--- a/include/scene.h	Thu Apr 24 18:12:32 2008 +0200
+++ b/include/scene.h	Sun Apr 27 09:44:49 2008 +0200
@@ -48,6 +48,23 @@
 };
 
 /**
+ * packet of 4 rays
+ */
+class RayPacket
+{
+public:
+	VectorPacket o, dir;
+
+	// index operator - get a ray
+	Ray operator[](int i) const
+	{
+		return Ray(
+			Vector3(o.x[i], o.y[i], o.z[i]),
+			Vector3(dir.x[i], dir.y[i], dir.z[i]));
+	};
+};
+
+/**
  * a camera
  */
 class Camera
@@ -78,65 +95,55 @@
 		return Ray(eye, dir);
 	};
 
-	void makeRayPacket(Sample *samples, Ray *rays)
+	void makeRayPacket(Sample *samples, RayPacket &rays)
 	{
 		__m128 m1x,m1y,m1z;
 		__m128 m2x,m2y,m2z;
 		__m128 m;
-		
+
 		// m1(xyz) = u * samples[i].x
-		m1x = _mm_set1_ps(u.x);
-		m1y = _mm_set1_ps(u.y);
-		m1z = _mm_set1_ps(u.z);
-		m = _mm_set_ps(samples[0].x, samples[1].x, samples[2].x, samples[3].x);
+		m1x = _mm_set_ps1(u.x);
+		m1y = _mm_set_ps1(u.y);
+		m1z = _mm_set_ps1(u.z);
+		m = _mm_set_ps(samples[3].x, samples[2].x, samples[1].x, samples[0].x);
 		m1x = _mm_mul_ps(m1x, m);
 		m1y = _mm_mul_ps(m1y, m);
 		m1z = _mm_mul_ps(m1z, m);
-		
+
 		// m2(xyz) = v * samples[i].y
-		m2x = _mm_set1_ps(v.x);
-		m2y = _mm_set1_ps(v.y);
-		m2z = _mm_set1_ps(v.z);
-		m = _mm_set_ps(samples[0].y, samples[1].y, samples[2].y, samples[3].y);
+		m2x = _mm_set_ps1(v.x);
+		m2y = _mm_set_ps1(v.y);
+		m2z = _mm_set_ps1(v.z);
+		m = _mm_set_ps(samples[3].y, samples[2].y, samples[1].y, samples[0].y);
 		m2x = _mm_mul_ps(m2x, m);
 		m2y = _mm_mul_ps(m2y, m);
 		m2z = _mm_mul_ps(m2z, m);
-		
+
 		// m1(xyz) = (m1 + m2) = (u*samples[i].x + v*samples[i].y)
 		m1x = _mm_add_ps(m1x, m2x);
 		m1y = _mm_add_ps(m1y, m2y);
 		m1z = _mm_add_ps(m1z, m2z);
-		
+
 		// m1(xyz) = m1*F = (u*samples[i].x + v*samples[i].y)*F
-		m = _mm_set_ps(F,F,F,F);
+		m = _mm_set_ps1(F);
 		m1x = _mm_mul_ps(m1x, m);
 		m1y = _mm_mul_ps(m1y, m);
 		m1z = _mm_mul_ps(m1z, m);
-		
+
 		// m1(xyz) = p - m1 = p - (u*samples[i].x + v*samples[i].y)*F = dir
-		m2x = _mm_set1_ps(p.x);
-		m2y = _mm_set1_ps(p.y);
-		m2z = _mm_set1_ps(p.z);
-		m2x = _mm_sub_ps(m2x, m1x);
-		m2y = _mm_sub_ps(m2y, m1y);
-		m2z = _mm_sub_ps(m2z, m1z);
-		
-		// normalize dir
-		m1x = _mm_mul_ps(m2x, m2x); // x*x
-		m1y = _mm_mul_ps(m2y, m2y); // y*y
-		m1z = _mm_mul_ps(m2z, m2z); // z*z
-		m = _mm_add_ps(m1x, m1y);   // x*x + y*y
-		m = _mm_add_ps(m, m1z);     // m = x*x + y*y + z*z
-		m = _mm_sqrt_ps(m);         // m = sqrt(m)
-		m2x = _mm_div_ps(m2x, m);   // dir(xyz) /= m
-		m2y = _mm_div_ps(m2y, m);
-		m2z = _mm_div_ps(m2z, m);
-		
-		for (int i = 0; i < 4; i++)
-		{
-			Vector3 dir(((float*)&m2x)[3-i], ((float*)&m2y)[3-i], ((float*)&m2z)[3-i]);
-			rays[i] = Ray(eye, dir);
-		}
+		m2x = _mm_set_ps1(p.x);
+		m2y = _mm_set_ps1(p.y);
+		m2z = _mm_set_ps1(p.z);
+		rays.dir.mx = _mm_sub_ps(m2x, m1x);
+		rays.dir.my = _mm_sub_ps(m2y, m1y);
+		rays.dir.mz = _mm_sub_ps(m2z, m1z);
+
+		// copy origin
+		rays.o.mx = _mm_set_ps1(eye.x);
+		rays.o.my = _mm_set_ps1(eye.y);
+		rays.o.mz = _mm_set_ps1(eye.z);
+
+		rays.dir.normalize();
 	};
 };
 
@@ -171,6 +178,13 @@
 	Float h() { return H.y-L.y; };
 	Float d() { return H.z-L.z; };
 	bool intersect(const Ray &ray, Float &a, Float &b);
+	bool intersect_packet(const RayPacket &rays, __m128 &a, __m128 &b)
+	{
+		return intersect(rays[0], ((float*)&a)[0], ((float*)&b)[0])
+			|| intersect(rays[1], ((float*)&a)[1], ((float*)&b)[1])
+			|| intersect(rays[2], ((float*)&a)[2], ((float*)&b)[2])
+			|| intersect(rays[3], ((float*)&a)[3], ((float*)&b)[3]);
+	};
 };
 
 #endif