src/smpi/colls/smpi_coll.cpp

   1 /* smpi_coll.c -- various optimized routing for collectives                 */
   2
   3 /* Copyright (c) 2009-2023. The SimGrid Team. All rights reserved.          */
   4
   5 /* This program is free software; you can redistribute it and/or modify it
   6  * under the terms of the license (GNU LGPL) which comes with this package. */
   7
   8 #include "smpi_coll.hpp"
   9 #include "private.hpp"
  10 #include "smpi_comm.hpp"
  11 #include "smpi_datatype.hpp"
  12 #include "smpi_op.hpp"
  13 #include "smpi_request.hpp"
  14 #include "xbt/config.hpp"
  15
  16 #include <iomanip>
  17 #include <map>
  18 #include <numeric>
  19 #include <sstream>
  20
  21 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_coll, smpi, "Logging specific to SMPI collectives.");
  22
  23 namespace simgrid::smpi {
  24
  25 std::map<std::string, std::vector<s_mpi_coll_description_t>, std::less<>> smpi_coll_descriptions(
  26     {{"gather",
  27       {{"default", "gather default collective", (void*)gather__default},
  28        {"ompi", "gather ompi collective", (void*)gather__ompi},
  29        {"ompi_basic_linear", "gather ompi_basic_linear collective", (void*)gather__ompi_basic_linear},
  30        {"ompi_binomial", "gather ompi_binomial collective", (void*)gather__ompi_binomial},
  31        {"ompi_linear_sync", "gather ompi_linear_sync collective", (void*)gather__ompi_linear_sync},
  32        {"mpich", "gather mpich collective", (void*)gather__mpich},
  33        {"mvapich2", "gather mvapich2 collective", (void*)gather__mvapich2},
  34        {"mvapich2_two_level", "gather mvapich2_two_level collective", (void*)gather__mvapich2_two_level},
  35        {"impi", "gather impi collective", (void*)gather__impi},
  36        {"automatic", "gather automatic collective", (void*)gather__automatic}}},
  37
  38      {"allgather",
  39       {{"default", "allgather default collective", (void*)allgather__default},
  40        {"2dmesh", "allgather 2dmesh collective", (void*)allgather__2dmesh},
  41        {"3dmesh", "allgather 3dmesh collective", (void*)allgather__3dmesh},
  42        {"bruck", "allgather bruck collective", (void*)allgather__bruck},
  43        {"GB", "allgather GB collective", (void*)allgather__GB},
  44        {"loosely_lr", "allgather loosely_lr collective", (void*)allgather__loosely_lr},
  45        {"NTSLR", "allgather NTSLR collective", (void*)allgather__NTSLR},
  46        {"NTSLR_NB", "allgather NTSLR_NB collective", (void*)allgather__NTSLR_NB},
  47        {"pair", "allgather pair collective", (void*)allgather__pair},
  48        {"rdb", "allgather rdb collective", (void*)allgather__rdb},
  49        {"rhv", "allgather rhv collective", (void*)allgather__rhv},
  50        {"ring", "allgather ring collective", (void*)allgather__ring},
  51        {"SMP_NTS", "allgather SMP_NTS collective", (void*)allgather__SMP_NTS},
  52        {"smp_simple", "allgather smp_simple collective", (void*)allgather__smp_simple},
  53        {"spreading_simple", "allgather spreading_simple collective", (void*)allgather__spreading_simple},
  54        {"ompi", "allgather ompi collective", (void*)allgather__ompi},
  55        {"ompi_neighborexchange", "allgather ompi_neighborexchange collective", (void*)allgather__ompi_neighborexchange},
  56        {"mvapich2", "allgather mvapich2 collective", (void*)allgather__mvapich2},
  57        {"mvapich2_smp", "allgather mvapich2_smp collective", (void*)allgather__mvapich2_smp},
  58        {"mpich", "allgather mpich collective", (void*)allgather__mpich},
  59        {"impi", "allgather impi collective", (void*)allgather__impi},
  60        {"automatic", "allgather automatic collective", (void*)allgather__automatic}}},
  61
  62      {"allgatherv",
  63       {{"default", "allgatherv default collective", (void*)allgatherv__default},
  64        {"GB", "allgatherv GB collective", (void*)allgatherv__GB},
  65        {"pair", "allgatherv pair collective", (void*)allgatherv__pair},
  66        {"ring", "allgatherv ring collective", (void*)allgatherv__ring},
  67        {"ompi", "allgatherv ompi collective", (void*)allgatherv__ompi},
  68        {"ompi_neighborexchange", "allgatherv ompi_neighborexchange collective",
  69         (void*)allgatherv__ompi_neighborexchange},
  70        {"ompi_bruck", "allgatherv ompi_bruck collective", (void*)allgatherv__ompi_bruck},
  71        {"mpich", "allgatherv mpich collective", (void*)allgatherv__mpich},
  72        {"mpich_rdb", "allgatherv mpich_rdb collective", (void*)allgatherv__mpich_rdb},
  73        {"mpich_ring", "allgatherv mpich_ring collective", (void*)allgatherv__mpich_ring},
  74        {"mvapich2", "allgatherv mvapich2 collective", (void*)allgatherv__mvapich2},
  75        {"impi", "allgatherv impi collective", (void*)allgatherv__impi},
  76        {"automatic", "allgatherv automatic collective", (void*)allgatherv__automatic}}},
  77
  78      {"allreduce",
  79       {{"default", "allreduce default collective", (void*)allreduce__default},
  80        {"lr", "allreduce lr collective", (void*)allreduce__lr},
  81        {"rab1", "allreduce rab1 collective", (void*)allreduce__rab1},
  82        {"rab2", "allreduce rab2 collective", (void*)allreduce__rab2},
  83        {"rab_rdb", "allreduce rab_rdb collective", (void*)allreduce__rab_rdb},
  84        {"rdb", "allreduce rdb collective", (void*)allreduce__rdb},
  85        {"smp_binomial", "allreduce smp_binomial collective", (void*)allreduce__smp_binomial},
  86        {"smp_binomial_pipeline", "allreduce smp_binomial_pipeline collective", (void*)allreduce__smp_binomial_pipeline},
  87        {"smp_rdb", "allreduce smp_rdb collective", (void*)allreduce__smp_rdb},
  88        {"smp_rsag", "allreduce smp_rsag collective", (void*)allreduce__smp_rsag},
  89        {"smp_rsag_lr", "allreduce smp_rsag_lr collective", (void*)allreduce__smp_rsag_lr},
  90        {"smp_rsag_rab", "allreduce smp_rsag_rab collective", (void*)allreduce__smp_rsag_rab},
  91        {"redbcast", "allreduce redbcast collective", (void*)allreduce__redbcast},
  92        {"ompi", "allreduce ompi collective", (void*)allreduce__ompi},
  93        {"ompi_ring_segmented", "allreduce ompi_ring_segmented collective", (void*)allreduce__ompi_ring_segmented},
  94        {"mpich", "allreduce mpich collective", (void*)allreduce__mpich},
  95        {"mvapich2", "allreduce mvapich2 collective", (void*)allreduce__mvapich2},
  96        {"mvapich2_rs", "allreduce mvapich2_rs collective", (void*)allreduce__mvapich2_rs},
  97        {"mvapich2_two_level", "allreduce mvapich2_two_level collective", (void*)allreduce__mvapich2_two_level},
  98        {"impi", "allreduce impi collective", (void*)allreduce__impi},
  99        {"rab", "allreduce rab collective", (void*)allreduce__rab},
 100        {"automatic", "allreduce automatic collective", (void*)allreduce__automatic}}},
 101
 102      {"reduce_scatter",
 103       {{"default", "reduce_scatter default collective", (void*)reduce_scatter__default},
 104        {"ompi", "reduce_scatter ompi collective", (void*)reduce_scatter__ompi},
 105        {"ompi_basic_recursivehalving", "reduce_scatter ompi_basic_recursivehalving collective",
 106         (void*)reduce_scatter__ompi_basic_recursivehalving},
 107        {"ompi_ring", "reduce_scatter ompi_ring collective", (void*)reduce_scatter__ompi_ring},
 108        {"ompi_butterfly", "reduce_scatter ompi_butterfly collective", (void*)reduce_scatter__ompi_butterfly},
 109        {"mpich", "reduce_scatter mpich collective", (void*)reduce_scatter__mpich},
 110        {"mpich_pair", "reduce_scatter mpich_pair collective", (void*)reduce_scatter__mpich_pair},
 111        {"mpich_rdb", "reduce_scatter mpich_rdb collective", (void*)reduce_scatter__mpich_rdb},
 112        {"mpich_noncomm", "reduce_scatter mpich_noncomm collective", (void*)reduce_scatter__mpich_noncomm},
 113        {"mvapich2", "reduce_scatter mvapich2 collective", (void*)reduce_scatter__mvapich2},
 114        {"impi", "reduce_scatter impi collective", (void*)reduce_scatter__impi},
 115        {"automatic", "reduce_scatter automatic collective", (void*)reduce_scatter__automatic}}},
 116
 117      {"scatter",
 118       {{"default", "scatter default collective", (void*)scatter__default},
 119        {"ompi", "scatter ompi collective", (void*)scatter__ompi},
 120        {"ompi_basic_linear", "scatter ompi_basic_linear collective", (void*)scatter__ompi_basic_linear},
 121        {"ompi_linear_nb", "scatter ompi_linear nonblocking collective", (void*)scatter__ompi_linear_nb},
 122        {"ompi_binomial", "scatter ompi_binomial collective", (void*)scatter__ompi_binomial},
 123        {"mpich", "scatter mpich collective", (void*)scatter__mpich},
 124        {"mvapich2", "scatter mvapich2 collective", (void*)scatter__mvapich2},
 125        {"mvapich2_two_level_binomial", "scatter mvapich2_two_level_binomial collective",
 126         (void*)scatter__mvapich2_two_level_binomial},
 127        {"mvapich2_two_level_direct", "scatter mvapich2_two_level_direct collective",
 128         (void*)scatter__mvapich2_two_level_direct},
 129        {"impi", "scatter impi collective", (void*)scatter__impi},
 130        {"automatic", "scatter automatic collective", (void*)scatter__automatic}}},
 131
 132      {"barrier",
 133       {{"default", "barrier default collective", (void*)barrier__default},
 134        {"ompi", "barrier ompi collective", (void*)barrier__ompi},
 135        {"ompi_basic_linear", "barrier ompi_basic_linear collective", (void*)barrier__ompi_basic_linear},
 136        {"ompi_two_procs", "barrier ompi_two_procs collective", (void*)barrier__ompi_two_procs},
 137        {"ompi_tree", "barrier ompi_tree collective", (void*)barrier__ompi_tree},
 138        {"ompi_bruck", "barrier ompi_bruck collective", (void*)barrier__ompi_bruck},
 139        {"ompi_recursivedoubling", "barrier ompi_recursivedoubling collective", (void*)barrier__ompi_recursivedoubling},
 140        {"ompi_doublering", "barrier ompi_doublering collective", (void*)barrier__ompi_doublering},
 141        {"mpich_smp", "barrier mpich_smp collective", (void*)barrier__mpich_smp},
 142        {"mpich", "barrier mpich collective", (void*)barrier__mpich},
 143        {"mvapich2_pair", "barrier mvapich2_pair collective", (void*)barrier__mvapich2_pair},
 144        {"mvapich2", "barrier mvapich2 collective", (void*)barrier__mvapich2},
 145        {"impi", "barrier impi collective", (void*)barrier__impi},
 146        {"automatic", "barrier automatic collective", (void*)barrier__automatic}}},
 147
 148      {"alltoall",
 149       {{"default", "alltoall default collective", (void*)alltoall__default},
 150        {"2dmesh", "alltoall 2dmesh collective", (void*)alltoall__2dmesh},
 151        {"3dmesh", "alltoall 3dmesh collective", (void*)alltoall__3dmesh},
 152        {"basic_linear", "alltoall basic_linear collective", (void*)alltoall__basic_linear},
 153        {"bruck", "alltoall bruck collective", (void*)alltoall__bruck},
 154        {"pair", "alltoall pair collective", (void*)alltoall__pair},
 155        {"pair_rma", "alltoall pair_rma collective", (void*)alltoall__pair_rma},
 156        {"pair_light_barrier", "alltoall pair_light_barrier collective", (void*)alltoall__pair_light_barrier},
 157        {"pair_mpi_barrier", "alltoall pair_mpi_barrier collective", (void*)alltoall__pair_mpi_barrier},
 158        {"pair_one_barrier", "alltoall pair_one_barrier collective", (void*)alltoall__pair_one_barrier},
 159        {"rdb", "alltoall rdb collective", (void*)alltoall__rdb},
 160        {"ring", "alltoall ring collective", (void*)alltoall__ring},
 161        {"ring_light_barrier", "alltoall ring_light_barrier collective", (void*)alltoall__ring_light_barrier},
 162        {"ring_mpi_barrier", "alltoall ring_mpi_barrier collective", (void*)alltoall__ring_mpi_barrier},
 163        {"ring_one_barrier", "alltoall ring_one_barrier collective", (void*)alltoall__ring_one_barrier},
 164        {"mvapich2", "alltoall mvapich2 collective", (void*)alltoall__mvapich2},
 165        {"mvapich2_scatter_dest", "alltoall mvapich2_scatter_dest collective", (void*)alltoall__mvapich2_scatter_dest},
 166        {"ompi", "alltoall ompi collective", (void*)alltoall__ompi},
 167        {"mpich", "alltoall mpich collective", (void*)alltoall__mpich},
 168        {"impi", "alltoall impi collective", (void*)alltoall__impi},
 169        {"automatic", "alltoall automatic collective", (void*)alltoall__automatic}}},
 170
 171      {"alltoallv",
 172       {{"default", "alltoallv default collective", (void*)alltoallv__default},
 173        {"bruck", "alltoallv bruck collective", (void*)alltoallv__bruck},
 174        {"pair", "alltoallv pair collective", (void*)alltoallv__pair},
 175        {"pair_light_barrier", "alltoallv pair_light_barrier collective", (void*)alltoallv__pair_light_barrier},
 176        {"pair_mpi_barrier", "alltoallv pair_mpi_barrier collective", (void*)alltoallv__pair_mpi_barrier},
 177        {"pair_one_barrier", "alltoallv pair_one_barrier collective", (void*)alltoallv__pair_one_barrier},
 178        {"ring", "alltoallv ring collective", (void*)alltoallv__ring},
 179        {"ring_light_barrier", "alltoallv ring_light_barrier collective", (void*)alltoallv__ring_light_barrier},
 180        {"ring_mpi_barrier", "alltoallv ring_mpi_barrier collective", (void*)alltoallv__ring_mpi_barrier},
 181        {"ring_one_barrier", "alltoallv ring_one_barrier collective", (void*)alltoallv__ring_one_barrier},
 182        {"ompi", "alltoallv ompi collective", (void*)alltoallv__ompi},
 183        {"mpich", "alltoallv mpich collective", (void*)alltoallv__mpich},
 184        {"ompi_basic_linear", "alltoallv ompi_basic_linear collective", (void*)alltoallv__ompi_basic_linear},
 185        {"mvapich2", "alltoallv mvapich2 collective", (void*)alltoallv__mvapich2},
 186        {"impi", "alltoallv impi collective", (void*)alltoallv__impi},
 187        {"automatic", "alltoallv automatic collective", (void*)alltoallv__automatic}}},
 188
 189      {"bcast",
 190       {{"default", "bcast default collective", (void*)bcast__default},
 191        {"arrival_pattern_aware", "bcast arrival_pattern_aware collective", (void*)bcast__arrival_pattern_aware},
 192        {"arrival_pattern_aware_wait", "bcast arrival_pattern_aware_wait collective",
 193         (void*)bcast__arrival_pattern_aware_wait},
 194        {"arrival_scatter", "bcast arrival_scatter collective", (void*)bcast__arrival_scatter},
 195        {"binomial_tree", "bcast binomial_tree collective", (void*)bcast__binomial_tree},
 196        {"flattree", "bcast flattree collective", (void*)bcast__flattree},
 197        {"flattree_pipeline", "bcast flattree_pipeline collective", (void*)bcast__flattree_pipeline},
 198        {"NTSB", "bcast NTSB collective", (void*)bcast__NTSB},
 199        {"NTSL", "bcast NTSL collective", (void*)bcast__NTSL},
 200        {"NTSL_Isend", "bcast NTSL_Isend collective", (void*)bcast__NTSL_Isend},
 201        {"scatter_LR_allgather", "bcast scatter_LR_allgather collective", (void*)bcast__scatter_LR_allgather},
 202        {"scatter_rdb_allgather", "bcast scatter_rdb_allgather collective", (void*)bcast__scatter_rdb_allgather},
 203        {"SMP_binary", "bcast SMP_binary collective", (void*)bcast__SMP_binary},
 204        {"SMP_binomial", "bcast SMP_binomial collective", (void*)bcast__SMP_binomial},
 205        {"SMP_linear", "bcast SMP_linear collective", (void*)bcast__SMP_linear},
 206        {"ompi", "bcast ompi collective", (void*)bcast__ompi},
 207        {"ompi_split_bintree", "bcast ompi_split_bintree collective", (void*)bcast__ompi_split_bintree},
 208        {"ompi_pipeline", "bcast ompi_pipeline collective", (void*)bcast__ompi_pipeline},
 209        {"mpich", "bcast mpich collective", (void*)bcast__mpich},
 210        {"mvapich2", "bcast mvapich2 collective", (void*)bcast__mvapich2},
 211        {"mvapich2_inter_node", "bcast mvapich2_inter_node collective", (void*)bcast__mvapich2_inter_node},
 212        {"mvapich2_intra_node", "bcast mvapich2_intra_node collective", (void*)bcast__mvapich2_intra_node},
 213        {"mvapich2_knomial_intra_node", "bcast mvapich2_knomial_intra_node collective",
 214         (void*)bcast__mvapich2_knomial_intra_node},
 215        {"impi", "bcast impi collective", (void*)bcast__impi},
 216        {"automatic", "bcast automatic collective", (void*)bcast__automatic}}},
 217
 218      {"reduce",
 219       {{"default", "reduce default collective", (void*)reduce__default},
 220        {"arrival_pattern_aware", "reduce arrival_pattern_aware collective", (void*)reduce__arrival_pattern_aware},
 221        {"binomial", "reduce binomial collective", (void*)reduce__binomial},
 222        {"flat_tree", "reduce flat_tree collective", (void*)reduce__flat_tree},
 223        {"NTSL", "reduce NTSL collective", (void*)reduce__NTSL},
 224        {"scatter_gather", "reduce scatter_gather collective", (void*)reduce__scatter_gather},
 225        {"ompi", "reduce ompi collective", (void*)reduce__ompi},
 226        {"ompi_chain", "reduce ompi_chain collective", (void*)reduce__ompi_chain},
 227        {"ompi_pipeline", "reduce ompi_pipeline collective", (void*)reduce__ompi_pipeline},
 228        {"ompi_basic_linear", "reduce ompi_basic_linear collective", (void*)reduce__ompi_basic_linear},
 229        {"ompi_in_order_binary", "reduce ompi_in_order_binary collective", (void*)reduce__ompi_in_order_binary},
 230        {"ompi_binary", "reduce ompi_binary collective", (void*)reduce__ompi_binary},
 231        {"ompi_binomial", "reduce ompi_binomial collective", (void*)reduce__ompi_binomial},
 232        {"mpich", "reduce mpich collective", (void*)reduce__mpich},
 233        {"mvapich2", "reduce mvapich2 collective", (void*)reduce__mvapich2},
 234        {"mvapich2_knomial", "reduce mvapich2_knomial collective", (void*)reduce__mvapich2_knomial},
 235        {"mvapich2_two_level", "reduce mvapich2_two_level collective", (void*)reduce__mvapich2_two_level},
 236        {"impi", "reduce impi collective", (void*)reduce__impi},
 237        {"rab", "reduce rab collective", (void*)reduce__rab},
 238        {"automatic", "reduce automatic collective", (void*)reduce__automatic}}}});
 239
 240 // Needed by the automatic selector weird implementation
 241 std::vector<s_mpi_coll_description_t>* colls::get_smpi_coll_descriptions(const std::string& name)
 242 {
 243   auto iter = smpi_coll_descriptions.find(name);
 244   xbt_assert(iter != smpi_coll_descriptions.end(), "No collective named %s. This is a bug.", name.c_str());
 245   return &iter->second;
 246 }
 247
 248 std::string colls::get_smpi_coll_help()
 249 {
 250   size_t max_name_len =
 251       std::accumulate(begin(smpi_coll_descriptions), end(smpi_coll_descriptions), 0, [](auto len, auto const& coll) {
 252         return std::max(len, std::accumulate(begin(coll.second), end(coll.second), 0, [](auto len, auto const& descr) {
 253                           return std::max<size_t>(len, descr.name.length());
 254                         }));
 255       });
 256   std::ostringstream oss;
 257   std::string title = "Available collective algorithms (select them with \"smpi/collective_name:algo_name\"):";
 258   oss << title << '\n' << std::setfill('=') << std::setw(title.length() + 1) << '\n';
 259   for (auto const& [coll, algos] : smpi_coll_descriptions) {
 260     std::string line = "Collective: \"" + coll + "\"";
 261     oss << line << '\n' << std::setfill('-') << std::setw(line.length() + 1) << '\n';
 262     oss << std::setfill(' ') << std::left;
 263     for (auto const& [name, descr, _] : algos)
 264       oss << "  " << std::setw(max_name_len) << name << " " << descr << "\n";
 265     oss << std::right << '\n';
 266   }
 267   oss << "Please see https://simgrid.org/doc/latest/app_smpi.html#available-algorithms for more information.\n";
 268   return oss.str();
 269 }
 270
 271 static s_mpi_coll_description_t* find_coll_description(const std::string& collective, const std::string& algo)
 272 {
 273   std::vector<s_mpi_coll_description_t>* table = colls::get_smpi_coll_descriptions(collective);
 274   xbt_assert(not table->empty(), "No registered algorithm for collective '%s'! This is a bug.", collective.c_str());
 275
 276   for (auto& desc : *table) {
 277     if (algo == desc.name) {
 278       if (desc.name != "default")
 279         XBT_INFO("Switch to algorithm %s for collective %s", desc.name.c_str(), collective.c_str());
 280       return &desc;
 281     }
 282   }
 283
 284   std::string name_list = table->at(0).name;
 285   for (unsigned long i = 1; i < table->size(); i++)
 286     name_list = name_list + ", " + table->at(i).name;
 287   xbt_die("Collective '%s' has no algorithm '%s'! Valid algorithms: %s. Please use --help-coll for details.",
 288           collective.c_str(), algo.c_str(), name_list.c_str());
 289 }
 290
 291 int (*colls::gather)(const void* send_buff, int send_count, MPI_Datatype send_type, void* recv_buff, int recv_count,
 292                      MPI_Datatype recv_type, int root, MPI_Comm comm);
 293 int (*colls::allgather)(const void* send_buff, int send_count, MPI_Datatype send_type, void* recv_buff, int recv_count,
 294                         MPI_Datatype recv_type, MPI_Comm comm);
 295 int (*colls::allgatherv)(const void* send_buff, int send_count, MPI_Datatype send_type, void* recv_buff,
 296                          const int* recv_count, const int* recv_disps, MPI_Datatype recv_type, MPI_Comm comm);
 297 int (*colls::alltoall)(const void* send_buff, int send_count, MPI_Datatype send_type, void* recv_buff, int recv_count,
 298                        MPI_Datatype recv_type, MPI_Comm comm);
 299 int (*colls::alltoallv)(const void* send_buff, const int* send_counts, const int* send_disps, MPI_Datatype send_type,
 300                         void* recv_buff, const int* recv_counts, const int* recv_disps, MPI_Datatype recv_type,
 301                         MPI_Comm comm);
 302 int (*colls::bcast)(void* buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm);
 303 int (*colls::reduce)(const void* buf, void* rbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm);
 304 int (*colls::allreduce)(const void* sbuf, void* rbuf, int rcount, MPI_Datatype dtype, MPI_Op op, MPI_Comm comm);
 305 int (*colls::reduce_scatter)(const void* sbuf, void* rbuf, const int* rcounts, MPI_Datatype dtype, MPI_Op op,
 306                              MPI_Comm comm);
 307 int (*colls::scatter)(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount,
 308                       MPI_Datatype recvtype, int root, MPI_Comm comm);
 309 int (*colls::barrier)(MPI_Comm comm);
 310
 311 void (*colls::smpi_coll_cleanup_callback)();
 312
 313 #define COLL_SETTER(cat, ret, args, args2)                                                                             \
 314   void colls::_XBT_CONCAT(set_, cat)(const std::string& name)                                                          \
 315   {                                                                                                                    \
 316     auto desc = find_coll_description(_XBT_STRINGIFY(cat), name);                                                      \
 317     cat       = reinterpret_cast<ret(*) args>(desc->coll);                                                             \
 318     xbt_assert(cat != nullptr, "Collective " _XBT_STRINGIFY(cat) " set to nullptr!");                                  \
 319   }
 320 COLL_APPLY(COLL_SETTER, COLL_GATHER_SIG, "")
 321 COLL_APPLY(COLL_SETTER,COLL_ALLGATHER_SIG,"")
 322 COLL_APPLY(COLL_SETTER,COLL_ALLGATHERV_SIG,"")
 323 COLL_APPLY(COLL_SETTER,COLL_REDUCE_SIG,"")
 324 COLL_APPLY(COLL_SETTER,COLL_ALLREDUCE_SIG,"")
 325 COLL_APPLY(COLL_SETTER,COLL_REDUCE_SCATTER_SIG,"")
 326 COLL_APPLY(COLL_SETTER,COLL_SCATTER_SIG,"")
 327 COLL_APPLY(COLL_SETTER,COLL_BARRIER_SIG,"")
 328 COLL_APPLY(COLL_SETTER,COLL_BCAST_SIG,"")
 329 COLL_APPLY(COLL_SETTER,COLL_ALLTOALL_SIG,"")
 330 COLL_APPLY(COLL_SETTER,COLL_ALLTOALLV_SIG,"")
 331
 332 void colls::set_collectives()
 333 {
 334   std::string selector_name = simgrid::config::get_value<std::string>("smpi/coll-selector");
 335   if (selector_name.empty())
 336     selector_name = "default";
 337
 338   std::pair<std::string, std::function<void(std::string)>> setter_callbacks[] = {
 339       {"gather", &colls::set_gather},         {"allgather", &colls::set_allgather},
 340       {"allgatherv", &colls::set_allgatherv}, {"allreduce", &colls::set_allreduce},
 341       {"alltoall", &colls::set_alltoall},     {"alltoallv", &colls::set_alltoallv},
 342       {"reduce", &colls::set_reduce},         {"reduce_scatter", &colls::set_reduce_scatter},
 343       {"scatter", &colls::set_scatter},       {"bcast", &colls::set_bcast},
 344       {"barrier", &colls::set_barrier}};
 345
 346   for (auto& elem : setter_callbacks) {
 347     std::string name = simgrid::config::get_value<std::string>(("smpi/" + elem.first).c_str());
 348     if (name.empty())
 349       name = selector_name;
 350
 351     (elem.second)(name);
 352   }
 353 }
 354
 355 //Implementations of the single algorithm collectives
 356
 357 int colls::gatherv(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, const int* recvcounts,
 358                    const int* displs, MPI_Datatype recvtype, int root, MPI_Comm comm)
 359 {
 360   MPI_Request request;
 361   colls::igatherv(sendbuf, sendcount, sendtype, recvbuf, recvcounts, displs, recvtype, root, comm, &request, 0);
 362   return Request::wait(&request, MPI_STATUS_IGNORE);
 363 }
 364
 365 int colls::scatterv(const void* sendbuf, const int* sendcounts, const int* displs, MPI_Datatype sendtype, void* recvbuf,
 366                     int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm)
 367 {
 368   MPI_Request request;
 369   colls::iscatterv(sendbuf, sendcounts, displs, sendtype, recvbuf, recvcount, recvtype, root, comm, &request, 0);
 370   return Request::wait(&request, MPI_STATUS_IGNORE);
 371 }
 372
 373 int colls::scan(const void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
 374 {
 375   int system_tag = -888;
 376   MPI_Aint lb      = 0;
 377   MPI_Aint dataext = 0;
 378
 379   int rank = comm->rank();
 380   int size = comm->size();
 381
 382   datatype->extent(&lb, &dataext);
 383
 384   // Local copy from self
 385   Datatype::copy(sendbuf, count, datatype, recvbuf, count, datatype);
 386
 387   // Send/Recv buffers to/from others
 388   auto* requests = new MPI_Request[size - 1];
 389   auto** tmpbufs = new unsigned char*[rank];
 390   int index = 0;
 391   for (int other = 0; other < rank; other++) {
 392     tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
 393     requests[index] = Request::irecv_init(tmpbufs[index], count, datatype, other, system_tag, comm);
 394     index++;
 395   }
 396   for (int other = rank + 1; other < size; other++) {
 397     requests[index] = Request::isend_init(sendbuf, count, datatype, other, system_tag, comm);
 398     index++;
 399   }
 400   // Wait for completion of all comms.
 401   Request::startall(size - 1, requests);
 402
 403   if(op != MPI_OP_NULL && op->is_commutative()){
 404     for (int other = 0; other < size - 1; other++) {
 405       index = Request::waitany(size - 1, requests, MPI_STATUS_IGNORE);
 406       if(index == MPI_UNDEFINED) {
 407         break;
 408       }
 409       if(index < rank) {
 410         // #Request is below rank: it's an irecv
 411         op->apply( tmpbufs[index], recvbuf, &count, datatype);
 412       }
 413     }
 414   }else{
 415     //non commutative case, wait in order
 416     for (int other = 0; other < size - 1; other++) {
 417       Request::wait(&(requests[other]), MPI_STATUS_IGNORE);
 418       if(index < rank && op!=MPI_OP_NULL) {
 419         op->apply( tmpbufs[other], recvbuf, &count, datatype);
 420       }
 421     }
 422   }
 423   for(index = 0; index < rank; index++) {
 424     smpi_free_tmp_buffer(tmpbufs[index]);
 425   }
 426   for(index = 0; index < size-1; index++) {
 427     Request::unref(&requests[index]);
 428   }
 429   delete[] tmpbufs;
 430   delete[] requests;
 431   return MPI_SUCCESS;
 432 }
 433
 434 int colls::exscan(const void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
 435 {
 436   int system_tag = -888;
 437   MPI_Aint lb         = 0;
 438   MPI_Aint dataext    = 0;
 439   int recvbuf_is_empty=1;
 440   int rank = comm->rank();
 441   int size = comm->size();
 442
 443   datatype->extent(&lb, &dataext);
 444
 445   // Send/Recv buffers to/from others
 446   auto* requests = new MPI_Request[size - 1];
 447   auto** tmpbufs = new unsigned char*[rank];
 448   int index = 0;
 449   for (int other = 0; other < rank; other++) {
 450     tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext);
 451     requests[index] = Request::irecv_init(tmpbufs[index], count, datatype, other, system_tag, comm);
 452     index++;
 453   }
 454   for (int other = rank + 1; other < size; other++) {
 455     requests[index] = Request::isend_init(sendbuf, count, datatype, other, system_tag, comm);
 456     index++;
 457   }
 458   // Wait for completion of all comms.
 459   Request::startall(size - 1, requests);
 460
 461   if(op != MPI_OP_NULL && op->is_commutative()){
 462     for (int other = 0; other < size - 1; other++) {
 463       index = Request::waitany(size - 1, requests, MPI_STATUS_IGNORE);
 464       if(index == MPI_UNDEFINED) {
 465         break;
 466       }
 467       if(index < rank) {
 468         if(recvbuf_is_empty){
 469           Datatype::copy(tmpbufs[index], count, datatype, recvbuf, count, datatype);
 470           recvbuf_is_empty=0;
 471         } else
 472           // #Request is below rank: it's an irecv
 473           op->apply( tmpbufs[index], recvbuf, &count, datatype);
 474       }
 475     }
 476   }else{
 477     //non commutative case, wait in order
 478     for (int other = 0; other < size - 1; other++) {
 479      Request::wait(&(requests[other]), MPI_STATUS_IGNORE);
 480       if(index < rank) {
 481         if (recvbuf_is_empty) {
 482           Datatype::copy(tmpbufs[other], count, datatype, recvbuf, count, datatype);
 483           recvbuf_is_empty = 0;
 484         } else
 485           if(op!=MPI_OP_NULL)
 486             op->apply( tmpbufs[other], recvbuf, &count, datatype);
 487       }
 488     }
 489   }
 490   for(index = 0; index < rank; index++) {
 491     smpi_free_tmp_buffer(tmpbufs[index]);
 492   }
 493   for(index = 0; index < size-1; index++) {
 494     Request::unref(&requests[index]);
 495   }
 496   delete[] tmpbufs;
 497   delete[] requests;
 498   return MPI_SUCCESS;
 499 }
 500
 501 int colls::alltoallw(const void* sendbuf, const int* sendcounts, const int* senddisps, const MPI_Datatype* sendtypes,
 502                      void* recvbuf, const int* recvcounts, const int* recvdisps, const MPI_Datatype* recvtypes,
 503                      MPI_Comm comm)
 504 {
 505   MPI_Request request;
 506   colls::ialltoallw(sendbuf, sendcounts, senddisps, sendtypes, recvbuf, recvcounts, recvdisps, recvtypes, comm,
 507                     &request, 0);
 508   return Request::wait(&request, MPI_STATUS_IGNORE);
 509 }
 510
 511 } // namespace simgrid::smpi