int main(int argc, char **argv)
{
int i,j;
long modj,oldj,oldxmodj;
int my_pe,n_pes;
size_t max_elements,max_elements_bytes;
static long *x;
shmem_init();
my_pe = shmem_my_pe();
n_pes = shmem_n_pes();
#ifdef HAVE_SET_CACHE_INV
shmem_set_cache_inv();
#endif
/* fail if trying to use only one processor */
if ( n_pes <= 1 ){
fprintf(stderr, "FAIL - test requires at least two PEs\n");
exit(1);
}
if(my_pe == 0)
fprintf(stderr, "shmem_long_finc(%s) n_pes=%d\n", argv[0],n_pes);
/* shmalloc x on all pes (only use the one on PE 0) */
max_elements_bytes = (size_t) (sizeof(long) * n_pes);
x = shmem_malloc( max_elements_bytes );
for(i=0; i<n_pes; i++)
x[i] = 0;
count = 0;
shmem_barrier_all();
for(i=0; i<ITER; i++) {
if (my_pe != 0) {
oldj = shmem_long_finc(&count, 0); /* get index oldj from PE 0 */
modj = (oldj % (n_pes-1)); /* PE 0 is just the counter/checker */
/* increment value in x[modj] */
oldxmodj = shmem_long_finc(&x[modj], 0);
/* printf("PE=%d,oldj=%ld,modj=%ld,oldxmodj=%ld\n",my_pe,oldj,modj,oldxmodj); */
}
}
shmem_barrier_all();
if (my_pe == 0) { /* check x[j] array on PE 0 */
for(j=1 ; j<n_pes; j++) {
if (x[j-1] != ITER)
fprintf(stderr, "FAIL PE %d of %d: x[%d] = %ld expected = %ld\n",
my_pe, n_pes, j-1, x[j-1], ITER);
}
}
shmem_barrier_all();
#ifdef NEEDS_FINALIZE
shmem_finalize();
#endif
return 0;
}
long FORTRANIFY (shmem_int8_finc) (long *target, int *pe)
{
return shmem_long_finc (target, *pe);
}
int main(int argc, char **argv)
{
int i,j;
short modjs, oldjs, oldxmodjs, oldxas, my_pes, vals;
int modji, oldji, oldxmodji, oldxai, my_pei, vali;
long modjl, oldjl, oldxmodjl, oldxal, my_pel, vall;
long long modjll,oldjll,oldxmodjll,oldxall,my_pell,valll;
int my_pe,n_pes;
size_t max_elements,max_elements_bytes;
static short *xs,*xas;
static int *xi,*xai;
static long *xl,*xal;
static long long *xll,*xall;
shmem_init();
my_pe = shmem_my_pe();
n_pes = shmem_n_pes();
my_pes = (short) my_pe;
my_pei = (int) my_pe;
my_pel = (long) my_pe;
my_pell = (long long) my_pe;
vals = 1; vali = 1; vall = 1; valll = 1;
#ifdef HAVE_SET_CACHE_INV
shmem_set_cache_inv();
#endif
/* fail if trying to use only one processor */
if ( n_pes <= 1 ){
fprintf(stderr, "FAIL - test requires at least two PEs\n");
exit(1);
}
if(my_pe == 0)
fprintf(stderr, "shmem_cswap(%s) n_pes=%d\n", argv[0],n_pes);
/* test shmem_short_finc & shmem_short_swap & shmem_short_cswap */
/* shmalloc xs & xas on all pes (only use the ones on PE 0) */
max_elements_bytes = (size_t) (sizeof(short) * n_pes);
xs = shmem_malloc( max_elements_bytes );
for(i=0; i<n_pes; i++)
xs[i] = 0;
max_elements_bytes = (size_t) (sizeof(short) * n_pes * ITER);
xas = shmem_malloc( max_elements_bytes );
for(i=0; i<n_pes*ITER; i++)
if (((i/(n_pes-1)) % 2) == 0) {
xas[i] = 1;
} else {
xas[i] = 0;
}
count_short = 0;
shmem_barrier_all();
for(i=0; i<ITER; i++) {
if (i == ITER-1) shmem_barrier_all(); /* all PEs participate last time */
if (my_pe != 0) {
oldjs = shmem_short_finc(&count_short, 0); /* get index oldjs from PE 0 */
modjs = (oldjs % (n_pes-1)); /* PE 0 is just the counter/checker */
/* conditionally record PE value in xas[oldjs] --
tells PE involved for each count */
oldxas = shmem_short_cswap(&xas[oldjs], vals, my_pes, 0);
/* printf("PE=%d,i=%d,oldjs=%d,oldxas=%d\n",my_pe,i,oldjs,oldxas); */
if (oldxas == 1) {
/* record PE value in xs[modjs] */
oldxmodjs = shmem_short_swap(&xs[modjs], my_pes, 0);
/* printf("PE=%d,oldjs=%d,modjs=%d,oldxmodjs=%d\n",
my_pe,oldjs,modjs,oldxmodjs); */
}
if (oldxas != 0 && oldxas != 1)
fprintf(stderr, "FAIL PE %d of %d: i=%d, oldxas = %d expected = 0\n",
my_pe, n_pes, i, oldxas);
}
}
shmem_barrier_all();
if (my_pe == 0) { /* check last xs[j] array PEs vs saved ans in xas[i] */
i = (ITER-2)*(n_pes-1);
for(j=1 ; j<n_pes; j++) {
/* printf("j=%d,xs[%d]=%d,xas[%d]=%d\n",j,j-1,xs[j-1],i,xas[i]); */
if (xs[j-1] != xas[i])
fprintf(stderr, "FAIL PE %d of %d: xs[%d] = %d expected = %d\n",
my_pe, n_pes, j-1, xs[j-1], xas[i]);
i++;
}
}
shmem_free(xs); shmem_free(xas);
/* test shmem_int_finc & shmem_int_swap & shmem_int_cswap */
/* shmalloc xi & xai on all pes (only use the ones on PE 0) */
max_elements_bytes = (size_t) (sizeof(int) * n_pes);
xi = shmem_malloc( max_elements_bytes );
for(i=0; i<n_pes; i++)
xi[i] = 0;
max_elements_bytes = (size_t) (sizeof(int) * n_pes * ITER);
xai = shmem_malloc( max_elements_bytes );
for(i=0; i<n_pes*ITER; i++)
if (((i/(n_pes-1)) % 2) == 0) {
xai[i] = 1;
} else {
xai[i] = 0;
}
count_int = 0;
shmem_barrier_all();
for(i=0; i<ITER; i++) {
if (i == ITER-1) shmem_barrier_all(); /* all PEs participate last time */
if (my_pe != 0) {
oldji = shmem_int_finc(&count_int, 0); /* get index oldji from PE 0 */
modji = (oldji % (n_pes-1)); /* PE 0 is just the counter/checker */
/* conditionally record PE value in xai[oldji] --
tells PE involved for each count */
oldxai = shmem_int_cswap(&xai[oldji], vali, my_pei, 0);
/* printf("PE=%d,i=%d,oldji=%d,oldxai=%d\n",my_pe,i,oldji,oldxai); */
if (oldxai == 1) {
/* record PE value in xi[modji] */
oldxmodji = shmem_int_swap(&xi[modji], my_pei, 0);
/* printf("PE=%d,oldji=%d,modji=%d,oldxmodji=%d\n",
my_pe,oldji,modji,oldxmodji); */
}
if (oldxai != 0 && oldxai != 1)
fprintf(stderr, "FAIL PE %d of %d: i=%d, oldxai = %d expected = 0\n",
my_pe, n_pes, i, oldxai);
}
}
shmem_barrier_all();
if (my_pe == 0) { /* check last xi[j] array PEs vs saved ans in xai[i] */
i = (ITER-2)*(n_pes-1);
for(j=1 ; j<n_pes; j++) {
/* printf("j=%d,xi[%d]=%d,xai[%d]=%d\n",j,j-1,xi[j-1],i,xai[i]); */
if (xi[j-1] != xai[i])
fprintf(stderr, "FAIL PE %d of %d: xi[%d] = %d expected = %d\n",
my_pe, n_pes, j-1, xi[j-1], xai[i]);
i++;
}
}
shmem_free(xi); shmem_free(xai);
/* test shmem_long_finc & shmem_long_swap & shmem_long_cswap */
/* shmalloc xl & xal on all pes (only use the ones on PE 0) */
max_elements_bytes = (size_t) (sizeof(long) * n_pes);
xl = shmem_malloc( max_elements_bytes );
for(i=0; i<n_pes; i++)
xl[i] = 0;
max_elements_bytes = (size_t) (sizeof(long) * n_pes * ITER);
xal = shmem_malloc( max_elements_bytes );
for(i=0; i<n_pes*ITER; i++)
if (((i/(n_pes-1)) % 2) == 0) {
xal[i] = 1;
} else {
xal[i] = 0;
}
count_long = 0;
shmem_barrier_all();
for(i=0; i<ITER; i++) {
if (i == ITER-1) shmem_barrier_all(); /* all PEs participate last time */
if (my_pe != 0) {
oldjl = shmem_long_finc(&count_long, 0); /* get index oldjl from PE 0 */
modjl = (oldjl % (n_pes-1)); /* PE 0 is just the counter/checker */
/* conditionally record PE value in xal[oldjl] --
tells PE involved for each count */
oldxal = shmem_long_cswap(&xal[oldjl], vall, my_pel, 0);
/* printf("PE=%d,i=%d,oldjl=%d,oldxal=%d\n",my_pe,i,oldjl,oldxal); */
if (oldxal == 1) {
/* record PE value in xl[modjl] */
oldxmodjl = shmem_long_swap(&xl[modjl], my_pel, 0);
/* printf("PE=%d,oldjl=%ld,modjl=%ld,oldxmodjl=%ld\n",
my_pe,oldjl,modjl,oldxmodjl); */
}
if (oldxal != 0 && oldxal != 1)
fprintf(stderr, "FAIL PE %d of %d: i=%d, oldxal = %ld expected = 0\n",
my_pe, n_pes, i, oldxal);
}
}
shmem_barrier_all();
if (my_pe == 0) { /* check last xl[j] array PEs vs saved ans in xal[i] */
i = (ITER-2)*(n_pes-1);
for(j=1 ; j<n_pes; j++) {
/* printf("j=%d,xl[%d]=%ld,xal[%d]=%ld\n",j,j-1,xl[j-1],i,xal[i]); */
if (xl[j-1] != xal[i])
fprintf(stderr, "FAIL PE %d of %d: xl[%d] = %ld expected = %ld\n",
my_pe, n_pes, j-1, xl[j-1], xal[i]);
i++;
}
}
shmem_free(xl); shmem_free(xal);
/* test shmem_longlong_finc & shmem_longlong_swap & shmem_longlong_cswap */
#ifdef HAVE_LONG_LONG
/* shmalloc xll & xall on all pes (only use the ones on PE 0) */
max_elements_bytes = (size_t) (sizeof(long long) * n_pes);
xll = shmem_malloc( max_elements_bytes );
for(i=0; i<n_pes; i++)
xll[i] = 0;
max_elements_bytes = (size_t) (sizeof(long long) * n_pes * ITER);
xall = shmem_malloc( max_elements_bytes );
for(i=0; i<n_pes*ITER; i++)
if (((i/(n_pes-1)) % 2) == 0) {
xall[i] = 1;
} else {
xall[i] = 0;
}
count_longlong = 0;
shmem_barrier_all();
for(i=0; i<ITER; i++) {
if (i == ITER-1) shmem_barrier_all(); /* all PEs participate last time */
if (my_pe != 0) {
oldjll = shmem_longlong_finc(&count_longlong, 0); /* get index oldjll from PE 0 */
modjll = (oldjll % (n_pes-1)); /* PE 0 is just the counter/checker */
/* conditionally record PE value in xall[oldjll] --
tells PE involved for each count */
oldxall = shmem_longlong_cswap(&xall[oldjll], valll, my_pell, 0);
/* printf("PE=%d,i=%d,oldjll=%d,oldxall=%d\n",my_pe,i,oldjll,oldxall); */
if (oldxall == 1) {
/* record PE value in xll[modjll] */
oldxmodjll = shmem_longlong_swap(&xll[modjll], my_pell, 0);
/* printf("PE=%d,oldjll=%ld,modjll=%ld,oldxmodjll=%ld\n",
my_pe,oldjll,modjll,oldxmodjll); */
}
if (oldxall != 0 && oldxall != 1)
fprintf(stderr, "FAIL PE %d of %d: i=%d, oldxall = %ld expected = 0\n",
my_pe, n_pes, i, oldxall);
}
}
shmem_barrier_all();
if (my_pe == 0) { /* check last xll[j] array PEs vs saved ans in xall[i] */
i = (ITER-2)*(n_pes-1);
for(j=1 ; j<n_pes; j++) {
/* printf("j=%d,xll[%d]=%ld,xall[%d]=%ld\n",j,j-1,xll[j-1],i,xall[i]); */
if (xll[j-1] != xall[i])
fprintf(stderr, "FAIL PE %d of %d: xll[%d] = %ld expected = %ld\n",
my_pe, n_pes, j-1, xll[j-1], xall[i]);
i++;
}
}
shmem_free(xll); shmem_free(xall);
#endif
#ifdef SHMEM_C_GENERIC_32
/* test shmem_finc & shmem_swap & shmem_cswap (GENERIC 32) */
/* shmalloc xi & xai on all pes (only use the ones on PE 0) */
max_elements_bytes = (size_t) (sizeof(int) * n_pes);
xi = shmem_malloc( max_elements_bytes );
for(i=0; i<n_pes; i++)
xi[i] = 0;
max_elements_bytes = (size_t) (sizeof(int) * n_pes * ITER);
xai = shmem_malloc( max_elements_bytes );
for(i=0; i<n_pes*ITER; i++)
if (((i/(n_pes-1)) % 2) == 0) {
xai[i] = 1;
} else {
xai[i] = 0;
}
count_int = 0;
shmem_barrier_all();
for(i=0; i<ITER; i++) {
if (i == ITER-1) shmem_barrier_all(); /* all PEs participate last time */
if (my_pe != 0) {
oldji = shmem_finc(&count_int, 0); /* get index oldji from PE 0 */
modji = (oldji % (n_pes-1)); /* PE 0 is just the counter/checker */
/* conditionally record PE value in xai[oldji] --
tells PE involved for each count */
oldxai = shmem_cswap(&xai[oldji], vali, my_pei, 0);
/* printf("PE=%d,i=%d,oldji=%d,oldxai=%d\n",my_pe,i,oldji,oldxai); */
if (oldxai == 1) {
/* record PE value in xi[modji] */
oldxmodji = shmem_swap(&xi[modji], my_pei, 0);
/* printf("PE=%d,oldji=%d,modji=%d,oldxmodji=%d\n",
my_pe,oldji,modji,oldxmodji); */
}
if (oldxai != 0 && oldxai != 1)
fprintf(stderr, "FAIL pe %d of %d: i=%d, oldxai = %d expected = 0\n",
my_pe, n_pes, i, oldxai);
}
}
shmem_barrier_all();
if (my_pe == 0) { /* check last xi[j] array PEs vs saved ans in xai[i] */
i = (ITER-2)*(n_pes-1);
for(j=1 ; j<n_pes; j++) {
/* printf("j=%d,xi[%d]=%d,xai[%d]=%d\n",j,j-1,xi[j-1],i,xai[i]); */
if (xi[j-1] != xai[i])
fprintf(stderr, "FAIL pe %d of %d: xi[%d] = %d expected = %d\n",
my_pe, n_pes, j-1, xi[j-1], xai[i]);
i++;
}
}
shmem_free(xi); shmem_free(xai);
#else
/* test shmem_finc & shmem_swap & shmem_cswap (GENERIC 64) */
/* shmalloc xl & xal on all pes (only use the ones on PE 0) */
max_elements_bytes = (size_t) (sizeof(long) * n_pes);
xl = shmem_malloc( max_elements_bytes );
for(i=0; i<n_pes; i++)
xl[i] = 0;
max_elements_bytes = (size_t) (sizeof(long) * n_pes * ITER);
xal = shmem_malloc( max_elements_bytes );
for(i=0; i<n_pes*ITER; i++)
if (((i/(n_pes-1)) % 2) == 0) {
xal[i] = 1;
} else {
xal[i] = 0;
}
count_long = 0;
shmem_barrier_all();
for(i=0; i<ITER; i++) {
if (i == ITER-1) shmem_barrier_all(); /* all PEs participate last time */
if (my_pe != 0) {
oldjl = shmem_finc(&count_long, 0); /* get index oldjl from PE 0 */
modjl = (oldjl % (n_pes-1)); /* PE 0 is just the counter/checker */
/* conditionally record PE value in xal[oldjl] --
tells PE involved for each count */
oldxal = shmem_cswap(&xal[oldjl], vall, my_pell, 0);
/* printf("PE=%d,i=%d,oldjl=%d,oldxal=%d\n",my_pe,i,oldjl,oldxal); */
if (oldxal == 1) {
/* record PE value in xl[modjl] */
oldxmodjl = shmem_swap(&xl[modjl], my_pell, 0);
/* printf("PE=%d,oldjl=%ld,modjl=%ld,oldxmodjl=%ld\n",
my_pe,oldjl,modjl,oldxmodjl); */
}
if (oldxal != 0 && oldxal != 1)
fprintf(stderr, "FAIL pe %d of %d: i=%d, oldxal = %ld expected = 0\n",
my_pe, n_pes, i, oldxal);
}
}
shmem_barrier_all();
if (my_pe == 0) { /* check last xl[j] array PEs vs saved ans in xal[i] */
i = (ITER-2)*(n_pes-1);
for(j=1 ; j<n_pes; j++) {
/* printf("j=%d,xl[%d]=%ld,xal[%d]=%ld\n",j,j-1,xl[j-1],i,xal[i]); */
if (xl[j-1] != xal[i])
fprintf(stderr, "FAIL pe %d of %d: xl[%d] = %ld expected = %10.2f\n",
my_pe, n_pes, j-1, xl[j-1], xal[i]);
i++;
}
}
shmem_free(xl); shmem_free(xal);
#endif
shmem_barrier_all();
#ifdef NEEDS_FINALIZE
shmem_finalize();
#endif
return 0;
}
int
main(int argc, char* argv[])
{
int me, num_pes, l, pe;
int Verbose = 0;
start_pes(0);
me = _my_pe();
num_pes = _num_pes();
for (l = 0 ; l < loops ; ++l) {
if ((src_int = shmalloc(sizeof(int))) == NULL) {
printf("PE-%d int shmalloc() failed?\n", me);
exit(1);
}
*src_int = 4;
dst_int = itmp = 0;
if ((src_long = shmalloc(sizeof(long))) == NULL) {
printf("PE-%d long shmalloc() failed?\n", me);
exit(1);
}
*src_long = 8;
dst_long = ltmp = 0;
if ((src_llong = shmalloc(sizeof(long long))) == NULL) {
printf("PE-%d long long shmalloc() failed?\n", me);
exit(1);
}
*src_llong = 16;
dst_llong = lltmp = 0;
//printf("PE-%d malloc()s done.\n",me);
shmem_barrier_all();
if ( me == 0 ) {
/* integer swap */
itmp = shmem_int_g(src_int,1);
Vprintf("PE-0 Initial Conditions(int) local %d rem(%d)\n",
dst_int,itmp);
dst_int = shmem_int_cswap(src_int,*src_int,0,1);
if (dst_int != 4) {
printf("PE-%d dst_int %d != 4?\n",me,dst_int);
exit(1);
}
/* verify remote data */
itmp = shmem_int_g(src_int,1);
if (itmp != 0) {
printf("PE-%d rem %d != 0?\n",me,itmp);
exit(1);
}
Vprintf("PE-0 1st int_cswap done: local %d rem(%d)\n",dst_int,itmp);
dst_int = shmem_int_cswap(src_int,0,dst_int,1);
if (dst_int != 0) {
printf("PE-%d dst_int %d != 0?\n",me,dst_int);
exit(1);
}
/* verify remote data */
itmp = shmem_int_g(src_int,1);
if (itmp != 4) {
printf("PE-%d rem %d != 4?\n",me,itmp);
exit(1);
}
Vprintf("PE-0 2nd int_swap done: local %d rem(%d)\n",dst_int,itmp);
/* cswap() should not swap as cond(0) != remote(4) */
dst_int = shmem_int_cswap(src_int,0,0,1);
if (dst_int != 4) {
printf("PE-%d int no-swap returned dst_int %d != 4?\n",
me,dst_int);
exit(1);
}
/* verify previous cswap() did not swap */
itmp = shmem_int_g(src_int,1);
if (itmp != 4) {
printf("PE-%d failed cond int_cswap() swapped? rem(%d) != 4?\n",
me,itmp);
exit(1);
}
/* long swap */
ltmp = shmem_long_g(src_long,1);
Vprintf("PE-0 Initial Conditions(long) local %ld rem(%ld)\n",
dst_long,ltmp);
dst_long = shmem_long_cswap(src_long,*src_long,0,1);
if (dst_long != 8) {
printf("PE-%d dst_long %ld != 8?\n",me,dst_long);
exit(1);
}
/* verify remote data */
ltmp = shmem_long_g(src_long,1);
if (ltmp != 0) {
printf("PE-%d long rem(%ld) != 0?\n",me,ltmp);
exit(1);
}
Vprintf("PE-0 1st long_cswap done: local %ld rem(%ld)\n",
dst_long,ltmp);
dst_long = shmem_long_cswap(src_long,0,dst_long,1);
if (dst_long != 0) {
printf("PE-%d dst_long %ld != 0?\n",me,dst_long);
exit(1);
}
/* verify remote data */
ltmp = shmem_long_g(src_long,1);
if (ltmp != 8) {
printf("PE-%d long rem(%ld) != 8?\n",me,ltmp);
exit(1);
}
Vprintf("PE-0 2nd long_swap done: local %ld rem(%ld)\n",
dst_long,ltmp);
/* cswap() should not swap as cond(0) != remote(8) */
dst_long = shmem_long_cswap(src_long,0,0,1);
if (dst_long != 8) {
printf("PE-%d long no-swap returned dst_long %ld != 8?\n",
me,dst_long);
exit(1);
}
/* verify previous cswap() did not swap */
ltmp = shmem_long_g(src_long,1);
if (ltmp != 8) {
printf("PE-%d failed cond long_cswap() swapped? rem(%ld) != 8?\n",
me,ltmp);
exit(1);
}
/* long long swap */
lltmp = shmem_longlong_g(src_llong,1);
Vprintf("PE-0 Initial Conditions(long long) local %lld rem(%lld)\n",
dst_llong,lltmp);
dst_llong = shmem_longlong_cswap(src_llong,*src_llong,0,1);
if (dst_llong != 16) {
printf("PE-%d dst_llong %lld != 16?\n",me,dst_llong);
exit(1);
}
/* verify remote data */
lltmp = shmem_longlong_g(src_llong,1);
if (lltmp != 0) {
printf("PE-%d longlong rem(%lld) != 0?\n",me,lltmp);
exit(1);
}
Vprintf("PE-0 1st longlong_cswap done: local %lld rem(%lld)\n",
dst_llong, lltmp);
dst_llong = shmem_longlong_cswap(src_llong,0,dst_llong,1);
if (dst_llong != 0) {
printf("PE-%d dst_llong %lld != 0?\n",me,dst_llong);
exit(1);
}
/* verify remote data */
lltmp = shmem_longlong_g(src_llong,1);
if (lltmp != 16) {
printf("PE-%d long long rem(%lld) != 16?\n",me,lltmp);
exit(1);
}
Vprintf("PE-0 2nd longlong_swap done: local %lld rem(%lld)\n",
dst_llong,lltmp);
/* cswap() should not swap as cond(0) != remote(8) */
dst_llong = shmem_longlong_cswap(src_llong,0,0,1);
if (dst_llong != 16) {
printf("PE-%d longlong no-swap returned dst_llong %lld != 16?\n",
me,dst_llong);
exit(1);
}
/* verify previous cswap() did not swap */
lltmp = shmem_longlong_g(src_llong,1);
if (lltmp != 16) {
printf("PE-0 failed cond longlong_cswap() swapped? rem(%lld) != 16?\n",
lltmp);
exit(1);
}
}
else {
if (!shmem_addr_accessible(src_int,0)) {
printf("PE-%d local src_int %p not accessible from PE-%d?\n",
me, (void*)src_int, 0);
exit(1);
}
if (!shmem_addr_accessible(src_long,0)) {
printf("PE-%d local src_long %p not accessible from PE-%d?\n",
me, (void*)src_long, 0);
exit(1);
}
if (!shmem_addr_accessible(src_llong,0)) {
printf("PE-%d local src_llong %p not accessible from PE-%d?\n",
me, (void*)src_llong, 0);
exit(1);
}
}
shmem_barrier_all();
/* shmem_*fadd() exercise */
if (me == 0) {
itmp = 0;
ltmp = 0;
lltmp = 0;
*src_int = 0;
*src_long = 0;
*src_llong = 0;
}
shmem_barrier_all();
(void)shmem_int_fadd( &itmp, me+1, 0 );
(void)shmem_long_fadd( <mp, me+1, 0 );
(void)shmem_longlong_fadd( &lltmp, me+1, 0 );
shmem_barrier_all();
if (me == 0) {
int tot;
for(pe=0,tot=0; pe < num_pes; pe++)
tot += pe+1;
if ( itmp != tot )
printf("fadd() total %d != expected %d?\n",itmp,tot);
if ( ltmp != (long)tot )
printf("fadd() total %ld != expected %d?\n",ltmp,tot);
if ( lltmp != (long long)tot )
printf("fadd() total %lld != expected %d?\n",lltmp,tot);
}
shmem_barrier_all();
(void)shmem_int_finc(src_int,0);
(void)shmem_long_finc(src_long,0);
(void)shmem_longlong_finc(src_llong,0);
shmem_barrier_all();
if (me == 0) {
int tot = num_pes;
if ( *src_int != tot )
printf("finc() total %d != expected %d?\n",*src_int,tot);
if ( *src_long != (long)tot )
printf("finc() total %ld != expected %d?\n",*src_long,tot);
if ( *src_llong != (long long)tot )
printf("finc() total %lld != expected %d?\n",*src_llong,tot);
}
shmem_barrier_all();
shfree(src_int);
shfree(src_long);
shfree(src_llong);
}
if (Verbose)
fprintf(stderr,"[%d] exit\n",_my_pe());
return 0;
}
int
main ()
{
int me, npes;
int *dest1;
float *dest2;
long *dest3;
double *dest4;
long long *dest5;
int swapped_val1, new_val1;
float swapped_val2, new_val2;
long swapped_val3, new_val3;
double swapped_val4, new_val4;
long long swapped_val5, new_val5;
int success = 1;
int success1_p1;
int success2_p1;
int success3_p1;
int success4_p1;
int success5_p1;
shmem_init ();
me = shmem_my_pe ();
npes = shmem_n_pes ();
shmem_barrier_all ();
/* Checks if there are atleast 2 executing PEs */
if (npes > 1) {
dest1 = (int *) shmem_malloc (sizeof (*dest1));
dest2 = (float *) shmem_malloc (sizeof (*dest2));
dest3 = (long *) shmem_malloc (sizeof (*dest3));
dest4 = (double *) shmem_malloc (sizeof (*dest4));
dest5 = (long long *) shmem_malloc (sizeof (*dest5));
*dest1 = *dest2 = *dest3 = *dest4 = *dest5 = me;
new_val1 = new_val2 = new_val3 = new_val4 = new_val5 = me;
success1_p1 = success1_p2 = success2_p1 = success2_p2 = success3_p1 =
success3_p2 = success4_p1 = success4_p2 = success5_p1 =
success5_p2 = -1;
shmem_barrier_all ();
swapped_val1 = shmem_int_swap (dest1, new_val1, (me + 1) % npes);
swapped_val2 = shmem_float_swap (dest2, new_val2, (me + 1) % npes);
swapped_val3 = shmem_long_swap (dest3, new_val3, (me + 1) % npes);
swapped_val4 = shmem_double_swap (dest4, new_val4, (me + 1) % npes);
swapped_val5 = shmem_longlong_swap (dest5, new_val5, (me + 1) % npes);
/* To validate the working of swap we need to check the value received
at the PE that initiated the swap as well as the dest PE */
if (me == 0) {
if (swapped_val1 == 1) {
success1_p1 = 1;
}
if (swapped_val2 == 1) {
success2_p1 = 1;
}
if (swapped_val3 == 1) {
success3_p1 = 1;
}
if (swapped_val4 == 1) {
success4_p1 = 1;
}
if (swapped_val5 == 1) {
success5_p1 = 1;
}
}
if (me == 1) {
if (*dest1 == 0) {
shmem_int_put (&success1_p2, &success, 1, 0);
}
if (*dest2 == 0) {
shmem_int_put (&success2_p2, &success, 1, 0);
}
if (*dest3 == 0) {
shmem_int_put (&success3_p2, &success, 1, 0);
}
if (*dest4 == 0) {
shmem_int_put (&success4_p2, &success, 1, 0);
}
if (*dest5 == 0) {
shmem_int_put (&success5_p2, &success, 1, 0);
}
}
shmem_barrier_all ();
if (me == 0) {
if (success1_p1 && success1_p2) {
printf ("Test shmem_int_swap: Passed\n");
}
else {
printf ("Test shmem_int_swap: Failed\n");
}
if (success2_p1 && success2_p2) {
printf ("Test shmem_float_swap: Passed\n");
}
else {
printf ("Test shmem_float_swap: Failed\n");
}
if (success3_p1 && success3_p2) {
printf ("Test shmem_long_swap: Passed\n");
}
else {
printf ("Test shmem_long_swap: Failed\n");
}
if (success4_p1 && success4_p2) {
printf ("Test shmem_double_swap: Passed\n");
}
else {
printf ("Test shmem_double_swap: Failed\n");
}
if (success5_p1 && success5_p2) {
printf ("Test shmem_longlong_swap: Passed\n");
}
else {
printf ("Test shmem_longlong_swap: Failed\n");
}
}
shmem_barrier_all ();
/* Test conditional swaps shmem_longlong_cswap, shmem_long_cswap,
shmem_int_cswap, */
*dest1 = *dest3 = *dest5 = me;
new_val1 = new_val3 = new_val5 = me;
success1_p1 = success1_p2 = success3_p1 = success3_p2 = success5_p1 =
success5_p2 = -1;
shmem_barrier_all ();
swapped_val1 = shmem_int_cswap (dest1, me + 1, (long) me, 1);
swapped_val3 = shmem_long_cswap (dest3, me + 1, (long) me, 1);
swapped_val5 = shmem_longlong_cswap (dest5, me + 1, (long) me, 1);
/* To validate the working of conditionalswap we need to check the
value received at the PE that initiated the conditional swap as
well as the dest PE */
if (me == 0) {
if (swapped_val1 == 1) {
success1_p1 = 1;
}
if (swapped_val3 == 1) {
success3_p1 = 1;
}
if (swapped_val5 == 1) {
success5_p1 = 1;
}
}
if (me == 1) {
if (*dest1 == 0) {
shmem_int_put (&success1_p2, &success, 1, 0);
}
if (*dest3 == 0) {
shmem_int_put (&success3_p2, &success, 1, 0);
}
if (*dest5 == 0) {
shmem_int_put (&success5_p2, &success, 1, 0);
}
}
shmem_barrier_all ();
if (me == 0) {
if (success1_p1 && success1_p2) {
printf ("Test shmem_int_cswap: Passed\n");
}
else {
printf ("Test shmem_int_cswap: Failed\n");
}
if (success3_p1 && success3_p2) {
printf ("Test shmem_long_cswap: Passed\n");
}
else {
printf ("Test shmem_long_cswap: Failed\n");
}
if (success5_p1 && success5_p2) {
printf ("Test shmem_longlong_cswap: Passed\n");
}
else {
printf ("Test shmem_longlong_cswap: Failed\n");
}
}
shmem_barrier_all ();
/* Test shmem_long_fadd, shmem_int_fadd, shmem_longlong_fadd */
*dest1 = *dest3 = *dest5 = me;
new_val1 = new_val3 = new_val5 = me;
success1_p1 = success1_p2 = success3_p1 = success3_p2 = success5_p1 =
success5_p2 = -1;
shmem_barrier_all ();
swapped_val1 = shmem_int_fadd (dest1, 1, 0);
swapped_val3 = shmem_long_fadd (dest3, 1, 0);
swapped_val5 = shmem_longlong_fadd (dest5, 1, 0);
/* To validate the working of fetch and add we need to check the old
value received at the PE that initiated the fetch and increment as
well as the new value on the dest PE */
if (me != 0) {
if (swapped_val1 == 0) {
success1_p1 = 1;
}
if (swapped_val3 == 0) {
success3_p1 = 1;
}
if (swapped_val5 == 0) {
success5_p1 = 1;
}
}
if (me == 0) {
if (*dest1 == npes - 1) {
shmem_int_put (&success1_p2, &success, 1, npes - 1);
}
if (*dest3 == npes - 1) {
shmem_int_put (&success3_p2, &success, 1, npes - 1);
}
if (*dest5 == npes - 1) {
shmem_int_put (&success5_p2, &success, 1, npes - 1);
}
}
shmem_barrier_all ();
if (me == npes - 1) {
if (success1_p1 && success1_p2) {
printf ("Test shmem_int_fadd: Passed\n");
}
else {
printf ("Test shmem_int_fadd: Failed\n");
}
if (success3_p1 && success3_p2) {
printf ("Test shmem_long_fadd: Passed\n");
}
else {
printf ("Test shmem_long_fadd: Failed\n");
}
if (success5_p1 && success5_p2) {
printf ("Test shmem_longlong_fadd: Passed\n");
}
else {
printf ("Test shmem_longlong_fadd: Failed\n");
}
}
shmem_barrier_all ();
/* Test shmem_long_finc, shmem_int_finc, shmem_longlong_finc */
*dest1 = *dest3 = *dest5 = me;
new_val1 = new_val3 = new_val5 = me;
success1_p1 = success1_p2 = success3_p1 = success3_p2 = success5_p1 =
success5_p2 = -1;
shmem_barrier_all ();
swapped_val1 = shmem_int_finc (dest1, 0);
swapped_val3 = shmem_long_finc (dest3, 0);
swapped_val5 = shmem_longlong_finc (dest5, 0);
/* To validate the working of fetch and increment we need to check the
old value received at the PE that initiated the fetch and increment
as well as the new value on the dest PE */
if (me != 0) {
if (swapped_val1 == 0) {
success1_p1 = 1;
}
if (swapped_val3 == 0) {
success3_p1 = 1;
}
if (swapped_val5 == 0) {
success5_p1 = 1;
}
}
if (me == 0) {
if (*dest1 == npes - 1) {
shmem_int_put (&success1_p2, &success, 1, npes - 1);
}
if (*dest3 == npes - 1) {
shmem_int_put (&success3_p2, &success, 1, npes - 1);
}
if (*dest5 == npes - 1) {
shmem_int_put (&success5_p2, &success, 1, npes - 1);
}
}
shmem_barrier_all ();
if (me == npes - 1) {
if (success1_p1 && success1_p2) {
printf ("Test shmem_int_finc: Passed\n");
}
else {
printf ("Test shmem_int_finc: Failed\n");
}
if (success3_p1 && success3_p2) {
printf ("Test shmem_long_finc: Passed\n");
}
else {
printf ("Test shmem_long_finc: Failed\n");
}
if (success5_p1 && success5_p2) {
printf ("Test shmem_longlong_finc: Passed\n");
}
else {
printf ("Test shmem_longlong_finc: Failed\n");
}
}
shmem_barrier_all ();
shmem_free (dest1);
shmem_free (dest2);
shmem_free (dest3);
shmem_free (dest4);
shmem_free (dest5);
}
else {
printf
("Number of PEs must be > 1 to test shmem atomics, test skipped\n");
}
shmem_finalize ();
return 0;
}
int main( int argc, char *argv[])
{
int rc=0, my_pe, npes, neighbor;
int loops=LOOPS;
int j;
size_t data_sz=sizeof(long) * 3;
double start_time;
long *data, lval=0;
if (argc > 1)
loops = atoi(argv[1]);
shmem_init();
my_pe = shmem_my_pe();
npes = shmem_n_pes();
data = shmem_malloc(data_sz);
if (!data) {
fprintf(stderr,"[%d] shmem_malloc(%ld) failure? %d\n",
my_pe,data_sz,errno);
shmem_global_exit(1);
}
memset((void*)data,0,data_sz);
shmem_barrier_all();
neighbor = (my_pe + 1) % npes;
start_time = shmemx_wtime();
for(j=0,elapsed=0.0; j < loops; j++) {
start_time = shmemx_wtime();
lval = shmem_long_finc( (void*)&data[1], neighbor );
elapsed += shmemx_wtime() - start_time;
if (lval != (long) j) {
fprintf(stderr,"[%d] Test: FAIL previous val %ld != %d Exit.\n",
my_pe, lval, j);
shmem_global_exit(1);
}
}
shmem_barrier_all();
rc = 0;
if (data[1] != (long)loops) {
fprintf(stderr,"[%d] finc neighbot: FAIL data[1](%p) %ld != %d Exit.\n",
my_pe, (void*)&data[1], data[1], loops);
rc--;
}
/* check if adjancent memory locations distrubed */
assert(data[0] == 0);
assert(data[2] == 0);
if (my_pe == 0 ) {
if (rc == 0 && Verbose)
fprintf(stderr,"[%d] finc neighbor: PASSED.\n",my_pe);
fprintf(stderr,"[%d] %d loops of shmem_long_finc() in %6.4f secs\n"
" %2.6f usecs per shmem_long_finc()\n",
my_pe,loops,elapsed,((elapsed*100000.0)/(double)loops));
}
shmem_free(data);
shmem_finalize();
return rc;
}
int main(int argc, char **argv)
{
int i,j;
short modjs, oldjs, oldxmodjs, valuejs;
int modji, oldji, oldxmodji, valueji;
long modjl, oldjl, oldxmodjl, valuejl;
long long modjll,oldjll,oldxmodjll, valuejll;
int my_pe,n_pes;
size_t max_elements,max_elements_bytes;
static short *xs;
static int *xi;
static long *xl;
static long long *xll;
shmem_init();
my_pe = shmem_my_pe();
n_pes = shmem_n_pes();
#ifdef HAVE_SET_CACHE_INV
shmem_set_cache_inv();
#endif
/* fail if trying to use only one processor */
if ( n_pes <= 1 ){
fprintf(stderr, "FAIL - test requires at least two PEs\n");
exit(1);
}
if(my_pe == 0)
fprintf(stderr, "shmem_fadd(%s) n_pes=%d\n", argv[0],n_pes);
/* test shmem_short_fadd */
#ifdef HAVE_SHORT
/* shmalloc xs on all pes (only use the one on PE 0) */
max_elements_bytes = (size_t) (sizeof(short) * n_pes);
xs = shmem_malloc( max_elements_bytes );
for(i=0; i<n_pes; i++)
xs[i] = 0;
count_short = 0;
shmem_barrier_all();
for(i=0; i<ITER; i++) {
if (my_pe != 0) {
oldjs = shmem_short_finc(&count_short, 0); /* get index oldjs from PE 0 */
modjs = (oldjs % (n_pes-1)); /* PE 0 is just the counter/checker */
/* add 10 to value in xs[modjs] */
valuejs = (short) 10;
oldxmodjs = shmem_short_fadd(&xs[modjs], valuejs, 0);
/* printf("PE=%d,oldjs=%d,modjs=%d,oldxmodjs=%d,valuejs=%d\n",
my_pe,oldjs,modjs,oldxmodjs,valuejs); */
}
}
shmem_barrier_all();
if (my_pe == 0) { /* check xs[j] array on PE 0 */
for(j=1 ; j<n_pes; j++) {
if (xs[j-1] != 10*ITER)
fprintf(stderr, "FAIL PE %d of %d: xs[%d] = %d expected = %d\n",
my_pe, n_pes, j-1, xs[j-1], ITER);
}
}
shmem_free(xs);
#endif
/* test shmem_int_fadd */
/* shmalloc xi on all pes (only use the one on PE 0) */
max_elements_bytes = (size_t) (sizeof(int) * n_pes);
xi = shmem_malloc( max_elements_bytes );
for(i=0; i<n_pes; i++)
xi[i] = 0;
count_int = 0;
shmem_barrier_all();
for(i=0; i<ITER; i++) {
if (my_pe != 0) {
oldji = shmem_int_finc(&count_int, 0); /* get index oldji from PE 0 */
modji = (oldji % (n_pes-1)); /* PE 0 is just the counter/checker */
/* add 10 to value in xi[modji] */
valueji = (int) 10;
oldxmodji = shmem_int_fadd(&xi[modji], valueji, 0);
/* printf("PE=%d,oldji=%d,modji=%d,oldxmodji=%d,valueji=%d\n",
my_pe,oldji,modji,oldxmodji,valueji); */
}
}
shmem_barrier_all();
if (my_pe == 0) { /* check xi[j] array on PE 0 */
for(j=1 ; j<n_pes; j++) {
if (xi[j-1] != 10*ITER)
fprintf(stderr, "FAIL PE %d of %d: xi[%d] = %d expected = %d\n",
my_pe, n_pes, j-1, xi[j-1], ITER);
}
}
shmem_free(xi);
/* test shmem_long_fadd */
/* shmalloc xl on all pes (only use the one on PE 0) */
max_elements_bytes = (size_t) (sizeof(long) * n_pes);
xl = shmem_malloc( max_elements_bytes );
for(i=0; i<n_pes; i++)
xl[i] = 0;
count_long = 0;
shmem_barrier_all();
for(i=0; i<ITER; i++) {
if (my_pe != 0) {
oldjl = shmem_long_finc(&count_long, 0); /* get index oldjl from PE 0 */
modjl = (oldjl % (n_pes-1)); /* PE 0 is just the counter/checker */
/* add 10 to value in xl[modjl] */
valuejl = (long) 10;
oldxmodjl = shmem_long_fadd(&xl[modjl], valuejl, 0);
/* printf("PE=%d,oldjl=%ld,modjl=%ld,oldxmodjl=%ld,valuejl=%ld\n",
my_pe,oldjl,modjl,oldxmodjl,valuejl); */
}
}
shmem_barrier_all();
if (my_pe == 0) { /* check xl[j] array on PE 0 */
for(j=1 ; j<n_pes; j++) {
if (xl[j-1] != 10*ITER)
fprintf(stderr, "FAIL PE %d of %d: xl[%d] = %ld expected = %ld\n",
my_pe, n_pes, j-1, xl[j-1], ITER);
}
}
shmem_free(xl);
/* test shmem_longlong_fadd */
#ifdef HAVE_LONG_LONG
/* shmalloc xll on all pes (only use the one on PE 0) */
max_elements_bytes = (size_t) (sizeof(long long) * n_pes);
xll = shmem_malloc( max_elements_bytes );
for(i=0; i<n_pes; i++)
xll[i] = 0;
count_longlong = 0;
shmem_barrier_all();
for(i=0; i<ITER; i++) {
if (my_pe != 0) {
oldjll = shmem_longlong_finc(&count_longlong, 0); /* get index oldjll from PE 0 */
modjll = (oldjll % (n_pes-1)); /* PE 0 is just the counter/checker */
/* add 10 to value in xll[modjll] */
valuejll = (long long) 10;
oldxmodjll = shmem_longlong_fadd(&xll[modjll], valuejll, 0);
/* printf("PE=%d,oldjll=%ld,modjll=%ld,oldxmodjll=%ld,valuejll=%ld\n",
my_pe,oldjll,modjll,oldxmodjll,valuejll); */
}
}
shmem_barrier_all();
if (my_pe == 0) { /* check xll[j] array on PE 0 */
for(j=1 ; j<n_pes; j++) {
if (xll[j-1] != 10*ITER)
fprintf(stderr, "FAIL PE %d of %d: xll[%d] = %ld expected = %ld\n",
my_pe, n_pes, j-1, xll[j-1], ITER);
}
}
shmem_free(xll);
#endif
#ifdef SHMEM_C_GENERIC_32
/* test shmem_fadd (GENERIC 32) */
/* shmalloc xi on all pes (only use the one on PE 0) */
max_elements_bytes = (size_t) (sizeof(int) * n_pes);
xi = shmem_malloc( max_elements_bytes );
for(i=0; i<n_pes; i++)
xi[i] = 0;
count_int = 0;
shmem_barrier_all();
for(i=0; i<ITER; i++) {
if (my_pe != 0) {
oldji = shmem_finc(&count_int, 0); /* get index oldji from PE 0 */
modji = (oldji % (n_pes-1)); /* PE 0 is just the counter/checker */
/* add 10 to value in xi[modji] */
valueji = (int) 10;
oldxmodji = shmem_fadd(&xi[modji], valueji, 0);
/* printf("PE=%d,oldji=%d,modji=%d,oldxmodji=%d,valueji=%d\n",
my_pe,oldji,modji,oldxmodji,valueji); */
}
}
shmem_barrier_all();
if (my_pe == 0) { /* check xi[j] array on PE 0 */
for(j=1 ; j<n_pes; j++) {
if (xi[j-1] != 10*ITER)
fprintf(stderr, "FAIL pe %d of %d: xi[%d] = %d expected = %d\n",
my_pe, n_pes, j-1, xi[j-1], ITER);
}
}
shmem_free(xi);
#else
/* test shmem_fadd (GENERIC 64) */
/* shmalloc xl on all pes (only use the one on PE 0) */
max_elements_bytes = (size_t) (sizeof(long) * n_pes);
xl = shmem_malloc( max_elements_bytes );
for(i=0; i<n_pes; i++)
xl[i] = 0;
count_long = 0;
shmem_barrier_all();
for(i=0; i<ITER; i++) {
if (my_pe != 0) {
#if (defined(__STDC_VERSION__) && __STDC_VERSION__ >= 201112L)
oldjl = shmem_finc(&count_long, 0); /* get index oldjl from PE 0 */
#else
oldjl = shmem_long_finc(&count_long, 0); /* get index oldjl from PE 0 */
#endif
modjl = (oldjl % (n_pes-1)); /* PE 0 is just the counter/checker */
/* add 10 to value in xl[modjl] */
valuejl = (long) 10;
#if (defined(__STDC_VERSION__) && __STDC_VERSION__ >= 201112L)
oldxmodjl = shmem_fadd(&xl[modjl], valuejl, 0);
#else
oldxmodjl = shmem_long_fadd(&xl[modjl], valuejl, 0);
#endif
/* printf("PE=%d,oldjl=%ld,modjl=%ld,oldxmodjl=%ld,valuejl=%ld\n",
my_pe,oldjl,modjl,oldxmodjl,valuejl); */
}
}
shmem_barrier_all();
if (my_pe == 0) { /* check xl[j] array on PE 0 */
for(j=1 ; j<n_pes; j++) {
if (xl[j-1] != 10*ITER)
fprintf(stderr, "FAIL pe %d of %d: xl[%d] = %ld expected = %ld\n",
my_pe, n_pes, j-1, xl[j-1], ITER);
}
}
shmem_free(xl);
#endif
shmem_barrier_all();
#ifdef NEEDS_FINALIZE
shmem_finalize();
#endif
return 0;
}
int main(int argc, char **argv)
{
int i,j;
long modj,oldj,oldxmodj,oldxa;
int my_pe,n_pes;
size_t max_elements,max_elements_bytes;
static long *x,*xa;
shmem_init();
my_pe = shmem_my_pe();
n_pes = shmem_n_pes();
#ifdef HAVE_SET_CACHE_INV
shmem_set_cache_inv();
#endif
/* fail if trying to use only one processor */
if ( n_pes <= 1 ){
fprintf(stderr, "FAIL - test requires at least two PEs\n");
exit(1);
}
if(my_pe == 0)
fprintf(stderr, "shmem_long_swap(%s) n_pes=%d\n", argv[0],n_pes);
/* shmalloc x & xa on all pes (only use the ones on PE 0) */
max_elements_bytes = (size_t) (sizeof(long) * n_pes);
x = shmem_malloc( max_elements_bytes );
for(i=0; i<n_pes; i++)
x[i] = 0;
max_elements_bytes = (size_t) (sizeof(long) * n_pes * ITER);
xa = shmem_malloc( max_elements_bytes );
for(i=0; i<n_pes*ITER; i++)
xa[i] = 0;
count = 0;
shmem_barrier_all();
for(i=0; i<ITER; i++) {
if (my_pe != 0) {
oldj = shmem_long_finc(&count, 0); /* get index oldj from PE 0 */
modj = (oldj % (n_pes-1)); /* PE 0 is just the counter/checker */
/* record PE value in x[modj] */
oldxmodj = shmem_long_swap(&x[modj], my_pe, 0);
/* printf("PE=%d,oldj=%ld,modj=%ld,oldxmodj=%ld\n",my_pe,oldj,modj,oldxmodj); */
/* record PE value in xa[oldj] -- tells PE involved for each count */
oldxa = shmem_long_swap(&xa[oldj], my_pe, 0);
/* printf("PE=%d,i=%d,oldj=%ld,oldxa=%ld\n",my_pe,i,oldj,oldxa); */
if (oldxa != 0)
fprintf(stderr, "FAIL PE %d of %d: i=%d, oldxa = %ld expected = 0\n",
my_pe, n_pes, i, oldxa);
}
}
shmem_barrier_all();
if (my_pe == 0) { /* check last x[j] array PEs vs saved ans in xa[i] */
i = (ITER-1)*(n_pes-1);
for(j=1 ; j<n_pes; j++) {
printf("j=%d,x[%d]=%ld,xa[%d]=%ld\n",j,j-1,x[j-1],i,xa[i]);
if (x[j-1] != xa[i])
fprintf(stderr, "FAIL PE %d of %d: x[%d] = %ld expected = %ld\n",
my_pe, n_pes, j-1, x[j-1], ITER);
i++;
}
}
shmem_barrier_all();
#ifdef NEEDS_FINALIZE
shmem_finalize();
#endif
return 0;
}