int cgs_solver::solve (sp_matrix_t matrix, spv_double sp_rhs, spv_double sp_sol)
{
#ifdef _DEBUG
BOSOUT (section::solvers, level::debug) << "CGS\n" << bs_end;
#endif
BS_ERROR (sp_rhs->size (), "cgs_solve");
BS_ERROR (sp_sol->size (), "cgs_solve");
BS_ERROR (prop, "cgs_solve");
t_double rho_1, rho_2 = 1, alpha = 1, beta, sigma;
int iter;
const double epsmac = 1e-24;
t_double r_norm, b_norm, den_norm;
//fp_type *x = solution;
t_double *rhs = &(*sp_rhs)[0];
t_double *sol = &(*sp_sol)[0];
const t_double one = 1.0;
//OMP_TIME_MEASURE_START (cgs_solve_timer);
t_double tol = prop->get_f (tol_idx);
tol *= tol;
//resid = prop->get_residuals ();
//convergence_rate = prop->get_convergence_rate ();
int max_iter = prop->get_i (max_iters_idx);
t_long n = matrix->get_n_rows () * matrix->get_n_block_size ();
BS_ASSERT (n == (t_long)sp_sol->size ());
t_double *p = &(*sp_p)[0];
t_double *phat = &(*sp_phat)[0];
t_double *v = &(*sp_v)[0];
t_double *tmp = &(*sp_tmp)[0];
t_double *q = &(*sp_q)[0];
t_double *u = &(*sp_u)[0];
t_double *d = &(*sp_d)[0];
t_double *dhat = &(*sp_dhat)[0];
t_double *r = &(*sp_r)[0];
t_double *rtilde = &(*sp_rtilde)[0];
//t_double *r_old = &(*sp_r_old)[0];
prop->set_b (success_idx, false);
// solution = {0}
//assign (solution, n, 0);
memset (sol, 0, sizeof (t_double) * n);
//solution.assign (n, 0);
sp_p->resize (n);
sp_phat->resize (n);
sp_v->resize (n);
sp_tmp->resize (n);
sp_q->resize (n);
sp_u->resize (n);
sp_d->resize (n);
sp_dhat->resize (n);
sp_r->resize (n);
sp_rtilde->resize (n);
sp_r_old->resize (n);
// r = {0}
//r.assign (n, 0);
memset (r, 0, sizeof (t_double) * n);
memset (tmp, 0, sizeof (t_double) * n);
memset (p, 0, sizeof (t_double) * n);
memset (v, 0, sizeof (t_double) * n);
memset (q, 0, sizeof (t_double) * n);
// TODO:paste
//tmp.assign (n, 0);
//p.assign (n, 0);
//v.assign (n, 0);
//q.assign (n, 0);
// p0 = u0 = r0;
//u.assign (r.begin (), r.end ());
memcpy (u, r, sizeof (t_double) * n);
// TODO:end
// r = Ax0 - b
matrix->calc_lin_comb (-1.0, 1.0, sp_sol, sp_rhs, sp_r);
//rtilde.assign (r.begin (), r.end ());
memcpy (rtilde, r, sizeof (t_double) * n);
//tools::save_seq_vector (tools::string_formater ("1_well_bhp.%s.txt", it->first).str).save (it->second);
r_norm = mv_vector_inner_product_n (r, r, n);
if (r_norm <= tol) // initial guess quite good
return 0;
rho_1 = r_norm;
b_norm = sqrt (mv_vector_inner_product_n (rhs, rhs, n));
// TODO:delete
//p.assign (r.begin (), r.end ());
//rtilde.assign (r.begin (), r.end ());
//v.assign (n, 0);
// TODO:end
if (b_norm > epsmac) // choose convergence criterion
{
// |r_i|/|b| <= eps if |b| > 0
tol *= b_norm;
den_norm = b_norm;
}
else // (r_norm > epsmac)
{
// |r_i|/|r0| <= eps if |b| = 0
tol *= r_norm;
den_norm = r_norm;
}
// set up initial norm and convergense factor
//prop->set_relative_factor (den_norm);
// main loop
for (iter = 0; iter < max_iter; ++iter)
{
//printf ("CGS iteration: %d, resid = %le\n", iter, r_norm);
//fflush (stdout);
// TODO: paste
if (iter)
{
//rho_1 = (r,rtilde)
rho_1 = mv_vector_inner_product_n (r, rtilde, n); //in first iter equals to r_norm
if (rho_1 == 0) // failure
{
if (den_norm > epsmac)
prop->set_f (final_res_idx, r_norm / den_norm);
else
prop->set_f (final_res_idx, r_norm);
bs_throw_exception ("CGS: Failure - rho_1 == 0");
}
}
beta = rho_1/rho_2; // beta = rho_n/rho_n-1
rho_2 = rho_1;
// u = r + beta*q
sum_vector_n (r, one, q, beta, u, n);
// tmp = q+beta*p_old
sum_vector_n (q, one, p, beta, tmp, n);
// p_new = u + beta*tmp
sum_vector_n (u, one, tmp, beta, p, n);
//temp_p.assign (p.begin (), p.end ());
if (prec)
{
if (prec->solve_prec (matrix, sp_p, sp_phat))
{
bs_throw_exception ("CGS: Preconditioner failed");
}
}
else // no precondition (preconditioner=identity_matrix)
{
memcpy (phat, p, sizeof (t_double) * n);
//phat.assign (p.begin (), p.end ());
}
// v = A * phat = A * p, if no precondition;
//v.assign (n, 0);
memset (v, 0, sizeof (t_double) * n);
matrix->matrix_vector_product (sp_phat, sp_v);
// sigma = (v,rtilde)
sigma = mv_vector_inner_product_n (rtilde, v, n);
if (sigma > epsmac || sigma < -epsmac)
// alpha = rho_1/sigma
alpha = rho_1 / sigma;
else // failure
{
if (den_norm > epsmac)
prop->set_f (final_res_idx, r_norm / den_norm);
else
prop->set_f (final_res_idx, r_norm);
bs_throw_exception ("CGS: Failure - sigma == 0");
}
// q = u - alpha*v
sum_vector_n (u, one, v, -alpha, q, n);
// d = u + q
sum_vector_n (u, one, q, one, d, n);
// dhat = M^(-1) * d;
//temp_d.assign (d.begin (), d.end ());
if (prec)
{
if(prec->solve_prec (matrix, sp_d, sp_dhat))
{
bs_throw_exception ("CGS: Preconditioner failed");
}
}
else // no precondition (preconditioner=identity_matrix)
{
//dhat.assign (d.begin (), d.end ());
memcpy (dhat, d, sizeof (t_double) * n);
}
//tmp.assign (n, 0);
memset (tmp, 0, sizeof (t_double) * n);
// tmp = A*d
matrix->matrix_vector_product (sp_dhat, sp_tmp);
// r = r - alpha*tmp
sum_vector_n (r, one, tmp, -alpha, r, n);
// x = x + alpha*dhat
sum_vector_n (sol, one, dhat, alpha, sol, n);
r_norm = mv_vector_inner_product_n (r, r, n);
if (r_norm <= tol) // && check_resid_for_matbalance (n_rows, nb, r, matb_tol))
break;
}
//tools::save_seq_vector ("solution.txt").save(solution);
//TODO: end
prop->set_i (iters_idx, iter + 1);
prop->set_b (success_idx, true);
/*
//additional checking convergence
mv_calc_lin_comb (matrix, -1.0, 1.0, solution, rhs, r);
r_norm = mv_vector_inner_product (r, r, n);
*/
if (den_norm > epsmac)
prop->set_f (final_res_idx, r_norm / den_norm);
else
prop->set_f (final_res_idx, r_norm);
//printf ("CGS OK! iters = %d, resid = %le\n", lprop->iters, lprop->final_resid);
//OMP_TIME_MEASURE_END (bicgstab_solve_timer);
return 0;
}
int tfqmr_solver::solve (sp_matrix_t matrix, spv_double sp_rhs, spv_double sp_sol)
{
BOSOUT (section::solvers, level::debug) << "TFQMR\n" << bs_end;
BS_ERROR (matrix, "tfqmr_solve");
BS_ERROR (sp_rhs->size (), "tfqmr_solve");
BS_ERROR (sp_sol->size (), "tfqmr_solve");
BS_ERROR (prop, "tfqmr_solve");
t_double rho_1, rho_2 = 1, alpha = 1, beta, sigma;
int iter;
const double epsmac = 1e-24;
t_double r_norm, b_norm, den_norm, w_norm, eta, nu, tau, c;
//fp_type *x = solution;
//OMP_TIME_MEASURE_START (tfqmr_solve_timer);
t_double *rhs = &(*sp_rhs)[0];
t_double *sol = &(*sp_sol)[0];
t_double tol = prop->get_f (tol_idx);
tol *= tol;
//resid = prop->get_residuals ();
//convergence_rate = prop->get_convergence_rate ();
int max_iter = prop->get_i (max_iters_idx);
t_long n = matrix->get_n_rows () * matrix->get_n_block_size ();
sp_p->resize (n);
sp_v->resize (n);
sp_w->resize (n);
sp_u->resize (n);
sp_q->resize (n);
sp_d->resize (n);
sp_res->resize (n);
sp_r->resize (n);
sp_rtilde->resize (n);
sp_tmp->resize (n);
sp_rhat->resize (n);
sp_y->resize (n);
//x_cgs = y + n;
prop->set_b (success_idx, false);
// solution = {0}
sp_sol->assign (0);
// r = {0}
sp_r->assign (0);
// TODO:paste
sp_tmp->assign (0);
sp_p->assign (0);
sp_v->assign (0);
sp_q->assign (0);
sp_d->assign (0);
t_double *r = &(*sp_r)[0];
t_double *p = &(*sp_p)[0];
t_double *v = &(*sp_v)[0];
t_double *w = &(*sp_w)[0];
t_double *u = &(*sp_u)[0];
t_double *q = &(*sp_q)[0];
t_double *d = &(*sp_d)[0];
t_double *res = &(*sp_res)[0];
t_double *rtilde = &(*sp_rtilde)[0];
t_double *tmp = &(*sp_tmp)[0];
t_double *rhat = &(*sp_rhat)[0];
t_double *y = &(*sp_y)[0];
// r = Ax0 - b
matrix->calc_lin_comb (-1.0, 1.0, sp_sol, sp_rhs, sp_r);
memcpy (rtilde, r, sizeof (t_double) * n);
//rtilde.assign (r.begin (), r.end ());
// p0 = u0 = r0;
//memcpy (p, r, n * sizeof (double));
memcpy (u, r, sizeof (t_double) * n);
memcpy (p, r, sizeof (t_double) * n);
memcpy (w, r, sizeof (t_double) * n);
//u.assign (r.begin (), r.end ());
//p.assign (r.begin (), r.end ());
//w.assign (r.begin (), r.end ());
// tmp = M^(-1) * u;
if (prec)
{
if (prec->solve_prec (matrix, sp_u, sp_tmp))
{
bs_throw_exception ("TFQMR: Preconditioner failed");
}
memcpy (u, tmp, sizeof (t_double) * n);
memcpy (p, tmp, sizeof (t_double) * n);
//u.assign (tmp.begin (), tmp.end ());
//p.assign (u.begin (), u.end ());
}
matrix->matrix_vector_product (sp_p, sp_v);
//tools::save_seq_vector (tools::string_formater ("1_well_bhp.%s.txt", it->first).str).save (it->second);
r_norm = mv_vector_inner_product_n (r, r, n);
if (r_norm <= tol) // initial guess quite good
return 0;
tau = sqrt (r_norm);
rho_1 = r_norm;
rho_2 = r_norm;
b_norm = sqrt (mv_vector_inner_product_n (rhs, rhs, n));
if (b_norm > epsmac) // choose convergence criterion
{
// |r_i|/|b| <= eps if |b| > 0
tol *= b_norm;
den_norm = b_norm;
}
else // (r_norm > epsmac)
{
// |r_i|/|r0| <= eps if |b| = 0
tol *= r_norm;
den_norm = r_norm;
}
// set up initial norm and convergense factor
//prop->set_relative_factor (den_norm);
int m, count;
// main loop
for (iter = 0; iter < max_iter; ++iter)
{
//printf ("TFQMR iteration: %d, resid = %le\n", iter, r_norm);
//fflush (stdout);
// TODO: paste
if (iter)
{
//rho_1 = mv_vector_inner_product (r, rtilde, n);//in first iter equals to r_norm
if (rho_1 == 0) // failure
{
if (den_norm > epsmac)
prop->set_f (final_res_idx, r_norm / den_norm);
else
prop->set_f (final_res_idx, r_norm);
bs_throw_exception ("TFQMR: Failure - rho_1 == 0");
}
sum_vector_n (u, (t_double)1., res, beta, p, n); //p[n] = u[n]+beta*res
//v.assign (n, 0);
memset (v, 0, sizeof (t_double) * n);
matrix->matrix_vector_product (sp_p, sp_v); //v[n]=Ap[n]
}
sigma = mv_vector_inner_product_n (rtilde, v, n); //sigma=(rtilde,v[n-1])
alpha = rho_1/sigma;
// tmp = M^(-1)*v
if (prec)
{
if (prec->solve_prec (matrix, sp_v, sp_tmp))
{
bs_throw_exception ("TFQMR: Preconditioner failed");
}
memcpy (v, tmp, sizeof (t_double) * n);
//v.assign (tmp.begin (), tmp.end ());
}
sum_vector_n (u, (t_double)1., v, -alpha, q, n); //q[n] = u[n-1]-alpha*v[n-1]
sum_vector_n (u, (t_double)1., q, (t_double)1., res, n); //res = u[n-1]+q[n]
//tmp.assign (n, 0);
memset (tmp, 0, sizeof (t_double) * n);
matrix->matrix_vector_product (sp_res, sp_tmp);// tmp=A*res
sum_vector_n (r, (t_double)1., tmp, -alpha, r, n);// r=r-alpha*res
//r_norm_old = r_norm;
r_norm = mv_vector_inner_product_n (r, r, n);
for (m = 1; m <= 2 ; m++)
{
if (m == 1) // m is odd
{
memcpy (y, u, sizeof (t_double) * n);
//y.assign (u.begin (), u.end ());
w_norm = sqrt(r_norm * r_norm);
}
else // m is even
{
memcpy (y, q, sizeof (t_double) * n);
//y.assign (q.begin (), q.end ());
w_norm = sqrt(r_norm);
}
sum_vector_n (y, (t_double)1., d, eta*nu*nu/alpha, d, n); //d[m] = y[m] + (eta[m-1]*nu[m-1]^2/alpha[n-1])*d[m-1]
nu = w_norm/tau; //nu[m]=||w[m+1]||/tau[m-1]
c = 1./sqrt (1. + nu*nu);
tau = tau*c*nu; //tau[m]=tau[m-1]nu[m]c[m]
eta = c*c*alpha; //eta[m]=c[m]^2*alpha[n-1]
//SUM_VECTOR(x,d,1,alpha,x_cgs,k,n); //x_cgs[n] = x[2n-1]+alpha[n-1]*d[2n]
sum_vector_n (sol, (t_double)1., d, eta, sol, n); //x[m] = x[m-1]+eta[m]*d[m]
if (r_norm <= tol)
{
count = 1;
break;
}
}
if (r_norm <= tol)
break;
rho_1 = mv_vector_inner_product_n (r, rtilde, n);//in first iter equals to r_norm
beta = rho_1 / rho_2;
// rhat = M^(-1) * r;
if (prec)
{
if (prec->solve_prec (matrix, sp_r, sp_rhat))
{
bs_throw_exception ("TFQMR: Preconditioner failed");
}
}
else // no precondition (preconditioner=identity_matrix)
{
memcpy (rhat, r, sizeof (t_double) * n);
//rhat.assign (r.begin (), r.end ());
}
sum_vector_n (rhat, (t_double)1., q, beta, u, n); //u[n] = r[n]+beta*q[n]
sum_vector_n (q, (t_double)1., p, beta, res, n); //res = q[n]+beta*p[n-1]
rho_2 = rho_1;
}
//TODO: end
prop->set_i (iters_idx, iter + 1);
prop->set_b (success_idx, true);
/*
//additional checking convergence
mv_calc_lin_comb (matrix, -1.0, 1.0, solution, rhs, r);
r_norm = mv_vector_inner_product (r, r, n);
*/
if (den_norm > epsmac)
prop->set_f (final_res_idx, r_norm / den_norm);
else
prop->set_f (final_res_idx, r_norm);
//printf ("TFQMR OK! iters = %d, resid = %le\n", lprop->iters, lprop->final_resid);
//OMP_TIME_MEASURE_END (tfqmr_solve_timer);
return 0;
}
int
bcsr_ilu_prec::solve(sp_matrix_t matrix, spv_double sp_rhs, spv_double sp_sol)
{
BS_ASSERT (matrix);
BS_ASSERT (sp_rhs->size ());
BS_ASSERT (sp_sol->size ());
BS_ASSERT (sp_rhs->size () == sp_sol->size ()) (sp_rhs->size ()) (sp_sol->size ());
t_long b_sqr;
t_long n;
t_long nb;
t_double *rhs = &(*sp_rhs)[0];
t_double *sol = &(*sp_sol)[0];
sp_bcsr_matrix_t ilu;
if (dynamic_cast<bcsr_matrix_iface_t*> (matrix.lock ()))
{
ilu = matrix;
BS_ASSERT (ilu);
}
//const item_array_t &values = ilu.get_values ();
bool ff = prop->get_b (use_internal_matrix);
spv_float spvalues = ff ? lu_matrix->get_values () : ilu->get_values ();
spv_long sp_rows = ff ? lu_matrix->get_rows_ptr () : ilu->get_rows_ptr ();
spv_long sp_cols = ff ? lu_matrix->get_cols_ind () : ilu->get_cols_ind ();
t_float *values = &(*spvalues)[0];
t_long *rows = &(*sp_rows)[0];
t_long *cols = &(*sp_cols)[0];
if (ff)
{
n = lu_matrix->get_n_rows ();
nb = lu_matrix->get_n_block_size ();
}
else
{
n = ilu->get_n_rows ();
nb = ilu->get_n_block_size ();
}
b_sqr = nb * nb;
t_long j, k, l;
t_long j1, j2;
//const fp_type *D_block, *M_block;
const t_float *D_block, *M_block;
t_double *v, *r;
memcpy (sol, rhs, sp_rhs->size () * sizeof (rhs[0]));
#ifdef _DEBUG
t_double *sol_ = &sol[0];
#endif
// solve Ly = b
r = &sol[0];
for (k = 0; k < n; ++k, r += nb)
{
// pointer to matrix row
j1 = rows[k];
j2 = j1;//diag_ind[k];
t_long j3 = rows[k + 1];
for (j = j1 + 1; j < j3; ++j)
{
l = cols[j];
if (l < k)
{
v = &sol[l * nb];
M_block = &values[j * b_sqr];
LU_FIND_ROOT_UPDATE (nb, M_block, v, r);
}
else
break;
}
BS_ASSERT (j1 == j2);
BS_ASSERT (cols[j2] == k) (cols[j2]) (k);
D_block = &values[j2 * b_sqr];
uLU_FIND_ROOT_L (nb, D_block, r);
}
// solve Ux = y
r = &sol [(n - 1) * nb];
for (k = n - 1; k >= 0; --k, r -= nb)
{
// find element in k column
t_long j0 = rows[k];
j1 = j0;//diag_ind[k];
j2 = rows[k + 1];
for (j = j0; j < j2; ++j)
{
l = cols[j];
if (l > k)
{
v = &sol[l * nb];
M_block = &values[j * b_sqr];
LU_FIND_ROOT_UPDATE (nb, M_block, v, r);
}
}
BS_ASSERT (j0 == j1);
BS_ASSERT (cols[j1] == k) (cols[j1]) (k);
D_block = &values[j1 * b_sqr];
uLU_FIND_ROOT_U (nb, D_block, r);
}
return 0;
}
int gs_solver::solve (sp_matrix_t matrix, spv_double sp_rhs, spv_double sp_sol)
{
BS_ERROR (matrix, "gs_solve");
BS_ERROR (sp_rhs->size (), "gs_solve");
BS_ERROR (sp_sol->size (), "gs_solve");
BS_ERROR (prop, "gs_solve");
int iter;
const double epsmac = 1e-24;
t_double r_norm, b_norm, den_norm;
t_double *rhs = &(*sp_rhs)[0];
//t_double *sol = &(*sp_sol)[0];
sp_bcsr_t bcsr;
if (!dynamic_cast<bcsr_t *> (matrix.lock ()))
{
bcsr = matrix;
BS_ASSERT (bcsr);
}
t_long n = matrix->get_n_rows () * matrix->get_n_block_size ();
spv_long flags;
t_double tol = prop->get_f (tol_idx);
tol *= tol;
int max_iter = prop->get_i (max_iters_idx);
prop->reset_i (cf_type_idx);
prop->set_b (success_idx, false);
matrix->init_vector (sp_r);
t_double *r = &(*sp_r)[0];
matrix->calc_lin_comb (-1.0, 1.0, sp_sol, sp_rhs, sp_r);
r_norm = mv_vector_inner_product_n (r, r, n);
b_norm = sqrt (mv_vector_inner_product_n (rhs, rhs, n));
if (b_norm > epsmac) // choose convergence criterion
{
// |r_i|/|b| <= eps if |b| > 0
tol *= b_norm;
den_norm = b_norm;
}
else // (r_norm > epsmac)
{
// |r_i|/|r0| <= eps if |b| = 0
tol *= r_norm;
den_norm = r_norm;
}
// main loop
for (iter = 0; iter < max_iter; ++iter)
{
smooth (bcsr, flags, 1, sp_rhs, sp_sol);
matrix->calc_lin_comb (-1.0, 1.0, sp_sol, sp_rhs, sp_r);
r_norm = mv_vector_inner_product_n (r, r, n);
if (r_norm <= tol) // initial guess quite good
break;
} // end of main loop
if (iter < max_iter)
{
prop->set_i (iters_idx, iter + 1);
prop->set_b (success_idx, true);
}
else
{
prop->set_i (iters_idx, iter + 1);
prop->set_b (success_idx, false);
}
if (den_norm > epsmac)
prop->set_f(final_res_idx, r_norm / den_norm);
else
prop->set_f (final_res_idx, r_norm);
BOSOUT (section::solvers, level::low) << "r_norm = " << r_norm << " r_norm / den_norm = " << r_norm / den_norm << " iter = " << iter << bs_end;
return 0;
}
/**
* \brief setup preconditioner (merge matrices if needed)
*
* \param matrix Various matrix
* \return 0 if success
*/
int
bcsr_ilu_prec::setup (sp_matrix_t matrix)
{
BS_ASSERT (matrix);
t_long n;
t_long nb;
t_long b_sqr;
bool ff = prop->get_b (use_internal_matrix);
sp_bcsr_matrix_t ilu;
if (dynamic_cast<bcsr_matrix_iface_t *> (matrix.lock ()))
{
ilu = matrix;
BS_ASSERT (ilu);
}
if (ff)
{
if (!lu_matrix)
{
lu_matrix = BS_KERNEL.create_object (matrix->bs_resolve_type ());
if (!lu_matrix)
{
bs_throw_exception ("Can not create matrix");
}
}
if (lu_matrix->copy (ilu))
{
bs_throw_exception ("Can not make matrix copy");
}
n = lu_matrix->get_n_rows ();
nb = lu_matrix->get_n_block_size ();
}
else
{
n = ilu->get_n_rows ();
nb = ilu->get_n_block_size ();
}
spv_long sp_ilu_rows = ff ? lu_matrix->get_rows_ptr () : ilu->get_rows_ptr ();
spv_long sp_ilu_cols = ff ? lu_matrix->get_cols_ind () : ilu->get_cols_ind ();
spv_float sp_ilu_values = ff ? lu_matrix->get_values () : ilu->get_values ();
t_long *ilu_rows = &(*sp_ilu_rows)[0];
t_long *ilu_cols = &(*sp_ilu_cols)[0];
t_float *ilu_values = &(*sp_ilu_values)[0];
b_sqr = nb * nb;
// common
//int r_code = 0;
t_long i, j, j1, j2, j3, cl, k;
t_float *block;
//ilu.ascii_write_in_csr_format ("ilu_set_internal");
// ilu.write_matrix_to_file ("matrix.out");
// ----------------------
// STAGE 3: build ilu factorization
t_long i_str;
//fp_type *d_block, *dd_block;
t_float *d_block, *dd_block;
t_double d;
t_long jj, jj1, jj2;
// loop through all rows
for (i = 0; i < n; ++i)
{
// update i-th row by 0..i-1 row
j1 = ilu_rows[i];
j2 = j1;//ilu_diag_ind[i];
j3 = ilu_rows[i + 1];
BS_ASSERT (j1 == j2) (j1) (j2);
// for all nonzero elements in i-th string (in L part)
// update it by diagonal element and update all other nonzero elements
// in i-th string
for (j = j1 + 1; j < j3; ++j)
{
// find corresponding diagonal element
i_str = ilu_cols[j];
if (i_str < i)
{
// update by diagonal element
block = &ilu_values[j * b_sqr];
d_block = &ilu_values[ilu_rows[i_str]/*ilu_diag_ind[i_str]*/ * b_sqr];
//BS_ASSERT (ilu_diag_ind[i_str] == ilu_rows[i_str]) (ilu_diag_ind[i_str]) (ilu_rows[i_str]);
uLU_SEEK_L (nb, d_block, block, d);
jj1 = ilu_rows[i_str];
jj2 = ilu_rows[i_str + 1];
k = j + 1;
for (jj = jj1; jj < jj2; ++jj)
{
cl = ilu_cols[jj];
if (cl <= i_str)
continue;
if (ilu_cols[j1] == cl)
{
// upgrade by corresponding element
d_block = &ilu_values[jj * b_sqr];
dd_block = &ilu_values[j1 * b_sqr];
LU_UPGRADE(nb, block, d_block, dd_block);
}
for (; k < j3 && ilu_cols[k] < cl; ++k)
;
if (k < j3 && ilu_cols[k] == cl)
{
// upgrade by corresponding element
d_block = &ilu_values[jj * b_sqr];
dd_block = &ilu_values[k * b_sqr];
LU_UPGRADE(nb, block, d_block, dd_block);
}
else if (k >= j3)
break;
}
}
else
break;
}
// factorize i-th string
block = &ilu_values[j2 * b_sqr];
uLU (nb, block, d);
for (j = j1; j < j3; ++j)
{
t_long cl = ilu_cols[j];
if (cl > i)
{
d_block = &ilu_values[j * b_sqr];
uLU_SEEK_U (nb, block, d_block);
}
}
}
// ----------------------
return 0;
}