; lenser -- FLI multicolor viewer, NTSC timing (self-contained)
;
; Same as fli.s but the inner loop is one NOP (2 cycles) longer so it self-syncs
; to the NTSC 65-cycle raster line (25 free CPU cycles per badline) vs PAL's 63.
;
; Re-points the VIC video matrix ($D018) and forces a badline ($D011 yscroll)
; on every visible raster line via a cycle-timed loop, giving per-line colour.
;
; Memory layout of appended data (loads from $4000):
;   $4000+L*$400  screen RAM for line L (L=0..7), 1000 bytes each
;   $6000         bitmap 8000   (offset $2000 in VIC bank 1)
;   $8000         colour RAM 1000  (copied to $D800)
;   $83E8         background 1
;
; assembled by viewer/assemble.py via xa

        ; BASIC autostart, SYS 2061
        * = $0801
        .word basicend
        .word 10
        .byte $9e
        .byte "2061"
        .byte 0
basicend:
        .word 0                   ; ML begins at $080D

start:
        sei

        ; copy colour RAM $8000 -> $D800 (1024 bytes)
        ldx #0
ccopy:
        lda $8000,x
        sta $d800,x
        lda $8100,x
        sta $d900,x
        lda $8200,x
        sta $da00,x
        lda $8300,x
        sta $db00,x
        inx
        bne ccopy

        ; build per-line tables. d018tab = ((line AND 7) << 4) OR $08
        ; d011tab = $38 OR (line AND 7)   = BMM+DEN+RSEL plus yscroll
        ldx #0
btab:
        txa
        and #$07
        asl
        asl
        asl
        asl
        ora #$08
        sta d018tab,x
        ; yscroll must equal (raster AND 7) to force a badline; the first
        ; displayed raster is 51, so line x maps to raster 51+x.
        txa
        clc
        adc #$03
        and #$07
        ora #$38
        sta d011tab,x
        inx
        bne btab

        ; VIC bank 1 ($4000-$7FFF)
        lda $dd00
        and #$fc
        ora #$02
        sta $dd00

        lda $83e8
        sta $d021                 ; background
        lda #$00
        sta $d020                 ; border black
        lda #$d8
        sta $d016                 ; multicolor on

        ; raster IRQ setup
        lda #<irq1
        sta $0314
        lda #>irq1
        sta $0315
        lda #$7f
        sta $dc0d                 ; disable CIA timer IRQs
        sta $dd0d
        lda $dc0d
        lda $dd0d
        lda #$01
        sta $d01a                 ; enable raster IRQ
        lda #$30
        sta $d012                 ; line 48
        lda $d011
        and #$7f
        sta $d011
        asl $d019                 ; ack
        cli
hang:
        jmp hang

; first IRQ, arrives with jitter on line 48, sets up the stabilised one
irq1:
        lda #<irq2
        sta $0314
        lda #>irq2
        sta $0315
        inc $d012                 ; fire again next line (49)
        asl $d019                 ; ack
        tsx
        cli
        ; burn the rest of the line so irq2 fires inside this NOP slide
        .dsb 40,$ea               ; 40 NOPs

; stabilised IRQ on line 49, runs the FLI loop
irq2:
        txs                       ; restore sp from irq1
        ; cancel the remaining 0/1-cycle jitter: read the raster twice; the
        ; branch is 2 or 3 cycles depending on alignment, normalising entry.
        lda $d012
        cmp $d012
        beq jl
jl:
        ; fixed delay so the first store lands before the c-access window
        ldx #$0d
d0:     dex
        bne d0
        nop
        nop

        ldx #$00
fliloop:
        lda d011tab,x             ; force badline (yscroll = line&7)
        sta $d011
        lda d018tab,x             ; point video matrix at screen[line]
        sta $d018
        nop                       ; NTSC extra 2 cycles for 65 cycle line
        inx
        cpx #200
        bne fliloop

        ; bottom border: leave bitmap on but stop forcing badlines
        lda #<irq1
        sta $0314
        lda #>irq1
        sta $0315
        lda #$30
        sta $d012
        asl $d019
        jmp $ea81                 ; restore regs + RTI

d018tab:
        .dsb 256,0
d011tab:
        .dsb 256,0