6 #include "smpi/private.h"
9 #define COLL_DESCRIPTION(cat, ret, args, name) \
11 # cat " " # name " collective",\
12 smpi_coll_tuned_ ## cat ## _ ## name}
14 #define COLL_PROTO(cat, ret, args, name) \
15 ret smpi_coll_tuned_ ## cat ## _ ## name(COLL_UNPAREN args);
16 #define COLL_UNPAREN(...) __VA_ARGS__
18 #define COLL_APPLY(action, sig, name) action(sig, name)
21 #define COLL_NOTHING(...)
26 #define COLL_GATHER_SIG gather, int, \
27 (void *send_buff, int send_count, MPI_Datatype send_type, \
28 void *recv_buff, int recv_count, MPI_Datatype recv_type, \
29 int root, MPI_Comm comm)
31 #define COLL_GATHERS(action, COLL_sep) \
32 COLL_APPLY(action, COLL_GATHER_SIG, ompi) COLL_sep \
33 COLL_APPLY(action, COLL_GATHER_SIG, ompi_basic_linear) COLL_sep \
34 COLL_APPLY(action, COLL_GATHER_SIG, ompi_binomial) COLL_sep \
35 COLL_APPLY(action, COLL_GATHER_SIG, ompi_linear_sync) COLL_sep \
36 COLL_APPLY(action, COLL_GATHER_SIG, mpich) COLL_sep \
37 COLL_APPLY(action, COLL_GATHER_SIG, automatic)
41 COLL_GATHERS(COLL_PROTO, COLL_NOsep)
46 #define COLL_ALLGATHER_SIG allgather, int, \
47 (void *send_buff, int send_count, MPI_Datatype send_type, \
48 void *recv_buff, int recv_count, MPI_Datatype recv_type, \
51 #define COLL_ALLGATHERS(action, COLL_sep) \
52 COLL_APPLY(action, COLL_ALLGATHER_SIG, 2dmesh) COLL_sep \
53 COLL_APPLY(action, COLL_ALLGATHER_SIG, 3dmesh) COLL_sep \
54 COLL_APPLY(action, COLL_ALLGATHER_SIG, bruck) COLL_sep \
55 COLL_APPLY(action, COLL_ALLGATHER_SIG, GB) COLL_sep \
56 COLL_APPLY(action, COLL_ALLGATHER_SIG, loosely_lr) COLL_sep \
57 COLL_APPLY(action, COLL_ALLGATHER_SIG, NTSLR) COLL_sep \
58 COLL_APPLY(action, COLL_ALLGATHER_SIG, NTSLR_NB) COLL_sep \
59 COLL_APPLY(action, COLL_ALLGATHER_SIG, pair) COLL_sep \
60 COLL_APPLY(action, COLL_ALLGATHER_SIG, rdb) COLL_sep \
61 COLL_APPLY(action, COLL_ALLGATHER_SIG, rhv) COLL_sep \
62 COLL_APPLY(action, COLL_ALLGATHER_SIG, ring) COLL_sep \
63 COLL_APPLY(action, COLL_ALLGATHER_SIG, SMP_NTS) COLL_sep \
64 COLL_APPLY(action, COLL_ALLGATHER_SIG, smp_simple) COLL_sep \
65 COLL_APPLY(action, COLL_ALLGATHER_SIG, spreading_simple) COLL_sep \
66 COLL_APPLY(action, COLL_ALLGATHER_SIG, ompi) COLL_sep \
67 COLL_APPLY(action, COLL_ALLGATHER_SIG, ompi_neighborexchange) COLL_sep \
68 COLL_APPLY(action, COLL_ALLGATHER_SIG, mpich) COLL_sep \
69 COLL_APPLY(action, COLL_ALLGATHER_SIG, automatic)
72 COLL_ALLGATHERS(COLL_PROTO, COLL_NOsep)
77 #define COLL_ALLGATHERV_SIG allgatherv, int, \
78 (void *send_buff, int send_count, MPI_Datatype send_type, \
79 void *recv_buff, int *recv_count, int *recv_disps, \
80 MPI_Datatype recv_type, MPI_Comm comm)
82 #define COLL_ALLGATHERVS(action, COLL_sep) \
83 COLL_APPLY(action, COLL_ALLGATHERV_SIG, GB) COLL_sep \
84 COLL_APPLY(action, COLL_ALLGATHERV_SIG, pair) COLL_sep \
85 COLL_APPLY(action, COLL_ALLGATHERV_SIG, ring) COLL_sep \
86 COLL_APPLY(action, COLL_ALLGATHERV_SIG, ompi) COLL_sep \
87 COLL_APPLY(action, COLL_ALLGATHERV_SIG, ompi_neighborexchange) COLL_sep \
88 COLL_APPLY(action, COLL_ALLGATHERV_SIG, ompi_bruck) COLL_sep \
89 COLL_APPLY(action, COLL_ALLGATHERV_SIG, mpich) COLL_sep \
90 COLL_APPLY(action, COLL_ALLGATHERV_SIG, mpich_rdb) COLL_sep \
91 COLL_APPLY(action, COLL_ALLGATHERV_SIG, mpich_ring) COLL_sep \
92 COLL_APPLY(action, COLL_ALLGATHERV_SIG, automatic)
94 COLL_ALLGATHERVS(COLL_PROTO, COLL_NOsep)
99 #define COLL_ALLREDUCE_SIG allreduce, int, \
100 (void *sbuf, void *rbuf, int rcount, \
101 MPI_Datatype dtype, MPI_Op op, MPI_Comm comm)
103 #define COLL_ALLREDUCES(action, COLL_sep) \
104 COLL_APPLY(action, COLL_ALLREDUCE_SIG, lr) COLL_sep \
105 COLL_APPLY(action, COLL_ALLREDUCE_SIG, rab1) COLL_sep \
106 COLL_APPLY(action, COLL_ALLREDUCE_SIG, rab2) COLL_sep \
107 COLL_APPLY(action, COLL_ALLREDUCE_SIG, rab_rdb) COLL_sep \
108 COLL_NOTHING(COLL_APPLY(action, COLL_ALLREDUCE_SIG, rab_reduce_scatter) COLL_sep) \
109 COLL_APPLY(action, COLL_ALLREDUCE_SIG, rab_rsag) COLL_sep \
110 COLL_APPLY(action, COLL_ALLREDUCE_SIG, rdb) COLL_sep \
111 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_binomial) COLL_sep \
112 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_binomial_pipeline) COLL_sep \
113 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_rdb) COLL_sep \
114 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_rsag) COLL_sep \
115 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_rsag_lr) COLL_sep \
116 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_rsag_rab) COLL_sep \
117 COLL_APPLY(action, COLL_ALLREDUCE_SIG, redbcast) COLL_sep \
118 COLL_APPLY(action, COLL_ALLREDUCE_SIG, ompi) COLL_sep \
119 COLL_APPLY(action, COLL_ALLREDUCE_SIG, ompi_ring_segmented) COLL_sep \
120 COLL_APPLY(action, COLL_ALLREDUCE_SIG, mpich) COLL_sep \
121 COLL_APPLY(action, COLL_ALLREDUCE_SIG, automatic)
123 COLL_ALLREDUCES(COLL_PROTO, COLL_NOsep)
129 #define COLL_ALLTOALL_SIG alltoall, int, \
130 (void *send_buff, int send_count, MPI_Datatype send_type, \
131 void *recv_buff, int recv_count, MPI_Datatype recv_type, \
134 #define COLL_ALLTOALLS(action, COLL_sep) \
135 COLL_APPLY(action, COLL_ALLTOALL_SIG, 2dmesh) COLL_sep \
136 COLL_APPLY(action, COLL_ALLTOALL_SIG, 3dmesh) COLL_sep \
137 COLL_NOTHING(COLL_APPLY(action, COLL_ALLTOALL_SIG, bruck) COLL_sep) \
138 COLL_APPLY(action, COLL_ALLTOALL_SIG, pair) COLL_sep \
139 COLL_APPLY(action, COLL_ALLTOALL_SIG, pair_light_barrier) COLL_sep \
140 COLL_APPLY(action, COLL_ALLTOALL_SIG, pair_mpi_barrier) COLL_sep \
141 COLL_APPLY(action, COLL_ALLTOALL_SIG, pair_one_barrier) COLL_sep \
142 COLL_APPLY(action, COLL_ALLTOALL_SIG, rdb) COLL_sep \
143 COLL_APPLY(action, COLL_ALLTOALL_SIG, ring) COLL_sep \
144 COLL_APPLY(action, COLL_ALLTOALL_SIG, ring_light_barrier) COLL_sep \
145 COLL_APPLY(action, COLL_ALLTOALL_SIG, ring_mpi_barrier) COLL_sep \
146 COLL_APPLY(action, COLL_ALLTOALL_SIG, ring_one_barrier) COLL_sep \
147 COLL_APPLY(action, COLL_ALLTOALL_SIG, simple) COLL_sep \
148 COLL_APPLY(action, COLL_ALLTOALL_SIG, ompi) COLL_sep \
149 COLL_APPLY(action, COLL_ALLTOALL_SIG, mpich) COLL_sep \
150 COLL_APPLY(action, COLL_ALLTOALL_SIG, ompi_pairwise) COLL_sep \
151 COLL_APPLY(action, COLL_ALLTOALL_SIG, automatic)
153 COLL_ALLTOALLS(COLL_PROTO, COLL_NOsep)
158 #define COLL_ALLTOALLV_SIG alltoallv, int, \
159 (void *send_buff, int *send_counts, int *send_disps, MPI_Datatype send_type, \
160 void *recv_buff, int *recv_counts, int *recv_disps, MPI_Datatype recv_type, \
163 #define COLL_ALLTOALLVS(action, COLL_sep) \
164 COLL_APPLY(action, COLL_ALLTOALLV_SIG, bruck) COLL_sep \
165 COLL_APPLY(action, COLL_ALLTOALLV_SIG, pair) COLL_sep \
166 COLL_APPLY(action, COLL_ALLTOALLV_SIG, pair_light_barrier) COLL_sep \
167 COLL_APPLY(action, COLL_ALLTOALLV_SIG, pair_mpi_barrier) COLL_sep \
168 COLL_APPLY(action, COLL_ALLTOALLV_SIG, pair_one_barrier) COLL_sep \
169 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ring) COLL_sep \
170 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ring_light_barrier) COLL_sep \
171 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ring_mpi_barrier) COLL_sep \
172 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ring_one_barrier) COLL_sep \
173 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ompi) COLL_sep \
174 COLL_APPLY(action, COLL_ALLTOALLV_SIG, mpich) COLL_sep \
175 COLL_APPLY(action, COLL_ALLTOALLV_SIG, ompi_basic_linear) COLL_sep \
176 COLL_APPLY(action, COLL_ALLTOALLV_SIG, automatic)
178 COLL_ALLTOALLVS(COLL_PROTO, COLL_NOsep)
183 #define COLL_BCAST_SIG bcast, int, \
184 (void *buf, int count, MPI_Datatype datatype, \
185 int root, MPI_Comm comm)
187 #define COLL_BCASTS(action, COLL_sep) \
188 COLL_APPLY(action, COLL_BCAST_SIG, arrival_nb) COLL_sep \
189 COLL_APPLY(action, COLL_BCAST_SIG, arrival_pattern_aware) COLL_sep \
190 COLL_APPLY(action, COLL_BCAST_SIG, arrival_pattern_aware_wait) COLL_sep \
191 COLL_APPLY(action, COLL_BCAST_SIG, arrival_scatter) COLL_sep \
192 COLL_APPLY(action, COLL_BCAST_SIG, binomial_tree) COLL_sep \
193 COLL_APPLY(action, COLL_BCAST_SIG, flattree) COLL_sep \
194 COLL_APPLY(action, COLL_BCAST_SIG, flattree_pipeline) COLL_sep \
195 COLL_APPLY(action, COLL_BCAST_SIG, NTSB) COLL_sep \
196 COLL_APPLY(action, COLL_BCAST_SIG, NTSL) COLL_sep \
197 COLL_APPLY(action, COLL_BCAST_SIG, NTSL_Isend) COLL_sep \
198 COLL_APPLY(action, COLL_BCAST_SIG, scatter_LR_allgather) COLL_sep \
199 COLL_APPLY(action, COLL_BCAST_SIG, scatter_rdb_allgather) COLL_sep \
200 COLL_APPLY(action, COLL_BCAST_SIG, SMP_binary) COLL_sep \
201 COLL_APPLY(action, COLL_BCAST_SIG, SMP_binomial) COLL_sep \
202 COLL_APPLY(action, COLL_BCAST_SIG, SMP_linear) COLL_sep \
203 COLL_APPLY(action, COLL_BCAST_SIG, ompi) COLL_sep \
204 COLL_APPLY(action, COLL_BCAST_SIG, ompi_split_bintree) COLL_sep \
205 COLL_APPLY(action, COLL_BCAST_SIG, ompi_pipeline) COLL_sep \
206 COLL_APPLY(action, COLL_BCAST_SIG, mpich) COLL_sep \
207 COLL_APPLY(action, COLL_BCAST_SIG, automatic)
209 COLL_BCASTS(COLL_PROTO, COLL_NOsep)
215 #define COLL_REDUCE_SIG reduce, int, \
216 (void *buf, void *rbuf, int count, MPI_Datatype datatype, \
217 MPI_Op op, int root, MPI_Comm comm)
219 #define COLL_REDUCES(action, COLL_sep) \
220 COLL_APPLY(action, COLL_REDUCE_SIG, arrival_pattern_aware) COLL_sep \
221 COLL_APPLY(action, COLL_REDUCE_SIG, binomial) COLL_sep \
222 COLL_APPLY(action, COLL_REDUCE_SIG, flat_tree) COLL_sep \
223 COLL_APPLY(action, COLL_REDUCE_SIG, NTSL) COLL_sep \
224 COLL_APPLY(action, COLL_REDUCE_SIG, scatter_gather) COLL_sep \
225 COLL_APPLY(action, COLL_REDUCE_SIG, ompi) COLL_sep \
226 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_chain) COLL_sep \
227 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_pipeline) COLL_sep \
228 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_basic_linear) COLL_sep \
229 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_in_order_binary) COLL_sep \
230 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_binary) COLL_sep \
231 COLL_APPLY(action, COLL_REDUCE_SIG, ompi_binomial) COLL_sep \
232 COLL_APPLY(action, COLL_REDUCE_SIG, mpich) COLL_sep \
233 COLL_APPLY(action, COLL_REDUCE_SIG, automatic)
235 COLL_REDUCES(COLL_PROTO, COLL_NOsep)
240 #define COLL_REDUCE_SCATTER_SIG reduce_scatter, int, \
241 (void *sbuf, void *rbuf, int *rcounts,\
242 MPI_Datatype dtype,MPI_Op op,MPI_Comm comm)
244 #define COLL_REDUCE_SCATTERS(action, COLL_sep) \
245 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, ompi) COLL_sep \
246 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, ompi_basic_recursivehalving) COLL_sep \
247 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, ompi_ring) COLL_sep \
248 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, mpich) COLL_sep \
249 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, mpich_pair) COLL_sep \
250 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, mpich_rdb) COLL_sep \
251 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, mpich_noncomm) COLL_sep \
252 COLL_APPLY(action, COLL_REDUCE_SCATTER_SIG, automatic)
256 COLL_REDUCE_SCATTERS(COLL_PROTO, COLL_NOsep)
262 #define COLL_SCATTER_SIG scatter, int, \
263 (void *sendbuf, int sendcount, MPI_Datatype sendtype,\
264 void *recvbuf, int recvcount, MPI_Datatype recvtype,\
265 int root, MPI_Comm comm)
267 #define COLL_SCATTERS(action, COLL_sep) \
268 COLL_APPLY(action, COLL_SCATTER_SIG, ompi) COLL_sep \
269 COLL_APPLY(action, COLL_SCATTER_SIG, ompi_basic_linear) COLL_sep \
270 COLL_APPLY(action, COLL_SCATTER_SIG, ompi_binomial) COLL_sep \
271 COLL_APPLY(action, COLL_SCATTER_SIG, mpich) COLL_sep \
272 COLL_APPLY(action, COLL_SCATTER_SIG, automatic)
274 COLL_SCATTERS(COLL_PROTO, COLL_NOsep)
279 #define COLL_BARRIER_SIG barrier, int, \
282 #define COLL_BARRIERS(action, COLL_sep) \
283 COLL_APPLY(action, COLL_BARRIER_SIG, ompi) COLL_sep \
284 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_basic_linear) COLL_sep \
285 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_two_procs) COLL_sep \
286 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_tree) COLL_sep \
287 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_bruck) COLL_sep \
288 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_recursivedoubling) COLL_sep \
289 COLL_APPLY(action, COLL_BARRIER_SIG, ompi_doublering) COLL_sep \
290 COLL_APPLY(action, COLL_BARRIER_SIG, mpich) COLL_sep \
291 COLL_APPLY(action, COLL_BARRIER_SIG, automatic)
293 COLL_BARRIERS(COLL_PROTO, COLL_NOsep)