dither4/dither4.s

SAVMSC = $58
VDSLST = $200
VDSLSTL = $200
VDSLSTH = $201
COLPF0 = $D016
COLPF1 = $D017
COLPF2 = $D018
COLPF3 = $D019
COLBK = $D01A

AUDC1 = $D201
DMACTL = $D400
DLISTL = $D402
DLISTH = $D403
WSYNC = $D40A
VCOUNT = $D40B
NMIEN = $D40E

temp1l = $80
temp1h = $81
temp1 = temp1l
temp2l = $82
temp2h = $83
temp2 = temp2l
sample_ptrl = $84
sample_ptrh = $85
sample_ptr = sample_ptrl
scanline = $86
frame_counter = $89

height = 160
bytes_per_line = 40
pages_per_frame = 32
lines_per_frame = 262
;scanline_offset = 31 + (40 - 24) / 2
scanline_offset = 46
scanline_max = (lines_per_frame - scanline_offset) / 2

.data

.import frame1_top
.import frame1_bottom
.import frame1_palette1_even
.import frame1_palette1_odd
.import frame1_palette2_even
.import frame1_palette2_odd
.import frame1_palette3_even
.import frame1_palette3_odd
.import displaylist

.code

.export start

.proc start
    ; Disable display DMA
    lda #$00
    sta DMACTL

    ; Disable VBI and DLI but allow Reset
    lda #$20
    sta NMIEN

    ; Set up the display list
    lda #.lobyte(displaylist)
    sta DLISTL
    lda #.hibyte(displaylist)
    sta DLISTH

    ; Set up the DLI handler
    lda #.lobyte(dli_handler)
    sta VDSLSTL
    lda #.hibyte(dli_handler)
    sta VDSLSTH

    ; Disable VBI but allow Reset and DLI
    lda #$a0
    sta NMIEN

    ; Manually wait for first scan line
wait_vblank:
    sta WSYNC
    lda VCOUNT
    bne wait_vblank

    ; Re-enable display DMA
    lda #$22
    sta DMACTL

wait_start:
    ; Wait for the vblank
    ; Resynchronize the scanline counter
wait_loop:
    ldy VCOUNT    ; 4 cycles
    bne wait_loop ; 2 cycles

    .macro inner_scanline frame_offset, line_offset
        ; Y should be VCOUNT at entry
        ; it'll fire on unused lines, but harmlessly
        ;ldy scanline ; 3 cyc
        ;inc scanline ; 5 cyc

        ; 23-26 cycles before break
        ; Leisurely memory fetches
        lda frame1_palette1_even + frame_offset + line_offset - scanline_offset / 2,y ; 4/5 @FIXME alternate
        pha                                                                           ; 3
        ldx frame1_palette2_even + frame_offset + line_offset - scanline_offset / 2,y ; 4/5
        lda frame1_palette3_even + frame_offset + line_offset - scanline_offset / 2,y ; 4/5
        tay                                                                           ; 2
        pla                                                                           ; 3
        ; Wait for horizontal blank
        sta WSYNC                                                                     ; 4

        ; 12 cycles after break
        ; Update color registers as fast as possible
        sta COLPF0 ; 4
        stx COLPF1 ; 4
        sty COLPF2 ; 4
    .endmacro

    .macro run_frame frame_offset
    .scope
        ; each scanline is 228 color clocks
        ; that's 114 CPU cycles
        ; minus 41-43 for DMA leaves 71-73 clock cycles per line

    each_scanline_pair:
        sty scanline ; 3 cycles
        inner_scanline frame_offset, 0 ; 23-26 cycles before break, 12 cycles after

        ldy scanline   ; 3 cycles
        inner_scanline frame_offset, 128 ; 23-26 cycles before break, 12 cycles after

        ; pair cleanup: 6 cycles
        ldy VCOUNT             ; 4 cycles
        bne each_scanline_pair ; 2 cycles

        ; frame cleanup: 11 cycles
        lda frame_counter ; 3 cycles
        eor #1            ; 2 cycles
        sta frame_counter ; 3 cycles
        ;jmp wait_start    ; 3 cycles
        jmp each_frame    ; 3 cycles
    .endscope
    .endmacro

each_frame:
run_frame1:
    run_frame 0

.endproc


.proc dli_handler
    lda #0
    sta frame_counter
    rti
.endproc