1 /* Copyright (c) 2013-2014. The SimGrid Team.
2 * All rights reserved. */
4 /* This program is free software; you can redistribute it and/or modify it
5 * under the terms of the license (GNU LGPL) which comes with this package. */
12 #include "smpi/private.h"
16 #define COLL_DESCRIPTION(cat, ret, args, name) \
18 # cat " " # name " collective",\
19 smpi_coll_tuned_ ## cat ## _ ## name}
21 #define COLL_PROTO(cat, ret, args, name) \
22 ret smpi_coll_tuned_ ## cat ## _ ## name(COLL_UNPAREN args);
23 #define COLL_UNPAREN(...) __VA_ARGS__
25 #define COLL_APPLY(action, sig, name) action(sig, name)
28 #define COLL_NOTHING(...)
33 #define COLL_GATHER_SIG gather, int, \
34 (void *send_buff, int send_count, MPI_Datatype send_type, \
35 void *recv_buff, int recv_count, MPI_Datatype recv_type, \
36 int root, MPI_Comm comm)
38 #define COLL_GATHERS(action, COLL_sep) \
39 COLL_APPLY(action, COLL_GATHER_SIG, ompi) COLL_sep \
40 COLL_APPLY(action, COLL_GATHER_SIG, ompi_basic_linear) COLL_sep \
41 COLL_APPLY(action, COLL_GATHER_SIG, ompi_binomial) COLL_sep \
42 COLL_APPLY(action, COLL_GATHER_SIG, ompi_linear_sync) COLL_sep \
43 COLL_APPLY(action, COLL_GATHER_SIG, mpich) COLL_sep \
44 COLL_APPLY(action, COLL_GATHER_SIG, mvapich2) COLL_sep \
45 COLL_APPLY(action, COLL_GATHER_SIG, automatic)
49 COLL_GATHERS(COLL_PROTO, COLL_NOsep)
54 #define COLL_ALLGATHER_SIG allgather, int, \
55 (void *send_buff, int send_count, MPI_Datatype send_type, \
56 void *recv_buff, int recv_count, MPI_Datatype recv_type, \
59 #define COLL_ALLGATHERS(action, COLL_sep) \
60 COLL_APPLY(action, COLL_ALLGATHER_SIG, 2dmesh) COLL_sep \
61 COLL_APPLY(action, COLL_ALLGATHER_SIG, 3dmesh) COLL_sep \
62 COLL_APPLY(action, COLL_ALLGATHER_SIG, bruck) COLL_sep \
63 COLL_APPLY(action, COLL_ALLGATHER_SIG, GB) COLL_sep \
64 COLL_APPLY(action, COLL_ALLGATHER_SIG, loosely_lr) COLL_sep \
65 COLL_APPLY(action, COLL_ALLGATHER_SIG, NTSLR) COLL_sep \
66 COLL_APPLY(action, COLL_ALLGATHER_SIG, NTSLR_NB) COLL_sep \
67 COLL_APPLY(action, COLL_ALLGATHER_SIG, pair) COLL_sep \
68 COLL_APPLY(action, COLL_ALLGATHER_SIG, rdb) COLL_sep \
69 COLL_APPLY(action, COLL_ALLGATHER_SIG, rhv) COLL_sep \
70 COLL_APPLY(action, COLL_ALLGATHER_SIG, ring) COLL_sep \
71 COLL_APPLY(action, COLL_ALLGATHER_SIG, SMP_NTS) COLL_sep \
72 COLL_APPLY(action, COLL_ALLGATHER_SIG, smp_simple) COLL_sep \
73 COLL_APPLY(action, COLL_ALLGATHER_SIG, spreading_simple) COLL_sep \
74 COLL_APPLY(action, COLL_ALLGATHER_SIG, ompi) COLL_sep \
75 COLL_APPLY(action, COLL_ALLGATHER_SIG, ompi_neighborexchange) COLL_sep \
76 COLL_APPLY(action, COLL_ALLGATHER_SIG, mvapich2) COLL_sep \
77 COLL_APPLY(action, COLL_ALLGATHER_SIG, mpich) COLL_sep \
78 COLL_APPLY(action, COLL_ALLGATHER_SIG, automatic)
81 COLL_ALLGATHERS(COLL_PROTO, COLL_NOsep)
86 #define COLL_ALLGATHERV_SIG allgatherv, int, \
87 (void *send_buff, int send_count, MPI_Datatype send_type, \
88 void *recv_buff, int *recv_count, int *recv_disps, \
89 MPI_Datatype recv_type, MPI_Comm comm)
91 #define COLL_ALLGATHERVS(action, COLL_sep) \
92 COLL_APPLY(action, COLL_ALLGATHERV_SIG, GB) COLL_sep \
93 COLL_APPLY(action, COLL_ALLGATHERV_SIG, pair) COLL_sep \
94 COLL_APPLY(action, COLL_ALLGATHERV_SIG, ring) COLL_sep \
95 COLL_APPLY(action, COLL_ALLGATHERV_SIG, ompi) COLL_sep \
96 COLL_APPLY(action, COLL_ALLGATHERV_SIG, ompi_neighborexchange) COLL_sep \
97 COLL_APPLY(action, COLL_ALLGATHERV_SIG, ompi_bruck) COLL_sep \
98 COLL_APPLY(action, COLL_ALLGATHERV_SIG, mpich) COLL_sep \
99 COLL_APPLY(action, COLL_ALLGATHERV_SIG, mpich_rdb) COLL_sep \
100 COLL_APPLY(action, COLL_ALLGATHERV_SIG, mpich_ring) COLL_sep \
101 COLL_APPLY(action, COLL_ALLGATHERV_SIG, mvapich2) COLL_sep \
102 COLL_APPLY(action, COLL_ALLGATHERV_SIG, automatic)
104 COLL_ALLGATHERVS(COLL_PROTO, COLL_NOsep)
109 #define COLL_ALLREDUCE_SIG allreduce, int, \
110 (void *sbuf, void *rbuf, int rcount, \
111 MPI_Datatype dtype, MPI_Op op, MPI_Comm comm)
113 #define COLL_ALLREDUCES(action, COLL_sep) \
114 COLL_APPLY(action, COLL_ALLREDUCE_SIG, lr) COLL_sep \
115 COLL_APPLY(action, COLL_ALLREDUCE_SIG, rab1) COLL_sep \
116 COLL_APPLY(action, COLL_ALLREDUCE_SIG, rab2) COLL_sep \
117 COLL_APPLY(action, COLL_ALLREDUCE_SIG, rab_rdb) COLL_sep \
118 COLL_APPLY(action, COLL_ALLREDUCE_SIG, rdb) COLL_sep \
119 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_binomial) COLL_sep \
120 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_binomial_pipeline) COLL_sep \
121 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_rdb) COLL_sep \
122 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_rsag) COLL_sep \
123 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_rsag_lr) COLL_sep \
124 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_rsag_rab) COLL_sep \
125 COLL_APPLY(action, COLL_ALLREDUCE_SIG, redbcast) COLL_sep \
126 COLL_APPLY(action, COLL_ALLREDUCE_SIG, ompi) COLL_sep \
127 COLL_APPLY(action, COLL_ALLREDUCE_SIG, ompi_ring_segmented) COLL_sep \
128 COLL_APPLY(action, COLL_ALLREDUCE_SIG, mpich) COLL_sep \
129 COLL_APPLY(action, COLL_ALLREDUCE_SIG, mvapich2) COLL_sep \
130 COLL_APPLY(action, COLL_ALLREDUCE_SIG, mvapich2_rs) COLL_sep \
131 COLL_APPLY(action, COLL_ALLREDUCE_SIG, automatic)
133 COLL_ALLREDUCES(COLL_PROTO, COLL_NOsep)
139 #define COLL_ALLTOALL_SIG alltoall, int, \
140 (void *send_buff, int send_count, MPI_Datatype send_type, \
141 void *recv_buff, int recv_count, MPI_Datatype recv_type, \
144 #define COLL_ALLTOALLS(action, COLL_sep) \
145 COLL_APPLY(action, COLL_ALLTOALL_SIG, 2dmesh) COLL_sep \
146 COLL_APPLY(action, COLL_ALLTOALL_SIG, 3dmesh) COLL_sep \
147 COLL_APPLY(action, COLL_ALLTOALL_SIG, bruck) COLL_sep \
148 COLL_APPLY(action, COLL_ALLTOALL_SIG, pair) COLL_sep \
149 COLL_APPLY(action, COLL_ALLTOALL_SIG, pair_light_barrier) COLL_sep \
150 COLL_APPLY(action, COLL_ALLTOALL_SIG, pair_mpi_barrier) COLL_sep \
151 COLL_APPLY(action, COLL_ALLTOALL_SIG, pair_one_barrier) COLL_sep \
152 COLL_APPLY(action, COLL_ALLTOALL_SIG, rdb) COLL_sep \
153 COLL_APPLY(action, COLL_ALLTOALL_SIG, ring) COLL_sep \
154 COLL_APPLY(action, COLL_ALLTOALL_SIG, ring_light_barrier) COLL_sep \
155 COLL_APPLY(action, COLL_ALLTOALL_SIG, ring_mpi_barrier) COLL_sep \
156 COLL_APPLY(action, COLL_ALLTOALL_SIG, ring_one_barrier) COLL_sep \
157 COLL_APPLY(action, COLL_ALLTOALL_SIG, mvapich2) COLL_sep \
158 COLL_APPLY(action, COLL_ALLTOALL_SIG, mvapich2_scatter_dest) COLL_sep \
159 COLL_APPLY(action, COLL_ALLTOALL_SIG, ompi) COLL_sep \
160 COLL_APPLY(action, COLL_ALLTOALL_SIG, mpich) COLL_sep \
161 COLL_APPLY(action, COLL_ALLTOALL_SIG, automatic)
163 COLL_ALLTOALLS(COLL_PROTO, COLL_NOsep)
168 #define COLL_ALLTOALLV_SIG alltoallv, int, \
169 (void *send_buff, int *send_counts, int *send_disps, MPI_Datatype send_type, \
170 void *recv_buff, int *recv_counts, int *recv_disps, MPI_Datatype recv_type, \
173 #define COLL_ALLTOALLVS(action, COLL_sep) \
174 COLL_APPLY(action, COLL_ALLTOALLV_SIG, bruck) COLL_sep \
175 COLL_APPLY(action, COLL_ALLTOALLV_SIG, pair) COLL_sep \
176 COLL_APPLY(action, COLL_ALLTOALLV_SIG, pair_light_barrier) COLL_sep \
177 COLL_APPLY(action, COLL_ALLTOALLV_SIG, pair_mpi_barrier) COLL_sep \
178 COLL_APPLY(action, COLL_ALLTOALLV_SIG, pair_one_barrier) COLL_sep \
179 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ring) COLL_sep \
180 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ring_light_barrier) COLL_sep \
181 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ring_mpi_barrier) COLL_sep \
182 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ring_one_barrier) COLL_sep \
183 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ompi) COLL_sep \
184 COLL_APPLY(action, COLL_ALLTOALLV_SIG, mpich) COLL_sep \
185 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ompi_basic_linear) COLL_sep \
186 COLL_APPLY(action, COLL_ALLTOALLV_SIG, mvapich2) COLL_sep \
187 COLL_APPLY(action, COLL_ALLTOALLV_SIG, automatic)
189 COLL_ALLTOALLVS(COLL_PROTO, COLL_NOsep)
194 #define COLL_BCAST_SIG bcast, int, \
195 (void *buf, int count, MPI_Datatype datatype, \
196 int root, MPI_Comm comm)
198 #define COLL_BCASTS(action, COLL_sep) \
199 COLL_APPLY(action, COLL_BCAST_SIG, arrival_pattern_aware) COLL_sep \
200 COLL_APPLY(action, COLL_BCAST_SIG, arrival_pattern_aware_wait) COLL_sep \
201 COLL_APPLY(action, COLL_BCAST_SIG, arrival_scatter) COLL_sep \
202 COLL_APPLY(action, COLL_BCAST_SIG, binomial_tree) COLL_sep \
203 COLL_APPLY(action, COLL_BCAST_SIG, flattree) COLL_sep \
204 COLL_APPLY(action, COLL_BCAST_SIG, flattree_pipeline) COLL_sep \
205 COLL_APPLY(action, COLL_BCAST_SIG, NTSB) COLL_sep \
206 COLL_APPLY(action, COLL_BCAST_SIG, NTSL) COLL_sep \
207 COLL_APPLY(action, COLL_BCAST_SIG, NTSL_Isend) COLL_sep \
208 COLL_APPLY(action, COLL_BCAST_SIG, scatter_LR_allgather) COLL_sep \
209 COLL_APPLY(action, COLL_BCAST_SIG, scatter_rdb_allgather) COLL_sep \
210 COLL_APPLY(action, COLL_BCAST_SIG, SMP_binary) COLL_sep \
211 COLL_APPLY(action, COLL_BCAST_SIG, SMP_binomial) COLL_sep \
212 COLL_APPLY(action, COLL_BCAST_SIG, SMP_linear) COLL_sep \
213 COLL_APPLY(action, COLL_BCAST_SIG, ompi) COLL_sep \
214 COLL_APPLY(action, COLL_BCAST_SIG, ompi_split_bintree) COLL_sep \
215 COLL_APPLY(action, COLL_BCAST_SIG, ompi_pipeline) COLL_sep \
216 COLL_APPLY(action, COLL_BCAST_SIG, mpich) COLL_sep \
217 COLL_APPLY(action, COLL_BCAST_SIG, mvapich2) COLL_sep \
218 COLL_APPLY(action, COLL_BCAST_SIG, automatic)
220 COLL_BCASTS(COLL_PROTO, COLL_NOsep)
226 #define COLL_REDUCE_SIG reduce, int, \
227 (void *buf, void *rbuf, int count, MPI_Datatype datatype, \
228 MPI_Op op, int root, MPI_Comm comm)
230 #define COLL_REDUCES(action, COLL_sep) \
231 COLL_APPLY(action, COLL_REDUCE_SIG, arrival_pattern_aware) COLL_sep \
232 COLL_APPLY(action, COLL_REDUCE_SIG, binomial) COLL_sep \
233 COLL_APPLY(action, COLL_REDUCE_SIG, flat_tree) COLL_sep \
234 COLL_APPLY(action, COLL_REDUCE_SIG, NTSL) COLL_sep \
235 COLL_APPLY(action, COLL_REDUCE_SIG, scatter_gather) COLL_sep \
236 COLL_APPLY(action, COLL_REDUCE_SIG, ompi) COLL_sep \
237 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_chain) COLL_sep \
238 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_pipeline) COLL_sep \
239 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_basic_linear) COLL_sep \
240 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_in_order_binary) COLL_sep \
241 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_binary) COLL_sep \
242 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_binomial) COLL_sep \
243 COLL_APPLY(action, COLL_REDUCE_SIG, mpich) COLL_sep \
244 COLL_APPLY(action, COLL_REDUCE_SIG, mvapich2) COLL_sep \
245 COLL_APPLY(action, COLL_REDUCE_SIG, mvapich2_knomial) COLL_sep \
246 COLL_APPLY(action, COLL_REDUCE_SIG, automatic)
248 COLL_REDUCES(COLL_PROTO, COLL_NOsep)
253 #define COLL_REDUCE_SCATTER_SIG reduce_scatter, int, \
254 (void *sbuf, void *rbuf, int *rcounts,\
255 MPI_Datatype dtype,MPI_Op op,MPI_Comm comm)
257 #define COLL_REDUCE_SCATTERS(action, COLL_sep) \
258 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, ompi) COLL_sep \
259 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, ompi_basic_recursivehalving) COLL_sep \
260 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, ompi_ring) COLL_sep \
261 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, mpich) COLL_sep \
262 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, mpich_pair) COLL_sep \
263 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, mpich_rdb) COLL_sep \
264 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, mpich_noncomm) COLL_sep \
265 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, mvapich2) COLL_sep \
266 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, automatic)
270 COLL_REDUCE_SCATTERS(COLL_PROTO, COLL_NOsep)
276 #define COLL_SCATTER_SIG scatter, int, \
277 (void *sendbuf, int sendcount, MPI_Datatype sendtype,\
278 void *recvbuf, int recvcount, MPI_Datatype recvtype,\
279 int root, MPI_Comm comm)
281 #define COLL_SCATTERS(action, COLL_sep) \
282 COLL_APPLY(action, COLL_SCATTER_SIG, ompi) COLL_sep \
283 COLL_APPLY(action, COLL_SCATTER_SIG, ompi_basic_linear) COLL_sep \
284 COLL_APPLY(action, COLL_SCATTER_SIG, ompi_binomial) COLL_sep \
285 COLL_APPLY(action, COLL_SCATTER_SIG, mpich) COLL_sep \
286 COLL_APPLY(action, COLL_SCATTER_SIG, mvapich2) COLL_sep \
287 COLL_APPLY(action, COLL_SCATTER_SIG, automatic)
289 COLL_SCATTERS(COLL_PROTO, COLL_NOsep)
294 #define COLL_BARRIER_SIG barrier, int, \
297 #define COLL_BARRIERS(action, COLL_sep) \
298 COLL_APPLY(action, COLL_BARRIER_SIG, ompi) COLL_sep \
299 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_basic_linear) COLL_sep \
300 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_two_procs) COLL_sep \
301 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_tree) COLL_sep \
302 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_bruck) COLL_sep \
303 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_recursivedoubling) COLL_sep \
304 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_doublering) COLL_sep \
305 COLL_APPLY(action, COLL_BARRIER_SIG, mpich) COLL_sep \
306 COLL_APPLY(action, COLL_BARRIER_SIG, mvapich2_pair) COLL_sep \
307 COLL_APPLY(action, COLL_BARRIER_SIG, mvapich2) COLL_sep \
308 COLL_APPLY(action, COLL_BARRIER_SIG, automatic)
310 COLL_BARRIERS(COLL_PROTO, COLL_NOsep)