194 

   195 static const __m128 zeros = _mm_set_ps1(0.);

   196 static const __m128 ones = _mm_set_ps1(1.);

   197 static const __m128 mEps = _mm_set_ps1(Eps);

   198 

   199 class VectorPacket

   200 {

   201 public:

   202 	union {

   203 		__m128 ma[3];

   204 		struct {

   205 			__m128 mx;

   206 			__m128 my;

   207 			__m128 mz;

   208 		};

   209 		struct {

   210 			float x[4];

   211 			float y[4];

   212 			float z[4];

   213 		};

   214 	};

   215 

   216 	VectorPacket() {};

   217 	VectorPacket(__m128 ax, __m128 ay, __m128 az):

   218 		mx(ax), my(ay), mz(az) {};

   219 

   220 	Vector3 getVector(int i) const

   221 	{

   222 		return Vector3(x[i], y[i], z[i]);

   223 	};

   224 

   225 	void normalize()

   226 	{

   227 		__m128 m,x,y,z;

   228 		x = _mm_mul_ps(mx, mx); // x*x

   229 		y = _mm_mul_ps(my, my); // y*y

   230 		z = _mm_mul_ps(mz, mz); // z*z

   231 		m = _mm_add_ps(x, y);

   232 		m = _mm_add_ps(m, z);     // x*x + y*y + z*z

   233 		m = _mm_sqrt_ps(m);

   234 		m = _mm_div_ps(ones, m);   // m = 1/sqrt(m)

   235 		mx = _mm_mul_ps(mx, m);

   236 		my = _mm_mul_ps(my, m);

   237 		mz = _mm_mul_ps(mz, m);

   238 	};

   239 

   240 	friend VectorPacket operator+(const VectorPacket &a, const VectorPacket &b)

   241 	{

   242 		return VectorPacket(

   243 			_mm_add_ps(a.mx, b.mx),

   244 			_mm_add_ps(a.my, b.my),

   245 			_mm_add_ps(a.mz, b.mz));

   246 	};

   247 

   248 	friend VectorPacket operator-(const VectorPacket &a, const VectorPacket &b)

   249 	{

   250 		return VectorPacket(

   251 			_mm_sub_ps(a.mx, b.mx),

   252 			_mm_sub_ps(a.my, b.my),

   253 			_mm_sub_ps(a.mz, b.mz));

   254 	};

   255 

   256 	friend VectorPacket operator*(const VectorPacket &v,  const __m128 &m)

   257 	{

   258 		return VectorPacket(

   259 			_mm_mul_ps(v.mx, m),

   260 			_mm_mul_ps(v.my, m),

   261 			_mm_mul_ps(v.mz, m));

   262 	};

   263 

   264 	friend VectorPacket operator/(const __m128 &m, const VectorPacket &v)

   265 	{

   266 		return VectorPacket(

   267 			_mm_div_ps(m, v.mx),

   268 			_mm_div_ps(m, v.my),

   269 			_mm_div_ps(m, v.mz));

   270 	};

   271 

   272 	// write to character stream

   273 	friend ostream & operator<<(ostream &st, const VectorPacket &v)

   274 	{

   275 		return st << "[" << v.getVector(0) << "," << v.getVector(1) 

   276 			<< "," << v.getVector(2) << "," << v.getVector(3) << ")";

   277 	};

   278 

   279 };

   280