core md, handle non-linear vdp colour mapping

diff --git a/pico/draw.c b/pico/draw.c
index 0e241af..e58d50f 100644 (file)
--- a/pico/draw.c
+++ b/pico/draw.c
@@ -74,11 +74,11 @@ u32 VdpSATCache[2*128];  // VDP sprite cache (1st 32 sprite attr bits)
 #define PXMASKH     0x39ce39ce  // 0x3def3def, all but MSB for all colours\r
 #elif defined(USE_BGR565)\r
 #define PXCONV(t)   ((t & 0x000e000e)<< 1) | ((t & 0x00e000e0)<<3) | ((t & 0x0e000e00)<<4)\r
-#define PXMASKL     0x08610861  // 0x18e318e3\r
+#define PXMASKL     0x08410841  // 0x18e318e3\r
 #define PXMASKH     0x738e738e  // 0x7bef7bef\r
 #else // RGB565\r
 #define PXCONV(t)   ((t & 0x000e000e)<<12) | ((t & 0x00e000e0)<<3) | ((t & 0x0e000e00)>>7)\r
-#define PXMASKL     0x08610861  // 0x18e318e3\r
+#define PXMASKL     0x08410841  // 0x18e318e3\r
 #define PXMASKH     0x738e738e  // 0x7bef7bef\r
 #endif\r
 \r
@@ -1438,22 +1438,27 @@ static void PicoDoHighPal555_8bit(int sh, int line, struct PicoEState *est)
     // otherwise intensity difference between this and s/h will be wrong\r
     t = PXCONV(t);\r
     t |= (t >> 4) & PXMASKL;\r
+    // take into account that MegaDrive RGB output isn't linear, there is a larger\r
+    // step for 0->1 and 6->7. While the latter isn't obviously noticeable, the\r
+    // former is clearly visible (see color test in Knuckles Chaotix)\r
+    t |= ((t ^ PXMASKL) & (/*t>>4|*/t>>3|t>>2) & PXMASKL) << 1;\r
     dpal[i] = t;\r
   }\r
 \r
   // norm: xxx0, sh: 0xxx, hi: 0xxx + 7\r
   if (sh)\r
   {\r
-    // shadowed pixels\r
+    // normal and shadowed pixels\r
     for (i = 0; i < 0x40 / 2; i++) {\r
       dpal[0xc0/2 + i] = dpal[i];\r
-      dpal[0x80/2 + i] = (dpal[i] >> 1) & PXMASKH;\r
+      // take into account that MegaDrive RGB output isn't linear, see above\r
+      t = (dpal[i] >> 1) & PXMASKH;\r
+      dpal[0x80/2 + i] = t + ((t>>2|t>>1|t>>0) & (PXMASKL<<1));\r
     }\r
     // hilighted pixels\r
     for (i = 0; i < 0x40 / 2; i++) {\r
-      t = ((dpal[i] >> 1) & PXMASKH) + PXMASKH;\r
-      t |= (t >> 4) & PXMASKL;\r
-      dpal[0x40/2 + i] = t;\r
+      t = (dpal[i] >> 1) & PXMASKH;\r
+      dpal[0x40/2 + i] = t + PXMASKH + PXMASKL;\r
     }\r
   }\r
 }\r
@@ -1475,6 +1480,8 @@ void PicoDoHighPal555(int sh, int line, struct PicoEState *est)
     // otherwise intensity difference between this and s/h will be wrong\r
     t = PXCONV(t);\r
     t |= (t >> 4) & PXMASKL;\r
+    // take into account that MegaDrive RGB output isn't linear, see above\r
+    t |= ((t ^ PXMASKL) & (/*t>>4|*/t>>3|t>>2) & PXMASKL) << 1;\r
     dpal[i] = dpal[0xc0/2 + i] = t;\r
   }\r
 \r
@@ -1482,13 +1489,15 @@ void PicoDoHighPal555(int sh, int line, struct PicoEState *est)
   if (sh)\r
   {\r
     // shadowed pixels\r
-    for (i = 0; i < 0x40 / 2; i++)\r
-      dpal[0x80/2 + i] = (dpal[i] >> 1) & PXMASKH;\r
+    for (i = 0; i < 0x40 / 2; i++) {\r
+      // take into account that MegaDrive RGB output isn't linear, see above\r
+      t = (dpal[i] >> 1) & PXMASKH;\r
+      dpal[0x80/2 + i] = t + ((t>>2|t>>1|t>>0) & (PXMASKL<<1));\r
+    }\r
     // hilighted pixels\r
     for (i = 0; i < 0x40 / 2; i++) {\r
-      t = ((dpal[i] >> 1) & PXMASKH) + PXMASKH;\r
-      t |= (t >> 4) & PXMASKL;\r
-      dpal[0x40/2 + i] = t;\r
+      t = (dpal[i] >> 1) & PXMASKH;\r
+      dpal[0x40/2 + i] = t + PXMASKH + PXMASKL;\r
     }\r
   }\r
 }\r

diff --git a/pico/draw_arm.S b/pico/draw_arm.S
index 05a2ece..d3f456e 100644 (file)
--- a/pico/draw_arm.S
+++ b/pico/draw_arm.S
@@ -1564,6 +1564,11 @@ DrawWindow:
 \r
     and     r2,   r8,   \reg,lsr #4\r
     orr     \reg, \reg, r2\r
+\r
+    orr     r3,   \reg, \reg,lsr #1\r
+    and     r3,   r3,   r8,lsl #2\r
+    bic     r3,   r3,   r2,lsl #2\r
+    orr     \reg, \reg, r3,lsr #1\r
 .endm\r
 \r
 @ trashes: r2-r8,r12,lr; r8 = 0x08610861; r0,r1 are advanced\r
@@ -1631,28 +1636,27 @@ PicoDoHighPal555:
 \r
     @ hilighted pixels (0x40-0x7f):\r
     @  t = ((dpal[i] >> 1) & 0x738e738e) + 0x738e738e;\r
-    @  t |= (t >> 4) & 0x08610861;\r
+    @  t += 0x08610861;\r
     @ r8=0x08610861\r
     mov     r12,    #0x008e\r
     orr     r12,r12,#0x7300\r
     orr     r12,r12,r12,lsl #16\r
+    add     r10,r12,r8\r
     mov     lr, #0x40/4\r
 .fl_loopcpRGB555_hi:\r
     ldmia   r3!, {r1,r6}\r
     and     r1, r12, r1, lsr #1\r
     and     r6, r12, r6, lsr #1\r
-    add     r1, r12, r1\r
-    add     r6, r12, r6\r
-    and     r5, r8, r1, lsr #4\r
-    and     r7, r8, r6, lsr #4\r
-    orr     r1, r1, r5\r
-    orr     r6, r6, r7\r
+    add     r1, r10, r1\r
+    add     r6, r10, r6\r
     stmia   r4!, {r1,r6}\r
     subs    lr, lr, #1\r
     bne     .fl_loopcpRGB555_hi\r
 \r
     sub     r3, r3, #0x40*2\r
     @ shadowed (0x80-0xbf), shadow|hilight (aka normal, 0xc0-0xff) pixels:\r
+    @ t = (dpal[i] >> 1) & 0x738e738e;\r
+    @ t += ((t>>2|t>>1|t>>0) & (0x08610861<<1));\r
     add     r5, r3, #0xc0*2\r
     mov     lr, #0x40/4\r
 .fl_loopcpRGB555_sh:\r
@@ -1660,7 +1664,15 @@ PicoDoHighPal555:
     subs    lr, lr, #1\r
     stmia   r5!, {r1,r6} @ 0xc0, normal\r
     and     r1, r12, r1, lsr #1\r
+    orr     r0, r1, r1, lsr #1\r
+    orr     r0, r0, r1, lsr #2\r
+    and     r0, r0, r8, lsl #1\r
+    add     r1, r1, r0\r
     and     r6, r12, r6, lsr #1\r
+    orr     r0, r6, r6, lsr #1\r
+    orr     r0, r0, r6, lsr #2\r
+    and     r0, r0, r8, lsl #1\r
+    add     r6, r6, r0\r
     stmia   r4!, {r1,r6}\r
     bne     .fl_loopcpRGB555_sh\r
 \r