choose_best_solver.cc

#include "drake/solvers/choose_best_solver.h"

#include "drake/common/never_destroyed.h"
#include "drake/solvers/clp_solver.h"
#include "drake/solvers/csdp_solver.h"
#include "drake/solvers/equality_constrained_qp_solver.h"
#include "drake/solvers/gurobi_solver.h"
#include "drake/solvers/ipopt_solver.h"
#include "drake/solvers/linear_system_solver.h"
#include "drake/solvers/moby_lcp_solver.h"
#include "drake/solvers/mosek_solver.h"
#include "drake/solvers/nlopt_solver.h"
#include "drake/solvers/osqp_solver.h"
#include "drake/solvers/scs_solver.h"
#include "drake/solvers/snopt_solver.h"

namespace drake {
namespace solvers {
namespace {

template <typename SomeSolver>
bool IsMatch(const MathematicalProgram& prog) {
  return SomeSolver::is_available() && SomeSolver::is_enabled() &&
      SomeSolver::ProgramAttributesSatisfied(prog);
}

}  // namespace

SolverId ChooseBestSolver(const MathematicalProgram& prog) {
  if (IsMatch<LinearSystemSolver>(prog)) {
    return LinearSystemSolver::id();
  } else if (IsMatch<EqualityConstrainedQPSolver>(prog)) {
    return EqualityConstrainedQPSolver::id();
  } else if (IsMatch<MosekSolver>(prog)) {
    // TODO(hongkai.dai@tri.global): based on my limited experience, Mosek is
    // faster than Gurobi for convex optimization problem. But we should run
    // a more thorough comparison.
    return MosekSolver::id();
  } else if (IsMatch<GurobiSolver>(prog)) {
    return GurobiSolver::id();
  } else if (IsMatch<ClpSolver>(prog)) {
    return ClpSolver::id();
  } else if (IsMatch<OsqpSolver>(prog)) {
    return OsqpSolver::id();
  } else if (IsMatch<MobyLCPSolver<double>>(prog)) {
    return MobyLcpSolverId::id();
  } else if (IsMatch<SnoptSolver>(prog)) {
    return SnoptSolver::id();
  } else if (IsMatch<IpoptSolver>(prog)) {
    return IpoptSolver::id();
  } else if (IsMatch<NloptSolver>(prog)) {
    return NloptSolver::id();
  } else if (IsMatch<CsdpSolver>(prog)) {
    return CsdpSolver::id();
  } else if (IsMatch<ScsSolver>(prog)) {
    // Use SCS as the last resort. SCS uses ADMM method, which converges fast to
    // modest accuracy quite fast, but then slows down significantly if the user
    // wants high accuracy.
    return ScsSolver::id();
  }
  throw std::invalid_argument(
      "There is no available solver for the optimization program");
}

const std::set<SolverId>& GetKnownSolvers() {
  static const never_destroyed<std::set<SolverId>> result{std::set<SolverId>{
      LinearSystemSolver::id(),
      EqualityConstrainedQPSolver::id(),
      MosekSolver::id(),
      GurobiSolver::id(),
      ClpSolver::id(),
      OsqpSolver::id(),
      MobyLcpSolverId::id(),
      SnoptSolver::id(),
      IpoptSolver::id(),
      NloptSolver::id(),
      CsdpSolver::id(),
      ScsSolver::id(),
  }};
  return result.access();
}

std::unique_ptr<SolverInterface> MakeSolver(const SolverId& id) {
  if (id == LinearSystemSolver::id()) {
    return std::make_unique<LinearSystemSolver>();
  } else if (id == EqualityConstrainedQPSolver::id()) {
    return std::make_unique<EqualityConstrainedQPSolver>();
  } else if (id == MosekSolver::id()) {
    return std::make_unique<MosekSolver>();
  } else if (id == GurobiSolver::id()) {
    return std::make_unique<GurobiSolver>();
  } else if (id == ClpSolver::id()) {
    return std::make_unique<ClpSolver>();
  } else if (id == OsqpSolver::id()) {
    return std::make_unique<OsqpSolver>();
  } else if (id == MobyLcpSolverId::id()) {
    return std::make_unique<MobyLCPSolver<double>>();
  } else if (id == SnoptSolver::id()) {
    return std::make_unique<SnoptSolver>();
  } else if (id == IpoptSolver::id()) {
    return std::make_unique<IpoptSolver>();
  } else if (id == NloptSolver::id()) {
    return std::make_unique<NloptSolver>();
  } else if (id == CsdpSolver::id()) {
    return std::make_unique<CsdpSolver>();
  } else if (id == ScsSolver::id()) {
    return std::make_unique<ScsSolver>();
  } else {
    throw std::invalid_argument("MakeSolver: no matching solver " + id.name());
  }
}

}  // namespace solvers
}  // namespace drake