MueLu_DroppingCommon.hpp

Go to the documentation of this file.

// @HEADER
// *****************************************************************************
//        MueLu: A package for multigrid based preconditioning
//
// Copyright 2012 NTESS and the MueLu contributors.
// SPDX-License-Identifier: BSD-3-Clause
// *****************************************************************************
// @HEADER

#ifndef MUELU_DROPPINGCOMMON_HPP
#define MUELU_DROPPINGCOMMON_HPP

// #define MUELU_COALESCE_DROP_DEBUG 1

#include "Kokkos_Core.hpp"
#if KOKKOS_VERSION >= 40799
#include "KokkosKernels_ArithTraits.hpp"
#else
#include "Kokkos_ArithTraits.hpp"
#endif
#include "Xpetra_Access.hpp"
#include "Xpetra_Matrix.hpp"
#include "Xpetra_VectorFactory.hpp"
#include "MueLu_Utilities.hpp"

namespace MueLu {

enum DecisionType : char {
  UNDECIDED = 0,  // no decision has been taken yet, used for initialization
  KEEP      = 1,  // keeep the entry
  DROP      = 2,  // drop it
  BOUNDARY  = 3   // entry is a boundary
};

namespace Misc {

template <class local_ordinal_type>
class NoOpFunctor {
 public:
  NoOpFunctor() {}

  KOKKOS_FORCEINLINE_FUNCTION
  void operator()(local_ordinal_type rlid) const {
  }
};

template <class local_matrix_type>
class PointwiseDropBoundaryFunctor {
 private:
  using scalar_type         = typename local_matrix_type::value_type;
  using local_ordinal_type  = typename local_matrix_type::ordinal_type;
  using memory_space        = typename local_matrix_type::memory_space;
  using results_view        = Kokkos::View<DecisionType*, memory_space>;
  using boundary_nodes_view = Kokkos::View<const bool*, memory_space>;

  local_matrix_type A;
  boundary_nodes_view boundaryNodes;
  results_view results;

 public:
  PointwiseDropBoundaryFunctor(local_matrix_type& A_, boundary_nodes_view boundaryNodes_, results_view& results_)
    : A(A_)
    , boundaryNodes(boundaryNodes_)
    , results(results_) {}

  KOKKOS_FORCEINLINE_FUNCTION
  void operator()(local_ordinal_type rlid) const {
    auto row                 = A.rowConst(rlid);
    const size_t offset      = A.graph.row_map(rlid);
    const bool isBoundaryRow = boundaryNodes(rlid);
    if (isBoundaryRow) {
      for (local_ordinal_type k = 0; k < row.length; ++k) {
        auto clid           = row.colidx(k);
        results(offset + k) = Kokkos::max(rlid == clid ? KEEP : DROP,
                                          results(offset + k));
      }
    }
  }
};

template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
class PointwiseSymmetricDropBoundaryFunctor {
 private:
  using local_matrix_type   = typename Xpetra::Matrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::local_matrix_type;
  using scalar_type         = typename local_matrix_type::value_type;
  using local_ordinal_type  = typename local_matrix_type::ordinal_type;
  using memory_space        = typename local_matrix_type::memory_space;
  using results_view        = Kokkos::View<DecisionType*, memory_space>;
  using boundary_nodes_view = Kokkos::View<bool*, memory_space>;

  local_matrix_type A;
  boundary_nodes_view boundaryNodes;
  boundary_nodes_view boundaryNodesCol;
  results_view results;

 public:
  PointwiseSymmetricDropBoundaryFunctor(Xpetra::Matrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>& A_, boundary_nodes_view boundaryNodes_, results_view& results_)
    : boundaryNodes(boundaryNodes_)
    , results(results_) {
    A                        = A_.getLocalMatrixDevice();
    auto boundaryNodesColumn = typename boundary_nodes_view::non_const_type("boundaryNodesColumn", A_.getColMap()->getLocalNumElements());
    auto boundaryNodesDomain = typename boundary_nodes_view::non_const_type("boundaryNodesDomain", A_.getDomainMap()->getLocalNumElements());
    Utilities<Scalar, LocalOrdinal, GlobalOrdinal, Node>::DetectDirichletColsAndDomains(A_, boundaryNodes, boundaryNodesColumn, boundaryNodesDomain);
    boundaryNodesCol = boundaryNodesColumn;
  }

  KOKKOS_FORCEINLINE_FUNCTION
  void operator()(local_ordinal_type rlid) const {
    auto row                 = A.rowConst(rlid);
    const size_t offset      = A.graph.row_map(rlid);
    const bool isBoundaryRow = boundaryNodes(rlid);
    for (local_ordinal_type k = 0; k < row.length; ++k) {
      auto clid = row.colidx(k);
      if (isBoundaryRow || boundaryNodesCol(clid))
        results(offset + k) = Kokkos::max(rlid == clid ? KEEP : DROP,
                                          results(offset + k));
    }
  }
};

template <class local_matrix_type>
class VectorDropBoundaryFunctor {
 private:
  using scalar_type             = typename local_matrix_type::value_type;
  using local_ordinal_type      = typename local_matrix_type::ordinal_type;
  using memory_space            = typename local_matrix_type::memory_space;
  using results_view            = Kokkos::View<DecisionType*, memory_space>;
  using boundary_nodes_view     = Kokkos::View<const bool*, memory_space>;
  using block_indices_view_type = Kokkos::View<local_ordinal_type*, memory_space>;

  local_matrix_type A;
  block_indices_view_type point_to_block;
  boundary_nodes_view boundaryNodes;
  results_view results;

 public:
  VectorDropBoundaryFunctor(local_matrix_type& A_, block_indices_view_type point_to_block_, boundary_nodes_view boundaryNodes_, results_view& results_)
    : A(A_)
    , point_to_block(point_to_block_)
    , boundaryNodes(boundaryNodes_)
    , results(results_) {}

  KOKKOS_FORCEINLINE_FUNCTION
  void operator()(local_ordinal_type rlid) const {
    auto row                 = A.rowConst(rlid);
    const size_t offset      = A.graph.row_map(rlid);
    const bool isBoundaryRow = boundaryNodes(point_to_block(rlid));
    if (isBoundaryRow) {
      for (local_ordinal_type k = 0; k < row.length; ++k) {
        auto clid           = row.colidx(k);
        results(offset + k) = Kokkos::max(rlid == clid ? KEEP : DROP,
                                          results(offset + k));
      }
    }
  }
};

template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
class VectorSymmetricDropBoundaryFunctor {
 private:
  using local_matrix_type       = typename Xpetra::Matrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>::local_matrix_type;
  using scalar_type             = typename local_matrix_type::value_type;
  using local_ordinal_type      = typename local_matrix_type::ordinal_type;
  using memory_space            = typename local_matrix_type::memory_space;
  using results_view            = Kokkos::View<DecisionType*, memory_space>;
  using boundary_nodes_view     = Kokkos::View<bool*, memory_space>;
  using block_indices_view_type = Kokkos::View<local_ordinal_type*, memory_space>;

  local_matrix_type A;
  block_indices_view_type point_to_block;
  block_indices_view_type ghosted_point_to_block;
  boundary_nodes_view boundaryNodes;
  boundary_nodes_view boundaryNodesCol;
  results_view results;

 public:
  VectorSymmetricDropBoundaryFunctor(Xpetra::Matrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>& A_, block_indices_view_type point_to_block_, block_indices_view_type ghosted_point_to_block_, boundary_nodes_view boundaryNodes_, results_view& results_)
    : point_to_block(point_to_block_)
    , ghosted_point_to_block(ghosted_point_to_block_)
    , boundaryNodes(boundaryNodes_)
    , results(results_) {
    A                        = A_.getLocalMatrixDevice();
    auto boundaryNodesColumn = typename boundary_nodes_view::non_const_type("boundaryNodesColumn", A_.getColMap()->getLocalNumElements());
    auto boundaryNodesDomain = typename boundary_nodes_view::non_const_type("boundaryNodesDomain", A_.getDomainMap()->getLocalNumElements());
    Utilities<Scalar, LocalOrdinal, GlobalOrdinal, Node>::DetectDirichletColsAndDomains(A_, boundaryNodes, boundaryNodesColumn, boundaryNodesDomain);
    boundaryNodesCol = boundaryNodesColumn;
  }

  KOKKOS_FORCEINLINE_FUNCTION
  void operator()(local_ordinal_type rlid) const {
    auto row                 = A.rowConst(rlid);
    const size_t offset      = A.graph.row_map(rlid);
    const bool isBoundaryRow = boundaryNodes(point_to_block(rlid));
    for (local_ordinal_type k = 0; k < row.length; ++k) {
      auto clid = row.colidx(k);
      if (isBoundaryRow || boundaryNodesCol(ghosted_point_to_block(clid))) {
        results(offset + k) = Kokkos::max(rlid == clid ? KEEP : DROP,
                                          results(offset + k));
      }
    }
  }
};

template <class local_matrix_type>
class KeepDiagonalFunctor {
 private:
  using scalar_type        = typename local_matrix_type::value_type;
  using local_ordinal_type = typename local_matrix_type::ordinal_type;
  using memory_space       = typename local_matrix_type::memory_space;
  using results_view       = Kokkos::View<DecisionType*, memory_space>;

  local_matrix_type A;
  results_view results;

 public:
  KeepDiagonalFunctor(local_matrix_type& A_, results_view& results_)
    : A(A_)
    , results(results_) {}

  KOKKOS_FORCEINLINE_FUNCTION
  void operator()(local_ordinal_type rlid) const {
    auto row            = A.rowConst(rlid);
    const size_t offset = A.graph.row_map(rlid);
    for (local_ordinal_type k = 0; k < row.length; ++k) {
      auto clid = row.colidx(k);
      if ((rlid == clid) && (results(offset + k) != BOUNDARY)) {
        results(offset + k) = KEEP;
        break;
      }
    }
  }
};

template <class local_matrix_type>
class DropOffRankFunctor {
 private:
  using scalar_type        = typename local_matrix_type::value_type;
  using local_ordinal_type = typename local_matrix_type::ordinal_type;
  using memory_space       = typename local_matrix_type::memory_space;
  using results_view       = Kokkos::View<DecisionType*, memory_space>;

  local_matrix_type A;
  results_view results;

 public:
  DropOffRankFunctor(local_matrix_type& A_, results_view& results_)
    : A(A_)
    , results(results_) {}

  KOKKOS_FORCEINLINE_FUNCTION
  void operator()(local_ordinal_type rlid) const {
    auto row            = A.rowConst(rlid);
    const size_t offset = A.graph.row_map(rlid);
    for (local_ordinal_type k = 0; k < row.length; ++k) {
      auto clid = row.colidx(k);
      if (clid >= A.numRows()) {
        results(offset + k) = Kokkos::max(DROP, results(offset + k));
      }
    }
  }
};

template <class local_matrix_type>
class MarkSingletonFunctor {
 private:
  using scalar_type        = typename local_matrix_type::value_type;
  using local_ordinal_type = typename local_matrix_type::ordinal_type;
  using memory_space       = typename local_matrix_type::memory_space;
  using results_view       = Kokkos::View<DecisionType*, memory_space>;

  using boundary_nodes_view = Kokkos::View<bool*, memory_space>;

  local_matrix_type A;
  boundary_nodes_view boundaryNodes;
  results_view results;

 public:
  MarkSingletonFunctor(local_matrix_type& A_, boundary_nodes_view boundaryNodes_, results_view& results_)
    : A(A_)
    , boundaryNodes(boundaryNodes_)
    , results(results_) {}

  KOKKOS_FORCEINLINE_FUNCTION
  void operator()(local_ordinal_type rlid) const {
    auto row            = A.rowConst(rlid);
    const size_t offset = A.graph.row_map(rlid);
    for (local_ordinal_type k = 0; k < row.length; ++k) {
      auto clid = row.colidx(k);
      if ((results(offset + k) == KEEP) && (rlid != clid))
        return;
    }
    boundaryNodes(rlid) = true;
    for (local_ordinal_type k = 0; k < row.length; ++k) {
      auto clid = row.colidx(k);
      if (rlid == clid)
        results(offset + k) = KEEP;
      else
        results(offset + k) = BOUNDARY;
    }
  }
};

template <class local_matrix_type>
class MarkSingletonVectorFunctor {
 private:
  using scalar_type             = typename local_matrix_type::value_type;
  using local_ordinal_type      = typename local_matrix_type::ordinal_type;
  using memory_space            = typename local_matrix_type::memory_space;
  using results_view            = Kokkos::View<DecisionType*, memory_space>;
  using block_indices_view_type = Kokkos::View<local_ordinal_type*, memory_space>;

  using boundary_nodes_view = Kokkos::View<bool*, memory_space>;

  local_matrix_type A;
  block_indices_view_type point_to_block;
  boundary_nodes_view boundaryNodes;
  results_view results;

 public:
  MarkSingletonVectorFunctor(local_matrix_type& A_, block_indices_view_type point_to_block_, boundary_nodes_view boundaryNodes_, results_view& results_)
    : A(A_)
    , point_to_block(point_to_block_)
    , boundaryNodes(boundaryNodes_)
    , results(results_) {}

  KOKKOS_FORCEINLINE_FUNCTION
  void operator()(local_ordinal_type rlid) const {
    auto row            = A.rowConst(rlid);
    const size_t offset = A.graph.row_map(rlid);
    for (local_ordinal_type k = 0; k < row.length; ++k) {
      auto clid = row.colidx(k);
      if ((results(offset + k) == KEEP) && (rlid != clid))
        return;
    }
    auto brlid           = point_to_block(rlid);
    boundaryNodes(brlid) = true;
    for (local_ordinal_type k = 0; k < row.length; ++k) {
      auto clid = row.colidx(k);
      if (rlid == clid)
        results(offset + k) = KEEP;
      else
        results(offset + k) = BOUNDARY;
    }
  }
};

template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
class BlockDiagonalizeFunctor {
 private:
  using matrix_type       = Xpetra::Matrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>;
  using local_matrix_type = typename matrix_type::local_matrix_type;

  using scalar_type        = typename local_matrix_type::value_type;
  using local_ordinal_type = typename local_matrix_type::ordinal_type;
  using memory_space       = typename local_matrix_type::memory_space;
  using results_view       = Kokkos::View<DecisionType*, memory_space>;

  using block_indices_type            = Xpetra::MultiVector<LocalOrdinal, LocalOrdinal, GlobalOrdinal, Node>;
  using local_block_indices_view_type = typename block_indices_type::dual_view_type_const::t_dev;

  local_matrix_type A;
  local_block_indices_view_type point_to_block;
  local_block_indices_view_type ghosted_point_to_block;
  results_view results;

 public:
  BlockDiagonalizeFunctor(matrix_type& A_, block_indices_type& point_to_block_, results_view& results_)
    : A(A_.getLocalMatrixDevice())
    , point_to_block(point_to_block_.getLocalViewDevice(Xpetra::Access::ReadOnly))
    , results(results_) {
    auto importer = A_.getCrsGraph()->getImporter();

    if (!importer.is_null()) {
      auto ghosted_point_to_blockMV = Xpetra::VectorFactory<LocalOrdinal, LocalOrdinal, GlobalOrdinal, Node>::Build(importer->getTargetMap());
      ghosted_point_to_blockMV->doImport(point_to_block_, *importer, Xpetra::INSERT);
      ghosted_point_to_block = ghosted_point_to_blockMV->getLocalViewDevice(Xpetra::Access::ReadOnly);
    } else
      ghosted_point_to_block = point_to_block;
  }

  KOKKOS_FORCEINLINE_FUNCTION
  void operator()(local_ordinal_type rlid) const {
    auto row            = A.rowConst(rlid);
    const size_t offset = A.graph.row_map(rlid);
    for (local_ordinal_type k = 0; k < row.length; ++k) {
      auto clid = row.colidx(k);
      if (point_to_block(rlid, 0) == ghosted_point_to_block(clid, 0)) {
        results(offset + k) = Kokkos::max(KEEP, results(offset + k));
      } else {
        results(offset + k) = Kokkos::max(DROP, results(offset + k));
      }
    }
  }
};

template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
class BlockDiagonalizeVectorFunctor {
 private:
  using matrix_type       = Xpetra::Matrix<Scalar, LocalOrdinal, GlobalOrdinal, Node>;
  using local_matrix_type = typename matrix_type::local_matrix_type;

  using scalar_type        = typename local_matrix_type::value_type;
  using local_ordinal_type = typename local_matrix_type::ordinal_type;
  using memory_space       = typename local_matrix_type::memory_space;
  using results_view       = Kokkos::View<DecisionType*, memory_space>;

  using block_indices_type            = Xpetra::Vector<LocalOrdinal, LocalOrdinal, GlobalOrdinal, Node>;
  using local_block_indices_view_type = typename block_indices_type::dual_view_type_const::t_dev;
  using id_translation_type           = Kokkos::View<local_ordinal_type*, memory_space>;
  using importer_type                 = Xpetra::Import<LocalOrdinal, GlobalOrdinal, Node>;

  local_matrix_type A;
  local_block_indices_view_type point_to_block;
  local_block_indices_view_type ghosted_point_to_block;
  results_view results;
  id_translation_type row_translation;
  id_translation_type col_translation;
  Teuchos::RCP<block_indices_type> ghosted_point_to_blockMV;

 public:
  BlockDiagonalizeVectorFunctor(matrix_type& A_, block_indices_type& point_to_block_, const RCP<const importer_type>& importer, results_view& results_, id_translation_type row_translation_, id_translation_type col_translation_)
    : A(A_.getLocalMatrixDevice())
    , point_to_block(point_to_block_.getLocalViewDevice(Xpetra::Access::ReadOnly))
    , results(results_)
    , row_translation(row_translation_)
    , col_translation(col_translation_) {
    if (!importer.is_null()) {
      ghosted_point_to_blockMV = Xpetra::VectorFactory<LocalOrdinal, LocalOrdinal, GlobalOrdinal, Node>::Build(importer->getTargetMap());
      ghosted_point_to_blockMV->doImport(point_to_block_, *importer, Xpetra::INSERT);
      ghosted_point_to_block = ghosted_point_to_blockMV->getLocalViewDevice(Xpetra::Access::ReadOnly);
    } else
      ghosted_point_to_block = point_to_block;
  }

  KOKKOS_FORCEINLINE_FUNCTION
  void operator()(local_ordinal_type rlid) const {
    auto row            = A.rowConst(rlid);
    const size_t offset = A.graph.row_map(rlid);
    auto brlid          = row_translation(rlid);
    for (local_ordinal_type k = 0; k < row.length; ++k) {
      auto clid  = row.colidx(k);
      auto bclid = col_translation(clid);
      if (point_to_block(brlid, 0) == ghosted_point_to_block(bclid, 0)) {
        results(offset + k) = Kokkos::max(KEEP, results(offset + k));
      } else {
        results(offset + k) = Kokkos::max(DROP, results(offset + k));
      }
    }
  }
};

template <class local_matrix_type>
class DebugFunctor {
 private:
  using scalar_type        = typename local_matrix_type::value_type;
  using local_ordinal_type = typename local_matrix_type::ordinal_type;
  using memory_space       = typename local_matrix_type::memory_space;
  using results_view       = Kokkos::View<DecisionType*, memory_space>;

  using boundary_nodes_view = Kokkos::View<bool*, memory_space>;

  local_matrix_type A;
  results_view results;

 public:
  DebugFunctor(local_matrix_type& A_, results_view& results_)
    : A(A_)
    , results(results_) {}

  KOKKOS_FORCEINLINE_FUNCTION
  void operator()(local_ordinal_type rlid) const {
    auto row            = A.rowConst(rlid);
    const size_t offset = A.graph.row_map(rlid);
    for (local_ordinal_type k = 0; k < row.length; ++k) {
      if (results(offset + k) == UNDECIDED) {
        Kokkos::printf("No dropping decision was taken for entry (%d, %d)\n", rlid, row.colidx(k));
        assert(false);
      }
    }
  }
};

template <class local_matrix_type>
class SymmetrizeFunctor {
 private:
  using scalar_type        = typename local_matrix_type::value_type;
  using local_ordinal_type = typename local_matrix_type::ordinal_type;
  using memory_space       = typename local_matrix_type::memory_space;
  using results_view       = Kokkos::View<DecisionType*, memory_space>;

  local_matrix_type A;
  results_view results;

 public:
  SymmetrizeFunctor(local_matrix_type& A_, results_view& results_)
    : A(A_)
    , results(results_) {}

  KOKKOS_FORCEINLINE_FUNCTION
  void operator()(local_ordinal_type rlid) const {
    auto row            = A.rowConst(rlid);
    const size_t offset = A.graph.row_map(rlid);
    for (local_ordinal_type k = 0; k < row.length; ++k) {
      if (results(offset + k) == KEEP) {
        auto clid = row.colidx(k);
        if (clid >= A.numRows())
          continue;
        auto row2            = A.rowConst(clid);
        const size_t offset2 = A.graph.row_map(clid);
        for (local_ordinal_type k2 = 0; k2 < row2.length; ++k2) {
          auto clid2 = row2.colidx(k2);
          if (clid2 == rlid) {
            if (results(offset2 + k2) == DROP)
              results(offset2 + k2) = KEEP;
            break;
          }
        }
      }
    }
  }
};

template <class view_type, class comparator_type>
KOKKOS_INLINE_FUNCTION void serialHeapSort(view_type& v, comparator_type comparator) {
  auto N       = v.extent(0);
  size_t start = N / 2;
  size_t end   = N;
  while (end > 1) {
    if (start > 0)
      start = start - 1;
    else {
      end       = end - 1;
      auto temp = v(0);
      v(0)      = v(end);
      v(end)    = temp;
    }
    size_t root = start;
    while (2 * root + 1 < end) {
      size_t child = 2 * root + 1;
      if ((child + 1 < end) and (comparator(v(child), v(child + 1))))
        ++child;

      if (comparator(v(root), v(child))) {
        auto temp = v(root);
        v(root)   = v(child);
        v(child)  = temp;
        root      = child;
      } else
        break;
    }
  }
}

enum StrengthMeasure : int {
  /*
  \f[
  \frac{|A_{ij}|^2}{|A_{ii}| |A_{jj}|} \le \theta^2
  \f]
   */
  SmoothedAggregationMeasure = 0,
  /*
  \f[
  \frac{-\operatorname{Re}A_{ij}}{| max_j -A_{ij}|} \le \theta
  \f]
  */
  SignedRugeStuebenMeasure = 1,

  /*
  \f[
  \frac{-\operatorname{sign}(A_{ij}) |A_{ij}|^2}{|A_{ii}| |A_{jj}|} \le \theta^2
  \f]
  */
  SignedSmoothedAggregationMeasure = 2,

  /*
  \f[
  |A_{ij}| \le \theta
  \f]
   */
  UnscaledMeasure = 3
};

}  // namespace Misc

}  // namespace MueLu

#define MUELU_ETI_SLGN_SoC(CLASSNAME, SC, LO, GO, NO)                                      \
  template class CLASSNAME<SC, LO, GO, NO, MueLu::Misc::SmoothedAggregationMeasure>;       \
  template class CLASSNAME<SC, LO, GO, NO, MueLu::Misc::SignedRugeStuebenMeasure>;         \
  template class CLASSNAME<SC, LO, GO, NO, MueLu::Misc::SignedSmoothedAggregationMeasure>; \
  template class CLASSNAME<SC, LO, GO, NO, MueLu::Misc::UnscaledMeasure>;

#endif