minor optimizations to gpu and gfx3d:

- sprwin was too big (256x256 instead of 256x192)
- line render buffer initializer was a per pixel loop instead of a 256 wide memset

desmume/src/GPU.cpp

@@ -50,10 +50,10 @@
 #include "debug.h"
 #include "render3D.h"
 #include "GPU_osd.h"
-#include "debug.h"
-
-//#define CHECKSPRITES
-
+#include "debug.h"
+
+//#define CHECKSPRITES
+
 #ifdef CHECKSPRITES
 #define CHECK_SPRITE(type) \
 if (!src) {\
@@ -67,7 +67,7 @@ continue;\
 #else
 #define CHECK_SPRITE(type) if (!src) { continue; };
 #endif
-
+
 
 ARM9_struct ARM9Mem;
 
@@ -235,7 +235,7 @@ static void GPU_resortBGs(GPU *gpu)
 	struct _DISPCNT * cnt = &gpu->dispx_st->dispx_DISPCNT.bits;
 	itemsForPriority_t * item;
 
-	memset(gpu->sprWin,0, 256*256);
+	memset(gpu->sprWin,0, 256*192);
 
 	// we don't need to check for windows here...
 // if we tick boxes, invisible layers become invisible & vice versa
@@ -2039,7 +2039,7 @@ void GPU_set_DISPCAPCNT(u32 val)
 		case 2:
 			gpu->dispCapCnt.capx = 256;
 			gpu->dispCapCnt.capy = 128;
-			break;
+			break;
 		case 3:
 			gpu->dispCapCnt.capx = 256;
 			gpu->dispCapCnt.capy = 192;
@@ -2060,8 +2060,8 @@ BOOL bright_init=FALSE;
 
 // comment this if want to use formulas instead
 // #define BRIGHT_TABLES
-
-#ifdef BRIGHT_TABLES
+
+#ifdef BRIGHT_TABLES
 static void calc_bright_colors() {
 	int base = 31 ;
 	int factor;
@@ -2094,9 +2094,9 @@ static void calc_bright_colors() {
 
 #undef FORMULA_MORE
 #undef FORMULA_LESS
-}
-#endif
-
+}
+#endif
+
 static INLINE void GPU_ligne_layer(NDS_Screen * screen, u16 l)
 {
 	GPU * gpu = screen->gpu;
@@ -2118,11 +2118,8 @@ static INLINE void GPU_ligne_layer(NDS_Screen * screen, u16 l)
 				!gpu->LayersEnable[4]) return;
 
 	// init background color & priorities
-	for(int i = 0; i< 256; ++i)
-	{
-		sprPrio[i]=0xFF;
-		gpu->sprWin[l][i]=0;
-	}
+	memset(sprPrio,0xFF,256);
+	memset(&gpu->sprWin[l],0,256);
 	
 	// init pixels priorities
 	for (int i=0; i<NB_PRIORITIES; i++) {
@@ -2187,13 +2184,13 @@ static INLINE void GPU_ligne_layer(NDS_Screen * screen, u16 l)
 			}
 		}
 	}
-}
-
+}
+
 // TODO: capture emulated not fully
 static INLINE void GPU_ligne_DispCapture(u16 l)
 {
-	GPU * gpu = MainScreen.gpu;
-	struct _DISPCNT * dispCnt = &(gpu->dispx_st)->dispx_DISPCNT.bits;
+	GPU * gpu = MainScreen.gpu;
+	struct _DISPCNT * dispCnt = &(gpu->dispx_st)->dispx_DISPCNT.bits;
 
 	if (l == 0)
 	{

desmume/src/gfx3d.cpp

@@ -1089,68 +1089,72 @@ u32 gfx3d_GetGXstatus()
 	return gxstat;
 }
 
-#define NOPARAMS() \
-	for (;;) \
-	{\
-		switch (clCmd & 0xFF)\
-		{\
-			case 0x00:\
-				{\
-					if (clInd > 0)\
-					{\
-						clCmd >>= 8;\
-						clInd--;\
-						continue;\
-					}\
-					clCmd = 0;\
-					break;\
-				}\
-			case 0x11:\
-				{\
-					GFX_FIFOadd(&MMU.gfx_fifo);\
-					*(u32 *)(ARM9Mem.ARM9_REG + 0x444) = val;\
-					gfx3d_glPushMatrix();\
-					clCmd >>= 8;\
-					clInd--;\
-					continue;\
-				}\
-			case 0x15:\
-				{\
-					GFX_FIFOadd(&MMU.gfx_fifo);\
-					*(u32 *)(ARM9Mem.ARM9_REG + 0x454) = val;\
-					gfx3d_glLoadIdentity();\
-					clCmd >>= 8;\
-					clInd--;\
-					continue;\
-				}\
-			case 0x41:\
-				{\
-					GFX_FIFOadd(&MMU.gfx_fifo);\
-					*(u32 *)(ARM9Mem.ARM9_REG + 0x504) = val;\
-					gfx3d_glEnd();\
-					clCmd >>= 8;\
-					clInd--;\
-					continue;\
-				}\
-		}\
-		break;\
+void NOPARAMS(u32 val)
+{
+	for (;;)
+	{
+		switch (clCmd & 0xFF)
+		{
+			case 0x00:
+				{
+					if (clInd > 0)
+					{
+						clCmd >>= 8;
+						clInd--;
+						continue;
+					}
+					clCmd = 0;
+					break;
+				}
+			case 0x11:
+				{
+					GFX_FIFOadd(&MMU.gfx_fifo);
+					*(u32 *)(ARM9Mem.ARM9_REG + 0x444) = val;
+					gfx3d_glPushMatrix();
+					clCmd >>= 8;
+					clInd--;
+					continue;
+				}
+			case 0x15:
+				{
+					GFX_FIFOadd(&MMU.gfx_fifo);
+					*(u32 *)(ARM9Mem.ARM9_REG + 0x454) = val;
+					gfx3d_glLoadIdentity();
+					clCmd >>= 8;
+					clInd--;
+					continue;
+				}
+			case 0x41:
+				{
+					GFX_FIFOadd(&MMU.gfx_fifo);
+					*(u32 *)(ARM9Mem.ARM9_REG + 0x504) = val;
+					gfx3d_glEnd();
+					clCmd >>= 8;
+					clInd--;
+					continue;
+				}
+		}
+		break;
 	}
+}
 
-#define SETCOUNTCOMMANDS() \
-	if (!clCmd) clInd = 0;\
-	else\
-	{\
-		u32 tmp_chk = 0xFF000000;\
-		for (int t = 4; t > 0; t--)\
-		{\
-			if ((clCmd & tmp_chk))\
-			{\
-				clInd = t;\
-				break;\
-			}\
-			tmp_chk >>= 8;\
-		}\
+void SETCOUNTCOMMANDS()
+{
+	if (!clCmd) clInd = 0;
+	else
+	{
+		u32 tmp_chk = 0xFF000000;
+		for (int t = 4; t > 0; t--)
+		{
+			if ((clCmd & tmp_chk))
+			{
+				clInd = t;
+				break;
+			}
+			tmp_chk >>= 8;
+		}
 	}
+}
 
 void gfx3d_Add_Command(u32 val)
 {
@@ -1160,7 +1164,7 @@ void gfx3d_Add_Command(u32 val)
 		if (val == 0) return;
 		clCmd = val;
 		SETCOUNTCOMMANDS();
-		NOPARAMS();
+		NOPARAMS(val);
 		return;
 	}
 
@@ -1377,7 +1381,7 @@ void gfx3d_Add_Command(u32 val)
 			return;
 	}
 	GFX_FIFOadd(&MMU.gfx_fifo);
-	NOPARAMS();
+	NOPARAMS(val);
 }
 
 void gfx3d_Add_Command_Direct(u32 cmd, u32 val)