/* compiled with -msse2, etc*/

//#include <mmintrin.h>  /* MMX */
//#include <xmmintrin.h> /* SSE */
#include <emmintrin.h> /* SSE2 */
//#include <pmmintrin.h> /* SSE3 */
//#include <tmmintrin.h> /* SSSE3 */
//#include <smmintrin.h> /* SSE4.2 SSE4.1 */

#include<sys/time.h>
#include<time.h>
#include<stdio.h>
#include <stdint.h>

typedef uint64_t ticks;

static __inline__ ticks ticks_start(void)
{
	unsigned a, d;
	// Serialize + read TSC
	asm volatile(
		             "cpuid\n\t"
			     "rdtsc\n\t"
			     : "=a"(a), "=d"(d)
			     : "a"(0)
			     : "rbx", "rcx", "memory"
		     );
	return ((ticks)d << 32) | a;
}

static __inline__ ticks ticks_end(void)
{
	unsigned a, d;
	// Read TSC + serialize (rdtscp is partially serializing; cpuid completes)
	asm volatile(
		             "rdtscp\n\t"
			     : "=a"(a), "=d"(d)
			     :
			     : "rcx", "memory"
		     );
	asm volatile(
		             "cpuid\n\t"
			     :
			     : "a"(0)
			     : "rbx", "rcx", "rdx", "memory"
		     );
	return ((ticks)d << 32) | a;
}


/*
typedef unsigned long long ticks;

static __inline__ ticks getticks(void)
{
     unsigned a, d;
     asm("cpuid");
     asm volatile("rdtsc" : "=a" (a), "=d" (d));

     return (((ticks)a) | (((ticks)d) << 32));
}
*/

int main()
{
  unsigned long long int t1, t2, t3;

  float z1[] = {1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0};
  float z2[] = {1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0};
  float z3[8] = {0.0};
  float z4[8];

  int i, j;

  t1 = ticks_start();
  
  for (j=0;j<100; j++) {
	  for(i=0; i<8; i++) {
		  z3[i] += z1[i] + z2[i];
	  }
  }
  t2 = ticks_end();
  printf("time = %lld\n", t2-t1);

  t1 = ticks_start();
  for (j=0; j<100; j++) {
	  /*
	  __m128 a0 = _mm_loadu_ps(&z1[0]);
	  __m128 b0 = _mm_loadu_ps(&z2[0]);
	  __m128 c0 = _mm_add_ps(a0, b0);
	  _mm_storeu_ps(&z3[0], c0);
	  _mm_storeu_ps(&z4[0], _mm_add_ps(_mm_loadu_ps(&z4[0]), c0));

	  __m128 a1 = _mm_loadu_ps(&z1[4]);
	  __m128 b1 = _mm_loadu_ps(&z2[4]);
	  __m128 c1 = _mm_add_ps(a1, b1);
	  _mm_storeu_ps(&z3[4], c1);
	  _mm_storeu_ps(&z4[4], _mm_add_ps(_mm_loadu_ps(&z4[4]), c1));
	  */
	  
	  __m128 *v_z1 = (__m128 *)z1;
	  __m128 *v_z2 = (__m128 *)z2;
	  __m128 *v_z3 = (__m128 *)z4;
	  __m128 *v_z4 = (__m128 *)z3;

	  *v_z4 = _mm_add_ps(*v_z1, *v_z2);
	  *v_z3 = _mm_add_ps(*v_z3, *v_z4);
	  v_z1++;
	  v_z2++;
	  v_z3++;
	  *v_z4 = _mm_add_ps(*v_z1, *v_z2);
	  *v_z3 = _mm_add_ps(*v_z3, *v_z4);
	  
  }
  t2 = ticks_end();

  printf("time = %lld\n", t2-t1);
  for (i=0; i<8; i++) {
    //    if (z3[i] != z4[i]) {printf("Wrong.\n"); exit(0);}
    printf("%d %f %f\n", i, z3[i], z4[i]);
  }
  printf("Correct, results matched.\n");
  return 0;
}