SolverAdapter.hpp

Go to the documentation of this file.

/*
  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_SOLVERADAPTER_HPP
#define OPM_SOLVERADAPTER_HPP
 
#include <dune/istl/bcrsmatrix.hh>
#include <dune/istl/operators.hh>
#include <dune/istl/paamg/pinfo.hh>
#include <dune/istl/schwarz.hh>
#include <dune/istl/solver.hh>
 
#include <opm/common/ErrorMacros.hpp>
 
#include <opm/simulators/linalg/gpuistl/GpuBlockPreconditioner.hpp>
#include <opm/simulators/linalg/gpuistl/GpuSparseMatrix.hpp>
#include <opm/simulators/linalg/gpuistl/GpuVector.hpp>
#include <opm/simulators/linalg/gpuistl/PreconditionerAdapter.hpp>
#include <opm/simulators/linalg/gpuistl/detail/has_function.hpp>
 
#if HAVE_MPI
#include <opm/simulators/linalg/gpuistl/GpuOwnerOverlapCopy.hpp>
#endif
 
#ifdef OPEN_MPI
#if OPEN_MPI
#include "mpi-ext.h"
#endif
#endif
 
#include <memory>
 
namespace Opm::gpuistl
{
template <class Operator, template <class> class UnderlyingSolver, class X>
class SolverAdapter : public Dune::IterativeSolver<X, X>
{
public:
    using typename Dune::IterativeSolver<X, X>::domain_type;
    using typename Dune::IterativeSolver<X, X>::range_type;
    using typename Dune::IterativeSolver<X, X>::field_type;
    using typename Dune::IterativeSolver<X, X>::real_type;
    using typename Dune::IterativeSolver<X, X>::scalar_real_type;
    static constexpr auto block_size = domain_type::block_type::dimension;
    using XGPU = Opm::gpuistl::GpuVector<real_type>;
 
 
    template<class Comm>
    SolverAdapter(Operator& op,
                  Dune::ScalarProduct<X>& sp,
                  std::shared_ptr<Dune::Preconditioner<X, X>> prec,
                  scalar_real_type reduction,
                  int maxit,
                  int verbose,
                  const Comm& comm)
        : Dune::IterativeSolver<X, X>(op, sp, *prec, reduction, maxit, verbose)
        , m_opOnCPUWithMatrix(op)
        , m_matrix(GpuSparseMatrix<real_type>::fromMatrix(op.getmat()))
        , m_underlyingSolver(constructSolver(prec, reduction, maxit, verbose, comm))
    {
        OPM_ERROR_IF(
            detail::is_a_well_operator<Operator>::value,
            "Currently we only support operators of type MatrixAdapter in the CUDA/HIP solver. "
            "Use --matrix-add-well-contributions=true. "
            "Using WellModelMatrixAdapter with SolverAdapter is not well-defined so it will not work well, or at all.");
    }
 
    virtual void apply(X& x, X& b, double reduction, Dune::InverseOperatorResult& res) override
    {
        // TODO: Can we do this without reimplementing the other function?
        // TODO: [perf] Do we need to update the matrix every time? Probably yes
        m_matrix.updateNonzeroValues(m_opOnCPUWithMatrix.getmat());
 
        if (!m_inputBuffer) {
            m_inputBuffer.reset(new XGPU(b.dim()));
            m_outputBuffer.reset(new XGPU(x.dim()));
        }
 
        m_inputBuffer->copyFromHost(b);
        // TODO: [perf] do we need to copy x here?
        m_outputBuffer->copyFromHost(x);
 
        m_underlyingSolver.apply(*m_outputBuffer, *m_inputBuffer, reduction, res);
 
        // TODO: [perf] do we need to copy b here?
        m_inputBuffer->copyToHost(b);
        m_outputBuffer->copyToHost(x);
    }
    virtual void apply(X& x, X& b, Dune::InverseOperatorResult& res) override
    {
        // TODO: [perf] Do we need to update the matrix every time? Probably yes
        m_matrix.updateNonzeroValues(m_opOnCPUWithMatrix.getmat());
 
        if (!m_inputBuffer) {
            m_inputBuffer.reset(new XGPU(b.dim()));
            m_outputBuffer.reset(new XGPU(x.dim()));
        }
 
        m_inputBuffer->copyFromHost(b);
        // TODO: [perf] do we need to copy x here?
        m_outputBuffer->copyFromHost(x);
 
        m_underlyingSolver.apply(*m_outputBuffer, *m_inputBuffer, res);
 
        // TODO: [perf] do we need to copy b here?
        m_inputBuffer->copyToHost(b);
        m_outputBuffer->copyToHost(x);
    }
 
private:
    Operator& m_opOnCPUWithMatrix;
    GpuSparseMatrix<real_type> m_matrix;
 
    UnderlyingSolver<XGPU> m_underlyingSolver;
 
 
    // TODO: Use a std::forward
    // This is the MPI parallel case (general communication object)
#if HAVE_MPI
    template <class Comm>
    UnderlyingSolver<XGPU> constructSolver(std::shared_ptr<Dune::Preconditioner<X, X>> prec,
                                           scalar_real_type reduction,
                                           int maxit,
                                           int verbose,
                                           const Comm& communication)
    {
        // TODO: See the below TODO over the definition of precHolder in the other overload of constructSolver
        // TODO: We are currently double wrapping preconditioners in the preconditioner factory to be extra
        //       compatible with CPU. Probably a cleaner way eventually would be to do more modifications to the
        //       flexible solver to accomodate the pure GPU better.
        auto precAsHolder = std::dynamic_pointer_cast<PreconditionerHolder<X, X>>(prec);
        if (!precAsHolder) {
            OPM_THROW(std::invalid_argument,
                      "The preconditioner needs to be a CUDA preconditioner (eg. GPUDILU) wrapped in a "
                      "Opm::gpuistl::PreconditionerAdapter wrapped in a "
                      "Opm::gpuistl::GpuBlockPreconditioner. If you are unsure what this means, set "
                      "preconditioner to 'GPUDILU'");
        }
 
        auto preconditionerAdapter = precAsHolder->getUnderlyingPreconditioner();
        auto preconditionerAdapterAsHolder
            = std::dynamic_pointer_cast<PreconditionerHolder<XGPU, XGPU>>(preconditionerAdapter);
        if (!preconditionerAdapterAsHolder) {
            OPM_THROW(std::invalid_argument,
                      "The preconditioner needs to be a CUDA preconditioner (eg. GPUDILU) wrapped in a "
                      "Opm::gpuistl::PreconditionerAdapter wrapped in a "
                      "Opm::gpuistl::GpuBlockPreconditioner. If you are unsure what this means, set "
                      "preconditioner to 'GPUDILU'");
        }
        // We need to get the underlying preconditioner:
        auto preconditionerReallyOnGPU = preconditionerAdapterAsHolder->getUnderlyingPreconditioner();
 
        // Temporary solution use the GPU Direct communication solely based on these prepcrosessor statements
        bool mpiSupportsCudaAwareAtCompileTime = false;
        bool mpiSupportsCudaAwareAtRunTime = false;
 
#if defined(MPIX_CUDA_AWARE_SUPPORT) && MPIX_CUDA_AWARE_SUPPORT
        mpiSupportsCudaAwareAtCompileTime = true;
#endif /* MPIX_CUDA_AWARE_SUPPORT */
 
#if defined(MPIX_CUDA_AWARE_SUPPORT)
        if (1 == MPIX_Query_cuda_support()) {
            mpiSupportsCudaAwareAtRunTime = true;
        }
#endif /* MPIX_CUDA_AWARE_SUPPORT */
 
 
        // TODO add typename Operator communication type as a named type with using
        std::shared_ptr<Opm::gpuistl::GPUSender<real_type, Comm>> gpuComm;
        if (mpiSupportsCudaAwareAtCompileTime && mpiSupportsCudaAwareAtRunTime) {
            gpuComm = std::make_shared<
                Opm::gpuistl::GPUAwareMPISender<real_type, block_size, Comm>>(
                communication);
        } else {
            gpuComm = std::make_shared<
                Opm::gpuistl::GPUObliviousMPISender<real_type, block_size, Comm>>(
                communication);
        }
 
        using CudaCommunication = GpuOwnerOverlapCopy<real_type, block_size, Comm>;
        using SchwarzOperator
            = Dune::OverlappingSchwarzOperator<GpuSparseMatrix<real_type>, XGPU, XGPU, CudaCommunication>;
        auto cudaCommunication = std::make_shared<CudaCommunication>(gpuComm);
 
        auto mpiPreconditioner = std::make_shared<GpuBlockPreconditioner<XGPU, XGPU, CudaCommunication>>(
            preconditionerReallyOnGPU, cudaCommunication);
 
        auto scalarProduct = std::make_shared<Dune::ParallelScalarProduct<XGPU, CudaCommunication>>(
            cudaCommunication, m_opOnCPUWithMatrix.category());
 
 
        // NOTE: Ownership of cudaCommunication is handled by mpiPreconditioner. However, just to make sure we
        //       remember this, we add this check. So remember that we hold one count in this scope, and one in the
        //       GpuBlockPreconditioner instance. We accommodate for the fact that it could be passed around in
        //       GpuBlockPreconditioner, hence we do not test for != 2
        OPM_ERROR_IF(cudaCommunication.use_count() < 2, "Internal error. Shared pointer not owned properly.");
        auto overlappingCudaOperator = std::make_shared<SchwarzOperator>(m_matrix, *cudaCommunication);
 
        return UnderlyingSolver<XGPU>(
            overlappingCudaOperator, scalarProduct, mpiPreconditioner, reduction, maxit, verbose);
    }
#endif
 
    // This is the serial case (specific overload for Dune::Amg::SequentialInformation)
    UnderlyingSolver<XGPU> constructSolver(std::shared_ptr<Dune::Preconditioner<X, X>> prec,
                                           scalar_real_type reduction,
                                           int maxit,
                                           int verbose,
                                           [[maybe_unused]] const Dune::Amg::SequentialInformation& communication)
    {
        // Dune::Amg::SequentialInformation is the serial case
        return constructSolver(prec, reduction, maxit, verbose);
    }
 
    // TODO: Use a std::forward
    UnderlyingSolver<XGPU> constructSolver(std::shared_ptr<Dune::Preconditioner<X, X>> prec,
                                           scalar_real_type reduction,
                                           int maxit,
                                           int verbose)
    {
        // TODO: Fix the reliance on casting here. This is a code smell to a certain degree, and assumes
        //       a certain setup beforehand. The only reason we do it this way is to minimize edits to the
        //       flexible solver. We could design it differently, but keep this for the time being until
        //       we figure out how we want to GPU-ify the rest of the system.
        auto precAsHolder = std::dynamic_pointer_cast<PreconditionerHolder<XGPU, XGPU>>(prec);
        if (!precAsHolder) {
            OPM_THROW(std::invalid_argument,
                      "The preconditioner needs to be a CUDA preconditioner wrapped in a "
                      "Opm::gpuistl::PreconditionerHolder (eg. GPUDILU).");
        }
        auto preconditionerOnGPU = precAsHolder->getUnderlyingPreconditioner();
 
        auto matrixOperator = std::make_shared<Dune::MatrixAdapter<GpuSparseMatrix<real_type>, XGPU, XGPU>>(m_matrix);
        auto scalarProduct = std::make_shared<Dune::SeqScalarProduct<XGPU>>();
        return UnderlyingSolver<XGPU>(matrixOperator, scalarProduct, preconditionerOnGPU, reduction, maxit, verbose);
    }
 
    std::unique_ptr<XGPU> m_inputBuffer;
    std::unique_ptr<XGPU> m_outputBuffer;
};
 
} // namespace Opm::gpuistl
 
#endif