diff --git a/examples/Demo/CMainMenu.cpp b/examples/Demo/CMainMenu.cpp
index ab98d19b..dcfa29c6 100644
--- a/examples/Demo/CMainMenu.cpp
+++ b/examples/Demo/CMainMenu.cpp
@@ -70,7 +70,7 @@ bool CMainMenu::run()
 	gui::IGUIListBox* box = guienv->addListBox(core::rect<int>(10,10,220,120), optTab, 1);
 	box->addItem(L"OpenGL 1.5");
 	box->addItem(L"Direct3D 9.0c");
-	box->addItem(L"Burning's Video 0.53");
+	box->addItem(L"Burning's Video 0.54");
 	box->addItem(L"Irrlicht Software Renderer 1.0");
 	switch (driverType )
 	{
diff --git a/source/Irrlicht/CSoftwareDriver2.cpp b/source/Irrlicht/CSoftwareDriver2.cpp
index b04e3030..4972f9d5 100644
--- a/source/Irrlicht/CSoftwareDriver2.cpp
+++ b/source/Irrlicht/CSoftwareDriver2.cpp
@@ -108,7 +108,6 @@ bool mat44_transposed_inverse(irr::core::CMatrix4<T>& out, const irr::core::CMat
 }
 #endif
 
-#if 0
 // difference to CMatrix4<T>::getInverse . higher precision in determinant. return identity on failure
 template <class T>
 bool mat44_inverse(irr::core::CMatrix4<T>& out, const irr::core::CMatrix4<T>& M)
@@ -153,7 +152,7 @@ bool mat44_inverse(irr::core::CMatrix4<T>& out, const irr::core::CMatrix4<T>& M)
 
 	return true;
 }
-#endif
+
 
 // void CMatrix4<T>::transformVec4(T *out, const T * in) const
 template <class T>
@@ -167,6 +166,15 @@ inline void transformVec4Vec4(const irr::core::CMatrix4<T>& m, T* burning_restri
 	out[3] = in[0] * M[3] + in[1] * M[7] + in[2] * M[11] + in[3] * M[15];
 }
 
+template <class T>
+inline void transformVec3Vec3(const irr::core::CMatrix4<T>& m, T* burning_restrict out, const T* burning_restrict in)
+{
+	const T* burning_restrict M = m.pointer();
+	out[0] = in[0] * M[0] + in[1] * M[4] + in[2] * M[8] + M[12];
+	out[1] = in[0] * M[1] + in[1] * M[5] + in[2] * M[9] + M[13];
+	out[2] = in[0] * M[2] + in[1] * M[6] + in[2] * M[10] + M[14];
+}
+
 #if 0
 // void CMatrix4<T>::transformVect(T *out, const core::vector3df &in) const
 template <class T>
@@ -256,6 +264,11 @@ static inline float dot(const irr::video::sVec4& a, const irr::video::sVec4& b)
 	return a.x * b.x + a.y * b.y + a.z * b.z + a.w * b.w;
 }
 
+static inline float inversesqrt(const float x )
+{
+	return x!= 0.f ? 1.f / sqrtf(x) : 0.f;
+}
+
 
 static inline irr::video::sVec4 operator-(const irr::video::sVec4& a)
 {
@@ -619,7 +632,7 @@ CBurningVideoDriver::CBurningVideoDriver(const irr::SIrrlichtCreationParameters&
 	BurningShader[ETR_TEXTURE_GOURAUD_ALPHA] = createTRTextureGouraudAlpha(this);
 	BurningShader[ETR_TEXTURE_GOURAUD_ALPHA_NOZ] = createTRTextureGouraudAlphaNoZ(this);
 
-	BurningShader[ETR_NORMAL_MAP_SOLID] = createTRNormalMap(this);
+	//BurningShader[ETR_NORMAL_MAP_SOLID] = createTRNormalMap(this, EMT_NORMAL_MAP_SOLID);
 	BurningShader[ETR_STENCIL_SHADOW] = createTRStencilShadow(this);
 	BurningShader[ETR_TEXTURE_BLEND] = createTRTextureBlend(this);
 
@@ -651,12 +664,28 @@ CBurningVideoDriver::CBurningVideoDriver(const irr::SIrrlichtCreationParameters&
 	addMaterialRenderer(tmr); // EMT_TRANSPARENT_ALPHA_CHANNEL_REF,
 	addMaterialRenderer(tmr); // EMT_TRANSPARENT_VERTEX_ALPHA,
 	addMaterialRenderer(tmr); // EMT_TRANSPARENT_REFLECTION_2_LAYER,
+
+#if 0
 	addMaterialRenderer(smr); // EMT_NORMAL_MAP_SOLID,
 	addMaterialRenderer(tmr); // EMT_NORMAL_MAP_TRANSPARENT_ADD_COLOR,
 	addMaterialRenderer(tmr); // EMT_NORMAL_MAP_TRANSPARENT_VERTEX_ALPHA,
 	addMaterialRenderer(smr); // EMT_PARALLAX_MAP_SOLID,
 	addMaterialRenderer(tmr); // EMT_PARALLAX_MAP_TRANSPARENT_ADD_COLOR,
 	addMaterialRenderer(tmr); // EMT_PARALLAX_MAP_TRANSPARENT_VERTEX_ALPHA,
+#else
+	// add normal map renderers
+	s32 tmp = 0;
+	video::IMaterialRenderer* renderer = 0;
+	renderer = createTRNormalMap(this, tmp, EMT_NORMAL_MAP_SOLID); renderer->drop();
+	renderer = createTRNormalMap(this, tmp, EMT_NORMAL_MAP_TRANSPARENT_ADD_COLOR); renderer->drop();
+	renderer = createTRNormalMap(this, tmp, EMT_NORMAL_MAP_TRANSPARENT_VERTEX_ALPHA); renderer->drop();
+
+	renderer = createTRParallaxMap(this, tmp, EMT_PARALLAX_MAP_SOLID); renderer->drop();
+	renderer = createTRParallaxMap(this, tmp, EMT_PARALLAX_MAP_TRANSPARENT_ADD_COLOR); renderer->drop();
+	renderer = createTRParallaxMap(this, tmp, EMT_PARALLAX_MAP_TRANSPARENT_VERTEX_ALPHA); renderer->drop();
+
+#endif
+
 	addMaterialRenderer(tmr); // EMT_ONETEXTURE_BLEND
 
 	smr->drop();
@@ -820,6 +849,9 @@ void CBurningVideoDriver::transform_calc(E_TRANSFORMATION_STATE_BURNING_VIDEO st
 	case ETS_NORMAL:
 		ok = flag[ETS_MODEL_VIEW] & ETF_VALID;
 		break;
+	case ETS_MODEL_INVERSE:
+		ok = flag[ETS_WORLD] & ETF_VALID;
+		break;
 	default:
 		break;
 	}
@@ -827,7 +859,7 @@ void CBurningVideoDriver::transform_calc(E_TRANSFORMATION_STATE_BURNING_VIDEO st
 	if (!ok)
 	{
 		char buf[256];
-		sprintf(buf, "transform_calc not valid for %d\n", state);
+		sprintf(buf, "transform_calc not valid for %d", state);
 		os::Printer::log(buf, ELL_WARNING);
 	}
 
@@ -862,6 +894,16 @@ void CBurningVideoDriver::transform_calc(E_TRANSFORMATION_STATE_BURNING_VIDEO st
 	case ETS_NORMAL:
 		mat33_transposed_inverse(matrix[state], matrix[ETS_MODEL_VIEW]);
 		break;
+	case ETS_MODEL_INVERSE:
+		if (flag[ETS_WORLD] & ETF_IDENTITY)
+		{
+			matrix[state] = matrix[ETS_WORLD];
+		}
+		else
+		{
+			mat44_inverse(matrix[state], matrix[ETS_WORLD]);
+		}
+		break;
 
 	default:
 		break;
@@ -917,6 +959,7 @@ void CBurningVideoDriver::setTransform(E_TRANSFORMATION_STATE state, const core:
 			flag[ETS_MODEL_VIEW_PROJ] &= ~ETF_VALID;
 			flag[ETS_MODEL_VIEW] &= ~ETF_VALID;
 			flag[ETS_NORMAL] &= ~ETF_VALID;
+			flag[ETS_MODEL_INVERSE] &= ~ETF_VALID;
 			break;
 		case ETS_TEXTURE_0:
 		case ETS_TEXTURE_1:
@@ -1739,7 +1782,6 @@ f32 MipmapLevel(const sVec2& uv, const sVec2& textureSize)
 }
 #endif
 
-//#define MAT_TEXTURE(tex) ( (video::CSoftwareTexture2*) Material.org.getTexture ( (u32)tex ) )
 #define MAT_TEXTURE(tex) ( (video::CSoftwareTexture2*) Material.org.TextureLayer[tex].Texture )
 
 //! clamp(value,0,1)
@@ -1904,46 +1946,59 @@ void CBurningVideoDriver::VertexCache_map_source_format()
 
 
 	SVSize* vSize = VertexShader.vSize;
-	//vSize[E4VT_STANDARD].Format = VERTEX4D_FORMAT_TEXTURE_1 | VERTEX4D_FORMAT_COLOR_1 | VERTEX4D_FORMAT_LIGHT_1 | VERTEX4D_FORMAT_SPECULAR;
 	vSize[E4VT_STANDARD].Format = VERTEX4D_FORMAT_TEXTURE_1 | VERTEX4D_FORMAT_COLOR_2_FOG;
 	vSize[E4VT_STANDARD].Pitch = sizeof(S3DVertex);
 	vSize[E4VT_STANDARD].TexSize = 1;
 	vSize[E4VT_STANDARD].TexCooSize = 1;
+	vSize[E4VT_STANDARD].ColSize = 2;
 
 	vSize[E4VT_2TCOORDS].Format = VERTEX4D_FORMAT_TEXTURE_2 | VERTEX4D_FORMAT_COLOR_2_FOG;
 	vSize[E4VT_2TCOORDS].Pitch = sizeof(S3DVertex2TCoords);
 	vSize[E4VT_2TCOORDS].TexSize = 2;
 	vSize[E4VT_2TCOORDS].TexCooSize = 2;
+	vSize[E4VT_2TCOORDS].ColSize = 2;
 
-	//vSize[E4VT_TANGENTS].Format = VERTEX4D_FORMAT_TEXTURE_2 | VERTEX4D_FORMAT_COLOR_1 | VERTEX4D_FORMAT_LIGHT_1 | VERTEX4D_FORMAT_BUMP_DOT3;
-	vSize[E4VT_TANGENTS].Format = VERTEX4D_FORMAT_TEXTURE_2 | VERTEX4D_FORMAT_COLOR_2_FOG | VERTEX4D_FORMAT_LIGHT_1 | VERTEX4D_FORMAT_BUMP_DOT3;
+	// 	EMT_NORMAL_MAP_SOLID,EMT_NORMAL_MAP_TRANSPARENT_ADD_COLOR,EMT_NORMAL_MAP_TRANSPARENT_VERTEX_ALPHA
+	vSize[E4VT_TANGENTS].Format = VERTEX4D_FORMAT_TEXTURE_2 | VERTEX4D_FORMAT_COLOR_4 | VERTEX4D_FORMAT_BUMP_DOT3;
 	vSize[E4VT_TANGENTS].Pitch = sizeof(S3DVertexTangents);
 	vSize[E4VT_TANGENTS].TexSize = 2;
 	vSize[E4VT_TANGENTS].TexCooSize = 2;
+	vSize[E4VT_TANGENTS].ColSize = 4;
 
-	// reflection map
+	// 	EMT_PARALLAX_MAP_SOLID,EMT_PARALLAX_MAP_TRANSPARENT_ADD_COLOR,EMT_PARALLAX_MAP_TRANSPARENT_VERTEX_ALPHA
+	vSize[E4VT_TANGENTS_PARALLAX].Format = VERTEX4D_FORMAT_TEXTURE_2 | VERTEX4D_FORMAT_COLOR_4 | VERTEX4D_FORMAT_PARALLAX;
+	vSize[E4VT_TANGENTS_PARALLAX].Pitch = sizeof(S3DVertexTangents);
+	vSize[E4VT_TANGENTS_PARALLAX].TexSize = 2;
+	vSize[E4VT_TANGENTS_PARALLAX].TexCooSize = 2;
+	vSize[E4VT_TANGENTS_PARALLAX].ColSize = 4;
+
+	// EMT_TRANSPARENT_REFLECTION_2_LAYER map
 	vSize[E4VT_REFLECTION_MAP].Format = VERTEX4D_FORMAT_TEXTURE_2 | VERTEX4D_FORMAT_COLOR_2_FOG;
 	vSize[E4VT_REFLECTION_MAP].Pitch = sizeof(S3DVertex);
 	vSize[E4VT_REFLECTION_MAP].TexSize = 2;
 	vSize[E4VT_REFLECTION_MAP].TexCooSize = 1; //TexCoo2 generated
+	vSize[E4VT_REFLECTION_MAP].ColSize = 2;
 
 	// shadow
 	vSize[E4VT_SHADOW].Format = 0;
 	vSize[E4VT_SHADOW].Pitch = sizeof(f32) * 3; // core::vector3df*
 	vSize[E4VT_SHADOW].TexSize = 0;
 	vSize[E4VT_SHADOW].TexCooSize = 0;
+	vSize[E4VT_SHADOW].ColSize = 0;
 
 	// color shading only (no texture)
-	vSize[E4VT_NO_TEXTURE].Format = VERTEX4D_FORMAT_COLOR_1 | VERTEX4D_FORMAT_LIGHT_1 | VERTEX4D_FORMAT_SPECULAR;
+	vSize[E4VT_NO_TEXTURE].Format = VERTEX4D_FORMAT_COLOR_2_FOG;
 	vSize[E4VT_NO_TEXTURE].Pitch = sizeof(S3DVertex);
 	vSize[E4VT_NO_TEXTURE].TexSize = 0;
 	vSize[E4VT_NO_TEXTURE].TexCooSize = 0;
+	vSize[E4VT_NO_TEXTURE].ColSize = 2;
 
 	//Line
 	vSize[E4VT_LINE].Format = VERTEX4D_FORMAT_COLOR_1;
 	vSize[E4VT_LINE].Pitch = sizeof(S3DVertex);
 	vSize[E4VT_LINE].TexSize = 0;
 	vSize[E4VT_LINE].TexCooSize = 0;
+	vSize[E4VT_LINE].ColSize = 1;
 
 	//verify with global defines
 	u32 size;
@@ -1952,7 +2007,7 @@ void CBurningVideoDriver::VertexCache_map_source_format()
 		u32& flag = vSize[i].Format;
 
 #if !defined(SOFTWARE_DRIVER_2_USE_SEPARATE_SPECULAR_COLOR)
-		flag &= ~VERTEX4D_FORMAT_SPECULAR;
+		//flag &= ~VERTEX4D_FORMAT_SPECULAR;
 #endif
 
 		if (vSize[i].TexSize > BURNING_MATERIAL_MAX_TEXTURES)
@@ -2036,9 +2091,29 @@ void CBurningVideoDriver::VertexCache_fill(const u32 sourceIndex, const u32 dest
 		sVec4 gl_Color; gl_Color.setA8R8G8B8(base->Color.color);
 
 		// Irrlicht TCoords and TCoords2 must be contiguous memory. baseTCoord has no 4 byte aligned start address!
-		sVec4 gl_MultiTexCoord0(base->TCoords.X, base->TCoords.Y, 1.f, 1.f);
-		sVec4 gl_MultiTexCoord1(((S3DVertex2TCoords*)source)->TCoords2.X, ((S3DVertex2TCoords*)source)->TCoords2.Y, 1.f, 1.f);
+		sVec4 gl_MultiTexCoord[4];
+		const sVec2Pack* baseTCoord = (const sVec2Pack*)&base->TCoords.X;
+		for (u32 m = 0; m < array_size(gl_MultiTexCoord); ++m)
+		{
+			if (m < VertexShader.vSize[VertexShader.vType].TexCooSize)
+			{
+				gl_MultiTexCoord[m].s = baseTCoord[m].x;
+				gl_MultiTexCoord[m].t = baseTCoord[m].y;
+			}
+			else
+			{
+				gl_MultiTexCoord[m].s = 0.f;
+				gl_MultiTexCoord[m].t = 0.f;
+			}
 
+			gl_MultiTexCoord[m].p = 1.f;
+			gl_MultiTexCoord[m].q = 1.f;
+		}
+
+#define gl_MultiTexCoord0 gl_MultiTexCoord[0]
+#define gl_MultiTexCoord1 gl_MultiTexCoord[1]
+#define gl_MultiTexCoord2 gl_MultiTexCoord[2]
+#define gl_MultiTexCoord3 gl_MultiTexCoord[3]
 #define gl_Position dest->Pos
 #define gl_TexCoord dest->Tex
 #define gl_FrontColor dest->Color[0]
@@ -2052,9 +2127,6 @@ void CBurningVideoDriver::VertexCache_fill(const u32 sourceIndex, const u32 dest
 #define gl_ModelViewProjectionMatrix matrix[ETS_MODEL_VIEW_PROJ]
 #define ftransform() (matrix[ETS_MODEL_VIEW_PROJ] * gl_Vertex)
 
-		//#define uniform(var,name) const var& name = shader->uniform_##var(#name)
-		//#define varying(var,name) var name = shader->varying_##var(#name)
-
 #define uniform(var,name) const var& name = (const var&)*shader->getUniform(#name,BL_VERTEX_FLOAT)
 #define varying(var,name) var& name = (var&)*shader->getUniform(#name,BL_FRAGMENT_FLOAT)
 
@@ -2198,12 +2270,66 @@ void CBurningVideoDriver::VertexCache_fill(const u32 sourceIndex, const u32 dest
 			normal = normalize(gl_Normal);
 			lightdir2 = normalize(lightdir);
 		}
+		else if (Material.VertexShader == STK_1303_0xd872cdb6)
+		{
+			// Shader based on work by Fabien Sanglard
+			// Released under the terms of CC-BY 3.0
 
+			varying(vec3,lightVec);
+			varying(vec3,halfVec);
+			varying(vec3,eyeVec);
+
+			uniform(vec3,lightdir);
+
+			//void main()
+			{
+
+				gl_TexCoord[0] = gl_MultiTexCoord0;
+
+				// Building the matrix Eye Space -> Tangent Space
+				vec3 n = normalize(gl_NormalMatrix * gl_Normal);
+				// gl_MultiTexCoord1.xyz
+				vec3 t = normalize(gl_NormalMatrix * vec3(1.0, 0.0, 0.0)); // tangent
+				vec3 b = cross(n, t);
+
+				vec3 vertexPosition = vec3(gl_ModelViewMatrix * gl_Vertex);
+
+				// transform light and half angle vectors by tangent basis
+				vec3 v;
+				v.x = dot(lightdir, t);
+				v.y = dot(lightdir, b);
+				v.z = dot(lightdir, n);
+				lightVec = normalize(v);
+
+				vertexPosition = normalize(vertexPosition);
+
+				eyeVec = normalize(-vertexPosition); // we are in Eye Coordinates, so EyePos is (0,0,0)
+
+				// Normalize the halfVector to pass it to the fragment shader
+
+				// No need to divide by two, the result is normalized anyway.
+				// vec3 halfVector = normalize((vertexPosition + lightDir) / 2.0);
+				vec3 halfVector = normalize(vertexPosition + lightdir);
+				v.x = dot(halfVector, t);
+				v.y = dot(halfVector, b);
+				v.z = dot(halfVector, n);
+
+				// No need to normalize, t,b,n and halfVector are normal vectors.
+				//normalize (v);
+				halfVec = v;
+
+				gl_Position = ftransform();
+			}
+		}
 #ifdef _MSC_VER
 #pragma warning (default: 4244) // conversion
 #pragma warning (default: 4305) // truncation
 #endif
 
+#undef gl_MultiTexCoord0
+#undef gl_MultiTexCoord1
+#undef gl_MultiTexCoord2
+#undef gl_MultiTexCoord3
 #undef vec2
 #undef vec3
 #undef vec4
@@ -2335,7 +2461,7 @@ fftransform:
 			break;
 		}
 
-		sVec4* a = dest->Color + ((VertexShader.vSize[VertexShader.vType].Format & VERTEX4D_FORMAT_COLOR_2_FOG) ? 1 : 0);
+		sVec4* a = dest->Color + (((VertexShader.vSize[VertexShader.vType].Format & VERTEX4D_FORMAT_MASK_COLOR)>=VERTEX4D_FORMAT_COLOR_2_FOG) ? 1 : 0);
 		a->a = clampf01(fog_factor);
 	}
 
@@ -2497,42 +2623,71 @@ fftransform:
 
 #if BURNING_MATERIAL_MAX_LIGHT_TANGENT > 0
 	if ((EyeSpace.TL_Flag & TL_LIGHT0_IS_NORMAL_MAP) &&
-		((VertexShader.vSize[VertexShader.vType].Format & VERTEX4D_FORMAT_MASK_TANGENT) == VERTEX4D_FORMAT_BUMP_DOT3)
+		((VertexShader.vSize[VertexShader.vType].Format & VERTEX4D_FORMAT_MASK_TANGENT) >= VERTEX4D_FORMAT_BUMP_DOT3)
 		)
 	{
 		const S3DVertexTangents* tangent = ((S3DVertexTangents*)source);
 
 		sVec4 vp;
+
 		sVec4 light_accu;
-		light_accu.x = 0.f;
-		light_accu.y = 0.f;
-		light_accu.z = 0.f;
-		light_accu.w = 0.f;
-		for (u32 i = 0; i < 2 && i < EyeSpace.Light.size(); ++i)
+		light_accu.set(0.f);
+
+		/*
+		* Color[0] lightcolor[0] a: vertexalpha
+		* Color[1] lightcolor[1] a: fogdistance
+		* Color[2] lightvector[0] 
+		* Color[3] lightvector[1] 
+		* LightVector[0] eyevector
+		*/
+
+		if ((VertexShader.vSize[VertexShader.vType].Format & VERTEX4D_FORMAT_MASK_TANGENT) >= VERTEX4D_FORMAT_PARALLAX)
+		{
+			vp.x = EyeSpace.leye.x - base->Pos.X;
+			vp.y = EyeSpace.leye.y - base->Pos.Y;
+			vp.z = EyeSpace.leye.z - base->Pos.Z;
+
+			light_accu.x = vp.dot(tangent->Tangent);
+			light_accu.y = -vp.dot(tangent->Binormal);
+			light_accu.z = -vp.dot(tangent->Normal);
+			light_accu.normalize_pack_xyz(dest->LightTangent[0], 1.f, 0.f);
+		}
+
+		const u32 maxLight = core::min_((u32)BURNING_MATERIAL_MAX_COLORS,(u32)2, EyeSpace.Light.size());
+		for (u32 i = 0; i < maxLight; ++i)
 		{
 			const SBurningShaderLight& light = EyeSpace.Light[i];
 			if (!light.LightIsOn)
 				continue;
 
-			// lightcolor with standard model
-			// but shader is different. treating light and vertex in same space
-#if 1
-			vp.x = light.pos.x - base->Pos.X;
-			vp.y = light.pos.y - base->Pos.Y;
-			vp.z = light.pos.z - base->Pos.Z;
-#else
-			vp.x = light.pos4.x - EyeSpace.vertex.x;
-			vp.y = light.pos4.y - EyeSpace.vertex.y;
-			vp.z = light.pos4.z - EyeSpace.vertex.z;
-#endif
+			vp.x = light.pos_local.x - base->Pos.X;
+			vp.y = light.pos_local.y - base->Pos.Y;
+			vp.z = light.pos_local.z - base->Pos.Z;
 
-			// transform by tangent matrix
+			f32 attenuation = inversesqrt(vp.dot_xyz(vp) * light.nmap_linearAttenuation);
+			dest->Color[i].r = light.DiffuseColor.r * attenuation;
+			dest->Color[i].g = light.DiffuseColor.g * attenuation;
+			dest->Color[i].b = light.DiffuseColor.b * attenuation;
+			dest->Color[i].clampf01();
+
+			// lightvector transform by tangent matrix
+#if BURNING_MATERIAL_MAX_COLORS >=4
+			dest->Color[2+i].x = vp.dot(tangent->Tangent);
+			dest->Color[2+i].y = vp.dot(tangent->Binormal);
+			dest->Color[2+i].z = vp.dot(tangent->Normal);
+			dest->Color[2 + i].normalize_dir_xyz_zero();
+#else
 			light_accu.x += (vp.x * tangent->Tangent.X + vp.y * tangent->Tangent.Y + vp.z * tangent->Tangent.Z);
 			light_accu.y += (vp.x * tangent->Binormal.X + vp.y * tangent->Binormal.Y + vp.z * tangent->Binormal.Z);
 			light_accu.z += (vp.x * tangent->Normal.X + vp.y * tangent->Normal.Y + vp.z * tangent->Normal.Z);
+#endif
 		}
+
+#if BURNING_MATERIAL_MAX_COLORS >=4
+#else
 		//normalize [-1,+1] to [0,1] -> obsolete
 		light_accu.normalize_pack_xyz(dest->LightTangent[0], 1.f, 0.f);
+#endif
 		dest->Tex[1].x = dest->Tex[0].x;
 		dest->Tex[1].y = dest->Tex[0].y;
 
@@ -2887,6 +3042,11 @@ int CBurningVideoDriver::VertexCache_reset(const void* vertices, u32 vertexCount
 
 	switch (Material.org.MaterialType) // (Material.Fallback_MaterialType)
 	{
+	case EMT_PARALLAX_MAP_SOLID:
+	case EMT_PARALLAX_MAP_TRANSPARENT_ADD_COLOR:
+	case EMT_PARALLAX_MAP_TRANSPARENT_VERTEX_ALPHA:
+		VertexShader.vType = vType == EVT_TANGENTS ? E4VT_TANGENTS_PARALLAX : (e4DVertexType)vType;
+		break;
 	case EMT_REFLECTION_2_LAYER:
 	case EMT_TRANSPARENT_REFLECTION_2_LAYER:
 		VertexShader.vType = vType == EVT_STANDARD ? E4VT_REFLECTION_MAP : (e4DVertexType)vType;
@@ -2937,12 +3097,7 @@ void CBurningVideoDriver::drawVertexPrimitiveList(const void* vertices, u32 vert
 		return;
 
 	pushShader(pType, 1);
-	/*
-		if ((u32)Material.org.MaterialType < MaterialRenderers.size())
-		{
-			MaterialRenderers[Material.org.MaterialType].Renderer->OnRender(this, vType);
-		}
-	*/
+
 	//Matrices needed for this primitive
 	transform_calc(ETS_MODEL_VIEW_PROJ);
 	if ((EyeSpace.TL_Flag & (TL_TEXTURE_TRANSFORM | TL_FOG | TL_LIGHT)) ||
@@ -2952,6 +3107,37 @@ void CBurningVideoDriver::drawVertexPrimitiveList(const void* vertices, u32 vert
 		transform_calc(ETS_NORMAL);
 	}
 
+	//objectspace
+	if (EyeSpace.TL_Flag & TL_LIGHT0_IS_NORMAL_MAP)
+	{
+		transform_calc(ETS_MODEL_INVERSE);
+
+		const core::matrix4* matrix = Transformation[TransformationStack];
+		if ((VertexShader.vSize[VertexShader.vType].Format & VERTEX4D_FORMAT_MASK_TANGENT) >= VERTEX4D_FORMAT_PARALLAX)
+		{
+			transform_calc(ETS_MODEL_VIEW);
+			mat44_inverse(EyeSpace.mvi, matrix[ETS_MODEL_VIEW]);
+
+			sVec4 eye(0.f, 0.f, 0.f, 1.f);
+			transformVec3Vec3(EyeSpace.mvi, &EyeSpace.leye.x, &eye.x);
+		}
+
+		const u32 maxLight = core::min_((u32)BURNING_MATERIAL_MAX_COLORS, (u32)2, EyeSpace.Light.size());
+		for (u32 i = 0; i < maxLight; ++i)
+		{
+			SBurningShaderLight& light = EyeSpace.Light[i];
+			if (!light.LightIsOn)
+				continue;
+
+			transformVec3Vec3(matrix[ETS_MODEL_INVERSE], &light.pos_local.x, &light.pos.x);
+		}
+
+	}
+
+	if ((u32)Material.org.MaterialType < MaterialRenderers.size())
+	{
+		MaterialRenderers[Material.org.MaterialType].Renderer->OnRender(this, (video::E_VERTEX_TYPE)VertexShader.vType);
+	}
 
 	s4DVertexPair* face[4];
 
@@ -3087,106 +3273,72 @@ void CBurningVideoDriver::drawVertexPrimitiveList(const void* vertices, u32 vert
 						continue; //not break; per clipper triangle
 
 
-					// mipmap2 
-#if 1
 				// select mipmap
 #if BURNING_MAX_MIP_CLIPPER == 1
-					if (probe == use_max_mip)
+				if (probe == use_max_mip)
 #endif
-						for (u32 m = 0; m < VertexShader.vSize[VertexShader.vType].TexSize; ++m)
-						{
-							video::CSoftwareTexture2* tex = MAT_TEXTURE(m);
-
-							const sVec2& v0 = (face[0] + s4DVertex_ofs(0))->Tex[m];
-							const sVec2& v1 = (face[1] + s4DVertex_ofs(0))->Tex[m];
-							const sVec2& v2 = (face[2] + s4DVertex_ofs(0))->Tex[m];
-
-							//todo: get triangle setup here
-							//bbox
-							t[0] = t[2] = v0.x;
-							t[1] = t[3] = v0.y;
-							if (v1.x < t[0]) t[0] = v1.x;
-							if (v1.y < t[1]) t[1] = v1.y;
-							if (v1.x > t[2]) t[2] = v1.x;
-							if (v1.y > t[3]) t[3] = v1.y;
-
-							if (v2.x < t[0]) t[0] = v2.x;
-							if (v2.y < t[1]) t[1] = v2.y;
-							if (v2.x > t[2]) t[2] = v2.x;
-							if (v2.y > t[3]) t[3] = v2.y;
-
-							f32 tex_area = fabsf((t[2] - t[0]) * (t[3] - t[1]));
-
-							//tex_area = |a.x * b.y - b.x * a.y| * 0.5; // crossproduct
-							//f32 tex_area = fabsf((v1.x - v0.x) * (v2.y - v0.y) - (v2.x - v0.x) * (v1.y - v0.y));
-
-							//various over and underflow cases
-							if (tex_area <= 0.000001f)
-								tex_area = 0.000001f;
-							else if (tex_area > 1.01f)
-								tex_area = 1.f / tex_area;
-
-							/* 2.f * tex_area * 1.6f bias. 1.6 too much for detailsmap3 */
-							//not dc_area * 0.5 cancel out 2 * TexBias
-							const u32 dc_area_over_tex_area = (u32)floorf( /*/tex_area > 0.0000001f ? */
-								fabsf(dc_area.f) * TexBias[TransformationStack] / tex_area
-								/*: 0.f*/
-							);
-
-							// get a near 1:1 ratio between pixel and texel
-							// tex_area * b[lodFactor].w * b[lodFactor].h > dc_area_abs
-
-							s32 lodFactor = 0;
-							const CSoftwareTexture2_Bound* b = tex->getTexBound_index();
-							while (lodFactor < SOFTWARE_DRIVER_2_MIPMAPPING_MAX &&
-								b[lodFactor].area > dc_area_over_tex_area
-								)
-							{
-								lodFactor += 1;
-							}
-
-							//clipped triangle should take single area based mipmap from unclipped face
-							//skybox,billboard test case 
-							//if (vertex_from_clipper) lodFactor -= 1;
-							if (has_vertex_run == 0) lod_max[m] = lodFactor;
-							else if (lodFactor < lod_max[m]) lod_max[m] = lodFactor;
-
-							//CurrentShader->setTextureParam(m, tex, lodFactor);
-							//select_polygon_mipmap_inside(face, m, tex->getTexBound());
-						}
-
-#else
-				//select mipmap ratio between drawing space and texture space (for multiply divide here)
-					dc_area.f = reciprocal_zero(dc_area.f);
-
-					// select mipmap
 					for (u32 m = 0; m < VertexShader.vSize[VertexShader.vType].TexSize; ++m)
 					{
 						video::CSoftwareTexture2* tex = MAT_TEXTURE(m);
 
-						//only guessing: take more detail (lower mipmap) in light+bump textures
-						f32 lod_bias = 0.100f;// core::clamp(map_value((f32)ScreenSize.Width * ScreenSize.Height, 160 * 120, 640 * 480, 1.f / 8.f, 1.f / 8.f), 0.01f, 1.f);
+						const sVec2& v0 = (face[0] + s4DVertex_ofs(0))->Tex[m];
+						const sVec2& v1 = (face[1] + s4DVertex_ofs(0))->Tex[m];
+						const sVec2& v2 = (face[2] + s4DVertex_ofs(0))->Tex[m];
 
-						//assume transparent add is ~50% transparent -> more detail
-						switch (Material.org.MaterialType)
+						//todo: get triangle setup here
+						//bbox
+						t[0] = t[2] = v0.x;
+						t[1] = t[3] = v0.y;
+						if (v1.x < t[0]) t[0] = v1.x;
+						if (v1.y < t[1]) t[1] = v1.y;
+						if (v1.x > t[2]) t[2] = v1.x;
+						if (v1.y > t[3]) t[3] = v1.y;
+
+						if (v2.x < t[0]) t[0] = v2.x;
+						if (v2.y < t[1]) t[1] = v2.y;
+						if (v2.x > t[2]) t[2] = v2.x;
+						if (v2.y > t[3]) t[3] = v2.y;
+
+						f32 tex_area = fabsf((t[2] - t[0]) * (t[3] - t[1]));
+
+						//tex_area = |a.x * b.y - b.x * a.y| * 0.5; // crossproduct
+						//f32 tex_area = fabsf((v1.x - v0.x) * (v2.y - v0.y) - (v2.x - v0.x) * (v1.y - v0.y));
+
+						//various over and underflow cases
+						if (tex_area <= 0.000001f)
+							tex_area = 0.000001f;
+						else if (tex_area > 1.01f)
+							tex_area = 1.f / tex_area;
+
+						/* 2.f * tex_area * 1.6f bias. 1.6 too much for detailsmap3 */
+						//not dc_area * 0.5 cancel out 2 * TexBias
+						const u32 dc_area_over_tex_area = (u32)floorf( /*/tex_area > 0.0000001f ? */
+							fabsf(dc_area.f) * TexBias[TransformationStack] / tex_area
+							/*: 0.f*/
+						);
+
+						// get a near 1:1 ratio between pixel and texel
+						// tex_area * b[lodFactor].w * b[lodFactor].h > dc_area_abs
+
+						s32 lodFactor = 0;
+						const CSoftwareTexture2_Bound* b = tex->getTexBound_index();
+						while (lodFactor < SOFTWARE_DRIVER_2_MIPMAPPING_MAX &&
+							b[lodFactor].area > dc_area_over_tex_area
+							)
 						{
-						case EMT_TRANSPARENT_ADD_COLOR:
-						case EMT_TRANSPARENT_ALPHA_CHANNEL:
-							lod_bias *= 0.5f;
-							break;
-						default:
-							break;
+							lodFactor += 1;
 						}
-						lod_bias *= tex->get_lod_bias();
-						//lod_bias += Material.org.TextureLayer[m].LODBias * 0.125f;
 
-						s32 lodFactor = lodFactor_inside(face, m, dc_area.f, lod_bias);
+						//clipped triangle should take single area based mipmap from unclipped face
+						//skybox,billboard test case 
+						//if (vertex_from_clipper) lodFactor -= 1;
+						if (has_vertex_run == 0) lod_max[m] = lodFactor;
+						else if (lodFactor < lod_max[m]) lod_max[m] = lodFactor;
 
-						CurrentShader->setTextureParam(m, tex, lodFactor);
-						//currently shader receives texture coordinate as Pixelcoo of 1 Texture
-						select_polygon_mipmap_inside(face, m, tex->getTexBound());
+						//CurrentShader->setTextureParam(m, tex, lodFactor);
+						//select_polygon_mipmap_inside(face, m, tex->getTexBound());
 					}
-#endif
+
 				}
 				//else /* if (VertexShader.primitiveHasVertex == 3) */
 #if BURNING_MAX_MIP_CLIPPER == 1
@@ -3349,6 +3501,7 @@ void CBurningVideoDriver::assignHardwareLight(SBurningShaderLight& l, const SLig
 	transformVec4Vec4(matrix[ETS_MODEL_VIEW], &l.pos4.x, &l.pos.x);
 	rotateMat44Vec3Vec4(matrix[ETS_MODEL_VIEW], &l.spotDirection4.x, &l.spotDirection.x);
 
+	l.nmap_linearAttenuation = 1.f / (0.001f + dl.Radius * dl.Radius);
 
 	/*
 	//case ELT_DIRECTIONAL:
@@ -3432,13 +3585,14 @@ void CBurningVideoDriver::setMaterial(const SMaterial& material)
 
 	const SMaterial& in = Material.org;
 
+	const u32 shaderid = (u32)in.MaterialType;
+
 	//basically set always. 2d does its own compare
 	//if (TransformationStack == ETF_STACK_2D ||	Material.resetRenderStates || compare_3d_material(Material.lastMaterial,in))
 	{
 		// ---------- Notify Shader
 		// unset old material
 		u32 shaderid_old = (u32)Material.lastMaterial.MaterialType;
-		u32 shaderid = (u32)in.MaterialType;
 
 		if (shaderid != shaderid_old && shaderid_old < MaterialRenderers.size())
 		{
@@ -3470,7 +3624,8 @@ void CBurningVideoDriver::setMaterial(const SMaterial& material)
 	EyeSpace.TL_Flag &= ~(TL_TEXTURE_TRANSFORM | TL_LIGHT0_IS_NORMAL_MAP);
 
 #ifdef SOFTWARE_DRIVER_2_TEXTURE_TRANSFORM
-	for (u32 m = 0; m < VertexShader.vSize[VertexShader.vType].TexSize; ++m)
+	//vertextype not set!
+	for (u32 m = 0; m < BURNING_MATERIAL_MAX_TEXTURES /*VertexShader.vSize[VertexShader.vType].TexSize*/; ++m)
 	{
 		flag[ETS_TEXTURE_0 + m] &= ~ETF_TEXGEN_MASK;
 		setTransform((E_TRANSFORMATION_STATE)(ETS_TEXTURE_0 + m), in.getTextureMatrix(m));
@@ -3520,7 +3675,7 @@ void CBurningVideoDriver::setMaterial(const SMaterial& material)
 
 	EBurningFFShader shader = Material.depth_test ? ETR_TEXTURE_GOURAUD : ETR_TEXTURE_GOURAUD_NOZ;
 
-	switch (Material.Fallback_MaterialType) //(Material.org.MaterialType)
+	switch (Material.Fallback_MaterialType) //(Material.org.MaterialType) // Material.Fallback_MaterialType)
 	{
 	case EMT_ONETEXTURE_BLEND:
 		shader = ETR_TEXTURE_BLEND;
@@ -3604,6 +3759,7 @@ void CBurningVideoDriver::setMaterial(const SMaterial& material)
 		{
 			shader = ETR_NORMAL_MAP_SOLID;
 			EyeSpace.TL_Flag |= TL_TEXTURE_TRANSFORM | TL_LIGHT0_IS_NORMAL_MAP;
+			EyeSpace.TL_Flag &= ~TL_LIGHT;
 		}
 		break;
 	case EMT_PARALLAX_MAP_SOLID:
@@ -3611,8 +3767,9 @@ void CBurningVideoDriver::setMaterial(const SMaterial& material)
 	case EMT_PARALLAX_MAP_TRANSPARENT_VERTEX_ALPHA:
 		if (texture1)
 		{
-			shader = ETR_NORMAL_MAP_SOLID;
+			shader = ETR_PARALLAX_MAP_SOLID;
 			EyeSpace.TL_Flag |= TL_TEXTURE_TRANSFORM | TL_LIGHT0_IS_NORMAL_MAP;
+			EyeSpace.TL_Flag &= ~TL_LIGHT;
 		}
 		break;
 
@@ -3653,6 +3810,11 @@ void CBurningVideoDriver::setMaterial(const SMaterial& material)
 
 	// switchToTriangleRenderer
 	CurrentShader = BurningShader[shader];
+	if (!CurrentShader && shaderid < MaterialRenderers.size())
+	{
+		CurrentShader = (IBurningShader *) MaterialRenderers[shaderid].Renderer;
+	}
+
 	if (CurrentShader)
 	{
 		CurrentShader->setTLFlag(EyeSpace.TL_Flag);
@@ -3664,12 +3826,13 @@ void CBurningVideoDriver::setMaterial(const SMaterial& material)
 	}
 
 
+#if 0
 	{
-		u32 shaderid = (u32)Material.org.MaterialType;
+		//u32 shaderid = (u32)Material.org.MaterialType;
 		if (shaderid < MaterialRenderers.size())
 			MaterialRenderers[shaderid].Renderer->OnRender(this, (video::E_VERTEX_TYPE)VertexShader.vType);
 	}
-
+#endif
 }
 
 
@@ -3913,7 +4076,7 @@ void CBurningVideoDriver::lightVertex_eye(s4DVertex* dest, const u32 vertexargb)
 	//separate specular
 	const sVec4& spec_mat = (EyeSpace.TL_Flag & TL_COLORMAT_SPECULAR) ? vertexColor : Material.SpecularColor;
 #if defined(SOFTWARE_DRIVER_2_USE_SEPARATE_SPECULAR_COLOR)
-	if ((VertexShader.vSize[VertexShader.vType].Format & VERTEX4D_FORMAT_COLOR_2_FOG))
+	if ((VertexShader.vSize[VertexShader.vType].Format & VERTEX4D_FORMAT_MASK_COLOR)>=VERTEX4D_FORMAT_COLOR_2_FOG)
 	{
 		specular.sat_mul_xyz(dest->Color[1], spec_mat);
 	}
@@ -4566,17 +4729,17 @@ void CBurningVideoDriver::pushShader(scene::E_PRIMITIVE_TYPE pType, int testCurr
 const wchar_t* CBurningVideoDriver::getName() const
 {
 #ifdef BURNINGVIDEO_RENDERER_BEAUTIFUL
-	return L"Burning's Video 0.53 beautiful";
+	return L"Burning's Video 0.54 beautiful";
 #elif defined(PATCH_SUPERTUX_8_0_1_with_1_9_0)
-	return L"Burning's Video 0.53 STK";
+	return L"Burning's Video 0.54 STK";
 #elif defined ( BURNINGVIDEO_RENDERER_ULTRA_FAST )
-	return L"Burning's Video 0.53 ultra fast";
+	return L"Burning's Video 0.54 ultra fast";
 #elif defined ( BURNINGVIDEO_RENDERER_FAST )
-	return L"Burning's Video 0.53 fast";
+	return L"Burning's Video 0.54 fast";
 #elif defined ( BURNINGVIDEO_RENDERER_CE )
-	return L"Burning's Video 0.53 CE";
+	return L"Burning's Video 0.54 CE";
 #else
-	return L"Burning's Video 0.53";
+	return L"Burning's Video 0.54";
 #endif
 }
 
diff --git a/source/Irrlicht/CSoftwareDriver2.h b/source/Irrlicht/CSoftwareDriver2.h
index ae645c7a..441ccd69 100644
--- a/source/Irrlicht/CSoftwareDriver2.h
+++ b/source/Irrlicht/CSoftwareDriver2.h
@@ -365,6 +365,8 @@ namespace video
 			ETS_MODEL_VIEW,
 			ETS_NORMAL, //3x3 ModelView Transposed Inverse
 
+			ETS_MODEL_INVERSE,	//normal,parallax
+
 			ETS_COUNT_BURNING = 16
 		};
 
diff --git a/source/Irrlicht/CTRGouraud2.cpp b/source/Irrlicht/CTRGouraud2.cpp
index 2a272e9c..b83e59ec 100644
--- a/source/Irrlicht/CTRGouraud2.cpp
+++ b/source/Irrlicht/CTRGouraud2.cpp
@@ -93,7 +93,7 @@ protected:
 
 //! constructor
 CTRGouraud2::CTRGouraud2(CBurningVideoDriver* driver)
-: IBurningShader(driver)
+: IBurningShader(driver, EMT_SOLID)
 {
 	#ifdef _DEBUG
 	setDebugName("CTRGouraud2");
diff --git a/source/Irrlicht/CTRGouraudAlpha2.cpp b/source/Irrlicht/CTRGouraudAlpha2.cpp
deleted file mode 100644
index 77028e6e..00000000
--- a/source/Irrlicht/CTRGouraudAlpha2.cpp
+++ /dev/null
@@ -1,651 +0,0 @@
-// Copyright (C) 2002-2012 Nikolaus Gebhardt / Thomas Alten
-// This file is part of the "Irrlicht Engine".
-// For conditions of distribution and use, see copyright notice in irrlicht.h
-
-#include "IrrCompileConfig.h"
-#include "IBurningShader.h"
-
-#ifdef _IRR_COMPILE_WITH_BURNINGSVIDEO_
-
-// compile flag for this file
-#undef USE_ZBUFFER
-#undef IPOL_Z
-#undef CMP_Z
-#undef WRITE_Z
-
-#undef IPOL_W
-#undef CMP_W
-#undef WRITE_W
-
-#undef SUBTEXEL
-#undef INVERSE_W
-
-#undef IPOL_C0
-#undef IPOL_T0
-#undef IPOL_T1
-
-// define render case
-#define SUBTEXEL
-#define INVERSE_W
-
-#define USE_ZBUFFER
-#define IPOL_W
-#define CMP_W
-#define WRITE_W
-
-#define IPOL_C0
-//#define IPOL_T0
-//#define IPOL_T1
-
-// apply global override
-#ifndef SOFTWARE_DRIVER_2_PERSPECTIVE_CORRECT
-	#undef INVERSE_W
-#endif
-
-#ifndef SOFTWARE_DRIVER_2_SUBTEXEL
-	#undef SUBTEXEL
-#endif
-
-#if BURNING_MATERIAL_MAX_COLORS < 1
-	#undef IPOL_C0
-#endif
-
-#if !defined ( SOFTWARE_DRIVER_2_USE_WBUFFER ) && defined ( USE_ZBUFFER )
-	#ifndef SOFTWARE_DRIVER_2_PERSPECTIVE_CORRECT
-		#undef IPOL_W
-	#endif
-	#define IPOL_Z
-
-	#ifdef CMP_W
-		#undef CMP_W
-		#define CMP_Z
-	#endif
-
-	#ifdef WRITE_W
-		#undef WRITE_W
-		#define WRITE_Z
-	#endif
-
-#endif
-
-
-
-namespace irr
-{
-
-namespace video
-{
-
-class CTRGouraudAlpha2 : public IBurningShader
-{
-public:
-
-	//! constructor
-	CTRGouraudAlpha2(CBurningVideoDriver* driver);
-
-	//! draws an indexed triangle list
-	virtual void drawTriangle ( const s4DVertex* burning_restrict a,const s4DVertex* burning_restrict b,const s4DVertex* burning_restrict c ) IRR_OVERRIDE;
-
-
-private:
-	void fragmentShader();
-
-};
-
-//! constructor
-CTRGouraudAlpha2::CTRGouraudAlpha2(CBurningVideoDriver* driver)
-: IBurningShader(driver)
-{
-	#ifdef _DEBUG
-	setDebugName("CTRGouraudAlpha2");
-	#endif
-}
-
-
-
-/*!
-*/
-void CTRGouraudAlpha2::fragmentShader()
-{
-	tVideoSample *dst;
-
-#ifdef USE_ZBUFFER
-	fp24 *z;
-#endif
-
-	s32 xStart;
-	s32 xEnd;
-	s32 dx;
-
-
-#ifdef SUBTEXEL
-	f32 subPixel;
-#endif
-
-#ifdef IPOL_Z
-	f32 slopeZ;
-#endif
-#ifdef IPOL_W
-	fp24 slopeW;
-#endif
-#ifdef IPOL_C0
-	sVec4 slopeC;
-#endif
-#ifdef IPOL_T0
-	sVec2 slopeT[0];
-#endif
-#ifdef IPOL_T1
-	sVec2 slopeT[1];
-#endif
-
-	// apply top-left fill-convention, left
-	xStart = fill_convention_left( line.x[0] );
-	xEnd = fill_convention_right( line.x[1] );
-
-	dx = xEnd - xStart;
-
-	if ( dx < 0 )
-		return;
-
-	// slopes
-	const f32 invDeltaX = fill_step_x( line.x[1] - line.x[0] );
-
-#ifdef IPOL_Z
-	slopeZ = (line.z[1] - line.z[0]) * invDeltaX;
-#endif
-#ifdef IPOL_W
-	slopeW = (line.w[1] - line.w[0]) * invDeltaX;
-#endif
-#ifdef IPOL_C0
-	slopeC = (line.c[0][1] - line.c[0][0]) * invDeltaX;
-#endif
-#ifdef IPOL_T0
-	slopeT[0] = (line.t[0][1] - line.t[0][0]) * invDeltaX;
-#endif
-#ifdef IPOL_T1
-	slopeT[1] = (line.t[1][1] - line.t[1][0]) * invDeltaX;
-#endif
-
-#ifdef SUBTEXEL
-	subPixel = ( (f32) xStart ) - line.x[0];
-#ifdef IPOL_Z
-	line.z[0] += slopeZ * subPixel;
-#endif
-#ifdef IPOL_W
-	line.w[0] += slopeW * subPixel;
-#endif
-#ifdef IPOL_C0
-	line.c[0][0] += slopeC * subPixel;
-#endif
-#ifdef IPOL_T0
-	line.t[0][0] += slopeT[0] * subPixel;
-#endif
-#ifdef IPOL_T1
-	line.t[1][0] += slopeT[1] * subPixel;
-#endif
-#endif
-
-	SOFTWARE_DRIVER_2_CLIPCHECK;
-	dst = (tVideoSample*)RenderTarget->getData() + ( line.y * RenderTarget->getDimension().Width ) + xStart;
-
-#ifdef USE_ZBUFFER
-	z = (fp24*) DepthBuffer->lock() + ( line.y * RenderTarget->getDimension().Width ) + xStart;
-#endif
-
-
-
-#ifdef IPOL_C0
-
-	f32 inversew = FIX_POINT_F32_MUL;
-
-	tFixPoint a0;
-	tFixPoint r0, g0, b0;
-	tFixPoint r1, g1, b1;
-	tFixPoint r2, g2, b2;
-#endif
-
-	for ( s32 i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
-	{
-#ifdef CMP_Z
-		if ( line.z[0] < z[i] )
-#endif
-#ifdef CMP_W
-		if ( line.w[0] >= z[i] )
-#endif
-
-		{
-#ifdef IPOL_C0
-#ifdef INVERSE_W
-			inversew = reciprocal_zero( line.w[0] );
-#endif
-			vec4_to_fix( a0, r0, g0, b0, line.c[0][0],inversew );
-
-			color_to_fix ( r1, g1, b1, dst[i] );
-
-			fix_color_norm(a0);
-			r2 = r1 + imulFix ( a0, r0 - r1 );
-			g2 = g1 + imulFix ( a0, g0 - g1 );
-			b2 = b1 + imulFix ( a0, b0 - b1 );
-
-			dst[i] = fix_to_sample( r2, g2, b2 );
-#else
-			dst[i] = PrimitiveColor;
-#endif
-#ifdef WRITE_Z
-			z[i] = line.z[0];
-#endif
-#ifdef WRITE_W
-			z[i] = line.w[0];
-#endif
-
-		}
-
-#ifdef IPOL_Z
-		line.z[0] += slopeZ;
-#endif
-#ifdef IPOL_W
-		line.w[0] += slopeW;
-#endif
-#ifdef IPOL_C0
-		line.c[0][0] += slopeC;
-#endif
-#ifdef IPOL_T0
-		line.t[0][0] += slopeT[0];
-#endif
-#ifdef IPOL_T1
-		line.t[1][0] += slopeT[1];
-#endif
-	}
-
-}
-
-void CTRGouraudAlpha2::drawTriangle(const s4DVertex* burning_restrict a, const s4DVertex* burning_restrict b, const s4DVertex* burning_restrict c)
-{
-	// sort on height, y
-	if ( F32_A_GREATER_B ( a->Pos.y , b->Pos.y ) ) swapVertexPointer(&a, &b);
-	if ( F32_A_GREATER_B ( b->Pos.y , c->Pos.y ) ) swapVertexPointer(&b, &c);
-	if ( F32_A_GREATER_B ( a->Pos.y , b->Pos.y ) ) swapVertexPointer(&a, &b);
-
-	const f32 ca = c->Pos.y - a->Pos.y;
-	const f32 ba = b->Pos.y - a->Pos.y;
-	const f32 cb = c->Pos.y - b->Pos.y;
-	// calculate delta y of the edges
-	scan.invDeltaY[0] = fill_step_y( ca );
-	scan.invDeltaY[1] = fill_step_y( ba );
-	scan.invDeltaY[2] = fill_step_y( cb );
-
-	if ( F32_LOWER_EQUAL_0 ( scan.invDeltaY[0] ) )
-		return;
-
-	// find if the major edge is left or right aligned
-	f32 temp[4];
-
-	temp[0] = a->Pos.x - c->Pos.x;
-	temp[1] = -ca;
-	temp[2] = b->Pos.x - a->Pos.x;
-	temp[3] = ba;
-
-	scan.left = ( temp[0] * temp[3] - temp[1] * temp[2] ) < 0.f ? 1 : 0;
-	scan.right = 1 - scan.left;
-
-	// calculate slopes for the major edge
-	scan.slopeX[0] = (c->Pos.x - a->Pos.x) * scan.invDeltaY[0];
-	scan.x[0] = a->Pos.x;
-
-#ifdef IPOL_Z
-	scan.slopeZ[0] = (c->Pos.z - a->Pos.z) * scan.invDeltaY[0];
-	scan.z[0] = a->Pos.z;
-#endif
-
-#ifdef IPOL_W
-	scan.slopeW[0] = (c->Pos.w - a->Pos.w) * scan.invDeltaY[0];
-	scan.w[0] = a->Pos.w;
-#endif
-
-#ifdef IPOL_C0
-	scan.slopeC[0][0] = (c->Color[0] - a->Color[0]) * scan.invDeltaY[0];
-	scan.c[0][0] = a->Color[0];
-#endif
-
-#ifdef IPOL_T0
-	scan.slopeT[0][0] = (c->Tex[0] - a->Tex[0]) * scan.invDeltaY[0];
-	scan.t[0][0] = a->Tex[0];
-#endif
-
-#ifdef IPOL_T1
-	scan.slopeT[1][0] = (c->Tex[1] - a->Tex[1]) * scan.invDeltaY[0];
-	scan.t[1][0] = a->Tex[1];
-#endif
-
-	// top left fill convention y run
-	s32 yStart;
-	s32 yEnd;
-
-#ifdef SUBTEXEL
-	f32 subPixel;
-#endif
-
-
-	// rasterize upper sub-triangle
-	if ( (f32) 0.0 != scan.invDeltaY[1]  )
-	{
-		// calculate slopes for top edge
-		scan.slopeX[1] = (b->Pos.x - a->Pos.x) * scan.invDeltaY[1];
-		scan.x[1] = a->Pos.x;
-
-#ifdef IPOL_Z
-		scan.slopeZ[1] = (b->Pos.z - a->Pos.z) * scan.invDeltaY[1];
-		scan.z[1] = a->Pos.z;
-#endif
-
-#ifdef IPOL_W
-		scan.slopeW[1] = (b->Pos.w - a->Pos.w) * scan.invDeltaY[1];
-		scan.w[1] = a->Pos.w;
-#endif
-
-#ifdef IPOL_C0
-		scan.slopeC[0][1] = (b->Color[0] - a->Color[0]) * scan.invDeltaY[1];
-		scan.c[0][1] = a->Color[0];
-#endif
-
-#ifdef IPOL_T0
-		scan.slopeT[0][1] = (b->Tex[0] - a->Tex[0]) * scan.invDeltaY[1];
-		scan.t[0][1] = a->Tex[0];
-#endif
-
-#ifdef IPOL_T1
-		scan.slopeT[1][1] = (b->Tex[1] - a->Tex[1]) * scan.invDeltaY[1];
-		scan.t[1][1] = a->Tex[1];
-#endif
-
-		// apply top-left fill convention, top part
-		yStart = fill_convention_top( a->Pos.y );
-		yEnd = fill_convention_down( b->Pos.y );
-
-#ifdef SUBTEXEL
-		subPixel = ( (f32) yStart ) - a->Pos.y;
-
-		// correct to pixel center
-		scan.x[0] += scan.slopeX[0] * subPixel;
-		scan.x[1] += scan.slopeX[1] * subPixel;
-
-#ifdef IPOL_Z
-		scan.z[0] += scan.slopeZ[0] * subPixel;
-		scan.z[1] += scan.slopeZ[1] * subPixel;
-#endif
-
-#ifdef IPOL_W
-		scan.w[0] += scan.slopeW[0] * subPixel;
-		scan.w[1] += scan.slopeW[1] * subPixel;
-#endif
-
-#ifdef IPOL_C0
-		scan.c[0][0] += scan.slopeC[0][0] * subPixel;
-		scan.c[0][1] += scan.slopeC[0][1] * subPixel;
-#endif
-
-#ifdef IPOL_T0
-		scan.t[0][0] += scan.slopeT[0][0] * subPixel;
-		scan.t[0][1] += scan.slopeT[0][1] * subPixel;
-#endif
-
-#ifdef IPOL_T1
-		scan.t[1][0] += scan.slopeT[1][0] * subPixel;
-		scan.t[1][1] += scan.slopeT[1][1] * subPixel;
-#endif
-
-#endif
-
-		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; line.y += SOFTWARE_DRIVER_2_STEP_Y)
-		{
-			line.x[scan.left] = scan.x[0];
-			line.x[scan.right] = scan.x[1];
-
-#ifdef IPOL_Z
-			line.z[scan.left] = scan.z[0];
-			line.z[scan.right] = scan.z[1];
-#endif
-
-#ifdef IPOL_W
-			line.w[scan.left] = scan.w[0];
-			line.w[scan.right] = scan.w[1];
-#endif
-
-#ifdef IPOL_C0
-			line.c[0][scan.left] = scan.c[0][0];
-			line.c[0][scan.right] = scan.c[0][1];
-#endif
-
-#ifdef IPOL_T0
-			line.t[0][scan.left] = scan.t[0][0];
-			line.t[0][scan.right] = scan.t[0][1];
-#endif
-
-#ifdef IPOL_T1
-			line.t[1][scan.left] = scan.t[1][0];
-			line.t[1][scan.right] = scan.t[1][1];
-#endif
-
-			// render a scanline
-			if_interlace_scanline fragmentShader();
-
-			scan.x[0] += scan.slopeX[0];
-			scan.x[1] += scan.slopeX[1];
-
-#ifdef IPOL_Z
-			scan.z[0] += scan.slopeZ[0];
-			scan.z[1] += scan.slopeZ[1];
-#endif
-
-#ifdef IPOL_W
-			scan.w[0] += scan.slopeW[0];
-			scan.w[1] += scan.slopeW[1];
-#endif
-
-#ifdef IPOL_C0
-			scan.c[0][0] += scan.slopeC[0][0];
-			scan.c[0][1] += scan.slopeC[0][1];
-#endif
-
-#ifdef IPOL_T0
-			scan.t[0][0] += scan.slopeT[0][0];
-			scan.t[0][1] += scan.slopeT[0][1];
-#endif
-
-#ifdef IPOL_T1
-			scan.t[1][0] += scan.slopeT[1][0];
-			scan.t[1][1] += scan.slopeT[1][1];
-#endif
-
-		}
-	}
-
-	// rasterize lower sub-triangle
-	if ( (f32) 0.0 != scan.invDeltaY[2] )
-	{
-		// advance to middle point
-		if( (f32) 0.0 != scan.invDeltaY[1] )
-		{
-			temp[0] = b->Pos.y - a->Pos.y;	// dy
-
-			scan.x[0] = a->Pos.x + scan.slopeX[0] * temp[0];
-#ifdef IPOL_Z
-			scan.z[0] = a->Pos.z + scan.slopeZ[0] * temp[0];
-#endif
-#ifdef IPOL_W
-			scan.w[0] = a->Pos.w + scan.slopeW[0] * temp[0];
-#endif
-#ifdef IPOL_C0
-			scan.c[0][0] = a->Color[0] + scan.slopeC[0][0] * temp[0];
-#endif
-#ifdef IPOL_T0
-			scan.t[0][0] = a->Tex[0] + scan.slopeT[0][0] * temp[0];
-#endif
-#ifdef IPOL_T1
-			scan.t[1][0] = a->Tex[1] + scan.slopeT[1][0] * temp[0];
-#endif
-
-		}
-
-		// calculate slopes for bottom edge
-		scan.slopeX[1] = (c->Pos.x - b->Pos.x) * scan.invDeltaY[2];
-		scan.x[1] = b->Pos.x;
-
-#ifdef IPOL_Z
-		scan.slopeZ[1] = (c->Pos.z - b->Pos.z) * scan.invDeltaY[2];
-		scan.z[1] = b->Pos.z;
-#endif
-
-#ifdef IPOL_W
-		scan.slopeW[1] = (c->Pos.w - b->Pos.w) * scan.invDeltaY[2];
-		scan.w[1] = b->Pos.w;
-#endif
-
-#ifdef IPOL_C0
-		scan.slopeC[0][1] = (c->Color[0] - b->Color[0]) * scan.invDeltaY[2];
-		scan.c[0][1] = b->Color[0];
-#endif
-
-#ifdef IPOL_T0
-		scan.slopeT[0][1] = (c->Tex[0] - b->Tex[0]) * scan.invDeltaY[2];
-		scan.t[0][1] = b->Tex[0];
-#endif
-
-#ifdef IPOL_T1
-		scan.slopeT[1][1] = (c->Tex[1] - b->Tex[1]) * scan.invDeltaY[2];
-		scan.t[1][1] = b->Tex[1];
-#endif
-
-		// apply top-left fill convention, top part
-		yStart = fill_convention_top( b->Pos.y );
-		yEnd = fill_convention_down( c->Pos.y );
-
-#ifdef SUBTEXEL
-		subPixel = ( (f32) yStart ) - b->Pos.y;
-
-		// correct to pixel center
-		scan.x[0] += scan.slopeX[0] * subPixel;
-		scan.x[1] += scan.slopeX[1] * subPixel;
-
-#ifdef IPOL_Z
-		scan.z[0] += scan.slopeZ[0] * subPixel;
-		scan.z[1] += scan.slopeZ[1] * subPixel;
-#endif
-
-#ifdef IPOL_W
-		scan.w[0] += scan.slopeW[0] * subPixel;
-		scan.w[1] += scan.slopeW[1] * subPixel;
-#endif
-
-#ifdef IPOL_C0
-		scan.c[0][0] += scan.slopeC[0][0] * subPixel;
-		scan.c[0][1] += scan.slopeC[0][1] * subPixel;
-#endif
-
-#ifdef IPOL_T0
-		scan.t[0][0] += scan.slopeT[0][0] * subPixel;
-		scan.t[0][1] += scan.slopeT[0][1] * subPixel;
-#endif
-
-#ifdef IPOL_T1
-		scan.t[1][0] += scan.slopeT[1][0] * subPixel;
-		scan.t[1][1] += scan.slopeT[1][1] * subPixel;
-#endif
-
-#endif
-
-		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; line.y += SOFTWARE_DRIVER_2_STEP_Y)
-		{
-			line.x[scan.left] = scan.x[0];
-			line.x[scan.right] = scan.x[1];
-
-#ifdef IPOL_Z
-			line.z[scan.left] = scan.z[0];
-			line.z[scan.right] = scan.z[1];
-#endif
-
-#ifdef IPOL_W
-			line.w[scan.left] = scan.w[0];
-			line.w[scan.right] = scan.w[1];
-#endif
-
-#ifdef IPOL_C0
-			line.c[0][scan.left] = scan.c[0][0];
-			line.c[0][scan.right] = scan.c[0][1];
-#endif
-
-#ifdef IPOL_T0
-			line.t[0][scan.left] = scan.t[0][0];
-			line.t[0][scan.right] = scan.t[0][1];
-#endif
-
-#ifdef IPOL_T1
-			line.t[1][scan.left] = scan.t[1][0];
-			line.t[1][scan.right] = scan.t[1][1];
-#endif
-
-			// render a scanline
-			if_interlace_scanline fragmentShader();
-
-			scan.x[0] += scan.slopeX[0];
-			scan.x[1] += scan.slopeX[1];
-
-#ifdef IPOL_Z
-			scan.z[0] += scan.slopeZ[0];
-			scan.z[1] += scan.slopeZ[1];
-#endif
-
-#ifdef IPOL_W
-			scan.w[0] += scan.slopeW[0];
-			scan.w[1] += scan.slopeW[1];
-#endif
-
-#ifdef IPOL_C0
-			scan.c[0][0] += scan.slopeC[0][0];
-			scan.c[0][1] += scan.slopeC[0][1];
-#endif
-
-#ifdef IPOL_T0
-			scan.t[0][0] += scan.slopeT[0][0];
-			scan.t[0][1] += scan.slopeT[0][1];
-#endif
-
-#ifdef IPOL_T1
-			scan.t[1][0] += scan.slopeT[1][0];
-			scan.t[1][1] += scan.slopeT[1][1];
-#endif
-
-		}
-	}
-
-}
-
-} // end namespace video
-} // end namespace irr
-
-#endif // _IRR_COMPILE_WITH_BURNINGSVIDEO_
-
-namespace irr
-{
-namespace video
-{
-
-//! creates a flat triangle renderer
-IBurningShader* createTriangleRendererGouraudAlpha2(CBurningVideoDriver* driver)
-{
-	// ETR_GOURAUD_ALPHA unused
-	#ifdef _IRR_COMPILE_WITH_BURNINGSVIDEO_
-	return new CTRGouraudAlpha2(driver);
-	#else
-	return 0;
-	#endif // _IRR_COMPILE_WITH_BURNINGSVIDEO_
-}
-
-
-} // end namespace video
-} // end namespace irr
-
-
-
diff --git a/source/Irrlicht/CTRGouraudAlphaNoZ2.cpp b/source/Irrlicht/CTRGouraudAlphaNoZ2.cpp
index ee4188d4..0db0f588 100644
--- a/source/Irrlicht/CTRGouraudAlphaNoZ2.cpp
+++ b/source/Irrlicht/CTRGouraudAlphaNoZ2.cpp
@@ -93,7 +93,7 @@ private:
 
 //! constructor
 CTRGouraudAlphaNoZ2::CTRGouraudAlphaNoZ2(CBurningVideoDriver* driver)
-: IBurningShader(driver)
+: IBurningShader(driver, EMT_TRANSPARENT_ALPHA_CHANNEL)
 {
 	#ifdef _DEBUG
 	setDebugName("CTRGouraudAlphaNoZ2");
diff --git a/source/Irrlicht/CTRGouraudNoZ2.cpp b/source/Irrlicht/CTRGouraudNoZ2.cpp
index 9bc9c16a..c7a21877 100644
--- a/source/Irrlicht/CTRGouraudNoZ2.cpp
+++ b/source/Irrlicht/CTRGouraudNoZ2.cpp
@@ -92,7 +92,7 @@ protected:
 
 //! constructor
 CTRGouraudNoZ2::CTRGouraudNoZ2(CBurningVideoDriver* driver)
-: IBurningShader(driver)
+: IBurningShader(driver, EMT_SOLID)
 {
 	#ifdef _DEBUG
 	setDebugName("CTRGouraudNoZ2");
diff --git a/source/Irrlicht/CTRNormalMap.cpp b/source/Irrlicht/CTRNormalMap.cpp
index e833729a..1a028b00 100644
--- a/source/Irrlicht/CTRNormalMap.cpp
+++ b/source/Irrlicht/CTRNormalMap.cpp
@@ -4,6 +4,7 @@
 
 #include "IrrCompileConfig.h"
 #include "IBurningShader.h"
+#include "CSoftwareDriver2.h"
 
 #ifdef _IRR_COMPILE_WITH_BURNINGSVIDEO_
 
@@ -22,6 +23,8 @@
 
 #undef IPOL_C0
 #undef IPOL_C1
+#undef IPOL_C2
+#undef IPOL_C3
 #undef IPOL_T0
 #undef IPOL_T1
 #undef IPOL_T2
@@ -38,9 +41,11 @@
 
 #define IPOL_C0
 #define IPOL_C1
+#define IPOL_C2
+#define IPOL_C3
 #define IPOL_T0
 #define IPOL_T1
-#define IPOL_L0
+//#define IPOL_L0
 
 // apply global override
 #ifndef SOFTWARE_DRIVER_2_PERSPECTIVE_CORRECT
@@ -90,26 +95,35 @@ class CTRNormalMap : public IBurningShader
 public:
 
 	//! constructor
-	CTRNormalMap(CBurningVideoDriver* driver);
+	CTRNormalMap(CBurningVideoDriver* driver,s32& outMaterialTypeNr, E_MATERIAL_TYPE baseMaterial);
+	~CTRNormalMap();
 
 	//! draws an indexed triangle list
 	virtual void drawTriangle(const s4DVertex* burning_restrict a, const s4DVertex* burning_restrict b, const s4DVertex* burning_restrict c) IRR_OVERRIDE;
 	virtual void OnSetMaterial(const SBurningShaderMaterial& material) IRR_OVERRIDE;
-
+	virtual void OnSetConstants(IMaterialRendererServices* services, s32 userData) IRR_OVERRIDE;
 private:
 	void fragmentShader();
 
 };
 
+
 //! constructor
-CTRNormalMap::CTRNormalMap(CBurningVideoDriver* driver)
-	: IBurningShader(driver)
+CTRNormalMap::CTRNormalMap(CBurningVideoDriver* driver, s32& outMaterialTypeNr, E_MATERIAL_TYPE baseMaterial)
+	: IBurningShader(driver, baseMaterial)
 {
 #ifdef _DEBUG
 	setDebugName("CTRNormalMap");
 #endif
+	CallBack = this;
+	outMaterialTypeNr = driver->addMaterialRenderer(this);
 }
 
+CTRNormalMap::~CTRNormalMap()
+{
+	if (CallBack == this)
+		CallBack = 0;
+}
 
 void CTRNormalMap::OnSetMaterial(const SBurningShaderMaterial& material)
 {
@@ -174,6 +188,12 @@ void CTRNormalMap::fragmentShader()
 #ifdef IPOL_C1
 	slopeC[1] = (line.c[1][1] - line.c[1][0]) * invDeltaX;
 #endif
+#ifdef IPOL_C2
+	slopeC[2] = (line.c[2][1] - line.c[2][0]) * invDeltaX;
+#endif
+#ifdef IPOL_C3
+	slopeC[3] = (line.c[3][1] - line.c[3][0]) * invDeltaX;
+#endif
 #ifdef IPOL_T0
 	slopeT[0] = (line.t[0][1] - line.t[0][0]) * invDeltaX;
 #endif
@@ -201,6 +221,12 @@ void CTRNormalMap::fragmentShader()
 #ifdef IPOL_C1
 	line.c[1][0] += slopeC[1] * subPixel;
 #endif
+#ifdef IPOL_C2
+	line.c[2][0] += slopeC[2] * subPixel;
+#endif
+#ifdef IPOL_C3
+	line.c[3][0] += slopeC[3] * subPixel;
+#endif
 #ifdef IPOL_T0
 	line.t[0][0] += slopeT[0] * subPixel;
 #endif
@@ -249,6 +275,10 @@ void CTRNormalMap::fragmentShader()
 	tFixPoint aFog = FIX_POINT_ONE;
 #endif
 
+#ifdef IPOL_C2
+	tFixPoint lx, ly, lz;
+#endif	
+
 
 	for (s32 i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
@@ -301,10 +331,13 @@ void CTRNormalMap::fragmentShader()
 				// normal map ( same texcoord0 but different mipmapping)
 				getSample_texture(r1, g1, b1, &IT[1], tx1, ty1);
 
+				// normal: xyz * 2 - 1
 				r1 = (r1 - FIX_POINT_HALF_COLOR) >> (COLOR_MAX_LOG2 - 1);
 				g1 = (g1 - FIX_POINT_HALF_COLOR) >> (COLOR_MAX_LOG2 - 1);
 				b1 = (b1 - FIX_POINT_HALF_COLOR) >> (COLOR_MAX_LOG2 - 1);
 
+
+				//lightvector
 #ifdef IPOL_L0
 				lx = tofix(line.l[0][0].x, inversew);
 				ly = tofix(line.l[0][0].y, inversew);
@@ -314,18 +347,35 @@ void CTRNormalMap::fragmentShader()
 				//max(dot(LightVector, Normal), 0.0);
 				ndotl = clampfix_mincolor((imulFix_simple(r1, lx) + imulFix_simple(g1, ly) + imulFix_simple(b1, lz)));
 #endif
+				lx = tofix(line.c[2][0].x, inversew);
+				ly = tofix(line.c[2][0].y, inversew);
+				lz = tofix(line.c[2][0].z, inversew);
+				//omit normalize
+				ndotl = clampfix_mincolor((imulFix_simple(r1, lx) + imulFix_simple(g1, ly) + imulFix_simple(b1, lz)));
 
 #ifdef IPOL_C0
 
-				//LightColor[0]
-				r3 = tofix(line.c[0][0].r, inversew);
-				g3 = tofix(line.c[0][0].g, inversew);
-				b3 = tofix(line.c[0][0].b, inversew);
+				//LightColor[0] * lambert
+				r3 = imulFix_simple(tofix(line.c[0][0].r, inversew), ndotl);
+				g3 = imulFix_simple(tofix(line.c[0][0].g, inversew), ndotl);
+				b3 = imulFix_simple(tofix(line.c[0][0].b, inversew), ndotl);
 
-				// Lambert * LightColor[0] * Diffuse Texture;
-				r2 = imulFix(imulFix_simple(r3, ndotl), r0);
-				g2 = imulFix(imulFix_simple(g3, ndotl), g0);
-				b2 = imulFix(imulFix_simple(b3, ndotl), b0);
+				//lightvector1
+				lx = tofix(line.c[3][0].x, inversew);
+				ly = tofix(line.c[3][0].y, inversew);
+				lz = tofix(line.c[3][0].z, inversew);
+				//omit normalize
+				ndotl = clampfix_mincolor((imulFix_simple(r1, lx) + imulFix_simple(g1, ly) + imulFix_simple(b1, lz)));
+
+				//LightColor[1] * lambert
+				r3 += imulFix_simple(tofix(line.c[1][0].r, inversew), ndotl);
+				g3 += imulFix_simple(tofix(line.c[1][0].g, inversew), ndotl);
+				b3 += imulFix_simple(tofix(line.c[1][0].b, inversew), ndotl);
+
+				// (Lambert0 * LightColor[0] + Lambert1 * LightColor[1]) *  Diffuse Texture;
+				r2 = clampfix_maxcolor(imulFix_simple(r3, r0));
+				g2 = clampfix_maxcolor(imulFix_simple(g3, g0));
+				b2 = clampfix_maxcolor(imulFix_simple(b3, b0));
 
 				//vertex alpha blend ( and omit depthwrite ,hacky..)
 				if (a3 + 2 < FIX_POINT_ONE)
@@ -369,6 +419,12 @@ void CTRNormalMap::fragmentShader()
 #ifdef IPOL_C1
 		line.c[1][0] += slopeC[1];
 #endif
+#ifdef IPOL_C2
+		line.c[2][0] += slopeC[2];
+#endif
+#ifdef IPOL_C3
+		line.c[3][0] += slopeC[3];
+#endif
 #ifdef IPOL_T0
 		line.t[0][0] += slopeT[0];
 #endif
@@ -438,6 +494,16 @@ void CTRNormalMap::drawTriangle(const s4DVertex* burning_restrict a, const s4DVe
 	scan.c[1][0] = a->Color[1];
 #endif
 
+#ifdef IPOL_C2
+	scan.slopeC[2][0] = (c->Color[2] - a->Color[2]) * scan.invDeltaY[0];
+	scan.c[2][0] = a->Color[2];
+#endif
+
+#ifdef IPOL_C3
+	scan.slopeC[3][0] = (c->Color[3] - a->Color[3]) * scan.invDeltaY[0];
+	scan.c[3][0] = a->Color[3];
+#endif
+
 #ifdef IPOL_T0
 	scan.slopeT[0][0] = (c->Tex[0] - a->Tex[0]) * scan.invDeltaY[0];
 	scan.t[0][0] = a->Tex[0];
@@ -494,6 +560,16 @@ void CTRNormalMap::drawTriangle(const s4DVertex* burning_restrict a, const s4DVe
 		scan.c[1][1] = a->Color[1];
 #endif
 
+#ifdef IPOL_C2
+		scan.slopeC[2][1] = (b->Color[2] - a->Color[2]) * scan.invDeltaY[1];
+		scan.c[2][1] = a->Color[2];
+#endif
+
+#ifdef IPOL_C3
+		scan.slopeC[3][1] = (b->Color[3] - a->Color[3]) * scan.invDeltaY[1];
+		scan.c[3][1] = a->Color[3];
+#endif
+
 #ifdef IPOL_T0
 		scan.slopeT[0][1] = (b->Tex[0] - a->Tex[0]) * scan.invDeltaY[1];
 		scan.t[0][1] = a->Tex[0];
@@ -545,6 +621,16 @@ void CTRNormalMap::drawTriangle(const s4DVertex* burning_restrict a, const s4DVe
 		scan.c[1][1] += scan.slopeC[1][1] * subPixel;
 #endif
 
+#ifdef IPOL_C2
+		scan.c[2][0] += scan.slopeC[2][0] * subPixel;
+		scan.c[2][1] += scan.slopeC[2][1] * subPixel;
+#endif
+
+#ifdef IPOL_C3
+		scan.c[3][0] += scan.slopeC[3][0] * subPixel;
+		scan.c[3][1] += scan.slopeC[3][1] * subPixel;
+#endif
+
 #ifdef IPOL_T0
 		scan.t[0][0] += scan.slopeT[0][0] * subPixel;
 		scan.t[0][1] += scan.slopeT[0][1] * subPixel;
@@ -593,6 +679,16 @@ void CTRNormalMap::drawTriangle(const s4DVertex* burning_restrict a, const s4DVe
 			line.c[1][scan.right] = scan.c[1][1];
 #endif
 
+#ifdef IPOL_C2
+			line.c[2][scan.left] = scan.c[2][0];
+			line.c[2][scan.right] = scan.c[2][1];
+#endif
+
+#ifdef IPOL_C3
+			line.c[3][scan.left] = scan.c[3][0];
+			line.c[3][scan.right] = scan.c[3][1];
+#endif
+
 #ifdef IPOL_T0
 			line.t[0][scan.left] = scan.t[0][0];
 			line.t[0][scan.right] = scan.t[0][1];
@@ -639,6 +735,16 @@ void CTRNormalMap::drawTriangle(const s4DVertex* burning_restrict a, const s4DVe
 			scan.c[1][1] += scan.slopeC[1][1];
 #endif
 
+#ifdef IPOL_C2
+			scan.c[2][0] += scan.slopeC[2][0];
+			scan.c[2][1] += scan.slopeC[2][1];
+#endif
+
+#ifdef IPOL_C3
+			scan.c[3][0] += scan.slopeC[3][0];
+			scan.c[3][1] += scan.slopeC[3][1];
+#endif
+
 #ifdef IPOL_T0
 			scan.t[0][0] += scan.slopeT[0][0];
 			scan.t[0][1] += scan.slopeT[0][1];
@@ -683,6 +789,12 @@ void CTRNormalMap::drawTriangle(const s4DVertex* burning_restrict a, const s4DVe
 #ifdef IPOL_C1
 			scan.c[1][0] = a->Color[1] + scan.slopeC[1][0] * temp[0];
 #endif
+#ifdef IPOL_C2
+			scan.c[2][0] = a->Color[2] + scan.slopeC[2][0] * temp[0];
+#endif
+#ifdef IPOL_C3
+			scan.c[3][0] = a->Color[3] + scan.slopeC[3][0] * temp[0];
+#endif
 #ifdef IPOL_T0
 			scan.t[0][0] = a->Tex[0] + scan.slopeT[0][0] * temp[0];
 #endif
@@ -722,6 +834,16 @@ void CTRNormalMap::drawTriangle(const s4DVertex* burning_restrict a, const s4DVe
 		scan.c[1][1] = b->Color[1];
 #endif
 
+#ifdef IPOL_C2
+		scan.slopeC[2][1] = (c->Color[2] - b->Color[2]) * scan.invDeltaY[2];
+		scan.c[2][1] = b->Color[2];
+#endif
+
+#ifdef IPOL_C3
+		scan.slopeC[3][1] = (c->Color[3] - b->Color[3]) * scan.invDeltaY[2];
+		scan.c[3][1] = b->Color[3];
+#endif
+
 #ifdef IPOL_T0
 		scan.slopeT[0][1] = (c->Tex[0] - b->Tex[0]) * scan.invDeltaY[2];
 		scan.t[0][1] = b->Tex[0];
@@ -773,6 +895,16 @@ void CTRNormalMap::drawTriangle(const s4DVertex* burning_restrict a, const s4DVe
 		scan.c[1][1] += scan.slopeC[1][1] * subPixel;
 #endif
 
+#ifdef IPOL_C2
+		scan.c[2][0] += scan.slopeC[2][0] * subPixel;
+		scan.c[2][1] += scan.slopeC[2][1] * subPixel;
+#endif
+
+#ifdef IPOL_C3
+		scan.c[3][0] += scan.slopeC[3][0] * subPixel;
+		scan.c[3][1] += scan.slopeC[3][1] * subPixel;
+#endif
+
 #ifdef IPOL_T0
 		scan.t[0][0] += scan.slopeT[0][0] * subPixel;
 		scan.t[0][1] += scan.slopeT[0][1] * subPixel;
@@ -821,6 +953,16 @@ void CTRNormalMap::drawTriangle(const s4DVertex* burning_restrict a, const s4DVe
 			line.c[1][scan.right] = scan.c[1][1];
 #endif
 
+#ifdef IPOL_C2
+			line.c[2][scan.left] = scan.c[2][0];
+			line.c[2][scan.right] = scan.c[2][1];
+#endif
+
+#ifdef IPOL_C3
+			line.c[3][scan.left] = scan.c[3][0];
+			line.c[3][scan.right] = scan.c[3][1];
+#endif
+
 #ifdef IPOL_T0
 			line.t[0][scan.left] = scan.t[0][0];
 			line.t[0][scan.right] = scan.t[0][1];
@@ -867,6 +1009,16 @@ void CTRNormalMap::drawTriangle(const s4DVertex* burning_restrict a, const s4DVe
 			scan.c[1][1] += scan.slopeC[1][1];
 #endif
 
+#ifdef IPOL_C2
+			scan.c[2][0] += scan.slopeC[2][0];
+			scan.c[2][1] += scan.slopeC[2][1];
+#endif
+
+#ifdef IPOL_C3
+			scan.c[3][0] += scan.slopeC[3][0];
+			scan.c[3][1] += scan.slopeC[3][1];
+#endif
+
 #ifdef IPOL_T0
 			scan.t[0][0] += scan.slopeT[0][0];
 			scan.t[0][1] += scan.slopeT[0][1];
@@ -887,10 +1039,62 @@ void CTRNormalMap::drawTriangle(const s4DVertex* burning_restrict a, const s4DVe
 #endif
 
 		}
-	}
+}
 
 }
 
+//! Called by the engine when the vertex and/or pixel shader constants for an
+//! material renderer should be set.
+void CTRNormalMap::OnSetConstants(IMaterialRendererServices* services, s32 userData)
+{
+#if 0
+	video::IVideoDriver* driver = services->getVideoDriver();
+
+	// set transposed world matrix
+	const core::matrix4& tWorld = driver->getTransform(video::ETS_WORLD).getTransposed();
+	services->setVertexShaderConstant(tWorld.pointer(), 0, 4);
+
+	// set transposed worldViewProj matrix
+	core::matrix4 worldViewProj(driver->getTransform(video::ETS_PROJECTION));
+	worldViewProj *= driver->getTransform(video::ETS_VIEW);
+	worldViewProj *= driver->getTransform(video::ETS_WORLD);
+	core::matrix4 tr(worldViewProj.getTransposed());
+	services->setVertexShaderConstant(tr.pointer(), 8, 4);
+
+	// here we fetch the fixed function lights from the driver
+	// and set them as constants
+
+	u32 cnt = driver->getDynamicLightCount();
+
+	// Load the inverse world matrix.
+	core::matrix4 invWorldMat;
+	driver->getTransform(video::ETS_WORLD).getInverse(invWorldMat);
+
+	for (u32 i = 0; i < 2; ++i)
+	{
+		video::SLight light;
+
+		if (i < cnt)
+			light = driver->getDynamicLight(i);
+		else
+		{
+			light.DiffuseColor.set(0, 0, 0); // make light dark
+			light.Radius = 1.0f;
+		}
+
+		light.DiffuseColor.a = 1.0f / (light.Radius * light.Radius); // set attenuation
+
+		// Transform the light by the inverse world matrix to get it into object space.
+		invWorldMat.transformVect(light.Position);
+
+		services->setVertexShaderConstant(
+			reinterpret_cast<const f32*>(&light.Position), 12 + (i * 2), 1);
+
+		services->setVertexShaderConstant(
+			reinterpret_cast<const f32*>(&light.DiffuseColor), 13 + (i * 2), 1);
+	}
+#endif
+}
 burning_namespace_end
 
 #endif // _IRR_COMPILE_WITH_BURNINGSVIDEO_
@@ -898,10 +1102,10 @@ burning_namespace_end
 burning_namespace_start
 
 //! creates a triangle renderer
-IBurningShader* createTRNormalMap(CBurningVideoDriver* driver)
+IBurningShader* createTRNormalMap(CBurningVideoDriver* driver, s32& outMaterialTypeNr, E_MATERIAL_TYPE baseMaterial)
 {
 #ifdef _IRR_COMPILE_WITH_BURNINGSVIDEO_
-	return new CTRNormalMap(driver);
+	return new CTRNormalMap(driver, outMaterialTypeNr,baseMaterial);
 #else
 	return 0;
 #endif // _IRR_COMPILE_WITH_BURNINGSVIDEO_
diff --git a/source/Irrlicht/CTRParallaxMap.cpp b/source/Irrlicht/CTRParallaxMap.cpp
new file mode 100644
index 00000000..fc9f555f
--- /dev/null
+++ b/source/Irrlicht/CTRParallaxMap.cpp
@@ -0,0 +1,1180 @@
+// Copyright (C) 2002-2012 Nikolaus Gebhardt / Thomas Alten
+// This file is part of the "Irrlicht Engine".
+// For conditions of distribution and use, see copyright notice in irrlicht.h
+
+#include "IrrCompileConfig.h"
+#include "IBurningShader.h"
+#include "CSoftwareDriver2.h"
+
+#ifdef _IRR_COMPILE_WITH_BURNINGSVIDEO_
+
+// compile flag for this file
+#undef USE_ZBUFFER
+#undef IPOL_Z
+#undef CMP_Z
+#undef WRITE_Z
+
+#undef IPOL_W
+#undef CMP_W
+#undef WRITE_W
+
+#undef SUBTEXEL
+#undef INVERSE_W
+
+#undef IPOL_C0
+#undef IPOL_C1
+#undef IPOL_C2
+#undef IPOL_C3
+#undef IPOL_T0
+#undef IPOL_T1
+#undef IPOL_T2
+#undef IPOL_L0
+
+// define render case
+#define SUBTEXEL
+#define INVERSE_W
+
+#define USE_ZBUFFER
+#define IPOL_W
+#define CMP_W
+#define WRITE_W
+
+#define IPOL_C0
+#define IPOL_C1
+#define IPOL_C2
+#define IPOL_C3
+#define IPOL_T0
+#define IPOL_T1
+#define IPOL_L0
+
+// apply global override
+#ifndef SOFTWARE_DRIVER_2_PERSPECTIVE_CORRECT
+#undef INVERSE_W
+#endif
+
+#ifndef SOFTWARE_DRIVER_2_SUBTEXEL
+#undef SUBTEXEL
+#endif
+
+#if BURNING_MATERIAL_MAX_COLORS < 1
+#undef IPOL_C0
+#endif
+
+#if BURNING_MATERIAL_MAX_COLORS < 2
+#undef IPOL_C1
+#endif
+
+#if BURNING_MATERIAL_MAX_COLORS < 3
+#undef IPOL_C2
+#endif
+
+#if BURNING_MATERIAL_MAX_COLORS < 4
+#undef IPOL_C3
+#endif
+
+#if BURNING_MATERIAL_MAX_LIGHT_TANGENT < 1
+#undef IPOL_L0
+#endif
+
+#if !defined ( SOFTWARE_DRIVER_2_USE_WBUFFER ) && defined ( USE_ZBUFFER )
+#ifndef SOFTWARE_DRIVER_2_PERSPECTIVE_CORRECT
+#undef IPOL_W
+#endif
+#define IPOL_Z
+
+#ifdef CMP_W
+#undef CMP_W
+#define CMP_Z
+#endif
+
+#ifdef WRITE_W
+#undef WRITE_W
+#define WRITE_Z
+#endif
+
+#endif
+
+
+burning_namespace_start
+
+
+class CBurningParallaxMap : public IBurningShader
+{
+public:
+
+	//! constructor
+	CBurningParallaxMap(CBurningVideoDriver* driver,s32& outMaterialTypeNr, E_MATERIAL_TYPE baseMaterial);
+	~CBurningParallaxMap();
+
+	//! draws an indexed triangle list
+	virtual void drawTriangle(const s4DVertex* burning_restrict a, const s4DVertex* burning_restrict b, const s4DVertex* burning_restrict c) IRR_OVERRIDE;
+	virtual void OnSetMaterial(const SBurningShaderMaterial& material) IRR_OVERRIDE;
+	virtual void OnSetMaterial(const video::SMaterial& material,
+		const video::SMaterial& lastMaterial,
+		bool resetAllRenderstates, video::IMaterialRendererServices* services) IRR_OVERRIDE;
+	virtual void OnSetConstants(IMaterialRendererServices* services, s32 userData) IRR_OVERRIDE;
+private:
+	void fragmentShader();
+	f32 CurrentScale;
+	tFixPoint CurrentScaleFix[2];
+};
+
+
+//! constructor
+CBurningParallaxMap::CBurningParallaxMap(CBurningVideoDriver* driver, s32& outMaterialTypeNr, E_MATERIAL_TYPE baseMaterial)
+	: IBurningShader(driver, baseMaterial)
+{
+#ifdef _DEBUG
+	setDebugName("CTRNormalMap");
+#endif
+	CallBack = this;
+	outMaterialTypeNr = driver->addMaterialRenderer(this);
+	CurrentScale = 0.02f;
+}
+
+CBurningParallaxMap::~CBurningParallaxMap()
+{
+	if (CallBack == this)
+		CallBack = 0;
+}
+
+void CBurningParallaxMap::OnSetMaterial(const SBurningShaderMaterial& material)
+{
+	CurrentScale = material.org.MaterialTypeParam;
+}
+
+void CBurningParallaxMap::OnSetMaterial(const video::SMaterial& material,
+	const video::SMaterial& lastMaterial,
+	bool resetAllRenderstates, video::IMaterialRendererServices* services)
+{
+	IBurningShader::OnSetMaterial(material, lastMaterial,
+		resetAllRenderstates, services);
+
+	CurrentScale = material.MaterialTypeParam;
+}
+
+/*!
+*/
+void CBurningParallaxMap::fragmentShader()
+{
+	tVideoSample* dst;
+
+#ifdef USE_ZBUFFER
+	fp24* z;
+#endif
+
+	s32 xStart;
+	s32 xEnd;
+	s32 dx;
+
+
+#ifdef SUBTEXEL
+	f32 subPixel;
+#endif
+
+#ifdef IPOL_Z
+	f32 slopeZ;
+#endif
+#ifdef IPOL_W
+	fp24 slopeW;
+#endif
+#ifdef IPOL_C0
+	sVec4 slopeC[BURNING_MATERIAL_MAX_COLORS];
+#endif
+#ifdef IPOL_T0
+	sVec2 slopeT[BURNING_MATERIAL_MAX_TEXTURES];
+#endif
+#ifdef IPOL_L0
+	sVec3Pack_unpack slopeL[BURNING_MATERIAL_MAX_LIGHT_TANGENT];
+#endif
+
+	// apply top-left fill-convention, left
+	xStart = fill_convention_left(line.x[0]);
+	xEnd = fill_convention_right(line.x[1]);
+
+	dx = xEnd - xStart;
+
+	if (dx < 0)
+		return;
+
+	// slopes
+	const f32 invDeltaX = fill_step_x(line.x[1] - line.x[0]);
+
+#ifdef IPOL_Z
+	slopeZ = (line.z[1] - line.z[0]) * invDeltaX;
+#endif
+#ifdef IPOL_W
+	slopeW = (line.w[1] - line.w[0]) * invDeltaX;
+#endif
+#ifdef IPOL_C0
+	slopeC[0] = (line.c[0][1] - line.c[0][0]) * invDeltaX;
+#endif
+#ifdef IPOL_C1
+	slopeC[1] = (line.c[1][1] - line.c[1][0]) * invDeltaX;
+#endif
+#ifdef IPOL_C2
+	slopeC[2] = (line.c[2][1] - line.c[2][0]) * invDeltaX;
+#endif
+#ifdef IPOL_C3
+	slopeC[3] = (line.c[3][1] - line.c[3][0]) * invDeltaX;
+#endif
+#ifdef IPOL_T0
+	slopeT[0] = (line.t[0][1] - line.t[0][0]) * invDeltaX;
+#endif
+#ifdef IPOL_T1
+	slopeT[1] = (line.t[1][1] - line.t[1][0]) * invDeltaX;
+#endif
+#ifdef IPOL_T2
+	slopeT[2] = (line.t[2][1] - line.t[2][0]) * invDeltaX;
+#endif
+#ifdef IPOL_L0
+	slopeL[0] = (line.l[0][1] - line.l[0][0]) * invDeltaX;
+#endif
+
+#ifdef SUBTEXEL
+	subPixel = ((f32)xStart) - line.x[0];
+#ifdef IPOL_Z
+	line.z[0] += slopeZ * subPixel;
+#endif
+#ifdef IPOL_W
+	line.w[0] += slopeW * subPixel;
+#endif
+#ifdef IPOL_C0
+	line.c[0][0] += slopeC[0] * subPixel;
+#endif
+#ifdef IPOL_C1
+	line.c[1][0] += slopeC[1] * subPixel;
+#endif
+#ifdef IPOL_C2
+	line.c[2][0] += slopeC[2] * subPixel;
+#endif
+#ifdef IPOL_C3
+	line.c[3][0] += slopeC[3] * subPixel;
+#endif
+#ifdef IPOL_T0
+	line.t[0][0] += slopeT[0] * subPixel;
+#endif
+#ifdef IPOL_T1
+	line.t[1][0] += slopeT[1] * subPixel;
+#endif
+#ifdef IPOL_T2
+	line.t[2][0] += slopeT[2] * subPixel;
+#endif
+#ifdef IPOL_L0
+	line.l[0][0] += slopeL[0] * subPixel;
+#endif
+#endif
+
+	SOFTWARE_DRIVER_2_CLIPCHECK;
+	dst = (tVideoSample*)RenderTarget->getData() + (line.y * RenderTarget->getDimension().Width) + xStart;
+
+#ifdef USE_ZBUFFER
+	z = (fp24*)DepthBuffer->lock() + (line.y * RenderTarget->getDimension().Width) + xStart;
+#endif
+
+
+	f32 inversew = FIX_POINT_F32_MUL;
+
+	tFixPoint tx0, ty0;
+
+#ifdef IPOL_T1
+	tFixPoint tx1, ty1;
+#endif
+
+	tFixPoint r0, g0, b0;
+	tFixPoint r1, g1, b1,a1;
+	tFixPoint r2, g2, b2;
+
+#ifdef IPOL_L0
+	tFixPoint lx, ly, lz;
+	sVec4 norm;
+#endif
+	tFixPoint ndotl = FIX_POINT_ONE;
+
+
+#ifdef IPOL_C0
+	tFixPoint a3, r3, g3, b3;
+#endif
+
+#ifdef IPOL_C1
+	tFixPoint aFog = FIX_POINT_ONE;
+#endif
+
+
+	for (s32 i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
+	{
+#ifdef CMP_Z
+		if (line.z[0] < z[i])
+#endif
+#ifdef CMP_W
+			if (line.w[0] >= z[i])
+#endif
+			{
+#ifdef INVERSE_W
+				inversew = fix_inverse32(line.w[0]);
+#endif
+
+#ifdef IPOL_C0
+				//vertex alpha blend ( and omit depthwrite ,hacky..)
+				a3 = tofix(line.c[0][0].a, inversew);
+				if (a3 + 2 >= FIX_POINT_ONE)
+				{
+#ifdef WRITE_Z
+					z[i] = line.z[0];
+#endif
+#ifdef WRITE_W
+					z[i] = line.w[0];
+#endif
+				}
+#endif
+
+#ifdef IPOL_C1
+				//complete inside fog
+				if (TL_Flag & TL_FOG)
+				{
+					aFog = tofix(line.c[1][0].a, inversew);
+					if (aFog <= 0)
+					{
+						dst[i] = fog_color_sample;
+						continue;
+					}
+				}
+#endif
+
+				tx1 = tofix(line.t[1][0].x, inversew);
+				ty1 = tofix(line.t[1][0].y, inversew);
+
+				// normal map height
+				getSample_texture(a1, &IT[1], tx1, ty1);
+
+				// xyz * 2 - 1
+				a1 = (a1 - FIX_POINT_HALF_COLOR) >> (COLOR_MAX_LOG2 - 1);
+				tFixPoint ofs = imulFix_simple(a1,CurrentScaleFix[0]);
+				//eyevector
+#ifdef IPOL_L0
+#if 0
+				norm.x = line.l[0][0].x * inversew;
+				norm.y = line.l[0][0].y * inversew;
+				norm.z = line.l[0][0].z * inversew;
+				norm.normalize_dir_xyz_zero();
+
+				lx = tofix(norm.x, FIX_POINT_F32_MUL);
+				ly = tofix(norm.y, FIX_POINT_F32_MUL);
+#else
+				lx = tofix(line.l[0][0].x, inversew);
+				ly = tofix(line.l[0][0].y, inversew);
+#endif
+				//lz = tofix(line.l[0][0].z, inversew);
+
+#endif
+				// vec2 TexCoord = EyeVector.xy * TempFetch.w + vTexCoord;
+				tx0 = tofix(line.t[0][0].x, inversew) + imulFix_simple(lx, ofs);
+				ty0 = tofix(line.t[0][0].y, inversew) + imulFix_simple(ly, ofs);
+				
+				// diffuse map
+				getSample_texture(r0, g0, b0, &IT[0], tx0, ty0);
+
+				ofs = imulFix_simple(a1, CurrentScaleFix[1]);
+				tx1 += imulFix_simple(lx, ofs);
+				ty1 += imulFix_simple(ly, ofs);
+
+				// normal map ( same texcoord0 but different mipmapping)
+				getSample_texture(r1, g1, b1, &IT[1], tx1, ty1);
+
+
+				// normal: xyz * 2 - 1
+				r1 = (r1 - FIX_POINT_HALF_COLOR) >> (COLOR_MAX_LOG2 - 1);
+				g1 = (g1 - FIX_POINT_HALF_COLOR) >> (COLOR_MAX_LOG2 - 1);
+				b1 = (b1 - FIX_POINT_HALF_COLOR) >> (COLOR_MAX_LOG2 - 1);
+
+
+				//lightvector
+				lx = tofix(line.c[2][0].x, inversew);
+				ly = tofix(line.c[2][0].y, inversew);
+				lz = tofix(line.c[2][0].z, inversew);
+				//omit normalize
+				//max(dot(LightVector, Normal), 0.0);
+				ndotl = clampfix_mincolor((imulFix_simple(r1, lx) + imulFix_simple(g1, ly) + imulFix_simple(b1, lz)));
+
+#ifdef IPOL_C0
+
+				//LightColor[0] * lambert
+				r3 = imulFix_simple(tofix(line.c[0][0].r, inversew), ndotl);
+				g3 = imulFix_simple(tofix(line.c[0][0].g, inversew), ndotl);
+				b3 = imulFix_simple(tofix(line.c[0][0].b, inversew), ndotl);
+
+				//lightvector1
+				lx = tofix(line.c[3][0].x, inversew);
+				ly = tofix(line.c[3][0].y, inversew);
+				lz = tofix(line.c[3][0].z, inversew);
+				//omit normalize
+				ndotl = clampfix_mincolor((imulFix_simple(r1, lx) + imulFix_simple(g1, ly) + imulFix_simple(b1, lz)));
+
+				//LightColor[1] * lambert
+				r3 += imulFix_simple(tofix(line.c[1][0].r, inversew), ndotl);
+				g3 += imulFix_simple(tofix(line.c[1][0].g, inversew), ndotl);
+				b3 += imulFix_simple(tofix(line.c[1][0].b, inversew), ndotl);
+
+				// (Lambert0 * LightColor[0] + Lambert1 * LightColor[1]) *  Diffuse Texture;
+				r2 = clampfix_maxcolor(imulFix_simple(r3, r0));
+				g2 = clampfix_maxcolor(imulFix_simple(g3, g0));
+				b2 = clampfix_maxcolor(imulFix_simple(b3, b0));
+
+				//vertex alpha blend ( and omit depthwrite ,hacky..)
+				if (a3 + 2 < FIX_POINT_ONE)
+				{
+					color_to_fix(r1, g1, b1, dst[i]);
+					r2 = r1 + imulFix(a3, r2 - r1);
+					g2 = g1 + imulFix(a3, g2 - g1);
+					b2 = b1 + imulFix(a3, b2 - b1);
+				}
+
+#ifdef IPOL_C1
+				//mix with distance
+				if (aFog < FIX_POINT_ONE) //TL_Flag & TL_FOG)
+				{
+					r2 = fog_color[1] + imulFix(aFog, r2 - fog_color[1]);
+					g2 = fog_color[2] + imulFix(aFog, g2 - fog_color[2]);
+					b2 = fog_color[3] + imulFix(aFog, b2 - fog_color[3]);
+				}
+#endif
+				dst[i] = fix_to_sample(r2, g2, b2);
+
+
+#else
+				r2 = imulFix_tex4(r0, r1);
+				g2 = imulFix_tex4(g0, g1);
+				b2 = imulFix_tex4(b0, b1);
+				dst[i] = fix_to_sample(r2, g2, b2);
+#endif
+
+			}
+
+#ifdef IPOL_Z
+		line.z[0] += slopeZ;
+#endif
+#ifdef IPOL_W
+		line.w[0] += slopeW;
+#endif
+#ifdef IPOL_C0
+		line.c[0][0] += slopeC[0];
+#endif
+#ifdef IPOL_C1
+		line.c[1][0] += slopeC[1];
+#endif
+#ifdef IPOL_C2
+		line.c[2][0] += slopeC[2];
+#endif
+#ifdef IPOL_C3
+		line.c[3][0] += slopeC[3];
+#endif
+#ifdef IPOL_T0
+		line.t[0][0] += slopeT[0];
+#endif
+#ifdef IPOL_T1
+		line.t[1][0] += slopeT[1];
+#endif
+#ifdef IPOL_T2
+		line.t[2][0] += slopeT[2];
+#endif
+#ifdef IPOL_L0
+		line.l[0][0] += slopeL[0];
+#endif
+	}
+
+}
+
+void CBurningParallaxMap::drawTriangle(const s4DVertex* burning_restrict a, const s4DVertex* burning_restrict b, const s4DVertex* burning_restrict c)
+{
+	// sort on height, y
+	if (F32_A_GREATER_B(a->Pos.y, b->Pos.y)) swapVertexPointer(&a, &b);
+	if (F32_A_GREATER_B(b->Pos.y, c->Pos.y)) swapVertexPointer(&b, &c);
+	if (F32_A_GREATER_B(a->Pos.y, b->Pos.y)) swapVertexPointer(&a, &b);
+
+	const f32 ca = c->Pos.y - a->Pos.y;
+	const f32 ba = b->Pos.y - a->Pos.y;
+	const f32 cb = c->Pos.y - b->Pos.y;
+	// calculate delta y of the edges
+	scan.invDeltaY[0] = fill_step_y(ca);
+	scan.invDeltaY[1] = fill_step_y(ba);
+	scan.invDeltaY[2] = fill_step_y(cb);
+
+	if (F32_LOWER_EQUAL_0(scan.invDeltaY[0]))
+		return;
+
+	CurrentScaleFix[0] = tofix(CurrentScale, FIX_POINT_F32_MUL) << (IT[0].pitchlog2- SOFTWARE_DRIVER_2_TEXTURE_GRANULARITY);
+	CurrentScaleFix[1] = tofix(CurrentScale, FIX_POINT_F32_MUL) << (IT[1].pitchlog2 - SOFTWARE_DRIVER_2_TEXTURE_GRANULARITY);
+
+	// find if the major edge is left or right aligned
+	f32 temp[4];
+
+	temp[0] = a->Pos.x - c->Pos.x;
+	temp[1] = -ca;
+	temp[2] = b->Pos.x - a->Pos.x;
+	temp[3] = ba;
+
+	scan.left = (temp[0] * temp[3] - temp[1] * temp[2]) < 0.f ? 1 : 0;
+	scan.right = 1 - scan.left;
+
+	// calculate slopes for the major edge
+	scan.slopeX[0] = (c->Pos.x - a->Pos.x) * scan.invDeltaY[0];
+	scan.x[0] = a->Pos.x;
+
+#ifdef IPOL_Z
+	scan.slopeZ[0] = (c->Pos.z - a->Pos.z) * scan.invDeltaY[0];
+	scan.z[0] = a->Pos.z;
+#endif
+
+#ifdef IPOL_W
+	scan.slopeW[0] = (c->Pos.w - a->Pos.w) * scan.invDeltaY[0];
+	scan.w[0] = a->Pos.w;
+#endif
+
+#ifdef IPOL_C0
+	scan.slopeC[0][0] = (c->Color[0] - a->Color[0]) * scan.invDeltaY[0];
+	scan.c[0][0] = a->Color[0];
+#endif
+
+#ifdef IPOL_C1
+	scan.slopeC[1][0] = (c->Color[1] - a->Color[1]) * scan.invDeltaY[0];
+	scan.c[1][0] = a->Color[1];
+#endif
+
+#ifdef IPOL_C2
+	scan.slopeC[2][0] = (c->Color[2] - a->Color[2]) * scan.invDeltaY[0];
+	scan.c[2][0] = a->Color[2];
+#endif
+
+#ifdef IPOL_C3
+	scan.slopeC[3][0] = (c->Color[3] - a->Color[3]) * scan.invDeltaY[0];
+	scan.c[3][0] = a->Color[3];
+#endif
+
+#ifdef IPOL_T0
+	scan.slopeT[0][0] = (c->Tex[0] - a->Tex[0]) * scan.invDeltaY[0];
+	scan.t[0][0] = a->Tex[0];
+#endif
+
+#ifdef IPOL_T1
+	scan.slopeT[1][0] = (c->Tex[1] - a->Tex[1]) * scan.invDeltaY[0];
+	scan.t[1][0] = a->Tex[1];
+#endif
+
+#ifdef IPOL_T2
+	scan.slopeT[2][0] = (c->Tex[2] - a->Tex[2]) * scan.invDeltaY[0];
+	scan.t[2][0] = a->Tex[2];
+#endif
+
+#ifdef IPOL_L0
+	scan.slopeL[0][0] = (c->LightTangent[0] - a->LightTangent[0]) * scan.invDeltaY[0];
+	scan.l[0][0] = a->LightTangent[0];
+#endif
+
+	// top left fill convention y run
+	s32 yStart;
+	s32 yEnd;
+
+#ifdef SUBTEXEL
+	f32 subPixel;
+#endif
+
+
+	// rasterize upper sub-triangle
+	if (F32_GREATER_0(scan.invDeltaY[1]))
+	{
+		// calculate slopes for top edge
+		scan.slopeX[1] = (b->Pos.x - a->Pos.x) * scan.invDeltaY[1];
+		scan.x[1] = a->Pos.x;
+
+#ifdef IPOL_Z
+		scan.slopeZ[1] = (b->Pos.z - a->Pos.z) * scan.invDeltaY[1];
+		scan.z[1] = a->Pos.z;
+#endif
+
+#ifdef IPOL_W
+		scan.slopeW[1] = (b->Pos.w - a->Pos.w) * scan.invDeltaY[1];
+		scan.w[1] = a->Pos.w;
+#endif
+
+#ifdef IPOL_C0
+		scan.slopeC[0][1] = (b->Color[0] - a->Color[0]) * scan.invDeltaY[1];
+		scan.c[0][1] = a->Color[0];
+#endif
+
+#ifdef IPOL_C1
+		scan.slopeC[1][1] = (b->Color[1] - a->Color[1]) * scan.invDeltaY[1];
+		scan.c[1][1] = a->Color[1];
+#endif
+
+#ifdef IPOL_C2
+		scan.slopeC[2][1] = (b->Color[2] - a->Color[2]) * scan.invDeltaY[1];
+		scan.c[2][1] = a->Color[2];
+#endif
+
+#ifdef IPOL_C3
+		scan.slopeC[3][1] = (b->Color[3] - a->Color[3]) * scan.invDeltaY[1];
+		scan.c[3][1] = a->Color[3];
+#endif
+
+#ifdef IPOL_T0
+		scan.slopeT[0][1] = (b->Tex[0] - a->Tex[0]) * scan.invDeltaY[1];
+		scan.t[0][1] = a->Tex[0];
+#endif
+
+#ifdef IPOL_T1
+		scan.slopeT[1][1] = (b->Tex[1] - a->Tex[1]) * scan.invDeltaY[1];
+		scan.t[1][1] = a->Tex[1];
+#endif
+
+#ifdef IPOL_T2
+		scan.slopeT[2][1] = (b->Tex[2] - a->Tex[2]) * scan.invDeltaY[1];
+		scan.t[2][1] = a->Tex[2];
+#endif
+
+#ifdef IPOL_L0
+		scan.slopeL[0][1] = (b->LightTangent[0] - a->LightTangent[0]) * scan.invDeltaY[1];
+		scan.l[0][1] = a->LightTangent[0];
+#endif
+
+		// apply top-left fill convention, top part
+		yStart = fill_convention_top(a->Pos.y);
+		yEnd = fill_convention_down(b->Pos.y);
+
+#ifdef SUBTEXEL
+		subPixel = ((f32)yStart) - a->Pos.y;
+
+		// correct to pixel center
+		scan.x[0] += scan.slopeX[0] * subPixel;
+		scan.x[1] += scan.slopeX[1] * subPixel;
+
+#ifdef IPOL_Z
+		scan.z[0] += scan.slopeZ[0] * subPixel;
+		scan.z[1] += scan.slopeZ[1] * subPixel;
+#endif
+
+#ifdef IPOL_W
+		scan.w[0] += scan.slopeW[0] * subPixel;
+		scan.w[1] += scan.slopeW[1] * subPixel;
+#endif
+
+#ifdef IPOL_C0
+		scan.c[0][0] += scan.slopeC[0][0] * subPixel;
+		scan.c[0][1] += scan.slopeC[0][1] * subPixel;
+#endif
+
+#ifdef IPOL_C1
+		scan.c[1][0] += scan.slopeC[1][0] * subPixel;
+		scan.c[1][1] += scan.slopeC[1][1] * subPixel;
+#endif
+
+#ifdef IPOL_C2
+		scan.c[2][0] += scan.slopeC[2][0] * subPixel;
+		scan.c[2][1] += scan.slopeC[2][1] * subPixel;
+#endif
+
+#ifdef IPOL_C3
+		scan.c[3][0] += scan.slopeC[3][0] * subPixel;
+		scan.c[3][1] += scan.slopeC[3][1] * subPixel;
+#endif
+
+#ifdef IPOL_T0
+		scan.t[0][0] += scan.slopeT[0][0] * subPixel;
+		scan.t[0][1] += scan.slopeT[0][1] * subPixel;
+#endif
+
+#ifdef IPOL_T1
+		scan.t[1][0] += scan.slopeT[1][0] * subPixel;
+		scan.t[1][1] += scan.slopeT[1][1] * subPixel;
+#endif
+
+#ifdef IPOL_T2
+		scan.t[2][0] += scan.slopeT[2][0] * subPixel;
+		scan.t[2][1] += scan.slopeT[2][1] * subPixel;
+#endif
+
+#ifdef IPOL_L0
+		scan.l[0][0] += scan.slopeL[0][0] * subPixel;
+		scan.l[0][1] += scan.slopeL[0][1] * subPixel;
+#endif
+
+#endif
+
+		// rasterize the edge scanlines
+		for (line.y = yStart; line.y <= yEnd; line.y += SOFTWARE_DRIVER_2_STEP_Y)
+		{
+			line.x[scan.left] = scan.x[0];
+			line.x[scan.right] = scan.x[1];
+
+#ifdef IPOL_Z
+			line.z[scan.left] = scan.z[0];
+			line.z[scan.right] = scan.z[1];
+#endif
+
+#ifdef IPOL_W
+			line.w[scan.left] = scan.w[0];
+			line.w[scan.right] = scan.w[1];
+#endif
+
+#ifdef IPOL_C0
+			line.c[0][scan.left] = scan.c[0][0];
+			line.c[0][scan.right] = scan.c[0][1];
+#endif
+
+#ifdef IPOL_C1
+			line.c[1][scan.left] = scan.c[1][0];
+			line.c[1][scan.right] = scan.c[1][1];
+#endif
+
+#ifdef IPOL_C2
+			line.c[2][scan.left] = scan.c[2][0];
+			line.c[2][scan.right] = scan.c[2][1];
+#endif
+
+#ifdef IPOL_C3
+			line.c[3][scan.left] = scan.c[3][0];
+			line.c[3][scan.right] = scan.c[3][1];
+#endif
+
+#ifdef IPOL_T0
+			line.t[0][scan.left] = scan.t[0][0];
+			line.t[0][scan.right] = scan.t[0][1];
+#endif
+
+#ifdef IPOL_T1
+			line.t[1][scan.left] = scan.t[1][0];
+			line.t[1][scan.right] = scan.t[1][1];
+#endif
+
+#ifdef IPOL_T2
+			line.t[2][scan.left] = scan.t[2][0];
+			line.t[2][scan.right] = scan.t[2][1];
+#endif
+
+#ifdef IPOL_L0
+			line.l[0][scan.left] = scan.l[0][0];
+			line.l[0][scan.right] = scan.l[0][1];
+#endif
+
+			// render a scanline
+			if_interlace_scanline fragmentShader();
+
+			scan.x[0] += scan.slopeX[0];
+			scan.x[1] += scan.slopeX[1];
+
+#ifdef IPOL_Z
+			scan.z[0] += scan.slopeZ[0];
+			scan.z[1] += scan.slopeZ[1];
+#endif
+
+#ifdef IPOL_W
+			scan.w[0] += scan.slopeW[0];
+			scan.w[1] += scan.slopeW[1];
+#endif
+
+#ifdef IPOL_C0
+			scan.c[0][0] += scan.slopeC[0][0];
+			scan.c[0][1] += scan.slopeC[0][1];
+#endif
+
+#ifdef IPOL_C1
+			scan.c[1][0] += scan.slopeC[1][0];
+			scan.c[1][1] += scan.slopeC[1][1];
+#endif
+
+#ifdef IPOL_C2
+			scan.c[2][0] += scan.slopeC[2][0];
+			scan.c[2][1] += scan.slopeC[2][1];
+#endif
+
+#ifdef IPOL_C3
+			scan.c[3][0] += scan.slopeC[3][0];
+			scan.c[3][1] += scan.slopeC[3][1];
+#endif
+
+#ifdef IPOL_T0
+			scan.t[0][0] += scan.slopeT[0][0];
+			scan.t[0][1] += scan.slopeT[0][1];
+#endif
+
+#ifdef IPOL_T1
+			scan.t[1][0] += scan.slopeT[1][0];
+			scan.t[1][1] += scan.slopeT[1][1];
+#endif
+
+#ifdef IPOL_T2
+			scan.t[2][0] += scan.slopeT[2][0];
+			scan.t[2][1] += scan.slopeT[2][1];
+#endif
+
+#ifdef IPOL_L0
+			scan.l[0][0] += scan.slopeL[0][0];
+			scan.l[0][1] += scan.slopeL[0][1];
+#endif
+
+		}
+	}
+
+	// rasterize lower sub-triangle
+	if (F32_GREATER_0(scan.invDeltaY[2]))
+	{
+		// advance to middle point
+		if (F32_GREATER_0(scan.invDeltaY[1]))
+		{
+			temp[0] = b->Pos.y - a->Pos.y;	// dy
+
+			scan.x[0] = a->Pos.x + scan.slopeX[0] * temp[0];
+#ifdef IPOL_Z
+			scan.z[0] = a->Pos.z + scan.slopeZ[0] * temp[0];
+#endif
+#ifdef IPOL_W
+			scan.w[0] = a->Pos.w + scan.slopeW[0] * temp[0];
+#endif
+#ifdef IPOL_C0
+			scan.c[0][0] = a->Color[0] + scan.slopeC[0][0] * temp[0];
+#endif
+#ifdef IPOL_C1
+			scan.c[1][0] = a->Color[1] + scan.slopeC[1][0] * temp[0];
+#endif
+#ifdef IPOL_C2
+			scan.c[2][0] = a->Color[2] + scan.slopeC[2][0] * temp[0];
+#endif
+#ifdef IPOL_C3
+			scan.c[3][0] = a->Color[3] + scan.slopeC[3][0] * temp[0];
+#endif
+#ifdef IPOL_T0
+			scan.t[0][0] = a->Tex[0] + scan.slopeT[0][0] * temp[0];
+#endif
+#ifdef IPOL_T1
+			scan.t[1][0] = a->Tex[1] + scan.slopeT[1][0] * temp[0];
+#endif
+#ifdef IPOL_T2
+			scan.t[2][0] = a->Tex[2] + scan.slopeT[2][0] * temp[0];
+#endif
+#ifdef IPOL_L0
+			scan.l[0][0] = sVec3Pack_unpack(a->LightTangent[0]) + scan.slopeL[0][0] * temp[0];
+#endif
+
+		}
+
+		// calculate slopes for bottom edge
+		scan.slopeX[1] = (c->Pos.x - b->Pos.x) * scan.invDeltaY[2];
+		scan.x[1] = b->Pos.x;
+
+#ifdef IPOL_Z
+		scan.slopeZ[1] = (c->Pos.z - b->Pos.z) * scan.invDeltaY[2];
+		scan.z[1] = b->Pos.z;
+#endif
+
+#ifdef IPOL_W
+		scan.slopeW[1] = (c->Pos.w - b->Pos.w) * scan.invDeltaY[2];
+		scan.w[1] = b->Pos.w;
+#endif
+
+#ifdef IPOL_C0
+		scan.slopeC[0][1] = (c->Color[0] - b->Color[0]) * scan.invDeltaY[2];
+		scan.c[0][1] = b->Color[0];
+#endif
+
+#ifdef IPOL_C1
+		scan.slopeC[1][1] = (c->Color[1] - b->Color[1]) * scan.invDeltaY[2];
+		scan.c[1][1] = b->Color[1];
+#endif
+
+#ifdef IPOL_C2
+		scan.slopeC[2][1] = (c->Color[2] - b->Color[2]) * scan.invDeltaY[2];
+		scan.c[2][1] = b->Color[2];
+#endif
+
+#ifdef IPOL_C3
+		scan.slopeC[3][1] = (c->Color[3] - b->Color[3]) * scan.invDeltaY[2];
+		scan.c[3][1] = b->Color[3];
+#endif
+
+#ifdef IPOL_T0
+		scan.slopeT[0][1] = (c->Tex[0] - b->Tex[0]) * scan.invDeltaY[2];
+		scan.t[0][1] = b->Tex[0];
+#endif
+
+#ifdef IPOL_T1
+		scan.slopeT[1][1] = (c->Tex[1] - b->Tex[1]) * scan.invDeltaY[2];
+		scan.t[1][1] = b->Tex[1];
+#endif
+
+#ifdef IPOL_T2
+		scan.slopeT[2][1] = (c->Tex[2] - b->Tex[2]) * scan.invDeltaY[2];
+		scan.t[2][1] = b->Tex[2];
+#endif
+
+#ifdef IPOL_L0
+		scan.slopeL[0][1] = (c->LightTangent[0] - b->LightTangent[0]) * scan.invDeltaY[2];
+		scan.l[0][1] = b->LightTangent[0];
+#endif
+
+		// apply top-left fill convention, top part
+		yStart = fill_convention_top(b->Pos.y);
+		yEnd = fill_convention_down(c->Pos.y);
+
+#ifdef SUBTEXEL
+		subPixel = ((f32)yStart) - b->Pos.y;
+
+		// correct to pixel center
+		scan.x[0] += scan.slopeX[0] * subPixel;
+		scan.x[1] += scan.slopeX[1] * subPixel;
+
+#ifdef IPOL_Z
+		scan.z[0] += scan.slopeZ[0] * subPixel;
+		scan.z[1] += scan.slopeZ[1] * subPixel;
+#endif
+
+#ifdef IPOL_W
+		scan.w[0] += scan.slopeW[0] * subPixel;
+		scan.w[1] += scan.slopeW[1] * subPixel;
+#endif
+
+#ifdef IPOL_C0
+		scan.c[0][0] += scan.slopeC[0][0] * subPixel;
+		scan.c[0][1] += scan.slopeC[0][1] * subPixel;
+#endif
+
+#ifdef IPOL_C1
+		scan.c[1][0] += scan.slopeC[1][0] * subPixel;
+		scan.c[1][1] += scan.slopeC[1][1] * subPixel;
+#endif
+
+#ifdef IPOL_C2
+		scan.c[2][0] += scan.slopeC[2][0] * subPixel;
+		scan.c[2][1] += scan.slopeC[2][1] * subPixel;
+#endif
+
+#ifdef IPOL_C3
+		scan.c[3][0] += scan.slopeC[3][0] * subPixel;
+		scan.c[3][1] += scan.slopeC[3][1] * subPixel;
+#endif
+
+#ifdef IPOL_T0
+		scan.t[0][0] += scan.slopeT[0][0] * subPixel;
+		scan.t[0][1] += scan.slopeT[0][1] * subPixel;
+#endif
+
+#ifdef IPOL_T1
+		scan.t[1][0] += scan.slopeT[1][0] * subPixel;
+		scan.t[1][1] += scan.slopeT[1][1] * subPixel;
+#endif
+
+#ifdef IPOL_T2
+		scan.t[2][0] += scan.slopeT[2][0] * subPixel;
+		scan.t[2][1] += scan.slopeT[2][1] * subPixel;
+#endif
+
+#ifdef IPOL_L0
+		scan.l[0][0] += scan.slopeL[0][0] * subPixel;
+		scan.l[0][1] += scan.slopeL[0][1] * subPixel;
+#endif
+
+#endif
+
+		// rasterize the edge scanlines
+		for (line.y = yStart; line.y <= yEnd; line.y += SOFTWARE_DRIVER_2_STEP_Y)
+		{
+			line.x[scan.left] = scan.x[0];
+			line.x[scan.right] = scan.x[1];
+
+#ifdef IPOL_Z
+			line.z[scan.left] = scan.z[0];
+			line.z[scan.right] = scan.z[1];
+#endif
+
+#ifdef IPOL_W
+			line.w[scan.left] = scan.w[0];
+			line.w[scan.right] = scan.w[1];
+#endif
+
+#ifdef IPOL_C0
+			line.c[0][scan.left] = scan.c[0][0];
+			line.c[0][scan.right] = scan.c[0][1];
+#endif
+
+#ifdef IPOL_C1
+			line.c[1][scan.left] = scan.c[1][0];
+			line.c[1][scan.right] = scan.c[1][1];
+#endif
+
+#ifdef IPOL_C2
+			line.c[2][scan.left] = scan.c[2][0];
+			line.c[2][scan.right] = scan.c[2][1];
+#endif
+
+#ifdef IPOL_C3
+			line.c[3][scan.left] = scan.c[3][0];
+			line.c[3][scan.right] = scan.c[3][1];
+#endif
+
+#ifdef IPOL_T0
+			line.t[0][scan.left] = scan.t[0][0];
+			line.t[0][scan.right] = scan.t[0][1];
+#endif
+
+#ifdef IPOL_T1
+			line.t[1][scan.left] = scan.t[1][0];
+			line.t[1][scan.right] = scan.t[1][1];
+#endif
+
+#ifdef IPOL_T2
+			line.t[2][scan.left] = scan.t[2][0];
+			line.t[2][scan.right] = scan.t[2][1];
+#endif
+
+#ifdef IPOL_L0
+			line.l[0][scan.left] = scan.l[0][0];
+			line.l[0][scan.right] = scan.l[0][1];
+#endif
+
+			// render a scanline
+			if_interlace_scanline fragmentShader();
+
+			scan.x[0] += scan.slopeX[0];
+			scan.x[1] += scan.slopeX[1];
+
+#ifdef IPOL_Z
+			scan.z[0] += scan.slopeZ[0];
+			scan.z[1] += scan.slopeZ[1];
+#endif
+
+#ifdef IPOL_W
+			scan.w[0] += scan.slopeW[0];
+			scan.w[1] += scan.slopeW[1];
+#endif
+
+#ifdef IPOL_C0
+			scan.c[0][0] += scan.slopeC[0][0];
+			scan.c[0][1] += scan.slopeC[0][1];
+#endif
+
+#ifdef IPOL_C1
+			scan.c[1][0] += scan.slopeC[1][0];
+			scan.c[1][1] += scan.slopeC[1][1];
+#endif
+
+#ifdef IPOL_C2
+			scan.c[2][0] += scan.slopeC[2][0];
+			scan.c[2][1] += scan.slopeC[2][1];
+#endif
+
+#ifdef IPOL_C3
+			scan.c[3][0] += scan.slopeC[3][0];
+			scan.c[3][1] += scan.slopeC[3][1];
+#endif
+
+#ifdef IPOL_T0
+			scan.t[0][0] += scan.slopeT[0][0];
+			scan.t[0][1] += scan.slopeT[0][1];
+#endif
+
+#ifdef IPOL_T1
+			scan.t[1][0] += scan.slopeT[1][0];
+			scan.t[1][1] += scan.slopeT[1][1];
+#endif
+#ifdef IPOL_T2
+			scan.t[2][0] += scan.slopeT[2][0];
+			scan.t[2][1] += scan.slopeT[2][1];
+#endif
+
+#ifdef IPOL_L0
+			scan.l[0][0] += scan.slopeL[0][0];
+			scan.l[0][1] += scan.slopeL[0][1];
+#endif
+
+		}
+	}
+
+}
+
+
+
+//! Called by the engine when the vertex and/or pixel shader constants for an
+//! material renderer should be set.
+void CBurningParallaxMap::OnSetConstants(IMaterialRendererServices* services, s32 userData)
+{
+#if 0
+	video::IVideoDriver* driver = services->getVideoDriver();
+
+	// set transposed world matrix
+	const core::matrix4& tWorld = driver->getTransform(video::ETS_WORLD).getTransposed();
+	services->setVertexShaderConstant(tWorld.pointer(), 0, 4);
+
+	// set transposed worldViewProj matrix
+	core::matrix4 worldViewProj(driver->getTransform(video::ETS_PROJECTION));
+	worldViewProj *= driver->getTransform(video::ETS_VIEW);
+	worldViewProj *= driver->getTransform(video::ETS_WORLD);
+	core::matrix4 tr(worldViewProj.getTransposed());
+	services->setVertexShaderConstant(tr.pointer(), 8, 4);
+
+	// here we fetch the fixed function lights from the driver
+	// and set them as constants
+
+	u32 cnt = driver->getDynamicLightCount();
+
+	// Load the inverse world matrix.
+	core::matrix4 invWorldMat;
+	driver->getTransform(video::ETS_WORLD).getInverse(invWorldMat);
+
+	for (u32 i = 0; i < 2; ++i)
+	{
+		video::SLight light;
+
+		if (i < cnt)
+			light = driver->getDynamicLight(i);
+		else
+		{
+			light.DiffuseColor.set(0, 0, 0); // make light dark
+			light.Radius = 1.0f;
+		}
+
+		light.DiffuseColor.a = 1.0f / (light.Radius * light.Radius); // set attenuation
+
+		// Transform the light by the inverse world matrix to get it into object space.
+		invWorldMat.transformVect(light.Position);
+
+		services->setVertexShaderConstant(
+			reinterpret_cast<const f32*>(&light.Position), 12 + (i * 2), 1);
+
+		services->setVertexShaderConstant(
+			reinterpret_cast<const f32*>(&light.DiffuseColor), 13 + (i * 2), 1);
+	}
+
+	// Obtain the view position by transforming 0,0,0 by the inverse view matrix
+	// and then multiply this by the inverse world matrix.
+	core::vector3df viewPos(0.0f, 0.0f, 0.0f);
+	core::matrix4 inverseView;
+	driver->getTransform(video::ETS_VIEW).getInverse(inverseView);
+	inverseView.transformVect(viewPos);
+	invWorldMat.transformVect(viewPos);
+	services->setVertexShaderConstant(reinterpret_cast<const f32*>(&viewPos.X), 16, 1);
+
+	// set scale factor
+	f32 factor = 0.02f; // default value
+	if (CurrentScale != 0.0f)
+		factor = CurrentScale;
+
+	f32 c6[] = { factor, factor, factor, factor };
+	services->setPixelShaderConstant(c6, 0, 1);
+#endif
+}
+
+burning_namespace_end
+
+#endif // _IRR_COMPILE_WITH_BURNINGSVIDEO_
+
+burning_namespace_start
+
+
+IBurningShader* createTRParallaxMap(CBurningVideoDriver* driver, s32& outMaterialTypeNr, E_MATERIAL_TYPE baseMaterial)
+{
+#ifdef _IRR_COMPILE_WITH_BURNINGSVIDEO_
+	return new CBurningParallaxMap(driver, outMaterialTypeNr, baseMaterial);
+#else
+	return 0;
+#endif // _IRR_COMPILE_WITH_BURNINGSVIDEO_
+}
+
+burning_namespace_end
diff --git a/source/Irrlicht/CTRStencilShadow.cpp b/source/Irrlicht/CTRStencilShadow.cpp
index 804ed128..b9edf1f6 100644
--- a/source/Irrlicht/CTRStencilShadow.cpp
+++ b/source/Irrlicht/CTRStencilShadow.cpp
@@ -93,7 +93,7 @@ private:
 
 //! constructor
 CTRStencilShadow::CTRStencilShadow(CBurningVideoDriver* driver)
-: IBurningShader(driver)
+: IBurningShader(driver, EMT_SOLID)
 {
 	#ifdef _DEBUG
 	setDebugName("CTRStencilShadow");
diff --git a/source/Irrlicht/CTRTextureBlend.cpp b/source/Irrlicht/CTRTextureBlend.cpp
index 592ec1f1..25f0c562 100644
--- a/source/Irrlicht/CTRTextureBlend.cpp
+++ b/source/Irrlicht/CTRTextureBlend.cpp
@@ -108,7 +108,7 @@ private:
 
 //! constructor
 CTRTextureBlend::CTRTextureBlend(CBurningVideoDriver* driver)
-: IBurningShader(driver)
+: IBurningShader(driver, EMT_ONETEXTURE_BLEND)
 {
 	#ifdef _DEBUG
 	setDebugName("CTRTextureBlend");
diff --git a/source/Irrlicht/CTRTextureDetailMap2.cpp b/source/Irrlicht/CTRTextureDetailMap2.cpp
index 22da5094..51c107fb 100644
--- a/source/Irrlicht/CTRTextureDetailMap2.cpp
+++ b/source/Irrlicht/CTRTextureDetailMap2.cpp
@@ -94,7 +94,7 @@ protected:
 
 //! constructor
 CTRTextureDetailMap2::CTRTextureDetailMap2(CBurningVideoDriver* driver)
-: IBurningShader(driver)
+: IBurningShader(driver, EMT_DETAIL_MAP)
 {
 	#ifdef _DEBUG
 	setDebugName("CTRTextureDetailMap2");
diff --git a/source/Irrlicht/CTRTextureGouraud2.cpp b/source/Irrlicht/CTRTextureGouraud2.cpp
index 3b8761b9..49b2a08d 100644
--- a/source/Irrlicht/CTRTextureGouraud2.cpp
+++ b/source/Irrlicht/CTRTextureGouraud2.cpp
@@ -112,7 +112,7 @@ private:
 
 //! constructor
 CTRTextureGouraud2::CTRTextureGouraud2(CBurningVideoDriver* driver)
-: IBurningShader(driver)
+: IBurningShader(driver,EMT_SOLID)
 {
 	#ifdef _DEBUG
 	setDebugName("CTRTextureGouraud2");
diff --git a/source/Irrlicht/CTRTextureGouraudAdd2.cpp b/source/Irrlicht/CTRTextureGouraudAdd2.cpp
index 0cc2f17d..66006b4e 100644
--- a/source/Irrlicht/CTRTextureGouraudAdd2.cpp
+++ b/source/Irrlicht/CTRTextureGouraudAdd2.cpp
@@ -93,7 +93,7 @@ private:
 
 //! constructor
 CTRTextureGouraudAdd2::CTRTextureGouraudAdd2(CBurningVideoDriver* driver)
-: IBurningShader(driver)
+: IBurningShader(driver, EMT_TRANSPARENT_ADD_COLOR)
 {
 	#ifdef _DEBUG
 	setDebugName("CTRTextureGouraudAdd2");
diff --git a/source/Irrlicht/CTRTextureGouraudAddNoZ2.cpp b/source/Irrlicht/CTRTextureGouraudAddNoZ2.cpp
index 78008f77..a02b1f0c 100644
--- a/source/Irrlicht/CTRTextureGouraudAddNoZ2.cpp
+++ b/source/Irrlicht/CTRTextureGouraudAddNoZ2.cpp
@@ -93,7 +93,7 @@ private:
 
 //! constructor
 CTRTextureGouraudAddNoZ2::CTRTextureGouraudAddNoZ2(CBurningVideoDriver* driver)
-: IBurningShader(driver)
+: IBurningShader(driver, EMT_TRANSPARENT_ADD_COLOR)
 {
 	#ifdef _DEBUG
 	setDebugName("CTRTextureGouraudAddNoZ2");
diff --git a/source/Irrlicht/CTRTextureGouraudAlpha.cpp b/source/Irrlicht/CTRTextureGouraudAlpha.cpp
index 41e40e35..bfb4cd53 100644
--- a/source/Irrlicht/CTRTextureGouraudAlpha.cpp
+++ b/source/Irrlicht/CTRTextureGouraudAlpha.cpp
@@ -100,7 +100,7 @@ private:
 
 //! constructor
 CTRTextureGouraudAlpha2::CTRTextureGouraudAlpha2(CBurningVideoDriver* driver)
-: IBurningShader(driver)
+: IBurningShader(driver, EMT_TRANSPARENT_ALPHA_CHANNEL)
 {
 	#ifdef _DEBUG
 	setDebugName("CTRTextureGouraudAlpha2");
diff --git a/source/Irrlicht/CTRTextureGouraudAlphaNoZ.cpp b/source/Irrlicht/CTRTextureGouraudAlphaNoZ.cpp
index 4c724de8..3b7e70ee 100644
--- a/source/Irrlicht/CTRTextureGouraudAlphaNoZ.cpp
+++ b/source/Irrlicht/CTRTextureGouraudAlphaNoZ.cpp
@@ -103,7 +103,7 @@ private:
 
 //! constructor
 CTRTextureGouraudAlphaNoZ::CTRTextureGouraudAlphaNoZ(CBurningVideoDriver* driver)
-: IBurningShader(driver)
+: IBurningShader(driver, EMT_TRANSPARENT_ALPHA_CHANNEL)
 {
 	#ifdef _DEBUG
 	setDebugName("CTRTextureGouraudAlphaNoZ");
@@ -125,7 +125,7 @@ void CTRTextureGouraudAlphaNoZ::OnSetMaterial(const SBurningShaderMaterial& mate
 
 	//check triangle on w = 1.f instead..
 #ifdef	SOFTWARE_DRIVER_2_BILINEAR
-	if (material.Fallback_MaterialType == EMT_TRANSPARENT_ALPHA_CHANNEL_REF)
+	if (material.org.MaterialType == EMT_TRANSPARENT_ALPHA_CHANNEL_REF)
 		fragmentShader = &CTRTextureGouraudAlphaNoZ::fragment_linear_alpharef;
 	else
 	if ( material.org.TextureLayer[0].BilinearFilter )
diff --git a/source/Irrlicht/CTRTextureGouraudNoZ2.cpp b/source/Irrlicht/CTRTextureGouraudNoZ2.cpp
index 85d96e65..9da8ee71 100644
--- a/source/Irrlicht/CTRTextureGouraudNoZ2.cpp
+++ b/source/Irrlicht/CTRTextureGouraudNoZ2.cpp
@@ -100,7 +100,7 @@ private:
 
 //! constructor
 CTRTextureGouraudNoZ2::CTRTextureGouraudNoZ2(CBurningVideoDriver* driver)
-	: IBurningShader(driver)
+	: IBurningShader(driver,EMT_SOLID)
 {
 #ifdef _DEBUG
 	setDebugName("CTRTextureGouraudNoZ2");
diff --git a/source/Irrlicht/CTRTextureGouraudVertexAlpha2.cpp b/source/Irrlicht/CTRTextureGouraudVertexAlpha2.cpp
index e90e752f..c44d3236 100644
--- a/source/Irrlicht/CTRTextureGouraudVertexAlpha2.cpp
+++ b/source/Irrlicht/CTRTextureGouraudVertexAlpha2.cpp
@@ -99,7 +99,7 @@ private:
 
 //! constructor
 CTRTextureVertexAlpha2::CTRTextureVertexAlpha2(CBurningVideoDriver* driver)
-: IBurningShader(driver)
+: IBurningShader(driver, EMT_TRANSPARENT_VERTEX_ALPHA)
 {
 	#ifdef _DEBUG
 	setDebugName("CTRTextureVertexAlpha2");
diff --git a/source/Irrlicht/CTRTextureLightMap2_Add.cpp b/source/Irrlicht/CTRTextureLightMap2_Add.cpp
index 5f53a9b9..0c874e75 100644
--- a/source/Irrlicht/CTRTextureLightMap2_Add.cpp
+++ b/source/Irrlicht/CTRTextureLightMap2_Add.cpp
@@ -92,7 +92,7 @@ private:
 
 //! constructor
 CTRTextureLightMap2_Add::CTRTextureLightMap2_Add(CBurningVideoDriver* driver)
-: IBurningShader(driver)
+: IBurningShader(driver, EMT_LIGHTMAP_ADD)
 {
 	#ifdef _DEBUG
 	setDebugName("CTRTextureLightMap2_Add");
diff --git a/source/Irrlicht/CTRTextureLightMap2_M1.cpp b/source/Irrlicht/CTRTextureLightMap2_M1.cpp
index aea774ec..382b5f9f 100644
--- a/source/Irrlicht/CTRTextureLightMap2_M1.cpp
+++ b/source/Irrlicht/CTRTextureLightMap2_M1.cpp
@@ -92,7 +92,7 @@ private:
 
 //! constructor
 CTRTextureLightMap2_M1::CTRTextureLightMap2_M1(CBurningVideoDriver* driver)
-: IBurningShader(driver)
+: IBurningShader(driver, EMT_LIGHTMAP)
 {
 	#ifdef _DEBUG
 	setDebugName("CTRTextureLightMap2_M1");
diff --git a/source/Irrlicht/CTRTextureLightMap2_M2.cpp b/source/Irrlicht/CTRTextureLightMap2_M2.cpp
index 56d0ccd5..433e78fc 100644
--- a/source/Irrlicht/CTRTextureLightMap2_M2.cpp
+++ b/source/Irrlicht/CTRTextureLightMap2_M2.cpp
@@ -92,7 +92,7 @@ private:
 
 //! constructor
 CTRTextureLightMap2_M2::CTRTextureLightMap2_M2(CBurningVideoDriver* driver)
-: IBurningShader(driver)
+: IBurningShader(driver, EMT_LIGHTMAP_LIGHTING_M2)
 {
 	#ifdef _DEBUG
 	setDebugName("CTRTextureLightMap2_M2");
diff --git a/source/Irrlicht/CTRTextureLightMap2_M4.cpp b/source/Irrlicht/CTRTextureLightMap2_M4.cpp
index 0e0620e9..f6af7cd1 100644
--- a/source/Irrlicht/CTRTextureLightMap2_M4.cpp
+++ b/source/Irrlicht/CTRTextureLightMap2_M4.cpp
@@ -93,7 +93,7 @@ private:
 
 //! constructor
 CTRTextureLightMap2_M4::CTRTextureLightMap2_M4(CBurningVideoDriver* driver)
-: IBurningShader(driver)
+: IBurningShader(driver, EMT_LIGHTMAP_M4)
 {
 	#ifdef _DEBUG
 	setDebugName("CTRTextureLightMap2_M4");
diff --git a/source/Irrlicht/CTRTextureLightMapGouraud2_M4.cpp b/source/Irrlicht/CTRTextureLightMapGouraud2_M4.cpp
index cc634e02..ba2010ad 100644
--- a/source/Irrlicht/CTRTextureLightMapGouraud2_M4.cpp
+++ b/source/Irrlicht/CTRTextureLightMapGouraud2_M4.cpp
@@ -92,7 +92,7 @@ private:
 
 //! constructor
 CTRGTextureLightMap2_M4::CTRGTextureLightMap2_M4(CBurningVideoDriver* driver)
-: IBurningShader(driver)
+: IBurningShader(driver, EMT_LIGHTMAP_LIGHTING_M4)
 {
 	#ifdef _DEBUG
 	setDebugName("CTRGTextureLightMap2_M4");
diff --git a/source/Irrlicht/CTRTextureWire2.cpp b/source/Irrlicht/CTRTextureWire2.cpp
index a33454d0..1b66e98c 100644
--- a/source/Irrlicht/CTRTextureWire2.cpp
+++ b/source/Irrlicht/CTRTextureWire2.cpp
@@ -98,7 +98,7 @@ protected:
 
 //! constructor
 CTRTextureWire2::CTRTextureWire2(CBurningVideoDriver* driver)
-	: IBurningShader(driver)
+	: IBurningShader(driver, EMT_SOLID)
 {
 #ifdef _DEBUG
 	setDebugName("CTRTextureWire2");
diff --git a/source/Irrlicht/CTR_transparent_reflection_2_layer.cpp b/source/Irrlicht/CTR_transparent_reflection_2_layer.cpp
index 4e209a2b..22856ff0 100644
--- a/source/Irrlicht/CTR_transparent_reflection_2_layer.cpp
+++ b/source/Irrlicht/CTR_transparent_reflection_2_layer.cpp
@@ -98,7 +98,7 @@ private:
 
 //! constructor
 CTR_transparent_reflection_2_layer::CTR_transparent_reflection_2_layer(CBurningVideoDriver* driver)
-	: IBurningShader(driver)
+	: IBurningShader(driver, EMT_REFLECTION_2_LAYER)
 {
 #ifdef _DEBUG
 	setDebugName("CTR_transparent_reflection_2_layer");
diff --git a/source/Irrlicht/IBurningShader.cpp b/source/Irrlicht/IBurningShader.cpp
index b0e5eaca..56d3359e 100644
--- a/source/Irrlicht/IBurningShader.cpp
+++ b/source/Irrlicht/IBurningShader.cpp
@@ -20,7 +20,7 @@ const tFixPointu IBurningShader::dithermask[] =
 	0xf0,0x70,0xd0,0x50
 };
 
-void IBurningShader::constructor_IBurningShader(CBurningVideoDriver* driver)
+void IBurningShader::constructor_IBurningShader(CBurningVideoDriver* driver, E_MATERIAL_TYPE baseMaterial)
 {
 #ifdef _DEBUG
 	setDebugName("IBurningShader");
@@ -62,16 +62,37 @@ void IBurningShader::constructor_IBurningShader(CBurningVideoDriver* driver)
 	stencilOp[1] = StencilOp_KEEP;
 	stencilOp[2] = StencilOp_KEEP;
 	AlphaRef = 0;
-	RenderPass_ShaderIsTransparent = 0;
 	PrimitiveColor = COLOR_BRIGHT_WHITE;
 	TL_Flag = 0;
 	fragment_draw_count = 0;
 	VertexShaderProgram_buildin = BVT_Fix;
+
+	//set default Transparent/Solid
+	BaseMaterial = baseMaterial;
+	switch (BaseMaterial)
+	{
+	case EMT_TRANSPARENT_ADD_COLOR:
+	case EMT_TRANSPARENT_ALPHA_CHANNEL:
+	case EMT_TRANSPARENT_ALPHA_CHANNEL_REF:
+	case EMT_TRANSPARENT_VERTEX_ALPHA:
+	case EMT_TRANSPARENT_REFLECTION_2_LAYER:
+	case EMT_NORMAL_MAP_TRANSPARENT_ADD_COLOR:
+	case EMT_NORMAL_MAP_TRANSPARENT_VERTEX_ALPHA:
+	case EMT_PARALLAX_MAP_TRANSPARENT_ADD_COLOR:
+	case EMT_PARALLAX_MAP_TRANSPARENT_VERTEX_ALPHA:
+	case EMT_ONETEXTURE_BLEND:
+		RenderPass_ShaderIsTransparent = 1;
+		break;
+	default:
+		RenderPass_ShaderIsTransparent = 0;
+		break;
+	}
+
 }
 
-IBurningShader::IBurningShader(CBurningVideoDriver* driver)
+IBurningShader::IBurningShader(CBurningVideoDriver* driver,E_MATERIAL_TYPE baseMaterial)
 {
-	constructor_IBurningShader(driver);
+	constructor_IBurningShader(driver, baseMaterial);
 }
 
 //! Constructor
@@ -94,32 +115,12 @@ IBurningShader::IBurningShader(
 	E_MATERIAL_TYPE baseMaterial,
 	s32 userData)
 {
-	constructor_IBurningShader(driver);
-	BaseMaterial = baseMaterial;
+	constructor_IBurningShader(driver, baseMaterial);
 	UserData = userData;
 	CallBack = callback;
 	if (CallBack)
 		CallBack->grab();
 
-	//set default Transparent/Solid
-	switch (BaseMaterial)
-	{
-	case EMT_TRANSPARENT_ADD_COLOR:
-	case EMT_TRANSPARENT_ALPHA_CHANNEL:
-	case EMT_TRANSPARENT_ALPHA_CHANNEL_REF:
-	case EMT_TRANSPARENT_VERTEX_ALPHA:
-	case EMT_TRANSPARENT_REFLECTION_2_LAYER:
-	case EMT_NORMAL_MAP_TRANSPARENT_ADD_COLOR:
-	case EMT_NORMAL_MAP_TRANSPARENT_VERTEX_ALPHA:
-	case EMT_PARALLAX_MAP_TRANSPARENT_ADD_COLOR:
-	case EMT_PARALLAX_MAP_TRANSPARENT_VERTEX_ALPHA:
-	case EMT_ONETEXTURE_BLEND:
-		RenderPass_ShaderIsTransparent = 1;
-		break;
-	default:
-		RenderPass_ShaderIsTransparent = 0;
-		break;
-	}
 
 	//v0.53 compile. only buildin
 	const c8* ip = vertexShaderProgram;
@@ -140,6 +141,7 @@ IBurningShader::IBurningShader(
 	else if (len == 958 && hash == 0xa048973b) VertexShaderProgram_buildin = STK_958_0xa048973b;	/* supertuxkart motion_blur.vert */
 	else if (len == 1309 && hash == 0x1fd689c2) VertexShaderProgram_buildin = STK_1309_0x1fd689c2;	/* supertuxkart normalmap.vert */
 	else if (len == 1204 && hash == 0x072a4094) VertexShaderProgram_buildin = STK_1204_0x072a4094;	/* supertuxkart splatting.vert */
+	else if (len == 1303 && hash == 0xd872cdb6) VertexShaderProgram_buildin = STK_1303_0xd872cdb6;	/* supertuxkart water.vert */
 
 
 	//VertexShaderProgram = vertexShaderProgram;
@@ -154,7 +156,8 @@ IBurningShader::IBurningShader(
 	FILE* f = fopen("shader_id.txt", run ? "a" : "wb");
 	if (f)
 	{
-		fprintf(f, "--- start outMaterialTypeNr:%d len:%d hash: 0x%08x\n", outMaterialTypeNr, len, hash);
+		fprintf(f, "--- start outMaterialTypeNr:%d len:%d hash: 0x%08x buildIn:%d\n"
+			, outMaterialTypeNr, len, hash, VertexShaderProgram_buildin);
 		fprintf(f, "%s", vertexShaderProgram);
 		fprintf(f, "\n-------------- end ---------------------------\n");
 		fclose(f);
@@ -335,43 +338,6 @@ void IBurningShader::setBasicRenderStates(const SMaterial& material, const SMate
 	Driver->setBasicRenderStates(material, lastMaterial, resetAllRenderstates);
 }
 
-#if 0
-const core::matrix4& IBurningShader::uniform_mat4(const c8* name)
-{
-	return (const core::matrix4&)*getUniform(name, BL_VERTEX_FLOAT);
-}
-
-video::sVec4 IBurningShader::uniform_vec4(const c8* name)
-{
-	const f32* v = getUniform(name, BL_VERTEX_FLOAT);
-	return video::sVec4(v[0], v[1], v[2], v[3]);
-}
-
-video::sVec4 IBurningShader::uniform_vec3(const c8* name)
-{
-	const f32* v = getUniform(name, BL_VERTEX_FLOAT);
-	return video::sVec4(v[0], v[1], v[2], 0.f);
-}
-
-
-core::matrix4& IBurningShader::varying_mat4(const c8* name)
-{
-	return (core::matrix4&)*getUniform(name, BL_FRAGMENT_FLOAT);
-}
-
-video::sVec4 IBurningShader::varying_vec4(const c8* name)
-{
-	const f32* v = getUniform(name, BL_FRAGMENT_FLOAT);
-	return video::sVec4(v[0], v[1], v[2], v[3]);
-}
-
-video::sVec4 IBurningShader::varying_vec3(const c8* name)
-{
-	const f32* v = getUniform(name, BL_FRAGMENT_FLOAT);
-	return video::sVec4(v[0], v[1], v[2], 0.f);
-}
-#endif
-
 static BurningUniform _empty = { "null",BL_VERTEX_FLOAT,{0.f,0.f,0.f,0.f} };
 const f32* IBurningShader::getUniform(const c8* name, EBurningUniformFlags flags) const
 {
@@ -430,15 +396,6 @@ bool IBurningShader::setShaderConstantID(EBurningUniformFlags flags, s32 index,
 {
 	if ((u32)index >= UniformInfo.size())
 		return false;
-#if 0
-	BurningUniform add;
-	while ((u32)index >= UniformInfo.size())
-	{
-		tiny_strcpy(add.name, tiny_itoa(UniformInfo.size(), 10));
-		add.type = flags;
-		UniformInfo.push_back(add);
-	}
-#endif
 
 	BurningUniform& use = UniformInfo[index];
 	use.type = flags;
@@ -472,7 +429,7 @@ void IBurningShader::setVertexShaderConstant(const f32* data, s32 startRegister,
 	c8 name[BL_ACTIVE_UNIFORM_MAX_LENGTH];
 	tiny_strcpy(name, tiny_itoa(startRegister, 10));
 
-	setShaderConstantID(BL_VERTEX_FLOAT, getShaderConstantID(BL_VERTEX_PROGRAM, name), data, constantAmount);
+	setShaderConstantID(BL_VERTEX_FLOAT, getShaderConstantID(BL_VERTEX_PROGRAM, name), data, constantAmount*4);
 }
 
 void IBurningShader::setPixelShaderConstant(const f32* data, s32 startRegister, s32 constantAmount)
@@ -480,7 +437,7 @@ void IBurningShader::setPixelShaderConstant(const f32* data, s32 startRegister,
 	c8 name[BL_ACTIVE_UNIFORM_MAX_LENGTH];
 	tiny_strcpy(name, tiny_itoa(startRegister, 10));
 
-	setShaderConstantID(BL_FRAGMENT_FLOAT, getShaderConstantID(BL_FRAGMENT_PROGRAM, name), data, constantAmount);
+	setShaderConstantID(BL_FRAGMENT_FLOAT, getShaderConstantID(BL_FRAGMENT_PROGRAM, name), data, constantAmount*4);
 }
 
 bool IBurningShader::setVertexShaderConstant(s32 index, const f32* floats, int count)
diff --git a/source/Irrlicht/IBurningShader.h b/source/Irrlicht/IBurningShader.h
index a3f5ecd7..11e3eeb3 100644
--- a/source/Irrlicht/IBurningShader.h
+++ b/source/Irrlicht/IBurningShader.h
@@ -20,6 +20,7 @@
 #include "IMaterialRenderer.h"
 #include "IMaterialRendererServices.h"
 #include "IGPUProgrammingServices.h"
+#include "IShaderConstantSetCallback.h"
 
 burning_namespace_start
 
@@ -47,6 +48,11 @@ struct SBurningShaderLight
 	sVec3Color AmbientColor;
 	sVec3Color DiffuseColor;
 	sVec3Color SpecularColor;
+
+	//normal,parallax
+	sVec4 pos_local;	//modelinverse
+	f32 nmap_linearAttenuation;
+
 	SBurningShaderLight()
 	{
 		LightIsOn = false;
@@ -105,6 +111,10 @@ struct SBurningShaderEyeSpace
 	f32 fog_scale; // 1 / (fog.end-fog.start)
 
 	size_t TL_Flag; // eTransformLightFlags
+
+	// objectspace
+	core::matrix4 mvi;	// inverse Model*View
+	sVec4 leye;	//eye vector unprojected
 };
 
 enum eBurningCullFlag
@@ -137,6 +147,7 @@ enum eBurningVertexShader
 	STK_958_0xa048973b,		/* supertuxkart motion_blur.vert */
 	STK_1204_0x072a4094,	/* supertuxkart splatting.vert */
 	STK_1309_0x1fd689c2,	/* supertuxkart normalmap.vert */
+	STK_1303_0xd872cdb6,	/* supertuxkart water.vert */
 };
 
 struct SBurningShaderMaterial
@@ -195,6 +206,7 @@ enum EBurningFFShader
 	ETR_TEXTURE_GOURAUD_ALPHA_NOZ_NOPERSPECTIVE_CORRECT,
 
 	ETR_NORMAL_MAP_SOLID,
+	ETR_PARALLAX_MAP_SOLID,
 	ETR_STENCIL_SHADOW,
 
 	ETR_TEXTURE_BLEND,
@@ -235,7 +247,7 @@ struct BurningUniform
 
 	bool operator==(const BurningUniform& other) const
 	{
-		return tiny_istoken(name, other.name);
+		return ((type & 3) == (other.type & 3)) && tiny_istoken(name, other.name);
 	}
 
 };
@@ -250,11 +262,11 @@ struct PushShaderData
 };
 
 class CBurningVideoDriver;
-class IBurningShader : public IMaterialRenderer, public IMaterialRendererServices
+class IBurningShader : public IMaterialRenderer, public IMaterialRendererServices, public IShaderConstantSetCallBack
 {
 public:
 	//! Constructor
-	IBurningShader(CBurningVideoDriver* driver);
+	IBurningShader(CBurningVideoDriver* driver, E_MATERIAL_TYPE baseMaterial );
 
 	//! Constructor
 	IBurningShader(
@@ -303,8 +315,10 @@ public:
 
 	void setStencilOp(eBurningStencilOp sfail, eBurningStencilOp dpfail, eBurningStencilOp dppass);
 
-	//IMaterialRenderer
+	//IShaderConstantSetCallBack
+	virtual void OnSetConstants(IMaterialRendererServices* services, s32 userData) IRR_OVERRIDE {};
 
+	//IMaterialRenderer
 	virtual void OnSetMaterial(const SMaterial& material, const SMaterial& lastMaterial,
 		bool resetAllRenderstates, IMaterialRendererServices* services) IRR_OVERRIDE;
 
@@ -380,21 +394,13 @@ public:
 	}
 
 	u32 fragment_draw_count;
-/*
-	const core::matrix4& uniform_mat4(const c8* name);
-	video::sVec4 uniform_vec4(const c8* name);
-	video::sVec4 uniform_vec3(const c8* name);
 
-	core::matrix4& varying_mat4(const c8* name);
-	video::sVec4 varying_vec4(const c8* name);
-	video::sVec4 varying_vec3(const c8* name);
-*/
 	const f32* getUniform(const c8* name, EBurningUniformFlags flags) const;
 
 protected:
 	//friend class CBurningVideoDriver;
 
-	void constructor_IBurningShader(CBurningVideoDriver* driver);
+	void constructor_IBurningShader(CBurningVideoDriver* driver, E_MATERIAL_TYPE baseMaterial);
 
 	CBurningVideoDriver* Driver;
 	IShaderConstantSetCallBack* CallBack;
@@ -462,7 +468,6 @@ IBurningShader* createTriangleRendererTextureVertexAlpha2(CBurningVideoDriver* d
 IBurningShader* createTriangleRendererTextureGouraudWire2(CBurningVideoDriver* driver);
 IBurningShader* createTriangleRendererGouraud2(CBurningVideoDriver* driver);
 IBurningShader* createTriangleRendererGouraudNoZ2(CBurningVideoDriver* driver);
-IBurningShader* createTriangleRendererGouraudAlpha2(CBurningVideoDriver* driver);
 IBurningShader* createTRGouraudAlphaNoZ2(CBurningVideoDriver* driver);
 IBurningShader* createTriangleRendererGouraudWire2(CBurningVideoDriver* driver);
 IBurningShader* createTriangleRendererTextureFlat2(CBurningVideoDriver* driver);
@@ -478,7 +483,8 @@ IBurningShader* createTRTextureGouraudAlphaNoZ(CBurningVideoDriver* driver);
 IBurningShader* createTRTextureBlend(CBurningVideoDriver* driver);
 IBurningShader* createTRTextureInverseAlphaBlend(CBurningVideoDriver* driver);
 
-IBurningShader* createTRNormalMap(CBurningVideoDriver* driver);
+IBurningShader* createTRNormalMap(CBurningVideoDriver* driver, s32& outMaterialTypeNr, E_MATERIAL_TYPE baseMaterial);
+IBurningShader* createTRParallaxMap(CBurningVideoDriver* driver, s32& outMaterialTypeNr, E_MATERIAL_TYPE baseMaterial);
 IBurningShader* createTRStencilShadow(CBurningVideoDriver* driver);
 
 IBurningShader* createTriangleRendererReference(CBurningVideoDriver* driver);
diff --git a/source/Irrlicht/Irrlicht.xcodeproj/project.pbxproj b/source/Irrlicht/Irrlicht.xcodeproj/project.pbxproj
index c6411697..98318c64 100644
--- a/source/Irrlicht/Irrlicht.xcodeproj/project.pbxproj
+++ b/source/Irrlicht/Irrlicht.xcodeproj/project.pbxproj
@@ -239,7 +239,6 @@
 		5E34CB931B7F6EC500F212E8 /* CSoftwareDriver2.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 5E34C9641B7F6A7600F212E8 /* CSoftwareDriver2.cpp */; };
 		5E34CB951B7F6EC500F212E8 /* CSoftwareTexture2.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 5E34C9661B7F6A7600F212E8 /* CSoftwareTexture2.cpp */; };
 		5E34CB971B7F6EC500F212E8 /* CTRGouraud2.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 5E34C9681B7F6A7600F212E8 /* CTRGouraud2.cpp */; };
-		5E34CB981B7F6EC500F212E8 /* CTRGouraudAlpha2.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 5E34C9691B7F6A7600F212E8 /* CTRGouraudAlpha2.cpp */; };
 		5E34CB991B7F6EC500F212E8 /* CTRGouraudAlphaNoZ2.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 5E34C96A1B7F6A7600F212E8 /* CTRGouraudAlphaNoZ2.cpp */; };
 		5E34CB9A1B7F6EC500F212E8 /* CTRNormalMap.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 5E34C96B1B7F6A7600F212E8 /* CTRNormalMap.cpp */; };
 		5E34CB9B1B7F6EC500F212E8 /* CTRStencilShadow.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 5E34C96C1B7F6A7600F212E8 /* CTRStencilShadow.cpp */; };
@@ -308,6 +307,7 @@
 		5E79089B1C10FEF900DFE7FE /* CB3DMeshWriter.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 5E7908981C10FE4A00DFE7FE /* CB3DMeshWriter.cpp */; };
 		5E8570BC1B7F9AC400B267D2 /* CIrrDeviceConsole.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 5E8570BA1B7F9AC400B267D2 /* CIrrDeviceConsole.cpp */; };
 		5E8570BD1B7F9AC400B267D2 /* CIrrDeviceConsole.h in Headers */ = {isa = PBXBuildFile; fileRef = 5E8570BB1B7F9AC400B267D2 /* CIrrDeviceConsole.h */; };
+		65E74BB3282193B50004CA1D /* CTRParallaxMap.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 65E74BB2282193B40004CA1D /* CTRParallaxMap.cpp */; };
 		8493BC02249DEBB30092813E /* IOctreeSceneNode.h in Headers */ = {isa = PBXBuildFile; fileRef = 8493BC01249DEBB20092813E /* IOctreeSceneNode.h */; };
 		8493BC04249DED3F0092813E /* IMemoryReadFile.h in Headers */ = {isa = PBXBuildFile; fileRef = 8493BC03249DED3F0092813E /* IMemoryReadFile.h */; };
 		84A219A5249F823900E189FC /* SOverrideMaterial.h in Headers */ = {isa = PBXBuildFile; fileRef = 84A219A4249F823800E189FC /* SOverrideMaterial.h */; };
@@ -926,7 +926,6 @@
 		5E34C9661B7F6A7600F212E8 /* CSoftwareTexture2.cpp */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.cpp.cpp; path = CSoftwareTexture2.cpp; sourceTree = "<group>"; };
 		5E34C9671B7F6A7600F212E8 /* CSoftwareTexture2.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; path = CSoftwareTexture2.h; sourceTree = "<group>"; };
 		5E34C9681B7F6A7600F212E8 /* CTRGouraud2.cpp */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.cpp.cpp; path = CTRGouraud2.cpp; sourceTree = "<group>"; };
-		5E34C9691B7F6A7600F212E8 /* CTRGouraudAlpha2.cpp */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.cpp.cpp; path = CTRGouraudAlpha2.cpp; sourceTree = "<group>"; };
 		5E34C96A1B7F6A7600F212E8 /* CTRGouraudAlphaNoZ2.cpp */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.cpp.cpp; path = CTRGouraudAlphaNoZ2.cpp; sourceTree = "<group>"; };
 		5E34C96B1B7F6A7600F212E8 /* CTRNormalMap.cpp */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.cpp.cpp; path = CTRNormalMap.cpp; sourceTree = "<group>"; };
 		5E34C96C1B7F6A7600F212E8 /* CTRStencilShadow.cpp */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.cpp.cpp; path = CTRStencilShadow.cpp; sourceTree = "<group>"; };
@@ -1047,6 +1046,7 @@
 		5E8570BA1B7F9AC400B267D2 /* CIrrDeviceConsole.cpp */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.cpp.cpp; path = CIrrDeviceConsole.cpp; sourceTree = "<group>"; };
 		5E8570BB1B7F9AC400B267D2 /* CIrrDeviceConsole.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; path = CIrrDeviceConsole.h; sourceTree = "<group>"; };
 		5EC24F5E1B8B861200DCA615 /* exampleHelper.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; name = exampleHelper.h; path = ../../include/exampleHelper.h; sourceTree = "<group>"; };
+		65E74BB2282193B40004CA1D /* CTRParallaxMap.cpp */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.cpp.cpp; path = CTRParallaxMap.cpp; sourceTree = "<group>"; };
 		8493BC01249DEBB20092813E /* IOctreeSceneNode.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = IOctreeSceneNode.h; path = ../../include/IOctreeSceneNode.h; sourceTree = "<group>"; };
 		8493BC03249DED3F0092813E /* IMemoryReadFile.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = IMemoryReadFile.h; path = ../../include/IMemoryReadFile.h; sourceTree = "<group>"; };
 		84A219A4249F823800E189FC /* SOverrideMaterial.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = SOverrideMaterial.h; path = ../../include/SOverrideMaterial.h; sourceTree = "<group>"; };
@@ -1918,10 +1918,10 @@
 				5E34C9661B7F6A7600F212E8 /* CSoftwareTexture2.cpp */,
 				5E34C9671B7F6A7600F212E8 /* CSoftwareTexture2.h */,
 				5E34C9681B7F6A7600F212E8 /* CTRGouraud2.cpp */,
-				5E34C9691B7F6A7600F212E8 /* CTRGouraudAlpha2.cpp */,
 				5E34C96A1B7F6A7600F212E8 /* CTRGouraudAlphaNoZ2.cpp */,
 				84A219B8249F834000E189FC /* CTRGouraudNoZ2.cpp */,
 				5E34C96B1B7F6A7600F212E8 /* CTRNormalMap.cpp */,
+				65E74BB2282193B40004CA1D /* CTRParallaxMap.cpp */,
 				5E34C96C1B7F6A7600F212E8 /* CTRStencilShadow.cpp */,
 				5E34C96D1B7F6A7600F212E8 /* CTRTextureBlend.cpp */,
 				5E34C96E1B7F6A7600F212E8 /* CTRTextureDetailMap2.cpp */,
@@ -2410,7 +2410,6 @@
 				5E34CB931B7F6EC500F212E8 /* CSoftwareDriver2.cpp in Sources */,
 				5E34CB951B7F6EC500F212E8 /* CSoftwareTexture2.cpp in Sources */,
 				5E34CB971B7F6EC500F212E8 /* CTRGouraud2.cpp in Sources */,
-				5E34CB981B7F6EC500F212E8 /* CTRGouraudAlpha2.cpp in Sources */,
 				5E34CB991B7F6EC500F212E8 /* CTRGouraudAlphaNoZ2.cpp in Sources */,
 				5E34CB9A1B7F6EC500F212E8 /* CTRNormalMap.cpp in Sources */,
 				5E34CB9B1B7F6EC500F212E8 /* CTRStencilShadow.cpp in Sources */,
@@ -2418,6 +2417,7 @@
 				5E34CB9D1B7F6EC500F212E8 /* CTRTextureDetailMap2.cpp in Sources */,
 				5E34CB9E1B7F6EC500F212E8 /* CTRTextureGouraud2.cpp in Sources */,
 				5E34CB9F1B7F6EC500F212E8 /* CTRTextureGouraudAdd2.cpp in Sources */,
+				65E74BB3282193B50004CA1D /* CTRParallaxMap.cpp in Sources */,
 				5E34CBA01B7F6EC500F212E8 /* CTRTextureGouraudAddNoZ2.cpp in Sources */,
 				5E34CBA11B7F6EC500F212E8 /* CTRTextureGouraudAlpha.cpp in Sources */,
 				5E34CBA21B7F6EC500F212E8 /* CTRTextureGouraudAlphaNoZ.cpp in Sources */,
diff --git a/source/Irrlicht/Irrlicht14.0.vcxproj b/source/Irrlicht/Irrlicht14.0.vcxproj
index 27faa97e..e355562d 100644
--- a/source/Irrlicht/Irrlicht14.0.vcxproj
+++ b/source/Irrlicht/Irrlicht14.0.vcxproj
@@ -849,30 +849,30 @@
     </Link>
   </ItemDefinitionGroup>
   <ItemGroup>
-    <ClInclude Include="..\..\include\EDeviceTypes.h" />  
+    <ClInclude Include="..\..\include\EDeviceTypes.h" />
     <ClInclude Include="..\..\include\EDriverFeatures.h" />
-    <ClInclude Include="..\..\include\EHardwareBufferFlags.h" />	
+    <ClInclude Include="..\..\include\EHardwareBufferFlags.h" />
     <ClInclude Include="..\..\include\EMaterialFlags.h" />
     <ClInclude Include="..\..\include\EMaterialTypes.h" />
     <ClInclude Include="..\..\include\EReadFileType.h" />
     <ClInclude Include="..\..\include\EShaderTypes.h" />
-    <ClInclude Include="..\..\include\fast_atof.h" />	
+    <ClInclude Include="..\..\include\fast_atof.h" />
     <ClInclude Include="..\..\include\IAnimatedMeshMD3.h" />
-    <ClInclude Include="..\..\include\IBillboardTextSceneNode.h" />	
-	<ClInclude Include="..\..\include\IBoneSceneNode.h" />
-    <ClInclude Include="..\..\include\IColladaMeshWriter.h" />	
+    <ClInclude Include="..\..\include\IBillboardTextSceneNode.h" />
+    <ClInclude Include="..\..\include\IBoneSceneNode.h" />
+    <ClInclude Include="..\..\include\IColladaMeshWriter.h" />
     <ClInclude Include="..\..\include\IDynamicMeshBuffer.h" />
     <ClInclude Include="..\..\include\IEventReceiver.h" />
-    <ClInclude Include="..\..\include\IFileArchive.h" />	
-	<ClInclude Include="..\..\include\IGUIFontBitmap.h" />
+    <ClInclude Include="..\..\include\IFileArchive.h" />
+    <ClInclude Include="..\..\include\IGUIFontBitmap.h" />
     <ClInclude Include="..\..\include\IGUITable.h" />
     <ClInclude Include="..\..\include\IImageWriter.h" />
     <ClInclude Include="..\..\include\IIndexBuffer.h" />
-    <ClInclude Include="..\..\include\ILightManager.h" />	
+    <ClInclude Include="..\..\include\ILightManager.h" />
     <ClInclude Include="..\..\include\IOctreeSceneNode.h" />
     <ClInclude Include="..\..\include\IProfiler.h" />
     <ClInclude Include="..\..\include\ILogger.h" />
-    <ClInclude Include="..\..\include\IMemoryReadFile.h" />	
+    <ClInclude Include="..\..\include\IMemoryReadFile.h" />
     <ClInclude Include="..\..\include\IOSOperator.h" />
     <ClInclude Include="..\..\include\IRandomizer.h" />
     <ClInclude Include="..\..\include\IReferenceCounted.h" />
@@ -882,9 +882,9 @@
     <ClInclude Include="..\..\include\leakHunter.h" />
     <ClInclude Include="..\..\include\IrrlichtDevice.h" />
     <ClInclude Include="..\..\include\irrTypes.h" />
-    <ClInclude Include="..\..\include\ISceneUserDataSerializer.h" />	
+    <ClInclude Include="..\..\include\ISceneUserDataSerializer.h" />
     <ClInclude Include="..\..\include\ITimer.h" />
-    <ClInclude Include="..\..\include\IVertexBuffer.h" />	
+    <ClInclude Include="..\..\include\IVertexBuffer.h" />
     <ClInclude Include="..\..\include\Keycodes.h" />
     <ClInclude Include="..\..\include\SIrrCreationParameters.h" />
     <ClInclude Include="..\..\include\SKeyMap.h" />
@@ -926,7 +926,7 @@
     <ClInclude Include="..\..\include\SOverrideMaterial.h" />
     <ClInclude Include="..\..\include\SSharedMeshBuffer.h" />
     <ClInclude Include="..\..\include\SVertexIndex.h" />
-    <ClInclude Include="..\..\include\SVertexManipulator.h" />	
+    <ClInclude Include="..\..\include\SVertexManipulator.h" />
     <ClInclude Include="..\..\include\triangle3d.h" />
     <ClInclude Include="..\..\include\vector2d.h" />
     <ClInclude Include="..\..\include\vector3d.h" />
@@ -1068,7 +1068,7 @@
     <ClInclude Include="COpenGLSLMaterialRenderer.h" />
     <ClInclude Include="CSceneManager.h" />
     <ClInclude Include="CWGLManager.h" />
-	<ClInclude Include="ISceneNodeAnimatorFinishing.h" />
+    <ClInclude Include="ISceneNodeAnimatorFinishing.h" />
     <ClInclude Include="Octree.h" />
     <ClInclude Include="CSMFMeshFileLoader.h" />
     <ClInclude Include="C3DSMeshFileLoader.h" />
@@ -1342,6 +1342,7 @@
     <ClCompile Include="CSkinnedMesh.cpp" />
     <ClCompile Include="CSTLMeshFileLoader.cpp" />
     <ClCompile Include="CTRGouraudNoZ2.cpp" />
+    <ClCompile Include="CTRParallaxMap.cpp" />
     <ClCompile Include="CTR_transparent_reflection_2_layer.cpp" />
     <ClCompile Include="CWGLManager.cpp" />
     <ClCompile Include="CXMeshFileLoader.cpp" />
@@ -1445,7 +1446,6 @@
     <ClCompile Include="CSoftwareDriver2.cpp" />
     <ClCompile Include="CSoftwareTexture2.cpp" />
     <ClCompile Include="CTRGouraud2.cpp" />
-    <ClCompile Include="CTRGouraudAlpha2.cpp" />
     <ClCompile Include="CTRGouraudAlphaNoZ2.cpp" />
     <ClCompile Include="CTRNormalMap.cpp" />
     <ClCompile Include="CTRStencilShadow.cpp" />
diff --git a/source/Irrlicht/Irrlicht14.0.vcxproj.filters b/source/Irrlicht/Irrlicht14.0.vcxproj.filters
index b4cd6a74..330b0260 100644
--- a/source/Irrlicht/Irrlicht14.0.vcxproj.filters
+++ b/source/Irrlicht/Irrlicht14.0.vcxproj.filters
@@ -1282,7 +1282,7 @@
     <ClInclude Include="..\..\include\EMaterialFlags.h">
       <Filter>include\video</Filter>
     </ClInclude>
-	<ClInclude Include="..\..\include\SSharedMeshBuffer.h">
+    <ClInclude Include="..\..\include\SSharedMeshBuffer.h">
       <Filter>include\scene</Filter>
     </ClInclude>
     <ClInclude Include="..\..\include\IRenderTarget.h">
@@ -1372,7 +1372,7 @@
     <ClInclude Include="CBlit.h">
       <Filter>Irrlicht\video\Null</Filter>
     </ClInclude>
-	<ClInclude Include="..\..\include\fast_atof.h">
+    <ClInclude Include="..\..\include\fast_atof.h">
       <Filter>include\core</Filter>
     </ClInclude>
     <ClInclude Include="..\..\include\ILightManager.h">
@@ -1437,7 +1437,7 @@
     </ClInclude>
     <ClInclude Include="..\..\include\EDeviceTypes.h">
       <Filter>include</Filter>
-    </ClInclude>	
+    </ClInclude>
   </ItemGroup>
   <ItemGroup>
     <None Include="..\..\changes.txt">
@@ -1838,9 +1838,6 @@
     <ClCompile Include="CTRGouraud2.cpp">
       <Filter>Irrlicht\video\Burning Video</Filter>
     </ClCompile>
-    <ClCompile Include="CTRGouraudAlpha2.cpp">
-      <Filter>Irrlicht\video\Burning Video</Filter>
-    </ClCompile>
     <ClCompile Include="CTRGouraudAlphaNoZ2.cpp">
       <Filter>Irrlicht\video\Burning Video</Filter>
     </ClCompile>
@@ -2384,6 +2381,9 @@
     <ClCompile Include="CTRGouraudNoZ2.cpp">
       <Filter>Irrlicht\video\Burning Video</Filter>
     </ClCompile>
+    <ClCompile Include="CTRParallaxMap.cpp">
+      <Filter>Irrlicht\video\Burning Video</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <ResourceCompile Include="Irrlicht.rc" />
diff --git a/source/Irrlicht/Irrlicht15.0.vcxproj b/source/Irrlicht/Irrlicht15.0.vcxproj
index a35d2acc..72cd3011 100644
--- a/source/Irrlicht/Irrlicht15.0.vcxproj
+++ b/source/Irrlicht/Irrlicht15.0.vcxproj
@@ -1343,6 +1343,7 @@
     <ClCompile Include="CSkinnedMesh.cpp" />
     <ClCompile Include="CSTLMeshFileLoader.cpp" />
     <ClCompile Include="CTRGouraudNoZ2.cpp" />
+		<ClCompile Include="CTRParallaxMap.cpp" />
     <ClCompile Include="CTR_transparent_reflection_2_layer.cpp" />
     <ClCompile Include="CWGLManager.cpp" />
     <ClCompile Include="CXMeshFileLoader.cpp" />
@@ -1446,7 +1447,6 @@
     <ClCompile Include="CSoftwareDriver2.cpp" />
     <ClCompile Include="CSoftwareTexture2.cpp" />
     <ClCompile Include="CTRGouraud2.cpp" />
-    <ClCompile Include="CTRGouraudAlpha2.cpp" />
     <ClCompile Include="CTRGouraudAlphaNoZ2.cpp" />
     <ClCompile Include="CTRNormalMap.cpp" />
     <ClCompile Include="CTRStencilShadow.cpp" />
diff --git a/source/Irrlicht/Makefile b/source/Irrlicht/Makefile
index 0208c015..4b4421d3 100644
--- a/source/Irrlicht/Makefile
+++ b/source/Irrlicht/Makefile
@@ -48,13 +48,15 @@ IRRDRVROBJ = CNullDriver.o COpenGLCacheHandler.o COpenGLDriver.o COpenGLNormalMa
 IRRIMAGEOBJ = CColorConverter.o CImage.o CImageLoaderBMP.o CImageLoaderDDS.o CImageLoaderJPG.o CImageLoaderPCX.o CImageLoaderPNG.o CImageLoaderPSD.o CImageLoaderPVR.o CImageLoaderTGA.o CImageLoaderPPM.o CImageLoaderWAL.o CImageLoaderRGB.o \
 	CImageWriterBMP.o CImageWriterJPG.o CImageWriterPCX.o CImageWriterPNG.o CImageWriterPPM.o CImageWriterPSD.o CImageWriterTGA.o
 IRRVIDEOOBJ = CVideoModeList.o CFPSCounter.o $(IRRDRVROBJ) $(IRRIMAGEOBJ)
-IRRSWRENDEROBJ = CSoftwareDriver.o CSoftwareTexture.o CTRFlat.o CTRFlatWire.o CTRGouraud.o CTRGouraudWire.o CTRNormalMap.o \
-	CTRStencilShadow.o CTRTextureFlat.o CTRTextureFlatWire.o CTRTextureGouraud.o CTRTextureGouraudAdd.o CTRTextureGouraudNoZ.o \
-	CTRTextureGouraudWire.o CZBuffer.o CTRTextureGouraudVertexAlpha2.o CTRTextureGouraudNoZ2.o CTRTextureLightMap2_M2.o \
-	CTRTextureLightMap2_M4.o CTRTextureLightMap2_M1.o CSoftwareDriver2.o CSoftwareTexture2.o CTRTextureGouraud2.o CTRGouraud2.o \
-	CTRGouraudAlpha2.o CTRGouraudAlphaNoZ2.o CTRTextureDetailMap2.o CTRTextureGouraudAdd2.o CTRTextureGouraudAddNoZ2.o CTRTextureWire2.o \
-	CTRTextureLightMap2_Add.o CTRTextureLightMapGouraud2_M4.o IBurningShader.o CTRTextureBlend.o CTRTextureGouraudAlpha.o \
-	CTRTextureGouraudAlphaNoZ.o CDepthBuffer.o CBurningShader_Raster_Reference.o CTR_transparent_reflection_2_layer.o CTRGouraudNoZ2.o burning_shader_color.o
+IRRSWRENDEROBJ = CSoftwareDriver.o CSoftwareTexture.o CTRFlat.o CTRFlatWire.o CTRGouraud.o CTRGouraudWire.o  \
+	CTRTextureFlat.o CTRTextureFlatWire.o CTRTextureGouraud.o CTRTextureGouraudAdd.o CTRTextureGouraudNoZ.o \
+	CTRTextureGouraudWire.o CZBuffer.o
+IRRBURNINGOBJ = CSoftwareDriver2.o CSoftwareTexture2.o CDepthBuffer.o IBurningShader.o CTRNormalMap.o CTRParallaxMap.o \
+	CTRStencilShadow.o CTRTextureGouraudVertexAlpha2.o CTRTextureGouraudNoZ2.o CTRTextureLightMap2_M2.o \
+	CTRTextureLightMap2_M1.o CTRTextureLightMapGouraud2_M4.o CTRTextureLightMap2_M4.o  CTRTextureGouraud2.o CTRGouraud2.o \
+	CTRGouraudAlphaNoZ2.o CTRTextureDetailMap2.o CTRTextureGouraudAdd2.o CTRTextureGouraudAddNoZ2.o CTRTextureWire2.o \
+	CTRTextureLightMap2_Add.o CTRTextureBlend.o CTRTextureGouraudAlpha.o burning_shader_color.o \
+	CTRTextureGouraudAlphaNoZ.o  CBurningShader_Raster_Reference.o CTR_transparent_reflection_2_layer.o CTRGouraudNoZ2.o
 IRRIOOBJ = CFileList.o CFileSystem.o CLimitReadFile.o CMemoryFile.o CReadFile.o CWriteFile.o CXMLReader.o CXMLWriter.o CWADReader.o CZipReader.o CPakReader.o CNPKReader.o CTarReader.o CMountPointReader.o irrXML.o CAttributes.o lzma/LzmaDec.o
 IRROTHEROBJ = CIrrDeviceSDL.o CIrrDeviceLinux.o CIrrDeviceConsole.o CIrrDeviceStub.o CIrrDeviceWin32.o CIrrDeviceFB.o CLogger.o COSOperator.o Irrlicht.o os.o leakHunter.o 	CProfiler.o utf8.o
 IRRGUIOBJ = CGUIButton.o CGUICheckBox.o CGUIComboBox.o CGUIContextMenu.o CGUIEditBox.o CGUIEnvironment.o CGUIFileOpenDialog.o CGUIFont.o CGUIImage.o CGUIInOutFader.o CGUIListBox.o CGUIMenu.o CGUIMeshViewer.o CGUIMessageBox.o CGUIModalScreen.o CGUIScrollBar.o CGUISpinBox.o CGUISkin.o CGUIStaticText.o CGUITabControl.o CGUITable.o CGUIToolBar.o CGUIWindow.o CGUIColorSelectDialog.o CDefaultGUIElementFactory.o CGUISpriteBank.o CGUIImageList.o CGUITreeView.o CGUIProfiler.o
@@ -67,7 +69,7 @@ BZIP2OBJ = bzip2/blocksort.o bzip2/huffman.o bzip2/crctable.o bzip2/randtable.o
 # Next variable is for additional scene nodes etc. of customized Irrlicht versions
 EXTRAOBJ =
 LINKOBJ = $(IRRMESHOBJ) $(IRROBJ) $(IRRPARTICLEOBJ) $(IRRANIMOBJ) \
-	$(IRRVIDEOOBJ) $(IRRSWRENDEROBJ) $(IRRIOOBJ) $(IRROTHEROBJ) \
+	$(IRRVIDEOOBJ) $(IRRSWRENDEROBJ) $(IRRBURNINGOBJ) $(IRRIOOBJ) $(IRROTHEROBJ) \
 	$(IRRGUIOBJ) $(ZLIBOBJ) $(JPEGLIBOBJ) $(LIBPNGOBJ) $(LIBAESGM) \
 	$(BZIP2OBJ) $(EXTRAOBJ)
 
diff --git a/source/Irrlicht/S4DVertex.h b/source/Irrlicht/S4DVertex.h
index a1d972ea..2498e0f4 100644
--- a/source/Irrlicht/S4DVertex.h
+++ b/source/Irrlicht/S4DVertex.h
@@ -166,9 +166,15 @@ struct ALIGN(16) sVec4
 		struct { f32 s, t, p, q; };
 	};
 
-	sVec4() {}
-	sVec4(f32 _x, f32 _y, f32 _z, f32 _w)
+#if __GNUC__
+	//have one warning i can't find
+	sVec4(f32 _x = 0.f, f32 _y = 0.f, f32 _z = 0.f, f32 _w = 0.f)
 		: x(_x), y(_y), z(_z), w(_w) {}
+#else
+	sVec4() {}
+	sVec4(f32 _x, f32 _y, f32 _z, f32 _w=0.f)
+		: x(_x), y(_y), z(_z), w(_w) {}
+#endif
 
 	// f = a * t + b * ( 1 - t )
 	REALINLINE void interpolate(const sVec4& burning_restrict a, const sVec4& burning_restrict b, const ipoltype t)
@@ -244,6 +250,11 @@ struct ALIGN(16) sVec4
 		return x * other.x + y * other.y + z * other.z;
 	}
 
+	REALINLINE f32 dot(const irr::core::vector3df& other) const
+	{
+		return x * other.X + y * other.Y + z * other.Z;
+	}
+
 	REALINLINE f32 dot_minus_xyz(const sVec4& other) const
 	{
 		return x * -other.x + y * -other.y + z * -other.z;
@@ -315,19 +326,19 @@ struct ALIGN(16) sVec4
 		//const f32 l = len * core::reciprocal_squareroot ( r * r + g * g + b * b );
 		f32 l = x * x + y * y + z * z;
 
-		l = l > 0.0000001f ? len / sqrtf(l) : 0.f;
+		l = l > 0.00000001f ? len / sqrtf(l) : 0.f;
 		out.x = (x * l) + ofs;
 		out.y = (y * l) + ofs;
 		out.z = (z * l) + ofs;
 	}
 
 	//shader suppport
-	sVec4(const sVec4& a, double w)
+	sVec4(const sVec4& a, double _w)
 	{
 		x = a.x;
 		y = a.y;
 		z = a.z;
-		this->w = (float)w;
+		w = (float)_w;
 	}
 	sVec4 xyz() const
 	{
@@ -385,15 +396,15 @@ struct ALIGN(16) sVec4
 	}
 
 	//sVec4 is a,r,g,b, alpha pass
-/*
-	void sat(sVec4& dest, const u32 argb) const
+#if 0
+	void sat_alpha_pass(sVec4& dest, const u32 argb) const
 	{
 		dest.a = ((argb & 0xFF000000) >> 24) * (1.f / 255.f);
 		dest.r = r <= 1.f ? r : 1.f;
 		dest.g = g <= 1.f ? g : 1.f;
 		dest.b = b <= 1.f ? b : 1.f;
 	}
-*/
+#endif
 	void sat_alpha_pass(sVec4& dest, const f32 vertex_alpha) const
 	{
 		dest.a = vertex_alpha;
@@ -434,81 +445,7 @@ typedef sVec4 sVec3Pack_unpack;
 
 typedef sVec4 sVec3Color;
 
-#if 0
-//!sVec4 is argb. sVec3Color is rgba
-struct sVec3Color
-{
-	f32 r, g, b, a;
-
-	void set(const f32 s)
-	{
-		r = s;
-		g = s;
-		b = s;
-		a = s;
-	}
-
-
-	void setColorf(const video::SColorf& color)
-	{
-		r = color.r;
-		g = color.g;
-		b = color.b;
-		a = color.a;
-	}
-
-	void add_rgb(const sVec3Color& other)
-	{
-		r += other.r;
-		g += other.g;
-		b += other.b;
-	}
-
-	void mad_rgb(const sVec3Color& other, const f32 v)
-	{
-		r += other.r * v;
-		g += other.g * v;
-		b += other.b * v;
-	}
-
-	void mad_rgbv(const sVec3Color& v0, const sVec3Color& v1)
-	{
-		r += v0.r * v1.r;
-		g += v0.g * v1.g;
-		b += v0.b * v1.b;
-	}
-
-	//sVec4 is a,r,g,b, alpha pass
-	void sat(sVec4& dest, const u32 argb) const
-	{
-		dest.a = ((argb & 0xFF000000) >> 24) * (1.f / 255.f);
-		dest.r = r <= 1.f ? r : 1.f;
-		dest.g = g <= 1.f ? g : 1.f;
-		dest.b = b <= 1.f ? b : 1.f;
-	}
-
-	void sat_xyz(sVec3Pack& dest, const sVec3Color& v1) const
-	{
-		f32 v;
-		v = r * v1.r;	dest.x = v < 1.f ? v : 1.f;
-		v = g * v1.g;	dest.y = v < 1.f ? v : 1.f;
-		v = b * v1.b;	dest.z = v < 1.f ? v : 1.f;
-	}
-
-	void sat_xyz(sVec4& dest, const sVec3Color& v1) const
-	{
-		f32 v;
-		dest.a = 1.f;
-		v = r * v1.r;	dest.r = v < 1.f ? v : 1.f;
-		v = g * v1.g;	dest.g = v < 1.f ? v : 1.f;
-		v = b * v1.b;	dest.b = v < 1.f ? v : 1.f;
-	}
-
-};
-
-#endif
-
-//internal BurningShaderFlag for a Vertex
+//internal BurningShaderFlag for a Vertex (Attributes)
 enum e4DVertexFlag
 {
 	VERTEX4D_CLIPMASK = 0x0000003F,
@@ -526,25 +463,26 @@ enum e4DVertexFlag
 
 	VERTEX4D_FORMAT_MASK = 0xFFFF0000,
 
-	VERTEX4D_FORMAT_MASK_TEXTURE = 0x000F0000,
-	VERTEX4D_FORMAT_TEXTURE_1 = 0x00010000,
-	VERTEX4D_FORMAT_TEXTURE_2 = 0x00020000,
-	VERTEX4D_FORMAT_TEXTURE_3 = 0x00030000,
-	VERTEX4D_FORMAT_TEXTURE_4 = 0x00040000,
+	VERTEX4D_FORMAT_MASK_TEXTURE	= 0x000F0000,
+	VERTEX4D_FORMAT_TEXTURE_1		= 0x00010000,
+	VERTEX4D_FORMAT_TEXTURE_2		= 0x00020000,
+	VERTEX4D_FORMAT_TEXTURE_3		= 0x00030000,
+	VERTEX4D_FORMAT_TEXTURE_4		= 0x00040000,
 
-	VERTEX4D_FORMAT_MASK_COLOR = 0x00F00000,
-	VERTEX4D_FORMAT_COLOR_1 = 0x00100000,
+	VERTEX4D_FORMAT_MASK_COLOR	= 0x00F00000,
+	VERTEX4D_FORMAT_COLOR_1		= 0x00100000,
 	VERTEX4D_FORMAT_COLOR_2_FOG = 0x00200000,
-	VERTEX4D_FORMAT_COLOR_3 = 0x00300000,
-	VERTEX4D_FORMAT_COLOR_4 = 0x00400000,
+	VERTEX4D_FORMAT_COLOR_3		= 0x00300000,
+	VERTEX4D_FORMAT_COLOR_4		= 0x00400000,
 
-	VERTEX4D_FORMAT_MASK_LIGHT = 0x0F000000,
-	VERTEX4D_FORMAT_LIGHT_1 = 0x01000000,
-	VERTEX4D_FORMAT_LIGHT_2 = 0x02000000,
+	VERTEX4D_FORMAT_MASK_LIGHT	= 0x0F000000,
+	VERTEX4D_FORMAT_LIGHT_1		= 0x01000000,
+	//VERTEX4D_FORMAT_LIGHT_2		= 0x02000000,
 
-	VERTEX4D_FORMAT_MASK_TANGENT = 0xF0000000,
-	VERTEX4D_FORMAT_BUMP_DOT3 = 0x10000000,
-	VERTEX4D_FORMAT_SPECULAR = 0x20000000,
+	VERTEX4D_FORMAT_MASK_TANGENT	= 0xF0000000,
+	VERTEX4D_FORMAT_BUMP_DOT3		= 0x10000000,
+	VERTEX4D_FORMAT_PARALLAX		= 0x20000000,
+	//VERTEX4D_FORMAT_SPECULAR		= 0x20000000,
 
 };
 
@@ -554,10 +492,13 @@ enum e4DVertexType
 	E4VT_STANDARD = 0,			// EVT_STANDARD, video::S3DVertex.
 	E4VT_2TCOORDS = 1,			// EVT_2TCOORDS, video::S3DVertex2TCoords.
 	E4VT_TANGENTS = 2,			// EVT_TANGENTS, video::S3DVertexTangents
-	E4VT_REFLECTION_MAP = 3,
-	E4VT_SHADOW = 4,			// float * 3
-	E4VT_NO_TEXTURE = 5,		// runtime if texture missing
-	E4VT_LINE = 6,
+
+	//encode attributes
+	E4VT_TANGENTS_PARALLAX = 4,
+	E4VT_REFLECTION_MAP = 5,
+	E4VT_SHADOW = 6,			// float * 3
+	E4VT_NO_TEXTURE = 7,		// runtime if texture missing
+	E4VT_LINE = 8,
 
 	E4VT_COUNT
 };
@@ -873,13 +814,14 @@ static REALINLINE void memcpy_s4DVertexPair(void* burning_restrict dst, const vo
 }
 
 
-//! hold info for different Vertex Types
+//! hold info for different Vertex Types (Attribute mapping)
 struct SVSize
 {
 	u32 Format;		// e4DVertexFlag VERTEX4D_FORMAT_MASK_TEXTURE
 	u32 Pitch;		// sizeof Vertex
 	u32 TexSize;	// amount Textures
-	u32 TexCooSize;	// sizeof TextureCoordinates
+	u32 TexCooSize;	// amount TextureCoordinates
+	u32 ColSize;	// amount Color Interpolators
 };
 
 
@@ -898,9 +840,6 @@ struct SVertexShader
 	SVertexShader() {}
 	~SVertexShader() {}
 
-	//VertexType
-	SVSize vSize[E4VT_COUNT];
-
 	// Transformed and lite, clipping state
 	// + Clipped, Projected
 	SAligned4DVertex mem;
@@ -919,6 +858,9 @@ struct SVertexShader
 	u32 primitiveHasVertex;
 	u32 primitiveRun;
 
+	//VertexType
+	SVSize vSize[E4VT_COUNT];
+
 	e4DVertexType vType;		//E_VERTEX_TYPE
 	scene::E_PRIMITIVE_TYPE pType;		//scene::E_PRIMITIVE_TYPE
 	e4DIndexType iType;		//E_INDEX_TYPE iType
diff --git a/source/Irrlicht/SoftwareDriver2_helper.h b/source/Irrlicht/SoftwareDriver2_helper.h
index 5e38d990..389e0896 100644
--- a/source/Irrlicht/SoftwareDriver2_helper.h
+++ b/source/Irrlicht/SoftwareDriver2_helper.h
@@ -1154,6 +1154,57 @@ static REALINLINE void getSample_texture(tFixPoint& a, tFixPoint& r, tFixPoint&
 
 }
 
+// get Sample bilinear
+static REALINLINE void getSample_texture(tFixPoint& a,
+	const sInternalTexture* burning_restrict tex, const tFixPointu tx, const tFixPointu ty
+)
+{
+
+	tFixPointu a00;
+	tFixPointu a01;
+	tFixPointu a10;
+	tFixPointu a11;
+
+	size_t o0, o1, o2, o3;
+	tVideoSample t00;
+
+	o0 = (((ty)&tex->textureYMask) >> FIX_POINT_PRE) << tex->pitchlog2;
+	o1 = (((ty + FIX_POINT_ONE) & tex->textureYMask) >> FIX_POINT_PRE) << tex->pitchlog2;
+	o2 = ((tx)&tex->textureXMask) >> (FIX_POINT_PRE - SOFTWARE_DRIVER_2_TEXTURE_GRANULARITY);
+	o3 = ((tx + FIX_POINT_ONE) & tex->textureXMask) >> (FIX_POINT_PRE - SOFTWARE_DRIVER_2_TEXTURE_GRANULARITY);
+
+	t00 = *((tVideoSample*)((u8*)tex->data + (o0 + o2)));
+	a00 = (t00 & MASK_A) >> SHIFT_A;
+
+	t00 = *((tVideoSample*)((u8*)tex->data + (o0 + o3)));
+	a10 = (t00 & MASK_A) >> SHIFT_A;
+
+	t00 = *((tVideoSample*)((u8*)tex->data + (o1 + o2)));
+	a01 = (t00 & MASK_A) >> SHIFT_A;
+
+	t00 = *((tVideoSample*)((u8*)tex->data + (o1 + o3)));
+	a11 = (t00 & MASK_A) >> SHIFT_A;
+
+	const tFixPointu txFract = tx & FIX_POINT_FRACT_MASK;
+	const tFixPointu txFractInv = FIX_POINT_ONE - txFract;
+
+	const tFixPointu tyFract = ty & FIX_POINT_FRACT_MASK;
+	const tFixPointu tyFractInv = FIX_POINT_ONE - tyFract;
+
+	const tFixPointu w00 = imulFixu(txFractInv, tyFractInv);
+	const tFixPointu w10 = imulFixu(txFract, tyFractInv);
+	const tFixPointu w01 = imulFixu(txFractInv, tyFract);
+	const tFixPointu w11 = imulFixu(txFract, tyFract);
+
+	a = (a00 * w00) +
+		(a01 * w01) +
+		(a10 * w10) +
+		(a11 * w11);
+
+	fix_alpha_color_max(a);
+
+}
+
 #else // SOFTWARE_DRIVER_2_BILINEAR
 
 // get Sample linear == getSample_fixpoint
diff --git a/source/Irrlicht/burning_shader_color.cpp b/source/Irrlicht/burning_shader_color.cpp
index 8c3657d4..58d3d490 100644
--- a/source/Irrlicht/burning_shader_color.cpp
+++ b/source/Irrlicht/burning_shader_color.cpp
@@ -1,4 +1,4 @@
-// Copyright (C) 2002-2012 Nikolaus Gebhardt / Thomas Alten
+// Copyright (C) 2022 Thomas Alten
 // This file is part of the "Irrlicht Engine".
 // For conditions of distribution and use, see copyright notice in irrlicht.h
 
diff --git a/source/Irrlicht/burning_shader_compile_fragment_default.h b/source/Irrlicht/burning_shader_compile_fragment_default.h
index eeae3d32..4cb3d224 100644
--- a/source/Irrlicht/burning_shader_compile_fragment_default.h
+++ b/source/Irrlicht/burning_shader_compile_fragment_default.h
@@ -36,7 +36,7 @@ private:
 
 //! constructor
 burning_shader_class::burning_shader_class(CBurningVideoDriver* driver)
-	: IBurningShader(driver)
+	: IBurningShader(driver,EMT_SOLID)
 {
 #ifdef _DEBUG
 	setDebugName(burning_stringify(burning_shader_class) );
diff --git a/tests/tests-last-passed-at.txt b/tests/tests-last-passed-at.txt
index 281666ac..c40c23b9 100644
--- a/tests/tests-last-passed-at.txt
+++ b/tests/tests-last-passed-at.txt
@@ -1,4 +1,4 @@
-Tests finished. 72 tests of 72 passed.
-Compiled as DEBUG
-Test suite pass at GMT Mon May  2 16:15:12 2022
-
+Tests finished. 72 tests of 72 passed.
+Compiled as RELEASE
+Test suite pass at GMT Tue May  3 20:01:55 2022
+