1 // Copyright (c) 2017-2024, 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 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 bps [PETSC_DIR=</path/to/petsc>] [CEED_DIR=</path/to/libceed>] 17 // 18 // Sample runs: 19 // 20 // ./bps -problem bp1 -degree 3 21 // ./bps -problem bp2 -degree 3 22 // ./bps -problem bp3 -degree 3 23 // ./bps -problem bp4 -degree 3 24 // ./bps -problem bp5 -degree 3 -ceed /cpu/self 25 // ./bps -problem bp6 -degree 3 -ceed /gpu/cuda 26 // 27 //TESTARGS(name="BP3, tet elements") -ceed {ceed_resource} -test -problem bp3 -degree 3 -ksp_max_it_clip 50,50 -simplex 28 //TESTARGS(name="BP5, hex elements") -ceed {ceed_resource} -test -problem bp5 -degree 3 -ksp_max_it_clip 18,18 29 30 /// @file 31 /// CEED BPs example using PETSc with DMPlex 32 /// See bpsraw.c for a "raw" implementation using a structured grid. 33 const char help[] = "Solve CEED BPs using PETSc with DMPlex\n"; 34 35 #include "bps.h" 36 37 #include <ceed.h> 38 #include <petscdmplex.h> 39 #include <petscksp.h> 40 #include <stdbool.h> 41 #include <string.h> 42 43 #include "include/bpsproblemdata.h" 44 #include "include/libceedsetup.h" 45 #include "include/matops.h" 46 #include "include/petscutils.h" 47 #include "include/petscversion.h" 48 #include "include/structs.h" 49 50 // ----------------------------------------------------------------------------- 51 // Main body of program, called in a loop for performance benchmarking purposes 52 // ----------------------------------------------------------------------------- 53 static PetscErrorCode RunWithDM(RunParams rp, DM dm, const char *ceed_resource) { 54 double my_rt_start, my_rt, rt_min, rt_max; 55 PetscInt xl_size, l_size, g_size; 56 Vec X, X_loc, rhs, rhs_loc; 57 Mat mat_O; 58 KSP ksp; 59 OperatorApplyContext op_apply_ctx, op_error_ctx; 60 Ceed ceed; 61 CeedData ceed_data; 62 CeedQFunction qf_error; 63 CeedOperator op_error; 64 CeedVector rhs_ceed, target; 65 VecType vec_type = VECSTANDARD; 66 PetscMemType mem_type; 67 68 PetscFunctionBeginUser; 69 // Set up libCEED 70 CeedInit(ceed_resource, &ceed); 71 CeedMemType mem_type_backend; 72 CeedGetPreferredMemType(ceed, &mem_type_backend); 73 74 // Set mesh vec_type 75 switch (mem_type_backend) { 76 case CEED_MEM_HOST: 77 vec_type = VECSTANDARD; 78 break; 79 case CEED_MEM_DEVICE: { 80 const char *resolved; 81 82 CeedGetResource(ceed, &resolved); 83 if (strstr(resolved, "/gpu/cuda")) vec_type = VECCUDA; 84 else if (strstr(resolved, "/gpu/hip/occa")) vec_type = VECSTANDARD; // https://github.com/CEED/libCEED/issues/678 85 else if (strstr(resolved, "/gpu/hip")) vec_type = VECHIP; 86 else vec_type = VECSTANDARD; 87 } 88 } 89 PetscCall(DMSetVecType(dm, vec_type)); 90 91 // Create global and local solution vectors 92 PetscCall(DMCreateGlobalVector(dm, &X)); 93 PetscCall(VecGetLocalSize(X, &l_size)); 94 PetscCall(VecGetSize(X, &g_size)); 95 PetscCall(DMCreateLocalVector(dm, &X_loc)); 96 PetscCall(VecGetSize(X_loc, &xl_size)); 97 PetscCall(VecDuplicate(X, &rhs)); 98 99 // Operator 100 PetscCall(PetscMalloc1(1, &op_apply_ctx)); 101 PetscCall(PetscMalloc1(1, &op_error_ctx)); 102 PetscCall(MatCreateShell(rp->comm, l_size, l_size, g_size, g_size, op_apply_ctx, &mat_O)); 103 PetscCall(MatShellSetOperation(mat_O, MATOP_MULT, (void (*)(void))MatMult_Ceed)); 104 PetscCall(MatShellSetOperation(mat_O, MATOP_GET_DIAGONAL, (void (*)(void))MatGetDiag)); 105 PetscCall(MatShellSetVecType(mat_O, vec_type)); 106 107 // Print summary 108 if (!rp->test_mode) { 109 PetscInt P = rp->degree + 1, Q = P + rp->q_extra; 110 111 const char *used_resource; 112 CeedGetResource(ceed, &used_resource); 113 114 VecType vec_type; 115 PetscCall(VecGetType(X, &vec_type)); 116 117 PetscInt c_start, c_end; 118 PetscCall(DMPlexGetHeightStratum(dm, 0, &c_start, &c_end)); 119 DMPolytopeType cell_type; 120 PetscCall(DMPlexGetCellType(dm, c_start, &cell_type)); 121 CeedElemTopology elem_topo = ElemTopologyP2C(cell_type); 122 PetscMPIInt comm_size; 123 PetscCall(MPI_Comm_size(rp->comm, &comm_size)); 124 PetscCall(PetscPrintf(rp->comm, 125 "\n-- CEED Benchmark Problem %" CeedInt_FMT " -- libCEED + PETSc --\n" 126 " MPI:\n" 127 " Hostname : %s\n" 128 " Total ranks : %d\n" 129 " Ranks per compute node : %d\n" 130 " PETSc:\n" 131 " PETSc Vec Type : %s\n" 132 " libCEED:\n" 133 " libCEED Backend : %s\n" 134 " libCEED Backend MemType : %s\n" 135 " Mesh:\n" 136 " Solution Order (P) : %" PetscInt_FMT "\n" 137 " Quadrature Order (Q) : %" PetscInt_FMT "\n" 138 " Additional quadrature points (q_extra) : %" PetscInt_FMT "\n" 139 " Global nodes : %" PetscInt_FMT "\n" 140 " Local Elements : %" PetscInt_FMT "\n" 141 " Element topology : %s\n" 142 " Owned nodes : %" PetscInt_FMT "\n" 143 " DoF per node : %" PetscInt_FMT "\n", 144 rp->bp_choice + 1, rp->hostname, comm_size, rp->ranks_per_node, vec_type, used_resource, CeedMemTypes[mem_type_backend], P, 145 Q, rp->q_extra, g_size / rp->num_comp_u, c_end - c_start, CeedElemTopologies[elem_topo], l_size / rp->num_comp_u, 146 rp->num_comp_u)); 147 } 148 149 // Create RHS vector 150 PetscCall(VecDuplicate(X_loc, &rhs_loc)); 151 PetscCall(VecZeroEntries(rhs_loc)); 152 CeedVectorCreate(ceed, xl_size, &rhs_ceed); 153 PetscCall(VecP2C(rhs_loc, &mem_type, rhs_ceed)); 154 155 PetscCall(PetscMalloc1(1, &ceed_data)); 156 PetscCall(SetupLibceedByDegree(dm, ceed, rp->degree, rp->dim, rp->q_extra, rp->dim, rp->num_comp_u, g_size, xl_size, bp_options[rp->bp_choice], 157 ceed_data, true, true, rhs_ceed, &target)); 158 159 // Gather RHS 160 PetscCall(VecC2P(rhs_ceed, mem_type, rhs_loc)); 161 PetscCall(VecZeroEntries(rhs)); 162 PetscCall(DMLocalToGlobal(dm, rhs_loc, ADD_VALUES, rhs)); 163 CeedVectorDestroy(&rhs_ceed); 164 165 // Create the error QFunction 166 CeedQFunctionCreateInterior(ceed, 1, bp_options[rp->bp_choice].error, bp_options[rp->bp_choice].error_loc, &qf_error); 167 CeedQFunctionAddInput(qf_error, "u", rp->num_comp_u, CEED_EVAL_INTERP); 168 CeedQFunctionAddInput(qf_error, "true_soln", rp->num_comp_u, CEED_EVAL_NONE); 169 CeedQFunctionAddInput(qf_error, "qdata", ceed_data->q_data_size, CEED_EVAL_NONE); 170 CeedQFunctionAddOutput(qf_error, "error", rp->num_comp_u, CEED_EVAL_INTERP); 171 172 // Create the error operator 173 CeedOperatorCreate(ceed, qf_error, CEED_QFUNCTION_NONE, CEED_QFUNCTION_NONE, &op_error); 174 CeedOperatorSetField(op_error, "u", ceed_data->elem_restr_u, ceed_data->basis_u, CEED_VECTOR_ACTIVE); 175 CeedOperatorSetField(op_error, "true_soln", ceed_data->elem_restr_u_i, CEED_BASIS_NONE, target); 176 CeedOperatorSetField(op_error, "qdata", ceed_data->elem_restr_qd_i, CEED_BASIS_NONE, ceed_data->q_data); 177 CeedOperatorSetField(op_error, "error", ceed_data->elem_restr_u, ceed_data->basis_u, CEED_VECTOR_ACTIVE); 178 179 // Set up apply operator context 180 PetscCall(SetupApplyOperatorCtx(rp->comm, dm, ceed, ceed_data, X_loc, op_apply_ctx)); 181 PetscCall(KSPCreate(rp->comm, &ksp)); 182 { 183 PC pc; 184 PetscCall(KSPGetPC(ksp, &pc)); 185 if (rp->bp_choice == CEED_BP1 || rp->bp_choice == CEED_BP2) { 186 PetscCall(PCSetType(pc, PCJACOBI)); 187 if (rp->simplex) { 188 PetscCall(PCJacobiSetType(pc, PC_JACOBI_DIAGONAL)); 189 } else { 190 PetscCall(PCJacobiSetType(pc, PC_JACOBI_ROWSUM)); 191 } 192 } else { 193 PetscCall(PCSetType(pc, PCNONE)); 194 } 195 PetscCall(KSPSetType(ksp, KSPCG)); 196 PetscCall(KSPSetNormType(ksp, KSP_NORM_NATURAL)); 197 PetscCall(KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, PETSC_DEFAULT)); 198 } 199 PetscCall(KSPSetOperators(ksp, mat_O, mat_O)); 200 201 // First run's performance log is not considered for benchmarking purposes 202 PetscCall(KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, 1)); 203 my_rt_start = MPI_Wtime(); 204 PetscCall(KSPSolve(ksp, rhs, X)); 205 my_rt = MPI_Wtime() - my_rt_start; 206 PetscCall(MPI_Allreduce(MPI_IN_PLACE, &my_rt, 1, MPI_DOUBLE, MPI_MIN, rp->comm)); 207 // Set maxits based on first iteration timing 208 if (my_rt > 0.02) { 209 PetscCall(KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, rp->ksp_max_it_clip[0])); 210 } else { 211 PetscCall(KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, rp->ksp_max_it_clip[1])); 212 } 213 PetscCall(KSPSetFromOptions(ksp)); 214 215 // Timed solve 216 PetscCall(VecZeroEntries(X)); 217 PetscCall(PetscBarrier((PetscObject)ksp)); 218 219 // -- Performance logging 220 PetscCall(PetscLogStagePush(rp->solve_stage)); 221 222 // -- Solve 223 my_rt_start = MPI_Wtime(); 224 PetscCall(KSPSolve(ksp, rhs, X)); 225 my_rt = MPI_Wtime() - my_rt_start; 226 227 // -- Performance logging 228 PetscCall(PetscLogStagePop()); 229 230 // Output results 231 { 232 KSPType ksp_type; 233 KSPConvergedReason reason; 234 PetscReal rnorm; 235 PetscInt its; 236 PetscCall(KSPGetType(ksp, &ksp_type)); 237 PetscCall(KSPGetConvergedReason(ksp, &reason)); 238 PetscCall(KSPGetIterationNumber(ksp, &its)); 239 PetscCall(KSPGetResidualNorm(ksp, &rnorm)); 240 if (!rp->test_mode || reason < 0 || rnorm > 1e-8) { 241 PetscCall(PetscPrintf(rp->comm, 242 " KSP:\n" 243 " KSP Type : %s\n" 244 " KSP Convergence : %s\n" 245 " Total KSP Iterations : %" PetscInt_FMT "\n" 246 " Final rnorm : %e\n", 247 ksp_type, KSPConvergedReasons[reason], its, (double)rnorm)); 248 } 249 if (!rp->test_mode) { 250 PetscCall(PetscPrintf(rp->comm, " Performance:\n")); 251 } 252 { 253 // Set up error operator context 254 PetscCall(SetupErrorOperatorCtx(rp->comm, dm, ceed, ceed_data, X_loc, op_error, op_error_ctx)); 255 PetscScalar l2_error; 256 PetscCall(ComputeL2Error(X, &l2_error, op_error_ctx)); 257 PetscReal tol = 5e-2; 258 if (!rp->test_mode || l2_error > tol) { 259 PetscCall(MPI_Allreduce(&my_rt, &rt_min, 1, MPI_DOUBLE, MPI_MIN, rp->comm)); 260 PetscCall(MPI_Allreduce(&my_rt, &rt_max, 1, MPI_DOUBLE, MPI_MAX, rp->comm)); 261 PetscCall(PetscPrintf(rp->comm, 262 " L2 Error : %e\n" 263 " CG Solve Time : %g (%g) sec\n", 264 (double)l2_error, rt_max, rt_min)); 265 } 266 } 267 if (!rp->test_mode) { 268 PetscCall(PetscPrintf(rp->comm, " DoFs/Sec in CG : %g (%g) million\n", 1e-6 * g_size * its / rt_max, 269 1e-6 * g_size * its / rt_min)); 270 } 271 } 272 273 if (rp->write_solution) { 274 PetscViewer vtk_viewer_soln; 275 276 PetscCall(PetscViewerCreate(rp->comm, &vtk_viewer_soln)); 277 PetscCall(PetscViewerSetType(vtk_viewer_soln, PETSCVIEWERVTK)); 278 PetscCall(PetscViewerFileSetName(vtk_viewer_soln, "solution.vtu")); 279 PetscCall(VecView(X, vtk_viewer_soln)); 280 PetscCall(PetscViewerDestroy(&vtk_viewer_soln)); 281 } 282 283 // Cleanup 284 PetscCall(VecDestroy(&X)); 285 PetscCall(VecDestroy(&X_loc)); 286 PetscCall(VecDestroy(&op_apply_ctx->Y_loc)); 287 PetscCall(VecDestroy(&op_error_ctx->Y_loc)); 288 PetscCall(MatDestroy(&mat_O)); 289 PetscCall(PetscFree(op_apply_ctx)); 290 PetscCall(PetscFree(op_error_ctx)); 291 PetscCall(CeedDataDestroy(0, ceed_data)); 292 293 PetscCall(VecDestroy(&rhs)); 294 PetscCall(VecDestroy(&rhs_loc)); 295 PetscCall(KSPDestroy(&ksp)); 296 CeedVectorDestroy(&target); 297 CeedQFunctionDestroy(&qf_error); 298 CeedOperatorDestroy(&op_error); 299 CeedDestroy(&ceed); 300 PetscFunctionReturn(PETSC_SUCCESS); 301 } 302 303 static PetscErrorCode Run(RunParams rp, PetscInt num_resources, char *const *ceed_resources, PetscInt num_bp_choices, const BPType *bp_choices) { 304 DM dm; 305 306 PetscFunctionBeginUser; 307 // Setup DM 308 PetscCall(CreateDistributedDM(rp, &dm)); 309 310 for (PetscInt b = 0; b < num_bp_choices; b++) { 311 DM dm_deg; 312 VecType vec_type; 313 PetscInt q_extra = rp->q_extra; 314 rp->bp_choice = bp_choices[b]; 315 rp->num_comp_u = bp_options[rp->bp_choice].num_comp_u; 316 rp->q_extra = q_extra < 0 ? bp_options[rp->bp_choice].q_extra : q_extra; 317 PetscCall(DMClone(dm, &dm_deg)); 318 PetscCall(DMGetVecType(dm, &vec_type)); 319 PetscCall(DMSetVecType(dm_deg, vec_type)); 320 // Create DM 321 PetscInt dim; 322 PetscCall(DMGetDimension(dm_deg, &dim)); 323 PetscCall(SetupDMByDegree(dm_deg, rp->degree, rp->q_extra, rp->num_comp_u, dim, bp_options[rp->bp_choice].enforce_bc)); 324 for (PetscInt r = 0; r < num_resources; r++) { 325 PetscCall(RunWithDM(rp, dm_deg, ceed_resources[r])); 326 } 327 PetscCall(DMDestroy(&dm_deg)); 328 rp->q_extra = q_extra; 329 } 330 331 PetscCall(DMDestroy(&dm)); 332 PetscFunctionReturn(PETSC_SUCCESS); 333 } 334 335 int main(int argc, char **argv) { 336 PetscMPIInt comm_size; 337 RunParams rp; 338 MPI_Comm comm; 339 char filename[PETSC_MAX_PATH_LEN]; 340 char *ceed_resources[30]; 341 PetscInt num_ceed_resources = 30; 342 char hostname[PETSC_MAX_PATH_LEN]; 343 344 PetscInt dim = 3, mesh_elem[3] = {3, 3, 3}; 345 PetscInt num_degrees = 30, degree[30] = {0}, num_local_nodes = 2, local_nodes[2] = {0}; 346 PetscMPIInt ranks_per_node; 347 PetscBool degree_set; 348 BPType bp_choices[10]; 349 PetscInt num_bp_choices = 10; 350 351 // Initialize PETSc 352 PetscCall(PetscInitialize(&argc, &argv, NULL, help)); 353 comm = PETSC_COMM_WORLD; 354 PetscCall(MPI_Comm_size(comm, &comm_size)); 355 #if defined(PETSC_HAVE_MPI_PROCESS_SHARED_MEMORY) 356 { 357 MPI_Comm splitcomm; 358 359 PetscCall(MPI_Comm_split_type(comm, MPI_COMM_TYPE_SHARED, 0, MPI_INFO_NULL, &splitcomm)); 360 PetscCall(MPI_Comm_size(splitcomm, &ranks_per_node)); 361 PetscCall(MPI_Comm_free(&splitcomm)); 362 } 363 #else 364 ranks_per_node = -1; // Unknown 365 #endif 366 367 // Setup all parameters needed in Run() 368 PetscCall(PetscMalloc1(1, &rp)); 369 rp->comm = comm; 370 371 // Read command line options 372 PetscOptionsBegin(comm, NULL, "CEED BPs in PETSc", NULL); 373 { 374 PetscBool set; 375 PetscCall(PetscOptionsEnumArray("-problem", "CEED benchmark problem to solve", NULL, bp_types, (PetscEnum *)bp_choices, &num_bp_choices, &set)); 376 if (!set) { 377 bp_choices[0] = CEED_BP1; 378 num_bp_choices = 1; 379 } 380 } 381 rp->test_mode = PETSC_FALSE; 382 PetscCall(PetscOptionsBool("-test", "Testing mode (do not print unless error is large)", NULL, rp->test_mode, &rp->test_mode, NULL)); 383 rp->write_solution = PETSC_FALSE; 384 PetscCall(PetscOptionsBool("-write_solution", "Write solution for visualization", NULL, rp->write_solution, &rp->write_solution, NULL)); 385 rp->simplex = PETSC_FALSE; 386 PetscCall(PetscOptionsBool("-simplex", "Element topology (default:hex)", NULL, rp->simplex, &rp->simplex, NULL)); 387 if ((bp_choices[0] == CEED_BP5 || bp_choices[0] == CEED_BP6) && (rp->simplex)) { 388 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "BP5/6 is not supported with simplex"); 389 } 390 degree[0] = rp->test_mode ? 3 : 2; 391 PetscCall(PetscOptionsIntArray("-degree", "Polynomial degree of tensor product basis", NULL, degree, &num_degrees, °ree_set)); 392 if (!degree_set) num_degrees = 1; 393 rp->q_extra = PETSC_DECIDE; 394 PetscCall(PetscOptionsInt("-q_extra", "Number of extra quadrature points (-1 for auto)", NULL, rp->q_extra, &rp->q_extra, NULL)); 395 { 396 PetscBool set; 397 PetscCall(PetscOptionsStringArray("-ceed", "CEED resource specifier (comma-separated list)", NULL, ceed_resources, &num_ceed_resources, &set)); 398 if (!set) { 399 PetscCall(PetscStrallocpy("/cpu/self", &ceed_resources[0])); 400 num_ceed_resources = 1; 401 } 402 } 403 PetscCall(PetscGetHostName(hostname, sizeof hostname)); 404 PetscCall(PetscOptionsString("-hostname", "Hostname for output", NULL, hostname, hostname, sizeof(hostname), NULL)); 405 rp->read_mesh = PETSC_FALSE; 406 PetscCall(PetscOptionsString("-mesh", "Read mesh from file", NULL, filename, filename, sizeof(filename), &rp->read_mesh)); 407 rp->filename = filename; 408 if (!rp->read_mesh) { 409 PetscInt tmp = dim; 410 PetscCall(PetscOptionsIntArray("-cells", "Number of cells per dimension", NULL, mesh_elem, &tmp, NULL)); 411 } 412 local_nodes[0] = 1000; 413 PetscCall(PetscOptionsIntArray("-local_nodes", 414 "Target number of locally owned nodes per " 415 "process (single value or min,max)", 416 NULL, local_nodes, &num_local_nodes, &rp->user_l_nodes)); 417 if (num_local_nodes < 2) local_nodes[1] = 2 * local_nodes[0]; 418 { 419 PetscInt two = 2; 420 rp->ksp_max_it_clip[0] = 5; 421 rp->ksp_max_it_clip[1] = 20; 422 PetscCall(PetscOptionsIntArray("-ksp_max_it_clip", "Min and max number of iterations to use during benchmarking", NULL, rp->ksp_max_it_clip, &two, 423 NULL)); 424 } 425 if (!degree_set) { 426 PetscInt max_degree = 8; 427 PetscCall(PetscOptionsInt("-max_degree", "Range of degrees [1, max_degree] to run with", NULL, max_degree, &max_degree, NULL)); 428 for (PetscInt i = 0; i < max_degree; i++) degree[i] = i + 1; 429 num_degrees = max_degree; 430 } 431 { 432 PetscBool flg; 433 PetscInt p = ranks_per_node; 434 PetscCall(PetscOptionsInt("-p", "Number of MPI ranks per node", NULL, p, &p, &flg)); 435 if (flg) ranks_per_node = p; 436 } 437 438 PetscOptionsEnd(); 439 440 // Register PETSc logging stage 441 PetscCall(PetscLogStageRegister("Solve Stage", &rp->solve_stage)); 442 443 rp->hostname = hostname; 444 rp->dim = dim; 445 rp->mesh_elem = mesh_elem; 446 rp->ranks_per_node = ranks_per_node; 447 448 for (PetscInt d = 0; d < num_degrees; d++) { 449 PetscInt deg = degree[d]; 450 for (PetscInt n = local_nodes[0]; n < local_nodes[1]; n *= 2) { 451 rp->degree = deg; 452 rp->local_nodes = n; 453 PetscCall(Run(rp, num_ceed_resources, ceed_resources, num_bp_choices, bp_choices)); 454 } 455 } 456 // Clear memory 457 PetscCall(PetscFree(rp)); 458 for (PetscInt i = 0; i < num_ceed_resources; i++) { 459 PetscCall(PetscFree(ceed_resources[i])); 460 } 461 return PetscFinalize(); 462 } 463