#include <mpi.h>
#include <rapidmind/platform.hpp>
using namespace rapidmind;
int main(int argc, char *argv[]) {
int n = ...;
MPI_Init(&argc, &argv);
int numproc, me;
MPI_Comm_size(MPI_COMM_WORLD, &numproc);
MPI_Comm_rank(MPI_COMM_WORLD, &me);
int p_left = -1, p_right = -1;
if (me > 0)
p_left = me-1;
if (me < numproc-1)
p_right = me+1;
int n_local0 = 1 + (me * (n-1)) / numproc;
int n_local1 = 1 + ((me+1) * (n-1)) / numproc;
rapidmind::init();
// allocate only local part + ghost zone of the arrays x,y
float *x, *y;
Array<1, Value1f> xr(n_local1 - n_local0 + 2);
Array<1, Value1f> yr(n_local1 - n_local0 + 2);
x = xr.write_data();
y = yr.write_data();
x -= (n_local0 - 1);
y -= (n_local0 - 1);
... // fill x, y
// fill ghost zone
MPI_Status s;
if (p_left != -1)
MPI_Send(&y[n_local0], 1, MPI_FLOAT, p_left,
1, MPI_COMM_WORLD);
if (p_right != -1) {
MPI_Recv(&y[n_local1], 1, MPI_FLOAT, p_right,
1, MPI_COMM_WORLD, &s);
MPI_Send(&y[n_local1-1], 1, MPI_FLOAT, p_right,
2, MPI_COMM_WORLD);
}
if (p_left != -1)
MPI_Recv(&y[n_local0-1], 1, MPI_FLOAT, p_left,
2, MPI_COMM_WORLD, &s);
Program add3 = RM_BEGIN {
In<Value1f> a;
In<Value1f> b;
In<Value1f> c;
Out<Value1f> d;
d = a + (b + c) * .5;
} RM_END;
Program add2 = RM_BEGIN {
In<Value1f> a;
In<Value1f> b;
Out<Value1f> c;
c = a + b;
} RM_END;
Program sqr = RM_BEGIN {
In<Value1f> a;
Out<Value1f> b;
b = a * a;
} RM_END;
Array<1, Value1f> x0 = add3(slice(xr,1,n_local1-n_local0),
offset(yr,2), take(yr,n_local1-n_local0));
xr = x0;
Array<1, Value1f> er = reduce(add2,
sqr(slice(yr,1,n_local1-n_local0));
float e = *er.read_data();
float e_local = e;
MPI_Allreduce(&e_local, &e, 1, MPI_FLOAT, MPI_SUM, MPI_COMM_WORLD);
... // output x, e
x += (n_local0 - 1);
y += (n_local0 - 1);
MPI_Finalize();
return 0;
}
|