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 "src/smpi/private.h"
18 #define COLL_DESCRIPTION(cat, ret, args, name) \
20 # cat " " # name " collective",\
21 (void*)smpi_coll_tuned_ ## cat ## _ ## name}
23 #define COLL_PROTO(cat, ret, args, name) \
24 ret smpi_coll_tuned_ ## cat ## _ ## name(COLL_UNPAREN args);
25 #define COLL_UNPAREN(...) __VA_ARGS__
27 #define COLL_APPLY(action, sig, name) action(sig, name)
30 #define COLL_NOTHING(...)
35 #define COLL_GATHER_SIG gather, int, \
36 (void *send_buff, int send_count, MPI_Datatype send_type, \
37 void *recv_buff, int recv_count, MPI_Datatype recv_type, \
38 int root, MPI_Comm comm)
40 #define COLL_GATHERS(action, COLL_sep) \
41 COLL_APPLY(action, COLL_GATHER_SIG, ompi) COLL_sep \
42 COLL_APPLY(action, COLL_GATHER_SIG, ompi_basic_linear) COLL_sep \
43 COLL_APPLY(action, COLL_GATHER_SIG, ompi_binomial) COLL_sep \
44 COLL_APPLY(action, COLL_GATHER_SIG, ompi_linear_sync) COLL_sep \
45 COLL_APPLY(action, COLL_GATHER_SIG, mpich) COLL_sep \
46 COLL_APPLY(action, COLL_GATHER_SIG, mvapich2) COLL_sep \
47 COLL_APPLY(action, COLL_GATHER_SIG, mvapich2_two_level) COLL_sep \
48 COLL_APPLY(action, COLL_GATHER_SIG, impi) COLL_sep \
49 COLL_APPLY(action, COLL_GATHER_SIG, automatic)
53 COLL_GATHERS(COLL_PROTO, COLL_NOsep)
58 #define COLL_ALLGATHER_SIG allgather, int, \
59 (void *send_buff, int send_count, MPI_Datatype send_type, \
60 void *recv_buff, int recv_count, MPI_Datatype recv_type, \
63 #define COLL_ALLGATHERS(action, COLL_sep) \
64 COLL_APPLY(action, COLL_ALLGATHER_SIG, 2dmesh) COLL_sep \
65 COLL_APPLY(action, COLL_ALLGATHER_SIG, 3dmesh) COLL_sep \
66 COLL_APPLY(action, COLL_ALLGATHER_SIG, bruck) COLL_sep \
67 COLL_APPLY(action, COLL_ALLGATHER_SIG, GB) COLL_sep \
68 COLL_APPLY(action, COLL_ALLGATHER_SIG, loosely_lr) COLL_sep \
69 COLL_APPLY(action, COLL_ALLGATHER_SIG, NTSLR) COLL_sep \
70 COLL_APPLY(action, COLL_ALLGATHER_SIG, NTSLR_NB) COLL_sep \
71 COLL_APPLY(action, COLL_ALLGATHER_SIG, pair) COLL_sep \
72 COLL_APPLY(action, COLL_ALLGATHER_SIG, rdb) COLL_sep \
73 COLL_APPLY(action, COLL_ALLGATHER_SIG, rhv) COLL_sep \
74 COLL_APPLY(action, COLL_ALLGATHER_SIG, ring) COLL_sep \
75 COLL_APPLY(action, COLL_ALLGATHER_SIG, SMP_NTS) COLL_sep \
76 COLL_APPLY(action, COLL_ALLGATHER_SIG, smp_simple) COLL_sep \
77 COLL_APPLY(action, COLL_ALLGATHER_SIG, spreading_simple) COLL_sep \
78 COLL_APPLY(action, COLL_ALLGATHER_SIG, ompi) COLL_sep \
79 COLL_APPLY(action, COLL_ALLGATHER_SIG, ompi_neighborexchange) COLL_sep \
80 COLL_APPLY(action, COLL_ALLGATHER_SIG, mvapich2) COLL_sep \
81 COLL_APPLY(action, COLL_ALLGATHER_SIG, mvapich2_smp) COLL_sep \
82 COLL_APPLY(action, COLL_ALLGATHER_SIG, mpich) COLL_sep \
83 COLL_APPLY(action, COLL_ALLGATHER_SIG, impi) COLL_sep \
84 COLL_APPLY(action, COLL_ALLGATHER_SIG, automatic)
87 COLL_ALLGATHERS(COLL_PROTO, COLL_NOsep)
92 #define COLL_ALLGATHERV_SIG allgatherv, int, \
93 (void *send_buff, int send_count, MPI_Datatype send_type, \
94 void *recv_buff, int *recv_count, int *recv_disps, \
95 MPI_Datatype recv_type, MPI_Comm comm)
97 #define COLL_ALLGATHERVS(action, COLL_sep) \
98 COLL_APPLY(action, COLL_ALLGATHERV_SIG, GB) COLL_sep \
99 COLL_APPLY(action, COLL_ALLGATHERV_SIG, pair) COLL_sep \
100 COLL_APPLY(action, COLL_ALLGATHERV_SIG, ring) COLL_sep \
101 COLL_APPLY(action, COLL_ALLGATHERV_SIG, ompi) COLL_sep \
102 COLL_APPLY(action, COLL_ALLGATHERV_SIG, ompi_neighborexchange) COLL_sep \
103 COLL_APPLY(action, COLL_ALLGATHERV_SIG, ompi_bruck) COLL_sep \
104 COLL_APPLY(action, COLL_ALLGATHERV_SIG, mpich) COLL_sep \
105 COLL_APPLY(action, COLL_ALLGATHERV_SIG, mpich_rdb) COLL_sep \
106 COLL_APPLY(action, COLL_ALLGATHERV_SIG, mpich_ring) COLL_sep \
107 COLL_APPLY(action, COLL_ALLGATHERV_SIG, mvapich2) COLL_sep \
108 COLL_APPLY(action, COLL_ALLGATHERV_SIG, impi) COLL_sep \
109 COLL_APPLY(action, COLL_ALLGATHERV_SIG, automatic)
111 COLL_ALLGATHERVS(COLL_PROTO, COLL_NOsep)
116 #define COLL_ALLREDUCE_SIG allreduce, int, \
117 (void *sbuf, void *rbuf, int rcount, \
118 MPI_Datatype dtype, MPI_Op op, MPI_Comm comm)
120 #define COLL_ALLREDUCES(action, COLL_sep) \
121 COLL_APPLY(action, COLL_ALLREDUCE_SIG, lr) COLL_sep \
122 COLL_APPLY(action, COLL_ALLREDUCE_SIG, rab1) COLL_sep \
123 COLL_APPLY(action, COLL_ALLREDUCE_SIG, rab2) COLL_sep \
124 COLL_APPLY(action, COLL_ALLREDUCE_SIG, rab_rdb) COLL_sep \
125 COLL_APPLY(action, COLL_ALLREDUCE_SIG, rdb) COLL_sep \
126 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_binomial) COLL_sep \
127 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_binomial_pipeline) COLL_sep \
128 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_rdb) COLL_sep \
129 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_rsag) COLL_sep \
130 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_rsag_lr) COLL_sep \
131 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_rsag_rab) COLL_sep \
132 COLL_APPLY(action, COLL_ALLREDUCE_SIG, redbcast) COLL_sep \
133 COLL_APPLY(action, COLL_ALLREDUCE_SIG, ompi) COLL_sep \
134 COLL_APPLY(action, COLL_ALLREDUCE_SIG, ompi_ring_segmented) COLL_sep \
135 COLL_APPLY(action, COLL_ALLREDUCE_SIG, mpich) COLL_sep \
136 COLL_APPLY(action, COLL_ALLREDUCE_SIG, mvapich2) COLL_sep \
137 COLL_APPLY(action, COLL_ALLREDUCE_SIG, mvapich2_rs) COLL_sep \
138 COLL_APPLY(action, COLL_ALLREDUCE_SIG, mvapich2_two_level) COLL_sep \
139 COLL_APPLY(action, COLL_ALLREDUCE_SIG, impi) COLL_sep \
140 COLL_APPLY(action, COLL_ALLREDUCE_SIG, rab) COLL_sep \
141 COLL_APPLY(action, COLL_ALLREDUCE_SIG, automatic)
143 COLL_ALLREDUCES(COLL_PROTO, COLL_NOsep)
149 #define COLL_ALLTOALL_SIG alltoall, int, \
150 (void *send_buff, int send_count, MPI_Datatype send_type, \
151 void *recv_buff, int recv_count, MPI_Datatype recv_type, \
154 #define COLL_ALLTOALLS(action, COLL_sep) \
155 COLL_APPLY(action, COLL_ALLTOALL_SIG, 2dmesh) COLL_sep \
156 COLL_APPLY(action, COLL_ALLTOALL_SIG, 3dmesh) COLL_sep \
157 COLL_APPLY(action, COLL_ALLTOALL_SIG, basic_linear) COLL_sep \
158 COLL_APPLY(action, COLL_ALLTOALL_SIG, bruck) COLL_sep \
159 COLL_APPLY(action, COLL_ALLTOALL_SIG, pair) COLL_sep \
160 COLL_APPLY(action, COLL_ALLTOALL_SIG, pair_rma) COLL_sep \
161 COLL_APPLY(action, COLL_ALLTOALL_SIG, pair_light_barrier) COLL_sep \
162 COLL_APPLY(action, COLL_ALLTOALL_SIG, pair_mpi_barrier) COLL_sep \
163 COLL_APPLY(action, COLL_ALLTOALL_SIG, pair_one_barrier) COLL_sep \
164 COLL_APPLY(action, COLL_ALLTOALL_SIG, rdb) COLL_sep \
165 COLL_APPLY(action, COLL_ALLTOALL_SIG, ring) COLL_sep \
166 COLL_APPLY(action, COLL_ALLTOALL_SIG, ring_light_barrier) COLL_sep \
167 COLL_APPLY(action, COLL_ALLTOALL_SIG, ring_mpi_barrier) COLL_sep \
168 COLL_APPLY(action, COLL_ALLTOALL_SIG, ring_one_barrier) COLL_sep \
169 COLL_APPLY(action, COLL_ALLTOALL_SIG, mvapich2) COLL_sep \
170 COLL_APPLY(action, COLL_ALLTOALL_SIG, mvapich2_scatter_dest) COLL_sep \
171 COLL_APPLY(action, COLL_ALLTOALL_SIG, ompi) COLL_sep \
172 COLL_APPLY(action, COLL_ALLTOALL_SIG, mpich) COLL_sep \
173 COLL_APPLY(action, COLL_ALLTOALL_SIG, impi) COLL_sep \
174 COLL_APPLY(action, COLL_ALLTOALL_SIG, automatic)
176 COLL_ALLTOALLS(COLL_PROTO, COLL_NOsep)
181 #define COLL_ALLTOALLV_SIG alltoallv, int, \
182 (void *send_buff, int *send_counts, int *send_disps, MPI_Datatype send_type, \
183 void *recv_buff, int *recv_counts, int *recv_disps, MPI_Datatype recv_type, \
186 #define COLL_ALLTOALLVS(action, COLL_sep) \
187 COLL_APPLY(action, COLL_ALLTOALLV_SIG, bruck) COLL_sep \
188 COLL_APPLY(action, COLL_ALLTOALLV_SIG, pair) COLL_sep \
189 COLL_APPLY(action, COLL_ALLTOALLV_SIG, pair_light_barrier) COLL_sep \
190 COLL_APPLY(action, COLL_ALLTOALLV_SIG, pair_mpi_barrier) COLL_sep \
191 COLL_APPLY(action, COLL_ALLTOALLV_SIG, pair_one_barrier) COLL_sep \
192 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ring) COLL_sep \
193 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ring_light_barrier) COLL_sep \
194 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ring_mpi_barrier) COLL_sep \
195 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ring_one_barrier) COLL_sep \
196 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ompi) COLL_sep \
197 COLL_APPLY(action, COLL_ALLTOALLV_SIG, mpich) COLL_sep \
198 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ompi_basic_linear) COLL_sep \
199 COLL_APPLY(action, COLL_ALLTOALLV_SIG, mvapich2) COLL_sep \
200 COLL_APPLY(action, COLL_ALLTOALLV_SIG, impi) COLL_sep \
201 COLL_APPLY(action, COLL_ALLTOALLV_SIG, automatic)
203 COLL_ALLTOALLVS(COLL_PROTO, COLL_NOsep)
208 #define COLL_BCAST_SIG bcast, int, \
209 (void *buf, int count, MPI_Datatype datatype, \
210 int root, MPI_Comm comm)
212 #define COLL_BCASTS(action, COLL_sep) \
213 COLL_APPLY(action, COLL_BCAST_SIG, arrival_pattern_aware) COLL_sep \
214 COLL_APPLY(action, COLL_BCAST_SIG, arrival_pattern_aware_wait) COLL_sep \
215 COLL_APPLY(action, COLL_BCAST_SIG, arrival_scatter) COLL_sep \
216 COLL_APPLY(action, COLL_BCAST_SIG, binomial_tree) COLL_sep \
217 COLL_APPLY(action, COLL_BCAST_SIG, flattree) COLL_sep \
218 COLL_APPLY(action, COLL_BCAST_SIG, flattree_pipeline) COLL_sep \
219 COLL_APPLY(action, COLL_BCAST_SIG, NTSB) COLL_sep \
220 COLL_APPLY(action, COLL_BCAST_SIG, NTSL) COLL_sep \
221 COLL_APPLY(action, COLL_BCAST_SIG, NTSL_Isend) COLL_sep \
222 COLL_APPLY(action, COLL_BCAST_SIG, scatter_LR_allgather) COLL_sep \
223 COLL_APPLY(action, COLL_BCAST_SIG, scatter_rdb_allgather) COLL_sep \
224 COLL_APPLY(action, COLL_BCAST_SIG, SMP_binary) COLL_sep \
225 COLL_APPLY(action, COLL_BCAST_SIG, SMP_binomial) COLL_sep \
226 COLL_APPLY(action, COLL_BCAST_SIG, SMP_linear) COLL_sep \
227 COLL_APPLY(action, COLL_BCAST_SIG, ompi) COLL_sep \
228 COLL_APPLY(action, COLL_BCAST_SIG, ompi_split_bintree) COLL_sep \
229 COLL_APPLY(action, COLL_BCAST_SIG, ompi_pipeline) COLL_sep \
230 COLL_APPLY(action, COLL_BCAST_SIG, mpich) COLL_sep \
231 COLL_APPLY(action, COLL_BCAST_SIG, mvapich2) COLL_sep \
232 COLL_APPLY(action, COLL_BCAST_SIG, mvapich2_inter_node) COLL_sep \
233 COLL_APPLY(action, COLL_BCAST_SIG, mvapich2_intra_node) COLL_sep \
234 COLL_APPLY(action, COLL_BCAST_SIG, mvapich2_knomial_intra_node) COLL_sep \
235 COLL_APPLY(action, COLL_BCAST_SIG, impi) COLL_sep \
236 COLL_APPLY(action, COLL_BCAST_SIG, automatic)
238 COLL_BCASTS(COLL_PROTO, COLL_NOsep)
244 #define COLL_REDUCE_SIG reduce, int, \
245 (void *buf, void *rbuf, int count, MPI_Datatype datatype, \
246 MPI_Op op, int root, MPI_Comm comm)
248 #define COLL_REDUCES(action, COLL_sep) \
249 COLL_APPLY(action, COLL_REDUCE_SIG, arrival_pattern_aware) COLL_sep \
250 COLL_APPLY(action, COLL_REDUCE_SIG, binomial) COLL_sep \
251 COLL_APPLY(action, COLL_REDUCE_SIG, flat_tree) COLL_sep \
252 COLL_APPLY(action, COLL_REDUCE_SIG, NTSL) COLL_sep \
253 COLL_APPLY(action, COLL_REDUCE_SIG, scatter_gather) COLL_sep \
254 COLL_APPLY(action, COLL_REDUCE_SIG, ompi) COLL_sep \
255 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_chain) COLL_sep \
256 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_pipeline) COLL_sep \
257 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_basic_linear) COLL_sep \
258 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_in_order_binary) COLL_sep \
259 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_binary) COLL_sep \
260 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_binomial) COLL_sep \
261 COLL_APPLY(action, COLL_REDUCE_SIG, mpich) COLL_sep \
262 COLL_APPLY(action, COLL_REDUCE_SIG, mvapich2) COLL_sep \
263 COLL_APPLY(action, COLL_REDUCE_SIG, mvapich2_knomial) COLL_sep \
264 COLL_APPLY(action, COLL_REDUCE_SIG, mvapich2_two_level) COLL_sep \
265 COLL_APPLY(action, COLL_REDUCE_SIG, impi) COLL_sep \
266 COLL_APPLY(action, COLL_REDUCE_SIG, rab) COLL_sep \
267 COLL_APPLY(action, COLL_REDUCE_SIG, automatic)
269 COLL_REDUCES(COLL_PROTO, COLL_NOsep)
274 #define COLL_REDUCE_SCATTER_SIG reduce_scatter, int, \
275 (void *sbuf, void *rbuf, int *rcounts,\
276 MPI_Datatype dtype,MPI_Op op,MPI_Comm comm)
278 #define COLL_REDUCE_SCATTERS(action, COLL_sep) \
279 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, ompi) COLL_sep \
280 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, ompi_basic_recursivehalving) COLL_sep \
281 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, ompi_ring) COLL_sep \
282 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, mpich) COLL_sep \
283 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, mpich_pair) COLL_sep \
284 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, mpich_rdb) COLL_sep \
285 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, mpich_noncomm) COLL_sep \
286 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, mvapich2) COLL_sep \
287 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, impi) COLL_sep \
288 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, automatic)
292 COLL_REDUCE_SCATTERS(COLL_PROTO, COLL_NOsep)
298 #define COLL_SCATTER_SIG scatter, int, \
299 (void *sendbuf, int sendcount, MPI_Datatype sendtype,\
300 void *recvbuf, int recvcount, MPI_Datatype recvtype,\
301 int root, MPI_Comm comm)
303 #define COLL_SCATTERS(action, COLL_sep) \
304 COLL_APPLY(action, COLL_SCATTER_SIG, ompi) COLL_sep \
305 COLL_APPLY(action, COLL_SCATTER_SIG, ompi_basic_linear) COLL_sep \
306 COLL_APPLY(action, COLL_SCATTER_SIG, ompi_binomial) COLL_sep \
307 COLL_APPLY(action, COLL_SCATTER_SIG, mpich) COLL_sep \
308 COLL_APPLY(action, COLL_SCATTER_SIG, mvapich2) COLL_sep \
309 COLL_APPLY(action, COLL_SCATTER_SIG, mvapich2_two_level_binomial) COLL_sep \
310 COLL_APPLY(action, COLL_SCATTER_SIG, mvapich2_two_level_direct) COLL_sep \
311 COLL_APPLY(action, COLL_SCATTER_SIG, impi) COLL_sep \
312 COLL_APPLY(action, COLL_SCATTER_SIG, automatic)
314 COLL_SCATTERS(COLL_PROTO, COLL_NOsep)
319 #define COLL_BARRIER_SIG barrier, int, \
322 #define COLL_BARRIERS(action, COLL_sep) \
323 COLL_APPLY(action, COLL_BARRIER_SIG, ompi) COLL_sep \
324 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_basic_linear) COLL_sep \
325 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_two_procs) COLL_sep \
326 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_tree) COLL_sep \
327 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_bruck) COLL_sep \
328 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_recursivedoubling) COLL_sep \
329 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_doublering) COLL_sep \
330 COLL_APPLY(action, COLL_BARRIER_SIG, mpich) COLL_sep \
331 COLL_APPLY(action, COLL_BARRIER_SIG, mvapich2_pair) COLL_sep \
332 COLL_APPLY(action, COLL_BARRIER_SIG, mvapich2) COLL_sep \
333 COLL_APPLY(action, COLL_BARRIER_SIG, impi) COLL_sep \
334 COLL_APPLY(action, COLL_BARRIER_SIG, automatic)
336 COLL_BARRIERS(COLL_PROTO, COLL_NOsep)