/usr/include/boost/mpi/collectives/reduce.hpp is in libboost1.46-dev 1.46.1-7ubuntu3.
This file is owned by root:root, with mode 0o644.
The actual contents of the file can be viewed below.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 | // Copyright (C) 2005-2006 Douglas Gregor <doug.gregor@gmail.com>.
// Copyright (C) 2004 The Trustees of Indiana University
// Use, modification and distribution is subject to the Boost Software
// License, Version 1.0. (See accompanying file LICENSE_1_0.txt or copy at
// http://www.boost.org/LICENSE_1_0.txt)
// Authors: Douglas Gregor
// Andrew Lumsdaine
// Message Passing Interface 1.1 -- Section 4.9.1. Reduce
#ifndef BOOST_MPI_REDUCE_HPP
#define BOOST_MPI_REDUCE_HPP
#include <boost/mpi/exception.hpp>
#include <boost/mpi/datatype.hpp>
// For (de-)serializing sends and receives
#include <boost/mpi/packed_oarchive.hpp>
#include <boost/mpi/packed_iarchive.hpp>
// For packed_[io]archive sends and receives
#include <boost/mpi/detail/point_to_point.hpp>
#include <boost/mpi/communicator.hpp>
#include <boost/mpi/environment.hpp>
#include <boost/mpi/detail/computation_tree.hpp>
#include <boost/mpi/operations.hpp>
#include <algorithm>
#include <exception>
#include <boost/assert.hpp>
#include <boost/scoped_array.hpp>
namespace boost { namespace mpi {
/************************************************************************
* Implementation details *
************************************************************************/
namespace detail {
/**********************************************************************
* Simple reduction with MPI_Reduce *
**********************************************************************/
// We are reducing at the root for a type that has an associated MPI
// datatype and operation, so we'll use MPI_Reduce directly.
template<typename T, typename Op>
void
reduce_impl(const communicator& comm, const T* in_values, int n,
T* out_values, Op op, int root, mpl::true_ /*is_mpi_op*/,
mpl::true_/*is_mpi_datatype*/)
{
BOOST_MPI_CHECK_RESULT(MPI_Reduce,
(const_cast<T*>(in_values), out_values, n,
boost::mpi::get_mpi_datatype<T>(*in_values),
(is_mpi_op<Op, T>::op()), root, comm));
}
// We are reducing to the root for a type that has an associated MPI
// datatype and operation, so we'll use MPI_Reduce directly.
template<typename T, typename Op>
void
reduce_impl(const communicator& comm, const T* in_values, int n, Op op,
int root, mpl::true_ /*is_mpi_op*/, mpl::true_/*is_mpi_datatype*/)
{
BOOST_MPI_CHECK_RESULT(MPI_Reduce,
(const_cast<T*>(in_values), 0, n,
boost::mpi::get_mpi_datatype<T>(*in_values),
(is_mpi_op<Op, T>::op()), root, comm));
}
/**********************************************************************
* User-defined reduction with MPI_Reduce *
**********************************************************************/
// We are reducing at the root for a type that has an associated MPI
// datatype but with a custom operation. We'll use MPI_Reduce
// directly, but we'll need to create an MPI_Op manually.
template<typename T, typename Op>
void
reduce_impl(const communicator& comm, const T* in_values, int n,
T* out_values, Op op, int root, mpl::false_ /*is_mpi_op*/,
mpl::true_/*is_mpi_datatype*/)
{
user_op<Op, T> mpi_op(op);
BOOST_MPI_CHECK_RESULT(MPI_Reduce,
(const_cast<T*>(in_values), out_values, n,
boost::mpi::get_mpi_datatype<T>(*in_values),
mpi_op.get_mpi_op(), root, comm));
}
// We are reducing to the root for a type that has an associated MPI
// datatype but with a custom operation. We'll use MPI_Reduce
// directly, but we'll need to create an MPI_Op manually.
template<typename T, typename Op>
void
reduce_impl(const communicator& comm, const T* in_values, int n, Op op,
int root, mpl::false_/*is_mpi_op*/, mpl::true_/*is_mpi_datatype*/)
{
user_op<Op, T> mpi_op(op);
BOOST_MPI_CHECK_RESULT(MPI_Reduce,
(const_cast<T*>(in_values), 0, n,
boost::mpi::get_mpi_datatype<T>(*in_values),
mpi_op.get_mpi_op(), root, comm));
}
/**********************************************************************
* User-defined, tree-based reduction for non-MPI data types *
**********************************************************************/
// Commutative reduction
template<typename T, typename Op>
void
tree_reduce_impl(const communicator& comm, const T* in_values, int n,
T* out_values, Op op, int root,
mpl::true_ /*is_commutative*/)
{
std::copy(in_values, in_values + n, out_values);
int size = comm.size();
int rank = comm.rank();
// The computation tree we will use.
detail::computation_tree tree(rank, size, root);
int tag = environment::collectives_tag();
MPI_Status status;
int children = 0;
for (int child = tree.child_begin();
children < tree.branching_factor() && child != root;
++children, child = (child + 1) % size) {
// Receive archive
packed_iarchive ia(comm);
detail::packed_archive_recv(comm, child, tag, ia, status);
T incoming;
for (int i = 0; i < n; ++i) {
ia >> incoming;
out_values[i] = op(out_values[i], incoming);
}
}
// For non-roots, send the result to the parent.
if (tree.parent() != rank) {
packed_oarchive oa(comm);
for (int i = 0; i < n; ++i)
oa << out_values[i];
detail::packed_archive_send(comm, tree.parent(), tag, oa);
}
}
// Commutative reduction from a non-root.
template<typename T, typename Op>
void
tree_reduce_impl(const communicator& comm, const T* in_values, int n, Op op,
int root, mpl::true_ /*is_commutative*/)
{
scoped_array<T> results(new T[n]);
detail::tree_reduce_impl(comm, in_values, n, results.get(), op, root,
mpl::true_());
}
// Non-commutative reduction
template<typename T, typename Op>
void
tree_reduce_impl(const communicator& comm, const T* in_values, int n,
T* out_values, Op op, int root,
mpl::false_ /*is_commutative*/)
{
int tag = environment::collectives_tag();
int left_child = root / 2;
int right_child = (root + comm.size()) / 2;
MPI_Status status;
if (left_child != root) {
// Receive value from the left child and merge it with the value
// we had incoming.
packed_iarchive ia(comm);
detail::packed_archive_recv(comm, left_child, tag, ia, status);
T incoming;
for (int i = 0; i < n; ++i) {
ia >> incoming;
out_values[i] = op(incoming, in_values[i]);
}
} else {
// There was no left value, so copy our incoming value.
std::copy(in_values, in_values + n, out_values);
}
if (right_child != root) {
// Receive value from the right child and merge it with the
// value we had incoming.
packed_iarchive ia(comm);
detail::packed_archive_recv(comm, right_child, tag, ia, status);
T incoming;
for (int i = 0; i < n; ++i) {
ia >> incoming;
out_values[i] = op(out_values[i], incoming);
}
}
}
// Non-commutative reduction from a non-root.
template<typename T, typename Op>
void
tree_reduce_impl(const communicator& comm, const T* in_values, int n, Op op,
int root, mpl::false_ /*is_commutative*/)
{
int size = comm.size();
int rank = comm.rank();
int tag = environment::collectives_tag();
// Determine our parents and children in the commutative binary
// computation tree.
int grandparent = root;
int parent = root;
int left_bound = 0;
int right_bound = size;
int left_child, right_child;
do {
left_child = (left_bound + parent) / 2;
right_child = (parent + right_bound) / 2;
if (rank < parent) {
// Go left.
grandparent = parent;
right_bound = parent;
parent = left_child;
} else if (rank > parent) {
// Go right.
grandparent = parent;
left_bound = parent + 1;
parent = right_child;
} else {
// We've found the parent
break;
}
} while (true);
// Our parent is the grandparent of our children. This is a slight
// abuse of notation, but it makes the send-to-parent below make
// more sense.
parent = grandparent;
MPI_Status status;
scoped_array<T> out_values(new T[n]);
if (left_child != rank) {
// Receive value from the left child and merge it with the value
// we had incoming.
packed_iarchive ia(comm);
detail::packed_archive_recv(comm, left_child, tag, ia, status);
T incoming;
for (int i = 0; i < n; ++i) {
ia >> incoming;
out_values[i] = op(incoming, in_values[i]);
}
} else {
// There was no left value, so copy our incoming value.
std::copy(in_values, in_values + n, out_values.get());
}
if (right_child != rank) {
// Receive value from the right child and merge it with the
// value we had incoming.
packed_iarchive ia(comm);
detail::packed_archive_recv(comm, right_child, tag, ia, status);
T incoming;
for (int i = 0; i < n; ++i) {
ia >> incoming;
out_values[i] = op(out_values[i], incoming);
}
}
// Send the combined value to our parent.
packed_oarchive oa(comm);
for (int i = 0; i < n; ++i)
oa << out_values[i];
detail::packed_archive_send(comm, parent, tag, oa);
}
// We are reducing at the root for a type that has no associated MPI
// datatype and operation, so we'll use a simple tree-based
// algorithm.
template<typename T, typename Op>
void
reduce_impl(const communicator& comm, const T* in_values, int n,
T* out_values, Op op, int root, mpl::false_ /*is_mpi_op*/,
mpl::false_ /*is_mpi_datatype*/)
{
detail::tree_reduce_impl(comm, in_values, n, out_values, op, root,
is_commutative<Op, T>());
}
// We are reducing to the root for a type that has no associated MPI
// datatype and operation, so we'll use a simple tree-based
// algorithm.
template<typename T, typename Op>
void
reduce_impl(const communicator& comm, const T* in_values, int n, Op op,
int root, mpl::false_ /*is_mpi_op*/,
mpl::false_ /*is_mpi_datatype*/)
{
detail::tree_reduce_impl(comm, in_values, n, op, root,
is_commutative<Op, T>());
}
} // end namespace detail
template<typename T, typename Op>
void
reduce(const communicator& comm, const T* in_values, int n, T* out_values,
Op op, int root)
{
if (comm.rank() == root)
detail::reduce_impl(comm, in_values, n, out_values, op, root,
is_mpi_op<Op, T>(), is_mpi_datatype<T>());
else
detail::reduce_impl(comm, in_values, n, op, root,
is_mpi_op<Op, T>(), is_mpi_datatype<T>());
}
template<typename T, typename Op>
void
reduce(const communicator& comm, const T* in_values, int n, Op op, int root)
{
BOOST_ASSERT(comm.rank() != root);
detail::reduce_impl(comm, in_values, n, op, root,
is_mpi_op<Op, T>(), is_mpi_datatype<T>());
}
template<typename T, typename Op>
void
reduce(const communicator& comm, const T& in_value, T& out_value, Op op,
int root)
{
if (comm.rank() == root)
detail::reduce_impl(comm, &in_value, 1, &out_value, op, root,
is_mpi_op<Op, T>(), is_mpi_datatype<T>());
else
detail::reduce_impl(comm, &in_value, 1, op, root,
is_mpi_op<Op, T>(), is_mpi_datatype<T>());
}
template<typename T, typename Op>
void reduce(const communicator& comm, const T& in_value, Op op, int root)
{
BOOST_ASSERT(comm.rank() != root);
detail::reduce_impl(comm, &in_value, 1, op, root,
is_mpi_op<Op, T>(), is_mpi_datatype<T>());
}
} } // end namespace boost::mpi
#endif // BOOST_MPI_REDUCE_HPP
|