GpuVector.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_GPUVECTOR_HEADER_HPP
#define OPM_GPUVECTOR_HEADER_HPP
#include <dune/common/fvector.hh>
#include <dune/istl/bvector.hh>

#include <fmt/core.h>
#include <opm/common/ErrorMacros.hpp>
#include <opm/simulators/linalg/gpuistl/detail/CuBlasHandle.hpp>
#include <opm/simulators/linalg/gpuistl/detail/safe_conversion.hpp>
#include <opm/simulators/linalg/gpuistl/detail/gpu_constants.hpp>
#include <vector>
#include <string>


namespace Opm::gpuistl
{

template <typename T>
class GpuVector
{
public:
    using field_type = T;
    using size_type = size_t;


    GpuVector(const GpuVector<T>& other);

    explicit GpuVector(const std::vector<T>& data);

    GpuVector& operator=(const GpuVector<T>& other);

    template<int BlockDimension>
    explicit GpuVector(const Dune::BlockVector<Dune::FieldVector<T, BlockDimension>>& bvector)
        : GpuVector(bvector.dim())
    {
        copyFromHost(bvector);
    }

    GpuVector& operator=(T scalar);

    GpuVector() : m_dataOnDevice(nullptr), m_numberOfElements(0), m_cuBlasHandle(detail::CuBlasHandle::getInstance()) {}

    explicit GpuVector(const size_t numberOfElements);


    GpuVector(const T* dataOnHost, const size_t numberOfElements);

    virtual ~GpuVector();

    T* data();

    const T* data() const;

    template <int BlockDimension>
    void copyFromHost(const Dune::BlockVector<Dune::FieldVector<T, BlockDimension>>& bvector)
    {
        // TODO: [perf] vector.dim() can be replaced by bvector.N() * BlockDimension
        if (detail::to_size_t(m_numberOfElements) != bvector.dim()) {
            OPM_THROW(std::runtime_error,
                      fmt::format("Given incompatible vector size. GpuVector has size {}, \n"
                                  "however, BlockVector has N() = {}, and dim = {}.",
                                  m_numberOfElements,
                                  bvector.N(),
                                  bvector.dim()));
        }
        const auto dataPointer = static_cast<const T*>(&(bvector[0][0]));
        copyFromHost(dataPointer, m_numberOfElements);
    }

    template <int BlockDimension>
    void copyFromHostAsync(const Dune::BlockVector<Dune::FieldVector<T, BlockDimension>>& bvector, cudaStream_t stream = detail::DEFAULT_STREAM)
    {
        // TODO: [perf] vector.dim() can be replaced by bvector.N() * BlockDimension
        if (detail::to_size_t(m_numberOfElements) != bvector.dim()) {
            OPM_THROW(std::runtime_error,
                      fmt::format("Given incompatible vector size. GpuVector has size {}, \n"
                                  "however, BlockVector has N() = {}, and dim = {}.",
                                  m_numberOfElements,
                                  bvector.N(),
                                  bvector.dim()));
        }
        const auto dataPointer = static_cast<const T*>(&(bvector[0][0]));
        copyFromHostAsync(dataPointer, m_numberOfElements, stream);
    }

    template <int BlockDimension>
    void copyToHost(Dune::BlockVector<Dune::FieldVector<T, BlockDimension>>& bvector) const
    {
        // TODO: [perf] vector.dim() can be replaced by bvector.N() * BlockDimension
        if (detail::to_size_t(m_numberOfElements) != bvector.dim()) {
            OPM_THROW(std::runtime_error,
                      fmt::format(fmt::runtime("Given incompatible vector size. GpuVector has size {},\n"
                                               "however, the BlockVector has has N() = {}, and dim() = {}."),
                                  m_numberOfElements,
                                  bvector.N(),
                                  bvector.dim()));
        }
        const auto dataPointer = static_cast<T*>(&(bvector[0][0]));
        copyToHost(dataPointer, m_numberOfElements);
    }

    template <int BlockDimension>
    void copyToHostAsync(Dune::BlockVector<Dune::FieldVector<T, BlockDimension>>& bvector, cudaStream_t stream = detail::DEFAULT_STREAM) const
    {
        // TODO: [perf] vector.dim() can be replaced by bvector.N() * BlockDimension
        if (detail::to_size_t(m_numberOfElements) != bvector.dim()) {
            OPM_THROW(std::runtime_error,
                      fmt::format("Given incompatible vector size. GpuVector has size {},\n however, the BlockVector "
                                  "has has N() = {}, and dim() = {}.",
                                  m_numberOfElements,
                                  bvector.N(),
                                  bvector.dim()));
        }
        const auto dataPointer = static_cast<T*>(&(bvector[0][0]));
        copyToHostAsync(dataPointer, m_numberOfElements, stream);
    }

    void copyFromHost(const T* dataPointer, size_t numberOfElements);

    void copyFromHostAsync(const T* dataPointer, size_t numberOfElements, cudaStream_t stream = detail::DEFAULT_STREAM);



    void copyToHost(T* dataPointer, size_t numberOfElements) const;

    void copyToHostAsync(T* dataPointer, size_t numberOfElements, cudaStream_t stream = detail::DEFAULT_STREAM) const;



    void copyFromHost(const std::vector<T>& data);

    void copyFromHostAsync(const std::vector<T>& data, cudaStream_t stream = detail::DEFAULT_STREAM);

    void copyToHost(std::vector<T>& data) const;

    void copyToHostAsync(std::vector<T>& data, cudaStream_t stream = detail::DEFAULT_STREAM) const;



    void copyFromDeviceToDevice(const GpuVector<T>& other) const;

    void prepareSendBuf(GpuVector<T>& buffer, const GpuVector<int>& indexSet) const;
    void syncFromRecvBuf(GpuVector<T>& buffer, const GpuVector<int>& indexSet) const;

    GpuVector<T>& operator*=(const T& scalar);

    GpuVector<T>& axpy(T alpha, const GpuVector<T>& y);

    GpuVector<T>& operator+=(const GpuVector<T>& other);

    GpuVector<T>& operator-=(const GpuVector<T>& other);

    T dot(const GpuVector<T>& other) const;

    T two_norm() const;

    T dot(const GpuVector<T>& other, const GpuVector<int>& indexSet, GpuVector<T>& buffer) const;

    T two_norm(const GpuVector<int>& indexSet, GpuVector<T>& buffer) const;


    T dot(const GpuVector<T>& other, const GpuVector<int>& indexSet) const;

    T two_norm(const GpuVector<int>& indexSet) const;


    size_type dim() const;

    void resize(size_t new_size);

    std::vector<T> asStdVector() const;

    template <int blockSize>
    Dune::BlockVector<Dune::FieldVector<T, blockSize>> asDuneBlockVector() const
    {
        OPM_ERROR_IF(dim() % blockSize != 0,
                     fmt::format("blockSize is not a multiple of dim(). Given blockSize = {}, and dim() = {}",
                                 blockSize,
                                 dim()));

        Dune::BlockVector<Dune::FieldVector<T, blockSize>> returnValue(dim() / blockSize);
        copyToHost(returnValue);
        return returnValue;
    }


    void setZeroAtIndexSet(const GpuVector<int>& indexSet);

    // Slow method that creates a string representation of a GpuVector for debug purposes
    std::string toDebugString()
    {
        std::vector<T> v = asStdVector();
        std::string res = "";
        for (T element : v){
            res += std::to_string(element) + " ";
        }
        res += std::to_string(v[v.size()-1]);
        return res;
    }

private:
    T* m_dataOnDevice = nullptr;

    // Note that we store this as int to make sure we are always cublas compatible.
    // This gives the added benefit that a size_t to int conversion error occurs during construction.
    int m_numberOfElements;
    detail::CuBlasHandle& m_cuBlasHandle;

    void assertSameSize(const GpuVector<T>& other) const;
    void assertSameSize(int size) const;

    void assertHasElements() const;
};

} // namespace Opm::gpuistl

// ADL bridge: convert GPU vector to Dune BlockVector and delegate to Dune's writer
namespace Opm::gpuistl
{
template <typename T>
inline void writeMatrixMarket(const GpuVector<T>& vectorOnDevice, std::ostream& ostr)
{
    const auto hostBlockVector = vectorOnDevice.template asDuneBlockVector<1>();
    writeMatrixMarket(hostBlockVector, ostr);
}
} // namespace Opm::gpuistl
#endif