xref: /libCEED/examples/petsc/multigrid.c (revision a171b6ef1192a8f5f3031330339aaab2ded4c8fb)
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 3-6 with Multigrid
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 multigrid [PETSC_DIR=</path/to/petsc>] [CEED_DIR=</path/to/libceed>]
17 //
18 // Sample runs:
19 //
20 //     multigrid -problem bp3
21 //     multigrid -problem bp4
22 //     multigrid -problem bp5 -ceed /cpu/self
23 //     multigrid -problem bp6 -ceed /gpu/cuda
24 //
25 //TESTARGS(name="BP3, hex elements") -ceed {ceed_resource} -test -problem bp3 -degree 3
26 //TESTARGS(name="BP3, tet elements") -ceed {ceed_resource} -test -problem bp3 -degree 3 -simplex
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 <ceed.h>
33 #include <petsc.h>
34 #include <petscdmplex.h>
35 #include <petscksp.h>
36 #include <petscsys.h>
37 #include <stdbool.h>
38 #include <string.h>
39 
40 #include "bps.h"
41 #include "include/bpsproblemdata.h"
42 #include "include/libceedsetup.h"
43 #include "include/matops.h"
44 #include "include/petscutils.h"
45 #include "include/petscversion.h"
46 #include "include/structs.h"
47 
48 int main(int argc, char **argv) {
49   MPI_Comm comm;
50   char     filename[PETSC_MAX_PATH_LEN], ceed_resource[PETSC_MAX_PATH_LEN] = "/cpu/self";
51   double   my_rt_start, my_rt, rt_min, rt_max;
52   PetscInt degree = 3, q_extra, *l_size, *xl_size, *g_size, dim = 3, fine_level, mesh_elem[3] = {3, 3, 3}, num_comp_u = 1, num_levels = degree,
53            *level_degrees;
54   PetscScalar           eps = 1.0;
55   PetscBool             test_mode, benchmark_mode, read_mesh, write_solution, simplex;
56   PetscLogStage         solve_stage;
57   PetscLogEvent         assemble_event;
58   DM                   *dm, dm_orig;
59   KSP                   ksp;
60   PC                    pc;
61   Mat                  *mat_O, *mat_pr, mat_coarse;
62   Vec                  *X, *X_loc, *mult, rhs, rhs_loc;
63   PetscMemType          mem_type;
64   OperatorApplyContext *op_apply_ctx, op_error_ctx;
65   ProlongRestrContext  *pr_restr_ctx;
66   Ceed                  ceed;
67   CeedData             *ceed_data;
68   CeedVector            rhs_ceed, target;
69   CeedQFunction         qf_error;
70   CeedOperator          op_error;
71   BPType                bp_choice;
72   CoarsenType           coarsen;
73 
74   PetscCall(PetscInitialize(&argc, &argv, NULL, help));
75   comm = PETSC_COMM_WORLD;
76 
77   // Parse command line options
78   PetscOptionsBegin(comm, NULL, "CEED BPs in PETSc", NULL);
79   bp_choice = CEED_BP3;
80   PetscCall(PetscOptionsEnum("-problem", "CEED benchmark problem to solve", NULL, bp_types, (PetscEnum)bp_choice, (PetscEnum *)&bp_choice, NULL));
81   num_comp_u = bp_options[bp_choice].num_comp_u;
82   test_mode  = PETSC_FALSE;
83   PetscCall(PetscOptionsBool("-test", "Testing mode (do not print unless error is large)", NULL, test_mode, &test_mode, NULL));
84   benchmark_mode = PETSC_FALSE;
85   PetscCall(PetscOptionsBool("-benchmark", "Benchmarking mode (prints benchmark statistics)", NULL, benchmark_mode, &benchmark_mode, NULL));
86   write_solution = PETSC_FALSE;
87   PetscCall(PetscOptionsBool("-write_solution", "Write solution for visualization", NULL, write_solution, &write_solution, NULL));
88   simplex = PETSC_FALSE;
89   PetscCall(PetscOptionsBool("-simplex", "Element topology (default:hex)", NULL, simplex, &simplex, NULL));
90   if ((bp_choice == CEED_BP5 || bp_choice == CEED_BP6) && (simplex)) {
91     SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "BP5/6 is not supported with simplex");
92   }
93   PetscCall(PetscOptionsScalar("-eps", "Epsilon parameter for Kershaw mesh transformation", NULL, eps, &eps, NULL));
94   if (eps > 1 || eps <= 0) SETERRQ(PETSC_COMM_WORLD, PETSC_ERR_ARG_OUTOFRANGE, "-eps %g must be (0,1]", (double)PetscRealPart(eps));
95   degree = test_mode ? 3 : 2;
96   PetscCall(PetscOptionsInt("-degree", "Polynomial degree of tensor product basis", NULL, degree, &degree, NULL));
97   if (degree < 1) SETERRQ(PETSC_COMM_WORLD, PETSC_ERR_ARG_OUTOFRANGE, "-degree %" PetscInt_FMT " must be at least 1", degree);
98   q_extra = bp_options[bp_choice].q_extra;
99   PetscCall(PetscOptionsInt("-q_extra", "Number of extra quadrature points", NULL, q_extra, &q_extra, NULL));
100   PetscCall(PetscOptionsString("-ceed", "CEED resource specifier", NULL, ceed_resource, ceed_resource, sizeof(ceed_resource), NULL));
101   coarsen = COARSEN_UNIFORM;
102   PetscCall(PetscOptionsEnum("-coarsen", "Coarsening strategy to use", NULL, coarsen_types, (PetscEnum)coarsen, (PetscEnum *)&coarsen, NULL));
103   read_mesh = PETSC_FALSE;
104   PetscCall(PetscOptionsString("-mesh", "Read mesh from file", NULL, filename, filename, sizeof(filename), &read_mesh));
105   if (!read_mesh) {
106     PetscInt tmp = dim;
107     PetscCall(PetscOptionsIntArray("-cells", "Number of cells per dimension", NULL, mesh_elem, &tmp, NULL));
108   }
109   PetscOptionsEnd();
110 
111   // Set up libCEED
112   CeedInit(ceed_resource, &ceed);
113   CeedMemType mem_type_backend;
114   CeedGetPreferredMemType(ceed, &mem_type_backend);
115 
116   // Setup DM
117   if (read_mesh) {
118     PetscCall(DMPlexCreateFromFile(PETSC_COMM_WORLD, filename, NULL, PETSC_TRUE, &dm_orig));
119   } else {
120     PetscCall(DMPlexCreateBoxMesh(PETSC_COMM_WORLD, dim, simplex, mesh_elem, NULL, NULL, NULL, PETSC_TRUE, &dm_orig));
121   }
122 
123   VecType vec_type;
124   switch (mem_type_backend) {
125     case CEED_MEM_HOST:
126       vec_type = VECSTANDARD;
127       break;
128     case CEED_MEM_DEVICE: {
129       const char *resolved;
130       CeedGetResource(ceed, &resolved);
131       if (strstr(resolved, "/gpu/cuda")) vec_type = VECCUDA;
132       else if (strstr(resolved, "/gpu/hip/occa")) vec_type = VECSTANDARD;  // https://github.com/CEED/libCEED/issues/678
133       else if (strstr(resolved, "/gpu/hip")) vec_type = VECHIP;
134       else vec_type = VECSTANDARD;
135     }
136   }
137   PetscCall(DMSetVecType(dm_orig, vec_type));
138   PetscCall(DMSetFromOptions(dm_orig));
139   PetscCall(DMViewFromOptions(dm_orig, NULL, "-dm_view"));
140 
141   // Apply Kershaw mesh transformation
142   PetscCall(Kershaw(dm_orig, eps));
143 
144   // Allocate arrays for PETSc objects for each level
145   switch (coarsen) {
146     case COARSEN_UNIFORM:
147       num_levels = degree;
148       break;
149     case COARSEN_LOGARITHMIC:
150       num_levels = ceil(log(degree) / log(2)) + 1;
151       break;
152   }
153   PetscCall(PetscMalloc1(num_levels, &level_degrees));
154   fine_level = num_levels - 1;
155 
156   switch (coarsen) {
157     case COARSEN_UNIFORM:
158       for (int i = 0; i < num_levels; i++) level_degrees[i] = i + 1;
159       break;
160     case COARSEN_LOGARITHMIC:
161       for (int i = 0; i < num_levels - 1; i++) level_degrees[i] = pow(2, i);
162       level_degrees[fine_level] = degree;
163       break;
164   }
165   PetscCall(PetscMalloc1(num_levels, &dm));
166   PetscCall(PetscMalloc1(num_levels, &X));
167   PetscCall(PetscMalloc1(num_levels, &X_loc));
168   PetscCall(PetscMalloc1(num_levels, &mult));
169   PetscCall(PetscMalloc1(num_levels, &op_apply_ctx));
170   PetscCall(PetscMalloc1(num_levels, &pr_restr_ctx));
171   PetscCall(PetscMalloc1(num_levels, &mat_O));
172   PetscCall(PetscMalloc1(num_levels, &mat_pr));
173   PetscCall(PetscMalloc1(num_levels, &l_size));
174   PetscCall(PetscMalloc1(num_levels, &xl_size));
175   PetscCall(PetscMalloc1(num_levels, &g_size));
176 
177   PetscInt c_start, c_end;
178   PetscCall(DMPlexGetHeightStratum(dm_orig, 0, &c_start, &c_end));
179   DMPolytopeType cell_type;
180   PetscCall(DMPlexGetCellType(dm_orig, c_start, &cell_type));
181   CeedElemTopology elem_topo = ElemTopologyP2C(cell_type);
182 
183   // Setup DM and Operator Mat Shells for each level
184   for (CeedInt i = 0; i < num_levels; i++) {
185     // Create DM
186     PetscCall(DMClone(dm_orig, &dm[i]));
187     PetscCall(DMGetVecType(dm_orig, &vec_type));
188     PetscCall(DMSetVecType(dm[i], vec_type));
189     PetscInt dim;
190     PetscCall(DMGetDimension(dm[i], &dim));
191     PetscCall(SetupDMByDegree(dm[i], level_degrees[fine_level], q_extra, num_comp_u, dim, bp_options[bp_choice].enforce_bc));
192 
193     // Create vectors
194     PetscCall(DMCreateGlobalVector(dm[i], &X[i]));
195     PetscCall(VecGetLocalSize(X[i], &l_size[i]));
196     PetscCall(VecGetSize(X[i], &g_size[i]));
197     PetscCall(DMCreateLocalVector(dm[i], &X_loc[i]));
198     PetscCall(VecGetSize(X_loc[i], &xl_size[i]));
199 
200     // Operator
201     PetscCall(PetscMalloc1(1, &op_apply_ctx[i]));
202     PetscCall(PetscMalloc1(1, &op_error_ctx));
203     PetscCall(MatCreateShell(comm, l_size[i], l_size[i], g_size[i], g_size[i], op_apply_ctx[i], &mat_O[i]));
204     PetscCall(MatShellSetOperation(mat_O[i], MATOP_MULT, (void (*)(void))MatMult_Ceed));
205     PetscCall(MatShellSetOperation(mat_O[i], MATOP_GET_DIAGONAL, (void (*)(void))MatGetDiag));
206     PetscCall(MatShellSetVecType(mat_O[i], vec_type));
207 
208     // Level transfers
209     if (i > 0) {
210       // Interp
211       PetscCall(PetscMalloc1(1, &pr_restr_ctx[i]));
212       PetscCall(MatCreateShell(comm, l_size[i], l_size[i - 1], g_size[i], g_size[i - 1], pr_restr_ctx[i], &mat_pr[i]));
213       PetscCall(MatShellSetOperation(mat_pr[i], MATOP_MULT, (void (*)(void))MatMult_Prolong));
214       PetscCall(MatShellSetOperation(mat_pr[i], MATOP_MULT_TRANSPOSE, (void (*)(void))MatMult_Restrict));
215       PetscCall(MatShellSetVecType(mat_pr[i], vec_type));
216     }
217   }
218   PetscCall(VecDuplicate(X[fine_level], &rhs));
219 
220   // Print global grid information
221   if (!test_mode) {
222     PetscInt P = degree + 1, Q = P + q_extra;
223 
224     const char *used_resource;
225     CeedGetResource(ceed, &used_resource);
226 
227     PetscCall(VecGetType(X[0], &vec_type));
228 
229     PetscCall(PetscPrintf(comm,
230                           "\n-- CEED Benchmark Problem %" CeedInt_FMT " -- libCEED + PETSc + PCMG --\n"
231                           "  PETSc:\n"
232                           "    PETSc Vec Type                          : %s\n"
233                           "  libCEED:\n"
234                           "    libCEED Backend                         : %s\n"
235                           "    libCEED Backend MemType                 : %s\n"
236                           "  Mesh:\n"
237                           "    Solution Order (P)                      : %" CeedInt_FMT "\n"
238                           "    Quadrature  Order (Q)                   : %" CeedInt_FMT "\n"
239                           "    Additional quadrature points (q_extra)  : %" CeedInt_FMT "\n"
240                           "    Global Nodes                            : %" PetscInt_FMT "\n"
241                           "    Owned Nodes                             : %" PetscInt_FMT "\n"
242                           "    DoF per node                            : %" PetscInt_FMT "\n"
243                           "    Element topology                        : %s\n"
244                           "  Multigrid:\n"
245                           "    Number of Levels                        : %" CeedInt_FMT "\n",
246                           bp_choice + 1, vec_type, used_resource, CeedMemTypes[mem_type_backend], P, Q, q_extra, g_size[fine_level] / num_comp_u,
247                           l_size[fine_level] / num_comp_u, num_comp_u, CeedElemTopologies[elem_topo], num_levels));
248   }
249 
250   // Create RHS vector
251   PetscCall(VecDuplicate(X_loc[fine_level], &rhs_loc));
252   PetscCall(VecZeroEntries(rhs_loc));
253   CeedVectorCreate(ceed, xl_size[fine_level], &rhs_ceed);
254   PetscCall(VecP2C(rhs_loc, &mem_type, rhs_ceed));
255 
256   // Set up libCEED operators on each level
257   PetscCall(PetscMalloc1(num_levels, &ceed_data));
258   for (PetscInt i = 0; i < num_levels; i++) {
259     // Print level information
260     if (!test_mode && (i == 0 || i == fine_level)) {
261       PetscCall(PetscPrintf(comm,
262                             "    Level %" PetscInt_FMT " (%s):\n"
263                             "      Solution Order (P)                    : %" CeedInt_FMT "\n"
264                             "      Global Nodes                          : %" PetscInt_FMT "\n"
265                             "      Owned Nodes                           : %" PetscInt_FMT "\n",
266                             i, (i ? "fine" : "coarse"), level_degrees[i] + 1, g_size[i] / num_comp_u, l_size[i] / num_comp_u));
267     }
268     PetscCall(PetscMalloc1(1, &ceed_data[i]));
269     PetscCall(SetupLibceedByDegree(dm[i], ceed, level_degrees[i], dim, q_extra, dim, num_comp_u, g_size[i], xl_size[i], bp_options[bp_choice],
270                                    ceed_data[i], i == (fine_level), rhs_ceed, &target));
271   }
272 
273   // Gather RHS
274   PetscCall(VecC2P(rhs_ceed, mem_type, rhs_loc));
275   PetscCall(VecZeroEntries(rhs));
276   PetscCall(DMLocalToGlobal(dm[fine_level], rhs_loc, ADD_VALUES, rhs));
277   CeedVectorDestroy(&rhs_ceed);
278 
279   // Create the error QFunction
280   CeedQFunctionCreateInterior(ceed, 1, bp_options[bp_choice].error, bp_options[bp_choice].error_loc, &qf_error);
281   CeedQFunctionAddInput(qf_error, "u", num_comp_u, CEED_EVAL_INTERP);
282   CeedQFunctionAddInput(qf_error, "true_soln", num_comp_u, CEED_EVAL_NONE);
283   CeedQFunctionAddInput(qf_error, "qdata", ceed_data[fine_level]->q_data_size, CEED_EVAL_NONE);
284   CeedQFunctionAddOutput(qf_error, "error", num_comp_u, CEED_EVAL_INTERP);
285 
286   // Create the error operator
287   CeedOperatorCreate(ceed, qf_error, CEED_QFUNCTION_NONE, CEED_QFUNCTION_NONE, &op_error);
288   CeedOperatorSetField(op_error, "u", ceed_data[fine_level]->elem_restr_u, ceed_data[fine_level]->basis_u, CEED_VECTOR_ACTIVE);
289   CeedOperatorSetField(op_error, "true_soln", ceed_data[fine_level]->elem_restr_u_i, CEED_BASIS_NONE, target);
290   CeedOperatorSetField(op_error, "qdata", ceed_data[fine_level]->elem_restr_qd_i, CEED_BASIS_NONE, ceed_data[fine_level]->q_data);
291   CeedOperatorSetField(op_error, "error", ceed_data[fine_level]->elem_restr_u, ceed_data[fine_level]->basis_u, CEED_VECTOR_ACTIVE);
292 
293   // Calculate multiplicity
294   for (int i = 0; i < num_levels; i++) {
295     PetscMemType mem_type;
296 
297     // CEED vector
298     PetscCall(VecZeroEntries(X_loc[i]));
299     PetscCall(VecP2C(X_loc[i], &mem_type, ceed_data[i]->x_ceed));
300 
301     // Multiplicity
302     CeedElemRestrictionGetMultiplicity(ceed_data[i]->elem_restr_u, ceed_data[i]->x_ceed);
303     CeedVectorSyncArray(ceed_data[i]->x_ceed, CEED_MEM_HOST);
304 
305     // Restore vector
306     PetscCall(VecC2P(ceed_data[i]->x_ceed, mem_type, X_loc[i]));
307 
308     // Creat mult vector
309     PetscCall(VecDuplicate(X_loc[i], &mult[i]));
310 
311     // Local-to-global
312     PetscCall(VecZeroEntries(X[i]));
313     PetscCall(DMLocalToGlobal(dm[i], X_loc[i], ADD_VALUES, X[i]));
314     PetscCall(VecZeroEntries(X_loc[i]));
315 
316     // Global-to-local
317     PetscCall(DMGlobalToLocal(dm[i], X[i], INSERT_VALUES, mult[i]));
318     PetscCall(VecZeroEntries(X[i]));
319 
320     // Multiplicity scaling
321     PetscCall(VecReciprocal(mult[i]));
322   }
323 
324   // Set up Mat
325   for (int i = 0; i < num_levels; i++) {
326     // Set up apply operator context
327     PetscCall(SetupApplyOperatorCtx(comm, dm[i], ceed, ceed_data[i], X_loc[i], op_apply_ctx[i]));
328 
329     if (i > 0) {
330       // Prolongation/Restriction Operator
331       PetscCall(CeedLevelTransferSetup(dm[i - 1], ceed, i, num_comp_u, ceed_data, bp_options[bp_choice], mult[i]));
332       pr_restr_ctx[i]->comm        = comm;
333       pr_restr_ctx[i]->dmf         = dm[i];
334       pr_restr_ctx[i]->dmc         = dm[i - 1];
335       pr_restr_ctx[i]->loc_vec_c   = X_loc[i - 1];
336       pr_restr_ctx[i]->loc_vec_f   = op_apply_ctx[i]->Y_loc;
337       pr_restr_ctx[i]->mult_vec    = mult[i];
338       pr_restr_ctx[i]->ceed_vec_c  = op_apply_ctx[i - 1]->x_ceed;
339       pr_restr_ctx[i]->ceed_vec_f  = op_apply_ctx[i]->y_ceed;
340       pr_restr_ctx[i]->op_prolong  = ceed_data[i]->op_prolong;
341       pr_restr_ctx[i]->op_restrict = ceed_data[i]->op_restrict;
342       pr_restr_ctx[i]->ceed        = ceed;
343     }
344   }
345 
346   // Assemble coarse grid Jacobian for AMG (or other sparse matrix) solve
347   PetscCall(DMCreateMatrix(dm[0], &mat_coarse));
348 
349   PetscCall(PetscLogEventRegister("AssembleMatrix", MAT_CLASSID, &assemble_event));
350   {
351     // Assemble matrix analytically
352     PetscCount num_entries;
353     CeedInt   *rows, *cols;
354     CeedVector coo_values;
355     CeedOperatorLinearAssembleSymbolic(op_apply_ctx[0]->op, &num_entries, &rows, &cols);
356     ISLocalToGlobalMapping ltog_row, ltog_col;
357     PetscCall(MatGetLocalToGlobalMapping(mat_coarse, &ltog_row, &ltog_col));
358     PetscCall(ISLocalToGlobalMappingApply(ltog_row, num_entries, rows, rows));
359     PetscCall(ISLocalToGlobalMappingApply(ltog_col, num_entries, cols, cols));
360     PetscCall(MatSetPreallocationCOO(mat_coarse, num_entries, rows, cols));
361     free(rows);
362     free(cols);
363     CeedVectorCreate(ceed, num_entries, &coo_values);
364     PetscCall(PetscLogEventBegin(assemble_event, mat_coarse, 0, 0, 0));
365     CeedOperatorLinearAssemble(op_apply_ctx[0]->op, coo_values);
366     const CeedScalar *values;
367     CeedVectorGetArrayRead(coo_values, CEED_MEM_HOST, &values);
368     PetscCall(MatSetValuesCOO(mat_coarse, values, ADD_VALUES));
369     CeedVectorRestoreArrayRead(coo_values, &values);
370     PetscCall(PetscLogEventEnd(assemble_event, mat_coarse, 0, 0, 0));
371     CeedVectorDestroy(&coo_values);
372   }
373 
374   // Set up KSP
375   PetscCall(KSPCreate(comm, &ksp));
376   {
377     PetscCall(KSPSetType(ksp, KSPCG));
378     PetscCall(KSPSetNormType(ksp, KSP_NORM_NATURAL));
379     PetscCall(KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, PETSC_DEFAULT));
380   }
381   PetscCall(KSPSetFromOptions(ksp));
382   PetscCall(KSPSetOperators(ksp, mat_O[fine_level], mat_O[fine_level]));
383 
384   // Set up PCMG
385   PetscCall(KSPGetPC(ksp, &pc));
386   PCMGCycleType pcmg_cycle_type = PC_MG_CYCLE_V;
387   {
388     PetscCall(PCSetType(pc, PCMG));
389 
390     // PCMG levels
391     PetscCall(PCMGSetLevels(pc, num_levels, NULL));
392     for (int i = 0; i < num_levels; i++) {
393       // Smoother
394       KSP smoother;
395       PC  smoother_pc;
396       PetscCall(PCMGGetSmoother(pc, i, &smoother));
397       PetscCall(KSPSetType(smoother, KSPCHEBYSHEV));
398       PetscCall(KSPChebyshevEstEigSet(smoother, 0, 0.1, 0, 1.1));
399       PetscCall(KSPChebyshevEstEigSetUseNoisy(smoother, PETSC_TRUE));
400       PetscCall(KSPSetOperators(smoother, mat_O[i], mat_O[i]));
401       PetscCall(KSPGetPC(smoother, &smoother_pc));
402       PetscCall(PCSetType(smoother_pc, PCJACOBI));
403       PetscCall(PCJacobiSetType(smoother_pc, PC_JACOBI_DIAGONAL));
404 
405       // Work vector
406       if (i < num_levels - 1) {
407         PetscCall(PCMGSetX(pc, i, X[i]));
408       }
409 
410       // Level transfers
411       if (i > 0) {
412         // Interpolation
413         PetscCall(PCMGSetInterpolation(pc, i, mat_pr[i]));
414       }
415 
416       // Coarse solve
417       KSP coarse;
418       PC  coarse_pc;
419       PetscCall(PCMGGetCoarseSolve(pc, &coarse));
420       PetscCall(KSPSetType(coarse, KSPPREONLY));
421       PetscCall(KSPSetOperators(coarse, mat_coarse, mat_coarse));
422 
423       PetscCall(KSPGetPC(coarse, &coarse_pc));
424       PetscCall(PCSetType(coarse_pc, PCGAMG));
425 
426       PetscCall(KSPSetOptionsPrefix(coarse, "coarse_"));
427       PetscCall(PCSetOptionsPrefix(coarse_pc, "coarse_"));
428       PetscCall(KSPSetFromOptions(coarse));
429       PetscCall(PCSetFromOptions(coarse_pc));
430     }
431 
432     // PCMG options
433     PetscCall(PCMGSetType(pc, PC_MG_MULTIPLICATIVE));
434     PetscCall(PCMGSetNumberSmooth(pc, 3));
435     PetscCall(PCMGSetCycleType(pc, pcmg_cycle_type));
436   }
437 
438   // First run, if benchmarking
439   if (benchmark_mode) {
440     PetscCall(KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, 1));
441     PetscCall(VecZeroEntries(X[fine_level]));
442     my_rt_start = MPI_Wtime();
443     PetscCall(KSPSolve(ksp, rhs, X[fine_level]));
444     my_rt = MPI_Wtime() - my_rt_start;
445     PetscCall(MPI_Allreduce(MPI_IN_PLACE, &my_rt, 1, MPI_DOUBLE, MPI_MIN, comm));
446     // Set maxits based on first iteration timing
447     if (my_rt > 0.02) {
448       PetscCall(KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, 5));
449     } else {
450       PetscCall(KSPSetTolerances(ksp, 1e-10, PETSC_DEFAULT, PETSC_DEFAULT, 20));
451     }
452   }
453 
454   // Timed solve
455   PetscCall(VecZeroEntries(X[fine_level]));
456   PetscCall(PetscBarrier((PetscObject)ksp));
457 
458   // -- Performance logging
459   PetscCall(PetscLogStageRegister("Solve Stage", &solve_stage));
460   PetscCall(PetscLogStagePush(solve_stage));
461 
462   // -- Solve
463   my_rt_start = MPI_Wtime();
464   PetscCall(KSPSolve(ksp, rhs, X[fine_level]));
465   my_rt = MPI_Wtime() - my_rt_start;
466 
467   // -- Performance logging
468   PetscCall(PetscLogStagePop());
469 
470   // Output results
471   {
472     KSPType            ksp_type;
473     PCMGType           pcmg_type;
474     KSPConvergedReason reason;
475     PetscReal          rnorm;
476     PetscInt           its;
477     PetscCall(KSPGetType(ksp, &ksp_type));
478     PetscCall(KSPGetConvergedReason(ksp, &reason));
479     PetscCall(KSPGetIterationNumber(ksp, &its));
480     PetscCall(KSPGetResidualNorm(ksp, &rnorm));
481     PetscCall(PCMGGetType(pc, &pcmg_type));
482     if (!test_mode || reason < 0 || rnorm > 1e-8) {
483       PetscCall(PetscPrintf(comm,
484                             "  KSP:\n"
485                             "    KSP Type                                : %s\n"
486                             "    KSP Convergence                         : %s\n"
487                             "    Total KSP Iterations                    : %" PetscInt_FMT "\n"
488                             "    Final rnorm                             : %e\n",
489                             ksp_type, KSPConvergedReasons[reason], its, (double)rnorm));
490       PetscCall(PetscPrintf(comm,
491                             "  PCMG:\n"
492                             "    PCMG Type                               : %s\n"
493                             "    PCMG Cycle Type                         : %s\n",
494                             PCMGTypes[pcmg_type], PCMGCycleTypes[pcmg_cycle_type]));
495     }
496     if (!test_mode) {
497       PetscCall(PetscPrintf(comm, "  Performance:\n"));
498     }
499     {
500       // Set up error operator context
501       PetscCall(SetupErrorOperatorCtx(comm, dm[fine_level], ceed, ceed_data[fine_level], X_loc[fine_level], op_error, op_error_ctx));
502       PetscScalar l2_error;
503       PetscCall(ComputeL2Error(X[fine_level], &l2_error, op_error_ctx));
504       PetscReal tol = 5e-2;
505       if (!test_mode || l2_error > tol) {
506         PetscCall(MPI_Allreduce(&my_rt, &rt_min, 1, MPI_DOUBLE, MPI_MIN, comm));
507         PetscCall(MPI_Allreduce(&my_rt, &rt_max, 1, MPI_DOUBLE, MPI_MAX, comm));
508         PetscCall(PetscPrintf(comm,
509                               "    L2 Error                                : %e\n"
510                               "    CG Solve Time                           : %g (%g) sec\n",
511                               (double)l2_error, rt_max, rt_min));
512       }
513     }
514     if (benchmark_mode && (!test_mode)) {
515       PetscCall(PetscPrintf(comm, "    DoFs/Sec in CG                            : %g (%g) million\n", 1e-6 * g_size[fine_level] * its / rt_max,
516                             1e-6 * g_size[fine_level] * its / rt_min));
517     }
518   }
519 
520   if (write_solution) {
521     PetscViewer vtk_viewer_soln;
522 
523     PetscCall(PetscViewerCreate(comm, &vtk_viewer_soln));
524     PetscCall(PetscViewerSetType(vtk_viewer_soln, PETSCVIEWERVTK));
525     PetscCall(PetscViewerFileSetName(vtk_viewer_soln, "solution.vtu"));
526     PetscCall(VecView(X[fine_level], vtk_viewer_soln));
527     PetscCall(PetscViewerDestroy(&vtk_viewer_soln));
528   }
529 
530   // Cleanup
531   for (int i = 0; i < num_levels; i++) {
532     PetscCall(VecDestroy(&X[i]));
533     PetscCall(VecDestroy(&X_loc[i]));
534     PetscCall(VecDestroy(&mult[i]));
535     PetscCall(VecDestroy(&op_apply_ctx[i]->Y_loc));
536     PetscCall(MatDestroy(&mat_O[i]));
537     PetscCall(PetscFree(op_apply_ctx[i]));
538     if (i > 0) {
539       PetscCall(MatDestroy(&mat_pr[i]));
540       PetscCall(PetscFree(pr_restr_ctx[i]));
541     }
542     PetscCall(CeedDataDestroy(i, ceed_data[i]));
543     PetscCall(DMDestroy(&dm[i]));
544   }
545   PetscCall(PetscFree(level_degrees));
546   PetscCall(PetscFree(dm));
547   PetscCall(PetscFree(X));
548   PetscCall(PetscFree(X_loc));
549   PetscCall(VecDestroy(&op_error_ctx->Y_loc));
550   PetscCall(PetscFree(mult));
551   PetscCall(PetscFree(mat_O));
552   PetscCall(PetscFree(mat_pr));
553   PetscCall(PetscFree(ceed_data));
554   PetscCall(PetscFree(op_apply_ctx));
555   PetscCall(PetscFree(op_error_ctx));
556   PetscCall(PetscFree(pr_restr_ctx));
557   PetscCall(PetscFree(l_size));
558   PetscCall(PetscFree(xl_size));
559   PetscCall(PetscFree(g_size));
560   PetscCall(VecDestroy(&rhs));
561   PetscCall(VecDestroy(&rhs_loc));
562   PetscCall(MatDestroy(&mat_coarse));
563   PetscCall(KSPDestroy(&ksp));
564   PetscCall(DMDestroy(&dm_orig));
565   CeedVectorDestroy(&target);
566   CeedQFunctionDestroy(&qf_error);
567   CeedOperatorDestroy(&op_error);
568   CeedDestroy(&ceed);
569   return PetscFinalize();
570 }
571