1 /* Copyright (c) 2013-2015. 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. */
7 #include "colls_private.h"
10 //attempt to do a quick autotuning version of the collective,
12 #define TRACE_AUTO_COLL(cat) if (TRACE_is_enabled()){\
13 type_t type = PJ_type_get_or_null (#cat, PJ_type_get_root());\
15 type=PJ_type_event_new(#cat, PJ_type_get_root());\
18 snprintf(cont_name,25, "rank-%d", smpi_process_index());\
19 val_t value = PJ_value_get_or_new(mpi_coll_##cat##_description[i].name,"1.0 1.0 1.0", type);\
20 new_pajeNewEvent (SIMIX_get_clock(), PJ_container_get(cont_name), type, value);\
24 #define AUTOMATIC_COLL_BENCH(cat, ret, args, args2)\
25 ret smpi_coll_tuned_ ## cat ## _ ## automatic(COLL_UNPAREN args)\
27 double time1, time2, time_min=DBL_MAX;\
28 int min_coll=-1, global_coll=-1;\
30 double buf_in, buf_out, max_min=DBL_MAX;\
31 for (i = 0; mpi_coll_##cat##_description[i].name; i++){\
32 if(!strcmp(mpi_coll_##cat##_description[i].name, "automatic"))continue;\
33 if(!strcmp(mpi_coll_##cat##_description[i].name, "default"))continue;\
34 smpi_mpi_barrier(comm);\
36 time1 = SIMIX_get_clock();\
39 mpi_coll_##cat##_description[i].coll) args2 ;\
41 catch (std::exception& ex) {\
44 time2 = SIMIX_get_clock();\
46 smpi_mpi_reduce((void*)&buf_out,(void*)&buf_in, 1, MPI_DOUBLE, MPI_MAX, 0,comm );\
47 if(time2-time1<time_min){\
49 time_min=time2-time1;\
51 if(smpi_comm_rank(comm)==0){\
58 if(smpi_comm_rank(comm)==0){\
59 XBT_WARN("For rank 0, the quickest was %s : %f , but global was %s : %f at max",mpi_coll_##cat##_description[min_coll].name, time_min,mpi_coll_##cat##_description[global_coll].name, max_min);\
61 XBT_WARN("The quickest %s was %s on rank %d and took %f",#cat,mpi_coll_##cat##_description[min_coll].name, smpi_comm_rank(comm), time_min);\
62 return (min_coll!=-1)?MPI_SUCCESS:MPI_ERR_INTERN;\
66 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_ALLGATHERV_SIG, (send_buff, send_count, send_type, recv_buff, recv_count, recv_disps, recv_type, comm));
67 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_ALLREDUCE_SIG, (sbuf, rbuf, rcount, dtype, op, comm));
68 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_GATHER_SIG, (send_buff, send_count, send_type, recv_buff, recv_count, recv_type, root, comm));
69 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_ALLGATHER_SIG, (send_buff,send_count,send_type,recv_buff,recv_count,recv_type,comm));
70 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_ALLTOALL_SIG,(send_buff, send_count, send_type, recv_buff, recv_count, recv_type,comm));
71 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_ALLTOALLV_SIG, (send_buff, send_counts, send_disps, send_type, recv_buff, recv_counts, recv_disps, recv_type, comm));
72 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_BCAST_SIG , (buf, count, datatype, root, comm));
73 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_REDUCE_SIG,(buf,rbuf, count, datatype, op, root, comm));
74 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_REDUCE_SCATTER_SIG ,(sbuf,rbuf, rcounts,dtype,op,comm));
75 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_SCATTER_SIG ,(sendbuf, sendcount, sendtype,recvbuf, recvcount, recvtype,root, comm));
76 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_BARRIER_SIG,(comm));