MPI_Comm NM_MPI_com(NumericsMatrix* A)
{
if (!A || (A && NM_linearSolverParams(A)->mpi_com == MPI_COMM_NULL))
{
int myid;
int argc = 0;
/* C99 requires that argv[argc] == NULL. With openmpi 1.8, we get a
* segfault if this is not true */
char *argv0 = NULL;
char **argv = &argv0;
CHECK_MPI(MPI_Init(&argc, &argv));
CHECK_MPI(MPI_Comm_rank(MPI_COMM_WORLD, &myid));
if (A)
{
NM_linearSolverParams(A)->mpi_com = MPI_COMM_WORLD;
NM_linearSolverParams(A)->mpi_com_init = 1;
}
}
if(A)
{
return NM_linearSolverParams(A)->mpi_com;
}
else
{
return MPI_COMM_WORLD;
}
}
int main(int argc, char ** argv)
{
int rc;
/* These are the desired and available thread support.
A hybrid code where all MPI calls are made from the main thread can used FUNNELED.
If threads are making MPI calls, MULTIPLE is appropriate. */
int requested = MPI_THREAD_FUNNELED, provided;
/* MPICH2 will be substantially more efficient than OpenMPI
for MPI_THREAD_{FUNNELED,SERIALIZED} but this is unlikely
to be a serious bottleneck. */
rc = MPI_Init_thread(&argc, &argv, requested, &provided); CHECK_MPI(rc);
if (provided<requested)
{
printf("MPI_Init_thread provided %s when %s was requested. Exiting. \n",
MPI_THREAD_STRING(provided), MPI_THREAD_STRING(requested) );
exit(1);
}
int world_size, world_rank;
rc = MPI_Comm_size(MPI_COMM_WORLD,&world_size); CHECK_MPI(rc);
rc = MPI_Comm_rank(MPI_COMM_WORLD,&world_rank); CHECK_MPI(rc);
int root = 0, count = 1;
/* the ternary is often branchless... */
long i, n = (argc>1 ? atol(argv[1]) : 100000);
rc = MPI_Bcast(&n, count, MPI_LONG, root, MPI_COMM_WORLD); CHECK_MPI(rc);
if (world_rank==0)
printf("%d: using %ld samples.\n", world_rank, world_size*n);
/* seed the RNG with something unique to a rank */
srand(world_rank);
long in = 0, total = 0;
for (i=0;i<n;i++)
{
register double x = (double)rand()/(double)RAND_MAX;
register double y = (double)rand()/(double)RAND_MAX;
register double z = x*x + y*y;
if (z<1.0) in++;
}
rc = MPI_Reduce(&in, &total, count, MPI_LONG, MPI_SUM, root, MPI_COMM_WORLD); CHECK_MPI(rc);
double pi = 4.0*(double)total/(world_size*n);
if (world_rank==0)
printf("%d: pi = %12.8lf.\n", world_rank, pi);
MPI_Finalize();
return 0;
}
void sopalin_launch_thread(void * sopalin_data_ref,
PASTIX_INT procnum, PASTIX_INT procnbr, void *ptr, PASTIX_INT verbose,
PASTIX_INT calc_thrdnbr, void * (*calc_routine)(void *), void *calc_data,
PASTIX_INT comm_thrdnbr, void * (*comm_routine)(void *), void *comm_data,
PASTIX_INT ooc_thrdnbr, void * (*ooc_routine)(void *), void *ooc_data)
{
sopthread_data_t *d = NULL;
pthread_t *calltab = NULL;
PASTIX_INT i;
PASTIX_INT ret;
PASTIX_INT thrdnbr;
PASTIX_INT thrdnbr_wo_ooc;
Sopalin_Data_t *sopalin_data = sopalin_data_ref;
(void)procnbr; (void)ptr;
thrdnbr = calc_thrdnbr + comm_thrdnbr + ooc_thrdnbr ;
thrdnbr_wo_ooc = calc_thrdnbr + comm_thrdnbr;
if (verbose > API_VERBOSE_NO)
print_one("Launching %d threads"
" (%d commputation, %d communication, %d out-of-core)\n",
(int) thrdnbr, (int)calc_thrdnbr,
(int)comm_thrdnbr, (int)ooc_thrdnbr);
MALLOC_INTERN(calltab, thrdnbr, pthread_t);
MALLOC_INTERN(d, thrdnbr, sopthread_data_t);
if (calc_thrdnbr > 1)
{
int comm_size;
CHECK_MPI(MPI_Comm_size(MPI_COMM_WORLD, &comm_size));
if (comm_size > 1)
CHECK_THREAD_LEVEL(sopalin_data->sopar->iparm[IPARM_THREAD_COMM_MODE]);
}
/* Lancement des threads de calcul */
for (i=0;i<calc_thrdnbr;i++)
{
pthread_attr_t attr;
pthread_attr_init(&attr);
#ifdef MARCEL2
{
int cpu = (i+procnum*thrdnbr)%sysconf(_SC_NPROCESSORS_ONLN);
if (thrdnbr <= sysconf(_SC_NPROCESSORS_ONLN))
marcel_attr_setvpset(&attr, MARCEL_VPSET_VP(cpu));
}
#endif
d[i].me = i;
d[i].data = calc_data;
ret = pthread_create(&calltab[i],&attr,calc_routine,(void *)&d[i]);
if (ret) {errorPrint("thread create."); EXIT(MOD_SOPALIN,THREAD_ERR);}
}
/* Lancement des threads de chargement ooc */
for (i=thrdnbr_wo_ooc; i<thrdnbr; i++)
{
pthread_attr_t attr;
pthread_attr_init(&attr);
#ifdef MARCEL2
{
int cpu = (i+procnum*thrdnbr)%sysconf(_SC_NPROCESSORS_ONLN);
if (thrdnbr <= sysconf(_SC_NPROCESSORS_ONLN))
marcel_attr_setvpset(&attr, MARCEL_VPSET_VP(cpu));
}
#endif
d[i].me = i;
d[i].data = ooc_data;
ret = pthread_create(&calltab[i],&attr,ooc_routine,(void *)&d[i]);
if (ret) {errorPrint("thread create."); EXIT(MOD_SOPALIN,THREAD_ERR);}
}
/* Lancement des threads de communication */
if ((comm_thrdnbr > 0) && (comm_routine != NULL))
{
/* print_one("-- Options Communication --\n"); */
/* print_one(" - Type : %d\n", sopar->type_comm); */
/* print_one(" - Nbthread : %d\n", sopar->nbthrdcomm); */
if (comm_thrdnbr > 1)
{
for (i=calc_thrdnbr;i<thrdnbr_wo_ooc;i++)
{
pthread_attr_t attr;
pthread_attr_init(&attr);
d[i].me = i;
d[i].data = comm_data;
ret = pthread_create(&calltab[i], &attr,
comm_routine, (void *)&d[i]);
if (ret)
{
errorPrint("thread create.");
EXIT(MOD_SOPALIN,THREAD_ERR);
}
}
}
else
{
d[calc_thrdnbr].me = calc_thrdnbr;
d[calc_thrdnbr].data = comm_data;
comm_routine((void*)&d[calc_thrdnbr]);
}
}
/* Recuperation de tous les threads lancés */
for (i=0;i<thrdnbr;i++)
{
/* On ne recupere pas le thread qd il a pas été lancé */
if ((comm_thrdnbr == 1) && (i == calc_thrdnbr)) continue;
ret = pthread_join(calltab[i],(void**)NULL);
if (ret) {errorPrint("thread join."); EXIT(MOD_SOPALIN,THREAD_ERR);}
}
memFree_null(calltab);
memFree_null(d);
}