src/smpi/include/smpi_coll.hpp

   1 /*High level handling of collective algorithms*/
   2 /* Copyright (c) 2009-2019. The SimGrid Team.
   3  * 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 #ifndef SMPI_COLL_HPP
   9 #define SMPI_COLL_HPP
  10
  11 #include "private.hpp"
  12 #include "xbt/base.h"
  13
  14 /** @brief MPI collective description */
  15
  16 #define COLL_DEFS(cat, ret, args, args2)                                                                               \
  17   static void set_##cat(const std::string& name);                                                                      \
  18   static s_mpi_coll_description_t mpi_coll_##cat##_description[];                                                      \
  19   static int(*cat) args;
  20
  21 #define COLL_SIG(cat, ret, args, args2)\
  22     static int cat args;
  23
  24
  25 #define COLL_DESCRIPTION(cat, ret, args, name) \
  26   {# name,\
  27    # cat " " # name " collective",\
  28    (void*) Coll_ ## cat ## _ ## name::cat }
  29
  30 #define COLL_PROTO(cat, ret, args, name) \
  31 class Coll_ ## cat ## _ ## name : public Coll { \
  32 public: \
  33 static ret cat  (COLL_UNPAREN args); \
  34 };
  35
  36 #define COLL_UNPAREN(...)  __VA_ARGS__
  37
  38 #define COLL_APPLY(action, sig, name) action(sig, name)
  39 #define COLL_COMMA ,
  40 #define COLL_NOsep
  41 #define COLL_NOTHING(...)
  42
  43 #define COLL_GATHER_SIG gather, int, \
  44                       (void *send_buff, int send_count, MPI_Datatype send_type, \
  45                        void *recv_buff, int recv_count, MPI_Datatype recv_type, \
  46                            int root, MPI_Comm comm)
  47 #define COLL_ALLGATHER_SIG allgather, int, \
  48                       (void *send_buff, int send_count, MPI_Datatype send_type, \
  49                        void *recv_buff, int recv_count, MPI_Datatype recv_type, \
  50                            MPI_Comm comm)
  51 #define COLL_ALLGATHERV_SIG allgatherv, int, \
  52                       (void *send_buff, int send_count, MPI_Datatype send_type, \
  53                        void *recv_buff, int *recv_count, int *recv_disps, \
  54                MPI_Datatype recv_type, MPI_Comm comm)
  55 #define COLL_ALLTOALL_SIG alltoall, int, \
  56                      (void *send_buff, int send_count, MPI_Datatype send_type, \
  57                       void *recv_buff, int recv_count, MPI_Datatype recv_type, \
  58                           MPI_Comm comm)
  59 #define COLL_ALLTOALLV_SIG alltoallv, int, \
  60                      (void *send_buff, int *send_counts, int *send_disps, MPI_Datatype send_type, \
  61                       void *recv_buff, int *recv_counts, int *recv_disps, MPI_Datatype recv_type, \
  62                           MPI_Comm comm)
  63 #define COLL_BCAST_SIG bcast, int, \
  64                   (void *buf, int count, MPI_Datatype datatype, \
  65                    int root, MPI_Comm comm)
  66 #define COLL_REDUCE_SIG reduce, int, \
  67                    (void *buf, void *rbuf, int count, MPI_Datatype datatype, \
  68                         MPI_Op op, int root, MPI_Comm comm)
  69 #define COLL_ALLREDUCE_SIG allreduce, int, \
  70                       (void *sbuf, void *rbuf, int rcount, \
  71                            MPI_Datatype dtype, MPI_Op op, MPI_Comm comm)
  72 #define COLL_REDUCE_SCATTER_SIG reduce_scatter, int, \
  73                       (void *sbuf, void *rbuf, int *rcounts,\
  74                     MPI_Datatype dtype,MPI_Op  op,MPI_Comm  comm)
  75 #define COLL_SCATTER_SIG scatter, int, \
  76                 (void *sendbuf, int sendcount, MPI_Datatype sendtype,\
  77                 void *recvbuf, int recvcount, MPI_Datatype recvtype,\
  78                 int root, MPI_Comm comm)
  79 #define COLL_BARRIER_SIG barrier, int, \
  80                 (MPI_Comm comm)
  81
  82 namespace simgrid{
  83 namespace smpi{
  84
  85 struct s_mpi_coll_description_t {
  86   std::string name;
  87   std::string description;
  88   void *coll;
  89 };
  90
  91 class Colls{
  92 public:
  93   static XBT_PUBLIC void coll_help(const char* category, s_mpi_coll_description_t* table);
  94   static XBT_PUBLIC int find_coll_description(s_mpi_coll_description_t* table, const std::string& name,
  95                                               const char* desc);
  96   static void set_collectives();
  97
  98   // for each collective type, create the set_* prototype, the description array and the function pointer
  99   COLL_APPLY(COLL_DEFS, COLL_GATHER_SIG, "");
 100   COLL_APPLY(COLL_DEFS, COLL_ALLGATHER_SIG, "");
 101   COLL_APPLY(COLL_DEFS, COLL_ALLGATHERV_SIG, "");
 102   COLL_APPLY(COLL_DEFS, COLL_REDUCE_SIG, "");
 103   COLL_APPLY(COLL_DEFS, COLL_ALLREDUCE_SIG, "");
 104   COLL_APPLY(COLL_DEFS, COLL_REDUCE_SCATTER_SIG, "");
 105   COLL_APPLY(COLL_DEFS, COLL_SCATTER_SIG, "");
 106   COLL_APPLY(COLL_DEFS, COLL_BARRIER_SIG, "");
 107   COLL_APPLY(COLL_DEFS, COLL_BCAST_SIG, "");
 108   COLL_APPLY(COLL_DEFS, COLL_ALLTOALL_SIG, "");
 109   COLL_APPLY(COLL_DEFS, COLL_ALLTOALLV_SIG, "");
 110
 111   // These fairly unused collectives only have one implementation in SMPI
 112   static int gatherv(void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int* recvcounts, int* displs,
 113                      MPI_Datatype recvtype, int root, MPI_Comm comm);
 114   static int scatterv(void* sendbuf, int* sendcounts, int* displs, MPI_Datatype sendtype, void* recvbuf, int recvcount,
 115                       MPI_Datatype recvtype, int root, MPI_Comm comm);
 116   static int scan(void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm);
 117   static int exscan(void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm);
 118   static int alltoallw
 119          (void* sendbuf, int* sendcounts, int* senddisps, MPI_Datatype* sendtypes, void* recvbuf, int* recvcounts,
 120           int* recvdisps, MPI_Datatype* recvtypes, MPI_Comm comm);
 121
 122   //async collectives
 123   static int ibarrier(MPI_Comm comm, MPI_Request* request);
 124   static int ibcast(void *buf, int count, MPI_Datatype datatype,
 125                    int root, MPI_Comm comm, MPI_Request* request);
 126   static int igather (void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount,
 127                                       MPI_Datatype recvtype, int root, MPI_Comm comm, MPI_Request *request);
 128   static int igatherv (void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf,
 129                                        int* recvcounts, int* displs, MPI_Datatype recvtype, int root, MPI_Comm comm, MPI_Request *request);
 130   static int iallgather (void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf,
 131                                          int recvcount, MPI_Datatype recvtype, MPI_Comm comm, MPI_Request *request);
 132   static int iallgatherv (void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf,
 133                                           int* recvcounts, int* displs, MPI_Datatype recvtype, MPI_Comm comm, MPI_Request *request);
 134   static int iscatter (void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf,
 135                                        int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm, MPI_Request *request);
 136   static int iscatterv (void* sendbuf, int* sendcounts, int* displs, MPI_Datatype sendtype,
 137                                         void* recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm, MPI_Request *request);
 138   static int ireduce
 139          (void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm, MPI_Request *request);
 140   static int iallreduce
 141          (void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm, MPI_Request *request);
 142   static int iscan
 143          (void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm, MPI_Request *request);
 144   static int iexscan
 145          (void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm, MPI_Request *request);
 146   static int ireduce_scatter
 147          (void* sendbuf, void* recvbuf, int* recvcounts, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm, MPI_Request *request);
 148   static int ireduce_scatter_block
 149          (void* sendbuf, void* recvbuf, int recvcount, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm, MPI_Request *request);
 150   static int ialltoall (void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf,
 151                                         int recvcount, MPI_Datatype recvtype, MPI_Comm comm, MPI_Request *request);
 152   static int ialltoallv
 153          (void* sendbuf, int* sendcounts, int* senddisps, MPI_Datatype sendtype, void* recvbuf, int* recvcounts,
 154           int* recvdisps, MPI_Datatype recvtype, MPI_Comm comm, MPI_Request *request);
 155   static int ialltoallw
 156          (void* sendbuf, int* sendcounts, int* senddisps, MPI_Datatype* sendtypes, void* recvbuf, int* recvcounts,
 157           int* recvdisps, MPI_Datatype* recvtypes, MPI_Comm comm, MPI_Request *request);
 158
 159
 160   static void (*smpi_coll_cleanup_callback)();
 161 };
 162
 163 class Coll {
 164 public:
 165   // for each collective type, create a function member
 166   COLL_APPLY(COLL_SIG, COLL_GATHER_SIG, "");
 167   COLL_APPLY(COLL_SIG, COLL_ALLGATHER_SIG, "");
 168   COLL_APPLY(COLL_SIG, COLL_ALLGATHERV_SIG, "");
 169   COLL_APPLY(COLL_SIG, COLL_REDUCE_SIG, "");
 170   COLL_APPLY(COLL_SIG, COLL_ALLREDUCE_SIG, "");
 171   COLL_APPLY(COLL_SIG, COLL_REDUCE_SCATTER_SIG, "");
 172   COLL_APPLY(COLL_SIG, COLL_SCATTER_SIG, "");
 173   COLL_APPLY(COLL_SIG, COLL_BARRIER_SIG, "");
 174   COLL_APPLY(COLL_SIG, COLL_BCAST_SIG, "");
 175   COLL_APPLY(COLL_SIG, COLL_ALLTOALL_SIG, "");
 176   COLL_APPLY(COLL_SIG, COLL_ALLTOALLV_SIG, "");
 177 };
 178
 179 /*************
 180  * GATHER *
 181  *************/
 182
 183 #define COLL_GATHERS(action, COLL_sep) \
 184 COLL_APPLY(action, COLL_GATHER_SIG, default) COLL_sep \
 185 COLL_APPLY(action, COLL_GATHER_SIG, ompi) COLL_sep \
 186 COLL_APPLY(action, COLL_GATHER_SIG, ompi_basic_linear) COLL_sep \
 187 COLL_APPLY(action, COLL_GATHER_SIG, ompi_binomial) COLL_sep \
 188 COLL_APPLY(action, COLL_GATHER_SIG, ompi_linear_sync) COLL_sep \
 189 COLL_APPLY(action, COLL_GATHER_SIG, mpich) COLL_sep \
 190 COLL_APPLY(action, COLL_GATHER_SIG, mvapich2) COLL_sep \
 191 COLL_APPLY(action, COLL_GATHER_SIG, mvapich2_two_level) COLL_sep \
 192 COLL_APPLY(action, COLL_GATHER_SIG, impi) COLL_sep \
 193 COLL_APPLY(action, COLL_GATHER_SIG, automatic)
 194
 195 COLL_GATHERS(COLL_PROTO, COLL_NOsep)
 196
 197 /*************
 198  * ALLGATHER *
 199  *************/
 200
 201 #define COLL_ALLGATHERS(action, COLL_sep) \
 202 COLL_APPLY(action, COLL_ALLGATHER_SIG, default) COLL_sep \
 203 COLL_APPLY(action, COLL_ALLGATHER_SIG, 2dmesh) COLL_sep \
 204 COLL_APPLY(action, COLL_ALLGATHER_SIG, 3dmesh) COLL_sep \
 205 COLL_APPLY(action, COLL_ALLGATHER_SIG, bruck) COLL_sep \
 206 COLL_APPLY(action, COLL_ALLGATHER_SIG, GB) COLL_sep \
 207 COLL_APPLY(action, COLL_ALLGATHER_SIG, loosely_lr) COLL_sep \
 208 COLL_APPLY(action, COLL_ALLGATHER_SIG, NTSLR) COLL_sep \
 209 COLL_APPLY(action, COLL_ALLGATHER_SIG, NTSLR_NB) COLL_sep \
 210 COLL_APPLY(action, COLL_ALLGATHER_SIG, pair) COLL_sep \
 211 COLL_APPLY(action, COLL_ALLGATHER_SIG, rdb) COLL_sep \
 212 COLL_APPLY(action, COLL_ALLGATHER_SIG, rhv) COLL_sep \
 213 COLL_APPLY(action, COLL_ALLGATHER_SIG, ring) COLL_sep \
 214 COLL_APPLY(action, COLL_ALLGATHER_SIG, SMP_NTS) COLL_sep \
 215 COLL_APPLY(action, COLL_ALLGATHER_SIG, smp_simple) COLL_sep \
 216 COLL_APPLY(action, COLL_ALLGATHER_SIG, spreading_simple) COLL_sep \
 217 COLL_APPLY(action, COLL_ALLGATHER_SIG, ompi) COLL_sep \
 218 COLL_APPLY(action, COLL_ALLGATHER_SIG, ompi_neighborexchange) COLL_sep \
 219 COLL_APPLY(action, COLL_ALLGATHER_SIG, mvapich2) COLL_sep \
 220 COLL_APPLY(action, COLL_ALLGATHER_SIG, mvapich2_smp) COLL_sep \
 221 COLL_APPLY(action, COLL_ALLGATHER_SIG, mpich) COLL_sep \
 222 COLL_APPLY(action, COLL_ALLGATHER_SIG, impi) COLL_sep \
 223 COLL_APPLY(action, COLL_ALLGATHER_SIG, automatic)
 224
 225 COLL_ALLGATHERS(COLL_PROTO, COLL_NOsep)
 226
 227 /**************
 228  * ALLGATHERV *
 229  **************/
 230
 231 #define COLL_ALLGATHERVS(action, COLL_sep) \
 232 COLL_APPLY(action, COLL_ALLGATHERV_SIG, default) COLL_sep \
 233 COLL_APPLY(action, COLL_ALLGATHERV_SIG, GB) COLL_sep \
 234 COLL_APPLY(action, COLL_ALLGATHERV_SIG, pair) COLL_sep \
 235 COLL_APPLY(action, COLL_ALLGATHERV_SIG, ring) COLL_sep \
 236 COLL_APPLY(action, COLL_ALLGATHERV_SIG, ompi) COLL_sep \
 237 COLL_APPLY(action, COLL_ALLGATHERV_SIG, ompi_neighborexchange) COLL_sep \
 238 COLL_APPLY(action, COLL_ALLGATHERV_SIG, ompi_bruck) COLL_sep \
 239 COLL_APPLY(action, COLL_ALLGATHERV_SIG, mpich) COLL_sep \
 240 COLL_APPLY(action, COLL_ALLGATHERV_SIG, mpich_rdb) COLL_sep \
 241 COLL_APPLY(action, COLL_ALLGATHERV_SIG, mpich_ring) COLL_sep \
 242 COLL_APPLY(action, COLL_ALLGATHERV_SIG, mvapich2) COLL_sep \
 243 COLL_APPLY(action, COLL_ALLGATHERV_SIG, impi) COLL_sep \
 244 COLL_APPLY(action, COLL_ALLGATHERV_SIG, automatic)
 245
 246 COLL_ALLGATHERVS(COLL_PROTO, COLL_NOsep)
 247
 248 /*************
 249  * ALLREDUCE *
 250  *************/
 251
 252 #define COLL_ALLREDUCES(action, COLL_sep) \
 253 COLL_APPLY(action, COLL_ALLREDUCE_SIG, default) COLL_sep \
 254 COLL_APPLY(action, COLL_ALLREDUCE_SIG, lr) COLL_sep \
 255 COLL_APPLY(action, COLL_ALLREDUCE_SIG, rab1) COLL_sep \
 256 COLL_APPLY(action, COLL_ALLREDUCE_SIG, rab2) COLL_sep \
 257 COLL_APPLY(action, COLL_ALLREDUCE_SIG, rab_rdb) COLL_sep \
 258 COLL_APPLY(action, COLL_ALLREDUCE_SIG, rdb) COLL_sep \
 259 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_binomial) COLL_sep \
 260 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_binomial_pipeline) COLL_sep \
 261 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_rdb) COLL_sep \
 262 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_rsag) COLL_sep \
 263 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_rsag_lr) COLL_sep \
 264 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_rsag_rab) COLL_sep \
 265 COLL_APPLY(action, COLL_ALLREDUCE_SIG, redbcast) COLL_sep \
 266 COLL_APPLY(action, COLL_ALLREDUCE_SIG, ompi) COLL_sep \
 267 COLL_APPLY(action, COLL_ALLREDUCE_SIG, ompi_ring_segmented) COLL_sep \
 268 COLL_APPLY(action, COLL_ALLREDUCE_SIG, mpich) COLL_sep \
 269 COLL_APPLY(action, COLL_ALLREDUCE_SIG, mvapich2) COLL_sep \
 270 COLL_APPLY(action, COLL_ALLREDUCE_SIG, mvapich2_rs) COLL_sep \
 271 COLL_APPLY(action, COLL_ALLREDUCE_SIG, mvapich2_two_level) COLL_sep \
 272 COLL_APPLY(action, COLL_ALLREDUCE_SIG, impi) COLL_sep \
 273 COLL_APPLY(action, COLL_ALLREDUCE_SIG, rab) COLL_sep \
 274 COLL_APPLY(action, COLL_ALLREDUCE_SIG, automatic)
 275
 276 COLL_ALLREDUCES(COLL_PROTO, COLL_NOsep)
 277
 278 /************
 279  * ALLTOALL *
 280  ************/
 281
 282 #define COLL_ALLTOALLS(action, COLL_sep) \
 283 COLL_APPLY(action, COLL_ALLTOALL_SIG, default) COLL_sep \
 284 COLL_APPLY(action, COLL_ALLTOALL_SIG, 2dmesh) COLL_sep \
 285 COLL_APPLY(action, COLL_ALLTOALL_SIG, 3dmesh) COLL_sep \
 286 COLL_APPLY(action, COLL_ALLTOALL_SIG, basic_linear) COLL_sep \
 287 COLL_APPLY(action, COLL_ALLTOALL_SIG, bruck) COLL_sep \
 288 COLL_APPLY(action, COLL_ALLTOALL_SIG, pair) COLL_sep \
 289 COLL_APPLY(action, COLL_ALLTOALL_SIG, pair_rma) COLL_sep \
 290 COLL_APPLY(action, COLL_ALLTOALL_SIG, pair_light_barrier) COLL_sep \
 291 COLL_APPLY(action, COLL_ALLTOALL_SIG, pair_mpi_barrier) COLL_sep \
 292 COLL_APPLY(action, COLL_ALLTOALL_SIG, pair_one_barrier) COLL_sep \
 293 COLL_APPLY(action, COLL_ALLTOALL_SIG, rdb) COLL_sep \
 294 COLL_APPLY(action, COLL_ALLTOALL_SIG, ring) COLL_sep \
 295 COLL_APPLY(action, COLL_ALLTOALL_SIG, ring_light_barrier) COLL_sep \
 296 COLL_APPLY(action, COLL_ALLTOALL_SIG, ring_mpi_barrier) COLL_sep \
 297 COLL_APPLY(action, COLL_ALLTOALL_SIG, ring_one_barrier) COLL_sep \
 298 COLL_APPLY(action, COLL_ALLTOALL_SIG, mvapich2) COLL_sep \
 299 COLL_APPLY(action, COLL_ALLTOALL_SIG, mvapich2_scatter_dest) COLL_sep \
 300 COLL_APPLY(action, COLL_ALLTOALL_SIG, ompi) COLL_sep \
 301 COLL_APPLY(action, COLL_ALLTOALL_SIG, mpich) COLL_sep \
 302 COLL_APPLY(action, COLL_ALLTOALL_SIG, impi) COLL_sep \
 303 COLL_APPLY(action, COLL_ALLTOALL_SIG, automatic)
 304
 305 COLL_ALLTOALLS(COLL_PROTO, COLL_NOsep)
 306
 307 /*************
 308  * ALLTOALLV *
 309  *************/
 310
 311 #define COLL_ALLTOALLVS(action, COLL_sep) \
 312 COLL_APPLY(action, COLL_ALLTOALLV_SIG, default) COLL_sep \
 313 COLL_APPLY(action, COLL_ALLTOALLV_SIG, bruck) COLL_sep \
 314 COLL_APPLY(action, COLL_ALLTOALLV_SIG, pair) COLL_sep \
 315 COLL_APPLY(action, COLL_ALLTOALLV_SIG, pair_light_barrier) COLL_sep \
 316 COLL_APPLY(action, COLL_ALLTOALLV_SIG, pair_mpi_barrier) COLL_sep \
 317 COLL_APPLY(action, COLL_ALLTOALLV_SIG, pair_one_barrier) COLL_sep \
 318 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ring) COLL_sep \
 319 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ring_light_barrier) COLL_sep \
 320 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ring_mpi_barrier) COLL_sep \
 321 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ring_one_barrier) COLL_sep \
 322 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ompi) COLL_sep \
 323 COLL_APPLY(action, COLL_ALLTOALLV_SIG, mpich) COLL_sep \
 324 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ompi_basic_linear) COLL_sep \
 325 COLL_APPLY(action, COLL_ALLTOALLV_SIG, mvapich2) COLL_sep \
 326 COLL_APPLY(action, COLL_ALLTOALLV_SIG, impi) COLL_sep \
 327 COLL_APPLY(action, COLL_ALLTOALLV_SIG, automatic)
 328
 329 COLL_ALLTOALLVS(COLL_PROTO, COLL_NOsep)
 330
 331 /*********
 332  * BCAST *
 333  *********/
 334
 335 #define COLL_BCASTS(action, COLL_sep) \
 336 COLL_APPLY(action, COLL_BCAST_SIG, default) COLL_sep \
 337 COLL_APPLY(action, COLL_BCAST_SIG, arrival_pattern_aware) COLL_sep \
 338 COLL_APPLY(action, COLL_BCAST_SIG, arrival_pattern_aware_wait) COLL_sep \
 339 COLL_APPLY(action, COLL_BCAST_SIG, arrival_scatter) COLL_sep \
 340 COLL_APPLY(action, COLL_BCAST_SIG, binomial_tree) COLL_sep \
 341 COLL_APPLY(action, COLL_BCAST_SIG, flattree) COLL_sep \
 342 COLL_APPLY(action, COLL_BCAST_SIG, flattree_pipeline) COLL_sep \
 343 COLL_APPLY(action, COLL_BCAST_SIG, NTSB) COLL_sep \
 344 COLL_APPLY(action, COLL_BCAST_SIG, NTSL) COLL_sep \
 345 COLL_APPLY(action, COLL_BCAST_SIG, NTSL_Isend) COLL_sep \
 346 COLL_APPLY(action, COLL_BCAST_SIG, scatter_LR_allgather) COLL_sep \
 347 COLL_APPLY(action, COLL_BCAST_SIG, scatter_rdb_allgather) COLL_sep \
 348 COLL_APPLY(action, COLL_BCAST_SIG, SMP_binary) COLL_sep \
 349 COLL_APPLY(action, COLL_BCAST_SIG, SMP_binomial) COLL_sep \
 350 COLL_APPLY(action, COLL_BCAST_SIG, SMP_linear) COLL_sep \
 351 COLL_APPLY(action, COLL_BCAST_SIG, ompi) COLL_sep \
 352 COLL_APPLY(action, COLL_BCAST_SIG, ompi_split_bintree) COLL_sep \
 353 COLL_APPLY(action, COLL_BCAST_SIG, ompi_pipeline) COLL_sep \
 354 COLL_APPLY(action, COLL_BCAST_SIG, mpich) COLL_sep \
 355 COLL_APPLY(action, COLL_BCAST_SIG, mvapich2)   COLL_sep \
 356 COLL_APPLY(action, COLL_BCAST_SIG, mvapich2_inter_node)   COLL_sep \
 357 COLL_APPLY(action, COLL_BCAST_SIG, mvapich2_intra_node)   COLL_sep \
 358 COLL_APPLY(action, COLL_BCAST_SIG, mvapich2_knomial_intra_node)   COLL_sep \
 359 COLL_APPLY(action, COLL_BCAST_SIG, impi)   COLL_sep \
 360 COLL_APPLY(action, COLL_BCAST_SIG, automatic)
 361
 362 COLL_BCASTS(COLL_PROTO, COLL_NOsep)
 363
 364 /**********
 365  * REDUCE *
 366  **********/
 367
 368 #define COLL_REDUCES(action, COLL_sep) \
 369 COLL_APPLY(action, COLL_REDUCE_SIG, default) COLL_sep \
 370 COLL_APPLY(action, COLL_REDUCE_SIG, arrival_pattern_aware) COLL_sep \
 371 COLL_APPLY(action, COLL_REDUCE_SIG, binomial) COLL_sep \
 372 COLL_APPLY(action, COLL_REDUCE_SIG, flat_tree) COLL_sep \
 373 COLL_APPLY(action, COLL_REDUCE_SIG, NTSL) COLL_sep \
 374 COLL_APPLY(action, COLL_REDUCE_SIG, scatter_gather) COLL_sep \
 375 COLL_APPLY(action, COLL_REDUCE_SIG, ompi) COLL_sep \
 376 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_chain) COLL_sep \
 377 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_pipeline) COLL_sep \
 378 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_basic_linear) COLL_sep \
 379 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_in_order_binary) COLL_sep \
 380 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_binary) COLL_sep \
 381 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_binomial) COLL_sep \
 382 COLL_APPLY(action, COLL_REDUCE_SIG, mpich) COLL_sep \
 383 COLL_APPLY(action, COLL_REDUCE_SIG, mvapich2) COLL_sep \
 384 COLL_APPLY(action, COLL_REDUCE_SIG, mvapich2_knomial) COLL_sep \
 385 COLL_APPLY(action, COLL_REDUCE_SIG, mvapich2_two_level) COLL_sep \
 386 COLL_APPLY(action, COLL_REDUCE_SIG, impi) COLL_sep \
 387 COLL_APPLY(action, COLL_REDUCE_SIG, rab) COLL_sep \
 388 COLL_APPLY(action, COLL_REDUCE_SIG, automatic)
 389
 390 COLL_REDUCES(COLL_PROTO, COLL_NOsep)
 391
 392 /*************
 393  * REDUCE_SCATTER *
 394  *************/
 395
 396 #define COLL_REDUCE_SCATTERS(action, COLL_sep) \
 397 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, default) COLL_sep \
 398 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, ompi) COLL_sep \
 399 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, ompi_basic_recursivehalving) COLL_sep \
 400 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, ompi_ring)  COLL_sep \
 401 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, mpich) COLL_sep \
 402 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, mpich_pair) COLL_sep \
 403 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, mpich_rdb) COLL_sep \
 404 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, mpich_noncomm) COLL_sep \
 405 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, mvapich2) COLL_sep \
 406 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, impi) COLL_sep \
 407 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, automatic)
 408
 409 COLL_REDUCE_SCATTERS(COLL_PROTO, COLL_NOsep)
 410
 411 /*************
 412  * SCATTER *
 413  *************/
 414
 415 #define COLL_SCATTERS(action, COLL_sep) \
 416 COLL_APPLY(action, COLL_SCATTER_SIG, default) COLL_sep \
 417 COLL_APPLY(action, COLL_SCATTER_SIG, ompi) COLL_sep \
 418 COLL_APPLY(action, COLL_SCATTER_SIG, ompi_basic_linear) COLL_sep \
 419 COLL_APPLY(action, COLL_SCATTER_SIG, ompi_binomial)  COLL_sep \
 420 COLL_APPLY(action, COLL_SCATTER_SIG, mpich)   COLL_sep \
 421 COLL_APPLY(action, COLL_SCATTER_SIG, mvapich2)   COLL_sep \
 422 COLL_APPLY(action, COLL_SCATTER_SIG, mvapich2_two_level_binomial)   COLL_sep \
 423 COLL_APPLY(action, COLL_SCATTER_SIG, mvapich2_two_level_direct)   COLL_sep \
 424 COLL_APPLY(action, COLL_SCATTER_SIG, impi)   COLL_sep \
 425 COLL_APPLY(action, COLL_SCATTER_SIG, automatic)
 426
 427 COLL_SCATTERS(COLL_PROTO, COLL_NOsep)
 428
 429 /*************
 430  * BARRIER *
 431  *************/
 432
 433 #define COLL_BARRIERS(action, COLL_sep) \
 434 COLL_APPLY(action, COLL_BARRIER_SIG, default) COLL_sep \
 435 COLL_APPLY(action, COLL_BARRIER_SIG, ompi) COLL_sep \
 436 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_basic_linear) COLL_sep \
 437 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_two_procs)  COLL_sep \
 438 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_tree)  COLL_sep \
 439 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_bruck)  COLL_sep \
 440 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_recursivedoubling) COLL_sep \
 441 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_doublering) COLL_sep \
 442 COLL_APPLY(action, COLL_BARRIER_SIG, mpich_smp)   COLL_sep \
 443 COLL_APPLY(action, COLL_BARRIER_SIG, mpich)   COLL_sep \
 444 COLL_APPLY(action, COLL_BARRIER_SIG, mvapich2_pair)   COLL_sep \
 445 COLL_APPLY(action, COLL_BARRIER_SIG, mvapich2)   COLL_sep \
 446 COLL_APPLY(action, COLL_BARRIER_SIG, impi)   COLL_sep \
 447 COLL_APPLY(action, COLL_BARRIER_SIG, automatic)
 448
 449 COLL_BARRIERS(COLL_PROTO, COLL_NOsep)
 450
 451 }
 452 }
 453 #endif