Capped sky rendering

src/r_drawers.cpp

@@ -107,6 +107,7 @@ extern "C"
 	void TriDrawNormal8_11_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriDrawNormal8_12_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriDrawNormal8_13_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
+	void TriDrawNormal8_14_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriDrawNormal32_0_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriDrawNormal32_1_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriDrawNormal32_2_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
@@ -121,6 +122,7 @@ extern "C"
 	void TriDrawNormal32_11_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriDrawNormal32_12_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriDrawNormal32_13_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
+	void TriDrawNormal32_14_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriFillNormal8_0_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriFillNormal8_1_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriFillNormal8_2_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
@@ -135,6 +137,7 @@ extern "C"
 	void TriFillNormal8_11_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriFillNormal8_12_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriFillNormal8_13_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
+	void TriFillNormal8_14_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriFillNormal32_0_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriFillNormal32_1_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriFillNormal32_2_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
@@ -149,6 +152,7 @@ extern "C"
 	void TriFillNormal32_11_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriFillNormal32_12_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriFillNormal32_13_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
+	void TriFillNormal32_14_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriDrawSubsector8_0_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriDrawSubsector8_1_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriDrawSubsector8_2_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
@@ -163,6 +167,7 @@ extern "C"
 	void TriDrawSubsector8_11_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriDrawSubsector8_12_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriDrawSubsector8_13_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
+	void TriDrawSubsector8_14_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriDrawSubsector32_0_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriDrawSubsector32_1_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriDrawSubsector32_2_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
@@ -177,6 +182,7 @@ extern "C"
 	void TriDrawSubsector32_11_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriDrawSubsector32_12_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriDrawSubsector32_13_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
+	void TriDrawSubsector32_14_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriFillSubsector8_0_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriFillSubsector8_1_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriFillSubsector8_2_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
@@ -191,6 +197,7 @@ extern "C"
 	void TriFillSubsector8_11_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriFillSubsector8_12_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriFillSubsector8_13_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
+	void TriFillSubsector8_14_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriFillSubsector32_0_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriFillSubsector32_1_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriFillSubsector32_2_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
@@ -205,6 +212,7 @@ extern "C"
 	void TriFillSubsector32_11_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriFillSubsector32_12_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriFillSubsector32_13_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
+	void TriFillSubsector32_14_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriStencil_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 	void TriStencilClose_SSE2(const TriDrawTriangleArgs *, WorkerThreadData *);
 }
@@ -289,6 +297,7 @@ Drawers::Drawers()
 	TriDrawNormal8.push_back(TriDrawNormal8_11_SSE2);
 	TriDrawNormal8.push_back(TriDrawNormal8_12_SSE2);
 	TriDrawNormal8.push_back(TriDrawNormal8_13_SSE2);
+	TriDrawNormal8.push_back(TriDrawNormal8_14_SSE2);
 	TriDrawNormal32.push_back(TriDrawNormal32_0_SSE2);
 	TriDrawNormal32.push_back(TriDrawNormal32_1_SSE2);
 	TriDrawNormal32.push_back(TriDrawNormal32_2_SSE2);
@@ -303,6 +312,7 @@ Drawers::Drawers()
 	TriDrawNormal32.push_back(TriDrawNormal32_11_SSE2);
 	TriDrawNormal32.push_back(TriDrawNormal32_12_SSE2);
 	TriDrawNormal32.push_back(TriDrawNormal32_13_SSE2);
+	TriDrawNormal32.push_back(TriDrawNormal32_14_SSE2);
 	TriFillNormal8.push_back(TriFillNormal8_0_SSE2);
 	TriFillNormal8.push_back(TriFillNormal8_1_SSE2);
 	TriFillNormal8.push_back(TriFillNormal8_2_SSE2);
@@ -317,6 +327,7 @@ Drawers::Drawers()
 	TriFillNormal8.push_back(TriFillNormal8_11_SSE2);
 	TriFillNormal8.push_back(TriFillNormal8_12_SSE2);
 	TriFillNormal8.push_back(TriFillNormal8_13_SSE2);
+	TriFillNormal8.push_back(TriFillNormal8_14_SSE2);
 	TriFillNormal32.push_back(TriFillNormal32_0_SSE2);
 	TriFillNormal32.push_back(TriFillNormal32_1_SSE2);
 	TriFillNormal32.push_back(TriFillNormal32_2_SSE2);
@@ -331,6 +342,7 @@ Drawers::Drawers()
 	TriFillNormal32.push_back(TriFillNormal32_11_SSE2);
 	TriFillNormal32.push_back(TriFillNormal32_12_SSE2);
 	TriFillNormal32.push_back(TriFillNormal32_13_SSE2);
+	TriFillNormal32.push_back(TriFillNormal32_14_SSE2);
 	TriDrawSubsector8.push_back(TriDrawSubsector8_0_SSE2);
 	TriDrawSubsector8.push_back(TriDrawSubsector8_1_SSE2);
 	TriDrawSubsector8.push_back(TriDrawSubsector8_2_SSE2);
@@ -345,6 +357,7 @@ Drawers::Drawers()
 	TriDrawSubsector8.push_back(TriDrawSubsector8_11_SSE2);
 	TriDrawSubsector8.push_back(TriDrawSubsector8_12_SSE2);
 	TriDrawSubsector8.push_back(TriDrawSubsector8_13_SSE2);
+	TriDrawSubsector8.push_back(TriDrawSubsector8_14_SSE2);
 	TriDrawSubsector32.push_back(TriDrawSubsector32_0_SSE2);
 	TriDrawSubsector32.push_back(TriDrawSubsector32_1_SSE2);
 	TriDrawSubsector32.push_back(TriDrawSubsector32_2_SSE2);
@@ -359,6 +372,7 @@ Drawers::Drawers()
 	TriDrawSubsector32.push_back(TriDrawSubsector32_11_SSE2);
 	TriDrawSubsector32.push_back(TriDrawSubsector32_12_SSE2);
 	TriDrawSubsector32.push_back(TriDrawSubsector32_13_SSE2);
+	TriDrawSubsector32.push_back(TriDrawSubsector32_14_SSE2);
 	TriFillSubsector8.push_back(TriFillSubsector8_0_SSE2);
 	TriFillSubsector8.push_back(TriFillSubsector8_1_SSE2);
 	TriFillSubsector8.push_back(TriFillSubsector8_2_SSE2);
@@ -373,6 +387,7 @@ Drawers::Drawers()
 	TriFillSubsector8.push_back(TriFillSubsector8_11_SSE2);
 	TriFillSubsector8.push_back(TriFillSubsector8_12_SSE2);
 	TriFillSubsector8.push_back(TriFillSubsector8_13_SSE2);
+	TriFillSubsector8.push_back(TriFillSubsector8_14_SSE2);
 	TriFillSubsector32.push_back(TriFillSubsector32_0_SSE2);
 	TriFillSubsector32.push_back(TriFillSubsector32_1_SSE2);
 	TriFillSubsector32.push_back(TriFillSubsector32_2_SSE2);
@@ -387,6 +402,7 @@ Drawers::Drawers()
 	TriFillSubsector32.push_back(TriFillSubsector32_11_SSE2);
 	TriFillSubsector32.push_back(TriFillSubsector32_12_SSE2);
 	TriFillSubsector32.push_back(TriFillSubsector32_13_SSE2);
+	TriFillSubsector32.push_back(TriFillSubsector32_14_SSE2);
 	TriStencil = TriStencil_SSE2;
 	TriStencilClose = TriStencilClose_SSE2;
 }

src/r_drawers.h

@@ -254,10 +254,11 @@ enum class TriBlendMode
 	TranslateAdd,   // blend_add(shade(translate(fg)), bg, srcalpha, calc_blend_bgalpha(fg, destalpha))
 	TranslateSub,   // blend_sub(shade(translate(fg)), bg, srcalpha, calc_blend_bgalpha(fg, destalpha))
 	TranslateRevSub,// blend_revsub(shade(translate(fg)), bg, srcalpha, calc_blend_bgalpha(fg, destalpha))
-	AddSrcColorOneMinusSrcColor // glBlendMode(GL_SRC_COLOR, GL_ONE_MINUS_SRC_COLOR) used by GZDoom's fullbright additive sprites
+	AddSrcColorOneMinusSrcColor, // glBlendMode(GL_SRC_COLOR, GL_ONE_MINUS_SRC_COLOR) used by GZDoom's fullbright additive sprites
+	Skycap          // Fade to sky color when the V texture coordinate go beyond the [-1, 1] range
 };
 
-inline int NumTriBlendModes() { return (int)TriBlendMode::AddSrcColorOneMinusSrcColor + 1; }
+inline int NumTriBlendModes() { return (int)TriBlendMode::Skycap + 1; }
 
 class Drawers
 {

src/r_poly_sky.cpp

@@ -66,20 +66,24 @@ void PolySkyDome::Render(const TriMatrix &worldToClip)
 	RenderCapColorRow(args, frontskytex, 0, false);
 	RenderCapColorRow(args, frontskytex, rc, true);
 
+	uint32_t topcapcolor = frontskytex->GetSkyCapColor(false);
+	uint32_t bottomcapcolor = frontskytex->GetSkyCapColor(true);
+
 	for (int i = 1; i <= mRows; i++)
 	{
-		RenderRow(args, i);
+		RenderRow(args, i, topcapcolor);
-		RenderRow(args, rc + i);
+		RenderRow(args, rc + i, bottomcapcolor);
 	}
 }
 
-void PolySkyDome::RenderRow(PolyDrawArgs &args, int row)
+void PolySkyDome::RenderRow(PolyDrawArgs &args, int row, uint32_t capcolor)
 {
 	args.vinput = &mVertices[mPrimStart[row]];
 	args.vcount = mPrimStart[row + 1] - mPrimStart[row];
 	args.mode = TriangleDrawMode::Strip;
 	args.ccw = false;
-	PolyTriangleDrawer::draw(args, TriDrawVariant::DrawNormal, TriBlendMode::Copy);
+	args.uniforms.color = capcolor;
+	PolyTriangleDrawer::draw(args, TriDrawVariant::DrawNormal, TriBlendMode::Skycap);
 }
 
 void PolySkyDome::RenderCapColorRow(PolyDrawArgs &args, FTexture *skytex, int row, bool bottomCap)
@@ -129,18 +133,6 @@ void PolySkyDome::CreateSkyHemisphere(bool zflip)
 	}
 }
 
-TriVertex PolySkyDome::SetVertex(float xx, float yy, float zz, float uu, float vv)
-{
-	TriVertex v;
-	v.x = xx;
-	v.y = yy;
-	v.z = zz;
-	v.w = 1.0f;
-	v.varying[0] = uu;
-	v.varying[1] = vv;
-	return v;
-}
-
 TriVertex PolySkyDome::SetVertexXYZ(float xx, float yy, float zz, float uu, float vv)
 {
 	TriVertex v;
@@ -166,7 +158,6 @@ void PolySkyDome::SkyVertex(int r, int c, bool zflip)
 	float z = (!zflip) ? scale * height : -scale * height;
 
 	float u, v;
-	//uint32_t color = r == 0 ? 0xffffff : 0xffffffff;
 
 	// And the texture coordinates.
 	if (!zflip)	// Flipped Y is for the lower hemisphere.
@@ -184,6 +175,6 @@ void PolySkyDome::SkyVertex(int r, int c, bool zflip)
 
 	// And finally the vertex.
 	TriVertex vert;
-	vert = SetVertexXYZ(-pos.X, z - 1.f, pos.Y, u * 4.0f, v * 1.2f + 0.5f/*, color*/);
+	vert = SetVertexXYZ(-pos.X, z - 1.f, pos.Y, u * 4.0f, v * 1.2f - 0.5f);
 	mVertices.Push(vert);
 }

src/r_poly_sky.h

@@ -38,9 +38,8 @@ private:
 	void SkyVertex(int r, int c, bool yflip);
 	void CreateSkyHemisphere(bool zflip);
 	void CreateDome();
-	void RenderRow(PolyDrawArgs &args, int row);
+	void RenderRow(PolyDrawArgs &args, int row, uint32_t capcolor);
 	void RenderCapColorRow(PolyDrawArgs &args, FTexture *skytex, int row, bool bottomCap);
 
-	TriVertex SetVertex(float xx, float yy, float zz, float uu = 0, float vv = 0);
 	TriVertex SetVertexXYZ(float xx, float yy, float zz, float uu = 0, float vv = 0);
 };

tools/drawergen/fixedfunction/drawtrianglecodegen.cpp

@@ -333,8 +333,6 @@ void DrawTriangleCodegen::SetupAffineBlock()
 	{
 		AffineVaryingPosX[i] = SSAInt(AffineVaryingPosY[i] * rcpW0, false);
 		AffineVaryingStepX[i] = (SSAInt((AffineVaryingPosY[i] + gradVaryingX[i]) * rcpW1, false) - AffineVaryingPosX[i]) / q;
-		AffineVaryingPosX[i] = AffineVaryingPosX[i] << 8;
-		AffineVaryingStepX[i] = AffineVaryingStepX[i] << 8;
 	}
 }
 
@@ -780,8 +778,8 @@ SSAInt DrawTriangleCodegen::Shade8(SSAInt c)
 
 SSAVec4i DrawTriangleCodegen::ProcessPixel32(SSAVec4i bg, SSAInt *varying)
 {
-	SSAInt ufrac = varying[0];
+	SSAInt ufrac = varying[0] << 8;
-	SSAInt vfrac = varying[1];
+	SSAInt vfrac = varying[1] << 8;
 
 	SSAInt upos = ((ufrac >> 16) * textureWidth) >> 16;
 	SSAInt vpos = ((vfrac >> 16) * textureHeight) >> 16;
@@ -848,6 +846,10 @@ SSAVec4i DrawTriangleCodegen::ProcessPixel32(SSAVec4i bg, SSAInt *varying)
 		fg = Sample32(uvoffset);
 		output = blend_add_srccolor_oneminussrccolor(shade_bgra_simple(fg, currentlight), bg);
 		break;
+	case TriBlendMode::Skycap:
+		fg = Sample32(uvoffset);
+		output = FadeOut(varying[1], fg);
+		break;
 	}
 
 	return output;
@@ -867,8 +869,8 @@ SSAInt DrawTriangleCodegen::ToPal8(SSAVec4i c)
 
 SSAInt DrawTriangleCodegen::ProcessPixel8(SSAInt bg, SSAInt *varying)
 {
-	SSAInt ufrac = varying[0];
+	SSAInt ufrac = varying[0] << 8;
-	SSAInt vfrac = varying[1];
+	SSAInt vfrac = varying[1] << 8;
 
 	SSAInt upos = ((ufrac >> 16) * textureWidth) >> 16;
 	SSAInt vpos = ((vfrac >> 16) * textureHeight) >> 16;
@@ -955,11 +957,28 @@ SSAInt DrawTriangleCodegen::ProcessPixel8(SSAInt bg, SSAInt *varying)
 		output = ToPal8(blend_add_srccolor_oneminussrccolor(fg, ToBgra(bg)));
 		output = (palindex == SSAInt(0)).select(bg, output);
 		break;
+	case TriBlendMode::Skycap:
+		fg = ToBgra(Sample8(uvoffset));
+		output = ToPal8(FadeOut(varying[1], fg));
+		break;
 	}
 
 	return output;
 }
 
+SSAVec4i DrawTriangleCodegen::FadeOut(SSAInt frac, SSAVec4i fg)
+{
+	int start_fade = 2; // How fast it should fade out
+
+	SSAInt alpha_top = SSAInt::MAX(SSAInt::MIN(frac.ashr(16 - start_fade), SSAInt(256)), SSAInt(0));
+	SSAInt alpha_bottom = SSAInt::MAX(SSAInt::MIN(((2 << 24) - frac).ashr(16 - start_fade), SSAInt(256)), SSAInt(0));
+	SSAInt alpha = SSAInt::MIN(alpha_top, alpha_bottom);
+	SSAInt inv_alpha = 256 - alpha;
+
+	fg = (fg * alpha + SSAVec4i::unpack(color) * inv_alpha) / 256;
+	return fg.insert(3, 255);
+}
+
 void DrawTriangleCodegen::SetStencilBlock(SSAInt block)
 {
 	StencilBlock = stencilValues[block * 64];

tools/drawergen/fixedfunction/drawtrianglecodegen.h

@@ -58,6 +58,8 @@ private:
 	SSAVec4i ToBgra(SSAInt index);
 	SSAInt ToPal8(SSAVec4i c);
 
+	SSAVec4i FadeOut(SSAInt frac, SSAVec4i color);
+
 	void SetStencilBlock(SSAInt block);
 	void StencilClear(SSAUByte value);
 	SSAUByte StencilGetSingle();