vdp rendering fixes (debug register, vscroll) for overdrive 2

diff --git a/pico/draw.c b/pico/draw.c
index eeeb553..dff5e07 100644 (file)
--- a/pico/draw.c
+++ b/pico/draw.c
@@ -29,6 +29,7 @@
  */\r
 \r
 #include "pico_int.h"\r
+#define FORCE  // layer forcing via debug register?\r
 \r
 int (*PicoScanBegin)(unsigned int num) = NULL;\r
 int (*PicoScanEnd)  (unsigned int num) = NULL;\r
@@ -222,6 +223,7 @@ TileFlipMakerAS(TileFlipSH_AS_onlyop_lp, pix_sh_as_onlyop)
 TileNormMakerAS(TileNormAS_onlymark, pix_sh_as_onlymark)\r
 TileFlipMakerAS(TileFlipAS_onlymark, pix_sh_as_onlymark)\r
 \r
+#ifdef FORCE\r
 // forced both layer draw (through debug reg)\r
 #define pix_and(x) \\r
   pd[x] = (pd[x] & 0xc0) | (pd[x] & (pal | t))\r
@@ -230,12 +232,16 @@ TileNormMaker(TileNorm_and, pix_and)
 TileFlipMaker(TileFlip_and, pix_and)\r
 \r
 // forced sprite draw (through debug reg)\r
-#define pix_sh_and(x) /* XXX is there S/H with forced draw? */ \\r
-  if (t>=0xe) pd[x]=(pd[x]&0x3f)|(t<<6); /* c0 shadow, 80 hilight */ \\r
-  else pd[x] = (pd[x] & 0xc0) | (pd[x] & (pal | t))\r
+#define pix_sh_as_and(x) /* XXX is there S/H with forced draw? */ \\r
+  if (m & (1<<(x+8))) { \\r
+    m &= ~(1<<(x+8)); \\r
+    if (t>=0xe) pd[x]=(pd[x]&0x3f)|(t<<6); /* c0 shadow, 80 hilight */ \\r
+    else pd[x] = (pd[x] & 0xc0) | (pd[x] & (pal | t)); \\r
+  }\r
  \r
-TileNormMaker(TileNormSH_and, pix_sh_and)\r
-TileFlipMaker(TileFlipSH_and, pix_sh_and)\r
+TileNormMakerAS(TileNormSH_AS_and, pix_sh_as_and)\r
+TileFlipMakerAS(TileFlipSH_AS_and, pix_sh_as_and)\r
+#endif\r
 \r
 // --------------------------------------------\r
 \r
@@ -311,6 +317,7 @@ static void DrawStripVSRam(struct TileStrip *ts, int plane_sh, int cellskip)
     int adj = ((ts->hscroll ^ dx) >> 3) & 1;\r
     cell -= adj + 1;\r
     ts->cells -= adj;\r
+    PicoMem.vsram[0x3e] = PicoMem.vsram[0x3f] = plane_sh >> 16;\r
   }\r
   cell+=cellskip;\r
   tilex+=cellskip;\r
@@ -479,7 +486,7 @@ static void DrawLayer(int plane_sh, int *hcache, int cellskip, int maxcells,
     // shit, we have 2-cell column based vscroll\r
     // luckily this doesn't happen too often\r
     ts.line=ymask|(shift[width]<<24); // save some stuff instead of line\r
-    PicoMem.vsram[(plane_sh & 1)+0x3e] = PicoMem.vsram[0x27]; // XXX really?\r
+    plane_sh |= PicoMem.vsram[0x26+(~plane_sh&1)] << 16;\r
     DrawStripVSRam(&ts, plane_sh, cellskip);\r
   } else {\r
     vscroll = PicoMem.vsram[plane_sh & 1]; // Get vertical scroll value\r
@@ -778,28 +785,6 @@ static void DrawSprite(int *sprite, int sh, int w)
 }\r
 #endif\r
 \r
-static NOINLINE void DrawTilesFromCacheForced(const int *hc)\r
-{\r
-  unsigned char *pd = Pico.est.HighCol;\r
-  int code, addr, dx;\r
-  unsigned int pack;\r
-  int pal;\r
-\r
-  // *ts->hc++ = code | (dx<<16) | (ty<<25);\r
-  while ((code = *hc++)) {\r
-    // Get tile address/2:\r
-    addr = (code & 0x7ff) << 4;\r
-    addr += (code >> 25) & 0x0e; // y offset into tile\r
-\r
-    dx = (code >> 16) & 0x1ff;\r
-    pal = ((code >> 9) & 0x30);\r
-    pack = *(unsigned int *)(PicoMem.vram + addr);\r
-\r
-    if (code & 0x0800) TileFlip_and(pd + dx, pack, pal);\r
-    else               TileNorm_and(pd + dx, pack, pal);\r
-  }\r
-}\r
-\r
 static void DrawSpriteInterlace(unsigned int *sprite)\r
 {\r
   unsigned char *pd = Pico.est.HighCol;\r
@@ -1040,16 +1025,181 @@ static void DrawSpritesHiAS(unsigned char *sprited, int sh)
   }\r
 }\r
 \r
+#ifdef FORCE\r
+static void DrawStripForced(struct TileStrip *ts, int lflags, int cellskip)\r
+{\r
+  unsigned char *pd = Pico.est.HighCol;\r
+  int tilex,dx,ty,code=0,addr=0,cells;\r
+  int oldcode=-1;\r
+  int pal=0,sh;\r
+\r
+  // Draw tiles across screen:\r
+  sh = (lflags & LF_SH) << 5; // 0x40\r
+  tilex=((-ts->hscroll)>>3)+cellskip;\r
+  ty=(ts->line&7)<<1; // Y-Offset into tile\r
+  dx=((ts->hscroll-1)&7)+1;\r
+  cells = ts->cells - cellskip;\r
+  if(dx != 8) cells++; // have hscroll, need to draw 1 cell more\r
+  dx+=cellskip<<3;\r
+\r
+  for (; cells > 0; dx+=8, tilex++, cells--)\r
+  {\r
+    unsigned int pack;\r
+\r
+    code = PicoMem.vram[ts->nametab + (tilex & ts->xmask)];\r
+\r
+    if (code!=oldcode) {\r
+      oldcode = code;\r
+      // Get tile address/2:\r
+      addr=(code&0x7ff)<<4;\r
+      addr+=ty;\r
+      if (code&0x1000) addr^=0xe; // Y-flip\r
+\r
+      pal=((code>>9)&0x30)|sh;\r
+    }\r
+\r
+    pack = *(unsigned int *)(PicoMem.vram + addr);\r
+\r
+    if (code & 0x0800) TileFlip_and(pd + dx, pack, pal);\r
+    else               TileNorm_and(pd + dx, pack, pal);\r
+  }\r
+}\r
+\r
+// this is messy\r
+static void DrawStripVSRamForced(struct TileStrip *ts, int plane_sh, int cellskip)\r
+{\r
+  unsigned char *pd = Pico.est.HighCol;\r
+  int tilex,dx,code=0,addr=0,cell=0;\r
+  int oldcode=-1;\r
+  int pal=0,scan=Pico.est.DrawScanline;\r
+\r
+  // Draw tiles across screen:\r
+  tilex=(-ts->hscroll)>>3;\r
+  dx=((ts->hscroll-1)&7)+1;\r
+  if (ts->hscroll & 0x0f) {\r
+    int adj = ((ts->hscroll ^ dx) >> 3) & 1;\r
+    cell -= adj + 1;\r
+    ts->cells -= adj;\r
+    PicoMem.vsram[0x3e] = PicoMem.vsram[0x3f] = plane_sh >> 16;\r
+  }\r
+  cell+=cellskip;\r
+  tilex+=cellskip;\r
+  dx+=cellskip<<3;\r
+\r
+  for (; cell < ts->cells; dx+=8,tilex++,cell++)\r
+  {\r
+    int nametabadd, ty;\r
+    unsigned int pack;\r
+\r
+    //if((cell&1)==0)\r
+    {\r
+      int line,vscroll;\r
+      vscroll=PicoMem.vsram[(plane_sh&1)+(cell&0x3e)];\r
+\r
+      // Find the line in the name table\r
+      line=(vscroll+scan)&ts->line&0xffff; // ts->line is really ymask ..\r
+      nametabadd=(line>>3)<<(ts->line>>24);    // .. and shift[width]\r
+      ty=(line&7)<<1; // Y-Offset into tile\r
+    }\r
+\r
+    code=PicoMem.vram[ts->nametab+nametabadd+(tilex&ts->xmask)];\r
+\r
+    if (code!=oldcode) {\r
+      oldcode = code;\r
+      // Get tile address/2:\r
+      addr=(code&0x7ff)<<4;\r
+\r
+      pal=((code>>9)&0x30)|((plane_sh<<5)&0x40);\r
+    }\r
+\r
+    if (code & 0x1000) ty ^= 0xe; // Y-flip\r
+    pack = *(unsigned int *)(PicoMem.vram + addr+ty);\r
+\r
+    if (code & 0x0800) TileFlip_and(pd + dx, pack, pal);\r
+    else               TileNorm_and(pd + dx, pack, pal);\r
+  }\r
+}\r
+\r
+static void DrawLayerForced(int plane_sh, int cellskip, int maxcells,\r
+  struct PicoEState *est)\r
+{\r
+  struct PicoVideo *pvid=&Pico.video;\r
+  const char shift[4]={5,6,5,7}; // 32,64 or 128 sized tilemaps (2 is invalid)\r
+  struct TileStrip ts;\r
+  int width, height, ymask;\r
+  int vscroll, htab;\r
+\r
+  ts.cells=maxcells;\r
+\r
+  // Work out the TileStrip to draw\r
+\r
+  // Work out the name table size: 32 64 or 128 tiles (0-3)\r
+  width=pvid->reg[16];\r
+  height=(width>>4)&3; width&=3;\r
+\r
+  ts.xmask=(1<<shift[width])-1; // X Mask in tiles (0x1f-0x7f)\r
+  ymask=(height<<8)|0xff;       // Y Mask in pixels\r
+  switch (width) {\r
+    case 1: ymask &= 0x1ff; break;\r
+    case 2: ymask =  0x007; break;\r
+    case 3: ymask =  0x0ff; break;\r
+  }\r
+\r
+  // Find name table:\r
+  if (plane_sh&1) ts.nametab=(pvid->reg[4]&0x07)<<12; // B\r
+  else            ts.nametab=(pvid->reg[2]&0x38)<< 9; // A\r
+\r
+  htab=pvid->reg[13]<<9; // Horizontal scroll table address\r
+  switch (pvid->reg[11]&3) {\r
+    case 1: htab += (est->DrawScanline<<1) &  0x0f; break;\r
+    case 2: htab += (est->DrawScanline<<1) & ~0x0f; break; // Offset by tile\r
+    case 3: htab += (est->DrawScanline<<1);         break; // Offset by line\r
+  }\r
+  htab+=plane_sh&1; // A or B\r
+\r
+  // Get horizontal scroll value, will be masked later\r
+  ts.hscroll = PicoMem.vram[htab & 0x7fff];\r
+\r
+  if((pvid->reg[12]&6) == 6) {\r
+    // interlace mode 2\r
+    vscroll = PicoMem.vsram[plane_sh & 1]; // Get vertical scroll value\r
+\r
+    // Find the line in the name table\r
+    ts.line=(vscroll+(est->DrawScanline<<1))&((ymask<<1)|1);\r
+    ts.nametab+=(ts.line>>4)<<shift[width];\r
+\r
+    DrawStripInterlace(&ts);\r
+  } else if( pvid->reg[11]&4) {\r
+    // shit, we have 2-cell column based vscroll\r
+    // luckily this doesn't happen too often\r
+    ts.line=ymask|(shift[width]<<24); // save some stuff instead of line\r
+    plane_sh |= PicoMem.vsram[0x26+(~plane_sh&1)] << 16;\r
+    DrawStripVSRamForced(&ts, plane_sh, cellskip);\r
+  } else {\r
+    vscroll = PicoMem.vsram[plane_sh & 1]; // Get vertical scroll value\r
+\r
+    // Find the line in the name table\r
+    ts.line=(vscroll+est->DrawScanline)&ymask;\r
+    ts.nametab+=(ts.line>>3)<<shift[width];\r
+\r
+    DrawStripForced(&ts, plane_sh, cellskip);\r
+  }\r
+}\r
+\r
+// rather messy (XXX revisit layer compositing)\r
 static void DrawSpritesForced(unsigned char *sprited)\r
 {\r
-  void (*fTileFunc)(unsigned char *pd, unsigned int pack, int pal);\r
+  unsigned (*fTileFunc)(unsigned char *pd, unsigned m, unsigned int pack, int pal);\r
   unsigned char *pd = Pico.est.HighCol;\r
-  unsigned char *p;\r
+  unsigned char mb[1+320+1];\r
+  unsigned char *p, *mp;\r
+  unsigned m;\r
   int entry, cnt;\r
 \r
   cnt = sprited[0] & 0x7f;\r
-  if (cnt == 0) return;\r
-\r
+  if (cnt == 0) { memset(pd, 0, sizeof(DefHighCol)); return; }\r
+  \r
+  memset(mb, 0xff, sizeof(mb));\r
   p = &sprited[4];\r
   if ((sprited[1] & (SPRL_TILE_OVFL|SPRL_HAVE_MASK0)) == (SPRL_TILE_OVFL|SPRL_HAVE_MASK0))\r
     return; // masking effective due to tile overflow\r
@@ -1065,8 +1215,8 @@ static void DrawSpritesForced(unsigned char *sprited)
     code = sprite[1];\r
     pal = (code>>9)&0x30;\r
 \r
-    if (code&0x800) fTileFunc = TileFlipSH_and;\r
-    else            fTileFunc = TileNormSH_and;\r
+    if (code&0x800) fTileFunc = TileFlipSH_AS_and;\r
+    else            fTileFunc = TileNormSH_AS_and;\r
 \r
     // parse remaining sprite data\r
     sy=sprite[0];\r
@@ -1087,7 +1237,8 @@ static void DrawSpritesForced(unsigned char *sprited)
     delta<<=4; // Delta of address\r
 \r
     if (entry+1 == cnt) width = p[entry+1]; // last sprite width limited?\r
-    for (; width; width--,sx+=8,tile+=delta)\r
+    mp = mb+(sx>>3);\r
+    for (m = *mp; width; width--, sx+=8, *mp++ = m, m >>= 8, tile+=delta)\r
     {\r
       unsigned int pack;\r
 \r
@@ -1095,10 +1246,25 @@ static void DrawSpritesForced(unsigned char *sprited)
       if(sx>=328) break; // Offscreen\r
 \r
       pack = *(unsigned int *)(PicoMem.vram + (tile & 0x7fff));\r
-      fTileFunc(pd + sx, pack, pal);\r
-    }\r
+\r
+      m |= mp[1] << 8; // next mask byte\r
+      // shift mask bits to bits 8-15 for easier load/store handling\r
+      m = fTileFunc(pd + sx, m << (8-(sx&0x7)), pack, pal) >> (8-(sx&0x7));\r
+    } \r
+    *mp = m; // write last mask byte\r
   }\r
+\r
+  // anything not covered by a sprite is off (XXX or bg?)\r
+  for (cnt = 1; cnt < sizeof(mb)-1; cnt++)\r
+    if (mb[cnt] == 0xff)\r
+      for (m = 0; m < 8; m++)\r
+        pd[8*cnt+m] = 0;\r
+    else if (mb[cnt])\r
+      for (m = 0; m < 8; m++)\r
+        if (mb[cnt] & (1<<m))\r
+          pd[8*cnt+m] = 0;\r
 }\r
+#endif\r
 \r
 \r
 // Index + 0  :    ----hhvv -lllllll -------y yyyyyyyy\r
@@ -1462,14 +1628,10 @@ static int DrawDisplay(int sh)
   /* - layer B low - */\r
   if (!(pvid->debug_p & PVD_KILL_B)) {\r
     lflags = LF_PLANE_1 | (sh << 1);\r
-    if (pvid->debug_p & PVD_FORCE_B)\r
-      lflags |= LF_FORCE;\r
     DrawLayer(lflags, HighCacheB, 0, maxcells, est);\r
   }\r
   /* - layer A low - */\r
   lflags = 0 | (sh << 1);\r
-  if (pvid->debug_p & PVD_FORCE_A)\r
-    lflags |= LF_FORCE;\r
   if (pvid->debug_p & PVD_KILL_A)\r
     ;\r
   else if (hvwind == 1)\r
@@ -1516,12 +1678,16 @@ static int DrawDisplay(int sh)
   else if (sprited[1] & SPRL_HAVE_HI)\r
     DrawAllSprites(sprited, 1, 0, est);\r
 \r
-  if (pvid->debug_p & PVD_FORCE_B)\r
-    DrawTilesFromCacheForced(HighCacheB);\r
-  else if (pvid->debug_p & PVD_FORCE_A)\r
-    DrawTilesFromCacheForced(HighCacheA);\r
-  else if (pvid->debug_p & PVD_FORCE_S)\r
+#ifdef FORCE\r
+  if (pvid->debug_p & PVD_FORCE_B) {\r
+    lflags = LF_PLANE_1 | (sh << 1);\r
+    DrawLayerForced(lflags, 0, maxcells, est);\r
+  } else if (pvid->debug_p & PVD_FORCE_A) {\r
+    lflags = (sh << 1);\r
+    DrawLayerForced(lflags, 0, maxcells, est);\r
+  } else if (pvid->debug_p & PVD_FORCE_S)\r
     DrawSpritesForced(sprited);\r
+#endif\r
 \r
 #if 0\r
   {\r

diff --git a/pico/draw_arm.S b/pico/draw_arm.S
index 8dc660c..1a0f351 100644 (file)
--- a/pico/draw_arm.S
+++ b/pico/draw_arm.S
@@ -545,8 +545,12 @@ DrawLayer:
     eor     r3, r3, r7\r
     sub     r10,r10, #1<<24       @ cell--  // start from negative for hscroll\r
     tst     r3, #0x08\r
+    add_c24 r1, lr, (OFS_PMEM_vsram-OFS_PMEM_vram)\r
+    ldr     r3, [r1, #0x4c]       @ r3=vsram[0x26..0x27]\r
     subne   r10,r10, #1<<16       @ cells--\r
     subne   r10,r10, #1<<24       @ cell--  // even more negative\r
+    ror     r3, r3, #16\r
+    str     r3, [r1, #0x7c]       @ vsram[0x3e..0x3f]=r3\r
 0:\r
     tst     r9, #1<<31\r
     mov     r3, #0\r
@@ -577,8 +581,8 @@ DrawLayer:
 \r
     @ calc offset and read tileline code to r7, also calc ty\r
     add_c24 r7, lr, (OFS_PMEM_vsram-OFS_PMEM_vram)\r
-    add     r7, r7, r10,asr #23   @ vsram + ((cell&~1)<<1)\r
-    bic     r7, r7, #3\r
+    and     r4, r10, #0x3e000000\r
+    add     r7, r7, r4, asr #23   @ vsram + ((cell&0x3e)<<1)\r
     tst     r10,#0x8000           @ plane1?\r
     addne   r7, r7, #2\r
     ldrh    r7, [r7]              @ r7=vscroll\r