all_reduce.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_all_reduce_hpp
#define cf3_common_PE_all_reduce_hpp


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

#include "common/PE/types.hpp"
#include "common/PE/operations.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, typename Op>
  inline void
  all_reduce_impl(const Communicator& comm,  Op, const T* in_values, const int in_n, const int *in_map, T* out_values, const int *out_map, const int stride)
  { 
    // get data type, op and some checkings
    Datatype type = get_mpi_datatype(*in_values);
    Operation op_= get_mpi_op<T,Op>::op();
    cf3_assert( stride>0 );

    // there is in_map
    T *in_buf=(T*)in_values;
    if (in_map!=0){
      if ( (in_buf=new T[stride*in_n+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_n; i++) in_buf[i]=in_values[in_map[i]]; }
      else { for(int i=0; i<in_n; 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 ((out_map!=0)||(in_values==out_values)) {
      if ( (out_buf=new T[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_Allreduce, ( in_buf, out_buf, in_n*stride, type, op_, comm ));

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

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


} // end namespace detail


template<typename T, typename Op>
inline T*
all_reduce(const Communicator& comm, const Op& op, const T* in_values, const int in_n, T* out_values, const int stride=1)
{
  // allocate out_buf if incoming pointer is null
  T* out_buf=out_values;
  if (out_values==0) {
    const int size=stride*in_n>1?stride*in_n:1;
    if ( (out_buf=new T[size]) == (T*)0 ) throw cf3::common::NotEnoughMemory(FromHere(),"Could not allocate temporary buffer.");
  }

  // call impl
  detail::all_reduce_impl(comm,op,in_values,in_n,0,out_buf,0,stride);
  return out_buf;
}


template<typename T, typename Op>
inline void
all_reduce(const Communicator& comm, const Op& op, const std::vector<T>& in_values, std::vector<T>& out_values, const int stride=1)
{
  // set out_values's sizes
  cf3_assert( in_values.size() % stride == 0 );
  out_values.resize(in_values.size());
  out_values.reserve(in_values.size());

  // call impl
  detail::all_reduce_impl(comm, op, (T*)(&in_values[0]), in_values.size()/stride, 0, (T*)(&out_values[0]), 0, stride);
}


template<typename T, typename Op>
inline T*
all_reduce(const Communicator& comm, const Op& op, const T* in_values, const int in_n, const int *in_map, T* out_values, const int *out_map, const int stride=1)
{
  // allocate out_buf if incoming pointer is null
  T* out_buf=out_values;
  if (out_values==0) {
    int out_sum=in_n;
    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 impl
  detail::all_reduce_impl(comm,op,in_values,in_n,in_map,out_buf,out_map,stride);
  return out_buf;
}


template<typename T, typename Op>
inline void
all_reduce(const Communicator& comm, const Op& op, const std::vector<T>& in_values, const std::vector<int>& in_map, std::vector<T>& out_values, const std::vector<int>& out_map, const int stride=1)
{
  // set out_values's sizes
  cf3_assert( in_values.size() % stride == 0 );

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

  // call impl
  detail::all_reduce_impl(comm, op, (T*)(&in_values[0]), in_map.size(), (in_map.empty() ? nullptr : &in_map[0]), (T*)(&out_values[0]), (out_map.empty() ? nullptr : &out_map[0]), stride);
}


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


#endif // cf3_common_PE_all_reduce_hpp