AR_Process::AR_Process(void)
{
// State Equation
F.resize(1,1);
F(0,0) = 0.8;
f.resize(1);
f(0) = 0.;
G.resize(1,1);
G.identity();
Qw.resize(1);
Qw(0,0)= 0.1;
// Observation noise
H.resize(1,1);
H(0,0) = 1;
h.resize(1);
h(0) = 0.;
Qv.resize(1);
Qv(0,0)=1;
// Init state
X0.resize(1);
X0(0) = 10.;
R0.resize(1);
R0.zero();
}
void l2r_rank_fun::grad(double *w, double *g)
{
int i;
int l=prob->l;
double *lg = new double[l];
double *tmp_vector = new double[l];
double *gtmp_vector = new double[global_l];
#pragma omp parallel for default(shared) private(i)
for (i=0;i<l;i++)
{
tmp_vector[i] = ATAQb[i] - ATe[i];
}
MPI_Allgatherv((void*)tmp_vector, l, MPI_DOUBLE, (void*)gtmp_vector, local_l, start_ptr, MPI_DOUBLE, MPI_COMM_WORLD);
Qv(gtmp_vector, lg);
MPI_Allgatherv((void*)lg, l, MPI_DOUBLE, (void*)g, local_l, start_ptr, MPI_DOUBLE, MPI_COMM_WORLD);
#pragma omp parallel for default(shared) private(i)
for(i=0;i<global_l;i++)
{
g[i] = gz[i] + 2*C*g[i];
}
delete[] tmp_vector;
delete[] gtmp_vector;
delete[] lg;
}
Van_Der_Pol::Van_Der_Pol(void)
{
lambda = 3.;
Qw.resize(1);
Qw(0,0)= 1.;
Qv.resize(1);
Qv(0,0)=0.1;
X0.resize(2);
X0(0) = 0.5;
X0(1) = 0.5;
R0.resize(2);
R0.zero();
R0(0,0)=0.;
R0(1,1)=.1;
Ts=.1;
}
void Vorticity::element_qoi_derivative( libMesh::DiffContext &context, const libMesh::QoISet & )
{
libMesh::FEMContext &c = libmesh_cast_ref<libMesh::FEMContext&>(context);
if( _subdomain_ids.find( (c.elem)->subdomain_id() ) != _subdomain_ids.end() )
{
// Element
libMesh::FEBase* element_fe;
c.get_element_fe<libMesh::Real>(this->_u_var, element_fe);
// Jacobian times integration weights
const std::vector<libMesh::Real> &JxW = element_fe->get_JxW();
// Grad of basis functions
const std::vector<std::vector<libMesh::RealGradient> >& du_phi =
c.element_fe_var[_u_var]->get_dphi();
const std::vector<std::vector<libMesh::RealGradient> >& dv_phi =
c.element_fe_var[_v_var]->get_dphi();
// Local DOF count and quadrature point count
const unsigned int n_T_dofs = c.dof_indices_var[0].size();
unsigned int n_qpoints = (c.get_element_qrule())->n_points();
// Warning: we assume here that vorticity is the only QoI!
// This should be consistent with the assertion in grins_mesh_adaptive_solver.C
/*! \todo Need to generalize this to the multiple QoI case */
libMesh::DenseSubVector<Number> &Qu = *c.elem_qoi_subderivatives[0][0];
libMesh::DenseSubVector<Number> &Qv = *c.elem_qoi_subderivatives[0][1];
// Integration loop
for( unsigned int qp = 0; qp != n_qpoints; qp++ )
{
for( unsigned int i = 0; i != n_T_dofs; i++ )
{
Qu(i) += - dv_phi[i][qp](1) * JxW[qp];
Qv(i) += du_phi[i][qp](0) * JxW[qp];
}
}
}
return;
}
void l2r_rank_fun::Hv(double *s, double *Hs)
{
int i,j,k;
int l=prob->l;
double *wa = new double[global_l];
double *lHs = new double[l];//
double *lwa = new double[l];//
selectiontree *T;
double *tmp_vector = new double[l];
double *gtmp_vector = new double[global_l];//
int tmp_value;
double gamma_tmp;
Qv(s, lwa);
MPI_Allgatherv((void*)lwa, l, MPI_DOUBLE, (void*)wa, local_l, start_ptr, MPI_DOUBLE, MPI_COMM_WORLD);//
#pragma omp parallel for private(i,j,k,T,gamma_tmp)
for (i=0;i<nr_subset;i++)
{
T=new selectiontree(nr_class[i]);
k=0;
for (j=0;j<count[i];j++)
{
while (k<count[i]&&(1-pi[i][j].value+pi[i][k].value>0))
{
T->insert_node(int_y[pi[i][k].id],lwa[pi[i][k].id]);
k++;
}
T->count_smaller(int_y[pi[i][j].id],&tmp_value, &tmp_vector[pi[i][j].id]);
}
delete T;
k=count[i]-1;
T = new selectiontree(nr_class[i]);
for (j=count[i]-1;j>=0;j--)
{
while (k>=0&&(1+pi[i][j].value-pi[i][k].value>0))
{
T->insert_node(int_y[pi[i][k].id],lwa[pi[i][k].id]);
k--;
}
T->count_larger(int_y[pi[i][j].id],&tmp_value, &gamma_tmp);
tmp_vector[pi[i][j].id] += gamma_tmp;
}
delete T;
}
#pragma omp parallel for default(shared) private(i)
for (i=0;i<l;i++)
{
tmp_vector[i]=wa[i+start_ptr[current_rank]]*((double)l_plus[i]+(double)l_minus[i])-tmp_vector[i];//
}
MPI_Allgatherv((void*)tmp_vector, l, MPI_DOUBLE, (void*)gtmp_vector, local_l, start_ptr, MPI_DOUBLE, MPI_COMM_WORLD);//
Qv(gtmp_vector, lHs);//
MPI_Allgatherv((void*)lHs, l, MPI_DOUBLE, (void*)Hs, local_l, start_ptr, MPI_DOUBLE, MPI_COMM_WORLD);//
#pragma omp parallel for default(shared) private(i)
for(i=0;i<global_l;i++)
{
Hs[i] = wa[i] + 2*C*Hs[i];//
}
delete[] wa;
delete[] lwa;//
delete[] tmp_vector;
delete[] gtmp_vector;
delete[] lHs;
}
double l2r_rank_fun::fun(double *w)// w is with the size of global_l
{
int i,j,k;
double f = 0.0;
double reg = 0.0;
int l=prob->l;
selectiontree *T;
Qv(w,z);
//generate gz via MPI_Allgatherv
MPI_Allgatherv((void*)z, l, MPI_DOUBLE, (void*)gz, local_l, start_ptr, MPI_DOUBLE, MPI_COMM_WORLD);
#pragma omp parallel for default(shared) private(i,j,k,T)
for (i=0;i<nr_subset;i++)
{
for (j=0;j<count[i];j++)
{
pi[i][j].id= perm[j+start[i]];
pi[i][j].value = z[perm[j+start[i]]];
}
qsort(pi[i], count[i], sizeof(id_and_value), compare_id_and_value);
T=new selectiontree(nr_class[i]);
k=0;
for (j=0;j<count[i];j++)
{
while (k<count[i]&&(1-pi[i][j].value+pi[i][k].value>0))
{
T->insert_node(int_y[pi[i][k].id],pi[i][k].value);
k++;
}
T->count_smaller(int_y[pi[i][j].id],&l_minus[pi[i][j].id], &gamma_minus[pi[i][j].id]);
}
delete T;
k=count[i]-1;
T = new selectiontree(nr_class[i]);
for (j=count[i]-1;j>=0;j--)
{
while (k>=0&&(1+pi[i][j].value-pi[i][k].value>0))
{
T->insert_node(int_y[pi[i][k].id],pi[i][k].value);
k--;
}
T->count_larger(int_y[pi[i][j].id],&l_plus[pi[i][j].id], &gamma_plus[pi[i][j].id]);
}
delete T;
}
//#pragma omp parallel for default(shared) private(i) reduction(+:f) schedule(dynamic)
for(i=0;i<global_l;i++)
{
f += w[i]*gz[i];
}
#pragma omp parallel for default(shared) private(i)
for (i=0;i<l;i++)
{
ATe[i] = l_minus[i] - l_plus[i];
ATAQb[i] = (l_plus[i]+l_minus[i])*gz[i+start_ptr[current_rank]]-gamma_plus[i]-gamma_minus[i];
}
//#pragma omp parallel for default(shared) //private(i) //reduction(+:reg) schedule(runtime)
for (int i=0;i<l;i++)
{
//#pragma omp atomic
reg += C*(gz[i+start_ptr[current_rank]]*(ATAQb[i] - 2 * ATe[i]) + l_minus[i]);
}
mpi_allreduce(®, 1, MPI_DOUBLE, MPI_SUM);
f /= 2.0;
f += reg;
si->obj=f;
return(f);
}
