From 724db457da104aa9298bdeb861e540e0e2ab42fd Mon Sep 17 00:00:00 2001
From: kub <derkub@gmail.com>
Date: Sat, 13 Jan 2024 12:07:13 +0100
Subject: [PATCH] core, handle background color DMA (aka fantom bitmap)

---
 pico/draw.c      | 86 +++++++++++++++++++++++++++++++++++++++++++++++-
 pico/pico.c      |  1 +
 pico/pico_int.h  |  1 +
 pico/videoport.c | 10 ++++++
 4 files changed, 97 insertions(+), 1 deletion(-)

diff --git a/pico/draw.c b/pico/draw.c
index 66b4fca5..40903e74 100644
--- a/pico/draw.c
+++ b/pico/draw.c
@@ -1599,7 +1599,59 @@ void BackFill(int bgc, int sh, struct PicoEState *est)
 
 // --------------------------------------------
 
-void PicoDoHighPal555_8bit(int sh, int line, struct PicoEState *est)
+static u16 *BgcDMAbase;
+static u32 BgcDMAsrc, BgcDMAmask;
+static int BgcDMAlen, BgcDMAoffs;
+
+#ifndef _ASM_DRAW_C
+static
+#endif
+// handle DMA to background color
+int BgcDMA(u16 *pd, int len, struct PicoEState *est)
+{
+  int xl = (len == 320 ? 38 : 33); // DMA slots during HSYNC
+  int upscale = (est->rendstatus & PDRAW_SOFTSCALE) && len < 320;
+
+  if (BgcDMAlen > 0) {
+    // BG color DMA under way. TODO for now handles the line as all background.
+    int i, l = len;
+    u16 *q = upscale ? DefOutBuff : pd;
+    u16 t;
+
+    if ((est->rendstatus & PDRAW_BORDER_32) && !upscale)
+      q += (320-len) / 2;
+
+    BgcDMAlen -= (l>>1)+xl;
+    if (BgcDMAlen < 0)
+      // partial line
+      l += 2*BgcDMAlen;
+
+    for (i = 0; i < l; i += 2) {
+      // TODO use ps to overwrite only real bg pixels
+      t = BgcDMAbase[BgcDMAsrc++ & BgcDMAmask];
+      q[i] = q[i+1] = PXCONV(t);
+    }
+    BgcDMAsrc += xl; // HSYNC DMA
+
+    t = est->HighPal[Pico.video.reg[7] & 0x3f];
+    while (i < len) q[i++] = t; // fill partial line with BG
+
+    if (upscale) {
+      switch (PicoIn.filter) {
+      case 3: h_upscale_bl4_4_5(pd, 320, q, 256, len, f_nop); break;
+      case 2: h_upscale_bl2_4_5(pd, 320, q, 256, len, f_nop); break;
+      case 1: h_upscale_snn_4_5(pd, 320, q, 256, len, f_nop); break;
+      default: h_upscale_nn_4_5(pd, 320, q, 256, len, f_nop); break;
+      }
+    }
+    return 1;
+  }
+  return 0;
+}
+
+// --------------------------------------------
+
+static void PicoDoHighPal555_8bit(int sh, int line, struct PicoEState *est)
 {
   unsigned int *spal, *dpal;
   unsigned int cnt = (sh ? 1 : est->SonicPalCount+1);
@@ -1695,6 +1747,9 @@ void FinalizeLine555(int sh, int line, struct PicoEState *est)
   else if ((PicoIn.AHW & PAHW_SMS) && (est->Pico->video.reg[0] & 0x20))
     len -= 8, ps += 8;
 
+  if (BgcDMA(pd, len, est))
+    return;
+
   if ((est->rendstatus & PDRAW_SOFTSCALE) && len < 320) {
     if (len >= 240 && len <= 256) {
       pd += (256-len)>>1;
@@ -2111,6 +2166,35 @@ void PicoDrawRefreshSprites(void)
   }
 }
 
+void PicoDrawBgcDMA(u16 *base, u32 source, u32 mask, int dlen, int sl)
+{
+  struct PicoEState *est = &Pico.est;
+  int len = (est->Pico->video.reg[12]&1) ? 320 : 256;
+  int xl = (est->Pico->video.reg[12]&1) ? 38 : 33; // DMA slots during HSYNC
+
+  BgcDMAbase = base;
+  BgcDMAsrc = source;
+  BgcDMAmask = mask;
+  BgcDMAlen = dlen;
+  BgcDMAoffs = 0;
+
+  // handle slot offset in 1st line
+  if (sl-12 > 0) // active display output only starts at slot 12
+    BgcDMAoffs = 2*(sl-12);
+  else if (sl < 0) { // DMA starts before active display
+    BgcDMAsrc += 2*-sl;
+    BgcDMAlen -= 2*-sl;
+  }
+
+  // skip 1st line if it had been drawn already
+  if (Pico.est.DrawScanline > Pico.m.scanline) {
+    len -= BgcDMAoffs;
+    BgcDMAsrc += (len>>1)+xl;
+    BgcDMAlen -= (len>>1)+xl;
+    BgcDMAoffs = 0;
+  }
+}
+
 // also works for fast renderer
 void PicoDrawUpdateHighPal(void)
 {
diff --git a/pico/pico.c b/pico/pico.c
index ced4091e..ad57b6ec 100644
--- a/pico/pico.c
+++ b/pico/pico.c
@@ -190,6 +190,7 @@ int PicoReset(void)
 
   // create an empty "dma" to cause 68k exec start at random frame location
   Pico.t.m68c_line_start = Pico.t.m68c_aim;
+  PicoDrawBgcDMA(NULL, 0, 0, 0, 0);
   PicoVideoFIFOWrite(rand() & 0x1fff, 0, 0, PVS_CPURD);
 
   SekFinishIdleDet();
diff --git a/pico/pico_int.h b/pico/pico_int.h
index e74a4ac2..0d336822 100644
--- a/pico/pico_int.h
+++ b/pico/pico_int.h
@@ -704,6 +704,7 @@ int CM_compareRun(int cyc, int is_sub);
 void PicoDrawInit(void);
 PICO_INTERNAL void PicoFrameStart(void);
 void PicoDrawRefreshSprites(void);
+void PicoDrawBgcDMA(u16 *base, u32 source, u32 mask, int len, int sl);
 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);
diff --git a/pico/videoport.c b/pico/videoport.c
index bd744a3b..fe8e5ebd 100644
--- a/pico/videoport.c
+++ b/pico/videoport.c
@@ -522,6 +522,7 @@ static void DmaSlow(int len, u32 source)
   u32 a = Pico.video.addr | (Pico.video.addr_u << 16), e;
   u16 *r, *base = NULL;
   u32 mask = 0x1ffff;
+  int lc = SekCyclesDone()-Pico.t.m68c_line_start;
 
   elprintf(EL_VDPDMA, "DmaSlow[%i] %06x->%04x len %i inc=%i blank %i [%u] @ %06x",
     Pico.video.type, source, a, len, inc, (Pico.video.status&SR_VB)||!(Pico.video.reg[1]&0x40),
@@ -603,6 +604,15 @@ static void DmaSlow(int len, u32 source)
     case 3: // cram
       Pico.m.dirtyPal = 1;
       r = PicoMem.cram;
+      if (inc == 0 && (Pico.video.reg[7] & 0x3f) == ((a/2) & 0x3f)) { // bg color DMA
+        PicoVideoSync(1);
+        int sl = VdpFIFO.fifo_hcounts[lc/clkdiv];
+        if (sl > VdpFIFO.fifo_hcounts[0]-5) // hint delay is 5 slots
+          sl = (s8)sl;
+        // TODO this is needed to cover timing inaccuracies
+        if (sl <= 12)  sl = -2;
+        PicoDrawBgcDMA(base, source, mask, len, sl);
+      }
       for (; len; len--)
       {
         r[(a / 2) & 0x3f] = base[source++ & mask] & 0xeee;
-- 
2.39.5