hyperbolic_system.h

Go to the documentation of this file.

//
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
// Copyright (C) 2020 - 2024 by the ryujin authors
//
 
#pragma once
 
#include <compile_time_options.h>
 
#include "equation_of_state_library.h"
 
#include <convenience_macros.h>
#include <discretization.h>
#include <multicomponent_vector.h>
#include <openmp.h>
#include <patterns_conversion.h>
#include <simd.h>
#include <state_vector.h>
 
#include <deal.II/base/parameter_acceptor.h>
#include <deal.II/base/tensor.h>
 
#include <array>
 
namespace ryujin
{
  namespace EulerAEOS
  {
    /*
     * For various divisions in the arbirtray equation of state module we
     * have a mathematical guarantee that the numerator and denominator are
     * nonnegative and the limit (of zero numerator and denominator) must
     * converge to zero. The following function takes care of rounding
     * issues when computing such quotients by (a) avoiding division by
     * zero and (b) ensuring non-negativity of the result.
     */
    template <typename Number>
    DEAL_II_ALWAYS_INLINE inline Number safe_division(const Number &numerator,
                                                      const Number &denominator)
    {
      using ScalarNumber = typename get_value_type<Number>::type;
      constexpr ScalarNumber min = std::numeric_limits<ScalarNumber>::min();
 
      return std::max(numerator, Number(0.)) /
             std::max(denominator, Number(min));
    }
 
 
    template <int dim, typename Number>
    class HyperbolicSystemView;
 
    class HyperbolicSystem final : public dealii::ParameterAcceptor
    {
    public:
      static inline std::string problem_name =
          "Compressible Euler equations (arbitrary EOS)";
 
      HyperbolicSystem(const std::string &subsection = "/HyperbolicSystem");
 
      template <int dim, typename Number>
      auto view() const
      {
        return HyperbolicSystemView<dim, Number>{*this};
      }
 
    private:
 
      std::string equation_of_state_;
      double reference_density_;
      double vacuum_state_relaxation_small_;
      double vacuum_state_relaxation_large_;
      bool compute_strict_bounds_;
 
      EquationOfStateLibrary::equation_of_state_list_type
          equation_of_state_list_;
 
      using EquationOfState = EquationOfStateLibrary::EquationOfState;
      std::shared_ptr<EquationOfState> selected_equation_of_state_;
 
      template <int dim, typename Number>
      friend class HyperbolicSystemView;
    }; /* HyperbolicSystem */
 
 
    template <int dim, typename Number>
    class HyperbolicSystemView
    {
    public:
      HyperbolicSystemView(const HyperbolicSystem &hyperbolic_system)
          : hyperbolic_system_(hyperbolic_system)
      {
      }
 
      template <int dim2, typename Number2>
      auto view() const
      {
        return HyperbolicSystemView<dim2, Number2>{hyperbolic_system_};
      }
 
      using ScalarNumber = typename get_value_type<Number>::type;
 
 
      DEAL_II_ALWAYS_INLINE inline const std::string &equation_of_state() const
      {
        return hyperbolic_system_.equation_of_state_;
      }
 
      DEAL_II_ALWAYS_INLINE inline ScalarNumber reference_density() const
      {
        return hyperbolic_system_.reference_density_;
      }
 
      DEAL_II_ALWAYS_INLINE inline ScalarNumber
      vacuum_state_relaxation_small() const
      {
        return hyperbolic_system_.vacuum_state_relaxation_small_;
      }
 
      DEAL_II_ALWAYS_INLINE inline ScalarNumber
      vacuum_state_relaxation_large() const
      {
        return hyperbolic_system_.vacuum_state_relaxation_large_;
      }
 
      DEAL_II_ALWAYS_INLINE inline bool compute_strict_bounds() const
      {
        return hyperbolic_system_.compute_strict_bounds_;
      }
 
 
 
      DEAL_II_ALWAYS_INLINE inline Number eos_pressure(const Number &rho,
                                                       const Number &e) const
      {
        const auto &eos = hyperbolic_system_.selected_equation_of_state_;
 
        if constexpr (std::is_same_v<ScalarNumber, Number>) {
          return ScalarNumber(eos->pressure(rho, e));
        } else {
          Number p;
          for (unsigned int k = 0; k < Number::size(); ++k) {
            p[k] = ScalarNumber(eos->pressure(rho[k], e[k]));
          }
          return p;
        }
      }
 
      DEAL_II_ALWAYS_INLINE inline Number
      eos_specific_internal_energy(const Number &rho, const Number &p) const
      {
        const auto &eos = hyperbolic_system_.selected_equation_of_state_;
 
        if constexpr (std::is_same_v<ScalarNumber, Number>) {
          return ScalarNumber(eos->specific_internal_energy(rho, p));
        } else {
          Number e;
          for (unsigned int k = 0; k < Number::size(); ++k) {
            e[k] = ScalarNumber(eos->specific_internal_energy(rho[k], p[k]));
          }
          return e;
        }
      }
 
      DEAL_II_ALWAYS_INLINE inline Number eos_temperature(const Number &rho,
                                                          const Number &e) const
      {
        const auto &eos = hyperbolic_system_.selected_equation_of_state_;
 
        if constexpr (std::is_same_v<ScalarNumber, Number>) {
          return ScalarNumber(eos->temperature(rho, e));
        } else {
          Number temp;
          for (unsigned int k = 0; k < Number::size(); ++k) {
            temp[k] = ScalarNumber(eos->temperature(rho[k], e[k]));
          }
          return temp;
        }
      }
 
      DEAL_II_ALWAYS_INLINE inline Number
      eos_speed_of_sound(const Number &rho, const Number &e) const
      {
        const auto &eos = hyperbolic_system_.selected_equation_of_state_;
 
        if constexpr (std::is_same_v<ScalarNumber, Number>) {
          return ScalarNumber(eos->speed_of_sound(rho, e));
        } else {
          Number c;
          for (unsigned int k = 0; k < Number::size(); ++k) {
            c[k] = ScalarNumber(eos->speed_of_sound(rho[k], e[k]));
          }
          return c;
        }
      }
 
      DEAL_II_ALWAYS_INLINE inline ScalarNumber eos_interpolation_b() const
      {
        const auto &eos = hyperbolic_system_.selected_equation_of_state_;
        return ScalarNumber(eos->interpolation_b());
      }
 
      DEAL_II_ALWAYS_INLINE inline ScalarNumber eos_interpolation_pinfty() const
      {
        const auto &eos = hyperbolic_system_.selected_equation_of_state_;
        return ScalarNumber(eos->interpolation_pinfty());
      }
 
      DEAL_II_ALWAYS_INLINE inline ScalarNumber eos_interpolation_q() const
      {
        const auto &eos = hyperbolic_system_.selected_equation_of_state_;
        return ScalarNumber(eos->interpolation_q());
      }
 
 
      static constexpr bool have_gamma = false;
 
      static constexpr bool have_eos_interpolation_b = true;
 
 
 
 
    private:
      const HyperbolicSystem &hyperbolic_system_;
 
    public:
 
 
      static constexpr unsigned int problem_dimension = 2 + dim;
 
      using state_type = dealii::Tensor<1, problem_dimension, Number>;
 
      using flux_type =
          dealii::Tensor<1, problem_dimension, dealii::Tensor<1, dim, Number>>;
 
      using flux_contribution_type = flux_type;
 
      static inline const auto component_names =
          []() -> std::array<std::string, problem_dimension> {
        if constexpr (dim == 1)
          return {"rho", "m", "E"};
        else if constexpr (dim == 2)
          return {"rho", "m_1", "m_2", "E"};
        else if constexpr (dim == 3)
          return {"rho", "m_1", "m_2", "m_3", "E"};
        __builtin_trap();
      }();
 
      static inline const auto primitive_component_names =
          []() -> std::array<std::string, problem_dimension> {
        if constexpr (dim == 1)
          return {"rho", "v", "e"};
        else if constexpr (dim == 2)
          return {"rho", "v_1", "v_2", "e"};
        else if constexpr (dim == 3)
          return {"rho", "v_1", "v_2", "v_3", "e"};
        __builtin_trap();
      }();
 
      static constexpr unsigned int n_precomputed_values = 4;
 
      using precomputed_type = std::array<Number, n_precomputed_values>;
 
      static inline const auto precomputed_names =
          std::array<std::string, n_precomputed_values>{
              {"p",
               "surrogate_gamma_min",
               "surrogate_specific_entropy",
               "surrogate_harten_entropy"}};
 
      static constexpr unsigned int n_initial_precomputed_values = 0;
 
      using initial_precomputed_type =
          std::array<Number, n_initial_precomputed_values>;
 
      static inline const auto initial_precomputed_names =
          std::array<std::string, n_initial_precomputed_values>{};
 
      using StateVector = Vectors::
          StateVector<ScalarNumber, problem_dimension, n_precomputed_values>;
 
      using HyperbolicVector =
          Vectors::MultiComponentVector<ScalarNumber, problem_dimension>;
 
      using PrecomputedVector =
          Vectors::MultiComponentVector<ScalarNumber, n_precomputed_values>;
 
      using InitialPrecomputedVector =
          Vectors::MultiComponentVector<ScalarNumber,
                                        n_initial_precomputed_values>;
 
 
 
      static constexpr unsigned int n_precomputation_cycles = 2;
 
      template <typename DISPATCH, typename SPARSITY>
      void precomputation_loop(unsigned int cycle,
                               const DISPATCH &dispatch_check,
                               const SPARSITY &sparsity_simd,
                               StateVector &state_vector,
                               unsigned int left,
                               unsigned int right,
                               const bool skip_constrained_dofs = true) const;
 
 
 
      static Number density(const state_type &U);
 
      Number filter_vacuum_density(const Number &rho) const;
 
      static dealii::Tensor<1, dim, Number> momentum(const state_type &U);
 
      static Number total_energy(const state_type &U);
 
      static Number internal_energy(const state_type &U);
 
      static state_type internal_energy_derivative(const state_type &U);
 
 
 
      Number surrogate_specific_entropy(const state_type &U,
                                        const Number &gamma_min) const;
 
      Number surrogate_harten_entropy(const state_type &U,
                                      const Number &gamma_min) const;
 
      state_type
      surrogate_harten_entropy_derivative(const state_type &U,
                                          const Number &eta,
                                          const Number &gamma_min) const;
 
      Number surrogate_gamma(const state_type &U, const Number &p) const;
 
      Number surrogate_pressure(const state_type &U, const Number &gamma) const;
 
      Number surrogate_speed_of_sound(const state_type &U,
                                      const Number &gamma) const;
 
      bool is_admissible(const state_type &U) const;
 
 
 
      template <int component>
      state_type prescribe_riemann_characteristic(
          const state_type &U,
          const Number &p,
          const state_type &U_bar,
          const Number &p_bar,
          const dealii::Tensor<1, dim, Number> &normal) const;
 
      template <typename Lambda>
      state_type
      apply_boundary_conditions(const dealii::types::boundary_id id,
                                const state_type &U,
                                const dealii::Tensor<1, dim, Number> &normal,
                                const Lambda &get_dirichlet_data) const;
 
 
 
      flux_type f(const state_type &U, const Number &p) const;
 
      flux_contribution_type
      flux_contribution(const PrecomputedVector &pv,
                        const InitialPrecomputedVector &piv,
                        const unsigned int i,
                        const state_type &U_i) const;
 
      flux_contribution_type
      flux_contribution(const PrecomputedVector &pv,
                        const InitialPrecomputedVector &piv,
                        const unsigned int *js,
                        const state_type &U_j) const;
 
      state_type
      flux_divergence(const flux_contribution_type &flux_i,
                      const flux_contribution_type &flux_j,
                      const dealii::Tensor<1, dim, Number> &c_ij) const;
 
      static constexpr bool have_high_order_flux = false;
 
      state_type high_order_flux_divergence(
          const flux_contribution_type &flux_i,
          const flux_contribution_type &flux_j,
          const dealii::Tensor<1, dim, Number> &c_ij) const = delete;
 
 
      static constexpr bool have_source_terms = false;
 
      state_type nodal_source(const PrecomputedVector &pv,
                              const unsigned int i,
                              const state_type &U_i,
                              const ScalarNumber tau) const = delete;
 
      state_type nodal_source(const PrecomputedVector &pv,
                              const unsigned int *js,
                              const state_type &U_j,
                              const ScalarNumber tau) const = delete;
 
 
 
      template <typename ST>
      state_type expand_state(const ST &state) const;
 
      template <typename ST>
      state_type from_initial_state(const ST &initial_state) const;
 
      state_type from_primitive_state(const state_type &primitive_state) const;
 
      state_type to_primitive_state(const state_type &state) const;
 
      template <typename Lambda>
      state_type apply_galilei_transform(const state_type &state,
                                         const Lambda &lambda) const;
    }; /* HyperbolicSystemView */
 
 
    /*
     * -------------------------------------------------------------------------
     * Inline definitions
     * -------------------------------------------------------------------------
     */
 
 
    inline HyperbolicSystem::HyperbolicSystem(
        const std::string &subsection /*= "HyperbolicSystem"*/)
        : ParameterAcceptor(subsection)
    {
      equation_of_state_ = "polytropic gas";
      add_parameter(
          "equation of state",
          equation_of_state_,
          "The equation of state. Valid names are given by any of the "
          "subsections defined below");
 
      compute_strict_bounds_ = true;
      add_parameter(
          "compute strict bounds",
          compute_strict_bounds_,
          "Compute strict, but significantly more expensive bounds at various "
          "places: (a) an expensive, but better upper wavespeed estimate in "
          "the approximate RiemannSolver; (b) entropy viscosity-commutator "
          "with correct gamma_min over the stencil; (c) mathematically correct "
          "surrogate specific entropy minimum with gamma_min over the "
          "stencil.");
 
      reference_density_ = 1.;
      add_parameter("reference density",
                    reference_density_,
                    "Problem specific density reference");
 
      vacuum_state_relaxation_small_ = 1.e2;
      add_parameter("vacuum state relaxation small",
                    vacuum_state_relaxation_small_,
                    "Problem specific vacuum relaxation parameter");
 
      vacuum_state_relaxation_large_ = 1.e4;
      add_parameter("vacuum state relaxation large",
                    vacuum_state_relaxation_large_,
                    "Problem specific vacuum relaxation parameter");
 
      /*
       * And finally populate the equation of state list with all equation of
       * state configurations defined in the EquationOfState namespace:
       */
      EquationOfStateLibrary::populate_equation_of_state_list(
          equation_of_state_list_, subsection);
 
      const auto populate_functions = [this]() {
        bool initialized = false;
        for (auto &it : equation_of_state_list_)
 
          /* Populate EOS-specific quantities and functions */
          if (it->name() == equation_of_state_) {
            selected_equation_of_state_ = it;
            problem_name =
                "Compressible Euler equations (" + it->name() + " EOS)";
            initialized = true;
            break;
          }
 
        AssertThrow(
            initialized,
            dealii::ExcMessage(
                "Could not find an equation of state description with name \"" +
                equation_of_state_ + "\""));
      };
 
      ParameterAcceptor::parse_parameters_call_back.connect(populate_functions);
      populate_functions();
    }
 
 
    template <int dim, typename Number>
    template <typename DISPATCH, typename SPARSITY>
    DEAL_II_ALWAYS_INLINE inline void
    HyperbolicSystemView<dim, Number>::precomputation_loop(
        unsigned int cycle [[maybe_unused]],
        const DISPATCH &dispatch_check,
        const SPARSITY &sparsity_simd,
        StateVector &state_vector,
        unsigned int left,
        unsigned int right,
        const bool skip_constrained_dofs /*= true*/) const
    {
      Assert(cycle == 0 || cycle == 1, dealii::ExcInternalError());
 
      const auto &U = std::get<0>(state_vector);
      auto &precomputed = std::get<1>(state_vector);
 
      /* We are inside a thread parallel context */
 
      const auto &eos = hyperbolic_system_.selected_equation_of_state_;
      unsigned int stride_size = get_stride_size<Number>;
 
      if (cycle == 0) {
        if (eos->prefer_vector_interface()) {
          /*
           * Set up temporary storage for p, rho, e and make two calls into
           * the eos library.
           */
          const auto offset = left;
          const auto size = right - left;
 
          static /* shared */ std::vector<double> p;
          static /* shared */ std::vector<double> rho;
          static /* shared */ std::vector<double> e;
          RYUJIN_OMP_SINGLE
          {
            p.resize(size);
            rho.resize(size);
            e.resize(size);
          }
 
          RYUJIN_OMP_FOR
          for (unsigned int i = 0; i < size; i += stride_size) {
            const auto U_i = U.template get_tensor<Number>(offset + i);
            const auto rho_i = density(U_i);
            const auto e_i = internal_energy(U_i) / rho_i;
            /*
             * Populate rho and e also for interpolated values from
             * constrainted degrees of freedom so that the vectors contain
             * physically admissible entries throughout.
             */
            write_entry<Number>(rho, rho_i, i);
            write_entry<Number>(e, e_i, i);
          }
 
          /* Make sure the call into eospac (and others) is single threaded. */
          RYUJIN_OMP_SINGLE
          {
            eos->pressure(p, rho, e);
          }
 
          RYUJIN_OMP_FOR
          for (unsigned int i = 0; i < size; i += stride_size) {
            /* Skip constrained degrees of freedom: */
            const unsigned int row_length = sparsity_simd.row_length(i);
            if (skip_constrained_dofs && row_length == 1)
              continue;
 
            dispatch_check(i);
 
            using PT = precomputed_type;
            const auto U_i = U.template get_tensor<Number>(offset + i);
            const auto p_i = get_entry<Number>(p, i);
            const auto gamma_i = surrogate_gamma(U_i, p_i);
            const PT prec_i{p_i, gamma_i, Number(0.), Number(0.)};
            precomputed.template write_tensor<Number>(prec_i, offset + i);
          }
        } else {
          /*
           * This is the variant with slightly better performance provided
           * that a call to the eos is not too expensive. This variant
           * calls into the eos library for every single degree of freedom.
           */
          RYUJIN_OMP_FOR
          for (unsigned int i = left; i < right; i += stride_size) {
            /* Skip constrained degrees of freedom: */
            const unsigned int row_length = sparsity_simd.row_length(i);
            if (skip_constrained_dofs && row_length == 1)
              continue;
 
            dispatch_check(i);
 
            const auto U_i = U.template get_tensor<Number>(i);
            const auto rho_i = density(U_i);
            const auto e_i = internal_energy(U_i) / rho_i;
            const auto p_i = eos_pressure(rho_i, e_i);
 
            const auto gamma_i = surrogate_gamma(U_i, p_i);
            using PT = precomputed_type;
            const PT prec_i{p_i, gamma_i, Number(0.), Number(0.)};
            precomputed.template write_tensor<Number>(prec_i, i);
          }
        } /* prefer_vector_interface */
      }   /* cycle == 0 */
 
      if (cycle == 1) {
        RYUJIN_OMP_FOR
        for (unsigned int i = left; i < right; i += stride_size) {
          using PT = precomputed_type;
 
          /* Skip constrained degrees of freedom: */
          const unsigned int row_length = sparsity_simd.row_length(i);
          if (skip_constrained_dofs && row_length == 1)
            continue;
 
          dispatch_check(i);
 
          const auto U_i = U.template get_tensor<Number>(i);
          auto prec_i = precomputed.template get_tensor<Number, PT>(i);
          auto &[p_i, gamma_min_i, s_i, eta_i] = prec_i;
 
          const unsigned int *js = sparsity_simd.columns(i) + stride_size;
          for (unsigned int col_idx = 1; col_idx < row_length;
               ++col_idx, js += stride_size) {
 
            const auto U_j = U.template get_tensor<Number>(js);
            const auto prec_j = precomputed.template get_tensor<Number, PT>(js);
            auto &[p_j, gamma_min_j, s_j, eta_j] = prec_j;
            const auto gamma_j = surrogate_gamma(U_j, p_j);
            gamma_min_i = std::min(gamma_min_i, gamma_j);
          }
 
          s_i = surrogate_specific_entropy(U_i, gamma_min_i);
          eta_i = surrogate_harten_entropy(U_i, gamma_min_i);
          precomputed.template write_tensor<Number>(prec_i, i);
        }
      }
    }
 
 
    template <int dim, typename Number>
    DEAL_II_ALWAYS_INLINE inline Number
    HyperbolicSystemView<dim, Number>::density(const state_type &U)
    {
      return U[0];
    }
 
 
    template <int dim, typename Number>
    DEAL_II_ALWAYS_INLINE inline Number
    HyperbolicSystemView<dim, Number>::filter_vacuum_density(
        const Number &rho) const
    {
      constexpr ScalarNumber eps = std::numeric_limits<ScalarNumber>::epsilon();
      const Number rho_cutoff_large =
          reference_density() * vacuum_state_relaxation_large() * eps;
 
      return dealii::compare_and_apply_mask<dealii::SIMDComparison::less_than>(
          std::abs(rho), rho_cutoff_large, Number(0.), rho);
    }
 
 
    template <int dim, typename Number>
    DEAL_II_ALWAYS_INLINE inline dealii::Tensor<1, dim, Number>
    HyperbolicSystemView<dim, Number>::momentum(const state_type &U)
    {
      dealii::Tensor<1, dim, Number> result;
      for (unsigned int i = 0; i < dim; ++i)
        result[i] = U[1 + i];
      return result;
    }
 
 
    template <int dim, typename Number>
    DEAL_II_ALWAYS_INLINE inline Number
    HyperbolicSystemView<dim, Number>::total_energy(const state_type &U)
    {
      return U[1 + dim];
    }
 
 
    template <int dim, typename Number>
    DEAL_II_ALWAYS_INLINE inline Number
    HyperbolicSystemView<dim, Number>::internal_energy(const state_type &U)
    {
      /*
       * rho e = (E - 1/2*m^2/rho)
       */
      const Number rho_inverse = ScalarNumber(1.) / density(U);
      const auto m = momentum(U);
      const Number E = total_energy(U);
      return E - ScalarNumber(0.5) * m.norm_square() * rho_inverse;
    }
 
 
    template <int dim, typename Number>
    DEAL_II_ALWAYS_INLINE inline auto
    HyperbolicSystemView<dim, Number>::internal_energy_derivative(
        const state_type &U) -> state_type
    {
      /*
       * With
       *   rho e = E - 1/2 |m|^2 / rho
       * we get
       *   (rho e)' = (1/2m^2/rho^2, -m/rho , 1 )^T
       */
 
      const Number rho_inverse = ScalarNumber(1.) / density(U);
      const auto u = momentum(U) * rho_inverse;
 
      state_type result;
 
      result[0] = ScalarNumber(0.5) * u.norm_square();
      for (unsigned int i = 0; i < dim; ++i) {
        result[1 + i] = -u[i];
      }
      result[dim + 1] = ScalarNumber(1.);
 
      return result;
    }
 
 
    template <int dim, typename Number>
    DEAL_II_ALWAYS_INLINE inline Number
    HyperbolicSystemView<dim, Number>::surrogate_specific_entropy(
        const state_type &U, const Number &gamma_min) const
    {
      const auto b = Number(eos_interpolation_b());
      const auto pinf = Number(eos_interpolation_pinfty());
      const auto q = Number(eos_interpolation_q());
 
      const auto rho = density(U);
      const auto rho_inverse = ScalarNumber(1.) / rho;
 
      const auto covolume = Number(1.) - b * rho;
 
      const auto shift = internal_energy(U) - rho * q - pinf * covolume;
 
      return shift * ryujin::pow(rho_inverse - b, gamma_min) / covolume;
    }
 
 
    template <int dim, typename Number>
    DEAL_II_ALWAYS_INLINE inline Number
    HyperbolicSystemView<dim, Number>::surrogate_harten_entropy(
        const state_type &U, const Number &gamma_min) const
    {
      const auto b = Number(eos_interpolation_b());
      const auto pinf = Number(eos_interpolation_pinfty());
      const auto q = Number(eos_interpolation_q());
 
      const auto rho = density(U);
      const auto m = momentum(U);
      const auto E = total_energy(U);
      const auto rho_rho_e_q =
          rho * E - ScalarNumber(0.5) * m.norm_square() - rho * rho * q;
 
      const auto exponent = ScalarNumber(1.) / (gamma_min + Number(1.));
 
      const auto covolume = Number(1.) - b * rho;
      const auto covolume_term = ryujin::pow(covolume, gamma_min - Number(1.));
 
      const auto rho_pinfcov = rho * pinf * covolume;
 
      return ryujin::pow(
          positive_part(rho_rho_e_q - rho_pinfcov) * covolume_term, exponent);
    }
 
 
    template <int dim, typename Number>
    DEAL_II_ALWAYS_INLINE inline auto
    HyperbolicSystemView<dim, Number>::surrogate_harten_entropy_derivative(
        const state_type &U, const Number &eta, const Number &gamma_min) const
        -> state_type
    {
      /*
       * With
       *   eta = (shift * (1-b*rho)^{gamma-1}) ^ {1/(gamma+1)},
       *   shift = rho * E - 1/2 |m|^2 - rho^2 * q - p_infty * rho * (1 - b rho)
       *
       *   shift' = [E - 2 * rho * q - p_infty * (1 - 2 b rho), -m, rho]^T
       *   factor = 1/(gamma+1) * (eta/(1-b rho))^-gamma / (1-b rho)^2
       *
       * we get
       *
       *   eta' = factor * (1-b*rho) * shift' -
       *          factor * shift * (gamma - 1) * b * [1, 0, 0]^T
       *
       */
      const auto b = Number(eos_interpolation_b());
      const auto pinf = Number(eos_interpolation_pinfty());
      const auto q = Number(eos_interpolation_q());
 
      const auto rho = density(U);
      const auto m = momentum(U);
      const auto E = total_energy(U);
 
      const auto covolume = Number(1.) - b * rho;
      const auto covolume_inverse = ScalarNumber(1.) / covolume;
 
      const auto shift = rho * E - ScalarNumber(0.5) * m.norm_square() -
                         rho * rho * q - rho * pinf * covolume;
 
      constexpr auto eps = std::numeric_limits<ScalarNumber>::epsilon();
      const auto regularization = m.norm() * eps;
      const auto max_val = ryujin::pow(
          std::max(regularization, eta * covolume_inverse), gamma_min);
      auto factor = safe_division(Number(1.0), max_val);
      factor *= fixed_power<2>(covolume_inverse) / (gamma_min + Number(1.));
 
      state_type result;
 
      const auto first_term = E - ScalarNumber(2.) * rho * q -
                              pinf * (Number(1.) - ScalarNumber(2.) * b * rho);
      const auto second_term = -(gamma_min - Number(1.)) * shift * b;
 
      result[0] = factor * (covolume * first_term + second_term);
      for (unsigned int i = 0; i < dim; ++i)
        result[1 + i] = -factor * covolume * m[i];
      result[dim + 1] = factor * covolume * rho;
 
      return result;
    }
 
 
    template <int dim, typename Number>
    DEAL_II_ALWAYS_INLINE inline Number
    HyperbolicSystemView<dim, Number>::surrogate_gamma(const state_type &U,
                                                       const Number &p) const
    {
      const auto b = Number(eos_interpolation_b());
      const auto pinf = Number(eos_interpolation_pinfty());
      const auto q = Number(eos_interpolation_q());
 
      const auto rho = density(U);
      const auto rho_e = internal_energy(U);
      const auto covolume = Number(1.) - b * rho;
 
      const auto numerator = (p + pinf) * covolume;
      const auto denominator = rho_e - rho * q - covolume * pinf;
      return Number(1.) + safe_division(numerator, denominator);
    }
 
 
    template <int dim, typename Number>
    DEAL_II_ALWAYS_INLINE inline Number
    HyperbolicSystemView<dim, Number>::surrogate_pressure(
        const state_type &U, const Number &gamma) const
    {
      const auto b = Number(eos_interpolation_b());
      const auto pinf = Number(eos_interpolation_pinfty());
      const auto q = Number(eos_interpolation_q());
 
      const auto rho = density(U);
      const auto rho_e = internal_energy(U);
      const auto covolume = Number(1.) - b * rho;
 
      return positive_part(gamma - Number(1.)) *
                 safe_division(rho_e - rho * q, covolume) -
             gamma * pinf;
    }
 
 
    template <int dim, typename Number>
    DEAL_II_ALWAYS_INLINE inline Number
    HyperbolicSystemView<dim, Number>::surrogate_speed_of_sound(
        const state_type &U, const Number &gamma) const
    {
      const auto b = Number(eos_interpolation_b());
      const auto pinf = Number(eos_interpolation_pinfty());
      const auto q = Number(eos_interpolation_q());
 
      const auto rho = density(U);
      const auto rho_e = internal_energy(U);
      const auto covolume = Number(1.) - b * rho;
 
      auto radicand =
          (rho_e - rho * q - pinf * covolume) / (covolume * covolume * rho);
      radicand *= gamma * (gamma - 1.);
      return std::sqrt(positive_part(radicand));
    }
 
 
    template <int dim, typename Number>
    DEAL_II_ALWAYS_INLINE inline bool
    HyperbolicSystemView<dim, Number>::is_admissible(const state_type &U) const
    {
      const auto b = Number(eos_interpolation_b());
      const auto pinf = Number(eos_interpolation_pinfty());
      const auto q = Number(eos_interpolation_q());
 
      const auto rho = density(U);
      const auto rho_e = internal_energy(U);
      const auto covolume = Number(1.) - b * rho;
 
      const auto shift = rho_e - rho * q - pinf * covolume;
 
      constexpr auto gt = dealii::SIMDComparison::greater_than;
      using T = Number;
      const auto test =
          dealii::compare_and_apply_mask<gt>(rho, T(0.), T(0.), T(-1.)) + //
          dealii::compare_and_apply_mask<gt>(shift, T(0.), T(0.), T(-1.));
 
#ifdef DEBUG_OUTPUT
      if (!(test == Number(0.))) {
        std::cout << std::fixed << std::setprecision(16);
        std::cout << "Bounds violation: Negative state [rho, e] detected!\n";
        std::cout << "\t\trho:           " << rho << "\n";
        std::cout << "\t\tint (shifted): " << shift << "\n";
      }
#endif
 
      return (test == Number(0.));
    }
 
 
    template <int dim, typename Number>
    template <int component>
    DEAL_II_ALWAYS_INLINE inline auto
    HyperbolicSystemView<dim, Number>::prescribe_riemann_characteristic(
        const state_type &U,
        const Number &p,
        const state_type &U_bar,
        const Number &p_bar,
        const dealii::Tensor<1, dim, Number> &normal) const -> state_type
    {
      static_assert(component == 1 || component == 2,
                    "component has to be 1 or 2");
 
      const auto b = Number(eos_interpolation_b());
      const auto pinf = Number(eos_interpolation_pinfty());
      const auto q = Number(eos_interpolation_q());
 
      /*
       * The "four" Riemann characteristics are formed under the assumption
       * of a locally isentropic flow. For this, we first transform both
       * states into {rho, vn, vperp, gamma, a}, where we use the NASG EOS
       * interpolation to derive a surrogate gamma and speed of sound a.
       *
       * See, e.g., https://arxiv.org/pdf/2004.08750, "Compressible flow in
       * a NOble-Abel Stiffened-Gas fluid", M. I. Radulescu.
       */
 
      const auto m = momentum(U);
      const auto rho = density(U);
      const auto vn = m * normal / rho;
 
      const auto gamma = surrogate_gamma(U, p);
      const auto a = surrogate_speed_of_sound(U, gamma);
      const auto covolume = 1. - b * rho;
 
      const auto m_bar = momentum(U_bar);
      const auto rho_bar = density(U_bar);
      const auto vn_bar = m_bar * normal / rho_bar;
 
      const auto gamma_bar = surrogate_gamma(U_bar, p_bar);
      const auto a_bar = surrogate_speed_of_sound(U_bar, gamma_bar);
      const auto covolume_bar = 1. - b * rho_bar;
 
      /*
       * Now compute the Riemann characteristics {R_1, R_2, vperp, s}:
       *   R_1 = v * n - 2 / (gamma - 1) * a * (1 - b * rho)
       *   R_2 = v * n + 2 / (gamma - 1) * a * (1 - b * rho)
       *   vperp
       *   S = (p + p_infty) / rho^gamma * (1 - b * rho)^gamma
       *
       * Here, we replace either R_1, or R_2 with values coming from U_bar:
       */
 
      const auto R_1 =
          component == 1 ? vn_bar - 2. * a_bar / (gamma_bar - 1.) * covolume_bar
                         : vn - 2. * a / (gamma - 1.) * covolume;
 
      const auto R_2 =
          component == 2 ? vn_bar + 2. * a_bar / (gamma_bar - 1.) * covolume_bar
                         : vn + 2. * a / (gamma - 1.) * covolume;
 
      /*
       * Note that we are really hoping for the best here... We require
       * that R_2 >= R_1 so that we can extract a valid sound speed...
       */
 
      Assert(
          R_2 >= R_1,
          dealii::ExcMessage("Encountered R_2 < R_1 in dynamic boundary value "
                             "enforcement. This implies that the interpolation "
                             "with Riemann characteristics failed."));
 
      const auto vperp = m / rho - vn * normal;
 
      const auto S = (p + pinf) * ryujin::pow(Number(1.) / rho - b, gamma);
 
      /*
       * Now, we have to reconstruct the actual conserved state U from the
       * Riemann characteristics R_1, R_2, vperp, and s. We first set up
       * {vn_new, vperp_new, a_new, S} and then solve for {rho_new, p_new}
       * with the help of the NASG EOS surrogate formulas:
       *
       *   S = (p + p_infty) / rho^gamma * (1 - b * rho)^gamma
       *
       *   a^2 = gamma * (p + p_infty) / (rho * cov)
       *
       *   This implies:
       *
       *   a^2 / (gamma * S) = rho^{gamma - 1} / (1 - b * rho)^{1 + gamma}
       */
 
      const auto vn_new = Number(0.5) * (R_1 + R_2);
 
      /*
       * Technically, we would need to solve for rho subject to a number of
       * nonlinear relationships:
       *
       *   a   = (gamma - 1) * (R_2 - R_1) / (4. * (1 - b * rho))
       *
       *   a^2 / (gamma * S) = rho^{gamma - 1} / (1 - b * rho)^{gamma + 1}
       *
       * This seems to be a bit expensive for the fact that our dynamic
       * boundary conditions are already terribly heuristic...
       *
       * So instead, we rewrite this system as:
       *
       *   a * (1 - b * rho) = (gamma - 1) * (R_2 - R_1) / 4.
       *
       *   a^2 / (gamma * S) (1 - b * rho)^2
       *                           = (rho / (1 - b * rho))^{gamma - 1}
       *
       * And compute the terms on the left simply with the old covolume and
       * solving an easier easier nonlinear equation for the density. The
       * resulting system reads:
       *
       *   a = (gamma - 1) * (R_2 - R_1) / (4. * (1 - b * rho_old))
       *   A = {a^2 / (gamma * S) (1 - b * rho_old)^{2 gamma}}^{1/(gamma - 1)}
       *
       *   rho = A / (1 + b * A)
       */
 
      const auto a_new_square =
          ryujin::fixed_power<2>((gamma - 1.) * (R_2 - R_1) / (4. * covolume));
 
      auto term = ryujin::pow(a_new_square / (gamma * S), 1. / (gamma - 1.));
      if (b != ScalarNumber(0.)) {
        term *= std::pow(covolume, 2. / (gamma - 1.));
      }
 
      const auto rho_new = term / (1. + b * term);
 
      const auto covolume_new = (1. - b * rho_new);
      const auto p_new = a_new_square / gamma * rho_new * covolume_new - pinf;
 
      /*
       * And translate back into conserved quantities:
       */
 
      const auto rho_e_new =
          rho_new * q + (p_new + gamma * pinf) * covolume_new / (gamma - 1.);
 
      state_type U_new;
      U_new[0] = rho_new;
      for (unsigned int d = 0; d < dim; ++d) {
        U_new[1 + d] = rho_new * (vn_new * normal + vperp)[d];
      }
      U_new[1 + dim] =
          rho_e_new + 0.5 * rho_new * (vn_new * vn_new + vperp.norm_square());
 
      return U_new;
    }
 
 
    template <int dim, typename Number>
    template <typename Lambda>
    DEAL_II_ALWAYS_INLINE inline auto
    HyperbolicSystemView<dim, Number>::apply_boundary_conditions(
        dealii::types::boundary_id id,
        const state_type &U,
        const dealii::Tensor<1, dim, Number> &normal,
        const Lambda &get_dirichlet_data) const -> state_type
    {
      state_type result = U;
 
      if (id == Boundary::dirichlet) {
        result = get_dirichlet_data();
 
      } else if (id == Boundary::dirichlet_momentum) {
        /* Only enforce Dirichlet conditions on the momentum: */
        auto m_dirichlet = momentum(get_dirichlet_data());
        for (unsigned int k = 0; k < dim; ++k)
          result[k + 1] = m_dirichlet[k];
 
      } else if (id == Boundary::slip) {
        auto m = momentum(U);
        m -= 1. * (m * normal) * normal;
        for (unsigned int k = 0; k < dim; ++k)
          result[k + 1] = m[k];
 
      } else if (id == Boundary::no_slip) {
        for (unsigned int k = 0; k < dim; ++k)
          result[k + 1] = Number(0.);
 
      } else if (id == Boundary::dynamic) {
        /*
         * On dynamic boundary conditions, we distinguish four cases:
         *
         *  - supersonic inflow: prescribe full state
         *  - subsonic inflow:
         *      decompose into Riemann invariants and leave R_2
         *      characteristic untouched.
         *  - supersonic outflow: do nothing
         *  - subsonic outflow:
         *      decompose into Riemann invariants and prescribe incoming
         *      R_1 characteristic.
         */
        const auto m = momentum(U);
        const auto rho = density(U);
        const auto rho_e = internal_energy(U);
 
        /*
         * We do not have precomputed values available. Thus, simply query
         * the pressure oracle and compute a surrogate speed of sound from
         * there:
         */
        const auto p = eos_pressure(rho, rho_e / rho);
        const auto gamma = surrogate_gamma(U, p);
        const auto a = surrogate_speed_of_sound(U, gamma);
        const auto vn = m * normal / rho;
 
        /* Supersonic inflow: */
        if (vn < -a) {
          result = get_dirichlet_data();
        }
 
        /* Subsonic inflow: */
        if (vn >= -a && vn <= 0.) {
          const auto U_dirichlet = get_dirichlet_data();
          const auto rho_dirichlet = density(U_dirichlet);
          const auto rho_e_dirichlet = internal_energy(U_dirichlet);
          const auto p_dirichlet =
              eos_pressure(rho_dirichlet, rho_e_dirichlet / rho_dirichlet);
 
          result = prescribe_riemann_characteristic<2>(
              U_dirichlet, p_dirichlet, U, p, normal);
        }
 
        /* Subsonic outflow: */
        if (vn > 0. && vn <= a) {
          const auto U_dirichlet = get_dirichlet_data();
          const auto rho_dirichlet = density(U_dirichlet);
          const auto rho_e_dirichlet = internal_energy(U_dirichlet);
          const auto p_dirichlet =
              eos_pressure(rho_dirichlet, rho_e_dirichlet / rho_dirichlet);
 
          result = prescribe_riemann_characteristic<1>(
              U, p, U_dirichlet, p_dirichlet, normal);
        }
        /* Supersonic outflow: do nothing, i.e., keep U as is */
 
      } else {
        AssertThrow(false, dealii::ExcNotImplemented());
      }
 
      return result;
    }
 
 
    template <int dim, typename Number>
    DEAL_II_ALWAYS_INLINE inline auto
    HyperbolicSystemView<dim, Number>::f(const state_type &U,
                                         const Number &p) const -> flux_type
    {
      const auto rho_inverse = ScalarNumber(1.) / density(U);
      const auto m = momentum(U);
      const auto E = total_energy(U);
 
      flux_type result;
 
      result[0] = m;
      for (unsigned int i = 0; i < dim; ++i) {
        result[1 + i] = m * (m[i] * rho_inverse);
        result[1 + i][i] += p;
      }
      result[dim + 1] = m * (rho_inverse * (E + p));
 
      return result;
    }
 
 
    template <int dim, typename Number>
    DEAL_II_ALWAYS_INLINE inline auto
    HyperbolicSystemView<dim, Number>::flux_contribution(
        const PrecomputedVector &pv,
        const InitialPrecomputedVector & /*piv*/,
        const unsigned int i,
        const state_type &U_i) const -> flux_contribution_type
    {
      const auto &[p_i, gamma_min_i, s_i, eta_i] =
          pv.template get_tensor<Number, precomputed_type>(i);
      return f(U_i, p_i);
    }
 
 
    template <int dim, typename Number>
    DEAL_II_ALWAYS_INLINE inline auto
    HyperbolicSystemView<dim, Number>::flux_contribution(
        const PrecomputedVector &pv,
        const InitialPrecomputedVector & /*piv*/,
        const unsigned int *js,
        const state_type &U_j) const -> flux_contribution_type
    {
      const auto &[p_j, gamma_min_j, s_j, eta_j] =
          pv.template get_tensor<Number, precomputed_type>(js);
      return f(U_j, p_j);
    }
 
 
    template <int dim, typename Number>
    DEAL_II_ALWAYS_INLINE inline auto
    HyperbolicSystemView<dim, Number>::flux_divergence(
        const flux_contribution_type &flux_i,
        const flux_contribution_type &flux_j,
        const dealii::Tensor<1, dim, Number> &c_ij) const -> state_type
    {
      return -contract(add(flux_i, flux_j), c_ij);
    }
 
 
    template <int dim, typename Number>
    template <typename ST>
    auto HyperbolicSystemView<dim, Number>::expand_state(const ST &state) const
        -> state_type
    {
      using T = typename ST::value_type;
      static_assert(std::is_same_v<Number, T>, "template mismatch");
 
      constexpr auto dim2 = ST::dimension - 2;
      static_assert(dim >= dim2,
                    "the space dimension of the argument state must not be "
                    "larger than the one of the target state");
 
      state_type result;
      result[0] = state[0];
      result[dim + 1] = state[dim2 + 1];
      for (unsigned int i = 1; i < dim2 + 1; ++i)
        result[i] = state[i];
 
      return result;
    }
 
 
    template <int dim, typename Number>
    template <typename ST>
    DEAL_II_ALWAYS_INLINE inline auto
    HyperbolicSystemView<dim, Number>::from_initial_state(
        const ST &initial_state) const -> state_type
    {
      auto primitive_state = expand_state(initial_state);
 
      /* pressure into specific internal energy: */
      const auto rho = density(primitive_state);
      const auto p = /*SIC!*/ total_energy(primitive_state);
      const auto e = eos_specific_internal_energy(rho, p);
      primitive_state[dim + 1] = e;
 
      return from_primitive_state(primitive_state);
    }
 
 
    template <int dim, typename Number>
    DEAL_II_ALWAYS_INLINE inline auto
    HyperbolicSystemView<dim, Number>::from_primitive_state(
        const state_type &primitive_state) const -> state_type
    {
      const auto rho = density(primitive_state);
      /* extract velocity: */
      const auto u = /*SIC!*/ momentum(primitive_state);
      /* extract specific internal energy: */
      const auto &e = /*SIC!*/ total_energy(primitive_state);
 
      auto state = primitive_state;
      /* Fix up momentum: */
      for (unsigned int i = 1; i < dim + 1; ++i)
        state[i] *= rho;
 
      /* Compute total energy: */
      state[dim + 1] = rho * e + Number(0.5) * rho * u * u;
 
      return state;
    }
 
 
    template <int dim, typename Number>
    DEAL_II_ALWAYS_INLINE inline auto
    HyperbolicSystemView<dim, Number>::to_primitive_state(
        const state_type &state) const -> state_type
    {
      const auto rho = density(state);
      const auto rho_inverse = Number(1.) / rho;
      const auto rho_e = internal_energy(state);
 
      auto primitive_state = state;
      /* Fix up velocity: */
      for (unsigned int i = 1; i < dim + 1; ++i)
        primitive_state[i] *= rho_inverse;
      /* Set specific internal energy: */
      primitive_state[dim + 1] = rho_e * rho_inverse;
 
      return primitive_state;
    }
 
 
    template <int dim, typename Number>
    template <typename Lambda>
    auto HyperbolicSystemView<dim, Number>::apply_galilei_transform(
        const state_type &state, const Lambda &lambda) const -> state_type
    {
      auto result = state;
      const auto M = lambda(momentum(state));
      for (unsigned int d = 0; d < dim; ++d)
        result[1 + d] = M[d];
      return result;
    }
  } // namespace EulerAEOS
} // namespace ryujin