core vdp, fix sprite rendering issues with Overdrive 1/2

pico/draw.c

@@ -97,7 +97,9 @@ u32 VdpSATCache[2*128];  // VDP sprite cache (1st 32 sprite attr bits)
 #define SPRL_HAVE_MASK0  0x02 // have sprite with x == 0 in 1st slot
 #define SPRL_MASKED      0x01 // lo prio masking by sprite with x == 0 active
 
-unsigned char HighLnSpr[240][4+MAX_LINE_SPRITES+1]; // sprite_count, ^flags, tile_count, sprites_total, [spritep]..., last_width
+// sprite cache. stores results of sprite parsing for each display line:
+// [visible_sprites_count, sprl_flags, tile_count, sprites_processed, sprite_idx[sprite_count], last_width]
+unsigned char HighLnSpr[240][4+MAX_LINE_SPRITES+1];
 
 int rendstatus_old;
 int rendlines;
@@ -290,7 +292,8 @@ TileFlipMaker(TileFlip_and, pix_and)
   pal |= 0xc0; /* leave s/h bits untouched in pixel "and" */ \
   if (likely(m & (1<<(x+8)))) { \
     m &= ~(1<<(x+8)); \
-    if (t<0xe) pd[x] &= pal|t; \
+    /* if (!t) pd[x] |= 0x40; as per titan hw notes? */ \
+    pd[x] &= pal|t; \
   }
  
 TileNormMakerAS(TileNormSH_AS_and, pix_sh_as_and)
@@ -977,7 +980,7 @@ static void DrawSpritesSHi(unsigned char *sprited, const struct PicoEState *est)
   unsigned char *p;
   int cnt, w;
 
-  cnt = sprited[0] & 0x7f;
+  cnt = sprited[0];
   if (cnt == 0) return;
 
   p = &sprited[4];
@@ -1046,7 +1049,7 @@ static void DrawSpritesHiAS(unsigned char *sprited, int sh)
   unsigned m;
   int entry, cnt;
 
-  cnt = sprited[0] & 0x7f;
+  cnt = sprited[0];
   if (cnt == 0) return;
 
   memset(mb, 0xff, sizeof(mb));
@@ -1323,7 +1326,7 @@ static void DrawSpritesForced(unsigned char *sprited)
   unsigned m;
   int entry, cnt;
 
-  cnt = sprited[0] & 0x7f;
+  cnt = sprited[0];
   if (cnt == 0) { memset(pd, 0, sizeof(DefHighCol)); return; }
   
   memset(mb, 0xff, sizeof(mb));
@@ -1382,7 +1385,9 @@ static void DrawSpritesForced(unsigned char *sprited)
     *mp = m; // write last mask byte
   }
 
-  // anything not covered by a sprite is off (XXX or bg?)
+  // anything not covered by a sprite is off 
+  // XXX Titan hw notes say that transparent pixels remove shadow. Is this also
+  // the case in areas where no sprites are displayed?
   for (cnt = 1; cnt < sizeof(mb)-1; cnt++)
     if (mb[cnt] == 0xff) {
       *(u32 *)(pd+8*cnt+0) = 0;
@@ -1401,7 +1406,7 @@ static void DrawSpritesForced(unsigned char *sprited)
 // Index + 0  :    hhhhvvvv ----hhvv yyyyyyyy yyyyyyyy // v, h: vert./horiz. size
 // Index + 4  :    xxxxxxxx xxxxxxxx pccvhnnn nnnnnnnn // x: x coord + 8
 
-static NOINLINE void ParseSprites(int max_lines)
+static NOINLINE void ParseSprites(int max_lines, int limit)
 {
   const struct PicoVideo *pvid=&Pico.video;
   const struct PicoEState *est=&Pico.est;
@@ -1417,6 +1422,10 @@ static NOINLINE void ParseSprites(int max_lines)
   if (max_lines > rendlines-1)
     max_lines = rendlines-1;
 
+  // look-ahead SAT parsing for next line and sprite pixel fetching for current
+  // line are limited if display was disabled during HBLANK before current line
+  if (limit) limit = 16; // max sprites/pixels processed
+
   if (!(Pico.video.reg[12]&1))
     max_sprites = 64, max_line_sprites = 16, max_width = 264;
   if (*est->PicoOpt & POPT_DIS_SPRITE_LIM)
@@ -1453,6 +1462,8 @@ static NOINLINE void ParseSprites(int max_lines)
     if (sy <= max_lines && sy + (height<<3) >= first_line) // sprite onscreen (y)?
     {
       int entry, y, w, sx_min, onscr_x, maybe_op = 0;
+      // omit look-ahead line if sprite parsing limit reached
+      int last_line = (limit && u >= 2*limit ? max_lines-1 : max_lines);
 
       sx_min = 8-(width<<3);
       onscr_x = sx_min < sx && sx < max_width;
@@ -1461,13 +1472,15 @@ static NOINLINE void ParseSprites(int max_lines)
 
       entry = ((pd - HighPreSpr) / 2) | ((code2>>8)&0x80);
       y = (sy >= first_line) ? sy : first_line;
-      for (; y < sy + (height<<3) && y <= max_lines; y++)
+      for (; y < sy + (height<<3) && y <= last_line; y++)
       {
         unsigned char *p = &HighLnSpr[y][0];
         int cnt = p[0];
-        if (p[3] >= max_line_sprites) continue;         // sprite limit?
         if (p[1] & SPRL_MASKED) continue;               // masked?
 
+        if (p[3] >= max_line_sprites) continue;         // sprite limit?
+        p[3] ++;
+
         w = width;
         if (p[2] + width > max_line_sprites*2) {        // tile limit?
           if (y+1 < 240) HighLnSpr[y+1][1] |= SPRL_TILE_OVFL;
@@ -1475,7 +1488,6 @@ static NOINLINE void ParseSprites(int max_lines)
           w = max_line_sprites*2 - p[2];
         }
         p[2] += w;
-        p[3] ++;
 
         if (sx == -0x78) {
           if (p[1] & (SPRL_HAVE_X|SPRL_TILE_OVFL))
@@ -1487,13 +1499,15 @@ static NOINLINE void ParseSprites(int max_lines)
 
         if (!onscr_x) continue; // offscreen x
 
-        p[4+cnt] = entry;
+        // sprite is (partly) visible, store info for renderer
-        p[5+cnt] = w; // width clipped by tile limit for sprite renderer
-        p[0] = cnt + 1;
         p[1] |= (entry & 0x80) ? SPRL_HAVE_HI : SPRL_HAVE_LO;
         p[1] |= maybe_op; // there might be op sprites on this line
         if (cnt > 0 && (code2 & 0x8000) && !(p[4+cnt-1]&0x80))
           p[1] |= SPRL_LO_ABOVE_HI;
+
+        p[4+cnt] = entry;
+        p[5+cnt] = w; // width clipped by tile limit for sprite renderer
+        p[0] = cnt + 1;
       }
     }
 
@@ -1506,6 +1520,23 @@ static NOINLINE void ParseSprites(int max_lines)
   }
   *pd = 0;
 
+  // fetching sprite pixels isn't done while display is disabled during HBLANK
+  if (limit) {
+    int w = 0;
+    unsigned char *sprited = &HighLnSpr[max_lines-1][0]; // current render line
+
+    for (u = 0; u < sprited[0]; u++) {
+      s32 *sp = HighPreSpr + (sprited[4+u] & 0x7f) * 2;
+      int sw = sp[0] >> 28;
+      if (w + sw > limit) {
+        sprited[0] = u;
+        sprited[4+u] = limit-w;
+        break;
+      }
+      w += sw;
+    }
+  }
+
 #if 0
   for (u = first_line; u <= max_lines; u++)
   {
@@ -1528,7 +1559,7 @@ static void DrawAllSprites(unsigned char *sprited, int prio, int sh,
   unsigned char *p;
   int cnt, w;
 
-  cnt = sprited[0] & 0x7f;
+  cnt = sprited[0];
   if (cnt == 0) return;
 
   p = &sprited[4];
@@ -1974,7 +2005,7 @@ static void PicoLine(int line, int offs, int sh, int bgc)
   Pico.est.DrawLineDest = (char *)Pico.est.DrawLineDest + DrawLineDestIncrement;
 }
 
-void PicoDrawSync(int to, int blank_last_line)
+void PicoDrawSync(int to, int blank_last_line, int limit_sprites)
 {
   struct PicoEState *est = &Pico.est;
   int line, offs = 0;
@@ -1990,7 +2021,7 @@ void PicoDrawSync(int to, int blank_last_line)
   }
   if (est->DrawScanline <= to && (est->rendstatus &
                 (PDRAW_SPRITES_MOVED|PDRAW_DIRTY_SPRITES|PDRAW_PARSE_SPRITES)))
-    ParseSprites(to + 1);
+    ParseSprites(to + 1, limit_sprites);
 
   for (line = est->DrawScanline; line < to; line++)
     PicoLine(line, offs, sh, bgc);

pico/pico.c

@@ -286,7 +286,7 @@ void PicoFrameDrawOnly(void)
 {
   if (!(PicoIn.AHW & PAHW_SMS)) {
     PicoFrameStart();
-    PicoDrawSync(Pico.m.pal?239:223, 0);
+    PicoDrawSync(Pico.m.pal?239:223, 0, 0);
   } else {
     PicoFrameDrawOnlyMS();
   }

pico/pico_cmn.c

@@ -167,7 +167,7 @@ static int PicoFrameHints(void)
   if (!skip)
   {
     if (Pico.est.DrawScanline < y)
-      PicoDrawSync(y - 1, 0);
+      PicoDrawSync(y - 1, 0, 0);
 #ifdef DRAW_FINISH_FUNC
     DRAW_FINISH_FUNC();
 #endif

pico/pico_int.h

@@ -692,7 +692,7 @@ int CM_compareRun(int cyc, int is_sub);
 // draw.c
 void PicoDrawInit(void);
 PICO_INTERNAL void PicoFrameStart(void);
-void PicoDrawSync(int to, int blank_last_line);
+void PicoDrawSync(int to, int blank_last_line, int limit_sprites);
 void BackFill(int reg7, int sh, struct PicoEState *est);
 void FinalizeLine555(int sh, int line, struct PicoEState *est);
 void FinalizeLine8bit(int sh, int line, struct PicoEState *est);

pico/videoport.c

@@ -139,6 +139,7 @@ void PicoVideoInit(void)
 
 
 static int blankline;           // display disabled for this line
+static int limitsprites;
 
 u32 SATaddr, SATmask;      // VRAM addr of sprite attribute table
 
@@ -793,6 +794,11 @@ static NOINLINE void CommandChange(struct PicoVideo *pvid)
 
 // VDP interface
  
+static inline int InHblank(int offs)
+{
+  return SekCyclesDone() - Pico.t.m68c_line_start <= 488-offs;
+}
+
 static void DrawSync(int skip)
 {
   int lines = Pico.video.reg[1]&0x08 ? 240 : 224;
@@ -802,10 +808,12 @@ static void DrawSync(int skip)
       !PicoIn.skipFrame && Pico.est.DrawScanline <= last) {
     //elprintf(EL_ANOMALY, "sync");
     if (blankline >= 0 && blankline < last) {
-      PicoDrawSync(blankline, 1);
+      PicoDrawSync(blankline, 1, 0);
       blankline = -1;
     }
-    PicoDrawSync(last, 0);
+    PicoDrawSync(last, 0, last == limitsprites);
+    if (last >= limitsprites)
+      limitsprites = -1;
   }
 }
 
@@ -820,18 +828,18 @@ PICO_INTERNAL_ASM void PicoVideoWrite(u32 a,unsigned short d)
   switch (a)
   {
   case 0x00: // Data port 0 or 2
-    // try avoiding the sync..
-    if (Pico.m.scanline < (pvid->reg[1]&0x08 ? 240 : 224) && (pvid->reg[1]&0x40) &&
-        !(!pvid->pending && ((pvid->command & 0xc00000f0) == 0x40000010 &&
-                        PicoMem.vsram[(pvid->addr>>1) & 0x3f] == (d & 0x7ff)))
-       )
-      DrawSync(0); // XXX  it's unclear when vscroll data is fetched from vsram?
-
     if (pvid->pending) {
       CommandChange(pvid);
       pvid->pending=0;
     }
 
+    // try avoiding the sync. can't easily do this for VRAM writes since they
+    // might update the SAT cache
+    if (Pico.m.scanline < (pvid->reg[1]&0x08 ? 240 : 224) && (pvid->reg[1]&0x40) &&
+        !(pvid->type == 3 && PicoMem.cram[(pvid->addr>>1) & 0x3f] == (d & 0xeee)) &&
+        !(pvid->type == 5 && PicoMem.vsram[(pvid->addr>>1) & 0x3f] == (d & 0x7ff)))
+      DrawSync(InHblank(440)); // experimentally, Overdrive 2
+
     if (!(PicoIn.opt&POPT_DIS_VDP_FIFO))
     {
       VdpFIFO.fifo_data[++VdpFIFO.fifo_dx&3] = d;
@@ -862,7 +870,7 @@ PICO_INTERNAL_ASM void PicoVideoWrite(u32 a,unsigned short d)
       CommandChange(pvid);
       // Check for dma:
       if (d & 0x80) {
-        DrawSync(SekCyclesDone() - Pico.t.m68c_line_start <= 488-390);
+        DrawSync(InHblank(390));
         CommandDma();
       }
     }
@@ -884,12 +892,17 @@ PICO_INTERNAL_ASM void PicoVideoWrite(u32 a,unsigned short d)
         if (num == 1 && ((pvid->reg[1]^d)&0x40)) {
           PicoVideoFIFOMode(d & 0x40, pvid->reg[12]&1);
           // handle line blanking before line rendering
-          if (SekCyclesDone() - Pico.t.m68c_line_start <= 488-390)
+          if (InHblank(390)) {
-            blankline = d&0x40 ? -1 : Pico.m.scanline;
+            // sprite rendering is limited if display is disabled and reenabled
+            // in HBLANK of the same scanline (Overdrive)
+            limitsprites = (d&0x40) && blankline == Pico.m.scanline ? Pico.m.scanline : -1;
+            blankline = (d&0x40) ? -1 : Pico.m.scanline;
+          }
         }
         if (num == 12 && ((pvid->reg[12]^d)&0x01))
           PicoVideoFIFOMode(pvid->reg[1]&0x40, d & 1);
-        DrawSync(SekCyclesDone() - Pico.t.m68c_line_start <= 488-390);
+        if (num <= 18) // no sync needed for DMA setup registers
+          DrawSync(InHblank(390));
         d &= 0xff;
         pvid->reg[num]=(unsigned char)d;
         switch (num)