diff --git a/gr-sim/gr-sim.c b/gr-sim/gr-sim.c
index a7a54aac..61abdbfb 100644
--- a/gr-sim/gr-sim.c
+++ b/gr-sim/gr-sim.c
@@ -595,3 +595,68 @@ int basic_vlin(int y1, int y2, int at) {
 
 	return 0;
 }
+
+
+short gr_addr_lookup[48]={
+	0x400,0x480,0x500,0x580,0x600,0x680,0x700,0x780,
+	0x428,0x4a8,0x528,0x5a8,0x628,0x6a8,0x728,0x7a8,
+	0x450,0x4d0,0x550,0x5d0,0x650,0x6d0,0x750,0x7d0,
+};
+
+int grsim_put_sprite(unsigned char *sprite_data, int xpos, int ypos) {
+
+	unsigned char i;
+	unsigned char *ptr;
+	short address;
+
+	ptr=sprite_data;
+	x=*ptr;
+	ptr++;
+	ram[CV]=*ptr;
+	ptr++;
+
+	while(1) {
+		address=gr_addr_lookup[ypos/2];
+		address+=xpos;
+		for(i=0;i<x;i++) {
+			a=*ptr;
+			if (a==0) {
+			}
+			else if ((a&0xf0)==0) {
+				ram[address]&=0xf0;
+				ram[address]|=a;
+			}
+			else if ((a&0x0f)==0) {
+				ram[address]&=0x0f;
+				ram[address]|=a;
+			}
+			else {
+				ram[address]=a;
+			}
+			ptr++;
+			address++;
+		}
+		ypos++;
+		ram[CV]--;
+		if (ram[CV]==0) break;
+	}
+
+	return 0;
+}
+
+int gr_copy(short source, short dest) {
+
+	short dest_addr,source_addr;
+	int i,j;
+
+	for(i=0;i<8;i++) {
+		source_addr=gr_addr_lookup[i]+0x400;
+		dest_addr=gr_addr_lookup[i];
+
+		for(j=0;j<120;j++) {
+			ram[dest_addr+j]=ram[source_addr+j];
+		}
+	}
+
+	return 0;
+}
diff --git a/gr-sim/gr-sim.h b/gr-sim/gr-sim.h
index 6de8b7d4..5bd7510b 100644
--- a/gr-sim/gr-sim.h
+++ b/gr-sim/gr-sim.h
@@ -11,4 +11,5 @@ int bload(char *filename, int address);
 int scrn(unsigned char xcoord, unsigned char ycoord);
 int grsim_unrle(unsigned char *rle_data, int address);
 int home(void);
-
+int grsim_put_sprite(unsigned char *sprite_data, int xpos, int ypos);
+int gr_copy(short source, short dest);
diff --git a/gr-sim/tfv.c b/gr-sim/tfv.c
index 55b9c94f..fdcfba36 100644
--- a/gr-sim/tfv.c
+++ b/gr-sim/tfv.c
@@ -32,23 +32,45 @@ static unsigned char title_rle[]=
 
 	};
 
+static unsigned char test_sprite[]={
+	0x8,0x4,
+	0x55,0x50,0x00,0x00,0x00,0x00,0x00,0x00,
+	0x55,0x55,0x55,0x00,0x00,0x00,0x00,0x00,
+	0xff,0x1f,0x4f,0x2f,0xff,0x22,0x20,0x00,
+	0x5f,0x5f,0x5f,0x5f,0xff,0xf2,0xf2,0xf2,
+};
+
 
 int main(int argc, char **argv) {
 
 	int ch;
+	int x,y;
 
 	grsim_init();
 
 	/* Title Screen */
-	grsim_unrle(title_rle,0x400);
+	grsim_unrle(title_rle,0x800);
+	gr_copy(0x800,0x400);
 
 	grsim_update();
 
+	x=20;	y=21;
+	color_equals(0);
+
 	while(1) {
 
 		ch=grsim_input();
 
 		if (ch=='q') break;
+		if (ch=='i') if (y>0) y-=2;
+		if (ch=='m') if (y<39) y+=2;
+		if (ch=='j') if (x>0) x--;
+		if (ch=='k') if (x<39) x++;
+
+		gr_copy(0x800,0x400);
+		grsim_put_sprite(test_sprite,x,y);
+
+		grsim_update();
 
 		usleep(100000);
 	}
diff --git a/tfv/Makefile b/tfv/Makefile
index e68266d5..7718061b 100644
--- a/tfv/Makefile
+++ b/tfv/Makefile
@@ -6,12 +6,12 @@ PNG2GR = ../gr-utils/png2gr
 all:	tfv.dsk
 
 tfv.dsk:	TITLE.GR TFV ED
-	$(DOS33) -y tfv.dsk BSAVE -a 0xc00 TFV
+	$(DOS33) -y tfv.dsk BSAVE -a 0x1000 TFV
 	$(DOS33) -y tfv.dsk BSAVE -a 0x400 TITLE.GR
 	$(DOS33) -y tfv.dsk BSAVE -a 0x900 ED
 
 TFV:	tfv.o
-	ld65 -o TFV tfv.o -C ./apple2_c00.inc
+	ld65 -o TFV tfv.o -C ./apple2_1000.inc
 
 tfv.o:	tfv.s title.inc
 	ca65 -o tfv.o tfv.s -l tfv.lst
@@ -27,3 +27,4 @@ TITLE.GR:	title.png
 
 clean:	
 	rm -f *~ TITLE.GR *.o *.lst ED
+
diff --git a/tfv/apple2_1000.inc b/tfv/apple2_1000.inc
new file mode 100644
index 00000000..e22976de
--- /dev/null
+++ b/tfv/apple2_1000.inc
@@ -0,0 +1,12 @@
+MEMORY {
+       ZP: start = $00, size = $1A, type = rw;
+       RAM: start = $1000, size = $8E00, file = %O;
+}
+
+SEGMENTS {
+CODE: load = RAM, type = ro;
+RODATA: load = RAM, type = ro;
+DATA: load = RAM, type = rw;
+BSS: load = RAM, type = bss, define = yes;
+ZEROPAGE: load = ZP, type = zp;
+}
diff --git a/tfv/tfv.dsk b/tfv/tfv.dsk
index 7dbb5d54..4d50fb8f 100644
Binary files a/tfv/tfv.dsk and b/tfv/tfv.dsk differ
diff --git a/tfv/tfv.s b/tfv/tfv.s
index 7ee52448..c26d4c51 100644
--- a/tfv/tfv.s
+++ b/tfv/tfv.s
@@ -1,14 +1,22 @@
 .define EQU =
 
+KEYPRESS	EQU	$C000
+KEYRESET	EQU	$C010
+
 ;; SOFT SWITCHES
-GR	EQU $C050
-TEXT	EQU $C051
-FULLGR	EQU $C052
-TEXTGR	EQU $C053
-PAGE0	EQU $C054
-PAGE1	EQU $C055
-LORES	EQU $C056
-HIRES	EQU $C057
+GR		EQU	$C050
+TEXT		EQU	$C051
+FULLGR		EQU	$C052
+TEXTGR		EQU	$C053
+PAGE0		EQU	$C054
+PAGE1		EQU	$C055
+LORES		EQU	$C056
+HIRES		EQU	$C057
+
+PADDLE_BUTTON0	EQU	$C061
+PADDL0		EQU	$C064
+PTRIG		EQU	$C070
+
 
 ;; MONITOR ROUTINES
 HLINE	EQU $F819			;; HLINE Y,$2C at A
@@ -38,11 +46,20 @@ H2	EQU $2C
 V2	EQU $2D
 MASK	EQU $2E
 COLOR	EQU $30
-FIRST	EQU $F0
 
-TEMP	EQU $FA
-RUN	EQU $FA
-TEMP2	EQU $FB
+FIRST		EQU	$F0
+LASTKEY		EQU	$F1
+PADDLE_STATUS	EQU	$F2
+XPOS		EQU	$F3
+YPOS		EQU	$F4
+TEMP		EQU	$FA
+RUN		EQU	$FA
+TEMP2		EQU	$FB
+TEMPY		EQU	$FB
+INL		EQU	$FC
+INH		EQU	$FD
+OUTL		EQU	$FE
+OUTH		EQU	$FF
 
 	;=============================
 	; set low-res graphics, page 0
@@ -55,6 +72,10 @@ TEMP2	EQU $FB
 	;=================
 	jsr     CLRTOP
 
+	lda	#$c
+	sta	BASH
+	lda	#$0
+	sta	BASL			; load image off-screen 0xc00
 
 	lda     #>(title_image)
         sta     GBASH
@@ -62,14 +83,71 @@ TEMP2	EQU $FB
         sta     GBASL
 	jsr	load_rle_gr
 
+	jsr	gr_copy
+
+	lda	#20
+	sta	YPOS
+	lda	#20
+	sta	XPOS
+
+main_loop:
+
+	jsr	gr_copy
+
+	jsr	put_sprite
+
+	jsr	wait_until_keypressed
+
+	lda	LASTKEY
+
+	cmp	#('Q')
+        beq	exit
+
+	cmp	#('I')
+	bne	check_down
+	dec	YPOS
+	dec	YPOS
+
+check_down:
+	cmp	#('M')
+	bne	check_left
+	inc	YPOS
+	inc	YPOS
+
+check_left:
+	cmp	#('J')
+	bne	check_right
+	dec	XPOS
+
+check_right:
+	cmp	#('K')
+	bne	check_done
+	inc	XPOS
+
+check_done:
+	jmp	main_loop
+
+
+
+exit:
+
 	lda	#$4
 	sta	BASH
 	lda	#$0
-	sta	BASL			; 0x400 is GR page 0
+	sta	BASL			; restore to 0x400 (page 0)
+					; copy to 0x400 (page 0)
+
+	; call home
+	jsr	HOME
+
 
 	; Return to BASIC?
 	rts
 
+;=====================================================================
+;= ROUTINES
+;=====================================================================
+
 	;=================
 	; load RLE image
 	;=================
@@ -188,4 +266,261 @@ set_gr_page0:
 	bit	GR			; set graphics
 	rts
 
+	;=========================================================
+	; gr_copy
+	;=========================================================
+	; for now copy 0xc00 to 0x400
+	; 2 + 8*38 + 4*80*23 + 4*120*26 + 13 = 20,159 = 20ms = 50Hz
+	; 
+gr_copy:
+	ldx	#0		; set y to zero				; 2
+
+gr_copy_loop:
+	stx	TEMP		; save y				; 3
+	txa			; move to A				; 2
+	asl			; mult by 2				; 2
+	tay			; put into Y				; 2
+	lda	gr_offsets,Y	; lookup low byte for line addr		; 5
+	sta	OUTL		; out and in are the same		; 3
+	sta	INL							; 3
+	lda	gr_offsets+1,Y	; lookup high byte for line addr	; 5
+	sta	OUTH							; 3
+	adc	#$8		; for now, fixed 0xc			; 2
+	sta	INH							; 3
+	ldx	TEMP		; restore y				; 3
+
+	ldy	#0		; set X counter to 0			; 2
+gr_copy_line:
+	lda     (INL),Y		; load a byte				; 5
+	sta     (OUTL),Y	; store a byte				; 6
+	iny			; increment pointer			; 2
+
+	cpx	#$4		; don't want to copy bottom 4*40	; 2
+	bcs	gr_copy_above4						; 3
+
+gr_copy_below4:
+	cpy	#120		; for early ones, copy 120 bytes	; 2
+	bne	gr_copy_line						; 3
+	beq	gr_copy_line_done					; 3
+
+gr_copy_above4:			; for last four, just copy 80 bytes
+	cpy	#80							; 2
+	bne	gr_copy_line						; 3
+
+gr_copy_line_done:
+	inx			; increment y value			; 2
+	cpx	#8		; there are 8 of them			; 2
+	bne	gr_copy_loop	; if not, loop				; 3
+	rts								; 6
+
+	;==========================================================
+	; Wait until keypressed
+	;==========================================================
+	;
+
+wait_until_keypressed:
+	lda	KEYPRESS			; check if keypressed
+	bpl	wait_until_keypressed		; if not, loop
+	jmp     figure_out_key
+
+
+	;==========================================================
+	; Get Key
+	;==========================================================
+	;
+
+get_key:
+
+check_paddle_button:
+
+	; check for paddle button
+
+	bit	PADDLE_BUTTON0
+	bpl	no_button
+	lda	#' '+128
+	jmp	save_key
+
+no_button:
+	lda	KEYPRESS
+	bpl	no_key
+
+figure_out_key:
+	cmp	#' '+128		; the mask destroys space
+	beq	save_key		; so handle it specially
+
+	and	#$5f			; mask, to make upper-case
+check_right_arrow:
+	cmp	#$15
+	bne	check_left_arrow
+	lda	#'K'
+check_left_arrow:
+	cmp	#$08
+	bne	check_up_arrow
+	lda	#'J'
+check_up_arrow:
+	cmp	#$0B
+	bne	check_down_arrow
+	lda	#'I'
+check_down_arrow:
+	cmp	#$0A
+	bne	check_escape
+	lda	#'M'
+check_escape:
+        cmp	#$1B
+	bne	save_key
+	lda	#'Q'
+	jmp	save_key
+
+no_key:
+	bit	PADDLE_STATUS
+	bpl	no_key_store
+
+	; check for paddle action
+	; code from http://web.pdx.edu/~heiss/technotes/aiie/tn.aiie.06.html
+
+	inc	PADDLE_STATUS
+	lda	PADDLE_STATUS
+	and	#$03
+	beq	check_paddles
+	jmp	no_key_store
+
+check_paddles:
+	lda	PADDLE_STATUS
+	and	#$80
+	sta	PADDLE_STATUS
+
+	ldx	#$0
+	LDA	PTRIG		;TRIGGER PADDLES
+	LDY	#0		;INIT COUNTER
+	NOP			;COMPENSATE FOR 1ST COUNT
+	NOP
+PREAD2:	LDA	PADDL0,X	;COUNT EVERY 11 uSEC.
+	BPL	RTS2D		;BRANCH WHEN TIMED OUT
+	INY			;INCREMENT COUNTER
+	BNE	PREAD2		;CONTINUE COUNTING
+	DEY			;COUNTER OVERFLOWED
+RTS2D:				;RETURN W/VALUE 0-255
+
+        cpy	#96
+        bmi	paddle_left
+	cpy	#160
+	bmi	no_key_store
+	lda	#'K'
+	jmp	save_key
+paddle_left:
+	lda	#'J'
+	jmp	save_key
+
+no_key_store:
+	lda	#0			; no key, so save a zero
+
+save_key:
+	sta	LASTKEY			; save the key to our buffer
+	bit	KEYRESET		; clear the keyboard buffer
+	rts
+
+	;=============================================
+	; put_sprite
+	;=============================================
+put_sprite:
+
+	lda	#>tb1_sprite	; hardcoded for tb1 for now
+	sta	INH
+	lda	#<tb1_sprite
+	sta	INL
+
+	ldy	#0		; byte 0 is xsize
+	lda	(INL),Y
+	sta	CH
+	iny
+
+	lda	(INL),Y		; byte 1 is ysize
+	sta	CV
+	iny
+
+	lda	YPOS		; make a copy of ypos
+	sta	TEMPY		; as we modify it
+
+put_sprite_loop:
+	sty	TEMP		; save sprite pointer
+
+	ldy	TEMPY
+	lda	gr_offsets,Y	; lookup low-res memory address		; 5
+	clc
+	adc	XPOS		; add in xpos
+	sta	OUTL		; store out low byte of addy		; 3
+	lda	gr_offsets+1,Y	; look up high byte			; 5
+	sta	OUTH		; and store it out			; 3
+	ldy	TEMP		; restore sprite pointer		; 3
+
+	ldx	CH
+
+put_sprite_pixel:
+	lda	(INL),Y			; get sprite colors
+	iny				; increment sprite pointer
+
+	sty	TEMP			; save sprite pointer
+	ldy	#$0
+
+	cmp	#$0			; if all zero, transparent
+	beq	put_sprite_done_draw	; don't draw it
+
+	sta	COLOR			; save color for later
+	and	#$f0			; check if top nibble zero
+	bne	put_sprite_bottom	; if not skip ahead
+
+	lda	#$0f			; setup mask
+	sta	MASK
+	bmi	put_sprite_mask
+
+put_sprite_bottom:
+	lda	COLOR			; re-load color
+	and	#$0f			; check if bottom nibble zero
+	bne	put_sprite_all		; if not, skip ahead
+	lda	#$0f
+	sta	MASK			; setup mask
+
+put_sprite_mask:
+	lda	(OUTL),Y		; get color at output
+	and	MASK			; mask off unneeded part
+	ora	COLOR			; or the color in
+	sta	(OUTL),Y		; store it back
+
+	jmp	put_sprite_done_draw	; we are done
+
+put_sprite_all:
+	lda	COLOR			; load color
+	sta	(OUTL),Y		; and write it out
+
+
+put_sprite_done_draw:
+
+	ldy	TEMP			; restore sprite pointer
+
+	inc	OUTL			; increment output pointer
+	dex				; decrement x counter
+	bne	put_sprite_pixel	; if not done, keep looping
+
+	inc	TEMPY			; each line has two y vars
+	inc	TEMPY
+	dec	CV			; decemenet total y count
+	bne	put_sprite_loop		; loop if not done
+
+	rts				; return
+
+
+	; waste memory with a lookup table
+	; maybe faster than using GBASCALC?
+gr_offsets:
+	.word	$400,$480,$500,$580,$600,$680,$700,$780
+	.word 	$428,$4a8,$528,$5a8,$628,$6a8,$728,$7a8
+	.word	$450,$4d0,$550,$5d0,$650,$6d0,$750,$7d0
+
+tb1_sprite:
+	.byte $8,$4
+	.byte $55,$50,$00,$00,$00,$00,$00,$00
+	.byte $55,$55,$55,$00,$00,$00,$00,$00
+	.byte $ff,$1f,$4f,$2f,$ff,$22,$20,$00
+	.byte $5f,$5f,$5f,$5f,$ff,$f2,$f2,$f2
+
 .include "title.inc"