boost/mpi/collectives/reduce.hpp

   1 // Copyright (C) 2005-2006 Douglas Gregor <doug.gregor@gmail.com>.
   2 // Copyright (C) 2004 The Trustees of Indiana University
   3
   4 // Use, modification and distribution is subject to the Boost Software
   5 // License, Version 1.0. (See accompanying file LICENSE_1_0.txt or copy at
   6 // http://www.boost.org/LICENSE_1_0.txt)
   7
   8 //   Authors: Douglas Gregor
   9 //            Andrew Lumsdaine
  10
  11 // Message Passing Interface 1.1 -- Section 4.9.1. Reduce
  12 #ifndef BOOST_MPI_REDUCE_HPP
  13 #define BOOST_MPI_REDUCE_HPP
  14
  15 #include <boost/mpi/exception.hpp>
  16 #include <boost/mpi/datatype.hpp>
  17
  18 // For (de-)serializing sends and receives
  19 #include <boost/mpi/packed_oarchive.hpp>
  20 #include <boost/mpi/packed_iarchive.hpp>
  21
  22 // For packed_[io]archive sends and receives
  23 #include <boost/mpi/detail/point_to_point.hpp>
  24
  25 #include <boost/mpi/communicator.hpp>
  26 #include <boost/mpi/environment.hpp>
  27 #include <boost/mpi/detail/computation_tree.hpp>
  28 #include <boost/mpi/operations.hpp>
  29 #include <algorithm>
  30 #include <exception>
  31 #include <boost/assert.hpp>
  32 #include <boost/scoped_array.hpp>
  33
  34 namespace boost { namespace mpi {
  35
  36
  37 /************************************************************************
  38  * Implementation details                                               *
  39  ************************************************************************/
  40 namespace detail {
  41   /**********************************************************************
  42    * Simple reduction with MPI_Reduce                                   *
  43    **********************************************************************/
  44   // We are reducing at the root for a type that has an associated MPI
  45   // datatype and operation, so we'll use MPI_Reduce directly.
  46   template<typename T, typename Op>
  47   void
  48   reduce_impl(const communicator& comm, const T* in_values, int n,
  49               T* out_values, Op op, int root, mpl::true_ /*is_mpi_op*/,
  50               mpl::true_/*is_mpi_datatype*/)
  51   {
  52     BOOST_MPI_CHECK_RESULT(MPI_Reduce,
  53                            (const_cast<T*>(in_values), out_values, n,
  54                             boost::mpi::get_mpi_datatype<T>(*in_values),
  55                             (is_mpi_op<Op, T>::op()), root, comm));
  56   }
  57
  58   // We are reducing to the root for a type that has an associated MPI
  59   // datatype and operation, so we'll use MPI_Reduce directly.
  60   template<typename T, typename Op>
  61   void
  62   reduce_impl(const communicator& comm, const T* in_values, int n, Op op,
  63               int root, mpl::true_ /*is_mpi_op*/, mpl::true_/*is_mpi_datatype*/)
  64   {
  65     BOOST_MPI_CHECK_RESULT(MPI_Reduce,
  66                            (const_cast<T*>(in_values), 0, n,
  67                             boost::mpi::get_mpi_datatype<T>(*in_values),
  68                             (is_mpi_op<Op, T>::op()), root, comm));
  69   }
  70
  71   /**********************************************************************
  72    * User-defined reduction with MPI_Reduce                             *
  73    **********************************************************************/
  74
  75   // We are reducing at the root for a type that has an associated MPI
  76   // datatype but with a custom operation. We'll use MPI_Reduce
  77   // directly, but we'll need to create an MPI_Op manually.
  78   template<typename T, typename Op>
  79   void
  80   reduce_impl(const communicator& comm, const T* in_values, int n,
  81               T* out_values, Op op, int root, mpl::false_ /*is_mpi_op*/,
  82               mpl::true_/*is_mpi_datatype*/)
  83   {
  84     user_op<Op, T> mpi_op(op);
  85     BOOST_MPI_CHECK_RESULT(MPI_Reduce,
  86                            (const_cast<T*>(in_values), out_values, n,
  87                             boost::mpi::get_mpi_datatype<T>(*in_values),
  88                             mpi_op.get_mpi_op(), root, comm));
  89   }
  90
  91   // We are reducing to the root for a type that has an associated MPI
  92   // datatype but with a custom operation. We'll use MPI_Reduce
  93   // directly, but we'll need to create an MPI_Op manually.
  94   template<typename T, typename Op>
  95   void
  96   reduce_impl(const communicator& comm, const T* in_values, int n, Op op,
  97               int root, mpl::false_/*is_mpi_op*/, mpl::true_/*is_mpi_datatype*/)
  98   {
  99     user_op<Op, T> mpi_op(op);
 100     BOOST_MPI_CHECK_RESULT(MPI_Reduce,
 101                            (const_cast<T*>(in_values), 0, n,
 102                             boost::mpi::get_mpi_datatype<T>(*in_values),
 103                             mpi_op.get_mpi_op(), root, comm));
 104   }
 105
 106   /**********************************************************************
 107    * User-defined, tree-based reduction for non-MPI data types          *
 108    **********************************************************************/
 109
 110   // Commutative reduction
 111   template<typename T, typename Op>
 112   void
 113   tree_reduce_impl(const communicator& comm, const T* in_values, int n,
 114                    T* out_values, Op op, int root,
 115                    mpl::true_ /*is_commutative*/)
 116   {
 117     std::copy(in_values, in_values + n, out_values);
 118
 119     int size = comm.size();
 120     int rank = comm.rank();
 121
 122     // The computation tree we will use.
 123     detail::computation_tree tree(rank, size, root);
 124
 125     int tag = environment::collectives_tag();
 126
 127     MPI_Status status;
 128     int children = 0;
 129     for (int child = tree.child_begin();
 130          children < tree.branching_factor() && child != root;
 131          ++children, child = (child + 1) % size) {
 132       // Receive archive
 133       packed_iarchive ia(comm);
 134       detail::packed_archive_recv(comm, child, tag, ia, status);
 135
 136       T incoming;
 137       for (int i = 0; i < n; ++i) {
 138         ia >> incoming;
 139         out_values[i] = op(out_values[i], incoming);
 140       }
 141     }
 142
 143     // For non-roots, send the result to the parent.
 144     if (tree.parent() != rank) {
 145       packed_oarchive oa(comm);
 146       for (int i = 0; i < n; ++i)
 147         oa << out_values[i];
 148       detail::packed_archive_send(comm, tree.parent(), tag, oa);
 149     }
 150   }
 151
 152   // Commutative reduction from a non-root.
 153   template<typename T, typename Op>
 154   void
 155   tree_reduce_impl(const communicator& comm, const T* in_values, int n, Op op,
 156                    int root, mpl::true_ /*is_commutative*/)
 157   {
 158     scoped_array<T> results(new T[n]);
 159     detail::tree_reduce_impl(comm, in_values, n, results.get(), op, root,
 160                              mpl::true_());
 161   }
 162
 163   // Non-commutative reduction
 164   template<typename T, typename Op>
 165   void
 166   tree_reduce_impl(const communicator& comm, const T* in_values, int n,
 167                    T* out_values, Op op, int root,
 168                    mpl::false_ /*is_commutative*/)
 169   {
 170     int tag = environment::collectives_tag();
 171
 172     int left_child = root / 2;
 173     int right_child = (root + comm.size()) / 2;
 174
 175     MPI_Status status;
 176     if (left_child != root) {
 177       // Receive value from the left child and merge it with the value
 178       // we had incoming.
 179       packed_iarchive ia(comm);
 180       detail::packed_archive_recv(comm, left_child, tag, ia, status);
 181       T incoming;
 182       for (int i = 0; i < n; ++i) {
 183         ia >> incoming;
 184         out_values[i] = op(incoming, in_values[i]);
 185       }
 186     } else {
 187       // There was no left value, so copy our incoming value.
 188       std::copy(in_values, in_values + n, out_values);
 189     }
 190
 191     if (right_child != root) {
 192       // Receive value from the right child and merge it with the
 193       // value we had incoming.
 194       packed_iarchive ia(comm);
 195       detail::packed_archive_recv(comm, right_child, tag, ia, status);
 196       T incoming;
 197       for (int i = 0; i < n; ++i) {
 198         ia >> incoming;
 199         out_values[i] = op(out_values[i], incoming);
 200       }
 201     }
 202   }
 203
 204   // Non-commutative reduction from a non-root.
 205   template<typename T, typename Op>
 206   void
 207   tree_reduce_impl(const communicator& comm, const T* in_values, int n, Op op,
 208                    int root, mpl::false_ /*is_commutative*/)
 209   {
 210     int size = comm.size();
 211     int rank = comm.rank();
 212
 213     int tag = environment::collectives_tag();
 214
 215     // Determine our parents and children in the commutative binary
 216     // computation tree.
 217     int grandparent = root;
 218     int parent = root;
 219     int left_bound = 0;
 220     int right_bound = size;
 221     int left_child, right_child;
 222     do {
 223       left_child = (left_bound + parent) / 2;
 224       right_child = (parent + right_bound) / 2;
 225
 226       if (rank < parent) {
 227         // Go left.
 228         grandparent = parent;
 229         right_bound = parent;
 230         parent = left_child;
 231       } else if (rank > parent) {
 232         // Go right.
 233         grandparent = parent;
 234         left_bound = parent + 1;
 235         parent = right_child;
 236       } else {
 237         // We've found the parent
 238         break;
 239       }
 240     } while (true);
 241
 242     // Our parent is the grandparent of our children. This is a slight
 243     // abuse of notation, but it makes the send-to-parent below make
 244     // more sense.
 245     parent = grandparent;
 246
 247     MPI_Status status;
 248     scoped_array<T> out_values(new T[n]);
 249     if (left_child != rank) {
 250       // Receive value from the left child and merge it with the value
 251       // we had incoming.
 252       packed_iarchive ia(comm);
 253       detail::packed_archive_recv(comm, left_child, tag, ia, status);
 254       T incoming;
 255       for (int i = 0; i < n; ++i) {
 256         ia >> incoming;
 257         out_values[i] = op(incoming, in_values[i]);
 258       }
 259     } else {
 260       // There was no left value, so copy our incoming value.
 261       std::copy(in_values, in_values + n, out_values.get());
 262     }
 263
 264     if (right_child != rank) {
 265       // Receive value from the right child and merge it with the
 266       // value we had incoming.
 267       packed_iarchive ia(comm);
 268       detail::packed_archive_recv(comm, right_child, tag, ia, status);
 269       T incoming;
 270       for (int i = 0; i < n; ++i) {
 271         ia >> incoming;
 272         out_values[i] = op(out_values[i], incoming);
 273       }
 274     }
 275
 276     // Send the combined value to our parent.
 277     packed_oarchive oa(comm);
 278     for (int i = 0; i < n; ++i)
 279       oa << out_values[i];
 280     detail::packed_archive_send(comm, parent, tag, oa);
 281   }
 282
 283   // We are reducing at the root for a type that has no associated MPI
 284   // datatype and operation, so we'll use a simple tree-based
 285   // algorithm.
 286   template<typename T, typename Op>
 287   void
 288   reduce_impl(const communicator& comm, const T* in_values, int n,
 289               T* out_values, Op op, int root, mpl::false_ /*is_mpi_op*/,
 290               mpl::false_ /*is_mpi_datatype*/)
 291   {
 292     detail::tree_reduce_impl(comm, in_values, n, out_values, op, root,
 293                              is_commutative<Op, T>());
 294   }
 295
 296   // We are reducing to the root for a type that has no associated MPI
 297   // datatype and operation, so we'll use a simple tree-based
 298   // algorithm.
 299   template<typename T, typename Op>
 300   void
 301   reduce_impl(const communicator& comm, const T* in_values, int n, Op op,
 302               int root, mpl::false_ /*is_mpi_op*/,
 303               mpl::false_ /*is_mpi_datatype*/)
 304   {
 305     detail::tree_reduce_impl(comm, in_values, n, op, root,
 306                              is_commutative<Op, T>());
 307   }
 308 } // end namespace detail
 309
 310 template<typename T, typename Op>
 311 void
 312 reduce(const communicator& comm, const T* in_values, int n, T* out_values,
 313        Op op, int root)
 314 {
 315   if (comm.rank() == root)
 316     detail::reduce_impl(comm, in_values, n, out_values, op, root,
 317                         is_mpi_op<Op, T>(), is_mpi_datatype<T>());
 318   else
 319     detail::reduce_impl(comm, in_values, n, op, root,
 320                         is_mpi_op<Op, T>(), is_mpi_datatype<T>());
 321 }
 322
 323 template<typename T, typename Op>
 324 void
 325 reduce(const communicator& comm, const T* in_values, int n, Op op, int root)
 326 {
 327   BOOST_ASSERT(comm.rank() != root);
 328
 329   detail::reduce_impl(comm, in_values, n, op, root,
 330                       is_mpi_op<Op, T>(), is_mpi_datatype<T>());
 331 }
 332
 333 template<typename T, typename Op>
 334 void
 335 reduce(const communicator& comm, const T& in_value, T& out_value, Op op,
 336        int root)
 337 {
 338   if (comm.rank() == root)
 339     detail::reduce_impl(comm, &in_value, 1, &out_value, op, root,
 340                         is_mpi_op<Op, T>(), is_mpi_datatype<T>());
 341   else
 342     detail::reduce_impl(comm, &in_value, 1, op, root,
 343                         is_mpi_op<Op, T>(), is_mpi_datatype<T>());
 344 }
 345
 346 template<typename T, typename Op>
 347 void reduce(const communicator& comm, const T& in_value, Op op, int root)
 348 {
 349   BOOST_ASSERT(comm.rank() != root);
 350
 351   detail::reduce_impl(comm, &in_value, 1, op, root,
 352                       is_mpi_op<Op, T>(), is_mpi_datatype<T>());
 353 }
 354
 355 } } // end namespace boost::mpi
 356
 357 #endif // BOOST_MPI_REDUCE_HPP