templatized more of the scalers; also introduced template struct ColorMasks (shortens/simplifies other code a bit)

svn-id: r10489

6 files changed, 91 insertions, 79 deletions

diff --git a/common/scaler.cpp b/common/scaler.cpp
index 8f618c472c..4bb51cc84e 100644
--- a/common/scaler.cpp
+++ b/common/scaler.cpp
@@ -23,24 +23,8 @@
 #include "common/scaler/intern.h"
 
 
-// TODO: get rid of the colorMask etc. variables and instead use templates.
-// This should give a respectable boost, since variable access (i.e. memory reads)
-// in the innermost loops of our operations would work with constant data instead.
-// That should help the inliner; reduce memory access; thus improve cache efficeny
-// etc. The drawback will be that each scaler will exist twice, once for 555 and
-// once for 555, resulting in the object file being twice as big (but thanks to
-// templates, no source code would be duplicated.
-
-
 int gBitFormat = 565;
 
-uint32 colorMask = 0xF7DEF7DE;
-uint32 lowPixelMask = 0x08210821;
-uint32 qcolorMask = 0xE79CE79C;
-uint32 qlowpixelMask = 0x18631863;
-static uint32 redblueMask = redblueMask_565;
-static uint32 greenMask = greenMask_565;
-
 // RGB-to-YUV lookup table
 int   RGBtoYUV[65536];
 
@@ -62,20 +46,8 @@ static void InitLUT(uint32 BitFormat);
 
 void InitScalers(uint32 BitFormat) {
 	if (BitFormat == 565) {
-		colorMask = 0xF7DEF7DE;
-		lowPixelMask = 0x08210821;
-		qcolorMask = 0xE79CE79C;
-		qlowpixelMask = 0x18631863;
-		redblueMask = redblueMask_565;
-		greenMask = greenMask_565;
 		dotmatrix = dotmatrix_565;
 	} else if (BitFormat == 555) {
-		colorMask = 0x7BDE7BDE;
-		lowPixelMask = 0x04210421;
-		qcolorMask = 0x739C739C;
-		qlowpixelMask = 0x0C630C63;
-		redblueMask = redblueMask_555;
-		greenMask = greenMask_555;
 		dotmatrix = dotmatrix_555;
 	} else {
 		error("Unknwon bit format %d\n", BitFormat);
@@ -261,6 +233,7 @@ void AdvMame3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPi
 	}
 }
 
+template<int bitFormat>
 void TV2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch, 
 					int width, int height) {
 	const uint32 nextlineSrc = srcPitch / sizeof(uint16);
@@ -286,6 +259,7 @@ void TV2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 		q += nextlineDst << 1;
 	}
 }
+MAKE_WRAPPER(TV2x)
 
 static inline uint16 DOT_16(uint16 c, int j, int i) {
   return c - ((c >> 2) & *(dotmatrix + ((j & 3) << 2) + (i & 3)));
diff --git a/common/scaler/2xsai.cpp b/common/scaler/2xsai.cpp
index c40c08ffcc..7cf02bba47 100644
--- a/common/scaler/2xsai.cpp
+++ b/common/scaler/2xsai.cpp
@@ -22,6 +22,8 @@
 
 #include "common/scaler/intern.h"
 
+
+
 static inline int GetResult(uint32 A, uint32 B, uint32 C, uint32 D) { 
 	const bool ac = (A==C);
 	const bool bc = (B==C);
@@ -41,18 +43,25 @@ static inline int GetResult(uint32 A, uint32 B, uint32 C, uint32 D) {
 	return rmap[y][x];
 }
 
+template<int bitFormat>
 static inline uint32 INTERPOLATE(uint32 A, uint32 B) {
-	return (((A & colorMask) >> 1) + ((B & colorMask) >> 1) + (A & B & lowPixelMask));
+
+	return (((A & highBits) >> 1) + ((B & highBits) >> 1) + (A & B & lowBits));
 }
 
+template<int bitFormat>
 static inline uint32 Q_INTERPOLATE(uint32 A, uint32 B, uint32 C, uint32 D) {
-	register uint32 x = ((A & qcolorMask) >> 2) + ((B & qcolorMask) >> 2) + ((C & qcolorMask) >> 2) + ((D & qcolorMask) >> 2);
-	register uint32 y = ((A & qlowpixelMask) + (B & qlowpixelMask) + (C & qlowpixelMask) + (D & qlowpixelMask)) >> 2;
+	register uint32 x = ((A & qhighBits) >> 2) + ((B & qhighBits) >> 2) + ((C & qhighBits) >> 2) + ((D & qhighBits) >> 2);
+	register uint32 y = ((A & qlowBits) + (B & qlowBits) + (C & qlowBits) + (D & qlowBits)) >> 2;
 
-	y &= qlowpixelMask;
+	y &= qlowBits;
 	return x + y;
 }
 
+#define INTERPOLATE		INTERPOLATE<bitFormat>
+#define Q_INTERPOLATE	Q_INTERPOLATE<bitFormat>
+
+template<int bitFormat>
 void Super2xSaI(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch, int width, int height) {
 	const uint16 *bP;
 	uint16 *dP;
@@ -159,6 +168,9 @@ void Super2xSaI(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstP
 	}
 }
 
+MAKE_WRAPPER(Super2xSaI)
+
+template<int bitFormat>
 void SuperEagle(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch, int width, int height) {
 	const uint16 *bP;
 	uint16 *dP;
@@ -267,6 +279,9 @@ void SuperEagle(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstP
 	}
 }
 
+MAKE_WRAPPER(SuperEagle)
+
+template<int bitFormat>
 void _2xSaI(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch, int width, int height) {
 	const uint16 *bP;
 	uint16 *dP;
@@ -399,3 +414,5 @@ void _2xSaI(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch
 		dstPtr += dstPitch * 2;
 	}
 }
+
+MAKE_WRAPPER(_2xSaI)
diff --git a/common/scaler/aspect.cpp b/common/scaler/aspect.cpp
index f337a7fb47..7a7dc492b2 100644
--- a/common/scaler/aspect.cpp
+++ b/common/scaler/aspect.cpp
@@ -32,7 +32,7 @@
 
 #if ASPECT_MODE == kSlowAndPerfectAspectMode
 
-template<int scale>
+template<int bitFormat, int scale>
 static inline uint16 interpolate5(uint16 A, uint16 B) {
 	uint16 r = (uint16)(((A & redblueMask & 0xFF00) * scale + (B & redblueMask & 0xFF00) * (5 - scale)) / 5);
 	uint16 g = (uint16)(((A & greenMask) * scale + (B & greenMask) * (5 - scale)) / 5);
@@ -42,30 +42,32 @@ static inline uint16 interpolate5(uint16 A, uint16 B) {
 }
 
 
-template<int scale>
+template<int bitFormat, int scale>
 static inline void interpolate5Line(uint16 *dst, const uint16 *srcA, const uint16 *srcB, int width) {
 	// Accurate but slightly slower code
 	while (width--) {
-		*dst++ = interpolate5<scale>(*srcA++, *srcB++);
+		*dst++ = interpolate5<bitFormat, scale>(*srcA++, *srcB++);
 	}
 }
 #endif
 
 #if ASPECT_MODE == kFastAndNiceAspectMode
 
+template<int bitFormat>
 static inline uint32 INTERPOLATE_1_1(uint32 A, uint32 B) {
-	return (((A & colorMask) >> 1) + ((B & colorMask) >> 1) + (A & B & lowPixelMask));
+	return (((A & highBits) >> 1) + ((B & highBits) >> 1) + (A & B & lowBits));
 }
 
+template<int bitFormat>
 static inline uint32 INTERPOLATE_1_3(uint32 A, uint32 B) {
-	register uint32 x = ((A & qcolorMask) >> 2) + ((B & qcolorMask) >> 2) * 3;
-	register uint32 y = ((A & qlowpixelMask) + (B & qlowpixelMask) * 3) >> 2;
+	register uint32 x = ((A & qhighBits) >> 2) + ((B & qhighBits) >> 2) * 3;
+	register uint32 y = ((A & qlowBits) + (B & qlowBits) * 3) >> 2;
 
-	y &= qlowpixelMask;
+	y &= qlowBits;
 	return x + y;
 }
 
-template<int scale>
+template<int bitFormat, int scale>
 static inline void interpolate5Line(uint16 *dst, const uint16 *srcA, const uint16 *srcB, int width) {
 	// For efficiency reasons we blit two pixels at a time, so it is important
 	// that makeRectStretchable() guarantees that the width is even and that
@@ -86,11 +88,11 @@ static inline void interpolate5Line(uint16 *dst, const uint16 *srcA, const uint1
 	uint32 *d = (uint32 *)dst;
 	if (scale == 1) {
 		while (width--) {
-			*d++ = INTERPOLATE_1_3(*sA++, *sB++);
+			*d++ = INTERPOLATE_1_3<bitFormat>(*sA++, *sB++);
 		}
 	} else {
 		while (width--) {
-			*d++ = INTERPOLATE_1_1(*sA++, *sB++);
+			*d++ = INTERPOLATE_1_1<bitFormat>(*sA++, *sB++);
 		}
 	}
 }
@@ -141,6 +143,7 @@ void makeRectStretchable(int &x, int &y, int &w, int &h) {
  * srcY + height - 1, and it should be stretched to Y coordinates srcY
  * through real2Aspect(srcY + height - 1).
  */
+template<int bitFormat>
 int stretch200To240(uint8 *buf, uint32 pitch, int width, int height, int srcX, int srcY, int origSrcY) {
 	int maxDstY = real2Aspect(origSrcY + height - 1);
 	int y;
@@ -163,16 +166,16 @@ int stretch200To240(uint8 *buf, uint32 pitch, int width, int height, int srcX, i
 				memcpy(dstPtr, srcPtr, width * 2);
 			break;
 		case 1:
-			interpolate5Line<1>((uint16 *)dstPtr, (const uint16 *)(srcPtr - pitch), (const uint16 *)srcPtr, width);
+			interpolate5Line<bitFormat, 1>((uint16 *)dstPtr, (const uint16 *)(srcPtr - pitch), (const uint16 *)srcPtr, width);
 			break;
 		case 2:
-			interpolate5Line<2>((uint16 *)dstPtr, (const uint16 *)(srcPtr - pitch), (const uint16 *)srcPtr, width);
+			interpolate5Line<bitFormat, 2>((uint16 *)dstPtr, (const uint16 *)(srcPtr - pitch), (const uint16 *)srcPtr, width);
 			break;
 		case 3:
-			interpolate5Line<2>((uint16 *)dstPtr, (const uint16 *)srcPtr, (const uint16 *)(srcPtr - pitch), width);
+			interpolate5Line<bitFormat, 2>((uint16 *)dstPtr, (const uint16 *)srcPtr, (const uint16 *)(srcPtr - pitch), width);
 			break;
 		case 4:
-			interpolate5Line<1>((uint16 *)dstPtr, (const uint16 *)srcPtr, (const uint16 *)(srcPtr - pitch), width);
+			interpolate5Line<bitFormat, 1>((uint16 *)dstPtr, (const uint16 *)srcPtr, (const uint16 *)(srcPtr - pitch), width);
 			break;
 		}
 #endif
@@ -181,3 +184,11 @@ int stretch200To240(uint8 *buf, uint32 pitch, int width, int height, int srcX, i
 
 	return 1 + maxDstY - srcY;
 }
+
+int stretch200To240(uint8 *buf, uint32 pitch, int width, int height, int srcX, int srcY, int origSrcY) {
+	if (gBitFormat == 565)
+		return stretch200To240<565>(buf, pitch, width, height, srcX, srcY, origSrcY);
+	else // gBitFormat == 555
+		return stretch200To240<555>(buf, pitch, width, height, srcX, srcY, origSrcY);
+}
+
diff --git a/common/scaler/hq2x.cpp b/common/scaler/hq2x.cpp
index 2505914c26..de134373dd 100644
--- a/common/scaler/hq2x.cpp
+++ b/common/scaler/hq2x.cpp
@@ -1929,9 +1929,4 @@ void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 	}
 }
 
-void HQ2x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch, int width, int height) {
-	if (gBitFormat == 565)
-		HQ2x<565>(srcPtr, srcPitch, dstPtr, dstPitch, width, height);
-	else // gBitFormat == 555
-		HQ2x<555>(srcPtr, srcPitch, dstPtr, dstPitch, width, height);
-}
+MAKE_WRAPPER(HQ2x)
diff --git a/common/scaler/hq3x.cpp b/common/scaler/hq3x.cpp
index f484e2eabb..6b40f676e2 100644
--- a/common/scaler/hq3x.cpp
+++ b/common/scaler/hq3x.cpp
@@ -2927,9 +2927,4 @@ void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch,
 	}
 }
 
-void HQ3x(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch, int width, int height) {
-	if (gBitFormat == 565)
-		HQ3x<565>(srcPtr, srcPitch, dstPtr, dstPitch, width, height);
-	else // gBitFormat == 555
-		HQ3x<555>(srcPtr, srcPitch, dstPtr, dstPitch, width, height);
-}
+MAKE_WRAPPER(HQ3x)
diff --git a/common/scaler/intern.h b/common/scaler/intern.h
index 5c24a2ce53..e65e6f311a 100644
--- a/common/scaler/intern.h
+++ b/common/scaler/intern.h
@@ -30,17 +30,36 @@
 #include "common/util.h"
 
 
-static const uint32 redblueMask_565 = 0xF81F;
-static const uint32 greenMask_565 = 0x07E0;
-static const uint32 redblueMask_555 = 0x7C1F;
-static const uint32 greenMask_555 = 0x03E0;
+template<int bitFormat>
+struct ColorMasks {
+};
+
+struct ColorMasks<565> {
+	static const int highBits = 0xF7DEF7DE;
+	static const int lowBits = 0x08210821;
+	static const int qhighBits = 0xE79CE79C;
+	static const int qlowBits = 0x18631863;
+	static const int redblueMask = 0xF81F;
+	static const int greenMask = 0x07E0;
+};
+
+struct ColorMasks<555> {
+	static const int highBits = 0x04210421;
+	static const int lowBits = 0x04210421;
+	static const int qhighBits = 0x739C739C;
+	static const int qlowBits = 0x0C630C63;
+	static const int redblueMask = 0x7C1F;
+	static const int greenMask = 0x03E0;
+};
+
+#define highBits	ColorMasks<bitFormat>::highBits
+#define lowBits		ColorMasks<bitFormat>::lowBits
+#define qhighBits	ColorMasks<bitFormat>::qhighBits
+#define qlowBits	ColorMasks<bitFormat>::qlowBits
+#define redblueMask	ColorMasks<bitFormat>::redblueMask
+#define greenMask	ColorMasks<bitFormat>::greenMask
 
 
-extern uint32 colorMask;
-extern uint32 lowPixelMask;
-extern uint32 qcolorMask;
-extern uint32 qlowpixelMask;
-
 extern int gBitFormat;
 
 
@@ -50,12 +69,8 @@ extern int gBitFormat;
  */
 template<int bitFormat, int w1, int w2>
 static inline uint16 interpolate16_2(uint16 p1, uint16 p2) {
-	if (bitFormat == 565)
-		return ((((p1 & redblueMask_565) * w1 + (p2 & redblueMask_565) * w2) / (w1 + w2)) & redblueMask_565) |
-		       ((((p1 & greenMask_565) * w1 + (p2 & greenMask_565) * w2) / (w1 + w2)) & greenMask_565);
-	else // bitFormat == 555
-		return ((((p1 & redblueMask_555) * w1 + (p2 & redblueMask_555) * w2) / (w1 + w2)) & redblueMask_555) |
-		       ((((p1 & greenMask_555) * w1 + (p2 & greenMask_555) * w2) / (w1 + w2)) & greenMask_555);
+	return ((((p1 & redblueMask) * w1 + (p2 & redblueMask) * w2) / (w1 + w2)) & redblueMask) |
+	       ((((p1 & greenMask) * w1 + (p2 & greenMask) * w2) / (w1 + w2)) & greenMask);
 }
 
 /**
@@ -64,12 +79,8 @@ static inline uint16 interpolate16_2(uint16 p1, uint16 p2) {
  */
 template<int bitFormat, int w1, int w2, int w3>
 static inline uint16 interpolate16_3(uint16 p1, uint16 p2, uint16 p3) {
-	if (bitFormat == 565)
-		return ((((p1 & redblueMask_565) * w1 + (p2 & redblueMask_565) * w2 + (p3 & redblueMask_565) * w3) / (w1 + w2 + w3)) & redblueMask_565) |
-		       ((((p1 & greenMask_565) * w1 + (p2 & greenMask_565) * w2 + (p3 & greenMask_565) * w3) / (w1 + w2 + w3)) & greenMask_565);
-	else // bitFormat == 555
-		return ((((p1 & redblueMask_555) * w1 + (p2 & redblueMask_555) * w2 + (p3 & redblueMask_555) * w3) / (w1 + w2 + w3)) & redblueMask_555) |
-		       ((((p1 & greenMask_555) * w1 + (p2 & greenMask_555) * w2 + (p3 & greenMask_555) * w3) / (w1 + w2 + w3)) & greenMask_555);
+	return ((((p1 & redblueMask) * w1 + (p2 & redblueMask) * w2 + (p3 & redblueMask) * w3) / (w1 + w2 + w3)) & redblueMask) |
+		   ((((p1 & greenMask) * w1 + (p2 & greenMask) * w2 + (p3 & greenMask) * w3) / (w1 + w2 + w3)) & greenMask);
 }
 
 
@@ -97,4 +108,13 @@ static inline bool diffYUV(int yuv1, int yuv2) {
  */
 extern int   RGBtoYUV[65536];
 
+/** Auxiliary macro to simplify creating those template function wrappers. */
+#define MAKE_WRAPPER(FUNC) \
+	void FUNC(const uint8 *srcPtr, uint32 srcPitch, uint8 *dstPtr, uint32 dstPitch, int width, int height) { \
+		if (gBitFormat == 565) \
+			FUNC<565>(srcPtr, srcPitch, dstPtr, dstPitch, width, height); \
+		else \
+			FUNC<555>(srcPtr, srcPitch, dstPtr, dstPitch, width, height); \
+	}
+
 #endif