Actual source code: ex3.c

  1: static char help[] = "Reduced formulation of the mother problem of PDE-constrained optimisation.\n";

  3: /*F
  4:   We solve the mother problem

  6:   min 1/2 || y - y_d ||^2_2 + \alpha/2 || u ||^2_2

  8:   subject to

 10:           - \laplace y = u          on \Omega
 11:                      y = 0          on \Gamma

 13:   where u is in L^2 and y is in H^1_0.

 15:   We formulate the reduced problem solely in terms of the control
 16:   by using the state equation to express y in terms of u, and then
 17:   apply LMVM/BLMVM to the resulting reduced problem.

 19:   Mesh independence is achieved by configuring the Riesz map for the control
 20:   space.

 22:   Example meshes where the Riesz map is crucial can be downloaded from the
 23:   http://people.maths.ox.ac.uk/~farrellp/meshes.tar.gz

 25:   Contributed by: Patrick Farrell <patrick.farrell@maths.ox.ac.uk>

 27:   Run with e.g.:
 28:   ./ex3 -laplace_ksp_type cg -laplace_pc_type hypre -tao_blmvm_mat_lmvm_J0_ksp_type cg -tao_blmvm_mat_lmvm_J0_pc_type gamg -laplace_ksp_monitor_true_residual -tao_monitor -petscspace_degree 1 -tao_converged_reason -tao_gatol 1.0e-9 -dm_view hdf5:solution.h5 -sol_view hdf5:solution.h5::append -use_riesz 1 -f $DATAFILESPATH/meshes/mesh-1.h5

 30:   and visualise in paraview with ../../../../petsc_gen_xdmf.py solution.h5.

 32:   Toggle the Riesz map (-use_riesz 0) to see the difference setting the Riesz maps makes.

 34:   TODO: broken for parallel runs
 35: F*/

 37: #include <petsc.h>
 38: #include <petscfe.h>
 39: #include <petscviewerhdf5.h>

 41: typedef struct {
 42:   DM           dm;
 43:   Mat          mass;
 44:   Vec          data;
 45:   Vec          state;
 46:   Vec          tmp1;
 47:   Vec          tmp2;
 48:   Vec          adjoint;
 49:   Mat          laplace;
 50:   KSP          ksp_laplace;
 51:   PetscInt     num_bc_dofs;
 52:   PetscInt    *bc_indices;
 53:   PetscScalar *bc_values;
 54:   PetscBool    use_riesz;
 55: } AppCtx;

 57: static PetscErrorCode CreateMesh(MPI_Comm comm, AppCtx *user, DM *dm)
 58: {
 59:   PetscBool flg;
 60:   char      filename[2048];

 62:   PetscFunctionBeginUser;
 63:   filename[0]     = '\0';
 64:   user->use_riesz = PETSC_TRUE;

 66:   PetscOptionsBegin(comm, "", "Poisson mother problem options", "DMPLEX");
 67:   PetscCall(PetscOptionsBool("-use_riesz", "Use the Riesz map to achieve mesh independence", "ex3.c", user->use_riesz, &user->use_riesz, NULL));
 68:   PetscCall(PetscOptionsString("-f", "filename to read", "ex3.c", filename, filename, sizeof(filename), &flg));
 69:   PetscOptionsEnd();

 71:   if (!flg) {
 72:     PetscCall(DMCreate(comm, dm));
 73:     PetscCall(DMSetType(*dm, DMPLEX));
 74:   } else {
 75:     /* TODO Eliminate this in favor of DMLoad() in new code */
 76: #if defined(PETSC_HAVE_HDF5)
 77:     const PetscInt vertices_per_cell = 3;
 78:     PetscViewer    viewer;
 79:     Vec            coordinates;
 80:     Vec            topology;
 81:     PetscInt       dim = 2, numCells;
 82:     PetscInt       numVertices;
 83:     PetscScalar   *coords;
 84:     PetscScalar   *topo_f;
 85:     PetscInt      *cells;
 86:     DMLabel        label;

 88:     /* Read in FEniCS HDF5 output */
 89:     PetscCall(PetscViewerHDF5Open(comm, filename, FILE_MODE_READ, &viewer));

 91:     /* create Vecs to read in the data from H5 */
 92:     PetscCall(VecCreate(comm, &coordinates));
 93:     PetscCall(PetscObjectSetName((PetscObject)coordinates, "coordinates"));
 94:     PetscCall(VecSetBlockSize(coordinates, dim));
 95:     PetscCall(VecCreate(comm, &topology));
 96:     PetscCall(PetscObjectSetName((PetscObject)topology, "topology"));
 97:     PetscCall(VecSetBlockSize(topology, vertices_per_cell));

 99:     /* navigate to the right group */
100:     PetscCall(PetscViewerHDF5PushGroup(viewer, "/Mesh/mesh"));

102:     /* Read the Vecs */
103:     PetscCall(VecLoad(coordinates, viewer));
104:     PetscCall(VecLoad(topology, viewer));

106:     /* do some ugly calculations */
107:     PetscCall(VecGetSize(topology, &numCells));
108:     numCells = numCells / vertices_per_cell;
109:     PetscCall(VecGetSize(coordinates, &numVertices));
110:     numVertices = numVertices / dim;

112:     PetscCall(VecGetArray(coordinates, &coords));
113:     PetscCall(VecGetArray(topology, &topo_f));
114:     /* and now we have to convert the double representation to integers to pass over, argh */
115:     PetscCall(PetscMalloc1(numCells * vertices_per_cell, &cells));
116:     for (PetscInt j = 0; j < numCells * vertices_per_cell; j++) cells[j] = (PetscInt)topo_f[j];

118:     /* Now create the DM */
119:     PetscCall(DMPlexCreateFromCellListPetsc(comm, dim, numCells, numVertices, vertices_per_cell, PETSC_TRUE, cells, dim, coords, dm));
120:     /* Check for flipped first cell */
121:     {
122:       PetscReal v0[3], J[9], invJ[9], detJ;

124:       PetscCall(DMPlexComputeCellGeometryFEM(*dm, 0, NULL, v0, J, invJ, &detJ));
125:       if (detJ < 0) {
126:         PetscCall(DMPlexOrientPoint(*dm, 0, -1));
127:         PetscCall(DMPlexComputeCellGeometryFEM(*dm, 0, NULL, v0, J, invJ, &detJ));
128:         PetscCheck(detJ >= 0, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Something is wrong");
129:       }
130:     }
131:     PetscCall(DMPlexOrient(*dm));
132:     PetscCall(DMCreateLabel(*dm, "marker"));
133:     PetscCall(DMGetLabel(*dm, "marker", &label));
134:     PetscCall(DMPlexMarkBoundaryFaces(*dm, 1, label));
135:     PetscCall(DMPlexLabelComplete(*dm, label));

137:     PetscCall(PetscViewerDestroy(&viewer));
138:     PetscCall(VecRestoreArray(coordinates, &coords));
139:     PetscCall(VecRestoreArray(topology, &topo_f));
140:     PetscCall(PetscFree(cells));
141:     PetscCall(VecDestroy(&coordinates));
142:     PetscCall(VecDestroy(&topology));
143: #else
144:     SETERRQ(PETSC_COMM_WORLD, PETSC_ERR_SUP, "Reconfigure PETSc with --download-hdf5");
145: #endif
146:   }
147:   PetscCall(DMSetFromOptions(*dm));
148:   PetscCall(DMViewFromOptions(*dm, NULL, "-dm_view"));
149:   PetscFunctionReturn(PETSC_SUCCESS);
150: }

152: void mass_kernel(PetscInt dim, PetscInt Nf, PetscInt NfAux, const PetscInt uOff[], const PetscInt uOff_x[], const PetscScalar u[], const PetscScalar u_t[], const PetscScalar u_x[], const PetscInt aOff[], const PetscInt aOff_x[], const PetscScalar a[], const PetscScalar a_t[], const PetscScalar a_x[], PetscReal t, PetscReal u_tShift, const PetscReal x[], PetscInt numConstants, const PetscScalar constants[], PetscScalar g0[])
153: {
154:   g0[0] = 1.0;
155: }

157: void laplace_kernel(PetscInt dim, PetscInt Nf, PetscInt NfAux, const PetscInt uOff[], const PetscInt uOff_x[], const PetscScalar u[], const PetscScalar u_t[], const PetscScalar u_x[], const PetscInt aOff[], const PetscInt aOff_x[], const PetscScalar a[], const PetscScalar a_t[], const PetscScalar a_x[], PetscReal t, PetscReal u_tShift, const PetscReal x[], PetscInt numConstants, const PetscScalar constants[], PetscScalar g3[])
158: {
159:   PetscInt d;
160:   for (d = 0; d < dim; ++d) g3[d * dim + d] = 1.0;
161: }

163: /* data we seek to match */
164: PetscErrorCode data_kernel(PetscInt dim, PetscReal time, const PetscReal x[], PetscInt Nf, PetscScalar *y, PetscCtx ctx)
165: {
166:   *y = 1.0 / (2 * PETSC_PI * PETSC_PI) * PetscSinReal(PETSC_PI * x[0]) * PetscSinReal(PETSC_PI * x[1]);
167:   /* the associated control is sin(pi*x[0])*sin(pi*x[1]) */
168:   return PETSC_SUCCESS;
169: }
170: PetscErrorCode zero(PetscInt dim, PetscReal time, const PetscReal x[], PetscInt Nf, PetscScalar *u, PetscCtx ctx)
171: {
172:   *u = 0.0;
173:   return PETSC_SUCCESS;
174: }

176: PetscErrorCode CreateCtx(DM dm, AppCtx *user)
177: {
178:   DM              dm_mass;
179:   DM              dm_laplace;
180:   PetscDS         prob_mass;
181:   PetscDS         prob_laplace;
182:   PetscFE         fe;
183:   DMLabel         label;
184:   PetscSection    section;
185:   PetscInt        n, k, p, d;
186:   PetscInt        dof, off;
187:   IS              is;
188:   const PetscInt *points;
189:   const PetscInt  dim = 2;
190:   PetscErrorCode (**wtf)(PetscInt dim, PetscReal time, const PetscReal x[], PetscInt Nf, PetscScalar *u, PetscCtx ctx);

192:   PetscFunctionBeginUser;
193:   /* make the data we seek to match */
194:   PetscCall(PetscFECreateDefault(PetscObjectComm((PetscObject)dm), dim, 1, PETSC_TRUE, NULL, 4, &fe));

196:   PetscCall(DMSetField(dm, 0, NULL, (PetscObject)fe));
197:   PetscCall(DMCreateDS(dm));
198:   PetscCall(DMCreateGlobalVector(dm, &user->data));

200:   /* ugh, this is hideous */
201:   /* y_d = interpolate(Expression("sin(x[0]) + .."), V) */
202:   PetscCall(PetscMalloc(1 * sizeof(void (*)(const PetscReal[], PetscScalar *, void *)), &wtf));
203:   wtf[0] = data_kernel;
204:   PetscCall(DMProjectFunction(dm, 0.0, wtf, NULL, INSERT_VALUES, user->data));
205:   PetscCall(PetscFree(wtf));

207:   /* assemble(inner(u, v)*dx), almost */
208:   PetscCall(DMClone(dm, &dm_mass));
209:   PetscCall(DMCopyDisc(dm, dm_mass));
210:   PetscCall(DMSetNumFields(dm_mass, 1));
211:   PetscCall(DMPlexCopyCoordinates(dm, dm_mass)); /* why do I have to do this separately? */
212:   PetscCall(DMGetDS(dm_mass, &prob_mass));
213:   PetscCall(PetscDSSetJacobian(prob_mass, 0, 0, mass_kernel, NULL, NULL, NULL));
214:   PetscCall(PetscDSSetDiscretization(prob_mass, 0, (PetscObject)fe));
215:   PetscCall(DMCreateMatrix(dm_mass, &user->mass));
216:   PetscCall(DMPlexSNESComputeJacobianFEM(dm_mass, user->data, user->mass, user->mass, NULL));
217:   PetscCall(MatSetOption(user->mass, MAT_SYMMETRIC, PETSC_TRUE));
218:   PetscCall(DMDestroy(&dm_mass));

220:   /* inner(grad(u), grad(v))*dx with homogeneous Dirichlet boundary conditions */
221:   PetscCall(DMClone(dm, &dm_laplace));
222:   PetscCall(DMCopyDisc(dm, dm_laplace));
223:   PetscCall(DMSetNumFields(dm_laplace, 1));
224:   PetscCall(DMPlexCopyCoordinates(dm, dm_laplace));
225:   PetscCall(DMGetDS(dm_laplace, &prob_laplace));
226:   PetscCall(PetscDSSetJacobian(prob_laplace, 0, 0, NULL, NULL, NULL, laplace_kernel));
227:   PetscCall(PetscDSSetDiscretization(prob_laplace, 0, (PetscObject)fe));
228:   PetscCall(DMCreateMatrix(dm_laplace, &user->laplace));
229:   PetscCall(DMPlexSNESComputeJacobianFEM(dm_laplace, user->data, user->laplace, user->laplace, NULL));

231:   PetscCall(DMGetLabel(dm_laplace, "marker", &label));
232:   PetscCall(DMGetLocalSection(dm_laplace, &section));
233:   PetscCall(DMLabelGetStratumSize(label, 1, &n));
234:   PetscCall(DMLabelGetStratumIS(label, 1, &is));
235:   PetscCall(ISGetIndices(is, &points));
236:   user->num_bc_dofs = 0;
237:   for (p = 0; p < n; ++p) {
238:     PetscCall(PetscSectionGetDof(section, points[p], &dof));
239:     user->num_bc_dofs += dof;
240:   }
241:   PetscCall(PetscMalloc1(user->num_bc_dofs, &user->bc_indices));
242:   for (p = 0, k = 0; p < n; ++p) {
243:     PetscCall(PetscSectionGetDof(section, points[p], &dof));
244:     PetscCall(PetscSectionGetOffset(section, points[p], &off));
245:     for (d = 0; d < dof; ++d) user->bc_indices[k++] = off + d;
246:   }
247:   PetscCall(ISRestoreIndices(is, &points));
248:   PetscCall(ISDestroy(&is));
249:   PetscCall(DMDestroy(&dm_laplace));

251:   /* This is how I handle boundary conditions. I can't figure out how to get
252:      plex to play with the way I want to impose the BCs. This loses symmetry,
253:      but not in a disastrous way. If someone can improve it, please do! */
254:   PetscCall(MatZeroRows(user->laplace, user->num_bc_dofs, user->bc_indices, 1.0, NULL, NULL));
255:   PetscCall(PetscCalloc1(user->num_bc_dofs, &user->bc_values));

257:   /* also create the KSP for solving the Laplace system */
258:   PetscCall(KSPCreate(PETSC_COMM_WORLD, &user->ksp_laplace));
259:   PetscCall(KSPSetOperators(user->ksp_laplace, user->laplace, user->laplace));
260:   PetscCall(KSPSetOptionsPrefix(user->ksp_laplace, "laplace_"));
261:   PetscCall(KSPSetFromOptions(user->ksp_laplace));

263:   /* A bit of setting up the user context */
264:   user->dm = dm;
265:   PetscCall(VecDuplicate(user->data, &user->state));
266:   PetscCall(VecDuplicate(user->data, &user->adjoint));
267:   PetscCall(VecDuplicate(user->data, &user->tmp1));
268:   PetscCall(VecDuplicate(user->data, &user->tmp2));

270:   PetscCall(PetscFEDestroy(&fe));
271:   PetscFunctionReturn(PETSC_SUCCESS);
272: }

274: PetscErrorCode DestroyCtx(AppCtx *user)
275: {
276:   PetscFunctionBeginUser;
277:   PetscCall(MatDestroy(&user->mass));
278:   PetscCall(MatDestroy(&user->laplace));
279:   PetscCall(KSPDestroy(&user->ksp_laplace));
280:   PetscCall(VecDestroy(&user->data));
281:   PetscCall(VecDestroy(&user->state));
282:   PetscCall(VecDestroy(&user->adjoint));
283:   PetscCall(VecDestroy(&user->tmp1));
284:   PetscCall(VecDestroy(&user->tmp2));
285:   PetscCall(PetscFree(user->bc_indices));
286:   PetscCall(PetscFree(user->bc_values));
287:   PetscFunctionReturn(PETSC_SUCCESS);
288: }

290: PetscErrorCode ReducedFunctionGradient(Tao tao, Vec u, PetscReal *func, Vec g, void *userv)
291: {
292:   AppCtx         *user  = (AppCtx *)userv;
293:   const PetscReal alpha = 1.0e-6; /* regularisation parameter */
294:   PetscReal       inner;

296:   PetscFunctionBeginUser;
297:   PetscCall(MatMult(user->mass, u, user->tmp1));
298:   PetscCall(VecDot(u, user->tmp1, &inner)); /* regularisation contribution to */
299:   *func = alpha * 0.5 * inner;              /* the functional                 */

301:   PetscCall(VecSet(g, 0.0));
302:   PetscCall(VecAXPY(g, alpha, user->tmp1)); /* regularisation contribution to the gradient */

304:   /* Now compute the forward state. */
305:   PetscCall(VecSetValues(user->tmp1, user->num_bc_dofs, user->bc_indices, user->bc_values, INSERT_VALUES));
306:   PetscCall(VecAssemblyBegin(user->tmp1));
307:   PetscCall(VecAssemblyEnd(user->tmp1));
308:   PetscCall(KSPSolve(user->ksp_laplace, user->tmp1, user->state)); /* forward solve */

310:   /* Now compute the adjoint state also. */
311:   PetscCall(VecCopy(user->state, user->tmp1));
312:   PetscCall(VecAXPY(user->tmp1, -1.0, user->data));
313:   PetscCall(MatMult(user->mass, user->tmp1, user->tmp2));
314:   PetscCall(VecDot(user->tmp1, user->tmp2, &inner)); /* misfit contribution to */
315:   *func += 0.5 * inner;                              /* the functional         */

317:   PetscCall(VecSetValues(user->tmp2, user->num_bc_dofs, user->bc_indices, user->bc_values, INSERT_VALUES));
318:   PetscCall(VecAssemblyBegin(user->tmp2));
319:   PetscCall(VecAssemblyEnd(user->tmp2));
320:   PetscCall(KSPSolve(user->ksp_laplace, user->tmp2, user->adjoint)); /* adjoint solve */

322:   /* And bring it home with the gradient. */
323:   PetscCall(MatMult(user->mass, user->adjoint, user->tmp1));
324:   PetscCall(VecAXPY(g, 1.0, user->tmp1)); /* adjoint contribution to the gradient */
325:   PetscFunctionReturn(PETSC_SUCCESS);
326: }

328: int main(int argc, char **argv)
329: {
330:   DM     dm;
331:   Tao    tao;
332:   Vec    u, lb, ub;
333:   AppCtx user;

335:   PetscFunctionBeginUser;
336:   PetscCall(PetscInitialize(&argc, &argv, NULL, help));
337:   PetscCall(CreateMesh(PETSC_COMM_WORLD, &user, &dm));
338:   PetscCall(CreateCtx(dm, &user));

340:   PetscCall(DMCreateGlobalVector(dm, &u));
341:   PetscCall(VecDuplicate(u, &lb));
342:   PetscCall(VecDuplicate(u, &ub));
343:   PetscCall(VecSet(ub, 0.8)); /* a nontrivial upper bound */

345:   PetscCall(TaoCreate(PETSC_COMM_WORLD, &tao));
346:   PetscCall(TaoSetSolution(tao, u));
347:   PetscCall(TaoSetObjectiveAndGradient(tao, NULL, ReducedFunctionGradient, &user));
348:   PetscCall(TaoSetVariableBounds(tao, lb, ub));
349:   PetscCall(TaoSetType(tao, TAOBLMVM));
350:   PetscCall(TaoSetFromOptions(tao));

352:   if (user.use_riesz) {
353:     PetscCall(TaoLMVMSetH0(tao, user.mass)); /* crucial for mesh independence */
354:     PetscCall(TaoSetGradientNorm(tao, user.mass));
355:   }

357:   PetscCall(TaoSolve(tao));

359:   PetscCall(DMViewFromOptions(dm, NULL, "-dm_view"));
360:   PetscCall(VecViewFromOptions(u, NULL, "-sol_view"));

362:   PetscCall(TaoDestroy(&tao));
363:   PetscCall(DMDestroy(&dm));
364:   PetscCall(VecDestroy(&u));
365:   PetscCall(VecDestroy(&lb));
366:   PetscCall(VecDestroy(&ub));
367:   PetscCall(DestroyCtx(&user));
368:   PetscCall(PetscFinalize());
369:   return 0;
370: }

372: /*TEST

374:     build:
375:       requires: !complex !single

377:     test:
378:       requires: hdf5 double datafilespath !defined(PETSC_USE_64BIT_INDICES) hypre !cuda
379:       args: -laplace_ksp_type cg -laplace_pc_type hypre -laplace_ksp_monitor_true_residual -laplace_ksp_max_it 5 -tao_blmvm_mat_lmvm_J0_ksp_type cg -tao_blmvm_mat_lmvm_J0_ksp_rtol 1e-5 -tao_blmvm_mat_lmvm_J0_pc_type gamg -tao_monitor -petscspace_degree 1 -tao_converged_reason -tao_gatol 1.0e-6 -dm_view hdf5:solution.h5 -sol_view hdf5:solution.h5::append -use_riesz 1 -f $DATAFILESPATH/meshes/mesh-1.h5
380:       filter: sed -e "s/-nan/nan/g"

382:     test:
383:       suffix: guess_pod
384:       requires: double triangle
385:       args: -laplace_ksp_type cg -laplace_pc_type gamg -laplace_ksp_max_it 8 -laplace_pc_gamg_esteig_ksp_type cg -laplace_pc_gamg_esteig_ksp_max_it 5 -laplace_mg_levels_ksp_chebyshev_esteig 0,0.25,0,1.0 -tao_blmvm_mat_lmvm_J0_ksp_type cg -tao_blmvm_mat_lmvm_J0_ksp_rtol 1e-5 -tao_blmvm_mat_lmvm_J0_pc_type gamg -tao_blmvm_mat_lmvm_J0_pc_gamg_esteig_ksp_type cg -tao_blmvm_mat_lmvm_J0_pc_gamg_esteig_ksp_max_it 3 -tao_monitor -petscspace_degree 1 -tao_converged_reason -dm_refine 0 -laplace_ksp_guess_type pod -tao_gatol 1e-6 -laplace_pc_gamg_aggressive_coarsening 0
386:       filter: sed -e "s/-nan/nan/g" -e "s/-NaN/nan/g" -e "s/NaN/nan/g" -e "s/CONVERGED_RTOL iterations 9/CONVERGED_RTOL iterations 8/g" -e "s/CONVERGED_RTOL iterations 4/CONVERGED_RTOL iterations 3/g"

388: TEST*/