예제 #1
0
void 
GenericEpetraProblem::outputResults(const NOX::Solver::Generic& solver, 
                   Teuchos::ParameterList& printParams)
{
  // Output the parameter list
  NOX::Utils utils(printParams);
  if (utils.isPrintType(NOX::Utils::Parameters)) {
    cout << endl << "Final Parameters" << endl
	 << "****************" << endl;
    solver.getList().print(cout);
    cout << endl;
  }

  // Get the Epetra_Vector with the final solution from the solver
  const NOX::Epetra::Group& finalGroup = 
      dynamic_cast<const NOX::Epetra::Group&>(solver.getSolutionGroup());
  const Epetra_Vector& finalSolution = 
      (dynamic_cast<const NOX::Epetra::Vector&>
        (finalGroup.getX())).getEpetraVector();

  // Print solution
  char file_name[25];
  FILE *ifp;
  int NumMyElements = finalSolution.Map().NumMyElements();
  (void) sprintf(file_name, "output.%d",finalSolution.Map().Comm().MyPID());
  ifp = fopen(file_name, "w");
  for (int i=0; i<NumMyElements; i++)
    fprintf(ifp,"%d  %E\n",finalSolution.Map().MinMyGID()+i,finalSolution[i]);
  fclose(ifp);
}
StatusType NormWRMS::
checkStatus(const NOX::Solver::Generic& problem,
        NOX::StatusTest::CheckType checkType)
{
  if (checkType == NOX::StatusTest::None) {
    status = Unevaluated;
    value = 1.0e+12;
    return status;
  }

  status = Unconverged;

  const Abstract::Group& soln = problem.getSolutionGroup();
  const Abstract::Group& oldsoln = problem.getPreviousSolutionGroup();
  const Abstract::Vector& x = soln.getScaledX();

  // On the first iteration, the old and current solution are the same so
  // we should return the test as unconverged until there is a valid
  // old solution (i.e. the number of iterations is greater than zero).
  int niters = problem.getNumIterations();
  if (niters == 0)
  {
    status = Unconverged;
    value = 1.0e+12;
    return status;
  }

  // **** Begin check for convergence criteria #1 ****

  // Create the working vectors if this is the first time this
  // operator is called.
  if (Teuchos::is_null(u))
    u = x.clone(NOX::ShapeCopy);
  if (Teuchos::is_null(v))
    v = x.clone(NOX::ShapeCopy);

  // Create the weighting vector u = RTOL |x| + ATOL
  // |x| is evaluated at the old time step
  v->abs(oldsoln.getScaledX());
  if (atolIsScalar)
  {
    u->init(1.0);
    u->update(rtol, *v, atol);
  }
  else
  {
    u->update(rtol, *v, 1.0, *atolVec, 0.0);
  }

  // v = 1/u (elementwise)
  v->reciprocal(*u);

  // u = x - oldx (i.e., the update)
  u->update(1.0, x, -1.0, oldsoln.getScaledX(), 0.0);

  // u = Cp * u @ v (where @ represents an elementwise multiply)
  u->scale(*v);

  // Turn off implicit scaling of norm if the vector supports it
  Teuchos::RCP<NOX::Abstract::ImplicitWeighting> iw_u;
  iw_u = Teuchos::rcp_dynamic_cast<NOX::Abstract::ImplicitWeighting>(u,false);
  bool saved_status = false;
  if (nonnull(iw_u) && m_disable_implicit_weighting) {
    saved_status = iw_u->getImplicitWeighting();
    iw_u->setImplicitWeighting(false);
  }

  // tmp = factor * sqrt (u * u / N)
  value = u->norm() * factor / sqrt(static_cast<double>(u->length()));

  // Set the implicit scaling back to original value
  if (nonnull(iw_u) && m_disable_implicit_weighting)
    iw_u->setImplicitWeighting(saved_status);

  StatusType status1 = Unconverged;
  if (value < tolerance)
    status1 = Converged;


  // **** Begin check for convergence criteria #2 ****
  StatusType status2 = Unconverged;

  // Determine if the Generic solver is a LineSearchBased solver
  // If it is not then return a "Converged" status
  const Solver::Generic* test = 0;
  test = dynamic_cast<const Solver::LineSearchBased*>(&problem);
  if (test == 0)
  {
    status2 = Converged;
  }
  else
  {
    printCriteria2Info = true;
    computedStepSize =
      (dynamic_cast<const Solver::LineSearchBased*>(&problem))->getStepSize();

    if (computedStepSize >= alpha)
      status2 = Converged;
  }

  // **** Begin check for convergence criteria #3 ****

  // First time through, make sure the output parameter list exists.
  // Since the list is const, a sublist call to a non-existent sublist
  // throws an error.  Therefore we have to check the existence of each
  // sublist before we call it.
  const Teuchos::ParameterList& p = problem.getList();
  if (niters == 1) {
    if (p.isSublist("Direction")) {
      if (p.sublist("Direction").isSublist("Newton")) {
    if (p.sublist("Direction").sublist("Newton").isSublist("Linear Solver")) {
      if (p.sublist("Direction").sublist("Newton").sublist("Linear Solver").isSublist("Output")) {

        const Teuchos::ParameterList& list = p.sublist("Direction").sublist("Newton").sublist("Linear Solver").sublist("Output");

        if (Teuchos::isParameterType<double>(list, "Achieved Tolerance")) {

          printCriteria3Info = true;


        }
      }
    }
      }
    }
  }

  StatusType status3 = Converged;
  if (printCriteria3Info) {
    achievedTol = const_cast<Teuchos::ParameterList&>(problem.getList()).
      sublist("Direction").sublist("Newton").sublist("Linear Solver").
      sublist("Output").get("Achieved Tolerance", -1.0);
    status3 = (achievedTol <= beta) ? Converged : Unconverged;
  }


  // Determine status of test
  if ((status1 == Converged) &&
      (status2 == Converged) &&
      (status3 == Converged))
    status = Converged;

  return status;
}