--- a/src/scene.cc	Fri May 02 13:27:47 2008 +0200
+++ b/src/scene.cc	Mon May 05 15:31:14 2008 +0200
@@ -73,7 +73,7 @@
 }
 
 /* http://www.siggraph.org/education/materials/HyperGraph/raytrace/rtinter3.htm */
-bool BBox::intersect(const Ray &ray, Float &a, Float &b)
+bool BBox::intersect(const Ray &ray, Float &a, Float &b) const
 {
 	register Float tnear = -Inf;
 	register Float tfar = Inf;
@@ -108,46 +108,33 @@
 	return true;
 }
 
-#ifndef NO_SSE
+#ifndef NO_SIMD
 // rewrite of BBox::intersect for ray packets
-__m128 BBox::intersect_packet(const RayPacket &rays, __m128 &a, __m128 &b)
+mfloat4 BBox::intersect_packet(const RayPacket &rays, mfloat4 &a, mfloat4 &b) const
 {
-	register __m128 tnear = mZero;
-	register __m128 tfar = mInf;
-	register __m128 t1, t2;
-	register __m128 mask = mAllSet;
-
-	for (int i = 0; i < 3; i++)
-	{
-		const __m128 mL = _mm_set_ps1(L[i]);
-		const __m128 mH = _mm_set_ps1(H[i]);
-		mask = _mm_and_ps(mask,
-		_mm_or_ps(
-		_mm_or_ps(_mm_cmplt_ps(rays.dir.ma[i], mMEps), _mm_cmpgt_ps(rays.dir.ma[i], mEps)),
-		_mm_and_ps(_mm_cmpge_ps(rays.o.ma[i], mL), _mm_cmple_ps(rays.o.ma[i], mH))
-		));
-		if (!_mm_movemask_ps(mask))
-			return mask;
+	mfloat4 origin = rays.o.ma[0];
+	mfloat4 invdir = rays.invdir.ma[0];
+	mfloat4 t1 = mmul(msub(mset1(L[0]), origin), invdir);
+	mfloat4 t2 = mmul(msub(mset1(H[0]), origin), invdir);
+	mfloat4 tmin = mmin(t1, t2);
+	mfloat4 tmax = mmax(t1, t2);
 
-		/* compute the intersection distance of the planes */
-		t1 = _mm_div_ps(_mm_sub_ps(mL, rays.o.ma[i]), rays.dir.ma[i]);
-		t2 = _mm_div_ps(_mm_sub_ps(mH, rays.o.ma[i]), rays.dir.ma[i]);
-
-		__m128 t = _mm_min_ps(t1, t2);
-		t2 = _mm_max_ps(t1, t2);
-		t1 = t;
+	origin = rays.o.ma[1];
+	invdir = rays.invdir.ma[1];
+	t1 = mmul(msub(mset1(L[1]), origin), invdir);
+	t2 = mmul(msub(mset1(H[1]), origin), invdir);
+	tmin = mmax(mmin(t1, t2), tmin);
+	tmax = mmin(mmax(t1, t2), tmax);
 
-		tnear = _mm_max_ps(tnear, t1);	/* want largest Tnear */
-		tfar = _mm_min_ps(tfar, t2);	/* want smallest Tfar */
+	origin = rays.o.ma[2];
+	invdir = rays.invdir.ma[2];
+	t1 = mmul(msub(mset1(L[2]), origin), invdir);
+	t2 = mmul(msub(mset1(H[2]), origin), invdir);
+	tmin = mmax(mmin(t1, t2), tmin);
+	tmax = mmin(mmax(t1, t2), tmax);
 
-		mask = _mm_and_ps(mask,
-			_mm_and_ps(_mm_cmple_ps(tnear, tfar), _mm_cmpge_ps(tfar, mZero)));
-		if (!_mm_movemask_ps(mask))
-			return mask;
-	}
-
-	a = tnear;
-	b = tfar;
-	return mask;
+	a = tmin;
+	b = tmax;
+	return mand(mcmplt(tmin, tmax), mcmpgt(tmax, mZero));
 }
 #endif