1 #include "colls_private.h"
3 #include "mc/mc_private.h"
5 //attempt to do a quick autotuning version of the collective,
8 #define TRACE_AUTO_COLL(cat) if (TRACE_is_enabled()){\
9 type_t type = PJ_type_get_or_null (#cat, PJ_type_get_root());\
11 type=PJ_type_event_new(#cat, PJ_type_get_root());\
13 char* cont_name=malloc(25*sizeof(char*));\
14 sprintf(cont_name, "rank-%d", smpi_process_index());\
15 val_t value = PJ_value_get_or_new(mpi_coll_##cat##_description[i].name,"1.0 1.0 1.0", type);\
16 new_pajeNewEvent (SIMIX_get_clock(), PJ_container_get(cont_name), type, value);\
19 #define TRACE_AUTO_COLL(cat)
23 #define AUTOMATIC_COLL_BENCH(cat, ret, args, args2)\
24 ret smpi_coll_tuned_ ## cat ## _ ## automatic(COLL_UNPAREN args)\
26 double time1, time2, time_min=INT_MAX;\
27 int min_coll=-1, global_coll=-1;\
29 double buf_in, buf_out, max_min=INT_MAX;\
30 for (i = 0; mpi_coll_##cat##_description[i].name; i++){\
31 if(!strcmp(mpi_coll_##cat##_description[i].name, "automatic"))continue;\
32 if(!strcmp(mpi_coll_##cat##_description[i].name, "default"))continue;\
33 smpi_mpi_barrier(comm);\
35 time1 = SIMIX_get_clock();\
37 mpi_coll_##cat##_description[i].coll) args2 ;\
38 time2 = SIMIX_get_clock();\
40 smpi_mpi_reduce((void*)&buf_out,(void*)&buf_in, 1, MPI_DOUBLE, MPI_MAX, 0,comm );\
41 if(time2-time1<time_min){\
43 time_min=time2-time1;\
45 if(smpi_comm_rank(comm)==0){\
52 if(smpi_comm_rank(comm)==0){\
53 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);\
55 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);\
56 return (min_coll!=-1)?MPI_SUCCESS:MPI_ERR_INTERN;\
60 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_ALLGATHERV_SIG, (send_buff, send_count, send_type, recv_buff, recv_count, recv_disps, recv_type, comm));
61 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_ALLREDUCE_SIG, (sbuf, rbuf, rcount, dtype, op, comm));
62 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_GATHER_SIG, (send_buff, send_count, send_type, recv_buff, recv_count, recv_type, root, comm));
63 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_ALLGATHER_SIG, (send_buff,send_count,send_type,recv_buff,recv_count,recv_type,comm));
64 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_ALLTOALL_SIG,(send_buff, send_count, send_type, recv_buff, recv_count, recv_type,comm));
65 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));
66 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_BCAST_SIG , (buf, count, datatype, root, comm));
67 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_REDUCE_SIG,(buf,rbuf, count, datatype, op, root, comm));
68 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_REDUCE_SCATTER_SIG ,(sbuf,rbuf, rcounts,dtype,op,comm));
69 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_SCATTER_SIG ,(sendbuf, sendcount, sendtype,recvbuf, recvcount, recvtype,root, comm));
70 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_BARRIER_SIG,(comm));