Amesos2_KLU2_def.hpp

Go to the documentation of this file.

// @HEADER
// *****************************************************************************
//           Amesos2: Templated Direct Sparse Solver Package
//
// Copyright 2011 NTESS and the Amesos2 contributors.
// SPDX-License-Identifier: BSD-3-Clause
// *****************************************************************************
// @HEADER

#ifndef AMESOS2_KLU2_DEF_HPP
#define AMESOS2_KLU2_DEF_HPP

#include <Teuchos_Tuple.hpp>
#include <Teuchos_ParameterList.hpp>
#include <Teuchos_StandardParameterEntryValidators.hpp>

#include "Amesos2_SolverCore_def.hpp"
#include "Amesos2_KLU2_decl.hpp"

namespace Amesos2 {


template <class Matrix, class Vector>
KLU2<Matrix,Vector>::KLU2(
  Teuchos::RCP<const Matrix> A,
  Teuchos::RCP<Vector>       X,
  Teuchos::RCP<const Vector> B )
  : SolverCore<Amesos2::KLU2,Matrix,Vector>(A, X, B)
  , transFlag_(0)
  , is_contiguous_(true)
  , use_gather_(true)
{
  ::KLU2::klu_defaults<klu2_dtype, local_ordinal_type> (&(data_.common_)) ;
  data_.symbolic_ = NULL;
  data_.numeric_ = NULL;

  // Override some default options
  // TODO: use data_ here to init
}


template <class Matrix, class Vector>
KLU2<Matrix,Vector>::~KLU2( )
{
  /* Free KLU2 data_types
   * - Matrices
   * - Vectors
   * - Other data
   */
  if (data_.symbolic_ != NULL)
      ::KLU2::klu_free_symbolic<klu2_dtype, local_ordinal_type>
                         (&(data_.symbolic_), &(data_.common_)) ;
  if (data_.numeric_ != NULL)
      ::KLU2::klu_free_numeric<klu2_dtype, local_ordinal_type>
                         (&(data_.numeric_), &(data_.common_)) ;

  // Storage is initialized in numericFactorization_impl()
  //if ( data_.A.Store != NULL ){
      // destoy
  //}

  // only root allocated these SuperMatrices.
  //if ( data_.L.Store != NULL ){       // will only be true for this->root_
      // destroy ..
  //}
}

template <class Matrix, class Vector>
bool
KLU2<Matrix,Vector>::single_proc_optimization() const {
  return (this->root_ && (this->matrixA_->getComm()->getSize() == 1) && is_contiguous_);
}

template<class Matrix, class Vector>
int
KLU2<Matrix,Vector>::preOrdering_impl()
{
  /* TODO: Define what it means for KLU2
   */
#ifdef HAVE_AMESOS2_TIMERS
    Teuchos::TimeMonitor preOrderTimer(this->timers_.preOrderTime_);
#endif

  return(0);
}


template <class Matrix, class Vector>
int
KLU2<Matrix,Vector>::symbolicFactorization_impl()
{
  if (data_.symbolic_ != NULL) {
      ::KLU2::klu_free_symbolic<klu2_dtype, local_ordinal_type>
                         (&(data_.symbolic_), &(data_.common_)) ;
  }

  if ( single_proc_optimization() ) {
    host_ordinal_type_array host_row_ptr_view;
    host_ordinal_type_array host_cols_view;
    this->matrixA_->returnRowPtr_kokkos_view(host_row_ptr_view);
    this->matrixA_->returnColInd_kokkos_view(host_cols_view);
    data_.symbolic_ = ::KLU2::klu_analyze<klu2_dtype, local_ordinal_type>
      ((local_ordinal_type)this->globalNumCols_, host_row_ptr_view.data(),
       host_cols_view.data(), &(data_.common_)) ;
  }
  else
  {
    data_.symbolic_ = ::KLU2::klu_analyze<klu2_dtype, local_ordinal_type>
      ((local_ordinal_type)this->globalNumCols_, host_col_ptr_view_.data(),
       host_rows_view_.data(), &(data_.common_)) ;

  } //end single_process_optim_check = false

  return(0);
}


template <class Matrix, class Vector>
int
KLU2<Matrix,Vector>::numericFactorization_impl()
{
  using Teuchos::as;

  // Cleanup old L and U matrices if we are not reusing a symbolic
  // factorization.  Stores and other data will be allocated in gstrf.
  // Only rank 0 has valid pointers, TODO: for KLU2

  int info = 0;
  if ( this->root_ ) {

    { // Do factorization
#ifdef HAVE_AMESOS2_TIMERS
      Teuchos::TimeMonitor numFactTimer(this->timers_.numFactTime_);
#endif

      if (data_.numeric_ != NULL) {
        ::KLU2::klu_free_numeric<klu2_dtype, local_ordinal_type>
          (&(data_.numeric_), &(data_.common_));
      }

      if ( single_proc_optimization() ) {
        host_ordinal_type_array host_row_ptr_view;
        host_ordinal_type_array host_cols_view;
        this->matrixA_->returnRowPtr_kokkos_view(host_row_ptr_view);
        this->matrixA_->returnColInd_kokkos_view(host_cols_view);
        this->matrixA_->returnValues_kokkos_view(host_nzvals_view_);
        klu2_dtype * pValues = function_map::convert_scalar(host_nzvals_view_.data());
        data_.numeric_ = ::KLU2::klu_factor<klu2_dtype, local_ordinal_type>
          (host_row_ptr_view.data(), host_cols_view.data(), pValues,
           data_.symbolic_, &(data_.common_));
      }
      else {
        klu2_dtype * pValues = function_map::convert_scalar(host_nzvals_view_.data());
        data_.numeric_ = ::KLU2::klu_factor<klu2_dtype, local_ordinal_type>
          (host_col_ptr_view_.data(), host_rows_view_.data(), pValues,
           data_.symbolic_, &(data_.common_));
      } //end single_process_optim_check = false

      // To have a test which confirms a throw, we need MPI to throw on all the
      // ranks. So we delay and broadcast first. Others throws in Amesos2 which
      // happen on just the root rank would also have the same problem if we
      // tested them but we decided to fix just this one for the present. This
      // is the only error/throw we currently have a unit test for.
      if(data_.numeric_ == nullptr) {
        info = 1;
      }

      // This is set after numeric factorization complete as pivoting can be used;
      // In this case, a discrepancy between symbolic and numeric nnz total can occur.
      if(info == 0) { // skip if error code so we don't segfault - will throw
        this->setNnzLU( as<size_t>((data_.numeric_)->lnz) + as<size_t>((data_.numeric_)->unz) );
      }
    } // end scope

  } // end this->root_

  /* All processes should have the same error code */
  Teuchos::broadcast(*(this->matrixA_->getComm()), 0, &info);

  TEUCHOS_TEST_FOR_EXCEPTION(info > 0, std::runtime_error,
      "KLU2 numeric factorization failed(info="+std::to_string(info)+")");

  return(info);
}

template <class Matrix, class Vector>
int
KLU2<Matrix,Vector>::solve_impl(
 const Teuchos::Ptr<MultiVecAdapter<Vector> >  X,
 const Teuchos::Ptr<const MultiVecAdapter<Vector> > B) const
{
  using Teuchos::as;
  int ierr = 0; // returned error code

  const global_size_type ld_rhs = this->root_ ? X->getGlobalLength() : 0;
  const size_t nrhs = X->getGlobalNumVectors();

  bool bDidAssignX;
  bool bDidAssignB;
  bool use_gather = use_gather_; // user param
  use_gather = (use_gather && this->matrixA_->getComm()->getSize() > 1); // only with multiple MPIs
  use_gather = (use_gather && (std::is_same<vector_scalar_type, float>::value ||
                               std::is_same<vector_scalar_type, double>::value)); // only for double or float vectors
  {
#ifdef HAVE_AMESOS2_TIMERS
    Teuchos::TimeMonitor mvConvTimer(this->timers_.vecConvTime_);
#endif
    const bool initialize_data = true;
    const bool do_not_initialize_data = false;
    if ( single_proc_optimization() && nrhs == 1 ) {
      // no msp creation
      bDidAssignB = Util::get_1d_copy_helper_kokkos_view<MultiVecAdapter<Vector>,
        host_solve_array_t>::do_get(initialize_data, B, bValues_, as<size_t>(ld_rhs));

      bDidAssignX = Util::get_1d_copy_helper_kokkos_view<MultiVecAdapter<Vector>,
        host_solve_array_t>::do_get(do_not_initialize_data, X, xValues_, as<size_t>(ld_rhs));
    }
    else {
      if (use_gather) {
        int rval = B->gather(bValues_, this->perm_g2l, this->recvCountRows, this->recvDisplRows,
                             (is_contiguous_ == true) ? ROOTED : CONTIGUOUS_AND_ROOTED);
        if (rval == 0) {
          X->gather(xValues_, this->perm_g2l, this->recvCountRows, this->recvDisplRows,
                    (is_contiguous_ == true) ? ROOTED : CONTIGUOUS_AND_ROOTED);
          bDidAssignB = true;  // TODO : find when we can avoid deep-copy
          bDidAssignX = false; // TODO : find when we can avoid scatter
        } else {
          use_gather = false;
        }
      }
      if (!use_gather) {
        bDidAssignB = Util::get_1d_copy_helper_kokkos_view<MultiVecAdapter<Vector>,
          host_solve_array_t>::do_get(initialize_data, B, bValues_,
            as<size_t>(ld_rhs),
            (is_contiguous_ == true) ? ROOTED : CONTIGUOUS_AND_ROOTED,
            this->rowIndexBase_);
        // see Amesos2_Tacho_def.hpp for an explanation of why we 'get' X
        bDidAssignX = Util::get_1d_copy_helper_kokkos_view<MultiVecAdapter<Vector>,
          host_solve_array_t>::do_get(do_not_initialize_data, X, xValues_,
            as<size_t>(ld_rhs),
            (is_contiguous_ == true) ? ROOTED : CONTIGUOUS_AND_ROOTED,
            this->rowIndexBase_);
      }

      // klu_tsolve is going to put the solution x into the input b.
      // Copy b to x then solve in x.
      // We do not want to solve in b, then copy to x, because if b was assigned
      // then the solve will change b permanently and mess up the next test cycle.
      // However if b was actually a copy (bDidAssignB = false) then we can avoid
      // this deep_copy and just assign xValues_ = bValues_.
      if(bDidAssignB) {
        Kokkos::deep_copy(xValues_, bValues_); // need deep_copy or solve will change adapter's b memory which should never happen
      }
      else {
        xValues_ = bValues_; // safe because bValues_ does not point straight to adapter's memory space
      }
    }
  }

  klu2_dtype * pxValues = function_map::convert_scalar(xValues_.data());
  klu2_dtype * pbValues = function_map::convert_scalar(bValues_.data());

  // can be null for non root
  if( this->root_) {
    TEUCHOS_TEST_FOR_EXCEPTION(pbValues == nullptr,
      std::runtime_error, "Amesos2 Runtime Error: b_vector returned null ");

    TEUCHOS_TEST_FOR_EXCEPTION(pxValues == nullptr,
      std::runtime_error, "Amesos2 Runtime Error: x_vector returned null ");
  }

  if ( single_proc_optimization() && nrhs == 1 ) {
#ifdef HAVE_AMESOS2_TIMERS
    Teuchos::TimeMonitor solveTimer(this->timers_.solveTime_);
#endif

    // For this case, Crs matrix raw pointers were used, so the non-transpose default solve
    // is actually the transpose solve as klu_solve expects Ccs matrix pointers
    // Thus, if the transFlag_ is true, the non-transpose solve should be used
    if (transFlag_ == 0)
    {
      ::KLU2::klu_tsolve2<klu2_dtype, local_ordinal_type>
        (data_.symbolic_, data_.numeric_,
         (local_ordinal_type)this->globalNumCols_,
         (local_ordinal_type)nrhs,
         pbValues, pxValues, &(data_.common_)) ;
    }
    else {
      ::KLU2::klu_solve2<klu2_dtype, local_ordinal_type>
        (data_.symbolic_, data_.numeric_,
         (local_ordinal_type)this->globalNumCols_,
         (local_ordinal_type)nrhs,
         pbValues, pxValues, &(data_.common_)) ;
    }

    /* All processes should have the same error code */
    // Teuchos::broadcast(*(this->getComm()), 0, &ierr);

  } // end single_process_optim_check && nrhs == 1
  else  // single proc optimizations but nrhs > 1,
        // or distributed over processes case
  {
    if ( this->root_ ) {
#ifdef HAVE_AMESOS2_TIMERS
      Teuchos::TimeMonitor solveTimer(this->timers_.solveTime_);
#endif
      if (transFlag_ == 0)
      {
        // For this case, Crs matrix raw pointers were used, so the non-transpose default solve
        // is actually the transpose solve as klu_solve expects Ccs matrix pointers
        // Thus, if the transFlag_ is true, the non-transpose solve should be used
        if ( single_proc_optimization() ) {
        ::KLU2::klu_tsolve<klu2_dtype, local_ordinal_type>
          (data_.symbolic_, data_.numeric_,
           (local_ordinal_type)this->globalNumCols_,
           (local_ordinal_type)nrhs,
           pxValues, &(data_.common_)) ;
        }
        else {
        ::KLU2::klu_solve<klu2_dtype, local_ordinal_type>
          (data_.symbolic_, data_.numeric_,
           (local_ordinal_type)this->globalNumCols_,
           (local_ordinal_type)nrhs,
           pxValues, &(data_.common_)) ;
        }
      }
      else
      {
        // For this case, Crs matrix raw pointers were used, so the non-transpose default solve
        // is actually the transpose solve as klu_solve expects Ccs matrix pointers
        // Thus, if the transFlag_ is true, the non-  transpose solve should be used
        if ( single_proc_optimization() ) {
        ::KLU2::klu_solve<klu2_dtype, local_ordinal_type>
          (data_.symbolic_, data_.numeric_,
           (local_ordinal_type)this->globalNumCols_,
           (local_ordinal_type)nrhs,
           pxValues, &(data_.common_)) ;
        }
        else {
        ::KLU2::klu_tsolve<klu2_dtype, local_ordinal_type>
          (data_.symbolic_, data_.numeric_,
           (local_ordinal_type)this->globalNumCols_,
           (local_ordinal_type)nrhs,
           pxValues, &(data_.common_)) ;
        }
      }
    } // end root_
  } //end else

  // if bDidAssignX, then we solved straight to the adapter's X memory space without
  // requiring additional memory allocation, so the x data is already in place.
  if(!bDidAssignX) {
#ifdef HAVE_AMESOS2_TIMERS
    Teuchos::TimeMonitor redistTimer( this->timers_.vecRedistTime_ );
#endif
    if (use_gather) {
      int rval = X->scatter(xValues_, this->perm_g2l, this->recvCountRows, this->recvDisplRows,
                            (is_contiguous_ == true) ? ROOTED : CONTIGUOUS_AND_ROOTED);
      if (rval != 0) use_gather = false;
    }
    if (!use_gather) {
      Util::put_1d_data_helper_kokkos_view<
        MultiVecAdapter<Vector>,host_solve_array_t>::do_put(X, xValues_,
          as<size_t>(ld_rhs),
          (is_contiguous_ == true) ? ROOTED : CONTIGUOUS_AND_ROOTED,
          this->rowIndexBase_);
    }
  }
  return(ierr);
}


template <class Matrix, class Vector>
bool
KLU2<Matrix,Vector>::matrixShapeOK_impl() const
{
  // The KLU2 factorization routines can handle square as well as
  // rectangular matrices, but KLU2 can only apply the solve routines to
  // square matrices, so we check the matrix for squareness.
  return( this->matrixA_->getGlobalNumRows() == this->matrixA_->getGlobalNumCols() );
}


template <class Matrix, class Vector>
void
KLU2<Matrix,Vector>::setParameters_impl(const Teuchos::RCP<Teuchos::ParameterList> & parameterList )
{
  using Teuchos::RCP;
  using Teuchos::getIntegralValue;
  using Teuchos::ParameterEntryValidator;

  RCP<const Teuchos::ParameterList> valid_params = getValidParameters_impl();

  transFlag_ = this->control_.useTranspose_ ? 1: 0;
  // The KLU2 transpose option can override the Amesos2 option
  if( parameterList->isParameter("Trans") ){
    RCP<const ParameterEntryValidator> trans_validator = valid_params->getEntry("Trans").validator();
    parameterList->getEntry("Trans").setValidator(trans_validator);

    transFlag_ = getIntegralValue<int>(*parameterList, "Trans");
  }

  if( parameterList->isParameter("IsContiguous") ){
    is_contiguous_ = parameterList->get<bool>("IsContiguous");
  }
  if( parameterList->isParameter("UseCustomGather") ){
    use_gather_ = parameterList->get<bool>("UseCustomGather");
  }
}


template <class Matrix, class Vector>
Teuchos::RCP<const Teuchos::ParameterList>
KLU2<Matrix,Vector>::getValidParameters_impl() const
{
  using std::string;
  using Teuchos::tuple;
  using Teuchos::ParameterList;
  using Teuchos::setStringToIntegralParameter;

  static Teuchos::RCP<const Teuchos::ParameterList> valid_params;

  if( is_null(valid_params) )
  {
    Teuchos::RCP<Teuchos::ParameterList> pl = Teuchos::parameterList();

    pl->set("Equil", true, "Whether to equilibrate the system before solve, does nothing now");
    pl->set("IsContiguous", true, "Whether GIDs contiguous");
    pl->set("UseCustomGather", true, "Whether to use new matrix-gather routine");

    setStringToIntegralParameter<int>("Trans", "NOTRANS",
                                      "Solve for the transpose system or not",
                                      tuple<string>("NOTRANS","TRANS","CONJ"),
                                      tuple<string>("Solve with transpose",
                                                    "Do not solve with transpose",
                                                    "Solve with the conjugate transpose"),
                                      tuple<int>(0, 1, 2),
                                      pl.getRawPtr());
    valid_params = pl;
  }

  return valid_params;
}


template <class Matrix, class Vector>
bool
KLU2<Matrix,Vector>::loadA_impl(EPhase current_phase)
{
  using Teuchos::as;
#ifdef HAVE_AMESOS2_TIMERS
  Teuchos::TimeMonitor convTimer(this->timers_.mtxConvTime_);
#endif

  if(current_phase == SOLVE)return(false);

  if ( single_proc_optimization() ) {
    // Do nothing in this case - Crs raw pointers will be used
  }
  else
  {
    // Only the root image needs storage allocated
    if( this->root_ ) {
      if (host_nzvals_view_.extent(0) != this->globalNumNonZeros_) 
        Kokkos::resize(host_nzvals_view_, this->globalNumNonZeros_);
      if (host_rows_view_.extent(0) != this->globalNumNonZeros_)
        Kokkos::resize(host_rows_view_, this->globalNumNonZeros_);
      if (host_col_ptr_view_.extent(0) != (this->globalNumRows_ + 1))
        Kokkos::resize(host_col_ptr_view_, this->globalNumRows_ + 1);
    }
    local_ordinal_type nnz_ret = -1;
    bool use_gather = use_gather_; // user param
    use_gather = (use_gather && this->matrixA_->getComm()->getSize() > 1); // only with multiple MPIs
    use_gather = (use_gather && (std::is_same<scalar_type, float>::value || std::is_same<scalar_type, double>::value)); // only for double or float
    {
#ifdef HAVE_AMESOS2_TIMERS
      Teuchos::TimeMonitor mtxRedistTimer( this->timers_.mtxRedistTime_ );
#endif
      if (use_gather) {
        bool column_major = true;
        if (!is_contiguous_) {
          auto contig_mat = this->matrixA_->reindex(this->contig_rowmap_, this->contig_colmap_, current_phase);
          nnz_ret = contig_mat->gather(host_nzvals_view_, host_rows_view_, host_col_ptr_view_, this->perm_g2l, this->recvCountRows, this->recvDisplRows, this->recvCounts, this->recvDispls,
                                       this->transpose_map, this->nzvals_t, column_major, current_phase);
        } else {
          nnz_ret = this->matrixA_->gather(host_nzvals_view_, host_rows_view_, host_col_ptr_view_, this->perm_g2l, this->recvCountRows, this->recvDisplRows, this->recvCounts, this->recvDispls,
                                           this->transpose_map, this->nzvals_t, column_major, current_phase);
        }
        // gather failed (e.g., not implemened for KokkosCrsMatrix)
        // in case of the failure, it falls back to the original "do_get"
        if (nnz_ret < 0) use_gather = false;
      }
      if (!use_gather) {
        Util::get_ccs_helper_kokkos_view<
          MatrixAdapter<Matrix>,host_value_type_array,host_ordinal_type_array,host_ordinal_type_array>
          ::do_get(this->matrixA_.ptr(), host_nzvals_view_, host_rows_view_, host_col_ptr_view_, nnz_ret,
            (is_contiguous_ == true) ? ROOTED : CONTIGUOUS_AND_ROOTED,
            ARBITRARY,
            this->rowIndexBase_);
      }
    }

    // gather return the total nnz_ret on every MPI process
    if (use_gather || this->root_) {
      TEUCHOS_TEST_FOR_EXCEPTION( nnz_ret != as<local_ordinal_type>(this->globalNumNonZeros_),
                          std::runtime_error,
                          "Amesos2_KLU2 loadA_impl: Did not get the expected number of non-zero vals("
                          +std::to_string(nnz_ret)+" vs "+std::to_string(this->globalNumNonZeros_)+")");
    }
  } //end else single_process_optim_check = false

  return true;
}


template<class Matrix, class Vector>
const char* KLU2<Matrix,Vector>::name = "KLU2";


} // end namespace Amesos2

#endif  // AMESOS2_KLU2_DEF_HPP