// matmul_thread4.java  4 threads for 4 or more cores, parallel execution
import java.lang.management.ManagementFactory;
import java.lang.management.ThreadMXBean;

public class matmul_thread4
{
  int nthread = 4;
  double cpu_start[] = new double[nthread]; // for each thread           
  double cpu_end[]   = new double[nthread];
  int DIM = 2048;
  double a[][] = new double[DIM][DIM];
  double b[][] = new double[DIM][DIM];
  double c[][] = new double[DIM][DIM];
  Thread t[] = new Thread[nthread];
  ThreadMXBean tb = ManagementFactory.getThreadMXBean();

  public matmul_thread4()
  {
    double twall = 0.0, tcpu = 0.0, total_tcpu;

    System.out.println("matmul_thread4.java running");
    System.out.println("DIM="+DIM+", nthread="+nthread);

    // initialize arrays (not timed)
    for(int i=0; i<DIM; i++)
    {
      for(int j=0; j<DIM; j++)
      {
        a[i][j] = (double)(i+j);
        b[i][j] = (double)(i+j+1);
      }
    }

    twall = System.currentTimeMillis()/1000.0;
    // start threads
    System.out.println("constructors starting");
    for(int j=0; j<nthread; j++) t[j] = new MyThread(j);
    for(int j=0; j<nthread; j++) t[j].start();
    try
    {
      for(int j=0; j<nthread; j++) t[j].join(); // wait for all to complete
    } // end try
    catch(InterruptedException e)
    {
      System.out.println("exception in join"+e);
    } // end catch

    twall = System.currentTimeMillis()/1000.0 - twall;
    System.out.println("c[1][1]="+c[1][1]+
		       ", c[2][2]="+c[2][2]);
    System.out.println("c[513][513]="+c[513][513]+
		       ",c[514][514]="+c[514][514]);
    System.out.println("c[1025][1025]="+c[1025][1025]+
		       ",c[1026][1026]="+c[1026][1026]);
    System.out.println("c[1538][1538]="+c[1538][1538]+
		       ",c[1539][1539]="+c[1539][1539]);
    
    total_tcpu = 0.0;
    for(int i=0; i<nthread; i++)
    {
      tcpu = cpu_end[i] - cpu_start[i];
      total_tcpu += tcpu;
      System.out.println("thread "+i+" took "+tcpu+" cpu sec");
    }

    System.out.println("took "+twall+" wall sec");
    System.out.println("average number of parallel threads "+(total_tcpu/twall));
 
  } // end matmul_thread4

  class MyThread extends Thread
  {
    // DIM, a, b, c, nthread, cpu_start, cpu_end    are global
    int id, part, i1, i2; // local to each thread
    double sum;

    MyThread(int part)
    {
      id = part;
      i1 = part*DIM/nthread;
      i2 = ((part+1)*DIM/nthread);
    } // end MyThread
    public void run()
    {
      System.out.println("id="+id+", i1="+i1+",i2="+i2);
      cpu_start[id] = tb.getCurrentThreadCpuTime()/1.0E9;

      for(int i=i1; i<i2; i++) // local to this thread
      {
        for(int j=0; j<DIM; j++)
        {
          sum=0.0;
          for(int k=0; k<DIM; k++)
          {
            sum += a[i][k]*b[k][j];
          } // end k
          c[i][j]=sum;
        } // end j
      } // end i
      System.out.println("id="+id+" finished");
      try
      {
        sleep(0);
      } // end try
      catch(InterruptedException e)
      {
        System.out.println("exception in MyThread run"+e);
      }	 // end catch
      cpu_end[id] = tb.getCurrentThreadCpuTime()/1.0E9;
    } // end run
  } // end class MyThread      

  public static void main (String[] args)
  {
    new matmul_thread4();
  }

} // end matmul_thread4.java