1 /*High level handling of collective algorithms*/
2 /* Copyright (c) 2009-2019. The SimGrid Team.
3 * All rights reserved. */
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. */
11 #include "private.hpp"
14 /** @brief MPI collective description */
16 #define COLL_DEFS(cat, ret, args, args2) \
17 void _XBT_CONCAT(set_, cat)(const std::string& name); \
18 extern int(*cat) args;
20 #define COLL_DESCRIPTION(cat, ret, args, name) \
22 _XBT_STRINGIFY(name) \
23 , _XBT_STRINGIFY(cat) " " _XBT_STRINGIFY(name) " collective", (void*)_XBT_CONCAT3(cat, __, name) \
26 #define COLL_UNPAREN(...) __VA_ARGS__
28 #define COLL_APPLY(action, sig, name) action(sig, name)
31 #define COLL_GATHER_SIG gather, int, \
32 (const void *send_buff, int send_count, MPI_Datatype send_type, \
33 void *recv_buff, int recv_count, MPI_Datatype recv_type, \
34 int root, MPI_Comm comm)
35 #define COLL_ALLGATHER_SIG allgather, int, \
36 (const void *send_buff, int send_count, MPI_Datatype send_type, \
37 void *recv_buff, int recv_count, MPI_Datatype recv_type, \
39 #define COLL_ALLGATHERV_SIG allgatherv, int, \
40 (const void *send_buff, int send_count, MPI_Datatype send_type, \
41 void *recv_buff, const int *recv_count, const int *recv_disps, \
42 MPI_Datatype recv_type, MPI_Comm comm)
43 #define COLL_ALLTOALL_SIG alltoall, int, \
44 (const void *send_buff, int send_count, MPI_Datatype send_type, \
45 void *recv_buff, int recv_count, MPI_Datatype recv_type, \
47 #define COLL_ALLTOALLV_SIG alltoallv, int, \
48 (const void *send_buff, const int *send_counts, const int *send_disps, MPI_Datatype send_type, \
49 void *recv_buff, const int *recv_counts, const int *recv_disps, MPI_Datatype recv_type, \
51 #define COLL_BCAST_SIG bcast, int, \
52 (void *buf, int count, MPI_Datatype datatype, \
53 int root, MPI_Comm comm)
54 #define COLL_REDUCE_SIG reduce, int, \
55 (const void *buf, void *rbuf, int count, MPI_Datatype datatype, \
56 MPI_Op op, int root, MPI_Comm comm)
57 #define COLL_ALLREDUCE_SIG allreduce, int, \
58 (const void *sbuf, void *rbuf, int rcount, \
59 MPI_Datatype dtype, MPI_Op op, MPI_Comm comm)
60 #define COLL_REDUCE_SCATTER_SIG reduce_scatter, int, \
61 (const void *sbuf, void *rbuf, const int *rcounts,\
62 MPI_Datatype dtype,MPI_Op op,MPI_Comm comm)
63 #define COLL_SCATTER_SIG scatter, int, \
64 (const void *sendbuf, int sendcount, MPI_Datatype sendtype,\
65 void *recvbuf, int recvcount, MPI_Datatype recvtype,\
66 int root, MPI_Comm comm)
67 #define COLL_BARRIER_SIG barrier, int, \
73 struct s_mpi_coll_description_t {
75 std::string description;
80 XBT_PUBLIC void coll_help(const char* category, s_mpi_coll_description_t* table);
81 XBT_PUBLIC int find_coll_description(s_mpi_coll_description_t* table, const std::string& name, const char* desc);
82 void set_collectives();
83 XBT_PUBLIC s_mpi_coll_description_t* get_smpi_coll_description(const char* name, int rank);
85 // for each collective type, create the set_* prototype and the function pointer
86 // void set_gather(const std::string& name);
87 // extern int(*gather)(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count,
88 // MPI_Datatype recv_type, int root, MPI_Comm comm);
89 COLL_APPLY(COLL_DEFS, COLL_GATHER_SIG, "")
90 COLL_APPLY(COLL_DEFS, COLL_ALLGATHER_SIG, "")
91 COLL_APPLY(COLL_DEFS, COLL_ALLGATHERV_SIG, "")
92 COLL_APPLY(COLL_DEFS, COLL_REDUCE_SIG, "")
93 COLL_APPLY(COLL_DEFS, COLL_ALLREDUCE_SIG, "")
94 COLL_APPLY(COLL_DEFS, COLL_REDUCE_SCATTER_SIG, "")
95 COLL_APPLY(COLL_DEFS, COLL_SCATTER_SIG, "")
96 COLL_APPLY(COLL_DEFS, COLL_BARRIER_SIG, "")
97 COLL_APPLY(COLL_DEFS, COLL_BCAST_SIG, "")
98 COLL_APPLY(COLL_DEFS, COLL_ALLTOALL_SIG, "")
99 COLL_APPLY(COLL_DEFS, COLL_ALLTOALLV_SIG, "")
101 // These fairly unused collectives only have one implementation in SMPI
102 int gatherv(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, const int* recvcounts,
103 const int* displs, MPI_Datatype recvtype, int root, MPI_Comm comm);
104 int scatterv(const void* sendbuf, const int* sendcounts, const int* displs, MPI_Datatype sendtype, void* recvbuf,
105 int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm);
106 int scan(const void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm);
107 int exscan(const void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm);
108 int alltoallw(const void* sendbuf, const int* sendcounts, const int* senddisps, const MPI_Datatype* sendtypes,
109 void* recvbuf, const int* recvcounts, const int* recvdisps, const MPI_Datatype* recvtypes, MPI_Comm comm);
112 int ibarrier(MPI_Comm comm, MPI_Request* request, int external = 1);
113 int ibcast(void* buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm, MPI_Request* request,
115 int igather(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount,
116 MPI_Datatype recvtype, int root, MPI_Comm comm, MPI_Request* request, int external = 1);
117 int igatherv(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, const int* recvcounts,
118 const int* displs, MPI_Datatype recvtype, int root, MPI_Comm comm, MPI_Request* request, int external = 1);
119 int iallgather(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount,
120 MPI_Datatype recvtype, MPI_Comm comm, MPI_Request* request, int external = 1);
121 int iallgatherv(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, const int* recvcounts,
122 const int* displs, MPI_Datatype recvtype, MPI_Comm comm, MPI_Request* request, int external = 1);
123 int iscatter(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 iscatterv(const void* sendbuf, const int* sendcounts, const int* displs, MPI_Datatype sendtype, void* recvbuf,
126 int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm, MPI_Request* request, int external = 1);
127 int ireduce(const void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm,
128 MPI_Request* request, int external = 1);
129 int iallreduce(const void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm,
130 MPI_Request* request, int external = 1);
131 int iscan(const void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm,
132 MPI_Request* request, int external = 1);
133 int iexscan(const void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm,
134 MPI_Request* request, int external = 1);
135 int ireduce_scatter(const void* sendbuf, void* recvbuf, const int* recvcounts, MPI_Datatype datatype, MPI_Op op,
136 MPI_Comm comm, MPI_Request* request, int external = 1);
137 int ireduce_scatter_block(const void* sendbuf, void* recvbuf, int recvcount, MPI_Datatype datatype, MPI_Op op,
138 MPI_Comm comm, MPI_Request* request, int external = 1);
139 int ialltoall(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount,
140 MPI_Datatype recvtype, MPI_Comm comm, MPI_Request* request, int external = 1);
141 int ialltoallv(const void* sendbuf, const int* sendcounts, const int* senddisps, MPI_Datatype sendtype, void* recvbuf,
142 const int* recvcounts, const int* recvdisps, MPI_Datatype recvtype, MPI_Comm comm, MPI_Request* request,
144 int ialltoallw(const void* sendbuf, const int* sendcounts, const int* senddisps, const MPI_Datatype* sendtypes,
145 void* recvbuf, const int* recvcounts, const int* recvdisps, const MPI_Datatype* recvtypes, MPI_Comm comm,
146 MPI_Request* request, int external = 1);
148 extern void (*smpi_coll_cleanup_callback)();
155 int gather__default(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, int root, MPI_Comm comm);
156 int gather__ompi(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, int root, MPI_Comm comm);
157 int gather__ompi_basic_linear(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, int root, MPI_Comm comm);
158 int gather__ompi_binomial(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, int root, MPI_Comm comm);
159 int gather__ompi_linear_sync(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, int root, MPI_Comm comm);
160 int gather__mpich(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, int root, MPI_Comm comm);
161 int gather__mvapich2(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, int root, MPI_Comm comm);
162 int gather__mvapich2_two_level(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, int root, MPI_Comm comm);
163 int gather__impi(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, int root, MPI_Comm comm);
164 int gather__automatic(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, int root, MPI_Comm comm);
166 #define COLL_GATHERS(action, COLL_sep) \
167 COLL_APPLY(action, COLL_GATHER_SIG, default) COLL_sep \
168 COLL_APPLY(action, COLL_GATHER_SIG, ompi) COLL_sep \
169 COLL_APPLY(action, COLL_GATHER_SIG, ompi_basic_linear) COLL_sep \
170 COLL_APPLY(action, COLL_GATHER_SIG, ompi_binomial) COLL_sep \
171 COLL_APPLY(action, COLL_GATHER_SIG, ompi_linear_sync) COLL_sep \
172 COLL_APPLY(action, COLL_GATHER_SIG, mpich) COLL_sep \
173 COLL_APPLY(action, COLL_GATHER_SIG, mvapich2) COLL_sep \
174 COLL_APPLY(action, COLL_GATHER_SIG, mvapich2_two_level) COLL_sep \
175 COLL_APPLY(action, COLL_GATHER_SIG, impi) COLL_sep \
176 COLL_APPLY(action, COLL_GATHER_SIG, automatic)
181 int allgather__default(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
182 int allgather__2dmesh(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
183 int allgather__3dmesh(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
184 int allgather__bruck(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
185 int allgather__GB(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
186 int allgather__loosely_lr(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
187 int allgather__NTSLR(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
188 int allgather__NTSLR_NB(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
189 int allgather__pair(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
190 int allgather__rdb(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
191 int allgather__rhv(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
192 int allgather__ring(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
193 int allgather__SMP_NTS(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
194 int allgather__smp_simple(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
195 int allgather__spreading_simple(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
196 int allgather__ompi(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
197 int allgather__ompi_neighborexchange (const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
198 int allgather__mvapich2(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
199 int allgather__mvapich2_smp(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
200 int allgather__mpich(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
201 int allgather__impi(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
202 int allgather__automatic(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
204 #define COLL_ALLGATHERS(action, COLL_sep) \
205 COLL_APPLY(action, COLL_ALLGATHER_SIG, default) COLL_sep \
206 COLL_APPLY(action, COLL_ALLGATHER_SIG, 2dmesh) COLL_sep \
207 COLL_APPLY(action, COLL_ALLGATHER_SIG, 3dmesh) COLL_sep \
208 COLL_APPLY(action, COLL_ALLGATHER_SIG, bruck) COLL_sep \
209 COLL_APPLY(action, COLL_ALLGATHER_SIG, GB) COLL_sep \
210 COLL_APPLY(action, COLL_ALLGATHER_SIG, loosely_lr) COLL_sep \
211 COLL_APPLY(action, COLL_ALLGATHER_SIG, NTSLR) COLL_sep \
212 COLL_APPLY(action, COLL_ALLGATHER_SIG, NTSLR_NB) COLL_sep \
213 COLL_APPLY(action, COLL_ALLGATHER_SIG, pair) COLL_sep \
214 COLL_APPLY(action, COLL_ALLGATHER_SIG, rdb) COLL_sep \
215 COLL_APPLY(action, COLL_ALLGATHER_SIG, rhv) COLL_sep \
216 COLL_APPLY(action, COLL_ALLGATHER_SIG, ring) COLL_sep \
217 COLL_APPLY(action, COLL_ALLGATHER_SIG, SMP_NTS) COLL_sep \
218 COLL_APPLY(action, COLL_ALLGATHER_SIG, smp_simple) COLL_sep \
219 COLL_APPLY(action, COLL_ALLGATHER_SIG, spreading_simple) COLL_sep \
220 COLL_APPLY(action, COLL_ALLGATHER_SIG, ompi) COLL_sep \
221 COLL_APPLY(action, COLL_ALLGATHER_SIG, ompi_neighborexchange) COLL_sep \
222 COLL_APPLY(action, COLL_ALLGATHER_SIG, mvapich2) COLL_sep \
223 COLL_APPLY(action, COLL_ALLGATHER_SIG, mvapich2_smp) COLL_sep \
224 COLL_APPLY(action, COLL_ALLGATHER_SIG, mpich) COLL_sep \
225 COLL_APPLY(action, COLL_ALLGATHER_SIG, impi) COLL_sep \
226 COLL_APPLY(action, COLL_ALLGATHER_SIG, automatic)
232 #define COLL_ALLGATHERVS(action, COLL_sep) \
233 COLL_APPLY(action, COLL_ALLGATHERV_SIG, default) COLL_sep \
234 COLL_APPLY(action, COLL_ALLGATHERV_SIG, GB) COLL_sep \
235 COLL_APPLY(action, COLL_ALLGATHERV_SIG, pair) COLL_sep \
236 COLL_APPLY(action, COLL_ALLGATHERV_SIG, ring) COLL_sep \
237 COLL_APPLY(action, COLL_ALLGATHERV_SIG, ompi) COLL_sep \
238 COLL_APPLY(action, COLL_ALLGATHERV_SIG, ompi_neighborexchange) COLL_sep \
239 COLL_APPLY(action, COLL_ALLGATHERV_SIG, ompi_bruck) COLL_sep \
240 COLL_APPLY(action, COLL_ALLGATHERV_SIG, mpich) COLL_sep \
241 COLL_APPLY(action, COLL_ALLGATHERV_SIG, mpich_rdb) COLL_sep \
242 COLL_APPLY(action, COLL_ALLGATHERV_SIG, mpich_ring) COLL_sep \
243 COLL_APPLY(action, COLL_ALLGATHERV_SIG, mvapich2) COLL_sep \
244 COLL_APPLY(action, COLL_ALLGATHERV_SIG, impi) COLL_sep \
245 COLL_APPLY(action, COLL_ALLGATHERV_SIG, automatic)
247 int allgatherv__default(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, const int *recv_count, const int *recv_disps, MPI_Datatype recv_type, MPI_Comm comm);
248 int allgatherv__GB(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, const int *recv_count, const int *recv_disps, MPI_Datatype recv_type, MPI_Comm comm);
249 int allgatherv__pair(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, const int *recv_count, const int *recv_disps, MPI_Datatype recv_type, MPI_Comm comm);
250 int allgatherv__ring(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, const int *recv_count, const int *recv_disps, MPI_Datatype recv_type, MPI_Comm comm);
251 int allgatherv__ompi(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, const int *recv_count, const int *recv_disps, MPI_Datatype recv_type, MPI_Comm comm);
252 int allgatherv__ompi_neighborexchange(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, const int *recv_count, const int *recv_disps, MPI_Datatype recv_type, MPI_Comm comm);
253 int allgatherv__ompi_bruck (const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, const int *recv_count, const int *recv_disps, MPI_Datatype recv_type, MPI_Comm comm);
254 int allgatherv__mpich(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, const int *recv_count, const int *recv_disps, MPI_Datatype recv_type, MPI_Comm comm);
255 int allgatherv__mpich_rdb(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, const int *recv_count, const int *recv_disps, MPI_Datatype recv_type, MPI_Comm comm);
256 int allgatherv__mpich_ring(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, const int *recv_count, const int *recv_disps, MPI_Datatype recv_type, MPI_Comm comm);
257 int allgatherv__mvapich2(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, const int *recv_count, const int *recv_disps, MPI_Datatype recv_type, MPI_Comm comm);
258 int allgatherv__impi(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, const int *recv_count, const int *recv_disps, MPI_Datatype recv_type, MPI_Comm comm);
259 int allgatherv__automatic(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, const int *recv_count, const int *recv_disps, MPI_Datatype recv_type, MPI_Comm comm);
265 #define COLL_ALLREDUCES(action, COLL_sep) \
266 COLL_APPLY(action, COLL_ALLREDUCE_SIG, default) COLL_sep \
267 COLL_APPLY(action, COLL_ALLREDUCE_SIG, lr) COLL_sep \
268 COLL_APPLY(action, COLL_ALLREDUCE_SIG, rab1) COLL_sep \
269 COLL_APPLY(action, COLL_ALLREDUCE_SIG, rab2) COLL_sep \
270 COLL_APPLY(action, COLL_ALLREDUCE_SIG, rab_rdb) COLL_sep \
271 COLL_APPLY(action, COLL_ALLREDUCE_SIG, rdb) COLL_sep \
272 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_binomial) COLL_sep \
273 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_binomial_pipeline) COLL_sep \
274 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_rdb) COLL_sep \
275 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_rsag) COLL_sep \
276 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_rsag_lr) COLL_sep \
277 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_rsag_rab) COLL_sep \
278 COLL_APPLY(action, COLL_ALLREDUCE_SIG, redbcast) COLL_sep \
279 COLL_APPLY(action, COLL_ALLREDUCE_SIG, ompi) COLL_sep \
280 COLL_APPLY(action, COLL_ALLREDUCE_SIG, ompi_ring_segmented) COLL_sep \
281 COLL_APPLY(action, COLL_ALLREDUCE_SIG, mpich) COLL_sep \
282 COLL_APPLY(action, COLL_ALLREDUCE_SIG, mvapich2) COLL_sep \
283 COLL_APPLY(action, COLL_ALLREDUCE_SIG, mvapich2_rs) COLL_sep \
284 COLL_APPLY(action, COLL_ALLREDUCE_SIG, mvapich2_two_level) COLL_sep \
285 COLL_APPLY(action, COLL_ALLREDUCE_SIG, impi) COLL_sep \
286 COLL_APPLY(action, COLL_ALLREDUCE_SIG, rab) COLL_sep \
287 COLL_APPLY(action, COLL_ALLREDUCE_SIG, automatic)
289 int allreduce__default(const void *sbuf, void *rbuf, int rcount, MPI_Datatype dtype, MPI_Op op, MPI_Comm comm);
290 int allreduce__lr(const void *sbuf, void *rbuf, int rcount, MPI_Datatype dtype, MPI_Op op, MPI_Comm comm);
291 int allreduce__rab1(const void *sbuf, void *rbuf, int rcount, MPI_Datatype dtype, MPI_Op op, MPI_Comm comm);
292 int allreduce__rab2(const void *sbuf, void *rbuf, int rcount, MPI_Datatype dtype, MPI_Op op, MPI_Comm comm);
293 int allreduce__rab_rdb(const void *sbuf, void *rbuf, int rcount, MPI_Datatype dtype, MPI_Op op, MPI_Comm comm);
294 int allreduce__rdb(const void *sbuf, void *rbuf, int rcount, MPI_Datatype dtype, MPI_Op op, MPI_Comm comm);
295 int allreduce__smp_binomial(const void *sbuf, void *rbuf, int rcount, MPI_Datatype dtype, MPI_Op op, MPI_Comm comm);
296 int allreduce__smp_binomial_pipeline(const void *sbuf, void *rbuf, int rcount, MPI_Datatype dtype, MPI_Op op, MPI_Comm comm);
297 int allreduce__smp_rdb(const void *sbuf, void *rbuf, int rcount, MPI_Datatype dtype, MPI_Op op, MPI_Comm comm);
298 int allreduce__smp_rsag(const void *sbuf, void *rbuf, int rcount, MPI_Datatype dtype, MPI_Op op, MPI_Comm comm);
299 int allreduce__smp_rsag_lr(const void *sbuf, void *rbuf, int rcount, MPI_Datatype dtype, MPI_Op op, MPI_Comm comm);
300 int allreduce__smp_rsag_rab(const void *sbuf, void *rbuf, int rcount, MPI_Datatype dtype, MPI_Op op, MPI_Comm comm);
301 int allreduce__redbcast(const void *sbuf, void *rbuf, int rcount, MPI_Datatype dtype, MPI_Op op, MPI_Comm comm);
302 int allreduce__ompi(const void *sbuf, void *rbuf, int rcount, MPI_Datatype dtype, MPI_Op op, MPI_Comm comm);
303 int allreduce__ompi_ring_segmented(const void *sbuf, void *rbuf, int rcount, MPI_Datatype dtype, MPI_Op op, MPI_Comm comm);
304 int allreduce__mpich(const void *sbuf, void *rbuf, int rcount, MPI_Datatype dtype, MPI_Op op, MPI_Comm comm);
305 int allreduce__mvapich2(const void *sbuf, void *rbuf, int rcount, MPI_Datatype dtype, MPI_Op op, MPI_Comm comm);
306 int allreduce__mvapich2_rs(const void *sbuf, void *rbuf, int rcount, MPI_Datatype dtype, MPI_Op op, MPI_Comm comm);
307 int allreduce__mvapich2_two_level(const void *sbuf, void *rbuf, int rcount, MPI_Datatype dtype, MPI_Op op, MPI_Comm comm);
308 int allreduce__impi(const void *sbuf, void *rbuf, int rcount, MPI_Datatype dtype, MPI_Op op, MPI_Comm comm);
309 int allreduce__rab(const void *sbuf, void *rbuf, int rcount, MPI_Datatype dtype, MPI_Op op, MPI_Comm comm);
310 int allreduce__automatic(const void *sbuf, void *rbuf, int rcount, MPI_Datatype dtype, MPI_Op op, MPI_Comm comm);
316 #define COLL_ALLTOALLS(action, COLL_sep) \
317 COLL_APPLY(action, COLL_ALLTOALL_SIG, default) COLL_sep \
318 COLL_APPLY(action, COLL_ALLTOALL_SIG, 2dmesh) COLL_sep \
319 COLL_APPLY(action, COLL_ALLTOALL_SIG, 3dmesh) COLL_sep \
320 COLL_APPLY(action, COLL_ALLTOALL_SIG, basic_linear) COLL_sep \
321 COLL_APPLY(action, COLL_ALLTOALL_SIG, bruck) COLL_sep \
322 COLL_APPLY(action, COLL_ALLTOALL_SIG, pair) COLL_sep \
323 COLL_APPLY(action, COLL_ALLTOALL_SIG, pair_rma) COLL_sep \
324 COLL_APPLY(action, COLL_ALLTOALL_SIG, pair_light_barrier) COLL_sep \
325 COLL_APPLY(action, COLL_ALLTOALL_SIG, pair_mpi_barrier) COLL_sep \
326 COLL_APPLY(action, COLL_ALLTOALL_SIG, pair_one_barrier) COLL_sep \
327 COLL_APPLY(action, COLL_ALLTOALL_SIG, rdb) COLL_sep \
328 COLL_APPLY(action, COLL_ALLTOALL_SIG, ring) COLL_sep \
329 COLL_APPLY(action, COLL_ALLTOALL_SIG, ring_light_barrier) COLL_sep \
330 COLL_APPLY(action, COLL_ALLTOALL_SIG, ring_mpi_barrier) COLL_sep \
331 COLL_APPLY(action, COLL_ALLTOALL_SIG, ring_one_barrier) COLL_sep \
332 COLL_APPLY(action, COLL_ALLTOALL_SIG, mvapich2) COLL_sep \
333 COLL_APPLY(action, COLL_ALLTOALL_SIG, mvapich2_scatter_dest) COLL_sep \
334 COLL_APPLY(action, COLL_ALLTOALL_SIG, ompi) COLL_sep \
335 COLL_APPLY(action, COLL_ALLTOALL_SIG, mpich) COLL_sep \
336 COLL_APPLY(action, COLL_ALLTOALL_SIG, impi) COLL_sep \
337 COLL_APPLY(action, COLL_ALLTOALL_SIG, automatic)
339 int alltoall__default(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
340 int alltoall__2dmesh(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
341 int alltoall__3dmesh(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
342 int alltoall__basic_linear(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
343 int alltoall__bruck(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
344 int alltoall__pair(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
345 int alltoall__pair_rma(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
346 int alltoall__pair_light_barrier(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
347 int alltoall__pair_mpi_barrier(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
348 int alltoall__pair_one_barrier(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
349 int alltoall__rdb(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
350 int alltoall__ring(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
351 int alltoall__ring_light_barrier(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
352 int alltoall__ring_mpi_barrier(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
353 int alltoall__ring_one_barrier(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
354 int alltoall__mvapich2(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
355 int alltoall__mvapich2_scatter_dest(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
356 int alltoall__ompi(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
357 int alltoall__mpich(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
358 int alltoall__impi(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
359 int alltoall__automatic(const void *send_buff, int send_count, MPI_Datatype send_type, void *recv_buff, int recv_count, MPI_Datatype recv_type, MPI_Comm comm);
365 #define COLL_ALLTOALLVS(action, COLL_sep) \
366 COLL_APPLY(action, COLL_ALLTOALLV_SIG, default) COLL_sep \
367 COLL_APPLY(action, COLL_ALLTOALLV_SIG, bruck) COLL_sep \
368 COLL_APPLY(action, COLL_ALLTOALLV_SIG, pair) COLL_sep \
369 COLL_APPLY(action, COLL_ALLTOALLV_SIG, pair_light_barrier) COLL_sep \
370 COLL_APPLY(action, COLL_ALLTOALLV_SIG, pair_mpi_barrier) COLL_sep \
371 COLL_APPLY(action, COLL_ALLTOALLV_SIG, pair_one_barrier) COLL_sep \
372 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ring) COLL_sep \
373 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ring_light_barrier) COLL_sep \
374 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ring_mpi_barrier) COLL_sep \
375 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ring_one_barrier) COLL_sep \
376 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ompi) COLL_sep \
377 COLL_APPLY(action, COLL_ALLTOALLV_SIG, mpich) COLL_sep \
378 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ompi_basic_linear) COLL_sep \
379 COLL_APPLY(action, COLL_ALLTOALLV_SIG, mvapich2) COLL_sep \
380 COLL_APPLY(action, COLL_ALLTOALLV_SIG, impi) COLL_sep \
381 COLL_APPLY(action, COLL_ALLTOALLV_SIG, automatic)
383 int alltoallv__default(const void *send_buff, const int *send_counts, const int *send_disps, MPI_Datatype send_type, void *recv_buff, const int *recv_counts, const int *recv_disps, MPI_Datatype recv_type, MPI_Comm comm);
384 int alltoallv__bruck(const void *send_buff, const int *send_counts, const int *send_disps, MPI_Datatype send_type, void *recv_buff, const int *recv_counts, const int *recv_disps, MPI_Datatype recv_type, MPI_Comm comm);
385 int alltoallv__pair(const void *send_buff, const int *send_counts, const int *send_disps, MPI_Datatype send_type, void *recv_buff, const int *recv_counts, const int *recv_disps, MPI_Datatype recv_type, MPI_Comm comm);
386 int alltoallv__pair_light_barrier(const void *send_buff, const int *send_counts, const int *send_disps, MPI_Datatype send_type, void *recv_buff, const int *recv_counts, const int *recv_disps, MPI_Datatype recv_type, MPI_Comm comm);
387 int alltoallv__pair_mpi_barrier(const void *send_buff, const int *send_counts, const int *send_disps, MPI_Datatype send_type, void *recv_buff, const int *recv_counts, const int *recv_disps, MPI_Datatype recv_type, MPI_Comm comm);
388 int alltoallv__pair_one_barrier(const void *send_buff, const int *send_counts, const int *send_disps, MPI_Datatype send_type, void *recv_buff, const int *recv_counts, const int *recv_disps, MPI_Datatype recv_type, MPI_Comm comm);
389 int alltoallv__ring(const void *send_buff, const int *send_counts, const int *send_disps, MPI_Datatype send_type, void *recv_buff, const int *recv_counts, const int *recv_disps, MPI_Datatype recv_type, MPI_Comm comm);
390 int alltoallv__ring_light_barrier(const void *send_buff, const int *send_counts, const int *send_disps, MPI_Datatype send_type, void *recv_buff, const int *recv_counts, const int *recv_disps, MPI_Datatype recv_type, MPI_Comm comm);
391 int alltoallv__ring_mpi_barrier(const void *send_buff, const int *send_counts, const int *send_disps, MPI_Datatype send_type, void *recv_buff, const int *recv_counts, const int *recv_disps, MPI_Datatype recv_type, MPI_Comm comm);
392 int alltoallv__ring_one_barrier(const void *send_buff, const int *send_counts, const int *send_disps, MPI_Datatype send_type, void *recv_buff, const int *recv_counts, const int *recv_disps, MPI_Datatype recv_type, MPI_Comm comm);
393 int alltoallv__ompi(const void *send_buff, const int *send_counts, const int *send_disps, MPI_Datatype send_type, void *recv_buff, const int *recv_counts, const int *recv_disps, MPI_Datatype recv_type, MPI_Comm comm);
394 int alltoallv__mpich(const void *send_buff, const int *send_counts, const int *send_disps, MPI_Datatype send_type, void *recv_buff, const int *recv_counts, const int *recv_disps, MPI_Datatype recv_type, MPI_Comm comm);
395 int alltoallv__ompi_basic_linear(const void *send_buff, const int *send_counts, const int *send_disps, MPI_Datatype send_type, void *recv_buff, const int *recv_counts, const int *recv_disps, MPI_Datatype recv_type, MPI_Comm comm);
396 int alltoallv__mvapich2(const void *send_buff, const int *send_counts, const int *send_disps, MPI_Datatype send_type, void *recv_buff, const int *recv_counts, const int *recv_disps, MPI_Datatype recv_type, MPI_Comm comm);
397 int alltoallv__impi(const void *send_buff, const int *send_counts, const int *send_disps, MPI_Datatype send_type, void *recv_buff, const int *recv_counts, const int *recv_disps, MPI_Datatype recv_type, MPI_Comm comm);
398 int alltoallv__automatic(const void *send_buff, const int *send_counts, const int *send_disps, MPI_Datatype send_type, void *recv_buff, const int *recv_counts, const int *recv_disps, MPI_Datatype recv_type, MPI_Comm comm);
404 #define COLL_BCASTS(action, COLL_sep) \
405 COLL_APPLY(action, COLL_BCAST_SIG, default) COLL_sep \
406 COLL_APPLY(action, COLL_BCAST_SIG, arrival_pattern_aware) COLL_sep \
407 COLL_APPLY(action, COLL_BCAST_SIG, arrival_pattern_aware_wait) COLL_sep \
408 COLL_APPLY(action, COLL_BCAST_SIG, arrival_scatter) COLL_sep \
409 COLL_APPLY(action, COLL_BCAST_SIG, binomial_tree) COLL_sep \
410 COLL_APPLY(action, COLL_BCAST_SIG, flattree) COLL_sep \
411 COLL_APPLY(action, COLL_BCAST_SIG, flattree_pipeline) COLL_sep \
412 COLL_APPLY(action, COLL_BCAST_SIG, NTSB) COLL_sep \
413 COLL_APPLY(action, COLL_BCAST_SIG, NTSL) COLL_sep \
414 COLL_APPLY(action, COLL_BCAST_SIG, NTSL_Isend) COLL_sep \
415 COLL_APPLY(action, COLL_BCAST_SIG, scatter_LR_allgather) COLL_sep \
416 COLL_APPLY(action, COLL_BCAST_SIG, scatter_rdb_allgather) COLL_sep \
417 COLL_APPLY(action, COLL_BCAST_SIG, SMP_binary) COLL_sep \
418 COLL_APPLY(action, COLL_BCAST_SIG, SMP_binomial) COLL_sep \
419 COLL_APPLY(action, COLL_BCAST_SIG, SMP_linear) COLL_sep \
420 COLL_APPLY(action, COLL_BCAST_SIG, ompi) COLL_sep \
421 COLL_APPLY(action, COLL_BCAST_SIG, ompi_split_bintree) COLL_sep \
422 COLL_APPLY(action, COLL_BCAST_SIG, ompi_pipeline) COLL_sep \
423 COLL_APPLY(action, COLL_BCAST_SIG, mpich) COLL_sep \
424 COLL_APPLY(action, COLL_BCAST_SIG, mvapich2) COLL_sep \
425 COLL_APPLY(action, COLL_BCAST_SIG, mvapich2_inter_node) COLL_sep \
426 COLL_APPLY(action, COLL_BCAST_SIG, mvapich2_intra_node) COLL_sep \
427 COLL_APPLY(action, COLL_BCAST_SIG, mvapich2_knomial_intra_node) COLL_sep \
428 COLL_APPLY(action, COLL_BCAST_SIG, impi) COLL_sep \
429 COLL_APPLY(action, COLL_BCAST_SIG, automatic)
431 int bcast__default(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm);
432 int bcast__arrival_pattern_aware(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm);
433 int bcast__arrival_pattern_aware_wait(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm);
434 int bcast__arrival_scatter(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm);
435 int bcast__binomial_tree(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm);
436 int bcast__flattree(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm);
437 int bcast__flattree_pipeline(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm);
438 int bcast__NTSB(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm);
439 int bcast__NTSL(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm);
440 int bcast__NTSL_Isend(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm);
441 int bcast__scatter_LR_allgather(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm);
442 int bcast__scatter_rdb_allgather(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm);
443 int bcast__SMP_binary(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm);
444 int bcast__SMP_binomial(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm);
445 int bcast__SMP_linear(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm);
446 int bcast__ompi(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm);
447 int bcast__ompi_split_bintree(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm);
448 int bcast__ompi_pipeline(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm);
449 int bcast__mpich(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm);
450 int bcast__mvapich2(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm);
451 int bcast__mvapich2_inter_node(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm);
452 int bcast__mvapich2_intra_node(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm);
453 int bcast__mvapich2_knomial_intra_node(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm);
454 int bcast__impi(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm);
455 int bcast__automatic(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm);
461 #define COLL_REDUCES(action, COLL_sep) \
462 COLL_APPLY(action, COLL_REDUCE_SIG, default) COLL_sep \
463 COLL_APPLY(action, COLL_REDUCE_SIG, arrival_pattern_aware) COLL_sep \
464 COLL_APPLY(action, COLL_REDUCE_SIG, binomial) COLL_sep \
465 COLL_APPLY(action, COLL_REDUCE_SIG, flat_tree) COLL_sep \
466 COLL_APPLY(action, COLL_REDUCE_SIG, NTSL) COLL_sep \
467 COLL_APPLY(action, COLL_REDUCE_SIG, scatter_gather) COLL_sep \
468 COLL_APPLY(action, COLL_REDUCE_SIG, ompi) COLL_sep \
469 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_chain) COLL_sep \
470 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_pipeline) COLL_sep \
471 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_basic_linear) COLL_sep \
472 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_in_order_binary) COLL_sep \
473 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_binary) COLL_sep \
474 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_binomial) COLL_sep \
475 COLL_APPLY(action, COLL_REDUCE_SIG, mpich) COLL_sep \
476 COLL_APPLY(action, COLL_REDUCE_SIG, mvapich2) COLL_sep \
477 COLL_APPLY(action, COLL_REDUCE_SIG, mvapich2_knomial) COLL_sep \
478 COLL_APPLY(action, COLL_REDUCE_SIG, mvapich2_two_level) COLL_sep \
479 COLL_APPLY(action, COLL_REDUCE_SIG, impi) COLL_sep \
480 COLL_APPLY(action, COLL_REDUCE_SIG, rab) COLL_sep \
481 COLL_APPLY(action, COLL_REDUCE_SIG, automatic)
483 int reduce__default(const void *buf, void *rbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm);
484 int reduce__arrival_pattern_aware(const void *buf, void *rbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm);
485 int reduce__binomial(const void *buf, void *rbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm);
486 int reduce__flat_tree(const void *buf, void *rbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm);
487 int reduce__NTSL(const void *buf, void *rbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm);
488 int reduce__scatter_gather(const void *buf, void *rbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm);
489 int reduce__ompi(const void *buf, void *rbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm);
490 int reduce__ompi_chain(const void *buf, void *rbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm);
491 int reduce__ompi_pipeline(const void *buf, void *rbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm);
492 int reduce__ompi_basic_linear(const void *buf, void *rbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm);
493 int reduce__ompi_in_order_binary(const void *buf, void *rbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm);
494 int reduce__ompi_binary(const void *buf, void *rbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm);
495 int reduce__ompi_binomial(const void *buf, void *rbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm);
496 int reduce__mpich(const void *buf, void *rbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm);
497 int reduce__mvapich2(const void *buf, void *rbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm);
498 int reduce__mvapich2_knomial(const void *buf, void *rbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm);
499 int reduce__mvapich2_two_level(const void *buf, void *rbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm);
500 int reduce__impi(const void *buf, void *rbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm);
501 int reduce__rab(const void *buf, void *rbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm);
502 int reduce__automatic(const void *buf, void *rbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm);
508 #define COLL_REDUCE_SCATTERS(action, COLL_sep) \
509 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, default) COLL_sep \
510 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, ompi) COLL_sep \
511 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, ompi_basic_recursivehalving) COLL_sep \
512 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, ompi_ring) COLL_sep \
513 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, mpich) COLL_sep \
514 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, mpich_pair) COLL_sep \
515 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, mpich_rdb) COLL_sep \
516 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, mpich_noncomm) COLL_sep \
517 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, mvapich2) COLL_sep \
518 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, impi) COLL_sep \
519 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, automatic)
521 int reduce_scatter__default(const void *sbuf, void *rbuf, const int *rcounts, MPI_Datatype dtype,MPI_Op op,MPI_Comm comm);
522 int reduce_scatter__ompi(const void *sbuf, void *rbuf, const int *rcounts, MPI_Datatype dtype,MPI_Op op,MPI_Comm comm);
523 int reduce_scatter__ompi_basic_recursivehalving(const void *sbuf, void *rbuf, const int *rcounts, MPI_Datatype dtype,MPI_Op op,MPI_Comm comm);
524 int reduce_scatter__ompi_ring(const void *sbuf, void *rbuf, const int *rcounts, MPI_Datatype dtype,MPI_Op op,MPI_Comm comm);
525 int reduce_scatter__mpich(const void *sbuf, void *rbuf, const int *rcounts, MPI_Datatype dtype,MPI_Op op,MPI_Comm comm);
526 int reduce_scatter__mpich_pair(const void *sbuf, void *rbuf, const int *rcounts, MPI_Datatype dtype,MPI_Op op,MPI_Comm comm);
527 int reduce_scatter__mpich_rdb(const void *sbuf, void *rbuf, const int *rcounts, MPI_Datatype dtype,MPI_Op op,MPI_Comm comm);
528 int reduce_scatter__mpich_noncomm(const void *sbuf, void *rbuf, const int *rcounts, MPI_Datatype dtype,MPI_Op op,MPI_Comm comm);
529 int reduce_scatter__mvapich2(const void *sbuf, void *rbuf, const int *rcounts, MPI_Datatype dtype,MPI_Op op,MPI_Comm comm);
530 int reduce_scatter__impi(const void *sbuf, void *rbuf, const int *rcounts, MPI_Datatype dtype,MPI_Op op,MPI_Comm comm);
531 int reduce_scatter__automatic(const void *sbuf, void *rbuf, const int *rcounts, MPI_Datatype dtype,MPI_Op op,MPI_Comm comm);
537 #define COLL_SCATTERS(action, COLL_sep) \
538 COLL_APPLY(action, COLL_SCATTER_SIG, default) COLL_sep \
539 COLL_APPLY(action, COLL_SCATTER_SIG, ompi) COLL_sep \
540 COLL_APPLY(action, COLL_SCATTER_SIG, ompi_basic_linear) COLL_sep \
541 COLL_APPLY(action, COLL_SCATTER_SIG, ompi_binomial) COLL_sep \
542 COLL_APPLY(action, COLL_SCATTER_SIG, mpich) COLL_sep \
543 COLL_APPLY(action, COLL_SCATTER_SIG, mvapich2) COLL_sep \
544 COLL_APPLY(action, COLL_SCATTER_SIG, mvapich2_two_level_binomial) COLL_sep \
545 COLL_APPLY(action, COLL_SCATTER_SIG, mvapich2_two_level_direct) COLL_sep \
546 COLL_APPLY(action, COLL_SCATTER_SIG, impi) COLL_sep \
547 COLL_APPLY(action, COLL_SCATTER_SIG, automatic)
549 int scatter__default(const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm);
550 int scatter__ompi(const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm);
551 int scatter__ompi_basic_linear(const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm);
552 int scatter__ompi_binomial(const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm);
553 int scatter__mpich(const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm);
554 int scatter__mvapich2(const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm);
555 int scatter__mvapich2_two_level_binomial(const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm);
556 int scatter__mvapich2_two_level_direct(const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm);
557 int scatter__impi (const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm);
558 int scatter__automatic (const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm);
564 #define COLL_BARRIERS(action, COLL_sep) \
565 COLL_APPLY(action, COLL_BARRIER_SIG, default) COLL_sep \
566 COLL_APPLY(action, COLL_BARRIER_SIG, ompi) COLL_sep \
567 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_basic_linear) COLL_sep \
568 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_two_procs) COLL_sep \
569 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_tree) COLL_sep \
570 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_bruck) COLL_sep \
571 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_recursivedoubling) COLL_sep \
572 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_doublering) COLL_sep \
573 COLL_APPLY(action, COLL_BARRIER_SIG, mpich_smp) COLL_sep \
574 COLL_APPLY(action, COLL_BARRIER_SIG, mpich) COLL_sep \
575 COLL_APPLY(action, COLL_BARRIER_SIG, mvapich2_pair) COLL_sep \
576 COLL_APPLY(action, COLL_BARRIER_SIG, mvapich2) COLL_sep \
577 COLL_APPLY(action, COLL_BARRIER_SIG, impi) COLL_sep \
578 COLL_APPLY(action, COLL_BARRIER_SIG, automatic)
580 int barrier__default(MPI_Comm comm);
581 int barrier__ompi(MPI_Comm comm);
582 int barrier__ompi_basic_linear(MPI_Comm comm);
583 int barrier__ompi_two_procs(MPI_Comm comm);
584 int barrier__ompi_tree(MPI_Comm comm);
585 int barrier__ompi_bruck(MPI_Comm comm);
586 int barrier__ompi_recursivedoubling(MPI_Comm comm);
587 int barrier__ompi_doublering(MPI_Comm comm);
588 int barrier__mpich_smp(MPI_Comm comm);
589 int barrier__mpich(MPI_Comm comm);
590 int barrier__mvapich2_pair(MPI_Comm comm);
591 int barrier__mvapich2 (MPI_Comm comm);
592 int barrier__impi(MPI_Comm comm);
593 int barrier__automatic(MPI_Comm comm);