GpuDILU.hpp

/*
  Copyright 2022-2023 SINTEF AS

  This file is part of the Open Porous Media project (OPM).

  OPM is free software: you can redistribute it and/or modify
  it under the terms of the GNU General Public License as published by
  the Free Software Foundation, either version 3 of the License, or
  (at your option) any later version.

  OPM is distributed in the hope that it will be useful,
  but WITHOUT ANY WARRANTY; without even the implied warranty of
  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  GNU General Public License for more details.

  You should have received a copy of the GNU General Public License
  along with OPM.  If not, see <http://www.gnu.org/licenses/>.
*/
#ifndef OPM_GPUDILU_HPP
#define OPM_GPUDILU_HPP

#include <map>
#include <memory>
#include <opm/grid/utility/SparseTable.hpp>
#include <opm/simulators/linalg/PreconditionerWithUpdate.hpp>
#include <opm/simulators/linalg/gpuistl/GpuBuffer.hpp>
#include <opm/simulators/linalg/gpuistl/GpuSparseMatrixWrapper.hpp>
#include <opm/simulators/linalg/gpuistl/GpuVector.hpp>
#include <opm/simulators/linalg/gpuistl/detail/kernel_enums.hpp>
#include <opm/simulators/linalg/gpuistl/detail/preconditionerKernels/ILU_variants_helper_kernels.hpp>
#include <opm/simulators/linalg/gpuistl/gpu_resources.hpp>
#include <utility>
#include <vector>


namespace Opm::gpuistl
{
template <class CPUMatrixT, class X, class Y, int l = 1>
class GpuDILU : public Dune::PreconditionerWithUpdate<X, Y>
{
public:
    using domain_type = X;
    using range_type = Y;
    using field_type = typename X::field_type;

    using GPUMatrix = GpuSparseMatrixWrapper<field_type>;
    using FloatMat = GpuSparseMatrixWrapper<float>;
    using FloatVec = GpuVector<float>;

    using matrix_type = GPUMatrix;


    explicit GpuDILU(const GPUMatrix& gpuMatrix,
                     const CPUMatrixT& cpuMatrix,
                     bool splitMatrix,
                     bool tuneKernels,
                     int mixedPrecisionScheme,
                     bool reorder);

    void pre(X& x, Y& b) override;

    void apply(X& v, const Y& d) override;

    void post(X& x) override;

    Dune::SolverCategory::Category category() const override;

    void update() final;

    void reorderAndSplitMatrix(int moveThreadBlockSize);

    void computeDiagonal(int factorizationThreadBlockSize);

    void tuneThreadBlockSizes();


    static constexpr bool shouldCallPre()
    {
        return false;
    }

    static constexpr bool shouldCallPost()
    {
        return false;
    }

    virtual bool hasPerfectUpdate() const override
    {
        return true;
    }


private:
    void apply(X& v, const Y& d, int lowerSolveThreadBlockSize, int upperSolveThreadBlockSize);
    void update(int moveThreadBlockSize, int factorizationThreadBlockSize);
    static constexpr const size_t blocksize_ = CPUMatrixT::block_type::cols;
    Opm::SparseTable<size_t> m_levelSets;
    std::vector<int> m_reorderedToNatural;
    std::vector<int> m_naturalToReordered;
    const GPUMatrix& m_gpuMatrix;
    std::unique_ptr<GPUMatrix> m_gpuMatrixReordered;
    std::unique_ptr<GPUMatrix> m_gpuMatrixReorderedLower;
    std::unique_ptr<GPUMatrix> m_gpuMatrixReorderedUpper;
    std::unique_ptr<GpuVector<field_type>> m_gpuMatrixReorderedDiag;
    std::unique_ptr<FloatMat> m_gpuMatrixReorderedLowerFloat;
    std::unique_ptr<FloatMat> m_gpuMatrixReorderedUpperFloat;
    std::unique_ptr<FloatVec> m_gpuMatrixReorderedDiagFloat;
    std::unique_ptr<FloatVec> m_gpuDInvFloat;
    GpuVector<int> m_gpuNaturalToReorder;
    GpuVector<int> m_gpuReorderToNatural;
    GpuBuffer<size_t> m_gpuLevelSets; // TODO: make this an optional to not spend time on the copy if not needed
    GpuVector<field_type> m_gpuDInv;
    bool m_splitMatrix;
    bool m_tuneThreadBlockSizes;
    const MatrixStorageMPScheme m_mixedPrecisionScheme;
    int m_upperSolveThreadBlockSize = -1;
    int m_lowerSolveThreadBlockSize = -1;
    int m_moveThreadBlockSize = -1;
    int m_DILUFactorizationThreadBlockSize = -1;
    bool m_reorder;

    GpuBuffer<size_t> m_diagIdxs;

    // Graphs for Apply
    std::map<std::pair<field_type*, const field_type*>, GPUGraph> m_apply_graphs;
    std::map<std::pair<field_type*, const field_type*>, GPUGraphExec> m_executableGraphs;

    // Stream for the DILU operations on the GPU
    GPUStream m_stream {};
    // Events for synchronization with main stream
    GPUEvent m_before {};
    GPUEvent m_after {};
};
} // end namespace Opm::gpuistl

#endif