Backport colour operations from Snes9x 1.60

6 files changed, 100 insertions, 125 deletions

diff --git a/Makefile b/Makefile
index 97f340e..57ed00f 100644
--- a/Makefile
+++ b/Makefile
@@ -1,5 +1,6 @@
-DEBUG=0
-LAGFIX=1
+DEBUG              = 0
+LAGFIX             = 1
+USE_OLD_COLOUR_OPS = 0
 TARGET_NAME = snes9x2002
 
 ifeq ($(platform),)
@@ -219,10 +220,9 @@ else ifeq ($(platform), gcw0)
    AR = /opt/gcw0-toolchain/usr/bin/mipsel-linux-ar
    fpic := -fPIC
    SHARED := -shared -Wl,--version-script=libretro/link.T -Wl,--no-undefined
-   DEFINES += -DFRAME_SKIP
-   CFLAGS += -std=c99 -ffast-math -march=mips32 -mtune=mips32r2 -mhard-float
+   CFLAGS += -std=c99 -fomit-frame-pointer -ffast-math -march=mips32 -mtune=mips32r2 -mhard-float
    CFLAGS += -fno-builtin -fno-exceptions
-   CFLAGS += -DPATH_MAX=256
+   CFLAGS += -DPATH_MAX=256 -DFAST_LSB_WORD_ACCESS
 
 # Windows MSVC 2010 x86
 else ifeq ($(platform), windows_msvc2010_x86)
diff --git a/Makefile.common b/Makefile.common
index b7078e2..5bde095 100644
--- a/Makefile.common
+++ b/Makefile.common
@@ -124,4 +124,8 @@ DEFINES += -DLAGFIX
 endif
 endif
 
+ifeq ($(USE_OLD_COLOUR_OPS),1)
+	FLAGS += -DUSE_OLD_COLOUR_OPS
+endif
+
 COMMON_DEFINES += $(CODE_DEFINES) $(WARNINGS_DEFINES) $(fpic)
diff --git a/src/gfx.c b/src/gfx.c
index 3842d1d..fa12da8 100644
--- a/src/gfx.c
+++ b/src/gfx.c
@@ -362,6 +362,8 @@ bool8_32 S9xGraphicsInit(void)
    DrawHiResClippedTilePtr = DrawHiResClippedTile16;
    S9xFixColourBrightness();
 
+#if defined(USE_OLD_COLOUR_OPS)
+   /* Pre-1.60 colour operations */
    if (!(GFX.X2 = (uint16*) malloc(sizeof(uint16) * 0x10000)))
       return (FALSE);
 
@@ -476,6 +478,10 @@ bool8_32 S9xGraphicsInit(void)
       }
    }
 #endif
+#else
+   if (!(GFX.ZERO = (uint16_t*) malloc(sizeof(uint16_t) * 0x10000)))
+      return false;
+#endif
 
    // Build a lookup table that if the top bit of the color value is zero
    // then the value is zero, otherwise its just the value.
@@ -514,6 +520,8 @@ bool8_32 S9xGraphicsInit(void)
 void S9xGraphicsDeinit(void)
 {
    // Free any memory allocated in S9xGraphicsInit
+#if defined(USE_OLD_COLOUR_OPS)
+   /* Pre-1.60 colour operations */
    if (GFX.X2)
    {
       free((char*) GFX.X2);
@@ -524,6 +532,7 @@ void S9xGraphicsDeinit(void)
       free((char*) GFX.ZERO_OR_X2);
       GFX.ZERO_OR_X2 = NULL;
    }
+#endif
    if (GFX.ZERO)
    {
       free((char*) GFX.ZERO);
diff --git a/src/gfx.h b/src/gfx.h
index f873531..362da1d 100644
--- a/src/gfx.h
+++ b/src/gfx.h
@@ -64,8 +64,11 @@ typedef struct
 
    // Setup in call to S9xGraphicsInit()
    int   Delta;
+#if defined(USE_OLD_COLOUR_OPS)
+   /* Pre-1.60 colour operations */
    uint16* X2;
    uint16* ZERO_OR_X2;
+#endif
    uint16* ZERO;
    uint8*  S;
    uint8*  DB;
@@ -202,6 +205,8 @@ extern uint8 mul_brightness [16][32];
 #define SUB_SCREEN_DEPTH 0
 #define MAIN_SCREEN_DEPTH 32
 
+#if defined(USE_OLD_COLOUR_OPS)
+/* Pre-1.60 colour operations */
 #if defined(OLD_COLOUR_BLENDING)
 #define COLOR_ADD(C1, C2) \
 GFX.X2 [((((C1) & RGB_REMOVE_LOW_BITS_MASK) + \
@@ -214,12 +219,34 @@ GFX.X2 [((((C1) & RGB_REMOVE_LOW_BITS_MASK) + \
     ((C1) & (C2) & RGB_LOW_BITS_MASK)] | \
  (((C1) ^ (C2)) & RGB_LOW_BITS_MASK))
 #endif
+#else
+static INLINE uint16_t COLOR_ADD(uint16_t C1, uint16_t C2)
+{
+	const int RED_MASK   = 0x1F << RED_SHIFT_BITS;
+	const int GREEN_MASK = 0x1F << GREEN_SHIFT_BITS;
+	const int BLUE_MASK  = 0x1F;
+
+	int rb          = (C1 & (RED_MASK | BLUE_MASK)) + (C2 & (RED_MASK | BLUE_MASK));
+	int rbcarry     = rb & ((0x20 << RED_SHIFT_BITS) | (0x20 << 0));
+	int g           = (C1 & (GREEN_MASK)) + (C2 & (GREEN_MASK));
+	int rgbsaturate = (((g & (0x20 << GREEN_SHIFT_BITS)) | rbcarry) >> 5) * 0x1f;
+	uint16_t retval = (rb & (RED_MASK | BLUE_MASK)) | (g & GREEN_MASK) | rgbsaturate;
+
+#if GREEN_SHIFT_BITS == 6
+	retval         |= (retval & 0x0400) >> 5;
+#endif
+
+	return retval;
+}
+#endif
 
 #define COLOR_ADD1_2(C1, C2) \
 (((((C1) & RGB_REMOVE_LOW_BITS_MASK) + \
           ((C2) & RGB_REMOVE_LOW_BITS_MASK)) >> 1) + \
          ((C1) & (C2) & RGB_LOW_BITS_MASK) | ALPHA_BITS_MASK)
 
+#if defined(USE_OLD_COLOUR_OPS)
+/* Pre-1.60 colour operations */
 #if defined(OLD_COLOUR_BLENDING)
 #define COLOR_SUB(C1, C2) \
 GFX.ZERO_OR_X2 [(((C1) | RGB_HI_BITS_MASKx2) - \
@@ -230,6 +257,24 @@ GFX.ZERO_OR_X2 [(((C1) | RGB_HI_BITS_MASKx2) - \
                   ((C2) & RGB_REMOVE_LOW_BITS_MASK)) >> 1] + \
 ((C1) & RGB_LOW_BITS_MASK) - ((C2) & RGB_LOW_BITS_MASK))
 #endif
+#else
+static INLINE uint16_t COLOR_SUB(uint16_t C1, uint16_t C2)
+{
+	int rb1         = (C1 & (THIRD_COLOR_MASK | FIRST_COLOR_MASK)) | ((0x20 << 0) | (0x20 << RED_SHIFT_BITS));
+	int rb2         = C2 & (THIRD_COLOR_MASK | FIRST_COLOR_MASK);
+	int rb          = rb1 - rb2;
+	int rbcarry     = rb & ((0x20 << RED_SHIFT_BITS) | (0x20 << 0));
+	int g           = ((C1 & (SECOND_COLOR_MASK)) | (0x20 << GREEN_SHIFT_BITS)) - (C2 & (SECOND_COLOR_MASK));
+	int rgbsaturate = (((g & (0x20 << GREEN_SHIFT_BITS)) | rbcarry) >> 5) * 0x1f;
+	uint16_t retval = ((rb & (THIRD_COLOR_MASK | FIRST_COLOR_MASK)) | (g & SECOND_COLOR_MASK)) & rgbsaturate;
+
+#if GREEN_SHIFT_BITS == 6
+	retval         |= (retval & 0x0400) >> 5;
+#endif
+
+	return retval;
+}
+#endif
 
 #define COLOR_SUB1_2(C1, C2) \
 GFX.ZERO [(((C1) | RGB_HI_BITS_MASKx2) - \
diff --git a/src/gfx16.c b/src/gfx16.c
index 51d1240..db35e6a 100644
--- a/src/gfx16.c
+++ b/src/gfx16.c
@@ -249,6 +249,7 @@ void DrawHiResTile16(uint32 Tile, uint32 Offset,
 
 bool8_32 S9xGraphicsInit()
 {
+   uint32 r, g, b;
    uint32 PixelOdd = 1;
    uint32 PixelEven = 2;
 
@@ -366,6 +367,8 @@ bool8_32 S9xGraphicsInit()
    DrawHiResClippedTilePtr = DrawHiResClippedTile16;
    S9xFixColourBrightness();
 
+#if defined(USE_OLD_COLOUR_OPS)
+   /* Pre-1.60 colour operations */
    if (!(GFX.X2 = (uint16*) malloc(sizeof(uint16) * 0x10000)))
       return (FALSE);
 
@@ -384,7 +387,6 @@ bool8_32 S9xGraphicsInit()
       }
       return (FALSE);
    }
-   uint32 r, g, b;
 
    // Build a lookup table that multiplies a packed RGB value by 2 with
    // saturation.
@@ -481,6 +483,10 @@ bool8_32 S9xGraphicsInit()
       }
    }
 #endif
+#else
+   if (!(GFX.ZERO = (uint16_t*) malloc(sizeof(uint16_t) * 0x10000)))
+      return false;
+#endif
 
    // Build a lookup table that if the top bit of the color value is zero
    // then the value is zero, otherwise its just the value.
@@ -519,6 +525,8 @@ bool8_32 S9xGraphicsInit()
 void S9xGraphicsDeinit(void)
 {
    // Free any memory allocated in S9xGraphicsInit
+#if defined(USE_OLD_COLOUR_OPS)
+   /* Pre-1.60 colour operations */
    if (GFX.X2)
    {
       free((char*) GFX.X2);
@@ -529,6 +537,7 @@ void S9xGraphicsDeinit(void)
       free((char*) GFX.ZERO_OR_X2);
       GFX.ZERO_OR_X2 = NULL;
    }
+#endif
    if (GFX.ZERO)
    {
       free((char*) GFX.ZERO);
diff --git a/src/pixform.h b/src/pixform.h
index f7e5e7a..1edd9c5 100644
--- a/src/pixform.h
+++ b/src/pixform.h
@@ -83,9 +83,11 @@ extern uint32 HIGH_BITS_SHIFTED_TWO_MASK;
 #define DECOMPOSE_PIXEL_RGB565(PIX,R,G,B) {(R) = (PIX) >> 11; (G) = ((PIX) >> 6) & 0x1f; (B) = (PIX) & 0x1f; }
 #define SPARE_RGB_BIT_MASK_RGB565 (1 << 5)
 
-#define MAX_RED_RGB565       31
-#define MAX_GREEN_RGB565     63
-#define MAX_BLUE_RGB565      31
+#define MAX_RED_RGB565            31
+#define MAX_GREEN_RGB565          63
+#define MAX_BLUE_RGB565           31
+#define RED_SHIFT_BITS_RGB565     11
+#define GREEN_SHIFT_BITS_RGB565   6
 #define RED_LOW_BIT_MASK_RGB565   0x0800
 #define GREEN_LOW_BIT_MASK_RGB565 0x0020
 #define BLUE_LOW_BIT_MASK_RGB565  0x0001
@@ -103,9 +105,11 @@ extern uint32 HIGH_BITS_SHIFTED_TWO_MASK;
 #define DECOMPOSE_PIXEL_RGB555(PIX,R,G,B) {(R) = (PIX) >> 10; (G) = ((PIX) >> 5) & 0x1f; (B) = (PIX) & 0x1f; }
 #define SPARE_RGB_BIT_MASK_RGB555 (1 << 15)
 
-#define MAX_RED_RGB555       31
-#define MAX_GREEN_RGB555     31
-#define MAX_BLUE_RGB555      31
+#define MAX_RED_RGB555            31
+#define MAX_GREEN_RGB555          31
+#define MAX_BLUE_RGB555           31
+#define RED_SHIFT_BITS_RGB555     10
+#define GREEN_SHIFT_BITS_RGB555   5
 #define RED_LOW_BIT_MASK_RGB555   0x0400
 #define GREEN_LOW_BIT_MASK_RGB555 0x0020
 #define BLUE_LOW_BIT_MASK_RGB555  0x0001
@@ -117,106 +121,6 @@ extern uint32 HIGH_BITS_SHIFTED_TWO_MASK;
 #define THIRD_COLOR_MASK_RGB555   0x001F
 #define ALPHA_BITS_MASK_RGB555    0x0000
 
-// BGR565 format
-#define BUILD_PIXEL_BGR565(R,G,B) (((int) (B) << 11) | ((int) (G) << 6) | (int) (R))
-#define BUILD_PIXEL2_BGR565(R,G,B) (((int) (B) << 11) | ((int) (G) << 5) | (int) (R))
-#define DECOMPOSE_PIXEL_BGR565(PIX,R,G,B) {(B) = (PIX) >> 11; (G) = ((PIX) >> 6) & 0x1f; (R) = (PIX) & 0x1f; }
-#define SPARE_RGB_BIT_MASK_BGR565 (1 << 5)
-
-#define MAX_RED_BGR565       31
-#define MAX_GREEN_BGR565     63
-#define MAX_BLUE_BGR565      31
-#define RED_LOW_BIT_MASK_BGR565   0x0001
-#define GREEN_LOW_BIT_MASK_BGR565 0x0040
-#define BLUE_LOW_BIT_MASK_BGR565  0x0800
-#define RED_HI_BIT_MASK_BGR565    0x0010
-#define GREEN_HI_BIT_MASK_BGR565  0x0400
-#define BLUE_HI_BIT_MASK_BGR565   0x8000
-#define FIRST_COLOR_MASK_BGR565   0xF800
-#define SECOND_COLOR_MASK_BGR565  0x07E0
-#define THIRD_COLOR_MASK_BGR565   0x001F
-#define ALPHA_BITS_MASK_BGR565    0x0000
-
-// BGR555 format
-#define BUILD_PIXEL_BGR555(R,G,B) (((int) (B) << 10) | ((int) (G) << 5) | (int) (R))
-#define BUILD_PIXEL2_BGR555(R,G,B) (((int) (B) << 10) | ((int) (G) << 5) | (int) (R))
-#define DECOMPOSE_PIXEL_BGR555(PIX,R,G,B) {(B) = (PIX) >> 10; (G) = ((PIX) >> 5) & 0x1f; (R) = (PIX) & 0x1f; }
-#define SPARE_RGB_BIT_MASK_BGR555 (1 << 15)
-
-#define MAX_RED_BGR555       31
-#define MAX_GREEN_BGR555     31
-#define MAX_BLUE_BGR555      31
-#define RED_LOW_BIT_MASK_BGR555   0x0001
-#define GREEN_LOW_BIT_MASK_BGR555 0x0020
-#define BLUE_LOW_BIT_MASK_BGR555  0x0400
-#define RED_HI_BIT_MASK_BGR555    0x0010
-#define GREEN_HI_BIT_MASK_BGR555  0x0200
-#define BLUE_HI_BIT_MASK_BGR555   0x4000
-#define FIRST_COLOR_MASK_BGR555   0x7C00
-#define SECOND_COLOR_MASK_BGR555  0x03E0
-#define THIRD_COLOR_MASK_BGR555   0x001F
-#define ALPHA_BITS_MASK_BGR555    0x0000
-
-// GBR565 format
-#define BUILD_PIXEL_GBR565(R,G,B) (((int) (G) << 11) | ((int) (B) << 6) | (int) (R))
-#define BUILD_PIXEL2_GBR565(R,G,B) (((int) (G) << 11) | ((int) (B) << 5) | (int) (R))
-#define DECOMPOSE_PIXEL_GBR565(PIX,R,G,B) {(G) = (PIX) >> 11; (B) = ((PIX) >> 6) & 0x1f; (R) = (PIX) & 0x1f; }
-#define SPARE_RGB_BIT_MASK_GBR565 (1 << 5)
-
-#define MAX_RED_GBR565            31
-#define MAX_BLUE_GBR565           63
-#define MAX_GREEN_GBR565          31
-#define RED_LOW_BIT_MASK_GBR565   0x0001
-#define BLUE_LOW_BIT_MASK_GBR565  0x0040
-#define GREEN_LOW_BIT_MASK_GBR565 0x0800
-#define RED_HI_BIT_MASK_GBR565    0x0010
-#define BLUE_HI_BIT_MASK_GBR565   0x0400
-#define GREEN_HI_BIT_MASK_GBR565  0x8000
-#define FIRST_COLOR_MASK_GBR565   0xF800
-#define SECOND_COLOR_MASK_GBR565  0x07E0
-#define THIRD_COLOR_MASK_GBR565   0x001F
-#define ALPHA_BITS_MASK_GBR565    0x0000
-
-// GBR555 format
-#define BUILD_PIXEL_GBR555(R,G,B) (((int) (G) << 10) | ((int) (B) << 5) | (int) (R))
-#define BUILD_PIXEL2_GBR555(R,G,B) (((int) (G) << 10) | ((int) (B) << 5) | (int) (R))
-#define DECOMPOSE_PIXEL_GBR555(PIX,R,G,B) {(G) = (PIX) >> 10; (B) = ((PIX) >> 5) & 0x1f; (R) = (PIX) & 0x1f; }
-#define SPARE_RGB_BIT_MASK_GBR555 (1 << 15)
-
-#define MAX_RED_GBR555            31
-#define MAX_BLUE_GBR555           31
-#define MAX_GREEN_GBR555          31
-#define RED_LOW_BIT_MASK_GBR555   0x0001
-#define BLUE_LOW_BIT_MASK_GBR555  0x0020
-#define GREEN_LOW_BIT_MASK_GBR555 0x0400
-#define RED_HI_BIT_MASK_GBR555    0x0010
-#define BLUE_HI_BIT_MASK_GBR555   0x0200
-#define GREEN_HI_BIT_MASK_GBR555  0x4000
-#define FIRST_COLOR_MASK_GBR555   0x7C00
-#define SECOND_COLOR_MASK_GBR555  0x03E0
-#define THIRD_COLOR_MASK_GBR555   0x001F
-#define ALPHA_BITS_MASK_GBR555    0x0000
-
-// RGB5551 format
-#define BUILD_PIXEL_RGB5551(R,G,B) (((int) (R) << 11) | ((int) (G) << 6) | (int) ((B) << 1) | 1)
-#define BUILD_PIXEL2_RGB5551(R,G,B) (((int) (R) << 11) | ((int) (G) << 6) | (int) ((B) << 1) | 1)
-#define DECOMPOSE_PIXEL_RGB5551(PIX,R,G,B) {(R) = (PIX) >> 11; (G) = ((PIX) >> 6) & 0x1f; (B) = ((PIX) >> 1) & 0x1f; }
-#define SPARE_RGB_BIT_MASK_RGB5551 (1)
-
-#define MAX_RED_RGB5551       31
-#define MAX_GREEN_RGB5551     31
-#define MAX_BLUE_RGB5551      31
-#define RED_LOW_BIT_MASK_RGB5551   0x0800
-#define GREEN_LOW_BIT_MASK_RGB5551 0x0040
-#define BLUE_LOW_BIT_MASK_RGB5551  0x0002
-#define RED_HI_BIT_MASK_RGB5551    0x8000
-#define GREEN_HI_BIT_MASK_RGB5551  0x0400
-#define BLUE_HI_BIT_MASK_RGB5551   0x0020
-#define FIRST_COLOR_MASK_RGB5551   0xf800
-#define SECOND_COLOR_MASK_RGB5551  0x07c0
-#define THIRD_COLOR_MASK_RGB5551   0x003e
-#define ALPHA_BITS_MASK_RGB5551    0x0001
-
 #ifndef GFX_MULTI_FORMAT
 #define CONCAT(X,Y) X##Y
 
@@ -230,24 +134,28 @@ extern uint32 HIGH_BITS_SHIFTED_TWO_MASK;
 #define BUILD_PIXEL2(R,G,B) BUILD_PIXEL2_D(PIXEL_FORMAT,R,G,B)
 #define DECOMPOSE_PIXEL(PIX,R,G,B) DECOMPOSE_PIXEL_D(PIXEL_FORMAT,PIX,R,G,B)
 
-#define MAX_RED_D(F)    CONCAT(MAX_RED_,F)
-#define MAX_BLUE_D(F)      CONCAT(MAX_BLUE_,F)
-#define MAX_GREEN_D(F)     CONCAT(MAX_GREEN_,F)
-#define RED_LOW_BIT_MASK_D(F) CONCAT(RED_LOW_BIT_MASK_,F)
-#define BLUE_LOW_BIT_MASK_D(F)   CONCAT(BLUE_LOW_BIT_MASK_,F)
-#define GREEN_LOW_BIT_MASK_D(F)  CONCAT(GREEN_LOW_BIT_MASK_,F)
-#define RED_HI_BIT_MASK_D(F)  CONCAT(RED_HI_BIT_MASK_,F)
-#define BLUE_HI_BIT_MASK_D(F) CONCAT(BLUE_HI_BIT_MASK_,F)
-#define GREEN_HI_BIT_MASK_D(F)   CONCAT(GREEN_HI_BIT_MASK_,F)
+#define MAX_RED_D(F)            CONCAT(MAX_RED_,F)
+#define MAX_BLUE_D(F)           CONCAT(MAX_BLUE_,F)
+#define MAX_GREEN_D(F)          CONCAT(MAX_GREEN_,F)
+#define RED_SHIFT_BITS_D(F)     CONCAT(RED_SHIFT_BITS_, F)
+#define GREEN_SHIFT_BITS_D(F)   CONCAT(GREEN_SHIFT_BITS_, F)
+#define RED_LOW_BIT_MASK_D(F)   CONCAT(RED_LOW_BIT_MASK_,F)
+#define BLUE_LOW_BIT_MASK_D(F)  CONCAT(BLUE_LOW_BIT_MASK_,F)
+#define GREEN_LOW_BIT_MASK_D(F) CONCAT(GREEN_LOW_BIT_MASK_,F)
+#define RED_HI_BIT_MASK_D(F)    CONCAT(RED_HI_BIT_MASK_,F)
+#define BLUE_HI_BIT_MASK_D(F)   CONCAT(BLUE_HI_BIT_MASK_,F)
+#define GREEN_HI_BIT_MASK_D(F)  CONCAT(GREEN_HI_BIT_MASK_,F)
 #define FIRST_COLOR_MASK_D(F)   CONCAT(FIRST_COLOR_MASK_,F)
 #define SECOND_COLOR_MASK_D(F)  CONCAT(SECOND_COLOR_MASK_,F)
 #define THIRD_COLOR_MASK_D(F)   CONCAT(THIRD_COLOR_MASK_,F)
 #define ALPHA_BITS_MASK_D(F)    CONCAT(ALPHA_BITS_MASK_,F)
 
-#define MAX_RED         MAX_RED_D(PIXEL_FORMAT)
-#define MAX_BLUE     MAX_BLUE_D(PIXEL_FORMAT)
-#define MAX_GREEN    MAX_GREEN_D(PIXEL_FORMAT)
-#define RED_LOW_BIT_MASK   RED_LOW_BIT_MASK_D(PIXEL_FORMAT)
+#define MAX_RED                 MAX_RED_D(PIXEL_FORMAT)
+#define MAX_BLUE                MAX_BLUE_D(PIXEL_FORMAT)
+#define MAX_GREEN               MAX_GREEN_D(PIXEL_FORMAT)
+#define RED_SHIFT_BITS          RED_SHIFT_BITS_D(PIXEL_FORMAT)
+#define GREEN_SHIFT_BITS        GREEN_SHIFT_BITS_D(PIXEL_FORMAT)
+#define RED_LOW_BIT_MASK        RED_LOW_BIT_MASK_D(PIXEL_FORMAT)
 #define BLUE_LOW_BIT_MASK       BLUE_LOW_BIT_MASK_D(PIXEL_FORMAT)
 #define GREEN_LOW_BIT_MASK      GREEN_LOW_BIT_MASK_D(PIXEL_FORMAT)
 #define RED_HI_BIT_MASK         RED_HI_BIT_MASK_D(PIXEL_FORMAT)