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_", nullptr};
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 PetscDefined(HAVE_SLEPC) && PetscDefined(HAVE_DYNAMIC_LIBRARIES) && PetscDefined(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: PetscFunctionBegin;
35: PetscOptionsHeadBegin(PetscOptionsObject, "KSPHPDDM options, cf. https://github.com/hpddm/hpddm");
36: i = (data->cntl[0] == static_cast<char>(PETSC_DECIDE) ? HPDDM_KRYLOV_METHOD_GMRES : data->cntl[0]);
37: PetscCall(PetscOptionsEList("-ksp_hpddm_type", "Type of Krylov method", "KSPHPDDMGetType", KSPHPDDMTypes, PETSC_STATIC_ARRAY_LENGTH(KSPHPDDMTypes), KSPHPDDMTypes[HPDDM_KRYLOV_METHOD_GMRES], &i, nullptr));
38: 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 */
39: data->cntl[0] = i;
40: PetscCall(PetscOptionsEnum("-ksp_hpddm_precision", "Precision in which Krylov bases are stored", "KSPHPDDM", KSPHPDDMPrecisionTypes, (PetscEnum)data->precision, (PetscEnum *)&data->precision, nullptr));
41: PetscCheck(data->precision != KSP_HPDDM_PRECISION_QUADRUPLE || PetscDefined(HAVE_REAL___FLOAT128), PetscObjectComm((PetscObject)ksp), PETSC_ERR_SUP_SYS, "Unsupported %s precision", KSPHPDDMPrecisionTypes[data->precision]);
42: PetscCheck(std::abs(data->precision - PETSC_KSPHPDDM_DEFAULT_PRECISION) <= 1, PetscObjectComm((PetscObject)ksp), PETSC_ERR_SUP, "Unhandled mixed %s and %s precisions", KSPHPDDMPrecisionTypes[data->precision], KSPHPDDMPrecisionTypes[PETSC_KSPHPDDM_DEFAULT_PRECISION]);
43: if (data->cntl[0] != HPDDM_KRYLOV_METHOD_NONE) {
44: if (data->cntl[0] != HPDDM_KRYLOV_METHOD_BCG && data->cntl[0] != HPDDM_KRYLOV_METHOD_BFBCG) {
45: i = (data->cntl[1] == static_cast<char>(PETSC_DECIDE) ? HPDDM_VARIANT_LEFT : data->cntl[1]);
46: if (ksp->pc_side_set == PC_SIDE_DEFAULT)
47: PetscCall(PetscOptionsEList("-ksp_hpddm_variant", "Left, right, or variable preconditioning", "KSPHPDDM", HPDDMVariant, PETSC_STATIC_ARRAY_LENGTH(HPDDMVariant), HPDDMVariant[HPDDM_VARIANT_LEFT], &i, nullptr));
48: else if (ksp->pc_side_set == PC_RIGHT) i = HPDDM_VARIANT_RIGHT;
49: data->cntl[1] = i;
50: if (i > 0) PetscCall(KSPSetPCSide(ksp, PC_RIGHT));
51: }
52: if (data->cntl[0] == HPDDM_KRYLOV_METHOD_BGMRES || data->cntl[0] == HPDDM_KRYLOV_METHOD_BGCRODR || data->cntl[0] == HPDDM_KRYLOV_METHOD_BFBCG) {
53: data->rcntl[0] = (PetscAbsReal(data->rcntl[0] - static_cast<PetscReal>(PETSC_DECIDE)) < PETSC_SMALL ? -1.0 : data->rcntl[0]);
54: PetscCall(PetscOptionsReal("-ksp_hpddm_deflation_tol", "Tolerance when deflating right-hand sides inside block methods", "KSPHPDDM", data->rcntl[0], data->rcntl, nullptr));
55: 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]));
56: PetscCall(PetscOptionsRangeInt("-ksp_hpddm_enlarge_krylov_subspace", "Split the initial right-hand side into multiple vectors", "KSPHPDDM", i, &i, nullptr, 1, std::numeric_limits<unsigned short>::max() - 1));
57: data->scntl[data->cntl[0] != HPDDM_KRYLOV_METHOD_BFBCG] = i;
58: } else data->scntl[data->cntl[0] != HPDDM_KRYLOV_METHOD_BCG] = 0;
59: 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) {
60: i = (data->cntl[2] == static_cast<char>(PETSC_DECIDE) ? HPDDM_ORTHOGONALIZATION_CGS : data->cntl[2] & 3);
61: PetscCall(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, nullptr));
62: j = (data->cntl[2] == static_cast<char>(PETSC_DECIDE) ? HPDDM_QR_CHOLQR : ((data->cntl[2] >> 2) & 7));
63: PetscCall(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, nullptr));
64: data->cntl[2] = static_cast<char>(i) + (static_cast<char>(j) << 2);
65: i = (data->scntl[0] == static_cast<unsigned short>(PETSC_DECIDE) ? PetscMin(30, ksp->max_it) : data->scntl[0]);
66: PetscCall(PetscOptionsRangeInt("-ksp_gmres_restart", "Maximum number of Arnoldi vectors generated per cycle", "KSPHPDDM", i, &i, nullptr, PetscMin(1, ksp->max_it), PetscMin(ksp->max_it, std::numeric_limits<unsigned short>::max() - 1)));
67: data->scntl[0] = i;
68: }
69: if (data->cntl[0] == HPDDM_KRYLOV_METHOD_BCG || data->cntl[0] == HPDDM_KRYLOV_METHOD_BFBCG) {
70: j = (data->cntl[1] == static_cast<char>(PETSC_DECIDE) ? HPDDM_QR_CHOLQR : data->cntl[1]);
71: PetscCall(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, nullptr));
72: data->cntl[1] = j;
73: }
74: if (data->cntl[0] == HPDDM_KRYLOV_METHOD_GCRODR || data->cntl[0] == HPDDM_KRYLOV_METHOD_BGCRODR) {
75: i = (data->icntl[0] == static_cast<int>(PETSC_DECIDE) ? PetscMin(20, data->scntl[0] - 1) : data->icntl[0]);
76: PetscCall(PetscOptionsRangeInt("-ksp_hpddm_recycle", "Number of harmonic Ritz vectors to compute", "KSPHPDDM", i, &i, nullptr, 1, data->scntl[0] - 1));
77: data->icntl[0] = i;
78: if (!PetscDefined(HAVE_SLEPC) || !PetscDefined(USE_SHARED_LIBRARIES) || data->cntl[0] == HPDDM_KRYLOV_METHOD_GCRODR) {
79: i = (data->cntl[3] == static_cast<char>(PETSC_DECIDE) ? HPDDM_RECYCLE_TARGET_SM : data->cntl[3]);
80: PetscCall(PetscOptionsEList("-ksp_hpddm_recycle_target", "Criterion to select harmonic Ritz vectors", "KSPHPDDM", HPDDMRecycleTarget, PETSC_STATIC_ARRAY_LENGTH(HPDDMRecycleTarget), HPDDMRecycleTarget[HPDDM_RECYCLE_TARGET_SM], &i, nullptr));
81: data->cntl[3] = i;
82: } else {
83: PetscCheck(data->precision == PETSC_KSPHPDDM_DEFAULT_PRECISION, PetscObjectComm((PetscObject)ksp), PETSC_ERR_ARG_INCOMP, "Cannot use SLEPc with a different precision than PETSc for harmonic Ritz eigensolves");
84: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)ksp), &size));
85: i = (data->cntl[3] == static_cast<char>(PETSC_DECIDE) ? 1 : data->cntl[3]);
86: PetscCall(PetscOptionsRangeInt("-ksp_hpddm_recycle_redistribute", "Number of processes used to solve eigenvalue problems when recycling in BGCRODR", "KSPHPDDM", i, &i, nullptr, 1, PetscMin(size, 192)));
87: data->cntl[3] = i;
88: }
89: i = (data->cntl[4] == static_cast<char>(PETSC_DECIDE) ? HPDDM_RECYCLE_STRATEGY_A : data->cntl[4]);
90: PetscCall(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, nullptr));
91: data->cntl[4] = i;
92: }
93: } else {
94: data->cntl[0] = HPDDM_KRYLOV_METHOD_NONE;
95: data->scntl[1] = 1;
96: }
97: PetscCheck(ksp->nmax >= std::numeric_limits<int>::min() && ksp->nmax <= std::numeric_limits<int>::max(), PetscObjectComm((PetscObject)ksp), PETSC_ERR_ARG_OUTOFRANGE, "KSPMatSolve() block size %" PetscInt_FMT " not representable by an integer, which is not handled by KSPHPDDM",
98: ksp->nmax);
99: data->icntl[1] = static_cast<int>(ksp->nmax);
100: PetscOptionsHeadEnd();
101: PetscFunctionReturn(PETSC_SUCCESS);
102: }
104: static PetscErrorCode KSPView_HPDDM(KSP ksp, PetscViewer viewer)
105: {
106: KSP_HPDDM *data = (KSP_HPDDM *)ksp->data;
107: HPDDM::PETScOperator *op = data->op;
108: const PetscScalar *array = op ? op->storage() : nullptr;
109: PetscBool ascii;
111: PetscFunctionBegin;
112: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &ascii));
113: if (op && ascii) {
114: PetscCall(PetscViewerASCIIPrintf(viewer, "HPDDM type: %s%s\n", KSPHPDDMTypes[std::min(static_cast<PetscInt>(data->cntl[0]), static_cast<PetscInt>(PETSC_STATIC_ARRAY_LENGTH(KSPHPDDMTypes) - 1))], data->cntl[1] == HPDDM_VARIANT_FLEXIBLE ? " (with support for variable preconditioning)" : ""));
115: PetscCall(PetscViewerASCIIPrintf(viewer, "precision: %s\n", KSPHPDDMPrecisionTypes[data->precision]));
116: if (data->cntl[0] == HPDDM_KRYLOV_METHOD_BGMRES || data->cntl[0] == HPDDM_KRYLOV_METHOD_BGCRODR || data->cntl[0] == HPDDM_KRYLOV_METHOD_BFBCG) {
117: if (PetscAbsReal(data->rcntl[0] - static_cast<PetscReal>(PETSC_DECIDE)) < PETSC_SMALL) PetscCall(PetscViewerASCIIPrintf(viewer, "no deflation at restarts\n"));
118: else PetscCall(PetscViewerASCIIPrintf(viewer, "deflation tolerance: %g\n", static_cast<double>(data->rcntl[0])));
119: }
120: if (data->cntl[0] == HPDDM_KRYLOV_METHOD_GCRODR || data->cntl[0] == HPDDM_KRYLOV_METHOD_BGCRODR) {
121: PetscCall(PetscViewerASCIIPrintf(viewer, "deflation subspace attached? %s\n", PetscBools[array ? PETSC_TRUE : PETSC_FALSE]));
122: if (!PetscDefined(HAVE_SLEPC) || !PetscDefined(USE_SHARED_LIBRARIES) || data->cntl[0] == HPDDM_KRYLOV_METHOD_GCRODR) PetscCall(PetscViewerASCIIPrintf(viewer, "deflation target: %s\n", HPDDMRecycleTarget[static_cast<PetscInt>(data->cntl[3])]));
123: else PetscCall(PetscViewerASCIIPrintf(viewer, "redistribution size: %d\n", static_cast<PetscMPIInt>(data->cntl[3])));
124: }
125: if (data->icntl[1] != static_cast<int>(PETSC_DECIDE)) PetscCall(PetscViewerASCIIPrintf(viewer, " block size is %d\n", data->icntl[1]));
126: }
127: PetscFunctionReturn(PETSC_SUCCESS);
128: }
130: static PetscErrorCode KSPSetUp_HPDDM(KSP ksp)
131: {
132: KSP_HPDDM *data = (KSP_HPDDM *)ksp->data;
133: Mat A;
134: PetscInt n, bs;
135: PetscBool match;
137: PetscFunctionBegin;
138: PetscCall(KSPGetOperators(ksp, &A, nullptr));
139: PetscCall(MatGetLocalSize(A, &n, nullptr));
140: PetscCall(MatGetBlockSize(A, &bs));
141: PetscCall(PetscObjectTypeCompareAny((PetscObject)A, &match, MATSEQKAIJ, MATMPIKAIJ, ""));
142: if (match) n /= bs;
143: data->op = new HPDDM::PETScOperator(ksp, n);
144: 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() */
145: PetscCall(PetscInfo(ksp, "KSPSetFromOptions() not called or uninitialized internal structure, hardwiring default KSPHPDDM options\n"));
146: if (data->cntl[0] == static_cast<char>(PETSC_DECIDE)) data->cntl[0] = 0; /* GMRES by default */
147: if (data->cntl[0] != HPDDM_KRYLOV_METHOD_NONE) { /* following options do not matter with PREONLY */
148: if (data->cntl[0] != HPDDM_KRYLOV_METHOD_BCG && data->cntl[0] != HPDDM_KRYLOV_METHOD_BFBCG) {
149: data->cntl[1] = HPDDM_VARIANT_LEFT; /* left preconditioning by default */
150: if (ksp->pc_side_set == PC_RIGHT) data->cntl[1] = HPDDM_VARIANT_RIGHT;
151: if (data->cntl[1] > 0) PetscCall(KSPSetPCSide(ksp, PC_RIGHT));
152: }
153: if (data->cntl[0] == HPDDM_KRYLOV_METHOD_BGMRES || data->cntl[0] == HPDDM_KRYLOV_METHOD_BGCRODR || data->cntl[0] == HPDDM_KRYLOV_METHOD_BFBCG) {
154: data->rcntl[0] = -1.0; /* no deflation by default */
155: data->scntl[data->cntl[0] != HPDDM_KRYLOV_METHOD_BFBCG] = 1; /* Krylov subspace not enlarged by default */
156: } else data->scntl[data->cntl[0] != HPDDM_KRYLOV_METHOD_BCG] = 0;
157: 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) {
158: data->cntl[2] = static_cast<char>(HPDDM_ORTHOGONALIZATION_CGS) + (static_cast<char>(HPDDM_QR_CHOLQR) << 2); /* CGS and CholQR by default */
159: data->scntl[0] = PetscMin(30, ksp->max_it); /* restart parameter of 30 by default */
160: }
161: if (data->cntl[0] == HPDDM_KRYLOV_METHOD_BCG || data->cntl[0] == HPDDM_KRYLOV_METHOD_BFBCG) { data->cntl[1] = HPDDM_QR_CHOLQR; /* CholQR by default */ }
162: if (data->cntl[0] == HPDDM_KRYLOV_METHOD_GCRODR || data->cntl[0] == HPDDM_KRYLOV_METHOD_BGCRODR) {
163: data->icntl[0] = PetscMin(20, data->scntl[0] - 1); /* recycled subspace of size 20 by default */
164: if (!PetscDefined(HAVE_SLEPC) || !PetscDefined(USE_SHARED_LIBRARIES) || data->cntl[0] == HPDDM_KRYLOV_METHOD_GCRODR) {
165: data->cntl[3] = HPDDM_RECYCLE_TARGET_SM; /* default recycling target */
166: } else {
167: data->cntl[3] = 1; /* redistribution parameter of 1 by default */
168: }
169: data->cntl[4] = HPDDM_RECYCLE_STRATEGY_A; /* default recycling strategy */
170: }
171: } else data->scntl[1] = 1;
172: }
173: PetscCheck(ksp->nmax >= std::numeric_limits<int>::min() && ksp->nmax <= std::numeric_limits<int>::max(), PetscObjectComm((PetscObject)ksp), PETSC_ERR_ARG_OUTOFRANGE, "KSPMatSolve() block size %" PetscInt_FMT " not representable by an integer, which is not handled by KSPHPDDM",
174: ksp->nmax);
175: data->icntl[1] = static_cast<int>(ksp->nmax);
176: PetscFunctionReturn(PETSC_SUCCESS);
177: }
179: static inline PetscErrorCode KSPReset_HPDDM_Private(KSP ksp)
180: {
181: KSP_HPDDM *data = (KSP_HPDDM *)ksp->data;
183: PetscFunctionBegin;
184: /* cast PETSC_DECIDE into the appropriate types to avoid compiler warnings */
185: std::fill_n(data->rcntl, PETSC_STATIC_ARRAY_LENGTH(data->rcntl), static_cast<PetscReal>(PETSC_DECIDE));
186: std::fill_n(data->icntl, PETSC_STATIC_ARRAY_LENGTH(data->icntl), static_cast<int>(PETSC_DECIDE));
187: std::fill_n(data->scntl, PETSC_STATIC_ARRAY_LENGTH(data->scntl), static_cast<unsigned short>(PETSC_DECIDE));
188: std::fill_n(data->cntl, PETSC_STATIC_ARRAY_LENGTH(data->cntl), static_cast<char>(PETSC_DECIDE));
189: data->precision = PETSC_KSPHPDDM_DEFAULT_PRECISION;
190: PetscFunctionReturn(PETSC_SUCCESS);
191: }
193: static PetscErrorCode KSPReset_HPDDM(KSP ksp)
194: {
195: KSP_HPDDM *data = (KSP_HPDDM *)ksp->data;
197: PetscFunctionBegin;
198: delete data->op;
199: data->op = nullptr;
200: PetscCall(KSPReset_HPDDM_Private(ksp));
201: PetscFunctionReturn(PETSC_SUCCESS);
202: }
204: static PetscErrorCode KSPDestroy_HPDDM(KSP ksp)
205: {
206: PetscFunctionBegin;
207: PetscCall(KSPReset_HPDDM(ksp));
208: PetscCall(KSPDestroyDefault(ksp));
209: PetscCall(PetscObjectComposeFunction((PetscObject)ksp, "KSPHPDDMSetDeflationMat_C", nullptr));
210: PetscCall(PetscObjectComposeFunction((PetscObject)ksp, "KSPHPDDMGetDeflationMat_C", nullptr));
211: PetscCall(PetscObjectComposeFunction((PetscObject)ksp, "KSPHPDDMSetType_C", nullptr));
212: PetscCall(PetscObjectComposeFunction((PetscObject)ksp, "KSPHPDDMGetType_C", nullptr));
213: PetscFunctionReturn(PETSC_SUCCESS);
214: }
216: template <PetscMemType type = PETSC_MEMTYPE_HOST>
217: static inline PetscErrorCode KSPSolve_HPDDM_Private(KSP ksp, const PetscScalar *b, PetscScalar *x, PetscInt n)
218: {
219: KSP_HPDDM *data = (KSP_HPDDM *)ksp->data;
220: KSPConvergedDefaultCtx *ctx = (KSPConvergedDefaultCtx *)ksp->cnvP;
221: const PetscInt N = data->op->getDof() * n;
222: PetscBool flg;
223: #if !PetscDefined(USE_REAL_DOUBLE) || PetscDefined(HAVE_F2CBLASLAPACK___FLOAT128_BINDINGS)
224: HPDDM::upscaled_type<PetscScalar> *high[2];
225: #endif
226: #if !PetscDefined(USE_REAL_SINGLE) || PetscDefined(HAVE_F2CBLASLAPACK___FP16_BINDINGS)
227: typedef HPDDM::downscaled_type<PetscReal> PetscDownscaledReal PETSC_ATTRIBUTE_MAY_ALIAS;
228: #if !PetscDefined(USE_COMPLEX)
229: PetscDownscaledReal *low[2];
230: #else
231: typedef PetscReal PetscAliasedReal PETSC_ATTRIBUTE_MAY_ALIAS;
232: HPDDM::downscaled_type<PetscScalar> *low[2];
233: PetscAliasedReal *x_r;
234: PetscDownscaledReal *low_r;
235: #endif
236: #endif
237: #if PetscDefined(HAVE_CUDA)
238: Mat A;
239: VecType vtype;
240: #endif
242: PetscFunctionBegin;
243: #if PetscDefined(HAVE_CUDA)
244: PetscCall(KSPGetOperators(ksp, &A, nullptr));
245: PetscCall(MatGetVecType(A, &vtype));
246: std::initializer_list<std::string> list = {VECCUDA, VECSEQCUDA, VECMPICUDA};
247: std::initializer_list<std::string>::const_iterator it = std::find(list.begin(), list.end(), std::string(vtype));
248: PetscCheck(type != PETSC_MEMTYPE_HOST || it == list.end(), PetscObjectComm((PetscObject)ksp), PETSC_ERR_SUP, "MatGetVecType() must return a Vec with the same PetscMemType as the right-hand side and solution, PetscMemType(%s) != %s", vtype, PetscMemTypeToString(type));
249: #endif
250: PetscCall(PCGetDiagonalScale(ksp->pc, &flg));
251: PetscCheck(!flg, PetscObjectComm((PetscObject)ksp), PETSC_ERR_SUP, "Krylov method %s does not support diagonal scaling", ((PetscObject)ksp)->type_name);
252: if (n > 1) {
253: if (ksp->converged == KSPConvergedDefault) {
254: PetscCheck(!ctx->mininitialrtol, PetscObjectComm((PetscObject)ksp), PETSC_ERR_SUP, "Krylov method %s does not support KSPConvergedDefaultSetUMIRNorm()", ((PetscObject)ksp)->type_name);
255: if (!ctx->initialrtol) {
256: PetscCall(PetscInfo(ksp, "Forcing KSPConvergedDefaultSetUIRNorm() since KSPConvergedDefault() cannot handle multiple norms\n"));
257: ctx->initialrtol = PETSC_TRUE;
258: }
259: } else PetscCall(PetscInfo(ksp, "Using a special \"converged\" callback, be careful, it is used in KSPHPDDM to track blocks of residuals\n"));
260: }
261: /* initial guess is always nonzero with recycling methods if there is a deflation subspace available */
262: if ((data->cntl[0] == HPDDM_KRYLOV_METHOD_GCRODR || data->cntl[0] == HPDDM_KRYLOV_METHOD_BGCRODR) && data->op->storage()) ksp->guess_zero = PETSC_FALSE;
263: ksp->its = 0;
264: ksp->reason = KSP_CONVERGED_ITERATING;
265: if (data->precision > PETSC_KSPHPDDM_DEFAULT_PRECISION) { /* Krylov basis stored in higher precision than PetscScalar */
266: #if !PetscDefined(USE_REAL_DOUBLE) || PetscDefined(HAVE_F2CBLASLAPACK___FLOAT128_BINDINGS)
267: if (type == PETSC_MEMTYPE_HOST) {
268: PetscCall(PetscMalloc2(N, high, N, high + 1));
269: HPDDM::copy_n(b, N, high[0]);
270: HPDDM::copy_n(x, N, high[1]);
271: PetscCall(HPDDM::IterativeMethod::solve(*data->op, high[0], high[1], n, PetscObjectComm((PetscObject)ksp)));
272: HPDDM::copy_n(high[1], N, x);
273: PetscCall(PetscFree2(high[0], high[1]));
274: } else {
275: PetscCheck(PetscDefined(HAVE_CUDA) && PetscDefined(USE_REAL_SINGLE), PetscObjectComm((PetscObject)ksp), PETSC_ERR_SUP, "CUDA in PETSc has no support for precisions other than single or double");
276: #if PetscDefined(HAVE_CUDA)
277: #if PetscDefined(HAVE_HPDDM)
278: PetscCall(KSPSolve_HPDDM_CUDA_Private(data, b, x, n, PetscObjectComm((PetscObject)ksp)));
279: #else
280: SETERRQ(PetscObjectComm((PetscObject)ksp), PETSC_ERR_SUP, "No CUDA support with --download-hpddm from SLEPc");
281: #endif
282: #endif
283: }
284: #else
285: PetscCheck(data->precision != KSP_HPDDM_PRECISION_QUADRUPLE, PetscObjectComm((PetscObject)ksp), PETSC_ERR_SUP, "Reconfigure with --download-f2cblaslapack --with-f2cblaslapack-float128-bindings");
286: #endif
287: } else if (data->precision < PETSC_KSPHPDDM_DEFAULT_PRECISION) { /* Krylov basis stored in lower precision than PetscScalar */
288: #if !PetscDefined(USE_REAL_SINGLE) || PetscDefined(HAVE_F2CBLASLAPACK___FP16_BINDINGS)
289: if (type == PETSC_MEMTYPE_HOST) {
290: PetscCall(PetscMalloc1(N, low));
291: #if !PetscDefined(USE_COMPLEX)
292: low[1] = reinterpret_cast<PetscDownscaledReal *>(x);
293: #else
294: low[1] = reinterpret_cast<HPDDM::downscaled_type<PetscScalar> *>(x);
295: #endif
296: std::copy_n(b, N, low[0]);
297: for (PetscInt i = 0; i < N; ++i) low[1][i] = x[i];
298: PetscCall(HPDDM::IterativeMethod::solve(*data->op, low[0], low[1], n, PetscObjectComm((PetscObject)ksp)));
299: #if !PetscDefined(USE_COMPLEX)
300: for (PetscInt i = N; i-- > 0;) x[i] = static_cast<PetscScalar>(low[1][i]);
301: #else
302: x_r = reinterpret_cast<PetscAliasedReal *>(x), low_r = reinterpret_cast<PetscDownscaledReal *>(x_r);
303: for (PetscInt i = 2 * N; i-- > 0;) x_r[i] = static_cast<PetscReal>(low_r[i]);
304: #endif
305: PetscCall(PetscFree(low[0]));
306: } else {
307: PetscCheck(PetscDefined(HAVE_CUDA) && PetscDefined(USE_REAL_DOUBLE), PetscObjectComm((PetscObject)ksp), PETSC_ERR_SUP, "CUDA in PETSc has no support for precisions other than single or double");
308: #if PetscDefined(HAVE_CUDA)
309: #if PetscDefined(HAVE_HPDDM)
310: PetscCall(KSPSolve_HPDDM_CUDA_Private(data, b, x, n, PetscObjectComm((PetscObject)ksp)));
311: #else
312: SETERRQ(PetscObjectComm((PetscObject)ksp), PETSC_ERR_SUP, "No CUDA support with --download-hpddm from SLEPc");
313: #endif
314: #endif
315: }
316: #else
317: PetscCheck(data->precision != KSP_HPDDM_PRECISION_HALF, PetscObjectComm((PetscObject)ksp), PETSC_ERR_SUP, "Reconfigure with --download-f2cblaslapack --with-f2cblaslapack-fp16-bindings");
318: #endif
319: } else { /* Krylov basis stored in the same precision as PetscScalar */
320: if (type == PETSC_MEMTYPE_HOST) PetscCall(HPDDM::IterativeMethod::solve(*data->op, b, x, n, PetscObjectComm((PetscObject)ksp)));
321: else {
322: PetscCheck(PetscDefined(USE_REAL_SINGLE) || PetscDefined(USE_REAL_DOUBLE), PetscObjectComm((PetscObject)ksp), PETSC_ERR_SUP, "CUDA in PETSc has no support for precisions other than single or double");
323: #if PetscDefined(HAVE_CUDA)
324: #if PetscDefined(HAVE_HPDDM)
325: PetscCall(KSPSolve_HPDDM_CUDA_Private(data, b, x, n, PetscObjectComm((PetscObject)ksp)));
326: #else
327: SETERRQ(PetscObjectComm((PetscObject)ksp), PETSC_ERR_SUP, "No CUDA support with --download-hpddm from SLEPc");
328: #endif
329: #endif
330: }
331: }
332: if (!ksp->reason) { /* KSPConvergedDefault() is still returning 0 (= KSP_CONVERGED_ITERATING) */
333: if (ksp->its >= ksp->max_it) ksp->reason = KSP_DIVERGED_ITS;
334: else ksp->reason = KSP_CONVERGED_RTOL; /* early exit by HPDDM, which only happens on breakdowns or convergence */
335: }
336: ksp->its = PetscMin(ksp->its, ksp->max_it);
337: PetscFunctionReturn(PETSC_SUCCESS);
338: }
340: static PetscErrorCode KSPSolve_HPDDM(KSP ksp)
341: {
342: KSP_HPDDM *data = (KSP_HPDDM *)ksp->data;
343: Mat A, B;
344: PetscScalar *x, *bt = nullptr, **ptr;
345: const PetscScalar *b;
346: PetscInt i, j, n;
347: PetscBool flg;
348: PetscMemType type[2];
350: PetscFunctionBegin;
351: PetscCall(PetscCitationsRegister(HPDDMCitation, &HPDDMCite));
352: PetscCall(KSPGetOperators(ksp, &A, nullptr));
353: PetscCall(PetscObjectTypeCompareAny((PetscObject)A, &flg, MATSEQKAIJ, MATMPIKAIJ, ""));
354: PetscCall(VecGetArrayWriteAndMemType(ksp->vec_sol, &x, type));
355: PetscCall(VecGetArrayReadAndMemType(ksp->vec_rhs, &b, type + 1));
356: PetscCheck(type[0] == type[1], PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_INCOMP, "Right-hand side and solution vectors must have the same PetscMemType, %s != %s", PetscMemTypeToString(type[0]), PetscMemTypeToString(type[1]));
357: if (!flg) {
358: if (PetscMemTypeCUDA(type[0])) PetscCall(KSPSolve_HPDDM_Private<PETSC_MEMTYPE_CUDA>(ksp, b, x, 1));
359: else {
360: PetscCheck(PetscMemTypeHost(type[0]), PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "PetscMemType (%s) is neither PETSC_MEMTYPE_HOST nor PETSC_MEMTYPE_CUDA", PetscMemTypeToString(type[0]));
361: PetscCall(KSPSolve_HPDDM_Private(ksp, b, x, 1));
362: }
363: } else {
364: PetscCheck(PetscMemTypeHost(type[0]), PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "PetscMemType (%s) is not PETSC_MEMTYPE_HOST", PetscMemTypeToString(type[0]));
365: PetscCall(MatKAIJGetScaledIdentity(A, &flg));
366: PetscCall(MatKAIJGetAIJ(A, &B));
367: PetscCall(MatGetBlockSize(A, &n));
368: PetscCall(MatGetLocalSize(B, &i, nullptr));
369: j = data->op->getDof();
370: if (!flg) i *= n; /* S and T are not scaled identities, cannot use block methods */
371: if (i != j) { /* switching between block and standard methods */
372: delete data->op;
373: data->op = new HPDDM::PETScOperator(ksp, i);
374: }
375: if (flg && n > 1) {
376: PetscCall(PetscMalloc1(i * n, &bt));
377: /* from row- to column-major to be consistent with HPDDM */
378: HPDDM::Wrapper<PetscScalar>::omatcopy<'T'>(i, n, b, n, bt, i);
379: ptr = const_cast<PetscScalar **>(&b);
380: std::swap(*ptr, bt);
381: HPDDM::Wrapper<PetscScalar>::imatcopy<'T'>(i, n, x, n, i);
382: }
383: PetscCall(KSPSolve_HPDDM_Private(ksp, b, x, flg ? n : 1));
384: if (flg && n > 1) {
385: std::swap(*ptr, bt);
386: PetscCall(PetscFree(bt));
387: /* from column- to row-major to be consistent with MatKAIJ format */
388: HPDDM::Wrapper<PetscScalar>::imatcopy<'T'>(n, i, x, i, n);
389: }
390: }
391: PetscCall(VecRestoreArrayReadAndMemType(ksp->vec_rhs, &b));
392: PetscCall(VecRestoreArrayWriteAndMemType(ksp->vec_sol, &x));
393: PetscFunctionReturn(PETSC_SUCCESS);
394: }
396: /*@
397: KSPHPDDMSetDeflationMat - Sets the deflation space used by Krylov methods in `KSPHPDDM` with recycling. This space is viewed as a set of vectors stored in
398: a `MATDENSE` (column major).
400: Input Parameters:
401: + ksp - iterative context
402: - U - deflation space to be used during KSPSolve()
404: Level: intermediate
406: .seealso: [](ch_ksp), `KSPHPDDM`, `KSPCreate()`, `KSPType`, `KSPHPDDMGetDeflationMat()`
407: @*/
408: PetscErrorCode KSPHPDDMSetDeflationMat(KSP ksp, Mat U)
409: {
410: PetscFunctionBegin;
413: PetscCheckSameComm(ksp, 1, U, 2);
414: PetscUseMethod(ksp, "KSPHPDDMSetDeflationMat_C", (KSP, Mat), (ksp, U));
415: PetscFunctionReturn(PETSC_SUCCESS);
416: }
418: /*@
419: KSPHPDDMGetDeflationMat - Gets the deflation space computed by Krylov methods in `KSPHPDDM` with recycling or `NULL` if `KSPSolve()` has not been called yet.
421: Input Parameter:
422: . ksp - iterative context
424: Output Parameter:
425: . U - deflation space generated during `KSPSolve()`
427: Level: intermediate
429: Note:
430: 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`.
432: .seealso: [](ch_ksp), `KSPHPDDM`, `KSPCreate()`, `KSPType`, `KSPHPDDMSetDeflationMat()`
433: @*/
434: PetscErrorCode KSPHPDDMGetDeflationMat(KSP ksp, Mat *U)
435: {
436: PetscFunctionBegin;
438: if (U) {
439: PetscAssertPointer(U, 2);
440: PetscUseMethod(ksp, "KSPHPDDMGetDeflationMat_C", (KSP, Mat *), (ksp, U));
441: }
442: PetscFunctionReturn(PETSC_SUCCESS);
443: }
445: static PetscErrorCode KSPHPDDMSetDeflationMat_HPDDM(KSP ksp, Mat U)
446: {
447: KSP_HPDDM *data = (KSP_HPDDM *)ksp->data;
448: HPDDM::PETScOperator *op = data->op;
449: Mat A;
450: const PetscScalar *array;
451: PetscScalar *copy;
452: PetscInt m1, M1, m2, M2, n2, N2, ldu;
453: PetscBool match;
455: PetscFunctionBegin;
456: if (!op) {
457: PetscCall(KSPSetUp(ksp));
458: op = data->op;
459: }
460: PetscCheck(data->precision == PETSC_KSPHPDDM_DEFAULT_PRECISION, PetscObjectComm((PetscObject)ksp), PETSC_ERR_SUP, "%s != %s", KSPHPDDMPrecisionTypes[data->precision], KSPHPDDMPrecisionTypes[PETSC_KSPHPDDM_DEFAULT_PRECISION]);
461: PetscCall(KSPGetOperators(ksp, &A, nullptr));
462: PetscCall(MatGetLocalSize(A, &m1, nullptr));
463: PetscCall(MatGetLocalSize(U, &m2, &n2));
464: PetscCall(MatGetSize(A, &M1, nullptr));
465: PetscCall(MatGetSize(U, &M2, &N2));
466: PetscCheck(m1 == m2 && M1 == M2, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Cannot use a deflation space with (m2,M2) = (%" PetscInt_FMT ",%" PetscInt_FMT ") for a linear system with (m1,M1) = (%" PetscInt_FMT ",%" PetscInt_FMT ")", m2, M2, m1, M1);
467: PetscCall(PetscObjectTypeCompareAny((PetscObject)U, &match, MATSEQDENSE, MATMPIDENSE, ""));
468: PetscCheck(match, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Provided deflation space not stored in a dense Mat");
469: PetscCall(MatDenseGetArrayRead(U, &array));
470: copy = op->allocate(m2, 1, N2);
471: PetscCheck(copy, PETSC_COMM_SELF, PETSC_ERR_POINTER, "Memory allocation error");
472: PetscCall(MatDenseGetLDA(U, &ldu));
473: HPDDM::Wrapper<PetscScalar>::omatcopy<'N'>(N2, m2, array, ldu, copy, m2);
474: PetscCall(MatDenseRestoreArrayRead(U, &array));
475: PetscFunctionReturn(PETSC_SUCCESS);
476: }
478: static PetscErrorCode KSPHPDDMGetDeflationMat_HPDDM(KSP ksp, Mat *U)
479: {
480: KSP_HPDDM *data = (KSP_HPDDM *)ksp->data;
481: HPDDM::PETScOperator *op = data->op;
482: Mat A;
483: const PetscScalar *array;
484: PetscScalar *copy;
485: PetscInt m1, M1, N2;
487: PetscFunctionBegin;
488: if (!op) {
489: PetscCall(KSPSetUp(ksp));
490: op = data->op;
491: }
492: PetscCheck(data->precision == PETSC_KSPHPDDM_DEFAULT_PRECISION, PetscObjectComm((PetscObject)ksp), PETSC_ERR_SUP, "%s != %s", KSPHPDDMPrecisionTypes[data->precision], KSPHPDDMPrecisionTypes[PETSC_KSPHPDDM_DEFAULT_PRECISION]);
493: array = op->storage();
494: N2 = op->k().first * op->k().second;
495: if (!array) *U = nullptr;
496: else {
497: PetscCall(KSPGetOperators(ksp, &A, nullptr));
498: PetscCall(MatGetLocalSize(A, &m1, nullptr));
499: PetscCall(MatGetSize(A, &M1, nullptr));
500: PetscCall(MatCreateDense(PetscObjectComm((PetscObject)ksp), m1, PETSC_DECIDE, M1, N2, nullptr, U));
501: PetscCall(MatDenseGetArrayWrite(*U, ©));
502: PetscCall(PetscArraycpy(copy, array, m1 * N2));
503: PetscCall(MatDenseRestoreArrayWrite(*U, ©));
504: }
505: PetscFunctionReturn(PETSC_SUCCESS);
506: }
508: static PetscErrorCode KSPMatSolve_HPDDM(KSP ksp, Mat B, Mat X)
509: {
510: KSP_HPDDM *data = (KSP_HPDDM *)ksp->data;
511: Mat A;
512: const PetscScalar *b;
513: PetscScalar *x;
514: PetscInt n, lda;
515: PetscMemType type[2];
517: PetscFunctionBegin;
518: PetscCall(PetscCitationsRegister(HPDDMCitation, &HPDDMCite));
519: if (!data->op) PetscCall(KSPSetUp(ksp));
520: PetscCall(KSPGetOperators(ksp, &A, nullptr));
521: PetscCall(MatGetLocalSize(B, &n, nullptr));
522: PetscCall(MatDenseGetLDA(B, &lda));
523: PetscCheck(n == lda, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Unhandled leading dimension lda = %" PetscInt_FMT " with n = %" PetscInt_FMT, lda, n);
524: PetscCall(MatGetLocalSize(A, &n, nullptr));
525: PetscCall(MatDenseGetLDA(X, &lda));
526: PetscCheck(n == lda, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Unhandled leading dimension lda = %" PetscInt_FMT " with n = %" PetscInt_FMT, lda, n);
527: PetscCall(MatGetSize(X, nullptr, &n));
528: PetscCall(MatDenseGetArrayWriteAndMemType(X, &x, type));
529: PetscCall(MatDenseGetArrayReadAndMemType(B, &b, type + 1));
530: PetscCheck(type[0] == type[1], PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_INCOMP, "Right-hand side and solution matrices must have the same PetscMemType, %s != %s", PetscMemTypeToString(type[0]), PetscMemTypeToString(type[1]));
531: if (PetscMemTypeCUDA(type[0])) PetscCall(KSPSolve_HPDDM_Private<PETSC_MEMTYPE_CUDA>(ksp, b, x, n));
532: else {
533: PetscCheck(PetscMemTypeHost(type[0]), PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "PetscMemType (%s) is neither PETSC_MEMTYPE_HOST nor PETSC_MEMTYPE_CUDA", PetscMemTypeToString(type[0]));
534: PetscCall(KSPSolve_HPDDM_Private(ksp, b, x, n));
535: }
536: PetscCall(MatDenseRestoreArrayReadAndMemType(B, &b));
537: PetscCall(MatDenseRestoreArrayWriteAndMemType(X, &x));
538: PetscFunctionReturn(PETSC_SUCCESS);
539: }
541: /*@
542: KSPHPDDMSetType - Sets the type of Krylov method used in `KSPHPDDM`.
544: Collective
546: Input Parameters:
547: + ksp - iterative context
548: - type - any of gmres, bgmres, cg, bcg, gcrodr, bgcrodr, bfbcg, or preonly
550: Level: intermediate
552: Notes:
553: Unlike `KSPReset()`, this function does not destroy any deflation space attached to the `KSP`.
555: As an example, in the following sequence\:
556: .vb
557: KSPHPDDMSetType(ksp, KSPGCRODR);
558: KSPSolve(ksp, b, x);
559: KSPHPDDMSetType(ksp, KSPGMRES);
560: KSPHPDDMSetType(ksp, KSPGCRODR);
561: KSPSolve(ksp, b, x);
562: .ve
563: the recycled space is reused in the second `KSPSolve()`.
565: .seealso: [](ch_ksp), `KSPCreate()`, `KSPType`, `KSPHPDDMType`, `KSPHPDDMGetType()`
566: @*/
567: PetscErrorCode KSPHPDDMSetType(KSP ksp, KSPHPDDMType type)
568: {
569: PetscFunctionBegin;
572: PetscUseMethod(ksp, "KSPHPDDMSetType_C", (KSP, KSPHPDDMType), (ksp, type));
573: PetscFunctionReturn(PETSC_SUCCESS);
574: }
576: /*@
577: KSPHPDDMGetType - Gets the type of Krylov method used in `KSPHPDDM`.
579: Input Parameter:
580: . ksp - iterative context
582: Output Parameter:
583: . type - any of gmres, bgmres, cg, bcg, gcrodr, bgcrodr, bfbcg, or preonly
585: Level: intermediate
587: .seealso: [](ch_ksp), `KSPCreate()`, `KSPType`, `KSPHPDDMType`, `KSPHPDDMSetType()`
588: @*/
589: PetscErrorCode KSPHPDDMGetType(KSP ksp, KSPHPDDMType *type)
590: {
591: PetscFunctionBegin;
593: if (type) {
594: PetscAssertPointer(type, 2);
595: PetscUseMethod(ksp, "KSPHPDDMGetType_C", (KSP, KSPHPDDMType *), (ksp, type));
596: }
597: PetscFunctionReturn(PETSC_SUCCESS);
598: }
600: static PetscErrorCode KSPHPDDMSetType_HPDDM(KSP ksp, KSPHPDDMType type)
601: {
602: KSP_HPDDM *data = (KSP_HPDDM *)ksp->data;
603: PetscInt i;
604: PetscBool flg = PETSC_FALSE;
606: PetscFunctionBegin;
607: for (i = 0; i < static_cast<PetscInt>(PETSC_STATIC_ARRAY_LENGTH(KSPHPDDMTypes)); ++i) {
608: PetscCall(PetscStrcmp(KSPHPDDMTypes[type], KSPHPDDMTypes[i], &flg));
609: if (flg) break;
610: }
611: PetscCheck(i != PETSC_STATIC_ARRAY_LENGTH(KSPHPDDMTypes), PetscObjectComm((PetscObject)ksp), PETSC_ERR_ARG_UNKNOWN_TYPE, "Unknown KSPHPDDMType %d", type);
612: if (data->cntl[0] != static_cast<char>(PETSC_DECIDE) && data->cntl[0] != i) PetscCall(KSPReset_HPDDM_Private(ksp));
613: data->cntl[0] = i;
614: PetscFunctionReturn(PETSC_SUCCESS);
615: }
617: static PetscErrorCode KSPHPDDMGetType_HPDDM(KSP ksp, KSPHPDDMType *type)
618: {
619: KSP_HPDDM *data = (KSP_HPDDM *)ksp->data;
621: PetscFunctionBegin;
622: PetscCheck(data->cntl[0] != static_cast<char>(PETSC_DECIDE), PETSC_COMM_SELF, PETSC_ERR_ORDER, "KSPHPDDMType not set yet");
623: /* need to shift by -1 for HPDDM_KRYLOV_METHOD_NONE */
624: *type = static_cast<KSPHPDDMType>(PetscMin(data->cntl[0], static_cast<char>(PETSC_STATIC_ARRAY_LENGTH(KSPHPDDMTypes) - 1)));
625: PetscFunctionReturn(PETSC_SUCCESS);
626: }
628: /*MC
629: 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.,
630: GCRODR [2006], a recycled Krylov method which is similar to `KSPLGMRES`, see [2016] for a comparison. ex75.c shows how to reproduce the results
631: from the aforementioned paper [2006]. A chronological bibliography of relevant publications linked with `KSP` available in HPDDM through `KSPHPDDM`,
632: and not available directly in PETSc, may be found below. The interface is explained in details in [2021].
634: Options Database Keys:
635: + -ksp_gmres_restart <restart, default=30> - see `KSPGMRES`
636: . -ksp_hpddm_type <type, default=gmres> - any of gmres, bgmres, cg, bcg, gcrodr, bgcrodr, bfbcg, or preonly, see `KSPHPDDMType`
637: . -ksp_hpddm_precision <value, default=same as PetscScalar> - any of half, single, double or quadruple, see `KSPHPDDMPrecision`
638: . -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)
639: . -ksp_hpddm_enlarge_krylov_subspace <p, default=1> - split the initial right-hand side into multiple vectors (only relevant with nonblock methods)
640: . -ksp_hpddm_orthogonalization <type, default=cgs> - any of cgs or mgs, see KSPGMRES
641: . -ksp_hpddm_qr <type, default=cholqr> - distributed QR factorizations with any of cholqr, cgs, or mgs (only relevant with block methods)
642: . -ksp_hpddm_variant <type, default=left> - any of left, right, or flexible (this option is superseded by `KSPSetPCSide()`)
643: . -ksp_hpddm_recycle <n, default=0> - number of harmonic Ritz vectors to compute (only relevant with GCRODR or BGCRODR)
644: . -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).
645: 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_
646: . -ksp_hpddm_recycle_strategy <type, default=A> - generalized eigenvalue problem A or B to solve for recycling (only relevant with flexible GCRODR or BGCRODR)
647: - -ksp_hpddm_recycle_symmetric <true, default=false> - symmetric generalized eigenproblems in BGCRODR, useful to switch to distributed solvers like EPSELEMENTAL or EPSSCALAPACK
648: (only relevant when PETSc is compiled with SLEPc)
650: Level: intermediate
652: References:
653: + 1980 - The block conjugate gradient algorithm and related methods. O'Leary. Linear Algebra and its Applications.
654: . 2006 - Recycling Krylov subspaces for sequences of linear systems. Parks, de Sturler, Mackey, Johnson, and Maiti. SIAM Journal on Scientific Computing
655: . 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.
656: Calandra, Gratton, Lago, Vasseur, and Carvalho. SIAM Journal on Scientific Computing.
657: . 2016 - Block iterative methods and recycling for improved scalability of linear solvers. Jolivet and Tournier. SC16.
658: . 2017 - A breakdown-free block conjugate gradient method. Ji and Li. BIT Numerical Mathematics.
659: - 2021 - KSPHPDDM and PCHPDDM: extending PETSc with advanced Krylov methods and robust multilevel overlapping Schwarz preconditioners. Jolivet, Roman, and Zampini.
660: Computer & Mathematics with Applications.
662: .seealso: [](ch_ksp), [](sec_flexibleksp), `KSPCreate()`, `KSPSetType()`, `KSPType`, `KSP`, `KSPGMRES`, `KSPCG`, `KSPLGMRES`, `KSPDGMRES`
663: M*/
665: PETSC_EXTERN PetscErrorCode KSPCreate_HPDDM(KSP ksp)
666: {
667: KSP_HPDDM *data;
668: PetscInt i;
669: const char *common[] = {KSPGMRES, KSPCG, KSPPREONLY};
670: PetscBool flg = PETSC_FALSE;
672: PetscFunctionBegin;
673: PetscCall(PetscNew(&data));
674: ksp->data = (void *)data;
675: PetscCall(KSPSetSupportedNorm(ksp, KSP_NORM_PRECONDITIONED, PC_LEFT, 2));
676: PetscCall(KSPSetSupportedNorm(ksp, KSP_NORM_UNPRECONDITIONED, PC_RIGHT, 1));
677: ksp->ops->solve = KSPSolve_HPDDM;
678: ksp->ops->matsolve = KSPMatSolve_HPDDM;
679: ksp->ops->setup = KSPSetUp_HPDDM;
680: ksp->ops->setfromoptions = KSPSetFromOptions_HPDDM;
681: ksp->ops->destroy = KSPDestroy_HPDDM;
682: ksp->ops->view = KSPView_HPDDM;
683: ksp->ops->reset = KSPReset_HPDDM;
684: PetscCall(KSPReset_HPDDM_Private(ksp));
685: for (i = 0; i < static_cast<PetscInt>(PETSC_STATIC_ARRAY_LENGTH(common)); ++i) {
686: PetscCall(PetscStrcmp(((PetscObject)ksp)->type_name, common[i], &flg));
687: if (flg) break;
688: }
689: if (!i) data->cntl[0] = HPDDM_KRYLOV_METHOD_GMRES;
690: else if (i == 1) data->cntl[0] = HPDDM_KRYLOV_METHOD_CG;
691: else if (i == 2) data->cntl[0] = HPDDM_KRYLOV_METHOD_NONE;
692: if (data->cntl[0] != static_cast<char>(PETSC_DECIDE)) PetscCall(PetscInfo(ksp, "Using the previously set KSPType %s\n", common[i]));
693: PetscCall(PetscObjectComposeFunction((PetscObject)ksp, "KSPHPDDMSetDeflationMat_C", KSPHPDDMSetDeflationMat_HPDDM));
694: PetscCall(PetscObjectComposeFunction((PetscObject)ksp, "KSPHPDDMGetDeflationMat_C", KSPHPDDMGetDeflationMat_HPDDM));
695: PetscCall(PetscObjectComposeFunction((PetscObject)ksp, "KSPHPDDMSetType_C", KSPHPDDMSetType_HPDDM));
696: PetscCall(PetscObjectComposeFunction((PetscObject)ksp, "KSPHPDDMGetType_C", KSPHPDDMGetType_HPDDM));
697: #if PetscDefined(HAVE_SLEPC) && PetscDefined(HAVE_DYNAMIC_LIBRARIES) && PetscDefined(USE_SHARED_LIBRARIES)
698: if (!loadedDL) PetscCall(HPDDMLoadDL_Private(&loadedDL));
699: #endif
700: data->precision = PETSC_KSPHPDDM_DEFAULT_PRECISION;
701: PetscFunctionReturn(PETSC_SUCCESS);
702: }