Logo AND Algorithmique Numérique Distribuée

Public GIT Repository
cleanups and simplifications to our build system
[simgrid.git] / src / smpi / colls / smpi_automatic_selector.c
1 /* Copyright (c) 2013-2014. The SimGrid Team.
2  * All rights reserved.                                                     */
3
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. */
6
7 #include "colls_private.h"
8 #ifdef HAVE_MC
9 #include "mc/mc_private.h"
10 #endif
11 #include <float.h>
12
13 //attempt to do a quick autotuning version of the collective,
14
15 #define TRACE_AUTO_COLL(cat) if (TRACE_is_enabled()){\
16         type_t type = PJ_type_get_or_null (#cat, PJ_type_get_root());\
17          if (!type){\
18              type=PJ_type_event_new(#cat, PJ_type_get_root());\
19          }\
20          char cont_name[25];\
21          sprintf(cont_name, "rank-%d", smpi_process_index());\
22          val_t value = PJ_value_get_or_new(mpi_coll_##cat##_description[i].name,"1.0 1.0 1.0", type);\
23          new_pajeNewEvent (SIMIX_get_clock(), PJ_container_get(cont_name), type, value);\
24       }
25
26
27 #define AUTOMATIC_COLL_BENCH(cat, ret, args, args2)\
28     ret smpi_coll_tuned_ ## cat ## _ ## automatic(COLL_UNPAREN args)\
29 {\
30   double time1, time2, time_min=DBL_MAX;\
31   volatile int min_coll=-1, global_coll=-1;\
32   volatile int i;\
33   xbt_ex_t ex;\
34   double buf_in, buf_out, max_min=DBL_MAX;\
35   for (i = 0; mpi_coll_##cat##_description[i].name; i++){\
36       if(!strcmp(mpi_coll_##cat##_description[i].name, "automatic"))continue;\
37       if(!strcmp(mpi_coll_##cat##_description[i].name, "default"))continue;\
38       smpi_mpi_barrier(comm);\
39       TRACE_AUTO_COLL(cat)\
40       time1 = SIMIX_get_clock();\
41       TRY{\
42       ((int (*) args)\
43           mpi_coll_##cat##_description[i].coll) args2 ;\
44       }\
45       CATCH(ex) {\
46         xbt_ex_free(ex);\
47         continue;\
48       }\
49       time2 = SIMIX_get_clock();\
50       buf_out=time2-time1;\
51       smpi_mpi_reduce((void*)&buf_out,(void*)&buf_in, 1, MPI_DOUBLE, MPI_MAX, 0,comm );\
52       if(time2-time1<time_min){\
53           min_coll=i;\
54           time_min=time2-time1;\
55       }\
56       if(smpi_comm_rank(comm)==0){\
57           if(buf_in<max_min){\
58               max_min=buf_in;\
59               global_coll=i;\
60           }\
61       }\
62   }\
63   if(smpi_comm_rank(comm)==0){\
64       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);\
65   }else\
66   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);\
67   return (min_coll!=-1)?MPI_SUCCESS:MPI_ERR_INTERN;\
68 }\
69
70
71 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_ALLGATHERV_SIG, (send_buff, send_count, send_type, recv_buff, recv_count, recv_disps, recv_type, comm));
72 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_ALLREDUCE_SIG, (sbuf, rbuf, rcount, dtype, op, comm));
73 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_GATHER_SIG, (send_buff, send_count, send_type, recv_buff, recv_count, recv_type, root, comm));
74 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_ALLGATHER_SIG, (send_buff,send_count,send_type,recv_buff,recv_count,recv_type,comm));
75 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_ALLTOALL_SIG,(send_buff, send_count, send_type, recv_buff, recv_count, recv_type,comm));
76 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));
77 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_BCAST_SIG , (buf, count, datatype, root, comm));
78 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_REDUCE_SIG,(buf,rbuf, count, datatype, op, root, comm));
79 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_REDUCE_SCATTER_SIG ,(sbuf,rbuf, rcounts,dtype,op,comm));
80 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_SCATTER_SIG ,(sendbuf, sendcount, sendtype,recvbuf, recvcount, recvtype,root, comm));
81 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_BARRIER_SIG,(comm));