NAG Library Manual, Mark 30.1
Interfaces:  FL   CL   CPP   AD 

NAG AD Library Introduction
Example description
/* nag::ad::c05ay Adjoint Example Program.
 */

#include <dco.hpp>
#include <iostream>
#include <nagad.h>

// Function which calls NAG AD Library routines.
template <typename T> void func(T r, T &x);

// Driver with the adjoint calls.
// Finds the zero point f(xv) = 0 for the function
// f(x) =  exp(-x) - r*x.
// Also, computes the gradient dxdr = dxv/dr using a symbolic adjoint of the
// root finder.
void driver(const double &rv, double &xv, double &dxdr);

int main()
{
  std::cout << " nag::ad::c05ay Adjoint Example Program Results\n";

  // Set problem parameters
  double rv = 2.0;
  // Solution x
  double xv;
  // Derivative of x
  double dxdr;

  // Call driver
  driver(rv, xv, dxdr);

  // sPrint outputs
  std::cout << "\n Derivatives calculated: First order adjoints\n";
  std::cout << " Computational mode    : symbolic (expert mode)\n";

  // Print derivatives
  std::cout.setf(std::ios::scientific, std::ios::floatfield);
  std::cout.precision(12);
  std::cout << "\n Solution:\n";
  std::cout << " x = " << xv << std::endl;
  std::cout << "\n Derivative of solution x w.r.t. parameter r:\n";
  std::cout << " dx/dr(x) = " << dxdr << std::endl;

  return 0;
}

// Driver with the adjoint calls.
// Finds the zero point f(xv) = 0 for the function
// f(x) =  exp(-x) - r*x.
// Also, computes the gradient dxdr = dxv/dr using a symbolic adjoint of the
// root finder
void driver(const double &rv, double &xv, double &dxdr)
{
  using T = dco::ga1s<double>::type;

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

  // Function parameter r
  T r = rv;

  // Register variables to differentiate w.r.t.
  dco::ga1s<double>::global_tape->register_variable(r);

  // Variable to differentiate
  T x;

  // Call the NAG AD Lib functions
  func(r, x);
  // Extract the computed solution
  xv = dco::value(x);

  dco::ga1s<double>::global_tape->register_output_variable(x);
  dco::derivative(x) = 1.0;
  dco::ga1s<double>::global_tape->interpret_adjoint();

  // Extract the derivatives
  dxdr = dco::derivative(r);

  // Remove tape
  dco::ga1s<double>::tape_t::remove(dco::ga1s<double>::global_tape);
}

// Function which calls NAG AD Library routines
template <typename T> void func(T r, T &x)
{
  // Active variables
  T a = 0.0, b = 1.0;
  T eps = 1.0e-5, eta = 0.0;
  // Create AD configuration data object
  Integer           ifail = 0;
  nag::ad::handle_t ad_handle;

  auto f = [&](nag::ad::handle_t const& handle, T const &x, T & z) {
             switch (handle.active_inputs())
               {
               case nag::ad::none:
                 z = exp(-dco::value(x)) - dco::value(x) * dco::value(r);
                 break;
               case nag::ad::state:
                 z = exp(-x) - x * dco::value(r);
                 break;
               case nag::ad::params:
                 z = exp(-dco::value(x)) - dco::value(x) * r;
                 break;
               case nag::ad::all:
                 z = exp(-x) - x * r;
                 break;
               }
           };

  // Set computational mode
  ad_handle.set_strategy(nag::ad::symbolic_expert);
  // Routine for computing an approximation to a simple zero of the function f
  ifail = 0;
  nag::ad::c05ay(ad_handle, a, b, eps, eta, f, x, ifail);
}