more rendering fixes and optimisations

pico/32x/draw.c

@@ -41,6 +41,13 @@ static void convert_pal555(int invert_prio)
   Pico32x.dirty_pal = 0;
 }
 
+// 555 conversion for direct color mode
+#if defined(USE_BGR555)
+#define DC555(t)	t
+#else
+#define DC555(t)	((t&m1) << 11) | ((t&m2) << 1) | ((t&m3) >> 10)
+#endif
+
 // direct color mode
 #define do_line_dc(pd, p32x, pmd, inv, pmd_draw_code)             \
 {                                                                 \
@@ -53,12 +60,12 @@ static void convert_pal555(int invert_prio)
   while (i > 0) {                                                 \
     for (; i > 0 && (*pmd & 0x3f) == mdbg; pd++, pmd++, i--) {    \
       t = *p32x++;                                                \
-      *pd = ((t&m1) << 11) | ((t&m2) << 1) | ((t&m3) >> 10);      \
+      *pd = DC555(t);                                             \
     }                                                             \
     for (; i > 0 && (*pmd & 0x3f) != mdbg; pd++, pmd++, i--) {    \
       t = *p32x++;                                                \
       if ((t ^ inv) & 0x8000)                                     \
-        *pd = ((t&m1) << 11) | ((t&m2) << 1) | ((t&m3) >> 10);    \
+        *pd = DC555(t);                                           \
       else                                                        \
         pmd_draw_code;                                            \
     }                                                             \

pico/draw.c

@@ -33,13 +33,14 @@
  *
  * low prio s/h rendering:
  * - plane and non-op sprite pixels have shadow
- * - sprite op pixel rendering is marked with hilight (deferred)
+ * - s/h sprite op pixel rendering is marked with hilight (deferred)
  * high prio s/h rendering:
  * - plane and non-op sprite pixels are normal
- * - all op sprite pixels (either marked or high prio) are rendered
+ * - all s/h sprite op pixels (either marked or high prio) are rendered
  *
  * not handled properly:
- * - high prio s/h sprite overlapping low prio sprite shows sprite, not A,B,G
+ * - high prio s/h sprite op overlapping low prio sprite shows sprite, not A,B,G
+ * - in debug sprite-masked, transparent high-prio sprite px don't remove shadow
  */
 
 #include "pico_int.h"
@@ -252,16 +253,18 @@ TileFlipMakerAS(TileFlipAS_onlymark, pix_sh_as_onlymark)
 // NB s/h already resolved by non-forced drawing
 // forced both layer draw (through debug reg)
 #define pix_and(x) \
-  pd[x] = (pd[x] & 0xc0) | (pd[x] & (pal | t))
+  pal |= 0xc0; /* leave s/h bits untouched in pixel "and" */ \
+  pd[x] &= pal | t
 
 TileNormMaker(TileNorm_and, pix_and)
 TileFlipMaker(TileFlip_and, pix_and)
 
 // forced sprite draw (through debug reg)
 #define pix_sh_as_and(x) \
+  pal |= 0xc0; /* leave s/h bits untouched in pixel "and" */ \
   if (m & (1<<(x+8))) { \
     m &= ~(1<<(x+8)); \
-    if (t<0xe) pd[x] = (pd[x] & 0xc0) | (pd[x] & (pal | t)); \
+    if (t<0xe) pd[x] &= pal | t; \
   }
  
 TileNormMakerAS(TileNormSH_AS_and, pix_sh_as_and)
@@ -356,11 +359,21 @@ static void DrawStripVSRam(struct TileStrip *ts, int plane_sh, int cellskip)
   dx+=cellskip<<3;
 
 //  int force = (plane_sh&LF_FORCE) << 13;
+  if ((cell&1)==1)
+  {
+    int line,vscroll;
+    vscroll = PicoMem.vsram[plane + (cell&0x3e)];
+
+    // Find the line in the name table
+    line=(vscroll+scan)&ts->line&0xffff; // ts->line is really ymask ..
+    nametabadd=(line>>3)<<(ts->line>>24);    // .. and shift[width]
+    ty=(line&7)<<1; // Y-Offset into tile
+  }
   for (; cell < ts->cells; dx+=8,tilex++,cell++)
   {
     u32 code, pack;
 
-    if ((cell&1)==0 || cell<0)
+    if ((cell&1)==0)
     {
       int line,vscroll;
       vscroll = PicoMem.vsram[plane + (cell&0x3e)];
@@ -1121,11 +1134,21 @@ static void DrawStripVSRamForced(struct TileStrip *ts, int plane_sh, int cellski
   tilex+=cellskip;
   dx+=cellskip<<3;
 
+  if ((cell&1)==1)
+  {
+    int line,vscroll;
+    vscroll = PicoMem.vsram[plane + (cell&0x3e)];
+
+    // Find the line in the name table
+    line=(vscroll+scan)&ts->line&0xffff; // ts->line is really ymask ..
+    nametabadd=(line>>3)<<(ts->line>>24);    // .. and shift[width]
+    ty=(line&7)<<1; // Y-Offset into tile
+  }
   for (; cell < ts->cells; dx+=8,tilex++,cell++)
   {
     unsigned int pack;
 
-    if ((cell&1)==0 || cell<0)
+    if ((cell&1)==0)
     {
       int line,vscroll;
       vscroll = PicoMem.vsram[plane + (cell&0x3e)];

pico/draw_arm.S

@@ -570,6 +570,32 @@ DrawLayer:
     @ r3=nametabadd[31:16]|must_be_0[15:8]|pal[7:0]
     @ r1=pd+dx r2=pack r6=hc r9=prevcode r11=HighCol r12=nametab lr=vram
     @ r4 & r7 are scratch in this loop
+
+    @ need to calc new ty?
+    lsls    r7, r10, #7           @ (cell&1)?
+    bmi     .dsloop_vs_subr1
+
+    @ calc offset and read tileline code to r7, also calc ty
+    add_c24 r7, lr, (OFS_PMEM_vsram-OFS_PMEM_vram)
+    and     r4, r10, #0x3e000000
+    add     r7, r7, r4, asr #23   @ vsram + ((cell&0x3e)<<1)
+    tst     r10,#0x8000           @ plane1?
+    addne   r7, r7, #2
+    ldrh    r7, [r7]              @ r7=vscroll
+
+    bic     r10,r10,#0xff         @ clear old ty
+    and     r4, r5, #0xff0000     @ scanline
+    add     r4, r4, r7, lsl #16   @ ... += vscroll
+    and     r4, r4, r5, lsl #16   @ ... &= ymask
+    and     r7, r4, #0x70000
+    orr     r10,r10,r7, lsr #15   @ new ty
+
+    mov     r4, r4, lsr #19
+    mov     r7, r5, lsr #24
+    mov     r4, r4, lsl r7        @ nametabadd
+    and     r3, r3, #0xff
+    orr     r3, r3, r4, lsl #16   @ r3=(nametabadd[31:16],pal[15:0])
+
 .dsloop_vs_subr1:
     sub     r1, r1, #8
 .dsloop_vs: @ 40-41 times
@@ -579,8 +605,7 @@ DrawLayer:
     ble     .dsloop_vs_exit
 
     @ need to calc new ty?
-    lsls    r7, r10, #7           @ (cell&1) && ...
+    lsls    r7, r10, #7           @ (cell&1)?
-    mvnmis  r7, r10               @ ... cell>=0
     bmi     0f
 
     @ calc offset and read tileline code to r7, also calc ty