#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <math.h>
#include <time.h>

#include <emmintrin.h> //sse2

//macro to print a __m128 vector
#define PVEC(a)						\
  {							\
    float tp[4];					\
    _mm_storeu_ps ( tp ,a);				\
    printf("VEC %7.1f %7.1f %7.1f %7.1f \n",		\
	   tp[0],					\
	   tp[1],					\
	   tp[2],					\
	   tp[3]);					\
  }							\
    

main(int argc, char** argv)
{


  __m128 vec1,vec2; //vectors containing 4 32-bits floats

  float a,b,c,d;
  a=12.0; b=13.0; c=15.0; d=29.0;


  vec1 = _mm_set_ps( a,b,c,d);    //  1) set the  value in the vector
  PVEC(vec1); // print the vector with the macro above

  vec2 = _mm_set1_ps( 10); //  2)set four times the same value in  second vector


  vec1 =  _mm_add_ps(vec1,vec2); // 3)  add the two vectors together

  printf(" + \n");
  PVEC(vec2);
  printf(" = \n");
  PVEC(vec1); // see result

  vec1 =  _mm_mul_ps(vec1,vec2); // 4) multiply the two vectors
  printf(" x \n");
  PVEC(vec2);
  printf(" = \n");

  PVEC(vec1); // see result

  float res[4];

  _mm_storeu_ps ( res ,vec1);// 5) store back result

  printf("%7.1f %7.1f %7.1f %7.1f \n",res[0],res[1],res[2],res[3]);


}