// ============================================================================
// ==                                                                        ==
// == AMDiS - Adaptive multidimensional simulations                          ==
// ==                                                                        ==
// ==  http://www.amdis-fem.org                                              ==
// ==                                                                        ==
// ============================================================================
//
// Software License for AMDiS
//
// Copyright (c) 2010 Dresden University of Technology 
// All rights reserved.
// Authors: Simon Vey, Thomas Witkowski et al.
//
// This file is part of AMDiS
//
// See also license.opensource.txt in the distribution.



/** \file MeshDistributor.h */

#ifndef AMDIS_MESHDISTRIBUTOR_H
#define AMDIS_MESHDISTRIBUTOR_H


#include <mpi.h>
#include "parallel/DofComm.h"
#include "parallel/ElementObjectDatabase.h"
#include "parallel/ParallelTypes.h"
#include "parallel/MeshLevelData.h"
#include "parallel/MeshPartitioner.h"
#include "parallel/InteriorBoundary.h"
#include "parallel/ParallelDofMapping.h"
#include "parallel/PeriodicMap.h"
#include "parallel/StdMpi.h"
#include "AMDiS_fwd.h"
#include "Containers.h"
#include "Global.h"
#include "ProblemTimeInterface.h"
#include "ProblemIterationInterface.h"
#include "FiniteElemSpace.h"
#include "Serializer.h"
#include "BoundaryManager.h"
#include "SystemVector.h"

namespace AMDiS {

  using namespace std;


  struct BoundaryDofInfo
  {
    map<GeoIndex, DofContainerSet> geoDofs;
  };


  class MeshDistributor
  {
  private:
    MeshDistributor();
	          
  public:
    ~MeshDistributor();

    void initParallelization();

    void exitParallelization();

    /// Adds a DOFVector to the set of \ref interchangeVecs. Thus, this vector 
    /// will be automatically interchanged between ranks when mesh is 
    /// repartitioned.
    void addInterchangeVector(DOFVector<double> *vec)
    {
      interchangeVectors.push_back(vec);
    }

    /// Adds all DOFVectors of a SystemVector to \ref interchangeVecs.
    void addInterchangeVector(SystemVector *vec)
    {
      for (int i = 0; i < vec->getSize(); i++)
	interchangeVectors.push_back(vec->getDOFVector(i));
    }
    
    /** \brief
     * This function checks if the mesh has changed on at least on rank. In 
     * this case, the interior boundaries are adapted on all ranks such that 
     * they fit together on all ranks. Furthermore the function 
     * \ref updateLocalGlobalNumbering() is called to update the DOF numberings 
     * and mappings on all rank due to the new mesh structure.
     *
     * \param[in]  tryRepartition   If this parameter is true, repartitioning 
     *                              may be done. This depends on several other 
     *                              parameters. If the parameter is false, the 
     *                              mesh is only checked and adapted but never 
     *                              repartitioned.
     */
    void checkMeshChange(bool tryRepartition = true);

    /// Checks if is required to repartition the mesh. If this is the case, a new
    /// partition will be created and the mesh will be redistributed between the
    /// ranks.
    void repartitionMesh();
    
    void getImbalanceFactor(double &imbalance, 
			    int &minDofs, 
			    int &maxDofs,
			    int &sumDofs);

    double getImbalanceFactor();

    /// Calculates the imbalancing factor and prints it to screen.
    void printImbalanceFactor();

    /// Test, if the mesh consists of macro elements only. The mesh partitioning 
    /// of the parallelization works for macro meshes only and would fail, if the 
    /// mesh is already refined in some way. Therefore, this function will exit
    /// the program if it finds a non macro element in the mesh.
    void testForMacroMesh();

    /// Set for each element on the partitioning level the number of 
    /// leaf elements.
    void setInitialElementWeights();

    inline string getName() 
    { 
      return name; 
    }

    inline Mesh* getMesh()
    {
      return mesh;
    }

    /// Returns an FE space from \ref feSpaces.
    inline const FiniteElemSpace* getFeSpace(unsigned int i = 0)
    {
      FUNCNAME("MeshDistributor::getFeSpace()");

      TEST_EXIT_DBG(i < feSpaces.size())("Should not happen!\n");

      return feSpaces[i];
    }

    /// Returns all FE spaces, thus \ref feSpaces.
    inline vector<const FiniteElemSpace*>& getFeSpaces()
    {
      return feSpaces;
    }

    /// Returns the DOF mapping object, \ref dofMap.
    inline ParallelDofMapping& getDofMap()
    {
      return dofMap;
    }

    inline ParallelDofMapping& getDofMapSd()
    {
      return dofMapSd;
    }

    /// Returns the periodic mapping handler, \ref periodicMap.
    inline PeriodicMap& getPeriodicMap()
    {
      return periodicMap;
    }

    DofComm& getDofComm()
    {
      return dofComm;
    }

    DofComm& getDofCommSd()
    {
      return dofCommSd;
    }

    InteriorBoundary& getIntBoundary()
    {
      return intBoundary;
    }

    InteriorBoundary& getIntBoundarySd()
    {
      return intBoundarySd;
    }

    inline long getLastMeshChangeIndex()
    {
      return lastMeshChangeIndex;
    }

    inline int getMpiRank()
    {
      return mpiRank;
    }

    inline int getMpiSize()
    {
      return mpiSize;
    }

    inline MPI::Intracomm& getMpiComm()
    {
      return mpiComm;
    }

    inline bool isInitialized()
    {
      return initialized;
    }

    // Writes all data of this object to an output stream.
    void serialize(ostream &out);

    // Reads the object data from an input stream.
    void deserialize(istream &in);

    /** \brief
     * This function must be used if the values of a DOFVector must be 
     * synchronised over all ranks. That means, that each rank sends the 
     * values of the DOFs, which are owned by the rank and lie on an interior 
     * bounday, to all other ranks also having these DOFs.
     *
     * This function must be used, for example, after the lineary system is 
     * solved, or after the DOFVector is set by some user defined functions, 
     * e.g., initial solution functions.
     */    
    template<typename T>
    void synchVector(DOFVector<T> &vec) 
    {
      StdMpi<vector<T> > stdMpi(mpiComm);

      const FiniteElemSpace *fe = vec.getFeSpace();

      for (DofComm::Iterator it(dofComm.getSendDofs(), fe); 
	   !it.end(); it.nextRank()) {
	vector<T> dofs;
	dofs.reserve(it.getDofs().size());
	
	for (; !it.endDofIter(); it.nextDof())
	  dofs.push_back(vec[it.getDofIndex()]);
	
	stdMpi.send(it.getRank(), dofs);
      }
	     
      for (DofComm::Iterator it(dofComm.getRecvDofs()); 
	   !it.end(); it.nextRank())
        stdMpi.recv(it.getRank());
	     
      stdMpi.startCommunication();

      for (DofComm::Iterator it(dofComm.getRecvDofs(), fe); 
	   !it.end(); it.nextRank())
	for (; !it.endDofIter(); it.nextDof())
	  vec[it.getDofIndex()] = 
	     stdMpi.getRecvData(it.getRank())[it.getDofCounter()];
    }
   
    /// Works in the same way as the function above defined for DOFVectors. Due
    /// to performance, this function does not call \ref synchVector for each 
    /// DOFVector, but instead sends all values of all DOFVectors all at once.
    void synchVector(SystemVector &vec, int level = 0);

    /// Works quite similar to the function \ref synchVector, but instead the 
    /// values of subdomain vectors are add along the boundaries.
    template<typename T>
    void synchAddVector(DOFVector<T> &vec)
    {
      StdMpi<vector<T> > stdMpi(mpiComm);

      const FiniteElemSpace *fe = vec.getFeSpace();

      for (DofComm::Iterator it(dofComm.getRecvDofs(), fe);
	   !it.end(); it.nextRank()) {
	vector<T> dofs;
	dofs.reserve(it.getDofs().size());
	
	for (; !it.endDofIter(); it.nextDof())
	  dofs.push_back(vec[it.getDofIndex()]);
	
	stdMpi.send(it.getRank(), dofs);
      }

      for (DofComm::Iterator it(dofComm.getSendDofs()); 
	   !it.end(); it.nextRank())
        stdMpi.recv(it.getRank());
	     
      stdMpi.startCommunication();

      for (DofComm::Iterator it(dofComm.getSendDofs(), fe); 
	   !it.end(); it.nextRank())
	for (; !it.endDofIter(); it.nextDof())
	  vec[it.getDofIndex()] += 
	     stdMpi.getRecvData(it.getRank())[it.getDofCounter()];

      synchVector(vec);
    }

    /// In 3D, a subdomain may not be a valid AMDiS mesh if it contains two
    /// parts which are only connected by an edge. In this case, the standard
    /// refinement algorithm does not work correctly, as two elements connected
    /// only on one edge are not neighours by definition. This functions checks
    /// for this situation and fix the problem. For this, the mesh is search for
    /// all edges connecting two elements that are otherwise not connected.
    void check3dValidMesh();

    void setBoundaryDofRequirement(Flag flag)
    {
      createBoundaryDofFlag = flag;
    }

    BoundaryDofInfo& getBoundaryDofInfo(const FiniteElemSpace *feSpace,
					int level)
    {
      FUNCNAME("MeshDistributor::getBoundaryDofInfo()");

      TEST_EXIT_DBG(level < static_cast<int>(boundaryDofInfo.size()))
	("Wrong level number: %d, whereas array size is %d!\n", 
	 level, boundaryDofInfo.size());

      return boundaryDofInfo[level][feSpace];
    }

    void getAllBoundaryDofs(const FiniteElemSpace *feSpace, 
			    int level,
			    DofContainer& dofs);

    const ElementObjectDatabase& getElementObjectDb() 
    {
      return elObjDb;
    }

    /// Adds a stationary problem to the global mesh distributor objects.
    static void addProblemStatGlobal(ProblemStatSeq *probStat);

    MeshLevelData& getMeshLevelData() 
    {
      return levelData;
    }

    void updateLocalGlobalNumbering();

    /// Updates the local and global DOF numbering after the mesh has been 
    /// changed.
    void updateLocalGlobalNumbering(ParallelDofMapping &dmap,
				    DofComm &dcom,
				    const FiniteElemSpace *feSpace);

  protected:
    void addProblemStat(ProblemStatSeq *probStat);

    /// Determines the interior boundaries, i.e. boundaries between ranks, and
    /// stores all information about them in \ref interiorBoundary.
    void createInteriorBoundary(bool firstCall);

    void createBoundaryDofs();

    /// Removes all macro elements from the mesh that are not part of ranks 
    /// partition.
    void removeMacroElements();

    /// Calls \ref createPeriodicMap(feSpace) for all FE spaces that are
    /// handled by the mesh distributor.
    void createPeriodicMap();

    /// Creates, for a specific FE space, to all DOFs in rank's partition that 
    /// are on a periodic boundary the mapping from dof index to the other 
    /// periodic dof indices. This information is stored in \ref periodicDofMap.  
    void createPeriodicMap(const FiniteElemSpace *feSpace);

    /// This function is called only once during the initialization when the
    /// whole macro mesh is available on all cores. It copies the pointers of all
    /// macro elements to \ref allMacroElements and stores all neighbour 
    /// information based on macro element indices (and not pointer based) in 
    /// \ref macroElementNeighbours. These information are then used to 
    /// reconstruct macro elements during mesh redistribution.
    void createMacroElementInfo();

    void updateMacroElementInfo();

    /** \brief
     * Checks for all given interior boundaries if the elements fit together on
     * both sides of the boundaries. If this is not the case, the mesh is 
     * adapted. Because refinement of a certain element may forces the 
     * refinement of other elements, it is not guaranteed that all rank's meshes
     * fit together after this function terminates. Hence, it must be called 
     * until a stable mesh refinement is reached.
     *
     * \param[in] allBound   Defines a map from rank to interior boundaries 
     *                       which should be checked.
     *
     * \return    If the mesh has  been changed by this function, it returns 
     *            true. Otherwise, it returns false, i.e., the given interior 
     *            boundaries fit together on both sides.
     */
    bool checkAndAdaptBoundary(RankToBoundMap &allBound);
  
    /// Sets \ref isRankDof to all matrices and rhs vectors in a given 
    /// stationary problem.
    void setRankDofs(ProblemStatSeq *probStat);

    /// Sets \ref isRankDof to all matrices and rhs vectors in all 
    /// stationary problems.
    void setRankDofs();

    /// Removes all periodic boundary condition information from all matrices and
    /// vectors of all stationary problems and from the mesh itself.
    void removePeriodicBoundaryConditions();

    /// Removes all periodic boundary condition information from all matrices and
    /// vector of a given stationary problem.
    void removePeriodicBoundaryConditions(ProblemStatSeq *probStat);

    // Removes all periodic boundaries from a given boundary map.
    void removePeriodicBoundaryConditions(BoundaryIndexMap& boundaryMap);

    void createMeshLevelStructure();

    /// Writes a vector of dof pointers to an output stream.
    void serialize(ostream &out, DofContainer &data);

    /// Writes a \ref RankToDofContainer to an output stream.
    void serialize(ostream &out, 
		   map<int, map<const FiniteElemSpace*, DofContainer> > &data);

    /// Reads a vector of dof pointers from an input stream.
    void deserialize(istream &in, DofContainer &data,
		     map<int, const DegreeOfFreedom*> &dofIndexMap);

    /// Reads a \ref RankToDofContainer from an input stream.
    void deserialize(istream &in, 
		     map<int, map<const FiniteElemSpace*, DofContainer> > &data,
		     map<const FiniteElemSpace*, map<int, const DegreeOfFreedom*> > &dofIndexMap);

    /// Writes a mapping from dof pointers to some values to an output stream.
    template<typename T>
    void serialize(ostream &out, map<const DegreeOfFreedom*, T> &data)
    {
      FUNCNAME("ParallelDomainBase::serialize()");

      int mapSize = data.size();
      SerUtil::serialize(out, mapSize);
      for (typename map<const DegreeOfFreedom*, T>::iterator it = data.begin();
	   it != data.end(); ++it) {
	int v1 = (*(it->first));
	T v2 = it->second;
	SerUtil::serialize(out, v1);
	SerUtil::serialize(out, v2);
      }
    }

    /// Reads a mapping from dof pointer to some values from an input stream.
    template<typename T>
    void deserialize(istream &in, map<const DegreeOfFreedom*, T> &data,
		     map<int, const DegreeOfFreedom*> &dofIndexMap)
    {
      FUNCNAME("ParallelDomainBase::deserialize()");

      int mapSize = 0;
      SerUtil::deserialize(in, mapSize);
      for (int i = 0; i < mapSize; i++) {
	int v1 = 0;
	T v2;
	SerUtil::deserialize(in, v1);
	SerUtil::deserialize(in, v2);

	TEST_EXIT_DBG(dofIndexMap.count(v1) != 0)
	  ("Cannot find DOF %d in map!\n", v1);

	data[dofIndexMap[v1]] = v2;
      }
    }

  protected:
    /// List of all stationary problems that are managed by this mesh 
    /// distributor.
    vector<ProblemStatSeq*> problemStat;

    /// If true, the mesh distributor is already initialized;
    bool initialized;

    /// The rank of the current process.
    int mpiRank;

    /// Overall number of processes.
    int mpiSize;

    /// MPI communicator collected all processes, which should be used for
    /// calculation. The Debug procces is not included in this communicator.
    MPI::Intracomm mpiComm;

    /// Name of the problem (as used in the init files)
    string name;

    /// Finite element spaces of the problem.
    vector<const FiniteElemSpace*> feSpaces;

    /// Mesh of the problem.
    Mesh *mesh;

    /// A refinement manager that should be used on the mesh. It is used to 
    /// refine elements at interior boundaries in order to fit together with 
    /// elements on the other side of the interior boundary.    
    RefinementManager *refineManager;

    /// Info level.
    int info;

    /// Pointer to a mesh partitioner that is used to partition the mesh to 
    /// the ranks.
    MeshPartitioner *partitioner;

    /// Weights for the elements, i.e., the number of leaf elements within 
    /// this element.
    map<int, double> elemWeights;

    /// Stores to every macro element index the number of the rank that owns this
    /// macro element.
    map<int, int> partitionMap;

    /// Mapping object to map from local DOF indices to global ones.
    ParallelDofMapping dofMap;

    ParallelDofMapping dofMapSd;

    /// Database to store and query all sub-objects of all elements of the 
    /// macro mesh.
    ElementObjectDatabase elObjDb;

    /// Defines the interior boundaries of the domain that result from 
    /// partitioning the whole mesh. 
    InteriorBoundary intBoundary;

    InteriorBoundary intBoundarySd;

    DofComm dofComm;

    DofComm dofCommSd;

    PeriodicMap periodicMap;

    /// This set of values must be interchanged between ranks when the mesh is 
    /// repartitioned.
    vector<DOFVector<double>*> interchangeVectors;
		        
    /// If the problem definition has been read from a serialization file, this 
    /// variable is true, otherwise it is false. This variable is used to stop the
    /// initialization function, if the problem definition has already been read
    /// from a serialization file.
    bool deserialized;

    /// Denotes whether there exists a filewriter for this object.
    bool writeSerializationFile;

    /// If true, it is possible to repartition the mesh during computations.
    bool repartitioningAllowed;

    /// Stores the number of mesh changes that must lie in between to 
    /// repartitionings.
    int repartitionIthChange;

    /// Counts the number of mesh changes after the last mesh repartitioning 
    /// was done.
    int nMeshChangesAfterLastRepartitioning;

    /// Countes the number of mesh repartitions that were done. Till now, this 
    /// variable is used only for debug outputs.
    int repartitioningCounter;

    /// If repartitioning of the mesh fail, this variable has a positive value
    /// that gives the number of mesh changes the mesh distributer will wait
    /// before trying new mesh repartitioning.
    int repartitioningFailed;

    /// Directory name where all debug output files should be written to.
    string debugOutputDir;

    /// Stores the mesh change index. This is used to recognize changes in the
    /// mesh structure (e.g. through refinement or coarsening managers).
    long lastMeshChangeIndex;

    /// Stores for all macro elements of the original macro mesh the
    /// neighbourhood information based on element indices. Thus, each macro
    /// element index is mapped to a vector containing all indices of 
    /// neighbouring macro elements.
    map<int, vector<int> > macroElementNeighbours;

    /// Store all macro elements of the overall mesh, i.e., before the
    /// mesh is redistributed for the first time.
    vector<MacroElement*> allMacroElements;

    Flag createBoundaryDofFlag;

    /// Stores on each mesh level for all FE spaces the information about
    /// all boundary DOFs.
    vector<map<const FiniteElemSpace*, BoundaryDofInfo> > boundaryDofInfo;

    MeshLevelData levelData;

    /// If there is no mesh adaptivity, the mesh distributor can remove some
    /// data structures which are only used if mesh changes or it must be
    /// redistributed due to some local adaptivity. By default, this variable
    /// is set to true, and thus no special assumption are made.
    bool meshAdaptivity;

  public:
    static bool sebastianMode;

    /// The boundary DOFs are sorted by subobject entities, i.e., first all
    /// face DOFs, edge DOFs and to the last vertex DOFs will be set to
    /// communication structure vectors, \ref sendDofs and \ref recvDofs.
    static const Flag BOUNDARY_SUBOBJ_SORTED;

    /// When boundary DOFs are created, \ref boundaryDofInfo is filled for
    /// all DOFs that this rank will send to other ranks (thus, rank 
    /// owned DOFs.
    static const Flag BOUNDARY_FILL_INFO_SEND_DOFS;

    /// When boundary DOFs are created, \ref boundaryDofInfo is filled for
    /// all DOFs that this rank will receive from other ranks (thus, DOFs
    /// that are owned by another rank).
    static const Flag BOUNDARY_FILL_INFO_RECV_DOFS;

    static MeshDistributor *globalMeshDistributor;

    friend class ParallelDebug;
  };
}

#endif // AMDIS_MESHDISTRIBUTOR_H