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 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 <stdbool.h> 33 #include <string.h> 34 #include <ceed.h> 35 #include <petsc.h> 36 #include <petscdmplex.h> 37 #include <petscksp.h> 38 #include <petscsys.h> 39 40 #include "bps.h" 41 #include "include/bpsproblemdata.h" 42 #include "include/petscutils.h" 43 #include "include/petscversion.h" 44 #include "include/matops.h" 45 #include "include/structs.h" 46 #include "include/libceedsetup.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 52 // include 'cublas_v2.h' will be needed to use 'petsccuda.h'. 53 #endif 54 #endif 55 56 int main(int argc, char **argv) { 57 PetscInt ierr; 58 MPI_Comm comm; 59 char filename[PETSC_MAX_PATH_LEN], 60 ceed_resource[PETSC_MAX_PATH_LEN] = "/cpu/self"; 61 double my_rt_start, my_rt, rt_min, rt_max; 62 PetscInt degree = 3, q_extra, *l_size, *xl_size, *g_size, dim = 3, fine_level, 63 mesh_elem[3] = {3, 3, 3}, num_comp_u = 1, num_levels = degree, *level_degrees; 64 PetscScalar *r; 65 PetscScalar eps = 1.0; 66 PetscBool test_mode, benchmark_mode, read_mesh, write_solution; 67 PetscLogStage solve_stage; 68 PetscLogEvent assemble_event; 69 DM *dm, dm_orig; 70 KSP ksp; 71 PC pc; 72 Mat *mat_O, *mat_pr, mat_coarse; 73 Vec *X, *X_loc, *mult, rhs, rhs_loc; 74 PetscMemType mem_type; 75 UserO *user_O; 76 UserProlongRestr *user_pr; 77 Ceed ceed; 78 CeedData *ceed_data; 79 CeedVector rhs_ceed, target; 80 CeedQFunction qf_error, qf_restrict, qf_prolong; 81 CeedOperator op_error; 82 BPType bp_choice; 83 CoarsenType coarsen; 84 85 ierr = PetscInitialize(&argc, &argv, NULL, help); 86 if (ierr) return ierr; 87 comm = PETSC_COMM_WORLD; 88 89 // Parse command line options 90 PetscOptionsBegin(comm, NULL, "CEED BPs in PETSc", NULL); 91 bp_choice = CEED_BP3; 92 ierr = PetscOptionsEnum("-problem", 93 "CEED benchmark problem to solve", NULL, 94 bp_types, (PetscEnum)bp_choice, (PetscEnum *)&bp_choice, 95 NULL); CHKERRQ(ierr); 96 num_comp_u = bp_options[bp_choice].num_comp_u; 97 test_mode = PETSC_FALSE; 98 ierr = PetscOptionsBool("-test", 99 "Testing mode (do not print unless error is large)", 100 NULL, test_mode, &test_mode, NULL); CHKERRQ(ierr); 101 benchmark_mode = PETSC_FALSE; 102 ierr = PetscOptionsBool("-benchmark", 103 "Benchmarking mode (prints benchmark statistics)", 104 NULL, benchmark_mode, &benchmark_mode, NULL); 105 CHKERRQ(ierr); 106 write_solution = PETSC_FALSE; 107 ierr = PetscOptionsBool("-write_solution", 108 "Write solution for visualization", 109 NULL, write_solution, &write_solution, NULL); 110 CHKERRQ(ierr); 111 ierr = PetscOptionsScalar("-eps", 112 "Epsilon parameter for Kershaw mesh transformation", 113 NULL, eps, &eps, NULL); 114 if (eps > 1 || eps <= 0) SETERRQ(PETSC_COMM_WORLD, PETSC_ERR_ARG_OUTOFRANGE, 115 "-eps %g must be (0,1]", (double)PetscRealPart(eps)); 116 degree = test_mode ? 3 : 2; 117 ierr = PetscOptionsInt("-degree", "Polynomial degree of tensor product basis", 118 NULL, degree, °ree, NULL); CHKERRQ(ierr); 119 if (degree < 1) SETERRQ(PETSC_COMM_WORLD, PETSC_ERR_ARG_OUTOFRANGE, 120 "-degree %" PetscInt_FMT " must be at least 1", degree); 121 q_extra = bp_options[bp_choice].q_extra; 122 ierr = PetscOptionsInt("-q_extra", "Number of extra quadrature points", 123 NULL, q_extra, &q_extra, NULL); CHKERRQ(ierr); 124 ierr = PetscOptionsString("-ceed", "CEED resource specifier", 125 NULL, ceed_resource, ceed_resource, 126 sizeof(ceed_resource), NULL); CHKERRQ(ierr); 127 coarsen = COARSEN_UNIFORM; 128 ierr = PetscOptionsEnum("-coarsen", 129 "Coarsening strategy to use", NULL, 130 coarsen_types, (PetscEnum)coarsen, 131 (PetscEnum *)&coarsen, NULL); CHKERRQ(ierr); 132 read_mesh = PETSC_FALSE; 133 ierr = PetscOptionsString("-mesh", "Read mesh from file", NULL, 134 filename, filename, sizeof(filename), &read_mesh); 135 CHKERRQ(ierr); 136 if (!read_mesh) { 137 PetscInt tmp = dim; 138 ierr = PetscOptionsIntArray("-cells","Number of cells per dimension", NULL, 139 mesh_elem, &tmp, NULL); CHKERRQ(ierr); 140 } 141 PetscOptionsEnd(); 142 143 // Set up libCEED 144 CeedInit(ceed_resource, &ceed); 145 CeedMemType mem_type_backend; 146 CeedGetPreferredMemType(ceed, &mem_type_backend); 147 148 // Setup DM 149 if (read_mesh) { 150 ierr = DMPlexCreateFromFile(PETSC_COMM_WORLD, filename, NULL, PETSC_TRUE, 151 &dm_orig); 152 CHKERRQ(ierr); 153 } else { 154 ierr = DMPlexCreateBoxMesh(PETSC_COMM_WORLD, dim, PETSC_FALSE, mesh_elem, NULL, 155 NULL, NULL, PETSC_TRUE, &dm_orig); CHKERRQ(ierr); 156 } 157 158 VecType vec_type; 159 switch (mem_type_backend) { 160 case CEED_MEM_HOST: vec_type = VECSTANDARD; break; 161 case CEED_MEM_DEVICE: { 162 const char *resolved; 163 CeedGetResource(ceed, &resolved); 164 if (strstr(resolved, "/gpu/cuda")) vec_type = VECCUDA; 165 else if (strstr(resolved, "/gpu/hip/occa")) 166 vec_type = VECSTANDARD; // https://github.com/CEED/libCEED/issues/678 167 else if (strstr(resolved, "/gpu/hip")) vec_type = VECHIP; 168 else vec_type = VECSTANDARD; 169 } 170 } 171 ierr = DMSetVecType(dm_orig, vec_type); CHKERRQ(ierr); 172 ierr = DMSetFromOptions(dm_orig); CHKERRQ(ierr); 173 ierr = DMViewFromOptions(dm_orig, NULL, "-dm_view"); CHKERRQ(ierr); 174 175 // Apply Kershaw mesh transformation 176 ierr = Kershaw(dm_orig, eps); CHKERRQ(ierr); 177 178 // Allocate arrays for PETSc objects for each level 179 switch (coarsen) { 180 case COARSEN_UNIFORM: 181 num_levels = degree; 182 break; 183 case COARSEN_LOGARITHMIC: 184 num_levels = ceil(log(degree)/log(2)) + 1; 185 break; 186 } 187 ierr = PetscMalloc1(num_levels, &level_degrees); CHKERRQ(ierr); 188 fine_level = num_levels - 1; 189 190 switch (coarsen) { 191 case COARSEN_UNIFORM: 192 for (int i=0; i<num_levels; i++) level_degrees[i] = i + 1; 193 break; 194 case COARSEN_LOGARITHMIC: 195 for (int i=0; i<num_levels - 1; i++) level_degrees[i] = pow(2,i); 196 level_degrees[fine_level] = degree; 197 break; 198 } 199 ierr = PetscMalloc1(num_levels, &dm); CHKERRQ(ierr); 200 ierr = PetscMalloc1(num_levels, &X); CHKERRQ(ierr); 201 ierr = PetscMalloc1(num_levels, &X_loc); CHKERRQ(ierr); 202 ierr = PetscMalloc1(num_levels, &mult); CHKERRQ(ierr); 203 ierr = PetscMalloc1(num_levels, &user_O); CHKERRQ(ierr); 204 ierr = PetscMalloc1(num_levels, &user_pr); CHKERRQ(ierr); 205 ierr = PetscMalloc1(num_levels, &mat_O); CHKERRQ(ierr); 206 ierr = PetscMalloc1(num_levels, &mat_pr); CHKERRQ(ierr); 207 ierr = PetscMalloc1(num_levels, &l_size); CHKERRQ(ierr); 208 ierr = PetscMalloc1(num_levels, &xl_size); CHKERRQ(ierr); 209 ierr = PetscMalloc1(num_levels, &g_size); CHKERRQ(ierr); 210 211 // Setup DM and Operator Mat Shells for each level 212 for (CeedInt i=0; i<num_levels; i++) { 213 // Create DM 214 ierr = DMClone(dm_orig, &dm[i]); CHKERRQ(ierr); 215 ierr = DMGetVecType(dm_orig, &vec_type); CHKERRQ(ierr); 216 ierr = DMSetVecType(dm[i], vec_type); CHKERRQ(ierr); 217 PetscInt dim; 218 ierr = DMGetDimension(dm[i], &dim); CHKERRQ(ierr); 219 ierr = SetupDMByDegree(dm[i], level_degrees[i], num_comp_u, dim, 220 bp_options[bp_choice].enforce_bc, bp_options[bp_choice].bc_func); 221 CHKERRQ(ierr); 222 223 // Create vectors 224 ierr = DMCreateGlobalVector(dm[i], &X[i]); CHKERRQ(ierr); 225 ierr = VecGetLocalSize(X[i], &l_size[i]); CHKERRQ(ierr); 226 ierr = VecGetSize(X[i], &g_size[i]); CHKERRQ(ierr); 227 ierr = DMCreateLocalVector(dm[i], &X_loc[i]); CHKERRQ(ierr); 228 ierr = VecGetSize(X_loc[i], &xl_size[i]); CHKERRQ(ierr); 229 230 // Operator 231 ierr = PetscMalloc1(1, &user_O[i]); CHKERRQ(ierr); 232 ierr = MatCreateShell(comm, l_size[i], l_size[i], g_size[i], g_size[i], 233 user_O[i], &mat_O[i]); CHKERRQ(ierr); 234 ierr = MatShellSetOperation(mat_O[i], MATOP_MULT, 235 (void(*)(void))MatMult_Ceed); CHKERRQ(ierr); 236 ierr = MatShellSetOperation(mat_O[i], MATOP_GET_DIAGONAL, 237 (void(*)(void))MatGetDiag); CHKERRQ(ierr); 238 ierr = MatShellSetVecType(mat_O[i], vec_type); CHKERRQ(ierr); 239 240 // Level transfers 241 if (i > 0) { 242 // Interp 243 ierr = PetscMalloc1(1, &user_pr[i]); CHKERRQ(ierr); 244 ierr = MatCreateShell(comm, l_size[i], l_size[i-1], g_size[i], g_size[i-1], 245 user_pr[i], &mat_pr[i]); CHKERRQ(ierr); 246 ierr = MatShellSetOperation(mat_pr[i], MATOP_MULT, 247 (void(*)(void))MatMult_Prolong); 248 CHKERRQ(ierr); 249 ierr = MatShellSetOperation(mat_pr[i], MATOP_MULT_TRANSPOSE, 250 (void(*)(void))MatMult_Restrict); 251 CHKERRQ(ierr); 252 ierr = MatShellSetVecType(mat_pr[i], vec_type); CHKERRQ(ierr); 253 } 254 } 255 ierr = VecDuplicate(X[fine_level], &rhs); CHKERRQ(ierr); 256 257 // Print global grid information 258 if (!test_mode) { 259 PetscInt P = degree + 1, Q = P + q_extra; 260 261 const char *used_resource; 262 CeedGetResource(ceed, &used_resource); 263 264 ierr = VecGetType(X[0], &vec_type); CHKERRQ(ierr); 265 266 ierr = PetscPrintf(comm, 267 "\n-- CEED Benchmark Problem %" CeedInt_FMT " -- libCEED + PETSc + PCMG --\n" 268 " PETSc:\n" 269 " PETSc Vec Type : %s\n" 270 " libCEED:\n" 271 " libCEED Backend : %s\n" 272 " libCEED Backend MemType : %s\n" 273 " Mesh:\n" 274 " Number of 1D Basis Nodes (p) : %" CeedInt_FMT "\n" 275 " Number of 1D Quadrature Points (q) : %" CeedInt_FMT "\n" 276 " Global Nodes : %" PetscInt_FMT "\n" 277 " Owned Nodes : %" PetscInt_FMT "\n" 278 " DoF per node : %" PetscInt_FMT "\n" 279 " Multigrid:\n" 280 " Number of Levels : %" CeedInt_FMT "\n", 281 bp_choice+1, vec_type, used_resource, 282 CeedMemTypes[mem_type_backend], 283 P, Q, g_size[fine_level]/num_comp_u, l_size[fine_level]/num_comp_u, 284 num_comp_u, num_levels); CHKERRQ(ierr); 285 } 286 287 // Create RHS vector 288 ierr = VecDuplicate(X_loc[fine_level], &rhs_loc); CHKERRQ(ierr); 289 ierr = VecZeroEntries(rhs_loc); CHKERRQ(ierr); 290 ierr = VecGetArrayAndMemType(rhs_loc, &r, &mem_type); CHKERRQ(ierr); 291 CeedVectorCreate(ceed, xl_size[fine_level], &rhs_ceed); 292 CeedVectorSetArray(rhs_ceed, MemTypeP2C(mem_type), CEED_USE_POINTER, r); 293 294 // Set up libCEED operators on each level 295 ierr = PetscMalloc1(num_levels, &ceed_data); CHKERRQ(ierr); 296 for (PetscInt i=0; i<num_levels; i++) { 297 // Print level information 298 if (!test_mode && (i == 0 || i == fine_level)) { 299 ierr = PetscPrintf(comm," Level %" PetscInt_FMT " (%s):\n" 300 " Number of 1D Basis Nodes (p) : %" CeedInt_FMT "\n" 301 " Global Nodes : %" PetscInt_FMT "\n" 302 " Owned Nodes : %" PetscInt_FMT "\n", 303 i, (i? "fine" : "coarse"), level_degrees[i] + 1, 304 g_size[i]/num_comp_u, l_size[i]/num_comp_u); CHKERRQ(ierr); 305 } 306 ierr = PetscMalloc1(1, &ceed_data[i]); CHKERRQ(ierr); 307 ierr = SetupLibceedByDegree(dm[i], ceed, level_degrees[i], dim, q_extra, 308 dim, num_comp_u, g_size[i], xl_size[i], bp_options[bp_choice], 309 ceed_data[i], i==(fine_level), rhs_ceed, &target); 310 CHKERRQ(ierr); 311 } 312 313 // Gather RHS 314 CeedVectorTakeArray(rhs_ceed, MemTypeP2C(mem_type), NULL); 315 ierr = VecRestoreArrayAndMemType(rhs_loc, &r); CHKERRQ(ierr); 316 ierr = VecZeroEntries(rhs); CHKERRQ(ierr); 317 ierr = DMLocalToGlobal(dm[fine_level], rhs_loc, ADD_VALUES, rhs); CHKERRQ(ierr); 318 CeedVectorDestroy(&rhs_ceed); 319 320 // Create the restriction/interpolation QFunction 321 CeedQFunctionCreateIdentity(ceed, num_comp_u, CEED_EVAL_NONE, CEED_EVAL_INTERP, 322 &qf_restrict); 323 CeedQFunctionCreateIdentity(ceed, num_comp_u, CEED_EVAL_INTERP, CEED_EVAL_NONE, 324 &qf_prolong); 325 326 // Set up libCEED level transfer operators 327 ierr = CeedLevelTransferSetup(ceed, num_levels, num_comp_u, ceed_data, 328 level_degrees, 329 qf_restrict, qf_prolong); CHKERRQ(ierr); 330 331 // Create the error QFunction 332 CeedQFunctionCreateInterior(ceed, 1, bp_options[bp_choice].error, 333 bp_options[bp_choice].error_loc, &qf_error); 334 CeedQFunctionAddInput(qf_error, "u", num_comp_u, CEED_EVAL_INTERP); 335 CeedQFunctionAddInput(qf_error, "true_soln", num_comp_u, CEED_EVAL_NONE); 336 CeedQFunctionAddOutput(qf_error, "error", num_comp_u, CEED_EVAL_NONE); 337 338 // Create the error operator 339 CeedOperatorCreate(ceed, qf_error, CEED_QFUNCTION_NONE, CEED_QFUNCTION_NONE, 340 &op_error); 341 CeedOperatorSetField(op_error, "u", ceed_data[fine_level]->elem_restr_u, 342 ceed_data[fine_level]->basis_u, CEED_VECTOR_ACTIVE); 343 CeedOperatorSetField(op_error, "true_soln", 344 ceed_data[fine_level]->elem_restr_u_i, 345 CEED_BASIS_COLLOCATED, target); 346 CeedOperatorSetField(op_error, "error", ceed_data[fine_level]->elem_restr_u_i, 347 CEED_BASIS_COLLOCATED, CEED_VECTOR_ACTIVE); 348 349 // Calculate multiplicity 350 for (int i=0; i<num_levels; i++) { 351 PetscScalar *x; 352 353 // CEED vector 354 ierr = VecZeroEntries(X_loc[i]); CHKERRQ(ierr); 355 ierr = VecGetArray(X_loc[i], &x); CHKERRQ(ierr); 356 CeedVectorSetArray(ceed_data[i]->x_ceed, CEED_MEM_HOST, CEED_USE_POINTER, x); 357 358 // Multiplicity 359 CeedElemRestrictionGetMultiplicity(ceed_data[i]->elem_restr_u, 360 ceed_data[i]->x_ceed); 361 CeedVectorSyncArray(ceed_data[i]->x_ceed, CEED_MEM_HOST); 362 363 // Restore vector 364 ierr = VecRestoreArray(X_loc[i], &x); CHKERRQ(ierr); 365 366 // Creat mult vector 367 ierr = VecDuplicate(X_loc[i], &mult[i]); CHKERRQ(ierr); 368 369 // Local-to-global 370 ierr = VecZeroEntries(X[i]); CHKERRQ(ierr); 371 ierr = DMLocalToGlobal(dm[i], X_loc[i], ADD_VALUES, X[i]); 372 CHKERRQ(ierr); 373 ierr = VecZeroEntries(X_loc[i]); CHKERRQ(ierr); 374 375 // Global-to-local 376 ierr = DMGlobalToLocal(dm[i], X[i], INSERT_VALUES, mult[i]); 377 CHKERRQ(ierr); 378 ierr = VecZeroEntries(X[i]); CHKERRQ(ierr); 379 380 // Multiplicity scaling 381 ierr = VecReciprocal(mult[i]); 382 } 383 384 // Set up Mat 385 for (int i=0; i<num_levels; i++) { 386 // User Operator 387 user_O[i]->comm = comm; 388 user_O[i]->dm = dm[i]; 389 user_O[i]->X_loc = X_loc[i]; 390 ierr = VecDuplicate(X_loc[i], &user_O[i]->Y_loc); CHKERRQ(ierr); 391 user_O[i]->x_ceed = ceed_data[i]->x_ceed; 392 user_O[i]->y_ceed = ceed_data[i]->y_ceed; 393 user_O[i]->op = ceed_data[i]->op_apply; 394 user_O[i]->ceed = ceed; 395 396 if (i > 0) { 397 // Prolongation/Restriction Operator 398 user_pr[i]->comm = comm; 399 user_pr[i]->dmf = dm[i]; 400 user_pr[i]->dmc = dm[i-1]; 401 user_pr[i]->loc_vec_c = X_loc[i-1]; 402 user_pr[i]->loc_vec_f = user_O[i]->Y_loc; 403 user_pr[i]->mult_vec = mult[i]; 404 user_pr[i]->ceed_vec_c = user_O[i-1]->x_ceed; 405 user_pr[i]->ceed_vec_f = user_O[i]->y_ceed; 406 user_pr[i]->op_prolong = ceed_data[i]->op_prolong; 407 user_pr[i]->op_restrict = ceed_data[i]->op_restrict; 408 user_pr[i]->ceed = ceed; 409 } 410 } 411 412 // Assemble coarse grid Jacobian for AMG (or other sparse matrix) solve 413 ierr = DMCreateMatrix(dm[0], &mat_coarse); CHKERRQ(ierr); 414 415 ierr = PetscLogEventRegister("AssembleMatrix", MAT_CLASSID, &assemble_event); 416 CHKERRQ(ierr); 417 { 418 // Assemble matrix analytically 419 PetscCount num_entries; 420 CeedInt *rows, *cols; 421 CeedVector coo_values; 422 CeedOperatorLinearAssembleSymbolic(user_O[0]->op, &num_entries, &rows, &cols); 423 ISLocalToGlobalMapping ltog_row, ltog_col; 424 ierr = MatGetLocalToGlobalMapping(mat_coarse, <og_row, <og_col); 425 CHKERRQ(ierr); 426 ierr = ISLocalToGlobalMappingApply(ltog_row, num_entries, rows, rows); 427 CHKERRQ(ierr); 428 ierr = ISLocalToGlobalMappingApply(ltog_col, num_entries, cols, cols); 429 CHKERRQ(ierr); 430 ierr = MatSetPreallocationCOO(mat_coarse, num_entries, rows, cols); 431 CHKERRQ(ierr); 432 free(rows); 433 free(cols); 434 CeedVectorCreate(ceed, num_entries, &coo_values); 435 ierr = PetscLogEventBegin(assemble_event, mat_coarse, 0, 0, 0); CHKERRQ(ierr); 436 CeedOperatorLinearAssemble(user_O[0]->op, coo_values); 437 const CeedScalar *values; 438 CeedVectorGetArrayRead(coo_values, CEED_MEM_HOST, &values); 439 ierr = MatSetValuesCOO(mat_coarse, values, ADD_VALUES); CHKERRQ(ierr); 440 CeedVectorRestoreArrayRead(coo_values, &values); 441 ierr = PetscLogEventEnd(assemble_event, mat_coarse, 0, 0, 0); CHKERRQ(ierr); 442 CeedVectorDestroy(&coo_values); 443 } 444 445 // Set up KSP 446 ierr = KSPCreate(comm, &ksp); CHKERRQ(ierr); 447 { 448 ierr = KSPSetType(ksp, KSPCG); CHKERRQ(ierr); 449 ierr = KSPSetNormType(ksp, KSP_NORM_NATURAL); CHKERRQ(ierr); 450 ierr = KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, 451 PETSC_DEFAULT); CHKERRQ(ierr); 452 } 453 ierr = KSPSetFromOptions(ksp); CHKERRQ(ierr); 454 ierr = KSPSetOperators(ksp, mat_O[fine_level], mat_O[fine_level]); 455 CHKERRQ(ierr); 456 457 // Set up PCMG 458 ierr = KSPGetPC(ksp, &pc); CHKERRQ(ierr); 459 PCMGCycleType pcmg_cycle_type = PC_MG_CYCLE_V; 460 { 461 ierr = PCSetType(pc, PCMG); CHKERRQ(ierr); 462 463 // PCMG levels 464 ierr = PCMGSetLevels(pc, num_levels, NULL); CHKERRQ(ierr); 465 for (int i=0; i<num_levels; i++) { 466 // Smoother 467 KSP smoother; 468 PC smoother_pc; 469 ierr = PCMGGetSmoother(pc, i, &smoother); CHKERRQ(ierr); 470 ierr = KSPSetType(smoother, KSPCHEBYSHEV); CHKERRQ(ierr); 471 ierr = KSPChebyshevEstEigSet(smoother, 0, 0.1, 0, 1.1); CHKERRQ(ierr); 472 ierr = KSPChebyshevEstEigSetUseNoisy(smoother, PETSC_TRUE); CHKERRQ(ierr); 473 ierr = KSPSetOperators(smoother, mat_O[i], mat_O[i]); CHKERRQ(ierr); 474 ierr = KSPGetPC(smoother, &smoother_pc); CHKERRQ(ierr); 475 ierr = PCSetType(smoother_pc, PCJACOBI); CHKERRQ(ierr); 476 ierr = PCJacobiSetType(smoother_pc, PC_JACOBI_DIAGONAL); CHKERRQ(ierr); 477 478 // Work vector 479 if (i < num_levels - 1) { 480 ierr = PCMGSetX(pc, i, X[i]); CHKERRQ(ierr); 481 } 482 483 // Level transfers 484 if (i > 0) { 485 // Interpolation 486 ierr = PCMGSetInterpolation(pc, i, mat_pr[i]); CHKERRQ(ierr); 487 } 488 489 // Coarse solve 490 KSP coarse; 491 PC coarse_pc; 492 ierr = PCMGGetCoarseSolve(pc, &coarse); CHKERRQ(ierr); 493 ierr = KSPSetType(coarse, KSPPREONLY); CHKERRQ(ierr); 494 ierr = KSPSetOperators(coarse, mat_coarse, mat_coarse); CHKERRQ(ierr); 495 496 ierr = KSPGetPC(coarse, &coarse_pc); CHKERRQ(ierr); 497 ierr = PCSetType(coarse_pc, PCGAMG); CHKERRQ(ierr); 498 499 ierr = KSPSetOptionsPrefix(coarse, "coarse_"); CHKERRQ(ierr); 500 ierr = PCSetOptionsPrefix(coarse_pc, "coarse_"); CHKERRQ(ierr); 501 ierr = KSPSetFromOptions(coarse); CHKERRQ(ierr); 502 ierr = PCSetFromOptions(coarse_pc); CHKERRQ(ierr); 503 } 504 505 // PCMG options 506 ierr = PCMGSetType(pc, PC_MG_MULTIPLICATIVE); CHKERRQ(ierr); 507 ierr = PCMGSetNumberSmooth(pc, 3); CHKERRQ(ierr); 508 ierr = PCMGSetCycleType(pc, pcmg_cycle_type); CHKERRQ(ierr); 509 } 510 511 // First run, if benchmarking 512 if (benchmark_mode) { 513 ierr = KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, 1); 514 CHKERRQ(ierr); 515 ierr = VecZeroEntries(X[fine_level]); CHKERRQ(ierr); 516 my_rt_start = MPI_Wtime(); 517 ierr = KSPSolve(ksp, rhs, X[fine_level]); CHKERRQ(ierr); 518 my_rt = MPI_Wtime() - my_rt_start; 519 ierr = MPI_Allreduce(MPI_IN_PLACE, &my_rt, 1, MPI_DOUBLE, MPI_MIN, comm); 520 CHKERRQ(ierr); 521 // Set maxits based on first iteration timing 522 if (my_rt > 0.02) { 523 ierr = KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, 5); 524 CHKERRQ(ierr); 525 } else { 526 ierr = KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, 20); 527 CHKERRQ(ierr); 528 } 529 } 530 531 // Timed solve 532 ierr = VecZeroEntries(X[fine_level]); CHKERRQ(ierr); 533 ierr = PetscBarrier((PetscObject)ksp); CHKERRQ(ierr); 534 535 // -- Performance logging 536 ierr = PetscLogStageRegister("Solve Stage", &solve_stage); CHKERRQ(ierr); 537 ierr = PetscLogStagePush(solve_stage); CHKERRQ(ierr); 538 539 // -- Solve 540 my_rt_start = MPI_Wtime(); 541 ierr = KSPSolve(ksp, rhs, X[fine_level]); CHKERRQ(ierr); 542 my_rt = MPI_Wtime() - my_rt_start; 543 544 545 // -- Performance logging 546 ierr = PetscLogStagePop(); 547 548 // Output results 549 { 550 KSPType ksp_type; 551 PCMGType pcmg_type; 552 KSPConvergedReason reason; 553 PetscReal rnorm; 554 PetscInt its; 555 ierr = KSPGetType(ksp, &ksp_type); CHKERRQ(ierr); 556 ierr = KSPGetConvergedReason(ksp, &reason); CHKERRQ(ierr); 557 ierr = KSPGetIterationNumber(ksp, &its); CHKERRQ(ierr); 558 ierr = KSPGetResidualNorm(ksp, &rnorm); CHKERRQ(ierr); 559 ierr = PCMGGetType(pc, &pcmg_type); CHKERRQ(ierr); 560 if (!test_mode || reason < 0 || rnorm > 1e-8) { 561 ierr = PetscPrintf(comm, 562 " KSP:\n" 563 " KSP Type : %s\n" 564 " KSP Convergence : %s\n" 565 " Total KSP Iterations : %" PetscInt_FMT "\n" 566 " Final rnorm : %e\n", 567 ksp_type, KSPConvergedReasons[reason], its, 568 (double)rnorm); CHKERRQ(ierr); 569 ierr = PetscPrintf(comm, 570 " PCMG:\n" 571 " PCMG Type : %s\n" 572 " PCMG Cycle Type : %s\n", 573 PCMGTypes[pcmg_type], 574 PCMGCycleTypes[pcmg_cycle_type]); CHKERRQ(ierr); 575 } 576 if (!test_mode) { 577 ierr = PetscPrintf(comm," Performance:\n"); CHKERRQ(ierr); 578 } 579 { 580 PetscReal max_error; 581 ierr = ComputeErrorMax(user_O[fine_level], op_error, X[fine_level], target, 582 &max_error); CHKERRQ(ierr); 583 PetscReal tol = 5e-2; 584 if (!test_mode || max_error > tol) { 585 ierr = MPI_Allreduce(&my_rt, &rt_min, 1, MPI_DOUBLE, MPI_MIN, comm); 586 CHKERRQ(ierr); 587 ierr = MPI_Allreduce(&my_rt, &rt_max, 1, MPI_DOUBLE, MPI_MAX, comm); 588 CHKERRQ(ierr); 589 ierr = PetscPrintf(comm, 590 " Pointwise Error (max) : %e\n" 591 " CG Solve Time : %g (%g) sec\n", 592 (double)max_error, rt_max, rt_min); CHKERRQ(ierr); 593 } 594 } 595 if (benchmark_mode && (!test_mode)) { 596 ierr = PetscPrintf(comm, 597 " DoFs/Sec in CG : %g (%g) million\n", 598 1e-6*g_size[fine_level]*its/rt_max, 599 1e-6*g_size[fine_level]*its/rt_min); 600 CHKERRQ(ierr); 601 } 602 } 603 604 if (write_solution) { 605 PetscViewer vtk_viewer_soln; 606 607 ierr = PetscViewerCreate(comm, &vtk_viewer_soln); CHKERRQ(ierr); 608 ierr = PetscViewerSetType(vtk_viewer_soln, PETSCVIEWERVTK); CHKERRQ(ierr); 609 ierr = PetscViewerFileSetName(vtk_viewer_soln, "solution.vtu"); CHKERRQ(ierr); 610 ierr = VecView(X[fine_level], vtk_viewer_soln); CHKERRQ(ierr); 611 ierr = PetscViewerDestroy(&vtk_viewer_soln); CHKERRQ(ierr); 612 } 613 614 // Cleanup 615 for (int i=0; i<num_levels; i++) { 616 ierr = VecDestroy(&X[i]); CHKERRQ(ierr); 617 ierr = VecDestroy(&X_loc[i]); CHKERRQ(ierr); 618 ierr = VecDestroy(&mult[i]); CHKERRQ(ierr); 619 ierr = VecDestroy(&user_O[i]->Y_loc); CHKERRQ(ierr); 620 ierr = MatDestroy(&mat_O[i]); CHKERRQ(ierr); 621 ierr = PetscFree(user_O[i]); CHKERRQ(ierr); 622 if (i > 0) { 623 ierr = MatDestroy(&mat_pr[i]); CHKERRQ(ierr); 624 ierr = PetscFree(user_pr[i]); CHKERRQ(ierr); 625 } 626 ierr = CeedDataDestroy(i, ceed_data[i]); CHKERRQ(ierr); 627 ierr = DMDestroy(&dm[i]); CHKERRQ(ierr); 628 } 629 ierr = PetscFree(level_degrees); CHKERRQ(ierr); 630 ierr = PetscFree(dm); CHKERRQ(ierr); 631 ierr = PetscFree(X); CHKERRQ(ierr); 632 ierr = PetscFree(X_loc); CHKERRQ(ierr); 633 ierr = PetscFree(mult); CHKERRQ(ierr); 634 ierr = PetscFree(mat_O); CHKERRQ(ierr); 635 ierr = PetscFree(mat_pr); CHKERRQ(ierr); 636 ierr = PetscFree(ceed_data); CHKERRQ(ierr); 637 ierr = PetscFree(user_O); CHKERRQ(ierr); 638 ierr = PetscFree(user_pr); CHKERRQ(ierr); 639 ierr = PetscFree(l_size); CHKERRQ(ierr); 640 ierr = PetscFree(xl_size); CHKERRQ(ierr); 641 ierr = PetscFree(g_size); CHKERRQ(ierr); 642 ierr = VecDestroy(&rhs); CHKERRQ(ierr); 643 ierr = VecDestroy(&rhs_loc); CHKERRQ(ierr); 644 ierr = MatDestroy(&mat_coarse); CHKERRQ(ierr); 645 ierr = KSPDestroy(&ksp); CHKERRQ(ierr); 646 ierr = DMDestroy(&dm_orig); CHKERRQ(ierr); 647 CeedVectorDestroy(&target); 648 CeedQFunctionDestroy(&qf_error); 649 CeedQFunctionDestroy(&qf_restrict); 650 CeedQFunctionDestroy(&qf_prolong); 651 CeedOperatorDestroy(&op_error); 652 CeedDestroy(&ceed); 653 return PetscFinalize(); 654 } 655