#include "pvm3.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 = ...;
if (pvm_parent() == PvmNoParent) {
#define N (4)
int tid[N];
pvm_spawn("program", argv, PvmTaskDefault, (char*)0, N, &tid[0]);
} else {
int mytid = pvm_mytid();
int *tids, me = -1;
int ntids = pvm_siblings(&tids);
for (int i=0; i<ntids; ++i)
if ( tids[i] == mytid) {
me = i;
break;
}
int p_left = -1, p_right = -1;
if (me > 0)
p_left = tids[me-1];
if (me < ntids-1)
p_right = tids[me+1];
int n_local0 = 1 + (me * (n-1)) / ntids;
int n_local1 = 1 + ((me+1) * (n-1)) / ntids;
pvm_joingroup("worker");
// allocate only local part + ghost zone of the arrays x,y
float *x, *y;
x = cache_aligned_allocator<float>().allocate(n_local1 - n_local0 + 2);
y = cache_aligned_allocator<float>().allocate(n_local1 - n_local0 + 2);
x -= (n_local0 - 1);
y -= (n_local0 - 1);
... // fill x, y
// fill ghost zone
if (p_left != -1) {
pvm_initsend(PvmDataDefault);
pvm_pkfloat(&y[n_local0], 1, 1);
int msgtag = 1;
pvm_send(p_left, msgtag);
}
if (p_right != -1) {
int msgtag = 1;
pvm_recv(p_right, msgtag);
pvm_upkfloat(&y[n_local1], 1, 1);
pvm_initsend(PvmDataDefault);
pvm_pkfloat(&y[n_local1-1], 1, 1);
msgtag = 2;
pvm_send(p_right, msgtag);
}
if (p_left != -1) {
int msgtag = 2;
pvm_recv(p_left, msgtag);
pvm_upkfloat(&y[n_local0-1], 1, 1);
}
thread1 t(x, y);
parallel_reduce(blocked_range<int>(n_local0, n_local1, 1024), t);
float e = t.ee;
int msgtag = 3;
pvm_reduce(PvmSum, &e, 1, PVM_FLOAT, msgtag, "worker", tids[0]);
msgtag = 4;
if (me==0) {
pvm_initsend(PvmDataDefault);
pvm_pkfloat(&e, 1, 1);
pvm_bcast("worker", msgtag);
} else {
pvm_recv(tids[0], msgtag);
pvm_upkfloat(&e, 1, 1);
}
... // output x, e
x += (n_local0 - 1);
y += (n_local0 - 1);
cache_aligned_allocator<float>().deallocate(x, n_local1 - n_local0 + 2);
cache_aligned_allocator<float>().deallocate(y, n_local1 - n_local0 + 2);
}
pvm_exit();
return 0;
}
|