code example
MPI message passing interface
two-sided communication.
POSIX threads
AltiVec PowerPC SIMD instructions

#include <mpi.h>
#include <pthread.h>
#include <altivec.h>

float *x, *y;
#define PROC (8)
float e_vec[PROC];
int n_thread0, n_thread1;

void *thread1(void *arg) {
  int p = (int)arg;
  int n0 = n_thread0 + (p * (n_thread1-n_thread0)) / PROC;
  int n1 = n_thread0 + ((p+1) * (n_thread1-n_thread0)) / PROC;
  // do computation
  float ve[4] = {0, 0, 0, 0};
  float e = 0;
  for (int i=n0; i<n1; i+=4) {
   float* yp = &y[i+1], y0 = &y[i], ym = &y[i-1];
   vec_st(vec_madd(
    vec_splats(.5),
    vec_add(
     vec_perm(vec_ld(0,ym), vec_ld(16,ym),
      vec_lvsl(0,ym)),
     vec_perm(vec_ld(0,yp), vec_ld(16,yp),
      vec_lvsl(0,yp))),
    vec_splats(0.)),
    0, &x[i]);
   vec_st(vec_add(
    vec_ld(0,&ve[0]),vec_madd(
     vec_ld(0,y0),
     vec_ld(0,y0),
     vec_splats(0.))),
    0, &ve[0]);
  }
  e += ve[0] + ve[1] + ve[2] + ve[3];
  e_vec[p] = e;
  return (void*) 0;
}

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
  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
  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);
 

  pthread_t threads[PROC];
  pthread_attr_t attr;
  pthread_attr_init(&attr);
  n_thread0 = n_local0;
  n_thread1 = n_local1;
  float e = 0;
  // start threads and wait for termination
  for (int p=0; p<PROC; ++p)
   pthread_create(&threads[p], &attr, thread1, (void *)p);
  for (int p=0; p<PROC; ++p) {
   pthread_join(threads[p], NULL);
   e += e_vec[p];
  }

  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);
  delete[] x, y;
  MPI_Finalize();
  return 0;
}

[start] [references] [download] [install]