Init code

When the NES is powered on or reset, the program should do the following within a fixed bank:

• Set IRQ ignore bit (not strictly necessary as the 6502 sets this flag on all interrupts, including RESET, but it allows program code to simulate a reset by JMP ($FFFC))
• Disable PPU NMIs and rendering
• Initialize stack pointer
• Initialize the mapper (if any)

The init code after this point may be placed either in the fixed bank or in a separate bank using a bankswitch followed by a JMP:

• Disable decimal mode (not strictly necessary as the 2A03 has no decimal mode, but it maintains compatibility with generic 6502 debuggers)
• If using a mapper that generates IRQs, disable APU timer IRQs
• Disable DMC IRQs [1]
• Set all RAM that your program uses to a known state. This often involves clearing internal RAM (@ $0000-$07FF) (and PRG RAM if needed (@ $6000-$7FFF)), except that which is intended to survive a reset (such as high scores). However, you don't need to set up an area of RAM that only one part of a program uses; you can set that in the same part where it's used, and in many cases, you'll need to set it up multiple times as part of the program is run multiple times.
• Wait at least 30,000 cycles (see PPU power up state) before reading or writing registers $2003 through $2007. This is commonly done by waiting for the PPU to signal the start of vertical blank twice through $2002.

Some mappers have no fixed bank because they switch all 32 KB of PRG at a time. These include AxROM, BxROM, GxROM, and some configurations of MMC1. You'll have to put the interrupt vectors and the code up to the end of the JMP in a separate section that is duplicated in each bank. Often, the 256-byte page $FF00-$FFFF contains the vectors, the start of the init code, and a "trampoline" for jumps from code in one bank to code in another.

Sample implementation:

reset:
    sei        ; ignore IRQs
    cld        ; disable decimal mode
    ldx #$40
    stx $4017  ; disable APU frame IRQ
    ldx #$ff
    txs        ; Set up stack
    inx        ; now X = 0
    stx $2000  ; disable NMI
    stx $2001  ; disable rendering
    stx $4010  ; disable DMC IRQs

    ; Optional (omitted):
    ; Set up mapper and jmp to further init code here.

    ; Clear the vblank flag, so we know that we are waiting for the
    ; start of a vertical blank and not powering on with the
    ; vblank flag spuriously set
    bit $2002

    ; First of two waits for vertical blank to make sure that the
    ; PPU has stabilized
@vblankwait1:  
    bit $2002
    bpl @vblankwait1

    ; We now have about 30,000 cycles to burn before the PPU stabilizes.
    ; One thing we can do with this time is put RAM in a known state.
    ; Here we fill it with $00, which matches what (say) a C compiler
    ; expects for BSS.  Conveniently, X is still 0.
    txa
@clrmem:
    sta $000,x
    sta $100,x
    sta $300,x
    sta $400,x
    sta $500,x
    sta $600,x
    sta $700,x  ; Remove this if you're storing reset-persistent data

    ; We skipped $200,x on purpose.  Usually, RAM page 2 is used for the
    ; display list to be copied to OAM.  OAM needs to be initialized to
    ; $EF-$FF, not 0, or you'll get a bunch of garbage sprites at (0, 0).

    inx
    bne @clrmem

    ; Other things you can do between vblank waits are set up audio
    ; or set up other mapper registers.
   
@vblankwait2:
    bit $2002
    bpl @vblankwait2

At this point, the program can fill the nametables, fill the pattern tables (if the board uses CHR RAM), fill the palette, and start displaying things.