diff --git a/src/include/86box/vid_svga_render_remap.h b/src/include/86box/vid_svga_render_remap.h
new file mode 100644
index 000000000..c2e48f1b2
--- /dev/null
+++ b/src/include/86box/vid_svga_render_remap.h
@@ -0,0 +1,126 @@
+/*Variables :
+        byte/word/doubleword mode
+        word has MA13/MA15->MA0
+        ET4000 treats doubleword as byte
+        row 0 -> MA13
+        row 1 -> MA14
+*/
+
+//S3 - enhanced mode mappings CR31.3 can force doubleword mode
+//Cirrus Logic handles SVGA writes seperately
+//S3, CL, TGUI blitters need checking
+
+//CL, S3, Mach64, ET4000, Banshee, TGUI all okay
+//Still to check - ViRGE, HT216
+#define VAR_BYTE_MODE      (0 << 0)
+#define VAR_WORD_MODE_MA13 (1 << 0)
+#define VAR_WORD_MODE_MA15 (2 << 0)
+#define VAR_DWORD_MODE     (3 << 0)
+#define VAR_MODE_MASK      (3 << 0)
+#define VAR_ROW0_MA13      (1 << 2)
+#define VAR_ROW1_MA14      (1 << 3)
+
+#define ADDRESS_REMAP_FUNC(nr) \
+        static uint32_t address_remap_func_ ## nr(svga_t *svga, uint32_t in_addr) \
+        {                                                                       \
+                uint32_t out_addr;                                              \
+                                                                                \
+                switch (nr & VAR_MODE_MASK)                                     \
+                {                                                               \
+                        case VAR_BYTE_MODE:                                     \
+                        out_addr = in_addr;                                     \
+                        break;                                                  \
+                                                                                \
+                        case VAR_WORD_MODE_MA13:                                \
+                        out_addr = ((in_addr << 1) & 0x1fff8) |                 \
+                                   ((in_addr >> 13) & 0x4) |                    \
+                                   (in_addr & ~0x1ffff);                        \
+                        break;                                                  \
+                                                                                \
+                        case VAR_WORD_MODE_MA15:                                \
+                        out_addr = ((in_addr << 1) & 0x1fff8) |                 \
+                                   ((in_addr >> 15) & 0x4) |                    \
+                                   (in_addr & ~0x1ffff);                        \
+                        break;                                                  \
+                                                                                \
+                        case VAR_DWORD_MODE:                                    \
+                        out_addr = ((in_addr << 2) & 0x3fff0) |                 \
+                                   ((in_addr >> 14) & 0xc) |                    \
+                                   (in_addr & ~0x3ffff);                        \
+                        break;                                                  \
+                }                                                               \
+																				\
+                if (nr & VAR_ROW0_MA13)                                         \
+                        out_addr = (out_addr & ~0x8000) |                \
+                                   ((svga->sc & 1) ? 0x8000 : 0);        \
+                if (nr & VAR_ROW1_MA14)                                         \
+                        out_addr = (out_addr & ~0x10000) |                \
+                                   ((svga->sc & 2) ? 0x10000 : 0);        \
+                                                                                \
+                return out_addr;                                                \
+        }
+
+ADDRESS_REMAP_FUNC(0)
+ADDRESS_REMAP_FUNC(1)
+ADDRESS_REMAP_FUNC(2)
+ADDRESS_REMAP_FUNC(3)
+ADDRESS_REMAP_FUNC(4)
+ADDRESS_REMAP_FUNC(5)
+ADDRESS_REMAP_FUNC(6)
+ADDRESS_REMAP_FUNC(7)
+ADDRESS_REMAP_FUNC(8)
+ADDRESS_REMAP_FUNC(9)
+ADDRESS_REMAP_FUNC(10)
+ADDRESS_REMAP_FUNC(11)
+ADDRESS_REMAP_FUNC(12)
+ADDRESS_REMAP_FUNC(13)
+ADDRESS_REMAP_FUNC(14)
+ADDRESS_REMAP_FUNC(15)
+
+static uint32_t (*address_remap_funcs[16])(svga_t *svga, uint32_t in_addr) =
+{
+        address_remap_func_0,
+        address_remap_func_1,
+        address_remap_func_2,
+        address_remap_func_3,
+        address_remap_func_4,
+        address_remap_func_5,
+        address_remap_func_6,
+        address_remap_func_7,
+        address_remap_func_8,
+        address_remap_func_9,
+        address_remap_func_10,
+        address_remap_func_11,
+        address_remap_func_12,
+        address_remap_func_13,
+        address_remap_func_14,
+        address_remap_func_15
+};
+
+void svga_recalc_remap_func(svga_t *svga)
+{
+        int func_nr;
+        
+        if (svga->fb_only)
+                func_nr = 0;
+        else {
+                if (svga->force_dword_mode)
+                        func_nr = VAR_DWORD_MODE;
+                else if (svga->crtc[0x14] & 0x40)
+                        func_nr = svga->packed_chain4 ? VAR_BYTE_MODE : VAR_DWORD_MODE;
+                else if (svga->crtc[0x17] & 0x40)
+                        func_nr = VAR_BYTE_MODE;
+                else if (svga->crtc[0x17] & 0x20)
+                        func_nr = VAR_WORD_MODE_MA15;
+                else
+                        func_nr = VAR_WORD_MODE_MA13;
+                        
+                if (!(svga->crtc[0x17] & 0x01))
+                        func_nr |= VAR_ROW0_MA13;
+                if (!(svga->crtc[0x17] & 0x02))
+                        func_nr |= VAR_ROW1_MA14;
+        }
+
+		svga->remap_required = (func_nr != 0);
+        svga->remap_func = address_remap_funcs[func_nr];
+}
diff --git a/src/video/vid_cl54xx.c b/src/video/vid_cl54xx.c
index 6a3331b1d..87952bb13 100644
--- a/src/video/vid_cl54xx.c
+++ b/src/video/vid_cl54xx.c
@@ -47,6 +47,7 @@
 #define BIOS_GD5422_PATH		"roms/video/cirruslogic/cl5422.bin"
 #define BIOS_GD5426_PATH		"roms/video/cirruslogic/Diamond SpeedStar PRO VLB v3.04.bin"
 #define BIOS_GD5428_ISA_PATH		"roms/video/cirruslogic/5428.bin"
+#define BIOS_GD5428_MCA_PATH		"roms/video/cirruslogic/SVGA141.ROM"
 #define BIOS_GD5428_PATH		"roms/video/cirruslogic/vlbusjapan.BIN"
 #define BIOS_GD5429_PATH		"roms/video/cirruslogic/5429.vbi"
 #define BIOS_GD5430_VLB_PATH		"roms/video/cirruslogic/diamondvlbus.bin"
@@ -150,7 +151,7 @@ typedef struct gd54xx_t
     mem_mapping_t 	aperture2_mapping;
     mem_mapping_t	vgablt_mapping;
 
-    svga_t		svga;
+    svga_t		svga, *mb_vga;
 
     int			has_bios, rev,
 			bit32;
@@ -214,7 +215,7 @@ typedef struct gd54xx_t
     int			card;
     
     uint8_t		pos_regs[8];
-    svga_t		*mb_vga;
+	int			vidsys_ena;
 
     uint32_t		lfb_base, vgablt_base;
 
@@ -691,7 +692,8 @@ gd54xx_out(uint16_t addr, uint8_t val, void *p)
 			}
 		}
 		svga->attrff ^= 1;
-                return;
+                return;				
+				
 	case 0x3c4:
 		svga->seqaddr = val;
 		break;
@@ -761,6 +763,8 @@ gd54xx_out(uint16_t addr, uint8_t val, void *p)
 						svga->hwcursor.addr = ((gd54xx->vram_size - 0x4000) + ((val & 0x3f) * 256));
 					break;
 				case 0x07:
+					svga->packed_chain4 = svga->seqregs[7] & 1;
+					svga_recalctimings(svga);
 					if (gd54xx_is_5422(svga))
 						gd543x_recalc_mapping(gd54xx);
 					else
@@ -878,7 +882,7 @@ gd54xx_out(uint16_t addr, uint8_t val, void *p)
 			}
 
 			svga->fast = (svga->gdcreg[8] == 0xff && !(svga->gdcreg[3] & 0x18) &&
-				     !svga->gdcreg[1]) && svga->chain4;
+				     !svga->gdcreg[1]) && ((svga->chain4 && svga->packed_chain4) || svga->fb_only);
 			if (((svga->gdcaddr == 5) && ((val ^ o) & 0x70)) ||
 			    ((svga->gdcaddr == 6) && ((val ^ o) & 1)))
 				svga_recalctimings(svga);
@@ -898,6 +902,8 @@ gd54xx_out(uint16_t addr, uint8_t val, void *p)
 						svga->adv_flags |= FLAG_EXT_WRITE;
 					if (svga->gdcreg[0xb] & 0x08)
 						svga->adv_flags |= FLAG_LATCH8;
+					if (svga->gdcreg[0xb] & 0x10)
+						svga->adv_flags |= FLAG_ADDR_BY16;
 					gd54xx_recalc_banking(gd54xx);
 					break;
 
@@ -1018,6 +1024,11 @@ gd54xx_out(uint16_t addr, uint8_t val, void *p)
 			}
 		}
 		return;
+		
+	case 0x46e8:
+		pclog("port 46e8 = %02x\n", val & 0x18);
+		break;
+		
 	case 0x3d4:
 		svga->crtcreg = val & gd54xx->crtcreg_mask;
 		return;
@@ -1595,8 +1606,11 @@ gd54xx_recalctimings(svga_t *svga)
     if (svga->seqregs[7] & CIRRUS_SR7_BPP_SVGA) {
 	if (linedbl)
 		svga->render = svga_render_8bpp_lowres;
-	else
+	else {
 		svga->render = svga_render_8bpp_highres;
+		if ((svga->dispend == 512) && !svga->interlace && gd54xx_is_5434(svga))
+			svga->hdisp <<= 1;
+	}
     } else if (svga->gdcreg[5] & 0x40)
 	svga->render = svga_render_8bpp_lowres;
 
@@ -1642,8 +1656,9 @@ gd54xx_recalctimings(svga_t *svga)
 						svga->render = svga_render_32bpp_lowres;
 					else
 						svga->render = svga_render_32bpp_highres;
-					if (svga->crtc[0x27] < CIRRUS_ID_CLGD5436)
+					if (svga->crtc[0x27] < CIRRUS_ID_CLGD5436) {
 						svga->rowoffset *= 2;
+					}
 				} else {
 					svga->bpp = 24;
 					if (linedbl)
@@ -1986,7 +2001,7 @@ gd54xx_write_modes45(svga_t *svga, uint8_t val, uint32_t addr)
 
     switch (svga->writemode) {
 	case 4:
-		if (svga->gdcreg[0xb] & 0x10) {
+		if (svga->adv_flags & FLAG_ADDR_BY16) {
 			addr <<= 2;
 			addr &= svga->decode_mask;
 
@@ -2008,7 +2023,7 @@ gd54xx_write_modes45(svga_t *svga, uint8_t val, uint32_t addr)
 		break;
 
 	case 5:
-		if (svga->gdcreg[0xb] & 0x10) {
+		if (svga->adv_flags & FLAG_ADDR_BY16) {
 			addr <<= 2;
 			addr &= svga->decode_mask;
 
@@ -3686,9 +3701,9 @@ cl_pci_write(int func, int addr, uint8_t val, void *p)
 static uint8_t 
 gd5428_mca_read(int port, void *p)
 {
-        gd54xx_t *gd54xx = (gd54xx_t *)p;
-
-        return gd54xx->pos_regs[port & 7];
+    gd54xx_t *gd54xx = (gd54xx_t *)p;
+	
+	return gd54xx->pos_regs[port & 7];
 }
 
 static void 
@@ -3698,17 +3713,15 @@ gd5428_mca_write(int port, uint8_t val, void *p)
 
         if (port < 0x102)
 		return;
-
-        gd54xx->pos_regs[port & 7] = val;
+	
+	gd54xx->pos_regs[port & 7] = val;
 	gd543x_recalc_mapping(gd54xx);
 }
 
 static uint8_t 
 gd5428_mca_feedb(void *p)
 {
-        gd54xx_t *gd54xx = (gd54xx_t *)p;
-
-        return gd54xx->pos_regs[2] & 1;
+    return 1;
 }
 
 static void
@@ -3761,6 +3774,8 @@ static void
 	case CIRRUS_ID_CLGD5428:
 		if (gd54xx->vlb)
 			romfn = BIOS_GD5428_PATH;
+		else if (gd54xx->mca)
+			romfn = BIOS_GD5428_MCA_PATH;
 		else
 			romfn = BIOS_GD5428_ISA_PATH;
 		break;
@@ -3953,7 +3968,8 @@ static void
     if (gd54xx->mca) {
 	gd54xx->pos_regs[0] = 0x7b;
 	gd54xx->pos_regs[1] = 0x91;
-	mca_add(gd5428_mca_read, gd5428_mca_write, gd5428_mca_feedb, NULL, gd54xx);
+	mca_add(gd5428_mca_read, gd5428_mca_write, gd5428_mca_feedb, NULL, gd54xx);		
+	io_sethandler(0x46e8, 0x0001, gd54xx_in, NULL, NULL, gd54xx_out, NULL, NULL, gd54xx);
     }
 
     if (gd54xx_is_5434(svga)) {
@@ -4013,6 +4029,12 @@ gd5428_isa_available(void)
     return rom_present(BIOS_GD5428_ISA_PATH);
 }
 
+static int
+gd5428_mca_available(void)
+{
+    return rom_present(BIOS_GD5428_MCA_PATH);
+}
+
 static int
 gd5429_available(void)
 {
@@ -4369,7 +4391,7 @@ const device_t gd5428_mca_device =
     gd54xx_init, 
     gd54xx_close, 
     NULL,
-    { gd5428_available },
+    { gd5428_mca_available },
     gd54xx_speed_changed,
     gd54xx_force_redraw,
     NULL
diff --git a/src/video/vid_et4000.c b/src/video/vid_et4000.c
index a4751d3ea..d22a6e6af 100644
--- a/src/video/vid_et4000.c
+++ b/src/video/vid_et4000.c
@@ -742,6 +742,8 @@ et4000_init(const device_t *info)
 
     dev->svga.translate_address = get_et4000_addr;
 
+	dev->svga.packed_chain4 = 1;
+
     return(dev);
 }
 
diff --git a/src/video/vid_et4000w32.c b/src/video/vid_et4000w32.c
index d2cc9b8e3..25e369cc3 100644
--- a/src/video/vid_et4000w32.c
+++ b/src/video/vid_et4000w32.c
@@ -46,16 +46,6 @@
 #define BIOS_ROM_PATH_W32P	"roms/video/et4000w32/ET4K_W32.BIN"
 #define BIOS_ROM_PATH_W32P_REVC	"roms/video/et4000w32/et4000w32pcardex.BIN"
 
-#define FIFO_SIZE		65536
-#define FIFO_MASK		(FIFO_SIZE - 1)
-#define FIFO_ENTRY_SIZE		(1 << 31)
-
-#define FIFO_ENTRIES		(et4000->fifo_write_idx - et4000->fifo_read_idx)
-#define FIFO_FULL		((et4000->fifo_write_idx - et4000->fifo_read_idx) >= (FIFO_SIZE-1))
-#define FIFO_EMPTY		(et4000->fifo_read_idx == et4000->fifo_write_idx)
-
-#define FIFO_TYPE		0xff000000
-#define FIFO_ADDR		0x00ffffff
 
 #define ACL_WRST		1
 #define ACL_RDST		2
@@ -73,20 +63,6 @@ enum
     ET4000W32P_DIAMOND
 };
 
-enum
-{
-    FIFO_INVALID    = (0x00 << 24),
-    FIFO_WRITE_BYTE = (0x01 << 24),
-    FIFO_WRITE_MMU  = (0x02 << 24)
-};
-
-
-typedef struct
-{
-    uint32_t addr_type;
-    uint32_t val;
-} fifo_entry_t;
-
 
 typedef struct et4000w32p_t
 {
@@ -138,15 +114,6 @@ typedef struct et4000w32p_t
 	uint32_t	base[3];
 	uint8_t		ctrl;
     } mmu;
-
-    fifo_entry_t	fifo[FIFO_SIZE];
-    volatile int	fifo_read_idx, fifo_write_idx;
-
-    thread_t *		fifo_thread;
-    event_t *		wake_fifo_thread;
-    event_t *		fifo_not_full_event;
-
-    uint64_t		blitter_time, status_time;
 } et4000w32p_t;
 
 
@@ -198,7 +165,7 @@ et4000w32p_out(uint16_t addr, uint8_t val, void *p)
     et4000w32p_t *et4000 = (et4000w32p_t *)p;
     svga_t *svga = &et4000->svga;
     uint8_t old;
-    uint32_t add2addr;
+    uint32_t add2addr = 0;
 
     if (((addr & 0xfff0) == 0x3d0 || (addr & 0xfff0) == 0x3b0) && !(svga->miscout & 1)) 
 	addr ^= 0x60;
@@ -287,18 +254,30 @@ et4000w32p_out(uint16_t addr, uint8_t val, void *p)
 	case 0x210b: case 0x211b: case 0x212b: case 0x213b:
 	case 0x214b: case 0x215b: case 0x216b: case 0x217b:
 		et4000->regs[et4000->index] = val;
-		svga->hwcursor.xsize = svga->hwcursor.ysize = (et4000->regs[0xEF] & 4) ? 128 : 64;
+		svga->hwcursor.xsize = svga->hwcursor.ysize = ((et4000->regs[0xEF] & 4) || (et4000->type == ET4000W32)) ? 128 : 64;
 		svga->hwcursor.x     = et4000->regs[0xE0] | ((et4000->regs[0xE1] & 7) << 8);
 		svga->hwcursor.y     = et4000->regs[0xE4] | ((et4000->regs[0xE5] & 7) << 8);
 		svga->hwcursor.ena   = !!(et4000->regs[0xF7] & 0x80);
+		svga->hwcursor.xoff  = et4000->regs[0xE2];
+		svga->hwcursor.yoff  = et4000->regs[0xE6];
+
+		if (et4000->type == ET4000W32) {
+			switch (svga->bpp) {
+				case 8:
+					svga->hwcursor.xoff += 32;
+					break;			
+			}
+		}
+
 		if (svga->hwcursor.xsize == 128) {
-			svga->hwcursor.xoff  = et4000->regs[0xE2] & 0x7f;
-			svga->hwcursor.yoff  = et4000->regs[0xE6] & 0x7f;
+			svga->hwcursor.xoff &= 0x7f;
+			svga->hwcursor.yoff &= 0x7f;
 		} else {
-			svga->hwcursor.xoff  = et4000->regs[0xE2] & 0x3f;
-			svga->hwcursor.yoff  = et4000->regs[0xE6] & 0x3f;
+			svga->hwcursor.xoff &= 0x3f;
+			svga->hwcursor.yoff &= 0x3f;
 		}
 		svga->hwcursor.addr  = (et4000->regs[0xe8] | (et4000->regs[0xe9] << 8) | ((et4000->regs[0xea] & 7) << 16)) << 2;
+
 		add2addr = svga->hwcursor.yoff * ((svga->hwcursor.xsize == 128) ? 32 : 16);
 		svga->hwcursor.addr += add2addr;
 		return;
@@ -385,28 +364,42 @@ et4000w32p_recalctimings(svga_t *svga)
     svga->clock = (cpuclock * (double)(1ull << 32)) / svga->getclock((svga->miscout >> 2) & 3, svga->clock_gen);
 
     if (svga->adv_flags & FLAG_NOSKEW) {
-	/* On the Cardex ET4000/W32p-based cards, adjust text mode clocks by 1. */
-	if (!(svga->gdcreg[6] & 1) && !(svga->attrregs[0x10] & 1)) {	/* Text mode */
-		svga->ma_latch--;
+		/* On the Cardex ET4000/W32p-based cards, adjust text mode clocks by 1. */
+		if (!(svga->gdcreg[6] & 1) && !(svga->attrregs[0x10] & 1)) {	/* Text mode */
+			svga->ma_latch--;
 
-		if ((svga->seqregs[1] & 8)) /*40 column*/
-			svga->hdisp += (svga->seqregs[1] & 1) ? 16 : 18;
-		else
-			svga->hdisp += (svga->seqregs[1] & 1) ? 8 : 9;
-	} else {
-		/* Also adjust the graphics mode clocks in some cases. */
-		if ((svga->gdcreg[5] & 0x40) && (svga->bpp != 32)) {
-			if ((svga->bpp == 15) || (svga->bpp == 16) || (svga->bpp == 24))
+			if ((svga->seqregs[1] & 8)) /*40 column*/
 				svga->hdisp += (svga->seqregs[1] & 1) ? 16 : 18;
 			else
 				svga->hdisp += (svga->seqregs[1] & 1) ? 8 : 9;
+		} else {
+			/* Also adjust the graphics mode clocks in some cases. */
+			if ((svga->gdcreg[5] & 0x40) && (svga->bpp != 32)) {
+				if ((svga->bpp == 15) || (svga->bpp == 16) || (svga->bpp == 24))
+					svga->hdisp += (svga->seqregs[1] & 1) ? 16 : 18;
+				else
+					svga->hdisp += (svga->seqregs[1] & 1) ? 8 : 9;
+			}
 		}
-	}
     }
 
-    et4000->adjust_cursor = 0;
+	if (et4000->type == ET4000W32) {
+		if ((svga->gdcreg[6] & 1) || (svga->attrregs[0x10] & 1)) {
+			if (svga->gdcreg[5] & 0x40) {
+				switch (svga->bpp) {
+					case 8:
+						if (svga->hdisp == 640 || svga->hdisp == 800 || svga->hdisp == 1024)
+							break;
+						svga->hdisp -= 24;
+						break;
+				}
+			}
+		}
+	}
 
-    switch (svga->bpp) {
+    et4000->adjust_cursor = 0;
+	
+	switch (svga->bpp) {
 	case 15: case 16:
 		svga->hdisp >>= 1;
 		if (et4000->type <= ET4000W32P_REVC)
@@ -417,7 +410,7 @@ et4000w32p_recalctimings(svga_t *svga)
 		if (et4000->type <= ET4000W32P_REVC)
 			et4000->adjust_cursor = 2;
 		break;
-    }
+	}
 
     svga->render = svga_render_blank;
     if (!svga->scrblank && svga->attr_palette_enable) {
@@ -655,7 +648,7 @@ et4000w32p_accel_write_mmu(et4000w32p_t *et4000, uint32_t addr, uint8_t val)
 		if (!(et4000->acl.internal.ctrl_routing & 0x37))
 			et4000w32p_blit(0xFFFFFF, ~0, 0, 0, et4000);
 		et4000->acl.cpu_input_num = 0;
-	}							
+	}
 
 	if (et4000->acl.internal.ctrl_routing & 7) {
 		et4000->acl.cpu_input = (et4000->acl.cpu_input &~ (0xFF << (et4000->acl.cpu_input_num << 3))) |
@@ -679,86 +672,6 @@ et4000w32p_accel_write_mmu(et4000w32p_t *et4000, uint32_t addr, uint8_t val)
 }
 
 
-static void
-fifo_thread(void *param)
-{
-    et4000w32p_t *et4000 = (et4000w32p_t *)param;
-
-    uint64_t start_time = 0;
-    uint64_t end_time = 0;
-
-    fifo_entry_t *fifo;
-
-    while (1) {
-	thread_set_event(et4000->fifo_not_full_event);
-	thread_wait_event(et4000->wake_fifo_thread, -1);
-	thread_reset_event(et4000->wake_fifo_thread);
-	et4000->blitter_busy = 1;
-	while (!FIFO_EMPTY) {
-		start_time = plat_timer_read();
-		fifo = &et4000->fifo[et4000->fifo_read_idx & FIFO_MASK];
-
-		switch (fifo->addr_type & FIFO_TYPE) {
-			case FIFO_WRITE_BYTE:
-				et4000w32p_accel_write_fifo(et4000, fifo->addr_type & FIFO_ADDR, fifo->val);
-				break;
-			case FIFO_WRITE_MMU:
-				et4000w32p_accel_write_mmu(et4000, fifo->addr_type & FIFO_ADDR, fifo->val);
-				break;
-		}
-
-		et4000->fifo_read_idx++;
-		fifo->addr_type = FIFO_INVALID;
-
-		if (FIFO_ENTRIES > 0xe000)
-			thread_set_event(et4000->fifo_not_full_event);
-
-		end_time = plat_timer_read();
-		et4000->blitter_time += end_time - start_time;
-	}
-	et4000->blitter_busy = 0;
-    }
-}
-
-
-static __inline void
-wake_fifo_thread(et4000w32p_t *et4000)
-{
-    thread_set_event(et4000->wake_fifo_thread);	/* Wake up FIFO thread if moving from idle */
-}
-
-
-static void
-et4000w32p_wait_fifo_idle(et4000w32p_t *et4000)
-{
-    while (!FIFO_EMPTY) {
-	wake_fifo_thread(et4000);
-	thread_wait_event(et4000->fifo_not_full_event, 1);
-    }
-}
-
-
-static void
-et4000w32p_queue(et4000w32p_t *et4000, uint32_t addr, uint32_t val, uint32_t type)
-{
-    fifo_entry_t *fifo = &et4000->fifo[et4000->fifo_write_idx & FIFO_MASK];
-
-    if (FIFO_FULL) {
-	thread_reset_event(et4000->fifo_not_full_event);
-	if (FIFO_FULL)
-		thread_wait_event(et4000->fifo_not_full_event, -1); /*Wait for room in ringbuffer*/
-    }
-
-    fifo->val = val;
-    fifo->addr_type = (addr & FIFO_ADDR) | type;
-
-    et4000->fifo_write_idx++;
-
-    if ((FIFO_ENTRIES > 0xe000) || (FIFO_ENTRIES < 8))
-	wake_fifo_thread(et4000);
-}
-
-
 void
 et4000w32p_mmu_write(uint32_t addr, uint8_t val, void *p)
 {
@@ -771,10 +684,7 @@ et4000w32p_mmu_write(uint32_t addr, uint8_t val, void *p)
 	case 0x4000:	/* MMU 2 */
 		et4000->bank = (addr >> 13) & 3;
 		if (et4000->mmu.ctrl & (1 << et4000->bank)) {
-			if (et4000->type >= ET4000W32P_REVC)
-				et4000w32p_queue(et4000, addr & 0x7fff, val, FIFO_WRITE_MMU);
-			else
-				et4000w32p_accel_write_mmu(et4000, addr & 0x7fff, val);
+			et4000w32p_accel_write_mmu(et4000, addr & 0x7fff, val);
 		} else {
 			if (((addr & 0x1fff) + et4000->mmu.base[et4000->bank]) < svga->vram_max) {
 				svga->vram[(addr & 0x1fff) + et4000->mmu.base[et4000->bank]] = val;
@@ -784,10 +694,7 @@ et4000w32p_mmu_write(uint32_t addr, uint8_t val, void *p)
 		break;
 	case 0x6000:
 		if ((addr & 0x7fff) >= 0x7f80) {
-			if (et4000->type >= ET4000W32P_REVC)
-				et4000w32p_queue(et4000, addr & 0x7fff, val, FIFO_WRITE_BYTE);
-			else
-				et4000w32p_accel_write_fifo(et4000, addr & 0x7fff, val);
+			et4000w32p_accel_write_fifo(et4000, addr & 0x7fff, val);
 		} else switch (addr & 0x7fff) {
 			case 0x7f00:	et4000->mmu.base[0] = (et4000->mmu.base[0] & 0xFFFFFF00) | val;		break;
 			case 0x7f01:	et4000->mmu.base[0] = (et4000->mmu.base[0] & 0xFFFF00FF) | (val << 8);	break;
@@ -822,9 +729,6 @@ et4000w32p_mmu_read(uint32_t addr, void *p)
 	case 0x4000:	/* MMU 2 */
 		bank = (addr >> 13) & 3;
 		if (et4000->mmu.ctrl & (1 << bank)) {
-			if (et4000->type >= ET4000W32P_REVC)
-				et4000w32p_wait_fifo_idle(et4000);
-
 			temp = 0xff;
 			if (et4000->acl.cpu_dat_pos) {
 				et4000->acl.cpu_dat_pos--;
@@ -844,10 +748,6 @@ et4000w32p_mmu_read(uint32_t addr, void *p)
 		return svga->vram[(addr&0x1fff) + et4000->mmu.base[bank]];
 
 	case 0x6000:
-		if ((addr & 0x7fff) >= 0x7f80) {
-			if (et4000->type >= ET4000W32P_REVC)
-				et4000w32p_wait_fifo_idle(et4000);
-		}
 		switch (addr & 0x7fff) {
 			case 0x7f00:	return et4000->mmu.base[0];
 			case 0x7f01:	return et4000->mmu.base[0] >> 8;
@@ -867,10 +767,6 @@ et4000w32p_mmu_read(uint32_t addr, void *p)
 				if (et4000->type >= ET4000W32P_REVC) {
 					temp = et4000->acl.status;
 					temp &= ~(ACL_RDST | ACL_WRST);
-					if (!FIFO_EMPTY)
-						temp |= ACL_RDST;
-					if (FIFO_FULL)
-						temp |= ACL_WRST;
 					if (temp == ACL_XYST && (et4000->acl.internal.ctrl_routing == 1 || et4000->acl.internal.ctrl_routing == 2))
 						temp |= ACL_RDST;
 				} else {
@@ -1280,13 +1176,29 @@ et4000w32p_hwcursor_draw(svga_t *svga, int displine)
     int shift = (et4000->adjust_cursor + 1);
     int width = (svga->hwcursor_latch.xsize - svga->hwcursor_latch.xoff);
     int pitch = (svga->hwcursor_latch.xsize == 128) ? 32 : 16;
+	int x_acc = 4;
+	int minus_width = 0;
     uint8_t dat;
     offset = svga->hwcursor_latch.xoff;
+	
+	if (et4000->type == ET4000W32) {
+		switch (svga->bpp) {
+			case 8:
+				minus_width = 0;
+				x_acc = 2;
+				break;
+			case 15: case 16:
+				minus_width = 64;
+				x_acc = 2;
+				break;
+		}
+	}
 
-    for (x = 0; x < width; x += 4) {
+    for (x = 0; x < (width - minus_width); x += x_acc) {
 	dat = svga->vram[svga->hwcursor_latch.addr + (offset >> 2)];
 
 	xx = svga->hwcursor_latch.x + svga->x_add + x;
+
 	if (!(xx % shift)) {
 		xx2 = xx / shift;
 		if (!(dat & 2))			buffer32->line[displine][xx2]  = (dat & 1) ? 0xFFFFFF : 0;
@@ -1318,7 +1230,7 @@ et4000w32p_hwcursor_draw(svga_t *svga, int displine)
 	offset += 4;
     }
 
-    svga->hwcursor_latch.addr += pitch;
+	svga->hwcursor_latch.addr += pitch;
 }
 
 
@@ -1375,10 +1287,10 @@ et4000w32p_pci_read(int func, int addr, void *p)
 
 	case 0x07:	return 1 << 1;	/* Medium DEVSEL timing */
 
-	case 0x08: return 0x06;	/* Revision ID */
+	case 0x08: return (et4000->rev);	/* Revision ID */
 	case 0x09: return 0;	/* Programming interface */
 
-	case 0x0a: return 0x00;	/* Supports VGA interface, XGA compatible */
+	case 0x0a: return 0x00;	/* Supports VGA interface */
 	case 0x0b: return 0x03;	/* This has to be done in order to make this card work with the two 486 PCI machines. */
 
 	case 0x10: return 0x00;	/* Linear frame buffer address */
@@ -1450,8 +1362,17 @@ et4000w32p_init(const device_t *info)
     et4000w32p_t *et4000 = malloc(sizeof(et4000w32p_t));
     memset(et4000, 0, sizeof(et4000w32p_t));
 
-    vram_size = device_get_config_int("memory");
-    et4000->interleaved = (vram_size == 2 && (info->local >= ET4000W32I)) ? 1 : 0;
+    et4000->pci = (info->flags & DEVICE_PCI) ? 0x80 : 0x00;
+    et4000->vlb = (info->flags & DEVICE_VLB) ? 0x40 : 0x00;
+	
+	/*The ET4000/W32i ISA BIOS seems to not support 2MB of VRAM*/
+	if ((info->local == ET4000W32) || ((info->local == ET4000W32I) && !(et4000->vlb)))
+		vram_size = 1;
+	else
+		vram_size = device_get_config_int("memory");
+
+	/*The interleaved VRAM was introduced by the ET4000/W32i*/
+    et4000->interleaved = ((vram_size == 2) && (info->local != ET4000W32)) ? 1 : 0;
 
     if (info->flags & DEVICE_PCI)
 	video_inform(VIDEO_FLAG_TYPE_SPECIAL, &timing_et4000w32_pci);
@@ -1467,22 +1388,22 @@ et4000w32p_init(const device_t *info)
 	      NULL);
 
     et4000->vram_mask = (vram_size << 20) - 1;
+	et4000->svga.decode_mask = (vram_size << 20) - 1;
 
     et4000->type = info->local;
 
-    et4000->pci = (info->flags & DEVICE_PCI) ? 0x80 : 0x00;
-    et4000->vlb = (info->flags & DEVICE_VLB) ? 0x40 : 0x00;
-
     switch(et4000->type) {
 	case ET4000W32:
 		/* ET4000/W32 */
+		et4000->rev = 0;
+		
 		rom_init(&et4000->bios_rom, BIOS_ROM_PATH_W32, 0xc0000, 0x8000, 0x7fff, 0,
 			 MEM_MAPPING_EXTERNAL);
 
 		et4000->svga.ramdac = device_add(&tseng_ics5301_ramdac_device);
 		et4000->svga.clock_gen = et4000->svga.ramdac;
 		et4000->svga.getclock = sdac_getclock;
-		break;			
+		break;	
 
 	case ET4000W32I:
 		/* ET4000/W32i rev B */
@@ -1577,12 +1498,7 @@ et4000w32p_init(const device_t *info)
     et4000->pci_regs[0x32] = 0x00;
     et4000->pci_regs[0x33] = 0xf0;
 
-    if (et4000->type >= ET4000W32P_REVC) {
-	et4000w32_log("Thread started\n");
-	et4000->wake_fifo_thread = thread_create_event();
-	et4000->fifo_not_full_event = thread_create_event();
-	et4000->fifo_thread = thread_create(fifo_thread, et4000);
-    }
+	et4000->svga.packed_chain4 = 1;
 
     return et4000;
 }
@@ -1644,12 +1560,6 @@ et4000w32p_close(void *p)
 
     svga_close(&et4000->svga);
 
-    if (et4000->type >= ET4000W32P_REVC) {
-	thread_kill(et4000->fifo_thread);
-	thread_destroy_event(et4000->wake_fifo_thread);
-	thread_destroy_event(et4000->fifo_not_full_event);
-    }
-
     free(et4000);
 }
 
@@ -1702,7 +1612,7 @@ const device_t et4000w32_device =
         { et4000w32_available },
         et4000w32p_speed_changed,
         et4000w32p_force_redraw,
-        et4000w32p_config
+        NULL
 };
 
 const device_t et4000w32_onboard_device =
@@ -1713,7 +1623,7 @@ const device_t et4000w32_onboard_device =
         { et4000w32_available },
         et4000w32p_speed_changed,
         et4000w32p_force_redraw,
-        et4000w32p_config
+        NULL
 };
 
 const device_t et4000w32i_isa_device =
@@ -1724,7 +1634,7 @@ const device_t et4000w32i_isa_device =
         { et4000w32i_isa_available },
         et4000w32p_speed_changed,
         et4000w32p_force_redraw,
-        et4000w32p_config
+        NULL
 };
 
 const device_t et4000w32i_vlb_device =
diff --git a/src/video/vid_ht216.c b/src/video/vid_ht216.c
index 96d2af838..ab1a86814 100644
--- a/src/video/vid_ht216.c
+++ b/src/video/vid_ht216.c
@@ -127,10 +127,20 @@ ht216_log(const char *fmt, ...)
 #define ht216_log(fmt, ...)
 #endif
 
+/*Remap address for chain-4/doubleword style layout*/
+static __inline uint32_t
+dword_remap(svga_t *svga, uint32_t in_addr)
+{
+	if (svga->packed_chain4)
+		return in_addr;
+	return ((in_addr & 0xfffc) << 2) | ((in_addr & 0x30000) >> 14) | (in_addr & ~0x3ffff);
+}
 
 static void
 ht216_recalc_bank_regs(ht216_t *ht216, int mode)
 {
+	svga_t *svga = &ht216->svga;
+	
 	if (mode) {
 	ht216->read_bank_reg[0] = ht216->ht_regs[0xe8];
 	ht216->write_bank_reg[0] = ht216->ht_regs[0xe8];
@@ -142,14 +152,14 @@ ht216_recalc_bank_regs(ht216_t *ht216, int mode)
 	ht216->write_bank_reg[0] = ((ht216->ht_regs[0xf6] & 0x3) << 6);
 	ht216->write_bank_reg[1] = ((ht216->ht_regs[0xf6] & 0x3) << 6);
 
-	if ((ht216->ht_regs[0xfc] & 0x20) || (ht216->ht_regs[0xfc] & 0x04)) {
+	if (svga->packed_chain4 || (ht216->ht_regs[0xfc] & HT_REG_FC_ECOLRE)) {
 		ht216->read_bank_reg[0]  |= (ht216->misc & 0x20);
 		ht216->read_bank_reg[1]  |= (ht216->misc & 0x20);
 		ht216->write_bank_reg[0] |= (ht216->misc & 0x20);
 		ht216->write_bank_reg[1] |= (ht216->misc & 0x20);
 	}
 
-	if ((ht216->ht_regs[0xfc] & 0x20) || ((ht216->ht_regs[0xfc] & 0x06) == 0x04)) {
+	if (svga->packed_chain4 || ((ht216->ht_regs[0xfc] & 0x06) == 0x04)) {
 		ht216->read_bank_reg[0]  |= ((ht216->ht_regs[0xf9] & 1) << 4);
 		ht216->read_bank_reg[1]  |= ((ht216->ht_regs[0xf9] & 1) << 4);
 		ht216->write_bank_reg[0] |= ((ht216->ht_regs[0xf9] & 1) << 4);
@@ -355,7 +365,8 @@ ht216_out(uint16_t addr, uint8_t val, void *p)
 					break;
 				
 				case 0xfc:
-					ht216_log("HT216 reg 0xfc write = %02x, mode = 0, chain4 = %x, bit 7 = %02x\n", val, svga->chain4, val & 0x80);
+					ht216_log("HT216 reg 0xfc write = %02x, mode = 0, chain4 = %x, bit 7 = %02x, packedchain = %02x\n", val, svga->chain4, val & 0x80, val & 0x20);
+					svga->packed_chain4 = !!(val & 0x20);
 					ht216_recalc_bank_regs(ht216, 0);
 					ht216_remap(ht216);
 					svga->fullchange = changeframecount;
@@ -383,6 +394,11 @@ ht216_out(uint16_t addr, uint8_t val, void *p)
 				else
 					svga->banked_mask = 0xffff;
 			}
+			
+			if (svga->gdcaddr <= 8) {
+				svga->fast = (svga->gdcreg[8] == 0xff && !(svga->gdcreg[3] & 0x18) &&
+						 !svga->gdcreg[1]) && svga->chain4 && svga->packed_chain4;
+			}
 			break;
 
 		case 0x3D4:
@@ -559,82 +575,6 @@ ht216_remap(ht216_t *ht216)
 }
 
 
-void
-ht216_1_2bpp_highres(svga_t *svga)
-{
-    int changed_offset, x;
-    int oddeven;
-    uint32_t addr, *p;
-    uint8_t edat[4];
-    uint8_t dat;
-
-    if ((svga->displine + svga->y_add) < 0)
-	return;
-
-    changed_offset = (svga->ma + (svga->sc & ~svga->crtc[0x17] & 3) * 0x8000) >> 12;
-
-    if (svga->changedvram[changed_offset] || svga->changedvram[changed_offset + 1] || svga->fullchange) {
-	p = &buffer32->line[svga->displine + svga->y_add][svga->x_add];
-
-	if (svga->firstline_draw == 2000) 
-		svga->firstline_draw = svga->displine;
-	svga->lastline_draw = svga->displine;
-
-	for (x = 0; x <= (svga->hdisp + svga->scrollcache); x += 8) {
-		addr = svga->ma;
-		oddeven = 0;
-
-		if (!(svga->crtc[0x17] & 0x40)) {
-			addr = (addr << 1) & svga->vram_mask;
-
-			if (svga->seqregs[1] & 4)
-				oddeven = (addr & 4) ? 1 : 0;
-
-			addr &= ~7;
-
-			if ((svga->crtc[0x17] & 0x20) && (svga->ma & 0x20000))
-				addr |= 4;
-			if (!(svga->crtc[0x17] & 0x20) && (svga->ma & 0x8000))
-				addr |= 4;
-		}
-
-		if (!(svga->crtc[0x17] & 0x01))
-			addr = (addr & ~0x8000) | ((svga->sc & 1) ? 0x8000 : 0);
-		if (!(svga->crtc[0x17] & 0x02))
-			addr = (addr & ~0x10000) | ((svga->sc & 2) ? 0x10000 : 0);
-
-		if (svga->seqregs[1] & 4) {
-			edat[0] = svga->vram[addr | oddeven];
-			edat[2] = svga->vram[addr | oddeven | 0x2];
-			edat[1] = edat[3] = 0;
-		} else {
-			*(uint32_t *)(&edat[0]) = *(uint32_t *)(&svga->vram[addr]);
-		}
-		svga->ma += 4;
-		svga->ma &= svga->vram_mask;
-
-		if (svga->crtc[0x17] & 0x80) {
-			dat = edatlookup[edat[0] >> 6][edat[1] >> 6] | (edatlookup[edat[2] >> 6][edat[3] >> 6] << 2);
-			p[0] = svga->pallook[svga->egapal[(dat >> 4) & svga->plane_mask]];
-			p[1] = svga->pallook[svga->egapal[dat & svga->plane_mask]];
-			dat = edatlookup[(edat[0] >> 4) & 3][(edat[1] >> 4) & 3] | (edatlookup[(edat[2] >> 4) & 3][(edat[3] >> 4) & 3] << 2);
-			p[2] = svga->pallook[svga->egapal[(dat >> 4) & svga->plane_mask]];
-			p[3] = svga->pallook[svga->egapal[dat & svga->plane_mask]];
-			dat = edatlookup[(edat[0] >> 2) & 3][(edat[1] >> 2) & 3] | (edatlookup[(edat[2] >> 2) & 3][(edat[3] >> 2) & 3] << 2);
-			p[4] = svga->pallook[svga->egapal[(dat >> 4) & svga->plane_mask]];
-			p[5] = svga->pallook[svga->egapal[dat & svga->plane_mask]];
-			dat = edatlookup[edat[0] & 3][edat[1] & 3] | (edatlookup[edat[2] & 3][edat[3] & 3] << 2);
-			p[6] = svga->pallook[svga->egapal[(dat >> 4) & svga->plane_mask]];
-			p[7] = svga->pallook[svga->egapal[dat & svga->plane_mask]];
-		} else
-			memset(p, 0x00, 8 * sizeof(uint32_t));
-
-		p += 8;
-	}
-    }
-}
-
-
 void
 ht216_recalctimings(svga_t *svga)
 {
@@ -674,7 +614,7 @@ ht216_recalctimings(svga_t *svga)
     
     if (svga->crtc[0x17] == 0xeb) {
 	svga->rowoffset <<= 1;
-	svga->render = ht216_1_2bpp_highres;
+	svga->render = svga_render_2bpp_headland_highres;
     }
 
     if (svga->bpp == 8) {
@@ -691,7 +631,7 @@ ht216_recalctimings(svga_t *svga)
 			ht216->adjust_cursor = 1;
 			svga->render = svga_render_8bpp_highres;
 		} else {
-			ht216_log("8bpp low\n");
+			ht216_log("8bpp low, packed = %02x, chain4 = %02x\n", svga->packed_chain4, svga->chain4);
 			svga->render = svga_render_8bpp_lowres;
 		}
 	} else if (ht216->ht_regs[0xfc] & HT_REG_FC_ECOLRE) {
@@ -702,6 +642,8 @@ ht216_recalctimings(svga_t *svga)
 			svga->hdisp++;
 			svga->hdisp *= (svga->seqregs[1] & 8) ? 16 : 8;
 			svga->rowoffset <<= 1;
+			if ((svga->crtc[0x17] & 0x60) == 0x20) /*Would result in a garbled screen with trailing cursor glitches*/
+				svga->crtc[0x17] |= 0x40;
 		}
 		svga->render = svga_render_8bpp_highres;
 	}
@@ -711,7 +653,7 @@ ht216_recalctimings(svga_t *svga)
 		svga->render = svga_render_15bpp_highres;
 	}
 
-    svga->ma_latch |= ((ht216->ht_regs[0xf6] & 0x30) << 14);
+	svga->ma_latch |= ((ht216->ht_regs[0xf6] & 0x30) << 14);
 
     if (svga->crtc[0x17] == 0xeb) /*Looks like that 1024x768 mono mode expects 512K of video memory*/
 	svga->vram_display_mask = 0x7ffff;
@@ -787,7 +729,6 @@ extalu(int op, uint8_t input_a, uint8_t input_b)
     return val;
 }
 
-
 static void
 ht216_dm_write(ht216_t *ht216, uint32_t addr, uint8_t cpu_dat, uint8_t cpu_dat_unexpanded)
 {
@@ -801,12 +742,12 @@ ht216_dm_write(ht216_t *ht216, uint32_t addr, uint8_t cpu_dat, uint8_t cpu_dat_u
 	writemask2 = svga->seqregs[2];
 
     if (!(svga->gdcreg[6] & 1))
-	svga->fullchange = 2;
+		svga->fullchange = 2;
 
-    if (svga->chain4 || svga->fb_only) {
+	if (svga->chain4) {
 	writemask2 = 1 << (addr & 3);
-	addr &= ~3;
-    } else if (svga->chain2_write && (svga->crtc[0x17] != 0xeb)) {
+	addr = dword_remap(svga, addr) & ~3;
+	} else if (svga->chain2_write && (svga->crtc[0x17] != 0xeb)) {
 	writemask2 &= ~0xa;
 	if (addr & 1)
 		writemask2 <<= 1;
@@ -984,9 +925,10 @@ ht216_dm_extalu_write(ht216_t *ht216, uint32_t addr, uint8_t cpu_dat, uint8_t bi
     uint8_t input_a = 0, input_b = 0;
     uint8_t fg, bg;
     uint8_t output;
+	uint32_t remapped_addr = dword_remap(svga, addr);
 
     if (ht216->ht_regs[0xcd] & HT_REG_CD_RMWMDE) /*RMW*/
-	input_b = svga->vram[addr];
+	input_b = svga->vram[remapped_addr];
     else
 	input_b = ht216->bg_latch[addr & 7];
 
@@ -1011,8 +953,8 @@ ht216_dm_extalu_write(ht216_t *ht216, uint32_t addr, uint8_t cpu_dat, uint8_t bi
     fg = extalu(ht216->ht_regs[0xce] >> 4, input_a, input_b);
     bg = extalu(ht216->ht_regs[0xce] & 0xf,  input_a, input_b);
     output = (fg & rop_select) | (bg & ~rop_select);
-    svga->vram[addr] = (svga->vram[addr] & ~bit_mask) | (output & bit_mask);
-    svga->changedvram[addr >> 12] = changeframecount;
+    svga->vram[addr] = (svga->vram[remapped_addr] & ~bit_mask) | (output & bit_mask);
+    svga->changedvram[remapped_addr >> 12] = changeframecount;
 }
 
 static void
@@ -1027,12 +969,12 @@ ht216_dm_masked_write(ht216_t *ht216, uint32_t addr, uint8_t val, uint8_t bit_ma
 	writemask2 = svga->seqregs[2];
 
     if (!(svga->gdcreg[6] & 1))
-	svga->fullchange = 2;
+		svga->fullchange = 2;
 
-    if ((svga->chain4 || svga->fb_only) && (svga->writemode < 4)) {
+	if (svga->chain4) {
 	writemask2 = 1 << (addr & 3);
-	addr &= ~3;
-    } else if (svga->chain2_write) {
+	addr = dword_remap(svga, addr) & ~3;
+	} else if (svga->chain2_write) {
 	writemask2 &= ~0xa;
 	if (addr & 1)
 		writemask2 <<= 1;
@@ -1304,7 +1246,7 @@ ht216_read_common(ht216_t *ht216, uint32_t addr)
 
     count = (1 << count);
 
-    if (svga->chain4 || svga->fb_only) {
+    if (svga->chain4 && svga->packed_chain4) {
 	addr &= svga->decode_mask;
 	if (addr >= svga->vram_max)
 		return 0xff;
@@ -1312,9 +1254,12 @@ ht216_read_common(ht216_t *ht216, uint32_t addr)
 	if (ht216->ht_regs[0xcd] & HT_REG_CD_ASTODE)
 		latch_addr += (svga->gdcreg[3] & 7);
 	for (i = 0; i < 8; i++)
-		ht216->bg_latch[i] = svga->vram[latch_addr + i];
-	return svga->vram[addr & svga->vram_mask];
-    } else if (svga->chain2_read && (svga->crtc[0x17] != 0xeb)) {
+		ht216->bg_latch[i] = svga->vram[dword_remap(svga, latch_addr + i)];
+	return svga->vram[dword_remap(svga, addr) & svga->vram_mask];
+    } else if (svga->chain4) {
+	readplane = addr & 3;
+	addr = ((addr & 0xfffc) << 2) | ((addr & 0x30000) >> 14) | (addr & ~0x3ffff);
+	} else if (svga->chain2_read && (svga->crtc[0x17] != 0xeb)) {
 	readplane = (readplane & 2) | (addr & 1);
 	addr &= ~1;
 	addr <<= 2;
diff --git a/src/video/vid_paradise.c b/src/video/vid_paradise.c
index 49f496942..567d7fd4c 100644
--- a/src/video/vid_paradise.c
+++ b/src/video/vid_paradise.c
@@ -47,6 +47,8 @@ typedef struct paradise_t
 		WD90C30
         } type;
 
+		uint32_t vram_mask;
+
         uint32_t read_bank[4], write_bank[4];
 } paradise_t;
 
@@ -112,17 +114,20 @@ void paradise_out(uint16_t addr, uint8_t val, void *p)
                         paradise_remap(paradise);
                         return;
                 }
-                if (svga->gdcaddr == 0x9 || svga->gdcaddr == 0xa)
+                if (svga->gdcaddr == 0x9 || svga->gdcaddr == 0xa || svga->gdcaddr == 0xb)
                 {
                         svga->gdcreg[svga->gdcaddr] = val;
                         paradise_remap(paradise);
                         return;
                 }
-                if (svga->gdcaddr == 0xe)
-                {
+				if (svga->gdcaddr == 0xd) {
+						svga->gdcreg[0xd] = val;
+						svga_recalctimings(svga);
+                        return;				
+				}
+                if (svga->gdcaddr == 0xe) {
                         svga->gdcreg[0xe] = val;
-                        paradise_remap(paradise);
-			svga_recalctimings(svga);
+						svga_recalctimings(svga);
                         return;
                 }
                 break;
@@ -152,6 +157,13 @@ void paradise_out(uint16_t addr, uint8_t val, void *p)
                                 svga_recalctimings(&paradise->svga);
                         }
                 }
+
+				switch (svga->crtcreg) {
+					case 0x2f:
+						if (paradise->type != PVGA1A)
+							svga->vram_display_mask = (val & 1) ? 0x3ffff : paradise->vram_mask;
+						break;
+				}
                 break;
         }
         svga_out(addr, val, svga);
@@ -219,8 +231,8 @@ void paradise_remap(paradise_t *paradise)
                 paradise->write_bank[0] = paradise->write_bank[2] =  (svga->gdcreg[0xa] & mask) << 12;
                 paradise->write_bank[1] = paradise->write_bank[3] = ((svga->gdcreg[0xa] & mask) << 12) + ((svga->gdcreg[6] & 0x08) ? 0 : 0x8000);
         }
-        else if (svga->gdcreg[0xe] & 0x08)
-        {
+        else if (svga->gdcreg[0xb] & 0x08)
+        {	
                 if (svga->gdcreg[0x6] & 0xc)
                 {
                         paradise->read_bank[0]  = paradise->read_bank[2]  =  (svga->gdcreg[0xa] & mask) << 12;
@@ -243,19 +255,37 @@ void paradise_remap(paradise_t *paradise)
                 paradise->write_bank[0] = paradise->write_bank[2] =  (svga->gdcreg[0x9] & mask) << 12;
                 paradise->write_bank[1] = paradise->write_bank[3] = ((svga->gdcreg[0x9] & mask) << 12) + ((svga->gdcreg[6] & 0x08) ? 0 : 0x8000);
         }
+
+		if (paradise->type == WD90C11) {
+			paradise->read_bank[1] &= 0x7ffff;
+			paradise->write_bank[1] &= 0x7ffff;
+		} else {
+			paradise->read_bank[1] &= 0xfffff;
+			paradise->write_bank[1] &= 0xfffff;			
+		}
 }
 
 void paradise_recalctimings(svga_t *svga)
 {
 	paradise_t *paradise = (paradise_t *) svga->p;
 
-	if (paradise->type == WD90C30)
-		svga->interlace = (svga->crtc[0x2d] & 0x20);
+	if (svga->gdcreg[0xd] & 0x08) svga->ma_latch |= 0x10000;
+	if (svga->gdcreg[0xd] & 0x10) svga->ma_latch |= 0x20000;
 
-	if (svga->gdcreg[0xe] & 0x01) {
-		svga->bpp = 8;
-		svga->lowres = 0;
-                svga->render = svga_render_8bpp_highres;
+	if (paradise->type == WD90C30) {
+		if (svga->crtc[0x3d] & 0x01) svga->vtotal |= 0x400;
+		if (svga->crtc[0x3d] & 0x02) svga->dispend |= 0x400;
+		if (svga->crtc[0x3d] & 0x04) svga->vsyncstart |= 0x400;
+		if (svga->crtc[0x3d] & 0x08) svga->vblankstart |= 0x400;
+		if (svga->crtc[0x3d] & 0x10) svga->split |= 0x400;
+
+		svga->interlace = (svga->crtc[0x2d] & 0x20);
+	}
+	
+	svga->lowres = !(svga->gdcreg[0xe] & 0x01);
+	
+	if (svga->bpp == 8 && !svga->lowres) {
+        svga->render = svga_render_8bpp_highres;
 	}
 }
 
@@ -263,7 +293,6 @@ static void paradise_write(uint32_t addr, uint8_t val, void *p)
 {
         paradise_t *paradise = (paradise_t *)p;
         addr = (addr & 0x7fff) + paradise->write_bank[(addr >> 15) & 3];
-
         svga_write_linear(addr, val, &paradise->svga);
 }
 static void paradise_writew(uint32_t addr, uint16_t val, void *p)
@@ -306,6 +335,7 @@ void *paradise_init(const device_t *info, uint32_t memsize)
 				   paradise_in, paradise_out,
 				   NULL,
 				   NULL);
+			paradise->vram_mask = memsize - 1;
 			break;
 		case WD90C11:
 			svga_init(info, &paradise->svga, paradise, 1 << 19, /*512kb*/
@@ -313,6 +343,7 @@ void *paradise_init(const device_t *info, uint32_t memsize)
 				   paradise_in, paradise_out,
 				   NULL,
 				   NULL);
+			paradise->vram_mask = (1 << 19) - 1;
 			break;
 		case WD90C30:
 			svga_init(info, &paradise->svga, paradise, memsize,
@@ -320,11 +351,12 @@ void *paradise_init(const device_t *info, uint32_t memsize)
 				   paradise_in, paradise_out,
 				   NULL,
 				   NULL);
+			paradise->vram_mask = memsize - 1;
 			break;
 	}
 
-        mem_mapping_set_handler(&paradise->svga.mapping, paradise_read, paradise_readw, NULL, paradise_write, paradise_writew, NULL);
-        mem_mapping_set_p(&paradise->svga.mapping, paradise);
+        mem_mapping_set_handler(&svga->mapping, paradise_read, paradise_readw, NULL, paradise_write, paradise_writew, NULL);
+        mem_mapping_set_p(&svga->mapping, paradise);
 
 	/* Common to all three types. */
 	svga->crtc[0x31] = 'W';
diff --git a/src/video/vid_s3.c b/src/video/vid_s3.c
index 5977537e3..bb8474ab9 100644
--- a/src/video/vid_s3.c
+++ b/src/video/vid_s3.c
@@ -353,6 +353,28 @@ static uint8_t s3_accel_in(uint16_t port, void *p);
 static uint8_t	s3_pci_read(int func, int addr, void *p);
 static void	s3_pci_write(int func, int addr, uint8_t val, void *p);
 
+/*Remap address for chain-4/doubleword style layout*/
+static __inline uint32_t 
+dword_remap(uint32_t in_addr)
+{
+	return ((in_addr << 2) & 0x3fff0) |
+		((in_addr >> 14) & 0xc) |
+		(in_addr & ~0x3fffc);
+}
+static __inline uint32_t 
+dword_remap_w(uint32_t in_addr)
+{
+	return ((in_addr << 2) & 0x1fff8) |
+		((in_addr >> 14) & 0x6) |
+		(in_addr & ~0x1fffe);
+}
+static __inline uint32_t 
+dword_remap_l(uint32_t in_addr)
+{
+	return ((in_addr << 2) & 0xfffc) |
+		((in_addr >> 14) & 0x3) |
+		(in_addr & ~0xffff);
+}
 
 static __inline void
 wake_fifo_thread(s3_t *s3)
@@ -1635,8 +1657,10 @@ s3_hwcursor_draw(svga_t *svga, int displine)
 
 	for (x = 0; x < 64; x += 16)
 	{
-		dat[0] = (svga->vram[svga->hwcursor_latch.addr]     << 8) | svga->vram[svga->hwcursor_latch.addr + 1];
-		dat[1] = (svga->vram[svga->hwcursor_latch.addr + 2] << 8) | svga->vram[svga->hwcursor_latch.addr + 3];
+		uint32_t remapped_addr = dword_remap(svga->hwcursor_latch.addr);
+	
+		dat[0] = (svga->vram[remapped_addr]     << 8) | svga->vram[remapped_addr + 1];
+		dat[1] = (svga->vram[remapped_addr + 2] << 8) | svga->vram[remapped_addr + 3];
                 if (svga->crtc[0x55] & 0x10) {
                         /*X11*/
                         for (xx = 0; xx < 16; xx++) {
@@ -2240,6 +2264,7 @@ s3_out(uint16_t addr, uint8_t val, void *p)
 		{
 			case 0x31:
 			s3->ma_ext = (s3->ma_ext & 0x1c) | ((val & 0x30) >> 4);
+			svga->force_dword_mode = val & 0x08;
 			break;
 			case 0x32:
 			if (svga->crtc[0x31] & 0x30)
@@ -3698,9 +3723,9 @@ polygon_setup(s3_t *s3)
 }
 
 
-#define READ(addr, dat) if (s3->bpp == 0)      dat = svga->vram[  (addr) & s3->vram_mask]; \
-			    else if (s3->bpp == 1) dat = vram_w[(addr) & (s3->vram_mask >> 1)]; \
-			    else		   dat = vram_l[(addr) & (s3->vram_mask >> 2)];
+#define READ(addr, dat) if (s3->bpp == 0)      dat = svga->vram[dword_remap(addr) & s3->vram_mask]; \
+			    else if (s3->bpp == 1) dat = vram_w[dword_remap_w(addr) & (s3->vram_mask >> 1)]; \
+			    else		   dat = vram_l[dword_remap_l(addr) & (s3->vram_mask >> 2)];
 
 #define MIX_READ {											       \
 			switch ((mix_dat & mix_mask) ? (s3->accel.frgd_mix & 0xf) : (s3->accel.bkgd_mix & 0xf)) \
@@ -4005,18 +4030,18 @@ polygon_setup(s3_t *s3)
 
 #define WRITE(addr, dat)     if (s3->bpp == 0)									       \
 			{											       \
-				svga->vram[(addr) & s3->vram_mask] = dat;					  \
-				svga->changedvram[((addr) & s3->vram_mask) >> 12] = changeframecount;		   \
+				svga->vram[dword_remap(addr) & s3->vram_mask] = dat;					  \
+				svga->changedvram[(dword_remap(addr) & s3->vram_mask) >> 12] = changeframecount;		   \
 			}											       \
 			else if (s3->bpp == 1)									  \
 			{											       \
-				vram_w[(addr) & (s3->vram_mask >> 1)] = dat;				       \
-				svga->changedvram[((addr) & (s3->vram_mask >> 1)) >> 11] = changeframecount;	    \
+				vram_w[dword_remap_w(addr) & (s3->vram_mask >> 1)] = dat;				       \
+				svga->changedvram[(dword_remap_w(addr) & (s3->vram_mask >> 1)) >> 11] = changeframecount;	    \
 			}											       \
 			else											    \
 			{											       \
-				vram_l[(addr) & (s3->vram_mask >> 2)] = dat;				       \
-				svga->changedvram[((addr) & (s3->vram_mask >> 2)) >> 10] = changeframecount;	    \
+				vram_l[dword_remap_l(addr) & (s3->vram_mask >> 2)] = dat;				       \
+				svga->changedvram[(dword_remap_l(addr) & (s3->vram_mask >> 2)) >> 10] = changeframecount;	    \
 			}
 
 
diff --git a/src/video/vid_s3_virge.c b/src/video/vid_s3_virge.c
index 4b37eda62..727b7f905 100644
--- a/src/video/vid_s3_virge.c
+++ b/src/video/vid_s3_virge.c
@@ -3886,6 +3886,8 @@ static void *s3_virge_init(const device_t *info)
         virge->i2c = i2c_gpio_init("ddc_s3_virge");
         virge->ddc = ddc_init(i2c_gpio_get_bus(virge->i2c));
  
+		virge->svga.packed_chain4 = 1;
+ 
         return virge;
 }
 
diff --git a/src/video/vid_svga.c b/src/video/vid_svga.c
index de84df1f0..2dbd1566f 100644
--- a/src/video/vid_svga.c
+++ b/src/video/vid_svga.c
@@ -178,7 +178,7 @@ svga_out(uint16_t addr, uint8_t val, void *p)
 				svga->chain2_write = !(val & 4);
 				svga->chain4 = val & 8;
 				svga->fast = (svga->gdcreg[8] == 0xff && !(svga->gdcreg[3] & 0x18) &&
-					      !svga->gdcreg[1]) && svga->chain4 && !(svga->adv_flags & FLAG_ADDR_BY8);
+					      !svga->gdcreg[1]) && ((svga->chain4 && svga->packed_chain4) || svga->fb_only) && !(svga->adv_flags & FLAG_ADDR_BY8);
 				break;
 		}
 		break;
@@ -261,7 +261,7 @@ svga_out(uint16_t addr, uint8_t val, void *p)
 		}
 		svga->gdcreg[svga->gdcaddr & 15] = val;                
 		svga->fast = (svga->gdcreg[8] == 0xff && !(svga->gdcreg[3] & 0x18) &&
-			     !svga->gdcreg[1]) && svga->chain4;
+			     !svga->gdcreg[1]) && ((svga->chain4 && svga->packed_chain4) || svga->fb_only);
 		if (((svga->gdcaddr & 15) == 5  && (val ^ o) & 0x70) ||
 		    ((svga->gdcaddr & 15) == 6 && (val ^ o) & 1))
 			svga_recalctimings(svga);
@@ -575,7 +575,9 @@ svga_recalctimings(svga_t *svga)
     if (svga->dispontime < TIMER_USEC)
 	svga->dispontime = TIMER_USEC;
     if (svga->dispofftime < TIMER_USEC)
-	svga->dispofftime = TIMER_USEC;
+		svga->dispofftime = TIMER_USEC;
+
+	svga_recalc_remap_func(svga);
 
     /* Inform the user interface of any DPMS mode changes. */
     if (svga->dpms) {
@@ -1058,9 +1060,14 @@ svga_write_common(uint32_t addr, uint8_t val, uint8_t linear, void *p)
 
     if ((svga->adv_flags & FLAG_ADDR_BY8) && (svga->writemode < 4))
 	addr <<= 3;
-    else if ((svga->chain4 || svga->fb_only) && (svga->writemode < 4)) {
+    else if (((svga->chain4 && svga->packed_chain4) || svga->fb_only) && (svga->writemode < 4)) {
 	writemask2 = 1 << (addr & 3);
 	addr &= ~3;
+	} else if (svga->chain4 && (svga->writemode < 4)) {
+	writemask2 = 1 << (addr & 3);
+	if (!linear)
+		addr &= ~3;
+	addr = ((addr & 0xfffc) << 2) | ((addr & 0x30000) >> 14) | (addr & ~0x3ffff);
     } else if (svga->chain2_write) {
 	writemask2 &= ~0xa;
 	if (addr & 1)
@@ -1239,7 +1246,7 @@ svga_read_common(uint32_t addr, uint8_t linear, void *p)
 
     if (svga->adv_flags & FLAG_ADDR_BY8)
 	addr <<= 3;
-    else if (svga->chain4 || svga->fb_only) {
+    else if ((svga->chain4 && svga->packed_chain4) || svga->fb_only) {
 	addr &= svga->decode_mask;
 	if (svga->translate_address)
 		addr = svga->translate_address(addr, p);
@@ -1249,6 +1256,9 @@ svga_read_common(uint32_t addr, uint8_t linear, void *p)
 	for (i = 0; i < count; i++)
 		svga->latch.b[i] = svga->vram[latch_addr | i];
 	return svga->vram[addr & svga->vram_mask];
+	} else if (svga->chain4) {
+	readplane = addr & 3;
+	addr = ((addr & 0xfffc) << 2) | ((addr & 0x30000) >> 14) | (addr & ~0x3ffff);
     } else if (svga->chain2_read) {
 	readplane = (readplane & 2) | (addr & 1);
 	addr &= ~1;
diff --git a/src/video/vid_svga_render.c b/src/video/vid_svga_render.c
index aee22a1fe..530b4e2c3 100644
--- a/src/video/vid_svga_render.c
+++ b/src/video/vid_svga_render.c
@@ -27,6 +27,7 @@
 #include <86box/video.h>
 #include <86box/vid_svga.h>
 #include <86box/vid_svga_render.h>
+#include <86box/vid_svga_render_remap.h>
 
 void
 svga_render_null(svga_t *svga)
@@ -129,11 +130,13 @@ svga_render_text_40(svga_t *svga)
 	xinc = (svga->seqregs[1] & 1) ? 16 : 18;
 
 	for (x = 0; x < (svga->hdisp + svga->scrollcache); x += xinc) {
+		uint32_t addr = svga->remap_func(svga, svga->ma) & svga->vram_display_mask;		
+		
 		drawcursor = ((svga->ma == svga->ca) && svga->con && svga->cursoron);
 
 		if (svga->crtc[0x17] & 0x80) {
-			chr  = svga->vram[(svga->ma << 1) & svga->vram_display_mask];
-			attr = svga->vram[((svga->ma << 1) + 1) & svga->vram_display_mask];
+			chr  = svga->vram[addr];
+			attr = svga->vram[addr+1];
 		} else
 			chr = attr = 0;
 
@@ -196,11 +199,13 @@ svga_render_text_80(svga_t *svga)
 	xinc = (svga->seqregs[1] & 1) ? 8 : 9;
 
 	for (x = 0; x < (svga->hdisp + svga->scrollcache); x += xinc) {
+		uint32_t addr = svga->remap_func(svga, svga->ma) & svga->vram_display_mask;		
+		
 		drawcursor = ((svga->ma == svga->ca) && svga->con && svga->cursoron);
 
 		if (svga->crtc[0x17] & 0x80) {
-			chr  = svga->vram[(svga->ma << 1) & svga->vram_display_mask];
-			attr = svga->vram[((svga->ma << 1) + 1) & svga->vram_display_mask];
+			chr  = svga->vram[addr];
+			attr = svga->vram[addr+1];
 		} else
 			chr = attr = 0;
 
@@ -263,11 +268,12 @@ svga_render_text_80_ksc5601(svga_t *svga)
 	xinc = (svga->seqregs[1] & 1) ? 8 : 9;
 
 	for (x = 0; x < (svga->hdisp + svga->scrollcache); x += xinc) {
+		uint32_t addr = svga->remap_func(svga, svga->ma) & svga->vram_display_mask;
 		drawcursor = ((svga->ma == svga->ca) && svga->con && svga->cursoron);
-		chr  = svga->vram[(svga->ma << 1) & svga->vram_display_mask];
-		nextchr = svga->vram[((svga->ma + 4) << 1) & svga->vram_display_mask];
+		chr  = svga->vram[addr];
+		nextchr = svga->vram[addr + 8];
 		if (svga->crtc[0x17] & 0x80)
-			attr = svga->vram[((svga->ma << 1) + 1) & svga->vram_display_mask];
+			attr = svga->vram[addr + 1];
 		else
 			attr = 0;
 
@@ -367,16 +373,17 @@ svga_render_text_80_ksc5601(svga_t *svga)
 void
 svga_render_2bpp_lowres(svga_t *svga)
 {
-    int changed_offset, x;
+    int x;
     uint8_t dat[2];
     uint32_t addr, *p;
+	uint32_t changed_addr;
 
     if ((svga->displine + svga->y_add) < 0)
 	return;
 
-    changed_offset = ((svga->ma << 1) + (svga->sc & ~svga->crtc[0x17] & 3) * 0x8000) >> 12;
+    changed_addr = svga->remap_func(svga, svga->ma);
 
-    if (svga->changedvram[changed_offset] || svga->changedvram[changed_offset + 1] || svga->fullchange) {
+    if (svga->changedvram[changed_addr >> 12] || svga->changedvram[(changed_addr >> 12) + 1] || svga->fullchange) {
 	p = &buffer32->line[svga->displine + svga->y_add][svga->x_add];
 
 	if (svga->firstline_draw == 2000)
@@ -384,24 +391,7 @@ svga_render_2bpp_lowres(svga_t *svga)
 	svga->lastline_draw = svga->displine;
 
 	for (x = 0; x <= (svga->hdisp + svga->scrollcache); x += 16) {
-		addr = svga->ma;
-
-		if (!(svga->crtc[0x17] & 0x40)) {
-			addr = (addr << 1) & svga->vram_mask;
-			addr &= ~7;
-
-			if ((svga->crtc[0x17] & 0x20) && (svga->ma & 0x20000))
-				addr |= 4;
-
-			if (!(svga->crtc[0x17] & 0x20) && (svga->ma & 0x8000))
-				addr |= 4;
-		}
-
-		if (!(svga->crtc[0x17] & 0x01))
-			addr = (addr & ~0x8000) | ((svga->sc & 1) ? 0x8000 : 0);
-
-		if (!(svga->crtc[0x17] & 0x02))
-			addr = (addr & ~0x10000) | ((svga->sc & 2) ? 0x10000 : 0);
+		addr = svga->remap_func(svga, svga->ma);
 
 		dat[0] = svga->vram[addr];
 		dat[1] = svga->vram[addr | 0x1];
@@ -433,16 +423,17 @@ svga_render_2bpp_lowres(svga_t *svga)
 void
 svga_render_2bpp_highres(svga_t *svga)
 {
-    int changed_offset, x;
+    int x;
     uint8_t dat[2];
     uint32_t addr, *p;
+	uint32_t changed_addr;
 
     if ((svga->displine + svga->y_add) < 0)
 	return;
 
-    changed_offset = ((svga->ma << 1) + (svga->sc & ~svga->crtc[0x17] & 3) * 0x8000) >> 12;
+    changed_addr = svga->remap_func(svga, svga->ma);
 
-    if (svga->changedvram[changed_offset] || svga->changedvram[changed_offset + 1] || svga->fullchange) {
+    if (svga->changedvram[changed_addr >> 12] || svga->changedvram[(changed_addr >> 12) + 1] || svga->fullchange) {
 	p = &buffer32->line[svga->displine + svga->y_add][svga->x_add];
 
 	if (svga->firstline_draw == 2000)
@@ -450,24 +441,7 @@ svga_render_2bpp_highres(svga_t *svga)
 	svga->lastline_draw = svga->displine;
 
 	for (x = 0; x <= (svga->hdisp + svga->scrollcache); x += 8) {
-		addr = svga->ma;
-
-		if (!(svga->crtc[0x17] & 0x40)) {
-			addr = (addr << 1) & svga->vram_mask;
-			addr &= ~7;
-
-			if ((svga->crtc[0x17] & 0x20) && (svga->ma & 0x20000))
-				addr |= 4;
-
-			if (!(svga->crtc[0x17] & 0x20) && (svga->ma & 0x8000))
-				addr |= 4;
-		}
-
-		if (!(svga->crtc[0x17] & 0x01))
-			addr = (addr & ~0x8000) | ((svga->sc & 1) ? 0x8000 : 0);
-
-		if (!(svga->crtc[0x17] & 0x02))
-			addr = (addr & ~0x10000) | ((svga->sc & 2) ? 0x10000 : 0);
+		addr = svga->remap_func(svga, svga->ma);
 
 		dat[0] = svga->vram[addr];
 		dat[1] = svga->vram[addr | 0x1];
@@ -496,6 +470,64 @@ svga_render_2bpp_highres(svga_t *svga)
 }
 
 
+void
+svga_render_2bpp_headland_highres(svga_t *svga)
+{
+    int x;
+    int oddeven;
+    uint32_t addr, *p;
+    uint8_t edat[4];
+    uint8_t dat;
+	uint32_t changed_addr;
+
+    if ((svga->displine + svga->y_add) < 0)
+	return;
+
+    changed_addr = svga->remap_func(svga, svga->ma);
+
+     if (svga->changedvram[changed_addr >> 12] || svga->changedvram[(changed_addr >> 12) + 1] || svga->fullchange) {
+	p = &buffer32->line[svga->displine + svga->y_add][svga->x_add];
+
+	if (svga->firstline_draw == 2000) 
+		svga->firstline_draw = svga->displine;
+	svga->lastline_draw = svga->displine;
+
+	for (x = 0; x <= (svga->hdisp + svga->scrollcache); x += 8) {
+		addr = svga->remap_func(svga, svga->ma);
+		oddeven = 0;
+
+		if (svga->seqregs[1] & 4) {
+			oddeven = (addr & 4) ? 1 : 0;
+			edat[0] = svga->vram[addr | oddeven];
+			edat[2] = svga->vram[addr | oddeven | 0x2];
+			edat[1] = edat[3] = 0;
+		} else {
+			*(uint32_t *)(&edat[0]) = *(uint32_t *)(&svga->vram[addr]);
+		}
+		svga->ma += 4;
+		svga->ma &= svga->vram_mask;
+
+		if (svga->crtc[0x17] & 0x80) {
+			dat = edatlookup[edat[0] >> 6][edat[1] >> 6] | (edatlookup[edat[2] >> 6][edat[3] >> 6] << 2);
+			p[0] = svga->pallook[svga->egapal[(dat >> 4) & svga->plane_mask]];
+			p[1] = svga->pallook[svga->egapal[dat & svga->plane_mask]];
+			dat = edatlookup[(edat[0] >> 4) & 3][(edat[1] >> 4) & 3] | (edatlookup[(edat[2] >> 4) & 3][(edat[3] >> 4) & 3] << 2);
+			p[2] = svga->pallook[svga->egapal[(dat >> 4) & svga->plane_mask]];
+			p[3] = svga->pallook[svga->egapal[dat & svga->plane_mask]];
+			dat = edatlookup[(edat[0] >> 2) & 3][(edat[1] >> 2) & 3] | (edatlookup[(edat[2] >> 2) & 3][(edat[3] >> 2) & 3] << 2);
+			p[4] = svga->pallook[svga->egapal[(dat >> 4) & svga->plane_mask]];
+			p[5] = svga->pallook[svga->egapal[dat & svga->plane_mask]];
+			dat = edatlookup[edat[0] & 3][edat[1] & 3] | (edatlookup[edat[2] & 3][edat[3] & 3] << 2);
+			p[6] = svga->pallook[svga->egapal[(dat >> 4) & svga->plane_mask]];
+			p[7] = svga->pallook[svga->egapal[dat & svga->plane_mask]];
+		} else
+			memset(p, 0x00, 8 * sizeof(uint32_t));
+
+		p += 8;
+	}
+    }
+}
+
 void
 svga_render_4bpp_lowres(svga_t *svga)
 {
@@ -503,11 +535,14 @@ svga_render_4bpp_lowres(svga_t *svga)
     uint32_t addr, *p;
     uint8_t edat[4];
     uint8_t dat;
+	uint32_t changed_addr;
 
     if ((svga->displine + svga->y_add) < 0)
-	return;
+		return;
 
-    if (svga->changedvram[svga->ma >> 12] || svga->changedvram[(svga->ma >> 12) + 1] || svga->fullchange) {
+	changed_addr = svga->remap_func(svga, svga->ma);
+
+    if (svga->changedvram[changed_addr >> 12] || svga->changedvram[(changed_addr >> 12) + 1] || svga->fullchange) {
 	p = &buffer32->line[svga->displine + svga->y_add][svga->x_add];
 
 	if (svga->firstline_draw == 2000) 
@@ -515,29 +550,11 @@ svga_render_4bpp_lowres(svga_t *svga)
 	svga->lastline_draw = svga->displine;
 
 	for (x = 0; x <= (svga->hdisp + svga->scrollcache); x += 16) {
-		addr = svga->ma;
+		addr = svga->remap_func(svga, svga->ma);
 		oddeven = 0;
 
-		if (!(svga->crtc[0x17] & 0x40)) {
-			addr = (addr << 1) & svga->vram_mask;
-
-			if (svga->seqregs[1] & 4)
-				oddeven = (addr & 4) ? 1 : 0;
-
-			addr &= ~7;
-
-			if ((svga->crtc[0x17] & 0x20) && (svga->ma & 0x20000))
-				addr |= 4;
-			if (!(svga->crtc[0x17] & 0x20) && (svga->ma & 0x8000))
-				addr |= 4;
-		}
-
-		if (!(svga->crtc[0x17] & 0x01))
-			addr = (addr & ~0x8000) | ((svga->sc & 1) ? 0x8000 : 0);
-		if (!(svga->crtc[0x17] & 0x02))
-			addr = (addr & ~0x10000) | ((svga->sc & 2) ? 0x10000 : 0);
-
 		if (svga->seqregs[1] & 4) {
+			oddeven = (addr & 4) ? 1 : 0;
 			edat[0] = svga->vram[addr | oddeven];
 			edat[2] = svga->vram[addr | oddeven | 0x2];
 				edat[1] = edat[3] = 0;
@@ -573,18 +590,18 @@ svga_render_4bpp_lowres(svga_t *svga)
 void
 svga_render_4bpp_highres(svga_t *svga)
 {
-    int changed_offset, x;
-    int oddeven;
+    int x, oddeven;
     uint32_t addr, *p;
     uint8_t edat[4];
     uint8_t dat;
+	uint32_t changed_addr;
 
     if ((svga->displine + svga->y_add) < 0)
-	return;
+		return;
 
-    changed_offset = (svga->ma + (svga->sc & ~svga->crtc[0x17] & 3) * 0x8000) >> 12;
+	changed_addr = svga->remap_func(svga, svga->ma);
 
-    if (svga->changedvram[changed_offset] || svga->changedvram[changed_offset + 1] || svga->fullchange) {
+    if (svga->changedvram[changed_addr >> 12] || svga->changedvram[(changed_addr >> 12) + 1] || svga->fullchange) {
 	p = &buffer32->line[svga->displine + svga->y_add][svga->x_add];
 
 	if (svga->firstline_draw == 2000) 
@@ -592,29 +609,11 @@ svga_render_4bpp_highres(svga_t *svga)
 	svga->lastline_draw = svga->displine;
 
 	for (x = 0; x <= (svga->hdisp + svga->scrollcache); x += 8) {
-		addr = svga->ma;
+		addr = svga->remap_func(svga, svga->ma);
 		oddeven = 0;
 
-		if (!(svga->crtc[0x17] & 0x40)) {
-			addr = (addr << 1) & svga->vram_mask;
-
-			if (svga->seqregs[1] & 4)
-				oddeven = (addr & 4) ? 1 : 0;
-
-			addr &= ~7;
-
-			if ((svga->crtc[0x17] & 0x20) && (svga->ma & 0x20000))
-				addr |= 4;
-			if (!(svga->crtc[0x17] & 0x20) && (svga->ma & 0x8000))
-				addr |= 4;
-		}
-
-		if (!(svga->crtc[0x17] & 0x01))
-			addr = (addr & ~0x8000) | ((svga->sc & 1) ? 0x8000 : 0);
-		if (!(svga->crtc[0x17] & 0x02))
-			addr = (addr & ~0x10000) | ((svga->sc & 2) ? 0x10000 : 0);
-
 		if (svga->seqregs[1] & 4) {
+			oddeven = (addr & 4) ? 1 : 0;
 			edat[0] = svga->vram[addr | oddeven];
 			edat[2] = svga->vram[addr | oddeven | 0x2];
 				edat[1] = edat[3] = 0;
@@ -653,27 +652,44 @@ svga_render_8bpp_lowres(svga_t *svga)
     int x;
     uint32_t *p;
     uint32_t dat;
+	uint32_t changed_addr;
+	uint32_t addr;
 
     if ((svga->displine + svga->y_add) < 0)
-	return;
+		return;
 
-    if (svga->changedvram[svga->ma >> 12] || svga->changedvram[(svga->ma >> 12) + 1] || svga->fullchange) {
+	changed_addr = svga->remap_func(svga, svga->ma);
+
+    if (svga->changedvram[changed_addr >> 12] || svga->changedvram[(changed_addr >> 12) + 1] || svga->fullchange) {
 	p = &buffer32->line[svga->displine + svga->y_add][svga->x_add];
 
 	if (svga->firstline_draw == 2000) 
 		svga->firstline_draw = svga->displine;
 	svga->lastline_draw = svga->displine;
+	
+	if (!svga->remap_required) {
+		for (x = 0; x <= (svga->hdisp + svga->scrollcache); x += 8) {
+			dat = *(uint32_t *)(&svga->vram[svga->ma & svga->vram_display_mask]);
+			p[0] = p[1] = svga->map8[dat & 0xff];
+			p[2] = p[3] = svga->map8[(dat >> 8) & 0xff];
+			p[4] = p[5] = svga->map8[(dat >> 16) & 0xff];
+			p[6] = p[7] = svga->map8[(dat >> 24) & 0xff];
 
-	for (x = 0; x <= (svga->hdisp + svga->scrollcache); x += 8) {
-		dat = *(uint32_t *)(&svga->vram[svga->ma & svga->vram_display_mask]);
+			svga->ma += 4;
+			p += 8;
+		}
+	} else {
+		for (x = 0; x <= (svga->hdisp + svga->scrollcache); x += 8) {
+			addr = svga->remap_func(svga, svga->ma);
+			dat = *(uint32_t *)(&svga->vram[addr & svga->vram_display_mask]);
+			p[0] = p[1] = svga->map8[dat & 0xff];
+			p[2] = p[3] = svga->map8[(dat >> 8) & 0xff];
+			p[4] = p[5] = svga->map8[(dat >> 16) & 0xff];
+			p[6] = p[7] = svga->map8[(dat >> 24) & 0xff];
 
-		p[0] = p[1] = svga->map8[dat & 0xff];
-		p[2] = p[3] = svga->map8[(dat >> 8) & 0xff];
-		p[4] = p[5] = svga->map8[(dat >> 16) & 0xff];
-		p[6] = p[7] = svga->map8[(dat >> 24) & 0xff];
-
-		svga->ma += 4;
-		p += 8;
+			svga->ma += 4;
+			p += 8;
+		}
 	}
 	svga->ma &= svga->vram_display_mask;
     }
@@ -686,38 +702,55 @@ svga_render_8bpp_highres(svga_t *svga)
     int x;
     uint32_t *p;
     uint32_t dat;
+	uint32_t changed_addr;
+	uint32_t addr;
 
     if ((svga->displine + svga->y_add) < 0)
-	return;
+		return;
 
-    if (svga->changedvram[svga->ma >> 12] || svga->changedvram[(svga->ma >> 12) + 1] || svga->fullchange) {
+	changed_addr = svga->remap_func(svga, svga->ma);
+
+    if (svga->changedvram[changed_addr >> 12] || svga->changedvram[(changed_addr >> 12) + 1] || svga->fullchange) {
 	p = &buffer32->line[svga->displine + svga->y_add][svga->x_add];
 
 	if (svga->firstline_draw == 2000) 
 		svga->firstline_draw = svga->displine;
 	svga->lastline_draw = svga->displine;
+	
+	if (!svga->remap_required) {
+		for (x = 0; x <= (svga->hdisp/* + svga->scrollcache*/); x += 8) {
+			dat = *(uint32_t *)(&svga->vram[svga->ma & svga->vram_display_mask]);
+			p[0] = svga->map8[dat & 0xff];
+			p[1] = svga->map8[(dat >> 8) & 0xff];
+			p[2] = svga->map8[(dat >> 16) & 0xff];
+			p[3] = svga->map8[(dat >> 24) & 0xff];
 
-	for (x = 0; x <= (svga->hdisp/* + svga->scrollcache*/); x += 8) {
-		dat = *(uint32_t *)(&svga->vram[svga->ma & svga->vram_display_mask]);
-		p[0] = svga->map8[dat & 0xff];
-		p[1] = svga->map8[(dat >> 8) & 0xff];
-		p[2] = svga->map8[(dat >> 16) & 0xff];
-		p[3] = svga->map8[(dat >> 24) & 0xff];
+			dat = *(uint32_t *)(&svga->vram[(svga->ma + 4) & svga->vram_display_mask]);
+			p[4] = svga->map8[dat & 0xff];
+			p[5] = svga->map8[(dat >> 8) & 0xff];
+			p[6] = svga->map8[(dat >> 16) & 0xff];
+			p[7] = svga->map8[(dat >> 24) & 0xff];
 
-		dat = *(uint32_t *)(&svga->vram[(svga->ma + 4) & svga->vram_display_mask]);
-		p[4] = svga->map8[dat & 0xff];
-		p[5] = svga->map8[(dat >> 8) & 0xff];
-		p[6] = svga->map8[(dat >> 16) & 0xff];
-		p[7] = svga->map8[(dat >> 24) & 0xff];
+			svga->ma += 8;
+			p += 8;
+		}
+	} else {
+		for (x = 0; x <= (svga->hdisp/* + svga->scrollcache*/); x += 4) {
+			addr = svga->remap_func(svga, svga->ma);
+			dat = *(uint32_t *)(&svga->vram[addr & svga->vram_display_mask]);
+			p[0] = svga->map8[dat & 0xff];
+			p[1] = svga->map8[(dat >> 8) & 0xff];
+			p[2] = svga->map8[(dat >> 16) & 0xff];
+			p[3] = svga->map8[(dat >> 24) & 0xff];
 
-		svga->ma += 8;
-		p += 8;
+			svga->ma += 4;
+			p += 4;
+		}
 	}
 	svga->ma &= svga->vram_display_mask;
     }
 }
 
-
 void
 svga_render_8bpp_tseng_lowres(svga_t *svga)
 {
@@ -833,32 +866,49 @@ svga_render_15bpp_lowres(svga_t *svga)
     int x;
     uint32_t *p;
     uint32_t dat;
+	uint32_t changed_addr, addr;
 
     if ((svga->displine + svga->y_add) < 0)
-	return;
+		return;
 
-    if (svga->changedvram[svga->ma >> 12] || svga->changedvram[(svga->ma >> 12) + 1] || svga->fullchange) {
+	changed_addr = svga->remap_func(svga, svga->ma);
+
+    if (svga->changedvram[changed_addr >> 12] || svga->changedvram[(changed_addr >> 12) + 1] || svga->fullchange) {
 	p = &buffer32->line[svga->displine + svga->y_add][svga->x_add];
 
 	if (svga->firstline_draw == 2000) 
 		svga->firstline_draw = svga->displine;
 	svga->lastline_draw = svga->displine;
+	
+	if (!svga->remap_required) {
+		for (x = 0; x <= (svga->hdisp + svga->scrollcache); x += 4) {
+			if (svga->crtc[0x17] & 0x80) {
+				dat = *(uint32_t *)(&svga->vram[(svga->ma + (x << 1)) & svga->vram_display_mask]);
 
-	for (x = 0; x <= (svga->hdisp + svga->scrollcache); x += 4) {
-		if (svga->crtc[0x17] & 0x80) {
-			dat = *(uint32_t *)(&svga->vram[(svga->ma + (x << 1)) & svga->vram_display_mask]);
+				*p++ = video_15to32[dat & 0xffff];
+				*p++ = video_15to32[dat >> 16];
 
-			p[(x << 1)]     = p[(x << 1) + 1] = video_15to32[dat & 0xffff];
-			p[(x << 1) + 2] = p[(x << 1) + 3] = video_15to32[dat >> 16];
+				dat = *(uint32_t *)(&svga->vram[(svga->ma + (x << 1) + 4) & svga->vram_display_mask]);
 
-			dat = *(uint32_t *)(&svga->vram[(svga->ma + (x << 1) + 4) & svga->vram_display_mask]);
+				*p++ = video_15to32[dat & 0xffff];
+				*p++ = video_15to32[dat >> 16];
+			} else
+				memset(&(p[(x << 1)]), 0x00, 8 * sizeof(uint32_t));
+		}
+		svga->ma += x << 1;
+	} else {
+		for (x = 0; x <= (svga->hdisp + svga->scrollcache); x += 2) {
+			if (svga->crtc[0x17] & 0x80) {
+				addr = svga->remap_func(svga, svga->ma);
+				dat = *(uint32_t *)(&svga->vram[addr & svga->vram_display_mask]);
 
-			p[(x << 1) + 4] = p[(x << 1) + 5] = video_15to32[dat & 0xffff];
-			p[(x << 1) + 6] = p[(x << 1) + 7] = video_15to32[dat >> 16];
-		} else
-			memset(&(p[(x << 1)]), 0x00, 8 * sizeof(uint32_t));
+				*p++ = video_15to32[dat & 0xffff];
+				*p++ = video_15to32[dat >> 16];
+			} else
+				memset(&(p[x]), 0x00, 2 * sizeof(uint32_t));
+		}
+		svga->ma += 4;
 	}
-	svga->ma += x << 1;
 	svga->ma &= svga->vram_display_mask;
     }
 }
@@ -870,38 +920,55 @@ svga_render_15bpp_highres(svga_t *svga)
     int x;
     uint32_t *p;
     uint32_t dat;
+	uint32_t changed_addr, addr;
 
     if ((svga->displine + svga->y_add) < 0)
-	return;
+		return;
 
-    if (svga->changedvram[svga->ma >> 12] || svga->changedvram[(svga->ma >> 12) + 1] || svga->fullchange) {
+	changed_addr = svga->remap_func(svga, svga->ma);
+
+    if (svga->changedvram[changed_addr >> 12] || svga->changedvram[(changed_addr >> 12) + 1] || svga->fullchange) {
 	p = &buffer32->line[svga->displine + svga->y_add][svga->x_add];
 
 	if (svga->firstline_draw == 2000) 
 		svga->firstline_draw = svga->displine;
 	svga->lastline_draw = svga->displine;
+	
+	if (!svga->remap_required) {
+		for (x = 0; x <= (svga->hdisp + svga->scrollcache); x += 8) {
+			if (svga->crtc[0x17] & 0x80) {
+				dat = *(uint32_t *)(&svga->vram[(svga->ma + (x << 1)) & svga->vram_display_mask]);
+				*p++ = video_15to32[dat & 0xffff];
+				*p++ = video_15to32[dat >> 16];
 
-	for (x = 0; x <= (svga->hdisp + svga->scrollcache); x += 8) {
-		if (svga->crtc[0x17] & 0x80) {
-			dat = *(uint32_t *)(&svga->vram[(svga->ma + (x << 1)) & svga->vram_display_mask]);
-			p[x]     = video_15to32[dat & 0xffff];
-			p[x + 1] = video_15to32[dat >> 16];
+				dat = *(uint32_t *)(&svga->vram[(svga->ma + (x << 1) + 4) & svga->vram_display_mask]);
+				*p++ = video_15to32[dat & 0xffff];
+				*p++ = video_15to32[dat >> 16];
 
-			dat = *(uint32_t *)(&svga->vram[(svga->ma + (x << 1) + 4) & svga->vram_display_mask]);
-			p[x + 2] = video_15to32[dat & 0xffff];
-			p[x + 3] = video_15to32[dat >> 16];
+				dat = *(uint32_t *)(&svga->vram[(svga->ma + (x << 1) + 8) & svga->vram_display_mask]);
+				*p++ = video_15to32[dat & 0xffff];
+				*p++ = video_15to32[dat >> 16];
 
-			dat = *(uint32_t *)(&svga->vram[(svga->ma + (x << 1) + 8) & svga->vram_display_mask]);
-			p[x + 4] = video_15to32[dat & 0xffff];
-			p[x + 5] = video_15to32[dat >> 16];
+				dat = *(uint32_t *)(&svga->vram[(svga->ma + (x << 1) + 12) & svga->vram_display_mask]);
+				*p++ = video_15to32[dat & 0xffff];
+				*p++ = video_15to32[dat >> 16];
+			} else
+				memset(&(p[x]), 0x00, 8 * sizeof(uint32_t));
+		}
+		svga->ma += x << 1;
+	} else {
+		for (x = 0; x <= (svga->hdisp + svga->scrollcache); x += 2) {
+			if (svga->crtc[0x17] & 0x80) {
+				addr = svga->remap_func(svga, svga->ma);
+				dat = *(uint32_t *)(&svga->vram[addr & svga->vram_display_mask]);
 
-			dat = *(uint32_t *)(&svga->vram[(svga->ma + (x << 1) + 12) & svga->vram_display_mask]);
-			p[x + 6] = video_15to32[dat & 0xffff];
-			p[x + 7] = video_15to32[dat >> 16];
-		} else
-			memset(&(p[x]), 0x00, 8 * sizeof(uint32_t));
+				*p++ = video_15to32[dat & 0xffff];
+				*p++ = video_15to32[dat >> 16];
+			} else
+				memset(&(p[x]), 0x00, 2 * sizeof(uint32_t));
+		}
+		svga->ma += 4;
 	}
-	svga->ma += x << 1; 
 	svga->ma &= svga->vram_display_mask;
     }
 }
@@ -999,30 +1066,49 @@ svga_render_16bpp_lowres(svga_t *svga)
     int x;
     uint32_t *p;
     uint32_t dat;
+	uint32_t changed_addr, addr;
 
     if ((svga->displine + svga->y_add) < 0)
-	return;
+		return;
 
-    if (svga->changedvram[svga->ma >> 12] || svga->changedvram[(svga->ma >> 12) + 1] || svga->fullchange) {
+	changed_addr = svga->remap_func(svga, svga->ma);
+
+    if (svga->changedvram[changed_addr >> 12] || svga->changedvram[(changed_addr >> 12) + 1] || svga->fullchange) {
 	p = &buffer32->line[svga->displine + svga->y_add][svga->x_add];
 
 	if (svga->firstline_draw == 2000)
 		svga->firstline_draw = svga->displine;
 	svga->lastline_draw = svga->displine;
+	
+	if (!svga->remap_required) {
+		for (x = 0; x <= (svga->hdisp + svga->scrollcache); x += 4) {
+			if (svga->crtc[0x17] & 0x80) {
+				dat = *(uint32_t *)(&svga->vram[(svga->ma + (x << 1)) & svga->vram_display_mask]);
 
-	for (x = 0; x <= (svga->hdisp + svga->scrollcache); x += 4) {
-		if (svga->crtc[0x17] & 0x80) {
-			dat = *(uint32_t *)(&svga->vram[(svga->ma + (x << 1)) & svga->vram_display_mask]);
-			p[(x << 1)]     = p[(x << 1) + 1] = video_16to32[dat & 0xffff];
-			p[(x << 1) + 2] = p[(x << 1) + 3] = video_16to32[dat >> 16];
+				*p++ = video_16to32[dat & 0xffff];
+				*p++ = video_16to32[dat >> 16];
 
-			dat = *(uint32_t *)(&svga->vram[(svga->ma + (x << 1) + 4) & svga->vram_display_mask]);
-			p[(x << 1) + 4] = p[(x << 1) + 5] = video_16to32[dat & 0xffff];
-			p[(x << 1) + 6] = p[(x << 1) + 7] = video_16to32[dat >> 16];
-		} else
-			memset(&(p[(x << 1)]), 0x00, 8 * sizeof(uint32_t));
+				dat = *(uint32_t *)(&svga->vram[(svga->ma + (x << 1) + 4) & svga->vram_display_mask]);
+
+				*p++ = video_16to32[dat & 0xffff];
+				*p++ = video_16to32[dat >> 16];
+			} else
+				memset(&(p[(x << 1)]), 0x00, 8 * sizeof(uint32_t));
+		}
+		svga->ma += x << 1;
+	} else {
+		for (x = 0; x <= (svga->hdisp + svga->scrollcache); x += 2) {
+			if (svga->crtc[0x17] & 0x80) {
+			addr = svga->remap_func(svga, svga->ma);
+			dat = *(uint32_t *)(&svga->vram[addr & svga->vram_display_mask]);
+
+			*p++ = video_16to32[dat & 0xffff];
+			*p++ = video_16to32[dat >> 16];
+			} else
+				memset(&(p[x]), 0x00, 2 * sizeof(uint32_t));
+		}
+		svga->ma += 4;
 	}
-	svga->ma += x << 1; 
 	svga->ma &= svga->vram_display_mask;
     }
 }
@@ -1033,38 +1119,57 @@ svga_render_16bpp_highres(svga_t *svga)
 {
     int x;
     uint32_t *p;
+	uint32_t dat;
+	uint32_t changed_addr, addr;
 
     if ((svga->displine + svga->y_add) < 0)
-	return;
+		return;
 
-    if (svga->changedvram[svga->ma >> 12] || svga->changedvram[(svga->ma >> 12) + 1] || svga->fullchange) {
+	changed_addr = svga->remap_func(svga, svga->ma);
+
+    if (svga->changedvram[changed_addr >> 12] || svga->changedvram[(changed_addr >> 12) + 1] || svga->fullchange) {
 	p = &buffer32->line[svga->displine + svga->y_add][svga->x_add];
 
 	if (svga->firstline_draw == 2000) 
 		svga->firstline_draw = svga->displine;
 	svga->lastline_draw = svga->displine;
+	
+	if (!svga->remap_required) {
+		for (x = 0; x <= (svga->hdisp + svga->scrollcache); x += 8) {
+			if (svga->crtc[0x17] & 0x80) {
+				dat = *(uint32_t *)(&svga->vram[(svga->ma + (x << 1)) & svga->vram_display_mask]);
+				*p++ = video_16to32[dat & 0xffff];
+				*p++ = video_16to32[dat >> 16];
 
-	for (x = 0; x <= (svga->hdisp + svga->scrollcache); x += 8) {
-		if (svga->crtc[0x17] & 0x80) {
-			uint32_t dat = *(uint32_t *)(&svga->vram[(svga->ma + (x << 1)) & svga->vram_display_mask]);
-			p[x]     = video_16to32[dat & 0xffff];
-			p[x + 1] = video_16to32[dat >> 16];
+				dat = *(uint32_t *)(&svga->vram[(svga->ma + (x << 1) + 4) & svga->vram_display_mask]);
+				*p++ = video_16to32[dat & 0xffff];
+				*p++ = video_16to32[dat >> 16];
 
-			dat = *(uint32_t *)(&svga->vram[(svga->ma + (x << 1) + 4) & svga->vram_display_mask]);
-			p[x + 2] = video_16to32[dat & 0xffff];
-			p[x + 3] = video_16to32[dat >> 16];
+				dat = *(uint32_t *)(&svga->vram[(svga->ma + (x << 1) + 8) & svga->vram_display_mask]);
+				*p++ = video_16to32[dat & 0xffff];
+				*p++ = video_16to32[dat >> 16];
 
-			dat = *(uint32_t *)(&svga->vram[(svga->ma + (x << 1) + 8) & svga->vram_display_mask]);
-			p[x + 4] = video_16to32[dat & 0xffff];
-			p[x + 5] = video_16to32[dat >> 16];
+				dat = *(uint32_t *)(&svga->vram[(svga->ma + (x << 1) + 12) & svga->vram_display_mask]);
+				*p++ = video_16to32[dat & 0xffff];
+				*p++ = video_16to32[dat >> 16];
+			} else
+				memset(&(p[x]), 0x00, 8 * sizeof(uint32_t));
+		}
+		svga->ma += x << 1;
+	} else {
+		for (x = 0; x <= (svga->hdisp + svga->scrollcache); x += 2) {
+			if (svga->crtc[0x17] & 0x80) {
+				addr = svga->remap_func(svga, svga->ma);
+				dat = *(uint32_t *)(&svga->vram[addr & svga->vram_display_mask]);
 
-			dat = *(uint32_t *)(&svga->vram[(svga->ma + (x << 1) + 12) & svga->vram_display_mask]);
-			p[x + 6] = video_16to32[dat & 0xffff];
-			p[x + 7] = video_16to32[dat >> 16];
-		} else
-			memset(&(p[x]), 0x00, 8 * sizeof(uint32_t));
+				*p++ = video_16to32[dat & 0xffff];
+				*p++ = video_16to32[dat >> 16];
+			} else
+				memset(&(p[x]), 0x00, 2 * sizeof(uint32_t));
+			
+			svga->ma += 4;
+		}
 	}
-	svga->ma += x << 1; 
 	svga->ma &= svga->vram_display_mask;
     }
 }
@@ -1074,27 +1179,60 @@ void
 svga_render_24bpp_lowres(svga_t *svga)
 {
     int x;
-    uint32_t fg;
+    uint32_t *p;
+	uint32_t changed_addr, addr;
+	uint32_t dat0, dat1, dat2;
 
     if ((svga->displine + svga->y_add) < 0)
-	return;
+		return;
+
+	changed_addr = svga->remap_func(svga, svga->ma);
+	
+    if (svga->changedvram[changed_addr >> 12] || svga->changedvram[(changed_addr >> 12) + 1] || svga->fullchange) {
+	p = &buffer32->line[svga->displine + svga->y_add][svga->x_add];
 
-    if (svga->changedvram[svga->ma >> 12] || svga->changedvram[(svga->ma >> 12) + 1] || svga->fullchange) {
 	if (svga->firstline_draw == 2000) 
 		svga->firstline_draw = svga->displine;
 	svga->lastline_draw = svga->displine;
+	
+	if (!svga->remap_required) {
+		for (x = 0; x <= (svga->hdisp + svga->scrollcache); x++) {
+			if (svga->crtc[0x17] & 0x80) {
+				dat0 = *(uint32_t *)(&svga->vram[svga->ma & svga->vram_display_mask]);
+				dat1 = *(uint32_t *)(&svga->vram[(svga->ma + 4) & svga->vram_display_mask]);
+				dat2 = *(uint32_t *)(&svga->vram[(svga->ma + 8) & svga->vram_display_mask]);
+			} else
+				dat0 = dat1 = dat2 = 0x00000000;
 
-	for (x = 0; x <= (svga->hdisp + svga->scrollcache); x++) {
-		if (svga->crtc[0x17] & 0x80)
-			fg = svga->vram[svga->ma] | (svga->vram[svga->ma + 1] << 8) | (svga->vram[svga->ma + 2] << 16);
-		else
-			fg = 0x00000000;
-		svga->ma += 3; 
-		svga->ma &= svga->vram_display_mask;
-		buffer32->line[svga->displine + svga->y_add][(x << 1) + svga->x_add] =
-		buffer32->line[svga->displine + svga->y_add][(x << 1) + 1 + svga->x_add] = fg;
+			p[0] = p[1] = dat0 & 0xffffff;
+			p[2] = p[3] = (dat0 >> 24) | ((dat1 & 0xffff) << 8);
+			p[4] = p[5] = (dat1 >> 16) | ((dat2 & 0xff) << 16);
+			p[6] = p[7] = dat2 >> 8;
+
+			svga->ma += 12;
+		}
+    } else {
+		for (x = 0; x <= (svga->hdisp + svga->scrollcache); x += 4) {
+			if (svga->crtc[0x17] & 0x80) {
+				addr = svga->remap_func(svga, svga->ma);
+				dat0 = *(uint32_t *)(&svga->vram[addr & svga->vram_display_mask]);
+				addr = svga->remap_func(svga, svga->ma + 4);
+				dat1 = *(uint32_t *)(&svga->vram[addr & svga->vram_display_mask]);
+				addr = svga->remap_func(svga, svga->ma + 8);
+				dat2 = *(uint32_t *)(&svga->vram[addr & svga->vram_display_mask]);
+			} else
+				dat0 = dat1 = dat2 = 0x00000000;
+
+			p[0] = p[1] = dat0 & 0xffffff;
+			p[2] = p[3] = (dat0 >> 24) | ((dat1 & 0xffff) << 8);
+			p[4] = p[5] = (dat1 >> 16) | ((dat2 & 0xff) << 16);
+			p[6] = p[7] = dat2 >> 8;			
+			
+			svga->ma += 12;
+		}		
+	}
+	svga->ma &= svga->vram_display_mask;
 	}
-    }
 }
 
 
@@ -1103,35 +1241,56 @@ svga_render_24bpp_highres(svga_t *svga)
 {
     int x;
     uint32_t *p;
-    uint32_t dat;
+	uint32_t changed_addr, addr;
+	uint32_t dat0, dat1, dat2;
 
     if ((svga->displine + svga->y_add) < 0)
-	return;
+		return;
 
-    if (svga->changedvram[svga->ma >> 12] || svga->changedvram[(svga->ma >> 12) + 1] || svga->fullchange) {
+	changed_addr = svga->remap_func(svga, svga->ma);
+
+    if (svga->changedvram[changed_addr >> 12] || svga->changedvram[(changed_addr >> 12) + 1] || svga->fullchange) {
 	p = &buffer32->line[svga->displine + svga->y_add][svga->x_add];
 
 	if (svga->firstline_draw == 2000) 
 		svga->firstline_draw = svga->displine;
 	svga->lastline_draw = svga->displine;
+	
+	if (!svga->remap_required) {
+		for (x = 0; x <= (svga->hdisp + svga->scrollcache); x += 4) {
+			if (svga->crtc[0x17] & 0x80) {
+				dat0 = *(uint32_t *)(&svga->vram[svga->ma & svga->vram_display_mask]);
+				dat1 = *(uint32_t *)(&svga->vram[(svga->ma + 4) & svga->vram_display_mask]);
+				dat2 = *(uint32_t *)(&svga->vram[(svga->ma + 8) & svga->vram_display_mask]);
 
-	for (x = 0; x <= (svga->hdisp + svga->scrollcache); x += 4) {
-		if (svga->crtc[0x17] & 0x80) {
-			dat = *(uint32_t *)(&svga->vram[svga->ma & svga->vram_display_mask]);
-			p[x] = dat & 0xffffff;
+				*p++ = dat0 & 0xffffff;
+				*p++ = (dat0 >> 24) | ((dat1 & 0xffff) << 8);
+				*p++ = (dat1 >> 16) | ((dat2 & 0xff) << 16);
+				*p++ = dat2 >> 8;
+			} else
+				memset(&(p[x]), 0x0, 4 * sizeof(uint32_t));
 
-			dat = *(uint32_t *)(&svga->vram[(svga->ma + 3) & svga->vram_display_mask]);
-			p[x + 1] = dat & 0xffffff;
+			svga->ma += 12;
+		}
+	} else {
+		for (x = 0; x <= (svga->hdisp + svga->scrollcache); x += 4) {
+			if (svga->crtc[0x17] & 0x80) {
+				addr = svga->remap_func(svga, svga->ma);
+				dat0 = *(uint32_t *)(&svga->vram[addr & svga->vram_display_mask]);
+				addr = svga->remap_func(svga, svga->ma + 4);
+				dat1 = *(uint32_t *)(&svga->vram[addr & svga->vram_display_mask]);
+				addr = svga->remap_func(svga, svga->ma + 8);
+				dat2 = *(uint32_t *)(&svga->vram[addr & svga->vram_display_mask]);
 
-			dat = *(uint32_t *)(&svga->vram[(svga->ma + 6) & svga->vram_display_mask]);
-			p[x + 2] = dat & 0xffffff;
+				*p++ = dat0 & 0xffffff;
+				*p++ = (dat0 >> 24) | ((dat1 & 0xffff) << 8);
+				*p++ = (dat1 >> 16) | ((dat2 & 0xff) << 16);
+				*p++ = dat2 >> 8;
+			} else
+				memset(&(p[x]), 0x0, 4 * sizeof(uint32_t));
 
-			dat = *(uint32_t *)(&svga->vram[(svga->ma + 9) & svga->vram_display_mask]);
-			p[x + 3] = dat & 0xffffff;
-		} else
-			memset(&(p[x]), 0x0, 4 * sizeof(uint32_t));
-
-		svga->ma += 12;
+			svga->ma += 12;
+		}		
 	}
 	svga->ma &= svga->vram_display_mask;
     }
@@ -1142,25 +1301,44 @@ void
 svga_render_32bpp_lowres(svga_t *svga)
 {
     int x;
-    uint32_t fg;
+	uint32_t *p;
+    uint32_t dat;
+	uint32_t changed_addr, addr;
 
     if ((svga->displine + svga->y_add) < 0)
-	return;
+		return;
 
-    if (svga->changedvram[svga->ma >> 12] || svga->changedvram[(svga->ma >> 12) + 1] || svga->fullchange) {
+	changed_addr = svga->remap_func(svga, svga->ma);
+
+    if (svga->changedvram[changed_addr >> 12] || svga->changedvram[(changed_addr >> 12) + 1] || svga->fullchange) {
+	p = &buffer32->line[svga->displine + svga->y_add][svga->x_add];	
+		
 	if (svga->firstline_draw == 2000) 
 		svga->firstline_draw = svga->displine;
 	svga->lastline_draw = svga->displine;
-
-	for (x = 0; x <= (svga->hdisp + svga->scrollcache); x++) {
-		if (svga->crtc[0x17] & 0x80)
-			fg = svga->vram[svga->ma] | (svga->vram[svga->ma + 1] << 8) | (svga->vram[svga->ma + 2] << 16);
-		else
-			fg = 0x00000000;
-		svga->ma += 4;
+	
+	if (!svga->remap_required) {
+		for (x = 0; x <= (svga->hdisp + svga->scrollcache); x++) {
+			if (svga->crtc[0x17] & 0x80)
+				dat = *(uint32_t *)(&svga->vram[(svga->ma + (x << 2)) & svga->vram_display_mask]);
+			else
+				dat = 0x00000000;
+			*p++ = dat & 0xffffff;
+			*p++ = dat & 0xffffff;
+		}
+		svga->ma += (x * 4);
+	} else {
+		for (x = 0; x <= (svga->hdisp + svga->scrollcache); x++) {
+			if (svga->crtc[0x17] & 0x80) {
+				addr = svga->remap_func(svga, svga->ma);
+				dat = *(uint32_t *)(&svga->vram[addr & svga->vram_display_mask]);
+			} else
+				dat = 0x00000000;
+			*p++ = dat & 0xffffff;
+			*p++ = dat & 0xffffff;
+			svga->ma += 4;
+		}
 		svga->ma &= svga->vram_display_mask;
-		buffer32->line[svga->displine + svga->y_add][(x << 1) + svga->x_add] =
-		buffer32->line[svga->displine + svga->y_add][(x << 1) + 1 + svga->x_add] = fg;
 	}
     }
 }
@@ -1172,25 +1350,41 @@ svga_render_32bpp_highres(svga_t *svga)
     int x;
     uint32_t *p;
     uint32_t dat;
+	uint32_t changed_addr, addr;
 
     if ((svga->displine + svga->y_add) < 0)
-	return;
+		return;
 
-    if (svga->changedvram[svga->ma >> 12] ||  svga->changedvram[(svga->ma >> 12) + 1] || svga->changedvram[(svga->ma >> 12) + 2] || svga->fullchange) {
+	changed_addr = svga->remap_func(svga, svga->ma);
+
+    if (svga->changedvram[changed_addr >> 12] || svga->changedvram[(changed_addr >> 12) + 1] || svga->fullchange) {
 	p = &buffer32->line[svga->displine + svga->y_add][svga->x_add];
         
 	if (svga->firstline_draw == 2000) 
 		svga->firstline_draw = svga->displine;
 	svga->lastline_draw = svga->displine;
+	
+	if (!svga->remap_required) {
+		for (x = 0; x <= (svga->hdisp + svga->scrollcache); x++) {
+			if (svga->crtc[0x17] & 0x80) {
+				dat = *(uint32_t *)(&svga->vram[(svga->ma + (x << 2)) & svga->vram_display_mask]);
+				*p++ = dat & 0xffffff;
+			} else
+				memset(&(p[x]), 0x0, 1 * sizeof(uint32_t));
+		}
+		svga->ma += (x * 4);
+	} else {
+		for (x = 0; x <= (svga->hdisp + svga->scrollcache); x++) {
+			if (svga->crtc[0x17] & 0x80) {
+				addr = svga->remap_func(svga, svga->ma);
+				dat = *(uint32_t *)(&svga->vram[addr & svga->vram_display_mask]);
+				*p++ = dat & 0xffffff;
+			} else
+				memset(&(p[x]), 0x0, 1 * sizeof(uint32_t));
 
-	for (x = 0; x <= (svga->hdisp + svga->scrollcache); x++) {
-		if (svga->crtc[0x17] & 0x80)
-			dat = *(uint32_t *)(&svga->vram[(svga->ma + (x << 2)) & svga->vram_display_mask]);
-		else
-			dat = 0x00000000;
-		p[x] = dat & 0xffffff;
+			svga->ma += 4;
+		}
 	}
-	svga->ma += 4; 
 	svga->ma &= svga->vram_display_mask;
     }
 }
@@ -1202,25 +1396,41 @@ svga_render_ABGR8888_highres(svga_t *svga)
     int x;
     uint32_t *p;
     uint32_t dat;
+	uint32_t changed_addr, addr;
 
     if ((svga->displine + svga->y_add) < 0)
-	return;
+		return;
 
-    if (svga->changedvram[svga->ma >> 12] ||  svga->changedvram[(svga->ma >> 12) + 1] || svga->changedvram[(svga->ma >> 12) + 2] || svga->fullchange) {
+	changed_addr = svga->remap_func(svga, svga->ma);
+
+    if (svga->changedvram[changed_addr >> 12] || svga->changedvram[(changed_addr >> 12) + 1] || svga->fullchange) {
 	p = &buffer32->line[svga->displine + svga->y_add][svga->x_add];
         
 	if (svga->firstline_draw == 2000) 
 		svga->firstline_draw = svga->displine;
 	svga->lastline_draw = svga->displine;
-
-	for (x = 0; x <= (svga->hdisp + svga->scrollcache); x++) {
-		if (svga->crtc[0x17] & 0x80)
-			dat = *(uint32_t *)(&svga->vram[(svga->ma + (x << 2)) & svga->vram_display_mask]);
-		else
-			dat = 0x00000000;
-		p[x] = ((dat & 0xff0000) >> 16) | (dat & 0x00ff00) | ((dat & 0x0000ff) << 16);
+	
+	if (!svga->remap_required) {
+		for (x = 0; x <= (svga->hdisp + svga->scrollcache); x++) {
+			if (svga->crtc[0x17] & 0x80) {
+				dat = *(uint32_t *)(&svga->vram[(svga->ma + (x << 2)) & svga->vram_display_mask]);
+				*p++ = ((dat & 0xff0000) >> 16) | (dat & 0x00ff00) | ((dat & 0x0000ff) << 16);
+			} else
+				memset(&(p[x]), 0x0, 1 * sizeof(uint32_t));
+		}
+		svga->ma += x*4;
+	} else {
+		for (x = 0; x <= (svga->hdisp + svga->scrollcache); x++) {
+			if (svga->crtc[0x17] & 0x80) {
+				addr = svga->remap_func(svga, svga->ma);
+				dat = *(uint32_t *)(&svga->vram[addr & svga->vram_display_mask]);
+				*p++ = ((dat & 0xff0000) >> 16) | (dat & 0x00ff00) | ((dat & 0x0000ff) << 16);
+			} else
+				memset(&(p[x]), 0x0, 1 * sizeof(uint32_t));
+			
+			svga->ma += 4;
+		}	
 	}
-	svga->ma += 4; 
 	svga->ma &= svga->vram_display_mask;
     }
 }
@@ -1232,25 +1442,41 @@ svga_render_RGBA8888_highres(svga_t *svga)
     int x;
     uint32_t *p;
     uint32_t dat;
+	uint32_t changed_addr, addr;
 
     if ((svga->displine + svga->y_add) < 0)
-	return;
+		return;
 
-    if (svga->changedvram[svga->ma >> 12] ||  svga->changedvram[(svga->ma >> 12) + 1] || svga->changedvram[(svga->ma >> 12) + 2] || svga->fullchange) {
+	changed_addr = svga->remap_func(svga, svga->ma);
+
+    if (svga->changedvram[changed_addr >> 12] || svga->changedvram[(changed_addr >> 12) + 1] || svga->fullchange) {
 	p = &buffer32->line[svga->displine + svga->y_add][svga->x_add];
         
 	if (svga->firstline_draw == 2000) 
 		svga->firstline_draw = svga->displine;
 	svga->lastline_draw = svga->displine;
+	
+	if (!svga->remap_required) {
+		for (x = 0; x <= (svga->hdisp + svga->scrollcache); x++) {
+			if (svga->crtc[0x17] & 0x80) {
+				dat = *(uint32_t *)(&svga->vram[(svga->ma + (x << 2)) & svga->vram_display_mask]);
+				*p++ = dat >> 8;
+			} else
+				memset(&(p[x]), 0x0, 1 * sizeof(uint32_t));
+		}
+		svga->ma += (x * 4);
+	} else {
+		for (x = 0; x <= (svga->hdisp + svga->scrollcache); x++) {
+			if (svga->crtc[0x17] & 0x80) {
+				addr = svga->remap_func(svga, svga->ma);
+				dat = *(uint32_t *)(&svga->vram[addr & svga->vram_display_mask]);
+				*p++ = dat >> 8;
+			} else
+				memset(&(p[x]), 0x0, 1 * sizeof(uint32_t));
 
-	for (x = 0; x <= (svga->hdisp + svga->scrollcache); x++) {
-		if (svga->crtc[0x17] & 0x80)
-			dat = *(uint32_t *)(&svga->vram[(svga->ma + (x << 2)) & svga->vram_display_mask]);
-		else
-			dat = 0x00000000;
-		p[x] = dat >> 8;
+			svga->ma += 4;
+		}
 	}
-	svga->ma += 4; 
 	svga->ma &= svga->vram_display_mask;
     }
 }
diff --git a/src/video/vid_tgui9440.c b/src/video/vid_tgui9440.c
index 9dd5fe081..9d2840ed7 100644
--- a/src/video/vid_tgui9440.c
+++ b/src/video/vid_tgui9440.c
@@ -70,8 +70,6 @@
 #include "cpu.h"
 #include <86box/plat.h>
 #include <86box/video.h>
-#include <86box/i2c.h>
-#include <86box/vid_ddc.h>
 #include <86box/vid_svga.h>
 #include <86box/vid_svga_render.h>
 
@@ -107,45 +105,19 @@
   but the Windows 3.1 driver always reads bytes and write words of 0xffff.*/  
 
 #define ROM_TGUI_9400CXI	"roms/video/tgui9440/9400CXI.vbi"
-#define ROM_TGUI_9440		"roms/video/tgui9440/9440.vbi"
+#define ROM_TGUI_9440		"roms/video/tgui9440/9440.VBI"
 
 #define EXT_CTRL_16BIT            0x01
 #define EXT_CTRL_MONO_EXPANSION   0x02
 #define EXT_CTRL_MONO_TRANSPARENT 0x04
 #define EXT_CTRL_LATCH_COPY       0x08
 
-#define FIFO_SIZE 65536
-#define FIFO_MASK (FIFO_SIZE - 1)
-#define FIFO_ENTRY_SIZE (1 << 31)
-
-#define FIFO_ENTRIES (tgui->fifo_write_idx - tgui->fifo_read_idx)
-#define FIFO_FULL    ((tgui->fifo_write_idx - tgui->fifo_read_idx) >= FIFO_SIZE)
-#define FIFO_EMPTY   (tgui->fifo_read_idx == tgui->fifo_write_idx)
-
-#define FIFO_TYPE 0xff000000
-#define FIFO_ADDR 0x00ffffff
-
 enum
 {
         TGUI_9400CXI = 0,
         TGUI_9440
 };
 
-enum
-{
-        FIFO_INVALID       = (0x00 << 24),
-        FIFO_WRITE_BYTE    = (0x01 << 24),
-        FIFO_WRITE_FB_BYTE = (0x04 << 24),
-        FIFO_WRITE_FB_WORD = (0x05 << 24),
-        FIFO_WRITE_FB_LONG = (0x06 << 24)
-};
-
-typedef struct
-{
-        uint32_t addr_type;
-        uint32_t val;
-} fifo_entry_t;
-
 typedef struct tgui_t
 {
         mem_mapping_t linear_mapping;
@@ -177,6 +149,7 @@ typedef struct tgui_t
         	int use_src;
 	
         	int pitch, bpp;
+			int val_access;
 
                 uint16_t tgui_pattern[8][8];
         } accel;
@@ -184,7 +157,7 @@ typedef struct tgui_t
         uint8_t ext_gdc_regs[16]; /*TGUI9400CXi only*/
         uint8_t copy_latch[16];
 
-        uint8_t tgui_3d8, tgui_3d9;
+        uint8_t tgui_3d8, tgui_3d9, port3CD;
         int oldmode;
         uint8_t oldctrl1;
         uint8_t oldctrl2,newctrl2;
@@ -197,21 +170,8 @@ typedef struct tgui_t
         int clock_m, clock_n, clock_k;
         
         uint32_t vram_size, vram_mask;
-
-        fifo_entry_t fifo[FIFO_SIZE];
-        volatile int fifo_read_idx, fifo_write_idx;
-
-        thread_t *fifo_thread;
-        event_t *wake_fifo_thread;
-        event_t *fifo_not_full_event;
-        
-        int blitter_busy;
-        uint64_t blitter_time;
-        uint64_t status_time;
         
         volatile int write_blitter;
-
-        void *i2c, *ddc;
 } tgui_t;
 
 video_timings_t timing_tgui_vlb = {VIDEO_BUS, 4,  8, 16,   4,  8, 16};
@@ -219,8 +179,6 @@ video_timings_t timing_tgui_pci = {VIDEO_PCI, 4,  8, 16,   4,  8, 16};
 
 void tgui_recalcmapping(tgui_t *tgui);
 
-static void fifo_thread(void *param);
-
 uint8_t tgui_accel_read(uint32_t addr, void *priv);
 uint16_t tgui_accel_read_w(uint32_t addr, void *priv);
 uint32_t tgui_accel_read_l(uint32_t addr, void *priv);
@@ -243,6 +201,23 @@ static void tgui_ext_write(uint32_t addr, uint8_t val, void *p);
 static void tgui_ext_writew(uint32_t addr, uint16_t val, void *p);
 static void tgui_ext_writel(uint32_t addr, uint32_t val, void *p);
 
+
+/*Remap address for chain-4/doubleword style layout*/
+static __inline uint32_t
+dword_remap(uint32_t in_addr)
+{
+        return ((in_addr << 2) & 0x3fff0) |
+                ((in_addr >> 14) & 0xc) |
+                (in_addr & ~0x3fffc);
+}
+static __inline uint32_t
+dword_remap_w(uint32_t in_addr)
+{
+        return ((in_addr << 2) & 0x1fff8) |
+                ((in_addr >> 14) & 0x6) |
+                (in_addr & ~0x1fffe);
+}
+
 void tgui_out(uint16_t addr, uint8_t val, void *p)
 {
         tgui_t *tgui = (tgui_t *)p;
@@ -309,9 +284,11 @@ void tgui_out(uint16_t addr, uint8_t val, void *p)
                                 svga->bpp = 8;
                                 break;
                         }
+						svga_recalctimings(svga);
                         return;
                 }
-		/*FALLTHROUGH*/
+				break;
+				
                 case 0x3C7: case 0x3C8: case 0x3C9:
                 if (tgui->type == TGUI_9400CXI)
                 {
@@ -319,7 +296,7 @@ void tgui_out(uint16_t addr, uint8_t val, void *p)
                         return;
                 }
                 tgui->ramdac_state = 0;
-		break;
+				break;
 
                 case 0x3CF:
                 if (svga->gdcaddr == 0x23)
@@ -384,16 +361,12 @@ void tgui_out(uint16_t addr, uint8_t val, void *p)
                                 {
                	                        tgui->linear_base = ((val & 0xf) | ((val >> 2) & 0x30)) << 20;
                        	                tgui->linear_size = (val & 0x10) ? 0x200000 : 0x100000;
-                                        tgui->svga.decode_mask = (val & 0x10) ? 0x1fffff : 0xfffff;
+                                        svga->decode_mask = (val & 0x10) ? 0x1fffff : 0xfffff;
                                 }
         			tgui_recalcmapping(tgui);
                         }
 			break;
 
-			case 0x37:
-			i2c_gpio_set(tgui->i2c, (val & 0x02) || !(val & 0x04), (val & 0x01) || !(val & 0x08));
-			break;
-
 			case 0x40: case 0x41: case 0x42: case 0x43:
 			case 0x44: case 0x45: case 0x46: case 0x47:
                         if (tgui->type >= TGUI_9440)
@@ -462,7 +435,7 @@ uint8_t tgui_in(uint16_t addr, void *p)
                                 case TGUI_9400CXI:
                                 return 0x93; /*TGUI9400CXi*/
                                 case TGUI_9440:
-                                return 0xe3; /*TGUI9440AGi*/
+                                return 0xa3; /*TGUI9440AGi*/
                         }
                 }
                 if ((svga->seqaddr & 0xf) == 0xd)
@@ -479,16 +452,19 @@ uint8_t tgui_in(uint16_t addr, void *p)
                 break;
                 case 0x3C6:
                 if (tgui->type == TGUI_9400CXI)
-                        return tkd8001_ramdac_in(addr, svga->ramdac, svga);
+					return tkd8001_ramdac_in(addr, svga->ramdac, svga);
                 if (tgui->ramdac_state == 4)
-                        return tgui->ramdac_ctrl;
-                tgui->ramdac_state++;
+					return tgui->ramdac_ctrl;
+				tgui->ramdac_state++;
                 break;
                 case 0x3C7: case 0x3C8: case 0x3C9:
                 if (tgui->type == TGUI_9400CXI)
                         return tkd8001_ramdac_in(addr, svga->ramdac, svga);
                 tgui->ramdac_state = 0;
                 break;
+				case 0x3CD:
+				return tgui->port3CD;				
+				
                 case 0x3CF:
                 if (tgui->type == TGUI_9400CXI && svga->gdcaddr >= 16 && svga->gdcaddr < 32)
                         return tgui->ext_gdc_regs[svga->gdcaddr & 15];
@@ -497,18 +473,6 @@ uint8_t tgui_in(uint16_t addr, void *p)
                 return svga->crtcreg;
                 case 0x3D5:
                 temp = svga->crtc[svga->crtcreg];
-                if (svga->crtcreg == 0x37) {
-                        if (!(svga->crtc[0x37] & 0x04)) {
-                                temp &= 0xfd;
-                                if (i2c_gpio_get_scl(tgui->i2c))
-                                        temp |= 0x02;
-                        }
-                        if (!(svga->crtc[0x37] & 0x08)) {
-                                temp &= 0xfe;
-                                if (i2c_gpio_get_sda(tgui->i2c))
-                                        temp |= 0x01;
-                        }
-                }
                 return temp;
                 case 0x3d8:
 		return tgui->tgui_3d8;
@@ -647,6 +611,7 @@ void tgui_recalcmapping(tgui_t *tgui)
 	if (svga->crtc[0x21] & 0x20)
 	{
                 mem_mapping_disable(&svga->mapping);
+				pclog("Linear size = %08x, linear base = %08x\n", tgui->linear_size, tgui->linear_base);
                 mem_mapping_set_addr(&tgui->linear_mapping, tgui->linear_base, tgui->linear_size);
                 if (tgui->type >= TGUI_9440)
                 {
@@ -688,7 +653,7 @@ void tgui_recalcmapping(tgui_t *tgui)
                         break;
                         case 0x4: /*64k at A0000*/
                         mem_mapping_set_addr(&svga->mapping, 0xa0000, 0x10000);
-                        mem_mapping_enable(&tgui->accel_mapping);
+						mem_mapping_enable(&tgui->accel_mapping);
                         svga->banked_mask = 0xffff;
                         break;
                         case 0x8: /*32k at B0000*/
@@ -711,12 +676,15 @@ void tgui_hwcursor_draw(svga_t *svga, int displine)
 	int pitch = (svga->hwcursor.xsize == 64) ? 16 : 8;
 	int byte, bit;
 	uint32_t color;
+	uint32_t remapped_addr;
         
         if (svga->interlace && svga->hwcursor_oddeven)
-                svga->hwcursor_latch.addr += pitch;
+			svga->hwcursor_latch.addr += pitch;
+
+		remapped_addr = dword_remap(svga->hwcursor_latch.addr);
 
         for (xx = 0; xx < svga->hwcursor.xsize; xx++) {
-		byte = svga->hwcursor_latch.addr + (xx >> 3);
+		byte = remapped_addr + (xx >> 3);
 		bit = (7 - (xx & 7));
 		dat[0] = (svga->vram[byte] >> bit) & 0x01;			/* AND */
 		dat[1] = (svga->vram[(pitch >> 1) + byte] >> bit) & 0x01;	/* XOR */
@@ -805,14 +773,14 @@ void tgui_pci_write(int func, int addr, uint8_t val, void *p)
                 case 0x12:
                 tgui->linear_base = (tgui->linear_base & 0xff000000) | ((val & 0xe0) << 16);
                 tgui->linear_size = 2 << 20;
-                tgui->svga.decode_mask = 0x1fffff;
+				svga->decode_mask = 0x1fffff;
                 svga->crtc[0x21] = (svga->crtc[0x21] & ~0xf) | (val >> 4);
                 tgui_recalcmapping(tgui);
                 break;
                 case 0x13:
                 tgui->linear_base = (tgui->linear_base & 0xe00000) | (val << 24);
                 tgui->linear_size = 2 << 20;
-                tgui->svga.decode_mask = 0x1fffff;
+				svga->decode_mask = 0x1fffff;
                 svga->crtc[0x21] = (svga->crtc[0x21] & ~0xc0) | (val >> 6);
                 tgui_recalcmapping(tgui);
                 break;
@@ -832,9 +800,13 @@ static uint8_t tgui_ext_linear_read(uint32_t addr, void *p)
                 return 0xff;
         
         addr &= ~0xf;
-        for (c = 0; c < 16; c++)
-                tgui->copy_latch[c] = svga->vram[addr+c];
+		addr = dword_remap(addr);
 
+        for (c = 0; c < 16; c++) {
+                tgui->copy_latch[c] = svga->vram[addr+c];
+				addr += ((c & 3) == 3) ? 13 : 1;
+		}
+		
         return svga->vram[addr & svga->vram_mask]; 
 }
 
@@ -862,7 +834,9 @@ static void tgui_ext_linear_write(uint32_t addr, uint8_t val, void *p)
         if (addr >= svga->vram_max)
                 return;
         addr &= svga->vram_mask;
-        addr &= ~0x7;
+        addr &= (tgui->ext_gdc_regs[0] & 8) ? ~0xf : ~0x7;
+
+        addr = dword_remap(addr);
         svga->changedvram[addr >> 12] = changeframecount;
         
         switch (tgui->ext_gdc_regs[0] & 0xf)
@@ -873,7 +847,7 @@ static void tgui_ext_linear_write(uint32_t addr, uint8_t val, void *p)
                 {
                         if (mask & (1 << c))
                                 *(uint8_t *)&svga->vram[addr] = (val & (1 << c)) ? fg[0] : bg[0];
-                        addr++;
+                        addr += (c == 4) ? 13 : 1;
                 }
                 break;
 
@@ -883,7 +857,7 @@ static void tgui_ext_linear_write(uint32_t addr, uint8_t val, void *p)
                 {
                         if (mask & (1 << c))
                                 *(uint8_t *)&svga->vram[addr] = (val & (1 << c)) ? fg[(c & 1) ^ 1] : bg[(c & 1) ^ 1];
-                        addr++;
+                        addr += (c == 4) ? 13 : 1;
                 }
                 break;
 
@@ -893,7 +867,7 @@ static void tgui_ext_linear_write(uint32_t addr, uint8_t val, void *p)
                 {
                         if ((val & mask) & (1 << c))
                                 *(uint8_t *)&svga->vram[addr] = fg[0];
-                        addr++;
+                        addr += (c == 4) ? 13 : 1;
                 }
                 break;
                 
@@ -903,15 +877,16 @@ static void tgui_ext_linear_write(uint32_t addr, uint8_t val, void *p)
                 {
                         if ((val & mask) & (1 << c))
                                 *(uint8_t *)&svga->vram[addr] = fg[(c & 1) ^ 1];
-                        addr++;
+						addr += (c == 4) ? 13 : 1;
                 }
                 break;
 
                 case 0x8: case 0x9: case 0xa: case 0xb:
                 case 0xc: case 0xd: case 0xe: case 0xf:
-                addr &= ~0xf;
-                for (c = 0; c < 16; c++)
-                        *(uint8_t *)&svga->vram[addr+c] = tgui->copy_latch[c];
+                for (c = 0; c < 16; c++) {
+                        *(uint8_t *)&svga->vram[addr] = tgui->copy_latch[c];
+                        addr += ((c & 3) == 3) ? 13 : 1;
+				}
                 break;
         }
 }
@@ -932,6 +907,8 @@ static void tgui_ext_linear_writew(uint32_t addr, uint16_t val, void *p)
                 return;
         addr &= svga->vram_mask;
         addr &= ~0xf;
+		
+		addr = dword_remap(addr);
         svga->changedvram[addr >> 12] = changeframecount;
         
         val = (val >> 8) | (val << 8);
@@ -944,7 +921,7 @@ static void tgui_ext_linear_writew(uint32_t addr, uint16_t val, void *p)
                 {
                         if (mask & (1 << c))
                                 *(uint8_t *)&svga->vram[addr] = (val & (1 << c)) ? fg[0] : bg[0];
-                        addr++;
+                        addr += (c & 3) ? 1 : 13;
                 }
                 break;
 
@@ -954,7 +931,7 @@ static void tgui_ext_linear_writew(uint32_t addr, uint16_t val, void *p)
                 {
                         if (mask & (1 << c))
                                 *(uint8_t *)&svga->vram[addr] = (val & (1 << c)) ? fg[(c & 1) ^ 1] : bg[(c & 1) ^ 1];
-                        addr++;
+                        addr += (c & 3) ? 1 : 13;
                 }
                 break;
 
@@ -964,7 +941,7 @@ static void tgui_ext_linear_writew(uint32_t addr, uint16_t val, void *p)
                 {
                         if ((val & mask) & (1 << c))
                                 *(uint8_t *)&svga->vram[addr] = fg[0];
-                        addr++;
+                        addr += (c & 3) ? 1 : 13;
                 }
                 break;
 
@@ -974,14 +951,16 @@ static void tgui_ext_linear_writew(uint32_t addr, uint16_t val, void *p)
                 {
                         if ((val & mask) & (1 << c))
                                 *(uint8_t *)&svga->vram[addr] = fg[(c & 1) ^ 1];
-                        addr++;
+                        addr += (c & 3) ? 1 : 13;
                 }
                 break;
                                 
                 case 0x8: case 0x9: case 0xa: case 0xb:
                 case 0xc: case 0xd: case 0xe: case 0xf:
-                for (c = 0; c < 16; c++)
+                for (c = 0; c < 16; c++) {
                         *(uint8_t *)&svga->vram[addr+c] = tgui->copy_latch[c];
+						addr += ((c & 3) == 3) ? 13 : 1;
+				}
                 break;
         }
 }
@@ -1034,8 +1013,8 @@ enum
 	TGUI_SOLIDFILL = 0x4000	/*Pattern all zero?*/
 };
 
-#define READ(addr, dat) if (tgui->accel.bpp == 0) dat = svga->vram[addr & 0x1fffff]; \
-                        else                     dat = vram_w[addr & 0xfffff];
+#define READ(addr, dat) if (tgui->accel.bpp == 0) dat = svga->vram[dword_remap(addr) & tgui->vram_mask]; \
+                        else dat = vram_w[dword_remap_w(addr) & (tgui->vram_mask >> 1)];
                         
 #define MIX() do \
 	{								\
@@ -1051,13 +1030,13 @@ enum
 
 #define WRITE(addr, dat)        if (tgui->accel.bpp == 0)                                                \
                                 {                                                                       \
-                                        svga->vram[addr & 0x1fffff] = dat;                                    \
-                                        svga->changedvram[((addr) & 0x1fffff) >> 12] = changeframecount;      \
+                                        svga->vram[dword_remap(addr) & tgui->vram_mask] = dat;                                    \
+                                        svga->changedvram[(dword_remap(addr) & tgui->vram_mask) >> 12] = changeframecount;      \
                                 }                                                                       \
                                 else                                                                    \
                                 {                                                                       \
-                                        vram_w[addr & 0xfffff] = dat;                                   \
-                                        svga->changedvram[((addr) & 0xfffff) >> 11] = changeframecount;        \
+                                        vram_w[dword_remap_w(addr) & (tgui->vram_mask >> 1)] = dat;                                   \
+                                        svga->changedvram[(dword_remap_w(addr) & (tgui->vram_mask >> 1)) >> 11] = changeframecount;        \
                                 }
                                 
 void tgui_accel_command(int count, uint32_t cpu_dat, tgui_t *tgui)
@@ -1127,6 +1106,8 @@ void tgui_accel_command(int count, uint32_t cpu_dat, tgui_t *tgui)
                         }
 		}
 	}
+	
+		pclog("Accel command = %i\n", tgui->accel.command);
         switch (tgui->accel.command)
 	{
 		case TGUI_BITBLT:
@@ -1137,7 +1118,7 @@ void tgui_accel_command(int count, uint32_t cpu_dat, tgui_t *tgui)
 			tgui->accel.pat_x = tgui->accel.dst_x;
 			tgui->accel.pat_y = tgui->accel.dst_y;
 		}
-
+		
 		switch (tgui->accel.flags & (TGUI_SRCMONO|TGUI_SRCDISP))
 		{
 			case TGUI_SRCCPU:
@@ -1149,19 +1130,19 @@ void tgui_accel_command(int count, uint32_t cpu_dat, tgui_t *tgui)
                                 }
 				if (tgui->accel.use_src)
                                         return;
+			} else {
+				count >>= 3;
 			}
-			else
-			     count >>= 3;
 			while (count)
 			{
 				if (tgui->accel.bpp == 0)
 				{
-                                        src_dat = cpu_dat >> 24;
+										src_dat = cpu_dat >> 24;
                                         cpu_dat <<= 8;
                                 }
                                 else
                                 {
-                                        src_dat = (cpu_dat >> 24) | ((cpu_dat >> 8) & 0xff00);
+										src_dat = (cpu_dat >> 24) | ((cpu_dat >> 8) & 0xff00);
                                         cpu_dat <<= 16;
                                         count--;
                                 }
@@ -1215,7 +1196,7 @@ void tgui_accel_command(int count, uint32_t cpu_dat, tgui_t *tgui)
 				if (tgui->accel.use_src)
                                         return;
 			}
-			while (count)
+			while (count--)
 			{
 				src_dat = ((cpu_dat >> 31) ? tgui->accel.fg_col : tgui->accel.bg_col);
 				if (tgui->accel.bpp == 0)
@@ -1258,12 +1239,11 @@ void tgui_accel_command(int count, uint32_t cpu_dat, tgui_t *tgui)
 					if (tgui->accel.use_src)
                                                 return;
 				}
-				count--;
 			}
 			break;
 
 			default:
-			while (count)
+			while (count--)
 			{
 				READ(tgui->accel.src, src_dat);
 				READ(tgui->accel.dst, dst_dat);                                                                
@@ -1295,7 +1275,6 @@ void tgui_accel_command(int count, uint32_t cpu_dat, tgui_t *tgui)
 					if (tgui->accel.y > tgui->accel.size_y)
 						return;
 				}
-				count--;
 			}
 			break;
 		}
@@ -1303,16 +1282,268 @@ void tgui_accel_command(int count, uint32_t cpu_dat, tgui_t *tgui)
 	}
 }
 
-static void tgui_accel_write_fifo(tgui_t *tgui, uint32_t addr, uint8_t val)
+void tgui_accel_out(uint16_t addr, uint8_t val, void *p)
 {
+        tgui_t *tgui = (tgui_t *)p;
+
+	switch (addr)
+	{
+		case 0x2122:
+		tgui->accel.ger22 = val;
+		tgui->accel.pitch = 512 << ((val >> 2) & 3);
+		tgui->accel.bpp = (val & 3) ? 1 : 0;
+		tgui->accel.pitch >>= tgui->accel.bpp;
+		break;
+                
+		case 0x2124: /*Command*/
+		tgui->accel.command = val;
+		tgui_accel_command(-1, 0, tgui);
+		break;
+		
+		case 0x2127: /*ROP*/
+		tgui->accel.rop = val;
+		tgui->accel.use_src = (val & 0x33) ^ ((val >> 2) & 0x33);
+		break;
+		
+		case 0x2128: /*Flags*/
+		tgui->accel.flags = (tgui->accel.flags & 0xff00) | val;
+		break;
+		case 0x2129: /*Flags*/
+		tgui->accel.flags = (tgui->accel.flags & 0xff) | (val << 8);
+		break;
+
+		case 0x212b:
+		tgui->accel.offset = val & 7;
+		break;
+		
+		case 0x212c: /*Foreground colour*/
+		tgui->accel.fg_col = (tgui->accel.fg_col & 0xff00) | val;
+		break;
+		case 0x212d: /*Foreground colour*/
+		tgui->accel.fg_col = (tgui->accel.fg_col & 0xff) | (val << 8);
+		break;
+
+		case 0x2130: /*Background colour*/
+		tgui->accel.bg_col = (tgui->accel.bg_col & 0xff00) | val;
+		break;
+		case 0x2131: /*Background colour*/
+		tgui->accel.bg_col = (tgui->accel.bg_col & 0xff) | (val << 8);
+		break;
+
+		case 0x2138: /*Dest X*/
+		tgui->accel.dst_x = (tgui->accel.dst_x & 0xff00) | val;
+		break;
+		case 0x2139: /*Dest X*/
+		tgui->accel.dst_x = (tgui->accel.dst_x & 0xff) | (val << 8);
+		break;
+		case 0x213a: /*Dest Y*/
+		tgui->accel.dst_y = (tgui->accel.dst_y & 0xff00) | val;
+		break;
+		case 0x213b: /*Dest Y*/
+		tgui->accel.dst_y = (tgui->accel.dst_y & 0xff) | (val << 8);
+		break;
+
+		case 0x213c: /*Src X*/
+		tgui->accel.src_x = (tgui->accel.src_x & 0xff00) | val;
+		break;
+		case 0x213d: /*Src X*/
+		tgui->accel.src_x = (tgui->accel.src_x & 0xff) | (val << 8);
+		break;
+		case 0x213e: /*Src Y*/
+		tgui->accel.src_y = (tgui->accel.src_y & 0xff00) | val;
+		break;
+		case 0x213f: /*Src Y*/
+		tgui->accel.src_y = (tgui->accel.src_y & 0xff) | (val << 8);
+		break;
+
+		case 0x2140: /*Size X*/
+		tgui->accel.size_x = (tgui->accel.size_x & 0xff00) | val;
+		break;
+		case 0x2141: /*Size X*/
+		tgui->accel.size_x = (tgui->accel.size_x & 0xff) | (val << 8);
+		break;
+		case 0x2142: /*Size Y*/
+		tgui->accel.size_y = (tgui->accel.size_y & 0xff00) | val;
+		break;
+		case 0x2143: /*Size Y*/
+		tgui->accel.size_y = (tgui->accel.size_y & 0xff) | (val << 8);
+		break;
+		
+		case 0x2180: case 0x2181: case 0x2182: case 0x2183:
+		case 0x2184: case 0x2185: case 0x2186: case 0x2187:
+		case 0x2188: case 0x2189: case 0x218a: case 0x218b:
+		case 0x218c: case 0x218d: case 0x218e: case 0x218f:
+		case 0x2190: case 0x2191: case 0x2192: case 0x2193:
+		case 0x2194: case 0x2195: case 0x2196: case 0x2197:
+		case 0x2198: case 0x2199: case 0x219a: case 0x219b:
+		case 0x219c: case 0x219d: case 0x219e: case 0x219f:
+		case 0x21a0: case 0x21a1: case 0x21a2: case 0x21a3:
+		case 0x21a4: case 0x21a5: case 0x21a6: case 0x21a7:
+		case 0x21a8: case 0x21a9: case 0x21aa: case 0x21ab:
+		case 0x21ac: case 0x21ad: case 0x21ae: case 0x21af:
+		case 0x21b0: case 0x21b1: case 0x21b2: case 0x21b3:
+		case 0x21b4: case 0x21b5: case 0x21b6: case 0x21b7:
+		case 0x21b8: case 0x21b9: case 0x21ba: case 0x21bb:
+		case 0x21bc: case 0x21bd: case 0x21be: case 0x21bf:
+		case 0x21c0: case 0x21c1: case 0x21c2: case 0x21c3:
+		case 0x21c4: case 0x21c5: case 0x21c6: case 0x21c7:
+		case 0x21c8: case 0x21c9: case 0x21ca: case 0x21cb:
+		case 0x21cc: case 0x21cd: case 0x21ce: case 0x21cf:
+		case 0x21d0: case 0x21d1: case 0x21d2: case 0x21d3:
+		case 0x21d4: case 0x21d5: case 0x21d6: case 0x21d7:
+		case 0x21d8: case 0x21d9: case 0x21da: case 0x21db:
+		case 0x21dc: case 0x21dd: case 0x21de: case 0x21df:
+		case 0x21e0: case 0x21e1: case 0x21e2: case 0x21e3:
+		case 0x21e4: case 0x21e5: case 0x21e6: case 0x21e7:
+		case 0x21e8: case 0x21e9: case 0x21ea: case 0x21eb:
+		case 0x21ec: case 0x21ed: case 0x21ee: case 0x21ef:
+		case 0x21f0: case 0x21f1: case 0x21f2: case 0x21f3:
+		case 0x21f4: case 0x21f5: case 0x21f6: case 0x21f7:
+		case 0x21f8: case 0x21f9: case 0x21fa: case 0x21fb:
+		case 0x21fc: case 0x21fd: case 0x21fe: case 0x21ff:
+		tgui->accel.pattern[addr & 0x7f] = val;
+		break;
+	}
+}
+
+void tgui_accel_out_w(uint16_t addr, uint16_t val, void *p)
+{
+        tgui_t *tgui = (tgui_t *)p;
+		tgui_accel_out(addr, val, tgui);
+		tgui_accel_out(addr + 1, val >> 8, tgui);
+}
+
+void tgui_accel_out_l(uint16_t addr, uint32_t val, void *p)
+{
+        tgui_t *tgui = (tgui_t *)p;
+		tgui_accel_out(addr, val, tgui);
+		tgui_accel_out(addr + 1, val >> 8, tgui);
+		tgui_accel_out(addr + 2, val >> 16, tgui);
+		tgui_accel_out(addr + 3, val >> 24, tgui);
+}
+
+uint8_t tgui_accel_in(uint16_t addr, void *p)
+{
+		tgui_t *tgui = (tgui_t *)p;
+
+	switch (addr)
+	{
+		case 0x2120: /*Status*/
+		return 0;
+		
+		case 0x2127: /*ROP*/
+		return tgui->accel.rop;
+		
+		case 0x2128: /*Flags*/
+		return tgui->accel.flags & 0xff;
+		case 0x2129: /*Flags*/
+		return tgui->accel.flags >> 8;
+
+		case 0x212b:
+		return tgui->accel.offset;
+		
+		case 0x212c: /*Background colour*/
+		return tgui->accel.bg_col & 0xff;
+		case 0x212d: /*Background colour*/
+		return tgui->accel.bg_col >> 8;
+
+		case 0x2130: /*Foreground colour*/
+		return tgui->accel.fg_col & 0xff;
+		case 0x2131: /*Foreground colour*/
+		return tgui->accel.fg_col >> 8;
+
+		case 0x2138: /*Dest X*/
+		return tgui->accel.dst_x & 0xff;
+		case 0x2139: /*Dest X*/
+		return tgui->accel.dst_x >> 8;
+		case 0x213a: /*Dest Y*/
+		return tgui->accel.dst_y & 0xff;
+		case 0x213b: /*Dest Y*/
+		return tgui->accel.dst_y >> 8;
+
+		case 0x213c: /*Src X*/
+		return tgui->accel.src_x & 0xff;
+		case 0x213d: /*Src X*/
+		return tgui->accel.src_x >> 8;
+		case 0x213e: /*Src Y*/
+		return tgui->accel.src_y & 0xff;
+		case 0x213f: /*Src Y*/
+		return tgui->accel.src_y >> 8;
+
+		case 0x2140: /*Size X*/
+		return tgui->accel.size_x & 0xff;
+		case 0x2141: /*Size X*/
+		return tgui->accel.size_x >> 8;
+		case 0x2142: /*Size Y*/
+		return tgui->accel.size_y & 0xff;
+		case 0x2143: /*Size Y*/
+		return tgui->accel.size_y >> 8;
+		
+		case 0x2180: case 0x2181: case 0x2182: case 0x2183:
+		case 0x2184: case 0x2185: case 0x2186: case 0x2187:
+		case 0x2188: case 0x2189: case 0x218a: case 0x218b:
+		case 0x218c: case 0x218d: case 0x218e: case 0x218f:
+		case 0x2190: case 0x2191: case 0x2192: case 0x2193:
+		case 0x2194: case 0x2195: case 0x2196: case 0x2197:
+		case 0x2198: case 0x2199: case 0x219a: case 0x219b:
+		case 0x219c: case 0x219d: case 0x219e: case 0x219f:
+		case 0x21a0: case 0x21a1: case 0x21a2: case 0x21a3:
+		case 0x21a4: case 0x21a5: case 0x21a6: case 0x21a7:
+		case 0x21a8: case 0x21a9: case 0x21aa: case 0x21ab:
+		case 0x21ac: case 0x21ad: case 0x21ae: case 0x21af:
+		case 0x21b0: case 0x21b1: case 0x21b2: case 0x21b3:
+		case 0x21b4: case 0x21b5: case 0x21b6: case 0x21b7:
+		case 0x21b8: case 0x21b9: case 0x21ba: case 0x21bb:
+		case 0x21bc: case 0x21bd: case 0x21be: case 0x21bf:
+		case 0x21c0: case 0x21c1: case 0x21c2: case 0x21c3:
+		case 0x21c4: case 0x21c5: case 0x21c6: case 0x21c7:
+		case 0x21c8: case 0x21c9: case 0x21ca: case 0x21cb:
+		case 0x21cc: case 0x21cd: case 0x21ce: case 0x21cf:
+		case 0x21d0: case 0x21d1: case 0x21d2: case 0x21d3:
+		case 0x21d4: case 0x21d5: case 0x21d6: case 0x21d7:
+		case 0x21d8: case 0x21d9: case 0x21da: case 0x21db:
+		case 0x21dc: case 0x21dd: case 0x21de: case 0x21df:
+		case 0x21e0: case 0x21e1: case 0x21e2: case 0x21e3:
+		case 0x21e4: case 0x21e5: case 0x21e6: case 0x21e7:
+		case 0x21e8: case 0x21e9: case 0x21ea: case 0x21eb:
+		case 0x21ec: case 0x21ed: case 0x21ee: case 0x21ef:
+		case 0x21f0: case 0x21f1: case 0x21f2: case 0x21f3:
+		case 0x21f4: case 0x21f5: case 0x21f6: case 0x21f7:
+		case 0x21f8: case 0x21f9: case 0x21fa: case 0x21fb:
+		case 0x21fc: case 0x21fd: case 0x21fe: case 0x21ff:
+		return tgui->accel.pattern[addr & 0x7f];
+	}
+	return 0;
+}
+
+uint16_t tgui_accel_in_w(uint16_t addr, void *p)
+{
+        tgui_t *tgui = (tgui_t *)p;
+	return tgui_accel_in(addr, tgui) | (tgui_accel_in(addr + 1, tgui) << 8);
+}
+
+uint32_t tgui_accel_in_l(uint16_t addr, void *p)
+{
+        tgui_t *tgui = (tgui_t *)p;
+	return tgui_accel_in_w(addr, tgui) | (tgui_accel_in_w(addr + 2, tgui) << 16);
+}
+
+
+void tgui_accel_write(uint32_t addr, uint8_t val, void *p)
+{
+        tgui_t *tgui = (tgui_t *)p;
+
+	if ((addr & ~0xff) != 0xbff00)
+		return;
+
 	switch (addr & 0xff)
 	{
-                case 0x22:
-                tgui->accel.ger22 = val;
-                tgui->accel.pitch = 512 << ((val >> 2) & 3);
-                tgui->accel.bpp = (val & 3) ? 1 : 0;
-                tgui->accel.pitch >>= tgui->accel.bpp;
-                break;
+		case 0x22:
+		tgui->accel.ger22 = val;
+		tgui->accel.pitch = 512 << ((val >> 2) & 3);
+		tgui->accel.bpp = (val & 3) ? 1 : 0;
+		tgui->accel.pitch >>= tgui->accel.bpp;
+		break;
                 
 		case 0x24: /*Command*/
 		tgui->accel.command = val;
@@ -1425,109 +1656,6 @@ static void tgui_accel_write_fifo(tgui_t *tgui, uint32_t addr, uint8_t val)
 	}
 }
 
-static void tgui_accel_write_fifo_fb_b(tgui_t *tgui, uint32_t addr, uint8_t val)
-{
-	tgui_accel_command(8, val << 24, tgui);
-}
-static void tgui_accel_write_fifo_fb_w(tgui_t *tgui, uint32_t addr, uint16_t val)
-{
-        tgui_accel_command(16, (((val & 0xff00) >> 8) | ((val & 0x00ff) << 8)) << 16, tgui);
-}
-static void tgui_accel_write_fifo_fb_l(tgui_t *tgui, uint32_t addr, uint32_t val)
-{
-	tgui_accel_command(32, ((val & 0xff000000) >> 24) | ((val & 0x00ff0000) >> 8) | ((val & 0x0000ff00) << 8) | ((val & 0x000000ff) << 24), tgui);
-}
-
-static void fifo_thread(void *param)
-{
-        tgui_t *tgui = (tgui_t *)param;
-        
-        while (1)
-        {
-                thread_set_event(tgui->fifo_not_full_event);
-                thread_wait_event(tgui->wake_fifo_thread, -1);
-                thread_reset_event(tgui->wake_fifo_thread);
-                tgui->blitter_busy = 1;
-                while (!FIFO_EMPTY)
-                {
-                        uint64_t start_time = plat_timer_read();
-                        uint64_t end_time;
-                        fifo_entry_t *fifo = &tgui->fifo[tgui->fifo_read_idx & FIFO_MASK];
-
-                        switch (fifo->addr_type & FIFO_TYPE)
-                        {
-                                case FIFO_WRITE_BYTE:
-                                tgui_accel_write_fifo(tgui, fifo->addr_type & FIFO_ADDR, fifo->val);
-                                break;
-                                case FIFO_WRITE_FB_BYTE:
-                                tgui_accel_write_fifo_fb_b(tgui, fifo->addr_type & FIFO_ADDR, fifo->val);
-                                break;
-                                case FIFO_WRITE_FB_WORD:
-                                tgui_accel_write_fifo_fb_w(tgui, fifo->addr_type & FIFO_ADDR, fifo->val);
-                                break;
-                                case FIFO_WRITE_FB_LONG:
-                                tgui_accel_write_fifo_fb_l(tgui, fifo->addr_type & FIFO_ADDR, fifo->val);
-                                break;
-                        }
-                                                
-                        tgui->fifo_read_idx++;
-                        fifo->addr_type = FIFO_INVALID;
-
-                        if (FIFO_ENTRIES > 0xe000)
-                                thread_set_event(tgui->fifo_not_full_event);
-
-                        end_time = plat_timer_read();
-                        tgui->blitter_time += end_time - start_time;
-                }
-                tgui->blitter_busy = 0;
-        }
-}
-
-static inline void wake_fifo_thread(tgui_t *tgui)
-{
-        thread_set_event(tgui->wake_fifo_thread); /*Wake up FIFO thread if moving from idle*/
-}
-
-static void tgui_wait_fifo_idle(tgui_t *tgui)
-{
-        while (!FIFO_EMPTY)
-        {
-                wake_fifo_thread(tgui);
-                thread_wait_event(tgui->fifo_not_full_event, 1);
-        }
-}
-
-static void tgui_queue(tgui_t *tgui, uint32_t addr, uint32_t val, uint32_t type)
-{
-        fifo_entry_t *fifo = &tgui->fifo[tgui->fifo_write_idx & FIFO_MASK];
-
-        if (FIFO_FULL)
-        {
-                thread_reset_event(tgui->fifo_not_full_event);
-                if (FIFO_FULL)
-                {
-                        thread_wait_event(tgui->fifo_not_full_event, -1); /*Wait for room in ringbuffer*/
-                }
-        }
-
-        fifo->val = val;
-        fifo->addr_type = (addr & FIFO_ADDR) | type;
-
-        tgui->fifo_write_idx++;
-        
-        if (FIFO_ENTRIES > 0xe000 || FIFO_ENTRIES < 8)
-                wake_fifo_thread(tgui);
-}
-
-
-void tgui_accel_write(uint32_t addr, uint8_t val, void *p)
-{
-        tgui_t *tgui = (tgui_t *)p;
-	if ((addr & ~0xff) != 0xbff00)
-		return;
-	tgui_queue(tgui, addr, val, FIFO_WRITE_BYTE);
-}
-
 void tgui_accel_write_w(uint32_t addr, uint16_t val, void *p)
 {
         tgui_t *tgui = (tgui_t *)p;
@@ -1549,13 +1677,9 @@ uint8_t tgui_accel_read(uint32_t addr, void *p)
         tgui_t *tgui = (tgui_t *)p;
 	if ((addr & ~0xff) != 0xbff00)
 		return 0xff;
-	if ((addr & 0xff) != 0x20)
-	       tgui_wait_fifo_idle(tgui);
 	switch (addr & 0xff)
 	{
 		case 0x20: /*Status*/
-		if (!FIFO_EMPTY)
-		      return 1 << 5;
 		return 0;
 		
 		case 0x27: /*ROP*/
@@ -1661,9 +1785,9 @@ void tgui_accel_write_fb_b(uint32_t addr, uint8_t val, void *p)
         tgui_t *tgui = (tgui_t *)svga->p;
 
         if (tgui->write_blitter)
-        	tgui_queue(tgui, addr, val, FIFO_WRITE_FB_BYTE);
+        	tgui_accel_command(8, val << 24, tgui);
         else
-                svga_write_linear(addr, val, svga);
+			svga_write_linear(addr, val, svga);
 }
 
 void tgui_accel_write_fb_w(uint32_t addr, uint16_t val, void *p)
@@ -1672,9 +1796,9 @@ void tgui_accel_write_fb_w(uint32_t addr, uint16_t val, void *p)
         tgui_t *tgui = (tgui_t *)svga->p;
 
         if (tgui->write_blitter)
-        	tgui_queue(tgui, addr, val, FIFO_WRITE_FB_WORD);
+        	tgui_accel_command(16, (((val & 0xff00) >> 8) | ((val & 0x00ff) << 8)) << 16, tgui);
         else
-                svga_writew_linear(addr, val, svga);
+			svga_writew_linear(addr, val, svga);
 }
 
 void tgui_accel_write_fb_l(uint32_t addr, uint32_t val, void *p)
@@ -1683,9 +1807,9 @@ void tgui_accel_write_fb_l(uint32_t addr, uint32_t val, void *p)
         tgui_t *tgui = (tgui_t *)svga->p;
 
         if (tgui->write_blitter)
-        	tgui_queue(tgui, addr, val, FIFO_WRITE_FB_LONG);
+        	tgui_accel_command(32, ((val & 0xff000000) >> 24) | ((val & 0x00ff0000) >> 8) | ((val & 0x0000ff00) << 8) | ((val & 0x000000ff) << 24), tgui);
         else
-                svga_writel_linear(addr, val, svga);
+            svga_writel_linear(addr, val, svga);
 }
 
 static void *tgui_init(const device_t *info)
@@ -1736,30 +1860,40 @@ static void *tgui_init(const device_t *info)
         mem_mapping_disable(&tgui->accel_mapping);
 
         io_sethandler(0x03c0, 0x0020, tgui_in, NULL, NULL, tgui_out, NULL, NULL, tgui);
-        if (tgui->type >= TGUI_9440)
+        if (tgui->type >= TGUI_9440) {
                 io_sethandler(0x43c8, 0x0002, tgui_in, NULL, NULL, tgui_out, NULL, NULL, tgui);
+				io_sethandler(0x2120, 0x0001, tgui_accel_in, tgui_accel_in_w, tgui_accel_in_l, tgui_accel_out, tgui_accel_out_w, tgui_accel_out_l, tgui);
+				io_sethandler(0x2122, 0x0001, tgui_accel_in, tgui_accel_in_w, tgui_accel_in_l, tgui_accel_out, tgui_accel_out_w, tgui_accel_out_l, tgui);
+				io_sethandler(0x2124, 0x0001, tgui_accel_in, tgui_accel_in_w, tgui_accel_in_l, tgui_accel_out, tgui_accel_out_w, tgui_accel_out_l, tgui);
+				io_sethandler(0x2127, 0x0001, tgui_accel_in, tgui_accel_in_w, tgui_accel_in_l, tgui_accel_out, tgui_accel_out_w, tgui_accel_out_l, tgui);
+				io_sethandler(0x2128, 0x0002, tgui_accel_in, tgui_accel_in_w, tgui_accel_in_l, tgui_accel_out, tgui_accel_out_w, tgui_accel_out_l, tgui);
+				io_sethandler(0x212b, 0x0001, tgui_accel_in, tgui_accel_in_w, tgui_accel_in_l, tgui_accel_out, tgui_accel_out_w, tgui_accel_out_l, tgui);
+				io_sethandler(0x212c, 0x0002, tgui_accel_in, tgui_accel_in_w, tgui_accel_in_l, tgui_accel_out, tgui_accel_out_w, tgui_accel_out_l, tgui);
+				io_sethandler(0x2130, 0x0002, tgui_accel_in, tgui_accel_in_w, tgui_accel_in_l, tgui_accel_out, tgui_accel_out_w, tgui_accel_out_l, tgui);
+				io_sethandler(0x2134, 0x0002, tgui_accel_in, tgui_accel_in_w, tgui_accel_in_l, tgui_accel_out, tgui_accel_out_w, tgui_accel_out_l, tgui);
+				io_sethandler(0x2138, 0x0002, tgui_accel_in, tgui_accel_in_w, tgui_accel_in_l, tgui_accel_out, tgui_accel_out_w, tgui_accel_out_l, tgui);
+				io_sethandler(0x213a, 0x0002, tgui_accel_in, tgui_accel_in_w, tgui_accel_in_l, tgui_accel_out, tgui_accel_out_w, tgui_accel_out_l, tgui);
+				io_sethandler(0x213c, 0x0002, tgui_accel_in, tgui_accel_in_w, tgui_accel_in_l, tgui_accel_out, tgui_accel_out_w, tgui_accel_out_l, tgui);
+				io_sethandler(0x213e, 0x0002, tgui_accel_in, tgui_accel_in_w, tgui_accel_in_l, tgui_accel_out, tgui_accel_out_w, tgui_accel_out_l, tgui);
+				io_sethandler(0x2140, 0x0002, tgui_accel_in, tgui_accel_in_w, tgui_accel_in_l, tgui_accel_out, tgui_accel_out_w, tgui_accel_out_l, tgui);
+				io_sethandler(0x2142, 0x0002, tgui_accel_in, tgui_accel_in_w, tgui_accel_in_l, tgui_accel_out, tgui_accel_out_w, tgui_accel_out_l, tgui);
+				io_sethandler(0x2180, 0x0080, tgui_accel_in, tgui_accel_in_w, tgui_accel_in_l, tgui_accel_out, tgui_accel_out_w, tgui_accel_out_l, tgui);
+		}
 
         if ((info->flags & DEVICE_PCI) && (tgui->type >= TGUI_9440))
                 pci_add_card(PCI_ADD_VIDEO, tgui_pci_read, tgui_pci_write, tgui);
 
-        tgui->i2c = i2c_gpio_init("ddc_tgui9440");
-        tgui->ddc = ddc_init(i2c_gpio_get_bus(tgui->i2c));
-
-        tgui->wake_fifo_thread = thread_create_event();
-        tgui->fifo_not_full_event = thread_create_event();
-        tgui->fifo_thread = thread_create(fifo_thread, tgui);
-
         return tgui;
 }
 
 static int tgui9400cxi_available()
 {
-        return rom_present("roms/video/tgui9440/9400CXI.vbi");
+        return rom_present(ROM_TGUI_9400CXI);
 }
 
 static int tgui9440_available()
 {
-        return rom_present("roms/video/tgui9440/9440.vbi");
+        return rom_present(ROM_TGUI_9440);
 }
 
 void tgui_close(void *p)
@@ -1768,13 +1902,6 @@ void tgui_close(void *p)
         
         svga_close(&tgui->svga);
 
-        ddc_close(tgui->ddc);
-        i2c_gpio_close(tgui->i2c);
-
-        thread_kill(tgui->fifo_thread);
-        thread_destroy_event(tgui->wake_fifo_thread);
-        thread_destroy_event(tgui->fifo_not_full_event);
-
         free(tgui);
 }