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 11 // CEED BP benchmark problems, see http://ceed.exascaleproject.org/bps, 12 // on a closed surface, such as the one of a discrete sphere. 13 // 14 // The code uses higher level communication protocols in DMPlex. 15 // 16 // Build with: 17 // 18 // make bpssphere [PETSC_DIR=</path/to/petsc>] [CEED_DIR=</path/to/libceed>] 19 // 20 // Sample runs: 21 // 22 // bpssphere -problem bp1 -degree 3 23 // bpssphere -problem bp2 -degree 3 24 // bpssphere -problem bp3 -degree 3 25 // bpssphere -problem bp4 -degree 3 26 // bpssphere -problem bp5 -degree 3 -ceed /cpu/self 27 // bpssphere -problem bp6 -degree 3 -ceed /gpu/cuda 28 // 29 //TESTARGS -ceed {ceed_resource} -test -problem bp3 -degree 3 -dm_refine 2 30 31 /// @file 32 /// CEED BPs example using PETSc with DMPlex 33 /// See bps.c for a "raw" implementation using a structured grid. 34 /// and bpsdmplex.c for an implementation using an unstructured grid. 35 static const char help[] = "Solve CEED BPs on a sphere using DMPlex in PETSc\n"; 36 37 #include <stdbool.h> 38 #include <string.h> 39 #include <ceed.h> 40 #include <petsc.h> 41 #include <petscdmplex.h> 42 #include <petscksp.h> 43 44 #include "bpssphere.h" 45 #include "include/sphereproblemdata.h" 46 #include "include/petscutils.h" 47 #include "include/petscversion.h" 48 #include "include/matops.h" 49 #include "include/libceedsetup.h" 50 51 52 #if PETSC_VERSION_LT(3,12,0) 53 #ifdef PETSC_HAVE_CUDA 54 #include <petsccuda.h> 55 // Note: With PETSc prior to version 3.12.0, providing the source path to 56 // include 'cublas_v2.h' will be needed to use 'petsccuda.h'. 57 #endif 58 #endif 59 60 int main(int argc, char **argv) { 61 PetscInt ierr; 62 MPI_Comm comm; 63 char ceed_resource[PETSC_MAX_PATH_LEN] = "/cpu/self", 64 filename[PETSC_MAX_PATH_LEN]; 65 double my_rt_start, my_rt, rt_min, rt_max; 66 PetscInt degree = 3, q_extra, l_size, g_size, topo_dim = 2, num_comp_x = 3, 67 num_comp_u = 1, xl_size; 68 PetscScalar *r; 69 PetscBool test_mode, benchmark_mode, read_mesh, write_solution, simplex; 70 PetscLogStage solve_stage; 71 Vec X, X_loc, rhs, rhs_loc; 72 Mat mat_O; 73 KSP ksp; 74 DM dm; 75 UserO user_O; 76 Ceed ceed; 77 CeedData ceed_data; 78 CeedQFunction qf_error; 79 CeedOperator op_error; 80 CeedVector rhs_ceed, target; 81 BPType bp_choice; 82 VecType vec_type; 83 PetscMemType mem_type; 84 85 ierr = PetscInitialize(&argc, &argv, NULL, help); 86 if (ierr) return ierr; 87 comm = PETSC_COMM_WORLD; 88 89 // Read command line options 90 PetscOptionsBegin(comm, NULL, "CEED BPs in PETSc", NULL); 91 bp_choice = CEED_BP1; 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 degree = test_mode ? 3 : 2; 112 ierr = PetscOptionsInt("-degree", "Polynomial degree of tensor product basis", 113 NULL, degree, °ree, NULL); CHKERRQ(ierr); 114 q_extra = bp_options[bp_choice].q_extra; 115 ierr = PetscOptionsInt("-q_extra", "Number of extra quadrature points", 116 NULL, q_extra, &q_extra, NULL); CHKERRQ(ierr); 117 ierr = PetscOptionsString("-ceed", "CEED resource specifier", 118 NULL, ceed_resource, ceed_resource, 119 sizeof(ceed_resource), NULL); CHKERRQ(ierr); 120 read_mesh = PETSC_FALSE; 121 ierr = PetscOptionsString("-mesh", "Read mesh from file", NULL, 122 filename, filename, sizeof(filename), &read_mesh); 123 CHKERRQ(ierr); 124 simplex = PETSC_FALSE; 125 ierr = PetscOptionsBool("-simplex", "Use simplices, or tensor product cells", 126 NULL, simplex, &simplex, NULL); CHKERRQ(ierr); 127 PetscOptionsEnd(); 128 129 // Setup DM 130 if (read_mesh) { 131 ierr = DMPlexCreateFromFile(PETSC_COMM_WORLD, filename, NULL, PETSC_TRUE, 132 &dm); 133 CHKERRQ(ierr); 134 } else { 135 // Create the mesh as a 0-refined sphere. This will create a cubic surface, 136 // not a box, and will snap to the unit sphere upon refinement. 137 ierr = DMPlexCreateSphereMesh(PETSC_COMM_WORLD, topo_dim, simplex, 1., &dm); 138 CHKERRQ(ierr); 139 // Set the object name 140 ierr = PetscObjectSetName((PetscObject)dm, "Sphere"); CHKERRQ(ierr); 141 // Refine DMPlex with uniform refinement using runtime option -dm_refine 142 ierr = DMPlexSetRefinementUniform(dm, PETSC_TRUE); CHKERRQ(ierr); 143 } 144 ierr = DMSetFromOptions(dm); CHKERRQ(ierr); 145 // View DMPlex via runtime option 146 ierr = DMViewFromOptions(dm, NULL, "-dm_view"); CHKERRQ(ierr); 147 148 // Create DM 149 ierr = SetupDMByDegree(dm, degree, num_comp_u, topo_dim, false, 150 (BCFunction)NULL); 151 CHKERRQ(ierr); 152 153 // Create vectors 154 ierr = DMCreateGlobalVector(dm, &X); CHKERRQ(ierr); 155 ierr = VecGetLocalSize(X, &l_size); CHKERRQ(ierr); 156 ierr = VecGetSize(X, &g_size); CHKERRQ(ierr); 157 ierr = DMCreateLocalVector(dm, &X_loc); CHKERRQ(ierr); 158 ierr = VecGetSize(X_loc, &xl_size); CHKERRQ(ierr); 159 ierr = VecDuplicate(X, &rhs); CHKERRQ(ierr); 160 161 // Operator 162 ierr = PetscMalloc1(1, &user_O); CHKERRQ(ierr); 163 ierr = MatCreateShell(comm, l_size, l_size, g_size, g_size, 164 user_O, &mat_O); CHKERRQ(ierr); 165 ierr = MatShellSetOperation(mat_O, MATOP_MULT, 166 (void(*)(void))MatMult_Ceed); CHKERRQ(ierr); 167 168 // Set up libCEED 169 CeedInit(ceed_resource, &ceed); 170 CeedMemType mem_type_backend; 171 CeedGetPreferredMemType(ceed, &mem_type_backend); 172 173 ierr = DMGetVecType(dm, &vec_type); CHKERRQ(ierr); 174 if (!vec_type) { // Not yet set by user -dm_vec_type 175 switch (mem_type_backend) { 176 case CEED_MEM_HOST: vec_type = VECSTANDARD; break; 177 case CEED_MEM_DEVICE: { 178 const char *resolved; 179 CeedGetResource(ceed, &resolved); 180 if (strstr(resolved, "/gpu/cuda")) vec_type = VECCUDA; 181 else if (strstr(resolved, "/gpu/hip/occa")) 182 vec_type = VECSTANDARD; // https://github.com/CEED/libCEED/issues/678 183 else if (strstr(resolved, "/gpu/hip")) vec_type = VECHIP; 184 else vec_type = VECSTANDARD; 185 } 186 } 187 ierr = DMSetVecType(dm, vec_type); CHKERRQ(ierr); 188 } 189 190 // Print summary 191 if (!test_mode) { 192 PetscInt P = degree + 1, Q = P + q_extra; 193 const char *used_resource; 194 CeedGetResource(ceed, &used_resource); 195 ierr = PetscPrintf(comm, 196 "\n-- CEED Benchmark Problem %" CeedInt_FMT 197 " on the Sphere -- libCEED + PETSc --\n" 198 " libCEED:\n" 199 " libCEED Backend : %s\n" 200 " libCEED Backend MemType : %s\n" 201 " Mesh:\n" 202 " Number of 1D Basis Nodes (p) : %" CeedInt_FMT "\n" 203 " Number of 1D Quadrature Points (q) : %" CeedInt_FMT "\n" 204 " Global nodes : %" PetscInt_FMT "\n", 205 bp_choice+1, ceed_resource, CeedMemTypes[mem_type_backend], P, Q, 206 g_size/num_comp_u); CHKERRQ(ierr); 207 } 208 209 // Create RHS vector 210 ierr = VecDuplicate(X_loc, &rhs_loc); CHKERRQ(ierr); 211 ierr = VecZeroEntries(rhs_loc); CHKERRQ(ierr); 212 ierr = VecGetArrayAndMemType(rhs_loc, &r, &mem_type); CHKERRQ(ierr); 213 CeedVectorCreate(ceed, xl_size, &rhs_ceed); 214 CeedVectorSetArray(rhs_ceed, MemTypeP2C(mem_type), CEED_USE_POINTER, r); 215 216 // Setup libCEED's objects 217 ierr = PetscMalloc1(1, &ceed_data); CHKERRQ(ierr); 218 ierr = SetupLibceedByDegree(dm, ceed, degree, topo_dim, q_extra, num_comp_x, 219 num_comp_u, g_size, xl_size, bp_options[bp_choice], 220 ceed_data, true, rhs_ceed, &target); CHKERRQ(ierr); 221 222 // Gather RHS 223 CeedVectorTakeArray(rhs_ceed, MemTypeP2C(mem_type), NULL); 224 ierr = VecRestoreArrayAndMemType(rhs_loc, &r); CHKERRQ(ierr); 225 ierr = VecZeroEntries(rhs); CHKERRQ(ierr); 226 ierr = DMLocalToGlobal(dm, rhs_loc, ADD_VALUES, rhs); CHKERRQ(ierr); 227 CeedVectorDestroy(&rhs_ceed); 228 229 // Create the error Q-function 230 CeedQFunctionCreateInterior(ceed, 1, bp_options[bp_choice].error, 231 bp_options[bp_choice].error_loc, &qf_error); 232 CeedQFunctionAddInput(qf_error, "u", num_comp_u, CEED_EVAL_INTERP); 233 CeedQFunctionAddInput(qf_error, "true_soln", num_comp_u, CEED_EVAL_NONE); 234 CeedQFunctionAddOutput(qf_error, "error", num_comp_u, CEED_EVAL_NONE); 235 236 // Create the error operator 237 CeedOperatorCreate(ceed, qf_error, NULL, NULL, &op_error); 238 CeedOperatorSetField(op_error, "u", ceed_data->elem_restr_u, 239 ceed_data->basis_u, CEED_VECTOR_ACTIVE); 240 CeedOperatorSetField(op_error, "true_soln", ceed_data->elem_restr_u_i, 241 CEED_BASIS_COLLOCATED, target); 242 CeedOperatorSetField(op_error, "error", ceed_data->elem_restr_u_i, 243 CEED_BASIS_COLLOCATED, CEED_VECTOR_ACTIVE); 244 245 // Set up Mat 246 user_O->comm = comm; 247 user_O->dm = dm; 248 user_O->X_loc = X_loc; 249 ierr = VecDuplicate(X_loc, &user_O->Y_loc); CHKERRQ(ierr); 250 user_O->x_ceed = ceed_data->x_ceed; 251 user_O->y_ceed = ceed_data->y_ceed; 252 user_O->op = ceed_data->op_apply; 253 user_O->ceed = ceed; 254 255 // Setup solver 256 ierr = KSPCreate(comm, &ksp); CHKERRQ(ierr); 257 { 258 PC pc; 259 ierr = KSPGetPC(ksp, &pc); CHKERRQ(ierr); 260 if (bp_choice == CEED_BP1 || bp_choice == CEED_BP2) { 261 ierr = PCSetType(pc, PCJACOBI); CHKERRQ(ierr); 262 ierr = PCJacobiSetType(pc, PC_JACOBI_ROWSUM); CHKERRQ(ierr); 263 } else { 264 ierr = PCSetType(pc, PCNONE); CHKERRQ(ierr); 265 MatNullSpace nullspace; 266 267 ierr = MatNullSpaceCreate(PETSC_COMM_WORLD, PETSC_TRUE, 0, 0, &nullspace); 268 CHKERRQ(ierr); 269 ierr = MatSetNullSpace(mat_O, nullspace); CHKERRQ(ierr); 270 ierr = MatNullSpaceDestroy(&nullspace); CHKERRQ(ierr); 271 } 272 ierr = KSPSetType(ksp, KSPCG); CHKERRQ(ierr); 273 ierr = KSPSetNormType(ksp, KSP_NORM_NATURAL); CHKERRQ(ierr); 274 ierr = KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, 275 PETSC_DEFAULT); CHKERRQ(ierr); 276 } 277 ierr = KSPSetFromOptions(ksp); CHKERRQ(ierr); 278 ierr = KSPSetOperators(ksp, mat_O, mat_O); CHKERRQ(ierr); 279 280 // First run, if benchmarking 281 if (benchmark_mode) { 282 ierr = KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, 1); 283 CHKERRQ(ierr); 284 my_rt_start = MPI_Wtime(); 285 ierr = KSPSolve(ksp, rhs, X); CHKERRQ(ierr); 286 my_rt = MPI_Wtime() - my_rt_start; 287 ierr = MPI_Allreduce(MPI_IN_PLACE, &my_rt, 1, MPI_DOUBLE, MPI_MIN, comm); 288 CHKERRQ(ierr); 289 // Set maxits based on first iteration timing 290 if (my_rt > 0.02) { 291 ierr = KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, 5); 292 CHKERRQ(ierr); 293 } else { 294 ierr = KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, 20); 295 CHKERRQ(ierr); 296 } 297 } 298 299 // Timed solve 300 ierr = VecZeroEntries(X); CHKERRQ(ierr); 301 ierr = PetscBarrier((PetscObject)ksp); CHKERRQ(ierr); 302 303 // -- Performance logging 304 ierr = PetscLogStageRegister("Solve Stage", &solve_stage); CHKERRQ(ierr); 305 ierr = PetscLogStagePush(solve_stage); CHKERRQ(ierr); 306 307 // -- Solve 308 my_rt_start = MPI_Wtime(); 309 ierr = KSPSolve(ksp, rhs, X); CHKERRQ(ierr); 310 my_rt = MPI_Wtime() - my_rt_start; 311 312 // -- Performance logging 313 ierr = PetscLogStagePop(); 314 315 // Output results 316 { 317 KSPType ksp_type; 318 KSPConvergedReason reason; 319 PetscReal rnorm; 320 PetscInt its; 321 ierr = KSPGetType(ksp, &ksp_type); CHKERRQ(ierr); 322 ierr = KSPGetConvergedReason(ksp, &reason); CHKERRQ(ierr); 323 ierr = KSPGetIterationNumber(ksp, &its); CHKERRQ(ierr); 324 ierr = KSPGetResidualNorm(ksp, &rnorm); CHKERRQ(ierr); 325 if (!test_mode || reason < 0 || rnorm > 1e-8) { 326 ierr = PetscPrintf(comm, 327 " KSP:\n" 328 " KSP Type : %s\n" 329 " KSP Convergence : %s\n" 330 " Total KSP Iterations : %" PetscInt_FMT "\n" 331 " Final rnorm : %e\n", 332 ksp_type, KSPConvergedReasons[reason], its, 333 (double)rnorm); CHKERRQ(ierr); 334 } 335 if (!test_mode) { 336 ierr = PetscPrintf(comm," Performance:\n"); CHKERRQ(ierr); 337 } 338 { 339 PetscReal max_error; 340 ierr = ComputeErrorMax(user_O, op_error, X, target, &max_error); 341 CHKERRQ(ierr); 342 PetscReal tol = 5e-4; 343 if (!test_mode || max_error > tol) { 344 ierr = MPI_Allreduce(&my_rt, &rt_min, 1, MPI_DOUBLE, MPI_MIN, comm); 345 CHKERRQ(ierr); 346 ierr = MPI_Allreduce(&my_rt, &rt_max, 1, MPI_DOUBLE, MPI_MAX, comm); 347 CHKERRQ(ierr); 348 ierr = PetscPrintf(comm, 349 " Pointwise Error (max) : %e\n" 350 " CG Solve Time : %g (%g) sec\n", 351 (double)max_error, rt_max, rt_min); CHKERRQ(ierr); 352 } 353 } 354 if (benchmark_mode && (!test_mode)) { 355 ierr = PetscPrintf(comm, 356 " DoFs/Sec in CG : %g (%g) million\n", 357 1e-6*g_size*its/rt_max, 1e-6*g_size*its/rt_min); CHKERRQ(ierr); 358 } 359 } 360 361 // Output solution 362 if (write_solution) { 363 PetscViewer vtk_viewer_soln; 364 365 ierr = PetscViewerCreate(comm, &vtk_viewer_soln); CHKERRQ(ierr); 366 ierr = PetscViewerSetType(vtk_viewer_soln, PETSCVIEWERVTK); CHKERRQ(ierr); 367 ierr = PetscViewerFileSetName(vtk_viewer_soln, "solution.vtu"); CHKERRQ(ierr); 368 ierr = VecView(X, vtk_viewer_soln); CHKERRQ(ierr); 369 ierr = PetscViewerDestroy(&vtk_viewer_soln); CHKERRQ(ierr); 370 } 371 372 // Cleanup 373 ierr = VecDestroy(&X); CHKERRQ(ierr); 374 ierr = VecDestroy(&X_loc); CHKERRQ(ierr); 375 ierr = VecDestroy(&user_O->Y_loc); CHKERRQ(ierr); 376 ierr = MatDestroy(&mat_O); CHKERRQ(ierr); 377 ierr = PetscFree(user_O); CHKERRQ(ierr); 378 ierr = CeedDataDestroy(0, ceed_data); CHKERRQ(ierr); 379 ierr = DMDestroy(&dm); CHKERRQ(ierr); 380 381 ierr = VecDestroy(&rhs); CHKERRQ(ierr); 382 ierr = VecDestroy(&rhs_loc); CHKERRQ(ierr); 383 ierr = KSPDestroy(&ksp); CHKERRQ(ierr); 384 CeedVectorDestroy(&target); 385 CeedQFunctionDestroy(&qf_error); 386 CeedOperatorDestroy(&op_error); 387 CeedDestroy(&ceed); 388 return PetscFinalize(); 389 } 390