/* F11JB_A1W_F C++ Header Example Program.
*
* Copyright 2024 Numerical Algorithms Group.
* Mark 30.2, 2024.
*/
#include <dco.hpp>
#include <iostream>
#include <nag.h>
#include <nagad.h>
#include <nagx04.h>
#include <stdio.h>
using namespace std;
extern "C"
{
static void NAG_CALL do_rcm(nag::ad::handle_t &ad_handle,
Integer & n,
Integer & nnz,
Integer irow[],
Integer icol[],
nagad_a1w_w_rtype a[],
nagad_a1w_w_rtype y[],
Integer istr[],
Integer perm_fwd[],
Integer perm_inv[],
Integer iwork[]);
}
int main()
{
int exit_status = 0;
nag::ad::handle_t ad_handle;
Integer ifail = 0;
cout << "F11JB_A1W_F C++ Header Example Program Results\n\n";
// Skip heading in data file
string mystr;
getline(cin, mystr);
// Read order of matrix and number of nonzero entries
Integer n, nnz;
cin >> n;
cin >> nnz;
Integer la = 3 * nnz;
Integer liwork = 2 * la + 7 * n + 1;
nagad_a1w_w_rtype *a = 0, *x = 0, *y = 0;
double * ar = 0, *yr = 0, *dxdy = 0;
Integer * icol = 0, *ipiv = 0, *irow = 0, *istr = 0, *iwork = 0;
Integer * perm_fwd = 0, *perm_inv = 0;
a = new nagad_a1w_w_rtype[la];
x = new nagad_a1w_w_rtype[n];
y = new nagad_a1w_w_rtype[n];
icol = new Integer[la];
ipiv = new Integer[n];
irow = new Integer[la];
istr = new Integer[n + 1];
iwork = new Integer[liwork];
perm_fwd = new Integer[n];
perm_inv = new Integer[n];
ar = new double[la];
yr = new double[n];
dxdy = new double[n * n];
// Read the matrix A
for (int i = 0; i < nnz; i++)
{
cin >> ar[i] >> irow[i] >> icol[i];
a[i] = ar[i];
}
// Create AD tape
dco::ga1s<double>::global_tape = dco::ga1s<double>::tape_t::create();
// Read the vector y
for (int i = 0; i < n; i++)
{
cin >> yr[i];
y[i] = yr[i];
dco::ga1s<double>::global_tape->register_variable(y[i]);
}
// Create AD configuration data object
ifail = 0;
// Calculate Cholesky factorization
Integer lfill = -1;
Integer nnzc, npivm;
nagad_a1w_w_rtype dscale, dtol;
dtol = 0.0;
dscale = 0.0;
// Compute reverse Cuthill-McKee permutation for bandwidth reduction
do_rcm(ad_handle, n, nnz, irow, icol, a, y, istr, perm_fwd, perm_inv, iwork);
ifail = 0;
nag::ad::f11ja(ad_handle, n, nnz, a, la, irow, icol, lfill, dtol, "N", dscale,
"M", ipiv, istr, nnzc, npivm, iwork, liwork, ifail);
// Check the output value of NPIVM
if (npivm > 0)
{
cout << " Factorization is not complete" << endl;
goto END;
}
// Solve P L D L^T P^T x = y
ifail = 0;
nag::ad::f11jb(ad_handle, n, a, la, irow, icol, ipiv, istr, "C", y, x, ifail);
// Output results
cout.setf(ios::scientific, ios::floatfield);
cout.precision(4);
cout << " Solution vector" << endl;
for (int i = 0; i < n; ++i)
{
cout.width(12);
cout << dco::value(x[perm_inv[i]]) << endl;
}
cout << "\n Derivatives calculated: First order adjoints\n";
cout << " Computational mode : algorithmic\n";
cout << "\n Derivatives of solution X w.r.t RHS Y:\n";
// Setup evaluation of derivatives via adjoints
for (int i = 0; i < n; i++)
{
// Reset adjoints, initialize derivative, and evaluate adjoint
dco::ga1s<double>::global_tape->zero_adjoints();
double inc = 1.0;
dco::derivative(x[perm_inv[i]]) += inc;
ifail = 0;
dco::ga1s<double>::global_tape->sparse_interpret() = true;
dco::ga1s<double>::global_tape->interpret_adjoint();
for (int j = 0; j < n; j++)
{
Integer k = i + j * n;
dxdy[k] = dco::derivative(y[perm_inv[j]]);
}
}
// Print derivatives
cout << endl;
NagError fail;
INIT_FAIL(fail);
x04cac(Nag_ColMajor, Nag_GeneralMatrix, Nag_NonUnitDiag, n, n, dxdy, n,
" dx_i/dy_j", 0, &fail);
END:
ifail = 0;
dco::ga1s<double>::tape_t::remove(dco::ga1s<double>::global_tape);
delete[] a;
delete[] x;
delete[] y;
delete[] icol;
delete[] ipiv;
delete[] irow;
delete[] istr;
delete[] iwork;
delete[] ar;
delete[] yr;
delete[] dxdy;
delete[] perm_fwd;
delete[] perm_inv;
return exit_status;
}
static void NAG_CALL do_rcm(nag::ad::handle_t &ad_handle,
Integer & n,
Integer & nnz,
Integer irow[],
Integer icol[],
nagad_a1w_w_rtype a[],
nagad_a1w_w_rtype y[],
Integer istr[],
Integer perm_fwd[],
Integer perm_inv[],
Integer iwork[])
{
logical lopts[5];
lopts[0] = 0;
lopts[1] = 0;
lopts[2] = 1;
lopts[3] = 1;
lopts[4] = 1;
nagad_a1w_w_rtype *rwork = 0;
Integer info[4], mask[1];
// SCS to CS, must add the upper triangle entries.
Integer j = nnz;
for (Integer i = 0; i < nnz; i++)
{
if (irow[i] > icol[i])
{
// strictly lower triangle, add the transposed
a[j] = a[i];
irow[j] = icol[i];
icol[j] = irow[i];
j++;
}
}
Integer nnz_cs = j;
// Reorder, CS to CCS, icolzp in istr
Integer ifail = 0;
nag::ad::f11za(ad_handle, n, nnz_cs, a, icol, irow, "F", "F", istr, iwork,
ifail);
// Calculate reverse Cuthill-McKee
ifail = 0;
nag::ad::f11ye(ad_handle, n, nnz_cs, istr, irow, lopts, mask, perm_fwd, info,
ifail);
// compute inverse perm, in perm_inv
for (int i = 0; i < n; i++)
{
perm_fwd[i] = perm_fwd[i] - 1;
perm_inv[perm_fwd[i]] = i;
}
// Apply permutation on column/row indices
Integer *iswapc = 0, *iswapr = 0;
iswapc = new Integer[nnz_cs];
iswapr = new Integer[nnz_cs];
for (int i = 0; i < nnz_cs; i++)
{
iswapc[i] = perm_inv[icol[i] - 1];
iswapr[i] = perm_inv[irow[i] - 1];
}
for (int i = 0; i < nnz_cs; i++)
{
icol[i] = iswapc[i] + 1;
irow[i] = iswapr[i] + 1;
}
delete[] iswapc;
delete[] iswapr;
// restrict to lower triangle, SCS format
// copying entries upwards
j = 0;
for (Integer i = 0; i < nnz_cs; i++)
{
if (irow[i] >= icol[i])
{
// non-upper triangle, bubble up
a[j] = a[i];
icol[j] = icol[i];
irow[j] = irow[i];
j++;
}
}
Integer nnz_scs = j;
// sort
ifail = 0;
nag::ad::f11zb(ad_handle, n, nnz_scs, a, irow, icol, "S", "K", istr, iwork,
ifail);
// permute rhs vector
rwork = new nagad_a1w_w_rtype[n];
for (int i = 0; i < n; i++)
{
rwork[i] = y[perm_fwd[i]];
}
for (int i = 0; i < n; i++)
{
y[i] = rwork[i];
}
delete[] rwork;
return;
}