/* this is a test program for my qcdoc communication routines */
/* Michael Creutz */

#define DEBUG 1
#define BLOCKSIZE 8

#include <stdio.h>
#include <time.h>
#include "include/qmemcpyoc.C"

int main() {
  int i,sum;
  double time, dsum, x ,y;
#define BUFFERSIZE 50  
  char buffer1[BUFFERSIZE];
  char buffer2[BUFFERSIZE];
  mycom.start();
  printf("testing stop and start\n");
  mycom.stop();
  mycom.start();
  printf("done with stop start test\n");
  /* put something in the buffers */
  sprintf(buffer1,"hello world from processor %d\n",processor);
  sprintf(buffer2,"buffer two\n");
  printf("   buffer 1: %s   buffer2: %s",buffer1,buffer2);

  printf("local transfer test\n");
  mycom.qmemcpy(processor,buffer2,processor,buffer1,BUFFERSIZE);
  mycom.finishcopy();
  printf("   buffer 1: %s   buffer2: %s",buffer1,buffer2);

  printf("remote fetch test:\n");
  mycom.qmemcpy(processor,buffer2,
		(processor+7)%NumNodes(),buffer1,BUFFERSIZE);
  mycom.finishcopy();
  printf("   buffer 1: %s   buffer2: %s",buffer1,buffer2);
  printf("%d: maxqueue = %d\n",processor, mycom.maxqueue);
  mycom.maxqueue=0;

  printf("single processor remote fetch test:\n");
  if (processor==0) mycom.qmemcpy(processor,buffer2,
				  (processor+2)%NumNodes(),buffer1,BUFFERSIZE);
  mycom.finishcopy();
  printf("   buffer 1: %s   buffer2: %s",buffer1,buffer2);
  printf("%d: maxqueue = %d\n",processor, mycom.maxqueue);
  mycom.maxqueue=0;

  printf("remote store test:\n");
  mycom.qmemcpy((processor+3)%NumNodes(),buffer2,
		processor              ,buffer1,BUFFERSIZE);
  mycom.finishcopy();
  printf("   buffer 1: %s   buffer2: %s",buffer1,buffer2);
  printf("%d: maxqueue = %d\n",processor, mycom.maxqueue);
  mycom.maxqueue=0;

  printf("remote to same remote test:\n");
  mycom.qmemcpy((processor+7)%NumNodes(),buffer2,
                (processor+7)%NumNodes(),buffer1,BUFFERSIZE);
  mycom.finishcopy();
  printf("   buffer 1: %s   buffer2: %s",buffer1,buffer2);
  printf("%d: maxqueue = %d\n",processor, mycom.maxqueue);
  mycom.maxqueue=0;

  printf("remote to different remote test:\n");
  mycom.qmemcpy((processor+13)%NumNodes(),buffer2,
                (processor+18)%NumNodes(),buffer1,BUFFERSIZE);
  mycom.finishcopy();
  printf("   buffer 1: %s   buffer2: %s",buffer1,buffer2);
  printf("%d: maxqueue = %d\n",processor, mycom.maxqueue);
  mycom.maxqueue=0;

  printf("integer global sum test:\n");
  time = (1.0 * clock()) / CLOCKS_PER_SEC;
  for (i=0;i<1000;i++)  sum=mycom.igsum(processor);
  time = (1.0 * clock()) / CLOCKS_PER_SEC - time;
  printf("sum of processor id's = %d\n",sum);
  printf(" 1000 integer global sums took %g seconds\n", time);
  printf("%d: maxqueue = %d\n",processor, mycom.maxqueue);
  mycom.maxqueue=0;

  printf("double global sum test:\n");
  time = (1.0 * clock()) / CLOCKS_PER_SEC;
  for (i=0;i<1000;i++)  dsum=mycom.dgsum(0.7*processor);
  time = (1.0 * clock()) / CLOCKS_PER_SEC - time;
  printf("sum of .7* processor id's = %g\n",dsum);
  printf(" 1000 double global sums took %g seconds\n", time);
  printf("%d: maxqueue = %d\n",processor, mycom.maxqueue);
  mycom.maxqueue=0;

  /* broadcast test */
  sprintf(buffer1,"processor %d says hello\n",processor);
  time = (1.0 * clock()) / CLOCKS_PER_SEC;
  for (i=0;i<1000;i++) mycom.bcast(4%NumNodes(),buffer1,BUFFERSIZE);
  time = (1.0 * clock()) / CLOCKS_PER_SEC - time;
  printf("broadcast test from processor %d:\n   buffer1: %s",
	 4%NumNodes(),buffer1);
  printf(" 1000 broadcasts of char buffer[%d] took %g seconds\n",
	 BUFFERSIZE,time);
  printf("%d: maxqueue = %d\n",processor, mycom.maxqueue);
  mycom.maxqueue=0;
  printf("sync_count=%d\n",mycom.sync_count);

  x=1.23;
  y=3.45;
  printf("addit test x=%g y=%g\n",x,y);
  if ((processor+3)%NumNodes()==0)
    mycom.addit(0,&x,y);
  mycom.finishcopy();
  printf("x + remote y = %g\n",x);

  x=1.0*processor;
  printf("gmax test: max processor id = %g\n", mycom.dgmax(x));

  mycom.stop();
  return 0;
}