GridFire/src/lib/solver/strategies/SpectralSolverStrategy.cpp

#include "gridfire/solver/strategies/SpectralSolverStrategy.h"

#include <sunlinsol/sunlinsol_dense.h>

#include "gridfire/utils/sundials.h"

#include "quill/LogMacros.h"
#include "sunmatrix/sunmatrix_dense.h"

namespace {
    std::pair<size_t, std::vector<double>> evaluate_bspline(
        double x,
        const gridfire::solver::SpectralSolverStrategy::SplineBasis& basis
    ) {
        const int p = basis.degree;
        const std::vector<double>& t = basis.knots;

        auto it = std::ranges::upper_bound(t, x);
        size_t i = std::distance(t.begin(), it) - 1;

        if (i < static_cast<size_t>(p)) i = p;
        if (i >= t.size() - 1 - p) i = t.size() - 2 - p;

        if (x >= t.back()) {
            i = t.size() - p - 2;
        }

        // Cox-de Boor algorithm
        std::vector<double> N(p + 1);
        std::vector<double> left(p + 1);
        std::vector<double> right(p + 1);

        N[0] = 1.0;

        for (int j = 1; j <= p; ++j) {
            left[j] = x - t[i + 1 - j];
            right[j] = t[i + j] - x;
            double saved = 0.0;

            for (int r = 0; r < j; ++r) {
                double temp = N[r] / (right[r + 1] + left[j - r]);
                N[r] = saved + right[r + 1] * temp;

                saved = left[j - r] * temp;
            }
            N[j] = saved;
        }
        return {i - p, N};
    }
}

namespace gridfire::solver {

    SpectralSolverStrategy::SpectralSolverStrategy(engine::DynamicEngine& engine) : MultiZoneNetworkSolverStrategy<engine::DynamicEngine> (engine) {
        LOG_INFO(m_logger, "Initializing SpectralSolverStrategy");

        utils::check_sundials_flag(SUNContext_Create(SUN_COMM_NULL, &m_sun_ctx), "SUNContext_Create", utils::SUNDIALS_RET_CODE_TYPES::CVODE);

        m_absTol = m_config->solver.spectral.absTol;
        m_relTol = m_config->solver.spectral.relTol;

        LOG_INFO(m_logger, "SpectralSolverStrategy initialized successfully");
    }

    SpectralSolverStrategy::~SpectralSolverStrategy() {
        LOG_INFO(m_logger, "Destroying SpectralSolverStrategy");


        if (m_cvode_mem) {
            CVodeFree(&m_cvode_mem);
            m_cvode_mem = nullptr;
        }

        if (m_LS) SUNLinSolFree(m_LS);
        if (m_J) SUNMatDestroy(m_J);
        if (m_Y) N_VDestroy(m_Y);
        if (m_constraints) N_VDestroy(m_constraints);

        if (m_sun_ctx) {
            SUNContext_Free(&m_sun_ctx);
            m_sun_ctx = nullptr;
        }

        if (m_T_coeffs) N_VDestroy(m_T_coeffs);
        if (m_rho_coeffs) N_VDestroy(m_rho_coeffs);

        LOG_INFO(m_logger, "SpectralSolverStrategy destroyed successfully");
    }

    ////////////////////////////////////////////////////////////////////////////////
    /// Main Evaluation Loop
    /////////////////////////////////////////////////////////////////////////////////

    std::vector<NetOut> SpectralSolverStrategy::evaluate(const std::vector<NetIn>& netIns, const std::vector<double>& mass_coords) {
        LOG_INFO(m_logger, "Starting spectral solver evaluation for {} zones", netIns.size());

        assert(std::ranges::all_of(netIns, [&netIns](const NetIn& in) { return in.tMax == netIns[0].tMax; }) && "All NetIn entries must have the same tMax for spectral solver evaluation.");

        std::vector<NetIn> updatedNetIns = netIns;
        for (auto& netIn : updatedNetIns) {
            netIn.composition = m_engine.update(netIn);
        }


        /////////////////////////////////////
        /// Evaluate the monitor function ///
        /////////////////////////////////////

        const std::vector<double> monitor_function = evaluate_monitor_function(updatedNetIns);
        m_current_basis = generate_basis_from_monitor(monitor_function, mass_coords);

        size_t num_basis_funcs = m_current_basis.knots.size() - m_current_basis.degree - 1;

        std::vector<BasisEval> shell_cache(updatedNetIns.size());
        for (size_t shellID = 0; shellID < shell_cache.size(); ++shellID) {
            auto [start, phi] = evaluate_bspline(mass_coords[shellID], m_current_basis);
            shell_cache[shellID] = {.start_idx=start, .phi=phi};
        }

        DenseLinearSolver proj_solver(num_basis_funcs, m_sun_ctx);
        proj_solver.init_from_cache(num_basis_funcs, shell_cache);

        if (m_T_coeffs) N_VDestroy(m_T_coeffs);
        m_T_coeffs = N_VNew_Serial(static_cast<sunindextype>(num_basis_funcs), m_sun_ctx);
        project_specific_variable(updatedNetIns, mass_coords, shell_cache, proj_solver, m_T_coeffs, 0, [](const NetIn& s) { return s.temperature; }, true);

        if (m_rho_coeffs) N_VDestroy(m_rho_coeffs);
        m_rho_coeffs = N_VNew_Serial(static_cast<sunindextype>(num_basis_funcs), m_sun_ctx);
        project_specific_variable(updatedNetIns, mass_coords, shell_cache, proj_solver, m_rho_coeffs, 0, [](const NetIn& s) { return s.density; }, true);

        size_t num_species = m_engine.getNetworkSpecies().size();
        size_t current_offset = 0;

        size_t total_coefficients = num_basis_funcs * (num_species + 1);

        if (m_Y) N_VDestroy(m_Y);
        if (m_constraints) N_VDestroy(m_constraints);

        m_Y = N_VNew_Serial(static_cast<sunindextype>(total_coefficients), m_sun_ctx);
        m_constraints = N_VClone(m_Y);
        N_VConst(0.0, m_constraints); // For now no constraints on coefficients

        for (const auto& sp : m_engine.getNetworkSpecies()) {
            project_specific_variable(
                updatedNetIns,
                mass_coords,
                shell_cache,
                proj_solver,
                m_Y,
                current_offset,
                [&sp](const NetIn& s) { return s.composition.getMolarAbundance(sp); },
                false
            );
            current_offset += num_basis_funcs;
        }

        sunrealtype* y_data = N_VGetArrayPointer(m_Y);
        const size_t energy_offset = num_species * num_basis_funcs;

        assert(energy_offset == current_offset && "Energy offset calculation mismatch in spectral solver initialization.");

        for (size_t i = 0; i < num_basis_funcs; ++i) {
            y_data[energy_offset + i] = 0.0;
        }

        DenseLinearSolver mass_solver(num_basis_funcs, m_sun_ctx);
        mass_solver.init_from_basis(num_basis_funcs, m_current_basis);

        /////////////////////////////////////
        /// CVODE Initialization          ///
        /////////////////////////////////////

        CVODEUserData data;
        data.solver_instance = this;
        data.engine = &m_engine;
        data.mass_matrix_solver_instance = &mass_solver;
        data.basis = &m_current_basis;

        const double absTol = m_absTol.value_or(1e-10);
        const double relTol = m_relTol.value_or(1e-6);

        const bool size_changed = m_last_size != total_coefficients;
        m_last_size = total_coefficients;

        if (m_cvode_mem == nullptr || size_changed) {
            if (m_cvode_mem) {
                CVodeFree(&m_cvode_mem);
                m_cvode_mem = nullptr;
            }
            if (m_LS) {
                SUNLinSolFree(m_LS);
                m_LS = nullptr;
            }
            if (m_J) {
                SUNMatDestroy(m_J);
                m_J = nullptr;
            }

            m_cvode_mem = CVodeCreate(CV_BDF, m_sun_ctx);
            utils::check_sundials_flag(m_cvode_mem == nullptr ? -1 : 0, "CVodeCreate", utils::SUNDIALS_RET_CODE_TYPES::CVODE);

            utils::check_sundials_flag(CVodeInit(m_cvode_mem, cvode_rhs_wrapper, 0.0, m_Y), "CVodeInit", utils::SUNDIALS_RET_CODE_TYPES::CVODE);
            m_J = SUNDenseMatrix(static_cast<sunindextype>(total_coefficients), static_cast<sunindextype>(total_coefficients), m_sun_ctx);
            m_LS = SUNLinSol_Dense(m_Y, m_J, m_sun_ctx);
            utils::check_sundials_flag(CVodeSetLinearSolver(m_cvode_mem, m_LS, m_J), "CVodeSetLinearSolver", utils::SUNDIALS_RET_CODE_TYPES::CVODE);

            // For now, we will not attach a Jacobian function, using finite differences
        } else {
            utils::check_sundials_flag(CVodeReInit(m_cvode_mem, 0.0, m_Y), "CVodeReInit", utils::SUNDIALS_RET_CODE_TYPES::CVODE);
        }

        utils::check_sundials_flag(CVodeSStolerances(m_cvode_mem, relTol, absTol), "CVodeSStolerances", utils::SUNDIALS_RET_CODE_TYPES::CVODE);
        utils::check_sundials_flag(CVodeSetUserData(m_cvode_mem, &data), "CVodeSetUserData", utils::SUNDIALS_RET_CODE_TYPES::CVODE);

        /////////////////////////////////////
        /// Time Integration Loop         ///
        /////////////////////////////////////

        const double target_time = updatedNetIns[0].tMax;
        double current_time = 0.0;

        while (current_time < target_time) {
            int flag = CVode(m_cvode_mem, target_time, m_Y, &current_time, CV_ONE_STEP);
            utils::check_sundials_flag(flag, "CVode", utils::SUNDIALS_RET_CODE_TYPES::CVODE);

            std::println("Advanced to time: {:10.4e} / {:10.4e}", current_time, target_time);
        }

        std::vector<NetOut> results = reconstruct_solution(updatedNetIns, mass_coords, m_Y, m_current_basis, target_time);
        return results;
    }

    void SpectralSolverStrategy::set_callback(const std::any &callback) {
        m_callback = std::any_cast<TimestepCallback>(callback);
    }

    std::vector<std::tuple<std::string, std::string>> SpectralSolverStrategy::describe_callback_context() const {
        throw std::runtime_error("SpectralSolverStrategy does not yet implement describe_callback_context.");
    }

    bool SpectralSolverStrategy::get_stdout_logging_enabled() const {
        return m_stdout_logging_enabled;
    }

    void SpectralSolverStrategy::set_stdout_logging_enabled(bool logging_enabled) {
        m_stdout_logging_enabled = logging_enabled;
    }

    ////////////////////////////////////////////////////////////////////////////////
    /// Static Wrappers for SUNDIALS Callbacks
    ////////////////////////////////////////////////////////////////////////////////

    int SpectralSolverStrategy::cvode_rhs_wrapper(
        const sunrealtype t,
        const N_Vector y_coeffs,
        const N_Vector ydot_coeffs,
        void *user_data
    ) {
        auto *data = static_cast<CVODEUserData*>(user_data);
        const auto *instance = data->solver_instance;

        try {
            return instance -> calculate_rhs(t, y_coeffs, ydot_coeffs, data);
        } catch (const std::exception& e) {
            LOG_CRITICAL(instance->m_logger, "Uncaught exception in Spectral Solver RHS wrapper at time {}: {}", t, e.what());
            return -1;
        } catch (...) {
            LOG_CRITICAL(instance->m_logger, "Unknown uncaught exception in Spectral Solver RHS wrapper at time {}", t);
            return -1;
        }
    }

    int SpectralSolverStrategy::cvode_jac_wrapper(
        const sunrealtype t,
        const N_Vector y,
        const N_Vector ydot,
        const SUNMatrix J,
        void *user_data,
        const N_Vector tmp1,
        const N_Vector tmp2,
        const N_Vector tmp3
    ) {
        const auto *data = static_cast<CVODEUserData*>(user_data);
        const auto *instance = data->solver_instance;

        try {
            LOG_WARNING_LIMIT_EVERY_N(1000, instance->m_logger, "Analytic Jacobian Generation not yet implemented, using finite difference approximation");
            return 0;
        } catch (const std::exception& e) {
            LOG_CRITICAL(instance->m_logger, "Uncaught exception in Spectral Solver Jacobian wrapper at time {}: {}", t, e.what());
            return -1;
        } catch (...) {
            LOG_CRITICAL(instance->m_logger, "Unknown uncaught exception in Spectral Solver Jacobian wrapper at time {}", t);
            return -1;
        }
    }

    ////////////////////////////////////////////////////////////////////////////////
    /// RHS implementation
    ////////////////////////////////////////////////////////////////////////////////

    int SpectralSolverStrategy::calculate_rhs(
        sunrealtype t,
        N_Vector y_coeffs,
        N_Vector ydot_coeffs,
        CVODEUserData* data
    ) const {
        const auto& basis = m_current_basis;
        DenseLinearSolver* mass_solver = data->mass_matrix_solver_instance;
        const auto& species_list = m_engine.getNetworkSpecies();

        const size_t num_basis_funcs = basis.knots.size() - basis.degree - 1;
        const size_t num_species = species_list.size();

        sunrealtype* rhs_data = N_VGetArrayPointer(ydot_coeffs);
        N_VConst(0.0, ydot_coeffs);

        // PERF: In future we can use openMP to parallelize over these basis functions once we make the engines thread safe
        for (size_t q = 0; q < basis.quadrature_nodes.size(); ++q) {
            double w_q = basis.quadrature_weights[q];

            const auto& [start_idx, phi] = basis.quad_evals[q];

            GridPoint gp = reconstruct_at_quadrature(y_coeffs, q, basis);
            std::expected<engine::StepDerivatives<double>, engine::EngineStatus> results = m_engine.calculateRHSAndEnergy(gp.composition, gp.T9, gp.rho, false);

            // PERF: When switching to parallel execution, we will need to protect this section with a mutex or use atomic operations since we cannot throw safely from multiple threads
            if (!results) {
                LOG_CRITICAL(m_logger, "Engine failed to calculate RHS at time {}: {}", t, EngineStatus_to_string(results.error()));
                return -1;
            }

            const auto& [dydt, eps_nuc, contributions, nu_loss, nu_flux] = results.value();

            for (size_t s = 0; s < num_species; ++s) {
                double rate = dydt.at(species_list[s]);
                size_t species_offset = s * num_basis_funcs;

                for (size_t k = 0; k < phi.size(); ++k) {
                    size_t global_idx = species_offset + start_idx + k;
                    rhs_data[global_idx] += w_q * phi[k] * rate;
                }
            }

            size_t energy_offset = num_species * num_basis_funcs;

            for (size_t k = 0; k < phi.size(); ++k) {
                size_t global_idx = energy_offset + start_idx + k;
                rhs_data[global_idx] += eps_nuc * w_q * phi[k];
            }
        }

        size_t total_vars = num_species + 1;
        mass_solver->solve_inplace(ydot_coeffs, total_vars, num_basis_funcs);

        return 0;
    }


    ////////////////////////////////////////////////////////////////////////////////
    /// Spectral Utilities
    ///     These include basis generation, monitor function evaluation
    ///     projection and reconstruction routines.
    ////////////////////////////////////////////////////////////////////////////////

    std::vector<double> SpectralSolverStrategy::evaluate_monitor_function(const std::vector<NetIn>& current_shells) const {
        const size_t n_shells = current_shells.size();
        if (n_shells < 3) {
            return std::vector<double>(n_shells, 1.0); // NOLINT(*-return-braced-init-list)
        }

        std::vector<double> M(n_shells, 1.0);

        auto accumulate_variable = [&](auto getter, double weight, bool use_log) {
            std::vector<double> data(n_shells);
            double min_val = std::numeric_limits<double>::max();
            double max_val = std::numeric_limits<double>::lowest();

            for (size_t i = 0 ; i < n_shells; ++i) {
                double val = getter(current_shells[i]);
                if (use_log) {
                    val = std::log10(std::max(val, 1e-100));
                }

                data[i] = val;

                if (val < min_val) min_val = val;
                if (val > max_val) max_val = val;
            }

            const double scale = max_val - min_val;
            if (scale < 1e-10) return;

            for (size_t i = 1; i < n_shells - 1; ++i) {
                const double v_prev = data[i-1];
                const double v_curr = data[i];
                const double v_next = data[i+1];

                // Finite difference estimates for first and second derivatives
                double d1 = std::abs(v_next - v_prev) / 2.0;
                double d2 = std::abs(v_next - 2.0 * v_curr + v_prev);

                d1 /= scale;
                d2 /= scale;

                const double alpha = m_config->solver.spectral.monitorFunction.alpha;
                const double beta = m_config->solver.spectral.monitorFunction.beta;

                M[i] += weight * (alpha * d1 + beta * d2);
            }
        };

        const double structure_weight = m_config->solver.spectral.monitorFunction.structure_weight;
        double abundance_weight = m_config->solver.spectral.monitorFunction.abundance_weight;
        accumulate_variable([](const NetIn& s) { return s.temperature; }, structure_weight, true);
        accumulate_variable([](const NetIn& s) { return s.density; }, structure_weight, true);

        for (const auto& sp : m_engine.getNetworkSpecies()) {
            accumulate_variable([&sp](const NetIn& s) { return s.composition.getMolarAbundance(sp); }, abundance_weight, false);
        }

        //////////////////////////////
        /// Smoothing the Monitor  ///
        //////////////////////////////

        std::vector<double> M_smooth = M;
        for (size_t i = 1; i < n_shells - 1; ++i) {
            M_smooth[i] = (M[i-1] + 2.0 * M[i] + M[i+1]) / 4.0;
        }

        M_smooth[0] = M_smooth[1];
        M_smooth[n_shells-1] = M_smooth[n_shells-2];

        return M_smooth;

    }

    SpectralSolverStrategy::SplineBasis SpectralSolverStrategy::generate_basis_from_monitor(
        const std::vector<double>& monitor_values,
        const std::vector<double>& mass_coordinates
    ) const {
        SplineBasis basis;
        basis.degree = 3; // Cubic Spline

        const size_t n_shells = monitor_values.size();

        std::vector<double> I(n_shells, 0.0);
        double current_integral = 0.0;

        for (size_t i = 1; i < n_shells; ++i) {
            const double dx = mass_coordinates[i] - mass_coordinates[i-1];
            double dI = 0.5 * (monitor_values[i] + monitor_values[i-1]) * dx;

            dI = std::max(dI, 1e-30);
            current_integral += dI;
            I[i] = current_integral;
        }

        const double total_integral = I.back();
        for (size_t i = 0; i < n_shells; ++i) {
            I[i] /= total_integral;
        }

        const size_t num_elements = m_config->solver.spectral.basis.num_elements;
        basis.knots.reserve(num_elements + 1 + 2 * basis.degree);

        // Note that these imply that mass_coordinates must be sorted in increasing order
        double min_mass = mass_coordinates.front();
        double max_mass = mass_coordinates.back();

        for (int i = 0; i < basis.degree; ++i) {
            basis.knots.push_back(min_mass);
        }

        for (size_t k = 1; k < num_elements; ++k) {
            double target_I = static_cast<double>(k) / static_cast<double>(num_elements);

            auto it = std::ranges::lower_bound(I, target_I);
            size_t idx = std::distance(I.begin(), it);

            if (idx == 0) idx = 1;
            if (idx >= n_shells) idx = n_shells - 1;

            double I0 = I[idx-1];
            double I1 = I[idx];
            double m0 = mass_coordinates[idx-1];
            double m1 = mass_coordinates[idx];

            double fraction = (target_I - I0) / (I1 - I0);
            double knot_location = m0 + fraction * (m1 - m0);

            basis.knots.push_back(knot_location);
        }

        for (int i = 0; i < basis.degree; ++i) {
            basis.knots.push_back(max_mass);
        }

        constexpr double sqrt_3_over_5 = 0.77459666924;
        constexpr double five_over_nine = 5.0 / 9.0;
        constexpr double eight_over_nine = 8.0 / 9.0;
        static constexpr std::array<double, 3> gl_nodes = {-sqrt_3_over_5, 0.0, sqrt_3_over_5};
        static constexpr std::array<double, 3> gl_weights = {five_over_nine, eight_over_nine, five_over_nine};

        basis.quadrature_nodes.clear();
        basis.quadrature_weights.clear();

        for (size_t i = basis.degree; i < basis.knots.size() - basis.degree - 1; ++i) {
            double a = basis.knots[i];
            double b = basis.knots[i+1];

            if ( b - a < 1e-14) continue;

            double mid = 0.5 * (a + b);
            double half_width = 0.5 * (b - a);

            for (size_t j = 0; j < gl_nodes.size(); ++j) {
                double phys_node = mid + gl_nodes[j] * half_width;
                double phys_weight = gl_weights[j] * half_width;

                basis.quadrature_nodes.push_back(phys_node);
                basis.quadrature_weights.push_back(phys_weight);

                auto [start, phi] = evaluate_bspline(phys_node, basis);
                basis.quad_evals.push_back({start, phi});
            }
        }

        return basis;
    }


    SpectralSolverStrategy::GridPoint SpectralSolverStrategy::reconstruct_at_quadrature(
        const N_Vector y_coeffs,
        const size_t quad_index,
        const SplineBasis &basis
    ) const {
        auto [start_idx, vals] = basis.quad_evals[quad_index];

        const sunrealtype* T_ptr = N_VGetArrayPointer(m_T_coeffs);
        const sunrealtype* rho_ptr = N_VGetArrayPointer(m_rho_coeffs);
        const sunrealtype* y_data = N_VGetArrayPointer(y_coeffs);

        const size_t num_basis_funcs = basis.knots.size() - basis.degree - 1;
        const std::vector<fourdst::atomic::Species>& species_list = m_engine.getNetworkSpecies();
        const size_t num_species = species_list.size();

        double logT = 0.0;
        double logRho = 0.0;

        for (size_t k = 0; k < vals.size(); ++k) {
            size_t idx = start_idx + k;
            logT += T_ptr[idx] * vals[k];
            logRho += rho_ptr[idx] * vals[k];
        }

        GridPoint result;
        result.T9 = std::pow(10.0, logT) / 1e9;
        result.rho = std::pow(10.0, logRho);

        for (size_t s = 0; s < num_species; ++s) {
            const fourdst::atomic::Species& species = species_list[s];
            double abundance = 0.0;
            const size_t offset = s * num_basis_funcs;

            for (size_t k = 0; k < vals.size(); ++k) {
                abundance += y_data[offset + start_idx + k] * vals[k];
            }

            // Note: It is possible this will lead to a loss of mass conservation. In future we may want to implement a better way to handle this.
            if (abundance < 0.0) abundance = 0.0;

            result.composition.registerSpecies(species);
            result.composition.setMolarAbundance(species, abundance);
        }

        return result;
    }

    std::vector<NetOut> SpectralSolverStrategy::reconstruct_solution(
        const std::vector<NetIn>& original_inputs,
        const std::vector<double>& mass_coordinates,
        const N_Vector final_coeffs,
        const SplineBasis& basis,
        const double dt
    ) const {
        const size_t n_shells = original_inputs.size();
        const size_t num_basis_funcs = basis.knots.size() - basis.degree - 1;

        std::vector<NetOut> outputs;
        outputs.reserve(n_shells);

        const sunrealtype* c_data = N_VGetArrayPointer(final_coeffs);

        const auto& species_list = m_engine.getNetworkSpecies();
        for (size_t shellID = 0; shellID < n_shells; ++shellID) {
            const double x = mass_coordinates[shellID];

            auto [start_idx, vals] = evaluate_bspline(x, basis);

            auto reconstruct_var = [&](const size_t coeff_offset) -> double {
                double result = 0.0;
                for (size_t i = 0; i < vals.size(); ++i) {
                    result += c_data[coeff_offset + start_idx + i] * vals[i];
                }
                return result;
            };

            fourdst::composition::Composition comp_new;

            for (size_t s_idx = 0; s_idx < species_list.size(); ++s_idx) {
                const fourdst::atomic::Species& sp = species_list[s_idx];
                comp_new.registerSpecies(sp);

                const size_t current_offset = s_idx * num_basis_funcs;
                double Y_val = reconstruct_var(current_offset);

                if (Y_val < 0.0 && Y_val > -1.0e-16) {
                    Y_val = 0.0;
                }

                if (Y_val < 0.0 && Y_val > -1e-16) Y_val = 0.0;
                if (Y_val >= 0.0) {
                    comp_new.setMolarAbundance(sp, Y_val);
                }
            }

            const double energy = reconstruct_var(species_list.size() * num_basis_funcs);

            NetOut netOut;
            netOut.composition = comp_new;
            netOut.energy = energy;
            netOut.num_steps = -1; // Not tracked in spectral solver

            outputs.push_back(std::move(netOut));
        }

        return outputs;
    }

    void SpectralSolverStrategy::project_specific_variable(
        const std::vector<NetIn> &current_shells,
        const std::vector<double> &mass_coordinates,
        const std::vector<BasisEval> &shell_cache,
        const DenseLinearSolver &linear_solver,
        N_Vector output_vec,
        size_t output_offset,
        const std::function<double(const NetIn &)> &getter,
        bool use_log
    ) {
        const size_t n_shells = current_shells.size();

        sunrealtype* out_ptr = N_VGetArrayPointer(output_vec);
        size_t basis_size = N_VGetLength(linear_solver.temp_vector);

        for (size_t i = 0; i < basis_size; ++i ) {
            out_ptr[output_offset + i] = 0.0;
        }

        for (size_t shellID = 0; shellID < n_shells; ++shellID) {
            double val = getter(current_shells[shellID]);
            if (use_log) val = std::log10(std::max(val, 1e-100));

            const auto& eval = shell_cache[shellID];

            for (size_t i = 0; i < eval.phi.size(); ++i) {
                out_ptr[output_offset + eval.start_idx + i] += val * eval.phi[i];
            }
        }

        sunrealtype* tmp_data = N_VGetArrayPointer(linear_solver.temp_vector);
        for (size_t i = 0; i < basis_size; ++i) tmp_data[i] = out_ptr[output_offset + i];

        SUNLinSolSolve(linear_solver.LS, linear_solver.A, linear_solver.temp_vector, linear_solver.temp_vector, 0.0);

        for (size_t i = 0; i < basis_size; ++i) out_ptr[output_offset + i] = tmp_data[i];
    }

    ///////////////////////////////////////////////////////////////////////////////
    /// SpectralSolverStrategy::MassMatrixSolver Implementation
    ///////////////////////////////////////////////////////////////////////////////

    SpectralSolverStrategy::DenseLinearSolver::DenseLinearSolver(
        size_t size,
        SUNContext sun_ctx
    ) : ctx(sun_ctx) {
        A = SUNDenseMatrix(size, size, sun_ctx);
        temp_vector = N_VNew_Serial(size, sun_ctx);

        LS = SUNLinSol_Dense(temp_vector, A, sun_ctx);

        if (!A || !temp_vector || !LS) {
            throw std::runtime_error("Failed to create MassMatrixSolver components.");
        }

        zero();
    }

    SpectralSolverStrategy::DenseLinearSolver::~DenseLinearSolver() {
        if (LS) SUNLinSolFree(LS);
        if (A) SUNMatDestroy(A);
        if (temp_vector) N_VDestroy(temp_vector);
    }

    void SpectralSolverStrategy::DenseLinearSolver::zero() const {
        SUNMatZero(A);
    }

    void SpectralSolverStrategy::DenseLinearSolver::init_from_cache(
        const size_t num_basis_funcs,
        const std::vector<BasisEval> &shell_cache
    ) const {
        sunrealtype* a_data = SUNDenseMatrix_Data(A);
        for (const auto&[start_idx, phi] : shell_cache) {
            for (size_t i = 0; i < phi.size(); ++i) {
                const size_t row = start_idx + i;

                for (size_t j = 0; j < phi.size(); ++j) {
                    const size_t col = start_idx + j;

                    a_data[col * num_basis_funcs + row] += phi[i] * phi[j];
                }
            }
        }
        setup();
    }

    void SpectralSolverStrategy::DenseLinearSolver::init_from_basis(
        const size_t num_basis_funcs,
        const SplineBasis &basis
    ) const {
        sunrealtype* m_data = SUNDenseMatrix_Data(A);
        for (size_t q = 0; q < basis.quadrature_nodes.size(); ++q) {
            double w_q = basis.quadrature_weights[q];
            const auto& eval = basis.quad_evals[q];

            for (size_t i = 0; i < eval.phi.size(); ++i) {
                size_t row = eval.start_idx + i;
                for (size_t j = 0; j < eval.phi.size(); ++j) {
                    size_t col = eval.start_idx + j;
                    m_data[col * num_basis_funcs + row] += w_q * eval.phi[j] * eval.phi[i];
                }
            }
        }
        setup();
    }

    void SpectralSolverStrategy::DenseLinearSolver::setup() const {
        utils::check_sundials_flag(SUNLinSolSetup(LS, A), "SUNLinSolSetup - Mass Matrix Solver", utils::SUNDIALS_RET_CODE_TYPES::CVODE);
    }

    // ReSharper disable once CppMemberFunctionMayBeConst
    void SpectralSolverStrategy::DenseLinearSolver::solve_inplace(const N_Vector x, const size_t num_vars, const size_t basis_size) const {
        sunrealtype* x_data = N_VGetArrayPointer(x);
        sunrealtype* tmp_data = N_VGetArrayPointer(temp_vector);

        for (size_t v = 0; v < num_vars; ++v) {
            const size_t offset = v * basis_size;

            for (size_t i = 0; i < basis_size; ++i) {
                tmp_data[i] = x_data[offset + i];
            }
            SUNLinSolSolve(LS, A, temp_vector, temp_vector, 0.0);

            for (size_t i = 0; i < basis_size; ++i) {
                x_data[offset + i] = tmp_data[i];
            }
        }
    }
}