diff --git a/megademo/Makefile b/megademo/Makefile
index 10433dc1..e059ce41 100644
--- a/megademo/Makefile
+++ b/megademo/Makefile
@@ -30,6 +30,7 @@ megademo.o:	megademo.s \
 	starring_people.s fs.inc FS_HGRC.BIN.lz4 deater.inc DEATER_HGRC.BIN.lz4\
 	check_email.s email_40_96.inc \
 	bird_mountain.s \
+	waterfall.s gr_unrolled_copy.s waterfall_page1.inc waterfall_page2.inc \
 	takeoff.s takeoff.inc takeoff.img.lz4 \
 	mode7.s \
 	fireworks.s fw_background.inc fw_state_machine.s fw.s \
diff --git a/megademo/bird_mountain.s b/megademo/bird_mountain.s
index 5d7dbfd0..32d63a3d 100644
--- a/megademo/bird_mountain.s
+++ b/megademo/bird_mountain.s
@@ -495,8 +495,8 @@ letters_bm:
 	.byte	3,16,"IS HARD!"
 	.byte	255
 
-.align	$100
-.include "tfv_sprites.inc"
+;.align	$100
+;.include "tfv_sprites.inc"
 
 katahdin:
 .incbin	"KATC.BIN.lz4",11		; skip the header
diff --git a/megademo/gr_unrolled_copy.s b/megademo/gr_unrolled_copy.s
new file mode 100644
index 00000000..4c89fc95
--- /dev/null
+++ b/megademo/gr_unrolled_copy.s
@@ -0,0 +1,648 @@
+	;=========================================================
+	; fast copy rows 22-36 from $C00 to $400
+	;=========================================================
+	;
+	; 6 + 7*8*40 = 2246 cycles
+	; 			6*7*40 = 1680 bytes of code?
+
+gr_copy_row22:
+
+;= y = 22 $5a8 $da8/$9a8 =========================
+	; x=0,y=22
+	lda	$da8
+	sta	$5a8
+	lda	$da9
+	sta	$5a9
+	lda	$daa
+	sta	$5aa
+	lda	$dab
+	sta	$5ab
+	lda	$dac
+	sta	$5ac
+	lda	$dad
+	sta	$5ad
+	lda	$dae
+	sta	$5ae
+	lda	$daf
+	sta	$5af
+
+	; x=8,y=22
+	lda	$db0
+	sta	$5b0
+	lda	$db1
+	sta	$5b1
+	lda	$db2
+	sta	$5b2
+	lda	$db3
+	sta	$5b3
+	lda	$db4
+	sta	$5b4
+	lda	$db5
+	sta	$5b5
+	lda	$db6
+	sta	$5b6
+	lda	$db7
+	sta	$5b7
+
+	; x=16,y=22
+	lda	$db8
+	sta	$5b8
+	lda	$db9
+	sta	$5b9
+	lda	$dba
+	sta	$5ba
+	lda	$dbb
+	sta	$5bb
+	lda	$dbc
+	sta	$5bc
+	lda	$dbd
+	sta	$5bd
+	lda	$dbe
+	sta	$5be
+	lda	$dbf
+	sta	$5bf
+
+	; x=24,y=22
+	lda	$dc0
+	sta	$5c0
+	lda	$dc1
+	sta	$5c1
+	lda	$dc2
+	sta	$5c2
+	lda	$dc3
+	sta	$5c3
+	lda	$dc4
+	sta	$5c4
+	lda	$dc5
+	sta	$5c5
+	lda	$dc6
+	sta	$5c6
+	lda	$dc7
+	sta	$5c7
+
+	; x=32,y=22
+	lda	$dc8
+	sta	$5c8
+	lda	$dc9
+	sta	$5c9
+	lda	$dca
+	sta	$5ca
+	lda	$dcb
+	sta	$5cb
+	lda	$dcc
+	sta	$5cc
+	lda	$dcd
+	sta	$5cd
+	lda	$dce
+	sta	$5ce
+	lda	$dcf
+	sta	$5cf
+
+;= y = 24 $628 $e28/$a28 =========================
+	; x=0,y=24
+	lda	$e28
+	sta	$628
+	lda	$e29
+	sta	$629
+	lda	$e2a
+	sta	$62a
+	lda	$e2b
+	sta	$62b
+	lda	$e2c
+	sta	$62c
+	lda	$e2d
+	sta	$62d
+	lda	$e2e
+	sta	$62e
+	lda	$e2f
+	sta	$62f
+
+	; x=8,y=24
+	lda	$e30
+	sta	$630
+	lda	$e31
+	sta	$631
+	lda	$e32
+	sta	$632
+	lda	$e33
+	sta	$633
+	lda	$e34
+	sta	$634
+	lda	$e35
+	sta	$635
+	lda	$e36
+	sta	$636
+	lda	$e37
+	sta	$637
+
+	; x=16,y=24
+	lda	$e38
+	sta	$638
+	lda	$e39
+	sta	$639
+	lda	$e3a
+	sta	$63a
+	lda	$e3b
+	sta	$63b
+	lda	$e3c
+	sta	$63c
+	lda	$e3d
+	sta	$63d
+	lda	$e3e
+	sta	$63e
+	lda	$e3f
+	sta	$63f
+
+	; x=24,y=24
+	lda	$e40
+	sta	$640
+	lda	$e41
+	sta	$641
+	lda	$e42
+	sta	$642
+	lda	$e43
+	sta	$643
+	lda	$e44
+	sta	$644
+	lda	$e45
+	sta	$645
+	lda	$e46
+	sta	$646
+	lda	$e47
+	sta	$647
+
+	; x=32,y=24
+	lda	$e48
+	sta	$648
+	lda	$e49
+	sta	$649
+	lda	$e4a
+	sta	$64a
+	lda	$e4b
+	sta	$64b
+	lda	$e4c
+	sta	$64c
+	lda	$e4d
+	sta	$64d
+	lda	$e4e
+	sta	$64e
+	lda	$e4f
+	sta	$64f
+
+;= y = 26 $6a8 $ea8/$aa8 =========================
+	; x=0,y=26
+	lda	$ea8
+	sta	$6a8
+	lda	$ea9
+	sta	$6a9
+	lda	$eaa
+	sta	$6aa
+	lda	$eab
+	sta	$6ab
+	lda	$eac
+	sta	$6ac
+	lda	$ead
+	sta	$6ad
+	lda	$eae
+	sta	$6ae
+	lda	$eaf
+	sta	$6af
+
+	; x=8,y=26
+	lda	$eb0
+	sta	$6b0
+	lda	$eb1
+	sta	$6b1
+	lda	$eb2
+	sta	$6b2
+	lda	$eb3
+	sta	$6b3
+	lda	$eb4
+	sta	$6b4
+	lda	$eb5
+	sta	$6b5
+	lda	$eb6
+	sta	$6b6
+	lda	$eb7
+	sta	$6b7
+
+	; x=16,y=26
+	lda	$eb8
+	sta	$6b8
+	lda	$eb9
+	sta	$6b9
+	lda	$eba
+	sta	$6ba
+	lda	$ebb
+	sta	$6bb
+	lda	$ebc
+	sta	$6bc
+	lda	$ebd
+	sta	$6bd
+	lda	$ebe
+	sta	$6be
+	lda	$ebf
+	sta	$6bf
+
+	; x=24,y=26
+	lda	$ec0
+	sta	$6c0
+	lda	$ec1
+	sta	$6c1
+	lda	$ec2
+	sta	$6c2
+	lda	$ec3
+	sta	$6c3
+	lda	$ec4
+	sta	$6c4
+	lda	$ec5
+	sta	$6c5
+	lda	$ec6
+	sta	$6c6
+	lda	$ec7
+	sta	$6c7
+
+	; x=32,y=26
+	lda	$ec8
+	sta	$6c8
+	lda	$ec9
+	sta	$6c9
+	lda	$eca
+	sta	$6ca
+	lda	$ecb
+	sta	$6cb
+	lda	$ecc
+	sta	$6cc
+	lda	$ecd
+	sta	$6cd
+	lda	$ece
+	sta	$6ce
+	lda	$ecf
+	sta	$6cf
+
+;= y = 28 $728 $f28/$b28 =========================
+	; x=0,y=28
+	lda	$f28
+	sta	$728
+	lda	$f29
+	sta	$729
+	lda	$f2a
+	sta	$72a
+	lda	$f2b
+	sta	$72b
+	lda	$f2c
+	sta	$72c
+	lda	$f2d
+	sta	$72d
+	lda	$f2e
+	sta	$72e
+	lda	$f2f
+	sta	$72f
+
+	; x=8,y=28
+	lda	$f30
+	sta	$730
+	lda	$f31
+	sta	$731
+	lda	$f32
+	sta	$732
+	lda	$f33
+	sta	$733
+	lda	$f34
+	sta	$734
+	lda	$f35
+	sta	$735
+	lda	$f36
+	sta	$736
+	lda	$f37
+	sta	$737
+
+	; x=16,y=28
+	lda	$f38
+	sta	$738
+	lda	$f39
+	sta	$739
+	lda	$f3a
+	sta	$73a
+	lda	$f3b
+	sta	$73b
+	lda	$f3c
+	sta	$73c
+	lda	$f3d
+	sta	$73d
+	lda	$f3e
+	sta	$73e
+	lda	$f3f
+	sta	$73f
+
+	; x=24,y=28
+	lda	$f40
+	sta	$740
+	lda	$f41
+	sta	$741
+	lda	$f42
+	sta	$742
+	lda	$f43
+	sta	$743
+	lda	$f44
+	sta	$744
+	lda	$f45
+	sta	$745
+	lda	$f46
+	sta	$746
+	lda	$f47
+	sta	$747
+
+	; x=32,y=28
+	lda	$f48
+	sta	$748
+	lda	$f49
+	sta	$749
+	lda	$f4a
+	sta	$74a
+	lda	$f4b
+	sta	$74b
+	lda	$f4c
+	sta	$74c
+	lda	$f4d
+	sta	$74d
+	lda	$f4e
+	sta	$74e
+	lda	$f4f
+	sta	$74f
+
+;= y = 30 $7a8 $fa8/$ba8 =========================
+	; x=0,y=30
+	lda	$fa8
+	sta	$7a8
+	lda	$fa9
+	sta	$7a9
+	lda	$faa
+	sta	$7aa
+	lda	$fab
+	sta	$7ab
+	lda	$fac
+	sta	$7ac
+	lda	$fad
+	sta	$7ad
+	lda	$fae
+	sta	$7ae
+	lda	$faf
+	sta	$7af
+
+	; x=8,y=30
+	lda	$fb0
+	sta	$7b0
+	lda	$fb1
+	sta	$7b1
+	lda	$fb2
+	sta	$7b2
+	lda	$fb3
+	sta	$7b3
+	lda	$fb4
+	sta	$7b4
+	lda	$fb5
+	sta	$7b5
+	lda	$fb6
+	sta	$7b6
+	lda	$fb7
+	sta	$7b7
+
+	; x=16,y=30
+	lda	$fb8
+	sta	$7b8
+	lda	$fb9
+	sta	$7b9
+	lda	$fba
+	sta	$7ba
+	lda	$fbb
+	sta	$7bb
+	lda	$fbc
+	sta	$7bc
+	lda	$fbd
+	sta	$7bd
+	lda	$fbe
+	sta	$7be
+	lda	$fbf
+	sta	$7bf
+
+	; x=24,y=30
+	lda	$fc0
+	sta	$7c0
+	lda	$fc1
+	sta	$7c1
+	lda	$fc2
+	sta	$7c2
+	lda	$fc3
+	sta	$7c3
+	lda	$fc4
+	sta	$7c4
+	lda	$fc5
+	sta	$7c5
+	lda	$fc6
+	sta	$7c6
+	lda	$fc7
+	sta	$7c7
+
+	; x=32,y=30
+	lda	$fc8
+	sta	$7c8
+	lda	$fc9
+	sta	$7c9
+	lda	$fca
+	sta	$7ca
+	lda	$fcb
+	sta	$7cb
+	lda	$fcc
+	sta	$7cc
+	lda	$fcd
+	sta	$7cd
+	lda	$fce
+	sta	$7ce
+	lda	$fcf
+	sta	$7cf
+
+;= y = 32 $450 $c50/$850 =========================
+	; x=0,y=32
+	lda	$c50
+	sta	$450
+	lda	$c51
+	sta	$451
+	lda	$c52
+	sta	$452
+	lda	$c53
+	sta	$453
+	lda	$c54
+	sta	$454
+	lda	$c55
+	sta	$455
+	lda	$c56
+	sta	$456
+	lda	$c57
+	sta	$457
+
+	; x=8,y=32
+	lda	$c58
+	sta	$458
+	lda	$c59
+	sta	$459
+	lda	$c5a
+	sta	$45a
+	lda	$c5b
+	sta	$45b
+	lda	$c5c
+	sta	$45c
+	lda	$c5d
+	sta	$45d
+	lda	$c5e
+	sta	$45e
+	lda	$c5f
+	sta	$45f
+
+	; x=16,y=32
+	lda	$c60
+	sta	$460
+	lda	$c61
+	sta	$461
+	lda	$c62
+	sta	$462
+	lda	$c63
+	sta	$463
+	lda	$c64
+	sta	$464
+	lda	$c65
+	sta	$465
+	lda	$c66
+	sta	$466
+	lda	$c67
+	sta	$467
+
+	; x=24,y=32
+	lda	$c68
+	sta	$468
+	lda	$c69
+	sta	$469
+	lda	$c6a
+	sta	$46a
+	lda	$c6b
+	sta	$46b
+	lda	$c6c
+	sta	$46c
+	lda	$c6d
+	sta	$46d
+	lda	$c6e
+	sta	$46e
+	lda	$c6f
+	sta	$46f
+
+	; x=32,y=32
+	lda	$c70
+	sta	$470
+	lda	$c71
+	sta	$471
+	lda	$c72
+	sta	$472
+	lda	$c73
+	sta	$473
+	lda	$c74
+	sta	$474
+	lda	$c75
+	sta	$475
+	lda	$c76
+	sta	$476
+	lda	$c77
+	sta	$477
+
+;= y = 34 $4d0 $cd0/$8d0 =========================
+	; x=0,y=34
+	lda	$cd0
+	sta	$4d0
+	lda	$cd1
+	sta	$4d1
+	lda	$cd2
+	sta	$4d2
+	lda	$cd3
+	sta	$4d3
+	lda	$cd4
+	sta	$4d4
+	lda	$cd5
+	sta	$4d5
+	lda	$cd6
+	sta	$4d6
+	lda	$cd7
+	sta	$4d7
+
+	; x=8,y=34
+	lda	$cd8
+	sta	$4d8
+	lda	$cd9
+	sta	$4d9
+	lda	$cda
+	sta	$4da
+	lda	$cdb
+	sta	$4db
+	lda	$cdc
+	sta	$4dc
+	lda	$cdd
+	sta	$4dd
+	lda	$cde
+	sta	$4de
+	lda	$cdf
+	sta	$4df
+
+	; x=16,y=34
+	lda	$ce0
+	sta	$4e0
+	lda	$ce1
+	sta	$4e1
+	lda	$ce2
+	sta	$4e2
+	lda	$ce3
+	sta	$4e3
+	lda	$ce4
+	sta	$4e4
+	lda	$ce5
+	sta	$4e5
+	lda	$ce6
+	sta	$4e6
+	lda	$ce7
+	sta	$4e7
+
+	; x=24,y=34
+	lda	$ce8
+	sta	$4e8
+	lda	$ce9
+	sta	$4e9
+	lda	$cea
+	sta	$4ea
+	lda	$ceb
+	sta	$4eb
+	lda	$cec
+	sta	$4ec
+	lda	$ced
+	sta	$4ed
+	lda	$cee
+	sta	$4ee
+	lda	$cef
+	sta	$4ef
+
+	; x=32,y=34
+	lda	$cf0
+	sta	$4f0
+	lda	$cf1
+	sta	$4f1
+	lda	$cf2
+	sta	$4f2
+	lda	$cf3
+	sta	$4f3
+	lda	$cf4
+	sta	$4f4
+	lda	$cf5
+	sta	$4f5
+	lda	$cf6
+	sta	$4f6
+	lda	$cf7
+	sta	$4f7
+
+	rts								; 6
+
diff --git a/megademo/megademo.s b/megademo/megademo.s
index 49ddc8c6..b5b68853 100644
--- a/megademo/megademo.s
+++ b/megademo/megademo.s
@@ -25,7 +25,7 @@ apple_iie:
 	;===================
 	jsr	HOME
 
-	jsr	rocket_takeoff
+	jsr	waterfall
 
 	; C64 Opening Sequence
 
@@ -49,9 +49,10 @@ apple_iie:
 	; Leaving house
 
 	; Riding bird
-	jsr	bird_mountain
+;	jsr	bird_mountain
 
 	; Waterfall
+	jsr	waterfall
 
 	; Enter ship
 	jsr	rocket_takeoff
@@ -94,9 +95,23 @@ loop_forever:
 .align $100
 	.include	"fireworks.s"
 	.include	"hgr.s"
-	.include	"bird_mountain.s"
+;	.include	"bird_mountain.s"
 	.include	"move_letters.s"
 .align $100
 	.include	"gr_putsprite.s"
 ;	.include	"mode7.s"
 	.include	"takeoff.s"
+	.include	"waterfall.s"
+
+
+;============================
+; Include Sprites
+;============================
+.align $100
+	.include "tfv_sprites.inc"
+
+;============================
+; Include Graphics
+;============================
+
+
diff --git a/megademo/tfv_sprites.inc b/megademo/tfv_sprites.inc
index dee1657f..80c14893 100644
--- a/megademo/tfv_sprites.inc
+++ b/megademo/tfv_sprites.inc
@@ -45,3 +45,32 @@ big_tree:
 	; 00=12 0X=0 X0=4 XX=9	
 	; = 33 + Y*(52)+ [30A + 64B + 69C + 54D]-Y
 	; = 33 + 5*52 + 30*12 + 69*4 + 54*9 - 5 = 1410
+
+bird_rider_stand_left:
+	.byte $6,$7
+	.byte $00,$00,$00,$bd,$dd,$00
+	.byte $90,$dd,$00,$db,$dd,$00
+	.byte $00,$dd,$02,$22,$00,$00
+	.byte $00,$dd,$22,$d2,$dd,$0d
+	.byte $00,$8d,$82,$0d,$00,$00
+	.byte $00,$00,$99,$00,$00,$00
+	.byte $00,$90,$99,$00,$00,$00
+	; 00=22 0X=3 X0=2 XX=15 
+	; = 33 + Y*(52)+ [30A + 64B + 69C + 54D]-Y
+	; = 33 + 7*52 + 30*22 + 64*3 + 69*2 + 54*15 - 7
+	; = 2190
+
+bird_rider_walk_left:
+	.byte $6,$7
+	.byte $00,$00,$00,$bd,$dd,$00
+	.byte $90,$dd,$00,$db,$dd,$00
+	.byte $00,$dd,$02,$22,$00,$00
+	.byte $00,$dd,$22,$d2,$dd,$0d
+	.byte $00,$8d,$82,$0d,$00,$00
+	.byte $00,$00,$99,$00,$00,$00
+	.byte $00,$99,$00,$99,$00,$00
+	; 00=22 0X=3 X0=1 XX=16
+	; = 33 + Y*(52)+ [30A + 64B + 69C + 54D]-Y
+	; = 33 + 7*52 + 30*22 + 64*3 + 69*1 + 54*16 - 7
+	; = 2175
+
diff --git a/megademo/waterfall.s b/megademo/waterfall.s
new file mode 100644
index 00000000..0f19bcb3
--- /dev/null
+++ b/megademo/waterfall.s
@@ -0,0 +1,1047 @@
+; Lo-res cycle-counting water transparency hack
+
+; by deater (Vince Weaver) <vince@deater.net>
+
+; Zero Page
+;FRAMEBUFFER	= $00	; $00 - $0F
+;YPOS		= $10
+YPOS_SIN	= $11
+BIRD_STATE	= $E0
+BIRD_DIR	= $E1
+OLD_XPOS	= $F4
+
+; location we don't care about
+
+DUMMY	= $300
+
+waterfall:
+
+	;===================
+	; init screen
+	jsr	TEXT
+	jsr	HOME
+	bit	PAGE0
+
+	;===================
+	; init vars
+	lda	#0
+	sta	BIRD_DIR
+	sta	BIRD_STATE
+	sta	FRAME
+	sta	OLD_XPOS
+
+	lda	#4
+	sta	DRAW_PAGE
+	sta	XPOS
+
+	;=============================
+	; Load foreground to graphic page1 (apple page2)
+
+	lda	#$0c
+	sta	BASH
+	lda	#$00
+	sta	BASL                    ; load image to $c00
+
+	lda	#<waterfall_page1
+	sta	GBASL
+	lda	#>waterfall_page1
+	sta	GBASH
+	jsr	load_rle_gr
+
+	jsr	gr_copy_to_current	; copy to page1
+
+	;=============================
+	; Load bg to memory
+
+	lda	#$0c
+	sta	BASH
+	lda	#$00
+	sta	BASL                    ; load image to $c00
+
+	lda	#<waterfall_page2
+	sta	GBASL
+	lda	#>waterfall_page2
+	sta	GBASH
+	jsr	load_rle_gr
+
+	;==============================
+	; setup graphics for vapor lock
+	;==============================
+
+	jsr	vapor_lock						; 6
+
+	; vapor lock returns with us at beginning of hsync in line
+	; 114 (7410 cycles), so with 5070 lines to go
+
+	jsr	gr_copy_to_current		; 6+ 9292
+
+	; now we have 322 left
+
+	; GR part
+	bit	LORES							; 4
+	bit	SET_GR							; 4
+	bit	FULLGR							; 4
+
+	; 322 - 12 = 310
+	; -3 for jmp
+	; 307
+
+	; Try X=9 Y=6 cycles=307
+
+	ldy	#6							; 2
+wfloopA:ldx	#9							; 2
+wfloopB:dex								; 2
+        bne	wfloopB							; 2nt/3
+        dey								; 2
+	bne	wfloopA							; 2nt/3
+
+        jmp     wf_display_loop
+
+wf_jump_table:
+	.word	(display_even-1)
+	.word	(display_odd-1)
+	.word	(display_three-1)
+	.word	(display_four-1)
+
+.align  $100
+
+
+	;================================================
+	; Display Loop
+	;================================================
+	; each scan line 65 cycles
+	;       1 cycle each byte (40cycles) + 25 for horizontal
+	;       Total of 12480 cycles to draw screen
+	; Vertical blank = 4550 cycles (70 scan lines)
+	; Total of 17030 cycles to get back to where was
+
+	; jump_table
+	; 		38 + display_odd
+
+	; we have 3 (the jmp back)
+	;		 = 3 cycles that need to be eaten by the vblank
+
+wf_display_loop:
+	inc	FRAME						; 5
+	lda	FRAME						; 3
+	and	#$30						; 2
+
+	; 0011 0000, so more or less div by 16, 3.75Hz
+
+	lsr							; 2
+	lsr							; 2
+	lsr							; 2
+	tay							; 2 ; 18
+
+	lda	wf_jump_table+1,y					; 4
+	pha							; 3
+	lda	wf_jump_table,y					; 4
+	pha							; 3
+	rts							; 6 ; 38
+
+wf_display_loop_return:
+
+vblank:
+	;======================================================
+	; We have 4550 cycles in the vblank, use them wisely
+	;======================================================
+	; do_nothing should be         4550
+	;				 -3 letfover from HBLANK code
+	;				** -9
+	;				-49 check for keypress
+	;			      -2252 copy screen
+	;			      -2231 draw sprite
+	;			=============
+	;			      15 cycles
+
+	; 15 cycles
+	inc	YPOS		; 5
+	nop			; 2
+	nop			; 2
+	lda	YPOS		; 3
+	lda	YPOS		; 3
+
+
+
+	;=========================
+	; Clear background
+	;=========================
+
+	jsr	gr_copy_row22					; 6+ 2246
+							;=========
+							;	2252
+
+
+	;==========================
+	; draw bird sprite
+	;==========================
+	; 					8 prefix
+	; bird_walk_right=	14 + 2175	2206 (need 17)
+	; bird_stand_right=	13 + 2190	2206 (need 3)
+	; bird_walk_left=	15 + 2175	2206 (need 16)
+	; bird_stand_left=	14 + 2190	2206 (need 2)
+	; call to sprite			17 postfix
+	;====================================================
+	;					2231
+
+	lda	BIRD_STATE						; 3
+	and	#1							; 2
+	ldx	BIRD_DIR						; 3
+	bne	bird_left
+
+bird_right:
+									; 2
+	cmp	#1							; 2
+	beq	bird_walk_right
+									; 2
+bird_stand_right:
+	ldx	#>bird_rider_stand_right				; 2
+	ldy	#<bird_rider_stand_right				; 2
+	lda	YPOS ; nop=3
+	jmp	wf_draw_bird						; 3
+bird_walk_right:
+									; 3
+	ldx	#>bird_rider_walk_right					; 2
+	ldy	#<bird_rider_walk_right					; 2
+	jmp	kill_less_time						; 3
+
+bird_left:
+									; 3
+	cmp	#1							; 2
+	beq	bird_walk_left
+
+bird_stand_left:
+									; 2
+	ldx	#>bird_rider_stand_left					; 2
+	ldy	#<bird_rider_stand_left					; 2
+	nop	; nop=2
+	jmp	wf_draw_bird						; 3
+bird_walk_left:
+									; 3
+	ldx	#>bird_rider_walk_left					; 2
+	ldy	#<bird_rider_walk_left					; 2
+
+wf_kill_time:
+	nop		; 2	; need to kill 16
+kill_less_time:
+	lda	YPOS ; nop=3	; need to kill 14
+	lda	YPOS ; nop=3
+	lda	YPOS ; nop=3
+	lda	YPOS ; nop=3
+	nop		; 2
+
+
+	jmp	wf_draw_bird						; 3
+
+
+
+wf_draw_bird:
+	stx	INH					; 3
+	sty	INL					; 3
+
+	lda	#22					; 2
+	sta	YPOS					; 3
+
+	jsr	put_sprite				; 6
+							;=========
+							; 17 + 2190
+
+
+
+	;====================
+	; Handle keypresses
+	; if no keypress, 9
+	; if keypress, 6+43 = 49
+
+	lda	KEYPRESS				; 4
+	bmi	wf_keypress
+							; 2
+wf_no_keypress:
+	; kill 40 cycles
+	ldx	#0					; 2
+	inc	YPOS,X					; 6
+	inc	YPOS,X					; 6
+	inc	YPOS,X					; 6
+	inc	YPOS,X					; 6
+	inc	YPOS,X					; 6
+	inc	YPOS,X					; 6
+	nop						; 2
+
+	jmp	wf_display_loop				; 3
+
+
+	;===================================================
+	; key was pressed handling takes 12 + 20 + 11 = 43
+
+wf_keypress:
+							; 1
+	bit	KEYRESET				; 4
+	inc	BIRD_STATE				; 5
+	and	#$5f		; mask keypress		; 2
+
+is_it_right:
+	cmp	#$15		; right arrow		; 2
+	bne	is_it_left
+							; 2
+	inc	XPOS					; 5
+	lda	#0					; 2
+	sta	BIRD_DIR				; 3
+	lda	YPOS		; 3-cycle nop		; 3
+convoluted:
+	jmp	adjust_xpos				; 3	; 20 if right
+
+is_it_left:
+							; 5
+	cmp	#$8		; left arrow		; 2
+	beq	is_left
+							; 2
+	nop						; 2
+	lda	YPOS		; 3-cycle nop
+	jmp	convoluted				; 3
+							; ------ 20 if neither
+
+is_left:
+							; 3
+	dec	XPOS					; 5	; 20 if left
+	lda	#1					; 2
+	sta	BIRD_DIR				; 3
+
+
+adjust_xpos:
+	lda	XPOS					; 3
+	and	#$1f		; keep in 0-31 range	; 2
+	sta	XPOS					; 3
+
+	jmp	wf_display_loop				; 3
+
+
+.align	$100
+.include "gr_unrolled_copy.s"
+
+.include "waterfall_page1.inc"
+.include "waterfall_page2.inc"
+
+
+.align	$100
+
+	;=================================
+	; Display Even
+	;=================================
+	; we have 65 cycles per line
+	; the first 25 are in hblank
+	; we come in already 21 cycles into things
+	; so the first scanline is a loss (but that's OK)
+
+	; first scanline: 21+ 2 (from ldy) so need to kill 65-23 = 42
+	; second scanline, again kill so 65 killed
+
+display_even:
+
+even_first_four_lines:
+
+	; line 0-3 = 65*4 = 260
+	;		    -38
+	;		     -2
+	;                    -2
+	;                   -25
+	;=========================
+	;		    193
+
+	lda	#193						; 2
+	jsr	delay_a						; 125
+
+	ldy	#4						; 2
+
+
+even_twinkle_stars:
+
+twinkle_loop_even:
+
+	; page1 for 4 lines, 65 - 4 = 61 -2 = 59 - 25 = 34
+
+	; line 0
+	bit	PAGE1						; 4
+	lda	#34						; 2
+	jsr	delay_a						; 25+34
+	; line 1
+	bit	PAGE1						; 4
+	lda	#34						; 2
+	jsr	delay_a						; 25+34
+	; line 2
+	bit	PAGE1						; 4
+	lda	#34						; 2
+	jsr	delay_a						; 25+34
+	; line 3
+	bit	PAGE1						; 4
+	lda	#27						; 2
+	jsr	delay_a						; 25+27
+
+	; below: 7 if not zero
+	; 	 7 (including ldy later) if is zero
+
+
+	dey							; 2
+	beq	twinkle_loop_even_done				; 3
+								;-1
+        jmp	twinkle_loop_even				; 3
+
+
+
+twinkle_loop_even_done:
+
+	ldy	#31						; 2
+falls_loop_even:
+
+; === line 0
+	bit	PAGE0						; 4
+	; delay 31
+	lda	#4						; 2
+	jsr	delay_a						; 25+4
+
+	; delay 11
+	bit	PAGE0						; 4
+	lda	YPOS						; 3
+	bit	PAGE0						; 4
+
+	; delay 19
+	nop							; 2
+	nop							; 2
+	asl	DUMMY						; 6
+	asl	DUMMY						; 6
+	lda	YPOS						; 3
+
+;=== line 1
+	bit	PAGE0						; 4
+	; delay 31
+	lda	#4						; 2
+	jsr	delay_a						; 25+4
+
+	; delay 11
+	bit	PAGE0						; 4
+	lda	YPOS						; 3
+	bit	PAGE0						; 4
+
+	; delay 19
+	nop							; 2
+	nop							; 2
+	asl	DUMMY						; 6
+	asl	DUMMY						; 6
+	lda	YPOS						; 3
+
+;=== line 2
+	bit	PAGE0						; 4
+	; delay 31
+	lda	#4						; 2
+	jsr	delay_a						; 25+4
+
+	; delay 11
+	bit	PAGE1						; 4
+	lda	YPOS						; 3
+	bit	PAGE0						; 4
+	; delay 19
+	nop							; 2
+	nop							; 2
+	asl	DUMMY						; 6
+	asl	DUMMY						; 6
+	lda	YPOS						; 3
+
+;== line 3
+	bit	PAGE0						; 4
+	; delay 31
+	lda	#4						; 2
+	jsr	delay_a						; 25+4
+
+	; delay 11
+	bit	PAGE1						; 4
+	lda	YPOS						; 3
+	bit	PAGE0						; 4 ; 46
+
+	; delay 21 - 7 from loop
+	nop							; 2
+	nop							; 2
+	asl	DUMMY						; 6
+	nop							; 2 ; 58
+
+	dey							; 2
+	beq	falls_loop_even_done				; 3
+								; -1
+	jmp	falls_loop_even					; 3
+falls_loop_even_done:
+
+	ldy	#12						; 2
+
+ground_loop_even:
+
+;==== line 0
+	bit	PAGE0						; 4
+	; delay 61
+	lda	#34						; 2
+	jsr	delay_a						; 25+34
+
+;==== line 1
+	bit	PAGE1						; 4
+	; delay 61
+	lda	#34						; 2
+	jsr	delay_a						; 25+34
+
+;==== line 2
+	bit	PAGE0						; 4
+	; delay 61
+	lda	#34						; 2
+	jsr	delay_a						; 25+34
+
+;=== line 3
+	bit	PAGE1						; 4
+	; delay 54
+	lda	#27						; 2
+	jsr	delay_a						; 25+27
+
+	dey							; 2
+	beq	ground_loop_even_done				; 3
+								; -1
+	jmp	ground_loop_even				; 3
+ground_loop_even_done:
+	nop							; 2
+
+	jmp	wf_display_loop_return				; 3
+
+.align	$100
+
+	;=================================
+	; Display Odd
+	;=================================
+	; we have 65 cycles per line
+	; the first 25 are in hblank
+
+display_odd:
+	; first scanline: comes in with 38
+
+odd_first_four_lines:
+
+	; line 0-3 = 65*4 = 260
+	;		    -38
+	;		     -2
+	;                    -2
+	;                   -25
+	;=========================
+	;		    193
+
+	lda	#193						; 2
+	jsr	delay_a						; 125
+
+	ldy	#4						; 2
+
+odd_twinkle_stars:
+
+twinkle_loop_odd:
+
+	; page1 for 4 lines, 65 - 4 = 61 -2 = 59 - 25 = 34
+
+	; line 0
+	bit	PAGE1						; 4
+	lda	#34						; 2
+	jsr	delay_a						; 25+34
+
+	; line 1
+	bit	PAGE1						; 4
+	lda	#34						; 2
+	jsr	delay_a						; 25+34
+
+	; line 2
+	bit	PAGE1						; 4
+	lda	#34						; 2
+	jsr	delay_a						; 25+34
+
+	; line 3
+	bit	PAGE1						; 4
+	lda	#27						; 2
+	jsr	delay_a						; 25+27
+
+	dey							; 2
+	beq	twinkle_loop_odd_done				; 3
+								; -1
+	jmp	twinkle_loop_odd				; 3
+twinkle_loop_odd_done:
+
+	ldy	#31						; 2
+falls_loop_odd:
+
+;=== line 0
+	bit	PAGE0						; 4
+	; delay 31
+	lda	#4						; 2
+	jsr	delay_a						; 25+4
+
+	; falls 11
+	bit	PAGE1						; 4
+	lda	YPOS						; 3
+	bit	PAGE0						; 4
+
+	; delay 19
+	nop							; 2
+	nop							; 2
+	asl	DUMMY						; 6
+	asl	DUMMY						; 6
+	lda	YPOS						; 3
+
+;==== line 1
+	bit	PAGE0						; 4
+	; delay 31
+	lda	#4						; 2
+	jsr	delay_a						; 25+4
+
+	; falls 8
+	bit	PAGE0						; 4
+	lda	YPOS						; 3
+	bit	PAGE0						; 4
+
+	; delay 19
+	nop							; 2
+	nop							; 2
+	asl	DUMMY						; 6
+	asl	DUMMY						; 6
+	lda	YPOS						; 3
+
+;=== line 2
+	bit	PAGE0						; 4
+	; delay 31
+	lda	#4						; 2
+	jsr	delay_a						; 25+4
+
+	; falls 8
+	bit	PAGE0						; 4
+	lda	YPOS						; 3
+	bit	PAGE0						; 4
+
+	; delay 19
+	nop							; 2
+	nop							; 2
+	asl	DUMMY						; 6
+	asl	DUMMY						; 6
+	lda	YPOS						; 3
+
+;===line 4
+	bit	PAGE0						; 4
+
+	; delay 31
+	lda	#4						; 2
+	jsr	delay_a						; 25+4
+
+	; falls 11
+	bit	PAGE1						; 4
+	lda	YPOS						; 3
+	bit	PAGE0						; 4 ; 44
+
+	; end falls
+	; delay 21 - 7 from loop
+	nop							; 2
+	nop							; 2
+	asl	DUMMY						; 6
+	nop							; 2 ; 58
+
+	dey							; 2
+	beq	falls_loop_odd_done				; 3
+								;-1
+	jmp	falls_loop_odd					; 3
+falls_loop_odd_done:
+								; 3
+	ldy	#12						; 2
+
+ground_loop_odd:
+
+;=== line 0
+	bit	PAGE1						; 4
+	; delay 61
+	lda	#34						; 2
+	jsr	delay_a						; 25+34
+
+;=== line 1
+	bit	PAGE0						; 4
+	; delay 61
+	lda	#34						; 2
+	jsr	delay_a						; 25+34
+
+;==== line 2
+	bit	PAGE1						; 4
+	; delay 61
+	lda	#34						; 2
+	jsr	delay_a						; 25+34
+
+;==== line 3
+	bit	PAGE0						; 4
+	; delay 54
+	lda	#27						; 2
+	jsr	delay_a						; 25+27
+
+	dey							; 2
+	beq	ground_loop_odd_done				; 3
+								;-1
+	jmp	ground_loop_odd					; 3
+ground_loop_odd_done:
+
+
+								; 3
+	nop							; 2
+
+	jmp	wf_display_loop_return				; 3
+
+	rts							; 6
+
+
+.align	$100
+
+	;=================================
+	; Display 3
+	;=================================
+	; we have 65 cycles per line
+	; the first 25 are in hblank
+	; we come in already 21 cycles into things
+	; so the first scanline is a loss (but that's OK)
+
+	; first scanline: 21+ 2 (from ldy) so need to kill 65-23 = 42
+	; second scanline, again kill so 65 killed
+
+display_three:
+
+three_first_four_lines:
+
+	; line 0-3 = 65*4 = 260
+	;		    -38
+	;		     -2
+	;                    -2
+	;                   -25
+	;=========================
+	;		    193
+
+	lda	#193						; 2
+	jsr	delay_a						; 125
+
+	ldy	#4						; 2
+
+
+three_twinkle_stars:
+
+twinkle_loop_three:
+
+	; line 0
+	bit	PAGE0						; 4
+	lda	#34						; 2
+	jsr	delay_a						; 25+34
+	; line 1
+	bit	PAGE1						; 4
+	lda	#34						; 2
+	jsr	delay_a						; 25+34
+	; line 2
+	bit	PAGE1						; 4
+	lda	#34						; 2
+	jsr	delay_a						; 25+34
+
+	; line 4
+	bit	PAGE0						; 4
+	lda	#27						; 2
+	jsr	delay_a						; 25+27
+
+	dey							; 2
+	beq	twinkle_loop_three_done				; 3
+								; -1
+	jmp	twinkle_loop_three				; 3
+
+
+twinkle_loop_three_done:
+
+	ldy	#31						; 2
+falls_loop_three:
+
+;=== line 0
+	bit	PAGE0						; 4
+	; delay 31
+	lda	#4						; 2
+	jsr	delay_a						; 25+4
+
+	; delay 11
+	bit	PAGE1						; 4
+	lda	YPOS						; 3
+	bit	PAGE0						; 4
+
+	; delay 19
+	nop							; 2
+	nop							; 2
+	asl	DUMMY						; 6
+	asl	DUMMY						; 6
+	lda	YPOS						; 3
+
+;=== line 1
+	bit	PAGE0						; 4
+	; delay 31
+	lda	#4						; 2
+	jsr	delay_a						; 25+4
+
+	; delay 11
+	bit	PAGE1						; 4
+	lda	YPOS						; 3
+	bit	PAGE0						; 4
+
+	; delay 19
+	nop
+	nop
+	asl	DUMMY						; 6
+	asl	DUMMY						; 6
+	lda	YPOS						; 3
+
+;=== line 2
+	bit	PAGE0						; 4
+	; delay 31
+	lda	#4						; 2
+	jsr	delay_a						; 25+4
+
+	; delay 11
+	bit	PAGE0						; 4
+	lda	YPOS						; 3
+	bit	PAGE0						; 4
+
+	; delay 19
+	nop							; 2
+	nop							; 2
+	asl	DUMMY						; 6
+	asl	DUMMY						; 6
+	lda	YPOS						; 3
+
+;=== line 4
+	bit	PAGE0						; 4
+	; delay 31
+	lda	#4						; 2
+	jsr	delay_a						; 25+4
+
+	; delay 11
+	bit	PAGE0						; 4
+	lda	YPOS						; 3
+	bit	PAGE0						; 4 ; 44
+
+	; delay 21 - 7 from loop
+	nop							; 2
+	nop							; 2
+	asl	DUMMY						; 6
+	nop							; 2 ; 58
+
+	dey							; 2
+	beq	falls_loop_three_done				; 3
+								; -1
+	jmp	falls_loop_three				; 3
+falls_loop_three_done:
+	ldy	#12						; 2
+
+ground_loop_three:
+
+;=== line 0
+	bit	PAGE1						; 4
+	; delay 61
+	lda	#34						; 2
+	jsr	delay_a						; 25+34
+
+;=== line 1
+	bit	PAGE0						; 4
+	; delay 61
+	lda	#34						; 2
+	jsr	delay_a						; 25+34
+
+;=== line 2
+	bit	PAGE1						; 4
+	; delay 61
+	lda	#34						; 2
+	jsr	delay_a						; 25+34
+
+;=== line 3
+	bit	PAGE0						; 4
+	; delay 54
+	lda	#27						; 2
+	jsr	delay_a						; 25+27
+
+	dey							; 2
+	beq	ground_loop_three_done				; 3
+								; -1
+	jmp	ground_loop_three				; 3
+ground_loop_three_done:
+
+	nop							; 2
+
+	jmp	wf_display_loop_return				; 3
+
+
+.align	$100
+
+	;=================================
+	; Display Four
+	;=================================
+	; we have 65 cycles per line
+	; the first 25 are in hblank
+	; we come in already 21 cycles into things
+	; so the first scanline is a loss (but that's OK)
+
+display_four:
+
+	; first scanline comes in with 38
+
+four_first_four_lines:
+	; line 0-3 = 65*4 = 260
+	;                   -38
+	;                    -2
+	;                    -2
+	;                   -25
+	;=========================
+	;                   193
+
+	lda	#193						; 2
+	jsr	delay_a						; 25+193
+
+	ldy     #4                                              ; 2
+
+four_twinkle_stars:
+
+twinkle_loop_four:
+
+;=== line 0
+	bit	PAGE1						; 4
+	lda	#34						; 2
+	jsr	delay_a						; 25+34
+
+;=== line 1
+	bit	PAGE1						; 4
+	lda	#34						; 2
+	jsr	delay_a						; 25+34
+
+;=== line 2
+	bit	PAGE1						; 4
+	lda	#34						; 2
+	jsr	delay_a						; 25+34
+
+
+;=== line 3
+	bit	PAGE1						; 4
+	lda	#27						; 2
+	jsr	delay_a						; 25+27
+
+
+	dey							; 2
+	beq	twinkle_loop_four_done				; 3
+								; -1
+	jmp	twinkle_loop_four				; 3
+twinkle_loop_four_done:
+
+	ldy	#31						; 2
+falls_loop_four:
+
+;=== line 0
+	bit	PAGE0						; 4
+	; delay 31
+	lda	#4						; 2
+	jsr	delay_a						; 25+4
+
+	; delay 11
+	bit	PAGE0						; 4
+	lda	YPOS						; 3
+	bit	PAGE0						; 4
+
+	; delay 19
+	nop							; 2
+	nop							; 2
+	asl	DUMMY						; 6
+	asl	DUMMY						; 6
+	lda	YPOS						; 3
+
+;=== line 1
+	bit	PAGE0						; 4
+	; delay 31
+	lda	#4						; 2
+	jsr	delay_a						; 25+4
+
+	; delay 11
+	bit	PAGE1						; 4
+	lda	YPOS						; 3
+	bit	PAGE0						; 4
+
+	; delay 19
+	nop							; 2
+	nop							; 2
+	asl	DUMMY						; 6
+	asl	DUMMY						; 6
+	lda	YPOS						; 3
+
+;=== line 2
+	bit	PAGE0						; 4
+	; delay 31
+	lda	#4						; 2
+	jsr	delay_a						; 25+4
+
+	; delay 11
+	bit	PAGE1						; 4
+	lda	YPOS						; 3
+	bit	PAGE0						; 4
+
+	; delay 19
+	nop							; 2
+	nop							; 2
+	asl	DUMMY						; 6
+	asl	DUMMY						; 6
+	lda	YPOS						; 3
+
+;=== line 3
+	bit	PAGE0						; 4
+	; delay 31
+	lda	#4						; 2
+	jsr	delay_a						; 25+4
+
+	; delay 11
+	bit	PAGE0						; 4
+	lda	YPOS						; 3
+	bit	PAGE0						; 4 ; 44
+
+	; delay 21 - 7 from loop
+	nop							; 2
+	nop							; 2
+	asl	DUMMY						; 6
+	nop							; 2 ; 58
+
+	dey							; 2
+	beq	falls_loop_four_done				; 3
+								; -1
+	jmp	falls_loop_four					; 3
+falls_loop_four_done:
+	ldy	#12						; 2
+
+ground_loop_four:
+
+;=== line 0
+	bit	PAGE0						; 4
+	; delay 61
+	lda	#34						; 2
+	jsr	delay_a						; 25+34
+
+;=== line 1
+	bit	PAGE1						; 4
+	; delay 61
+	lda	#34						; 2
+	jsr	delay_a						; 25+34
+
+;=== line 2
+	bit	PAGE0						; 4
+	; delay 61
+	lda	#34						; 2
+	jsr	delay_a						; 25+34
+
+
+;=== line 3
+	bit	PAGE1						; 4
+	; delay 54
+	lda	#27						; 2
+	jsr	delay_a						; 25+27
+
+	dey							; 2
+	beq	ground_loop_four_done				; 3
+								; -1
+	jmp	ground_loop_four				; 3
+ground_loop_four_done:
+	nop							; 2
+
+	jmp	wf_display_loop_return				; 3
+
+
diff --git a/megademo/waterfall_page1.inc b/megademo/waterfall_page1.inc
new file mode 100644
index 00000000..d854dfbf
--- /dev/null
+++ b/megademo/waterfall_page1.inc
@@ -0,0 +1,52 @@
+waterfall_page1:	.byte $28 ; ysize=48
+	.byte $AD,$00, $F0, $A9,$00, $04, $A0,$10,$44, $00, $0F
+	.byte $AC,$00, $50,$50, $A3,$00, $80, $A3,$00, $A3,$54, $AE,$44
+	.byte $A6,$00, $07, $00, $A3,$50, $55, $85,$85, $88,$88
+	.byte $EE, $77, $EE, $88, $EE, $77, $EE
+	.byte $A3,$88, $85,$85, $55, $A7,$85, $55, $A3,$85, $A4,$50
+	.byte $55, $A3,$85, $A3,$88, $55, $A4,$88, $EE, $77
+	.byte $EE, $78, $EE, $77, $EE, $A5,$88, $55
+	.byte $A7,$88, $55, $A7,$88, $55, $A6,$88, $55, $A4,$88
+	.byte $EE, $77, $EE, $77, $EE, $77, $EE
+	.byte $A5,$88, $55, $A7,$88, $55, $A7,$88, $55, $A6,$88
+	.byte $55, $A4,$88, $EE, $77, $EE, $77, $EE
+	.byte $77, $EE, $A5,$88, $55, $A6,$88, $55, $A8,$88
+	.byte $55, $A6,$88, $55, $A4,$88, $EE, $77, $EE
+	.byte $77, $EE, $77, $EE, $A5,$88, $55, $A6,$88
+	.byte $55, $A8,$88, $85, $58, $A5,$88, $55, $A4,$88
+	.byte $EE, $77, $EE, $77, $EE, $77, $EE
+	.byte $A4,$88, $58, $85, $A6,$88, $55, $A9,$88, $55
+	.byte $A5,$88, $55, $A4,$88, $EE, $77, $EE, $77
+	.byte $EE, $77, $EE, $A4,$88, $55, $A7,$88, $55
+	.byte $A9,$88, $55, $A6,$88, $55, $A3,$88, $EE, $77
+	.byte $EE, $77, $EE, $77, $EE, $A4,$88, $55
+	.byte $A6,$88, $58, $85, $A9,$88, $55, $A6,$88, $55
+	.byte $A3,$88, $EE, $77, $EE, $77, $EE, $77
+	.byte $EE, $A4,$88, $55, $A6,$88, $55, $AA,$88, $55
+	.byte $A6,$88, $55, $A3,$88, $EE, $77, $EE, $77
+	.byte $EE, $77, $EE, $A4,$88, $55, $A6,$88, $55
+	.byte $A4,$88, $55, $A5,$88, $85, $58, $A5,$88, $55
+	.byte $A3,$88, $EE, $77, $EE, $77, $EE, $77
+	.byte $EE, $A4,$88, $55, $A6,$88, $55, $A4,$88, $55
+	.byte $A6,$88, $55, $A5,$88, $55, $A3,$88, $EE, $77
+	.byte $EE, $77, $EE, $77, $EE, $A3,$88, $58
+	.byte $85, $A6,$88, $55, $A4,$88, $55, $A6,$88, $55
+	.byte $A5,$88, $55, $A3,$88, $EE, $77, $EE, $77
+	.byte $EE, $77, $EE, $A3,$88, $55, $A7,$88, $55
+	.byte $A4,$88, $55, $A6,$88, $55, $A5,$88, $55, $88
+	.byte $58,$58, $EE, $77, $EE, $77, $EE, $77
+	.byte $EE, $A3,$58, $55, $A7,$88, $55, $A3,$88, $58
+	.byte $85, $A6,$88, $55, $88,$88, $A3,$58, $A4,$55, $EE
+	.byte $77, $EE, $77, $EE, $77, $EE, $A6,$55
+	.byte $A4,$58, $88, $55, $A3,$88, $55, $88, $A4,$58
+	.byte $AA,$55, $F5, $FF,$FF, $77, $EE, $77, $EE
+	.byte $F7, $FE, $F5, $AB,$55, $A3,$58, $55, $58
+	.byte $AD,$55, $A3,$FF, $FE, $FF, $8E, $88, $8E
+	.byte $A4,$FF, $A0,$1D,$55, $6F, $EF,$EF, $6F, $6E, $68
+	.byte $A4,$FF, $A0,$1C,$55, $65, $A3,$E6, $66,$66, $6E,$6E, $66
+	.byte $EF, $E6, $66, $A0,$1A,$55, $65, $A3,$E6, $66,$66
+	.byte $A4,$6E, $A5,$E6, $65, $A0,$16,$55, $65, $A4,$E6, $66
+	.byte $A5,$6E, $66, $E6, $86, $88, $86, $E6
+	.byte $66, $A0,$14,$55, $65, $A5,$E6, $66, $A5,$6E, $66
+	.byte $A7,$E6, $66, $AD,$55
+	.byte $A1
diff --git a/megademo/waterfall_page2.inc b/megademo/waterfall_page2.inc
new file mode 100644
index 00000000..0fc680a4
--- /dev/null
+++ b/megademo/waterfall_page2.inc
@@ -0,0 +1,35 @@
+waterfall_page2:	.byte $28 ; ysize=48
+	.byte $AD,$00, $50, $A9,$00, $04, $A0,$10,$44, $00, $05
+	.byte $AC,$00, $50,$50, $A3,$00, $80, $A3,$00, $A3,$54, $AE,$44
+	.byte $A6,$00, $05, $00, $A3,$50, $55, $85,$85, $88,$88
+	.byte $8E, $87, $8E, $88, $8E, $87, $8E
+	.byte $A3,$88, $85,$85, $55, $A7,$85, $55, $A3,$85, $A4,$50
+	.byte $55, $A3,$85, $A3,$88, $55, $A0,$10,$88, $55, $A7,$88
+	.byte $55, $A7,$88, $55, $A6,$88, $55, $A0,$10,$88, $55
+	.byte $A7,$88, $55, $A7,$88, $55, $A6,$88, $55, $A0,$10,$88
+	.byte $55, $A6,$88, $55, $A8,$88, $55, $A6,$88, $55
+	.byte $A0,$10,$88, $55, $A6,$88, $55, $A8,$88, $85, $58
+	.byte $A5,$88, $55, $AF,$88, $58, $85, $A6,$88, $55
+	.byte $A9,$88, $55, $A5,$88, $55, $AF,$88, $55, $A7,$88
+	.byte $55, $A9,$88, $55, $A6,$88, $55, $AE,$88, $55
+	.byte $A6,$88, $58, $85, $A9,$88, $55, $A6,$88, $55
+	.byte $AE,$88, $55, $A6,$88, $55, $AA,$88, $55, $A6,$88
+	.byte $55, $AE,$88, $55, $A6,$88, $55, $A4,$88, $55
+	.byte $A5,$88, $85, $58, $A5,$88, $55, $AE,$88, $55
+	.byte $A6,$88, $55, $A4,$88, $55, $A6,$88, $55, $A5,$88
+	.byte $55, $AD,$88, $58, $85, $A6,$88, $55, $A4,$88
+	.byte $55, $A6,$88, $55, $A5,$88, $55, $AD,$88, $55
+	.byte $A7,$88, $55, $A4,$88, $55, $A6,$88, $55, $A5,$88
+	.byte $55, $88, $AC,$58, $55, $A7,$88, $55, $A3,$88
+	.byte $58, $85, $A6,$88, $55, $88,$88, $A3,$58, $A0,$11,$55
+	.byte $A4,$58, $88, $55, $A3,$88, $55, $88, $A4,$58
+	.byte $A0,$1F,$55, $A3,$58, $55, $58, $AE,$55, $F5, $FF
+	.byte $F5, $FF, $85, $88, $85, $FF,$FF, $5F
+	.byte $A0,$1E,$55, $A4,$6F, $E6, $E8, $A4,$FF, $A0,$1C,$55, $65
+	.byte $66, $A3,$6E, $A3,$E6, $66, $6F, $6E, $66
+	.byte $A0,$1A,$55, $65, $66, $A3,$6E, $A4,$E6, $66,$66, $A3,$6E
+	.byte $66, $65, $A0,$16,$55, $65, $66, $A3,$6E, $66
+	.byte $A5,$E6, $66, $6E, $8E, $88, $8E, $6E
+	.byte $66, $A0,$14,$55, $65, $66, $A4,$6E, $66, $A5,$E6
+	.byte $66, $A6,$6E, $66,$66, $AD,$55
+	.byte $A1