- Begin reimplementing rgb555 again.

src/r_draw_pal.cpp

@@ -43,6 +43,9 @@
 #include "v_video.h"
 #include "r_draw_pal.h"
 
+// [SP] r_blendmode - false = rgb555 matching (ZDoom classic), true = rgb666 (refactored)
+CVAR(Bool, r_blendmode, false, CVAR_GLOBALCONFIG | CVAR_ARCHIVE)
+
 /*
 	[RH] This translucency algorithm is based on DOSDoom 0.65's, but uses
 	a 32k RGB table instead of an 8k one. At least on my machine, it's
@@ -303,19 +306,39 @@ namespace swrenderer
 		fracstep *= thread->num_cores;
 		pitch *= thread->num_cores;
 
-		do
+		if (!r_blendmode)
 		{
-			uint8_t pix = source[frac >> bits];
+			do
-			if (pix != 0)
 			{
-				uint32_t r = MIN(GPalette.BaseColors[colormap[pix]].r + GPalette.BaseColors[*dest].r, 255);
+				uint8_t pix = source[frac >> bits];
-				uint32_t g = MIN(GPalette.BaseColors[colormap[pix]].g + GPalette.BaseColors[*dest].g, 255);
+				if (pix != 0)
-				uint32_t b = MIN(GPalette.BaseColors[colormap[pix]].b + GPalette.BaseColors[*dest].b, 255);
+				{
-				*dest = RGB256k.RGB[r>>2][g>>2][b>>2];
+					uint32_t fg = fg2rgb[colormap[pix]];
-			}
+					uint32_t bg = bg2rgb[*dest];
-			frac += fracstep;
+					fg = (fg + bg) | 0x1f07c1f;
-			dest += pitch;
+					*dest = RGB32k.All[fg & (fg >> 15)];
-		} while (--count);
+				}
+				frac += fracstep;
+				dest += pitch;
+			} while (--count);
+
+		}
+		else
+		{
+			do
+			{
+				uint8_t pix = source[frac >> bits];
+				if (pix != 0)
+				{
+					uint32_t r = MIN(GPalette.BaseColors[colormap[pix]].r + GPalette.BaseColors[*dest].r, 255);
+					uint32_t g = MIN(GPalette.BaseColors[colormap[pix]].g + GPalette.BaseColors[*dest].g, 255);
+					uint32_t b = MIN(GPalette.BaseColors[colormap[pix]].b + GPalette.BaseColors[*dest].b, 255);
+					*dest = RGB256k.RGB[r>>2][g>>2][b>>2];
+				}
+				frac += fracstep;
+				dest += pitch;
+			} while (--count);
+		}
 	}
 
 	void DrawWallAdd4PalCommand::Execute(DrawerThread *thread)
@@ -341,22 +364,44 @@ namespace swrenderer
 		}
 		pitch *= thread->num_cores;
 
-		do
+		if (!r_blendmode)
 		{
-			for (int i = 0; i < 4; ++i)
+			do
 			{
-				uint8_t pix = _source[i][dc_wall_texturefrac[i] >> bits];
+				for (int i = 0; i < 4; ++i)
-				if (pix != 0)
 				{
-					uint32_t r = MIN(GPalette.BaseColors[_colormap[i][pix]].r + GPalette.BaseColors[dest[i]].r, 255);
+					uint8_t pix = _source[i][dc_wall_texturefrac[i] >> bits];
-					uint32_t g = MIN(GPalette.BaseColors[_colormap[i][pix]].g + GPalette.BaseColors[dest[i]].g, 255);
+					if (pix != 0)
-					uint32_t b = MIN(GPalette.BaseColors[_colormap[i][pix]].b + GPalette.BaseColors[dest[i]].b, 255);
+					{
-					dest[i] = RGB256k.RGB[r>>2][g>>2][b>>2];
+						uint32_t fg = fg2rgb[_colormap[i][pix]];
+						uint32_t bg = bg2rgb[dest[i]];
+						fg = (fg + bg) | 0x1f07c1f;
+						dest[i] = RGB32k.All[fg & (fg >> 15)];
+					}
+					dc_wall_texturefrac[i] += dc_wall_iscale[i];
 				}
-				dc_wall_texturefrac[i] += dc_wall_iscale[i];
+				dest += pitch;
-			}
+			} while (--count);
-			dest += pitch;
+		}
-		} while (--count);
+		else
+		{
+			do
+			{
+				for (int i = 0; i < 4; ++i)
+				{
+					uint8_t pix = _source[i][dc_wall_texturefrac[i] >> bits];
+					if (pix != 0)
+					{
+						uint32_t r = MIN(GPalette.BaseColors[_colormap[i][pix]].r + GPalette.BaseColors[dest[i]].r, 255);
+						uint32_t g = MIN(GPalette.BaseColors[_colormap[i][pix]].g + GPalette.BaseColors[dest[i]].g, 255);
+						uint32_t b = MIN(GPalette.BaseColors[_colormap[i][pix]].b + GPalette.BaseColors[dest[i]].b, 255);
+						dest[i] = RGB256k.RGB[r>>2][g>>2][b>>2];
+					}
+					dc_wall_texturefrac[i] += dc_wall_iscale[i];
+				}
+				dest += pitch;
+			} while (--count);
+		}
 	}
 
 	void DrawWallAddClamp1PalCommand::Execute(DrawerThread *thread)
@@ -379,19 +424,43 @@ namespace swrenderer
 		fracstep *= thread->num_cores;
 		pitch *= thread->num_cores;
 
-		do
+		if (!r_blendmode)
 		{
-			uint8_t pix = source[frac >> bits];
+			do
-			if (pix != 0)
 			{
-				uint32_t r = MIN(GPalette.BaseColors[colormap[pix]].r + GPalette.BaseColors[*dest].r, 255);
+				uint8_t pix = source[frac >> bits];
-				uint32_t g = MIN(GPalette.BaseColors[colormap[pix]].g + GPalette.BaseColors[*dest].g, 255);
+				if (pix != 0)
-				uint32_t b = MIN(GPalette.BaseColors[colormap[pix]].b + GPalette.BaseColors[*dest].b, 255);
+				{
-				*dest = RGB256k.RGB[r>>2][g>>2][b>>2];
+					uint32_t a = fg2rgb[colormap[pix]] + bg2rgb[*dest];
-			}
+					uint32_t b = a;
-			frac += fracstep;
+
-			dest += pitch;
+					a |= 0x01f07c1f;
-		} while (--count);
+					b &= 0x40100400;
+					a &= 0x3fffffff;
+					b = b - (b >> 5);
+					a |= b;
+					*dest = RGB32k.All[a & (a >> 15)];
+				}
+				frac += fracstep;
+				dest += pitch;
+			} while (--count);
+		}
+		else
+		{
+			do
+			{
+				uint8_t pix = source[frac >> bits];
+				if (pix != 0)
+				{
+					uint32_t r = MIN(GPalette.BaseColors[colormap[pix]].r + GPalette.BaseColors[*dest].r, 255);
+					uint32_t g = MIN(GPalette.BaseColors[colormap[pix]].g + GPalette.BaseColors[*dest].g, 255);
+					uint32_t b = MIN(GPalette.BaseColors[colormap[pix]].b + GPalette.BaseColors[*dest].b, 255);
+					*dest = RGB256k.RGB[r>>2][g>>2][b>>2];
+				}
+				frac += fracstep;
+				dest += pitch;
+			} while (--count);
+		}
 	}
 
 	void DrawWallAddClamp4PalCommand::Execute(DrawerThread *thread)

src/v_video.cpp

@@ -143,6 +143,7 @@ extern "C" {
 DWORD Col2RGB8[65][256];
 DWORD *Col2RGB8_LessPrecision[65];
 DWORD Col2RGB8_Inverse[65][256];
+ColorTable32k RGB32k;
 ColorTable256k RGB256k;
 }
 
@@ -346,6 +347,8 @@ void DCanvas::Dim (PalEntry color, float damount, int x1, int y1, int w, int h)
 	if (damount == 0.f)
 		return;
 
+	DWORD *bg2rgb;
+	DWORD fg;
 	int gap;
 	BYTE *spot;
 	int x, y;
@@ -367,6 +370,16 @@ void DCanvas::Dim (PalEntry color, float damount, int x1, int y1, int w, int h)
 		return;
 	}
 
+	{
+		int amount;
+
+		amount = (int)(damount * 64);
+		bg2rgb = Col2RGB8[64-amount];
+
+		fg = (((color.r * amount) >> 4) << 20) |
+			  ((color.g * amount) >> 4) |
+			 (((color.b * amount) >> 4) << 10);
+	}
 
 	spot = Buffer + x1 + y1*Pitch;
 	gap = Pitch - w;
@@ -376,17 +389,37 @@ void DCanvas::Dim (PalEntry color, float damount, int x1, int y1, int w, int h)
 	int dimmedcolor_r = color.r * alpha;
 	int dimmedcolor_g = color.g * alpha;
 	int dimmedcolor_b = color.b * alpha;
-	for (y = h; y != 0; y--)
+
+	if (!r_blendmode)
 	{
-		for (x = w; x != 0; x--)
+		for (y = h; y != 0; y--)
 		{
-			uint32_t r = (dimmedcolor_r + GPalette.BaseColors[*spot].r * ialpha) >> 8;
+			for (x = w; x != 0; x--)
-			uint32_t g = (dimmedcolor_g + GPalette.BaseColors[*spot].g * ialpha) >> 8;
+			{
-			uint32_t b = (dimmedcolor_b + GPalette.BaseColors[*spot].b * ialpha) >> 8;
+				DWORD bg;
-			*spot = (BYTE)RGB256k.RGB[r][g][b];
+
-			spot++;
+				bg = bg2rgb[(*spot)&0xff];
+				bg = (fg+bg) | 0x1f07c1f;
+				*spot = RGB32k.All[bg&(bg>>15)];
+				spot++;
+			}
+			spot += gap;
+		}
+	}
+	else
+	{
+		for (y = h; y != 0; y--)
+		{
+			for (x = w; x != 0; x--)
+			{
+				uint32_t r = (dimmedcolor_r + GPalette.BaseColors[*spot].r * ialpha) >> 8;
+				uint32_t g = (dimmedcolor_g + GPalette.BaseColors[*spot].g * ialpha) >> 8;
+				uint32_t b = (dimmedcolor_b + GPalette.BaseColors[*spot].b * ialpha) >> 8;
+				*spot = (BYTE)RGB256k.RGB[r][g][b];
+				spot++;
+			}
+			spot += gap;
 		}
-		spot += gap;
 	}
 }
 
@@ -658,12 +691,47 @@ static void BuildTransTable (const PalEntry *palette)
 {
 	int r, g, b;
 
+	// create the RGB555 lookup table
+	for (r = 0; r < 32; r++)
+		for (g = 0; g < 32; g++)
+			for (b = 0; b < 32; b++)
+				RGB32k.RGB[r][g][b] = ColorMatcher.Pick ((r<<3)|(r>>2), (g<<3)|(g>>2), (b<<3)|(b>>2));
 	// create the RGB666 lookup table
 	for (r = 0; r < 64; r++)
 		for (g = 0; g < 64; g++)
 			for (b = 0; b < 64; b++)
 				RGB256k.RGB[r][g][b] = ColorMatcher.Pick ((r<<2)|(r>>4), (g<<2)|(g>>4), (b<<2)|(b>>4));
 
+	int x, y;
+
+	// create the swizzled palette
+	for (x = 0; x < 65; x++)
+		for (y = 0; y < 256; y++)
+			Col2RGB8[x][y] = (((palette[y].r*x)>>4)<<20) |
+							  ((palette[y].g*x)>>4) |
+							 (((palette[y].b*x)>>4)<<10);
+
+	// create the swizzled palette with the lsb of red and blue forced to 0
+	// (for green, a 1 is okay since it never gets added into)
+	for (x = 1; x < 64; x++)
+	{
+		Col2RGB8_LessPrecision[x] = Col2RGB8_2[x-1];
+		for (y = 0; y < 256; y++)
+		{
+			Col2RGB8_2[x-1][y] = Col2RGB8[x][y] & 0x3feffbff;
+		}
+	}
+	Col2RGB8_LessPrecision[0] = Col2RGB8[0];
+	Col2RGB8_LessPrecision[64] = Col2RGB8[64];
+
+	// create the inverse swizzled palette
+	for (x = 0; x < 65; x++)
+		for (y = 0; y < 256; y++)
+		{
+			Col2RGB8_Inverse[x][y] = (((((255-palette[y].r)*x)>>4)<<20) |
+									  (((255-palette[y].g)*x)>>4) |
+									  ((((255-palette[y].b)*x)>>4)<<10)) & 0x3feffbff;
+		}
 }
 
 //==========================================================================

src/v_video.h

@@ -448,6 +448,20 @@ EXTERN_CVAR (Float, Gamma)
 
 // Translucency tables
 
+// RGB32k is a normal R5G5B5 -> palette lookup table.
+
+// Use a union so we can "overflow" without warnings.
+// Otherwise, we get stuff like this from Clang (when compiled
+// with -fsanitize=bounds) while running:
+//   src/v_video.cpp:390:12: runtime error: index 1068 out of bounds for type 'BYTE [32]'
+//   src/r_draw.cpp:273:11: runtime error: index 1057 out of bounds for type 'BYTE [32]'
+union ColorTable32k
+{
+	BYTE RGB[32][32][32];
+	BYTE All[32 *32 *32];
+};
+extern "C" ColorTable32k RGB32k;
+
 // [SP] RGB666 support
 union ColorTable256k
 {
@@ -456,6 +470,31 @@ union ColorTable256k
 };
 extern "C" ColorTable256k RGB256k;
 
+// Col2RGB8 is a pre-multiplied palette for color lookup. It is stored in a
+// special R10B10G10 format for efficient blending computation.
+//		--RRRRRrrr--BBBBBbbb--GGGGGggg--   at level 64
+//		--------rrrr------bbbb------gggg   at level 1
+extern "C" DWORD Col2RGB8[65][256];
+
+// Col2RGB8_LessPrecision is the same as Col2RGB8, but the LSB for red
+// and blue are forced to zero, so if the blend overflows, it won't spill
+// over into the next component's value.
+//		--RRRRRrrr-#BBBBBbbb-#GGGGGggg--  at level 64
+//      --------rrr#------bbb#------gggg  at level 1
+extern "C" DWORD *Col2RGB8_LessPrecision[65];
+
+// Col2RGB8_Inverse is the same as Col2RGB8_LessPrecision, except the source
+// palette has been inverted.
+extern "C" DWORD Col2RGB8_Inverse[65][256];
+
+// "Magic" numbers used during the blending:
+//		--000001111100000111110000011111	= 0x01f07c1f
+//		-0111111111011111111101111111111	= 0x3FEFFBFF
+//		-1000000000100000000010000000000	= 0x40100400
+//		------10000000001000000000100000	= 0x40100400 >> 5
+//		--11111-----11111-----11111-----	= 0x40100400 - (0x40100400 >> 5) aka "white"
+//		--111111111111111111111111111111	= 0x3FFFFFFF
+
 // Allocates buffer screens, call before R_Init.
 void V_Init (bool restart);