diff --git a/Makefile.mak b/Makefile.mak
index 6fc5e707..f3a39b46 100644
--- a/Makefile.mak
+++ b/Makefile.mak
@@ -133,7 +133,7 @@ bin\th04\res_huma.com: th04\res_huma.cpp
 $**
 | masters.lib
 
-bin\th04\op.exe: bin\th04\op.obj th04\m_char.cpp bin\th01\vplanset.obj bin\frmdely1.obj bin\th03\pi_put.obj bin\th03\pi_load.obj bin\hfliplut.obj bin\th04\input_w.obj bin\th04\vector.obj bin\th04\snd_pmdr.obj bin\th04\snd_mmdr.obj bin\th04\snd_kaja.obj bin\th04\cdg_put.obj bin\th04\exit.obj bin\th04\initop.obj bin\th04\cdg_p_na.obj bin\th04\input_s.obj bin\th04\snd_se_r.obj bin\th04\snd_se.obj bin\th04\egcrect.obj bin\th04\bgimage.obj bin\th04\bgimager.obj bin\th04\cdg_load.obj th04\frmdely2.c
+bin\th04\op.exe: bin\th04\op.obj th04\m_char.cpp bin\th01\vplanset.obj bin\frmdely1.obj bin\th03\pi_put.obj bin\th03\pi_load.obj bin\hfliplut.obj bin\th04\input_w.obj bin\th04\vector.obj bin\th04\snd_pmdr.obj bin\th04\snd_mmdr.obj bin\th04\snd_kaja.obj bin\th04\cdg_p_nc.obj bin\th04\cdg_put.obj bin\th04\exit.obj bin\th04\initop.obj bin\th04\cdg_p_na.obj bin\th04\input_s.obj bin\th04\snd_se_r.obj bin\th04\snd_se.obj bin\th04\egcrect.obj bin\th04\bgimage.obj bin\th04\bgimager.obj bin\th04\cdg_load.obj th04\frmdely2.c
 	$(CC) $(CFLAGS) $(LARGE_LFLAGS) -DGAME=4 -DBINARY='O' -3 -Z -d -nbin\th04\ -eOP.EXE @&&|
 $**
 |
diff --git a/Tupfile b/Tupfile
index 1e52605e..f3ac72b2 100644
--- a/Tupfile
+++ b/Tupfile
@@ -89,6 +89,7 @@ BMP2ARR = bin\\Pipeline\\bmp2arr.exe
 : th04_memchk.asm |> !as |> bin\\th04\\memchk.obj
 : th04\\scoreupd.asm |> !as4 |>
 : th04\\cdg_put.asm |> !as4 |>
+: th04\\cdg_p_nc.asm |> !as4 |>
 : th04\\cdg_p_pr.asm |> !as4 |>
 : th04\\input_s.asm |> !as4 |>
 : th04\\cdg_load.asm |> !as4 |>
diff --git a/Tupfile.bat b/Tupfile.bat
index 239d27bc..4fcbd5c8 100644
--- a/Tupfile.bat
+++ b/Tupfile.bat
@@ -38,6 +38,7 @@ tasm32 /m /mx /kh32768 /t th04_zuninit.asm bin\th04\zuninit.obj
 tasm32 /m /mx /kh32768 /t th04_memchk.asm bin\th04\memchk.obj
 tasm32 /m /mx /kh32768 /t /dGAME=4 th04\scoreupd.asm bin\th04\scoreupd.obj
 tasm32 /m /mx /kh32768 /t /dGAME=4 th04\cdg_put.asm bin\th04\cdg_put.obj
+tasm32 /m /mx /kh32768 /t /dGAME=4 th04\cdg_p_nc.asm bin\th04\cdg_p_nc.obj
 tasm32 /m /mx /kh32768 /t /dGAME=4 th04\cdg_p_pr.asm bin\th04\cdg_p_pr.obj
 tasm32 /m /mx /kh32768 /t /dGAME=4 th04\input_s.asm bin\th04\input_s.obj
 tasm32 /m /mx /kh32768 /t /dGAME=4 th04\cdg_load.asm bin\th04\cdg_load.obj
diff --git a/th04/cdg_p_nc.asm b/th04/cdg_p_nc.asm
new file mode 100644
index 00000000..0fbe69f9
--- /dev/null
+++ b/th04/cdg_p_nc.asm
@@ -0,0 +1 @@
+include th04/formats/cdg_p_nc.asm
diff --git a/th04/formats/cdg_p_nc.asm b/th04/formats/cdg_p_nc.asm
new file mode 100644
index 00000000..f3f1fe24
--- /dev/null
+++ b/th04/formats/cdg_p_nc.asm
@@ -0,0 +1,60 @@
+	.386
+	locals
+
+include pc98.inc
+include th03/formats/cdg.inc
+
+	extrn _cdg_slots:cdg_t:CDG_SLOT_COUNT
+
+SHARED	segment word public 'CODE' use16
+	assume cs:SHARED
+
+; Identical to the barely decompilable TH05 version, which doesn't rely on
+; self-modifying code for the width.
+public CDG_PUT_NOCOLORS_8
+cdg_put_nocolors_8 proc far
+	; (PASCAL calling convention, parameter list needs to be reversed here)
+	arg @@slot:word, @@top:word, @@left:word
+	@@stride_backwards	equ <dx>
+
+	push	bp
+	mov	bp, sp
+	push	si
+	push	di
+
+	cdg_slot_offset	si, @@slot
+	cdg_dst_offset	di, si, @@left
+
+	mov	ax, [si+cdg_t.vram_dword_w]
+	mov	@@width, ax
+	jmp	short $+2
+	shl	ax, 2
+	add	ax, ROW_SIZE
+	mov	@@stride_backwards, ax
+
+	cdg_dst_segment	es, @@top, bx
+
+	push	ds
+	mov	ds, [si+cdg_t.seg_alpha]
+	xor	si, si
+	cld
+	even
+
+@@next_row:
+	@@width = word ptr $+1
+	mov	cx, 1234h
+	rep movsd
+	sub	di, @@stride_backwards
+	jns	short @@next_row
+
+	pop	ds
+	pop	di
+	pop	si
+	pop	bp
+	retf	6
+cdg_put_nocolors_8 endp
+	even
+
+SHARED	ends
+
+	end
diff --git a/th04/formats/cdg_put_nocolors.asm b/th04/formats/cdg_put_nocolors.asm
deleted file mode 100644
index 65106f44..00000000
--- a/th04/formats/cdg_put_nocolors.asm
+++ /dev/null
@@ -1,44 +0,0 @@
-; Identical to the barely decompilable TH05 version, which doesn't rely on
-; self-modifying code for the width.
-public CDG_PUT_NOCOLORS_8
-cdg_put_nocolors_8 proc far
-
-@@slot	= word ptr  6
-@@top 	= word ptr  8
-@@left	= word ptr  10
-
-	push	bp
-	mov	bp, sp
-	push	si
-	push	di
-
-	cdg_slot_offset	si, [bp+@@slot]
-	cdg_dst_offset	di, si, [bp+@@left]
-
-	mov	ax, [si+cdg_t.vram_dword_w]
-	mov	word ptr cs:@@width+1, ax
-	jmp	short $+2
-	shl	ax, 2
-	add	ax, (640 / 8)
-	mov	dx, ax
-
-	cdg_dst_segment	es, [bp+@@top], bx
-
-	push	ds
-	mov	ds, [si+cdg_t.seg_alpha]
-	xor	si, si
-	cld
-	align 2
-
-@@width:
-	mov	cx, 1234h
-	rep movsd
-	sub	di, dx
-	jns	short @@width
-	pop	ds
-	pop	di
-	pop	si
-	pop	bp
-	retf	6
-cdg_put_nocolors_8 endp
-	align 2
diff --git a/th04_op.asm b/th04_op.asm
index f90b45ff..e83e5e5b 100644
--- a/th04_op.asm
+++ b/th04_op.asm
@@ -2642,13 +2642,13 @@ SHARED	segment	word public 'CODE' use16
 	extern _snd_pmd_resident:proc
 	extern _snd_mmd_resident:proc
 	extern SND_KAJA_INTERRUPT:proc
+	extern CDG_PUT_NOCOLORS_8:proc
 SHARED	ends
 
 SHARED_	segment	word public 'CODE' use16
 		assume cs:g_SHARED
 		assume es:nothing, ss:nothing, ds:_DATA, fs:nothing, gs:nothing
 
-include th04/formats/cdg_put_nocolors.asm
 include th04/snd/detmodes.asm
 include th03/snd/delaymea.asm
 include th02/exit_dos.asm