#include <stdio.h>
#include <sys/times.h>
#include <sys/types.h>
#include <time.h>

#define CACHE_MIN (1024)
#define CACHE_MAX (1024*1024)
#define SAMPLE 10
int x[CACHE_MAX];

double get_seconds() {
  struct tms rusage;
  times(&rusage);
return (double) (rusage.tms_utime)/CLK_TCK;
}
double get_secondsr() {
  struct tms rusage;
  return (double) times(&rusage)/CLK_TCK;
}

void main()
{
     int register i, index, stride, limit, temp;
     int steps, tsteps, csize;
     double sec0, sec; /* timing variables */
     char buf[1000];
     FILE *f;

     for (csize = CACHE_MIN; csize <= CACHE_MAX; csize=csize*2) {
	 sprintf (buf, "u%d.dat", csize);
	 f = fopen (buf, "w");
	 fprintf (f, "# Size %d, Stride, Read+write (user)\n", csize * sizeof(int));

       for (stride = 1; stride <= csize/2; stride=stride*2) {

	 sec = 0;
	 limit = csize - stride + 1;

	 steps = 0;
	 do {
	   sec0 = get_seconds();
	   for (i=SAMPLE*stride; i != 0; i--) {
	     for (index = 0; index < limit; index = index + stride) {
	       x[index] = x[index] + 1;
	     }
	   }
	   steps = steps + 1;
	   sec = sec + (get_seconds() - sec0);
	 } while (sec < 1.0);

	 tsteps = 0;
	 do {
	   sec0 = get_seconds();
	   for (i=SAMPLE*stride; i != 0; i--) {
	     for (index = 0; index < limit; index = index + stride) {
	       temp = temp + index;
	     }
	   }
	   tsteps = tsteps + 1;
	   sec = sec + (get_seconds() - sec0);
	 } while (tsteps < steps);
	 fprintf(f, "%7d %14.0f\n", stride*sizeof(int),
		(double)sec*1e9/(steps*SAMPLE*stride*((limit-1)/stride+1)));
	 printf("Size: %7d Stride: %7d read+write: %14.0f ns\n",
		csize*sizeof(int), stride*sizeof(int),
		(double)sec*1e9/(steps*SAMPLE*stride*((limit-1)/stride+1)));
       }
       fclose (f);
     }
     for (csize = CACHE_MIN; csize <= CACHE_MAX; csize=csize*2) {
	 sprintf (buf, "r%d.dat", csize);
	 f = fopen (buf, "w");
	 fprintf (f, "# Size %d, Stride, Read+write (real)\n", csize * sizeof(int));

       for (stride = 1; stride <= csize/2; stride=stride*2) {

	 sec = 0;
	 limit = csize - stride + 1;

	 steps = 0;
	 do {
	   sec0 = get_secondsr();
	   for (i=SAMPLE*stride; i != 0; i--) {
	     for (index = 0; index < limit; index = index + stride) {
	       x[index] = x[index] + 1;
	     }
	   }
	   steps = steps + 1;
	   sec = sec + (get_secondsr() - sec0);
	 } while (sec < 1.0);

	 tsteps = 0;
	 do {
	   sec0 = get_secondsr();
	   for (i=SAMPLE*stride; i != 0; i--) {
	     for (index = 0; index < limit; index = index + stride) {
	       temp = temp + index;
	     }
	   }
	   tsteps = tsteps + 1;
	   sec = sec + (get_secondsr() - sec0);
	 } while (tsteps < steps);
	 fprintf(f, "%7d %14.0f\n", stride*sizeof(int),
		(double)sec*1e9/(steps*SAMPLE*stride*((limit-1)/stride+1)));
	 printf("Size: %7d Stride: %7d read+write: %14.0f ns\n",
		csize*sizeof(int), stride*sizeof(int),
		(double)sec*1e9/(steps*SAMPLE*stride*((limit-1)/stride+1)));
       }
       fclose (f);
     }
     f = fopen ("cache.plt", "w");
     fprintf (f, "plot \\\n");
     for (csize = CACHE_MIN; csize <= CACHE_MAX; csize=csize*2) {
       fprintf (f, "'u%d.dat' with linespoints, \\\n", csize);
       fprintf (f, "'r%d.dat' with linespoints%c \\\n", csize,
		csize == CACHE_MAX ? ' ' : ',');
     }
     fclose (f);
}