Burningsvideo 0.52

- scissor for 2D
- downscaled internal rendertargets,interlaced drawing
- supertuxkart gui ok

git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6154 dfc29bdd-3216-0410-991c-e03cc46cb475

changes.txt

@@ -13,15 +13,16 @@ Changes in 1.9 (not yet released)
 - Improvements to B3D writer for speed, readability and handling of low framerate animations.
   Thanks @JLouisB for the patch (For more info, see: http://irrlicht.sourceforge.net/forum/viewtopic.php?f=2&t=50067&start=15)
 - Add another render pass ESNRP_GUI which is drawn last and is p.E. useful for rendering gui nodes in the scenemanager.
-- BurningVideo: 0.51
+- BurningVideo: 0.52
   - 10 year anniversary update
   - Lighting model reworked. moved to eyespace like openGL. [Specular Highlights, Fog, Sphere/Reflection Map] 
   - increased internal s4DVertex to support 4 Textures and 4 Colors [switchable]
   - Textures are handled as sRGB during Mipmap Generation. More accurate, less visual disruption
-  - 2D is drawn as 3D like hardware drivers. [switchable]. enables viewport scaling, material2D
+  - 2D is drawn as 3D like hardware drivers. [switchable]. enables viewport scaling, material2D, scissor
   - Texture Spatial Resolution Limiting working. [lower memory consumption,SOFTWARE_DRIVER_2_TEXTURE_MAXSIZE]
   - NormalMap for 1 Light accurate. still all lights are accumulated
   - SuperTuxKart 8.0.1 playable
+  - Internal Backbuffer Scaling and Interlacing
   - Known Problems
     - Depthbuffer range not equal to Hardware Drivers. Problems with Orthographic Stencil Shadows
 	- Triangle MipMap Selection. Wrong for TextureAtlas and Billboards

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 0.51");
+	box->addItem(L"Burning's Video 0.52");
 	box->addItem(L"Irrlicht Software Renderer 1.0");
 	switch (driverType )
 	{

include/IImage.h

@@ -29,6 +29,9 @@ public:
 	//! constructor
 	IImage(ECOLOR_FORMAT format, const core::dimension2d<u32>& size, bool deleteMemory) :
 		Format(format), Size(size), Data(0), MipMapsData(0), BytesPerPixel(0), Pitch(0), DeleteMemory(deleteMemory), DeleteMipMapsMemory(false)
+#if defined(IRRLICHT_sRGB)
+		,Format_sRGB(1)
+#endif
 	{
 		BytesPerPixel = getBitsPerPixelFromFormat(Format) / 8;
 		Pitch = BytesPerPixel * Size.Width;
@@ -50,6 +53,18 @@ public:
 		return Format;
 	}
 
+#if defined(IRRLICHT_sRGB)
+	//! Texture is linear/sRGB (should be part of ColorFormat: default yes)
+	int get_sRGB() const
+	{
+		return Format_sRGB;
+	}
+	void set_sRGB(int val)
+	{
+		Format_sRGB = val;
+	}
+#endif
+
 	//! Returns width and height of image data.
 	const core::dimension2d<u32>& getDimension() const
 	{
@@ -531,6 +546,22 @@ public:
 		return false;
 	}
 
+#if defined(PATCH_SUPERTUX_8_0_1_with_1_9_0)
+	static bool isRenderTargetOnlyFormat(const ECOLOR_FORMAT format)
+	{
+		switch (format)
+		{
+		case ECF_A1R5G5B5:
+		case ECF_R5G6B5:
+		case ECF_R8G8B8:
+		case ECF_A8R8G8B8:
+			return false;
+		default:
+			return true;
+		}
+	}
+#endif
+
 protected:
 	ECOLOR_FORMAT Format;
 	core::dimension2d<u32> Size;
@@ -545,8 +576,12 @@ protected:
 	bool DeleteMipMapsMemory;
 
 	core::irrAllocator<u8> Allocator;
+#if defined(IRRLICHT_sRGB)
+	int Format_sRGB;
+#endif
 };
 
+
 } // end namespace video
 } // end namespace irr
 

source/Irrlicht/CBurningShader_Raster_Reference.cpp

@@ -698,7 +698,7 @@ REALINLINE void CBurningShader_Raster_Reference::scanline2()
 		return;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2( line.x[1] - line.x[0] );
+	const f32 invDeltaX = fill_step_x( line.x[1] - line.x[0] );
 	const f32 subPixel = ( (f32) pShader.xStart ) - line.x[0];
 
 	// store slopes in endpoint, and correct first pixel
@@ -722,8 +722,8 @@ REALINLINE void CBurningShader_Raster_Reference::scanline2()
 	}
 
 	SOFTWARE_DRIVER_2_CLIPCHECK_REF;
-	pShader.dst = (tVideoSample*) ( (u8*) RenderTarget->getData() + ( line.y * RenderTarget->getPitch() ) + ( pShader.xStart << VIDEO_SAMPLE_GRANULARITY ) );
-	pShader.z = (fp24*) ( (u8*) DepthBuffer->lock() + ( line.y * DepthBuffer->getPitch() ) + ( pShader.xStart << VIDEO_SAMPLE_GRANULARITY ) );
+	pShader.dst = (tVideoSample*) ( (u8*) RenderTarget->getData() + ( line.y * RenderTarget->getPitch() ) + ( pShader.xStart << SOFTWARE_DRIVER_2_RENDERTARGET_GRANULARITY) );
+	pShader.z = (fp24*) ( (u8*) DepthBuffer->lock() + ( line.y * DepthBuffer->getPitch() ) + ( pShader.xStart << SOFTWARE_DRIVER_2_RENDERTARGET_GRANULARITY) );
 
 	for ( pShader.i = 0; pShader.i <= pShader.dx; ++pShader.i )
 	{
@@ -764,10 +764,10 @@ REALINLINE void CBurningShader_Raster_Reference::scanline ()
 		return;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2( line.x[1] - line.x[0] );
+	const f32 invDeltaX = fill_step_x( line.x[1] - line.x[0] );
 
 	// search z-buffer for first not occulled pixel
-	pShader.z = (fp24*) ( (u8*) DepthBuffer->lock() + ( line.y * DepthBuffer->getPitch() ) + ( pShader.xStart << VIDEO_SAMPLE_GRANULARITY ) );
+	pShader.z = (fp24*) ( (u8*) DepthBuffer->lock() + ( line.y * DepthBuffer->getPitch() ) + ( pShader.xStart << SOFTWARE_DRIVER_2_RENDERTARGET_GRANULARITY) );
 
 	// subTexel
 	const f32 subPixel = ( (f32) pShader.xStart ) - line.x[0];
@@ -807,7 +807,7 @@ REALINLINE void CBurningShader_Raster_Reference::scanline ()
 	line.w[0] = a;
 	line.w[1] = b;
 
-	pShader.dst = (tVideoSample*) ( (u8*) RenderTarget->getData() + ( line.y * RenderTarget->getPitch() ) + ( pShader.xStart << VIDEO_SAMPLE_GRANULARITY ) );
+	pShader.dst = (tVideoSample*) ( (u8*) RenderTarget->getData() + ( line.y * RenderTarget->getPitch() ) + ( pShader.xStart << SOFTWARE_DRIVER_2_RENDERTARGET_GRANULARITY) );
 
 	a = (f32) pShader.i + subPixel;
 
@@ -863,9 +863,9 @@ void CBurningShader_Raster_Reference::drawTriangle(const s4DVertex* burning_rest
 
 
 	// calculate delta y of the edges
-	scan.invDeltaY[0] = reciprocal_zero2( c->Pos.y - a->Pos.y );
-	scan.invDeltaY[1] = reciprocal_zero2( b->Pos.y - a->Pos.y );
-	scan.invDeltaY[2] = reciprocal_zero2( c->Pos.y - b->Pos.y );
+	scan.invDeltaY[0] = fill_step_y( c->Pos.y - a->Pos.y );
+	scan.invDeltaY[1] = fill_step_y( b->Pos.y - a->Pos.y );
+	scan.invDeltaY[2] = fill_step_y( c->Pos.y - b->Pos.y );
 
 	if ( F32_LOWER_EQUAL_0 ( scan.invDeltaY[0] )  )
 		return;
@@ -959,7 +959,7 @@ void CBurningShader_Raster_Reference::drawTriangle(const s4DVertex* burning_rest
 		}
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		for( line.y = yStart; line.y <= yEnd; line.y += SOFTWARE_DRIVER_2_STEP_Y)
 		{
 			line.x[scan.left] = scan.x[0];
 			line.w[scan.left] = scan.w[0];
@@ -1076,7 +1076,7 @@ void CBurningShader_Raster_Reference::drawTriangle(const s4DVertex* burning_rest
 		}
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		for( line.y = yStart; line.y <= yEnd; line.y += SOFTWARE_DRIVER_2_STEP_Y)
 		{
 			line.x[scan.left] = scan.x[0];
 			line.w[scan.left] = scan.w[0];

source/Irrlicht/CDepthBuffer.cpp

@@ -40,7 +40,7 @@ CDepthBuffer::~CDepthBuffer()
 
 
 //! clears the zbuffer
-void CDepthBuffer::clear(f32 value)
+void CDepthBuffer::clear(f32 value, interlaced_control interlaced)
 {
 	ieee754 zMaxValue;
 
@@ -50,7 +50,7 @@ void CDepthBuffer::clear(f32 value)
 	zMaxValue.f = value;
 #endif
 
-	memset32 ( Buffer, zMaxValue.u, TotalSize );
+	memset32_interlaced(Buffer, zMaxValue.u, Pitch, Size.Height, interlaced);
 }
 
 
@@ -66,9 +66,10 @@ void CDepthBuffer::setSize(const core::dimension2d<u32>& size)
 	delete [] Buffer;
 
 	Pitch = size.Width * sizeof ( fp24 );
-	TotalSize = Pitch * size.Height;
+	size_t TotalSize = Pitch * size.Height;
 	Buffer = new u8[align_next(TotalSize,16)];
-	clear ();
+
+	clear( 1.f, interlace_disabled());
 }
 
 
@@ -107,7 +108,7 @@ CStencilBuffer::~CStencilBuffer()
 
 
 //! clears the buffer
-void CStencilBuffer::clear(u8 value)
+void CStencilBuffer::clear(u32 value, const interlaced_control interlaced)
 {
 	u32 set = value;
 	if (Bit == 8)
@@ -115,7 +116,7 @@ void CStencilBuffer::clear(u8 value)
 		set |= set << 8;
 		set |= set << 16;
 	}
-	memset32 ( Buffer, set, TotalSize );
+	memset32_interlaced ( Buffer, set, Pitch,Size.Height,interlaced );
 }
 
 
@@ -131,9 +132,10 @@ void CStencilBuffer::setSize(const core::dimension2d<u32>& size)
 	delete [] Buffer;
 
 	Pitch = size.Width * sizeof (tStencilSample);
-	TotalSize = Pitch * size.Height;
+	size_t TotalSize = Pitch * size.Height;
 	Buffer = new u8[align_next(TotalSize,16)];
-	clear ();
+
+	clear(0, interlace_disabled());
 }
 
 

source/Irrlicht/CDepthBuffer.h

@@ -23,7 +23,7 @@ namespace video
 		virtual ~CDepthBuffer();
 
 		//! clears the zbuffer
-		virtual void clear(f32 value=1.f) _IRR_OVERRIDE_;
+		virtual void clear(f32 value, const interlaced_control interlaced) _IRR_OVERRIDE_;
 
 		//! sets the new size of the zbuffer
 		virtual void setSize(const core::dimension2d<u32>& size) _IRR_OVERRIDE_;
@@ -45,7 +45,6 @@ namespace video
 
 		u8* Buffer;
 		core::dimension2d<u32> Size;
-		u32 TotalSize;
 		u32 Pitch;
 	};
 
@@ -61,7 +60,7 @@ namespace video
 		virtual ~CStencilBuffer();
 
 		//! clears the zbuffer
-		virtual void clear(u8 value=0) _IRR_OVERRIDE_;
+		virtual void clear(u32 value, const interlaced_control interlaced) _IRR_OVERRIDE_;
 
 		//! sets the new size of the zbuffer
 		virtual void setSize(const core::dimension2d<u32>& size) _IRR_OVERRIDE_;
@@ -82,7 +81,6 @@ namespace video
 	private:
 		u8* Buffer;
 		core::dimension2d<u32> Size;
-		u32 TotalSize;
 		u32 Pitch;
 		u32 Bit;
 	};

source/Irrlicht/CSoftwareDriver2.h

@@ -50,6 +50,18 @@ namespace video
 		virtual bool beginScene(u16 clearFlag, SColor clearColor, f32 clearDepth, u8 clearStencil,
 			const SExposedVideoData& videoData, core::rect<s32>* sourceRect) _IRR_OVERRIDE_;
 
+#if defined(PATCH_SUPERTUX_8_0_1_with_1_9_0)
+		virtual bool beginScene(bool backBuffer, bool zBuffer, SColor color,
+			const SExposedVideoData& videoData, core::rect<s32>* sourceRect)
+		{
+			u16 flag = 0;
+			if (backBuffer)	flag |= ECBF_COLOR;
+			if (zBuffer) flag |= ECBF_DEPTH;
+			return beginScene(flag, color, 1.f, 0, videoData, sourceRect);
+		}
+		virtual bool setRenderTarget(video::ITexture* texture, bool clearBackBuffer, bool clearZBuffer, SColor color);
+#endif
+
 		virtual bool endScene() _IRR_OVERRIDE_;
 
 		//! Only used by the internal engine. Used to notify the driver that
@@ -152,7 +164,12 @@ namespace video
 
 		//! Creates a render target texture.
 		virtual ITexture* addRenderTargetTexture(const core::dimension2d<u32>& size,
-			const io::path& name, const ECOLOR_FORMAT format = ECF_UNKNOWN) _IRR_OVERRIDE_;
+			const io::path& name, const ECOLOR_FORMAT format = ECF_UNKNOWN
+#if defined(PATCH_SUPERTUX_8_0_1_with_1_9_0)
+			, const bool useStencil = false
+#endif
+		) _IRR_OVERRIDE_;
+
 
 		virtual void clearBuffers(u16 flag, SColor color, f32 depth, u8 stencil) _IRR_OVERRIDE_;
 
@@ -190,8 +207,10 @@ namespace video
 		//! Check if the driver supports creating textures with the given color format
 		virtual bool queryTextureFormat(ECOLOR_FORMAT format) const _IRR_OVERRIDE_;
 
+#if !defined(PATCH_SUPERTUX_8_0_1_with_1_9_0)
 		//! Used by some SceneNodes to check if a material should be rendered in the transparent render pass
 		virtual bool needsTransparentRenderPass(const irr::video::SMaterial& material) const _IRR_OVERRIDE_;
+#endif
 
 		IDepthBuffer * getDepthBuffer () { return DepthBuffer; }
 		IStencilBuffer * getStencilBuffer () { return StencilBuffer; }
@@ -252,6 +271,33 @@ namespace video
 		
 		virtual void setPixelShaderConstant(const f32* data, s32 startRegister, s32 constantAmount) _IRR_OVERRIDE_;
 
+#if defined(PATCH_SUPERTUX_8_0_1_with_1_9_0)
+		virtual bool setVertexShaderConstant(const c8* name, const f32* floats, int count)
+		{
+			return setVertexShaderConstant(getVertexShaderConstantID(name), floats, count);
+		}
+		virtual bool setVertexShaderConstant(const c8* name, const bool* bools, int count)
+		{
+			return setVertexShaderConstant(getVertexShaderConstantID(name), (const s32*)bools, count);
+		}
+		virtual bool setVertexShaderConstant(const c8* name, const s32* ints, int count)
+		{
+			return setVertexShaderConstant(getVertexShaderConstantID(name), ints, count);
+		}
+
+		virtual bool setPixelShaderConstant(const c8* name, const f32* floats, int count)
+		{
+			return setPixelShaderConstant(getPixelShaderConstantID(name), floats, count);
+		}
+		virtual bool setPixelShaderConstant(const c8* name, const bool* bools, int count)
+		{
+			return setPixelShaderConstant(getPixelShaderConstantID(name), (const s32*)bools, count);
+		}
+		virtual bool setPixelShaderConstant(const c8* name, const s32* ints, int count)
+		{
+			return setPixelShaderConstant(getPixelShaderConstantID(name), ints, count);
+		}
+#endif
 		//! Get pointer to the IVideoDriver interface
 		/** \return Pointer to the IVideoDriver interface */
 		virtual IVideoDriver* getVideoDriver() _IRR_OVERRIDE_;
@@ -261,7 +307,7 @@ namespace video
 		void saveBuffer();
 
 		//! sets a render target
-		void setRenderTargetImage(video::CImage* image);
+		void setRenderTargetImage2(video::IImage* color, video::IImage* depth=0, video::IImage* stencil=0);
 
 		//! sets the current Texture
 		//bool setTexture(u32 stage, video::ITexture* texture);
@@ -278,6 +324,7 @@ namespace video
 		video::ITexture* RenderTargetTexture;
 		video::IImage* RenderTargetSurface;
 		core::dimension2d<u32> RenderTargetSize;
+		sVec4 RatioRenderTargetScreen; // Smaller Render Target
 
 		IBurningShader* CurrentShader;
 		IBurningShader* BurningShader[ETR2_COUNT];
@@ -302,6 +349,8 @@ namespace video
 			ETS_COUNT_BURNING = 16
 		};
 
+		// align manually to 16 byte start address
+		//u8 _pack_0[8];
 		enum E_TRANSFORMATION_FLAG
 		{
 			ETF_VALID = 1,
@@ -311,12 +360,12 @@ namespace video
 			ETF_TEXGEN_WRAP = 16,
 			ETF_TEXGEN_MASK = ETF_TEXGEN_CAMERA_SPHERE | ETF_TEXGEN_CAMERA_REFLECTION | ETF_TEXGEN_WRAP
 		};
+		size_t TransformationStack; // 0 .. 3D , 1 .. 2D
 		core::matrix4 Transformation[2][ETS_COUNT_BURNING];
 		size_t TransformationFlag[2][ETS_COUNT_BURNING]; // E_TRANSFORMATION_FLAG
 		
-		size_t TransformationStack; // 0 .. 3D , 1 .. 2D
 
-		void setRenderStates2DMode(const video::SColor& color,video::ITexture* texture,bool useAlphaChannelOfTexture);
+		void setRenderStates2DMode(const video::SColor& color,const video::ITexture* texture,bool useAlphaChannelOfTexture);
 		void setRenderStates3DMode();
 
 		//ETS_CLIPSCALE, // moved outside to stay at 16 matrices
@@ -366,7 +415,7 @@ namespace video
 
 		//const is misleading. **v is const that true, but not *v..
 		f32 screenarea_inside (const s4DVertexPair* burning_restrict const face[] ) const;
-		s32 lodFactor_inside ( const s4DVertexPair* burning_restrict const face[], const size_t tex, f32 dc_area, f32 lod_bias ) const;
+		s32 lodFactor_inside ( const s4DVertexPair* burning_restrict const face[], const size_t tex, const f32 dc_area, const f32 lod_bias ) const;
 		void select_polygon_mipmap_inside ( s4DVertex* burning_restrict face[], const size_t tex, const CSoftwareTexture2_Bound& b ) const;
 
 		void getCameraPosWorldSpace();
@@ -377,6 +426,20 @@ namespace video
 
 		//! Built-in 2D quad for 2D rendering.
 		S3DVertex Quad2DVertices[4];
+		interlaced_control Interlaced;
+
+#if defined(PATCH_SUPERTUX_8_0_1_with_1_9_0)
+		core::array<IRenderTarget*> RenderTargets;
+
+		inline bool getWriteZBuffer(const SMaterial& material) const
+		{
+			return material.ZWriteEnable && (AllowZWriteOnTransparent || !material.isTransparent());
+		}
+		virtual video::ITexture* createDeviceDependentTexture(IImage* surface, const io::path& name, void* mipmapData = 0)
+		{
+			return createDeviceDependentTexture(name, surface);
+		}
+#endif
 
 	};
 

source/Irrlicht/CSoftwareTexture2.cpp

@@ -20,6 +20,30 @@ namespace video
 //! stretches srcRect src to dstRect dst, applying a sliding window box filter in linear color space (sRGB->linear->sRGB)
 void Resample_subSampling(eBlitter op, video::IImage* dst, const core::rect<s32>* dstRect, const video::IImage* src, const core::rect<s32>* srcRect, size_t flags);
 
+//nearest pow of 2 ( 257 will be 256 not 512 )
+static inline core::dimension2d<u32> getOptimalSize(const core::dimension2d<u32>& original, const u32 allowNonPowerOfTwo, const u32 maxSize)
+{
+	u32 w, h;
+	if (allowNonPowerOfTwo)
+	{
+		w = original.Width;
+		h = original.Height;
+	}
+	else
+	{
+		w = 1;
+		while (w * 2 < original.Width) w *= 2;
+		if (w * 2 - original.Width < original.Width - w) w *= 2;
+
+		h = 1;
+		while (h * 2 < original.Height) h *= 2;
+		if (h * 2 - original.Height < original.Height - h) h *= 2;
+	}
+	if (maxSize && w > maxSize) w = maxSize;
+	if (maxSize && h > maxSize) h = maxSize;
+	return core::dimension2d<u32>(w, h);
+}
+
 //! constructor
 CSoftwareTexture2::CSoftwareTexture2(IImage* image, const io::path& name, u32 flags, CBurningVideoDriver* driver)
 	: ITexture(name
@@ -27,24 +51,24 @@ CSoftwareTexture2::CSoftwareTexture2(IImage* image, const io::path& name, u32 fl
 		, ETT_2D
 #endif
 	)
-	,MipMapLOD(0), Flags(flags), Driver(driver)
+	, MipMapLOD(0), Flags(flags), Driver(driver)
 {
-	#ifdef _DEBUG
+#ifdef _DEBUG
 	setDebugName("CSoftwareTexture2");
-	#endif
+#endif
 
-#ifndef SOFTWARE_DRIVER_2_MIPMAPPING
-	Flags &= ~(GEN_MIPMAP| GEN_MIPMAP_AUTO);
+#if SOFTWARE_DRIVER_2_MIPMAPPING_MAX <= 1
+	Flags &= ~(GEN_MIPMAP | GEN_MIPMAP_AUTO);
 #endif
 	//set baseclass properties
 	DriverType = EDT_BURNINGSVIDEO;
-	ColorFormat = BURNINGSHADER_COLOR_FORMAT;
+	ColorFormat = (Flags & IS_RENDERTARGET) ? SOFTWARE_DRIVER_2_RENDERTARGET_COLOR_FORMAT : SOFTWARE_DRIVER_2_TEXTURE_COLOR_FORMAT;
 	IsRenderTarget = (Flags & IS_RENDERTARGET) != 0;
 	HasMipMaps = (Flags & GEN_MIPMAP) != 0;
 	MipMap0_Area[0] = 1;
 	MipMap0_Area[1] = 1;
-	LodBIAS = 0.75f;
-	for ( size_t i = 0; i < SOFTWARE_DRIVER_2_MIPMAPPING_MAX; ++i ) MipMap[i] = 0;
+	LodBIAS = 1.f;
+	for (size_t i = 0; i < array_size(MipMap); ++i) MipMap[i] = 0;
 	if (!image) return;
 
 	OriginalSize = image->getDimension();
@@ -52,7 +76,7 @@ CSoftwareTexture2::CSoftwareTexture2(IImage* image, const io::path& name, u32 fl
 
 
 #if defined(IRRLICHT_sRGB)
-	if ( Flags & IMAGE_IS_LINEAR ) image->set_sRGB(0);
+	if (Flags & IMAGE_IS_LINEAR) image->set_sRGB(0);
 #else
 	//guessing linear image
 	if (name.find("light") >= 0 ||
@@ -72,47 +96,65 @@ CSoftwareTexture2::CSoftwareTexture2(IImage* image, const io::path& name, u32 fl
 
 	//visual studio code warning
 	u32 maxTexSize = SOFTWARE_DRIVER_2_TEXTURE_MAXSIZE;
-	core::dimension2d<u32> optSize(	OriginalSize.getOptimalSize(
+
+#if defined(PATCH_SUPERTUX_8_0_1_with_1_9_0)
+	if (IsRenderTarget && name.find("RaceGUI::markers") >= 0)
+	{
+		maxTexSize = 0;
+	}
+#endif
+	/*
+		core::dimension2d<u32> optSize(OriginalSize.getOptimalSize(
 			(Flags & ALLOW_NPOT) ? 0 : 1, // requirePowerOfTwo
 			false, // requireSquare
-			(Flags & ALLOW_NPOT) ? 1 : maxTexSize == 0, // larger
+			(Flags & ALLOW_NPOT) ? 1 : maxTexSize == SOFTWARE_DRIVER_2_TEXTURE_MAXSIZE, // larger
 			(Flags & ALLOW_NPOT) ? 0 : maxTexSize // maxValue
 		)
 		);
-
+	*/
+	core::dimension2d<u32> optSize(getOptimalSize(OriginalSize, Flags & ALLOW_NPOT, maxTexSize));
 	if (OriginalSize == optSize)
 	{
-		MipMap[0] = new CImage(BURNINGSHADER_COLOR_FORMAT, image->getDimension());
+		MipMap[0] = new CImage(ColorFormat, image->getDimension());
 #if defined(IRRLICHT_sRGB)
-		MipMap[0]->set_sRGB( (Flags & TEXTURE_IS_LINEAR ) ? 0 : image->get_sRGB()  );
+		MipMap[0]->set_sRGB((Flags & TEXTURE_IS_LINEAR) ? 0 : image->get_sRGB());
 #endif
-		if (!isCompressed)
+		if (!isCompressed && image->getData())
 			image->copyTo(MipMap[0]);
 	}
 	else
 	{
-		char buf[256];
-		core::stringw showName ( name );
-		snprintf_irr ( buf, sizeof(buf), "Burningvideo: Warning Texture %ls reformat %ux%u,%d -> %ux%u,%d",
-						showName.c_str(),
-						OriginalSize.Width, OriginalSize.Height, OriginalColorFormat,
-						optSize.Width, optSize.Height,BURNINGSHADER_COLOR_FORMAT
-					);
-
-		os::Printer::log ( buf, ELL_WARNING );
-		MipMap[0] = new CImage(BURNINGSHADER_COLOR_FORMAT, optSize);
+		MipMap[0] = new CImage(ColorFormat, optSize);
 #if defined(IRRLICHT_sRGB)
-		MipMap[0]->set_sRGB( (Flags & TEXTURE_IS_LINEAR ) ? 0 : image->get_sRGB()  );
+		MipMap[0]->set_sRGB((Flags & TEXTURE_IS_LINEAR) ? 0 : image->get_sRGB());
 #endif
 		if (!isCompressed)
 		{
 			//image->copyToScalingBoxFilter ( MipMap[0],0, false );
-			Resample_subSampling(BLITTER_TEXTURE,MipMap[0],0,image,0, Flags);
+			Resample_subSampling(BLITTER_TEXTURE, MipMap[0], 0, image, 0, Flags);
 		}
 		// if Original Size is used for calculation ( 2D position, font) it will be wrong
 		//OriginalSize = optSize;
 	}
 
+	// Show Information about resizing
+	if (OriginalSize != optSize ||
+		(	OriginalColorFormat != ColorFormat &&
+			!((OriginalColorFormat == ECF_R8G8B8 || OriginalColorFormat == ECF_A1R5G5B5) && ColorFormat == ECF_A8R8G8B8)
+		)
+	)
+	{
+		char buf[256];
+		core::stringw showName(name);
+		snprintf_irr(buf, sizeof(buf), "Burningvideo: Texture '%ls' reformat %ux%u,%s -> %ux%u,%s",
+			showName.c_str(),
+			OriginalSize.Width, OriginalSize.Height, ColorFormatNames[OriginalColorFormat],
+			optSize.Width, optSize.Height, ColorFormatNames[ColorFormat]
+		);
+		os::Printer::log(buf, ELL_DEBUG);
+	}
+
+
 	//select highest mipmap 0
 	regenerateMipMapLevels(image->getMipMapsData());
 }
@@ -121,9 +163,9 @@ CSoftwareTexture2::CSoftwareTexture2(IImage* image, const io::path& name, u32 fl
 //! destructor
 CSoftwareTexture2::~CSoftwareTexture2()
 {
-	for ( size_t i = 0; i < SOFTWARE_DRIVER_2_MIPMAPPING_MAX; ++i )
+	for (size_t i = 0; i < array_size(MipMap); ++i)
 	{
-		if ( MipMap[i] )
+		if (MipMap[i])
 		{
 			MipMap[i]->drop();
 			MipMap[i] = 0;
@@ -132,7 +174,6 @@ CSoftwareTexture2::~CSoftwareTexture2()
 }
 
 
-
 //! Regenerates the mip map levels of the texture. Useful after locking and
 //! modifying the texture
 #if !defined(PATCH_SUPERTUX_8_0_1_with_1_9_0)
@@ -144,9 +185,9 @@ void CSoftwareTexture2::regenerateMipMapLevels(void* data)
 	int i;
 
 	// release
-	for ( i = 1; i < SOFTWARE_DRIVER_2_MIPMAPPING_MAX; ++i )
+	for (i = 1; i < array_size(MipMap); ++i)
 	{
-		if ( MipMap[i] )
+		if (MipMap[i])
 		{
 			MipMap[i]->drop();
 			MipMap[i] = 0;
@@ -155,10 +196,10 @@ void CSoftwareTexture2::regenerateMipMapLevels(void* data)
 
 	core::dimension2d<u32> newSize;
 
-	if (HasMipMaps && ( (Flags & GEN_MIPMAP_AUTO) || 0 == data ) )
+	if (HasMipMaps && ((Flags & GEN_MIPMAP_AUTO) || 0 == data))
 	{
 		//need memory also if autogen mipmap disabled
-		for (i = 1; i < SOFTWARE_DRIVER_2_MIPMAPPING_MAX; ++i)
+		for (i = 1; i < array_size(MipMap); ++i)
 		{
 			const core::dimension2du& upperDim = MipMap[i - 1]->getDimension();
 			//isotropic
@@ -167,7 +208,7 @@ void CSoftwareTexture2::regenerateMipMapLevels(void* data)
 			if (upperDim == newSize)
 				break;
 
-			MipMap[i] = new CImage(BURNINGSHADER_COLOR_FORMAT, newSize);
+			MipMap[i] = new CImage(ColorFormat, newSize);
 #if defined(IRRLICHT_sRGB)
 			MipMap[i]->set_sRGB(MipMap[i - 1]->get_sRGB());
 #endif
@@ -192,13 +233,13 @@ void CSoftwareTexture2::regenerateMipMapLevels(void* data)
 			if (origSize.Height > 1) origSize.Height >>= 1;
 			mip_end += IImage::getDataSizeFromFormat(OriginalColorFormat, origSize.Width, origSize.Height);
 			i += 1;
-		} while ((origSize.Width != 1 || origSize.Height != 1) && i < SOFTWARE_DRIVER_2_MIPMAPPING_MAX);
+		} while ((origSize.Width != 1 || origSize.Height != 1) && i < array_size(MipMap));
 
 		//TODO: this is not true
-		LodBIAS = i*0.75f;
+		LodBIAS = i * 2.f;
 
 		origSize = OriginalSize;
-		for (i = 1; i < SOFTWARE_DRIVER_2_MIPMAPPING_MAX && mip_current < mip_end; ++i)
+		for (i = 1; i < array_size(MipMap) && mip_current < mip_end; ++i)
 		{
 			const core::dimension2du& upperDim = MipMap[i - 1]->getDimension();
 			//isotropic
@@ -210,10 +251,10 @@ void CSoftwareTexture2::regenerateMipMapLevels(void* data)
 			if (origSize.Width > 1) origSize.Width >>= 1;
 			if (origSize.Height > 1) origSize.Height >>= 1;
 
-			if (OriginalColorFormat != BURNINGSHADER_COLOR_FORMAT)
+			if (OriginalColorFormat != ColorFormat)
 			{
 				IImage* tmpImage = new CImage(OriginalColorFormat, origSize, mip_current, true, false);
-				MipMap[i] = new CImage(BURNINGSHADER_COLOR_FORMAT, newSize);
+				MipMap[i] = new CImage(ColorFormat, newSize);
 				if (origSize == newSize)
 					tmpImage->copyTo(MipMap[i]);
 				else
@@ -223,11 +264,11 @@ void CSoftwareTexture2::regenerateMipMapLevels(void* data)
 			else
 			{
 				if (origSize == newSize)
-					MipMap[i] = new CImage(BURNINGSHADER_COLOR_FORMAT, newSize, mip_current, false);
+					MipMap[i] = new CImage(ColorFormat, newSize, mip_current, false);
 				else
 				{
-					MipMap[i] = new CImage(BURNINGSHADER_COLOR_FORMAT, newSize);
-					IImage* tmpImage = new CImage(BURNINGSHADER_COLOR_FORMAT, origSize, mip_current, true, false);
+					MipMap[i] = new CImage(ColorFormat, newSize);
+					IImage* tmpImage = new CImage(ColorFormat, origSize, mip_current, true, false);
 					tmpImage->copyToScalingBoxFilter(MipMap[i]);
 					tmpImage->drop();
 				}
@@ -238,33 +279,33 @@ void CSoftwareTexture2::regenerateMipMapLevels(void* data)
 
 
 	//visualize mipmap
-	for (i=1; i < 0 && i < SOFTWARE_DRIVER_2_MIPMAPPING_MAX; ++i)
+	for (i = 1; i < 0 && i < array_size(MipMap); ++i)
 	{
-/*
 		static u32 color[] = {
-			0x30bf7f00,0x3040bf00,0x30bf00bf,0x3000bf00,
-			0x300080bf,0x30bf4000,0x300040bf,0x307f00bf,
-			0x30bf0000,0x3000bfbf,0x304000bf,0x307fbf00,
-			0x8000bf7f,0x80bf0040,0x80bfbf00,0x800000bf
-		};
-*/
-		static u32 color[] = {
-			0xFFFFFFFF,0xFFFF0000,0xFF00FF00,0xFF0000FF,
-			0xFFFFFF00,0xFF00FFFF,0xFFFF00FF,0xFF0000FF,
-			0xFF0000FF,0xFF0000FF,0xFF0000FF,0xFF0000FF,
-			0xFF0000FF,0xFF0000FF,0xFF0000FF,0xFFFF00FF
+			0xFFFF0000,
+			0xFFFF0000,0xFF00FF00,0xFF0000FF,
+			0xFFFFFF00,0xFF00FFFF,0xFFFF00FF,
+			0xFFff6600,0xFF00ff66,0xFF6600FF,
+			0xFF66ff00,0xFF0066ff,0xFFff0066,
+			0xFF33ff00,0xFF0033ff,0xFF3300ff,
+			0xFF0000FF,0xFF0000FF,0xFF0000FF
 		};
 
-		if ( MipMap[i] )
+		if (MipMap[i])
 		{
-			core::rect<s32> p (core::position2di(0,0),MipMap[i]->getDimension());
+			int border = 0;
+			const core::dimension2du& d = MipMap[i]->getDimension();
+			core::rect<s32> p(0, 0, d.Width, d.Height);
 			SColor c((color[i & 15] & 0x00FFFFFF) | 0xFF000000);
-			Blit(BLITTER_TEXTURE_ALPHA_COLOR_BLEND, MipMap[i], 0, 0, MipMap[i], &p, c.color);
+
+			core::rect<s32> dclip(border, border, d.Width - border, d.Height - border);
+			
+			Blit(BLITTER_TEXTURE_ALPHA_COLOR_BLEND, MipMap[i], &dclip, 0, MipMap[i], &p, c.color);
 		}
 	}
 
 	//save mipmap chain
-	if ( 0 )
+	if (0)
 	{
 		char buf[256];
 		const char* name = getName().getPath().c_str();
@@ -277,11 +318,11 @@ void CSoftwareTexture2::regenerateMipMapLevels(void* data)
 			//if (name[i] == '.') ext = i;
 			i += 1;
 		}
-		for (i = 0; i < SOFTWARE_DRIVER_2_MIPMAPPING_MAX; ++i)
+		for (i = 0; i < array_size(MipMap); ++i)
 		{
 			if (MipMap[i])
 			{
-				snprintf_irr(buf, sizeof(buf),"mip/%s_%02d.png", name + filename,i);
+				snprintf_irr(buf, sizeof(buf), "mip/%s_%02d.png", name + filename, i);
 				Driver->writeImageToFile(MipMap[i], buf);
 			}
 		}
@@ -300,12 +341,16 @@ void CSoftwareTexture2::calcDerivative()
 		MipMap0_Area[0] = dim.Width;
 		MipMap0_Area[1] = dim.Height; // screensize of a triangle
 
+		//TA: try to mimic openGL mipmap. ( don't do this!)
+		//if (MipMap0_Area[0] < 32) MipMap0_Area[0] = 32;
+		//if (MipMap0_Area[1] < 32) MipMap0_Area[1] = 32;
+
 		Size = dim; // MipMap[MipMapLOD]->getDimension();
 		Pitch = MipMap[MipMapLOD]->getPitch();
 	}
 
 	//preCalc mipmap texel center boundaries
-	for ( s32 i = 0; i < SOFTWARE_DRIVER_2_MIPMAPPING_MAX; ++i )
+	for (s32 i = 0; i < array_size(MipMap); ++i)
 	{
 		CSoftwareTexture2_Bound& b = TexBound[i];
 		if (MipMap[i])
@@ -335,7 +380,7 @@ void CSoftwareTexture2::calcDerivative()
 
 CSoftwareRenderTarget2::CSoftwareRenderTarget2(CBurningVideoDriver* driver) : Driver(driver)
 #if defined(PATCH_SUPERTUX_8_0_1_with_1_9_0)
-, IRenderTarget(0)
+	, IRenderTarget(0)
 #endif
 {
 	DriverType = EDT_BURNINGSVIDEO;
@@ -378,12 +423,6 @@ void CSoftwareRenderTarget2::setTexture(const core::array<ITexture*>& texture, I
 	}
 }
 
-ITexture* CSoftwareRenderTarget2::getTexture() const
-{
-	return Texture[0];
-}
-
-
 
 static const float srgb_8bit_to_linear_float[1 << 8] = {
 	0.0f, 3.03527e-4f, 6.07054e-4f, 9.10581e-4f,
@@ -474,7 +513,7 @@ u32 linear_to_srgb_8bit(const float v)
 	ieee754 c;
 	c.f = v;
 	const size_t x = c.u;
-	const u32 *table = (u32*)srgb_8bit_to_linear_float;
+	const u32* table = (u32*)srgb_8bit_to_linear_float;
 	u32 y = 0;
 	y += table[y + 128] <= x ? 128 : 0;
 	y += table[y + 64] <= x ? 64 : 0;
@@ -497,7 +536,7 @@ struct absrect2
 	s32 y1;
 };
 
-static inline int clipTest(absrect2 &o, const core::rect<s32>* a, const absrect2& b)
+static inline int clipTest(absrect2& o, const core::rect<s32>* a, const absrect2& b)
 {
 	if (a == 0)
 	{
@@ -522,19 +561,19 @@ static inline int clipTest(absrect2 &o, const core::rect<s32>* a, const absrect2
 //! stretches srcRect src to dstRect dst, applying a sliding window box filter in linear color space (sRGB->linear->sRGB)
 // todo: texture jumps (mip selection problem)
 void Resample_subSampling(eBlitter op, video::IImage* dst, const core::rect<s32>* dstRect,
-	const video::IImage* src, const core::rect<s32>* srcRect,size_t flags)
+	const video::IImage* src, const core::rect<s32>* srcRect, size_t flags)
 {
+	u8* dstData = (u8*)dst->getData();
 	const absrect2 dst_clip = { 0,0,(s32)dst->getDimension().Width,(s32)dst->getDimension().Height };
 	absrect2 dc;
-	if (clipTest(dc, dstRect, dst_clip)) return;
+	if (clipTest(dc, dstRect, dst_clip) || !dstData) return;
 	const video::ECOLOR_FORMAT dstFormat = dst->getColorFormat();
-	u8* dstData = (u8*)dst->getData();
 
+	const u8* srcData = (u8*)src->getData();
 	const absrect2 src_clip = { 0,0,(s32)src->getDimension().Width,(s32)src->getDimension().Height };
 	absrect2 sc;
-	if (clipTest(sc, srcRect, src_clip)) return;
+	if (clipTest(sc, srcRect, src_clip) || !srcData) return;
 	const video::ECOLOR_FORMAT srcFormat = src->getColorFormat();
-	const u8* srcData = (u8*)src->getData();
 
 #if defined(IRRLICHT_sRGB)
 	const int dst_sRGB = dst->get_sRGB();
@@ -560,14 +599,14 @@ void Resample_subSampling(eBlitter op, video::IImage* dst, const core::rect<s32>
 	for (int dy = dc.y0; dy < dc.y1; ++dy)
 	{
 		f[1] = f[3];
-		f[3] = sc.y0 + (dy + 1 - dc.y0)*scale[1];
+		f[3] = sc.y0 + (dy + 1 - dc.y0) * scale[1];
 		if (f[3] >= sc.y1) f[3] = sc.y1 - 0.001f; //todo:1.f/dim should be enough
 
 		f[2] = (float)sc.x0;
 		for (int dx = dc.x0; dx < dc.x1; ++dx)
 		{
 			f[0] = f[2];
-			f[2] = sc.x0 + (dx + 1 - dc.x0)*scale[0];
+			f[2] = sc.x0 + (dx + 1 - dc.x0) * scale[0];
 			if (f[2] >= sc.x1) f[2] = sc.x1 - 0.001f;
 
 			//accumulate linear color
@@ -599,12 +638,12 @@ void Resample_subSampling(eBlitter op, video::IImage* dst, const core::rect<s32>
 
 					switch (srcFormat)
 					{
-					case video::ECF_A1R5G5B5: sbgra = video::A1R5G5B5toA8R8G8B8(*(u16*)(srcData + (fy*src_clip.x1) * 2 + (fx * 2))); break;
-					case video::ECF_R5G6B5: sbgra = video::R5G6B5toA8R8G8B8(*(u16*)(srcData + (fy*src_clip.x1) * 2 + (fx * 2))); break;
-					case video::ECF_A8R8G8B8: sbgra = *(u32*)(srcData + (fy*src_clip.x1) * 4 + (fx * 4)); break;
+					case video::ECF_A1R5G5B5: sbgra = video::A1R5G5B5toA8R8G8B8(*(u16*)(srcData + (fy * src_clip.x1) * 2 + (fx * 2))); break;
+					case video::ECF_R5G6B5: sbgra = video::R5G6B5toA8R8G8B8(*(u16*)(srcData + (fy * src_clip.x1) * 2 + (fx * 2))); break;
+					case video::ECF_A8R8G8B8: sbgra = *(u32*)(srcData + (fy * src_clip.x1) * 4 + (fx * 4)); break;
 					case video::ECF_R8G8B8:
 					{
-						const u8* p = srcData + (fy*src_clip.x1) * 3 + (fx * 3);
+						const u8* p = srcData + (fy * src_clip.x1) * 3 + (fx * 3);
 						sbgra = 0xFF000000 | p[0] << 16 | p[1] << 8 | p[2];
 					} break;
 					default: break;
@@ -650,16 +689,16 @@ void Resample_subSampling(eBlitter op, video::IImage* dst, const core::rect<s32>
 			}
 			switch (dstFormat)
 			{
-			case video::ECF_A8R8G8B8: *(u32*)(dstData + (dy*dst_clip.x1) * 4 + (dx * 4)) = sbgra; break;
+			case video::ECF_A8R8G8B8: *(u32*)(dstData + (dy * dst_clip.x1) * 4 + (dx * 4)) = sbgra; break;
 			case video::ECF_R8G8B8:
 			{
-				u8* p = dstData + (dy*dst_clip.x1) * 3 + (dx * 3);
+				u8* p = dstData + (dy * dst_clip.x1) * 3 + (dx * 3);
 				p[2] = (sbgra) & 0xFF;
 				p[1] = (sbgra >> 8) & 0xFF;
 				p[0] = (sbgra >> 16) & 0xFF;
 			} break;
-			case video::ECF_A1R5G5B5: *(u16*)(dstData + (dy*dst_clip.x1) * 2 + (dx * 2)) = video::A8R8G8B8toA1R5G5B5(sbgra); break;
-			case video::ECF_R5G6B5:   *(u16*)(dstData + (dy*dst_clip.x1) * 2 + (dx * 2)) = video::A8R8G8B8toR5G6B5(sbgra); break;
+			case video::ECF_A1R5G5B5: *(u16*)(dstData + (dy * dst_clip.x1) * 2 + (dx * 2)) = video::A8R8G8B8toA1R5G5B5(sbgra); break;
+			case video::ECF_R5G6B5:   *(u16*)(dstData + (dy * dst_clip.x1) * 2 + (dx * 2)) = video::A8R8G8B8toR5G6B5(sbgra); break;
 			default:
 				break;
 			}

source/Irrlicht/CSoftwareTexture2.h

@@ -8,7 +8,11 @@
 #include "SoftwareDriver2_compile_config.h"
 
 #include "ITexture.h"
+#if defined(PATCH_SUPERTUX_8_0_1_with_1_9_0)
+#include "IVideoDriver.h"
+#else
 #include "IRenderTarget.h"
+#endif
 #include "CImage.h"
 
 namespace irr
@@ -51,14 +55,18 @@ public:
 	u32 getMipmapLevel(s32 newLevel) const
 	{
 		if ( newLevel < 0 ) newLevel = 0;
-		else if ( newLevel >= SOFTWARE_DRIVER_2_MIPMAPPING_MAX ) newLevel = SOFTWARE_DRIVER_2_MIPMAPPING_MAX - 1;
+		else if ( newLevel >= array_size(MipMap)) newLevel = array_size(MipMap) - 1;
 
 		while ( newLevel > 0 && MipMap[newLevel] == 0 ) newLevel -= 1;
 		return newLevel;
 	}
 
 	//! lock function
+#if defined(PATCH_SUPERTUX_8_0_1_with_1_9_0)
+	virtual void* lock(E_TEXTURE_LOCK_MODE mode, u32 mipmapLevel)
+#else
 	virtual void* lock(E_TEXTURE_LOCK_MODE mode, u32 mipmapLevel, u32 layer, E_TEXTURE_LOCK_FLAGS lockFlags = ETLF_FLIP_Y_UP_RTT) _IRR_OVERRIDE_
+#endif
 	{
 		if (Flags & GEN_MIPMAP)
 		{
@@ -75,18 +83,20 @@ public:
 	virtual void unlock() _IRR_OVERRIDE_
 	{
 	}
-
+/*
 	//! compare the area drawn with the area of the texture
 	f32 getLODFactor( const f32 texArea ) const
 	{
 		return MipMap0_Area[0]* MipMap0_Area[1] * 0.5f * texArea;
 		//return MipMap[0]->getImageDataSizeInPixels () * texArea;
 	}
+*/
 
 	const u32* getMipMap0_Area() const
 	{
 		return MipMap0_Area;
 	}
+	f32 get_lod_bias() const { return LodBIAS; }
 
 	//! returns unoptimized surface (misleading name. burning can scale down originalimage)
 	virtual CImage* getImage() const
@@ -106,9 +116,33 @@ public:
 		return TexBound[MipMapLOD];
 	}
 
+#if !defined(PATCH_SUPERTUX_8_0_1_with_1_9_0)
 	virtual void regenerateMipMapLevels(void* data = 0, u32 layer = 0) _IRR_OVERRIDE_;
+#else
+	virtual void regenerateMipMapLevels(void* data = 0);
+#endif
+
+
+#if defined(PATCH_SUPERTUX_8_0_1_with_1_9_0)
+	const core::dimension2d<u32>& getOriginalSize() const { return OriginalSize; };
+	const core::dimension2d<u32>& getSize() const { return Size; };
+	E_DRIVER_TYPE getDriverType() const { return DriverType; };
+	ECOLOR_FORMAT getColorFormat() const { return ColorFormat; };
+	ECOLOR_FORMAT getOriginalColorFormat() const { return OriginalColorFormat; };
+	u32 getPitch() const { return Pitch; };
+	bool hasMipMaps() const { return HasMipMaps; }
+	bool isRenderTarget() const { return IsRenderTarget; }
+
+	core::dimension2d<u32> OriginalSize;
+	core::dimension2d<u32> Size;
+	E_DRIVER_TYPE DriverType;
+	ECOLOR_FORMAT OriginalColorFormat;
+	ECOLOR_FORMAT ColorFormat;
+	u32 Pitch;
+	bool HasMipMaps;
+	bool IsRenderTarget;
+#endif
 
-	f32 get_lod_bias() const { return LodBIAS; }
 private:
 	void calcDerivative();
 
@@ -120,7 +154,7 @@ private:
 	CImage* MipMap[SOFTWARE_DRIVER_2_MIPMAPPING_MAX];
 	CSoftwareTexture2_Bound TexBound[SOFTWARE_DRIVER_2_MIPMAPPING_MAX];
 	u32 MipMap0_Area[2];
-	f32 LodBIAS;
+	f32 LodBIAS;	// Tweak mipmap selection
 };
 
 /*!
@@ -134,7 +168,10 @@ public:
 
 	virtual void setTexture(const core::array<ITexture*>& texture, ITexture* depthStencil, const core::array<E_CUBE_SURFACE>& cubeSurfaces) _IRR_OVERRIDE_;
 
-	ITexture* getTexture() const;
+#if defined(PATCH_SUPERTUX_8_0_1_with_1_9_0)
+	E_DRIVER_TYPE DriverType;
+	core::array<ITexture*> Texture;
+#endif
 
 protected:
 	CBurningVideoDriver* Driver;

source/Irrlicht/CTRGouraud2.cpp

@@ -87,7 +87,7 @@ public:
 	//virtual bool canWireFrame () { return true; }
 
 protected:
-	virtual void scanline_bilinear ();
+	virtual void fragmentShader();
 
 };
 
@@ -104,7 +104,7 @@ CTRGouraud2::CTRGouraud2(CBurningVideoDriver* driver)
 
 /*!
 */
-void CTRGouraud2::scanline_bilinear ()
+void CTRGouraud2::fragmentShader()
 {
 	tVideoSample *dst;
 
@@ -143,7 +143,7 @@ void CTRGouraud2::scanline_bilinear ()
 		return;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2( line.x[1] - line.x[0] );
+	const f32 invDeltaX = fill_step_x( line.x[1] - line.x[0] );
 
 #ifdef IPOL_Z
 	slopeZ = (line.z[1] - line.z[0]) * invDeltaX;
@@ -196,10 +196,11 @@ void CTRGouraud2::scanline_bilinear ()
 
 #endif
 
-	for ( s32 i = 0; i <= dx; ++i )
+	for ( s32 i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 		//if test active only first pixel
-		if ( (0 == EdgeTestPass) & i ) break;
+		if ((0 == EdgeTestPass) & (i > line.x_edgetest)) break;
+
 #ifdef CMP_Z
 		if ( line.z[0] < z[i] )
 #endif
@@ -384,8 +385,9 @@ void CTRGouraud2::drawTriangle(const s4DVertex* burning_restrict a, const s4DVer
 
 #endif
 
+		line.x_edgetest = fill_convention_edge(scan.slopeX[scan.left]);
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -416,7 +418,7 @@ void CTRGouraud2::drawTriangle(const s4DVertex* burning_restrict a, const s4DVer
 #endif
 
 			// render a scanline
-			scanline_bilinear ();
+			interlace_scanline fragmentShader();
 			if ( EdgeTestPass & edge_test_first_line ) break;
 
 			scan.x[0] += scan.slopeX[0];
@@ -511,7 +513,6 @@ void CTRGouraud2::drawTriangle(const s4DVertex* burning_restrict a, const s4DVer
 		yEnd = fill_convention_right( c->Pos.y );
 
 #ifdef SUBTEXEL
-
 		subPixel = ( (f32) yStart ) - b->Pos.y;
 
 		// correct to pixel center
@@ -545,8 +546,9 @@ void CTRGouraud2::drawTriangle(const s4DVertex* burning_restrict a, const s4DVer
 
 #endif
 
+		line.x_edgetest = fill_convention_edge(scan.slopeX[scan.left]);
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -577,7 +579,7 @@ void CTRGouraud2::drawTriangle(const s4DVertex* burning_restrict a, const s4DVer
 #endif
 
 			// render a scanline
-			scanline_bilinear ();
+			interlace_scanline fragmentShader();
 			if ( EdgeTestPass & edge_test_first_line ) break;
 
 			scan.x[0] += scan.slopeX[0];

source/Irrlicht/CTRGouraudAlpha2.cpp

@@ -88,7 +88,7 @@ public:
 
 
 private:
-	void scanline_bilinear ();
+	void fragmentShader();
 
 };
 
@@ -105,7 +105,7 @@ CTRGouraudAlpha2::CTRGouraudAlpha2(CBurningVideoDriver* driver)
 
 /*!
 */
-void CTRGouraudAlpha2::scanline_bilinear ()
+void CTRGouraudAlpha2::fragmentShader()
 {
 	tVideoSample *dst;
 
@@ -148,7 +148,7 @@ void CTRGouraudAlpha2::scanline_bilinear ()
 		return;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2( line.x[1] - line.x[0] );
+	const f32 invDeltaX = fill_step_x( line.x[1] - line.x[0] );
 
 #ifdef IPOL_Z
 	slopeZ = (line.z[1] - line.z[0]) * invDeltaX;
@@ -204,7 +204,7 @@ void CTRGouraudAlpha2::scanline_bilinear ()
 	tFixPoint r2, g2, b2;
 #endif
 
-	for ( s32 i = 0; i <= dx; ++i )
+	for ( s32 i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef CMP_Z
 		if ( line.z[0] < z[i] )
@@ -270,9 +270,9 @@ void CTRGouraudAlpha2::drawTriangle(const s4DVertex* burning_restrict a, const s
 	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] = reciprocal_zero( ca );
-	scan.invDeltaY[1] = reciprocal_zero( ba );
-	scan.invDeltaY[2] = reciprocal_zero( cb );
+	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;
@@ -397,7 +397,7 @@ void CTRGouraudAlpha2::drawTriangle(const s4DVertex* burning_restrict a, const s
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -428,7 +428,7 @@ void CTRGouraudAlpha2::drawTriangle(const s4DVertex* burning_restrict a, const s
 #endif
 
 			// render a scanline
-			scanline_bilinear ( );
+			interlace_scanline fragmentShader();
 
 			scan.x[0] += scan.slopeX[0];
 			scan.x[1] += scan.slopeX[1];
@@ -522,7 +522,6 @@ void CTRGouraudAlpha2::drawTriangle(const s4DVertex* burning_restrict a, const s
 		yEnd = fill_convention_right( c->Pos.y );
 
 #ifdef SUBTEXEL
-
 		subPixel = ( (f32) yStart ) - b->Pos.y;
 
 		// correct to pixel center
@@ -557,7 +556,7 @@ void CTRGouraudAlpha2::drawTriangle(const s4DVertex* burning_restrict a, const s
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -588,7 +587,7 @@ void CTRGouraudAlpha2::drawTriangle(const s4DVertex* burning_restrict a, const s
 #endif
 
 			// render a scanline
-			scanline_bilinear ( );
+			interlace_scanline fragmentShader();
 
 			scan.x[0] += scan.slopeX[0];
 			scan.x[1] += scan.slopeX[1];

source/Irrlicht/CTRGouraudAlphaNoZ2.cpp

@@ -88,7 +88,7 @@ public:
 
 
 private:
-	void scanline_bilinear ();
+	void fragmentShader();
 };
 
 //! constructor
@@ -104,7 +104,7 @@ CTRGouraudAlphaNoZ2::CTRGouraudAlphaNoZ2(CBurningVideoDriver* driver)
 
 /*!
 */
-void CTRGouraudAlphaNoZ2::scanline_bilinear ()
+void CTRGouraudAlphaNoZ2::fragmentShader()
 {
 	tVideoSample *dst;
 
@@ -144,7 +144,7 @@ void CTRGouraudAlphaNoZ2::scanline_bilinear ()
 		return;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2( line.x[1] - line.x[0] );
+	const f32 invDeltaX = fill_step_x( line.x[1] - line.x[0] );
 
 #ifdef IPOL_Z
 	slopeZ = (line.z[1] - line.z[0]) * invDeltaX;
@@ -201,7 +201,7 @@ void CTRGouraudAlphaNoZ2::scanline_bilinear ()
 	tFixPoint r2, g2, b2;
 #endif
 
-	for ( s32 i = 0; i <= dx; ++i )
+	for ( s32 i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef CMP_Z
 		if ( line.z[0] < z[i] )
@@ -271,9 +271,9 @@ void CTRGouraudAlphaNoZ2::drawTriangle(const s4DVertex* burning_restrict a, cons
 	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] = reciprocal_zero( ca );
-	scan.invDeltaY[1] = reciprocal_zero( ba );
-	scan.invDeltaY[2] = reciprocal_zero( cb );
+	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;
@@ -397,7 +397,7 @@ void CTRGouraudAlphaNoZ2::drawTriangle(const s4DVertex* burning_restrict a, cons
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -428,7 +428,7 @@ void CTRGouraudAlphaNoZ2::drawTriangle(const s4DVertex* burning_restrict a, cons
 #endif
 
 			// render a scanline
-			scanline_bilinear ();
+			interlace_scanline fragmentShader();
 
 			scan.x[0] += scan.slopeX[0];
 			scan.x[1] += scan.slopeX[1];
@@ -557,7 +557,7 @@ void CTRGouraudAlphaNoZ2::drawTriangle(const s4DVertex* burning_restrict a, cons
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -588,7 +588,7 @@ void CTRGouraudAlphaNoZ2::drawTriangle(const s4DVertex* burning_restrict a, cons
 #endif
 
 			// render a scanline
-			scanline_bilinear ();
+			interlace_scanline fragmentShader();
 
 			scan.x[0] += scan.slopeX[0];
 			scan.x[1] += scan.slopeX[1];

source/Irrlicht/CTRGouraudNoZ2.cpp

@@ -87,7 +87,7 @@ public:
 	virtual bool canWireFrame () { return true; }
 
 protected:
-	virtual void scanline_bilinear ();
+	virtual void fragmentShader();
 };
 
 //! constructor
@@ -103,7 +103,7 @@ CTRGouraudNoZ2::CTRGouraudNoZ2(CBurningVideoDriver* driver)
 
 /*!
 */
-void CTRGouraudNoZ2::scanline_bilinear ()
+void CTRGouraudNoZ2::fragmentShader()
 {
 	tVideoSample *dst;
 
@@ -142,7 +142,7 @@ void CTRGouraudNoZ2::scanline_bilinear ()
 		return;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2( line.x[1] - line.x[0] );
+	const f32 invDeltaX = fill_step_x( line.x[1] - line.x[0] );
 
 #ifdef IPOL_Z
 	slopeZ = (line.z[1] - line.z[0]) * invDeltaX;
@@ -195,9 +195,9 @@ void CTRGouraudNoZ2::scanline_bilinear ()
 
 #endif
 
-	for ( s32 i = 0; i <= dx; ++i )
+	for ( s32 i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
-		if ( (0 == EdgeTestPass) & i ) break;
+		if ((0 == EdgeTestPass) & (i > line.x_edgetest)) break;
 
 #ifdef CMP_Z
 		if ( line.z[0] < z[i] )
@@ -383,8 +383,10 @@ void CTRGouraudNoZ2::drawTriangle(const s4DVertex* burning_restrict a, const s4D
 
 #endif
 
+		line.x_edgetest = fill_convention_edge(scan.slopeX[scan.left]);
+
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -415,7 +417,7 @@ void CTRGouraudNoZ2::drawTriangle(const s4DVertex* burning_restrict a, const s4D
 #endif
 
 			// render a scanline
-			scanline_bilinear ();
+			interlace_scanline fragmentShader();
 			if ( EdgeTestPass & edge_test_first_line ) break;
 
 			scan.x[0] += scan.slopeX[0];
@@ -509,7 +511,6 @@ void CTRGouraudNoZ2::drawTriangle(const s4DVertex* burning_restrict a, const s4D
 		yEnd = fill_convention_right( c->Pos.y );
 
 #ifdef SUBTEXEL
-
 		subPixel = ( (f32) yStart ) - b->Pos.y;
 
 		// correct to pixel center
@@ -542,9 +543,10 @@ void CTRGouraudNoZ2::drawTriangle(const s4DVertex* burning_restrict a, const s4D
 #endif
 
 #endif
+		line.x_edgetest = fill_convention_edge(scan.slopeX[scan.left]);
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -575,7 +577,7 @@ void CTRGouraudNoZ2::drawTriangle(const s4DVertex* burning_restrict a, const s4D
 #endif
 
 			// render a scanline
-			scanline_bilinear ();
+			interlace_scanline fragmentShader();
 			if ( EdgeTestPass & edge_test_first_line ) break;
 
 			scan.x[0] += scan.slopeX[0];

source/Irrlicht/CTRNormalMap.cpp

@@ -164,7 +164,7 @@ void CTRNormalMap::fragmentShader()
 		return;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2( line.x[1] - line.x[0] );
+	const f32 invDeltaX = fill_step_x( line.x[1] - line.x[0] );
 
 #ifdef IPOL_Z
 	slopeZ = (line.z[1] - line.z[0]) * invDeltaX;
@@ -254,7 +254,7 @@ void CTRNormalMap::fragmentShader()
 #endif
 
 
-	for ( s32 i = 0; i <= dx; i++ )
+	for ( s32 i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef CMP_Z
 		if ( line.z[0] < z[i] )
@@ -400,9 +400,9 @@ void CTRNormalMap::drawTriangle(const s4DVertex* burning_restrict a, const s4DVe
 	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] = reciprocal_zero( ca );
-	scan.invDeltaY[1] = reciprocal_zero( ba );
-	scan.invDeltaY[2] = reciprocal_zero( cb );
+	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;
@@ -572,7 +572,7 @@ void CTRNormalMap::drawTriangle(const s4DVertex* burning_restrict a, const s4DVe
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -618,7 +618,7 @@ void CTRNormalMap::drawTriangle(const s4DVertex* burning_restrict a, const s4DVe
 #endif
 
 			// render a scanline
-			fragmentShader ();
+			interlace_scanline fragmentShader ();
 
 			scan.x[0] += scan.slopeX[0];
 			scan.x[1] += scan.slopeX[1];
@@ -751,7 +751,6 @@ void CTRNormalMap::drawTriangle(const s4DVertex* burning_restrict a, const s4DVe
 		yEnd = fill_convention_right( c->Pos.y );
 
 #ifdef SUBTEXEL
-
 		subPixel = ( (f32) yStart ) - b->Pos.y;
 
 		// correct to pixel center
@@ -801,7 +800,7 @@ void CTRNormalMap::drawTriangle(const s4DVertex* burning_restrict a, const s4DVe
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -847,7 +846,7 @@ void CTRNormalMap::drawTriangle(const s4DVertex* burning_restrict a, const s4DVe
 #endif
 
 			// render a scanline
-			fragmentShader();
+			interlace_scanline fragmentShader();
 
 			scan.x[0] += scan.slopeX[0];
 			scan.x[1] += scan.slopeX[1];

source/Irrlicht/CTRStencilShadow.cpp

@@ -139,7 +139,7 @@ void CTRStencilShadow::fragmentShader()
 	SOFTWARE_DRIVER_2_CLIPCHECK;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2( line.x[1] - line.x[0] );
+	const f32 invDeltaX = fill_step_x( line.x[1] - line.x[0] );
 
 #ifdef IPOL_Z
 	slopeZ = (line.z[1] - line.z[0]) * invDeltaX;
@@ -174,7 +174,7 @@ void CTRStencilShadow::fragmentShader()
 #endif
 
 	s32 i;
-	for (i = 0; i <= dx; i++)
+	for (i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef CMP_Z
 		if (line.z[0] < z[i])
@@ -226,9 +226,9 @@ void CTRStencilShadow::drawTriangle(const s4DVertex* burning_restrict a, const s
 	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] = reciprocal_zero( ca );
-	scan.invDeltaY[1] = reciprocal_zero( ba );
-	scan.invDeltaY[2] = reciprocal_zero( cb );
+	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;
@@ -382,7 +382,7 @@ void CTRStencilShadow::drawTriangle(const s4DVertex* burning_restrict a, const s
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -423,7 +423,7 @@ void CTRStencilShadow::drawTriangle(const s4DVertex* burning_restrict a, const s
 #endif
 
 			// render a scanline
-			fragmentShader ();
+			interlace_scanline fragmentShader ();
 
 			scan.x[0] += scan.slopeX[0];
 			scan.x[1] += scan.slopeX[1];
@@ -545,7 +545,6 @@ void CTRStencilShadow::drawTriangle(const s4DVertex* burning_restrict a, const s
 		yEnd = fill_convention_right( c->Pos.y );
 
 #ifdef SUBTEXEL
-
 		subPixel = ( (f32) yStart ) - b->Pos.y;
 
 		// correct to pixel center
@@ -590,7 +589,7 @@ void CTRStencilShadow::drawTriangle(const s4DVertex* burning_restrict a, const s
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -631,7 +630,7 @@ void CTRStencilShadow::drawTriangle(const s4DVertex* burning_restrict a, const s
 #endif
 
 			// render a scanline
-			fragmentShader ();
+			interlace_scanline fragmentShader ();
 
 			scan.x[0] += scan.slopeX[0];
 			scan.x[1] += scan.slopeX[1];

source/Irrlicht/CTRTextureBlend.cpp

@@ -87,22 +87,6 @@ namespace video
 		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_;
 
-#if defined(PATCH_SUPERTUX_8_0_1)
-
-		virtual void setZCompareFunc(u32 func)
-		{
-			depth_func = (E_COMPARISON_FUNC)func;
-		}
-		virtual void setParam(u32 index, f32 value)
-		{
-			SBurningShaderMaterial material;
-			material.org.ZBuffer = depth_func;
-			material.org.MaterialTypeParam = value;
-			OnSetMaterial(material);
-	}
-
-#endif
-
 private:
 	// fragment shader
 	typedef void (CTRTextureBlend::*tFragmentShader) ();
@@ -269,7 +253,7 @@ void CTRTextureBlend::fragment_dst_color_src_alpha ()
 		return;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2( line.x[1] - line.x[0] );
+	const f32 invDeltaX = fill_step_x( line.x[1] - line.x[0] );
 
 #ifdef IPOL_Z
 	slopeZ = (line.z[1] - line.z[0]) * invDeltaX;
@@ -325,7 +309,7 @@ void CTRTextureBlend::fragment_dst_color_src_alpha ()
 	{
 	default:
 	case ECFN_LESSEQUAL:
-	for ( i = 0; i <= dx; ++i )
+	for ( i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef CMP_W
 		if ( line.w[0] >= z[i] )
@@ -368,7 +352,7 @@ void CTRTextureBlend::fragment_dst_color_src_alpha ()
 	break;
 
 	case 2:
-	for ( i = 0; i <= dx; ++i )
+	for ( i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef CMP_W
 		if ( line.w[0] == z[i] )
@@ -454,7 +438,7 @@ void CTRTextureBlend::fragment_src_color_src_alpha ()
 		return;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2( line.x[1] - line.x[0] );
+	const f32 invDeltaX = fill_step_x( line.x[1] - line.x[0] );
 
 #ifdef IPOL_Z
 	slopeZ = (line.z[1] - line.z[0]) * invDeltaX;
@@ -514,7 +498,7 @@ void CTRTextureBlend::fragment_src_color_src_alpha ()
 	{
 	default:
 	case ECFN_LESSEQUAL:
-	for ( i = 0; i <= dx; ++i )
+	for ( i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef CMP_W
 		if ( line.w[0] >= z[i] )
@@ -561,7 +545,7 @@ void CTRTextureBlend::fragment_src_color_src_alpha ()
 	break;
 
 	case 2:
-	for ( i = 0; i <= dx; ++i )
+	for ( i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef CMP_W
 		if ( line.w[0] == z[i] )
@@ -649,7 +633,7 @@ void CTRTextureBlend::fragment_one_one_minus_src_alpha()
 		return;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2 ( line.x[1] - line.x[0] );
+	const f32 invDeltaX = fill_step_x( line.x[1] - line.x[0] );
 
 #ifdef IPOL_Z
 	slopeZ = (line.z[1] - line.z[0]) * invDeltaX;
@@ -707,7 +691,7 @@ void CTRTextureBlend::fragment_one_one_minus_src_alpha()
 	{
 	default:
 	case ECFN_LESSEQUAL:
-	for ( i = 0; i <= dx; ++i )
+	for ( i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef CMP_W
 		if ( line.w[0] >= z[i] )
@@ -757,7 +741,7 @@ void CTRTextureBlend::fragment_one_one_minus_src_alpha()
 	break;
 
 	case 2:
-	for ( i = 0; i <= dx; ++i )
+	for ( i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef CMP_W
 		if ( line.w[0] == z[i] )
@@ -850,7 +834,7 @@ void CTRTextureBlend::fragment_one_minus_dst_alpha_one ()
 		return;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2( line.x[1] - line.x[0] );
+	const f32 invDeltaX = fill_step_x( line.x[1] - line.x[0] );
 
 #ifdef IPOL_Z
 	slopeZ = (line.z[1] - line.z[0]) * invDeltaX;
@@ -908,7 +892,7 @@ void CTRTextureBlend::fragment_one_minus_dst_alpha_one ()
 	{
 	default:
 	case ECFN_LESSEQUAL:
-	for ( i = 0; i <= dx; ++i )
+	for ( i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef CMP_W
 		if ( line.w[0] >= z[i] )
@@ -957,7 +941,7 @@ void CTRTextureBlend::fragment_one_minus_dst_alpha_one ()
 	break;
 
 	case 2:
-	for ( i = 0; i <= dx; ++i )
+	for ( i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef CMP_W
 		if ( line.w[0] == z[i] )
@@ -1050,7 +1034,7 @@ void CTRTextureBlend::fragment_src_alpha_one ()
 		return;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2( line.x[1] - line.x[0] );
+	const f32 invDeltaX = fill_step_x( line.x[1] - line.x[0] );
 
 #ifdef IPOL_Z
 	slopeZ = (line.z[1] - line.z[0]) * invDeltaX;
@@ -1108,7 +1092,7 @@ void CTRTextureBlend::fragment_src_alpha_one ()
 	{
 	default:
 	case ECFN_LESSEQUAL:
-	for ( i = 0; i <= dx; ++i )
+	for ( i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef CMP_W
 		if ( line.w[0] >= z[i] )
@@ -1174,7 +1158,7 @@ void CTRTextureBlend::fragment_src_alpha_one ()
 	break;
 
 	case 2:
-	for ( i = 0; i <= dx; ++i )
+	for ( i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef CMP_W
 		if ( line.w[0] == z[i] )
@@ -1280,7 +1264,7 @@ void CTRTextureBlend::fragment_src_alpha_one_minus_src_alpha()
 		return;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2(line.x[1] - line.x[0]);
+	const f32 invDeltaX = fill_step_x(line.x[1] - line.x[0]);
 
 #ifdef IPOL_Z
 	slopeZ = (line.z[1] - line.z[0]) * invDeltaX;
@@ -1337,7 +1321,7 @@ void CTRTextureBlend::fragment_src_alpha_one_minus_src_alpha()
 	{
 	default:
 	case ECFN_LESSEQUAL:
-		for (i = 0; i <= dx; ++i)
+		for (i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 		{
 #ifdef CMP_W
 			if (line.w[0] >= z[i])
@@ -1434,7 +1418,7 @@ void CTRTextureBlend::fragment_dst_color_one_minus_dst_alpha ()
 		return;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2( line.x[1] - line.x[0] );
+	const f32 invDeltaX = fill_step_x( line.x[1] - line.x[0] );
 
 #ifdef IPOL_Z
 	slopeZ = (line.z[1] - line.z[0]) * invDeltaX;
@@ -1492,7 +1476,7 @@ void CTRTextureBlend::fragment_dst_color_one_minus_dst_alpha ()
 	{
 	default:
 	case ECFN_LESSEQUAL:
-	for ( i = 0; i <= dx; ++i )
+	for ( i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef CMP_W
 		if ( line.w[0] >= z[i] )
@@ -1541,7 +1525,7 @@ void CTRTextureBlend::fragment_dst_color_one_minus_dst_alpha ()
 	break;
 
 	case 2:
-	for ( i = 0; i <= dx; ++i )
+	for ( i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef CMP_W
 		if ( line.w[0] == z[i] )
@@ -1633,7 +1617,7 @@ void CTRTextureBlend::fragment_dst_color_zero ()
 		return;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2( line.x[1] - line.x[0] );
+	const f32 invDeltaX = fill_step_x( line.x[1] - line.x[0] );
 
 #ifdef IPOL_Z
 	slopeZ = (line.z[1] - line.z[0]) * invDeltaX;
@@ -1691,7 +1675,7 @@ void CTRTextureBlend::fragment_dst_color_zero ()
 	{
 	default:
 	case ECFN_LESSEQUAL:
-	for ( i = 0; i <= dx; ++i )
+	for ( i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef CMP_W
 		if ( line.w[0] >= z[i] )
@@ -1739,7 +1723,7 @@ void CTRTextureBlend::fragment_dst_color_zero ()
 	break;
 
 	case 2:
-	for ( i = 0; i <= dx; ++i )
+	for ( i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef CMP_W
 		if ( line.w[0] == z[i] )
@@ -1830,7 +1814,7 @@ void CTRTextureBlend::fragment_dst_color_one ()
 		return;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2( line.x[1] - line.x[0] );
+	const f32 invDeltaX = fill_step_x( line.x[1] - line.x[0] );
 
 #ifdef IPOL_Z
 	slopeZ = (line.z[1] - line.z[0]) * invDeltaX;
@@ -1888,7 +1872,7 @@ void CTRTextureBlend::fragment_dst_color_one ()
 	{
 	default:
 	case ECFN_LESSEQUAL:
-	for ( i = 0; i <= dx; ++i )
+	for ( i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef CMP_W
 		if ( line.w[0] >= z[i] )
@@ -1937,7 +1921,7 @@ void CTRTextureBlend::fragment_dst_color_one ()
 	break;
 
 	case 2:
-	for ( i = 0; i <= dx; ++i )
+	for ( i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef CMP_W
 		if ( line.w[0] == z[i] )
@@ -2030,7 +2014,7 @@ void CTRTextureBlend::fragment_zero_one_minus_scr_color ()
 		return;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2( line.x[1] - line.x[0] );
+	const f32 invDeltaX = fill_step_x( line.x[1] - line.x[0] );
 
 #ifdef IPOL_Z
 	slopeZ = (line.z[1] - line.z[0]) * invDeltaX;
@@ -2088,7 +2072,7 @@ void CTRTextureBlend::fragment_zero_one_minus_scr_color ()
 	{
 	default:
 	case ECFN_LESSEQUAL:
-	for ( i = 0; i <= dx; ++i )
+	for ( i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef CMP_W
 		if ( line.w[0] >= z[i] )
@@ -2137,7 +2121,7 @@ void CTRTextureBlend::fragment_zero_one_minus_scr_color ()
 	break;
 
 	case 2:
-	for ( i = 0; i <= dx; ++i )
+	for ( i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef CMP_W
 		if ( line.w[0] == z[i] )
@@ -2202,9 +2186,9 @@ void CTRTextureBlend::drawTriangle(const s4DVertex* burning_restrict a, const s4
 	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] = reciprocal_zero( ca );
-	scan.invDeltaY[1] = reciprocal_zero( ba );
-	scan.invDeltaY[2] = reciprocal_zero( cb );
+	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;
@@ -2328,7 +2312,7 @@ void CTRTextureBlend::drawTriangle(const s4DVertex* burning_restrict a, const s4
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -2359,7 +2343,7 @@ void CTRTextureBlend::drawTriangle(const s4DVertex* burning_restrict a, const s4
 #endif
 
 			// render a scanline
-			(this->*fragmentShader) ();
+			interlace_scanline (this->*fragmentShader) ();
 
 			scan.x[0] += scan.slopeX[0];
 			scan.x[1] += scan.slopeX[1];
@@ -2453,7 +2437,6 @@ void CTRTextureBlend::drawTriangle(const s4DVertex* burning_restrict a, const s4
 		yEnd = fill_convention_right( c->Pos.y );
 
 #ifdef SUBTEXEL
-
 		subPixel = ( (f32) yStart ) - b->Pos.y;
 
 		// correct to pixel center
@@ -2488,7 +2471,7 @@ void CTRTextureBlend::drawTriangle(const s4DVertex* burning_restrict a, const s4
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -2519,7 +2502,7 @@ void CTRTextureBlend::drawTriangle(const s4DVertex* burning_restrict a, const s4
 #endif
 
 			// render a scanline
-			(this->*fragmentShader) ();
+			interlace_scanline (this->*fragmentShader) ();
 
 			scan.x[0] += scan.slopeX[0];
 			scan.x[1] += scan.slopeX[1];

source/Irrlicht/CTRTextureDetailMap2.cpp

@@ -88,7 +88,7 @@ public:
 	virtual bool canWireFrame () { return true; }
 
 protected:
-	virtual void scanline_bilinear ();
+	virtual void fragmentShader();
 
 };
 
@@ -105,7 +105,7 @@ CTRTextureDetailMap2::CTRTextureDetailMap2(CBurningVideoDriver* driver)
 
 /*!
 */
-void CTRTextureDetailMap2::scanline_bilinear ()
+void CTRTextureDetailMap2::fragmentShader()
 {
 	tVideoSample *dst;
 
@@ -140,12 +140,11 @@ void CTRTextureDetailMap2::scanline_bilinear ()
 	xEnd = fill_convention_right( line.x[1] );
 
 	dx = xEnd - xStart;
-
 	if ( dx < 0 )
 		return;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2( line.x[1] - line.x[0] );
+	const f32 invDeltaX = fill_step_x( line.x[1] - line.x[0] );
 
 #ifdef IPOL_Z
 	slopeZ = (line.z[1] - line.z[0]) * invDeltaX;
@@ -200,9 +199,9 @@ void CTRTextureDetailMap2::scanline_bilinear ()
 	tFixPoint r2, g2, b2;
 
 
-	for ( s32 i = 0; i <= dx; ++i )
+	for ( s32 i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
-		if ( (0 == EdgeTestPass) & i ) break;
+		if ( (0 == EdgeTestPass) & (i > line.x_edgetest)) break;
 
 #ifdef CMP_Z
 		if ( line.z[0] < z[i] )
@@ -326,7 +325,7 @@ void CTRTextureDetailMap2::drawTriangle(const s4DVertex* burning_restrict a, con
 #endif
 
 	// rasterize upper sub-triangle
-	if ( (f32) 0.0 != scan.invDeltaY[1]  )
+	if (F32_GREATER_0(scan.invDeltaY[1])  )
 	{
 		// calculate slopes for top edge
 		scan.slopeX[1] = (b->Pos.x - a->Pos.x) * scan.invDeltaY[1];
@@ -394,9 +393,10 @@ void CTRTextureDetailMap2::drawTriangle(const s4DVertex* burning_restrict a, con
 #endif
 
 #endif
+		line.x_edgetest = fill_convention_edge(scan.slopeX[scan.left]);
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -427,7 +427,8 @@ void CTRTextureDetailMap2::drawTriangle(const s4DVertex* burning_restrict a, con
 #endif
 
 			// render a scanline
-			scanline_bilinear ();
+			interlace_scanline fragmentShader();
+			if (EdgeTestPass & edge_test_first_line) break;
 
 			scan.x[0] += scan.slopeX[0];
 			scan.x[1] += scan.slopeX[1];
@@ -461,10 +462,10 @@ void CTRTextureDetailMap2::drawTriangle(const s4DVertex* burning_restrict a, con
 	}
 
 	// rasterize lower sub-triangle
-	if ( (f32) 0.0 != scan.invDeltaY[2] )
+	if (F32_GREATER_0(scan.invDeltaY[2]) )
 	{
 		// advance to middle point
-		if( (f32) 0.0 != scan.invDeltaY[1] )
+		if(F32_GREATER_0(scan.invDeltaY[1]) )
 		{
 			temp[0] = b->Pos.y - a->Pos.y;	// dy
 
@@ -521,7 +522,6 @@ void CTRTextureDetailMap2::drawTriangle(const s4DVertex* burning_restrict a, con
 		yEnd = fill_convention_right( c->Pos.y );
 
 #ifdef SUBTEXEL
-
 		subPixel = ( (f32) yStart ) - b->Pos.y;
 
 		// correct to pixel center
@@ -555,8 +555,10 @@ void CTRTextureDetailMap2::drawTriangle(const s4DVertex* burning_restrict a, con
 
 #endif
 
+		line.x_edgetest = fill_convention_edge(scan.slopeX[scan.left]);
+
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -587,7 +589,8 @@ void CTRTextureDetailMap2::drawTriangle(const s4DVertex* burning_restrict a, con
 #endif
 
 			// render a scanline
-			scanline_bilinear ();
+			interlace_scanline fragmentShader();
+			if (EdgeTestPass & edge_test_first_line) break;
 
 			scan.x[0] += scan.slopeX[0];
 			scan.x[1] += scan.slopeX[1];

source/Irrlicht/CTRTextureGouraud2.cpp

@@ -167,7 +167,7 @@ void CTRTextureGouraud2::fragmentShader ()
 		return;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2( line.x[1] - line.x[0] );
+	const f32 invDeltaX = fill_step_x( line.x[1] - line.x[0] );
 
 #ifdef IPOL_Z
 	slopeZ = (line.z[1] - line.z[0]) * invDeltaX;
@@ -254,10 +254,10 @@ void CTRTextureGouraud2::fragmentShader ()
 	u32 dIndex = ( line.y & 3 ) << 2;
 #endif
 
-	for ( s32 i = 0; i <= dx; ++i )
+	for ( s32 i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 		//if test active only first pixel
-		if ( (0 == EdgeTestPass) & i ) break;
+		if ((0 == EdgeTestPass) & (i > line.x_edgetest)) break;
 
 #ifdef CMP_Z
 		if ( line.z[0] < z[i] )
@@ -382,9 +382,9 @@ void CTRTextureGouraud2::drawTriangle(const s4DVertex* burning_restrict a, const
 	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] = reciprocal_zero( ca );
-	scan.invDeltaY[1] = reciprocal_zero( ba );
-	scan.invDeltaY[2] = reciprocal_zero( cb );
+	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;
@@ -551,9 +551,10 @@ void CTRTextureGouraud2::drawTriangle(const s4DVertex* burning_restrict a, const
 #endif
 
 #endif
+		line.x_edgetest = fill_convention_edge(scan.slopeX[scan.left]);
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -599,7 +600,7 @@ void CTRTextureGouraud2::drawTriangle(const s4DVertex* burning_restrict a, const
 #endif
 
 			// render a scanline
-			fragmentShader ();
+			interlace_scanline fragmentShader ();
 			if ( EdgeTestPass & edge_test_first_line ) break;
 
 
@@ -734,7 +735,6 @@ void CTRTextureGouraud2::drawTriangle(const s4DVertex* burning_restrict a, const
 		yEnd = fill_convention_right( c->Pos.y );
 
 #ifdef SUBTEXEL
-
 		subPixel = ( (f32) yStart ) - b->Pos.y;
 
 		// correct to pixel center
@@ -761,7 +761,7 @@ void CTRTextureGouraud2::drawTriangle(const s4DVertex* burning_restrict a, const
 		scan.c[1][1] += scan.slopeC[1][1] * subPixel;
 #endif
 
-#ifdef IPOL_C1
+#ifdef IPOL_C2
 		scan.c[2][0] += scan.slopeC[2][0] * subPixel;
 		scan.c[2][1] += scan.slopeC[2][1] * subPixel;
 #endif
@@ -782,9 +782,10 @@ void CTRTextureGouraud2::drawTriangle(const s4DVertex* burning_restrict a, const
 #endif
 
 #endif
+		line.x_edgetest = fill_convention_edge(scan.slopeX[scan.left]);
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -830,7 +831,7 @@ void CTRTextureGouraud2::drawTriangle(const s4DVertex* burning_restrict a, const
 #endif
 
 			// render a scanline
-			fragmentShader ();
+			interlace_scanline fragmentShader ();
 			if ( EdgeTestPass & edge_test_first_line ) break;
 
 

source/Irrlicht/CTRTextureGouraudAdd2.cpp

@@ -88,7 +88,7 @@ public:
 
 
 private:
-	void scanline_bilinear ();
+	void fragmentShader();
 };
 
 //! constructor
@@ -104,7 +104,7 @@ CTRTextureGouraudAdd2::CTRTextureGouraudAdd2(CBurningVideoDriver* driver)
 
 /*!
 */
-void CTRTextureGouraudAdd2::scanline_bilinear ()
+void CTRTextureGouraudAdd2::fragmentShader()
 {
 	tVideoSample *dst;
 
@@ -144,7 +144,7 @@ void CTRTextureGouraudAdd2::scanline_bilinear ()
 		return;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2( line.x[1] - line.x[0] );
+	const f32 invDeltaX = fill_step_x( line.x[1] - line.x[0] );
 
 #ifdef IPOL_Z
 	slopeZ = (line.z[1] - line.z[0]) * invDeltaX;
@@ -203,7 +203,7 @@ void CTRTextureGouraudAdd2::scanline_bilinear ()
 #endif
 
 
-	for ( s32 i = 0; i <= dx; ++i )
+	for ( s32 i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef CMP_Z
 		if ( line.z[0] < z[i] )
@@ -280,9 +280,9 @@ void CTRTextureGouraudAdd2::drawTriangle(const s4DVertex* burning_restrict a, co
 	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] = reciprocal_zero( ca );
-	scan.invDeltaY[1] = reciprocal_zero( ba );
-	scan.invDeltaY[2] = reciprocal_zero( cb );
+	scan.invDeltaY[0] = fill_step_y( ca );
+	scan.invDeltaY[1] = fill_step_y( ba );
+	scan.invDeltaY[2] = fill_step_y( cb );
 
 	// find if the major edge is left or right aligned
 	f32 temp[4];
@@ -403,7 +403,7 @@ void CTRTextureGouraudAdd2::drawTriangle(const s4DVertex* burning_restrict a, co
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -434,7 +434,7 @@ void CTRTextureGouraudAdd2::drawTriangle(const s4DVertex* burning_restrict a, co
 #endif
 
 			// render a scanline
-			scanline_bilinear ();
+			interlace_scanline fragmentShader();
 
 			scan.x[0] += scan.slopeX[0];
 			scan.x[1] += scan.slopeX[1];
@@ -563,7 +563,7 @@ void CTRTextureGouraudAdd2::drawTriangle(const s4DVertex* burning_restrict a, co
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -594,7 +594,7 @@ void CTRTextureGouraudAdd2::drawTriangle(const s4DVertex* burning_restrict a, co
 #endif
 
 			// render a scanline
-			scanline_bilinear ();
+			interlace_scanline fragmentShader();
 
 			scan.x[0] += scan.slopeX[0];
 			scan.x[1] += scan.slopeX[1];

source/Irrlicht/CTRTextureGouraudAddNoZ2.cpp

@@ -143,7 +143,7 @@ void CTRTextureGouraudAddNoZ2::fragmentShader()
 		return;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2( line.x[1] - line.x[0] );
+	const f32 invDeltaX = fill_step_x( line.x[1] - line.x[0] );
 
 #ifdef IPOL_Z
 	slopeZ = (line.z[1] - line.z[0]) * invDeltaX;
@@ -200,7 +200,7 @@ void CTRTextureGouraudAddNoZ2::fragmentShader()
 	tFixPoint r2, g2, b2;
 #endif
 
-	for ( s32 i = 0; i <= dx; ++i )
+	for ( s32 i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef CMP_Z
 		if ( line.z[0] < z[i] )
@@ -275,9 +275,9 @@ void CTRTextureGouraudAddNoZ2::drawTriangle(const s4DVertex* burning_restrict a,
 	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] = reciprocal_zero( ca );
-	scan.invDeltaY[1] = reciprocal_zero( ba );
-	scan.invDeltaY[2] = reciprocal_zero( cb );
+	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;
@@ -401,7 +401,7 @@ void CTRTextureGouraudAddNoZ2::drawTriangle(const s4DVertex* burning_restrict a,
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -432,7 +432,7 @@ void CTRTextureGouraudAddNoZ2::drawTriangle(const s4DVertex* burning_restrict a,
 #endif
 
 			// render a scanline
-			fragmentShader();
+			interlace_scanline fragmentShader();
 
 			scan.x[0] += scan.slopeX[0];
 			scan.x[1] += scan.slopeX[1];
@@ -526,7 +526,6 @@ void CTRTextureGouraudAddNoZ2::drawTriangle(const s4DVertex* burning_restrict a,
 		yEnd = fill_convention_right( c->Pos.y );
 
 #ifdef SUBTEXEL
-
 		subPixel = ( (f32) yStart ) - b->Pos.y;
 
 		// correct to pixel center
@@ -561,7 +560,7 @@ void CTRTextureGouraudAddNoZ2::drawTriangle(const s4DVertex* burning_restrict a,
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -592,7 +591,7 @@ void CTRTextureGouraudAddNoZ2::drawTriangle(const s4DVertex* burning_restrict a,
 #endif
 
 			// render a scanline
-			fragmentShader( );
+			interlace_scanline fragmentShader();
 
 			scan.x[0] += scan.slopeX[0];
 			scan.x[1] += scan.slopeX[1];

source/Irrlicht/CTRTextureGouraudAlpha.cpp

@@ -88,7 +88,7 @@ public:
 	virtual void OnSetMaterial(const SBurningShaderMaterial& material) _IRR_OVERRIDE_;
 
 private:
-	void scanline_bilinear ();
+	void fragmentShader();
 
 };
 
@@ -115,7 +115,7 @@ void CTRTextureGouraudAlpha2::OnSetMaterial(const SBurningShaderMaterial& materi
 
 /*!
 */
-void CTRTextureGouraudAlpha2::scanline_bilinear ()
+void CTRTextureGouraudAlpha2::fragmentShader()
 {
 	tVideoSample *dst;
 
@@ -155,7 +155,7 @@ void CTRTextureGouraudAlpha2::scanline_bilinear ()
 		return;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2( line.x[1] - line.x[0] );
+	const f32 invDeltaX = fill_step_x( line.x[1] - line.x[0] );
 
 #ifdef IPOL_Z
 	slopeZ = (line.z[1] - line.z[0]) * invDeltaX;
@@ -213,7 +213,7 @@ void CTRTextureGouraudAlpha2::scanline_bilinear ()
 #endif
 #endif
 
-	for ( s32 i = 0; i <= dx; ++i )
+	for ( s32 i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef CMP_Z
 		if ( line.z[0] < z[i] )
@@ -337,9 +337,9 @@ void CTRTextureGouraudAlpha2::drawTriangle(const s4DVertex* burning_restrict a,
 	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] = reciprocal_zero( ca );
-	scan.invDeltaY[1] = reciprocal_zero( ba );
-	scan.invDeltaY[2] = reciprocal_zero( cb );
+	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;
@@ -463,7 +463,7 @@ void CTRTextureGouraudAlpha2::drawTriangle(const s4DVertex* burning_restrict a,
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -494,7 +494,7 @@ void CTRTextureGouraudAlpha2::drawTriangle(const s4DVertex* burning_restrict a,
 #endif
 
 			// render a scanline
-			scanline_bilinear ( );
+			interlace_scanline fragmentShader();
 
 			scan.x[0] += scan.slopeX[0];
 			scan.x[1] += scan.slopeX[1];
@@ -588,7 +588,6 @@ void CTRTextureGouraudAlpha2::drawTriangle(const s4DVertex* burning_restrict a,
 		yEnd = fill_convention_right( c->Pos.y );
 
 #ifdef SUBTEXEL
-
 		subPixel = ( (f32) yStart ) - b->Pos.y;
 
 		// correct to pixel center
@@ -623,7 +622,7 @@ void CTRTextureGouraudAlpha2::drawTriangle(const s4DVertex* burning_restrict a,
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -654,7 +653,7 @@ void CTRTextureGouraudAlpha2::drawTriangle(const s4DVertex* burning_restrict a,
 #endif
 
 			// render a scanline
-			scanline_bilinear ( );
+			interlace_scanline fragmentShader();
 
 			scan.x[0] += scan.slopeX[0];
 			scan.x[1] += scan.slopeX[1];

source/Irrlicht/CTRTextureGouraudAlphaNoZ.cpp

@@ -93,20 +93,6 @@ public:
 	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_;
 
-#if defined(PATCH_SUPERTUX_8_0_1)
-	SBurningShaderMaterial Material;
-
-	virtual void setMaterial(const SBurningShaderMaterial &material)
-	{
-		Material = material;
-	}
-
-	virtual void setParam(u32 index, f32 value)
-	{
-		OnSetMaterial(Material);
-}
-#endif
-
 private:
 
 	// fragment shader
@@ -198,7 +184,7 @@ void CTRTextureGouraudAlphaNoZ::fragment_linear()
 		return;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2( line.x[1] - line.x[0] );
+	const f32 invDeltaX = fill_step_x( line.x[1] - line.x[0] );
 
 #ifdef IPOL_Z
 	slopeZ = (line.z[1] - line.z[0]) * invDeltaX;
@@ -268,7 +254,7 @@ void CTRTextureGouraudAlphaNoZ::fragment_linear()
 	tFixPoint a3,r3, g3, b3;
 #endif
 
-	for ( s32 i = 0; i <= dx; ++i )
+	for ( s32 i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef CMP_Z
 		if ( line.z[0] < z[i] )
@@ -276,7 +262,7 @@ void CTRTextureGouraudAlphaNoZ::fragment_linear()
 #ifdef CMP_W
 		if ( line.w[0] >= z[i] )
 #endif
-
+		scissor_test_x
 		{
 
 #if defined(BURNINGVIDEO_RENDERER_FAST) && COLOR_MAX==0xff
@@ -418,7 +404,7 @@ void CTRTextureGouraudAlphaNoZ::fragment_linear_test()
 		return;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2(line.x[1] - line.x[0]);
+	const f32 invDeltaX = fill_step_x(line.x[1] - line.x[0]);
 
 #ifdef IPOL_Z
 	slopeZ = (line.z[1] - line.z[0]) * invDeltaX;
@@ -488,7 +474,7 @@ void CTRTextureGouraudAlphaNoZ::fragment_linear_test()
 	tFixPoint a3, r3, g3, b3;
 #endif
 
-	for (s32 i = 0; i <= dx; ++i)
+	for (s32 i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef CMP_Z
 		if (line.z[0] < z[i])
@@ -496,7 +482,7 @@ void CTRTextureGouraudAlphaNoZ::fragment_linear_test()
 #ifdef CMP_W
 		if (line.w[0] >= z[i])
 #endif
-
+		scissor_test_x
 		{
 
 #if defined(BURNINGVIDEO_RENDERER_FAST) && COLOR_MAX==0xff
@@ -639,7 +625,7 @@ void CTRTextureGouraudAlphaNoZ::fragment_point_noz()
 		return;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2(line.x[1] - line.x[0]);
+	const f32 invDeltaX = fill_step_x(line.x[1] - line.x[0]);
 
 #ifdef IPOL_Z
 	slopeZ = (line.z[1] - line.z[0]) * invDeltaX;
@@ -708,15 +694,15 @@ void CTRTextureGouraudAlphaNoZ::fragment_point_noz()
 #ifdef IPOL_C1
 	tFixPoint a3, r3, g3, b3;
 #endif
-	for (s32 i = 0; i <= dx; ++i)
+	for (s32 i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef CMP_Z
-		if (line.z[0] < z[i])
+		//if (line.z[0] < z[i])
 #endif
 #ifdef CMP_W
 //			if (line.w[0] >= z[i])
 #endif
-
+			scissor_test_x
 			{
 
 #if defined(BURNINGVIDEO_RENDERER_FAST) && COLOR_MAX==0xff
@@ -827,9 +813,9 @@ void CTRTextureGouraudAlphaNoZ::drawTriangle(const s4DVertex* burning_restrict a
 	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] = reciprocal_zero( ca );
-	scan.invDeltaY[1] = reciprocal_zero( ba );
-	scan.invDeltaY[2] = reciprocal_zero( cb );
+	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;
@@ -968,7 +954,7 @@ void CTRTextureGouraudAlphaNoZ::drawTriangle(const s4DVertex* burning_restrict a
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -1004,6 +990,8 @@ void CTRTextureGouraudAlphaNoZ::drawTriangle(const s4DVertex* burning_restrict a
 #endif
 
 			// render a scanline
+			interlace_scanline
+			scissor_test_y
 			(this->*fragmentShader) ();
 
 			scan.x[0] += scan.slopeX[0];
@@ -1112,7 +1100,6 @@ void CTRTextureGouraudAlphaNoZ::drawTriangle(const s4DVertex* burning_restrict a
 		yEnd = fill_convention_right( c->Pos.y );
 
 #ifdef SUBTEXEL
-
 		subPixel = ( (f32) yStart ) - b->Pos.y;
 
 		// correct to pixel center
@@ -1152,7 +1139,7 @@ void CTRTextureGouraudAlphaNoZ::drawTriangle(const s4DVertex* burning_restrict a
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -1188,6 +1175,8 @@ void CTRTextureGouraudAlphaNoZ::drawTriangle(const s4DVertex* burning_restrict a
 #endif
 
 			// render a scanline
+			interlace_scanline
+			scissor_test_y
 			(this->*fragmentShader) ();
 
 			scan.x[0] += scan.slopeX[0];

source/Irrlicht/CTRTextureGouraudNoZ2.cpp

@@ -89,11 +89,16 @@ public:
 
 	//! 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_;
 
 private:
-	void scanline_bilinear ();
 
+	// fragment shader
+	typedef void (CTRTextureGouraudNoZ2::* tFragmentShader) ();
+	void fragment_bilinear();
+	void fragment_no_filter();
+
+	tFragmentShader fragmentShader;
 };
 
 //! constructor
@@ -103,13 +108,32 @@ CTRTextureGouraudNoZ2::CTRTextureGouraudNoZ2(CBurningVideoDriver* driver)
 	#ifdef _DEBUG
 	setDebugName("CTRTextureGouraudNoZ2");
 	#endif
+
+	fragmentShader = &CTRTextureGouraudNoZ2::fragment_bilinear;
 }
 
-
-
 /*!
 */
-void CTRTextureGouraudNoZ2::scanline_bilinear ( )
+void CTRTextureGouraudNoZ2::OnSetMaterial(const SBurningShaderMaterial& material)
+{
+
+	if (material.org.TextureLayer[0].BilinearFilter ||
+		material.org.TextureLayer[0].TrilinearFilter ||
+		material.org.TextureLayer[0].AnisotropicFilter
+		)
+	{
+		fragmentShader = &CTRTextureGouraudNoZ2::fragment_bilinear;
+	}
+	else
+	{
+		fragmentShader = &CTRTextureGouraudNoZ2::fragment_no_filter;
+	}
+
+}
+
+/*!
+*/
+void CTRTextureGouraudNoZ2::fragment_bilinear()
 {
 	tVideoSample *dst;
 
@@ -149,7 +173,7 @@ void CTRTextureGouraudNoZ2::scanline_bilinear ( )
 		return;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2( line.x[1] - line.x[0] );
+	const f32 invDeltaX = fill_step_x( line.x[1] - line.x[0] );
 
 #ifdef IPOL_Z
 	slopeZ = (line.z[1] - line.z[0]) * invDeltaX;
@@ -168,7 +192,7 @@ void CTRTextureGouraudNoZ2::scanline_bilinear ( )
 #endif
 
 #ifdef SUBTEXEL
-	subPixel = ( (f32) xStart ) - line.x[0];
+	subPixel = ((f32) xStart) - line.x[0];
 #ifdef IPOL_Z
 	line.z[0] += slopeZ * subPixel;
 #endif
@@ -200,7 +224,7 @@ void CTRTextureGouraudNoZ2::scanline_bilinear ( )
 	tFixPoint ty0;
 	tFixPoint r0, g0, b0;
 
-	for ( s32 i = 0; i <= dx; ++i )
+	for ( s32 i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef CMP_Z
 		if ( line.z[0] < z[i] )
@@ -208,7 +232,7 @@ void CTRTextureGouraudNoZ2::scanline_bilinear ( )
 #ifdef CMP_W
 		if ( line.w[0] >= z[i] )
 #endif
-
+		scissor_test_x
 		{
 #ifdef INVERSE_W
 			inversew = fix_inverse32 ( line.w[0] );
@@ -248,6 +272,146 @@ void CTRTextureGouraudNoZ2::scanline_bilinear ( )
 
 }
 
+/*!
+*/
+void CTRTextureGouraudNoZ2::fragment_no_filter()
+{
+	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[BURNING_MATERIAL_MAX_TEXTURES];
+#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[1] - line.c[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] += 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
+
+
+	f32 inversew = FIX_POINT_F32_MUL;
+
+	tFixPoint tx0;
+	tFixPoint ty0;
+	//tFixPoint r0, g0, b0;
+
+	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
+			//scissor_test_x
+			{
+	#ifdef INVERSE_W
+				inversew = fix_inverse32(line.w[0]);
+	#endif
+				tx0 = tofix(line.t[0][0].x,inversew);
+				ty0 = tofix(line.t[0][0].y,inversew);
+				//skybox
+				dst[i] = getTexel_plain(&IT[0], tx0, ty0);
+
+				//getSample_texture ( r0, g0, b0, IT+0, tx0, ty0 );
+				//dst[i] = fix_to_sample( r0, g0, b0 );
+
+	#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] += slopeC;
+#endif
+#ifdef IPOL_T0
+		line.t[0][0] += slopeT[0];
+#endif
+#ifdef IPOL_T1
+		line.t[1][0] += slopeT[1];
+#endif
+	}
+
+}
+
 void CTRTextureGouraudNoZ2::drawTriangle(const s4DVertex* burning_restrict a, const s4DVertex* burning_restrict b, const s4DVertex* burning_restrict c)
 {
 	// sort on height, y
@@ -258,10 +422,11 @@ void CTRTextureGouraudNoZ2::drawTriangle(const s4DVertex* burning_restrict a, co
 	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] = reciprocal_zero( ca );
-	scan.invDeltaY[1] = reciprocal_zero( ba );
-	scan.invDeltaY[2] = reciprocal_zero( cb );
+	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;
@@ -315,7 +480,7 @@ void CTRTextureGouraudNoZ2::drawTriangle(const s4DVertex* burning_restrict a, co
 #endif
 
 	// rasterize upper sub-triangle
-	if ( (f32) 0.0 != scan.invDeltaY[1]  )
+	if (F32_GREATER_0(scan.invDeltaY[1])  )
 	{
 		// calculate slopes for top edge
 		scan.slopeX[1] = (b->Pos.x - a->Pos.x) * scan.invDeltaY[1];
@@ -385,7 +550,7 @@ void CTRTextureGouraudNoZ2::drawTriangle(const s4DVertex* burning_restrict a, co
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -415,8 +580,11 @@ void CTRTextureGouraudNoZ2::drawTriangle(const s4DVertex* burning_restrict a, co
 			line.t[1][scan.right] = scan.t[1][1];
 #endif
 
+
 			// render a scanline
-			scanline_bilinear ( );
+			interlace_scanline
+			scissor_test_y
+			(this->*fragmentShader) ();
 
 			scan.x[0] += scan.slopeX[0];
 			scan.x[1] += scan.slopeX[1];
@@ -450,10 +618,10 @@ void CTRTextureGouraudNoZ2::drawTriangle(const s4DVertex* burning_restrict a, co
 	}
 
 	// rasterize lower sub-triangle
-	if ( (f32) 0.0 != scan.invDeltaY[2] )
+	if (F32_GREATER_0(scan.invDeltaY[2]) )
 	{
 		// advance to middle point
-		if( (f32) 0.0 != scan.invDeltaY[1] )
+		if(F32_GREATER_0(scan.invDeltaY[1]) )
 		{
 			temp[0] = b->Pos.y - a->Pos.y;	// dy
 
@@ -545,7 +713,7 @@ void CTRTextureGouraudNoZ2::drawTriangle(const s4DVertex* burning_restrict a, co
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -576,7 +744,9 @@ void CTRTextureGouraudNoZ2::drawTriangle(const s4DVertex* burning_restrict a, co
 #endif
 
 			// render a scanline
-			scanline_bilinear ();
+			interlace_scanline
+			scissor_test_y
+			(this->*fragmentShader) ();
 
 			scan.x[0] += scan.slopeX[0];
 			scan.x[1] += scan.slopeX[1];

source/Irrlicht/CTRTextureGouraudVertexAlpha2.cpp

@@ -93,7 +93,7 @@ public:
 
 
 private:
-	void scanline_bilinear ();
+	void fragmentShader();
 
 };
 
@@ -110,7 +110,7 @@ CTRTextureVertexAlpha2::CTRTextureVertexAlpha2(CBurningVideoDriver* driver)
 
 /*!
 */
-void CTRTextureVertexAlpha2::scanline_bilinear (  )
+void CTRTextureVertexAlpha2::fragmentShader( )
 {
 	tVideoSample *dst;
 
@@ -150,7 +150,7 @@ void CTRTextureVertexAlpha2::scanline_bilinear (  )
 		return;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2( line.x[1] - line.x[0] );
+	const f32 invDeltaX = fill_step_x( line.x[1] - line.x[0] );
 
 #ifdef IPOL_Z
 	slopeZ = (line.z[1] - line.z[0]) * invDeltaX;
@@ -219,7 +219,7 @@ void CTRTextureVertexAlpha2::scanline_bilinear (  )
 #endif
 
 
-	for ( s32 i = 0; i <= dx; ++i )
+	for ( s32 i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef CMP_Z
 		if ( line.z[0] < z[i] )
@@ -332,9 +332,9 @@ void CTRTextureVertexAlpha2::drawTriangle(const s4DVertex* burning_restrict a, c
 	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] = reciprocal_zero( ca );
-	scan.invDeltaY[1] = reciprocal_zero( ba );
-	scan.invDeltaY[2] = reciprocal_zero( cb );
+	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;
@@ -473,7 +473,7 @@ void CTRTextureVertexAlpha2::drawTriangle(const s4DVertex* burning_restrict a, c
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -509,7 +509,7 @@ void CTRTextureVertexAlpha2::drawTriangle(const s4DVertex* burning_restrict a, c
 #endif
 
 			// render a scanline
-			scanline_bilinear ();
+			interlace_scanline fragmentShader();
 
 			scan.x[0] += scan.slopeX[0];
 			scan.x[1] += scan.slopeX[1];
@@ -656,7 +656,7 @@ void CTRTextureVertexAlpha2::drawTriangle(const s4DVertex* burning_restrict a, c
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -692,7 +692,7 @@ void CTRTextureVertexAlpha2::drawTriangle(const s4DVertex* burning_restrict a, c
 #endif
 
 			// render a scanline
-			scanline_bilinear ();
+			interlace_scanline fragmentShader();
 
 			scan.x[0] += scan.slopeX[0];
 			scan.x[1] += scan.slopeX[1];

source/Irrlicht/CTRTextureLightMap2_Add.cpp

@@ -86,7 +86,7 @@ public:
 
 
 private:
-	void scanline_bilinear ();
+	void fragmentShader();
 
 };
 
@@ -103,7 +103,7 @@ CTRTextureLightMap2_Add::CTRTextureLightMap2_Add(CBurningVideoDriver* driver)
 
 /*!
 */
-REALINLINE void CTRTextureLightMap2_Add::scanline_bilinear ()
+REALINLINE void CTRTextureLightMap2_Add::fragmentShader()
 {
 	tVideoSample *dst;
 
@@ -143,7 +143,7 @@ REALINLINE void CTRTextureLightMap2_Add::scanline_bilinear ()
 		return;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2( line.x[1] - line.x[0] );
+	const f32 invDeltaX = fill_step_x( line.x[1] - line.x[0] );
 
 #ifdef IPOL_Z
 	slopeZ = (line.z[1] - line.z[0]) * invDeltaX;
@@ -199,7 +199,7 @@ REALINLINE void CTRTextureLightMap2_Add::scanline_bilinear ()
 #endif
 
 
-	for ( s32 i = 0; i <= dx; i++ )
+	for ( s32 i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef CMP_Z
 		if ( line.z[0] < z[i] )
@@ -274,9 +274,9 @@ void CTRTextureLightMap2_Add::drawTriangle(const s4DVertex* burning_restrict a,
 	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] = reciprocal_zero( ca );
-	scan.invDeltaY[1] = reciprocal_zero( ba );
-	scan.invDeltaY[2] = reciprocal_zero( cb );
+	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;
@@ -400,7 +400,7 @@ void CTRTextureLightMap2_Add::drawTriangle(const s4DVertex* burning_restrict a,
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -431,7 +431,7 @@ void CTRTextureLightMap2_Add::drawTriangle(const s4DVertex* burning_restrict a,
 #endif
 
 			// render a scanline
-			scanline_bilinear ();
+			interlace_scanline fragmentShader();
 
 			scan.x[0] += scan.slopeX[0];
 			scan.x[1] += scan.slopeX[1];
@@ -560,7 +560,7 @@ void CTRTextureLightMap2_Add::drawTriangle(const s4DVertex* burning_restrict a,
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -591,7 +591,7 @@ void CTRTextureLightMap2_Add::drawTriangle(const s4DVertex* burning_restrict a,
 #endif
 
 			// render a scanline
-			scanline_bilinear ();
+			interlace_scanline fragmentShader();
 
 			scan.x[0] += scan.slopeX[0];
 			scan.x[1] += scan.slopeX[1];

source/Irrlicht/CTRTextureLightMap2_M1.cpp

@@ -121,7 +121,7 @@ REALINLINE void CTRTextureLightMap2_M1::scanline_bilinear2 ()
 		return;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2( line.x[1] - line.x[0] );
+	const f32 invDeltaX = fill_step_x( line.x[1] - line.x[0] );
 
 	// search z-buffer for first not occulled pixel
 	z = (fp24*) DepthBuffer->lock() + ( line.y * RenderTarget->getDimension().Width ) + xStart;
@@ -193,7 +193,7 @@ REALINLINE void CTRTextureLightMap2_M1::scanline_bilinear2 ()
 #endif
 
 
-	for ( ;i <= dx; i++ )
+	for ( ;i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef IPOL_W
 		if ( line.w[0] >= z[i] )
@@ -255,9 +255,9 @@ void CTRTextureLightMap2_M1::drawTriangle(const s4DVertex* burning_restrict a, c
 	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] = reciprocal_zero( ca );
-	scan.invDeltaY[1] = reciprocal_zero( ba );
-	scan.invDeltaY[2] = reciprocal_zero( cb );
+	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;
@@ -381,7 +381,7 @@ void CTRTextureLightMap2_M1::drawTriangle(const s4DVertex* burning_restrict a, c
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -542,7 +542,7 @@ void CTRTextureLightMap2_M1::drawTriangle(const s4DVertex* burning_restrict a, c
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];

source/Irrlicht/CTRTextureLightMap2_M2.cpp

@@ -121,7 +121,7 @@ REALINLINE void CTRTextureLightMap2_M2::scanline_bilinear2 ()
 		return;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2( line.x[1] - line.x[0] );
+	const f32 invDeltaX = fill_step_x( line.x[1] - line.x[0] );
 
 	// search z-buffer for first not occulled pixel
 	z = (fp24*) DepthBuffer->lock() + ( line.y * RenderTarget->getDimension().Width ) + xStart;
@@ -254,9 +254,9 @@ void CTRTextureLightMap2_M2::drawTriangle(const s4DVertex* burning_restrict a, c
 	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] = reciprocal_zero( ca );
-	scan.invDeltaY[1] = reciprocal_zero( ba );
-	scan.invDeltaY[2] = reciprocal_zero( cb );
+	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;
@@ -380,7 +380,7 @@ void CTRTextureLightMap2_M2::drawTriangle(const s4DVertex* burning_restrict a, c
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -411,7 +411,7 @@ void CTRTextureLightMap2_M2::drawTriangle(const s4DVertex* burning_restrict a, c
 #endif
 
 			// render a scanline
-			scanline_bilinear2 ();
+			interlace_scanline scanline_bilinear2 ();
 
 			scan.x[0] += scan.slopeX[0];
 			scan.x[1] += scan.slopeX[1];
@@ -541,7 +541,7 @@ void CTRTextureLightMap2_M2::drawTriangle(const s4DVertex* burning_restrict a, c
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -572,7 +572,7 @@ void CTRTextureLightMap2_M2::drawTriangle(const s4DVertex* burning_restrict a, c
 #endif
 
 			// render a scanline
-			scanline_bilinear2 ();
+			interlace_scanline scanline_bilinear2 ();
 
 			scan.x[0] += scan.slopeX[0];
 			scan.x[1] += scan.slopeX[1];

source/Irrlicht/CTRTextureLightMap2_M4.cpp

@@ -93,10 +93,10 @@ private:
 	void scanline_bilinear2_mag ();
 	void scanline_bilinear2_min ();
 #else
-	#define scanline_bilinear2_mag scanline_bilinear
+	#define scanline_bilinear2_mag fragmentShader
 #endif
 
-	void scanline_bilinear ();
+	void fragmentShader();
 
 };
 
@@ -130,7 +130,7 @@ void CTRTextureLightMap2_M4::scanline_bilinear2_mag ()
 	SOFTWARE_DRIVER_2_CLIPCHECK;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2( line.x[1] - line.x[0] );
+	const f32 invDeltaX = fill_step_x( line.x[1] - line.x[0] );
 
 	// search z-buffer for first not occulled pixel
 	i = ( line.y * RenderTarget->getDimension().Width ) + xStart;
@@ -202,7 +202,7 @@ void CTRTextureLightMap2_M4::scanline_bilinear2_mag ()
 #endif
 
 
-	for ( ;i <= dx; i++ )
+	for ( ;i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef IPOL_W
 		if ( line.w[0] >= z[i] )
@@ -271,7 +271,7 @@ void CTRTextureLightMap2_M4::scanline_bilinear2_min ()
 	SOFTWARE_DRIVER_2_CLIPCHECK;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2( line.x[1] - line.x[0] );
+	const f32 invDeltaX = fill_step_x( line.x[1] - line.x[0] );
 
 	// search z-buffer for first not occulled pixel
 	z = (fp24*) DepthBuffer->lock() + ( line.y * RenderTarget->getDimension().Width ) + xStart;
@@ -334,7 +334,7 @@ void CTRTextureLightMap2_M4::scanline_bilinear2_min ()
 	tFixPoint r1, g1, b1;
 
 
-	for ( ;i <= dx; i++ )
+	for ( ;i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef IPOL_W
 		if ( line.w[0] >= z[i] )
@@ -393,9 +393,9 @@ void CTRTextureLightMap2_M4::drawTriangle_Min ( const s4DVertex* burning_restric
 	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] = reciprocal_zero( ca );
-	scan.invDeltaY[1] = reciprocal_zero( ba );
-	scan.invDeltaY[2] = reciprocal_zero( cb );
+	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;
@@ -519,7 +519,7 @@ void CTRTextureLightMap2_M4::drawTriangle_Min ( const s4DVertex* burning_restric
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -550,7 +550,7 @@ void CTRTextureLightMap2_M4::drawTriangle_Min ( const s4DVertex* burning_restric
 #endif
 
 			// render a scanline
-			scanline_bilinear2_min ();
+			interlace_scanline scanline_bilinear2_min ();
 
 			scan.x[0] += scan.slopeX[0];
 			scan.x[1] += scan.slopeX[1];
@@ -680,7 +680,7 @@ void CTRTextureLightMap2_M4::drawTriangle_Min ( const s4DVertex* burning_restric
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -711,7 +711,7 @@ void CTRTextureLightMap2_M4::drawTriangle_Min ( const s4DVertex* burning_restric
 #endif
 
 			// render a scanline
-			scanline_bilinear2_min ();
+			interlace_scanline scanline_bilinear2_min ();
 
 			scan.x[0] += scan.slopeX[0];
 			scan.x[1] += scan.slopeX[1];
@@ -769,9 +769,9 @@ void CTRTextureLightMap2_M4::drawTriangle(const s4DVertex* burning_restrict a, c
 		return;
 
 	// calculate delta y of the edges
-	scan.invDeltaY[0] = reciprocal_zero( ca );
-	scan.invDeltaY[1] = reciprocal_zero( ba );
-	scan.invDeltaY[2] = reciprocal_zero( cb );
+	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;
@@ -896,7 +896,7 @@ void CTRTextureLightMap2_M4::drawTriangle(const s4DVertex* burning_restrict a, c
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -927,7 +927,7 @@ void CTRTextureLightMap2_M4::drawTriangle(const s4DVertex* burning_restrict a, c
 #endif
 
 			// render a scanline
-			scanline_bilinear2_mag ();
+			interlace_scanline scanline_bilinear2_mag ();
 
 			scan.x[0] += scan.slopeX[0];
 			scan.x[1] += scan.slopeX[1];
@@ -1057,7 +1057,7 @@ void CTRTextureLightMap2_M4::drawTriangle(const s4DVertex* burning_restrict a, c
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -1088,7 +1088,7 @@ void CTRTextureLightMap2_M4::drawTriangle(const s4DVertex* burning_restrict a, c
 #endif
 
 			// render a scanline
-			scanline_bilinear2_mag ();
+			interlace_scanline scanline_bilinear2_mag ();
 
 			scan.x[0] += scan.slopeX[0];
 			scan.x[1] += scan.slopeX[1];

source/Irrlicht/CTRTextureLightMapGouraud2_M4.cpp

@@ -86,7 +86,7 @@ public:
 
 
 private:
-	void scanline_bilinear ();
+	void fragmentShader();
 
 };
 
@@ -103,7 +103,7 @@ CTRGTextureLightMap2_M4::CTRGTextureLightMap2_M4(CBurningVideoDriver* driver)
 
 /*!
 */
-void CTRGTextureLightMap2_M4::scanline_bilinear ()
+void CTRGTextureLightMap2_M4::fragmentShader()
 {
 	tVideoSample *dst;
 
@@ -143,7 +143,7 @@ void CTRGTextureLightMap2_M4::scanline_bilinear ()
 		return;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2( line.x[1] - line.x[0] );
+	const f32 invDeltaX = fill_step_x( line.x[1] - line.x[0] );
 
 #ifdef IPOL_Z
 	slopeZ = (line.z[1] - line.z[0]) * invDeltaX;
@@ -201,7 +201,7 @@ void CTRGTextureLightMap2_M4::scanline_bilinear ()
 	tFixPoint r3, g3, b3;
 #endif
 
-	for ( s32 i = 0; i <= dx; i++ )
+	for ( s32 i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X )
 	{
 #ifdef CMP_Z
 		if ( line.z[0] < z[i] )
@@ -283,9 +283,9 @@ void CTRGTextureLightMap2_M4::drawTriangle(const s4DVertex* burning_restrict a,
 	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] = reciprocal_zero( ca );
-	scan.invDeltaY[1] = reciprocal_zero( ba );
-	scan.invDeltaY[2] = reciprocal_zero( cb );
+	scan.invDeltaY[0] = fill_step_y( ca );
+	scan.invDeltaY[1] = fill_step_y( ba );
+	scan.invDeltaY[2] = fill_step_y( cb );
 
 	if ( F32_LOWER_0 ( scan.invDeltaY[0] )  )
 		return;
@@ -411,7 +411,7 @@ void CTRGTextureLightMap2_M4::drawTriangle(const s4DVertex* burning_restrict a,
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -442,7 +442,7 @@ void CTRGTextureLightMap2_M4::drawTriangle(const s4DVertex* burning_restrict a,
 #endif
 
 			// render a scanline
-			scanline_bilinear ();
+			interlace_scanline fragmentShader();
 
 			scan.x[0] += scan.slopeX[0];
 			scan.x[1] += scan.slopeX[1];
@@ -573,7 +573,7 @@ void CTRGTextureLightMap2_M4::drawTriangle(const s4DVertex* burning_restrict a,
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -604,7 +604,7 @@ void CTRGTextureLightMap2_M4::drawTriangle(const s4DVertex* burning_restrict a,
 #endif
 
 			// render a scanline
-			scanline_bilinear ();
+			interlace_scanline fragmentShader();
 
 			scan.x[0] += scan.slopeX[0];
 			scan.x[1] += scan.slopeX[1];

source/Irrlicht/CTRTextureWire2.cpp

@@ -91,7 +91,7 @@ public:
 	virtual bool canPointCloud() _IRR_OVERRIDE_  { return true; }
 
 protected:
-	virtual void scanline_bilinear ();
+	virtual void fragmentShader();
 
 	void renderAlphaLine ( const s4DVertex *a,const s4DVertex *b ) const;
 	void renderLine ( const s4DVertex *a,const s4DVertex *b, int renderZero = 0 ) const;
@@ -114,7 +114,7 @@ CTRTextureWire2::CTRTextureWire2(CBurningVideoDriver* driver)
 */
 void CTRTextureWire2::renderLine ( const s4DVertex *a,const s4DVertex *b, int renderZero) const
 {
-	int pitch0 = RenderTarget->getDimension().Width << VIDEO_SAMPLE_GRANULARITY;
+	int pitch0 = RenderTarget->getDimension().Width << SOFTWARE_DRIVER_2_RENDERTARGET_GRANULARITY;
 #ifdef USE_ZBUFFER
 	int pitch1 = RenderTarget->getDimension().Width << 2;
 #endif
@@ -138,7 +138,7 @@ void CTRTextureWire2::renderLine ( const s4DVertex *a,const s4DVertex *b, int re
 	fp24 *z;
 #endif
 
-	int xInc0 = 1 << VIDEO_SAMPLE_GRANULARITY;
+	int xInc0 = 1 << SOFTWARE_DRIVER_2_RENDERTARGET_GRANULARITY;
 	int yInc0 = pitch0;
 
 #ifdef USE_ZBUFFER
@@ -148,7 +148,7 @@ void CTRTextureWire2::renderLine ( const s4DVertex *a,const s4DVertex *b, int re
 
 	if ( dx < 0 )
 	{
-		xInc0 = - ( 1 << VIDEO_SAMPLE_GRANULARITY);
+		xInc0 = - ( 1 << SOFTWARE_DRIVER_2_RENDERTARGET_GRANULARITY);
 #ifdef USE_ZBUFFER
 		xInc1 = -4;
 #endif
@@ -173,7 +173,7 @@ void CTRTextureWire2::renderLine ( const s4DVertex *a,const s4DVertex *b, int re
 	}
 
 	SOFTWARE_DRIVER_2_CLIPCHECK_WIRE;
-	dst = (tVideoSample*) ( (u8*) RenderTarget->getData() + ( aposy * pitch0 ) + (aposx* (1<< VIDEO_SAMPLE_GRANULARITY) ) );
+	dst = (tVideoSample*) ( (u8*) RenderTarget->getData() + ( aposy * pitch0 ) + (aposx* (1<< SOFTWARE_DRIVER_2_RENDERTARGET_GRANULARITY) ) );
 #ifdef USE_ZBUFFER
 	z = (fp24*) ( (u8*) (fp24*) DepthBuffer->lock() + ( aposy * pitch1 ) + (aposx << 2 ) );
 #endif
@@ -182,7 +182,7 @@ void CTRTextureWire2::renderLine ( const s4DVertex *a,const s4DVertex *b, int re
 	m = dy << 1;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2( (f32)dx );
+	const f32 invDeltaX = fill_step_x( (f32)dx );
 
 #ifdef IPOL_Z
 	f32 slopeZ = (b->Pos.z - a->Pos.z) * invDeltaX;
@@ -283,7 +283,7 @@ void CTRTextureWire2::renderLine ( const s4DVertex *a,const s4DVertex *b, int re
 
 }
 
-void CTRTextureWire2::scanline_bilinear()
+void CTRTextureWire2::fragmentShader()
 {
 }
 

source/Irrlicht/CTR_transparent_reflection_2_layer.cpp

@@ -149,7 +149,7 @@ void CTR_transparent_reflection_2_layer::fragmentShader()
 		return;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2( line.x[1] - line.x[0] );
+	const f32 invDeltaX = fill_step_x( line.x[1] - line.x[0] );
 
 #ifdef IPOL_Z
 	slopeZ = (line.z[1] - line.z[0]) * invDeltaX;
@@ -206,7 +206,7 @@ void CTR_transparent_reflection_2_layer::fragmentShader()
 	switch(MaterialType) {
 	default:
 	case EMT_REFLECTION_2_LAYER:
-	for (s32 i = 0; i <= dx; ++i)
+	for (s32 i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef CMP_Z
 		if (line.z[0] < z[i])
@@ -263,7 +263,7 @@ void CTR_transparent_reflection_2_layer::fragmentShader()
 	break;
 
 	case EMT_TRANSPARENT_REFLECTION_2_LAYER:
-	for (s32 i = 0; i <= dx; ++i)
+	for (s32 i = 0; i <= dx; i += SOFTWARE_DRIVER_2_STEP_X)
 	{
 #ifdef CMP_Z
 	if (line.z[0] < z[i])
@@ -344,9 +344,9 @@ void CTR_transparent_reflection_2_layer::drawTriangle ( const s4DVertex* burning
 	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] = reciprocal_zero( ca );
-	scan.invDeltaY[1] = reciprocal_zero( ba );
-	scan.invDeltaY[2] = reciprocal_zero( cb );
+	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;
@@ -470,7 +470,7 @@ void CTR_transparent_reflection_2_layer::drawTriangle ( const s4DVertex* burning
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -501,6 +501,7 @@ void CTR_transparent_reflection_2_layer::drawTriangle ( const s4DVertex* burning
 #endif
 
 			// render a scanline
+			interlace_scanline
 			fragmentShader();
 
 			scan.x[0] += scan.slopeX[0];
@@ -630,7 +631,7 @@ void CTR_transparent_reflection_2_layer::drawTriangle ( const s4DVertex* burning
 #endif
 
 		// rasterize the edge scanlines
-		for( line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -661,6 +662,7 @@ void CTR_transparent_reflection_2_layer::drawTriangle ( const s4DVertex* burning
 #endif
 
 			// render a scanline
+			interlace_scanline
 			fragmentShader();
 
 			scan.x[0] += scan.slopeX[0];

source/Irrlicht/IBurningShader.cpp

@@ -12,29 +12,37 @@
 
 namespace irr
 {
-
-
 namespace video
 {
 
-	const tFixPointu IBurningShader::dithermask[] =
-	{
+const tFixPointu IBurningShader::dithermask[] =
+{
 	0x00,0x80,0x20,0xa0,
 	0xc0,0x40,0xe0,0x60,
 	0x30,0xb0,0x10,0x90,
 	0xf0,0x70,0xd0,0x50
-	};
+};
 
 void IBurningShader::constructor_IBurningShader(CBurningVideoDriver* driver)
 {
-	#ifdef _DEBUG
+#ifdef _DEBUG
 	setDebugName("IBurningShader");
-	#endif
+#endif
+
+	if (((unsigned long long)&scan & 15) || ((unsigned long long)&line & 15))
+	{
+		os::Printer::log("BurningVideo Shader not 16 byte aligned", ELL_ERROR);
+		_IRR_DEBUG_BREAK_IF(1);
+	}
+
+	Interlaced.enable = 0;
+	Interlaced.bypass = 1;
+	Interlaced.nr = 0;
 
 	EdgeTestPass = edge_test_pass;
 	EdgeTestPass_stack = edge_test_pass;
 
-	for ( u32 i = 0; i < BURNING_MATERIAL_MAX_TEXTURES; ++i )
+	for (u32 i = 0; i < BURNING_MATERIAL_MAX_TEXTURES; ++i)
 	{
 		IT[i].Texture = 0;
 	}
@@ -44,12 +52,12 @@ void IBurningShader::constructor_IBurningShader(CBurningVideoDriver* driver)
 
 	RenderTarget = 0;
 	ColorMask = COLOR_BRIGHT_WHITE;
-	DepthBuffer = (CDepthBuffer*) driver->getDepthBuffer ();
-	if ( DepthBuffer )
+	DepthBuffer = (CDepthBuffer*)driver->getDepthBuffer();
+	if (DepthBuffer)
 		DepthBuffer->grab();
 
-	Stencil = (CStencilBuffer*) driver->getStencilBuffer ();
-	if ( Stencil )
+	Stencil = (CStencilBuffer*)driver->getStencilBuffer();
+	if (Stencil)
 		Stencil->grab();
 
 	stencilOp[0] = StencilOp_KEEP;
@@ -76,7 +84,7 @@ IBurningShader::IBurningShader(
 	const c8* pixelShaderProgram,
 	const c8* pixelShaderEntryPointName,
 	E_PIXEL_SHADER_TYPE psCompileTarget,
-	const c8* geometryShaderProgram ,
+	const c8* geometryShaderProgram,
 	const c8* geometryShaderEntryPointName,
 	E_GEOMETRY_SHADER_TYPE gsCompileTarget,
 	scene::E_PRIMITIVE_TYPE inType,
@@ -109,9 +117,9 @@ IBurningShader::~IBurningShader()
 	if (Stencil)
 		Stencil->drop();
 
-	for ( u32 i = 0; i != BURNING_MATERIAL_MAX_TEXTURES; ++i )
+	for (u32 i = 0; i != BURNING_MATERIAL_MAX_TEXTURES; ++i)
 	{
-		if ( IT[i].Texture )
+		if (IT[i].Texture)
 			IT[i].Texture->drop();
 	}
 
@@ -121,12 +129,14 @@ IBurningShader::~IBurningShader()
 }
 
 //! sets a render target
-void IBurningShader::setRenderTarget(video::IImage* surface, const core::rect<s32>& viewPort)
+void IBurningShader::setRenderTarget(video::IImage* surface, const core::rect<s32>& viewPort, const interlaced_control interlaced)
 {
+	Interlaced = interlaced;
+
 	if (RenderTarget)
 		RenderTarget->drop();
 
-	RenderTarget = (video::CImage* ) surface;
+	RenderTarget = (video::CImage*) surface;
 
 	if (RenderTarget)
 	{
@@ -138,16 +148,16 @@ void IBurningShader::setRenderTarget(video::IImage* surface, const core::rect<s3
 
 
 //! sets the Texture
-void IBurningShader::setTextureParam( const size_t stage, video::CSoftwareTexture2* texture, s32 lodFactor)
+void IBurningShader::setTextureParam(const size_t stage, video::CSoftwareTexture2* texture, s32 lodFactor)
 {
-	sInternalTexture *it = &IT[stage];
+	sInternalTexture* it = &IT[stage];
 
-	if ( it->Texture)
+	if (it->Texture)
 		it->Texture->drop();
 
 	it->Texture = texture;
 
-	if ( it->Texture)
+	if (it->Texture)
 	{
 		it->Texture->grab();
 
@@ -156,48 +166,52 @@ void IBurningShader::setTextureParam( const size_t stage, video::CSoftwareTextur
 
 		//only mipmap chain (means positive lodFactor)
 		u32 existing_level = it->Texture->getMipmapLevel(lodFactor);
-		it->data = (tVideoSample*) it->Texture->lock(ETLM_READ_ONLY, existing_level, 0);
+#if !defined(PATCH_SUPERTUX_8_0_1_with_1_9_0)
+		it->data = (tVideoSample*)it->Texture->lock(ETLM_READ_ONLY, existing_level, 0);
+#else
+		it->data = (tVideoSample*)it->Texture->lock(ETLM_READ_ONLY, existing_level);
+#endif
 
 		// prepare for optimal fixpoint
-		it->pitchlog2 = s32_log2_s32 ( it->Texture->getPitch() );
+		it->pitchlog2 = s32_log2_s32(it->Texture->getPitch());
 
-		const core::dimension2d<u32> &dim = it->Texture->getSize();
-		it->textureXMask = s32_to_fixPoint ( dim.Width - 1 ) & FIX_POINT_UNSIGNED_MASK;
-		it->textureYMask = s32_to_fixPoint ( dim.Height - 1 ) & FIX_POINT_UNSIGNED_MASK;
+		const core::dimension2d<u32>& dim = it->Texture->getSize();
+		it->textureXMask = s32_to_fixPoint(dim.Width - 1) & FIX_POINT_UNSIGNED_MASK;
+		it->textureYMask = s32_to_fixPoint(dim.Height - 1) & FIX_POINT_UNSIGNED_MASK;
 	}
 }
 
 //emulate a line with degenerate triangle and special shader mode (not perfect...)
-void IBurningShader::drawLine ( const s4DVertex *a,const s4DVertex *b)
+void IBurningShader::drawLine(const s4DVertex* a, const s4DVertex* b)
 {
 	sVec2 d;
 	d.x = b->Pos.x - a->Pos.x;	d.x *= d.x;
 	d.y = b->Pos.y - a->Pos.y;	d.y *= d.y;
 	//if ( d.x * d.y < 0.001f ) return;
 
-	if ( a->Pos.x > b->Pos.x ) swapVertexPointer(&a, &b);
+	if (a->Pos.x > b->Pos.x) swapVertexPointer(&a, &b);
 
 	s4DVertex c = *a;
 
-	const f32 w = (f32)RenderTarget->getDimension().Width-1;
-	const f32 h = (f32)RenderTarget->getDimension().Height-1;
+	const f32 w = (f32)RenderTarget->getDimension().Width - 1;
+	const f32 h = (f32)RenderTarget->getDimension().Height - 1;
 
-	if ( d.x < 2.f ) { c.Pos.x = b->Pos.x + 1.f + d.y; if ( c.Pos.x > w ) c.Pos.x = w; }
+	if (d.x < 2.f) { c.Pos.x = b->Pos.x + 1.f + d.y; if (c.Pos.x > w) c.Pos.x = w; }
 	else c.Pos.x = b->Pos.x;
-	if ( d.y < 2.f ) { c.Pos.y = b->Pos.y + 1.f; if ( c.Pos.y > h ) c.Pos.y = h; EdgeTestPass |= edge_test_first_line; }
+	if (d.y < 2.f) { c.Pos.y = b->Pos.y + 1.f; if (c.Pos.y > h) c.Pos.y = h; EdgeTestPass |= edge_test_first_line; }
 
-	drawTriangle ( a,b,&c );
+	drawTriangle(a, b, &c);
 	EdgeTestPass &= ~edge_test_first_line;
 
 }
 
-void IBurningShader::drawPoint(const s4DVertex *a)
+void IBurningShader::drawPoint(const s4DVertex* a)
 {
 }
 
-void IBurningShader::drawWireFrameTriangle ( const s4DVertex *a,const s4DVertex *b,const s4DVertex *c )
+void IBurningShader::drawWireFrameTriangle(const s4DVertex* a, const s4DVertex* b, const s4DVertex* c)
 {
-	if ( EdgeTestPass & edge_test_pass ) drawTriangle(a, b, c);
+	if (EdgeTestPass & edge_test_pass) drawTriangle(a, b, c);
 	else if (EdgeTestPass & edge_test_point)
 	{
 		drawPoint(a);
@@ -256,7 +270,7 @@ void IBurningShader::setBasicRenderStates(const SMaterial& material, const SMate
 	Driver->setBasicRenderStates(material, lastMaterial, resetAllRenderstates);
 }
 
-s32 IBurningShader::getShaderConstantID(EBurningUniformFlags flags,const c8* name)
+s32 IBurningShader::getShaderConstantID(EBurningUniformFlags flags, const c8* name)
 {
 	if (!name || !name[0])
 		return -1;
@@ -285,7 +299,7 @@ const char* tiny_itoa(s32 value, int base)
 
 	b[p] = '\0';
 	do {
-		b[--p] = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"[value%base];
+		b[--p] = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"[value % base];
 		value /= base;
 	} while (value && p > 0);
 
@@ -302,7 +316,7 @@ bool IBurningShader::setShaderConstantID(EBurningUniformFlags flags, s32 index,
 	BurningUniform add;
 	while ((u32)index >= UniformInfo.size())
 	{
-		tiny_strcpy(add.name, tiny_itoa(UniformInfo.size(),10));
+		tiny_strcpy(add.name, tiny_itoa(UniformInfo.size(), 10));
 		add.type = flags;
 		UniformInfo.push_back(add);
 	}
@@ -340,7 +354,7 @@ 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);
 }
 
 void IBurningShader::setPixelShaderConstant(const f32* data, s32 startRegister, s32 constantAmount)

source/Irrlicht/IBurningShader.h

@@ -215,7 +215,6 @@ namespace video
 
 	};
 
-
 	class CBurningVideoDriver;
 	class IBurningShader : public IMaterialRenderer, public IMaterialRendererServices
 	{
@@ -247,7 +246,7 @@ namespace video
 		virtual ~IBurningShader();
 
 		//! sets a render target
-		virtual void setRenderTarget(video::IImage* surface, const core::rect<s32>& viewPort);
+		virtual void setRenderTarget(video::IImage* surface, const core::rect<s32>& viewPort, const interlaced_control interlaced);
 
 		//! sets the Texture
 		virtual void setTextureParam( const size_t stage, video::CSoftwareTexture2* texture, s32 lodFactor);
@@ -299,14 +298,38 @@ namespace video
 		virtual bool setPixelShaderConstant(s32 index, const u32* ints, int count)  _IRR_OVERRIDE_;
 		virtual IVideoDriver* getVideoDriver() _IRR_OVERRIDE_;
 
+#if defined(PATCH_SUPERTUX_8_0_1_with_1_9_0)
+		virtual bool setVertexShaderConstant(const c8* name, const f32* floats, int count)
+		{
+			return setVertexShaderConstant(getVertexShaderConstantID(name), floats, count);
+		}
+		virtual bool setVertexShaderConstant(const c8* name, const bool* bools, int count)
+		{
+			return setVertexShaderConstant(getVertexShaderConstantID(name), (const s32*)bools, count);
+		}
+		virtual bool setVertexShaderConstant(const c8* name, const s32* ints, int count)
+		{
+			return setVertexShaderConstant(getVertexShaderConstantID(name), ints, count);
+		}
+
+		virtual bool setPixelShaderConstant(const c8* name, const f32* floats, int count)
+		{
+			return setPixelShaderConstant(getPixelShaderConstantID(name), floats, count);
+		}
+		virtual bool setPixelShaderConstant(const c8* name, const bool* bools, int count)
+		{
+			return setPixelShaderConstant(getPixelShaderConstantID(name), (const s32*)bools, count);
+		}
+		virtual bool setPixelShaderConstant(const c8* name, const s32* ints, int count)
+		{
+			return setPixelShaderConstant(getPixelShaderConstantID(name), ints, count);
+		}
+#endif
+
 		//used if no color interpolation is defined
 		void setPrimitiveColor(const video::SColor& color)
 		{
-			#if BURNINGSHADER_COLOR_FORMAT == ECF_A8R8G8B8
-				PrimitiveColor = color.color;
-			#else
-				PrimitiveColor = color.toA1R5G5B5();
-			#endif
+			PrimitiveColor = color_to_sample(color);
 		}
 		void setTLFlag(size_t in /*eTransformLightFlags*/)
 		{
@@ -314,11 +337,7 @@ namespace video
 		}
 		void setFog(SColor color_fog)
 		{
-#if BURNINGSHADER_COLOR_FORMAT == ECF_A8R8G8B8
-			fog_color_sample = color_fog.color;
-#else
-			fog_color_sample = color_fog.toA1R5G5B5();
-#endif
+			fog_color_sample = color_to_sample(color_fog);
 			color_to_fix(fog_color, fog_color_sample);
 		}
 		void setScissor(const AbsRectangle& scissor)
@@ -349,13 +368,15 @@ namespace video
 		static const tFixPointu dithermask[ 4 * 4];
 
 		//draw degenerate triangle as line (left edge) drawTriangle -> holes,drawLine dda/bresenham
-		int EdgeTestPass; //edge_test_flag
-		int EdgeTestPass_stack;
+		size_t EdgeTestPass; //edge_test_flag
+		size_t EdgeTestPass_stack;
+		interlaced_control Interlaced; // passed from driver
 
 		eBurningStencilOp stencilOp[4];
 		tFixPoint AlphaRef;
 		int RenderPass_ShaderIsTransparent;
 
+		u8 _unused_pack[4];
 		sScanConvertData scan;
 		sScanLineData line;
 		tVideoSample PrimitiveColor; //used if no color interpolation is defined
@@ -365,6 +386,17 @@ namespace video
 		tVideoSample fog_color_sample;
 
 		AbsRectangle Scissor;
+
+		inline tVideoSample color_to_sample(const video::SColor& color) const
+		{
+			//RenderTarget->getColorFormat()
+#if SOFTWARE_DRIVER_2_RENDERTARGET_COLOR_FORMAT == ECF_A8R8G8B8
+			return color.color;
+#else
+			return color.toA1R5G5B5();
+#endif
+		}
+
 	};
 
 

source/Irrlicht/IDepthBuffer.h

@@ -21,7 +21,7 @@ namespace video
 		virtual ~IDepthBuffer() {};
 
 		//! clears the zbuffer
-		virtual void clear(f32 value) = 0;
+		virtual void clear(f32 value, const interlaced_control interlaced) = 0;
 
 		//! sets the new size of the zbuffer
 		virtual void setSize(const core::dimension2d<u32>& size) = 0;
@@ -52,7 +52,7 @@ namespace video
 		virtual ~IStencilBuffer() {};
 
 		//! clears the stencil buffer
-		virtual void clear(u8 value) = 0;
+		virtual void clear(u32 value, const interlaced_control interlaced) = 0;
 
 		//! sets the new size of the zbuffer
 		virtual void setSize(const core::dimension2d<u32>& size) = 0;

source/Irrlicht/S4DVertex.h

@@ -79,6 +79,13 @@ struct sVec2
 
 #include "irrpack.h"
 
+//! sVec2Pack is Irrlicht S3DVertex,S3DVertex2TCoords,S3DVertexTangents Texutre Coordinates.
+// Start address is not 4 byte aligned
+struct sVec2Pack
+{
+	f32 x, y;
+};
+
 //! sVec3Pack used in BurningShader, packed direction
 struct sVec3Pack
 {
@@ -553,8 +560,11 @@ struct SAligned4DVertex
 	}
 
 	s4DVertex* data;	//align to 16 byte
-	u8* mem;
 	size_t ElementSize;
+
+private:
+
+	u8* mem;
 };
 
 //#define memcpy_s4DVertexPair(dst,src) memcpy(dst,src,sizeof_s4DVertex * 2)
@@ -732,7 +742,7 @@ struct sScanConvertData
 {
 	u32 left;			// major edge left/right
 	u32 right;			// !left
-	u32 _unused_pack[2];
+	u8 _unused_pack[8];
 
 	f32 invDeltaY[4];	// inverse edge delta for screen space sorted triangle 
 
@@ -767,7 +777,7 @@ struct sScanConvertData
 struct sScanLineData
 {
 	s32 y;				// y position of scanline
-	u32 _unused_pack[1];
+	u8 _unused_pack[4];
 	f32 x[2];			// x start, x end of scanline
 
 #if defined ( SOFTWARE_DRIVER_2_USE_WBUFFER ) || defined ( SOFTWARE_DRIVER_2_PERSPECTIVE_CORRECT )
@@ -776,7 +786,8 @@ struct sScanLineData
 	f32 z[2];			// z start, z end of scanline
 #endif
 
-	u32 _unused_pack_1[2];
+	s32 x_edgetest;		// slope x
+	u8 _unused_pack_1[4];
 
 #if BURNING_MATERIAL_MAX_COLORS > 0
 	sVec4 c[BURNING_MATERIAL_MAX_COLORS][2];			// color start, color end of scanline

source/Irrlicht/SoftwareDriver2_compile_config.h

@@ -10,140 +10,167 @@
 // Generic Render Flags for burning's video rasterizer
 // defined now in irrlicht compile config
 
-#if defined(PATCH_SUPERTUX_8_0_1)
-	#undef BURNINGVIDEO_RENDERER_BEAUTIFUL
+#if 1 && defined(PATCH_SUPERTUX_8_0_1_with_1_9_0)
+#undef BURNINGVIDEO_RENDERER_BEAUTIFUL
+
+#define SOFTWARE_DRIVER_2_PERSPECTIVE_CORRECT
+#define SOFTWARE_DRIVER_2_SUBTEXEL
+#define SOFTWARE_DRIVER_2_BILINEAR
+#define SOFTWARE_DRIVER_2_LIGHTING
+#define SOFTWARE_DRIVER_2_USE_VERTEX_COLOR
+//#define SOFTWARE_DRIVER_2_USE_SEPARATE_SPECULAR_COLOR
+#define SOFTWARE_DRIVER_2_USE_WBUFFER
+#define SOFTWARE_DRIVER_2_32BIT
+#define	SOFTWARE_DRIVER_2_TEXTURE_COLOR_FORMAT	ECF_A8R8G8B8
+#define	SOFTWARE_DRIVER_2_RENDERTARGET_COLOR_FORMAT	ECF_A8R8G8B8
+#define SOFTWARE_DRIVER_2_TEXTURE_MAXSIZE		256
+#define SOFTWARE_DRIVER_2_TEXTURE_TRANSFORM
+#define SOFTWARE_DRIVER_2_MIPMAPPING_MAX		8
+#define SOFTWARE_DRIVER_2_MIPMAPPING_MIN_SIZE   1
+#define SOFTWARE_DRIVER_2_SCANLINE_MAG_MIN
+#define SOFTWARE_DRIVER_2_CLIPPING
+#define SOFTWARE_DRIVER_2_2D_AS_3D
+#define SOFTWARE_DRIVER_2_INTERLACED
 
-	//#define SOFTWARE_DRIVER_2_PERSPECTIVE_CORRECT
-	#define SOFTWARE_DRIVER_2_SUBTEXEL
-	//#define SOFTWARE_DRIVER_2_BILINEAR
-	#define SOFTWARE_DRIVER_2_LIGHTING
-	#define SOFTWARE_DRIVER_2_USE_VERTEX_COLOR
-	#define SOFTWARE_DRIVER_2_USE_SEPARATE_SPECULAR_COLOR
-	#define SOFTWARE_DRIVER_2_32BIT
-	#define SOFTWARE_DRIVER_2_MIPMAPPING
-	#define SOFTWARE_DRIVER_2_USE_WBUFFER
-	#define SOFTWARE_DRIVER_2_TEXTURE_TRANSFORM
-	#define SOFTWARE_DRIVER_2_TEXTURE_MAXSIZE		256
-	#define SOFTWARE_DRIVER_2_SCANLINE_MAG_MIN
-	#define SOFTWARE_DRIVER_2_CLIPPING
-	#define SOFTWARE_DRIVER_2_2D_AS_2D
 #endif
 
-
 #ifdef BURNINGVIDEO_RENDERER_BEAUTIFUL
-	#define SOFTWARE_DRIVER_2_PERSPECTIVE_CORRECT
-	#define SOFTWARE_DRIVER_2_SUBTEXEL
-	#define SOFTWARE_DRIVER_2_BILINEAR
-	#define SOFTWARE_DRIVER_2_LIGHTING
-	#define SOFTWARE_DRIVER_2_USE_VERTEX_COLOR
-	#define SOFTWARE_DRIVER_2_USE_SEPARATE_SPECULAR_COLOR
-	#define SOFTWARE_DRIVER_2_32BIT
-	#define SOFTWARE_DRIVER_2_MIPMAPPING
-	#define SOFTWARE_DRIVER_2_USE_WBUFFER
-	#define SOFTWARE_DRIVER_2_TEXTURE_TRANSFORM
-	#define SOFTWARE_DRIVER_2_TEXTURE_MAXSIZE		0
-	#define SOFTWARE_DRIVER_2_SCANLINE_MAG_MIN
-	#define SOFTWARE_DRIVER_2_CLIPPING
-	#define SOFTWARE_DRIVER_2_2D_AS_3D
+#define SOFTWARE_DRIVER_2_PERSPECTIVE_CORRECT
+#define SOFTWARE_DRIVER_2_SUBTEXEL
+#define SOFTWARE_DRIVER_2_BILINEAR
+#define SOFTWARE_DRIVER_2_LIGHTING
+#define SOFTWARE_DRIVER_2_USE_VERTEX_COLOR
+#define SOFTWARE_DRIVER_2_USE_SEPARATE_SPECULAR_COLOR
+#define SOFTWARE_DRIVER_2_USE_WBUFFER
+#define SOFTWARE_DRIVER_2_32BIT
+#define	SOFTWARE_DRIVER_2_TEXTURE_COLOR_FORMAT	ECF_A8R8G8B8
+#define	SOFTWARE_DRIVER_2_RENDERTARGET_COLOR_FORMAT	ECF_A8R8G8B8
+#define SOFTWARE_DRIVER_2_TEXTURE_MAXSIZE		0x100000
+#define SOFTWARE_DRIVER_2_TEXTURE_TRANSFORM
+#define SOFTWARE_DRIVER_2_MIPMAPPING_MAX		16
+#define SOFTWARE_DRIVER_2_MIPMAPPING_MIN_SIZE   1
+#define SOFTWARE_DRIVER_2_SCANLINE_MAG_MIN
+#define SOFTWARE_DRIVER_2_CLIPPING
+#define SOFTWARE_DRIVER_2_2D_AS_3D
+#define SOFTWARE_DRIVER_2_INTERLACED
 #endif
 
 //! Set Flags for Windows Mobile
 #ifdef BURNINGVIDEO_RENDERER_CE
-	#define SOFTWARE_DRIVER_2_PERSPECTIVE_CORRECT
-	#define SOFTWARE_DRIVER_2_SUBTEXEL
-	//#define SOFTWARE_DRIVER_2_BILINEAR
-	//#define SOFTWARE_DRIVER_2_LIGHTING
-	#define SOFTWARE_DRIVER_2_USE_VERTEX_COLOR
-	#define SOFTWARE_DRIVER_2_16BIT
-	#define SOFTWARE_DRIVER_2_MIPMAPPING
-	#define SOFTWARE_DRIVER_2_USE_WBUFFER
-	//#define SOFTWARE_DRIVER_2_TEXTURE_TRANSFORM
-	#define SOFTWARE_DRIVER_2_TEXTURE_MAXSIZE		64
-	#define SOFTWARE_DRIVER_2_SCANLINE_MAG_MIN
-	//#define SOFTWARE_DRIVER_2_CLIPPING
-	#define SOFTWARE_DRIVER_2_2D_AS_2D
+#define SOFTWARE_DRIVER_2_PERSPECTIVE_CORRECT
+#define SOFTWARE_DRIVER_2_SUBTEXEL
+//#define SOFTWARE_DRIVER_2_BILINEAR
+//#define SOFTWARE_DRIVER_2_LIGHTING
+#define SOFTWARE_DRIVER_2_USE_VERTEX_COLOR
+#define SOFTWARE_DRIVER_2_USE_WBUFFER
+#define SOFTWARE_DRIVER_2_16BIT
+#define	SOFTWARE_DRIVER_2_TEXTURE_COLOR_FORMAT	ECF_A1R5G5B5
+#define	SOFTWARE_DRIVER_2_RENDERTARGET_COLOR_FORMAT	ECF_A1R5G5B5
+#define SOFTWARE_DRIVER_2_TEXTURE_MAXSIZE		64
+//#define SOFTWARE_DRIVER_2_TEXTURE_TRANSFORM
+#define SOFTWARE_DRIVER_2_MIPMAPPING_MAX		4
+#define SOFTWARE_DRIVER_2_MIPMAPPING_MIN_SIZE   8
+#define SOFTWARE_DRIVER_2_SCANLINE_MAG_MIN
+//#define SOFTWARE_DRIVER_2_CLIPPING
+#define SOFTWARE_DRIVER_2_2D_AS_2D
 #endif
 
 #ifdef BURNINGVIDEO_RENDERER_FAST
-	#define SOFTWARE_DRIVER_2_PERSPECTIVE_CORRECT
-	#define SOFTWARE_DRIVER_2_SUBTEXEL
-	//#define SOFTWARE_DRIVER_2_BILINEAR
-	//#define SOFTWARE_DRIVER_2_LIGHTING
-	#define SOFTWARE_DRIVER_2_USE_VERTEX_COLOR
-	//#define SOFTWARE_DRIVER_2_32BIT
-	#define SOFTWARE_DRIVER_2_16BIT
-	#define SOFTWARE_DRIVER_2_MIPMAPPING
-	#define SOFTWARE_DRIVER_2_USE_WBUFFER
-	#define SOFTWARE_DRIVER_2_TEXTURE_MAXSIZE		256
-	#define SOFTWARE_DRIVER_2_SCANLINE_MAG_MIN
-	#define SOFTWARE_DRIVER_2_CLIPPING
-	#define SOFTWARE_DRIVER_2_2D_AS_2D
+#define SOFTWARE_DRIVER_2_PERSPECTIVE_CORRECT
+#define SOFTWARE_DRIVER_2_SUBTEXEL
+//#define SOFTWARE_DRIVER_2_BILINEAR
+//#define SOFTWARE_DRIVER_2_LIGHTING
+#define SOFTWARE_DRIVER_2_USE_VERTEX_COLOR
+#define SOFTWARE_DRIVER_2_USE_WBUFFER
+#define SOFTWARE_DRIVER_2_16BIT
+#define SOFTWARE_DRIVER_2_TEXTURE_MAXSIZE		256
+#define	SOFTWARE_DRIVER_2_TEXTURE_COLOR_FORMAT	ECF_A1R5G5B5
+#define	SOFTWARE_DRIVER_2_RENDERTARGET_COLOR_FORMAT	ECF_A1R5G5B5
+#define SOFTWARE_DRIVER_2_MIPMAPPING_MAX		4
+#define SOFTWARE_DRIVER_2_MIPMAPPING_MIN_SIZE   8
+#define SOFTWARE_DRIVER_2_SCANLINE_MAG_MIN
+#define SOFTWARE_DRIVER_2_CLIPPING
+#define SOFTWARE_DRIVER_2_2D_AS_2D
+#define SOFTWARE_DRIVER_2_INTERLACED
 #endif
 
 #ifdef BURNINGVIDEO_RENDERER_ULTRA_FAST
-	#define BURNINGVIDEO_RENDERER_FAST
+#define BURNINGVIDEO_RENDERER_FAST
 
-	//#define SOFTWARE_DRIVER_2_PERSPECTIVE_CORRECT
-	//#define SOFTWARE_DRIVER_2_SUBTEXEL
-	//#define SOFTWARE_DRIVER_2_BILINEAR
-	//#define SOFTWARE_DRIVER_2_LIGHTING
-	#define SOFTWARE_DRIVER_2_USE_VERTEX_COLOR
-	//#define SOFTWARE_DRIVER_2_32BIT
-	#define SOFTWARE_DRIVER_2_16BIT
-	//#define SOFTWARE_DRIVER_2_MIPMAPPING
-	//#define SOFTWARE_DRIVER_2_USE_WBUFFER
-	//#define SOFTWARE_DRIVER_2_TEXTURE_TRANSFORM
-	#define SOFTWARE_DRIVER_2_TEXTURE_MAXSIZE		128
-	#define SOFTWARE_DRIVER_2_SCANLINE_MAG_MIN
-	//#define SOFTWARE_DRIVER_2_CLIPPING
-	#define SOFTWARE_DRIVER_2_2D_AS_2D
+//#define SOFTWARE_DRIVER_2_PERSPECTIVE_CORRECT
+//#define SOFTWARE_DRIVER_2_SUBTEXEL
+//#define SOFTWARE_DRIVER_2_BILINEAR
+//#define SOFTWARE_DRIVER_2_LIGHTING
+#define SOFTWARE_DRIVER_2_USE_VERTEX_COLOR
+//#define SOFTWARE_DRIVER_2_USE_WBUFFER
+#define SOFTWARE_DRIVER_2_16BIT
+#define SOFTWARE_DRIVER_2_TEXTURE_MAXSIZE		128
+#define	SOFTWARE_DRIVER_2_TEXTURE_COLOR_FORMAT	ECF_A1R5G5B5
+#define	SOFTWARE_DRIVER_2_RENDERTARGET_COLOR_FORMAT	ECF_A1R5G5B5
+//#define SOFTWARE_DRIVER_2_TEXTURE_TRANSFORM
+#define SOFTWARE_DRIVER_2_MIPMAPPING_MAX		1
+#define SOFTWARE_DRIVER_2_MIPMAPPING_MIN_SIZE   1
+#define SOFTWARE_DRIVER_2_SCANLINE_MAG_MIN
+//#define SOFTWARE_DRIVER_2_CLIPPING
+#define SOFTWARE_DRIVER_2_2D_AS_2D
+#define SOFTWARE_DRIVER_2_INTERLACED
 #endif
 
-// Derivate flags
-
-// texture format
-#ifdef SOFTWARE_DRIVER_2_32BIT
-	#define	BURNINGSHADER_COLOR_FORMAT	ECF_A8R8G8B8
-#else
-	#define	BURNINGSHADER_COLOR_FORMAT	ECF_A1R5G5B5
-#endif
-
-// mip mapping - precalculated texture filter
-#if defined ( SOFTWARE_DRIVER_2_MIPMAPPING )
-	#if defined( BURNINGVIDEO_RENDERER_BEAUTIFUL )
-		#define SOFTWARE_DRIVER_2_MIPMAPPING_MAX		16
-		#define SOFTWARE_DRIVER_2_MIPMAPPING_MIN_SIZE   1
-	#elif defined ( BURNINGVIDEO_RENDERER_CE )
-		#define SOFTWARE_DRIVER_2_MIPMAPPING_MAX		4
-		#define SOFTWARE_DRIVER_2_MIPMAPPING_MIN_SIZE   8
-	#else
-		#define SOFTWARE_DRIVER_2_MIPMAPPING_MAX		4
-		#define SOFTWARE_DRIVER_2_MIPMAPPING_MIN_SIZE   8
-	#endif
-#else
-	#define SOFTWARE_DRIVER_2_MIPMAPPING_MAX		1
-	#define SOFTWARE_DRIVER_2_MIPMAPPING_MIN_SIZE   1
-#endif
-
-
 
 #ifndef REALINLINE
-	#ifdef _MSC_VER
-		#define REALINLINE __forceinline
-	#else
-		#define REALINLINE inline
-	#endif
+#ifdef _MSC_VER
+#define REALINLINE __forceinline
+#else
+#define REALINLINE inline
+#endif
 #endif
 
+#define reciprocal_zero(x) ((x) != 0.f ? 1.f / (x):0.f)
+#define reciprocal_one(x) ((x) != 0.f ? 1.f / (x):1.f)
+
+//Control Scanline output
+#define SOFTWARE_DRIVER_2_STEP_X 1
+#define SOFTWARE_DRIVER_2_STEP_Y 1
 
 // null check necessary (burningvideo only)
-#define reciprocal_zero(x) ((x) != 0.f ? 1.f / (x):0.f)
-static inline float reciprocal_zero2(float x) { return x != 0.f ? 1.f / x : 0.f; }
-#define reciprocal_one(x) ((x) != 0.f ? 1.f / (x):1.f)
+#define fill_step_y(y) ((y) != 0.f ? (float)1.f / (y):0.f)
+static inline float fill_step_x(float x) { return x != 0.f ? (float)SOFTWARE_DRIVER_2_STEP_X / x : 0.f; }
+
+#define interlace_control_bit 1
+#define interlace_control_mask ((1<<interlace_control_bit)-1)
+struct interlaced_control
+{
+	unsigned enable : 1;
+	unsigned bypass : 1;
+	unsigned nr : interlace_control_bit;
+};
+struct interlace_scanline_data { unsigned int y; };
+
+static inline interlaced_control interlace_disabled()
+{
+	interlaced_control v;
+	v.enable = 0;
+	v.bypass = 1;
+	v.nr = 0;
+	return v;
+}
+#if defined(SOFTWARE_DRIVER_2_INTERLACED)
+#define interlace_scanline if ( Interlaced.bypass | ((line.y & interlace_control_mask) == Interlaced.nr) )
+#define interlace_scanline_enabled if ( (line.y & interlace_control_mask) == Interlaced.nr )
+//#define interlace_scanline if ( Interlaced.disabled | (((line.y >> (interlace_control_bit-1) ) & 1) == (Interlaced.nr & 1)) )
+//#define interlace_scanline
+#else
+#define interlace_scanline
+#define interlace_scanline_enabled
+#endif
+
+#define scissor_test_y if ((~TL_Flag & TL_SCISSOR) || ((line.y >= Scissor.y0) & (line.y <= Scissor.y1)))
+#define scissor_test_x if ((~TL_Flag & TL_SCISSOR) || ((i+xStart >= Scissor.x0) & (i+xStart <= Scissor.x1)))
 
 #define fill_convention_left(x) (s32) ceilf(x)
 #define fill_convention_right(x) ((s32) ceilf(x))-1
 #define fill_convention_none(x) (s32) (x)
+#define fill_convention_edge(x) (s32) floorf(fabsf(x)+0.f)
 //#define fill_convention_left(x) 65536 - int(65536.0f - x)
 //#define fill_convention_right(x) 65535 - int(65536.0f - x)
 
@@ -157,7 +184,7 @@ static inline float reciprocal_zero2(float x) { return x != 0.f ? 1.f / x : 0.f;
 
 inline float reciprocal_zero_no(const float x)
 {
-	if (x*x <= 0.00001f) __debugbreak();
+	if (x * x <= 0.00001f) __debugbreak();
 	return 1.f / x;
 }
 #else
@@ -183,49 +210,49 @@ enum edge_test_flag
 #define fix_color_norm(x) x = (x+1) >> COLOR_MAX_LOG2
 
 //! from 1 bit to 5 bit
-#ifdef SOFTWARE_DRIVER_2_32BIT
-	#define fix_alpha_color_max(x)
+#if defined(SOFTWARE_DRIVER_2_32BIT)
+#define fix_alpha_color_max(x)
 #else
-	#define fix_alpha_color_max(x) if (x) x = (x << COLOR_MAX_LOG2) - 1
+#define fix_alpha_color_max(x) if (x) x = (x << COLOR_MAX_LOG2) - 1
 #endif
 
 // Check windows
 #if _WIN32 || _WIN64
 #if _WIN64
-	#define ENV64BIT
+#define ENV64BIT
 #else
-	#define ENV32BIT
+#define ENV32BIT
 #endif
 #endif
 
 // Check GCC
 #if __GNUC__
 #if __x86_64__ || __ppc64__
-	#define ENV64BIT
+#define ENV64BIT
 #else
-	#define ENV32BIT
+#define ENV32BIT
 #endif
 #endif
 
 #if defined(ENV64BIT) && defined(BURNINGVIDEO_RENDERER_BEAUTIFUL)
-	typedef float ipoltype;
+typedef float ipoltype;
 #else
-	typedef float ipoltype;
+typedef float ipoltype;
 #endif
 
 #define	ipol_lower_equal_0(n)	((n) <= (ipoltype)0.0)
 #define	ipol_greater_0(n)		((n) >  (ipoltype)0.0)
 
 #if	(_MSC_VER > 1700 )
-	#define burning_restrict __restrict
+#define burning_restrict __restrict
 #else
-	#define burning_restrict
+#define burning_restrict
 #endif
 
 /*
 	if (condition) state |= mask; else state &= ~mask;
 */
-static inline void burning_setbit(size_t &state, int condition, size_t mask)
+static inline void burning_setbit(size_t& state, int condition, size_t mask)
 {
 	if (condition) state |= mask;
 	else state &= ~mask;
@@ -234,7 +261,7 @@ static inline void burning_setbit(size_t &state, int condition, size_t mask)
 /*
 	if (condition) state |= m; else state &= ~m;
 */
-REALINLINE void burning_setbit32(unsigned int &state, int condition, const unsigned int mask)
+REALINLINE void burning_setbit32(unsigned int& state, int condition, const unsigned int mask)
 {
 	// 0, or any positive to mask
 	//s32 conmask = -condition >> 31;
@@ -246,44 +273,37 @@ REALINLINE void burning_setbit32(unsigned int &state, int condition, const unsig
 #define burning_create(s) burning_create_indirect(s)
 
 
-#if defined(PATCH_SUPERTUX_8_0_1)
-#define getData lock
+#if defined(PATCH_SUPERTUX_8_0_1_with_1_9_0)
 #define snprintf_irr sprintf_s
-
-#ifdef SOFTWARE_DRIVER_2_USE_VERTEX_COLOR
-#ifdef SOFTWARE_DRIVER_2_USE_SEPARATE_SPECULAR_COLOR
-#define BURNING_MATERIAL_MAX_COLORS 2
-#else
-#define BURNING_MATERIAL_MAX_COLORS 1
-#endif
-#else
-#define BURNING_MATERIAL_MAX_COLORS 0
-#endif
-
-#ifndef _IRR_OVERRIDE_
-#define _IRR_OVERRIDE_ /**/
-#endif
-
-#define fix_to_color fix_to_sample
-#define fix4_to_color fix4_to_sample
-#define vec4_to_fix getSample_color
-#define SOFTWARE_DRIVER_2_MIPMAPPING_LOD_BIAS 0
+#define EVDF_DEPTH_CLAMP 43
+#define E_CUBE_SURFACE int
+#define ECFN_DISABLED 0
 
 namespace irr {
+	namespace video {
 
-	REALINLINE void memcpy32_small(void * dest, const void *source, size_t bytesize)
+		//! Enum for the flags of clear buffer
+		enum E_CLEAR_BUFFER_FLAG
 		{
-		size_t c = bytesize >> 2;
+			ECBF_NONE = 0,
+			ECBF_COLOR = 1,
+			ECBF_DEPTH = 2,
+			ECBF_STENCIL = 4,
+			ECBF_ALL = ECBF_COLOR | ECBF_DEPTH | ECBF_STENCIL
+		};
 
-		do
+		//! For SMaterial.ZWriteEnable
+		enum E_ZWRITE
 		{
-			((unsigned int *)dest)[c - 1] = ((unsigned int *)source)[c - 1];
-		} while (--c);
-
+			EZW_OFF = 0,
+			EZW_AUTO,
+			EZW_ON
+		};
 	}
+}
+#endif // PATCH_SUPERTUX_8_0_1_with_1_9_0
 
-} // namespace irr
-#endif // #if defined(PATCH_SUPERTUX_8_0_1)
-
+//! Size of a static C-style array.
+#define array_size(_arr)  ((sizeof(_arr)/sizeof(*_arr)))
 
 #endif // __S_VIDEO_2_SOFTWARE_COMPILE_CONFIG_H_INCLUDED__

source/Irrlicht/SoftwareDriver2_helper.h

@@ -22,7 +22,7 @@ namespace irr
 
 // supporting different packed pixel needs many defines...
 
-#ifdef SOFTWARE_DRIVER_2_32BIT
+#if defined(SOFTWARE_DRIVER_2_32BIT)
 	typedef u32	tVideoSample;
 	typedef u32	tStencilSample;
 
@@ -40,8 +40,8 @@ namespace irr
 	#define	COLOR_MAX_LOG2				8
 	#define	COLOR_BRIGHT_WHITE			0xFFFFFFFF
 
-	#define VIDEO_SAMPLE_GRANULARITY	(unsigned)2
-
+	#define SOFTWARE_DRIVER_2_TEXTURE_GRANULARITY	(unsigned)2
+	#define SOFTWARE_DRIVER_2_RENDERTARGET_GRANULARITY	(unsigned)2
 #else
 	typedef u16	tVideoSample;
 	typedef u8	tStencilSample;
@@ -59,7 +59,8 @@ namespace irr
 	#define	COLOR_MAX					0x1F
 	#define	COLOR_MAX_LOG2				5
 	#define	COLOR_BRIGHT_WHITE			0xFFFF
-	#define VIDEO_SAMPLE_GRANULARITY	(unsigned)1
+	#define SOFTWARE_DRIVER_2_TEXTURE_GRANULARITY	(unsigned)1
+	#define SOFTWARE_DRIVER_2_RENDERTARGET_GRANULARITY	(unsigned)1
 
 #endif
 
@@ -140,6 +141,19 @@ inline void memset16(void * dest, const u16 value, size_t bytesize)
 	}
 }
 
+//! memset interleaved
+inline void memset32_interlaced(void* dest, const u32 value, size_t pitch,u32 height,const interlaced_control Interlaced)
+{
+	if (Interlaced.bypass) return memset32(dest, value, pitch * height);
+
+	u8* dst = (u8*)dest;
+	interlace_scanline_data line;
+	for (line.y = 0; line.y < height; line.y += SOFTWARE_DRIVER_2_STEP_Y)
+	{
+		interlace_scanline_enabled memset32(dst, value, pitch);
+		dst += pitch;
+	}
+}
 
 // byte-align structures
 #include "irrpack.h"
@@ -577,7 +591,7 @@ REALINLINE tFixPoint imulFix2(const tFixPoint x, const tFixPoint y)
 */
 REALINLINE tFixPoint imulFix_tex1(const tFixPoint x, const tFixPoint y)
 {
-#ifdef SOFTWARE_DRIVER_2_32BIT
+#if SOFTWARE_DRIVER_2_TEXTURE_COLOR_FORMAT == ECF_A8R8G8B8
 	return (((tFixPointu)x >> 2)*(((tFixPointu)y + FIX_POINT_ONE) >> 2)) >> (tFixPointu) (FIX_POINT_PRE + 4);
 #else
 	return (x * (y+ FIX_POINT_ONE)) >> (FIX_POINT_PRE + 5);
@@ -598,7 +612,7 @@ REALINLINE tFixPoint imulFix_tex2(const tFixPoint x, const tFixPoint y)
 
 REALINLINE tFixPoint imulFix_tex4(const tFixPoint x, const tFixPoint y)
 {
-#ifdef SOFTWARE_DRIVER_2_32BIT
+#if SOFTWARE_DRIVER_2_TEXTURE_COLOR_FORMAT == ECF_A8R8G8B8
 	tFixPoint a = (((tFixPointu)x >> 2)*(((tFixPointu)y + FIX_POINT_ONE) >> 2)) >> (tFixPointu)(FIX_POINT_PRE + 2);
 #else
 	tFixPoint a = (x * (y + FIX_POINT_ONE)) >> (FIX_POINT_PRE + 3);
@@ -608,17 +622,6 @@ REALINLINE tFixPoint imulFix_tex4(const tFixPoint x, const tFixPoint y)
 }
 
 
-#if 0
-#define imulFix_tex1(x,y) ((((tFixPointu)x >> 2) * ((tFixPointu)y >> 2)) >> (tFixPointu)(FIX_POINT_PRE + 4))
-#define imulFix_tex2(x,y) ((((tFixPointu)x >> 2) * ((tFixPointu)y >> 2)) >> (tFixPointu)(FIX_POINT_PRE + 3))
-
-#ifdef SOFTWARE_DRIVER_2_32BIT
-#define imulFix_tex4(x,y) ( ( (tFixPointu) x >> 2 ) * ( (tFixPointu) y >> 2 ) ) >> (tFixPointu) ( FIX_POINT_PRE + 2 )
-#else
-#define imulFix_tex4(x,y) ( x * y) >> ( FIX_POINT_PRE + ( VIDEO_SAMPLE_GRANULARITY * 3 ) )
-#endif
-#endif
-
 /*!
 	clamp FixPoint to maxcolor in FixPoint, min(a,COLOR_MAX)
 */
@@ -831,7 +834,7 @@ static inline tVideoSample getTexel_plain ( const sInternalTexture* t, const tFi
 	size_t ofs;
 
 	ofs = ( ( ty & t->textureYMask ) >> FIX_POINT_PRE ) << t->pitchlog2;
-	ofs |= ( tx & t->textureXMask ) >> ( FIX_POINT_PRE - VIDEO_SAMPLE_GRANULARITY );
+	ofs |= ( tx & t->textureXMask ) >> ( FIX_POINT_PRE - SOFTWARE_DRIVER_2_TEXTURE_GRANULARITY );
 
 	// texel
 	return *((tVideoSample*)( (u8*) t->data + ofs ));
@@ -845,7 +848,7 @@ inline void getTexel_fix ( tFixPoint &r, tFixPoint &g, tFixPoint &b,
 	size_t ofs;
 
 	ofs = ( ((ty + FIX_POINT_ZERO_DOT_FIVE) & t->textureYMask ) >> FIX_POINT_PRE ) << t->pitchlog2;
-	ofs |= ((tx + FIX_POINT_ZERO_DOT_FIVE) & t->textureXMask ) >> ( FIX_POINT_PRE - VIDEO_SAMPLE_GRANULARITY );
+	ofs |= ((tx + FIX_POINT_ZERO_DOT_FIVE) & t->textureXMask ) >> ( FIX_POINT_PRE - SOFTWARE_DRIVER_2_TEXTURE_GRANULARITY );
 
 	// texel
 	tVideoSample t00;
@@ -865,7 +868,7 @@ inline void getTexel_fix(tFixPoint &a, tFixPoint &r, tFixPoint &g, tFixPoint &b,
 	size_t ofs;
 
 	ofs = (((ty+ FIX_POINT_ZERO_DOT_FIVE) & t->textureYMask) >> FIX_POINT_PRE) << t->pitchlog2;
-	ofs |= ((tx+ FIX_POINT_ZERO_DOT_FIVE) & t->textureXMask) >> (FIX_POINT_PRE - VIDEO_SAMPLE_GRANULARITY);
+	ofs |= ((tx+ FIX_POINT_ZERO_DOT_FIVE) & t->textureXMask) >> (FIX_POINT_PRE - SOFTWARE_DRIVER_2_TEXTURE_GRANULARITY);
 
 	// texel
 	tVideoSample t00;
@@ -886,7 +889,7 @@ static REALINLINE void getTexel_fix ( tFixPoint &a,
 	size_t ofs;
 
 	ofs = ( ((ty + FIX_POINT_ZERO_DOT_FIVE) & t->textureYMask ) >> FIX_POINT_PRE ) << t->pitchlog2;
-	ofs |= ((tx + FIX_POINT_ZERO_DOT_FIVE) & t->textureXMask ) >> ( FIX_POINT_PRE - VIDEO_SAMPLE_GRANULARITY );
+	ofs |= ((tx + FIX_POINT_ZERO_DOT_FIVE) & t->textureXMask ) >> ( FIX_POINT_PRE - SOFTWARE_DRIVER_2_TEXTURE_GRANULARITY );
 
 	// texel
 	tVideoSample t00;
@@ -913,7 +916,7 @@ static REALINLINE void getSample_texture(tFixPoint &r, tFixPoint &g, tFixPoint &
 		size_t ofs;
 
 		ofs = (((ty + FIX_POINT_ZERO_DOT_FIVE) & t->textureYMask) >> FIX_POINT_PRE) << t->pitchlog2;
-		ofs += ((tx + FIX_POINT_ZERO_DOT_FIVE) & t->textureXMask) >> (FIX_POINT_PRE - VIDEO_SAMPLE_GRANULARITY);
+		ofs += ((tx + FIX_POINT_ZERO_DOT_FIVE) & t->textureXMask) >> (FIX_POINT_PRE - SOFTWARE_DRIVER_2_TEXTURE_GRANULARITY);
 
 		// texel
 		tVideoSample t00;
@@ -937,8 +940,8 @@ static REALINLINE void getSample_texture(tFixPoint &r, tFixPoint &g, tFixPoint &
 	//wraps positive (ignoring negative)
 	o0 = (((ty)& t->textureYMask) >> FIX_POINT_PRE) << t->pitchlog2;
 	o1 = (((ty + FIX_POINT_ONE) & t->textureYMask) >> FIX_POINT_PRE) << t->pitchlog2;
-	o2 = ((tx)& t->textureXMask) >> (FIX_POINT_PRE - VIDEO_SAMPLE_GRANULARITY);
-	o3 = ((tx + FIX_POINT_ONE) & t->textureXMask) >> (FIX_POINT_PRE - VIDEO_SAMPLE_GRANULARITY);
+	o2 = ((tx)& t->textureXMask) >> (FIX_POINT_PRE - SOFTWARE_DRIVER_2_TEXTURE_GRANULARITY);
+	o3 = ((tx + FIX_POINT_ONE) & t->textureXMask) >> (FIX_POINT_PRE - SOFTWARE_DRIVER_2_TEXTURE_GRANULARITY);
 
 	t00 = *((tVideoSample*)((u8*)t->data + (o0 + o2)));
 	r00 = (t00 & MASK_R) >> SHIFT_R;
@@ -1001,7 +1004,7 @@ static REALINLINE void getSample_texture(tFixPoint &r, tFixPoint &g, tFixPoint &
 		//nearest neighbor
 		size_t ofs;
 		ofs = (((ty + FIX_POINT_ZERO_DOT_FIVE) & tex->textureYMask) >> FIX_POINT_PRE) << tex->pitchlog2;
-		ofs += ((tx + FIX_POINT_ZERO_DOT_FIVE) & tex->textureXMask) >> (FIX_POINT_PRE - VIDEO_SAMPLE_GRANULARITY);
+		ofs += ((tx + FIX_POINT_ZERO_DOT_FIVE) & tex->textureXMask) >> (FIX_POINT_PRE - SOFTWARE_DRIVER_2_TEXTURE_GRANULARITY);
 
 		tVideoSample t00;
 		t00 = *((tVideoSample*)((u8*)tex->data + ofs));
@@ -1029,8 +1032,8 @@ static REALINLINE void getSample_texture(tFixPoint &r, tFixPoint &g, tFixPoint &
 		size_t o0, o1, o2, o3;
 		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) >> (unsigned)(FIX_POINT_PRE - VIDEO_SAMPLE_GRANULARITY);
-		o3 = ((tx + FIX_POINT_ONE) & tex->textureXMask) >> (unsigned)(FIX_POINT_PRE - VIDEO_SAMPLE_GRANULARITY);
+		o2 = ((tx)& tex->textureXMask) >> (unsigned)(FIX_POINT_PRE - SOFTWARE_DRIVER_2_TEXTURE_GRANULARITY);
+		o3 = ((tx + FIX_POINT_ONE) & tex->textureXMask) >> (unsigned)(FIX_POINT_PRE - SOFTWARE_DRIVER_2_TEXTURE_GRANULARITY);
 
 		t[0] = *((tVideoSample*)((u8*)tex->data + (o0 + o2)));
 		t[1] = *((tVideoSample*)((u8*)tex->data + (o0 + o3)));
@@ -1072,8 +1075,8 @@ static REALINLINE void getSample_texture(tFixPoint &a, tFixPoint &r, tFixPoint &
 
 	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 - VIDEO_SAMPLE_GRANULARITY);
-	o3 = ((tx + FIX_POINT_ONE) & tex->textureXMask) >> (FIX_POINT_PRE - VIDEO_SAMPLE_GRANULARITY);
+	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;
@@ -1144,7 +1147,7 @@ static REALINLINE void getSample_texture(tFixPoint &r, tFixPoint &g, tFixPoint &
 {
 	size_t ofs;
 	ofs = (((ty + FIX_POINT_ZERO_DOT_FIVE) & t->textureYMask) >> FIX_POINT_PRE) << t->pitchlog2;
-	ofs += ((tx + FIX_POINT_ZERO_DOT_FIVE) & t->textureXMask) >> (FIX_POINT_PRE - VIDEO_SAMPLE_GRANULARITY);
+	ofs += ((tx + FIX_POINT_ZERO_DOT_FIVE) & t->textureXMask) >> (FIX_POINT_PRE - SOFTWARE_DRIVER_2_TEXTURE_GRANULARITY);
 
 	// texel
 	const tVideoSample t00 = *((tVideoSample*)((u8*)t->data + ofs));
@@ -1160,7 +1163,7 @@ static REALINLINE void getSample_texture(tFixPoint &a, tFixPoint &r, tFixPoint &
 {
 	size_t ofs;
 	ofs = (((ty + FIX_POINT_ZERO_DOT_FIVE) & t->textureYMask) >> FIX_POINT_PRE) << t->pitchlog2;
-	ofs += ((tx + FIX_POINT_ZERO_DOT_FIVE) & t->textureXMask) >> (FIX_POINT_PRE - VIDEO_SAMPLE_GRANULARITY);
+	ofs += ((tx + FIX_POINT_ZERO_DOT_FIVE) & t->textureXMask) >> (FIX_POINT_PRE - SOFTWARE_DRIVER_2_TEXTURE_GRANULARITY);
 
 	// texel
 	const tVideoSample t00 = *((tVideoSample*)((u8*)t->data + ofs));
@@ -1233,8 +1236,6 @@ static inline int tiny_isequal(const char *s1, const char *s2, size_t n)
 }
 
 #define tiny_istoken(a, b) tiny_isequal(a,b,sizeof(a)-1) != 0
-//! Size of a static C-style array.
-#define array_size(_arr)  ((sizeof(_arr)/sizeof(*_arr)))
 
 
 } // end namespace irr

source/Irrlicht/burning_shader_compile_fragment_start.h

@@ -40,7 +40,7 @@ void burning_shader_class::burning_shader_fragment()
 		return;
 
 	// slopes
-	const f32 invDeltaX = reciprocal_zero2(line.x[1] - line.x[0]);
+	const f32 invDeltaX = fill_step_x(line.x[1] - line.x[0]);
 
 #ifdef IPOL_Z
 	slopeZ = (line.z[1] - line.z[0]) * invDeltaX;
@@ -96,9 +96,9 @@ void burning_shader_class::burning_shader_fragment()
 	tFixPoint r1, g1, b1;
 #endif
 
-	for (s32 i = 0; i <= dx; ++i)
+	for (s32 i = 0; i <= dx; i+= SOFTWARE_DRIVER_2_STEP_X)
 	{
-		if ((0 == EdgeTestPass) & i) break;
+		if ((0 == EdgeTestPass) & (i > line.x_edgetest)) break;
 
 #ifdef CMP_Z
 		if (line.z[0] < z[i])

source/Irrlicht/burning_shader_compile_triangle.h

@@ -139,8 +139,10 @@ void burning_shader_class::drawTriangle(const s4DVertex* burning_restrict a, con
 
 #endif
 
+		line.x_edgetest = fill_convention_edge(scan.slopeX[scan.left]);
+
 		// rasterize the edge scanlines
-		for (line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -171,6 +173,7 @@ void burning_shader_class::drawTriangle(const s4DVertex* burning_restrict a, con
 #endif
 
 			// render a scanline
+			interlace_scanline
 			(this->*fragmentShader) ();
 			if (EdgeTestPass & edge_test_first_line) break;
 
@@ -265,7 +268,6 @@ void burning_shader_class::drawTriangle(const s4DVertex* burning_restrict a, con
 		yEnd = fill_convention_right(c->Pos.y);
 
 #ifdef SUBTEXEL
-
 		subPixel = ((f32)yStart) - b->Pos.y;
 
 		// correct to pixel center
@@ -299,8 +301,10 @@ void burning_shader_class::drawTriangle(const s4DVertex* burning_restrict a, con
 
 #endif
 
+		line.x_edgetest = fill_convention_edge(scan.slopeX[scan.left]);
+
 		// rasterize the edge scanlines
-		for (line.y = yStart; line.y <= yEnd; ++line.y)
+		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];
@@ -331,6 +335,7 @@ void burning_shader_class::drawTriangle(const s4DVertex* burning_restrict a, con
 #endif
 
 			// render a scanline
+			interlace_scanline
 			(this->*fragmentShader) ();
 			if (EdgeTestPass & edge_test_first_line) break;
 

tests/flyCircleAnimator.cpp

@@ -48,7 +48,7 @@ bool flyCircleAnimator(void)
 	{
 		smgr->drawAll();
 		driver->endScene();
-		result = takeScreenshotAndCompareAgainstReference(driver, "-flyCircleAnimator.png", 100);
+		result = takeScreenshotAndCompareAgainstReference(driver, "-flyCircleAnimator.png");
 	}
 
 	device->closeDevice();

tests/tests-last-passed-at.txt

@@ -1,4 +1,4 @@
 Tests finished. 72 tests of 72 passed.
 Compiled as RELEASE
-Test suite pass at GMT Sun Oct 25 15:03:26 2020
+Test suite pass at GMT Tue Nov 10 18:42:40 2020