diff --git a/mos-platform/nes/nesdoug/CMakeLists.txt b/mos-platform/nes/nesdoug/CMakeLists.txt
index 6d4f4b483..9800f35df 100644
--- a/mos-platform/nes/nesdoug/CMakeLists.txt
+++ b/mos-platform/nes/nesdoug/CMakeLists.txt
@@ -5,9 +5,13 @@ install(FILES
 TYPE INCLUDE)
 
 add_platform_library(nes-nesdoug
+  metatile.c
+  metatile.s
   nesdoug.s
   nesdoug.c
   padlib.s
+  vram_buffer.c
+  vram_buffer.s
   zaplib.s
 )
 target_include_directories(nes-nesdoug BEFORE PUBLIC ..)
diff --git a/mos-platform/nes/nesdoug/metatile.c b/mos-platform/nes/nesdoug/metatile.c
new file mode 100644
index 000000000..56cbffaf7
--- /dev/null
+++ b/mos-platform/nes/nesdoug/metatile.c
@@ -0,0 +1,15 @@
+// Copyright 2022 LLVM-MOS Project
+// Licensed under the Apache License, Version 2.0 with LLVM Exceptions.
+// See /~https://github.com/llvm-mos/llvm-mos-sdk/blob/main/LICENSE for license
+// information.
+
+// Keeping these functions in C LTOs them in, which informs LTO code generation
+// that the ZP regions aren't available. This saves users of the library from
+// having to manually reserve ZP space from LTO.
+__attribute__((section(".zp.meta_ptr"))) const char *META_PTR;
+__attribute__((section(".zp.data_ptr"))) const char *DATA_PTR;
+
+void set_data_pointer(const void *data) { DATA_PTR = (const char *)data; }
+void set_mt_pointer(const void *metatiles) {
+  META_PTR = (const char *)metatiles;
+}
diff --git a/mos-platform/nes/nesdoug/metatile.s b/mos-platform/nes/nesdoug/metatile.s
new file mode 100644
index 000000000..6a6e3358d
--- /dev/null
+++ b/mos-platform/nes/nesdoug/metatile.s
@@ -0,0 +1,270 @@
+;written by Doug Fraker
+;version 1.2, 1/1/2022
+
+.zeropage VRAM_BUF, META_PTR, DATA_PTR
+
+
+
+;void buffer_4_mt(int ppu_address, char index);
+.section .text.buffer_4_mt,"ax",@progbits
+.globl buffer_4_mt
+buffer_4_mt:
+	sta __rc3
+	lda __rc2
+	;a is the index into the data, get 4 metatiles
+
+	and #$ee ;sanitize, x and y should be even
+	tay
+	lda (DATA_PTR), y
+	sta __rc6
+	iny
+	lda (DATA_PTR), y
+	sta __rc7
+		tya
+		clc
+		adc #15
+			cmp #$f0 ;too far, data set only 240 bytes
+			bcs .Lskip
+		tay
+	lda (DATA_PTR), y
+	sta __rc8
+	iny
+	lda (DATA_PTR), y
+	sta __rc9
+.Lskip:
+;metatiles are in _rc6 - rc9 now
+	lda __rc3
+	and #$9c ;sanitize, should be top left
+	sta __rc11
+	stx __rc12 ;save for later, ppu_address
+
+	sta __rc4
+	txa
+	ora #$40	;NT_UPD_HORZ
+	sta __rc5
+
+;buffer the ppu_address
+
+	lda #0
+	sta __rc10 ;loop count, index to the metatiles
+	ldx VRAM_INDEX
+.Lbuffer_4_mt_loop:
+	lda __rc4 ;low byte
+	sta VRAM_BUF+1, x
+	lda __rc5 ;high byte
+	sta VRAM_BUF,x
+	jsr .Lsub1 ;adds $20 to the address for next time
+
+		lda #4 ;tell the system 4 bytes in a row
+		sta VRAM_BUF+2,x
+		sta VRAM_BUF+9,x ;loops twice, so, it does this twice
+
+	jsr .Lsub2 ;gets y is which metatile
+
+	lda (META_PTR), y
+	sta VRAM_BUF+3,x ;		buffer the 4 tiles
+	iny
+	lda (META_PTR), y
+	sta VRAM_BUF+4,x
+	iny
+	lda (META_PTR), y
+	sta VRAM_BUF+10,x
+	iny
+	lda (META_PTR), y
+	sta VRAM_BUF+11,x
+	jsr .Lsub4 ;get attrib bits, shift into place
+
+;same, but for right side
+	lda __rc4 ;low byte ppu address, again
+	sta VRAM_BUF+8,x
+	lda __rc5 ;high byte
+	sta VRAM_BUF+7,x
+	jsr .Lsub1
+
+	inc __rc10 ;count and index
+	jsr .Lsub2
+	lda (META_PTR), y
+	sta VRAM_BUF+5,x ;		buffer the 4 tiles
+	iny
+	lda (META_PTR), y
+	sta VRAM_BUF+6,x
+	iny
+	lda (META_PTR), y
+	sta VRAM_BUF+12,x
+	iny
+	lda (META_PTR), y
+	sta VRAM_BUF+13,x
+	jsr .Lsub4
+
+	txa ;adjust the vram index to the next set
+	clc
+	adc #14
+	tax
+
+	jsr .Lsub3 ;check if lowest y on screen, skip the la
+	bne .Lloop_done
+
+	inc __rc10
+	ldy __rc10
+	cpy #4
+	bcc .Lbuffer_4_mt_loop
+
+.Lloop_done:
+
+
+
+;now push 1 attribute byte to the vram buffer
+;first, shift the bits to get an attribute address
+;we stored the original at TEMP+7,8, 8 is high byte
+;a bunch of bit shifting to get 3 bits from x and 3 y
+	lsr __rc12 ;high byte
+	ror __rc11
+	lsr __rc12
+	ror __rc11
+	lda __rc11
+	pha ;save
+	and #7 ;just the x bits
+	sta __rc4
+	pla
+	lsr a ;just the y bits
+	lsr a
+	and #$f8
+	ora #$c0
+	ora __rc4
+	sta __rc4 ;low byte
+;now high byte
+	lda __rc12
+	asl a
+	asl a
+	ora #$23
+	sta __rc5 ;high byte, and the low byte is in TEMP
+
+
+;finally, push it all to the vram_buffer as a single byte
+	lda __rc5 ;high byte
+	sta VRAM_BUF,x
+	inx
+	lda __rc4 ;low byte
+	sta VRAM_BUF,x
+	inx
+	lda __rc14
+	sta VRAM_BUF,x
+	inx
+	lda #$ff ;=NT_UPD_EOF
+	sta VRAM_BUF,x
+	stx VRAM_INDEX
+	rts
+
+
+.Lsub1:	;add $20 is a 1 down on the screen
+	tay ;high byte
+	lda __rc4
+	clc
+	adc #$20
+	sta __rc4
+	bcc .Lsub1b
+	iny
+.Lsub1b:
+	sty __rc5
+	rts
+
+
+.Lsub2:	;get the next metatile offset
+	ldy __rc10
+	lda __rc6, y ;metatile
+;multiply by 5
+	sta __rc13
+	asl a
+	asl a ;x4 = 4 bytes per
+	clc
+	adc __rc13
+	tay
+	rts
+
+
+.Lsub3: ;check make sure we're not at the lowest y and overflowing
+	ldy #0 ;x is forbidden
+	lda __rc12 ;high byte
+	and #$03
+	cmp #$03
+	bne .Lnot_overflow
+	lda __rc11
+	cmp #$80 ;last row of mt
+	bcc .Lnot_overflow
+	iny
+	lsr __rc14 ;make sure the attrib bits in correct position
+	lsr __rc14
+	lsr __rc14
+	lsr __rc14
+.Lnot_overflow:
+	tya ;set flag
+	rts
+
+
+.Lsub4: ;get attrib bits, roll them in place
+	iny
+	lda (META_PTR), y ;5th byte = attribute
+	and #3 ;just need 2 bits
+	ror a ;bit to carry
+	ror __rc14 ;shift carry in
+	ror a ;bit to carry
+	ror __rc14 ;roll the a.t. bits in the high 2 bits
+	rts
+
+
+
+
+;void buffer_1_mt(int ppu_address, char metatile);
+.section .text.buffer_1_mt,"ax",@progbits
+.globl buffer_1_mt
+buffer_1_mt:
+	and #$de ;sanitize, should be even x and y
+	sta __rc4
+	txa
+	ora #$40 ;NT_UPD_HORZ
+	sta __rc5
+
+	ldx VRAM_INDEX
+	lda __rc4				;ppu address
+	sta VRAM_BUF+1,x
+		clc
+		adc #$20 ;shouldn't be rollover
+	sta VRAM_BUF+6,x
+
+	lda __rc5
+	sta VRAM_BUF,x
+	sta VRAM_BUF+5,x
+
+	lda #2 ;tell the system 2 bytes in a row
+	sta VRAM_BUF+2,x
+	sta VRAM_BUF+7,x
+
+	lda __rc2 ;which metatile
+	asl a
+	asl a
+	clc
+	adc __rc2 ;multiply 5
+	tay
+	lda (META_PTR), y ;tile
+	sta VRAM_BUF+3,x
+	iny
+	lda (META_PTR), y ;tile
+	sta VRAM_BUF+4,x
+	iny
+	lda (META_PTR), y ;tile
+	sta VRAM_BUF+8,x
+	iny
+	lda (META_PTR), y ;tile
+	sta VRAM_BUF+9,x
+
+	txa
+	clc
+	adc #10
+	sta VRAM_INDEX
+	tax
+	lda #$ff ;=NT_UPD_EOF
+	sta VRAM_BUF,x
+	rts
+
+
+
diff --git a/mos-platform/nes/nesdoug/nesdoug.c b/mos-platform/nes/nesdoug/nesdoug.c
index 88b638a79..50147bec6 100644
--- a/mos-platform/nes/nesdoug/nesdoug.c
+++ b/mos-platform/nes/nesdoug/nesdoug.c
@@ -7,21 +7,6 @@
 // based on code by Groepaz/Hitmen <groepaz@gmx.net>, Ullrich von Bassewitz
 // <uz@cc65.org>
 
-// Keeping these functions in C LTOs them in, which informs LTO code generation
-// that the ZP regions aren't available. This saves users of the library from
-// having to manually reserve ZP space from LTO.
-__attribute__((section(".zp.vram_index"))) char VRAM_INDEX;
-__attribute__((section(".zp.meta_ptr"))) const char *META_PTR;
-__attribute__((section(".zp.data_ptr"))) const char *DATA_PTR;
-
-extern char VRAM_BUF[];
-void set_vram_update(const void *buf);
-void set_vram_buffer(void) {
-  VRAM_BUF[0] = 0xff;
-  VRAM_INDEX = 0;
-  set_vram_update(VRAM_BUF);
-}
-
 extern char PAD_STATET[];
 char get_pad_new(char pad) { return PAD_STATET[pad]; }
 
@@ -44,11 +29,6 @@ void set_scroll_y(unsigned y) {
   PPUCTRL_VAR = PPUCTRL_VAR & 0xfd | (y >> 8 & 0x01) << 1;
 }
 
-void set_data_pointer(const void *data) { DATA_PTR = (const char *)data; }
-void set_mt_pointer(const void *metatiles) {
-  META_PTR = (const char *)metatiles;
-}
-
 extern volatile char PPUMASK_VAR;
 void color_emphasis(char color) {
   PPUMASK_VAR = PPUMASK_VAR & 0x1f | color & 0xe0;
diff --git a/mos-platform/nes/nesdoug/nesdoug.s b/mos-platform/nes/nesdoug/nesdoug.s
index 66876ce8a..a98cee8eb 100644
--- a/mos-platform/nes/nesdoug/nesdoug.s
+++ b/mos-platform/nes/nesdoug/nesdoug.s
@@ -5,125 +5,6 @@
 .include "nes.inc"
 .include "neslib.inc"
 
-.zeropage VRAM_INDEX, META_PTR, DATA_PTR
-
-.section .noinit.vram_buf,"aw",@nobits
-.globl VRAM_BUF
-.balign 128
-VRAM_BUF:
-  .zero 128
-
-.text
-.global __post_vram_update
-__post_vram_update:
-	ldx #$ff
-	stx VRAM_BUF
-	inx ;x=0
-	stx VRAM_INDEX
-	rts
-
-;void multi_vram_buffer_horz(char * data, char len, int ppu_address);
-.section .text.multi_vram_buffer_horz,"ax",@progbits
-.globl multi_vram_buffer_horz
-multi_vram_buffer_horz:
-	;     A - len
-	;     X - <ppu_address
-	; __rc2 - <data
-	; __rc3 - >data
-	; __rc4 - >ppu_address
-	ldy VRAM_INDEX
-
-	sta __rc5         ; save len for loop comparison
-	sta VRAM_BUF+2, y ; store len in header
-
-	lda #$40          ; load horizontal flag
-
-multi_vram_buffer_common:
-	ora __rc4         ; combine direction flag with upper address byte
-	sta VRAM_BUF+0, y ; store upper byte in header
-	txa
-	sta VRAM_BUF+1, y ; store lower byte in header
-
-	; get data index past header
-	tya
-	clc
-	adc #3
-	tax ;need y for source, x is for dest and for vram_index
-
-	ldy #0
-
-  ; TODO: unroll?
-.Lmulti_vram_buffer_common_loop:
-	lda (__rc2), y
-	sta VRAM_BUF, x
-	inx
-	iny
-	cpy __rc5
-	bne .Lmulti_vram_buffer_common_loop
-	lda #$ff ;=NT_UPD_EOF
-	sta VRAM_BUF, x
-	stx VRAM_INDEX
-	rts
-
-
-
-
-;void multi_vram_buffer_vert(char * data, char len, int ppu_address);
-.section .text.multi_vram_buffer_vert,"ax",@progbits
-.globl multi_vram_buffer_vert
-multi_vram_buffer_vert:
-	;     A - len
-	;     X - <ppu_address
-	; __rc2 - <data
-	; __rc3 - >data
-	; __rc4 - >ppu_address
-	ldy VRAM_INDEX
-
-	sta __rc5         ; save len for loop comparison
-	sta VRAM_BUF+2, y ; store len in header
-
-	lda #$80          ; load vertical flag
-	bne multi_vram_buffer_common ; always taken
-
-
-
-
-;void one_vram_buffer(char data, int ppu_address);
-.section .text.one_vram_buffer,"ax",@progbits
-.globl one_vram_buffer
-one_vram_buffer:
-	;     A - data
-	;     X - <ppu_address
-	; __rc2 - >ppu_address
-	ldy VRAM_INDEX
-
-	sta VRAM_BUF+2, y ; data
-	lda __rc2
-	sta VRAM_BUF, y   ; address hi
-	txa
-	sta VRAM_BUF+1, y ; adderess lo
-	iny
-	iny
-	iny
-	lda #$ff ;=NT_UPD_EOF
-	sta VRAM_BUF, y
-	sty VRAM_INDEX
-	rts
-
-
-
-
-;void clear_vram_buffer(void);
-;_clear_vram_buffer:
-;	lda #0
-;	sta VRAM_INDEX
-;	lda #$ff
-;	sta VRAM_BUF
-;	rts
-
-
-
-
 ;char check_collision(void * object1, void * object2);
 .section .text.check_collision,"ax",@progbits
 .globl check_collision
@@ -370,270 +251,6 @@ get_at_addr:
 
 
 
-;void buffer_4_mt(int ppu_address, char index);
-.section .text.buffer_4_mt,"ax",@progbits
-.globl buffer_4_mt
-buffer_4_mt:
-	sta __rc3
-	lda __rc2
-	;a is the index into the data, get 4 metatiles
-
-	and #$ee ;sanitize, x and y should be even
-	tay
-	lda (DATA_PTR), y
-	sta __rc6
-	iny
-	lda (DATA_PTR), y
-	sta __rc7
-		tya
-		clc
-		adc #15
-			cmp #$f0 ;too far, data set only 240 bytes
-			bcs .Lskip
-		tay
-	lda (DATA_PTR), y
-	sta __rc8
-	iny
-	lda (DATA_PTR), y
-	sta __rc9
-.Lskip:
-;metatiles are in _rc6 - rc9 now
-	lda __rc3
-	and #$9c ;sanitize, should be top left
-	sta __rc11
-	stx __rc12 ;save for later, ppu_address
-
-	sta __rc4
-	txa
-	ora #$40	;NT_UPD_HORZ
-	sta __rc5
-
-;buffer the ppu_address
-
-	lda #0
-	sta __rc10 ;loop count, index to the metatiles
-	ldx VRAM_INDEX
-.Lbuffer_4_mt_loop:
-	lda __rc4 ;low byte
-	sta VRAM_BUF+1, x
-	lda __rc5 ;high byte
-	sta VRAM_BUF,x
-	jsr .Lsub1 ;adds $20 to the address for next time
-
-		lda #4 ;tell the system 4 bytes in a row
-		sta VRAM_BUF+2,x
-		sta VRAM_BUF+9,x ;loops twice, so, it does this twice
-
-	jsr .Lsub2 ;gets y is which metatile
-
-	lda (META_PTR), y
-	sta VRAM_BUF+3,x ;		buffer the 4 tiles
-	iny
-	lda (META_PTR), y
-	sta VRAM_BUF+4,x
-	iny
-	lda (META_PTR), y
-	sta VRAM_BUF+10,x
-	iny
-	lda (META_PTR), y
-	sta VRAM_BUF+11,x
-	jsr .Lsub4 ;get attrib bits, shift into place
-
-;same, but for right side
-	lda __rc4 ;low byte ppu address, again
-	sta VRAM_BUF+8,x
-	lda __rc5 ;high byte
-	sta VRAM_BUF+7,x
-	jsr .Lsub1
-
-	inc __rc10 ;count and index
-	jsr .Lsub2
-	lda (META_PTR), y
-	sta VRAM_BUF+5,x ;		buffer the 4 tiles
-	iny
-	lda (META_PTR), y
-	sta VRAM_BUF+6,x
-	iny
-	lda (META_PTR), y
-	sta VRAM_BUF+12,x
-	iny
-	lda (META_PTR), y
-	sta VRAM_BUF+13,x
-	jsr .Lsub4
-
-	txa ;adjust the vram index to the next set
-	clc
-	adc #14
-	tax
-
-	jsr .Lsub3 ;check if lowest y on screen, skip the la
-	bne .Lloop_done
-
-	inc __rc10
-	ldy __rc10
-	cpy #4
-	bcc .Lbuffer_4_mt_loop
-
-.Lloop_done:
-
-
-
-;now push 1 attribute byte to the vram buffer
-;first, shift the bits to get an attribute address
-;we stored the original at TEMP+7,8, 8 is high byte
-;a bunch of bit shifting to get 3 bits from x and 3 y
-	lsr __rc12 ;high byte
-	ror __rc11
-	lsr __rc12
-	ror __rc11
-	lda __rc11
-	pha ;save
-	and #7 ;just the x bits
-	sta __rc4
-	pla
-	lsr a ;just the y bits
-	lsr a
-	and #$f8
-	ora #$c0
-	ora __rc4
-	sta __rc4 ;low byte
-;now high byte
-	lda __rc12
-	asl a
-	asl a
-	ora #$23
-	sta __rc5 ;high byte, and the low byte is in TEMP
-
-
-;finally, push it all to the vram_buffer as a single byte
-	lda __rc5 ;high byte
-	sta VRAM_BUF,x
-	inx
-	lda __rc4 ;low byte
-	sta VRAM_BUF,x
-	inx
-	lda __rc14
-	sta VRAM_BUF,x
-	inx
-	lda #$ff ;=NT_UPD_EOF
-	sta VRAM_BUF,x
-	stx VRAM_INDEX
-	rts
-
-
-.Lsub1:	;add $20 is a 1 down on the screen
-	tay ;high byte
-	lda __rc4
-	clc
-	adc #$20
-	sta __rc4
-	bcc .Lsub1b
-	iny
-.Lsub1b:
-	sty __rc5
-	rts
-
-
-.Lsub2:	;get the next metatile offset
-	ldy __rc10
-	lda __rc6, y ;metatile
-;multiply by 5
-	sta __rc13
-	asl a
-	asl a ;x4 = 4 bytes per
-	clc
-	adc __rc13
-	tay
-	rts
-
-
-.Lsub3: ;check make sure we're not at the lowest y and overflowing
-	ldy #0 ;x is forbidden
-	lda __rc12 ;high byte
-	and #$03
-	cmp #$03
-	bne .Lnot_overflow
-	lda __rc11
-	cmp #$80 ;last row of mt
-	bcc .Lnot_overflow
-	iny
-	lsr __rc14 ;make sure the attrib bits in correct position
-	lsr __rc14
-	lsr __rc14
-	lsr __rc14
-.Lnot_overflow:
-	tya ;set flag
-	rts
-
-
-.Lsub4: ;get attrib bits, roll them in place
-	iny
-	lda (META_PTR), y ;5th byte = attribute
-	and #3 ;just need 2 bits
-	ror a ;bit to carry
-	ror __rc14 ;shift carry in
-	ror a ;bit to carry
-	ror __rc14 ;roll the a.t. bits in the high 2 bits
-	rts
-
-
-
-
-;void buffer_1_mt(int ppu_address, char metatile);
-.section .text.buffer_1_mt,"ax",@progbits
-.globl buffer_1_mt
-buffer_1_mt:
-	and #$de ;sanitize, should be even x and y
-	sta __rc4
-	txa
-	ora #$40 ;NT_UPD_HORZ
-	sta __rc5
-
-	ldx VRAM_INDEX
-	lda __rc4				;ppu address
-	sta VRAM_BUF+1,x
-		clc
-		adc #$20 ;shouldn't be rollover
-	sta VRAM_BUF+6,x
-
-	lda __rc5
-	sta VRAM_BUF,x
-	sta VRAM_BUF+5,x
-
-	lda #2 ;tell the system 2 bytes in a row
-	sta VRAM_BUF+2,x
-	sta VRAM_BUF+7,x
-
-	lda __rc2 ;which metatile
-	asl a
-	asl a
-	clc
-	adc __rc2 ;multiply 5
-	tay
-	lda (META_PTR), y ;tile
-	sta VRAM_BUF+3,x
-	iny
-	lda (META_PTR), y ;tile
-	sta VRAM_BUF+4,x
-	iny
-	lda (META_PTR), y ;tile
-	sta VRAM_BUF+8,x
-	iny
-	lda (META_PTR), y ;tile
-	sta VRAM_BUF+9,x
-
-	txa
-	clc
-	adc #10
-	sta VRAM_INDEX
-	tax
-	lda #$ff ;=NT_UPD_EOF
-	sta VRAM_BUF,x
-	rts
-
-
-
-
 ;void xy_split(unsigned x, unsigned y);
 .section .text.xy_split,"ax",@progbits
 .globl xy_split
diff --git a/mos-platform/nes/nesdoug/vram_buffer.c b/mos-platform/nes/nesdoug/vram_buffer.c
new file mode 100644
index 000000000..8f46ce41f
--- /dev/null
+++ b/mos-platform/nes/nesdoug/vram_buffer.c
@@ -0,0 +1,13 @@
+// written by Doug Fraker
+// version 1.2, 1/1/2022
+
+__attribute__((section(".zp.vram_index"))) char VRAM_INDEX;
+
+extern char VRAM_BUF[];
+
+void set_vram_update(const void *buf);
+void set_vram_buffer(void) {
+  VRAM_BUF[0] = 0xff;
+  VRAM_INDEX = 0;
+  set_vram_update(VRAM_BUF);
+}
diff --git a/mos-platform/nes/nesdoug/vram_buffer.s b/mos-platform/nes/nesdoug/vram_buffer.s
new file mode 100644
index 000000000..5115fc1d9
--- /dev/null
+++ b/mos-platform/nes/nesdoug/vram_buffer.s
@@ -0,0 +1,118 @@
+;written by Doug Fraker
+;version 1.2, 1/1/2022
+
+.zeropage VRAM_INDEX
+
+.section .noinit.vram_buf,"aw",@nobits
+.globl VRAM_BUF
+.balign 128
+VRAM_BUF:
+  .zero 128
+
+.text
+.global __post_vram_update
+__post_vram_update:
+	ldx #$ff
+	stx VRAM_BUF
+	inx ;x=0
+	stx VRAM_INDEX
+	rts
+
+;void multi_vram_buffer_horz(char * data, char len, int ppu_address);
+.section .text.multi_vram_buffer_horz,"ax",@progbits
+.globl multi_vram_buffer_horz
+multi_vram_buffer_horz:
+	;     A - len
+	;     X - <ppu_address
+	; __rc2 - <data
+	; __rc3 - >data
+	; __rc4 - >ppu_address
+	ldy VRAM_INDEX
+
+	sta __rc5         ; save len for loop comparison
+	sta VRAM_BUF+2, y ; store len in header
+
+	lda #$40          ; load horizontal flag
+
+multi_vram_buffer_common:
+	ora __rc4         ; combine direction flag with upper address byte
+	sta VRAM_BUF+0, y ; store upper byte in header
+	txa
+	sta VRAM_BUF+1, y ; store lower byte in header
+
+	; get data index past header
+	tya
+	clc
+	adc #3
+	tax ;need y for source, x is for dest and for vram_index
+
+	ldy #0
+
+  ; TODO: unroll?
+.Lmulti_vram_buffer_common_loop:
+	lda (__rc2), y
+	sta VRAM_BUF, x
+	inx
+	iny
+	cpy __rc5
+	bne .Lmulti_vram_buffer_common_loop
+	lda #$ff ;=NT_UPD_EOF
+	sta VRAM_BUF, x
+	stx VRAM_INDEX
+	rts
+
+
+
+
+;void multi_vram_buffer_vert(char * data, char len, int ppu_address);
+.section .text.multi_vram_buffer_vert,"ax",@progbits
+.globl multi_vram_buffer_vert
+multi_vram_buffer_vert:
+	;     A - len
+	;     X - <ppu_address
+	; __rc2 - <data
+	; __rc3 - >data
+	; __rc4 - >ppu_address
+	ldy VRAM_INDEX
+
+	sta __rc5         ; save len for loop comparison
+	sta VRAM_BUF+2, y ; store len in header
+
+	lda #$80          ; load vertical flag
+	bne multi_vram_buffer_common ; always taken
+
+
+
+
+;void one_vram_buffer(char data, int ppu_address);
+.section .text.one_vram_buffer,"ax",@progbits
+.globl one_vram_buffer
+one_vram_buffer:
+	;     A - data
+	;     X - <ppu_address
+	; __rc2 - >ppu_address
+	ldy VRAM_INDEX
+
+	sta VRAM_BUF+2, y ; data
+	lda __rc2
+	sta VRAM_BUF, y   ; address hi
+	txa
+	sta VRAM_BUF+1, y ; adderess lo
+	iny
+	iny
+	iny
+	lda #$ff ;=NT_UPD_EOF
+	sta VRAM_BUF, y
+	sty VRAM_INDEX
+	rts
+
+
+
+
+;void clear_vram_buffer(void);
+;_clear_vram_buffer:
+;	lda #0
+;	sta VRAM_INDEX
+;	lda #$ff
+;	sta VRAM_BUF
+;	rts
diff --git a/mos-platform/nes/neslib/CMakeLists.txt b/mos-platform/nes/neslib/CMakeLists.txt
index b5bd9c94b..b599d4419 100644
--- a/mos-platform/nes/neslib/CMakeLists.txt
+++ b/mos-platform/nes/neslib/CMakeLists.txt
@@ -8,7 +8,15 @@ add_platform_library(nes-neslib
   neslib.s
   ntsc.c
   ntsc.s
+  oam_update.c
+  oam_update.s
+  pal_bright.c
+  pal_bright.s
+  pal_update.c
+  pal_update.s
   rand.c
-  rand.s)
+  rand.s
+  vram_update.c
+  vram_update.s)
 target_include_directories(nes-neslib BEFORE PUBLIC ..)
 target_link_libraries(nes-neslib PRIVATE common-asminc nes-neslib-asminc)
diff --git a/mos-platform/nes/neslib/neslib.c b/mos-platform/nes/neslib/neslib.c
index 25d6afff8..e72b9e0f6 100644
--- a/mos-platform/nes/neslib/neslib.c
+++ b/mos-platform/nes/neslib/neslib.c
@@ -16,11 +16,6 @@
 __attribute__((section(".zp.frame_cnt1"))) volatile char FRAME_CNT1;
 __attribute__((section(".zp.frame_cnt2"))) volatile char FRAME_CNT2;
 __attribute__((section(".zp.vram_update"))) volatile char VRAM_UPDATE;
-__attribute__((section(".zp.name_upd_adr"))) volatile const char *NAME_UPD_ADR;
-__attribute__((section(".zp.name_upd_enable"))) volatile char NAME_UPD_ENABLE;
-__attribute__((section(".zp.pal_update"))) volatile char PAL_UPDATE;
-__attribute__((section(".zp.pal_bg_ptr"))) volatile const char *PAL_BG_PTR;
-__attribute__((section(".zp.pal_spr_ptr"))) volatile const char *PAL_SPR_PTR;
 __attribute__((section(".zp.scroll_x"))) volatile char SCROLL_X;
 __attribute__((section(".zp.scroll_y"))) volatile char SCROLL_Y;
 __attribute__((section(".zp.scroll_x1"))) volatile char SCROLL_X1;
@@ -31,30 +26,6 @@ __attribute__((section(".zp.pad_statet"))) char PAD_STATET[2];
 __attribute__((section(".zp.ppu_ctrl_var"))) volatile char PPUCTRL_VAR;
 __attribute__((section(".zp.ppu_ctrl_var1"))) volatile char PPUCTRL_VAR1;
 __attribute__((section(".zp.ppu_mask_var"))) volatile char PPUMASK_VAR;
-__attribute__((section(".zp.sprid"))) unsigned SPRID;
-
-extern volatile char PAL_BUF[];
-void pal_col(char index, char color) {
-  PAL_BUF[index & 0x1f] = color;
-  ++PAL_UPDATE;
-}
-
-extern const char __palBrightTableL[];
-extern const char __palBrightTableH[];
-void pal_spr_bright(char bright) {
-  PAL_SPR_PTR = (const char *)(__palBrightTableH[bright] << 8 |
-                               __palBrightTableL[bright]);
-  PAL_UPDATE = 1;
-}
-void pal_bg_bright(char bright) {
-  PAL_BG_PTR = (const char *)(__palBrightTableH[bright] << 8 |
-                              __palBrightTableL[bright]);
-  PAL_UPDATE = 1;
-}
-void pal_bright(char bright) {
-  pal_spr_bright(bright);
-  pal_bg_bright(bright);
-}
 
 void ppu_wait_nmi(void);
 void ppu_off(void) {
@@ -78,10 +49,6 @@ void ppu_on_spr(void) {
 
 void ppu_mask(char mask) { PPUMASK_VAR = mask; }
 
-void oam_set(char index) { SPRID = index & 0xfc; }
-
-char oam_get(void) { return SPRID; }
-
 char pad_poll(char pad);
 char pad_trigger(char pad) {
   pad_poll(pad);
@@ -90,11 +57,6 @@ char pad_trigger(char pad) {
 
 char pad_state(char pad) { return PAD_STATE[pad]; }
 
-void set_vram_update(const void *buf) {
-  NAME_UPD_ADR = buf;
-  NAME_UPD_ENABLE = NAME_UPD_ADR != 0;
-}
-
 void vram_adr(unsigned adr) {
   PPU.vram.address = adr >> 8;
   PPU.vram.address = adr & 0xff;
diff --git a/mos-platform/nes/neslib/neslib.inc b/mos-platform/nes/neslib/neslib.inc
index ae2e40c0f..a6592d514 100644
--- a/mos-platform/nes/neslib/neslib.inc
+++ b/mos-platform/nes/neslib/neslib.inc
@@ -1,4 +1,4 @@
-.zeropage FRAME_CNT1, FRAME_CNT2, VRAM_UPDATE, NAME_UPD_ADR, NAME_UPD_ENABLE
-.zeropage PAL_UPDATE, PAL_BG_PTR, PAL_SPR_PTR, SCROLL_X, SCROLL_Y, SCROLL_X1
+.zeropage FRAME_CNT1, FRAME_CNT2, VRAM_UPDATE
+.zeropage SCROLL_X, SCROLL_Y, SCROLL_X1
 .zeropage SCROLL_Y1, PAD_STATE, PAD_STATEP, PAD_STATET, PPUCTRL_VAR
-.zeropage PPUCTRL_VAR1, PPUMASK_VAR, SPRID
+.zeropage PPUCTRL_VAR1, PPUMASK_VAR
diff --git a/mos-platform/nes/neslib/neslib.s b/mos-platform/nes/neslib/neslib.s
index 01592f37d..731638548 100644
--- a/mos-platform/nes/neslib/neslib.s
+++ b/mos-platform/nes/neslib/neslib.s
@@ -16,17 +16,6 @@
 .include "neslib.inc"
 .include "ntsc.inc"
 
-; Reserve space at beginning of RAM for OAM buffer.
-.section .noinit.oam_buf,"a",@nobits
-.globl OAM_BUF
-.align 256
-OAM_BUF:
-  .space 256
-
-; Place the pallette buffer at the very bottom 32 bytes of the hard stack.
-.globl PAL_BUF
-PAL_BUF = 0x0100
-
 .section .init.100,"axR",@progbits
 clearRAM:
     lda #0
@@ -44,7 +33,7 @@ clearRAM:
     bne 1b
 
 
-.section .init.270,"axR",@progbits
+.section .init.255,"axR",@progbits
 clearPalette:
 	lda #$3f
 	sta PPUADDR
@@ -69,95 +58,37 @@ clearVRAM:
 	dey
 	bne 1b
 
-	lda #4
-	jsr pal_bright
-	jsr pal_clear
-	jsr oam_clear
-
-	lda #0b10000000
-	sta PPUCTRL_VAR
-	sta PPUCTRL		;enable NMI
-	lda #0b00000110
-	sta PPUMASK_VAR
 
+.section .init.400,"axR",@progbits
+.globl neslib_final_init
+neslib_final_init:
 	lda #0
 	sta PPUSCROLL
 	sta PPUSCROLL
 
-.section .nmi.050,"axR",@progbits
-	jsr neslib_nmi
 
-.section .text.neslib_nmi,"ax",@progbits
-.globl neslib_nmi
-neslib_nmi:
+.section .nmi.050,"axR",@progbits
+.globl neslib_nmi_begin
+neslib_nmi_begin:
 	lda PPUMASK_VAR	;if rendering is disabled, do not access the VRAM at all
 	and #0b00011000
 	bne .LrenderingOn
-	jmp	.LskipAll
+	jmp neslib_nmi_skip
 
 .LrenderingOn:
 	lda VRAM_UPDATE ;is the frame complete?
 	bne .LdoUpdate
-	jmp .LskipAll ;skipUpd
+	jmp neslib_nmi_skip
 
 .LdoUpdate:
 	lda #0
 	sta VRAM_UPDATE
 
-	lda #>OAM_BUF		;update OAM
-	sta OAMDMA
-
-	lda PAL_UPDATE		;update palette if needed
-	bne .LupdPal
-	jmp .LupdVRAM
-
-.LupdPal:
-
-	ldx #0
-	stx PAL_UPDATE
-
-	lda #$3f
-	sta PPUADDR
-	stx PPUADDR
-
-	ldy PAL_BUF				;background color, remember it in X
-	lda (PAL_BG_PTR),y
-	sta PPUDATA
-	tax
-
-	.irp i,0,1,2
-	ldy PAL_BUF+1+\i
-	lda (PAL_BG_PTR),y
-	sta PPUDATA
-	.endr
-
-	.irp j,0,1,2
-	stx PPUDATA			;background color
-	.irp i,0,1,2
-	ldy PAL_BUF+5+(\j*4)+\i
-	lda (PAL_BG_PTR),y
-	sta PPUDATA
-	.endr
-	.endr
-
-	.irp j,0,1,2,3
-	stx PPUDATA			;background color
-	.irp i,0,1,2
-	ldy PAL_BUF+17+(\j*4)+\i
-	lda (PAL_SPR_PTR),y
-	sta PPUDATA
-	.endr
-	.endr
-
-.LupdVRAM:
 
-	lda NAME_UPD_ENABLE
-	beq .LskipUpd
-
-	jsr flush_vram_update2
-
-.LskipUpd:
 
+.section .nmi.075,"axR",@progbits
+.globl neslib_nmi_end
+neslib_nmi_end:
 	lda #0
 	sta PPUADDR
 	sta PPUADDR
@@ -170,8 +101,7 @@ neslib_nmi:
 	lda PPUCTRL_VAR
 	sta PPUCTRL
 
-.LskipAll:
-
+neslib_nmi_skip:
 	lda PPUMASK_VAR
 	sta PPUMASK
 
@@ -183,203 +113,6 @@ neslib_nmi:
 	lda #0
 	sta FRAME_CNT2
 .LskipNtsc:
-	rts
-
-;void pal_all(const char *data);
-.section .text.pal_all,"ax",@progbits
-.global pal_all
-pal_all:
-
-	ldx #$00
-	lda #$20
-
-.Lpal_copy:
-
-	sta __rc4
-
-	ldy #$00
-
-0:
-
-	lda (__rc2),y
-	sta PAL_BUF,x
-	inx
-	iny
-	dec __rc4
-	bne 0b
-
-	inc PAL_UPDATE
-
-	rts
-
-
-
-;void pal_bg(const char *data);
-.section .text.pal_bg,"ax",@progbits
-.globl pal_bg
-pal_bg:
-
-	ldx #$00
-	lda #$10
-	jmp .Lpal_copy
-
-
-
-;void pal_spr(const char *data);
-.section .text.pal_spr,"ax",@progbits
-.globl pal_spr
-pal_spr:
-
-	ldx #$10
-	txa
-	jmp .Lpal_copy
-
-
-
-;void pal_clear(void);
-.section .text.pal_clear,"ax",@progbits
-.globl pal_clear
-pal_clear:
-
-	lda #$0f
-	ldx #0
-
-0:
-
-	sta PAL_BUF,x
-	inx
-	cpx #$20
-	bne 0b
-	stx PAL_UPDATE
-	rts
-
-
-
-;void oam_clear(void);
-.section .text.oam_clear,"ax",@progbits
-.globl oam_clear
-oam_clear:
-
-	ldx #0
-	stx SPRID ; automatically sets sprid to zero
-	lda #$ff
-0:
-	sta OAM_BUF,x
-	inx
-	inx
-	inx
-	inx
-	bne 0b
-	rts
-
-
-;void oam_size(unsigned char size);
-.section .text.oam_size,"ax",@progbits
-.globl oam_size
-oam_size:
-
-	asl a
-	asl a
-	asl a
-	asl a
-	asl a
-	and #$20
-	sta __rc2
-	lda PPUCTRL_VAR
-	and #$df
-	ora __rc2
-	sta PPUCTRL_VAR
-
-	rts
-
-
-
-;void oam_spr(unsigned char x,unsigned char y,unsigned char chrnum,unsigned char attr);
-;sprid removed
-.section .text.oam_spr,"ax",@progbits
-.globl oam_spr
-oam_spr:
-	ldy SPRID
-	sta OAM_BUF+3,y
-
-	txa
-	sta OAM_BUF+0,y
-
-	lda __rc2
-	sta OAM_BUF+1,y
-
-	lda __rc3
-	sta OAM_BUF+2,y
-
-	tya
-	clc ; if we can prove Carry is always false, then we don't need this
-	adc #$04
-	sta SPRID
-	rts
-
-
-
-;void oam_meta_spr(unsigned char x,unsigned char y,const unsigned char *data);
-;sprid removed
-.section .text.oam_meta_spr,"ax",@progbits
-.globl oam_meta_spr
-oam_meta_spr:
-
-	sta __rc4
-	stx __rc5
-	ldx SPRID
-	ldy #0
-1:
-	lda (__rc2),y		;x offset
-	cmp #$80
-	beq 2f
-	iny
-	clc
-	adc __rc4
-	sta OAM_BUF+3,x
-	lda (__rc2),y		;y offset
-	iny
-	clc
-	adc __rc5
-	sta OAM_BUF+0,x
-	lda (__rc2),y		;tile
-	iny
-	sta OAM_BUF+1,x
-	lda (__rc2),y		;attribute
-	iny
-	sta OAM_BUF+2,x
-	inx
-	inx
-	inx
-	inx
-	jmp 1b
-2:
-	stx SPRID
-	rts
-
-
-
-;void oam_hide_rest(void);
-;sprid removed
-.section .text.oam_hide_rest,"ax",@progbits
-.globl oam_hide_rest
-oam_hide_rest:
-
-	ldx SPRID
-	lda #240
-
-0:
-
-	sta OAM_BUF,x
-	inx
-	inx
-	inx
-	inx
-	bne 0b
-	;x is zero
-	stx SPRID
-	rts
-
 
 
 ;void ppu_wait_frame(void);
@@ -709,94 +442,6 @@ pad_poll:
 
 
 
-;void flush_vram_update(unsigned char *buf);
-.section .text.flush_vram_update,"ax",@progbits
-.globl flush_vram_update
-flush_vram_update:
-	lda __rc2
-	sta NAME_UPD_ADR+0
-	lda __rc3
-	sta NAME_UPD_ADR+1
-
-.globl flush_vram_update2
-flush_vram_update2: ;minor changes %
-
-	ldy #0
-
-.LupdName:
-	; First byte is upper PPU address or #$ff if done
-	lda (NAME_UPD_ADR),y
-	iny
-	cmp #$40            ; bits 6 and 7 indicate sequential ops
-	bcc .LupdSingle
-
-	; save upper address byte for arithmetic
-	tax
-	lda PPUCTRL_VAR
-	cpx #$80             ; below 80 is horizontal
-	bmi .LupdHorzSeq
-	cpx #$ff
-	beq .LupdDone
-
-.LupdVertSeq:
-	; Set control bit for vertical traversal
-	ora #$04         ; TODO constants for ctrl flags?
-	bne .LupdNameSeq ;always taken
-
-.LupdSingle:
-	sta PPUADDR
-	lda (NAME_UPD_ADR),y ; address lo
-	iny
-	sta PPUADDR
-	lda (NAME_UPD_ADR),y ; data
-	iny
-	sta PPUDATA
-	bne .LupdName    ; always taken. Assumes index never wraps
-
-.LupdHorzSeq:
-	; Clear control bit for vertical traversal
-	and #$fb
-
-.LupdNameSeq:
-	; Store new control value
-	sta PPUCTRL
-
-	; Mask out top 2 bits of upper address byte
-	txa
-	and #$3F
-
-	sta PPUADDR
-	lda (NAME_UPD_ADR),y  ; address lo
-	iny
-	sta PPUADDR
-	lda (NAME_UPD_ADR),y  ; size
-	iny
-
-	; store size in counter
-	tax
-
-.LupdNameLoop:
-
-	lda (NAME_UPD_ADR),y
-	iny
-	sta PPUDATA
-	dex
-	bne .LupdNameLoop
-
-	lda PPUCTRL_VAR
-	sta PPUCTRL
-
-	jmp .LupdName
-
-.LupdDone:
-	jmp __post_vram_update
-
-.weak __post_vram_update
-__post_vram_update:
-	rts
-
-
-
 ;void vram_fill(unsigned char n,unsigned int len);
 .section .text.vram_fill,"ax",@progbits
 .globl vram_fill
@@ -854,40 +499,3 @@ vram_inc:
 	rts
 
 
-
-.section .rodata.bright_table,"a",@progbits
-.globl __palBrightTableL
-.globl __palBrightTableH
-__palBrightTableL:
-
-	.byte palBrightTable0@mos16lo,palBrightTable1@mos16lo,palBrightTable2@mos16lo
-	.byte palBrightTable3@mos16lo,palBrightTable4@mos16lo,palBrightTable5@mos16lo
-	.byte palBrightTable6@mos16lo,palBrightTable7@mos16lo,palBrightTable8@mos16lo
-
-__palBrightTableH:
-
-	.byte palBrightTable0@mos16hi,palBrightTable1@mos16hi,palBrightTable2@mos16hi
-	.byte palBrightTable3@mos16hi,palBrightTable4@mos16hi,palBrightTable5@mos16hi
-	.byte palBrightTable6@mos16hi,palBrightTable7@mos16hi,palBrightTable8@mos16hi
-
-palBrightTable0:
-	.byte $0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f	;black
-palBrightTable1:
-	.byte $0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f
-palBrightTable2:
-	.byte $0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f
-palBrightTable3:
-	.byte $0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f
-palBrightTable4:
-	.byte $00,$01,$02,$03,$04,$05,$06,$07,$08,$09,$0a,$0b,$0c,$0f,$0f,$0f	;normal colors
-palBrightTable5:
-	.byte $10,$11,$12,$13,$14,$15,$16,$17,$18,$19,$1a,$1b,$1c,$00,$00,$00
-palBrightTable6:
-	.byte $10,$21,$22,$23,$24,$25,$26,$27,$28,$29,$2a,$2b,$2c,$10,$10,$10	;$10 because $20 is the same as $30
-palBrightTable7:
-	.byte $30,$31,$32,$33,$34,$35,$36,$37,$38,$39,$3a,$3b,$3c,$20,$20,$20
-palBrightTable8:
-	.byte $30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30	;white
-	.byte $30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30
-	.byte $30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30
-	.byte $30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30
diff --git a/mos-platform/nes/neslib/ntsc.s b/mos-platform/nes/neslib/ntsc.s
index c4780a529..893b8fa45 100644
--- a/mos-platform/nes/neslib/ntsc.s
+++ b/mos-platform/nes/neslib/ntsc.s
@@ -6,6 +6,12 @@
 .section .init.275,"axR",@progbits
 .globl __do_init_ntsc_mode
 __do_init_ntsc_mode:
+	lda #0b10000000
+	sta PPUCTRL_VAR
+	sta PPUCTRL		;enable NMI
+	lda #0b00000110
+	sta PPUMASK_VAR
+
 waitSync3:
 	lda FRAME_CNT1
 1:
@@ -27,12 +33,3 @@ detectNTSC:
 
 	jsr ppu_off
 
-	lda #0
-	sta __rc2
-	sta __rc3
-	jsr set_vram_update
-
-	lda #0
-	sta PPUSCROLL
-	sta PPUSCROLL
-
diff --git a/mos-platform/nes/neslib/oam_update.c b/mos-platform/nes/neslib/oam_update.c
new file mode 100644
index 000000000..18e02036a
--- /dev/null
+++ b/mos-platform/nes/neslib/oam_update.c
@@ -0,0 +1,10 @@
+// Copyright 2023 LLVM-MOS Project
+// Licensed under the Apache License, Version 2.0 with LLVM Exceptions.
+// See /~https://github.com/llvm-mos/llvm-mos-sdk/blob/main/LICENSE for license
+// information.
+
+__attribute__((section(".zp.sprid"))) unsigned SPRID;
+
+void oam_set(char index) { SPRID = index & 0xfc; }
+
+char oam_get(void) { return SPRID; }
diff --git a/mos-platform/nes/neslib/oam_update.s b/mos-platform/nes/neslib/oam_update.s
new file mode 100644
index 000000000..55709445f
--- /dev/null
+++ b/mos-platform/nes/neslib/oam_update.s
@@ -0,0 +1,135 @@
+; Copyright 2023 LLVM-MOS Project
+; Licensed under the Apache License, Version 2.0 with LLVM Exceptions.
+; See /~https://github.com/llvm-mos/llvm-mos-sdk/blob/main/LICENSE for license
+; information.
+
+.zeropage SPRID
+
+.include "nes.inc"
+
+; Reserve space at beginning of RAM for OAM buffer.
+.section .noinit.oam_buf,"a",@nobits
+.globl OAM_BUF
+.align 256
+OAM_BUF:
+  .space 256
+
+
+.section .init.270,"axR",@progbits
+.globl oam_update_init
+oam_update_init:
+	jsr oam_clear
+
+
+.section .nmi.055,"axR",@progbits
+.globl oam_update_nmi
+oam_update_nmi:
+        lda #>OAM_BUF
+        sta OAMDMA
+
+
+
+;void oam_clear(void);
+.section .text.oam_clear,"ax",@progbits
+.globl oam_clear
+oam_clear:
+
+	ldx #0
+	stx SPRID ; automatically sets sprid to zero
+	lda #$ff
+0:
+	sta OAM_BUF,x
+	inx
+	inx
+	inx
+	inx
+	bne 0b
+	rts
+
+
+
+;void oam_spr(unsigned char x,unsigned char y,unsigned char chrnum,unsigned char attr);
+;sprid removed
+.section .text.oam_spr,"ax",@progbits
+.globl oam_spr
+oam_spr:
+	ldy SPRID
+	sta OAM_BUF+3,y
+
+	txa
+	sta OAM_BUF+0,y
+
+	lda __rc2
+	sta OAM_BUF+1,y
+
+	lda __rc3
+	sta OAM_BUF+2,y
+
+	tya
+	clc ; if we can prove Carry is always false, then we don't need this
+	adc #$04
+	sta SPRID
+	rts
+
+
+
+;void oam_meta_spr(unsigned char x,unsigned char y,const unsigned char *data);
+;sprid removed
+.section .text.oam_meta_spr,"ax",@progbits
+.globl oam_meta_spr
+oam_meta_spr:
+
+	sta __rc4
+	stx __rc5
+	ldx SPRID
+	ldy #0
+1:
+	lda (__rc2),y		;x offset
+	cmp #$80
+	beq 2f
+	iny
+	clc
+	adc __rc4
+	sta OAM_BUF+3,x
+	lda (__rc2),y		;y offset
+	iny
+	clc
+	adc __rc5
+	sta OAM_BUF+0,x
+	lda (__rc2),y		;tile
+	iny
+	sta OAM_BUF+1,x
+	lda (__rc2),y		;attribute
+	iny
+	sta OAM_BUF+2,x
+	inx
+	inx
+	inx
+	inx
+	jmp 1b
+2:
+	stx SPRID
+	rts
+
+;void oam_hide_rest(void);
+;sprid removed
+.section .text.oam_hide_rest,"ax",@progbits
+.globl oam_hide_rest
+oam_hide_rest:
+
+	ldx SPRID
+	lda #240
+
+0:
+
+	sta OAM_BUF,x
+	inx
+	inx
+	inx
+	inx
+	bne 0b
+	;x is zero
+	stx SPRID
+	rts
+
+
diff --git a/mos-platform/nes/neslib/pal_bright.c b/mos-platform/nes/neslib/pal_bright.c
new file mode 100644
index 000000000..c1697f53a
--- /dev/null
+++ b/mos-platform/nes/neslib/pal_bright.c
@@ -0,0 +1,25 @@
+// Copyright 2023 LLVM-MOS Project
+// Licensed under the Apache License, Version 2.0 with LLVM Exceptions.
+// See /~https://github.com/llvm-mos/llvm-mos-sdk/blob/main/LICENSE for license
+// information.
+
+__attribute__((section(".zp.pal_bg_ptr"))) volatile const char *PAL_BG_PTR;
+__attribute__((section(".zp.pal_spr_ptr"))) volatile const char *PAL_SPR_PTR;
+__attribute__((section(".zp.pal_update"))) extern volatile char PAL_UPDATE;
+
+extern const char __palBrightTableL[];
+extern const char __palBrightTableH[];
+void pal_spr_bright(char bright) {
+  PAL_SPR_PTR = (const char *)(__palBrightTableH[bright] << 8 |
+                               __palBrightTableL[bright]);
+  PAL_UPDATE = 1;
+}
+void pal_bg_bright(char bright) {
+  PAL_BG_PTR = (const char *)(__palBrightTableH[bright] << 8 |
+                              __palBrightTableL[bright]);
+  PAL_UPDATE = 1;
+}
+void pal_bright(char bright) {
+  pal_spr_bright(bright);
+  pal_bg_bright(bright);
+}
diff --git a/mos-platform/nes/neslib/pal_bright.s b/mos-platform/nes/neslib/pal_bright.s
new file mode 100644
index 000000000..d97be511c
--- /dev/null
+++ b/mos-platform/nes/neslib/pal_bright.s
@@ -0,0 +1,48 @@
+; Copyright 2023 LLVM-MOS Project
+; Licensed under the Apache License, Version 2.0 with LLVM Exceptions.
+; See /~https://github.com/llvm-mos/llvm-mos-sdk/blob/main/LICENSE for license
+; information.
+
+.section .init.260,"axR",@progbits
+.globl pal_bright_init
+pal_bright_init:
+	lda #4
+	jsr pal_bright
+
+
+.section .rodata.bright_table,"a",@progbits
+.globl __palBrightTableL
+.globl __palBrightTableH
+__palBrightTableL:
+
+	.byte palBrightTable0@mos16lo,palBrightTable1@mos16lo,palBrightTable2@mos16lo
+	.byte palBrightTable3@mos16lo,palBrightTable4@mos16lo,palBrightTable5@mos16lo
+	.byte palBrightTable6@mos16lo,palBrightTable7@mos16lo,palBrightTable8@mos16lo
+
+__palBrightTableH:
+
+	.byte palBrightTable0@mos16hi,palBrightTable1@mos16hi,palBrightTable2@mos16hi
+	.byte palBrightTable3@mos16hi,palBrightTable4@mos16hi,palBrightTable5@mos16hi
+	.byte palBrightTable6@mos16hi,palBrightTable7@mos16hi,palBrightTable8@mos16hi
+
+palBrightTable0:
+	.byte $0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f	;black
+palBrightTable1:
+	.byte $0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f
+palBrightTable2:
+	.byte $0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f
+palBrightTable3:
+	.byte $0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f,$0f
+palBrightTable4:
+	.byte $00,$01,$02,$03,$04,$05,$06,$07,$08,$09,$0a,$0b,$0c,$0f,$0f,$0f	;normal colors
+palBrightTable5:
+	.byte $10,$11,$12,$13,$14,$15,$16,$17,$18,$19,$1a,$1b,$1c,$00,$00,$00
+palBrightTable6:
+	.byte $10,$21,$22,$23,$24,$25,$26,$27,$28,$29,$2a,$2b,$2c,$10,$10,$10	;$10 because $20 is the same as $30
+palBrightTable7:
+	.byte $30,$31,$32,$33,$34,$35,$36,$37,$38,$39,$3a,$3b,$3c,$20,$20,$20
+palBrightTable8:
+	.byte $30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30	;white
+	.byte $30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30
+	.byte $30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30
+	.byte $30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30,$30
diff --git a/mos-platform/nes/neslib/pal_update.c b/mos-platform/nes/neslib/pal_update.c
new file mode 100644
index 000000000..88801ae35
--- /dev/null
+++ b/mos-platform/nes/neslib/pal_update.c
@@ -0,0 +1,13 @@
+// Copyright 2023 LLVM-MOS Project
+// Licensed under the Apache License, Version 2.0 with LLVM Exceptions.
+// See /~https://github.com/llvm-mos/llvm-mos-sdk/blob/main/LICENSE for license
+// information.
+
+__attribute__((section(".zp.pal_update"))) volatile char PAL_UPDATE;
+
+
+extern volatile char PAL_BUF[];
+void pal_col(char index, char color) {
+  PAL_BUF[index & 0x1f] = color;
+  ++PAL_UPDATE;
+}
diff --git a/mos-platform/nes/neslib/pal_update.s b/mos-platform/nes/neslib/pal_update.s
new file mode 100644
index 000000000..9392d536a
--- /dev/null
+++ b/mos-platform/nes/neslib/pal_update.s
@@ -0,0 +1,156 @@
+; Copyright 2023 LLVM-MOS Project
+; Licensed under the Apache License, Version 2.0 with LLVM Exceptions.
+; See /~https://github.com/llvm-mos/llvm-mos-sdk/blob/main/LICENSE for license
+; information.
+
+.zeropage PAL_UPDATE, PAL_BG_PTR, PAL_SPR_PTR
+
+.include "nes.inc"
+
+; Place the pallette buffer at the very bottom 32 bytes of the hard stack.
+.globl PAL_BUF
+PAL_BUF = 0x0100
+
+.section .init.265,"axR",@progbits
+.globl pal_update_init
+pal_update_init:
+	jsr pal_clear
+
+
+.section .nmi.060,"axR",@progbits
+.globl pal_update_nmi
+pal_update_nmi:
+	lda PAL_UPDATE		;update palette if needed
+	bne .LupdPal
+	jmp .LskipUpd
+
+.LupdPal:
+
+	ldx #0
+	stx PAL_UPDATE
+
+	lda #$3f
+	sta PPUADDR
+	stx PPUADDR
+
+	ldy PAL_BUF				;background color, remember it in X
+	lda (PAL_BG_PTR),y
+	sta PPUDATA
+	tax
+
+	.irp i,0,1,2
+	ldy PAL_BUF+1+\i
+	lda (PAL_BG_PTR),y
+	sta PPUDATA
+	.endr
+
+	.irp j,0,1,2
+	stx PPUDATA			;background color
+	.irp i,0,1,2
+	ldy PAL_BUF+5+(\j*4)+\i
+	lda (PAL_BG_PTR),y
+	sta PPUDATA
+	.endr
+	.endr
+
+	.irp j,0,1,2,3
+	stx PPUDATA			;background color
+	.irp i,0,1,2
+	ldy PAL_BUF+17+(\j*4)+\i
+	lda (PAL_SPR_PTR),y
+	sta PPUDATA
+	.endr
+	.endr
+.LskipUpd:
+
+
+
+;void pal_all(const char *data);
+.section .text.pal_all,"ax",@progbits
+.global pal_all
+pal_all:
+
+	ldx #$00
+	lda #$20
+
+.Lpal_copy:
+
+	sta __rc4
+
+	ldy #$00
+
+0:
+
+	lda (__rc2),y
+	sta PAL_BUF,x
+	inx
+	iny
+	dec __rc4
+	bne 0b
+
+	inc PAL_UPDATE
+
+	rts
+
+
+
+;void pal_bg(const char *data);
+.section .text.pal_bg,"ax",@progbits
+.globl pal_bg
+pal_bg:
+
+	ldx #$00
+	lda #$10
+	jmp .Lpal_copy
+
+
+
+;void pal_spr(const char *data);
+.section .text.pal_spr,"ax",@progbits
+.globl pal_spr
+pal_spr:
+
+	ldx #$10
+	txa
+	jmp .Lpal_copy
+
+
+
+;void pal_clear(void);
+.section .text.pal_clear,"ax",@progbits
+.globl pal_clear
+pal_clear:
+
+	lda #$0f
+	ldx #0
+
+0:
+
+	sta PAL_BUF,x
+	inx
+	cpx #$20
+	bne 0b
+	stx PAL_UPDATE
+	rts
+
+
+;void oam_size(unsigned char size);
+.section .text.oam_size,"ax",@progbits
+.globl oam_size
+oam_size:
+
+	asl a
+	asl a
+	asl a
+	asl a
+	asl a
+	and #$20
+	sta __rc2
+	lda PPUCTRL_VAR
+	and #$df
+	ora __rc2
+	sta PPUCTRL_VAR
+
+	rts
+
+
diff --git a/mos-platform/nes/neslib/rand.s b/mos-platform/nes/neslib/rand.s
index 0fdae8c8a..5a383f35e 100644
--- a/mos-platform/nes/neslib/rand.s
+++ b/mos-platform/nes/neslib/rand.s
@@ -1,6 +1,6 @@
  .include "neslib.inc"
 
- .section .init.270,"axR",@progbits
+ .section .init.255,"axR",@progbits
 	lda #$fd
 	sta RAND_SEED
 	sta RAND_SEED+1
diff --git a/mos-platform/nes/neslib/vram_update.c b/mos-platform/nes/neslib/vram_update.c
new file mode 100644
index 000000000..3d05b1783
--- /dev/null
+++ b/mos-platform/nes/neslib/vram_update.c
@@ -0,0 +1,15 @@
+// Copyright 2023 LLVM-MOS Project
+// Licensed under the Apache License, Version 2.0 with LLVM Exceptions.
+// See /~https://github.com/llvm-mos/llvm-mos-sdk/blob/main/LICENSE for license
+// information.
+
+__attribute__((section(".zp.name_upd_adr"))) volatile const char *NAME_UPD_ADR;
+__attribute__((section(".zp.name_upd_enable"))) volatile char NAME_UPD_ENABLE;
+
+void set_vram_update(const void *buf) {
+  NAME_UPD_ADR = buf;
+  NAME_UPD_ENABLE = NAME_UPD_ADR != 0;
+}
+
+// Bring in assembly module
+asm (".globl vram_update_nmi");
diff --git a/mos-platform/nes/neslib/vram_update.s b/mos-platform/nes/neslib/vram_update.s
new file mode 100644
index 000000000..4b924820c
--- /dev/null
+++ b/mos-platform/nes/neslib/vram_update.s
@@ -0,0 +1,113 @@
+; Copyright 2023 LLVM-MOS Project
+; Licensed under the Apache License, Version 2.0 with LLVM Exceptions.
+; See /~https://github.com/llvm-mos/llvm-mos-sdk/blob/main/LICENSE for license
+; information.
+
+.zeropage NAME_UPD_ADR, NAME_UPD_ENABLE
+
+.include "nes.inc"
+.include "neslib.inc"
+
+.section .init.280,"axR",@progbits
+.globl vram_update_init
+vram_update_init:
+	lda #0
+	sta __rc2
+	sta __rc3
+	jsr set_vram_update
+
+
+.section .nmi.065,"axR",@progbits
+.globl vram_update_nmi
+vram_update_nmi:
+	lda NAME_UPD_ENABLE
+	beq .LskipUpd
+	jsr flush_vram_update2
+.LskipUpd:
+
+
+;void flush_vram_update(unsigned char *buf);
+.section .text.flush_vram_update,"axR",@progbits
+.globl flush_vram_update
+flush_vram_update:
+	lda __rc2
+	sta NAME_UPD_ADR+0
+	lda __rc3
+	sta NAME_UPD_ADR+1
+
+.globl flush_vram_update2
+flush_vram_update2: ;minor changes %
+
+	ldy #0
+
+.LupdName:
+	; First byte is upper PPU address or #$ff if done
+	lda (NAME_UPD_ADR),y
+	iny
+	cmp #$40            ; bits 6 and 7 indicate sequential ops
+	bcc .LupdSingle
+
+	; save upper address byte for arithmetic
+	tax
+	lda PPUCTRL_VAR
+	cpx #$80             ; below 80 is horizontal
+	bmi .LupdHorzSeq
+	cpx #$ff
+	beq .LupdDone
+
+.LupdVertSeq:
+	; Set control bit for vertical traversal
+	ora #$04         ; TODO constants for ctrl flags?
+	bne .LupdNameSeq ;always taken
+
+.LupdSingle:
+	sta PPUADDR
+	lda (NAME_UPD_ADR),y ; address lo
+	iny
+	sta PPUADDR
+	lda (NAME_UPD_ADR),y ; data
+	iny
+	sta PPUDATA
+	bne .LupdName    ; always taken. Assumes index never wraps
+
+.LupdHorzSeq:
+	; Clear control bit for vertical traversal
+	and #$fb
+
+.LupdNameSeq:
+	; Store new control value
+	sta PPUCTRL
+
+	; Mask out top 2 bits of upper address byte
+	txa
+	and #$3F
+
+	sta PPUADDR
+	lda (NAME_UPD_ADR),y  ; address lo
+	iny
+	sta PPUADDR
+	lda (NAME_UPD_ADR),y  ; size
+	iny
+
+	; store size in counter
+	tax
+
+.LupdNameLoop:
+
+	lda (NAME_UPD_ADR),y
+	iny
+	sta PPUDATA
+	dex
+	bne .LupdNameLoop
+
+	lda PPUCTRL_VAR
+	sta PPUCTRL
+
+	jmp .LupdName
+
+.LupdDone:
+	jmp __post_vram_update
+
+.weak __post_vram_update
+__post_vram_update:
+	rts