Port color operation code from Snes9x 1.60

4 files changed, 49 insertions, 205 deletions

diff --git a/source/gfx.c b/source/gfx.c
index 00fb4d1..3508011 100644
--- a/source/gfx.c
+++ b/source/gfx.c
@@ -226,81 +226,12 @@ bool S9xInitGFX(void)
    GFX.PPLx2 = GFX.Pitch;
    S9xFixColourBrightness();
 
-   if (!(GFX.X2 = (uint16_t*) malloc(sizeof(uint16_t) * 0x10000)))
-      return false;
-
-   if (!(GFX.ZERO_OR_X2 = (uint16_t*) malloc(sizeof(uint16_t) * 0x10000)) || !(GFX.ZERO = (uint16_t*) malloc(sizeof(uint16_t) * 0x10000)))
+   if (!(GFX.ZERO = (uint16_t*) malloc(sizeof(uint16_t) * 0x10000)))
    {
-      if (GFX.ZERO_OR_X2)
-      {
-         free(GFX.ZERO_OR_X2);
-         GFX.ZERO_OR_X2 = NULL;
-      }
-      if (GFX.X2)
-      {
-         free(GFX.X2);
-         GFX.X2 = NULL;
-      }
       return false;
    }
 
-   /* Build a lookup table that multiplies a packed RGB value by 2 with
-    * saturation. */
-   for (r = 0; r <= MAX_RED; r++)
-   {
-      uint32_t r2 = r << 1;
-      if (r2 > MAX_RED)
-         r2 = MAX_RED;
-      for (g = 0; g <= MAX_GREEN; g++)
-      {
-         uint32_t g2 = g << 1;
-         if (g2 > MAX_GREEN)
-            g2 = MAX_GREEN;
-         for (b = 0; b <= MAX_BLUE; b++)
-         {
-            uint32_t b2 = b << 1;
-            if (b2 > MAX_BLUE)
-               b2 = MAX_BLUE;
-            GFX.X2 [BUILD_PIXEL2(r, g, b)] = BUILD_PIXEL2(r2, g2, b2);
-            GFX.X2 [BUILD_PIXEL2(r, g, b) & ~ALPHA_BITS_MASK] = BUILD_PIXEL2(r2, g2, b2);
-         }
-      }
-   }
    memset(GFX.ZERO, 0, 0x10000 * sizeof(uint16_t));
-   memset(GFX.ZERO_OR_X2, 0, 0x10000 * sizeof(uint16_t));
-   /* Build a lookup table that if the top bit of the color value is zero
-    * then the value is zero, otherwise multiply the value by 2. Used by
-    * the color subtraction code. */
-   for (r = 0; r <= MAX_RED; r++)
-   {
-      uint32_t r2 = r;
-      if ((r2 & 0x10) == 0)
-         r2 = 0;
-      else
-         r2 = (r2 << 1) & MAX_RED;
-
-      for (g = 0; g <= MAX_GREEN; g++)
-      {
-         uint32_t g2 = g;
-         if ((g2 & GREEN_HI_BIT) == 0)
-            g2 = 0;
-         else
-            g2 = (g2 << 1) & MAX_GREEN;
-
-         for (b = 0; b <= MAX_BLUE; b++)
-         {
-            uint32_t b2 = b;
-            if ((b2 & 0x10) == 0)
-               b2 = 0;
-            else
-               b2 = (b2 << 1) & MAX_BLUE;
-
-            GFX.ZERO_OR_X2 [BUILD_PIXEL2(r, g, b)] = BUILD_PIXEL2(r2, g2, b2);
-            GFX.ZERO_OR_X2 [BUILD_PIXEL2(r, g, b) & ~ALPHA_BITS_MASK] = BUILD_PIXEL2(MAX(1, r2), MAX(1, g2), MAX(1, b2));
-         }
-      }
-   }
-
    /* 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. */
    for (r = 0; r <= MAX_RED; r++)
@@ -337,16 +268,6 @@ bool S9xInitGFX(void)
 void S9xDeinitGFX(void)
 {
    /* Free any memory allocated in S9xInitGFX */
-   if (GFX.X2)
-   {
-      free(GFX.X2);
-      GFX.X2 = NULL;
-   }
-   if (GFX.ZERO_OR_X2)
-   {
-      free(GFX.ZERO_OR_X2);
-      GFX.ZERO_OR_X2 = NULL;
-   }
    if (GFX.ZERO)
    {
       free(GFX.ZERO);
diff --git a/source/gfx.h b/source/gfx.h
index 582f7c2..fc4806b 100644
--- a/source/gfx.h
+++ b/source/gfx.h
@@ -33,8 +33,6 @@ typedef struct
    uint32_t    Pitch;
 
    int32_t     Delta;
-   uint16_t*   X2;
-   uint16_t*   ZERO_OR_X2;
    uint16_t*   ZERO;
    uint32_t    RealPitch;  /* True pitch of Screen buffer. */
    uint32_t    Pitch2;     /* Same as RealPitch except while using speed up hack for Glide. */
@@ -148,28 +146,50 @@ extern uint8_t mul_brightness [16][32];
 
 static INLINE uint16_t COLOR_ADD(uint16_t C1, uint16_t C2)
 {
-   if (C1 == 0)
-      return C2;
-   else if (C2 == 0)
-      return C1;
-   else
-      return GFX.X2[(((C1 & RGB_REMOVE_LOW_BITS_MASK) + (C2 & RGB_REMOVE_LOW_BITS_MASK)) >> 1) + (C1 & C2 & RGB_LOW_BITS_MASK)] | ((C1 ^ C2) & RGB_LOW_BITS_MASK);
+	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);
+	rb += 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;
+}
+
+static INLINE uint16_t COLOR_ADD1_2(uint16_t C1, uint16_t C2)
+{
+	return ((((((C1)&RGB_REMOVE_LOW_BITS_MASK) + ((C2)&RGB_REMOVE_LOW_BITS_MASK)) >> 1) +
+		 ((C1) & (C2)&RGB_LOW_BITS_MASK)) |
+		ALPHA_BITS_MASK);
 }
 
-#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))
+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
 
-#define COLOR_SUB(C1, C2) \
-(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))
+	return retval;
+}
 
-#define COLOR_SUB1_2(C1, C2) \
-GFX.ZERO [(((C1) | RGB_HI_BITS_MASKx2) - \
-           ((C2) & RGB_REMOVE_LOW_BITS_MASK)) >> 1]
+static INLINE uint16_t COLOR_SUB1_2(uint16_t C1, uint16_t C2)
+{
+	return GFX.ZERO[(((C1) | RGB_HI_BITS_MASKx2) - ((C2)&RGB_REMOVE_LOW_BITS_MASK)) >> 1];
+}
 
 typedef void (*NormalTileRenderer)(uint32_t Tile, int32_t Offset, uint32_t StartLine, uint32_t LineCount);
 typedef void (*ClippedTileRenderer)(uint32_t Tile, int32_t Offset, uint32_t StartPixel, uint32_t Width, uint32_t StartLine, uint32_t LineCount);
diff --git a/source/pixform.h b/source/pixform.h
index f9c075c..4a61ae2 100644
--- a/source/pixform.h
+++ b/source/pixform.h
@@ -12,6 +12,8 @@
 #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
@@ -32,6 +34,8 @@
 #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
@@ -43,107 +47,6 @@
 #define THIRD_COLOR_MASK_RGB555   0x001F
 #define ALPHA_BITS_MASK_RGB555    0x0000
 
-/* BGR565 format */
-#define BUILD_PIXEL_BGR565(R,G,B) (((int32_t) (B) << 11) | ((int32_t) (G) << 6) | (int32_t) (R))
-#define BUILD_PIXEL2_BGR565(R,G,B) (((int32_t) (B) << 11) | ((int32_t) (G) << 5) | (int32_t) (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) (((int32_t) (B) << 10) | ((int32_t) (G) << 5) | (int32_t) (R))
-#define BUILD_PIXEL2_BGR555(R,G,B) (((int32_t) (B) << 10) | ((int32_t) (G) << 5) | (int32_t) (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) (((int32_t) (G) << 11) | ((int32_t) (B) << 6) | (int32_t) (R))
-#define BUILD_PIXEL2_GBR565(R,G,B) (((int32_t) (G) << 11) | ((int32_t) (B) << 5) | (int32_t) (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) (((int32_t) (G) << 10) | ((int32_t) (B) << 5) | (int32_t) (R))
-#define BUILD_PIXEL2_GBR555(R,G,B) (((int32_t) (G) << 10) | ((int32_t) (B) << 5) | (int32_t) (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) (((int32_t) (R) << 11) | ((int32_t) (G) << 6) | (int32_t) ((B) << 1) | 1)
-#define BUILD_PIXEL2_RGB5551(R,G,B) (((int32_t) (R) << 11) | ((int32_t) (G) << 6) | (int32_t) ((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
-
-
 #define CONCAT(X,Y) X##Y
 
 /* C pre-processor needs a two stage macro define to enable it to concat
@@ -159,6 +62,8 @@
 #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)
@@ -173,6 +78,8 @@
 #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)
diff --git a/source/port.h b/source/port.h
index 06c5f5e..6c3425c 100644
--- a/source/port.h
+++ b/source/port.h
@@ -14,11 +14,7 @@
 #include <sys/types.h>
 #include <retro_inline.h>
 
-#ifdef PSP
-#define PIXEL_FORMAT BGR555
-#else
 #define PIXEL_FORMAT RGB565
-#endif
 /* The above is used to disable the 16-bit graphics mode checks sprinkled
  * throughout the code, if the pixel format is always 16-bit. */