// tiArray performance tester, Dan Bonachea

public class arrperf {
  public static void report(String desc, Timer t, int iters) {
    System.out.println(desc + ": " + t.micros()/iters + " us per iteration ("
                       + t.secs() + " secs total)");
  }

  public single static void main(String[]args) {
    int iters = 10000;
    int sz = 100;
    if (args.length > 0) 
      try { iters = Integer.parseInt(args[0]); } catch (Throwable exn) {}
    if (args.length > 1) 
      try { sz = Integer.parseInt(args[1]); } catch (Throwable exn) {}
    System.out.println("Running " + iters + " iterations of tiArray performance test(sz="+sz+")...");
    Timer t = new Timer();

    RectDomain<1> rd1 =  [1:sz];
    RectDomain<1> rd1a = [2:sz-1];
    RectDomain<2> rd2 =  [[1,1]:[sz,sz]];
    RectDomain<2> rd2a = [[2,2]:[sz-1,sz-1]];
    RectDomain<3> rd3 =  [[1,1,1]:[sz,sz,sz]];
    RectDomain<3> rd3a = [[2,2,2]:[sz-1,sz-1,sz-1]];
    //------------------------------------------------------------------------------------
    // non-contiguous copy benchmark

    int [1d] local ltiarr1a = new int[rd1];
    int [2d] local ltiarr2a = new int[rd2];
    int [3d] local ltiarr3a = new int[rd3];

    int [1d] local ltiarr1b = new int[rd1];
    int [2d] local ltiarr2b = new int[rd2];
    int [3d] local ltiarr3b = new int[rd3];

    int [1d] tiarr1a = broadcast ltiarr1a from 0;
    int [2d] tiarr2a = broadcast ltiarr2a from 0;
    int [3d] tiarr3a = broadcast ltiarr3a from 0;

    int [1d] tiarr1b = broadcast ltiarr1b from 0;
    int [2d] tiarr2b = broadcast ltiarr2b from 0;
    int [3d] tiarr3b = broadcast ltiarr3b from 0;

    ltiarr1b = ltiarr1b.restrict(rd1a);
    ltiarr2b = ltiarr2b.restrict(rd2a);
    ltiarr3b = ltiarr3b.restrict(rd3a);

    tiarr1b = tiarr1b.restrict(rd1a);
    tiarr2b = tiarr2b.restrict(rd2a);
    tiarr3b = tiarr3b.restrict(rd3a);

    t.reset();
    t.start();
    for (int i=0; i < iters; i++) {
      tiarr1a.copy(tiarr1b);
    }
    t.stop();
    report("non-contiguous local 1-d global array copy", t, iters);

    t.reset();
    t.start();
    for (int i=0; i < iters; i++) {
      ltiarr1a.copy(ltiarr1b);
    }
    t.stop();
    report("non-contiguous local 1-d local array copy", t, iters);

    t.reset();
    t.start();
    for (int i=0; i < iters; i++) {
      tiarr2a.copy(tiarr2b);
    }
    t.stop();
    report("non-contiguous local 2-d global array copy", t, iters);

    t.reset();
    t.start();
    for (int i=0; i < iters; i++) {
      ltiarr2a.copy(ltiarr2b);
    }
    t.stop();
    report("non-contiguous local 2-d local array copy", t, iters);

    t.reset();
    t.start();
    for (int i=0; i < iters; i++) {
      tiarr3a.copy(tiarr3b);
    }
    t.stop();
    report("non-contiguous local 3-d global array copy", t, iters);

    t.reset();
    t.start();
    for (int i=0; i < iters; i++) {
      ltiarr3a.copy(ltiarr3b);
    }
    t.stop();
    report("non-contiguous local 3-d local array copy", t, iters);

  }}