From: kub <derkub@gmail.com>
Date: Fri, 22 Mar 2019 22:03:26 +0000 (+0100)
Subject: reworked palette and buffer handling due to some 32X bugs
X-Git-Tag: v2.00~876
X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=b1a047c9267065bb85095761a3192ce2874d7383;p=picodrive.git

reworked palette and buffer handling due to some 32X bugs
---

diff --git a/pico/32x/draw.c b/pico/32x/draw.c
index 2287e246..4bdbc89a 100644
--- a/pico/32x/draw.c
+++ b/pico/32x/draw.c
@@ -11,6 +11,9 @@ int (*PicoScan32xBegin)(unsigned int num);
 int (*PicoScan32xEnd)(unsigned int num);
 int Pico32xDrawMode;
 
+void *DrawLineDestBase32x;
+int DrawLineDestIncrement32x;
+
 static void convert_pal555(int invert_prio)
 {
   unsigned int *ps = (void *)Pico32xMem->pal;
@@ -228,13 +231,11 @@ void PicoDraw32xLayer(int offs, int lines, int md_bg)
   int lines_sft_offs;
   int which_func;
 
-  Pico.est.DrawLineDest = (char *)DrawLineDestBase + offs * DrawLineDestIncrement;
+  Pico.est.DrawLineDest = (char *)DrawLineDestBase32x + offs * DrawLineDestIncrement32x;
   dram = Pico32xMem->dram[Pico32x.vdp_regs[0x0a/2] & P32XV_FS];
 
-  if (Pico32xDrawMode == PDM32X_BOTH) {
-    if (Pico.m.dirtyPal)
-      PicoDrawUpdateHighPal();
-  }
+  if (Pico32xDrawMode == PDM32X_BOTH)
+    PicoDrawUpdateHighPal();
 
   if ((Pico32x.vdp_regs[0] & P32XV_Mx) == 2)
   {
@@ -273,20 +274,21 @@ do_it:
 void PicoDraw32xLayerMdOnly(int offs, int lines)
 {
   int have_scan = PicoScan32xBegin != NULL && PicoScan32xEnd != NULL;
-  unsigned short *dst = (void *)((char *)DrawLineDestBase + offs * DrawLineDestIncrement);
+  unsigned short *dst = (void *)((char *)DrawLineDestBase32x + offs * DrawLineDestIncrement32x);
   unsigned char  *pmd = Pico.est.Draw2FB + 328 * offs + 8;
   unsigned short *pal = Pico.est.HighPal;
   int poffs = 0, plen = 320;
   int l, p;
 
   if (!(Pico.video.reg[12] & 1)) {
-    // 32col mode
+    // 32col mode. for some render modes MD pixel data carries an offset
+    if (!(PicoIn.opt & (POPT_ALT_RENDERER|POPT_DIS_32C_BORDER)))
+      pmd += 32;
     poffs = 32;
     plen = 256;
   }
 
-  if (Pico.m.dirtyPal)
-    PicoDrawUpdateHighPal();
+  PicoDrawUpdateHighPal();
 
   dst += poffs;
   for (l = 0; l < lines; l++) {
@@ -300,7 +302,7 @@ void PicoDraw32xLayerMdOnly(int offs, int lines)
       dst[p + 2] = pal[*pmd++];
       dst[p + 3] = pal[*pmd++];
     }
-    dst = (void *)((char *)dst + DrawLineDestIncrement);
+    dst = (void *)((char *)dst + DrawLineDestIncrement32x);
     pmd += 328 - plen;
     if (have_scan)
       PicoScan32xEnd(l + offs);
@@ -314,16 +316,32 @@ void PicoDrawSetOutFormat32x(pdso_t which, int use_32x_line_mode)
   Pico32xNativePal = Pico32xMem->pal_native;
 #endif
 
-  if (which == PDF_RGB555 && use_32x_line_mode) {
-    // we'll draw via FinalizeLine32xRGB555 (rare)
+  if (which == PDF_RGB555) {
+    // need CLUT pixels in PicoDraw2FB for layer transparency
+    PicoDrawSetInternalBuf(Pico.est.Draw2FB, 328);
+    PicoDrawSetOutBufMD(DrawLineDestBase32x, DrawLineDestIncrement32x);
+  } else {
+    // use the same layout as alt renderer
     PicoDrawSetInternalBuf(NULL, 0);
-    Pico32xDrawMode = PDM32X_OFF;
-    return;
+    PicoDrawSetOutBufMD(Pico.est.Draw2FB + 8, 328);
   }
 
-  // use the same layout as alt renderer
-  PicoDrawSetInternalBuf(Pico.est.Draw2FB, 328);
-  Pico32xDrawMode = (which == PDF_RGB555) ? PDM32X_32X_ONLY : PDM32X_BOTH;
+  if (use_32x_line_mode)
+    // we'll draw via FinalizeLine32xRGB555 (rare)
+    Pico32xDrawMode = PDM32X_OFF;
+  else
+    // in RGB555 mode the 32x layer is drawn over the MD layer, in the other
+    // modes 32x and MD layer are merged together by the 32x renderer
+    Pico32xDrawMode = (which == PDF_RGB555) ? PDM32X_32X_ONLY : PDM32X_BOTH;
+}
+
+void PicoDrawSetOutBuf32X(void *dest, int increment)
+{
+  DrawLineDestBase32x = dest;
+  DrawLineDestIncrement32x = increment;
+  // in RGB555 mode this buffer is also used by the MD renderer
+  if (Pico32xDrawMode != PDM32X_BOTH)
+    PicoDrawSetOutBufMD(DrawLineDestBase32x, DrawLineDestIncrement32x);
 }
 
 // vim:shiftwidth=2:ts=2:expandtab
diff --git a/pico/draw.c b/pico/draw.c
index 680de3da..e345a28d 100644
--- a/pico/draw.c
+++ b/pico/draw.c
@@ -1239,6 +1239,49 @@ void BackFill(int reg7, int sh, struct PicoEState *est)
 
 // --------------------------------------------
 
+void PicoDoHighPal555_8bit(int sh, int line, struct PicoEState *est)
+{
+  unsigned int *spal, *dpal;
+  unsigned int cnt = (sh ? 1 : est->SonicPalCount+1);
+  unsigned int t, i;
+
+  // reset dirty only if there are no outstanding changes
+  if (Pico.m.dirtyPal == 2)
+    Pico.m.dirtyPal = 0;
+
+  // In Sonic render mode palettes were backuped in SonicPal
+  spal = (void *)est->SonicPal;
+  dpal = (void *)est->HighPal;
+
+  // additional palettes stored after in-frame changes
+  for (i = 0; i < cnt * 0x40 / 2; i++) {
+    t = spal[i];
+#ifdef USE_BGR555
+    t = ((t & 0x000e000e)<< 1) | ((t & 0x00e000e0)<<3) | ((t & 0x0e000e00)<<4);
+#else
+    t = ((t & 0x000e000e)<<12) | ((t & 0x00e000e0)<<3) | ((t & 0x0e000e00)>>7);
+#endif
+    // treat it like it was 4-bit per channel, since in s/h mode it somewhat is that.
+    // otherwise intensity difference between this and s/h will be wrong
+    t |= (t >> 4) & 0x08610861; // 0x18e318e3
+    dpal[i] = t;
+  }
+
+  // norm: xxx0, sh: 0xxx, hi: 0xxx + 7
+  if (sh)
+  {
+    // shadowed pixels
+    for (i = 0; i < 0x40 / 2; i++)
+      dpal[0x40/2 | i] = dpal[0xc0/2 | i] = (dpal[i] >> 1) & 0x738e738e;
+    // hilighted pixels
+    for (i = 0; i < 0x40 / 2; i++) {
+      t = ((dpal[i] >> 1) & 0x738e738e) + 0x738e738e; // 0x7bef7bef;
+      t |= (t >> 4) & 0x08610861;
+      dpal[0x80/2 | i] = t;
+    }
+  }
+}
+
 #ifndef _ASM_DRAW_C
 void PicoDoHighPal555(int sh, int line, struct PicoEState *est)
 {
@@ -1285,8 +1328,7 @@ void FinalizeLine555(int sh, int line, struct PicoEState *est)
   unsigned short *pal=est->HighPal;
   int len;
 
-  if (Pico.m.dirtyPal)
-    PicoDoHighPal555(sh, line, est);
+  PicoDrawUpdateHighPal();
 
   if (Pico.video.reg[12]&1) {
     len = 320;
@@ -1315,22 +1357,21 @@ void FinalizeLine555(int sh, int line, struct PicoEState *est)
 static void FinalizeLine8bit(int sh, int line, struct PicoEState *est)
 {
   unsigned char *pd = est->DrawLineDest;
-  int len, rs = est->rendstatus;
-  static int dirty_count;
+  int len;
+  static int dirty_line;
 
-  if (!sh && Pico.m.dirtyPal == 1)
+  if (Pico.m.dirtyPal == 1)
   {
     // a hack for mid-frame palette changes
-    if (!(rs & PDRAW_SONIC_MODE))
-         dirty_count = 1;
-    else dirty_count++;
-    rs |= PDRAW_SONIC_MODE;
-    est->rendstatus = rs;
-    if (dirty_count == 3) {
-      blockcpy(est->HighPal, PicoMem.cram, 0x40*2);
-    } else if (dirty_count == 11) {
-      blockcpy(est->HighPal+0x40, PicoMem.cram, 0x40*2);
+    if (!(est->rendstatus & PDRAW_SONIC_MODE) || line - dirty_line > 4) {
+      // store a maximum of 3 additional palettes in SonicPal
+      if (est->SonicPalCount < 3)
+        est->SonicPalCount ++;
+      dirty_line = line;
+      est->rendstatus |= PDRAW_SONIC_MODE;
     }
+    blockcpy(est->SonicPal+est->SonicPalCount*0x40, PicoMem.cram, 0x40*2);
+    Pico.m.dirtyPal = 2;
   }
 
   if (Pico.video.reg[12]&1) {
@@ -1341,12 +1382,9 @@ static void FinalizeLine8bit(int sh, int line, struct PicoEState *est)
     len = 256;
   }
 
-  if (!sh && (rs & PDRAW_SONIC_MODE)) {
-    if (dirty_count >= 11) {
-      blockcpy_or(pd, est->HighCol+8, len, 0x80);
-    } else {
-      blockcpy_or(pd, est->HighCol+8, len, 0x40);
-    }
+  if (!sh && (est->rendstatus & PDRAW_SONIC_MODE)) {
+    // select active backup palette
+    blockcpy_or(pd, est->HighCol+8, len, est->SonicPalCount*0x40);
   } else {
     blockcpy(pd, est->HighCol+8, len);
   }
@@ -1478,6 +1516,7 @@ static int DrawDisplay(int sh)
 PICO_INTERNAL void PicoFrameStart(void)
 {
   int offs = 8, lines = 224;
+  int dirty = ((Pico.est.rendstatus & PDRAW_SONIC_MODE) || Pico.m.dirtyPal);
 
   // prepare to do this frame
   Pico.est.rendstatus = 0;
@@ -1503,11 +1542,16 @@ PICO_INTERNAL void PicoFrameStart(void)
   Pico.est.DrawScanline = 0;
   skip_next_line = 0;
 
+  if (FinalizeLine == FinalizeLine8bit) {
+    // make a backup of the current palette in case Sonic mode is detected later
+    Pico.est.SonicPalCount = 0;
+    Pico.m.dirtyPal = (dirty ? 2 : 0); // mark as dirty but already copied
+    blockcpy(Pico.est.SonicPal, PicoMem.cram, 0x40*2);
+  }
+
   if (PicoIn.opt & POPT_ALT_RENDERER)
     return;
 
-  if (Pico.m.dirtyPal)
-    Pico.m.dirtyPal = 2; // reset dirty if needed
   PrepareSprites(1);
 }
 
@@ -1598,15 +1642,21 @@ void PicoDrawSync(int to, int blank_last_line)
 void PicoDrawUpdateHighPal(void)
 {
   struct PicoEState *est = &Pico.est;
-  int sh = (Pico.video.reg[0xC] & 8) >> 3; // shadow/hilight?
-  if (PicoIn.opt & POPT_ALT_RENDERER)
-    sh = 0; // no s/h support
+  if (Pico.m.dirtyPal) {
+    int sh = (Pico.video.reg[0xC] & 8) >> 3; // shadow/hilight?
+    if ((PicoIn.opt & POPT_ALT_RENDERER) | (est->rendstatus & PDRAW_SONIC_MODE))
+      sh = 0; // no s/h support
+
+    if (FinalizeLine == FinalizeLine8bit)
+      PicoDoHighPal555_8bit(sh, 0, est);
+    else
+      PicoDoHighPal555(sh, 0, est);
 
-  PicoDoHighPal555(sh, 0, &Pico.est);
-  if (est->rendstatus & PDRAW_SONIC_MODE) {
-    // FIXME?
-    memcpy(est->HighPal + 0x40, est->HighPal, 0x40*2);
-    memcpy(est->HighPal + 0x80, est->HighPal, 0x40*2);
+    // cover for sprite priority bits if not in s/h or sonic mode
+    if (!sh && !(est->rendstatus & PDRAW_SONIC_MODE)) {
+      blockcpy(est->HighPal+0x40, est->HighPal, 0x40*2);
+      blockcpy(est->HighPal+0x80, est->HighPal, 0x80*2);
+    }
   }
 }
 
@@ -1629,17 +1679,33 @@ void PicoDrawSetOutFormat(pdso_t which, int use_32x_line_mode)
       FinalizeLine = NULL;
       break;
   }
-  PicoDrawSetOutFormat32x(which, use_32x_line_mode);
+  if (PicoIn.AHW & PAHW_32X)
+    PicoDrawSetOutFormat32x(which, use_32x_line_mode);
   PicoDrawSetOutputMode4(which);
   rendstatus_old = -1;
 }
 
+void PicoDrawSetOutBufMD(void *dest, int increment)
+{
+  if (dest != NULL) {
+    DrawLineDestBase = dest;
+    DrawLineDestIncrement = increment;
+    Pico.est.DrawLineDest = DrawLineDestBase + Pico.est.DrawScanline * increment;
+  }
+  else {
+    DrawLineDestBase = DefOutBuff;
+    DrawLineDestIncrement = 0;
+    Pico.est.DrawLineDest = DefOutBuff;
+  }
+}
+
 // note: may be called on the middle of frame
 void PicoDrawSetOutBuf(void *dest, int increment)
 {
-  DrawLineDestBase = dest;
-  DrawLineDestIncrement = increment;
-  Pico.est.DrawLineDest = (char *)DrawLineDestBase + Pico.est.DrawScanline * increment;
+  if (PicoIn.AHW & PAHW_32X)
+    PicoDrawSetOutBuf32X(dest, increment);
+  else
+    PicoDrawSetOutBufMD(dest, increment);
 }
 
 void PicoDrawSetInternalBuf(void *dest, int increment)
@@ -1652,6 +1718,7 @@ void PicoDrawSetInternalBuf(void *dest, int increment)
   else {
     HighColBase = DefHighCol;
     HighColIncrement = 0;
+    Pico.est.HighCol = DefHighCol;
   }
 }
 
diff --git a/pico/draw_arm.S b/pico/draw_arm.S
index 29af1c13..3bc27033 100644
--- a/pico/draw_arm.S
+++ b/pico/draw_arm.S
@@ -1498,11 +1498,9 @@ vidConvCpyRGB565: @ void *to, void *from, int pixels
 PicoDoHighPal555:
     stmfd   sp!, {r4-r10,lr}
     mov     r10,r2               @ est
-    mov     r1, #0
     ldr     r8, [r10, #OFS_EST_Pico]
 
-PicoDoHighPal555_nopush:
-    orr     r9, r1, r0, lsl #31  @ 0:called from FinalizeLine555, 31: s/h
+    mov     r9, r0
 
     add     r0, r10, #OFS_EST_HighPal
 
@@ -1517,7 +1515,7 @@ PicoDoHighPal555_nopush:
 
     vidConvCpyRGB565_local
 
-    tst     r9, #(1<<31)
+    cmp     r9, #0
     beq     PicoDoHighPal555_end
 
     add     r3, r10, #OFS_EST_HighPal
@@ -1560,11 +1558,7 @@ PicoDoHighPal555_nopush:
     mov     r0, #1
 
 PicoDoHighPal555_end:
-    tst     r9, #1
-    ldmeqfd sp!, {r4-r10,pc}
-
-    ldr     r8, [r10, #OFS_EST_Pico]
-    b       FinalizeLineRGB555_pal_done
+    ldmfd   sp!, {r4-r10,pc}
 
 
 @ void FinalizeLine555(int sh, int line, struct PicoEState *est)
@@ -1576,19 +1570,11 @@ FinalizeLine555:
     mov     r10,r2               @ est
     ldr     r8, [r10, #OFS_EST_Pico]
 
-    ldrb    r2, [r8, #OFS_Pico_m_dirtyPal]
-    mov     r1, #1
-    tst     r2, r2
-    bne     PicoDoHighPal555_nopush
+    bl      PicoDrawUpdateHighPal
 
-FinalizeLineRGB555_pal_done:
     add     r3, r10, #OFS_EST_HighPal
 
-    ldr     r12, [r10, #OFS_EST_rendstatus]
-    eors    r0, r0, #1           @ sh is 0
     mov     lr, #0xff
-    tstne   r12,#PDRAW_ACC_SPRITES
-    movne   lr, #0x3f
 
     ldr     r1, [r10, #OFS_EST_HighCol]
     ldr     r0, [r10, #OFS_EST_DrawLineDest]
diff --git a/pico/pico_int.h b/pico/pico_int.h
index cca7f954..f6d8b37f 100644
--- a/pico/pico_int.h
+++ b/pico/pico_int.h
@@ -356,6 +356,8 @@ struct PicoEState
   unsigned int  *PicoOpt;
   unsigned char *Draw2FB;
   unsigned short HighPal[0x100];
+  unsigned short SonicPal[0x100];
+  int SonicPalCount;
 };
 
 struct PicoMem
@@ -923,6 +925,7 @@ void p32x_sh2_poll_event(SH2 *sh2, unsigned int flags, unsigned int m68k_cycles)
 
 // 32x/draw.c
 void PicoDrawSetOutFormat32x(pdso_t which, int use_32x_line_mode);
+void PicoDrawSetOutBuf32X(void *dest, int increment);
 void FinalizeLine32xRGB555(int sh, int line, struct PicoEState *est);
 void PicoDraw32xLayer(int offs, int lines, int mdbg);
 void PicoDraw32xLayerMdOnly(int offs, int lines);
diff --git a/pico/videoport.c b/pico/videoport.c
index cd76dc04..d18c2cf9 100644
--- a/pico/videoport.c
+++ b/pico/videoport.c
@@ -41,7 +41,7 @@ static void VideoWrite(u16 d)
             if (a - ((unsigned)(Pico.video.reg[5]&0x7f) << 9) < 0x400)
               Pico.est.rendstatus |= PDRAW_DIRTY_SPRITES;
             break;
-    case 3: Pico.m.dirtyPal = 1;
+    case 3: if (PicoMem.cram [(a >> 1) & 0x3f] != d) Pico.m.dirtyPal = 1;
             PicoMem.cram [(a >> 1) & 0x3f] = d; break;
     case 5: PicoMem.vsram[(a >> 1) & 0x3f] = d; break;
     case 0x81:
@@ -441,7 +441,7 @@ PICO_INTERNAL_ASM void PicoVideoWrite(unsigned int a,unsigned short d)
             break;
           case 0x0c:
             // renderers should update their palettes if sh/hi mode is changed
-            if ((d^dold)&8) Pico.m.dirtyPal = 2;
+            if ((d^dold)&8) Pico.m.dirtyPal = 1;
             break;
         }
         return;
diff --git a/platform/common/common.mak b/platform/common/common.mak
index 197e8677..2f676abc 100644
--- a/platform/common/common.mak
+++ b/platform/common/common.mak
@@ -70,7 +70,7 @@ SRCS_COMMON += $(R)pico/cd/memory_arm.S
 endif
 ifeq "$(asm_32xdraw)" "1"
 DEFINES += _ASM_32X_DRAW
-SRCS_COMMON += $(R)pico/32x/draw_arm.s
+SRCS_COMMON += $(R)pico/32x/draw_arm.S
 endif
 ifeq "$(asm_mix)" "1"
 SRCS_COMMON += $(R)pico/sound/mix_arm.S
diff --git a/platform/common/emu.c b/platform/common/emu.c
index 0a9f0890..da03bff1 100644
--- a/platform/common/emu.c
+++ b/platform/common/emu.c
@@ -1407,8 +1407,10 @@ void emu_loop(void)
 			{
 				notice_msg_time = 0;
 				plat_status_msg_clear();
+#ifndef __GP2X__
 				plat_video_flip();
 				plat_status_msg_clear(); /* Do it again in case of double buffering */
+#endif
 				notice_msg = NULL;
 			}
 			else {
diff --git a/platform/gizmondo/emu.c b/platform/gizmondo/emu.c
index 86c473c2..fcf27125 100644
--- a/platform/gizmondo/emu.c
+++ b/platform/gizmondo/emu.c
@@ -155,7 +155,7 @@ static void blit(const char *fps, const char *notice)
 		}
 		// a hack for VR
 		if (PicoIn.AHW & PAHW_SVP)
-			memset32((int *)(Pico.est.Draw2FB+328*8+328*223), 0xe0e0e0e0, 328);
+			memset((int *)(Pico.est.Draw2FB+328*8+328*223), 0xe0e0e0e0, 328*4);
 		if (!(Pico.video.reg[12]&1)) lines_flags|=0x10000;
 		if (currentConfig.EmuOpt&0x4000)
 			lines_flags|=0x40000; // (Pico.m.frame_count&1)?0x20000:0x40000;
@@ -166,22 +166,25 @@ static void blit(const char *fps, const char *notice)
 		int lines_flags;
 		// 8bit accurate renderer
 		if (Pico.m.dirtyPal) {
-			Pico.m.dirtyPal = 0;
-			vidConvCpyRGB565(localPal, Pico.cram, 0x40);
+			if (Pico.m.dirtyPal == 2)
+				Pico.m.dirtyPal = 0;
+			/* no support
+			switch (Pico.est.SonicPalCount) {
+			case 3: vidConvCpyRGB565(localPal+0xc0, Pico.est.SonicPal+0xc0, 0x40);
+			case 2: vidConvCpyRGB565(localPal+0x80, Pico.est.SonicPal+0x80, 0x40);
+			case 1: vidConvCpyRGB565(localPal+0x40, Pico.est.SonicPal+0x40, 0x40);
+			default://vidConvCpyRGB565(localPal, Pico.est.SonicPal, 0x40);
+			} */
+			vidConvCpyRGB565(localPal, Pico.est.SonicPal, 0x40);
 			if (Pico.video.reg[0xC]&8) { // shadow/hilight mode
-				//vidConvCpyRGB32sh(localPal+0x40, Pico.cram, 0x40);
-				//vidConvCpyRGB32hi(localPal+0x80, Pico.cram, 0x40); // TODO?
-				memcpy32((void *)(localPal+0xc0), (void *)(localPal+0x40), 0x40*2/4);
+				//vidConvCpyRGB32sh(localPal+0x40, Pico.est.SonicPal, 0x40);
+				//vidConvCpyRGB32hi(localPal+0x80, Pico.est.SonicPal, 0x40); // TODO?
+				memcpy((void *)(localPal+0xc0), (void *)(localPal+0x40), 0x40*2);
 				localPal[0xc0] = 0x0600;
 				localPal[0xd0] = 0xc000;
 				localPal[0xe0] = 0x0000; // reserved pixels for OSD
 				localPal[0xf0] = 0xffff;
 			}
-			/* no support
-			else if (rendstatus & 0x20) { // mid-frame palette changes
-				vidConvCpyRGB565(localPal+0x40, HighPal, 0x40);
-				vidConvCpyRGB565(localPal+0x80, HighPal+0x40, 0x40);
-			} */
 		}
 		lines_flags = (Pico.video.reg[1]&8) ? 240 : 224;
 		if (!(Pico.video.reg[12]&1)) lines_flags|=0x10000;
diff --git a/platform/gizmondo/menu.c b/platform/gizmondo/menu.c
index 51f032f0..1045f47b 100644
--- a/platform/gizmondo/menu.c
+++ b/platform/gizmondo/menu.c
@@ -54,7 +54,7 @@ static unsigned int inp_prev = 0;
 void menu_draw_begin(int use_bgbuff)
 {
 	if (use_bgbuff)
-		memcpy32((int *)menu_screen, (int *)bg_buffer, 321*240*2/4);
+		memcpy((int *)menu_screen, (int *)bg_buffer, 321*240*2);
 }
 
 
@@ -66,7 +66,7 @@ void menu_draw_end(void)
 		lprintf("%s: Framework2D_LockBuffer() returned NULL\n", __FUNCTION__);
 		return;
 	}
-	memcpy32(giz_screen, (int *)menu_screen, 321*240*2/4);
+	memcpy(giz_screen, (int *)menu_screen, 321*240*2);
 	fb_unlock();
 	giz_screen = NULL;
 	fb_flip();
diff --git a/platform/gp2x/emu.c b/platform/gp2x/emu.c
index 18d8a57e..450ac080 100644
--- a/platform/gp2x/emu.c
+++ b/platform/gp2x/emu.c
@@ -291,32 +291,45 @@ static int EmuScanEnd16_ld(unsigned int num)
 }
 
 static int localPal[0x100];
+static int localPalSize;
+
 static void (*vidcpyM2)(void *dest, void *src, int m32col, int with_32c_border);
 static int (*make_local_pal)(int fast_mode);
 
 static int make_local_pal_md(int fast_mode)
 {
-	int pallen = 0xc0;
-
-	bgr444_to_rgb32(localPal, PicoMem.cram);
-	if (fast_mode)
-		return 0x40;
+	int pallen = 0x100;
 
-	if (Pico.video.reg[0xC] & 8) { // shadow/hilight mode
-		bgr444_to_rgb32_sh(localPal, PicoMem.cram);
-		localPal[0xc0] = 0x0000c000;
-		localPal[0xd0] = 0x00c00000;
-		localPal[0xe0] = 0x00000000; // reserved pixels for OSD
-		localPal[0xf0] = 0x00ffffff;
-		pallen = 0x100;
+	if (fast_mode) {
+		bgr444_to_rgb32(localPal, PicoMem.cram);
+		pallen = 0x40;
+		Pico.m.dirtyPal = 0;
 	}
 	else if (Pico.est.rendstatus & PDRAW_SONIC_MODE) { // mid-frame palette changes
-		bgr444_to_rgb32(localPal+0x40, Pico.est.HighPal);
-		bgr444_to_rgb32(localPal+0x80, Pico.est.HighPal+0x40);
+		switch (Pico.est.SonicPalCount) {
+		case 3: bgr444_to_rgb32(localPal+0xc0, Pico.est.SonicPal+0xc0);
+		case 2: bgr444_to_rgb32(localPal+0x80, Pico.est.SonicPal+0x80);
+		case 1: bgr444_to_rgb32(localPal+0x40, Pico.est.SonicPal+0x40);
+		default:bgr444_to_rgb32(localPal, Pico.est.SonicPal);
+		}
+		pallen = (Pico.est.SonicPalCount+1)*0x40;
 	}
-	else
-		memcpy(localPal + 0x80, localPal, 0x40 * 4); // for spr prio mess
+	else if (Pico.video.reg[0xC] & 8) { // shadow/hilight mode
+		bgr444_to_rgb32(localPal, Pico.est.SonicPal);
+		bgr444_to_rgb32_sh(localPal, Pico.est.SonicPal);
+	}
+	else {
+		bgr444_to_rgb32(localPal, Pico.est.SonicPal);
+		memcpy(localPal+0x40, localPal, 0x40*4); // for spr prio mess
+		memcpy(localPal+0x80, localPal, 0x80*4); // for spr prio mess
+	}
+	localPal[0xc0] = 0x0000c000;
+	localPal[0xd0] = 0x00c00000;
+	localPal[0xe0] = 0x00000000; // reserved pixels for OSD
+	localPal[0xf0] = 0x00ffffff;
 
+        if (Pico.m.dirtyPal == 2)
+		Pico.m.dirtyPal = 0;
 	return pallen;
 }
 
@@ -334,25 +347,21 @@ static int make_local_pal_sms(int fast_mode)
 		*dpal++ = t;
 	}
 
+	Pico.m.dirtyPal = 0;
 	return 0x40;
 }
 
 void pemu_finalize_frame(const char *fps, const char *notice)
 {
 	int emu_opt = currentConfig.EmuOpt;
-	int ret;
 
 	if (PicoIn.AHW & PAHW_32X)
-		; // nothing to do
+		localPalSize = 0; // nothing to do
 	else if (get_renderer() == RT_8BIT_FAST)
 	{
 		// 8bit fast renderer
-		if (Pico.m.dirtyPal) {
-			Pico.m.dirtyPal = 0;
-			ret = make_local_pal(1);
-			// feed new palette to our device
-			gp2x_video_setpalette(localPal, ret);
-		}
+		if (Pico.m.dirtyPal)
+			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);
@@ -364,12 +373,9 @@ void pemu_finalize_frame(const char *fps, const char *notice)
 	{
 		// 8bit accurate renderer
 		if (Pico.m.dirtyPal)
-		{
-			Pico.m.dirtyPal = 0;
-			ret = make_local_pal(0);
-			gp2x_video_setpalette(localPal, ret);
-		}
+			localPalSize = make_local_pal(0);
 	}
+	else	localPalSize = 0; // no palette in 16bit mode
 
 	if (notice)
 		osd_text(4, osd_y, notice);
@@ -385,6 +391,10 @@ void plat_video_flip(void)
 {
 	int stride = g_screen_width;
 	gp2x_video_flip();
+	// switching the palette takes immediate effect, whilst flipping only
+	// takes effect with the next vsync; unavoidable flicker may occur!
+	if (localPalSize)
+		gp2x_video_setpalette(localPal, localPalSize);
 
 	if (is_16bit_mode())
 		stride *= 2;
@@ -502,9 +512,6 @@ static void vid_reset_mode(void)
 		if (renderer == RT_16BIT && (currentConfig.EmuOpt & EOPT_WIZ_TEAR_FIX)) {
 			PicoDrawSetOutFormat(PDF_RGB555, 1);
 		}
-		else {
-			PicoDrawSetOutFormat(PDF_NONE, 0);
-		}
 		PicoDrawSetOutBuf(g_screen_ptr, g_screen_width * 2);
 		gp2x_mode = 16;
 	}
@@ -537,10 +544,7 @@ static void vid_reset_mode(void)
 		localPal[0xe0] = 0x00000000; // reserved pixels for OSD
 		localPal[0xf0] = 0x00ffffff;
 		gp2x_video_setpalette(localPal, 0x100);
-		gp2x_memset_all_buffers(0, 0xe0, 320*240);
 	}
-	else
-		gp2x_memset_all_buffers(0, 0, 320*240*2);
 
 	if (currentConfig.EmuOpt & EOPT_WIZ_TEAR_FIX)
 		gp2x_mode = -gp2x_mode;
@@ -723,6 +727,8 @@ void pemu_forced_frame(int no_scale, int do_emu)
 	PicoDrawSetCallbacks(NULL, NULL);
 	Pico.m.dirtyPal = 1;
 
+	if (!no_scale)
+		no_scale = currentConfig.scaling == EOPT_SCALE_NONE;
 	emu_cmn_forced_frame(no_scale, do_emu);
 
 	g_menubg_src_ptr = g_screen_ptr;
diff --git a/platform/linux/emu.c b/platform/linux/emu.c
index 5d4432fa..8af5afa8 100644
--- a/platform/linux/emu.c
+++ b/platform/linux/emu.c
@@ -71,8 +71,8 @@ void pemu_finalize_frame(const char *fps, const char *notice)
 		unsigned char *ps = Pico.est.Draw2FB + 328*8 + 8;
 		unsigned short *pal = Pico.est.HighPal;
 		int i, x;
-		if (Pico.m.dirtyPal)
-			PicoDrawUpdateHighPal();
+
+		PicoDrawUpdateHighPal();
 		for (i = 0; i < 224; i++, ps += 8)
 			for (x = 0; x < 320; x++)
 				*pd++ = pal[*ps++];
@@ -109,6 +109,8 @@ static void apply_renderer(void)
 
 	if (PicoIn.AHW & PAHW_32X)
 		PicoDrawSetOutBuf(g_screen_ptr, g_screen_ppitch * 2);
+
+	Pico.m.dirtyPal = 1;
 }
 
 void plat_video_toggle_renderer(int change, int is_menu)
diff --git a/platform/psp/emu.c b/platform/psp/emu.c
index 5c0cb57f..917ecc8d 100644
--- a/platform/psp/emu.c
+++ b/platform/psp/emu.c
@@ -201,13 +201,22 @@ static void do_pal_update(int allow_sh, int allow_as)
 
 	//for (i = 0x3f/2; i >= 0; i--)
 	//	dpal[i] = ((spal[i]&0x000f000f)<< 1)|((spal[i]&0x00f000f0)<<3)|((spal[i]&0x0f000f00)<<4);
-	do_pal_convert(localPal, Pico.cram, currentConfig.gamma, currentConfig.gamma2);
-
-	Pico.m.dirtyPal = 0;
-	need_pal_upload = 1;
-
-	if (allow_sh && (Pico.video.reg[0xC]&8)) // shadow/hilight?
+	if ((currentConfig.EmuOpt&0x80) || (PicoOpt&0x10)) {
+		do_pal_convert(localPal, Pico.cram, currentConfig.gamma, currentConfig.gamma2);
+		Pico.m.dirtyPal = 0;
+	}
+	else if (Pico.est.rendstatus&0x20)
+	{
+		switch (Pico.est.SonicPalCount) {
+		case 3: do_pal_convert(localPal+0xc0, Pico.est.SonicPal+0xc0, currentConfig.gamma, currentConfig.gamma2);
+		case 2: do_pal_convert(localPal+0x80, Pico.est.SonicPal+0x80, currentConfig.gamma, currentConfig.gamma2);
+		case 1:	do_pal_convert(localPal+0x40, Pico.est.SonicPal+0x40, currentConfig.gamma, currentConfig.gamma2);
+		default:do_pal_convert(localPal, Pico.est.SonicPal, currentConfig.gamma, currentConfig.gamma2);
+		}
+	}
+	else if (allow_sh && (Pico.video.reg[0xC]&8)) // shadow/hilight?
 	{
+		do_pal_convert(localPal, Pico.est.SonicPal, currentConfig.gamma, currentConfig.gamma2);
 		// shadowed pixels
 		for (i = 0x3f/2; i >= 0; i--)
 			dpal[0x20|i] = dpal[0x60|i] = (dpal[i]>>1)&0x7bcf7bcf;
@@ -223,6 +232,16 @@ static void do_pal_update(int allow_sh, int allow_as)
 		localPal[0xe0] = 0;
 		localPal[0xf0] = 0x001f;
 	}
+ 	else if (allow_as && (Pico.est.rendstatus & PDRAW_SPR_LO_ON_HI))
+ 	{
+		do_pal_convert(localPal, Pico.est.SonicPal, currentConfig.gamma, currentConfig.gamma2);
+		memcpy((int *)dpal+0x40/2, (void *)localPal, 0x40*2);
+		memcpy((int *)dpal+0x80/2, (void *)localPal, 0x80*2);
+	}
+
+	if (Pico.m.dirtyPal == 2)
+		Pico.m.dirtyPal = 0;
+	need_pal_upload = 1;
 }
 
 static void do_slowmode_lines(int line_to)
@@ -639,7 +658,7 @@ static void writeSound(int len)
 
 	PicoIn.sndOut += len / 2;
 	/*if (PicoIn.sndOut > sndBuffer_endptr) {
-		memcpy32((int *)(void *)sndBuffer, (int *)endptr, (PicoIn.sndOut - endptr + 1) / 2);
+		memcpy((int *)(void *)sndBuffer, (int *)endptr, (PicoIn.sndOut - endptr + 1) * 2);
 		PicoIn.sndOut = &sndBuffer[PicoIn.sndOut - endptr];
 		lprintf("mov\n");
 	}
diff --git a/platform/psp/menu.c b/platform/psp/menu.c
index ab022f97..fc31b8e7 100644
--- a/platform/psp/menu.c
+++ b/platform/psp/menu.c
@@ -59,7 +59,7 @@ void menu_draw_begin(void)
 	// int i;
 
 	// for (i = 272; i >= 0; i--, dst += 512, src += 480)
-	//	memcpy32((int *)dst, (int *)src, 480*2/4);
+	//	memcpy((int *)dst, (int *)src, 480*2);
 
 	sceGuSync(0,0); // sync with prev
 	sceGuStart(GU_DIRECT, guCmdList);