#include <mpi.h>
#include <boost/thread/thread.hpp>
using namespace boost;
#define PROC (8)
// 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;
}
struct thread1 {
float *x, *y, *ep;
int i0, i1, p;
thread1(float *xx, float *yy, float *ee, int pp, int ii0, int ii1) :
x(xx), y(yy), ep(ee), p(pp), i0(ii0), i1(ii1) {}
void operator()() {
#define BLOCK (8190)
float xStream<BLOCK>;
float xoutStream<BLOCK>;
float yStream<BLOCK+2>;
float einStream<BLOCK+2>;
float eStream<1>;
float e = 0;
float ee = 0;
for (int i=i0; i<i1; 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;
}
e += ee;
*ep = e;
}
};
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;
// allocate only local part + ghost zone of the arrays x,y
float *x, *y;
MPI_Alloc_mem(sizeof(float) * (n_local1 - n_local0 + 2),
MPI_INFO_NULL, &x);
MPI_Alloc_mem(sizeof(float) * (n_local1 - n_local0 + 2),
MPI_INFO_NULL, &y);
x -= (n_local0 - 1);
y -= (n_local0 - 1);
MPI_Win win;
MPI_Win_create(&y[n_local0], sizeof(float) * (n_local1-n_local0+2),
sizeof(float), MPI_INFO_NULL, MPI_COMM_WORLD, &win);
... // fill x, y
// fill ghost zone
MPI_Win_fence(0, win);
if (p_left != -1)
MPI_Put(&y[n_local0], 1, MPI_FLOAT, p_left,
n_local1-n_local0+1, 1, MPI_FLOAT, win);
if (p_right != -1)
MPI_Put(&y[n_local1-1], 1, MPI_FLOAT, p_right,
0, 1, MPI_FLOAT, win);
MPI_Win_fence(0, win);
float e = 0;
float e_vec[PROC];
thread_group grp;
// start threads and wait for termination
for (int i=0; i<PROC; ++i) {
thread1 t(x, y, &e_vec[i], i,
n_local0+((n_local1-n_local0)*i)/PROC, n_local0+((n_local1-n_local0)*(i+1))/PROC);
grp.create_thread(t);
}
grp.join_all();
for (int i=0; i<PROC; ++i)
e += e_vec[i];
float e_local = e;
MPI_Allreduce(&e_local, &e, 1, MPI_FLOAT, MPI_SUM, MPI_COMM_WORLD);
... // output x, e
MPI_Win_free(&win);
x += (n_local0 - 1);
y += (n_local0 - 1);
MPI_Free_mem(y);
MPI_Free_mem(x);
MPI_Finalize();
return 0;
}
|