#include "pvm3.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 = ...;
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 = new float[n_local1 - n_local0 + 2];
y = new float[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);
}
float e = 0;
#pragma hmpp code1 callsite
sub1(n_local1-n_local0, &x[n_local0], &y[n_local0-1], &e);
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);
delete[] x, y;
}
pvm_exit();
return 0;
}
|