From: kub <derkub@gmail.com>
Date: Tue, 11 Mar 2025 06:39:06 +0000 (+0100)
Subject: core md, handle non-linear vdp colour mapping
X-Git-Tag: v2.04~19
X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=02d3e670b9091188681180be0b3c97f87dfce1a9;p=picodrive.git

core md, handle non-linear vdp colour mapping
---

diff --git a/pico/draw.c b/pico/draw.c
index 0e241af1..e58d50fd 100644
--- 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
 #elif defined(USE_BGR565)
 #define PXCONV(t)   ((t & 0x000e000e)<< 1) | ((t & 0x00e000e0)<<3) | ((t & 0x0e000e00)<<4)
-#define PXMASKL     0x08610861  // 0x18e318e3
+#define PXMASKL     0x08410841  // 0x18e318e3
 #define PXMASKH     0x738e738e  // 0x7bef7bef
 #else // RGB565
 #define PXCONV(t)   ((t & 0x000e000e)<<12) | ((t & 0x00e000e0)<<3) | ((t & 0x0e000e00)>>7)
-#define PXMASKL     0x08610861  // 0x18e318e3
+#define PXMASKL     0x08410841  // 0x18e318e3
 #define PXMASKH     0x738e738e  // 0x7bef7bef
 #endif
 
@@ -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
     t = PXCONV(t);
     t |= (t >> 4) & PXMASKL;
+    // take into account that MegaDrive RGB output isn't linear, there is a larger
+    // step for 0->1 and 6->7. While the latter isn't obviously noticeable, the
+    // former is clearly visible (see color test in Knuckles Chaotix)
+    t |= ((t ^ PXMASKL) & (/*t>>4|*/t>>3|t>>2) & PXMASKL) << 1;
     dpal[i] = t;
   }
 
   // norm: xxx0, sh: 0xxx, hi: 0xxx + 7
   if (sh)
   {
-    // shadowed pixels
+    // normal and shadowed pixels
     for (i = 0; i < 0x40 / 2; i++) {
       dpal[0xc0/2 + i] = dpal[i];
-      dpal[0x80/2 + i] = (dpal[i] >> 1) & PXMASKH;
+      // take into account that MegaDrive RGB output isn't linear, see above
+      t = (dpal[i] >> 1) & PXMASKH;
+      dpal[0x80/2 + i] = t + ((t>>2|t>>1|t>>0) & (PXMASKL<<1));
     }
     // hilighted pixels
     for (i = 0; i < 0x40 / 2; i++) {
-      t = ((dpal[i] >> 1) & PXMASKH) + PXMASKH;
-      t |= (t >> 4) & PXMASKL;
-      dpal[0x40/2 + i] = t;
+      t = (dpal[i] >> 1) & PXMASKH;
+      dpal[0x40/2 + i] = t + PXMASKH + PXMASKL;
     }
   }
 }
@@ -1475,6 +1480,8 @@ void PicoDoHighPal555(int sh, int line, struct PicoEState *est)
     // otherwise intensity difference between this and s/h will be wrong
     t = PXCONV(t);
     t |= (t >> 4) & PXMASKL;
+    // take into account that MegaDrive RGB output isn't linear, see above
+    t |= ((t ^ PXMASKL) & (/*t>>4|*/t>>3|t>>2) & PXMASKL) << 1;
     dpal[i] = dpal[0xc0/2 + i] = t;
   }
 
@@ -1482,13 +1489,15 @@ void PicoDoHighPal555(int sh, int line, struct PicoEState *est)
   if (sh)
   {
     // shadowed pixels
-    for (i = 0; i < 0x40 / 2; i++)
-      dpal[0x80/2 + i] = (dpal[i] >> 1) & PXMASKH;
+    for (i = 0; i < 0x40 / 2; i++) {
+      // take into account that MegaDrive RGB output isn't linear, see above
+      t = (dpal[i] >> 1) & PXMASKH;
+      dpal[0x80/2 + i] = t + ((t>>2|t>>1|t>>0) & (PXMASKL<<1));
+    }
     // hilighted pixels
     for (i = 0; i < 0x40 / 2; i++) {
-      t = ((dpal[i] >> 1) & PXMASKH) + PXMASKH;
-      t |= (t >> 4) & PXMASKL;
-      dpal[0x40/2 + i] = t;
+      t = (dpal[i] >> 1) & PXMASKH;
+      dpal[0x40/2 + i] = t + PXMASKH + PXMASKL;
     }
   }
 }
diff --git a/pico/draw_arm.S b/pico/draw_arm.S
index 05a2ecec..d3f456ed 100644
--- a/pico/draw_arm.S
+++ b/pico/draw_arm.S
@@ -1564,6 +1564,11 @@ DrawWindow:
 
     and     r2,   r8,   \reg,lsr #4
     orr     \reg, \reg, r2
+
+    orr     r3,   \reg, \reg,lsr #1
+    and     r3,   r3,   r8,lsl #2
+    bic     r3,   r3,   r2,lsl #2
+    orr     \reg, \reg, r3,lsr #1
 .endm
 
 @ trashes: r2-r8,r12,lr; r8 = 0x08610861; r0,r1 are advanced
@@ -1631,28 +1636,27 @@ PicoDoHighPal555:
 
     @ hilighted pixels (0x40-0x7f):
     @  t = ((dpal[i] >> 1) & 0x738e738e) + 0x738e738e;
-    @  t |= (t >> 4) & 0x08610861;
+    @  t += 0x08610861;
     @ r8=0x08610861
     mov     r12,    #0x008e
     orr     r12,r12,#0x7300
     orr     r12,r12,r12,lsl #16
+    add     r10,r12,r8
     mov     lr, #0x40/4
 .fl_loopcpRGB555_hi:
     ldmia   r3!, {r1,r6}
     and     r1, r12, r1, lsr #1
     and     r6, r12, r6, lsr #1
-    add     r1, r12, r1
-    add     r6, r12, r6
-    and     r5, r8, r1, lsr #4
-    and     r7, r8, r6, lsr #4
-    orr     r1, r1, r5
-    orr     r6, r6, r7
+    add     r1, r10, r1
+    add     r6, r10, r6
     stmia   r4!, {r1,r6}
     subs    lr, lr, #1
     bne     .fl_loopcpRGB555_hi
 
     sub     r3, r3, #0x40*2
     @ shadowed (0x80-0xbf), shadow|hilight (aka normal, 0xc0-0xff) pixels:
+    @ t = (dpal[i] >> 1) & 0x738e738e;
+    @ t += ((t>>2|t>>1|t>>0) & (0x08610861<<1));
     add     r5, r3, #0xc0*2
     mov     lr, #0x40/4
 .fl_loopcpRGB555_sh:
@@ -1660,7 +1664,15 @@ PicoDoHighPal555:
     subs    lr, lr, #1
     stmia   r5!, {r1,r6} @ 0xc0, normal
     and     r1, r12, r1, lsr #1
+    orr     r0, r1, r1, lsr #1
+    orr     r0, r0, r1, lsr #2
+    and     r0, r0, r8, lsl #1
+    add     r1, r1, r0
     and     r6, r12, r6, lsr #1
+    orr     r0, r6, r6, lsr #1
+    orr     r0, r0, r6, lsr #2
+    and     r0, r0, r8, lsl #1
+    add     r6, r6, r0
     stmia   r4!, {r1,r6}
     bne     .fl_loopcpRGB555_sh