diff --git a/src/map.s b/src/map.s
index 6c75fb4..e34d652 100644
--- a/src/map.s
+++ b/src/map.s
@@ -4,13 +4,16 @@ INCLUDE "util.inc"
 
 SECTION "Map Data", WRAM0
 
+DEF ROW_BUFFER_SIZE EQUS "SCRN_X_B + 4"
+DEF COL_BUFFER_SIZE EQUS "SCRN_Y_B + 4"
+
 SCROLL_DIR:: DB
 
 PENDING_ROW_PTR:: DW ; Where to write pending row data (0 = no write)
-PENDING_ROW_DATA:: DS SCRN_VX_B ; Row to be written
+PENDING_ROW_DATA:: DS ROW_BUFFER_SIZE ; Row to be written
 
 PENDING_COL_PTR:: DW ; Where to write pending column data (0 = no write)
-PENDING_COL_DATA: DS SCRN_VY_B ; Column to be written
+PENDING_COL_DATA: DS COL_BUFFER_SIZE ; Column to be written
 
 CURRENT_DATA_START::
 CURRENT_TILE_PTR:: DW ; Location of tile data
@@ -82,77 +85,6 @@ Map_Load::
 
     ret
 
-; Loads the map X-coordinate into B
-MACRO LOAD_MAPX
-    ; if SCX <= 0x7f
-    ; ld a, [rSCX]
-    ; cp $80
-    ; jr nc, .wrap\@
-
-    ; B = CAMERA_X - SCX/8
-    ld a, [rSCX]
-    ld b, a
-    ld a, [CURRENT_CAMERA_X]
-    srln b, 3
-    sub b
-    ld b, a
-    ; jr .done\@
-
-    ; else SCX > 0x7f
-; .wrap\@:
-;     ; B = CAMERA_X - PAGE_X/8
-;     ld a, [PAGEX]
-;     ld b, a
-;     ld a, [CURRENT_CAMERA_X]
-;     srln b, 3
-;     sub b
-;     ld b, a
-
-;     ; HL += PAGE_X/8
-;     ld a, [PAGEX]
-;     srln a, 3
-;     ADD16 hl
-
-; .done\@:
-ENDM
-
-; Loads the map Y-coordinate into C
-MACRO LOAD_MAPY
-    ; if SCY <= 0x7f
-    ; ld a, [rSCY]
-    ; cp $80
-    ; jr nc, .wrap\@
-
-    ; C = CAMERA_Y - SCY/8
-    ld a, [rSCY]
-    ld c, a
-    ld a, [CURRENT_CAMERA_Y]
-    srln c, 3
-    sub c
-    ld c, a
-;     jr .done\@
-
-;     ; else SCY > 0x7f
-; .wrap\@:
-;     ; C = CAMERA_Y - SCY/8
-;     ld a, [PAGEY]
-;     ld c, a
-;     ld a, [CURRENT_CAMERA_Y]
-;     srln c, 3
-;     sub c
-;     ld c, a
-
-;     ; HL += 32 * PAGE_Y/8
-;     ld d, 0
-;     ld a, [PAGEY]
-;     ld e, a
-;     REPT 5
-;         SLA16 de
-;     ENDR
-;     add hl, de
-; .done\@:
-ENDM
-
 ; Update map state based on SCX/SCY
 Map_Scroll::
     ; Check scroll up/down
@@ -165,7 +97,10 @@ Map_Scroll::
     and %111
     ret nz
 
-    LOAD_MAPX
+    ; B = CAMERA_X - 2
+    ld a, [CURRENT_CAMERA_X]
+    sub 2
+    ld b, a
 
     ld a, [SCROLL_DIR]
     cp DIRB_UP
@@ -176,12 +111,10 @@ Map_Scroll::
     sub 2
     ld c, a
 
-    ; HL = _SCRN0 + 32 * ((SCX/8 - 2) % 32)
-    ld a, [rSCX]
+    ; MAP Y = SCY/8 - 2
+    ld a, [rSCY]
     srln a, 3
     sub 2
-    and %11111
-    call get_row_ptr
 
     jr .write_row
 
@@ -191,13 +124,24 @@ Map_Scroll::
     add SCRN_Y_B + 1
     ld c, a
 
-    ; HL = _SCRN0 + 32 * (SCY/8 + SCRN_Y_B + 1)
+    ; MAP Y = SCY/8 + SCRN_Y_B + 1
     ld a, [rSCY]
     srln a, 3
     add SCRN_Y_B + 1
-    call get_row_ptr
 
 .write_row:
+    and %11111
+    call get_row_ptr
+
+    ld a, [rSCX]
+    sub 16
+    srl a
+    srl a
+    srl a
+    ld e, a
+    ld d, 0
+    add hl, de
+
     call enqueue_row_write
     ret
 
@@ -207,43 +151,51 @@ Map_Scroll::
     and %111
     ret nz
 
-    LOAD_MAPY
+    ; C = CAMERA_Y - 2
+    ld a, [CURRENT_CAMERA_Y]
+    sub 2
+    ld c, a
+
+    ld a, [rSCY]
+    sub 16
+    srl a
+    srl a
+    srl a
+    call get_row_ptr
 
     ld a, [SCROLL_DIR]
     cp DIRB_LEFT
     jr nz, .scroll_right
 
-    ; HL = VRAM + (SCX/8 - 2) % 32
-    ld hl, _SCRN0
-    ld a, [rSCX]
-    srln a, 3
-    sub 2
-    and %11111
-    ADD16 hl
-
     ; B = CAMERA_X - 2
     ld a, [CURRENT_CAMERA_X]
     sub 2
     ld b, a
 
+    ; HL = VRAM + (SCX/8 - 2)
+    ld a, [rSCX]
+    srln a, 3
+    sub 2
+
     jr .write_col
 
 .scroll_right:
+    ; B = CAMERA_X + 20 + 1
+    ld a, [CURRENT_CAMERA_X]
+    add SCRN_X_B + 1
+    ld b, a
+
     ; HL = VRAM + SCX/8 + SCRN_X_B + 1
-    ld hl, _SCRN0
     ld a, [rSCX]
     srl a
     srl a
     srl a
     add SCRN_X_B + 1
-    ADD16 hl
-
-    ; B = CAMERA_X + 20 + 1
-    ld a, [CURRENT_CAMERA_X]
-    add SCRN_X_B + 1
-    ld b, a
 
 .write_col:
+    and %11111
+    ADD16 hl
+
     call enqueue_col_write
     ret
 
@@ -257,7 +209,7 @@ Map_Update::
     jr z, .update_col
 
     ld hl, PENDING_ROW_DATA
-    ld d, SCRN_VX_B
+    ld d, ROW_BUFFER_SIZE
 .copy_row:
     ld a, [hl+]
     ld [bc], a
@@ -289,7 +241,7 @@ Map_Update::
     or c
     ret z
 
-    ld d, SCRN_VY_B
+    ld d, COL_BUFFER_SIZE
     ld hl, PENDING_COL_DATA
 .update_col_loop:
     ld a, [hl+]
@@ -367,7 +319,7 @@ enqueue_row_write:
 
 .copy_map_row:
     ; C = BYTES_LEFT
-    ld c, SCRN_VX_B
+    ld c, ROW_BUFFER_SIZE
     ld de, PENDING_ROW_DATA
 
     ; If X > 0, increment map pointer by X
@@ -378,15 +330,18 @@ enqueue_row_write:
     ADD16 hl
     jr .copy_middle
 
-    ; Note: Can skip checking BYTES_LEFT > 0 in this loop. If there were
-    ; SCRN_VX_B zeros to write, then Y would be 1 greater and we would have
-    ; jumped into .zero_row before reaching this code
 .pad_left:
     ; Check X < 0
     bit 7, b
     jr z, .copy_middle
 
+    ; Check BYTES_LEFT > 0
+    ld a, c
+    or a
+    ret z
+
     ; *ROW++ = 0
+    xor a
     ld [de], a
     inc de
 
@@ -436,7 +391,7 @@ enqueue_row_write:
 .zero_row:
     ld hl, PENDING_ROW_DATA
     xor a
-    ld c, SCRN_VX_B
+    ld c, ROW_BUFFER_SIZE
 .zero_row_loop:
     ld [hl+], a
     dec c
@@ -456,17 +411,7 @@ write_map_row:
     push de
     push hl
 
-    call enqueue_row_write
-
-    ld a, [PENDING_ROW_PTR]
-    ld c, a
-    ld a, [PENDING_ROW_PTR + 1]
-    ld b, a
-
-    ; TODO: Fix
-    ld hl, PENDING_ROW_DATA
-    ld d, SCRN_VX_B
-    MEMCPY bc, hl, d
+    call init_row_write
 
     pop hl
     pop de
@@ -493,43 +438,30 @@ enqueue_col_write:
     ; If X >= MAP_WIDTH, write 0s
     ld a, [CURRENT_MAP_WIDTH]
     dec a
-    cp c
+    cp b
     jr c, .zero_row
 
-    ; HL = CURRENT_MAP_PTR
+    ; HL = CURRENT_MAP_PTR + X
     ld a, [CURRENT_MAP_PTR]
     ld l, a
     ld a, [CURRENT_MAP_PTR + 1]
     ld h, a
-
-    ; HL = CURRENT_MAP_PTR + Y * MAP_WIDTH + X
-    ld d, 0
-    ld a, [CURRENT_MAP_WIDTH]
-    ld e, a
-    ld a, c
-
-.get_map_col_ptr:
-    or a
-    jr z, .copy_map_column
-    add hl, de
-    dec a
-    jr .get_map_col_ptr
-
-.copy_map_column:
     ld a, b
     ADD16 hl
 
     ; B = BYTES_LEFT
-    ld b, SCRN_VY_B
+    ld b, COL_BUFFER_SIZE
     ld de, PENDING_COL_DATA
 
-    ; Note: Can skip checking BYTES_LEFT > 0 in this loop. If there were
-    ; SCRN_VY_B zeros to write, then X would be 1 greater and we would have
-    ; jumped into .zero_row before reaching this code
 .pad_left:
     ; Check Y < 0
     bit 7, c
-    jr z, .copy_middle
+    jr z, .adjust_hl
+
+    ; Check BYTES_LEFT > 0
+    ld a, b
+    or a
+    ret z
 
     ; *ROW++ = 0
     xor a
@@ -541,6 +473,25 @@ enqueue_col_write:
     dec b
     jr .pad_left
 
+.adjust_hl:
+    push de
+
+    ; HL += Y * MAP_WIDTH
+    ld d, 0
+    ld a, [CURRENT_MAP_WIDTH]
+    ld e, a
+    ld a, c
+
+.get_map_col_ptr:
+    or a
+    jr z, .done_get_col_map_ptr
+    add hl, de
+    dec a
+    jr .get_map_col_ptr
+
+.done_get_col_map_ptr:
+    pop de
+
 .copy_middle:
     ; Check Y < MAP_HEIGHT
     ld a, [CURRENT_MAP_HEIGHT]
@@ -586,43 +537,133 @@ enqueue_col_write:
 .zero_row:
     ld hl, PENDING_COL_DATA
     xor a
-    ld c, SCRN_VY_B
+    ld c, COL_BUFFER_SIZE
 .zero_row_loop:
     ld [hl+], a
     dec c
     jr nz, .zero_row_loop
     ret
 
-; Compute the distance (modulo 32) between two values
-; @param b Value 1
-; @param c Value 2
-mmd2:
-    push bc
 
-    ld a, b
-    sub c
+; Basically enqueue_col_write but for entire row, used during map load
+; TODO: Deduplicate
+; @param b Map X coordinate (signed)
+; @param c Map Y coordinate (signed)
+; @param hl Where to write the column in map VRAM
+; @destroy All registers
+init_row_write:
+    ; If Y < 0, write 0s
+    bit 7, c
+    jr nz, .zero_row
 
-    cpl
-    ld b, a
-    cpl
+    ; If Y >= MAP_HEIGHT, write 0s
+    ld a, [CURRENT_MAP_HEIGHT]
+    dec a
+    cp c
+    jr c, .zero_row
+
+    ; DE = HL
+    ld d, h
+    ld e, l
+
+    ; HL = CURRENT_MAP_PTR
+    ld a, [CURRENT_MAP_PTR]
+    ld l, a
+    ld a, [CURRENT_MAP_PTR + 1]
+    ld h, a
+
+    push de
+
+    ; HL = CURRENT_MAP_PTR + Y * MAP_WIDTH
+    ld d, 0
+    ld a, [CURRENT_MAP_WIDTH]
+    ld e, a
+    ld a, c
 
+.get_map_row_ptr:
+    or a
+    jr z, .copy_map_row
+    add hl, de
+    dec a
+    jr .get_map_row_ptr
+
+.copy_map_row:
+    pop de
+
+    ; C = BYTES_LEFT
+    ld c, SCRN_VX_B
+
+    ; If X > 0, increment map pointer by X
+    ; TODO: Remove this branch by adding X before copy_middle
     bit 7, b
-    jr z, .0
+    jr nz, .pad_left
 
-    ld b, a
+    ld a, b
+    ADD16 hl
+    jr .copy_middle
+
+.pad_left:
+    ; Check X < 0
+    bit 7, b
+    jr z, .copy_middle
+
+    ; Check BYTES_LEFT > 0
+    ld a, c
+    or a
+    ret z
 
-.0:
-    ld a, 32
-    sub b
+    ; *ROW++ = 0
+    ld [de], a
+    inc de
 
+    ; X++, BYTES_LEFT--
+    inc b
+    dec c
+    jr .pad_left
+
+.copy_middle:
+    ; Check X < MAP_WIDTH
+    ld a, [CURRENT_MAP_WIDTH]
+    dec a
     cp b
-    jr nc, .1
+    jr c, .pad_right
 
-    pop bc
-    ret
+    ; Check BYTES_LEFT > 0
+    ld a, c
+    or a
+    ret z
 
-.1:
-    ld a, b
+    ; *ROW++ = *MAP++
+    ld a, [hl+]
+    ld [de], a
+    inc de
 
-    pop bc
-    ret
+    ; X++, BYTES_LEFT--
+    inc b
+    dec c
+    jr .copy_middle
+
+.pad_right:
+    ; Check BYTES_LEFT > 0
+    ld a, c
+    or a
+    ret z
+
+    ; *ROW++ = 0
+    xor a
+    ld [de], a
+    inc de
+
+    ; X++, BYTES_LEFT--
+    inc b
+    dec c
+    jr .pad_right
+
+.zero_row:
+    xor a
+    ld c, SCRN_VX_B
+.zero_row_loop:
+    ld [hl+], a
+    dec c
+    jr nz, .zero_row_loop
+    ret
\ No newline at end of file