tiny softrasterizer speedups: change texcache to produce different pixel formats depending on 3d renderer, and optimize getline

desmume/src/OGLRender.cpp

@@ -536,7 +536,7 @@ static void setTexture(unsigned int format, unsigned int texpal)
 	}
 
 
-	TexCache_SetTexture(format, texpal);
+	TexCache_SetTexture<TexFormat_32bpp>(format, texpal);
 }
 
 

desmume/src/gfx3d.cpp

@@ -79,6 +79,11 @@ CACHE_ALIGN const u8 material_3bit_to_8bit[] = {
 	0x00, 0x24, 0x49, 0x6D, 0x92, 0xB6, 0xDB, 0xFF
 };
 
+//maybe not very precise
+CACHE_ALIGN const u8 material_3bit_to_5bit[] = {
+	0, 4, 8, 13, 17, 22, 26, 31
+};
+
 CACHE_ALIGN const u8 alpha_5bit_to_4bit[] = {
 	0x00, 0x00,
 	0x01, 0x01,

desmume/src/gfx3d.h

@@ -33,6 +33,9 @@
 //produce a 32bpp color from a ds RGB15 plus an 8bit alpha, using a table
 #define RGB15TO32(col,alpha8) ( ((alpha8)<<24) | color_15bit_to_24bit[col&0x7FFF] )
 
+//produce a 5555 32bit color from a ds RGB15 plus an 5bit alpha
+#define RGB15TO5555(col,alpha5) (((alpha5)<<24) | ((((col) & 0x7C00)>>10)<<16) | ((((col) & 0x3E0)>>5)<<8) | (((col) & 0x1F)))
+
 //produce a 24bpp color from a ds RGB15, using a table
 #define RGB15TO24_REVERSE(col) ( color_15bit_to_24bit_reverse[col&0x7FFF] )
 
@@ -183,6 +186,7 @@ extern CACHE_ALIGN u16 color_15bit_to_16bit_reverse[32768];
 extern CACHE_ALIGN u8 mixTable555[32][32][32];
 extern CACHE_ALIGN const int material_5bit_to_31bit[32];
 extern CACHE_ALIGN const u8 material_5bit_to_8bit[32];
+extern CACHE_ALIGN const u8 material_3bit_to_5bit[8];
 extern CACHE_ALIGN const u8 material_3bit_to_8bit[8];
 extern CACHE_ALIGN const u8 alpha_5bit_to_4bit[32];
 

desmume/src/rasterize.cpp

@@ -309,12 +309,7 @@ static struct Sampler
 		dowrap(iu,iv);
 
 		Fragment::Color color;
-		u32 col32 = ((u32*)textures.currentData)[(iv<<wshift)+iu];
+		color.color = ((u32*)textures.currentData)[(iv<<wshift)+iu];
-		//todo - teach texcache how to provide these already in 5555
-		col32 >>= 3;
-		col32 &= 0x1F1F1F1F;
-		color.color = col32;
-
 		return color;
 	}
 
@@ -940,18 +935,24 @@ static void SoftRastVramReconfigureSignal() {
 	TexCache_Invalidate();
 }
 
+CACHE_ALIGN static const u16 alpha_lookup[] = {
+	0x0000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,
+	0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,
+	0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,
+	0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000,0x8000};
+
 static void SoftRastGetLine(int line, u16* dst, u8* dstAlpha)
 {
 	Fragment* src = screen+((line)<<8);
 	for(int i=0;i<256;i++)
 	{
 		const bool testRenderAlpha = false;
-		u8 r = src->color.components.r;
+		const u8 r = src->color.components.r;
-		u8 g = src->color.components.g;
+		const u8 g = src->color.components.g;
-		u8 b = src->color.components.b;
+		const u8 b = src->color.components.b;
 		*dst = R5G5B5TORGB15(r,g,b);
-		if(src->color.components.a > 0) 
+
-			*dst |= 0x8000;
+		*dst |= alpha_lookup[src->color.components.a];
 		*dstAlpha = alpha_5bit_to_4bit[src->color.components.a];
 
 		if(testRenderAlpha)
@@ -971,12 +972,11 @@ static void SoftRastGetLineCaptured(int line, u16* dst) {
 	Fragment* src = screen+((line)<<8);
 	for(int i=0;i<256;i++)
 	{
-		u8 r = src->color.components.r;
+		const u8 r = src->color.components.r;
-		u8 g = src->color.components.g;
+		const u8 g = src->color.components.g;
-		u8 b = src->color.components.b;
+		const u8 b = src->color.components.b;
 		*dst = R5G5B5TORGB15(r,g,b);
-		if(src->color.components.a > 0) 
+		*dst |= alpha_lookup[src->color.components.a];
-			*dst |= 0x8000;
 		src++;
 		dst++;
 	}
@@ -1269,7 +1269,7 @@ static void SoftRastRender()
 	
 		if(needInitTexture || lastTextureFormat != poly->texParam || lastTexturePalette != poly->texPalette)
 		{
-			TexCache_SetTexture(poly->texParam,poly->texPalette);
+			TexCache_SetTexture<TexFormat_15bpp>(poly->texParam,poly->texPalette);
 			sampler.setup(poly->texParam);
 			lastTextureFormat = poly->texParam;
 			lastTexturePalette = poly->texPalette;

desmume/src/texcache.cpp

@@ -12,6 +12,7 @@
 using std::min;
 using std::max;
 
+//only dump this from ogl renderer. for now, softrasterizer creates things in an incompatible pixel format
 //#define DEBUG_DUMP_TEXTURE
 
 //This class represents a number of regions of memory which should be viewed as contiguous
@@ -150,7 +151,11 @@ static void DebugDumpTexture(int which)
 
 
 static int lastTexture = -1;
-void TexCache_SetTexture(unsigned int format, unsigned int texpal)
+
+#define CONVERT(color,alpha) ((TEXFORMAT == TexFormat_32bpp)?(RGB15TO32(color,alpha)):RGB15TO5555(color,alpha))
+
+template<TexCache_TexFormat TEXFORMAT>
+void TexCache_SetTexture(u32 format, u32 texpal)
 {
 	//for each texformat, number of palette entries
 	const int palSizes[] = {0, 32, 4, 16, 256, 0, 8, 0};
@@ -309,7 +314,8 @@ REJECT:
 	//INFO("Texture %03i - format=%08X; pal=%04X (mode %X, width %04i, height %04i)\n",i, texcache[i].frm, texcache[i].pal, texcache[i].mode, sizeX, sizeY);
 
 	//============================================================================ Texture conversion
-	u32 palZeroTransparent = (1-((format>>29)&1))*255;						// shash: CONVERT THIS TO A TABLE :)
+	const u32 opaqueColor = TEXFORMAT==TexFormat_32bpp?255:31;
+	u32 palZeroTransparent = (1-((format>>29)&1))*opaqueColor;
 
 	switch (texcache[tx].mode)
 	{
@@ -321,6 +327,9 @@ REJECT:
 				{
 					u16 c = pal[*adr&31];
 					u8 alpha = *adr>>5;
+					if(TEXFORMAT == TexFormat_15bpp)
+						*dwdst++ = RGB15TO5555(c,material_3bit_to_5bit[alpha]);
+					else
 						*dwdst++ = RGB15TO32(c,material_3bit_to_8bit[alpha]);
 					adr++;
 				}
@@ -339,19 +348,19 @@ REJECT:
 
 					bits = (*adr)&0x3;
 					c = pal[bits];
-					*dwdst++ = RGB15TO32(c,(bits == 0) ? palZeroTransparent : 255);
+					*dwdst++ = CONVERT(c,(bits == 0) ? palZeroTransparent : opaqueColor);
 
 					bits = ((*adr)>>2)&0x3;
 					c = pal[bits];
-					*dwdst++ = RGB15TO32(c,(bits == 0) ? palZeroTransparent : 255);
+					*dwdst++ = CONVERT(c,(bits == 0) ? palZeroTransparent : opaqueColor);
 
 					bits = ((*adr)>>4)&0x3;
 					c = pal[bits];
-					*dwdst++ = RGB15TO32(c,(bits == 0) ? palZeroTransparent : 255);
+					*dwdst++ = CONVERT(c,(bits == 0) ? palZeroTransparent : opaqueColor);
 
 					bits = ((*adr)>>6)&0x3;
 					c = pal[bits];
-					*dwdst++ = RGB15TO32(c,(bits == 0) ? palZeroTransparent : 255);
+					*dwdst++ = CONVERT(c,(bits == 0) ? palZeroTransparent : opaqueColor);
 
 					adr++;
 				}
@@ -369,11 +378,11 @@ REJECT:
 
 					bits = (*adr)&0xF;
 					c = pal[bits];
-					*dwdst++ = RGB15TO32(c,(bits == 0) ? palZeroTransparent : 255);
+					*dwdst++ = CONVERT(c,(bits == 0) ? palZeroTransparent : opaqueColor);
 
 					bits = ((*adr)>>4);
 					c = pal[bits];
-					*dwdst++ = RGB15TO32(c,(bits == 0) ? palZeroTransparent : 255);
+					*dwdst++ = CONVERT(c,(bits == 0) ? palZeroTransparent : opaqueColor);
 					adr++;
 				}
 			}
@@ -386,7 +395,7 @@ REJECT:
 				for(u32 x = 0; x < ms.items[j].len; ++x)
 				{
 					u16 c = pal[*adr];
-					*dwdst++ = RGB15TO32(c,(*adr == 0) ? palZeroTransparent : 255);
+					*dwdst++ = CONVERT(c,(*adr == 0) ? palZeroTransparent : opaqueColor);
 					adr++;
 				}
 			}
@@ -492,6 +501,17 @@ REJECT:
 						}
 					}
 
+					if(TEXFORMAT==TexFormat_15bpp)
+					{
+						for(int i=0;i<4;i++)
+						{
+							tmp_col[i] >>= 3;
+							tmp_col[i] &= 0x1F1F1F1F;
+						}
+					}
+
+					//TODO - this could be more precise for 32bpp mode (run it through the color separation table)
+
 					//set all 16 texels
 					for (int sy = 0; sy < 4; sy++)
 					{
@@ -520,6 +540,9 @@ REJECT:
 				{
 					u16 c = pal[*adr&0x07];
 					u8 alpha = (*adr>>3);
+					if(TEXFORMAT == TexFormat_15bpp)
+						*dwdst++ = RGB15TO5555(c,alpha);
+					else
 						*dwdst++ = RGB15TO32(c,material_5bit_to_8bit[alpha]);
 					adr++;
 				}
@@ -534,7 +557,7 @@ REJECT:
 				{
 					u16 c = map[x];
 					int alpha = ((c&0x8000)?255:0);
-					*dwdst++ = RGB15TO32(c&0x7FFF,alpha);
+					*dwdst++ = CONVERT(c&0x7FFF,alpha);
 				}
 			}
 			break;
@@ -585,3 +608,7 @@ void TexCache_Invalidate()
 
 void (*TexCache_BindTexture)(u32 texnum) = NULL;
 void (*TexCache_BindTextureData)(u32 texnum, u8* data);
+
+//these templates needed to be instantiated manually
+template void TexCache_SetTexture<TexFormat_32bpp>(u32 format, u32 texpal);
+template void TexCache_SetTexture<TexFormat_15bpp>(u32 format, u32 texpal);

desmume/src/texcache.h

@@ -3,6 +3,12 @@
 
 #include "common.h"
 
+enum TexCache_TexFormat
+{
+	TexFormat_32bpp,
+	TexFormat_15bpp
+};
+
 #define MAX_TEXTURE 500
 #ifndef NOSSE2
 struct ALIGN(16) TextureCache
@@ -36,7 +42,10 @@ extern void (*TexCache_BindTexture)(u32 texnum);
 extern void (*TexCache_BindTextureData)(u32 texnum, u8* data);
 
 void TexCache_Reset();
-void TexCache_SetTexture(unsigned int format, unsigned int texpal);
+
+template<TexCache_TexFormat format>
+void TexCache_SetTexture(u32 format, u32 texpal);
+
 void TexCache_Invalidate();
 
 extern u8 TexCache_texMAP[1024*2048*4];