diff --git a/tools/ps2test/main.cpp b/tools/ps2test/main.cpp
index 5ca95ff..330400a 100755
--- a/tools/ps2test/main.cpp
+++ b/tools/ps2test/main.cpp
@@ -28,7 +28,20 @@ int *__errno() { return &errno; }
 // NONINTERLACED has half the vertical units
 
 uint128 packetbuf[128];
-uint128 chainbuf[128];
+uint128 vuXYZScale;
+uint128 vuXYZOffset;
+extern uint32 geometryCall[];
+extern uint32 skinPipe[];
+
+uint128 *curVifPtr;
+uint128 lightpacket[128];
+int32 numLightQ;
+
+
+
+rw::World *world;
+rw::Camera *camera;
+
 
 int frames;
 
@@ -40,6 +53,22 @@ printquad(uint128 p)
 	printf("%016lx %016lx\n", lp[1], lp[0]);
 }
 
+void
+printquad4(uint128 p)
+{
+	uint32 *lp;
+	lp = (uint32*)&p;
+	printf("%08x %08x %08x %08x\n", lp[0], lp[1], lp[2], lp[3]);
+}
+
+void
+dump4(uint128 *p, int n)
+{
+printf("data at %p\n", p);
+	while(n--)
+		printquad4(*p++);
+}
+
 struct DmaChannel {
 	uint32 chcr; uint32 pad0[3];
 	uint32 madr; uint32 pad1[3];
@@ -393,29 +422,6 @@ drawtri(void)
 	toGIFchain(packetbuf);
 }
 
-
-rw::Matrix projMat, viewMat, worldMat;
-
-extern uint32 MyDmaPacket[];
-extern rw::RawMatrix vuMat;
-extern rw::RawMatrix vuLightMat;
-extern float  vuXYZScale[];
-extern float  vuXYZOffset[];
-extern float  vuOffset[];
-extern uint64 vuGIFtag[];
-extern float  vuMatcolor[];
-extern float  vuSurfProps[];
-extern float  vuAmbLight[];
-extern uint32 vuGeometry[];
-extern uint32 mpgCall[];
-extern uint32 textureCall[];
-extern uint32 geometryCall[];
-extern uint32 defaultPipe[];
-extern uint32 skinPipe[];
-
-rw::World *world;
-rw::Camera *camera;
-
 void
 printMatrix(rw::Matrix *m)
 {
@@ -436,59 +442,120 @@ printMatrix(rw::Matrix *m)
                   m->flags);
 }
 
+// This is not proper data, just for testing
+void
+setupLight(rw::Atomic *atomic)
+{
+	using namespace rw;
+	Matrix *lightmat;
+	float32 *lp;
+
+	numLightQ = 0;
+	lp = (float32*)lightpacket;
+
+	// TODO: this is the wrong matrix. we actually want to
+	// transform the light, not all normals.
+	lightmat = atomic->getFrame()->getLTM();
+	*lp++ = lightmat->right.x;
+	*lp++ = lightmat->right.y;
+	*lp++ = lightmat->right.z;
+	*lp++ = 0.0f;
+	*lp++ = lightmat->up.x;
+	*lp++ = lightmat->up.y;
+	*lp++ = lightmat->up.z;
+	*lp++ = 0.0f;
+	*lp++ = lightmat->at.x;
+	*lp++ = lightmat->at.y;
+	*lp++ = lightmat->at.z;
+	*lp++ = 0.0f;
+	*lp++ = lightmat->pos.x;
+	*lp++ = lightmat->pos.y;
+	*lp++ = lightmat->pos.z;
+	*lp++ = 1.0f;
+	// TODO: make a proper light block
+	// ambient
+	*lp++ = 80.0f;
+	*lp++ = 80.0f;
+	*lp++ = 80.0f;
+	*lp++ = 0.0f;
+	// directional
+	*lp++ = 0.5f;
+	*lp++ = -0.5f;
+	*lp++ = -0.7071f;
+	*lp++ = 0.0f;
+	numLightQ = 6;
+}
+
+void
+setupTransform(rw::Atomic *atomic, rw::Matrix *trans)
+{
+	rw::Matrix::mult(trans, atomic->getFrame()->getLTM(), &camera->viewMatrix);
+}
+
+enum {
+	DMAcnt       = 0x10000000,
+	DMAref       = 0x30000000,
+	DMAcall      = 0x50000000,
+	DMAret       = 0x60000000,
+	DMAend       = 0x70000000,
+
+	V4_32 = 0x6C
+};
+
+#define UNPACK(type, nq, offset) ((type)<<24 | (nq)<<16 | (offset))
+#define STCYCL(WL,CL) (0x01000000 | (WL)<<8 | (CL))
 
 void
 drawAtomic(rw::Atomic *atomic)
 {
 	using namespace rw;
 
-	int i;
+	Matrix trans;
 	Geometry *geo;
-	Matrix *vp;
-
-	printf("view matrix\n");
-	printMatrix(&camera->viewMatrix);
-	printf("camera matrix\n");
-	printMatrix(camera->getFrame()->getLTM());
-	printf("atomic ltm\n");
-	printMatrix(atomic->getFrame()->getLTM());
-
-/*
-	RpAtomicPS2AllGetMeshHeaderMeshCache(atomic, ps2AllPipeData);
-	RpAtomicPS2AllGatherObjMetrics(atomic);
-	RpAtomicPS2AllMorphSetup(atomic, ps2AllPipeData);
-	RpAtomicPS2AllObjInstanceTest(atomic, ps2AllPipeData);
-	RpAtomicPS2AllClear(atomic);
-	RpAtomicPS2AllTransformSetup(atomic, transform);
-	RpAtomicPS2AllFrustumTest(atomic, &inFrustum);
-	RpAtomicPS2AllPrimTypeTransTypeSetup(ps2AllPipeData, inFrustum);
-	RpAtomicPS2AllMatModulateSetup(atomic, ps2AllPipeData);
-	RpAtomicPS2AllLightingSetup(ps2AllPipeData);
-*/
-
-	// Transform Setup
-	Matrix::mult((Matrix*)&vuMat, atomic->getFrame()->getLTM(), &camera->viewMatrix);
-	vuMat.rightw = vuMat.right.z;
-	vuMat.upw = vuMat.up.z;
-	vuMat.atw = vuMat.at.z;
-	vuMat.posw = vuMat.pos.z;
-	
-	*(Matrix*)&vuLightMat = *atomic->getFrame()->getLTM();
-	vuLightMat.rightw = 0.0f;
-	vuLightMat.upw = 0.0f;
-	vuLightMat.atw = 0.0f;
-	vuLightMat.posw = 1.0f;
+	ps2::ObjPipeline *pipe;
+	ps2::MatPipeline *matpipe;
+	Material *material;
+	uint128 tmp, *lp;
+	uint32 *vec;
+	RGBAf color;
+	int i;
 
 	geo = atomic->geometry;
-	if(!(geo->flags & Geometry::NATIVE)){
-		printf("not instanced!\n");
+	pipe = (ps2::ObjPipeline*)atomic->getPipeline();
+	if(pipe->platform != PLATFORM_PS2)
 		return;
-	}
 
-	vuAmbLight[0] = 80;
-	vuAmbLight[1] = 80;
-	vuAmbLight[2] = 80;
-	vuAmbLight[3] = 0;
+	setupLight(atomic);
+	setupTransform(atomic, &trans);
+
+	curVifPtr = packetbuf;
+	// upload lights
+	MAKEQ(tmp, DMAcnt | numLightQ+8, 0, STCYCL(4,4), UNPACK(V4_32, numLightQ, 0x3d0));
+	*curVifPtr++ = tmp;
+	for(lp = lightpacket; numLightQ--;)
+		*curVifPtr++ = *lp++;
+
+	// upload transformation matrix
+	MAKEQ(tmp, 0, 0, STCYCL(4,4), UNPACK(V4_32, 4, 0x3f0));
+	*curVifPtr++ = tmp;
+	vec = (uint32*)&trans.right;
+	MAKEQ(tmp, vec[0], vec[1], vec[2], vec[2]);
+	*curVifPtr++ = tmp;
+	vec = (uint32*)&trans.up;
+	MAKEQ(tmp, vec[0], vec[1], vec[2], vec[2]);
+	*curVifPtr++ = tmp;
+	vec = (uint32*)&trans.at;
+	MAKEQ(tmp, vec[0], vec[1], vec[2], vec[2]);
+	*curVifPtr++ = tmp;
+	vec = (uint32*)&trans.pos;
+	MAKEQ(tmp, vec[0], vec[1], vec[2], vec[2]);
+	*curVifPtr++ = tmp;
+
+	// upload camera/screen info
+	MAKEQ(tmp, 0, 0, STCYCL(4,4), UNPACK(V4_32, 2, 0x3f7));
+	*curVifPtr++ = tmp;
+	*curVifPtr++ = vuXYZScale;
+	*curVifPtr++ = vuXYZOffset;
 
 	assert(geo->instData != NULL);
 	rw::ps2::InstanceDataHeader *instData =
@@ -496,27 +563,49 @@ drawAtomic(rw::Atomic *atomic)
 	rw::MeshHeader *meshHeader = geo->meshHeader;
 	rw::Mesh *mesh;
 	for(i = 0; i < instData->numMeshes; i++){
-		geometryCall[1] = (uint32)instData->instanceMeshes[i].data;
-		vuMatcolor[0] = 1.0f;
-		vuMatcolor[1] = 1.0f;
-		vuMatcolor[2] = 1.0f;
-		vuMatcolor[3] = 1.0f;
+		material = instData->instanceMeshes[i].material;
+		matpipe = pipe->groupPipeline;
+		if(matpipe == nil)
+			matpipe = (ps2::MatPipeline*)material->pipeline;
+		if(matpipe == nil)
+			matpipe = ps2::defaultMatPipe;
 
-		vuGIFtag[0] = GIF_MAKE_TAG(0, 1, 1, GS_MAKE_PRIM(GS_PRIM_TRI_STRIP,1,0,0,0,0,0,0,0), GIF_PACKED, 3);
-		vuGIFtag[1] = 0x412;
-
-		geometryCall[3] = 0x020000DC;
-		mpgCall[1] = (uint32)skinPipe;
-//		geometryCall[3] = 0x02000114;
-//		mpgCall[1] = (uint32)defaultPipe;
-		toVIF1chain(MyDmaPacket);
+		// call vu code
+		MAKEQ(tmp, DMAcall, (uint32)skinPipe, 0, 0);
+		*curVifPtr++ = tmp;
+		// unpack GIF tag, material color, surface properties
+		MAKEQ(tmp, DMAcnt | 3, 0, STCYCL(4,4), UNPACK(V4_32, 3, 0x3fa));
+		*curVifPtr++ = tmp;
+		MAKE128(tmp, 0x412,
+			GIF_MAKE_TAG(0, 1, 1, GS_MAKE_PRIM(GS_PRIM_TRI_STRIP,1,0,0,0,0,0,0,0), GIF_PACKED, 3));
+		*curVifPtr++ = tmp;
+		convColor(&color, &material->color);
+		color.alpha *= 128.0f/255.0f;
+		MAKEQ(tmp, *(uint32*)&color.red, *(uint32*)&color.green,
+			*(uint32*)&color.blue, *(uint32*)&color.alpha);
+		*curVifPtr++ = tmp;
+		MAKEQ(tmp, *(uint32*)&material->surfaceProps.ambient,
+			*(uint32*)&material->surfaceProps.specular,
+			*(uint32*)&material->surfaceProps.diffuse,
+			0.0f);	// extra
+		*curVifPtr++ = tmp;
+		// call geometry
+		MAKEQ(tmp, DMAcall, (uint32)instData->instanceMeshes[i].data, 0x03000000, 0x02000000 | matpipe->vifOffset);
+		*curVifPtr++ = tmp;
 	}
+	MAKEQ(tmp, DMAend, 0, 0, 0);
+	*curVifPtr++ = tmp;
+	for(lp = packetbuf; lp < curVifPtr; lp++)
+		printquad4(*lp);
+	toVIF1chain(packetbuf);
 }
 
 void
 beginCamera(void)
 {
 	uint128 *p, tmp;
+	float32 *f;
+
 	p = packetbuf;
 	MAKE128(tmp, 0xe, GIF_MAKE_TAG(2, 1, 0, 0, GIF_PACKED, 1));
 	*p++ = tmp;
@@ -525,14 +614,16 @@ beginCamera(void)
 	MAKE128(tmp, GS_TEST_1, GS_MAKE_TEST(0, 0, 0, 0, 0, 0, 1, 2));
 	*p++ = tmp;
 	toGIF(packetbuf, 3);
-	vuXYZScale[0] = WIDTH;
-	vuXYZScale[1] = HEIGHT;
-	vuXYZScale[2] = camera->zScale;
-	vuXYZScale[3] = 0.0f;
-	vuXYZOffset[0] = 2048.0f;
-	vuXYZOffset[1] = 2048.0f;
-	vuXYZOffset[2] = camera->zShift;
-	vuXYZOffset[3] = 0.0f;
+	f = (float32*)&vuXYZScale;
+	f[0] = WIDTH;
+	f[1] = HEIGHT;
+	f[2] = camera->zScale;
+	f[3] = 0.0f;
+	f = (float32*)&vuXYZOffset;
+	f[0] = 2048.0f;
+	f[1] = 2048.0f;
+	f[2] = camera->zShift;
+	f[3] = 0.0f;
 }
 
 rw::EngineStartParams engineStartParams;
@@ -636,12 +727,9 @@ main()
 	GsPutDispCtx(&gsCtx.disp[1]);
 	// PCSX2 needs a delay for some reason
 	{ int i; for(i = 0; i < 1000000; i++); }
-	clearscreen(0, 0, 0);
-	drawtest();
-	drawtri();
-
-	vuOffset[0] = 0.0f;
-	vuOffset[1] = 0.0f;
+	clearscreen(0x80, 0x80, 0x80);
+//	drawtest();
+//	drawtri();
 
 	camera->beginUpdate();
 	beginCamera();
diff --git a/tools/ps2test/ps2.h b/tools/ps2test/ps2.h
index 86122da..ced015a 100755
--- a/tools/ps2test/ps2.h
+++ b/tools/ps2test/ps2.h
@@ -7,10 +7,10 @@ typedef unsigned int uint128 __attribute__((mode(TI)));
 
 #define MAKE128(RES,MSB,LSB) \
 	__asm__ ( "pcpyld %0, %1, %2" : "=r" (RES) : "r" ((uint64)MSB), "r" ((uint64)LSB))
-#define UINT64(LOW,HIGH) ((uint64)HIGH<<32 | (uint64)LOW)
+#define UINT64(LOW,HIGH) (((uint64)HIGH)<<32 | ((uint64)LOW))
 #define MAKEQ(RES,W0,W1,W2,W3) MAKE128(RES,UINT64(W2,W3),UINT64(W0,W1))
 
-#define BIT64(v,s) ((uint64)(v) << (s))
+#define BIT64(v,s) (((uint64)(v)) << (s))
 
 #include "mem.h"
 #include "gs.h"
diff --git a/tools/ps2test/vu/defaultpipe.dsm b/tools/ps2test/vu/defaultpipe.dsm
index bc971bf..ae3fa39 100755
--- a/tools/ps2test/vu/defaultpipe.dsm
+++ b/tools/ps2test/vu/defaultpipe.dsm
@@ -11,34 +11,15 @@
 .equ outBuf2, (outBuf1+outSize)
 
 .equ lightMat,		0x3d0
-.equ lightDir,		0x3d4
+.equ ambientLight, 	0x3d4
+.equ lightDir,		0x3d5
+
 .equ matrix, 		0x3f0
 .equ XYZScale, 		0x3f7
 .equ XYZOffset, 	0x3f8
 .equ gifTag, 		0x3fa
 .equ matColor, 		0x3fb
-.equ ambientLight, 	0x3fd
-
-/* This is the the projection matrix we start with:
- * 1/2w       0    ox/2w + 1/2   -ox/2w
- *    0   -1/2h   -oy/2h + 1/2    oy/2h
- *    0       0              1        0
- *    0       0              1        0
- * To get rid of the +1/2 in the combined matrix we
- * subtract the z-row/2 from the x- and y-rows.
- *
- * The z-row is then set to [0 0 0 1] such that multiplication
- * by XYZscale gives [0 0 0 zScale]. After perspective division
- * and addition of XYZoffset we then get zScale/w + zShift for z.
- *
- * XYZScale scales xy to the resolution and z by zScale.
- * XYZOffset translates xy to the GS coordinate system (where
- * [2048, 2048] is the center of the frame buffer) and add zShift to z.
- */
-
-; constant:
-;	VF28-VF31	transformation matrix
-;	VF25		XYZ offset
+.equ surfProps, 	0x3fc
 
 
 .balign 16,0
@@ -47,23 +28,7 @@ DMAret *
 MPG 0, *
 .vu
 Start:
-	SUB.z  VF28, VF28, VF28		LOI 0.5				; right.z = 0
-	SUB.z  VF29, VF29, VF29		LQ VF28, matrix(VI00)		; up.z = 0 - load matrix
-	SUB.z  VF30, VF30, VF30		LQ VF29, matrix+1(VI00)		; at.z = 0 - load matrix
-	ADDw.z VF31, VF00, VF00		LQ VF30, matrix+2(VI00)		; at.z = 1 - load matrix
-	NOP				LQ VF31, matrix+3(VI00)		;          - load matrix
-	MULi.w VF20, VF28, I		LQ.xyz VF01, XYZScale(VI00)	; fix matrix - load scale
-	MULi.w VF21, VF29, I		NOP				; fix matrix
-	MULi.w VF22, VF30, I		NOP				; fix matrix
-	MULi.w VF23, VF31, I		NOP				; fix matrix
-	SUBw.xy VF28, VF28, VF20	NOP				; fix matrix
-	SUBw.xy VF29, VF29, VF21	NOP				; fix matrix
-	SUBw.xy VF30, VF30, VF22	NOP				; fix matrix
-	SUBw.xy VF31, VF31, VF23	NOP				; fix matrix
-	MUL.xy  VF28, VF28, VF01	LQ.xyz VF25, XYZOffset(VI00)	; scale matrix
-	MUL.xy  VF29, VF29, VF01	IADDIU VI12, VI00, outBuf1	; scale matrix
-	MUL.xy  VF30, VF30, VF01	IADDIU VI13, VI00, outBuf2	; scale matrix
-	MUL.xyz VF31, VF31, VF01	NOP				; scale matrix
+#include "setup_persp.vu"
 Cnt:
 	NOP				XTOP VI02			; input pointer
 	NOP				LQ VF01, gifTag(VI00)
diff --git a/tools/ps2test/vu/light.vu b/tools/ps2test/vu/light.vu
index cd941d6..8f7a31c 100755
--- a/tools/ps2test/vu/light.vu
+++ b/tools/ps2test/vu/light.vu
@@ -1,94 +1,94 @@
 ; Ambient light:
-	NOP	LQ VF26, ambientLight(VI00)
-	NOP	XITOP VI01
-	NOP	IADDIU VI03, VI12, 2
+	NOP				LQ VF26, ambientLight(VI00)
+	NOP				XITOP VI01
+	NOP				IADDIU VI03, VI12, 2
 Ambloop:
-	NOP	LQ VF03, 0(VI03)	; output color
-	NOP	NOP
-	NOP	NOP
-	NOP	NOP
-	ITOF0 VF03, VF03	NOP
-	NOP	NOP
-	NOP	NOP
-	NOP	NOP
-	ADD VF03, VF03, VF26	NOP
-	NOP	NOP
-	NOP	NOP
-	NOP	NOP
-	FTOI0 VF03, VF03	NOP
-	NOP	IADDI VI01, VI01, -1
-	NOP	IADDIU VI03, VI03, numOutAttribs
-	NOP	IBNE VI01, VI00, Ambloop
-	NOP	SQ VF03, -numOutAttribs(VI03)
+	NOP				LQ VF03, 0(VI03)	; output color
+	NOP				NOP
+	NOP				NOP
+	NOP				NOP
+	ITOF0 VF03, VF03		NOP
+	NOP				NOP
+	NOP				NOP
+	NOP				NOP
+	ADD.xyz VF03, VF03, VF26	NOP
+	NOP				NOP
+	NOP				NOP
+	NOP				NOP
+	FTOI0 VF03, VF03		NOP
+	NOP				IADDI VI01, VI01, -1
+	NOP				IADDIU VI03, VI03, numOutAttribs
+	NOP				IBNE VI01, VI00, Ambloop
+	NOP				SQ VF03, -numOutAttribs(VI03)
 ; end amblight
 
 ; Direct Light
-	NOP	LQ VF26, lightDir(VI00)
-	NOP	XITOP VI01
-	NOP	XTOP VI02
-	NOP	IADDIU VI03, VI12, 2
+	NOP				LQ VF26, lightDir(VI00)
+	NOP				XITOP VI01
+	NOP				XTOP VI02
+	NOP				IADDIU VI03, VI12, 2
 	SUB.xyz VF26, VF00, VF26	NOP
 Dirloop:
-	NOP	LQ VF01, 3(VI02);	; normal
-	NOP	LQ VF02, 0(VI03);	; output color
-	NOP	NOP
-	NOP	NOP
-	MUL VF03, VF01, VF26	NOP
-	ITOF0 VF02, VF02	NOP
-	NOP	NOP
-	NOP	NOP
+	NOP				LQ VF01, 3(VI02);	; normal
+	NOP				LQ VF02, 0(VI03);	; output color
+	NOP				NOP
+	NOP				NOP
+	MUL VF03, VF01, VF26		NOP
+	ITOF0 VF02, VF02		NOP
+	NOP				NOP
+	NOP				NOP
 	ADDy.x VF03, VF03, VF03y	NOP
-	NOP	NOP
-	NOP	NOP
-	NOP	NOP
+	NOP				NOP
+	NOP				NOP
+	NOP				NOP
 	ADDz.x VF03, VF03, VF03z	NOP
-	NOP	NOP
-	NOP	NOP
-	NOP	NOP
-	MAX.x VF03, VF00, VF03	NOP	; clamp to 0
-	NOP[I]	LOI 255
-	NOP	NOP
-	NOP	NOP
-	MULi.x VF03, VF03, I	NOP
-	NOP	NOP
-	NOP	NOP
-	NOP	NOP
+	NOP				NOP
+	NOP				NOP
+	NOP				NOP
+	MAX.x VF03, VF00, VF03		NOP	; clamp to 0
+	NOP[I]				LOI 255
+	NOP				NOP
+	NOP				NOP
+	MULi.x VF03, VF03, I		NOP
+	NOP				NOP
+	NOP				NOP
+	NOP				NOP
 	ADDx.xyz VF02, VF02, VF03x	NOP
-	NOP	NOP
-	NOP	NOP
-	NOP	NOP
-	FTOI0 VF02, VF02	NOP
-	NOP	IADDI VI01, VI01, -1
-	NOP	IADDIU VI02, VI02, numInAttribs
-	NOP	IADDIU VI03, VI03, numOutAttribs
-	NOP	IBNE VI01, VI00, Dirloop
-	NOP	SQ VF02, -numOutAttribs(VI03)
+	NOP				NOP
+	NOP				NOP
+	NOP				NOP
+	FTOI0 VF02, VF02		NOP
+	NOP				IADDI VI01, VI01, -1
+	NOP				IADDIU VI02, VI02, numInAttribs
+	NOP				IADDIU VI03, VI03, numOutAttribs
+	NOP				IBNE VI01, VI00, Dirloop
+	NOP				SQ VF02, -numOutAttribs(VI03)
 ; end dirlight
 
 ; Material color and clamp
-	NOP	LQ VF27, matColor(VI00)
-	NOP	XITOP VI01
-	NOP	IADDIU VI03, VI12, 2
+	NOP				LQ VF27, matColor(VI00)
+	NOP				XITOP VI01
+	NOP				IADDIU VI03, VI12, 2
 Colorloop:
-	NOP	LQ VF03, 0(VI03)
-	NOP	NOP
-	NOP	NOP
-	NOP	NOP
-	ITOF0 VF03, VF03	NOP
-	NOP	NOP
-	NOP	NOP
-	NOP	NOP
-;;	MUL VF03, VF03, VF27	NOP
-	NOP[I]	LOI 255
-	NOP	NOP
-	NOP	NOP
-	MINIi VF03, VF03, I	NOP
-	NOP	NOP
-	NOP	NOP
-	NOP	NOP
-	FTOI0 VF03, VF03	NOP
-	NOP	IADDI VI01, VI01, -1
-	NOP	IADDIU VI03, VI03, numOutAttribs
-	NOP	IBNE VI01, VI00, Colorloop
-	NOP	SQ VF03, -numOutAttribs(VI03)
+	NOP				LQ VF03, 0(VI03)
+	NOP				NOP
+	NOP				NOP
+	NOP				NOP
+	ITOF0 VF03, VF03		NOP
+	NOP				NOP
+	NOP				NOP
+	NOP				NOP
+	MUL VF03, VF03, VF27		NOP
+	NOP[I]				LOI 255
+	NOP				NOP
+	NOP				NOP
+	MINIi VF03, VF03, I		NOP
+	NOP				NOP
+	NOP				NOP
+	NOP				NOP
+	FTOI0 VF03, VF03		NOP
+	NOP				IADDI VI01, VI01, -1
+	NOP				IADDIU VI03, VI03, numOutAttribs
+	NOP				IBNE VI01, VI00, Colorloop
+	NOP				SQ VF03, -numOutAttribs(VI03)
 ; end material color
diff --git a/tools/ps2test/vu/setup_persp.vu b/tools/ps2test/vu/setup_persp.vu
new file mode 100755
index 0000000..b9ea42f
--- /dev/null
+++ b/tools/ps2test/vu/setup_persp.vu
@@ -0,0 +1,39 @@
+/* This is the the projection matrix we start with:
+ * 1/2w       0    ox/2w + 1/2   -ox/2w
+ *    0   -1/2h   -oy/2h + 1/2    oy/2h
+ *    0       0              1        0
+ *    0       0              1        0
+ * To get rid of the +1/2 in the combined matrix we
+ * subtract the z-row/2 from the x- and y-rows.
+ *
+ * The z-row is then set to [0 0 0 1] such that multiplication
+ * by XYZscale gives [0 0 0 zScale]. After perspective division
+ * and addition of XYZoffset we then get zScale/w + zShift for z.
+ *
+ * XYZScale scales xy to the resolution and z by zScale.
+ * XYZOffset translates xy to the GS coordinate system (where
+ * [2048, 2048] is the center of the frame buffer) and add zShift to z.
+ */
+
+; constant:
+;	VF28-VF31	transformation matrix
+;	VF25		XYZ offset
+
+
+	SUB.z  VF28, VF28, VF28		LOI 0.5				; right.z = 0
+	SUB.z  VF29, VF29, VF29		LQ VF28, matrix(VI00)		; up.z = 0 - load matrix
+	SUB.z  VF30, VF30, VF30		LQ VF29, matrix+1(VI00)		; at.z = 0 - load matrix
+	ADDw.z VF31, VF00, VF00		LQ VF30, matrix+2(VI00)		; at.z = 1 - load matrix
+	NOP				LQ VF31, matrix+3(VI00)		;          - load matrix
+	MULi.w VF20, VF28, I		LQ.xyz VF01, XYZScale(VI00)	; fix matrix - load scale
+	MULi.w VF21, VF29, I		NOP				; fix matrix
+	MULi.w VF22, VF30, I		NOP				; fix matrix
+	MULi.w VF23, VF31, I		NOP				; fix matrix
+	SUBw.xy VF28, VF28, VF20	NOP				; fix matrix
+	SUBw.xy VF29, VF29, VF21	NOP				; fix matrix
+	SUBw.xy VF30, VF30, VF22	NOP				; fix matrix
+	SUBw.xy VF31, VF31, VF23	NOP				; fix matrix
+	MUL.xy  VF28, VF28, VF01	LQ.xyz VF25, XYZOffset(VI00)	; scale matrix
+	MUL.xy  VF29, VF29, VF01	IADDIU VI12, VI00, outBuf1	; scale matrix
+	MUL.xy  VF30, VF30, VF01	IADDIU VI13, VI00, outBuf2	; scale matrix
+	MUL.xyz VF31, VF31, VF01	NOP				; scale matrix
diff --git a/tools/ps2test/vu/skinpipe.dsm b/tools/ps2test/vu/skinpipe.dsm
index a3a8122..18536cd 100755
--- a/tools/ps2test/vu/skinpipe.dsm
+++ b/tools/ps2test/vu/skinpipe.dsm
@@ -11,34 +11,15 @@
 .equ outBuf2, (outBuf1+outSize)
 
 .equ lightMat,		0x3d0
-.equ lightDir,		0x3d4
+.equ ambientLight, 	0x3d4
+.equ lightDir,		0x3d5
+
 .equ matrix, 		0x3f0
 .equ XYZScale, 		0x3f7
 .equ XYZOffset, 	0x3f8
 .equ gifTag, 		0x3fa
 .equ matColor, 		0x3fb
-.equ ambientLight, 	0x3fd
-
-/* This is the the projection matrix we start with:
- * 1/2w       0    ox/2w + 1/2   -ox/2w
- *    0   -1/2h   -oy/2h + 1/2    oy/2h
- *    0       0              1        0
- *    0       0              1        0
- * To get rid of the +1/2 in the combined matrix we
- * subtract the z-row/2 from the x- and y-rows.
- *
- * The z-row is then set to [0 0 0 1] such that multiplication
- * by XYZscale gives [0 0 0 zScale]. After perspective division
- * and addition of XYZoffset we then get zScale/w + zShift for z.
- *
- * XYZScale scales xy to the resolution and z by zScale.
- * XYZOffset translates xy to the GS coordinate system (where
- * [2048, 2048] is the center of the frame buffer) and add zShift to z.
- */
-
-; constant:
-;	VF28-VF31	transformation matrix
-;	VF25		XYZ offset
+.equ surfProps, 	0x3fc
 
 
 .balign 16,0
@@ -47,23 +28,7 @@ DMAret *
 MPG 0, *
 .vu
 Start:
-	SUB.z  VF28, VF28, VF28		LOI 0.5				; right.z = 0
-	SUB.z  VF29, VF29, VF29		LQ VF28, matrix(VI00)		; up.z = 0 - load matrix
-	SUB.z  VF30, VF30, VF30		LQ VF29, matrix+1(VI00)		; at.z = 0 - load matrix
-	ADDw.z VF31, VF00, VF00		LQ VF30, matrix+2(VI00)		; at.z = 1 - load matrix
-	NOP				LQ VF31, matrix+3(VI00)		;          - load matrix
-	MULi.w VF20, VF28, I		LQ.xyz VF01, XYZScale(VI00)	; fix matrix - load scale
-	MULi.w VF21, VF29, I		NOP				; fix matrix
-	MULi.w VF22, VF30, I		NOP				; fix matrix
-	MULi.w VF23, VF31, I		NOP				; fix matrix
-	SUBw.xy VF28, VF28, VF20	NOP				; fix matrix
-	SUBw.xy VF29, VF29, VF21	NOP				; fix matrix
-	SUBw.xy VF30, VF30, VF22	NOP				; fix matrix
-	SUBw.xy VF31, VF31, VF23	NOP				; fix matrix
-	MUL.xy  VF28, VF28, VF01	LQ.xyz VF25, XYZOffset(VI00)	; scale matrix
-	MUL.xy  VF29, VF29, VF01	IADDIU VI12, VI00, outBuf1	; scale matrix
-	MUL.xy  VF30, VF30, VF01	IADDIU VI13, VI00, outBuf2	; scale matrix
-	MUL.xyz VF31, VF31, VF01	NOP				; scale matrix
+#include "setup_persp.vu"
 Cnt:
 	NOP				XTOP VI02			; input pointer
 	NOP				LQ VF01, gifTag(VI00)
diff --git a/tools/ps2test/vu/vu.dsm b/tools/ps2test/vu/vu.dsm
deleted file mode 100755
index 58c10d6..0000000
--- a/tools/ps2test/vu/vu.dsm
+++ /dev/null
@@ -1,98 +0,0 @@
-.data
-.global MyDmaPacket
-.global vuLightMat
-.global vuMat
-.global vuXYZScale
-.global vuXYZOffset
-.global vuOffset
-.global vuGIFtag
-.global vuMatcolor
-.global vuSurfProps
-.global vuAmbLight
-.global mpgCall
-.global textureCall
-.global geometryCall
-
-.align 4
-
-MyDmaPacket:
-DMAcnt *
-.EndDmaData
-
-mpgCall:
-DMAcall *, 0 ;vuProg
-.EndDmaData
-
-/*
-#define vuSDLightOffset     0x3d0
-#define vuSDBlockLow        0x3c3
-#define vuSDBlockHigh       0x3f0
-#define vuSDmat0            vuSDBlockHigh
-#define vuSDmat1            vuSDBlockHigh+1
-#define vuSDmat2            vuSDBlockHigh+2
-#define vuSDmat3            vuSDBlockHigh+3
-#define vuSDnearClip        vuSDBlockHigh+4
-#define vuSDfarClip         vuSDBlockHigh+5
-#define vuSDxMaxyMax        vuSDBlockHigh+6
-#define vuSDcamWcamHzScale  vuSDBlockHigh+7
-#define vuSDoffXoffYzShift  vuSDBlockHigh+8
-#define vuSDrealOffset      vuSDBlockHigh+9
-#define vuSDgifTag          vuSDBlockHigh+10
-#define vuSDcolScale        vuSDBlockHigh+11
-#define vuSDsurfProps       vuSDBlockHigh+12
-#define vuSDpingPongAddr    vuSDBlockHigh+13    
-#define vuSDpingPongCount   vuSDBlockHigh+14    
-#define vuSDClipvec1        vuSDBlockHigh+13                                     
-#define vuSDClipvec2        vuSDBlockHigh+14                                        
-#define vuSDVUSwitch        vuSDBlockHigh+15    
-*/
-
-DMAcnt *
-	UNPACK 4, 4, V4_32, 0x3d0, *
-vuLightMat:
-	.float 1.0, 0.0, 0.0, 0.0
-	.float 0.0, 1.0, 0.0, 0.0
-	.float 0.0, 0.0, 1.0, 0.0
-	.float 0.0, 0.0, 0.0, 1.0
-lightDir:
-	.float 0.5, -0.5, -0.70710, 0.0
-	.EndUnpack
-
-	UNPACK 4, 4, V4_32, 0x3f0, *
-vuMat:
-	.float 0.0, 0.0, 0.0, 0.0
-	.float 0.0, 0.0, 0.0, 0.0
-	.float 0.0, 0.0, 0.0, 0.0
-	.float 0.0, 0.0, 0.0, 0.0
-	.EndUnpack
-
-	UNPACK 4, 4, V4_32, 0x3f7, *
-vuXYZScale:
-	.float 0.0, 0.0, 0.0, 0.0
-vuXYZOffset:
-	.float 0.0, 0.0, 0.0, 0.0
-vuOffset:
-	.float 0.0, 0.0, 0.0, 0.0
-vuGIFtag:
-	.int 0x00008000, 0x3005C000, 0x0000000412, 0x00000000
-vuMatcolor:
-	.float 1.0, 1.0, 1.0, 0.5
-vuSurfProps:
-	.float 1.0, 1.0, 1.0, 1.0
-vuAmbLight:
-	.float 0, 0, 0, 0
-	.EndUnpack
-.EndDmaData
-
-;;textureCall:
-;;DMAcall *, 0
-;;.EndDmaData
-
-geometryCall:
-DMAcall *, 0 ;vuGeometry
-	BASE 0
-	OFFSET 0x0
-.EndDmaData
-
-DMAend
-