#include <mpp/shmem.h>
#include <omp.h>
float e;
// kernel
kernel void sub1(float x[], float y[],
out float xout<>, out float eout<>)
{
// do computation
float i = 1 + indexof(xout).x;
xout = x[i] + (y[i+1] + y[i-1])*.5f;
eout = y[i] * y[i];
}
reduce void sub2(float y<>, reduce float e<>)
{
e += y;
}
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();
e = 0;
#pragma omp parallel
{
int p = omp_get_thread_num();
int num = omp_get_num_threads();
#define BLOCK (8190)
float xStream<BLOCK>;
float xoutStream<BLOCK>;
float yStream<BLOCK+2>;
float einStream<BLOCK+2>;
float eStream<1>;
int n0 = 1+((n_local1-n_local0)*p)/num;
int n1 = 1+((n_local1-n_local0)*(p+1))/num;
float ee = 0;
for (int i=n0; i<n1; i+=BLOCK) {
streamRead(xStream, x+i);
streamRead(yStream, y+i-1);
sub1(xStream, yStream, xoutStream, einStream);
streamWrite(xoutStream, x+i);
sub2(einStream, eStream);
float e_local;
streamWrite(eStream, &e_local);
ee += e_local;
}
#pragma omp atomic
e += ee;
}
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;
}
|