#ifndef _SSE_H
#define _SSE_H
typedef REAL vReal __attribute__((mode(V4SF),aligned(16)));
typedef struct {
    REAL re, im;
} complex;

typedef struct {
    vReal re, im;
} vcomplex;
typedef struct SU3 {
    complex v[3][3];
} SU3;

typedef struct {
    vcomplex v[3][3];
} vSU3;
typedef struct {
    vcomplex f[4][3];
} vFermion;

typedef struct {
    vcomplex f[2][3];
} vHalfFermion;
typedef struct {
    vFermion f;
} vEvenFermion;

typedef struct {
    vFermion f;
} vOddFermion;
static inline vReal vmk1(REAL a) {
     vReal v = __builtin_ia32_loadss((float *)&a);
     asm("shufps\t$0,%0,%0" : "+x" (v));
     return v;
}
static inline vReal vmk4(REAL a0, REAL a1, REAL a2, REAL a3) {
  vReal v;
  REAL *r = (REAL *)&v;
  r[0] = a0;
  r[1] = a1;
  r[2] = a2;
  r[3] = a3;
  return v;
}
static inline REAL vsum(vReal v)
{
   REAL *vv = (REAL *)&v;
   return vv[0] + vv[1] + vv[2] + vv[3];
}
static inline void vput_3(vReal *v, REAL a3)
{
   ((REAL *)v)[3] = a3;
}

static inline void vput_0(vReal *v, REAL a0)
{
   ((REAL *)v)[0] = a0;
}
static inline vReal shift_up1(vReal a, vReal b)
{
   vReal x = a;
   vReal y = b;
   asm("shufps\t$0x30,%0,%1\n\t"
       "shufps\t$0x29,%1,%0"
       : "+x" (x), "+x" (y));
   return x;
}
static inline vReal shift_up2(vReal a, vReal b)
{
   vReal x = a;
   asm("shufps\t$0x4e,%1,%0"
       : "+x" (x): "x" (b));
   return x;
}
static inline vReal shift_up3(vReal a, vReal b)
{
   vReal x = a;
   asm("shufps\t$0x03,%1,%0\n\t"
       "shufps\t$0x9c,%1,%0"
       : "+x" (x): "x" (b));
   return x;
}
static inline vReal shift_down1(vReal a, vReal b)
{
   return shift_up3(a, b);
}
static inline vReal shift_down2(vReal a, vReal b)
{
   return shift_up2(a, b);
}
static inline vReal shift_down3(vReal a, vReal b)
{
   return shift_up1(a, b);
}
static inline void vhfzero(vHalfFermion *v)
{
   vReal z = vmk1(0.0);

   v->f[0][0].re = v->f[0][0].im = 
   v->f[0][1].re = v->f[0][1].im = 
   v->f[0][2].re = v->f[0][2].im = 
   v->f[1][0].re = v->f[1][0].im = 
   v->f[1][1].re = v->f[1][1].im = 
   v->f[1][2].re = v->f[1][2].im = z;
}
#endif