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   void _XBT_CONCAT(set_, cat)(const std::string& name);                                                                \
  18   extern int(*cat) args;
  19
  20 #define COLL_SIG(cat, ret, args, args2) int cat args;
  21
  22 #define COLL_DESCRIPTION(cat, ret, args, name)                                                                         \
  23   {                                                                                                                    \
  24     _XBT_STRINGIFY(name)                                                                                               \
  25     , _XBT_STRINGIFY(cat) " " _XBT_STRINGIFY(name) " collective", (void*)_XBT_CONCAT3(cat, __, name)        \
  26   }
  27
  28 #define COLL_PROTO(cat, ret, args, name)                                                                               \
  29   ret _XBT_CONCAT3(cat, __, name) args;
  30
  31 #define COLL_UNPAREN(...)  __VA_ARGS__
  32
  33 #define COLL_APPLY(action, sig, name) action(sig, name)
  34 #define COLL_COMMA ,
  35 #define COLL_NOsep
  36 #define COLL_NOTHING(...)
  37
  38 #define COLL_GATHER_SIG gather, int, \
  39                       (const void *send_buff, int send_count, MPI_Datatype send_type, \
  40                        void *recv_buff, int recv_count, MPI_Datatype recv_type, \
  41                            int root, MPI_Comm comm)
  42 #define COLL_ALLGATHER_SIG allgather, int, \
  43                       (const void *send_buff, int send_count, MPI_Datatype send_type, \
  44                        void *recv_buff, int recv_count, MPI_Datatype recv_type, \
  45                            MPI_Comm comm)
  46 #define COLL_ALLGATHERV_SIG allgatherv, int, \
  47                       (const void *send_buff, int send_count, MPI_Datatype send_type, \
  48                        void *recv_buff, const int *recv_count, const int *recv_disps, \
  49                MPI_Datatype recv_type, MPI_Comm comm)
  50 #define COLL_ALLTOALL_SIG alltoall, int, \
  51                      (const void *send_buff, int send_count, MPI_Datatype send_type, \
  52                       void *recv_buff, int recv_count, MPI_Datatype recv_type, \
  53                           MPI_Comm comm)
  54 #define COLL_ALLTOALLV_SIG alltoallv, int, \
  55                      (const void *send_buff, const int *send_counts, const int *send_disps, MPI_Datatype send_type, \
  56                       void *recv_buff, const int *recv_counts, const int *recv_disps, MPI_Datatype recv_type, \
  57                           MPI_Comm comm)
  58 #define COLL_BCAST_SIG bcast, int, \
  59                   (void *buf, int count, MPI_Datatype datatype, \
  60                    int root, MPI_Comm comm)
  61 #define COLL_REDUCE_SIG reduce, int, \
  62                    (const void *buf, void *rbuf, int count, MPI_Datatype datatype, \
  63                         MPI_Op op, int root, MPI_Comm comm)
  64 #define COLL_ALLREDUCE_SIG allreduce, int, \
  65                       (const void *sbuf, void *rbuf, int rcount, \
  66                            MPI_Datatype dtype, MPI_Op op, MPI_Comm comm)
  67 #define COLL_REDUCE_SCATTER_SIG reduce_scatter, int, \
  68                       (const void *sbuf, void *rbuf, const int *rcounts,\
  69                     MPI_Datatype dtype,MPI_Op  op,MPI_Comm  comm)
  70 #define COLL_SCATTER_SIG scatter, int, \
  71                 (const void *sendbuf, int sendcount, MPI_Datatype sendtype,\
  72                 void *recvbuf, int recvcount, MPI_Datatype recvtype,\
  73                 int root, MPI_Comm comm)
  74 #define COLL_BARRIER_SIG barrier, int, \
  75                 (MPI_Comm comm)
  76
  77 namespace simgrid{
  78 namespace smpi{
  79
  80 struct s_mpi_coll_description_t {
  81   std::string name;
  82   std::string description;
  83   void *coll;
  84 };
  85
  86 namespace colls {
  87 XBT_PUBLIC void coll_help(const char* category, s_mpi_coll_description_t* table);
  88 XBT_PUBLIC int find_coll_description(s_mpi_coll_description_t* table, const std::string& name, const char* desc);
  89 void set_collectives();
  90 XBT_PUBLIC s_mpi_coll_description_t* get_smpi_coll_description(const char* name, int rank);
  91
  92 // for each collective type, create the set_* prototype, the description array and the function pointer
  93 //  void set_gather(const std::string& name);
  94 //  extern int(*gather)(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count,
  95 //  MPI_Datatype recv_type,
  96 //                      int root, MPI_Comm comm);
  97 COLL_APPLY(COLL_DEFS, COLL_GATHER_SIG, "")
  98 COLL_APPLY(COLL_DEFS, COLL_ALLGATHER_SIG, "")
  99 COLL_APPLY(COLL_DEFS, COLL_ALLGATHERV_SIG, "")
 100 COLL_APPLY(COLL_DEFS, COLL_REDUCE_SIG, "")
 101 COLL_APPLY(COLL_DEFS, COLL_ALLREDUCE_SIG, "")
 102 COLL_APPLY(COLL_DEFS, COLL_REDUCE_SCATTER_SIG, "")
 103 COLL_APPLY(COLL_DEFS, COLL_SCATTER_SIG, "")
 104 COLL_APPLY(COLL_DEFS, COLL_BARRIER_SIG, "")
 105 COLL_APPLY(COLL_DEFS, COLL_BCAST_SIG, "")
 106 COLL_APPLY(COLL_DEFS, COLL_ALLTOALL_SIG, "")
 107 COLL_APPLY(COLL_DEFS, COLL_ALLTOALLV_SIG, "")
 108
 109 // These fairly unused collectives only have one implementation in SMPI
 110 int gatherv(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, const int* recvcounts,
 111             const int* displs, MPI_Datatype recvtype, int root, MPI_Comm comm);
 112 int scatterv(const void* sendbuf, const int* sendcounts, const int* displs, MPI_Datatype sendtype, void* recvbuf,
 113              int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm);
 114 int scan(const void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm);
 115 int exscan(const void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm);
 116 int alltoallw(const void* sendbuf, const int* sendcounts, const int* senddisps, const MPI_Datatype* sendtypes,
 117               void* recvbuf, const int* recvcounts, const int* recvdisps, const MPI_Datatype* recvtypes, MPI_Comm comm);
 118
 119 // async collectives
 120 int ibarrier(MPI_Comm comm, MPI_Request* request, int external = 1);
 121 int ibcast(void* buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm, MPI_Request* request,
 122            int external = 1);
 123 int igather(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount,
 124             MPI_Datatype recvtype, int root, MPI_Comm comm, MPI_Request* request, int external = 1);
 125 int igatherv(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, const int* recvcounts,
 126              const int* displs, MPI_Datatype recvtype, int root, MPI_Comm comm, MPI_Request* request, int external = 1);
 127 int iallgather(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount,
 128                MPI_Datatype recvtype, MPI_Comm comm, MPI_Request* request, int external = 1);
 129 int iallgatherv(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, const int* recvcounts,
 130                 const int* displs, MPI_Datatype recvtype, MPI_Comm comm, MPI_Request* request, int external = 1);
 131 int iscatter(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount,
 132              MPI_Datatype recvtype, int root, MPI_Comm comm, MPI_Request* request, int external = 1);
 133 int iscatterv(const void* sendbuf, const int* sendcounts, const int* displs, MPI_Datatype sendtype, void* recvbuf,
 134               int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm, MPI_Request* request, int external = 1);
 135 int ireduce(const void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm,
 136             MPI_Request* request, int external = 1);
 137 int iallreduce(const void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm,
 138                MPI_Request* request, int external = 1);
 139 int iscan(const void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm,
 140           MPI_Request* request, int external = 1);
 141 int iexscan(const void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm,
 142             MPI_Request* request, int external = 1);
 143 int ireduce_scatter(const void* sendbuf, void* recvbuf, const int* recvcounts, MPI_Datatype datatype, MPI_Op op,
 144                     MPI_Comm comm, MPI_Request* request, int external = 1);
 145 int ireduce_scatter_block(const void* sendbuf, void* recvbuf, int recvcount, MPI_Datatype datatype, MPI_Op op,
 146                           MPI_Comm comm, MPI_Request* request, int external = 1);
 147 int ialltoall(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount,
 148               MPI_Datatype recvtype, MPI_Comm comm, MPI_Request* request, int external = 1);
 149 int ialltoallv(const void* sendbuf, const int* sendcounts, const int* senddisps, MPI_Datatype sendtype, void* recvbuf,
 150                const int* recvcounts, const int* recvdisps, MPI_Datatype recvtype, MPI_Comm comm, MPI_Request* request,
 151                int external = 1);
 152 int ialltoallw(const void* sendbuf, const int* sendcounts, const int* senddisps, const MPI_Datatype* sendtypes,
 153                void* recvbuf, const int* recvcounts, const int* recvdisps, const MPI_Datatype* recvtypes, MPI_Comm comm,
 154                MPI_Request* request, int external = 1);
 155
 156 extern void (*smpi_coll_cleanup_callback)();
 157 };
 158
 159 /*************
 160  * GATHER *
 161  *************/
 162
 163 #define COLL_GATHERS(action, COLL_sep) \
 164 COLL_APPLY(action, COLL_GATHER_SIG, default) COLL_sep \
 165 COLL_APPLY(action, COLL_GATHER_SIG, ompi) COLL_sep \
 166 COLL_APPLY(action, COLL_GATHER_SIG, ompi_basic_linear) COLL_sep \
 167 COLL_APPLY(action, COLL_GATHER_SIG, ompi_binomial) COLL_sep \
 168 COLL_APPLY(action, COLL_GATHER_SIG, ompi_linear_sync) COLL_sep \
 169 COLL_APPLY(action, COLL_GATHER_SIG, mpich) COLL_sep \
 170 COLL_APPLY(action, COLL_GATHER_SIG, mvapich2) COLL_sep \
 171 COLL_APPLY(action, COLL_GATHER_SIG, mvapich2_two_level) COLL_sep \
 172 COLL_APPLY(action, COLL_GATHER_SIG, impi) COLL_sep \
 173 COLL_APPLY(action, COLL_GATHER_SIG, automatic)
 174
 175 COLL_GATHERS(COLL_PROTO, COLL_NOsep)
 176
 177 /*************
 178  * ALLGATHER *
 179  *************/
 180
 181 #define COLL_ALLGATHERS(action, COLL_sep) \
 182 COLL_APPLY(action, COLL_ALLGATHER_SIG, default) COLL_sep \
 183 COLL_APPLY(action, COLL_ALLGATHER_SIG, 2dmesh) COLL_sep \
 184 COLL_APPLY(action, COLL_ALLGATHER_SIG, 3dmesh) COLL_sep \
 185 COLL_APPLY(action, COLL_ALLGATHER_SIG, bruck) COLL_sep \
 186 COLL_APPLY(action, COLL_ALLGATHER_SIG, GB) COLL_sep \
 187 COLL_APPLY(action, COLL_ALLGATHER_SIG, loosely_lr) COLL_sep \
 188 COLL_APPLY(action, COLL_ALLGATHER_SIG, NTSLR) COLL_sep \
 189 COLL_APPLY(action, COLL_ALLGATHER_SIG, NTSLR_NB) COLL_sep \
 190 COLL_APPLY(action, COLL_ALLGATHER_SIG, pair) COLL_sep \
 191 COLL_APPLY(action, COLL_ALLGATHER_SIG, rdb) COLL_sep \
 192 COLL_APPLY(action, COLL_ALLGATHER_SIG, rhv) COLL_sep \
 193 COLL_APPLY(action, COLL_ALLGATHER_SIG, ring) COLL_sep \
 194 COLL_APPLY(action, COLL_ALLGATHER_SIG, SMP_NTS) COLL_sep \
 195 COLL_APPLY(action, COLL_ALLGATHER_SIG, smp_simple) COLL_sep \
 196 COLL_APPLY(action, COLL_ALLGATHER_SIG, spreading_simple) COLL_sep \
 197 COLL_APPLY(action, COLL_ALLGATHER_SIG, ompi) COLL_sep \
 198 COLL_APPLY(action, COLL_ALLGATHER_SIG, ompi_neighborexchange) COLL_sep \
 199 COLL_APPLY(action, COLL_ALLGATHER_SIG, mvapich2) COLL_sep \
 200 COLL_APPLY(action, COLL_ALLGATHER_SIG, mvapich2_smp) COLL_sep \
 201 COLL_APPLY(action, COLL_ALLGATHER_SIG, mpich) COLL_sep \
 202 COLL_APPLY(action, COLL_ALLGATHER_SIG, impi) COLL_sep \
 203 COLL_APPLY(action, COLL_ALLGATHER_SIG, automatic)
 204
 205 COLL_ALLGATHERS(COLL_PROTO, COLL_NOsep)
 206
 207 /**************
 208  * ALLGATHERV *
 209  **************/
 210
 211 #define COLL_ALLGATHERVS(action, COLL_sep) \
 212 COLL_APPLY(action, COLL_ALLGATHERV_SIG, default) COLL_sep \
 213 COLL_APPLY(action, COLL_ALLGATHERV_SIG, GB) COLL_sep \
 214 COLL_APPLY(action, COLL_ALLGATHERV_SIG, pair) COLL_sep \
 215 COLL_APPLY(action, COLL_ALLGATHERV_SIG, ring) COLL_sep \
 216 COLL_APPLY(action, COLL_ALLGATHERV_SIG, ompi) COLL_sep \
 217 COLL_APPLY(action, COLL_ALLGATHERV_SIG, ompi_neighborexchange) COLL_sep \
 218 COLL_APPLY(action, COLL_ALLGATHERV_SIG, ompi_bruck) COLL_sep \
 219 COLL_APPLY(action, COLL_ALLGATHERV_SIG, mpich) COLL_sep \
 220 COLL_APPLY(action, COLL_ALLGATHERV_SIG, mpich_rdb) COLL_sep \
 221 COLL_APPLY(action, COLL_ALLGATHERV_SIG, mpich_ring) COLL_sep \
 222 COLL_APPLY(action, COLL_ALLGATHERV_SIG, mvapich2) COLL_sep \
 223 COLL_APPLY(action, COLL_ALLGATHERV_SIG, impi) COLL_sep \
 224 COLL_APPLY(action, COLL_ALLGATHERV_SIG, automatic)
 225
 226 COLL_ALLGATHERVS(COLL_PROTO, COLL_NOsep)
 227
 228 /*************
 229  * ALLREDUCE *
 230  *************/
 231
 232 #define COLL_ALLREDUCES(action, COLL_sep) \
 233 COLL_APPLY(action, COLL_ALLREDUCE_SIG, default) COLL_sep \
 234 COLL_APPLY(action, COLL_ALLREDUCE_SIG, lr) COLL_sep \
 235 COLL_APPLY(action, COLL_ALLREDUCE_SIG, rab1) COLL_sep \
 236 COLL_APPLY(action, COLL_ALLREDUCE_SIG, rab2) COLL_sep \
 237 COLL_APPLY(action, COLL_ALLREDUCE_SIG, rab_rdb) COLL_sep \
 238 COLL_APPLY(action, COLL_ALLREDUCE_SIG, rdb) COLL_sep \
 239 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_binomial) COLL_sep \
 240 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_binomial_pipeline) COLL_sep \
 241 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_rdb) COLL_sep \
 242 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_rsag) COLL_sep \
 243 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_rsag_lr) COLL_sep \
 244 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_rsag_rab) COLL_sep \
 245 COLL_APPLY(action, COLL_ALLREDUCE_SIG, redbcast) COLL_sep \
 246 COLL_APPLY(action, COLL_ALLREDUCE_SIG, ompi) COLL_sep \
 247 COLL_APPLY(action, COLL_ALLREDUCE_SIG, ompi_ring_segmented) COLL_sep \
 248 COLL_APPLY(action, COLL_ALLREDUCE_SIG, mpich) COLL_sep \
 249 COLL_APPLY(action, COLL_ALLREDUCE_SIG, mvapich2) COLL_sep \
 250 COLL_APPLY(action, COLL_ALLREDUCE_SIG, mvapich2_rs) COLL_sep \
 251 COLL_APPLY(action, COLL_ALLREDUCE_SIG, mvapich2_two_level) COLL_sep \
 252 COLL_APPLY(action, COLL_ALLREDUCE_SIG, impi) COLL_sep \
 253 COLL_APPLY(action, COLL_ALLREDUCE_SIG, rab) COLL_sep \
 254 COLL_APPLY(action, COLL_ALLREDUCE_SIG, automatic)
 255
 256 COLL_ALLREDUCES(COLL_PROTO, COLL_NOsep)
 257
 258 /************
 259  * ALLTOALL *
 260  ************/
 261
 262 #define COLL_ALLTOALLS(action, COLL_sep) \
 263 COLL_APPLY(action, COLL_ALLTOALL_SIG, default) COLL_sep \
 264 COLL_APPLY(action, COLL_ALLTOALL_SIG, 2dmesh) COLL_sep \
 265 COLL_APPLY(action, COLL_ALLTOALL_SIG, 3dmesh) COLL_sep \
 266 COLL_APPLY(action, COLL_ALLTOALL_SIG, basic_linear) COLL_sep \
 267 COLL_APPLY(action, COLL_ALLTOALL_SIG, bruck) COLL_sep \
 268 COLL_APPLY(action, COLL_ALLTOALL_SIG, pair) COLL_sep \
 269 COLL_APPLY(action, COLL_ALLTOALL_SIG, pair_rma) COLL_sep \
 270 COLL_APPLY(action, COLL_ALLTOALL_SIG, pair_light_barrier) COLL_sep \
 271 COLL_APPLY(action, COLL_ALLTOALL_SIG, pair_mpi_barrier) COLL_sep \
 272 COLL_APPLY(action, COLL_ALLTOALL_SIG, pair_one_barrier) COLL_sep \
 273 COLL_APPLY(action, COLL_ALLTOALL_SIG, rdb) COLL_sep \
 274 COLL_APPLY(action, COLL_ALLTOALL_SIG, ring) COLL_sep \
 275 COLL_APPLY(action, COLL_ALLTOALL_SIG, ring_light_barrier) COLL_sep \
 276 COLL_APPLY(action, COLL_ALLTOALL_SIG, ring_mpi_barrier) COLL_sep \
 277 COLL_APPLY(action, COLL_ALLTOALL_SIG, ring_one_barrier) COLL_sep \
 278 COLL_APPLY(action, COLL_ALLTOALL_SIG, mvapich2) COLL_sep \
 279 COLL_APPLY(action, COLL_ALLTOALL_SIG, mvapich2_scatter_dest) COLL_sep \
 280 COLL_APPLY(action, COLL_ALLTOALL_SIG, ompi) COLL_sep \
 281 COLL_APPLY(action, COLL_ALLTOALL_SIG, mpich) COLL_sep \
 282 COLL_APPLY(action, COLL_ALLTOALL_SIG, impi) COLL_sep \
 283 COLL_APPLY(action, COLL_ALLTOALL_SIG, automatic)
 284
 285 COLL_ALLTOALLS(COLL_PROTO, COLL_NOsep)
 286
 287 /*************
 288  * ALLTOALLV *
 289  *************/
 290
 291 #define COLL_ALLTOALLVS(action, COLL_sep) \
 292 COLL_APPLY(action, COLL_ALLTOALLV_SIG, default) COLL_sep \
 293 COLL_APPLY(action, COLL_ALLTOALLV_SIG, bruck) COLL_sep \
 294 COLL_APPLY(action, COLL_ALLTOALLV_SIG, pair) COLL_sep \
 295 COLL_APPLY(action, COLL_ALLTOALLV_SIG, pair_light_barrier) COLL_sep \
 296 COLL_APPLY(action, COLL_ALLTOALLV_SIG, pair_mpi_barrier) COLL_sep \
 297 COLL_APPLY(action, COLL_ALLTOALLV_SIG, pair_one_barrier) COLL_sep \
 298 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ring) COLL_sep \
 299 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ring_light_barrier) COLL_sep \
 300 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ring_mpi_barrier) COLL_sep \
 301 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ring_one_barrier) COLL_sep \
 302 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ompi) COLL_sep \
 303 COLL_APPLY(action, COLL_ALLTOALLV_SIG, mpich) COLL_sep \
 304 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ompi_basic_linear) COLL_sep \
 305 COLL_APPLY(action, COLL_ALLTOALLV_SIG, mvapich2) COLL_sep \
 306 COLL_APPLY(action, COLL_ALLTOALLV_SIG, impi) COLL_sep \
 307 COLL_APPLY(action, COLL_ALLTOALLV_SIG, automatic)
 308
 309 COLL_ALLTOALLVS(COLL_PROTO, COLL_NOsep)
 310
 311 /*********
 312  * BCAST *
 313  *********/
 314
 315 #define COLL_BCASTS(action, COLL_sep) \
 316 COLL_APPLY(action, COLL_BCAST_SIG, default) COLL_sep \
 317 COLL_APPLY(action, COLL_BCAST_SIG, arrival_pattern_aware) COLL_sep \
 318 COLL_APPLY(action, COLL_BCAST_SIG, arrival_pattern_aware_wait) COLL_sep \
 319 COLL_APPLY(action, COLL_BCAST_SIG, arrival_scatter) COLL_sep \
 320 COLL_APPLY(action, COLL_BCAST_SIG, binomial_tree) COLL_sep \
 321 COLL_APPLY(action, COLL_BCAST_SIG, flattree) COLL_sep \
 322 COLL_APPLY(action, COLL_BCAST_SIG, flattree_pipeline) COLL_sep \
 323 COLL_APPLY(action, COLL_BCAST_SIG, NTSB) COLL_sep \
 324 COLL_APPLY(action, COLL_BCAST_SIG, NTSL) COLL_sep \
 325 COLL_APPLY(action, COLL_BCAST_SIG, NTSL_Isend) COLL_sep \
 326 COLL_APPLY(action, COLL_BCAST_SIG, scatter_LR_allgather) COLL_sep \
 327 COLL_APPLY(action, COLL_BCAST_SIG, scatter_rdb_allgather) COLL_sep \
 328 COLL_APPLY(action, COLL_BCAST_SIG, SMP_binary) COLL_sep \
 329 COLL_APPLY(action, COLL_BCAST_SIG, SMP_binomial) COLL_sep \
 330 COLL_APPLY(action, COLL_BCAST_SIG, SMP_linear) COLL_sep \
 331 COLL_APPLY(action, COLL_BCAST_SIG, ompi) COLL_sep \
 332 COLL_APPLY(action, COLL_BCAST_SIG, ompi_split_bintree) COLL_sep \
 333 COLL_APPLY(action, COLL_BCAST_SIG, ompi_pipeline) COLL_sep \
 334 COLL_APPLY(action, COLL_BCAST_SIG, mpich) COLL_sep \
 335 COLL_APPLY(action, COLL_BCAST_SIG, mvapich2)   COLL_sep \
 336 COLL_APPLY(action, COLL_BCAST_SIG, mvapich2_inter_node)   COLL_sep \
 337 COLL_APPLY(action, COLL_BCAST_SIG, mvapich2_intra_node)   COLL_sep \
 338 COLL_APPLY(action, COLL_BCAST_SIG, mvapich2_knomial_intra_node)   COLL_sep \
 339 COLL_APPLY(action, COLL_BCAST_SIG, impi)   COLL_sep \
 340 COLL_APPLY(action, COLL_BCAST_SIG, automatic)
 341
 342 COLL_BCASTS(COLL_PROTO, COLL_NOsep)
 343
 344 /**********
 345  * REDUCE *
 346  **********/
 347
 348 #define COLL_REDUCES(action, COLL_sep) \
 349 COLL_APPLY(action, COLL_REDUCE_SIG, default) COLL_sep \
 350 COLL_APPLY(action, COLL_REDUCE_SIG, arrival_pattern_aware) COLL_sep \
 351 COLL_APPLY(action, COLL_REDUCE_SIG, binomial) COLL_sep \
 352 COLL_APPLY(action, COLL_REDUCE_SIG, flat_tree) COLL_sep \
 353 COLL_APPLY(action, COLL_REDUCE_SIG, NTSL) COLL_sep \
 354 COLL_APPLY(action, COLL_REDUCE_SIG, scatter_gather) COLL_sep \
 355 COLL_APPLY(action, COLL_REDUCE_SIG, ompi) COLL_sep \
 356 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_chain) COLL_sep \
 357 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_pipeline) COLL_sep \
 358 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_basic_linear) COLL_sep \
 359 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_in_order_binary) COLL_sep \
 360 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_binary) COLL_sep \
 361 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_binomial) COLL_sep \
 362 COLL_APPLY(action, COLL_REDUCE_SIG, mpich) COLL_sep \
 363 COLL_APPLY(action, COLL_REDUCE_SIG, mvapich2) COLL_sep \
 364 COLL_APPLY(action, COLL_REDUCE_SIG, mvapich2_knomial) COLL_sep \
 365 COLL_APPLY(action, COLL_REDUCE_SIG, mvapich2_two_level) COLL_sep \
 366 COLL_APPLY(action, COLL_REDUCE_SIG, impi) COLL_sep \
 367 COLL_APPLY(action, COLL_REDUCE_SIG, rab) COLL_sep \
 368 COLL_APPLY(action, COLL_REDUCE_SIG, automatic)
 369
 370 COLL_REDUCES(COLL_PROTO, COLL_NOsep)
 371
 372 /*************
 373  * REDUCE_SCATTER *
 374  *************/
 375
 376 #define COLL_REDUCE_SCATTERS(action, COLL_sep) \
 377 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, default) COLL_sep \
 378 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, ompi) COLL_sep \
 379 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, ompi_basic_recursivehalving) COLL_sep \
 380 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, ompi_ring)  COLL_sep \
 381 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, mpich) COLL_sep \
 382 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, mpich_pair) COLL_sep \
 383 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, mpich_rdb) COLL_sep \
 384 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, mpich_noncomm) COLL_sep \
 385 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, mvapich2) COLL_sep \
 386 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, impi) COLL_sep \
 387 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, automatic)
 388
 389 COLL_REDUCE_SCATTERS(COLL_PROTO, COLL_NOsep)
 390
 391 /*************
 392  * SCATTER *
 393  *************/
 394
 395 #define COLL_SCATTERS(action, COLL_sep) \
 396 COLL_APPLY(action, COLL_SCATTER_SIG, default) COLL_sep \
 397 COLL_APPLY(action, COLL_SCATTER_SIG, ompi) COLL_sep \
 398 COLL_APPLY(action, COLL_SCATTER_SIG, ompi_basic_linear) COLL_sep \
 399 COLL_APPLY(action, COLL_SCATTER_SIG, ompi_binomial)  COLL_sep \
 400 COLL_APPLY(action, COLL_SCATTER_SIG, mpich)   COLL_sep \
 401 COLL_APPLY(action, COLL_SCATTER_SIG, mvapich2)   COLL_sep \
 402 COLL_APPLY(action, COLL_SCATTER_SIG, mvapich2_two_level_binomial)   COLL_sep \
 403 COLL_APPLY(action, COLL_SCATTER_SIG, mvapich2_two_level_direct)   COLL_sep \
 404 COLL_APPLY(action, COLL_SCATTER_SIG, impi)   COLL_sep \
 405 COLL_APPLY(action, COLL_SCATTER_SIG, automatic)
 406
 407 COLL_SCATTERS(COLL_PROTO, COLL_NOsep)
 408
 409 /*************
 410  * BARRIER *
 411  *************/
 412
 413 #define COLL_BARRIERS(action, COLL_sep) \
 414 COLL_APPLY(action, COLL_BARRIER_SIG, default) COLL_sep \
 415 COLL_APPLY(action, COLL_BARRIER_SIG, ompi) COLL_sep \
 416 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_basic_linear) COLL_sep \
 417 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_two_procs)  COLL_sep \
 418 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_tree)  COLL_sep \
 419 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_bruck)  COLL_sep \
 420 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_recursivedoubling) COLL_sep \
 421 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_doublering) COLL_sep \
 422 COLL_APPLY(action, COLL_BARRIER_SIG, mpich_smp)   COLL_sep \
 423 COLL_APPLY(action, COLL_BARRIER_SIG, mpich)   COLL_sep \
 424 COLL_APPLY(action, COLL_BARRIER_SIG, mvapich2_pair)   COLL_sep \
 425 COLL_APPLY(action, COLL_BARRIER_SIG, mvapich2)   COLL_sep \
 426 COLL_APPLY(action, COLL_BARRIER_SIG, impi)   COLL_sep \
 427 COLL_APPLY(action, COLL_BARRIER_SIG, automatic)
 428
 429 COLL_BARRIERS(COLL_PROTO, COLL_NOsep)
 430
 431 }
 432 }
 433 #endif