diff --git a/Makefile b/Makefile
index 10ee95e97..4f3ba2489 100644
--- a/Makefile
+++ b/Makefile
@@ -100,6 +100,7 @@ OBJS = \
 	d_mapinfo.o \
 	sh2_fixed.o \
 	sh2_draw.o \
+	sh2_drawlow.o \
 	sh2_mixer.o \
 	r_cache.o \
 	m_fire.o \
@@ -117,7 +118,7 @@ m68k.bin:
 
 $(TARGET).32x: $(TARGET).elf
 	$(OBJC) -O binary $< temp2.bin
-	$(DD) if=temp2.bin of=temp.bin bs=198K conv=sync
+	$(DD) if=temp2.bin of=temp.bin bs=200K conv=sync
 	rm -f temp3.bin
 	cat temp.bin $(WAD) >>temp3.bin
 	$(DD) if=temp3.bin of=$@ bs=512K conv=sync
diff --git a/marsdraw.c b/marsdraw.c
index 5b9a9ad29..cbceb7339 100644
--- a/marsdraw.c
+++ b/marsdraw.c
@@ -523,7 +523,7 @@ void I_DrawSpanPotatoLow(int ds_y, int ds_x1, int ds_x2, int light, fixed_t ds_x
 {
 	pixel_t* dest, pix;
 	unsigned count;
-	short* dc_colormap;
+	int8_t *dc_colormap;
 
 #ifdef RANGECHECK
 	if (ds_x2 < ds_x1 || ds_x1<0 || ds_x2 >= viewportWidth || ds_y>viewportHeight)
@@ -536,8 +536,9 @@ void I_DrawSpanPotatoLow(int ds_y, int ds_x1, int ds_x2, int light, fixed_t ds_x
 	count = ds_x2 - ds_x1 + 1;
 
 	dest = viewportbuffer + ds_y * 320 / 2 + ds_x1;
-	dc_colormap = (int16_t *)dc_colormaps + light;
-	pix = dc_colormap[(int8_t)ds_source[513]];
+	dc_colormap = (int8_t *)dc_colormaps + light;
+	pix = (uint8_t)dc_colormap[(int8_t)ds_source[513]];
+	pix = ((uint8_t)pix << 8) | pix;
 
 	do {
 		*dest++ = pix;
diff --git a/marsnew.c b/marsnew.c
index dbf4b7ab1..96670acbd 100644
--- a/marsnew.c
+++ b/marsnew.c
@@ -713,13 +713,14 @@ pixel_t* I_OverwriteBuffer(void)
 int I_ViewportYPos(void)
 {
 	const int fbh = I_FrameBufferHeight();
+	int width = viewportWidth * (lowres ? 2 : 1);
 
 	if (splitscreen)
 		return (fbh - viewportHeight) / 2;
 
-	if (viewportWidth < 160)
+	if (width < 160)
 		return (fbh - jo_stbar_height - viewportHeight) / 2;
-	if (viewportWidth == 320)
+	if (width == 320)
 		return (fbh - jo_stbar_height - viewportHeight);
 	return (fbh - jo_stbar_height - viewportHeight) / 2;
 }
@@ -728,14 +729,15 @@ pixel_t	*I_ViewportBuffer (void)
 {
 	int x = 0;
 	pixel_t *vb = (pixel_t * )framebuffer;
-	
+	int width = viewportWidth * (lowres ? 2 : 1);
+
 	if (splitscreen)
 	{
 		x = vd->displayplayer ? 160 : 0;
 	}
 	else
 	{
-		x = (320 - viewportWidth) / 2;
+		x = (320 - width) / 2;
 	}
 
 	vb += I_ViewportYPos() * 320 / 2 + x / 2;
diff --git a/marssave.c b/marssave.c
index 311108d64..dab15bf0d 100644
--- a/marssave.c
+++ b/marssave.c
@@ -65,6 +65,7 @@ typedef struct __attribute((packed))
 	int8_t yabcdpad;
 	int8_t sfxdriver;
 	char spcmDir[9];
+	int8_t lowres;
 } saveopts_t;
 
 static char saveslotguard[SRAM_SLOTSIZE - sizeof(savegame_t)] __attribute__((unused));
@@ -192,6 +193,7 @@ static void SaveOptions(void)
 	so.yabcdpad = yabcdpad;
 	so.magic1 = SRAM_MAGIC1;
 	so.magic2 = SRAM_MAGIC2;
+	so.lowres = lowres;
 	D_snprintf(so.spcmDir, sizeof(so.spcmDir), "%s", spcmDir);
 
 	Mars_WriteSRAM((void*)&so, optslotoffset, sizeof(saveopts_t));
@@ -247,6 +249,7 @@ static void ReadOptions(void)
 	anamorphicview = so.anamorphic;
 	sfxdriver = so.sfxdriver;
 	yabcdpad = so.yabcdpad;
+	lowres = so.lowres;
 
 	so.spcmDir[sizeof(so.spcmDir)-1] = '\0';
 	D_snprintf(spcmDir, sizeof(spcmDir), "%s", so.spcmDir);
@@ -274,6 +277,7 @@ void ReadEEProm(void)
 	anamorphicview = 0;
 	sfxdriver = 0;
 	detailmode = detmode_normal;
+	lowres = false;
 
 	ReadOptions();
 }
diff --git a/o_main.c b/o_main.c
index 6fa757632..7a6e335f6 100644
--- a/o_main.c
+++ b/o_main.c
@@ -39,6 +39,7 @@ typedef enum
 	mi_resolution,
 	mi_anamorphic,
 	mi_detailmode,
+	mi_lowres,
 
 	mi_controltype,
 	mi_yabcdpad,
@@ -247,6 +248,10 @@ void O_Init (void)
 	menuitem[mi_detailmode].x = ITEMX;
 	menuitem[mi_detailmode].y = STARTY + ITEMSPACE * 3;
 
+	D_memcpy(menuitem[mi_lowres].name, "Low detail", 11);
+	menuitem[mi_lowres].x = ITEMX;
+	menuitem[mi_lowres].y = STARTY + ITEMSPACE * 4;
+
 	D_memcpy(menuitem[mi_controltype].name, "Gamepad", 8);
 	menuitem[mi_controltype].x = ITEMX;
 	menuitem[mi_controltype].y = STARTY;
@@ -293,7 +298,7 @@ void O_Init (void)
 
 	D_memcpy(menuscreen[ms_video].name, "Video", 6);
 	menuscreen[ms_video].firstitem = mi_resolution;
-	menuscreen[ms_video].numitems = mi_detailmode - mi_resolution + 1;
+	menuscreen[ms_video].numitems = mi_lowres - mi_resolution + 1;
 
 	D_memcpy(menuscreen[ms_controls].name, "Controls", 9);
 	menuscreen[ms_controls].firstitem = mi_controltype;
@@ -724,6 +729,13 @@ void O_Control (player_t *player)
 						if (detailmode != detmode_potato)
 							detailmode = detmode_potato;
 						break;
+					case mi_lowres:
+						if (!lowres)
+						{
+							lowres = true;
+							R_SetViewportSize(viewportNum);
+						}
+						break;
 					}
 				}
 
@@ -738,6 +750,13 @@ void O_Control (player_t *player)
 						if (detailmode != detmode_normal)
 							detailmode = detmode_normal;
 						break;
+					case mi_lowres:
+						if (lowres)
+						{
+							lowres = false;
+							R_SetViewportSize(viewportNum);
+						}
+						break;
 					}
 				}
 
@@ -908,6 +927,15 @@ void O_Drawer (void)
 			print(menuitem[mi_detailmode].x + 160, menuitem[mi_detailmode].y, "off");
 			break;
 		}
+
+		switch (lowres) {
+		case true:
+			print(menuitem[mi_lowres].x + 160, menuitem[mi_lowres].y, "on");
+			break;
+		default:
+			print(menuitem[mi_lowres].x + 160, menuitem[mi_lowres].y, "off");
+			break;
+		}
 	}
 
 	if (screenpos == ms_help)
diff --git a/p_tick.c b/p_tick.c
index 38058e7d7..f82e195ca 100644
--- a/p_tick.c
+++ b/p_tick.c
@@ -540,9 +540,11 @@ void P_Drawer (void)
 		clearscreen = 2;
 
 	if (clearscreen > 0) {
+		int width = viewportWidth * (lowres ? 2 : 1);
+
 		I_ResetLineTable();
 
-		if (viewportWidth == 320)
+		if (width == 320)
 			I_ClearFrameBuffer();
 		else
 			DrawTiledBackground();
diff --git a/r_local.h b/r_local.h
index d500eb2ab..fa4ed81c9 100644
--- a/r_local.h
+++ b/r_local.h
@@ -12,6 +12,7 @@ extern fixed_t stretch;
 extern fixed_t stretchX;
 extern VINT weaponYpos;
 extern fixed_t weaponXScale;
+extern boolean lowres;
 
 #define	PROJECTION			centerXFrac
 
diff --git a/r_main.c b/r_main.c
index e33236dfd..4378f83f8 100644
--- a/r_main.c
+++ b/r_main.c
@@ -14,6 +14,7 @@ fixed_t stretch;
 fixed_t stretchX;
 VINT weaponYpos;
 fixed_t weaponXScale;
+boolean lowres = false;
 
 VINT detailmode;
 
@@ -260,6 +261,8 @@ void R_SetViewportSize(int num)
 
 	width = viewports[num][splitscreen][0];
 	height = viewports[num][splitscreen][1];
+	if (lowres)
+		width /= 2;
 
 	viewportNum = num;
 	viewportWidth = width;
@@ -281,7 +284,7 @@ void R_SetViewportSize(int num)
 		//stretch = (fixed_t)((160.0f / width) * ((float)height / 180.0f) * 2.2f * FRACUNIT);
 		stretch = ((FRACUNIT * 16 * height) / 180 * 22) / width;
 	}
-	weaponXScale = FRACUNIT;
+	weaponXScale = FRACUNIT / (lowres ? 2 : 1);
 	stretchX = stretch * centerX;
 
 	weaponYpos = 180;
@@ -319,11 +322,22 @@ void R_SetDrawFuncs(void)
 	if (detailmode < detmode_potato || detailmode >= MAXDETAILMODES)
 		detailmode = detmode_normal;
 
-	drawcol = I_DrawColumn;
-	drawcolnpo2 = I_DrawColumnNPo2;
-	drawfuzzcol = I_DrawFuzzColumn;
-	drawspan = detailmode == detmode_potato ? I_DrawSpanPotato : I_DrawSpan;
+	if (lowres)
+	{
+		drawcol = I_DrawColumnLow;
+		drawcolnpo2 = I_DrawColumnNPo2Low;
+		drawfuzzcol = I_DrawFuzzColumnLow;
+		drawspan = detailmode == detmode_potato ? I_DrawSpanPotatoLow : I_DrawSpanLow;
+	}
+	else
+	{
+		drawcol = I_DrawColumn;
+		drawcolnpo2 = I_DrawColumnNPo2;
+		drawfuzzcol = I_DrawFuzzColumn;
+		drawspan = detailmode == detmode_potato ? I_DrawSpanPotato : I_DrawSpan;
+	}
 
+	Mars_ClearCache();
 #ifdef MARS
 	Mars_CommSlaveClearCache();
 #endif
diff --git a/r_phase3.c b/r_phase3.c
index 20b1b6f4e..1d9f26beb 100644
--- a/r_phase3.c
+++ b/r_phase3.c
@@ -217,7 +217,7 @@ static void R_PrepPSprite(pspdef_t *psp)
       return;
 
    xscale = weaponXScale;
-   center = centerXFrac - 160 * weaponXScale;
+   center = centerXFrac - (lowres ? 320 - viewportWidth : 160) * weaponXScale;
 
    tx = psp->sx + center;
    topoffset = (((fixed_t)BIGSHORT(patch->topoffset) - weaponYpos) << FRACBITS) - psp->sy;
diff --git a/sh2_drawlow.s b/sh2_drawlow.s
index 2f6dc9231..bef103666 100644
--- a/sh2_drawlow.s
+++ b/sh2_drawlow.s
@@ -7,16 +7,14 @@
 
 ! Draw a vertical column of pixels from a projected wall texture.
 ! Source is the top of the column to scale.
-! Low detail (doubl-wide pixels) mode.
 !
-!void I_DrawColumnLow(int dc_x, int dc_yl, int dc_yh, int light, fixed_t frac,
+!void I_DrawColumn(int dc_x, int dc_yl, int dc_yh, int light, fixed_t frac,
 !                  fixed_t fracstep, inpixel_t *dc_source, int dc_texheight)
-
-        .align  4
+        .align 2
         .global _I_DrawColumnLowA
 _I_DrawColumnLowA:
-	add	#1,r6
 
+	add	#1,r6
 0:
         cmp/ge  r6,r5
         bf/s    1f
@@ -28,70 +26,74 @@ _I_DrawColumnLowA:
 1:
         mov.l   r8,@-r15
         mov.l   r9,@-r15
+        mov.l   r10,@-r15
         mov.l   @(DOOMTLS_COLORMAP, gbr),r0
-        add     r7,r7
         add     r0,r7           /* dc_colormap = colormap + light */
         mov.l   draw_fb,r8
         mov.l   @r8,r8          /* frame buffer start */
-        add     r4,r8
-        add     r4,r8           /* fb += dc_x*2 */
+        add     r4,r8           /* fb += dc_x */
+        add     r4,r8           /* fb += dc_x */
         shll8   r5
         add     r5,r8
         shlr2   r5
         add     r5,r8           /* fb += (dc_yl*256 + dc_yl*64) */
-        mov.l   @(8,r15),r2     /* frac */
-        mov.l   @(12,r15),r3    /* fracstep */
-        mov.l   @(16,r15),r5    /* dc_source */
-        mov.l   @(20,r15),r4
+        mov.l   @(12,r15),r2     /* frac */
+        mov.l   @(16,r15),r3    /* fracstep */
+        mov.l   @(20,r15),r5    /* dc_source */
+        mov.l   @(24,r15),r4
         mov.l   draw_width,r1
         add     #-1,r4          /* heightmask = texheight - 1 */
+
         swap.w  r2,r0           /* (frac >> 16) */
         and     r4,r0           /* (frac >> 16) & heightmask */
 
         /* test if count & 1 */
         shlr    r6
         movt    r9              /* 1 if count was odd */
-        bt/s    do_col_loop_low_1px
+        bt/s    do_col_loop_1px
         add     r9,r6
 
         .p2alignw 2, 0x0009
-do_col_loop_low:
+do_col_loop:
         mov.b   @(r0,r5),r0     /* pix = dc_source[(frac >> 16) & heightmask] */
-        add     r0,r0
-        mov.w   @(r0,r7),r9     /* dpix = dc_colormap[pix] */
         add     r3,r2           /* frac += fracstep */
+        mov.b   @(r0,r7),r9     /* dpix = dc_colormap[pix] */
         swap.w  r2,r0           /* (frac >> 16) */
-        and     r4,r0           /* (frac >> 16) & heightmask */
+        extu.b  r9,r9
+        swap.b  r9,r10
+        or      r10,r9
         mov.w   r9,@r8          /* *fb = dpix */
         add     r1,r8           /* fb += SCREENWIDTH */
-do_col_loop_low_1px:
-        mov.b   @(r0,r5),r0     /* pix = dc_source[(frac >> 16) & heightmask] */
-        add     r0,r0
-        mov.w   @(r0,r7),r9     /* dpix = dc_colormap[pix] */
+        and     r4,r0           /* (frac >> 16) & heightmask */
+do_col_loop_1px:
         add     r3,r2           /* frac += fracstep */
+        mov.b   @(r0,r5),r0     /* pix = dc_source[(frac >> 16) & heightmask] */
         dt      r6              /* count-- */
+        mov.b   @(r0,r7),r9     /* dpix = dc_colormap[pix] */
         swap.w  r2,r0           /* (frac >> 16) */
-        mov.w   r9,@r8          /* *fb = dpix */
         and     r4,r0           /* (frac >> 16) & heightmask */
-        bf/s    do_col_loop_low
+        extu.b  r9,r9
+        swap.b  r9,r10
+        or      r10,r9
+        mov.w   r9,@r8          /* *fb = dpix */
+        bf/s    do_col_loop
         add     r1,r8           /* fb += SCREENWIDTH */
 
+        mov.l   @r15+,r10
         mov.l   @r15+,r9
         rts
         mov.l   @r15+,r8
 
 ! Draw a vertical column of pixels from a projected wall texture.
 ! Non-power of 2 texture height.
-! Low detail (doubl-wide pixels) mode.
 !
-!void I_DrawColumnNPO2Low(int dc_x, int dc_yl, int dc_yh, int light, fixed_t frac,
+!void I_DrawColumnNPO2(int dc_x, int dc_yl, int dc_yh, int light, fixed_t frac,
 !                      fixed_t fracstep, inpixel_t *dc_source, int dc_texheight)
-
-        .align  4
+        .align 2
         .global _I_DrawColumnNPo2LowA
 _I_DrawColumnNPo2LowA:
-	add	#1,r6
 
+	add	#1,r6
 0:
         cmp/ge  r6,r5
         bf/s    1f
@@ -102,21 +104,21 @@ _I_DrawColumnNPo2LowA:
         nop
 1:
         mov.l   r8,@-r15
+        mov.l   r9,@-r15 
         mov.l   @(DOOMTLS_COLORMAP, gbr),r0
-        add     r7,r7
         add     r0,r7           /* dc_colormap = colormap + light */
         mov.l   draw_fb,r8
         mov.l   @r8,r8          /* frame buffer start */
-        add     r4,r8
-        add     r4,r8           /* fb += dc_x*2 */
+        add     r4,r8           /* fb += dc_x */
+        add     r4,r8           /* fb += dc_x */
         shll8   r5
         add     r5,r8
         shlr2   r5
         add     r5,r8           /* fb += (dc_yl*256 + dc_yl*64) */
-        mov.l   @(4,r15),r2     /* frac */
-        mov.l   @(8,r15),r3     /* fracstep */
-        mov.l   @(12,r15),r5    /* dc_source */
-        mov.l   @(16,r15),r4
+        mov.l   @(8,r15),r2     /* frac */
+        mov.l   @(12,r15),r3     /* fracstep */
+        mov.l   @(16,r15),r5    /* dc_source */
+        mov.l   @(20,r15),r4
         shll16  r4              /* heightmask = texheight << FRACBITS */
 
         mov     #0,r0
@@ -138,26 +140,28 @@ _I_DrawColumnNPo2LowA:
         bt      4b
 5:
         mov.l   draw_width,r1
-        sub     r1,r8           /* fb -= SCREENWIDTH */
 
         .p2alignw 2, 0x0009
-do_cnp_loop_low:
+do_cnp_loop:
         mov     r2,r0
         shlr16  r0              /* frac >> 16 */
         mov.b   @(r0,r5),r0     /* pix = dc_source[frac >> 16] */
-        add     r1,r8           /* fb += SCREENWIDTH */
-        add     r3,r2           /* frac += fracstep */
         cmp/ge  r4,r2
-        add     r0,r0
+        mov.b   @(r0,r7),r0     /* dpix = dc_colormap[pix] */
         bf/s    1f
-        mov.w   @(r0,r7),r0     /* dpix = dc_colormap[pix] */
+        add     r3,r2           /* frac += fracstep */
         /* if (frac >= heightmask) */
         sub     r4,r2           /* frac -= heightmask */
 1:
         dt      r6              /* count-- */
-        bf/s    do_cnp_loop_low
+        extu.b  r0,r0
+        swap.b  r0,r9
+        or      r9,r0
         mov.w   r0,@r8          /* *fb = dpix */
+        bf/s    do_cnp_loop
+        add     r1,r8           /* fb += SCREENWIDTH */
 
+        mov.l   @r15+,r9
         rts
         mov.l   @r15+,r8
 
@@ -165,12 +169,12 @@ do_cnp_loop_low:
 ! Source is the top of the column to scale.
 ! Low detail (doubl-wide pixels) mode.
 !
-!void I_DrawFuzzColumnLow(int dc_x, int dc_yl, int dc_yh, int light, fixed_t frac,
+!void I_DrawFuzzColumn(int dc_x, int dc_yl, int dc_yh, int light, fixed_t frac,
 !       fixed_t fracstep, inpixel_t *dc_source, int dc_texheight)
-
-        .align  4
+        .align 2
         .global _I_DrawFuzzColumnLowA
 _I_DrawFuzzColumnLowA:
+
         mov     #0,r0
         cmp/eq  r0,r5
         bf      0f
@@ -195,12 +199,12 @@ _I_DrawFuzzColumnLowA:
 3:
         mov.l   @(DOOMTLS_COLORMAP, gbr),r0
         mov.l   r8,@-r15
-        add     r7,r7
+        mov.l   r9,@-r15
         add     r0,r7           /* dc_colormap = colormap + light */
         mov.l   draw_fb,r8
         mov.l   @r8,r8          /* frame buffer start */
-        add     r4,r8
-        add     r4,r8           /* fb += dc_x*2 */
+        add     r4,r8           /* fb += dc_x */
+        add     r4,r8           /* fb += dc_x */
         shll8   r5
         add     r5,r8
         shlr2   r5
@@ -213,37 +217,38 @@ _I_DrawFuzzColumnLowA:
         and     #126,r0         /* fuzzpos &= FUZZMASK */
 
         .p2alignw 2, 0x0009
-do_fuzz_col_loop_low:
+do_fuzz_col_loop:
         mov.w   @(r0,r5),r0     /* pix = fuzztable[fuzzpos] */
-        dt      r6              /* count-- */
-        mov.b   @(r0,r8),r0     /* pix = dest[pix] */
-        add     r0,r0
-        mov.w   @(r0,r7),r4     /* dpix = dc_colormap[pix] */
         add     #2,r3
+        mov.b   @(r0,r8),r0     /* pix = dest[pix] */
+        dt      r6              /* count-- */
+        mov.b   @(r0,r7),r4     /* dpix = dc_colormap[pix] */
         mov     r3,r0
         and     #126,r0         /* fuzzpos &= FUZZMASK */
-        mov.w   r4,@r8          /* *fb = dpix */
-        bf/s    do_fuzz_col_loop_low
+        extu.b  r4,r4
+        swap.b  r4,r9
+        or      r9,r4
+        mov.b   r4,@r8          /* *fb = dpix */
+        bf/s    do_fuzz_col_loop
         add     r1,r8           /* fb += SCREENWIDTH */
 
         shlr    r3
         mov     r3,r0
         mov.l   r0,@(DOOMTLS_FUZZPOS, gbr)
+        mov.l   @r15+,r9
         rts
         mov.l   @r15+,r8
 
 ! Draw a horizontal row of pixels from a projected flat (floor/ceiling) texture.
-! Low detail (doubl-wide pixels) mode.
 !
-!void I_DrawSpanLow(int ds_y, int ds_x1, int ds_x2, int light, fixed_t ds_xfrac,
+!void I_DrawSpan(int ds_y, int ds_x1, int ds_x2, int light, fixed_t ds_xfrac,
 !                fixed_t ds_yfrac, fixed_t ds_xstep, fixed_t ds_ystep,
 !                inpixel_t *ds_source, int ds_height)
-
-        .align  4
+        .align 2
         .global _I_DrawSpanLowA
 _I_DrawSpanLowA:
-	add	#1,r6
 
+	add	#1,r6
 0:
         cmp/ge  r6,r5
         bf/s    1f
@@ -252,80 +257,100 @@ _I_DrawSpanLowA:
         /* dc_yl >= dc_yh, exit */
         rts
         nop
+
 1:
         mov.l   r8,@-r15
         mov.l   r9,@-r15
         mov.l   r10,@-r15
         mov.l   r11,@-r15
         mov.l   r12,@-r15
+        mov.l   r13,@-r15
         mov.l   @(DOOMTLS_COLORMAP, gbr),r0
-        add     r7,r7
         add     r0,r7           /* ds_colormap = colormap + light */
         mov.l   draw_fb,r8
+
         mov.l   @r8,r8          /* frame buffer start */
         add     r5,r8
-        add     r5,r8           /* fb += ds_x1*2 */
+        add     r5,r8
+        add     r6,r8
+        add     r6,r8
         shll8   r4
         add     r4,r8
         shlr2   r4
         add     r4,r8           /* fb += (ds_y*256 + ds_y*64) */
-        mov.l   @(20,r15),r2    /* xfrac */
-        mov.l   @(24,r15),r4    /* yfrac */
-        mov.l   @(28,r15),r3    /* xstep */
-        mov.l   @(32,r15),r5    /* ystep */
-        mov.l   @(36,r15),r9    /* ds_source */
-        mov.l   @(40,r15),r12   /* ds_height */
-
-        mov     r12,r11
-        dt      r11
-        mulu.w  r12,r11         /* (ds_height-1) * ds_height */
+
+        mov.l   @(40,r15),r9    /* ds_source */
+        mov.l   @(44,r15),r11   /* ds_height */
+
+        mov     r11,r12
         dt      r12             /* (ds_height-1) */
+        mulu.w  r12,r11         /* (ds_height-1) * ds_height */
+
+        mov.l   @(24,r15),r2    /* xfrac */
+        mov.l   @(32,r15),r3    /* xstep */
+
         sts     macl,r11        /* draw_flat_ymask */
 
-        swap.w  r4,r1           /* (yfrac >> 16) */
-        and     r11,r1          /* (yfrac >> 16) & 63*64 */
+        dmuls.l r3,r6
+
+        mov.l   @(28,r15),r4    /* yfrac */
+        mov.l   @(36,r15),r5    /* ystep */
+
+        sts     macl,r10
+        add     r10,r2          /* xfrac += xstep * count */
+
+        dmuls.l r5,r6
+
         swap.w  r2,r0           /* (xfrac >> 16) */
         and     r12,r0          /* (xfrac >> 16) & 63 */
+
+        sts     macl,r10
+        add     r10,r4          /* yfrac += ystep * count */
+
+        swap.w  r4,r1           /* (yfrac >> 16) */
+        and     r11,r1          /* (yfrac >> 16) & 63*64 */
+
         or      r1,r0           /* spot = ((yfrac >> 16) & *64) | ((xfrac >> 16) & 63) */
 
         /* test if count & 1 */
         shlr    r6
         movt    r1              /* 1 if count was odd */
-        bt/s    do_span_low_loop_1px
+        bt/s    do_span_loop_1px
         add     r1,r6
 
         .p2alignw 2, 0x0009
-do_span_low_loop:
+do_span_loop:
         mov.b   @(r0,r9),r0     /* pix = ds_source[spot] */
-        add     r3,r2           /* xfrac += xstep */
-        add     r5,r4           /* yfrac += ystep */
-        add     r0,r0
-        mov.w   @(r0,r7),r10    /* dpix = ds_colormap[pix] */
+        sub     r3,r2           /* xfrac -= xstep */
+        sub     r5,r4           /* yfrac -= ystep */
         swap.w  r4,r1           /* (yfrac >> 16) */
-        and     r11,r1          /* (yfrac >> 16) & 63*64 */
+        mov.b   @(r0,r7),r10    /* dpix = ds_colormap[pix] */
         swap.w  r2,r0           /* (xfrac >> 16) */
         and     r12,r0          /* (xfrac >> 16) & 63 */
+        extu.b  r10,r10
+        swap.b  r10,r13
+        or      r13,r10
+        mov.w   r10,@-r8        /* *--fb = dpix */
+        and     r11,r1          /* (yfrac >> 16) & 63*64 */
         or      r1,r0           /* spot = ((yfrac >> 16) & *64) | ((xfrac >> 16) & 63) */
-        mov.w   r10,@r8         /* *fb = dpix */
-        add     #2,r8           /* fb++ */
-
-do_span_low_loop_1px:
+do_span_loop_1px:
+        dt      r6              /* count-- */
         mov.b   @(r0,r9),r0     /* pix = ds_source[spot] */
-        add     r3,r2           /* xfrac += xstep */
-        add     r5,r4           /* yfrac += ystep */
-        add     r0,r0
-        mov.w   @(r0,r7),r10    /* dpix = ds_colormap[pix] */
+        sub     r3,r2           /* xfrac -= xstep */
+        sub     r5,r4           /* yfrac -= ystep */
         swap.w  r4,r1           /* (yfrac >> 16) */
-        and     r11,r1          /* (yfrac >> 16) & 63*64 */
+        mov.b   @(r0,r7),r10    /* dpix = ds_colormap[pix] */
         swap.w  r2,r0           /* (xfrac >> 16) */
         and     r12,r0          /* (xfrac >> 16) & 63 */
+        extu.b  r10,r10
+        swap.b  r10,r13
+        or      r13,r10
+        mov.w   r10,@-r8        /* *--fb = dpix */
+        and     r11,r1          /* (yfrac >> 16) & 63*64 */
+        bf/s    do_span_loop
         or      r1,r0           /* spot = ((yfrac >> 16) & *64) | ((xfrac >> 16) & 63) */
-        mov.w   r10,@r8         /* *fb = dpix */
-
-        dt      r6              /* count-- */
-        bf/s    do_span_low_loop
-        add     #2,r8           /* fb++ */
 
+        mov.l   @r15+,r13
         mov.l   @r15+,r12
         mov.l   @r15+,r11
         mov.l   @r15+,r10
@@ -340,7 +365,5 @@ draw_width:
         .long   320
 draw_height:
         .long   _viewportHeight
-draw_flat_ymask:
-        .long   4032
 draw_fuzzoffset:
         .long   _fuzzoffset
diff --git a/wadbase.s b/wadbase.s
index 6c3e29e19..a13583202 100644
--- a/wadbase.s
+++ b/wadbase.s
@@ -4,5 +4,5 @@
 
         .global _wadBase
 _wadBase:
-        .long   0x2031800
+        .long   0x2032000