int
main(int argc, char* argv[])
{
long source[10] = { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10 };
static long target[10] = { 11, 12, 13, 14, 15, 16, 17, 18, 19, 20 };
int i, bad = 0;
shmem_init();
if (shmem_n_pes() == 1) {
printf("%s: Requires number of PEs > 1\n", argv[0]);
shmem_finalize();
return 0;
}
if (shmem_my_pe() == 0) {
for (i = 0 ; i < 10 ; ++i) {
source[i] = shmem_long_swap(target + i, source[i], 1);
}
}
shmem_barrier_all(); /* sync sender and receiver */
if (shmem_my_pe() == 0) {
for (i = 0 ; i < 10 ; ++i) {
if (source[i] != i + 11) bad = 1;
}
} else if (shmem_my_pe() == 1) {
for (i = 0 ; i < 10 ; ++i) {
if (target[i] != i + 1) bad = 1;
}
}
if (bad) {
fprintf(stderr, "[%02d] error: ", shmem_my_pe());
for (i = 0 ; i < 10 ; ++i) {
fprintf(stderr, "%ld ", (shmem_my_pe() == 0) ? source[i] : target[i]);
}
fprintf(stderr, "\n");
}
shmem_finalize();
return bad;
}
long FORTRANIFY (shmem_swap) (long *target, long *value, int *pe)
{
return shmem_long_swap (target, *value, *pe);
}
int main(int argc, char **argv)
{
int i,j;
short oldjs, oldxs, my_pes;
int oldji, oldxi, my_pei;
long oldjl, oldxl, my_pel;
long long oldjll,oldxll,my_pell;
float oldjf, oldxf, my_pef;
double oldjd, oldxd, my_ped;
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;
static float *xf;
static double *xd;
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;
my_pef = (float) my_pe;
my_ped = (double) my_pe;
#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_swap(%s) n_pes=%d\n", argv[0],n_pes);
/* test shmem_short_swap */
/* shmalloc xs on all pes (only check 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;
shmem_barrier_all();
oldjs = 0;
for(i=0; i<ITER; i++) {
if (my_pe != 0) {
my_pes = my_pes + (short) 1;
/* record PE value in xs[my_pe] -- save PE number */
oldxs = shmem_short_swap(&xs[my_pe], my_pes, 0);
/* printf("PE=%d,i=%d,my_pes=%d,oldxs=%d\n",my_pe,i,my_pes,oldxs); */
if (oldxs != oldjs)
fprintf(stderr, "FAIL PE %d of %d: i=%d, oldxs = %d expected = %d\n",
my_pe, n_pes, i, oldxs, oldjs);
oldjs = my_pes;
}
}
shmem_barrier_all();
if (my_pe == 0) { /* check xs[j] array vs PE# + ITER */
i = (int) ITER + 1;
for(j=1 ; j<n_pes; j++) {
/* printf("j=%d,xs[%d]=%d,i=%d\n",j,j,xs[j],i); */
if (xs[j] != (short) i)
fprintf(stderr, "FAIL PE %d of %d: xs[%d] = %d expected = %d\n",
my_pe, n_pes, j, xs[j],i);
i++;
}
}
shmem_free(xs);
/* test shmem_int_swap */
/* shmalloc xi on all pes (only check 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;
shmem_barrier_all();
oldji = 0;
for(i=0; i<ITER; i++) {
if (my_pe != 0) {
my_pei = my_pei + (int) 1;
/* record PE value in xi[my_pe] -- save PE number */
oldxi = shmem_int_swap(&xi[my_pe], my_pei, 0);
/* printf("PE=%d,i=%d,my_pei=%d,oldxi=%d\n",my_pe,i,my_pei,oldxi); */
if (oldxi != oldji)
fprintf(stderr, "FAIL PE %d of %d: i=%d, oldxi = %d expected = %d\n",
my_pe, n_pes, i, oldxi, oldji);
oldji = my_pei;
}
}
shmem_barrier_all();
if (my_pe == 0) { /* check xi[j] array vs PE# + ITER */
i = (int) ITER + 1;
for(j=1 ; j<n_pes; j++) {
/* printf("j=%d,xi[%d]=%d,i=%d\n",j,j,xi[j],i); */
if (xi[j] != i)
fprintf(stderr, "FAIL PE %d of %d: xi[%d] = %d expected = %d\n",
my_pe, n_pes, j, xi[j],i);
i++;
}
}
shmem_free(xi);
/* test shmem_long_swap */
/* shmalloc xl on all pes (only check 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;
shmem_barrier_all();
oldjl = 0;
for(i=0; i<ITER; i++) {
if (my_pe != 0) {
my_pel = my_pel + (long) 1;
/* record PE value in xl[my_pe] -- save PE number */
oldxl = shmem_long_swap(&xl[my_pe], my_pel, 0);
/* printf("PE=%d,i=%d,my_pel=%d,oldxl=%d\n",my_pe,i,my_pel,oldxl); */
if (oldxl != oldjl)
fprintf(stderr, "FAIL PE %d of %d: i=%d, oldxl = %d expected = %d\n",
my_pe, n_pes, i, oldxl, oldjl);
oldjl = my_pel;
}
}
shmem_barrier_all();
if (my_pe == 0) { /* check xl[j] array vs PE# + ITER */
i = (int) ITER + 1;
for(j=1 ; j<n_pes; j++) {
/* printf("j=%d,xl[%d]=%d,i=%d\n",j,j,xl[j],i); */
if (xl[j] != (long)i)
fprintf(stderr, "FAIL PE %d of %d: xl[%d] = %ld expected = %d\n",
my_pe, n_pes, j, xl[j],i);
i++;
}
}
shmem_free(xl);
/* test shmem_longlong_swap */
#ifdef HAVE_LONG_LONG
/* shmalloc xll on all pes (only check 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;
shmem_barrier_all();
oldjll = 0;
for(i=0; i<ITER; i++) {
if (my_pe != 0) {
my_pell = my_pell + (long long) 1;
/* record PE value in xll[my_pe] -- save PE number */
oldxll = shmem_longlong_swap(&xll[my_pe], my_pell, 0);
/* printf("PE=%d,i=%d,my_pell=%ld,oldxll=%d\n",my_pe,i,my_pell,oldxll); */
if (oldxll != (long long) oldjll)
fprintf(stderr, "FAIL PE %d of %d: i=%d, oldxll = %ld expected = %ld\n",
my_pe, n_pes, i, oldxll, oldjll);
oldjll = my_pell;
}
}
shmem_barrier_all();
if (my_pe == 0) { /* check xll[j] array vs PE# + ITER */
i = (int) ITER + 1;
for(j=1 ; j<n_pes; j++) {
/* printf("j=%d,xll[%d]=%ld,i=%d\n",j,j,xll[j],i); */
if (xll[j] != (long long) i)
fprintf(stderr, "FAIL PE %d of %d: xll[%d] = %d expected = %d\n",
my_pe, n_pes, j, xll[j],i);
i++;
}
}
shmem_free(xll);
#endif
/* test shmem_float_swap */
/* shmalloc xf on all pes (only use the ones on PE 0) */
max_elements_bytes = (size_t) (sizeof(float) * n_pes);
xf = shmem_malloc( max_elements_bytes );
for(i=0; i<n_pes; i++)
xf[i] = (float) 0;
shmem_barrier_all();
oldjf = (float) 0;
for(i=0; i<ITER; i++) {
if (my_pe != 0) {
my_pef = my_pef + (float) 1;
/* record PE value in xf[my_pe] -- save PE number */
oldxf = shmem_float_swap(&xf[my_pe], my_pef, 0);
/* printf("PE=%d,i=%d,my_pef=%10.2f,oldxf=%10.2f\n",my_pe,i,my_pef,oldxf); */
if (oldxf != oldjf)
fprintf(stderr, "FAIL PE %d of %d: i=%d, oldxf = %10.2f expected = %10.2f\n",
my_pe, n_pes, i, oldxf, oldjf);
oldjf = my_pef;
}
}
shmem_barrier_all();
if (my_pe == 0) { /* check xs[j] array vs PE# + ITER */
i = (int) ITER + 1;
for(j=1 ; j<n_pes; j++) {
/* printf("j=%d,xf[%d]=%10.2f,i=%d\n",j,j,xf[j],i); */
if (xf[j] != (float) i)
fprintf(stderr, "FAIL PE %d of %d: xf[%d] = %10.2f expected = %10.2f\n",
my_pe, n_pes, j-1, xf[j], (float)i);
i++;
}
}
shmem_free(xf);
/* test shmem_double_swap */
/* shmalloc xd on all pes (only use the ones on PE 0) */
max_elements_bytes = (size_t) (sizeof(double) * n_pes);
xd = shmem_malloc( max_elements_bytes );
for(i=0; i<n_pes; i++)
xd[i] = (double) 0;
shmem_barrier_all();
oldjd = (double) 0;
for(i=0; i<ITER; i++) {
if (my_pe != 0) {
my_ped = my_ped + (double) 1;
/* record PE value in xd[my_pe] -- save PE number */
oldxd = shmem_double_swap(&xd[my_pe], my_ped, 0);
/* printf("PE=%d,i=%d,my_ped=%10.2f,oldxd=%10.2f\n",my_pe,i,my_ped,oldxd);
*/
if (oldxd != oldjd)
fprintf(stderr, "FAIL PE %d of %d: i=%d, oldxd = %10.2f expected = %10.2f\n",
my_pe, n_pes, i, oldxd, oldjd);
oldjd = my_ped;
}
}
shmem_barrier_all();
if (my_pe == 0) { /* check xd[j] array vs PE# + ITER */
i = (int) ITER + 1;
for(j=1 ; j<n_pes; j++) {
/* printf("j=%d,xd[%d]=%10.2f,i=%d\n",j,j,xd[j],i); */
if (xd[j] != (double) i)
fprintf(stderr, "FAIL PE %d of %d: xd[%d] = %10.2f expected = %10.2f\n",
my_pe, n_pes, j, xd[j], (double)i);
i++;
}
}
shmem_free(xd);
#ifdef SHMEM_C_GENERIC_32
/* test shmem_swap (GENERIC 32) */
my_pei = (int) my_pe;
/* shmalloc xi on all pes (only check 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;
shmem_barrier_all();
oldji = 0;
for(i=0; i<ITER; i++) {
if (my_pe != 0) {
my_pei = my_pei + (int) 1;
/* record PE value in xi[my_pe] -- save PE number */
oldxi = shmem_swap(&xi[my_pe], my_pei, 0);
/* printf("PE=%d,i=%d,my_pei=%d,oldxi=%d\n",my_pe,i,my_pei,oldxi); */
if (oldxi != oldji)
fprintf(stderr, "FAIL pe %d of %d: i=%d, oldxi = %d expected = %d\n",
my_pe, n_pes, i, oldxi, oldji);
oldji = my_pei;
}
}
shmem_barrier_all();
if (my_pe == 0) { /* check xi[j] array vs PE# + ITER */
i = (int) ITER + 1;
for(j=1 ; j<n_pes; j++) {
/* printf("j=%d,xi[%d]=%d,i=%d\n",j,j,xi[j],i); */
if (xi[j] != i)
fprintf(stderr, "FAIL pe %d of %d: xi[%d] = %d expected = %d\n",
my_pe, n_pes, j, xi[j],i);
i++;
}
}
shmem_free(xi);
#else
/* test shmem_swap (GENERIC 64) */
my_pel = (long) my_pe;
/* shmalloc xl on all pes (only check 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;
shmem_barrier_all();
oldjl = 0;
for(i=0; i<ITER; i++) {
if (my_pe != 0) {
my_pel = my_pel + (long) 1;
/* record PE value in xl[my_pe] -- save PE number */
oldxl = shmem_swap(&xl[my_pe], my_pel, 0);
/* printf("PE=%d,i=%d,my_pel=%d,oldxl=%d\n",my_pe,i,my_pel,oldxl); */
if (oldxl != oldjl)
fprintf(stderr, "FAIL pe %d of %d: i=%d, oldxl = %d expected = %d\n",
my_pe, n_pes, i, oldxl, oldjl);
oldjl = my_pel;
}
}
shmem_barrier_all();
if (my_pe == 0) { /* check xl[j] array vs PE# + ITER */
i = (int) ITER + 1;
for(j=1 ; j<n_pes; j++) {
/* printf("j=%d,xl[%d]=%d,i=%d\n",j,j,xl[j],i); */
if (xl[j] != (long)i)
fprintf(stderr, "FAIL pe %d of %d: xl[%d] = %ld expected = %d\n",
my_pe, n_pes, j, xl[j],i);
i++;
}
}
shmem_free(xl);
#endif
shmem_barrier_all();
#ifdef NEEDS_FINALIZE
shmem_finalize();
#endif
return 0;
}
int
main(int argc, char* argv[])
{
int me, num_procs, l, j;
int Verbose = 0;
start_pes(0);
me = _my_pe();
num_procs = _num_pes();
if ( num_procs < 2 ) {
if (me ==0)
printf("PE[0] requires 2 or more PEs?\n");
return 1;
}
for (l = 0 ; l < loops ; ++l) {
if ((src_int = shmalloc(2*num_procs*sizeof(int))) == NULL) {
printf("PE-%d int shmalloc() failed?\n", me);
exit(1);
}
dst_int = &src_int[num_procs];
for(j=0; j < num_procs; j++) {
src_int[j] = 4;
dst_int[j] = 0;
}
if ((src_float = shmalloc(2*num_procs*sizeof(float))) == NULL) {
printf("PE-%d float shmalloc() failed?\n", me);
exit(1);
}
dst_float = &src_float[num_procs];
for(j=0; j < num_procs; j++) {
src_float[j] = 4.0;
dst_float[j] = 0.0;
}
if ((src_double = shmalloc(2*num_procs*sizeof(double))) == NULL) {
printf("PE-%d double shmalloc() failed?\n", me);
exit(1);
}
dst_double = &src_double[num_procs];
for(j=0; j < num_procs; j++) {
src_double[j] = 8.0;
dst_double[j] = 0.0;
}
if ((src_long = shmalloc(2*num_procs*sizeof(long))) == NULL) {
printf("PE-%d long shmalloc() failed?\n", me);
exit(1);
}
dst_long = &src_long[num_procs];
for(j=0; j < num_procs; j++) {
src_long[j] = 8;
dst_long[j] = 0;
}
if ((src_llong = shmalloc(2*num_procs*sizeof(long long))) == NULL) {
printf("PE-%d long shmalloc() failed?\n", me);
exit(1);
}
dst_llong = &src_llong[num_procs];
for(j=0; j < num_procs; j++) {
src_llong[j] = 16;
dst_llong[j] = 0;
}
shmem_barrier_all();
if ( me != 0 ) {
/* is 'src_*' accessible from PE0? should be. */
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_float,0)) {
printf("PE-%d local src_float %p not accessible from PE-%d?\n",
me, (void*)src_float, 0);
exit(1);
}
if (!shmem_addr_accessible(src_double,0)) {
printf("PE-%d local src_double %p not accessible from PE-%d?\n",
me, (void*)src_double, 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();
if ( me == 0 ) {
shmem_quiet();
for(j=1; j < num_procs; j++) {
dst_int[j] = shmem_int_swap(src_int+j,0,j);
if (dst_int[j] != 4) {
printf("PE-%d dst_int[%d] %d != 4?\n",me,j,dst_int[j]);
exit(1);
}
}
shmem_barrier_all();
/* verify remote data */
for(j=1; j < num_procs; j++) {
itmp = shmem_int_g(src_int+j,j);
if (itmp != 0) {
printf("PE-0 int PE[%d] rem(%d) != 0?\n",j,itmp);
exit(1);
}
/* swap back */
dst_int[j] = shmem_int_swap(src_int+j,dst_int[j],j);
if (dst_int[j] != 0) {
printf("PE-0 dst_int[%d] %d != 0?\n",j,dst_int[j]);
exit(1);
}
itmp = shmem_int_g(src_int+j,j);
if (itmp != 4) {
printf("PE-0 PE[%d] rem %d != 4?\n",j,itmp);
exit(1);
}
}
for(j=1; j < num_procs; j++) {
dst_float[j] = shmem_float_swap(src_float+j,0.0,j);
if (dst_float[j] != 4.0) {
printf("PE-0 dst_float[%d] %f != 4.0?\n",j,dst_float[j]);
exit(1);
}
/* verify remote data */
ftmp = shmem_float_g(src_float+j,j);
if (ftmp != 0.0) {
printf("PE-0 float rem(%f) != 0.0?\n",ftmp);
exit(1);
}
/* swap back */
dst_float[j] = shmem_float_swap(src_float+j,dst_float[j],j);
if (dst_float[j] != 0.0) {
printf("PE-0 dst_float[%d] %f != 0.0?\n",j,dst_float[j]);
exit(1);
}
ftmp = shmem_float_g(src_float+j,j);
if (ftmp != 4.0) {
printf("PE-%d float rem(%f) != 4.0?\n",me,ftmp);
exit(1);
}
}
for(j=1; j < num_procs; j++) {
dst_double[j] = shmem_double_swap(src_double+j,0.0,j);
if (dst_double[j] != 8.0) {
printf("PE-0 dst_double[%d] %f != 8.0?\n",j,dst_double[j]);
exit(1);
}
/* verify remote data */
dtmp = shmem_double_g(src_double+j,j);
if (dtmp != 0.0) {
printf("PE-0 float rem(%f) != 0.0?\n",dtmp);
exit(1);
}
dst_double[j] = shmem_double_swap(src_double+j,dst_double[j],j);
if (dst_double[j] != 0.0) {
printf("PE-0 dst_double[%d] %f != 0.0?\n",j,dst_double[j]);
exit(1);
}
dtmp = shmem_double_g(src_double+j,j);
if (dtmp != 8.0) {
printf("PE-0 double rem(%f) != 8.0?\n",dtmp);
exit(1);
}
}
for(j=1; j < num_procs; j++) {
dst_long[j] = shmem_long_swap(src_long+j,0,j);
if (dst_long[j] != 8) {
printf("PE-0 dst_long[%d] %ld != 8?\n",j,dst_long[j]);
exit(1);
}
}
shmem_barrier_all();
/* verify remote data */
for(j=1; j < num_procs; j++) {
ltmp = shmem_long_g(src_long+j,j);
if (ltmp != 0) {
printf("PE-0 PE[%d]long rem(%ld) != 0?\n",j,ltmp);
exit(1);
}
/* swap back */
dst_long[j] = shmem_long_swap(src_long+j,dst_long[j],j);
if (dst_long[j] != 0) {
printf("PE-%d dst_long[%d] %ld != 0?\n",me,j,dst_long[j]);
exit(1);
}
ltmp = shmem_long_g(src_long+j,j);
if (ltmp != 8) {
printf("PE-%d long rem(%ld) != 8?\n",me,ltmp);
exit(1);
}
}
for(j=1; j < num_procs; j++) {
dst_llong[j] = shmem_longlong_swap(src_llong+j,0,j);
if (dst_llong[j] != 16) {
printf("PE-%d dst_llong[%d] %lld != 16?\n",me,j,dst_llong[j]);
exit(1);
}
}
shmem_barrier_all();
/* verify remote data */
for(j=1; j < num_procs; j++) {
lltmp = shmem_longlong_g(src_llong+j,j);
if (lltmp != 0) {
printf("PE-%d long long rem(%lld) != 0?\n",me,lltmp);
exit(1);
}
/* swap back */
dst_llong[j] = shmem_longlong_swap(src_llong+j,dst_llong[j],j);
if (dst_llong[j] != 0) {
printf("PE-%d dst_llong[%d] %lld != 0?\n", me,j,dst_llong[j]);
exit(1);
}
lltmp = shmem_longlong_g(src_llong+j,j);
if (lltmp != 16) {
printf("PE-%d longlong rem(%lld) != 16?\n",me,lltmp);
exit(1);
}
}
}
else {
shmem_int_wait_until(&src_int[me],SHMEM_CMP_EQ,0);
shmem_barrier_all();
shmem_long_wait_until(&src_long[me],SHMEM_CMP_EQ,0);
shmem_barrier_all();
shmem_longlong_wait_until(&src_llong[me],SHMEM_CMP_EQ,0);
shmem_barrier_all();
}
shmem_barrier_all();
shfree(src_int);
shfree(src_float);
shfree(src_double);
shfree(src_long);
shfree(src_llong);
}
if (Verbose)
fprintf(stderr,"[%d] exit\n",_my_pe());
return 0;
}
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 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;
}
/**
* untyped variant
*/
long
shmem_swap (long *target, long value, int pe)
{
return shmem_long_swap (target, value, pe);
}
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;
}