_R) { /* old ksp is not reusable, so reset it */
6081: PetscCall(KSPReset(pcbddc->ksp_R));
6082: PetscCall(MatDestroy(&A_RR));
6083: reuse = MAT_INITIAL_MATRIX;
6084: } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */
6085: if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */
6086: PetscCall(MatDestroy(&A_RR));
6087: reuse = MAT_INITIAL_MATRIX;
6088: } else { /* safe to reuse the matrix */
6089: reuse = MAT_REUSE_MATRIX;
6090: }
6091: }
6092: /* last check */
6093: if (pc->flag == DIFFERENT_NONZERO_PATTERN) {
6094: PetscCall(MatDestroy(&A_RR));
6095: reuse = MAT_INITIAL_MATRIX;
6096: }
6097: } else { /* first time, so we need to create the matrix */
6098: reuse = MAT_INITIAL_MATRIX;
6099: }
6100: /* convert pcbddc->local_mat if needed later in PCBDDCSetUpCorrection
6101: TODO: Get Rid of these conversions */
6102: PetscCall(MatGetBlockSize(pcbddc->local_mat, &mbs));
6103: PetscCall(ISGetBlockSize(pcbddc->is_R_local, &ibs));
6104: PetscCall(PetscObjectTypeCompare((PetscObject)pcbddc->local_mat, MATSEQSBAIJ, &issbaij));
6105: if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */
6106: if (matis->A == pcbddc->local_mat) {
6107: PetscCall(MatDestroy(&pcbddc->local_mat));
6108: PetscCall(MatConvert(matis->A, MATSEQAIJ, MAT_INITIAL_MATRIX, &pcbddc->local_mat));
6109: } else {
6110: PetscCall(MatConvert(pcbddc->local_mat, MATSEQAIJ, MAT_INPLACE_MATRIX, &pcbddc->local_mat));
6111: }
6112: } else if (issbaij) { /* need to convert to BAIJ to get off-diagonal blocks */
6113: if (matis->A == pcbddc->local_mat) {
6114: PetscCall(MatDestroy(&pcbddc->local_mat));
6115: PetscCall(MatConvert(matis->A, mbs > 1 ? MATSEQBAIJ : MATSEQAIJ, MAT_INITIAL_MATRIX, &pcbddc->local_mat));
6116: } else {
6117: PetscCall(MatConvert(pcbddc->local_mat, mbs > 1 ? MATSEQBAIJ : MATSEQAIJ, MAT_INPLACE_MATRIX, &pcbddc->local_mat));
6118: }
6119: }
6120: /* extract A_RR */
6121: if (reuse_neumann_solver) {
6122: PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
6124: if (pcbddc->dbg_flag) { /* we need A_RR to test the solver later */
6125: PetscCall(MatDestroy(&A_RR));
6126: if (reuse_solver->benign_n) { /* we are not using the explicit change of basis on the pressures */
6127: PetscCall(PCBDDCBenignProject(pc, pcbddc->is_R_local, pcbddc->is_R_local, &A_RR));
6128: } else {
6129: PetscCall(MatCreateSubMatrix(pcbddc->local_mat, pcbddc->is_R_local, pcbddc->is_R_local, MAT_INITIAL_MATRIX, &A_RR));
6130: }
6131: } else {
6132: PetscCall(MatDestroy(&A_RR));
6133: PetscCall(PCGetOperators(reuse_solver->correction_solver, &A_RR, NULL));
6134: PetscCall(PetscObjectReference((PetscObject)A_RR));
6135: }
6136: } else { /* we have to build the neumann solver, so we need to extract the relevant matrix */
6137: PetscCall(MatCreateSubMatrix(pcbddc->local_mat, pcbddc->is_R_local, pcbddc->is_R_local, reuse, &A_RR));
6138: }
6139: PetscCall(MatIsSymmetricKnown(pcbddc->local_mat, &isset, &issym));
6140: if (isset) PetscCall(MatSetOption(A_RR, MAT_SYMMETRIC, issym));
6141: opts = PETSC_FALSE;
6142: if (!pcbddc->ksp_R) { /* create object if not present */
6143: opts = PETSC_TRUE;
6144: PetscCall(KSPCreate(PETSC_COMM_SELF, &pcbddc->ksp_R));
6145: PetscCall(KSPSetNestLevel(pcbddc->ksp_R, pc->kspnestlevel));
6146: PetscCall(PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R, (PetscObject)pc, 1));
6147: /* default */
6148: PetscCall(KSPSetType(pcbddc->ksp_R, KSPPREONLY));
6149: PetscCall(KSPSetOptionsPrefix(pcbddc->ksp_R, neu_prefix));
6150: PetscCall(KSPGetPC(pcbddc->ksp_R, &pc_temp));
6151: PetscCall(PetscObjectTypeCompare((PetscObject)A_RR, MATSEQSBAIJ, &issbaij));
6152: if (issbaij) PetscCall(PCSetType(pc_temp, PCCHOLESKY));
6153: else PetscCall(PCSetType(pc_temp, PCLU));
6154: PetscCall(KSPSetErrorIfNotConverged(pcbddc->ksp_R, pc->erroriffailure));
6155: }
6156: PetscCall(MatSetOptionsPrefix(A_RR, ((PetscObject)pcbddc->ksp_R)->prefix));
6157: PetscCall(MatViewFromOptions(A_RR, NULL, "-mat_view"));
6158: PetscCall(KSPSetOperators(pcbddc->ksp_R, A_RR, A_RR));
6159: if (opts) { /* Allow user's customization once */
6160: PetscCall(KSPSetFromOptions(pcbddc->ksp_R));
6161: }
6162: PetscCall(MatGetNearNullSpace(A_RR, &nnsp));
6163: if (pcbddc->NullSpace_corr[2] && !nnsp) { /* approximate solver, propagate NearNullSpace */
6164: PetscCall(MatNullSpacePropagateAny_Private(pcbddc->local_mat, pcbddc->is_R_local, A_RR));
6165: }
6166: PetscCall(MatGetNearNullSpace(A_RR, &nnsp));
6167: PetscCall(KSPGetPC(pcbddc->ksp_R, &pc_temp));
6168: PetscCall(PetscObjectHasFunction((PetscObject)pc_temp, "PCSetCoordinates_C", &f));
6169: if (f && pcbddc->mat_graph->cloc && !nnsp) {
6170: PetscReal *coords = pcbddc->mat_graph->coords, *scoords;
6171: const PetscInt *idxs;
6172: PetscInt cdim = pcbddc->mat_graph->cdim, nl, i, d;
6174: PetscCall(ISGetLocalSize(pcbddc->is_R_local, &nl));
6175: PetscCall(ISGetIndices(pcbddc->is_R_local, &idxs));
6176: PetscCall(PetscMalloc1(nl * cdim, &scoords));
6177: for (i = 0; i < nl; i++) {
6178: for (d = 0; d < cdim; d++) scoords[i * cdim + d] = coords[idxs[i] * cdim + d];
6179: }
6180: PetscCall(ISRestoreIndices(pcbddc->is_R_local, &idxs));
6181: PetscCall(PCSetCoordinates(pc_temp, cdim, nl, scoords));
6182: PetscCall(PetscFree(scoords));
6183: }
6185: /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */
6186: if (!n_R) {
6187: PetscCall(KSPGetPC(pcbddc->ksp_R, &pc_temp));
6188: PetscCall(PCSetType(pc_temp, PCNONE));
6189: }
6190: /* Reuse solver if it is present */
6191: if (reuse_neumann_solver) {
6192: PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
6194: PetscCall(KSPSetPC(pcbddc->ksp_R, reuse_solver->correction_solver));
6195: }
6196: PetscCall(KSPSetUp(pcbddc->ksp_R));
6197: }
6199: if (pcbddc->dbg_flag) {
6200: PetscCall(PetscViewerFlush(pcbddc->dbg_viewer));
6201: PetscCall(PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer));
6202: PetscCall(PetscViewerASCIIPrintf(pcbddc->dbg_viewer, "--------------------------------------------------\n"));
6203: }
6204: PetscCall(PetscLogEventEnd(PC_BDDC_LocalSolvers[pcbddc->current_level], pc, 0, 0, 0));
6206: /* adapt Dirichlet and Neumann solvers if a nullspace correction has been requested */
6207: if (pcbddc->NullSpace_corr[0]) PetscCall(PCBDDCSetUseExactDirichlet(pc, PETSC_FALSE));
6208: if (dirichlet && pcbddc->NullSpace_corr[0] && !pcbddc->switch_static) PetscCall(PCBDDCNullSpaceAssembleCorrection(pc, PETSC_TRUE, pcbddc->NullSpace_corr[1]));
6209: if (neumann && pcbddc->NullSpace_corr[2]) PetscCall(PCBDDCNullSpaceAssembleCorrection(pc, PETSC_FALSE, pcbddc->NullSpace_corr[3]));
6210: /* check Dirichlet and Neumann solvers */
6211: if (pcbddc->dbg_flag) {
6212: if (dirichlet) { /* Dirichlet */
6213: PetscCall(VecSetRandom(pcis->vec1_D, NULL));
6214: PetscCall(MatMult(pcis->A_II, pcis->vec1_D, pcis->vec2_D));
6215: PetscCall(KSPSolve(pcbddc->ksp_D, pcis->vec2_D, pcis->vec2_D));
6216: PetscCall(KSPCheckSolve(pcbddc->ksp_D, pc, pcis->vec2_D));
6217: PetscCall(VecAXPY(pcis->vec1_D, m_one, pcis->vec2_D));
6218: PetscCall(VecNorm(pcis->vec1_D, NORM_INFINITY, &value));
6219: PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, "Subdomain %04d infinity error for Dirichlet solve (%s) = % 1.14e \n", PetscGlobalRank, ((PetscObject)pcbddc->ksp_D)->prefix, (double)value));
6220: PetscCall(PetscViewerFlush(pcbddc->dbg_viewer));
6221: }
6222: if (neumann) { /* Neumann */
6223: PetscCall(VecSetRandom(pcbddc->vec1_R, NULL));
6224: PetscCall(MatMult(A_RR, pcbddc->vec1_R, pcbddc->vec2_R));
6225: PetscCall(KSPSolve(pcbddc->ksp_R, pcbddc->vec2_R, pcbddc->vec2_R));
6226: PetscCall(KSPCheckSolve(pcbddc->ksp_R, pc, pcbddc->vec2_R));
6227: PetscCall(VecAXPY(pcbddc->vec1_R, m_one, pcbddc->vec2_R));
6228: PetscCall(VecNorm(pcbddc->vec1_R, NORM_INFINITY, &value));
6229: PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, "Subdomain %04d infinity error for Neumann solve (%s) = % 1.14e\n", PetscGlobalRank, ((PetscObject)pcbddc->ksp_R)->prefix, (double)value));
6230: PetscCall(PetscViewerFlush(pcbddc->dbg_viewer));
6231: }
6232: }
6233: /* free Neumann problem's matrix */
6234: PetscCall(MatDestroy(&A_RR));
6235: PetscFunctionReturn(PETSC_SUCCESS);
6236: }
6238: static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose)
6239: {
6240: PC_BDDC *pcbddc = (PC_BDDC *)pc->data;
6241: PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs;
6242: PetscBool reuse_solver = sub_schurs ? (sub_schurs->reuse_solver ? PETSC_TRUE : PETSC_FALSE) : PETSC_FALSE;
6244: PetscFunctionBegin;
6245: if (!reuse_solver) PetscCall(VecSet(pcbddc->vec1_R, 0.));
6246: if (!pcbddc->switch_static) {
6247: if (applytranspose && pcbddc->local_auxmat1) {
6248: PetscCall(MatMultTranspose(pcbddc->local_auxmat2, inout_B, pcbddc->vec1_C));
6249: PetscCall(MatMultTransposeAdd(pcbddc->local_auxmat1, pcbddc->vec1_C, inout_B, inout_B));
6250: }
6251: if (!reuse_solver) {
6252: PetscCall(VecScatterBegin(pcbddc->R_to_B, inout_B, pcbddc->vec1_R, INSERT_VALUES, SCATTER_REVERSE));
6253: PetscCall(VecScatterEnd(pcbddc->R_to_B, inout_B, pcbddc->vec1_R, INSERT_VALUES, SCATTER_REVERSE));
6254: } else {
6255: PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
6257: PetscCall(VecScatterBegin(reuse_solver->correction_scatter_B, inout_B, reuse_solver->rhs_B, INSERT_VALUES, SCATTER_FORWARD));
6258: PetscCall(VecScatterEnd(reuse_solver->correction_scatter_B, inout_B, reuse_solver->rhs_B, INSERT_VALUES, SCATTER_FORWARD));
6259: }
6260: } else {
6261: PetscCall(VecScatterBegin(pcbddc->R_to_B, inout_B, pcbddc->vec1_R, INSERT_VALUES, SCATTER_REVERSE));
6262: PetscCall(VecScatterEnd(pcbddc->R_to_B, inout_B, pcbddc->vec1_R, INSERT_VALUES, SCATTER_REVERSE));
6263: PetscCall(VecScatterBegin(pcbddc->R_to_D, inout_D, pcbddc->vec1_R, INSERT_VALUES, SCATTER_REVERSE));
6264: PetscCall(VecScatterEnd(pcbddc->R_to_D, inout_D, pcbddc->vec1_R, INSERT_VALUES, SCATTER_REVERSE));
6265: if (applytranspose && pcbddc->local_auxmat1) {
6266: PetscCall(MatMultTranspose(pcbddc->local_auxmat2, pcbddc->vec1_R, pcbddc->vec1_C));
6267: PetscCall(MatMultTransposeAdd(pcbddc->local_auxmat1, pcbddc->vec1_C, inout_B, inout_B));
6268: PetscCall(VecScatterBegin(pcbddc->R_to_B, inout_B, pcbddc->vec1_R, INSERT_VALUES, SCATTER_REVERSE));
6269: PetscCall(VecScatterEnd(pcbddc->R_to_B, inout_B, pcbddc->vec1_R, INSERT_VALUES, SCATTER_REVERSE));
6270: }
6271: }
6272: PetscCall(PetscLogEventBegin(PC_BDDC_Solves[pcbddc->current_level][1], pc, 0, 0, 0));
6273: if (!reuse_solver || pcbddc->switch_static) {
6274: if (applytranspose) {
6275: PetscCall(KSPSolveTranspose(pcbddc->ksp_R, pcbddc->vec1_R, pcbddc->vec1_R));
6276: } else {
6277: PetscCall(KSPSolve(pcbddc->ksp_R, pcbddc->vec1_R, pcbddc->vec1_R));
6278: }
6279: PetscCall(KSPCheckSolve(pcbddc->ksp_R, pc, pcbddc->vec1_R));
6280: } else {
6281: PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
6283: if (applytranspose) {
6284: PetscCall(MatFactorSolveSchurComplementTranspose(reuse_solver->F, reuse_solver->rhs_B, reuse_solver->sol_B));
6285: } else {
6286: PetscCall(MatFactorSolveSchurComplement(reuse_solver->F, reuse_solver->rhs_B, reuse_solver->sol_B));
6287: }
6288: }
6289: PetscCall(PetscLogEventEnd(PC_BDDC_Solves[pcbddc->current_level][1], pc, 0, 0, 0));
6290: PetscCall(VecSet(inout_B, 0.));
6291: if (!pcbddc->switch_static) {
6292: if (!reuse_solver) {
6293: PetscCall(VecScatterBegin(pcbddc->R_to_B, pcbddc->vec1_R, inout_B, INSERT_VALUES, SCATTER_FORWARD));
6294: PetscCall(VecScatterEnd(pcbddc->R_to_B, pcbddc->vec1_R, inout_B, INSERT_VALUES, SCATTER_FORWARD));
6295: } else {
6296: PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
6298: PetscCall(VecScatterBegin(reuse_solver->correction_scatter_B, reuse_solver->sol_B, inout_B, INSERT_VALUES, SCATTER_REVERSE));
6299: PetscCall(VecScatterEnd(reuse_solver->correction_scatter_B, reuse_solver->sol_B, inout_B, INSERT_VALUES, SCATTER_REVERSE));
6300: }
6301: if (!applytranspose && pcbddc->local_auxmat1) {
6302: PetscCall(MatMult(pcbddc->local_auxmat1, inout_B, pcbddc->vec1_C));
6303: PetscCall(MatMultAdd(pcbddc->local_auxmat2, pcbddc->vec1_C, inout_B, inout_B));
6304: }
6305: } else {
6306: PetscCall(VecScatterBegin(pcbddc->R_to_B, pcbddc->vec1_R, inout_B, INSERT_VALUES, SCATTER_FORWARD));
6307: PetscCall(VecScatterEnd(pcbddc->R_to_B, pcbddc->vec1_R, inout_B, INSERT_VALUES, SCATTER_FORWARD));
6308: PetscCall(VecScatterBegin(pcbddc->R_to_D, pcbddc->vec1_R, inout_D, INSERT_VALUES, SCATTER_FORWARD));
6309: PetscCall(VecScatterEnd(pcbddc->R_to_D, pcbddc->vec1_R, inout_D, INSERT_VALUES, SCATTER_FORWARD));
6310: if (!applytranspose && pcbddc->local_auxmat1) {
6311: PetscCall(MatMult(pcbddc->local_auxmat1, inout_B, pcbddc->vec1_C));
6312: PetscCall(MatMultAdd(pcbddc->local_auxmat2, pcbddc->vec1_C, pcbddc->vec1_R, pcbddc->vec1_R));
6313: }
6314: PetscCall(VecScatterBegin(pcbddc->R_to_B, pcbddc->vec1_R, inout_B, INSERT_VALUES, SCATTER_FORWARD));
6315: PetscCall(VecScatterEnd(pcbddc->R_to_B, pcbddc->vec1_R, inout_B, INSERT_VALUES, SCATTER_FORWARD));
6316: PetscCall(VecScatterBegin(pcbddc->R_to_D, pcbddc->vec1_R, inout_D, INSERT_VALUES, SCATTER_FORWARD));
6317: PetscCall(VecScatterEnd(pcbddc->R_to_D, pcbddc->vec1_R, inout_D, INSERT_VALUES, SCATTER_FORWARD));
6318: }
6319: PetscFunctionReturn(PETSC_SUCCESS);
6320: }
6322: /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */
6323: PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose)
6324: {
6325: PC_BDDC *pcbddc = (PC_BDDC *)pc->data;
6326: PC_IS *pcis = (PC_IS *)pc->data;
6327: const PetscScalar zero = 0.0;
6329: PetscFunctionBegin;
6330: /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */
6331: if (!pcbddc->benign_apply_coarse_only) {
6332: if (applytranspose) {
6333: PetscCall(MatMultTranspose(pcbddc->coarse_phi_B, pcis->vec1_B, pcbddc->vec1_P));
6334: if (pcbddc->switch_static) PetscCall(MatMultTransposeAdd(pcbddc->coarse_phi_D, pcis->vec1_D, pcbddc->vec1_P, pcbddc->vec1_P));
6335: } else {
6336: PetscCall(MatMultTranspose(pcbddc->coarse_psi_B, pcis->vec1_B, pcbddc->vec1_P));
6337: if (pcbddc->switch_static) PetscCall(MatMultTransposeAdd(pcbddc->coarse_psi_D, pcis->vec1_D, pcbddc->vec1_P, pcbddc->vec1_P));
6338: }
6339: } else {
6340: PetscCall(VecSet(pcbddc->vec1_P, zero));
6341: }
6343: /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */
6344: if (pcbddc->benign_n) {
6345: PetscScalar *array;
6346: PetscInt j;
6348: PetscCall(VecGetArray(pcbddc->vec1_P, &array));
6349: for (j = 0; j < pcbddc->benign_n; j++) array[pcbddc->local_primal_size - pcbddc->benign_n + j] += pcbddc->benign_p0[j];
6350: PetscCall(VecRestoreArray(pcbddc->vec1_P, &array));
6351: }
6353: /* start communications from local primal nodes to rhs of coarse solver */
6354: PetscCall(VecSet(pcbddc->coarse_vec, zero));
6355: PetscCall(PCBDDCScatterCoarseDataBegin(pc, ADD_VALUES, SCATTER_FORWARD));
6356: PetscCall(PCBDDCScatterCoarseDataEnd(pc, ADD_VALUES, SCATTER_FORWARD));
6358: /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */
6359: PetscCall(PetscLogEventBegin(PC_BDDC_Solves[pcbddc->current_level][2], pc, 0, 0, 0));
6360: if (pcbddc->coarse_ksp) {
6361: Mat coarse_mat;
6362: Vec rhs, sol;
6363: MatNullSpace nullsp;
6364: PetscBool isbddc = PETSC_FALSE;
6366: if (pcbddc->benign_have_null) {
6367: PC coarse_pc;
6369: PetscCall(KSPGetPC(pcbddc->coarse_ksp, &coarse_pc));
6370: PetscCall(PetscObjectTypeCompare((PetscObject)coarse_pc, PCBDDC, &isbddc));
6371: /* we need to propagate to coarser levels the need for a possible benign correction */
6372: if (isbddc && pcbddc->benign_apply_coarse_only && !pcbddc->benign_skip_correction) {
6373: PC_BDDC *coarsepcbddc = (PC_BDDC *)coarse_pc->data;
6374: coarsepcbddc->benign_skip_correction = PETSC_FALSE;
6375: coarsepcbddc->benign_apply_coarse_only = PETSC_TRUE;
6376: }
6377: }
6378: PetscCall(KSPGetRhs(pcbddc->coarse_ksp, &rhs));
6379: PetscCall(KSPGetSolution(pcbddc->coarse_ksp, &sol));
6380: PetscCall(KSPGetOperators(pcbddc->coarse_ksp, &coarse_mat, NULL));
6381: if (applytranspose) {
6382: PetscCheck(!pcbddc->benign_apply_coarse_only, PetscObjectComm((PetscObject)pcbddc->coarse_ksp), PETSC_ERR_SUP, "Not yet implemented");
6383: PetscCall(KSPSolveTranspose(pcbddc->coarse_ksp, rhs, sol));
6384: PetscCall(KSPCheckSolve(pcbddc->coarse_ksp, pc, sol));
6385: PetscCall(MatGetTransposeNullSpace(coarse_mat, &nullsp));
6386: if (nullsp) PetscCall(MatNullSpaceRemove(nullsp, sol));
6387: } else {
6388: PetscCall(MatGetNullSpace(coarse_mat, &nullsp));
6389: if (pcbddc->benign_apply_coarse_only && isbddc) { /* need just to apply the coarse preconditioner during presolve */
6390: PC coarse_pc;
6392: if (nullsp) PetscCall(MatNullSpaceRemove(nullsp, rhs));
6393: PetscCall(KSPGetPC(pcbddc->coarse_ksp, &coarse_pc));
6394: PetscCall(PCPreSolve(coarse_pc, pcbddc->coarse_ksp));
6395: PetscCall(PCBDDCBenignRemoveInterior(coarse_pc, rhs, sol));
6396: PetscCall(PCPostSolve(coarse_pc, pcbddc->coarse_ksp));
6397: } else {
6398: PetscCall(KSPSolve(pcbddc->coarse_ksp, rhs, sol));
6399: PetscCall(KSPCheckSolve(pcbddc->coarse_ksp, pc, sol));
6400: if (nullsp) PetscCall(MatNullSpaceRemove(nullsp, sol));
6401: }
6402: }
6403: /* we don't need the benign correction at coarser levels anymore */
6404: if (pcbddc->benign_have_null && isbddc) {
6405: PC coarse_pc;
6406: PC_BDDC *coarsepcbddc;
6408: PetscCall(KSPGetPC(pcbddc->coarse_ksp, &coarse_pc));
6409: coarsepcbddc = (PC_BDDC *)coarse_pc->data;
6410: coarsepcbddc->benign_skip_correction = PETSC_TRUE;
6411: coarsepcbddc->benign_apply_coarse_only = PETSC_FALSE;
6412: }
6413: }
6414: PetscCall(PetscLogEventEnd(PC_BDDC_Solves[pcbddc->current_level][2], pc, 0, 0, 0));
6416: /* Local solution on R nodes */
6417: if (!pcbddc->benign_apply_coarse_only) PetscCall(PCBDDCSolveSubstructureCorrection(pc, pcis->vec1_B, pcis->vec1_D, applytranspose));
6418: /* communications from coarse sol to local primal nodes */
6419: PetscCall(PCBDDCScatterCoarseDataBegin(pc, INSERT_VALUES, SCATTER_REVERSE));
6420: PetscCall(PCBDDCScatterCoarseDataEnd(pc, INSERT_VALUES, SCATTER_REVERSE));
6422: /* Sum contributions from the two levels */
6423: if (!pcbddc->benign_apply_coarse_only) {
6424: if (applytranspose) {
6425: PetscCall(MatMultAdd(pcbddc->coarse_psi_B, pcbddc->vec1_P, pcis->vec1_B, pcis->vec1_B));
6426: if (pcbddc->switch_static) PetscCall(MatMultAdd(pcbddc->coarse_psi_D, pcbddc->vec1_P, pcis->vec1_D, pcis->vec1_D));
6427: } else {
6428: PetscCall(MatMultAdd(pcbddc->coarse_phi_B, pcbddc->vec1_P, pcis->vec1_B, pcis->vec1_B));
6429: if (pcbddc->switch_static) PetscCall(MatMultAdd(pcbddc->coarse_phi_D, pcbddc->vec1_P, pcis->vec1_D, pcis->vec1_D));
6430: }
6431: /* store p0 */
6432: if (pcbddc->benign_n) {
6433: PetscScalar *array;
6434: PetscInt j;
6436: PetscCall(VecGetArray(pcbddc->vec1_P, &array));
6437: for (j = 0; j < pcbddc->benign_n; j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size - pcbddc->benign_n + j];
6438: PetscCall(VecRestoreArray(pcbddc->vec1_P, &array));
6439: }
6440: } else { /* expand the coarse solution */
6441: if (applytranspose) {
6442: PetscCall(MatMult(pcbddc->coarse_psi_B, pcbddc->vec1_P, pcis->vec1_B));
6443: } else {
6444: PetscCall(MatMult(pcbddc->coarse_phi_B, pcbddc->vec1_P, pcis->vec1_B));
6445: }
6446: }
6447: PetscFunctionReturn(PETSC_SUCCESS);
6448: }
6450: PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc, InsertMode imode, ScatterMode smode)
6451: {
6452: PC_BDDC *pcbddc = (PC_BDDC *)pc->data;
6453: Vec from, to;
6454: const PetscScalar *array;
6456: PetscFunctionBegin;
6457: if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */
6458: from = pcbddc->coarse_vec;
6459: to = pcbddc->vec1_P;
6460: if (pcbddc->coarse_ksp) { /* get array from coarse processes */
6461: Vec tvec;
6463: PetscCall(KSPGetRhs(pcbddc->coarse_ksp, &tvec));
6464: PetscCall(VecResetArray(tvec));
6465: PetscCall(KSPGetSolution(pcbddc->coarse_ksp, &tvec));
6466: PetscCall(VecGetArrayRead(tvec, &array));
6467: PetscCall(VecPlaceArray(from, array));
6468: PetscCall(VecRestoreArrayRead(tvec, &array));
6469: }
6470: } else { /* from local to global -> put data in coarse right-hand side */
6471: from = pcbddc->vec1_P;
6472: to = pcbddc->coarse_vec;
6473: }
6474: PetscCall(VecScatterBegin(pcbddc->coarse_loc_to_glob, from, to, imode, smode));
6475: PetscFunctionReturn(PETSC_SUCCESS);
6476: }
6478: PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode)
6479: {
6480: PC_BDDC *pcbddc = (PC_BDDC *)pc->data;
6481: Vec from, to;
6482: const PetscScalar *array;
6484: PetscFunctionBegin;
6485: if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */
6486: from = pcbddc->coarse_vec;
6487: to = pcbddc->vec1_P;
6488: } else { /* from local to global -> put data in coarse right-hand side */
6489: from = pcbddc->vec1_P;
6490: to = pcbddc->coarse_vec;
6491: }
6492: PetscCall(VecScatterEnd(pcbddc->coarse_loc_to_glob, from, to, imode, smode));
6493: if (smode == SCATTER_FORWARD) {
6494: if (pcbddc->coarse_ksp) { /* get array from coarse processes */
6495: Vec tvec;
6497: PetscCall(KSPGetRhs(pcbddc->coarse_ksp, &tvec));
6498: PetscCall(VecGetArrayRead(to, &array));
6499: PetscCall(VecPlaceArray(tvec, array));
6500: PetscCall(VecRestoreArrayRead(to, &array));
6501: }
6502: } else {
6503: if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */
6504: PetscCall(VecResetArray(from));
6505: }
6506: }
6507: PetscFunctionReturn(PETSC_SUCCESS);
6508: }
6510: PetscErrorCode PCBDDCConstraintsSetUp(PC pc)
6511: {
6512: PC_IS *pcis = (PC_IS *)pc->data;
6513: PC_BDDC *pcbddc = (PC_BDDC *)pc->data;
6514: Mat_IS *matis = (Mat_IS *)pc->pmat->data;
6515: /* one and zero */
6516: PetscScalar one = 1.0, zero = 0.0;
6517: /* space to store constraints and their local indices */
6518: PetscScalar *constraints_data;
6519: PetscInt *constraints_idxs, *constraints_idxs_B;
6520: PetscInt *constraints_idxs_ptr, *constraints_data_ptr;
6521: PetscInt *constraints_n;
6522: /* iterators */
6523: PetscInt i, j, k, total_counts, total_counts_cc, cum;
6524: /* BLAS integers */
6525: PetscBLASInt lwork, lierr;
6526: PetscBLASInt Blas_N, Blas_M, Blas_K, Blas_one = 1;
6527: PetscBLASInt Blas_LDA, Blas_LDB, Blas_LDC;
6528: /* reuse */
6529: PetscInt olocal_primal_size, olocal_primal_size_cc;
6530: PetscInt *olocal_primal_ref_node, *olocal_primal_ref_mult;
6531: /* change of basis */
6532: PetscBool qr_needed;
6533: PetscBT change_basis, qr_needed_idx;
6534: /* auxiliary stuff */
6535: PetscInt *nnz, *is_indices;
6536: PetscInt ncc;
6537: /* some quantities */
6538: PetscInt n_vertices, total_primal_vertices, valid_constraints;
6539: PetscInt size_of_constraint, max_size_of_constraint = 0, max_constraints, temp_constraints;
6540: PetscReal tol; /* tolerance for retaining eigenmodes */
6542: PetscFunctionBegin;
6543: tol = PetscSqrtReal(PETSC_SMALL);
6544: /* Destroy Mat objects computed previously */
6545: PetscCall(MatDestroy(&pcbddc->ChangeOfBasisMatrix));
6546: PetscCall(MatDestroy(&pcbddc->ConstraintMatrix));
6547: PetscCall(MatDestroy(&pcbddc->switch_static_change));
6548: /* save info on constraints from previous setup (if any) */
6549: olocal_primal_size = pcbddc->local_primal_size;
6550: olocal_primal_size_cc = pcbddc->local_primal_size_cc;
6551: PetscCall(PetscMalloc2(olocal_primal_size_cc, &olocal_primal_ref_node, olocal_primal_size_cc, &olocal_primal_ref_mult));
6552: PetscCall(PetscArraycpy(olocal_primal_ref_node, pcbddc->local_primal_ref_node, olocal_primal_size_cc));
6553: PetscCall(PetscArraycpy(olocal_primal_ref_mult, pcbddc->local_primal_ref_mult, olocal_primal_size_cc));
6554: PetscCall(PetscFree2(pcbddc->local_primal_ref_node, pcbddc->local_primal_ref_mult));
6555: PetscCall(PetscFree(pcbddc->primal_indices_local_idxs));
6557: if (!pcbddc->adaptive_selection) {
6558: IS ISForVertices, *ISForFaces, *ISForEdges;
6559: MatNullSpace nearnullsp;
6560: const Vec *nearnullvecs;
6561: Vec *localnearnullsp;
6562: PetscScalar *array;
6563: PetscInt n_ISForFaces, n_ISForEdges, nnsp_size, o_nf, o_ne;
6564: PetscBool nnsp_has_cnst;
6565: /* LAPACK working arrays for SVD or POD */
6566: PetscBool skip_lapack, boolforchange;
6567: PetscScalar *work;
6568: PetscReal *singular_vals;
6569: #if defined(PETSC_USE_COMPLEX)
6570: PetscReal *rwork;
6571: #endif
6572: PetscScalar *temp_basis = NULL, *correlation_mat = NULL;
6573: PetscBLASInt dummy_int = 1;
6574: PetscScalar dummy_scalar = 1.;
6575: PetscBool use_pod = PetscDefined(MISSING_LAPACK_GESVD) || PetscDefined(HAVE_MKL_LIBS) ? PETSC_TRUE : PETSC_FALSE; /* MKL SVD with same input gives different results on different processes! */
6577: /* Get index sets for faces, edges and vertices from graph */
6578: PetscCall(PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph, &n_ISForFaces, &ISForFaces, &n_ISForEdges, &ISForEdges, &ISForVertices));
6579: o_nf = n_ISForFaces;
6580: o_ne = n_ISForEdges;
6581: n_vertices = 0;
6582: if (ISForVertices) PetscCall(ISGetSize(ISForVertices, &n_vertices));
6583: /* print some info */
6584: if (pcbddc->dbg_flag && (!pcbddc->sub_schurs || pcbddc->sub_schurs_rebuild)) {
6585: if (!pcbddc->dbg_viewer) pcbddc->dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pc));
6586: PetscCall(PCBDDCGraphASCIIView(pcbddc->mat_graph, pcbddc->dbg_flag, pcbddc->dbg_viewer));
6587: PetscCall(PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer));
6588: PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, "--------------------------------------------------------------\n"));
6589: PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, "Subdomain %04d got %02" PetscInt_FMT " local candidate vertices (%d)\n", PetscGlobalRank, n_vertices, pcbddc->use_vertices));
6590: PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, "Subdomain %04d got %02" PetscInt_FMT " local candidate edges (%d)\n", PetscGlobalRank, n_ISForEdges, pcbddc->use_edges));
6591: PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, "Subdomain %04d got %02" PetscInt_FMT " local candidate faces (%d)\n", PetscGlobalRank, n_ISForFaces, pcbddc->use_faces));
6592: PetscCall(PetscViewerFlush(pcbddc->dbg_viewer));
6593: PetscCall(PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer));
6594: }
6596: if (!pcbddc->use_vertices) n_vertices = 0;
6597: if (!pcbddc->use_edges) n_ISForEdges = 0;
6598: if (!pcbddc->use_faces) n_ISForFaces = 0;
6600: /* check if near null space is attached to global mat */
6601: if (pcbddc->use_nnsp) PetscCall(MatGetNearNullSpace(pc->pmat, &nearnullsp));
6602: else nearnullsp = NULL;
6604: if (nearnullsp) {
6605: PetscCall(MatNullSpaceGetVecs(nearnullsp, &nnsp_has_cnst, &nnsp_size, &nearnullvecs));
6606: /* remove any stored info */
6607: PetscCall(MatNullSpaceDestroy(&pcbddc->onearnullspace));
6608: PetscCall(PetscFree(pcbddc->onearnullvecs_state));
6609: /* store information for BDDC solver reuse */
6610: PetscCall(PetscObjectReference((PetscObject)nearnullsp));
6611: pcbddc->onearnullspace = nearnullsp;
6612: PetscCall(PetscMalloc1(nnsp_size, &pcbddc->onearnullvecs_state));
6613: for (i = 0; i < nnsp_size; i++) PetscCall(PetscObjectStateGet((PetscObject)nearnullvecs[i], &pcbddc->onearnullvecs_state[i]));
6614: } else { /* if near null space is not provided BDDC uses constants by default */
6615: nnsp_size = 0;
6616: nnsp_has_cnst = PETSC_TRUE;
6617: }
6618: /* get max number of constraints on a single cc */
6619: max_constraints = nnsp_size;
6620: if (nnsp_has_cnst) max_constraints++;
6622: /*
6623: Evaluate maximum storage size needed by the procedure
6624: - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]"
6625: - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]"
6626: There can be multiple constraints per connected component
6627: */
6628: ncc = n_vertices + n_ISForFaces + n_ISForEdges;
6629: PetscCall(PetscMalloc3(ncc + 1, &constraints_idxs_ptr, ncc + 1, &constraints_data_ptr, ncc, &constraints_n));
6631: total_counts = n_ISForFaces + n_ISForEdges;
6632: total_counts *= max_constraints;
6633: total_counts += n_vertices;
6634: PetscCall(PetscBTCreate(total_counts, &change_basis));
6636: total_counts = 0;
6637: max_size_of_constraint = 0;
6638: for (i = 0; i < n_ISForEdges + n_ISForFaces; i++) {
6639: IS used_is;
6640: if (i < n_ISForEdges) {
6641: used_is = ISForEdges[i];
6642: } else {
6643: used_is = ISForFaces[i - n_ISForEdges];
6644: }
6645: PetscCall(ISGetSize(used_is, &j));
6646: total_counts += j;
6647: max_size_of_constraint = PetscMax(j, max_size_of_constraint);
6648: }
6649: PetscCall(PetscMalloc3(total_counts * max_constraints + n_vertices, &constraints_data, total_counts + n_vertices, &constraints_idxs, total_counts + n_vertices, &constraints_idxs_B));
6651: /* get local part of global near null space vectors */
6652: PetscCall(PetscMalloc1(nnsp_size, &localnearnullsp));
6653: for (k = 0; k < nnsp_size; k++) {
6654: PetscCall(VecDuplicate(pcis->vec1_N, &localnearnullsp[k]));
6655: PetscCall(VecScatterBegin(matis->rctx, nearnullvecs[k], localnearnullsp[k], INSERT_VALUES, SCATTER_FORWARD));
6656: PetscCall(VecScatterEnd(matis->rctx, nearnullvecs[k], localnearnullsp[k], INSERT_VALUES, SCATTER_FORWARD));
6657: }
6659: /* whether or not to skip lapack calls */
6660: skip_lapack = PETSC_TRUE;
6661: if (n_ISForFaces + n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE;
6663: /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */
6664: if (!skip_lapack) {
6665: PetscScalar temp_work;
6667: if (use_pod) {
6668: /* Proper Orthogonal Decomposition (POD) using the snapshot method */
6669: PetscCall(PetscMalloc1(max_constraints * max_constraints, &correlation_mat));
6670: PetscCall(PetscMalloc1(max_constraints, &singular_vals));
6671: PetscCall(PetscMalloc1(max_size_of_constraint * max_constraints, &temp_basis));
6672: #if defined(PETSC_USE_COMPLEX)
6673: PetscCall(PetscMalloc1(3 * max_constraints, &rwork));
6674: #endif
6675: /* now we evaluate the optimal workspace using query with lwork=-1 */
6676: PetscCall(PetscBLASIntCast(max_constraints, &Blas_N));
6677: PetscCall(PetscBLASIntCast(max_constraints, &Blas_LDA));
6678: lwork = -1;
6679: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
6680: #if !defined(PETSC_USE_COMPLEX)
6681: PetscCallBLAS("LAPACKsyev", LAPACKsyev_("V", "U", &Blas_N, correlation_mat, &Blas_LDA, singular_vals, &temp_work, &lwork, &lierr));
6682: #else
6683: PetscCallBLAS("LAPACKsyev", LAPACKsyev_("V", "U", &Blas_N, correlation_mat, &Blas_LDA, singular_vals, &temp_work, &lwork, rwork, &lierr));
6684: #endif
6685: PetscCall(PetscFPTrapPop());
6686: PetscCheck(!lierr, PETSC_COMM_SELF, PETSC_ERR_LIB, "Error in query to SYEV Lapack routine %" PetscBLASInt_FMT, lierr);
6687: } else {
6688: #if !defined(PETSC_MISSING_LAPACK_GESVD)
6689: /* SVD */
6690: PetscInt max_n, min_n;
6691: max_n = max_size_of_constraint;
6692: min_n = max_constraints;
6693: if (max_size_of_constraint < max_constraints) {
6694: min_n = max_size_of_constraint;
6695: max_n = max_constraints;
6696: }
6697: PetscCall(PetscMalloc1(min_n, &singular_vals));
6698: #if defined(PETSC_USE_COMPLEX)
6699: PetscCall(PetscMalloc1(5 * min_n, &rwork));
6700: #endif
6701: /* now we evaluate the optimal workspace using query with lwork=-1 */
6702: lwork = -1;
6703: PetscCall(PetscBLASIntCast(max_n, &Blas_M));
6704: PetscCall(PetscBLASIntCast(min_n, &Blas_N));
6705: PetscCall(PetscBLASIntCast(max_n, &Blas_LDA));
6706: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
6707: #if !defined(PETSC_USE_COMPLEX)
6708: PetscCallBLAS("LAPACKgesvd", LAPACKgesvd_("O", "N", &Blas_M, &Blas_N, &constraints_data[0], &Blas_LDA, singular_vals, &dummy_scalar, &dummy_int, &dummy_scalar, &dummy_int, &temp_work, &lwork, &lierr));
6709: #else
6710: PetscCallBLAS("LAPACKgesvd", LAPACKgesvd_("O", "N", &Blas_M, &Blas_N, &constraints_data[0], &Blas_LDA, singular_vals, &dummy_scalar, &dummy_int, &dummy_scalar, &dummy_int, &temp_work, &lwork, rwork, &lierr));
6711: #endif
6712: PetscCall(PetscFPTrapPop());
6713: PetscCheck(!lierr, PETSC_COMM_SELF, PETSC_ERR_LIB, "Error in query to GESVD Lapack routine %" PetscBLASInt_FMT, lierr);
6714: #else
6715: SETERRQ(PETSC_COMM_SELF, PETSC_ERR_LIB, "This should not happen");
6716: #endif /* on missing GESVD */
6717: }
6718: /* Allocate optimal workspace */
6719: PetscCall(PetscBLASIntCast((PetscInt)PetscRealPart(temp_work), &lwork));
6720: PetscCall(PetscMalloc1(lwork, &work));
6721: }
6722: /* Now we can loop on constraining sets */
6723: total_counts = 0;
6724: constraints_idxs_ptr[0] = 0;
6725: constraints_data_ptr[0] = 0;
6726: /* vertices */
6727: if (n_vertices) {
6728: PetscCall(ISGetIndices(ISForVertices, (const PetscInt **)&is_indices));
6729: PetscCall(PetscArraycpy(constraints_idxs, is_indices, n_vertices));
6730: for (i = 0; i < n_vertices; i++) {
6731: constraints_n[total_counts] = 1;
6732: constraints_data[total_counts] = 1.0;
6733: constraints_idxs_ptr[total_counts + 1] = constraints_idxs_ptr[total_counts] + 1;
6734: constraints_data_ptr[total_counts + 1] = constraints_data_ptr[total_counts] + 1;
6735: total_counts++;
6736: }
6737: PetscCall(ISRestoreIndices(ISForVertices, (const PetscInt **)&is_indices));
6738: }
6740: /* edges and faces */
6741: total_counts_cc = total_counts;
6742: for (ncc = 0; ncc < n_ISForEdges + n_ISForFaces; ncc++) {
6743: IS used_is;
6744: PetscBool idxs_copied = PETSC_FALSE;
6746: if (ncc < n_ISForEdges) {
6747: used_is = ISForEdges[ncc];
6748: boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */
6749: } else {
6750: used_is = ISForFaces[ncc - n_ISForEdges];
6751: boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */
6752: }
6753: temp_constraints = 0; /* zero the number of constraints I have on this conn comp */
6755: PetscCall(ISGetSize(used_is, &size_of_constraint));
6756: if (!size_of_constraint) continue;
6757: PetscCall(ISGetIndices(used_is, (const PetscInt **)&is_indices));
6758: if (nnsp_has_cnst) {
6759: PetscScalar quad_value;
6761: PetscCall(PetscArraycpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc], is_indices, size_of_constraint));
6762: idxs_copied = PETSC_TRUE;
6764: if (!pcbddc->use_nnsp_true) {
6765: quad_value = (PetscScalar)(1.0 / PetscSqrtReal((PetscReal)size_of_constraint));
6766: } else {
6767: quad_value = 1.0;
6768: }
6769: for (j = 0; j < size_of_constraint; j++) constraints_data[constraints_data_ptr[total_counts_cc] + j] = quad_value;
6770: temp_constraints++;
6771: total_counts++;
6772: }
6773: for (k = 0; k < nnsp_size; k++) {
6774: PetscReal real_value;
6775: PetscScalar *ptr_to_data;
6777: PetscCall(VecGetArrayRead(localnearnullsp[k], (const PetscScalar **)&array));
6778: ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc] + temp_constraints * size_of_constraint];
6779: for (j = 0; j < size_of_constraint; j++) ptr_to_data[j] = array[is_indices[j]];
6780: PetscCall(VecRestoreArrayRead(localnearnullsp[k], (const PetscScalar **)&array));
6781: /* check if array is null on the connected component */
6782: PetscCall(PetscBLASIntCast(size_of_constraint, &Blas_N));
6783: PetscCallBLAS("BLASasum", real_value = BLASasum_(&Blas_N, ptr_to_data, &Blas_one));
6784: if (real_value > tol * size_of_constraint) { /* keep indices and values */
6785: temp_constraints++;
6786: total_counts++;
6787: if (!idxs_copied) {
6788: PetscCall(PetscArraycpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc], is_indices, size_of_constraint));
6789: idxs_copied = PETSC_TRUE;
6790: }
6791: }
6792: }
6793: PetscCall(ISRestoreIndices(used_is, (const PetscInt **)&is_indices));
6794: valid_constraints = temp_constraints;
6795: if (!pcbddc->use_nnsp_true && temp_constraints) {
6796: if (temp_constraints == 1) { /* just normalize the constraint */
6797: PetscScalar norm, *ptr_to_data;
6799: ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]];
6800: PetscCall(PetscBLASIntCast(size_of_constraint, &Blas_N));
6801: PetscCallBLAS("BLASdot", norm = BLASdot_(&Blas_N, ptr_to_data, &Blas_one, ptr_to_data, &Blas_one));
6802: norm = 1.0 / PetscSqrtReal(PetscRealPart(norm));
6803: PetscCallBLAS("BLASscal", BLASscal_(&Blas_N, &norm, ptr_to_data, &Blas_one));
6804: } else { /* perform SVD */
6805: PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]];
6807: if (use_pod) {
6808: /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag
6809: POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2)
6810: -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined
6811: the constraints basis will differ (by a complex factor with absolute value equal to 1)
6812: from that computed using LAPACKgesvd
6813: -> This is due to a different computation of eigenvectors in LAPACKheev
6814: -> The quality of the POD-computed basis will be the same */
6815: PetscCall(PetscArrayzero(correlation_mat, temp_constraints * temp_constraints));
6816: /* Store upper triangular part of correlation matrix */
6817: PetscCall(PetscBLASIntCast(size_of_constraint, &Blas_N));
6818: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
6819: for (j = 0; j < temp_constraints; j++) {
6820: for (k = 0; k < j + 1; k++) PetscCallBLAS("BLASdot", correlation_mat[j * temp_constraints + k] = BLASdot_(&Blas_N, ptr_to_data + k * size_of_constraint, &Blas_one, ptr_to_data + j * size_of_constraint, &Blas_one));
6821: }
6822: /* compute eigenvalues and eigenvectors of correlation matrix */
6823: PetscCall(PetscBLASIntCast(temp_constraints, &Blas_N));
6824: PetscCall(PetscBLASIntCast(temp_constraints, &Blas_LDA));
6825: #if !defined(PETSC_USE_COMPLEX)
6826: PetscCallBLAS("LAPACKsyev", LAPACKsyev_("V", "U", &Blas_N, correlation_mat, &Blas_LDA, singular_vals, work, &lwork, &lierr));
6827: #else
6828: PetscCallBLAS("LAPACKsyev", LAPACKsyev_("V", "U", &Blas_N, correlation_mat, &Blas_LDA, singular_vals, work, &lwork, rwork, &lierr));
6829: #endif
6830: PetscCall(PetscFPTrapPop());
6831: PetscCheck(!lierr, PETSC_COMM_SELF, PETSC_ERR_LIB, "Error in SYEV Lapack routine %" PetscBLASInt_FMT, lierr);
6832: /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */
6833: j = 0;
6834: while (j < temp_constraints && singular_vals[j] / singular_vals[temp_constraints - 1] < tol) j++;
6835: total_counts = total_counts - j;
6836: valid_constraints = temp_constraints - j;
6837: /* scale and copy POD basis into used quadrature memory */
6838: PetscCall(PetscBLASIntCast(size_of_constraint, &Blas_M));
6839: PetscCall(PetscBLASIntCast(temp_constraints, &Blas_N));
6840: PetscCall(PetscBLASIntCast(temp_constraints, &Blas_K));
6841: PetscCall(PetscBLASIntCast(size_of_constraint, &Blas_LDA));
6842: PetscCall(PetscBLASIntCast(temp_constraints, &Blas_LDB));
6843: PetscCall(PetscBLASIntCast(size_of_constraint, &Blas_LDC));
6844: if (j < temp_constraints) {
6845: PetscInt ii;
6846: for (k = j; k < temp_constraints; k++) singular_vals[k] = 1.0 / PetscSqrtReal(singular_vals[k]);
6847: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
6848: PetscCallBLAS("BLASgemm", BLASgemm_("N", "N", &Blas_M, &Blas_N, &Blas_K, &one, ptr_to_data, &Blas_LDA, correlation_mat, &Blas_LDB, &zero, temp_basis, &Blas_LDC));
6849: PetscCall(PetscFPTrapPop());
6850: for (k = 0; k < temp_constraints - j; k++) {
6851: for (ii = 0; ii < size_of_constraint; ii++) ptr_to_data[k * size_of_constraint + ii] = singular_vals[temp_constraints - 1 - k] * temp_basis[(temp_constraints - 1 - k) * size_of_constraint + ii];
6852: }
6853: }
6854: } else {
6855: #if !defined(PETSC_MISSING_LAPACK_GESVD)
6856: PetscCall(PetscBLASIntCast(size_of_constraint, &Blas_M));
6857: PetscCall(PetscBLASIntCast(temp_constraints, &Blas_N));
6858: PetscCall(PetscBLASIntCast(size_of_constraint, &Blas_LDA));
6859: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
6860: #if !defined(PETSC_USE_COMPLEX)
6861: PetscCallBLAS("LAPACKgesvd", LAPACKgesvd_("O", "N", &Blas_M, &Blas_N, ptr_to_data, &Blas_LDA, singular_vals, &dummy_scalar, &dummy_int, &dummy_scalar, &dummy_int, work, &lwork, &lierr));
6862: #else
6863: PetscCallBLAS("LAPACKgesvd", LAPACKgesvd_("O", "N", &Blas_M, &Blas_N, ptr_to_data, &Blas_LDA, singular_vals, &dummy_scalar, &dummy_int, &dummy_scalar, &dummy_int, work, &lwork, rwork, &lierr));
6864: #endif
6865: PetscCheck(!lierr, PETSC_COMM_SELF, PETSC_ERR_LIB, "Error in GESVD Lapack routine %" PetscBLASInt_FMT, lierr);
6866: PetscCall(PetscFPTrapPop());
6867: /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */
6868: k = temp_constraints;
6869: if (k > size_of_constraint) k = size_of_constraint;
6870: j = 0;
6871: while (j < k && singular_vals[k - j - 1] / singular_vals[0] < tol) j++;
6872: valid_constraints = k - j;
6873: total_counts = total_counts - temp_constraints + valid_constraints;
6874: #else
6875: SETERRQ(PETSC_COMM_SELF, PETSC_ERR_LIB, "This should not happen");
6876: #endif /* on missing GESVD */
6877: }
6878: }
6879: }
6880: /* update pointers information */
6881: if (valid_constraints) {
6882: constraints_n[total_counts_cc] = valid_constraints;
6883: constraints_idxs_ptr[total_counts_cc + 1] = constraints_idxs_ptr[total_counts_cc] + size_of_constraint;
6884: constraints_data_ptr[total_counts_cc + 1] = constraints_data_ptr[total_counts_cc] + size_of_constraint * valid_constraints;
6885: /* set change_of_basis flag */
6886: if (boolforchange) PetscCall(PetscBTSet(change_basis, total_counts_cc));
6887: total_counts_cc++;
6888: }
6889: }
6890: /* free workspace */
6891: if (!skip_lapack) {
6892: PetscCall(PetscFree(work));
6893: #if defined(PETSC_USE_COMPLEX)
6894: PetscCall(PetscFree(rwork));
6895: #endif
6896: PetscCall(PetscFree(singular_vals));
6897: PetscCall(PetscFree(correlation_mat));
6898: PetscCall(PetscFree(temp_basis));
6899: }
6900: for (k = 0; k < nnsp_size; k++) PetscCall(VecDestroy(&localnearnullsp[k]));
6901: PetscCall(PetscFree(localnearnullsp));
6902: /* free index sets of faces, edges and vertices */
6903: PetscCall(PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph, &o_nf, &ISForFaces, &o_ne, &ISForEdges, &ISForVertices));
6904: } else {
6905: PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs;
6907: total_counts = 0;
6908: n_vertices = 0;
6909: if (sub_schurs->is_vertices && pcbddc->use_vertices) PetscCall(ISGetLocalSize(sub_schurs->is_vertices, &n_vertices));
6910: max_constraints = 0;
6911: total_counts_cc = 0;
6912: for (i = 0; i < sub_schurs->n_subs + n_vertices; i++) {
6913: total_counts += pcbddc->adaptive_constraints_n[i];
6914: if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++;
6915: max_constraints = PetscMax(max_constraints, pcbddc->adaptive_constraints_n[i]);
6916: }
6917: constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr;
6918: constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr;
6919: constraints_idxs = pcbddc->adaptive_constraints_idxs;
6920: constraints_data = pcbddc->adaptive_constraints_data;
6921: /* constraints_n differs from pcbddc->adaptive_constraints_n */
6922: PetscCall(PetscMalloc1(total_counts_cc, &constraints_n));
6923: total_counts_cc = 0;
6924: for (i = 0; i < sub_schurs->n_subs + n_vertices; i++) {
6925: if (pcbddc->adaptive_constraints_n[i]) constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i];
6926: }
6928: max_size_of_constraint = 0;
6929: for (i = 0; i < total_counts_cc; i++) max_size_of_constraint = PetscMax(max_size_of_constraint, constraints_idxs_ptr[i + 1] - constraints_idxs_ptr[i]);
6930: PetscCall(PetscMalloc1(constraints_idxs_ptr[total_counts_cc], &constraints_idxs_B));
6931: /* Change of basis */
6932: PetscCall(PetscBTCreate(total_counts_cc, &change_basis));
6933: if (pcbddc->use_change_of_basis) {
6934: for (i = 0; i < sub_schurs->n_subs; i++) {
6935: if (PetscBTLookup(sub_schurs->is_edge, i) || pcbddc->use_change_on_faces) PetscCall(PetscBTSet(change_basis, i + n_vertices));
6936: }
6937: }
6938: }
6939: pcbddc->local_primal_size = total_counts;
6940: PetscCall(PetscMalloc1(pcbddc->local_primal_size + pcbddc->benign_n, &pcbddc->primal_indices_local_idxs));
6942: /* map constraints_idxs in boundary numbering */
6943: if (pcbddc->use_change_of_basis) {
6944: PetscCall(ISGlobalToLocalMappingApply(pcis->BtoNmap, IS_GTOLM_DROP, constraints_idxs_ptr[total_counts_cc], constraints_idxs, &i, constraints_idxs_B));
6945: PetscCheck(i == constraints_idxs_ptr[total_counts_cc], PETSC_COMM_SELF, PETSC_ERR_PLIB, "Error in boundary numbering for constraints indices %" PetscInt_FMT " != %" PetscInt_FMT, constraints_idxs_ptr[total_counts_cc], i);
6946: }
6948: /* Create constraint matrix */
6949: PetscCall(MatCreate(PETSC_COMM_SELF, &pcbddc->ConstraintMatrix));
6950: PetscCall(MatSetType(pcbddc->ConstraintMatrix, MATAIJ));
6951: PetscCall(MatSetSizes(pcbddc->ConstraintMatrix, pcbddc->local_primal_size, pcis->n, pcbddc->local_primal_size, pcis->n));
6953: /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */
6954: /* determine if a QR strategy is needed for change of basis */
6955: qr_needed = pcbddc->use_qr_single;
6956: PetscCall(PetscBTCreate(total_counts_cc, &qr_needed_idx));
6957: total_primal_vertices = 0;
6958: pcbddc->local_primal_size_cc = 0;
6959: for (i = 0; i < total_counts_cc; i++) {
6960: size_of_constraint = constraints_idxs_ptr[i + 1] - constraints_idxs_ptr[i];
6961: if (size_of_constraint == 1 && pcbddc->mat_graph->custom_minimal_size) {
6962: pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]];
6963: pcbddc->local_primal_size_cc += 1;
6964: } else if (PetscBTLookup(change_basis, i)) {
6965: for (k = 0; k < constraints_n[i]; k++) pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i] + k];
6966: pcbddc->local_primal_size_cc += constraints_n[i];
6967: if (constraints_n[i] > 1 || pcbddc->use_qr_single) {
6968: PetscCall(PetscBTSet(qr_needed_idx, i));
6969: qr_needed = PETSC_TRUE;
6970: }
6971: } else {
6972: pcbddc->local_primal_size_cc += 1;
6973: }
6974: }
6975: /* note that the local variable n_vertices used below stores the number of pointwise constraints */
6976: pcbddc->n_vertices = total_primal_vertices;
6977: /* permute indices in order to have a sorted set of vertices */
6978: PetscCall(PetscSortInt(total_primal_vertices, pcbddc->primal_indices_local_idxs));
6979: PetscCall(PetscMalloc2(pcbddc->local_primal_size_cc + pcbddc->benign_n, &pcbddc->local_primal_ref_node, pcbddc->local_primal_size_cc + pcbddc->benign_n, &pcbddc->local_primal_ref_mult));
6980: PetscCall(PetscArraycpy(pcbddc->local_primal_ref_node, pcbddc->primal_indices_local_idxs, total_primal_vertices));
6981: for (i = 0; i < total_primal_vertices; i++) pcbddc->local_primal_ref_mult[i] = 1;
6983: /* nonzero structure of constraint matrix */
6984: /* and get reference dof for local constraints */
6985: PetscCall(PetscMalloc1(pcbddc->local_primal_size, &nnz));
6986: for (i = 0; i < total_primal_vertices; i++) nnz[i] = 1;
6988: j = total_primal_vertices;
6989: total_counts = total_primal_vertices;
6990: cum = total_primal_vertices;
6991: for (i = n_vertices; i < total_counts_cc; i++) {
6992: if (!PetscBTLookup(change_basis, i)) {
6993: pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]];
6994: pcbddc->local_primal_ref_mult[cum] = constraints_n[i];
6995: cum++;
6996: size_of_constraint = constraints_idxs_ptr[i + 1] - constraints_idxs_ptr[i];
6997: for (k = 0; k < constraints_n[i]; k++) {
6998: pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i] + k];
6999: nnz[j + k] = size_of_constraint;
7000: }
7001: j += constraints_n[i];
7002: }
7003: }
7004: PetscCall(MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix, 0, nnz));
7005: PetscCall(MatSetOption(pcbddc->ConstraintMatrix, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_TRUE));
7006: PetscCall(MatSetOption(pcbddc->ConstraintMatrix, MAT_IGNORE_ZERO_ENTRIES, PETSC_TRUE));
7007: PetscCall(PetscFree(nnz));
7009: /* set values in constraint matrix */
7010: for (i = 0; i < total_primal_vertices; i++) PetscCall(MatSetValue(pcbddc->ConstraintMatrix, i, pcbddc->local_primal_ref_node[i], 1.0, INSERT_VALUES));
7011: total_counts = total_primal_vertices;
7012: for (i = n_vertices; i < total_counts_cc; i++) {
7013: if (!PetscBTLookup(change_basis, i)) {
7014: PetscInt *cols;
7016: size_of_constraint = constraints_idxs_ptr[i + 1] - constraints_idxs_ptr[i];
7017: cols = constraints_idxs + constraints_idxs_ptr[i];
7018: for (k = 0; k < constraints_n[i]; k++) {
7019: PetscInt row = total_counts + k;
7020: PetscScalar *vals;
7022: vals = constraints_data + constraints_data_ptr[i] + k * size_of_constraint;
7023: PetscCall(MatSetValues(pcbddc->ConstraintMatrix, 1, &row, size_of_constraint, cols, vals, INSERT_VALUES));
7024: }
7025: total_counts += constraints_n[i];
7026: }
7027: }
7028: /* assembling */
7029: PetscCall(MatAssemblyBegin(pcbddc->ConstraintMatrix, MAT_FINAL_ASSEMBLY));
7030: PetscCall(MatAssemblyEnd(pcbddc->ConstraintMatrix, MAT_FINAL_ASSEMBLY));
7031: PetscCall(MatViewFromOptions(pcbddc->ConstraintMatrix, (PetscObject)pc, "-pc_bddc_constraint_mat_view"));
7033: /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */
7034: if (pcbddc->use_change_of_basis) {
7035: /* dual and primal dofs on a single cc */
7036: PetscInt dual_dofs, primal_dofs;
7037: /* working stuff for GEQRF */
7038: PetscScalar *qr_basis = NULL, *qr_tau = NULL, *qr_work = NULL, lqr_work_t;
7039: PetscBLASInt lqr_work;
7040: /* working stuff for UNGQR */
7041: PetscScalar *gqr_work = NULL, lgqr_work_t = 0.0;
7042: PetscBLASInt lgqr_work;
7043: /* working stuff for TRTRS */
7044: PetscScalar *trs_rhs = NULL;
7045: PetscBLASInt Blas_NRHS;
7046: /* pointers for values insertion into change of basis matrix */
7047: PetscInt *start_rows, *start_cols;
7048: PetscScalar *start_vals;
7049: /* working stuff for values insertion */
7050: PetscBT is_primal;
7051: PetscInt *aux_primal_numbering_B;
7052: /* matrix sizes */
7053: PetscInt global_size, local_size;
7054: /* temporary change of basis */
7055: Mat localChangeOfBasisMatrix;
7056: /* extra space for debugging */
7057: PetscScalar *dbg_work = NULL;
7059: PetscCall(MatCreate(PETSC_COMM_SELF, &localChangeOfBasisMatrix));
7060: PetscCall(MatSetType(localChangeOfBasisMatrix, MATAIJ));
7061: PetscCall(MatSetSizes(localChangeOfBasisMatrix, pcis->n, pcis->n, pcis->n, pcis->n));
7062: /* nonzeros for local mat */
7063: PetscCall(PetscMalloc1(pcis->n, &nnz));
7064: if (!pcbddc->benign_change || pcbddc->fake_change) {
7065: for (i = 0; i < pcis->n; i++) nnz[i] = 1;
7066: } else {
7067: const PetscInt *ii;
7068: PetscInt n;
7069: PetscBool flg_row;
7070: PetscCall(MatGetRowIJ(pcbddc->benign_change, 0, PETSC_FALSE, PETSC_FALSE, &n, &ii, NULL, &flg_row));
7071: for (i = 0; i < n; i++) nnz[i] = ii[i + 1] - ii[i];
7072: PetscCall(MatRestoreRowIJ(pcbddc->benign_change, 0, PETSC_FALSE, PETSC_FALSE, &n, &ii, NULL, &flg_row));
7073: }
7074: for (i = n_vertices; i < total_counts_cc; i++) {
7075: if (PetscBTLookup(change_basis, i)) {
7076: size_of_constraint = constraints_idxs_ptr[i + 1] - constraints_idxs_ptr[i];
7077: if (PetscBTLookup(qr_needed_idx, i)) {
7078: for (j = 0; j < size_of_constraint; j++) nnz[constraints_idxs[constraints_idxs_ptr[i] + j]] = size_of_constraint;
7079: } else {
7080: nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint;
7081: for (j = 1; j < size_of_constraint; j++) nnz[constraints_idxs[constraints_idxs_ptr[i] + j]] = 2;
7082: }
7083: }
7084: }
7085: PetscCall(MatSeqAIJSetPreallocation(localChangeOfBasisMatrix, 0, nnz));
7086: PetscCall(MatSetOption(localChangeOfBasisMatrix, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_TRUE));
7087: PetscCall(PetscFree(nnz));
7088: /* Set interior change in the matrix */
7089: if (!pcbddc->benign_change || pcbddc->fake_change) {
7090: for (i = 0; i < pcis->n; i++) PetscCall(MatSetValue(localChangeOfBasisMatrix, i, i, 1.0, INSERT_VALUES));
7091: } else {
7092: const PetscInt *ii, *jj;
7093: PetscScalar *aa;
7094: PetscInt n;
7095: PetscBool flg_row;
7096: PetscCall(MatGetRowIJ(pcbddc->benign_change, 0, PETSC_FALSE, PETSC_FALSE, &n, &ii, &jj, &flg_row));
7097: PetscCall(MatSeqAIJGetArray(pcbddc->benign_change, &aa));
7098: for (i = 0; i < n; i++) PetscCall(MatSetValues(localChangeOfBasisMatrix, 1, &i, ii[i + 1] - ii[i], jj + ii[i], aa + ii[i], INSERT_VALUES));
7099: PetscCall(MatSeqAIJRestoreArray(pcbddc->benign_change, &aa));
7100: PetscCall(MatRestoreRowIJ(pcbddc->benign_change, 0, PETSC_FALSE, PETSC_FALSE, &n, &ii, &jj, &flg_row));
7101: }
7103: if (pcbddc->dbg_flag) {
7104: PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, "--------------------------------------------------------------\n"));
7105: PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, "Checking change of basis computation for subdomain %04d\n", PetscGlobalRank));
7106: }
7108: /* Now we loop on the constraints which need a change of basis */
7109: /*
7110: Change of basis matrix is evaluated similarly to the FIRST APPROACH in
7111: Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1)
7113: Basic blocks of change of basis matrix T computed:
7115: - By using the following block transformation if there is only a primal dof on the cc (and -pc_bddc_use_qr_single is not specified)
7117: | 1 0 ... 0 s_1/S |
7118: | 0 1 ... 0 s_2/S |
7119: | ... |
7120: | 0 ... 1 s_{n-1}/S |
7121: | -s_1/s_n ... -s_{n-1}/s_n s_n/S |
7123: with S = \sum_{i=1}^n s_i^2
7124: NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering
7125: in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering
7127: - QR decomposition of constraints otherwise
7128: */
7129: if (qr_needed && max_size_of_constraint) {
7130: /* space to store Q */
7131: PetscCall(PetscMalloc1(max_size_of_constraint * max_size_of_constraint, &qr_basis));
7132: /* array to store scaling factors for reflectors */
7133: PetscCall(PetscMalloc1(max_constraints, &qr_tau));
7134: /* first we issue queries for optimal work */
7135: PetscCall(PetscBLASIntCast(max_size_of_constraint, &Blas_M));
7136: PetscCall(PetscBLASIntCast(max_constraints, &Blas_N));
7137: PetscCall(PetscBLASIntCast(max_size_of_constraint, &Blas_LDA));
7138: lqr_work = -1;
7139: PetscCallBLAS("LAPACKgeqrf", LAPACKgeqrf_(&Blas_M, &Blas_N, qr_basis, &Blas_LDA, qr_tau, &lqr_work_t, &lqr_work, &lierr));
7140: PetscCheck(!lierr, PETSC_COMM_SELF, PETSC_ERR_LIB, "Error in query to GEQRF Lapack routine %" PetscBLASInt_FMT, lierr);
7141: PetscCall(PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t), &lqr_work));
7142: PetscCall(PetscMalloc1(lqr_work, &qr_work));
7143: lgqr_work = -1;
7144: PetscCall(PetscBLASIntCast(max_size_of_constraint, &Blas_M));
7145: PetscCall(PetscBLASIntCast(max_size_of_constraint, &Blas_N));
7146: PetscCall(PetscBLASIntCast(max_constraints, &Blas_K));
7147: PetscCall(PetscBLASIntCast(max_size_of_constraint, &Blas_LDA));
7148: if (Blas_K > Blas_M) Blas_K = Blas_M; /* adjust just for computing optimal work */
7149: PetscCallBLAS("LAPACKorgqr", LAPACKorgqr_(&Blas_M, &Blas_N, &Blas_K, qr_basis, &Blas_LDA, qr_tau, &lgqr_work_t, &lgqr_work, &lierr));
7150: PetscCheck(!lierr, PETSC_COMM_SELF, PETSC_ERR_LIB, "Error in query to ORGQR/UNGQR Lapack routine %" PetscBLASInt_FMT, lierr);
7151: PetscCall(PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t), &lgqr_work));
7152: PetscCall(PetscMalloc1(lgqr_work, &gqr_work));
7153: /* array to store rhs and solution of triangular solver */
7154: PetscCall(PetscMalloc1(max_constraints * max_constraints, &trs_rhs));
7155: /* allocating workspace for check */
7156: if (pcbddc->dbg_flag) PetscCall(PetscMalloc1(max_size_of_constraint * (max_constraints + max_size_of_constraint), &dbg_work));
7157: }
7158: /* array to store whether a node is primal or not */
7159: PetscCall(PetscBTCreate(pcis->n_B, &is_primal));
7160: PetscCall(PetscMalloc1(total_primal_vertices, &aux_primal_numbering_B));
7161: PetscCall(ISGlobalToLocalMappingApply(pcis->BtoNmap, IS_GTOLM_DROP, total_primal_vertices, pcbddc->local_primal_ref_node, &i, aux_primal_numbering_B));
7162: PetscCheck(i == total_primal_vertices, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Error in boundary numbering for BDDC vertices! %" PetscInt_FMT " != %" PetscInt_FMT, total_primal_vertices, i);
7163: for (i = 0; i < total_primal_vertices; i++) PetscCall(PetscBTSet(is_primal, aux_primal_numbering_B[i]));
7164: PetscCall(PetscFree(aux_primal_numbering_B));
7166: /* loop on constraints and see whether or not they need a change of basis and compute it */
7167: for (total_counts = n_vertices; total_counts < total_counts_cc; total_counts++) {
7168: size_of_constraint = constraints_idxs_ptr[total_counts + 1] - constraints_idxs_ptr[total_counts];
7169: if (PetscBTLookup(change_basis, total_counts)) {
7170: /* get constraint info */
7171: primal_dofs = constraints_n[total_counts];
7172: dual_dofs = size_of_constraint - primal_dofs;
7174: if (pcbddc->dbg_flag) PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, "Constraints %" PetscInt_FMT ": %" PetscInt_FMT " need a change of basis (size %" PetscInt_FMT ")\n", total_counts, primal_dofs, size_of_constraint));
7176: if (PetscBTLookup(qr_needed_idx, total_counts)) { /* QR */
7178: /* copy quadrature constraints for change of basis check */
7179: if (pcbddc->dbg_flag) PetscCall(PetscArraycpy(dbg_work, &constraints_data[constraints_data_ptr[total_counts]], size_of_constraint * primal_dofs));
7180: /* copy temporary constraints into larger work vector (in order to store all columns of Q) */
7181: PetscCall(PetscArraycpy(qr_basis, &constraints_data[constraints_data_ptr[total_counts]], size_of_constraint * primal_dofs));
7183: /* compute QR decomposition of constraints */
7184: PetscCall(PetscBLASIntCast(size_of_constraint, &Blas_M));
7185: PetscCall(PetscBLASIntCast(primal_dofs, &Blas_N));
7186: PetscCall(PetscBLASIntCast(size_of_constraint, &Blas_LDA));
7187: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
7188: PetscCallBLAS("LAPACKgeqrf", LAPACKgeqrf_(&Blas_M, &Blas_N, qr_basis, &Blas_LDA, qr_tau, qr_work, &lqr_work, &lierr));
7189: PetscCheck(!lierr, PETSC_COMM_SELF, PETSC_ERR_LIB, "Error in GEQRF Lapack routine %" PetscBLASInt_FMT, lierr);
7190: PetscCall(PetscFPTrapPop());
7192: /* explicitly compute R^-T */
7193: PetscCall(PetscArrayzero(trs_rhs, primal_dofs * primal_dofs));
7194: for (j = 0; j < primal_dofs; j++) trs_rhs[j * (primal_dofs + 1)] = 1.0;
7195: PetscCall(PetscBLASIntCast(primal_dofs, &Blas_N));
7196: PetscCall(PetscBLASIntCast(primal_dofs, &Blas_NRHS));
7197: PetscCall(PetscBLASIntCast(size_of_constraint, &Blas_LDA));
7198: PetscCall(PetscBLASIntCast(primal_dofs, &Blas_LDB));
7199: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
7200: PetscCallBLAS("LAPACKtrtrs", LAPACKtrtrs_("U", "T", "N", &Blas_N, &Blas_NRHS, qr_basis, &Blas_LDA, trs_rhs, &Blas_LDB, &lierr));
7201: PetscCheck(!lierr, PETSC_COMM_SELF, PETSC_ERR_LIB, "Error in TRTRS Lapack routine %" PetscBLASInt_FMT, lierr);
7202: PetscCall(PetscFPTrapPop());
7204: /* explicitly compute all columns of Q (Q = [Q1 | Q2]) overwriting QR factorization in qr_basis */
7205: PetscCall(PetscBLASIntCast(size_of_constraint, &Blas_M));
7206: PetscCall(PetscBLASIntCast(size_of_constraint, &Blas_N));
7207: PetscCall(PetscBLASIntCast(primal_dofs, &Blas_K));
7208: PetscCall(PetscBLASIntCast(size_of_constraint, &Blas_LDA));
7209: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
7210: PetscCallBLAS("LAPACKorgqr", LAPACKorgqr_(&Blas_M, &Blas_N, &Blas_K, qr_basis, &Blas_LDA, qr_tau, gqr_work, &lgqr_work, &lierr));
7211: PetscCheck(!lierr, PETSC_COMM_SELF, PETSC_ERR_LIB, "Error in ORGQR/UNGQR Lapack routine %" PetscBLASInt_FMT, lierr);
7212: PetscCall(PetscFPTrapPop());
7214: /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints
7215: i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below)
7216: where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */
7217: PetscCall(PetscBLASIntCast(size_of_constraint, &Blas_M));
7218: PetscCall(PetscBLASIntCast(primal_dofs, &Blas_N));
7219: PetscCall(PetscBLASIntCast(primal_dofs, &Blas_K));
7220: PetscCall(PetscBLASIntCast(size_of_constraint, &Blas_LDA));
7221: PetscCall(PetscBLASIntCast(primal_dofs, &Blas_LDB));
7222: PetscCall(PetscBLASIntCast(size_of_constraint, &Blas_LDC));
7223: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
7224: PetscCallBLAS("BLASgemm", BLASgemm_("N", "N", &Blas_M, &Blas_N, &Blas_K, &one, qr_basis, &Blas_LDA, trs_rhs, &Blas_LDB, &zero, constraints_data + constraints_data_ptr[total_counts], &Blas_LDC));
7225: PetscCall(PetscFPTrapPop());
7226: PetscCall(PetscArraycpy(qr_basis, &constraints_data[constraints_data_ptr[total_counts]], size_of_constraint * primal_dofs));
7228: /* insert values in change of basis matrix respecting global ordering of new primal dofs */
7229: start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]];
7230: /* insert cols for primal dofs */
7231: for (j = 0; j < primal_dofs; j++) {
7232: start_vals = &qr_basis[j * size_of_constraint];
7233: start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts] + j];
7234: PetscCall(MatSetValues(localChangeOfBasisMatrix, size_of_constraint, start_rows, 1, start_cols, start_vals, INSERT_VALUES));
7235: }
7236: /* insert cols for dual dofs */
7237: for (j = 0, k = 0; j < dual_dofs; k++) {
7238: if (!PetscBTLookup(is_primal, constraints_idxs_B[constraints_idxs_ptr[total_counts] + k])) {
7239: start_vals = &qr_basis[(primal_dofs + j) * size_of_constraint];
7240: start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts] + k];
7241: PetscCall(MatSetValues(localChangeOfBasisMatrix, size_of_constraint, start_rows, 1, start_cols, start_vals, INSERT_VALUES));
7242: j++;
7243: }
7244: }
7246: /* check change of basis */
7247: if (pcbddc->dbg_flag) {
7248: PetscInt ii, jj;
7249: PetscBool valid_qr = PETSC_TRUE;
7250: PetscCall(PetscBLASIntCast(primal_dofs, &Blas_M));
7251: PetscCall(PetscBLASIntCast(size_of_constraint, &Blas_N));
7252: PetscCall(PetscBLASIntCast(size_of_constraint, &Blas_K));
7253: PetscCall(PetscBLASIntCast(size_of_constraint, &Blas_LDA));
7254: PetscCall(PetscBLASIntCast(size_of_constraint, &Blas_LDB));
7255: PetscCall(PetscBLASIntCast(primal_dofs, &Blas_LDC));
7256: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
7257: PetscCallBLAS("BLASgemm", BLASgemm_("T", "N", &Blas_M, &Blas_N, &Blas_K, &one, dbg_work, &Blas_LDA, qr_basis, &Blas_LDB, &zero, &dbg_work[size_of_constraint * primal_dofs], &Blas_LDC));
7258: PetscCall(PetscFPTrapPop());
7259: for (jj = 0; jj < size_of_constraint; jj++) {
7260: for (ii = 0; ii < primal_dofs; ii++) {
7261: if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint * primal_dofs + jj * primal_dofs + ii]) > 1.e-12) valid_qr = PETSC_FALSE;
7262: if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint * primal_dofs + jj * primal_dofs + ii] - (PetscReal)1) > 1.e-12) valid_qr = PETSC_FALSE;
7263: }
7264: }
7265: if (!valid_qr) {
7266: PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, "\t-> wrong change of basis!\n"));
7267: for (jj = 0; jj < size_of_constraint; jj++) {
7268: for (ii = 0; ii < primal_dofs; ii++) {
7269: if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint * primal_dofs + jj * primal_dofs + ii]) > 1.e-12) {
7270: PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, "\tQr basis function %" PetscInt_FMT " is not orthogonal to constraint %" PetscInt_FMT " (%1.14e)!\n", jj, ii, (double)PetscAbsScalar(dbg_work[size_of_constraint * primal_dofs + jj * primal_dofs + ii])));
7271: }
7272: if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint * primal_dofs + jj * primal_dofs + ii] - (PetscReal)1) > 1.e-12) {
7273: PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, "\tQr basis function %" PetscInt_FMT " is not unitary w.r.t constraint %" PetscInt_FMT " (%1.14e)!\n", jj, ii, (double)PetscAbsScalar(dbg_work[size_of_constraint * primal_dofs + jj * primal_dofs + ii])));
7274: }
7275: }
7276: }
7277: } else {
7278: PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, "\t-> right change of basis!\n"));
7279: }
7280: }
7281: } else { /* simple transformation block */
7282: PetscInt row, col;
7283: PetscScalar val, norm;
7285: PetscCall(PetscBLASIntCast(size_of_constraint, &Blas_N));
7286: PetscCallBLAS("BLASdot", norm = BLASdot_(&Blas_N, constraints_data + constraints_data_ptr[total_counts], &Blas_one, constraints_data + constraints_data_ptr[total_counts], &Blas_one));
7287: for (j = 0; j < size_of_constraint; j++) {
7288: PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts] + j];
7289: row = constraints_idxs[constraints_idxs_ptr[total_counts] + j];
7290: if (!PetscBTLookup(is_primal, row_B)) {
7291: col = constraints_idxs[constraints_idxs_ptr[total_counts]];
7292: PetscCall(MatSetValue(localChangeOfBasisMatrix, row, row, 1.0, INSERT_VALUES));
7293: PetscCall(MatSetValue(localChangeOfBasisMatrix, row, col, constraints_data[constraints_data_ptr[total_counts] + j] / norm, INSERT_VALUES));
7294: } else {
7295: for (k = 0; k < size_of_constraint; k++) {
7296: col = constraints_idxs[constraints_idxs_ptr[total_counts] + k];
7297: if (row != col) {
7298: val = -constraints_data[constraints_data_ptr[total_counts] + k] / constraints_data[constraints_data_ptr[total_counts]];
7299: } else {
7300: val = constraints_data[constraints_data_ptr[total_counts]] / norm;
7301: }
7302: PetscCall(MatSetValue(localChangeOfBasisMatrix, row, col, val, INSERT_VALUES));
7303: }
7304: }
7305: }
7306: if (pcbddc->dbg_flag) PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, "\t-> using standard change of basis\n"));
7307: }
7308: } else {
7309: if (pcbddc->dbg_flag) PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, "Constraint %" PetscInt_FMT " does not need a change of basis (size %" PetscInt_FMT ")\n", total_counts, size_of_constraint));
7310: }
7311: }
7313: /* free workspace */
7314: if (qr_needed) {
7315: if (pcbddc->dbg_flag) PetscCall(PetscFree(dbg_work));
7316: PetscCall(PetscFree(trs_rhs));
7317: PetscCall(PetscFree(qr_tau));
7318: PetscCall(PetscFree(qr_work));
7319: PetscCall(PetscFree(gqr_work));
7320: PetscCall(PetscFree(qr_basis));
7321: }
7322: PetscCall(PetscBTDestroy(&is_primal));
7323: PetscCall(MatAssemblyBegin(localChangeOfBasisMatrix, MAT_FINAL_ASSEMBLY));
7324: PetscCall(MatAssemblyEnd(localChangeOfBasisMatrix, MAT_FINAL_ASSEMBLY));
7326: /* assembling of global change of variable */
7327: if (!pcbddc->fake_change) {
7328: Mat tmat;
7330: PetscCall(VecGetSize(pcis->vec1_global, &global_size));
7331: PetscCall(VecGetLocalSize(pcis->vec1_global, &local_size));
7332: PetscCall(MatDuplicate(pc->pmat, MAT_DO_NOT_COPY_VALUES, &tmat));
7333: PetscCall(MatISSetLocalMat(tmat, localChangeOfBasisMatrix));
7334: PetscCall(MatAssemblyBegin(tmat, MAT_FINAL_ASSEMBLY));
7335: PetscCall(MatAssemblyEnd(tmat, MAT_FINAL_ASSEMBLY));
7336: PetscCall(MatConvert(tmat, MATAIJ, MAT_INITIAL_MATRIX, &pcbddc->ChangeOfBasisMatrix));
7337: PetscCall(MatDestroy(&tmat));
7338: PetscCall(VecSet(pcis->vec1_global, 0.0));
7339: PetscCall(VecSet(pcis->vec1_N, 1.0));
7340: PetscCall(VecScatterBegin(matis->rctx, pcis->vec1_N, pcis->vec1_global, ADD_VALUES, SCATTER_REVERSE));
7341: PetscCall(VecScatterEnd(matis->rctx, pcis->vec1_N, pcis->vec1_global, ADD_VALUES, SCATTER_REVERSE));
7342: PetscCall(VecReciprocal(pcis->vec1_global));
7343: PetscCall(MatDiagonalScale(pcbddc->ChangeOfBasisMatrix, pcis->vec1_global, NULL));
7345: /* check */
7346: if (pcbddc->dbg_flag) {
7347: PetscReal error;
7348: Vec x, x_change;
7350: PetscCall(VecDuplicate(pcis->vec1_global, &x));
7351: PetscCall(VecDuplicate(pcis->vec1_global, &x_change));
7352: PetscCall(VecSetRandom(x, NULL));
7353: PetscCall(VecCopy(x, pcis->vec1_global));
7354: PetscCall(VecScatterBegin(matis->rctx, x, pcis->vec1_N, INSERT_VALUES, SCATTER_FORWARD));
7355: PetscCall(VecScatterEnd(matis->rctx, x, pcis->vec1_N, INSERT_VALUES, SCATTER_FORWARD));
7356: PetscCall(MatMult(localChangeOfBasisMatrix, pcis->vec1_N, pcis->vec2_N));
7357: PetscCall(VecScatterBegin(matis->rctx, pcis->vec2_N, x, INSERT_VALUES, SCATTER_REVERSE));
7358: PetscCall(VecScatterEnd(matis->rctx, pcis->vec2_N, x, INSERT_VALUES, SCATTER_REVERSE));
7359: PetscCall(MatMult(pcbddc->ChangeOfBasisMatrix, pcis->vec1_global, x_change));
7360: PetscCall(VecAXPY(x, -1.0, x_change));
7361: PetscCall(VecNorm(x, NORM_INFINITY, &error));
7362: PetscCheck(error <= PETSC_SMALL, PetscObjectComm((PetscObject)pc), PETSC_ERR_PLIB, "Error global vs local change on N: %1.6e", (double)error);
7363: PetscCall(VecDestroy(&x));
7364: PetscCall(VecDestroy(&x_change));
7365: }
7366: /* adapt sub_schurs computed (if any) */
7367: if (pcbddc->use_deluxe_scaling) {
7368: PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs;
7370: PetscCheck(!pcbddc->use_change_of_basis || !pcbddc->adaptive_userdefined, PetscObjectComm((PetscObject)pc), PETSC_ERR_SUP, "Cannot mix automatic change of basis, adaptive selection and user-defined constraints");
7371: if (sub_schurs && sub_schurs->S_Ej_all) {
7372: Mat S_new, tmat;
7373: IS is_all_N, is_V_Sall = NULL;
7375: PetscCall(ISLocalToGlobalMappingApplyIS(pcis->BtoNmap, sub_schurs->is_Ej_all, &is_all_N));
7376: PetscCall(MatCreateSubMatrix(localChangeOfBasisMatrix, is_all_N, is_all_N, MAT_INITIAL_MATRIX, &tmat));
7377: if (pcbddc->deluxe_zerorows) {
7378: ISLocalToGlobalMapping NtoSall;
7379: IS is_V;
7380: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, pcbddc->n_vertices, pcbddc->local_primal_ref_node, PETSC_COPY_VALUES, &is_V));
7381: PetscCall(ISLocalToGlobalMappingCreateIS(is_all_N, &NtoSall));
7382: PetscCall(ISGlobalToLocalMappingApplyIS(NtoSall, IS_GTOLM_DROP, is_V, &is_V_Sall));
7383: PetscCall(ISLocalToGlobalMappingDestroy(&NtoSall));
7384: PetscCall(ISDestroy(&is_V));
7385: }
7386: PetscCall(ISDestroy(&is_all_N));
7387: PetscCall(MatPtAP(sub_schurs->S_Ej_all, tmat, MAT_INITIAL_MATRIX, 1.0, &S_new));
7388: PetscCall(MatDestroy(&sub_schurs->S_Ej_all));
7389: PetscCall(PetscObjectReference((PetscObject)S_new));
7390: if (pcbddc->deluxe_zerorows) {
7391: const PetscScalar *array;
7392: const PetscInt *idxs_V, *idxs_all;
7393: PetscInt i, n_V;
7395: PetscCall(MatZeroRowsColumnsIS(S_new, is_V_Sall, 1., NULL, NULL));
7396: PetscCall(ISGetLocalSize(is_V_Sall, &n_V));
7397: PetscCall(ISGetIndices(is_V_Sall, &idxs_V));
7398: PetscCall(ISGetIndices(sub_schurs->is_Ej_all, &idxs_all));
7399: PetscCall(VecGetArrayRead(pcis->D, &array));
7400: for (i = 0; i < n_V; i++) {
7401: PetscScalar val;
7402: PetscInt idx;
7404: idx = idxs_V[i];
7405: val = array[idxs_all[idxs_V[i]]];
7406: PetscCall(MatSetValue(S_new, idx, idx, val, INSERT_VALUES));
7407: }
7408: PetscCall(MatAssemblyBegin(S_new, MAT_FINAL_ASSEMBLY));
7409: PetscCall(MatAssemblyEnd(S_new, MAT_FINAL_ASSEMBLY));
7410: PetscCall(VecRestoreArrayRead(pcis->D, &array));
7411: PetscCall(ISRestoreIndices(sub_schurs->is_Ej_all, &idxs_all));
7412: PetscCall(ISRestoreIndices(is_V_Sall, &idxs_V));
7413: }
7414: sub_schurs->S_Ej_all = S_new;
7415: PetscCall(MatDestroy(&S_new));
7416: if (sub_schurs->sum_S_Ej_all) {
7417: PetscCall(MatPtAP(sub_schurs->sum_S_Ej_all, tmat, MAT_INITIAL_MATRIX, 1.0, &S_new));
7418: PetscCall(MatDestroy(&sub_schurs->sum_S_Ej_all));
7419: PetscCall(PetscObjectReference((PetscObject)S_new));
7420: if (pcbddc->deluxe_zerorows) PetscCall(MatZeroRowsColumnsIS(S_new, is_V_Sall, 1., NULL, NULL));
7421: sub_schurs->sum_S_Ej_all = S_new;
7422: PetscCall(MatDestroy(&S_new));
7423: }
7424: PetscCall(ISDestroy(&is_V_Sall));
7425: PetscCall(MatDestroy(&tmat));
7426: }
7427: /* destroy any change of basis context in sub_schurs */
7428: if (sub_schurs && sub_schurs->change) {
7429: PetscInt i;
7431: for (i = 0; i < sub_schurs->n_subs; i++) PetscCall(KSPDestroy(&sub_schurs->change[i]));
7432: PetscCall(PetscFree(sub_schurs->change));
7433: }
7434: }
7435: if (pcbddc->switch_static) { /* need to save the local change */
7436: pcbddc->switch_static_change = localChangeOfBasisMatrix;
7437: } else {
7438: PetscCall(MatDestroy(&localChangeOfBasisMatrix));
7439: }
7440: /* determine if any process has changed the pressures locally */
7441: pcbddc->change_interior = pcbddc->benign_have_null;
7442: } else { /* fake change (get back change of basis into ConstraintMatrix and info on qr) */
7443: PetscCall(MatDestroy(&pcbddc->ConstraintMatrix));
7444: pcbddc->ConstraintMatrix = localChangeOfBasisMatrix;
7445: pcbddc->use_qr_single = qr_needed;
7446: }
7447: } else if (pcbddc->user_ChangeOfBasisMatrix || pcbddc->benign_saddle_point) {
7448: if (!pcbddc->benign_have_null && pcbddc->user_ChangeOfBasisMatrix) {
7449: PetscCall(PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix));
7450: pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix;
7451: } else {
7452: Mat benign_global = NULL;
7453: if (pcbddc->benign_have_null) {
7454: Mat M;
7456: pcbddc->change_interior = PETSC_TRUE;
7457: PetscCall(VecCopy(matis->counter, pcis->vec1_N));
7458: PetscCall(VecReciprocal(pcis->vec1_N));
7459: PetscCall(MatDuplicate(pc->pmat, MAT_DO_NOT_COPY_VALUES, &benign_global));
7460: if (pcbddc->benign_change) {
7461: PetscCall(MatDuplicate(pcbddc->benign_change, MAT_COPY_VALUES, &M));
7462: PetscCall(MatDiagonalScale(M, pcis->vec1_N, NULL));
7463: } else {
7464: PetscCall(MatCreateSeqAIJ(PETSC_COMM_SELF, pcis->n, pcis->n, 1, NULL, &M));
7465: PetscCall(MatDiagonalSet(M, pcis->vec1_N, INSERT_VALUES));
7466: }
7467: PetscCall(MatISSetLocalMat(benign_global, M));
7468: PetscCall(MatDestroy(&M));
7469: PetscCall(MatAssemblyBegin(benign_global, MAT_FINAL_ASSEMBLY));
7470: PetscCall(MatAssemblyEnd(benign_global, MAT_FINAL_ASSEMBLY));
7471: }
7472: if (pcbddc->user_ChangeOfBasisMatrix) {
7473: PetscCall(MatMatMult(pcbddc->user_ChangeOfBasisMatrix, benign_global, MAT_INITIAL_MATRIX, PETSC_DETERMINE, &pcbddc->ChangeOfBasisMatrix));
7474: PetscCall(MatDestroy(&benign_global));
7475: } else if (pcbddc->benign_have_null) {
7476: pcbddc->ChangeOfBasisMatrix = benign_global;
7477: }
7478: }
7479: if (pcbddc->switch_static && pcbddc->ChangeOfBasisMatrix) { /* need to save the local change */
7480: IS is_global;
7481: const PetscInt *gidxs;
7483: PetscCall(ISLocalToGlobalMappingGetIndices(matis->rmapping, &gidxs));
7484: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)pc), pcis->n, gidxs, PETSC_COPY_VALUES, &is_global));
7485: PetscCall(ISLocalToGlobalMappingRestoreIndices(matis->rmapping, &gidxs));
7486: PetscCall(MatCreateSubMatrixUnsorted(pcbddc->ChangeOfBasisMatrix, is_global, is_global, &pcbddc->switch_static_change));
7487: PetscCall(ISDestroy(&is_global));
7488: }
7489: }
7490: if (!pcbddc->fake_change && pcbddc->ChangeOfBasisMatrix && !pcbddc->work_change) PetscCall(VecDuplicate(pcis->vec1_global, &pcbddc->work_change));
7492: if (!pcbddc->fake_change) {
7493: /* add pressure dofs to set of primal nodes for numbering purposes */
7494: for (i = 0; i < pcbddc->benign_n; i++) {
7495: pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i];
7496: pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i];
7497: pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1;
7498: pcbddc->local_primal_size_cc++;
7499: pcbddc->local_primal_size++;
7500: }
7502: /* check if a new primal space has been introduced (also take into account benign trick) */
7503: pcbddc->new_primal_space_local = PETSC_TRUE;
7504: if (olocal_primal_size == pcbddc->local_primal_size) {
7505: PetscCall(PetscArraycmp(pcbddc->local_primal_ref_node, olocal_primal_ref_node, olocal_primal_size_cc, &pcbddc->new_primal_space_local));
7506: pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local);
7507: if (!pcbddc->new_primal_space_local) {
7508: PetscCall(PetscArraycmp(pcbddc->local_primal_ref_mult, olocal_primal_ref_mult, olocal_primal_size_cc, &pcbddc->new_primal_space_local));
7509: pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local);
7510: }
7511: }
7512: /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */
7513: PetscCallMPI(MPIU_Allreduce(&pcbddc->new_primal_space_local, &pcbddc->new_primal_space, 1, MPI_C_BOOL, MPI_LOR, PetscObjectComm((PetscObject)pc)));
7514: }
7515: PetscCall(PetscFree2(olocal_primal_ref_node, olocal_primal_ref_mult));
7517: /* flush dbg viewer */
7518: if (pcbddc->dbg_flag) PetscCall(PetscViewerFlush(pcbddc->dbg_viewer));
7520: /* free workspace */
7521: PetscCall(PetscBTDestroy(&qr_needed_idx));
7522: PetscCall(PetscBTDestroy(&change_basis));
7523: if (!pcbddc->adaptive_selection) {
7524: PetscCall(PetscFree3(constraints_idxs_ptr, constraints_data_ptr, constraints_n));
7525: PetscCall(PetscFree3(constraints_data, constraints_idxs, constraints_idxs_B));
7526: } else {
7527: PetscCall(PetscFree5(pcbddc->adaptive_constraints_n, pcbddc->adaptive_constraints_idxs_ptr, pcbddc->adaptive_constraints_data_ptr, pcbddc->adaptive_constraints_idxs, pcbddc->adaptive_constraints_data));
7528: PetscCall(PetscFree(constraints_n));
7529: PetscCall(PetscFree(constraints_idxs_B));
7530: }
7531: PetscFunctionReturn(PETSC_SUCCESS);
7532: }
7534: PetscErrorCode PCBDDCAnalyzeInterface(PC pc)
7535: {
7536: ISLocalToGlobalMapping map;
7537: PC_BDDC *pcbddc = (PC_BDDC *)pc->data;
7538: Mat_IS *matis = (Mat_IS *)pc->pmat->data;
7539: PetscInt i, N;
7540: PetscBool rcsr = PETSC_FALSE;
7542: PetscFunctionBegin;
7543: if (pcbddc->recompute_topography) {
7544: pcbddc->graphanalyzed = PETSC_FALSE;
7545: /* Reset previously computed graph */
7546: PetscCall(PCBDDCGraphReset(pcbddc->mat_graph));
7547: /* Init local Graph struct */
7548: PetscCall(MatGetSize(pc->pmat, &N, NULL));
7549: PetscCall(MatISGetLocalToGlobalMapping(pc->pmat, &map, NULL));
7550: PetscCall(PCBDDCGraphInit(pcbddc->mat_graph, map, N, pcbddc->graphmaxcount));
7552: if (pcbddc->user_primal_vertices_local && !pcbddc->user_primal_vertices) PetscCall(PCBDDCConsistencyCheckIS(pc, MPI_LOR, &pcbddc->user_primal_vertices_local));
7553: /* Check validity of the csr graph passed in by the user */
7554: PetscCheck(!pcbddc->mat_graph->nvtxs_csr || pcbddc->mat_graph->nvtxs_csr == pcbddc->mat_graph->nvtxs, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Invalid size of local CSR graph! Found %" PetscInt_FMT ", expected %" PetscInt_FMT, pcbddc->mat_graph->nvtxs_csr,
7555: pcbddc->mat_graph->nvtxs);
7557: /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */
7558: if (!pcbddc->mat_graph->xadj && pcbddc->use_local_adj) {
7559: PetscInt *xadj, *adjncy;
7560: PetscInt nvtxs;
7561: PetscBool flg_row;
7562: Mat A;
7564: PetscCall(PetscObjectReference((PetscObject)matis->A));
7565: A = matis->A;
7566: for (PetscInt i = 0; i < pcbddc->local_adj_square; i++) {
7567: Mat AtA;
7569: PetscCall(MatProductCreate(A, A, NULL, &AtA));
7570: PetscCall(MatSetOptionsPrefix(AtA, "pc_bddc_graph_"));
7571: PetscCall(MatProductSetType(AtA, MATPRODUCT_AtB));
7572: PetscCall(MatProductSetFromOptions(AtA));
7573: PetscCall(MatProductSymbolic(AtA));
7574: PetscCall(MatProductClear(AtA));
7575: /* we only need the sparsity, cheat and tell PETSc the matrix has been assembled */
7576: AtA->assembled = PETSC_TRUE;
7577: PetscCall(MatDestroy(&A));
7578: A = AtA;
7579: }
7580: PetscCall(MatGetRowIJ(A, 0, PETSC_TRUE, PETSC_FALSE, &nvtxs, (const PetscInt **)&xadj, (const PetscInt **)&adjncy, &flg_row));
7581: if (flg_row) {
7582: PetscCall(PCBDDCSetLocalAdjacencyGraph(pc, nvtxs, xadj, adjncy, PETSC_COPY_VALUES));
7583: pcbddc->computed_rowadj = PETSC_TRUE;
7584: PetscCall(MatRestoreRowIJ(A, 0, PETSC_TRUE, PETSC_FALSE, &nvtxs, (const PetscInt **)&xadj, (const PetscInt **)&adjncy, &flg_row));
7585: rcsr = PETSC_TRUE;
7586: }
7587: PetscCall(MatDestroy(&A));
7588: }
7589: if (pcbddc->dbg_flag) PetscCall(PetscViewerFlush(pcbddc->dbg_viewer));
7591: if (pcbddc->mat_graph->cdim && !pcbddc->mat_graph->cloc) {
7592: PetscReal *lcoords;
7593: PetscInt n;
7594: MPI_Datatype dimrealtype;
7595: PetscMPIInt cdimi;
7597: /* TODO: support for blocked */
7598: PetscCheck(pcbddc->mat_graph->cnloc == pc->pmat->rmap->n, PETSC_COMM_SELF, PETSC_ERR_USER, "Invalid number of local coordinates! Got %" PetscInt_FMT ", expected %" PetscInt_FMT, pcbddc->mat_graph->cnloc, pc->pmat->rmap->n);
7599: PetscCall(MatGetLocalSize(matis->A, &n, NULL));
7600: PetscCall(PetscMalloc1(pcbddc->mat_graph->cdim * n, &lcoords));
7601: PetscCall(PetscMPIIntCast(pcbddc->mat_graph->cdim, &cdimi));
7602: PetscCallMPI(MPI_Type_contiguous(cdimi, MPIU_REAL, &dimrealtype));
7603: PetscCallMPI(MPI_Type_commit(&dimrealtype));
7604: PetscCall(PetscSFBcastBegin(matis->sf, dimrealtype, pcbddc->mat_graph->coords, lcoords, MPI_REPLACE));
7605: PetscCall(PetscSFBcastEnd(matis->sf, dimrealtype, pcbddc->mat_graph->coords, lcoords, MPI_REPLACE));
7606: PetscCallMPI(MPI_Type_free(&dimrealtype));
7607: PetscCall(PetscFree(pcbddc->mat_graph->coords));
7609: pcbddc->mat_graph->coords = lcoords;
7610: pcbddc->mat_graph->cloc = PETSC_TRUE;
7611: pcbddc->mat_graph->cnloc = n;
7612: }
7613: PetscCheck(!pcbddc->mat_graph->cnloc || pcbddc->mat_graph->cnloc == pcbddc->mat_graph->nvtxs, PETSC_COMM_SELF, PETSC_ERR_USER, "Invalid number of local subdomain coordinates! Got %" PetscInt_FMT ", expected %" PetscInt_FMT, pcbddc->mat_graph->cnloc,
7614: pcbddc->mat_graph->nvtxs);
7615: pcbddc->mat_graph->active_coords = (PetscBool)(pcbddc->corner_selection && pcbddc->mat_graph->cdim && !pcbddc->corner_selected);
7617: /* attach info on disconnected subdomains if present */
7618: if (pcbddc->n_local_subs) {
7619: PetscInt *local_subs, n, totn;
7621: PetscCall(MatGetLocalSize(matis->A, &n, NULL));
7622: PetscCall(PetscMalloc1(n, &local_subs));
7623: for (i = 0; i < n; i++) local_subs[i] = pcbddc->n_local_subs;
7624: for (i = 0; i < pcbddc->n_local_subs; i++) {
7625: const PetscInt *idxs;
7626: PetscInt nl, j;
7628: PetscCall(ISGetLocalSize(pcbddc->local_subs[i], &nl));
7629: PetscCall(ISGetIndices(pcbddc->local_subs[i], &idxs));
7630: for (j = 0; j < nl; j++) local_subs[idxs[j]] = i;
7631: PetscCall(ISRestoreIndices(pcbddc->local_subs[i], &idxs));
7632: }
7633: for (i = 0, totn = 0; i < n; i++) totn = PetscMax(totn, local_subs[i]);
7634: pcbddc->mat_graph->n_local_subs = totn + 1;
7635: pcbddc->mat_graph->local_subs = local_subs;
7636: }
7638: /* Setup of Graph */
7639: PetscCall(PCBDDCGraphSetUp(pcbddc->mat_graph, pcbddc->vertex_size, pcbddc->NeumannBoundariesLocal, pcbddc->DirichletBoundariesLocal, pcbddc->n_ISForDofsLocal, pcbddc->ISForDofsLocal, pcbddc->user_primal_vertices_local));
7640: }
7642: if (!pcbddc->graphanalyzed) {
7643: /* Graph's connected components analysis */
7644: PetscCall(PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph));
7645: pcbddc->graphanalyzed = PETSC_TRUE;
7646: pcbddc->corner_selected = pcbddc->corner_selection;
7647: }
7648: if (rcsr) pcbddc->mat_graph->nvtxs_csr = 0;
7649: PetscFunctionReturn(PETSC_SUCCESS);
7650: }
7652: PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt *nio, Vec vecs[])
7653: {
7654: PetscInt i, j, n;
7655: PetscScalar *alphas;
7656: PetscReal norm, *onorms;
7658: PetscFunctionBegin;
7659: n = *nio;
7660: if (!n) PetscFunctionReturn(PETSC_SUCCESS);
7661: PetscCall(PetscMalloc2(n, &alphas, n, &onorms));
7662: PetscCall(VecNormalize(vecs[0], &norm));
7663: if (norm < PETSC_SMALL) {
7664: onorms[0] = 0.0;
7665: PetscCall(VecSet(vecs[0], 0.0));
7666: } else {
7667: onorms[0] = norm;
7668: }
7670: for (i = 1; i < n; i++) {
7671: PetscCall(VecMDot(vecs[i], i, vecs, alphas));
7672: for (j = 0; j < i; j++) alphas[j] = PetscConj(-alphas[j]);
7673: PetscCall(VecMAXPY(vecs[i], i, alphas, vecs));
7674: PetscCall(VecNormalize(vecs[i], &norm));
7675: if (norm < PETSC_SMALL) {
7676: onorms[i] = 0.0;
7677: PetscCall(VecSet(vecs[i], 0.0));
7678: } else {
7679: onorms[i] = norm;
7680: }
7681: }
7682: /* push nonzero vectors at the beginning */
7683: for (i = 0; i < n; i++) {
7684: if (onorms[i] == 0.0) {
7685: for (j = i + 1; j < n; j++) {
7686: if (onorms[j] != 0.0) {
7687: PetscCall(VecCopy(vecs[j], vecs[i]));
7688: onorms[i] = onorms[j];
7689: onorms[j] = 0.0;
7690: break;
7691: }
7692: }
7693: }
7694: }
7695: for (i = 0, *nio = 0; i < n; i++) *nio += onorms[i] != 0.0 ? 1 : 0;
7696: PetscCall(PetscFree2(alphas, onorms));
7697: PetscFunctionReturn(PETSC_SUCCESS);
7698: }
7700: static PetscErrorCode PCBDDCMatISGetSubassemblingPattern(Mat mat, PetscInt *n_subdomains, PetscInt redprocs, IS *is_sends, PetscBool *have_void)
7701: {
7702: ISLocalToGlobalMapping mapping;
7703: Mat A;
7704: PetscInt n_neighs, *neighs, *n_shared, **shared;
7705: PetscMPIInt size, rank, color;
7706: PetscInt *xadj, *adjncy;
7707: PetscInt *adjncy_wgt, *v_wgt, *ranks_send_to_idx;
7708: PetscInt im_active, active_procs, N, n, i, j, threshold = 2;
7709: PetscInt void_procs, *procs_candidates = NULL;
7710: PetscInt xadj_count, *count;
7711: PetscBool ismatis, use_vwgt = PETSC_FALSE;
7712: PetscSubcomm psubcomm;
7713: MPI_Comm subcomm;
7715: PetscFunctionBegin;
7717: PetscCall(PetscObjectTypeCompare((PetscObject)mat, MATIS, &ismatis));
7718: PetscCheck(ismatis, PetscObjectComm((PetscObject)mat), PETSC_ERR_SUP, "Cannot use %s on a matrix object which is not of type MATIS", PETSC_FUNCTION_NAME);
7721: PetscCheck(*n_subdomains > 0, PetscObjectComm((PetscObject)mat), PETSC_ERR_ARG_WRONG, "Invalid number of subdomains requested %" PetscInt_FMT, *n_subdomains);
7723: if (have_void) *have_void = PETSC_FALSE;
7724: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)mat), &size));
7725: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)mat), &rank));
7726: PetscCall(MatISGetLocalMat(mat, &A));
7727: PetscCall(MatGetLocalSize(A, &n, NULL));
7728: im_active = !!n;
7729: PetscCallMPI(MPIU_Allreduce(&im_active, &active_procs, 1, MPIU_INT, MPI_SUM, PetscObjectComm((PetscObject)mat)));
7730: void_procs = size - active_procs;
7731: /* get ranks of non-active processes in mat communicator */
7732: if (void_procs) {
7733: PetscInt ncand;
7735: if (have_void) *have_void = PETSC_TRUE;
7736: PetscCall(PetscMalloc1(size, &procs_candidates));
7737: PetscCallMPI(MPI_Allgather(&im_active, 1, MPIU_INT, procs_candidates, 1, MPIU_INT, PetscObjectComm((PetscObject)mat)));
7738: for (i = 0, ncand = 0; i < size; i++) {
7739: if (!procs_candidates[i]) procs_candidates[ncand++] = i;
7740: }
7741: /* force n_subdomains to be not greater that the number of non-active processes */
7742: *n_subdomains = PetscMin(void_procs, *n_subdomains);
7743: }
7745: /* number of subdomains requested greater than active processes or matrix size -> just shift the matrix
7746: number of subdomains requested 1 -> send to rank-0 or first candidate in voids */
7747: PetscCall(MatGetSize(mat, &N, NULL));
7748: if (active_procs < *n_subdomains || *n_subdomains == 1 || N <= *n_subdomains) {
7749: PetscInt issize, isidx, dest;
7750: PetscBool default_sub;
7752: if (*n_subdomains == 1) dest = 0;
7753: else dest = rank;
7754: if (im_active) {
7755: issize = 1;
7756: if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */
7757: isidx = procs_candidates[dest];
7758: } else {
7759: isidx = dest;
7760: }
7761: } else {
7762: issize = 0;
7763: isidx = rank;
7764: }
7765: if (*n_subdomains != 1) *n_subdomains = active_procs;
7766: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)mat), issize, &isidx, PETSC_COPY_VALUES, is_sends));
7767: default_sub = (PetscBool)(isidx == rank);
7768: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &default_sub, 1, MPI_C_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
7769: if (default_sub) PetscCall(PetscObjectSetName((PetscObject)*is_sends, "default subassembling"));
7770: PetscCall(PetscFree(procs_candidates));
7771: PetscFunctionReturn(PETSC_SUCCESS);
7772: }
7773: PetscCall(PetscOptionsGetBool(NULL, ((PetscObject)A)->prefix, "-mat_is_partitioning_use_vwgt", &use_vwgt, NULL));
7774: PetscCall(PetscOptionsGetInt(NULL, ((PetscObject)A)->prefix, "-mat_is_partitioning_threshold", &threshold, NULL));
7775: threshold = PetscMax(threshold, 2);
7777: /* Get info on mapping */
7778: PetscCall(MatISGetLocalToGlobalMapping(mat, &mapping, NULL));
7779: PetscCall(ISLocalToGlobalMappingGetInfo(mapping, &n_neighs, &neighs, &n_shared, &shared));
7781: /* build local CSR graph of subdomains' connectivity */
7782: PetscCall(PetscMalloc1(2, &xadj));
7783: xadj[0] = 0;
7784: xadj[1] = PetscMax(n_neighs - 1, 0);
7785: PetscCall(PetscMalloc1(xadj[1], &adjncy));
7786: PetscCall(PetscMalloc1(xadj[1], &adjncy_wgt));
7787: PetscCall(PetscCalloc1(n, &count));
7788: for (i = 1; i < n_neighs; i++)
7789: for (j = 0; j < n_shared[i]; j++) count[shared[i][j]] += 1;
7791: xadj_count = 0;
7792: for (i = 1; i < n_neighs; i++) {
7793: for (j = 0; j < n_shared[i]; j++) {
7794: if (count[shared[i][j]] < threshold) {
7795: adjncy[xadj_count] = neighs[i];
7796: adjncy_wgt[xadj_count] = n_shared[i];
7797: xadj_count++;
7798: break;
7799: }
7800: }
7801: }
7802: xadj[1] = xadj_count;
7803: PetscCall(PetscFree(count));
7804: PetscCall(ISLocalToGlobalMappingRestoreInfo(mapping, &n_neighs, &neighs, &n_shared, &shared));
7805: PetscCall(PetscSortIntWithArray(xadj[1], adjncy, adjncy_wgt));
7807: PetscCall(PetscMalloc1(1, &ranks_send_to_idx));
7809: /* Restrict work on active processes only */
7810: PetscCall(PetscMPIIntCast(im_active, &color));
7811: if (void_procs) {
7812: PetscCall(PetscSubcommCreate(PetscObjectComm((PetscObject)mat), &psubcomm));
7813: PetscCall(PetscSubcommSetNumber(psubcomm, 2)); /* 2 groups, active process and not active processes */
7814: PetscCall(PetscSubcommSetTypeGeneral(psubcomm, color, rank));
7815: subcomm = PetscSubcommChild(psubcomm);
7816: } else {
7817: psubcomm = NULL;
7818: subcomm = PetscObjectComm((PetscObject)mat);
7819: }
7821: v_wgt = NULL;
7822: if (!color) {
7823: PetscCall(PetscFree(xadj));
7824: PetscCall(PetscFree(adjncy));
7825: PetscCall(PetscFree(adjncy_wgt));
7826: } else {
7827: Mat subdomain_adj;
7828: IS new_ranks, new_ranks_contig;
7829: MatPartitioning partitioner;
7830: PetscInt rstart, rend;
7831: PetscMPIInt irstart = 0, irend = 0;
7832: PetscInt *is_indices, *oldranks;
7833: PetscMPIInt size;
7834: PetscBool aggregate;
7836: PetscCallMPI(MPI_Comm_size(subcomm, &size));
7837: if (void_procs) {
7838: PetscInt prank = rank;
7839: PetscCall(PetscMalloc1(size, &oldranks));
7840: PetscCallMPI(MPI_Allgather(&prank, 1, MPIU_INT, oldranks, 1, MPIU_INT, subcomm));
7841: for (i = 0; i < xadj[1]; i++) PetscCall(PetscFindInt(adjncy[i], size, oldranks, &adjncy[i]));
7842: PetscCall(PetscSortIntWithArray(xadj[1], adjncy, adjncy_wgt));
7843: } else {
7844: oldranks = NULL;
7845: }
7846: aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE);
7847: if (aggregate) { /* TODO: all this part could be made more efficient */
7848: PetscInt lrows, row, ncols, *cols;
7849: PetscMPIInt nrank;
7850: PetscScalar *vals;
7852: PetscCallMPI(MPI_Comm_rank(subcomm, &nrank));
7853: lrows = 0;
7854: if (nrank < redprocs) {
7855: lrows = size / redprocs;
7856: if (nrank < size % redprocs) lrows++;
7857: }
7858: PetscCall(MatCreateAIJ(subcomm, lrows, lrows, size, size, 50, NULL, 50, NULL, &subdomain_adj));
7859: PetscCall(MatGetOwnershipRange(subdomain_adj, &rstart, &rend));
7860: PetscCall(PetscMPIIntCast(rstart, &irstart));
7861: PetscCall(PetscMPIIntCast(rend, &irend));
7862: PetscCall(MatSetOption(subdomain_adj, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_FALSE));
7863: PetscCall(MatSetOption(subdomain_adj, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_FALSE));
7864: row = nrank;
7865: ncols = xadj[1] - xadj[0];
7866: cols = adjncy;
7867: PetscCall(PetscMalloc1(ncols, &vals));
7868: for (i = 0; i < ncols; i++) vals[i] = adjncy_wgt[i];
7869: PetscCall(MatSetValues(subdomain_adj, 1, &row, ncols, cols, vals, INSERT_VALUES));
7870: PetscCall(MatAssemblyBegin(subdomain_adj, MAT_FINAL_ASSEMBLY));
7871: PetscCall(MatAssemblyEnd(subdomain_adj, MAT_FINAL_ASSEMBLY));
7872: PetscCall(PetscFree(xadj));
7873: PetscCall(PetscFree(adjncy));
7874: PetscCall(PetscFree(adjncy_wgt));
7875: PetscCall(PetscFree(vals));
7876: if (use_vwgt) {
7877: Vec v;
7878: const PetscScalar *array;
7879: PetscInt nl;
7881: PetscCall(MatCreateVecs(subdomain_adj, &v, NULL));
7882: PetscCall(VecSetValue(v, row, (PetscScalar)n, INSERT_VALUES));
7883: PetscCall(VecAssemblyBegin(v));
7884: PetscCall(VecAssemblyEnd(v));
7885: PetscCall(VecGetLocalSize(v, &nl));
7886: PetscCall(VecGetArrayRead(v, &array));
7887: PetscCall(PetscMalloc1(nl, &v_wgt));
7888: for (i = 0; i < nl; i++) v_wgt[i] = (PetscInt)PetscRealPart(array[i]);
7889: PetscCall(VecRestoreArrayRead(v, &array));
7890: PetscCall(VecDestroy(&v));
7891: }
7892: } else {
7893: PetscCall(MatCreateMPIAdj(subcomm, 1, size, xadj, adjncy, adjncy_wgt, &subdomain_adj));
7894: if (use_vwgt) {
7895: PetscCall(PetscMalloc1(1, &v_wgt));
7896: v_wgt[0] = n;
7897: }
7898: }
7899: /* PetscCall(MatView(subdomain_adj,0)); */
7901: /* Partition */
7902: PetscCall(MatPartitioningCreate(subcomm, &partitioner));
7903: #if defined(PETSC_HAVE_PTSCOTCH)
7904: PetscCall(MatPartitioningSetType(partitioner, MATPARTITIONINGPTSCOTCH));
7905: #elif defined(PETSC_HAVE_PARMETIS)
7906: PetscCall(MatPartitioningSetType(partitioner, MATPARTITIONINGPARMETIS));
7907: #else
7908: PetscCall(MatPartitioningSetType(partitioner, MATPARTITIONINGAVERAGE));
7909: #endif
7910: PetscCall(MatPartitioningSetAdjacency(partitioner, subdomain_adj));
7911: if (v_wgt) PetscCall(MatPartitioningSetVertexWeights(partitioner, v_wgt));
7912: *n_subdomains = PetscMin(size, *n_subdomains);
7913: PetscCall(MatPartitioningSetNParts(partitioner, *n_subdomains));
7914: PetscCall(MatPartitioningSetFromOptions(partitioner));
7915: PetscCall(MatPartitioningApply(partitioner, &new_ranks));
7916: /* PetscCall(MatPartitioningView(partitioner,0)); */
7918: /* renumber new_ranks to avoid "holes" in new set of processors */
7919: PetscCall(ISRenumber(new_ranks, NULL, NULL, &new_ranks_contig));
7920: PetscCall(ISDestroy(&new_ranks));
7921: PetscCall(ISGetIndices(new_ranks_contig, (const PetscInt **)&is_indices));
7922: if (!aggregate) {
7923: if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */
7924: PetscAssert(oldranks, PETSC_COMM_SELF, PETSC_ERR_PLIB, "This should not happen");
7925: ranks_send_to_idx[0] = procs_candidates[oldranks[is_indices[0]]];
7926: } else if (oldranks) {
7927: ranks_send_to_idx[0] = oldranks[is_indices[0]];
7928: } else {
7929: ranks_send_to_idx[0] = is_indices[0];
7930: }
7931: } else {
7932: PetscInt idx = 0;
7933: PetscMPIInt tag;
7934: MPI_Request *reqs;
7936: PetscCall(PetscObjectGetNewTag((PetscObject)subdomain_adj, &tag));
7937: PetscCall(PetscMalloc1(rend - rstart, &reqs));
7938: for (PetscMPIInt i = irstart; i < irend; i++) PetscCallMPI(MPIU_Isend(is_indices + i - rstart, 1, MPIU_INT, i, tag, subcomm, &reqs[i - rstart]));
7939: PetscCallMPI(MPIU_Recv(&idx, 1, MPIU_INT, MPI_ANY_SOURCE, tag, subcomm, MPI_STATUS_IGNORE));
7940: PetscCallMPI(MPI_Waitall(irend - irstart, reqs, MPI_STATUSES_IGNORE));
7941: PetscCall(PetscFree(reqs));
7942: if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */
7943: PetscAssert(oldranks, PETSC_COMM_SELF, PETSC_ERR_PLIB, "This should not happen");
7944: ranks_send_to_idx[0] = procs_candidates[oldranks[idx]];
7945: } else if (oldranks) {
7946: ranks_send_to_idx[0] = oldranks[idx];
7947: } else {
7948: ranks_send_to_idx[0] = idx;
7949: }
7950: }
7951: PetscCall(ISRestoreIndices(new_ranks_contig, (const PetscInt **)&is_indices));
7952: /* clean up */
7953: PetscCall(PetscFree(oldranks));
7954: PetscCall(ISDestroy(&new_ranks_contig));
7955: PetscCall(MatDestroy(&subdomain_adj));
7956: PetscCall(MatPartitioningDestroy(&partitioner));
7957: }
7958: PetscCall(PetscSubcommDestroy(&psubcomm));
7959: PetscCall(PetscFree(procs_candidates));
7961: /* assemble parallel IS for sends */
7962: i = 1;
7963: if (!color) i = 0;
7964: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)mat), i, ranks_send_to_idx, PETSC_OWN_POINTER, is_sends));
7965: PetscFunctionReturn(PETSC_SUCCESS);
7966: }
7968: typedef enum {
7969: MATDENSE_PRIVATE = 0,
7970: MATAIJ_PRIVATE,
7971: MATBAIJ_PRIVATE,
7972: MATSBAIJ_PRIVATE
7973: } MatTypePrivate;
7975: static PetscErrorCode PCBDDCMatISSubassemble(Mat mat, IS is_sends, PetscInt n_subdomains, PetscBool restrict_comm, PetscBool restrict_full, PetscBool reuse, Mat *mat_n, PetscInt nis, IS isarray[], PetscInt nvecs, Vec nnsp_vec[])
7976: {
7977: Mat local_mat;
7978: IS is_sends_internal;
7979: PetscInt rows, cols, new_local_rows;
7980: PetscInt i, bs, buf_size_idxs, buf_size_idxs_is, buf_size_vals, buf_size_vecs;
7981: PetscBool ismatis, isdense, newisdense, destroy_mat;
7982: ISLocalToGlobalMapping l2gmap;
7983: PetscInt *l2gmap_indices;
7984: const PetscInt *is_indices;
7985: MatType new_local_type;
7986: /* buffers */
7987: PetscInt *ptr_idxs, *send_buffer_idxs, *recv_buffer_idxs;
7988: PetscInt *ptr_idxs_is, *send_buffer_idxs_is, *recv_buffer_idxs_is;
7989: PetscInt *recv_buffer_idxs_local;
7990: PetscScalar *ptr_vals, *recv_buffer_vals;
7991: const PetscScalar *send_buffer_vals;
7992: PetscScalar *ptr_vecs, *send_buffer_vecs, *recv_buffer_vecs;
7993: /* MPI */
7994: MPI_Comm comm, comm_n;
7995: PetscSubcomm subcomm;
7996: PetscMPIInt n_sends, n_recvs, size;
7997: PetscMPIInt *iflags, *ilengths_idxs, *ilengths_vals, *ilengths_idxs_is;
7998: PetscMPIInt *onodes, *onodes_is, *olengths_idxs, *olengths_idxs_is, *olengths_vals;
7999: PetscMPIInt len, tag_idxs, tag_idxs_is, tag_vals, tag_vecs, source_dest;
8000: MPI_Request *send_req_idxs, *send_req_idxs_is, *send_req_vals, *send_req_vecs;
8001: MPI_Request *recv_req_idxs, *recv_req_idxs_is, *recv_req_vals, *recv_req_vecs;
8003: PetscFunctionBegin;
8005: PetscCall(PetscObjectTypeCompare((PetscObject)mat, MATIS, &ismatis));
8006: PetscCheck(ismatis, PetscObjectComm((PetscObject)mat), PETSC_ERR_SUP, "Cannot use %s on a matrix object which is not of type MATIS", PETSC_FUNCTION_NAME);
8013: if (nvecs) {
8014: PetscCheck(nvecs <= 1, PetscObjectComm((PetscObject)mat), PETSC_ERR_SUP, "Just 1 vector supported");
8016: }
8017: /* further checks */
8018: PetscCall(MatISGetLocalMat(mat, &local_mat));
8019: PetscCall(PetscObjectTypeCompare((PetscObject)local_mat, MATSEQDENSE, &isdense));
8020: PetscCheck(isdense, PetscObjectComm((PetscObject)mat), PETSC_ERR_SUP, "Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE");
8022: PetscCall(MatGetSize(local_mat, &rows, &cols));
8023: PetscCheck(rows == cols, PetscObjectComm((PetscObject)mat), PETSC_ERR_SUP, "Local MATIS matrices should be square");
8024: if (reuse && *mat_n) {
8025: PetscInt mrows, mcols, mnrows, mncols;
8027: PetscCall(PetscObjectTypeCompare((PetscObject)*mat_n, MATIS, &ismatis));
8028: PetscCheck(ismatis, PetscObjectComm((PetscObject)*mat_n), PETSC_ERR_SUP, "Cannot reuse a matrix which is not of type MATIS");
8029: PetscCall(MatGetSize(mat, &mrows, &mcols));
8030: PetscCall(MatGetSize(*mat_n, &mnrows, &mncols));
8031: PetscCheck(mrows == mnrows, PetscObjectComm((PetscObject)mat), PETSC_ERR_SUP, "Cannot reuse matrix! Wrong number of rows %" PetscInt_FMT " != %" PetscInt_FMT, mrows, mnrows);
8032: PetscCheck(mcols == mncols, PetscObjectComm((PetscObject)mat), PETSC_ERR_SUP, "Cannot reuse matrix! Wrong number of cols %" PetscInt_FMT " != %" PetscInt_FMT, mcols, mncols);
8033: }
8034: PetscCall(MatGetBlockSize(local_mat, &bs));
8037: /* prepare IS for sending if not provided */
8038: if (!is_sends) {
8039: PetscCheck(n_subdomains, PetscObjectComm((PetscObject)mat), PETSC_ERR_SUP, "You should specify either an IS or a target number of subdomains");
8040: PetscCall(PCBDDCMatISGetSubassemblingPattern(mat, &n_subdomains, 0, &is_sends_internal, NULL));
8041: } else {
8042: PetscCall(PetscObjectReference((PetscObject)is_sends));
8043: is_sends_internal = is_sends;
8044: }
8046: /* get comm */
8047: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
8049: /* compute number of sends */
8050: PetscCall(ISGetLocalSize(is_sends_internal, &i));
8051: PetscCall(PetscMPIIntCast(i, &n_sends));
8053: /* compute number of receives */
8054: PetscCallMPI(MPI_Comm_size(comm, &size));
8055: PetscCall(PetscMalloc1(size, &iflags));
8056: PetscCall(PetscArrayzero(iflags, size));
8057: PetscCall(ISGetIndices(is_sends_internal, &is_indices));
8058: for (i = 0; i < n_sends; i++) iflags[is_indices[i]] = 1;
8059: PetscCall(PetscGatherNumberOfMessages(comm, iflags, NULL, &n_recvs));
8060: PetscCall(PetscFree(iflags));
8062: /* restrict comm if requested */
8063: subcomm = NULL;
8064: destroy_mat = PETSC_FALSE;
8065: if (restrict_comm) {
8066: PetscMPIInt color, subcommsize;
8068: color = 0;
8069: if (restrict_full) {
8070: if (!n_recvs) color = 1; /* processes not receiving anything will not participate in new comm (full restriction) */
8071: } else {
8072: if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not participate in new comm */
8073: }
8074: PetscCallMPI(MPIU_Allreduce(&color, &subcommsize, 1, MPI_INT, MPI_SUM, comm));
8075: subcommsize = size - subcommsize;
8076: /* check if reuse has been requested */
8077: if (reuse) {
8078: if (*mat_n) {
8079: PetscMPIInt subcommsize2;
8080: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n), &subcommsize2));
8081: PetscCheck(subcommsize == subcommsize2, PetscObjectComm((PetscObject)*mat_n), PETSC_ERR_PLIB, "Cannot reuse matrix! wrong subcomm size %d != %d", subcommsize, subcommsize2);
8082: comm_n = PetscObjectComm((PetscObject)*mat_n);
8083: } else {
8084: comm_n = PETSC_COMM_SELF;
8085: }
8086: } else { /* MAT_INITIAL_MATRIX */
8087: PetscMPIInt rank;
8089: PetscCallMPI(MPI_Comm_rank(comm, &rank));
8090: PetscCall(PetscSubcommCreate(comm, &subcomm));
8091: PetscCall(PetscSubcommSetNumber(subcomm, 2));
8092: PetscCall(PetscSubcommSetTypeGeneral(subcomm, color, rank));
8093: comm_n = PetscSubcommChild(subcomm);
8094: }
8095: /* flag to destroy *mat_n if not significative */
8096: if (color) destroy_mat = PETSC_TRUE;
8097: } else {
8098: comm_n = comm;
8099: }
8101: /* prepare send/receive buffers */
8102: PetscCall(PetscMalloc1(size, &ilengths_idxs));
8103: PetscCall(PetscArrayzero(ilengths_idxs, size));
8104: PetscCall(PetscMalloc1(size, &ilengths_vals));
8105: PetscCall(PetscArrayzero(ilengths_vals, size));
8106: if (nis) PetscCall(PetscCalloc1(size, &ilengths_idxs_is));
8108: /* Get data from local matrices */
8109: PetscCheck(isdense, PetscObjectComm((PetscObject)mat), PETSC_ERR_SUP, "Subassembling of AIJ local matrices not yet implemented");
8110: /* TODO: See below some guidelines on how to prepare the local buffers */
8111: /*
8112: send_buffer_vals should contain the raw values of the local matrix
8113: send_buffer_idxs should contain:
8114: - MatType_PRIVATE type
8115: - PetscInt size_of_l2gmap
8116: - PetscInt global_row_indices[size_of_l2gmap]
8117: - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values
8118: */
8119: {
8120: ISLocalToGlobalMapping mapping;
8122: PetscCall(MatISGetLocalToGlobalMapping(mat, &mapping, NULL));
8123: PetscCall(MatDenseGetArrayRead(local_mat, &send_buffer_vals));
8124: PetscCall(ISLocalToGlobalMappingGetSize(mapping, &i));
8125: PetscCall(PetscMalloc1(i + 2, &send_buffer_idxs));
8126: send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE;
8127: send_buffer_idxs[1] = i;
8128: PetscCall(ISLocalToGlobalMappingGetIndices(mapping, (const PetscInt **)&ptr_idxs));
8129: PetscCall(PetscArraycpy(&send_buffer_idxs[2], ptr_idxs, i));
8130: PetscCall(ISLocalToGlobalMappingRestoreIndices(mapping, (const PetscInt **)&ptr_idxs));
8131: PetscCall(PetscMPIIntCast(i, &len));
8132: for (i = 0; i < n_sends; i++) {
8133: ilengths_vals[is_indices[i]] = len * len;
8134: ilengths_idxs[is_indices[i]] = len + 2;
8135: }
8136: }
8137: PetscCall(PetscGatherMessageLengths2(comm, n_sends, n_recvs, ilengths_idxs, ilengths_vals, &onodes, &olengths_idxs, &olengths_vals));
8138: /* additional is (if any) */
8139: if (nis) {
8140: PetscMPIInt psum;
8141: PetscInt j;
8142: for (j = 0, psum = 0; j < nis; j++) {
8143: PetscInt plen;
8144: PetscCall(ISGetLocalSize(isarray[j], &plen));
8145: PetscCall(PetscMPIIntCast(plen, &len));
8146: psum += len + 1; /* indices + length */
8147: }
8148: PetscCall(PetscMalloc1(psum, &send_buffer_idxs_is));
8149: for (j = 0, psum = 0; j < nis; j++) {
8150: PetscInt plen;
8151: const PetscInt *is_array_idxs;
8152: PetscCall(ISGetLocalSize(isarray[j], &plen));
8153: send_buffer_idxs_is[psum] = plen;
8154: PetscCall(ISGetIndices(isarray[j], &is_array_idxs));
8155: PetscCall(PetscArraycpy(&send_buffer_idxs_is[psum + 1], is_array_idxs, plen));
8156: PetscCall(ISRestoreIndices(isarray[j], &is_array_idxs));
8157: psum += plen + 1; /* indices + length */
8158: }
8159: for (i = 0; i < n_sends; i++) ilengths_idxs_is[is_indices[i]] = psum;
8160: PetscCall(PetscGatherMessageLengths(comm, n_sends, n_recvs, ilengths_idxs_is, &onodes_is, &olengths_idxs_is));
8161: }
8162: PetscCall(MatISRestoreLocalMat(mat, &local_mat));
8164: buf_size_idxs = 0;
8165: buf_size_vals = 0;
8166: buf_size_idxs_is = 0;
8167: buf_size_vecs = 0;
8168: for (i = 0; i < n_recvs; i++) {
8169: buf_size_idxs += olengths_idxs[i];
8170: buf_size_vals += olengths_vals[i];
8171: if (nis) buf_size_idxs_is += olengths_idxs_is[i];
8172: if (nvecs) buf_size_vecs += olengths_idxs[i];
8173: }
8174: PetscCall(PetscMalloc1(buf_size_idxs, &recv_buffer_idxs));
8175: PetscCall(PetscMalloc1(buf_size_vals, &recv_buffer_vals));
8176: PetscCall(PetscMalloc1(buf_size_idxs_is, &recv_buffer_idxs_is));
8177: PetscCall(PetscMalloc1(buf_size_vecs, &recv_buffer_vecs));
8179: /* get new tags for clean communications */
8180: PetscCall(PetscObjectGetNewTag((PetscObject)mat, &tag_idxs));
8181: PetscCall(PetscObjectGetNewTag((PetscObject)mat, &tag_vals));
8182: PetscCall(PetscObjectGetNewTag((PetscObject)mat, &tag_idxs_is));
8183: PetscCall(PetscObjectGetNewTag((PetscObject)mat, &tag_vecs));
8185: /* allocate for requests */
8186: PetscCall(PetscMalloc1(n_sends, &send_req_idxs));
8187: PetscCall(PetscMalloc1(n_sends, &send_req_vals));
8188: PetscCall(PetscMalloc1(n_sends, &send_req_idxs_is));
8189: PetscCall(PetscMalloc1(n_sends, &send_req_vecs));
8190: PetscCall(PetscMalloc1(n_recvs, &recv_req_idxs));
8191: PetscCall(PetscMalloc1(n_recvs, &recv_req_vals));
8192: PetscCall(PetscMalloc1(n_recvs, &recv_req_idxs_is));
8193: PetscCall(PetscMalloc1(n_recvs, &recv_req_vecs));
8195: /* communications */
8196: ptr_idxs = recv_buffer_idxs;
8197: ptr_vals = recv_buffer_vals;
8198: ptr_idxs_is = recv_buffer_idxs_is;
8199: ptr_vecs = recv_buffer_vecs;
8200: for (i = 0; i < n_recvs; i++) {
8201: PetscCallMPI(MPIU_Irecv(ptr_idxs, olengths_idxs[i], MPIU_INT, onodes[i], tag_idxs, comm, &recv_req_idxs[i]));
8202: PetscCallMPI(MPIU_Irecv(ptr_vals, olengths_vals[i], MPIU_SCALAR, onodes[i], tag_vals, comm, &recv_req_vals[i]));
8203: ptr_idxs += olengths_idxs[i];
8204: ptr_vals += olengths_vals[i];
8205: if (nis) {
8206: PetscCallMPI(MPIU_Irecv(ptr_idxs_is, olengths_idxs_is[i], MPIU_INT, onodes_is[i], tag_idxs_is, comm, &recv_req_idxs_is[i]));
8207: ptr_idxs_is += olengths_idxs_is[i];
8208: }
8209: if (nvecs) {
8210: PetscCallMPI(MPIU_Irecv(ptr_vecs, olengths_idxs[i] - 2, MPIU_SCALAR, onodes[i], tag_vecs, comm, &recv_req_vecs[i]));
8211: ptr_vecs += olengths_idxs[i] - 2;
8212: }
8213: }
8214: for (i = 0; i < n_sends; i++) {
8215: PetscCall(PetscMPIIntCast(is_indices[i], &source_dest));
8216: PetscCallMPI(MPIU_Isend(send_buffer_idxs, ilengths_idxs[source_dest], MPIU_INT, source_dest, tag_idxs, comm, &send_req_idxs[i]));
8217: PetscCallMPI(MPIU_Isend(send_buffer_vals, ilengths_vals[source_dest], MPIU_SCALAR, source_dest, tag_vals, comm, &send_req_vals[i]));
8218: if (nis) PetscCallMPI(MPIU_Isend(send_buffer_idxs_is, ilengths_idxs_is[source_dest], MPIU_INT, source_dest, tag_idxs_is, comm, &send_req_idxs_is[i]));
8219: if (nvecs) {
8220: PetscCall(VecGetArray(nnsp_vec[0], &send_buffer_vecs));
8221: PetscCallMPI(MPIU_Isend(send_buffer_vecs, ilengths_idxs[source_dest] - 2, MPIU_SCALAR, source_dest, tag_vecs, comm, &send_req_vecs[i]));
8222: }
8223: }
8224: PetscCall(ISRestoreIndices(is_sends_internal, &is_indices));
8225: PetscCall(ISDestroy(&is_sends_internal));
8227: /* assemble new l2g map */
8228: PetscCallMPI(MPI_Waitall(n_recvs, recv_req_idxs, MPI_STATUSES_IGNORE));
8229: ptr_idxs = recv_buffer_idxs;
8230: new_local_rows = 0;
8231: for (i = 0; i < n_recvs; i++) {
8232: new_local_rows += *(ptr_idxs + 1); /* second element is the local size of the l2gmap */
8233: ptr_idxs += olengths_idxs[i];
8234: }
8235: PetscCall(PetscMalloc1(new_local_rows, &l2gmap_indices));
8236: ptr_idxs = recv_buffer_idxs;
8237: new_local_rows = 0;
8238: for (i = 0; i < n_recvs; i++) {
8239: PetscCall(PetscArraycpy(&l2gmap_indices[new_local_rows], ptr_idxs + 2, *(ptr_idxs + 1)));
8240: new_local_rows += *(ptr_idxs + 1); /* second element is the local size of the l2gmap */
8241: ptr_idxs += olengths_idxs[i];
8242: }
8243: PetscCall(PetscSortRemoveDupsInt(&new_local_rows, l2gmap_indices));
8244: PetscCall(ISLocalToGlobalMappingCreate(comm_n, 1, new_local_rows, l2gmap_indices, PETSC_COPY_VALUES, &l2gmap));
8245: PetscCall(PetscFree(l2gmap_indices));
8247: /* infer new local matrix type from received local matrices type */
8248: /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */
8249: /* it also assumes that if the block size is set, than it is the same among all local matrices (see checks at the beginning of the function) */
8250: if (n_recvs) {
8251: MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0];
8252: ptr_idxs = recv_buffer_idxs;
8253: for (i = 0; i < n_recvs; i++) {
8254: if ((PetscInt)new_local_type_private != *ptr_idxs) {
8255: new_local_type_private = MATAIJ_PRIVATE;
8256: break;
8257: }
8258: ptr_idxs += olengths_idxs[i];
8259: }
8260: switch (new_local_type_private) {
8261: case MATDENSE_PRIVATE:
8262: new_local_type = MATSEQAIJ;
8263: bs = 1;
8264: break;
8265: case MATAIJ_PRIVATE:
8266: new_local_type = MATSEQAIJ;
8267: bs = 1;
8268: break;
8269: case MATBAIJ_PRIVATE:
8270: new_local_type = MATSEQBAIJ;
8271: break;
8272: case MATSBAIJ_PRIVATE:
8273: new_local_type = MATSEQSBAIJ;
8274: break;
8275: default:
8276: SETERRQ(comm, PETSC_ERR_SUP, "Unsupported private type %d in %s", new_local_type_private, PETSC_FUNCTION_NAME);
8277: }
8278: } else { /* by default, new_local_type is seqaij */
8279: new_local_type = MATSEQAIJ;
8280: bs = 1;
8281: }
8283: /* create MATIS object if needed */
8284: if (!reuse) {
8285: PetscCall(MatGetSize(mat, &rows, &cols));
8286: PetscCall(MatCreateIS(comm_n, bs, PETSC_DECIDE, PETSC_DECIDE, rows, cols, l2gmap, l2gmap, mat_n));
8287: } else {
8288: /* it also destroys the local matrices */
8289: if (*mat_n) PetscCall(MatSetLocalToGlobalMapping(*mat_n, l2gmap, l2gmap));
8290: else PetscCall(MatCreateIS(comm_n, bs, PETSC_DECIDE, PETSC_DECIDE, rows, cols, l2gmap, l2gmap, mat_n)); /* this is a fake object */
8291: }
8292: PetscCall(MatISGetLocalMat(*mat_n, &local_mat));
8293: PetscCall(MatSetType(local_mat, new_local_type));
8295: PetscCallMPI(MPI_Waitall(n_recvs, recv_req_vals, MPI_STATUSES_IGNORE));
8297: /* Global to local map of received indices */
8298: PetscCall(PetscMalloc1(buf_size_idxs, &recv_buffer_idxs_local)); /* needed for values insertion */
8299: PetscCall(ISGlobalToLocalMappingApply(l2gmap, IS_GTOLM_MASK, buf_size_idxs, recv_buffer_idxs, &i, recv_buffer_idxs_local));
8300: PetscCall(ISLocalToGlobalMappingDestroy(&l2gmap));
8302: /* restore attributes -> type of incoming data and its size */
8303: buf_size_idxs = 0;
8304: for (i = 0; i < n_recvs; i++) {
8305: recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs];
8306: recv_buffer_idxs_local[buf_size_idxs + 1] = recv_buffer_idxs[buf_size_idxs + 1];
8307: buf_size_idxs += olengths_idxs[i];
8308: }
8309: PetscCall(PetscFree(recv_buffer_idxs));
8311: /* set preallocation */
8312: PetscCall(PetscObjectTypeCompare((PetscObject)local_mat, MATSEQDENSE, &newisdense));
8313: if (!newisdense) {
8314: PetscInt *new_local_nnz = NULL;
8316: ptr_idxs = recv_buffer_idxs_local;
8317: if (n_recvs) PetscCall(PetscCalloc1(new_local_rows, &new_local_nnz));
8318: for (i = 0; i < n_recvs; i++) {
8319: PetscInt j;
8320: if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */
8321: for (j = 0; j < *(ptr_idxs + 1); j++) new_local_nnz[*(ptr_idxs + 2 + j)] += *(ptr_idxs + 1);
8322: } else {
8323: /* TODO */
8324: }
8325: ptr_idxs += olengths_idxs[i];
8326: }
8327: if (new_local_nnz) {
8328: for (i = 0; i < new_local_rows; i++) new_local_nnz[i] = PetscMin(new_local_nnz[i], new_local_rows);
8329: PetscCall(MatSeqAIJSetPreallocation(local_mat, 0, new_local_nnz));
8330: for (i = 0; i < new_local_rows; i++) new_local_nnz[i] /= bs;
8331: PetscCall(MatSeqBAIJSetPreallocation(local_mat, bs, 0, new_local_nnz));
8332: for (i = 0; i < new_local_rows; i++) new_local_nnz[i] = PetscMax(new_local_nnz[i] - i, 0);
8333: PetscCall(MatSeqSBAIJSetPreallocation(local_mat, bs, 0, new_local_nnz));
8334: } else {
8335: PetscCall(MatSetUp(local_mat));
8336: }
8337: PetscCall(PetscFree(new_local_nnz));
8338: } else {
8339: PetscCall(MatSetUp(local_mat));
8340: }
8342: /* set values */
8343: ptr_vals = recv_buffer_vals;
8344: ptr_idxs = recv_buffer_idxs_local;
8345: for (i = 0; i < n_recvs; i++) {
8346: if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */
8347: PetscCall(MatSetOption(local_mat, MAT_ROW_ORIENTED, PETSC_FALSE));
8348: PetscCall(MatSetValues(local_mat, *(ptr_idxs + 1), ptr_idxs + 2, *(ptr_idxs + 1), ptr_idxs + 2, ptr_vals, ADD_VALUES));
8349: PetscCall(MatAssemblyBegin(local_mat, MAT_FLUSH_ASSEMBLY));
8350: PetscCall(MatAssemblyEnd(local_mat, MAT_FLUSH_ASSEMBLY));
8351: PetscCall(MatSetOption(local_mat, MAT_ROW_ORIENTED, PETSC_TRUE));
8352: } else {
8353: /* TODO */
8354: }
8355: ptr_idxs += olengths_idxs[i];
8356: ptr_vals += olengths_vals[i];
8357: }
8358: PetscCall(MatAssemblyBegin(local_mat, MAT_FINAL_ASSEMBLY));
8359: PetscCall(MatAssemblyEnd(local_mat, MAT_FINAL_ASSEMBLY));
8360: PetscCall(MatISRestoreLocalMat(*mat_n, &local_mat));
8361: PetscCall(MatAssemblyBegin(*mat_n, MAT_FINAL_ASSEMBLY));
8362: PetscCall(MatAssemblyEnd(*mat_n, MAT_FINAL_ASSEMBLY));
8363: PetscCall(PetscFree(recv_buffer_vals));
8365: #if 0
8366: if (!restrict_comm) { /* check */
8367: Vec lvec,rvec;
8368: PetscReal infty_error;
8370: PetscCall(MatCreateVecs(mat,&rvec,&lvec));
8371: PetscCall(VecSetRandom(rvec,NULL));
8372: PetscCall(MatMult(mat,rvec,lvec));
8373: PetscCall(VecScale(lvec,-1.0));
8374: PetscCall(MatMultAdd(*mat_n,rvec,lvec,lvec));
8375: PetscCall(VecNorm(lvec,NORM_INFINITY,&infty_error));
8376: PetscCall(PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error));
8377: PetscCall(VecDestroy(&rvec));
8378: PetscCall(VecDestroy(&lvec));
8379: }
8380: #endif
8382: /* assemble new additional is (if any) */
8383: if (nis) {
8384: PetscInt **temp_idxs, *count_is, j, psum;
8386: PetscCallMPI(MPI_Waitall(n_recvs, recv_req_idxs_is, MPI_STATUSES_IGNORE));
8387: PetscCall(PetscCalloc1(nis, &count_is));
8388: ptr_idxs = recv_buffer_idxs_is;
8389: psum = 0;
8390: for (i = 0; i < n_recvs; i++) {
8391: for (j = 0; j < nis; j++) {
8392: PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */
8393: count_is[j] += plen; /* increment counting of buffer for j-th IS */
8394: psum += plen;
8395: ptr_idxs += plen + 1; /* shift pointer to received data */
8396: }
8397: }
8398: PetscCall(PetscMalloc1(nis, &temp_idxs));
8399: PetscCall(PetscMalloc1(psum, &temp_idxs[0]));
8400: for (i = 1; i < nis; i++) temp_idxs[i] = PetscSafePointerPlusOffset(temp_idxs[i - 1], count_is[i - 1]);
8401: PetscCall(PetscArrayzero(count_is, nis));
8402: ptr_idxs = recv_buffer_idxs_is;
8403: for (i = 0; i < n_recvs; i++) {
8404: for (j = 0; j < nis; j++) {
8405: PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */
8406: PetscCall(PetscArraycpy(&temp_idxs[j][count_is[j]], ptr_idxs + 1, plen));
8407: count_is[j] += plen; /* increment starting point of buffer for j-th IS */
8408: ptr_idxs += plen + 1; /* shift pointer to received data */
8409: }
8410: }
8411: for (i = 0; i < nis; i++) {
8412: PetscCall(ISDestroy(&isarray[i]));
8413: PetscCall(PetscSortRemoveDupsInt(&count_is[i], temp_idxs[i]));
8414: PetscCall(ISCreateGeneral(comm_n, count_is[i], temp_idxs[i], PETSC_COPY_VALUES, &isarray[i]));
8415: }
8416: PetscCall(PetscFree(count_is));
8417: PetscCall(PetscFree(temp_idxs[0]));
8418: PetscCall(PetscFree(temp_idxs));
8419: }
8420: /* free workspace */
8421: PetscCall(PetscFree(recv_buffer_idxs_is));
8422: PetscCallMPI(MPI_Waitall(n_sends, send_req_idxs, MPI_STATUSES_IGNORE));
8423: PetscCall(PetscFree(send_buffer_idxs));
8424: PetscCallMPI(MPI_Waitall(n_sends, send_req_vals, MPI_STATUSES_IGNORE));
8425: if (isdense) {
8426: PetscCall(MatISGetLocalMat(mat, &local_mat));
8427: PetscCall(MatDenseRestoreArrayRead(local_mat, &send_buffer_vals));
8428: PetscCall(MatISRestoreLocalMat(mat, &local_mat));
8429: } else {
8430: /* PetscCall(PetscFree(send_buffer_vals)); */
8431: }
8432: if (nis) {
8433: PetscCallMPI(MPI_Waitall(n_sends, send_req_idxs_is, MPI_STATUSES_IGNORE));
8434: PetscCall(PetscFree(send_buffer_idxs_is));
8435: }
8437: if (nvecs) {
8438: PetscCallMPI(MPI_Waitall(n_recvs, recv_req_vecs, MPI_STATUSES_IGNORE));
8439: PetscCallMPI(MPI_Waitall(n_sends, send_req_vecs, MPI_STATUSES_IGNORE));
8440: PetscCall(VecRestoreArray(nnsp_vec[0], &send_buffer_vecs));
8441: PetscCall(VecDestroy(&nnsp_vec[0]));
8442: PetscCall(VecCreate(comm_n, &nnsp_vec[0]));
8443: PetscCall(VecSetSizes(nnsp_vec[0], new_local_rows, PETSC_DECIDE));
8444: PetscCall(VecSetType(nnsp_vec[0], VECSTANDARD));
8445: /* set values */
8446: ptr_vals = recv_buffer_vecs;
8447: ptr_idxs = recv_buffer_idxs_local;
8448: PetscCall(VecGetArray(nnsp_vec[0], &send_buffer_vecs));
8449: for (i = 0; i < n_recvs; i++) {
8450: PetscInt j;
8451: for (j = 0; j < *(ptr_idxs + 1); j++) send_buffer_vecs[*(ptr_idxs + 2 + j)] += *(ptr_vals + j);
8452: ptr_idxs += olengths_idxs[i];
8453: ptr_vals += olengths_idxs[i] - 2;
8454: }
8455: PetscCall(VecRestoreArray(nnsp_vec[0], &send_buffer_vecs));
8456: PetscCall(VecAssemblyBegin(nnsp_vec[0]));
8457: PetscCall(VecAssemblyEnd(nnsp_vec[0]));
8458: }
8460: PetscCall(PetscFree(recv_buffer_vecs));
8461: PetscCall(PetscFree(recv_buffer_idxs_local));
8462: PetscCall(PetscFree(recv_req_idxs));
8463: PetscCall(PetscFree(recv_req_vals));
8464: PetscCall(PetscFree(recv_req_vecs));
8465: PetscCall(PetscFree(recv_req_idxs_is));
8466: PetscCall(PetscFree(send_req_idxs));
8467: PetscCall(PetscFree(send_req_vals));
8468: PetscCall(PetscFree(send_req_vecs));
8469: PetscCall(PetscFree(send_req_idxs_is));
8470: PetscCall(PetscFree(ilengths_vals));
8471: PetscCall(PetscFree(ilengths_idxs));
8472: PetscCall(PetscFree(olengths_vals));
8473: PetscCall(PetscFree(olengths_idxs));
8474: PetscCall(PetscFree(onodes));
8475: if (nis) {
8476: PetscCall(PetscFree(ilengths_idxs_is));
8477: PetscCall(PetscFree(olengths_idxs_is));
8478: PetscCall(PetscFree(onodes_is));
8479: }
8480: PetscCall(PetscSubcommDestroy(&subcomm));
8481: if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not participate */
8482: PetscCall(MatDestroy(mat_n));
8483: for (i = 0; i < nis; i++) PetscCall(ISDestroy(&isarray[i]));
8484: if (nvecs) { /* need to match VecDestroy nnsp_vec called in the other code path */
8485: PetscCall(VecDestroy(&nnsp_vec[0]));
8486: }
8487: *mat_n = NULL;
8488: }
8489: PetscFunctionReturn(PETSC_SUCCESS);
8490: }
8492: /* temporary hack into ksp private data structure */
8493: #include <petsc/private/kspimpl.h>
8495: PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc, Mat coarse_submat)
8496: {
8497: PC_BDDC *pcbddc = (PC_BDDC *)pc->data;
8498: PC_IS *pcis = (PC_IS *)pc->data;
8499: PCBDDCGraph graph = pcbddc->mat_graph;
8500: Mat coarse_mat, coarse_mat_is;
8501: Mat coarsedivudotp = NULL;
8502: Mat coarseG, t_coarse_mat_is;
8503: MatNullSpace CoarseNullSpace = NULL;
8504: ISLocalToGlobalMapping coarse_islg;
8505: IS coarse_is, *isarray, corners;
8506: PetscInt i, im_active = -1, active_procs = -1;
8507: PetscInt nis, nisdofs, nisneu, nisvert;
8508: PetscInt coarse_eqs_per_proc, coarsening_ratio;
8509: PC pc_temp;
8510: PCType coarse_pc_type;
8511: KSPType coarse_ksp_type;
8512: PetscBool multilevel_requested, multilevel_allowed;
8513: PetscBool coarse_reuse, multi_element = graph->multi_element;
8514: PetscInt ncoarse, nedcfield;
8515: PetscBool compute_vecs = PETSC_FALSE;
8516: PetscScalar *array;
8517: MatReuse coarse_mat_reuse;
8518: PetscBool restr, full_restr, have_void;
8519: PetscMPIInt size;
8521: PetscFunctionBegin;
8522: PetscCall(PetscLogEventBegin(PC_BDDC_CoarseSetUp[pcbddc->current_level], pc, 0, 0, 0));
8523: /* Assign global numbering to coarse dofs */
8524: if (pcbddc->new_primal_space || pcbddc->coarse_size == -1) { /* a new primal space is present or it is the first initialization, so recompute global numbering */
8525: PetscInt ocoarse_size;
8526: compute_vecs = PETSC_TRUE;
8528: pcbddc->new_primal_space = PETSC_TRUE;
8529: ocoarse_size = pcbddc->coarse_size;
8530: PetscCall(PetscFree(pcbddc->global_primal_indices));
8531: PetscCall(PCBDDCComputePrimalNumbering(pc, &pcbddc->coarse_size, &pcbddc->global_primal_indices));
8532: /* see if we can avoid some work */
8533: if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */
8534: /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */
8535: if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) {
8536: PetscCall(KSPReset(pcbddc->coarse_ksp));
8537: coarse_reuse = PETSC_FALSE;
8538: } else { /* we can safely reuse already computed coarse matrix */
8539: coarse_reuse = PETSC_TRUE;
8540: }
8541: } else { /* there's no coarse ksp, so we need to create the coarse matrix too */
8542: coarse_reuse = PETSC_FALSE;
8543: }
8544: /* reset any subassembling information */
8545: if (!coarse_reuse || pcbddc->recompute_topography) PetscCall(ISDestroy(&pcbddc->coarse_subassembling));
8546: } else { /* primal space is unchanged, so we can reuse coarse matrix */
8547: coarse_reuse = PETSC_TRUE;
8548: }
8549: if (coarse_reuse && pcbddc->coarse_ksp) {
8550: PetscCall(KSPGetOperators(pcbddc->coarse_ksp, &coarse_mat, NULL));
8551: PetscCall(PetscObjectReference((PetscObject)coarse_mat));
8552: coarse_mat_reuse = MAT_REUSE_MATRIX;
8553: } else {
8554: coarse_mat = NULL;
8555: coarse_mat_reuse = MAT_INITIAL_MATRIX;
8556: }
8558: /* creates temporary l2gmap and IS for coarse indexes */
8559: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)pc), pcbddc->local_primal_size, pcbddc->global_primal_indices, PETSC_COPY_VALUES, &coarse_is));
8560: PetscCall(ISLocalToGlobalMappingCreateIS(coarse_is, &coarse_islg));
8562: /* creates temporary MATIS object for coarse matrix */
8563: PetscCall(MatCreate(PetscObjectComm((PetscObject)pc), &t_coarse_mat_is));
8564: PetscCall(MatSetType(t_coarse_mat_is, MATIS));
8565: PetscCall(MatSetSizes(t_coarse_mat_is, PETSC_DECIDE, PETSC_DECIDE, pcbddc->coarse_size, pcbddc->coarse_size));
8566: PetscCall(MatISSetAllowRepeated(t_coarse_mat_is, multi_element));
8567: PetscCall(MatSetLocalToGlobalMapping(t_coarse_mat_is, coarse_islg, coarse_islg));
8568: PetscCall(MatISSetLocalMat(t_coarse_mat_is, coarse_submat));
8569: PetscCall(MatAssemblyBegin(t_coarse_mat_is, MAT_FINAL_ASSEMBLY));
8570: PetscCall(MatAssemblyEnd(t_coarse_mat_is, MAT_FINAL_ASSEMBLY));
8571: PetscCall(MatViewFromOptions(t_coarse_mat_is, (PetscObject)pc, "-pc_bddc_coarse_mat_is_view"));
8573: /* count "active" (i.e. with positive local size) and "void" processes */
8574: im_active = !!pcis->n;
8575: PetscCallMPI(MPIU_Allreduce(&im_active, &active_procs, 1, MPIU_INT, MPI_SUM, PetscObjectComm((PetscObject)pc)));
8577: /* determine number of processes partecipating to coarse solver and compute subassembling pattern */
8578: /* restr : whether we want to exclude senders (which are not receivers) from the subassembling pattern */
8579: /* full_restr : just use the receivers from the subassembling pattern */
8580: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)pc), &size));
8581: coarse_mat_is = NULL;
8582: multilevel_allowed = PETSC_FALSE;
8583: multilevel_requested = PETSC_FALSE;
8584: coarse_eqs_per_proc = PetscMin(PetscMax(pcbddc->coarse_size, 1), pcbddc->coarse_eqs_per_proc);
8585: if (coarse_eqs_per_proc < 0 || size == 1) coarse_eqs_per_proc = PetscMax(pcbddc->coarse_size, 1);
8586: if (pcbddc->current_level < pcbddc->max_levels) multilevel_requested = PETSC_TRUE;
8587: if (pcbddc->coarse_size <= pcbddc->coarse_eqs_limit) multilevel_requested = PETSC_FALSE;
8588: coarsening_ratio = multi_element ? 1 : pcbddc->coarsening_ratio;
8589: if (multilevel_requested) {
8590: ncoarse = active_procs / coarsening_ratio;
8591: restr = PETSC_FALSE;
8592: full_restr = PETSC_FALSE;
8593: } else {
8594: ncoarse = pcbddc->coarse_size / coarse_eqs_per_proc + !!(pcbddc->coarse_size % coarse_eqs_per_proc);
8595: restr = PETSC_TRUE;
8596: full_restr = PETSC_TRUE;
8597: }
8598: if (!pcbddc->coarse_size || (size == 1 && !multi_element)) multilevel_allowed = multilevel_requested = restr = full_restr = PETSC_FALSE;
8599: ncoarse = PetscMax(1, ncoarse);
8600: if (!pcbddc->coarse_subassembling) {
8601: if (coarsening_ratio > 1) {
8602: if (multilevel_requested) {
8603: PetscCall(PCBDDCMatISGetSubassemblingPattern(pc->pmat, &ncoarse, pcbddc->coarse_adj_red, &pcbddc->coarse_subassembling, &have_void));
8604: } else {
8605: PetscCall(PCBDDCMatISGetSubassemblingPattern(t_coarse_mat_is, &ncoarse, pcbddc->coarse_adj_red, &pcbddc->coarse_subassembling, &have_void));
8606: }
8607: } else {
8608: PetscMPIInt rank;
8610: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)pc), &rank));
8611: have_void = (active_procs == size) ? PETSC_FALSE : PETSC_TRUE;
8612: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)pc), 1, rank, 1, &pcbddc->coarse_subassembling));
8613: PetscCall(PetscObjectSetName((PetscObject)pcbddc->coarse_subassembling, "default subassembling"));
8614: }
8615: } else { /* if a subassembling pattern exists, then we can reuse the coarse ksp and compute the number of process involved */
8616: PetscInt psum;
8617: if (pcbddc->coarse_ksp) psum = 1;
8618: else psum = 0;
8619: PetscCallMPI(MPIU_Allreduce(&psum, &ncoarse, 1, MPIU_INT, MPI_SUM, PetscObjectComm((PetscObject)pc)));
8620: have_void = ncoarse < size ? PETSC_TRUE : PETSC_FALSE;
8621: }
8622: /* determine if we can go multilevel */
8623: if (multilevel_requested) {
8624: if (ncoarse > 1) multilevel_allowed = PETSC_TRUE; /* found enough processes */
8625: else restr = full_restr = PETSC_TRUE; /* 1 subdomain, use a direct solver */
8626: }
8627: if (multilevel_allowed && have_void) restr = PETSC_TRUE;
8629: /* dump subassembling pattern */
8630: if (pcbddc->dbg_flag && multilevel_allowed) PetscCall(ISView(pcbddc->coarse_subassembling, pcbddc->dbg_viewer));
8631: /* compute dofs splitting and neumann boundaries for coarse dofs */
8632: nedcfield = -1;
8633: corners = NULL;
8634: if (multilevel_allowed && !coarse_reuse && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal || pcbddc->nedclocal || pcbddc->corner_selected)) { /* protects from unneeded computations */
8635: PetscInt *tidxs, *tidxs2, nout, tsize, i;
8636: const PetscInt *idxs;
8637: ISLocalToGlobalMapping tmap;
8639: /* create map between primal indices (in local representative ordering) and local primal numbering */
8640: PetscCall(ISLocalToGlobalMappingCreate(PETSC_COMM_SELF, 1, pcbddc->local_primal_size, pcbddc->primal_indices_local_idxs, PETSC_COPY_VALUES, &tmap));
8641: /* allocate space for temporary storage */
8642: PetscCall(PetscMalloc1(pcbddc->local_primal_size, &tidxs));
8643: PetscCall(PetscMalloc1(pcbddc->local_primal_size, &tidxs2));
8644: /* allocate for IS array */
8645: nisdofs = pcbddc->n_ISForDofsLocal;
8646: if (pcbddc->nedclocal) {
8647: if (pcbddc->nedfield > -1) {
8648: nedcfield = pcbddc->nedfield;
8649: } else {
8650: nedcfield = 0;
8651: PetscCheck(!nisdofs, PetscObjectComm((PetscObject)pc), PETSC_ERR_PLIB, "This should not happen (%" PetscInt_FMT ")", nisdofs);
8652: nisdofs = 1;
8653: }
8654: }
8655: nisneu = !!pcbddc->NeumannBoundariesLocal;
8656: nisvert = 0; /* nisvert is not used */
8657: nis = nisdofs + nisneu + nisvert;
8658: PetscCall(PetscMalloc1(nis, &isarray));
8659: /* dofs splitting */
8660: for (i = 0; i < nisdofs; i++) {
8661: /* PetscCall(ISView(pcbddc->ISForDofsLocal[i],0)); */
8662: if (nedcfield != i) {
8663: PetscCall(ISGetLocalSize(pcbddc->ISForDofsLocal[i], &tsize));
8664: PetscCall(ISGetIndices(pcbddc->ISForDofsLocal[i], &idxs));
8665: PetscCall(ISGlobalToLocalMappingApply(tmap, IS_GTOLM_DROP, tsize, idxs, &nout, tidxs));
8666: PetscCall(ISRestoreIndices(pcbddc->ISForDofsLocal[i], &idxs));
8667: } else {
8668: PetscCall(ISGetLocalSize(pcbddc->nedclocal, &tsize));
8669: PetscCall(ISGetIndices(pcbddc->nedclocal, &idxs));
8670: PetscCall(ISGlobalToLocalMappingApply(tmap, IS_GTOLM_DROP, tsize, idxs, &nout, tidxs));
8671: PetscCheck(tsize == nout, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Failed when mapping coarse nedelec field! %" PetscInt_FMT " != %" PetscInt_FMT, tsize, nout);
8672: PetscCall(ISRestoreIndices(pcbddc->nedclocal, &idxs));
8673: }
8674: PetscCall(ISLocalToGlobalMappingApply(coarse_islg, nout, tidxs, tidxs2));
8675: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)pc), nout, tidxs2, PETSC_COPY_VALUES, &isarray[i]));
8676: /* PetscCall(ISView(isarray[i],0)); */
8677: }
8678: /* neumann boundaries */
8679: if (pcbddc->NeumannBoundariesLocal) {
8680: /* PetscCall(ISView(pcbddc->NeumannBoundariesLocal,0)); */
8681: PetscCall(ISGetLocalSize(pcbddc->NeumannBoundariesLocal, &tsize));
8682: PetscCall(ISGetIndices(pcbddc->NeumannBoundariesLocal, &idxs));
8683: PetscCall(ISGlobalToLocalMappingApply(tmap, IS_GTOLM_DROP, tsize, idxs, &nout, tidxs));
8684: PetscCall(ISRestoreIndices(pcbddc->NeumannBoundariesLocal, &idxs));
8685: PetscCall(ISLocalToGlobalMappingApply(coarse_islg, nout, tidxs, tidxs2));
8686: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)pc), nout, tidxs2, PETSC_COPY_VALUES, &isarray[nisdofs]));
8687: /* PetscCall(ISView(isarray[nisdofs],0)); */
8688: }
8689: /* coordinates */
8690: if (pcbddc->corner_selected) {
8691: PetscCall(PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph, NULL, NULL, NULL, NULL, &corners));
8692: PetscCall(ISGetLocalSize(corners, &tsize));
8693: PetscCall(ISGetIndices(corners, &idxs));
8694: PetscCall(ISGlobalToLocalMappingApply(tmap, IS_GTOLM_DROP, tsize, idxs, &nout, tidxs));
8695: PetscCheck(tsize == nout, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Failed when mapping corners! %" PetscInt_FMT " != %" PetscInt_FMT, tsize, nout);
8696: PetscCall(ISRestoreIndices(corners, &idxs));
8697: PetscCall(PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph, NULL, NULL, NULL, NULL, &corners));
8698: PetscCall(ISLocalToGlobalMappingApply(coarse_islg, nout, tidxs, tidxs2));
8699: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)pc), nout, tidxs2, PETSC_COPY_VALUES, &corners));
8700: }
8701: PetscCall(PetscFree(tidxs));
8702: PetscCall(PetscFree(tidxs2));
8703: PetscCall(ISLocalToGlobalMappingDestroy(&tmap));
8704: } else {
8705: nis = 0;
8706: nisdofs = 0;
8707: nisneu = 0;
8708: nisvert = 0;
8709: isarray = NULL;
8710: }
8711: /* destroy no longer needed map */
8712: PetscCall(ISLocalToGlobalMappingDestroy(&coarse_islg));
8714: /* subassemble */
8715: if (multilevel_allowed) {
8716: Vec vp[1];
8717: PetscInt nvecs = 0;
8718: PetscBool reuse;
8720: vp[0] = NULL;
8721: /* XXX HDIV also */
8722: if (pcbddc->benign_have_null) { /* propagate no-net-flux quadrature to coarser level */
8723: PetscCall(VecCreate(PetscObjectComm((PetscObject)pc), &vp[0]));
8724: PetscCall(VecSetSizes(vp[0], pcbddc->local_primal_size, PETSC_DECIDE));
8725: PetscCall(VecSetType(vp[0], VECSTANDARD));
8726: nvecs = 1;
8728: if (pcbddc->divudotp) {
8729: Mat B, loc_divudotp;
8730: Vec v, p;
8731: IS dummy;
8732: PetscInt np;
8734: PetscCall(MatISGetLocalMat(pcbddc->divudotp, &loc_divudotp));
8735: PetscCall(MatGetSize(loc_divudotp, &np, NULL));
8736: PetscCall(ISCreateStride(PETSC_COMM_SELF, np, 0, 1, &dummy));
8737: PetscCall(MatCreateSubMatrix(loc_divudotp, dummy, pcis->is_B_local, MAT_INITIAL_MATRIX, &B));
8738: PetscCall(MatCreateVecs(B, &v, &p));
8739: PetscCall(VecSet(p, 1.));
8740: PetscCall(MatMultTranspose(B, p, v));
8741: PetscCall(VecDestroy(&p));
8742: PetscCall(MatDestroy(&B));
8743: PetscCall(VecGetArray(vp[0], &array));
8744: PetscCall(VecPlaceArray(pcbddc->vec1_P, array));
8745: PetscCall(MatMultTranspose(pcbddc->coarse_phi_B, v, pcbddc->vec1_P));
8746: PetscCall(VecResetArray(pcbddc->vec1_P));
8747: PetscCall(VecRestoreArray(vp[0], &array));
8748: PetscCall(ISDestroy(&dummy));
8749: PetscCall(VecDestroy(&v));
8750: }
8751: }
8752: if (coarse_mat) reuse = PETSC_TRUE;
8753: else reuse = PETSC_FALSE;
8754: if (multi_element) {
8755: PetscCall(PetscObjectReference((PetscObject)t_coarse_mat_is));
8756: coarse_mat_is = t_coarse_mat_is;
8757: } else {
8758: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &reuse, 1, MPI_C_BOOL, MPI_LOR, PetscObjectComm((PetscObject)pc)));
8759: if (reuse) {
8760: PetscCall(PCBDDCMatISSubassemble(t_coarse_mat_is, pcbddc->coarse_subassembling, 0, restr, full_restr, PETSC_TRUE, &coarse_mat, nis, isarray, nvecs, vp));
8761: } else {
8762: PetscCall(PCBDDCMatISSubassemble(t_coarse_mat_is, pcbddc->coarse_subassembling, 0, restr, full_restr, PETSC_FALSE, &coarse_mat_is, nis, isarray, nvecs, vp));
8763: }
8764: if (vp[0]) { /* vp[0] could have been placed on a different set of processes */
8765: PetscScalar *arraym;
8766: const PetscScalar *arrayv;
8767: PetscInt nl;
8768: PetscCall(VecGetLocalSize(vp[0], &nl));
8769: PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, 1, nl, NULL, &coarsedivudotp));
8770: PetscCall(MatDenseGetArray(coarsedivudotp, &arraym));
8771: PetscCall(VecGetArrayRead(vp[0], &arrayv));
8772: PetscCall(PetscArraycpy(arraym, arrayv, nl));
8773: PetscCall(VecRestoreArrayRead(vp[0], &arrayv));
8774: PetscCall(MatDenseRestoreArray(coarsedivudotp, &arraym));
8775: PetscCall(VecDestroy(&vp[0]));
8776: } else {
8777: PetscCall(MatCreateSeqAIJ(PETSC_COMM_SELF, 0, 0, 1, NULL, &coarsedivudotp));
8778: }
8779: }
8780: } else {
8781: PetscBool default_sub;
8783: PetscCall(PetscStrcmp(((PetscObject)pcbddc->coarse_subassembling)->name, "default subassembling", &default_sub));
8784: if (!default_sub) PetscCall(PCBDDCMatISSubassemble(t_coarse_mat_is, pcbddc->coarse_subassembling, 0, restr, full_restr, PETSC_FALSE, &coarse_mat_is, 0, NULL, 0, NULL));
8785: else {
8786: PetscCall(PetscObjectReference((PetscObject)t_coarse_mat_is));
8787: coarse_mat_is = t_coarse_mat_is;
8788: }
8789: }
8790: if (coarse_mat_is || coarse_mat) {
8791: if (!multilevel_allowed) {
8792: PetscCall(MatConvert(coarse_mat_is, MATAIJ, coarse_mat_reuse, &coarse_mat));
8793: } else {
8794: /* if this matrix is present, it means we are not reusing the coarse matrix */
8795: if (coarse_mat_is) {
8796: PetscCheck(!coarse_mat, PetscObjectComm((PetscObject)coarse_mat_is), PETSC_ERR_PLIB, "This should not happen");
8797: PetscCall(PetscObjectReference((PetscObject)coarse_mat_is));
8798: coarse_mat = coarse_mat_is;
8799: }
8800: }
8801: }
8802: PetscCall(MatDestroy(&t_coarse_mat_is));
8803: PetscCall(MatDestroy(&coarse_mat_is));
8805: /* create local to global scatters for coarse problem */
8806: if (compute_vecs) {
8807: PetscInt lrows;
8808: PetscCall(VecDestroy(&pcbddc->coarse_vec));
8809: if (coarse_mat) {
8810: PetscCall(MatGetLocalSize(coarse_mat, &lrows, NULL));
8811: } else {
8812: lrows = 0;
8813: }
8814: PetscCall(VecCreate(PetscObjectComm((PetscObject)pc), &pcbddc->coarse_vec));
8815: PetscCall(VecSetSizes(pcbddc->coarse_vec, lrows, PETSC_DECIDE));
8816: PetscCall(VecSetType(pcbddc->coarse_vec, coarse_mat ? coarse_mat->defaultvectype : VECSTANDARD));
8817: PetscCall(VecScatterDestroy(&pcbddc->coarse_loc_to_glob));
8818: PetscCall(VecScatterCreate(pcbddc->vec1_P, NULL, pcbddc->coarse_vec, coarse_is, &pcbddc->coarse_loc_to_glob));
8819: }
8820: PetscCall(ISDestroy(&coarse_is));
8822: /* set defaults for coarse KSP and PC */
8823: if (multilevel_allowed) {
8824: coarse_ksp_type = KSPRICHARDSON;
8825: coarse_pc_type = PCBDDC;
8826: } else {
8827: coarse_ksp_type = KSPPREONLY;
8828: coarse_pc_type = PCREDUNDANT;
8829: }
8831: /* print some info if requested */
8832: if (pcbddc->dbg_flag) {
8833: if (!multilevel_allowed) {
8834: PetscCall(PetscViewerASCIIPrintf(pcbddc->dbg_viewer, "--------------------------------------------------\n"));
8835: if (multilevel_requested) {
8836: PetscCall(PetscViewerASCIIPrintf(pcbddc->dbg_viewer, "Not enough active processes on level %" PetscInt_FMT " (active processes %" PetscInt_FMT ", coarsening ratio %" PetscInt_FMT ")\n", pcbddc->current_level, active_procs, coarsening_ratio));
8837: } else if (pcbddc->max_levels) {
8838: PetscCall(PetscViewerASCIIPrintf(pcbddc->dbg_viewer, "Maximum number of requested levels reached (%" PetscInt_FMT ")\n", pcbddc->max_levels));
8839: }
8840: PetscCall(PetscViewerFlush(pcbddc->dbg_viewer));
8841: }
8842: }
8844: /* communicate coarse discrete gradient */
8845: coarseG = NULL;
8846: if (pcbddc->nedcG && multilevel_allowed) {
8847: MPI_Comm ccomm;
8848: if (coarse_mat) {
8849: ccomm = PetscObjectComm((PetscObject)coarse_mat);
8850: } else {
8851: ccomm = MPI_COMM_NULL;
8852: }
8853: PetscCall(MatMPIAIJRestrict(pcbddc->nedcG, ccomm, &coarseG));
8854: }
8856: /* create the coarse KSP object only once with defaults */
8857: if (coarse_mat) {
8858: PetscBool isredundant, isbddc, force, valid;
8859: PetscViewer dbg_viewer = NULL;
8860: PetscBool isset, issym, isher, isspd;
8862: if (pcbddc->dbg_flag) {
8863: dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat));
8864: PetscCall(PetscViewerASCIIAddTab(dbg_viewer, 2 * pcbddc->current_level));
8865: }
8866: if (!pcbddc->coarse_ksp) {
8867: char prefix[256], str_level[16];
8868: size_t len;
8870: PetscCall(KSPCreate(PetscObjectComm((PetscObject)coarse_mat), &pcbddc->coarse_ksp));
8871: PetscCall(KSPSetNestLevel(pcbddc->coarse_ksp, pc->kspnestlevel));
8872: PetscCall(KSPSetErrorIfNotConverged(pcbddc->coarse_ksp, pc->erroriffailure));
8873: PetscCall(PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp, (PetscObject)pc, 1));
8874: PetscCall(KSPSetTolerances(pcbddc->coarse_ksp, PETSC_CURRENT, PETSC_CURRENT, PETSC_CURRENT, 1));
8875: PetscCall(KSPSetOperators(pcbddc->coarse_ksp, coarse_mat, coarse_mat));
8876: PetscCall(KSPSetType(pcbddc->coarse_ksp, coarse_ksp_type));
8877: PetscCall(KSPSetNormType(pcbddc->coarse_ksp, KSP_NORM_NONE));
8878: PetscCall(KSPGetPC(pcbddc->coarse_ksp, &pc_temp));
8879: /* TODO is this logic correct? should check for coarse_mat type */
8880: PetscCall(PCSetType(pc_temp, coarse_pc_type));
8881: /* prefix */
8882: PetscCall(PetscStrncpy(prefix, "", sizeof(prefix)));
8883: PetscCall(PetscStrncpy(str_level, "", sizeof(str_level)));
8884: if (!pcbddc->current_level) {
8885: PetscCall(PetscStrncpy(prefix, ((PetscObject)pc)->prefix, sizeof(prefix)));
8886: PetscCall(PetscStrlcat(prefix, "pc_bddc_coarse_", sizeof(prefix)));
8887: } else {
8888: PetscCall(PetscStrlen(((PetscObject)pc)->prefix, &len));
8889: if (pcbddc->current_level > 1) len -= 3; /* remove "lX_" with X level number */
8890: if (pcbddc->current_level > 10) len -= 1; /* remove another char from level number */
8891: /* Nonstandard use of PetscStrncpy() to copy only a portion of the string */
8892: PetscCall(PetscStrncpy(prefix, ((PetscObject)pc)->prefix, len + 1));
8893: PetscCall(PetscSNPrintf(str_level, sizeof(str_level), "l%" PetscInt_FMT "_", pcbddc->current_level));
8894: PetscCall(PetscStrlcat(prefix, str_level, sizeof(prefix)));
8895: }
8896: PetscCall(KSPSetOptionsPrefix(pcbddc->coarse_ksp, prefix));
8897: /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */
8898: PetscCall(PCBDDCSetLevel(pc_temp, pcbddc->current_level + 1));
8899: PetscCall(PCBDDCSetCoarseningRatio(pc_temp, pcbddc->coarsening_ratio));
8900: PetscCall(PCBDDCSetLevels(pc_temp, pcbddc->max_levels));
8901: /* allow user customization */
8902: PetscCall(KSPSetFromOptions(pcbddc->coarse_ksp));
8903: /* get some info after set from options */
8904: PetscCall(KSPGetPC(pcbddc->coarse_ksp, &pc_temp));
8905: /* multilevel cannot be done with coarse PC different from BDDC, NN, HPDDM, unless forced to */
8906: force = PETSC_FALSE;
8907: PetscCall(PetscOptionsGetBool(NULL, ((PetscObject)pc_temp)->prefix, "-pc_type_forced", &force, NULL));
8908: PetscCall(PetscObjectTypeCompareAny((PetscObject)pc_temp, &valid, PCBDDC, PCNN, PCHPDDM, ""));
8909: PetscCall(PetscObjectTypeCompare((PetscObject)pc_temp, PCBDDC, &isbddc));
8910: if (multilevel_allowed && !force && !valid) {
8911: isbddc = PETSC_TRUE;
8912: PetscCall(PCSetType(pc_temp, PCBDDC));
8913: PetscCall(PCBDDCSetLevel(pc_temp, pcbddc->current_level + 1));
8914: PetscCall(PCBDDCSetCoarseningRatio(pc_temp, pcbddc->coarsening_ratio));
8915: PetscCall(PCBDDCSetLevels(pc_temp, pcbddc->max_levels));
8916: if (pc_temp->ops->setfromoptions) { /* need to setfromoptions again, skipping the pc_type */
8917: PetscObjectOptionsBegin((PetscObject)pc_temp);
8918: PetscCall((*pc_temp->ops->setfromoptions)(pc_temp, PetscOptionsObject));
8919: PetscCall(PetscObjectProcessOptionsHandlers((PetscObject)pc_temp, PetscOptionsObject));
8920: PetscOptionsEnd();
8921: pc_temp->setfromoptionscalled++;
8922: }
8923: }
8924: }
8925: /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */
8926: PetscCall(KSPGetPC(pcbddc->coarse_ksp, &pc_temp));
8927: if (nisdofs) {
8928: PetscCall(PCBDDCSetDofsSplitting(pc_temp, nisdofs, isarray));
8929: for (i = 0; i < nisdofs; i++) PetscCall(ISDestroy(&isarray[i]));
8930: }
8931: if (nisneu) {
8932: PetscCall(PCBDDCSetNeumannBoundaries(pc_temp, isarray[nisdofs]));
8933: PetscCall(ISDestroy(&isarray[nisdofs]));
8934: }
8935: if (nisvert) {
8936: PetscCall(PCBDDCSetPrimalVerticesIS(pc_temp, isarray[nis - 1]));
8937: PetscCall(ISDestroy(&isarray[nis - 1]));
8938: }
8939: if (coarseG) PetscCall(PCBDDCSetDiscreteGradient(pc_temp, coarseG, 1, nedcfield, PETSC_FALSE, PETSC_TRUE));
8941: /* get some info after set from options */
8942: PetscCall(PetscObjectTypeCompare((PetscObject)pc_temp, PCBDDC, &isbddc));
8944: /* multilevel can only be requested via -pc_bddc_levels or PCBDDCSetLevels */
8945: if (isbddc && !multilevel_allowed) PetscCall(PCSetType(pc_temp, coarse_pc_type));
8946: /* multilevel cannot be done with coarse PC different from BDDC, NN, HPDDM, unless forced to */
8947: force = PETSC_FALSE;
8948: PetscCall(PetscOptionsGetBool(NULL, ((PetscObject)pc_temp)->prefix, "-pc_type_forced", &force, NULL));
8949: PetscCall(PetscObjectTypeCompareAny((PetscObject)pc_temp, &valid, PCBDDC, PCNN, PCHPDDM, ""));
8950: if (multilevel_requested && multilevel_allowed && !valid && !force) PetscCall(PCSetType(pc_temp, PCBDDC));
8951: PetscCall(PetscObjectTypeCompare((PetscObject)pc_temp, PCREDUNDANT, &isredundant));
8952: if (isredundant) {
8953: KSP inner_ksp;
8954: PC inner_pc;
8956: PetscCall(PCRedundantGetKSP(pc_temp, &inner_ksp));
8957: PetscCall(KSPGetPC(inner_ksp, &inner_pc));
8958: }
8960: /* parameters which miss an API */
8961: PetscCall(PetscObjectTypeCompare((PetscObject)pc_temp, PCBDDC, &isbddc));
8962: if (isbddc) {
8963: PC_BDDC *pcbddc_coarse = (PC_BDDC *)pc_temp->data;
8965: pcbddc_coarse->detect_disconnected = PETSC_TRUE;
8966: pcbddc_coarse->coarse_eqs_per_proc = pcbddc->coarse_eqs_per_proc;
8967: pcbddc_coarse->coarse_eqs_limit = pcbddc->coarse_eqs_limit;
8968: pcbddc_coarse->benign_saddle_point = pcbddc->benign_have_null;
8969: if (pcbddc_coarse->benign_saddle_point) {
8970: Mat coarsedivudotp_is;
8971: ISLocalToGlobalMapping l2gmap, rl2g, cl2g;
8972: IS row, col;
8973: const PetscInt *gidxs;
8974: PetscInt n, st, M, N;
8976: PetscCall(MatGetSize(coarsedivudotp, &n, NULL));
8977: PetscCallMPI(MPI_Scan(&n, &st, 1, MPIU_INT, MPI_SUM, PetscObjectComm((PetscObject)coarse_mat)));
8978: st = st - n;
8979: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)coarse_mat), 1, st, 1, &row));
8980: PetscCall(MatISGetLocalToGlobalMapping(coarse_mat, &l2gmap, NULL));
8981: PetscCall(ISLocalToGlobalMappingGetSize(l2gmap, &n));
8982: PetscCall(ISLocalToGlobalMappingGetIndices(l2gmap, &gidxs));
8983: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)coarse_mat), n, gidxs, PETSC_COPY_VALUES, &col));
8984: PetscCall(ISLocalToGlobalMappingRestoreIndices(l2gmap, &gidxs));
8985: PetscCall(ISLocalToGlobalMappingCreateIS(row, &rl2g));
8986: PetscCall(ISLocalToGlobalMappingCreateIS(col, &cl2g));
8987: PetscCall(ISGetSize(row, &M));
8988: PetscCall(MatGetSize(coarse_mat, &N, NULL));
8989: PetscCall(ISDestroy(&row));
8990: PetscCall(ISDestroy(&col));
8991: PetscCall(MatCreate(PetscObjectComm((PetscObject)coarse_mat), &coarsedivudotp_is));
8992: PetscCall(MatSetType(coarsedivudotp_is, MATIS));
8993: PetscCall(MatSetSizes(coarsedivudotp_is, PETSC_DECIDE, PETSC_DECIDE, M, N));
8994: PetscCall(MatSetLocalToGlobalMapping(coarsedivudotp_is, rl2g, cl2g));
8995: PetscCall(ISLocalToGlobalMappingDestroy(&rl2g));
8996: PetscCall(ISLocalToGlobalMappingDestroy(&cl2g));
8997: PetscCall(MatISSetLocalMat(coarsedivudotp_is, coarsedivudotp));
8998: PetscCall(MatDestroy(&coarsedivudotp));
8999: PetscCall(PCBDDCSetDivergenceMat(pc_temp, coarsedivudotp_is, PETSC_FALSE, NULL));
9000: PetscCall(MatDestroy(&coarsedivudotp_is));
9001: pcbddc_coarse->adaptive_userdefined = PETSC_TRUE;
9002: if (pcbddc->adaptive_threshold[0] == 0.0) pcbddc_coarse->deluxe_zerorows = PETSC_TRUE;
9003: }
9004: }
9006: /* propagate symmetry info of coarse matrix */
9007: PetscCall(MatSetOption(coarse_mat, MAT_STRUCTURALLY_SYMMETRIC, PETSC_TRUE));
9008: PetscCall(MatIsSymmetricKnown(pc->pmat, &isset, &issym));
9009: if (isset) PetscCall(MatSetOption(coarse_mat, MAT_SYMMETRIC, issym));
9010: PetscCall(MatIsHermitianKnown(pc->pmat, &isset, &isher));
9011: if (isset) PetscCall(MatSetOption(coarse_mat, MAT_HERMITIAN, isher));
9012: PetscCall(MatIsSPDKnown(pc->pmat, &isset, &isspd));
9013: if (isset) PetscCall(MatSetOption(coarse_mat, MAT_SPD, isspd));
9015: if (pcbddc->benign_saddle_point && !pcbddc->benign_have_null) PetscCall(MatSetOption(coarse_mat, MAT_SPD, PETSC_TRUE));
9016: /* set operators */
9017: PetscCall(MatViewFromOptions(coarse_mat, (PetscObject)pc, "-pc_bddc_coarse_mat_view"));
9018: PetscCall(MatSetOptionsPrefix(coarse_mat, ((PetscObject)pcbddc->coarse_ksp)->prefix));
9019: PetscCall(KSPSetOperators(pcbddc->coarse_ksp, coarse_mat, coarse_mat));
9020: if (pcbddc->dbg_flag) PetscCall(PetscViewerASCIISubtractTab(dbg_viewer, 2 * pcbddc->current_level));
9021: }
9022: PetscCall(MatDestroy(&coarseG));
9023: PetscCall(PetscFree(isarray));
9024: #if 0
9025: {
9026: PetscViewer viewer;
9027: char filename[256];
9028: PetscCall(PetscSNPrintf(filename, PETSC_STATIC_ARRAY_LENGTH(filename), "coarse_mat_level%d.m",pcbddc->current_level));
9029: PetscCall(PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer));
9030: PetscCall(PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB));
9031: PetscCall(MatView(coarse_mat,viewer));
9032: PetscCall(PetscViewerPopFormat(viewer));
9033: PetscCall(PetscViewerDestroy(&viewer));
9034: }
9035: #endif
9037: if (corners) {
9038: Vec gv;
9039: IS is;
9040: const PetscInt *idxs;
9041: PetscInt i, d, N, n, cdim = pcbddc->mat_graph->cdim;
9042: PetscScalar *coords;
9044: PetscCheck(pcbddc->mat_graph->cloc, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Missing local coordinates");
9045: PetscCall(VecGetSize(pcbddc->coarse_vec, &N));
9046: PetscCall(VecGetLocalSize(pcbddc->coarse_vec, &n));
9047: PetscCall(VecCreate(PetscObjectComm((PetscObject)pcbddc->coarse_vec), &gv));
9048: PetscCall(VecSetBlockSize(gv, cdim));
9049: PetscCall(VecSetSizes(gv, n * cdim, N * cdim));
9050: PetscCall(VecSetType(gv, VECSTANDARD));
9051: PetscCall(VecSetFromOptions(gv));
9052: PetscCall(VecSet(gv, PETSC_MAX_REAL)); /* we only propagate coordinates from vertices constraints */
9054: PetscCall(PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph, NULL, NULL, NULL, NULL, &is));
9055: PetscCall(ISGetLocalSize(is, &n));
9056: PetscCall(ISGetIndices(is, &idxs));
9057: PetscCall(PetscMalloc1(n * cdim, &coords));
9058: for (i = 0; i < n; i++) {
9059: for (d = 0; d < cdim; d++) coords[cdim * i + d] = pcbddc->mat_graph->coords[cdim * idxs[i] + d];
9060: }
9061: PetscCall(ISRestoreIndices(is, &idxs));
9062: PetscCall(PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph, NULL, NULL, NULL, NULL, &is));
9064: PetscCall(ISGetLocalSize(corners, &n));
9065: PetscCall(ISGetIndices(corners, &idxs));
9066: PetscCall(VecSetValuesBlocked(gv, n, idxs, coords, INSERT_VALUES));
9067: PetscCall(ISRestoreIndices(corners, &idxs));
9068: PetscCall(PetscFree(coords));
9069: PetscCall(VecAssemblyBegin(gv));
9070: PetscCall(VecAssemblyEnd(gv));
9071: PetscCall(VecGetArray(gv, &coords));
9072: if (pcbddc->coarse_ksp) {
9073: PC coarse_pc;
9074: PetscBool isbddc;
9076: PetscCall(KSPGetPC(pcbddc->coarse_ksp, &coarse_pc));
9077: PetscCall(PetscObjectTypeCompare((PetscObject)coarse_pc, PCBDDC, &isbddc));
9078: if (isbddc) { /* coarse coordinates have PETSC_MAX_REAL, specific for BDDC */
9079: PetscReal *realcoords;
9081: PetscCall(VecGetLocalSize(gv, &n));
9082: #if defined(PETSC_USE_COMPLEX)
9083: PetscCall(PetscMalloc1(n, &realcoords));
9084: for (i = 0; i < n; i++) realcoords[i] = PetscRealPart(coords[i]);
9085: #else
9086: realcoords = coords;
9087: #endif
9088: PetscCall(PCSetCoordinates(coarse_pc, cdim, n / cdim, realcoords));
9089: #if defined(PETSC_USE_COMPLEX)
9090: PetscCall(PetscFree(realcoords));
9091: #endif
9092: }
9093: }
9094: PetscCall(VecRestoreArray(gv, &coords));
9095: PetscCall(VecDestroy(&gv));
9096: }
9097: PetscCall(ISDestroy(&corners));
9099: if (pcbddc->coarse_ksp) {
9100: Vec crhs, csol;
9102: PetscCall(KSPGetSolution(pcbddc->coarse_ksp, &csol));
9103: PetscCall(KSPGetRhs(pcbddc->coarse_ksp, &crhs));
9104: if (!csol) PetscCall(MatCreateVecs(coarse_mat, &pcbddc->coarse_ksp->vec_sol, NULL));
9105: if (!crhs) PetscCall(MatCreateVecs(coarse_mat, NULL, &pcbddc->coarse_ksp->vec_rhs));
9106: }
9107: PetscCall(MatDestroy(&coarsedivudotp));
9109: /* compute null space for coarse solver if the benign trick has been requested */
9110: if (pcbddc->benign_null) {
9111: PetscCall(VecSet(pcbddc->vec1_P, 0.));
9112: for (i = 0; i < pcbddc->benign_n; i++) PetscCall(VecSetValue(pcbddc->vec1_P, pcbddc->local_primal_size - pcbddc->benign_n + i, 1.0, INSERT_VALUES));
9113: PetscCall(VecAssemblyBegin(pcbddc->vec1_P));
9114: PetscCall(VecAssemblyEnd(pcbddc->vec1_P));
9115: PetscCall(VecScatterBegin(pcbddc->coarse_loc_to_glob, pcbddc->vec1_P, pcbddc->coarse_vec, INSERT_VALUES, SCATTER_FORWARD));
9116: PetscCall(VecScatterEnd(pcbddc->coarse_loc_to_glob, pcbddc->vec1_P, pcbddc->coarse_vec, INSERT_VALUES, SCATTER_FORWARD));
9117: if (coarse_mat) {
9118: Vec nullv;
9119: PetscScalar *array, *array2;
9120: PetscInt nl;
9122: PetscCall(MatCreateVecs(coarse_mat, &nullv, NULL));
9123: PetscCall(VecGetLocalSize(nullv, &nl));
9124: PetscCall(VecGetArrayRead(pcbddc->coarse_vec, (const PetscScalar **)&array));
9125: PetscCall(VecGetArray(nullv, &array2));
9126: PetscCall(PetscArraycpy(array2, array, nl));
9127: PetscCall(VecRestoreArray(nullv, &array2));
9128: PetscCall(VecRestoreArrayRead(pcbddc->coarse_vec, (const PetscScalar **)&array));
9129: PetscCall(VecNormalize(nullv, NULL));
9130: PetscCall(MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat), PETSC_FALSE, 1, &nullv, &CoarseNullSpace));
9131: PetscCall(VecDestroy(&nullv));
9132: }
9133: }
9134: PetscCall(PetscLogEventEnd(PC_BDDC_CoarseSetUp[pcbddc->current_level], pc, 0, 0, 0));
9136: PetscCall(PetscLogEventBegin(PC_BDDC_CoarseSolver[pcbddc->current_level], pc, 0, 0, 0));
9137: if (pcbddc->coarse_ksp) {
9138: PetscBool ispreonly;
9140: if (CoarseNullSpace) {
9141: PetscBool isnull;
9143: PetscCall(MatNullSpaceTest(CoarseNullSpace, coarse_mat, &isnull));
9144: if (isnull) PetscCall(MatSetNullSpace(coarse_mat, CoarseNullSpace));
9145: /* TODO: add local nullspaces (if any) */
9146: }
9147: /* setup coarse ksp */
9148: PetscCall(KSPSetUp(pcbddc->coarse_ksp));
9149: /* Check coarse problem if in debug mode or if solving with an iterative method */
9150: PetscCall(PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp, KSPPREONLY, &ispreonly));
9151: if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates)) {
9152: KSP check_ksp;
9153: KSPType check_ksp_type;
9154: PC check_pc;
9155: Vec check_vec, coarse_vec;
9156: PetscReal abs_infty_error, infty_error, lambda_min = 1.0, lambda_max = 1.0;
9157: PetscInt its;
9158: PetscBool compute_eigs;
9159: PetscReal *eigs_r, *eigs_c;
9160: PetscInt neigs;
9161: const char *prefix;
9163: /* Create ksp object suitable for estimation of extreme eigenvalues */
9164: PetscCall(KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp), &check_ksp));
9165: PetscCall(KSPSetNestLevel(check_ksp, pc->kspnestlevel));
9166: PetscCall(PetscObjectIncrementTabLevel((PetscObject)check_ksp, (PetscObject)pcbddc->coarse_ksp, 0));
9167: PetscCall(KSPSetErrorIfNotConverged(pcbddc->coarse_ksp, PETSC_FALSE));
9168: PetscCall(KSPSetOperators(check_ksp, coarse_mat, coarse_mat));
9169: PetscCall(KSPSetTolerances(check_ksp, 1.e-12, 1.e-12, PETSC_CURRENT, pcbddc->coarse_size));
9170: /* prevent from setup unneeded object */
9171: PetscCall(KSPGetPC(check_ksp, &check_pc));
9172: PetscCall(PCSetType(check_pc, PCNONE));
9173: if (ispreonly) {
9174: check_ksp_type = KSPPREONLY;
9175: compute_eigs = PETSC_FALSE;
9176: } else {
9177: check_ksp_type = KSPGMRES;
9178: compute_eigs = PETSC_TRUE;
9179: }
9180: PetscCall(KSPSetType(check_ksp, check_ksp_type));
9181: PetscCall(KSPSetComputeSingularValues(check_ksp, compute_eigs));
9182: PetscCall(KSPSetComputeEigenvalues(check_ksp, compute_eigs));
9183: PetscCall(KSPGMRESSetRestart(check_ksp, pcbddc->coarse_size + 1));
9184: PetscCall(KSPGetOptionsPrefix(pcbddc->coarse_ksp, &prefix));
9185: PetscCall(KSPSetOptionsPrefix(check_ksp, prefix));
9186: PetscCall(KSPAppendOptionsPrefix(check_ksp, "check_"));
9187: PetscCall(KSPSetFromOptions(check_ksp));
9188: PetscCall(KSPSetUp(check_ksp));
9189: PetscCall(KSPGetPC(pcbddc->coarse_ksp, &check_pc));
9190: PetscCall(KSPSetPC(check_ksp, check_pc));
9191: /* create random vec */
9192: PetscCall(MatCreateVecs(coarse_mat, &coarse_vec, &check_vec));
9193: PetscCall(VecSetRandom(check_vec, NULL));
9194: PetscCall(MatMult(coarse_mat, check_vec, coarse_vec));
9195: /* solve coarse problem */
9196: PetscCall(KSPSolve(check_ksp, coarse_vec, coarse_vec));
9197: PetscCall(KSPCheckSolve(check_ksp, pc, coarse_vec));
9198: /* set eigenvalue estimation if preonly has not been requested */
9199: if (compute_eigs) {
9200: PetscCall(PetscMalloc1(pcbddc->coarse_size + 1, &eigs_r));
9201: PetscCall(PetscMalloc1(pcbddc->coarse_size + 1, &eigs_c));
9202: PetscCall(KSPComputeEigenvalues(check_ksp, pcbddc->coarse_size + 1, eigs_r, eigs_c, &neigs));
9203: if (neigs) {
9204: lambda_max = eigs_r[neigs - 1];
9205: lambda_min = eigs_r[0];
9206: if (pcbddc->use_coarse_estimates) {
9207: if (lambda_max >= lambda_min) { /* using PETSC_SMALL since lambda_max == lambda_min is not allowed by KSPChebyshevSetEigenvalues */
9208: PetscCall(KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp, lambda_max + PETSC_SMALL, lambda_min));
9209: PetscCall(KSPRichardsonSetScale(pcbddc->coarse_ksp, 2.0 / (lambda_max + lambda_min)));
9210: }
9211: }
9212: }
9213: }
9215: /* check coarse problem residual error */
9216: if (pcbddc->dbg_flag) {
9217: PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp));
9218: PetscCall(PetscViewerASCIIAddTab(dbg_viewer, 2 * (pcbddc->current_level + 1)));
9219: PetscCall(VecAXPY(check_vec, -1.0, coarse_vec));
9220: PetscCall(VecNorm(check_vec, NORM_INFINITY, &infty_error));
9221: PetscCall(MatMult(coarse_mat, check_vec, coarse_vec));
9222: PetscCall(VecNorm(coarse_vec, NORM_INFINITY, &abs_infty_error));
9223: PetscCall(PetscViewerASCIIPrintf(dbg_viewer, "Coarse problem details (use estimates %d)\n", pcbddc->use_coarse_estimates));
9224: PetscCall(PetscObjectPrintClassNamePrefixType((PetscObject)pcbddc->coarse_ksp, dbg_viewer));
9225: PetscCall(PetscObjectPrintClassNamePrefixType((PetscObject)check_pc, dbg_viewer));
9226: PetscCall(PetscViewerASCIIPrintf(dbg_viewer, "Coarse problem exact infty_error : %1.6e\n", (double)infty_error));
9227: PetscCall(PetscViewerASCIIPrintf(dbg_viewer, "Coarse problem residual infty_error: %1.6e\n", (double)abs_infty_error));
9228: if (CoarseNullSpace) PetscCall(PetscViewerASCIIPrintf(dbg_viewer, "Coarse problem is singular\n"));
9229: if (compute_eigs) {
9230: PetscReal lambda_max_s, lambda_min_s;
9231: KSPConvergedReason reason;
9232: PetscCall(KSPGetType(check_ksp, &check_ksp_type));
9233: PetscCall(KSPGetIterationNumber(check_ksp, &its));
9234: PetscCall(KSPGetConvergedReason(check_ksp, &reason));
9235: PetscCall(KSPComputeExtremeSingularValues(check_ksp, &lambda_max_s, &lambda_min_s));
9236: PetscCall(PetscViewerASCIIPrintf(dbg_viewer, "Coarse problem eigenvalues (estimated with %" PetscInt_FMT " iterations of %s, conv reason %d): %1.6e %1.6e (%1.6e %1.6e)\n", its, check_ksp_type, reason, (double)lambda_min, (double)lambda_max, (double)lambda_min_s, (double)lambda_max_s));
9237: for (i = 0; i < neigs; i++) PetscCall(PetscViewerASCIIPrintf(dbg_viewer, "%1.6e %1.6ei\n", (double)eigs_r[i], (double)eigs_c[i]));
9238: }
9239: PetscCall(PetscViewerFlush(dbg_viewer));
9240: PetscCall(PetscViewerASCIISubtractTab(dbg_viewer, 2 * (pcbddc->current_level + 1)));
9241: }
9242: PetscCall(VecDestroy(&check_vec));
9243: PetscCall(VecDestroy(&coarse_vec));
9244: PetscCall(KSPDestroy(&check_ksp));
9245: if (compute_eigs) {
9246: PetscCall(PetscFree(eigs_r));
9247: PetscCall(PetscFree(eigs_c));
9248: }
9249: }
9250: }
9251: PetscCall(MatNullSpaceDestroy(&CoarseNullSpace));
9252: /* print additional info */
9253: if (pcbddc->dbg_flag) {
9254: /* waits until all processes reaches this point */
9255: PetscCall(PetscBarrier((PetscObject)pc));
9256: PetscCall(PetscViewerASCIIPrintf(pcbddc->dbg_viewer, "Coarse solver setup completed at level %" PetscInt_FMT "\n", pcbddc->current_level));
9257: PetscCall(PetscViewerFlush(pcbddc->dbg_viewer));
9258: }
9260: /* free memory */
9261: PetscCall(MatDestroy(&coarse_mat));
9262: PetscCall(PetscLogEventEnd(PC_BDDC_CoarseSolver[pcbddc->current_level], pc, 0, 0, 0));
9263: PetscFunctionReturn(PETSC_SUCCESS);
9264: }
9266: PetscErrorCode PCBDDCComputePrimalNumbering(PC pc, PetscInt *coarse_size_n, PetscInt **local_primal_indices_n)
9267: {
9268: PC_BDDC *pcbddc = (PC_BDDC *)pc->data;
9269: PC_IS *pcis = (PC_IS *)pc->data;
9270: IS subset, subset_mult, subset_n;
9271: PetscInt local_size, coarse_size = 0;
9272: PetscInt *local_primal_indices = NULL;
9273: const PetscInt *t_local_primal_indices;
9275: PetscFunctionBegin;
9276: /* Compute global number of coarse dofs */
9277: PetscCheck(!pcbddc->local_primal_size || pcbddc->local_primal_ref_node, PETSC_COMM_SELF, PETSC_ERR_PLIB, "BDDC ConstraintsSetUp should be called first");
9278: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)pc->pmat), pcbddc->local_primal_size_cc, pcbddc->local_primal_ref_node, PETSC_COPY_VALUES, &subset_n));
9279: PetscCall(ISLocalToGlobalMappingApplyIS(pcis->mapping, subset_n, &subset));
9280: PetscCall(ISDestroy(&subset_n));
9281: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)pc->pmat), pcbddc->local_primal_size_cc, pcbddc->local_primal_ref_mult, PETSC_COPY_VALUES, &subset_mult));
9282: PetscCall(ISRenumber(subset, subset_mult, &coarse_size, &subset_n));
9283: PetscCall(ISDestroy(&subset));
9284: PetscCall(ISDestroy(&subset_mult));
9285: PetscCall(ISGetLocalSize(subset_n, &local_size));
9286: PetscCheck(local_size == pcbddc->local_primal_size, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Invalid number of local primal indices computed %" PetscInt_FMT " != %" PetscInt_FMT, local_size, pcbddc->local_primal_size);
9287: PetscCall(PetscMalloc1(local_size, &local_primal_indices));
9288: PetscCall(ISGetIndices(subset_n, &t_local_primal_indices));
9289: PetscCall(PetscArraycpy(local_primal_indices, t_local_primal_indices, local_size));
9290: PetscCall(ISRestoreIndices(subset_n, &t_local_primal_indices));
9291: PetscCall(ISDestroy(&subset_n));
9293: if (pcbddc->dbg_flag) {
9294: PetscCall(PetscViewerFlush(pcbddc->dbg_viewer));
9295: PetscCall(PetscViewerASCIIPrintf(pcbddc->dbg_viewer, "--------------------------------------------------\n"));
9296: PetscCall(PetscViewerASCIIPrintf(pcbddc->dbg_viewer, "Size of coarse problem is %" PetscInt_FMT "\n", coarse_size));
9297: PetscCall(PetscViewerFlush(pcbddc->dbg_viewer));
9298: }
9300: /* get back data */
9301: *coarse_size_n = coarse_size;
9302: *local_primal_indices_n = local_primal_indices;
9303: PetscFunctionReturn(PETSC_SUCCESS);
9304: }
9306: PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx, Vec gwork, Vec lwork, IS globalis, IS *localis)
9307: {
9308: IS localis_t;
9309: PetscInt i, lsize, *idxs, n;
9310: PetscScalar *vals;
9312: PetscFunctionBegin;
9313: /* get indices in local ordering exploiting local to global map */
9314: PetscCall(ISGetLocalSize(globalis, &lsize));
9315: PetscCall(PetscMalloc1(lsize, &vals));
9316: for (i = 0; i < lsize; i++) vals[i] = 1.0;
9317: PetscCall(ISGetIndices(globalis, (const PetscInt **)&idxs));
9318: PetscCall(VecSet(gwork, 0.0));
9319: PetscCall(VecSet(lwork, 0.0));
9320: if (idxs) { /* multilevel guard */
9321: PetscCall(VecSetOption(gwork, VEC_IGNORE_NEGATIVE_INDICES, PETSC_TRUE));
9322: PetscCall(VecSetValues(gwork, lsize, idxs, vals, INSERT_VALUES));
9323: }
9324: PetscCall(VecAssemblyBegin(gwork));
9325: PetscCall(ISRestoreIndices(globalis, (const PetscInt **)&idxs));
9326: PetscCall(PetscFree(vals));
9327: PetscCall(VecAssemblyEnd(gwork));
9328: /* now compute set in local ordering */
9329: PetscCall(VecScatterBegin(g2l_ctx, gwork, lwork, INSERT_VALUES, SCATTER_FORWARD));
9330: PetscCall(VecScatterEnd(g2l_ctx, gwork, lwork, INSERT_VALUES, SCATTER_FORWARD));
9331: PetscCall(VecGetArrayRead(lwork, (const PetscScalar **)&vals));
9332: PetscCall(VecGetSize(lwork, &n));
9333: for (i = 0, lsize = 0; i < n; i++) {
9334: if (PetscRealPart(vals[i]) > 0.5) lsize++;
9335: }
9336: PetscCall(PetscMalloc1(lsize, &idxs));
9337: for (i = 0, lsize = 0; i < n; i++) {
9338: if (PetscRealPart(vals[i]) > 0.5) idxs[lsize++] = i;
9339: }
9340: PetscCall(VecRestoreArrayRead(lwork, (const PetscScalar **)&vals));
9341: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)gwork), lsize, idxs, PETSC_OWN_POINTER, &localis_t));
9342: *localis = localis_t;
9343: PetscFunctionReturn(PETSC_SUCCESS);
9344: }
9346: PetscErrorCode PCBDDCComputeFakeChange(PC pc, PetscBool constraints, PCBDDCGraph graph, PCBDDCSubSchurs schurs, Mat *change, IS *change_primal, IS *change_primal_mult, PetscBool *change_with_qr)
9347: {
9348: PC_IS *pcis = (PC_IS *)pc->data;
9349: PC_BDDC *pcbddc = (PC_BDDC *)pc->data;
9350: PC_IS *pcisf;
9351: PC_BDDC *pcbddcf;
9352: PC pcf;
9354: PetscFunctionBegin;
9355: PetscCall(PCCreate(PetscObjectComm((PetscObject)pc), &pcf));
9356: PetscCall(PCSetOperators(pcf, pc->mat, pc->pmat));
9357: PetscCall(PCSetType(pcf, PCBDDC));
9359: pcisf = (PC_IS *)pcf->data;
9360: pcbddcf = (PC_BDDC *)pcf->data;
9362: pcisf->is_B_local = pcis->is_B_local;
9363: pcisf->vec1_N = pcis->vec1_N;
9364: pcisf->BtoNmap = pcis->BtoNmap;
9365: pcisf->n = pcis->n;
9366: pcisf->n_B = pcis->n_B;
9368: PetscCall(PetscFree(pcbddcf->mat_graph));
9369: PetscCall(PetscFree(pcbddcf->sub_schurs));
9370: pcbddcf->mat_graph = graph ? graph : pcbddc->mat_graph;
9371: pcbddcf->sub_schurs = schurs;
9372: pcbddcf->adaptive_selection = schurs ? PETSC_TRUE : PETSC_FALSE;
9373: pcbddcf->adaptive_threshold[0] = pcbddc->adaptive_threshold[0];
9374: pcbddcf->adaptive_threshold[1] = pcbddc->adaptive_threshold[1];
9375: pcbddcf->adaptive_nmin = pcbddc->adaptive_nmin;
9376: pcbddcf->adaptive_nmax = pcbddc->adaptive_nmax;
9377: pcbddcf->use_faces = PETSC_TRUE;
9378: pcbddcf->use_change_of_basis = (PetscBool)!constraints;
9379: pcbddcf->use_change_on_faces = (PetscBool)!constraints;
9380: pcbddcf->use_qr_single = (PetscBool)!constraints;
9381: pcbddcf->fake_change = PETSC_TRUE;
9382: pcbddcf->dbg_flag = pcbddc->dbg_flag;
9384: PetscCall(PCBDDCAdaptiveSelection(pcf));
9385: PetscCall(PCBDDCConstraintsSetUp(pcf));
9387: *change = pcbddcf->ConstraintMatrix;
9388: if (change_primal) PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)pc->pmat), pcbddcf->local_primal_size_cc, pcbddcf->local_primal_ref_node, PETSC_COPY_VALUES, change_primal));
9389: if (change_primal_mult) PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)pc->pmat), pcbddcf->local_primal_size_cc, pcbddcf->local_primal_ref_mult, PETSC_COPY_VALUES, change_primal_mult));
9390: if (change_with_qr) *change_with_qr = pcbddcf->use_qr_single;
9392: if (schurs) pcbddcf->sub_schurs = NULL;
9393: pcbddcf->ConstraintMatrix = NULL;
9394: pcbddcf->mat_graph = NULL;
9395: pcisf->is_B_local = NULL;
9396: pcisf->vec1_N = NULL;
9397: pcisf->BtoNmap = NULL;
9398: PetscCall(PCDestroy(&pcf));
9399: PetscFunctionReturn(PETSC_SUCCESS);
9400: }
9402: PetscErrorCode PCBDDCSetUpSubSchurs(PC pc)
9403: {
9404: PC_IS *pcis = (PC_IS *)pc->data;
9405: PC_BDDC *pcbddc = (PC_BDDC *)pc->data;
9406: PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs;
9407: Mat S_j;
9408: PetscInt *used_xadj, *used_adjncy;
9409: PetscBool free_used_adj;
9411: PetscFunctionBegin;
9412: PetscCall(PetscLogEventBegin(PC_BDDC_Schurs[pcbddc->current_level], pc, 0, 0, 0));
9413: /* decide the adjacency to be used for determining internal problems for local schur on subsets */
9414: free_used_adj = PETSC_FALSE;
9415: if (pcbddc->sub_schurs_layers == -1) {
9416: used_xadj = NULL;
9417: used_adjncy = NULL;
9418: } else {
9419: if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) {
9420: used_xadj = pcbddc->mat_graph->xadj;
9421: used_adjncy = pcbddc->mat_graph->adjncy;
9422: } else if (pcbddc->computed_rowadj) {
9423: used_xadj = pcbddc->mat_graph->xadj;
9424: used_adjncy = pcbddc->mat_graph->adjncy;
9425: } else {
9426: PetscBool flg_row = PETSC_FALSE;
9427: const PetscInt *xadj, *adjncy;
9428: PetscInt nvtxs;
9430: PetscCall(MatGetRowIJ(pcbddc->local_mat, 0, PETSC_TRUE, PETSC_FALSE, &nvtxs, &xadj, &adjncy, &flg_row));
9431: if (flg_row) {
9432: PetscCall(PetscMalloc2(nvtxs + 1, &used_xadj, xadj[nvtxs], &used_adjncy));
9433: PetscCall(PetscArraycpy(used_xadj, xadj, nvtxs + 1));
9434: PetscCall(PetscArraycpy(used_adjncy, adjncy, xadj[nvtxs]));
9435: free_used_adj = PETSC_TRUE;
9436: } else {
9437: pcbddc->sub_schurs_layers = -1;
9438: used_xadj = NULL;
9439: used_adjncy = NULL;
9440: }
9441: PetscCall(MatRestoreRowIJ(pcbddc->local_mat, 0, PETSC_TRUE, PETSC_FALSE, &nvtxs, &xadj, &adjncy, &flg_row));
9442: }
9443: }
9445: /* setup sub_schurs data */
9446: PetscCall(MatCreateSchurComplement(pcis->A_II, pcis->pA_II, pcis->A_IB, pcis->A_BI, pcis->A_BB, &S_j));
9447: if (!sub_schurs->schur_explicit) {
9448: /* pcbddc->ksp_D up to date only if not using MatFactor with Schur complement support */
9449: PetscCall(MatSchurComplementSetKSP(S_j, pcbddc->ksp_D));
9450: PetscCall(PCBDDCSubSchursSetUp(sub_schurs, NULL, S_j, PETSC_FALSE, used_xadj, used_adjncy, pcbddc->sub_schurs_layers, NULL, pcbddc->adaptive_selection, PETSC_FALSE, PETSC_FALSE, 0, NULL, NULL, NULL, NULL));
9451: } else {
9452: Mat change = NULL;
9453: Vec scaling = NULL;
9454: IS change_primal = NULL, iP;
9455: PetscInt benign_n;
9456: PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis;
9457: PetscBool need_change = PETSC_FALSE;
9458: PetscBool discrete_harmonic = PETSC_FALSE;
9460: if (!pcbddc->use_vertices && reuse_solvers) {
9461: PetscInt n_vertices;
9463: PetscCall(ISGetLocalSize(sub_schurs->is_vertices, &n_vertices));
9464: reuse_solvers = (PetscBool)!n_vertices;
9465: }
9466: if (!pcbddc->benign_change_explicit) {
9467: benign_n = pcbddc->benign_n;
9468: } else {
9469: benign_n = 0;
9470: }
9471: /* sub_schurs->change is a local object; instead, PCBDDCConstraintsSetUp and the quantities used in the test below are logically collective on pc.
9472: We need a global reduction to avoid possible deadlocks.
9473: We assume that sub_schurs->change is created once, and then reused for different solves, unless the topography has been recomputed */
9474: if (pcbddc->adaptive_userdefined || (pcbddc->deluxe_zerorows && !pcbddc->use_change_of_basis)) {
9475: PetscBool have_loc_change = (PetscBool)(!!sub_schurs->change);
9476: PetscCallMPI(MPIU_Allreduce(&have_loc_change, &need_change, 1, MPI_C_BOOL, MPI_LOR, PetscObjectComm((PetscObject)pc)));
9477: need_change = (PetscBool)(!need_change);
9478: }
9479: /* If the user defines additional constraints, we import them here */
9480: if (need_change) {
9481: PetscCheck(!pcbddc->sub_schurs_rebuild, PETSC_COMM_SELF, PETSC_ERR_SUP, "Cannot compute change of basis with a different graph");
9482: PetscCall(PCBDDCComputeFakeChange(pc, PETSC_FALSE, NULL, NULL, &change, &change_primal, NULL, &sub_schurs->change_with_qr));
9483: }
9484: if (!pcbddc->use_deluxe_scaling) scaling = pcis->D;
9486: PetscCall(PetscObjectQuery((PetscObject)pc, "__KSPFETIDP_iP", (PetscObject *)&iP));
9487: if (iP) {
9488: PetscOptionsBegin(PetscObjectComm((PetscObject)iP), sub_schurs->prefix, "BDDC sub_schurs options", "PC");
9489: PetscCall(PetscOptionsBool("-sub_schurs_discrete_harmonic", NULL, NULL, discrete_harmonic, &discrete_harmonic, NULL));
9490: PetscOptionsEnd();
9491: }
9492: if (discrete_harmonic) {
9493: Mat A;
9494: PetscCall(MatDuplicate(pcbddc->local_mat, MAT_COPY_VALUES, &A));
9495: PetscCall(MatZeroRowsColumnsIS(A, iP, 1.0, NULL, NULL));
9496: PetscCall(PetscObjectCompose((PetscObject)A, "__KSPFETIDP_iP", (PetscObject)iP));
9497: PetscCall(PCBDDCSubSchursSetUp(sub_schurs, A, S_j, pcbddc->sub_schurs_exact_schur, used_xadj, used_adjncy, pcbddc->sub_schurs_layers, scaling, pcbddc->adaptive_selection, reuse_solvers, pcbddc->benign_saddle_point, benign_n, pcbddc->benign_p0_lidx,
9498: pcbddc->benign_zerodiag_subs, change, change_primal));
9499: PetscCall(MatDestroy(&A));
9500: } else {
9501: PetscCall(PCBDDCSubSchursSetUp(sub_schurs, pcbddc->local_mat, S_j, pcbddc->sub_schurs_exact_schur, used_xadj, used_adjncy, pcbddc->sub_schurs_layers, scaling, pcbddc->adaptive_selection, reuse_solvers, pcbddc->benign_saddle_point, benign_n,
9502: pcbddc->benign_p0_lidx, pcbddc->benign_zerodiag_subs, change, change_primal));
9503: }
9504: PetscCall(MatDestroy(&change));
9505: PetscCall(ISDestroy(&change_primal));
9506: }
9507: PetscCall(MatDestroy(&S_j));
9509: /* free adjacency */
9510: if (free_used_adj) PetscCall(PetscFree2(used_xadj, used_adjncy));
9511: PetscCall(PetscLogEventEnd(PC_BDDC_Schurs[pcbddc->current_level], pc, 0, 0, 0));
9512: PetscFunctionReturn(PETSC_SUCCESS);
9513: }
9515: PetscErrorCode PCBDDCInitSubSchurs(PC pc)
9516: {
9517: PC_IS *pcis = (PC_IS *)pc->data;
9518: PC_BDDC *pcbddc = (PC_BDDC *)pc->data;
9519: PCBDDCGraph graph;
9521: PetscFunctionBegin;
9522: /* attach interface graph for determining subsets */
9523: if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */
9524: IS verticesIS, verticescomm;
9525: PetscInt vsize, *idxs;
9527: PetscCall(PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph, NULL, NULL, NULL, NULL, &verticesIS));
9528: PetscCall(ISGetSize(verticesIS, &vsize));
9529: PetscCall(ISGetIndices(verticesIS, (const PetscInt **)&idxs));
9530: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)pc), vsize, idxs, PETSC_COPY_VALUES, &verticescomm));
9531: PetscCall(ISRestoreIndices(verticesIS, (const PetscInt **)&idxs));
9532: PetscCall(PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph, NULL, NULL, NULL, NULL, &verticesIS));
9533: PetscCall(PCBDDCGraphCreate(&graph));
9534: PetscCall(PCBDDCGraphInit(graph, pcbddc->mat_graph->l2gmap, pcbddc->mat_graph->nvtxs_global, pcbddc->graphmaxcount));
9535: PetscCall(PCBDDCGraphSetUp(graph, pcbddc->mat_graph->custom_minimal_size, NULL, pcbddc->DirichletBoundariesLocal, 0, NULL, verticescomm));
9536: PetscCall(ISDestroy(&verticescomm));
9537: PetscCall(PCBDDCGraphComputeConnectedComponents(graph));
9538: } else {
9539: graph = pcbddc->mat_graph;
9540: }
9541: /* print some info */
9542: if (pcbddc->dbg_flag && !pcbddc->sub_schurs_rebuild) {
9543: IS vertices;
9544: PetscInt nv, nedges, nfaces;
9545: PetscCall(PCBDDCGraphASCIIView(graph, pcbddc->dbg_flag, pcbddc->dbg_viewer));
9546: PetscCall(PCBDDCGraphGetCandidatesIS(graph, &nfaces, NULL, &nedges, NULL, &vertices));
9547: PetscCall(ISGetSize(vertices, &nv));
9548: PetscCall(PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer));
9549: PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, "--------------------------------------------------------------\n"));
9550: PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, "Subdomain %04d got %02" PetscInt_FMT " local candidate vertices (%d)\n", PetscGlobalRank, nv, pcbddc->use_vertices));
9551: PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, "Subdomain %04d got %02" PetscInt_FMT " local candidate edges (%d)\n", PetscGlobalRank, nedges, pcbddc->use_edges));
9552: PetscCall(PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer, "Subdomain %04d got %02" PetscInt_FMT " local candidate faces (%d)\n", PetscGlobalRank, nfaces, pcbddc->use_faces));
9553: PetscCall(PetscViewerFlush(pcbddc->dbg_viewer));
9554: PetscCall(PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer));
9555: PetscCall(PCBDDCGraphRestoreCandidatesIS(graph, &nfaces, NULL, &nedges, NULL, &vertices));
9556: }
9558: /* sub_schurs init */
9559: if (!pcbddc->sub_schurs) PetscCall(PCBDDCSubSchursCreate(&pcbddc->sub_schurs));
9560: PetscCall(PCBDDCSubSchursInit(pcbddc->sub_schurs, ((PetscObject)pc)->prefix, pcis->is_I_local, pcis->is_B_local, graph, pcis->BtoNmap, pcbddc->sub_schurs_rebuild, PETSC_FALSE));
9562: /* free graph struct */
9563: if (pcbddc->sub_schurs_rebuild) PetscCall(PCBDDCGraphDestroy(&graph));
9564: PetscFunctionReturn(PETSC_SUCCESS);
9565: }
9567: static PetscErrorCode PCBDDCViewGlobalIS(PC pc, IS is, PetscViewer viewer)
9568: {
9569: Mat_IS *matis = (Mat_IS *)pc->pmat->data;
9570: PetscInt n = pc->pmat->rmap->n, ln, ni, st;
9571: const PetscInt *idxs;
9572: IS gis;
9574: PetscFunctionBegin;
9575: if (!is) PetscFunctionReturn(PETSC_SUCCESS);
9576: PetscCall(MatGetOwnershipRange(pc->pmat, &st, NULL));
9577: PetscCall(MatGetLocalSize(matis->A, NULL, &ln));
9578: PetscCall(PetscArrayzero(matis->sf_leafdata, ln));
9579: PetscCall(PetscArrayzero(matis->sf_rootdata, n));
9580: PetscCall(ISGetLocalSize(is, &ni));
9581: PetscCall(ISGetIndices(is, &idxs));
9582: for (PetscInt i = 0; i < ni; i++) {
9583: if (idxs[i] < 0 || idxs[i] >= ln) continue;
9584: matis->sf_leafdata[idxs[i]] = 1;
9585: }
9586: PetscCall(ISRestoreIndices(is, &idxs));
9587: PetscCall(PetscSFReduceBegin(matis->sf, MPIU_INT, matis->sf_leafdata, matis->sf_rootdata, MPI_SUM));
9588: PetscCall(PetscSFReduceEnd(matis->sf, MPIU_INT, matis->sf_leafdata, matis->sf_rootdata, MPI_SUM));
9589: ln = 0;
9590: for (PetscInt i = 0; i < n; i++) {
9591: if (matis->sf_rootdata[i]) matis->sf_rootdata[ln++] = i + st;
9592: }
9593: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)pc), ln, matis->sf_rootdata, PETSC_USE_POINTER, &gis));
9594: PetscCall(ISView(gis, viewer));
9595: PetscCall(ISDestroy(&gis));
9596: PetscFunctionReturn(PETSC_SUCCESS);
9597: }
9599: PetscErrorCode PCBDDCLoadOrViewCustomization(PC pc, PetscBool load, const char *outfile)
9600: {
9601: PetscInt header[11];
9602: PC_BDDC *pcbddc = (PC_BDDC *)pc->data;
9603: PetscViewer viewer;
9604: MPI_Comm comm = PetscObjectComm((PetscObject)pc);
9606: PetscFunctionBegin;
9607: PetscCall(PetscViewerBinaryOpen(comm, outfile ? outfile : "bddc_dump.dat", load ? FILE_MODE_READ : FILE_MODE_WRITE, &viewer));
9608: if (load) {
9609: IS is;
9610: Mat A;
9612: PetscCall(PetscViewerBinaryRead(viewer, header, PETSC_STATIC_ARRAY_LENGTH(header), NULL, PETSC_INT));
9613: PetscCheck(header[0] == 0 || header[0] == 1, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Not a BDDC dump next in file");
9614: PetscCheck(header[1] == 0 || header[1] == 1, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Not a BDDC dump next in file");
9615: PetscCheck(header[2] >= 0, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Not a BDDC dump next in file");
9616: PetscCheck(header[3] == 0 || header[3] == 1, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Not a BDDC dump next in file");
9617: PetscCheck(header[4] == 0 || header[4] == 1, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Not a BDDC dump next in file");
9618: PetscCheck(header[5] >= 0, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Not a BDDC dump next in file");
9619: PetscCheck(header[7] == 0 || header[7] == 1, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Not a BDDC dump next in file");
9620: PetscCheck(header[8] == 0 || header[8] == 1, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Not a BDDC dump next in file");
9621: PetscCheck(header[9] == 0 || header[9] == 1, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Not a BDDC dump next in file");
9622: PetscCheck(header[10] == 0 || header[10] == 1, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Not a BDDC dump next in file");
9623: if (header[0]) {
9624: PetscCall(ISCreate(comm, &is));
9625: PetscCall(ISLoad(is, viewer));
9626: PetscCall(PCBDDCSetDirichletBoundaries(pc, is));
9627: PetscCall(ISDestroy(&is));
9628: }
9629: if (header[1]) {
9630: PetscCall(ISCreate(comm, &is));
9631: PetscCall(ISLoad(is, viewer));
9632: PetscCall(PCBDDCSetNeumannBoundaries(pc, is));
9633: PetscCall(ISDestroy(&is));
9634: }
9635: if (header[2]) {
9636: IS *isarray;
9638: PetscCall(PetscMalloc1(header[2], &isarray));
9639: for (PetscInt i = 0; i < header[2]; i++) {
9640: PetscCall(ISCreate(comm, &isarray[i]));
9641: PetscCall(ISLoad(isarray[i], viewer));
9642: }
9643: PetscCall(PCBDDCSetDofsSplitting(pc, header[2], isarray));
9644: for (PetscInt i = 0; i < header[2]; i++) PetscCall(ISDestroy(&isarray[i]));
9645: PetscCall(PetscFree(isarray));
9646: }
9647: if (header[3]) {
9648: PetscCall(ISCreate(comm, &is));
9649: PetscCall(ISLoad(is, viewer));
9650: PetscCall(PCBDDCSetPrimalVerticesIS(pc, is));
9651: PetscCall(ISDestroy(&is));
9652: }
9653: if (header[4]) {
9654: PetscCall(MatCreate(comm, &A));
9655: PetscCall(MatSetType(A, MATAIJ));
9656: PetscCall(MatLoad(A, viewer));
9657: PetscCall(PCBDDCSetDiscreteGradient(pc, A, header[5], header[6], (PetscBool)header[7], (PetscBool)header[8]));
9658: PetscCall(MatDestroy(&A));
9659: }
9660: if (header[9]) {
9661: PetscCall(MatCreate(comm, &A));
9662: PetscCall(MatSetType(A, MATIS));
9663: PetscCall(MatLoad(A, viewer));
9664: PetscCall(PCBDDCSetDivergenceMat(pc, A, (PetscBool)header[10], NULL));
9665: PetscCall(MatDestroy(&A));
9666: }
9667: } else {
9668: header[0] = (PetscInt)!!pcbddc->DirichletBoundariesLocal;
9669: header[1] = (PetscInt)!!pcbddc->NeumannBoundariesLocal;
9670: header[2] = pcbddc->n_ISForDofsLocal;
9671: header[3] = (PetscInt)!!pcbddc->user_primal_vertices_local;
9672: header[4] = (PetscInt)!!pcbddc->discretegradient;
9673: header[5] = pcbddc->nedorder;
9674: header[6] = pcbddc->nedfield;
9675: header[7] = (PetscInt)pcbddc->nedglobal;
9676: header[8] = (PetscInt)pcbddc->conforming;
9677: header[9] = (PetscInt)!!pcbddc->divudotp;
9678: header[10] = (PetscInt)pcbddc->divudotp_trans;
9679: if (header[4]) header[3] = 0;
9681: PetscCall(PetscViewerBinaryWrite(viewer, header, PETSC_STATIC_ARRAY_LENGTH(header), PETSC_INT));
9682: PetscCall(PCBDDCViewGlobalIS(pc, pcbddc->DirichletBoundariesLocal, viewer));
9683: PetscCall(PCBDDCViewGlobalIS(pc, pcbddc->NeumannBoundariesLocal, viewer));
9684: for (PetscInt i = 0; i < header[2]; i++) PetscCall(PCBDDCViewGlobalIS(pc, pcbddc->ISForDofsLocal[i], viewer));
9685: if (header[3]) PetscCall(PCBDDCViewGlobalIS(pc, pcbddc->user_primal_vertices_local, viewer));
9686: if (header[4]) PetscCall(MatView(pcbddc->discretegradient, viewer));
9687: if (header[9]) PetscCall(MatView(pcbddc->divudotp, viewer));
9688: }
9689: PetscCall(PetscViewerDestroy(&viewer));
9690: PetscFunctionReturn(PETSC_SUCCESS);
9691: }
9693: #include <../src/mat/impls/aij/mpi/mpiaij.h>
9694: static PetscErrorCode MatMPIAIJRestrict(Mat A, MPI_Comm ccomm, Mat *B)
9695: {
9696: Mat At;
9697: IS rows;
9698: PetscInt rst, ren;
9699: PetscLayout rmap;
9701: PetscFunctionBegin;
9702: rst = ren = 0;
9703: if (ccomm != MPI_COMM_NULL) {
9704: PetscCall(PetscLayoutCreate(ccomm, &rmap));
9705: PetscCall(PetscLayoutSetSize(rmap, A->rmap->N));
9706: PetscCall(PetscLayoutSetBlockSize(rmap, 1));
9707: PetscCall(PetscLayoutSetUp(rmap));
9708: PetscCall(PetscLayoutGetRange(rmap, &rst, &ren));
9709: }
9710: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)A), ren - rst, rst, 1, &rows));
9711: PetscCall(MatCreateSubMatrix(A, rows, NULL, MAT_INITIAL_MATRIX, &At));
9712: PetscCall(ISDestroy(&rows));
9714: if (ccomm != MPI_COMM_NULL) {
9715: Mat_MPIAIJ *a, *b;
9716: IS from, to;
9717: Vec gvec;
9718: PetscInt lsize;
9720: PetscCall(MatCreate(ccomm, B));
9721: PetscCall(MatSetSizes(*B, ren - rst, PETSC_DECIDE, PETSC_DECIDE, At->cmap->N));
9722: PetscCall(MatSetType(*B, MATAIJ));
9723: PetscCall(PetscLayoutDestroy(&(*B)->rmap));
9724: PetscCall(PetscLayoutSetUp((*B)->cmap));
9725: a = (Mat_MPIAIJ *)At->data;
9726: b = (Mat_MPIAIJ *)(*B)->data;
9727: PetscCallMPI(MPI_Comm_size(ccomm, &b->size));
9728: PetscCallMPI(MPI_Comm_rank(ccomm, &b->rank));
9729: PetscCall(PetscObjectReference((PetscObject)a->A));
9730: PetscCall(PetscObjectReference((PetscObject)a->B));
9731: b->A = a->A;
9732: b->B = a->B;
9734: b->donotstash = a->donotstash;
9735: b->roworiented = a->roworiented;
9736: b->rowindices = NULL;
9737: b->rowvalues = NULL;
9738: b->getrowactive = PETSC_FALSE;
9740: (*B)->rmap = rmap;
9741: (*B)->factortype = A->factortype;
9742: (*B)->assembled = PETSC_TRUE;
9743: (*B)->insertmode = NOT_SET_VALUES;
9744: (*B)->preallocated = PETSC_TRUE;
9746: if (a->colmap) {
9747: #if defined(PETSC_USE_CTABLE)
9748: PetscCall(PetscHMapIDuplicate(a->colmap, &b->colmap));
9749: #else
9750: PetscCall(PetscMalloc1(At->cmap->N, &b->colmap));
9751: PetscCall(PetscArraycpy(b->colmap, a->colmap, At->cmap->N));
9752: #endif
9753: } else b->colmap = NULL;
9754: if (a->garray) {
9755: PetscInt len;
9756: len = a->B->cmap->n;
9757: PetscCall(PetscMalloc1(len + 1, &b->garray));
9758: if (len) PetscCall(PetscArraycpy(b->garray, a->garray, len));
9759: } else b->garray = NULL;
9761: PetscCall(PetscObjectReference((PetscObject)a->lvec));
9762: b->lvec = a->lvec;
9764: /* cannot use VecScatterCopy */
9765: PetscCall(VecGetLocalSize(b->lvec, &lsize));
9766: PetscCall(ISCreateGeneral(ccomm, lsize, b->garray, PETSC_USE_POINTER, &from));
9767: PetscCall(ISCreateStride(PETSC_COMM_SELF, lsize, 0, 1, &to));
9768: PetscCall(MatCreateVecs(*B, &gvec, NULL));
9769: PetscCall(VecScatterCreate(gvec, from, b->lvec, to, &b->Mvctx));
9770: PetscCall(ISDestroy(&from));
9771: PetscCall(ISDestroy(&to));
9772: PetscCall(VecDestroy(&gvec));
9773: }
9774: PetscCall(MatDestroy(&At));
9775: PetscFunctionReturn(PETSC_SUCCESS);
9776: }
9778: /* same as MatCreateSubMatrix(A, rows, NULL,...) but allows repeated rows */
9779: static PetscErrorCode MatAIJExtractRows(Mat A, IS rows, Mat *sA)
9780: {
9781: PetscBool isaij;
9782: MPI_Comm comm;
9784: PetscFunctionBegin;
9785: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
9786: PetscCall(PetscObjectBaseTypeCompareAny((PetscObject)A, &isaij, MATSEQAIJ, MATMPIAIJ, ""));
9787: PetscCheck(isaij, comm, PETSC_ERR_SUP, "Not implemented");
9788: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATSEQAIJ, &isaij));
9789: if (isaij) { /* SeqAIJ supports repeated rows */
9790: PetscCall(MatCreateSubMatrix(A, rows, NULL, MAT_INITIAL_MATRIX, sA));
9791: } else {
9792: Mat A_loc;
9793: Mat_SeqAIJ *da;
9794: PetscSF sf;
9795: PetscInt ni, *di, *dj, m = A->rmap->n, c, *ldata, *rdata;
9796: PetscScalar *daa;
9797: const PetscInt *idxs;
9798: const PetscSFNode *iremotes;
9799: PetscSFNode *remotes;
9801: /* SF for incoming rows */
9802: PetscCall(PetscSFCreate(comm, &sf));
9803: PetscCall(ISGetLocalSize(rows, &ni));
9804: PetscCall(ISGetIndices(rows, &idxs));
9805: PetscCall(PetscSFSetGraphLayout(sf, A->rmap, ni, NULL, PETSC_USE_POINTER, idxs));
9806: PetscCall(ISRestoreIndices(rows, &idxs));
9808: PetscCall(MatMPIAIJGetLocalMat(A, MAT_INITIAL_MATRIX, &A_loc));
9809: da = (Mat_SeqAIJ *)A_loc->data;
9810: PetscCall(PetscMalloc2(2 * ni, &ldata, 2 * m, &rdata));
9811: for (PetscInt i = 0; i < m; i++) {
9812: rdata[2 * i + 0] = da->i[i + 1] - da->i[i];
9813: rdata[2 * i + 1] = da->i[i];
9814: }
9815: PetscCall(PetscSFBcastBegin(sf, MPIU_2INT, rdata, ldata, MPI_REPLACE));
9816: PetscCall(PetscSFBcastEnd(sf, MPIU_2INT, rdata, ldata, MPI_REPLACE));
9817: PetscCall(PetscMalloc1(ni + 1, &di));
9818: di[0] = 0;
9819: for (PetscInt i = 0; i < ni; i++) di[i + 1] = di[i] + ldata[2 * i + 0];
9820: PetscCall(PetscMalloc1(di[ni], &dj));
9821: PetscCall(PetscMalloc1(di[ni], &daa));
9822: PetscCall(PetscMalloc1(di[ni], &remotes));
9824: PetscCall(PetscSFGetGraph(sf, NULL, NULL, NULL, &iremotes));
9826: /* SF graph for nonzeros */
9827: c = 0;
9828: for (PetscInt i = 0; i < ni; i++) {
9829: const PetscInt rank = iremotes[i].rank;
9830: const PetscInt rsize = ldata[2 * i];
9831: for (PetscInt j = 0; j < rsize; j++) {
9832: remotes[c].rank = rank;
9833: remotes[c].index = ldata[2 * i + 1] + j;
9834: c++;
9835: }
9836: }
9837: PetscCheck(c == di[ni], PETSC_COMM_SELF, PETSC_ERR_PLIB, "Invalid number of local nonzeros %" PetscInt_FMT " != %" PetscInt_FMT, c, di[ni]);
9838: PetscCall(PetscSFSetGraph(sf, da->i[m], di[ni], NULL, PETSC_USE_POINTER, remotes, PETSC_USE_POINTER));
9839: PetscCall(PetscSFBcastBegin(sf, MPIU_INT, da->j, dj, MPI_REPLACE));
9840: PetscCall(PetscSFBcastEnd(sf, MPIU_INT, da->j, dj, MPI_REPLACE));
9841: PetscCall(PetscSFBcastBegin(sf, MPIU_SCALAR, da->a, daa, MPI_REPLACE));
9842: PetscCall(PetscSFBcastEnd(sf, MPIU_SCALAR, da->a, daa, MPI_REPLACE));
9844: PetscCall(MatCreateMPIAIJWithArrays(comm, ni, A->cmap->n, PETSC_DECIDE, A->cmap->N, di, dj, daa, sA));
9845: PetscCall(MatDestroy(&A_loc));
9846: PetscCall(PetscSFDestroy(&sf));
9847: PetscCall(PetscFree(di));
9848: PetscCall(PetscFree(dj));
9849: PetscCall(PetscFree(daa));
9850: PetscCall(PetscFree(remotes));
9851: PetscCall(PetscFree2(ldata, rdata));
9852: }
9853: PetscFunctionReturn(PETSC_SUCCESS);
9854: }