#include <mpp/shmem.h>
#include <tbb/task_scheduler_init.h>
#include <tbb/blocked_range.h>
#include <tbb/parallel_reduce.h>
#include <tbb/cache_aligned_allocator.h>
using namespace tbb;
#include <xmmintrin.h>
struct thread1 {
float ee;
float *x, *y;
thread1(float *xx, float *yy) : ee(0), x(xx), y(yy) {}
thread1(thread1& s, split) { ee = 0; x = s.x; y = s.y; }
void operator() (const blocked_range<int> & r) {
// do computation
float ve[4] = {0, 0, 0, 0};
float e = ee;
for (int i=r.begin(); i!=r.end(); 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];
ee = e;
}
void join(thread1& s) { ee += s.ee; }
};
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();
thread1 t(x, y);
parallel_reduce(blocked_range<int>(n_local0, n_local1, 1024), t);
float e = t.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;
}
|