26 files changed, 2330 insertions, 35 deletions

diff --git a/src/audio/oal/stream.cpp b/src/audio/oal/stream.cpp
index f2a9eb22..23a0a0c9 100644
--- a/src/audio/oal/stream.cpp
+++ b/src/audio/oal/stream.cpp
@@ -8,13 +8,8 @@
 #include <opusfile.h>
 #else
 #ifdef _WIN32
-
-// TODO: This is due to version difference of 32-bit libmpg123 and 64-bit libmpg123, fix it
-#ifndef _WIN64
-typedef long ssize_t;
-#endif
 #pragma comment( lib, "libsndfile-1.lib" )
-#pragma comment( lib, "libmpg123.lib" )
+#pragma comment( lib, "libmpg123-0.lib" )
 #else
 #include "crossplatform.h"
 #endif
diff --git a/src/control/RoadBlocks.cpp b/src/control/RoadBlocks.cpp
index 010dec3f..e8630bde 100644
--- a/src/control/RoadBlocks.cpp
+++ b/src/control/RoadBlocks.cpp
@@ -111,6 +111,10 @@ CRoadBlocks::GenerateRoadBlocks(void)
 { 
 	CMatrix tmp1, tmp2;
 	static int16 unk;
+#ifdef SQUEEZE_PERFORMANCE
+	if (FindPlayerPed()->m_pWanted->m_RoadblockDensity == 0)
+		return;
+#endif
 	uint32 frame = CTimer::GetFrameCounter() & 0xF;
 	int16 nRoadblockNode = (int16)(NUMROADBLOCKS * frame) / 16;
 	const int16 maxRoadBlocks = (int16)(NUMROADBLOCKS * (frame + 1)) / 16;
diff --git a/src/core/Accident.cpp b/src/core/Accident.cpp
index 1fd6c123..c8611323 100644
--- a/src/core/Accident.cpp
+++ b/src/core/Accident.cpp
@@ -53,6 +53,10 @@ CAccidentManager::ReportAccident(CPed *ped)
 void
 CAccidentManager::Update()
 {
+#ifdef SQUEEZE_PERFORMANCE
+	// Handled after injury registered.
+	return;
+#endif
 	int32 e;
 	if (CEventList::GetEvent(EVENT_INJURED_PED, &e)) {
 		CPed *ped = CPools::GetPed(gaEvent[e].entityRef);
diff --git a/src/core/Accident.h b/src/core/Accident.h
index 949d5fb9..568e1149 100644
--- a/src/core/Accident.h
+++ b/src/core/Accident.h
@@ -1,5 +1,4 @@
 #pragma once
-#include "common.h"
 #include "config.h"
 
 class CPed;
diff --git a/src/core/Collision.cpp b/src/core/Collision.cpp
index 268e3c5e..c2037dc5 100644
--- a/src/core/Collision.cpp
+++ b/src/core/Collision.cpp
@@ -1,5 +1,6 @@
 #include "common.h"
 
+#include "VuVector.h"
 #include "main.h"
 #include "Lists.h"
 #include "Game.h"
@@ -25,6 +26,346 @@
 
 //--MIAMI: file done
 
+
+// TODO: where do these go?
+
+#ifdef VU_COLLISION
+
+struct VuTriangle
+{
+	// Compressed int16 but unpacked
+#ifdef GTA_PS2
+	uint128 v0;
+	uint128 v1;
+	uint128 v2;
+	uint128 plane;
+#else
+	int32 v0[4];
+	int32 v1[4];
+	int32 v2[4];
+	int32 plane[4];
+#endif
+};
+
+#ifndef GTA_PS2
+static int16 vi01;
+static CVuVector vf01;
+static CVuVector vf02;
+static CVuVector vf03;
+
+CVuVector
+DistanceBetweenSphereAndLine(const CVuVector &center, const CVuVector &p0, const CVuVector &line)
+{
+	// center  VF12
+	// p0      VF14
+	// line    VF15
+	CVuVector ret;	// VF16
+	CVuVector p1 = p0+line;
+	CVuVector dist0 = center - p0;	// VF20
+	CVuVector dist1 = center - p1;	// VF25
+	float lenSq = line.MagnitudeSqr();	// VF21
+	float distSq0 = dist0.MagnitudeSqr();	// VF22
+	float distSq1 = dist1.MagnitudeSqr();
+	float dot = DotProduct(dist0, line);	// VF23
+	if(dot < 0.0f){
+		// not above line, closest to p0
+		ret = p0;
+		ret.w = distSq0;
+		return ret;
+	}
+	float t = dot/lenSq;	// param of nearest point on infinite line
+	if(t > 1.0f){
+		// not above line, closest to p1
+		ret = p1;
+		ret.w = distSq1;
+		return ret;
+	}
+	// closest to line
+	ret = p0 + line*t;
+	ret.w = (ret - center).MagnitudeSqr();
+	return ret;
+}
+inline int SignFlags(const CVector &v)
+{
+	int f = 0;
+	if(v.x < 0.0f) f |= 1;
+	if(v.y < 0.0f) f |= 2;
+	if(v.z < 0.0f) f |= 4;
+	return f;
+}
+#endif
+
+extern "C" void
+LineToTriangleCollision(const CVuVector &p0, const CVuVector &p1,
+	const CVuVector &v0, const CVuVector &v1, const CVuVector &v2,
+	const CVuVector &plane)
+{
+#ifdef GTA_PS2
+	__asm__ volatile (
+		".set noreorder\n"
+		"lqc2	vf12, 0x0(%0)\n"
+		"lqc2	vf13, 0x0(%1)\n"
+		"lqc2	vf14, 0x0(%2)\n"
+		"lqc2	vf15, 0x0(%3)\n"
+		"lqc2	vf16, 0x0(%4)\n"
+		"lqc2	vf17, 0x0(%5)\n"
+		"vcallms	Vu0LineToTriangleCollisionStart\n"
+		".set reorder\n"
+		:
+		: "r" (&p0), "r" (&p1), "r" (&v0), "r" (&v1), "r" (&v2), "r" (&plane)
+	);
+#else
+	float dot0 = DotProduct(plane, p0);
+	float dot1 = DotProduct(plane, p1);
+	float dist0 = plane.w - dot0;
+	float dist1 = plane.w - dot1;
+
+	// if points are on the same side, no collision
+	if(dist0 * dist1 > 0.0f){
+		vi01 = 0;
+		return;
+	}
+
+	CVuVector diff = p1 - p0;
+	float t = dist0/(dot1 - dot0);
+	CVuVector p = p0 + diff*t;
+	p.w = 0.0f;
+	vf01 = p;
+	vf03.x = t;
+
+	// Check if point is inside
+	CVector cross1 = CrossProduct(p-v0, v1-v0);
+	CVector cross2 = CrossProduct(p-v1, v2-v1);
+	CVector cross3 = CrossProduct(p-v2, v0-v2);
+	// Only check relevant directions
+	int flagmask = 0;
+	if(Abs(plane.x) > 0.5f) flagmask |= 1;
+	if(Abs(plane.y) > 0.5f) flagmask |= 2;
+	if(Abs(plane.z) > 0.5f) flagmask |= 4;
+	int flags1 = SignFlags(cross1) & flagmask;
+	int flags2 = SignFlags(cross2) & flagmask;
+	int flags3 = SignFlags(cross3) & flagmask;
+	// inside if on the same side of all edges
+	if(flags1 != flags2 || flags1 != flags3){
+		vi01 = 0;
+		return;
+	}
+	vi01 = 1;
+	vf02 = plane;
+	return;
+#endif
+}
+
+extern "C" void
+LineToTriangleCollisionCompressed(const CVuVector &p0, const CVuVector &p1, VuTriangle &tri)
+{
+#ifdef GTA_PS2
+	__asm__ volatile (
+		".set noreorder\n"
+		"lqc2	vf12, 0x0(%0)\n"
+		"lqc2	vf13, 0x0(%1)\n"
+		"lqc2	vf14, 0x0(%2)\n"
+		"lqc2	vf15, 0x10(%2)\n"
+		"lqc2	vf16, 0x20(%2)\n"
+		"lqc2	vf17, 0x30(%2)\n"
+		"vcallms	Vu0LineToTriangleCollisionCompressedStart\n"
+		".set reorder\n"
+		:
+		: "r" (&p0), "r" (&p1), "r" (&tri)
+	);
+#else
+	CVuVector v0, v1, v2, plane;
+	v0.x = tri.v0[0]/128.0f;
+	v0.y = tri.v0[1]/128.0f;
+	v0.z = tri.v0[2]/128.0f;
+	v0.w = tri.v0[3]/128.0f;
+	v1.x = tri.v1[0]/128.0f;
+	v1.y = tri.v1[1]/128.0f;
+	v1.z = tri.v1[2]/128.0f;
+	v1.w = tri.v1[3]/128.0f;
+	v2.x = tri.v2[0]/128.0f;
+	v2.y = tri.v2[1]/128.0f;
+	v2.z = tri.v2[2]/128.0f;
+	v2.w = tri.v2[3]/128.0f;
+	plane.x = tri.plane[0]/4096.0f;
+	plane.y = tri.plane[1]/4096.0f;
+	plane.z = tri.plane[2]/4096.0f;
+	plane.w = tri.plane[3]/128.0f;
+	LineToTriangleCollision(p0, p1, v0, v1, v2, plane);
+#endif
+}
+
+extern "C" void
+SphereToTriangleCollision(const CVuVector &sph,
+	const CVuVector &v0, const CVuVector &v1, const CVuVector &v2,
+	const CVuVector &plane)
+{
+#ifdef GTA_PS2
+	__asm__ volatile (
+		".set noreorder\n"
+		"lqc2	vf12, 0x0(%0)\n"
+		"lqc2	vf14, 0x0(%1)\n"
+		"lqc2	vf15, 0x0(%2)\n"
+		"lqc2	vf16, 0x0(%3)\n"
+		"lqc2	vf17, 0x0(%4)\n"
+		"vcallms	Vu0SphereToTriangleCollisionStart\n"
+		".set reorder\n"
+		:
+		: "r" (&sph), "r" (&v0), "r" (&v1), "r" (&v2), "r" (&plane)
+	);
+#else
+	float planedist = DotProduct(plane, sph) - plane.w;	// VF02
+	if(Abs(planedist) > sph.w){
+		vi01 = 0;
+		return;
+	}
+	// point on plane
+	CVuVector p = sph - planedist*plane;
+	p.w = 0.0f;
+	vf01 = p;
+	planedist = Abs(planedist);
+	// edges
+	CVuVector v01 = v1 - v0;
+	CVuVector v12 = v2 - v1;
+	CVuVector v20 = v0 - v2;
+	// VU code calculates normal again for some weird reason...
+	// Check sides of point
+	CVector cross1 = CrossProduct(p-v0, v01);
+	CVector cross2 = CrossProduct(p-v1, v12);
+	CVector cross3 = CrossProduct(p-v2, v20);
+	// Only check relevant directions
+	int flagmask = 0;
+	if(Abs(plane.x) > 0.1f) flagmask |= 1;
+	if(Abs(plane.y) > 0.1f) flagmask |= 2;
+	if(Abs(plane.z) > 0.1f) flagmask |= 4;
+	int nflags = SignFlags(plane) & flagmask;
+	int flags1 = SignFlags(cross1) & flagmask;
+	int flags2 = SignFlags(cross2) & flagmask;
+	int flags3 = SignFlags(cross3) & flagmask;
+	int testcase = 0;
+	CVuVector closest(0.0f, 0.0f, 0.0f);	// VF04
+	if(flags1 == nflags){
+		closest += v2;
+		testcase++;
+	}
+	if(flags2 == nflags){
+		closest += v0;
+		testcase++;
+	}
+	if(flags3 == nflags){
+		closest += v1;
+		testcase++;
+	}
+	if(testcase == 3){
+		// inside triangle - dist to plane already checked
+		vf02 = plane;
+		vf02.w = vf03.x = planedist;
+		vi01 = 1;
+	}else if(testcase == 1){
+		// outside two sides - closest to point opposide inside edge
+		vf01 = closest;
+		vf02 = sph - closest;
+		float distSq = vf02.MagnitudeSqr();
+		vi01 = sph.w*sph.w > distSq;
+		vf03.x = Sqrt(distSq);
+		vf02 *= 1.0f/vf03.x;
+	}else{
+		// inside two sides - closest to third edge
+		if(flags1 != nflags)
+			closest = DistanceBetweenSphereAndLine(sph, v0, v01);
+		else if(flags2 != nflags)
+			closest = DistanceBetweenSphereAndLine(sph, v1, v12);
+		else
+			closest = DistanceBetweenSphereAndLine(sph, v2, v20);
+		vi01 = sph.w*sph.w > closest.w;
+		vf01 = closest;
+		vf02 = sph - closest;
+		vf03.x = Sqrt(closest.w);
+		vf02 *= 1.0f/vf03.x;
+	}
+#endif
+}
+
+extern "C" void
+SphereToTriangleCollisionCompressed(const CVuVector &sph, VuTriangle &tri)
+{
+#ifdef GTA_PS2
+	__asm__ volatile (
+		".set noreorder\n"
+		"lqc2	vf12, 0x0(%0)\n"
+		"lqc2	vf14, 0x0(%1)\n"
+		"lqc2	vf15, 0x10(%1)\n"
+		"lqc2	vf16, 0x20(%1)\n"
+		"lqc2	vf17, 0x30(%1)\n"
+		"vcallms	Vu0SphereToTriangleCollisionCompressedStart\n"
+		".set reorder\n"
+		:
+		: "r" (&sph), "r" (&tri)
+	);
+#else
+	CVuVector v0, v1, v2, plane;
+	v0.x = tri.v0[0]/128.0f;
+	v0.y = tri.v0[1]/128.0f;
+	v0.z = tri.v0[2]/128.0f;
+	v0.w = tri.v0[3]/128.0f;
+	v1.x = tri.v1[0]/128.0f;
+	v1.y = tri.v1[1]/128.0f;
+	v1.z = tri.v1[2]/128.0f;
+	v1.w = tri.v1[3]/128.0f;
+	v2.x = tri.v2[0]/128.0f;
+	v2.y = tri.v2[1]/128.0f;
+	v2.z = tri.v2[2]/128.0f;
+	v2.w = tri.v2[3]/128.0f;
+	plane.x = tri.plane[0]/4096.0f;
+	plane.y = tri.plane[1]/4096.0f;
+	plane.z = tri.plane[2]/4096.0f;
+	plane.w = tri.plane[3]/128.0f;
+	SphereToTriangleCollision(sph, v0, v1, v2, plane);
+#endif
+}
+
+inline int
+GetVUresult(void)
+{
+#ifdef GTA_PS2
+	int ret;
+	__asm__ volatile (
+		"cfc2.i	%0,vi01\n"	// .i important! wait for VU0 to finish
+		: "=r" (ret)
+	);
+	return ret;
+#else
+	return vi01;
+#endif
+}
+
+inline int
+GetVUresult(CVuVector &point, CVuVector &normal, float &dist)
+{
+#ifdef GTA_PS2
+	int ret;
+	__asm__ volatile (
+		"cfc2.i	%0,vi01\n"	// .i important! wait for VU0 to finish
+		"sqc2	vf01,(%1)\n"
+		"sqc2	vf02,(%2)\n"
+		"qmfc2	$12,vf03\n"
+		"sw	$12,(%3)\n"
+		: "=r" (ret)
+		: "r" (&point), "r" (&normal), "r" (&dist)
+		: "$12"
+	);
+	return ret;
+#else
+	point = vf01;
+	normal = vf02;
+	dist = vf03.x;
+	return vi01;
+#endif
+}
+
+#endif
+
+
 enum Direction
 {
 	DIR_X_POS,
@@ -252,6 +593,20 @@ CCollision::TestVerticalLineBox(const CColLine &line, const CBox &box)
 bool
 CCollision::TestLineTriangle(const CColLine &line, const CompressedVector *verts, const CColTriangle &tri, const CColTrianglePlane &plane)
 {
+#ifdef VU_COLLISION
+	// not used in favour of optimized loops
+	VuTriangle vutri;
+	verts[tri.a].Unpack(vutri.v0);
+	verts[tri.b].Unpack(vutri.v1);
+	verts[tri.c].Unpack(vutri.v2);
+	plane.Unpack(vutri.plane);
+
+	LineToTriangleCollisionCompressed(*(CVuVector*)&line.p0, *(CVuVector*)&line.p1, vutri);
+
+	if(GetVUresult())
+		return true;
+	return false;
+#else
 	float t;
 	CVector normal;
 	plane.GetNormal(normal);
@@ -325,6 +680,7 @@ CCollision::TestLineTriangle(const CColLine &line, const CompressedVector *verts
 	if(CrossProduct2D(vec3-vec1, vect-vec1) > 0.0f) return false;
 	if(CrossProduct2D(vec3-vec2, vect-vec2) < 0.0f) return false;
 	return true;
+#endif
 }
 
 // Test if line segment intersects with sphere.
@@ -362,6 +718,20 @@ bool
 CCollision::TestSphereTriangle(const CColSphere &sphere,
 	const CompressedVector *verts, const CColTriangle &tri, const CColTrianglePlane &plane)
 {
+#ifdef VU_COLLISION
+	// not used in favour of optimized loops
+	VuTriangle vutri;
+	verts[tri.a].Unpack(vutri.v0);
+	verts[tri.b].Unpack(vutri.v1);
+	verts[tri.c].Unpack(vutri.v2);
+	plane.Unpack(vutri.plane);
+
+	SphereToTriangleCollisionCompressed(*(CVuVector*)&sphere, vutri);
+
+	if(GetVUresult())
+		return true;
+	return false;
+#else
 	// If sphere and plane don't intersect, no collision
 	float planedist = plane.CalcPoint(sphere.center);
 	if(Abs(planedist) > sphere.radius)
@@ -375,7 +745,9 @@ CCollision::TestSphereTriangle(const CColSphere &sphere,
 	CVector vec2 = vb - va;
 	float len = vec2.Magnitude();
 	vec2 = vec2 * (1.0f/len);
-	CVector vec1 = CrossProduct(vec2, plane.normal);
+	CVector normal;
+	plane.GetNormal(normal);
+	CVector vec1 = CrossProduct(vec2, normal);
 
 	// We know A has local coordinate [0,0] and B has [0,len].
 	// Now calculate coordinates on triangle for these two vectors:
@@ -420,11 +792,78 @@ CCollision::TestSphereTriangle(const CColSphere &sphere,
 	}
 
 	return dist < sphere.radius;
+#endif
 }
 
 bool
 CCollision::TestLineOfSight(const CColLine &line, const CMatrix &matrix, CColModel &model, bool ignoreSeeThrough, bool ignoreShootThrough)
 {
+#ifdef VU_COLLISION
+	CMatrix matTransform;
+	int i;
+
+	// transform line to model space
+	Invert(matrix, matTransform);
+	CVuVector newline[2];
+	TransformPoints(newline, 2, matTransform, (RwV3d*)&line.p0, sizeof(CColLine)/2);
+
+	// If we don't intersect with the bounding box, no chance on the rest
+	if(!TestLineBox(*(CColLine*)newline, model.boundingBox))
+		return false;
+
+	for(i = 0; i < model.numSpheres; i++){
+		if(ignoreSeeThrough && IsSeeThrough(model.spheres[i].surface)) continue;
+		if(TestLineSphere(*(CColLine*)newline, model.spheres[i]))
+			return true;
+	}
+
+	for(i = 0; i < model.numBoxes; i++){
+		if(ignoreSeeThrough && IsSeeThrough(model.boxes[i].surface)) continue;
+		if(TestLineBox(*(CColLine*)newline, model.boxes[i]))
+			return true;
+	}
+
+	CalculateTrianglePlanes(&model);
+	int lastTest = -1;
+	VuTriangle vutri;
+	for(i = 0; i < model.numTriangles; i++){
+		if(ignoreSeeThrough && IsSeeThrough(model.triangles[i].surface)) continue;
+
+		CColTriangle *tri = &model.triangles[i];
+		model.vertices[tri->a].Unpack(vutri.v0);
+		model.vertices[tri->b].Unpack(vutri.v1);
+		model.vertices[tri->c].Unpack(vutri.v2);
+		model.trianglePlanes[i].Unpack(vutri.plane);
+
+		LineToTriangleCollisionCompressed(newline[0], newline[1], vutri);
+		lastTest = i;
+		break;
+	}
+#ifdef FIX_BUGS
+	// no need to check first again
+	i++;
+#endif
+	for(; i < model.numTriangles; i++){
+		if(ignoreSeeThrough && IsSeeThrough(model.triangles[i].surface)) continue;
+
+		CColTriangle *tri = &model.triangles[i];
+		model.vertices[tri->a].Unpack(vutri.v0);
+		model.vertices[tri->b].Unpack(vutri.v1);
+		model.vertices[tri->c].Unpack(vutri.v2);
+		model.trianglePlanes[i].Unpack(vutri.plane);
+
+		if(GetVUresult())
+			return true;
+
+		LineToTriangleCollisionCompressed(newline[0], newline[1], vutri);
+		lastTest = i;
+
+	}
+	if(lastTest != -1 && GetVUresult())
+		return true;
+
+	return false;
+#else
 	static CMatrix matTransform;
 	int i;
 
@@ -459,6 +898,7 @@ CCollision::TestLineOfSight(const CColLine &line, const CMatrix &matrix, CColMod
 	}
 
 	return false;
+#endif
 }
 
 // TODO: TestPillWithSpheresInColModel, but only called from overloaded CWeapon::FireMelee which isn't used
@@ -475,20 +915,24 @@ CCollision::ProcessSphereSphere(const CColSphere &s1, const CColSphere &s2, CCol
 {
 	CVector dist = s1.center - s2.center;
 	float d = dist.Magnitude() - s2.radius;	// distance from s1's center to s2
-	float dc = d < 0.0f ? 0.0f : d;		// clamp to zero, i.e. if s1's center is inside s2
+	float depth = s1.radius - d;	// sphere overlap
+	if(d < 0.0f) d = 0.0f;		// clamp to zero, i.e. if s1's center is inside s2
 	// no collision if sphere is not close enough
-	if(mindistsq <= dc*dc || s1.radius <= dc)
-		return false;
-	dist.Normalise();
-	point.point = s1.center - dist*dc;
-	point.normal = dist;
-	point.surfaceA = s1.surface;
-	point.pieceA = s1.piece;
-	point.surfaceB = s2.surface;
-	point.pieceB = s2.piece;
-	point.depth = s1.radius - d;	// sphere overlap
-	mindistsq = dc*dc;		// collision radius
-	return true;
+	if(d*d < mindistsq && d < s1.radius){
+		dist.Normalise();
+		point.point = s1.center - dist*d;
+		point.normal = dist;
+#ifndef VU_COLLISION
+		point.surfaceA = s1.surface;
+		point.pieceA = s1.piece;
+		point.surfaceB = s2.surface;
+		point.pieceB = s2.piece;
+#endif
+		point.depth = depth;	
+		mindistsq = d*d;		// collision radius
+		return true;
+	}
+	return false;
 }
 
 bool
@@ -528,10 +972,12 @@ CCollision::ProcessSphereBox(const CColSphere &sph, const CColBox &box, CColPoin
 		if(lensq < mindistsq){
 			point.normal = dist * (1.0f/Sqrt(lensq));
 			point.point = sph.center - point.normal;
+#ifndef VU_COLLISION
 			point.surfaceA = sph.surface;
 			point.pieceA = sph.piece;
 			point.surfaceB = box.surface;
 			point.pieceB = box.piece;
+#endif
 
 			// find absolute distance to the closer side in each dimension
 			float dx = dist.x > 0.0f ?
@@ -571,10 +1017,12 @@ CCollision::ProcessSphereBox(const CColSphere &sph, const CColBox &box, CColPoin
 			float len = Sqrt(lensq);
 			point.point = p;
 			point.normal = dist * (1.0f/len);
+#ifndef VU_COLLISION
 			point.surfaceA = sph.surface;
 			point.pieceA = sph.piece;
 			point.surfaceB = box.surface;
 			point.pieceB = box.piece;
+#endif
 			point.depth = sph.radius - len;
 			mindistsq = lensq;
 			return true;
@@ -689,10 +1137,12 @@ CCollision::ProcessLineBox(const CColLine &line, const CColBox &box, CColPoint &
 
 	point.point = p;
 	point.normal = normal;
+#ifndef VU_COLLISION
 	point.surfaceA = 0;
 	point.pieceA = 0;
 	point.surfaceB = box.surface;
 	point.pieceB = box.piece;
+#endif
 	mindist = mint;
 
 	return true;
@@ -722,10 +1172,12 @@ CCollision::ProcessLineSphere(const CColLine &line, const CColSphere &sphere, CC
 	point.point = line.p0 + v01*t;
 	point.normal = point.point - sphere.center;
 	point.normal.Normalise();
+#ifndef VU_COLLISION
 	point.surfaceA = 0;
 	point.pieceA = 0;
 	point.surfaceB = sphere.surface;
 	point.pieceB = sphere.piece;
+#endif
 	mindist = t;
 	return true;
 }
@@ -736,6 +1188,17 @@ CCollision::ProcessVerticalLineTriangle(const CColLine &line,
 	const CompressedVector *verts, const CColTriangle &tri, const CColTrianglePlane &plane,
 	CColPoint &point, float &mindist, CStoredCollPoly *poly)
 {
+#ifdef VU_COLLISION
+	// not used in favour of optimized loops
+	bool res = ProcessLineTriangle(line, verts, tri, plane, point, mindist);
+	if(res && poly){
+		poly->verts[0] = verts[tri.a].Get();
+		poly->verts[1] = verts[tri.b].Get();
+		poly->verts[2] = verts[tri.c].Get();
+		poly->valid = true;
+	}
+	return res;
+#else
 	float t;
 	CVector normal;
 
@@ -822,11 +1285,40 @@ CCollision::ProcessVerticalLineTriangle(const CColLine &line,
 	}
 	mindist = t;
 	return true;
+#endif
 }
 
 bool
 CCollision::IsStoredPolyStillValidVerticalLine(const CVector &pos, float z, CColPoint &point, CStoredCollPoly *poly)
 {
+#ifdef VU_COLLISION
+	if(!poly->valid)
+		return false;
+
+	CVuVector p0 = pos;
+	CVuVector p1 = pos;
+	p1.z = z;
+
+	CVector v01 = poly->verts[1] - poly->verts[0];
+	CVector v02 = poly->verts[2] - poly->verts[0];
+	CVuVector plane = CrossProduct(v02, v01);
+	plane.Normalise();
+	plane.w = DotProduct(plane, poly->verts[0]);
+
+	LineToTriangleCollision(p0, p1, poly->verts[0], poly->verts[1], poly->verts[2], plane);
+
+	CVuVector pnt;
+	float dist;
+	if(!GetVUresult(pnt, plane, dist))
+#ifdef FIX_BUGS
+		// perhaps not needed but be safe
+		return poly->valid = false;
+#else
+		return false;
+#endif
+	point.point = pnt;
+	return true;
+#else
 	float t;
 
 	if(!poly->valid)
@@ -849,7 +1341,9 @@ CCollision::IsStoredPolyStillValidVerticalLine(const CVector &pos, float z, CCol
 		return poly->valid = false;
 
 	// intersection parameter on line
-	t = -plane.CalcPoint(p0) / DotProduct(p1 - p0, plane.normal);
+	CVector normal;
+	plane.GetNormal(normal);
+	t = -plane.CalcPoint(p0) / DotProduct(p1 - p0, normal);
 	// find point of intersection
 	CVector p = p0 + (p1-p0)*t;
 
@@ -899,13 +1393,36 @@ CCollision::IsStoredPolyStillValidVerticalLine(const CVector &pos, float z, CCol
 	if(CrossProduct2D(vec3-vec2, vect-vec2) < 0.0f) return poly->valid = false;
 	point.point = p;
 	return poly->valid = true;
+#endif
 }
 
 bool
-CCollision::ProcessLineTriangle(const CColLine &line ,
+CCollision::ProcessLineTriangle(const CColLine &line,
 	const CompressedVector *verts, const CColTriangle &tri, const CColTrianglePlane &plane,
 	CColPoint &point, float &mindist, CStoredCollPoly *poly)
 {
+#ifdef VU_COLLISION
+	// not used in favour of optimized loops
+	VuTriangle vutri;
+	verts[tri.a].Unpack(vutri.v0);
+	verts[tri.b].Unpack(vutri.v1);
+	verts[tri.c].Unpack(vutri.v2);
+	plane.Unpack(vutri.plane);
+
+	LineToTriangleCollisionCompressed(*(CVuVector*)&line.p0, *(CVuVector*)&line.p1, vutri);
+
+	CVuVector pnt, normal;
+	float dist;
+	if(GetVUresult(pnt, normal, dist)){
+		if(dist < mindist){
+			point.point = pnt;
+			point.normal = normal;
+			mindist = dist;
+			return true;
+		}
+	}
+	return false;
+#else
 	float t;
 	CVector normal;
 	plane.GetNormal(normal);
@@ -985,6 +1502,7 @@ CCollision::ProcessLineTriangle(const CColLine &line ,
 	}
 	mindist = t;
 	return true;
+#endif
 }
 
 bool
@@ -992,6 +1510,30 @@ CCollision::ProcessSphereTriangle(const CColSphere &sphere,
 	const CompressedVector *verts, const CColTriangle &tri, const CColTrianglePlane &plane,
 	CColPoint &point, float &mindistsq)
 {
+#ifdef VU_COLLISION
+	// not used in favour of optimized loops
+	VuTriangle vutri;
+	verts[tri.a].Unpack(vutri.v0);
+	verts[tri.b].Unpack(vutri.v1);
+	verts[tri.c].Unpack(vutri.v2);
+	plane.Unpack(vutri.plane);
+
+	SphereToTriangleCollisionCompressed(*(CVuVector*)&sphere, vutri);
+
+	CVuVector pnt, normal;
+	float dist;
+	if(GetVUresult(pnt, normal, dist) && dist*dist < mindistsq){
+		float depth = sphere.radius - dist;
+		if(depth > point.depth){
+			point.point = pnt;
+			point.normal = normal;
+			point.depth = depth;
+			mindistsq = dist*dist;
+			return true;
+		}
+	}
+	return false;
+#else
 	// If sphere and plane don't intersect, no collision
 	float planedist = plane.CalcPoint(sphere.center);
 	float distsq = planedist*planedist;
@@ -1061,13 +1603,16 @@ CCollision::ProcessSphereTriangle(const CColSphere &sphere,
 	point.point = p;
 	point.normal = sphere.center - p;
 	point.normal.Normalise();
+#ifndef VU_COLLISION
 	point.surfaceA = sphere.surface;
 	point.pieceA = sphere.piece;
 	point.surfaceB = tri.surface;
 	point.pieceB = 0;
+#endif
 	point.depth = sphere.radius - dist;
 	mindistsq = dist*dist;
 	return true;
+#endif
 }
 
 bool
@@ -1075,6 +1620,94 @@ CCollision::ProcessLineOfSight(const CColLine &line,
 	const CMatrix &matrix, CColModel &model,
 	CColPoint &point, float &mindist, bool ignoreSeeThrough, bool ignoreShootThrough)
 {
+#ifdef VU_COLLISION
+	CMatrix matTransform;
+	int i;
+
+	// transform line to model space
+	Invert(matrix, matTransform);
+	CVuVector newline[2];
+	TransformPoints(newline, 2, matTransform, (RwV3d*)&line.p0, sizeof(CColLine)/2);
+
+	if(mindist < 1.0f)
+		newline[1] = newline[0] + (newline[1] - newline[0])*mindist;
+
+	// If we don't intersect with the bounding box, no chance on the rest
+	if(!TestLineBox(*(CColLine*)newline, model.boundingBox))
+		return false;
+
+	float coldist = 1.0f;
+	for(i = 0; i < model.numSpheres; i++){
+		if(ignoreSeeThrough && IsSeeThrough(model.spheres[i].surface)) continue;
+		if(ProcessLineSphere(*(CColLine*)newline, model.spheres[i], point, coldist))
+			point.Set(0, 0, model.spheres[i].surface, model.spheres[i].piece);
+	}
+
+	for(i = 0; i < model.numBoxes; i++){
+		if(ignoreSeeThrough && IsSeeThrough(model.boxes[i].surface)) continue;
+		if(ProcessLineBox(*(CColLine*)newline, model.boxes[i], point, coldist))
+			point.Set(0, 0, model.boxes[i].surface, model.boxes[i].piece);
+	}
+
+	CalculateTrianglePlanes(&model);
+	VuTriangle vutri;
+	CColTriangle *lasttri = nil;
+	for(i = 0; i < model.numTriangles; i++){
+		if(ignoreSeeThrough && IsSeeThrough(model.triangles[i].surface)) continue;
+
+		CColTriangle *tri = &model.triangles[i];
+		model.vertices[tri->a].Unpack(vutri.v0);
+		model.vertices[tri->b].Unpack(vutri.v1);
+		model.vertices[tri->c].Unpack(vutri.v2);
+		model.trianglePlanes[i].Unpack(vutri.plane);
+
+		LineToTriangleCollisionCompressed(newline[0], newline[1], vutri);
+		lasttri = tri;
+		break;
+	}
+#ifdef FIX_BUGS
+	// no need to check first again
+	i++;
+#endif
+	CVuVector pnt, normal;
+	float dist;
+	for(; i < model.numTriangles; i++){
+		if(ignoreSeeThrough && IsSeeThrough(model.triangles[i].surface)) continue;
+
+		CColTriangle *tri = &model.triangles[i];
+		model.vertices[tri->a].Unpack(vutri.v0);
+		model.vertices[tri->b].Unpack(vutri.v1);
+		model.vertices[tri->c].Unpack(vutri.v2);
+		model.trianglePlanes[i].Unpack(vutri.plane);
+
+		if(GetVUresult(pnt, normal, dist))
+			if(dist < coldist){
+				point.point = pnt;
+				point.normal = normal;
+				point.Set(0, 0, lasttri->surface, 0);
+				coldist = dist;
+			}
+
+		LineToTriangleCollisionCompressed(newline[0], newline[1], vutri);
+		lasttri = tri;
+	}
+	if(lasttri && GetVUresult(pnt, normal, dist))
+		if(dist < coldist){
+			point.point = pnt;
+			point.normal = normal;
+			point.Set(0, 0, lasttri->surface, 0);
+			coldist = dist;
+		}
+
+
+	if(coldist < 1.0f){
+		point.point = matrix * point.point;
+		point.normal = Multiply3x3(matrix, point.normal);
+		mindist *= coldist;
+		return true;
+	}
+	return false;
+#else
 	static CMatrix matTransform;
 	int i;
 
@@ -1113,6 +1746,7 @@ CCollision::ProcessLineOfSight(const CColLine &line,
 		return true;
 	}
 	return false;
+#endif
 }
 
 bool
@@ -1120,6 +1754,125 @@ CCollision::ProcessVerticalLine(const CColLine &line,
 	const CMatrix &matrix, CColModel &model,
 	CColPoint &point, float &mindist, bool ignoreSeeThrough, bool ignoreShootThrough, CStoredCollPoly *poly)
 {
+#ifdef VU_COLLISION
+	static CStoredCollPoly TempStoredPoly;
+	CMatrix matTransform;
+	int i;
+
+	// transform line to model space
+	Invert(matrix, matTransform);
+	CVuVector newline[2];
+	TransformPoints(newline, 2, matTransform, (RwV3d*)&line.p0, sizeof(CColLine)/2);
+
+	if(mindist < 1.0f)
+		newline[1] = newline[0] + (newline[1] - newline[0])*mindist;
+
+	if(!TestLineBox(*(CColLine*)newline, model.boundingBox))
+		return false;
+
+	float coldist = 1.0f;
+	for(i = 0; i < model.numSpheres; i++){
+		if(ignoreSeeThrough && IsSeeThrough(model.spheres[i].surface)) continue;
+		if(ProcessLineSphere(*(CColLine*)newline, model.spheres[i], point, coldist))
+			point.Set(0, 0, model.spheres[i].surface, model.spheres[i].piece);
+	}
+
+	for(i = 0; i < model.numBoxes; i++){
+		if(ignoreSeeThrough && IsSeeThrough(model.boxes[i].surface)) continue;
+		if(ProcessLineBox(*(CColLine*)newline, model.boxes[i], point, coldist))
+			point.Set(0, 0, model.boxes[i].surface, model.boxes[i].piece);
+	}
+
+	CalculateTrianglePlanes(&model);
+	TempStoredPoly.valid = false;
+	if(model.numTriangles){
+		bool registeredCol;
+		CColTriangle *lasttri = nil;
+		VuTriangle vutri;
+		for(i = 0; i < model.numTriangles; i++){
+			if(ignoreSeeThrough && IsSeeThrough(model.triangles[i].surface)) continue;
+
+			CColTriangle *tri = &model.triangles[i];
+			model.vertices[tri->a].Unpack(vutri.v0);
+			model.vertices[tri->b].Unpack(vutri.v1);
+			model.vertices[tri->c].Unpack(vutri.v2);
+			model.trianglePlanes[i].Unpack(vutri.plane);
+
+			LineToTriangleCollisionCompressed(newline[0], newline[1], vutri);
+			lasttri = tri;
+			break;
+		}
+#ifdef FIX_BUGS
+		// no need to check first again
+		i++;
+#endif
+		CVuVector pnt, normal;
+		float dist;
+		for(; i < model.numTriangles; i++){
+			if(ignoreSeeThrough && IsSeeThrough(model.triangles[i].surface)) continue;
+
+			CColTriangle *tri = &model.triangles[i];
+			model.vertices[tri->a].Unpack(vutri.v0);
+			model.vertices[tri->b].Unpack(vutri.v1);
+			model.vertices[tri->c].Unpack(vutri.v2);
+			model.trianglePlanes[i].Unpack(vutri.plane);
+
+			if(GetVUresult(pnt, normal, dist)){
+				if(dist < coldist){
+					point.point = pnt;
+					point.normal = normal;
+					point.Set(0, 0, lasttri->surface, 0);
+					coldist = dist;
+					registeredCol = true;
+				}else
+					registeredCol = false;
+			}else
+				registeredCol = false;
+
+			if(registeredCol){
+				TempStoredPoly.verts[0] = model.vertices[lasttri->a].Get();
+				TempStoredPoly.verts[1] = model.vertices[lasttri->b].Get();
+				TempStoredPoly.verts[2] = model.vertices[lasttri->c].Get();
+				TempStoredPoly.valid = true;
+			}
+
+			LineToTriangleCollisionCompressed(newline[0], newline[1], vutri);
+			lasttri = tri;
+		}
+		if(lasttri && GetVUresult(pnt, normal, dist)){
+			if(dist < coldist){
+				point.point = pnt;
+				point.normal = normal;
+				point.Set(0, 0, lasttri->surface, 0);
+				coldist = dist;
+				registeredCol = true;
+			}else
+				registeredCol = false;
+		}else
+			registeredCol = false;
+
+		if(registeredCol){
+			TempStoredPoly.verts[0] = model.vertices[lasttri->a].Get();
+			TempStoredPoly.verts[1] = model.vertices[lasttri->b].Get();
+			TempStoredPoly.verts[2] = model.vertices[lasttri->c].Get();
+			TempStoredPoly.valid = true;
+		}
+	}
+
+	if(coldist < 1.0f){
+		point.point = matrix * point.point;
+		point.normal = Multiply3x3(matrix, point.normal);
+		if(TempStoredPoly.valid && poly){
+			*poly = TempStoredPoly;
+			poly->verts[0] = matrix * CVector(poly->verts[0]);
+			poly->verts[1] = matrix * CVector(poly->verts[1]);
+			poly->verts[2] = matrix * CVector(poly->verts[2]);
+		}
+		mindist *= coldist;
+		return true;
+	}
+	return false;
+#else
 	static CStoredCollPoly TempStoredPoly;
 	int i;
 
@@ -1162,6 +1915,7 @@ CCollision::ProcessVerticalLine(const CColLine &line,
 		return true;
 	}
 	return false;
+#endif
 }
 
 enum {
@@ -1171,6 +1925,15 @@ enum {
 	MAXNUMTRIS = 600
 };
 
+#ifdef VU_COLLISION
+#ifdef GTA_PS2
+#define SPR(off) ((uint8*)(0x70000000 + (off)))
+#else
+static uint8 fakeSPR[16*1024];
+#define SPR(off) ((uint8*)(fakeSPR + (off)))
+#endif
+#endif
+
 // This checks model A's spheres and lines against model B's spheres, boxes and triangles.
 // Returns the number of A's spheres that collide.
 // Returned ColPoints are in world space.
@@ -1180,6 +1943,308 @@ CCollision::ProcessColModels(const CMatrix &matrixA, CColModel &modelA,
 	const CMatrix &matrixB, CColModel &modelB,
 	CColPoint *spherepoints, CColPoint *linepoints, float *linedists)
 {
+#ifdef VU_COLLISION
+	CVuVector *aSpheresA = (CVuVector*)SPR(0x0000);
+	CVuVector *aSpheresB = (CVuVector*)SPR(0x0800);
+	CVuVector *aLinesA = (CVuVector*)SPR(0x1000);
+	int32 *aSphereIndicesA = (int32*)SPR(0x1200);
+	int32 *aSphereIndicesB = (int32*)SPR(0x1400);
+	int32 *aBoxIndicesB = (int32*)SPR(0x1600);
+	int32 *aTriangleIndicesB = (int32*)SPR(0x1680);
+	bool *aCollided = (bool*)SPR(0x1FE0);
+	CMatrix &matAB = *(CMatrix*)SPR(0x1FF0);
+	CMatrix &matBA = *(CMatrix*)SPR(0x2040);
+	int i, j, k;
+
+	// From model A space to model B space
+	Invert(matrixB, matAB);
+	matAB *= matrixA;
+
+	CVuVector bsphereAB;	// bounding sphere of A in B space
+	TransformPoint(bsphereAB, matAB, *(RwV3d*)modelA.boundingSphere.center);	// inlined
+	bsphereAB.w = modelA.boundingSphere.radius;
+	if(!TestSphereBox(*(CColSphere*)&bsphereAB, modelB.boundingBox))
+		return 0;
+
+	// transform modelA's spheres and lines to B space
+	TransformPoints(aSpheresA, modelA.numSpheres, matAB, (RwV3d*)&modelA.spheres->center, sizeof(CColSphere));
+	for(i = 0; i < modelA.numSpheres; i++)
+		aSpheresA[i].w = modelA.spheres[i].radius;
+	TransformPoints(aLinesA, modelA.numLines*2, matAB, (RwV3d*)&modelA.lines->p0, sizeof(CColLine)/2);
+
+	// Test them against model B's bounding volumes
+	int numSpheresA = 0;
+	for(i = 0; i < modelA.numSpheres; i++)
+		if(TestSphereBox(*(CColSphere*)&aSpheresA[i], modelB.boundingBox))
+			aSphereIndicesA[numSpheresA++] = i;
+	// No collision
+	if(numSpheresA == 0 && modelA.numLines == 0)
+		return 0;
+
+
+	// B to A space
+	Invert(matrixA, matBA);
+	matBA *= matrixB;
+
+	// transform modelB's spheres to A space
+	TransformPoints(aSpheresB, modelB.numSpheres, matBA, (RwV3d*)&modelB.spheres->center, sizeof(CColSphere));
+	for(i = 0; i < modelB.numSpheres; i++)
+		aSpheresB[i].w = modelB.spheres[i].radius;
+
+	// Check model B against A's bounding volumes
+	int numSpheresB = 0;
+	int numBoxesB = 0;
+	int numTrianglesB = 0;
+	for(i = 0; i < modelB.numSpheres; i++)
+		if(TestSphereBox(*(CColSphere*)&aSpheresB[i], modelA.boundingBox))
+			aSphereIndicesB[numSpheresB++] = i;
+	for(i = 0; i < modelB.numBoxes; i++)
+		if(TestSphereBox(*(CColSphere*)&bsphereAB, modelB.boxes[i]))
+			aBoxIndicesB[numBoxesB++] = i;
+	CalculateTrianglePlanes(&modelB);
+	if(modelB.numTriangles){
+		VuTriangle vutri;
+		// process the first triangle
+		CColTriangle *tri = &modelB.triangles[0];
+		modelB.vertices[tri->a].Unpack(vutri.v0);
+		modelB.vertices[tri->b].Unpack(vutri.v1);
+		modelB.vertices[tri->c].Unpack(vutri.v2);
+		modelB.trianglePlanes[0].Unpack(vutri.plane);
+
+		SphereToTriangleCollisionCompressed(bsphereAB, vutri);
+
+		for(i = 1; i < modelB.numTriangles; i++){
+			// set up the next triangle while VU0 is running
+			tri = &modelB.triangles[i];
+			modelB.vertices[tri->a].Unpack(vutri.v0);
+			modelB.vertices[tri->b].Unpack(vutri.v1);
+			modelB.vertices[tri->c].Unpack(vutri.v2);
+			modelB.trianglePlanes[i].Unpack(vutri.plane);
+
+			// check previous result
+			if(GetVUresult())
+				aTriangleIndicesB[numTrianglesB++] = i-1;
+
+			// kick off this one
+			SphereToTriangleCollisionCompressed(bsphereAB, vutri);
+		}
+
+		// check last result
+		if(GetVUresult())
+			aTriangleIndicesB[numTrianglesB++] = i-1;
+	}
+	// No collision
+	if(numSpheresB == 0 && numBoxesB == 0 && numTrianglesB == 0)
+		return 0;
+
+	// We now have the collision volumes in A and B that are worth processing.
+
+	// Process A's spheres against B's collision volumes
+	int numCollisions = 0;
+	spherepoints[numCollisions].depth = -1.0f;
+	for(i = 0; i < numSpheresA; i++){
+		float coldist = 1.0e24f;
+		bool hasCollided = false;
+		CColSphere *sphA = &modelA.spheres[aSphereIndicesA[i]];
+		CVuVector *vusphA = &aSpheresA[aSphereIndicesA[i]];
+
+		for(j = 0; j < numSpheresB; j++)
+			// This actually looks like something was inlined here
+			if(ProcessSphereSphere(*(CColSphere*)vusphA, modelB.spheres[aSphereIndicesB[j]],
+					spherepoints[numCollisions], coldist)){
+				spherepoints[numCollisions].Set(
+					sphA->surface, sphA->piece,
+					modelB.spheres[aSphereIndicesB[j]].surface, modelB.spheres[aSphereIndicesB[j]].piece);
+				hasCollided = true;
+			}
+		for(j = 0; j < numBoxesB; j++)
+			if(ProcessSphereBox(*(CColSphere*)vusphA, modelB.boxes[aBoxIndicesB[j]],
+					spherepoints[numCollisions], coldist)){
+				spherepoints[numCollisions].Set(
+					sphA->surface, sphA->piece,
+					modelB.boxes[aBoxIndicesB[j]].surface, modelB.boxes[aBoxIndicesB[j]].piece);
+				hasCollided = true;
+			}
+		if(numTrianglesB){
+			CVuVector point, normal;
+			float depth;
+			bool registeredCol;
+			CColTriangle *lasttri;
+
+			VuTriangle vutri;
+			// process the first triangle
+			k = aTriangleIndicesB[0];
+			CColTriangle *tri = &modelB.triangles[k];
+			modelB.vertices[tri->a].Unpack(vutri.v0);
+			modelB.vertices[tri->b].Unpack(vutri.v1);
+			modelB.vertices[tri->c].Unpack(vutri.v2);
+			modelB.trianglePlanes[k].Unpack(vutri.plane);
+
+			SphereToTriangleCollisionCompressed(*vusphA, vutri);
+			lasttri = tri;
+
+			for(j = 1; j < numTrianglesB; j++){
+				k = aTriangleIndicesB[j];
+				// set up the next triangle while VU0 is running
+				tri = &modelB.triangles[k];
+				modelB.vertices[tri->a].Unpack(vutri.v0);
+				modelB.vertices[tri->b].Unpack(vutri.v1);
+				modelB.vertices[tri->c].Unpack(vutri.v2);
+				modelB.trianglePlanes[k].Unpack(vutri.plane);
+
+				// check previous result
+				// TODO: this looks inlined but spherepoints[numCollisions] does not...
+				if(GetVUresult(point, normal, depth)){
+					depth = sphA->radius - depth;
+					if(depth > spherepoints[numCollisions].depth){
+						spherepoints[numCollisions].point = point;
+						spherepoints[numCollisions].normal = normal;
+						spherepoints[numCollisions].Set(depth,
+							sphA->surface, sphA->piece, lasttri->surface, 0);
+						registeredCol = true;
+					}else
+						registeredCol = false;
+				}else
+					registeredCol = false;
+
+				if(registeredCol)
+					hasCollided = true;
+
+				// kick off this one
+				SphereToTriangleCollisionCompressed(*vusphA, vutri);
+				lasttri = tri;
+			}
+
+			// check last result
+			// TODO: this looks inlined but spherepoints[numCollisions] does not...
+			if(GetVUresult(point, normal, depth)){
+				depth = sphA->radius - depth;
+				if(depth > spherepoints[numCollisions].depth){
+					spherepoints[numCollisions].point = point;
+					spherepoints[numCollisions].normal = normal;
+					spherepoints[numCollisions].Set(depth,
+						sphA->surface, sphA->piece, lasttri->surface, 0);
+					registeredCol = true;
+				}else
+					registeredCol = false;
+			}else
+				registeredCol = false;
+
+			if(registeredCol)
+				hasCollided = true;
+		}
+
+		if(hasCollided){
+			numCollisions++;
+			if(numCollisions == MAX_COLLISION_POINTS)
+				break;
+			spherepoints[numCollisions].depth = -1.0f;
+		}
+	}
+	for(i = 0; i < numCollisions; i++){
+		// TODO: both VU0 macros
+		spherepoints[i].point = matrixB * spherepoints[i].point;
+		spherepoints[i].normal = Multiply3x3(matrixB, spherepoints[i].normal);
+	}
+
+	// And the same thing for the lines in A
+	for(i = 0; i < modelA.numLines; i++){
+		aCollided[i] = false;
+		CVuVector *lineA = &aLinesA[i*2];
+
+		for(j = 0; j < numSpheresB; j++)
+			if(ProcessLineSphere(*(CColLine*)lineA, modelB.spheres[aSphereIndicesB[j]],
+					linepoints[i], linedists[i])){
+				linepoints[i].Set(0, 0,
+#ifdef FIX_BUGS
+					modelB.spheres[aSphereIndicesB[j]].surface, modelB.spheres[aSphereIndicesB[j]].piece);
+#else
+					modelB.spheres[j].surface, modelB.spheres[j].piece);
+#endif
+				aCollided[i] = true;
+			}
+		for(j = 0; j < numBoxesB; j++)
+			if(ProcessLineBox(*(CColLine*)lineA, modelB.boxes[aBoxIndicesB[j]],
+					linepoints[i], linedists[i])){
+				linepoints[i].Set(0, 0,
+					modelB.boxes[aBoxIndicesB[j]].surface, modelB.boxes[aBoxIndicesB[j]].piece);
+				aCollided[i] = true;
+			}
+		if(numTrianglesB){
+			CVuVector point, normal;
+			float dist;
+			bool registeredCol;
+			CColTriangle *lasttri;
+
+			VuTriangle vutri;
+			// process the first triangle
+			k = aTriangleIndicesB[0];
+			CColTriangle *tri = &modelB.triangles[k];
+			modelB.vertices[tri->a].Unpack(vutri.v0);
+			modelB.vertices[tri->b].Unpack(vutri.v1);
+			modelB.vertices[tri->c].Unpack(vutri.v2);
+			modelB.trianglePlanes[k].Unpack(vutri.plane);
+
+			LineToTriangleCollisionCompressed(lineA[0], lineA[1], vutri);
+			lasttri = tri;
+
+			for(j = 1; j < numTrianglesB; j++){
+				k = aTriangleIndicesB[j];
+				// set up the next triangle while VU0 is running
+				CColTriangle *tri = &modelB.triangles[k];
+				modelB.vertices[tri->a].Unpack(vutri.v0);
+				modelB.vertices[tri->b].Unpack(vutri.v1);
+				modelB.vertices[tri->c].Unpack(vutri.v2);
+				modelB.trianglePlanes[k].Unpack(vutri.plane);
+
+				// check previous result
+				// TODO: this again somewhat looks inlined
+				if(GetVUresult(point, normal, dist)){
+					if(dist < linedists[i]){
+						linepoints[i].point = point;
+						linepoints[i].normal = normal;
+						linedists[i] = dist;
+						linepoints[i].Set(0, 0, lasttri->surface, 0);
+						registeredCol = true;
+					}else
+						registeredCol = false;
+				}else
+					registeredCol = false;
+
+				if(registeredCol)
+					aCollided[i] = true;
+
+				// kick of this one
+				LineToTriangleCollisionCompressed(lineA[0], lineA[1], vutri);
+				lasttri = tri;
+			}
+
+			// check last result
+			if(GetVUresult(point, normal, dist)){
+				if(dist < linedists[i]){
+					linepoints[i].point = point;
+					linepoints[i].normal = normal;
+					linedists[i] = dist;
+					linepoints[i].Set(0, 0, lasttri->surface, 0);
+					registeredCol = true;
+				}else
+					registeredCol = false;
+			}else
+				registeredCol = false;
+
+			if(registeredCol)
+				aCollided[i] = true;
+		}
+
+		if(aCollided[i]){
+			// TODO: both VU0 macros
+			linepoints[i].point = matrixB * linepoints[i].point;
+			linepoints[i].normal = Multiply3x3(matrixB, linepoints[i].normal);
+		}
+	}
+
+	return numCollisions;	// sphere collisions
+#else
 	static int aSphereIndicesA[MAXNUMSPHERES];
 	static int aLineIndicesA[MAXNUMLINES];
 	static int aSphereIndicesB[MAXNUMSPHERES];
@@ -1196,14 +2261,16 @@ CCollision::ProcessColModels(const CMatrix &matrixA, CColModel &modelA,
 	assert(modelA.numLines <= MAXNUMLINES);
 
 	// From model A space to model B space
-	matAB = Invert(matrixB, matAB) * matrixA;
+	Invert(matrixB, matAB);
+	matAB *= matrixA;
 
 	CColSphere bsphereAB;	// bounding sphere of A in B space
 	bsphereAB.Set(modelA.boundingSphere.radius, matAB * modelA.boundingSphere.center);
 	if(!TestSphereBox(bsphereAB, modelB.boundingBox))
 		return 0;
 	// B to A space
-	matBA = Invert(matrixA, matBA) * matrixB;
+	matBA = Invert(matrixA, matBA);
+	matBA *= matrixB;
 
 	// transform modelA's spheres and lines to B space
 	for(i = 0; i < modelA.numSpheres; i++){
@@ -1316,6 +2383,7 @@ CCollision::ProcessColModels(const CMatrix &matrixA, CColModel &modelA,
 		}
 
 	return numCollisions;	// sphere collisions
+#endif
 }
 
 
@@ -1858,6 +2926,19 @@ CColTriangle::Set(const CompressedVector *, int a, int b, int c, uint8 surf, uin
 	this->surface = surf;
 }
 
+#ifdef VU_COLLISION
+void
+CColTrianglePlane::Set(const CVector &va, const CVector &vb, const CVector &vc)
+{
+	CVector norm = CrossProduct(vc-va, vb-va);
+	norm.Normalise();
+	float d = DotProduct(norm, va);
+	normal.x = norm.x*4096.0f;
+	normal.y = norm.y*4096.0f;
+	normal.z = norm.z*4096.0f;
+	dist = d*128.0f;
+}
+#else
 void
 CColTrianglePlane::Set(const CVector &va, const CVector &vb, const CVector &vc)
 {
@@ -1873,6 +2954,7 @@ CColTrianglePlane::Set(const CVector &va, const CVector &vb, const CVector &vc)
 	else
 		dir = normal.z < 0.0f ? DIR_Z_NEG : DIR_Z_POS;
 }
+#endif
 
 CColPoint&
 CColPoint::operator=(const CColPoint& other)
diff --git a/src/core/Collision.h b/src/core/Collision.h
index 1dc29efb..fd079028 100644
--- a/src/core/Collision.h
+++ b/src/core/Collision.h
@@ -2,9 +2,12 @@
 
 #include "templates.h"
 #include "Game.h"	// for eLevelName
+#ifdef VU_COLLISION
+#include "VuVector.h"
+#endif
 
 // If you spawn many tanks at once, you will see that collisions of two entity exceeds 32.
-#ifdef FIX_BUGS
+#if defined(FIX_BUGS) && !defined(SQUEEZE_PERFORMANCE)
 #define MAX_COLLISION_POINTS 64
 #else
 #define MAX_COLLISION_POINTS 32
@@ -16,6 +19,28 @@ struct CompressedVector
 	int16 x, y, z;
 	CVector Get(void) const { return CVector(x, y, z)/128.0f; };
 	void Set(float x, float y, float z) { this->x = x*128.0f; this->y = y*128.0f; this->z = z*128.0f; };
+#ifdef GTA_PS2
+	void Unpack(uint128 &qword) const {
+		__asm__ volatile (
+			"lh      $8, 0(%1)\n"
+			"lh      $9, 2(%1)\n"
+			"lh      $10, 4(%1)\n"
+			"pextlw  $10, $8\n"
+			"pextlw  $2, $9, $10\n"
+			"sq      $2, %0\n"
+			: "=m" (qword)
+			: "r" (this)
+			: "$8", "$9", "$10", "$2"
+		);
+	}
+#else
+	void Unpack(int32 *qword) const {
+		qword[0] = x;
+		qword[1] = y;
+		qword[2] = z;
+		qword[3] = 0;	// junk
+	}
+#endif
 #else
 	float x, y, z;
 	CVector Get(void) const { return CVector(x, y, z); };
@@ -25,6 +50,7 @@ struct CompressedVector
 
 struct CSphere
 {
+	// NB: this has to be compatible with a CVuVector
 	CVector center;
 	float radius;
 	void Set(float radius, const CVector &center) { this->center = center; this->radius = radius; }
@@ -59,6 +85,7 @@ struct CColBox : public CBox
 
 struct CColLine
 {
+	// NB: this has to be compatible with two CVuVectors
 	CVector p0;
 	int pad0;
 	CVector p1;
@@ -81,6 +108,39 @@ struct CColTriangle
 
 struct CColTrianglePlane
 {
+#ifdef VU_COLLISION
+	CompressedVector normal;
+	int16 dist;
+
+	void Set(const CVector &va, const CVector &vb, const CVector &vc);
+	void Set(const CompressedVector *v, CColTriangle &tri) { Set(v[tri.a].Get(), v[tri.b].Get(), v[tri.c].Get()); }
+	void GetNormal(CVector &n) const { n.x = normal.x/4096.0f; n.y = normal.y/4096.0f; n.z = normal.z/4096.0f; }
+	float CalcPoint(const CVector &v) const { CVector n; GetNormal(n); return DotProduct(n, v) - dist/128.0f; };
+#ifdef GTA_PS2
+	void Unpack(uint128 &qword) const {
+		__asm__ volatile (
+			"lh      $8, 0(%1)\n"
+			"lh      $9, 2(%1)\n"
+			"lh      $10, 4(%1)\n"
+			"lh      $11, 6(%1)\n"
+			"pextlw  $10, $8\n"
+			"pextlw  $11, $9\n"
+			"pextlw  $2, $11, $10\n"
+			"sq      $2, %0\n"
+			: "=m" (qword)
+			: "r" (this)
+			: "$8", "$9", "$10", "$11", "$2"
+		);
+	}
+#else
+	void Unpack(int32 *qword) const {
+		qword[0] = normal.x;
+		qword[1] = normal.y;
+		qword[2] = normal.z;
+		qword[3] = dist;
+	}
+#endif
+#else
 	CVector normal;
 	float dist;
 	uint8 dir;
@@ -92,6 +152,7 @@ struct CColTrianglePlane
 	float GetNormalY() const { return normal.y; }
 	float GetNormalZ() const { return normal.z; }
 	float CalcPoint(const CVector &v) const { return DotProduct(normal, v) - dist; };
+#endif
 };
 
 struct CColPoint
@@ -112,7 +173,11 @@ struct CColPoint
 
 struct CStoredCollPoly
 {
+#ifdef VU_COLLISION
+	CVuVector verts[3];
+#else
 	CVector verts[3];
+#endif
 	bool valid;
 };
 
diff --git a/src/core/EventList.cpp b/src/core/EventList.cpp
index e5f264c7..da0a5967 100644
--- a/src/core/EventList.cpp
+++ b/src/core/EventList.cpp
@@ -8,6 +8,7 @@
 #include "Messages.h"
 #include "Text.h"
 #include "main.h"
+#include "Accident.h"
 
 int32 CEventList::ms_nFirstFreeSlotIndex;
 CEvent gaEvent[NUMEVENTS];
@@ -63,6 +64,13 @@ CEventList::RegisterEvent(eEventType type, eEventEntity entityType, CEntity *ent
 	int ref;
 	bool copsDontCare;
 
+#ifdef SQUEEZE_PERFORMANCE
+	if (type == EVENT_INJURED_PED) {
+		gAccidentManager.ReportAccident((CPed*)ent);
+		return;
+	}
+#endif
+
 	copsDontCare = false;
 	switch(entityType){
 	case EVENT_ENTITY_PED:
diff --git a/src/core/Game.cpp b/src/core/Game.cpp
index 7700d321..6825c247 100644
--- a/src/core/Game.cpp
+++ b/src/core/Game.cpp
@@ -109,6 +109,7 @@ int gameTxdSlot;
 
 bool DoRWStuffStartOfFrame(int16 TopRed, int16 TopGreen, int16 TopBlue, int16 BottomRed, int16 BottomGreen, int16 BottomBlue, int16 Alpha);
 void DoRWStuffEndOfFrame(void);
+#ifdef PS2_MENU
 void MessageScreen(char *msg)
 {
 	//TODO: stretch_screen
@@ -142,6 +143,7 @@ void MessageScreen(char *msg)
 	
 	DoRWStuffEndOfFrame();
 }
+#endif
 
 bool
 CGame::InitialiseOnceBeforeRW(void)
diff --git a/src/core/World.cpp b/src/core/World.cpp
index 7126b26f..fb2dbad3 100644
--- a/src/core/World.cpp
+++ b/src/core/World.cpp
@@ -1933,6 +1933,11 @@ CWorld::Process(void)
 	} else {
 		for(CPtrNode *node = ms_listMovingEntityPtrs.first; node; node = node->next) {
 			CEntity *movingEnt = (CEntity *)node->item;
+#ifdef SQUEEZE_PERFORMANCE
+			if (movingEnt->bRemoveFromWorld) {
+				RemoveEntityInsteadOfProcessingIt(movingEnt);
+			} else
+#endif
 			if(movingEnt->m_rwObject && RwObjectGetType(movingEnt->m_rwObject) == rpCLUMP &&
 			   RpAnimBlendClumpGetFirstAssociation(movingEnt->GetClump())) {
 // TODO(MIAMI): doRender argument
diff --git a/src/core/Zones.cpp b/src/core/Zones.cpp
index 3095b36d..96f5dc62 100644
--- a/src/core/Zones.cpp
+++ b/src/core/Zones.cpp
@@ -7,6 +7,7 @@
 #include "Clock.h"
 #include "Text.h"
 #include "World.h"
+#include "Timer.h"
 
 eLevelName CTheZones::m_CurrLevel;
 int16 CTheZones::FindIndex;
@@ -127,6 +128,10 @@ CTheZones::Init(void)
 void
 CTheZones::Update(void)
 {
+#ifdef SQUEEZE_PERFORMANCE
+	if (CTimer::GetFrameCounter() % 5 != 0)
+		return;
+#endif
 	CVector pos;
 	pos = FindPlayerCoors();
 	m_CurrLevel = GetLevelFromPosition(&pos);
diff --git a/src/core/common.h b/src/core/common.h
index d22c1715..ba1d24ef 100644
--- a/src/core/common.h
+++ b/src/core/common.h
@@ -89,6 +89,16 @@ typedef uint16_t wchar;
 #include <rpskin.h>
 #endif
 
+#ifdef __GNUC__
+#define TYPEALIGN(n) __attribute__ ((aligned (n)))
+#else
+#ifdef _MSC_VER
+#define TYPEALIGN(n) __declspec(align(n))
+#else
+#define TYPEALIGN(n)	// unknown compiler...ignore
+#endif
+#endif
+
 #define ALIGNPTR(p) (void*)((((uintptr)(void*)p) + sizeof(void*)-1) & ~(sizeof(void*)-1))
 
 // PDP-10 like byte functions
diff --git a/src/core/config.h b/src/core/config.h
index 4cd4cbe1..fa00bf67 100644
--- a/src/core/config.h
+++ b/src/core/config.h
@@ -167,7 +167,7 @@ enum Config {
 #if defined GTA_PS2
 #	define GTA_PS2_STUFF
 #	define RANDOMSPLASH
-#	define COMPRESSED_COL_VECTORS
+#	define VU_COLLISION
 #elif defined GTA_PC
 #	define GTA3_1_1_PATCH
 //#	define GTA3_STEAM_PATCH
@@ -180,6 +180,10 @@ enum Config {
 #elif defined GTA_XBOX
 #endif
 
+#ifdef VU_COLLISION
+#define COMPRESSED_COL_VECTORS	// current need compressed vectors in this code
+#endif
+
 #ifdef MASTER
 	// only in master builds
 #else
@@ -297,4 +301,13 @@ enum Config {
 #ifndef AUDIO_OAL // is not working yet for openal
 #define AUDIO_CACHE // cache sound lengths to speed up the cold boot
 #endif
-//#define PS2_AUDIO   // changes audio paths for cutscenes and radio to PS2 paths, needs vbdec to support VB with MSS
\ No newline at end of file
+//#define PS2_AUDIO   // changes audio paths for cutscenes and radio to PS2 paths, needs vbdec to support VB with MSS
+
+
+//#define SQUEEZE_PERFORMANCE
+#ifdef SQUEEZE_PERFORMANCE
+	#undef PS2_ALPHA_TEST
+	#undef NO_ISLAND_LOADING
+	#define PC_PARTICLE
+	#define VC_PED_PORTS // To not process collisions always. But should be tested if that's really beneficial
+#endif
\ No newline at end of file
diff --git a/src/core/re3.cpp b/src/core/re3.cpp
index 81b2bfc0..10f0e032 100644
--- a/src/core/re3.cpp
+++ b/src/core/re3.cpp
@@ -489,9 +489,7 @@ DebugMenuPopulate(void)
 
 		DebugMenuAddVarBool8("Render", "Draw hud", &CHud::m_Wants_To_Draw_Hud, nil);
 		DebugMenuAddVarBool8("Render", "Backface Culling", &gBackfaceCulling, nil);
-#ifdef LIBRW
 		DebugMenuAddVarBool8("Render", "PS2 Alpha test Emu", &gPS2alphaTest, nil);
-#endif
 		DebugMenuAddVarBool8("Render", "Frame limiter", &FrontEndMenuManager.m_PrefsFrameLimiter, nil);
 		DebugMenuAddVarBool8("Render", "VSynch", &FrontEndMenuManager.m_PrefsVsync, nil);
 		DebugMenuAddVar("Render", "Max FPS", &RsGlobal.maxFPS, nil, 1, 1, 1000, nil);
diff --git a/src/core/vu0Collision.dsm b/src/core/vu0Collision.dsm
new file mode 100644
index 00000000..657c8b81
--- /dev/null
+++ b/src/core/vu0Collision.dsm
@@ -0,0 +1,21 @@
+.align 4
+.global Vu0CollisionDmaTag
+Vu0CollisionDmaTag:
+DMAcnt *
+MPG 0, *
+.vu
+.include "vu0Collision_1.s"
+.EndMPG
+.EndDmaData
+DMAend
+
+.global Vu0Collision2DmaTag
+Vu0Collision2DmaTag:
+DMAcnt *
+MPG 0, *
+.vu
+.include "vu0Collision_2.s"
+.EndMPG
+.EndDmaData
+DMAend
+.end
diff --git a/src/core/vu0Collision_1.s b/src/core/vu0Collision_1.s
new file mode 100644
index 00000000..055c8640
--- /dev/null
+++ b/src/core/vu0Collision_1.s
@@ -0,0 +1,610 @@
+QuitAndFail:
+	NOP[E]                                   IADDIU VI01, VI00, 0
+	NOP                                      NOP
+
+
+QuitAndSucceed:
+	NOP[E]                                   IADDIU VI01, VI00, 1
+	NOP                                      NOP
+
+
+; 20 -- unused
+; VF12, VF13 xyz: sphere centers
+; VF14, VF15 x: sphere radii
+; out:
+;  VI01: set when collision
+;  VF01: supposed to be intersection point?
+;  VF02: normal (pointing towards s1, not normalized)
+.globl Vu0SphereToSphereCollision
+Vu0SphereToSphereCollision:
+	SUB.xyz VF02, VF13, VF12                 NOP                     ; dist of centers
+	ADD.x VF04, VF14, VF15                   NOP                     ; s = sum of radii
+	MUL.xyzw VF03, VF02, VF02                NOP                     ;
+	MUL.x VF04, VF04, VF04                   DIV Q, VF14x, VF04x     ; square s
+	NOP                                      NOP                     ;
+	NOP                                      NOP                     ;
+	MULAx.w ACC, VF00, VF03                  NOP                     ;
+	MADDAy.w ACC, VF00, VF03                 NOP                     ;
+	MADDz.w VF03, VF00, VF03                 NOP                     ; d = DistSq of centers
+	NOP                                      NOP                     ;
+	MULAw.xyz ACC, VF12, VF00                NOP                     ;
+	MADDq.xyz VF01, VF02, Q                  NOP                     ; intersection, but wrong
+	CLIPw.xyz VF04, VF03                     NOP                     ; compare s and d
+	SUB.xyz VF02, VF00, VF02                 NOP                     ; compute normal
+	NOP                                      NOP                     ;
+	NOP                                      NOP                     ;
+	NOP                                      FCAND VI01, 0x3         ; 0x2 cannot be set here
+	NOP[E]                                   NOP                     ;
+	NOP                                      NOP                     ;
+
+
+; B8 -- unused
+; VF12:
+; VF13: radius
+; VF14:
+; VF15: box dimensions (?)
+.globl Vu0SphereToAABBCollision
+Vu0SphereToAABBCollision:
+	SUB.xyz VF03, VF12, VF14                 LOI 0.5
+	MULi.xyz VF15, VF15, I                   NOP
+	MUL.x VF13, VF13, VF13                   NOP
+	SUB.xyz VF04, VF03, VF15                 NOP
+	ADD.xyz VF05, VF03, VF15                 MR32.xyzw VF16, VF15
+	CLIPw.xyz VF03, VF16                     MR32.xyzw VF17, VF16
+	MUL.xyz VF04, VF04, VF04                 NOP
+	MUL.xyz VF05, VF05, VF05                 NOP
+	CLIPw.xyz VF03, VF17                     MR32.xyzw VF16, VF17
+	NOP                                      FCAND VI01, 0x1
+	MINI.xyz VF04, VF04, VF05                MFIR.x VF09, VI01
+	NOP                                      NOP
+	CLIPw.xyz VF03, VF16                     FCAND VI01, 0x4
+	NOP                                      MFIR.y VF09, VI01
+	NOP                                      NOP
+	MULAx.w ACC, VF00, VF00                  NOP
+	ADD.xyz VF01, VF00, VF03                 FCAND VI01, 0x10
+	NOP                                      MFIR.z VF09, VI01
+	NOP                                      LOI 2
+	NOP                                      FCAND VI01, 0x30
+	SUBAw.xyz ACC, VF00, VF00                IADD VI04, VI00, VI01
+	ITOF0.xyz VF09, VF09                     FCAND VI01, 0x300
+	NOP                                      IADD VI03, VI00, VI01
+	NOP                                      FCAND VI01, 0x3000
+	NOP                                      IADD VI02, VI00, VI01
+	MADDi.xyzw VF09, VF09, I                 NOP
+	NOP                                      IBEQ VI04, VI00, IgnoreZValue
+	NOP                                      NOP
+	MADDAz.w ACC, VF00, VF04                 NOP
+	MUL.z VF01, VF09, VF15                   NOP
+IgnoreZValue:
+	NOP                                      IBEQ VI03, VI00, IgnoreYValue
+	NOP                                      NOP
+	MADDAy.w ACC, VF00, VF04                 NOP
+	MUL.y VF01, VF09, VF15                   NOP
+IgnoreYValue:
+	NOP                                      IBEQ VI02, VI00, IgnoreXValue
+	NOP                                      NOP
+	MADDAx.w ACC, VF00, VF04                 NOP
+	MUL.x VF01, VF09, VF15                   NOP
+IgnoreXValue:
+	MADDx.w VF06, VF00, VF00                 NOP
+	SUB.xyz VF02, VF03, VF01                 NOP
+	ADD.xyz VF01, VF01, VF14                 NOP
+	MULx.w VF01, VF00, VF00                  NOP
+	CLIPw.xyz VF13, VF06                     NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      FCAND VI01, 0x1
+QuitMicrocode:
+	NOP[E]                                   NOP
+	NOP                                      NOP
+
+
+; 240
+.globl Vu0LineToSphereCollision
+Vu0LineToSphereCollision:
+	SUB.xyzw VF01, VF13, VF12                NOP
+	SUB.xyzw VF02, VF14, VF12                NOP
+	MUL.xyz VF03, VF01, VF02                 NOP
+	MUL.xyz VF04, VF01, VF01                 NOP
+	MUL.x VF15, VF15, VF15                   NOP
+	MUL.xyz VF02, VF02, VF02                 NOP
+	MULAx.w ACC, VF00, VF03                  NOP
+	MADDAy.w ACC, VF00, VF03                 NOP
+	MADDz.w VF03, VF00, VF03                 NOP
+	MULAx.w ACC, VF00, VF04                  NOP
+	MADDAy.w ACC, VF00, VF04                 NOP
+	MADDz.w VF01, VF00, VF04                 NOP
+	MULAx.w ACC, VF00, VF02                  NOP
+	MADDAy.w ACC, VF00, VF02                 NOP
+	MADDz.w VF02, VF00, VF02                 NOP
+	MULA.w ACC, VF03, VF03                   NOP
+	MADDAx.w ACC, VF01, VF15                 NOP
+	MSUB.w VF05, VF01, VF02                  NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      IADDIU VI02, VI00, 0x10
+	NOP                                      FMAND VI01, VI02
+	NOP                                      IBNE VI01, VI00, QuitAndFail
+	NOP                                      NOP
+	CLIPw.xyz VF15, VF02                     SQRT Q, VF05w
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      FCAND VI01, 0x1
+	NOP                                      IBNE VI00, VI01, LineStartInsideSphere
+	NOP                                      NOP
+	SUBq.w VF05, VF03, Q                     NOP
+	SUB.w VF05, VF05, VF01                   DIV Q, VF05w, VF01w
+	NOP                                      FMAND VI01, VI02
+	NOP                                      IBNE VI01, VI00, QuitAndFail
+	NOP                                      NOP
+	NOP                                      FMAND VI01, VI02
+	NOP                                      IBEQ VI01, VI00, QuitAndFail
+	NOP                                      NOP
+	ADDA.xyz ACC, VF12, VF00                 NOP
+	MADDq.xyz VF01, VF01, Q                  NOP
+	MULx.w VF01, VF00, VF00                  NOP
+	SUB.xyz VF02, VF01, VF14                 NOP
+	NOP[E]                                   NOP
+	NOP                                      NOP
+LineStartInsideSphere:
+	NOP                                      MOVE.xyzw VF01, VF12
+	NOP[E]                                   IADDIU VI01, VI00, 0x1
+	NOP                                      NOP
+
+
+; 3C0
+.globl Vu0LineToAABBCollision
+Vu0LineToAABBCollision:
+	SUB.xyzw VF08, VF13, VF12                LOI 0.5
+	MULi.xyz VF15, VF15, I                   IADDIU VI08, VI00, 0x0
+	SUB.xyzw VF12, VF12, VF14                NOP
+	SUB.xyzw VF13, VF13, VF14                NOP
+	NOP                                      DIV Q, VF00w, VF08x
+	NOP                                      MR32.xyzw VF03, VF15
+	SUB.xyz VF06, VF15, VF12                 NOP
+	ADD.xyz VF07, VF15, VF12                 NOP
+	NOP                                      NOP
+	CLIPw.xyz VF12, VF03                     MR32.xyzw VF04, VF03
+	NOP                                      NOP
+	ADDq.x VF09, VF00, Q                     DIV Q, VF00w, VF08y
+	NOP                                      NOP
+	CLIPw.xyz VF12, VF04                     MR32.xyzw VF05, VF04
+	SUB.xyz VF07, VF00, VF07                 IADDIU VI06, VI00, 0xCC
+	NOP                                      IADDIU VI07, VI00, 0x30
+	NOP                                      NOP
+	CLIPw.xyz VF12, VF05                     FCGET VI02
+	NOP                                      IAND VI02, VI02, VI06
+	ADDq.y VF09, VF00, Q                     DIV Q, VF00w, VF08z
+	SUB.xyz VF10, VF00, VF10                 NOP
+	CLIPw.xyz VF13, VF03                     FCGET VI03
+	CLIPw.xyz VF13, VF04                     IAND VI03, VI03, VI07
+	CLIPw.xyz VF13, VF05                     FCAND VI01, 0x3330
+	NOP                                      IBEQ VI01, VI00, StartPointInsideAABB
+	NOP                                      NOP
+	ADDq.z VF09, VF00, Q                     FCGET VI04
+	NOP                                      FCGET VI05
+	NOP                                      IAND VI04, VI04, VI06
+	NOP                                      IAND VI05, VI05, VI07
+	MULx.xyz VF17, VF08, VF09                NOP
+	MULy.xyz VF18, VF08, VF09                IADDIU VI07, VI00, 0x80
+	MULz.xyz VF19, VF08, VF09                IAND VI06, VI02, VI07
+	MUL.w VF10, VF00, VF00                   IAND VI07, VI04, VI07
+	NOP                                      NOP
+	NOP                                      IBEQ VI06, VI07, CheckMaxXSide
+	NOP                                      NOP
+	MULAx.xyz ACC, VF17, VF07                NOP
+	MADDw.xyz VF16, VF12, VF00               NOP
+	MUL.x VF10, VF07, VF09                   NOP
+	CLIPw.xyz VF16, VF04                     NOP
+	CLIPw.xyz VF16, VF05                     NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      FCAND VI01, 0x330
+	NOP                                      IBNE VI01, VI00, CheckMaxXSide
+	NOP                                      NOP
+	MULx.w VF10, VF00, VF10                  IADDIU VI08, VI00, 0x1
+	ADD.yz VF02, VF00, VF00                  MOVE.xyzw VF01, VF16
+	SUBw.x VF02, VF00, VF00                  NOP
+CheckMaxXSide:
+	MULAx.xyz ACC, VF17, VF06                IADDIU VI07, VI00, 0x40
+	MADDw.xyz VF16, VF12, VF00               IAND VI06, VI02, VI07
+	MUL.x VF10, VF06, VF09                   IAND VI07, VI04, VI07
+	NOP                                      NOP
+	NOP                                      IBEQ VI06, VI07, CheckMinYSide
+	NOP                                      NOP
+	CLIPw.xyz VF16, VF04                     NOP
+	CLIPw.xyz VF16, VF05                     NOP
+	CLIPw.xyz VF10, VF10                     NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      FCAND VI01, 0xCC03
+	NOP                                      IBNE VI01, VI00, CheckMinYSide
+	NOP                                      NOP
+	MULx.w VF10, VF00, VF10                  IADDIU VI08, VI00, 0x1
+	ADD.yz VF02, VF00, VF00                  MOVE.xyzw VF01, VF16
+	ADDw.x VF02, VF00, VF00                  NOP
+CheckMinYSide:
+	MULAy.xyz ACC, VF18, VF07                IADDIU VI07, VI00, 0x8
+	MADDw.xyz VF16, VF12, VF00               IAND VI06, VI02, VI07
+	MUL.y VF10, VF07, VF09                   IAND VI07, VI04, VI07
+	NOP                                      NOP
+	NOP                                      IBEQ VI06, VI07, CheckMaxYSide
+	NOP                                      NOP
+	CLIPw.xyz VF16, VF03                     NOP
+	CLIPw.xyz VF16, VF05                     NOP
+	CLIPw.xyz VF10, VF10                     NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      FCAND VI01, 0x3C0C
+	NOP                                      IBNE VI01, VI00, CheckMaxYSide
+	NOP                                      NOP
+	MULy.w VF10, VF00, VF10                  IADDIU VI08, VI00, 0x1
+	ADD.xz VF02, VF00, VF00                  MOVE.xyzw VF01, VF16
+	SUBw.y VF02, VF00, VF00                  NOP
+CheckMaxYSide:
+	MULAy.xyz ACC, VF18, VF06                IADDIU VI07, VI00, 0x4
+	MADDw.xyz VF16, VF12, VF00               IAND VI06, VI02, VI07
+	MUL.y VF10, VF06, VF09                   IAND VI07, VI04, VI07
+	NOP                                      NOP
+	NOP                                      IBEQ VI06, VI07, CheckMinZSide
+	NOP                                      NOP
+	CLIPw.xyz VF16, VF03                     NOP
+	CLIPw.xyz VF16, VF05                     NOP
+	CLIPw.xyz VF10, VF10                     NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      FCAND VI01, 0x3C0C
+	NOP                                      IBNE VI01, VI00, CheckMinZSide
+	NOP                                      NOP
+	MULy.w VF10, VF00, VF10                  IADDIU VI08, VI00, 0x1
+	ADD.xz VF02, VF00, VF00                  MOVE.xyzw VF01, VF16
+	ADDw.y VF02, VF00, VF00                  NOP
+CheckMinZSide:
+	MULAz.xyz ACC, VF19, VF07                IADDIU VI07, VI00, 0x20
+	MADDw.xyz VF16, VF12, VF00               IAND VI06, VI03, VI07
+	MUL.z VF10, VF07, VF09                   IAND VI07, VI05, VI07
+	NOP                                      NOP
+	NOP                                      IBEQ VI06, VI07, CheckMaxZSide
+	NOP                                      NOP
+	CLIPw.xyz VF16, VF03                     NOP
+	CLIPw.xyz VF16, VF04                     NOP
+	CLIPw.xyz VF10, VF10                     NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      FCAND VI01, 0x3330
+	NOP                                      IBNE VI01, VI00, CheckMaxZSide
+	NOP                                      NOP
+	MULz.w VF10, VF00, VF10                  IADDIU VI08, VI00, 0x1
+	ADD.xy VF02, VF00, VF00                  MOVE.xyzw VF01, VF16
+	SUBw.z VF02, VF00, VF00                  NOP
+CheckMaxZSide:
+	MULAz.xyz ACC, VF19, VF06                IADDIU VI07, VI00, 0x10
+	MADDw.xyz VF16, VF12, VF00               IAND VI06, VI03, VI07
+	MUL.z VF10, VF06, VF09                   IAND VI07, VI05, VI07
+	NOP                                      NOP
+	NOP                                      IBEQ VI06, VI07, DoneAllChecks
+	NOP                                      NOP
+	CLIPw.xyz VF16, VF03                     NOP
+	CLIPw.xyz VF16, VF04                     NOP
+	CLIPw.xyz VF10, VF10                     NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      FCAND VI01, 0x3330
+	NOP                                      IBNE VI01, VI00, DoneAllChecks
+	NOP                                      NOP
+	MULz.w VF10, VF00, VF10                  IADDIU VI08, VI00, 0x1
+	ADD.xy VF02, VF00, VF00                  MOVE.xyzw VF01, VF16
+	ADDw.z VF02, VF00, VF00                  NOP
+DoneAllChecks:
+	ADD.xyz VF01, VF01, VF14                 IADD VI01, VI00, VI08
+	NOP[E]                                   NOP
+	NOP                                      NOP
+StartPointInsideAABB:
+	ADD.xyz VF01, VF12, VF14                 WAITQ
+	NOP                                      IADDIU VI01, VI00, 0x1
+	NOP[E]                                   NOP
+	NOP                                      NOP
+
+
+; 860
+.globl Vu0LineToTriangleCollisionCompressedStart
+Vu0LineToTriangleCollisionCompressedStart:
+	ITOF0.xyzw VF17, VF17                    LOI 0.000244140625	; 1.0/4096.0
+	ITOF0.xyzw VF14, VF14                    NOP
+	ITOF0.xyzw VF15, VF15                    NOP
+	ITOF0.xyzw VF16, VF16                    NOP
+	MULi.xyz VF17, VF17, I                   LOI 0.0078125		; 1.0/128.0
+	MULi.w VF17, VF17, I                     NOP
+	MULi.xyzw VF14, VF14, I                  NOP
+	MULi.xyzw VF15, VF15, I                  NOP
+	MULi.xyzw VF16, VF16, I                  NOP
+; fall through
+
+; 8A8
+; VF12: point0
+; VF13: point1
+; VF14-16: verts
+; VF17: plane
+; out:
+;  VF01: intersection point
+;  VF02: triangle normal
+;  VF03 x: intersection parameter
+.globl Vu0LineToTriangleCollisionStart
+Vu0LineToTriangleCollisionStart:
+	MUL.xyz VF10, VF17, VF12                 LOI 0.5
+	MUL.xyz VF11, VF17, VF13                 NOP
+	SUB.xyz VF02, VF13, VF12                 NOP                            ; line dist
+	ADD.xyz VF17, VF17, VF00                 NOP
+	MULi.w VF03, VF00, I                     NOP
+	MULAx.w ACC, VF00, VF10                  NOP
+	MADDAy.w ACC, VF00, VF10                 IADDIU VI06, VI00, 0xE0
+	MADDz.w VF10, VF00, VF10                 FMAND VI05, VI06               ; -- normal sign flags, unused
+	MULAx.w ACC, VF00, VF11                  NOP
+	MADDAy.w ACC, VF00, VF11                 NOP
+	MADDz.w VF11, VF00, VF11                 NOP
+	SUB.w VF09, VF17, VF10                   NOP                            ; plane-pos 0
+	CLIPw.xyz VF17, VF03                     NOP                            ; compare normal against 0.5 to figure out which in which dimension to compare
+	NOP                                      IADDIU VI02, VI00, 0x10        ; Sw flag
+	SUBA.w ACC, VF17, VF11                   NOP                            ; plane-pos 1
+	SUB.w VF08, VF11, VF10                   FMAND VI01, VI02
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      FMAND VI02, VI02
+	NOP                                      IBEQ VI01, VI02, QuitAndFail   ; if on same side, no collision
+	NOP                                      NOP
+	NOP                                      DIV Q, VF09w, VF08w            ; parameter of intersection
+	NOP                                      FCAND VI01, 0x3                ; check x direction
+	NOP                                      IADDIU VI02, VI01, 0x7F
+	NOP                                      IADDIU VI06, VI00, 0x80
+	NOP                                      IAND VI02, VI02, VI06          ; Sx flag
+	NOP                                      FCAND VI01, 0xC                ; check y direction
+	NOP                                      IADDIU VI03, VI01, 0x3F
+	MULAw.xyz ACC, VF12, VF00                IADDIU VI06, VI00, 0x40
+	MADDq.xyz VF01, VF02, Q                  IAND VI03, VI03, VI06          ; point of intersection -- Sy flag
+	MULx.w VF01, VF00, VF00                  FCAND VI01, 0x30               ; -- check z direction
+	ADDq.x VF03, VF00, Q                     IADDIU VI04, VI01, 0x1F        ; output parameter
+	SUB.xyz VF05, VF15, VF14                 IADDIU VI06, VI00, 0x20        ; edge vectors
+	SUB.xyz VF08, VF01, VF14                 IAND VI04, VI04, VI06          ; edge vectors -- Sz flag
+	SUB.xyz VF06, VF16, VF15                 IADD VI06, VI02, VI03          ; edge vectors
+	SUB.xyz VF09, VF01, VF15                 IADD VI06, VI06, VI04          ; edge vectors -- combine flags
+	SUB.xyz VF07, VF14, VF16                 NOP                            ; edge vectors
+	SUB.xyz VF10, VF01, VF16                 NOP                            ; edge vectors
+	OPMULA.xyz ACC, VF08, VF05               NOP
+	OPMSUB.xyz VF18, VF05, VF08              NOP                            ; cross1
+	OPMULA.xyz ACC, VF09, VF06               NOP
+	OPMSUB.xyz VF19, VF06, VF09              NOP                            ; cross2
+	OPMULA.xyz ACC, VF10, VF07               NOP
+	OPMSUB.xyz VF20, VF07, VF10              FMAND VI02, VI06               ; cross3
+	NOP                                      NOP
+	NOP                                      FMAND VI03, VI06
+	NOP                                      NOP
+	NOP                                      FMAND VI04, VI06
+	NOP                                      NOP
+	NOP                                      IBNE VI03, VI02, QuitAndFail   ; point has to lie on the same side of all edges (i.e. inside)
+	NOP                                      NOP
+	NOP                                      IBNE VI04, VI02, QuitAndFail
+	NOP                                      NOP
+	MULw.xyz VF02, VF17, VF00                IADDIU VI01, VI00, 0x1         ; success
+	NOP[E]                                   NOP
+	NOP                                      NOP
+
+
+; A68
+; VF12: center
+; VF14: line origin
+; VF15: line vector to other point
+; out: VF16 xyz: nearest point on line; w: distance to that point
+DistanceBetweenSphereAndLine:
+	SUB.xyz VF20, VF12, VF14                 NOP
+	MUL.xyz VF21, VF15, VF15                 NOP
+	ADDA.xyz ACC, VF14, VF15                 NOP
+	MSUBw.xyz VF25, VF12, VF00               NOP                                        ; VF25 = VF12 - (VF14+VF15)
+	MUL.xyz VF22, VF20, VF20                 NOP
+	MUL.xyz VF23, VF20, VF15                 NOP
+	MULAx.w ACC, VF00, VF21                  NOP
+	MADDAy.w ACC, VF00, VF21                 NOP
+	MADDz.w VF21, VF00, VF21                 NOP	                                    ; MagSq VF15 (line length)
+	MULAx.w ACC, VF00, VF23                  NOP
+	MADDAy.w ACC, VF00, VF23                 NOP
+	MADDz.w VF23, VF00, VF23                 NOP                                        ; dot(VF12-VF14, VF15)
+	MULAx.w ACC, VF00, VF22                  NOP
+	MADDAy.w ACC, VF00, VF22                 NOP
+	MADDz.w VF22, VF00, VF22                 IADDIU VI08, VI00, 0x10                    ; MagSq VF12-VF14 -- Sw bit
+	MUL.xyz VF25, VF25, VF25                 FMAND VI08, VI08
+	NOP                                      DIV Q, VF23w, VF21w
+	NOP                                      IBNE VI00, VI08, NegativeRatio
+	NOP                                      NOP
+	ADDA.xyz ACC, VF00, VF14                 NOP
+	MADDq.xyz VF16, VF15, Q                  WAITQ                                     ; nearest point on infinte line
+	ADDq.x VF24, VF00, Q                     NOP                                       ; ratio
+	NOP                                      NOP
+	NOP                                      NOP
+	SUB.xyz VF26, VF16, VF12                 NOP
+	CLIPw.xyz VF24, VF00                     NOP                                       ; compare ratio to 1.0
+	NOP                                      NOP
+	NOP                                      NOP
+	MUL.xyz VF26, VF26, VF26                 NOP
+	NOP                                      FCAND VI01, 0x1
+	NOP                                      IBNE VI00, VI01, RatioGreaterThanOne
+	NOP                                      NOP
+	MULAx.w ACC, VF00, VF26                  NOP
+	MADDAy.w ACC, VF00, VF26                 NOP
+	MADDz.w VF16, VF00, VF26                 NOP                                      ; distance
+	NOP                                      JR VI15
+	NOP                                      NOP
+NegativeRatio:
+	ADD.xyz VF16, VF00, VF14                 NOP                 ; return line origin
+	MUL.w VF16, VF00, VF22                   NOP                 ; and DistSq to it
+	NOP                                      JR VI15
+	NOP                                      NOP
+RatioGreaterThanOne:
+	MULAx.w ACC, VF00, VF25                  NOP
+	MADDAy.w ACC, VF00, VF25                 NOP
+	MADDz.w VF16, VF00, VF25                 NOP
+	ADD.xyz VF16, VF14, VF15                 NOP                ; return toerh line point
+	NOP                                      JR VI15
+	NOP                                      NOP
+
+
+; BE0
+.globl Vu0SphereToTriangleCollisionCompressedStart
+Vu0SphereToTriangleCollisionCompressedStart:
+	ITOF0.xyzw VF17, VF17                    LOI 0.000244140625	; 1.0/4096.0
+	ITOF0.xyzw VF14, VF14                    NOP
+	ITOF0.xyzw VF15, VF15                    NOP
+	ITOF0.xyzw VF16, VF16                    NOP
+	MULi.xyz VF17, VF17, I                   LOI 0.0078125		; 1.0/128.0
+	MULi.w VF17, VF17, I                     NOP
+	MULi.xyzw VF14, VF14, I                  NOP
+	MULi.xyzw VF15, VF15, I                  NOP
+	MULi.xyzw VF16, VF16, I                  NOP
+; fall through
+
+; C28
+; VF12: sphere
+; VF14-16: verts
+; VF17: plane
+; out:
+;  VF01: intersection point
+;  VF02: triangle normal
+;  VF03 x: intersection parameter
+.globl Vu0SphereToTriangleCollisionStart
+Vu0SphereToTriangleCollisionStart:
+	MUL.xyz VF02, VF12, VF17                 LOI 0.1
+	ADD.xyz VF17, VF17, VF00                 NOP
+	ADDw.x VF13, VF00, VF12                  NOP
+	NOP                                      NOP
+	MULAx.w ACC, VF00, VF02                  IADDIU VI06, VI00, 0xE0
+	MADDAy.w ACC, VF00, VF02                 FMAND VI05, VI06                  ; normal sign flags
+	MADDAz.w ACC, VF00, VF02                 NOP
+	MSUB.w VF02, VF00, VF17                  NOP                               ; center plane pos
+	MULi.w VF03, VF00, I                     MOVE.xyzw VF04, VF03
+	NOP                                      NOP
+	NOP                                      NOP
+	CLIPw.xyz VF13, VF02                     NOP                               ; compare dist and radius
+	CLIPw.xyz VF17, VF03                     NOP
+	MULAw.xyz ACC, VF12, VF00                IADDIU VI07, VI00, 0x0            ; -- clear test case
+	MSUBw.xyz VF01, VF17, VF02               NOP
+	MULx.w VF01, VF00, VF00                  FCAND VI01, 0x3                   ; projected center on plane
+	ABS.w VF02, VF02                         IBEQ VI00, VI01, QuitAndFail      ; no intersection
+	NOP                                      NOP
+	NOP                                      FCAND VI01, 0x3                   ; -- check x direction
+	SUB.xyz VF02, VF12, VF01                 IADDIU VI02, VI01, 0x7F
+	NOP                                      IADDIU VI06, VI00, 0x80
+	SUB.xyz VF05, VF15, VF14                 IAND VI02, VI02, VI06
+	SUB.xyz VF08, VF01, VF14                 FCAND VI01, 0xC                   ; -- check y direction
+	SUB.xyz VF06, VF16, VF15                 IADDIU VI03, VI01, 0x3F
+	SUB.xyz VF09, VF01, VF15                 IADDIU VI06, VI00, 0x40
+	SUB.xyz VF07, VF14, VF16                 IAND VI03, VI03, VI06
+	SUB.xyz VF10, VF01, VF16                 FCAND VI01, 0x30                  ; -- check z direction
+	MUL.xyz VF03, VF02, VF02                 IADDIU VI04, VI01, 0x1F
+	OPMULA.xyz ACC, VF08, VF05               IADDIU VI06, VI00, 0x20
+	OPMSUB.xyz VF18, VF05, VF08              IAND VI04, VI04, VI06
+	OPMULA.xyz ACC, VF09, VF06               NOP
+	OPMSUB.xyz VF19, VF06, VF09              IADD VI06, VI02, VI03
+	OPMULA.xyz ACC, VF10, VF07               IADD VI06, VI06, VI04             ; -- combine flags
+	OPMSUB.xyz VF20, VF07, VF10              FMAND VI02, VI06                  ; -- cross 1 flags
+	MULAx.w ACC, VF00, VF03                  IAND VI05, VI05, VI06
+	MADDAy.w ACC, VF00, VF03                 FMAND VI03, VI06                  ; -- cross 2 flags
+	MADDz.w VF03, VF00, VF03                 IADDIU VI08, VI00, 0x3
+	NOP                                      FMAND VI04, VI06                  ; -- cross 3 flags
+	NOP                                      NOP
+	NOP                                      IBNE VI02, VI05, CheckSide2
+	NOP                                      RSQRT Q, VF00w, VF03w
+	ADD.xyz VF04, VF00, VF16                 IADDIU VI07, VI07, 0x1                          ; inside side 1
+CheckSide2:
+	NOP                                      IBNE VI03, VI05, CheckSide3
+	NOP                                      NOP
+	ADD.xyz VF04, VF00, VF14                 IADDIU VI07, VI07, 0x1                          ; inside side 2
+CheckSide3:
+	NOP                                      IBNE VI04, VI05, FinishCheckingSides
+	NOP                                      NOP
+	ADD.xyz VF04, VF00, VF15                 IADDIU VI07, VI07, 0x1                          ; inside side 3
+	NOP                                      NOP
+	NOP                                      IBEQ VI07, VI08, TotallyInsideTriangle
+	NOP                                      NOP
+FinishCheckingSides:
+	MUL.x VF13, VF13, VF13                   IADDIU VI08, VI00, 0x2
+	MULq.xyz VF02, VF02, Q                   WAITQ
+	NOP                                      IBNE VI07, VI08, IntersectionOutsideTwoSides
+	NOP                                      NOP
+	NOP                                      IBEQ VI02, VI05, CheckDistanceSide2
+	NOP                                      NOP
+	NOP                                      MOVE.xyzw VF15, VF05
+	NOP                                      BAL VI15, DistanceBetweenSphereAndLine
+	NOP                                      NOP
+	NOP                                      B ProcessLineResult
+	NOP                                      NOP
+CheckDistanceSide2:
+	NOP                                      IBEQ VI03, VI05, CheckDistanceSide3
+	NOP                                      NOP
+	NOP                                      MOVE.xyzw VF14, VF15
+	NOP                                      MOVE.xyzw VF15, VF06
+	NOP                                      BAL VI15, DistanceBetweenSphereAndLine
+	NOP                                      NOP
+	NOP                                      B ProcessLineResult
+	NOP                                      NOP
+CheckDistanceSide3:
+	NOP                                      MOVE.xyzw VF14, VF16
+	NOP                                      MOVE.xyzw VF15, VF07
+	NOP                                      BAL VI15, DistanceBetweenSphereAndLine
+	NOP                                      NOP
+	NOP                                      B ProcessLineResult
+	NOP                                      NOP
+IntersectionOutsideTwoSides:
+	SUB.xyz VF05, VF04, VF12                 NOP
+	ADD.xyz VF01, VF00, VF04                 NOP                                ; col point
+	SUB.xyz VF02, VF12, VF04                 NOP
+	NOP                                      NOP
+	MUL.xyz VF05, VF05, VF05                 NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	MULAx.w ACC, VF00, VF05                  NOP
+	MADDAy.w ACC, VF00, VF05                 NOP
+	MADDz.w VF05, VF00, VF05                 NOP                                ; distSq to vertex
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	CLIPw.xyz VF13, VF05                     SQRT Q, VF05w                      ; compare radiusSq and distSq
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      FCAND VI01, 0x1
+	ADDq.x VF03, VF00, Q                     WAITQ                              ; dist to vertex
+	NOP                                      IBEQ VI00, VI01, QuitAndFail       ; too far
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      DIV Q, VF00w, VF03x
+	MULq.xyz VF02, VF02, Q                   WAITQ                              ; col normal
+	NOP[E]                                   NOP
+	NOP                                      NOP
+TotallyInsideTriangle:
+	ADDw.x VF03, VF00, VF02                  WAITQ
+	MULq.xyz VF02, VF02, Q                   NOP
+	NOP[E]                                   IADDIU VI01, VI00, 0x1
+	NOP                                      NOP
+ProcessLineResult:
+	CLIPw.xyz VF13, VF16                     SQRT Q, VF16w
+	ADD.xyz VF01, VF00, VF16                 NOP
+	SUB.xyz VF02, VF12, VF16                 NOP
+	NOP                                      NOP
+	NOP                                      FCAND VI01, 0x1
+	ADDq.x VF03, VF00, Q                     WAITQ
+	NOP                                      IBEQ VI00, VI01, QuitAndFail
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      DIV Q, VF00w, VF03x
+	MULq.xyz VF02, VF02, Q                   WAITQ
+	NOP[E]                                   NOP
+	NOP                                      NOP
+
+EndOfMicrocode:
diff --git a/src/core/vu0Collision_2.s b/src/core/vu0Collision_2.s
new file mode 100644
index 00000000..716c29ac
--- /dev/null
+++ b/src/core/vu0Collision_2.s
@@ -0,0 +1,191 @@
+QuitAndFail2:
+	NOP[E]                                   IADDIU VI01, VI00, 0x0
+	NOP                                      NOP
+
+
+QuitAndSucceed2:
+	NOP[E]                                   IADDIU VI01, VI00, 0x1
+	NOP                                      NOP
+
+
+; 20
+GetBBVertices:
+	MULw.xy VF02, VF01, VF00                 NOP
+	MUL.z VF02, VF01, VF11                   NOP
+	MULw.xz VF03, VF01, VF00                 NOP
+	MUL.y VF03, VF01, VF11                   NOP
+	MULw.x VF04, VF01, VF00                  NOP
+	MUL.yz VF04, VF01, VF11                  NOP
+	NOP                                      JR VI15
+	NOP                                      NOP
+
+
+; 60
+Vu0OBBToOBBCollision:
+	SUBw.xyz VF11, VF00, VF00                LOI 0.5
+	MULi.xyz VF12, VF12, I                   NOP
+	MULi.xyz VF13, VF13, I                   NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      MOVE.xyz VF01, VF12
+	NOP                                      BAL VI15, GetBBVertices
+	NOP                                      NOP
+	MULAx.xyz ACC, VF14, VF01                NOP
+	MADDAy.xyz ACC, VF15, VF01               NOP
+	MADDz.xyz VF01, VF16, VF01               NOP
+	MULAx.xyz ACC, VF14, VF02                NOP
+	MADDAy.xyz ACC, VF15, VF02               NOP
+	MADDz.xyz VF02, VF16, VF02               NOP
+	MULAx.xyz ACC, VF14, VF03                NOP
+	MADDAy.xyz ACC, VF15, VF03               NOP
+	MADDz.xyz VF03, VF16, VF03               NOP
+	MULAx.xyz ACC, VF14, VF04                NOP
+	MADDAy.xyz ACC, VF15, VF04               NOP
+	MADDz.xyz VF04, VF16, VF04               NOP
+	ABS.xyz VF05, VF01                       NOP
+	ABS.xyz VF06, VF02                       NOP
+	ABS.xyz VF07, VF03                       NOP
+	ABS.xyz VF08, VF04                       NOP
+	NOP                                      NOP
+	MAX.xyz VF05, VF05, VF06                 NOP
+	NOP                                      NOP
+	MAX.xyz VF07, VF07, VF08                 NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	MAX.xyz VF05, VF05, VF07                 NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	ADD.xyz VF09, VF05, VF13                 NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	MULx.w VF05, VF00, VF09                  NOP
+	MULy.w VF06, VF00, VF09                  NOP
+	MULz.w VF07, VF00, VF09                  NOP
+	CLIPw.xyz VF17, VF05                     NOP
+	CLIPw.xyz VF17, VF06                     NOP
+	CLIPw.xyz VF17, VF07                     MOVE.xyz VF01, VF13
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      FCAND VI01, 0x3330
+	NOP                                      IBNE VI01, VI00, QuitAndFail2
+	NOP                                      NOP
+	NOP                                      BAL VI15, GetBBVertices
+	NOP                                      NOP
+	MULAx.xyz ACC, VF18, VF01                NOP
+	MADDAy.xyz ACC, VF19, VF01               NOP
+	MADDz.xyz VF01, VF20, VF01               NOP
+	MULAx.xyz ACC, VF18, VF02                NOP
+	MADDAy.xyz ACC, VF19, VF02               NOP
+	MADDz.xyz VF02, VF20, VF02               NOP
+	MULAx.xyz ACC, VF18, VF03                NOP
+	MADDAy.xyz ACC, VF19, VF03               NOP
+	MADDz.xyz VF03, VF20, VF03               NOP
+	MULAx.xyz ACC, VF18, VF04                NOP
+	MADDAy.xyz ACC, VF19, VF04               NOP
+	MADDz.xyz VF04, VF20, VF04               NOP
+	ABS.xyz VF05, VF01                       NOP
+	ABS.xyz VF06, VF02                       NOP
+	ABS.xyz VF07, VF03                       NOP
+	ABS.xyz VF08, VF04                       NOP
+	NOP                                      NOP
+	MAX.xyz VF05, VF05, VF06                 NOP
+	NOP                                      NOP
+	MAX.xyz VF07, VF07, VF08                 NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	MAX.xyz VF05, VF05, VF07                 NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	ADD.xyz VF09, VF05, VF12                 NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	MULx.w VF05, VF00, VF09                  NOP
+	MULy.w VF06, VF00, VF09                  NOP
+	MULz.w VF07, VF00, VF09                  NOP
+	CLIPw.xyz VF21, VF05                     NOP
+	CLIPw.xyz VF21, VF06                     NOP
+	CLIPw.xyz VF21, VF07                     NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      NOP
+	NOP                                      FCAND VI01, 0x3330
+	NOP                                      IBNE VI01, VI00, QuitAndFail2
+	NOP                                      NOP
+	SUB.xyz VF06, VF02, VF01                 NOP
+	SUB.xyz VF07, VF03, VF01                 NOP
+	ADD.xyz VF08, VF04, VF01                 NOP
+	ADD.x VF09, VF00, VF12                   NOP
+	ADD.yz VF09, VF00, VF00                  NOP
+	ADD.y VF10, VF00, VF12                   NOP
+	ADD.xz VF10, VF00, VF00                  NOP
+	ADD.z VF11, VF00, VF12                   IADDI VI04, VI00, 0x0
+	ADD.xy VF11, VF00, VF00                  IADD VI02, VI00, VI00
+	OPMULA.xyz ACC, VF06, VF09               NOP
+	OPMSUB.xyz VF01, VF09, VF06              NOP
+	OPMULA.xyz ACC, VF06, VF10               NOP
+	OPMSUB.xyz VF02, VF10, VF06              NOP
+	OPMULA.xyz ACC, VF06, VF11               NOP
+	OPMSUB.xyz VF03, VF11, VF06              SQI.xyzw VF01, (VI02++)
+	OPMULA.xyz ACC, VF07, VF09               NOP
+	OPMSUB.xyz VF01, VF09, VF07              SQI.xyzw VF02, (VI02++)
+	OPMULA.xyz ACC, VF07, VF10               NOP
+	OPMSUB.xyz VF02, VF10, VF07              SQI.xyzw VF03, (VI02++)
+	OPMULA.xyz ACC, VF07, VF11               NOP
+	OPMSUB.xyz VF03, VF11, VF07              SQI.xyzw VF01, (VI02++)
+	OPMULA.xyz ACC, VF08, VF09               NOP
+	OPMSUB.xyz VF01, VF09, VF08              SQI.xyzw VF02, (VI02++)
+	OPMULA.xyz ACC, VF08, VF10               NOP
+	OPMSUB.xyz VF02, VF10, VF08              SQI.xyzw VF03, (VI02++)
+	OPMULA.xyz ACC, VF08, VF11               LOI 0.5
+	OPMSUB.xyz VF01, VF11, VF08              SQI.xyzw VF01, (VI02++)
+	MULi.xyz VF06, VF06, I                   NOP
+	MULi.xyz VF07, VF07, I                   SQI.xyzw VF02, (VI02++)
+	MULi.xyz VF08, VF08, I                   NOP
+	MUL.xyz VF02, VF21, VF01                 NOP
+	MUL.xyz VF03, VF12, VF01                 NOP
+	MUL.xyz VF09, VF06, VF01                 NOP
+	MUL.xyz VF10, VF07, VF01                 NOP
+	MUL.xyz VF11, VF08, VF01                 NOP
+	ABS.xyz VF03, VF03                       NOP
+	ADDy.x VF05, VF09, VF09                  NOP
+	ADDx.y VF05, VF10, VF10                  NOP
+	ADDx.z VF05, VF11, VF11                  NOP
+	NOP                                      NOP
+EdgePairLoop:
+	ADDz.x VF05, VF05, VF09                  NOP
+	ADDz.y VF05, VF05, VF10                  NOP
+	ADDy.z VF05, VF05, VF11                  NOP
+	MULAx.w ACC, VF00, VF02                  IADD VI03, VI02, VI00
+	MADDAy.w ACC, VF00, VF02                 LQD.xyzw VF01, (--VI02)
+	MADDz.w VF02, VF00, VF02                 NOP
+	ABS.xyz VF05, VF05                       NOP
+	MULAx.w ACC, VF00, VF03                  NOP
+	MADDAy.w ACC, VF00, VF03                 NOP
+	MADDAz.w ACC, VF00, VF03                 NOP
+	MADDAx.w ACC, VF00, VF05                 NOP
+	MADDAy.w ACC, VF00, VF05                 NOP
+	MADDz.w VF03, VF00, VF05                 NOP
+	ADDw.x VF04, VF00, VF02                  NOP
+	MUL.xyz VF02, VF21, VF01                 NOP
+	MUL.xyz VF03, VF12, VF01                 NOP
+	MUL.xyz VF09, VF06, VF01                 NOP
+	CLIPw.xyz VF04, VF03                     NOP
+	MUL.xyz VF10, VF07, VF01                 NOP
+	MUL.xyz VF11, VF08, VF01                 NOP
+	ABS.xyz VF03, VF03                       NOP
+	ADDy.x VF05, VF09, VF09                  FCAND VI01, 0x3
+	ADDx.y VF05, VF10, VF10                  IBNE VI01, VI00, QuitAndFail2
+	ADDx.z VF05, VF11, VF11                  NOP
+	NOP                                      IBNE VI03, VI00, EdgePairLoop
+	NOP                                      NOP
+	NOP[E]                                   IADDIU VI01, VI00, 0x1
+	NOP                                      NOP
+
+EndOfMicrocode2:
diff --git a/src/math/Matrix.h b/src/math/Matrix.h
index da20ea18..73870fe5 100644
--- a/src/math/Matrix.h
+++ b/src/math/Matrix.h
@@ -90,6 +90,7 @@ public:
 		m_matrix.pos.z += rhs.m_matrix.pos.z;
 		return *this;
 	}
+	CMatrix& operator*=(CMatrix const &rhs);
 
 	const CVector &GetPosition(void) const { return *(CVector*)&m_matrix.pos; }
 	CVector& GetPosition(void) { return *(CVector*)&m_matrix.pos; }
diff --git a/src/math/VuVector.h b/src/math/VuVector.h
new file mode 100644
index 00000000..f90818e0
--- /dev/null
+++ b/src/math/VuVector.h
@@ -0,0 +1,41 @@
+#pragma once
+
+class TYPEALIGN(16) CVuVector : public CVector
+{
+public:
+	float w;
+	CVuVector(void) {}
+	CVuVector(float x, float y, float z) : CVector(x, y, z) {}
+	CVuVector(float x, float y, float z, float w) : CVector(x, y, z), w(w) {}
+	CVuVector(const CVector &v) : CVector(v.x, v.y, v.z) {}
+#ifdef RWCORE_H
+	CVuVector(const RwV3d &v) : CVector(v.x, v.y, v.z) {}
+
+	operator RwV3d (void) const {
+		RwV3d vecRw = { this->x, this->y, this->z };
+		return vecRw;
+	}
+	
+	operator RwV3d *(void) {
+		return (RwV3d*)this;
+	}
+#endif
+/*
+	void Normalise(void) {
+		float sq = MagnitudeSqr();
+		// TODO: VU0 code
+		if(sq > 0.0f){
+			float invsqrt = RecipSqrt(sq);
+			x *= invsqrt;
+			y *= invsqrt;
+			z *= invsqrt;
+		}else
+			x = 1.0f;
+	}
+*/
+};
+
+void TransformPoint(CVuVector &out, const CMatrix &mat, const CVuVector &in);
+void TransformPoint(CVuVector &out, const CMatrix &mat, const RwV3d &in);
+void TransformPoints(CVuVector *out, int n, const CMatrix &mat, const RwV3d *in, int stride);
+void TransformPoints(CVuVector *out, int n, const CMatrix &mat, const CVuVector *in);
diff --git a/src/math/math.cpp b/src/math/math.cpp
index 0661dcac..d231b147 100644
--- a/src/math/math.cpp
+++ b/src/math/math.cpp
@@ -1,9 +1,124 @@
 #include "common.h"
 
 #include "Quaternion.h"
+#include "VuVector.h"
 
 // TODO: move more stuff into here
 
+
+void TransformPoint(CVuVector &out, const CMatrix &mat, const CVuVector &in)
+{
+#ifdef GTA_PS2
+	__asm__ __volatile__("\n\
+		lqc2    vf01,0x0(%2)\n\
+		lqc2    vf02,0x0(%1)\n\
+		lqc2    vf03,0x10(%1)\n\
+		lqc2    vf04,0x20(%1)\n\
+		lqc2    vf05,0x30(%1)\n\
+		vmulax.xyz	ACC,   vf02,vf01\n\
+		vmadday.xyz	ACC,   vf03,vf01\n\
+		vmaddaz.xyz	ACC,   vf04,vf01\n\
+		vmaddw.xyz	vf06,vf05,vf00\n\
+		sqc2    vf06,0x0(%0)\n\
+		": : "r" (&out) , "r" (&mat) ,"r" (&in): "memory");
+#else
+	out = mat * in;
+#endif
+}
+
+void TransformPoint(CVuVector &out, const CMatrix &mat, const RwV3d &in)
+{
+#ifdef GTA_PS2
+	__asm__ __volatile__("\n\
+		ldr	$8,0x0(%2)\n\
+		ldl	$8,0x7(%2)\n\
+		lw	$9,0x8(%2)\n\
+		pcpyld	$10,$9,$8\n\
+		qmtc2	$10,vf01\n\
+		lqc2    vf02,0x0(%1)\n\
+		lqc2    vf03,0x10(%1)\n\
+		lqc2    vf04,0x20(%1)\n\
+		lqc2    vf05,0x30(%1)\n\
+		vmulax.xyz	ACC,   vf02,vf01\n\
+		vmadday.xyz	ACC,   vf03,vf01\n\
+		vmaddaz.xyz	ACC,   vf04,vf01\n\
+		vmaddw.xyz	vf06,vf05,vf00\n\
+		sqc2    vf06,0x0(%0)\n\
+		": : "r" (&out) , "r" (&mat) ,"r" (&in): "memory");
+#else
+	out = mat * in;
+#endif
+}
+
+void TransformPoints(CVuVector *out, int n, const CMatrix &mat, const RwV3d *in, int stride)
+{
+#ifdef GTA_PS3
+	__asm__ __volatile__("\n\
+		paddub	$3,%4,$0\n\
+		lqc2    vf02,0x0(%2)\n\
+		lqc2    vf03,0x10(%2)\n\
+		lqc2    vf04,0x20(%2)\n\
+		lqc2    vf05,0x30(%2)\n\
+		ldr	$8,0x0(%3)\n\
+		ldl	$8,0x7(%3)\n\
+		lw	$9,0x8(%3)\n\
+		pcpyld	$10,$9,$8\n\
+		qmtc2	$10,vf01\n\
+	1:	vmulax.xyz	ACC,   vf02,vf01\n\
+		vmadday.xyz	ACC,   vf03,vf01\n\
+		vmaddaz.xyz	ACC,   vf04,vf01\n\
+		vmaddw.xyz	vf06,vf05,vf00\n\
+		add	%3,%3,$3\n\
+		ldr	$8,0x0(%3)\n\
+		ldl	$8,0x7(%3)\n\
+		lw	$9,0x8(%3)\n\
+		pcpyld	$10,$9,$8\n\
+		qmtc2	$10,vf01\n\
+		addi	%1,%1,-1\n\
+		addiu	%0,%0,0x10\n\
+		sqc2    vf06,-0x10(%0)\n\
+		bnez	%1,1b\n\
+		": : "r" (out) , "r" (n), "r" (&mat), "r" (in), "r" (stride): "memory");
+#else
+	while(n--){
+		*out = mat * *in;
+		in = (RwV3d*)((uint8*)in + stride);
+		out++;
+	}
+#endif
+}
+
+void TransformPoints(CVuVector *out, int n, const CMatrix &mat, const CVuVector *in)
+{
+#ifdef GTA_PS2
+	__asm__ __volatile__("\n\
+		lqc2    vf02,0x0(%2)\n\
+		lqc2    vf03,0x10(%2)\n\
+		lqc2    vf04,0x20(%2)\n\
+		lqc2    vf05,0x30(%2)\n\
+		lqc2    vf01,0x0(%3)\n\
+		nop\n\
+	1:	vmulax.xyz	ACC,   vf02,vf01\n\
+		vmadday.xyz	ACC,   vf03,vf01\n\
+		vmaddaz.xyz	ACC,   vf04,vf01\n\
+		vmaddw.xyz	vf06,vf05,vf00\n\
+		lqc2	vf01,0x10(%3)\n\
+		addiu	%3,%3,0x10\n\
+		addi	%1,%1,-1\n\
+		addiu	%0,%0,0x10\n\
+		sqc2    vf06,-0x10(%0)\n\
+		bnez	%1,1b\n\
+		": : "r" (out) , "r" (n), "r" (&mat) ,"r" (in): "memory");
+#else
+	while(n--){
+		*out = mat * *in;
+		in++;
+		out++;
+	}
+#endif
+}
+
+
 void
 CMatrix::SetRotate(float xAngle, float yAngle, float zAngle)
 {
@@ -74,6 +189,7 @@ CMatrix::Reorthogonalise(void)
 CMatrix&
 Invert(const CMatrix &src, CMatrix &dst)
 {
+	// TODO: VU0 code
 	// GTA handles this as a raw 4x4 orthonormal matrix
 	// and trashes the RW flags, let's not do that
 	// actual copy of librw code:
@@ -104,6 +220,7 @@ Invert(const CMatrix &src, CMatrix &dst)
 CVector
 operator*(const CMatrix &mat, const CVector &vec)
 {
+	// TODO: VU0 code
 	return CVector(
 		mat.m_matrix.right.x * vec.x + mat.m_matrix.up.x * vec.y + mat.m_matrix.at.x * vec.z + mat.m_matrix.pos.x,
 		mat.m_matrix.right.y * vec.x + mat.m_matrix.up.y * vec.y + mat.m_matrix.at.y * vec.z + mat.m_matrix.pos.y,
@@ -113,6 +230,7 @@ operator*(const CMatrix &mat, const CVector &vec)
 CMatrix
 operator*(const CMatrix &m1, const CMatrix &m2)
 {
+	// TODO: VU0 code
 	CMatrix out;
 	RwMatrix *dst = &out.m_matrix;
 	const RwMatrix *src1 = &m1.m_matrix;
@@ -132,9 +250,18 @@ operator*(const CMatrix &m1, const CMatrix &m2)
 	return out;
 }
 
+CMatrix&
+CMatrix::operator*=(CMatrix const &rhs)
+{
+	// TODO: VU0 code
+	*this = *this * rhs;
+	return *this;
+}
+
 const CVector
 Multiply3x3(const CMatrix &mat, const CVector &vec)
 {
+	// TODO: VU0 code
 	return CVector(
 		mat.m_matrix.right.x * vec.x + mat.m_matrix.up.x * vec.y + mat.m_matrix.at.x * vec.z,
 		mat.m_matrix.right.y * vec.x + mat.m_matrix.up.y * vec.y + mat.m_matrix.at.y * vec.z,
@@ -166,6 +293,7 @@ CQuaternion::Slerp(const CQuaternion &q1, const CQuaternion &q2, float theta, fl
 			w1 = Sin((1.0f - t) * theta) * invSin;
 			w2 = Sin(t * theta) * invSin;
 		}
+		// TODO: VU0 code
 		*this = w1*q1 + w2*q2;
 	}
 }
diff --git a/src/peds/Population.cpp b/src/peds/Population.cpp
index 70caf1f2..f7e0973f 100644
--- a/src/peds/Population.cpp
+++ b/src/peds/Population.cpp
@@ -1004,7 +1004,11 @@ CPopulation::ManagePopulation(void)
 	}
 
 	int pedPoolSize = CPools::GetPedPool()->GetSize();
+#ifndef SQUEEZE_PERFORMANCE
 	for (int poolIndex = pedPoolSize-1; poolIndex >= 0; poolIndex--) {
+#else
+	for (int poolIndex = (pedPoolSize * (frameMod32 + 1) / 32) - 1; poolIndex >= pedPoolSize * frameMod32 / 32; poolIndex--) {
+#endif
 		CPed *ped = CPools::GetPedPool()->GetSlot(poolIndex);
 
 		if (ped && !ped->IsPlayer() && ped->CanBeDeleted() && !ped->bInVehicle) {
diff --git a/src/render/Renderer.cpp b/src/render/Renderer.cpp
index 0640f25e..03ef1064 100644
--- a/src/render/Renderer.cpp
+++ b/src/render/Renderer.cpp
@@ -74,8 +74,12 @@ CRenderer::PreRender(void)
 	for(i = 0; i < ms_nNoOfVisibleEntities; i++)
 		ms_aVisibleEntityPtrs[i]->PreRender();
 
-	for(i = 0; i < ms_nNoOfInVisibleEntities; i++)
+	for (i = 0; i < ms_nNoOfInVisibleEntities; i++) {
+#ifdef SQUEEZE_PERFORMANCE
+		if (ms_aInVisibleEntityPtrs[i]->IsVehicle() && ((CVehicle*)ms_aInVisibleEntityPtrs[i])->IsHeli())
+#endif
 		ms_aInVisibleEntityPtrs[i]->PreRender();
+	}
 
 	for(node = CVisibilityPlugins::m_alphaEntityList.head.next;
 	    node != &CVisibilityPlugins::m_alphaEntityList.tail;
diff --git a/src/render/Rubbish.h b/src/render/Rubbish.h
index 52050e20..5a4e479b 100644
--- a/src/render/Rubbish.h
+++ b/src/render/Rubbish.h
@@ -4,7 +4,11 @@ class CVehicle;
 
 enum {
 	// NB: not all values are allowed, check the code
+#ifdef SQUEEZE_PERFORMANCE
+	NUM_RUBBISH_SHEETS = 32
+#else
 	NUM_RUBBISH_SHEETS = 64
+#endif
 };
 
 class COneSheet
diff --git a/src/render/Shadows.cpp b/src/render/Shadows.cpp
index c22afa76..9b864d27 100644
--- a/src/render/Shadows.cpp
+++ b/src/render/Shadows.cpp
@@ -1585,7 +1585,9 @@ CShadows::CastShadowEntityXY(CEntity *pEntity,  float fStartX, float fStartY, fl
 		CColTrianglePlane *pColTriPlanes = pCol->trianglePlanes;
 		ASSERT(pColTriPlanes != nil);
 
-		if ( Abs(pColTriPlanes[i].normal.z) > 0.1f )
+		CVector normal;
+		pColTriPlanes[i].GetNormal(normal);
+		if ( Abs(normal.z) > 0.1f )
 		{
 			CColTriangle *pColTri = pCol->triangles;
 			ASSERT(pColTri != nil);
diff --git a/src/weapons/BulletInfo.cpp b/src/weapons/BulletInfo.cpp
index 129e4b2b..41d9c730 100644
--- a/src/weapons/BulletInfo.cpp
+++ b/src/weapons/BulletInfo.cpp
@@ -24,6 +24,10 @@
 #include "World.h"
 #include "SurfaceTable.h"
 
+#ifdef SQUEEZE_PERFORMANCE
+uint32 bulletInfoInUse;
+#endif
+
 #define BULLET_LIFETIME (1000)
 #define NUM_PED_BLOOD_PARTICLES (8)
 #define BLOOD_PARTICLE_OFFSET (CVector(0.0f, 0.0f, 0.0f))
@@ -47,6 +51,9 @@ void CBulletInfo::Initialise(void)
 		gaBulletInfo[i].m_pSource = nil;
 	}
 	debug("CBulletInfo ready\n");
+#ifdef SQUEEZE_PERFORMANCE
+	bulletInfoInUse = 0;
+#endif
 }
 
 void CBulletInfo::Shutdown(void)
@@ -71,11 +78,19 @@ bool CBulletInfo::AddBullet(CEntity* pSource, eWeaponType type, CVector vecPosit
 	gaBulletInfo[i].m_vecSpeed = vecSpeed;
 	gaBulletInfo[i].m_fTimer = CTimer::GetTimeInMilliseconds() + BULLET_LIFETIME;
 	gaBulletInfo[i].m_bInUse = true;
+
+#ifdef SQUEEZE_PERFORMANCE
+	bulletInfoInUse++;
+#endif
 	return true;
 }
 
 void CBulletInfo::Update(void)
 {
+#ifdef SQUEEZE_PERFORMANCE
+	if (bulletInfoInUse == 0)
+		return;
+#endif
 	bPlayerSniperBullet = false;
 	for (int i = 0; i < NUM_BULLETS; i++) {
 		CBulletInfo* pBullet = &gaBulletInfo[i];
@@ -83,8 +98,12 @@ void CBulletInfo::Update(void)
 			pBullet->m_pSource = nil;
 		if (!pBullet->m_bInUse)
 			continue;
-		if (CTimer::GetTimeInMilliseconds() > pBullet->m_fTimer)
+		if (CTimer::GetTimeInMilliseconds() > pBullet->m_fTimer) {
 			pBullet->m_bInUse = false;
+#ifdef SQUEEZE_PERFORMANCE
+			bulletInfoInUse--;
+#endif
+		}
 		CVector vecOldPos = pBullet->m_vecPosition;
 		CVector vecNewPos = pBullet->m_vecPosition + pBullet->m_vecSpeed * CTimer::GetTimeStep() * 0.5f;
 		CWorld::bIncludeCarTyres = true;
@@ -108,6 +127,9 @@ void CBulletInfo::Update(void)
 					pPed->InflictDamage(pBullet->m_pSource, pBullet->m_eWeaponType, pBullet->m_nDamage, (ePedPieceTypes)point.pieceB, pPed->GetLocalDirection(pPed->GetPosition() - point.point));
 					CEventList::RegisterEvent(pPed->m_nPedType == PEDTYPE_COP ? EVENT_SHOOT_COP : EVENT_SHOOT_PED, EVENT_ENTITY_PED, pPed, (CPed*)pBullet->m_pSource, 1000);
 					pBullet->m_bInUse = false;
+#ifdef SQUEEZE_PERFORMANCE
+						bulletInfoInUse--;
+#endif
 					vecNewPos = point.point;
 				}
 				if (CGame::nastyGame) {
@@ -130,6 +152,9 @@ void CBulletInfo::Update(void)
 						}
 					}
 					pBullet->m_bInUse = false;
+#ifdef SQUEEZE_PERFORMANCE
+					bulletInfoInUse--;
+#endif
 					vecNewPos = point.point;
 				}
 			}
@@ -144,6 +169,9 @@ void CBulletInfo::Update(void)
 				}
 #ifdef FIX_BUGS
 				pBullet->m_bInUse = false;
+#ifdef SQUEEZE_PERFORMANCE
+				bulletInfoInUse--;
+#endif
 				vecNewPos = point.point;
 #endif
 			}
@@ -163,6 +191,9 @@ void CBulletInfo::Update(void)
 				}
 #ifdef FIX_BUGS
 				pBullet->m_bInUse = false;
+#ifdef SQUEEZE_PERFORMANCE
+				bulletInfoInUse--;
+#endif
 				vecNewPos = point.point;
 #endif
 			}
@@ -213,8 +244,12 @@ void CBulletInfo::Update(void)
 		}
 		pBullet->m_vecPosition = vecNewPos;
 		if (pBullet->m_vecPosition.x < -MAP_BORDER || pBullet->m_vecPosition.x > MAP_BORDER ||
-			pBullet->m_vecPosition.y < -MAP_BORDER || pBullet->m_vecPosition.y > MAP_BORDER)
+			pBullet->m_vecPosition.y < -MAP_BORDER || pBullet->m_vecPosition.y > MAP_BORDER) {
 			pBullet->m_bInUse = false;
+#ifdef SQUEEZE_PERFORMANCE
+			bulletInfoInUse--;
+#endif
+		}
 	}
 }
 
diff --git a/src/weapons/ProjectileInfo.cpp b/src/weapons/ProjectileInfo.cpp
index 35e55b36..4150852a 100644
--- a/src/weapons/ProjectileInfo.cpp
+++ b/src/weapons/ProjectileInfo.cpp
@@ -13,6 +13,10 @@
 #include "Weapon.h"
 #include "World.h"
 
+#ifdef SQUEEZE_PERFORMANCE
+uint32 projectileInUse;
+#endif
+
 CProjectileInfo gaProjectileInfo[NUM_PROJECTILES];
 CProjectile *CProjectileInfo::ms_apProjectile[NUM_PROJECTILES];
 
@@ -30,6 +34,10 @@ CProjectileInfo::Initialise()
 	}
 
 	debug("CProjectileInfo ready\n");
+
+#ifdef SQUEEZE_PERFORMANCE
+	projectileInUse = 0;
+#endif
 }
 
 void
@@ -163,6 +171,10 @@ CProjectileInfo::AddProjectile(CEntity *entity, eWeaponType weapon, CVector pos,
 	ms_apProjectile[i]->m_fElasticity = elasticity;
 	ms_apProjectile[i]->m_nSpecialCollisionResponseCases = SpecialCollisionResponseCase;
 
+#ifdef SQUEEZE_PERFORMANCE
+	projectileInUse++;
+#endif
+
 	gaProjectileInfo[i].m_bInUse = true;
 	CWorld::Add(ms_apProjectile[i]);
 
@@ -178,6 +190,9 @@ void
 CProjectileInfo::RemoveProjectile(CProjectileInfo *info, CProjectile *projectile)
 {
 	RemoveNotAdd(info->m_pSource, info->m_eWeaponType, projectile->GetPosition());
+#ifdef SQUEEZE_PERFORMANCE
+	projectileInUse--;
+#endif
 
 	info->m_bInUse = false;
 	CWorld::Remove(projectile);
@@ -205,6 +220,11 @@ CProjectileInfo::RemoveNotAdd(CEntity *entity, eWeaponType weaponType, CVector p
 void
 CProjectileInfo::Update()
 {
+#ifdef SQUEEZE_PERFORMANCE
+	if (projectileInUse == 0)
+		return;
+#endif
+
 	for (int i = 0; i < ARRAY_SIZE(gaProjectileInfo); i++) {
 		if (!gaProjectileInfo[i].m_bInUse) continue;
 
@@ -213,6 +233,10 @@ CProjectileInfo::Update()
 			gaProjectileInfo[i].m_pSource = nil;
 
 		if (ms_apProjectile[i] == nil) {
+#ifdef SQUEEZE_PERFORMANCE
+			projectileInUse--;
+#endif
+
 			gaProjectileInfo[i].m_bInUse = false;
 			continue;
 		}
@@ -276,6 +300,10 @@ CProjectileInfo::IsProjectileInRange(float x1, float x2, float y1, float y2, flo
 				if (pos.x >= x1 && pos.x <= x2 && pos.y >= y1 && pos.y <= y2 && pos.z >= z1 && pos.z <= z2) {
 					result = true;
 					if (remove) {
+#ifdef SQUEEZE_PERFORMANCE
+						projectileInUse--;
+#endif
+
 						gaProjectileInfo[i].m_bInUse = false;
 						CWorld::Remove(ms_apProjectile[i]);
 						delete ms_apProjectile[i];
@@ -304,8 +332,17 @@ CProjectileInfo::RemoveDetonatorProjectiles()
 void
 CProjectileInfo::RemoveAllProjectiles()
 {
+#ifdef SQUEEZE_PERFORMANCE
+	if (projectileInUse == 0)
+		return;
+#endif
+
 	for (int i = 0; i < ARRAY_SIZE(ms_apProjectile); i++) {
 		if (gaProjectileInfo[i].m_bInUse) {
+#ifdef SQUEEZE_PERFORMANCE
+			projectileInUse--;
+#endif
+
 			gaProjectileInfo[i].m_bInUse = false;
 			CWorld::Remove(ms_apProjectile[i]);
 			delete ms_apProjectile[i];
@@ -316,12 +353,21 @@ CProjectileInfo::RemoveAllProjectiles()
 bool
 CProjectileInfo::RemoveIfThisIsAProjectile(CObject *object)
 {
+#ifdef SQUEEZE_PERFORMANCE
+	if (projectileInUse == 0)
+		return false;
+#endif
+
 	int i = 0;
 	while (ms_apProjectile[i++] != object) {
 		if (i >= ARRAY_SIZE(ms_apProjectile))
 			return false;
 	}
 
+#ifdef SQUEEZE_PERFORMANCE
+	projectileInUse--;
+#endif
+
 	gaProjectileInfo[i].m_bInUse = false;
 	CWorld::Remove(ms_apProjectile[i]);
 	delete ms_apProjectile[i];
diff --git a/src/weapons/ShotInfo.cpp b/src/weapons/ShotInfo.cpp
index 6fc6341d..a70f3f7b 100644
--- a/src/weapons/ShotInfo.cpp
+++ b/src/weapons/ShotInfo.cpp
@@ -13,6 +13,9 @@
 CShotInfo gaShotInfo[NUMSHOTINFOS];
 float CShotInfo::ms_afRandTable[20];
 
+#ifdef SQUEEZE_PERFORMANCE
+uint32 shotInfoInUse;
+#endif
 
 /*
 	Used for flamethrower. I don't know why it's name is CShotInfo.
@@ -41,6 +44,9 @@ CShotInfo::Initialise()
 		nextVal += 0.005f;
 	}
 	debug("CShotInfo ready\n");
+#ifdef SQUEEZE_PERFORMANCE
+	shotInfoInUse = 0;
+#endif
 }
 
 bool
@@ -54,6 +60,10 @@ CShotInfo::AddShot(CEntity *sourceEntity, eWeaponType weapon, CVector startPos,
 	if (slot == ARRAY_SIZE(gaShotInfo))
 		return false;
 
+#ifdef SQUEEZE_PERFORMANCE
+	shotInfoInUse++;
+#endif
+
 	gaShotInfo[slot].m_inUse = true;
 	gaShotInfo[slot].m_weapon = weapon;
 	gaShotInfo[slot].m_startPos = startPos;
@@ -87,6 +97,10 @@ CShotInfo::Shutdown()
 void
 CShotInfo::Update()
 {
+#ifdef SQUEEZE_PERFORMANCE
+	if (shotInfoInUse == 0)
+		return;
+#endif
 	for (int slot = 0; slot < ARRAY_SIZE(gaShotInfo); slot++) {
 		CShotInfo &shot = gaShotInfo[slot];
 		if (shot.m_sourceEntity && shot.m_sourceEntity->IsPed() && !((CPed*)shot.m_sourceEntity)->IsPointerValid())
@@ -96,8 +110,12 @@ CShotInfo::Update()
 			continue;
 
 		CWeaponInfo *weaponInfo = CWeaponInfo::GetWeaponInfo(shot.m_weapon);
-		if (CTimer::GetTimeInMilliseconds() > shot.m_timeout)
+		if (CTimer::GetTimeInMilliseconds() > shot.m_timeout) {
+#ifdef SQUEEZE_PERFORMANCE
+			shotInfoInUse--;
+#endif
 			shot.m_inUse = false;
+		}
 
 		if (weaponInfo->m_bSlowsDown)
 			shot.m_areaAffected *= pow(0.96, CTimer::GetTimeStep()); // FRAMERATE