diff --git a/src/charset.cpp b/src/charset.cpp
index 601115e..f9208b3 100644
--- a/src/charset.cpp
+++ b/src/charset.cpp
@@ -56,7 +56,9 @@ void
 Charset::close(void)
 {
 	rwFree(indices);
+	indices = nil;
 	rwFree(vertices);
+	vertices = nil;
 }
 
 Charset*
diff --git a/src/d3d/d3d9matfx.cpp b/src/d3d/d3d9matfx.cpp
index 500107e..35a4adf 100644
--- a/src/d3d/d3d9matfx.cpp
+++ b/src/d3d/d3d9matfx.cpp
@@ -159,7 +159,7 @@ matfxRenderCB_Shader(Atomic *atomic, InstanceDataHeader *header)
 	bool normals = !!(atomic->geometry->flags & Geometry::NORMALS);
 
 	float surfProps[4];
-	surfProps[3] = atomic->geometry->flags&Geometry::PRELIT ? 1.0f : 0.0f;
+	surfProps[3] = 0.0f;
 
 	InstanceData *inst = header->inst;
 	for(uint32 i = 0; i < header->numMeshes; i++){
diff --git a/src/d3d/d3d9render.cpp b/src/d3d/d3d9render.cpp
index 494cc92..e182e98 100644
--- a/src/d3d/d3d9render.cpp
+++ b/src/d3d/d3d9render.cpp
@@ -89,10 +89,9 @@ defaultRenderCB_Fix(Atomic *atomic, InstanceDataHeader *header)
 	convMatrix(&world, f->getLTM());
 	d3ddevice->SetTransform(D3DTS_WORLD, (D3DMATRIX*)&world);
 
-	setStreamSource(0, (IDirect3DVertexBuffer9*)header->vertexStream[0].vertexBuffer,
-	                           0, header->vertexStream[0].stride);
-	setIndices((IDirect3DIndexBuffer9*)header->indexBuffer);
-	setVertexDeclaration((IDirect3DVertexDeclaration9*)header->vertexDeclaration);
+	setStreamSource(0, header->vertexStream[0].vertexBuffer, 0, header->vertexStream[0].stride);
+	setIndices(header->indexBuffer);
+	setVertexDeclaration(header->vertexDeclaration);
 
 	InstanceData *inst = header->inst;
 	for(uint32 i = 0; i < header->numMeshes; i++){
@@ -145,10 +144,9 @@ void
 defaultRenderCB_Shader(Atomic *atomic, InstanceDataHeader *header)
 {
 	int vsBits;
-	setStreamSource(0, (IDirect3DVertexBuffer9*)header->vertexStream[0].vertexBuffer,
-	                           0, header->vertexStream[0].stride);
-	setIndices((IDirect3DIndexBuffer9*)header->indexBuffer);
-	setVertexDeclaration((IDirect3DVertexDeclaration9*)header->vertexDeclaration);
+	setStreamSource(0, header->vertexStream[0].vertexBuffer, 0, header->vertexStream[0].stride);
+	setIndices(header->indexBuffer);
+	setVertexDeclaration(header->vertexDeclaration);
 
 	vsBits = lightingCB_Shader(atomic);
 	uploadMatrices(atomic->getFrame()->getLTM());
@@ -165,7 +163,7 @@ defaultRenderCB_Shader(Atomic *atomic, InstanceDataHeader *header)
 		setVertexShader(default_all_VS);
 
 	float surfProps[4];
-	surfProps[3] = atomic->geometry->flags&Geometry::PRELIT ? 1.0f : 0.0f;
+	surfProps[3] = 0.0f;
 
 	InstanceData *inst = header->inst;
 	for(uint32 i = 0; i < header->numMeshes; i++){
diff --git a/src/d3d/d3d9skin.cpp b/src/d3d/d3d9skin.cpp
index b99d795..e45d68a 100644
--- a/src/d3d/d3d9skin.cpp
+++ b/src/d3d/d3d9skin.cpp
@@ -288,7 +288,7 @@ skinRenderCB(Atomic *atomic, InstanceDataHeader *header)
 		setVertexShader(skin_all_VS);
 
 	float surfProps[4];
-	surfProps[3] = atomic->geometry->flags&Geometry::PRELIT ? 1.0f : 0.0f;
+	surfProps[3] = 0.0f;
 
 	InstanceData *inst = header->inst;
 	for(uint32 i = 0; i < header->numMeshes; i++){
diff --git a/src/d3d/d3dimmed.cpp b/src/d3d/d3dimmed.cpp
index 7a5e47b..db18af3 100644
--- a/src/d3d/d3dimmed.cpp
+++ b/src/d3d/d3dimmed.cpp
@@ -37,6 +37,8 @@ static IDirect3DVertexDeclaration9 *im2ddecl;
 static IDirect3DVertexBuffer9 *im2dvertbuf;
 static IDirect3DIndexBuffer9 *im2dindbuf;
 
+void *im2dOverridePS;
+
 void
 openIm2D(void)
 {
@@ -111,7 +113,9 @@ im2DRenderPrimitive(PrimitiveType primType, void *vertices, int32 numVertices)
 	setStreamSource(0, im2dvertbuf, 0, sizeof(Im2DVertex));
 	setVertexDeclaration(im2ddecl);
 
-	if(engine->device.getRenderState(TEXTURERASTER))
+	if(im2dOverridePS)
+		setPixelShader(im2dOverridePS);
+	else if(engine->device.getRenderState(TEXTURERASTER))
 		setPixelShader(im2d_tex_PS);
 	else
 		setPixelShader(im2d_PS);
@@ -163,7 +167,9 @@ im2DRenderIndexedPrimitive(PrimitiveType primType,
 	setIndices(im2dindbuf);
 	setVertexDeclaration(im2ddecl);
 
-	if(engine->device.getRenderState(TEXTURERASTER))
+	if(im2dOverridePS)
+		setPixelShader(im2dOverridePS);
+	else if(engine->device.getRenderState(TEXTURERASTER))
 		setPixelShader(im2d_tex_PS);
 	else
 		setPixelShader(im2d_PS);
diff --git a/src/d3d/rwd3d.h b/src/d3d/rwd3d.h
index 06bd6ba..9ee270d 100644
--- a/src/d3d/rwd3d.h
+++ b/src/d3d/rwd3d.h
@@ -263,6 +263,8 @@ void uploadMatrices(void);	// no world transform
 void uploadMatrices(Matrix *worldMat);
 int32 uploadLights(WorldLights *lightData);	// called by lightingCB_Shader
 
+extern void *im2dOverridePS;
+
 extern void *default_amb_VS;
 extern void *default_amb_dir_VS;
 extern void *default_all_VS;
diff --git a/src/d3d/shaders/matfx_env_VS.hlsl b/src/d3d/shaders/matfx_env_VS.hlsl
index 4521855..73179ab 100644
--- a/src/d3d/shaders/matfx_env_VS.hlsl
+++ b/src/d3d/shaders/matfx_env_VS.hlsl
@@ -23,11 +23,11 @@ VS_out main(in VS_in input)
 	VS_out output;
 
 	output.Position = mul(combinedMat, input.Position);
-	float3 Vertex = mul(worldMat, input.Position).xyz;
-	float3 Normal = mul(normalMat, input.Normal);
+	float3 V = mul(worldMat, input.Position).xyz;
+	float3 N = mul(normalMat, input.Normal);
 
 	output.TexCoord0.xy = input.TexCoord;
-	output.TexCoord1 = mul(texMat, float4(Normal, 1.0)).xy;
+	output.TexCoord1 = mul(texMat, float4(N, 1.0)).xy;
 
 	output.Color = input.Prelight;
 	output.Color.rgb += ambientLight.rgb * surfAmbient;
@@ -35,15 +35,15 @@ VS_out main(in VS_in input)
 	int i;
 #ifdef DIRECTIONALS
 	for(i = 0; i < numDirLights; i++)
-		output.Color.xyz += DoDirLight(lights[i+firstDirLight], Normal)*surfDiffuse;
+		output.Color.xyz += DoDirLight(lights[i+firstDirLight], N)*surfDiffuse;
 #endif
 #ifdef POINTLIGHTS
 	for(i = 0; i < numPointLights; i++)
-		output.Color.xyz += DoPointLight(lights[i+firstPointLight], Vertex.xyz, Normal)*surfDiffuse;
+		output.Color.xyz += DoPointLight(lights[i+firstPointLight], V, N)*surfDiffuse;
 #endif
 #ifdef SPOTLIGHTS
 	for(i = 0; i < numSpotLights; i++)
-		output.Color.xyz += DoSpotLight(lights[i+firstSpotLight], Vertex.xyz, Normal)*surfDiffuse;
+		output.Color.xyz += DoSpotLight(lights[i+firstSpotLight], V, N)*surfDiffuse;
 #endif
 	// PS2 clamps before material color
 	output.Color = clamp(output.Color, 0.0, 1.0);
diff --git a/src/engine.cpp b/src/engine.cpp
index 052dab6..310c89d 100644
--- a/src/engine.cpp
+++ b/src/engine.cpp
@@ -293,6 +293,10 @@ Engine::term(void)
 
 	PluginList::close();
 
+	// This has to be reset because it won't be opened again otherwise
+	// TODO: maybe reset more stuff here?
+	d3d::nativeRasterOffset = 0;
+
 #ifdef TRACK_ALLOCATIONS
 	FORLIST(lnk, allocations){
 		MemoryBlock *mem = LLLinkGetData(lnk, MemoryBlock, inAllocList);
diff --git a/src/gl/gl3device.cpp b/src/gl/gl3device.cpp
index 328f171..53c5446 100644
--- a/src/gl/gl3device.cpp
+++ b/src/gl/gl3device.cpp
@@ -62,10 +62,13 @@ struct UniformState
 {
 	int32   alphaFunc;
 	float32 alphaRef;
-	int32   fogEnable;
 	float32 fogStart;
 	float32 fogEnd;
-	int32   pad[3];
+
+	float32 fogRange;
+	float32 fogDisable;
+	int32   pad[2];
+
 	RGBAf   fogColor;
 };
 
@@ -77,14 +80,10 @@ struct UniformScene
 
 struct UniformLight
 {
-	V3d     position;
-	float32 w;
-	V3d     direction;
-	int32   pad1;
-	RGBAf   color;
-	float32 radius;
-	float32 minusCosAngle;
-	int32   pad2[2];
+	float32 enabled, radius, minusCosAngle, hardSpot;
+	V3d position; int32 pad0;
+	V3d direction; int32 pad1;
+	RGBAf color;
 };
 
 #define MAX_LIGHTS 8
@@ -93,9 +92,9 @@ struct UniformObject
 {
 	RawMatrix    world;
 	RGBAf        ambLight;
-	int32        numLights;
-	int32        pad[3];
-	UniformLight lights[MAX_LIGHTS];
+	UniformLight directLights[MAX_LIGHTS];
+	UniformLight pointLights[MAX_LIGHTS];
+	UniformLight spotLights[MAX_LIGHTS];
 };
 
 static GLuint vao;
@@ -106,9 +105,9 @@ static UniformScene uniformScene;
 static UniformObject uniformObject;
 
 int32 u_matColor;
-int32 u_surfaceProps;
+int32 u_surfProps;
 
-Shader *simpleShader;
+Shader *defaultShader;
 
 static bool32 stateDirty = 1;
 static bool32 sceneDirty = 1;
@@ -133,6 +132,9 @@ struct RwStateCache {
 	uint32 zwrite;
 	uint32 ztest;
 	uint32 cullmode;
+	uint32 fogEnable;
+	float32 fogStart;
+	float32 fogEnd;
 	RwRasterStateCache texstage[MAXNUMSTAGES];
 };
 static RwStateCache rwStateCache;
@@ -155,6 +157,41 @@ static uint32 blendMap[] = {
 	GL_SRC_ALPHA_SATURATE,
 };
 
+static void
+setDepthTest(bool32 enable)
+{
+	if(rwStateCache.ztest != enable){
+		rwStateCache.ztest = enable;
+		if(rwStateCache.zwrite && !enable){
+			// If we still want to write, enable but set mode to always
+			glEnable(GL_DEPTH_TEST);
+			glDepthFunc(GL_ALWAYS);
+		}else{
+			if(rwStateCache.ztest)
+				glEnable(GL_DEPTH_TEST);
+			else
+				glDisable(GL_DEPTH_TEST);
+			glDepthFunc(GL_LEQUAL);
+		}
+
+	}
+}
+
+static void
+setDepthWrite(bool32 enable)
+{
+	enable = enable ? GL_TRUE : GL_FALSE;
+	if(rwStateCache.zwrite != enable){
+		rwStateCache.zwrite = enable;
+		if(enable && !rwStateCache.ztest){
+			// Have to switch on ztest so writing can work
+			glEnable(GL_DEPTH_TEST);
+			glDepthFunc(GL_ALWAYS);
+		}
+		glDepthMask(rwStateCache.zwrite);
+	}
+}
+
 static void
 setAlphaTest(bool32 enable)
 {
@@ -412,23 +449,14 @@ setRenderState(int32 state, void *pvalue)
 		}
 		break;
 	case ZTESTENABLE:
-		if(rwStateCache.ztest != value){
-			rwStateCache.ztest = value;
-			if(rwStateCache.ztest)
-				glEnable(GL_DEPTH_TEST);
-			else
-				glDisable(GL_DEPTH_TEST);
-		}
+		setDepthTest(value);
 		break;
 	case ZWRITEENABLE:
-		if(rwStateCache.zwrite != (value ? GL_TRUE : GL_FALSE)){
-			rwStateCache.zwrite = value ? GL_TRUE : GL_FALSE;
-			glDepthMask(rwStateCache.zwrite);
-		}
+		setDepthWrite(value);
 		break;
 	case FOGENABLE:
-		if(uniformState.fogEnable != value){
-			uniformState.fogEnable = value;
+		if(rwStateCache.fogEnable != value){
+			rwStateCache.fogEnable = value;
 			stateDirty = 1;
 		}
 		break;
@@ -506,7 +534,7 @@ getRenderState(int32 state)
 		val = rwStateCache.zwrite;
 		break;
 	case FOGENABLE:
-		val = uniformState.fogEnable;
+		val = rwStateCache.fogEnable;
 		break;
 	case FOGCOLOR:
 		convColor(&rgba, &uniformState.fogColor);
@@ -534,8 +562,10 @@ resetRenderState(void)
 	rwStateCache.alphaFunc = ALPHAGREATEREQUAL;
 	uniformState.alphaFunc = 0;
 	uniformState.alphaRef = 10.0f/255.0f;
-	uniformState.fogEnable = 0;
+	uniformState.fogDisable = 1.0f;
 	uniformState.fogStart = 0.0f;
+	uniformState.fogEnd = 0.0f;
+	uniformState.fogRange = 0.0f;
 	uniformState.fogColor = { 1.0f, 1.0f, 1.0f, 1.0f };
 	stateDirty = 1;
 
@@ -573,40 +603,66 @@ setWorldMatrix(Matrix *mat)
 	objectDirty = 1;
 }
 
-void
-setAmbientLight(RGBAf *amb)
+int32
+setLights(WorldLights *lightData)
 {
-	uniformObject.ambLight = *amb;
-	objectDirty = 1;
-}
+	int i, np, ns;
+	Light *l;
+	int32 bits;
 
-void
-setNumLights(int32 n)
-{
-	uniformObject.numLights = n;
-	objectDirty = 1;
-}
+	uniformObject.ambLight = lightData->ambient;
+	memset(uniformObject.directLights, 0, sizeof(uniformObject.directLights));
+	memset(uniformObject.pointLights, 0, sizeof(uniformObject.pointLights));
+	memset(uniformObject.spotLights, 0, sizeof(uniformObject.spotLights));
 
-void
-setLight(int32 n, Light *light)
-{
-	UniformLight *l;
-	Frame *f;
-	Matrix *m;
+	bits = 0;
 
-	l = &uniformObject.lights[n];
-	f = light->getFrame();
-	if(f){
-		m = f->getLTM();
-		l->position  = m->pos;
-		l->direction = m->at;
+	for(i = 0; i < lightData->numDirectionals && i < 8; i++){
+		l = lightData->directionals[i];
+		uniformObject.directLights[i].enabled = 1.0f;
+		uniformObject.directLights[i].color = l->color;
+		uniformObject.directLights[i].direction = l->getFrame()->getLTM()->at;
+		bits |= VSLIGHT_POINT;
 	}
-	// light has position
-	l->w = light->getType() >= Light::POINT ? 1.0f : 0.0f;
-	l->color = light->color;
-	l->radius = light->radius;
-	l->minusCosAngle = light->minusCosAngle;
+
+	np = 0;
+	ns = 0;
+	for(i = 0; i < lightData->numLocals; i++){
+		l = lightData->locals[i];
+		switch(l->getType()){
+		case Light::POINT:
+			if(np >= 8)
+				continue;
+			uniformObject.pointLights[np].enabled = 1.0f;
+			uniformObject.pointLights[np].color = l->color;
+			uniformObject.pointLights[np].position = l->getFrame()->getLTM()->pos;
+			uniformObject.pointLights[np].radius = l->radius;
+			np++;
+			bits |= VSLIGHT_POINT;
+			break;
+
+		case Light::SPOT:
+		case Light::SOFTSPOT:
+			if(np >= 8)
+				continue;
+			uniformObject.spotLights[ns].enabled = 1.0f;
+			uniformObject.spotLights[ns].color = l->color;
+			uniformObject.spotLights[ns].position = l->getFrame()->getLTM()->pos;
+			uniformObject.spotLights[ns].direction = l->getFrame()->getLTM()->at;
+			uniformObject.spotLights[ns].radius = l->radius;
+			uniformObject.spotLights[ns].minusCosAngle = l->minusCosAngle;
+			if(l->getType() == Light::SOFTSPOT)
+				uniformObject.spotLights[ns].hardSpot = 0.0f;
+			else
+				uniformObject.spotLights[ns].hardSpot = 1.0f;
+			ns++;
+			bits |= VSLIGHT_SPOT;
+			break;
+		}
+	}
+
 	objectDirty = 1;
+	return 0;
 }
 
 void
@@ -639,6 +695,10 @@ flushCache(void)
 		sceneDirty = 0;
 	}
 	if(stateDirty){
+		uniformState.fogDisable = rwStateCache.fogEnable ? 0.0f : 1.0f;
+		uniformState.fogStart = rwStateCache.fogStart;
+		uniformState.fogEnd = rwStateCache.fogEnd;
+		uniformState.fogRange = 1.0f/(rwStateCache.fogStart - rwStateCache.fogEnd);
 		glBindBuffer(GL_UNIFORM_BUFFER, ubo_state);
 		glBufferSubData(GL_UNIFORM_BUFFER, 0, sizeof(UniformState),
 				&uniformState);
@@ -762,12 +822,12 @@ beginUpdate(Camera *cam)
 	memcpy(&cam->devProj, &proj, sizeof(RawMatrix));
 	setProjectionMatrix(proj);
 
-	if(uniformState.fogStart != cam->fogPlane){
-		uniformState.fogStart = cam->fogPlane;
+	if(rwStateCache.fogStart != cam->fogPlane){
+		rwStateCache.fogStart = cam->fogPlane;
 		stateDirty = 1;
 	}
-	if(uniformState.fogEnd != cam->farPlane){
-		uniformState.fogEnd = cam->farPlane;
+	if(rwStateCache.fogEnd != cam->farPlane){
+		rwStateCache.fogEnd = cam->farPlane;
 		stateDirty = 1;
 	}
 
@@ -993,7 +1053,7 @@ initOpenGL(void)
 	registerBlock("Object");
 	registerBlock("State");
 	u_matColor = registerUniform("u_matColor");
-	u_surfaceProps = registerUniform("u_surfaceProps");
+	u_surfProps = registerUniform("u_surfProps");
 
 	glClearColor(0.25, 0.25, 0.25, 1.0);
 
@@ -1031,9 +1091,12 @@ initOpenGL(void)
 	             GL_DYNAMIC_DRAW);
 	glBindBuffer(GL_UNIFORM_BUFFER, 0);
 
-#include "shaders/simple_vs_gl3.inc"
+#include "shaders/default_vs_gl3.inc"
 #include "shaders/simple_fs_gl3.inc"
-	simpleShader = Shader::fromStrings(simple_vert_src, simple_frag_src);
+	const char *vs[] = { header_vert_src, default_vert_src, nil };
+	const char *fs[] = { simple_frag_src, nil };
+	defaultShader = Shader::create(vs, fs);
+	assert(defaultShader);
 
 	openIm2D();
 	openIm3D();
diff --git a/src/gl/gl3immed.cpp b/src/gl/gl3immed.cpp
index 90cb69a..8cb62d9 100644
--- a/src/gl/gl3immed.cpp
+++ b/src/gl/gl3immed.cpp
@@ -1,6 +1,7 @@
 #include <cstdio>
 #include <cstdlib>
 #include <cstring>
+#include <cassert>
 
 #include "../rwbase.h"
 #include "../rwerror.h"
@@ -50,7 +51,10 @@ openIm2D(void)
 
 #include "shaders/im2d_gl3.inc"
 #include "shaders/simple_fs_gl3.inc"
-	im2dShader = Shader::fromStrings(im2d_vert_src, simple_frag_src);
+	const char *vs[] = { header_vert_src, im2d_vert_src, nil };
+	const char *fs[] = { simple_frag_src, nil };
+	im2dShader = Shader::create(vs, fs);
+	assert(im2dShader);
 
 	glGenBuffers(1, &im2DIbo);
 	glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, im2DIbo);
@@ -176,7 +180,10 @@ openIm3D(void)
 {
 #include "shaders/im3d_gl3.inc"
 #include "shaders/simple_fs_gl3.inc"
-	im3dShader = Shader::fromStrings(im3d_vert_src, simple_frag_src);
+	const char *vs[] = { header_vert_src, im3d_vert_src, nil };
+	const char *fs[] = { simple_frag_src, nil };
+	im3dShader = Shader::create(vs, fs);
+	assert(im3dShader);
 
 	glGenBuffers(1, &im3DIbo);
 	glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, im3DIbo);
diff --git a/src/gl/gl3matfx.cpp b/src/gl/gl3matfx.cpp
index 51aa60f..8f9ecba 100644
--- a/src/gl/gl3matfx.cpp
+++ b/src/gl/gl3matfx.cpp
@@ -37,9 +37,14 @@ matfxOpen(void *o, int32, int32)
 {
 	u_texMatrix = registerUniform("u_texMatrix");
 	u_coefficient = registerUniform("u_coefficient");
-#include "shaders/matfx_gl3.inc"
 	matFXGlobals.pipelines[PLATFORM_GL3] = makeMatFXPipeline();
-	envShader = Shader::fromStrings(matfx_env_vert_src, matfx_env_frag_src);
+
+#include "shaders/matfx_gl3.inc"
+	const char *vs[] = { header_vert_src, matfx_env_vert_src, nil };
+	const char *fs[] = { matfx_env_frag_src, nil };
+	envShader = Shader::create(vs, fs);
+	assert(envShader);
+
 	return o;
 }
 
@@ -65,7 +70,7 @@ matfxDefaultRender(InstanceDataHeader *header, InstanceData *inst)
 	GLfloat surfProps[4];
 	m = inst->material;
 
-	simpleShader->use();
+	defaultShader->use();
 
 	convColor(&col, &m->color);
 	glUniform4fv(U(u_matColor), 1, (GLfloat*)&col);
@@ -74,7 +79,7 @@ matfxDefaultRender(InstanceDataHeader *header, InstanceData *inst)
 	surfProps[1] = m->surfaceProps.specular;
 	surfProps[2] = m->surfaceProps.diffuse;
 	surfProps[3] = 0.0f;
-	glUniform4fv(U(u_surfaceProps), 1, surfProps);
+	glUniform4fv(U(u_surfProps), 1, surfProps);
 
 	setTexture(0, m->texture);
 
@@ -85,82 +90,54 @@ matfxDefaultRender(InstanceDataHeader *header, InstanceData *inst)
 	               GL_UNSIGNED_SHORT, (void*)(uintptr)inst->offset);
 }
 
+static Frame *lastEnvFrame;
+
+static RawMatrix normal2texcoord = {
+	{ 0.5f,  0.0f, 0.0f }, 0.0f,
+	{ 0.0f, -0.5f, 0.0f }, 0.0f,
+	{ 0.0f,  0.0f, 1.0f }, 0.0f,
+	{ 0.5f,  0.5f, 0.0f }, 1.0f
+};
+
 void
-calcEnvTexMatrix(Frame *f, float32 *mat)
+uploadEnvMatrix(Frame *frame)
 {
-	Matrix cam;
-	if(f){
-		// PS2 style - could be simplified by the shader
-		cam = *((Camera*)engine->currentCamera)->getFrame()->getLTM();
-		cam.pos = { 0.0, 0.0, 0.0 };
-		cam.right.x = -cam.right.x;
-		cam.right.y = -cam.right.y;
-		cam.right.z = -cam.right.z;
+	Matrix invMat;
+	if(frame == nil)
+		frame = engine->currentCamera->getFrame();
 
-		Matrix inv;
-		Matrix::invert(&inv, f->getLTM());
-		inv.pos = { -1.0, -1.0, -1.0 };
-		inv.right.x *= -0.5f;
-		inv.right.y *= -0.5f;
-		inv.right.z *= -0.5f;
-		inv.up.x *= -0.5f;
-		inv.up.y *= -0.5f;
-		inv.up.z *= -0.5f;
-		inv.at.x *= -0.5f;
-		inv.at.y *= -0.5f;
-		inv.at.z *= -0.5f;
-		inv.pos.x *= -0.5f;
-		inv.pos.y *= -0.5f;
-		inv.pos.z *= -0.5f;
+	// cache the matrix across multiple meshes
+	static RawMatrix envMtx;
+	if(frame != lastEnvFrame){
+		lastEnvFrame = frame;
 
-		Matrix m;
-		Matrix::mult(&m, &cam, &inv);
-
-		memcpy(mat, &m, 64);
-		mat[3] = mat[7] = mat[11] = 0.0f;
-		mat[15] = 1.0f;
-
-	}else{
-		// D3D - TODO: find out what PS2 does
-		mat[0]  = 0.5f;
-		mat[1]  = 0.0f;
-		mat[2]  = 0.0f;
-		mat[3]  = 0.0f;
-
-		mat[4]  = 0.0f;
-		mat[5]  = -0.5f;
-		mat[6]  = 0.0f;
-		mat[7]  = 0.0f;
-
-		mat[8]  = 0.0f;
-		mat[9]  = 0.0f;
-		mat[10] = 1.0f;
-		mat[11] = 0.0f;
-
-		mat[12] = 0.5f;
-		mat[13] = 0.5f;
-		mat[14] = 0.0f;
-		mat[15] = 0.0f;
+		RawMatrix invMtx;
+		Matrix::invert(&invMat, frame->getLTM());
+		convMatrix(&invMtx, &invMat);
+		RawMatrix::mult(&envMtx, &invMtx, &normal2texcoord);
 	}
+	glUniformMatrix4fv(U(u_texMatrix), 1, GL_FALSE, (float*)&envMtx);
 }
 
 void
-matfxEnvRender(InstanceDataHeader *header, InstanceData *inst)
+matfxEnvRender(InstanceDataHeader *header, InstanceData *inst, MatFX::Env *env)
 {
 	Material *m;
 	RGBAf col;
 	GLfloat surfProps[4];
-	float32 texMat[16];
 	m = inst->material;
 
-	matfxDefaultRender(header, inst);
-
-	MatFX *fx = MatFX::get(m);
-	int32 idx = fx->getEffectIndex(MatFX::ENVMAP);
-	MatFX::Env *env = &fx->fx[idx].env;
+	if(env->tex == nil || env->coefficient == 0.0f){
+		matfxDefaultRender(header, inst);
+		return;
+	}
 
 	envShader->use();
 
+	setTexture(0, m->texture);
+	setTexture(1, env->tex);
+	uploadEnvMatrix(env->frame);
+
 	convColor(&col, &m->color);
 	glUniform4fv(U(u_matColor), 1, (GLfloat*)&col);
 
@@ -168,52 +145,47 @@ matfxEnvRender(InstanceDataHeader *header, InstanceData *inst)
 	surfProps[1] = m->surfaceProps.specular;
 	surfProps[2] = m->surfaceProps.diffuse;
 	surfProps[3] = 0.0f;
-	glUniform4fv(U(u_surfaceProps), 1, surfProps);
+	glUniform4fv(U(u_surfProps), 1, surfProps);
 
 	glUniform1fv(U(u_coefficient), 1, &env->coefficient);
 
-	calcEnvTexMatrix(env->frame, texMat);
-	glUniformMatrix4fv(U(u_texMatrix), 1, GL_FALSE, texMat);
-
-	setTexture(0, env->tex);
-
 	rw::SetRenderState(VERTEXALPHA, 1);
 	rw::SetRenderState(SRCBLEND, BLENDONE);
-	rw::SetRenderState(DESTBLEND, BLENDONE);
 
 	flushCache();
 	glDrawElements(header->primType, inst->numIndex,
 	               GL_UNSIGNED_SHORT, (void*)(uintptr)inst->offset);
 
 	rw::SetRenderState(SRCBLEND, BLENDSRCALPHA);
-	rw::SetRenderState(DESTBLEND, BLENDINVSRCALPHA);
 }
 
 void
 matfxRenderCB(Atomic *atomic, InstanceDataHeader *header)
 {
 	setWorldMatrix(atomic->getFrame()->getLTM());
-	lightingCB(!!(atomic->geometry->flags & Geometry::NORMALS));
+	lightingCB(atomic);
 
 	glBindBuffer(GL_ARRAY_BUFFER, header->vbo);
 	glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, header->ibo);
 	setAttribPointers(header->attribDesc, header->numAttribs);
 
+	lastEnvFrame = nil;
+
 	InstanceData *inst = header->inst;
 	int32 n = header->numMeshes;
 
-//	rw::SetRenderState(ALPHATESTFUNC, 1);
-//	rw::SetRenderState(ALPHATESTREF, 50);
-
-	int32 fx;
 	while(n--){
-		fx = MatFX::getEffects(inst->material);
-		switch(fx){
+		MatFX *matfx = MatFX::get(inst->material);
+
+		if(matfx == nil)
+			matfxDefaultRender(header, inst);
+		else switch(matfx->type){
 		case MatFX::ENVMAP:
-			matfxEnvRender(header, inst);
+			matfxEnvRender(header, inst, &matfx->fx[0].env);
 			break;
 		default:
 			matfxDefaultRender(header, inst);
+			break;
 		}
 		inst++;
 	}
diff --git a/src/gl/gl3raster.cpp b/src/gl/gl3raster.cpp
index 7bb833f..6b9064d 100644
--- a/src/gl/gl3raster.cpp
+++ b/src/gl/gl3raster.cpp
@@ -44,7 +44,7 @@ rasterCreateTexture(Raster *raster)
 		// TODO: check if this is correct
 		natras->internalFormat = GL_RGBA;
 		natras->format = GL_RGBA;
-		natras->type = GL_UNSIGNED_SHORT_5_5_5_1;
+		natras->type = GL_UNSIGNED_SHORT_1_5_5_5_REV;
 		natras->hasAlpha = 1;
 		break;
 	default:
@@ -96,7 +96,7 @@ rasterCreateCameraTexture(Raster *raster)
 		// TODO: check if this is correct
 		natras->internalFormat = GL_RGBA;
 		natras->format = GL_RGBA;
-		natras->type = GL_UNSIGNED_SHORT_5_5_5_1;
+		natras->type = GL_UNSIGNED_SHORT_1_5_5_5_REV;
 		natras->hasAlpha = 1;
 		break;
 	}
@@ -242,12 +242,23 @@ imageFindRasterFormat(Image *img, int32 type,
 	return 1;
 }
 
+static uint8*
+flipImage(Image *image)
+{
+	int i;
+	uint8 *newPx = (uint8*)rwMalloc(image->stride*image->height, 0);
+	for(i = 0; i < image->height; i++)
+		memcpy(&newPx[i*image->stride], &image->pixels[(image->height-1-i)*image->stride], image->stride);
+	return newPx;
+}
+
 bool32
 rasterFromImage(Raster *raster, Image *image)
 {
 	if((raster->type&0xF) != Raster::TEXTURE)
 		return 0;
 
+#ifdef RW_OPENGL
 	// Unpalettize image if necessary but don't change original
 	Image *truecolimg = nil;
 	if(image->depth <= 8){
@@ -259,18 +270,28 @@ rasterFromImage(Raster *raster, Image *image)
 		image = truecolimg;
 	}
 
+	// NB: important to set the format of the input data here!
 	Gl3Raster *natras = PLUGINOFFSET(Gl3Raster, raster, nativeRasterOffset);
 	switch(image->depth){
 	case 32:
 		if(raster->format != Raster::C8888 &&
 		   raster->format != Raster::C888)
 			goto err;
+		natras->format = GL_RGBA;
+		natras->type = GL_UNSIGNED_BYTE;
+		natras->hasAlpha = 1;
 		break;
 	case 24:
 		if(raster->format != Raster::C888) goto err;
+		natras->format = GL_RGB;
+		natras->type = GL_UNSIGNED_BYTE;
+		natras->hasAlpha = 0;
 		break;
 	case 16:
 		if(raster->format != Raster::C1555) goto err;
+		natras->format = GL_RGBA;
+		natras->type = GL_UNSIGNED_SHORT_1_5_5_5_REV;
+		natras->hasAlpha = 1;
 		break;
 
 	case 8:
@@ -283,12 +304,16 @@ rasterFromImage(Raster *raster, Image *image)
 
 	natras->hasAlpha = image->hasAlpha();
 
-#ifdef RW_OPENGL
+
+	uint8 *flipped = flipImage(image);
+
 	glBindTexture(GL_TEXTURE_2D, natras->texid);
 	glTexImage2D(GL_TEXTURE_2D, 0, natras->internalFormat,
 	             raster->width, raster->height,
-	             0, natras->format, natras->type, image->pixels);
+	             0, natras->format, natras->type, flipped);
 	glBindTexture(GL_TEXTURE_2D, 0);
+
+	rwFree(flipped);
 #endif
 	return 1;
 }
diff --git a/src/gl/gl3render.cpp b/src/gl/gl3render.cpp
index e5b2f1a..4d6039e 100644
--- a/src/gl/gl3render.cpp
+++ b/src/gl/gl3render.cpp
@@ -40,31 +40,29 @@ disableAttribPointers(AttribDesc *attribDescs, int32 numAttribs)
 		glDisableVertexAttribArray(a->index);
 }
 
-void
-lightingCB(bool32 normals)
+int32
+lightingCB(Atomic *atomic)
 {
-	World *world;
-	RGBAf ambLight = { 0.0, 0.0, 0.0, 1.0 };
-	int n = 0;
+	WorldLights lightData;
+	Light *directionals[8];
+	Light *locals[8];
+	lightData.directionals = directionals;
+	lightData.numDirectionals = 8;
+	lightData.locals = locals;
+	lightData.numLocals = 8;
 
-	world = (World*)engine->currentWorld;
-	// only unpositioned lights right now
-	FORLIST(lnk, world->globalLights){
-		Light *l = Light::fromWorld(lnk);
-		if(normals &&
-		   l->getType() == Light::DIRECTIONAL &&
-		   l->getFlags() & Light::LIGHTATOMICS){
-			if(n >= MAX_LIGHTS)
-				continue;
-			setLight(n++, l);
-		}else if(l->getType() == Light::AMBIENT){
-			ambLight.red   += l->color.red;
-			ambLight.green += l->color.green;
-			ambLight.blue  += l->color.blue;
+	if(atomic->geometry->flags & rw::Geometry::LIGHT){
+		((World*)engine->currentWorld)->enumerateLights(atomic, &lightData);
+		if((atomic->geometry->flags & rw::Geometry::NORMALS) == 0){
+			// Get rid of lights that need normals when we don't have any
+			lightData.numDirectionals = 0;
+			lightData.numLocals = 0;
 		}
+		return setLights(&lightData);
+	}else{
+		memset(&lightData, 0, sizeof(lightData));
+		return setLights(&lightData);
 	}
-	setNumLights(n);
-	setAmbientLight(&ambLight);
 }
 
 #define U(i) currentShader->uniformLocations[i]
@@ -77,7 +75,7 @@ defaultRenderCB(Atomic *atomic, InstanceDataHeader *header)
 	GLfloat surfProps[4];
 
 	setWorldMatrix(atomic->getFrame()->getLTM());
-	lightingCB(!!(atomic->geometry->flags & Geometry::NORMALS));
+	lightingCB(atomic);
 
 	glBindBuffer(GL_ARRAY_BUFFER, header->vbo);
 	glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, header->ibo);
@@ -86,7 +84,7 @@ defaultRenderCB(Atomic *atomic, InstanceDataHeader *header)
 	InstanceData *inst = header->inst;
 	int32 n = header->numMeshes;
 
-	simpleShader->use();
+	defaultShader->use();
 
 	while(n--){
 		m = inst->material;
@@ -98,7 +96,7 @@ defaultRenderCB(Atomic *atomic, InstanceDataHeader *header)
 		surfProps[1] = m->surfaceProps.specular;
 		surfProps[2] = m->surfaceProps.diffuse;
 		surfProps[3] = 0.0f;
-		glUniform4fv(U(u_surfaceProps), 1, surfProps);
+		glUniform4fv(U(u_surfProps), 1, surfProps);
 
 		setTexture(0, m->texture);
 
diff --git a/src/gl/gl3shader.cpp b/src/gl/gl3shader.cpp
index 19dfc34..a3bd08f 100644
--- a/src/gl/gl3shader.cpp
+++ b/src/gl/gl3shader.cpp
@@ -16,6 +16,8 @@
 namespace rw {
 namespace gl3 {
 
+#include "shaders/header_vs.inc"
+
 UniformRegistry uniformRegistry;
 
 int32
@@ -89,14 +91,17 @@ loadfile(const char *path)
 }
 
 static int
-compileshader(GLenum type, const char *src, GLuint *shader)
+compileshader(GLenum type, const char **src, GLuint *shader)
 {
+	GLint n;
 	GLint shdr, success;
 	GLint len;
 	char *log;
 
+	for(n = 0; src[n]; n++);
+
 	shdr = glCreateShader(type);
-	glShaderSource(shdr, 1, &src, nil);
+	glShaderSource(shdr, n, src, nil);
 	glCompileShader(shdr);
 	glGetShaderiv(shdr, GL_COMPILE_STATUS, &success);
 	if(!success){
@@ -139,6 +144,58 @@ linkprogram(GLint vs, GLint fs, GLuint *program)
 	return 0;
 }
 
+Shader*
+Shader::create(const char **vsrc, const char **fsrc)
+{
+	GLuint vs, fs, program;
+	int i;
+	int fail;
+
+	fail = compileshader(GL_VERTEX_SHADER, vsrc, &vs);
+	if(fail)
+		return nil;
+
+	fail = compileshader(GL_FRAGMENT_SHADER, fsrc, &fs);
+	if(fail)
+		return nil;
+
+	fail = linkprogram(vs, fs, &program);
+	if(fail)
+		return nil;
+	glDeleteProgram(vs);
+	glDeleteProgram(fs);
+
+	Shader *sh = rwNewT(Shader, 1, MEMDUR_EVENT | ID_DRIVER);	 // or global?
+
+	// set uniform block binding
+	for(i = 0; i < uniformRegistry.numBlocks; i++){
+		int idx = glGetUniformBlockIndex(program,
+		                                 uniformRegistry.blockNames[i]);
+		if(idx >= 0)
+			glUniformBlockBinding(program, idx, i);
+	}
+
+	// query uniform locations
+	sh->program = program;
+	sh->uniformLocations = rwNewT(GLint, uniformRegistry.numUniforms, MEMDUR_EVENT | ID_DRIVER);
+	for(i = 0; i < uniformRegistry.numUniforms; i++)
+		sh->uniformLocations[i] = glGetUniformLocation(program,
+			uniformRegistry.uniformNames[i]);
+
+	// set samplers
+	glUseProgram(program);
+	char name[64];
+	GLint loc;
+	for(i = 0; i < 4; i++){
+		sprintf(name, "tex%d", i);
+		loc = glGetUniformLocation(program, name);
+		glUniform1i(loc, i);
+	}
+
+	return sh;
+}
+
+#if 0
 Shader*
 Shader::fromStrings(const char *vsrc, const char *fsrc)
 {
@@ -191,6 +248,7 @@ Shader::fromFiles(const char *vspath, const char *fspath)
 	rwFree(fsrc);
 	return s;
 }
+#endif
 
 void
 Shader::use(void)
diff --git a/src/gl/gl3skin.cpp b/src/gl/gl3skin.cpp
index 30b0aae..2169781 100644
--- a/src/gl/gl3skin.cpp
+++ b/src/gl/gl3skin.cpp
@@ -35,10 +35,15 @@ static void*
 skinOpen(void *o, int32, int32)
 {
 	u_boneMatrices = registerUniform("u_boneMatrices");
+	skinGlobals.pipelines[PLATFORM_GL3] = makeSkinPipeline();
+
 #include "shaders/simple_fs_gl3.inc"
 #include "shaders/skin_gl3.inc"
-	skinGlobals.pipelines[PLATFORM_GL3] = makeSkinPipeline();
-	skinShader = Shader::fromStrings(skin_vert_src, simple_frag_src);
+	const char *vs[] = { header_vert_src, skin_vert_src, nil };
+	const char *fs[] = { simple_frag_src, nil };
+	skinShader = Shader::create(vs, fs);
+	assert(skinShader);
+
 	return o;
 }
 
@@ -216,23 +221,34 @@ skinUninstanceCB(Geometry *geo, InstanceDataHeader *header)
 static float skinMatrices[64*16];
 
 void
-updateSkinMatrices(Atomic *a)
+uploadSkinMatrices(Atomic *a)
 {
+	int i;
 	Skin *skin = Skin::get(a->geometry);
 	HAnimHierarchy *hier = Skin::getHierarchy(a);
 	Matrix *invMats = (Matrix*)skin->inverseMatrices;
+	Matrix tmp;
 
-	float *m;
-	m = (float*)skinMatrices;
-	for(int i = 0; i < hier->numNodes; i++){
-		invMats[i].flags = 0;
-		Matrix::mult((Matrix*)m, &invMats[i], &hier->matrices[i]);
-		m[3] = 0.0f;
-		m[7] = 0.0f;
-		m[11] = 0.0f;
-		m[15] = 1.0f;
-		m += 16;
+	Matrix *m = (Matrix*)skinMatrices;
+
+	if(hier->flags & HAnimHierarchy::LOCALSPACEMATRICES){
+		for(i = 0; i < hier->numNodes; i++){
+			invMats[i].flags = 0;
+			Matrix::mult(m, &invMats[i], &hier->matrices[i]);
+			m++;
+		}
+	}else{
+		Matrix invAtmMat;
+		Matrix::invert(&invAtmMat, a->getFrame()->getLTM());
+		for(i = 0; i < hier->numNodes; i++){
+			invMats[i].flags = 0;
+			Matrix::mult(&tmp, &hier->matrices[i], &invAtmMat);
+			Matrix::mult(m, &invMats[i], &tmp);
+			m++;
+		}
 	}
+	glUniformMatrix4fv(U(u_boneMatrices), 64, GL_FALSE,
+	                   (GLfloat*)skinMatrices);
 }
 
 void
@@ -243,7 +259,7 @@ skinRenderCB(Atomic *atomic, InstanceDataHeader *header)
 	GLfloat surfProps[4];
 
 	setWorldMatrix(atomic->getFrame()->getLTM());
-	lightingCB(!!(atomic->geometry->flags & Geometry::NORMALS));
+	lightingCB(atomic);
 
 	glBindBuffer(GL_ARRAY_BUFFER, header->vbo);
 	glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, header->ibo);
@@ -252,14 +268,9 @@ skinRenderCB(Atomic *atomic, InstanceDataHeader *header)
 	InstanceData *inst = header->inst;
 	int32 n = header->numMeshes;
 
-//	rw::setRenderState(ALPHATESTFUNC, 1);
-//	rw::setRenderState(ALPHATESTREF, 50);
-
 	skinShader->use();
 
-	updateSkinMatrices(atomic);
-	glUniformMatrix4fv(U(u_boneMatrices), 64, GL_FALSE,
-	                   (GLfloat*)skinMatrices);
+	uploadSkinMatrices(atomic);
 
 	while(n--){
 		m = inst->material;
@@ -271,7 +282,7 @@ skinRenderCB(Atomic *atomic, InstanceDataHeader *header)
 		surfProps[1] = m->surfaceProps.specular;
 		surfProps[2] = m->surfaceProps.diffuse;
 		surfProps[3] = 0.0f;
-		glUniform4fv(U(u_surfaceProps), 1, surfProps);
+		glUniform4fv(U(u_surfProps), 1, surfProps);
 
 		setTexture(0, m->texture);
 
diff --git a/src/gl/rwgl3.h b/src/gl/rwgl3.h
index 09772d7..91e462c 100644
--- a/src/gl/rwgl3.h
+++ b/src/gl/rwgl3.h
@@ -57,7 +57,7 @@ enum AttribIndices
 
 // default uniform indices
 extern int32 u_matColor;
-extern int32 u_surfaceProps;
+extern int32 u_surfProps;
 
 struct InstanceData
 {
@@ -92,7 +92,7 @@ struct InstanceDataHeader : rw::InstanceDataHeader
 
 struct Shader;
 
-extern Shader *simpleShader;
+extern Shader *defaultShader;
 
 struct Im3DVertex
 {
@@ -146,15 +146,27 @@ void disableAttribPointers(AttribDesc *attribDescs, int32 numAttribs);
 
 // Render state
 
+// Vertex shader bits
+enum
+{
+	// These should be low so they could be used as indices
+	VSLIGHT_DIRECT	= 1,
+	VSLIGHT_POINT	= 2,
+	VSLIGHT_SPOT	= 4,
+	VSLIGHT_MASK	= 7,	// all the above
+	// less critical
+	VSLIGHT_AMBIENT = 8,
+};
+
+extern const char *header_vert_src;
+
 // per Scene
 void setProjectionMatrix(float32*);
 void setViewMatrix(float32*);
 
 // per Object
 void setWorldMatrix(Matrix*);
-void setAmbientLight(RGBAf*);
-void setNumLights(int32 n);
-void setLight(int32 n, Light*);
+int32 setLights(WorldLights *lightData);
 
 // per Mesh
 void setTexture(int32 n, Texture *tex);
@@ -176,7 +188,7 @@ public:
 void defaultInstanceCB(Geometry *geo, InstanceDataHeader *header);
 void defaultUninstanceCB(Geometry *geo, InstanceDataHeader *header);
 void defaultRenderCB(Atomic *atomic, InstanceDataHeader *header);
-void lightingCB(bool32 normals);
+int32 lightingCB(Atomic *atomic);
 
 ObjPipeline *makeDefaultPipeline(void);
 
diff --git a/src/gl/rwgl3shader.h b/src/gl/rwgl3shader.h
index c209d58..72f736f 100644
--- a/src/gl/rwgl3shader.h
+++ b/src/gl/rwgl3shader.h
@@ -31,8 +31,9 @@ struct Shader
 	// same number of elements as UniformRegistry::numUniforms
 	GLint *uniformLocations;
 
-	static Shader *fromFiles(const char *vs, const char *fs);
-	static Shader *fromStrings(const char *vsrc, const char *fsrc);
+	static Shader *create(const char **vsrc, const char **fsrc);
+//	static Shader *fromFiles(const char *vs, const char *fs);
+//	static Shader *fromStrings(const char *vsrc, const char *fsrc);
 	void use(void);
 	void destroy(void);
 };
diff --git a/src/gl/shaders/Makefile b/src/gl/shaders/Makefile
index ecb2e5b..cec56d3 100644
--- a/src/gl/shaders/Makefile
+++ b/src/gl/shaders/Makefile
@@ -1,4 +1,9 @@
-all: im2d_gl3.inc im3d_gl3.inc simple_vs_gl3.inc simple_fs_gl3.inc matfx_gl3.inc skin_gl3.inc
+all: header_vs.inc im2d_gl3.inc im3d_gl3.inc default_vs_gl3.inc simple_fs_gl3.inc matfx_gl3.inc skin_gl3.inc
+
+header_vs.inc: header.vert
+	(echo 'const char *header_vert_src =';\
+	 sed 's/..*/"&\\n"/' header.vert;\
+	 echo ';') >header_vs.inc
 
 im2d_gl3.inc: im2d.vert
 	(echo 'const char *im2d_vert_src =';\
@@ -10,10 +15,10 @@ im3d_gl3.inc: im3d.vert
 	 sed 's/..*/"&\\n"/' im3d.vert;\
 	 echo ';') >im3d_gl3.inc
 
-simple_vs_gl3.inc: simple.vert
-	(echo 'const char *simple_vert_src =';\
-	 sed 's/..*/"&\\n"/' simple.vert;\
-	 echo ';') >simple_vs_gl3.inc
+default_vs_gl3.inc: default.vert
+	(echo 'const char *default_vert_src =';\
+	 sed 's/..*/"&\\n"/' default.vert;\
+	 echo ';') >default_vs_gl3.inc
 
 simple_fs_gl3.inc: simple.frag
 	(echo 'const char *simple_frag_src =';\
diff --git a/src/gl/shaders/default.vert b/src/gl/shaders/default.vert
new file mode 100644
index 0000000..cebf153
--- /dev/null
+++ b/src/gl/shaders/default.vert
@@ -0,0 +1,48 @@
+layout(location = 0) in vec3 in_pos;
+layout(location = 1) in vec3 in_normal;
+layout(location = 2) in vec4 in_color;
+layout(location = 3) in vec2 in_tex0;
+
+out vec4 v_color;
+out vec2 v_tex0;
+out float v_fog;
+
+void
+main(void)
+{
+	vec4 V = u_world * vec4(in_pos, 1.0);
+	gl_Position = u_proj * u_view * V;
+	vec3 N = mat3(u_world) * in_normal;
+
+	v_tex0 = in_tex0;
+
+	v_color = in_color;
+	v_color.rgb += u_ambLight.rgb*surfAmbient;
+
+#ifdef DIRECTIONALS
+	for(int i = 0; i < MAX_LIGHTS; i++){
+		if(u_directLights[i].enabled == 0.0)
+			break;
+		v_color.rgb += DoDirLight(u_directLights[i], N)*surfDiffuse;
+	}
+#endif
+#ifdef POINTLIGHTS
+	for(int i = 0; i < MAX_LIGHTS; i++){
+		if(u_pointLights[i].enabled == 0.0)
+			break;
+		v_color.rgb += DoPointLight(u_pointLights[i], V.xyz, N)*surfDiffuse;
+	}
+#endif
+#ifdef SPOTLIGHTS
+	for(int i = 0; i < MAX_LIGHTS; i++){
+		if(u_spotLights[i].enabled == 0.0)
+			break;
+		v_color.rgb += DoSpotLight(u_spotLights[i], V.xyz, N)*surfDiffuse;
+	}
+#endif
+
+	v_color = clamp(v_color, 0.0f, 1.0);
+	v_color *= u_matColor;
+
+	v_fog = DoFog(gl_Position.w);
+}
diff --git a/src/gl/shaders/default_vs_gl3.inc b/src/gl/shaders/default_vs_gl3.inc
new file mode 100644
index 0000000..cc3965a
--- /dev/null
+++ b/src/gl/shaders/default_vs_gl3.inc
@@ -0,0 +1,50 @@
+const char *default_vert_src =
+"layout(location = 0) in vec3 in_pos;\n"
+"layout(location = 1) in vec3 in_normal;\n"
+"layout(location = 2) in vec4 in_color;\n"
+"layout(location = 3) in vec2 in_tex0;\n"
+
+"out vec4 v_color;\n"
+"out vec2 v_tex0;\n"
+"out float v_fog;\n"
+
+"void\n"
+"main(void)\n"
+"{\n"
+"	vec4 V = u_world * vec4(in_pos, 1.0);\n"
+"	gl_Position = u_proj * u_view * V;\n"
+"	vec3 N = mat3(u_world) * in_normal;\n"
+
+"	v_tex0 = in_tex0;\n"
+
+"	v_color = in_color;\n"
+"	v_color.rgb += u_ambLight.rgb*surfAmbient;\n"
+
+"#ifdef DIRECTIONALS\n"
+"	for(int i = 0; i < MAX_LIGHTS; i++){\n"
+"		if(u_directLights[i].enabled == 0.0)\n"
+"			break;\n"
+"		v_color.rgb += DoDirLight(u_directLights[i], N)*surfDiffuse;\n"
+"	}\n"
+"#endif\n"
+"#ifdef POINTLIGHTS\n"
+"	for(int i = 0; i < MAX_LIGHTS; i++){\n"
+"		if(u_pointLights[i].enabled == 0.0)\n"
+"			break;\n"
+"		v_color.rgb += DoPointLight(u_pointLights[i], V.xyz, N)*surfDiffuse;\n"
+"	}\n"
+"#endif\n"
+"#ifdef SPOTLIGHTS\n"
+"	for(int i = 0; i < MAX_LIGHTS; i++){\n"
+"		if(u_spotLights[i].enabled == 0.0)\n"
+"			break;\n"
+"		v_color.rgb += DoSpotLight(u_spotLights[i], V.xyz, N)*surfDiffuse;\n"
+"	}\n"
+"#endif\n"
+
+"	v_color = clamp(v_color, 0.0f, 1.0);\n"
+"	v_color *= u_matColor;\n"
+
+"	v_fog = DoFog(gl_Position.w);\n"
+"}\n"
+;
diff --git a/src/gl/shaders/header.vert b/src/gl/shaders/header.vert
new file mode 100644
index 0000000..8048156
--- /dev/null
+++ b/src/gl/shaders/header.vert
@@ -0,0 +1,88 @@
+#version 330
+
+layout(std140) uniform State
+{
+	int   u_alphaTest;
+	float u_alphaRef;
+
+	float u_fogStart;
+	float u_fogEnd;
+	float u_fogRange;
+	float u_fogDisable;
+	vec4  u_fogColor;
+};
+
+layout(std140) uniform Scene
+{
+	mat4 u_proj;
+	mat4 u_view;
+};
+
+#define MAX_LIGHTS 8
+struct Light {
+	float enabled;
+	float radius;
+	float minusCosAngle;
+	float hardSpot;
+	vec4  position;
+	vec4  direction;
+	vec4  color;
+};
+
+layout(std140) uniform Object
+{
+	mat4  u_world;
+	vec4  u_ambLight;
+	Light u_directLights[MAX_LIGHTS];
+	Light u_pointLights[MAX_LIGHTS];
+	Light u_spotLights[MAX_LIGHTS];
+};
+
+uniform vec4 u_matColor;
+uniform vec4 u_surfProps;	// amb, spec, diff, extra
+
+#define surfAmbient (u_surfProps.x)
+#define surfSpecular (u_surfProps.y)
+#define surfDiffuse (u_surfProps.z)
+
+vec3 DoDirLight(Light L, vec3 N)
+{
+	float l = max(0.0, dot(N, -L.direction.xyz));
+	return l*L.color.rgb;
+}
+
+vec3 DoPointLight(Light L, vec3 V, vec3 N)
+{
+	// As on PS2
+	vec3 dir = V - L.position.xyz;
+	float dist = length(dir);
+	float atten = max(0.0, (1.0 - dist/L.radius));
+	float l = max(0.0, dot(N, -normalize(dir)));
+	return l*L.color.rgb*atten;
+}
+
+vec3 DoSpotLight(Light L, vec3 V, vec3 N)
+{
+	// As on PS2
+	vec3 dir = V - L.position.xyz;
+	float dist = length(dir);
+	float atten = max(0.0, (1.0 - dist/L.radius));
+	dir /= dist;
+	float l = max(0.0, dot(N, -dir));
+	float pcos = dot(dir, L.direction.xyz);	// cos to point
+	float ccos = -L.minusCosAngle;
+	float falloff = (pcos-ccos)/(1.0-ccos);
+	if(falloff < 0)	// outside of cone
+		l = 0;
+	l *= max(falloff, L.hardSpot);
+	return l*L.color.xyz*atten;
+}
+
+float DoFog(float w)
+{
+	return clamp((w - u_fogEnd)*u_fogRange, u_fogDisable, 1.0);
+}
+
+#define DIRECTIONALS
+//#define POINTLIGHTS
+//#define SPOTLIGHTS
\ No newline at end of file
diff --git a/src/gl/shaders/header_vs.inc b/src/gl/shaders/header_vs.inc
new file mode 100644
index 0000000..2f884dc
--- /dev/null
+++ b/src/gl/shaders/header_vs.inc
@@ -0,0 +1,89 @@
+const char *header_vert_src =
+"#version 330\n"
+
+"layout(std140) uniform State\n"
+"{\n"
+"	int   u_alphaTest;\n"
+"	float u_alphaRef;\n"
+
+"	float u_fogStart;\n"
+"	float u_fogEnd;\n"
+"	float u_fogRange;\n"
+"	float u_fogDisable;\n"
+"	vec4  u_fogColor;\n"
+"};\n"
+
+"layout(std140) uniform Scene\n"
+"{\n"
+"	mat4 u_proj;\n"
+"	mat4 u_view;\n"
+"};\n"
+
+"#define MAX_LIGHTS 8\n"
+"struct Light {\n"
+"	float enabled;\n"
+"	float radius;\n"
+"	float minusCosAngle;\n"
+"	float hardSpot;\n"
+"	vec4  position;\n"
+"	vec4  direction;\n"
+"	vec4  color;\n"
+"};\n"
+
+"layout(std140) uniform Object\n"
+"{\n"
+"	mat4  u_world;\n"
+"	vec4  u_ambLight;\n"
+"	Light u_directLights[MAX_LIGHTS];\n"
+"	Light u_pointLights[MAX_LIGHTS];\n"
+"	Light u_spotLights[MAX_LIGHTS];\n"
+"};\n"
+
+"uniform vec4 u_matColor;\n"
+"uniform vec4 u_surfProps;	// amb, spec, diff, extra\n"
+
+"#define surfAmbient (u_surfProps.x)\n"
+"#define surfSpecular (u_surfProps.y)\n"
+"#define surfDiffuse (u_surfProps.z)\n"
+
+"vec3 DoDirLight(Light L, vec3 N)\n"
+"{\n"
+"	float l = max(0.0, dot(N, -L.direction.xyz));\n"
+"	return l*L.color.rgb;\n"
+"}\n"
+
+"vec3 DoPointLight(Light L, vec3 V, vec3 N)\n"
+"{\n"
+"	// As on PS2\n"
+"	vec3 dir = V - L.position.xyz;\n"
+"	float dist = length(dir);\n"
+"	float atten = max(0.0, (1.0 - dist/L.radius));\n"
+"	float l = max(0.0, dot(N, -normalize(dir)));\n"
+"	return l*L.color.rgb*atten;\n"
+"}\n"
+
+"vec3 DoSpotLight(Light L, vec3 V, vec3 N)\n"
+"{\n"
+"	// As on PS2\n"
+"	vec3 dir = V - L.position.xyz;\n"
+"	float dist = length(dir);\n"
+"	float atten = max(0.0, (1.0 - dist/L.radius));\n"
+"	dir /= dist;\n"
+"	float l = max(0.0, dot(N, -dir));\n"
+"	float pcos = dot(dir, L.direction.xyz);	// cos to point\n"
+"	float ccos = -L.minusCosAngle;\n"
+"	float falloff = (pcos-ccos)/(1.0-ccos);\n"
+"	if(falloff < 0)	// outside of cone\n"
+"		l = 0;\n"
+"	l *= max(falloff, L.hardSpot);\n"
+"	return l*L.color.xyz*atten;\n"
+"}\n"
+
+"float DoFog(float w)\n"
+"{\n"
+"	return clamp((w - u_fogEnd)*u_fogRange, u_fogDisable, 1.0);\n"
+"}\n"
+
+"#define DIRECTIONALS\n"
+"//#define POINTLIGHTS\n"
+"//#define SPOTLIGHTS\n";
diff --git a/src/gl/shaders/im2d.vert b/src/gl/shaders/im2d.vert
index 46c5c20..cff0870 100644
--- a/src/gl/shaders/im2d.vert
+++ b/src/gl/shaders/im2d.vert
@@ -1,16 +1,3 @@
-#version 330
-
-layout(std140) uniform State
-{
-	int   u_alphaTest;
-	float u_alphaRef;
-
-	int   u_fogEnable;
-	float u_fogStart;
-	float u_fogEnd;
-	vec4  u_fogColor;
-};
-
 uniform vec4 u_xform;
 
 layout(location = 0) in vec4 in_pos;
@@ -26,8 +13,8 @@ main(void)
 {
 	gl_Position = in_pos;
 	gl_Position.xy = gl_Position.xy * u_xform.xy + u_xform.zw;
+	v_fog = DoFog(gl_Position.z);
 	gl_Position.xyz *= gl_Position.w;
 	v_color = in_color;
 	v_tex0 = in_tex0;
-	v_fog = clamp((gl_Position.z - u_fogEnd)/(u_fogStart - u_fogEnd), 0.0, 1.0);
 }
diff --git a/src/gl/shaders/im2d_gl3.inc b/src/gl/shaders/im2d_gl3.inc
index 11f2e80..3212919 100644
--- a/src/gl/shaders/im2d_gl3.inc
+++ b/src/gl/shaders/im2d_gl3.inc
@@ -1,17 +1,4 @@
 const char *im2d_vert_src =
-"#version 330\n"
-
-"layout(std140) uniform State\n"
-"{\n"
-"	int   u_alphaTest;\n"
-"	float u_alphaRef;\n"
-
-"	int   u_fogEnable;\n"
-"	float u_fogStart;\n"
-"	float u_fogEnd;\n"
-"	vec4  u_fogColor;\n"
-"};\n"
-
 "uniform vec4 u_xform;\n"
 
 "layout(location = 0) in vec4 in_pos;\n"
@@ -27,9 +14,9 @@ const char *im2d_vert_src =
 "{\n"
 "	gl_Position = in_pos;\n"
 "	gl_Position.xy = gl_Position.xy * u_xform.xy + u_xform.zw;\n"
+"	v_fog = DoFog(gl_Position.z);\n"
 "	gl_Position.xyz *= gl_Position.w;\n"
 "	v_color = in_color;\n"
 "	v_tex0 = in_tex0;\n"
-"	v_fog = clamp((gl_Position.z - u_fogEnd)/(u_fogStart - u_fogEnd), 0.0, 1.0);\n"
 "}\n"
 ;
diff --git a/src/gl/shaders/im3d.vert b/src/gl/shaders/im3d.vert
index 55091d6..0276370 100644
--- a/src/gl/shaders/im3d.vert
+++ b/src/gl/shaders/im3d.vert
@@ -1,27 +1,3 @@
-#version 330
-
-layout(std140) uniform State
-{
-	int   u_alphaTest;
-	float u_alphaRef;
-
-	int   u_fogEnable;
-	float u_fogStart;
-	float u_fogEnd;
-	vec4  u_fogColor;
-};
-
-layout(std140) uniform Scene
-{
-	mat4 u_proj;
-	mat4 u_view;
-};
-
-layout(std140) uniform Object
-{
-	mat4  u_world;
-};
-
 layout(location = 0) in vec3 in_pos;
 layout(location = 2) in vec4 in_color;
 layout(location = 3) in vec2 in_tex0;
@@ -38,5 +14,5 @@ main(void)
 	gl_Position = u_proj * cV;
 	v_color = in_color;
 	v_tex0 = in_tex0;
-	v_fog = clamp((cV.z - u_fogEnd)/(u_fogStart - u_fogEnd), 0.0, 1.0);
+	v_fog = DoFog(gl_Position.w);
 }
diff --git a/src/gl/shaders/im3d_gl3.inc b/src/gl/shaders/im3d_gl3.inc
index 2c878c0..7459606 100644
--- a/src/gl/shaders/im3d_gl3.inc
+++ b/src/gl/shaders/im3d_gl3.inc
@@ -1,28 +1,4 @@
 const char *im3d_vert_src =
-"#version 330\n"
-
-"layout(std140) uniform State\n"
-"{\n"
-"	int   u_alphaTest;\n"
-"	float u_alphaRef;\n"
-
-"	int   u_fogEnable;\n"
-"	float u_fogStart;\n"
-"	float u_fogEnd;\n"
-"	vec4  u_fogColor;\n"
-"};\n"
-
-"layout(std140) uniform Scene\n"
-"{\n"
-"	mat4 u_proj;\n"
-"	mat4 u_view;\n"
-"};\n"
-
-"layout(std140) uniform Object\n"
-"{\n"
-"	mat4  u_world;\n"
-"};\n"
-
 "layout(location = 0) in vec3 in_pos;\n"
 "layout(location = 2) in vec4 in_color;\n"
 "layout(location = 3) in vec2 in_tex0;\n"
@@ -35,10 +11,10 @@ const char *im3d_vert_src =
 "main(void)\n"
 "{\n"
 "	vec4 V = u_world * vec4(in_pos, 1.0);\n"
-"	vec4 cV = u_view * V;   \n"
+"	vec4 cV = u_view * V;\n"
 "	gl_Position = u_proj * cV;\n"
 "	v_color = in_color;\n"
 "	v_tex0 = in_tex0;\n"
-"	v_fog = clamp((cV.z - u_fogEnd)/(u_fogStart - u_fogEnd), 0.0, 1.0);\n"
+"	v_fog = DoFog(gl_Position.w);\n"
 "}\n"
 ;
diff --git a/src/gl/shaders/matfx_env.frag b/src/gl/shaders/matfx_env.frag
index ed90362..87ec846 100644
--- a/src/gl/shaders/matfx_env.frag
+++ b/src/gl/shaders/matfx_env.frag
@@ -4,20 +4,38 @@ layout(std140) uniform State
 {
 	int   u_alphaTest;
 	float u_alphaRef;
-	// fog etc.
+
+	float u_fogStart;
+	float u_fogEnd;
+	float u_fogRange;
+	float u_fogDisable;
+	vec4  u_fogColor;
 };
 
-uniform sampler2D tex;
+uniform sampler2D tex0;
+uniform sampler2D tex1;
+
+uniform float u_coefficient;
 
 in vec4 v_color;
 in vec2 v_tex0;
+in vec2 v_tex1;
+in float v_fog;
 
 out vec4 color;
 
 void
 main(void)
 {
-	color = v_color*texture(tex, vec2(v_tex0.x, v_tex0.y));
+	vec4 pass1 = v_color;
+	vec4 envColor = pass1;	// TODO: colorClamp
+
+	pass1 *= texture(tex0, vec2(v_tex0.x, 1.0-v_tex0.y));
+	vec4 pass2 = envColor*u_coefficient*texture(tex1, vec2(v_tex1.x, 1.0-v_tex1.y));
+
+	color.rgb = pass1.rgb*pass1.a + pass2.rgb;
+	color.a = pass1.a;
+
 	switch(u_alphaTest){
 	default:
 	case 0: break;
@@ -31,4 +49,3 @@ main(void)
 		break;
 	}
 }
-
diff --git a/src/gl/shaders/matfx_env.vert b/src/gl/shaders/matfx_env.vert
index d63ada2..40cf146 100644
--- a/src/gl/shaders/matfx_env.vert
+++ b/src/gl/shaders/matfx_env.vert
@@ -1,33 +1,4 @@
-#version 330
-
-layout(std140) uniform Scene
-{
-	mat4 u_proj;
-	mat4 u_view;
-};
-
-#define MAX_LIGHTS 8
-struct Light {
-	vec4  position;
-	vec4  direction;
-	vec4  color;
-	float radius;
-	float minusCosAngle;
-};
-
-layout(std140) uniform Object
-{
-	mat4  u_world;
-	vec4  u_ambLight;
-	int   u_numLights;
-	Light u_lights[MAX_LIGHTS];
-};
-
-uniform vec4 u_matColor;
-uniform vec4 u_surfaceProps;	// amb, spec, diff, extra
-
 uniform mat4 u_texMatrix;
-uniform float u_coefficient;
 
 layout(location = 0) in vec3 in_pos;
 layout(location = 1) in vec3 in_normal;
@@ -36,6 +7,8 @@ layout(location = 3) in vec2 in_tex0;
 
 out vec4 v_color;
 out vec2 v_tex0;
+out vec2 v_tex1;
+out float v_fog;
 
 void
 main(void)
@@ -44,15 +17,36 @@ main(void)
 	gl_Position = u_proj * u_view * V;
 	vec3 N = mat3(u_world) * in_normal;
 
+	v_tex0 = in_tex0;
+	v_tex1 = (u_texMatrix * vec4(N, 1.0)).xy;
+
 	v_color = in_color;
-	for(int i = 0; i < u_numLights; i++){
-		float L = max(0.0, dot(N, -normalize(u_lights[i].direction.xyz)));
-		v_color.rgb += u_lights[i].color.rgb*L*u_surfaceProps.z;
+	v_color.rgb += u_ambLight.rgb*surfAmbient;
+
+#ifdef DIRECTIONALS
+	for(int i = 0; i < MAX_LIGHTS; i++){
+		if(u_directLights[i].enabled == 0.0)
+			break;
+		v_color.rgb += DoDirLight(u_directLights[i], N)*surfDiffuse;
 	}
-	v_color.rgb += u_ambLight.rgb*u_surfaceProps.x;
+#endif
+#ifdef POINTLIGHTS
+	for(int i = 0; i < MAX_LIGHTS; i++){
+		if(u_pointLights[i].enabled == 0.0)
+			break;
+		v_color.rgb += DoPointLight(u_pointLights[i], V.xyz, N)*surfDiffuse;
+	}
+#endif
+#ifdef SPOTLIGHTS
+	for(int i = 0; i < MAX_LIGHTS; i++){
+		if(u_spotLights[i].enabled == 0.0)
+			break;
+		v_color.rgb += DoSpotLight(u_spotLights[i], V.xyz, N)*surfDiffuse;
+	}
+#endif
+
+	v_color = clamp(v_color, 0.0f, 1.0);
 	v_color *= u_matColor;
 
-	v_color *= u_coefficient;
-	vec3 cN = mat3(u_view) * in_normal;
-	v_tex0 = (u_texMatrix * vec4(cN, 1.0)).xy;
+	v_fog = DoFog(gl_Position.w);
 }
diff --git a/src/gl/shaders/matfx_gl3.inc b/src/gl/shaders/matfx_gl3.inc
index 8b65b64..ac351fa 100644
--- a/src/gl/shaders/matfx_gl3.inc
+++ b/src/gl/shaders/matfx_gl3.inc
@@ -1,34 +1,5 @@
 const char *matfx_env_vert_src =
-"#version 330\n"
-
-"layout(std140) uniform Scene\n"
-"{\n"
-"	mat4 u_proj;\n"
-"	mat4 u_view;\n"
-"};\n"
-
-"#define MAX_LIGHTS 8\n"
-"struct Light {\n"
-"	vec4  position;\n"
-"	vec4  direction;\n"
-"	vec4  color;\n"
-"	float radius;\n"
-"	float minusCosAngle;\n"
-"};\n"
-
-"layout(std140) uniform Object\n"
-"{\n"
-"	mat4  u_world;\n"
-"	vec4  u_ambLight;\n"
-"	int   u_numLights;\n"
-"	Light u_lights[MAX_LIGHTS];\n"
-"};\n"
-
-"uniform vec4 u_matColor;\n"
-"uniform vec4 u_surfaceProps;	// amb, spec, diff, extra\n"
-
 "uniform mat4 u_texMatrix;\n"
-"uniform float u_coefficient;\n"
 
 "layout(location = 0) in vec3 in_pos;\n"
 "layout(location = 1) in vec3 in_normal;\n"
@@ -37,6 +8,8 @@ const char *matfx_env_vert_src =
 
 "out vec4 v_color;\n"
 "out vec2 v_tex0;\n"
+"out vec2 v_tex1;\n"
+"out float v_fog;\n"
 
 "void\n"
 "main(void)\n"
@@ -45,17 +18,38 @@ const char *matfx_env_vert_src =
 "	gl_Position = u_proj * u_view * V;\n"
 "	vec3 N = mat3(u_world) * in_normal;\n"
 
+"	v_tex0 = in_tex0;\n"
+"	v_tex1 = (u_texMatrix * vec4(N, 1.0)).xy;\n"
+
 "	v_color = in_color;\n"
-"	for(int i = 0; i < u_numLights; i++){\n"
-"		float L = max(0.0, dot(N, -normalize(u_lights[i].direction.xyz)));\n"
-"		v_color.rgb += u_lights[i].color.rgb*L*u_surfaceProps.z;\n"
+"	v_color.rgb += u_ambLight.rgb*surfAmbient;\n"
+
+"#ifdef DIRECTIONALS\n"
+"	for(int i = 0; i < MAX_LIGHTS; i++){\n"
+"		if(u_directLights[i].enabled == 0.0)\n"
+"			break;\n"
+"		v_color.rgb += DoDirLight(u_directLights[i], N)*surfDiffuse;\n"
 "	}\n"
-"	v_color.rgb += u_ambLight.rgb*u_surfaceProps.x;\n"
+"#endif\n"
+"#ifdef POINTLIGHTS\n"
+"	for(int i = 0; i < MAX_LIGHTS; i++){\n"
+"		if(u_pointLights[i].enabled == 0.0)\n"
+"			break;\n"
+"		v_color.rgb += DoPointLight(u_pointLights[i], V.xyz, N)*surfDiffuse;\n"
+"	}\n"
+"#endif\n"
+"#ifdef SPOTLIGHTS\n"
+"	for(int i = 0; i < MAX_LIGHTS; i++){\n"
+"		if(u_spotLights[i].enabled == 0.0)\n"
+"			break;\n"
+"		v_color.rgb += DoSpotLight(u_spotLights[i], V.xyz, N)*surfDiffuse;\n"
+"	}\n"
+"#endif\n"
+
+"	v_color = clamp(v_color, 0.0f, 1.0);\n"
 "	v_color *= u_matColor;\n"
 
-"	v_color *= u_coefficient;\n"
-"	vec3 cN = mat3(u_view) * in_normal;\n"
-"	v_tex0 = (u_texMatrix * vec4(cN, 1.0)).xy;\n"
+"	v_fog = DoFog(gl_Position.w);\n"
 "}\n"
 ;
 const char *matfx_env_frag_src =
@@ -65,20 +59,38 @@ const char *matfx_env_frag_src =
 "{\n"
 "	int   u_alphaTest;\n"
 "	float u_alphaRef;\n"
-"	// fog etc.\n"
+
+"	float u_fogStart;\n"
+"	float u_fogEnd;\n"
+"	float u_fogRange;\n"
+"	float u_fogDisable;\n"
+"	vec4  u_fogColor;\n"
 "};\n"
 
-"uniform sampler2D tex;\n"
+"uniform sampler2D tex0;\n"
+"uniform sampler2D tex1;\n"
+
+"uniform float u_coefficient;\n"
 
 "in vec4 v_color;\n"
 "in vec2 v_tex0;\n"
+"in vec2 v_tex1;\n"
+"in float v_fog;\n"
 
 "out vec4 color;\n"
 
 "void\n"
 "main(void)\n"
 "{\n"
-"	color = v_color*texture(tex, vec2(v_tex0.x, v_tex0.y));\n"
+"	vec4 pass1 = v_color;\n"
+"	vec4 envColor = pass1;	// TODO: colorClamp\n"
+
+"	pass1 *= texture(tex0, vec2(v_tex0.x, 1.0-v_tex0.y));\n"
+"	vec4 pass2 = envColor*u_coefficient*texture(tex1, vec2(v_tex1.x, 1.0-v_tex1.y));\n"
+
+"	color.rgb = pass1.rgb*pass1.a + pass2.rgb;\n"
+"	color.a = pass1.a;\n"
+
 "	switch(u_alphaTest){\n"
 "	default:\n"
 "	case 0: break;\n"
@@ -92,5 +104,4 @@ const char *matfx_env_frag_src =
 "		break;\n"
 "	}\n"
 "}\n"
-
 ;
diff --git a/src/gl/shaders/simple.frag b/src/gl/shaders/simple.frag
index 770d62b..9be99ad 100644
--- a/src/gl/shaders/simple.frag
+++ b/src/gl/shaders/simple.frag
@@ -5,13 +5,14 @@ layout(std140) uniform State
 	int   u_alphaTest;
 	float u_alphaRef;
 
-	int   u_fogEnable;
 	float u_fogStart;
 	float u_fogEnd;
+	float u_fogRange;
+	float u_fogDisable;
 	vec4  u_fogColor;
 };
 
-uniform sampler2D tex;
+uniform sampler2D tex0;
 
 in vec4 v_color;
 in vec2 v_tex0;
@@ -22,9 +23,8 @@ out vec4 color;
 void
 main(void)
 {
-	color = v_color*texture(tex, vec2(v_tex0.x, v_tex0.y));
-	if(u_fogEnable != 0)
-		color.rgb = mix(u_fogColor.rgb, color.rgb, v_fog);
+	color = v_color*texture(tex0, vec2(v_tex0.x, 1.0-v_tex0.y));
+	color.rgb = mix(u_fogColor.rgb, color.rgb, v_fog);
 	switch(u_alphaTest){
 	default:
 	case 0: break;
diff --git a/src/gl/shaders/simple.vert b/src/gl/shaders/simple.vert
deleted file mode 100644
index 2cd7f1a..0000000
--- a/src/gl/shaders/simple.vert
+++ /dev/null
@@ -1,68 +0,0 @@
-#version 330
-
-layout(std140) uniform State
-{
-	int   u_alphaTest;
-	float u_alphaRef;
-
-	int   u_fogEnable;
-	float u_fogStart;
-	float u_fogEnd;
-	vec4  u_fogColor;
-};
-
-layout(std140) uniform Scene
-{
-	mat4 u_proj;
-	mat4 u_view;
-};
-
-#define MAX_LIGHTS 8
-struct Light {
-	vec4  position;
-	vec4  direction;
-	vec4  color;
-	float radius;
-	float minusCosAngle;
-};
-
-layout(std140) uniform Object
-{
-	mat4  u_world;
-	vec4  u_ambLight;
-	int   u_numLights;
-	Light u_lights[MAX_LIGHTS];
-};
-
-uniform vec4 u_matColor;
-uniform vec4 u_surfaceProps;	// amb, spec, diff, extra
-
-layout(location = 0) in vec3 in_pos;
-layout(location = 1) in vec3 in_normal;
-layout(location = 2) in vec4 in_color;
-layout(location = 3) in vec2 in_tex0;
-
-out vec4 v_color;
-out vec2 v_tex0;
-out float v_fog;
-
-void
-main(void)
-{
-	vec4 V = u_world * vec4(in_pos, 1.0);
-	vec4 cV = u_view * V;
-	gl_Position = u_proj * cV;
-	vec3 N = mat3(u_world) * in_normal;
-
-	v_color = in_color;
-	for(int i = 0; i < u_numLights; i++){
-		float L = max(0.0, dot(N, -normalize(u_lights[i].direction.xyz)));
-		v_color.rgb += u_lights[i].color.rgb*L*u_surfaceProps.z;
-	}
-	v_color.rgb += u_ambLight.rgb*u_surfaceProps.x;
-	v_color *= u_matColor;
-
-	v_tex0 = in_tex0;
-
-	v_fog = clamp((cV.z - u_fogEnd)/(u_fogStart - u_fogEnd), 0.0, 1.0);
-}
diff --git a/src/gl/shaders/simple_fs_gl3.inc b/src/gl/shaders/simple_fs_gl3.inc
index 6cf1d7f..74b4ec0 100644
--- a/src/gl/shaders/simple_fs_gl3.inc
+++ b/src/gl/shaders/simple_fs_gl3.inc
@@ -6,13 +6,14 @@ const char *simple_frag_src =
 "	int   u_alphaTest;\n"
 "	float u_alphaRef;\n"
 
-"	int   u_fogEnable;\n"
 "	float u_fogStart;\n"
 "	float u_fogEnd;\n"
+"	float u_fogRange;\n"
+"	float u_fogDisable;\n"
 "	vec4  u_fogColor;\n"
 "};\n"
 
-"uniform sampler2D tex;\n"
+"uniform sampler2D tex0;\n"
 
 "in vec4 v_color;\n"
 "in vec2 v_tex0;\n"
@@ -23,9 +24,8 @@ const char *simple_frag_src =
 "void\n"
 "main(void)\n"
 "{\n"
-"	color = v_color*texture(tex, vec2(v_tex0.x, v_tex0.y));\n"
-"	if(u_fogEnable != 0)\n"
-"		color.rgb = mix(u_fogColor.rgb, color.rgb, v_fog);\n"
+"	color = v_color*texture(tex0, vec2(v_tex0.x, 1.0-v_tex0.y));\n"
+"	color.rgb = mix(u_fogColor.rgb, color.rgb, v_fog);\n"
 "	switch(u_alphaTest){\n"
 "	default:\n"
 "	case 0: break;\n"
diff --git a/src/gl/shaders/simple_vs_gl3.inc b/src/gl/shaders/simple_vs_gl3.inc
deleted file mode 100644
index 5332b28..0000000
--- a/src/gl/shaders/simple_vs_gl3.inc
+++ /dev/null
@@ -1,72 +0,0 @@
-const char *simple_vert_src =
-"#version 330\n"
-
-"layout(std140) uniform State\n"
-"{\n"
-"	int   u_alphaTest;\n"
-"	float u_alphaRef;\n"
-
-"	int   u_fogEnable;\n"
-"	float u_fogStart;\n"
-"	float u_fogEnd;\n"
-"	vec4  u_fogColor;\n"
-"};\n"
-
-"layout(std140) uniform Scene\n"
-"{\n"
-"	mat4 u_proj;\n"
-"	mat4 u_view;\n"
-"};\n"
-
-"#define MAX_LIGHTS 8\n"
-"struct Light {\n"
-"	vec4  position;\n"
-"	vec4  direction;\n"
-"	vec4  color;\n"
-"	float radius;\n"
-"	float minusCosAngle;\n"
-"};\n"
-
-"layout(std140) uniform Object\n"
-"{\n"
-"	mat4  u_world;\n"
-"	vec4  u_ambLight;\n"
-"	int   u_numLights;\n"
-"	Light u_lights[MAX_LIGHTS];\n"
-"};\n"
-
-"uniform vec4 u_matColor;\n"
-"uniform vec4 u_surfaceProps;	// amb, spec, diff, extra\n"
-
-"layout(location = 0) in vec3 in_pos;\n"
-"layout(location = 1) in vec3 in_normal;\n"
-"layout(location = 2) in vec4 in_color;\n"
-"layout(location = 3) in vec2 in_tex0;\n"
-
-"out vec4 v_color;\n"
-"out vec2 v_tex0;\n"
-"out float v_fog;\n"
-
-"void\n"
-"main(void)\n"
-"{\n"
-"	vec3 lightdir = vec3(1.0, 1.0, -1.0);\n"
-
-"	vec4 V = u_world * vec4(in_pos, 1.0);\n"
-"	vec4 cV = u_view * V;   \n"
-"	gl_Position = u_proj * cV;\n"
-"	vec3 N = mat3(u_world) * in_normal;\n"
-
-"	v_color = in_color;\n"
-"	for(int i = 0; i < u_numLights; i++){\n"
-"		float L = max(0.0, dot(N, -normalize(u_lights[i].direction.xyz)));\n"
-"		v_color.rgb += u_lights[i].color.rgb*L*u_surfaceProps.z;\n"
-"	}\n"
-"	v_color.rgb += u_ambLight.rgb*u_surfaceProps.x;\n"
-"	v_color *= u_matColor;\n"
-
-"	v_tex0 = in_tex0;\n"
-
-"        v_fog = clamp((cV.z - u_fogEnd)/(u_fogStart - u_fogEnd), 0.0, 1.0);\n"
-"}\n"
-;
diff --git a/src/gl/shaders/skin.vert b/src/gl/shaders/skin.vert
index 8861561..1a90898 100644
--- a/src/gl/shaders/skin.vert
+++ b/src/gl/shaders/skin.vert
@@ -1,41 +1,4 @@
-#version 330
-
-layout(std140) uniform State
-{
-	int   u_alphaTest;
-	float u_alphaRef;
-
-	int   u_fogEnable;
-	float u_fogStart;
-	float u_fogEnd;
-	vec4  u_fogColor;
-};
-
-layout(std140) uniform Scene
-{
-	mat4 u_proj;
-	mat4 u_view;
-};
-
-#define MAX_LIGHTS 8
-struct Light {
-	vec4  position;
-	vec4  direction;
-	vec4  color;
-	float radius;
-	float minusCosAngle;
-};
-
-layout(std140) uniform Object
-{
-	mat4  u_world;
-	vec4  u_ambLight;
-	int   u_numLights;
-	Light u_lights[MAX_LIGHTS];
-};
-
-uniform vec4 u_matColor;
-uniform vec4 u_surfaceProps;	// amb, spec, diff, extra
+uniform mat4 u_boneMatrices[64];
 
 layout(location = 0) in vec3 in_pos;
 layout(location = 1) in vec3 in_normal;
@@ -44,8 +7,6 @@ layout(location = 3) in vec2 in_tex0;
 layout(location = 11) in vec4 in_weights;
 layout(location = 12) in vec4 in_indices;
 
-uniform mat4 u_boneMatrices[64];
-
 out vec4 v_color;
 out vec2 v_tex0;
 out float v_fog;
@@ -53,26 +14,44 @@ out float v_fog;
 void
 main(void)
 {
-	mat4 m = u_boneMatrices[int(in_indices[0])] * in_weights[0];
-	m +=     u_boneMatrices[int(in_indices[1])] * in_weights[1];
-	m +=     u_boneMatrices[int(in_indices[2])] * in_weights[2];
-	m +=     u_boneMatrices[int(in_indices[3])] * in_weights[3];
-	mat4 world = u_world * m;
+	vec3 SkinVertex = vec3(0.0, 0.0, 0.0);
+	vec3 SkinNormal = vec3(0.0, 0.0, 0.0);
+	for(int i = 0; i < 4; i++){
+		SkinVertex += (u_boneMatrices[int(in_indices[i])] * vec4(in_pos, 1.0)).xyz * in_weights[i];
+		SkinNormal += (mat3(u_boneMatrices[int(in_indices[i])]) * in_normal) * in_weights[i];
+	}
 
-	vec4 V = world * vec4(in_pos, 1.0);
-	vec4 cV = u_view * V;
-	gl_Position = u_proj * cV;
-	vec3 N = mat3(world) * in_normal;
+	vec4 V = u_world * vec4(SkinVertex, 1.0);
+	gl_Position = u_proj * u_view * V;
+	vec3 N = mat3(u_world) * SkinNormal;
 
 	v_color = in_color;
-	for(int i = 0; i < u_numLights; i++){
-		float L = max(0.0, dot(N, -normalize(u_lights[i].direction.xyz)));
-		v_color.rgb += u_lights[i].color.rgb*L*u_surfaceProps.z;
+	v_color.rgb += u_ambLight.rgb*surfAmbient;
+
+#ifdef DIRECTIONALS
+	for(int i = 0; i < MAX_LIGHTS; i++){
+		if(u_directLights[i].enabled == 0.0)
+			break;
+		v_color.rgb += DoDirLight(u_directLights[i], N)*surfDiffuse;
 	}
-	v_color.rgb += u_ambLight.rgb*u_surfaceProps.x;
+#endif
+#ifdef POINTLIGHTS
+	for(int i = 0; i < MAX_LIGHTS; i++){
+		if(u_pointLights[i].enabled == 0.0)
+			break;
+		v_color.rgb += DoPointLight(u_pointLights[i], V.xyz, N)*surfDiffuse;
+	}
+#endif
+#ifdef SPOTLIGHTS
+	for(int i = 0; i < MAX_LIGHTS; i++){
+		if(u_spotLights[i].enabled == 0.0)
+			break;
+		v_color.rgb += DoSpotLight(u_spotLights[i], V.xyz, N)*surfDiffuse;
+	}
+#endif
 	v_color *= u_matColor;
 
 	v_tex0 = in_tex0;
 
-	v_fog = clamp((cV.z - u_fogEnd)/(u_fogStart - u_fogEnd), 0.0, 1.0);
+	v_fog = DoFog(gl_Position.z);
 }
diff --git a/src/gl/shaders/skin_gl3.inc b/src/gl/shaders/skin_gl3.inc
index 4730087..0880a54 100644
--- a/src/gl/shaders/skin_gl3.inc
+++ b/src/gl/shaders/skin_gl3.inc
@@ -1,42 +1,5 @@
 const char *skin_vert_src =
-"#version 330\n"
-
-"layout(std140) uniform State\n"
-"{\n"
-"	int   u_alphaTest;\n"
-"	float u_alphaRef;\n"
-
-"	int   u_fogEnable;\n"
-"	float u_fogStart;\n"
-"	float u_fogEnd;\n"
-"	vec4  u_fogColor;\n"
-"};\n"
-
-"layout(std140) uniform Scene\n"
-"{\n"
-"	mat4 u_proj;\n"
-"	mat4 u_view;\n"
-"};\n"
-
-"#define MAX_LIGHTS 8\n"
-"struct Light {\n"
-"	vec4  position;\n"
-"	vec4  direction;\n"
-"	vec4  color;\n"
-"	float radius;\n"
-"	float minusCosAngle;\n"
-"};\n"
-
-"layout(std140) uniform Object\n"
-"{\n"
-"	mat4  u_world;\n"
-"	vec4  u_ambLight;\n"
-"	int   u_numLights;\n"
-"	Light u_lights[MAX_LIGHTS];\n"
-"};\n"
-
-"uniform vec4 u_matColor;\n"
-"uniform vec4 u_surfaceProps;	// amb, spec, diff, extra\n"
+"uniform mat4 u_boneMatrices[64];\n"
 
 "layout(location = 0) in vec3 in_pos;\n"
 "layout(location = 1) in vec3 in_normal;\n"
@@ -45,8 +8,6 @@ const char *skin_vert_src =
 "layout(location = 11) in vec4 in_weights;\n"
 "layout(location = 12) in vec4 in_indices;\n"
 
-"uniform mat4 u_boneMatrices[64];\n"
-
 "out vec4 v_color;\n"
 "out vec2 v_tex0;\n"
 "out float v_fog;\n"
@@ -54,27 +15,45 @@ const char *skin_vert_src =
 "void\n"
 "main(void)\n"
 "{\n"
-"	mat4 m = u_boneMatrices[int(in_indices[0])] * in_weights[0];\n"
-"	m +=     u_boneMatrices[int(in_indices[1])] * in_weights[1];\n"
-"	m +=     u_boneMatrices[int(in_indices[2])] * in_weights[2];\n"
-"	m +=     u_boneMatrices[int(in_indices[3])] * in_weights[3];\n"
-"	mat4 world = u_world * m;\n"
+"	vec3 SkinVertex = vec3(0.0, 0.0, 0.0);\n"
+"	vec3 SkinNormal = vec3(0.0, 0.0, 0.0);\n"
+"	for(int i = 0; i < 4; i++){\n"
+"		SkinVertex += (u_boneMatrices[int(in_indices[i])] * vec4(in_pos, 1.0)).xyz * in_weights[i];\n"
+"		SkinNormal += (mat3(u_boneMatrices[int(in_indices[i])]) * in_normal) * in_weights[i];\n"
+"	}\n"
 
-"	vec4 V = world * vec4(in_pos, 1.0);\n"
-"	vec4 cV = u_view * V;   \n"
-"	gl_Position = u_proj * cV;\n"
-"	vec3 N = mat3(world) * in_normal;\n"
+"	vec4 V = u_world * vec4(SkinVertex, 1.0);\n"
+"	gl_Position = u_proj * u_view * V;\n"
+"	vec3 N = mat3(u_world) * SkinNormal;\n"
 
 "	v_color = in_color;\n"
-"	for(int i = 0; i < u_numLights; i++){\n"
-"		float L = max(0.0, dot(N, -normalize(u_lights[i].direction.xyz)));\n"
-"		v_color.rgb += u_lights[i].color.rgb*L*u_surfaceProps.z;\n"
+"	v_color.rgb += u_ambLight.rgb*surfAmbient;\n"
+
+"#ifdef DIRECTIONALS\n"
+"	for(int i = 0; i < MAX_LIGHTS; i++){\n"
+"		if(u_directLights[i].enabled == 0.0)\n"
+"			break;\n"
+"		v_color.rgb += DoDirLight(u_directLights[i], N)*surfDiffuse;\n"
 "	}\n"
-"	v_color.rgb += u_ambLight.rgb*u_surfaceProps.x;\n"
+"#endif\n"
+"#ifdef POINTLIGHTS\n"
+"	for(int i = 0; i < MAX_LIGHTS; i++){\n"
+"		if(u_pointLights[i].enabled == 0.0)\n"
+"			break;\n"
+"		v_color.rgb += DoPointLight(u_pointLights[i], V.xyz, N)*surfDiffuse;\n"
+"	}\n"
+"#endif\n"
+"#ifdef SPOTLIGHTS\n"
+"	for(int i = 0; i < MAX_LIGHTS; i++){\n"
+"		if(u_spotLights[i].enabled == 0.0)\n"
+"			break;\n"
+"		v_color.rgb += DoSpotLight(u_spotLights[i], V.xyz, N)*surfDiffuse;\n"
+"	}\n"
+"#endif\n"
 "	v_color *= u_matColor;\n"
 
 "	v_tex0 = in_tex0;\n"
 
-"        v_fog = clamp((cV.z - u_fogEnd)/(u_fogStart - u_fogEnd), 0.0, 1.0);\n"
+"	v_fog = DoFog(gl_Position.z);\n"
 "}\n"
 ;
diff --git a/src/image.cpp b/src/image.cpp
index 6a350a4..0e86730 100644
--- a/src/image.cpp
+++ b/src/image.cpp
@@ -21,7 +21,13 @@ namespace rw {
 
 int32 Image::numAllocated;
 
-// TODO: full 16 bit support
+// Image formats are as follows:
+//  32 bit has 4 bytes: 8888 RGBA
+//  24 bit has 3 bytes: 888 RGB
+//  16 bit has 2 bytes: 1555 ARGB stored in platform native order (TODO?)
+// palettes always have 4 bytes: r, g, b, a
+//   8 bit has 1 byte: x
+//   4 bit has 1 byte per two pixels: 0xLR, where L and R are the left and right pixel resp.
 
 Image*
 Image::create(int32 width, int32 height, int32 depth)