1 // Copyright (c) 2017, Lawrence Livermore National Security, LLC. Produced at 2 // the Lawrence Livermore National Laboratory. LLNL-CODE-734707. All Rights 3 // reserved. See files LICENSE and NOTICE for details. 4 // 5 // This file is part of CEED, a collection of benchmarks, miniapps, software 6 // libraries and APIs for efficient high-order finite element and spectral 7 // element discretizations for exascale applications. For more information and 8 // source code availability see http://github.com/ceed. 9 // 10 // The CEED research is supported by the Exascale Computing Project 17-SC-20-SC, 11 // a collaborative effort of two U.S. Department of Energy organizations (Office 12 // of Science and the National Nuclear Security Administration) responsible for 13 // the planning and preparation of a capable exascale ecosystem, including 14 // software, applications, hardware, advanced system engineering and early 15 // testbed platforms, in support of the nation's exascale computing imperative. 16 17 // libCEED + PETSc Example: CEED BPs 18 // 19 // This example demonstrates a simple usage of libCEED with PETSc to solve the 20 // CEED BP benchmark problems, see http://ceed.exascaleproject.org/bps, 21 // on a closed surface, such as the one of a discrete sphere. 22 // 23 // The code uses higher level communication protocols in DMPlex. 24 // 25 // Build with: 26 // 27 // make bpssphere [PETSC_DIR=</path/to/petsc>] [CEED_DIR=</path/to/libceed>] 28 // 29 // Sample runs: 30 // 31 // bpssphere -problem bp1 -degree 3 32 // bpssphere -problem bp2 -degree 3 33 // bpssphere -problem bp3 -degree 3 34 // bpssphere -problem bp4 -degree 3 35 // bpssphere -problem bp5 -degree 3 -ceed /cpu/self 36 // bpssphere -problem bp6 -degree 3 -ceed /gpu/cuda 37 // 38 //TESTARGS -ceed {ceed_resource} -test -problem bp3 -degree 3 -dm_refine 2 39 40 /// @file 41 /// CEED BPs example using PETSc with DMPlex 42 /// See bps.c for a "raw" implementation using a structured grid. 43 /// and bpsdmplex.c for an implementation using an unstructured grid. 44 static const char help[] = "Solve CEED BPs on a sphere using DMPlex in PETSc\n"; 45 46 #include <stdbool.h> 47 #include <string.h> 48 #include <ceed.h> 49 #include <petsc.h> 50 #include <petscdmplex.h> 51 #include <petscksp.h> 52 53 #include "bpssphere.h" 54 #include "include/sphereproblemdata.h" 55 #include "include/petscutils.h" 56 #include "include/petscversion.h" 57 #include "include/matops.h" 58 #include "include/libceedsetup.h" 59 60 61 #if PETSC_VERSION_LT(3,12,0) 62 #ifdef PETSC_HAVE_CUDA 63 #include <petsccuda.h> 64 // Note: With PETSc prior to version 3.12.0, providing the source path to 65 // include 'cublas_v2.h' will be needed to use 'petsccuda.h'. 66 #endif 67 #endif 68 69 int main(int argc, char **argv) { 70 PetscInt ierr; 71 MPI_Comm comm; 72 char ceed_resource[PETSC_MAX_PATH_LEN] = "/cpu/self", 73 filename[PETSC_MAX_PATH_LEN]; 74 double my_rt_start, my_rt, rt_min, rt_max; 75 PetscInt degree = 3, q_extra, l_size, g_size, topo_dim = 2, num_comp_x = 3, 76 num_comp_u = 1, xl_size; 77 PetscScalar *r; 78 PetscBool test_mode, benchmark_mode, read_mesh, write_solution, simplex; 79 PetscLogStage solve_stage; 80 Vec X, X_loc, rhs, rhs_loc; 81 Mat mat_O; 82 KSP ksp; 83 DM dm; 84 UserO user_O; 85 Ceed ceed; 86 CeedData ceed_data; 87 CeedQFunction qf_error; 88 CeedOperator op_error; 89 CeedVector rhs_ceed, target; 90 BPType bp_choice; 91 VecType vec_type; 92 PetscMemType mem_type; 93 94 ierr = PetscInitialize(&argc, &argv, NULL, help); 95 if (ierr) return ierr; 96 comm = PETSC_COMM_WORLD; 97 98 // Read command line options 99 ierr = PetscOptionsBegin(comm, NULL, "CEED BPs in PETSc", NULL); CHKERRQ(ierr); 100 bp_choice = CEED_BP1; 101 ierr = PetscOptionsEnum("-problem", 102 "CEED benchmark problem to solve", NULL, 103 bp_types, (PetscEnum)bp_choice, (PetscEnum *)&bp_choice, 104 NULL); CHKERRQ(ierr); 105 num_comp_u = bp_options[bp_choice].num_comp_u; 106 test_mode = PETSC_FALSE; 107 ierr = PetscOptionsBool("-test", 108 "Testing mode (do not print unless error is large)", 109 NULL, test_mode, &test_mode, NULL); CHKERRQ(ierr); 110 benchmark_mode = PETSC_FALSE; 111 ierr = PetscOptionsBool("-benchmark", 112 "Benchmarking mode (prints benchmark statistics)", 113 NULL, benchmark_mode, &benchmark_mode, NULL); 114 CHKERRQ(ierr); 115 write_solution = PETSC_FALSE; 116 ierr = PetscOptionsBool("-write_solution", 117 "Write solution for visualization", 118 NULL, write_solution, &write_solution, NULL); 119 CHKERRQ(ierr); 120 degree = test_mode ? 3 : 2; 121 ierr = PetscOptionsInt("-degree", "Polynomial degree of tensor product basis", 122 NULL, degree, °ree, NULL); CHKERRQ(ierr); 123 q_extra = bp_options[bp_choice].q_extra; 124 ierr = PetscOptionsInt("-q_extra", "Number of extra quadrature points", 125 NULL, q_extra, &q_extra, NULL); CHKERRQ(ierr); 126 ierr = PetscOptionsString("-ceed", "CEED resource specifier", 127 NULL, ceed_resource, ceed_resource, 128 sizeof(ceed_resource), NULL); CHKERRQ(ierr); 129 read_mesh = PETSC_FALSE; 130 ierr = PetscOptionsString("-mesh", "Read mesh from file", NULL, 131 filename, filename, sizeof(filename), &read_mesh); 132 CHKERRQ(ierr); 133 simplex = PETSC_FALSE; 134 ierr = PetscOptionsBool("-simplex", "Use simplices, or tensor product cells", 135 NULL, simplex, &simplex, NULL); CHKERRQ(ierr); 136 ierr = PetscOptionsEnd(); CHKERRQ(ierr); 137 138 // Setup DM 139 if (read_mesh) { 140 ierr = DMPlexCreateFromFile(PETSC_COMM_WORLD, filename, NULL, PETSC_TRUE, 141 &dm); 142 CHKERRQ(ierr); 143 } else { 144 // Create the mesh as a 0-refined sphere. This will create a cubic surface, 145 // not a box, and will snap to the unit sphere upon refinement. 146 ierr = DMPlexCreateSphereMesh(PETSC_COMM_WORLD, topo_dim, simplex, 1., &dm); 147 CHKERRQ(ierr); 148 // Set the object name 149 ierr = PetscObjectSetName((PetscObject)dm, "Sphere"); CHKERRQ(ierr); 150 // Refine DMPlex with uniform refinement using runtime option -dm_refine 151 ierr = DMPlexSetRefinementUniform(dm, PETSC_TRUE); CHKERRQ(ierr); 152 } 153 ierr = DMSetFromOptions(dm); CHKERRQ(ierr); 154 // View DMPlex via runtime option 155 ierr = DMViewFromOptions(dm, NULL, "-dm_view"); CHKERRQ(ierr); 156 157 // Create DM 158 ierr = SetupDMByDegree(dm, degree, num_comp_u, topo_dim, false, 159 (BCFunction)NULL); 160 CHKERRQ(ierr); 161 162 // Create vectors 163 ierr = DMCreateGlobalVector(dm, &X); CHKERRQ(ierr); 164 ierr = VecGetLocalSize(X, &l_size); CHKERRQ(ierr); 165 ierr = VecGetSize(X, &g_size); CHKERRQ(ierr); 166 ierr = DMCreateLocalVector(dm, &X_loc); CHKERRQ(ierr); 167 ierr = VecGetSize(X_loc, &xl_size); CHKERRQ(ierr); 168 ierr = VecDuplicate(X, &rhs); CHKERRQ(ierr); 169 170 // Operator 171 ierr = PetscMalloc1(1, &user_O); CHKERRQ(ierr); 172 ierr = MatCreateShell(comm, l_size, l_size, g_size, g_size, 173 user_O, &mat_O); CHKERRQ(ierr); 174 ierr = MatShellSetOperation(mat_O, MATOP_MULT, 175 (void(*)(void))MatMult_Ceed); CHKERRQ(ierr); 176 177 // Set up libCEED 178 CeedInit(ceed_resource, &ceed); 179 CeedMemType mem_type_backend; 180 CeedGetPreferredMemType(ceed, &mem_type_backend); 181 182 ierr = DMGetVecType(dm, &vec_type); CHKERRQ(ierr); 183 if (!vec_type) { // Not yet set by user -dm_vec_type 184 switch (mem_type_backend) { 185 case CEED_MEM_HOST: vec_type = VECSTANDARD; break; 186 case CEED_MEM_DEVICE: { 187 const char *resolved; 188 CeedGetResource(ceed, &resolved); 189 if (strstr(resolved, "/gpu/cuda")) vec_type = VECCUDA; 190 else if (strstr(resolved, "/gpu/hip/occa")) 191 vec_type = VECSTANDARD; // https://github.com/CEED/libCEED/issues/678 192 else if (strstr(resolved, "/gpu/hip")) vec_type = VECHIP; 193 else vec_type = VECSTANDARD; 194 } 195 } 196 ierr = DMSetVecType(dm, vec_type); CHKERRQ(ierr); 197 } 198 199 // Print summary 200 if (!test_mode) { 201 PetscInt P = degree + 1, Q = P + q_extra; 202 const char *used_resource; 203 CeedGetResource(ceed, &used_resource); 204 ierr = PetscPrintf(comm, 205 "\n-- CEED Benchmark Problem %d on the Sphere -- libCEED + PETSc --\n" 206 " libCEED:\n" 207 " libCEED Backend : %s\n" 208 " libCEED Backend MemType : %s\n" 209 " Mesh:\n" 210 " Number of 1D Basis Nodes (p) : %d\n" 211 " Number of 1D Quadrature Points (q) : %d\n" 212 " Global nodes : %D\n", 213 bp_choice+1, ceed_resource, CeedMemTypes[mem_type_backend], P, Q, 214 g_size/num_comp_u); CHKERRQ(ierr); 215 } 216 217 // Create RHS vector 218 ierr = VecDuplicate(X_loc, &rhs_loc); CHKERRQ(ierr); 219 ierr = VecZeroEntries(rhs_loc); CHKERRQ(ierr); 220 ierr = VecGetArrayAndMemType(rhs_loc, &r, &mem_type); CHKERRQ(ierr); 221 CeedVectorCreate(ceed, xl_size, &rhs_ceed); 222 CeedVectorSetArray(rhs_ceed, MemTypeP2C(mem_type), CEED_USE_POINTER, r); 223 224 // Setup libCEED's objects 225 ierr = PetscMalloc1(1, &ceed_data); CHKERRQ(ierr); 226 ierr = SetupLibceedByDegree(dm, ceed, degree, topo_dim, q_extra, num_comp_x, 227 num_comp_u, g_size, xl_size, bp_options[bp_choice], 228 ceed_data, true, rhs_ceed, &target); CHKERRQ(ierr); 229 230 // Gather RHS 231 CeedVectorTakeArray(rhs_ceed, MemTypeP2C(mem_type), NULL); 232 ierr = VecRestoreArrayAndMemType(rhs_loc, &r); CHKERRQ(ierr); 233 ierr = VecZeroEntries(rhs); CHKERRQ(ierr); 234 ierr = DMLocalToGlobal(dm, rhs_loc, ADD_VALUES, rhs); CHKERRQ(ierr); 235 CeedVectorDestroy(&rhs_ceed); 236 237 // Create the error Q-function 238 CeedQFunctionCreateInterior(ceed, 1, bp_options[bp_choice].error, 239 bp_options[bp_choice].error_loc, &qf_error); 240 CeedQFunctionAddInput(qf_error, "u", num_comp_u, CEED_EVAL_INTERP); 241 CeedQFunctionAddInput(qf_error, "true_soln", num_comp_u, CEED_EVAL_NONE); 242 CeedQFunctionAddOutput(qf_error, "error", num_comp_u, CEED_EVAL_NONE); 243 244 // Create the error operator 245 CeedOperatorCreate(ceed, qf_error, NULL, NULL, &op_error); 246 CeedOperatorSetField(op_error, "u", ceed_data->elem_restr_u, 247 ceed_data->basis_u, CEED_VECTOR_ACTIVE); 248 CeedOperatorSetField(op_error, "true_soln", ceed_data->elem_restr_u_i, 249 CEED_BASIS_COLLOCATED, target); 250 CeedOperatorSetField(op_error, "error", ceed_data->elem_restr_u_i, 251 CEED_BASIS_COLLOCATED, CEED_VECTOR_ACTIVE); 252 253 // Set up Mat 254 user_O->comm = comm; 255 user_O->dm = dm; 256 user_O->X_loc = X_loc; 257 ierr = VecDuplicate(X_loc, &user_O->Y_loc); CHKERRQ(ierr); 258 user_O->x_ceed = ceed_data->x_ceed; 259 user_O->y_ceed = ceed_data->y_ceed; 260 user_O->op = ceed_data->op_apply; 261 user_O->ceed = ceed; 262 263 // Setup solver 264 ierr = KSPCreate(comm, &ksp); CHKERRQ(ierr); 265 { 266 PC pc; 267 ierr = KSPGetPC(ksp, &pc); CHKERRQ(ierr); 268 if (bp_choice == CEED_BP1 || bp_choice == CEED_BP2) { 269 ierr = PCSetType(pc, PCJACOBI); CHKERRQ(ierr); 270 ierr = PCJacobiSetType(pc, PC_JACOBI_ROWSUM); CHKERRQ(ierr); 271 } else { 272 ierr = PCSetType(pc, PCNONE); CHKERRQ(ierr); 273 MatNullSpace nullspace; 274 275 ierr = MatNullSpaceCreate(PETSC_COMM_WORLD, PETSC_TRUE, 0, 0, &nullspace); 276 CHKERRQ(ierr); 277 ierr = MatSetNullSpace(mat_O, nullspace); CHKERRQ(ierr); 278 ierr = MatNullSpaceDestroy(&nullspace); CHKERRQ(ierr); 279 } 280 ierr = KSPSetType(ksp, KSPCG); CHKERRQ(ierr); 281 ierr = KSPSetNormType(ksp, KSP_NORM_NATURAL); CHKERRQ(ierr); 282 ierr = KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, 283 PETSC_DEFAULT); CHKERRQ(ierr); 284 } 285 ierr = KSPSetFromOptions(ksp); CHKERRQ(ierr); 286 ierr = KSPSetOperators(ksp, mat_O, mat_O); CHKERRQ(ierr); 287 288 // First run, if benchmarking 289 if (benchmark_mode) { 290 ierr = KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, 1); 291 CHKERRQ(ierr); 292 my_rt_start = MPI_Wtime(); 293 ierr = KSPSolve(ksp, rhs, X); CHKERRQ(ierr); 294 my_rt = MPI_Wtime() - my_rt_start; 295 ierr = MPI_Allreduce(MPI_IN_PLACE, &my_rt, 1, MPI_DOUBLE, MPI_MIN, comm); 296 CHKERRQ(ierr); 297 // Set maxits based on first iteration timing 298 if (my_rt > 0.02) { 299 ierr = KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, 5); 300 CHKERRQ(ierr); 301 } else { 302 ierr = KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, 20); 303 CHKERRQ(ierr); 304 } 305 } 306 307 // Timed solve 308 ierr = VecZeroEntries(X); CHKERRQ(ierr); 309 ierr = PetscBarrier((PetscObject)ksp); CHKERRQ(ierr); 310 311 // -- Performance logging 312 ierr = PetscLogStageRegister("Solve Stage", &solve_stage); CHKERRQ(ierr); 313 ierr = PetscLogStagePush(solve_stage); CHKERRQ(ierr); 314 315 // -- Solve 316 my_rt_start = MPI_Wtime(); 317 ierr = KSPSolve(ksp, rhs, X); CHKERRQ(ierr); 318 my_rt = MPI_Wtime() - my_rt_start; 319 320 // -- Performance logging 321 ierr = PetscLogStagePop(); 322 323 // Output results 324 { 325 KSPType ksp_type; 326 KSPConvergedReason reason; 327 PetscReal rnorm; 328 PetscInt its; 329 ierr = KSPGetType(ksp, &ksp_type); CHKERRQ(ierr); 330 ierr = KSPGetConvergedReason(ksp, &reason); CHKERRQ(ierr); 331 ierr = KSPGetIterationNumber(ksp, &its); CHKERRQ(ierr); 332 ierr = KSPGetResidualNorm(ksp, &rnorm); CHKERRQ(ierr); 333 if (!test_mode || reason < 0 || rnorm > 1e-8) { 334 ierr = PetscPrintf(comm, 335 " KSP:\n" 336 " KSP Type : %s\n" 337 " KSP Convergence : %s\n" 338 " Total KSP Iterations : %D\n" 339 " Final rnorm : %e\n", 340 ksp_type, KSPConvergedReasons[reason], its, 341 (double)rnorm); CHKERRQ(ierr); 342 } 343 if (!test_mode) { 344 ierr = PetscPrintf(comm," Performance:\n"); CHKERRQ(ierr); 345 } 346 { 347 PetscReal max_error; 348 ierr = ComputeErrorMax(user_O, op_error, X, target, &max_error); 349 CHKERRQ(ierr); 350 PetscReal tol = 5e-4; 351 if (!test_mode || max_error > tol) { 352 ierr = MPI_Allreduce(&my_rt, &rt_min, 1, MPI_DOUBLE, MPI_MIN, comm); 353 CHKERRQ(ierr); 354 ierr = MPI_Allreduce(&my_rt, &rt_max, 1, MPI_DOUBLE, MPI_MAX, comm); 355 CHKERRQ(ierr); 356 ierr = PetscPrintf(comm, 357 " Pointwise Error (max) : %e\n" 358 " CG Solve Time : %g (%g) sec\n", 359 (double)max_error, rt_max, rt_min); CHKERRQ(ierr); 360 } 361 } 362 if (benchmark_mode && (!test_mode)) { 363 ierr = PetscPrintf(comm, 364 " DoFs/Sec in CG : %g (%g) million\n", 365 1e-6*g_size*its/rt_max, 1e-6*g_size*its/rt_min); CHKERRQ(ierr); 366 } 367 } 368 369 // Output solution 370 if (write_solution) { 371 PetscViewer vtk_viewer_soln; 372 373 ierr = PetscViewerCreate(comm, &vtk_viewer_soln); CHKERRQ(ierr); 374 ierr = PetscViewerSetType(vtk_viewer_soln, PETSCVIEWERVTK); CHKERRQ(ierr); 375 ierr = PetscViewerFileSetName(vtk_viewer_soln, "solution.vtu"); CHKERRQ(ierr); 376 ierr = VecView(X, vtk_viewer_soln); CHKERRQ(ierr); 377 ierr = PetscViewerDestroy(&vtk_viewer_soln); CHKERRQ(ierr); 378 } 379 380 // Cleanup 381 ierr = VecDestroy(&X); CHKERRQ(ierr); 382 ierr = VecDestroy(&X_loc); CHKERRQ(ierr); 383 ierr = VecDestroy(&user_O->Y_loc); CHKERRQ(ierr); 384 ierr = MatDestroy(&mat_O); CHKERRQ(ierr); 385 ierr = PetscFree(user_O); CHKERRQ(ierr); 386 ierr = CeedDataDestroy(0, ceed_data); CHKERRQ(ierr); 387 ierr = DMDestroy(&dm); CHKERRQ(ierr); 388 389 ierr = VecDestroy(&rhs); CHKERRQ(ierr); 390 ierr = VecDestroy(&rhs_loc); CHKERRQ(ierr); 391 ierr = KSPDestroy(&ksp); CHKERRQ(ierr); 392 CeedVectorDestroy(&target); 393 CeedQFunctionDestroy(&qf_error); 394 CeedOperatorDestroy(&op_error); 395 CeedDestroy(&ceed); 396 return PetscFinalize(); 397 } 398