MueLu_Aggregates_kokkos_def.hpp

Go to the documentation of this file.

// @HEADER
//
// ***********************************************************************
//
//        MueLu: A package for multigrid based preconditioning
//                  Copyright 2012 Sandia Corporation
//
// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
// the U.S. Government retains certain rights in this software.
//
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions are
// met:
//
// 1. Redistributions of source code must retain the above copyright
// notice, this list of conditions and the following disclaimer.
//
// 2. Redistributions in binary form must reproduce the above copyright
// notice, this list of conditions and the following disclaimer in the
// documentation and/or other materials provided with the distribution.
//
// 3. Neither the name of the Corporation nor the names of the
// contributors may be used to endorse or promote products derived from
// this software without specific prior written permission.
//
// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
//
// Questions? Contact
//                    Jonathan Hu       (jhu@sandia.gov)
//                    Andrey Prokopenko (aprokop@sandia.gov)
//                    Ray Tuminaro      (rstumin@sandia.gov)
//                    Tobias Wiesner    (tawiesn@sandia.gov)
//
// ***********************************************************************
//
// @HEADER
#ifndef MUELU_AGGREGATES_KOKKOS_DEF_HPP
#define MUELU_AGGREGATES_KOKKOS_DEF_HPP

#include <Xpetra_Map.hpp>
#include <Xpetra_Vector.hpp>
#include <Xpetra_MultiVectorFactory.hpp>
#include <Xpetra_VectorFactory.hpp>

#include "MueLu_LWGraph_kokkos.hpp"
#include "MueLu_Utilities_decl.hpp"
#include "MueLu_Aggregates_kokkos_decl.hpp"

namespace MueLu {

  template <class LocalOrdinal, class GlobalOrdinal, class DeviceType>
  Aggregates_kokkos<LocalOrdinal, GlobalOrdinal, Tpetra::KokkosCompat::KokkosDeviceWrapperNode<DeviceType> >::
  Aggregates_kokkos(LWGraph_kokkos graph) {
    numAggregates_  = 0;
    numGlobalAggregates_  = 0;

    vertex2AggId_ = LOMultiVectorFactory::Build(graph.GetImportMap(), 1);
    vertex2AggId_->putScalar(MUELU_UNAGGREGATED);

    procWinner_ = LOVectorFactory::Build(graph.GetImportMap());
    procWinner_->putScalar(MUELU_UNASSIGNED);

    isRoot_ = Kokkos::View<bool*, device_type>(Kokkos::ViewAllocateWithoutInitializing("roots"), graph.GetImportMap()->getLocalNumElements());
    Kokkos::deep_copy(isRoot_, false);

    // slow but safe, force TentativePFactory to build column map for P itself
    aggregatesIncludeGhosts_ = true;
  }

  template <class LocalOrdinal, class GlobalOrdinal, class DeviceType>
  Aggregates_kokkos<LocalOrdinal, GlobalOrdinal, Tpetra::KokkosCompat::KokkosDeviceWrapperNode<DeviceType> >::
  Aggregates_kokkos(const RCP<const Map>& map) {
    numAggregates_ = 0;
    numGlobalAggregates_  = 0;

    vertex2AggId_ = LOMultiVectorFactory::Build(map, 1);
    vertex2AggId_->putScalar(MUELU_UNAGGREGATED);

    procWinner_ = LOVectorFactory::Build(map);
    procWinner_->putScalar(MUELU_UNASSIGNED);

    isRoot_ = Kokkos::View<bool*,device_type>(Kokkos::ViewAllocateWithoutInitializing("roots"), map->getLocalNumElements());
    Kokkos::deep_copy(isRoot_, false);

    // slow but safe, force TentativePFactory to build column map for P itself
    aggregatesIncludeGhosts_ = true;
  }

  template <class LocalOrdinal, class GlobalOrdinal, class DeviceType>
  typename Aggregates_kokkos<LocalOrdinal, GlobalOrdinal, Tpetra::KokkosCompat::KokkosDeviceWrapperNode<DeviceType> >::aggregates_sizes_type::const_type
  Aggregates_kokkos<LocalOrdinal, GlobalOrdinal, Tpetra::KokkosCompat::KokkosDeviceWrapperNode<DeviceType> >::ComputeAggregateSizes(bool forceRecompute) const {
    if (aggregateSizes_.size() && !forceRecompute) {
      return aggregateSizes_;

    } else {
      // It is necessary to initialize this to 0
      aggregates_sizes_type aggregateSizes("aggregates", numAggregates_);

      int myPID = GetMap()->getComm()->getRank();

      auto vertex2AggId = vertex2AggId_->getDeviceLocalView(Xpetra::Access::ReadOnly);
      auto procWinner   = procWinner_  ->getDeviceLocalView(Xpetra::Access::ReadOnly);

      typename AppendTrait<decltype(aggregateSizes_), Kokkos::Atomic>::type aggregateSizesAtomic = aggregateSizes;
      Kokkos::parallel_for("MueLu:Aggregates:ComputeAggregateSizes:for", range_type(0,procWinner.size()),
        KOKKOS_LAMBDA(const LO i) {
          if (procWinner(i, 0) == myPID)
            aggregateSizesAtomic(vertex2AggId(i, 0))++;
        });

      aggregateSizes_ = aggregateSizes;

      return aggregateSizes;
    }

  }

  template <class LocalOrdinal, class GlobalOrdinal, class DeviceType>
  typename Aggregates_kokkos<LocalOrdinal, GlobalOrdinal, Tpetra::KokkosCompat::KokkosDeviceWrapperNode<DeviceType> >::local_graph_type
  Aggregates_kokkos<LocalOrdinal, GlobalOrdinal, Tpetra::KokkosCompat::KokkosDeviceWrapperNode<DeviceType> >::GetGraph() const {
    using row_map_type = typename local_graph_type::row_map_type;
    using entries_type = typename local_graph_type::entries_type;
    using size_type    = typename local_graph_type::size_type;

    auto numAggregates = numAggregates_;

    if (static_cast<LO>(graph_.numRows()) == numAggregates)
      return graph_;

    auto vertex2AggId = vertex2AggId_->getDeviceLocalView(Xpetra::Access::ReadOnly);
    auto procWinner   = procWinner_  ->getDeviceLocalView(Xpetra::Access::ReadOnly);
    auto sizes        = ComputeAggregateSizes();

    // FIXME_KOKKOS: replace by ViewAllocateWithoutInitializing + rows(0) = 0.
    typename row_map_type::non_const_type rows("Agg_rows", numAggregates+1);  // rows(0) = 0 automatically

    // parallel_scan (exclusive)
    Kokkos::parallel_scan("MueLu:Aggregates:GetGraph:compute_rows", range_type(0, numAggregates),
      KOKKOS_LAMBDA(const LO i, LO& update, const bool& final_pass) {
        update += sizes(i);
        if (final_pass)
          rows(i+1) = update;
      });

    decltype(rows) offsets(Kokkos::ViewAllocateWithoutInitializing("Agg_offsets"), numAggregates+1); // +1 is just for ease
    Kokkos::deep_copy(offsets, rows);

    int myPID = GetMap()->getComm()->getRank();

    size_type numNNZ;
    {
      Kokkos::View<size_type, device_type> numNNZ_device = Kokkos::subview(rows, numAggregates);
      typename Kokkos::View<size_type, device_type>::HostMirror numNNZ_host = Kokkos::create_mirror_view(numNNZ_device);
      Kokkos::deep_copy(numNNZ_host, numNNZ_device);
      numNNZ = numNNZ_host();
    }
    typename entries_type::non_const_type cols(Kokkos::ViewAllocateWithoutInitializing("Agg_cols"), numNNZ);
    size_t realnnz = 0;
    Kokkos::parallel_reduce("MueLu:Aggregates:GetGraph:compute_cols", range_type(0, procWinner.size()),
      KOKKOS_LAMBDA(const LO i, size_t& nnz) {
        if (procWinner(i, 0) == myPID) {
          typedef typename std::remove_reference< decltype( offsets(0) ) >::type atomic_incr_type;
          auto idx = Kokkos::atomic_fetch_add( &offsets(vertex2AggId(i,0)), atomic_incr_type(1));
          cols(idx) = i;
          nnz++;
        }
      }, realnnz);
    TEUCHOS_TEST_FOR_EXCEPTION(realnnz != numNNZ, Exceptions::RuntimeError,
                               "MueLu: Internal error: Something is wrong with aggregates graph construction: numNNZ = " << numNNZ << " != " << realnnz << " = realnnz");

    graph_ = local_graph_type(cols, rows);

    return graph_;
  }
  
  template <class LocalOrdinal, class GlobalOrdinal, class DeviceType>
  void 
  Aggregates_kokkos<LocalOrdinal, GlobalOrdinal, Tpetra::KokkosCompat::KokkosDeviceWrapperNode<DeviceType> >::ComputeNodesInAggregate(LO_view & aggPtr, LO_view & aggNodes, LO_view & unaggregated) const {
    LO numAggs  = GetNumAggregates();
    LO numNodes = vertex2AggId_->getLocalLength();
    auto vertex2AggId = vertex2AggId_->getDeviceLocalView(Xpetra::Access::ReadOnly);
    typename aggregates_sizes_type::const_type aggSizes = ComputeAggregateSizes(true);
    LO INVALID = Teuchos::OrdinalTraits<LO>::invalid();

    aggPtr = LO_view("aggPtr",numAggs+1);
    aggNodes = LO_view("aggNodes",numNodes);
    LO_view aggCurr("agg curr",numAggs+1);

    // Construct the "rowptr" and the counter
    Kokkos::parallel_scan("MueLu:Aggregates:ComputeNodesInAggregate:scan", range_type(0,numAggs+1),
      KOKKOS_LAMBDA(const LO aggIdx, LO& aggOffset, bool final_pass) {
        LO count = 0;
        if(aggIdx < numAggs)
          count = aggSizes(aggIdx);
        if(final_pass) {
          aggPtr(aggIdx) = aggOffset;
          aggCurr(aggIdx) = aggOffset;
          if(aggIdx==numAggs)
            aggCurr(numAggs) = 0; // use this for counting unaggregated nodes
        }
        aggOffset += count;
      });

    // Preallocate unaggregated to the correct size
    LO numUnaggregated = 0;
    Kokkos::parallel_reduce("MueLu:Aggregates:ComputeNodesInAggregate:unaggregatedSize", range_type(0,numNodes),
      KOKKOS_LAMBDA(const LO nodeIdx, LO & count) { 
        if(vertex2AggId(nodeIdx,0)==INVALID) 
          count++;
      }, numUnaggregated);
    unaggregated = LO_view("unaggregated",numUnaggregated);

    // Stick the nodes in each aggregate's spot
    Kokkos::parallel_for("MueLu:Aggregates:ComputeNodesInAggregate:for", range_type(0,numNodes),
      KOKKOS_LAMBDA(const LO nodeIdx) {
        LO aggIdx = vertex2AggId(nodeIdx,0);
        if(aggIdx != INVALID) {
          // atomic postincrement aggCurr(aggIdx) each time
          aggNodes(Kokkos::atomic_fetch_add(&aggCurr(aggIdx),1)) = nodeIdx;
        } else {
          // same, but using last entry of aggCurr for unaggregated nodes
          unaggregated(Kokkos::atomic_fetch_add(&aggCurr(numAggs),1)) = nodeIdx;
        }
      });
    
  }

  template <class LocalOrdinal, class GlobalOrdinal, class DeviceType>
  std::string Aggregates_kokkos<LocalOrdinal, GlobalOrdinal, Tpetra::KokkosCompat::KokkosDeviceWrapperNode<DeviceType> >::description() const {
    if (numGlobalAggregates_ == -1) return BaseClass::description() + "{nGlobalAggregates = not computed}"; 
    else                            return BaseClass::description() + "{nGlobalAggregates = " + toString(numGlobalAggregates_) + "}";
  }

  template <class LocalOrdinal, class GlobalOrdinal, class DeviceType>
  void Aggregates_kokkos<LocalOrdinal, GlobalOrdinal, Tpetra::KokkosCompat::KokkosDeviceWrapperNode<DeviceType> >::print(Teuchos::FancyOStream& out, const Teuchos::EVerbosityLevel verbLevel) const {
    MUELU_DESCRIBE;

    if (verbLevel & Statistics1) {
      if (numGlobalAggregates_ == -1) out0 << "Global number of aggregates: not computed " << std::endl;
      else                            out0 << "Global number of aggregates: " << numGlobalAggregates_ << std::endl;
    }
  }

  template <class LocalOrdinal, class GlobalOrdinal, class DeviceType>
  GlobalOrdinal Aggregates_kokkos<LocalOrdinal, GlobalOrdinal, Tpetra::KokkosCompat::KokkosDeviceWrapperNode<DeviceType> >::GetNumGlobalAggregatesComputeIfNeeded() {

    if (numGlobalAggregates_ != -1) {
      LO nAggregates = GetNumAggregates();
      GO nGlobalAggregates;
      MueLu_sumAll(vertex2AggId_->getMap()->getComm(), (GO)nAggregates, nGlobalAggregates);
      SetNumGlobalAggregates(nGlobalAggregates);
    }
    return numGlobalAggregates_;
  }

  template <class LocalOrdinal, class GlobalOrdinal, class DeviceType>
  const RCP<const Xpetra::Map<LocalOrdinal,GlobalOrdinal, Tpetra::KokkosCompat::KokkosDeviceWrapperNode<DeviceType>> >
  Aggregates_kokkos<LocalOrdinal, GlobalOrdinal, Tpetra::KokkosCompat::KokkosDeviceWrapperNode<DeviceType>>::GetMap() const {
    return vertex2AggId_->getMap();
  }

} //namespace MueLu

#endif // MUELU_AGGREGATES_KOKKOS_DEF_HPP