// Titanium micro-benchmarker
// Dan Bonachea <boanchea@cs.berkeley.edu>
// 10/2/2001

public class perf {
  public int peerid;
  public boolean activeproc;

  public void report(String s) {
    System.out.println("P"+Ti.thisProc()+"-P" + peerid +": " + s);
  }

public Point<1> [1d] makePtArray(int sz, RectDomain<1> arrSz) {
  Point<1> [1d] retval = new Point<1>[1:sz];
  int max = arrSz.max()[1];
  foreach (p in retval.domain()) {
    retval[p] = [((int)(Math.random() * max)) + 1];
  }
  return retval;
}

  public sglobal void gatherbandwidth(int MBpertest, int maxiters, double sparsity) {
    boolean dense = sparsity == 0.0;
    for (int single sz = 1; sz <= 1048576; sz*=2) {
      RectDomain<1> arrSz = [1:(dense?sz:sz*(int)(1/sparsity))];
      double [1d] single [1d] allArrs = new double[0:Ti.numProcs()-1][1d];
      int iters = (MBpertest * 1048576) / (8*sz);
      if (iters > maxiters) iters = maxiters;
      double [1d] myArr = new double[arrSz];

      allArrs.exchange(myArr);
      double [1d] peerArr = allArrs[peerid];

      Ti.barrier();
      if (activeproc) {
        Point<1> [1d] ptArray;
        if (sparsity == 0.0) ptArray = ((Domain<1>)arrSz).PointList();
        else ptArray = makePtArray(sz,arrSz);
        double [1d] packedArr = new double[1:sz];

        Timer t = new Timer();
        t.start();
        for (int i = 0; i < iters; i++) {
          peerArr.gather(packedArr,ptArray);
        }
        t.stop();
        double totalsec = t.secs();
        double totalMB = ((double)iters) * sz * 8 / 1048576.0;
        report("gather bandwidth ("+sz*8+" byte packs) = " + (totalMB/totalsec) + " MB/sec");
      }
    }
  }

  public sglobal void scatterbandwidth(int MBpertest, int maxiters, double sparsity) {
    boolean dense = sparsity == 0.0;
    for (int single sz = 1; sz <= 1048576; sz*=2) {
      RectDomain<1> arrSz = [1:(dense?sz:sz*(int)(1/sparsity))];
      double [1d] single [1d] allArrs = new double[0:Ti.numProcs()-1][1d];
      int iters = (MBpertest * 1048576) / (8*sz);
      if (iters > maxiters) iters = maxiters;
      double [1d] myArr = new double[arrSz];

      allArrs.exchange(myArr);
      double [1d] peerArr = allArrs[peerid];

      Ti.barrier();
      if (activeproc) {
        Point<1> [1d] ptArray;
        if (sparsity == 0.0) ptArray = ((Domain<1>)arrSz).PointList();
        else ptArray = makePtArray(sz,arrSz);
        double [1d] packedArr = new double[1:sz];

        Timer t = new Timer();
        t.start();
        for (int i = 0; i < iters; i++) {
          peerArr.scatter(packedArr,ptArray);
        }
        t.stop();
        double totalsec = t.secs();
        double totalMB = ((double)iters) * sz * 8 / 1048576.0;
        report("scatter bandwidth ("+sz*8+" byte packs) = " + (totalMB/totalsec) + " MB/sec");
      }
    }
  }

  public sglobal void sparsecopybandwidth(int MBpertest, int maxiters, double sparsity) {
    boolean dense = sparsity == 0.0;
    for (int single sz = 1; sz <= 1048576; sz*=2) {
      RectDomain<1> arrSz = [1:(dense?sz:sz*(int)(1/sparsity))];
      double [1d] single [1d] allArrs = new double[0:Ti.numProcs()-1][1d];
      int iters = (MBpertest * 1048576) / (8*sz);
      if (iters > maxiters) iters = maxiters;
      double [1d] myArr = new double[arrSz];

      allArrs.exchange(myArr);
      double [1d] peerArr = allArrs[peerid];

      Ti.barrier();
      if (activeproc) {
        Point<1> [1d] ptArray;
        if (sparsity == 0.0) ptArray = ((Domain<1>)arrSz).PointList(); 
        else ptArray = makePtArray(sz,arrSz);

        Timer t = new Timer();
        t.start();
        for (int i = 0; i < iters; i++) {
          peerArr.copy(myArr,ptArray);
        }
        t.stop();
        double totalsec = t.secs();
        double totalMB = ((double)iters) * sz * 8 / 1048576.0;
        report("sparse copy bandwidth ("+sz*8+" byte packs) = " + (totalMB/totalsec) + " MB/sec");
      }
    }
  }


  // usage: perf [numiterations] [maxMBperiteration]

  public static void main(String [] args) {
    int iters;
    int maxMB;

    { int iters0 = 10000;
      int maxMB0 = 10;
      if (args.length > 0) 
        try { iters0 = Integer.parseInt(args[0]); } catch (Throwable exn) {}
      if (args.length > 1) 
        try { maxMB0 = Integer.parseInt(args[1]); } catch (Throwable exn) {}
      iters = broadcast iters0 from 0;
      maxMB = broadcast maxMB0 from 0;
    }

    perf single p = new perf();
    p.peerid = ( Ti.thisProc() + 1 ) % Ti.numProcs();
    p.activeproc = (Ti.thisProc() % 2 == 0);

    if (Ti.thisProc() == 0)
      System.out.println("Running bandwidth test (iterations at each chunk size: MAX("+iters+" iterations, "+maxMB+" MB))");

    if (Ti.thisProc() == 0)
      System.out.println("*** Remote bandwidth (dense) ***");

    p.gatherbandwidth(maxMB, iters, 0.0);
    p.scatterbandwidth(maxMB, iters, 0.0);
    p.sparsecopybandwidth(maxMB, iters, 0.0);

    for (int single i=1;i<=4;i++) {
      if (Ti.thisProc() == 0)
        System.out.println("*** Remote bandwidth (sparsity="+i*0.25+") ***");
      p.gatherbandwidth(maxMB, iters, i*0.25);
      p.scatterbandwidth(maxMB, iters, i*0.25);
      p.sparsecopybandwidth(maxMB, iters, i*0.25);
    }

    p.peerid = Ti.thisProc();
    p.activeproc = true;

    if (Ti.thisProc() == 0)
      System.out.println("*** Local bandwidth (dense) ***");

    p.gatherbandwidth(maxMB, iters, 0.0);
    p.scatterbandwidth(maxMB, iters, 0.0);
    p.sparsecopybandwidth(maxMB, iters, 0.0);

    for (int single i=1;i<=4;i++) {
      if (Ti.thisProc() == 0)
        System.out.println("*** Local bandwidth (sparsity="+i*0.25+") ***");
      p.gatherbandwidth(maxMB, iters, i*0.25);
      p.scatterbandwidth(maxMB, iters, i*0.25);
      p.sparsecopybandwidth(maxMB, iters, i*0.25);
    }


  }
}