gather.hpp

Go to the documentation of this file.

// Copyright (C) 2010-2013 von Karman Institute for Fluid Dynamics, Belgium
//
// This software is distributed under the terms of the
// GNU Lesser General Public License version 3 (LGPLv3).
// See doc/lgpl.txt and doc/gpl.txt for the license text.

#ifndef cf3_common_PE_gather_hpp
#define cf3_common_PE_gather_hpp


#include "common/Foreach.hpp"
#include "common/BasicExceptions.hpp"

#include "common/PE/types.hpp"
#include "common/PE/datatype.hpp"

// #include "common/PE/debug.hpp" // for debugging mpi



namespace cf3 {
  namespace common {
    namespace PE {


namespace detail {


  template<typename T>
  inline void
  gatherc_impl(const Communicator& comm, const T* in_values, const int in_n, T* out_values, const int root, const  int stride )
  {
    // get data type and number of processors
    Datatype type = PE::get_mpi_datatype(*in_values);
    int nproc,irank;
    MPI_CHECK_RESULT(MPI_Comm_size,(comm,&nproc));
    MPI_CHECK_RESULT(MPI_Comm_rank,(comm,&irank));

    // if stride is greater than one
    cf3_assert( stride>0 );

    // set up out_buf
    T* out_buf=out_values;
    if (irank==root) if (in_values==out_values) {
      if ( (out_buf=new T[nproc*in_n*stride+1]) == (T*)0 ) throw cf3::common::NotEnoughMemory(FromHere(),"Could not allocate temporary buffer."); // +1 for avoiding possible zero allocation
    }

    // do the communication
    MPI_CHECK_RESULT(MPI_Gather, (const_cast<T*>(in_values), in_n*stride, type, out_buf, in_n*stride, type, root, comm));

    // deal with out_buf
    if (irank==root) if (in_values==out_values) {
      memcpy(out_values,out_buf,nproc*in_n*stride*sizeof(T));
      delete[] out_buf;
    }
  }


  template<typename T>
  inline void
  gathervm_impl(const Communicator& comm, const T* in_values, const int in_n, const int *in_map, T* out_values, const int *out_n, const int *out_map, const int root, const int stride )
  {
    // get data type and number of processors
    Datatype type = PE::get_mpi_datatype(*in_values);
    int nproc,irank;
    MPI_CHECK_RESULT(MPI_Comm_size,(comm,&nproc));
    MPI_CHECK_RESULT(MPI_Comm_rank,(comm,&irank));

    // if stride is smaller than one and unsupported functionality
    cf3_assert( stride>0 );

    // compute displacements both on send an receive side
    // also compute stride-multiplied send and receive counts
    int *out_nstride=0;
    int *out_disp=0;
    int out_sum=0;
    if (irank==root) {
      out_nstride=new int[nproc];
      out_disp=new int[nproc];
      out_disp[0]=0;
      for(int i=0; i<nproc-1; i++) {
        out_nstride[i]=stride*out_n[i];
        out_disp[i+1]=out_disp[i]+out_nstride[i];
      }
      out_nstride[nproc-1]=out_n[nproc-1]*stride;
      out_sum=out_disp[nproc-1]+stride*out_n[nproc-1];
    }

    // compute total number of send and receive items
    const int in_sum=stride*in_n;

    // set up in_buf
    T *in_buf=(T*)in_values;
    if (in_map!=0) {
      if ( (in_buf=new T[in_sum+1]) == (T*)0 ) throw cf3::common::NotEnoughMemory(FromHere(),"Could not allocate temporary buffer."); // +1 for avoiding possible zero allocation
      if (stride==1) { for(int i=0; i<in_sum; i++) in_buf[i]=in_values[in_map[i]]; }
      else { for(int i=0; i<in_sum/stride; i++) memcpy(&in_buf[stride*i],&in_values[stride*in_map[i]],stride*sizeof(T)); }
    }

    // set up out_buf
    T *out_buf=out_values;
    if (irank==root) if ((out_map!=0)||(in_values==out_values)) {
      if ( (out_buf=new T[out_sum+1]) == (T*)0 ) throw cf3::common::NotEnoughMemory(FromHere(),"Could not allocate temporary buffer."); // +1 for avoiding possible zero allocation
    }

    // do the communication
    MPI_CHECK_RESULT(MPI_Gatherv, (in_buf, in_sum, type, out_buf, out_nstride, out_disp, type, root, comm));

    // re-populate out_values
    if (irank==root) {
      if (out_map!=0) {
        if (stride==1) { for(int i=0; i<out_sum; i++) out_values[out_map[i]]=out_buf[i]; }
        else { for(int i=0; i<out_sum/stride; i++) memcpy(&out_values[stride*out_map[i]],&out_buf[stride*i],stride*sizeof(T)); }
        delete[] out_buf;
      } else if (in_values==out_values) {
        memcpy(out_values,out_buf,out_sum*sizeof(T));
        delete[] out_buf;
      }
    }

    // free internal memory
    if (in_map!=0) delete[] in_buf;
    delete[] out_disp;
    delete[] out_nstride;
  }


} // end namespace detail


template<typename T>
inline T*
gather(const Communicator& comm, const T* in_values, const int in_n, T* out_values, const int root, const int stride=1)
{
  // get nproc, irank
  int nproc,irank;
  MPI_CHECK_RESULT(MPI_Comm_size,(comm,&nproc));
  MPI_CHECK_RESULT(MPI_Comm_rank,(comm,&irank));

  // allocate out_buf if incoming pointer is null
  T* out_buf=out_values;
  if (irank==root) if (out_values==0) {
    const int size=stride*nproc*in_n>1?stride*nproc*in_n:1;
    if ( (out_buf=new T[size]) == (T*)0 ) throw cf3::common::NotEnoughMemory(FromHere(),"Could not allocate temporary buffer.");
  }

  // call c_impl
  if (irank==root) {
    detail::gatherc_impl(comm, in_values, in_n, out_buf, root, stride);
  } else {
    detail::gatherc_impl(comm, in_values, in_n, (T*)0, root, stride);
  }
  return out_buf;
}


template<typename T>
inline void
gather(const Communicator& comm, const std::vector<T>& in_values, std::vector<T>& out_values, const int root, const int stride=1)
{
  // get nproc, irank
  int nproc,irank;
  MPI_CHECK_RESULT(MPI_Comm_size,(comm,&nproc));
  MPI_CHECK_RESULT(MPI_Comm_rank,(comm,&irank));

  // set out_values's sizes
  cf3_assert( in_values.size() % (stride) == 0 );
  int in_n=(int)in_values.size();
  if (irank==root) {
    out_values.resize(in_n*nproc);
    out_values.reserve(in_n*nproc);
  }

  // call c_impl
  if (irank==root) {
    detail::gatherc_impl(comm, (T*)(&in_values[0]), in_n/(stride), (T*)(&out_values[0]), root, stride);
  } else {
    detail::gatherc_impl(comm, (T*)(&in_values[0]), in_n/(stride), (T*)0, root, stride);
  }
}


//needs a forward
template<typename T>
inline T*
gather(const Communicator& comm, const T* in_values, const int in_n, const int *in_map, T* out_values, int *out_n, const int *out_map, const int root, const int stride=1);

template<typename T>
inline T*
gather(const Communicator& comm, const T* in_values, const int in_n, T* out_values, int *out_n, const int root, const int stride=1)
{
  // call mapped variable gather
  return gather(comm,in_values,in_n,0,out_values,out_n,0,root,stride);
}


//needs a forward
template<typename T>
inline void
gather(const Communicator& comm, const std::vector<T>& in_values, const int in_n, const std::vector<int>& in_map, std::vector<T>& out_values, std::vector<int>& out_n, const std::vector<int>& out_map, const int root, const int stride=1);

template<typename T>
inline void
gather(const Communicator& comm, const std::vector<T>& in_values, const int in_n, std::vector<T>& out_values, std::vector<int>& out_n, const int root, const int stride=1)
{
  // call mapped variable gather
  std::vector<int> in_map(0);
  std::vector<int> out_map(0);

  int irank;
  MPI_CHECK_RESULT(MPI_Comm_rank,(comm,&irank));

  if ((irank==root)&&(&in_values[0]==&out_values[0]))
  {
    std::vector<T> out_tmp(0);
    gather(comm,in_values,in_n,in_map,out_tmp,out_n,out_map,root,stride);
    out_values.assign(out_tmp.begin(),out_tmp.end());
  }
  else
  {
    gather(comm,in_values,in_n,in_map,out_values,out_n,out_map,root,stride);
  }

}


template<typename T>
inline T*
gather(const Communicator& comm, const T* in_values, const int in_n, const int *in_map, T* out_values, int *out_n, const int *out_map, const int root, const int stride)
{
  // get nproc, irank
  int nproc,irank;
  MPI_CHECK_RESULT(MPI_Comm_size,(comm,&nproc));
  MPI_CHECK_RESULT(MPI_Comm_rank,(comm,&irank));

  // if out_n consist of -1s then communicate for number of receives
  int out_sum=0;
  for (int i=0; i<nproc; i++) out_sum+=out_n[i];
  if (out_sum==-nproc) {
    if (out_map!=0) throw cf3::common::ParallelError(FromHere(),"Trying to perform communication with receive map while receive counts are unknown, this is bad usage of parallel environment.");
    detail::gatherc_impl(comm,&in_n,1,out_n,root,1);
    out_sum=0;
    if (irank==root) for (int i=0; i<nproc; i++) out_sum+=out_n[i];
  }

  // allocate out_buf if incoming pointer is null
  T* out_buf=out_values;
  if (irank==root) {
    if (out_values==0) {
      if (out_map!=0){
        int out_sum_tmp=0;
        for (int i=0; i<out_sum; i++) out_sum_tmp=out_map[i]>out_sum_tmp?out_map[i]:out_sum_tmp;
        out_sum=out_sum_tmp+1;
      }
      if ( (out_buf=new T[stride*out_sum]) == (T*)0 ) throw cf3::common::NotEnoughMemory(FromHere(),"Could not allocate temporary buffer.");
    }
  }

  // call vm_impl
  if (irank==root) {
    detail::gathervm_impl(comm, in_values, in_n, in_map, out_buf, out_n, out_map, root, stride);
  } else {
    detail::gathervm_impl(comm, in_values, in_n, in_map, (T*)0, (int*)0, (int*)0, root, stride);
  }
  return out_buf;
}


template<typename T>
inline void
gather(const Communicator& comm, const std::vector<T>& in_values, const int in_n, const std::vector<int>& in_map, std::vector<T>& out_values, std::vector<int>& out_n, const std::vector<int>& out_map, const int root, const int stride)
{
  // number of processes and checking in_n and out_n (out_n deliberately throws exception because the vector can arrive from arbitrary previous usage)
  int nproc,irank;
  MPI_CHECK_RESULT(MPI_Comm_size,(comm,&nproc));
  MPI_CHECK_RESULT(MPI_Comm_rank,(comm,&irank));
  if ((int)out_n.size()!=nproc) cf3::common::BadValue(FromHere(),"Size of vector for number of items to be received does not match to number of processes.");

  // compute number of send and receive
  int out_sum=0;
  boost_foreach( int i, out_n ) out_sum+=i;

  // if necessary, do communication for out_n
  if (out_sum == -nproc){
    if (out_map.size()!=0) throw cf3::common::ParallelError(FromHere(),"Trying to perform communication with receive map while receive counts are unknown, this is bad usage of parallel environment.");
    detail::gatherc_impl(comm,&in_n,1,&out_n[0],root,1);
    out_sum=0;
    if (irank==root) boost_foreach( int & i, out_n ) out_sum+=i;
  }

  // resize out_values if vector size is zero
  if (irank==root) {
    if (out_values.size() == 0 ){
      if (out_map.size()!=0) {
        out_sum=0;
        boost_foreach( int i, out_map ) out_sum=i>out_sum?i:out_sum;
        if (out_sum!=0) out_sum++;
      }
      out_values.resize(stride*out_sum);
      out_values.reserve(stride*out_sum);
    }
  }

  // call vm_impl
  if (irank==root) {
    detail::gathervm_impl(comm, (T*)(&in_values[0]), in_n, (in_map.empty() ? nullptr : &in_map[0]), (T*)(&out_values[0]), &out_n[0], (out_map.empty() ? nullptr : &out_map[0]), root, stride);
  } else {
    detail::gathervm_impl(comm, (T*)(&in_values[0]), in_n, (in_map.empty() ? nullptr : &in_map[0]), (T*)0, (int*)0, (int*)0, root, stride);
  }
}


} // namespace PE
} // namespace common
} // namespace cf3


#endif // cf3_common_PE_gather_hpp