Actual source code: asm.c
1: /*
2: This file defines an additive Schwarz preconditioner for any Mat implementation.
4: Note that each processor may have any number of subdomains. But in order to
5: deal easily with the VecScatter(), we treat each processor as if it has the
6: same number of subdomains.
8: n - total number of true subdomains on all processors
9: n_local_true - actual number of subdomains on this processor
10: n_local = maximum over all processors of n_local_true
11: */
13: #include <petsc/private/pcasmimpl.h>
14: #include <petsc/private/matimpl.h>
16: static PetscErrorCode PCView_ASM(PC pc, PetscViewer viewer)
17: {
18: PC_ASM *osm = (PC_ASM *)pc->data;
19: PetscMPIInt rank;
20: PetscInt i, bsz;
21: PetscBool iascii, isstring;
22: PetscViewer sviewer;
23: PetscViewerFormat format;
24: const char *prefix;
26: PetscFunctionBegin;
27: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &iascii));
28: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERSTRING, &isstring));
29: if (iascii) {
30: char overlaps[256] = "user-defined overlap", blocks[256] = "total subdomain blocks not yet set";
31: if (osm->overlap >= 0) PetscCall(PetscSNPrintf(overlaps, sizeof(overlaps), "amount of overlap = %" PetscInt_FMT, osm->overlap));
32: if (osm->n > 0) PetscCall(PetscSNPrintf(blocks, sizeof(blocks), "total subdomain blocks = %" PetscInt_FMT, osm->n));
33: PetscCall(PetscViewerASCIIPrintf(viewer, " %s, %s\n", blocks, overlaps));
34: PetscCall(PetscViewerASCIIPrintf(viewer, " restriction/interpolation type - %s\n", PCASMTypes[osm->type]));
35: if (osm->dm_subdomains) PetscCall(PetscViewerASCIIPrintf(viewer, " Additive Schwarz: using DM to define subdomains\n"));
36: if (osm->loctype != PC_COMPOSITE_ADDITIVE) PetscCall(PetscViewerASCIIPrintf(viewer, " Additive Schwarz: local solve composition type - %s\n", PCCompositeTypes[osm->loctype]));
37: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)pc), &rank));
38: PetscCall(PetscViewerGetFormat(viewer, &format));
39: if (format != PETSC_VIEWER_ASCII_INFO_DETAIL) {
40: if (osm->ksp) {
41: PetscCall(PetscViewerASCIIPrintf(viewer, " Local solver information for first block is in the following KSP and PC objects on rank 0:\n"));
42: PetscCall(PCGetOptionsPrefix(pc, &prefix));
43: PetscCall(PetscViewerASCIIPrintf(viewer, " Use -%sksp_view ::ascii_info_detail to display information for all blocks\n", prefix ? prefix : ""));
44: PetscCall(PetscViewerGetSubViewer(viewer, PETSC_COMM_SELF, &sviewer));
45: if (rank == 0) {
46: PetscCall(PetscViewerASCIIPushTab(sviewer));
47: PetscCall(KSPView(osm->ksp[0], sviewer));
48: PetscCall(PetscViewerASCIIPopTab(sviewer));
49: }
50: PetscCall(PetscViewerRestoreSubViewer(viewer, PETSC_COMM_SELF, &sviewer));
51: }
52: } else {
53: PetscCall(PetscViewerASCIIPushSynchronized(viewer));
54: PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, " [%d] number of local blocks = %" PetscInt_FMT "\n", (int)rank, osm->n_local_true));
55: PetscCall(PetscViewerFlush(viewer));
56: PetscCall(PetscViewerASCIIPrintf(viewer, " Local solver information for each block is in the following KSP and PC objects:\n"));
57: PetscCall(PetscViewerASCIIPushTab(viewer));
58: PetscCall(PetscViewerASCIIPrintf(viewer, "- - - - - - - - - - - - - - - - - -\n"));
59: PetscCall(PetscViewerGetSubViewer(viewer, PETSC_COMM_SELF, &sviewer));
60: for (i = 0; i < osm->n_local_true; i++) {
61: PetscCall(ISGetLocalSize(osm->is[i], &bsz));
62: PetscCall(PetscViewerASCIIPrintf(sviewer, "[%d] local block number %" PetscInt_FMT ", size = %" PetscInt_FMT "\n", (int)rank, i, bsz));
63: PetscCall(KSPView(osm->ksp[i], sviewer));
64: PetscCall(PetscViewerASCIIPrintf(sviewer, "- - - - - - - - - - - - - - - - - -\n"));
65: }
66: PetscCall(PetscViewerRestoreSubViewer(viewer, PETSC_COMM_SELF, &sviewer));
67: PetscCall(PetscViewerASCIIPopTab(viewer));
68: PetscCall(PetscViewerASCIIPopSynchronized(viewer));
69: }
70: } else if (isstring) {
71: PetscCall(PetscViewerStringSPrintf(viewer, " blocks=%" PetscInt_FMT ", overlap=%" PetscInt_FMT ", type=%s", osm->n, osm->overlap, PCASMTypes[osm->type]));
72: PetscCall(PetscViewerGetSubViewer(viewer, PETSC_COMM_SELF, &sviewer));
73: if (osm->ksp) PetscCall(KSPView(osm->ksp[0], sviewer));
74: PetscCall(PetscViewerRestoreSubViewer(viewer, PETSC_COMM_SELF, &sviewer));
75: }
76: PetscFunctionReturn(PETSC_SUCCESS);
77: }
79: static PetscErrorCode PCASMPrintSubdomains(PC pc)
80: {
81: PC_ASM *osm = (PC_ASM *)pc->data;
82: const char *prefix;
83: char fname[PETSC_MAX_PATH_LEN + 1];
84: PetscViewer viewer, sviewer;
85: char *s;
86: PetscInt i, j, nidx;
87: const PetscInt *idx;
88: PetscMPIInt rank, size;
90: PetscFunctionBegin;
91: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)pc), &size));
92: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)pc), &rank));
93: PetscCall(PCGetOptionsPrefix(pc, &prefix));
94: PetscCall(PetscOptionsGetString(NULL, prefix, "-pc_asm_print_subdomains", fname, sizeof(fname), NULL));
95: if (fname[0] == 0) PetscCall(PetscStrncpy(fname, "stdout", sizeof(fname)));
96: PetscCall(PetscViewerASCIIOpen(PetscObjectComm((PetscObject)pc), fname, &viewer));
97: for (i = 0; i < osm->n_local; i++) {
98: if (i < osm->n_local_true) {
99: PetscCall(ISGetLocalSize(osm->is[i], &nidx));
100: PetscCall(ISGetIndices(osm->is[i], &idx));
101: /* Print to a string viewer; no more than 15 characters per index plus 512 char for the header.*/
102: #define len 16 * (nidx + 1) + 512
103: PetscCall(PetscMalloc1(len, &s));
104: PetscCall(PetscViewerStringOpen(PETSC_COMM_SELF, s, len, &sviewer));
105: #undef len
106: PetscCall(PetscViewerStringSPrintf(sviewer, "[%d:%d] Subdomain %" PetscInt_FMT " with overlap:\n", rank, size, i));
107: for (j = 0; j < nidx; j++) PetscCall(PetscViewerStringSPrintf(sviewer, "%" PetscInt_FMT " ", idx[j]));
108: PetscCall(ISRestoreIndices(osm->is[i], &idx));
109: PetscCall(PetscViewerStringSPrintf(sviewer, "\n"));
110: PetscCall(PetscViewerDestroy(&sviewer));
111: PetscCall(PetscViewerASCIIPushSynchronized(viewer));
112: PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "%s", s));
113: PetscCall(PetscViewerFlush(viewer));
114: PetscCall(PetscViewerASCIIPopSynchronized(viewer));
115: PetscCall(PetscFree(s));
116: if (osm->is_local) {
117: /* Print to a string viewer; no more than 15 characters per index plus 512 char for the header.*/
118: #define len 16 * (nidx + 1) + 512
119: PetscCall(PetscMalloc1(len, &s));
120: PetscCall(PetscViewerStringOpen(PETSC_COMM_SELF, s, len, &sviewer));
121: #undef len
122: PetscCall(PetscViewerStringSPrintf(sviewer, "[%d:%d] Subdomain %" PetscInt_FMT " without overlap:\n", rank, size, i));
123: PetscCall(ISGetLocalSize(osm->is_local[i], &nidx));
124: PetscCall(ISGetIndices(osm->is_local[i], &idx));
125: for (j = 0; j < nidx; j++) PetscCall(PetscViewerStringSPrintf(sviewer, "%" PetscInt_FMT " ", idx[j]));
126: PetscCall(ISRestoreIndices(osm->is_local[i], &idx));
127: PetscCall(PetscViewerStringSPrintf(sviewer, "\n"));
128: PetscCall(PetscViewerDestroy(&sviewer));
129: PetscCall(PetscViewerASCIIPushSynchronized(viewer));
130: PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "%s", s));
131: PetscCall(PetscViewerFlush(viewer));
132: PetscCall(PetscViewerASCIIPopSynchronized(viewer));
133: PetscCall(PetscFree(s));
134: }
135: } else {
136: /* Participate in collective viewer calls. */
137: PetscCall(PetscViewerASCIIPushSynchronized(viewer));
138: PetscCall(PetscViewerFlush(viewer));
139: PetscCall(PetscViewerASCIIPopSynchronized(viewer));
140: /* Assume either all ranks have is_local or none do. */
141: if (osm->is_local) {
142: PetscCall(PetscViewerASCIIPushSynchronized(viewer));
143: PetscCall(PetscViewerFlush(viewer));
144: PetscCall(PetscViewerASCIIPopSynchronized(viewer));
145: }
146: }
147: }
148: PetscCall(PetscViewerFlush(viewer));
149: PetscCall(PetscViewerDestroy(&viewer));
150: PetscFunctionReturn(PETSC_SUCCESS);
151: }
153: static PetscErrorCode PCSetUp_ASM(PC pc)
154: {
155: PC_ASM *osm = (PC_ASM *)pc->data;
156: PetscBool flg;
157: PetscInt i, m, m_local;
158: MatReuse scall = MAT_REUSE_MATRIX;
159: IS isl;
160: KSP ksp;
161: PC subpc;
162: const char *prefix, *pprefix;
163: Vec vec;
164: DM *domain_dm = NULL;
165: MatNullSpace *nullsp = NULL;
167: PetscFunctionBegin;
168: if (!pc->setupcalled) {
169: PetscInt m;
171: /* Note: if subdomains have been set either via PCASMSetTotalSubdomains() or via PCASMSetLocalSubdomains(), osm->n_local_true will not be PETSC_DECIDE */
172: if (osm->n_local_true == PETSC_DECIDE) {
173: /* no subdomains given */
174: /* try pc->dm first, if allowed */
175: if (osm->dm_subdomains && pc->dm) {
176: PetscInt num_domains, d;
177: char **domain_names;
178: IS *inner_domain_is, *outer_domain_is;
179: PetscCall(DMCreateDomainDecomposition(pc->dm, &num_domains, &domain_names, &inner_domain_is, &outer_domain_is, &domain_dm));
180: osm->overlap = -1; /* We do not want to increase the overlap of the IS.
181: A future improvement of this code might allow one to use
182: DM-defined subdomains and also increase the overlap,
183: but that is not currently supported */
184: if (num_domains) PetscCall(PCASMSetLocalSubdomains(pc, num_domains, outer_domain_is, inner_domain_is));
185: for (d = 0; d < num_domains; ++d) {
186: if (domain_names) PetscCall(PetscFree(domain_names[d]));
187: if (inner_domain_is) PetscCall(ISDestroy(&inner_domain_is[d]));
188: if (outer_domain_is) PetscCall(ISDestroy(&outer_domain_is[d]));
189: }
190: PetscCall(PetscFree(domain_names));
191: PetscCall(PetscFree(inner_domain_is));
192: PetscCall(PetscFree(outer_domain_is));
193: }
194: if (osm->n_local_true == PETSC_DECIDE) {
195: /* still no subdomains; use one subdomain per processor */
196: osm->n_local_true = 1;
197: }
198: }
199: { /* determine the global and max number of subdomains */
200: struct {
201: PetscInt max, sum;
202: } inwork, outwork;
203: PetscMPIInt size;
205: inwork.max = osm->n_local_true;
206: inwork.sum = osm->n_local_true;
207: PetscCall(MPIU_Allreduce(&inwork, &outwork, 1, MPIU_2INT, MPIU_MAXSUM_OP, PetscObjectComm((PetscObject)pc)));
208: osm->n_local = outwork.max;
209: osm->n = outwork.sum;
211: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)pc), &size));
212: if (outwork.max == 1 && outwork.sum == size) {
213: /* osm->n_local_true = 1 on all processes, set this option may enable use of optimized MatCreateSubMatrices() implementation */
214: PetscCall(MatSetOption(pc->pmat, MAT_SUBMAT_SINGLEIS, PETSC_TRUE));
215: }
216: }
217: if (!osm->is) { /* create the index sets */
218: PetscCall(PCASMCreateSubdomains(pc->pmat, osm->n_local_true, &osm->is));
219: }
220: if (osm->n_local_true > 1 && !osm->is_local) {
221: PetscCall(PetscMalloc1(osm->n_local_true, &osm->is_local));
222: for (i = 0; i < osm->n_local_true; i++) {
223: if (osm->overlap > 0) { /* With positive overlap, osm->is[i] will be modified */
224: PetscCall(ISDuplicate(osm->is[i], &osm->is_local[i]));
225: PetscCall(ISCopy(osm->is[i], osm->is_local[i]));
226: } else {
227: PetscCall(PetscObjectReference((PetscObject)osm->is[i]));
228: osm->is_local[i] = osm->is[i];
229: }
230: }
231: }
232: PetscCall(PCGetOptionsPrefix(pc, &prefix));
233: if (osm->overlap > 0) {
234: /* Extend the "overlapping" regions by a number of steps */
235: PetscCall(MatIncreaseOverlap(pc->pmat, osm->n_local_true, osm->is, osm->overlap));
236: }
237: if (osm->sort_indices) {
238: for (i = 0; i < osm->n_local_true; i++) {
239: PetscCall(ISSort(osm->is[i]));
240: if (osm->is_local) PetscCall(ISSort(osm->is_local[i]));
241: }
242: }
243: flg = PETSC_FALSE;
244: PetscCall(PetscOptionsHasName(NULL, prefix, "-pc_asm_print_subdomains", &flg));
245: if (flg) PetscCall(PCASMPrintSubdomains(pc));
246: if (!osm->ksp) {
247: /* Create the local solvers */
248: PetscCall(PetscMalloc1(osm->n_local_true, &osm->ksp));
249: if (domain_dm) PetscCall(PetscInfo(pc, "Setting up ASM subproblems using the embedded DM\n"));
250: for (i = 0; i < osm->n_local_true; i++) {
251: PetscCall(KSPCreate(PETSC_COMM_SELF, &ksp));
252: PetscCall(KSPSetNestLevel(ksp, pc->kspnestlevel));
253: PetscCall(KSPSetErrorIfNotConverged(ksp, pc->erroriffailure));
254: PetscCall(PetscObjectIncrementTabLevel((PetscObject)ksp, (PetscObject)pc, 1));
255: PetscCall(KSPSetType(ksp, KSPPREONLY));
256: PetscCall(KSPGetPC(ksp, &subpc));
257: PetscCall(PCGetOptionsPrefix(pc, &prefix));
258: PetscCall(KSPSetOptionsPrefix(ksp, prefix));
259: PetscCall(KSPAppendOptionsPrefix(ksp, "sub_"));
260: if (domain_dm) {
261: PetscCall(KSPSetDM(ksp, domain_dm[i]));
262: PetscCall(KSPSetDMActive(ksp, PETSC_FALSE));
263: PetscCall(DMDestroy(&domain_dm[i]));
264: }
265: osm->ksp[i] = ksp;
266: }
267: if (domain_dm) PetscCall(PetscFree(domain_dm));
268: }
270: PetscCall(ISConcatenate(PETSC_COMM_SELF, osm->n_local_true, osm->is, &osm->lis));
271: PetscCall(ISSortRemoveDups(osm->lis));
272: PetscCall(ISGetLocalSize(osm->lis, &m));
274: scall = MAT_INITIAL_MATRIX;
275: } else {
276: /*
277: Destroy the blocks from the previous iteration
278: */
279: if (pc->flag == DIFFERENT_NONZERO_PATTERN) {
280: PetscCall(MatGetNullSpaces(osm->n_local_true, osm->pmat, &nullsp));
281: PetscCall(MatDestroyMatrices(osm->n_local_true, &osm->pmat));
282: scall = MAT_INITIAL_MATRIX;
283: }
284: }
286: /* Destroy previous submatrices of a different type than pc->pmat since MAT_REUSE_MATRIX won't work in that case */
287: if (scall == MAT_REUSE_MATRIX && osm->sub_mat_type) {
288: PetscCall(MatGetNullSpaces(osm->n_local_true, osm->pmat, &nullsp));
289: if (osm->n_local_true > 0) PetscCall(MatDestroySubMatrices(osm->n_local_true, &osm->pmat));
290: scall = MAT_INITIAL_MATRIX;
291: }
293: /*
294: Extract out the submatrices
295: */
296: PetscCall(MatCreateSubMatrices(pc->pmat, osm->n_local_true, osm->is, osm->is, scall, &osm->pmat));
297: if (scall == MAT_INITIAL_MATRIX) {
298: PetscCall(PetscObjectGetOptionsPrefix((PetscObject)pc->pmat, &pprefix));
299: for (i = 0; i < osm->n_local_true; i++) PetscCall(PetscObjectSetOptionsPrefix((PetscObject)osm->pmat[i], pprefix));
300: if (nullsp) PetscCall(MatRestoreNullSpaces(osm->n_local_true, osm->pmat, &nullsp));
301: }
303: /* Convert the types of the submatrices (if needbe) */
304: if (osm->sub_mat_type) {
305: for (i = 0; i < osm->n_local_true; i++) PetscCall(MatConvert(osm->pmat[i], osm->sub_mat_type, MAT_INPLACE_MATRIX, &osm->pmat[i]));
306: }
308: if (!pc->setupcalled) {
309: VecType vtype;
311: /* Create the local work vectors (from the local matrices) and scatter contexts */
312: PetscCall(MatCreateVecs(pc->pmat, &vec, NULL));
314: PetscCheck(!osm->is_local || osm->n_local_true == 1 || (osm->type != PC_ASM_INTERPOLATE && osm->type != PC_ASM_NONE), PetscObjectComm((PetscObject)pc), PETSC_ERR_SUP, "Cannot use interpolate or none PCASMType if is_local was provided to PCASMSetLocalSubdomains() with more than a single subdomain");
315: if (osm->is_local && osm->type != PC_ASM_BASIC && osm->loctype == PC_COMPOSITE_ADDITIVE) PetscCall(PetscMalloc1(osm->n_local_true, &osm->lprolongation));
316: PetscCall(PetscMalloc1(osm->n_local_true, &osm->lrestriction));
317: PetscCall(PetscMalloc1(osm->n_local_true, &osm->x));
318: PetscCall(PetscMalloc1(osm->n_local_true, &osm->y));
320: PetscCall(ISGetLocalSize(osm->lis, &m));
321: PetscCall(ISCreateStride(PETSC_COMM_SELF, m, 0, 1, &isl));
322: PetscCall(MatGetVecType(osm->pmat[0], &vtype));
323: PetscCall(VecCreate(PETSC_COMM_SELF, &osm->lx));
324: PetscCall(VecSetSizes(osm->lx, m, m));
325: PetscCall(VecSetType(osm->lx, vtype));
326: PetscCall(VecDuplicate(osm->lx, &osm->ly));
327: PetscCall(VecScatterCreate(vec, osm->lis, osm->lx, isl, &osm->restriction));
328: PetscCall(ISDestroy(&isl));
330: for (i = 0; i < osm->n_local_true; ++i) {
331: ISLocalToGlobalMapping ltog;
332: IS isll;
333: const PetscInt *idx_is;
334: PetscInt *idx_lis, nout;
336: PetscCall(ISGetLocalSize(osm->is[i], &m));
337: PetscCall(MatCreateVecs(osm->pmat[i], &osm->x[i], NULL));
338: PetscCall(VecDuplicate(osm->x[i], &osm->y[i]));
340: /* generate a scatter from ly to y[i] picking all the overlapping is[i] entries */
341: PetscCall(ISLocalToGlobalMappingCreateIS(osm->lis, <og));
342: PetscCall(ISGetLocalSize(osm->is[i], &m));
343: PetscCall(ISGetIndices(osm->is[i], &idx_is));
344: PetscCall(PetscMalloc1(m, &idx_lis));
345: PetscCall(ISGlobalToLocalMappingApply(ltog, IS_GTOLM_DROP, m, idx_is, &nout, idx_lis));
346: PetscCheck(nout == m, PETSC_COMM_SELF, PETSC_ERR_PLIB, "is not a subset of lis");
347: PetscCall(ISRestoreIndices(osm->is[i], &idx_is));
348: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, m, idx_lis, PETSC_OWN_POINTER, &isll));
349: PetscCall(ISLocalToGlobalMappingDestroy(<og));
350: PetscCall(ISCreateStride(PETSC_COMM_SELF, m, 0, 1, &isl));
351: PetscCall(VecScatterCreate(osm->ly, isll, osm->y[i], isl, &osm->lrestriction[i]));
352: PetscCall(ISDestroy(&isll));
353: PetscCall(ISDestroy(&isl));
354: if (osm->lprolongation) { /* generate a scatter from y[i] to ly picking only the non-overlapping is_local[i] entries */
355: ISLocalToGlobalMapping ltog;
356: IS isll, isll_local;
357: const PetscInt *idx_local;
358: PetscInt *idx1, *idx2, nout;
360: PetscCall(ISGetLocalSize(osm->is_local[i], &m_local));
361: PetscCall(ISGetIndices(osm->is_local[i], &idx_local));
363: PetscCall(ISLocalToGlobalMappingCreateIS(osm->is[i], <og));
364: PetscCall(PetscMalloc1(m_local, &idx1));
365: PetscCall(ISGlobalToLocalMappingApply(ltog, IS_GTOLM_DROP, m_local, idx_local, &nout, idx1));
366: PetscCall(ISLocalToGlobalMappingDestroy(<og));
367: PetscCheck(nout == m_local, PETSC_COMM_SELF, PETSC_ERR_PLIB, "is_local not a subset of is");
368: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, m_local, idx1, PETSC_OWN_POINTER, &isll));
370: PetscCall(ISLocalToGlobalMappingCreateIS(osm->lis, <og));
371: PetscCall(PetscMalloc1(m_local, &idx2));
372: PetscCall(ISGlobalToLocalMappingApply(ltog, IS_GTOLM_DROP, m_local, idx_local, &nout, idx2));
373: PetscCall(ISLocalToGlobalMappingDestroy(<og));
374: PetscCheck(nout == m_local, PETSC_COMM_SELF, PETSC_ERR_PLIB, "is_local not a subset of lis");
375: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, m_local, idx2, PETSC_OWN_POINTER, &isll_local));
377: PetscCall(ISRestoreIndices(osm->is_local[i], &idx_local));
378: PetscCall(VecScatterCreate(osm->y[i], isll, osm->ly, isll_local, &osm->lprolongation[i]));
380: PetscCall(ISDestroy(&isll));
381: PetscCall(ISDestroy(&isll_local));
382: }
383: }
384: PetscCall(VecDestroy(&vec));
385: }
387: if (osm->loctype == PC_COMPOSITE_MULTIPLICATIVE) {
388: IS *cis;
389: PetscInt c;
391: PetscCall(PetscMalloc1(osm->n_local_true, &cis));
392: for (c = 0; c < osm->n_local_true; ++c) cis[c] = osm->lis;
393: PetscCall(MatCreateSubMatrices(pc->pmat, osm->n_local_true, osm->is, cis, scall, &osm->lmats));
394: PetscCall(PetscFree(cis));
395: }
397: /* Return control to the user so that the submatrices can be modified (e.g., to apply
398: different boundary conditions for the submatrices than for the global problem) */
399: PetscCall(PCModifySubMatrices(pc, osm->n_local_true, osm->is, osm->is, osm->pmat, pc->modifysubmatricesP));
401: /*
402: Loop over subdomains putting them into local ksp
403: */
404: PetscCall(KSPGetOptionsPrefix(osm->ksp[0], &prefix));
405: for (i = 0; i < osm->n_local_true; i++) {
406: PetscCall(KSPSetOperators(osm->ksp[i], osm->pmat[i], osm->pmat[i]));
407: PetscCall(MatSetOptionsPrefix(osm->pmat[i], prefix));
408: if (!pc->setupcalled) PetscCall(KSPSetFromOptions(osm->ksp[i]));
409: }
410: PetscFunctionReturn(PETSC_SUCCESS);
411: }
413: static PetscErrorCode PCSetUpOnBlocks_ASM(PC pc)
414: {
415: PC_ASM *osm = (PC_ASM *)pc->data;
416: PetscInt i;
417: KSPConvergedReason reason;
419: PetscFunctionBegin;
420: for (i = 0; i < osm->n_local_true; i++) {
421: PetscCall(KSPSetUp(osm->ksp[i]));
422: PetscCall(KSPGetConvergedReason(osm->ksp[i], &reason));
423: if (reason == KSP_DIVERGED_PC_FAILED) pc->failedreason = PC_SUBPC_ERROR;
424: }
425: PetscFunctionReturn(PETSC_SUCCESS);
426: }
428: static PetscErrorCode PCApply_ASM(PC pc, Vec x, Vec y)
429: {
430: PC_ASM *osm = (PC_ASM *)pc->data;
431: PetscInt i, n_local_true = osm->n_local_true;
432: ScatterMode forward = SCATTER_FORWARD, reverse = SCATTER_REVERSE;
434: PetscFunctionBegin;
435: /*
436: support for limiting the restriction or interpolation to only local
437: subdomain values (leaving the other values 0).
438: */
439: if (!(osm->type & PC_ASM_RESTRICT)) {
440: forward = SCATTER_FORWARD_LOCAL;
441: /* have to zero the work RHS since scatter may leave some slots empty */
442: PetscCall(VecSet(osm->lx, 0.0));
443: }
444: if (!(osm->type & PC_ASM_INTERPOLATE)) reverse = SCATTER_REVERSE_LOCAL;
446: PetscCheck(osm->loctype == PC_COMPOSITE_MULTIPLICATIVE || osm->loctype == PC_COMPOSITE_ADDITIVE, PetscObjectComm((PetscObject)pc), PETSC_ERR_ARG_WRONG, "Invalid local composition type: %s", PCCompositeTypes[osm->loctype]);
447: /* zero the global and the local solutions */
448: PetscCall(VecSet(y, 0.0));
449: PetscCall(VecSet(osm->ly, 0.0));
451: /* copy the global RHS to local RHS including the ghost nodes */
452: PetscCall(VecScatterBegin(osm->restriction, x, osm->lx, INSERT_VALUES, forward));
453: PetscCall(VecScatterEnd(osm->restriction, x, osm->lx, INSERT_VALUES, forward));
455: /* restrict local RHS to the overlapping 0-block RHS */
456: PetscCall(VecScatterBegin(osm->lrestriction[0], osm->lx, osm->x[0], INSERT_VALUES, forward));
457: PetscCall(VecScatterEnd(osm->lrestriction[0], osm->lx, osm->x[0], INSERT_VALUES, forward));
459: /* do the local solves */
460: for (i = 0; i < n_local_true; ++i) {
461: /* solve the overlapping i-block */
462: PetscCall(PetscLogEventBegin(PC_ApplyOnBlocks, osm->ksp[i], osm->x[i], osm->y[i], 0));
463: PetscCall(KSPSolve(osm->ksp[i], osm->x[i], osm->y[i]));
464: PetscCall(KSPCheckSolve(osm->ksp[i], pc, osm->y[i]));
465: PetscCall(PetscLogEventEnd(PC_ApplyOnBlocks, osm->ksp[i], osm->x[i], osm->y[i], 0));
467: if (osm->lprolongation && osm->type != PC_ASM_INTERPOLATE) { /* interpolate the non-overlapping i-block solution to the local solution (only for restrictive additive) */
468: PetscCall(VecScatterBegin(osm->lprolongation[i], osm->y[i], osm->ly, ADD_VALUES, forward));
469: PetscCall(VecScatterEnd(osm->lprolongation[i], osm->y[i], osm->ly, ADD_VALUES, forward));
470: } else { /* interpolate the overlapping i-block solution to the local solution */
471: PetscCall(VecScatterBegin(osm->lrestriction[i], osm->y[i], osm->ly, ADD_VALUES, reverse));
472: PetscCall(VecScatterEnd(osm->lrestriction[i], osm->y[i], osm->ly, ADD_VALUES, reverse));
473: }
475: if (i < n_local_true - 1) {
476: /* restrict local RHS to the overlapping (i+1)-block RHS */
477: PetscCall(VecScatterBegin(osm->lrestriction[i + 1], osm->lx, osm->x[i + 1], INSERT_VALUES, forward));
478: PetscCall(VecScatterEnd(osm->lrestriction[i + 1], osm->lx, osm->x[i + 1], INSERT_VALUES, forward));
480: if (osm->loctype == PC_COMPOSITE_MULTIPLICATIVE) {
481: /* update the overlapping (i+1)-block RHS using the current local solution */
482: PetscCall(MatMult(osm->lmats[i + 1], osm->ly, osm->y[i + 1]));
483: PetscCall(VecAXPBY(osm->x[i + 1], -1., 1., osm->y[i + 1]));
484: }
485: }
486: }
487: /* add the local solution to the global solution including the ghost nodes */
488: PetscCall(VecScatterBegin(osm->restriction, osm->ly, y, ADD_VALUES, reverse));
489: PetscCall(VecScatterEnd(osm->restriction, osm->ly, y, ADD_VALUES, reverse));
490: PetscFunctionReturn(PETSC_SUCCESS);
491: }
493: static PetscErrorCode PCMatApply_ASM(PC pc, Mat X, Mat Y)
494: {
495: PC_ASM *osm = (PC_ASM *)pc->data;
496: Mat Z, W;
497: Vec x;
498: PetscInt i, m, N;
499: ScatterMode forward = SCATTER_FORWARD, reverse = SCATTER_REVERSE;
501: PetscFunctionBegin;
502: PetscCheck(osm->n_local_true <= 1, PetscObjectComm((PetscObject)pc), PETSC_ERR_SUP, "Not yet implemented");
503: /*
504: support for limiting the restriction or interpolation to only local
505: subdomain values (leaving the other values 0).
506: */
507: if (!(osm->type & PC_ASM_RESTRICT)) {
508: forward = SCATTER_FORWARD_LOCAL;
509: /* have to zero the work RHS since scatter may leave some slots empty */
510: PetscCall(VecSet(osm->lx, 0.0));
511: }
512: if (!(osm->type & PC_ASM_INTERPOLATE)) reverse = SCATTER_REVERSE_LOCAL;
513: PetscCall(VecGetLocalSize(osm->x[0], &m));
514: PetscCall(MatGetSize(X, NULL, &N));
515: PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, m, N, NULL, &Z));
517: PetscCheck(osm->loctype == PC_COMPOSITE_MULTIPLICATIVE || osm->loctype == PC_COMPOSITE_ADDITIVE, PetscObjectComm((PetscObject)pc), PETSC_ERR_ARG_WRONG, "Invalid local composition type: %s", PCCompositeTypes[osm->loctype]);
518: /* zero the global and the local solutions */
519: PetscCall(MatZeroEntries(Y));
520: PetscCall(VecSet(osm->ly, 0.0));
522: for (i = 0; i < N; ++i) {
523: PetscCall(MatDenseGetColumnVecRead(X, i, &x));
524: /* copy the global RHS to local RHS including the ghost nodes */
525: PetscCall(VecScatterBegin(osm->restriction, x, osm->lx, INSERT_VALUES, forward));
526: PetscCall(VecScatterEnd(osm->restriction, x, osm->lx, INSERT_VALUES, forward));
527: PetscCall(MatDenseRestoreColumnVecRead(X, i, &x));
529: PetscCall(MatDenseGetColumnVecWrite(Z, i, &x));
530: /* restrict local RHS to the overlapping 0-block RHS */
531: PetscCall(VecScatterBegin(osm->lrestriction[0], osm->lx, x, INSERT_VALUES, forward));
532: PetscCall(VecScatterEnd(osm->lrestriction[0], osm->lx, x, INSERT_VALUES, forward));
533: PetscCall(MatDenseRestoreColumnVecWrite(Z, i, &x));
534: }
535: PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, m, N, NULL, &W));
536: /* solve the overlapping 0-block */
537: PetscCall(PetscLogEventBegin(PC_ApplyOnBlocks, osm->ksp[0], Z, W, 0));
538: PetscCall(KSPMatSolve(osm->ksp[0], Z, W));
539: PetscCall(KSPCheckSolve(osm->ksp[0], pc, NULL));
540: PetscCall(PetscLogEventEnd(PC_ApplyOnBlocks, osm->ksp[0], Z, W, 0));
541: PetscCall(MatDestroy(&Z));
543: for (i = 0; i < N; ++i) {
544: PetscCall(VecSet(osm->ly, 0.0));
545: PetscCall(MatDenseGetColumnVecRead(W, i, &x));
546: if (osm->lprolongation && osm->type != PC_ASM_INTERPOLATE) { /* interpolate the non-overlapping 0-block solution to the local solution (only for restrictive additive) */
547: PetscCall(VecScatterBegin(osm->lprolongation[0], x, osm->ly, ADD_VALUES, forward));
548: PetscCall(VecScatterEnd(osm->lprolongation[0], x, osm->ly, ADD_VALUES, forward));
549: } else { /* interpolate the overlapping 0-block solution to the local solution */
550: PetscCall(VecScatterBegin(osm->lrestriction[0], x, osm->ly, ADD_VALUES, reverse));
551: PetscCall(VecScatterEnd(osm->lrestriction[0], x, osm->ly, ADD_VALUES, reverse));
552: }
553: PetscCall(MatDenseRestoreColumnVecRead(W, i, &x));
555: PetscCall(MatDenseGetColumnVecWrite(Y, i, &x));
556: /* add the local solution to the global solution including the ghost nodes */
557: PetscCall(VecScatterBegin(osm->restriction, osm->ly, x, ADD_VALUES, reverse));
558: PetscCall(VecScatterEnd(osm->restriction, osm->ly, x, ADD_VALUES, reverse));
559: PetscCall(MatDenseRestoreColumnVecWrite(Y, i, &x));
560: }
561: PetscCall(MatDestroy(&W));
562: PetscFunctionReturn(PETSC_SUCCESS);
563: }
565: static PetscErrorCode PCApplyTranspose_ASM(PC pc, Vec x, Vec y)
566: {
567: PC_ASM *osm = (PC_ASM *)pc->data;
568: PetscInt i, n_local_true = osm->n_local_true;
569: ScatterMode forward = SCATTER_FORWARD, reverse = SCATTER_REVERSE;
571: PetscFunctionBegin;
572: /*
573: Support for limiting the restriction or interpolation to only local
574: subdomain values (leaving the other values 0).
576: Note: these are reversed from the PCApply_ASM() because we are applying the
577: transpose of the three terms
578: */
580: if (!(osm->type & PC_ASM_INTERPOLATE)) {
581: forward = SCATTER_FORWARD_LOCAL;
582: /* have to zero the work RHS since scatter may leave some slots empty */
583: PetscCall(VecSet(osm->lx, 0.0));
584: }
585: if (!(osm->type & PC_ASM_RESTRICT)) reverse = SCATTER_REVERSE_LOCAL;
587: /* zero the global and the local solutions */
588: PetscCall(VecSet(y, 0.0));
589: PetscCall(VecSet(osm->ly, 0.0));
591: /* Copy the global RHS to local RHS including the ghost nodes */
592: PetscCall(VecScatterBegin(osm->restriction, x, osm->lx, INSERT_VALUES, forward));
593: PetscCall(VecScatterEnd(osm->restriction, x, osm->lx, INSERT_VALUES, forward));
595: /* Restrict local RHS to the overlapping 0-block RHS */
596: PetscCall(VecScatterBegin(osm->lrestriction[0], osm->lx, osm->x[0], INSERT_VALUES, forward));
597: PetscCall(VecScatterEnd(osm->lrestriction[0], osm->lx, osm->x[0], INSERT_VALUES, forward));
599: /* do the local solves */
600: for (i = 0; i < n_local_true; ++i) {
601: /* solve the overlapping i-block */
602: PetscCall(PetscLogEventBegin(PC_ApplyOnBlocks, osm->ksp[i], osm->x[i], osm->y[i], 0));
603: PetscCall(KSPSolveTranspose(osm->ksp[i], osm->x[i], osm->y[i]));
604: PetscCall(KSPCheckSolve(osm->ksp[i], pc, osm->y[i]));
605: PetscCall(PetscLogEventEnd(PC_ApplyOnBlocks, osm->ksp[i], osm->x[i], osm->y[i], 0));
607: if (osm->lprolongation && osm->type != PC_ASM_RESTRICT) { /* interpolate the non-overlapping i-block solution to the local solution */
608: PetscCall(VecScatterBegin(osm->lprolongation[i], osm->y[i], osm->ly, ADD_VALUES, forward));
609: PetscCall(VecScatterEnd(osm->lprolongation[i], osm->y[i], osm->ly, ADD_VALUES, forward));
610: } else { /* interpolate the overlapping i-block solution to the local solution */
611: PetscCall(VecScatterBegin(osm->lrestriction[i], osm->y[i], osm->ly, ADD_VALUES, reverse));
612: PetscCall(VecScatterEnd(osm->lrestriction[i], osm->y[i], osm->ly, ADD_VALUES, reverse));
613: }
615: if (i < n_local_true - 1) {
616: /* Restrict local RHS to the overlapping (i+1)-block RHS */
617: PetscCall(VecScatterBegin(osm->lrestriction[i + 1], osm->lx, osm->x[i + 1], INSERT_VALUES, forward));
618: PetscCall(VecScatterEnd(osm->lrestriction[i + 1], osm->lx, osm->x[i + 1], INSERT_VALUES, forward));
619: }
620: }
621: /* Add the local solution to the global solution including the ghost nodes */
622: PetscCall(VecScatterBegin(osm->restriction, osm->ly, y, ADD_VALUES, reverse));
623: PetscCall(VecScatterEnd(osm->restriction, osm->ly, y, ADD_VALUES, reverse));
624: PetscFunctionReturn(PETSC_SUCCESS);
625: }
627: static PetscErrorCode PCReset_ASM(PC pc)
628: {
629: PC_ASM *osm = (PC_ASM *)pc->data;
630: PetscInt i;
632: PetscFunctionBegin;
633: if (osm->ksp) {
634: for (i = 0; i < osm->n_local_true; i++) PetscCall(KSPReset(osm->ksp[i]));
635: }
636: if (osm->pmat) {
637: if (osm->n_local_true > 0) PetscCall(MatDestroySubMatrices(osm->n_local_true, &osm->pmat));
638: }
639: if (osm->lrestriction) {
640: PetscCall(VecScatterDestroy(&osm->restriction));
641: for (i = 0; i < osm->n_local_true; i++) {
642: PetscCall(VecScatterDestroy(&osm->lrestriction[i]));
643: if (osm->lprolongation) PetscCall(VecScatterDestroy(&osm->lprolongation[i]));
644: PetscCall(VecDestroy(&osm->x[i]));
645: PetscCall(VecDestroy(&osm->y[i]));
646: }
647: PetscCall(PetscFree(osm->lrestriction));
648: if (osm->lprolongation) PetscCall(PetscFree(osm->lprolongation));
649: PetscCall(PetscFree(osm->x));
650: PetscCall(PetscFree(osm->y));
651: }
652: PetscCall(PCASMDestroySubdomains(osm->n_local_true, osm->is, osm->is_local));
653: PetscCall(ISDestroy(&osm->lis));
654: PetscCall(VecDestroy(&osm->lx));
655: PetscCall(VecDestroy(&osm->ly));
656: if (osm->loctype == PC_COMPOSITE_MULTIPLICATIVE) PetscCall(MatDestroyMatrices(osm->n_local_true, &osm->lmats));
658: PetscCall(PetscFree(osm->sub_mat_type));
660: osm->is = NULL;
661: osm->is_local = NULL;
662: PetscFunctionReturn(PETSC_SUCCESS);
663: }
665: static PetscErrorCode PCDestroy_ASM(PC pc)
666: {
667: PC_ASM *osm = (PC_ASM *)pc->data;
668: PetscInt i;
670: PetscFunctionBegin;
671: PetscCall(PCReset_ASM(pc));
672: if (osm->ksp) {
673: for (i = 0; i < osm->n_local_true; i++) PetscCall(KSPDestroy(&osm->ksp[i]));
674: PetscCall(PetscFree(osm->ksp));
675: }
676: PetscCall(PetscFree(pc->data));
678: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCASMSetLocalSubdomains_C", NULL));
679: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCASMSetTotalSubdomains_C", NULL));
680: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCASMSetOverlap_C", NULL));
681: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCASMSetType_C", NULL));
682: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCASMGetType_C", NULL));
683: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCASMSetLocalType_C", NULL));
684: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCASMGetLocalType_C", NULL));
685: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCASMSetSortIndices_C", NULL));
686: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCASMGetSubKSP_C", NULL));
687: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCASMGetSubMatType_C", NULL));
688: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCASMSetSubMatType_C", NULL));
689: PetscFunctionReturn(PETSC_SUCCESS);
690: }
692: static PetscErrorCode PCSetFromOptions_ASM(PC pc, PetscOptionItems *PetscOptionsObject)
693: {
694: PC_ASM *osm = (PC_ASM *)pc->data;
695: PetscInt blocks, ovl;
696: PetscBool flg;
697: PCASMType asmtype;
698: PCCompositeType loctype;
699: char sub_mat_type[256];
701: PetscFunctionBegin;
702: PetscOptionsHeadBegin(PetscOptionsObject, "Additive Schwarz options");
703: PetscCall(PetscOptionsBool("-pc_asm_dm_subdomains", "Use DMCreateDomainDecomposition() to define subdomains", "PCASMSetDMSubdomains", osm->dm_subdomains, &osm->dm_subdomains, &flg));
704: PetscCall(PetscOptionsInt("-pc_asm_blocks", "Number of subdomains", "PCASMSetTotalSubdomains", osm->n, &blocks, &flg));
705: if (flg) {
706: PetscCall(PCASMSetTotalSubdomains(pc, blocks, NULL, NULL));
707: osm->dm_subdomains = PETSC_FALSE;
708: }
709: PetscCall(PetscOptionsInt("-pc_asm_local_blocks", "Number of local subdomains", "PCASMSetLocalSubdomains", osm->n_local_true, &blocks, &flg));
710: if (flg) {
711: PetscCall(PCASMSetLocalSubdomains(pc, blocks, NULL, NULL));
712: osm->dm_subdomains = PETSC_FALSE;
713: }
714: PetscCall(PetscOptionsInt("-pc_asm_overlap", "Number of grid points overlap", "PCASMSetOverlap", osm->overlap, &ovl, &flg));
715: if (flg) {
716: PetscCall(PCASMSetOverlap(pc, ovl));
717: osm->dm_subdomains = PETSC_FALSE;
718: }
719: flg = PETSC_FALSE;
720: PetscCall(PetscOptionsEnum("-pc_asm_type", "Type of restriction/extension", "PCASMSetType", PCASMTypes, (PetscEnum)osm->type, (PetscEnum *)&asmtype, &flg));
721: if (flg) PetscCall(PCASMSetType(pc, asmtype));
722: flg = PETSC_FALSE;
723: PetscCall(PetscOptionsEnum("-pc_asm_local_type", "Type of local solver composition", "PCASMSetLocalType", PCCompositeTypes, (PetscEnum)osm->loctype, (PetscEnum *)&loctype, &flg));
724: if (flg) PetscCall(PCASMSetLocalType(pc, loctype));
725: PetscCall(PetscOptionsFList("-pc_asm_sub_mat_type", "Subsolve Matrix Type", "PCASMSetSubMatType", MatList, NULL, sub_mat_type, 256, &flg));
726: if (flg) PetscCall(PCASMSetSubMatType(pc, sub_mat_type));
727: PetscOptionsHeadEnd();
728: PetscFunctionReturn(PETSC_SUCCESS);
729: }
731: static PetscErrorCode PCASMSetLocalSubdomains_ASM(PC pc, PetscInt n, IS is[], IS is_local[])
732: {
733: PC_ASM *osm = (PC_ASM *)pc->data;
734: PetscInt i;
736: PetscFunctionBegin;
737: PetscCheck(n >= 1, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Each process must have 1 or more blocks, n = %" PetscInt_FMT, n);
738: PetscCheck(!pc->setupcalled || (n == osm->n_local_true && !is), PetscObjectComm((PetscObject)pc), PETSC_ERR_ARG_WRONGSTATE, "PCASMSetLocalSubdomains() should be called before calling PCSetUp().");
740: if (!pc->setupcalled) {
741: if (is) {
742: for (i = 0; i < n; i++) PetscCall(PetscObjectReference((PetscObject)is[i]));
743: }
744: if (is_local) {
745: for (i = 0; i < n; i++) PetscCall(PetscObjectReference((PetscObject)is_local[i]));
746: }
747: PetscCall(PCASMDestroySubdomains(osm->n_local_true, osm->is, osm->is_local));
749: if (osm->ksp && osm->n_local_true != n) {
750: for (i = 0; i < osm->n_local_true; i++) PetscCall(KSPDestroy(&osm->ksp[i]));
751: PetscCall(PetscFree(osm->ksp));
752: }
754: osm->n_local_true = n;
755: osm->is = NULL;
756: osm->is_local = NULL;
757: if (is) {
758: PetscCall(PetscMalloc1(n, &osm->is));
759: for (i = 0; i < n; i++) osm->is[i] = is[i];
760: /* Flag indicating that the user has set overlapping subdomains so PCASM should not increase their size. */
761: osm->overlap = -1;
762: }
763: if (is_local) {
764: PetscCall(PetscMalloc1(n, &osm->is_local));
765: for (i = 0; i < n; i++) osm->is_local[i] = is_local[i];
766: if (!is) {
767: PetscCall(PetscMalloc1(osm->n_local_true, &osm->is));
768: for (i = 0; i < osm->n_local_true; i++) {
769: if (osm->overlap > 0) { /* With positive overlap, osm->is[i] will be modified */
770: PetscCall(ISDuplicate(osm->is_local[i], &osm->is[i]));
771: PetscCall(ISCopy(osm->is_local[i], osm->is[i]));
772: } else {
773: PetscCall(PetscObjectReference((PetscObject)osm->is_local[i]));
774: osm->is[i] = osm->is_local[i];
775: }
776: }
777: }
778: }
779: }
780: PetscFunctionReturn(PETSC_SUCCESS);
781: }
783: static PetscErrorCode PCASMSetTotalSubdomains_ASM(PC pc, PetscInt N, IS *is, IS *is_local)
784: {
785: PC_ASM *osm = (PC_ASM *)pc->data;
786: PetscMPIInt rank, size;
787: PetscInt n;
789: PetscFunctionBegin;
790: PetscCheck(N >= 1, PetscObjectComm((PetscObject)pc), PETSC_ERR_ARG_OUTOFRANGE, "Number of total blocks must be > 0, N = %" PetscInt_FMT, N);
791: PetscCheck(!is && !is_local, PetscObjectComm((PetscObject)pc), PETSC_ERR_SUP, "Use PCASMSetLocalSubdomains() to set specific index sets, they cannot be set globally yet.");
793: /*
794: Split the subdomains equally among all processors
795: */
796: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)pc), &rank));
797: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)pc), &size));
798: n = N / size + ((N % size) > rank);
799: PetscCheck(n, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Process %d must have at least one block: total processors %d total blocks %" PetscInt_FMT, (int)rank, (int)size, N);
800: PetscCheck(!pc->setupcalled || n == osm->n_local_true, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "PCASMSetTotalSubdomains() should be called before PCSetUp().");
801: if (!pc->setupcalled) {
802: PetscCall(PCASMDestroySubdomains(osm->n_local_true, osm->is, osm->is_local));
804: osm->n_local_true = n;
805: osm->is = NULL;
806: osm->is_local = NULL;
807: }
808: PetscFunctionReturn(PETSC_SUCCESS);
809: }
811: static PetscErrorCode PCASMSetOverlap_ASM(PC pc, PetscInt ovl)
812: {
813: PC_ASM *osm = (PC_ASM *)pc->data;
815: PetscFunctionBegin;
816: PetscCheck(ovl >= 0, PetscObjectComm((PetscObject)pc), PETSC_ERR_ARG_OUTOFRANGE, "Negative overlap value requested");
817: PetscCheck(!pc->setupcalled || ovl == osm->overlap, PetscObjectComm((PetscObject)pc), PETSC_ERR_ARG_WRONGSTATE, "PCASMSetOverlap() should be called before PCSetUp().");
818: if (!pc->setupcalled) osm->overlap = ovl;
819: PetscFunctionReturn(PETSC_SUCCESS);
820: }
822: static PetscErrorCode PCASMSetType_ASM(PC pc, PCASMType type)
823: {
824: PC_ASM *osm = (PC_ASM *)pc->data;
826: PetscFunctionBegin;
827: osm->type = type;
828: osm->type_set = PETSC_TRUE;
829: PetscFunctionReturn(PETSC_SUCCESS);
830: }
832: static PetscErrorCode PCASMGetType_ASM(PC pc, PCASMType *type)
833: {
834: PC_ASM *osm = (PC_ASM *)pc->data;
836: PetscFunctionBegin;
837: *type = osm->type;
838: PetscFunctionReturn(PETSC_SUCCESS);
839: }
841: static PetscErrorCode PCASMSetLocalType_ASM(PC pc, PCCompositeType type)
842: {
843: PC_ASM *osm = (PC_ASM *)pc->data;
845: PetscFunctionBegin;
846: PetscCheck(type == PC_COMPOSITE_ADDITIVE || type == PC_COMPOSITE_MULTIPLICATIVE, PetscObjectComm((PetscObject)pc), PETSC_ERR_SUP, "Only supports additive or multiplicative as the local type");
847: osm->loctype = type;
848: PetscFunctionReturn(PETSC_SUCCESS);
849: }
851: static PetscErrorCode PCASMGetLocalType_ASM(PC pc, PCCompositeType *type)
852: {
853: PC_ASM *osm = (PC_ASM *)pc->data;
855: PetscFunctionBegin;
856: *type = osm->loctype;
857: PetscFunctionReturn(PETSC_SUCCESS);
858: }
860: static PetscErrorCode PCASMSetSortIndices_ASM(PC pc, PetscBool doSort)
861: {
862: PC_ASM *osm = (PC_ASM *)pc->data;
864: PetscFunctionBegin;
865: osm->sort_indices = doSort;
866: PetscFunctionReturn(PETSC_SUCCESS);
867: }
869: static PetscErrorCode PCASMGetSubKSP_ASM(PC pc, PetscInt *n_local, PetscInt *first_local, KSP **ksp)
870: {
871: PC_ASM *osm = (PC_ASM *)pc->data;
873: PetscFunctionBegin;
874: PetscCheck(osm->n_local_true >= 1, PetscObjectComm((PetscObject)pc), PETSC_ERR_ORDER, "Need to call PCSetUp() on PC (or KSPSetUp() on the outer KSP object) before calling here");
876: if (n_local) *n_local = osm->n_local_true;
877: if (first_local) {
878: PetscCallMPI(MPI_Scan(&osm->n_local_true, first_local, 1, MPIU_INT, MPI_SUM, PetscObjectComm((PetscObject)pc)));
879: *first_local -= osm->n_local_true;
880: }
881: if (ksp) *ksp = osm->ksp;
882: PetscFunctionReturn(PETSC_SUCCESS);
883: }
885: static PetscErrorCode PCASMGetSubMatType_ASM(PC pc, MatType *sub_mat_type)
886: {
887: PC_ASM *osm = (PC_ASM *)pc->data;
889: PetscFunctionBegin;
891: PetscAssertPointer(sub_mat_type, 2);
892: *sub_mat_type = osm->sub_mat_type;
893: PetscFunctionReturn(PETSC_SUCCESS);
894: }
896: static PetscErrorCode PCASMSetSubMatType_ASM(PC pc, MatType sub_mat_type)
897: {
898: PC_ASM *osm = (PC_ASM *)pc->data;
900: PetscFunctionBegin;
902: PetscCall(PetscFree(osm->sub_mat_type));
903: PetscCall(PetscStrallocpy(sub_mat_type, (char **)&osm->sub_mat_type));
904: PetscFunctionReturn(PETSC_SUCCESS);
905: }
907: /*@C
908: PCASMSetLocalSubdomains - Sets the local subdomains (for this processor only) for the additive Schwarz preconditioner `PCASM`.
910: Collective
912: Input Parameters:
913: + pc - the preconditioner context
914: . n - the number of subdomains for this processor (default value = 1)
915: . is - the index set that defines the subdomains for this processor (or `NULL` for PETSc to determine subdomains)
916: the values of the `is` array are copied so you can free the array (not the `IS` in the array) after this call
917: - is_local - the index sets that define the local part of the subdomains for this processor, not used unless `PCASMType` is `PC_ASM_RESTRICT`
918: (or `NULL` to not provide these). The values of the `is_local` array are copied so you can free the array
919: (not the `IS` in the array) after this call
921: Options Database Key:
922: . -pc_asm_local_blocks <blks> - Sets number of local blocks
924: Level: advanced
926: Notes:
927: The `IS` numbering is in the parallel, global numbering of the vector for both `is` and `is_local`
929: By default the `PCASM` preconditioner uses 1 block per processor.
931: Use `PCASMSetTotalSubdomains()` to set the subdomains for all processors.
933: If `is_local` is provided and `PCASMType` is `PC_ASM_RESTRICT` then the solution only over the `is_local` region is interpolated
934: back to form the global solution (this is the standard restricted additive Schwarz method, RASM)
936: If `is_local` is provided and `PCASMType` is `PC_ASM_INTERPOLATE` or `PC_ASM_NONE` then an error is generated since there is
937: no code to handle that case.
939: .seealso: [](ch_ksp), `PCASM`, `PCASMSetTotalSubdomains()`, `PCASMSetOverlap()`, `PCASMGetSubKSP()`,
940: `PCASMCreateSubdomains2D()`, `PCASMGetLocalSubdomains()`, `PCASMType`, `PCASMSetType()`, `PCGASM`
941: @*/
942: PetscErrorCode PCASMSetLocalSubdomains(PC pc, PetscInt n, IS is[], IS is_local[])
943: {
944: PetscFunctionBegin;
946: PetscTryMethod(pc, "PCASMSetLocalSubdomains_C", (PC, PetscInt, IS[], IS[]), (pc, n, is, is_local));
947: PetscFunctionReturn(PETSC_SUCCESS);
948: }
950: /*@C
951: PCASMSetTotalSubdomains - Sets the subdomains for all processors for the
952: additive Schwarz preconditioner, `PCASM`.
954: Collective, all MPI ranks must pass in the same array of `IS`
956: Input Parameters:
957: + pc - the preconditioner context
958: . N - the number of subdomains for all processors
959: . is - the index sets that define the subdomains for all processors (or `NULL` to ask PETSc to determine the subdomains)
960: - is_local - the index sets that define the local part of the subdomains for this processor (or `NULL` to not provide this information)
962: Options Database Key:
963: . -pc_asm_blocks <blks> - Sets total blocks
965: Level: advanced
967: Notes:
968: Currently you cannot use this to set the actual subdomains with the argument `is` or `is_local`.
970: By default the `PCASM` preconditioner uses 1 block per processor.
972: These index sets cannot be destroyed until after completion of the
973: linear solves for which the `PCASM` preconditioner is being used.
975: Use `PCASMSetLocalSubdomains()` to set local subdomains.
977: The `IS` numbering is in the parallel, global numbering of the vector for both is and is_local
979: .seealso: [](ch_ksp), `PCASM`, `PCASMSetLocalSubdomains()`, `PCASMSetOverlap()`, `PCASMGetSubKSP()`,
980: `PCASMCreateSubdomains2D()`, `PCGASM`
981: @*/
982: PetscErrorCode PCASMSetTotalSubdomains(PC pc, PetscInt N, IS is[], IS is_local[])
983: {
984: PetscFunctionBegin;
986: PetscTryMethod(pc, "PCASMSetTotalSubdomains_C", (PC, PetscInt, IS[], IS[]), (pc, N, is, is_local));
987: PetscFunctionReturn(PETSC_SUCCESS);
988: }
990: /*@
991: PCASMSetOverlap - Sets the overlap between a pair of subdomains for the
992: additive Schwarz preconditioner, `PCASM`.
994: Logically Collective
996: Input Parameters:
997: + pc - the preconditioner context
998: - ovl - the amount of overlap between subdomains (ovl >= 0, default value = 1)
1000: Options Database Key:
1001: . -pc_asm_overlap <ovl> - Sets overlap
1003: Level: intermediate
1005: Notes:
1006: By default the `PCASM` preconditioner uses 1 block per processor. To use
1007: multiple blocks per perocessor, see `PCASMSetTotalSubdomains()` and
1008: `PCASMSetLocalSubdomains()` (and the option -pc_asm_blocks <blks>).
1010: The overlap defaults to 1, so if one desires that no additional
1011: overlap be computed beyond what may have been set with a call to
1012: `PCASMSetTotalSubdomains()` or `PCASMSetLocalSubdomains()`, then ovl
1013: must be set to be 0. In particular, if one does not explicitly set
1014: the subdomains an application code, then all overlap would be computed
1015: internally by PETSc, and using an overlap of 0 would result in an `PCASM`
1016: variant that is equivalent to the block Jacobi preconditioner.
1018: The default algorithm used by PETSc to increase overlap is fast, but not scalable,
1019: use the option -mat_increase_overlap_scalable when the problem and number of processes is large.
1021: One can define initial index sets with any overlap via
1022: `PCASMSetLocalSubdomains()`; the routine
1023: `PCASMSetOverlap()` merely allows PETSc to extend that overlap further
1024: if desired.
1026: .seealso: [](ch_ksp), `PCASM`, `PCASMSetTotalSubdomains()`, `PCASMSetLocalSubdomains()`, `PCASMGetSubKSP()`,
1027: `PCASMCreateSubdomains2D()`, `PCASMGetLocalSubdomains()`, `MatIncreaseOverlap()`, `PCGASM`
1028: @*/
1029: PetscErrorCode PCASMSetOverlap(PC pc, PetscInt ovl)
1030: {
1031: PetscFunctionBegin;
1034: PetscTryMethod(pc, "PCASMSetOverlap_C", (PC, PetscInt), (pc, ovl));
1035: PetscFunctionReturn(PETSC_SUCCESS);
1036: }
1038: /*@
1039: PCASMSetType - Sets the type of restriction and interpolation used
1040: for local problems in the additive Schwarz method, `PCASM`.
1042: Logically Collective
1044: Input Parameters:
1045: + pc - the preconditioner context
1046: - type - variant of `PCASM`, one of
1047: .vb
1048: PC_ASM_BASIC - full interpolation and restriction
1049: PC_ASM_RESTRICT - full restriction, local processor interpolation (default)
1050: PC_ASM_INTERPOLATE - full interpolation, local processor restriction
1051: PC_ASM_NONE - local processor restriction and interpolation
1052: .ve
1054: Options Database Key:
1055: . -pc_asm_type [basic,restrict,interpolate,none] - Sets `PCASMType`
1057: Level: intermediate
1059: Note:
1060: if the is_local arguments are passed to `PCASMSetLocalSubdomains()` then they are used when `PC_ASM_RESTRICT` has been selected
1061: to limit the local processor interpolation
1063: .seealso: [](ch_ksp), `PCASM`, `PCASMSetTotalSubdomains()`, `PCASMGetSubKSP()`,
1064: `PCASMCreateSubdomains2D()`, `PCASMType`, `PCASMSetLocalType()`, `PCASMGetLocalType()`, `PCGASM`
1065: @*/
1066: PetscErrorCode PCASMSetType(PC pc, PCASMType type)
1067: {
1068: PetscFunctionBegin;
1071: PetscTryMethod(pc, "PCASMSetType_C", (PC, PCASMType), (pc, type));
1072: PetscFunctionReturn(PETSC_SUCCESS);
1073: }
1075: /*@
1076: PCASMGetType - Gets the type of restriction and interpolation used
1077: for local problems in the additive Schwarz method, `PCASM`.
1079: Logically Collective
1081: Input Parameter:
1082: . pc - the preconditioner context
1084: Output Parameter:
1085: . type - variant of `PCASM`, one of
1086: .vb
1087: PC_ASM_BASIC - full interpolation and restriction
1088: PC_ASM_RESTRICT - full restriction, local processor interpolation
1089: PC_ASM_INTERPOLATE - full interpolation, local processor restriction
1090: PC_ASM_NONE - local processor restriction and interpolation
1091: .ve
1093: Options Database Key:
1094: . -pc_asm_type [basic,restrict,interpolate,none] - Sets `PCASM` type
1096: Level: intermediate
1098: .seealso: [](ch_ksp), `PCASM`, `PCASMSetTotalSubdomains()`, `PCASMGetSubKSP()`, `PCGASM`,
1099: `PCASMCreateSubdomains2D()`, `PCASMType`, `PCASMSetType()`, `PCASMSetLocalType()`, `PCASMGetLocalType()`
1100: @*/
1101: PetscErrorCode PCASMGetType(PC pc, PCASMType *type)
1102: {
1103: PetscFunctionBegin;
1105: PetscUseMethod(pc, "PCASMGetType_C", (PC, PCASMType *), (pc, type));
1106: PetscFunctionReturn(PETSC_SUCCESS);
1107: }
1109: /*@
1110: PCASMSetLocalType - Sets the type of composition used for local problems in the additive Schwarz method, `PCASM`.
1112: Logically Collective
1114: Input Parameters:
1115: + pc - the preconditioner context
1116: - type - type of composition, one of
1117: .vb
1118: PC_COMPOSITE_ADDITIVE - local additive combination
1119: PC_COMPOSITE_MULTIPLICATIVE - local multiplicative combination
1120: .ve
1122: Options Database Key:
1123: . -pc_asm_local_type [additive,multiplicative] - Sets local solver composition type
1125: Level: intermediate
1127: .seealso: [](ch_ksp), `PCASM`, `PCASMSetType()`, `PCASMGetType()`, `PCASMGetLocalType()`, `PCASMType`, `PCCompositeType`
1128: @*/
1129: PetscErrorCode PCASMSetLocalType(PC pc, PCCompositeType type)
1130: {
1131: PetscFunctionBegin;
1134: PetscTryMethod(pc, "PCASMSetLocalType_C", (PC, PCCompositeType), (pc, type));
1135: PetscFunctionReturn(PETSC_SUCCESS);
1136: }
1138: /*@
1139: PCASMGetLocalType - Gets the type of composition used for local problems in the additive Schwarz method, `PCASM`.
1141: Logically Collective
1143: Input Parameter:
1144: . pc - the preconditioner context
1146: Output Parameter:
1147: . type - type of composition, one of
1148: .vb
1149: PC_COMPOSITE_ADDITIVE - local additive combination
1150: PC_COMPOSITE_MULTIPLICATIVE - local multiplicative combination
1151: .ve
1153: Options Database Key:
1154: . -pc_asm_local_type [additive,multiplicative] - Sets local solver composition type
1156: Level: intermediate
1158: .seealso: [](ch_ksp), `PCASM`, `PCASMSetType()`, `PCASMGetType()`, `PCASMSetLocalType()`, `PCASMType`, `PCCompositeType`
1159: @*/
1160: PetscErrorCode PCASMGetLocalType(PC pc, PCCompositeType *type)
1161: {
1162: PetscFunctionBegin;
1164: PetscAssertPointer(type, 2);
1165: PetscUseMethod(pc, "PCASMGetLocalType_C", (PC, PCCompositeType *), (pc, type));
1166: PetscFunctionReturn(PETSC_SUCCESS);
1167: }
1169: /*@
1170: PCASMSetSortIndices - Determines whether subdomain indices are sorted.
1172: Logically Collective
1174: Input Parameters:
1175: + pc - the preconditioner context
1176: - doSort - sort the subdomain indices
1178: Level: intermediate
1180: .seealso: [](ch_ksp), `PCASM`, `PCASMSetLocalSubdomains()`, `PCASMSetTotalSubdomains()`, `PCASMGetSubKSP()`,
1181: `PCASMCreateSubdomains2D()`
1182: @*/
1183: PetscErrorCode PCASMSetSortIndices(PC pc, PetscBool doSort)
1184: {
1185: PetscFunctionBegin;
1188: PetscTryMethod(pc, "PCASMSetSortIndices_C", (PC, PetscBool), (pc, doSort));
1189: PetscFunctionReturn(PETSC_SUCCESS);
1190: }
1192: /*@C
1193: PCASMGetSubKSP - Gets the local `KSP` contexts for all blocks on
1194: this processor.
1196: Collective iff first_local is requested
1198: Input Parameter:
1199: . pc - the preconditioner context
1201: Output Parameters:
1202: + n_local - the number of blocks on this processor or `NULL`
1203: . first_local - the global number of the first block on this processor or `NULL`, all processors must request or all must pass `NULL`
1204: - ksp - the array of `KSP` contexts
1206: Level: advanced
1208: Notes:
1209: After `PCASMGetSubKSP()` the array of `KSP`s is not to be freed.
1211: You must call `KSPSetUp()` before calling `PCASMGetSubKSP()`.
1213: Fortran Notes:
1214: The output argument 'ksp' must be an array of sufficient length or `PETSC_NULL_KSP`. The latter can be used to learn the necessary length.
1216: .seealso: [](ch_ksp), `PCASM`, `PCASMSetTotalSubdomains()`, `PCASMSetOverlap()`,
1217: `PCASMCreateSubdomains2D()`,
1218: @*/
1219: PetscErrorCode PCASMGetSubKSP(PC pc, PetscInt *n_local, PetscInt *first_local, KSP *ksp[])
1220: {
1221: PetscFunctionBegin;
1223: PetscUseMethod(pc, "PCASMGetSubKSP_C", (PC, PetscInt *, PetscInt *, KSP **), (pc, n_local, first_local, ksp));
1224: PetscFunctionReturn(PETSC_SUCCESS);
1225: }
1227: /*MC
1228: PCASM - Use the (restricted) additive Schwarz method, each block is (approximately) solved with
1229: its own `KSP` object, {cite}`dryja1987additive` and {cite}`1sbg`
1231: Options Database Keys:
1232: + -pc_asm_blocks <blks> - Sets total blocks. Defaults to one block per MPI process.
1233: . -pc_asm_overlap <ovl> - Sets overlap
1234: . -pc_asm_type [basic,restrict,interpolate,none] - Sets `PCASMType`, default is restrict. See `PCASMSetType()`
1235: . -pc_asm_dm_subdomains <bool> - use subdomains defined by the `DM` with `DMCreateDomainDecomposition()`
1236: - -pc_asm_local_type [additive, multiplicative] - Sets `PCCompositeType`, default is additive. See `PCASMSetLocalType()`
1238: Level: beginner
1240: Notes:
1241: If you run with, for example, 3 blocks on 1 processor or 3 blocks on 3 processors you
1242: will get a different convergence rate due to the default option of `-pc_asm_type restrict`. Use
1243: `-pc_asm_type basic` to get the same convergence behavior
1245: Each processor can have one or more blocks, but a block cannot be shared by more
1246: than one processor. Use `PCGASM` for subdomains shared by multiple processes.
1248: To set options on the solvers for each block append `-sub_` to all the `KSP`, and `PC`
1249: options database keys. For example, `-sub_pc_type ilu -sub_pc_factor_levels 1 -sub_ksp_type preonly`
1251: To set the options on the solvers separate for each block call `PCASMGetSubKSP()`
1252: and set the options directly on the resulting `KSP` object (you can access its `PC` with `KSPGetPC()`)
1254: If the `PC` has an associated `DM`, then, by default, `DMCreateDomainDecomposition()` is used to create the subdomains
1256: .seealso: [](ch_ksp), `PCCreate()`, `PCSetType()`, `PCType`, `PC`, `PCASMType`, `PCCompositeType`,
1257: `PCBJACOBI`, `PCASMGetSubKSP()`, `PCASMSetLocalSubdomains()`, `PCASMType`, `PCASMGetType()`, `PCASMSetLocalType()`, `PCASMGetLocalType()`
1258: `PCASMSetTotalSubdomains()`, `PCSetModifySubMatrices()`, `PCASMSetOverlap()`, `PCASMSetType()`, `PCCompositeType`
1259: M*/
1261: PETSC_EXTERN PetscErrorCode PCCreate_ASM(PC pc)
1262: {
1263: PC_ASM *osm;
1265: PetscFunctionBegin;
1266: PetscCall(PetscNew(&osm));
1268: osm->n = PETSC_DECIDE;
1269: osm->n_local = 0;
1270: osm->n_local_true = PETSC_DECIDE;
1271: osm->overlap = 1;
1272: osm->ksp = NULL;
1273: osm->restriction = NULL;
1274: osm->lprolongation = NULL;
1275: osm->lrestriction = NULL;
1276: osm->x = NULL;
1277: osm->y = NULL;
1278: osm->is = NULL;
1279: osm->is_local = NULL;
1280: osm->mat = NULL;
1281: osm->pmat = NULL;
1282: osm->type = PC_ASM_RESTRICT;
1283: osm->loctype = PC_COMPOSITE_ADDITIVE;
1284: osm->sort_indices = PETSC_TRUE;
1285: osm->dm_subdomains = PETSC_FALSE;
1286: osm->sub_mat_type = NULL;
1288: pc->data = (void *)osm;
1289: pc->ops->apply = PCApply_ASM;
1290: pc->ops->matapply = PCMatApply_ASM;
1291: pc->ops->applytranspose = PCApplyTranspose_ASM;
1292: pc->ops->setup = PCSetUp_ASM;
1293: pc->ops->reset = PCReset_ASM;
1294: pc->ops->destroy = PCDestroy_ASM;
1295: pc->ops->setfromoptions = PCSetFromOptions_ASM;
1296: pc->ops->setuponblocks = PCSetUpOnBlocks_ASM;
1297: pc->ops->view = PCView_ASM;
1298: pc->ops->applyrichardson = NULL;
1300: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCASMSetLocalSubdomains_C", PCASMSetLocalSubdomains_ASM));
1301: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCASMSetTotalSubdomains_C", PCASMSetTotalSubdomains_ASM));
1302: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCASMSetOverlap_C", PCASMSetOverlap_ASM));
1303: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCASMSetType_C", PCASMSetType_ASM));
1304: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCASMGetType_C", PCASMGetType_ASM));
1305: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCASMSetLocalType_C", PCASMSetLocalType_ASM));
1306: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCASMGetLocalType_C", PCASMGetLocalType_ASM));
1307: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCASMSetSortIndices_C", PCASMSetSortIndices_ASM));
1308: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCASMGetSubKSP_C", PCASMGetSubKSP_ASM));
1309: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCASMGetSubMatType_C", PCASMGetSubMatType_ASM));
1310: PetscCall(PetscObjectComposeFunction((PetscObject)pc, "PCASMSetSubMatType_C", PCASMSetSubMatType_ASM));
1311: PetscFunctionReturn(PETSC_SUCCESS);
1312: }
1314: /*@C
1315: PCASMCreateSubdomains - Creates the index sets for the overlapping Schwarz
1316: preconditioner, `PCASM`, for any problem on a general grid.
1318: Collective
1320: Input Parameters:
1321: + A - The global matrix operator
1322: - n - the number of local blocks
1324: Output Parameter:
1325: . outis - the array of index sets defining the subdomains
1327: Level: advanced
1329: Note:
1330: This generates nonoverlapping subdomains; the `PCASM` will generate the overlap
1331: from these if you use `PCASMSetLocalSubdomains()`
1333: Fortran Notes:
1334: You must provide the array outis[] already allocated of length n.
1336: .seealso: [](ch_ksp), `PCASM`, `PCASMSetLocalSubdomains()`, `PCASMDestroySubdomains()`
1337: @*/
1338: PetscErrorCode PCASMCreateSubdomains(Mat A, PetscInt n, IS *outis[])
1339: {
1340: MatPartitioning mpart;
1341: const char *prefix;
1342: PetscInt i, j, rstart, rend, bs;
1343: PetscBool hasop, isbaij = PETSC_FALSE, foundpart = PETSC_FALSE;
1344: Mat Ad = NULL, adj;
1345: IS ispart, isnumb, *is;
1347: PetscFunctionBegin;
1349: PetscAssertPointer(outis, 3);
1350: PetscCheck(n >= 1, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "number of local blocks must be > 0, n = %" PetscInt_FMT, n);
1352: /* Get prefix, row distribution, and block size */
1353: PetscCall(MatGetOptionsPrefix(A, &prefix));
1354: PetscCall(MatGetOwnershipRange(A, &rstart, &rend));
1355: PetscCall(MatGetBlockSize(A, &bs));
1356: PetscCheck(rstart / bs * bs == rstart && rend / bs * bs == rend, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "bad row distribution [%" PetscInt_FMT ",%" PetscInt_FMT ") for matrix block size %" PetscInt_FMT, rstart, rend, bs);
1358: /* Get diagonal block from matrix if possible */
1359: PetscCall(MatHasOperation(A, MATOP_GET_DIAGONAL_BLOCK, &hasop));
1360: if (hasop) PetscCall(MatGetDiagonalBlock(A, &Ad));
1361: if (Ad) {
1362: PetscCall(PetscObjectBaseTypeCompare((PetscObject)Ad, MATSEQBAIJ, &isbaij));
1363: if (!isbaij) PetscCall(PetscObjectBaseTypeCompare((PetscObject)Ad, MATSEQSBAIJ, &isbaij));
1364: }
1365: if (Ad && n > 1) {
1366: PetscBool match, done;
1367: /* Try to setup a good matrix partitioning if available */
1368: PetscCall(MatPartitioningCreate(PETSC_COMM_SELF, &mpart));
1369: PetscCall(PetscObjectSetOptionsPrefix((PetscObject)mpart, prefix));
1370: PetscCall(MatPartitioningSetFromOptions(mpart));
1371: PetscCall(PetscObjectTypeCompare((PetscObject)mpart, MATPARTITIONINGCURRENT, &match));
1372: if (!match) PetscCall(PetscObjectTypeCompare((PetscObject)mpart, MATPARTITIONINGSQUARE, &match));
1373: if (!match) { /* assume a "good" partitioner is available */
1374: PetscInt na;
1375: const PetscInt *ia, *ja;
1376: PetscCall(MatGetRowIJ(Ad, 0, PETSC_TRUE, isbaij, &na, &ia, &ja, &done));
1377: if (done) {
1378: /* Build adjacency matrix by hand. Unfortunately a call to
1379: MatConvert(Ad,MATMPIADJ,MAT_INITIAL_MATRIX,&adj) will
1380: remove the block-aij structure and we cannot expect
1381: MatPartitioning to split vertices as we need */
1382: PetscInt i, j, len, nnz, cnt, *iia = NULL, *jja = NULL;
1383: const PetscInt *row;
1384: nnz = 0;
1385: for (i = 0; i < na; i++) { /* count number of nonzeros */
1386: len = ia[i + 1] - ia[i];
1387: row = ja + ia[i];
1388: for (j = 0; j < len; j++) {
1389: if (row[j] == i) { /* don't count diagonal */
1390: len--;
1391: break;
1392: }
1393: }
1394: nnz += len;
1395: }
1396: PetscCall(PetscMalloc1(na + 1, &iia));
1397: PetscCall(PetscMalloc1(nnz, &jja));
1398: nnz = 0;
1399: iia[0] = 0;
1400: for (i = 0; i < na; i++) { /* fill adjacency */
1401: cnt = 0;
1402: len = ia[i + 1] - ia[i];
1403: row = ja + ia[i];
1404: for (j = 0; j < len; j++) {
1405: if (row[j] != i) { /* if not diagonal */
1406: jja[nnz + cnt++] = row[j];
1407: }
1408: }
1409: nnz += cnt;
1410: iia[i + 1] = nnz;
1411: }
1412: /* Partitioning of the adjacency matrix */
1413: PetscCall(MatCreateMPIAdj(PETSC_COMM_SELF, na, na, iia, jja, NULL, &adj));
1414: PetscCall(MatPartitioningSetAdjacency(mpart, adj));
1415: PetscCall(MatPartitioningSetNParts(mpart, n));
1416: PetscCall(MatPartitioningApply(mpart, &ispart));
1417: PetscCall(ISPartitioningToNumbering(ispart, &isnumb));
1418: PetscCall(MatDestroy(&adj));
1419: foundpart = PETSC_TRUE;
1420: }
1421: PetscCall(MatRestoreRowIJ(Ad, 0, PETSC_TRUE, isbaij, &na, &ia, &ja, &done));
1422: }
1423: PetscCall(MatPartitioningDestroy(&mpart));
1424: }
1426: PetscCall(PetscMalloc1(n, &is));
1427: *outis = is;
1429: if (!foundpart) {
1430: /* Partitioning by contiguous chunks of rows */
1432: PetscInt mbs = (rend - rstart) / bs;
1433: PetscInt start = rstart;
1434: for (i = 0; i < n; i++) {
1435: PetscInt count = (mbs / n + ((mbs % n) > i)) * bs;
1436: PetscCall(ISCreateStride(PETSC_COMM_SELF, count, start, 1, &is[i]));
1437: start += count;
1438: }
1440: } else {
1441: /* Partitioning by adjacency of diagonal block */
1443: const PetscInt *numbering;
1444: PetscInt *count, nidx, *indices, *newidx, start = 0;
1445: /* Get node count in each partition */
1446: PetscCall(PetscMalloc1(n, &count));
1447: PetscCall(ISPartitioningCount(ispart, n, count));
1448: if (isbaij && bs > 1) { /* adjust for the block-aij case */
1449: for (i = 0; i < n; i++) count[i] *= bs;
1450: }
1451: /* Build indices from node numbering */
1452: PetscCall(ISGetLocalSize(isnumb, &nidx));
1453: PetscCall(PetscMalloc1(nidx, &indices));
1454: for (i = 0; i < nidx; i++) indices[i] = i; /* needs to be initialized */
1455: PetscCall(ISGetIndices(isnumb, &numbering));
1456: PetscCall(PetscSortIntWithPermutation(nidx, numbering, indices));
1457: PetscCall(ISRestoreIndices(isnumb, &numbering));
1458: if (isbaij && bs > 1) { /* adjust for the block-aij case */
1459: PetscCall(PetscMalloc1(nidx * bs, &newidx));
1460: for (i = 0; i < nidx; i++) {
1461: for (j = 0; j < bs; j++) newidx[i * bs + j] = indices[i] * bs + j;
1462: }
1463: PetscCall(PetscFree(indices));
1464: nidx *= bs;
1465: indices = newidx;
1466: }
1467: /* Shift to get global indices */
1468: for (i = 0; i < nidx; i++) indices[i] += rstart;
1470: /* Build the index sets for each block */
1471: for (i = 0; i < n; i++) {
1472: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, count[i], &indices[start], PETSC_COPY_VALUES, &is[i]));
1473: PetscCall(ISSort(is[i]));
1474: start += count[i];
1475: }
1477: PetscCall(PetscFree(count));
1478: PetscCall(PetscFree(indices));
1479: PetscCall(ISDestroy(&isnumb));
1480: PetscCall(ISDestroy(&ispart));
1481: }
1482: PetscFunctionReturn(PETSC_SUCCESS);
1483: }
1485: /*@C
1486: PCASMDestroySubdomains - Destroys the index sets created with
1487: `PCASMCreateSubdomains()`. Should be called after setting subdomains with `PCASMSetLocalSubdomains()`.
1489: Collective
1491: Input Parameters:
1492: + n - the number of index sets
1493: . is - the array of index sets
1494: - is_local - the array of local index sets, can be `NULL`
1496: Level: advanced
1498: .seealso: [](ch_ksp), `PCASM`, `PCASMCreateSubdomains()`, `PCASMSetLocalSubdomains()`
1499: @*/
1500: PetscErrorCode PCASMDestroySubdomains(PetscInt n, IS is[], IS is_local[])
1501: {
1502: PetscInt i;
1504: PetscFunctionBegin;
1505: if (n <= 0) PetscFunctionReturn(PETSC_SUCCESS);
1506: if (is) {
1507: PetscAssertPointer(is, 2);
1508: for (i = 0; i < n; i++) PetscCall(ISDestroy(&is[i]));
1509: PetscCall(PetscFree(is));
1510: }
1511: if (is_local) {
1512: PetscAssertPointer(is_local, 3);
1513: for (i = 0; i < n; i++) PetscCall(ISDestroy(&is_local[i]));
1514: PetscCall(PetscFree(is_local));
1515: }
1516: PetscFunctionReturn(PETSC_SUCCESS);
1517: }
1519: /*@C
1520: PCASMCreateSubdomains2D - Creates the index sets for the overlapping Schwarz
1521: preconditioner, `PCASM`, for a two-dimensional problem on a regular grid.
1523: Not Collective
1525: Input Parameters:
1526: + m - the number of mesh points in the x direction
1527: . n - the number of mesh points in the y direction
1528: . M - the number of subdomains in the x direction
1529: . N - the number of subdomains in the y direction
1530: . dof - degrees of freedom per node
1531: - overlap - overlap in mesh lines
1533: Output Parameters:
1534: + Nsub - the number of subdomains created
1535: . is - array of index sets defining overlapping (if overlap > 0) subdomains
1536: - is_local - array of index sets defining non-overlapping subdomains
1538: Level: advanced
1540: Note:
1541: Presently `PCAMSCreateSubdomains2d()` is valid only for sequential
1542: preconditioners. More general related routines are
1543: `PCASMSetTotalSubdomains()` and `PCASMSetLocalSubdomains()`.
1545: Fortran Notes:
1546: The `IS` must be declared as an array of length long enough to hold `Nsub` entries
1548: .seealso: [](ch_ksp), `PCASM`, `PCASMSetTotalSubdomains()`, `PCASMSetLocalSubdomains()`, `PCASMGetSubKSP()`,
1549: `PCASMSetOverlap()`
1550: @*/
1551: PetscErrorCode PCASMCreateSubdomains2D(PetscInt m, PetscInt n, PetscInt M, PetscInt N, PetscInt dof, PetscInt overlap, PetscInt *Nsub, IS **is, IS **is_local)
1552: {
1553: PetscInt i, j, height, width, ystart, xstart, yleft, yright, xleft, xright, loc_outer;
1554: PetscInt nidx, *idx, loc, ii, jj, count;
1556: PetscFunctionBegin;
1557: PetscCheck(dof == 1, PETSC_COMM_SELF, PETSC_ERR_SUP, "dof must be 1");
1559: *Nsub = N * M;
1560: PetscCall(PetscMalloc1(*Nsub, is));
1561: PetscCall(PetscMalloc1(*Nsub, is_local));
1562: ystart = 0;
1563: loc_outer = 0;
1564: for (i = 0; i < N; i++) {
1565: height = n / N + ((n % N) > i); /* height of subdomain */
1566: PetscCheck(height >= 2, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Too many N subdomains for mesh dimension n");
1567: yleft = ystart - overlap;
1568: if (yleft < 0) yleft = 0;
1569: yright = ystart + height + overlap;
1570: if (yright > n) yright = n;
1571: xstart = 0;
1572: for (j = 0; j < M; j++) {
1573: width = m / M + ((m % M) > j); /* width of subdomain */
1574: PetscCheck(width >= 2, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Too many M subdomains for mesh dimension m");
1575: xleft = xstart - overlap;
1576: if (xleft < 0) xleft = 0;
1577: xright = xstart + width + overlap;
1578: if (xright > m) xright = m;
1579: nidx = (xright - xleft) * (yright - yleft);
1580: PetscCall(PetscMalloc1(nidx, &idx));
1581: loc = 0;
1582: for (ii = yleft; ii < yright; ii++) {
1583: count = m * ii + xleft;
1584: for (jj = xleft; jj < xright; jj++) idx[loc++] = count++;
1585: }
1586: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, nidx, idx, PETSC_COPY_VALUES, (*is) + loc_outer));
1587: if (overlap == 0) {
1588: PetscCall(PetscObjectReference((PetscObject)(*is)[loc_outer]));
1590: (*is_local)[loc_outer] = (*is)[loc_outer];
1591: } else {
1592: for (loc = 0, ii = ystart; ii < ystart + height; ii++) {
1593: for (jj = xstart; jj < xstart + width; jj++) idx[loc++] = m * ii + jj;
1594: }
1595: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, loc, idx, PETSC_COPY_VALUES, *is_local + loc_outer));
1596: }
1597: PetscCall(PetscFree(idx));
1598: xstart += width;
1599: loc_outer++;
1600: }
1601: ystart += height;
1602: }
1603: for (i = 0; i < *Nsub; i++) PetscCall(ISSort((*is)[i]));
1604: PetscFunctionReturn(PETSC_SUCCESS);
1605: }
1607: /*@C
1608: PCASMGetLocalSubdomains - Gets the local subdomains (for this processor
1609: only) for the additive Schwarz preconditioner, `PCASM`.
1611: Not Collective
1613: Input Parameter:
1614: . pc - the preconditioner context
1616: Output Parameters:
1617: + n - if requested, the number of subdomains for this processor (default value = 1)
1618: . is - if requested, the index sets that define the subdomains for this processor
1619: - is_local - if requested, the index sets that define the local part of the subdomains for this processor (can be `NULL`)
1621: Level: advanced
1623: Note:
1624: The `IS` numbering is in the parallel, global numbering of the vector.
1626: .seealso: [](ch_ksp), `PCASM`, `PCASMSetTotalSubdomains()`, `PCASMSetOverlap()`, `PCASMGetSubKSP()`,
1627: `PCASMCreateSubdomains2D()`, `PCASMSetLocalSubdomains()`, `PCASMGetLocalSubmatrices()`
1628: @*/
1629: PetscErrorCode PCASMGetLocalSubdomains(PC pc, PetscInt *n, IS *is[], IS *is_local[])
1630: {
1631: PC_ASM *osm = (PC_ASM *)pc->data;
1632: PetscBool match;
1634: PetscFunctionBegin;
1636: if (n) PetscAssertPointer(n, 2);
1637: if (is) PetscAssertPointer(is, 3);
1638: if (is_local) PetscAssertPointer(is_local, 4);
1639: PetscCall(PetscObjectTypeCompare((PetscObject)pc, PCASM, &match));
1640: PetscCheck(match, PetscObjectComm((PetscObject)pc), PETSC_ERR_ARG_WRONG, "PC is not a PCASM");
1641: if (n) *n = osm->n_local_true;
1642: if (is) *is = osm->is;
1643: if (is_local) *is_local = osm->is_local;
1644: PetscFunctionReturn(PETSC_SUCCESS);
1645: }
1647: /*@C
1648: PCASMGetLocalSubmatrices - Gets the local submatrices (for this processor
1649: only) for the additive Schwarz preconditioner, `PCASM`.
1651: Not Collective
1653: Input Parameter:
1654: . pc - the preconditioner context
1656: Output Parameters:
1657: + n - if requested, the number of matrices for this processor (default value = 1)
1658: - mat - if requested, the matrices
1660: Level: advanced
1662: Notes:
1663: Call after `PCSetUp()` (or `KSPSetUp()`) but before `PCApply()` and before `PCSetUpOnBlocks()`)
1665: Usually one would use `PCSetModifySubMatrices()` to change the submatrices in building the preconditioner.
1667: .seealso: [](ch_ksp), `PCASM`, `PCASMSetTotalSubdomains()`, `PCASMSetOverlap()`, `PCASMGetSubKSP()`,
1668: `PCASMCreateSubdomains2D()`, `PCASMSetLocalSubdomains()`, `PCASMGetLocalSubdomains()`, `PCSetModifySubMatrices()`
1669: @*/
1670: PetscErrorCode PCASMGetLocalSubmatrices(PC pc, PetscInt *n, Mat *mat[])
1671: {
1672: PC_ASM *osm;
1673: PetscBool match;
1675: PetscFunctionBegin;
1677: if (n) PetscAssertPointer(n, 2);
1678: if (mat) PetscAssertPointer(mat, 3);
1679: PetscCheck(pc->setupcalled, PetscObjectComm((PetscObject)pc), PETSC_ERR_ARG_WRONGSTATE, "Must call after KSPSetUp() or PCSetUp().");
1680: PetscCall(PetscObjectTypeCompare((PetscObject)pc, PCASM, &match));
1681: if (!match) {
1682: if (n) *n = 0;
1683: if (mat) *mat = NULL;
1684: } else {
1685: osm = (PC_ASM *)pc->data;
1686: if (n) *n = osm->n_local_true;
1687: if (mat) *mat = osm->pmat;
1688: }
1689: PetscFunctionReturn(PETSC_SUCCESS);
1690: }
1692: /*@
1693: PCASMSetDMSubdomains - Indicates whether to use `DMCreateDomainDecomposition()` to define the subdomains, whenever possible.
1695: Logically Collective
1697: Input Parameters:
1698: + pc - the preconditioner
1699: - flg - boolean indicating whether to use subdomains defined by the `DM`
1701: Options Database Key:
1702: . -pc_asm_dm_subdomains <bool> - use subdomains defined by the `DM` with `DMCreateDomainDecomposition()`
1704: Level: intermediate
1706: Note:
1707: `PCASMSetTotalSubdomains()` and `PCASMSetOverlap()` take precedence over `PCASMSetDMSubdomains()`,
1708: so setting either of the first two effectively turns the latter off.
1710: Developer Note:
1711: This should be `PCASMSetUseDMSubdomains()`, similarly for the options database key
1713: .seealso: [](ch_ksp), `PCASM`, `PCASMGetDMSubdomains()`, `PCASMSetTotalSubdomains()`, `PCASMSetOverlap()`
1714: `PCASMCreateSubdomains2D()`, `PCASMSetLocalSubdomains()`, `PCASMGetLocalSubdomains()`
1715: @*/
1716: PetscErrorCode PCASMSetDMSubdomains(PC pc, PetscBool flg)
1717: {
1718: PC_ASM *osm = (PC_ASM *)pc->data;
1719: PetscBool match;
1721: PetscFunctionBegin;
1724: PetscCheck(!pc->setupcalled, ((PetscObject)pc)->comm, PETSC_ERR_ARG_WRONGSTATE, "Not for a setup PC.");
1725: PetscCall(PetscObjectTypeCompare((PetscObject)pc, PCASM, &match));
1726: if (match) osm->dm_subdomains = flg;
1727: PetscFunctionReturn(PETSC_SUCCESS);
1728: }
1730: /*@
1731: PCASMGetDMSubdomains - Returns flag indicating whether to use `DMCreateDomainDecomposition()` to define the subdomains, whenever possible.
1733: Not Collective
1735: Input Parameter:
1736: . pc - the preconditioner
1738: Output Parameter:
1739: . flg - boolean indicating whether to use subdomains defined by the `DM`
1741: Level: intermediate
1743: Developer Note:
1744: This should be `PCASMSetUseDMSubdomains()`
1746: .seealso: [](ch_ksp), `PCASM`, `PCASMSetDMSubdomains()`, `PCASMSetTotalSubdomains()`, `PCASMSetOverlap()`
1747: `PCASMCreateSubdomains2D()`, `PCASMSetLocalSubdomains()`, `PCASMGetLocalSubdomains()`
1748: @*/
1749: PetscErrorCode PCASMGetDMSubdomains(PC pc, PetscBool *flg)
1750: {
1751: PC_ASM *osm = (PC_ASM *)pc->data;
1752: PetscBool match;
1754: PetscFunctionBegin;
1756: PetscAssertPointer(flg, 2);
1757: PetscCall(PetscObjectTypeCompare((PetscObject)pc, PCASM, &match));
1758: if (match) *flg = osm->dm_subdomains;
1759: else *flg = PETSC_FALSE;
1760: PetscFunctionReturn(PETSC_SUCCESS);
1761: }
1763: /*@C
1764: PCASMGetSubMatType - Gets the matrix type used for `PCASM` subsolves, as a string.
1766: Not Collective
1768: Input Parameter:
1769: . pc - the `PC`
1771: Output Parameter:
1772: . sub_mat_type - name of matrix type
1774: Level: advanced
1776: .seealso: [](ch_ksp), `PCASM`, `PCASMSetSubMatType()`, `PCSetType()`, `VecSetType()`, `MatType`, `Mat`
1777: @*/
1778: PetscErrorCode PCASMGetSubMatType(PC pc, MatType *sub_mat_type)
1779: {
1780: PetscFunctionBegin;
1782: PetscTryMethod(pc, "PCASMGetSubMatType_C", (PC, MatType *), (pc, sub_mat_type));
1783: PetscFunctionReturn(PETSC_SUCCESS);
1784: }
1786: /*@C
1787: PCASMSetSubMatType - Set the type of matrix used for `PCASM` subsolves
1789: Collective
1791: Input Parameters:
1792: + pc - the `PC` object
1793: - sub_mat_type - the `MatType`
1795: Options Database Key:
1796: . -pc_asm_sub_mat_type <sub_mat_type> - Sets the matrix type used for subsolves, for example, seqaijviennacl.
1797: If you specify a base name like aijviennacl, the corresponding sequential type is assumed.
1799: Note:
1800: See `MatType` for available types
1802: Level: advanced
1804: .seealso: [](ch_ksp), `PCASM`, `PCASMGetSubMatType()`, `PCSetType()`, `VecSetType()`, `MatType`, `Mat`
1805: @*/
1806: PetscErrorCode PCASMSetSubMatType(PC pc, MatType sub_mat_type)
1807: {
1808: PetscFunctionBegin;
1810: PetscTryMethod(pc, "PCASMSetSubMatType_C", (PC, MatType), (pc, sub_mat_type));
1811: PetscFunctionReturn(PETSC_SUCCESS);
1812: }