#include <mpp/shmem.h>
#include <pthread.h>
#include <xmmintrin.h>
float *x, *y;
#define PROC (8)
float e_vec[PROC];
int n_thread0, n_thread1;
void *thread1(void *arg) {
int p = (int)arg;
int n0 = n_thread0 + (p * (n_thread1-n_thread0)) / PROC;
int n1 = n_thread0 + ((p+1) * (n_thread1-n_thread0)) / PROC;
// do computation
float ve[4] = {0, 0, 0, 0};
float e = 0;
for (int i=n0; i<n1; i+=4) {
float half =.5;
_mm_store_ps(&x[i],
_mm_mul_ps(_mm_load1_ps(&half),
_mm_add_ps(_mm_loadu_ps(&y[i+1]),
_mm_loadu_ps(&y[i-1]))));
_mm_store_ps(&ve[0],
_mm_add_ps(_mm_load_ps(&ve[0]),
_mm_mul_ps(_mm_load_ps(&y[i]),
_mm_load_ps(&y[i]))));
}
e += ve[0] + ve[1] + ve[2] + ve[3];
e_vec[p] = e;
return (void*) 0;
}
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
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();
pthread_t threads[PROC];
pthread_attr_t attr;
pthread_attr_init(&attr);
n_thread0 = n_local0;
n_thread1 = n_local1;
float e = 0;
// start threads and wait for termination
for (int p=0; p<PROC; ++p)
pthread_create(&threads[p], &attr, thread1, (void *)p);
for (int p=0; p<PROC; ++p) {
pthread_join(threads[p], NULL);
e += e_vec[p];
}
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;
}
|