Changed rgba renderer to use a command queue system for its drawers

src/CMakeLists.txt

@@ -882,6 +882,7 @@ set( FASTMATH_PCH_SOURCES
 	r_3dfloors.cpp
 	r_bsp.cpp
 	r_draw.cpp
+	r_draw_rgba.cpp
 	r_drawt.cpp
 	r_drawt_rgba.cpp
 	r_main.cpp

src/r_draw.h

@@ -24,6 +24,13 @@
 #define __R_DRAW__
 
 #include "r_defs.h"
+#include <vector>
+
+// Spectre/Invisibility.
+#define FUZZTABLE	50
+extern "C" int 	fuzzoffset[FUZZTABLE + 1];	// [RH] +1 for the assembly routine
+extern "C" int 	fuzzpos;
+extern "C" int	fuzzviewheight;
 
 struct FColormap;
 struct ShadeConstants;
@@ -173,7 +180,6 @@ void rt_copy4cols_RGBA_c (int sx, int yl, int yh);
 
 void rt_shaded1col_RGBA_c (int hx, int sx, int yl, int yh);
 void rt_shaded4cols_RGBA_c (int sx, int yl, int yh);
-void rt_shaded4cols_RGBA_SSE (int sx, int yl, int yh);
 
 void rt_map1col_RGBA_c (int hx, int sx, int yl, int yh);
 void rt_add1col_RGBA_c (int hx, int sx, int yl, int yh);
@@ -188,11 +194,8 @@ void rt_tlatesubclamp1col_RGBA_c (int hx, int sx, int yl, int yh);
 void rt_tlaterevsubclamp1col_RGBA_c (int hx, int sx, int yl, int yh);
 
 void rt_map4cols_RGBA_c (int sx, int yl, int yh);
-void rt_map4cols_RGBA_SSE (int sx, int yl, int yh);
 void rt_add4cols_RGBA_c (int sx, int yl, int yh);
-void rt_add4cols_RGBA_SSE (int sx, int yl, int yh);
 void rt_addclamp4cols_RGBA_c (int sx, int yl, int yh);
-void rt_addclamp4cols_RGBA_SSE (int sx, int yl, int yh);
 void rt_subclamp4cols_RGBA_c (int sx, int yl, int yh);
 void rt_revsubclamp4cols_RGBA_c (int sx, int yl, int yh);
 
@@ -235,6 +238,7 @@ extern void (*rt_tlatesubclamp4cols)(int sx, int yl, int yh);
 extern void (*rt_tlaterevsubclamp4cols)(int sx, int yl, int yh);
 
 extern void (*rt_initcols)(BYTE *buffer);
+extern void (*rt_span_coverage)(int x, int start, int stop);
 
 void rt_draw4cols (int sx);
 
@@ -242,6 +246,8 @@ void rt_draw4cols (int sx);
 void rt_initcols_pal (BYTE *buffer);
 void rt_initcols_rgba (BYTE *buffer);
 
+void rt_span_coverage_pal(int x, int start, int stop);
+void rt_span_coverage_rgba(int x, int start, int stop);
 
 extern void (*R_DrawFogBoundary)(int x1, int x2, short *uclip, short *dclip);
 
@@ -277,9 +283,40 @@ void	R_DrawFuzzColumnP_RGBA_C (void);
 void	R_DrawTranslatedColumnP_RGBA_C (void);
 void	R_DrawShadedColumnP_RGBA_C (void);
 void	R_DrawSpanP_RGBA_C (void);
-void	R_DrawSpanP_RGBA_SSE (void);
 void	R_DrawSpanMaskedP_RGBA_C (void);
 
+void	R_DrawSpanTranslucentP_RGBA_C();
+void	R_DrawSpanMaskedTranslucentP_RGBA_C();
+void	R_DrawSpanAddClampP_RGBA_C();
+void	R_DrawSpanMaskedAddClampP_RGBA_C();
+void	R_FillColumnP_RGBA();
+void	R_FillAddColumn_RGBA_C();
+void	R_FillAddClampColumn_RGBA();
+void	R_FillSubClampColumn_RGBA();
+void	R_FillRevSubClampColumn_RGBA();
+void	R_DrawAddColumnP_RGBA_C();
+void	R_DrawTlatedAddColumnP_RGBA_C();
+void	R_DrawAddClampColumnP_RGBA_C();
+void	R_DrawAddClampTranslatedColumnP_RGBA_C();
+void	R_DrawSubClampColumnP_RGBA_C();
+void	R_DrawSubClampTranslatedColumnP_RGBA_C();
+void	R_DrawRevSubClampColumnP_RGBA_C();
+void	R_DrawRevSubClampTranslatedColumnP_RGBA_C();
+void	R_FillSpan_RGBA();
+void	R_DrawFogBoundary_RGBA(int x1, int x2, short *uclip, short *dclip);
+fixed_t	tmvline1_add_RGBA();
+void	tmvline4_add_RGBA();
+fixed_t	tmvline1_addclamp_RGBA();
+void	tmvline4_addclamp_RGBA();
+fixed_t	tmvline1_subclamp_RGBA();
+void	tmvline4_subclamp_RGBA();
+fixed_t	tmvline1_revsubclamp_RGBA();
+void	tmvline4_revsubclamp_RGBA();
+DWORD	vlinec1_RGBA();
+void	vlinec4_RGBA();
+DWORD	mvlinec1_RGBA();
+void	mvlinec4_RGBA();
+
 void	R_DrawSpanTranslucentP_C (void);
 void	R_DrawSpanMaskedTranslucentP_C (void);
 
@@ -403,4 +440,52 @@ void R_SetDSColorMapLight(FColormap *base_colormap, float light, int shade);
 
 void R_SetTranslationMap(lighttable_t *translation);
 
+// Wait until all drawers finished executing
+void R_FinishDrawerCommands();
+
+class DrawerThread
+{
+public:
+	int core = 0;
+	int num_cores = 1;
+
+	uint32_t dc_temp_rgbabuff_rgba[MAXHEIGHT * 4];
+	uint32_t *dc_temp_rgba;
+};
+
+class DrawerCommand
+{
+public:
+	virtual void Execute(DrawerThread *thread) = 0;
+};
+
+class DrawerCommandQueue
+{
+	enum { memorypool_size = 4 * 1024 * 1024 };
+	char memorypool[memorypool_size];
+	size_t memorypool_pos = 0;
+
+	std::vector<DrawerCommand *> commands;
+
+	static DrawerCommandQueue *Instance();
+
+public:
+	// Allocate memory valid for the duration of a command execution
+	static void* AllocMemory(size_t size);
+
+	// Queue command to be executed by drawer worker threads
+	template<typename T, typename... Types>
+	static void QueueCommand(Types &&... args)
+	{
+		void *ptr = AllocMemory(sizeof(T));
+		T *command = new (ptr)T(std::forward<Types>(args)...);
+		if (!command)
+			return;
+		Instance()->commands.push_back(command);
+	}
+
+	// Wait until all worker threads finished executing commands
+	static void Finish();
+};
+
 #endif

src/r_drawt.cpp

@@ -1019,6 +1019,14 @@ void rt_initcols_pal (BYTE *buff)
 		horizspan[y] = dc_ctspan[y] = &dc_tspans[y][0];
 }
 
+void rt_span_coverage_pal(int x, int start, int stop)
+{
+	unsigned int **tspan = &dc_ctspan[x & 3];
+	(*tspan)[0] = start;
+	(*tspan)[1] = stop;
+	*tspan += 2;
+}
+
 // Stretches a column into a temporary buffer which is later
 // drawn to the screen along with up to three other columns.
 void R_DrawColumnHorizP_C (void)

src/r_main.cpp

@@ -979,6 +979,8 @@ void R_RenderViewToCanvas (AActor *actor, DCanvas *canvas,
 	R_SetupBuffer ();
 	screen->Unlock ();
 
+	R_FinishDrawerCommands();
+
 	viewactive = savedviewactive;
 	r_swtruecolor = savedoutputformat;
 

src/r_swrenderer.cpp

@@ -165,6 +165,7 @@ void FSoftwareRenderer::RenderView(player_t *player)
 	R_RenderActorView (player->mo);
 	// [RH] Let cameras draw onto textures that were visible this frame.
 	FCanvasTextureInfo::UpdateAll ();
+	R_FinishDrawerCommands();
 }
 
 //==========================================================================

src/r_things.cpp

@@ -706,10 +706,7 @@ void R_DrawVisVoxel(vissprite_t *spr, int minslabz, int maxslabz, short *cliptop
 				}
 				else
 				{
-					unsigned int **tspan = &dc_ctspan[x & 3];
-					(*tspan)[0] = span->Start;
-					(*tspan)[1] = span->Stop - 1;
-					*tspan += 2;
+					rt_span_coverage(x, span->Start, span->Stop - 1);
 				}
 			}
 			if (!(flags & DVF_SPANSONLY) && (x & 3) == 3)