distribute.hpp 50.5 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
#ifndef _DISTRIBUTE_HPP_
#define _DISTRIBUTE_HPP_

/**
 * \brief Distributed distribute algorithm
 * \file
 *
 * \author Quentin Khan
 */

#include "inria/algorithm/distributed/mpi.hpp"
#include "inria/io.hpp"
#include "inria/utils.hpp"
#include "inria/meta.hpp"

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
#include <numeric>
#include <iterator>
#include <vector>

/**
 * \brief Inria namespace
 *
 */
namespace inria {

namespace details {

/**
 * \brief A non owning memory chunk of an array.
 *
 * \tparam T Type of the array's elements
 * \tparam Int Size type
 */
template<class T, class Int>
struct chunk_t : std::pair<T*,Int> {
    using std::pair<T*,Int>::pair;

    /**
     * \brief Chunk element count
     */
    Int size() const noexcept {
        return this->second;
    }

    /**
     * \brief Increase chunk size
     */
    void expand(Int value = 1) {
        this->second += value;
    }

    /**
     * \brief Reseat chunk underlying pointer
     */
    void reset(T* addr) {
        this->first = addr;
    }

    /**
     * \brief Change span size
     *
     * \param value New chunk size
     */
    void resize(Int value) {
        this->second = value;
    }

    /**
     * \brief Get underlying data
     */
    T* data() noexcept { return this->first; }
    /** \copydoc data */
    const T* data() const noexcept { return this->first; }

    /**
     * \brief Get begin iterator
     */
    T* begin() noexcept { return this->first; }
    /** \copydoc begin */
    const T* begin() const noexcept { return this->first; }
    /** \copydoc begin */
    const T* cbegin() const noexcept { return this->first; }

    /**
     * \brief Get end iterator
     */
    T* end() noexcept { return std::next(this->first, this->second); }
    /** \copydoc end */
    const T* end() const noexcept { return std::next(this->first, this->second); }
    /** \copydoc end */
    const T* cend() const noexcept { return std::next(this->first, this->second); }

};

/**
 * \brief Implements inter process distribution main steps.
 *
 * Distribution is done in 3 steps:
 *
 *   - Local data initialisation: gather information about the environment, how
 *     distribution is going to be decided, to which process.
 *   - Remote data and communication setup: decide which piece of data send to
 *     which process, asks other processes what data they will send.
 *   - Data communication: send and receive data.
 *
 * \tparam T Type of the objects to send
 * \tparam ForwardIt Local range iterator
 * \tparam Distribution Distribution configuration.
 * \parblock
 * An instance of Distribution must provide an call operator taking a `const T&`
 * and returning the rank of the target process for that object.
 *
 * ~~~{.cpp}
 * int Distribution::operator()(const T&);
 * ~~~
 * \endparblock
 */
template<class T, class ForwardIt, class Distribution>
struct distribution_state {

    distribution_state()                                     = delete;
    distribution_state(const distribution_state&)            = delete;
    distribution_state(distribution_state&&)                 = delete;
    distribution_state& operator=(const distribution_state&) = delete;
    distribution_state& operator=(distribution_state&&)      = delete;

    // MPI configuration

    /// MPI environment configuration
    inria::mpi_config conf;
    /// Datatype management
    inria::mpi::datatype_commit_guard datatype_guard;
    /// MPI communication tag for successful comms
    const int ok_tag;

    /// Distribution object
    Distribution distrib;

    // Communication data

    /**
     * \brief Copy of the range to send
     *
     * One buffer per target process
     */
    std::vector<std::vector<T>> send_buffers;
    /**
     * \brief Reception buffer
     *
     * Only one buffer, the communication sizes are expected to be
     * known. Therefore the reception offsets can be computed and written
     * directly in the final buffer.
     */
    std::vector<T> recv_buffer;

    /// Total reception size, to initialise #recv_buffer
    std::uint64_t recv_size;

    /// View type that holds pointer to data and element count
    using chunk_t = details::chunk_t<T, std::uint64_t>;
    /**
     * \brief Send information buffer
     *
     * Contains the 'send' information for the data to be sent to each
     * process. Each process receives the size of the data
     * communication. #send_buffers is not used directly because MPI needs
     * pointers to existing storage to send data.
     */
    std::vector<chunk_t> send_chunks;
    /**
     * \brief Receive information buffer
     *
     * Holds the size and target pointer to the span of recv_buffer that will
     * receive data from each process.
     */
    std::vector<chunk_t> recv_chunks;

    /**
     * \brief Intialise local data and setup environment
     *
     * On construction, the distribution_state performs the following:
     *
     *   - setup the MPI environment: a datatype is created is the configuration
     *     specifies MPI_DATATYPE_NULL.
     *   - copy the input range to the send buffer
     *   - compute the local range weightn, the global range weight, the
     *     cumulated range weight up to local range (excluded) and the target
     *     cumulated weight for all process.
     *
     * The final distribution is computed from the proportions
     * argument. Proportion must expose an operator[rank] that returns the
     * proportion for the process of given rank. The proportion maximum is
     *
     *     sum(0 <= i < nb_proc, propotions[i]);
     *
     * \param conf_ MPI environment configuration
     * \param first Iterator to local range first element
     * \param last  Local range end iterator
     * \param distribution Distribution object
     * \parblock
     *
     * Must provide a call operator returning the rank of the process the given
     * element belongs to.
     *
     * ~~~{.cpp}
     * int Distribution::operator()(const T&);
     * ~~~
     * \endparblock
     */
    distribution_state(inria::mpi_config conf_, ForwardIt first, ForwardIt last, Distribution& distribution)
        : conf(conf_),
          datatype_guard(mpi::create_datatype_if_null<T>(conf.datatype)),
          ok_tag(conf.base_tag + 153),
          distrib(distribution),
          send_buffers(conf.comm.size()),
          send_chunks(conf.comm.size(), {nullptr, 0}),
          recv_chunks(conf.comm.size(), {nullptr, 0})
    {
        setup_send_data(first, last);
        setup_recv_data();
    }


    /**
     * \brief Setup send data
     *
     * Fills the #send_buffers and sets the #send_chunks spans. Each element of
     * the range is passed to the distribution and inserted in the relevant
     * send_buffer.
     *
     * \param first Iterator to local range first element
     * \param last  Local range end iterator
     */
    void setup_send_data(ForwardIt first, ForwardIt last) {
        // Copy each element to the corresponding send buffer
        while(first != last) {
            std::uint64_t target_rank = this->distrib(*first);
            assert(target_rank < this->send_buffers.size());
            this->send_buffers[target_rank].push_back(*first);
            ++first;
        }
        // Setup span views
244
        for(std::size_t i = 0; i < this->send_buffers.size(); ++i) {
245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498
            send_chunks[i].resize(this->send_buffers[i].size());
            send_chunks[i].reset(this->send_buffers[i].data());
        }
    }


    /**
     * \brief Setup send data and prepare reception buffer.
     *
     * All processes communicate to all processes the amount of data they will
     * send. The reception buffer size is computed and sliced to simultaneously
     * receive from several processes at the same time.
     */
    void setup_recv_data() {
        using std::begin;
        using std::end;
        // Setup recv data
        //
        // Send/receive future comms metadata to each other process j:
        //    - send the number of elements to send to j
        //    - receive the number of elements to receive
        //
        // Create MPI datatype [****,    ]
        //              uint64_t ^     ^ empty space to reach total size of chunk_t
        MPI_Datatype mpi_chunk_size;
        MPI_Type_create_resized(MPI_UINT64_T, 0, sizeof(chunk_t), &mpi_chunk_size);
        MPI_Type_commit(&mpi_chunk_size);
        // Actual communication round
        conf.comm.alltoall(&(this->send_chunks[0].second), 1, mpi_chunk_size,
                           &(this->recv_chunks[0].second), 1, mpi_chunk_size);
        MPI_Type_free(&mpi_chunk_size);

        // Compute reception buffer size
        this->recv_size = std::accumulate(
            begin(this->recv_chunks), end(this->recv_chunks), 0ul,
            [](const std::uint64_t& res, const chunk_t& chunk) {
                return res + chunk.size();
            });
        recv_buffer.resize(recv_size);
        // Set the receive chunks start pointer
        recv_chunks[0].first = recv_buffer.data();
        std::partial_sum(begin(recv_chunks), end(recv_chunks),
                         begin(recv_chunks),
                         [](chunk_t& a, const chunk_t& b) {
                             return chunk_t{a.end(), b.size()};
                         });
    }


    /**
     * \brief Send and receive data to all processes individually.
     */
    void do_comms() {
        const int proc_count = conf.comm.size();
        const int rank = conf.comm.rank();

        std::vector<mpi::request> requests;
        requests.reserve(2 * proc_count);

        for(int i = 0; i < proc_count; ++i) {
            if(0 == recv_chunks[i].size() || i == rank) {
                continue; // Skip empty or local comm
            }
            requests.push_back(conf.comm.irecv(recv_chunks[i].data(),
                                               static_cast<int>(recv_chunks[i].size()),
                                               conf.datatype, i, this->ok_tag));
        }
        for(int i = 0; i < proc_count; ++i) {
            if(0 == send_chunks[i].size() || i == rank) {
                continue; // Skip empty or local comm
            }
            requests.push_back(conf.comm.isend(send_chunks[i].data(),
                                               static_cast<int>(send_chunks[i].size()),
                                               conf.datatype, i, this->ok_tag));
        }
        // Copy the elements that are already on the current process
        std::copy(std::begin(send_chunks[rank]), std::end(send_chunks[rank]),
                  std::begin(recv_chunks[rank]));

        mpi::request::waitall(requests.size(), requests.data());
    }
};


/**
 * \brief Helper function to deduce the distribution_state type from arguments.
 *
 * \warning Since the copy constructor of distribution_state is deleted, this
 * method cannot be used to create a distribution state. C++17 lifts this
 * restriction with garanteed copy elision.
 *
 * See documentation of details::distribution_state.
 *
 * \param conf  MPI environment configuration
 * \param first Iterator to local range first element
 * \param last  Local range end iterator
 * \param distrib Distribution object
 *
 * \tparam T Type of the objects to send
 * \tparam ForwardIt Local range iterator
 * \tparam Distribution Distribution configuration.
 */
template<class ForwardIt, class Distribution, class T = typename std::iterator_traits<ForwardIt>::value_type>
details::distribution_state<T, ForwardIt, Distribution>
create_state(inria::mpi_config conf, ForwardIt first, ForwardIt last, Distribution&& distrib) {
    return {conf, first, last, distrib};
}

} // close namespace [inria]::details

/**
 * \brief Function object returning 1 for any argument
 */
struct unit_weight {
    /**
     * \brief Call operator, returns 1
     *
     * \tparam Ts Parameter types
     */
    template<class... Ts>
    double operator()(const Ts& ...) {return 1;}
};

/**
 * \brief Indexable object returning 1 for any argument
 */
struct uniform_proportions {
    /**
     * \brief Index operator, returns 1
     *
     * \tparam T Parameter type
     */
    template<class T>
    double operator[](const T&) {return 1;}
};

/**
 * \brief Callable object for custom proportional distribution.
 *
 * Implements the `Distribute` concept for the inria::distribute function. Upon
 * creation, this object scans the target range to compute the final ideal
 * weight distribution. Each subsequent call to `operator()` adds the given
 * element weight to an accummulator and returns the element selected target
 * process rank.
 *
 * \tparam Weight Weight function object used to find out elements distribution.
 *
 * \see details::distribution_state
 *
 * ~~~{.cpp}
 * std::vector<T> objs = {...};
 * // For 3 processes, 16.67%, 33.33%, 50%
 * std::array<int, 3> proportions = {100,200,300}
 * proportional_distribution<> distrib(conf, objs, proportions);
 * distribute(conf, objs, distrib);
 * ~~~
 */
template<class Weight = unit_weight>
struct proportional_distribution {
    /// Weight function object
    Weight weight;
    /// Local range weight
    std::uint64_t local_weight = 0;
    /// Range weight of all processes up to current one (excluded)
    std::uint64_t cumul_weight = 0;
    /// Global range weight
    std::uint64_t total_weight = 0;
    /// Ideal cumulated weight for each process
    std::vector<std::uint64_t> target_weight;

    /// Target rank returned by operator()
    std::uint64_t target_rank = 0;
    /// Number of processes
    int proc_count = 0;

    /**
     * \brief Create a propotional distribution for a distributed range
     *
     * Scans a distributed range to compute the global weight. Succesive calls
     * to operator() yield the process rank of given elements.
     *
     * \param conf  MPI configuration object.
     * \param first Iterator to local rank first element.
     * \param last  Iterator passed local rank last element.
     * \param proportions Array like object that holds proportions for each object.
     * \param w_    Function object to get an element weight
     *
     * \tparam ElemIt Iterator type of the element range
     * \tparam Proportion
     * \parblock
     * Array like type
     *
     * Proportion must specify an operator[] that returns the final proportion
     * of the global range that given process must hold. The sum of the
     * proportions is used.
     * \endparblock
     */
    template<class ElemIt, class Proportion>
    proportional_distribution(
        mpi_config conf,
        ElemIt first,
        ElemIt last,
        Proportion&& proportions,
        Weight w_ = {}
    ) :
        weight(w_),
        target_weight(static_cast<std::uint64_t>(conf.comm.size())),
        proc_count{conf.comm.size()}
    {
        // Compute all the weights
        using T = typename std::iterator_traits<ElemIt>::value_type;
        auto add_weight = [&](std::uint64_t w, const T& e) {
            return w + get_weight(e);
        };
        // Compute local weight
        local_weight = std::accumulate(first, last, 0ul, add_weight);
        // Sum weights of previous and current processes
        conf.comm.scan(&local_weight, &cumul_weight, 1, MPI_UINT64_T, MPI_SUM);
        // Last process broadcasts its cumulated weight
        total_weight = cumul_weight;
        conf.comm.bcast(&total_weight, 1, MPI_UINT64_T, proc_count-1);
        // Correct cumulated weight to not include local range weight
        cumul_weight -= local_weight;

        using proportion_t = typename std::decay<decltype(proportions[0])>::type;
        proportion_t total_distri = 0;
        for(std::uint64_t i = 0; i < target_weight.size(); ++i) {
            total_distri += proportions[i];
        }

        proportion_t total_weight_p_t = static_cast<proportion_t>(total_weight);

        target_weight[0] = static_cast<std::uint64_t>(total_weight_p_t * proportions[0] / total_distri);
        for(std::uint64_t i = 1; i < target_weight.size(); ++i) {
            target_weight[i] = target_weight[i-1] +
                static_cast<std::uint64_t>(total_weight_p_t * proportions[i] / total_distri);
        }

        auto remaining_weight = total_weight - target_weight[proc_count-1];
        if(remaining_weight > 0) {
            auto remaining_div = remaining_weight / proc_count;
            auto remaining_mod = remaining_weight % proc_count;
            for(std::uint64_t i = 0; i < target_weight.size(); ++i) {
                target_weight[i] += (remaining_div * (i+1)) + std::min(i+1, remaining_mod);
            }
        }
    }

    /**
     * \brief Create a propotional distribution for a distributed range
     *
     * Scans a distributed range to compute the global weight. Succesive calls
     * to operator() yield the process rank of given elements.
     *
     * \param conf MPI configuration object.
     * \param r    Local range
     * \param p    Array like object that holds proportions for each object.
     * \param w    Function object to get an element weight
     *
     * \tparam ElemIt Iterator type of the element range
     * \tparam Proportion
     * \parblock
     * Array like type
     *
     * Proportion must specify an operator[] that returns the final proportion
     * of the global range that given process must hold. The sum of the
     * proportions is used.
     * \endparblock
     */
    template<class Range, class Proportion>
    proportional_distribution(mpi_config conf, const Range& r, Proportion&& p, Weight w = {})
        : proportional_distribution(conf, std::begin(r), std::end(r), std::forward<Proportion>(p), w)
    {}

    /**
     * \brief Compute target process of an element
     *
     * The object's accumulator #cumul_weight is increased by `e`'s weight at
     * each call. The target rank is then chosen according to #cumul_weight.
     *
     * \param e Element of which to compute the target
     *
     * \tparam T Type of `e`
     */
    template<class T>
    std::uint64_t operator()(const T& e) {
        const std::uint64_t proc_count_minus_one = proc_count-1;
        while((target_rank < proc_count_minus_one) &&
              (this->cumul_weight >= target_weight[target_rank])) {
            ++target_rank;
        }
        this->cumul_weight += get_weight(e);
        return target_rank;
    }

    /**
     * \brief Helper to call #weight
     *
     * This mainly factorizes the static_cast of `weight(e)`
     *
     * \param e Element to get the weight of
     *
     * \tparam T Type of `e`
     */
    template<class T> std::uint64_t get_weight(const T& e) {
        return static_cast<std::uint64_t>(weight(e));
    }
};

/**
 * \brief Callable object for uniform distribution.
 *
 * Implements the `Distribute` concept for the inria::distribute function.
 *
 * \tparam Weight Weight function object used to find out elements distribution.
 *
 * \see details::proportional_distribution
 * \see details::distribution_state
 *
 * ~~~[cpp]
 * std::vector<T> objs = {...};
 * distribute(conf, objs, uniform_distribution{conf, objs});
 * ~~~
 */
template<class Weight = unit_weight>
struct uniform_distribution : proportional_distribution<Weight> {

    /**
     * \brief Constructor from range iterators
     *
     * \param conf  MPI configuration object.
     * \param first Iterator to local rank first element.
     * \param last  Iterator passed local rank last element.
     * \param w     Function object to get an element weight
     *
     * \tparam ElemIt Iterator type
     */
    template<class ElemIt>
    uniform_distribution(mpi_config conf, ElemIt first, ElemIt last, Weight w = {})
        : proportional_distribution<Weight>(conf, first, last, uniform_proportions{}, w)
    {}

    /**
     * \brief Constructor from range
     *
     * \param conf MPI configuration object.
     * \param r    Local range
     * \param w    Function object to get an element weight
     *
     * \tparam Range Range type
     */
    template<class Range>
    uniform_distribution(mpi_config conf, const Range& r, Weight w = {})
        : proportional_distribution<Weight>(conf, r, uniform_proportions{}, w)
    {}
};

/**
 * \brief Redistribute global range among processes
 *
 * The global range is shared among the processes and must be
 * redistributed. Each process part of the global range is called the local
 * range.
 *
 * The order of the elements is kept. No guarantee is made for elements that
 * compare equal.
 *
 * \note Current implementation keeps local order and if two elements compare
 * equal, the one that was originally on the lowest ranked process is placed
 * first.
 *
 * The elements are copied to the output range. The original local range is kept
 * untouched.
 *
 * \warning It is the developper responsibility to provide an output range large
 * enough to store the results. A check is done using `out_last - out_first`
 *
 * \param conf  MPI environment
 * \param first Local range begin iterator
 * \param last  Local range end iterator
 * \param out_first Output local range begin iterator
 * \param out_last  Output local range end iterator
 * \param distrib Distribution object
 *
 * \tparam ForwardIt Forward input iterator
 * \tparam ForwardOutputIterator Forward output iterator
 * \tparam Distribution Distribution object type
 * \parblock
 * Must provide a call operator returning the rank of the process the given
 * element belongs to.
 *
 * ~~~{.cpp}
 * int Distribution::operator()(const T&);
 * ~~~
 * \endparblock
 *
 * \return An iterator to the new end of the output range.
 *
 * \exception std::runtime_error is thrown if the output range is too small.
 *
 * Example:
 *
 * ~~~{.cpp}
 * // Distribute data from process 0 to the others
 *
 * // MPI communicator
 * auto world = inria::mpi::communicator::world();
 *
 * // All processes have a large enough buffer
 * int output[100] = {};
 * // Input is an empty vector, no allocation is done
 * std::vector<int> input{};
 *
 * // Process 0 fills input with data
 * if(world.rank() == 0) {
 *     input.resize(100);
 *     std::iota(begin(input), end(input), 0);
 * }
 *
 * // Distribution object, keeps all even elements on 0, all odd elements on 1
 * struct dist {
 *     int operator()(const int& i) {
 *         return i % 2;
 *     }
 * };
 *
 * try {
 *     distribute(world, begin(input), end(input), output, output + 100, dist{});
 * } catch(std::out_of_range& e) {
 *     std::cerr << e.what() << '\n';
 * }
 * ~~~
 */
template<class ForwardIt, class ForwardOutputIterator, class Distribution,
         #ifndef DOXYGEN_DOCUMENTATION
         class T = typename std::iterator_traits<ForwardIt>::value_type,
         class = disable_if_t<is_range<ForwardIt>::value>,
         class = enable_if_t<is_algo_distribution<typename std::remove_reference<Distribution>::type, T>::value>
         #endif
         >
ForwardOutputIterator distribute(
    inria::mpi_config conf,
    ForwardIt first, ForwardIt last,
    ForwardOutputIterator out_first, ForwardOutputIterator out_last,
    Distribution&& distrib)
{
    using state_t = decltype(details::create_state(conf, first, last, distrib));
    state_t state{conf, first, last, distrib};

    // Once comms are set up, check whether the output range is big enough.
    std::uint64_t output_capacity = std::distance(out_first, out_last);
    char output_large_enough = output_capacity >= state.recv_size;
    char global_large_enough = 1;
    conf.comm.allreduce(&output_large_enough, &global_large_enough,
                        1, MPI_CHAR, MPI_LAND);
    if(! global_large_enough) {
        if(! output_large_enough) {
            throw std::runtime_error("Process output range is too small.");
        } else {
            throw std::runtime_error("Other process output range is too small.");
        }
    }

    state.do_comms();

    return std::copy(std::begin(state.recv_buffer), std::end(state.recv_buffer),
                     out_first);
}

/**
 * \brief Evenly redistribute global range among processes
 *
 * The global range is shared among the processes and must be evenly
 * redistributed. Each process part of the global range is called the local
 * range.
 *
 * The order of the elements is kept. No guarantee is made for elements that
 * compare equal.
 *
 * \note Current implementation keeps local order and if two elements compare
 * equal, the one that was originally on the lowest ranked process is placed
 * first.
 *
 * The elements are copied to the output range. The original local range is kept
 * untouched.
 *
 * \warning It is the developper responsibility to provide an output range large
 * enough to store the results. A check is done using `out_last - out_first`
 *
 * \param conf  MPI environment
 * \param first Local range begin iterator
 * \param last  Local range end iterator
 * \param out_first Output local range begin iterator
 * \param out_last  Output local range end iterator
 *
 * \tparam ForwardIt Forward input iterator
 * \tparam OutputIterator Forward output iterator
 *
 * \return An iterator to the new end of the output range.
 *
 * \exception std::runtime_error is thrown if the output range is too small.
 *
 * Example:
 *
 * ~~~{.cpp}
 * // Distribute data from process 0 to the others
 *
 * // MPI communicator
 * auto world = inria::mpi::communicator::world();
 *
 * // All processes have a large enough buffer
 * int output[100] = {};
 * // Input is an empty vector, no allocation is done
 * std::vector<int> input{};
 *
 * // Process 0 fills input with data
 * if(world.rank() == 0) {
 *     input.resize(100);
 *     std::iota(begin(input), end(input), 0);
 * }
 *
 * try {
 *     distribute(world, begin(input), end(input), output, output + 100);
 * } catch(std::out_of_range& e) {
 *     std::cerr << e.what() << '\n';
 * }
 * ~~~

 */
template<class ForwardIt, class OutputIterator,
         #ifndef DOXYGEN_DOCUMENTATION
         class T = typename std::iterator_traits<ForwardIt>::value_type,
         class = disable_if_t<is_range<ForwardIt>::value>
         #endif
         >
OutputIterator distribute(
    inria::mpi_config conf, ForwardIt first, ForwardIt last,
    OutputIterator out_first, OutputIterator out_last)
{
    return distribute(conf, first, last, out_first, out_last, uniform_distribution<>{conf, first, last});
}

/**
 * \brief Redistribute global range among processes
 *
 * The global range is shared among the processes and must be
 * redistributed. Each process part of the global range is called the local
 * range.
 *
 * The order of the elements is kept. No guarantee is made for elements that
 * compare equal.
 *
 * \note Current implementation keeps local order and if two elements compare
 * equal, the one that was originally on the lowest ranked process is placed
 * first.
 *
 * The elements are copied to the output range. The original local range is kept
 * untouched.
 *
 * \warning It is the developper responsibility to provide an output range large
 * enough to store the results. A check is done using `out_last - out_first`
 *
 * \param conf MPI environment
 * \param in   Input range
 * \param out_first Output local range begin iterator
 * \param out_last  Output local range end iterator
 * \param distrib Distribution object
 *
 * \tparam InRange Input range
 * \parblock
 * Must satisfy the usual range requirements, tested using inria::is_range.
 * \endparblock
 * \tparam OutputIterator Forward output iterator
 * \tparam Distribution Distribution object type
 * \parblock
 * Must provide a call operator returning the rank of the process the given
 * element belongs to.
 *
 * ~~~{.cpp}
 * int Distribution::operator()(const T&);
 * ~~~
 * \endparblock
 *
 * \return An iterator to the new end of the output range.
 *
 * \exception std::runtime_error is thrown if the output range is too small.
 */
template<class InRange, class OutputIterator, class Distribution,
         #ifndef DOXYGEN_DOCUMENTATION
         class = enable_if_t<is_range<typename std::remove_reference<InRange>::type>::value>,
         class T = range_element_t<typename std::remove_reference<InRange>::type>,
         class = disable_if_t<is_range<OutputIterator>::value>,
         class = enable_if_t<is_algo_distribution<typename std::remove_reference<Distribution>::type, T>::value>
         #endif
         >
OutputIterator distribute(
    inria::mpi_config conf, InRange&& in,
    OutputIterator out_first, OutputIterator out_last,
    Distribution&& distrib)
{
    using std::begin;
    using std::end;
    return distribute(conf, begin(in), end(in), out_first, out_last, std::forward<Distribution>(distrib));
}

/**
 * \brief Evenly redistribute global range among processes
 *
 * The global range is shared among the processes and must be
 * redistributed. Each process part of the global range is called the local
 * range.
 *
 * The order of the elements is kept. No guarantee is made for elements that
 * compare equal.
 *
 * \note Current implementation keeps local order and if two elements compare
 * equal, the one that was originally on the lowest ranked process is placed
 * first.
 *
 * The elements are copied to the output range. The original local range is kept
 * untouched.
 *
 * \warning It is the developper responsibility to provide an output range large
 * enough to store the results. A check is done using `out_last - out_first`
 *
 * \param conf MPI environment
 * \param in   Input range
 * \param out_first Output local range begin iterator
 * \param out_last  Output local range end iterator
 *
 * \tparam InRange Input range
 * \parblock
 * Must satisfy the usual range requirements, tested using inria::is_range.
 * \endparblock
 * \tparam OutputIterator Forward output iterator
 *
 * \return An iterator to the new end of the output range.
 *
 * \exception std::runtime_error is thrown if the output range is too small.
 */
template<class InRange, class OutputIterator,
         #ifndef DOXYGEN_DOCUMENTATION
         class = enable_if_t<is_range<typename std::remove_reference<InRange>::type>::value>,
         class T = range_element_t<typename std::remove_reference<InRange>::type>,
         class = disable_if_t<is_range<OutputIterator>::value>
         #endif
         >
OutputIterator distribute(
    inria::mpi_config conf, InRange&& in,
    OutputIterator out_first, OutputIterator out_last)
{
    return distribute(conf, in, out_first, out_last, uniform_distribution<>{conf, in});
}

/**
 * \brief Redistribute global range among processes
 *
 * The global range is shared among the processes and must be
 * redistributed. Each process part of the global range is called the local
 * range.
 *
 * The order of the elements is kept. No guarantee is made for elements that
 * compare equal.
 *
 * \note Current implementation keeps local order and if two elements compare
 * equal, the one that was originally on the lowest ranked process is placed
 * first.
 *
 * The elements are copied to the output range. The original local range is kept
 * untouched.
 *
 * \warning It is the developper responsibility to provide an output range large
 * enough to store the results. A check is done using `out_last - out_first`
 *
 * \param conf  MPI environment
 * \param first Local range begin iterator
 * \param last  Local range end iterator
 * \param out   Output range
 * \param distrib Distribution object
 *
 * \tparam ForwardIt Forward input iterator
 * \tparam AssignableOutRange Output range
 * \parblock
 * Must satisfy the usual range requirements, tested using inria::is_range and
 * the provide an `assign` method that sets the range from an iterator defined
 * range.
 * \endparblock
 * \tparam Distribution Distribution object type
 * \parblock
 * Must provide a call operator returning the rank of the process the given
 * element belongs to.
 *
 * ~~~{.cpp}
 * int Distribution::operator()(const T&);
 * ~~~
 * \endparblock
 */
template<class ForwardIt, class AssignableOutRange, class Distribution,
         #ifndef DOXYGEN_DOCUMENTATION
         class T = typename std::iterator_traits<ForwardIt>::value_type,
         class = enable_if_t<is_range<AssignableOutRange>::value>,
         class = enable_if_t<is_assignable<AssignableOutRange, ForwardIt, ForwardIt>::value>,
         class = enable_if_t<is_algo_distribution<typename std::remove_reference<Distribution>::type, T>::value>
         #endif
         >
void distribute(inria::mpi_config conf,
                ForwardIt first, ForwardIt last,
                AssignableOutRange& out,
                Distribution&& distrib)
{
    using std::begin;
    using std::end;
    using state_t = decltype(details::create_state(conf, first, last, distrib));
    state_t state{conf, first, last, distrib};
    state.do_comms();
    out.assign(begin(state.recv_buffer), end(state.recv_buffer));
}

/**
 * \brief Evenly redistribute global range among processes
 *
 * The global range is shared among the processes and must be
 * redistributed. Each process part of the global range is called the local
 * range.
 *
 * The order of the elements is kept. No guarantee is made for elements that
 * compare equal.
 *
 * \note Current implementation keeps local order and if two elements compare
 * equal, the one that was originally on the lowest ranked process is placed
 * first.
 *
 * The elements are copied to the output range. The original local range is kept
 * untouched.
 *
 * \warning It is the developper responsibility to provide an output range large
 * enough to store the results. A check is done using `out_last - out_first`
 *
 * \param conf  MPI environment
 * \param first Local range begin iterator
 * \param last  Local range end iterator
 * \param out   Output range
 *
 * \tparam ForwardIt Forward input iterator
 * \tparam AssignableOutRange Output range
 * \parblock
 * Must satisfy the usual range requirements, tested using inria::is_range and
 * the provide an `assign` method that sets the range from an iterator defined
 * range.
 * \endparblock
 *
 * Example:
 *
 * ~~~{.cpp}
 * // Distribute data from process 0 to the others
 *
 * // MPI communicator
 * auto world = inria::mpi::communicator::world();
 *
 * // All processes have an assignable output buffer
 * std::vector<int> output{};
  * // Input is an empty vector, no allocation is done
 * std::vector<int> input{};
 *
 * // Process 0 fills input with data
 * if(world.rank() == 0) {
 *     input.resize(100);
 *     std::iota(begin(input), end(input), 0);
 * }
 *
 * distribute(world, begin(input), end(input), output);
 */
template<class ForwardIt, class AssignableOutRange,
         #ifndef DOXYGEN_DOCUMENTATION
         class T = typename std::iterator_traits<ForwardIt>::value_type,
         class = enable_if_t<is_range<AssignableOutRange>::value>,
         class = enable_if_t<is_assignable<AssignableOutRange, ForwardIt, ForwardIt>::value>
         #endif
         >
void distribute(inria::mpi_config conf,
                ForwardIt first, ForwardIt last,
                AssignableOutRange& out)
{
    distribute(conf, first, last, out, uniform_distribution<>{conf, first, last});
}

/**
 * \brief Redistribute global range among processes
 *
 * The global range is shared among the processes and must be
 * redistributed. Each process part of the global range is called the local
 * range.
 *
 * The order of the elements is kept. No guarantee is made for elements that
 * compare equal.
 *
 * \note Current implementation keeps local order and if two elements compare
 * equal, the one that was originally on the lowest ranked process is placed
 * first.
 *
 * The elements are copied to the output range. The original local range is kept
 * untouched.
 *
 * \warning It is the developper responsibility to provide an output range large
 * enough to store the results. A check is done using `out_last - out_first`
 *
 * \param conf MPI environment
 * \param in   Input range
 * \param out  Output range
 * \param distrib Distribution object
 *
 * \tparam InRange Input range
 * \parblock
 *
 * Must satisfy the usual range requirements, tested using inria::is_range.
 * \endparblock
 * \tparam AssignableOutRange Output range
 * \parblock
 *
 * Must satisfy the usual range requirements, tested using inria::is_range and
 * the provide an `assign` method that sets the range from an iterator defined
 * range.
 * \endparblock
 * \tparam Distribution Distribution object type
 * \parblock
 *
 * Must provide a call operator returning the rank of the process the given
 * element belongs to.
 *
 * ~~~{.cpp}
 * int Distribution::operator()(const T&);
 * ~~~
 * \endparblock
 */
template<class InRange, class AssignableOutRange, class Distribution,
         #ifndef DOXYGEN_DOCUMENTATION
         class = enable_if_t<is_range<typename std::remove_reference<InRange>::type>::value>,
         class T = range_element_t<typename std::remove_reference<InRange>::type>,
         class = enable_if_t<is_range<AssignableOutRange>::value>,
         class = enable_if_t<is_assignable<AssignableOutRange, T*, T*>::value>,
         class = enable_if_t<is_algo_distribution<typename std::decay<Distribution>::type, T>::value>
         #endif
         >
void distribute(inria::mpi_config conf,
                InRange&& in,
                AssignableOutRange& out,
                Distribution&& distrib)
{
    using std::begin;
    using std::end;
    distribute(conf, begin(in), end(in), out, distrib);
}

/**
 * \brief Redistribute global range among processes
 *
 * The global range is shared among the processes and must be
 * redistributed. Each process part of the global range is called the local
 * range.
 *
 * The order of the elements is kept. No guarantee is made for elements that
 * compare equal.
 *
 * \note Current implementation keeps local order and if two elements compare
 * equal, the one that was originally on the lowest ranked process is placed
 * first.
 *
 * The elements are copied to the output range. The original local range is kept
 * untouched.
 *
 * \warning It is the developper responsibility to provide an output range large
 * enough to store the results. A check is done using `out_last - out_first`
 *
 * \param conf MPI environment
 * \param in   Input range
 * \param out  Output range
 *
 * \tparam InRange Input range
 * \parblock
 *
 * Must satisfy the usual range requirements, tested using inria::is_range.
 * \endparblock
 * \tparam AssignableOutRange Output range
 * \parblock
 *
 * Must satisfy the usual range requirements, tested using inria::is_range and
 * the provide an `assign` method that sets the range from an iterator defined
 * range.
 * \endparblock
 */
template<class InRange, class AssignableOutRange,
         #ifndef DOXYGEN_DOCUMENTATION
         class = enable_if_t<is_range<typename std::remove_reference<InRange>::type>::value>,
         class T = range_element_t<typename std::remove_reference<InRange>::type>,
         class = enable_if_t<is_range<AssignableOutRange>::value>,
         class = enable_if_t<is_assignable<AssignableOutRange, T*, T*>::value>
         #endif
         >
void distribute(inria::mpi_config conf,
                InRange&& in,
                AssignableOutRange& out)
{
    using std::begin;
    using std::end;
    distribute(conf, begin(in), end(in), out, uniform_distribution<>{conf, begin(in), end(in)});
}

/**
 * \brief Redistribute global range among processes
 *
 * The global range is shared among the processes and must be
 * redistributed. Each process part of the global range is called the local
 * range.
 *
 * The order of the elements is kept. No guarantee is made for elements that
 * compare equal.
 *
 * \note Current implementation keeps local order and if two elements compare
 * equal, the one that was originally on the lowest ranked process is placed
 * first.
 *
 * The elements are copied to the output range. The original local range is kept
 * untouched.
 *
 * \warning It is the developper responsibility to provide an output range large
 * enough to store the results. A check is done using `out_last - out_first`
 *
 * \param conf  MPI environment
 * \param inout Input and output range
 * \param distrib Distribution object
 *
 * \tparam AssignableInOutRange Input and output range
 * \parblock
 *
 * Must satisfy the usual range requirements, tested using inria::is_range and
 * the provide an `assign` method that sets the range from an iterator defined
 * range.
 * \endparblock
 * \tparam Distribution Distribution object type
 * \parblock
 *
 * Must provide a call operator returning the rank of the process the given
 * element belongs to.
 *
 * ~~~{.cpp}
 * int Distribution::operator()(const T&);
 * ~~~
 * \endparblock
 */
template<class AssignableInOutRange, class Distribution,
         #ifndef DOXYGEN_DOCUMENTATION
         class = enable_if_t<is_range<AssignableInOutRange>::value>,
         class T = range_element_t<AssignableInOutRange>,
         class = enable_if_t<is_assignable<AssignableInOutRange, T*, T*>::value>,
         class = enable_if_t<is_algo_distribution<typename std::remove_reference<Distribution>::type, T>::value>
         #endif
         >
void distribute(inria::mpi_config conf,
                AssignableInOutRange& inout,
                Distribution&& distrib)
{
    distribute(conf, inout, inout, distrib);
}

/**
 * \brief Evenly redistribute global range among processes
 *
 * The global range is shared among the processes and must be
 * redistributed. Each process part of the global range is called the local
 * range.
 *
 * The order of the elements is kept. No guarantee is made for elements that
 * compare equal.
 *
 * \note Current implementation keeps local order and if two elements compare
 * equal, the one that was originally on the lowest ranked process is placed
 * first.
 *
 * The elements are copied to the output range. The original local range is kept
 * untouched.
 *
 * \warning It is the developper responsibility to provide an output range large
 * enough to store the results. A check is done using `out_last - out_first`
 *
 * \param conf  MPI environment
 * \param inout Input and output range
 *
 * \tparam AssignableInOutRange Input and output range
 * \parblock
 *
 * Must satisfy the usual range requirements, tested using inria::is_range and
 * the provide an `assign` method that sets the range from an iterator defined
 * range.
 * \endparblock
 */
template<class AssignableInOutRange,
         #ifndef DOXYGEN_DOCUMENTATION
         class = enable_if_t<is_range<AssignableInOutRange>::value>,
         class T = range_element_t<AssignableInOutRange>,
         class = enable_if_t<is_assignable<AssignableInOutRange, T*, T*>::value>
         #endif
         >
void distribute(inria::mpi_config conf,
                AssignableInOutRange& inout)
{
    distribute(conf, inout, inout, uniform_distribution<>{conf, begin(inout), end(inout)});
}

/**
 * \brief Redistribute global range among processes
 *
 * The global range is shared among the processes and must be
 * redistributed. Each process part of the global range is called the local
 * range.
 *
 * The order of the elements is kept. No guarantee is made for elements that
 * compare equal.
 *
 * \note Current implementation keeps local order and if two elements compare
 * equal, the one that was originally on the lowest ranked process is placed
 * first.
 *
 * The elements are copied to the output range. The original local range is kept
 * untouched.
 *
 * \warning It is the developper responsibility to provide an output range large
 * enough to store the results. A check is done using `out_last - out_first`
 *
 * \param conf  MPI environment
 * \param first Local range begin iterator
 * \param last  Local range end iterator
 * \param out   Output local range
 * \param distrib Distribution object
 *
 * \tparam ForwardIt Forward input iterator
 * \tparam NonAssignableOutRange Output range
 * \parblock
 *
 * Must satisfy the usual range requirements, tested using inria::is_range.
 * \endparblock
 * \tparam Distribution Distribution object type
 * \parblock
 * Must provide a call operator returning the rank of the process the given
 * element belongs to.
 *
 * ~~~{.cpp}
 * int Distribution::operator()(const T&);
 * ~~~
 * \endparblock
 *
 * \return An iterator to the new end of the output range.
 *
 * \exception std::runtime_error is thrown if the output range is too small.
 */
template<class ForwardIt, class NonAssignableOutRange, class Distribution,
         #ifndef DOXYGEN_DOCUMENTATION
         class T = typename std::iterator_traits<ForwardIt>::value_type,
         class = enable_if_t<is_range<NonAssignableOutRange>::value>,
         class = disable_if_t<is_assignable<NonAssignableOutRange, ForwardIt, ForwardIt>::value>,
         class = enable_if_t<is_algo_distribution<typename std::remove_reference<Distribution>::type, T>::value>,
         class AvoidTemplateRedeclarationError = void
         #endif
         >
auto distribute(inria::mpi_config conf,
                ForwardIt first, ForwardIt last,
                NonAssignableOutRange& out,
                Distribution&& distrib)
    -> decltype(std::begin(out))
{
    return distribute(conf, first, last, std::begin(out), std::end(out), distrib);
}

/**
 * \brief Evenly redistribute global range among processes
 *
 * The global range is shared among the processes and must be
 * redistributed. Each process part of the global range is called the local
 * range.
 *
 * The order of the elements is kept. No guarantee is made for elements that
 * compare equal.
 *
 * \note Current implementation keeps local order and if two elements compare
 * equal, the one that was originally on the lowest ranked process is placed
 * first.
 *
 * The elements are copied to the output range. The original local range is kept
 * untouched.
 *
 * \warning It is the developper responsibility to provide an output range large
 * enough to store the results. A check is done using `out_last - out_first`
 *
 * \param conf  MPI environment
 * \param first Local range begin iterator
 * \param last  Local range end iterator
 * \param out   Output local range
 *
 * \tparam ForwardIt Forward input iterator
 * \tparam NonAssignableOutRange Output range
 * \parblock
 *
 * Must satisfy the usual range requirements, tested using inria::is_range.
 * \endparblock
 *
 * \return An iterator to the new end of the output range.
 *
 * \exception std::runtime_error is thrown if the output range is too small.
 */
template<class ForwardIt, class NonAssignableOutRange,
         #ifndef DOXYGEN_DOCUMENTATION
         class T = typename std::iterator_traits<ForwardIt>::value_type,
         class = enable_if_t<is_range<NonAssignableOutRange>::value>,
         class = disable_if_t<is_assignable<NonAssignableOutRange, ForwardIt, ForwardIt>::value>,
         class AvoidTemplateRedeclarationError = void
         #endif
         >
auto distribute(inria::mpi_config conf,
                ForwardIt first, ForwardIt last,
                NonAssignableOutRange& out)
    -> decltype(std::begin(out))
{
    return distribute(conf, first, last, std::begin(out), std::end(out), uniform_distribution<>{conf, first, last});
}

/**
 * \brief Redistribute global range among processes
 *
 * The global range is shared among the processes and must be
 * redistributed. Each process part of the global range is called the local
 * range.
 *
 * The order of the elements is kept. No guarantee is made for elements that
 * compare equal.
 *
 * \note Current implementation keeps local order and if two elements compare
 * equal, the one that was originally on the lowest ranked process is placed
 * first.
 *
 * The elements are copied to the output range. The original local range is kept
 * untouched.
 *
 * \warning It is the developper responsibility to provide an output range large
 * enough to store the results. A check is done using `out_last - out_first`
 *
 * \param conf  MPI environment
 * \param in   Input range
 * \param out   Output local range
 * \param distrib Distribution object
 *
 * \tparam InRange Input range
 * \parblock
 *
 * Must satisfy the usual range requirements, tested using inria::is_range.
 * \endparblock
 * \tparam NonAssignableOutRange Output range
 * \parblock
 *
 * Must satisfy the usual range requirements, tested using inria::is_range.
 * \endparblock
 * \tparam Distribution Distribution object type
 * \parblock
 *
 * Must provide a call operator returning the rank of the process the given
 * element belongs to.
 *
 * ~~~{.cpp}
 * int Distribution::operator()(const T&);
 * ~~~
 * \endparblock
 *
 * \return An iterator to the new end of the output range.
 *
 * \exception std::runtime_error is thrown if the output range is too small.
 */
template<class InRange, class NonAssignableOutRange, class Distribution,
         #ifndef DOXYGEN_DOCUMENTATION
         class = enable_if_t<is_range<typename std::remove_reference<InRange>::type>::value>,
         class T = range_element_t<typename std::remove_reference<InRange>::type>,
         class = enable_if_t<is_range<NonAssignableOutRange>::value>,
         class = disable_if_t<is_assignable<NonAssignableOutRange, T*, T*>::value>,
         class = enable_if_t<is_algo_distribution<typename std::remove_reference<Distribution>::type, T>::value>,
         class AvoidTemplateRedeclarationError = void
         #endif
         >
auto distribute(inria::mpi_config conf,
                InRange&& in,
                NonAssignableOutRange& out,
                Distribution&& distrib)
    -> decltype(std::begin(out))
{
    return distribute(conf, std::begin(in), std::end(in), std::begin(out), std::end(out), distrib);
}

/**
 * \brief Evenly redistribute global range among processes
 *
 * The global range is shared among the processes and must be
 * redistributed. Each process part of the global range is called the local
 * range.
 *
 * The order of the elements is kept. No guarantee is made for elements that
 * compare equal.
 *
 * \note Current implementation keeps local order and if two elements compare
 * equal, the one that was originally on the lowest ranked process is placed
 * first.
 *
 * The elements are copied to the output range. The original local range is kept
 * untouched.
 *
 * \warning It is the developper responsibility to provide an output range large
 * enough to store the results. A check is done using `out_last - out_first`
 *
 * \param conf  MPI environment
 * \param in   Input range
 * \param out   Output local range
 *
 * \tparam InRange Input range
 * \parblock
 *
 * Must satisfy the usual range requirements, tested using inria::is_range.
 * \endparblock
 * \tparam NonAssignableOutRange Output range
 * \parblock
 *
 * Must satisfy the usual range requirements, tested using inria::is_range.
 * \endparblock
 *
 * \return An iterator to the new end of the output range.
 *
 * \exception std::runtime_error is thrown if the output range is too small.
 */
template<class InRange, class NonAssignableOutRange,
         #ifndef DOXYGEN_DOCUMENTATION
         class = enable_if_t<is_range<typename std::remove_reference<InRange>::type>::value>,
         class T = range_element_t<typename std::remove_reference<InRange>::type>,
         class = enable_if_t<is_range<NonAssignableOutRange>::value>,
         class = disable_if_t<is_assignable<NonAssignableOutRange, T*, T*>::value>,
         class AvoidTemplateRedeclarationError = void
         #endif
         >
auto distribute(inria::mpi_config conf,
                InRange&& in,
                NonAssignableOutRange& out)
    -> decltype(std::begin(out))
{
    using std::begin;
    using std::end;
    return distribute(conf, begin(in), end(in), begin(out), end(out), uniform_distribution<>{conf, begin(in), end(in)});
}

}

#endif /* _DISTRIBUTE_HPP_ */