Amesos2_cuSOLVER_decl.hpp

// @HEADER
// *****************************************************************************
//           Amesos2: Templated Direct Sparse Solver Package
//
// Copyright 2011 NTESS and the Amesos2 contributors.
// SPDX-License-Identifier: BSD-3-Clause
// *****************************************************************************
// @HEADER
 
#ifndef AMESOS2_CUSOLVER_DECL_HPP
#define AMESOS2_CUSOLVER_DECL_HPP
 
#include "Amesos2_SolverTraits.hpp"
#include "Amesos2_SolverCore.hpp"
#include "Amesos2_cuSOLVER_FunctionMap.hpp"
 
namespace Amesos2 {
 
template <class Matrix,
          class Vector>
class cuSOLVER : public SolverCore<Amesos2::cuSOLVER, Matrix, Vector>
{
  friend class SolverCore<Amesos2::cuSOLVER,Matrix,Vector>;
 
public:
 
  static const char* name;      // declaration. Initialization outside.
 
  typedef cuSOLVER<Matrix,Vector>                                        type;
  typedef SolverCore<Amesos2::cuSOLVER,Matrix,Vector>              super_type;
 
  typedef typename super_type::scalar_type                        scalar_type;
  typedef typename super_type::local_ordinal_type          local_ordinal_type;
  typedef typename super_type::global_ordinal_type        global_ordinal_type;
  typedef typename super_type::global_size_type              global_size_type;
  typedef typename super_type::node_type                            node_type;
 
  typedef TypeMap<Amesos2::cuSOLVER,scalar_type>                     type_map;
 
  typedef typename type_map::type                               cusolver_type;
  typedef typename type_map::magnitude_type                    magnitude_type;
 
  typedef FunctionMap<Amesos2::cuSOLVER,cusolver_type>           function_map;
 
  #ifdef KOKKOS_ENABLE_CUDA
    // solver will be UVM off
    typedef Kokkos::Device<Kokkos::Cuda, Kokkos::CudaSpace>       device_type;
  #else
    typedef Kokkos::DefaultExecutionSpace::device_type            device_type;
  #endif
 
  typedef int size_type;
  typedef int ordinal_type;
  typedef Kokkos::View<size_type*, device_type>        device_size_type_array;
  typedef Kokkos::View<ordinal_type*, device_type>  device_ordinal_type_array;
  typedef Kokkos::View<cusolver_type*, device_type>   device_value_type_array;
 
 
 
  cuSOLVER(Teuchos::RCP<const Matrix> A,
          Teuchos::RCP<Vector>       X,
          Teuchos::RCP<const Vector> B);
 
 
  ~cuSOLVER( );
 
 
private:
 
  int preOrdering_impl();
 
  int symbolicFactorization_impl();
 
  int numericFactorization_impl();
 
  int solve_impl(const Teuchos::Ptr<MultiVecAdapter<Vector> >       X,
                 const Teuchos::Ptr<const MultiVecAdapter<Vector> > B) const;
 
  bool matrixShapeOK_impl() const;
 
  void setParameters_impl(
    const Teuchos::RCP<Teuchos::ParameterList> & parameterList );
 
  Teuchos::RCP<const Teuchos::ParameterList> getValidParameters_impl() const;
 
  bool loadA_impl(EPhase current_phase);
 
  void describe_impl(Teuchos::FancyOStream &out,
                     const Teuchos::EVerbosityLevel verbLevel) const;
 
  bool do_optimization() const;
 
  // cuSOLVER/cuBLAS handles and options for both threshold-selected paths.
  mutable struct cuSolverData {
    cusolverDnHandle_t dn_handle;
    cusolverSpHandle_t sp_handle;
    csrcholInfo_t      chol_info;
    cusparseMatDescr_t desc;
    cublasHandle_t     blas_handle;
    bool bReorder = false;
    int small_matrix_threshold = 2500;
  } data_;
 
  typedef Kokkos::View<cusolver_type**, Kokkos::LayoutLeft, device_type>
      device_value_type_matrix;
 
  typedef Kokkos::View<cusolver_type**, Kokkos::LayoutLeft, device_type>
      device_solve_array_t;
 
  // Scratch n×n matrix for LU factorization (overwritten by getrf, not used in solve)
  mutable device_value_type_matrix device_matrix_;
 
  // Cached explicit inverse A^{-1} (n×n, column-major); used every solve via GEMM
  mutable device_value_type_matrix device_inverse_;
 
  // Pivot indices from LU factorization (length n); only needed during numericFactorization
  mutable Kokkos::View<int*, device_type> device_ipiv_;
 
  // Scalar device integer for cusolverDn status output
  mutable Kokkos::View<int, device_type> device_info_;
 
  // Factorization workspace (length determined by bufferSize query)
  mutable device_value_type_array buffer_;
 
  // RHS and solution vectors (column-major, n × nrhs)
  mutable device_solve_array_t xValues_;
  mutable device_solve_array_t bValues_;
 
  device_value_type_array device_nzvals_view_;
  device_size_type_array device_row_ptr_view_;
  device_ordinal_type_array device_cols_view_;
  size_t sorted_nnz;
 
  // data for reordering
  typedef Kokkos::View<ordinal_type*, device_type> permute_array_t;
  permute_array_t device_perm_;
  permute_array_t device_peri_;
  mutable device_solve_array_t permute_result_;
 
  mutable Teuchos::RCP<const Tpetra::Map<local_ordinal_type,
                                          global_ordinal_type,
                                          node_type>> distributionMap_;
 
};                              // End class cuSOLVER
 
template <>
struct solver_traits<cuSOLVER> {
#ifdef HAVE_TEUCHOS_COMPLEX
  typedef Meta::make_list6<float, double,
                           std::complex<float>, std::complex<double>,
                           Kokkos::complex<float>, Kokkos::complex<double>>
                           supported_scalars;
#else
  typedef Meta::make_list2<float, double> supported_scalars;
#endif
};
 
template <typename Scalar, typename LocalOrdinal, typename ExecutionSpace>
struct solver_supports_matrix<cuSOLVER,
  KokkosSparse::CrsMatrix<Scalar, LocalOrdinal, ExecutionSpace>> {
  static const bool value = true;
};
 
} // end namespace Amesos2
 
#endif  // AMESOS2_CUSOLVER_DECL_HPP