void
FORTRANIFY (shmem_fcollect4) (void *target, const void *source, int *nelems,
int *PE_start, int *logPE_stride, int *PE_size,
int *pSync)
{
shmem_fcollect32 (target, source,
*nelems, *PE_start, *logPE_stride, *PE_size,
(long *) pSync);
}
void
fcollect(int *target, int *src, int elements, int me, int npes, int loops)
{
int i;
double start_time, elapsed_time;
long total_bytes = loops * elements * sizeof(*src);
long *ps, *pSync, *pSync1;
pSync = (long*) shmem_malloc( 2 * sizeof(long) * _SHMEM_COLLECT_SYNC_SIZE );
pSync1 = &pSync[_SHMEM_COLLECT_SYNC_SIZE];
for (i = 0; i < _SHMEM_COLLECT_SYNC_SIZE; i++) {
pSync[i] = pSync1[i] = _SHMEM_SYNC_VALUE;
}
target = (int *) shmem_malloc( elements * sizeof(*target) * npes );
if (me==0 && Verbose) {
fprintf(stdout,"%s: %d loops of fcollect32(%ld bytes) over %d PEs: ",
__FUNCTION__,loops,(elements*sizeof(*src)),npes);
fflush(stdout);
}
shmem_barrier_all();
start_time = shmemx_wtime();
for(i = 0; i < loops; i++) {
ps = &pSync[(i&1)];
shmem_fcollect32( target, src, elements, 0, 0, npes, ps );
}
elapsed_time = shmemx_wtime() - start_time;
if (me==0 && Verbose) {
printf("%7.3f secs\n", elapsed_time);
printf(" %7.5f usecs / fcollect32(), %ld Kbytes @ %7.4f MB/sec\n\n",
(elapsed_time/((double)loops*npes))*1000000.0,
(total_bytes/1024),
((double)total_bytes/(1024.0*1024.0)) / elapsed_time );
}
shmem_barrier_all();
shmem_free(target);
shmem_free( pSync );
shmem_barrier_all();
}
int
main (void)
{
int i;
int *target;
int *source;
int me, npes;
struct timeval start, end;
long time_taken, start_time, end_time;
shmem_init ();
me = shmem_my_pe ();
npes = shmem_n_pes ();
source = (int *) shmem_malloc (N_ELEMENTS * sizeof (*source));
time_taken = 0;
for (i = 0; i < N_ELEMENTS; i += 1) {
source[i] = (i + 1) * 10 + me;
}
target = (int *) shmem_malloc (N_ELEMENTS * sizeof (*target) * npes);
for (i = 0; i < N_ELEMENTS * npes; i += 1) {
target[i] = -90;
}
for (i = 0; i < _SHMEM_COLLECT_SYNC_SIZE; i += 1) {
pSyncA[i] = _SHMEM_SYNC_VALUE;
pSyncB[i] = _SHMEM_SYNC_VALUE;
}
shmem_barrier_all ();
for (i = 0; i < 10000; i++) {
gettimeofday (&start, NULL);
start_time = (start.tv_sec * 1000000.0) + start.tv_usec;
/* alternate between 2 pSync arrays to synchronize consequent
collectives of even and odd iterations */
if (i % 2) {
shmem_fcollect32 (target, source, N_ELEMENTS, 0, 0, npes, pSyncA);
}
else {
shmem_fcollect32 (target, source, N_ELEMENTS, 0, 0, npes, pSyncB);
}
gettimeofday (&end, NULL);
end_time = (end.tv_sec * 1000000.0) + end.tv_usec;
if (me == 0) {
time_taken = time_taken + (end_time - start_time);
}
}
if (me == 0) {
printf
("Time required to collect %d bytes of data, with %d PEs is %ld microseconds\n",
(4 * N_ELEMENTS * npes), npes, time_taken / 10000);
}
shmem_barrier_all ();
shmem_free (target);
shmem_free (source);
shmem_finalize ();
return 0;
}
int
main (int argc, char *argv[])
{
int *sray, *rray;
int *sdisp, *scounts, *rdisp, *rcounts, *rcounts_full;
int ssize, rsize, i, k, j;
float z;
init_it (&argc, &argv);
scounts = (int *) shmem_malloc (sizeof (int) * numnodes);
rcounts = (int *) shmem_malloc (sizeof (int) * numnodes);
rcounts_full = (int *) shmem_malloc (sizeof (int) * numnodes * numnodes);
sdisp = (int *) shmem_malloc (sizeof (int) * numnodes);
rdisp = (int *) shmem_malloc (sizeof (int) * numnodes);
/*
! seed the random number generator with a ! different number on each
processor */
seed_random (myid);
/* find out how much data to send */
for (i = 0; i < numnodes; i++) {
random_number (&z);
scounts[i] = (int) (5.0 * z) + 1;
}
printf ("myid= %d scounts=%d %d %d %d\n", myid, scounts[0], scounts[1],
scounts[2], scounts[3]);
printf ("\n");
/* tell the other processors how much data is coming */
// mpi_err = MPI_Alltoall(scounts,1,MPI_INT, rcounts,1,MPI_INT,
// MPI_COMM_WORLD);
static long psync[_SHMEM_COLLECT_SYNC_SIZE];
for (i = 0; i < _SHMEM_COLLECT_SYNC_SIZE; i++)
psync[i] = _SHMEM_SYNC_VALUE;
shmem_barrier_all ();
int other, j1;
shmem_fcollect32 (rcounts_full, scounts, 4, 0, 0, numnodes, psync);
for (i = 0; i < numnodes; i++) {
rcounts[i] = rcounts_full[i * numnodes + myid];
}
printf ("-----myid= %d rcounts=", myid);
for (i = 0; i < numnodes; i++)
printf ("%d ", rcounts[i]);
printf ("\n");
/* write(*,*)"myid= ",myid," rcounts= ",rcounts */
/* calculate displacements and the size of the arrays */
sdisp[0] = 0;
for (i = 1; i < numnodes; i++) {
sdisp[i] = scounts[i - 1] + sdisp[i - 1];
}
rdisp[0] = 0;
for (i = 1; i < numnodes; i++) {
rdisp[i] = rcounts[i - 1] + rdisp[i - 1];
}
ssize = 0;
rsize = 0;
for (i = 0; i < numnodes; i++) {
ssize = ssize + scounts[i];
rsize = rsize + rcounts[i];
}
/* allocate send and rec arrays */
sray = (int *) shmem_malloc (sizeof (int) * 20);
rray = (int *) shmem_malloc (sizeof (int) * 20);
for (i = 0; i < ssize; i++) {
sray[i] = myid;
}
/* send/rec different amounts of data to/from each processor */
// mpi_err = MPI_Alltoallv(sray,scounts,sdisp,MPI_INT,
// rray,rcounts,rdisp,MPI_INT, MPI_COMM_WORLD);
shmem_barrier_all ();
for (j1 = 0; j1 < numnodes; j1++) {
int k1 = sdisp[j1];
static int k2;
shmem_int_get (&k2, &rdisp[myid], 1, j1);
shmem_int_put (rray + k2, sray + k1, scounts[j1], j1);
}
shmem_barrier_all ();
// not possible, coz even though the rcounts[myid] will be different on
// each PE, the elements collected
// by PE0 from other PE's will be constant.
// shmem_collect32(rray_full,sray,rcounts[myid],0,0,numnodes,psync);
printf ("myid= %d rray=", myid);
for (i = 0; i < rsize; i++)
printf ("%d ", rray[i]);
printf ("\n");
// mpi_err = MPI_Finalize();
shmem_finalize ();
return 0;
}