finite_differences.cpp

/*
 *    This file is part of CasADi.
 *
 *    CasADi -- A symbolic framework for dynamic optimization.
 *    Copyright (C) 2010-2023 Joel Andersson, Joris Gillis, Moritz Diehl,
 *                            KU Leuven. All rights reserved.
 *    Copyright (C) 2011-2014 Greg Horn
 *
 *    CasADi is free software; you can redistribute it and/or
 *    modify it under the terms of the GNU Lesser General Public
 *    License as published by the Free Software Foundation; either
 *    version 3 of the License, or (at your option) any later version.
 *
 *    CasADi is distributed in the hope that it will be useful,
 *    but WITHOUT ANY WARRANTY; without even the implied warranty of
 *    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 *    Lesser General Public License for more details.
 *
 *    You should have received a copy of the GNU Lesser General Public
 *    License along with CasADi; if not, write to the Free Software
 *    Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
 *
 */
 
 
#include "finite_differences.hpp"
 
namespace casadi {
 
std::string to_string(FdMode v) {
  switch (v) {
  case FdMode::FORWARD: return "forward";
  case FdMode::BACKWARD: return "backward";
  case FdMode::CENTRAL: return "central";
  case FdMode::SMOOTHING: return "smoothing";
  default: break;
  }
  return "";
}
 
casadi_int n_fd_points(FdMode v) {
  switch (v) {
    case FdMode::FORWARD:
    case FdMode::BACKWARD: return 2;
    case FdMode::CENTRAL: return 3;
    case FdMode::SMOOTHING: return 5;
    default: break;
  }
  return -1;
}
 
casadi_int fd_offset(FdMode v) {
  switch (v) {
    case FdMode::FORWARD: return 0;
    case FdMode::BACKWARD:
    case FdMode::CENTRAL: return 1;
    case FdMode::SMOOTHING: return 2;
    default: break;
  }
  return -1;
}
 
FiniteDiff::FiniteDiff(const std::string& name, casadi_int n)
  : FunctionInternal(name), n_(n) {
}
 
FiniteDiff::~FiniteDiff() {
  clear_mem();
}
 
const Options FiniteDiff::options_
= {{&FunctionInternal::options_},
    {{"second_order_stepsize",
      {OT_DOUBLE,
      "Second order perturbation size [default: 1e-3]"}},
    {"h",
      {OT_DOUBLE,
      "Step size [default: computed from abstol]"}},
    {"h_max",
      {OT_DOUBLE,
      "Maximum step size [default 0]"}},
    {"h_min",
      {OT_DOUBLE,
      "Minimum step size [default inf]"}},
    {"smoothing",
      {OT_DOUBLE,
      "Smoothing regularization [default: machine precision]"}},
    {"reltol",
      {OT_DOUBLE,
      "Accuracy of function inputs [default: query object]"}},
    {"abstol",
      {OT_DOUBLE,
      "Accuracy of function outputs [default: query object]"}},
    {"u_aim",
      {OT_DOUBLE,
      "Target ratio of roundoff error to truncation error [default: 100.]"}},
    {"h_iter",
      {OT_INT,
      "Number of iterations to improve on the step-size "
      "[default: 1 if error estimate available, otherwise 0]"}},
    }
};
 
void FiniteDiff::init(const Dict& opts) {
  // Call the initialization method of the base class
  FunctionInternal::init(opts);
 
  // Default options
  h_min_ = 0;
  h_max_ = inf;
  m_.smoothing = eps;
  m_.reltol = derivative_of_->get_reltol();
  m_.abstol = derivative_of_->get_abstol();
  h_ = calc_stepsize(m_.abstol);
  u_aim_ = 100;
  h_iter_ = has_err() ? 1 : 0;
 
  // Read options
  for (auto&& op : opts) {
    if (op.first=="h") {
      h_ = op.second;
    } else if (op.first=="h_min") {
      h_min_ = op.second;
    } else if (op.first=="h_max") {
      h_max_ = op.second;
    } else if (op.first=="reltol") {
      m_.reltol = op.second;
    } else if (op.first=="abstol") {
      m_.abstol = op.second;
    } else if (op.first=="smoothing") {
      m_.smoothing = op.second;
    } else if (op.first=="u_aim") {
      u_aim_ = op.second;
    } else if (op.first=="h_iter") {
      h_iter_ = op.second;
    }
  }
 
  // Check h_iter for consistency
  if (h_iter_!=0 && !has_err()) {
    casadi_error("Perturbation size refinement requires an error estimate, "
    "which is not available for the class '" + class_name() + "'. "
    "Choose a different differencing scheme.");
  }
 
  // Allocate work vector for (perturbed) inputs and outputs
  n_z_ = derivative_of_.nnz_in();
  n_y_ = derivative_of_.nnz_out();
  alloc_res(n_pert(), true); // yk
  alloc_w((n_pert() + 3) * n_y_, true); // yk[:], y0, y, J
  alloc_w(n_z_, true); // z
 
  // Dimensions
  if (verbose_) {
    casadi_message("Finite differences (" + class_name() + ") with "
                    + str(n_z_) + " inputs, " + str(n_y_)
                    + " outputs and " + str(n_) + " directional derivatives.");
  }
 
  // Allocate sufficient temporary memory for function evaluation
  alloc(derivative_of_);
}
 
Sparsity FiniteDiff::get_sparsity_in(casadi_int i) {
  casadi_int n_in = derivative_of_.n_in(), n_out = derivative_of_.n_out();
  if (i<n_in) {
    // Non-differentiated input
    return derivative_of_.sparsity_in(i);
  } else if (i<n_in+n_out) {
    // Non-differentiated output
    return derivative_of_.sparsity_out(i-n_in);
  } else {
    // Seeds
    casadi_int ii = i - n_in - n_out;
    if (is_diff_in_[i]) {
      return repmat(derivative_of_.sparsity_in(ii), 1, n_);
    } else {
      return Sparsity(derivative_of_.size1_in(ii),
                      derivative_of_.size2_in(ii)*n_);
    }
  }
}
 
Sparsity FiniteDiff::get_sparsity_out(casadi_int i) {
  return repmat(derivative_of_.sparsity_out(i), 1, n_);
}
 
double FiniteDiff::get_default_in(casadi_int ind) const {
  if (ind<derivative_of_.n_in()) {
    return derivative_of_.default_in(ind);
  } else {
    return 0;
  }
}
 
size_t FiniteDiff::get_n_in() {
  return derivative_of_.n_in() + derivative_of_.n_out() + derivative_of_.n_in();
}
 
size_t FiniteDiff::get_n_out() {
  return derivative_of_.n_out();
}
 
std::string FiniteDiff::get_name_in(casadi_int i) {
  casadi_int n_in = derivative_of_.n_in(), n_out = derivative_of_.n_out();
  if (i<n_in) {
    return derivative_of_.name_in(i);
  } else if (i<n_in+n_out) {
    return "out_" + derivative_of_.name_out(i-n_in);
  } else {
    return "fwd_" + derivative_of_.name_in(i-n_in-n_out);
  }
}
 
std::string FiniteDiff::get_name_out(casadi_int i) {
  return "fwd_" + derivative_of_.name_out(i);
}
 
Function CentralDiff::get_forward(casadi_int nfwd, const std::string& name,
                                  const std::vector<std::string>& inames,
                                  const std::vector<std::string>& onames,
                                  const Dict& opts) const {
  // Commented out, does not work well
#if 0
  // The second order derivative is calculated as the backwards derivative
  // of the forward derivative, which is equivalent to central differences
  // of second order
  std::string f_name = "fd_" + name;
  Dict f_opts = {{"derivative_of", derivative_of_}};
  Function f = Function::create(new ForwardDiff(f_name, n_, h_), f_opts);
  // Calculate backwards derivative of f
  f_opts["derivative_of"] = f;
  return Function::create(new ForwardDiff(name, nfwd, -h_), f_opts);
#endif
  return Function::create(new CentralDiff(name, nfwd), opts);
}
 
int FiniteDiff::eval(const double** arg, double** res,
    casadi_int* iw, double* w, void* mem) const {
  setup(mem, arg, res, iw, w);
  // Shorthands
  casadi_int n_in = derivative_of_.n_in(), n_out = derivative_of_.n_out();
  casadi_int n_pert = this->n_pert();
 
  // Non-differentiated input
  const double** x0 = arg;
  arg += n_in;
 
  // Non-differentiated output
  double* y0 = w;
  for (casadi_int j=0; j<n_out; ++j) {
    const casadi_int nnz = derivative_of_.nnz_out(j);
    casadi_copy(*arg++, nnz, w);
    w += nnz;
  }
 
  // Forward seeds
  const double** seed = arg;
  arg += n_in;
 
  // Forward sensitivities
  double** sens = res;
  res += n_out;
 
  // Finite difference approximation
  double* J = w;
  w += n_y_;
 
  // Perturbed function values
  double** yk = res;
  res += n_pert;
  for (casadi_int j=0; j<n_pert; ++j) {
    yk[j] = w, w += n_y_;
  }
 
  // Setup arg and z for evaluation
  double *z = w;
  for (casadi_int j=0; j<n_in; ++j) {
    arg[j] = w;
    w += derivative_of_.nnz_in(j);
  }
 
  // Setup res and y for evaluation
  double *y = w;
  for (casadi_int j=0; j<n_out; ++j) {
    res[j] = w;
    w += derivative_of_.nnz_out(j);
  }
 
  // For all sensitivity directions
  for (casadi_int i=0; i<n_; ++i) {
    // Initial stepsize
    double h = h_;
    // Perform finite difference algorithm with different step sizes
    for (casadi_int iter=0; iter<1+h_iter_; ++iter) {
      // Calculate perturbed function values
      for (casadi_int k=0; k<n_pert; ++k) {
        // Perturb inputs
        casadi_int off = 0;
        for (casadi_int j=0; j<n_in; ++j) {
          casadi_int nnz = derivative_of_.nnz_in(j);
          casadi_copy(x0[j], nnz, z + off);
          if (seed[j] && is_diff_in_[j]) casadi_axpy(nnz, pert(k, h), seed[j] + i*nnz, z + off);
          off += nnz;
        }
        // Evaluate
        if (derivative_of_(arg, res, iw, w)) return 1;
        // Save outputs
        casadi_copy(y, n_y_, yk[k]);
      }
      // Finite difference calculation with error estimate
      double u = calc_fd(yk, y0, J, h);
      if (iter==h_iter_) break;
 
      // Update step size
      if (u < 0) {
        // Perturbation failed, try a smaller step size
        h /= u_aim_;
      } else {
        // Update h to get u near the target ratio
        h *= sqrt(u_aim_ / fmax(1., u));
      }
      // Make sure h stays in the range [h_min_,h_max_]
      h = fmin(fmax(h, h_min_), h_max_);
    }
 
    // Gather sensitivities
    casadi_int off = 0;
    for (casadi_int j=0; j<n_out; ++j) {
      casadi_int nnz = derivative_of_.nnz_out(j);
      if (sens[j]) casadi_copy(J + off, nnz, sens[j] + i*nnz);
      off += nnz;
    }
  }
  return 0;
}
 
double ForwardDiff::calc_fd(double** yk, double* y0, double* J, double h) const {
  return casadi_forward_diff_old(yk, y0, J, h, n_y_, &m_);
}
 
double CentralDiff::calc_fd(double** yk, double* y0, double* J, double h) const {
  return casadi_central_diff_old(yk, y0, J, h, n_y_, &m_);
}
 
void FiniteDiff::codegen_declarations(CodeGenerator& g) const {
  g.add_dependency(derivative_of_);
  g.add_auxiliary(CodeGenerator::AUX_FINITE_DIFF);
}
 
void FiniteDiff::codegen_body(CodeGenerator& g) const {
  // Shorthands
  casadi_int n_in = derivative_of_.n_in(), n_out = derivative_of_.n_out();
  casadi_int n_pert = this->n_pert();
 
  g.comment("Non-differentiated input");
  g.local("x0", "const casadi_real", "**");
  g << "x0 = arg, arg += " << n_in << ";\n";
 
  g.comment("Non-differentiated output");
  g.local("y0", "casadi_real", "*");
  g << "y0 = w;\n";
  for (casadi_int j=0; j<n_out; ++j) {
    const casadi_int nnz = derivative_of_.nnz_out(j);
    g << g.copy("*arg++", nnz, "w") << " w += " << nnz << ";\n";
  }
 
  g.comment("Forward seeds");
  g.local("seed", "const casadi_real", "**");
  g << "seed = arg, arg += " << n_in << ";\n";
 
  g.comment("Forward sensitivities");
  g.local("sens", "casadi_real", "**");
  g << "sens = res, res += " << n_out << ";\n";
 
  g.comment("Finite difference approximation");
  g.local("J", "casadi_real", "*");
  g << "J = w, w += " << n_y_ << ";\n";
 
  g.comment("Perturbed function value");
  g.local("yk", "casadi_real", "**");
  g << "yk = res, res += " << n_pert << ";\n";
  g.local("j", "casadi_int");
  g << "for (j=0; j<" << n_pert << "; ++j) yk[j] = w, w += " << n_y_ << ";\n";
 
  g.comment("Setup arg and z for evaluation");
  g.local("z", "casadi_real", "*");
  g << "z = w;\n";
  for (casadi_int j=0; j<n_in; ++j) {
    g << g.arg(j) << " = w, w += " << derivative_of_.nnz_in(j) << ";\n";
  }
 
  g.comment("Setup res and y for evaluation");
  g.local("y", "casadi_real", "*");
  g << "y = w;\n";
  for (casadi_int j=0; j<n_out; ++j) {
    g << g.res(j) << " = w, w += " << derivative_of_.nnz_out(j) << ";\n";
  }
 
  g.comment("For all sensitivity directions");
  g.local("i", "casadi_int");
  g << "for (i=0; i<" << n_ << "; ++i) {\n";
 
  g.comment("Initial stepsize");
  g.local("h", "casadi_real");
  g << "h = " << g.constant(h_) << ";\n";
 
  g.comment("Perform finite difference algorithm with different step sizes");
  g.local("iter", "casadi_int");
  g << "for (iter=0; iter<" << 1+h_iter_ << "; ++iter) {\n";
 
  g.comment("Calculate perturbed function values");
  g.local("k", "casadi_int");
  g << "for (k=0; k<" << n_pert << "; ++k) {\n";
 
  g.comment("Perturb inputs");
  casadi_int off=0;
  for (casadi_int j=0; j<n_in; ++j) {
    casadi_int nnz = derivative_of_.nnz_in(j);
    std::string s = "seed[" + str(j) + "]";
    g << g.copy("x0[" + str(j) + "]", nnz, "z+" + str(off)) << "\n";
    if (is_diff_in_[j]) {
      g << "if ("+s+") " << g.axpy(nnz, pert("k", "h"),
                                    s+"+i*"+str(nnz), "z+" + str(off)) << "\n";
    }
    off += nnz;
  }
 
  g.comment("Evaluate");
  g << "if (" << g(derivative_of_, "arg", "res", "iw", "w") << ") return 1;\n";
 
  g.comment("Save outputs");
  g << g.copy("y", n_y_, "yk[k]") << "\n";
 
  g << "}\n"; // for (k=0, ...)
 
  g.comment("Finite difference calculation with error estimate");
  g.local("u", "casadi_real");
  g.local("m", "const struct casadi_finite_diff_mem");
  g.init_local("m", "{" + g.constant(m_.reltol) + ", "
                        + g.constant(m_.abstol) + ", "
                        + g.constant(m_.smoothing) + "}");
  g << "u = " << calc_fd() << "(yk, y0, J, h, " << n_y_ << ", &m);\n";
  g << "if (iter==" << h_iter_ << ") break;\n";
 
  g.comment("Update step size");
  g << "if (u < 0) {\n";
  // Perturbation failed, try a smaller step size
  g << "h /= " << u_aim_ << ";\n";
  g << "} else {\n";
  // Update h to get u near the target ratio
  g << "h *= sqrt(" << u_aim_ << " / fmax(1., u));\n";
  g << "}\n";
  // Make sure h stays in the range [h_min_,h_max_]
  if (h_min_>0 || isfinite(h_max_)) {
    std::string h = "h";
    if (h_min_>0) h = "fmax(" + h + ", " + g.constant(h_min_) + ")";
    if (isfinite(h_max_)) h = "fmin(" + h + ", " + g.constant(h_max_) + ")";
    g << "h = " << h << ";\n";
  }
 
  g << "}\n"; // for (iter=0, ...)
 
  g.comment("Gather sensitivities");
  off = 0;
  for (casadi_int j=0; j<n_out; ++j) {
    casadi_int nnz = derivative_of_.nnz_out(j);
    std::string s = "sens[" + str(j) + "]";
    g << "if (" << s << ") " << g.copy("J+" + str(off), nnz, s + "+i*" + str(nnz)) << "\n";
    off += nnz;
  }
  g << "}\n"; // for (i=0, ...)
}
 
std::string Smoothing::pert(const std::string& k, const std::string& h) const {
  std::string sign = "(2*(" + k + "/2)-1)";
  std::string len = "(" + k + "%2+1)";
  return len + "*" + sign + "*" + h;
}
 
double Smoothing::pert(casadi_int k, double h) const {
  casadi_int sign = 2*(k/2)-1;
  casadi_int len = k%2+1;
  return static_cast<double>(len*sign)*h;
}
 
double Smoothing::calc_fd(double** yk, double* y0, double* J, double h) const {
  return casadi_smoothing_diff_old(yk, y0, J, h, n_y_, &m_);
}
 
Function ForwardDiff::get_forward(casadi_int nfwd, const std::string& name,
    const std::vector<std::string>& inames,
    const std::vector<std::string>& onames,
    const Dict& opts) const {
  return Function::create(new ForwardDiff(name, nfwd), opts);
}
 
Function Smoothing::get_forward(casadi_int nfwd, const std::string& name,
    const std::vector<std::string>& inames,
    const std::vector<std::string>& onames,
    const Dict& opts) const {
  return Function::create(new Smoothing(name, nfwd), opts);
}
 
} // namespace casadi