public class scattergatherbw{

    public static void runGather(int sz){
	int maxMB = 100;
	int MBpertest = maxMB;
	int maxiters = 10000;
	int numProcs = Ti.numProcs();
	int halfProcs = numProcs / 2;
	int startIndex;
	int thisProc = Ti.thisProc();

	if (Ti.thisProc() < halfProcs){
	    startIndex = halfProcs;
	}
	else{
	    startIndex = 0;
	}

	double [1d] single [1d] allArrs = new double[0:Ti.numProcs()-1][1d];
	double [1d] myArr = new double[1:sz];
	
	int iters = (MBpertest * 1048576) / (sz*8);
	if (iters > maxiters) iters = maxiters;

	allArrs.exchange(myArr);
	double [1d] local dest = new double[1:sz];
	Point <1>[1d] local points = new Point<1>[dest.domain()];
	foreach (p in points.domain()){
	    points[p] = p;
	}
	Ti.barrier();
	
	if (Ti.thisProc() < halfProcs){
	    long starttime = System.currentTimeMillis();
	    for (int i = 0; i < iters; i++) {
		for (int j = 0; j < halfProcs; j++){
		    double [1d] peerArr = allArrs[(j + thisProc) % halfProcs + startIndex];
		    peerArr.gather(dest, points);
		}
	    }
	    long endtime = System.currentTimeMillis();
	    double totalsec = ((double)endtime - starttime) / 1000.0;
	    double totalMB = ((double)iters) * (sz * 8 * halfProcs) / 1048576.0;
	    System.out.println("gather bandwidth ("+ (sz * halfProcs) +" doubles per thread) = " + (totalMB/totalsec) + " MB/sec");
	}

	Ti.barrier();
    }

    public static void runScatter(int sz){
	int maxMB = 100;
	int MBpertest = maxMB;
	int maxiters = 10000;
	int numProcs = Ti.numProcs();
	int halfProcs = numProcs / 2;
	int startIndex;
	int thisProc = Ti.thisProc();

	if (Ti.thisProc() < halfProcs){
	    startIndex = halfProcs;
	}
	else{
	    startIndex = 0;
	}

	double [1d] single [1d] allArrs = new double[0:Ti.numProcs()-1][1d];
	double [1d] myArr = new double[1:sz];
	
	int iters = (MBpertest * 1048576) / (sz*8);
	if (iters > maxiters) iters = maxiters;

	allArrs.exchange(myArr);
	double [1d] local dest = new double[1:sz];
	Point <1>[1d] local points = new Point<1>[dest.domain()];
	foreach (p in points.domain()){
	    points[p] = p;
	}
	Ti.barrier();
	
	if (Ti.thisProc() < halfProcs){
	    long starttime = System.currentTimeMillis();
	    for (int i = 0; i < iters; i++) {
		for (int j = 0; j < halfProcs; j++){
		    double [1d] peerArr = allArrs[(j + thisProc) % halfProcs + startIndex];
		    
		    peerArr.scatter(dest, points);
		}
	    }
	    long endtime = System.currentTimeMillis();
	    double totalsec = ((double)endtime - starttime) / 1000.0;
	    double totalMB = ((double)iters) * (sz * 8 * halfProcs) / 1048576.0;
	    System.out.println("scatter bandwidth ("+ (sz * halfProcs) +" doubles per thread) = " + (totalMB/totalsec) + " MB/sec");
	}

	Ti.barrier();
    }

    public static single void main(String[] args){
	int single max;
	int single min;
	int single sz;
	max = 1048576;
	min = 256;
	for (sz = min; sz <= max; sz*=2){
	    runGather(sz);
	}

	if (Ti.thisProc() == 0){
	    System.out.println();
	}
	Ti.barrier();
	for (sz = min; sz <= max; sz*=2){
	    runScatter(sz);
	}
    } 
}