---

LevelDataI.H

Go to the documentation of this file.

/* _______              __
  / ___/ /  ___  __ _  / /  ___
 / /__/ _ \/ _ \/  ' \/ _ \/ _ \
 \___/_//_/\___/_/_/_/_.__/\___/ 
*/
//
// This software is copyright (C) by the Lawrence Berkeley
// National Laboratory.  Permission is granted to reproduce
// this software for non-commercial purposes provided that
// this notice is left intact.
// 
// It is acknowledged that the U.S. Government has rights to
// this software under Contract DE-AC03-765F00098 between
// the U.S.  Department of Energy and the University of
// California.
//
// This software is provided as a professional and academic
// contribution for joint exchange. Thus it is experimental,
// is provided ``as is'', with no warranties of any kind
// whatsoever, no support, no promise of updates, or printed
// documentation. By using this software, you acknowledge
// that the Lawrence Berkeley National Laboratory and
// Regents of the University of California shall have no
// liability with respect to the infringement of other
// copyrights by any part of this software.
//

#include <cstdlib>
#include <algorithm>
#include "parstream.H"
#include "memtrack.H"

using std::sort;

template<class T> inline
LevelData<T>::LevelData()
  :m_sendbuffer(NULL), m_sendcapacity(0),
   m_recbuffer(NULL), m_reccapacity(0)
{
#ifdef MPI
  numSends = 0;
  numReceives = 0;
#endif
}

template<class T> inline
LevelData<T>::LevelData(const DisjointBoxLayout& dp, int comps, const IntVect& ghost,
                        const DataFactory<T>& a_factory)
  : m_disjointBoxLayout(dp), m_ghost(ghost), m_sendbuffer(NULL),
    m_sendcapacity(0),  m_recbuffer(NULL),m_reccapacity(0)
{ 
#ifdef MPI
  numSends = 0;
  numReceives = 0;
#endif
  m_boxLayout = dp; 
  m_comps = comps;
  m_isdefined = true;

  if(!dp.isClosed())
    {
      MayDay::Error("non-disjoint DisjointBoxLayout: LevelData<T>::LevelData(const DisjointBoxLayout& dp, int comps)");
    }

  Interval interval(0, comps-1);
  allocateGhostVector(a_factory, ghost);
  setVector(*this, interval, interval);
  
 
}



// Since I need to thwart the user from invoking the
// 'define' methods that use a general BoxLayout, I cannot
// get at said functions myself now. Ha!  So, I have to recode
// them here.

template<class T> inline
void LevelData<T>::define(const DisjointBoxLayout& dp, int comps, const IntVect& ghost,
                          const DataFactory<T> & a_factory)
{
  m_isdefined = true;
  if(!dp.isClosed())
    {
      MayDay::Error("non-disjoint DisjointBoxLayout: LevelData<T>::define(const DisjointBoxLayout& dp,....)");
    }
  if(comps<=0)
    {
      MayDay::Error("LevelData::LevelData(const BoxLayout& dp, int comps)  comps<=0");
    }
  m_comps = comps;
  m_boxLayout = dp;

  m_disjointBoxLayout = dp;
  m_ghost = ghost;

  Interval interval(0, comps-1);
  allocateGhostVector(a_factory, ghost);
  setVector(*this, interval, interval);

#ifdef MPI  
  m_fromMe.resize(0);
  m_toMe.resize(0);
#endif

}


template<class T> inline
void LevelData<T>::define(const LevelData<T>& da,  const DataFactory<T> & a_factory)
{
  m_isdefined = true;
  if(this == &da) return;
  m_disjointBoxLayout = da.m_disjointBoxLayout;
  m_boxLayout  = da.m_disjointBoxLayout;
  m_comps     = da.m_comps;
  m_ghost     = da.m_ghost;

  Interval srcAnddest(0, m_comps-1);

  allocateGhostVector(a_factory, m_ghost);
  setVector(da, srcAnddest, srcAnddest);

#ifdef MPI
  m_fromMe.resize(0);
  m_toMe.resize(0);
#endif
}


template<class T> inline
void LevelData<T>::define(const LevelData<T>& da, const Interval& comps,
                          const DataFactory<T>& a_factory)
{
  m_isdefined = true;
  if(this == &da){
    MayDay::Error(" LevelData<T>::define(const LevelData<T>& da, const Interval& comps) called with 'this'");
  }
  assert(comps.size()>0);
  // this line doesn't make any sense!
  //assert(comps.end()<=m_comps);
  assert(comps.begin()>=0);

  m_disjointBoxLayout = da.m_disjointBoxLayout;
  m_boxLayout  = da.m_disjointBoxLayout;

  m_comps = comps.size();

  m_ghost = da.m_ghost;

  Interval dest(0, m_comps-1);

  allocateGhostVector(a_factory, m_ghost);

  setVector(da, comps, dest);

#ifdef MPI
  m_fromMe.resize(0);
  m_toMe.resize(0);
#endif
}

template<class T> inline
void LevelData<T>::copyTo(const Interval& srcComps, 
                          BoxLayoutData<T>& dest,
                          const Interval& destComps) const
{
  if((BoxLayoutData<T>*)this == &dest) return;

  if(boxLayout() == dest.boxLayout())
    {
      // parallel direct copy here, no communication issues
      for(DataIterator it(dataIterator()); it.ok(); ++it)
        {
          dest[it()].copy(box(it()), 
                          destComps,
                          box(it()),
                          this->operator[](it()),
                          srcComps);
        }
      return;
    }

  Copier copier(m_disjointBoxLayout, dest.boxLayout());
  copyTo(srcComps, dest, destComps, copier);
}

template<class T> inline
void LevelData<T>::copyTo(const Interval& srcComps, 
                          LevelData<T>& dest,
                          const Interval& destComps) const
{
  if(this == &dest){
    MayDay::Error("src == dest in copyTo function. Perhaps you want exchange ?");
  }

  if(boxLayout() == dest.boxLayout()  && dest.ghostVect() == IntVect::Zero)
    {
      // parallel direct copy here, no communication issues
      for(DataIterator it(dataIterator()); it.ok(); ++it)
        {
          dest[it()].copy(box(it()), 
                          destComps,
                          box(it()),
                          this->operator[](it()),
                          srcComps);
        }
      return;
    }

  Copier copier(m_disjointBoxLayout, dest.getBoxes(), dest.m_ghost);
  copyTo(srcComps, dest, destComps, copier);
}


template<class T> inline
void LevelData<T>::copyTo(const Interval& srcComps, 
                          BoxLayoutData<T>& dest,
                          const Interval& destComps,
                          const Copier& copier) const
{

  makeItSo(srcComps, *this, dest, destComps, copier);

}

template<class T> inline
void LevelData<T>::copyTo(const Interval& srcComps, 
                          LevelData<T>& dest,
                          const Interval& destComps,
                          const Copier& copier) const
{

  makeItSo(srcComps, *this, dest, destComps, copier);

}

template<class T> inline 
void LevelData<T>::exchange(const Interval& comps)
{
  // later on we can code this part as a direct algorithm
  // by copying and pasting the code from the Copier::define code
  // for now, just do the easy to debug approach.
  Copier copier(m_disjointBoxLayout, m_disjointBoxLayout, m_ghost);
  exchange(comps, copier);

  // if there aren't any ghost cells, there isn't really anything
  // to do here (also, if m_ghost == Zero, m_exchangeCopier
  // wasn't defined!
  //if (m_ghost != IntVect::Zero)
  //makeItSo(comps, *this, *this, comps, m_exchangeCopier);
}

template<class T> inline
void LevelData<T>::exchange(const Interval& comps,
                            const Copier& copier)
{
  makeItSo(comps, *this, *this, comps, copier);

}

template<class T> inline
void LevelData<T>::makeItSo(const Interval&   a_srcComps, 
                            const LevelData<T>& a_src,
                            BoxLayoutData<T>& a_dest,
                            const Interval&   a_destComps,
                            const Copier&     a_copier) const
{
  // The following five functions are nullOps in uniprocessor mode

  // Instead of doing this here, do it an end of makeItSo (ndk)
  //completePendingSends(); // wait for sends from possible previous operation

  // new evil logic to determine under what conditions we can just
  // re-use our messaging pattern and buffers from the last call to
  // makeItSo.  pretty elaborate, I know.  bvs
#ifdef MPI
  static Copier* lastCopier=NULL;
  
    if(T::preAllocatable() == 2 || !a_copier.bufferAllocated() || 
         (m_fromMe.size() + m_toMe.size() == 0) ||lastCopier != &a_copier){
        allocateBuffers(a_src,  a_srcComps,
                                        a_dest, a_destComps,
                                        a_copier);  //monkey with buffers, set up 'fromMe' and 'toMe' queues
                a_copier.setBufferAllocated(true);
          }
        lastCopier = (Copier*)(&a_copier);
#endif

  writeSendDataFromMeIntoBuffers(a_src, a_srcComps);


  // If there is nothing to recv/send, don't go into these functions 
  // and allocate memory that will not be freed later.  (ndk)
  // The #ifdef MPI is for the m_toMe and m_fromMe
#ifdef MPI
  if (m_toMe.size() > 0) {
    postReceivesToMe(); // all non-blocking
  }

  if (m_fromMe.size() > 0) {
    postSendsFromMe();  // all non-blocking
  }
#endif

  //  computation that could occur during communication should really
  //  go here somehow.  while all the non-blocking sending and receiving is
  //  going on.  
  //
  //  my thought is to return from this function at this point an object
  //  that encapsulates the argument list above.  
  //  a "ChomboMessaging" object.
  //  The user can keep a reference
  //  to this object and do computations.  When they reach the limit of what
  //  they can compute without this communication completing, they call the
  //  "finalize()" function of their ChomboMessaging object and the rest of this
  //  code below gets executed.
  //  a real question though is: is there really enough computation to do while
  //  messaging is going on to justify the effort, and what machines really have
  //  good asynchronous messaging to make the work worthwhile.
  //  
  //  the other approach is to more finely decompose the overlapping of
  //  messaging and computation by using the ChomboMessaging object in the
  //  DataIterator construction.  The DataIterator returns T objects as they
  //  are completed from messaging.  This preserves almost all of the Chombo
  //  code as is but would be mucho tricky to actually implement and might only
  //  gain little.  This would not be a thing to try unitl Chombo is
  //  heavily instrumented for performance measuring.  in this design, unpackRecievesToMe()
  //  would become a complicated process interwoven with a DataIterator.

  //  postReceivesToMe();

  // perform local copy
  for(CopyIterator it(a_copier, CopyIterator::LOCAL); it.ok(); ++it)
    {
      const MotionItem& item = it();
      a_dest[item.toIndex].copy(item.fromRegion, 
                                a_destComps,
                                item.toRegion,
                                a_src[item.fromIndex],
                                a_srcComps);
    }

  // Uncomment and Move this out of unpackReceivesToMe()  (ndk)
  completePendingSends(); // wait for sends from possible previous operation

  unpackReceivesToMe(a_dest, a_destComps); // nullOp in uniprocessor mode

}

template<class T> inline
void LevelData<T>::define(const BoxLayout& dp, int comps,  const DataFactory<T>& a_factory)
{
  MayDay::Error("LevelData<T>::define called with BoxLayout input");
}

template<class T> inline
void LevelData<T>::define(const BoxLayout& dp)
{
  MayDay::Error("LevelData<T>::define called with BoxLayout input");
}

template<class T> inline
void LevelData<T>::define(const BoxLayoutData<T>& da, const DataFactory<T>& a_factory )
{
  MayDay::Error("LevelData<T>::define called with BoxLayout input");
}

template<class T> inline
void LevelData<T>::define(const BoxLayoutData<T>& da, const Interval& comps,
                          const DataFactory<T>& a_factory)
{
  MayDay::Error("LevelData<T>::define called with BoxLayout input");
}

template<class T> inline
LevelData<T>::~LevelData()
{
  completePendingSends();
  free(m_sendbuffer);
  free(m_recbuffer);
}

#ifndef MPI
// uniprocessor version of all these nullop functions.
template<class T> inline
void LevelData<T>::completePendingSends() const
{;}

template<class T> inline
void LevelData<T>::allocateBuffers(const LevelData<T>& a_src, 
                                   const Interval& a_srcComps,
                                   const BoxLayoutData<T>& a_dest,
                                   const Interval& a_destComps,
                                   const Copier&   a_copier) const
{;}  

template<class T> inline
void LevelData<T>::writeSendDataFromMeIntoBuffers(const LevelData<T>& a_src, 
                                                  const Interval&     a_srcComps) const
{;}

template<class T> inline
void LevelData<T>::postSendsFromMe() const
{;}

template<class T> inline
void LevelData<T>::postReceivesToMe() const
{;}

template<class T> inline
void LevelData<T>::unpackReceivesToMe(BoxLayoutData<T>& a_dest, 
                                      const Interval&   a_destComps) const
{;}

#else

// MPI versions of the above codes.

template<class T> inline
void LevelData<T>::completePendingSends() const
{
  if(numSends > 0){
    int result = MPI_Waitall(numSends, m_sendRequests, m_sendStatus);
    if(result != MPI_SUCCESS)
      {
        //hell if I know what to do about failed messaging here
      }

    delete[] m_sendRequests;
    delete[] m_sendStatus;
  }
  numSends = 0;
}

template<class T> inline
void LevelData<T>::allocateBuffers(const LevelData<T>& a_src, 
                                   const Interval& a_srcComps,
                                   const BoxLayoutData<T>& a_dest,
                                   const Interval& a_destComps,
                                   const Copier&   a_copier) const
{
  m_fromMe.resize(0);
  m_toMe.resize(0);
  size_t sendBufferSize = 0;
  size_t recBufferSize  = 0;
  // two versions of code here.  one for preAllocatable T, one not.

  //  T dummy;
  for(CopyIterator it(a_copier, CopyIterator::FROM); it.ok(); ++it)
    {
      const MotionItem& item = it();
      bufEntry b;
      b.item = &item;
      b.size = a_src[item.fromIndex].size(item.fromRegion, a_srcComps);
      sendBufferSize+=b.size;
      b.procID = item.procID;
      m_fromMe.push_back(b);
    }
  sort(m_fromMe.begin(), m_fromMe.end());
  for(CopyIterator it(a_copier, CopyIterator::TO); it.ok(); ++it)
    {
      const MotionItem& item = it();
      bufEntry b;
      b.item = &item;
      if (!(T::preAllocatable() == 2))
        {
          b.size = a_dest[item.toIndex].size(item.toRegion, a_destComps);
          recBufferSize+=b.size;
        }
      b.procID = item.procID;
      m_toMe.push_back(b);
    }
  sort(m_toMe.begin(), m_toMe.end());

  if(T::preAllocatable() == 2) // dynamic allocatable, need two pass
    {
      // in the non-preallocatable case, I need to message the
      // values for the m_toMe[*].size 
      if(m_fromMe.size() > 0)
        {
          MPI_Request nullrequest;

          int lastProc = -1;
          int messageIndex = 0;
          for(int i=0; i<m_fromMe.size(); ++i)
            {
              bufEntry& b = m_fromMe[i];
              if(b.procID == lastProc) messageIndex++;
              else                     messageIndex = 0;
              lastProc = b.procID;
              MPI_Isend(&(b.size), 1, MPI_INT, b.procID, 
                        messageIndex, Chombo_MPI::comm, &(nullrequest));
              MPI_Request_free(&(nullrequest));  // we don't wait on these sends, assume
              //    this memory really doesn't go anywhere.  Most MPI systems will fast message
              //    a single integer message
            }
        }
      if(m_toMe.size() > 0)
        {
          m_receiveRequests = new MPI_Request[m_toMe.size()];
          m_receiveStatus   = new MPI_Status[m_toMe.size()];
          int lastProc = -1;
          int messageIndex = 0;
          for(int i=0; i<m_toMe.size(); ++i)
            {
              bufEntry& b = m_toMe[i];
              if(b.procID == lastProc) messageIndex++;
              else                     messageIndex = 0;
              lastProc = b.procID;
              MPI_Irecv(&(b.size), 1, MPI_INT, b.procID, 
                        messageIndex, Chombo_MPI::comm, m_receiveRequests+i);
            }

          int result = MPI_Waitall(m_toMe.size(), m_receiveRequests, m_receiveStatus);
          if(result != MPI_SUCCESS)
            {
              MayDay::Error("First pass of two-phase communication failed");
            }
          for(int i=0; i<m_toMe.size(); ++i)  recBufferSize+= m_toMe[i].size;
          delete[] m_receiveRequests;
          delete[] m_receiveStatus;
        }
    }

  // allocate send and receveive buffer space.

  if(sendBufferSize > m_sendcapacity)
    {
      free(m_sendbuffer);
      m_sendbuffer = malloc(sendBufferSize);
      if(m_sendbuffer == NULL)
        {
          MayDay::Error("Out of memory in LevelData::allocatebuffers");
        }
      m_sendcapacity = sendBufferSize;
    }

  if(recBufferSize > m_reccapacity)
    {
      free(m_recbuffer);
      m_recbuffer = malloc(recBufferSize);
      if(m_recbuffer == NULL)
        {
          MayDay::Error("Out of memory in LevelData::allocatebuffers");
        }
      m_reccapacity = recBufferSize;
    }


  /*
  pout()<<"\n";
  for(int i=0; i<m_fromMe.size(); i++) 
    pout()<<m_fromMe[i].item->region<<"{"<<m_fromMe[i].procID<<"}"<<" ";
  pout() <<"::::";
  for(int i=0; i<m_toMe.size(); i++) 
    pout()<<m_toMe[i].item->region<<"{"<<m_toMe[i].procID<<"}"<<" ";
  pout() << endl;
  */

  char* nextFree = (char*)m_sendbuffer;
  if(m_fromMe.size() > 0)
  {
    for(unsigned int i=0; i<m_fromMe.size(); ++i)
      {
        m_fromMe[i].bufPtr = nextFree;
        nextFree += m_fromMe[i].size;
      }
  }

  nextFree = (char*)m_recbuffer;
  if(m_toMe.size() > 0)
  {
    for(unsigned int i=0; i<m_toMe.size(); ++i)
      {
        m_toMe[i].bufPtr = nextFree;
        nextFree += m_toMe[i].size;
      }
  }

  // since fromMe and toMe are sorted based on procID, messages can now be grouped
  // together on a per-processor basis.

}



template<class T> inline
void LevelData<T>::writeSendDataFromMeIntoBuffers(const LevelData<T>& a_src, 
                                                  const Interval&     a_srcComps) const
{


  for(unsigned int i=0; i<m_fromMe.size(); ++i)
    {
      const bufEntry& entry = m_fromMe[i];
      a_src[entry.item->fromIndex].linearOut(entry.bufPtr, entry.item->fromRegion, a_srcComps);
    }

}

template<class T> inline
void LevelData<T>::postSendsFromMe() const
{

  // now we get the magic of message coalescence
  // fromMe has already been sorted in the allocateBuffers() step.

  numSends = m_fromMe.size();
  if(numSends > 1){
  for(unsigned int i=m_fromMe.size()-1; i>0; --i)
    {
      if(m_fromMe[i].procID == m_fromMe[i-1].procID) 
        {
          numSends--;
          m_fromMe[i-1].size+=m_fromMe[i].size;
          m_fromMe[i].size = 0;
        }
    }
  }
  m_sendRequests = new MPI_Request[numSends];
  m_sendStatus = new MPI_Status[numSends];


  unsigned int next=0;
  for(int i=0; i<numSends; ++i)
    {
      const bufEntry& entry = m_fromMe[next];
      // cout<<procID()<< ": sending message of "<<entry.size;
      // cout<<" to proc "<<  entry.procID<<endl;
      MPI_Isend(entry.bufPtr, entry.size, MPI_BYTE, entry.procID, 
                0, Chombo_MPI::comm, m_sendRequests+i);
      ++next;
      while(next < m_fromMe.size() && m_fromMe[next].size == 0) ++next;
    }


}

template<class T> inline
void LevelData<T>::postReceivesToMe() const
{
  numReceives = m_toMe.size();

  if(numReceives > 1){
  for(unsigned int i=m_toMe.size()-1; i>0; --i)
    {
      if(m_toMe[i].procID == m_toMe[i-1].procID) 
        {
          numReceives--;
          m_toMe[i-1].size+=m_toMe[i].size;
          m_toMe[i].size = 0;
        }
    }
  }
  m_receiveRequests = new MPI_Request[numReceives];
  m_receiveStatus = new MPI_Status[numReceives];


  unsigned int next=0;
  for(int i=0; i<numReceives; ++i)
    {
      const bufEntry& entry = m_toMe[next];
      //cout<<procID()<< ": receiving message of "<<entry.size;
      //cout<<" from proc "<<  entry.procID<<endl;
      MPI_Irecv(entry.bufPtr, entry.size, MPI_BYTE, entry.procID, 
                0, Chombo_MPI::comm, m_receiveRequests+i);
      ++next;
      while(next < m_toMe.size() && m_toMe[next].size == 0) ++next;
    }

}


template<class T> inline
void LevelData<T>::unpackReceivesToMe(BoxLayoutData<T>& a_dest, 
                                      const Interval&   a_destComps) const
{

  if(numReceives > 0){
    int result = MPI_Waitall(numReceives, m_receiveRequests, m_receiveStatus);
    if(result != MPI_SUCCESS)
      {
        //hell if I know what to do about failed messaging here
      }

    for(unsigned int i=0; i<m_toMe.size(); ++i)
      {
        const bufEntry& entry = m_toMe[i];
        a_dest[entry.item->toIndex].linearIn(entry.bufPtr, entry.item->toRegion, a_destComps);
      }

    delete[] m_receiveRequests;
    delete[] m_receiveStatus;
  }
  numReceives = 0;
}

#endif

---