diff --git a/Makefile.mak b/Makefile.mak
index e7c37a79..fe1aed04 100644
--- a/Makefile.mak
+++ b/Makefile.mak
@@ -152,7 +152,7 @@ bin\th05\res_kso.com: th05\res_kso.cpp
 $**
 | masters.lib
 
-bin\th05\op.exe: th05\op010.cpp bin\th05\op.obj th05\op011.cpp th05\m_char.cpp bin\th05\pi_cpp_2.obj bin\th05\pi_asm_2.obj bin\th05\initop.obj bin\th05\input_s.obj bin\th05\inp_h_w.obj bin\th05\snd_dlym.obj th05\cdg_p_nc.cpp bin\th05\frmdelay.obj bin\th04\cdg_load.obj bin\th05\egcrect.obj bin\hfliplut.obj
+bin\th05\op.exe: th05\op010.cpp bin\th05\op.obj th05\op011.cpp th05\m_char.cpp bin\th05\pi_asm_1.obj bin\th05\pi_cpp_2.obj bin\th05\pi_asm_2.obj bin\th05\initop.obj bin\th05\input_s.obj bin\th05\inp_h_w.obj bin\th05\snd_dlym.obj th05\cdg_p_nc.cpp bin\th05\frmdelay.obj bin\th04\cdg_load.obj bin\th05\egcrect.obj bin\hfliplut.obj
 	$(CC) $(CFLAGS) $(LARGE_LFLAGS) -DGAME=5 -DBINARY='O' -3 -Z -nbin\th05\ -eOP.EXE @&&|
 $**
 |
@@ -162,7 +162,7 @@ bin\th05\main.exe: bin\th05\main.obj th05\main010.cpp th05\main011.cpp th05\p_co
 $**
 |
 
-bin\th05\maine.exe: bin\th05\maine.obj th05\maine011.cpp th05\regist.cpp th05\staff.cpp bin\th05\pi_cpp_2.obj bin\th05\pi_asm_2.obj bin\th05\initmain.obj bin\th05\input_s.obj bin\th05\inp_h_w.obj bin\th05\snd_dlym.obj bin\th05\frmdelay.obj bin\th04\cdg_load.obj bin\th05\egcrect.obj bin\hfliplut.obj
+bin\th05\maine.exe: bin\th05\maine.obj th05\maine011.cpp th05\regist.cpp th05\staff.cpp bin\th05\pi_asm_1.obj bin\th05\pi_cpp_2.obj bin\th05\pi_asm_2.obj bin\th05\initmain.obj bin\th05\input_s.obj bin\th05\inp_h_w.obj bin\th05\snd_dlym.obj bin\th05\frmdelay.obj bin\th04\cdg_load.obj bin\th05\egcrect.obj bin\hfliplut.obj
 	$(CC) $(CFLAGS) $(LARGE_LFLAGS) -DGAME=5 -DBINARY='E' -Z -nbin\th05\ -eMAINE.EXE @&&|
 $**
 |
diff --git a/Tupfile b/Tupfile
index acdbd901..ffe314aa 100644
--- a/Tupfile
+++ b/Tupfile
@@ -109,6 +109,7 @@ BMP2ARR = bin\\Pipeline\\bmp2arr.exe
 : th05\\player.asm |> !as5 |>
 : th05\\hud_bar.asm |> !as5 |>
 : th05\\bullet.asm |> !as5 |>
+: th05\\pi_asm_1.asm |> !as5 |>
 : th05\\pi_asm_2.asm |> !as5 |>
 : th05_op.asm | \
 	th05/sprites/piano_l.asp \
diff --git a/Tupfile.bat b/Tupfile.bat
index d3396f18..af8e9c67 100644
--- a/Tupfile.bat
+++ b/Tupfile.bat
@@ -50,6 +50,7 @@ tasm32 /m /mx /kh32768 /t /dGAME=5 th04\scoreupd.asm bin\th05\scoreupd.obj
 tasm32 /m /mx /kh32768 /t /dGAME=5 th05\player.asm bin\th05\player.obj
 tasm32 /m /mx /kh32768 /t /dGAME=5 th05\hud_bar.asm bin\th05\hud_bar.obj
 tasm32 /m /mx /kh32768 /t /dGAME=5 th05\bullet.asm bin\th05\bullet.obj
+tasm32 /m /mx /kh32768 /t /dGAME=5 th05\pi_asm_1.asm bin\th05\pi_asm_1.obj
 tasm32 /m /mx /kh32768 /t /dGAME=5 th05\pi_asm_2.asm bin\th05\pi_asm_2.obj
 tasm32 /m /mx /kh32768 /t th05_op.asm bin\th05\op.obj
 tasm32 /m /mx /kh32768 /t th05_main.asm bin\th05\main.obj
diff --git a/libs/master.lib/master.inc b/libs/master.lib/master.inc
index 0191f636..8f40d3c7 100644
--- a/libs/master.lib/master.inc
+++ b/libs/master.lib/master.inc
@@ -5,6 +5,7 @@
 
 include libs/master.lib/macros.inc
 
+EGC_OFF procdesc pascal far
 GAIJI_PUTSA procdesc pascal far \
 	x:word, y:word, strp_seg:word, strp_off:word, atrb:word
 GRAPH_PACK_PUT_8_NOCLIP procdesc pascal far \
diff --git a/th03/formats/pi_put_masked[data].asm b/th03/formats/pi_put_masked[data].asm
index 6058a293..121b6bdc 100644
--- a/th03/formats/pi_put_masked[data].asm
+++ b/th03/formats/pi_put_masked[data].asm
@@ -1,3 +1,5 @@
+PI_MASK_H = 4
+
 public _PI_MASKS
 _PI_MASKS	label word
 ; 0
diff --git a/th05/formats/pi_asm_1.asm b/th05/formats/pi_asm_1.asm
new file mode 100644
index 00000000..8f953b5c
--- /dev/null
+++ b/th05/formats/pi_asm_1.asm
@@ -0,0 +1,100 @@
+; First TH05 .PI assembly translation unit.
+
+	.386
+	.model large
+	locals
+
+include pc98.inc
+include libs/master.lib/master.inc
+
+	extrn _PI_MASKS:word
+	extrn _pi_mask_ptr:word
+	extrn _pi_mask_y:word
+	extrn _pi_put_masked_vram_offset:word
+
+; Apparently necessary to avoid fixup overflows. In this case, the call to
+; pi_masked_egc_setup_copy() would be bizarrely offset otherwise, which TLINK
+; wouldn't even report?! I did very much not enjoy debugging this.
+g_SHARED group SHARED, SHARED_
+SHARED	segment byte public 'CODE' use16
+SHARED	ends
+
+SHARED_	segment word public 'CODE' use16
+	assume cs:g_SHARED
+
+	_pi_mask_setup_egc_and_advance procdesc near
+
+public PI_PUT_MASKED_8_ROWLOOP
+pi_put_masked_8_rowloop proc near
+; Can't use ARG, because the function doesn't `PUSH BP`!
+@@stride_packed	= word ptr [bp+2]
+@@w            	= word ptr [bp+4]
+@@top          	= word ptr [bp+6]
+@@left         	= word ptr [bp+8]
+@@mask_id equ ax
+@@h equ di
+
+; Each original row is first blitted to this fixed row in VRAM, before being
+; blitted to its actual destination with the EGC active.
+TEMP_ROW = RES_Y
+
+	shl	@@mask_id, 3	; *= PI_MASK_H * word
+	add	@@mask_id, offset _PI_MASKS
+	mov	_pi_mask_ptr, @@mask_id
+	mov	bp, sp
+	mov	dx, @@left
+	shr	dx, 3
+	mov	ax, @@top
+	shl	ax, 6
+	add	dx, ax
+	shr	ax, 2
+	add	dx, ax
+	mov	_pi_put_masked_vram_offset, dx
+	mov	_pi_mask_y, 0
+
+@@put_row:
+	push	es
+	call	graph_pack_put_8_noclip pascal, 0, TEMP_ROW, es, si, @@w
+	push	ds
+	push	@@h
+	push	si
+	mov	di, _pi_put_masked_vram_offset
+	add	_pi_put_masked_vram_offset, ROW_SIZE
+	cmp	_pi_put_masked_vram_offset, PLANE_SIZE
+	jb	short @@next_row
+	sub	_pi_put_masked_vram_offset, PLANE_SIZE
+
+@@next_row:
+	call	_pi_mask_setup_egc_and_advance
+	mov	ax, SEG_PLANE_B
+	mov	es, ax
+	assume es:nothing
+	mov	ds, ax
+	assume ds:nothing
+	mov	si, (ROW_SIZE * TEMP_ROW)
+	mov	cx, @@w
+	shr	cx, 4
+	rep movsw
+	call	egc_off
+	pop	si
+	pop	@@h
+	pop	ds
+	pop	es
+	add	si, @@stride_packed
+
+	; .PI pointer normalization, see pi_buffer_p_normalize()
+	mov	ax, si
+	shr	ax, 4
+	mov	dx, es
+	add	dx, ax
+	mov	es, dx
+	and	si, 0Fh
+
+	dec	@@h
+	jnz	short @@put_row
+	retn	8
+pi_put_masked_8_rowloop endp
+
+SHARED_	ends
+
+	end
diff --git a/th05/formats/pi_put_masked.asm b/th05/formats/pi_put_masked.asm
index c9d2f873..48f530eb 100644
--- a/th05/formats/pi_put_masked.asm
+++ b/th05/formats/pi_put_masked.asm
@@ -1,4 +1,4 @@
-_pi_mask_setup_egc_and_advance procdesc near
+PI_PUT_MASKED_8_ROWLOOP procdesc near
 
 public PI_PUT_MASKED_8
 pi_put_masked_8 proc far
@@ -81,77 +81,3 @@ pi_put_quarter_masked_8 proc far
 	retf	0Ah
 pi_put_quarter_masked_8 endp
 	even
-
-; ---------------------------------------------------------------------------
-
-; void pascal pi_put_masked_8_rowloop(
-;	int mask_id<ax>,
-;	void far *pi_buf<es:si>,
-;	pixel_t h<di>,
-;	screen_x_t left, vram_y_t top, pixel_t w, size_t stride_packed
-; );
-pi_put_masked_8_rowloop proc near
-@@stride_packed	= word ptr [bp+2]
-@@w	= word ptr [bp+4]
-@@top	= word ptr [bp+6]
-@@left	= word ptr [bp+8]
-@@mask_id equ ax
-@@h equ di
-
-TEMP_ROW = RES_Y
-
-	shl	@@mask_id, 3
-	add	@@mask_id, offset _PI_MASKS
-	mov	_pi_mask_ptr, @@mask_id
-	mov	bp, sp
-	mov	dx, @@left
-	shr	dx, 3
-	mov	ax, @@top
-	shl	ax, 6
-	add	dx, ax
-	shr	ax, 2
-	add	dx, ax
-	mov	_pi_put_masked_vram_offset, dx
-	mov	_pi_mask_y, 0
-
-@@put_row:
-	push	es
-	call	graph_pack_put_8_noclip pascal, 0, TEMP_ROW, es, si, @@w
-	push	ds
-	push	@@h
-	push	si
-	mov	di, _pi_put_masked_vram_offset
-	add	_pi_put_masked_vram_offset, ROW_SIZE
-	cmp	_pi_put_masked_vram_offset, PLANE_SIZE
-	jb	short @@next_row
-	sub	_pi_put_masked_vram_offset, PLANE_SIZE
-
-@@next_row:
-	call	_pi_mask_setup_egc_and_advance
-	mov	ax, GRAM_400
-	mov	es, ax
-	assume es:nothing
-	mov	ds, ax
-	assume ds:nothing
-	mov	si, (ROW_SIZE * TEMP_ROW)
-	mov	cx, @@w
-	shr	cx, 4
-	rep movsw
-	call	egc_off
-	pop	si
-	pop	@@h
-	pop	ds
-	assume ds:_DATA
-	pop	es
-	assume es:nothing
-	add	si, @@stride_packed
-	mov	ax, si
-	shr	ax, 4
-	mov	dx, es
-	add	dx, ax
-	mov	es, dx
-	and	si, 0Fh
-	dec	@@h
-	jnz	short @@put_row
-	retn	8
-pi_put_masked_8_rowloop endp
diff --git a/th05/pi_asm_1.asm b/th05/pi_asm_1.asm
new file mode 100644
index 00000000..6bf3e0cd
--- /dev/null
+++ b/th05/pi_asm_1.asm
@@ -0,0 +1 @@
+include th05/formats/pi_asm_1.asm