+mov ax, cx ; ax = idx
+mov bl, COLS ; bl = COLS
+div bl ; al = idx / COLS, ah = idx % COLS
+
+mov [bp - 1], al ; row = idx / COLS
+mov [bp - 2], ah ; row = idx % COLS
+mov byte [bp - 3], 0 ; neighbours = 0
+
+mov byte [bp - 4], -1 ; i = -1
+.i: ; do {
+
+ mov byte [bp - 5], -1 ; j = -1
+.j: ; do {
+
+ mov al, [bp - 4]
+ or al, [bp - 5] ; if (!i && !j)
+ jz .continue ; continue
+
+ mov al, [bp - 1]
+ add al, [bp - 4] ; al = row + i
+
+ cmp al, ROWS ; if ((unsigned)(row + i) >= ROWS)
+ jae .continue ; continue
+
+ mov ah, [bp - 2]
+ add ah, [bp - 5] ; ah = col + j
+
+ cmp ah, COLS ; if ((unsigned)(col + j) >= COLS)
+ jae .continue ; continue
+
+ movzx bx, ah ; bx = col + j
+
+ mov ah, COLS
+ mul ah ; ax = (row + i) * COLS
+
+ add bx, ax ; bx = cell index
+ shl bx, 1 ; bx = grid offset
+
+ mov dx, [es:si + bx] ; dl = cell state
+
+ cmp dl, ALIVE ; if (!ALIVE)
+ jne .continue ; continue
+
+ inc byte [bp - 3] ; neighbours++
+
+.continue:
+ inc byte [bp - 5]
+ cmp byte [bp - 5], 1
+ jle .j ; } while (++j <= 1)
+
+ inc byte [bp - 4]
+ cmp byte [bp - 4], 1
+ jle .i ; } while (++i <= 1)
+
+movzx ax, byte [bp - 3] ; ax = # alive neighbours
+
+leave
+ret
+
+; update_cell() - write next cell state ---------------------------------------
+
+; input:
+; es:si -> current grid
+; es:di -> next grid
+; cx -> cell index
+
+; output:
+; [es:di + cx * 2] = updated cell state
+
+; clobbers ax, bx, dx
+
+update_cell:
+
+call alive_neighbours ; ax = # alive neighbours
+
+mov bx, cx
+shl bx, 1
+
+mov dx, [es:si + bx] ; dl = cell current state
+
+cmp dl, ALIVE ; if (ALIVE) {
+jne .else
+
+shr ax, 1
+cmp ax, 1 ; if (n != 2 || n != 3)
+jne .dead ; return DEAD
+
+jmp .alive ; return ALIVE
+
+.else: ; }
+
+cmp ax, 3 ; if (n != 3)
+jne .dead ; return DEAD
+
+.alive: ; return ALIVE
+mov dl, ALIVE
+jmp .write
+
+.dead:
+mov dl, DEAD
+
+.write:
+mov [es:di + bx], dx
+
+ret
+
+; update_grid() - write next grid (inactive vga page) -------------------------
+
+; clobbers di, si, ax, bx, cx, dx
+
+update_grid:
+
+mov si, [currvgapg] ; es:si -> current page
+
+mov di, VGAPGSZ
+xor di, si ; es:di -> next page
+
+mov cx, COLS * ROWS - 1 ; for each cell (index)
+
+.update_loop: ; do {
+
+ call update_cell
+
+ dec cx
+ jns .update_loop ; } while (--cx > 0)
+
+ret
+
+; halt() - stop program execution ---------------------------------------------