Data Parallel

Parallel Arrays in Java

Currently, take a pure class library approach, similar to A++/P++, HPC++ parallel STL, Adlib, etc.

We parametrize an array by a member of the Array class, similar to an HPF template. Index ranges of array distributed over dimensions of process grid, represented by Procs class.

Parallel Java application written as a class extending library class Node: maintains some global information and provides collective operations on arrays.

Data Parallel

Data-parallel example: initialization

  Procs p = new Procs(this, 2, 2) ;

  Range x = new Range(N, p, 0) ;  // distrib over 1st dim of `p'
  Range y = new Range(N, p, 1) ;  // distrib over 2nd dim of `p'

  Array r = new Array(p, x, y) ;

  byte [] w = new byte [r.seg()] ;  // main data array

  ... declare neighbour arrays, `cn_', `cp_', etc, similarly

  // initialize the ``life'' board

  for(r.forall() ; r.test() ; r.next())
    w [r.sub()] = fun(x.idx(), y.idx()) ;

  for (int k=0; k<NITER; k++) {

    // Get neighbours

    shift(cn_, w, r, 0,  1, CYCLIC);
    shift(cp_, w, r, 0, -1, CYCLIC);
    ... etc, copy arrays for 8 neighbours

    // Life update rule

    for(int i=0; i<w.length; i++) {
      switch (cn_[i] + cp_[i] + c_n[i] + c_p[i] +
              cnn[i] + cnp[i] + cpn[i] + cpp[i]) {
        case 2 : break;
        case 3 : w[i] = 1; break;
        default: w[i] = 0; break;
      }
    }
  }

  ArrayFloat a = new ArrayFloat(p, x, y) ;
  ArrayFloat b = new ArrayFloat(p, x, y) ;
  ArrayInt c = new ArrayInt(p, x, y) ;

  a = MATMUL(b, c) ;