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 11 // CEED BP benchmark problems, see http://ceed.exascaleproject.org/bps. 12 // 13 // The code uses higher level communication protocols in DMPlex. 14 // 15 // Build with: 16 // 17 // make multigrid [PETSC_DIR=</path/to/petsc>] [CEED_DIR=</path/to/libceed>] 18 // 19 // Sample runs: 20 // 21 // multigrid -problem bp3 22 // multigrid -problem bp4 23 // multigrid -problem bp5 -ceed /cpu/self 24 // multigrid -problem bp6 -ceed /gpu/cuda 25 // 26 //TESTARGS -ceed {ceed_resource} -test -problem bp3 -degree 3 27 //TESTARGS -ceed {ceed_resource} -test -problem bp3 -degree 3 -simplex 28 29 /// @file 30 /// CEED BPs 1-6 multigrid example using PETSc 31 const char help[] = "Solve CEED BPs using p-multigrid with PETSc and DMPlex\n"; 32 33 #include <ceed.h> 34 #include <petsc.h> 35 #include <petscdmplex.h> 36 #include <petscksp.h> 37 #include <petscsys.h> 38 #include <stdbool.h> 39 #include <string.h> 40 41 #include "bps.h" 42 #include "include/bpsproblemdata.h" 43 #include "include/libceedsetup.h" 44 #include "include/matops.h" 45 #include "include/petscutils.h" 46 #include "include/petscversion.h" 47 #include "include/structs.h" 48 49 #if PETSC_VERSION_LT(3, 12, 0) 50 #ifdef PETSC_HAVE_CUDA 51 #include <petsccuda.h> 52 // Note: With PETSc prior to version 3.12.0, providing the source path to 53 // include 'cublas_v2.h' will be needed to use 'petsccuda.h'. 54 #endif 55 #endif 56 57 int main(int argc, char **argv) { 58 MPI_Comm comm; 59 char filename[PETSC_MAX_PATH_LEN], ceed_resource[PETSC_MAX_PATH_LEN] = "/cpu/self"; 60 double my_rt_start, my_rt, rt_min, rt_max; 61 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, 62 *level_degrees; 63 PetscScalar *r; 64 PetscScalar eps = 1.0; 65 PetscBool test_mode, benchmark_mode, read_mesh, write_solution, simplex; 66 PetscLogStage solve_stage; 67 PetscLogEvent assemble_event; 68 DM *dm, dm_orig; 69 KSP ksp; 70 PC pc; 71 Mat *mat_O, *mat_pr, mat_coarse; 72 Vec *X, *X_loc, *mult, rhs, rhs_loc; 73 PetscMemType mem_type; 74 OperatorApplyContext *op_apply_ctx, op_error_ctx; 75 ProlongRestrContext *pr_restr_ctx; 76 Ceed ceed; 77 CeedData *ceed_data; 78 CeedVector rhs_ceed, target; 79 CeedQFunction qf_error; 80 CeedOperator op_error; 81 BPType bp_choice; 82 CoarsenType coarsen; 83 84 PetscCall(PetscInitialize(&argc, &argv, NULL, help)); 85 comm = PETSC_COMM_WORLD; 86 87 // Parse command line options 88 PetscOptionsBegin(comm, NULL, "CEED BPs in PETSc", NULL); 89 bp_choice = CEED_BP3; 90 PetscCall(PetscOptionsEnum("-problem", "CEED benchmark problem to solve", NULL, bp_types, (PetscEnum)bp_choice, (PetscEnum *)&bp_choice, NULL)); 91 num_comp_u = bp_options[bp_choice].num_comp_u; 92 test_mode = PETSC_FALSE; 93 PetscCall(PetscOptionsBool("-test", "Testing mode (do not print unless error is large)", NULL, test_mode, &test_mode, NULL)); 94 benchmark_mode = PETSC_FALSE; 95 PetscCall(PetscOptionsBool("-benchmark", "Benchmarking mode (prints benchmark statistics)", NULL, benchmark_mode, &benchmark_mode, NULL)); 96 write_solution = PETSC_FALSE; 97 PetscCall(PetscOptionsBool("-write_solution", "Write solution for visualization", NULL, write_solution, &write_solution, NULL)); 98 simplex = PETSC_FALSE; 99 PetscCall(PetscOptionsBool("-simplex", "Element topology (default:hex)", NULL, simplex, &simplex, NULL)); 100 if ((bp_choice == CEED_BP5 || bp_choice == CEED_BP6) && (simplex)) { 101 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "BP5/6 is not supported with simplex"); 102 } 103 PetscCall(PetscOptionsScalar("-eps", "Epsilon parameter for Kershaw mesh transformation", NULL, eps, &eps, NULL)); 104 if (eps > 1 || eps <= 0) SETERRQ(PETSC_COMM_WORLD, PETSC_ERR_ARG_OUTOFRANGE, "-eps %g must be (0,1]", (double)PetscRealPart(eps)); 105 degree = test_mode ? 3 : 2; 106 PetscCall(PetscOptionsInt("-degree", "Polynomial degree of tensor product basis", NULL, degree, °ree, NULL)); 107 if (degree < 1) SETERRQ(PETSC_COMM_WORLD, PETSC_ERR_ARG_OUTOFRANGE, "-degree %" PetscInt_FMT " must be at least 1", degree); 108 q_extra = bp_options[bp_choice].q_extra; 109 PetscCall(PetscOptionsInt("-q_extra", "Number of extra quadrature points", NULL, q_extra, &q_extra, NULL)); 110 PetscCall(PetscOptionsString("-ceed", "CEED resource specifier", NULL, ceed_resource, ceed_resource, sizeof(ceed_resource), NULL)); 111 coarsen = COARSEN_UNIFORM; 112 PetscCall(PetscOptionsEnum("-coarsen", "Coarsening strategy to use", NULL, coarsen_types, (PetscEnum)coarsen, (PetscEnum *)&coarsen, NULL)); 113 read_mesh = PETSC_FALSE; 114 PetscCall(PetscOptionsString("-mesh", "Read mesh from file", NULL, filename, filename, sizeof(filename), &read_mesh)); 115 if (!read_mesh) { 116 PetscInt tmp = dim; 117 PetscCall(PetscOptionsIntArray("-cells", "Number of cells per dimension", NULL, mesh_elem, &tmp, NULL)); 118 } 119 PetscOptionsEnd(); 120 121 // Set up libCEED 122 CeedInit(ceed_resource, &ceed); 123 CeedMemType mem_type_backend; 124 CeedGetPreferredMemType(ceed, &mem_type_backend); 125 126 // Setup DM 127 if (read_mesh) { 128 PetscCall(DMPlexCreateFromFile(PETSC_COMM_WORLD, filename, NULL, PETSC_TRUE, &dm_orig)); 129 } else { 130 PetscCall(DMPlexCreateBoxMesh(PETSC_COMM_WORLD, dim, simplex, mesh_elem, NULL, NULL, NULL, PETSC_TRUE, &dm_orig)); 131 } 132 133 VecType vec_type; 134 switch (mem_type_backend) { 135 case CEED_MEM_HOST: 136 vec_type = VECSTANDARD; 137 break; 138 case CEED_MEM_DEVICE: { 139 const char *resolved; 140 CeedGetResource(ceed, &resolved); 141 if (strstr(resolved, "/gpu/cuda")) vec_type = VECCUDA; 142 else if (strstr(resolved, "/gpu/hip/occa")) vec_type = VECSTANDARD; // https://github.com/CEED/libCEED/issues/678 143 else if (strstr(resolved, "/gpu/hip")) vec_type = VECHIP; 144 else vec_type = VECSTANDARD; 145 } 146 } 147 PetscCall(DMSetVecType(dm_orig, vec_type)); 148 PetscCall(DMSetFromOptions(dm_orig)); 149 PetscCall(DMViewFromOptions(dm_orig, NULL, "-dm_view")); 150 151 // Apply Kershaw mesh transformation 152 PetscCall(Kershaw(dm_orig, eps)); 153 154 // Allocate arrays for PETSc objects for each level 155 switch (coarsen) { 156 case COARSEN_UNIFORM: 157 num_levels = degree; 158 break; 159 case COARSEN_LOGARITHMIC: 160 num_levels = ceil(log(degree) / log(2)) + 1; 161 break; 162 } 163 PetscCall(PetscMalloc1(num_levels, &level_degrees)); 164 fine_level = num_levels - 1; 165 166 switch (coarsen) { 167 case COARSEN_UNIFORM: 168 for (int i = 0; i < num_levels; i++) level_degrees[i] = i + 1; 169 break; 170 case COARSEN_LOGARITHMIC: 171 for (int i = 0; i < num_levels - 1; i++) level_degrees[i] = pow(2, i); 172 level_degrees[fine_level] = degree; 173 break; 174 } 175 PetscCall(PetscMalloc1(num_levels, &dm)); 176 PetscCall(PetscMalloc1(num_levels, &X)); 177 PetscCall(PetscMalloc1(num_levels, &X_loc)); 178 PetscCall(PetscMalloc1(num_levels, &mult)); 179 PetscCall(PetscMalloc1(num_levels, &op_apply_ctx)); 180 PetscCall(PetscMalloc1(num_levels, &pr_restr_ctx)); 181 PetscCall(PetscMalloc1(num_levels, &mat_O)); 182 PetscCall(PetscMalloc1(num_levels, &mat_pr)); 183 PetscCall(PetscMalloc1(num_levels, &l_size)); 184 PetscCall(PetscMalloc1(num_levels, &xl_size)); 185 PetscCall(PetscMalloc1(num_levels, &g_size)); 186 187 PetscInt c_start, c_end; 188 PetscCall(DMPlexGetHeightStratum(dm_orig, 0, &c_start, &c_end)); 189 DMPolytopeType cell_type; 190 PetscCall(DMPlexGetCellType(dm_orig, c_start, &cell_type)); 191 CeedElemTopology elem_topo = ElemTopologyP2C(cell_type); 192 193 // Setup DM and Operator Mat Shells for each level 194 for (CeedInt i = 0; i < num_levels; i++) { 195 // Create DM 196 PetscCall(DMClone(dm_orig, &dm[i])); 197 PetscCall(DMGetVecType(dm_orig, &vec_type)); 198 PetscCall(DMSetVecType(dm[i], vec_type)); 199 PetscInt dim; 200 PetscCall(DMGetDimension(dm[i], &dim)); 201 PetscCall(SetupDMByDegree(dm[i], level_degrees[fine_level], q_extra, num_comp_u, dim, bp_options[bp_choice].enforce_bc)); 202 203 // Create vectors 204 PetscCall(DMCreateGlobalVector(dm[i], &X[i])); 205 PetscCall(VecGetLocalSize(X[i], &l_size[i])); 206 PetscCall(VecGetSize(X[i], &g_size[i])); 207 PetscCall(DMCreateLocalVector(dm[i], &X_loc[i])); 208 PetscCall(VecGetSize(X_loc[i], &xl_size[i])); 209 210 // Operator 211 PetscCall(PetscMalloc1(1, &op_apply_ctx[i])); 212 PetscCall(PetscMalloc1(1, &op_error_ctx)); 213 PetscCall(MatCreateShell(comm, l_size[i], l_size[i], g_size[i], g_size[i], op_apply_ctx[i], &mat_O[i])); 214 PetscCall(MatShellSetOperation(mat_O[i], MATOP_MULT, (void (*)(void))MatMult_Ceed)); 215 PetscCall(MatShellSetOperation(mat_O[i], MATOP_GET_DIAGONAL, (void (*)(void))MatGetDiag)); 216 PetscCall(MatShellSetVecType(mat_O[i], vec_type)); 217 218 // Level transfers 219 if (i > 0) { 220 // Interp 221 PetscCall(PetscMalloc1(1, &pr_restr_ctx[i])); 222 PetscCall(MatCreateShell(comm, l_size[i], l_size[i - 1], g_size[i], g_size[i - 1], pr_restr_ctx[i], &mat_pr[i])); 223 PetscCall(MatShellSetOperation(mat_pr[i], MATOP_MULT, (void (*)(void))MatMult_Prolong)); 224 PetscCall(MatShellSetOperation(mat_pr[i], MATOP_MULT_TRANSPOSE, (void (*)(void))MatMult_Restrict)); 225 PetscCall(MatShellSetVecType(mat_pr[i], vec_type)); 226 } 227 } 228 PetscCall(VecDuplicate(X[fine_level], &rhs)); 229 230 // Print global grid information 231 if (!test_mode) { 232 PetscInt P = degree + 1, Q = P + q_extra; 233 234 const char *used_resource; 235 CeedGetResource(ceed, &used_resource); 236 237 PetscCall(VecGetType(X[0], &vec_type)); 238 239 PetscCall(PetscPrintf(comm, 240 "\n-- CEED Benchmark Problem %" CeedInt_FMT " -- libCEED + PETSc + PCMG --\n" 241 " PETSc:\n" 242 " PETSc Vec Type : %s\n" 243 " libCEED:\n" 244 " libCEED Backend : %s\n" 245 " libCEED Backend MemType : %s\n" 246 " Mesh:\n" 247 " Solution Order (P) : %" CeedInt_FMT "\n" 248 " Quadrature Order (Q) : %" CeedInt_FMT "\n" 249 " Additional quadrature points (q_extra) : %" CeedInt_FMT "\n" 250 " Global Nodes : %" PetscInt_FMT "\n" 251 " Owned Nodes : %" PetscInt_FMT "\n" 252 " DoF per node : %" PetscInt_FMT "\n" 253 " Element topology : %s\n" 254 " Multigrid:\n" 255 " Number of Levels : %" CeedInt_FMT "\n", 256 bp_choice + 1, vec_type, used_resource, CeedMemTypes[mem_type_backend], P, Q, q_extra, g_size[fine_level] / num_comp_u, 257 l_size[fine_level] / num_comp_u, num_comp_u, CeedElemTopologies[elem_topo], num_levels)); 258 } 259 260 // Create RHS vector 261 PetscCall(VecDuplicate(X_loc[fine_level], &rhs_loc)); 262 PetscCall(VecZeroEntries(rhs_loc)); 263 PetscCall(VecGetArrayAndMemType(rhs_loc, &r, &mem_type)); 264 CeedVectorCreate(ceed, xl_size[fine_level], &rhs_ceed); 265 CeedVectorSetArray(rhs_ceed, MemTypeP2C(mem_type), CEED_USE_POINTER, r); 266 267 // Set up libCEED operators on each level 268 PetscCall(PetscMalloc1(num_levels, &ceed_data)); 269 for (PetscInt i = 0; i < num_levels; i++) { 270 // Print level information 271 if (!test_mode && (i == 0 || i == fine_level)) { 272 PetscCall(PetscPrintf(comm, 273 " Level %" PetscInt_FMT " (%s):\n" 274 " Solution Order (P) : %" CeedInt_FMT "\n" 275 " Global Nodes : %" PetscInt_FMT "\n" 276 " Owned Nodes : %" PetscInt_FMT "\n", 277 i, (i ? "fine" : "coarse"), level_degrees[i] + 1, g_size[i] / num_comp_u, l_size[i] / num_comp_u)); 278 } 279 PetscCall(PetscMalloc1(1, &ceed_data[i])); 280 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], 281 ceed_data[i], i == (fine_level), rhs_ceed, &target)); 282 } 283 284 // Gather RHS 285 CeedVectorTakeArray(rhs_ceed, MemTypeP2C(mem_type), NULL); 286 PetscCall(VecRestoreArrayAndMemType(rhs_loc, &r)); 287 PetscCall(VecZeroEntries(rhs)); 288 PetscCall(DMLocalToGlobal(dm[fine_level], rhs_loc, ADD_VALUES, rhs)); 289 CeedVectorDestroy(&rhs_ceed); 290 291 // Create the error QFunction 292 CeedQFunctionCreateInterior(ceed, 1, bp_options[bp_choice].error, bp_options[bp_choice].error_loc, &qf_error); 293 CeedQFunctionAddInput(qf_error, "u", num_comp_u, CEED_EVAL_INTERP); 294 CeedQFunctionAddInput(qf_error, "true_soln", num_comp_u, CEED_EVAL_NONE); 295 CeedQFunctionAddInput(qf_error, "qdata", ceed_data[fine_level]->q_data_size, CEED_EVAL_NONE); 296 CeedQFunctionAddOutput(qf_error, "error", num_comp_u, CEED_EVAL_INTERP); 297 298 // Create the error operator 299 CeedOperatorCreate(ceed, qf_error, CEED_QFUNCTION_NONE, CEED_QFUNCTION_NONE, &op_error); 300 CeedOperatorSetField(op_error, "u", ceed_data[fine_level]->elem_restr_u, ceed_data[fine_level]->basis_u, CEED_VECTOR_ACTIVE); 301 CeedOperatorSetField(op_error, "true_soln", ceed_data[fine_level]->elem_restr_u_i, CEED_BASIS_COLLOCATED, target); 302 CeedOperatorSetField(op_error, "qdata", ceed_data[fine_level]->elem_restr_qd_i, CEED_BASIS_COLLOCATED, ceed_data[fine_level]->q_data); 303 CeedOperatorSetField(op_error, "error", ceed_data[fine_level]->elem_restr_u, ceed_data[fine_level]->basis_u, CEED_VECTOR_ACTIVE); 304 305 // Calculate multiplicity 306 for (int i = 0; i < num_levels; i++) { 307 PetscScalar *x; 308 309 // CEED vector 310 PetscCall(VecZeroEntries(X_loc[i])); 311 PetscCall(VecGetArray(X_loc[i], &x)); 312 CeedVectorSetArray(ceed_data[i]->x_ceed, CEED_MEM_HOST, CEED_USE_POINTER, x); 313 314 // Multiplicity 315 CeedElemRestrictionGetMultiplicity(ceed_data[i]->elem_restr_u, ceed_data[i]->x_ceed); 316 CeedVectorSyncArray(ceed_data[i]->x_ceed, CEED_MEM_HOST); 317 318 // Restore vector 319 PetscCall(VecRestoreArray(X_loc[i], &x)); 320 321 // Creat mult vector 322 PetscCall(VecDuplicate(X_loc[i], &mult[i])); 323 324 // Local-to-global 325 PetscCall(VecZeroEntries(X[i])); 326 PetscCall(DMLocalToGlobal(dm[i], X_loc[i], ADD_VALUES, X[i])); 327 PetscCall(VecZeroEntries(X_loc[i])); 328 329 // Global-to-local 330 PetscCall(DMGlobalToLocal(dm[i], X[i], INSERT_VALUES, mult[i])); 331 PetscCall(VecZeroEntries(X[i])); 332 333 // Multiplicity scaling 334 PetscCall(VecReciprocal(mult[i])); 335 } 336 337 // Set up Mat 338 for (int i = 0; i < num_levels; i++) { 339 // Set up apply operator context 340 PetscCall(SetupApplyOperatorCtx(comm, dm[i], ceed, ceed_data[i], X_loc[i], op_apply_ctx[i])); 341 342 if (i > 0) { 343 // Prolongation/Restriction Operator 344 PetscCall(CeedLevelTransferSetup(dm[i - 1], ceed, i, num_comp_u, ceed_data, bp_options[bp_choice], mult[i])); 345 pr_restr_ctx[i]->comm = comm; 346 pr_restr_ctx[i]->dmf = dm[i]; 347 pr_restr_ctx[i]->dmc = dm[i - 1]; 348 pr_restr_ctx[i]->loc_vec_c = X_loc[i - 1]; 349 pr_restr_ctx[i]->loc_vec_f = op_apply_ctx[i]->Y_loc; 350 pr_restr_ctx[i]->mult_vec = mult[i]; 351 pr_restr_ctx[i]->ceed_vec_c = op_apply_ctx[i - 1]->x_ceed; 352 pr_restr_ctx[i]->ceed_vec_f = op_apply_ctx[i]->y_ceed; 353 pr_restr_ctx[i]->op_prolong = ceed_data[i]->op_prolong; 354 pr_restr_ctx[i]->op_restrict = ceed_data[i]->op_restrict; 355 pr_restr_ctx[i]->ceed = ceed; 356 } 357 } 358 359 // Assemble coarse grid Jacobian for AMG (or other sparse matrix) solve 360 PetscCall(DMCreateMatrix(dm[0], &mat_coarse)); 361 362 PetscCall(PetscLogEventRegister("AssembleMatrix", MAT_CLASSID, &assemble_event)); 363 { 364 // Assemble matrix analytically 365 PetscCount num_entries; 366 CeedInt *rows, *cols; 367 CeedVector coo_values; 368 CeedOperatorLinearAssembleSymbolic(op_apply_ctx[0]->op, &num_entries, &rows, &cols); 369 ISLocalToGlobalMapping ltog_row, ltog_col; 370 PetscCall(MatGetLocalToGlobalMapping(mat_coarse, <og_row, <og_col)); 371 PetscCall(ISLocalToGlobalMappingApply(ltog_row, num_entries, rows, rows)); 372 PetscCall(ISLocalToGlobalMappingApply(ltog_col, num_entries, cols, cols)); 373 PetscCall(MatSetPreallocationCOO(mat_coarse, num_entries, rows, cols)); 374 free(rows); 375 free(cols); 376 CeedVectorCreate(ceed, num_entries, &coo_values); 377 PetscCall(PetscLogEventBegin(assemble_event, mat_coarse, 0, 0, 0)); 378 CeedOperatorLinearAssemble(op_apply_ctx[0]->op, coo_values); 379 const CeedScalar *values; 380 CeedVectorGetArrayRead(coo_values, CEED_MEM_HOST, &values); 381 PetscCall(MatSetValuesCOO(mat_coarse, values, ADD_VALUES)); 382 CeedVectorRestoreArrayRead(coo_values, &values); 383 PetscCall(PetscLogEventEnd(assemble_event, mat_coarse, 0, 0, 0)); 384 CeedVectorDestroy(&coo_values); 385 } 386 387 // Set up KSP 388 PetscCall(KSPCreate(comm, &ksp)); 389 { 390 PetscCall(KSPSetType(ksp, KSPCG)); 391 PetscCall(KSPSetNormType(ksp, KSP_NORM_NATURAL)); 392 PetscCall(KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, PETSC_DEFAULT)); 393 } 394 PetscCall(KSPSetFromOptions(ksp)); 395 PetscCall(KSPSetOperators(ksp, mat_O[fine_level], mat_O[fine_level])); 396 397 // Set up PCMG 398 PetscCall(KSPGetPC(ksp, &pc)); 399 PCMGCycleType pcmg_cycle_type = PC_MG_CYCLE_V; 400 { 401 PetscCall(PCSetType(pc, PCMG)); 402 403 // PCMG levels 404 PetscCall(PCMGSetLevels(pc, num_levels, NULL)); 405 for (int i = 0; i < num_levels; i++) { 406 // Smoother 407 KSP smoother; 408 PC smoother_pc; 409 PetscCall(PCMGGetSmoother(pc, i, &smoother)); 410 PetscCall(KSPSetType(smoother, KSPCHEBYSHEV)); 411 PetscCall(KSPChebyshevEstEigSet(smoother, 0, 0.1, 0, 1.1)); 412 PetscCall(KSPChebyshevEstEigSetUseNoisy(smoother, PETSC_TRUE)); 413 PetscCall(KSPSetOperators(smoother, mat_O[i], mat_O[i])); 414 PetscCall(KSPGetPC(smoother, &smoother_pc)); 415 PetscCall(PCSetType(smoother_pc, PCJACOBI)); 416 PetscCall(PCJacobiSetType(smoother_pc, PC_JACOBI_DIAGONAL)); 417 418 // Work vector 419 if (i < num_levels - 1) { 420 PetscCall(PCMGSetX(pc, i, X[i])); 421 } 422 423 // Level transfers 424 if (i > 0) { 425 // Interpolation 426 PetscCall(PCMGSetInterpolation(pc, i, mat_pr[i])); 427 } 428 429 // Coarse solve 430 KSP coarse; 431 PC coarse_pc; 432 PetscCall(PCMGGetCoarseSolve(pc, &coarse)); 433 PetscCall(KSPSetType(coarse, KSPPREONLY)); 434 PetscCall(KSPSetOperators(coarse, mat_coarse, mat_coarse)); 435 436 PetscCall(KSPGetPC(coarse, &coarse_pc)); 437 PetscCall(PCSetType(coarse_pc, PCGAMG)); 438 439 PetscCall(KSPSetOptionsPrefix(coarse, "coarse_")); 440 PetscCall(PCSetOptionsPrefix(coarse_pc, "coarse_")); 441 PetscCall(KSPSetFromOptions(coarse)); 442 PetscCall(PCSetFromOptions(coarse_pc)); 443 } 444 445 // PCMG options 446 PetscCall(PCMGSetType(pc, PC_MG_MULTIPLICATIVE)); 447 PetscCall(PCMGSetNumberSmooth(pc, 3)); 448 PetscCall(PCMGSetCycleType(pc, pcmg_cycle_type)); 449 } 450 451 // First run, if benchmarking 452 if (benchmark_mode) { 453 PetscCall(KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, 1)); 454 PetscCall(VecZeroEntries(X[fine_level])); 455 my_rt_start = MPI_Wtime(); 456 PetscCall(KSPSolve(ksp, rhs, X[fine_level])); 457 my_rt = MPI_Wtime() - my_rt_start; 458 PetscCall(MPI_Allreduce(MPI_IN_PLACE, &my_rt, 1, MPI_DOUBLE, MPI_MIN, comm)); 459 // Set maxits based on first iteration timing 460 if (my_rt > 0.02) { 461 PetscCall(KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, 5)); 462 } else { 463 PetscCall(KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, 20)); 464 } 465 } 466 467 // Timed solve 468 PetscCall(VecZeroEntries(X[fine_level])); 469 PetscCall(PetscBarrier((PetscObject)ksp)); 470 471 // -- Performance logging 472 PetscCall(PetscLogStageRegister("Solve Stage", &solve_stage)); 473 PetscCall(PetscLogStagePush(solve_stage)); 474 475 // -- Solve 476 my_rt_start = MPI_Wtime(); 477 PetscCall(KSPSolve(ksp, rhs, X[fine_level])); 478 my_rt = MPI_Wtime() - my_rt_start; 479 480 // -- Performance logging 481 PetscCall(PetscLogStagePop()); 482 483 // Output results 484 { 485 KSPType ksp_type; 486 PCMGType pcmg_type; 487 KSPConvergedReason reason; 488 PetscReal rnorm; 489 PetscInt its; 490 PetscCall(KSPGetType(ksp, &ksp_type)); 491 PetscCall(KSPGetConvergedReason(ksp, &reason)); 492 PetscCall(KSPGetIterationNumber(ksp, &its)); 493 PetscCall(KSPGetResidualNorm(ksp, &rnorm)); 494 PetscCall(PCMGGetType(pc, &pcmg_type)); 495 if (!test_mode || reason < 0 || rnorm > 1e-8) { 496 PetscCall(PetscPrintf(comm, 497 " KSP:\n" 498 " KSP Type : %s\n" 499 " KSP Convergence : %s\n" 500 " Total KSP Iterations : %" PetscInt_FMT "\n" 501 " Final rnorm : %e\n", 502 ksp_type, KSPConvergedReasons[reason], its, (double)rnorm)); 503 PetscCall(PetscPrintf(comm, 504 " PCMG:\n" 505 " PCMG Type : %s\n" 506 " PCMG Cycle Type : %s\n", 507 PCMGTypes[pcmg_type], PCMGCycleTypes[pcmg_cycle_type])); 508 } 509 if (!test_mode) { 510 PetscCall(PetscPrintf(comm, " Performance:\n")); 511 } 512 { 513 // Set up error operator context 514 PetscCall(SetupErrorOperatorCtx(comm, dm[fine_level], ceed, ceed_data[fine_level], X_loc[fine_level], op_error, op_error_ctx)); 515 PetscScalar l2_error; 516 PetscCall(ComputeL2Error(X[fine_level], &l2_error, op_error_ctx)); 517 PetscReal tol = 5e-2; 518 if (!test_mode || l2_error > tol) { 519 PetscCall(MPI_Allreduce(&my_rt, &rt_min, 1, MPI_DOUBLE, MPI_MIN, comm)); 520 PetscCall(MPI_Allreduce(&my_rt, &rt_max, 1, MPI_DOUBLE, MPI_MAX, comm)); 521 PetscCall(PetscPrintf(comm, 522 " L2 Error : %e\n" 523 " CG Solve Time : %g (%g) sec\n", 524 (double)l2_error, rt_max, rt_min)); 525 } 526 } 527 if (benchmark_mode && (!test_mode)) { 528 PetscCall(PetscPrintf(comm, " DoFs/Sec in CG : %g (%g) million\n", 1e-6 * g_size[fine_level] * its / rt_max, 529 1e-6 * g_size[fine_level] * its / rt_min)); 530 } 531 } 532 533 if (write_solution) { 534 PetscViewer vtk_viewer_soln; 535 536 PetscCall(PetscViewerCreate(comm, &vtk_viewer_soln)); 537 PetscCall(PetscViewerSetType(vtk_viewer_soln, PETSCVIEWERVTK)); 538 PetscCall(PetscViewerFileSetName(vtk_viewer_soln, "solution.vtu")); 539 PetscCall(VecView(X[fine_level], vtk_viewer_soln)); 540 PetscCall(PetscViewerDestroy(&vtk_viewer_soln)); 541 } 542 543 // Cleanup 544 for (int i = 0; i < num_levels; i++) { 545 PetscCall(VecDestroy(&X[i])); 546 PetscCall(VecDestroy(&X_loc[i])); 547 PetscCall(VecDestroy(&mult[i])); 548 PetscCall(VecDestroy(&op_apply_ctx[i]->Y_loc)); 549 PetscCall(MatDestroy(&mat_O[i])); 550 PetscCall(PetscFree(op_apply_ctx[i])); 551 if (i > 0) { 552 PetscCall(MatDestroy(&mat_pr[i])); 553 PetscCall(PetscFree(pr_restr_ctx[i])); 554 } 555 PetscCall(CeedDataDestroy(i, ceed_data[i])); 556 PetscCall(DMDestroy(&dm[i])); 557 } 558 PetscCall(PetscFree(level_degrees)); 559 PetscCall(PetscFree(dm)); 560 PetscCall(PetscFree(X)); 561 PetscCall(PetscFree(X_loc)); 562 PetscCall(VecDestroy(&op_error_ctx->Y_loc)); 563 PetscCall(PetscFree(mult)); 564 PetscCall(PetscFree(mat_O)); 565 PetscCall(PetscFree(mat_pr)); 566 PetscCall(PetscFree(ceed_data)); 567 PetscCall(PetscFree(op_apply_ctx)); 568 PetscCall(PetscFree(op_error_ctx)); 569 PetscCall(PetscFree(pr_restr_ctx)); 570 PetscCall(PetscFree(l_size)); 571 PetscCall(PetscFree(xl_size)); 572 PetscCall(PetscFree(g_size)); 573 PetscCall(VecDestroy(&rhs)); 574 PetscCall(VecDestroy(&rhs_loc)); 575 PetscCall(MatDestroy(&mat_coarse)); 576 PetscCall(KSPDestroy(&ksp)); 577 PetscCall(DMDestroy(&dm_orig)); 578 CeedVectorDestroy(&target); 579 CeedQFunctionDestroy(&qf_error); 580 CeedOperatorDestroy(&op_error); 581 CeedDestroy(&ceed); 582 return PetscFinalize(); 583 } 584