void NonlinearMatrixSolver<Scalar>::solve_linear_system()
{
// store the previous solution to previous_sln_vector.
memcpy(this->previous_sln_vector, this->sln_vector, sizeof(Scalar)*this->problem_size);
// Solve, if the solver is iterative, give him the initial guess.
this->linear_matrix_solver->solve(this->use_initial_guess_for_iterative_solvers ? this->sln_vector : nullptr);
// 1. store the solution.
double solution_change_norm = this->update_solution_return_change_norm(this->linear_matrix_solver->get_sln_vector());
// 2. store the solution change.
this->get_parameter_value(this->p_solution_change_norms).push_back(solution_change_norm);
// 3. store the solution norm.
this->get_parameter_value(this->p_solution_norms).push_back(get_l2_norm(this->sln_vector, this->problem_size));
}
bool NonlinearMatrixSolver<Scalar>::do_initial_step_return_finished()
{
// Store the initial norm.
this->get_parameter_value(this->p_solution_norms).push_back(get_l2_norm(this->sln_vector, this->problem_size));
// Assemble the system.
if (this->jacobian_reusable && this->constant_jacobian)
{
this->linear_matrix_solver->set_reuse_scheme(HERMES_REUSE_MATRIX_STRUCTURE_COMPLETELY);
this->assemble_residual(false);
this->get_parameter_value(this->p_iterations_with_recalculated_jacobian).push_back(false);
}
else
{
this->linear_matrix_solver->set_reuse_scheme(HERMES_CREATE_STRUCTURE_FROM_SCRATCH);
this->assemble(false, false);
this->jacobian_reusable = true;
this->get_parameter_value(this->p_iterations_with_recalculated_jacobian).push_back(true);
}
// Get the residual norm and act upon it.
double residual_norm = this->calculate_residual_norm();
this->info("\n\tNonlinearSolver: initial residual norm: %g", residual_norm);
this->get_parameter_value(this->p_residual_norms).push_back(residual_norm);
this->solve_linear_system();
if (this->handle_convergence_state_return_finished(this->get_convergence_state()))
return true;
if (this->on_initial_step_end() == false)
{
this->info("\tNonlinearSolver: aborted.");
this->finalize_solving();
return true;
}
return false;
}
int main() {
// Create space, set Dirichlet BC, enumerate basis functions.
Space* space = new Space(A, B, NELEM, DIR_BC_LEFT, DIR_BC_RIGHT, P_INIT, NEQ);
info("ndof = %d", Space::get_num_dofs(space));
// Initialize the weak formulation.
WeakForm wf(4);
wf.add_matrix_form(0, 0, jacobian_1_1);
wf.add_matrix_form(0, 2, jacobian_1_3);
wf.add_matrix_form(0, 3, jacobian_1_4);
wf.add_matrix_form(1, 1, jacobian_2_2);
wf.add_matrix_form(1, 2, jacobian_2_3);
wf.add_matrix_form(1, 3, jacobian_2_4);
wf.add_matrix_form(2, 0, jacobian_3_1);
wf.add_matrix_form(2, 1, jacobian_3_2);
wf.add_matrix_form(2, 2, jacobian_3_3);
wf.add_matrix_form(3, 0, jacobian_4_1);
wf.add_matrix_form(3, 1, jacobian_4_2);
wf.add_matrix_form(3, 3, jacobian_4_4);
wf.add_vector_form(0, residual_1);
wf.add_vector_form(1, residual_2);
wf.add_vector_form(2, residual_3);
wf.add_vector_form(3, residual_4);
wf.add_matrix_form_surf(0, 0, jacobian_surf_right_U_Re, BOUNDARY_RIGHT);
wf.add_matrix_form_surf(0, 2, jacobian_surf_right_U_Im, BOUNDARY_RIGHT);
wf.add_matrix_form_surf(1, 1, jacobian_surf_right_I_Re, BOUNDARY_RIGHT);
wf.add_matrix_form_surf(1, 3, jacobian_surf_right_I_Im, BOUNDARY_RIGHT);
// Initialize the FE problem.
bool is_linear = false;
DiscreteProblem *dp = new DiscreteProblem(&wf, space, is_linear);
// Set zero initial condition.
double *coeff_vec = new double[Space::get_num_dofs(space)];
set_zero(coeff_vec, Space::get_num_dofs(space));
// Set up the solver, matrix, and rhs according to the solver selection.
SparseMatrix* matrix = create_matrix(matrix_solver);
Vector* rhs = create_vector(matrix_solver);
Solver* solver = create_linear_solver(matrix_solver, matrix, rhs);
int it = 1;
while (1) {
// Assemble the Jacobian matrix and residual vector.
dp->assemble(coeff_vec, matrix, rhs);
// Calculate the l2-norm of residual vector.
double res_l2_norm = get_l2_norm(rhs);
// Info for user.
info("---- Newton iter %d, ndof %d, res. l2 norm %g", it, Space::get_num_dofs(space), res_l2_norm);
// If l2 norm of the residual vector is within tolerance, then quit.
// NOTE: at least one full iteration forced
// here because sometimes the initial
// residual on fine mesh is too small.
if(res_l2_norm < NEWTON_TOL && it > 1) break;
// Multiply the residual vector with -1 since the matrix
// equation reads J(Y^n) \deltaY^{n+1} = -F(Y^n).
for(int i = 0; i < Space::get_num_dofs(space); i++) rhs->set(i, -rhs->get(i));
// Solve the linear system.
if(!solver->solve())
error ("Matrix solver failed.\n");
// Add \deltaY^{n+1} to Y^n.
for (int i = 0; i < Space::get_num_dofs(space); i++) coeff_vec[i] += solver->get_solution()[i];
// If the maximum number of iteration has been reached, then quit.
if (it >= NEWTON_MAX_ITER) error ("Newton method did not converge.");
it++;
}
// Plot the solution.
Linearizer l(space);
l.plot_solution("solution.gp");
info("Done.");
return 0;
}
void compute_trajectory(Space *space, DiscreteProblem *dp)
{
info("alpha = (%g, %g, %g, %g), zeta = (%g, %g, %g, %g)",
alpha_ctrl[0], alpha_ctrl[1],
alpha_ctrl[2], alpha_ctrl[3], zeta_ctrl[0],
zeta_ctrl[1], zeta_ctrl[2], zeta_ctrl[3]);
// Newton's loop.
// Fill vector coeff_vec using dof and coeffs arrays in elements.
double *coeff_vec = new double[Space::get_num_dofs(space)];
get_coeff_vector(space, coeff_vec);
// Set up the solver, matrix, and rhs according to the solver selection.
SparseMatrix* matrix = create_matrix(matrix_solver);
Vector* rhs = create_vector(matrix_solver);
Solver* solver = create_linear_solver(matrix_solver, matrix, rhs);
int it = 1;
while (1)
{
// Obtain the number of degrees of freedom.
int ndof = Space::get_num_dofs(space);
// Assemble the Jacobian matrix and residual vector.
dp->assemble(coeff_vec, matrix, rhs);
// Calculate the l2-norm of residual vector.
double res_l2_norm = get_l2_norm(rhs);
// Info for user.
info("---- Newton iter %d, ndof %d, res. l2 norm %g", it, Space::get_num_dofs(space), res_l2_norm);
// If l2 norm of the residual vector is within tolerance, then quit.
// NOTE: at least one full iteration forced
// here because sometimes the initial
// residual on fine mesh is too small.
if(res_l2_norm < NEWTON_TOL && it > 1) break;
// Multiply the residual vector with -1 since the matrix
// equation reads J(Y^n) \deltaY^{n+1} = -F(Y^n).
for(int i=0; i<ndof; i++) rhs->set(i, -rhs->get(i));
// Solve the linear system.
if(!solver->solve())
error ("Matrix solver failed.\n");
// Add \deltaY^{n+1} to Y^n.
for (int i = 0; i < ndof; i++) coeff_vec[i] += solver->get_solution()[i];
// If the maximum number of iteration has been reached, then quit.
if (it >= NEWTON_MAX_ITER) error ("Newton method did not converge.");
// Copy coefficients from vector y to elements.
set_coeff_vector(coeff_vec, space);
it++;
}
// Cleanup.
delete matrix;
delete rhs;
delete solver;
delete [] coeff_vec;
}
int main()
{
// Create space.
// Transform input data to the format used by the "Space" constructor.
SpaceData *md = new SpaceData(verbose);
Space* space = new Space(md->N_macroel, md->interfaces, md->poly_orders, md->material_markers, md->subdivisions, N_GRP, N_SLN);
delete md;
// Enumerate basis functions, info for user.
int ndof = Space::get_num_dofs(space);
info("ndof: %d", ndof);
// Plot the space.
space->plot("space.gp");
// Initial approximation of the dominant eigenvalue.
double K_EFF = 1.0;
// Initial approximation of the dominant eigenvector.
double init_val = 1.0;
for (int g = 0; g < N_GRP; g++)
{
set_vertex_dofs_constant(space, init_val, g);
space->set_bc_right_dirichlet(g, flux_right_surf[g]);
}
// Initialize the weak formulation.
WeakForm wf(2);
wf.add_matrix_form(0, 0, jacobian_fuel_0_0, NULL, fuel);
wf.add_matrix_form(0, 0, jacobian_water_0_0, NULL, water);
wf.add_matrix_form(0, 1, jacobian_fuel_0_1, NULL, fuel);
wf.add_matrix_form(0, 1, jacobian_water_0_1, NULL, water);
wf.add_matrix_form(1, 0, jacobian_fuel_1_0, NULL, fuel);
wf.add_matrix_form(1, 0, jacobian_water_1_0, NULL, water);
wf.add_matrix_form(1, 1, jacobian_fuel_1_1, NULL, fuel);
wf.add_matrix_form(1, 1, jacobian_water_1_1, NULL, water);
wf.add_vector_form(0, residual_fuel_0, NULL, fuel);
wf.add_vector_form(0, residual_water_0, NULL, water);
wf.add_vector_form(1, residual_fuel_1, NULL, fuel);
wf.add_vector_form(1, residual_water_1, NULL, water);
wf.add_vector_form_surf(0, residual_surf_left_0, BOUNDARY_LEFT);
wf.add_vector_form_surf(1, residual_surf_left_1, BOUNDARY_LEFT);
// Initialize the FE problem.
bool is_linear = false;
DiscreteProblem *dp = new DiscreteProblem(&wf, space, is_linear);
Linearizer l(space);
char solution_file[32];
// Source iteration
int i;
int current_solution = 0, previous_solution = 1;
double K_EFF_old;
for (i = 0; i < Max_SI; i++)
{
// Plot the critical (i.e. steady-state) flux in the actual iteration.
sprintf(solution_file, "solution_%d.gp", i);
l.plot_solution(solution_file);
// Store the previous solution (used at the right-hand side).
for (int g = 0; g < N_GRP; g++)
copy_dofs(current_solution, previous_solution, space, g);
// Obtain the number of degrees of freedom.
int ndof = Space::get_num_dofs(space);
// Fill vector coeff_vec using dof and coeffs arrays in elements.
double *coeff_vec = new double[Space::get_num_dofs(space)];
get_coeff_vector(space, coeff_vec);
// Set up the solver, matrix, and rhs according to the solver selection.
SparseMatrix* matrix = create_matrix(matrix_solver);
Vector* rhs = create_vector(matrix_solver);
Solver* solver = create_linear_solver(matrix_solver, matrix, rhs);
int it = 1;
while (1)
{
// Obtain the number of degrees of freedom.
int ndof = Space::get_num_dofs(space);
// Assemble the Jacobian matrix and residual vector.
dp->assemble(coeff_vec, matrix, rhs);
// Calculate the l2-norm of residual vector.
double res_l2_norm = get_l2_norm(rhs);
// Info for user.
info("---- Newton iter %d, ndof %d, res. l2 norm %g", it, Space::get_num_dofs(space), res_l2_norm);
// If l2 norm of the residual vector is within tolerance, then quit.
// NOTE: at least one full iteration forced
// here because sometimes the initial
// residual on fine mesh is too small.
if(res_l2_norm < NEWTON_TOL && it > 1) break;
// Multiply the residual vector with -1 since the matrix
// equation reads J(Y^n) \deltaY^{n+1} = -F(Y^n).
for(int i=0; i<ndof; i++) rhs->set(i, -rhs->get(i));
// Solve the linear system.
if(!solver->solve())
error ("Matrix solver failed.\n");
// Add \deltaY^{n+1} to Y^n.
for (int i = 0; i < ndof; i++) coeff_vec[i] += solver->get_solution()[i];
// If the maximum number of iteration has been reached, then quit.
if (it >= NEWTON_MAX_ITER) error ("Newton method did not converge.");
// Copy coefficients from vector y to elements.
set_coeff_vector(coeff_vec, space);
it++;
}
// Cleanup.
delete matrix;
delete rhs;
delete solver;
delete [] coeff_vec;
// Update the eigenvalue.
K_EFF_old = K_EFF;
K_EFF = calc_total_reaction_rate(space, nSf, 0., 40.);
// Convergence test.
if (fabs(K_EFF - K_EFF_old)/K_EFF < TOL_SI) break;
// Normalize total neutron flux to one fission neutron.
multiply_dofs_with_constant(space, 1./K_EFF, current_solution);
if (verbose) info("K_EFF_%d = %.8f", i+1, K_EFF);
}
// Print the converged eigenvalue.
info("K_EFF = %.8f, err= %.8f%%", K_EFF, 100*(K_EFF-1));
// Plot the converged critical neutron flux.
sprintf(solution_file, "solution.gp");
l.plot_solution(solution_file);
// Comparison with analytical results (see the reference above).
double flux[N_GRP], J[N_GRP], R;
get_solution_at_point(space, 0.0, flux, J);
R = flux[0]/flux[1];
info("phi_fast/phi_therm at x=0 : %.4f, err = %.2f%%", R, 100*(R-2.5332)/2.5332);
get_solution_at_point(space, 40.0, flux, J);
R = flux[0]/flux[1];
info("phi_fast/phi_therm at x=40 : %.4f, err = %.2f%%", R, 100*(R-1.5162)/1.5162);
info("Done.");
return 0;
}
int main(int argc, char* argv[])
{
// Load the mesh.
Mesh mesh;
H2DReader mloader;
mloader.load("square.mesh", &mesh);
// Initial mesh refinements.
for(int i = 0; i < INIT_GLOB_REF_NUM; i++) mesh.refine_all_elements();
mesh.refine_towards_boundary(1, INIT_BDY_REF_NUM);
// Create an H1 space with default shapeset.
H1Space space(&mesh, bc_types, essential_bc_values, P_INIT);
int ndof = Space::get_num_dofs(&space);
info("ndof = %d.", ndof);
// Solution for the previous time level.
Solution u_prev_time;
// Initialize the weak formulation.
WeakForm wf;
if (TIME_INTEGRATION == 1) {
wf.add_matrix_form(jac_euler, jac_ord, HERMES_UNSYM, HERMES_ANY, &u_prev_time);
wf.add_vector_form(res_euler, res_ord, HERMES_ANY, &u_prev_time);
}
else {
wf.add_matrix_form(jac_cranic, jac_ord, HERMES_UNSYM, HERMES_ANY, &u_prev_time);
wf.add_vector_form(res_cranic, res_ord, HERMES_ANY, &u_prev_time);
}
// Project the function init_cond() on the FE space
// to obtain initial coefficient vector for the Newton's method.
scalar* coeff_vec = new scalar[Space::get_num_dofs(&space)];
info("Projecting initial condition to obtain initial vector for the Newton's method.");
OGProjection::project_global(&space, init_cond, coeff_vec, matrix_solver);
Solution::vector_to_solution(coeff_vec, &space, &u_prev_time);
// Initialize views.
ScalarView sview("Solution", 0, 0, 500, 400);
OrderView oview("Mesh", 520, 0, 450, 400);
oview.show(&space);
sview.show(&u_prev_time);
// Time stepping loop:
double current_time = 0.0;
int t_step = 1;
do {
info("---- Time step %d, t = %g s.", t_step, current_time); t_step++;
// Initialize the FE problem.
bool is_linear = false;
DiscreteProblem dp(&wf, &space, is_linear);
// Set up the solver, matrix, and rhs according to the solver selection.
SparseMatrix* matrix = create_matrix(matrix_solver);
Vector* rhs = create_vector(matrix_solver);
Solver* solver = create_linear_solver(matrix_solver, matrix, rhs);
// Perform Newton's iteration.
int it = 1;
while (1)
{
// Obtain the number of degrees of freedom.
int ndof = Space::get_num_dofs(&space);
// Assemble the Jacobian matrix and residual vector.
dp.assemble(coeff_vec, matrix, rhs, false);
// Multiply the residual vector with -1 since the matrix
// equation reads J(Y^n) \deltaY^{n+1} = -F(Y^n).
for (int i = 0; i < ndof; i++) rhs->set(i, -rhs->get(i));
// Calculate the l2-norm of residual vector.
double res_l2_norm = get_l2_norm(rhs);
// Info for user.
info("---- Newton iter %d, ndof %d, res. l2 norm %g", it, Space::get_num_dofs(&space), res_l2_norm);
// If l2 norm of the residual vector is within tolerance, or the maximum number
// of iteration has been reached, then quit.
if (res_l2_norm < NEWTON_TOL || it > NEWTON_MAX_ITER) break;
// Solve the linear system.
if(!solver->solve())
error ("Matrix solver failed.\n");
// Add \deltaY^{n+1} to Y^n.
for (int i = 0; i < ndof; i++) coeff_vec[i] += solver->get_solution()[i];
if (it >= NEWTON_MAX_ITER)
error ("Newton method did not converge.");
it++;
}
// Translate the resulting coefficient vector into the Solution sln.
Solution::vector_to_solution(coeff_vec, &space, &u_prev_time);
// Cleanup.
delete matrix;
delete rhs;
delete solver;
// Update time.
current_time += TAU;
// Show the new time level solution.
char title[100];
sprintf(title, "Solution, t = %g", current_time);
sview.set_title(title);
sview.show(&u_prev_time);
} while (current_time < T_FINAL);
delete [] coeff_vec;
// Wait for all views to be closed.
View::wait();
return 0;
}
bool rk_time_step(double current_time, double time_step, ButcherTable* const bt,
scalar* coeff_vec, scalar* err_vec, DiscreteProblem* dp, MatrixSolverType matrix_solver,
bool verbose, bool is_linear, double newton_tol, int newton_max_iter,
double newton_damping_coeff, double newton_max_allowed_residual_norm)
{
// Check for not implemented features.
if (matrix_solver != SOLVER_UMFPACK)
error("Sorry, rk_time_step() still only works with UMFpack.");
if (dp->get_weak_formulation()->get_neq() > 1)
error("Sorry, rk_time_step() does not work with systems yet.");
// Get number of stages from the Butcher's table.
int num_stages = bt->get_size();
// Check whether the user provided a second B-row if he wants
// err_vec.
if(err_vec != NULL) {
double b2_coeff_sum = 0;
for (int i=0; i < num_stages; i++) b2_coeff_sum += fabs(bt->get_B2(i));
if (b2_coeff_sum < 1e-10)
error("err_vec != NULL but the B2 row in the Butcher's table is zero in rk_time_step().");
}
// Matrix for the time derivative part of the equation (left-hand side).
UMFPackMatrix* matrix_left = new UMFPackMatrix();
// Matrix and vector for the rest (right-hand side).
UMFPackMatrix* matrix_right = new UMFPackMatrix();
UMFPackVector* vector_right = new UMFPackVector();
// Create matrix solver.
Solver* solver = create_linear_solver(matrix_solver, matrix_right, vector_right);
// Get original space, mesh, and ndof.
dp->get_space(0);
Mesh* mesh = dp->get_space(0)->get_mesh();
int ndof = dp->get_space(0)->get_num_dofs();
// Create spaces for stage solutions. This is necessary
// to define a num_stages x num_stages block weak formulation.
Hermes::vector<Space*> stage_spaces;
stage_spaces.push_back(dp->get_space(0));
for (int i = 1; i < num_stages; i++) {
stage_spaces.push_back(dp->get_space(0)->dup(mesh));
}
Space::assign_dofs(stage_spaces);
// Create a multistage weak formulation.
WeakForm stage_wf_left; // For the matrix M (size ndof times ndof).
WeakForm stage_wf_right(num_stages); // For the rest of equation (written on the right),
// size num_stages*ndof times num_stages*ndof.
create_stage_wf(current_time, time_step, bt, dp, &stage_wf_left, &stage_wf_right);
// Initialize discrete problems for the assembling of the
// matrix M and the stage Jacobian matrix and residual.
DiscreteProblem stage_dp_left(&stage_wf_left, dp->get_space(0));
DiscreteProblem stage_dp_right(&stage_wf_right, stage_spaces);
// Vector K_vector of length num_stages * ndof. will represent
// the 'k_i' vectors in the usual R-K notation.
scalar* K_vector = new scalar[num_stages*ndof];
memset(K_vector, 0, num_stages * ndof * sizeof(scalar));
// Vector u_prev_vec will represent y_n + h \sum_{j=1}^s a_{ij}k_i
// in the usual R-K notation.
scalar* u_prev_vec = new scalar[num_stages*ndof];
// Vector for the left part of the residual.
scalar* vector_left = new scalar[num_stages*ndof];
// Prepare residuals of stage solutions.
Hermes::vector<Solution*> residuals;
Hermes::vector<bool> add_dir_lift;
for (int i = 0; i < num_stages; i++) {
residuals.push_back(new Solution(mesh));
add_dir_lift.push_back(false);
}
// Assemble the block-diagonal mass matrix M of size ndof times ndof.
// The corresponding part of the global residual vector is obtained
// just by multiplication.
stage_dp_left.assemble(matrix_left);
// The Newton's loop.
double residual_norm;
int it = 1;
while (true)
{
// Prepare vector Y_n + h\sum_{j=1}^s a_{ij} K_j.
for (int i = 0; i < num_stages; i++) { // block row
for (int idx = 0; idx < ndof; idx++) {
scalar increment = 0;
for (int j = 0; j < num_stages; j++) {
increment += bt->get_A(i, j) * K_vector[j*ndof + idx];
}
u_prev_vec[i*ndof + idx] = coeff_vec[idx] + time_step * increment;
}
}
multiply_as_diagonal_block_matrix(matrix_left, num_stages,
K_vector, vector_left);
// Assemble the block Jacobian matrix of the stationary residual F
// Diagonal blocks are created even if empty, so that matrix_left
// can be added later.
bool rhs_only = false;
bool force_diagonal_blocks = true;
stage_dp_right.assemble(u_prev_vec, matrix_right, vector_right,
rhs_only, force_diagonal_blocks);
matrix_right->add_to_diagonal_blocks(num_stages, matrix_left);
vector_right->add_vector(vector_left);
// Multiply the residual vector with -1 since the matrix
// equation reads J(Y^n) \deltaY^{n+1} = -F(Y^n).
vector_right->change_sign();
// Measure the residual norm.
if (HERMES_RESIDUAL_AS_VECTOR_RK) {
// Calculate the l2-norm of residual vector.
residual_norm = get_l2_norm(vector_right);
}
else {
// Translate residual vector into residual functions.
Solution::vector_to_solutions(vector_right, stage_dp_right.get_spaces(),
residuals, add_dir_lift);
residual_norm = calc_norms(residuals);
}
// Info for the user.
if (verbose) info("---- Newton iter %d, ndof %d, residual norm %g",
it, ndof, residual_norm);
// If maximum allowed residual norm is exceeded, fail.
if (residual_norm > newton_max_allowed_residual_norm) {
if (verbose) {
info("Current residual norm: %g", residual_norm);
info("Maximum allowed residual norm: %g", newton_max_allowed_residual_norm);
info("Newton solve not successful, returning false.");
}
return false;
}
// If residual norm is within tolerance, or the maximum number
// of iteration has been reached, or the problem is linear, then quit.
if ((residual_norm < newton_tol || it > newton_max_iter) && it > 1) break;
// Solve the linear system.
if(!solver->solve()) error ("Matrix solver failed.\n");
// Add \deltaY^{n+1} to Y^n.
for (int i = 0; i < num_stages*ndof; i++) {
K_vector[i] += newton_damping_coeff * solver->get_solution()[i];
}
// If the problem is linear, quit.
if (is_linear) {
if (verbose) {
info("Terminating Newton's loop as problem is linear.");
}
break;
}
// Increase iteration counter.
it++;
}
// If max number of iterations was exceeded, fail.
if (it >= newton_max_iter) {
if (verbose) info("Maximum allowed number of Newton iterations exceeded, returning false.");
return false;
}
// Calculate the vector \sum_{j=1}^s b_j k_j.
Vector* rk_increment_vector = create_vector(matrix_solver);
rk_increment_vector->alloc(ndof);
for (int i = 0; i < ndof; i++) {
rk_increment_vector->set(i, 0);
for (int j = 0; j < num_stages; j++) {
rk_increment_vector->add(i, bt->get_B(j) * K_vector[j*ndof + i]);
}
}
// Calculate Y^{n+1} = Y^n + h \sum_{j=1}^s b_j k_j.
for (int i = 0; i < ndof; i++) coeff_vec[i] += time_step * rk_increment_vector->get(i);
// If err_vec is not NULL, use the second B-row in the Butcher's
// table to calculate the second approximation Y_{n+1}. Then
// subtract the original one from it, and return this as an
// error vector err_vec.
if (err_vec != NULL) {
for (int i = 0; i < ndof; i++) {
rk_increment_vector->set(i, 0);
for (int j = 0; j < num_stages; j++) {
rk_increment_vector->add(i, bt->get_B2(j) * K_vector[j*ndof + i]);
}
}
for (int i = 0; i < ndof; i++) err_vec[i] = time_step * rk_increment_vector->get(i);
for (int i = 0; i < ndof; i++) err_vec[i] = err_vec[i] - coeff_vec[i];
}
// Clean up.
delete matrix_left;
delete matrix_right;
delete vector_right;
delete solver;
delete rk_increment_vector;
// Delete stage spaces, but not the first (original) one.
for (int i = 1; i < num_stages; i++) delete stage_spaces[i];
// Delete all residuals.
for (int i = 0; i < num_stages; i++) delete residuals[i];
// TODO: Delete stage_wf, in particular its external solutions
// stage_time_sol[i], i = 0, 1, ..., num_stages-1.
// Delete stage_vec and u_prev_vec.
delete [] K_vector;
delete [] u_prev_vec;
// debug
delete [] vector_left;
return true;
}
int main()
{
// Create space.
// Transform input data to the format used by the "Space" constructor.
SpaceData *md = new SpaceData();
Space* space = new Space(md->N_macroel, md->interfaces, md->poly_orders, md->material_markers, md->subdivisions, N_GRP, N_SLN);
delete md;
// Enumerate basis functions, info for user.
int ndof = Space::get_num_dofs(space);
info("ndof: %d", ndof);
// Plot the space.
space->plot("space.gp");
for (int g = 0; g < N_GRP; g++)
{
space->set_bc_left_dirichlet(g, flux_left_surf[g]);
space->set_bc_right_dirichlet(g, flux_right_surf[g]);
}
// Initialize the weak formulation.
WeakForm wf(2);
wf.add_matrix_form(0, 0, jacobian_mat1_0_0, NULL, mat1);
wf.add_matrix_form(0, 0, jacobian_mat2_0_0, NULL, mat2);
wf.add_matrix_form(0, 0, jacobian_mat3_0_0, NULL, mat3);
wf.add_matrix_form(0, 1, jacobian_mat1_0_1, NULL, mat1);
wf.add_matrix_form(0, 1, jacobian_mat2_0_1, NULL, mat2);
wf.add_matrix_form(0, 1, jacobian_mat3_0_1, NULL, mat3);
wf.add_matrix_form(1, 0, jacobian_mat1_1_0, NULL, mat1);
wf.add_matrix_form(1, 0, jacobian_mat2_1_0, NULL, mat2);
wf.add_matrix_form(1, 0, jacobian_mat3_1_0, NULL, mat3);
wf.add_matrix_form(1, 1, jacobian_mat1_1_1, NULL, mat1);
wf.add_matrix_form(1, 1, jacobian_mat2_1_1, NULL, mat2);
wf.add_matrix_form(1, 1, jacobian_mat3_1_1, NULL, mat3);
wf.add_vector_form(0, residual_mat1_0, NULL, mat1);
wf.add_vector_form(0, residual_mat2_0, NULL, mat2);
wf.add_vector_form(0, residual_mat3_0, NULL, mat3);
wf.add_vector_form(1, residual_mat1_1, NULL, mat1);
wf.add_vector_form(1, residual_mat2_1, NULL, mat2);
wf.add_vector_form(1, residual_mat3_1, NULL, mat3);
// Initialize the FE problem.
bool is_linear = false;
DiscreteProblem *dp = new DiscreteProblem(&wf, space, is_linear);
// Newton's loop.
// Fill vector coeff_vec using dof and coeffs arrays in elements.
double *coeff_vec = new double[Space::get_num_dofs(space)];
get_coeff_vector(space, coeff_vec);
// Set up the solver, matrix, and rhs according to the solver selection.
SparseMatrix* matrix = create_matrix(matrix_solver);
Vector* rhs = create_vector(matrix_solver);
Solver* solver = create_linear_solver(matrix_solver, matrix, rhs);
int it = 1;
while (1)
{
// Obtain the number of degrees of freedom.
int ndof = Space::get_num_dofs(space);
// Assemble the Jacobian matrix and residual vector.
dp->assemble(coeff_vec, matrix, rhs);
// Calculate the l2-norm of residual vector.
double res_l2_norm = get_l2_norm(rhs);
// Info for user.
info("---- Newton iter %d, ndof %d, res. l2 norm %g", it, Space::get_num_dofs(space), res_l2_norm);
// If l2 norm of the residual vector is within tolerance, then quit.
// NOTE: at least one full iteration forced
// here because sometimes the initial
// residual on fine mesh is too small.
if(res_l2_norm < NEWTON_TOL && it > 1) break;
// Multiply the residual vector with -1 since the matrix
// equation reads J(Y^n) \deltaY^{n+1} = -F(Y^n).
for(int i=0; i<ndof; i++) rhs->set(i, -rhs->get(i));
// Solve the linear system.
if(!solver->solve())
error ("Matrix solver failed.\n");
// Add \deltaY^{n+1} to Y^n.
for (int i = 0; i < ndof; i++) coeff_vec[i] += solver->get_solution()[i];
// If the maximum number of iteration has been reached, then quit.
if (it >= NEWTON_MAX_ITER) error ("Newton method did not converge.");
// Copy coefficients from vector y to elements.
set_coeff_vector(coeff_vec, space);
it++;
}
// Plot the solution.
Linearizer l(space);
l.plot_solution("solution.gp");
info("Done.");
return 0;
}
int main()
{
// Time measurement.
TimePeriod cpu_time;
cpu_time.tick();
// Create space, set Dirichlet BC, enumerate basis functions.
Space* space = new Space(A, B, NELEM, DIR_BC_LEFT, DIR_BC_RIGHT, P_INIT, NEQ);
int ndof = Space::get_num_dofs(space);
info("ndof: %d", ndof);
// Initialize the weak formulation.
WeakForm wf;
wf.add_matrix_form(jacobian);
wf.add_vector_form(residual);
// Initialize the FE problem.
bool is_linear = false;
DiscreteProblem *dp = new DiscreteProblem(&wf, space, is_linear);
// Set zero initial condition.
double *coeff_vec = new double[ndof];
set_zero(coeff_vec, ndof);
// Set up the solver, matrix, and rhs according to the solver selection.
SparseMatrix* matrix = create_matrix(matrix_solver);
Vector* rhs = create_vector(matrix_solver);
Solver* solver = create_linear_solver(matrix_solver, matrix, rhs);
int it = 1;
bool success = false;
while (1)
{
// Obtain the number of degrees of freedom.
int ndof = Space::get_num_dofs(space);
// Assemble the Jacobian matrix and residual vector.
dp->assemble(coeff_vec, matrix, rhs);
// Calculate the l2-norm of residual vector.
double res_l2_norm = get_l2_norm(rhs);
// Info for user.
info("---- Newton iter %d, ndof %d, res. l2 norm %g", it, Space::get_num_dofs(space), res_l2_norm);
// If l2 norm of the residual vector is within tolerance, then quit.
// NOTE: at least one full iteration forced
// here because sometimes the initial
// residual on fine mesh is too small.
if(res_l2_norm < NEWTON_TOL && it > 1) break;
// Multiply the residual vector with -1 since the matrix
// equation reads J(Y^n) \deltaY^{n+1} = -F(Y^n).
for(int i=0; i<ndof; i++) rhs->set(i, -rhs->get(i));
// Solve the linear system.
if(!(success = solver->solve()))
error ("Matrix solver failed.\n");
// Add \deltaY^{n+1} to Y^n.
for (int i = 0; i < ndof; i++) coeff_vec[i] += solver->get_solution()[i];
// If the maximum number of iteration has been reached, then quit.
if (it >= NEWTON_MAX_ITER) error ("Newton method did not converge.");
it++;
}
info("Total running time: %g s", cpu_time.accumulated());
// Test variable.
info("ndof = %d.", Space::get_num_dofs(space));
if (success)
{
info("Success!");
return ERROR_SUCCESS;
}
else
{
info("Failure!");
return ERROR_FAILURE;
}
}
int main() {
// Three macroelements are defined above via the interfaces[] array.
// poly_orders[]... initial poly degrees of macroelements.
// material_markers[]... material markers of macroelements.
// subdivisions[]... equidistant subdivision of macroelements.
int poly_orders[N_MAT] = {P_init_inner, P_init_outer, P_init_reflector };
int material_markers[N_MAT] = {Marker_inner, Marker_outer, Marker_reflector };
int subdivisions[N_MAT] = {N_subdiv_inner, N_subdiv_outer, N_subdiv_reflector };
// Create space.
Space* space = new Space(N_MAT, interfaces, poly_orders, material_markers, subdivisions, N_GRP, N_SLN);
// Enumerate basis functions, info for user.
info("N_dof = %d", Space::get_num_dofs(space));
// Initial approximation: u = 1.
double K_EFF_old;
double init_val = 1.0;
set_vertex_dofs_constant(space, init_val, 0);
// Initialize the weak formulation.
WeakForm wf;
wf.add_matrix_form(jacobian_vol_inner, NULL, Marker_inner);
wf.add_matrix_form(jacobian_vol_outer, NULL, Marker_outer);
wf.add_matrix_form(jacobian_vol_reflector, NULL, Marker_reflector);
wf.add_vector_form(residual_vol_inner, NULL, Marker_inner);
wf.add_vector_form(residual_vol_outer, NULL, Marker_outer);
wf.add_vector_form(residual_vol_reflector, NULL, Marker_reflector);
wf.add_vector_form_surf(residual_surf_left, BOUNDARY_LEFT);
wf.add_matrix_form_surf(jacobian_surf_right, BOUNDARY_RIGHT);
wf.add_vector_form_surf(residual_surf_right, BOUNDARY_RIGHT);
// Initialize the FE problem.
bool is_linear = false;
DiscreteProblem *dp = new DiscreteProblem(&wf, space, is_linear);
// Source iteration (power method).
for (int i = 0; i < Max_SI; i++)
{
// Obtain fission source.
int current_solution = 0, previous_solution = 1;
copy_dofs(current_solution, previous_solution, space);
// Obtain the number of degrees of freedom.
int ndof = Space::get_num_dofs(space);
// Fill vector coeff_vec using dof and coeffs arrays in elements.
double *coeff_vec = new double[Space::get_num_dofs(space)];
get_coeff_vector(space, coeff_vec);
// Set up the solver, matrix, and rhs according to the solver selection.
SparseMatrix* matrix = create_matrix(matrix_solver);
Vector* rhs = create_vector(matrix_solver);
Solver* solver = create_linear_solver(matrix_solver, matrix, rhs);
int it = 1;
while (1) {
// Obtain the number of degrees of freedom.
int ndof = Space::get_num_dofs(space);
// Assemble the Jacobian matrix and residual vector.
dp->assemble(matrix, rhs);
// Calculate the l2-norm of residual vector.
double res_l2_norm = get_l2_norm(rhs);
// Info for user.
info("---- Newton iter %d, ndof %d, res. l2 norm %g", it, Space::get_num_dofs(space), res_l2_norm);
// If l2 norm of the residual vector is within tolerance, then quit.
// NOTE: at least one full iteration forced
// here because sometimes the initial
// residual on fine mesh is too small.
if(res_l2_norm < NEWTON_TOL && it > 1) break;
// Multiply the residual vector with -1 since the matrix
// equation reads J(Y^n) \deltaY^{n+1} = -F(Y^n).
for(int i=0; i<ndof; i++) rhs->set(i, -rhs->get(i));
// Solve the linear system.
if(!solver->solve())
error ("Matrix solver failed.\n");
// Add \deltaY^{n+1} to Y^n.
for (int i = 0; i < ndof; i++) coeff_vec[i] += solver->get_solution()[i];
// If the maximum number of iteration has been reached, then quit.
if (it >= NEWTON_MAX_ITER) error ("Newton method did not converge.");
// Copy coefficients from vector y to elements.
set_coeff_vector(coeff_vec, space);
it++;
}
// Cleanup.
delete matrix;
delete rhs;
delete solver;
delete [] coeff_vec;
// Update the eigenvalue.
K_EFF_old = K_EFF;
K_EFF = calc_fission_yield(space);
info("K_EFF_%d = %f", i, K_EFF);
if (fabs(K_EFF - K_EFF_old)/K_EFF < TOL_SI) break;
}
// Plot the critical (i.e. steady-state) neutron flux.
Linearizer l(space);
l.plot_solution("solution.gp");
// Normalize so that the absolute neutron flux generates 320 Watts of energy
// (note that, using the symmetry condition at the origin, we've solved for
// flux only in the right half of the reactor).
normalize_to_power(space, 320/2.);
// Plot the solution and space.
l.plot_solution("solution_320W.gp");
space->plot("space.gp");
info("K_EFF = %f", K_EFF);
info("Done.");
return 0;
}
bool rk_time_step(double current_time, double time_step, ButcherTable* const bt,
Solution* sln_time_prev, Solution* sln_time_new, Solution* error_fn,
DiscreteProblem* dp, MatrixSolverType matrix_solver,
bool verbose, bool is_linear, double newton_tol, int newton_max_iter,
double newton_damping_coeff, double newton_max_allowed_residual_norm)
{
// Check for not implemented features.
if (matrix_solver != SOLVER_UMFPACK)
error("Sorry, rk_time_step() still only works with UMFpack.");
if (dp->get_weak_formulation()->get_neq() > 1)
error("Sorry, rk_time_step() does not work with systems yet.");
// Get number of stages from the Butcher's table.
int num_stages = bt->get_size();
// Check whether the user provided a nonzero B2-row if he wants temporal error estimation.
if(error_fn != NULL) if (bt->is_embedded() == false) {
error("rk_time_step(): R-K method must be embedded if temporal error estimate is requested.");
}
// Matrix for the time derivative part of the equation (left-hand side).
UMFPackMatrix* matrix_left = new UMFPackMatrix();
// Matrix and vector for the rest (right-hand side).
UMFPackMatrix* matrix_right = new UMFPackMatrix();
UMFPackVector* vector_right = new UMFPackVector();
// Create matrix solver.
Solver* solver = create_linear_solver(matrix_solver, matrix_right, vector_right);
// Get space, mesh, and ndof for the stage solutions in the R-K method (K_i vectors).
Space* K_space = dp->get_space(0);
Mesh* K_mesh = K_space->get_mesh();
int ndof = K_space->get_num_dofs();
// Create spaces for stage solutions K_i. This is necessary
// to define a num_stages x num_stages block weak formulation.
Hermes::vector<Space*> stage_spaces;
stage_spaces.push_back(K_space);
for (int i = 1; i < num_stages; i++) {
stage_spaces.push_back(K_space->dup(K_mesh));
}
Space::assign_dofs(stage_spaces);
// Create a multistage weak formulation.
WeakForm stage_wf_left; // For the matrix M (size ndof times ndof).
WeakForm stage_wf_right(num_stages); // For the rest of equation (written on the right),
// size num_stages*ndof times num_stages*ndof.
Solution** stage_time_sol = new Solution*[num_stages];
// This array will be filled by artificially created
// solutions to represent stage times.
create_stage_wf(current_time, time_step, bt, dp, &stage_wf_left, &stage_wf_right, stage_time_sol);
// Initialize discrete problems for the assembling of the
// matrix M and the stage Jacobian matrix and residual.
DiscreteProblem stage_dp_left(&stage_wf_left, K_space);
DiscreteProblem stage_dp_right(&stage_wf_right, stage_spaces);
// Vector K_vector of length num_stages * ndof. will represent
// the 'K_i' vectors in the usual R-K notation.
scalar* K_vector = new scalar[num_stages*ndof];
memset(K_vector, 0, num_stages * ndof * sizeof(scalar));
// Vector u_ext_vec will represent h \sum_{j=1}^s a_{ij} K_i.
scalar* u_ext_vec = new scalar[num_stages*ndof];
// Vector for the left part of the residual.
scalar* vector_left = new scalar[num_stages*ndof];
// Prepare residuals of stage solutions.
Hermes::vector<Solution*> residuals;
Hermes::vector<bool> add_dir_lift;
for (int i = 0; i < num_stages; i++) {
residuals.push_back(new Solution(K_mesh));
add_dir_lift.push_back(false);
}
// Assemble the block-diagonal mass matrix M of size ndof times ndof.
// The corresponding part of the global residual vector is obtained
// just by multiplication.
stage_dp_left.assemble(matrix_left);
// The Newton's loop.
double residual_norm;
int it = 1;
while (true)
{
// Prepare vector h\sum_{j=1}^s a_{ij} K_j.
for (int i = 0; i < num_stages; i++) { // block row
for (int idx = 0; idx < ndof; idx++) {
scalar increment = 0;
for (int j = 0; j < num_stages; j++) {
increment += bt->get_A(i, j) * K_vector[j*ndof + idx];
}
u_ext_vec[i*ndof + idx] = time_step * increment;
}
}
multiply_as_diagonal_block_matrix(matrix_left, num_stages, K_vector, vector_left);
// Assemble the block Jacobian matrix of the stationary residual F
// Diagonal blocks are created even if empty, so that matrix_left
// can be added later.
bool rhs_only = false;
bool force_diagonal_blocks = true;
stage_dp_right.assemble(u_ext_vec, matrix_right, vector_right,
rhs_only, force_diagonal_blocks, false); // false = do not add Dirichlet lift while
// converting u_ext_vec into Solutions.
matrix_right->add_to_diagonal_blocks(num_stages, matrix_left);
vector_right->add_vector(vector_left);
// Multiply the residual vector with -1 since the matrix
// equation reads J(Y^n) \deltaY^{n+1} = -F(Y^n).
vector_right->change_sign();
// Measure the residual norm.
if (HERMES_RESIDUAL_AS_VECTOR_RK) {
// Calculate the l2-norm of residual vector.
residual_norm = get_l2_norm(vector_right);
}
else {
// Translate residual vector into residual functions.
Solution::vector_to_solutions(vector_right, stage_dp_right.get_spaces(),
residuals, add_dir_lift);
residual_norm = calc_norms(residuals);
}
// Info for the user.
if (verbose) info("---- Newton iter %d, ndof %d, residual norm %g",
it, ndof, residual_norm);
// If maximum allowed residual norm is exceeded, fail.
if (residual_norm > newton_max_allowed_residual_norm) {
if (verbose) {
info("Current residual norm: %g", residual_norm);
info("Maximum allowed residual norm: %g", newton_max_allowed_residual_norm);
info("Newton solve not successful, returning false.");
}
return false;
}
// If residual norm is within tolerance, or the maximum number
// of iteration has been reached, or the problem is linear, then quit.
if ((residual_norm < newton_tol || it > newton_max_iter) && it > 1) break;
// Solve the linear system.
if(!solver->solve()) error ("Matrix solver failed.\n");
// Add \deltaK^{n+1} to K^n.
for (int i = 0; i < num_stages*ndof; i++) {
K_vector[i] += newton_damping_coeff * solver->get_solution()[i];
}
// If the problem is linear, quit.
if (is_linear) {
if (verbose) {
info("Terminating Newton's loop as problem is linear.");
}
break;
}
// Increase iteration counter.
it++;
}
// If max number of iterations was exceeded, fail.
if (it >= newton_max_iter) {
if (verbose) info("Maximum allowed number of Newton iterations exceeded, returning false.");
return false;
}
// Project previous time level solution on the stage space,
// to be able to add them together. The result of the projection
// will be stored in the vector coeff_vec.
// FIXME - this projection is slow and it is not needed when the
// spaces are the same (if spatial adaptivity does not take place).
scalar* coeff_vec = new scalar[ndof];
OGProjection::project_global(K_space, sln_time_prev, coeff_vec, matrix_solver);
// Calculate new time level solution in the stage space (u_{n+1} = u_n + h \sum_{j=1}^s b_j k_j).
for (int i = 0; i < ndof; i++) {
for (int j = 0; j < num_stages; j++) {
coeff_vec[i] += time_step * bt->get_B(j) * K_vector[j*ndof + i];
}
}
Solution::vector_to_solution(coeff_vec, K_space, sln_time_new);
// If error_fn is not NULL, use the B2-row in the Butcher's
// table to calculate the temporal error estimate.
if (error_fn != NULL) {
for (int i = 0; i < ndof; i++) {
coeff_vec[i] = 0;
for (int j = 0; j < num_stages; j++) {
coeff_vec[i] += (bt->get_B(j) - bt->get_B2(j)) * K_vector[j*ndof + i];
}
coeff_vec[i] *= time_step;
}
Solution::vector_to_solution(coeff_vec, K_space, error_fn, false);
}
// Clean up.
delete matrix_left;
delete matrix_right;
delete vector_right;
delete solver;
// Delete stage spaces, but not the first (original) one.
for (int i = 1; i < num_stages; i++) delete stage_spaces[i];
// Delete all residuals.
for (int i = 0; i < num_stages; i++) delete residuals[i];
// Delete artificial Solutions with stage times.
for (int i = 0; i < num_stages; i++)
delete stage_time_sol[i];
delete [] stage_time_sol;
// Clean up.
delete [] K_vector;
delete [] u_ext_vec;
delete [] coeff_vec;
delete [] vector_left;
return true;
}
double NonlinearMatrixSolver<Scalar>::calculate_residual_norm()
{
return get_l2_norm(this->get_residual());
}
void NonlinearMatrixSolver<Scalar>::solve(Scalar* coeff_vec)
{
// Initialization.
this->init_solving(coeff_vec);
#pragma region parameter_setup
// Initialize parameters.
unsigned int successful_steps_damping = 0;
unsigned int successful_steps_jacobian = 0;
std::vector<bool> iterations_with_recalculated_jacobian;
std::vector<double> residual_norms;
std::vector<double> solution_norms;
std::vector<double> solution_change_norms;
std::vector<double> damping_factors;
bool residual_norm_drop;
unsigned int iteration = 1;
// Initial damping factor.
damping_factors.push_back(this->manual_damping ? manual_damping_factor : initial_auto_damping_factor);
// Link parameters.
this->set_parameter_value(this->p_residual_norms, &residual_norms);
this->set_parameter_value(this->p_solution_norms, &solution_norms);
this->set_parameter_value(this->p_solution_change_norms, &solution_change_norms);
this->set_parameter_value(this->p_successful_steps_damping, &successful_steps_damping);
this->set_parameter_value(this->p_successful_steps_jacobian, &successful_steps_jacobian);
this->set_parameter_value(this->p_residual_norm_drop, &residual_norm_drop);
this->set_parameter_value(this->p_damping_factors, &damping_factors);
this->set_parameter_value(this->p_iterations_with_recalculated_jacobian, &iterations_with_recalculated_jacobian);
this->set_parameter_value(this->p_iteration, &iteration);
#pragma endregion
// Does the initial step & checks the convergence & deallocates.
if (this->do_initial_step_return_finished())
return;
// Main Nonlinear loop
while (true)
{
// Handle the event of step beginning.
if (!this->on_step_begin())
{
this->info("\tNonlinearSolver: aborted.");
this->finalize_solving();
return;
}
#pragma region damping_factor_loop
this->info("\n\tNonlinearSolver: Damping factor handling:");
// Loop searching for the damping factor.
do
{
// Assemble just the residual.
this->assemble_residual(false);
// Current residual norm.
this->get_parameter_value(this->p_residual_norms).push_back(this->calculate_residual_norm());
// Test convergence - if in this loop we found a solution.
this->info("\tNonlinearSolver: Convergence test...");
if (this->handle_convergence_state_return_finished(this->get_convergence_state()))
return;
// Inspect the damping factor.
this->info("\tNonlinearSolver: Probing the damping factor...");
try
{
// Calculate damping factor, and return whether or not was this a successful step.
residual_norm_drop = this->calculate_damping_factor(successful_steps_damping);
}
catch (Exceptions::NonlinearException& e)
{
this->finalize_solving();
throw e;
return;
}
if (!residual_norm_drop)
{
// Delete the previous residual and solution norm.
residual_norms.pop_back();
solution_norms.pop_back();
// Adjust the previous solution change norm.
solution_change_norms.back() /= this->auto_damping_ratio;
// Try with the different damping factor.
// Important thing here is the factor used that must be calculated from the current one and the previous one.
// This results in the following relation (since the damping factor is only updated one way).
for (int i = 0; i < this->problem_size; i++)
this->sln_vector[i] = this->previous_sln_vector[i] + (this->sln_vector[i] - this->previous_sln_vector[i]) / this->auto_damping_ratio;
// Add new_ solution norm.
solution_norms.push_back(get_l2_norm(this->sln_vector, this->problem_size));
}
} while (!residual_norm_drop);
#pragma endregion
// Damping factor was updated, handle the event.
this->on_damping_factor_updated();
#pragma region jacobian_reusage_loop
this->info("\n\tNonlinearSolver: Jacobian handling:");
// Loop until jacobian is not reusable anymore.
// The whole loop is skipped if the jacobian is not suitable for being reused at all.
while (this->jacobian_reusable && (this->constant_jacobian || force_reuse_jacobian_values(successful_steps_jacobian)))
{
this->residual_back->set_vector(this->get_residual());
// Info & handle the situation as necessary.
this->info("\tNonlinearSolver: Reusing Jacobian.");
this->on_reused_jacobian_step_begin();
// Solve the system.
this->linear_matrix_solver->set_reuse_scheme(HERMES_REUSE_MATRIX_STRUCTURE_COMPLETELY);
this->solve_linear_system();
// Assemble next residual for both reusage and convergence test.
this->assemble_residual(false);
// Current residual norm.
this->get_parameter_value(this->p_residual_norms).push_back(this->calculate_residual_norm());
// Test whether it was okay to reuse the jacobian.
if (!this->jacobian_reused_okay(successful_steps_jacobian))
{
this->warn("\tNonlinearSolver: Reused Jacobian disapproved.");
this->get_parameter_value(this->p_residual_norms).pop_back();
this->get_parameter_value(this->p_solution_norms).pop_back();
this->get_parameter_value(this->p_solution_change_norms).pop_back();
memcpy(this->sln_vector, this->previous_sln_vector, sizeof(Scalar)*this->problem_size);
this->get_residual()->set_vector(residual_back);
break;
}
// Increase the iteration count.
this->get_parameter_value(this->p_iterations_with_recalculated_jacobian).push_back(false);
this->get_parameter_value(this->p_iteration)++;
// Output successful reuse info.
this->step_info();
// Handle the event of end of a step.
this->on_reused_jacobian_step_end();
if (!this->on_step_end())
{
this->info("\tNonlinearSolver: aborted.");
this->finalize_solving();
return;
}
// Test convergence - if in this iteration we found a solution.
if (this->handle_convergence_state_return_finished(this->get_convergence_state()))
return;
}
#pragma endregion
// Reassemble the jacobian once not reusable anymore.
this->info("\tNonlinearSolver: Re-calculating Jacobian.");
// Set factorization scheme.
this->assemble_jacobian(true);
this->linear_matrix_solver->set_reuse_scheme(HERMES_CREATE_STRUCTURE_FROM_SCRATCH);
// Solve the system, state that the jacobian is reusable should it be desirable.
this->solve_linear_system();
this->jacobian_reusable = true;
// Increase the iteration count.
this->get_parameter_value(this->p_iterations_with_recalculated_jacobian).push_back(true);
this->get_parameter_value(this->p_iteration)++;
// Output info.
this->step_info();
// Handle the event of end of a step.
if (!this->on_step_end())
{
this->info("\tNonlinearSolver: aborted.");
this->finalize_solving();
return;
}
// Test convergence - if in this iteration we found a solution.
if (this->handle_convergence_state_return_finished(this->get_convergence_state()))
return;
}
}
int main()
{
// Time measurement.
TimePeriod cpu_time;
cpu_time.tick();
// Create coarse mesh, set Dirichlet BC, enumerate basis functions.
Space* space = new Space(A, B, NELEM, DIR_BC_LEFT, DIR_BC_RIGHT, P_INIT, NEQ);
// Enumerate basis functions, info for user.
int ndof = Space::get_num_dofs(space);
info("ndof: %d", ndof);
// Initialize the weak formulation.
WeakForm wf;
wf.add_matrix_form(jacobian);
wf.add_vector_form(residual);
double elem_errors[MAX_ELEM_NUM]; // This array decides what
// elements will be refined.
ElemPtr2 ref_elem_pairs[MAX_ELEM_NUM]; // To store element pairs from the
// FTR solution. Decides how
// elements will be hp-refined.
for (int i=0; i < MAX_ELEM_NUM; i++)
{
ref_elem_pairs[i][0] = new Element();
ref_elem_pairs[i][1] = new Element();
}
// DOF and CPU convergence graphs.
SimpleGraph graph_dof_exact, graph_cpu_exact;
/// Adaptivity loop:
int as = 1;
bool done = false;
do
{
info("---- Adaptivity step %d:", as);
// Initialize the FE problem.
bool is_linear = false;
DiscreteProblem *dp_coarse = new DiscreteProblem(&wf, space, is_linear);
// Newton's loop on coarse mesh.
// Fill vector coeff_vec using dof and coeffs arrays in elements.
double *coeff_vec_coarse = new double[Space::get_num_dofs(space)];
get_coeff_vector(space, coeff_vec_coarse);
// Set up the solver, matrix, and rhs according to the solver selection.
SparseMatrix* matrix_coarse = create_matrix(matrix_solver);
Vector* rhs_coarse = create_vector(matrix_solver);
Solver* solver_coarse = create_linear_solver(matrix_solver, matrix_coarse, rhs_coarse);
int it = 1;
while (1)
{
// Obtain the number of degrees of freedom.
int ndof_coarse = Space::get_num_dofs(space);
// Assemble the Jacobian matrix and residual vector.
dp_coarse->assemble(coeff_vec_coarse, matrix_coarse, rhs_coarse);
// Calculate the l2-norm of residual vector.
double res_l2_norm = get_l2_norm(rhs_coarse);
// Info for user.
info("---- Newton iter %d, ndof %d, res. l2 norm %g", it, Space::get_num_dofs(space), res_l2_norm);
// If l2 norm of the residual vector is within tolerance, then quit.
// NOTE: at least one full iteration forced
// here because sometimes the initial
// residual on fine mesh is too small.
if(res_l2_norm < NEWTON_TOL_COARSE && it > 1) break;
// Multiply the residual vector with -1 since the matrix
// equation reads J(Y^n) \deltaY^{n+1} = -F(Y^n).
for(int i=0; i<ndof_coarse; i++) rhs_coarse->set(i, -rhs_coarse->get(i));
// Solve the linear system.
if(!solver_coarse->solve())
error ("Matrix solver failed.\n");
// Add \deltaY^{n+1} to Y^n.
for (int i = 0; i < ndof_coarse; i++) coeff_vec_coarse[i] += solver_coarse->get_solution()[i];
// If the maximum number of iteration has been reached, then quit.
if (it >= NEWTON_MAX_ITER) error ("Newton method did not converge.");
// Copy coefficients from vector y to elements.
set_coeff_vector(coeff_vec_coarse, space);
it++;
}
// Cleanup.
delete matrix_coarse;
delete rhs_coarse;
delete solver_coarse;
delete [] coeff_vec_coarse;
delete dp_coarse;
// For every element perform its fast trial refinement (FTR),
// calculate the norm of the difference between the FTR
// solution and the coarse space solution, and store the
// error in the elem_errors[] array.
int n_elem = space->get_n_active_elem();
for (int i=0; i < n_elem; i++)
{
info("=== Starting FTR of Elem [%d].", i);
// Replicate coarse space including solution.
Space *space_ref_local = space->replicate();
// Perform FTR of element 'i'
space_ref_local->reference_refinement(i, 1);
info("Elem [%d]: fine space created (%d DOF).",
i, space_ref_local->assign_dofs());
// Initialize the FE problem.
bool is_linear = false;
DiscreteProblem* dp = new DiscreteProblem(&wf, space_ref_local, is_linear);
// Set up the solver, matrix, and rhs according to the solver selection.
SparseMatrix* matrix = create_matrix(matrix_solver);
Vector* rhs = create_vector(matrix_solver);
Solver* solver = create_linear_solver(matrix_solver, matrix, rhs);
// Newton's loop on the FTR space.
// Fill vector coeff_vec using dof and coeffs arrays in elements.
double *coeff_vec = new double[Space::get_num_dofs(space_ref_local)];
get_coeff_vector(space_ref_local, coeff_vec);
memset(coeff_vec, 0, Space::get_num_dofs(space_ref_local)*sizeof(double));
int it = 1;
while (1)
{
// Obtain the number of degrees of freedom.
int ndof = Space::get_num_dofs(space_ref_local);
// Assemble the Jacobian matrix and residual vector.
dp->assemble(coeff_vec, matrix, rhs);
// Calculate the l2-norm of residual vector.
double res_l2_norm = get_l2_norm(rhs);
// Info for user.
info("---- Newton iter %d, ndof %d, res. l2 norm %g", it, Space::get_num_dofs(space_ref_local), res_l2_norm);
// If l2 norm of the residual vector is within tolerance, then quit.
// NOTE: at least one full iteration forced
// here because sometimes the initial
// residual on fine mesh is too small.
if(res_l2_norm < NEWTON_TOL_REF && it > 1) break;
// Multiply the residual vector with -1 since the matrix
// equation reads J(Y^n) \deltaY^{n+1} = -F(Y^n).
for(int i=0; i<ndof; i++) rhs->set(i, -rhs->get(i));
// Solve the linear system.
if(!solver->solve())
error ("Matrix solver failed.\n");
// Add \deltaY^{n+1} to Y^n.
for (int i = 0; i < ndof; i++) coeff_vec[i] += solver->get_solution()[i];
// If the maximum number of iteration has been reached, then quit.
if (it >= NEWTON_MAX_ITER) error ("Newton method did not converge.");
// Copy coefficients from vector y to elements.
set_coeff_vector(coeff_vec, space_ref_local);
it++;
}
// Cleanup.
delete matrix;
delete rhs;
delete solver;
delete dp;
delete [] coeff_vec;
// Print FTR solution (enumerated).
//Linearizer *lxx = new Linearizer(space_ref_local);
//char out_filename[255];
//sprintf(out_filename, "solution_ref_%d.gp", i);
//lxx->plot_solution(out_filename);
//delete lxx;
// Calculate norm of the difference between the coarse space
// and FTR solutions.
// NOTE: later we want to look at the difference in some quantity
// of interest rather than error in global norm.
double err_est_array[MAX_ELEM_NUM];
elem_errors[i] = calc_err_est(NORM, space, space_ref_local, err_est_array) * 100;
info("Elem [%d]: absolute error (est) = %g%%", i, elem_errors[i]);
// Copy the reference element pair for element 'i'.
// into the ref_elem_pairs[i][] array
Iterator *I = new Iterator(space);
Iterator *I_ref = new Iterator(space_ref_local);
Element *e, *e_ref;
while (1)
{
e = I->next_active_element();
e_ref = I_ref->next_active_element();
if (e->id == i)
{
e_ref->copy_into(ref_elem_pairs[e->id][0]);
// coarse element 'e' was split in space.
if (e->level != e_ref->level)
{
e_ref = I_ref->next_active_element();
e_ref->copy_into(ref_elem_pairs[e->id][1]);
}
break;
}
}
delete I;
delete I_ref;
delete space_ref_local;
}
// Time measurement.
cpu_time.tick();
// If exact solution available, also calculate exact error.
if (EXACT_SOL_PROVIDED)
{
// Calculate element errors wrt. exact solution.
double err_exact_rel = calc_err_exact(NORM, space, exact_sol, NEQ, A, B) * 100;
// Info for user.
info("Relative error (exact) = %g %%", err_exact_rel);
// Add entry to DOF and CPU convergence graphs.
graph_dof_exact.add_values(Space::get_num_dofs(space), err_exact_rel);
graph_cpu_exact.add_values(cpu_time.accumulated(), err_exact_rel);
}
// Calculate max FTR error.
double max_ftr_error = 0;
for (int i=0; i < space->get_n_active_elem(); i++)
{
if (elem_errors[i] > max_ftr_error) max_ftr_error = elem_errors[i];
}
info("Max FTR error = %g%%.", max_ftr_error);
// Decide whether the max. FTR error is sufficiently small.
if(max_ftr_error < TOL_ERR_FTR) break;
// debug
//if (as >= 1) break;
// Returns updated coarse space with the last solution on it.
adapt(NORM, ADAPT_TYPE, THRESHOLD, elem_errors, space, ref_elem_pairs);
// Plot spaces, results, and errors.
//adapt_plotting(space, ref_elem_pairs, NORM, EXACT_SOL_PROVIDED, exact_sol);
as++;
}
while (done == false);
info("Total running time: %g s", cpu_time.accumulated());
// Save convergence graphs.
graph_dof_exact.save("conv_dof_exact.dat");
graph_cpu_exact.save("conv_cpu_exact.dat");
// Test variable.
bool success = true;
info("ndof = %d.", Space::get_num_dofs(space));
if (Space::get_num_dofs(space) > 35) success = false;
if (success)
{
info("Success!");
return ERROR_SUCCESS;
}
else
{
info("Failure!");
return ERROR_FAILURE;
}
}
int main()
{
// Create space.
// Transform input data to the format used by the "Space" constructor.
SpaceData *md = new SpaceData();
// Boundary conditions.
Hermes::vector<BCSpec *>DIR_BC_LEFT;
Hermes::vector<BCSpec *>DIR_BC_RIGHT;
for (int g = 0; g < N_GRP; g++)
{
DIR_BC_LEFT.push_back(new BCSpec(g,flux_left_surf[g]));
DIR_BC_RIGHT.push_back(new BCSpec(g,flux_right_surf[g]));
}
Space* space = new Space(md->N_macroel, md->interfaces, md->poly_orders, md->material_markers, md->subdivisions,
DIR_BC_LEFT, DIR_BC_RIGHT, N_GRP, N_SLN);
delete md;
// Enumerate basis functions, info for user.
int ndof = Space::get_num_dofs(space);
info("ndof: %d", ndof);
// Plot the space.
// Initialize the weak formulation.
WeakForm wf(2);
wf.add_matrix_form(0, 0, jacobian_mat1_0_0, NULL, mat1);
wf.add_matrix_form(0, 0, jacobian_mat2_0_0, NULL, mat2);
wf.add_matrix_form(0, 0, jacobian_mat3_0_0, NULL, mat3);
wf.add_matrix_form(0, 1, jacobian_mat1_0_1, NULL, mat1);
wf.add_matrix_form(0, 1, jacobian_mat2_0_1, NULL, mat2);
wf.add_matrix_form(0, 1, jacobian_mat3_0_1, NULL, mat3);
wf.add_matrix_form(1, 0, jacobian_mat1_1_0, NULL, mat1);
wf.add_matrix_form(1, 0, jacobian_mat2_1_0, NULL, mat2);
wf.add_matrix_form(1, 0, jacobian_mat3_1_0, NULL, mat3);
wf.add_matrix_form(1, 1, jacobian_mat1_1_1, NULL, mat1);
wf.add_matrix_form(1, 1, jacobian_mat2_1_1, NULL, mat2);
wf.add_matrix_form(1, 1, jacobian_mat3_1_1, NULL, mat3);
wf.add_vector_form(0, residual_mat1_0, NULL, mat1);
wf.add_vector_form(0, residual_mat2_0, NULL, mat2);
wf.add_vector_form(0, residual_mat3_0, NULL, mat3);
wf.add_vector_form(1, residual_mat1_1, NULL, mat1);
wf.add_vector_form(1, residual_mat2_1, NULL, mat2);
wf.add_vector_form(1, residual_mat3_1, NULL, mat3);
// Initialize the FE problem.
bool is_linear = false;
DiscreteProblem *dp = new DiscreteProblem(&wf, space, is_linear);
// Newton's loop.
// Fill vector coeff_vec using dof and coeffs arrays in elements.
double *coeff_vec = new double[Space::get_num_dofs(space)];
get_coeff_vector(space, coeff_vec);
// Set up the solver, matrix, and rhs according to the solver selection.
SparseMatrix* matrix = create_matrix(matrix_solver);
Vector* rhs = create_vector(matrix_solver);
Solver* solver = create_linear_solver(matrix_solver, matrix, rhs);
int it = 1;
while (1)
{
// Obtain the number of degrees of freedom.
int ndof = Space::get_num_dofs(space);
// Assemble the Jacobian matrix and residual vector.
dp->assemble(coeff_vec, matrix, rhs);
#include "../../../hermes_common/solver/umfpack_solver.h"
CSCMatrix *mm = static_cast<CSCMatrix*>(matrix);
UMFPackVector *mv = static_cast<UMFPackVector*>(rhs);
FILE *fp = fopen("data.m", "wt");
mm->dump(fp, "A");
mv->dump(fp, "b");
fclose(fp);
// Calculate the l2-norm of residual vector.
double res_l2_norm = get_l2_norm(rhs);
// Info for user.
info("---- Newton iter %d, ndof %d, res. l2 norm %g", it, Space::get_num_dofs(space), res_l2_norm);
// If l2 norm of the residual vector is within tolerance, then quit.
// NOTE: at least one full iteration forced
// here because sometimes the initial
// residual on fine mesh is too small.
if(res_l2_norm < NEWTON_TOL && it > 1) break;
// Multiply the residual vector with -1 since the matrix
// equation reads J(Y^n) \deltaY^{n+1} = -F(Y^n).
for(int i=0; i<ndof; i++) rhs->set(i, -rhs->get(i));
// Solve the linear system.
if(!solver->solve())
error ("Matrix solver failed.\n");
// Add \deltaY^{n+1} to Y^n.
for (int i = 0; i < ndof; i++) coeff_vec[i] += solver->get_solution()[i];
// If the maximum number of iteration has been reached, then quit.
if (it >= NEWTON_MAX_ITER) error ("Newton method did not converge.");
// Copy coefficients from vector y to elements.
set_coeff_vector(coeff_vec, space);
it++;
}
// Plot the solution.
Linearizer l(space);
l.plot_solution("solution.gp");
// Cleanup.
for(unsigned i = 0; i < DIR_BC_LEFT.size(); i++)
delete DIR_BC_LEFT[i];
DIR_BC_LEFT.clear();
for(unsigned i = 0; i < DIR_BC_RIGHT.size(); i++)
delete DIR_BC_RIGHT[i];
DIR_BC_RIGHT.clear();
delete matrix;
delete rhs;
delete solver;
delete[] coeff_vec;
delete dp;
delete space;
info("Done.");
return 0;
}
int main()
{
// Create space, set Dirichlet BC, enumerate basis functions.
Space* space = new Space(A, B, NELEM, DIR_BC_LEFT, DIR_BC_RIGHT, P_INIT, NEQ);
int ndof = Space::get_num_dofs(space);
info("ndof: %d", ndof);
// Initialize the weak formulation.
WeakForm wf;
wf.add_matrix_form(jacobian_vol);
wf.add_vector_form(residual_vol);
wf.add_vector_form_surf(0, residual_surf_right, BOUNDARY_RIGHT);
// Initialize the FE problem.
bool is_linear = false;
DiscreteProblem *dp = new DiscreteProblem(&wf, space, is_linear);
// Newton's loop.
// Fill vector coeff_vec using dof and coeffs arrays in elements.
double *coeff_vec = new double[Space::get_num_dofs(space)];
get_coeff_vector(space, coeff_vec);
// Set up the solver, matrix, and rhs according to the solver selection.
SparseMatrix* matrix = create_matrix(matrix_solver);
Vector* rhs = create_vector(matrix_solver);
Solver* solver = create_linear_solver(matrix_solver, matrix, rhs);
int it = 1;
while (1) {
// Obtain the number of degrees of freedom.
int ndof = Space::get_num_dofs(space);
// Assemble the Jacobian matrix and residual vector.
dp->assemble(coeff_vec, matrix, rhs);
// Calculate the l2-norm of residual vector.
double res_l2_norm = get_l2_norm(rhs);
// Info for user.
info("---- Newton iter %d, ndof %d, res. l2 norm %g", it, Space::get_num_dofs(space), res_l2_norm);
// If l2 norm of the residual vector is within tolerance, then quit.
// NOTE: at least one full iteration forced
// here because sometimes the initial
// residual on fine mesh is too small.
if(res_l2_norm < NEWTON_TOL && it > 1) break;
// Multiply the residual vector with -1 since the matrix
// equation reads J(Y^n) \deltaY^{n+1} = -F(Y^n).
for(int i=0; i<ndof; i++) rhs->set(i, -rhs->get(i));
// Solve the linear system.
if(!solver->solve())
error ("Matrix solver failed.\n");
// Add \deltaY^{n+1} to Y^n.
for (int i = 0; i < ndof; i++) coeff_vec[i] += solver->get_solution()[i];
// If the maximum number of iteration has been reached, then quit.
if (it >= NEWTON_MAX_ITER) error ("Newton method did not converge.");
// Copy coefficients from vector y to elements.
set_coeff_vector(coeff_vec, space);
it++;
}
// Plot the solution.
Linearizer l(space);
l.plot_solution("solution.gp");
// Plot the resulting space.
space->plot("space.gp");
info("Done.");
return 0;
}
int main()
{
// Time measurement.
TimePeriod cpu_time;
cpu_time.tick();
// Create coarse mesh, set Dirichlet BC, enumerate basis functions.
Space* space = new Space(A, B, NELEM, DIR_BC_LEFT, DIR_BC_RIGHT, P_INIT, NEQ);
// Enumerate basis functions, info for user.
int ndof = Space::get_num_dofs(space);
info("ndof: %d", ndof);
// Initialize the weak formulation.
WeakForm wf;
wf.add_matrix_form(jacobian);
wf.add_vector_form(residual);
// Initialize the FE problem.
bool is_linear = false;
DiscreteProblem *dp_coarse = new DiscreteProblem(&wf, space, is_linear);
if(JFNK == 0)
{
// Newton's loop on coarse mesh.
// Fill vector coeff_vec using dof and coeffs arrays in elements.
double *coeff_vec_coarse = new double[Space::get_num_dofs(space)];
get_coeff_vector(space, coeff_vec_coarse);
// Set up the solver, matrix, and rhs according to the solver selection.
SparseMatrix* matrix_coarse = create_matrix(matrix_solver);
Vector* rhs_coarse = create_vector(matrix_solver);
Solver* solver_coarse = create_linear_solver(matrix_solver, matrix_coarse, rhs_coarse);
int it = 1;
while (1)
{
// Obtain the number of degrees of freedom.
int ndof_coarse = Space::get_num_dofs(space);
// Assemble the Jacobian matrix and residual vector.
dp_coarse->assemble(coeff_vec_coarse, matrix_coarse, rhs_coarse);
// Calculate the l2-norm of residual vector.
double res_l2_norm = get_l2_norm(rhs_coarse);
// Info for user.
info("---- Newton iter %d, ndof %d, res. l2 norm %g", it, Space::get_num_dofs(space), res_l2_norm);
// If l2 norm of the residual vector is within tolerance, then quit.
// NOTE: at least one full iteration forced
// here because sometimes the initial
// residual on fine mesh is too small.
if(res_l2_norm < NEWTON_TOL_COARSE && it > 1) break;
// Multiply the residual vector with -1 since the matrix
// equation reads J(Y^n) \deltaY^{n+1} = -F(Y^n).
for(int i = 0; i < ndof_coarse; i++) rhs_coarse->set(i, -rhs_coarse->get(i));
// Solve the linear system.
if(!solver_coarse->solve())
error ("Matrix solver failed.\n");
// Add \deltaY^{n+1} to Y^n.
for (int i = 0; i < ndof_coarse; i++) coeff_vec_coarse[i] += solver_coarse->get_solution()[i];
// If the maximum number of iteration has been reached, then quit.
if (it >= NEWTON_MAX_ITER) error ("Newton method did not converge.");
// Copy coefficients from vector y to elements.
set_coeff_vector(coeff_vec_coarse, space);
it++;
}
// Cleanup.
delete matrix_coarse;
delete rhs_coarse;
delete solver_coarse;
delete [] coeff_vec_coarse;
}
else
jfnk_cg(dp_coarse, space, MATRIX_SOLVER_TOL, MATRIX_SOLVER_MAXITER,
JFNK_EPSILON, NEWTON_TOL_COARSE, NEWTON_MAX_ITER, matrix_solver);
// Cleanup.
delete dp_coarse;
// DOF and CPU convergence graphs.
SimpleGraph graph_dof_est, graph_cpu_est;
SimpleGraph graph_dof_exact, graph_cpu_exact;
// Adaptivity loop:
int as = 1;
double ftr_errors[MAX_ELEM_NUM]; // This array decides what
// elements will be refined.
bool done = false;
do
{
info("---- Adaptivity step %d:", as);
// Construct globally refined reference mesh and setup reference space.
Space* ref_space = construct_refined_space(space);
// Initialize the FE problem.
bool is_linear = false;
DiscreteProblem* dp = new DiscreteProblem(&wf, ref_space, is_linear);
if(JFNK == 0)
{
// Set up the solver, matrix, and rhs according to the solver selection.
SparseMatrix* matrix = create_matrix(matrix_solver);
Vector* rhs = create_vector(matrix_solver);
Solver* solver = create_linear_solver(matrix_solver, matrix, rhs);
// Newton's loop on the fine mesh.
info("Solving on fine mesh:");
// Fill vector coeff_vec using dof and coeffs arrays in elements.
double *coeff_vec = new double[Space::get_num_dofs(ref_space)];
get_coeff_vector(ref_space, coeff_vec);
int it = 1;
while (1)
{
// Obtain the number of degrees of freedom.
int ndof = Space::get_num_dofs(ref_space);
// Assemble the Jacobian matrix and residual vector.
dp->assemble(coeff_vec, matrix, rhs);
// Calculate the l2-norm of residual vector.
double res_l2_norm = get_l2_norm(rhs);
// Info for user.
info("---- Newton iter %d, ndof %d, res. l2 norm %g", it, Space::get_num_dofs(ref_space), res_l2_norm);
// If l2 norm of the residual vector is within tolerance, then quit.
// NOTE: at least one full iteration forced
// here because sometimes the initial
// residual on fine mesh is too small.
if(res_l2_norm < NEWTON_TOL_REF && it > 1) break;
// Multiply the residual vector with -1 since the matrix
// equation reads J(Y^n) \deltaY^{n+1} = -F(Y^n).
for(int i = 0; i < ndof; i++) rhs->set(i, -rhs->get(i));
// Solve the linear system.
if(!solver->solve())
error ("Matrix solver failed.\n");
// Add \deltaY^{n+1} to Y^n.
for (int i = 0; i < ndof; i++) coeff_vec[i] += solver->get_solution()[i];
// If the maximum number of iteration has been reached, then quit.
if (it >= NEWTON_MAX_ITER) error ("Newton method did not converge.");
// Copy coefficients from vector y to elements.
set_coeff_vector(coeff_vec, ref_space);
it++;
}
// Cleanup.
delete matrix;
delete rhs;
delete solver;
delete [] coeff_vec;
}
else
jfnk_cg(dp, ref_space, MATRIX_SOLVER_TOL, MATRIX_SOLVER_MAXITER,
JFNK_EPSILON, NEWTON_TOL_COARSE, NEWTON_MAX_ITER, matrix_solver);
// Cleanup.
delete dp;
// Starting with second adaptivity step, obtain new coarse
// mesh solution via projecting the fine mesh solution.
if(as > 1)
{
info("Projecting the fine mesh solution onto the coarse mesh.");
// Project the fine mesh solution (defined on space_ref) onto the coarse mesh (defined on space).
OGProjection::project_global(space, ref_space, matrix_solver);
}
double max_qoi_err_est = 0;
for (int i=0; i < space->get_n_active_elem(); i++)
{
if (GOAL_ORIENTED == 1)
{
// Use quantity of interest.
double qoi_est = quantity_of_interest(space, X_QOI);
double qoi_ref_est = quantity_of_interest(ref_space, X_QOI);
ftr_errors[i] = fabs(qoi_ref_est - qoi_est);
}
else
{
// Use global norm
double err_est_array[MAX_ELEM_NUM];
ftr_errors[i] = calc_err_est(NORM, space, ref_space, err_est_array);
}
// Info for user.
info("Elem [%d]: absolute error (est) = %g%%", i, ftr_errors[i]);
// Time measurement.
cpu_time.tick();
// Calculating maximum of QOI FTR error for plotting purposes
if (GOAL_ORIENTED == 1)
{
if (ftr_errors[i] > max_qoi_err_est)
max_qoi_err_est = ftr_errors[i];
}
else
{
double qoi_est = quantity_of_interest(space, X_QOI);
double qoi_ref_est = quantity_of_interest(ref_space, X_QOI);
double err_est = fabs(qoi_ref_est - qoi_est);
if (err_est > max_qoi_err_est)
max_qoi_err_est = err_est;
}
}
// Add entries to convergence graphs.
if (EXACT_SOL_PROVIDED)
{
double qoi_est = quantity_of_interest(space, X_QOI);
double u[MAX_EQN_NUM], dudx[MAX_EQN_NUM];
exact_sol(X_QOI, u, dudx);
double err_qoi_exact = fabs(u[0] - qoi_est);
// Info for user.
info("Relative error (exact) = %g %%", err_qoi_exact);
// Add entry to DOF and CPU convergence graphs.
graph_dof_exact.add_values(Space::get_num_dofs(space), err_qoi_exact);
graph_cpu_exact.add_values(cpu_time.accumulated(), err_qoi_exact);
}
// Add entry to DOF and CPU convergence graphs.
graph_dof_est.add_values(Space::get_num_dofs(space), max_qoi_err_est);
graph_cpu_est.add_values(cpu_time.accumulated(), max_qoi_err_est);
// Decide whether the max. FTR error in the quantity of interest
// is sufficiently small.
if(max_qoi_err_est < TOL_ERR_QOI) break;
// Returns updated coarse and fine meshes, with the last
// coarse and fine mesh solutions on them, respectively.
// The coefficient vectors and numbers of degrees of freedom
// on both meshes are also updated.
adapt(NORM, ADAPT_TYPE, THRESHOLD, ftr_errors, space, ref_space);
as++;
// Plot meshes, results, and errors.
adapt_plotting(space, ref_space, NORM, EXACT_SOL_PROVIDED, exact_sol);
// Cleanup.
delete ref_space;
}
while (done == false);
info("Total running time: %g s", cpu_time.accumulated());
// Save convergence graphs.
graph_dof_est.save("conv_dof_est.dat");
graph_cpu_est.save("conv_cpu_est.dat");
graph_dof_exact.save("conv_dof_exact.dat");
graph_cpu_exact.save("conv_cpu_exact.dat");
// Test variable.
bool success = true;
info("ndof = %d.", Space::get_num_dofs(space));
if (Space::get_num_dofs(space) > 150) success = false;
if (success)
{
info("Success!");
return ERROR_SUCCESS;
}
else
{
info("Failure!");
return ERROR_FAILURE;
}
}
int main() {
// Time measurement.
TimePeriod cpu_time;
cpu_time.tick();
// Create coarse mesh, set Dirichlet BC, enumerate basis functions.
Space* space = new Space(A, B, NELEM, DIR_BC_LEFT, DIR_BC_RIGHT, P_INIT, NEQ);
// Enumerate basis functions, info for user.
int ndof = Space::get_num_dofs(space);
info("ndof: %d", ndof);
// Initialize the weak formulation.
WeakForm wf;
wf.add_matrix_form(jacobian);
wf.add_vector_form(residual);
// Initialize the FE problem.
bool is_linear = false;
DiscreteProblem *dp_coarse = new DiscreteProblem(&wf, space, is_linear);
// Newton's loop on coarse mesh.
// Fill vector coeff_vec using dof and coeffs arrays in elements.
double *coeff_vec_coarse = new double[Space::get_num_dofs(space)];
get_coeff_vector(space, coeff_vec_coarse);
// Set up the solver, matrix, and rhs according to the solver selection.
SparseMatrix* matrix_coarse = create_matrix(matrix_solver);
Vector* rhs_coarse = create_vector(matrix_solver);
Solver* solver_coarse = create_linear_solver(matrix_solver, matrix_coarse, rhs_coarse);
int it = 1;
while (1) {
// Obtain the number of degrees of freedom.
int ndof_coarse = Space::get_num_dofs(space);
// Assemble the Jacobian matrix and residual vector.
dp_coarse->assemble(coeff_vec_coarse, matrix_coarse, rhs_coarse);
// Calculate the l2-norm of residual vector.
double res_l2_norm = get_l2_norm(rhs_coarse);
// Info for user.
info("---- Newton iter %d, ndof %d, res. l2 norm %g", it, Space::get_num_dofs(space), res_l2_norm);
// If l2 norm of the residual vector is within tolerance, then quit.
// NOTE: at least one full iteration forced
// here because sometimes the initial
// residual on fine mesh is too small.
if(res_l2_norm < NEWTON_TOL_COARSE && it > 1) break;
// Multiply the residual vector with -1 since the matrix
// equation reads J(Y^n) \deltaY^{n+1} = -F(Y^n).
for(int i=0; i < ndof_coarse; i++) rhs_coarse->set(i, -rhs_coarse->get(i));
// Solve the linear system.
if(!solver_coarse->solve())
error ("Matrix solver failed.\n");
// Add \deltaY^{n+1} to Y^n.
for (int i = 0; i < ndof_coarse; i++) coeff_vec_coarse[i] += solver_coarse->get_solution()[i];
// If the maximum number of iteration has been reached, then quit.
if (it >= NEWTON_MAX_ITER) error ("Newton method did not converge.");
// Copy coefficients from vector y to elements.
set_coeff_vector(coeff_vec_coarse, space);
it++;
}
// Cleanup.
delete matrix_coarse;
delete rhs_coarse;
delete solver_coarse;
delete [] coeff_vec_coarse;
delete dp_coarse;
// DOF and CPU convergence graphs.
SimpleGraph graph_dof_est, graph_cpu_est;
SimpleGraph graph_dof_exact, graph_cpu_exact;
// Adaptivity loop:
int as = 1;
bool done = false;
do
{
info("---- Adaptivity step %d:", as);
// Construct globally refined reference mesh and setup reference space.
Space* ref_space = construct_refined_space(space);
// Initialize the FE problem.
bool is_linear = false;
DiscreteProblem* dp = new DiscreteProblem(&wf, ref_space, is_linear);
// Set up the solver, matrix, and rhs according to the solver selection.
SparseMatrix* matrix = create_matrix(matrix_solver);
Vector* rhs = create_vector(matrix_solver);
Solver* solver = create_linear_solver(matrix_solver, matrix, rhs);
// Newton's loop on the fine mesh.
info("Solving on fine mesh:");
// Fill vector coeff_vec using dof and coeffs arrays in elements.
double *coeff_vec = new double[Space::get_num_dofs(ref_space)];
get_coeff_vector(ref_space, coeff_vec);
int it = 1;
while (1)
{
// Obtain the number of degrees of freedom.
int ndof = Space::get_num_dofs(ref_space);
// Assemble the Jacobian matrix and residual vector.
dp->assemble(coeff_vec, matrix, rhs);
// Calculate the l2-norm of residual vector.
double res_l2_norm = get_l2_norm(rhs);
// Info for user.
info("---- Newton iter %d, ndof %d, res. l2 norm %g", it, Space::get_num_dofs(ref_space), res_l2_norm);
// If l2 norm of the residual vector is within tolerance, then quit.
// NOTE: at least one full iteration forced
// here because sometimes the initial
// residual on fine mesh is too small.
if(res_l2_norm < NEWTON_TOL_REF && it > 1) break;
// Multiply the residual vector with -1 since the matrix
// equation reads J(Y^n) \deltaY^{n+1} = -F(Y^n).
for(int i = 0; i < ndof; i++) rhs->set(i, -rhs->get(i));
// Solve the linear system.
if(!solver->solve())
error ("Matrix solver failed.\n");
// Add \deltaY^{n+1} to Y^n.
for (int i = 0; i < ndof; i++) coeff_vec[i] += solver->get_solution()[i];
// If the maximum number of iteration has been reached, then quit.
if (it >= NEWTON_MAX_ITER) error ("Newton method did not converge.");
// Copy coefficients from vector y to elements.
set_coeff_vector(coeff_vec, ref_space);
it++;
}
// Starting with second adaptivity step, obtain new coarse
// mesh solution via projecting the fine mesh solution.
if(as > 1)
{
info("Projecting the fine mesh solution onto the coarse mesh.");
// Project the fine mesh solution (defined on space_ref) onto the coarse mesh (defined on space).
OGProjection::project_global(space, ref_space, matrix_solver);
}
// Calculate element errors and total error estimate.
info("Calculating error estimate.");
double err_est_array[MAX_ELEM_NUM];
double err_est_rel = calc_err_est(NORM, space, ref_space, err_est_array) * 100;
// Report results.
info("ndof_coarse: %d, ndof_fine: %d, err_est_rel: %g%%",
Space::get_num_dofs(space), Space::get_num_dofs(ref_space), err_est_rel);
// Time measurement.
cpu_time.tick();
// If exact solution available, also calculate exact error.
if (EXACT_SOL_PROVIDED)
{
// Calculate element errors wrt. exact solution.
double err_exact_rel = calc_err_exact(NORM, space, exact_sol, NEQ, A, B) * 100;
// Info for user.
info("Relative error (exact) = %g %%", err_exact_rel);
// Add entry to DOF and CPU convergence graphs.
graph_dof_exact.add_values(Space::get_num_dofs(space), err_exact_rel);
graph_cpu_exact.add_values(cpu_time.accumulated(), err_exact_rel);
}
// Add entry to DOF and CPU convergence graphs.
graph_dof_est.add_values(Space::get_num_dofs(space), err_est_rel);
graph_cpu_est.add_values(cpu_time.accumulated(), err_est_rel);
// If err_est_rel too large, adapt the mesh.
if (err_est_rel < NEWTON_TOL_REF) done = true;
else
{
info("Adapting the coarse mesh.");
adapt(NORM, ADAPT_TYPE, THRESHOLD, err_est_array, space, ref_space);
}
as++;
// Plot meshes, results, and errors.
adapt_plotting(space, ref_space, NORM, EXACT_SOL_PROVIDED, exact_sol);
// Cleanup.
delete solver;
delete matrix;
delete rhs;
delete ref_space;
delete dp;
delete [] coeff_vec;
}
while (done == false);
info("Total running time: %g s", cpu_time.accumulated());
// Save convergence graphs.
graph_dof_est.save("conv_dof_est.dat");
graph_cpu_est.save("conv_cpu_est.dat");
graph_dof_exact.save("conv_dof_exact.dat");
graph_cpu_exact.save("conv_cpu_exact.dat");
return 0;
}
int main()
{
// Create space, set Dirichlet BC, enumerate basis functions.
Space* space = new Space(A, B, NELEM, DIR_BC_LEFT, DIR_BC_RIGHT, P_INIT, NEQ);
int ndof = Space::get_num_dofs(space);
info("ndof: %d", ndof);
// Initialize the weak formulation.
WeakForm wf;
wf.add_matrix_form(jacobian);
wf.add_vector_form(residual);
// Initialize the FE problem.
DiscreteProblem *dp = new DiscreteProblem(&wf, space);
// Allocate coefficient vector.
double *coeff_vec = new double[ndof];
memset(coeff_vec, 0, ndof*sizeof(double));
// Set up the solver, matrix, and rhs according to the solver selection.
SparseMatrix* matrix = create_matrix(matrix_solver);
Vector* rhs = create_vector(matrix_solver);
Solver* solver = create_linear_solver(matrix_solver, matrix, rhs);
// Time stepping loop.
double current_time = 0.0;
while (current_time < T_FINAL)
{
// Newton's loop.
// Fill vector coeff_vec using dof and coeffs arrays in elements.
get_coeff_vector(space, coeff_vec);
int it = 1;
while (true)
{
// Assemble the Jacobian matrix and residual vector.
dp->assemble(coeff_vec, matrix, rhs);
// Calculate the l2-norm of residual vector.
double res_l2_norm = get_l2_norm(rhs);
// Info for user.
info("---- Newton iter %d, ndof %d, res. l2 norm %g", it, Space::get_num_dofs(space), res_l2_norm);
// If l2 norm of the residual vector is within tolerance, then quit.
// NOTE: at least one full iteration forced
// here because sometimes the initial
// residual on fine mesh is too small.
if(res_l2_norm < NEWTON_TOL && it > 1) break;
// Multiply the residual vector with -1 since the matrix
// equation reads J(Y^n) \deltaY^{n+1} = -F(Y^n).
for(int i=0; i<ndof; i++) rhs->set(i, -rhs->get(i));
// Solve the linear system.
if(!solver->solve())
error ("Matrix solver failed.\n");
// Add \deltaY^{n+1} to Y^n.
for (int i = 0; i < ndof; i++) coeff_vec[i] += solver->get_solution()[i];
// If the maximum number of iteration has been reached, then quit.
if (it >= NEWTON_MAX_ITER) error ("Newton method did not converge.");
// Copy coefficients from vector y to elements.
set_coeff_vector(coeff_vec, space);
it++;
}
// Plot the solution.
Linearizer l(space);
char filename[100];
sprintf(filename, "solution_%g.gp", current_time);
l.plot_solution(filename);
info("Solution %s written.", filename);
current_time += TAU;
}
// Plot the resulting space.
space->plot("space.gp");
// Cleaning
delete dp;
delete rhs;
delete solver;
delete[] coeff_vec;
delete space;
delete matrix;
info("Done.");
return 0;
}
