dense.hpp

/*******************************<GINKGO LICENSE>******************************
Copyright (c) 2017-2019, the Ginkgo authors
All rights reserved.

Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions
are met:

1. Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.

2. Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in the
documentation and/or other materials provided with the distribution.

3. Neither the name of the copyright holder nor the names of its
contributors may be used to endorse or promote products derived from
this software without specific prior written permission.

THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
******************************<GINKGO LICENSE>*******************************/

#ifndef GKO_CORE_MATRIX_DENSE_HPP_
#define GKO_CORE_MATRIX_DENSE_HPP_


#include <ginkgo/core/base/array.hpp>
#include <ginkgo/core/base/executor.hpp>
#include <ginkgo/core/base/lin_op.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 <initializer_list>


namespace gko {
namespace matrix {


template <typename ValueType, typename IndexType>
class Coo;

template <typename ValueType, typename IndexType>
class Csr;

template <typename ValueType, typename IndexType>
class Ell;

template <typename ValueType, typename IndexType>
class Hybrid;

template <typename ValueType, typename IndexType>
class Sellp;


template <typename ValueType = default_precision>
class Dense : public EnableLinOp<Dense<ValueType>>,
              public EnableCreateMethod<Dense<ValueType>>,
              public ConvertibleTo<Coo<ValueType, int32>>,
              public ConvertibleTo<Coo<ValueType, int64>>,
              public ConvertibleTo<Csr<ValueType, int32>>,
              public ConvertibleTo<Csr<ValueType, int64>>,
              public ConvertibleTo<Ell<ValueType, int32>>,
              public ConvertibleTo<Ell<ValueType, int64>>,
              public ConvertibleTo<Hybrid<ValueType, int32>>,
              public ConvertibleTo<Hybrid<ValueType, int64>>,
              public ConvertibleTo<Sellp<ValueType, int32>>,
              public ConvertibleTo<Sellp<ValueType, int64>>,
              public ReadableFromMatrixData<ValueType, int32>,
              public ReadableFromMatrixData<ValueType, int64>,
              public WritableToMatrixData<ValueType, int32>,
              public WritableToMatrixData<ValueType, int64>,
              public Transposable {
    friend class EnableCreateMethod<Dense>;
    friend class EnablePolymorphicObject<Dense, LinOp>;
    friend class Coo<ValueType, int32>;
    friend class Coo<ValueType, int64>;
    friend class Csr<ValueType, int32>;
    friend class Csr<ValueType, int64>;
    friend class Ell<ValueType, int32>;
    friend class Ell<ValueType, int64>;
    friend class Hybrid<ValueType, int32>;
    friend class Hybrid<ValueType, int64>;
    friend class Sellp<ValueType, int32>;
    friend class Sellp<ValueType, int64>;

public:
    using EnableLinOp<Dense>::convert_to;
    using EnableLinOp<Dense>::move_to;

    using value_type = ValueType;
    using index_type = int64;
    using mat_data = gko::matrix_data<ValueType, int64>;
    using mat_data32 = gko::matrix_data<ValueType, int32>;

    using row_major_range = gko::range<gko::accessor::row_major<ValueType, 2>>;

    static std::unique_ptr<Dense> create_with_config_of(const Dense *other)
    {
        // De-referencing `other` before calling the functions (instead of
        // using operator `->`) is currently required to be compatible with
        // CUDA 10.1.
        // Otherwise, it results in a compile error.
        // TODO Check if the compiler error is fixed and revert to `operator->`.
        return Dense::create((*other).get_executor(), (*other).get_size(),
                             (*other).get_stride());
    }

    void convert_to(Coo<ValueType, int32> *result) const override;

    void move_to(Coo<ValueType, int32> *result) override;

    void convert_to(Coo<ValueType, int64> *result) const override;

    void move_to(Coo<ValueType, int64> *result) override;

    void convert_to(Csr<ValueType, int32> *result) const override;

    void move_to(Csr<ValueType, int32> *result) override;

    void convert_to(Csr<ValueType, int64> *result) const override;

    void move_to(Csr<ValueType, int64> *result) override;

    void convert_to(Ell<ValueType, int32> *result) const override;

    void move_to(Ell<ValueType, int32> *result) override;

    void convert_to(Ell<ValueType, int64> *result) const override;

    void move_to(Ell<ValueType, int64> *result) override;

    void convert_to(Hybrid<ValueType, int32> *result) const override;

    void move_to(Hybrid<ValueType, int32> *result) override;

    void convert_to(Hybrid<ValueType, int64> *result) const override;

    void move_to(Hybrid<ValueType, int64> *result) override;

    void convert_to(Sellp<ValueType, int32> *result) const override;

    void move_to(Sellp<ValueType, int32> *result) override;

    void convert_to(Sellp<ValueType, int64> *result) const override;

    void move_to(Sellp<ValueType, int64> *result) override;

    void read(const mat_data &data) override;

    void read(const mat_data32 &data) override;

    void write(mat_data &data) const override;

    void write(mat_data32 &data) const override;

    std::unique_ptr<LinOp> transpose() const override;

    std::unique_ptr<LinOp> conj_transpose() const override;

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

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

    size_type get_stride() const noexcept { return stride_; }

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

    value_type &at(size_type row, size_type col) noexcept
    {
        return values_.get_data()[linearize_index(row, col)];
    }

    value_type at(size_type row, size_type col) const noexcept
    {
        return values_.get_const_data()[linearize_index(row, col)];
    }

    ValueType &at(size_type idx) noexcept
    {
        return values_.get_data()[linearize_index(idx)];
    }

    ValueType at(size_type idx) const noexcept
    {
        return values_.get_const_data()[linearize_index(idx)];
    }

    void scale(const LinOp *alpha)
    {
        auto exec = this->get_executor();
        this->scale_impl(make_temporary_clone(exec, alpha).get());
    }

    void add_scaled(const LinOp *alpha, const LinOp *b)
    {
        auto exec = this->get_executor();
        this->add_scaled_impl(make_temporary_clone(exec, alpha).get(),
                              make_temporary_clone(exec, b).get());
    }

    void compute_dot(const LinOp *b, LinOp *result) const
    {
        auto exec = this->get_executor();
        this->compute_dot_impl(make_temporary_clone(exec, b).get(),
                               make_temporary_clone(exec, result).get());
    }

    void compute_norm2(LinOp *result) const
    {
        auto exec = this->get_executor();
        this->compute_norm2_impl(make_temporary_clone(exec, result).get());
    }

    std::unique_ptr<Dense> create_submatrix(const span &rows,
                                            const span &columns,
                                            const size_type stride)
    {
        row_major_range range_this{this->get_values(), this->get_size()[0],
                                   this->get_size()[1], this->get_stride()};
        auto range_result = range_this(rows, columns);
        // TODO: can result in HUGE padding - which will be copied with the
        // vector
        return Dense::create(
            this->get_executor(),
            dim<2>{range_result.length(0), range_result.length(1)},
            Array<ValueType>::view(
                this->get_executor(),
                range_result.length(0) * range_this.length(1) - columns.begin,
                range_result->data),
            stride);
    }

    /*
     * Create a submatrix from the original matrix.
     *
     * @param rows     row span
     * @param columns  column span
     */
    std::unique_ptr<Dense> create_submatrix(const span &rows,
                                            const span &columns)
    {
        return create_submatrix(rows, columns, this->get_stride());
    }

protected:
    Dense(std::shared_ptr<const Executor> exec, const dim<2> &size = dim<2>{})
        : Dense(std::move(exec), size, size[1])
    {}

    Dense(std::shared_ptr<const Executor> exec, const dim<2> &size,
          size_type stride)
        : EnableLinOp<Dense>(exec, size),
          values_(exec, size[0] * stride),
          stride_(stride)
    {}

    template <typename ValuesArray>
    Dense(std::shared_ptr<const Executor> exec, const dim<2> &size,
          ValuesArray &&values, size_type stride)
        : EnableLinOp<Dense>(exec, size),
          values_{exec, std::forward<ValuesArray>(values)},
          stride_{stride}
    {
        GKO_ENSURE_IN_BOUNDS((size[0] - 1) * stride + size[1] - 1,
                             values_.get_num_elems());
    }

    virtual void scale_impl(const LinOp *alpha);

    virtual void add_scaled_impl(const LinOp *alpha, const LinOp *b);

    virtual void compute_dot_impl(const LinOp *b, LinOp *result) const;

    virtual void compute_norm2_impl(LinOp *result) const;

    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;

    size_type linearize_index(size_type row, size_type col) const noexcept
    {
        return row * stride_ + col;
    }

    size_type linearize_index(size_type idx) const noexcept
    {
        return linearize_index(idx / this->get_size()[1],
                               idx % this->get_size()[1]);
    }

private:
    Array<value_type> values_;
    size_type stride_;
};  // namespace matrix


}  // namespace matrix


template <typename Matrix, typename... TArgs>
std::unique_ptr<Matrix> initialize(
    size_type stride, std::initializer_list<typename Matrix::value_type> vals,
    std::shared_ptr<const Executor> exec, TArgs &&... create_args)
{
    using dense = matrix::Dense<typename Matrix::value_type>;
    size_type num_rows = vals.size();
    auto tmp = dense::create(exec->get_master(), dim<2>{num_rows, 1}, stride);
    size_type idx = 0;
    for (const auto &elem : vals) {
        tmp->at(idx) = elem;
        ++idx;
    }
    auto mtx = Matrix::create(exec, std::forward<TArgs>(create_args)...);
    tmp->move_to(mtx.get());
    return mtx;
}

template <typename Matrix, typename... TArgs>
std::unique_ptr<Matrix> initialize(
    std::initializer_list<typename Matrix::value_type> vals,
    std::shared_ptr<const Executor> exec, TArgs &&... create_args)
{
    return initialize<Matrix>(1, vals, std::move(exec),
                              std::forward<TArgs>(create_args)...);
}


template <typename Matrix, typename... TArgs>
std::unique_ptr<Matrix> initialize(
    size_type stride,
    std::initializer_list<std::initializer_list<typename Matrix::value_type>>
        vals,
    std::shared_ptr<const Executor> exec, TArgs &&... create_args)
{
    using dense = matrix::Dense<typename Matrix::value_type>;
    size_type num_rows = vals.size();
    size_type num_cols = num_rows > 0 ? begin(vals)->size() : 1;
    auto tmp =
        dense::create(exec->get_master(), dim<2>{num_rows, num_cols}, stride);
    size_type ridx = 0;
    for (const auto &row : vals) {
        size_type cidx = 0;
        for (const auto &elem : row) {
            tmp->at(ridx, cidx) = elem;
            ++cidx;
        }
        ++ridx;
    }
    auto mtx = Matrix::create(exec, std::forward<TArgs>(create_args)...);
    tmp->move_to(mtx.get());
    return mtx;
}


template <typename Matrix, typename... TArgs>
std::unique_ptr<Matrix> initialize(
    std::initializer_list<std::initializer_list<typename Matrix::value_type>>
        vals,
    std::shared_ptr<const Executor> exec, TArgs &&... create_args)
{
    return initialize<Matrix>(vals.size() > 0 ? begin(vals)->size() : 0, vals,
                              std::move(exec),
                              std::forward<TArgs>(create_args)...);
}


}  // namespace gko


#endif  // GKO_CORE_MATRIX_DENSE_HPP_