From 352479001c82fc5c9abbcc1501c6e78ddc691a14 Mon Sep 17 00:00:00 2001
From: kub <derkub@gmail.com>
Date: Tue, 4 Aug 2020 22:24:56 +0200
Subject: [PATCH] vdp, optimisation for 8bit renderers

---
 pico/draw.c          |  9 ++++++++-
 pico/mode4.c         | 18 ++++++++----------
 platform/gp2x/emu.c  |  2 +-
 platform/linux/emu.c |  2 +-
 4 files changed, 18 insertions(+), 13 deletions(-)

diff --git a/pico/draw.c b/pico/draw.c
index 43bbea76..ed818546 100644
--- a/pico/draw.c
+++ b/pico/draw.c
@@ -1582,7 +1582,10 @@ static void FinalizeLine8bit(int sh, int line, struct PicoEState *est)
     len = 256;
   }
 
-  if (!sh && (est->rendstatus & PDRAW_SONIC_MODE)) {
+  if (DrawLineDestBase == HighColBase) {
+    if (!sh && (est->rendstatus & PDRAW_SONIC_MODE))
+      blockcpy_or(pd+8, est->HighCol+8, len, est->SonicPalCount*0x40);
+  } else if (!sh && (est->rendstatus & PDRAW_SONIC_MODE)) {
     // select active backup palette
     blockcpy_or(pd, est->HighCol+8, len, est->SonicPalCount*0x40);
   } else {
@@ -1892,6 +1895,10 @@ void PicoDrawSetOutFormat(pdso_t which, int use_32x_line_mode)
 
 void PicoDrawSetOutBufMD(void *dest, int increment)
 {
+  if (FinalizeLine == FinalizeLine8bit && increment == 328) {
+    // kludge for no-copy mode
+    PicoDrawSetInternalBuf(dest, increment);
+  }
   if (dest != NULL) {
     DrawLineDestBase = dest;
     DrawLineDestIncrement = increment;
diff --git a/pico/mode4.c b/pico/mode4.c
index cb1e8e0c..fa4407b5 100644
--- a/pico/mode4.c
+++ b/pico/mode4.c
@@ -17,7 +17,7 @@
 
 static void (*FinalizeLineM4)(int line);
 static int skip_next_line;
-static int screen_offset;
+static int screen_offset, line_offset;
 
 static void TileBGM4(int sx, int pal)
 {
@@ -110,8 +110,7 @@ static void draw_sprites(int scanline)
 
   if (pv->reg[0] & 8)
     xoff = 0;
-  if (!FinalizeLineM4 && !(PicoIn.opt & POPT_DIS_32C_BORDER))
-    xoff += 32;
+  xoff += line_offset;
 
   sat = (unsigned char *)PicoMem.vram + ((pv->reg[5] & 0x7e) << 7);
   if (pv->reg[1] & 2) {
@@ -252,8 +251,7 @@ static void DrawDisplayM4(int scanline)
   if (dx != 8)
     cells++; // have hscroll, need to draw 1 cell more
   dx += cellskip << 3;
-  if (!FinalizeLineM4 && !(PicoIn.opt & POPT_DIS_32C_BORDER))
-    dx += 32;
+  dx += line_offset;
 
   // low priority tiles
   if (!(pv->debug_p & PVD_KILL_B))
@@ -365,18 +363,18 @@ static void FinalizeLine8bitM4(int line)
 {
   unsigned char *pd = Pico.est.DrawLineDest;
 
-  if (!(PicoIn.opt & POPT_DIS_32C_BORDER))
-    pd += 32;
-
-  memcpy(pd, Pico.est.HighCol + 8, 256);
+  if (HighColBase != DrawLineDestBase)
+    memcpy(pd + line_offset, Pico.est.HighCol + line_offset + 8, 256);
 }
 
 void PicoDrawSetOutputMode4(pdso_t which)
 {
+  line_offset = PicoIn.opt & POPT_DIS_32C_BORDER ? 0 : 32;
   switch (which)
   {
     case PDF_8BIT:   FinalizeLineM4 = FinalizeLine8bitM4; break;
-    case PDF_RGB555: FinalizeLineM4 = FinalizeLineRGB555M4; break;
+    case PDF_RGB555: FinalizeLineM4 = FinalizeLineRGB555M4;
+                     line_offset = 0 /* done in FinalizeLine */; break;
     default:         FinalizeLineM4 = NULL;
                      PicoDrawSetInternalBuf(Pico.est.Draw2FB, 328); break;
   }
diff --git a/platform/gp2x/emu.c b/platform/gp2x/emu.c
index 1deb84da..42e34ee3 100644
--- a/platform/gp2x/emu.c
+++ b/platform/gp2x/emu.c
@@ -364,7 +364,7 @@ void pemu_finalize_frame(const char *fps, const char *notice)
 			localPalSize = make_local_pal(1);
 		// a hack for VR
 		if (PicoIn.AHW & PAHW_SVP)
-			memset32((int *)(Pico.est.Draw2FB+328*8+328*223), 0xe0e0e0e0, 328);
+			memset32((int *)(Pico.est.Draw2FB+328*8+328*223), 0xe0e0e0e0, 328/4);
 		// do actual copy
 		vidcpyM2(g_screen_ptr, Pico.est.Draw2FB+328*8,
 			!(Pico.video.reg[12] & 1), !(PicoIn.opt & POPT_DIS_32C_BORDER));
diff --git a/platform/linux/emu.c b/platform/linux/emu.c
index 4ef08a79..8c86471d 100644
--- a/platform/linux/emu.c
+++ b/platform/linux/emu.c
@@ -107,7 +107,7 @@ static void apply_renderer(void)
 	case RT_8BIT_ACC:
 		PicoIn.opt &= ~POPT_ALT_RENDERER;
 		PicoDrawSetOutFormat(PDF_8BIT, 0);
-		PicoDrawSetOutBuf(Pico.est.Draw2FB + 8, 328);
+		PicoDrawSetOutBuf(Pico.est.Draw2FB, 328);
 		break;
 	case RT_8BIT_FAST:
 		PicoIn.opt |=  POPT_ALT_RENDERER;
-- 
2.39.5