int
main(int argc, char* argv[])
{
long source[10] = { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10 };
long *target;
int *flag;
int i, num_pes;
int failed = 0;
shmem_init();
target = (long*) shmem_malloc(sizeof(long) * 10);
flag = (int*) shmem_malloc(sizeof(int));
*flag = 0;
num_pes=shmem_n_pes();
memset(target, 0, sizeof(long)*10);
shmem_barrier_all();
if (shmem_my_pe() == 0) {
for(i = 0; i < num_pes; i++) {
shmem_long_put_nbi(target, source, 10, i);
shmem_fence();
shmem_int_inc(flag, i);
}
}
shmem_int_wait_until(flag, SHMEM_CMP_EQ, 1);
for (i = 0; i < 10; i++) {
if (target[i] != source[i]) {
fprintf(stderr,"[%d] target[%d] = %ld, expected %ld\n",
shmem_my_pe(), i, target[i], source[i]);
failed = 1;
}
}
shmem_free(target);
shmem_free(flag);
shmem_finalize();
return failed;
}
int main(int argc, char **argv)
{
int i,j,iter;
int my_pe,n_pes;
int *flag,*one;
size_t max_elements,max_elements_bytes;
size_t elements[16] = {1,2,4,8,12,16,24,32,64,128,256,512,1024,2048,4096,8192};
int num_elements = 16;
short *srce_short,*targ_short;
int *srce_int,*targ_int;
long *srce_long,*targ_long;
float *srce_float,*targ_float;
double *srce_double,*targ_double;
shmem_init();
my_pe = shmem_my_pe();
n_pes = shmem_n_pes();
flag = shmem_malloc((size_t) sizeof(int));
one = shmem_malloc((size_t) sizeof(int));
*one = 1;
/* fail if trying to use odd number of processors */
if ( (n_pes % 2) != 0 ){
fprintf(stderr, "FAIL - test requires even number of PEs\n");
exit(1);
}
if(my_pe == 0)
fprintf(stderr, "shmem_both_put_nb_size(%s)\n", argv[0]);
/* alloc arrays */
max_elements = (size_t) (MAX_SIZE / sizeof(int));
max_elements_bytes = (size_t) (sizeof(int)*max_elements);
if(my_pe == 0)
fprintf(stderr,"shmem_int_put_nb max_elements = %d\n",max_elements);
srce_int = shmem_malloc(max_elements_bytes);
targ_int = shmem_malloc(max_elements_bytes);
if((srce_int == NULL) || (targ_int == NULL))
shmalloc_error();
max_elements = (size_t) (MAX_SIZE / sizeof(short));
max_elements_bytes = (size_t) (sizeof(short)*max_elements);
if(my_pe == 0)
fprintf(stderr,"shmem_short_put max_elements = %d\n",max_elements);
srce_short = shmem_malloc(max_elements_bytes);
targ_short = shmem_malloc(max_elements_bytes);
if((srce_short == NULL) || (targ_short == NULL))
shmalloc_error();
max_elements = (size_t) (MAX_SIZE / sizeof(long));
max_elements_bytes = (size_t) (sizeof(long)*max_elements);
if(my_pe == 0)
fprintf(stderr,"shmem_long_put_nb max_elements = %d\n",max_elements);
srce_long = shmem_malloc(max_elements_bytes);
targ_long = shmem_malloc(max_elements_bytes);
if((srce_long == NULL) || (targ_long == NULL))
shmalloc_error();
max_elements = (size_t) (MAX_SIZE / sizeof(float));
max_elements_bytes = (size_t) (sizeof(float)*max_elements);
if(my_pe == 0)
fprintf(stderr,"shmem_float_put_nb max_elements = %d\n",max_elements);
srce_float = shmem_malloc(max_elements_bytes);
targ_float = shmem_malloc(max_elements_bytes);
if((srce_float == NULL) || (targ_float == NULL))
shmalloc_error();
max_elements = (size_t) (MAX_SIZE / sizeof(double));
max_elements_bytes = (size_t) (sizeof(double)*max_elements);
if(my_pe == 0)
fprintf(stderr,"shmem_double_put_nb max_elements = %d\n",max_elements);
srce_double = shmem_malloc(max_elements_bytes);
targ_double = shmem_malloc(max_elements_bytes);
if((srce_double == NULL) || (targ_double == NULL))
shmalloc_error();
if(my_pe == 0)
fprintf(stderr,"Actual value used for max_elements = %d\n",max_elements);
/* try the different sizes MAX_ITER times */
for (iter = 0; iter < MAX_ITER; iter++) {
for (i = 0; i < num_elements; i++) {
*flag = 0;
if (elements[i] <= max_elements) {
if ( (my_pe % 2) == 0 )
for(j = 0; j < elements[i]; j++) {
srce_short[j] = (short)(my_pe+j);
srce_int[j] = (int)(iter*10000+elements[i]*100+my_pe+j);
srce_long[j] = (long)(iter*10000+elements[i]*100+my_pe+j);
srce_float[j] = (float)(iter*10000+elements[i]*100+my_pe+j);
srce_double[j] = (double)(iter*10000+elements[i]*100+my_pe+j);
}
else
for(j = 0; j < elements[i]; j++) {
targ_short[j] = (short)(my_pe+j);
targ_int[j] = (int)(iter*10000+elements[i]*100+my_pe+j);
targ_long[j] = (long)(iter*10000+elements[i]*100+my_pe+j);
targ_float[j] = (float)(iter*10000+elements[i]*100+my_pe+j);
targ_double[j] = (double)(iter*10000+elements[i]*100+my_pe+j);
}
shmem_barrier_all();
if ( (my_pe % 2) == 0 ) {
#ifndef OPENSHMEM
shmemx_int_put_nb(targ_int,srce_int,elements[i],my_pe+1,NULL);
shmemx_long_put_nb(targ_long,srce_long,elements[i],my_pe+1,NULL);
shmemx_float_put_nb(targ_float,srce_float,elements[i],my_pe+1,NULL);
shmemx_double_put_nb(targ_double,srce_double,elements[i],my_pe+1,NULL);
#else
shmem_int_put_nbi(targ_int,srce_int,elements[i],my_pe+1);
shmem_long_put_nbi(targ_long,srce_long,elements[i],my_pe+1);
shmem_float_put_nbi(targ_float,srce_float,elements[i],my_pe+1);
shmem_double_put_nbi(targ_double,srce_double,elements[i],my_pe+1);
#endif
/* this one is blocking */
shmem_short_put(targ_short,srce_short,elements[i],my_pe+1);
shmem_quiet();
shmem_int_put(flag,one,(size_t)1,my_pe+1);
} else {
shmem_int_wait(flag,0);
for(j = 0; j < elements[i]; j++) {
if ( targ_short[j] != (short)(my_pe+j-1) )
fprintf(stderr,
"FAIL: PE [%d] iter=%d i=%d targ_short[%d]=%d not equal %d\n",
my_pe,iter,i,j,targ_short[j],my_pe+j-1);
if ( targ_int[j] != (int)(iter*10000+elements[i]*100+my_pe+j-1) )
fprintf(stderr,
"FAIL: PE [%d] iter=%d i=%d targ_int[%d]=%d not equal %d\n",
my_pe,iter,i,j,targ_int[j],iter*10000+elements[i]*100+my_pe+j-1);
if ( targ_long[j] != (long)(iter*10000+elements[i]*100+my_pe+j-1) )
fprintf(stderr,
"FAIL: PE [%d] iter=%d i=%d targ_long[%d]=%d not equal %d\n",
my_pe,iter,i,j,targ_long[j],iter*10000+elements[i]*100+my_pe+j-1);
if ( targ_float[j] != (float)(iter*10000+elements[i]*100+my_pe+j-1) )
fprintf(stderr,
"FAIL: PE [%d] iter=%d i=%d targ_long[%d]=%f not equal %d\n",
my_pe,iter,i,j,targ_float[j],iter*10000+elements[i]*100+my_pe+j-1);
if ( targ_double[j] != (double)(iter*10000+elements[i]*100+my_pe+j-1) )
fprintf(stderr,
"FAIL: PE [%d] iter=%d i=%d targ_double[%d]=%f not equal %d\n",
my_pe,iter,i,j,targ_double[j],iter*10000+elements[i]*100+my_pe+j-1);
}
}
}
}
}
shmem_free(srce_short); shmem_free(targ_short);
shmem_free(srce_int); shmem_free(targ_int);
shmem_free(srce_long); shmem_free(targ_long);
shmem_free(srce_float); shmem_free(targ_float);
shmem_free(srce_double); shmem_free(targ_double);
#ifdef NEEDS_FINALIZE
shmem_finalize();
#endif
return 0;
}