csr.hpp

// SPDX-FileCopyrightText: 2017 - 2025 The Ginkgo authors
//
// SPDX-License-Identifier: BSD-3-Clause
 
#ifndef GKO_PUBLIC_CORE_MATRIX_CSR_HPP_
#define GKO_PUBLIC_CORE_MATRIX_CSR_HPP_
 
 
#include <ginkgo/core/base/array.hpp>
#include <ginkgo/core/base/index_set.hpp>
#include <ginkgo/core/base/lin_op.hpp>
#include <ginkgo/core/base/math.hpp>
#include <ginkgo/core/matrix/permutation.hpp>
#include <ginkgo/core/matrix/scaled_permutation.hpp>
 
 
namespace gko {
namespace matrix {
 
 
template <typename ValueType>
class Dense;
 
template <typename ValueType>
class Diagonal;
 
template <typename ValueType, typename IndexType>
class Coo;
 
template <typename ValueType, typename IndexType>
class Ell;
 
template <typename ValueType, typename IndexType>
class Hybrid;
 
template <typename ValueType, typename IndexType>
class Sellp;
 
template <typename ValueType, typename IndexType>
class SparsityCsr;
 
template <typename ValueType, typename IndexType>
class Csr;
 
template <typename ValueType, typename IndexType>
class Fbcsr;
 
template <typename ValueType, typename IndexType>
class CsrBuilder;
 
template <typename IndexType>
class Permutation;
 
 
namespace detail {
 
 
template <typename ValueType = default_precision, typename IndexType = int32>
void strategy_rebuild_helper(Csr<ValueType, IndexType>* result);
 
 
}  // namespace detail
 

template <typename ValueType = default_precision, typename IndexType = int32>
class Csr : public EnableLinOp<Csr<ValueType, IndexType>>,
            public ConvertibleTo<Csr<next_precision<ValueType>, IndexType>>,
#if GINKGO_ENABLE_HALF || GINKGO_ENABLE_BFLOAT16
            public ConvertibleTo<Csr<next_precision<ValueType, 2>, IndexType>>,
#endif
#if GINKGO_ENABLE_HALF && GINKGO_ENABLE_BFLOAT16
            public ConvertibleTo<Csr<next_precision<ValueType, 3>, IndexType>>,
#endif
            public ConvertibleTo<Dense<ValueType>>,
            public ConvertibleTo<Coo<ValueType, IndexType>>,
            public ConvertibleTo<Ell<ValueType, IndexType>>,
            public ConvertibleTo<Fbcsr<ValueType, IndexType>>,
            public ConvertibleTo<Hybrid<ValueType, IndexType>>,
            public ConvertibleTo<Sellp<ValueType, IndexType>>,
            public ConvertibleTo<SparsityCsr<ValueType, IndexType>>,
            public DiagonalExtractable<ValueType>,
            public ReadableFromMatrixData<ValueType, IndexType>,
            public WritableToMatrixData<ValueType, IndexType>,
            public Transposable,
            public Permutable<IndexType>,
            public EnableAbsoluteComputation<
                remove_complex<Csr<ValueType, IndexType>>>,
            public ScaledIdentityAddable {
    friend class EnablePolymorphicObject<Csr, LinOp>;
    friend class Coo<ValueType, IndexType>;
    friend class Dense<ValueType>;
    friend class Diagonal<ValueType>;
    friend class Ell<ValueType, IndexType>;
    friend class Hybrid<ValueType, IndexType>;
    friend class Sellp<ValueType, IndexType>;
    friend class SparsityCsr<ValueType, IndexType>;
    friend class Fbcsr<ValueType, IndexType>;
    friend class CsrBuilder<ValueType, IndexType>;
    friend class Csr<to_complex<ValueType>, IndexType>;
    GKO_ASSERT_SUPPORTED_VALUE_AND_INDEX_TYPE;
 
public:
    using EnableLinOp<Csr>::convert_to;
    using EnableLinOp<Csr>::move_to;
    using ConvertibleTo<Csr<next_precision<ValueType>, IndexType>>::convert_to;
    using ConvertibleTo<Csr<next_precision<ValueType>, IndexType>>::move_to;
    using ConvertibleTo<Dense<ValueType>>::convert_to;
    using ConvertibleTo<Dense<ValueType>>::move_to;
    using ConvertibleTo<Coo<ValueType, IndexType>>::convert_to;
    using ConvertibleTo<Coo<ValueType, IndexType>>::move_to;
    using ConvertibleTo<Ell<ValueType, IndexType>>::convert_to;
    using ConvertibleTo<Ell<ValueType, IndexType>>::move_to;
    using ConvertibleTo<Fbcsr<ValueType, IndexType>>::convert_to;
    using ConvertibleTo<Fbcsr<ValueType, IndexType>>::move_to;
    using ConvertibleTo<Hybrid<ValueType, IndexType>>::convert_to;
    using ConvertibleTo<Hybrid<ValueType, IndexType>>::move_to;
    using ConvertibleTo<Sellp<ValueType, IndexType>>::convert_to;
    using ConvertibleTo<Sellp<ValueType, IndexType>>::move_to;
    using ConvertibleTo<SparsityCsr<ValueType, IndexType>>::convert_to;
    using ConvertibleTo<SparsityCsr<ValueType, IndexType>>::move_to;
    using ReadableFromMatrixData<ValueType, IndexType>::read;
 
    using value_type = ValueType;
    using index_type = IndexType;
    using transposed_type = Csr<ValueType, IndexType>;
    using mat_data = matrix_data<ValueType, IndexType>;
    using device_mat_data = device_matrix_data<ValueType, IndexType>;
    using absolute_type = remove_complex<Csr>;
 
    class automatical;

    class strategy_type {
        friend class automatical;
 
    public:
        strategy_type(std::string name) : name_(name) {}
 
        virtual ~strategy_type() = default;

        std::string get_name() { return name_; }

        virtual void process(const array<index_type>& mtx_row_ptrs,
                             array<index_type>* mtx_srow) = 0;

        virtual int64_t clac_size(const int64_t nnz) = 0;

        virtual std::shared_ptr<strategy_type> copy() = 0;
 
    protected:
        void set_name(std::string name) { name_ = name; }
 
    private:
        std::string name_;
    };

    class classical : public strategy_type {
    public:
        classical() : strategy_type("classical"), max_length_per_row_(0) {}
 
        void process(const array<index_type>& mtx_row_ptrs,
                     array<index_type>* mtx_srow) override
        {
            auto host_mtx_exec = mtx_row_ptrs.get_executor()->get_master();
            array<index_type> row_ptrs_host(host_mtx_exec);
            const bool is_mtx_on_host{host_mtx_exec ==
                                      mtx_row_ptrs.get_executor()};
            const index_type* row_ptrs{};
            if (is_mtx_on_host) {
                row_ptrs = mtx_row_ptrs.get_const_data();
            } else {
                row_ptrs_host = mtx_row_ptrs;
                row_ptrs = row_ptrs_host.get_const_data();
            }
            auto num_rows = mtx_row_ptrs.get_size() - 1;
            max_length_per_row_ = 0;
            for (size_type i = 0; i < num_rows; i++) {
                max_length_per_row_ = std::max(max_length_per_row_,
                                               row_ptrs[i + 1] - row_ptrs[i]);
            }
        }
 
        int64_t clac_size(const int64_t nnz) override { return 0; }
 
        index_type get_max_length_per_row() const noexcept
        {
            return max_length_per_row_;
        }
 
        std::shared_ptr<strategy_type> copy() override
        {
            return std::make_shared<classical>();
        }
 
    private:
        index_type max_length_per_row_;
    };

    class merge_path : public strategy_type {
    public:
        merge_path() : strategy_type("merge_path") {}
 
        void process(const array<index_type>& mtx_row_ptrs,
                     array<index_type>* mtx_srow) override
        {}
 
        int64_t clac_size(const int64_t nnz) override { return 0; }
 
        std::shared_ptr<strategy_type> copy() override
        {
            return std::make_shared<merge_path>();
        }
    };

    class cusparse : public strategy_type {
    public:
        cusparse() : strategy_type("cusparse") {}
 
        void process(const array<index_type>& mtx_row_ptrs,
                     array<index_type>* mtx_srow) override
        {}
 
        int64_t clac_size(const int64_t nnz) override { return 0; }
 
        std::shared_ptr<strategy_type> copy() override
        {
            return std::make_shared<cusparse>();
        }
    };

    class sparselib : public strategy_type {
    public:
        sparselib() : strategy_type("sparselib") {}
 
        void process(const array<index_type>& mtx_row_ptrs,
                     array<index_type>* mtx_srow) override
        {}
 
        int64_t clac_size(const int64_t nnz) override { return 0; }
 
        std::shared_ptr<strategy_type> copy() override
        {
            return std::make_shared<sparselib>();
        }
    };

    class load_balance : public strategy_type {
    public:
        [[deprecated]] load_balance()
            : load_balance(std::move(
                  gko::CudaExecutor::create(0, gko::OmpExecutor::create())))
        {}

        load_balance(std::shared_ptr<const CudaExecutor> exec)
            : load_balance(exec->get_num_warps(), exec->get_warp_size())
        {}

        load_balance(std::shared_ptr<const HipExecutor> exec)
            : load_balance(exec->get_num_warps(), exec->get_warp_size(), false)
        {}

        load_balance(std::shared_ptr<const DpcppExecutor> exec)
            : load_balance(exec->get_num_subgroups(), 32, false, "intel")
        {}

        load_balance(int64_t nwarps, int warp_size = 32,
                     bool cuda_strategy = true,
                     std::string strategy_name = "none")
            : strategy_type("load_balance"),
              nwarps_(nwarps),
              warp_size_(warp_size),
              cuda_strategy_(cuda_strategy),
              strategy_name_(strategy_name)
        {}
 
        void process(const array<index_type>& mtx_row_ptrs,
                     array<index_type>* mtx_srow) override
        {
            auto nwarps = mtx_srow->get_size();
 
            if (nwarps > 0) {
                auto host_srow_exec = mtx_srow->get_executor()->get_master();
                auto host_mtx_exec = mtx_row_ptrs.get_executor()->get_master();
                const bool is_srow_on_host{host_srow_exec ==
                                           mtx_srow->get_executor()};
                const bool is_mtx_on_host{host_mtx_exec ==
                                          mtx_row_ptrs.get_executor()};
                array<index_type> row_ptrs_host(host_mtx_exec);
                array<index_type> srow_host(host_srow_exec);
                const index_type* row_ptrs{};
                index_type* srow{};
                if (is_srow_on_host) {
                    srow = mtx_srow->get_data();
                } else {
                    srow_host = *mtx_srow;
                    srow = srow_host.get_data();
                }
                if (is_mtx_on_host) {
                    row_ptrs = mtx_row_ptrs.get_const_data();
                } else {
                    row_ptrs_host = mtx_row_ptrs;
                    row_ptrs = row_ptrs_host.get_const_data();
                }
                for (size_type i = 0; i < nwarps; i++) {
                    srow[i] = 0;
                }
                const auto num_rows = mtx_row_ptrs.get_size() - 1;
                const auto num_elems = row_ptrs[num_rows];
                const auto bucket_divider =
                    num_elems > 0 ? ceildiv(num_elems, warp_size_) : 1;
                for (size_type i = 0; i < num_rows; i++) {
                    auto bucket =
                        ceildiv((ceildiv(row_ptrs[i + 1], warp_size_) * nwarps),
                                bucket_divider);
                    if (bucket < nwarps) {
                        srow[bucket]++;
                    }
                }
                // find starting row for thread i
                for (size_type i = 1; i < nwarps; i++) {
                    srow[i] += srow[i - 1];
                }
                if (!is_srow_on_host) {
                    *mtx_srow = srow_host;
                }
            }
        }
 
        int64_t clac_size(const int64_t nnz) override
        {
            if (warp_size_ > 0) {
                int multiple = 8;
                if (nnz >= static_cast<int64_t>(2e8)) {
                    multiple = 2048;
                } else if (nnz >= static_cast<int64_t>(2e7)) {
                    multiple = 512;
                } else if (nnz >= static_cast<int64_t>(2e6)) {
                    multiple = 128;
                } else if (nnz >= static_cast<int64_t>(2e5)) {
                    multiple = 32;
                }
                if (strategy_name_ == "intel") {
                    multiple = 8;
                    if (nnz >= static_cast<int64_t>(2e8)) {
                        multiple = 256;
                    } else if (nnz >= static_cast<int64_t>(2e7)) {
                        multiple = 32;
                    }
                }
#if GINKGO_HIP_PLATFORM_HCC
                if (!cuda_strategy_) {
                    multiple = 8;
                    if (nnz >= static_cast<int64_t>(1e7)) {
                        multiple = 64;
                    } else if (nnz >= static_cast<int64_t>(1e6)) {
                        multiple = 16;
                    }
                }
#endif  // GINKGO_HIP_PLATFORM_HCC
 
                auto nwarps = nwarps_ * multiple;
                return min(ceildiv(nnz, warp_size_), nwarps);
            } else {
                return 0;
            }
        }
 
        std::shared_ptr<strategy_type> copy() override
        {
            return std::make_shared<load_balance>(
                nwarps_, warp_size_, cuda_strategy_, strategy_name_);
        }
 
    private:
        int64_t nwarps_;
        int warp_size_;
        bool cuda_strategy_;
        std::string strategy_name_;
    };
 
    class automatical : public strategy_type {
    public:
        /* Use imbalance strategy when the maximum number of nonzero per row is
         * more than 1024 on NVIDIA hardware */
        const index_type nvidia_row_len_limit = 1024;
        /* Use imbalance strategy when the matrix has more more than 1e6 on
         * NVIDIA hardware */
        const index_type nvidia_nnz_limit{static_cast<index_type>(1e6)};
        /* Use imbalance strategy when the maximum number of nonzero per row is
         * more than 768 on AMD hardware */
        const index_type amd_row_len_limit = 768;
        /* Use imbalance strategy when the matrix has more more than 1e8 on AMD
         * hardware */
        const index_type amd_nnz_limit{static_cast<index_type>(1e8)};
        /* Use imbalance strategy when the maximum number of nonzero per row is
         * more than 25600 on Intel hardware */
        const index_type intel_row_len_limit = 25600;
        /* Use imbalance strategy when the matrix has more more than 3e8 on
         * Intel hardware */
        const index_type intel_nnz_limit{static_cast<index_type>(3e8)};
 
    public:
        [[deprecated]] automatical()
            : automatical(std::move(
                  gko::CudaExecutor::create(0, gko::OmpExecutor::create())))
        {}

        automatical(std::shared_ptr<const CudaExecutor> exec)
            : automatical(exec->get_num_warps(), exec->get_warp_size())
        {}

        automatical(std::shared_ptr<const HipExecutor> exec)
            : automatical(exec->get_num_warps(), exec->get_warp_size(), false)
        {}

        automatical(std::shared_ptr<const DpcppExecutor> exec)
            : automatical(exec->get_num_subgroups(), 32, false, "intel")
        {}

        automatical(int64_t nwarps, int warp_size = 32,
                    bool cuda_strategy = true,
                    std::string strategy_name = "none")
            : strategy_type("automatical"),
              nwarps_(nwarps),
              warp_size_(warp_size),
              cuda_strategy_(cuda_strategy),
              strategy_name_(strategy_name),
              max_length_per_row_(0)
        {}
 
        void process(const array<index_type>& mtx_row_ptrs,
                     array<index_type>* mtx_srow) override
        {
            // if the number of stored elements is larger than <nnz_limit> or
            // the maximum number of stored elements per row is larger than
            // <row_len_limit>, use load_balance otherwise use classical
            index_type nnz_limit = nvidia_nnz_limit;
            index_type row_len_limit = nvidia_row_len_limit;
            if (strategy_name_ == "intel") {
                nnz_limit = intel_nnz_limit;
                row_len_limit = intel_row_len_limit;
            }
#if GINKGO_HIP_PLATFORM_HCC
            if (!cuda_strategy_) {
                nnz_limit = amd_nnz_limit;
                row_len_limit = amd_row_len_limit;
            }
#endif  // GINKGO_HIP_PLATFORM_HCC
            auto host_mtx_exec = mtx_row_ptrs.get_executor()->get_master();
            const bool is_mtx_on_host{host_mtx_exec ==
                                      mtx_row_ptrs.get_executor()};
            array<index_type> row_ptrs_host(host_mtx_exec);
            const index_type* row_ptrs{};
            if (is_mtx_on_host) {
                row_ptrs = mtx_row_ptrs.get_const_data();
            } else {
                row_ptrs_host = mtx_row_ptrs;
                row_ptrs = row_ptrs_host.get_const_data();
            }
            const auto num_rows = mtx_row_ptrs.get_size() - 1;
            if (row_ptrs[num_rows] > nnz_limit) {
                load_balance actual_strategy(nwarps_, warp_size_,
                                             cuda_strategy_, strategy_name_);
                if (is_mtx_on_host) {
                    actual_strategy.process(mtx_row_ptrs, mtx_srow);
                } else {
                    actual_strategy.process(row_ptrs_host, mtx_srow);
                }
                this->set_name(actual_strategy.get_name());
            } else {
                index_type maxnum = 0;
                for (size_type i = 0; i < num_rows; i++) {
                    maxnum = std::max(maxnum, row_ptrs[i + 1] - row_ptrs[i]);
                }
                if (maxnum > row_len_limit) {
                    load_balance actual_strategy(
                        nwarps_, warp_size_, cuda_strategy_, strategy_name_);
                    if (is_mtx_on_host) {
                        actual_strategy.process(mtx_row_ptrs, mtx_srow);
                    } else {
                        actual_strategy.process(row_ptrs_host, mtx_srow);
                    }
                    this->set_name(actual_strategy.get_name());
                } else {
                    classical actual_strategy;
                    if (is_mtx_on_host) {
                        actual_strategy.process(mtx_row_ptrs, mtx_srow);
                        max_length_per_row_ =
                            actual_strategy.get_max_length_per_row();
                    } else {
                        actual_strategy.process(row_ptrs_host, mtx_srow);
                        max_length_per_row_ =
                            actual_strategy.get_max_length_per_row();
                    }
                    this->set_name(actual_strategy.get_name());
                }
            }
        }
 
        int64_t clac_size(const int64_t nnz) override
        {
            return std::make_shared<load_balance>(
                       nwarps_, warp_size_, cuda_strategy_, strategy_name_)
                ->clac_size(nnz);
        }
 
        index_type get_max_length_per_row() const noexcept
        {
            return max_length_per_row_;
        }
 
        std::shared_ptr<strategy_type> copy() override
        {
            return std::make_shared<automatical>(
                nwarps_, warp_size_, cuda_strategy_, strategy_name_);
        }
 
    private:
        int64_t nwarps_;
        int warp_size_;
        bool cuda_strategy_;
        std::string strategy_name_;
        index_type max_length_per_row_;
    };
 
    friend class Csr<previous_precision<ValueType>, IndexType>;
 
    void convert_to(
        Csr<next_precision<ValueType>, IndexType>* result) const override;
 
    void move_to(Csr<next_precision<ValueType>, IndexType>* result) override;
 
#if GINKGO_ENABLE_HALF || GINKGO_ENABLE_BFLOAT16
    friend class Csr<previous_precision<ValueType, 2>, IndexType>;
    using ConvertibleTo<
        Csr<next_precision<ValueType, 2>, IndexType>>::convert_to;
    using ConvertibleTo<Csr<next_precision<ValueType, 2>, IndexType>>::move_to;
 
    void convert_to(
        Csr<next_precision<ValueType, 2>, IndexType>* result) const override;
 
    void move_to(Csr<next_precision<ValueType, 2>, IndexType>* result) override;
#endif
 
#if GINKGO_ENABLE_HALF && GINKGO_ENABLE_BFLOAT16
    friend class Csr<previous_precision<ValueType, 3>, IndexType>;
    using ConvertibleTo<
        Csr<next_precision<ValueType, 3>, IndexType>>::convert_to;
    using ConvertibleTo<Csr<next_precision<ValueType, 3>, IndexType>>::move_to;
 
    void convert_to(
        Csr<next_precision<ValueType, 3>, IndexType>* result) const override;
 
    void move_to(Csr<next_precision<ValueType, 3>, IndexType>* result) override;
#endif
 
    void convert_to(Dense<ValueType>* other) const override;
 
    void move_to(Dense<ValueType>* other) override;
 
    void convert_to(Coo<ValueType, IndexType>* result) const override;
 
    void move_to(Coo<ValueType, IndexType>* result) override;
 
    void convert_to(Ell<ValueType, IndexType>* result) const override;
 
    void move_to(Ell<ValueType, IndexType>* result) override;
 
    void convert_to(Fbcsr<ValueType, IndexType>* result) const override;
 
    void move_to(Fbcsr<ValueType, IndexType>* result) override;
 
    void convert_to(Hybrid<ValueType, IndexType>* result) const override;
 
    void move_to(Hybrid<ValueType, IndexType>* result) override;
 
    void convert_to(Sellp<ValueType, IndexType>* result) const override;
 
    void move_to(Sellp<ValueType, IndexType>* result) override;
 
    void convert_to(SparsityCsr<ValueType, IndexType>* result) const override;
 
    void move_to(SparsityCsr<ValueType, IndexType>* result) override;
 
    void read(const mat_data& data) override;
 
    void read(const device_mat_data& data) override;
 
    void read(device_mat_data&& data) override;
 
    void write(mat_data& data) const override;
 
    std::unique_ptr<LinOp> transpose() const override;
 
    std::unique_ptr<LinOp> conj_transpose() const override;

    class multiply_reuse_info {
        friend class Csr;
 
    public:
        explicit multiply_reuse_info();
 
        ~multiply_reuse_info();
 
        multiply_reuse_info(const multiply_reuse_info&) = delete;
 
        multiply_reuse_info(multiply_reuse_info&&) noexcept;
 
        multiply_reuse_info& operator=(const multiply_reuse_info&) = delete;
 
        multiply_reuse_info& operator=(multiply_reuse_info&&) noexcept;

        void update_values(ptr_param<const Csr> mtx1, ptr_param<const Csr> mtx2,
                           ptr_param<Csr> out) const;
 
    private:
        struct lookup_data;
 
        explicit multiply_reuse_info(std::unique_ptr<lookup_data> data);
 
        std::unique_ptr<lookup_data> internal;
    };

    std::unique_ptr<Csr> multiply(ptr_param<const Csr> other) const;

    std::pair<std::unique_ptr<Csr>, multiply_reuse_info> multiply_reuse(
        ptr_param<const Csr> other) const;

    class multiply_add_reuse_info {
        friend class Csr;
 
    public:
        explicit multiply_add_reuse_info();
 
        ~multiply_add_reuse_info();
 
        multiply_add_reuse_info(const multiply_add_reuse_info&) = delete;
 
        multiply_add_reuse_info(multiply_add_reuse_info&&) noexcept;
 
        multiply_add_reuse_info& operator=(const multiply_add_reuse_info&) =
            delete;
 
        multiply_add_reuse_info& operator=(multiply_add_reuse_info&&) noexcept;

        void update_values(ptr_param<const Csr> mtx,
                           ptr_param<const Dense<value_type>> scale_mult,
                           ptr_param<const Csr> mtx_mult,
                           ptr_param<const Dense<value_type>> scale_add,
                           ptr_param<const Csr> mtx_add,
                           ptr_param<Csr> out) const;
 
    private:
        struct lookup_data;
 
        explicit multiply_add_reuse_info(std::unique_ptr<lookup_data> data);
 
        std::unique_ptr<lookup_data> internal;
    };

    std::unique_ptr<Csr> multiply_add(
        ptr_param<const Dense<value_type>> scale_mult,
        ptr_param<const Csr> mtx_mult,
        ptr_param<const Dense<value_type>> scale_add,
        ptr_param<const Csr> mtx_add) const;

    std::pair<std::unique_ptr<Csr>, multiply_add_reuse_info> multiply_add_reuse(
        ptr_param<const Dense<value_type>> scale_mult,
        ptr_param<const Csr> mtx_mult,
        ptr_param<const Dense<value_type>> scale_add,
        ptr_param<const Csr> mtx_add) const;

    class scale_add_reuse_info {
        friend class Csr;
 
    public:
        explicit scale_add_reuse_info();
 
        ~scale_add_reuse_info();
 
        scale_add_reuse_info(const scale_add_reuse_info&) = delete;
 
        scale_add_reuse_info(scale_add_reuse_info&&) noexcept;
 
        scale_add_reuse_info& operator=(const scale_add_reuse_info&) = delete;
 
        scale_add_reuse_info& operator=(scale_add_reuse_info&&) noexcept;

        void update_values(ptr_param<const Dense<value_type>> scale1,
                           ptr_param<const Csr> mtx1,
                           ptr_param<const Dense<value_type>> scale2,
                           ptr_param<const Csr> mtx2, ptr_param<Csr> out) const;
 
    private:
        struct lookup_data;
 
        explicit scale_add_reuse_info(std::unique_ptr<lookup_data> data);
 
        std::unique_ptr<lookup_data> internal;
    };

    std::unique_ptr<Csr> scale_add(
        ptr_param<const Dense<value_type>> scale_this,
        ptr_param<const Dense<value_type>> scale_other,
        ptr_param<const Csr> mtx_other) const;

    std::pair<std::unique_ptr<Csr>, scale_add_reuse_info> add_scale_reuse(
        ptr_param<const Dense<value_type>> scale_this,
        ptr_param<const Dense<value_type>> scale_other,
        ptr_param<const Csr> mtx_other) const;

    struct permuting_reuse_info {
        explicit permuting_reuse_info();

        explicit permuting_reuse_info(
            std::unique_ptr<Permutation<index_type>> value_permutation);

        void update_values(ptr_param<const Csr> input,
                           ptr_param<Csr> output) const;
 
        std::unique_ptr<Permutation<IndexType>> value_permutation;
    };

    std::pair<std::unique_ptr<Csr>, permuting_reuse_info> transpose_reuse()
        const;

    std::unique_ptr<Csr> permute(
        ptr_param<const Permutation<index_type>> permutation,
        permute_mode mode = permute_mode::symmetric) const;

    std::unique_ptr<Csr> permute(
        ptr_param<const Permutation<index_type>> row_permutation,
        ptr_param<const Permutation<index_type>> column_permutation,
        bool invert = false) const;

    std::pair<std::unique_ptr<Csr>, permuting_reuse_info> permute_reuse(
        ptr_param<const Permutation<index_type>> permutation,
        permute_mode mode = permute_mode::symmetric) const;

    std::pair<std::unique_ptr<Csr>, permuting_reuse_info> permute_reuse(
        ptr_param<const Permutation<index_type>> row_permutation,
        ptr_param<const Permutation<index_type>> column_permutation,
        bool invert = false) const;

    std::unique_ptr<Csr> scale_permute(
        ptr_param<const ScaledPermutation<value_type, index_type>> permutation,
        permute_mode = permute_mode::symmetric) const;

    std::unique_ptr<Csr> scale_permute(
        ptr_param<const ScaledPermutation<value_type, index_type>>
            row_permutation,
        ptr_param<const ScaledPermutation<value_type, index_type>>
            column_permutation,
        bool invert = false) const;
 
    std::unique_ptr<LinOp> permute(
        const array<IndexType>* permutation_indices) const override;
 
    std::unique_ptr<LinOp> inverse_permute(
        const array<IndexType>* inverse_permutation_indices) const override;
 
    std::unique_ptr<LinOp> row_permute(
        const array<IndexType>* permutation_indices) const override;
 
    std::unique_ptr<LinOp> column_permute(
        const array<IndexType>* permutation_indices) const override;
 
    std::unique_ptr<LinOp> inverse_row_permute(
        const array<IndexType>* inverse_permutation_indices) const override;
 
    std::unique_ptr<LinOp> inverse_column_permute(
        const array<IndexType>* inverse_permutation_indices) const override;
 
    std::unique_ptr<Diagonal<ValueType>> extract_diagonal() const override;
 
    std::unique_ptr<absolute_type> compute_absolute() const override;
 
    void compute_absolute_inplace() override;

    void sort_by_column_index();
 
    /*
     * Tests if all row entry pairs (value, col_idx) are sorted by column index
     *
     * @returns True if all row entry pairs (value, col_idx) are sorted by
     *          column index
     */
    bool is_sorted_by_column_index() const;

    value_type* get_values() noexcept { return values_.get_data(); }

    const value_type* get_const_values() const noexcept
    {
        return values_.get_const_data();
    }

    std::unique_ptr<Dense<ValueType>> create_value_view();

    std::unique_ptr<const Dense<ValueType>> create_const_value_view() const;

    index_type* get_col_idxs() noexcept { return col_idxs_.get_data(); }

    const index_type* get_const_col_idxs() const noexcept
    {
        return col_idxs_.get_const_data();
    }

    index_type* get_row_ptrs() noexcept { return row_ptrs_.get_data(); }

    const index_type* get_const_row_ptrs() const noexcept
    {
        return row_ptrs_.get_const_data();
    }

    index_type* get_srow() noexcept { return srow_.get_data(); }

    const index_type* get_const_srow() const noexcept
    {
        return srow_.get_const_data();
    }

    size_type get_num_srow_elements() const noexcept
    {
        return srow_.get_size();
    }

    size_type get_num_stored_elements() const noexcept
    {
        return values_.get_size();
    }

    std::shared_ptr<strategy_type> get_strategy() const noexcept
    {
        return strategy_;
    }

    void set_strategy(std::shared_ptr<strategy_type> strategy)
    {
        strategy_ = std::move(strategy->copy());
        this->make_srow();
    }

    void scale(ptr_param<const LinOp> alpha)
    {
        auto exec = this->get_executor();
        GKO_ASSERT_EQUAL_DIMENSIONS(alpha, dim<2>(1, 1));
        this->scale_impl(make_temporary_clone(exec, alpha).get());
    }

    void inv_scale(ptr_param<const LinOp> alpha)
    {
        auto exec = this->get_executor();
        GKO_ASSERT_EQUAL_DIMENSIONS(alpha, dim<2>(1, 1));
        this->inv_scale_impl(make_temporary_clone(exec, alpha).get());
    }

    static std::unique_ptr<Csr> create(std::shared_ptr<const Executor> exec,
                                       std::shared_ptr<strategy_type> strategy);

    static std::unique_ptr<Csr> create(
        std::shared_ptr<const Executor> exec, const dim<2>& size = {},
        size_type num_nonzeros = {},
        std::shared_ptr<strategy_type> strategy = nullptr);

    static std::unique_ptr<Csr> create(
        std::shared_ptr<const Executor> exec, const dim<2>& size,
        array<value_type> values, array<index_type> col_idxs,
        array<index_type> row_ptrs,
        std::shared_ptr<strategy_type> strategy = nullptr);

    template <typename InputValueType, typename InputColumnIndexType,
              typename InputRowPtrType>
    GKO_DEPRECATED(
        "explicitly construct the gko::array argument instead of passing "
        "initializer lists")
    static std::unique_ptr<Csr> create(
        std::shared_ptr<const Executor> exec, const dim<2>& size,
        std::initializer_list<InputValueType> values,
        std::initializer_list<InputColumnIndexType> col_idxs,
        std::initializer_list<InputRowPtrType> row_ptrs)
    {
        return create(exec, size, array<value_type>{exec, std::move(values)},
                      array<index_type>{exec, std::move(col_idxs)},
                      array<index_type>{exec, std::move(row_ptrs)});
    }

    static std::unique_ptr<const Csr> create_const(
        std::shared_ptr<const Executor> exec, const dim<2>& size,
        gko::detail::const_array_view<ValueType>&& values,
        gko::detail::const_array_view<IndexType>&& col_idxs,
        gko::detail::const_array_view<IndexType>&& row_ptrs,
        std::shared_ptr<strategy_type> strategy = nullptr);

    std::unique_ptr<Csr<ValueType, IndexType>> create_submatrix(
        const index_set<IndexType>& row_index_set,
        const index_set<IndexType>& column_index_set) const;

    std::unique_ptr<Csr<ValueType, IndexType>> create_submatrix(
        const span& row_span, const span& column_span) const;

    Csr& operator=(const Csr&);

    Csr& operator=(Csr&&);

    Csr(const Csr&);

    Csr(Csr&&);
 
protected:
    Csr(std::shared_ptr<const Executor> exec, const dim<2>& size = {},
        size_type num_nonzeros = {},
        std::shared_ptr<strategy_type> strategy = nullptr);
 
    Csr(std::shared_ptr<const Executor> exec, const dim<2>& size,
        array<value_type> values, array<index_type> col_idxs,
        array<index_type> row_ptrs,
        std::shared_ptr<strategy_type> strategy = nullptr);
 
    void apply_impl(const LinOp* b, LinOp* x) const override;
 
    void apply_impl(const LinOp* alpha, const LinOp* b, const LinOp* beta,
                    LinOp* x) const override;
 
    // TODO: This provides some more sane settings. Please fix this!
    static std::shared_ptr<strategy_type> make_default_strategy(
        std::shared_ptr<const Executor> exec)
    {
        auto cuda_exec = std::dynamic_pointer_cast<const CudaExecutor>(exec);
        auto hip_exec = std::dynamic_pointer_cast<const HipExecutor>(exec);
        auto dpcpp_exec = std::dynamic_pointer_cast<const DpcppExecutor>(exec);
        std::shared_ptr<strategy_type> new_strategy;
        if (cuda_exec) {
            new_strategy = std::make_shared<automatical>(cuda_exec);
        } else if (hip_exec) {
            new_strategy = std::make_shared<automatical>(hip_exec);
        } else if (dpcpp_exec) {
            new_strategy = std::make_shared<automatical>(dpcpp_exec);
        } else {
            new_strategy = std::make_shared<classical>();
        }
        return new_strategy;
    }
 
    // TODO clean this up as soon as we improve strategy_type
    template <typename CsrType>
    void convert_strategy_helper(CsrType* result) const
    {
        auto strat = this->get_strategy().get();
        std::shared_ptr<typename CsrType::strategy_type> new_strat;
        if (dynamic_cast<classical*>(strat)) {
            new_strat = std::make_shared<typename CsrType::classical>();
        } else if (dynamic_cast<merge_path*>(strat)) {
            new_strat = std::make_shared<typename CsrType::merge_path>();
        } else if (dynamic_cast<cusparse*>(strat)) {
            new_strat = std::make_shared<typename CsrType::cusparse>();
        } else if (dynamic_cast<sparselib*>(strat)) {
            new_strat = std::make_shared<typename CsrType::sparselib>();
        } else {
            auto rexec = result->get_executor();
            auto cuda_exec =
                std::dynamic_pointer_cast<const CudaExecutor>(rexec);
            auto hip_exec = std::dynamic_pointer_cast<const HipExecutor>(rexec);
            auto dpcpp_exec =
                std::dynamic_pointer_cast<const DpcppExecutor>(rexec);
            auto lb = dynamic_cast<load_balance*>(strat);
            if (cuda_exec) {
                if (lb) {
                    new_strat =
                        std::make_shared<typename CsrType::load_balance>(
                            cuda_exec);
                } else {
                    new_strat = std::make_shared<typename CsrType::automatical>(
                        cuda_exec);
                }
            } else if (hip_exec) {
                if (lb) {
                    new_strat =
                        std::make_shared<typename CsrType::load_balance>(
                            hip_exec);
                } else {
                    new_strat = std::make_shared<typename CsrType::automatical>(
                        hip_exec);
                }
            } else if (dpcpp_exec) {
                if (lb) {
                    new_strat =
                        std::make_shared<typename CsrType::load_balance>(
                            dpcpp_exec);
                } else {
                    new_strat = std::make_shared<typename CsrType::automatical>(
                        dpcpp_exec);
                }
            } else {
                // Try to preserve this executor's configuration
                auto this_cuda_exec =
                    std::dynamic_pointer_cast<const CudaExecutor>(
                        this->get_executor());
                auto this_hip_exec =
                    std::dynamic_pointer_cast<const HipExecutor>(
                        this->get_executor());
                auto this_dpcpp_exec =
                    std::dynamic_pointer_cast<const DpcppExecutor>(
                        this->get_executor());
                if (this_cuda_exec) {
                    if (lb) {
                        new_strat =
                            std::make_shared<typename CsrType::load_balance>(
                                this_cuda_exec);
                    } else {
                        new_strat =
                            std::make_shared<typename CsrType::automatical>(
                                this_cuda_exec);
                    }
                } else if (this_hip_exec) {
                    if (lb) {
                        new_strat =
                            std::make_shared<typename CsrType::load_balance>(
                                this_hip_exec);
                    } else {
                        new_strat =
                            std::make_shared<typename CsrType::automatical>(
                                this_hip_exec);
                    }
                } else if (this_dpcpp_exec) {
                    if (lb) {
                        new_strat =
                            std::make_shared<typename CsrType::load_balance>(
                                this_dpcpp_exec);
                    } else {
                        new_strat =
                            std::make_shared<typename CsrType::automatical>(
                                this_dpcpp_exec);
                    }
                } else {
                    // FIXME: this changes strategies.
                    // We had a load balance or automatical strategy from a non
                    // HIP or Cuda executor and are moving to a non HIP or Cuda
                    // executor.
                    new_strat = std::make_shared<typename CsrType::classical>();
                }
            }
        }
        result->set_strategy(new_strat);
    }

    void make_srow()
    {
        srow_.resize_and_reset(strategy_->clac_size(values_.get_size()));
        strategy_->process(row_ptrs_, &srow_);
    }

    virtual void scale_impl(const LinOp* alpha);

    virtual void inv_scale_impl(const LinOp* alpha);
 
private:
    std::shared_ptr<strategy_type> strategy_;
    array<value_type> values_;
    array<index_type> col_idxs_;
    array<index_type> row_ptrs_;
    array<index_type> srow_;
 
    void add_scaled_identity_impl(const LinOp* a, const LinOp* b) override;
};
 
 
namespace detail {
 

template <typename ValueType, typename IndexType>
void strategy_rebuild_helper(Csr<ValueType, IndexType>* result)
{
    using load_balance = typename Csr<ValueType, IndexType>::load_balance;
    using automatical = typename Csr<ValueType, IndexType>::automatical;
    auto strategy = result->get_strategy();
    auto executor = result->get_executor();
    if (std::dynamic_pointer_cast<load_balance>(strategy)) {
        if (auto exec =
                std::dynamic_pointer_cast<const HipExecutor>(executor)) {
            result->set_strategy(std::make_shared<load_balance>(exec));
        } else if (auto exec = std::dynamic_pointer_cast<const CudaExecutor>(
                       executor)) {
            result->set_strategy(std::make_shared<load_balance>(exec));
        }
    } else if (std::dynamic_pointer_cast<automatical>(strategy)) {
        if (auto exec =
                std::dynamic_pointer_cast<const HipExecutor>(executor)) {
            result->set_strategy(std::make_shared<automatical>(exec));
        } else if (auto exec = std::dynamic_pointer_cast<const CudaExecutor>(
                       executor)) {
            result->set_strategy(std::make_shared<automatical>(exec));
        }
    }
}
 
 
}  // namespace detail
}  // namespace matrix
}  // namespace gko
 
 
#endif  // GKO_PUBLIC_CORE_MATRIX_CSR_HPP_