#include <array>
#include <cmath>
#include <limits>
#include <memory>

#include <catch2/catch_approx.hpp>
#include <catch2/catch_test_macros.hpp>
#include "mfem/linalg/densemat.hpp"
#include "serif/backend/mfem/solver/preconditioner/border.hpp"
#include "serif/backend/mfem/solver/normalization/operators.hpp"

namespace
{
    namespace backend = serif::solver::preconditioner::backend::mfem;
    using serif::solver::preconditioner::ApplicationContract;

    /** @brief Build independent dense examples in the legacy tests' row-major notation. */
    [[nodiscard]] mfem::DenseMatrix matrix(int size, std::initializer_list<double> entries)
    {
        REQUIRE(entries.size() == static_cast<std::size_t>(size * size));
        mfem::DenseMatrix result(size);
        auto entry = entries.begin();
        for (int row = 0; row < size; ++row)
        {
            for (int column = 0; column < size; ++column)
            {
                result(row, column) = *entry++;
            }
        }
        return result;
    }

    /** @brief Keep test vectors explicit without relying on MFEM initializer-list support. */
    [[nodiscard]] mfem::Vector vector(std::initializer_list<double> values)
    {
        mfem::Vector result(static_cast<int>(values.size()));
        int index = 0;
        for (double value : values)
        {
            result[index++] = value;
        }
        return result;
    }

    /** @brief Compare complete actions by their absolute error norm. */
    [[nodiscard]] double difference(const mfem::Vector& left, const mfem::Vector& right)
    {
        mfem::Vector result(left);
        result -= right;
        return result.Norml2();
    }

    /** @brief Count physical work while retaining independently specified dense actions. */
    class CountedOperator final : public mfem::Operator
    {
    public:
        /** @brief Freeze the coefficients while allowing failure and work-count instrumentation. */
        explicit CountedOperator(mfem::DenseMatrix coefficients) : Operator(coefficients.Height(), coefficients.Width()), coefficients(std::move(coefficients))
        {
        }

        /** @brief Apply the exact dense action and optionally simulate a rejected numerical callback. */
        void Mult(const mfem::Vector& input, mfem::Vector& output) const override
        {
            ++applications;
            coefficients.Mult(input, output);
            if (nonfinite)
            {
                output[0] = std::numeric_limits<double>::infinity();
            }
        }

        mfem::DenseMatrix coefficients;
        mutable int applications{0};
        bool nonfinite{false};
    };

    /** @brief A nonlinear inner solve exposes the difference between cached linear responses and fresh solves. */
    class VariableInverse final : public mfem::Operator
    {
    public:
        /** @brief Use one unknown so the independently derived correction is unambiguous. */
        VariableInverse() : Operator(1)
        {
        }

        /** @brief Model a right-hand-side-dependent inverse with a known analytic response. */
        void Mult(const mfem::Vector& input, mfem::Vector& output) const override
        {
            ++applications;
            output[0] = evaluate(input[0]);
        }

        /** @brief Evaluate the independent scalar nonlinear response. */
        [[nodiscard]] static double evaluate(double value)
        {
            return value / (2.0 + std::abs(value));
        }

        mutable int applications{0};
    };
} // namespace

TEST_CASE("Dense borders preserve exact factorizations and stationary response work counts", "[preconditioner][border][unit]")
{
    // MeanField specification_border.cpp:845 retains one, two, and four generated scalars and both preparations.
    for (const int count : {1, 2, 4})
    {
        CAPTURE(count);
        for (const int generation : {0, 1})
        {
            CAPTURE(generation);
            auto fields = std::make_shared<CountedOperator>(matrix(3, {1.0 / (2 + 0.25 * generation), 0, 0, 0, 1.0 / (3 + 0.5 * generation), 0, 0, 0, 1.0 / (5 + 0.75 * generation)}));
            auto forward = std::make_shared<mfem::DenseMatrix>(3, count);
            auto reverse = std::make_shared<mfem::DenseMatrix>(count, 3);
            auto diagonal = std::make_shared<mfem::DenseMatrix>(count);
            mfem::DenseMatrix complete(3 + count);
            complete = 0.0;
            complete(0, 0) = 2 + 0.25 * generation;
            complete(1, 1) = 3 + 0.5 * generation;
            complete(2, 2) = 5 + 0.75 * generation;
            for (int row = 0; row < count; ++row)
            {
                for (int column = 0; column < count; ++column)
                {
                    (*diagonal)(row, column) = row == column ? 2.0 + row + 0.5 * generation : 0.01 * (row + column + 1);
                    complete(3 + row, 3 + column) = (*diagonal)(row, column);
                }
                for (int column = 0; column < 3; ++column)
                {
                    (*forward)(column, row) = -0.03 * ((column + 1) * (row + 2));
                    (*reverse)(row, column) = 0.04 * ((row + 1) * (column + 2));
                    complete(column, 3 + row) = (*forward)(column, row);
                    complete(3 + row, column) = (*reverse)(row, column);
                }
            }
            auto prepared = backend::try_bordered_inverse({.action = fields}, {.first_from_second = forward, .second_from_first = reverse}, diagonal);
            REQUIRE(prepared);
            CHECK(fields->applications == count);
            mfem::Vector rhs(3 + count), actual(3 + count), expected(3 + count);
            for (int index = 0; index < rhs.Size(); ++index)
            {
                rhs[index] = 0.25 + 0.17 * (index + 1);
            }
            mfem::DenseMatrixInverse exact(complete);
            exact.Mult(rhs, expected);
            prepared->action->Mult(rhs, actual);
            CHECK(difference(actual, expected) / std::max(1.0, expected.Norml2()) <= 2.0e-13);
            CHECK(fields->applications == count + 1);
            prepared->action->Mult(rhs, rhs);
            CHECK(difference(rhs, expected) / std::max(1.0, expected.Norml2()) <= 2.0e-13);
            CHECK(fields->applications == count + 2);
            if (count == 2)
            {
                auto flexible = backend::try_bordered_inverse({.action = fields, .contract = ApplicationContract::variable}, {.first_from_second = forward, .second_from_first = reverse}, diagonal);
                REQUIRE(flexible);
                const int before = fields->applications;
                flexible->action->Mult(vector({0.42, 0.59, 0.76, 0.93, 1.10}), actual);
                CHECK(difference(actual, expected) / std::max(1.0, expected.Norml2()) <= 2.0e-13);
                CHECK(fields->applications == before + 2);
            }
        }
    }
}

TEST_CASE("Variable border inverses recompute the corrected field solve", "[preconditioner][border][unit]")
{
    const auto fields = std::make_shared<VariableInverse>();
    const auto forward = std::make_shared<mfem::DenseMatrix>(matrix(1, {0.75}));
    const auto reverse = std::make_shared<mfem::DenseMatrix>(matrix(1, {1.25}));
    const auto diagonal = std::make_shared<mfem::DenseMatrix>(matrix(1, {4.0}));
    auto prepared = backend::try_bordered_inverse({.action = fields, .contract = ApplicationContract::variable}, {.first_from_second = forward, .second_from_first = reverse}, diagonal);
    REQUIRE(prepared);
    CHECK(prepared->contract == ApplicationContract::variable);
    CHECK(fields->applications == 1);
    mfem::Vector rhs = vector({2.5, -0.3});
    mfem::Vector actual(2);
    const double response = VariableInverse::evaluate(0.75);
    const double first = VariableInverse::evaluate(rhs[0]);
    const double border = (rhs[1] - 1.25 * first) / (4.0 - 1.25 * response);
    const double corrected = VariableInverse::evaluate(rhs[0] - 0.75 * border);
    prepared->action->Mult(rhs, actual);
    CHECK(fields->applications == 3);
    CHECK(actual[0] == Catch::Approx(corrected).margin(2.0e-14));
    CHECK(actual[1] == Catch::Approx(border).margin(2.0e-14));
    CHECK(std::abs(actual[0] - (first - response * border)) > 1.0e-3);
}

TEST_CASE("An empty border preserves the original field inverse", "[preconditioner][border][unit]")
{
    const backend::Inverse fields{.action = std::make_shared<mfem::DenseMatrix>(matrix(1, {0.5}))};
    auto prepared = backend::try_bordered_inverse(fields, {.first_from_second = std::make_shared<mfem::DenseMatrix>(1, 0), .second_from_first = std::make_shared<mfem::DenseMatrix>(0, 1)}, std::make_shared<mfem::DenseMatrix>(0));
    REQUIRE(prepared);
    CHECK(prepared->action == fields.action);
}

TEST_CASE("Unbordered field selection preserves nonunit normalization and typed packing", "[preconditioner][border][coordinates]")
{
    namespace normalization = serif::solver::normalization;
    namespace scaled = normalization::backend::mfem;

    // Original eight-coordinate fixture: three uneven field groups and two excluded scalar borders.
    auto packed = std::make_shared<mfem::DenseMatrix>(8);
    *packed = 0.0;
    for (int index = 0; index < 6; ++index)
    {
        (*packed)(index, index) = index + 2.0;
    }
    (*packed)(0, 5) = 0.4;
    (*packed)(4, 1) = -0.6;
    const std::array output_order{5, 2, 3, 4, 6, 0, 1, 7};
    const std::array input_order{4, 5, 1, 2, 3, 6, 0, 7};
    auto physical = backend::try_restrict_operator(packed, output_order, input_order);
    REQUIRE(physical);
    auto factors = normalization::make_diagonal_normalization({1.0 / 23, 1.0 / 29, 1.0 / 31, 1.0 / 37, 1.0 / 41, 1.0 / 43, 1.0 / 47, 1.0 / 53}, {0.5, 1.0 / 3, 0.2, 1.0 / 7, 1.0 / 11, 1.0 / 13, 1.0 / 17, 1.0 / 19});
    REQUIRE(factors);
    scaled::ScaledInverseOperator inverse(**physical, *factors);
    mfem::Vector rhs = vector({1, 2, 3, 4, 5, 6, 7, 8});
    const mfem::Vector expected = vector({42.0 / 23, 112.0 / 29, 275.0 / 31, 3.0 / 37, 0, 240.4 / 43, 45.0 / 47, 0});
    mfem::Vector actual(8);
    inverse.Mult(rhs, actual);
    CHECK(difference(actual, expected) < 1.0e-13);
    inverse.Mult(rhs, rhs);
    CHECK(difference(rhs, expected) < 1.0e-13);
}