#include <mpp/shmem.h>
#pragma hmpp code1 codelet, &
#pragma hmpp code1 target=CUDA:BROOK:PTHREADSSE:PTHREAD:SSE, &
#pragma hmpp code1 args[x,e].io=inout
static void sub1(int n, float x[n], float y[n+2], float e[1]) {
int i;
float e_local = e[0];
// do computation
for (i=0; i<n; ++i) {
x[i] += ( y[i+2] + y[i] )*.5;
e_local += y[i+1] * y[i+1];
}
e[0] = e_local;
}
int main(int argc, char *argv[]) {
int n = ...;
start_pes(0);
int nn = (n-1) / _num_pes();
int n_local0 = 1 + _my_pe() * nn;
int n_local1 = 1 + (_my_pe()+1) * nn;
// allocate only local part + ghost zone of the arrays x,y
float *x, *y;
x = (float*)shmalloc((n_local1 - n_local0 + 2)*sizeof(float));
y = (float*)shmalloc((n_local1 - n_local0 + 2)*sizeof(float));
x -= (n_local0 - 1);
y -= (n_local0 - 1);
shmem_barrier_all();
... // fill x, y
// fill ghost zone
if (_my_pe() > 0)
shmem_float_put(&y[n_local1], &y[n_local0], 1, _my_pe()-1);
if (_my_pe() < _num_pes()-1)
shmem_float_put(&y[n_local0-1], &y[n_local1-1], 1, _my_pe()+1);
shmem_barrier_all();
float e = 0;
#pragma hmpp code1 callsite
sub1(n_local1-n_local0, &x[n_local0], &y[n_local0-1], &e);
static float work[_SHMEM_REDUCE_MIN_WRKDATA_SIZE];
static long sync[_SHMEM_REDUCE_MIN_WRKDATA_SIZE];
static float el, es;
el = e;
shmem_float_sum_to_all(&es, &el, 1,
0, 0, _num_pes(), work, sync);
e = es;
... // output x, e
x += (n_local0 - 1);
y += (n_local0 - 1);
shfree(x);
shfree(y);
return 0;
}
|