code example
MPI-2 message passing interface, version 2
one-sided communication defined in MPI-2
POSIX threads
SSE IA SIMD instructions

#include <mpi.h>
#include <pthread.h>
#include <xmmintrin.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 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];
  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
  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);

  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

  MPI_Win_free(&win);
  x += (n_local0 - 1);
  y += (n_local0 - 1);
  MPI_Free_mem(y);
  MPI_Free_mem(x);
  MPI_Finalize();
  return 0;
}

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