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(...)
27 #define COLL_ALLGATHER_SIG allgather, int, \
28 (void *send_buff, int send_count, MPI_Datatype send_type, \
29 void *recv_buff, int recv_count, MPI_Datatype recv_type, \
32 #define COLL_ALLGATHERS(action, COLL_sep) \
33 COLL_NOTHING(COLL_APPLY(action, COLL_ALLGATHER_SIG, 2dmesh) COLL_sep) \
34 COLL_NOTHING(COLL_APPLY(action, COLL_ALLGATHER_SIG, 3dmesh) COLL_sep) \
35 COLL_NOTHING(COLL_APPLY(action, COLL_ALLGATHER_SIG, bruck) COLL_sep) \
36 COLL_APPLY(action, COLL_ALLGATHER_SIG, GB) COLL_sep \
37 COLL_APPLY(action, COLL_ALLGATHER_SIG, loosely_lr) COLL_sep \
38 COLL_APPLY(action, COLL_ALLGATHER_SIG, lr) COLL_sep \
39 COLL_APPLY(action, COLL_ALLGATHER_SIG, NTSLR) COLL_sep \
40 COLL_APPLY(action, COLL_ALLGATHER_SIG, NTSLR_NB) COLL_sep \
41 COLL_APPLY(action, COLL_ALLGATHER_SIG, pair) COLL_sep \
42 COLL_APPLY(action, COLL_ALLGATHER_SIG, rdb) COLL_sep \
43 COLL_APPLY(action, COLL_ALLGATHER_SIG, rhv) COLL_sep \
44 COLL_APPLY(action, COLL_ALLGATHER_SIG, ring) COLL_sep \
45 COLL_APPLY(action, COLL_ALLGATHER_SIG, SMP_NTS) COLL_sep \
46 COLL_APPLY(action, COLL_ALLGATHER_SIG, smp_simple) COLL_sep \
47 COLL_APPLY(action, COLL_ALLGATHER_SIG, spreading_simple)
49 COLL_ALLGATHERS(COLL_PROTO, COLL_NOsep)
55 #define COLL_ALLREDUCE_SIG allreduce, int, \
56 (void *sbuf, void *rbuf, int rcount, \
57 MPI_Datatype dtype, MPI_Op op, MPI_Comm comm)
59 #define COLL_ALLREDUCES(action, COLL_sep) \
60 COLL_APPLY(action, COLL_ALLREDUCE_SIG, lr) COLL_sep \
61 COLL_APPLY(action, COLL_ALLREDUCE_SIG, NTS) COLL_sep \
62 COLL_APPLY(action, COLL_ALLREDUCE_SIG, rab1) COLL_sep \
63 COLL_APPLY(action, COLL_ALLREDUCE_SIG, rab2) COLL_sep \
64 COLL_NOTHING(COLL_APPLY(action, COLL_ALLREDUCE_SIG, rab_rdb) COLL_sep) \
65 COLL_NOTHING(COLL_APPLY(action, COLL_ALLREDUCE_SIG, rab_reduce_scatter) COLL_sep) \
66 COLL_APPLY(action, COLL_ALLREDUCE_SIG, rab_rsag) COLL_sep \
67 COLL_APPLY(action, COLL_ALLREDUCE_SIG, rdb) COLL_sep \
68 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_binomial) COLL_sep \
69 COLL_NOTHING(COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_binomial_pipeline) COLL_sep) \
70 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_rdb) COLL_sep \
71 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_rsag) COLL_sep \
72 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_rsag_lr) COLL_sep \
73 COLL_APPLY(action, COLL_ALLREDUCE_SIG, smp_rsag_rab) COLL_sep \
74 COLL_APPLY(action, COLL_ALLREDUCE_SIG, redbcast)
76 COLL_ALLREDUCES(COLL_PROTO, COLL_NOsep)
82 #define COLL_ALLTOALL_SIG alltoall, int, \
83 (void *send_buff, int send_count, MPI_Datatype send_type, \
84 void *recv_buff, int recv_count, MPI_Datatype recv_type, \
87 #define COLL_ALLTOALLS(action, COLL_sep) \
88 COLL_APPLY(action, COLL_ALLTOALL_SIG, 2dmesh) COLL_sep \
89 COLL_APPLY(action, COLL_ALLTOALL_SIG, 3dmesh) COLL_sep \
90 COLL_NOTHING(COLL_APPLY(action, COLL_ALLTOALL_SIG, bruck) COLL_sep) \
91 COLL_APPLY(action, COLL_ALLTOALL_SIG, pair) COLL_sep \
92 COLL_APPLY(action, COLL_ALLTOALL_SIG, pair_light_barrier) COLL_sep \
93 COLL_APPLY(action, COLL_ALLTOALL_SIG, pair_mpi_barrier) COLL_sep \
94 COLL_APPLY(action, COLL_ALLTOALL_SIG, pair_one_barrier) COLL_sep \
95 COLL_APPLY(action, COLL_ALLTOALL_SIG, rdb) COLL_sep \
96 COLL_APPLY(action, COLL_ALLTOALL_SIG, ring) COLL_sep \
97 COLL_APPLY(action, COLL_ALLTOALL_SIG, ring_light_barrier) COLL_sep \
98 COLL_APPLY(action, COLL_ALLTOALL_SIG, ring_mpi_barrier) COLL_sep \
99 COLL_APPLY(action, COLL_ALLTOALL_SIG, ring_one_barrier) COLL_sep \
100 COLL_APPLY(action, COLL_ALLTOALL_SIG, simple)
102 COLL_ALLTOALLS(COLL_PROTO, COLL_NOsep)
107 #define COLL_ALLTOALLV_SIG alltoallv, int, \
108 (void *send_buff, int *send_counts, int *send_disps, MPI_Datatype send_type, \
109 void *recv_buff, int *recv_counts, int *recv_disps, MPI_Datatype recv_type, \
112 #define COLL_ALLTOALLVS(action, COLL_sep) \
113 COLL_APPLY(action, COLL_ALLTOALLV_SIG, bruck) COLL_sep \
114 COLL_APPLY(action, COLL_ALLTOALLV_SIG, pair) COLL_sep \
115 COLL_APPLY(action, COLL_ALLTOALLV_SIG, pair_light_barrier) COLL_sep \
116 COLL_APPLY(action, COLL_ALLTOALLV_SIG, pair_mpi_barrier) COLL_sep \
117 COLL_APPLY(action, COLL_ALLTOALLV_SIG, pair_one_barrier) COLL_sep \
118 COLL_APPLY(action, COLL_ALLTOALLV_SIG, pairwise)
120 COLL_ALLTOALLVS(COLL_PROTO, COLL_NOsep)
125 #define COLL_BCAST_SIG bcast, int, \
126 (void *buf, int count, MPI_Datatype datatype, \
127 int root, MPI_Comm comm)
129 #define COLL_BCASTS(action, COLL_sep) \
130 COLL_APPLY(action, COLL_BCAST_SIG, arrival_nb) COLL_sep \
131 COLL_APPLY(action, COLL_BCAST_SIG, arrival_pattern_aware) COLL_sep \
132 COLL_APPLY(action, COLL_BCAST_SIG, arrival_pattern_aware_wait) COLL_sep \
133 COLL_APPLY(action, COLL_BCAST_SIG, arrival_scatter) COLL_sep \
134 COLL_APPLY(action, COLL_BCAST_SIG, binomial_tree) COLL_sep \
135 COLL_APPLY(action, COLL_BCAST_SIG, flattree) COLL_sep \
136 COLL_APPLY(action, COLL_BCAST_SIG, flattree_pipeline) COLL_sep \
137 COLL_APPLY(action, COLL_BCAST_SIG, NTSB) COLL_sep \
138 COLL_APPLY(action, COLL_BCAST_SIG, NTSL) COLL_sep \
139 COLL_APPLY(action, COLL_BCAST_SIG, NTSL_Isend) COLL_sep \
140 COLL_APPLY(action, COLL_BCAST_SIG, scatter_LR_allgather) COLL_sep \
141 COLL_APPLY(action, COLL_BCAST_SIG, scatter_rdb_allgather) COLL_sep \
142 COLL_APPLY(action, COLL_BCAST_SIG, SMP_binary) COLL_sep \
143 COLL_APPLY(action, COLL_BCAST_SIG, SMP_binomial) COLL_sep \
144 COLL_APPLY(action, COLL_BCAST_SIG, SMP_linear)
146 COLL_BCASTS(COLL_PROTO, COLL_NOsep)
152 #define COLL_REDUCE_SIG reduce, int, \
153 (void *buf, void *rbuf, int count, MPI_Datatype datatype, \
154 MPI_Op op, int root, MPI_Comm comm)
156 #define COLL_REDUCES(action, COLL_sep) \
157 COLL_APPLY(action, COLL_REDUCE_SIG, arrival_pattern_aware) COLL_sep \
158 COLL_APPLY(action, COLL_REDUCE_SIG, binomial) COLL_sep \
159 COLL_APPLY(action, COLL_REDUCE_SIG, flat_tree) COLL_sep \
160 COLL_APPLY(action, COLL_REDUCE_SIG, NTSL) COLL_sep \
161 COLL_APPLY(action, COLL_REDUCE_SIG, scatter_gather)
163 COLL_REDUCES(COLL_PROTO, COLL_NOsep)