core vdp, avoid rendering if no changes to RAMs/registers

mainly benefits 32X in faster/fastest rendering mode, 0-5% saved

pico/32x/32x.c

@@ -587,6 +587,8 @@ void PicoFrame32x(void)
     pcd_prepare_frame();
 
   PicoFrameStart();
+  if (Pico32xDrawMode != PDM32X_BOTH)
+    Pico.est.rendstatus |= PDRAW_SYNC_NEEDED;
   PicoFrameHints();
 
   elprintf(EL_32X, "poll: %02x %02x %02x",

pico/draw.c

@@ -1910,6 +1910,7 @@ PICO_INTERNAL void PicoFrameStart(void)
   int loffs = 8, lines = 224, coffs = 0, columns = 320;
   int sprep = est->rendstatus & PDRAW_DIRTY_SPRITES;
   int skipped = est->rendstatus & PDRAW_SKIP_FRAME;
+  int sync = est->rendstatus & (PDRAW_SYNC_NEEDED | PDRAW_SYNC_NEXT);
 
   // prepare to do this frame
   est->rendstatus = 0;
@@ -1935,7 +1936,12 @@ PICO_INTERNAL void PicoFrameStart(void)
     // mode_change() might reset rendstatus_old by calling SetColorFormat
     emu_video_mode_change(loffs, lines, coffs, columns);
     rendstatus_old = est->rendstatus & (PDRAW_INTERLACE|PDRAW_32_COLS|PDRAW_30_ROWS);
+    // mode_change() might clear buffers, redraw needed
+    est->rendstatus |= PDRAW_SYNC_NEEDED;
   }
+
+  if (sync | skipped)
+    est->rendstatus |= PDRAW_SYNC_NEEDED;
   if (PicoIn.skipFrame) // preserve this until something is rendered at last
     est->rendstatus |= PDRAW_SKIP_FRAME;
   if (sprep | skipped)
@@ -2030,6 +2036,14 @@ void PicoDrawSync(int to, int blank_last_line, int limit_sprites)
   if (est->DrawScanline <= to &&
                 (est->rendstatus & (PDRAW_DIRTY_SPRITES|PDRAW_PARSE_SPRITES)))
     ParseSprites(to + 1, limit_sprites);
+  else if (!(est->rendstatus & PDRAW_SYNC_NEEDED)) {
+    // nothing has changed in VDP/VRAM and buffer is the same -> no sync needed
+    int count = to+1 - est->DrawScanline;
+    est->HighCol += count*HighColIncrement;
+    est->DrawLineDest = (char *)est->DrawLineDest + count*DrawLineDestIncrement;
+    est->DrawScanline = to+1;
+    return;
+  }
 
   for (line = est->DrawScanline; line < to; line++)
     PicoLine(line, offs, sh, bgc);
@@ -2135,6 +2149,8 @@ void PicoDrawSetOutBufMD(void *dest, int increment)
     PicoDrawSetInternalBuf(dest, increment); // needed for SMS
     PicoDraw2SetOutBuf(dest, increment);
   } else if (dest != NULL) {
+    if (dest != DrawLineDestBase)
+      Pico.est.rendstatus |= PDRAW_SYNC_NEEDED;
     DrawLineDestBase = dest;
     DrawLineDestIncrement = increment;
     Pico.est.DrawLineDest = (char *)DrawLineDestBase + Pico.est.DrawScanline * increment;
@@ -2157,6 +2173,8 @@ void PicoDrawSetOutBuf(void *dest, int increment)
 void PicoDrawSetInternalBuf(void *dest, int increment)
 {
   if (dest != NULL) {
+    if (dest != HighColBase)
+      Pico.est.rendstatus |= PDRAW_SYNC_NEEDED;
     HighColBase = dest;
     HighColIncrement = increment;
     Pico.est.HighCol = HighColBase + Pico.est.DrawScanline * increment;

pico/pico.h

@@ -223,6 +223,7 @@ void vidConvCpyRGB565(void *to, void *from, int pixels);
 #endif
 void PicoDoHighPal555(int sh, int line, struct PicoEState *est);
 // internals, NB must keep in sync with ASM draw functions
+#define PDRAW_SYNC_NEEDED   (1<<0) // redraw needed
 #define PDRAW_WND_DIFF_PRIO (1<<1) // not all window tiles use same priority
 #define PDRAW_PARSE_SPRITES (1<<2) // SAT needs parsing
 #define PDRAW_INTERLACE     (1<<3)
@@ -236,6 +237,7 @@ void PicoDoHighPal555(int sh, int line, struct PicoEState *est);
 #define PDRAW_30_ROWS      (1<<11) // 30 rows mode (240 lines)
 #define PDRAW_32X_SCALE    (1<<12) // scale CLUT layer for 32X
 #define PDRAW_SMS_BLANK_1  (1<<13) // 1st column blanked
+#define PDRAW_SYNC_NEXT    (1<<14) // sync next frame
 extern int rendstatus_old;
 extern int rendlines;
 

pico/pico_cmn.c

@@ -168,10 +168,12 @@ static int PicoFrameHints(void)
     {
       // find the right moment for frame renderer, when display is no longer blanked
       if ((pv->reg[1]&0x40) || y > 100) {
+        if (Pico.est.rendstatus & PDRAW_SYNC_NEEDED)
           PicoFrameFull();
 #ifdef DRAW_FINISH_FUNC
         DRAW_FINISH_FUNC();
 #endif
+        Pico.est.rendstatus &= ~PDRAW_SYNC_NEEDED;
         skip = 1;
       }
     }
@@ -191,10 +193,11 @@ static int PicoFrameHints(void)
   if (!skip)
   {
     if (Pico.est.DrawScanline < y)
-      PicoDrawSync(y - 1, 0, 0);
+      PicoVideoSync(-1);
 #ifdef DRAW_FINISH_FUNC
     DRAW_FINISH_FUNC();
 #endif
+    Pico.est.rendstatus &= ~PDRAW_SYNC_NEEDED;
   }
 #ifdef PICO_32X
   p32x_render_frame();

pico/pico_int.h

@@ -917,6 +917,7 @@ int PicoVideoFIFOHint(void);
 void PicoVideoFIFOMode(int active, int h40);
 int PicoVideoFIFOWrite(int count, int byte_p, unsigned sr_mask, unsigned sr_flags);
 void PicoVideoInit(void);
+void PicoVideoSync(int skip);
 void PicoVideoSave(void);
 void PicoVideoLoad(void);
 void PicoVideoCacheSAT(int load);

pico/videoport.c

@@ -809,13 +809,17 @@ static inline int InHblank(int offs)
   return SekCyclesDone() - Pico.t.m68c_line_start <= offs;
 }
 
-static void DrawSync(int skip)
+void PicoVideoSync(int skip)
 {
   int lines = Pico.video.reg[1]&0x08 ? 240 : 224;
-  int last = Pico.m.scanline - (skip || blankline == Pico.m.scanline);
+  int last = Pico.m.scanline - ((skip > 0) || blankline == Pico.m.scanline);
+
+  if (!(PicoIn.opt & POPT_ALT_RENDERER) && !PicoIn.skipFrame) {
+    if (last >= lines)
+      last = lines-1;
+    else // in active display, need to sync next frame as well
+      Pico.est.rendstatus |= PDRAW_SYNC_NEXT;
 
-  if (last < lines && !(PicoIn.opt & POPT_ALT_RENDERER) &&
-      !PicoIn.skipFrame && Pico.est.DrawScanline <= last) {
     //elprintf(EL_ANOMALY, "sync");
     if (blankline >= 0 && blankline < last) {
       PicoDrawSync(blankline, 1, 0);
@@ -825,6 +829,8 @@ static void DrawSync(int skip)
     if (last >= limitsprites)
       limitsprites = -1;
   }
+  if (skip >= 0)
+    Pico.est.rendstatus |= PDRAW_SYNC_NEEDED;
 }
 
 PICO_INTERNAL_ASM void PicoVideoWrite(u32 a,unsigned short d)
@@ -849,7 +855,11 @@ PICO_INTERNAL_ASM void PicoVideoWrite(u32 a,unsigned short d)
         !(pvid->type == 1 && !(pvid->addr&1) && ((pvid->addr^SATaddr)&SATmask) && PicoMem.vram[pvid->addr>>1] == d) &&
         !(pvid->type == 3 && PicoMem.cram[(pvid->addr>>1) & 0x3f] == (d & 0xeee)) &&
         !(pvid->type == 5 && PicoMem.vsram[(pvid->addr>>1) & 0x3f] == (d & 0x7ff)))
-      DrawSync(InHblank(48)); // experimentally, Overdrive 2
+      // the vertical scroll value for this line must be read from VSRAM early,
+      // since the A/B tile row to be read depends on it. E.g. Skitchin, OD2
+      // in contrast, CRAM writes would have an immediate effect on the current
+      // pixel. XXX think about different offset values for different RAM types
+      PicoVideoSync(InHblank(30));
 
     if (!(PicoIn.opt&POPT_DIS_VDP_FIFO))
     {
@@ -881,7 +891,7 @@ PICO_INTERNAL_ASM void PicoVideoWrite(u32 a,unsigned short d)
       CommandChange(pvid);
       // Check for dma:
       if (d & 0x80) {
-        DrawSync(InHblank(93));
+        PicoVideoSync(InHblank(93));
         CommandDma();
       }
     }
@@ -913,8 +923,9 @@ PICO_INTERNAL_ASM void PicoVideoWrite(u32 a,unsigned short d)
         }
         if (num == 12 && ((pvid->reg[12]^d)&0x01))
           PicoVideoFIFOMode(pvid->reg[1]&0x40, d & 1);
-        if (num <= 18 && pvid->reg[num] != d) // no sync for DMA setup
+        if (((1<<num) & 0x738ff) && pvid->reg[num] != d)
-          DrawSync(InHblank(93)); // Toy Story
+          // VDP regs 0-7,11-13,16-18 influence rendering, ignore all others
+          PicoVideoSync(InHblank(93)); // Toy Story
         pvid->reg[num] = d;
 
         switch (num)