Dynamic lights on flats

src/r_draw_rgba.cpp

@@ -100,6 +100,11 @@ namespace swrenderer
 			args.flags |= DrawSpanArgs::simple_shade;
 		if (!sampler_setup(args.source, args.xbits, args.ybits, ds_source_mipmapped))
 			args.flags |= DrawSpanArgs::nearest_filter;
+
+		args.viewpos_x = dc_viewpos.X;
+		args.step_viewpos_x = dc_viewpos_step.X;
+		args.dynlights = dc_lights;
+		args.num_dynlights = dc_num_lights;
 	}
 
 	void DrawSpanLLVMCommand::Execute(DrawerThread *thread)

src/r_drawers.h

@@ -139,6 +139,10 @@ struct DrawSpanArgs
 		nearest_filter = 2
 	};
 
+	float viewpos_x, step_viewpos_x;
+	TriLight *dynlights;
+	uint32_t num_dynlights;
+
 	FString ToString();
 };
 

src/r_plane.cpp

@@ -259,11 +259,11 @@ void R_MapPlane (int y, int x1)
 	if (r_swtruecolor)
 	{
 		// Find row position in view space
-		float zspan = (float)((CenterY - y - 0.5) * InvZtoScale / planeheight);
+		float zspan = planeheight / (fabs(y + 0.5 - CenterY) / InvZtoScale);
 		dc_viewpos.X = (float)((x1 + 0.5 - CenterX) / CenterX * zspan);
 		dc_viewpos.Y = zspan;
-		dc_viewpos.Z = (float)planeheight;
+		dc_viewpos.Z = (float)((CenterY - y - 0.5) / InvZtoScale * zspan);
-		dc_viewpos_step.X = (float)(-zspan / InvZtoScale);
+		dc_viewpos_step.X = (float)(zspan / CenterX);
 
 		static TriLight lightbuffer[64 * 1024];
 		static int nextlightindex = 0;
@@ -286,9 +286,9 @@ void R_MapPlane (int y, int x1)
 			auto &light = dc_lights[dc_num_lights++];
 			light.x = (float)(lightX * ViewSin - lightY * ViewCos);
 			light.y = (float)(lightX * ViewTanCos + lightY * ViewTanSin) - dc_viewpos.Y;
-			light.z = (float)(lightZ - dc_viewpos.Z);
+			light.z = (float)lightZ - dc_viewpos.Z;
 			light.radius = 256.0f / cur_node->lightsource->GetRadius();
-			light.color = 0xff000000 | (red << 16) | (green << 8) | blue;
+			light.color = (red << 16) | (green << 8) | blue;
 
 			// Precalculate the constant part of the dot here so the drawer doesn't have to.
 			light.y = light.y * light.y + light.z * light.z;

src/r_walldraw.cpp

@@ -575,7 +575,7 @@ static void Draw1Column(int x, int y1, int y2, WallSampler &sampler, void(*draw1
 				light.y = (float)(lightX * ViewTanCos + lightY * ViewTanSin) - dc_viewpos.Y;
 				light.z = (float)lightZ;
 				light.radius = 256.0f / cur_node->lightsource->GetRadius();
-				light.color = 0xff000000 | (red << 16) | (green << 8) | blue;
+				light.color = (red << 16) | (green << 8) | blue;
 
 				// Precalculate the constant part of the dot here so the drawer doesn't have to.
 				light.x = light.x * light.x + light.y * light.y;

tools/drawergen/fixedfunction/drawspancodegen.cpp

@@ -59,6 +59,10 @@ void DrawSpanCodegen::Generate(DrawSpanVariant variant, SSAValue args)
 	SSAShort fade_blue = args[0][22].load(true);
 	SSAShort desaturate = args[0][23].load(true);
 	SSAInt flags = args[0][24].load(true);
+	start_viewpos_x = args[0][25].load(true);
+	step_viewpos_x = args[0][26].load(true);
+	dynlights = args[0][27].load(true);
+	num_dynlights = args[0][28].load(true);
 	shade_constants.light = SSAVec4i(light_blue.zext_int(), light_green.zext_int(), light_red.zext_int(), light_alpha.zext_int());
 	shade_constants.fade = SSAVec4i(fade_blue.zext_int(), fade_green.zext_int(), fade_red.zext_int(), fade_alpha.zext_int());
 	shade_constants.desaturate = desaturate.zext_int();
@@ -115,6 +119,7 @@ SSAInt DrawSpanCodegen::Loop4x(DrawSpanVariant variant, bool isSimpleShade, bool
 {
 	SSAInt sseLength = count / 4;
 	stack_index.store(SSAInt(0));
+	stack_viewpos_x.store(start_viewpos_x);
 	{
 		SSAForBlock loop;
 		SSAInt index = stack_index.load();
@@ -136,9 +141,11 @@ SSAInt DrawSpanCodegen::Loop4x(DrawSpanVariant variant, bool isSimpleShade, bool
 		{
 			SSAInt xfrac = stack_xfrac.load();
 			SSAInt yfrac = stack_yfrac.load();
+			viewpos_x = stack_viewpos_x.load();
 
 			colors[i] = Blend(Shade(Sample(xfrac, yfrac, isNearestFilter, is64x64), isSimpleShade), bgcolors[i], variant);
 
+			stack_viewpos_x.store(viewpos_x + step_viewpos_x);
 			stack_xfrac.store(xfrac + xstep);
 			stack_yfrac.store(yfrac + ystep);
 		}
@@ -158,6 +165,7 @@ void DrawSpanCodegen::Loop(SSAInt start, DrawSpanVariant variant, bool isSimpleS
 	{
 		SSAForBlock loop;
 		SSAInt index = stack_index.load();
+		viewpos_x = stack_viewpos_x.load();
 		loop.loop_block(index < count);
 
 		SSAInt xfrac = stack_xfrac.load();
@@ -167,6 +175,7 @@ void DrawSpanCodegen::Loop(SSAInt start, DrawSpanVariant variant, bool isSimpleS
 		SSAVec4i color = Blend(Shade(Sample(xfrac, yfrac, isNearestFilter, is64x64), isSimpleShade), bgcolor, variant);
 		data[index * 4].store_vec4ub(color);
 
+		stack_viewpos_x.store(viewpos_x + step_viewpos_x);
 		stack_index.store(index.add(SSAInt(1), true, true));
 		stack_xfrac.store(xfrac + xstep);
 		stack_yfrac.store(yfrac + ystep);
@@ -222,10 +231,41 @@ SSAVec4i DrawSpanCodegen::SampleLinear(SSAUBytePtr texture, SSAInt xfrac, SSAInt
 
 SSAVec4i DrawSpanCodegen::Shade(SSAVec4i fg, bool isSimpleShade)
 {
+	SSAVec4i c;
 	if (isSimpleShade)
-		return shade_bgra_simple(fg, light);
+		c = shade_bgra_simple(fg, light);
 	else
-		return shade_bgra_advanced(fg, light, shade_constants);
+		c = shade_bgra_advanced(fg, light, shade_constants);
+
+	stack_lit_color.store(c);
+	stack_light_index.store(SSAInt(0));
+
+	SSAForBlock block;
+	SSAInt light_index = stack_light_index.load();
+	SSAVec4i lit_color = stack_lit_color.load();
+	block.loop_block(light_index < num_dynlights);
+	{
+		SSAVec4i light_color = SSAUBytePtr(dynlights[light_index][0].v).load_vec4ub(true);
+		SSAFloat light_x = dynlights[light_index][1].load(true);
+		SSAFloat light_y = dynlights[light_index][2].load(true);
+		//SSAFloat light_z = dynlights[light_index][3].load(true);
+		SSAFloat light_rcp_radius = dynlights[light_index][4].load(true);
+
+		// L = light-pos
+		// dist = sqrt(dot(L, L))
+		// attenuation = 1 - MIN(dist * (1/radius), 1)
+		SSAFloat Lyz2 = light_y; // L.y*L.y + L.z*L.z
+		SSAFloat Lx = light_x - viewpos_x;
+		SSAFloat dist = SSAFloat::sqrt(Lyz2 + Lx * Lx);
+		SSAInt attenuation = SSAInt(SSAFloat(256.0f) - SSAFloat::MIN(dist * light_rcp_radius, SSAFloat(256.0f)), true);
+		SSAVec4i contribution = (light_color * fg * attenuation) >> 16;
+
+		stack_lit_color.store(lit_color + contribution);
+		stack_light_index.store(light_index + 1);
+	}
+	block.end_block();
+
+	return stack_lit_color.load();
 }
 
 SSAVec4i DrawSpanCodegen::Blend(SSAVec4i fg, SSAVec4i bg, DrawSpanVariant variant)

tools/drawergen/fixedfunction/drawspancodegen.h

@@ -49,7 +49,9 @@ private:
 	SSAVec4i Shade(SSAVec4i fg, bool isSimpleShade);
 	SSAVec4i Blend(SSAVec4i fg, SSAVec4i bg, DrawSpanVariant variant);
 
-	SSAStack<SSAInt> stack_index, stack_xfrac, stack_yfrac;
+	SSAStack<SSAInt> stack_index, stack_xfrac, stack_yfrac, stack_light_index;
+	SSAStack<SSAVec4i> stack_lit_color;
+	SSAStack<SSAFloat> stack_viewpos_x;
 
 	SSAUBytePtr destorg;
 	SSAUBytePtr source;
@@ -73,4 +75,9 @@ private:
 	SSABool is_simple_shade;
 	SSABool is_nearest_filter;
 	SSAShadeConstants shade_constants;
+
+	SSAFloat start_viewpos_x, step_viewpos_x;
+	SSAValue dynlights; // TriLight*
+	SSAInt num_dynlights;
+	SSAFloat viewpos_x;
 };

tools/drawergen/llvmdrawers.cpp

@@ -265,6 +265,10 @@ llvm::Type *LLVMDrawers::GetDrawSpanArgsStruct(llvm::LLVMContext &context)
 	elements.push_back(llvm::Type::getInt16Ty(context)); // uint16_t fade_blue;
 	elements.push_back(llvm::Type::getInt16Ty(context)); // uint16_t desaturate;
 	elements.push_back(llvm::Type::getInt32Ty(context)); // uint32_t flags;
+	elements.push_back(llvm::Type::getFloatTy(context)); // float viewpos_x;
+	elements.push_back(llvm::Type::getFloatTy(context)); // float step_viewpos_x;
+	elements.push_back(GetTriLightStruct(context));      // TriLight *dynlights;
+	elements.push_back(llvm::Type::getInt32Ty(context)); // uint32_t num_dynlights;
 	DrawSpanArgsStruct = llvm::StructType::create(context, elements, "DrawSpanArgs", false)->getPointerTo();
 	return DrawSpanArgsStruct;
 }