/* F06PA_A1W_F C++ Header Example Program.
 *
 * Copyright 2021 Numerical Algorithms Group.
 * Mark 27.2, 2021.
 */

#include <dco.hpp>
#include <iostream>
#include <math.h>
#include <nag.h>
#include <nagad.h>
#include <nagx04.h>
#include <stdio.h>
#include <string>
using namespace std;

int main(void)
{
  int     exit_status = 0;
  void *  ad_handle   = 0;
  Integer ifail       = 0;

  cout << "F06PA_A1W_F C++ Header Example Program Results\n\n";
  // Skip heading in data file
  string mystr;
  getline(cin, mystr);

  // Read problem size and parameters
  Integer m, n;
  cin >> m >> n;
  double alphar, betar;
  cin >> alphar >> betar;
  nagad_a1w_w_rtype alpha, beta;
  alpha = alphar;
  beta  = betar;

  // Allocate arrays containing A and its factorized form, B
  // and the solution X.
  nagad_a1w_w_rtype *a = 0, *b = 0, *y = 0;
  double *           ar = 0, *br = 0, *yr = 0, *da = 0, *db = 0;
  a  = new nagad_a1w_w_rtype[m * n];
  b  = new nagad_a1w_w_rtype[m];
  y  = new nagad_a1w_w_rtype[n];
  ar = new double[m * n];
  br = new double[m];
  yr = new double[n];
  da = new double[n];
  db = new double[n];

  // Read the matrix A
  double dd;
  for (int i = 0; i < m; ++i)
    {
      for (int j = 0; j < n; ++j)
        {
          cin >> dd;
          Integer k = i + j * m;
          ar[k]     = dd;
          a[k]      = dd;
        }
    }
  for (int i = 0; i < m; ++i)
    {
      cin >> dd;
      br[i] = dd;
      b[i]  = dd;
    }
  for (int i = 0; i < n; ++i)
    {
      yr[i] = 1.0;
      y[i]  = 1.0;
    }

  // Create AD tape
  dco::ga1s<double>::global_tape = dco::ga1s<double>::tape_t::create();

  // Create AD configuration data object
  ifail = 0;
  nag::ad::x10aa(ad_handle, ifail);

  dco::ga1s<double>::global_tape->register_variable(alpha);
  dco::ga1s<double>::global_tape->register_variable(beta);

  for (int i = 0; i < n; ++i)
    {
      yr[i] = dco::value(y[i]);
    }

  //  y = beta*y + alpha*A^T*b
  ifail = 0;
  nag::ad::f06pa(ad_handle, "T", m, n, alpha, a, m, b, 1, beta, y, 1, ifail);

  for (int i = 0; i < n; ++i)
    {
      yr[i] = dco::value(y[i]);
    }

  cout << endl;
  NagError fail;
  INIT_FAIL(fail);
  x04cac(Nag_ColMajor, Nag_GeneralMatrix, Nag_NonUnitDiag, n, 1, yr, n,
         "  Array y after update", 0, &fail);

  cout << "\n\n Derivatives calculated: First order adjoints\n";
  cout << " Computational mode    : algorithmic\n";
  cout << "\n Derivatives of y w.r.t alpha and beta:\n";

  for (int i = 0; i < n; ++i)
    {
      dco::ga1s<double>::global_tape->zero_adjoints();
      double inc = 1.0;
      dco::derivative(y[i]) += inc;
      ifail                                              = 0;
      dco::ga1s<double>::global_tape->sparse_interpret() = true;
      dco::ga1s<double>::global_tape->interpret_adjoint();
      da[i] = dco::derivative(alpha);
      db[i] = dco::derivative(beta);
    }

  // Print derivatives
  cout << endl;
  x04cac(Nag_ColMajor, Nag_GeneralMatrix, Nag_NonUnitDiag, 1, n, da, 1,
         "           dy/dalpha", 0, &fail);
  cout << endl;
  x04cac(Nag_ColMajor, Nag_GeneralMatrix, Nag_NonUnitDiag, 1, n, db, 1,
         "           dy/dbeta", 0, &fail);

  // Remove computational data object and tape
  nag::ad::x10ab(ad_handle, ifail);
  dco::ga1s<double>::tape_t::remove(dco::ga1s<double>::global_tape);

  return exit_status;
}