1 // Copyright (c) 2017-2022, Lawrence Livermore National Security, LLC and other CEED contributors. 2 // All Rights Reserved. See the top-level LICENSE and NOTICE files for details. 3 // 4 // SPDX-License-Identifier: BSD-2-Clause 5 // 6 // This file is part of CEED: http://github.com/ceed 7 8 // libCEED + PETSc Example: CEED BPs 3-6 with Multigrid 9 // 10 // This example demonstrates a simple usage of libCEED with PETSc to solve the CEED BP benchmark problems, see http://ceed.exascaleproject.org/bps. 11 // 12 // The code uses higher level communication protocols in DMPlex. 13 // 14 // Build with: 15 // 16 // make multigrid [PETSC_DIR=</path/to/petsc>] [CEED_DIR=</path/to/libceed>] 17 // 18 // Sample runs: 19 // 20 // multigrid -problem bp3 21 // multigrid -problem bp4 22 // multigrid -problem bp5 -ceed /cpu/self 23 // multigrid -problem bp6 -ceed /gpu/cuda 24 // 25 //TESTARGS -ceed {ceed_resource} -test -problem bp3 -degree 3 26 //TESTARGS -ceed {ceed_resource} -test -problem bp3 -degree 3 -simplex 27 28 /// @file 29 /// CEED BPs 1-6 multigrid example using PETSc 30 const char help[] = "Solve CEED BPs using p-multigrid with PETSc and DMPlex\n"; 31 32 #include <ceed.h> 33 #include <petsc.h> 34 #include <petscdmplex.h> 35 #include <petscksp.h> 36 #include <petscsys.h> 37 #include <stdbool.h> 38 #include <string.h> 39 40 #include "bps.h" 41 #include "include/bpsproblemdata.h" 42 #include "include/libceedsetup.h" 43 #include "include/matops.h" 44 #include "include/petscutils.h" 45 #include "include/petscversion.h" 46 #include "include/structs.h" 47 48 #if PETSC_VERSION_LT(3, 12, 0) 49 #ifdef PETSC_HAVE_CUDA 50 #include <petsccuda.h> 51 // Note: With PETSc prior to version 3.12.0, providing the source path to include 'cublas_v2.h' will be needed to use 'petsccuda.h'. 52 #endif 53 #endif 54 55 int main(int argc, char **argv) { 56 MPI_Comm comm; 57 char filename[PETSC_MAX_PATH_LEN], ceed_resource[PETSC_MAX_PATH_LEN] = "/cpu/self"; 58 double my_rt_start, my_rt, rt_min, rt_max; 59 PetscInt degree = 3, q_extra, *l_size, *xl_size, *g_size, dim = 3, fine_level, mesh_elem[3] = {3, 3, 3}, num_comp_u = 1, num_levels = degree, 60 *level_degrees; 61 PetscScalar *r; 62 PetscScalar eps = 1.0; 63 PetscBool test_mode, benchmark_mode, read_mesh, write_solution, simplex; 64 PetscLogStage solve_stage; 65 PetscLogEvent assemble_event; 66 DM *dm, dm_orig; 67 KSP ksp; 68 PC pc; 69 Mat *mat_O, *mat_pr, mat_coarse; 70 Vec *X, *X_loc, *mult, rhs, rhs_loc; 71 PetscMemType mem_type; 72 OperatorApplyContext *op_apply_ctx, op_error_ctx; 73 ProlongRestrContext *pr_restr_ctx; 74 Ceed ceed; 75 CeedData *ceed_data; 76 CeedVector rhs_ceed, target; 77 CeedQFunction qf_error; 78 CeedOperator op_error; 79 BPType bp_choice; 80 CoarsenType coarsen; 81 82 PetscCall(PetscInitialize(&argc, &argv, NULL, help)); 83 comm = PETSC_COMM_WORLD; 84 85 // Parse command line options 86 PetscOptionsBegin(comm, NULL, "CEED BPs in PETSc", NULL); 87 bp_choice = CEED_BP3; 88 PetscCall(PetscOptionsEnum("-problem", "CEED benchmark problem to solve", NULL, bp_types, (PetscEnum)bp_choice, (PetscEnum *)&bp_choice, NULL)); 89 num_comp_u = bp_options[bp_choice].num_comp_u; 90 test_mode = PETSC_FALSE; 91 PetscCall(PetscOptionsBool("-test", "Testing mode (do not print unless error is large)", NULL, test_mode, &test_mode, NULL)); 92 benchmark_mode = PETSC_FALSE; 93 PetscCall(PetscOptionsBool("-benchmark", "Benchmarking mode (prints benchmark statistics)", NULL, benchmark_mode, &benchmark_mode, NULL)); 94 write_solution = PETSC_FALSE; 95 PetscCall(PetscOptionsBool("-write_solution", "Write solution for visualization", NULL, write_solution, &write_solution, NULL)); 96 simplex = PETSC_FALSE; 97 PetscCall(PetscOptionsBool("-simplex", "Element topology (default:hex)", NULL, simplex, &simplex, NULL)); 98 if ((bp_choice == CEED_BP5 || bp_choice == CEED_BP6) && (simplex)) { 99 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "BP5/6 is not supported with simplex"); 100 } 101 PetscCall(PetscOptionsScalar("-eps", "Epsilon parameter for Kershaw mesh transformation", NULL, eps, &eps, NULL)); 102 if (eps > 1 || eps <= 0) SETERRQ(PETSC_COMM_WORLD, PETSC_ERR_ARG_OUTOFRANGE, "-eps %g must be (0,1]", (double)PetscRealPart(eps)); 103 degree = test_mode ? 3 : 2; 104 PetscCall(PetscOptionsInt("-degree", "Polynomial degree of tensor product basis", NULL, degree, °ree, NULL)); 105 if (degree < 1) SETERRQ(PETSC_COMM_WORLD, PETSC_ERR_ARG_OUTOFRANGE, "-degree %" PetscInt_FMT " must be at least 1", degree); 106 q_extra = bp_options[bp_choice].q_extra; 107 PetscCall(PetscOptionsInt("-q_extra", "Number of extra quadrature points", NULL, q_extra, &q_extra, NULL)); 108 PetscCall(PetscOptionsString("-ceed", "CEED resource specifier", NULL, ceed_resource, ceed_resource, sizeof(ceed_resource), NULL)); 109 coarsen = COARSEN_UNIFORM; 110 PetscCall(PetscOptionsEnum("-coarsen", "Coarsening strategy to use", NULL, coarsen_types, (PetscEnum)coarsen, (PetscEnum *)&coarsen, NULL)); 111 read_mesh = PETSC_FALSE; 112 PetscCall(PetscOptionsString("-mesh", "Read mesh from file", NULL, filename, filename, sizeof(filename), &read_mesh)); 113 if (!read_mesh) { 114 PetscInt tmp = dim; 115 PetscCall(PetscOptionsIntArray("-cells", "Number of cells per dimension", NULL, mesh_elem, &tmp, NULL)); 116 } 117 PetscOptionsEnd(); 118 119 // Set up libCEED 120 CeedInit(ceed_resource, &ceed); 121 CeedMemType mem_type_backend; 122 CeedGetPreferredMemType(ceed, &mem_type_backend); 123 124 // Setup DM 125 if (read_mesh) { 126 PetscCall(DMPlexCreateFromFile(PETSC_COMM_WORLD, filename, NULL, PETSC_TRUE, &dm_orig)); 127 } else { 128 PetscCall(DMPlexCreateBoxMesh(PETSC_COMM_WORLD, dim, simplex, mesh_elem, NULL, NULL, NULL, PETSC_TRUE, &dm_orig)); 129 } 130 131 VecType vec_type; 132 switch (mem_type_backend) { 133 case CEED_MEM_HOST: 134 vec_type = VECSTANDARD; 135 break; 136 case CEED_MEM_DEVICE: { 137 const char *resolved; 138 CeedGetResource(ceed, &resolved); 139 if (strstr(resolved, "/gpu/cuda")) vec_type = VECCUDA; 140 else if (strstr(resolved, "/gpu/hip/occa")) vec_type = VECSTANDARD; // https://github.com/CEED/libCEED/issues/678 141 else if (strstr(resolved, "/gpu/hip")) vec_type = VECHIP; 142 else vec_type = VECSTANDARD; 143 } 144 } 145 PetscCall(DMSetVecType(dm_orig, vec_type)); 146 PetscCall(DMSetFromOptions(dm_orig)); 147 PetscCall(DMViewFromOptions(dm_orig, NULL, "-dm_view")); 148 149 // Apply Kershaw mesh transformation 150 PetscCall(Kershaw(dm_orig, eps)); 151 152 // Allocate arrays for PETSc objects for each level 153 switch (coarsen) { 154 case COARSEN_UNIFORM: 155 num_levels = degree; 156 break; 157 case COARSEN_LOGARITHMIC: 158 num_levels = ceil(log(degree) / log(2)) + 1; 159 break; 160 } 161 PetscCall(PetscMalloc1(num_levels, &level_degrees)); 162 fine_level = num_levels - 1; 163 164 switch (coarsen) { 165 case COARSEN_UNIFORM: 166 for (int i = 0; i < num_levels; i++) level_degrees[i] = i + 1; 167 break; 168 case COARSEN_LOGARITHMIC: 169 for (int i = 0; i < num_levels - 1; i++) level_degrees[i] = pow(2, i); 170 level_degrees[fine_level] = degree; 171 break; 172 } 173 PetscCall(PetscMalloc1(num_levels, &dm)); 174 PetscCall(PetscMalloc1(num_levels, &X)); 175 PetscCall(PetscMalloc1(num_levels, &X_loc)); 176 PetscCall(PetscMalloc1(num_levels, &mult)); 177 PetscCall(PetscMalloc1(num_levels, &op_apply_ctx)); 178 PetscCall(PetscMalloc1(num_levels, &pr_restr_ctx)); 179 PetscCall(PetscMalloc1(num_levels, &mat_O)); 180 PetscCall(PetscMalloc1(num_levels, &mat_pr)); 181 PetscCall(PetscMalloc1(num_levels, &l_size)); 182 PetscCall(PetscMalloc1(num_levels, &xl_size)); 183 PetscCall(PetscMalloc1(num_levels, &g_size)); 184 185 PetscInt c_start, c_end; 186 PetscCall(DMPlexGetHeightStratum(dm_orig, 0, &c_start, &c_end)); 187 DMPolytopeType cell_type; 188 PetscCall(DMPlexGetCellType(dm_orig, c_start, &cell_type)); 189 CeedElemTopology elem_topo = ElemTopologyP2C(cell_type); 190 191 // Setup DM and Operator Mat Shells for each level 192 for (CeedInt i = 0; i < num_levels; i++) { 193 // Create DM 194 PetscCall(DMClone(dm_orig, &dm[i])); 195 PetscCall(DMGetVecType(dm_orig, &vec_type)); 196 PetscCall(DMSetVecType(dm[i], vec_type)); 197 PetscInt dim; 198 PetscCall(DMGetDimension(dm[i], &dim)); 199 PetscCall(SetupDMByDegree(dm[i], level_degrees[fine_level], q_extra, num_comp_u, dim, bp_options[bp_choice].enforce_bc)); 200 201 // Create vectors 202 PetscCall(DMCreateGlobalVector(dm[i], &X[i])); 203 PetscCall(VecGetLocalSize(X[i], &l_size[i])); 204 PetscCall(VecGetSize(X[i], &g_size[i])); 205 PetscCall(DMCreateLocalVector(dm[i], &X_loc[i])); 206 PetscCall(VecGetSize(X_loc[i], &xl_size[i])); 207 208 // Operator 209 PetscCall(PetscMalloc1(1, &op_apply_ctx[i])); 210 PetscCall(PetscMalloc1(1, &op_error_ctx)); 211 PetscCall(MatCreateShell(comm, l_size[i], l_size[i], g_size[i], g_size[i], op_apply_ctx[i], &mat_O[i])); 212 PetscCall(MatShellSetOperation(mat_O[i], MATOP_MULT, (void (*)(void))MatMult_Ceed)); 213 PetscCall(MatShellSetOperation(mat_O[i], MATOP_GET_DIAGONAL, (void (*)(void))MatGetDiag)); 214 PetscCall(MatShellSetVecType(mat_O[i], vec_type)); 215 216 // Level transfers 217 if (i > 0) { 218 // Interp 219 PetscCall(PetscMalloc1(1, &pr_restr_ctx[i])); 220 PetscCall(MatCreateShell(comm, l_size[i], l_size[i - 1], g_size[i], g_size[i - 1], pr_restr_ctx[i], &mat_pr[i])); 221 PetscCall(MatShellSetOperation(mat_pr[i], MATOP_MULT, (void (*)(void))MatMult_Prolong)); 222 PetscCall(MatShellSetOperation(mat_pr[i], MATOP_MULT_TRANSPOSE, (void (*)(void))MatMult_Restrict)); 223 PetscCall(MatShellSetVecType(mat_pr[i], vec_type)); 224 } 225 } 226 PetscCall(VecDuplicate(X[fine_level], &rhs)); 227 228 // Print global grid information 229 if (!test_mode) { 230 PetscInt P = degree + 1, Q = P + q_extra; 231 232 const char *used_resource; 233 CeedGetResource(ceed, &used_resource); 234 235 PetscCall(VecGetType(X[0], &vec_type)); 236 237 PetscCall(PetscPrintf(comm, 238 "\n-- CEED Benchmark Problem %" CeedInt_FMT " -- libCEED + PETSc + PCMG --\n" 239 " PETSc:\n" 240 " PETSc Vec Type : %s\n" 241 " libCEED:\n" 242 " libCEED Backend : %s\n" 243 " libCEED Backend MemType : %s\n" 244 " Mesh:\n" 245 " Solution Order (P) : %" CeedInt_FMT "\n" 246 " Quadrature Order (Q) : %" CeedInt_FMT "\n" 247 " Additional quadrature points (q_extra) : %" CeedInt_FMT "\n" 248 " Global Nodes : %" PetscInt_FMT "\n" 249 " Owned Nodes : %" PetscInt_FMT "\n" 250 " DoF per node : %" PetscInt_FMT "\n" 251 " Element topology : %s\n" 252 " Multigrid:\n" 253 " Number of Levels : %" CeedInt_FMT "\n", 254 bp_choice + 1, vec_type, used_resource, CeedMemTypes[mem_type_backend], P, Q, q_extra, g_size[fine_level] / num_comp_u, 255 l_size[fine_level] / num_comp_u, num_comp_u, CeedElemTopologies[elem_topo], num_levels)); 256 } 257 258 // Create RHS vector 259 PetscCall(VecDuplicate(X_loc[fine_level], &rhs_loc)); 260 PetscCall(VecZeroEntries(rhs_loc)); 261 PetscCall(VecGetArrayAndMemType(rhs_loc, &r, &mem_type)); 262 CeedVectorCreate(ceed, xl_size[fine_level], &rhs_ceed); 263 CeedVectorSetArray(rhs_ceed, MemTypeP2C(mem_type), CEED_USE_POINTER, r); 264 265 // Set up libCEED operators on each level 266 PetscCall(PetscMalloc1(num_levels, &ceed_data)); 267 for (PetscInt i = 0; i < num_levels; i++) { 268 // Print level information 269 if (!test_mode && (i == 0 || i == fine_level)) { 270 PetscCall(PetscPrintf(comm, 271 " Level %" PetscInt_FMT " (%s):\n" 272 " Solution Order (P) : %" CeedInt_FMT "\n" 273 " Global Nodes : %" PetscInt_FMT "\n" 274 " Owned Nodes : %" PetscInt_FMT "\n", 275 i, (i ? "fine" : "coarse"), level_degrees[i] + 1, g_size[i] / num_comp_u, l_size[i] / num_comp_u)); 276 } 277 PetscCall(PetscMalloc1(1, &ceed_data[i])); 278 PetscCall(SetupLibceedByDegree(dm[i], ceed, level_degrees[i], dim, q_extra, dim, num_comp_u, g_size[i], xl_size[i], bp_options[bp_choice], 279 ceed_data[i], i == (fine_level), rhs_ceed, &target)); 280 } 281 282 // Gather RHS 283 CeedVectorTakeArray(rhs_ceed, MemTypeP2C(mem_type), NULL); 284 PetscCall(VecRestoreArrayAndMemType(rhs_loc, &r)); 285 PetscCall(VecZeroEntries(rhs)); 286 PetscCall(DMLocalToGlobal(dm[fine_level], rhs_loc, ADD_VALUES, rhs)); 287 CeedVectorDestroy(&rhs_ceed); 288 289 // Create the error QFunction 290 CeedQFunctionCreateInterior(ceed, 1, bp_options[bp_choice].error, bp_options[bp_choice].error_loc, &qf_error); 291 CeedQFunctionAddInput(qf_error, "u", num_comp_u, CEED_EVAL_INTERP); 292 CeedQFunctionAddInput(qf_error, "true_soln", num_comp_u, CEED_EVAL_NONE); 293 CeedQFunctionAddInput(qf_error, "qdata", ceed_data[fine_level]->q_data_size, CEED_EVAL_NONE); 294 CeedQFunctionAddOutput(qf_error, "error", num_comp_u, CEED_EVAL_INTERP); 295 296 // Create the error operator 297 CeedOperatorCreate(ceed, qf_error, CEED_QFUNCTION_NONE, CEED_QFUNCTION_NONE, &op_error); 298 CeedOperatorSetField(op_error, "u", ceed_data[fine_level]->elem_restr_u, ceed_data[fine_level]->basis_u, CEED_VECTOR_ACTIVE); 299 CeedOperatorSetField(op_error, "true_soln", ceed_data[fine_level]->elem_restr_u_i, CEED_BASIS_COLLOCATED, target); 300 CeedOperatorSetField(op_error, "qdata", ceed_data[fine_level]->elem_restr_qd_i, CEED_BASIS_COLLOCATED, ceed_data[fine_level]->q_data); 301 CeedOperatorSetField(op_error, "error", ceed_data[fine_level]->elem_restr_u, ceed_data[fine_level]->basis_u, CEED_VECTOR_ACTIVE); 302 303 // Calculate multiplicity 304 for (int i = 0; i < num_levels; i++) { 305 PetscScalar *x; 306 307 // CEED vector 308 PetscCall(VecZeroEntries(X_loc[i])); 309 PetscCall(VecGetArray(X_loc[i], &x)); 310 CeedVectorSetArray(ceed_data[i]->x_ceed, CEED_MEM_HOST, CEED_USE_POINTER, x); 311 312 // Multiplicity 313 CeedElemRestrictionGetMultiplicity(ceed_data[i]->elem_restr_u, ceed_data[i]->x_ceed); 314 CeedVectorSyncArray(ceed_data[i]->x_ceed, CEED_MEM_HOST); 315 316 // Restore vector 317 PetscCall(VecRestoreArray(X_loc[i], &x)); 318 319 // Creat mult vector 320 PetscCall(VecDuplicate(X_loc[i], &mult[i])); 321 322 // Local-to-global 323 PetscCall(VecZeroEntries(X[i])); 324 PetscCall(DMLocalToGlobal(dm[i], X_loc[i], ADD_VALUES, X[i])); 325 PetscCall(VecZeroEntries(X_loc[i])); 326 327 // Global-to-local 328 PetscCall(DMGlobalToLocal(dm[i], X[i], INSERT_VALUES, mult[i])); 329 PetscCall(VecZeroEntries(X[i])); 330 331 // Multiplicity scaling 332 PetscCall(VecReciprocal(mult[i])); 333 } 334 335 // Set up Mat 336 for (int i = 0; i < num_levels; i++) { 337 // Set up apply operator context 338 PetscCall(SetupApplyOperatorCtx(comm, dm[i], ceed, ceed_data[i], X_loc[i], op_apply_ctx[i])); 339 340 if (i > 0) { 341 // Prolongation/Restriction Operator 342 PetscCall(CeedLevelTransferSetup(dm[i - 1], ceed, i, num_comp_u, ceed_data, bp_options[bp_choice], mult[i])); 343 pr_restr_ctx[i]->comm = comm; 344 pr_restr_ctx[i]->dmf = dm[i]; 345 pr_restr_ctx[i]->dmc = dm[i - 1]; 346 pr_restr_ctx[i]->loc_vec_c = X_loc[i - 1]; 347 pr_restr_ctx[i]->loc_vec_f = op_apply_ctx[i]->Y_loc; 348 pr_restr_ctx[i]->mult_vec = mult[i]; 349 pr_restr_ctx[i]->ceed_vec_c = op_apply_ctx[i - 1]->x_ceed; 350 pr_restr_ctx[i]->ceed_vec_f = op_apply_ctx[i]->y_ceed; 351 pr_restr_ctx[i]->op_prolong = ceed_data[i]->op_prolong; 352 pr_restr_ctx[i]->op_restrict = ceed_data[i]->op_restrict; 353 pr_restr_ctx[i]->ceed = ceed; 354 } 355 } 356 357 // Assemble coarse grid Jacobian for AMG (or other sparse matrix) solve 358 PetscCall(DMCreateMatrix(dm[0], &mat_coarse)); 359 360 PetscCall(PetscLogEventRegister("AssembleMatrix", MAT_CLASSID, &assemble_event)); 361 { 362 // Assemble matrix analytically 363 PetscCount num_entries; 364 CeedInt *rows, *cols; 365 CeedVector coo_values; 366 CeedOperatorLinearAssembleSymbolic(op_apply_ctx[0]->op, &num_entries, &rows, &cols); 367 ISLocalToGlobalMapping ltog_row, ltog_col; 368 PetscCall(MatGetLocalToGlobalMapping(mat_coarse, <og_row, <og_col)); 369 PetscCall(ISLocalToGlobalMappingApply(ltog_row, num_entries, rows, rows)); 370 PetscCall(ISLocalToGlobalMappingApply(ltog_col, num_entries, cols, cols)); 371 PetscCall(MatSetPreallocationCOO(mat_coarse, num_entries, rows, cols)); 372 free(rows); 373 free(cols); 374 CeedVectorCreate(ceed, num_entries, &coo_values); 375 PetscCall(PetscLogEventBegin(assemble_event, mat_coarse, 0, 0, 0)); 376 CeedOperatorLinearAssemble(op_apply_ctx[0]->op, coo_values); 377 const CeedScalar *values; 378 CeedVectorGetArrayRead(coo_values, CEED_MEM_HOST, &values); 379 PetscCall(MatSetValuesCOO(mat_coarse, values, ADD_VALUES)); 380 CeedVectorRestoreArrayRead(coo_values, &values); 381 PetscCall(PetscLogEventEnd(assemble_event, mat_coarse, 0, 0, 0)); 382 CeedVectorDestroy(&coo_values); 383 } 384 385 // Set up KSP 386 PetscCall(KSPCreate(comm, &ksp)); 387 { 388 PetscCall(KSPSetType(ksp, KSPCG)); 389 PetscCall(KSPSetNormType(ksp, KSP_NORM_NATURAL)); 390 PetscCall(KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, PETSC_DEFAULT)); 391 } 392 PetscCall(KSPSetFromOptions(ksp)); 393 PetscCall(KSPSetOperators(ksp, mat_O[fine_level], mat_O[fine_level])); 394 395 // Set up PCMG 396 PetscCall(KSPGetPC(ksp, &pc)); 397 PCMGCycleType pcmg_cycle_type = PC_MG_CYCLE_V; 398 { 399 PetscCall(PCSetType(pc, PCMG)); 400 401 // PCMG levels 402 PetscCall(PCMGSetLevels(pc, num_levels, NULL)); 403 for (int i = 0; i < num_levels; i++) { 404 // Smoother 405 KSP smoother; 406 PC smoother_pc; 407 PetscCall(PCMGGetSmoother(pc, i, &smoother)); 408 PetscCall(KSPSetType(smoother, KSPCHEBYSHEV)); 409 PetscCall(KSPChebyshevEstEigSet(smoother, 0, 0.1, 0, 1.1)); 410 PetscCall(KSPChebyshevEstEigSetUseNoisy(smoother, PETSC_TRUE)); 411 PetscCall(KSPSetOperators(smoother, mat_O[i], mat_O[i])); 412 PetscCall(KSPGetPC(smoother, &smoother_pc)); 413 PetscCall(PCSetType(smoother_pc, PCJACOBI)); 414 PetscCall(PCJacobiSetType(smoother_pc, PC_JACOBI_DIAGONAL)); 415 416 // Work vector 417 if (i < num_levels - 1) { 418 PetscCall(PCMGSetX(pc, i, X[i])); 419 } 420 421 // Level transfers 422 if (i > 0) { 423 // Interpolation 424 PetscCall(PCMGSetInterpolation(pc, i, mat_pr[i])); 425 } 426 427 // Coarse solve 428 KSP coarse; 429 PC coarse_pc; 430 PetscCall(PCMGGetCoarseSolve(pc, &coarse)); 431 PetscCall(KSPSetType(coarse, KSPPREONLY)); 432 PetscCall(KSPSetOperators(coarse, mat_coarse, mat_coarse)); 433 434 PetscCall(KSPGetPC(coarse, &coarse_pc)); 435 PetscCall(PCSetType(coarse_pc, PCGAMG)); 436 437 PetscCall(KSPSetOptionsPrefix(coarse, "coarse_")); 438 PetscCall(PCSetOptionsPrefix(coarse_pc, "coarse_")); 439 PetscCall(KSPSetFromOptions(coarse)); 440 PetscCall(PCSetFromOptions(coarse_pc)); 441 } 442 443 // PCMG options 444 PetscCall(PCMGSetType(pc, PC_MG_MULTIPLICATIVE)); 445 PetscCall(PCMGSetNumberSmooth(pc, 3)); 446 PetscCall(PCMGSetCycleType(pc, pcmg_cycle_type)); 447 } 448 449 // First run, if benchmarking 450 if (benchmark_mode) { 451 PetscCall(KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, 1)); 452 PetscCall(VecZeroEntries(X[fine_level])); 453 my_rt_start = MPI_Wtime(); 454 PetscCall(KSPSolve(ksp, rhs, X[fine_level])); 455 my_rt = MPI_Wtime() - my_rt_start; 456 PetscCall(MPI_Allreduce(MPI_IN_PLACE, &my_rt, 1, MPI_DOUBLE, MPI_MIN, comm)); 457 // Set maxits based on first iteration timing 458 if (my_rt > 0.02) { 459 PetscCall(KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, 5)); 460 } else { 461 PetscCall(KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, 20)); 462 } 463 } 464 465 // Timed solve 466 PetscCall(VecZeroEntries(X[fine_level])); 467 PetscCall(PetscBarrier((PetscObject)ksp)); 468 469 // -- Performance logging 470 PetscCall(PetscLogStageRegister("Solve Stage", &solve_stage)); 471 PetscCall(PetscLogStagePush(solve_stage)); 472 473 // -- Solve 474 my_rt_start = MPI_Wtime(); 475 PetscCall(KSPSolve(ksp, rhs, X[fine_level])); 476 my_rt = MPI_Wtime() - my_rt_start; 477 478 // -- Performance logging 479 PetscCall(PetscLogStagePop()); 480 481 // Output results 482 { 483 KSPType ksp_type; 484 PCMGType pcmg_type; 485 KSPConvergedReason reason; 486 PetscReal rnorm; 487 PetscInt its; 488 PetscCall(KSPGetType(ksp, &ksp_type)); 489 PetscCall(KSPGetConvergedReason(ksp, &reason)); 490 PetscCall(KSPGetIterationNumber(ksp, &its)); 491 PetscCall(KSPGetResidualNorm(ksp, &rnorm)); 492 PetscCall(PCMGGetType(pc, &pcmg_type)); 493 if (!test_mode || reason < 0 || rnorm > 1e-8) { 494 PetscCall(PetscPrintf(comm, 495 " KSP:\n" 496 " KSP Type : %s\n" 497 " KSP Convergence : %s\n" 498 " Total KSP Iterations : %" PetscInt_FMT "\n" 499 " Final rnorm : %e\n", 500 ksp_type, KSPConvergedReasons[reason], its, (double)rnorm)); 501 PetscCall(PetscPrintf(comm, 502 " PCMG:\n" 503 " PCMG Type : %s\n" 504 " PCMG Cycle Type : %s\n", 505 PCMGTypes[pcmg_type], PCMGCycleTypes[pcmg_cycle_type])); 506 } 507 if (!test_mode) { 508 PetscCall(PetscPrintf(comm, " Performance:\n")); 509 } 510 { 511 // Set up error operator context 512 PetscCall(SetupErrorOperatorCtx(comm, dm[fine_level], ceed, ceed_data[fine_level], X_loc[fine_level], op_error, op_error_ctx)); 513 PetscScalar l2_error; 514 PetscCall(ComputeL2Error(X[fine_level], &l2_error, op_error_ctx)); 515 PetscReal tol = 5e-2; 516 if (!test_mode || l2_error > tol) { 517 PetscCall(MPI_Allreduce(&my_rt, &rt_min, 1, MPI_DOUBLE, MPI_MIN, comm)); 518 PetscCall(MPI_Allreduce(&my_rt, &rt_max, 1, MPI_DOUBLE, MPI_MAX, comm)); 519 PetscCall(PetscPrintf(comm, 520 " L2 Error : %e\n" 521 " CG Solve Time : %g (%g) sec\n", 522 (double)l2_error, rt_max, rt_min)); 523 } 524 } 525 if (benchmark_mode && (!test_mode)) { 526 PetscCall(PetscPrintf(comm, " DoFs/Sec in CG : %g (%g) million\n", 1e-6 * g_size[fine_level] * its / rt_max, 527 1e-6 * g_size[fine_level] * its / rt_min)); 528 } 529 } 530 531 if (write_solution) { 532 PetscViewer vtk_viewer_soln; 533 534 PetscCall(PetscViewerCreate(comm, &vtk_viewer_soln)); 535 PetscCall(PetscViewerSetType(vtk_viewer_soln, PETSCVIEWERVTK)); 536 PetscCall(PetscViewerFileSetName(vtk_viewer_soln, "solution.vtu")); 537 PetscCall(VecView(X[fine_level], vtk_viewer_soln)); 538 PetscCall(PetscViewerDestroy(&vtk_viewer_soln)); 539 } 540 541 // Cleanup 542 for (int i = 0; i < num_levels; i++) { 543 PetscCall(VecDestroy(&X[i])); 544 PetscCall(VecDestroy(&X_loc[i])); 545 PetscCall(VecDestroy(&mult[i])); 546 PetscCall(VecDestroy(&op_apply_ctx[i]->Y_loc)); 547 PetscCall(MatDestroy(&mat_O[i])); 548 PetscCall(PetscFree(op_apply_ctx[i])); 549 if (i > 0) { 550 PetscCall(MatDestroy(&mat_pr[i])); 551 PetscCall(PetscFree(pr_restr_ctx[i])); 552 } 553 PetscCall(CeedDataDestroy(i, ceed_data[i])); 554 PetscCall(DMDestroy(&dm[i])); 555 } 556 PetscCall(PetscFree(level_degrees)); 557 PetscCall(PetscFree(dm)); 558 PetscCall(PetscFree(X)); 559 PetscCall(PetscFree(X_loc)); 560 PetscCall(VecDestroy(&op_error_ctx->Y_loc)); 561 PetscCall(PetscFree(mult)); 562 PetscCall(PetscFree(mat_O)); 563 PetscCall(PetscFree(mat_pr)); 564 PetscCall(PetscFree(ceed_data)); 565 PetscCall(PetscFree(op_apply_ctx)); 566 PetscCall(PetscFree(op_error_ctx)); 567 PetscCall(PetscFree(pr_restr_ctx)); 568 PetscCall(PetscFree(l_size)); 569 PetscCall(PetscFree(xl_size)); 570 PetscCall(PetscFree(g_size)); 571 PetscCall(VecDestroy(&rhs)); 572 PetscCall(VecDestroy(&rhs_loc)); 573 PetscCall(MatDestroy(&mat_coarse)); 574 PetscCall(KSPDestroy(&ksp)); 575 PetscCall(DMDestroy(&dm_orig)); 576 CeedVectorDestroy(&target); 577 CeedQFunctionDestroy(&qf_error); 578 CeedOperatorDestroy(&op_error); 579 CeedDestroy(&ceed); 580 return PetscFinalize(); 581 } 582