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 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, on 11 // a closed surface, such as the one of a discrete sphere. 12 // 13 // The code uses higher level communication protocols in DMPlex. 14 // 15 // Build with: 16 // 17 // make bpssphere [PETSC_DIR=</path/to/petsc>] [CEED_DIR=</path/to/libceed>] 18 // 19 // Sample runs: 20 // 21 // bpssphere -problem bp1 -degree 3 22 // bpssphere -problem bp2 -degree 3 23 // bpssphere -problem bp3 -degree 3 24 // bpssphere -problem bp4 -degree 3 25 // bpssphere -problem bp5 -degree 3 -ceed /cpu/self 26 // bpssphere -problem bp6 -degree 3 -ceed /gpu/cuda 27 // 28 //TESTARGS -ceed {ceed_resource} -test -problem bp3 -degree 3 -dm_refine 2 29 30 /// @file 31 /// CEED BPs example using PETSc with DMPlex 32 /// See bps.c for a "raw" implementation using a structured grid and bpsdmplex.c for an implementation using an unstructured grid. 33 static const char help[] = "Solve CEED BPs on a sphere using DMPlex in PETSc\n"; 34 35 #include "bpssphere.h" 36 37 #include <ceed.h> 38 #include <petsc.h> 39 #include <petscdmplex.h> 40 #include <petscksp.h> 41 #include <stdbool.h> 42 #include <string.h> 43 44 #include "include/libceedsetup.h" 45 #include "include/matops.h" 46 #include "include/petscutils.h" 47 #include "include/petscversion.h" 48 #include "include/sphereproblemdata.h" 49 50 #if PETSC_VERSION_LT(3, 12, 0) 51 #ifdef PETSC_HAVE_CUDA 52 #include <petsccuda.h> 53 // 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'. 54 #endif 55 #endif 56 57 int main(int argc, char **argv) { 58 MPI_Comm comm; 59 char ceed_resource[PETSC_MAX_PATH_LEN] = "/cpu/self", filename[PETSC_MAX_PATH_LEN]; 60 double my_rt_start, my_rt, rt_min, rt_max; 61 PetscInt degree = 3, q_extra, l_size, g_size, topo_dim = 2, num_comp_x = 3, num_comp_u = 1, xl_size; 62 PetscScalar *r; 63 PetscBool test_mode, benchmark_mode, read_mesh, write_solution, simplex; 64 PetscLogStage solve_stage; 65 Vec X, X_loc, rhs, rhs_loc; 66 Mat mat_O; 67 KSP ksp; 68 DM dm; 69 OperatorApplyContext op_apply_ctx, op_error_ctx; 70 Ceed ceed; 71 CeedData ceed_data; 72 CeedQFunction qf_error; 73 CeedOperator op_error; 74 CeedVector rhs_ceed, target; 75 BPType bp_choice; 76 VecType vec_type; 77 PetscMemType mem_type; 78 79 PetscCall(PetscInitialize(&argc, &argv, NULL, help)); 80 comm = PETSC_COMM_WORLD; 81 82 // Read command line options 83 PetscOptionsBegin(comm, NULL, "CEED BPs in PETSc", NULL); 84 bp_choice = CEED_BP1; 85 PetscCall(PetscOptionsEnum("-problem", "CEED benchmark problem to solve", NULL, bp_types, (PetscEnum)bp_choice, (PetscEnum *)&bp_choice, NULL)); 86 num_comp_u = bp_options[bp_choice].num_comp_u; 87 test_mode = PETSC_FALSE; 88 PetscCall(PetscOptionsBool("-test", "Testing mode (do not print unless error is large)", NULL, test_mode, &test_mode, NULL)); 89 benchmark_mode = PETSC_FALSE; 90 PetscCall(PetscOptionsBool("-benchmark", "Benchmarking mode (prints benchmark statistics)", NULL, benchmark_mode, &benchmark_mode, NULL)); 91 write_solution = PETSC_FALSE; 92 PetscCall(PetscOptionsBool("-write_solution", "Write solution for visualization", NULL, write_solution, &write_solution, NULL)); 93 degree = test_mode ? 3 : 2; 94 PetscCall(PetscOptionsInt("-degree", "Polynomial degree of tensor product basis", NULL, degree, °ree, NULL)); 95 q_extra = bp_options[bp_choice].q_extra; 96 PetscCall(PetscOptionsInt("-q_extra", "Number of extra quadrature points", NULL, q_extra, &q_extra, NULL)); 97 PetscCall(PetscOptionsString("-ceed", "CEED resource specifier", NULL, ceed_resource, ceed_resource, sizeof(ceed_resource), NULL)); 98 read_mesh = PETSC_FALSE; 99 PetscCall(PetscOptionsString("-mesh", "Read mesh from file", NULL, filename, filename, sizeof(filename), &read_mesh)); 100 simplex = PETSC_FALSE; 101 PetscCall(PetscOptionsBool("-simplex", "Use simplices, or tensor product cells", NULL, simplex, &simplex, NULL)); 102 PetscOptionsEnd(); 103 104 // Setup DM 105 if (read_mesh) { 106 PetscCall(DMPlexCreateFromFile(PETSC_COMM_WORLD, filename, NULL, PETSC_TRUE, &dm)); 107 } else { 108 // Create the mesh as a 0-refined sphere. 109 // This will create a cubic surface, not a box, and will snap to the unit sphere upon refinement. 110 PetscCall(DMPlexCreateSphereMesh(PETSC_COMM_WORLD, topo_dim, simplex, 1., &dm)); 111 // Set the object name 112 PetscCall(PetscObjectSetName((PetscObject)dm, "Sphere")); 113 // Refine DMPlex with uniform refinement using runtime option -dm_refine 114 PetscCall(DMPlexSetRefinementUniform(dm, PETSC_TRUE)); 115 } 116 PetscCall(DMSetFromOptions(dm)); 117 // View DMPlex via runtime option 118 PetscCall(DMViewFromOptions(dm, NULL, "-dm_view")); 119 120 // Create DM 121 PetscCall(SetupDMByDegree(dm, degree, q_extra, num_comp_u, topo_dim, false)); 122 123 // Create vectors 124 PetscCall(DMCreateGlobalVector(dm, &X)); 125 PetscCall(VecGetLocalSize(X, &l_size)); 126 PetscCall(VecGetSize(X, &g_size)); 127 PetscCall(DMCreateLocalVector(dm, &X_loc)); 128 PetscCall(VecGetSize(X_loc, &xl_size)); 129 PetscCall(VecDuplicate(X, &rhs)); 130 131 // Operator 132 PetscCall(PetscMalloc1(1, &op_apply_ctx)); 133 PetscCall(PetscMalloc1(1, &op_error_ctx)); 134 PetscCall(MatCreateShell(comm, l_size, l_size, g_size, g_size, op_apply_ctx, &mat_O)); 135 PetscCall(MatShellSetOperation(mat_O, MATOP_MULT, (void (*)(void))MatMult_Ceed)); 136 137 // Set up libCEED 138 CeedInit(ceed_resource, &ceed); 139 CeedMemType mem_type_backend; 140 CeedGetPreferredMemType(ceed, &mem_type_backend); 141 142 PetscCall(DMGetVecType(dm, &vec_type)); 143 if (!vec_type) { // Not yet set by user -dm_vec_type 144 switch (mem_type_backend) { 145 case CEED_MEM_HOST: 146 vec_type = VECSTANDARD; 147 break; 148 case CEED_MEM_DEVICE: { 149 const char *resolved; 150 CeedGetResource(ceed, &resolved); 151 if (strstr(resolved, "/gpu/cuda")) vec_type = VECCUDA; 152 else if (strstr(resolved, "/gpu/hip/occa")) vec_type = VECSTANDARD; // https://github.com/CEED/libCEED/issues/678 153 else if (strstr(resolved, "/gpu/hip")) vec_type = VECHIP; 154 else vec_type = VECSTANDARD; 155 } 156 } 157 PetscCall(DMSetVecType(dm, vec_type)); 158 } 159 160 // Print summary 161 if (!test_mode) { 162 PetscInt P = degree + 1, Q = P + q_extra; 163 const char *used_resource; 164 CeedGetResource(ceed, &used_resource); 165 PetscCall(PetscPrintf(comm, 166 "\n-- CEED Benchmark Problem %" CeedInt_FMT " on the Sphere -- libCEED + PETSc --\n" 167 " libCEED:\n" 168 " libCEED Backend : %s\n" 169 " libCEED Backend MemType : %s\n" 170 " Mesh:\n" 171 " Solution Order (P) : %" CeedInt_FMT "\n" 172 " Quadrature Order (Q) : %" CeedInt_FMT "\n" 173 " Additional quadrature points (q_extra) : %" CeedInt_FMT "\n" 174 " Global nodes : %" PetscInt_FMT "\n", 175 bp_choice + 1, ceed_resource, CeedMemTypes[mem_type_backend], P, Q, q_extra, g_size / num_comp_u)); 176 } 177 178 // Create RHS vector 179 PetscCall(VecDuplicate(X_loc, &rhs_loc)); 180 PetscCall(VecZeroEntries(rhs_loc)); 181 PetscCall(VecGetArrayAndMemType(rhs_loc, &r, &mem_type)); 182 CeedVectorCreate(ceed, xl_size, &rhs_ceed); 183 CeedVectorSetArray(rhs_ceed, MemTypeP2C(mem_type), CEED_USE_POINTER, r); 184 185 // Setup libCEED's objects 186 PetscCall(PetscMalloc1(1, &ceed_data)); 187 PetscCall(SetupLibceedByDegree(dm, ceed, degree, topo_dim, q_extra, num_comp_x, num_comp_u, g_size, xl_size, bp_options[bp_choice], ceed_data, true, 188 rhs_ceed, &target)); 189 190 // Gather RHS 191 CeedVectorTakeArray(rhs_ceed, MemTypeP2C(mem_type), NULL); 192 PetscCall(VecRestoreArrayAndMemType(rhs_loc, &r)); 193 PetscCall(VecZeroEntries(rhs)); 194 PetscCall(DMLocalToGlobal(dm, rhs_loc, ADD_VALUES, rhs)); 195 CeedVectorDestroy(&rhs_ceed); 196 197 // Create the error Q-function 198 CeedQFunctionCreateInterior(ceed, 1, bp_options[bp_choice].error, bp_options[bp_choice].error_loc, &qf_error); 199 CeedQFunctionAddInput(qf_error, "u", num_comp_u, CEED_EVAL_INTERP); 200 CeedQFunctionAddInput(qf_error, "true_soln", num_comp_u, CEED_EVAL_NONE); 201 CeedQFunctionAddInput(qf_error, "qdata", ceed_data->q_data_size, CEED_EVAL_NONE); 202 CeedQFunctionAddOutput(qf_error, "error", num_comp_u, CEED_EVAL_INTERP); 203 204 // Create the error operator 205 CeedOperatorCreate(ceed, qf_error, NULL, NULL, &op_error); 206 CeedOperatorSetField(op_error, "u", ceed_data->elem_restr_u, ceed_data->basis_u, CEED_VECTOR_ACTIVE); 207 CeedOperatorSetField(op_error, "true_soln", ceed_data->elem_restr_u_i, CEED_BASIS_COLLOCATED, target); 208 CeedOperatorSetField(op_error, "qdata", ceed_data->elem_restr_qd_i, CEED_BASIS_COLLOCATED, ceed_data->q_data); 209 CeedOperatorSetField(op_error, "error", ceed_data->elem_restr_u, ceed_data->basis_u, CEED_VECTOR_ACTIVE); 210 211 // Set up apply operator context 212 PetscCall(SetupApplyOperatorCtx(comm, dm, ceed, ceed_data, X_loc, op_apply_ctx)); 213 214 // Setup solver 215 PetscCall(KSPCreate(comm, &ksp)); 216 { 217 PC pc; 218 PetscCall(KSPGetPC(ksp, &pc)); 219 if (bp_choice == CEED_BP1 || bp_choice == CEED_BP2) { 220 PetscCall(PCSetType(pc, PCJACOBI)); 221 PetscCall(PCJacobiSetType(pc, PC_JACOBI_ROWSUM)); 222 } else { 223 PetscCall(PCSetType(pc, PCNONE)); 224 MatNullSpace nullspace; 225 226 PetscCall(MatNullSpaceCreate(PETSC_COMM_WORLD, PETSC_TRUE, 0, 0, &nullspace)); 227 PetscCall(MatSetNullSpace(mat_O, nullspace)); 228 PetscCall(MatNullSpaceDestroy(&nullspace)); 229 } 230 PetscCall(KSPSetType(ksp, KSPCG)); 231 PetscCall(KSPSetNormType(ksp, KSP_NORM_NATURAL)); 232 PetscCall(KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, PETSC_DEFAULT)); 233 } 234 PetscCall(KSPSetFromOptions(ksp)); 235 PetscCall(KSPSetOperators(ksp, mat_O, mat_O)); 236 237 // First run, if benchmarking 238 if (benchmark_mode) { 239 PetscCall(KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, 1)); 240 my_rt_start = MPI_Wtime(); 241 PetscCall(KSPSolve(ksp, rhs, X)); 242 my_rt = MPI_Wtime() - my_rt_start; 243 PetscCall(MPI_Allreduce(MPI_IN_PLACE, &my_rt, 1, MPI_DOUBLE, MPI_MIN, comm)); 244 // Set maxits based on first iteration timing 245 if (my_rt > 0.02) { 246 PetscCall(KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, 5)); 247 } else { 248 PetscCall(KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, 20)); 249 } 250 } 251 252 // Timed solve 253 PetscCall(VecZeroEntries(X)); 254 PetscCall(PetscBarrier((PetscObject)ksp)); 255 256 // -- Performance logging 257 PetscCall(PetscLogStageRegister("Solve Stage", &solve_stage)); 258 PetscCall(PetscLogStagePush(solve_stage)); 259 260 // -- Solve 261 my_rt_start = MPI_Wtime(); 262 PetscCall(KSPSolve(ksp, rhs, X)); 263 my_rt = MPI_Wtime() - my_rt_start; 264 265 // -- Performance logging 266 PetscCall(PetscLogStagePop()); 267 268 // Output results 269 { 270 KSPType ksp_type; 271 KSPConvergedReason reason; 272 PetscReal rnorm; 273 PetscInt its; 274 PetscCall(KSPGetType(ksp, &ksp_type)); 275 PetscCall(KSPGetConvergedReason(ksp, &reason)); 276 PetscCall(KSPGetIterationNumber(ksp, &its)); 277 PetscCall(KSPGetResidualNorm(ksp, &rnorm)); 278 if (!test_mode || reason < 0 || rnorm > 1e-8) { 279 PetscCall(PetscPrintf(comm, 280 " KSP:\n" 281 " KSP Type : %s\n" 282 " KSP Convergence : %s\n" 283 " Total KSP Iterations : %" PetscInt_FMT "\n" 284 " Final rnorm : %e\n", 285 ksp_type, KSPConvergedReasons[reason], its, (double)rnorm)); 286 } 287 if (!test_mode) { 288 PetscCall(PetscPrintf(comm, " Performance:\n")); 289 } 290 { 291 // Set up error operator context 292 PetscCall(SetupErrorOperatorCtx(comm, dm, ceed, ceed_data, X_loc, op_error, op_error_ctx)); 293 PetscScalar l2_error; 294 PetscCall(ComputeL2Error(X, &l2_error, op_error_ctx)); 295 PetscReal tol = 5e-4; 296 if (!test_mode || l2_error > tol) { 297 PetscCall(MPI_Allreduce(&my_rt, &rt_min, 1, MPI_DOUBLE, MPI_MIN, comm)); 298 PetscCall(MPI_Allreduce(&my_rt, &rt_max, 1, MPI_DOUBLE, MPI_MAX, comm)); 299 PetscCall(PetscPrintf(comm, 300 " L2 Error : %e\n" 301 " CG Solve Time : %g (%g) sec\n", 302 (double)l2_error, rt_max, rt_min)); 303 } 304 } 305 if (benchmark_mode && (!test_mode)) { 306 PetscCall(PetscPrintf(comm, " DoFs/Sec in CG : %g (%g) million\n", 1e-6 * g_size * its / rt_max, 307 1e-6 * g_size * its / rt_min)); 308 } 309 } 310 311 // Output solution 312 if (write_solution) { 313 PetscViewer vtk_viewer_soln; 314 315 PetscCall(PetscViewerCreate(comm, &vtk_viewer_soln)); 316 PetscCall(PetscViewerSetType(vtk_viewer_soln, PETSCVIEWERVTK)); 317 PetscCall(PetscViewerFileSetName(vtk_viewer_soln, "solution.vtu")); 318 PetscCall(VecView(X, vtk_viewer_soln)); 319 PetscCall(PetscViewerDestroy(&vtk_viewer_soln)); 320 } 321 322 // Cleanup 323 PetscCall(VecDestroy(&X)); 324 PetscCall(VecDestroy(&X_loc)); 325 PetscCall(VecDestroy(&op_apply_ctx->Y_loc)); 326 PetscCall(VecDestroy(&op_error_ctx->Y_loc)); 327 PetscCall(MatDestroy(&mat_O)); 328 PetscCall(PetscFree(op_apply_ctx)); 329 PetscCall(PetscFree(op_error_ctx)); 330 PetscCall(CeedDataDestroy(0, ceed_data)); 331 PetscCall(DMDestroy(&dm)); 332 333 PetscCall(VecDestroy(&rhs)); 334 PetscCall(VecDestroy(&rhs_loc)); 335 PetscCall(KSPDestroy(&ksp)); 336 CeedVectorDestroy(&target); 337 CeedQFunctionDestroy(&qf_error); 338 CeedOperatorDestroy(&op_error); 339 CeedDestroy(&ceed); 340 return PetscFinalize(); 341 } 342