/******************************************************************************
* FILE: loop12.c
* DESCRIPTION:
*   Demonstrates loop unrolling
******************************************************************************/

#include <stdio.h>
#include <sys/time.h>
#include <time.h>
#define	SEED		1995
#define	ARRAY_SIZE	400

float a, aa, aaa, aaaa, aaaaa, a1, a2, a3, a4, a5, a6, a7, a8;
float a9, a10, a11, a12, a13, a14, a15, a16;
float b[ARRAY_SIZE], bb[ARRAY_SIZE], bbb[ARRAY_SIZE], bbbb[ARRAY_SIZE], bbbbb[ARRAY_SIZE];
float c[ARRAY_SIZE], cc[ARRAY_SIZE], ccc[ARRAY_SIZE], cccc[ARRAY_SIZE], ccccc[ARRAY_SIZE];
struct timeval start_time, end_time;

main()
{
   int i, j;
   initialize_data();

/*****************************************************************************/
/*				Untuned Loop			             */
/*****************************************************************************/
   gettimeofday(&start_time, (struct timeval*)0);

   for(i=0;i<ARRAY_SIZE;i++)
      for(j=0;j<ARRAY_SIZE;j++)
         a = a + b[j] * c[i];

   gettimeofday(&end_time, (struct timeval*)0);  /* after time */
   printf("Execution time for the uncollapsed loop:\n");
   print_execution_times();

/*****************************************************************************/
/*				Tuned Loop - depth 2                         */
/*****************************************************************************/
   gettimeofday(&start_time, (struct timeval*)0);

   a1 = a2 = 0.0;
   for(i=0;i<ARRAY_SIZE;i++)
      for(j=0;j<ARRAY_SIZE;j+=2)
      {
         a1 = a1 + bb[j] * cc[i];
         a2 = a2 + bb[j+1] * cc[i];
      }
   aa = a1 + a2;

   gettimeofday(&end_time, (struct timeval*)0);  /* after time */
   printf("Execution time for the collapsed loop:\n");
   print_execution_times();
/*****************************************************************************/
/*				Tuned Loop - depth 4                         */
/*****************************************************************************/
   gettimeofday(&start_time, (struct timeval*)0);

   a1 = a2 = a3 = a4 = 0.0;
   for(i=0;i<ARRAY_SIZE;i++)
      for(j=0;j<ARRAY_SIZE;j+=4)
      {
         a1 = a1 + bbb[j]   * ccc[i];
         a2 = a2 + bbb[j+1] * ccc[i];
         a3 = a3 + bbb[j+2] * ccc[i];
         a4 = a4 + bbb[j+3] * ccc[i];
      }
   aa = a1 + a2 + a3 + a4;

   gettimeofday(&end_time, (struct timeval*)0);  /* after time */
   printf("Execution time for the collapsed loop:\n");
   print_execution_times();
/*****************************************************************************/
/*				Tuned Loop - depth 8                         */
/*****************************************************************************/
   gettimeofday(&start_time, (struct timeval*)0);

   a1 = a2 = a3 = a4 = a5 = a6 = a7 = a8 = 0.0;
   for(i=0;i<ARRAY_SIZE;i++)
      for(j=0;j<ARRAY_SIZE;j+=8)
      {
         a1 = a1 + bbbb[j] * cccc[i];
         a2 = a2 + bbbb[j+1] * cccc[i];
         a3 = a3 + bbbb[j+2] * cccc[i];
         a4 = a4 + bbbb[j+3] * cccc[i];
         a5 = a5 + bbbb[j+4] * cccc[i];
         a6 = a6 + bbbb[j+5] * cccc[i];
         a7 = a7 + bbbb[j+6] * cccc[i];
         a8 = a8 + bbbb[j+7] * cccc[i];
      }
   aa = a1 + a2 + a3 + a4 + a5 + a6 + a7 + a8;

   gettimeofday(&end_time, (struct timeval*)0);  /* after time */
   printf("Execution time for the collapsed loop:\n");
   print_execution_times();
/*****************************************************************************/
/*				Tuned Loop - depth 16                        */
/*****************************************************************************/
   gettimeofday(&start_time, (struct timeval*)0);

   a1 = a2 = a3 = a4 = a5 = a6 = a7 = a8 = 0.0;
   a9 = a10 = a11 = a12 = a13 = a14 = a15 = a16 = 0.0;
   for(i=0;i<ARRAY_SIZE;i++)
      for(j=0;j<ARRAY_SIZE;j+=16)
      {
         a1 = a1 + bbbbb[j] * ccccc[i];
         a2 = a2 + bbbbb[j+1] * ccccc[i];
         a3 = a3 + bbbbb[j+2] * ccccc[i];
         a4 = a4 + bbbbb[j+3] * ccccc[i];
         a5 = a5 + bbbbb[j+4] * ccccc[i];
         a6 = a6 + bbbbb[j+5] * ccccc[i];
         a7 = a7 + bbbbb[j+6] * ccccc[i];
         a8 = a8 + bbbbb[j+7] * ccccc[i];
         a9 = a9 + bbbbb[j+8] * ccccc[i];
         a10 = a10 + bbbbb[j+9] * ccccc[i];
         a11 = a11 + bbbbb[j+10] * ccccc[i];
         a12 = a12 + bbbbb[j+11] * ccccc[i];
         a13 = a13 + bbbbb[j+12] * ccccc[i];
         a14 = a14 + bbbbb[j+13] * ccccc[i];
         a15 = a15 + bbbbb[j+14] * ccccc[i];
         a16 = a16 + bbbbb[j+15] * ccccc[i];
      }
   aa = a1 + a2 + a3 + a4 + a5 + a6 + a7 + a8;
   aa+= a9 + a10 + a11 + a12 + a13 + a14 + a15 + a16;

   gettimeofday(&end_time, (struct timeval*)0);  /* after time */
   printf("Execution time for the collapsed loop:\n");
   print_execution_times();
   finalize_data();
}

initialize_data()
{
   int i, j, k;
   srandom(SEED);
   for(i=0;i<ARRAY_SIZE;i++)
   {
      b[i] = bb[i] = bbb[i] = bbbb[i] = bbbbb[i] = (float) (random() % 100000);
      c[i] = cc[i] = ccc[i] = cccc[i] = ccccc[i] = (float) (random() % 100000);
   }
   a = aa = aaa = aaaa = aaaaa = 0.0;
}

finalize_data()
{
   int i, j;
   FILE *fp, *fopen();

   fp = fopen("/dev/null", "w");
   fprintf(fp, "%f\n", a);
   fprintf(fp, "%f\n", aa);
   fprintf(fp, "%f\n", aaa);
   fprintf(fp, "%f\n", aaaa);
   fprintf(fp, "%f\n", aaaaa);
   fclose(fp);
}

print_execution_times()
{
   int total_usecs;
   float total_time;
   total_usecs = (end_time.tv_sec-start_time.tv_sec) * 1000000 +
         (end_time.tv_usec-start_time.tv_usec);
   printf("Total time was %.2f mSec\n ", ((float) total_usecs) / 1000.0);
   total_time = ((float) total_usecs) / 1000000.0;
}