batch_dense.hpp

// SPDX-FileCopyrightText: 2017 - 2025 The Ginkgo authors
//
// SPDX-License-Identifier: BSD-3-Clause
 
#ifndef GKO_PUBLIC_CORE_MATRIX_BATCH_DENSE_HPP_
#define GKO_PUBLIC_CORE_MATRIX_BATCH_DENSE_HPP_
 
 
#include <initializer_list>
#include <vector>
 
#include <ginkgo/core/base/array.hpp>
#include <ginkgo/core/base/batch_lin_op.hpp>
#include <ginkgo/core/base/batch_multi_vector.hpp>
#include <ginkgo/core/base/executor.hpp>
#include <ginkgo/core/base/mtx_io.hpp>
#include <ginkgo/core/base/range_accessors.hpp>
#include <ginkgo/core/base/types.hpp>
#include <ginkgo/core/base/utils.hpp>
#include <ginkgo/core/matrix/dense.hpp>
 
 
namespace gko {
namespace batch {
namespace matrix {
 
 
template <typename ValueType = default_precision>
class Dense final : public EnableBatchLinOp<Dense<ValueType>>,
#if GINKGO_ENABLE_HALF || GINKGO_ENABLE_BFLOAT16
                    public ConvertibleTo<Dense<next_precision<ValueType, 2>>>,
#endif
#if GINKGO_ENABLE_HALF && GINKGO_ENABLE_BFLOAT16
                    public ConvertibleTo<Dense<next_precision<ValueType, 3>>>,
#endif
                    public ConvertibleTo<Dense<next_precision<ValueType>>> {
    friend class EnablePolymorphicObject<Dense, BatchLinOp>;
    friend class Dense<to_complex<ValueType>>;
    friend class Dense<previous_precision<ValueType>>;
 
public:
    using EnableBatchLinOp<Dense>::convert_to;
    using EnableBatchLinOp<Dense>::move_to;
 
    using value_type = ValueType;
    using index_type = int32;
    using transposed_type = Dense<ValueType>;
    using unbatch_type = gko::matrix::Dense<ValueType>;
    using absolute_type = remove_complex<Dense>;
    using complex_type = to_complex<Dense>;
 
    void convert_to(Dense<next_precision<ValueType>>* result) const override;
 
    void move_to(Dense<next_precision<ValueType>>* result) override;
 
#if GINKGO_ENABLE_HALF || GINKGO_ENABLE_BFLOAT16
    friend class Dense<previous_precision<ValueType, 2>>;
    using ConvertibleTo<Dense<next_precision<ValueType, 2>>>::convert_to;
    using ConvertibleTo<Dense<next_precision<ValueType, 2>>>::move_to;
 
    void convert_to(Dense<next_precision<ValueType, 2>>* result) const override;
 
    void move_to(Dense<next_precision<ValueType, 2>>* result) override;
#endif
 
#if GINKGO_ENABLE_HALF && GINKGO_ENABLE_BFLOAT16
    friend class Dense<previous_precision<ValueType, 3>>;
    using ConvertibleTo<Dense<next_precision<ValueType, 3>>>::convert_to;
    using ConvertibleTo<Dense<next_precision<ValueType, 3>>>::move_to;
 
    void convert_to(Dense<next_precision<ValueType, 3>>* result) const override;
 
    void move_to(Dense<next_precision<ValueType, 3>>* result) override;
#endif
 
    std::unique_ptr<unbatch_type> create_view_for_item(size_type item_id);
 
    std::unique_ptr<const unbatch_type> create_const_view_for_item(
        size_type item_id) const;
 
    size_type get_cumulative_offset(size_type batch_id) const
    {
        GKO_ASSERT(batch_id < this->get_num_batch_items());
        return batch_id * this->get_common_size()[0] *
               this->get_common_size()[1];
    }
 
    value_type* get_values() noexcept { return values_.get_data(); }
 
    const value_type* get_const_values() const noexcept
    {
        return values_.get_const_data();
    }
 
    value_type& at(size_type batch_id, size_type row, size_type col)
    {
        GKO_ASSERT(batch_id < this->get_num_batch_items());
        return values_.get_data()[linearize_index(batch_id, row, col)];
    }
 
    value_type at(size_type batch_id, size_type row, size_type col) const
    {
        GKO_ASSERT(batch_id < this->get_num_batch_items());
        return values_.get_const_data()[linearize_index(batch_id, row, col)];
    }
 
    ValueType& at(size_type batch_id, size_type idx) noexcept
    {
        GKO_ASSERT(batch_id < this->get_num_batch_items());
        return values_.get_data()[linearize_index(batch_id, idx)];
    }
 
    ValueType at(size_type batch_id, size_type idx) const noexcept
    {
        GKO_ASSERT(batch_id < this->get_num_batch_items());
        return values_.get_const_data()[linearize_index(batch_id, idx)];
    }
 
    value_type* get_values_for_item(size_type batch_id) noexcept
    {
        GKO_ASSERT(batch_id < this->get_num_batch_items());
        return values_.get_data() + this->get_cumulative_offset(batch_id);
    }
 
    const value_type* get_const_values_for_item(
        size_type batch_id) const noexcept
    {
        GKO_ASSERT(batch_id < this->get_num_batch_items());
        return values_.get_const_data() + this->get_cumulative_offset(batch_id);
    }
 
    size_type get_num_stored_elements() const noexcept
    {
        return values_.get_size();
    }
 
    size_type get_num_elements_per_item() const noexcept
    {
        return this->get_num_stored_elements() / this->get_num_batch_items();
    }
 
    static std::unique_ptr<Dense> create(
        std::shared_ptr<const Executor> exec,
        const batch_dim<2>& size = batch_dim<2>{});
 
    static std::unique_ptr<Dense> create(std::shared_ptr<const Executor> exec,
                                         const batch_dim<2>& size,
                                         array<value_type> values);
 
    template <typename InputValueType>
    GKO_DEPRECATED(
        "explicitly construct the gko::array argument instead of passing an"
        "initializer list")
    static std::unique_ptr<Dense> create(
        std::shared_ptr<const Executor> exec, const batch_dim<2>& size,
        std::initializer_list<InputValueType> values)
    {
        return create(exec, size, array<value_type>{exec, std::move(values)});
    }
 
    static std::unique_ptr<const Dense> create_const(
        std::shared_ptr<const Executor> exec, const batch_dim<2>& sizes,
        gko::detail::const_array_view<ValueType>&& values);
 
    Dense* apply(ptr_param<const MultiVector<value_type>> b,
                 ptr_param<MultiVector<value_type>> x);
 
    Dense* apply(ptr_param<const MultiVector<value_type>> alpha,
                 ptr_param<const MultiVector<value_type>> b,
                 ptr_param<const MultiVector<value_type>> beta,
                 ptr_param<MultiVector<value_type>> x);
 
    const Dense* apply(ptr_param<const MultiVector<value_type>> b,
                       ptr_param<MultiVector<value_type>> x) const;
 
    const Dense* apply(ptr_param<const MultiVector<value_type>> alpha,
                       ptr_param<const MultiVector<value_type>> b,
                       ptr_param<const MultiVector<value_type>> beta,
                       ptr_param<MultiVector<value_type>> x) const;
 
    void scale(const array<value_type>& row_scale,
               const array<value_type>& col_scale);
 
    void scale_add(ptr_param<const MultiVector<value_type>> alpha,
                   ptr_param<const batch::matrix::Dense<value_type>> b);
 
    void add_scaled_identity(ptr_param<const MultiVector<value_type>> alpha,
                             ptr_param<const MultiVector<value_type>> beta);
 
private:
    inline size_type compute_num_elems(const batch_dim<2>& size)
    {
        return size.get_num_batch_items() * size.get_common_size()[0] *
               size.get_common_size()[1];
    }
 
    Dense(std::shared_ptr<const Executor> exec,
          const batch_dim<2>& size = batch_dim<2>{});
 
    Dense(std::shared_ptr<const Executor> exec, const batch_dim<2>& size,
          array<value_type> values);
 
    void apply_impl(const MultiVector<value_type>* b,
                    MultiVector<value_type>* x) const;
 
    void apply_impl(const MultiVector<value_type>* alpha,
                    const MultiVector<value_type>* b,
                    const MultiVector<value_type>* beta,
                    MultiVector<value_type>* x) const;
 
    size_type linearize_index(size_type batch, size_type row,
                              size_type col) const noexcept
    {
        return this->get_cumulative_offset(batch) +
               row * this->get_size().get_common_size()[1] + col;
    }
 
    size_type linearize_index(size_type batch, size_type idx) const noexcept
    {
        return linearize_index(batch, idx / this->get_common_size()[1],
                               idx % this->get_common_size()[1]);
    }
 
    array<value_type> values_;
};
 
 
}  // namespace matrix
}  // namespace batch
}  // namespace gko
 
 
#endif  // GKO_PUBLIC_CORE_MATRIX_BATCH_DENSE_HPP_