Actual source code: hpddm.cxx
1: #define HPDDM_MIXED_PRECISION 1
2: #include <petsc/private/petschpddm.h>
4: const char *const KSPHPDDMTypes[] = {KSPGMRES, "bgmres", KSPCG, "bcg", "gcrodr", "bgcrodr", "bfbcg", KSPPREONLY};
5: const char *const KSPHPDDMPrecisionTypes[] = {"HALF", "SINGLE", "DOUBLE", "QUADRUPLE", "KSPHPDDMPrecisionType", "KSP_HPDDM_PRECISION_", NULL};
6: const char *const HPDDMOrthogonalization[] = {"cgs", "mgs"};
7: const char *const HPDDMQR[] = {"cholqr", "cgs", "mgs"};
8: const char *const HPDDMVariant[] = {"left", "right", "flexible"};
9: const char *const HPDDMRecycleTarget[] = {"SM", "LM", "SR", "LR", "SI", "LI"};
10: const char *const HPDDMRecycleStrategy[] = {"A", "B"};
12: PetscBool HPDDMCite = PETSC_FALSE;
13: const char HPDDMCitation[] = "@article{jolivet2020petsc,\n"
14: " Author = {Jolivet, Pierre and Roman, Jose E. and Zampini, Stefano},\n"
15: " Title = {{KSPHPDDM} and {PCHPDDM}: Extending {PETSc} with Robust Overlapping {Schwarz} Preconditioners and Advanced {Krylov} Methods},\n"
16: " Year = {2021},\n"
17: " Publisher = {Elsevier},\n"
18: " Journal = {Computer \\& Mathematics with Applications},\n"
19: " Volume = {84},\n"
20: " Pages = {277--295},\n"
21: " Url = {https://github.com/prj-/jolivet2020petsc}\n"
22: "}\n";
24: #if defined(PETSC_HAVE_SLEPC) && defined(PETSC_USE_SHARED_LIBRARIES)
25: static PetscBool loadedDL = PETSC_FALSE;
26: #endif
28: static PetscErrorCode KSPSetFromOptions_HPDDM(KSP ksp, PetscOptionItems *PetscOptionsObject)
29: {
30: KSP_HPDDM *data = (KSP_HPDDM *)ksp->data;
31: PetscInt i, j;
32: PetscMPIInt size;
34: PetscOptionsHeadBegin(PetscOptionsObject, "KSPHPDDM options, cf. https://github.com/hpddm/hpddm");
35: i = (data->cntl[0] == static_cast<char>(PETSC_DECIDE) ? HPDDM_KRYLOV_METHOD_GMRES : data->cntl[0]);
36: PetscOptionsEList("-ksp_hpddm_type", "Type of Krylov method", "KSPHPDDMGetType", KSPHPDDMTypes, PETSC_STATIC_ARRAY_LENGTH(KSPHPDDMTypes), KSPHPDDMTypes[HPDDM_KRYLOV_METHOD_GMRES], &i, NULL);
37: if (i == PETSC_STATIC_ARRAY_LENGTH(KSPHPDDMTypes) - 1) i = HPDDM_KRYLOV_METHOD_NONE; /* need to shift the value since HPDDM_KRYLOV_METHOD_RICHARDSON is not registered in PETSc */
38: data->cntl[0] = i;
39: PetscOptionsEnum("-ksp_hpddm_precision", "Precision in which Krylov bases are stored", "KSPHPDDM", KSPHPDDMPrecisionTypes, (PetscEnum)data->precision, (PetscEnum *)&data->precision, NULL);
42: if (data->cntl[0] != HPDDM_KRYLOV_METHOD_NONE) {
43: if (data->cntl[0] != HPDDM_KRYLOV_METHOD_BCG && data->cntl[0] != HPDDM_KRYLOV_METHOD_BFBCG) {
44: i = (data->cntl[1] == static_cast<char>(PETSC_DECIDE) ? HPDDM_VARIANT_LEFT : data->cntl[1]);
45: if (ksp->pc_side_set == PC_SIDE_DEFAULT)
46: PetscOptionsEList("-ksp_hpddm_variant", "Left, right, or variable preconditioning", "KSPHPDDM", HPDDMVariant, PETSC_STATIC_ARRAY_LENGTH(HPDDMVariant), HPDDMVariant[HPDDM_VARIANT_LEFT], &i, NULL);
47: else if (ksp->pc_side_set == PC_RIGHT) i = HPDDM_VARIANT_RIGHT;
48: data->cntl[1] = i;
49: if (i > 0) KSPSetPCSide(ksp, PC_RIGHT);
50: }
51: if (data->cntl[0] == HPDDM_KRYLOV_METHOD_BGMRES || data->cntl[0] == HPDDM_KRYLOV_METHOD_BGCRODR || data->cntl[0] == HPDDM_KRYLOV_METHOD_BFBCG) {
52: data->rcntl[0] = (PetscAbsReal(data->rcntl[0] - static_cast<PetscReal>(PETSC_DECIDE)) < PETSC_SMALL ? -1.0 : data->rcntl[0]);
53: PetscOptionsReal("-ksp_hpddm_deflation_tol", "Tolerance when deflating right-hand sides inside block methods", "KSPHPDDM", data->rcntl[0], data->rcntl, NULL);
54: i = (data->scntl[data->cntl[0] != HPDDM_KRYLOV_METHOD_BFBCG] == static_cast<unsigned short>(PETSC_DECIDE) ? 1 : PetscMax(1, data->scntl[data->cntl[0] != HPDDM_KRYLOV_METHOD_BFBCG]));
55: PetscOptionsRangeInt("-ksp_hpddm_enlarge_krylov_subspace", "Split the initial right-hand side into multiple vectors", "KSPHPDDM", i, &i, NULL, 1, std::numeric_limits<unsigned short>::max() - 1);
56: data->scntl[data->cntl[0] != HPDDM_KRYLOV_METHOD_BFBCG] = i;
57: } else data->scntl[data->cntl[0] != HPDDM_KRYLOV_METHOD_BCG] = 0;
58: if (data->cntl[0] == HPDDM_KRYLOV_METHOD_GMRES || data->cntl[0] == HPDDM_KRYLOV_METHOD_BGMRES || data->cntl[0] == HPDDM_KRYLOV_METHOD_GCRODR || data->cntl[0] == HPDDM_KRYLOV_METHOD_BGCRODR) {
59: i = (data->cntl[2] == static_cast<char>(PETSC_DECIDE) ? HPDDM_ORTHOGONALIZATION_CGS : data->cntl[2] & 3);
60: PetscOptionsEList("-ksp_hpddm_orthogonalization", "Classical (faster) or Modified (more robust) Gram--Schmidt process", "KSPHPDDM", HPDDMOrthogonalization, PETSC_STATIC_ARRAY_LENGTH(HPDDMOrthogonalization), HPDDMOrthogonalization[HPDDM_ORTHOGONALIZATION_CGS], &i, NULL);
61: j = (data->cntl[2] == static_cast<char>(PETSC_DECIDE) ? HPDDM_QR_CHOLQR : ((data->cntl[2] >> 2) & 7));
62: PetscOptionsEList("-ksp_hpddm_qr", "Distributed QR factorizations computed with Cholesky QR, Classical or Modified Gram--Schmidt process", "KSPHPDDM", HPDDMQR, PETSC_STATIC_ARRAY_LENGTH(HPDDMQR), HPDDMQR[HPDDM_QR_CHOLQR], &j, NULL);
63: data->cntl[2] = static_cast<char>(i) + (static_cast<char>(j) << 2);
64: i = (data->scntl[0] == static_cast<unsigned short>(PETSC_DECIDE) ? PetscMin(30, ksp->max_it) : data->scntl[0]);
65: PetscOptionsRangeInt("-ksp_gmres_restart", "Maximum number of Arnoldi vectors generated per cycle", "KSPHPDDM", i, &i, NULL, PetscMin(1, ksp->max_it), PetscMin(ksp->max_it, std::numeric_limits<unsigned short>::max() - 1));
66: data->scntl[0] = i;
67: }
68: if (data->cntl[0] == HPDDM_KRYLOV_METHOD_BCG || data->cntl[0] == HPDDM_KRYLOV_METHOD_BFBCG) {
69: j = (data->cntl[1] == static_cast<char>(PETSC_DECIDE) ? HPDDM_QR_CHOLQR : data->cntl[1]);
70: PetscOptionsEList("-ksp_hpddm_qr", "Distributed QR factorizations computed with Cholesky QR, Classical or Modified Gram--Schmidt process", "KSPHPDDM", HPDDMQR, PETSC_STATIC_ARRAY_LENGTH(HPDDMQR), HPDDMQR[HPDDM_QR_CHOLQR], &j, NULL);
71: data->cntl[1] = j;
72: }
73: if (data->cntl[0] == HPDDM_KRYLOV_METHOD_GCRODR || data->cntl[0] == HPDDM_KRYLOV_METHOD_BGCRODR) {
74: i = (data->icntl[0] == static_cast<int>(PETSC_DECIDE) ? PetscMin(20, data->scntl[0] - 1) : data->icntl[0]);
75: PetscOptionsRangeInt("-ksp_hpddm_recycle", "Number of harmonic Ritz vectors to compute", "KSPHPDDM", i, &i, NULL, 1, data->scntl[0] - 1);
76: data->icntl[0] = i;
77: if (!PetscDefined(HAVE_SLEPC) || !PetscDefined(USE_SHARED_LIBRARIES) || data->cntl[0] == HPDDM_KRYLOV_METHOD_GCRODR) {
78: i = (data->cntl[3] == static_cast<char>(PETSC_DECIDE) ? HPDDM_RECYCLE_TARGET_SM : data->cntl[3]);
79: PetscOptionsEList("-ksp_hpddm_recycle_target", "Criterion to select harmonic Ritz vectors", "KSPHPDDM", HPDDMRecycleTarget, PETSC_STATIC_ARRAY_LENGTH(HPDDMRecycleTarget), HPDDMRecycleTarget[HPDDM_RECYCLE_TARGET_SM], &i, NULL);
80: data->cntl[3] = i;
81: } else {
83: MPI_Comm_size(PetscObjectComm((PetscObject)ksp), &size);
84: i = (data->cntl[3] == static_cast<char>(PETSC_DECIDE) ? 1 : data->cntl[3]);
85: PetscOptionsRangeInt("-ksp_hpddm_recycle_redistribute", "Number of processes used to solve eigenvalue problems when recycling in BGCRODR", "KSPHPDDM", i, &i, NULL, 1, PetscMin(size, 192));
86: data->cntl[3] = i;
87: }
88: i = (data->cntl[4] == static_cast<char>(PETSC_DECIDE) ? HPDDM_RECYCLE_STRATEGY_A : data->cntl[4]);
89: PetscOptionsEList("-ksp_hpddm_recycle_strategy", "Generalized eigenvalue problem to solve for recycling", "KSPHPDDM", HPDDMRecycleStrategy, PETSC_STATIC_ARRAY_LENGTH(HPDDMRecycleStrategy), HPDDMRecycleStrategy[HPDDM_RECYCLE_STRATEGY_A], &i, NULL);
90: data->cntl[4] = i;
91: }
92: } else {
93: data->cntl[0] = HPDDM_KRYLOV_METHOD_NONE;
94: data->scntl[1] = 1;
95: }
97: ksp->nmax);
98: data->icntl[1] = static_cast<int>(ksp->nmax);
99: PetscOptionsHeadEnd();
100: return 0;
101: }
103: static PetscErrorCode KSPView_HPDDM(KSP ksp, PetscViewer viewer)
104: {
105: KSP_HPDDM *data = (KSP_HPDDM *)ksp->data;
106: HPDDM::PETScOperator *op = data->op;
107: const PetscScalar *array = op ? op->storage() : NULL;
108: PetscBool ascii;
110: PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &ascii);
111: if (op && ascii) {
112: PetscViewerASCIIPrintf(viewer, "HPDDM type: %s\n", KSPHPDDMTypes[std::min(static_cast<PetscInt>(data->cntl[0]), static_cast<PetscInt>(PETSC_STATIC_ARRAY_LENGTH(KSPHPDDMTypes) - 1))]);
113: PetscViewerASCIIPrintf(viewer, "precision: %s\n", KSPHPDDMPrecisionTypes[data->precision]);
114: if (data->cntl[0] == HPDDM_KRYLOV_METHOD_BGMRES || data->cntl[0] == HPDDM_KRYLOV_METHOD_BGCRODR || data->cntl[0] == HPDDM_KRYLOV_METHOD_BFBCG) {
115: if (PetscAbsReal(data->rcntl[0] - static_cast<PetscReal>(PETSC_DECIDE)) < PETSC_SMALL) PetscViewerASCIIPrintf(viewer, "no deflation at restarts\n");
116: else PetscViewerASCIIPrintf(viewer, "deflation tolerance: %g\n", static_cast<double>(data->rcntl[0]));
117: }
118: if (data->cntl[0] == HPDDM_KRYLOV_METHOD_GCRODR || data->cntl[0] == HPDDM_KRYLOV_METHOD_BGCRODR) {
119: PetscViewerASCIIPrintf(viewer, "deflation subspace attached? %s\n", PetscBools[array ? PETSC_TRUE : PETSC_FALSE]);
120: if (!PetscDefined(HAVE_SLEPC) || !PetscDefined(USE_SHARED_LIBRARIES) || data->cntl[0] == HPDDM_KRYLOV_METHOD_GCRODR) PetscViewerASCIIPrintf(viewer, "deflation target: %s\n", HPDDMRecycleTarget[static_cast<PetscInt>(data->cntl[3])]);
121: else PetscViewerASCIIPrintf(viewer, "redistribution size: %d\n", static_cast<PetscMPIInt>(data->cntl[3]));
122: }
123: if (data->icntl[1] != static_cast<int>(PETSC_DECIDE)) PetscViewerASCIIPrintf(viewer, " block size is %d\n", data->icntl[1]);
124: }
125: return 0;
126: }
128: static PetscErrorCode KSPSetUp_HPDDM(KSP ksp)
129: {
130: KSP_HPDDM *data = (KSP_HPDDM *)ksp->data;
131: Mat A;
132: PetscInt n, bs;
133: PetscBool match;
135: KSPGetOperators(ksp, &A, NULL);
136: MatGetLocalSize(A, &n, NULL);
137: MatGetBlockSize(A, &bs);
138: PetscObjectTypeCompareAny((PetscObject)A, &match, MATSEQKAIJ, MATMPIKAIJ, "");
139: if (match) n /= bs;
140: data->op = new HPDDM::PETScOperator(ksp, n);
141: if (PetscUnlikely(!ksp->setfromoptionscalled || data->cntl[0] == static_cast<char>(PETSC_DECIDE))) { /* what follows is basically a copy/paste of KSPSetFromOptions_HPDDM, with no call to PetscOptions() */
142: PetscInfo(ksp, "KSPSetFromOptions() not called or uninitialized internal structure, hardwiring default KSPHPDDM options\n");
143: if (data->cntl[0] == static_cast<char>(PETSC_DECIDE)) data->cntl[0] = 0; /* GMRES by default */
144: if (data->cntl[0] != HPDDM_KRYLOV_METHOD_NONE) { /* following options do not matter with PREONLY */
145: if (data->cntl[0] != HPDDM_KRYLOV_METHOD_BCG && data->cntl[0] != HPDDM_KRYLOV_METHOD_BFBCG) {
146: data->cntl[1] = HPDDM_VARIANT_LEFT; /* left preconditioning by default */
147: if (ksp->pc_side_set == PC_RIGHT) data->cntl[1] = HPDDM_VARIANT_RIGHT;
148: if (data->cntl[1] > 0) KSPSetPCSide(ksp, PC_RIGHT);
149: }
150: if (data->cntl[0] == HPDDM_KRYLOV_METHOD_BGMRES || data->cntl[0] == HPDDM_KRYLOV_METHOD_BGCRODR || data->cntl[0] == HPDDM_KRYLOV_METHOD_BFBCG) {
151: data->rcntl[0] = -1.0; /* no deflation by default */
152: data->scntl[data->cntl[0] != HPDDM_KRYLOV_METHOD_BFBCG] = 1; /* Krylov subspace not enlarged by default */
153: } else data->scntl[data->cntl[0] != HPDDM_KRYLOV_METHOD_BCG] = 0;
154: if (data->cntl[0] == HPDDM_KRYLOV_METHOD_GMRES || data->cntl[0] == HPDDM_KRYLOV_METHOD_BGMRES || data->cntl[0] == HPDDM_KRYLOV_METHOD_GCRODR || data->cntl[0] == HPDDM_KRYLOV_METHOD_BGCRODR) {
155: data->cntl[2] = static_cast<char>(HPDDM_ORTHOGONALIZATION_CGS) + (static_cast<char>(HPDDM_QR_CHOLQR) << 2); /* CGS and CholQR by default */
156: data->scntl[0] = PetscMin(30, ksp->max_it); /* restart parameter of 30 by default */
157: }
158: if (data->cntl[0] == HPDDM_KRYLOV_METHOD_BCG || data->cntl[0] == HPDDM_KRYLOV_METHOD_BFBCG) { data->cntl[1] = HPDDM_QR_CHOLQR; /* CholQR by default */ }
159: if (data->cntl[0] == HPDDM_KRYLOV_METHOD_GCRODR || data->cntl[0] == HPDDM_KRYLOV_METHOD_BGCRODR) {
160: data->icntl[0] = PetscMin(20, data->scntl[0] - 1); /* recycled subspace of size 20 by default */
161: if (!PetscDefined(HAVE_SLEPC) || !PetscDefined(USE_SHARED_LIBRARIES) || data->cntl[0] == HPDDM_KRYLOV_METHOD_GCRODR) {
162: data->cntl[3] = HPDDM_RECYCLE_TARGET_SM; /* default recycling target */
163: } else {
164: data->cntl[3] = 1; /* redistribution parameter of 1 by default */
165: }
166: data->cntl[4] = HPDDM_RECYCLE_STRATEGY_A; /* default recycling strategy */
167: }
168: } else data->scntl[1] = 1;
169: }
171: ksp->nmax);
172: data->icntl[1] = static_cast<int>(ksp->nmax);
173: return 0;
174: }
176: static inline PetscErrorCode KSPHPDDMReset_Private(KSP ksp)
177: {
178: KSP_HPDDM *data = (KSP_HPDDM *)ksp->data;
180: /* cast PETSC_DECIDE into the appropriate types to avoid compiler warnings */
181: std::fill_n(data->rcntl, PETSC_STATIC_ARRAY_LENGTH(data->rcntl), static_cast<PetscReal>(PETSC_DECIDE));
182: std::fill_n(data->icntl, PETSC_STATIC_ARRAY_LENGTH(data->icntl), static_cast<int>(PETSC_DECIDE));
183: std::fill_n(data->scntl, PETSC_STATIC_ARRAY_LENGTH(data->scntl), static_cast<unsigned short>(PETSC_DECIDE));
184: std::fill_n(data->cntl, PETSC_STATIC_ARRAY_LENGTH(data->cntl), static_cast<char>(PETSC_DECIDE));
185: data->precision = PETSC_KSPHPDDM_DEFAULT_PRECISION;
186: return 0;
187: }
189: static PetscErrorCode KSPReset_HPDDM(KSP ksp)
190: {
191: KSP_HPDDM *data = (KSP_HPDDM *)ksp->data;
193: if (data->op) {
194: delete data->op;
195: data->op = NULL;
196: }
197: KSPHPDDMReset_Private(ksp);
198: return 0;
199: }
201: static PetscErrorCode KSPDestroy_HPDDM(KSP ksp)
202: {
203: KSPReset_HPDDM(ksp);
204: KSPDestroyDefault(ksp);
205: PetscObjectComposeFunction((PetscObject)ksp, "KSPHPDDMSetDeflationMat_C", NULL);
206: PetscObjectComposeFunction((PetscObject)ksp, "KSPHPDDMGetDeflationMat_C", NULL);
207: PetscObjectComposeFunction((PetscObject)ksp, "KSPHPDDMSetType_C", NULL);
208: PetscObjectComposeFunction((PetscObject)ksp, "KSPHPDDMGetType_C", NULL);
209: return 0;
210: }
212: static inline PetscErrorCode KSPSolve_HPDDM_Private(KSP ksp, const PetscScalar *b, PetscScalar *x, PetscInt n)
213: {
214: KSP_HPDDM *data = (KSP_HPDDM *)ksp->data;
215: KSPConvergedDefaultCtx *ctx = (KSPConvergedDefaultCtx *)ksp->cnvP;
216: const PetscInt N = data->op->getDof() * n;
217: PetscBool scale;
218: #if !PetscDefined(USE_REAL_DOUBLE) || PetscDefined(HAVE_F2CBLASLAPACK___FLOAT128_BINDINGS)
219: HPDDM::upscaled_type<PetscScalar> *high[2];
220: #endif
221: #if !PetscDefined(USE_REAL_SINGLE) || PetscDefined(HAVE_F2CBLASLAPACK___FP16_BINDINGS)
222: HPDDM::downscaled_type<PetscScalar> *low[2];
223: #endif
225: PCGetDiagonalScale(ksp->pc, &scale);
227: if (n > 1) {
228: if (ksp->converged == KSPConvergedDefault) {
230: if (!ctx->initialrtol) {
231: PetscInfo(ksp, "Forcing KSPConvergedDefaultSetUIRNorm() since KSPConvergedDefault() cannot handle multiple norms\n");
232: ctx->initialrtol = PETSC_TRUE;
233: }
234: } else PetscInfo(ksp, "Using a special \"converged\" callback, be careful, it is used in KSPHPDDM to track blocks of residuals\n");
235: }
236: /* initial guess is always nonzero with recycling methods if there is a deflation subspace available */
237: if ((data->cntl[0] == HPDDM_KRYLOV_METHOD_GCRODR || data->cntl[0] == HPDDM_KRYLOV_METHOD_BGCRODR) && data->op->storage()) ksp->guess_zero = PETSC_FALSE;
238: ksp->its = 0;
239: ksp->reason = KSP_CONVERGED_ITERATING;
240: if (data->precision > PETSC_KSPHPDDM_DEFAULT_PRECISION) { /* Krylov basis stored in higher precision than PetscScalar */
241: #if !PetscDefined(USE_REAL_DOUBLE) || PetscDefined(HAVE_F2CBLASLAPACK___FLOAT128_BINDINGS)
242: PetscMalloc2(N, high, N, high + 1);
243: HPDDM::copy_n(b, N, high[0]);
244: HPDDM::copy_n(x, N, high[1]);
245: HPDDM::IterativeMethod::solve(*data->op, high[0], high[1], n, PetscObjectComm((PetscObject)ksp));
246: HPDDM::copy_n(high[1], N, x);
247: PetscFree2(high[0], high[1]);
248: #else
250: #endif
251: } else if (data->precision < PETSC_KSPHPDDM_DEFAULT_PRECISION) { /* Krylov basis stored in lower precision than PetscScalar */
252: #if !PetscDefined(USE_REAL_SINGLE) || PetscDefined(HAVE_F2CBLASLAPACK___FP16_BINDINGS)
253: PetscMalloc1(N, low);
254: low[1] = reinterpret_cast<HPDDM::downscaled_type<PetscScalar> *>(x);
255: std::copy_n(b, N, low[0]);
256: for (PetscInt i = 0; i < N; ++i) low[1][i] = x[i];
257: HPDDM::IterativeMethod::solve(*data->op, low[0], low[1], n, PetscObjectComm((PetscObject)ksp));
258: if (N) {
259: low[0][0] = low[1][0];
260: std::copy_backward(low[1] + 1, low[1] + N, x + N);
261: x[0] = low[0][0];
262: }
263: PetscFree(low[0]);
264: #else
266: #endif
267: } else HPDDM::IterativeMethod::solve(*data->op, b, x, n, PetscObjectComm((PetscObject)ksp)); /* Krylov basis stored in the same precision as PetscScalar */
268: if (!ksp->reason) { /* KSPConvergedDefault() is still returning 0 (= KSP_CONVERGED_ITERATING) */
269: if (ksp->its >= ksp->max_it) ksp->reason = KSP_DIVERGED_ITS;
270: else ksp->reason = KSP_CONVERGED_RTOL; /* early exit by HPDDM, which only happens on breakdowns or convergence */
271: }
272: ksp->its = PetscMin(ksp->its, ksp->max_it);
273: return 0;
274: }
276: static PetscErrorCode KSPSolve_HPDDM(KSP ksp)
277: {
278: KSP_HPDDM *data = (KSP_HPDDM *)ksp->data;
279: Mat A, B;
280: PetscScalar *x, *bt = NULL, **ptr;
281: const PetscScalar *b;
282: PetscInt i, j, n;
283: PetscBool flg;
285: PetscCitationsRegister(HPDDMCitation, &HPDDMCite);
286: KSPGetOperators(ksp, &A, NULL);
287: PetscObjectTypeCompareAny((PetscObject)A, &flg, MATSEQKAIJ, MATMPIKAIJ, "");
288: VecGetArrayWrite(ksp->vec_sol, &x);
289: VecGetArrayRead(ksp->vec_rhs, &b);
290: if (!flg) KSPSolve_HPDDM_Private(ksp, b, x, 1);
291: else {
292: MatKAIJGetScaledIdentity(A, &flg);
293: MatKAIJGetAIJ(A, &B);
294: MatGetBlockSize(A, &n);
295: MatGetLocalSize(B, &i, NULL);
296: j = data->op->getDof();
297: if (!flg) i *= n; /* S and T are not scaled identities, cannot use block methods */
298: if (i != j) { /* switching between block and standard methods */
299: delete data->op;
300: data->op = new HPDDM::PETScOperator(ksp, i);
301: }
302: if (flg && n > 1) {
303: PetscMalloc1(i * n, &bt);
304: /* from row- to column-major to be consistent with HPDDM */
305: HPDDM::Wrapper<PetscScalar>::omatcopy<'T'>(i, n, b, n, bt, i);
306: ptr = const_cast<PetscScalar **>(&b);
307: std::swap(*ptr, bt);
308: HPDDM::Wrapper<PetscScalar>::imatcopy<'T'>(i, n, x, n, i);
309: }
310: KSPSolve_HPDDM_Private(ksp, b, x, flg ? n : 1);
311: if (flg && n > 1) {
312: std::swap(*ptr, bt);
313: PetscFree(bt);
314: /* from column- to row-major to be consistent with MatKAIJ format */
315: HPDDM::Wrapper<PetscScalar>::imatcopy<'T'>(n, i, x, i, n);
316: }
317: }
318: VecRestoreArrayRead(ksp->vec_rhs, &b);
319: VecRestoreArrayWrite(ksp->vec_sol, &x);
320: return 0;
321: }
323: /*@
324: KSPHPDDMSetDeflationMat - Sets the deflation space used by Krylov methods in `KSPHPDDM` with recycling. This space is viewed as a set of vectors stored in
325: a `MATDENSE` (column major).
327: Input Parameters:
328: + ksp - iterative context
329: - U - deflation space to be used during KSPSolve()
331: Level: intermediate
333: .seealso: [](chapter_ksp), `KSPHPDDM`, `KSPCreate()`, `KSPType`, `KSPHPDDMGetDeflationMat()`
334: @*/
335: PetscErrorCode KSPHPDDMSetDeflationMat(KSP ksp, Mat U)
336: {
340: PetscUseMethod(ksp, "KSPHPDDMSetDeflationMat_C", (KSP, Mat), (ksp, U));
341: return 0;
342: }
344: /*@
345: KSPHPDDMGetDeflationMat - Gets the deflation space computed by Krylov methods in `KSPHPDDM` with recycling or NULL if `KSPSolve()` has not been called yet.
346: This space is viewed as a set of vectors stored in a `MATDENSE` (column major). It is the responsibility of the user to free the returned `Mat`.
348: Input Parameter:
349: . ksp - iterative context
351: Output Parameter:
352: . U - deflation space generated during `KSPSolve()`
354: Level: intermediate
356: .seealso: [](chapter_ksp), `KSPHPDDM`, `KSPCreate()`, `KSPType`, `KSPHPDDMSetDeflationMat()`
357: @*/
358: PetscErrorCode KSPHPDDMGetDeflationMat(KSP ksp, Mat *U)
359: {
361: if (U) {
363: PetscUseMethod(ksp, "KSPHPDDMGetDeflationMat_C", (KSP, Mat *), (ksp, U));
364: }
365: return 0;
366: }
368: static PetscErrorCode KSPHPDDMSetDeflationMat_HPDDM(KSP ksp, Mat U)
369: {
370: KSP_HPDDM *data = (KSP_HPDDM *)ksp->data;
371: HPDDM::PETScOperator *op = data->op;
372: Mat A;
373: const PetscScalar *array;
374: PetscScalar *copy;
375: PetscInt m1, M1, m2, M2, n2, N2, ldu;
376: PetscBool match;
378: if (!op) {
379: KSPSetUp(ksp);
380: op = data->op;
381: }
383: KSPGetOperators(ksp, &A, NULL);
384: MatGetLocalSize(A, &m1, NULL);
385: MatGetLocalSize(U, &m2, &n2);
386: MatGetSize(A, &M1, NULL);
387: MatGetSize(U, &M2, &N2);
389: PetscObjectTypeCompareAny((PetscObject)U, &match, MATSEQDENSE, MATMPIDENSE, "");
391: MatDenseGetArrayRead(U, &array);
392: copy = op->allocate(m2, 1, N2);
394: MatDenseGetLDA(U, &ldu);
395: HPDDM::Wrapper<PetscScalar>::omatcopy<'N'>(N2, m2, array, ldu, copy, m2);
396: MatDenseRestoreArrayRead(U, &array);
397: return 0;
398: }
400: static PetscErrorCode KSPHPDDMGetDeflationMat_HPDDM(KSP ksp, Mat *U)
401: {
402: KSP_HPDDM *data = (KSP_HPDDM *)ksp->data;
403: HPDDM::PETScOperator *op = data->op;
404: Mat A;
405: const PetscScalar *array;
406: PetscScalar *copy;
407: PetscInt m1, M1, N2;
409: if (!op) {
410: KSPSetUp(ksp);
411: op = data->op;
412: }
414: array = op->storage();
415: N2 = op->k().first * op->k().second;
416: if (!array) *U = NULL;
417: else {
418: KSPGetOperators(ksp, &A, NULL);
419: MatGetLocalSize(A, &m1, NULL);
420: MatGetSize(A, &M1, NULL);
421: MatCreateDense(PetscObjectComm((PetscObject)ksp), m1, PETSC_DECIDE, M1, N2, NULL, U);
422: MatDenseGetArrayWrite(*U, ©);
423: PetscArraycpy(copy, array, m1 * N2);
424: MatDenseRestoreArrayWrite(*U, ©);
425: }
426: return 0;
427: }
429: static PetscErrorCode KSPMatSolve_HPDDM(KSP ksp, Mat B, Mat X)
430: {
431: KSP_HPDDM *data = (KSP_HPDDM *)ksp->data;
432: HPDDM::PETScOperator *op = data->op;
433: Mat A;
434: const PetscScalar *b;
435: PetscScalar *x;
436: PetscInt n, lda;
438: PetscCitationsRegister(HPDDMCitation, &HPDDMCite);
439: if (!op) {
440: KSPSetUp(ksp);
441: op = data->op;
442: }
443: KSPGetOperators(ksp, &A, NULL);
444: MatGetLocalSize(B, &n, NULL);
445: MatDenseGetLDA(B, &lda);
447: MatGetLocalSize(A, &n, NULL);
448: MatDenseGetLDA(X, &lda);
450: MatDenseGetArrayRead(B, &b);
451: MatDenseGetArrayWrite(X, &x);
452: MatGetSize(X, NULL, &n);
453: KSPSolve_HPDDM_Private(ksp, b, x, n);
454: MatDenseRestoreArrayWrite(X, &x);
455: MatDenseRestoreArrayRead(B, &b);
456: return 0;
457: }
459: /*@
460: KSPHPDDMSetType - Sets the type of Krylov method used in `KSPHPDDM`.
462: Collective
464: Input Parameters:
465: + ksp - iterative context
466: - type - any of gmres, bgmres, cg, bcg, gcrodr, bgcrodr, bfbcg, or preonly
468: Level: intermediate
470: Notes:
471: Unlike `KSPReset()`, this function does not destroy any deflation space attached to the `KSP`.
473: As an example, in the following sequence:
474: .vb
475: KSPHPDDMSetType(ksp, KSPGCRODR);
476: KSPSolve(ksp, b, x);
477: KSPHPDDMSetType(ksp, KSPGMRES);
478: KSPHPDDMSetType(ksp, KSPGCRODR);
479: KSPSolve(ksp, b, x);
480: .ve
481: the recycled space is reused in the second `KSPSolve()`.
483: .seealso: [](chapter_ksp), `KSPCreate()`, `KSPType`, `KSPHPDDMType`, `KSPHPDDMGetType()`
484: @*/
485: PetscErrorCode KSPHPDDMSetType(KSP ksp, KSPHPDDMType type)
486: {
489: PetscUseMethod(ksp, "KSPHPDDMSetType_C", (KSP, KSPHPDDMType), (ksp, type));
490: return 0;
491: }
493: /*@
494: KSPHPDDMGetType - Gets the type of Krylov method used in `KSPHPDDM`.
496: Input Parameter:
497: . ksp - iterative context
499: Output Parameter:
500: . type - any of gmres, bgmres, cg, bcg, gcrodr, bgcrodr, bfbcg, or preonly
502: Level: intermediate
504: .seealso: [](chapter_ksp), `KSPCreate()`, `KSPType`, `KSPHPDDMType`, `KSPHPDDMSetType()`
505: @*/
506: PetscErrorCode KSPHPDDMGetType(KSP ksp, KSPHPDDMType *type)
507: {
509: if (type) {
511: PetscUseMethod(ksp, "KSPHPDDMGetType_C", (KSP, KSPHPDDMType *), (ksp, type));
512: }
513: return 0;
514: }
516: static PetscErrorCode KSPHPDDMSetType_HPDDM(KSP ksp, KSPHPDDMType type)
517: {
518: KSP_HPDDM *data = (KSP_HPDDM *)ksp->data;
519: PetscInt i;
520: PetscBool flg = PETSC_FALSE;
522: for (i = 0; i < static_cast<PetscInt>(PETSC_STATIC_ARRAY_LENGTH(KSPHPDDMTypes)); ++i) {
523: PetscStrcmp(KSPHPDDMTypes[type], KSPHPDDMTypes[i], &flg);
524: if (flg) break;
525: }
527: if (data->cntl[0] != static_cast<char>(PETSC_DECIDE) && data->cntl[0] != i) KSPHPDDMReset_Private(ksp);
528: data->cntl[0] = i;
529: return 0;
530: }
532: static PetscErrorCode KSPHPDDMGetType_HPDDM(KSP ksp, KSPHPDDMType *type)
533: {
534: KSP_HPDDM *data = (KSP_HPDDM *)ksp->data;
537: /* need to shift by -1 for HPDDM_KRYLOV_METHOD_NONE */
538: *type = static_cast<KSPHPDDMType>(PetscMin(data->cntl[0], static_cast<char>(PETSC_STATIC_ARRAY_LENGTH(KSPHPDDMTypes) - 1)));
539: return 0;
540: }
542: /*MC
543: KSPHPDDM - Interface with the HPDDM library. This `KSP` may be used to further select methods that are currently not implemented natively in PETSc, e.g.,
544: GCRODR [2006], a recycled Krylov method which is similar to `KSPLGMRES`, see [2016] for a comparison. ex75.c shows how to reproduce the results
545: from the aforementioned paper [2006]. A chronological bibliography of relevant publications linked with `KSP` available in HPDDM through `KSPHPDDM`,
546: and not available directly in PETSc, may be found below. The interface is explained in details in [2021].
548: Options Database Keys:
549: + -ksp_gmres_restart <restart, default=30> - see `KSPGMRES`
550: . -ksp_hpddm_type <type, default=gmres> - any of gmres, bgmres, cg, bcg, gcrodr, bgcrodr, bfbcg, or preonly, see `KSPHPDDMType`
551: . -ksp_hpddm_precision <value, default=same as PetscScalar> - any of single or double, see `KSPHPDDMPrecision`
552: . -ksp_hpddm_deflation_tol <eps, default=\-1.0> - tolerance when deflating right-hand sides inside block methods (no deflation by default, only relevant with block methods)
553: . -ksp_hpddm_enlarge_krylov_subspace <p, default=1> - split the initial right-hand side into multiple vectors (only relevant with nonblock methods)
554: . -ksp_hpddm_orthogonalization <type, default=cgs> - any of cgs or mgs, see KSPGMRES
555: . -ksp_hpddm_qr <type, default=cholqr> - distributed QR factorizations with any of cholqr, cgs, or mgs (only relevant with block methods)
556: . -ksp_hpddm_variant <type, default=left> - any of left, right, or flexible (this option is superseded by `KSPSetPCSide()`)
557: . -ksp_hpddm_recycle <n, default=0> - number of harmonic Ritz vectors to compute (only relevant with GCRODR or BGCRODR)
558: . -ksp_hpddm_recycle_target <type, default=SM> - criterion to select harmonic Ritz vectors using either SM, LM, SR, LR, SI, or LI (only relevant with GCRODR or BGCRODR).
559: For BGCRODR, if PETSc is compiled with SLEPc, this option is not relevant, since SLEPc is used instead. Options are set with the prefix -ksp_hpddm_recycle_eps_
560: . -ksp_hpddm_recycle_strategy <type, default=A> - generalized eigenvalue problem A or B to solve for recycling (only relevant with flexible GCRODR or BGCRODR)
561: - -ksp_hpddm_recycle_symmetric <true, default=false> - symmetric generalized eigenproblems in BGCRODR, useful to switch to distributed solvers like EPSELEMENTAL or EPSSCALAPACK
562: (only relevant when PETSc is compiled with SLEPc)
564: Level: intermediate
566: References:
567: + 1980 - The block conjugate gradient algorithm and related methods. O'Leary. Linear Algebra and its Applications.
568: . 2006 - Recycling Krylov subspaces for sequences of linear systems. Parks, de Sturler, Mackey, Johnson, and Maiti. SIAM Journal on Scientific Computing
569: . 2013 - A modified block flexible GMRES method with deflation at each iteration for the solution of non-Hermitian linear systems with multiple right-hand sides.
570: Calandra, Gratton, Lago, Vasseur, and Carvalho. SIAM Journal on Scientific Computing.
571: . 2016 - Block iterative methods and recycling for improved scalability of linear solvers. Jolivet and Tournier. SC16.
572: . 2017 - A breakdown-free block conjugate gradient method. Ji and Li. BIT Numerical Mathematics.
573: - 2021 - KSPHPDDM and PCHPDDM: extending PETSc with advanced Krylov methods and robust multilevel overlapping Schwarz preconditioners. Jolivet, Roman, and Zampini.
574: Computer & Mathematics with Applications.
576: .seealso: [](chapter_ksp), [](sec_flexibleksp), `KSPCreate()`, `KSPSetType()`, `KSPType`, `KSP`, `KSPGMRES`, `KSPCG`, `KSPLGMRES`, `KSPDGMRES`
577: M*/
579: PETSC_EXTERN PetscErrorCode KSPCreate_HPDDM(KSP ksp)
580: {
581: KSP_HPDDM *data;
582: PetscInt i;
583: const char *common[] = {KSPGMRES, KSPCG, KSPPREONLY};
584: PetscBool flg = PETSC_FALSE;
586: PetscNew(&data);
587: ksp->data = (void *)data;
588: KSPSetSupportedNorm(ksp, KSP_NORM_PRECONDITIONED, PC_LEFT, 2);
589: KSPSetSupportedNorm(ksp, KSP_NORM_UNPRECONDITIONED, PC_RIGHT, 1);
590: ksp->ops->solve = KSPSolve_HPDDM;
591: ksp->ops->matsolve = KSPMatSolve_HPDDM;
592: ksp->ops->setup = KSPSetUp_HPDDM;
593: ksp->ops->setfromoptions = KSPSetFromOptions_HPDDM;
594: ksp->ops->destroy = KSPDestroy_HPDDM;
595: ksp->ops->view = KSPView_HPDDM;
596: ksp->ops->reset = KSPReset_HPDDM;
597: KSPHPDDMReset_Private(ksp);
598: for (i = 0; i < static_cast<PetscInt>(PETSC_STATIC_ARRAY_LENGTH(common)); ++i) {
599: PetscStrcmp(((PetscObject)ksp)->type_name, common[i], &flg);
600: if (flg) break;
601: }
602: if (!i) data->cntl[0] = HPDDM_KRYLOV_METHOD_GMRES;
603: else if (i == 1) data->cntl[0] = HPDDM_KRYLOV_METHOD_CG;
604: else if (i == 2) data->cntl[0] = HPDDM_KRYLOV_METHOD_NONE;
605: if (data->cntl[0] != static_cast<char>(PETSC_DECIDE)) PetscInfo(ksp, "Using the previously set KSPType %s\n", common[i]);
606: PetscObjectComposeFunction((PetscObject)ksp, "KSPHPDDMSetDeflationMat_C", KSPHPDDMSetDeflationMat_HPDDM);
607: PetscObjectComposeFunction((PetscObject)ksp, "KSPHPDDMGetDeflationMat_C", KSPHPDDMGetDeflationMat_HPDDM);
608: PetscObjectComposeFunction((PetscObject)ksp, "KSPHPDDMSetType_C", KSPHPDDMSetType_HPDDM);
609: PetscObjectComposeFunction((PetscObject)ksp, "KSPHPDDMGetType_C", KSPHPDDMGetType_HPDDM);
610: #if defined(PETSC_HAVE_SLEPC) && PetscDefined(HAVE_DYNAMIC_LIBRARIES) && defined(PETSC_USE_SHARED_LIBRARIES)
611: if (!loadedDL) HPDDMLoadDL_Private(&loadedDL);
612: #endif
613: data->precision = PETSC_KSPHPDDM_DEFAULT_PRECISION;
614: return 0;
615: }