Moved RGBA draw stuff to its own header file

src/r_draw.cpp

@@ -39,6 +39,7 @@
 #include "v_palette.h"
 #include "r_data/colormaps.h"
 #include "r_plane.h"
+#include "r_draw_rgba.h"
 
 #include "gi.h"
 #include "stats.h"
@@ -2295,34 +2296,34 @@ void R_InitColumnDrawers ()
 			domvline4_saved = domvline4;
 		}
 
-		R_DrawColumnHoriz			= R_DrawColumnHorizP_RGBA_C;
+		R_DrawColumnHoriz			= R_DrawColumnHorizP_RGBA;
-		R_DrawColumn				= R_DrawColumnP_RGBA_C;
+		R_DrawColumn				= R_DrawColumnP_RGBA;
-		R_DrawFuzzColumn			= R_DrawFuzzColumnP_RGBA_C;
+		R_DrawFuzzColumn			= R_DrawFuzzColumnP_RGBA;
-		R_DrawTranslatedColumn		= R_DrawTranslatedColumnP_RGBA_C;
+		R_DrawTranslatedColumn		= R_DrawTranslatedColumnP_RGBA;
-		R_DrawShadedColumn			= R_DrawShadedColumnP_RGBA_C;
+		R_DrawShadedColumn			= R_DrawShadedColumnP_RGBA;
-		R_DrawSpanMasked			= R_DrawSpanMaskedP_RGBA_C;
+		R_DrawSpanMasked			= R_DrawSpanMaskedP_RGBA;
-		R_DrawSpan					= R_DrawSpanP_RGBA_C;
+		R_DrawSpan					= R_DrawSpanP_RGBA;
 
-		R_DrawSpanTranslucent		= R_DrawSpanTranslucentP_RGBA_C;
+		R_DrawSpanTranslucent		= R_DrawSpanTranslucentP_RGBA;
-		R_DrawSpanMaskedTranslucent = R_DrawSpanMaskedTranslucentP_RGBA_C;
+		R_DrawSpanMaskedTranslucent = R_DrawSpanMaskedTranslucentP_RGBA;
-		R_DrawSpanAddClamp			= R_DrawSpanAddClampP_RGBA_C;
+		R_DrawSpanAddClamp			= R_DrawSpanAddClampP_RGBA;
-		R_DrawSpanMaskedAddClamp	= R_DrawSpanMaskedAddClampP_RGBA_C;
+		R_DrawSpanMaskedAddClamp	= R_DrawSpanMaskedAddClampP_RGBA;
 		R_FillColumn				= R_FillColumnP_RGBA;
-		R_FillAddColumn				= R_FillAddColumn_RGBA_C;
+		R_FillAddColumn				= R_FillAddColumn_RGBA;
 		R_FillAddClampColumn		= R_FillAddClampColumn_RGBA;
 		R_FillSubClampColumn		= R_FillSubClampColumn_RGBA;
 		R_FillRevSubClampColumn		= R_FillRevSubClampColumn_RGBA;
-		R_DrawAddColumn				= R_DrawAddColumnP_RGBA_C;
+		R_DrawAddColumn				= R_DrawAddColumnP_RGBA;
-		R_DrawTlatedAddColumn		= R_DrawTlatedAddColumnP_RGBA_C;
+		R_DrawTlatedAddColumn		= R_DrawTlatedAddColumnP_RGBA;
-		R_DrawAddClampColumn		= R_DrawAddClampColumnP_RGBA_C;
+		R_DrawAddClampColumn		= R_DrawAddClampColumnP_RGBA;
-		R_DrawAddClampTranslatedColumn = R_DrawAddClampTranslatedColumnP_RGBA_C;
+		R_DrawAddClampTranslatedColumn = R_DrawAddClampTranslatedColumnP_RGBA;
-		R_DrawSubClampColumn		= R_DrawSubClampColumnP_RGBA_C;
+		R_DrawSubClampColumn		= R_DrawSubClampColumnP_RGBA;
-		R_DrawSubClampTranslatedColumn = R_DrawSubClampTranslatedColumnP_RGBA_C;
+		R_DrawSubClampTranslatedColumn = R_DrawSubClampTranslatedColumnP_RGBA;
-		R_DrawRevSubClampColumn		= R_DrawRevSubClampColumnP_RGBA_C;
+		R_DrawRevSubClampColumn		= R_DrawRevSubClampColumnP_RGBA;
-		R_DrawRevSubClampTranslatedColumn = R_DrawRevSubClampTranslatedColumnP_RGBA_C;
+		R_DrawRevSubClampTranslatedColumn = R_DrawRevSubClampTranslatedColumnP_RGBA;
 		R_FillSpan					= R_FillSpan_RGBA;
 		R_DrawFogBoundary			= R_DrawFogBoundary_RGBA;
-		R_FillColumnHoriz			= R_FillColumnHorizP_RGBA_C;
+		R_FillColumnHoriz			= R_FillColumnHorizP_RGBA;
 
 		R_DrawFogBoundary			= R_DrawFogBoundary_RGBA;
 		R_MapTiltedPlane			= R_MapTiltedPlane_RGBA;
@@ -2338,30 +2339,30 @@ void R_InitColumnDrawers ()
 		tmvline1_revsubclamp		= tmvline1_revsubclamp_RGBA;
 		tmvline4_revsubclamp		= tmvline4_revsubclamp_RGBA;
 
-		rt_copy1col					= rt_copy1col_RGBA_c;
+		rt_copy1col					= rt_copy1col_RGBA;
-		rt_copy4cols				= rt_copy4cols_RGBA_c;
+		rt_copy4cols				= rt_copy4cols_RGBA;
-		rt_map1col					= rt_map1col_RGBA_c;
+		rt_map1col					= rt_map1col_RGBA;
-		rt_map4cols					= rt_map4cols_RGBA_c;
+		rt_map4cols					= rt_map4cols_RGBA;
-		rt_shaded1col				= rt_shaded1col_RGBA_c;
+		rt_shaded1col				= rt_shaded1col_RGBA;
-		rt_shaded4cols				= rt_shaded4cols_RGBA_c;
+		rt_shaded4cols				= rt_shaded4cols_RGBA;
-		rt_add1col					= rt_add1col_RGBA_c;
+		rt_add1col					= rt_add1col_RGBA;
-		rt_add4cols					= rt_add4cols_RGBA_c;
+		rt_add4cols					= rt_add4cols_RGBA;
-		rt_addclamp1col				= rt_addclamp1col_RGBA_c;
+		rt_addclamp1col				= rt_addclamp1col_RGBA;
-		rt_addclamp4cols			= rt_addclamp4cols_RGBA_c;
+		rt_addclamp4cols			= rt_addclamp4cols_RGBA;
-		rt_subclamp1col				= rt_subclamp1col_RGBA_c;
+		rt_subclamp1col				= rt_subclamp1col_RGBA;
-		rt_revsubclamp1col			= rt_revsubclamp1col_RGBA_c;
+		rt_revsubclamp1col			= rt_revsubclamp1col_RGBA;
-		rt_tlate1col				= rt_tlate1col_RGBA_c;
+		rt_tlate1col				= rt_tlate1col_RGBA;
-		rt_tlateadd1col				= rt_tlateadd1col_RGBA_c;
+		rt_tlateadd1col				= rt_tlateadd1col_RGBA;
-		rt_tlateaddclamp1col		= rt_tlateaddclamp1col_RGBA_c;
+		rt_tlateaddclamp1col		= rt_tlateaddclamp1col_RGBA;
-		rt_tlatesubclamp1col		= rt_tlatesubclamp1col_RGBA_c;
+		rt_tlatesubclamp1col		= rt_tlatesubclamp1col_RGBA;
-		rt_tlaterevsubclamp1col		= rt_tlaterevsubclamp1col_RGBA_c;
+		rt_tlaterevsubclamp1col		= rt_tlaterevsubclamp1col_RGBA;
-		rt_subclamp4cols			= rt_subclamp4cols_RGBA_c;
+		rt_subclamp4cols			= rt_subclamp4cols_RGBA;
-		rt_revsubclamp4cols			= rt_revsubclamp4cols_RGBA_c;
+		rt_revsubclamp4cols			= rt_revsubclamp4cols_RGBA;
-		rt_tlate4cols				= rt_tlate4cols_RGBA_c;
+		rt_tlate4cols				= rt_tlate4cols_RGBA;
-		rt_tlateadd4cols			= rt_tlateadd4cols_RGBA_c;
+		rt_tlateadd4cols			= rt_tlateadd4cols_RGBA;
-		rt_tlateaddclamp4cols		= rt_tlateaddclamp4cols_RGBA_c;
+		rt_tlateaddclamp4cols		= rt_tlateaddclamp4cols_RGBA;
-		rt_tlatesubclamp4cols		= rt_tlatesubclamp4cols_RGBA_c;
+		rt_tlatesubclamp4cols		= rt_tlatesubclamp4cols_RGBA;
-		rt_tlaterevsubclamp4cols	= rt_tlaterevsubclamp4cols_RGBA_c;
+		rt_tlaterevsubclamp4cols	= rt_tlaterevsubclamp4cols_RGBA;
 		rt_initcols					= rt_initcols_rgba;
 		rt_span_coverage			= rt_span_coverage_rgba;
 

src/r_draw.h

@@ -24,11 +24,6 @@
 #define __R_DRAW__
 
 #include "r_defs.h"
-#include <vector>
-#include <memory>
-#include <thread>
-#include <mutex>
-#include <condition_variable>
 
 // Spectre/Invisibility.
 #define FUZZTABLE	50
@@ -175,39 +170,6 @@ void rt_map4cols_asm1 (int sx, int yl, int yh);
 void rt_map4cols_asm2 (int sx, int yl, int yh);
 void rt_add4cols_asm (int sx, int yl, int yh);
 void rt_addclamp4cols_asm (int sx, int yl, int yh);
-
-///
-
-void rt_copy1col_RGBA_c (int hx, int sx, int yl, int yh);
-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_map1col_RGBA_c (int hx, int sx, int yl, int yh);
-void rt_add1col_RGBA_c (int hx, int sx, int yl, int yh);
-void rt_addclamp1col_RGBA_c (int hx, int sx, int yl, int yh);
-void rt_subclamp1col_RGBA_c (int hx, int sx, int yl, int yh);
-void rt_revsubclamp1col_RGBA_c (int hx, int sx, int yl, int yh);
-
-void rt_tlate1col_RGBA_c (int hx, int sx, int yl, int yh);
-void rt_tlateadd1col_RGBA_c (int hx, int sx, int yl, int yh);
-void rt_tlateaddclamp1col_RGBA_c (int hx, int sx, int yl, int yh);
-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_add4cols_RGBA_c (int sx, int yl, int yh);
-void rt_addclamp4cols_RGBA_c (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);
-
-void rt_tlate4cols_RGBA_c (int sx, int yl, int yh);
-void rt_tlateadd4cols_RGBA_c (int sx, int yl, int yh);
-void rt_tlateaddclamp4cols_RGBA_c (int sx, int yl, int yh);
-void rt_tlatesubclamp4cols_RGBA_c (int sx, int yl, int yh);
-void rt_tlaterevsubclamp4cols_RGBA_c (int sx, int yl, int yh);
-
 }
 
 extern void (*rt_copy1col)(int hx, int sx, int yl, int yh);
@@ -247,10 +209,8 @@ void rt_draw4cols (int sx);
 
 // [RH] Preps the temporary horizontal buffer.
 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);
 
@@ -280,46 +240,6 @@ void	R_DrawSpanMaskedP_C (void);
 
 #endif
 
-void	R_DrawColumnHorizP_RGBA_C (void);
-void	R_DrawColumnP_RGBA_C (void);
-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_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);
 
@@ -348,9 +268,6 @@ void	R_FillColumnP_C (void);
 void	R_FillColumnHorizP_C (void);
 void	R_FillSpan_C (void);
 
-void	R_FillColumnHorizP_RGBA_C(void);
-void	R_FillSpan_RGBA_C(void);
-
 #ifdef X86_ASM
 #define R_SetupDrawSlab R_SetupDrawSlabA
 #define R_DrawSlab R_DrawSlabA
@@ -443,160 +360,4 @@ void R_SetDSColorMapLight(FColormap *base_colormap, float light, int shade);
 
 void R_SetTranslationMap(lighttable_t *translation);
 
-// Redirect drawer commands to worker threads
-void R_BeginDrawerCommands();
-
-// Wait until all drawers finished executing
-void R_EndDrawerCommands();
-
-class DrawerCommandQueue;
-
-class DrawerThread
-{
-public:
-	std::thread thread;
-
-	// Thread line index of this thread
-	int core = 0;
-
-	// Number of active threads
-	int num_cores = 1;
-
-	// Range of rows processed this pass
-	int pass_start_y = 0;
-	int pass_end_y = MAXHEIGHT;
-
-	uint32_t dc_temp_rgbabuff_rgba[MAXHEIGHT * 4];
-	uint32_t *dc_temp_rgba;
-
-	// Checks if a line is rendered by this thread
-	bool line_skipped_by_thread(int line)
-	{
-		return line < pass_start_y || line >= pass_end_y || line % num_cores != core;
-	}
-
-	// The number of lines to skip to reach the first line to be rendered by this thread
-	int skipped_by_thread(int first_line)
-	{
-		int pass_skip = MAX(pass_start_y - first_line, 0);
-		int core_skip = (num_cores - (first_line + pass_skip - core) % num_cores) % num_cores;
-		return pass_skip + core_skip;
-	}
-
-	// The number of lines to be rendered by this thread
-	int count_for_thread(int first_line, int count)
-	{
-		int lines_until_pass_end = MAX(pass_end_y - first_line, 0);
-		count = MIN(count, lines_until_pass_end);
-		int c = (count - skipped_by_thread(first_line) + num_cores - 1) / num_cores;
-		return MAX(c, 0);
-	}
-
-	// Calculate the dest address for the first line to be rendered by this thread
-	uint32_t *dest_for_thread(int first_line, int pitch, uint32_t *dest)
-	{
-		return dest + skipped_by_thread(first_line) * pitch;
-	}
-};
-
-class DrawerCommand
-{
-protected:
-	int dc_dest_y;
-
-public:
-	DrawerCommand()
-	{
-		dc_dest_y = static_cast<int>((dc_dest - dc_destorg) / (dc_pitch * 4));
-	}
-
-	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;
-
-	std::vector<DrawerThread> threads;
-
-	std::mutex start_mutex;
-	std::condition_variable start_condition;
-	std::vector<DrawerCommand *> active_commands;
-	bool shutdown_flag = false;
-	int run_id = 0;
-
-	std::mutex end_mutex;
-	std::condition_variable end_condition;
-	size_t finished_threads = 0;
-
-	int threaded_render = 0;
-	DrawerThread single_core_thread;
-	int num_passes = 2;
-	int rows_in_pass = 540;
-
-	void StartThreads();
-	void StopThreads();
-	void Finish();
-
-	static DrawerCommandQueue *Instance();
-
-	~DrawerCommandQueue();
-
-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)
-	{
-		auto queue = Instance();
-		if (queue->threaded_render == 0)
-		{
-			T command(std::forward<Types>(args)...);
-			command.Execute(&queue->single_core_thread);
-		}
-		else
-		{
-			void *ptr = AllocMemory(sizeof(T));
-			if (!ptr)
-				return;
-			T *command = new (ptr)T(std::forward<Types>(args)...);
-			queue->commands.push_back(command);
-		}
-	}
-	
-	// Redirects all drawing commands to worker threads until End is called
-	// Begin/End blocks can be nested.
-	static void Begin();
-
-	// End redirection and wait until all worker threads finished executing
-	static void End();
-
-	// Waits until all worker threads finished executing
-	static void WaitForWorkers();
-};
-
-class ApplySpecialColormapRGBACommand : public DrawerCommand
-{
-	BYTE *buffer;
-	int pitch;
-	int width;
-	int height;
-	int start_red;
-	int start_green;
-	int start_blue;
-	int end_red;
-	int end_green;
-	int end_blue;
-
-public:
-	ApplySpecialColormapRGBACommand(FSpecialColormap *colormap, DFrameBuffer *screen);
-	void Execute(DrawerThread *thread) override;
-};
-
 #endif

src/r_draw_rgba.cpp

@@ -37,6 +37,7 @@
 #include "v_palette.h"
 #include "r_data/colormaps.h"
 #include "r_plane.h"
+#include "r_draw_rgba.h"
 
 #include "gi.h"
 #include "stats.h"
@@ -3655,7 +3656,7 @@ void R_EndDrawerCommands()
 	DrawerCommandQueue::End();
 }
 
-void R_DrawColumnP_RGBA_C()
+void R_DrawColumnP_RGBA()
 {
 	DrawerCommandQueue::QueueCommand<DrawColumnRGBACommand>();
 }
@@ -3665,7 +3666,7 @@ void R_FillColumnP_RGBA()
 	DrawerCommandQueue::QueueCommand<FillColumnRGBACommand>();
 }
 
-void R_FillAddColumn_RGBA_C()
+void R_FillAddColumn_RGBA()
 {
 	DrawerCommandQueue::QueueCommand<FillAddColumnRGBACommand>();
 }
@@ -3685,88 +3686,88 @@ void R_FillRevSubClampColumn_RGBA()
 	DrawerCommandQueue::QueueCommand<FillRevSubClampColumnRGBACommand>();
 }
 
-void R_DrawFuzzColumnP_RGBA_C()
+void R_DrawFuzzColumnP_RGBA()
 {
 	DrawerCommandQueue::QueueCommand<DrawFuzzColumnRGBACommand>();
 	fuzzpos = (fuzzpos + dc_yh - dc_yl) % FUZZTABLE;
 }
 
-void R_DrawAddColumnP_RGBA_C()
+void R_DrawAddColumnP_RGBA()
 {
 	DrawerCommandQueue::QueueCommand<DrawAddColumnRGBACommand>();
 }
 
-void R_DrawTranslatedColumnP_RGBA_C()
+void R_DrawTranslatedColumnP_RGBA()
 {
 	DrawerCommandQueue::QueueCommand<DrawTranslatedColumnRGBACommand>();
 }
 
-void R_DrawTlatedAddColumnP_RGBA_C()
+void R_DrawTlatedAddColumnP_RGBA()
 {
 	DrawerCommandQueue::QueueCommand<DrawTlatedAddColumnRGBACommand>();
 }
 
-void R_DrawShadedColumnP_RGBA_C()
+void R_DrawShadedColumnP_RGBA()
 {
 	DrawerCommandQueue::QueueCommand<DrawShadedColumnRGBACommand>();
 }
 
-void R_DrawAddClampColumnP_RGBA_C()
+void R_DrawAddClampColumnP_RGBA()
 {
 	DrawerCommandQueue::QueueCommand<DrawAddClampColumnRGBACommand>();
 }
 
-void R_DrawAddClampTranslatedColumnP_RGBA_C()
+void R_DrawAddClampTranslatedColumnP_RGBA()
 {
 	DrawerCommandQueue::QueueCommand<DrawAddClampTranslatedColumnRGBACommand>();
 }
 
-void R_DrawSubClampColumnP_RGBA_C()
+void R_DrawSubClampColumnP_RGBA()
 {
 	DrawerCommandQueue::QueueCommand<DrawSubClampColumnRGBACommand>();
 }
 
-void R_DrawSubClampTranslatedColumnP_RGBA_C()
+void R_DrawSubClampTranslatedColumnP_RGBA()
 {
 	DrawerCommandQueue::QueueCommand<DrawSubClampTranslatedColumnRGBACommand>();
 }
 
-void R_DrawRevSubClampColumnP_RGBA_C()
+void R_DrawRevSubClampColumnP_RGBA()
 {
 	DrawerCommandQueue::QueueCommand<DrawRevSubClampColumnRGBACommand>();
 }
 
-void R_DrawRevSubClampTranslatedColumnP_RGBA_C()
+void R_DrawRevSubClampTranslatedColumnP_RGBA()
 {
 	DrawerCommandQueue::QueueCommand<DrawRevSubClampTranslatedColumnRGBACommand>();
 }
 
-void R_DrawSpanP_RGBA_C()
+void R_DrawSpanP_RGBA()
 {
 	DrawerCommandQueue::QueueCommand<DrawSpanRGBACommand>();
 }
 
-void R_DrawSpanMaskedP_RGBA_C()
+void R_DrawSpanMaskedP_RGBA()
 {
 	DrawerCommandQueue::QueueCommand<DrawSpanMaskedRGBACommand>();
 }
 
-void R_DrawSpanTranslucentP_RGBA_C()
+void R_DrawSpanTranslucentP_RGBA()
 {
 	DrawerCommandQueue::QueueCommand<DrawSpanTranslucentRGBACommand>();
 }
 
-void R_DrawSpanMaskedTranslucentP_RGBA_C()
+void R_DrawSpanMaskedTranslucentP_RGBA()
 {
 	DrawerCommandQueue::QueueCommand<DrawSpanMaskedTranslucentRGBACommand>();
 }
 
-void R_DrawSpanAddClampP_RGBA_C()
+void R_DrawSpanAddClampP_RGBA()
 {
 	DrawerCommandQueue::QueueCommand<DrawSpanAddClampRGBACommand>();
 }
 
-void R_DrawSpanMaskedAddClampP_RGBA_C()
+void R_DrawSpanMaskedAddClampP_RGBA()
 {
 	DrawerCommandQueue::QueueCommand<DrawSpanMaskedAddClampRGBACommand>();
 }

src/r_draw_rgba.h

@@ -0,0 +1,276 @@
+// Emacs style mode select	 -*- C++ -*- 
+//-----------------------------------------------------------------------------
+//
+// $Id:$
+//
+// Copyright (C) 1993-1996 by id Software, Inc.
+//
+// This source is available for distribution and/or modification
+// only under the terms of the DOOM Source Code License as
+// published by id Software. All rights reserved.
+//
+// The source is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// FITNESS FOR A PARTICULAR PURPOSE. See the DOOM Source Code License
+// for more details.
+//
+// DESCRIPTION:
+//		System specific interface stuff.
+//
+//-----------------------------------------------------------------------------
+
+
+#ifndef __R_DRAW_RGBA__
+#define __R_DRAW_RGBA__
+
+#include "r_draw.h"
+#include <vector>
+#include <memory>
+#include <thread>
+#include <mutex>
+#include <condition_variable>
+
+/////////////////////////////////////////////////////////////////////////////
+// Drawer functions:
+
+void rt_initcols_rgba(BYTE *buffer);
+void rt_span_coverage_rgba(int x, int start, int stop);
+
+void rt_copy1col_RGBA(int hx, int sx, int yl, int yh);
+void rt_copy4cols_RGBA(int sx, int yl, int yh);
+void rt_shaded1col_RGBA(int hx, int sx, int yl, int yh);
+void rt_shaded4cols_RGBA(int sx, int yl, int yh);
+void rt_map1col_RGBA(int hx, int sx, int yl, int yh);
+void rt_add1col_RGBA(int hx, int sx, int yl, int yh);
+void rt_addclamp1col_RGBA(int hx, int sx, int yl, int yh);
+void rt_subclamp1col_RGBA(int hx, int sx, int yl, int yh);
+void rt_revsubclamp1col_RGBA(int hx, int sx, int yl, int yh);
+void rt_tlate1col_RGBA(int hx, int sx, int yl, int yh);
+void rt_tlateadd1col_RGBA(int hx, int sx, int yl, int yh);
+void rt_tlateaddclamp1col_RGBA(int hx, int sx, int yl, int yh);
+void rt_tlatesubclamp1col_RGBA(int hx, int sx, int yl, int yh);
+void rt_tlaterevsubclamp1col_RGBA(int hx, int sx, int yl, int yh);
+void rt_map4cols_RGBA(int sx, int yl, int yh);
+void rt_add4cols_RGBA(int sx, int yl, int yh);
+void rt_addclamp4cols_RGBA(int sx, int yl, int yh);
+void rt_subclamp4cols_RGBA(int sx, int yl, int yh);
+void rt_revsubclamp4cols_RGBA(int sx, int yl, int yh);
+void rt_tlate4cols_RGBA(int sx, int yl, int yh);
+void rt_tlateadd4cols_RGBA(int sx, int yl, int yh);
+void rt_tlateaddclamp4cols_RGBA(int sx, int yl, int yh);
+void rt_tlatesubclamp4cols_RGBA(int sx, int yl, int yh);
+void rt_tlaterevsubclamp4cols_RGBA(int sx, int yl, int yh);
+
+void R_DrawColumnHorizP_RGBA();
+void R_DrawColumnP_RGBA();
+void R_DrawFuzzColumnP_RGBA();
+void R_DrawTranslatedColumnP_RGBA();
+void R_DrawShadedColumnP_RGBA();
+
+void R_FillColumnP_RGBA();
+void R_FillAddColumn_RGBA();
+void R_FillAddClampColumn_RGBA();
+void R_FillSubClampColumn_RGBA();
+void R_FillRevSubClampColumn_RGBA();
+void R_DrawAddColumnP_RGBA();
+void R_DrawTlatedAddColumnP_RGBA();
+void R_DrawAddClampColumnP_RGBA();
+void R_DrawAddClampTranslatedColumnP_RGBA();
+void R_DrawSubClampColumnP_RGBA();
+void R_DrawSubClampTranslatedColumnP_RGBA();
+void R_DrawRevSubClampColumnP_RGBA();
+void R_DrawRevSubClampTranslatedColumnP_RGBA();
+
+void R_DrawSpanP_RGBA(void);
+void R_DrawSpanMaskedP_RGBA(void);
+void R_DrawSpanTranslucentP_RGBA();
+void R_DrawSpanMaskedTranslucentP_RGBA();
+void R_DrawSpanAddClampP_RGBA();
+void R_DrawSpanMaskedAddClampP_RGBA();
+void R_FillSpan_RGBA();
+
+void R_DrawFogBoundary_RGBA(int x1, int x2, short *uclip, short *dclip);
+
+DWORD vlinec1_RGBA();
+void vlinec4_RGBA();
+DWORD mvlinec1_RGBA();
+void mvlinec4_RGBA();
+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();
+
+void R_FillColumnHorizP_RGBA();
+void R_FillSpan_RGBA();
+
+/////////////////////////////////////////////////////////////////////////////
+// Multithreaded rendering infrastructure:
+
+// Redirect drawer commands to worker threads
+void R_BeginDrawerCommands();
+
+// Wait until all drawers finished executing
+void R_EndDrawerCommands();
+
+struct FSpecialColormap;
+class DrawerCommandQueue;
+
+// Worker data for each thread executing drawer commands
+class DrawerThread
+{
+public:
+	std::thread thread;
+
+	// Thread line index of this thread
+	int core = 0;
+
+	// Number of active threads
+	int num_cores = 1;
+
+	// Range of rows processed this pass
+	int pass_start_y = 0;
+	int pass_end_y = MAXHEIGHT;
+
+	uint32_t dc_temp_rgbabuff_rgba[MAXHEIGHT * 4];
+	uint32_t *dc_temp_rgba;
+
+	// Checks if a line is rendered by this thread
+	bool line_skipped_by_thread(int line)
+	{
+		return line < pass_start_y || line >= pass_end_y || line % num_cores != core;
+	}
+
+	// The number of lines to skip to reach the first line to be rendered by this thread
+	int skipped_by_thread(int first_line)
+	{
+		int pass_skip = MAX(pass_start_y - first_line, 0);
+		int core_skip = (num_cores - (first_line + pass_skip - core) % num_cores) % num_cores;
+		return pass_skip + core_skip;
+	}
+
+	// The number of lines to be rendered by this thread
+	int count_for_thread(int first_line, int count)
+	{
+		int lines_until_pass_end = MAX(pass_end_y - first_line, 0);
+		count = MIN(count, lines_until_pass_end);
+		int c = (count - skipped_by_thread(first_line) + num_cores - 1) / num_cores;
+		return MAX(c, 0);
+	}
+
+	// Calculate the dest address for the first line to be rendered by this thread
+	uint32_t *dest_for_thread(int first_line, int pitch, uint32_t *dest)
+	{
+		return dest + skipped_by_thread(first_line) * pitch;
+	}
+};
+
+// Task to be executed by each worker thread
+class DrawerCommand
+{
+protected:
+	int dc_dest_y;
+
+public:
+	DrawerCommand()
+	{
+		dc_dest_y = static_cast<int>((dc_dest - dc_destorg) / (dc_pitch * 4));
+	}
+
+	virtual void Execute(DrawerThread *thread) = 0;
+};
+
+// Manages queueing up commands and executing them on worker threads
+class DrawerCommandQueue
+{
+	enum { memorypool_size = 4 * 1024 * 1024 };
+	char memorypool[memorypool_size];
+	size_t memorypool_pos = 0;
+
+	std::vector<DrawerCommand *> commands;
+
+	std::vector<DrawerThread> threads;
+
+	std::mutex start_mutex;
+	std::condition_variable start_condition;
+	std::vector<DrawerCommand *> active_commands;
+	bool shutdown_flag = false;
+	int run_id = 0;
+
+	std::mutex end_mutex;
+	std::condition_variable end_condition;
+	size_t finished_threads = 0;
+
+	int threaded_render = 0;
+	DrawerThread single_core_thread;
+	int num_passes = 2;
+	int rows_in_pass = 540;
+
+	void StartThreads();
+	void StopThreads();
+	void Finish();
+
+	static DrawerCommandQueue *Instance();
+
+	~DrawerCommandQueue();
+
+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)
+	{
+		auto queue = Instance();
+		if (queue->threaded_render == 0)
+		{
+			T command(std::forward<Types>(args)...);
+			command.Execute(&queue->single_core_thread);
+		}
+		else
+		{
+			void *ptr = AllocMemory(sizeof(T));
+			if (!ptr)
+				return;
+			T *command = new (ptr)T(std::forward<Types>(args)...);
+			queue->commands.push_back(command);
+		}
+	}
+
+	// Redirects all drawing commands to worker threads until End is called
+	// Begin/End blocks can be nested.
+	static void Begin();
+
+	// End redirection and wait until all worker threads finished executing
+	static void End();
+
+	// Waits until all worker threads finished executing
+	static void WaitForWorkers();
+};
+
+/////////////////////////////////////////////////////////////////////////////
+// Drawer commands:
+
+class ApplySpecialColormapRGBACommand : public DrawerCommand
+{
+	BYTE *buffer;
+	int pitch;
+	int width;
+	int height;
+	int start_red;
+	int start_green;
+	int start_blue;
+	int end_red;
+	int end_green;
+	int end_blue;
+
+public:
+	ApplySpecialColormapRGBACommand(FSpecialColormap *colormap, DFrameBuffer *screen);
+	void Execute(DrawerThread *thread) override;
+};
+
+#endif

src/r_drawt_rgba.cpp

@@ -42,6 +42,7 @@
 #include "r_main.h"
 #include "r_things.h"
 #include "v_video.h"
+#include "r_draw_rgba.h"
 #ifndef NO_SSE
 #include <emmintrin.h>
 #endif
@@ -1628,171 +1629,171 @@ public:
 /////////////////////////////////////////////////////////////////////////////
 
 // Copies one span at hx to the screen at sx.
-void rt_copy1col_RGBA_c (int hx, int sx, int yl, int yh)
+void rt_copy1col_RGBA (int hx, int sx, int yl, int yh)
 {
 	DrawerCommandQueue::QueueCommand<RtCopy1colRGBACommand>(hx, sx, yl, yh);
 }
 
 // Copies all four spans to the screen starting at sx.
-void rt_copy4cols_RGBA_c (int sx, int yl, int yh)
+void rt_copy4cols_RGBA (int sx, int yl, int yh)
 {
 	// To do: we could do this with SSE using __m128i
-	rt_copy1col_RGBA_c(0, sx, yl, yh);
+	rt_copy1col_RGBA(0, sx, yl, yh);
-	rt_copy1col_RGBA_c(1, sx + 1, yl, yh);
+	rt_copy1col_RGBA(1, sx + 1, yl, yh);
-	rt_copy1col_RGBA_c(2, sx + 2, yl, yh);
+	rt_copy1col_RGBA(2, sx + 2, yl, yh);
-	rt_copy1col_RGBA_c(3, sx + 3, yl, yh);
+	rt_copy1col_RGBA(3, sx + 3, yl, yh);
 }
 
 // Maps one span at hx to the screen at sx.
-void rt_map1col_RGBA_c (int hx, int sx, int yl, int yh)
+void rt_map1col_RGBA (int hx, int sx, int yl, int yh)
 {
 	DrawerCommandQueue::QueueCommand<RtMap1colRGBACommand>(hx, sx, yl, yh);
 }
 
 // Maps all four spans to the screen starting at sx.
-void rt_map4cols_RGBA_c (int sx, int yl, int yh)
+void rt_map4cols_RGBA (int sx, int yl, int yh)
 {
 	DrawerCommandQueue::QueueCommand<RtMap4colsRGBACommand>(sx, yl, yh);
 }
 
-void rt_Translate1col_RGBA_c(const BYTE *translation, int hx, int yl, int yh)
+void rt_Translate1col_RGBA(const BYTE *translation, int hx, int yl, int yh)
 {
 	DrawerCommandQueue::QueueCommand<RtTranslate1colRGBACommand>(translation, hx, yl, yh);
 }
 
-void rt_Translate4cols_RGBA_c(const BYTE *translation, int yl, int yh)
+void rt_Translate4cols_RGBA(const BYTE *translation, int yl, int yh)
 {
 	DrawerCommandQueue::QueueCommand<RtTranslate4colsRGBACommand>(translation, yl, yh);
 }
 
 // Translates one span at hx to the screen at sx.
-void rt_tlate1col_RGBA_c (int hx, int sx, int yl, int yh)
+void rt_tlate1col_RGBA (int hx, int sx, int yl, int yh)
 {
-	rt_Translate1col_RGBA_c(dc_translation, hx, yl, yh);
+	rt_Translate1col_RGBA(dc_translation, hx, yl, yh);
 	rt_map1col(hx, sx, yl, yh);
 }
 
 // Translates all four spans to the screen starting at sx.
-void rt_tlate4cols_RGBA_c (int sx, int yl, int yh)
+void rt_tlate4cols_RGBA (int sx, int yl, int yh)
 {
-	rt_Translate4cols_RGBA_c(dc_translation, yl, yh);
+	rt_Translate4cols_RGBA(dc_translation, yl, yh);
 	rt_map4cols(sx, yl, yh);
 }
 
 // Adds one span at hx to the screen at sx without clamping.
-void rt_add1col_RGBA_c (int hx, int sx, int yl, int yh)
+void rt_add1col_RGBA (int hx, int sx, int yl, int yh)
 {
 	DrawerCommandQueue::QueueCommand<RtAdd1colRGBACommand>(hx, sx, yl, yh);
 }
 
 // Adds all four spans to the screen starting at sx without clamping.
-void rt_add4cols_RGBA_c (int sx, int yl, int yh)
+void rt_add4cols_RGBA (int sx, int yl, int yh)
 {
 	DrawerCommandQueue::QueueCommand<RtAdd4colsRGBACommand>(sx, yl, yh);
 }
 
 // Translates and adds one span at hx to the screen at sx without clamping.
-void rt_tlateadd1col_RGBA_c (int hx, int sx, int yl, int yh)
+void rt_tlateadd1col_RGBA (int hx, int sx, int yl, int yh)
 {
-	rt_Translate1col_RGBA_c(dc_translation, hx, yl, yh);
+	rt_Translate1col_RGBA(dc_translation, hx, yl, yh);
 	rt_add1col(hx, sx, yl, yh);
 }
 
 // Translates and adds all four spans to the screen starting at sx without clamping.
-void rt_tlateadd4cols_RGBA_c(int sx, int yl, int yh)
+void rt_tlateadd4cols_RGBA(int sx, int yl, int yh)
 {
-	rt_Translate4cols_RGBA_c(dc_translation, yl, yh);
+	rt_Translate4cols_RGBA(dc_translation, yl, yh);
 	rt_add4cols(sx, yl, yh);
 }
 
 // Shades one span at hx to the screen at sx.
-void rt_shaded1col_RGBA_c (int hx, int sx, int yl, int yh)
+void rt_shaded1col_RGBA (int hx, int sx, int yl, int yh)
 {
 	DrawerCommandQueue::QueueCommand<RtShaded1colRGBACommand>(hx, sx, yl, yh);
 }
 
 // Shades all four spans to the screen starting at sx.
-void rt_shaded4cols_RGBA_c (int sx, int yl, int yh)
+void rt_shaded4cols_RGBA (int sx, int yl, int yh)
 {
 	DrawerCommandQueue::QueueCommand<RtShaded4colsRGBACommand>(sx, yl, yh);
 }
 
 // Adds one span at hx to the screen at sx with clamping.
-void rt_addclamp1col_RGBA_c (int hx, int sx, int yl, int yh)
+void rt_addclamp1col_RGBA (int hx, int sx, int yl, int yh)
 {
 	DrawerCommandQueue::QueueCommand<RtAddClamp1colRGBACommand>(hx, sx, yl, yh);
 }
 
 // Adds all four spans to the screen starting at sx with clamping.
-void rt_addclamp4cols_RGBA_c (int sx, int yl, int yh)
+void rt_addclamp4cols_RGBA (int sx, int yl, int yh)
 {
 	DrawerCommandQueue::QueueCommand<RtAddClamp4colsRGBACommand>(sx, yl, yh);
 }
 
 // Translates and adds one span at hx to the screen at sx with clamping.
-void rt_tlateaddclamp1col_RGBA_c (int hx, int sx, int yl, int yh)
+void rt_tlateaddclamp1col_RGBA (int hx, int sx, int yl, int yh)
 {
-	rt_Translate1col_RGBA_c(dc_translation, hx, yl, yh);
+	rt_Translate1col_RGBA(dc_translation, hx, yl, yh);
-	rt_addclamp1col_RGBA_c(hx, sx, yl, yh);
+	rt_addclamp1col_RGBA(hx, sx, yl, yh);
 }
 
 // Translates and adds all four spans to the screen starting at sx with clamping.
-void rt_tlateaddclamp4cols_RGBA_c (int sx, int yl, int yh)
+void rt_tlateaddclamp4cols_RGBA (int sx, int yl, int yh)
 {
-	rt_Translate4cols_RGBA_c(dc_translation, yl, yh);
+	rt_Translate4cols_RGBA(dc_translation, yl, yh);
 	rt_addclamp4cols(sx, yl, yh);
 }
 
 // Subtracts one span at hx to the screen at sx with clamping.
-void rt_subclamp1col_RGBA_c (int hx, int sx, int yl, int yh)
+void rt_subclamp1col_RGBA (int hx, int sx, int yl, int yh)
 {
 	DrawerCommandQueue::QueueCommand<RtSubClamp1colRGBACommand>(hx, sx, yl, yh);
 }
 
 // Subtracts all four spans to the screen starting at sx with clamping.
-void rt_subclamp4cols_RGBA_c (int sx, int yl, int yh)
+void rt_subclamp4cols_RGBA (int sx, int yl, int yh)
 {
 	DrawerCommandQueue::QueueCommand<RtSubClamp4colsRGBACommand>(sx, yl, yh);
 }
 
 // Translates and subtracts one span at hx to the screen at sx with clamping.
-void rt_tlatesubclamp1col_RGBA_c (int hx, int sx, int yl, int yh)
+void rt_tlatesubclamp1col_RGBA (int hx, int sx, int yl, int yh)
 {
-	rt_Translate1col_RGBA_c(dc_translation, hx, yl, yh);
+	rt_Translate1col_RGBA(dc_translation, hx, yl, yh);
-	rt_subclamp1col_RGBA_c(hx, sx, yl, yh);
+	rt_subclamp1col_RGBA(hx, sx, yl, yh);
 }
 
 // Translates and subtracts all four spans to the screen starting at sx with clamping.
-void rt_tlatesubclamp4cols_RGBA_c (int sx, int yl, int yh)
+void rt_tlatesubclamp4cols_RGBA (int sx, int yl, int yh)
 {
-	rt_Translate4cols_RGBA_c(dc_translation, yl, yh);
+	rt_Translate4cols_RGBA(dc_translation, yl, yh);
-	rt_subclamp4cols_RGBA_c(sx, yl, yh);
+	rt_subclamp4cols_RGBA(sx, yl, yh);
 }
 
 // Subtracts one span at hx from the screen at sx with clamping.
-void rt_revsubclamp1col_RGBA_c (int hx, int sx, int yl, int yh)
+void rt_revsubclamp1col_RGBA (int hx, int sx, int yl, int yh)
 {
 	DrawerCommandQueue::QueueCommand<RtRevSubClamp1colRGBACommand>(hx, sx, yl, yh);
 }
 
 // Subtracts all four spans from the screen starting at sx with clamping.
-void rt_revsubclamp4cols_RGBA_c (int sx, int yl, int yh)
+void rt_revsubclamp4cols_RGBA (int sx, int yl, int yh)
 {
 	DrawerCommandQueue::QueueCommand<RtRevSubClamp4colsRGBACommand>(sx, yl, yh);
 }
 
 // Translates and subtracts one span at hx from the screen at sx with clamping.
-void rt_tlaterevsubclamp1col_RGBA_c (int hx, int sx, int yl, int yh)
+void rt_tlaterevsubclamp1col_RGBA (int hx, int sx, int yl, int yh)
 {
-	rt_Translate1col_RGBA_c(dc_translation, hx, yl, yh);
+	rt_Translate1col_RGBA(dc_translation, hx, yl, yh);
-	rt_revsubclamp1col_RGBA_c(hx, sx, yl, yh);
+	rt_revsubclamp1col_RGBA(hx, sx, yl, yh);
 }
 
 // Translates and subtracts all four spans from the screen starting at sx with clamping.
-void rt_tlaterevsubclamp4cols_RGBA_c (int sx, int yl, int yh)
+void rt_tlaterevsubclamp4cols_RGBA (int sx, int yl, int yh)
 {
-	rt_Translate4cols_RGBA_c(dc_translation, yl, yh);
+	rt_Translate4cols_RGBA(dc_translation, yl, yh);
-	rt_revsubclamp4cols_RGBA_c(sx, yl, yh);
+	rt_revsubclamp4cols_RGBA(sx, yl, yh);
 }
 
 // Before each pass through a rendering loop that uses these routines,
@@ -1815,7 +1816,7 @@ void rt_span_coverage_rgba(int x, int start, int stop)
 
 // Stretches a column into a temporary buffer which is later
 // drawn to the screen along with up to three other columns.
-void R_DrawColumnHorizP_RGBA_C (void)
+void R_DrawColumnHorizP_RGBA (void)
 {
 	if (dc_count <= 0)
 		return;
@@ -1830,7 +1831,7 @@ void R_DrawColumnHorizP_RGBA_C (void)
 }
 
 // [RH] Just fills a column with a given color
-void R_FillColumnHorizP_RGBA_C (void)
+void R_FillColumnHorizP_RGBA (void)
 {
 	if (dc_count <= 0)
 		return;

src/r_main.cpp

@@ -40,6 +40,7 @@
 #include "r_segs.h"
 #include "r_3dfloors.h"
 #include "r_sky.h"
+#include "r_draw_rgba.h"
 #include "st_stuff.h"
 #include "c_cvars.h"
 #include "c_dispatch.h"

src/r_swrenderer.cpp

@@ -42,6 +42,7 @@
 #include "r_3dfloors.h"
 #include "textures/textures.h"
 #include "r_data/voxels.h"
+#include "r_draw_rgba.h"
 
 EXTERN_CVAR(Bool, r_shadercolormaps)
 

src/r_things.cpp

@@ -58,6 +58,7 @@
 #include "r_plane.h"
 #include "r_segs.h"
 #include "r_3dfloors.h"
+#include "r_draw_rgba.h"
 #include "v_palette.h"
 #include "r_data/r_translate.h"
 #include "r_data/colormaps.h"