GBATEK - GBA/NDS Technical Info

GBATEK
Gameboy Advance / Nintendo DS / DSi / 3DS - Technical Info - Extracted from no$gba version 3.05
About this Document

GBA Reference

Overview
GBA Technical Data
GBA Memory Map
GBA I/O Map

Hardware Programming
GBA LCD Video Controller
GBA Sound Controller
GBA Timers
GBA DMA Transfers
GBA Communication Ports
GBA Keypad Input
GBA Interrupt Control
GBA System Control
GBA Cartridges
GBA Unpredictable Things

Other
ARM CPU Reference
BIOS Functions
External Connectors

NDS Reference

Overview
DS Technical Data
DS I/O Maps
DS Memory Maps

Hardware Programming
DS Memory Control
DS Video
DS 3D Video
DS Sound
DS System and Built-in Peripherals
DS Cartridges, Encryption, Firmware
DS Xboo
DS Wireless Communications

Other
BIOS Functions
ARM CPU Reference
External Connectors

3DS Reference

3DS Reference (under construction)

DSi Reference

DSi Basic Differences to NDS

New Hardware Features
DSi I/O Map
DSi Control Registers (SCFG)
DSi XpertTeak (DSP)
DSi New Shared WRAM
DSi New DMA (NDMA)
DSi Microphone and SoundExt
DSi Advanced Encryption Standard (AES)
DSi Cartridge Header
DSi Touchscreen/Sound Controller
DSi I2C Bus
DSi Cameras
DSi SD/MMC Protocol and I/O Ports
DSi SD/MMC Filesystem
DSi Atheros Wifi SDIO Interface
DSi Atheros Wifi Internal Hardware
DSi GPIO Registers
DSi Console IDs
DSi Unknown Registers
DSi Notes
DSi Exploits
DSi Regions

CPU Reference

General ARM7TDMI Information
ARM CPU Overview
ARM CPU Register Set
ARM CPU Flags & Condition Field (cond)
ARM CPU 26bit Memory Interface
ARM CPU Exceptions
ARM CPU Memory Alignments

Further Information
ARM Pseudo Instructions and Directives
ARM CP15 System Control Coprocessor
ARM CPU Instruction Cycle Times
ARM CPU Versions
ARM CPU Data Sheet

CPU 32bit ARM Mode

ARM 32bit Opcodes (ARM Code)
ARM Instruction Summary
ARM Branch and Branch with Link (B, BL, BX, BLX, SWI, BKPT)
ARM Data Processing (ALU)
ARM Multiply and Multiply-Accumulate (MUL, MLA)
ARM Special ARM9 Instructions (CLZ, QADD/QSUB)
ARM PSR Transfer (MRS, MSR)
ARM Memory: Single Data Transfer (LDR, STR, PLD)
ARM Memory: Halfword, Doubleword, and Signed Data Transfer
ARM Memory: Block Data Transfer (LDM, STM)
ARM Memory: Single Data Swap (SWP)
ARM Coprocessor (MRC/MCR, LDC/STC, CDP, MCRR/MRRC)

CPU 16bit THUMB Mode

ARM 16bit Opcodes (THUMB Code)
When operating in THUMB state, cut-down 16bit opcodes are used.
THUMB is supported on T-variants of ARMv4 and up, ie. ARMv4T, ARMv5T, etc.
THUMB Instruction Summary
THUMB Register Operations (ALU, BX)
THUMB Memory Load/Store (LDR/STR)
THUMB Memory Addressing (ADD PC/SP)
THUMB Memory Multiple Load/Store (PUSH/POP and LDM/STM)
THUMB Jumps and Calls

GBA Reference

GBA Technical Data

CPU Modes

  ARM Mode     ARM7TDMI 32bit RISC CPU, 16.78MHz, 32bit opcodes (GBA)
  THUMB Mode   ARM7TDMI 32bit RISC CPU, 16.78MHz, 16bit opcodes (GBA)
  CGB Mode     Z80/8080-style 8bit CPU, 4.2MHz or 8.4MHz  (CGB compatibility)
  DMG Mode     Z80/8080-style 8bit CPU, 4.2MHz (monochrome gameboy compatib.)

Internal Memory

  BIOS ROM     16 KBytes
  Work RAM     288 KBytes (Fast 32K on-chip, plus Slow 256K on-board)
  VRAM         96 KBytes
  OAM          1 KByte (128 OBJs 3x16bit, 32 OBJ-Rotation/Scalings 4x16bit)
  Palette RAM  1 KByte (256 BG colors, 256 OBJ colors)

Video

  Display      240x160 pixels (2.9 inch TFT color LCD display)
  BG layers    4 background layers
  BG types     Tile/map based, or Bitmap based
  BG colors    256 colors, or 16 colors/16 palettes, or 32768 colors
  OBJ colors   256 colors, or 16 colors/16 palettes
  OBJ size     12 types (in range 8x8 up to 64x64 dots)
  OBJs/Screen  max. 128 OBJs of any size (up to 64x64 dots each)
  OBJs/Line    max. 128 OBJs of 8x8 dots size (under best circumstances)
  Priorities   OBJ/OBJ: 0-127, OBJ/BG: 0-3, BG/BG: 0-3
  Effects      Rotation/Scaling, alpha blending, fade-in/out, mosaic, window
  Backlight    GBA SP only (optionally by light on/off toggle button)

Sound

  Analogue     4 channel CGB compatible (3x square wave, 1x noise)
  Digital      2 DMA sound channels
  Output       Built-in speaker (mono), or headphones socket (stereo)

Controls

  Gamepad      4 Direction Keys, 6 Buttons

Communication Ports

  Serial Port  Various transfer modes, 4-Player Link, Single Game Pak play

External Memory

  GBA Game Pak max. 32MB ROM or flash ROM + max 64K SRAM
  CGB Game Pak max. 32KB ROM + 8KB SRAM (more memory requires banking)

Case Dimensions

  Size (mm)    GBA: 145x81x25 - GBA SP: 82x82x24 (closed), 155x82x24 (stretch)

Power Supply

  Battery GBA  GBA: 2x1.5V DC (AA), Life-time approx. 15 hours
  Battery SP   GBA SP: Built-in rechargeable Lithium ion battery, 3.7V 600mAh
  External     GBA: 3.3V DC 350mA - GBA SP: 5.2V DC 320mA

 ----------------------------------------------------------------------------

Original Gameboy Advance (GBA)

        ____._____________...___.____
   ____/    :  CARTRIDGE  SIO   :    \____
  | L       _____________________  LED  R |
  |        |                     |        |
  |  _||_  |   2.9" TFT SCREEN   |    (A) |
  | |_  _| | 240x160pix  61x40mm | (B)    |
  |   ||   |    NO BACKLIGHT     |  ::::  |
  |        |                     | SPEAKR |
  | STRT() |_____________________|  ::::  |
  | SLCT()     GAME BOY ADVANCE    VOLUME |
  |____  OFF-ON  BATTERY 2xAA PHONES  _==_|
       \__.##.__________________,,___/

GBA SP (GBA SP)

   _______________________                                 _
  | _____________________ |                               / /
  ||                     ||                              / /
  ||   2.9" TFT SCREEN   ||                             / /
  || 240x160pix  61x40mm ||                            / /
  ||   WITH BACKLIGHT    ||                           / /
  ||                     ||     GBA SP SIDE VIEWS    / /
  ||_____________________||                         / /
  |  GAME BOY ADVANCE SP  |   _____________________(_)
  |_______________________|  |. . . . . . . .'.'.   _|
  |_|________|________|_|_|  |_CARTRIDGE_:_BATT._:_|_| <-- EXT1/EXT2
  |L    EXT1     EXT2    R|
  |          (*)      LEDSo   _____________________ _
  (VOL_||_           (A)  o  |_____________________(_)
  |  |_  _| ,,,,,(B)      |  |. . . . . . . .'.'.   _|
  |    ||   ;SPK;         |  |_CARTRIDGE_:_BATT._:_|_| <-- EXT1/EXT2
  |         '''''      ON #                         _ _____________________
  |       SLCT STRT    OFF#   _____________________(_)_____________________|
  | CART.  ()   ()        |  |. . . . . . . .'.'.   _|
  |_:___________________:_|  |_CARTRIDGE_:_BATT._:_|_| <-- EXT1/EXT2

Gameboy Micro (GBA Micro)

      ________________SIO_______________
     | L      __________________      R |
     |       |     GBA-MICRO    |       |
     | _||_  |  2.0" TFT SCREEN |    (A)| +
     ||_  _| |240x160pix 42x28mm| (B)   |VOL
     |  ||   |     BACKLIGHT    |       | -
     |       |__________________|  ...  |
     |___________SELECT__START__________|
       PWR   <--- CARTRIDGE SLOT ---> PHONES

Nintendo DS (NDS)

     _____________________________________
    |        _____________________        |
    |       |                     |       |
    |       |    3" TFT SCREEN    |       |
    |       | 256x192pix  61x46mm |       |
    |       |      BACKLIGHT      |       |
    | ::::: |    Original NDS     | ::::: |
    | ::::: |_____________________| ::::: |
   _|        _          ______   _        |_  <-- gap between screens: 22mm
  |L|_______| |________|      |_| |_______|R|     (equivalent to 90 pixels)
  |_______   _____________________   _______|
  |  PWR  | |                     | |SEL STA|
  |   _   | |    3" TFT SCREEN    | |       |
  | _| |_ | | 256x192pix  61x46mm | |   X   |
  ||_   _|| |      BACKLIGHT      | | Y   A |
  |  |_|  | |    TOUCH SCREEN     | |   B   |
  |       | |_____________________| |       |
  |_______|             NintendoDS  |_______|
  |         MIC                LEDS         |
  |_________________________________________|
       VOL        SLOT2(GBA)     MIC/PHONES

Nintendo DS Lite (NDS-Lite)

     _____________________________________
    |        _____________________        |
    |       |                     |       |
    |       |    3" TFT SCREEN    |       |
    |  ...  | 256x192pix  61x46mm |  ...  |
    |  ...  |      BACKLIGHT      |  ...  |
    |       |      NDS-LITE       |       |
    |       |_____________________|       |
    |___  _ _ _ _ _ _ _ _ _ _ _ _ _ _ ____|   <-- gap between screens: 23mm
   L| _ |_____________MIC____________|LEDS|R
    |   _    _____________________        |
    | _| |_ |                     |   X   |
    ||_   _||    3" TFT SCREEN    | Y   A |PWR
    |  |_|  | 256x192pix  61x46mm |   B   |
    |       |      BACKLIGHT      |       |
    |       |    TOUCH SCREEN     |oSTART |
    |       |_____________________|oSELECT|
    |_____________________________________|
       VOL        SLOT2(GBA)     MIC/PHONES

Nintendo DSi (DSi)

     _____________________________________
    |        _____________________        |
    |       |                     |   O o | <-- CAM (O) and LED (o)
    |       |   3.25" TFT SCREEN  |       |     (on backside)
    |       | 256x192pix  66x50mm |       |
    |       |      BACKLIGHT      |       |
    |  __   |         DSi         |   __  |
    | (__)  |_____________________|  (__) |
    |___  _ _ _ _ _ _ _ _ _ _ _ _ _ _ ____|  <-- gap between screens: 23mm
   L|LEDS|__________CAM__MIC_________| __ |R                   (88 pixels)
  + |   _    _____________________        |
 VOL| _| |_ |                     |   X   | <-- SD Card Slot
  - ||_   _||   3.25" TFT SCREEN  | Y   A |
    |  |_|  | 256x192pix  66x50mm |   B   |
    |       |      BACKLIGHT      |       |
    |       |    TOUCH SCREEN     |oSTART |
    | POWERo|_____________________|oSELECT|
    |_____________________________________|
                                 MIC/PHONES

Nintendo DSi XL

  As DSi, but bigger case, and bigger 4.2" screens

Gameboy Player (Gamecube Joypad) (GBA Player)

                   _________
       L____-------         -------____R
       /   ___   \           /   (Y)   \Z
      /   / O \   | (START) |        (X)\   Z      = Gameboy Player Menu
     |    \___/    \_______/      (A)    |  X or Y = Select button
     |\         _   \     /    (B)      /|
     | \___   _| |_  \   /   ___    ___/ |  optionally X/Y can be
     |    |\ |_   _| /   \  / C \  /|    |  swapped with L/R (?)
     |    | \  |_|  /     \ \___/ / |    |
     |    |  \_____/       \_____/  |    |  analogue sticks = ?
      \__/                           \__/

Gameboy Player (Gamecube Bongos) (GBA Player)

       _______     _______
      /   Y   \   /   X   \   Y/B = left bongo rear/front side
     | . . . . |_| . . . . |  X/A = right bongo rear/front side
     |    B    |R|    A    |  S   = start/pause button
     |\_______/|_|\_______/|  R   = microphone (triggers R button)
     |\_______/|S|\_______/|
     |         |_|         |  (the X/Y inputs can be assigned to
     |\_______/| |\_______/|  GBA R/L inputs in GBA player setup)
      \_______/   \_______/

The GBA's separate 8bit/32bit CPU modes cannot be operated simultaneously. Switching is allowed between ARM and THUMB modes only (that are the two GBA modes).
This manual does not describe CGB and DMG modes, both are completely different than GBA modes, and both cannot be accessed from inside of GBA modes anyways.

Gameboy Player
An GBA Adapter for the Gamecube console; allowing to play GBA games on a television set.
GBA Gameboy Player

GBA SP Notes
Deluxe version of the original GBA. With backlight, new folded laptop-style case, and built-in rechargeable battery. Appears to be 100% compatible with GBA, there seems to be no way to detect SPs by software.

Gameboy Micro (GBA Micro)
Minituarized GBA. Supports 32bit GBA games only (no 8bit DMG/CGB games). The 256K Main RAM is a bit slower than usually (cannot be "overclocked via port 4000800h).

Nintendo DS (Dual Screen) Notes
New handheld with two screens, backwards compatible with GBA games, it is NOT backwards compatible with older 8bit games (mono/color gameboys) though..
Also, the DS has no link port, so that GBA games will thus work only in single player mode, link-port accessoires like printers cannot be used, and most unfortunately multiboot won't work (trying to press Select+Start at powerup will just lock up the DS).

iQue Notes
iQue is a brand name used by Nintendo in China, iQue GBA and iQue DS are essentially same as Nintendo GBA and Nintendo DS.
The iQue DS contains a larger firmware chip (the charset additionally contains about 6700 simplified chinese characters), the bootmenu still allows to select (only) six languages (japanese has been replaced by chinese). The iQue DS can play normal international NDS games, plus chinese dedicated games. The latter ones won't work on normal NDS consoles (that, reportedly simply due to a firmware-version check contained in chinese dedicated games, aside from that check, the games should be fully compatible with NDS consoles).

GBA Memory Map

General Internal Memory

  00000000-00003FFF   BIOS - System ROM         (16 KBytes)
  00004000-01FFFFFF   Not used
  02000000-0203FFFF   WRAM - On-board Work RAM  (256 KBytes) 2 Wait
  02040000-02FFFFFF   Not used
  03000000-03007FFF   WRAM - On-chip Work RAM   (32 KBytes)
  03008000-03FFFFFF   Not used
  04000000-040003FE   I/O Registers
  04000400-04FFFFFF   Not used

Internal Display Memory

  05000000-050003FF   BG/OBJ Palette RAM        (1 Kbyte)
  05000400-05FFFFFF   Not used
  06000000-06017FFF   VRAM - Video RAM          (96 KBytes)
  06018000-06FFFFFF   Not used
  07000000-070003FF   OAM - OBJ Attributes      (1 Kbyte)
  07000400-07FFFFFF   Not used

External Memory (Game Pak)

  08000000-09FFFFFF   Game Pak ROM/FlashROM (max 32MB) - Wait State 0
  0A000000-0BFFFFFF   Game Pak ROM/FlashROM (max 32MB) - Wait State 1
  0C000000-0DFFFFFF   Game Pak ROM/FlashROM (max 32MB) - Wait State 2
  0E000000-0E00FFFF   Game Pak SRAM    (max 64 KBytes) - 8bit Bus width
  0E010000-0FFFFFFF   Not used

Unused Memory Area

  10000000-FFFFFFFF   Not used (upper 4bits of address bus unused)

Default WRAM Usage
By default, the 256 bytes at 03007F00h-03007FFFh in Work RAM are reserved for Interrupt vector, Interrupt Stack, and BIOS Call Stack. The remaining WRAM is free for whatever use (including User Stack, which is initially located at 03007F00h).

Address Bus Width and CPU Read/Write Access Widths
Shows the Bus-Width, supported read and write widths, and the clock cycles for 8/16/32bit accesses.

  Region        Bus   Read      Write     Cycles
  BIOS ROM      32    8/16/32   -         1/1/1
  Work RAM 32K  32    8/16/32   8/16/32   1/1/1
  I/O           32    8/16/32   8/16/32   1/1/1
  OAM           32    8/16/32   16/32     1/1/1 *
  Work RAM 256K 16    8/16/32   8/16/32   3/3/6 **
  Palette RAM   16    8/16/32   16/32     1/1/2 *
  VRAM          16    8/16/32   16/32     1/1/2 *
  GamePak ROM   16    8/16/32   -         5/5/8 **/***
  GamePak Flash 16    8/16/32   16/32     5/5/8 **/***
  GamePak SRAM  8     8         8         5     **

Timing Notes:

  *   Plus 1 cycle if GBA accesses video memory at the same time.
  **  Default waitstate settings, see System Control chapter.
  *** Separate timings for sequential, and non-sequential accesses.
  One cycle equals approx. 59.59ns (ie. 16.78MHz clock).

All memory (except GamePak SRAM) can be accessed by 16bit and 32bit DMA.

GamePak Memory
Only DMA3 (and the CPU of course) may access GamePak ROM. GamePak SRAM can be accessed by the CPU only - restricted to bytewise 8bit transfers. The SRAM region is supposed for as external FLASH backup memory, or for battery-backed SRAM.
For details about configuration of GamePak Waitstates, see:
GBA System Control

VRAM, OAM, and Palette RAM Access
These memory regions can be accessed during H-Blank or V-Blank only (unless display is disabled by Forced Blank bit in DISPCNT register).
There is an additional restriction for OAM memory: Accesses during H-Blank are allowed only if 'H-Blank Interval Free' in DISPCNT is set (which'd reduce number of display-able OBJs though).
The CPU appears to be able to access VRAM/OAM/Palette at any time, a waitstate (one clock cycle) being inserted automatically in case that the display controller was accessing memory simultaneously. (Ie. unlike as in old 8bit gameboy, the data will not get lost.)

CPU Mode Performance
Note that the GamePak ROM bus is limited to 16bits, thus executing ARM instructions (32bit opcodes) from inside of GamePak ROM would result in a not so good performance. So, it'd be more recommended to use THUMB instruction (16bit opcodes) which'd allow each opcode to be read at once.
(ARM instructions can be used at best performance by copying code from GamePak ROM into internal Work RAM)

Data Format
Even though the ARM CPU itself would allow to select between Little-Endian and Big-Endian format by using an external circuit, in the GBA no such circuit exists, and the data format is always Little-Endian. That is, when accessing 16bit or 32bit data in memory, the least significant bits are stored in the first byte (smallest address), and the most significant bits in the last byte. (Ie. same as for 80x86 and Z80 CPUs.)

GBA I/O Map

LCD I/O Registers

  4000000h  2    R/W  DISPCNT   LCD Control
  4000002h  2    R/W  -         Undocumented - Green Swap
  4000004h  2    R/W  DISPSTAT  General LCD Status (STAT,LYC)
  4000006h  2    R    VCOUNT    Vertical Counter (LY)
  4000008h  2    R/W  BG0CNT    BG0 Control
  400000Ah  2    R/W  BG1CNT    BG1 Control
  400000Ch  2    R/W  BG2CNT    BG2 Control
  400000Eh  2    R/W  BG3CNT    BG3 Control
  4000010h  2    W    BG0HOFS   BG0 X-Offset
  4000012h  2    W    BG0VOFS   BG0 Y-Offset
  4000014h  2    W    BG1HOFS   BG1 X-Offset
  4000016h  2    W    BG1VOFS   BG1 Y-Offset
  4000018h  2    W    BG2HOFS   BG2 X-Offset
  400001Ah  2    W    BG2VOFS   BG2 Y-Offset
  400001Ch  2    W    BG3HOFS   BG3 X-Offset
  400001Eh  2    W    BG3VOFS   BG3 Y-Offset
  4000020h  2    W    BG2PA     BG2 Rotation/Scaling Parameter A (dx)
  4000022h  2    W    BG2PB     BG2 Rotation/Scaling Parameter B (dmx)
  4000024h  2    W    BG2PC     BG2 Rotation/Scaling Parameter C (dy)
  4000026h  2    W    BG2PD     BG2 Rotation/Scaling Parameter D (dmy)
  4000028h  4    W    BG2X      BG2 Reference Point X-Coordinate
  400002Ch  4    W    BG2Y      BG2 Reference Point Y-Coordinate
  4000030h  2    W    BG3PA     BG3 Rotation/Scaling Parameter A (dx)
  4000032h  2    W    BG3PB     BG3 Rotation/Scaling Parameter B (dmx)
  4000034h  2    W    BG3PC     BG3 Rotation/Scaling Parameter C (dy)
  4000036h  2    W    BG3PD     BG3 Rotation/Scaling Parameter D (dmy)
  4000038h  4    W    BG3X      BG3 Reference Point X-Coordinate
  400003Ch  4    W    BG3Y      BG3 Reference Point Y-Coordinate
  4000040h  2    W    WIN0H     Window 0 Horizontal Dimensions
  4000042h  2    W    WIN1H     Window 1 Horizontal Dimensions
  4000044h  2    W    WIN0V     Window 0 Vertical Dimensions
  4000046h  2    W    WIN1V     Window 1 Vertical Dimensions
  4000048h  2    R/W  WININ     Inside of Window 0 and 1
  400004Ah  2    R/W  WINOUT    Inside of OBJ Window & Outside of Windows
  400004Ch  2    W    MOSAIC    Mosaic Size
  400004Eh       -    -         Not used
  4000050h  2    R/W  BLDCNT    Color Special Effects Selection
  4000052h  2    R/W  BLDALPHA  Alpha Blending Coefficients
  4000054h  2    W    BLDY      Brightness (Fade-In/Out) Coefficient
  4000056h       -    -         Not used

Sound Registers

  4000060h  2  R/W  SOUND1CNT_L Channel 1 Sweep register       (NR10)
  4000062h  2  R/W  SOUND1CNT_H Channel 1 Duty/Length/Envelope (NR11, NR12)
  4000064h  2  R/W  SOUND1CNT_X Channel 1 Frequency/Control    (NR13, NR14)
  4000066h     -    -           Not used
  4000068h  2  R/W  SOUND2CNT_L Channel 2 Duty/Length/Envelope (NR21, NR22)
  400006Ah     -    -           Not used
  400006Ch  2  R/W  SOUND2CNT_H Channel 2 Frequency/Control    (NR23, NR24)
  400006Eh     -    -           Not used
  4000070h  2  R/W  SOUND3CNT_L Channel 3 Stop/Wave RAM select (NR30)
  4000072h  2  R/W  SOUND3CNT_H Channel 3 Length/Volume        (NR31, NR32)
  4000074h  2  R/W  SOUND3CNT_X Channel 3 Frequency/Control    (NR33, NR34)
  4000076h     -    -           Not used
  4000078h  2  R/W  SOUND4CNT_L Channel 4 Length/Envelope      (NR41, NR42)
  400007Ah     -    -           Not used
  400007Ch  2  R/W  SOUND4CNT_H Channel 4 Frequency/Control    (NR43, NR44)
  400007Eh     -    -           Not used
  4000080h  2  R/W  SOUNDCNT_L  Control Stereo/Volume/Enable   (NR50, NR51)
  4000082h  2  R/W  SOUNDCNT_H  Control Mixing/DMA Control
  4000084h  2  R/W  SOUNDCNT_X  Control Sound on/off           (NR52)
  4000086h     -    -           Not used
  4000088h  2  BIOS SOUNDBIAS   Sound PWM Control
  400008Ah  ..   -    -         Not used
  4000090h 2x10h R/W  WAVE_RAM  Channel 3 Wave Pattern RAM (2 banks!!)
  40000A0h  4    W    FIFO_A    Channel A FIFO, Data 0-3
  40000A4h  4    W    FIFO_B    Channel B FIFO, Data 0-3
  40000A8h       -    -         Not used

DMA Transfer Channels

  40000B0h  4    W    DMA0SAD   DMA 0 Source Address
  40000B4h  4    W    DMA0DAD   DMA 0 Destination Address
  40000B8h  2    W    DMA0CNT_L DMA 0 Word Count
  40000BAh  2    R/W  DMA0CNT_H DMA 0 Control
  40000BCh  4    W    DMA1SAD   DMA 1 Source Address
  40000C0h  4    W    DMA1DAD   DMA 1 Destination Address
  40000C4h  2    W    DMA1CNT_L DMA 1 Word Count
  40000C6h  2    R/W  DMA1CNT_H DMA 1 Control
  40000C8h  4    W    DMA2SAD   DMA 2 Source Address
  40000CCh  4    W    DMA2DAD   DMA 2 Destination Address
  40000D0h  2    W    DMA2CNT_L DMA 2 Word Count
  40000D2h  2    R/W  DMA2CNT_H DMA 2 Control
  40000D4h  4    W    DMA3SAD   DMA 3 Source Address
  40000D8h  4    W    DMA3DAD   DMA 3 Destination Address
  40000DCh  2    W    DMA3CNT_L DMA 3 Word Count
  40000DEh  2    R/W  DMA3CNT_H DMA 3 Control
  40000E0h       -    -         Not used

Timer Registers

  4000100h  2    R/W  TM0CNT_L  Timer 0 Counter/Reload
  4000102h  2    R/W  TM0CNT_H  Timer 0 Control
  4000104h  2    R/W  TM1CNT_L  Timer 1 Counter/Reload
  4000106h  2    R/W  TM1CNT_H  Timer 1 Control
  4000108h  2    R/W  TM2CNT_L  Timer 2 Counter/Reload
  400010Ah  2    R/W  TM2CNT_H  Timer 2 Control
  400010Ch  2    R/W  TM3CNT_L  Timer 3 Counter/Reload
  400010Eh  2    R/W  TM3CNT_H  Timer 3 Control
  4000110h       -    -         Not used

Serial Communication (1)

  4000120h  4    R/W  SIODATA32 SIO Data (Normal-32bit Mode; shared with below)
  4000120h  2    R/W  SIOMULTI0 SIO Data 0 (Parent)    (Multi-Player Mode)
  4000122h  2    R/W  SIOMULTI1 SIO Data 1 (1st Child) (Multi-Player Mode)
  4000124h  2    R/W  SIOMULTI2 SIO Data 2 (2nd Child) (Multi-Player Mode)
  4000126h  2    R/W  SIOMULTI3 SIO Data 3 (3rd Child) (Multi-Player Mode)
  4000128h  2    R/W  SIOCNT    SIO Control Register
  400012Ah  2    R/W  SIOMLT_SEND SIO Data (Local of MultiPlayer; shared below)
  400012Ah  2    R/W  SIODATA8  SIO Data (Normal-8bit and UART Mode)
  400012Ch       -    -         Not used

Keypad Input

  4000130h  2    R    KEYINPUT  Key Status
  4000132h  2    R/W  KEYCNT    Key Interrupt Control

Serial Communication (2)

  4000134h  2    R/W  RCNT      SIO Mode Select/General Purpose Data
  4000136h  -    -    IR        Ancient - Infrared Register (Prototypes only)
  4000138h       -    -         Not used
  4000140h  2    R/W  JOYCNT    SIO JOY Bus Control
  4000142h       -    -         Not used
  4000150h  4    R/W  JOY_RECV  SIO JOY Bus Receive Data
  4000154h  4    R/W  JOY_TRANS SIO JOY Bus Transmit Data
  4000158h  2    R/?  JOYSTAT   SIO JOY Bus Receive Status
  400015Ah       -    -         Not used

Interrupt, Waitstate, and Power-Down Control

  4000200h  2    R/W  IE        Interrupt Enable Register
  4000202h  2    R/W  IF        Interrupt Request Flags / IRQ Acknowledge
  4000204h  2    R/W  WAITCNT   Game Pak Waitstate Control
  4000206h       -    -         Not used
  4000208h  2    R/W  IME       Interrupt Master Enable Register
  400020Ah       -    -         Not used
  4000300h  1    R/W  POSTFLG   Undocumented - Post Boot Flag
  4000301h  1    W    HALTCNT   Undocumented - Power Down Control
  4000302h       -    -         Not used
  4000410h  ?    ?    ?         Undocumented - Purpose Unknown / Bug ??? 0FFh
  4000411h       -    -         Not used
  4000800h  4    R/W  ?         Undocumented - Internal Memory Control (R/W)
  4000804h       -    -         Not used
  4xx0800h  4    R/W  ?         Mirrors of 4000800h (repeated each 64K)
  4700000h  4    W    (3DS)     Disable ARM7 bootrom overlay (3DS only)

All further addresses at 4XXXXXXh are unused and do not contain mirrors of the I/O area, with the only exception that 4000800h is repeated each 64K (ie. mirrored at 4010800h, 4020800h, etc.)

GBA LCD Video Controller

Registers
LCD I/O Display Control
LCD I/O Interrupts and Status
LCD I/O BG Control
LCD I/O BG Scrolling
LCD I/O BG Rotation/Scaling
LCD I/O Window Feature
LCD I/O Mosaic Function
LCD I/O Color Special Effects

VRAM
LCD VRAM Overview
LCD VRAM Character Data
LCD VRAM BG Screen Data Format (BG Map)
LCD VRAM Bitmap BG Modes

Sprites
LCD OBJ - Overview
LCD OBJ - OAM Attributes
LCD OBJ - OAM Rotation/Scaling Parameters
LCD OBJ - VRAM Character (Tile) Mapping

Other
LCD Color Palettes
LCD Dimensions and Timings

LCD I/O Display Control

4000000h - DISPCNT - LCD Control (Read/Write)

  Bit   Expl.
  0-2   BG Mode                (0-5=Video Mode 0-5, 6-7=Prohibited)
  3     Reserved / CGB Mode    (0=GBA, 1=CGB; can be set only by BIOS opcodes)
  4     Display Frame Select   (0-1=Frame 0-1) (for BG Modes 4,5 only)
  5     H-Blank Interval Free  (1=Allow access to OAM during H-Blank)
  6     OBJ Character VRAM Mapping (0=Two dimensional, 1=One dimensional)
  7     Forced Blank           (1=Allow FAST access to VRAM,Palette,OAM)
  8     Screen Display BG0  (0=Off, 1=On)
  9     Screen Display BG1  (0=Off, 1=On)
  10    Screen Display BG2  (0=Off, 1=On)
  11    Screen Display BG3  (0=Off, 1=On)
  12    Screen Display OBJ  (0=Off, 1=On)
  13    Window 0 Display Flag   (0=Off, 1=On)
  14    Window 1 Display Flag   (0=Off, 1=On)
  15    OBJ Window Display Flag (0=Off, 1=On)

The table summarizes the facilities of the separate BG modes (video modes).

  Mode  Rot/Scal Layers Size               Tiles Colors       Features
  0     No       0123   256x256..512x515   1024  16/16..256/1 SFMABP
  1     Mixed    012-   (BG0,BG1 as above Mode 0, BG2 as below Mode 2)
  2     Yes      --23   128x128..1024x1024 256   256/1        S-MABP
  3     Yes      --2-   240x160            1     32768        --MABP
  4     Yes      --2-   240x160            2     256/1        --MABP
  5     Yes      --2-   160x128            2     32768        --MABP

Features: S)crolling, F)lip, M)osaic, A)lphaBlending, B)rightness, P)riority.

BG Modes 0-2 are Tile/Map-based. BG Modes 3-5 are Bitmap-based, in these modes 1 or 2 Frames (ie. bitmaps, or 'full screen tiles') exists, if two frames exist, either one can be displayed, and the other one can be redrawn in background.

Blanking Bits
Setting Forced Blank (Bit 7) causes the video controller to display white lines, and all VRAM, Palette RAM, and OAM may be accessed.
"When the internal HV synchronous counter cancels a forced blank during a display period, the display begins from the beginning, following the display of two vertical lines." What ?
Setting H-Blank Interval Free (Bit 5) allows to access OAM during H-Blank time - using this feature reduces the number of sprites that can be displayed per line.

Display Enable Bits
By default, BG0-3 and OBJ Display Flags (Bit 8-12) are used to enable/disable BGs and OBJ. When enabling Window 0 and/or 1 (Bit 13-14), color special effects may be used, and BG0-3 and OBJ are controlled by the window(s).

Frame Selection
In BG Modes 4 and 5 (Bitmap modes), either one of the two bitmaps/frames may be displayed (Bit 4), allowing the user to update the other (invisible) frame in background. In BG Mode 3, only one frame exists.
In BG Modes 0-2 (Tile/Map based modes), a similar effect may be gained by altering the base address(es) of BG Map and/or BG Character data.

4000002h - Undocumented - Green Swap (R/W)
Normally, red green blue intensities for a group of two pixels is output as BGRbgr (uppercase for left pixel at even xloc, lowercase for right pixel at odd xloc). When the Green Swap bit is set, each pixel group is output as BgRbGr (ie. green intensity of each two pixels exchanged).

  Bit   Expl.
  0     Green Swap  (0=Normal, 1=Swap)
  1-15  Not used

This feature appears to be applied to the final picture (ie. after mixing the separate BG and OBJ layers). Eventually intended for other display types (with other pin-outs). With normal GBA hardware it is just producing an interesting dirt effect.
The NDS DISPCNT registers are 32bit (4000000h..4000003h), so Green Swap doesn't exist in NDS mode, however, the NDS does support Green Swap in GBA mode.

LCD I/O Interrupts and Status

4000004h - DISPSTAT - General LCD Status (Read/Write)
Display status and Interrupt control. The H-Blank conditions are generated once per scanline, including for the 'hidden' scanlines during V-Blank.

  Bit   Expl.
  0     V-Blank flag   (Read only) (1=VBlank) (set in line 160..226; not 227)
  1     H-Blank flag   (Read only) (1=HBlank) (toggled in all lines, 0..227)
  2     V-Counter flag (Read only) (1=Match)  (set in selected line)     (R)
  3     V-Blank IRQ Enable         (1=Enable)                          (R/W)
  4     H-Blank IRQ Enable         (1=Enable)                          (R/W)
  5     V-Counter IRQ Enable       (1=Enable)                          (R/W)
  6     Not used (0) / DSi: LCD Initialization Ready (0=Busy, 1=Ready)   (R)
  7     Not used (0) / NDS: MSB of V-Vcount Setting (LYC.Bit8) (0..262)(R/W)
  8-15  V-Count Setting (LYC)      (0..227)                            (R/W)

The V-Count-Setting value is much the same as LYC of older gameboys, when its value is identical to the content of the VCOUNT register then the V-Counter flag is set (Bit 2), and (if enabled in Bit 5) an interrupt is requested.
Although the drawing time is only 960 cycles (240*4), the H-Blank flag is "0" for a total of 1006 cycles.

4000006h - VCOUNT - Vertical Counter (Read only)
Indicates the currently drawn scanline, values in range from 160..227 indicate 'hidden' scanlines within VBlank area.

  Bit   Expl.
  0-7   Current Scanline (LY)      (0..227)                              (R)
  8     Not used (0) / NDS: MSB of Current Scanline (LY.Bit8) (0..262)   (R)
  9-15  Not Used (0)

Note: This is much the same than the 'LY' register of older gameboys.

LCD I/O BG Control

4000008h - BG0CNT - BG0 Control (R/W) (BG Modes 0,1 only)
400000Ah - BG1CNT - BG1 Control (R/W) (BG Modes 0,1 only)
400000Ch - BG2CNT - BG2 Control (R/W) (BG Modes 0,1,2 only)
400000Eh - BG3CNT - BG3 Control (R/W) (BG Modes 0,2 only)

  Bit   Expl.
  0-1   BG Priority           (0-3, 0=Highest)
  2-3   Character Base Block  (0-3, in units of 16 KBytes) (=BG Tile Data)
  4-5   Not used (must be zero) (except in NDS mode: MSBs of char base)
  6     Mosaic                (0=Disable, 1=Enable)
  7     Colors/Palettes       (0=16/16, 1=256/1)
  8-12  Screen Base Block     (0-31, in units of 2 KBytes) (=BG Map Data)
  13    BG0/BG1: Not used (except in NDS mode: Ext Palette Slot for BG0/BG1)
  13    BG2/BG3: Display Area Overflow (0=Transparent, 1=Wraparound)
  14-15 Screen Size (0-3)

Internal Screen Size (dots) and size of BG Map (bytes):

  Value  Text Mode      Rotation/Scaling Mode
  0      256x256 (2K)   128x128   (256 bytes)
  1      512x256 (4K)   256x256   (1K)
  2      256x512 (4K)   512x512   (4K)
  3      512x512 (8K)   1024x1024 (16K)

In case that some or all BGs are set to same priority then BG0 is having the highest, and BG3 the lowest priority.

In 'Text Modes', the screen size is organized as follows: The screen consists of one or more 256x256 pixel (32x32 tiles) areas. When Size=0: only 1 area (SC0), when Size=1 or Size=2: two areas (SC0,SC1 either horizontally or vertically arranged next to each other), when Size=3: four areas (SC0,SC1 in upper row, SC2,SC3 in lower row). Whereas SC0 is defined by the normal BG Map base address (Bit 8-12 of BGxCNT), SC1 uses same address +2K, SC2 address +4K, SC3 address +6K. When the screen is scrolled it'll always wraparound.

In 'Rotation/Scaling Modes', the screen size is organized as follows, only one area (SC0) of variable size 128x128..1024x1024 pixels (16x16..128x128 tiles) exists. When the screen is rotated/scaled (or scrolled?) so that the LCD viewport reaches outside of the background/screen area, then BG may be either displayed as transparent or wraparound (Bit 13 of BGxCNT).

LCD I/O BG Scrolling

4000010h - BG0HOFS - BG0 X-Offset (W)
4000012h - BG0VOFS - BG0 Y-Offset (W)

  Bit   Expl.
  0-8   Offset (0-511)
  9-15  Not used

Specifies the coordinate of the upperleft first visible dot of BG0 background layer, ie. used to scroll the BG0 area.

4000014h - BG1HOFS - BG1 X-Offset (W)
4000016h - BG1VOFS - BG1 Y-Offset (W)
Same as above BG0HOFS and BG0VOFS for BG1 respectively.

4000018h - BG2HOFS - BG2 X-Offset (W)
400001Ah - BG2VOFS - BG2 Y-Offset (W)
Same as above BG0HOFS and BG0VOFS for BG2 respectively.

400001Ch - BG3HOFS - BG3 X-Offset (W)
400001Eh - BG3VOFS - BG3 Y-Offset (W)
Same as above BG0HOFS and BG0VOFS for BG3 respectively.

The above BG scrolling registers are exclusively used in Text modes, ie. for all layers in BG Mode 0, and for the first two layers in BG mode 1.
In other BG modes (Rotation/Scaling and Bitmap modes) above registers are ignored. Instead, the screen may be scrolled by modifying the BG Rotation/Scaling Reference Point registers.

LCD I/O BG Rotation/Scaling

4000028h - BG2X_L - BG2 Reference Point X-Coordinate, lower 16 bit (W)
400002Ah - BG2X_H - BG2 Reference Point X-Coordinate, upper 12 bit (W)
400002Ch - BG2Y_L - BG2 Reference Point Y-Coordinate, lower 16 bit (W)
400002Eh - BG2Y_H - BG2 Reference Point Y-Coordinate, upper 12 bit (W)
These registers are replacing the BG scrolling registers which are used for Text mode, ie. the X/Y coordinates specify the source position from inside of the BG Map/Bitmap of the pixel to be displayed at upper left of the GBA display. The normal BG scrolling registers are ignored in Rotation/Scaling and Bitmap modes.

  Bit   Expl.
  0-7   Fractional portion (8 bits)
  8-26  Integer portion    (19 bits)
  27    Sign               (1 bit)
  28-31 Not used

Because values are shifted left by eight, fractional portions may be specified in steps of 1/256 pixels (this would be relevant only if the screen is actually rotated or scaled). Normal signed 32bit values may be written to above registers (the most significant bits will be ignored and the value will be cut-down to 28bits, but this is no actual problem because signed values have set all MSBs to the same value).

Internal Reference Point Registers
The above reference points are automatically copied to internal registers during each vblank, specifying the origin for the first scanline. The internal registers are then incremented by dmx and dmy after each scanline.
Caution: Writing to a reference point register by software outside of the Vblank period does immediately copy the new value to the corresponding internal register, that means: in the current frame, the new value specifies the origin of the <current> scanline (instead of the topmost scanline).

4000020h - BG2PA - BG2 Rotation/Scaling Parameter A (alias dx) (W)
4000022h - BG2PB - BG2 Rotation/Scaling Parameter B (alias dmx) (W)
4000024h - BG2PC - BG2 Rotation/Scaling Parameter C (alias dy) (W)
4000026h - BG2PD - BG2 Rotation/Scaling Parameter D (alias dmy) (W)

  Bit   Expl.
  0-7   Fractional portion (8 bits)
  8-14  Integer portion    (7 bits)
  15    Sign               (1 bit)

See below for details.

400003Xh - BG3X_L/H, BG3Y_L/H, BG3PA-D - BG3 Rotation/Scaling Parameters
Same as above BG2 Reference Point, and Rotation/Scaling Parameters, for BG3 respectively.

dx (PA) and dy (PC)
When transforming a horizontal line, dx and dy specify the resulting gradient and magnification for that line. For example:
Horizontal line, length=100, dx=1, and dy=1. The resulting line would be drawn at 45 degrees, f(y)=1/1*x. Note that this would involve that line is magnified, the new length is SQR(100^2+100^2)=141.42. Yup, exactly - that's the old a^2 + b^2 = c^2 formula.

dmx (PB) and dmy (PD)
These values define the resulting gradient and magnification for transformation of vertical lines. However, when rotating a square area (which is surrounded by horizontal and vertical lines), then the desired result should be usually a rotated <square> area (ie. not a parallelogram, for example).
Thus, dmx and dmy must be defined in direct relationship to dx and dy, taking the example above, we'd have to set dmx=-1, and dmy=1, f(x)=-1/1*y.

Area Overflow
In result of rotation/scaling it may often happen that areas outside of the actual BG area become moved into the LCD viewport. Depending of the Area Overflow bit (BG2CNT and BG3CNT, Bit 13) these areas may be either displayed (by wrapping the BG area), or may be displayed transparent.
This works only in BG modes 1 and 2. The area overflow is ignored in Bitmap modes (BG modes 3-5), the outside of the Bitmaps is always transparent.

--- more details and confusing or helpful formulas ---

The following parameters are required for Rotation/Scaling

  Rotation Center X and Y Coordinates (x0,y0)
  Rotation Angle                      (alpha)
  Magnification X and Y Values        (xMag,yMag)

The display is rotated by 'alpha' degrees around the center.
The displayed picture is magnified by 'xMag' along x-Axis (Y=y0) and 'yMag' along y-Axis (X=x0).

Calculating Rotation/Scaling Parameters A-D

  A = Cos (alpha) / xMag    ;distance moved in direction x, same line
  B = Sin (alpha) / xMag    ;distance moved in direction x, next line
  C = Sin (alpha) / yMag    ;distance moved in direction y, same line
  D = Cos (alpha) / yMag    ;distance moved in direction y, next line

Calculating the position of a rotated/scaled dot
Using the following expressions,

  x0,y0    Rotation Center
  x1,y1    Old Position of a pixel (before rotation/scaling)
  x2,y2    New position of above pixel (after rotation scaling)
  A,B,C,D  BG2PA-BG2PD Parameters (as calculated above)

the following formula can be used to calculate x2,y2:

  x2 = A(x1-x0) + B(y1-y0) + x0
  y2 = C(x1-x0) + D(y1-y0) + y0

LCD I/O Window Feature

The Window Feature may be used to split the screen into four regions. The BG0-3,OBJ layers and Color Special Effects can be separately enabled or disabled in each of these regions.

The DISPCNT Register
DISPCNT Bits 13-15 are used to enable Window 0, Window 1, and/or OBJ Window regions, if any of these regions is enabled then the "Outside of Windows" region is automatically enabled, too.
DISPCNT Bits 8-12 are kept used as master enable bits for the BG0-3,OBJ layers, a layer is displayed only if both DISPCNT and WININ/OUT enable bits are set.

4000040h - WIN0H - Window 0 Horizontal Dimensions (W)
4000042h - WIN1H - Window 1 Horizontal Dimensions (W)

  Bit   Expl.
  0-7   X2, Rightmost coordinate of window, plus 1
  8-15  X1, Leftmost coordinate of window

Garbage values of X2>240 or X1>X2 are interpreted as X2=240.

4000044h - WIN0V - Window 0 Vertical Dimensions (W)
4000046h - WIN1V - Window 1 Vertical Dimensions (W)

  Bit   Expl.
  0-7   Y2, Bottom-most coordinate of window, plus 1
  8-15  Y1, Top-most coordinate of window

Garbage values of Y2>160 or Y1>Y2 are interpreted as Y2=160.

4000048h - WININ - Control of Inside of Window(s) (R/W)

  Bit   Expl.
  0-3   Window 0 BG0-BG3 Enable Bits     (0=No Display, 1=Display)
  4     Window 0 OBJ Enable Bit          (0=No Display, 1=Display)
  5     Window 0 Color Special Effect    (0=Disable, 1=Enable)
  6-7   Not used
  8-11  Window 1 BG0-BG3 Enable Bits     (0=No Display, 1=Display)
  12    Window 1 OBJ Enable Bit          (0=No Display, 1=Display)
  13    Window 1 Color Special Effect    (0=Disable, 1=Enable)
  14-15 Not used

400004Ah - WINOUT - Control of Outside of Windows & Inside of OBJ Window (R/W)

  Bit   Expl.
  0-3   Outside BG0-BG3 Enable Bits      (0=No Display, 1=Display)
  4     Outside OBJ Enable Bit           (0=No Display, 1=Display)
  5     Outside Color Special Effect     (0=Disable, 1=Enable)
  6-7   Not used
  8-11  OBJ Window BG0-BG3 Enable Bits   (0=No Display, 1=Display)
  12    OBJ Window OBJ Enable Bit        (0=No Display, 1=Display)
  13    OBJ Window Color Special Effect  (0=Disable, 1=Enable)
  14-15 Not used

The OBJ Window
The dimension of the OBJ Window is specified by OBJs which are having the "OBJ Mode" attribute being set to "OBJ Window". Any non-transparent dots of any such OBJs are marked as OBJ Window area. The OBJ itself is not displayed.
The color, palette, and display priority of these OBJs are ignored. Both DISPCNT Bits 12 and 15 must be set when defining OBJ Window region(s).

Window Priority
In case that more than one window is enabled, and that these windows do overlap, Window 0 is having highest priority, Window 1 medium, and Obj Window lowest priority. Outside of Window is having zero priority, it is used for all dots which are not inside of any window region.

LCD I/O Mosaic Function

400004Ch - MOSAIC - Mosaic Size (W)
The Mosaic function can be separately enabled/disabled for BG0-BG3 by BG0CNT-BG3CNT Registers, as well as for each OBJ0-127 by OBJ attributes in OAM memory. Also, setting all of the bits below to zero effectively disables the mosaic function.

  Bit   Expl.
  0-3   BG Mosaic H-Size  (minus 1)
  4-7   BG Mosaic V-Size  (minus 1)
  8-11  OBJ Mosaic H-Size (minus 1)
  12-15 OBJ Mosaic V-Size (minus 1)
  16-31 Not used

Example: When setting H-Size to 5, then pixels 0-5 of each display row are colorized as pixel 0, pixels 6-11 as pixel 6, pixels 12-17 as pixel 12, and so on.

Normally, a 'mosaic-pixel' is colorized by the color of the upperleft covered pixel. In many cases it might be more desireful to use the color of the pixel in the center of the covered area - this effect may be gained by scrolling the background (or by adjusting the OBJ position, as far as upper/left rows/columns of OBJ are transparent).

LCD I/O Color Special Effects

Two types of Special Effects are supported: Alpha Blending (Semi-Transparency) allows to combine colors of two selected surfaces. Brightness Increase/Decrease adjust the brightness of the selected surface.

4000050h - BLDCNT - Color Special Effects Selection (R/W)

  Bit   Expl.
  0     BG0 1st Target Pixel (Background 0)
  1     BG1 1st Target Pixel (Background 1)
  2     BG2 1st Target Pixel (Background 2)
  3     BG3 1st Target Pixel (Background 3)
  4     OBJ 1st Target Pixel (Top-most OBJ pixel)
  5     BD  1st Target Pixel (Backdrop)
  6-7   Color Special Effect (0-3, see below)
         0 = None                (Special effects disabled)
         1 = Alpha Blending      (1st+2nd Target mixed)
         2 = Brightness Increase (1st Target becomes whiter)
         3 = Brightness Decrease (1st Target becomes blacker)
  8     BG0 2nd Target Pixel (Background 0)
  9     BG1 2nd Target Pixel (Background 1)
  10    BG2 2nd Target Pixel (Background 2)
  11    BG3 2nd Target Pixel (Background 3)
  12    OBJ 2nd Target Pixel (Top-most OBJ pixel)
  13    BD  2nd Target Pixel (Backdrop)
  14-15 Not used

Selects the 1st Target layer(s) for special effects. For Alpha Blending/Semi-Transparency, it does also select the 2nd Target layer(s), which should have next lower display priority as the 1st Target.
However, any combinations are possible, including that all layers may be selected as both 1st+2nd target, in that case the top-most pixel will be used as 1st target, and the next lower pixel as 2nd target.

4000052h - BLDALPHA - Alpha Blending Coefficients (R/W) (not W)
Used for Color Special Effects Mode 1, and for Semi-Transparent OBJs.

  Bit   Expl.
  0-4   EVA Coefficient (1st Target) (0..16 = 0/16..16/16, 17..31=16/16)
  5-7   Not used
  8-12  EVB Coefficient (2nd Target) (0..16 = 0/16..16/16, 17..31=16/16)
  13-15 Not used

For this effect, the top-most non-transparent pixel must be selected as 1st Target, and the next-lower non-transparent pixel must be selected as 2nd Target, if so - and only if so, then color intensities of 1st and 2nd Target are mixed together by using the parameters in BLDALPHA register, for each pixel each R, G, B intensities are calculated separately:

  I = MIN ( 31, I1st*EVA + I2nd*EVB )

Otherwise - for example, if only one target exists, or if a non-transparent non-2nd-target pixel is moved between the two targets, or if 2nd target has higher display priority than 1st target - then only the top-most pixel is displayed (at normal intensity, regardless of BLDALPHA).

4000054h - BLDY - Brightness (Fade-In/Out) Coefficient (W) (not R/W)
Used for Color Special Effects Modes 2 and 3.

  Bit   Expl.
  0-4   EVY Coefficient (Brightness) (0..16 = 0/16..16/16, 17..31=16/16)
  5-31  Not used

For each pixel each R, G, B intensities are calculated separately:

  I = I1st + (31-I1st)*EVY   ;For Brightness Increase
  I = I1st - (I1st)*EVY      ;For Brightness Decrease

The color intensities of any selected 1st target surface(s) are increased or decreased by using the parameter in BLDY register.

Semi-Transparent OBJs
OBJs that are defined as 'Semi-Transparent' in OAM memory are always selected as 1st Target (regardless of BLDCNT Bit 4), and are always using Alpha Blending mode (regardless of BLDCNT Bit 6-7).
The BLDCNT register may be used to perform Brightness effects on the OBJ (and/or other BG/BD layers). However, if a semi-transparent OBJ pixel does overlap a 2nd target pixel, then semi-transparency becomes priority, and the brightness effect will not take place (neither on 1st, nor 2nd target).

The OBJ Layer
Before special effects are applied, the display controller computes the OBJ priority ordering, and isolates the top-most OBJ pixel. In result, only the top-most OBJ pixel is recursed at the time when processing special effects. Ie. alpha blending and semi-transparency can be used for OBJ-to-BG or BG-to-OBJ , but not for OBJ-to-OBJ.

LCD VRAM Overview

The GBA contains 96 Kbytes VRAM built-in, located at address 06000000-06017FFF, depending on the BG Mode used as follows:

BG Mode 0,1,2 (Tile/Map based Modes)

  06000000-0600FFFF  64 KBytes shared for BG Map and Tiles
  06010000-06017FFF  32 KBytes OBJ Tiles

The shared 64K area can be split into BG Map area(s), and BG Tiles area(s), the respective addresses for Map and Tile areas are set up by BG0CNT-BG3CNT registers. The Map address may be specified in units of 2K (steps of 800h), the Tile address in units of 16K (steps of 4000h).

BG Mode 0,1 (Tile/Map based Text mode)
The tiles may have 4bit or 8bit color depth, minimum map size is 32x32 tiles, maximum is 64x64 tiles, up to 1024 tiles can be used per map.

  Item        Depth     Required Memory
  One Tile    4bit      20h bytes
  One Tile    8bit      40h bytes
  1024 Tiles  4bit      8000h (32K)
  1024 Tiles  8bit      10000h (64K) - excluding some bytes for BG map
  BG Map      32x32     800h (2K)
  BG Map      64x64     2000h (8K)

BG Mode 1,2 (Tile/Map based Rotation/Scaling mode)
The tiles may have 8bit color depth only, minimum map size is 16x16 tiles, maximum is 128x128 tiles, up to 256 tiles can be used per map.

  Item        Depth     Required Memory
  One Tile    8bit      40h bytes
  256  Tiles  8bit      4000h (16K)
  BG Map      16x16     100h bytes
  BG Map      128x128   4000h (16K)

BG Mode 3 (Bitmap based Mode for still images)

  06000000-06013FFF  80 KBytes Frame 0 buffer (only 75K actually used)
  06014000-06017FFF  16 KBytes OBJ Tiles

BG Mode 4,5 (Bitmap based Modes)

  06000000-06009FFF  40 KBytes Frame 0 buffer (only 37.5K used in Mode 4)
  0600A000-06013FFF  40 KBytes Frame 1 buffer (only 37.5K used in Mode 4)
  06014000-06017FFF  16 KBytes OBJ Tiles

Note
Additionally to the above VRAM, the GBA also contains 1 KByte Palette RAM (at 05000000h) and 1 KByte OAM (at 07000000h) which are both used by the display controller as well.

LCD VRAM Character Data

Each character (tile) consists of 8x8 dots (64 dots in total). The color depth may be either 4bit or 8bit (see BG0CNT-BG3CNT).

4bit depth (16 colors, 16 palettes)
Each tile occupies 32 bytes of memory, the first 4 bytes for the topmost row of the tile, and so on. Each byte representing two dots, the lower 4 bits define the color for the left (!) dot, the upper 4 bits the color for the right dot.

8bit depth (256 colors, 1 palette)
Each tile occupies 64 bytes of memory, the first 8 bytes for the topmost row of the tile, and so on. Each byte selects the palette entry for each dot.

LCD VRAM BG Screen Data Format (BG Map)

The display background consists of 8x8 dot tiles, the arrangement of these tiles is specified by the BG Screen Data (BG Map). The separate entries in this map are as follows:

Text BG Screen (2 bytes per entry)
Specifies the tile number and attributes. Note that BG tile numbers are always specified in steps of 1 (unlike OBJ tile numbers which are using steps of two in 256 color/1 palette mode).

  Bit   Expl.
  0-9   Tile Number     (0-1023) (a bit less in 256 color mode, because
                           there'd be otherwise no room for the bg map)
  10    Horizontal Flip (0=Normal, 1=Mirrored)
  11    Vertical Flip   (0=Normal, 1=Mirrored)
  12-15 Palette Number  (0-15)    (Not used in 256 color/1 palette mode)

A Text BG Map always consists of 32x32 entries (256x256 pixels), 400h entries = 800h bytes. However, depending on the BG Size, one, two, or four of these Maps may be used together, allowing to create backgrounds of 256x256, 512x256, 256x512, or 512x512 pixels, if so, the first map (SC0) is located at base+0, the next map (SC1) at base+800h, and so on.

Rotation/Scaling BG Screen (1 byte per entry)
In this mode, only 256 tiles can be used. There are no x/y-flip attributes, the color depth is always 256 colors/1 palette.

  Bit   Expl.
  0-7   Tile Number     (0-255)

The dimensions of Rotation/Scaling BG Maps depend on the BG size. For size 0-3 that are: 16x16 tiles (128x128 pixels), 32x32 tiles (256x256 pixels), 64x64 tiles (512x512 pixels), or 128x128 tiles (1024x1024 pixels).

The size and VRAM base address of the separate BG maps for BG0-3 are set up by BG0CNT-BG3CNT registers.

LCD VRAM Bitmap BG Modes

In BG Modes 3-5 the background is defined in form of a bitmap (unlike as for Tile/Map based BG modes). Bitmaps are implemented as BG2, with Rotation/Scaling support. As bitmap modes are occupying 80KBytes of BG memory, only 16KBytes of VRAM can be used for OBJ tiles.

BG Mode 3 - 240x160 pixels, 32768 colors
Two bytes are associated to each pixel, directly defining one of the 32768 colors (without using palette data, and thus not supporting a 'transparent' BG color).

  Bit   Expl.
  0-4   Red Intensity   (0-31)
  5-9   Green Intensity (0-31)
  10-14 Blue Intensity  (0-31)
  15    Not used in GBA Mode (in NDS Mode: Alpha=0=Transparent, Alpha=1=Normal)

The first 480 bytes define the topmost line, the next 480 the next line, and so on. The background occupies 75 KBytes (06000000-06012BFF), most of the 80 Kbytes BG area, not allowing to redraw an invisible second frame in background, so this mode is mostly recommended for still images only.

BG Mode 4 - 240x160 pixels, 256 colors (out of 32768 colors)
One byte is associated to each pixel, selecting one of the 256 palette entries. Color 0 (backdrop) is transparent, and OBJs may be displayed behind the bitmap.
The first 240 bytes define the topmost line, the next 240 the next line, and so on. The background occupies 37.5 KBytes, allowing two frames to be used (06000000-060095FF for Frame 0, and 0600A000-060135FF for Frame 1).

BG Mode 5 - 160x128 pixels, 32768 colors
Colors are defined as for Mode 3 (see above), but horizontal and vertical size are cut down to 160x128 pixels only - smaller than the physical dimensions of the LCD screen.
The background occupies exactly 40 KBytes, so that BG VRAM may be split into two frames (06000000-06009FFF for Frame 0, and 0600A000-06013FFF for Frame 1).

In BG modes 4,5, one Frame may be displayed (selected by DISPCNT Bit 4), the other Frame is invisible and may be redrawn in background.

LCD OBJ - Overview

General
Objects (OBJs) are moveable sprites. Up to 128 OBJs (of any size, up to 64x64 dots each) can be displayed per screen, and under best circumstances up to 128 OBJs (of small 8x8 dots size) can be displayed per horizontal display line.

Maximum Number of Sprites per Line
The total available OBJ rendering cycles per line are

  1210  (=304*4-6)   If "H-Blank Interval Free" bit in DISPCNT register is 0
  954   (=240*4-6)   If "H-Blank Interval Free" bit in DISPCNT register is 1

The required rendering cycles are (depending on horizontal OBJ size)

  Cycles per <n> Pixels    OBJ Type              OBJ Type Screen Pixel Range
  n*1 cycles               Normal OBJs           8..64 pixels
  10+n*2 cycles            Rotation/Scaling OBJs 8..64 pixels   (area clipped)
  10+n*2 cycles            Rotation/Scaling OBJs 16..128 pixels (double size)

Caution:
The maximum number of OBJs per line is also affected by undisplayed (offscreen) OBJs which are having higher priority than displayed OBJs.
To avoid this, move displayed OBJs to the begin of OAM memory (ie. OBJ0 has highest priority, OBJ127 lowest).
Otherwise (in case that the program logic expects OBJs at fixed positions in OAM) at least take care to set the OBJ size of undisplayed OBJs to 8x8 with Rotation/Scaling disabled (this reduces the overload).
Does the above also apply for VERTICALLY OFFSCREEN (or VERTICALLY not on CURRENT LINE) sprites ?

VRAM - Character Data
OBJs are always combined of one or more 8x8 pixel Tiles (much like BG Tiles in BG Modes 0-2). However, OBJ Tiles are stored in a separate area in VRAM: 06010000-06017FFF (32 KBytes) in BG Mode 0-2, or 06014000-06017FFF (16 KBytes) in BG Mode 3-5.
Depending on the size of the above area (16K or 32K), and on the OBJ color depth (4bit or 8bit), 256-1024 8x8 dots OBJ Tiles can be defined.

OAM - Object Attribute Memory
This memory area contains Attributes which specify position, size, color depth, etc. appearance for each of the 128 OBJs. Additionally, it contains 32 OBJ Rotation/Scaling Parameter groups. OAM is located at 07000000-070003FF (sized 1 KByte).

LCD OBJ - OAM Attributes

OBJ Attributes
There are 128 entries in OAM for each OBJ0-OBJ127. Each entry consists of 6 bytes (three 16bit Attributes). Attributes for OBJ0 are located at 07000000, for OBJ1 at 07000008, OBJ2 at 07000010, and so on.

As you can see, there are blank spaces at 07000006, 0700000E, 07000016, etc. - these 16bit values are used for OBJ Rotation/Scaling (as described in the next chapter) - they are not directly related to the separate OBJs.

OBJ Attribute 0 (R/W)

  Bit   Expl.
  0-7   Y-Coordinate           (0-255)
  8     Rotation/Scaling Flag  (0=Off, 1=On)
  When Rotation/Scaling used (Attribute 0, bit 8 set):
    9     Double-Size Flag     (0=Normal, 1=Double)
  When Rotation/Scaling not used (Attribute 0, bit 8 cleared):
    9     OBJ Disable          (0=Normal, 1=Not displayed)
  10-11 OBJ Mode  (0=Normal, 1=Semi-Transparent, 2=OBJ Window, 3=Prohibited)
  12    OBJ Mosaic             (0=Off, 1=On)
  13    Colors/Palettes        (0=16/16, 1=256/1)
  14-15 OBJ Shape              (0=Square,1=Horizontal,2=Vertical,3=Prohibited)

Caution: A very large OBJ (of 128 pixels vertically, ie. a 64 pixels OBJ in a Double Size area) located at Y>128 will be treated as at Y>-128, the OBJ is then displayed parts offscreen at the TOP of the display, it is then NOT displayed at the bottom.

OBJ Attribute 1 (R/W)

  Bit   Expl.
  0-8   X-Coordinate           (0-511)
  When Rotation/Scaling used (Attribute 0, bit 8 set):
    9-13  Rotation/Scaling Parameter Selection (0-31)
          (Selects one of the 32 Rotation/Scaling Parameters that
          can be defined in OAM, for details read next chapter.)
  When Rotation/Scaling not used (Attribute 0, bit 8 cleared):
    9-11  Not used
    12    Horizontal Flip      (0=Normal, 1=Mirrored)
    13    Vertical Flip        (0=Normal, 1=Mirrored)
  14-15 OBJ Size               (0..3, depends on OBJ Shape, see Attr 0)
          Size  Square   Horizontal  Vertical
          0     8x8      16x8        8x16
          1     16x16    32x8        8x32
          2     32x32    32x16       16x32
          3     64x64    64x32       32x64

OBJ Attribute 2 (R/W)

  Bit   Expl.
  0-9   Character Name          (0-1023=Tile Number)
  10-11 Priority relative to BG (0-3; 0=Highest)
  12-15 Palette Number   (0-15) (Not used in 256 color/1 palette mode)

Notes:

OBJ Mode
The OBJ Mode may be Normal, Semi-Transparent, or OBJ Window.
Semi-Transparent means that the OBJ is used as 'Alpha Blending 1st Target' (regardless of BLDCNT register, for details see chapter about Color Special Effects).
OBJ Window means that the OBJ is not displayed, instead, dots with non-zero color are used as mask for the OBJ Window, see DISPCNT and WINOUT for details.

OBJ Tile Number
There are two situations which may divide the amount of available tiles by two (by four if both situations apply):

1. When using the 256 Colors/1 Palette mode, only each second tile may be used, the lower bit of the tile number should be zero (in 2-dimensional mapping mode, the bit is completely ignored).

2. When using BG Mode 3-5 (Bitmap Modes), only tile numbers 512-1023 may be used. That is because lower 16K of OBJ memory are used for BG. Attempts to use tiles 0-511 are ignored (not displayed).

Priority
In case that the 'Priority relative to BG' is the same than the priority of one of the background layers, then the OBJ becomes higher priority and is displayed on top of that BG layer.
Caution: Take care not to mess up BG Priority and OBJ priority. For example, the following would cause garbage to be displayed:

  OBJ No. 0 with Priority relative to BG=1   ;hi OBJ prio, lo BG prio
  OBJ No. 1 with Priority relative to BG=0   ;lo OBJ prio, hi BG prio

That is, OBJ0 is always having priority above OBJ1-127, so assigning a lower BG Priority to OBJ0 than for OBJ1-127 would be a bad idea.

LCD OBJ - OAM Rotation/Scaling Parameters

As described in the previous chapter, there are blank spaces between each of the 128 OBJ Attribute Fields in OAM memory. These 128 16bit gaps are used to store OBJ Rotation/Scaling Parameters.

Location of Rotation/Scaling Parameters in OAM
Four 16bit parameters (PA,PB,PC,PD) are required to define a complete group of Rotation/Scaling data. These are spread across OAM as such:

  1st Group - PA=07000006, PB=0700000E, PC=07000016, PD=0700001E
  2nd Group - PA=07000026, PB=0700002E, PC=07000036, PD=0700003E
  etc.

By using all blank space (128 x 16bit), up to 32 of these groups (4 x 16bit each) can be defined in OAM.

OBJ Rotation/Scaling PA,PB,PC,PD Parameters (R/W)
Each OBJ that uses Rotation/Scaling may select between any of the above 32 parameter groups. For details, refer to the previous chapter about OBJ Attributes.
The meaning of the separate PA,PB,PC,PD values is identical as for BG, for details read the chapter about BG Rotation/Scaling.

OBJ Reference Point & Rotation Center
The OBJ Reference Point is the upper left of the OBJ, ie. OBJ X/Y coordinates: X+0, Y+0.
The OBJ Rotation Center is always (or should be usually?) in the middle of the object, ie. for a 8x32 pixel OBJ, this would be at the OBJ X/Y coordinates: X+4, and Y+16.

OBJ Double-Size Bit (for OBJs that use Rotation/Scaling)
When Double-Size is zero: The sprite is rotated, and then display inside of the normal-sized (not rotated) rectangular area - the edges of the rotated sprite will become invisible if they reach outside of that area.
When Double-Size is set: The sprite is rotated, and then display inside of the double-sized (not rotated) rectangular area - this ensures that the edges of the rotated sprite remain visible even if they would reach outside of the normal-sized area. (Except that, for example, rotating a 8x32 pixel sprite by 90 degrees would still cut off parts of the sprite as the double-size area isn't large enough.)

LCD OBJ - VRAM Character (Tile) Mapping

Each OBJ tile consists of 8x8 dots, however, bigger OBJs can be displayed by combining several 8x8 tiles. The horizontal and vertical size for each OBJ may be separately defined in OAM, possible H/V sizes are 8,16,32,64 dots - allowing 'square' OBJs to be used (such like 8x8, 16x16, etc) as well as 'rectangular' OBJs (such like 8x32, 64x16, etc.)

When displaying an OBJ that contains of more than one 8x8 tile, one of the following two mapping modes can be used. In either case, the tile number of the upperleft tile must be specified in OAM memory.

Two Dimensional Character Mapping (DISPCNT Bit 6 cleared)
This mapping mode assumes that the 1024 OBJ tiles are arranged as a matrix of 32x32 tiles / 256x256 pixels (In 256 color mode: 16x32 tiles / 128x256 pixels). Ie. the upper row of this matrix contains tiles 00h-1Fh, the next row tiles 20h-3Fh, and so on.
For example, when displaying a 16x16 pixel OBJ, with tile number set to 04h; The upper row of the OBJ will consist of tile 04h and 05h, the next row of 24h and 25h. (In 256 color mode: 04h and 06h, 24h and 26h.)

One Dimensional Character Mapping (DISPCNT Bit 6 set)
In this mode, tiles are mapped each after each other from 00h-3FFh.
Using the same example as above, the upper row of the OBJ will consist of tile 04h and 05h, the next row of tile 06h and 07h. (In 256 color mode: 04h and 06h, 08h and 0Ah.)

LCD Color Palettes

Color Palette RAM
BG and OBJ palettes are using separate memory regions:

  05000000-050001FF - BG Palette RAM (512 bytes, 256 colors)
  05000200-050003FF - OBJ Palette RAM (512 bytes, 256 colors)

Each BG and OBJ palette RAM may be either split into 16 palettes with 16 colors each, or may be used as a single palette with 256 colors.
Note that some OBJs may access palette RAM in 16 color mode, while other OBJs may use 256 color mode at the same time. Same for BG0-BG3 layers.

Transparent Colors
Color 0 of all BG and OBJ palettes is transparent. Even though palettes are described as 16 (256) color palettes, only 15 (255) colors are actually visible.

Backdrop Color
Color 0 of BG Palette 0 is used as backdrop color. This color is displayed if an area of the screen is not covered by any non-transparent BG or OBJ dots.

Color Definitions
Each color occupies two bytes (same as for 32768 color BG modes):

  Bit   Expl.
  0-4   Red Intensity   (0-31)
  5-9   Green Intensity (0-31)
  10-14 Blue Intensity  (0-31)
  15    Not used

Intensities
Under normal circumstances (light source/viewing angle), the intensities 0-14 are practically all black, and only intensities 15-31 are resulting in visible medium..bright colors.

Note: The intensity problem appears in the 8bit CGB "compatibility" mode either. The original CGB display produced the opposite effect: Intensities 0-14 resulted in dark..medium colors, and intensities 15-31 resulted in bright colors. Any "medium" colors of CGB games will appear invisible/black on GBA hardware, and only very bright colors will be visible.

LCD Dimensions and Timings

Horizontal Dimensions
The drawing time for each dot is 4 CPU cycles.

  Visible     240 dots,  57.221 us,    960 cycles - 78% of h-time
  H-Blanking   68 dots,  16.212 us,    272 cycles - 22% of h-time
  Total       308 dots,  73.433 us,   1232 cycles - ca. 13.620 kHz

VRAM and Palette RAM may be accessed during H-Blanking. OAM can accessed only if "H-Blank Interval Free" bit in DISPCNT register is set.

Vertical Dimensions

  Visible (*) 160 lines, 11.749 ms, 197120 cycles - 70% of v-time
  V-Blanking   68 lines,  4.994 ms,  83776 cycles - 30% of v-time
  Total       228 lines, 16.743 ms, 280896 cycles - ca. 59.737 Hz

All VRAM, OAM, and Palette RAM may be accessed during V-Blanking.
Note that no H-Blank interrupts are generated within V-Blank period.

System Clock
The system clock is 16.78MHz (16*1024*1024 Hz), one cycle is thus approx. 59.59ns.

(*) Even though vertical screen size is 160 lines, the upper 8 lines are not <really> visible, these lines are covered by a shadow when holding the GBA orientated towards a light source, the lines are effectively black - and should not be used to display important information.

Interlace
The LCD display is using some sort of interlace in which even scanlines are dimmed in each second frame, and odd scanlines are dimmed in each other frame (it does always render ALL lines in ALL frames, but half of them are dimmed).
The effect can be seen when displaying some horizontal lines in each second frame, and hiding them in each other frame: the hardware will randomly show the lines in dimmed or non-dimmed form (depending on whether the test was started in an even or odd frame).
Unknown if it's possible to determine the even/off frame state by software (or possibly to reset the hardware to this or that state by software).
Note: The NDS is applying some sort of frameskip to GBA games, about every 3 seconds there will by a missing (or maybe: inserted) frame, ie. a GBA game that is updating the display in sync with GBA interlace will get offsync on NDS consoles.

GBA Sound Controller

The GBA supplies four 'analogue' sound channels for Tone and Noise (mostly compatible to CGB sound), as well as two 'digital' sound channels (which can be used to replay 8bit DMA sample data).

GBA Sound Channel 1 - Tone & Sweep
GBA Sound Channel 2 - Tone
GBA Sound Channel 3 - Wave Output
GBA Sound Channel 4 - Noise
GBA Sound Channel A and B - DMA Sound

GBA Sound Control Registers
GBA Comparison of CGB and GBA Sound

The GBA includes only a single (mono) speaker built-in, each channel may be output to either left and/or right channels by using the external line-out connector (for stereo headphones, etc).

GBA Sound Channel 1 - Tone & Sweep

4000060h - SOUND1CNT_L (NR10) - Channel 1 Sweep register (R/W)

  Bit        Expl.
  0-2   R/W  Number of sweep shift      (n=0-7)
  3     R/W  Sweep Frequency Direction  (0=Increase, 1=Decrease)
  4-6   R/W  Sweep Time; units of 7.8ms (0-7, min=7.8ms, max=54.7ms)
  7-15  -    Not used

Sweep is disabled by setting Sweep Time to zero, if so, the direction bit should be set.
The change of frequency (NR13,NR14) at each shift is calculated by the following formula where X(0) is initial freq & X(t-1) is last freq:

  X(t) = X(t-1) +/- X(t-1)/2^n

4000062h - SOUND1CNT_H (NR11, NR12) - Channel 1 Duty/Len/Envelope (R/W)

  Bit        Expl.
  0-5   W    Sound length; units of (64-n)/256s  (0-63)
  6-7   R/W  Wave Pattern Duty                   (0-3, see below)
  8-10  R/W  Envelope Step-Time; units of n/64s  (1-7, 0=No Envelope)
  11    R/W  Envelope Direction                  (0=Decrease, 1=Increase)
  12-15 R/W  Initial Volume of envelope          (1-15, 0=No Sound)

Wave Duty:

  0: 12.5% ( -_______-_______-_______ )
  1: 25%   ( --______--______--______ )
  2: 50%   ( ----____----____----____ ) (normal)
  3: 75%   ( ------__------__------__ )

The Length value is used only if Bit 6 in NR14 is set.

4000064h - SOUND1CNT_X (NR13, NR14) - Channel 1 Frequency/Control (R/W)

  Bit        Expl.
  0-10  W    Frequency; 131072/(2048-n)Hz  (0-2047)
  11-13 -    Not used
  14    R/W  Length Flag  (1=Stop output when length in NR11 expires)
  15    W    Initial      (1=Restart Sound)
  16-31 -    Not used

GBA Sound Channel 2 - Tone

This sound channel works exactly as channel 1, except that it doesn't have a Tone Envelope/Sweep Register.

4000068h - SOUND2CNT_L (NR21, NR22) - Channel 2 Duty/Length/Envelope (R/W)
400006Ah - Not used
400006Ch - SOUND2CNT_H (NR23, NR24) - Channel 2 Frequency/Control (R/W)
For details, refer to channel 1 description.

GBA Sound Channel 3 - Wave Output

This channel can be used to output digital sound, the length of the sample buffer (Wave RAM) can be either 32 or 64 digits (4bit samples). This sound channel can be also used to output normal tones when initializing the Wave RAM by a square wave. This channel doesn't have a volume envelope register.

4000070h - SOUND3CNT_L (NR30) - Channel 3 Stop/Wave RAM select (R/W)

  Bit        Expl.
  0-4   -    Not used
  5     R/W  Wave RAM Dimension   (0=One bank/32 digits, 1=Two banks/64 digits)
  6     R/W  Wave RAM Bank Number (0-1, see below)
  7     R/W  Sound Channel 3 Off  (0=Stop, 1=Playback)
  8-15  -    Not used

The currently selected Bank Number (Bit 6) will be played back, while reading/writing to/from wave RAM will address the other (not selected) bank. When dimension is set to two banks, output will start by replaying the currently selected bank.

4000072h - SOUND3CNT_H (NR31, NR32) - Channel 3 Length/Volume (R/W)

  Bit        Expl.
  0-7   W    Sound length; units of (256-n)/256s  (0-255)
  8-12  -    Not used.
  13-14 R/W  Sound Volume  (0=Mute/Zero, 1=100%, 2=50%, 3=25%)
  15    R/W  Force Volume  (0=Use above, 1=Force 75% regardless of above)

The Length value is used only if Bit 6 in NR34 is set.

4000074h - SOUND3CNT_X (NR33, NR34) - Channel 3 Frequency/Control (R/W)

  Bit        Expl.
  0-10  W    Sample Rate; 2097152/(2048-n) Hz   (0-2047)
  11-13 -    Not used
  14    R/W  Length Flag  (1=Stop output when length in NR31 expires)
  15    W    Initial      (1=Restart Sound)
  16-31 -    Not used

The above sample rate specifies the number of wave RAM digits per second, the actual tone frequency depends on the wave RAM content, for example:

  Wave RAM, single bank 32 digits   Tone Frequency
  FFFFFFFFFFFFFFFF0000000000000000  65536/(2048-n) Hz
  FFFFFFFF00000000FFFFFFFF00000000  131072/(2048-n) Hz
  FFFF0000FFFF0000FFFF0000FFFF0000  262144/(2048-n) Hz
  FF00FF00FF00FF00FF00FF00FF00FF00  524288/(2048-n) Hz
  F0F0F0F0F0F0F0F0F0F0F0F0F0F0F0F0  1048576/(2048-n) Hz

4000090h - WAVE_RAM0_L - Channel 3 Wave Pattern RAM (W/R)
4000092h - WAVE_RAM0_H - Channel 3 Wave Pattern RAM (W/R)
4000094h - WAVE_RAM1_L - Channel 3 Wave Pattern RAM (W/R)
4000096h - WAVE_RAM1_H - Channel 3 Wave Pattern RAM (W/R)
4000098h - WAVE_RAM2_L - Channel 3 Wave Pattern RAM (W/R)
400009Ah - WAVE_RAM2_H - Channel 3 Wave Pattern RAM (W/R)
400009Ch - WAVE_RAM3_L - Channel 3 Wave Pattern RAM (W/R)
400009Eh - WAVE_RAM3_H - Channel 3 Wave Pattern RAM (W/R)
This area contains 16 bytes (32 x 4bits) Wave Pattern data which is output by channel 3. Data is played back ordered as follows: MSBs of 1st byte, followed by LSBs of 1st byte, followed by MSBs of 2nd byte, and so on - this results in a confusing ordering when filling Wave RAM in units of 16bit data - ie. samples would be then located in Bits 4-7, 0-3, 12-15, 8-11.

In the GBA, two Wave Patterns exists (each 32 x 4bits), either one may be played (as selected in NR30 register), the other bank may be accessed by the users. After all 32 samples have been played, output of the same bank (or other bank, as specified in NR30) will be automatically restarted.

Internally, Wave RAM is a giant shift-register, there is no pointer which is addressing the currently played digit. Instead, the entire 128 bits are shifted, and the 4 least significant bits are output.
Thus, when reading from Wave RAM, data might have changed its position. And, when writing to Wave RAM all data should be updated (it'd be no good idea to assume that old data is still located at the same position where it has been written to previously).

GBA Sound Channel 4 - Noise

This channel is used to output white noise. This is done by randomly switching the amplitude between high and low at a given frequency. Depending on the frequency the noise will appear 'harder' or 'softer'.

It is also possible to influence the function of the random generator, so the that the output becomes more regular, resulting in a limited ability to output Tone instead of Noise.

4000078h - SOUND4CNT_L (NR41, NR42) - Channel 4 Length/Envelope (R/W)

  Bit        Expl.
  0-5   W    Sound length; units of (64-n)/256s  (0-63)
  6-7   -    Not used
  8-10  R/W  Envelope Step-Time; units of n/64s  (1-7, 0=No Envelope)
  11    R/W  Envelope Direction                  (0=Decrease, 1=Increase)
  12-15 R/W  Initial Volume of envelope          (1-15, 0=No Sound)
  16-31 -    Not used

The Length value is used only if Bit 6 in NR44 is set.

400007Ch - SOUND4CNT_H (NR43, NR44) - Channel 4 Frequency/Control (R/W)
The amplitude is randomly switched between high and low at the given frequency. A higher frequency will make the noise to appear 'softer'.
When Bit 3 is set, the output will become more regular, and some frequencies will sound more like Tone than Noise.

  Bit        Expl.
  0-2   R/W  Dividing Ratio of Frequencies (r)
  3     R/W  Counter Step/Width (0=15 bits, 1=7 bits)
  4-7   R/W  Shift Clock Frequency (s)
  8-13  -    Not used
  14    R/W  Length Flag  (1=Stop output when length in NR41 expires)
  15    W    Initial      (1=Restart Sound)
  16-31 -    Not used

Frequency = 524288 Hz / r / 2^(s+1) ;For r=0 assume r=0.5 instead

Noise Random Generator (aka Polynomial Counter)
Noise randomly switches between HIGH and LOW levels, the output levels are calculated by a shift register (X), at the selected frequency, as such:

  7bit:  X=X SHR 1, IF carry THEN Out=HIGH, X=X XOR 60h ELSE Out=LOW
  15bit: X=X SHR 1, IF carry THEN Out=HIGH, X=X XOR 6000h ELSE Out=LOW

The initial value when (re-)starting the sound is X=40h (7bit) or X=4000h (15bit). The data stream repeats after 7Fh (7bit) or 7FFFh (15bit) steps.

GBA Sound Channel A and B - DMA Sound

The GBA contains two DMA sound channels (A and B), each allowing to replay digital sound (signed 8bit data, ie. -128..+127). Data can be transferred from INTERNAL memory (not sure if EXTERNAL memory works also ?) to FIFO by using DMA channel 1 or 2, the sample rate is generated by using one of the Timers.

40000A0h - FIFO_A_L - Sound A FIFO, Data 0 and Data 1 (W)
40000A2h - FIFO_A_H - Sound A FIFO, Data 2 and Data 3 (W)
These two registers may receive 32bit (4 bytes) of audio data (Data 0-3, Data 0 being located in least significant byte which is replayed first).
Internally, the capacity of the FIFO is 8 x 32bit (32 bytes), allowing to buffer a small amount of samples. As the name says (First In First Out), oldest data is replayed first.

40000A4h - FIFO_B_L - Sound B FIFO, Data 0 and Data 1 (W)
40000A6h - FIFO_B_H - Sound B FIFO, Data 2 and Data 3 (W)
Same as above, for Sound B.

Initializing DMA-Sound Playback
- Select Timer 0 or 1 in SOUNDCNT_H control register.
- Clear the FIFO.
- Manually write a sample byte to the FIFO.
- Initialize transfer mode for DMA 1 or 2.
- Initialize DMA Sound settings in sound control register.
- Start the timer.

DMA-Sound Playback Procedure
The pseudo-procedure below is automatically repeated.

  If Timer overflows then
    Move 8bit data from FIFO to sound circuit.
    If FIFO contains only 4 x 32bits (16 bytes) then
      Request more data per DMA
      Receive 4 x 32bit (16 bytes) per DMA
    Endif
  Endif

This playback mechanism will be repeated forever, regardless of the actual length of the sample buffer.

Synchronizing Sample Buffers
The buffer-end may be determined by counting sound Timer IRQs (each sample byte), or sound DMA IRQs (each 16th sample byte). Both methods would require a lot of CPU time (IRQ processing), and both would fail if interrupts are disabled for a longer period.
Better solutions would be to synchronize the sample rate/buffer length with V-blanks, or to use a second timer (in count up/slave mode) which produces an IRQ after the desired number of samples.

The Sample Rate
The GBA hardware does internally re-sample all sound output to 32.768kHz (default SOUNDBIAS setting). It'd thus do not make much sense to use higher DMA/Timer rates. Best re-sampling accuracy can be gained by using DMA/Timer rates of 32.768kHz, 16.384kHz, or 8.192kHz (ie. fragments of the physical output rate).

GBA Sound Control Registers

4000080h - SOUNDCNT_L (NR50, NR51) - Channel L/R Volume/Enable (R/W)

  Bit        Expl.
  0-2   R/W  Sound 1-4 Master Volume RIGHT (0-7)
  3     -    Not used
  4-6   R/W  Sound 1-4 Master Volume LEFT (0-7)
  7     -    Not used
  8-11  R/W  Sound 1-4 Enable Flags RIGHT (each Bit 8-11, 0=Disable, 1=Enable)
  12-15 R/W  Sound 1-4 Enable Flags LEFT (each Bit 12-15, 0=Disable, 1=Enable)

4000082h - SOUNDCNT_H (GBA only) - DMA Sound Control/Mixing (R/W)

  Bit        Expl.
  0-1   R/W  Sound # 1-4 Volume   (0=25%, 1=50%, 2=100%, 3=Prohibited)
  2     R/W  DMA Sound A Volume   (0=50%, 1=100%)
  3     R/W  DMA Sound B Volume   (0=50%, 1=100%)
  4-7   -    Not used
  8     R/W  DMA Sound A Enable RIGHT (0=Disable, 1=Enable)
  9     R/W  DMA Sound A Enable LEFT  (0=Disable, 1=Enable)
  10    R/W  DMA Sound A Timer Select (0=Timer 0, 1=Timer 1)
  11    W?   DMA Sound A Reset FIFO   (1=Reset)
  12    R/W  DMA Sound B Enable RIGHT (0=Disable, 1=Enable)
  13    R/W  DMA Sound B Enable LEFT  (0=Disable, 1=Enable)
  14    R/W  DMA Sound B Timer Select (0=Timer 0, 1=Timer 1)
  15    W?   DMA Sound B Reset FIFO   (1=Reset)

4000084h - SOUNDCNT_X (NR52) - Sound on/off (R/W)
Bits 0-3 are automatically set when starting sound output, and are automatically cleared when a sound ends. (Ie. when the length expires, as far as length is enabled. The bits are NOT reset when an volume envelope ends.)

  Bit        Expl.
  0     R    Sound 1 ON flag (Read Only)
  1     R    Sound 2 ON flag (Read Only)
  2     R    Sound 3 ON flag (Read Only)
  3     R    Sound 4 ON flag (Read Only)
  4-6   -    Not used
  7     R/W  PSG/FIFO Master Enable (0=Disable, 1=Enable) (Read/Write)
  8-31  -    Not used

While Bit 7 is cleared, both PSG and FIFO sounds are disabled, and all PSG registers at 4000060h..4000081h are reset to zero (and must be re-initialized after re-enabling sound). However, registers 4000082h and 4000088h are kept read/write-able (of which, 4000082h has no function when sound is off, whilst 4000088h does work even when sound is off).

4000088h - SOUNDBIAS - Sound PWM Control (R/W, see below)
This register controls the final sound output. The default setting is 0200h, it is normally not required to change this value.

  Bit        Expl.
  0     -    Not used
  1-9   R/W  Bias Level (Default=100h, converting signed samples into unsigned)
  10-13 -    Not used
  14-15 R/W  Amplitude Resolution/Sampling Cycle (Default=0, see below)
  16-31 -    Not used

Amplitude Resolution/Sampling Cycle (0-3):

  0  9bit / 32.768kHz   (Default, best for DMA channels A,B)
  1  8bit / 65.536kHz
  2  7bit / 131.072kHz
  3  6bit / 262.144kHz  (Best for PSG channels 1-4)

For more information on this register, read the descriptions below.

400008Ch - Not used
400008Eh - Not used

Max Output Levels (with max volume settings)
Each of the two FIFOs can span the FULL output range (+/-200h).
Each of the four PSGs can span one QUARTER of the output range (+/-80h).
The current output levels of all six channels are added together by hardware.
So together, the FIFOs and PSGs, could reach THRICE the range (+/-600h).
The BIAS value is added to that signed value. With default BIAS (200h), the possible range becomes -400h..+800h, however, values that exceed the unsigned 10bit output range of 0..3FFh are clipped to MinMax(0,3FFh).

Resampling to 32.768kHz / 9bit (default)
The PSG channels 1-4 are internally generated at 262.144kHz, and DMA sound A-B could be theoretically generated at timer rates up to 16.78MHz. However, the final sound output is resampled to a rate of 32.768kHz, at 9bit depth (the above 10bit value, divided by two). If necessary, rates higher than 32.768kHz can be selected in the SOUNDBIAS register, that would result in a depth smaller than 9bit though.

PWM (Pulse Width Modulation) Output 16.78MHz / 1bit
Okay, now comes the actual output. The GBA can output only two voltages (low and high), these 'bits' are output at system clock speed (16.78MHz). If using the default 32.768kHz sampling rate, then 512 bits are output per sample (512*32K=16M). Each sample value (9bit range, N=0..511), would be then output as N low bits, followed by 512-N high bits. The resulting 'noise' is smoothed down by capacitors, by the speaker, and by human hearing, so that it will effectively sound like clean D/A converted 9bit voltages at 32kHz sampling rate.

Changing the BIAS Level
Normally use 200h for clean sound output. A value of 000h might make sense during periods when no sound is output (causing the PWM circuit to output low-bits only, which is eventually reducing the power consumption, and/or preventing 32KHz noise). Note: Using the SoundBias function (SWI 19h) allows to change the level by slowly incrementing or decrementing it (without hard scratch noise).

Low Power Mode
When not using sound output, power consumption can be reduced by setting both 4000084h (PSG/FIFO) and 4000088h (BIAS) to zero.

GBA Comparison of CGB and GBA Sound

The GBA sound controller is mostly the same than that of older monochrome gameboy and CGB. The following changes have been done:

New Sound Channels
Two new sound channels have been added that may be used to replay 8bit digital sound. Sample rate and sample data must be supplied by using a Timer and a DMA channel.

New Control Registers
The SOUNDCNT_H register controls the new DMA channels - as well as mixing with the four old channels. The SOUNDBIAS register controls the final sound output.

Sound Channel 3 Changes
The length of the Wave RAM is doubled by dividing it into two banks of 32 digits each, either one or both banks may be replayed (one after each other), for details check NR30 Bit 5-6. Optionally, the sound may be output at 75% volume, for details check NR32 Bit 7.

Changed Control Registers
NR50 is not supporting Vin signals (that's been an external sound input from cartridge).

Changed I/O Addresses
The GBAs sound register are located at 04000060-040000AE instead of at FF10-FF3F as in CGB and monochrome gameboy. However, note that there have been new blank spaces inserted between some of the separate registers - therefore it is NOT possible to port CGB software to GBA just by changing the sound base address.

Accessing I/O Registers
In some cases two of the old 8bit registers are packed into a 16bit register and may be accessed as such.

GBA Timers

The GBA includes four incrementing 16bit timers.
Timer 0 and 1 can be used to supply the sample rate for DMA sound channel A and/or B.

4000100h - TM0CNT_L - Timer 0 Counter/Reload (R/W)
4000104h - TM1CNT_L - Timer 1 Counter/Reload (R/W)
4000108h - TM2CNT_L - Timer 2 Counter/Reload (R/W)
400010Ch - TM3CNT_L - Timer 3 Counter/Reload (R/W)
Writing to these registers initializes the <reload> value (but does not directly affect the current counter value). Reading returns the current <counter> value (or the recent/frozen counter value if the timer has been stopped).
The reload value is copied into the counter only upon following two situations: Automatically upon timer overflows, or when the timer start bit becomes changed from 0 to 1.
Note: When simultaneously changing the start bit from 0 to 1, and setting the reload value at the same time (by a single 32bit I/O operation), then the newly written reload value is recognized as new counter value.

4000102h - TM0CNT_H - Timer 0 Control (R/W)
4000106h - TM1CNT_H - Timer 1 Control (R/W)
400010Ah - TM2CNT_H - Timer 2 Control (R/W)
400010Eh - TM3CNT_H - Timer 3 Control (R/W)

  Bit   Expl.
  0-1   Prescaler Selection (0=F/1, 1=F/64, 2=F/256, 3=F/1024)
  2     Count-up Timing   (0=Normal, 1=See below)  ;Not used in TM0CNT_H
  3-5   Not used
  6     Timer IRQ Enable  (0=Disable, 1=IRQ on Timer overflow)
  7     Timer Start/Stop  (0=Stop, 1=Operate)
  8-15  Not used

When Count-up Timing is enabled, the prescaler value is ignored, instead the time is incremented each time when the previous counter overflows. This function cannot be used for Timer 0 (as it is the first timer).
F = System Clock (16.78MHz).

GBA DMA Transfers

Overview
The GBA includes four DMA channels, the highest priority is assigned to DMA0, followed by DMA1, DMA2, and DMA3. DMA Channels with lower priority are paused until channels with higher priority have completed.
The CPU is paused when DMA transfers are active, however, the CPU is operating during the periods when Sound/Blanking DMA transfers are paused.

Special features of the separate DMA channels
DMA0 - highest priority, best for timing critical transfers (eg. HBlank DMA).
DMA1 and DMA2 - can be used to feed digital sample data to the Sound FIFOs.
DMA3 - can be used to write to Game Pak ROM/FlashROM (but not GamePak SRAM).
Beside for that, each DMA 0-3 may be used for whatever general purposes.

40000B0h,0B2h - DMA0SAD - DMA 0 Source Address (W) (internal memory)
40000BCh,0BEh - DMA1SAD - DMA 1 Source Address (W) (any memory)
40000C8h,0CAh - DMA2SAD - DMA 2 Source Address (W) (any memory)
40000D4h,0D6h - DMA3SAD - DMA 3 Source Address (W) (any memory)
The most significant address bits are ignored, only the least significant 27 or 28 bits are used (max 07FFFFFFh internal memory, or max 0FFFFFFFh any memory - except SRAM ?!).

40000B4h,0B6h - DMA0DAD - DMA 0 Destination Address (W) (internal memory)
40000C0h,0C2h - DMA1DAD - DMA 1 Destination Address (W) (internal memory)
40000CCh,0CEh - DMA2DAD - DMA 2 Destination Address (W) (internal memory)
40000D8h,0DAh - DMA3DAD - DMA 3 Destination Address (W) (any memory)
The most significant address bits are ignored, only the least significant 27 or 28 bits are used (max. 07FFFFFFh internal memory or 0FFFFFFFh any memory - except SRAM ?!).

40000B8h - DMA0CNT_L - DMA 0 Word Count (W) (14 bit, 1..4000h)
40000C4h - DMA1CNT_L - DMA 1 Word Count (W) (14 bit, 1..4000h)
40000D0h - DMA2CNT_L - DMA 2 Word Count (W) (14 bit, 1..4000h)
40000DCh - DMA3CNT_L - DMA 3 Word Count (W) (16 bit, 1..10000h)
Specifies the number of data units to be transferred, each unit is 16bit or 32bit depending on the transfer type, a value of zero is treated as max length (ie. 4000h, or 10000h for DMA3).

40000BAh - DMA0CNT_H - DMA 0 Control (R/W)
40000C6h - DMA1CNT_H - DMA 1 Control (R/W)
40000D2h - DMA2CNT_H - DMA 2 Control (R/W)
40000DEh - DMA3CNT_H - DMA 3 Control (R/W)

  Bit   Expl.
  0-4   Not used
  5-6   Dest Addr Control  (0=Increment,1=Decrement,2=Fixed,3=Increment/Reload)
  7-8   Source Adr Control (0=Increment,1=Decrement,2=Fixed,3=Prohibited)
  9     DMA Repeat                   (0=Off, 1=On) (Must be zero if Bit 11 set)
  10    DMA Transfer Type            (0=16bit, 1=32bit)
  11    Game Pak DRQ  - DMA3 only -  (0=Normal, 1=DRQ <from> Game Pak, DMA3)
  12-13 DMA Start Timing  (0=Immediately, 1=VBlank, 2=HBlank, 3=Special)
          The 'Special' setting (Start Timing=3) depends on the DMA channel:
          DMA0=Prohibited, DMA1/DMA2=Sound FIFO, DMA3=Video Capture
  14    IRQ upon end of Word Count   (0=Disable, 1=Enable)
  15    DMA Enable                   (0=Off, 1=On)

After changing the Enable bit from 0 to 1, wait 2 clock cycles before accessing any DMA related registers.

When accessing OAM (7000000h) or OBJ VRAM (6010000h) by HBlank Timing, then the "H-Blank Interval Free" bit in DISPCNT register must be set.

Source and Destination Address and Word Count Registers
The SAD, DAD, and CNT_L registers are holding the initial start addresses, and initial length. The hardware does NOT change the content of these registers during or after the transfer.
The actual transfer takes place by using internal pointer/counter registers. The initial values are copied into internal regs under the following circumstances:
Upon DMA Enable (Bit 15) changing from 0 to 1: Reloads SAD, DAD, CNT_L.
Upon Repeat: Reloads CNT_L, and optionally DAD (Increment+Reload).

DMA Repeat bit
If the Repeat bit is cleared: The Enable bit is automatically cleared after the specified number of data units has been transferred.
If the Repeat bit is set: The Enable bit remains set after the transfer, and the transfer will be restarted each time when the Start condition (eg. HBlank, Fifo) becomes true. The specified number of data units is transferred <each> time when the transfer is (re-)started. The transfer will be repeated forever, until it gets stopped by software.

Sound DMA (FIFO Timing Mode) (DMA1 and DMA2 only)
In this mode, the DMA Repeat bit must be set, and the destination address must be FIFO_A (040000A0h) or FIFO_B (040000A4h).
Upon DMA request from sound controller, 4 units of 32bits (16 bytes) are transferred (both Word Count register and DMA Transfer Type bit are ignored). The destination address will not be incremented in FIFO mode.
Keep in mind that DMA channels of higher priority may offhold sound DMA. For example, when using a 64 kHz sample rate, 16 bytes of sound DMA data are requested each 0.25ms (4 kHz), at this time another 16 bytes are still in the FIFO so that there's still 0.25ms time to satisfy the DMA request. Thus DMAs with higher priority should not be operated for longer than 0.25ms. (This problem does not arise for HBlank transfers as HBlank time is limited to 16.212us.)

Game Pak DMA
Only DMA 3 may be used to transfer data to/from Game Pak ROM or Flash ROM - it cannot access Game Pak SRAM though (as SRAM data bus is limited to 8bit units). In normal mode, DMA is requested as long until Word Count becomes zero. When setting the 'Game Pack DRQ' bit, then the cartridge must contain an external circuit which outputs a /DREQ signal. Note that there is only one pin for /DREQ and /IREQ, thus the cartridge may not supply /IREQs while using DRQ mode.

Video Capture Mode (DMA3 only)
Intended to copy a bitmap from memory (or from external hardware/camera) to VRAM. When using this transfer mode, set the repeat bit, and write the number of data units (per scanline) to the word count register. Capture works similar like HBlank DMA, however, the transfer is started when VCOUNT=2, it is then repeated each scanline, and it gets stopped when VCOUNT=162.

Transfer End
The DMA Enable flag (Bit 15) is automatically cleared upon completion of the transfer. The user may also clear this bit manually in order to stop the transfer (obviously this is possible for Sound/Blanking DMAs only, in all other cases the CPU is stopped until the transfer completes by itself).

Transfer Rate/Timing
Except for the first data unit, all units are transferred by sequential reads and writes. For n data units, the DMA transfer time is:

  2N+2(n-1)S+xI

Of which, 1N+(n-1)S are read cycles, and the other 1N+(n-1)S are write cycles, actual number of cycles depends on the waitstates and bus-width of the source and destination areas (as described in CPU Instruction Cycle Times chapter). Internal time for DMA processing is 2I (normally), or 4I (if both source and destination are in gamepak memory area).

DMA lockup when stopping while starting ???
Capture delayed, Capture Enable=AutoCleared ???

GBA Communication Ports

The GBAs Serial Port may be used in various different communication modes. Normal mode may exchange data between two GBAs (or to transfer data from master GBA to several slave GBAs in one-way direction).
Multi-player mode may exchange data between up to four GBAs. UART mode works much like a RS232 interface. JOY Bus mode uses a standardized Nintendo protocol. And General Purpose mode allows to mis-use the 'serial' port as bi-directional 4bit parallel port.
Note: The Nintendo DS does not include a Serial Port.

SIO Normal Mode
SIO Multi-Player Mode
SIO UART Mode
SIO JOY BUS Mode
SIO General-Purpose Mode
SIO Control Registers Summary

Wireless Adapter
GBA Wireless Adapter

Infrared Communication Adapters
Even though early GBA prototypes have been intended to support IR communication, this feature has been removed.
However, Nintendo is apparently considering to provide an external IR adapter (to be connected to the SIO connector, being accessed in General Purpose mode).
Also, it'd be theoretically possible to include IR ports built-in in game cartridges (as done for some older 8bit/monochrome Hudson games).

SIO Normal Mode

This mode is used to communicate between two units.
Transfer rates of 256Kbit/s or 2Mbit/s can be selected, however, the fast 2Mbit/s is intended ONLY for special hardware expansions that are DIRECTLY connected to the GBA link port (ie. without a cable being located between the GBA and expansion hardware). In normal cases, always use 256Kbit/s transfer rate which provides stable results.
Transfer lengths of 8bit or 32bit may be used, the 8bit mode is the same as for older DMG/CGB gameboys, however, the voltages for "GBA cartridges in GBAs" are different as for "DMG/CGB cartridges in DMG/CGB/GBAs", ie. it is not possible to communicate between DMG/CGB games and GBA games.

4000134h - RCNT (R) - Mode Selection, in Normal/Multiplayer/UART modes (R/W)

  Bit   Expl.
  0-3   Undocumented (current SC,SD,SI,SO state, as for General Purpose mode)
  4-8   Not used     (Should be 0, bits are read/write-able though)
  9-13  Not used     (Always 0, read only)
  14    Not used     (Should be 0, bit is read/write-able though)
  15    Must be zero (0) for Normal/Multiplayer/UART modes

4000128h - SIOCNT - SIO Control, usage in NORMAL Mode (R/W)

  Bit   Expl.
  0     Shift Clock (SC)        (0=External, 1=Internal)
  1     Internal Shift Clock    (0=256KHz, 1=2MHz)
  2     SI State (opponents SO) (0=Low, 1=High/None) --- (Read Only)
  3     SO during inactivity    (0=Low, 1=High) (applied ONLY when Bit7=0)
  4-6   Not used                (Read only, always 0 ?)
  7     Start Bit               (0=Inactive/Ready, 1=Start/Active)
  8-11  Not used                (R/W, should be 0)
  12    Transfer Length         (0=8bit, 1=32bit)
  13    Must be "0" for Normal Mode
  14    IRQ Enable              (0=Disable, 1=Want IRQ upon completion)
  15    Not used                (Read only, always 0)

The Start bit is automatically reset when the transfer completes, ie. when all 8 or 32 bits are transferred, at that time an IRQ may be generated.

400012Ah - SIODATA8 - SIO Normal Communication 8bit Data (R/W)
For 8bit normal mode. Contains 8bit data (only lower 8bit are used). Outgoing data should be written to this register before starting the transfer. During transfer, transmitted bits are shifted-out (MSB first), and received bits are shifted-in simultaneously. Upon transfer completion, the register contains the received 8bit value.

4000120h - SIODATA32_L - SIO Normal Communication lower 16bit data (R/W)
4000122h - SIODATA32_H - SIO Normal Communication upper 16bit data (R/W)
Same as above SIODATA8, for 32bit normal transfer mode respectively.
SIOCNT/RCNT must be set to 32bit normal mode <before> writing to SIODATA32.

Initialization
First, initialize RCNT register. Second, set mode/clock bits in SIOCNT with startbit cleared. For master: select internal clock, and (in most cases) specify 256KHz as transfer rate. For slave: select external clock, the local transfer rate selection is then ignored, as the transfer rate is supplied by the remote GBA (or other computer, which might supply custom transfer rates).
Third, set the startbit in SIOCNT with mode/clock bits unchanged.

Recommended Communication Procedure for SLAVE unit (external clock)
- Initialize data which is to be sent to master.
- Set Start flag.
- Set SO to LOW to indicate that master may start now.
- Wait for IRQ (or for Start bit to become zero). (Check timeout here!)
- Set SO to HIGH to indicate that we are not ready.
- Process received data.
- Repeat procedure if more data is to be transferred.
(or is so=high done automatically? would be fine - more stable - otherwise master may still need delay)

Recommended Communication Procedure for SLAVE unit (external clock)
- Initialize data which is to be sent to master.
- Set Start=0 and SO=0 (SO=LOW indicates that slave is (almost) ready).
- Set Start=1 and SO=1 (SO=HIGH indicates not ready, applied after transfer).

  (Expl. Old SO=LOW kept output until 1st clock bit received).
  (Expl. New SO=HIGH is automatically output at transfer completion).

- Set SO to LOW to indicate that master may start now.
- Wait for IRQ (or for Start bit to become zero). (Check timeout here!)
- Process received data.
- Repeat procedure if more data is to be transferred.

Recommended Communication Procedure for MASTER unit (internal clock)
- Initialize data which is to be sent to slave.
- Wait for SI to become LOW (slave ready). (Check timeout here!)
- Set Start flag.
- Wait for IRQ (or for Start bit to become zero).
- Process received data.
- Repeat procedure if more data is to be transferred.

Cable Protocol
During inactive transfer, the shift clock (SC) is high. The transmit (SO) and receive (SI) data lines may be manually controlled as described above.
When master sends SC=LOW, each master and slave must output the next outgoing data bit to SO. When master sends SC=HIGH, each master and slave must read out the opponents data bit from SI. This is repeated for each of the 8 or 32 bits, and when completed SC will be kept high again.

Transfer Rates
Either 256KHz or 2MHz rates can be selected for SC, so max 32KBytes (256Kbit) or 128KBytes (2Mbit) can be transferred per second. However, the software must process each 8bit or 32bit of transmitted data separately, so the actual transfer rate will be reduced by the time spent on handling each data unit.
Only 256KHz provides stable results in most cases (such like when linking between two GBAs). The 2MHz rate is intended for special expansion hardware (with very short wires) only.

Using Normal mode for One-Way Multiplayer communication
When using normal mode with multiplay-cables, data isn't exchanged between first and second GBA as usually. Instead, data is shifted from first to last GBA (the first GBA receives zero, because master SI is shortcut to GND).
This behaviour may be used for fast ONE-WAY data transfer from master to all other GBAs. For example (3 GBAs linked):

  Step         Sender      1st Recipient   2nd Recipient
  Transfer 1:  DATA #0 --> UNDEF      -->  UNDEF     -->
  Transfer 2:  DATA #1 --> DATA #0    -->  UNDEF     -->
  Transfer 3:  DATA #2 --> DATA #1    -->  DATA #0   -->
  Transfer 4:  DATA #3 --> DATA #2    -->  DATA #1   -->

The recipients should not output any own data, instead they should forward the previously received data to the next recipient during next transfer (just keep the incoming data unmodified in the data register).
Due to the delayed forwarding, 2nd recipient should ignore the first incoming data. After the last transfer, the sender must send one (or more) dummy data unit(s), so that the last data is forwarded to the 2nd (or further) recipient(s).

SIO Multi-Player Mode

Multi-Player mode can be used to communicate between up to 4 units.

4000134h - RCNT (R) - Mode Selection, in Normal/Multiplayer/UART modes (R/W)

  Bit   Expl.
  0-3   Undocumented (current SC,SD,SI,SO state, as for General Purpose mode)
  4-8   Not used     (Should be 0, bits are read/write-able though)
  9-13  Not used     (Always 0, read only)
  14    Not used     (Should be 0, bit is read/write-able though)
  15    Must be zero (0) for Normal/Multiplayer/UART modes

Note: Even though undocumented, many Nintendo games are using Bit 0 to test current SC state in multiplay mode.

4000128h - SIOCNT - SIO Control, usage in MULTI-PLAYER Mode (R/W)

  Bit   Expl.
  0-1   Baud Rate     (0-3: 9600,38400,57600,115200 bps)
  2     SI-Terminal   (0=Parent, 1=Child)                  (Read Only)
  3     SD-Terminal   (0=Bad connection, 1=All GBAs Ready) (Read Only)
  4-5   Multi-Player ID     (0=Parent, 1-3=1st-3rd child)  (Read Only)
  6     Multi-Player Error  (0=Normal, 1=Error)            (Read Only)
  7     Start/Busy Bit      (0=Inactive, 1=Start/Busy) (Read Only for Slaves)
  8-11  Not used            (R/W, should be 0)
  12    Must be "0" for Multi-Player mode
  13    Must be "1" for Multi-Player mode
  14    IRQ Enable          (0=Disable, 1=Want IRQ upon completion)
  15    Not used            (Read only, always 0)

The ID Bits are undefined until the first transfer has completed.

400012Ah - SIOMLT_SEND - Data Send Register (R/W)
Outgoing data (16 bit) which is to be sent to the other GBAs.

4000120h - SIOMULTI0 - SIO Multi-Player Data 0 (Parent) (R/W)
4000122h - SIOMULTI1 - SIO Multi-Player Data 1 (1st child) (R/W)
4000124h - SIOMULTI2 - SIO Multi-Player Data 2 (2nd child) (R/W)
4000126h - SIOMULTI3 - SIO Multi-Player Data 3 (3rd child) (R/W)
These registers are automatically reset to FFFFh upon transfer start.
After transfer, these registers contain incoming data (16bit each) from all remote GBAs (if any / otherwise still FFFFh), as well as the local outgoing SIOMLT_SEND data.
Ie. after the transfer, all connected GBAs will contain the same values in their SIOMULTI0-3 registers.

Initialization
- Initialize RCNT Bit 14-15 and SIOCNT Bit 12-13 to select Multi-Player mode.
- Read SIOCNT Bit 3 to verify that all GBAs are in Multi-Player mode.
- Read SIOCNT Bit 2 to detect whether this is the Parent/Master unit.

Recommended Transmission Procedure
- Write outgoing data to SIODATA_SEND.
- Master must set Start bit.
- All units must process received data in SIOMULTI0-3 when transfer completed.
- After the first successful transfer, ID Bits in SIOCNT are valid.
- If more data is to be transferred, repeat procedure.
The parent unit blindly sends data regardless of whether childs have already processed old data/supplied new data. So, parent unit might be required to insert delays between each transfer, and/or perform error checking.
Also, slave units may signalize that they are not ready by temporarily switching into another communication mode (which does not output SD High, as Multi-Player mode does during inactivity).

Transfer Protocol
Beginning
- The masters SI pin is always LOW.
- When all GBAs are in Multiplayer mode (ready) SD is HIGH.
- When master starts the transfer, it sets SC=LOW, slaves receive Busy bit.
Step A
- ID Bits in master unit are set to 0.
- Master outputs Startbit (LOW), 16bit Data, Stopbit (HIGH) through SD.
- This data is written to SIOMULTI0 of all GBAs (including master).
- Master forwards LOW from its SO to 1st childs SI.
- Transfer ends if next child does not output data after certain time.
Step B
- ID Bits in 1st child unit are set to 1.
- 1st Child outputs Startbit (LOW), 16bit Data, Stopbit (HIGH) through SD.
- This data is written to SIOMULTI1 of all GBAs (including 1st child).
- 1st child forwards LOW from its SO to 2nd childs SI.
- Transfer ends if next child does not output data after certain time.
Step C
- ID Bits in 2nd child unit are set to 2.
- 2nd Child outputs Startbit (LOW), 16bit Data, Stopbit (HIGH) through SD.
- This data is written to SIOMULTI2 of all GBAs (including 2nd child).
- 2nd child forwards LOW from its SO to 3rd childs SI.
- Transfer ends if next child does not output data after certain time.
Step D
- ID Bits in 3rd child unit are set to 3.
- 3rd Child outputs Startbit (LOW), 16bit Data, Stopbit (HIGH) through SD.
- This data is written to SIOMULTI3 of all GBAs (including 3rd child).
- Transfer ends (this was the last child).
Transfer end
- Master sets SC=HIGH, all GBAs set SO=HIGH.
- The Start/Busy bits of all GBAs are automatically cleared.
- Interrupts are requested in all GBAs (as far as enabled).

Error Bit
This bit is set when a slave did not receive SI=LOW even though SC=LOW signalized a transfer (this might happen when connecting more than 4 GBAs, or when the previous child is not connected). Also, the bit is set when a Stopbit wasn't HIGH.
The error bit may be undefined during active transfer - read only after transfer completion (the transfer continues and completes as normal even if errors have occurred for some or all GBAs).
Don't know: The bit is automatically reset/initialized with each transfer, or must be manually reset?

Transmission Time
The transmission time depends on the selected Baud rate. And on the amount of Bits (16 data bits plus start/stop bits for each GBA), delays between data for each GBA, plus final timeout (if less than 4 GBAs). That is, depending on the number of connected GBAs:

  GBAs    Bits    Delays   Timeout
  1       18      None     Yes
  2       36      1        Yes
  3       54      2        Yes
  4       72      3        None

(The average Delay and Timeout periods are unknown?)
Above is not counting the additional CPU time that must be spent on initiating and processing each transfer.

Fast One-Way Transmission
Beside for the actual SIO Multiplayer mode, you can also use SIO Normal mode for fast one-way data transfer from Master unit to all Child unit(s). See chapter about SIO Normal mode for details.

SIO UART Mode

This mode works much like a RS232 port, however, the voltages are unknown, probably 0/3V rather than +/-12V ?. SI and SO are data lines (with crossed wires), SC and SD signalize Clear to Send (with crossed wires also, which requires special cable when linking between two GBAs ?)

4000134h - RCNT (R) - Mode Selection, in Normal/Multiplayer/UART modes (R/W)

  Bit   Expl.
  0-3   Undocumented (current SC,SD,SI,SO state, as for General Purpose mode)
  4-8   Not used     (Should be 0, bits are read/write-able though)
  9-13  Not used     (Always 0, read only)
  14    Not used     (Should be 0, bit is read/write-able though)
  15    Must be zero (0) for Normal/Multiplayer/UART modes

4000128h - SCCNT_L - SIO Control, usage in UART Mode (R/W)

  Bit   Expl.
  0-1   Baud Rate  (0-3: 9600,38400,57600,115200 bps)
  2     CTS Flag   (0=Send always/blindly, 1=Send only when SC=LOW)
  3     Parity Control (0=Even, 1=Odd)
  4     Send Data Flag      (0=Not Full,  1=Full)    (Read Only)
  5     Receive Data Flag   (0=Not Empty, 1=Empty)   (Read Only)
  6     Error Flag          (0=No Error,  1=Error)   (Read Only)
  7     Data Length         (0=7bits,   1=8bits)
  8     FIFO Enable Flag    (0=Disable, 1=Enable)
  9     Parity Enable Flag  (0=Disable, 1=Enable)
  10    Send Enable Flag    (0=Disable, 1=Enable)
  11    Receive Enable Flag (0=Disable, 1=Enable)
  12    Must be "1" for UART mode
  13    Must be "1" for UART mode
  14    IRQ Enable          (0=Disable, 1=IRQ when any Bit 4/5/6 become set)
  15    Not used            (Read only, always 0)

400012Ah - SIODATA8 - usage in UART Mode (R/W)
Addresses the send/receive shift register, or (when FIFO is used) the send/receive FIFO. In either case only the lower 8bit of SIODATA8 are used, the upper 8bit are not used.
The send/receive FIFO may store up to four 8bit data units each. For example, while 1 unit is still transferred from the send shift register, it is possible to deposit another 4 units in the send FIFO, which are then automatically moved to the send shift register one after each other.

Send/Receive Enable, CTS Feedback
The receiver outputs SD=LOW (which is input as SC=LOW at the remote side) when it is ready to receive data (that is, when Receive Enable is set, and the Receive shift register (or receive FIFO) isn't full.
When CTS flag is set to always/blindly, then the sender transmits data immediately when Send Enable is set, otherwise data is transmitted only when Send Enable is set and SC is LOW.

Error Flag
The error flag is set when a bad stop bit has been received (stop bit must be 0), when a parity error has occurred (if enabled), or when new data has been completely received while the receive data register (or receive FIFO) is already full.
The error flag is automatically reset when reading from SIOCNT register.

Init & Initback
The content of the FIFO is reset when FIFO is disabled in UART mode, thus, when entering UART mode initially set FIFO=disabled.
The Send/Receive enable bits must be reset before switching from UART mode into another SIO mode!

SIO JOY BUS Mode

This communication mode uses Nintendo's standardized JOY Bus protocol. When using this communication mode, the GBA is always operated as SLAVE!

In this mode, SI and SO pins are data lines (apparently synchronized by Start/Stop bits?), SC and SD are set to low (including during active transfer?), the transfer rate is unknown?

4000134h - RCNT (R) - Mode Selection, in JOY BUS mode (R/W)

  Bit   Expl.
  0-3   Undocumented (current SC,SD,SI,SO state, as for General Purpose mode)
  4-8   Not used     (Should be 0, bits are read/write-able though)
  9-13  Not used     (Always 0, read only)
  14    Must be "1" for JOY BUS Mode
  15    Must be "1" for JOY BUS Mode

4000128h - SIOCNT - SIO Control, not used in JOY BUS Mode
This register is not used in JOY BUS mode.

4000140h - JOYCNT - JOY BUS Control Register (R/W)

  Bit   Expl.
  0     Device Reset Flag     (Command FFh)          (Read/Acknowledge)
  1     Receive Complete Flag (Command 14h or 15h?)  (Read/Acknowledge)
  2     Send Complete Flag    (Command 15h or 14h?)  (Read/Acknowledge)
  3-5   Not used
  6     IRQ when receiving a Device Reset Command  (0=Disable, 1=Enable)
  7-31  Not used

Bit 0-2 are working much like the bits in the IF register: Write a "1" bit to reset (acknowledge) the respective bit.
UNCLEAR: Interrupts can be requested for Send/Receive commands also?

4000150h - JOY_RECV_L - Receive Data Register low (R/W)
4000152h - JOY_RECV_H - Receive Data Register high (R/W)
4000154h - JOY_TRANS_L - Send Data Register low (R/W)
4000156h - JOY_TRANS_H - Send Data Register high (R/W)
Send/receive data registers.

4000158h - JOYSTAT - Receive Status Register (R/W)

  Bit   Expl.
  0     Not used
  1     Receive Status Flag   (0=Remote GBA is/was receiving) (Read Only?)
  2     Not used
  3     Send Status Flag      (1=Remote GBA is/was sending)   (Read Only?)
  4-5   General Purpose Flag  (Not assigned, may be used for whatever purpose)
  6-31  Not used

Bit 1 is automatically set when writing to local JOY_TRANS.
Bit 3 is automatically reset when reading from local JOY_RECV.

Below are the four possible commands which can be received by the GBA. Note that the GBA (slave) cannot send any commands itself, all it can do is to read incoming data, and to provide 'reply' data which may (or may not) be read out by the master unit.

Command FFh - Device Reset

  Receive FFh (Command)
  Send    00h (GBA Type number LSB (or MSB?))
  Send    04h (GBA Type number MSB (or LSB?))
  Send    XXh (lower 8bits of SIOSTAT register)

Command 00h - Type/Status Data Request

  Receive 00h (Command)
  Send    00h (GBA Type number LSB (or MSB?))
  Send    04h (GBA Type number MSB (or LSB?))
  Send    XXh (lower 8bits of SIOSTAT register)

Command 15h - GBA Data Write (to GBA)

  Receive 15h (Command)
  Receive XXh (Lower 8bits of JOY_RECV_L)
  Receive XXh (Upper 8bits of JOY_RECV_L)
  Receive XXh (Lower 8bits of JOY_RECV_H)
  Receive XXh (Upper 8bits of JOY_RECV_H)
  Send    XXh (lower 8bits of SIOSTAT register)

Command 14h - GBA Data Read (from GBA)

  Receive 14h (Command)
  Send    XXh (Lower 8bits of JOY_TRANS_L)
  Send    XXh (Upper 8bits of JOY_TRANS_L)
  Send    XXh (Lower 8bits of JOY_TRANS_H)
  Send    XXh (Upper 8bits of JOY_TRANS_H)
  Send    XXh (lower 8bits of SIOSTAT register)

SIO General-Purpose Mode

In this mode, the SIO is 'misused' as a 4bit bi-directional parallel port, each of the SI,SO,SC,SD pins may be directly controlled, each can be separately declared as input (with internal pull-up) or as output signal.

4000134h - RCNT (R) - SIO Mode, usage in GENERAL-PURPOSE Mode (R/W)
Interrupts can be requested when SI changes from HIGH to LOW, as General Purpose mode does not require a serial shift clock, this interrupt may be produced even when the GBA is in Stop (low power standby) state.

  Bit   Expl.
  0     SC Data Bit         (0=Low, 1=High)
  1     SD Data Bit         (0=Low, 1=High)
  2     SI Data Bit         (0=Low, 1=High)
  3     SO Data Bit         (0=Low, 1=High)
  4     SC Direction        (0=Input, 1=Output)
  5     SD Direction        (0=Input, 1=Output)
  6     SI Direction        (0=Input, 1=Output, but see below)
  7     SO Direction        (0=Input, 1=Output)
  8     SI Interrupt Enable (0=Disable, 1=Enable)
  9-13  Not used
  14    Must be "0" for General-Purpose Mode
  15    Must be "1" for General-Purpose or JOYBUS Mode

SI should be always used as Input to avoid problems with other hardware which does not expect data to be output there.

4000128h - SIOCNT - SIO Control, not used in GENERAL-PURPOSE Mode
This register is not used in general purpose mode. That is, the separate bits of SIOCNT still exist and are read- and/or write-able in the same manner as for Normal, Multiplay, or UART mode (depending on SIOCNT Bit 12,13), but are having no effect on data being output to the link port.

SIO Control Registers Summary

Mode Selection (by RCNT.15-14 and SIOCNT.13-12)

  R.15 R.14 S.13 S.12 Mode
  0    x    0    0    Normal 8bit
  0    x    0    1    Normal 32bit
  0    x    1    0    Multiplay 16bit
  0    x    1    1    UART (RS232)
  1    0    x    x    General Purpose
  1    1    x    x    JOY BUS

SIOCNT

  Bit    0      1    2     3      4 5 6   7     8    9      10   11
  Normal Master Rate SI/In SO/Out - - -   Start -    -      -    -
  Multi  Baud   Baud SI/In SD/In  ID# Err Start -    -      -    -
  UART   Baud   Baud CTS   Parity S R Err Bits  FIFO Parity Send Recv

GBA Wireless Adapter

GBA Wireless Adapter (AGB-015 or OXY-004)
GBA Wireless Adapter Games
GBA Wireless Adapter Login
GBA Wireless Adapter Commands
GBA Wireless Adapter Component Lists

GBA Wireless Adapter Games

GBA Wireless Adapter compatible Games

  bit Generations series (Japan only)
  Boktai 2: Solar Boy Django (Konami)
  Boktai 3: Sabata's Counterattack
  Classic NES Series: Donkey Kong
  Classic NES Series: Dr. Mario
  Classic NES Series: Ice Climber
  Classic NES Series: Pac-Man
  Classic NES Series: Super Mario Bros.
  Classic NES Series: Xevious
  Digimon Racing (Bandai) (No Wireless Adapter support in European release)
  Dragon Ball Z: Buu's Fury (Atari)
  Famicom Mini Series: #13 Balloon Fight
  Famicom Mini Series: #12 Clu Clu Land
  Famicom Mini Series: #16 Dig Dug
  Famicom Mini Series: #02 Donkey Kong
  Famicom Mini Series: #15 Dr. Mario
  Famicom Mini Series: #03 Ice Climber
  Famicom Mini Series: #18 Makaimura
  Famicom Mini Series: #08 Mappy
  Famicom Mini Series: #11 Mario Bros.
  Famicom Mini Series: #06 Pac-Man
  Famicom Mini Series: #30 SD Gundam World Scramble Wars
  Famicom Mini Series: #01 Super Mario Bros.
  Famicom Mini Series: #21 Super Mario Bros.
  Famicom Mini Series: #19 Twin Bee
  Famicom Mini Series: #14 Wrecking Crew
  Famicom Mini Series: #07 Xevious
  Hamtaro: Ham-Ham Games (Nintendo)
  Lord of the Rings: The Third Age, The (EA Games)
  Mario Golf: Advance Tour (Nintendo)
  Mario Tennis: Power Tour (Nintendo)
  Mega Man Battle Network 5: Team Protoman (Capcom)
  Mega Man Battle Network 5: Team Colonel (Capcom)
  Mega Man Battle Network 6: Cybeast Falzar
  Mega Man Battle Network 6: Cybeast Gregar
  Momotaro Dentetsu G: Make a Gold Deck! (Japan only)
  Pokemon Emerald (Nintendo)
  Pokemon FireRed (Nintendo)
  Pokemon LeafGreen (Nintendo)
  Sennen Kazoku (Japan only)
  Shrek SuperSlam
  Sonic Advance 3

GBA Wireless Adapter Login

GBA Wireless Adapter Login

  rcnt=8000h    ;\
  rcnt=80A0h    ;
  rcnt=80A2h    ; reset adapter or so
  wait          ;
  rcnt=80A0h    ;/
  siocnt=5003h  ;\set 32bit normal mode, 2MHz internal clock
  rcnt=0000h    ;/
  passes=0, index=0
 @@lop:
  passes=passes+1, if passes>32 then ERROR  ;give up (usually only 10 passses)
  recv.lo=siodata AND FFFFh    ;response from adapter
  recv.hi=siodata/10000h       ;adapter's own "NI" data
  if send.hi<>recv.lo then index=0, goto @@stuck  ;<-- fallback to index=0
  if (send.lo XOR FFFFh)<>recv.lo then goto @@stuck
  if (send.hi XOR FFFFh)<>recv.hi then goto @@stuck
  index=index+1
 @@stuck:
  send.lo=halfword[@@key_string+index*2]
  send.hi=recv.hi XOR FFFFh
  siodata=send.lo+(send.hi*10000h)
  siocnt.bit7=1                        ;<-- start transmission
  if index<4 then goto @@lop
  ret
 @@key_string db 'NINTENDO',01h,80h    ;10 bytes (5 halfwords; index=0..4)

Data exchanged during Login

               GBA                         ADAPTER
               xxxx494E ;\     <-->        xxxxxxxx
               xxxx494E ; "NI" <--> "NI"/; 494EB6B1 ;\
  NOT("NI") /; B6B1494E ;/     <-->     \; 494EB6B1 ; NOT("NI")
            \; B6B1544E ;\"NT" <--> "NT"/; 544EB6B1 ;/
  NOT("NT") /; ABB1544E ;/     <-->     \; 544EABB1 ;\NOT("NT")
            \; ABB14E45 ;\"EN" <--> "EN"/; 4E45ABB1 ;/
  NOT("EN") /; B1BA4E45 ;/     <-->     \; 4E45B1BA ;\NOT("EN")
            \; B1BA4F44 ;\"DO" <--> "DO"/; 4F44B1BA ;/
  NOT("DO") /; B0BB4F44 ;/     <-->     \; 4F44B0BB ;\NOT("DO")
            \; B0BB8001 ;-fin  <-->  fin-; 8001B0BB ;/
                 \   \                      \   \
                  \   LSBs=Own               \   LSBs=Inverse of
                   \   Data.From.Gba          \   Prev.Data.From.Gba
                    \                          \
                     MSBs=Inverse of            MSBs=Own
                      Prev.Data.From.Adapter     Data.From.Adapter

GBA Wireless Adapter Commands

Wireless Command/Parameter Transmission

  GBA       Adapter
  9966ppcch 80000000h   ;-send command (cc), and num param_words (pp)
  <param01> 80000000h   ;\
  <param02> 80000000h   ; send "pp" parameter word(s), if any
  ...       ...         ;/
  80000000h 9966rraah   ;-recv ack (aa=cc+80h), and num response_words (rr)
  80000000? <reply01>   ;\
  80000000? <reply02>   ; recv "rr" response word(s), if any
  ...       ...         ;/

Wireless 32bit Transfers

  wait until [4000128h].Bit2=0  ;want SI=0
  set [4000128h].Bit3=1         ;set SO=1
  wait until [4000128h].Bit2=1  ;want SI=1
  set [4000128h].Bit3=0,Bit7=1  ;set SO=0 and start 32bit transfer

All command/param/reply transfers should be done at Internal Clock (except, Response Words for command 25h,27h,35h,37h should use External Clock).

Wireless Commands

  Cmd Para Reply Name
  10h -    -     Hello (send immediately after login)
  11h -    1     Good/Bad response to cmd 16h ?
  12h
  13h -    1
  14h
  15h
  16h 6    -     Introduce (send game/user name)
  17h 1    -     Config (send after Hello) (eg. param=003C0420h or 003C043Ch)
  18h
  19h
  1Ah
  1Bh
  1Ch -    -
  1Dh -    NN    Get Directory? (receive list of game/user names?)
  1Eh -    NN    Get Directory? (receive list of game/user names?)
  1Fh 1    -     Select Game for Download (send 16bit Game_ID)

  20h -    1
  21h -    1     Good/Bad response to cmd 1Fh ?
  22h
  23h
  24h -    -
  25h                                       ;use EXT clock!
  26h -    -
  27h -    -     Begin Download ?           ;use EXT clock!
  28h
  29h
  2Ah
  2Bh
  2Ch
  2Dh
  2Eh
  2Fh

  30h 1    -
  31h
  32h
  33h
  34h
  35h                                       ;use EXT clock!
  36h
  37h                                       ;use EXT clock!
  38h
  39h
  3Ah
  3Bh
  3Ch
  3Dh -    -     Bye (return to language select)
  3Eh
  3Fh

Special Response 996601EEh for error or so? (only at software side?)

GBA Wireless Adapter Component Lists

Main Chipset

  U1 32pin Freescale MC13190 (2.4 GHz ISM band transceiver)
  U2 48pin Freescale CT3000 or CT3001 (depending on adapter version)
  X3  2pin 9.5MHz crystal

The MC13190 is a Short-Range, Low-Power 2.4 GHz ISM band transceiver.
The processor is Motorola's 32-bit M-Core RISC engine. (?) MCT3000 (?)
See also: http://www.eetimes.com/document.asp?doc_id=1271943

Version with GERMAN Postal Code on sticker:

  Sticker on Case:
    "GAME BOY advance, WIRELESS ADAPTER"
    "Pat.Pend.Made in Philipines, CE0125(!)B"
    "MODEL NO./MODELE NO.AGB-015 D-63760 Grossosteim P/AGB-A-WA-EUR-2 E3"
  PCB: "19-C046-04, A-7" (top side) and "B-7" and Microchip ",\\" (bottom side)
  PCB: white stamp "3104, 94V-0, RU, TW-15"
  PCB: black stamp "22FDE"
  U1 32pin "Freescale 13190, 4WFQ" (MC13190) (2.4 GHz ISM band transceiver)
  U2 48pin "Freescale CT3001, XAC0445"  (bottom side)
  X3  2pin "D959L4I" (9.5MHz)           (top side) (ca. 19 clks per 2us)

Further components... top side (A-7)

  D1   5pin "D6F, 44"   (top side, below X3)
  U71  6pin ".., () 2"  (top side, right of X3, tiny black chip)
  B71  6pin "[]"        (top side, right of X3, small white chip)
  ANT  2pin on-board copper wings
  Q?   3pin             (top side, above CN1)
  Q?   3pin             (top side, above CN1)
  D?   2pin "72"        (top side, above CN1)
  D3   2pin "F2"        (top side, above CN1)
  U200 4pin "MSV"       (top side, above CN1)
  U202 5pin "LXKA"      (top side, right of CN1)
  U203 4pin "M6H"       (top side, right of CN1)
  CN1  6pin connector to GBA link port (top side)

Further components... bottom side (B-7)

  U201 5pin "LXVB"      (bottom side, near CN1)
  U72  4pin "BMs"       (bottom side, near ANT, tiny black chip)
  FL70 ?pin "[] o26"    (bottom side, near ANT, bigger white chip)
  B70  6pin "[]"        (bottom side, near ANT, small white chip)

Plus, resistors and capacitors (without any markings).

Version WITHOUT sticker:

  Sticker on Case: N/A
  PCB: "19-C046-03, A-1" (top side) and "B-1" and Microchip ",\\" (bottom side)
  PCB: white stamp "3204, TW-15, RU, 94V-0"
  PCB: black stamp "23MN" or "23NH" or so (smeared)
  U1 32pin "Freescale 13190, 4FGD"      (top side)
  U2 48pin "Freescale CT3000, XAB0425"  (bottom side) ;CT3000 (not CT3001)
  X3  2pin "9.5SKSS4GT"                 (top side)

Further components... top side (A-1)

  D1   5pin "D6F, 31"   (top side, below X3)
  U71  6pin "P3, () 2"  (top side, right of X3, tiny black chip)
  B71  6pin "[]"        (top side, right of X3, small white chip)
  ANT  2pin on-board copper wings
  Q70  3pin             (top side, above CN1)
  D?   2pin "72"        (top side, above CN1)
  D3   2pin "F2"        (top side, above CN1)
  U200 4pin "MSV"       (top side, above CN1)
  U202 5pin "LXKH"      (top side, right of CN1)
  U203 4pin "M6H"       (top side, right of CN1)
  CN1  6pin connector to GBA link port (top side)

Further components... bottom side (B-1)

  U201 5pin "LXV2"      (bottom side, near CN1)
  U70  6pin "AAG"       (bottom side, near ANT, tiny black chip)
  FL70 ?pin "[] o26"    (bottom side, near ANT, bigger white chip)
  B70  6pin "[]"        (bottom side, near ANT, small white chip)

Plus, resistors and capacitors (without any markings).

Major Differences

  Sticker      "N/A"                     vs "Grossosteim P/AGB-A-WA-EUR-2 E3"
  PCB-markings "19-C046-03, A-1, 3204"   vs "19-C046-04, A-7, 3104"
  U1           "CT3000, XAB0425"         vs "CT3001, XAC0445"
  Transistors  One transistor (Q70)      vs Two transistors (both nameless)
  U70/U72      U70 "AAG" (6pin)          vs U72 "BMs" (4pin)

Purpose of the changes is unknown (either older/newer revisions, or different regions with different FCC regulations).

GBA Infrared Communication

Early GBA prototypes have been intended to include a built-in IR port for sending and receiving IR signals. Among others, this port could have been used to communicate with other GBAs, or older CGB models, or TV Remote Controls, etc.

[ THE INFRARED COMMUNICATION FEATURE IS -NOT- SUPPORTED ANYMORE ]
Anyways, the prototype specifications have been as shown below...

Keep in mind that the IR signal may be interrupted by whatever objects moved between sender and receiver - the IR port isn't recommended for programs that require realtime data exchange (such like action games).

4000136h - IR - Infrared Register (R/W)

  Bit   Expl.
  0     Transmission Data  (0=LED Off, 1=LED On)
  1     READ Enable        (0=Disable, 1=Enable)
  2     Reception Data     (0=None, 1=Signal received) (Read only)
  3     AMP Operation      (0=Off, 1=On)
  4     IRQ Enable Flag    (0=Disable, 1=Enable)
  5-15  Not used

When IRQ is enabled, an interrupt is requested if the incoming signal was 0.119us Off (2 cycles), followed by 0.536us On (9 cycles) - minimum timing periods each.

Transmission Notes
When transmitting an IR signal, note that it'd be not a good idea to keep the LED turned On for a very long period (such like sending a 1 second synchronization pulse). The recipient's circuit would treat such a long signal as "normal IR pollution which is in the air" after a while, and thus ignore the signal.

Reception Notes
Received data is internally latched. Latched data may be read out by setting both READ and AMP bits.
Note: Provided that you don't want to receive your own IR signal, be sure to set Bit 0 to zero before attempting to receive data.

Power-consumption
After using the IR port, be sure to reset the register to zero in order to reduce battery power consumption.

GBA Keypad Input

The built-in GBA gamepad has 4 direction keys, and 6 buttons.

4000130h - KEYINPUT - Key Status (R)

  Bit   Expl.
  0     Button A        (0=Pressed, 1=Released)
  1     Button B        (etc.)
  2     Select          (etc.)
  3     Start           (etc.)
  4     Right           (etc.)
  5     Left            (etc.)
  6     Up              (etc.)
  7     Down            (etc.)
  8     Button R        (etc.)
  9     Button L        (etc.)
  10-15 Not used

It'd be usually recommended to read-out this register only once per frame, and to store the current state in memory. As a side effect, this method avoids problems caused by switch bounce when a key is newly released or pressed.

4000132h - KEYCNT - Key Interrupt Control (R/W)
The keypad IRQ function is intended to terminate the very-low-power Stop mode, it is not suitable for processing normal user input, to do this, most programs are invoking their keypad handlers from within VBlank IRQ.

  Bit   Expl.
  0     Button A        (0=Ignore, 1=Select)
  1     Button B        (etc.)
  2     Select          (etc.)
  3     Start           (etc.)
  4     Right           (etc.)
  5     Left            (etc.)
  6     Up              (etc.)
  7     Down            (etc.)
  8     Button R        (etc.)
  9     Button L        (etc.)
  10-13 Not used
  14    Button IRQ Enable      (0=Disable, 1=Enable)
  15    Button IRQ Condition   (0=Logical OR, 1=Logical AND)

In logical OR mode, an interrupt is requested when at least one of the selected buttons is pressed.
In logical AND mode, an interrupt is requested when ALL of the selected buttons are pressed.

Notes
In 8bit gameboy compatibility mode, L and R Buttons are used to toggle the screen size between normal 160x144 pixels and stretched 240x144 pixels.
The GBA SP is additionally having a * Button used to toggle the backlight on and off (controlled by separate hardware logic, there's no way to detect or change the current backlight state by software).

GBA Interrupt Control

4000208h - IME - Interrupt Master Enable Register (R/W)

  Bit   Expl.
  0     Disable all interrupts         (0=Disable All, 1=See IE register)
  1-31  Not used

4000200h - IE - Interrupt Enable Register (R/W)

  Bit   Expl.
  0     LCD V-Blank                    (0=Disable)
  1     LCD H-Blank                    (etc.)
  2     LCD V-Counter Match            (etc.)
  3     Timer 0 Overflow               (etc.)
  4     Timer 1 Overflow               (etc.)
  5     Timer 2 Overflow               (etc.)
  6     Timer 3 Overflow               (etc.)
  7     Serial Communication           (etc.)
  8     DMA 0                          (etc.)
  9     DMA 1                          (etc.)
  10    DMA 2                          (etc.)
  11    DMA 3                          (etc.)
  12    Keypad                         (etc.)
  13    Game Pak (external IRQ source) (etc.)
  14-15 Not used

Note that there is another 'master enable flag' directly in the CPUs Status Register (CPSR) accessible in privileged modes, see CPU reference for details.

4000202h - IF - Interrupt Request Flags / IRQ Acknowledge (R/W, see below)

  Bit   Expl.
  0     LCD V-Blank                    (1=Request Interrupt)
  1     LCD H-Blank                    (etc.)
  2     LCD V-Counter Match            (etc.)
  3     Timer 0 Overflow               (etc.)
  4     Timer 1 Overflow               (etc.)
  5     Timer 2 Overflow               (etc.)
  6     Timer 3 Overflow               (etc.)
  7     Serial Communication           (etc.)
  8     DMA 0                          (etc.)
  9     DMA 1                          (etc.)
  10    DMA 2                          (etc.)
  11    DMA 3                          (etc.)
  12    Keypad                         (etc.)
  13    Game Pak (external IRQ source) (etc.)
  14-15 Not used

Interrupts must be manually acknowledged by writing a "1" to one of the IRQ bits, the IRQ bit will then be cleared.

"[Cautions regarding clearing IME and IE]
A corresponding interrupt could occur even while a command to clear IME or each flag of the IE register is being executed. When clearing a flag of IE, you need to clear IME in advance so that mismatching of interrupt checks will not occur." ?

"[When multiple interrupts are used]
When the timing of clearing of IME and the timing of an interrupt agree, multiple interrupts will not occur during that interrupt. Therefore, set (enable) IME after saving IME during the interrupt routine." ?

BIOS Interrupt handling
Upon interrupt execution, the CPU is switched into IRQ mode, and the physical interrupt vector is called - as this address is located in BIOS ROM, the BIOS will always execute the following code before it forwards control to the user handler:

  00000018  b      128h                ;IRQ vector: jump to actual BIOS handler
  00000128  stmfd  r13!,r0-r3,r12,r14  ;save registers to SP_irq
  0000012C  mov    r0,4000000h         ;ptr+4 to 03FFFFFC (mirror of 03007FFC)
  00000130  add    r14,r15,0h          ;retadr for USER handler $+8=138h
  00000134  ldr    r15,[r0,-4h]        ;jump to [03FFFFFC] USER handler
  00000138  ldmfd  r13!,r0-r3,r12,r14  ;restore registers from SP_irq
  0000013C  subs   r15,r14,4h          ;return from IRQ (PC=LR-4, CPSR=SPSR)

As shown above, a pointer to the 32bit/ARM-code user handler must be setup in [03007FFCh]. By default, 160 bytes of memory are reserved for interrupt stack at 03007F00h-03007F9Fh.

Recommended User Interrupt handling
- If necessary switch to THUMB state manually (handler is called in ARM state)
- Determine reason(s) of interrupt by examining IF register
- User program may freely assign priority to each reason by own logic
- Process the most important reason of your choice
- User MUST manually acknowledge by writing to IF register
- If user wants to allow nested interrupts, save SPSR_irq, then enable IRQs.
- If using other registers than BIOS-pushed R0-R3, manually save R4-R11 also.
- Note that Interrupt Stack is used (which may have limited size)
- So, for memory consuming stack operations use system mode (=user stack).
- When calling subroutines in system mode, save LSR_usr also.
- Restore SPSR_irq and/or R4-R11 if you've saved them above.
- Finally, return to BIOS handler by BX LR (R14_irq) instruction.

Default memory usage at 03007FXX (and mirrored to 03FFFFXX)

  Addr.    Size Expl.
  3007FFCh 4    Pointer to user IRQ handler (32bit ARM code)
  3007FF8h 2    Interrupt Check Flag (for IntrWait/VBlankIntrWait functions)
  3007FF4h 4    Allocated Area
  3007FF0h 4    Pointer to Sound Buffer
  3007FE0h 16   Allocated Area
  3007FA0h 64   Default area for SP_svc Supervisor Stack (4 words/time)
  3007F00h 160  Default area for SP_irq Interrupt Stack (6 words/time)

Memory below 7F00h is free for User Stack and user data. The three stack pointers are initially initialized at the TOP of the respective areas:

  SP_svc=03007FE0h
  SP_irq=03007FA0h
  SP_usr=03007F00h

The user may redefine these addresses and move stacks into other locations, however, the addresses for system data at 7FE0h-7FFFh are fixed.

Not sure, is following free for user ?
Registers R8-R12_fiq, R13_fiq, R14_fiq, SPSR_fiq
Registers R13-R14_abt, SPSR_abt
Registers R13-R14_und, SPSR_und

Fast Interrupt (FIQ)
The ARM CPU provides two interrupt sources, IRQ and FIQ. In the GBA only IRQ is used. In normal GBAs, the FIQ signal is shortcut to VDD35, ie. the signal is always high, and there is no way to generate a FIQ by hardware. The registers R8..12_fiq could be used by software (when switching into FIQ mode by writing to CPSR) - however, this might make the game incompatible with hardware debuggers (which are reportedly using FIQs for debugging purposes).

GBA System Control

4000204h - WAITCNT - Waitstate Control (R/W)
This register is used to configure game pak access timings. The game pak ROM is mirrored to three address regions at 08000000h, 0A000000h, and 0C000000h, these areas are called Wait State 0-2. Different access timings may be assigned to each area (this might be useful in case that a game pak contains several ROM chips with different access times each).

  Bit   Expl.
  0-1   SRAM Wait Control          (0..3 = 4,3,2,8 cycles)
  2-3   Wait State 0 First Access  (0..3 = 4,3,2,8 cycles)
  4     Wait State 0 Second Access (0..1 = 2,1 cycles)
  5-6   Wait State 1 First Access  (0..3 = 4,3,2,8 cycles)
  7     Wait State 1 Second Access (0..1 = 4,1 cycles; unlike above WS0)
  8-9   Wait State 2 First Access  (0..3 = 4,3,2,8 cycles)
  10    Wait State 2 Second Access (0..1 = 8,1 cycles; unlike above WS0,WS1)
  11-12 PHI Terminal Output        (0..3 = Disable, 4.19MHz, 8.38MHz, 16.78MHz)
  13    Not used
  14    Game Pak Prefetch Buffer (Pipe) (0=Disable, 1=Enable)
  15    Game Pak Type Flag  (Read Only) (0=GBA, 1=CGB) (IN35 signal)
  16-31 Not used

At startup, the default setting is 0000h. Currently manufactured cartridges are using the following settings: WS0/ROM=3,1 clks; SRAM=8 clks; WS2/EEPROM: 8,8 clks; prefetch enabled; that is, WAITCNT=4317h, for more info see "GBA Cartridges" chapter.

First Access (Non-sequential) and Second Access (Sequential) define the waitstates for N and S cycles, the actual access time is 1 clock cycle PLUS the number of waitstates.
GamePak uses 16bit data bus, so that a 32bit access is split into TWO 16bit accesses (of which, the second fragment is always sequential, even if the first fragment was non-sequential).

GBA GamePak Prefetch

NOTES:
The GBA forcefully uses non-sequential timing at the beginning of each 128K-block of gamepak ROM, eg. "LDMIA [801fff8h],r0-r7" will have non-sequential timing at 8020000h.
The PHI Terminal output (PHI Pin of Gamepak Bus) should be disabled.

4000300h - POSTFLG - BYTE - Undocumented - Post Boot / Debug Control (R/W)
After initial reset, the GBA BIOS initializes the register to 01h, and any further execution of the Reset vector (00000000h) will pass control to the Debug vector (0000001Ch) when sensing the register to be still set to 01h.

  Bit   Expl.
  0     Undocumented. First Boot Flag  (0=First, 1=Further)
  1-7   Undocumented. Not used.

Normally the debug handler rejects control unless it detects Debug flags in cartridge header, in that case it may redirect to a cut-down boot procedure (bypassing Nintendo logo and boot delays, much like nocash burst boot for multiboot software). I am not sure if it is possible to reset the GBA externally without automatically resetting register 300h though.

4000301h - HALTCNT - BYTE - Undocumented - Low Power Mode Control (W)
Writing to this register switches the GBA into battery saving mode.
In Halt mode, the CPU is paused as long as (IE AND IF)=0, this should be used to reduce power-consumption during periods when the CPU is waiting for interrupt events.
In Stop mode, most of the hardware including sound and video are paused, this very-low-power mode could be used much like a screensaver.

  Bit   Expl.
  0-6   Undocumented. Not used.
  7     Undocumented. Power Down Mode  (0=Halt, 1=Stop)

The current GBA BIOS addresses only the upper eight bits of this register (by writing 00h or 80h to address 04000301h), however, as the register isn't officially documented, some or all of the bits might have different meanings in future GBA models.
For best forwards compatibility, it'd generally be more recommended to use the BIOS Functions SWI 2 (Halt) or SWI 3 (Stop) rather than writing to this register directly.

4000410h - Undocumented - Purpose Unknown ? 8bit (W)
The BIOS writes the 8bit value 0FFh to this address. Purpose Unknown.
Probably just another bug in the BIOS.

4000800h - 32bit - Undocumented - Internal Memory Control (R/W)
Supported by GBA and GBA SP only - NOT supported by DS (even in GBA mode).
Also supported by GBA Micro - but crashes on "overclocked" WRAM setting.
Initialized to 0D000020h (by hardware). Unlike all other I/O registers, this register is mirrored across the whole I/O area (in increments of 64K, ie. at 4000800h, 4010800h, 4020800h, ..., 4FF0800h)

  Bit   Expl.
  0     Disable 32K+256K WRAM (0=Normal, 1=Disable) (when off: empty/prefetch)
         From endrift: bit0 swaps 00000000h-01FFFFFFh and 02000000h-03FFFFFFh
         in GBA mode (but keeps BIOS protection)
  1     Unknown          (Read/Write-able)
  2     Unknown          (Read/Write-able)
  3     Unknown, CGB?    (Read/Write-able)
         From shinyquagsire23: bit3 seems to disable the CGB bootrom (carts
         without SRAM will typically boot with Nintendo logo skipped, and
         carts with SRAM will typically crash somehow)
  4     Unused (0)
  5     Enable 256K WRAM (0=Disable, 1=Normal) (when off: mirror of 32K WRAM)
  6-23  Unused (0)
  24-27 Wait Control WRAM 256K (0-14 = 15..1 Waitstates, 15=Lockup)
  28-31 Unknown          (Read/Write-able)

The default value 0Dh in Bits 24-27 selects 2 waitstates for 256K WRAM (ie. 3/3/6 cycles 8/16/32bit accesses). The fastest possible setting would be 0Eh (1 waitstate, 2/2/4 cycles for 8/16/32bit), that works on GBA and GBA SP only, the GBA Micro locks up with that setting (it's on-chip RAM is too slow, and works only with 2 or more waitstates).

Note: One cycle equals approx. 59.59ns (ie. 16.78MHz clock).

GBA GamePak Prefetch

GamePak Prefetch can be enabled in WAITCNT register. When prefetch buffer is enabled, the GBA attempts to read opcodes from Game Pak ROM during periods when the CPU is not using the bus (if any). Memory access is then performed with 0 Waits if the CPU requests data which is already stored in the buffer. The prefetch buffer stores up to eight 16bit values.

GamePak ROM Opcodes
The prefetch feature works only with <opcodes> fetched from GamePak ROM. Opcodes executed in RAM or BIOS are not affected by the prefetch feature (even if that opcodes read <data> from GamePak ROM).

Prefetch Enable
For GamePak ROM opcodes, prefetch may occur in two situations:

  1) opcodes with internal cycles (I) which do not change R15, shift/rotate
     register-by-register, load opcodes (ldr,ldm,pop,swp), multiply opcodes
  2) opcodes that load/store memory (ldr,str,ldm,stm,etc.)

Prefetch Disable Bug
When Prefetch is disabled, the Prefetch Disable Bug will occur for all

  "Opcodes in GamePak ROM with Internal Cycles which do not change R15"

for those opcodes, the bug changes the opcode fetch time from 1S to 1N.
Note: Affected opcodes (with I cycles) are: Shift/rotate register-by-register opcodes, multiply opcodes, and load opcodes (ldr,ldm,pop,swp).

GBA Cartridges

ROM
GBA Cartridge Header
GBA Cartridge ROM

Backup Media
Aside from ROM, cartridges may also include one of the following backup medias, used to store game positions, highscore tables, options, or other data.
GBA Cart Backup IDs
GBA Cart Backup SRAM/FRAM
GBA Cart Backup EEPROM
GBA Cart Backup Flash ROM
GBA Cart Backup DACS

Add-Ons
GBA Cart I/O Port (GPIO)
GBA Cart Real-Time Clock (RTC)
GBA Cart Solar Sensor
GBA Cart Tilt Sensor
GBA Cart Gyro Sensor
GBA Cart Rumble
GBA Cart e-Reader
GBA Cart Unknown Devices
GBA Cart Protections

Other Accessoires
GBA Flashcards
GBA Cheat Devices

GBA Cartridge Header

The first 192 bytes at 8000000h-80000BFh in ROM are used as cartridge header. The same header is also used for Multiboot images at 2000000h-20000BFh (plus some additional multiboot entries at 20000C0h and up).

Header Overview

  Address Bytes Expl.
  000h    4     ROM Entry Point  (32bit ARM branch opcode, eg. "B rom_start")
  004h    156   Nintendo Logo    (compressed bitmap, required!)
  0A0h    12    Game Title       (uppercase ascii, max 12 characters)
  0ACh    4     Game Code        (uppercase ascii, 4 characters)
  0B0h    2     Maker Code       (uppercase ascii, 2 characters)
  0B2h    1     Fixed value      (must be 96h, required!)
  0B3h    1     Main unit code   (00h for current GBA models)
  0B4h    1     Device type      (usually 00h) (bit7=DACS/debug related)
  0B5h    7     Reserved Area    (should be zero filled)
  0BCh    1     Software version (usually 00h)
  0BDh    1     Complement check (header checksum, required!)
  0BEh    2     Reserved Area    (should be zero filled)
  --- Additional Multiboot Header Entries ---
  0C0h    4     RAM Entry Point  (32bit ARM branch opcode, eg. "B ram_start")
  0C4h    1     Boot mode        (init as 00h - BIOS overwrites this value!)
  0C5h    1     Slave ID Number  (init as 00h - BIOS overwrites this value!)
  0C6h    26    Not used         (seems to be unused)
  0E0h    4     JOYBUS Entry Pt. (32bit ARM branch opcode, eg. "B joy_start")

Note: With all entry points, the CPU is initially set into system mode.

000h - Entry Point, 4 Bytes
Space for a single 32bit ARM opcode that redirects to the actual startaddress of the cartridge, this should be usually a "B <start>" instruction.
Note: This entry is ignored by Multiboot slave GBAs (in fact, the entry is then overwritten and redirected to a separate Multiboot Entry Point, as described below).

004h..09Fh - Nintendo Logo, 156 Bytes
Contains the Nintendo logo which is displayed during the boot procedure. Cartridge won't work if this data is missing or modified.
In detail: This area contains Huffman compression data (but excluding the compression header which is hardcoded in the BIOS, so that it'd be probably not possible to hack the GBA by producing de-compression buffer overflows).
A copy of the compression data is stored in the BIOS, the GBA will compare this data and lock-up itself if the BIOS data isn't exactly the same as in the cartridge (or multiboot header). The only exception are the two entries below which are allowed to have variable settings in some bits.

09Ch Bit 2,7 - Debugging Enable
This is part of the above Nintendo Logo area, and must be commonly set to 21h, however, Bit 2 and Bit 7 may be set to other values.
When both bits are set (ie. A5h), the FIQ/Undefined Instruction handler in the BIOS becomes unlocked, the handler then forwards these exceptions to the user handler in cartridge ROM (entry point defined in 80000B4h, see below).
Other bit combinations currently do not seem to have special functions.

09Eh Bit 0,1 - Cartridge Key Number MSBs
This is part of the above Nintendo Logo area, and must be commonly set to F8h, however, Bit 0-1 may be set to other values.
During startup, the BIOS performs some dummy-reads from a stream of pre-defined addresses, even though these reads seem to be meaningless, they might be intended to unlock a read-protection inside of commercial cartridge. There are 16 pre-defined address streams - selected by a 4bit key number - of which the upper two bits are gained from 800009Eh Bit 0-1, and the lower two bits from a checksum across header bytes 09Dh..0B7h (bytewise XORed, divided by 40h).

0A0h - Game Title, Uppercase Ascii, max 12 characters
Space for the game title, padded with 00h (if less than 12 chars).

0ACh - Game Code, Uppercase Ascii, 4 characters
This is the same code as the AGB-UTTD code which is printed on the package and sticker on (commercial) cartridges (excluding the leading "AGB-" part).

  U  Unique Code          (usually "A" or "B" or special meaning)
  TT Short Title          (eg. "PM" for Pac Man)
  D  Destination/Language (usually "J" or "E" or "P" or specific language)

The first character (U) is usually "A" or "B", in detail:

  A  Normal game; Older titles (mainly 2001..2003)
  B  Normal game; Newer titles (2003..)
  C  Normal game; Not used yet, but might be used for even newer titles
  F  Famicom/Classic NES Series (software emulated NES games)
  K  Yoshi and Koro Koro Puzzle (acceleration sensor)
  P  e-Reader (dot-code scanner) (or NDS PassMe image when gamecode="PASS")
  R  Warioware Twisted (cartridge with rumble and z-axis gyro sensor)
  U  Boktai 1 and 2 (cartridge with RTC and solar sensor)
  V  Drill Dozer (cartridge with rumble)

The second/third characters (TT) are:

  Usually an abbreviation of the game title (eg. "PM" for "Pac Man") (unless
  that gamecode was already used for another game, then TT is just random)

The fourth character (D) indicates Destination/Language:

  J  Japan             P  Europe/Elsewhere   F  French          S  Spanish
  E  USA/English       D  German             I  Italian

0B0h - Maker code, Uppercase Ascii, 2 characters
Identifies the (commercial) developer. For example, "01"=Nintendo.

0B2h - Fixed value, 1 Byte
Must be 96h.

0B3h - Main unit code, 1 Byte
Identifies the required hardware. Should be 00h for current GBA models.

0B4h - Device type, 1 Byte
Normally, this entry should be zero. With Nintendo's hardware debugger Bit 7 identifies the debugging handlers entry point and size of DACS (Debugging And Communication System) memory: Bit7=0: 9FFC000h/8MBIT DACS, Bit7=1: 9FE2000h/1MBIT DACS. The debugging handler can be enabled in 800009Ch (see above), normal cartridges do not have any memory (nor any mirrors) at these addresses though.

0B5h - Reserved Area, 7 Bytes
Reserved, zero filled.

0BCh - Software version number
Version number of the game. Usually zero.

0BDh - Complement check, 1 Byte
Header checksum, cartridge won't work if incorrect. Calculate as such:
chk=0:for i=0A0h to 0BCh:chk=chk-[i]:next:chk=(chk-19h) and 0FFh

0BEh - Reserved Area, 2 Bytes
Reserved, zero filled.

Below required for Multiboot/slave programs only. For Multiboot, the above 192 bytes are required to be transferred as header-block (loaded to 2000000h-20000BFh), and some additional header-information must be located at the beginning of the actual program/data-block (loaded to 20000C0h and up). This extended header consists of Multiboot Entry point(s) which must be set up correctly, and of two reserved bytes which are overwritten by the boot procedure:

0C0h - Normal/Multiplay mode Entry Point
This entry is used only if the GBA has been booted by using Normal or Multiplay transfer mode (but not by Joybus mode).
Typically deposit a ARM-32bit "B <start>" branch opcode at this location, which is pointing to your actual initialization procedure.

0C4h (BYTE) - Boot mode
The slave GBA download procedure overwrites this byte by a value which is indicating the used multiboot transfer mode.

  Value  Expl.
  01h    Joybus mode
  02h    Normal mode
  03h    Multiplay mode

Typically set this byte to zero by inserting DCB 00h in your source.
Be sure that your uploaded program does not contain important program code or data at this location, or at the ID-byte location below.

0C5h (BYTE) - Slave ID Number
If the GBA has been booted in Normal or Multiplay mode, this byte becomes overwritten by the slave ID number of the local GBA (that'd be always 01h for normal mode).

  Value  Expl.
  01h    Slave #1
  02h    Slave #2
  03h    Slave #3

Typically set this byte to zero by inserting DCB 00h in your source.
When booted in Joybus mode, the value is NOT changed and remains the same as uploaded from the master GBA.

0C6h..0DFh - Not used
Appears to be unused.

0E0h - Joybus mode Entry Point
If the GBA has been booted by using Joybus transfer mode, then the entry point is located at this address rather than at 20000C0h. Either put your initialization procedure directly at this address, or redirect to the actual boot procedure by depositing a "B <start>" opcode here (either one using 32bit ARM code). Or, if you are not intending to support joybus mode (which is probably rarely used), ignore this entry.

GBA Cartridge ROM

ROM Size
The games F-ZERO and Super Mario Advance use ROMs of 4 MBytes each. Zelda uses 8 MBytes. Not sure if other sizes are manufactured.

ROM Waitstates
The GBA starts the cartridge with 4,2 waitstates (N,S) and prefetch disabled. The program may change these settings by writing to WAITCNT, the games F-ZERO and Super Mario Advance use 3,1 waitstates (N,S) each, with prefetch enabled.
Third-party flashcards are reportedly running unstable with these settings. Also, prefetch and shorter waitstates are allowing to read more data and opcodes from ROM is less time, the downside is that it increases the power consumption.

ROM Chip
Because of how 24bit addresses are squeezed through the Gampak bus, the cartridge must include a circuit that latches the lower 16 address bits on non-sequential access, and that increments these bits on sequential access. Nintendo includes this circuit directly in the ROM chip.
Also, the ROM must have 16bit data bus (or a circuit which converts two 8bit data units into one 16bit unit - by not exceeding the waitstate timings).

GBA Cart Backup IDs

Nintendo didn't include a backup-type entry in the ROM header, however, the required type can be detected by ID strings in the ROM-image. Nintendo's tools are automatically inserting these strings (as part of their library headers). When using other tools, you may insert ID strings by hand.

ID Strings
The ID string must be located at a word-aligned memory location, the string length should be a multiple of 4 bytes (padded with zero's).

  EEPROM_Vnnn    EEPROM 512 bytes or 8 Kbytes (4Kbit or 64Kbit)
  SRAM_Vnnn      SRAM 32 Kbytes (256Kbit)
  FLASH_Vnnn     FLASH 64 Kbytes (512Kbit) (ID used in older files)
  FLASH512_Vnnn  FLASH 64 Kbytes (512Kbit) (ID used in newer files)
  FLASH1M_Vnnn   FLASH 128 Kbytes (1Mbit)

For Nintendo's tools, "nnn" is a 3-digit library version number. When using other tools, best keep it set to "nnn" rather than inserting numeric digits.

Notes
No$gba does auto-detect most backup types, even without ID strings, except for 128K FLASH (without ID "FLASH1M_Vnnn", the FLASH size defaults to 64K). Ideally, for faster detection, the ID should be put into the first some bytes of the ROM-image (ie. somewhere right after the ROM header).

GBA Cart Backup SRAM/FRAM

SRAM - 32 KBytes (256Kbit) Lifetime: Depends on back-up battery
FRAM - 32 KBytes (256Kbit) Lifetime: 10,000,000,000 read/write per bit

Hyundai GM76V256CLLFW10 SRAM (Static RAM) (eg. F-Zero)
Fujitsu MB85R256 FRAM (Ferroelectric RAM) (eg. Warioware Twisted)

Addressing and Waitstates
SRAM/FRAM is mapped to E000000h-E007FFFh, it should be accessed with 8 waitstates (write a value of 3 into Bit0-1 of WAITCNT).

Databus Width
The SRAM/FRAM databus is restricted to 8 bits, it should be accessed by LDRB, LDRSB, and STRB opcodes only.

Reading and Writing
Reading from SRAM/FRAM should be performed by code executed in WRAM only (but not by code executed in ROM). There is no such restriction for writing.

Preventing Data Loss
The GBA SRAM/FRAM carts do not include a write-protect function (unlike older 8bit gameboy carts). This seems to be a problem and may cause data loss when a cartridge is removed or inserted while the GBA is still turned on. As far as I understand, this is not so much a hardware problem, but rather a software problem, ie. theoretically you could remove/insert the cartridge as many times as you want, but you should take care that your program does not crash (and write blindly into memory).

Recommended Workaround
Enable the Gamepak Interrupt (it'll most likely get triggered when removing the cartridge), and hang-up the GBA in an endless loop when your interrupt handler senses a Gamepak IRQ. For obvious reason, your interrupt handler should be located in WRAM, ie. not in the (removed) ROM cartridge. The handler should process Gamepak IRQs at highest priority. Periods during which interrupts are disabled should be kept as short as possible, if necessary allow nested interrupts.

When to use the above Workaround
A program that relies wholly on code and data in WRAM, and that does not crash even when ROM is removed, may keep operating without having to use the above mechanism.
Do NOT use the workaround for programs that run without a cartridge inserted (ie. single gamepak/multiboot slaves), or for programs that use Gamepak IRQ/DMA for other purposes.
All other programs should use it. It'd be eventually a good idea to include it even in programs that do not use SRAM/FRAM themselves (eg. otherwise removing a SRAM/FRAM-less cartridge may lock up the GBA, and may cause it to destroy backup data when inserting a SRAM/FRAM cartridge).

SRAM vs FRAM
FRAM (Ferroelectric RAM) is a newer technology, used in newer GBA carts, unlike SRAM (Static RAM), it doesn't require a battery to hold the data. At software side, it is accessed exactly like SRAM, ie. unlike EEPROM/FLASH, it doesn't require any Write/Erase commands/delays.

Note
In SRAM/FRAM cartridges, the /REQ pin (Pin 31 of Gamepak bus) should be a little bit shorter as than the other pins; when removing the cartridge, this causes the gamepak IRQ signal to get triggered before the other pins are disconnected.

GBA Cart Backup EEPROM

9853 - EEPROM 512 Bytes (0200h) (4Kbit) (eg. used by Super Mario Advance)
9854 - EEPROM 8 KBytes (2000h) (64Kbit) (eg. used by Boktai)
Lifetime: 100,000 writes per address

Addressing and Waitstates
The eeprom is connected to Bit0 of the data bus, and to the upper 1 bit (or upper 17 bits in case of large 32MB ROM) of the cartridge ROM address bus, communication with the chip takes place serially.
The eeprom must be used with 8 waitstates (set WAITCNT=X3XXh; 8,8 clks in WS2 area), the eeprom can be then addressed at DFFFF00h..DFFFFFFh.
Respectively, with eeprom, ROM is restricted to 8000000h-9FFFeFFh (max. 1FFFF00h bytes = 32MB minus 256 bytes). On carts with 16MB or smaller ROM, eeprom can be alternately accessed anywhere at D000000h-DFFFFFFh.

Data and Address Width
Data can be read from (or written to) the EEPROM in units of 64bits (8 bytes). Writing automatically erases the old 64bits of data. Addressing works in units of 64bits respectively, that is, for 512 Bytes EEPROMS: an address range of 0-3Fh, 6bit bus width; and for 8KByte EEPROMs: a range of 0-3FFh, 14bit bus width (only the lower 10 address bits are used, upper 4 bits should be zero).

Set Address (For Reading)
Prepare the following bitstream in memory:

  2 bits "11" (Read Request)
  n bits eeprom address (MSB first, 6 or 14 bits, depending on EEPROM)
  1 bit "0"

Then transfer the stream to eeprom by using DMA.

Read Data
Read a stream of 68 bits from EEPROM by using DMA,
then decipher the received data as follows:

  4 bits - ignore these
 64 bits - data (conventionally MSB first)

Write Data to Address
Prepare the following bitstream in memory, then transfer the stream to eeprom by using DMA, it'll take ca. 108368 clock cycles (ca. 6.5ms) until the old data is erased and new data is programmed.

  2 bits "10" (Write Request)
  n bits eeprom address (MSB first, 6 or 14 bits, depending on EEPROM)
 64 bits data (conventionally MSB first)
  1 bit "0"

After the DMA, keep reading from the chip, by normal LDRH [DFFFF00h], until Bit 0 of the returned data becomes "1" (Ready). To prevent your program from locking up in case of malfunction, generate a timeout if the chip does not reply after 10ms or longer.

Using DMA
Transferring a bitstreams to/from the EEPROM must be done via DMA3 (manual transfers via LDRH/STRH won't work; probably because they don't keep /CS=LOW and A23=HIGH throughout the transfer).
For using DMA, a buffer in memory must be used (that buffer would be typically allocated temporarily on stack, one halfword for each bit, bit1-15 of the halfwords are don't care, only bit0 is used).
The buffer must be transfered as a whole to/from EEPROM by using DMA3 (DMA0-2 can't access external memory), use 16bit transfer mode, both source and destination address incrementing (ie. DMA3CNT=80000000h+length).
DMA channels of higher priority should be disabled during the transfer (ie. H/V-Blank or Sound FIFO DMAs). And, of course any interrupts that might mess with DMA registers should be disabled.

Pin-Outs
The EEPROM chips are having only 8 pins, these are connected, Pin 1..8, to ROMCS, RD, WR, AD0, GND, GND, A23, VDD of the GamePak bus. Carts with 32MB ROM must have A7..A22 logically ANDed with A23.

Notes
There seems to be no autodection mechanism, so that a hardcoded bus width must be used.

GBA Cart Backup Flash ROM

64 KBytes - 512Kbits Flash ROM - Lifetime: 10,000 writes per sector
128 KBytes - 1Mbit Flash ROM - Lifetime: ??? writes per sector

Chip Identification (all device types)

  [E005555h]=AAh, [E002AAAh]=55h, [E005555h]=90h  (enter ID mode)
  dev=[E000001h], man=[E000000h]                  (get device & manufacturer)
  [E005555h]=AAh, [E002AAAh]=55h, [E005555h]=F0h  (terminate ID mode)

Used to detect the type (and presence) of FLASH chips. See Device Types below.

Reading Data Bytes (all device types)

  dat=[E00xxxxh]                                  (read byte from address xxxx)

Erase Entire Chip (all device types)

  [E005555h]=AAh, [E002AAAh]=55h, [E005555h]=80h  (erase command)
  [E005555h]=AAh, [E002AAAh]=55h, [E005555h]=10h  (erase entire chip)
  wait until [E000000h]=FFh (or timeout)

Erases all memory in chip, erased memory is FFh-filled.

Erase 4Kbyte Sector (all device types, except Atmel)

  [E005555h]=AAh, [E002AAAh]=55h, [E005555h]=80h  (erase command)
  [E005555h]=AAh, [E002AAAh]=55h, [E00n000h]=30h  (erase sector n)
  wait until [E00n000h]=FFh (or timeout)

Erases memory at E00n000h..E00nFFFh, erased memory is FFh-filled.

Erase-and-Write 128 Bytes Sector (only Atmel devices)

  old=IME, IME=0                                  (disable interrupts)
  [E005555h]=AAh, [E002AAAh]=55h, [E005555h]=A0h  (erase/write sector command)
  [E00xxxxh+00h..7Fh]=dat[00h..7Fh]               (write 128 bytes)
  IME=old                                         (restore old IME state)
  wait until [E00xxxxh+7Fh]=dat[7Fh] (or timeout)

Interrupts (and DMAs) should be disabled during command/write phase. Target address must be a multiple of 80h.

Write Single Data Byte (all device types, except Atmel)

  [E005555h]=AAh, [E002AAAh]=55h, [E005555h]=A0h  (write byte command)
  [E00xxxxh]=dat                                  (write byte to address xxxx)
  wait until [E00xxxxh]=dat (or timeout)

The target memory location must have been previously erased.

Terminate Command after Timeout (only Macronix devices, ID=1CC2h)

  [E005555h]=F0h                            (force end of write/erase command)

Use if timeout occurred during "wait until" periods, for Macronix devices only.

Bank Switching (devices bigger than 64K only)

  [E005555h]=AAh, [E002AAAh]=55h, [E005555h]=B0h  (select bank command)
  [E000000h]=bnk                                  (write bank number 0..1)

Specifies 64K bank number for read/write/erase operations.
Required because gamepak flash/sram addressbus is limited to 16bit width.

Device Types
Nintendo puts different FLASH chips in commercial game cartridges. Developers should thus detect & support all chip types. For Atmel chips it'd be recommended to simulate 4K sectors by software, though reportedly Nintendo doesn't use Atmel chips in newer games anymore. Also mind that different timings should not disturb compatibility and performance.

  ID     Name       Size  Sectors  AverageTimings  Timeouts/ms   Waits
  D4BFh  SST        64K   16x4K    20us?,?,?       10,  40, 200  3,2
  1CC2h  Macronix   64K   16x4K    ?,?,?           10,2000,2000  8,3
  1B32h  Panasonic  64K   16x4K    ?,?,?           10, 500, 500  4,2
  3D1Fh  Atmel      64K   512x128  ?,?,?           ...40..,  40  8,8
  1362h  Sanyo      128K  ?        ?,?,?           ?    ?    ?    ?
  09C2h  Macronix   128K  ?        ?,?,?           ?    ?    ?    ?

Identification Codes MSB=Device Type, LSB=Manufacturer.
Size in bytes, and numbers of sectors * sector size in bytes.
Average medium Write, Erase Sector, Erase Chips timings are unknown?
Timeouts in milliseconds for Write, Erase Sector, Erase Chips.
Waitstates for Writes, and Reads in clock cycles.

Accessing FLASH Memory
FLASH memory is located in the "SRAM" area at E000000h..E00FFFFh, which is restricted to 16bit address and 8bit data buswidths. Respectively, the memory can be accessed <only> by 8bit read/write LDRB/STRB opcodes.
Also, reading anything (data or status/busy information) can be done <only> by opcodes executed in WRAM (not from opcodes in ROM) (there's no such restriction for writing).

FLASH Waitstates
Use 8 clk waitstates for initial detection (WAITCNT Bits 0,1 both set). After detection of certain device types smaller wait values may be used for write/erase, and even smaller wait values for raw reading, see Device Types table.
In practice, games seem to use smaller values only for write/erase (even though those operations are slow anyways), whilst raw reads are always done at 8 clk waits (even though reads could actually benefit slightly from smaller wait values).

Verify Write/Erase and Retry
Even though device signalizes the completion of write/erase operations, it'd be recommended to read/confirm the content of the changed memory area by software. In practice, Nintendo's "erase-write-verify-retry" function typically repeats the operation up to three times in case of errors.
Also, for SST devices only, the "erase-write" and "erase-write-verify-retry" functions repeat the erase command up to 80 times, additionally followed by one further erase command if no retries were needed, otherwise followed by six further erase commands.

Note
FLASH (64Kbytes) is used by the game Sonic Advance, and possibly others.

GBA Cart Backup DACS

128 KBytes - 1Mbit DACS - Lifetime: 100,000 writes.
1024 KBytes - 8Mbit DACS - Lifetime: 100,000 writes.

DACS (Debugging And Communication System) is used in Nintendo's hardware debugger only, DACS is NOT used in normal game cartridges.

Parts of DACS memory is used to store the debugging exception handlers (entry point/size defined in cartridge header), the remaining memory could be used to store game positions or other data. The address space is the upper end of the 32MB ROM area, the memory can be read directly by the CPU, including for ability to execute program code in this area.

GBA Cart I/O Port (GPIO)

4bit General Purpose I/O Port (GPIO) - contained in the ROM-chip

Used by Boktai for RTC and Solar Sensor:
GBA Cart Real-Time Clock (RTC)
GBA Cart Solar Sensor
And by Warioware Twisted for Rumble and Z-Axis Sensor:
GBA Cart Rumble
GBA Cart Gyro Sensor
Might be also used by other games for other purposes, such like other sensors, or SRAM bank switching, etc.

The I/O registers are mapped to a 6-byte region in the ROM-area at 80000C4h, the 6-byte region should be zero-filled in the ROM-image. In Boktai, the size of the zero-filled region is 0E0h bytes - that probably due to an incorrect definition (the additional bytes do not contain any extra ports, nor mirrors of the ports in the 6-byte region). Observe that ROM-bus writes are limited to 16bit/32bit access (STRB opcodes are ignored; that, only in DS mode?).

80000C4h - I/O Port Data (selectable W or R/W)

  bit0-3  Data Bits 0..3 (0=Low, 1=High)
  bit4-15 not used (0)

80000C6h - I/O Port Direction (for above Data Port) (selectable W or R/W)

  bit0-3  Direction for Data Port Bits 0..3 (0=In, 1=Out)
  bit4-15 not used (0)

80000C8h - I/O Port Control (selectable W or R/W)

  bit0    Register 80000C4h..80000C8h Control (0=Write-Only, 1=Read/Write)
  bit1-15 not used (0)

In write-only mode, reads return 00h (or possible other data, if the rom contains non-zero data at that location).

Connection Examples

  GPIO       | Boktai  | Wario
  Bit Pin    | RTC SOL | GYR RBL
  -----------+---------+---------
  0   ROM.1  | SCK CLK | RES -
  1   ROM.2  | SIO RST | CLK -
  2   ROM.21 | CS  -   | DTA -
  3   ROM.22 | -   FLG | -   MOT
  -----------+---------+---------
  IRQ ROM.43 | IRQ -   | -   -

Aside from the I/O Port, the ROM-chip also includes an inverter (used for inverting the RTC /IRQ signal), and some sort of an (unused) address decoder output (which appears to be equal or related to A23 signal) (ie. reacting on ROM A23, or SRAM D7, which share the same pin on GBA slot).

GBA Cart Real-Time Clock (RTC)

S3511 - 8pin RTC with 3-wire serial bus (used in Boktai)

The RTC chip is (almost) the same as used in NDS consoles:
DS Real-Time Clock (RTC)
The chip is accessed via 4bit I/O port (only 3bits are used for RTC):
GBA Cart I/O Port (GPIO)

Comparision of RTC Registers

  NDS_________GBA_________GBA/Params___
  stat2       control     (1-byte)
  datetime    datetime    (7-byte)
  time        time        (3-byte)
  stat1       force reset (0-byte)
  clkadjust   force irq   (0-byte)
  alarm1/int1 always FFh  (boktai contains code for writing 1-byte to it)
  alarm2      always FFh  (unused)
  free        always FFh  (unused)

Control Register

  Bit Dir Expl.
  0   -   Not used
  1   R/W IRQ duty/hold related?
  2   -   Not used
  3   R/W Per Minute IRQ (30s duty)        (0=Disable, 1=Enable)
  4   -   Not used
  5   R/W Unknown?
  6   R/W 12/24-hour Mode                  (0=12h, 1=24h) (usually 1)
  7   R   Power-Off (auto cleared on read) (0=Normal, 1=Failure, time lost)

Setting after Battery-Shortcut is 82h. Setting after Force-Reset is 00h.
Unused bits seem to be always zero, but might be read-only or write-only?

Datetime and Time Registers
Same as NDS, except AM/PM flag moved from hour.bit6 (NDS) to hour.bit7 (GBA).

Force Reset/Irq Registers
Used to reset all RTC registers (all used registers become 00h, except day/month which become 01h), or to drag the IRQ output LOW for a short moment. These registers are strobed by ANY access to them, ie. by both writing to, as well as reading from these registers.

RTC Games

  Boktai series     ;which/how many titles?
  P-Letter series   ;which/how many titles?
  Rockman EXE 4.5 Real Operation

Pin-Outs / IRQ Signal
The package has identical pin-outs as in NDS, although it is slightly larger than the miniature chip in the DS.
For whatever reason, the RTC's /IRQ output is passed through an inverter (contained in the ROM-chip), the inverted signal is then passed to the /IRQ pin on the cartridge slot. So, IRQ's will be triggered on the "wrong" edge - possible somehow in relation with detecting cartridge-removal IRQs?

GBA Cart Solar Sensor

Uses a Photo Diode as Solar Sensor (used in Boktai, allowing to defeat vampires when the cartridge is exposed to sunlight). The cartridge comes in transparent case, and it's slightly longer than normal carts, so the sensor reaches out of the cartridge slot. According to the manual, the sensor works only with sunlight, but actually it works with any strong light source (eg. a 100 Watt bulb at 1-2 centimeters distance). The sensor is accessed via 4bit I/O port (only 3bits used), which is contained in the ROM-chip.
GBA Cart I/O Port (GPIO)

A/D Conversion
The cartridge uses a self-made digital-ramp converter A/D converter, which is (maybe) better than measuring a capacitor charge-up time, and/or less expensive than a real ADC-chip:
It contains a 74LV4040 12bit binary counter (clocked by CPU via the I/O port), of which only the lower 8bit are used, which are passed to a resistor ladder-type D/A converter, which is generating a linear increasing voltage, which is passed to a TLV272 voltage comparator, which is passing a signal to the I/O port when the counter voltage becomes greater than the sensor voltage.

Example Code

  strh  0001h,[80000c8h] ;-enable R/W mode
  strh  0007h,[80000c6h] ;-init I/O direction
  strh  0002h,[80000c4h] ;-reset counter to zero (high=reset) (I/O bit0)
  strh  0000h,[80000c4h] ;-clear reset (low=normal)
  mov   r0,0             ;-initial level
 @@lop:
  strh  0001h,[80000c4h] ;-clock high ;\increase counter      (I/O bit1)
  strh  0000h,[80000c4h] ;-clock low  ;/
  ldrh  r1,[80000c4h]    ;-read port                          (I/O bit3)
  tst   r1,08h           ;\
  addeq r0,1             ; loop until voltage match (exit with r0=00h..FFh),
  tsteq r0,100h          ; or until failure/timeout (exit with r0=100h)
  beq   @@lop            ;/

The results vary depending on the clock rate used. In above example, ensure that IRQs or DMAs do not interrupt the function. Alternately, use a super-slow clock rate (eg. like 666Hz used in Boktai) so that additional small IRQ/DMA delays have little effect on the overall timing. Results should be somewhat:

  E8h  total darkness (including LED light, or daylight on rainy days)
  Dxh  close to a 100 Watt Bulb
  5xh  reaches max level in boktai's solar gauge
  00h  close to a tactical nuclear bomb dropped on your city

The exact values may change from cartridge to cartridge, so it'd be recommened to include a darkness calibration function, prompting the user to cover the sensor for a moment (in Boktai, access Options by pressing left/right in title screen) (alternately, auto-calibration could theoretically memorize the darkest in-game level ever seen).

GBA Cart Tilt Sensor

Yoshi's Universal Gravitation / Yoshi Topsy Turvy (X/Y-Axis)
Koro Koro Puzzle (probably same as Yoshi, X/Y-Axis, too) (?)

Yoshi-Type (X/Y-Axis)
All of the registers are one byte wide, mapped into the top "half" of the SRAM memory range.

  E008000h (W) Write 55h to start sampling
  E008100h (W) Write AAh to start sampling
  E008200h (R) Lower 8 bits of X axis
  E008300h (R) Upper 4 bits of X axis, and Bit7: ADC Status (0=Busy, 1=Ready)
  E008400h (R) Lower 8 bits of Y axis
  E008500h (R) Upper 4 bits of Y axis

You must set SRAM wait control to 8 clocks to access it correctly.
You must also set the cartridge PHI terminal to 4 MHz to make it work.
Sampling routine (typically executed once a frame during VBlank):

  wait until [E008300h].Bit7=1 or until timeout ;wait ready
  x = ([E008300h] AND 0Fh)*100h + [E008200h]    ;get x
  y = ([E008500h] AND 0Fh)*100h + [E008400h]    ;get y
  [E008000h]=55h, [E008100h]=AAh                ;start next conversion

Example values (may vary on different carts and on temperature, etc):

  X ranged between 0x2AF to 0x477, center at 0x392.    Huh?
  Y ranged between 0x2C3 to 0x480, center at 0x3A0.    Huh?

Thanks to Flubba for Yoshi-Type information.
Unknown if the Yoshi-Type sensors are sensing rotation, or orientation, or motion, or something else? In case of rotation, rotation around X-axis would result in motion in Y-direction, so not too sure whether X and Y have which meaning?
Most probably, the sensors are measuring (both) static acceleration (gravity), and dynamic acceleration (eg. shaking the device left/right).
The X/Y values are likely to be mirrored depending on using a back-loading cartridge slot (original GBA), or front-loading cartridge slot (newer GBA SP, and NDS, and NDS-Lite).

GBA Cart Gyro Sensor

Warioware Twisted (Z-Axis Gyro Sensor, plus Rumble)

Wario-Type (Z-Axis)
Uses a single-axis sensor, which senses rotation around the Z-axis. The sensor is connected to an analogue-in, serial-out ADC chip, which is accessed via lower 3 bits of the GPIO,
GBA Cart I/O Port (GPIO)
The four I/O Lines are connected like so,

  GPIO.Bit0 (W) Start Conversion
  GPIO.Bit1 (W) Serial Clock
  GPIO.Bit2 (R) Serial Data
  GPIO.Bit3 (W) Used for Rumble (not gyro related)

There should be at least <three sequential 32bit ARM opcodes executed in WS0 region> between the STRH opcodes which toggle the CLK signal. Wario uses WAITCNT=45B7h (SRAM=8clks, WS0/WS1/WS2=3,1clks, Prefetch=On, PHI=Off).
The data stream consists of: 4 dummy bits (usually zero), followed by 12 data bits, followed by endless unused bits (usually zero).

 read_gyro:
  mov  r1,8000000h      ;-cartridge base address
  mov  r0,01h           ;\enable R/W access
  strh r0,[r1,0c8h]     ;/
  mov  r0,0bh           ;\init direction (gpio2=input, others=output)
  strh r0,[r1,0c6h]     ;/
  ldrh r2,[r1,0c4h]     ;-get current state (for keeping gpio3=rumble)
  orr  r2,3                     ;\
  strh r2,[r1,0c4h] ;gpio0=1    ; start ADC conversion
  bic  r2,1                     ;
  strh r2,[r1,0c4h] ;gpio0=0    ;/
  mov  r0,00010000h ;stop-bit           ;\
  bic  r2,2                             ;
 @@lop:                                 ;
  ldrh r3,[r1,0c4h] ;get gpio2=data     ; read 16 bits
  strh r2,[r1,0c4h] ;gpio1=0=clk=low    ; (4 dummy bits, plus 12 data bits)
  movs r3,r3,lsr 3  ;gpio2 to cy=data   ;
  adcs r0,r0,r0     ;merge data, cy=done;
  orr  r3,r2,2      ;set bit1 and delay ;
  strh r3,[r1,0c4h] ;gpio1=1=clk=high   ;
  bcc  @@lop                            ;/
  bic  r0,0f000h                 ;-strip upper 4 dummy bits (isolate 12bit adc)
  bx   lr

Example values (may vary on different carts, battery charge, temperature, etc):

  354h  rotated in anti-clockwise direction (shock-speed)
  64Dh  rotated in anti-clockwise direction (normal fast)
  6A3h  rotated in anti-clockwise direction (slow)
  6C0h  no rotation                         (stopped)
  6DAh  rotation in clockwise direction     (slow)
  73Ah  rotation in clockwise direction     (normal fast)
  9E3h  rotation in clockwise direction     (shock-speed)

For detection, values 000h and FFFh would indicate that there's no sensor.
The Z-axis always points into same direction; no matter of frontloading or backloading cartridge slots.
Thanks to Momo Vampire for contributing a Wario cartridge.

X/Y/Z-Axes
X-Axis and Y-Axis are meant to be following the screens X and Y coordinates, so the Z-Axis would point into the screens depth direction.

DSi Cameras
DSi consoles can mis-use the built-in cameras as Gyro sensor (as done by the System Flaw DSi game).

GBA Cart Rumble

Warioware Twisted (Rumble, plus Z-Axis Gyro Sensor)
Drill Dozer (Rumble only) <-- and ALSO supports Gameboy Player rumble?

GBA Rumble Carts are containing a small motor, which is causing some vibration when/while it is switched on (that, unlike DS Rumble, which must be repeatedly toggled on/off).

In Warioware Twisted, rumble is controlled via GPIO.Bit3 (Data 0=Low=Off, 1=High=On) (and Direction 1=Output), the other GPIO Bits are used for the gyro sensor.
GBA Cart I/O Port (GPIO)
Note: GPIO3 is connected to an external pulldown resistor (so the HighZ level gets dragged to Low=Off when direction is set to Input).

Unknown if Drill Dozer is controlled via GPIO.Bit3, too?

DS Rumble Pak
Additionally, there's a Rumble Pak for the NDS, which connects to the GBA slot, so it can be used also for GBA games (provided that the game doesn't require the GBA slot, eg. GBA multiboot games).
DS Cart Rumble Pak

Gamecube Rumble
Moreover, GBA games that are running on a Gameboy Player are having access to the Rumble function of Gamecube joypads.
GBA Gameboy Player

GBA Cart e-Reader

GBA Cart e-Reader Overview
GBA Cart e-Reader I/O Ports
GBA Cart e-Reader Dotcode Format
GBA Cart e-Reader Data Format
GBA Cart e-Reader Program Code
GBA Cart e-Reader API Functions
GBA Cart e-Reader VPK Decompression
GBA Cart e-Reader Error Correction
GBA Cart e-Reader File Formats

   ________________
  |   ShortStrip   |
  |L              L|
  |o    Center    o|
  |n    Region    n|
  |g              g|
  |  may contain   |
  |S   pictures,  S|
  |t instructions t|
  |r     etc.     r|
  |i              i|
  |p              p|
  |___ShortStrip___|

GBA Cart e-Reader Overview

The e-Reader is a large GBA cartridge (about as big as the GBA console), with built-in dotcode scanning hardware. Dotcodes are tiny strips of black and white pixels printed on the edges of cardboard cards. The cards have to be pulled through a slot on the e-Reader, which is giving it a feeling like using a magnet card reader. The binary data on the dotcodes contains small games, either in native GBA code (ARM/THUMB), or in software emulated 8bit Z80 or NES/Famicom (6502) code.

The e-Reader Hardware
The hardware consists of regular 8MByte ROM and 128KByte FLASH chips, two link ports, a custom PGA chip, the camera module (with two red LEDs, used as light source), and some analogue components for generating the LED voltages, etc. The camera supports 402x302 pixels with 7bit monochrome color depth, but the PGA clips it to max 320 pixels per scanline with 1bit color depth.

Link Port Plug/Socket
The e-Reader's two link ports are simply interconnected with each other; without connection to the rest of the e-Reader hardware. These ports are used only on the original GBA (where the large e-Reader cartridge would be covering the GBA's link socket). When trying to insert the e-Reader into an original NDS (or GBA-Micro), then the e-Reader's link plug will hit against the case of the NDS, so it works only with some minor modification to the hardware. There's no such problem with GBA-SP and NDS-Lite.

Region/Version
There are 3 different e-Reader's: Japanese/Original, Japanese/Plus, and Non-Japanese. The Original version has only 64K FLASH, no Link Port, and reportedly supports only Z80 code, but no NES/GBA code. The Plus and Non-Japanese versions should be almost identical, except that they reject cards from the wrong region, and that the title strings aren't ASCII in Japan, the Plus version should be backwards compatible to the Original one.

The Problem
Nintendo's current programmers are definetly unable to squeeze a Pac-Man style game into less than 4MBytes. Their solution has been: MORE memory. That is, they've put a whopping 8MByte BIOS ROM into the e-Reader, which contains the User Interface, and software emulation for running some of their 20 years old 8bit NES and Game&Watch titles, which do fit on a few dotcode strips.

GBA Cart e-Reader I/O Ports

DF80000h Useless Register (R/W)

  0     Output to PGA.Pin93 (which seems to be not connected to anything)
  1-3   Unknown, read/write-able (not used by e-Reader BIOS)
  4-15  Always zero (0)

DFA0000h Reset Register (R/W)

  0    Always zero              (0)
  1    Reset Something?         (0=Normal, 1=Reset)
  2    Unknown, always set      (1)
  3    Unknown, read/write-able (not used by e-Reader BIOS)
  4-7  Always zero              (0)
  8    Unknown, read/write-able (not used by e-Reader BIOS)
  9-15 Always zero              (0)

DFC0000h..DFC0027h Scanline Data (R)
Scanline data (40 bytes, for 320 pixels, 1bit per pixel, 0=black, 1=white).
The first (leftmost) pixel is located in the LSB of the LAST byte.
Port E00FFB1h.Bit1 (and [4000202h].Bit13) indicates when a new scanline is present, the data should be then transferred to RAM via DMA3 (SAD=DFC0000h, DAD=buf+y*28h, CNT=80000014h; a slower non-DMA transfer method would result in missed scanlines). After the DMA, software must reset E00FFB1h.Bit1.
Note: The scanning resolution is 1000 DPI.

DFC0028h+(0..2Fh*2) Brightest Pixels of 8x6 Blocks (R)

  0-6  Max Brightness (00h..7Fh; 00h=All black, 7Fh=One or more white)
  7-15 Always zero

Can be used to adjust the Port E00FF80h..E00FFAFh settings.

DFC0088h Darkest Pixel of whole Image (R)

  0-7  Max Darkness   (00h..7Fh; 00h=One or more black, 7Fh=All white)
  8-15 Always zero

Can be used to adjust the Port E00FF80h..E00FFAFh settings.

E00FF80h..E00FFAFh Intensity Boundaries for 8x6 Blocks (R/W)
The 320x246 pixel camera input is split into 8x6 blocks (40x41 pixels each), with Block00h=Upper-right, Block07h=Upper-left, ..., Block27h=Lower-left. The boundary values for the separate blocks are used for 128-grayscale to 2-color conversion, probably done like "IF Pixel>Boundary THEN white ELSE black".

  0-6  Block Intensity Boundaries (0..7Fh; 7Fh=Whole block gets black)
  7    Always zero

The default boundary values are stored in FLASH memory, the values are typically ranging from 28h (outer edges) to 34h (center image), that in respect to the light source (the two LEDs are emitting more light to the center region).

E00FFB0h Control Register 0 (R/W)

  0    Serial Data       (Low/High)
  1    Serial Clock      (Low/High)
  2    Serial Direction  (0=Input, 1=Output)
  3    Led/Irq Enable    (0=Off, 1=On; Enable LED and Gamepak IRQ)
  4    Start Scan        (0=Off, 1=Start) (0-to-1 --> Resync line 0)
  5    Phi 16MHz Output  (0=Off, 1=On; Enable Clock for Camera, and for LED)
  6    Power 3V Enable   (0=Off, 1=On; Enable 3V Supply for Camera)
  7    Not used          (always 0) (sometimes 1) (Read only)

E00FFB1h Control Register 1 (R/W)

  0    Not used          (always 0)
  1    Scanline Flag     (1=Scanline Received, 0=Acknowledge)
  2-3  Not used          (always 0)
  4    Strange Bit       (0=Normal, 1=Force Resync/Line0 on certain interval?)
  5    LED Anode Voltage (0=3.0V, 1=5.1V; requires E00FFB0h.Bit3+5 to be set)
  6    Not used          (always 0)
  7    Input from PGA.Pin22, always high (not used by e-Reader) (Read Only)

Bit1 can be SET by hardware only, software can only RESET that bit, the Gamepak IRQ flag (Port 4000202h.Bit13) becomes set on 0-to-1 transitions.

E00FFB2h Light Source LED Kathode Duration (LSB) (R/W)
E00FFB3h Light Source LED Kathode Duration (MSB) (R/W)
Selects the LED Kathode=LOW Duration, aka the LED=ON Duration. That does act as pulse width modulated LED brightness selection (the camera seems to react slowly enough to view the light as being dimmed to medium, rather than seeing the actual light ON and OFF states). The PWM timer seems to be clocked at 8MHz. The hardware clips timer values 2000h..FFFFh to max 2000h (=1ms). Additionally, the e-Reader BIOS clips values to max 11B3h. Default setting is found in FLASH calibration data. A value of 0000h disables the LED.

Serial Port Registers (Camera Type 1) (DV488800) (calib_data[3Ch]=1)
All 16bit values are ordered MSB,LSB. All registers are whole 8bit Read/Write-able, except 00h,57h-5Ah (read only), and 53h-55h (2bit only).

  Port     Expl.               (e-Reader Setting)
  00h      Maybe Chip ID (12h) (not used by e-Reader BIOS) (Read Only)
  01h                          (05h)    ;-Bit0: 1=auto-repeat scanning?
  02h                          (0Eh)
  10h-11h  Vertical Scroll     (calib_data[30h]+7)
  12h-13h  Horizontal Scroll   (0030h)
  14h-15h  Vertical Size       (00F6h=246)
  16h-17h  Horizontal Size     (0140h=320)
  20h-21h  H-Blank Duration    (00C4h)
  22h-23h                      (0400h)  ;-Upper-Blanking in dot-clock units?
  25h                          (var)    ;-bit1: 0=enable [57h..5Ah] ?
  26h                          (var)    ;\maybe a 16bit value
  27h                          (var)    ;/
  28h                          (00h)
  30h      Brightness/contrast (calib_data[31h]+/-nn)
  31h-33h                      (014h,014h,014h)
  34h      Brightness/contrast (02h)
  50h-52h  8bit Read/Write     (not used by e-Reader BIOS)
  53h-55h  2bit Read/Write     (not used by e-Reader BIOS)
  56h      8bit Read/Write     (not used by e-Reader BIOS)
  57h-58h  16bit value, used to autodetect/adjust register[30h] (Read Only)
  59h-5Ah  16bit value, used to autodetect/adjust register[30h] (Read Only)
  80h-FFh  Mirrors of 00h..7Fh (not used by e-Reader BIOS)

All other ports are unused, writes to those ports are ignored, and reads are returning data mirrored from other ports; that is typically data from 2 or more ports, ORed together.

Serial Port Registers (Camera Type 2) (calib_data[3Ch]=2)
All 16bit values are using more conventional LSB,MSB ordering, and port numbers are arranged in a more reasonable way. The e-Reader BIOS doesn't support (or doesn't require) brightness adjustment for this camera module.

  Port     Expl.             (e-Reader Setting)
  00h                        (22h)
  01h                        (50h)
  02h-03h  Vertical Scroll   (calib_data[30h]+28h)
  04h-05h  Horizontal Scroll (001Eh)
  06h-07h  Vertical Size     (00F6h)    ;=246
  08h-09h  Horizontal Size   (0140h)    ;=320
  0Ah-0Ch                    (not used by e-Reader BIOS)
  0Dh                        (01h)
  0Eh-0Fh                    (01EAh)    ;=245*2
  10h-11h                    (00F5h)    ;=245
  12h-13h                    (20h,F0h)  ;maybe min/max values?
  14h-15h                    (31h,C0h)  ;maybe min/max values?
  16h                        (00h)
  17h-18h                    (77h,77h)
  19h-1Ch                    (30h,30h,30h,30h)
  1Dh-20h                    (80h,80h,80h,80h)
  21h-FFh                    (not used by e-Reader BIOS)

This appears to be a Micron (aka Aptina) camera (resembling the DSi cameras).
My own e-Reader uses a Type 1 camera module. Not sure if Nintendo has ever manufactured any e-Readers with Type 2 cameras?

Calibration Data in FLASH Memory (Bank 0, Sector 0Dh)

  E00D000 14h  ID String ('Card-E Reader 2001',0,0)
  E00D014 2    Sector Checksum (NOT(x+x/10000h); x=sum of all other halfwords)

Begin of actual data (40h bytes)

  E00D016 8x6  [00h] Intensity Boundaries for 8x6 blocks ;see E00FF80h..AFh
  E00D046 1    [30h] Vertical scroll (0..36h)  ;see type1.reg10h/type2.reg02h
  E00D047 1    [31h] Brightness or contrast    ;see type1.reg30h
  E00D048 2    [32h] LED Duration              ;see E00FFB2h..B3h
  E00D04A 2    [34h] Not used?   (0000h)
  E00D04C 2    [36h] Signed value, related to adjusting the 8x6 blocks
  E00D04E 4    [38h] Not used?   (00000077h)
  E00D052 4    [3Ch] Camera Type (0=none,1=DV488800,2=Whatever?)

Remaining bytes in this Sector...

  E00D056 FAAh Not used (zerofilled) (included in above checksum)

Flowchart for Overall Camera Access
ereader_scan_camera:

 call ereader_power_on
 call ereader_initialize
 for z=1 to number_of_frames
  for y=0 to 245
   Wait until E00FFB1h.Bit1 gets set by hardware (can be handled by IRQ)
   Copy 14h halfwords from DFC0000h to buf+y*28h via DMA3
   Reset E00FFB1h.Bit1 by software
  next y
  ;(could now check DFC0028h..DFC0086h/DFC0088h for adjusting E00FF00h..2Fh)
  ;(could now show image on screen, that may require to stop/pause scanning)
 next z
 call ereader_power_off
 Ret

ereader_power_on:

 [4000204h]=5803h   ;Init waitstates, and enable Phi 16MHz
 [DFA0000h].Bit1=1
 Wait(10ms)
 [E00FFB0h]=40h     ;Enable Power3V and reset other bits
 [DFA0000h].Bit1=0
 [E00FFB1h]=20h     ;Enable Power5V and reset other bits
 Wait(40ms)
 [E00FFB1h].Bit4=0  ;...should be already 0 ?
 [E00FFB0h]=40h+27h ;Phi16MHz=On, SioDtaClkDir=HighHighOut
 Ret

ereader_power_off:

 [E00FFB0h]=04h    ;Power3V=Off, Disable Everything, SioDtaClkDir=LowLowOut
 [DFA0000h].Bit1=0 ;...should be already 0
 [E00FFB1h].Bit5=0 ;Power5V=Off
 Ret

ereader_initialize:

 IF calib_data[3Ch] AND 03h = 1 THEN init_camera_type1
 [E00FFB0h].Bit4=1 ;ScanStart
 IF calib_data[3Ch] AND 03h = 2 THEN init_camera_type2
 Copy calib_data[00h..2Fh] to [E00FF80h+00h..2Fh]  ;Intensity Boundaries
 Copy calib_data[32h..33h] to [E00FFB2h+00h..01h]  ;LED Duration LSB,MSB
 [E00FFB0h].Bit3=1                                 ;LedIrqOn
 Ret

init_camera_type1:

 x=MIN(0,calib_data[31h]-0Bh)
 Set Sio Registers (as shown for Camera Type 1, except below values...)
 Set Sio Registers [30h]=x [25h]=04h, [26h]=58h, [27h]=6Ch
 ;(could now detect/adjust <x> based on Sio Registers [57h..5Ah])
 Set Sio Registers [30h]=x [25h]=06h, [26h]=E8h, [27h]=6Ch
 Ret

init_camera_type2:

 Wait(0.5ms)
 Set Sio Registers (as shown for Camera Type 2)
 Ret

Accessing Serial Registers via E00FFB0h

      Begin   Write(A) Write(B) Read(C) Read(D) End     Idle    PwrOff
  Dir ooooooo ooooooo  ooooooo  iiiiiii iiiiiii ooooooo ooooooo ooooooo
  Dta ---____ AAAAAAA  BBBBBBB  xxxxxCx xxxxxDx ______- ------- _______
  Clk ------_ ___---_  ___---_  ___---_ ___---_ ___---- ------- _______

Flowchart for accessing Serial Registers via E00FFB0h (looks like I2C bus)

 Delay:
  Wait circa 2.5us, Ret
 SioBegin:
  SioDta=1, SioDir=Out, SioClk=1, Delay, SioDta=0, Delay, SioClk=0, Ret
 SioEnd:
  SioDta=0, SioDir=Out, Delay, SioClk=1, Delay, SioDta=1, Ret
 SioRead1bit:   ;out: databit
  SioDir=In, Delay, SioClk=1, Delay, databit=SioDta, SioClk=0, Ret
 SioWrite1bit:  ;in: databit
  SioDta=databit, SioDir=Out, Delay, SioClk=1, Delay, SioClk=0, Ret
 SioReadByte:   ;in: endflag - out: data
  for i=7 to 0, data.bit<i>=SioRead1bit, next i, SioWrite1bit(endflag), Ret
 SioWriteByte:  ;in: data - out: errorflag
  for i=7 to 0, Delay(huh/why?), SioWrite1bit(data.bit<i>), next i
  errorflag=SioRead1bit, SioDir=Out(huh/why?), Ret
 SioWriteRegisters:  ;in: index, len, buffer
  SioBegin
  SioWriteByte(22h)        ;command (set_index) (and write_data)
  SioWriteByte(index)      ;index
  for i=0 to len-1
   SioWriteByte(buffer[i]) ;write data (and auto-increment index)
  next
  SioEnd
  ret
 SioReadRegisters:   ;in: index, len - out: buffer
  SioBegin
  SioWriteByte(22h)        ;command (set_index) (without any write_data here)
  SioWriteByte(index)      ;index
  SioBegin
  SioWriteByte(23h)        ;command (read_data) (using above index)
  for i=0 to len-1
   if i=len-1 then endflag=1 else endflag=0
   buffer[i]=SioReadByte(endflag)  ;read data (and auto-increment index)
  next
  SioEnd
  Ret

Caution: Accessing the SIO registers appears highly unstable, and seems to require error handling with retries. Not sure what is causing that problem, possibly the registers cannot be accessed during camera-data-scans...?

WAITCNT
The e-Reader BIOS uses WAITCNT [4000204h]=5803h when accessing the PGA, that is, gamepak 16.78MHz phi output (bit11-12=3), 8 waits for SRAM region (bit0-1=3), gamepak prefetch enabled (bit14=1), also sets WS0 to 4,2 waits (bit2-4=0), and sets WS2 to odd 4,8 waits (bit8-10=0). The WS2 (probably WS0 too) settings are nonsense, and should work with faster timings (the e-Reader can be accessed in NDS mode, which doesn't support that slow timings).

e-Reader Memory and I/O Map (with all used/unused/mirrored regions)

  C000000h-C7FFFFFh  ROM (8MB)
  C800000h-DF7FFFFh  Open Bus
  DF80000h-DF80001h  Useless Register (R/W)
  DF80002h-DF9FFFFh  Mirrors of DF80000h-DF80001h
  DFA0000h-DFA0001h  Reset Register (R/W)
  DFA0002h-DFBFFFFh  Mirrors of DFA0000h-DFA0001h
  DFC0000h-DFC0027h  Scanline Data (320 Pixels) (R)
  DFC0028h-DFC0087h  Brightest Pixels of 8x6 Blocks (R)
  DFC0088h           Darkest Pixel of whole Image (R)
  DFC0089h-DFC00FFh  Always zero
  DFC0100h-DFDFFFFh  Mirrors of DFC0000h-DFC00FFh
  DFE0000h-DFFFFFFh  Open Bus
  E000000h-E00CFFFh  FLASH Bank 0 - Data
  E00D000h-E00DFFFh  FLASH Bank 0 - Calibration Data
  E00E000h-E00EFFFh  FLASH Bank 0 - Copy of Calibration Data
  E00F000h-E00FF7Fh  FLASH Bank 0 - Unused region
  E000000h-E00EFFFh  FLASH Bank 1 - Data
  E00F000h-E00FF7Fh  FLASH Bank 1 - Unused region
  E00FF80h-E00FFAFh  Intensity Boundaries for 8x6 Blocks (R/W)
  E00FFB0h           Control Register 0 (R/W)
  E00FFB1h           Control Register 1 (R/W)
  E00FFB2h-E00FFB3h  LED Duration (16bit) (R/W)
  E00FFB4h-E00FFBFh  Always zero
  E00FFC0h-E00FFFFh  Mirror of E00FF80h-E00FFBFh

Mind that WS2 should be accessed by LDRH/STRH, and SRAM region by LDRB/STRB.
Additionally about 32 serial bus registers are contained in the camera module.

Camera Module Notes
The Type 1 initial setting on power-on is 402x302 pixels, the e-Reader uses only 320x246 pixels. The full vertical resolution could be probably used without problems. Port DFC0000h-DFC0027h are restricted to 320 pixels, so larger horizontal resolutions could be probably obtained only by changing the horizontal scroll offset on each 2nd scan.
The camera output is 128 grayscales (via parallel 7bit databus), but the PGA converts it to 2 colors (1bit depth). For still images, it might be possible to get 4 grayshades via 3 scans with different block intensity boundary settings.
No idea if the camera supports serial commands other than 22h and 23h. Namely, it <would> be a quite obvious and basic feature to allow to receive the bitmap via the 2-wire serial bus (alternately to the 7bit databus), if supported, it'd allow to get 7bit images, bypassing 1bit PGA conversion.
When used as actual camera (by cutting an opening in the case), the main problem is the 1bit color depth, which allows only black and white schemes, when/if solving that problem, focusing might be also a problem.

Either the camera or the PGA seem to have a problem on white-to-black transitions in vertical direction, the upper some black pixels are sorts of getting striped or dithered. For example, scanning the large sync marks appears as:

  Actual Shape    Scanned Shape
     XXXXX            X  X
    XXXXXXX           X  X X
   XXXXXXXXX        X X  X XX
   XXXXXXXXX        X X  X XX
    XXXXXXX          XXXXXXX
     XXXXX            XXXXX

That appears only on large black shapes (the smaller data dots look better). Probably the image is scanned from bottom upwards (and the camera senses only the initial transition at the bottom, and then looses track of what it is doing).

GBA Cart e-Reader Dotcode Format

Resolution is 342.39 DPI (almost 10 blocks per inch).
Resolution is 134.8 dots/cm (almost 4 blocks per centimeter).
The width and height of each block, and the spacing to the bottom edge of the card is ca. 1/10 inch, or ca. 4 millimeters.

   XXX            BLOCK 1             XXX            BLOCK 2             XXX
  XXXXX                              XXXXX                              XXXXX
  XXXXX   X X X X X X  X X X X X X   XXXXX   X X X X X X  X X X X X X   XXXXX
  XXXXX                              XXXXX                              XXXXX
   XXX   HHHHHHHHHHHHHHHHHHHH......   XXX   HHHHHHHHHHHHHHHHHHHH......   XXX
         ..........................         ..........................
         ...... 3 short lines .....         ..........................
    A..................................A..................................A..
    A....      26 long lines       ....A........ X = Sync Marks   ........A..
    A....  (each 34 data dots)     ....A........ H = Block Header ........A..
    A....(not all lines shown here)....A........ . = Data Bits    ........A..
    A..................................A........ A = Address Bits ........A..
         ...... 3 short lines .....         ..........................
         ...(each 26 data dots)....         ..........................
   XXX   ..........................   XXX   ..........................   XXX
  XXXXX                              XXXXX                              XXXXX
  XXXXX   X X X X X X  X X X X X X   XXXXX   X X X X X X  X X X X X X   XXXXX
  XXXXX                              XXXXX                              XXXXX
   XXX                                XXX                                XXX
             <ca. 35 blank lines>
  ___Snip____________________________________________________________________

Address Columns
Each Column consists of 26 dots. From top to bottom: 1 black dot, 8 blank dots, 16 address dots (MSB topmost), and 1 blank dot. The 16bit address values can be calculated as:

  addr[0] = 03FFh
  for i = 1 to 53
    addr[i] = addr[i-1] xor ((i and (-i)) * 769h)
    if (i and 07h)=0 then addr[i] = addr[i] xor (769h)
    if (i and 0Fh)=0 then addr[i] = addr[i] xor (769h*2)
    if (i and 1Fh)=0 then addr[i] = addr[i] xor (769h*4) xor (769h)
  next i

Short strips use addr[1..19], long strips use addr[25..53], left to right.

Block Header
The 18h-byte Block Header is taken from the 1st two bytes (20 dots) of the 1st 0Ch blocks (and is then repeated in the 1st two bytes of further blocks).

  00h      Unknown              (00h)
  01h      Dotcode type         (02h=Short, 03h=Long)
  02h      Unknown              (00h)
  03h      Address of 1st Block (01h=Short, 19h=Long)
  04h      Total Fragment Size  (40h) ;64 bytes per fragment, of which,
                                      ;48 bytes are actual data, the remaining
  05h      Error-Info Size      (10h) ;16 bytes are error-info
  06h      Unknown              (00h)
  07h      Interleave Value     (1Ch=Short, 2Ch=Long)
  08h..17h 16 bytes Reed-solomon error correction info for Block Header

Data 4-Bit to 5-bit Conversion
In the Block Header (HHHHH), and Data Region (.....), each 4bit are expanded to 5bit, so one byte occupies 10 dots, and each block (1040 data dots) contains 104 bytes.

  4bit  00h 01h 02h 03h 04h 05h 06h 07h 08h 09h 0Ah 0Bh 0Ch 0Dh 0Eh 0Fh
  5bit  00h 01h 02h 12h 04h 05h 06h 16h 08h 09h 0Ah 14h 0Ch 0Dh 11h 10h

That formatting ensures that there are no more than two continous black dots (in horizontal direction), neither inside of a 5bit value, nor between two 5bit values, however, the address bars are violating that rule, and up to 5 continous black dots can appear at the (..A..) block boundaries.

Data Order
Data starts with the upper bit of the 5bit value for the upper 4bit of the first byte, which is located at the leftmost dot of the upper line of the leftmost block, it does then extend towards rightmost dot of that block, and does then continue in the next line, until reaching the bottom of the block, and does then continue in the next block. The 1st two bytes of each block contain a portion of the Block Header, the remaining 102 bytes in each block contain data.

Data Size
A long strip consists of 28 blocks (28*104 = 2912 bytes), a short strip of 18 blocks (18*104 = 1872 bytes). Of which, less than 75% can be actually used for program code, the remaining data contains error correction info, and various headers. See Data Format for more info.

Interleaved Fragments
The Interleave Value (I) specifies the number of fragments, and does also specify the step to the next byte inside of a fragment; except that, at the block boundaries (every 104 bytes), the step is 2 bigger (for skipping the next two Block Header bytes).

  RAW Offset  Content
  000h..001h  1st 2 bytes of RAW Header
  002h        1st byte of 1st fragment
  003h        1st byte of 2nd fragment
  ...         ...
  002h+I-1    1st byte of last fragment
  002h+I      2nd byte of 1st fragment
  003h+I      2nd byte of 2nd fragment
  ...         ...
  002h+I*2-1  2nd byte of last fragment
  ...         ...

Each fragment consists of 48 actual data bytes, followed by 16 error correction bytes, followed by 0..2 unused bytes (since I*40h doesn't exactly match num_blocks*102).

GBA Cart e-Reader Data Format

Data Strip Format
The size of the data region is I*48 bytes (I=Interleave Value, see Dotcode Format), the first 48-byte fragment contains the Data Header, the remaining (I-1) fragments are Data Fragments (which contain title(s), and VPK compressed program code).

First Strip

  Data Header  (48 bytes)
  Main-Title   (17 bytes, or 33 bytes)
  Sub-Title(s) (3+18 bytes, or 33 bytes) (for each strip) (optional)
  VPK Size     (2 byte value, total length of VPK Data in ALL strips)
  NULL Value   (4 bytes, contained ONLY in 1st strip of GBA strips)
  VPK Data     (length as defined in VPK Size entry, see above)

Further Strip(s)

  Data Header  (48 bytes)
  Main-Title   (17 bytes, or 33 bytes)
  Sub-Title(s) (3+18 bytes, or 33 bytes) (for each strip) (optional)
  VPK Data     (continued from previous strip)

Data Header (30h bytes) (1st fragment)

  00h-01h  Fixed         (00h,30h)
  02h      Fixed         (01h)     ;01h="Do not calculate Global Checksum" ?
  03h      Primary Type  (see below)
  04h-05h  Fixed         (00h,01h) (don't care)
  06h-07h  Strip Size    (0510h=Short, 0810h=Long Strip) ((I-1)*30h) (MSB,LSB)
  08h-0Bh  Fixed         (00h,00h,10h,12h)
  0Ch-0Dh  Region/Type   (see below)
  0Eh      Strip Type    (02h=Short Strip, 01h=Long Strip) (don't care)
  0Fh      Fixed         (00h) (don't care)
  10h-11h  Unknown       (whatever) (don't care)
  12h      Fixed         (10h)     ;10h="Do calculate Data Checksum" ?
  13h-14h  Data Checksum (see below) (MSB,LSB)
  15h-19h  Fixed         (19h,00h,00h,00h,08h)
  1Ah-21h  ID String     ('NINTENDO')
  22h-25h  Fixed         (00h,22h,00h,09h)
  26h-29h  Size Info     (see below)
  2Ah-2Dh  Flags         (see below)
  2Eh      Header Checksum (entries [0Ch-0Dh,10h-11h,26h-2Dh] XORed together)
  2Fh      Global Checksum (see below)

Primary Type [03h] is 8bit,

  0      Card Type (upper bit) (see below)
  1      Unknown (usually opposite of Bit0) (don't care)
  2-7    Unknown (usually zero)

Region/Type [0Ch..0Dh] is 16bit,

  0-3    Unknown (don't care)
  4-7    Card Type (lower bits) (see below)
  8-11   Region/Version (0=Japan/Original, 1=Non-japan, 2=Japan/Plus)
  12-15  Unknown (don't care)

Size Info [26h-29h] is 32bit,

  0      Unknown            (don't care)
  1-4    Strip Number       (01h..Number of strips)
  5-8    Number of Strips   (01h..0Ch) (01h..08h for Japan/Original version)
  9-23   Size of all Strips (excluding Headers and Main/Sub-Titles)
         (same as "VPK Size", but also including the 2-byte "VPK Size" value,
         plus the 4-byte NULL value; if it is present)
  24-31  Fixed              (02h) (don't care)

Flags [2Ah-2Dh] is 32bit,

  0      Permission to save (0=Start Immediately, 1=Prompt for FLASH Saving)
  1      Sub-Title Flag     (0=Yes, 1=None)    (Japan/Original: always 0=Yes)
  2      Application Type   (0=GBA/Z80, 1=NES) (Japan/Original: always 0=Z80)
  3-31   Zero (0) (don't care)

Data Checksum [13h-14h] is the complement (NOT) of the sum of all halfwords in all Data Fragments, however, it's all done in reversed byte order: checksum is calculated with halfwords that are read in MSB,LSB order, and the resulting checksum is stored in MSB,LSB order in the Header Fragment.
Global Checksum [2Fh] is the complement (NOT) of the sum of the first 2Fh bytes in the Data Header plus the sum of all Data Fragment checksums; the Data Fragment checksums are all 30h bytes in a fragment XORed with each other.

Titles (3+N bytes, or N bytes)
Titles can be 33 bytes for both Main and Sub (Format 0Eh), or Main=17 bytes and Sub=3+18 bytes (Formats 02h..05h). In the 3+N bytes form, the first 3 bytes (24bit) are are used to display "stats" information in form of "HP: h1 ID: i1-i2-i3", defined as:

  Bit    Expl.
  0-3    h1, values 1..15 shown as "10..150", value 0 is not displayed
  4-6    i3, values 0..7 shown as "A..G,#"
  7-13   i2, values 0..98 shown as "01..99" values 99..127 as "A0..C8"
  14-18  i1, values 0..31 shown as "A..Z,-,_,{HP},.,{ID?},:"
  19-22  Unknown
  23     Disable stats (0=Show as "HP: h1 ID: i1-i2-i3", 1=Don't show it)

The N bytes portion contains the actual title, which must be terminated by 00h (so the max length is N-1 characters, if it is shorter than N-1, then the unused bytes are padded by further 00h's). The character set is normal ASCII for non-Japan (see Region/Version entry in header), and 2-byte SHIFT-JIS for Japanese long-titles (=max 16 2-byte chars) with values as so:

  00h          --> end-byte
  81h,40h      --> SPC
  81h,43h..97h --> punctuation marks
  82h,4Fh..58h --> "0..9"
  82h,60h..79h --> "A..Z"
  82h,81h..9Ah --> "a..z"

And 1-byte chars for Japanese short-titles,

  00     = end-byte
  01     = spc
  02..0B = 0..9
  0C..AF = japanese
  B0..B4 = dash, male, female, comma, round-dot
  B5..C0 = !"%&~?/+-:.'
  C1..DA = A..Z
  DB..DF = unused (blank)
  E0..E5 = japanese
  E6..FF = a..z
  N/A    = #$()*;<=>@[\]^_`{|}

Additionally to the Main-Title, optional Sub-Titles for each strip can be included (see Sub-Title Flag in header). If enabled, then ALL strip titles are included in each strip (allowing to show a preview of which strips have/haven't been scanned yet).
The e-Reader can display maximum of 8 sub-titles, if the data consists of more than 8 strips, then sub-titles aren't displayed (so it'd be waste of space to include them in the dotcodes).
The Main Title gets clipped to 128 pixels width (that are, circa 22 characters), and, the e-Reader BIOS acts confused on multi-strip games with Main Titles longer than 26 characters (so the full 33 bytes may be used only in Japan; with 16bit charset).
If the title is empty (00h-filled), and there is only one card in the application, then the application is started immediately. That, without allowing the user to save it in FLASH memory.
Caution: Although shorter Titles do save memory, they do act unpleasant: the text "(C) P-Letter" will be displayed at the bottom of the loading screen.
On Japanese/Original, 8bit sub-titles can be up to 18 characters (without any end-byte) (or less when stats are enabled, due to limited screen width).

Card Types (Primary Type.Bit0 and Region/Type.Bit12-15)

  00h..01h  Blank Screen (?)
  02h..03h  Dotcode Application with 17byte-title, with stats, load music A
  04h..05h  Dotcode Application with 17byte-title, with stats, load music B
  06h..07h  P-Letter Attacks
  08h..09h  Construction Escape
  0Ah..0Bh  Construction Action
  0Ch..0Dh  Construction Melody Box
  0Eh       Dotcode Application with 33byte-title, without stats, load music A
  0Fh       Game specific cards
  10h..1Dh  P-Letter Viewer
  1Eh..1Fh  Same as 0Eh and 0Fh (see above)

The 'Application' types are meant to be executable GBA/Z80/NES programs.

GBA Cart e-Reader Program Code

The GBA/Z80/NES program code is stored in the VPK compressed area.
NES-type is indicated by header [2Ah].Bit2, GBA-type is indicated by the NULL value inserted between VPK Size and VPK Data, otherwise Z80-type is used.

GBA Format
Load Address and Entrypoint are at 2000000h (in ARM state). The 32bit word at 2000008h is eventually destroyed by the e-Reader. Namely,

  IF e-Reader is Non-Japanese,
  AND [2000008h] is outside of range of 2000000h..20000E3h,
  AND only if booted from camera (not when booted from FLASH?),
  THEN [2000008h]=[2000008h]-0001610Ch ELSE [2000008h] kept intact

Existing multiboot-able GBA binaries can be converted to e-Reader format by,

  Store "B 20000C0h" at 2000000h   ;redirect to RAM-entrypoint
  Zerofill 2000004h..20000BFh      ;erase header (for better compression rate)
  Store 01h,01h at 20000C4h        ;indicate RAM boot

The GBA code has full access to the GBA hardware, and may additionally use whatever API functions contained in the e-Reader BIOS. With the incoming LR register value, "mov r0,N, bx lr" returns to the e-Reader BIOS (with N being 0=Restart, or 2=To_Menu). No idea if it's necessary to preserve portions of RAM when returning to the e-Reader BIOS?
Caution: Unlike for normal GBA cartridges/multiboot files, the hardware is left uninitialized when booting dotcodes (among others: sound DMA is active, and brightness is set to zero), use "mov r0,0feh, swi 010000h" to get the normal settings.

NES Format
Emulates a NES (Nintendo Entertainment System) console (aka Family Computer).
The visible 240x224 pixel NES/NTSC screen resolution is resampled to 240x160 to match the smaller vertical resolution of the GBA hardware. So, writing e-Reader games in NES format will result in blurred screen output. The screen/sound/joypad is accessed via emulated NES I/O ports, program code is running on an emulated 6502 8bit CPU, for more info on the NES hardware, see no$nes debugger specifications, or

  http://problemkaputt.de/everynes.htm

The e-Reader's NES emulator supports only 16K PRG ROM, followed by 8K VROM. The emulation accuracy is very low, barely working with some of Nintendo's own NES titles; running the no$nes diagnostics program on it has successfully failed on ALL hardware tests ;-)
The load address for the 16K PRG-ROM is C000h, the 16bit NMI vector at [FFFAh] is encrypted like so:

  for i=17h to 0
   for j=07h to 0, nmi = nmi shr 1, if carry then nmi = nmi xor 8646h, next j
   nmi = nmi xor (byte[dmca_data+i] shl 8)
  next i
  dmca_data: db 0,0,'DMCA NINTENDO E-READER'

The 16bit reset vector at [FFFCh] contains:

  Bit0-14  Lower bits of Entrypoint (0..7FFFh = Address 8000h..FFFFh)
  Bit15    Nametable Mode (0=Vertical Mirroring, 1=Horizontal Mirroring)

reportedly,

   (NES limitations, 1 16K program rom + 1-2 8K CHR rom, mapper 0 and 1)
   ines mapper 1 would be MMC1, rather than CNROM (ines mapper 3)?
   but, there are more or less NONE games that have 16K PRG ROM + 16K VROM?

The L+R Button key-combination allows to reset the NES, however, there seems to be no way to return to the e-Reader BIOS.

Z80/8080 Format
The e-Reader doesn't support the following Z80 opcodes:

  CB [Prefix]     E0 RET PO   E2 JP PO,nn   E4 CALL PO,nn   27 DAA    76 HALT
  ED [Prefix]     E8 RET PE   EA JP PE,nn   EC CALL PE,nn   D3 OUT (n),A
  DD [IX Prefix]  F3 DI       08 EX AF,AF'  F4 CALL P,nn    DB IN A,(n)
  FD [IY Prefix]  FB EI       D9 EXX        FC CALL M,nn    xx RST 00h..38h

That is leaving not more than six supported Z80 opcodes (DJNZ, JR, JR c/nc/z/nz), everything else are 8080 opcodes. Custom opcodes are:

  76 WAIT A frames, D3 WAIT n frames, and C7/CF RST 0/8 used for API calls.

The load address and entrypoint are at 0100h in the emulated Z80 address space. The Z80 doesn't have direct access to the GBA hardware, instead video/sound/joypad are accessed via API functions, invoked via RST 0 and RST 8 opcodes, followed by an 8bit data byte, and with parameters in the Z80 CPU registers. For example, "ld a,02h, rst 8, db 00h" does return to the e-Reader BIOS.
The Z80/8080 emulation is incredibly inefficient, written in HLL code, developed by somebody whom knew nothing about emulation nor about ARM nor about Z80/8080 processors.

Running GBA-code on Japanese/Original e-Reader
Original e-Reader supports Z80 code only, but can be tweaked to run GBA-code:

  retry:
   ld bc,data // ld hl,00c8h      ;src/dst
  lop:
   ld a,[bc] // inc bc // ld e,a  ;lsb
   ld a,[bc] // inc bc // ld d,a  ;msb
   dw 0bcfh ;aka rst 8 // db 0bh  ;[4000000h+hl]=de (DMA registers)
   inc hl // inc  hl // ld a,l
   cp a,0dch // jr nz,lop
  mod1 equ $+1
   dw 37cfh ;aka rst 8 // db 37h  ;bx 3E700F0h
  ;below executed only on jap/plus... on jap/plus, above 37cfh is hl=[400010Ch]
   ld a,3Ah // ld [mod1],a                  ;bx 3E700F0h (3Ah instead 37h)
   ld hl,1 // ld [mod2],hl // ld [mod3],hl  ;base (0200010Ch instead 0201610Ch)
   jr retry
  data:
  mod2 equ $+1
   dd loader         ;40000C8h dma2sad (loader)            ;\
   dd 030000F0h      ;40000CCh dma2dad (mirrored 3E700F0h) ; relocate loader
   dd 8000000ah      ;40000D0h dma2cnt (copy 0Ah x 16bit)  ;/
  mod3 equ $+1
   dd main           ;40000D4h dma3sad (main)              ;\prepare main reloc
   dd 02000000h      ;40000D8h dma3dad (2000000h)          ;/dma3cnt see loader
   .align 2          ;alignment for 16bit-halfword
  org $+201600ch     ;jap/plus: adjusted to org $+200000ch
  loader:
   mov r0,80000000h  ;(dma3cnt, copy 10000h x 16bit)
   mov r1,04000000h  ;i/o base
   strb r1,[r1,208h] ;ime=0 (better disable ime before moving ram)
   str r0,[r1,0DCh]  ;dma3cnt (relocate to 2000000h)
   mov r15,2000000h  ;start relocated code at 2000000h in ARM state
  main:
   ;...insert/append whatever ARM code here...
   end

GBA Cart e-Reader API Functions

Z80 Interface (Special Opcodes)

  db 76h       ;Wait8bit A
  db D3h,xxh   ;Wait8bit xxh
  db C7h,xxh   ;RST0_xxh
  db CFh,xxh   ;RST8_xxh
  ld r,[00xxh]       ;get system values (addresses differ on jap/ori)
  ld r,[00C2h..C3h]  ;GetKeyStateSticky (jap/ori: 9F02h..9F03h)
  ld r,[00C4h..C5h]  ;GetKeyStateRaw    (jap/ori: 9F04h..9F05h)
  ld r,[00C0h..C1h]  ;see Exit and ExitRestart
  ld r,[00D0h..D3h]  ;see Mul16bit

For jap/ori, 9Fxxh isn't forwards compatible with jap/plus, so it'd be better to check joypad via IoRead.

GBA Interface

  bx [30075FCh] ;ApiVector ;in: r0=func_no,r1,r2,r3,[sp+0],[sp+4],[sp+8]=params
  bx lr         ;Exit      ;in: r0 (0=Restart, 2=To_Menu)

Wait8bit/Wait16bit
The various Wait opcodes and functions are waiting as many frames as specified. Many API functions have no effect until the next Wait occurs.

Z80 RST0_xxh Functions / GBA Functions 02xxh

  RST0_00h FadeIn, A speed, number of frames (0..x)
  RST0_01h FadeOut
  RST0_02h BlinkWhite
  RST0_03h  (?)
  RST0_04h  (?) blend_func_unk1
  RST0_05h  (?)
  RST0_06h  (?)
  RST0_07h  (?)
  RST0_08h  (?)
  RST0_09h  (?) _020264CC_check
  RST0_0Ah  (?) _020264CC_free
  RST0_0Bh N/A (bx 0)
  RST0_0Ch N/A (bx 0)
  RST0_0Dh N/A (bx 0)
  RST0_0Eh N/A (bx 0)
  RST0_0Fh N/A (bx 0)
  RST0_10h LoadSystemBackground, A number of background (1..101), E bg# (0..3)
  RST0_11h SetBackgroundOffset, A=bg# (0..3), DE=X, BC=Y
  RST0_12h SetBackgroundAutoScroll
  RST0_13h SetBackgroundMirrorToggle
  RST0_14h  (?)
  RST0_15h  (?)
  RST0_16h  (?) write_000000FF_to_02029494_
  RST0_17h  (?)
  RST0_18h  (?)
  RST0_19h SetBackgroundMode, A=mode (0..2)
  RST0_1Ah  (?)
  RST0_1Bh  (?)
  RST0_1Ch  (?)
  RST0_1Dh  (?)
  RST0_1Eh  (?)
  RST0_1Fh  (?)
  RST0_20h LayerShow
  RST0_21h LayerHide
  RST0_22h  (?)
  RST0_23h  (?)
  RST0_24h ... [20264DCh+A*20h+1Ah]=DE, [20264DCh+A*20h+1Ch]=BC
  RST0_25h  (?)
  RST0_26h  (?)
  RST0_27h  (?)
  RST0_28h  (?)
  RST0_29h  (?)
  RST0_2Ah  (?)
  RST0_2Bh  (?)
  RST0_2Ch  (?)
  RST0_2Dh LoadCustomBackground, A bg# (0..3), DE pointer to struct_background,
           max. tile data size = 3000h bytes, max. map data size = 1000h bytes
  RST0_2Eh GBA: N/A - Z80: (?)
  RST0_2Fh  (?)
  RST0_30h CreateSystemSprite, - -   (what "- -" ???)
  RST0_31h SpriteFree, HL sprite handle
  RST0_32h SetSpritePos, HL=sprite handle, DE=X, BC=Y
  RST0_33h  (?) sprite_unk2
  RST0_34h SpriteFrameNext
  RST0_35h SpriteFramePrev
  RST0_36h SetSpriteFrame, HL=sprite handle, E=frame number (0..x)
  RST0_37h  (?) sprite_unk3
  RST0_38h  (?) sprite_unk4
  RST0_39h SetSpriteAutoMove, HL=sprite handle, DE=X, BC=Y
  RST0_3Ah  (?) sprite_unk5
  RST0_3Bh  (?) sprite_unk6
  RST0_3Ch SpriteAutoAnimate
  RST0_3Dh  (?) sprite_unk7
  RST0_3Eh SpriteAutoRotateUntilAngle
  RST0_3Fh SpriteAutoRotateByAngle
  RST0_40h SpriteAutoRotateByTime
  RST0_41h  (?) sprite_unk8
  RST0_42h SetSpriteAutoMoveHorizontal
  RST0_43h SetSpriteAutoMoveVertical
  RST0_44h  (?) sprite_unk9
  RST0_45h SpriteDrawOnBackground
  RST0_46h SpriteShow, HL=sprite handle
  RST0_47h SpriteHide, HL=sprite handle
  RST0_48h SpriteMirrorToggle
  RST0_49h  (?) sprite_unk10
  RST0_4Ah  (?) sprite_unk11
  RST0_4Bh  (?) sprite_unk12
  RST0_4Ch GetSpritePos
  RST0_4Dh CreateCustomSprite
  RST0_4Eh  (?)
  RST0_4Fh  (?) sprite_unk14
  RST0_50h  (?) sprite_unk15
  RST0_51h  (?) sprite_unk16
  RST0_52h  (?) sprite_unk17
  RST0_53h  (?) sprite_unk18
  RST0_54h  (?)
  RST0_55h  (?) sprite_unk20
  RST0_56h  (?)
  RST0_57h SpriteMove
  RST0_58h  (?) sprite_unk22
  RST0_59h  (?) sprite_unk23
  RST0_5Ah  (?) sprite_unk24
  RST0_5Bh SpriteAutoScaleUntilSize, C=speed (higher value is slower),
           HL=sprite handle, DE=size (0100h = normal size,
           lower value = larger, higher value = smaller)
  RST0_5Ch SpriteAutoScaleBySize
  RST0_5Dh SpriteAutoScaleWidthUntilSize
  RST0_5Eh SpriteAutoScaleHeightBySize
  RST0_5Fh  (?)
  RST0_60h  (?)
  RST0_61h  (?)
  RST0_62h  (?)
  RST0_63h  (?)
  RST0_64h hl=[[2024D28h+a*4]+12h]
  RST0_65h  (?) sprite_unk25
  RST0_66h SetSpriteVisible, HL=sprite handle, E=(0=not visible, 1=visible)
  RST0_67h  (?) sprite_unk26
  RST0_68h  (?) set_sprite_unk27
  RST0_69h  (?) get_sprite_unk27
  RST0_6Ah  (?)
  RST0_6Bh  (?)
  RST0_6Ch  (?)
  RST0_6Dh  (?)
  RST0_6Eh hl=[hl+000Ah]  ;r0=[r1+0Ah]
  RST0_6Fh  (?)
  RST0_70h  (?)
  RST0_71h  (?)
  RST0_72h  (?)
  RST0_73h  (?)
  RST0_74h  (?)
  RST0_75h  (?)
  RST0_76h  (?)
  RST0_77h  (?)
  RST0_78h  (?)
  RST0_79h  (?)
  RST0_7Ah  (?)
  RST0_7Bh  (?)
  RST0_7Ch  (?) _0202FD2C_unk12
  RST0_7Dh Wait16bit ;HL=num_frames (16bit variant of Wait8bit opcode/function)
  RST0_7Eh SetBackgroundPalette, HL=src_addr, DE=offset, C=num_colors (1..x)
  RST0_7Fh GetBackgroundPalette(a,b,c)
  RST0_80h SetSpritePalette, HL=src_addr, DE=offset, C=num_colors (1..x)
  RST0_81h GetSpritePalette(a,b,c)
  RST0_82h ClearPalette
  RST0_83h  (?) _0202FD2C_unk11
  RST0_84h  (?)
  RST0_85h  (?)
  RST0_86h  (?)
  RST0_87h  (?) _0202FD2C_unk8
  RST0_88h  (?) _0202FD2C_unk7
  RST0_89h  (?)
  RST0_8Ah  (?) _0202FD2C_unk6
  RST0_8Bh  (?) _0202FD2C_unk5
  RST0_8Ch GBA: N/A - Z80: (?)
  RST0_8Dh GBA: N/A - Z80: (?)
  RST0_8Eh  (?)
  RST0_8Fh WindowHide
  RST0_90h CreateRegion, H=bg# (0..3), L=palbank# (0..15),
           D,E,B,C=x1,y1,cx,cy (in tiles), return: n/a (no$note: n/a ???)
  RST0_91h SetRegionColor
  RST0_92h ClearRegion
  RST0_93h SetPixel
  RST0_94h GetPixel
  RST0_95h DrawLine
  RST0_96h DrawRect
  RST0_97h  (?) _0202FD2C_unk4
  RST0_98h SetTextColor, A=region handle, D=color foreground (0..15),
           E=color background (0..15)
  RST0_99h DrawText, A=region handle, BC=pointer to text, D=X, E=Y
           (non-japan uses ASCII text, but japanese e-reader's use STH ELSE?)
  RST0_9Ah SetTextSize
  RST0_9Bh  (?) RegionUnk7
  RST0_9Ch  (?) _0202FD2C_unk3
  RST0_9Dh  (?) _0202FD2C_unk2
  RST0_9Eh  (?) _0202FD2C_unk1
  RST0_9Fh Z80: (?) - GBA: SetBackgroundModeRaw
  RST0_A0h  (?)
  RST0_A1h  (?)
  RST0_A2h  (?) RegionUnk6
  RST0_A3h GBA: N/A - Z80: (?)
  RST0_A4h GBA: N/A - Z80: (?)
  RST0_A5h  (?)
  RST0_A6h  (?)
  RST0_A7h  (?)
  RST0_A8h  (?)
  RST0_A9h  (?)
  RST0_AAh  (?)
  RST0_ABh  (?)
  RST0_ACh  (?)
  RST0_ADh  (?) RegionUnk5
  RST0_AEh [202FD2Ch+122h]=A
  RST0_AFh [202FD2Ch+123h]=A
  RST0_B0h [202FD2Ch+124h]=A
  RST0_B1h  (?)
  RST0_B2h  (?)
  RST0_B3h GBA: N/A - Z80: Sqrt   ;hl=sqrt(hl)
  RST0_B4h GBA: N/A - Z80: ArcTan ;hl=ArcTan2(hl,de)
  RST0_B5h Sine                   ;hl=sin(a)*de
  RST0_B6h Cosine                 ;hl=cos(a)*de
  RST0_B7h  (?)
  RST0_B8h  (?)
  RST0_B9h N/A (bx 0)
  RST0_BAh N/A (bx 0)
  RST0_BBh N/A (bx 0)
  RST0_BCh N/A (bx 0)
  RST0_BDh N/A (bx 0)
  RST0_BEh N/A (bx 0)
  RST0_BFh N/A (bx 0)
  Below Non-Japan and Japan/Plus only (not Japan/Ori)
  RST0_C0h GetTextWidth(a,b)
  RST0_C1h GetTextWidthEx(a,b,c)
  RST0_C2h  (?)
  RST0_C3h Z80: N/A (bx 0) - GBA: (?)
  RST0_C4h  (?)
  RST0_C5h  (?)
  RST0_C6h  (?)
  RST0_C7h  (?)
  RST0_C8h  (?)
  RST0_C9h  (?)
  RST0_CAh  (?)
  RST0_CBh  (?)
  RST0_CCh  (?)
  RST0_CDh N/A (bx lr)
  RST0_CEh ;same as RST0_3Bh, but with 16bit mask
  RST0_CFh ;same as RST0_3Eh, but with 16bit de
  RST0_D0h ;same as RST0_3Fh, but with 16bit de
  RST0_D1h ;same as RST0_5Bh, but with 16bit de
  RST0_D2h ;same as RST0_5Ch, but with 16bit de
  RST0_D3h ;same as RST0_5Dh, but with 16bit de
  RST0_D4h ;same as RST0_5Eh, but with 16bit de
  RST0_D5h  (?)
  RST0_D6h  (?)
  RST0_D7h ;[202FD2Ch+125h]=A
  RST0_D8h  (?)
  RST0_D9h  (?)
  RST0_DAh  (?)
  RST0_DBh ;A=[3003E51h]
  RST0_DCh ;[3004658h]=01h
  RST0_DDh DecompressVPKorNonVPK
  RST0_DEh FlashWriteSectorSingle(a,b)
  RST0_DFh FlashReadSectorSingle(a,b)
  RST0_E0h SoftReset
  RST0_E1h GetCartridgeHeader     ;[hl+0..BFh]=[8000000h..80000BFh]
  RST0_E2h GBA: N/A - Z80: bx hl  ;in: hl=addr, af,bc,de,sp=param, out: a
  RST0_E3h Z80: N/A (bx 0) - GBA: (?)
  RST0_E4h  (?)
  RST0_E5h  (?)
  RST0_E6h  (?)
  RST0_E7h  (?)
  RST0_E8h  (?)
  RST0_E9h ;[2029498h]=0000h
  RST0_EAh Z80: N/A (bx 0) - GBA: InitMemory(a)
  RST0_EBh  (?) BL_irq_sio_dma3
  RST0_ECh ;hl = [3003E30h]*100h + [3003E34h]
  RST0_EDh FlashWriteSectorMulti(a,b,c)
  RST0_EEh FlashReadPart(a,b,c)
  RST0_EFh ;A=((-([2029416h] xor 1)) OR (+([2029416h] xor 1))) SHR 31
  RST0_F0h  (?) _unk1
  RST0_F1h RandomInit     ;in: hl=random_seed
  RST0_F2h                         (?)
  Below Japan/Plus only
  RST0_F3h  (?)
  RST0_F4h  (?)
  RST0_F5h  (?)
  RST0_F6h  (?)
  RST0_F7h GBA: N/A - Z80: (?)
  Below is undefined/garbage (values as so in Z80 mode)
  Jap/Ori: RST0_C0h      N/A (bx 0)
  Jap/Ori: RST0_C1h..FFh Overlaps RST8 jump list
  Non-Jap: RST0_F3h..FFh Overlaps RST8 jump list
  Jap/Pls: RST0_F8h..FFh Overlaps RST8 jump list

Z80 RST8_xxh Functions / GBA Functions 01xxh

  RST8_00h GBA: N/A - Z80: Exit       ;[00C0h]=a ;(1=restart, 2=exit)
  RST8_01h GBA: N/A - Z80: Mul8bit    ;hl=a*e
  RST8_02h GBA: N/A - Z80: Mul16bit   ;hl=hl*de, s32[00D0h]=hl*de
  RST8_03h Div                        ;hl=hl/de
  RST8_04h DivRem                     ;hl=hl mod de
  RST8_05h PlaySystemSound            ;in: hl=sound_number
  RST8_06h  (?) sound_unk1
  RST8_07h Random8bit                 ;a=random(0..FFh)
  RST8_08h SetSoundVolume
  RST8_09h BcdTime                    ;[de+0..5]=hhmmss(hl*bc)
  RST8_0Ah BcdNumber                  ;[de+0..4]=BCD(hl), [de+5]=00h
  RST8_0Bh IoWrite                    ;[4000000h+hl]=de
  RST8_0Ch IoRead                     ;de=[4000000h+hl]
  RST8_0Dh GBA: N/A - Z80:  (?)
  RST8_0Eh GBA: N/A - Z80:  (?)
  RST8_0Fh GBA: N/A - Z80:  (?)
  RST8_10h GBA: N/A - Z80:  (?)
  RST8_11h DivSigned                  ;hl=hl/de, signed
  RST8_12h RandomMax                  ;a=random(0..a-1)
  RST8_13h SetSoundSpeed
  RST8_14h  hl=[202FD20h]=[2024CACh]
  RST8_15h  hl=[2024CACh]-[202FD20h]
  RST8_16h SoundPause
  RST8_17h SoundResume
  RST8_18h PlaySystemSoundEx
  RST8_19h IsSoundPlaying
  RST8_1Ah  (?)
  RST8_1Bh  (?)
  RST8_1Ch  (?)
  RST8_1Dh GetExitCount               ;a=[2032D34h]
  RST8_1Eh Permille                   ;hl=de*1000/hl
  RST8_1Fh GBA: N/A - Z80: ExitRestart;[2032D38h]=a, [00C0h]=0001h  ;a=?
  RST8_20h GBA: N/A - Z80: WaitJoypad ;wait until joypad<>0, set hl=joypad
  RST8_21h GBA: N/A - Z80:  (?)
  RST8_22h  (?) _sound_unk7
  RST8_23h  (?) _sound_unk8
  RST8_24h  (?) _sound_unk9
  RST8_25h  (?) _sound_unk10
  RST8_26h Mosaic     ;bg<n>cnt.bit6=a.bit<n>, [400004Ch]=de
  RST8_27h  (?)
  RST8_28h  (?)
  RST8_29h  (?)
  RST8_2Ah  (?) get_8bit_from_2030110h
  RST8_2Bh  (?)
  RST8_2Ch  (?) get_16bit_from_2030112h ;jap/ori: hl=[20077B2h]
  RST8_2Dh  (?) get_16bit_from_2030114h ;jap/ori: hl=[20077B4h]
  RST8_2Eh  (?)
  RST8_2Fh PlayCustomSound(a,b)
  Below not for Japanese/Original
  (the renumbered functions can be theoretically used on japanese/original)
  (but, doing so would blow forwards compatibility with japanese/plus)
  RST8_30h (ori: none)      GBA: N/A - Z80: (?)
  RST8_31h (ori: none)      PlayCustomSoundEx(a,b,c)
  RST8_32h (ori: RST8_30h)  BrightnessHalf   ;[4000050h]=00FFh,[4000054h]=0008h
  RST8_33h (ori: RST8_31h)  BrightnessNormal ;[4000050h]=0000h
  RST8_34h (ori: RST8_32h)  N/A (bx lr)
  RST8_35h (ori: RST8_33h)   (?)
  RST8_36h (ori: RST8_34h)  ResetTimer ;[400010Ch]=00000000h, [400010Eh]=A+80h
  RST8_37h (ori: RST8_35h)  GetTimer   ;hl=[400010Ch]
  RST8_38h (ori: none)      GBA: N/A - Z80:  (?)
  Below is undefined/reserved/garbage (values as so in Z80 mode)
  (can be used to tweak jap/ori to start GBA-code from inside of Z80-code)
  (that, after relocating code to 3000xxxh via DMA via IoWrite function)
  RST8_39h (ori: RST8_36h)  bx 0140014h
  RST8_3Ah (ori: RST8_37h)  bx 3E700F0h
  RST8_3Bh (ori: RST8_38h)  bx 3E70000h+1
  RST8_3Ch (ori: RST8_39h)  bx 3E703E6h+1
  RST8_3Dh (ori: RST8_3Ah)  bx 3E703E6h+1
  RST8_3Eh (ori: RST8_3Bh)  bx 3E703E6h+1
  RST8_3Fh (ori: RST8_3Ch)  bx 3E703E6h+1
  40h-FFh  (ori: 3Dh-FFh)   bx ...

GBA Functions 03xxh (none such in Z80 mode)

  RSTX_00h Wait8bit  ;for 16bit: RST0_7Dh
  RSTX_01h GetKeyStateSticky()
  RSTX_02h GetKeyStateRaw()
  RSTX_03h  (?)
  RSTX_04h  (?)

GBA Cart e-Reader VPK Decompression

vpk_decompress(src,dest)

  collected32bit=80000000h  ;initially empty (endflag in bit31)
  for i=0 to 3, id[i]=read_bits(8), next i, if id[0..3]<>'vpk0' then error
  dest_end=dest+read_bits(32)     ;size of decompressed data (of all strips)
  method=read_bits(8), if method>1 then error
  tree_index=0, load_huffman_tree, disproot=tree_index
  tree_index=tree_index+1, load_huffman_tree, lenroot=tree_index
  ;above stuff is contained only in the first strip. below loop starts at
  ;current location in first strip, and does then continue in further strips.
 decompress_loop:
  if read_bits(1)=0 then                   ;copy one uncompressed data byte,
    [dest]=read_bits(8), dest=dest+1       ;does work without huffman trees
  else
    if disproot=-1 or lenroot=-1 then error  ;compression does require trees
    disp=read_tree(disproot)
    if method=1   ;disp*4 is good for 32bit ARM opcodes
      if disp>2 then disp=disp*4-8 else disp=disp+4*read_tree(disproot)-7
    len=read_tree(lenroot)
    if len=0 or disp<=0 or dest+len-1>dest_end then error ;whoops
    for j=1 to len, [dest]=[dest-disp], dest=dest+1, next j
  if dest<dest_end then decompress_loop
  ret

read_bits(num)

  mov  data=0
  for i=1 to num
    shl collected32bit,1   ;move next bit to carry, or set zeroflag if empty
    if zeroflag
      collected32bit=[src+0]*1000000h+[src+1]*10000h+[src+2]*100h+[src+3]
      src=src+4            ;read data in 32bit units, in reversed byte-order
      carryflag=1          ;endbit
      rcl collected32bit,1 ;move bit31 to carry (and endbit to bit0)
    rcl data,1             ;move carry to data
  next i
  ret(data)

read_tree(root_index)

  i=root_index
  while node[i].right<>-1  ;loop until reaching data node
    if read_bits(1)=1 then i=node[i].right else i=node[i].left
  i=node[i].left           ;get number of bits
  i=read_bits(i)           ;read that number of bits
  ret(i)                   ;return that value

load_huffman_tree

  stacktop=sp
  if read_bits(1)=1 then tree_index=-1, ret  ;exit (empty)
  node[tree_index].right=-1                  ;indicate data node
  node[tree_index].left=read_bits(8)         ;store data value
  if read_bits(1)=1 then ret                 ;exit (only 1 data node at root)
  push tree_index                     ;save previous (child) node
  tree_index=tree_index+1
  jmp data_injump
 load_loop:
  push tree_index                     ;save previous (child) node
  tree_index=tree_index+1
  if read_bits(1)=1 then parent_node
 data_injump:
  node[tree_index].right=-1           ;indicate data node
  node[tree_index].left=read_bits(8)  ;store data value
  jmp load_loop
 parent_node:
  pop node[tree_index].right          ;store 1st child
  pop node[tree_index].left           ;store 2nd child
  if sp<>stacktop then jmp load_loop
  if read_bits(1)=0 then error        ;end bit (must be 1)
  ret

The best values for the huffman trees that I've found are 6,9,12-bit displacements for method 0 (best for NES/Z80 code), and two less for method 1, ie. 4,7,10-bit (best for GBA code). And 2,4,10-bit for the length values. The smallest value in node 0, and the other values in node 10 and 11.

Notes
The decompression works similar to the GBA BIOS'es LZ77 decompression function, but without using fixed bit-widths of length=4bit and displacement=12bit, instead, the bit-widths are read from huffman trees (which can also define fixed bit-widths; if data is located directly in the root node).
Unlike the GBA BIOS'es Huffman decompression function, the trees are starting with data entries, end are ending with the root entry. The above load function deciphers the data, and returns the root index.
With the variable bit-widths, the VPK compression rate is quite good, only, it's a pity that the length/disp values are zero-based, eg. for 2bit and 4bit lengths, it'd be much better to assign 2bit as 2..5, and 4bit as 6..21.

Non-VPK
The e-Reader additionally supports an alternate decompression function, indicated by the absence of the "vpk0" ID, which supports compression of increasing byte-values, which isn't useful for program code.
Bit15 of the VPK Size value seems to disable (de-)compression, the VPK Data field is then containing plain uncompressed data.

GBA Cart e-Reader Error Correction

The Error Correction Information that is appended at the end of the Block Header & Data Fragments consists of standard Reed-Solomon codes, which are also used for CD/DVD disks, DSL modems, and digital DVB television signals. That info allows to locate and repair a number of invalid data bytes.

Below code shows how to create and verify error-info (but not how to do the actual error correction). The dtalen,errlen values should be 18h,10h for the Block Header, and 40h,10h for Data Fragments; the latter settings might be possible to get changed to other values though?

append_error_info(data,dtalen,errlen)

  reverse_byte_order(data,dtalen)
  zerofill_error_bytes(data,errlen)
  for i=dtalen-1 to errlen  ;loop across data portion
    z = rev[ data[i] xor data[errlen-1] ] ;
    for j=errlen-1 to 0     ;loop across error-info portion
    if j=0 then x=00h else x=data[j-1]
      if z<>FFh then
        y=gg[j], if y<>FFh then
          y=y+z, if y>=FFh then y=y-FFh
          x=x xor pow[y]
      data[j]=x
    next j
  next i
  invert_error_bytes(data,errlen)
  reverse_byte_order(data,dtalen)

verify_error_info(data,dtalen,errlen)

  reverse_byte_order(data,dtalen)
  invert_error_bytes(data,errlen)
  make_rev(data,dtalen)
  for i=78h to 78h+errlen-1
    x=0, z=0
    for j=0 to dtalen-1
      y=data[j]
      if y<>FFh then
        y=y+z, if y>=FFh then y=y-FFh
        x=x xor pow[y]
      z=z+i, if z>=FFh then z=z-FFh
    next j
    if x<>0 then error
  next i
  ;(if errors occured, could correct them now)
  make_pow(data,dtalen)
  invert_error_bytes(data,errlen)
  reverse_byte_order(data,dtalen)

make_rev(data,len)

  for i=0 to len-1, data[i]=rev[data[i]], next i

make_pow(data,len)

  for i=0 to len-1, data[i]=pow[data[i]], next i

invert_error_bytes(data,len)

  for i=0 to len-1, data[i]=data[i] xor FFh, next i

zerofill_error_bytes(data,len)

  for i=0 to len-1, data[i]=00h, next i

reverse_byte_order(data,len)

  for i=0 to (len-1)/2, x=data[i], data[i]=data[len-i], data[len-i]=x, next i

create_pow_and_rev_tables

  x=01h, pow[FFh]=00h, rev[00h]=FFh
  for i=00h to FEh
    pow[i]=x, rev[x]=i, x=x*2, if x>=100h then x=x xor 187h
  next i

create_gg_table

  gg[0]=pow[78h]
  for i=1 to errlen-1
    gg[i]=01h
    for j=i downto 0
      if j=0 then y=00h else y=gg[j-1]
      x=gg[j], if x<>00h then
        x=rev[x]+78h+i, if x>=FFh then x=x-FFh
        y=y xor pow[x]
      gg[j]=y
    next j
  next i
  make_rev(gg,errlen)

With above value of 78h, and errlen=10h, gg[00h..0Fh] will be always:

  00h,4Bh,EBh,D5h,EFh,4Ch,71h,00h,F4h,00h,71h,4Ch,EFh,D5h,EBh,4Bh

So using a hardcoded table should take up less memory than calculating it.

Notes
The actual error correction should be able to fix up to "errlen" errors at known locations (eg. data from blocks that haven't been scanned, or whose 5bit-to-4bit conversion had failed due to an invalid 5bit value), or up to "errlen/2" errors at unknown locations. The corrected data isn't guaranteed to be correct (even if it looks okay to the "verify" function), so the Data Header checksums should be checked, too.

More Info
For more info, I've found Reed-Solomon source code from Simon Rockliff, and an updated version from Robert Morelos-Zaragoza and Hari Thirumoorthy to be useful. For getting started with that source, some important relationships & differences are:

  pow = alpha_to, but generated as shown above
  rev = index_of, dito
  b0  = 78h
  nn  = dtalen
  kk  = dtalen-errlen
  %nn = MOD FFh (for the ereader that isn't MOD dtalen)
  -1  = FFh

And, the ereader processes data/errinfo backwards, starting at the last byte.

GBA Cart e-Reader File Formats

.BMP Files (homebrew 300 DPI strips)
Contains a picture of the whole dotcode strip with address bars and sync marks (see Dotcode chapter) in Microsoft's Bitmap format. The image is conventionally surrounded by a blank 2-pixel border, resulting in a size of 989x44 pixels for long strips. The file should should have 1bit color depth. The pixels per meter entry should match the desired printing resolution, either 300 DPI or 360 DPI. But, resolution of printer hardware is typically specified in inch rather than in meters, so an exact match isn't supported by Microsoft. Most homebrew .BMP files contain nonsense resolutions like 200 DPI, or 300 dots per meter (ca. 8 DPI).

.JPG Files (scanned 1200 DPI strips)
Same as BMP, but should contain a dotcode scanned at 1200 DPI, with correct orientation (the card-edge side at the bottom of the image), and containing only the dotcode (not the whole card), so the JPG size should be about 3450x155 pixels for long strips.
No$gba currently doesn't work with progressive JPGs. Scans with white background can be saved as monochrome JPG. Scans with red/yellow background should contain a correct RED layer (due to the red LED light source) (the brightness of the green/blue layers can be set to zero for better compression).

.RAW Files
Contains the "raw" information from the BMP format, that is, 2-byte block header, 102-byte data, 2-byte block header, 102-byte data, etc. The data portion is interleaved, and includes the full 48-byte data header, titles, vpk compressed data, error-info, and unused bytes. RAW files are excluding Address Bars, Sync Marks, and 4bit-to-5bit encoding.
Each RAW file contains one or more strip(s), so the RAW filesize is either 18*104 bytes (short strip), or 28*104 bytes (long strip), or a multiple thereof (if it contains more than one strip) (although multi-strip games are often stored in separate files for each strip; named file1.raw, file2.raw, etc).

.BIN Files
Filesize should be I*30h, with I=1Ch for short strips, and I=2Ch for long strips, or a multiple thereof (if it contains more than one strip). Each strip consists of the 48-byte Data Header, followed by title(s), and vpk compressed data. Unlike .RAW files, .BIN files aren't interleaved, and do not contain Block Headers, nor error-info, nor unused bytes (in last block). The files do contain padding bytes to match a full strip-size of I*30h.
Caution: Older .BIN files have been using a size-reduced 12-byte header (taken from entries 0Dh, 0Ch, 10h-11h, 26h-2Dh of the 48-byte Data Header; in that order), that files have never contained more than one strip per file, so the filesize should be exactly I*30h-36, the size-reduced header doesn't contain a Primary Type entry, so it's everyone's bet which Card Type is to be used (hint: the 12-byte headers were based on the assumption that Primary Type would be always 01h on Short Strips, and 02h on Long Strips).

.SAV Files
Contains a copy of the e-Reader's 128Kbyte FLASH memory. With the saved e-Reader application being located in the 2nd 64K-bank, the data consists of a header with title and gba/nes/z80 format info, followed by the vpk compressed data. The FLASH memory does also contain e-Reader calibration settings, the remaining 100Kbytes are typically FFh-filled.

GBA Cart Unknown Devices

GBA Infra-Red Port (AGB-006)
No info?

GBA Cart Protections

Classic NES Series
These are some NES/Famicom games ported or emulated to work on GBA. The games are doing some uncommon stuff that can cause compatibility problems when not using original GBA consoles or cartridges.
- CPU pipeline (selfmodifying code that shall NOT affect prefetched opcodes)
- STMDA write to I/O ports (writes in INCREASING order, not DECREASING order)
- SRAM detection (refuses to run if SRAM exists; the games do contain EEPROM)
- ROM mirrors (instead of the usual increasing numbers in unused ROM area)
- RAM mirrors (eg. main RAM accessed at 2F00000h instead of 2000000h)
Note: These games can be detected by checking [80000ACh]="F" (ie. game code="Fxxx").

GBA Flashcards

Flashcards are re-writable cartridges using FLASH memory, allowing to test even multiboot-incompatible GBA software on real hardware, providing a good development environment when used in combination with a reasonable software debugger.

The carts can be written to from external tools, or directly from GBA programs.
Below are pseudo code flowcharts for detect, erase, and write operations.
All flash reads/writes are meant to be 16bit (ldrh/strh) memory accesses.

detect_flashcard:

 configure_flashcard(9E2468Ah,9413h)    ;unlock flash advance cards
 turbo=1, send_command(8000000h,90h)    ;enter ID mode (both chips, if any)
 maker=[8000000h], device=[8000000h+2]
 IF maker=device THEN device=[8000000h+4] ELSE turbo=0
 flashcard_read_mode                    ;exit ID mode
 search (maker+device*10000h) in device_list
 total/erase/write_block_size = list_entry SHL turbo

flashcard_erase(dest,len):

 FOR x=1 to len/erase_block_size
  send_command(dest,20h)        ;erase sector command
  send_command(dest,D0h)        ;confirm erase sector
  dest=dest+erase_block_size
 IF wait_busy=okay THEN NEXT x
 enter_read_mode                ;exit erase/status mode

flashcard_write(src,dest,len):

 siz=write_block_size
 FOR x=1 to len/siz
  IF siz=2 THEN send_command(dest,10h)  ;write halfword command
  IF siz>2 THEN send_command(dest,E8h)  ;write to buffer command
  IF siz>2 THEN send_command(dest,16-1) ;buffer size 16 halfwords (per chip)
  FOR y=1 TO siz/2
   [dest]=[src], dest=dest+2, src=src+2 ;write data to buffer
  NEXT y
  IF siz>2 THEN send_command(dest,D0h)  ;confirm write to buffer
 IF wait_busy=okay THEN NEXT x
 enter_read_mode                        ;exit write/status mode

send_command(adr,val):

 [adr]=val
 IF turbo THEN [adr+2]=val

enter_read_mode:

 send_command(8000000h,FFh)     ;exit status mode
 send_command(8000000h,FFh)     ;again maybe more stable (as in jeff's source)

flashcard_wait_busy:

 start=time
 REPEAT
  stat=[8000000h] XOR 80h
  IF turbo THEN stat=stat OR ([8000000h+2] XOR 80h)
  IF (stat AND 7Fh)>0 THEN error
  IF (stat AND 80h)=0 THEN ready
  IF time-start>5secs THEN timeout
 UNTIL ready OR error OR timeout
 IF error OR timeout THEN send_command(8000000h,50h)    ;clear status

configure_flashcard(adr,val): ;required for Flash Advance cards only

 [930ECA8h]=5354h
 [802468Ah]=1234h, repeated 500 times
 [800ECA8h]=5354h
 [802468Ah]=5354h
 [802468Ah]=5678h, repeated 500 times
 [930ECA8h]=5354h
 [802468Ah]=5354h
 [8ECA800h]=5678h
 [80268A0h]=1234h
 [802468Ah]=ABCDh, repeated 500 times
 [930ECA8h]=5354h
 [adr]=val

init_backup: ;no info how to use that exactly

 configure_flashcard(942468Ah,???)

device_list: (id code, total/erase/write sizes in bytes)

  ID Code    Total   Erase  Write  Name
  -??-00DCh      ?       ?      ?  Hudson Cart (???)
  00160089h     4M    128K     32  Intel i28F320J3A (Flash Advance)
  00170089h     8M    128K     32  Intel i28F640J3A (Flash Advance)
  00180089h    16M    128K     32  Intel i28F128J3A (Flash Advance)
  00E200B0h      ?     64K      2  Sharp LH28F320BJE ? (Nintendo)

Notes
All flashcards should work at 4,2 waitstates (power on default), most commercial games change waits to 3,1 which may work unstable with some/older FA flashcards. Intel FLASH specified to have a lifetime of 100,000 erases, and average block erase time 1 second (up to 5 second in worst cases).
Aside from the main FLASH memory, Flash Advance (FA) (aka Visoly) cards additionally contain battery buffered SRAM backup, and FLASH backup, and in some cases also EEPROM backup.
Turbo FA cards are containing two chips interlaced (at odd/even halfword addresses), allowing to write/erase both chips simultaneously, resulting in twice as fast programming time.
Standard Nintendo flash carts have to be modified before you can actually write to them. This is done by removing resistor R7 and putting it at empty location R8.
Mind that write/erase/detect modes output status information in ROM area, so that in that modes all GBA program code (and any interrupt handlers) must be executed in WRAM, not in ROM.

Thanks to Jeff Frohwein for his FAQ and CARTLIB sample in FLGBA at devrs.com

GBA Cheat Devices

Codebreaker (US) aka Xploder (EUR).
Gameshark (US) aka Action Replay (EUR).

GBA Cheat Codes - General Info
GBA Cheat Codes - Codebreaker/Xploder
GBA Cheat Codes - Gameshark/Action Replay V1/V2
GBA Cheat Codes - Pro Action Replay V3

GBA Cheat Codes - General Info

Cheat devices are external adapters, connected between the GBA and the game cartridge. The devices include a BIOS ROM which is, among others, used to prompt the user to enter cheat codes.
These codes are used to patch specified memory locations for a certain GBA game, allowing the user to gain goodies such like Infinite sex, 255 Cigarettes, etc.

ROM and RAM Patches
For ROM Patches, the device watches the address bus, if it matches a specified address then it outputs a patched value to the data bus, that mechanism is implemented by hardware, aside from the Hook Enable Code some devices also allow a limited number of cheats to use ROM patches.
Most cheat codes are RAM patches, each time when the hook procedure is executed it will process all codes and overwrite the specified addresses in RAM (or VRAM or I/O area) by the desired values.

Enable Codes (Must Be On)
Enable codes usually consist of the Game ID, Hook Address, and eventually a third code used to encrypt all following codes. The Game ID is used to confirm that the correct cartridge is inserted, just a verification, though the device may insist on the ID code.
The Hook Address specifies an address in cartridge ROM, and should point to an opcode which is executed several times per second (eg. once per frame, many codes place the hook in the joypad handler). At the hook address, the device redirects to its own BIOS, processes the RAM patches, and does then return control to the game cartridge.
Note: The hook address should not point to opcodes with relative addressing (eg. B, BL, LDR Rd,=Imm, ADD Rd,=Imm opcodes - which are all relative to PC program counter register).

Alignment
Addresses for 16bit or 32bit values should be properly aligned.

GBA Cheat Codes - Codebreaker/Xploder

Codebreaker Codes

  0000xxxx 000y  Enable Code 1 - Game ID
  1aaaaaaa 000z  Enable Code 2 - Hook Address
  2aaaaaaa yyyy  [aaaaaaa]=[aaaaaaa] OR yyyy
  3aaaaaaa 00yy  [aaaaaaa]=yy
  4aaaaaaa yyyy  [aaaaaaa+0..(cccc-1)*ssss]=yyyy+0..(cccc-1)*ssss
  iiiicccc ssss  parameters for above code
  5aaaaaaa cccc  [aaaaaaa+0..(cccc-1)]=11,22,33,44,etc.
  11223344 5566  parameter bytes 1..6 for above code (example)
  77880000 0000  parameter bytes 7..8 for above code (padded with zero)
  6aaaaaaa yyyy  [aaaaaaa]=[aaaaaaa] AND yyyy
  7aaaaaaa yyyy  IF [aaaaaaa]=yyyy THEN (next code)
  8aaaaaaa yyyy  [aaaaaaa]=yyyy
  9xyyxxxx xxxx  Enable Code 0 - Encrypt all following codes (optional)
  Aaaaaaaa yyyy  IF [aaaaaaa]<>yyyy THEN (next code)
  Baaaaaaa yyyy  IF [aaaaaaa]>yyyy THEN (next code) (signed comparison)
  Caaaaaaa yyyy  IF [aaaaaaa]<yyyy THEN (next code) (signed comparison)
  D0000020 yyyy  IF [joypad] AND yyyy = 0 THEN (next code)
  Eaaaaaaa yyyy  [aaaaaaa]=[aaaaaaa]+yyyy
  Faaaaaaa yyyy  IF [aaaaaaa] AND yyyy THEN (next code)

Codebreaker Enable Codes
Hook Address 'aaaaaaa' is a 25bit offset in ROM-image (0-1FFFFFFh).
Flag byte 'y' (usually 0Ah), Bit1=Disable IRQs, Bit3=CRC Exists.
Code Handler Store Address 'z' (0-7, usually 7) (8000100h+z*400000h).
Checksum 'xxxx' for first 64Kbytes of cartridge (no$gba pads by FFh if ROM is smaller than 64K). Calculated, by using unsigned 16bit values, as such:

  crc=FFFFh
  for i=0 to FFFFh
   x=byte[i] xor (crc/100h)
   x=x xor (x/10h)
   crc=(crc*100h) xor (x*1001h) xor (x*20h)
  next i

Codebreaker Encryption
codebreaker_change_encryption:
Encryption can be (optionally) activated by code "9xyyxxxx xxxx",

  for i=0 to 2Fh, swaplist[i]=i, next i
  randomizer = 1111h xor byte[code+4]                              ;LSB value
  for i=0 to 4Fh
    exchange swaplist[random MOD 30h] with swaplist[random MOD 30h]
  next i
  halfword[seedlist+0] = halfword[code+0]                          ;LSW address
  randomizer = 4EFAD1C3h
  for i=0 to byte[code+3]-91h, randomizer=random, next i           ;MSB address
  word[seedlist+2]=random, halfword[seedlist+6]=random
  randomizer = F254h xor byte[code+5]                              ;MSB value
  for i=0 to byte[code+5]-01h, randomizer=random, next i           ;MSB value
  word[seedlist+8]=random, halfword[seedlist+12]=random
  ;note: byte[code+2] = don't care
  ret

The above random function works like so:

  randomizer=randomizer*41C64E6Dh+3039h, x=(randomizer SHL 14 AND C0000000h)
  randomizer=randomizer*41C64E6Dh+3039h, x=(randomizer SHR 1  AND 3FFF8000h)+x
  randomizer=randomizer*41C64E6Dh+3039h, x=(randomizer SHR 16 AND 00007FFFh)+x
  return(x)

Once when encryption is activated, all following codes are decrypted like so:

  for i=2Fh to 0
    j=swaplist[i]
    bitno1=(i AND 7), index1=xlatlist[i/8]
    bitno2=(j AND 7), index2=xlatlist[j/8]
    exchange [code+index1].bitno1 with [code+index2].bitno2
  next i
  word[code+0] = word[code+0] xor word[seedlist+8]
  i = (byte[code+3]*1010000h + byte[code+0]*100h + byte[code+5])
  i = (halfword[code+1]*10001h) xor (word[seedlist+2]) xor i
  i = (byte[seedlist+0]*1010101h) xor (byte[seedlist+1]*1000000h) xor i
  j = (byte[code+5] + (byte[code+0] xor byte[code+4])*100h)
  j = (byte[seedlist+0]*101h) xor halfword[seedlist+6] xor j
  word[code+0] = i, halfword[code+4] = j

The above xlatlist is fixed: xlatlist[0..5] = 3,2,1,0,5,4

GBA Cheat Codes - Gameshark/Action Replay V1/V2

Gameshark RAW Codes (These codes must be encrypted before using them)

  0aaaaaaa 000000xx  [aaaaaaa]=xx
  1aaaaaaa 0000xxxx  [aaaaaaa]=xxxx
  2aaaaaaa xxxxxxxx  [aaaaaaa]=xxxxxxxx
  3000cccc xxxxxxxx  write xxxxxxxx to (cccc-1) addresses (list in next codes)
  aaaaaaaa aaaaaaaa  parameter for above code, containing two addresses each
  aaaaaaaa 00000000  last parameter for above, zero-padded if only one address
  60aaaaaa y000xxxx  [8000000h+aaaaaa*2]=xxxx (ROM Patch)
  8a1aaaaa 000000xx  IF GS_Button_Down THEN [a0aaaaa]=xx
  8a2aaaaa 0000xxxx  IF GS_Button_Down THEN [a0aaaaa]=xxxx
  80F00000 0000xxxx  IF GS_Button_Down THEN slowdown xxxx * ? cycles per hook
  Daaaaaaa 0000xxxx  IF [aaaaaaa]=xxxx THEN (next code)
  E0zzxxxx 0aaaaaaa  IF [aaaaaaa]=xxxx THEN (next 'zz' codes)
  Faaaaaaa 00000x0y  Enable Code - Hook Routine
  xxxxxxxx 001DC0DE  Enable Code - Game Code ID (value at [0ACh] in cartridge)
  DEADFACE 0000xxyy  Change Encryption Seeds

Enable Code - Hook Routine
Hook Address 'aaaaaaa' is a 28bit ROM address (8FFFFFFh-9FFFFFFh).
Used to insert the GS code handler routine where it will be executed at
least 20 times per second. Without this code, GSA can not write to RAM.

 y=1 - Executes code handler without backing up the LR register.
 y=2 - Executes code handler and backs up the LR register.
 y=3 - Replaces a 32-bit pointer used for long-branches.
 x=0 - Must turn GSA off before loading game.
 x=1 - Must not do that.

ROM Patch
This type allows GSA to intercept ROM reads and returns the value xxxx.

 y=0 wait for the code handler to enable the patch
 y=1 patch is enabled before the game starts
 y=2 unknown ?

Note: V1/V2 hardware can only have up to 1 user-defined rom patch max. V3 can have up to 4. Some enable code types can shorten the amount of user-defined rom patches available.

Gameshark Encryption
A=Left half, and V=Right half of code.

  FOR I=1 TO 32
    A=A + (V*16+S0) XOR (V+I*9E3779B9h) XOR (V/32+S1)
    V=V + (A*16+S2) XOR (A+I*9E3779B9h) XOR (A/32+S3)
  NEXT I

Upon startup, the initial encryption seeds are:

  S0=09F4FBBDh S1=9681884Ah S2=352027E9h S3=F3DEE5A7h

Upon DEADFACE 0000xxyy, the S0..S3 seeds are changed like so:

  FOR y=0 TO 3
   FOR x=0 TO 3
    z = T1[(xx+x) AND FFh] + T2[(yy+y) AND FFh]
    Sy = Sy*100h + (z AND FFh)
   NEXT x
  NEXT y

All calculations truncated to unsigned 32bit integer values.
T1 and T2 are translation tables contained in the gameshark cartridge.

GBA Cheat Codes - Pro Action Replay V3

Pro Action Replay V3 - RAW Codes

  C4aaaaaa 0000yyyy  Enable Code - Hook Routine at [8aaaaaa]
  xxxxxxxx 001DC0DE  Enable Code - ID Code [080000AC]
  DEADFACE 0000xxxx  Enable Code - Change Encryption Seeds
  00aaaaaa xxxxxxyy  [a0aaaaa..a0aaaaa+xxxxxx]=yy
  02aaaaaa xxxxyyyy  [a0aaaaa..a0aaaaa+xxxx*2]=yyyy
  04aaaaaa yyyyyyyy  [a0aaaaa]=yyyyyyyy
  40aaaaaa xxxxxxyy  [ [a0aaaaa] + xxxxxx ]=yy   (Indirect)
  42aaaaaa xxxxyyyy  [ [a0aaaaa] + xxxx*2 ]=yyyy (Indirect)
  44aaaaaa yyyyyyyy  [ [a0aaaaa] ]=yyyyyyyy      (Indirect)
  80aaaaaa 000000yy  [a0aaaaa]=[a0aaaaa]+yy
  82aaaaaa 0000yyyy  [a0aaaaa]=[a0aaaaa]+yyyy
  84aaaaaa yyyyyyyy  [a0aaaaa]=[a0aaaaa]+yyyyyyyy
  C6aaaaaa 0000yyyy  [4aaaaaa]=yyyy              (I/O Area)
  C7aaaaaa yyyyyyyy  [4aaaaaa]=yyyyyyyy          (I/O Area)
  iiaaaaaa yyyyyyyy  IF [a0aaaaa] <cond> <value> THEN <action>
  00000000 60000000  ELSE (?)
  00000000 40000000  ENDIF (?)
  00000000 0800xx00  AR Slowdown : loops the AR xx times
  00000000 00000000  End of the code list
  00000000 10aaaaaa 000000zz 00000000  IF AR_BUTTON THEN [a0aaaaa]=zz
  00000000 12aaaaaa 0000zzzz 00000000  IF AR_BUTTON THEN [a0aaaaa]=zzzz
  00000000 14aaaaaa zzzzzzzz 00000000  IF AR_BUTTON THEN [a0aaaaa]=zzzzzzzz
  00000000 18aaaaaa 0000zzzz 00000000  [8000000+aaaaaa*2]=zzzz  (ROM Patch 1)
  00000000 1Aaaaaaa 0000zzzz 00000000  [8000000+aaaaaa*2]=zzzz  (ROM Patch 2)
  00000000 1Caaaaaa 0000zzzz 00000000  [8000000+aaaaaa*2]=zzzz  (ROM Patch 3)
  00000000 1Eaaaaaa 0000zzzz 00000000  [8000000+aaaaaa*2]=zzzz  (ROM Patch 4)

  00000000 80aaaaaa 000000yy ssccssss  repeat cc times [a0aaaaa]=yy
   (with yy=yy+ss, a0aaaaa=a0aaaaa+ssss after each step)

  00000000 82aaaaaa 0000yyyy ssccssss  repeat cc times [a0aaaaa]=yyyy
   (with yyyy=yyyy+ss, a0aaaaa=a0aaaaa+ssss*2 after each step)

  00000000 84aaaaaa yyyyyyyy ssccssss  repeat cc times [a0aaaaa]=yyyyyyyy
   (with yyyy=yyyy+ss, a0aaaaa=a0aaaaa+ssss*4 after each step)

Warning: There is a bug on the real AR (v2 upgraded to v3, and maybe on real v3) with the 32bit Increment Slide code. You HAVE to add a code (best choice is 80000000 00000000 : add 0 to value at address 0) right after it, else the AR will erase the 2 last 8 digits lines of the 32 Bits Inc. Slide code when you enter it !!!

Final Notes
The 'turn off all codes' makes an infinite loop (that can't be broken, unless the condition becomes True). - How? By Interrupt? Huh?
ROM Patch1 works on real V3 and, on V1/V2 upgraded to V3.
ROM Patch2,3,4 work on real V3 hardware only.

Pro Action Replay V3 Conditional Codes - iiaaaaaa yyyyyyyy
The 'ii' is composed of <cond> + <value> + <action>.

  <cond>           <value>            <action>
  08 Equal =       00 8bit zz         00 execute next code
  10 Not equal <>  02 16bit zzzz      40 execute next two codes
  18 Signed <      04 32bit zzzzzzzz  80 execute all following
  20 Signed >      06 (always false)     codes until ELSE or ENDIF
  28 Unsigned <                       C0 normal ELSE turn off all codes
  30 Unsigned >
  38 Logical AND

For example, ii=18h+02h+40h=5Ah, produces IF [a0aaaaa]<zzzz THEN next 2 codes.

Always... Codes

  For the "Always..." codes:
  - XXXXXXXX can be any authorised address except 00000000 (eg. use 02000000).
  - ZZZZZZZZ can be anything.
  - The "y" in the code data must be in the [1-7] range (which means not 0).
  typ=y,sub=0,siz=3   Always skip next line.
  typ=y,sub=1,siz=3   Always skip next 2 lines.
  typ=y,sub=2,siz=3   Always Stops executing all the codes below.
  typ=y,sub=3,siz=3   Always turn off all codes.

Code Format (ttaaaaaa xxxxyyzz)

 adr mask = 003FFFFF
 n/a mask = 00C00000 ;not used
 xtr mask = 01000000 ;used only by I/O write, and MSB of Hook
 siz mask = 06000000
 typ mask = 38000000 ;0=normal, other=conditional
 sub mask = C0000000

Pro Action Replay V3 Encryption
Works exactly as for Gameshark Encryption, but with different initial seeds,

  S0=7AA9648Fh S1=7FAE6994h S2=C0EFAAD5h S3=42712C57h

And, the T1 and T2 translation tables are different, too.

GBA Gameboy Player

The Gameboy Player is an "adapter" for the Gamecube console. It's basicly is a GBA in a black box without LCD screen and without buttons, connected to an expansion port at the bottom of the Gamecube. The Gamecube is then capturing the GBA video output (and passing it to the television set), and in the other direction, passing the Gamecube joypad input to the GBA inputs.

Unlocking and Detecting Gameboy Player Functions
Both unlocking and detection requires to display the 240x160 pixel Gameboy Player logo (44 colors) for a number of frames... maybe at least 3-4 frames? not sure if it checks the color of the logo... so maybe it can be hidden by using dark gray on black background?
While displaying this logo, the joypad data will switch between values 03FFh (2 frames duration) and 030Fh (1 frame duration). The latter value (left, right, up, down all pressed) indicates that it's a Gameboy Player.

Palette
Knowing Nintendo, they've probably not reproduced the blurred GBA colors (?), so the games won't look as desired on the TV screen. Unless the game does detect the Gameboy Player, and adjust the colors accordingly by software.

Rumble
The only known existing special function is the joypad rumble function, controlled by sending data through the serial port (the normal GBA port, even though it also has the connectors).

The Game Boy Player added a rumble feature to certain Game Boy Advance games when played with a GameCube controller. Those games included:

 Drill Dozer (supports BOTH handheld-rumble and GBP-rumble?)
 Mario & Luigi: Superstar Saga
 Pokemon Pinball: Ruby & Sapphire
 Shikakui Atama wo Marukusuru Advance: Kokugo Sansu Rika Shakai
 Shikakui Atama wo Marukusuru Advance: Kanji Keisan
 Summon Night Craft Sword Monogatari: Hajimari no Ishi
 Super Mario Advance 4: Super Mario Bros. 3

Fredrik Olsson (aka Flubba) has implemented rumble in 3 applications now RumblePong (FluBBA) (homebrew)

  Remudvance (FluBBA) (homebrew)
  Goomba (FluBBA) (8bit Gameboy Color Emulator for 32bit GBA) (homebrew)
  and, supposedly in "Tetanus on Drugs" (Tepples) (homebrew)

The GBP can also use some of the extra controllers for the GC like the Bongas
from Donkey Konga.

The logo requires at least 256 colors, it doesn't matter if you use a tiled
screen mode or a bitmapped one, the logo can be ripped from either
"Pokemon Pinball" or "Super Mario Advance 4".

Rumble
After detecting/unlocking the Gameboy Player, init RCNT and SIOCNT to 32bit normal mode, external clock, SO=high, with IRQ enabled, and set the transfer start bit. You should then receive the following sequence (about once per frame), and your serial IRQ handler should send responses accordingly:

  Receive  Response
  0000494E 494EB6B1
  xxxx494E 494EB6B1
  B6B1494E 544EB6B1
  B6B1544E 544EABB1
  ABB1544E 4E45ABB1
  ABB14E45 4E45B1BA
  B1BA4E45 4F44B1BA
  B1BA4F44 4F44B0BB
  B0BB4F44 8000B0BB
  B0BB8002 10000010
  10000010 20000013
  20000013 40000004
  30000003 40000004
  30000003 40000004
  30000003 40000004
  30000003 400000yy
  30000003 40000004

The first part of the transfer just contains the string "NINTENDO" split into 16bit fragments, and bitwise inversions thereof (eg. 494Eh="NI", and B6B1h=NOT 494Eh). In the second part, <yy> should be 04h=RumbleOff, or 26h=RumbleOn.

Note
If it's having a similar range of functions as the 8bit Super Gameboy, then the Gameboy Player might be also able to access analogue joypad input, and to access other features of the Gamecube hardware, up to possibly executing code on the Gamecube CPU...?

GBA Unpredictable Things

Forward
Most of the below is caused by 'traces' from previous operations which have used the databus. No promises that the results are stable on all current or future GBA models, and/or under all temperature and interference circumstances.
Also, below specifies 32bit data accesses only. When reading units less than 32bit, data is rotated depending on the alignment of the originally specified address, and 8bit or 16bit are then isolated from the 32bit value as usually.

Reading from BIOS Memory (00000000-00003FFF)
The BIOS memory is protected against reading, the GBA allows to read opcodes or data only if the program counter is located inside of the BIOS area. If the program counter is not in the BIOS area, reading will return the most recent successfully fetched BIOS opcode (eg. the opcode at [00DCh+8] after startup and SoftReset, the opcode at [0134h+8] during IRQ execution, and opcode at [013Ch+8] after IRQ execution, and opcode at [0188h+8] after SWI execution).

Reading from Unused Memory (00004000-01FFFFFF,10000000-FFFFFFFF)
Accessing unused memory at 00004000h-01FFFFFFh, and 10000000h-FFFFFFFFh (and 02000000h-03FFFFFFh when RAM is disabled via Port 4000800h) returns the recently pre-fetched opcode. For ARM code this is simply:

  WORD = [$+8]

For THUMB code the result consists of two 16bit fragments and depends on the address area and alignment where the opcode was stored.
For THUMB code in Main RAM, Palette Memory, VRAM, and Cartridge ROM this is:

  LSW = [$+4], MSW = [$+4]

For THUMB code in BIOS or OAM (and in 32K-WRAM on Original-NDS (in GBA mode)):

  LSW = [$+4], MSW = [$+6]   ;for opcodes at 4-byte aligned locations
  LSW = [$+2], MSW = [$+4]   ;for opcodes at non-4-byte aligned locations

For THUMB code in 32K-WRAM on GBA, GBA SP, GBA Micro, NDS-Lite (but not NDS):

  LSW = [$+4], MSW = OldHI   ;for opcodes at 4-byte aligned locations
  LSW = OldLO, MSW = [$+4]   ;for opcodes at non-4-byte aligned locations

Whereas OldLO/OldHI are usually:

  OldLO=[$+2], OldHI=[$+2]

Unless the previous opcode's prefetch was overwritten; that can happen if the previous opcode was itself an LDR opcode, ie. if it was itself reading data:

  OldLO=LSW(data), OldHI=MSW(data)
  Theoretically, this might also change if a DMA transfer occurs.

Note: Additionally, as usually, the 32bit data value will be rotated if the data address wasn't 4-byte aligned, and the upper bits of the 32bit value will be masked in case of LDRB/LDRH reads.
Note: The opcode prefetch is caused by the prefetch pipeline in the CPU itself, not by the external gamepak prefetch, ie. it works for code in ROM and RAM as well.

Reading from Unused or Write-Only I/O Ports
Works like above Unused Memory when the entire 32bit memory fragment is Unused (eg. 0E0h) and/or Write-Only (eg. DMA0SAD). And otherwise, returns zero if the lower 16bit fragment is readable (eg. 04Ch=MOSAIC, 04Eh=NOTUSED/ZERO).

Reading from GamePak ROM when no Cartridge is inserted
Because Gamepak uses the same signal-lines for both 16bit data and for lower 16bit halfword address, the entire gamepak ROM area is effectively filled by incrementing 16bit values (Address/2 AND FFFFh).

Memory Mirrors
Most internal memory is mirrored across the whole 24bit/16MB address space in which it is located: Slow On-board RAM at 2XXXXXX, Fast On-Chip RAM at 3XXXXXXh, Palette RAM at 5XXXXXXh, VRAM at 6XXXXXXh, and OAM at 7XXXXXXh. Even though VRAM is sized 96K (64K+32K), it is repeated in steps of 128K (64K+32K+32K, the two 32K blocks itself being mirrors of each other).
BIOS ROM, Normal ROM Cartridges, and I/O area are NOT mirrored, the only exception is the undocumented I/O port at 4000800h (repeated each 64K).
The 64K SRAM area is mirrored across the whole 32MB area at E000000h-FFFFFFFh, also, inside of the 64K SRAM field, 32K SRAM chips are repeated twice.

Writing 8bit Data to Video Memory
Video Memory (BG, OBJ, OAM, Palette) can be written to in 16bit and 32bit units only. Attempts to write 8bit data (by STRB opcode) won't work:
Writes to OBJ (6010000h-6017FFFh) (or 6014000h-6017FFFh in Bitmap mode) and to OAM (7000000h-70003FFh) are ignored, the memory content remains unchanged.
Writes to BG (6000000h-600FFFFh) (or 6000000h-6013FFFh in Bitmap mode) and to Palette (5000000h-50003FFh) are writing the new 8bit value to BOTH upper and lower 8bits of the addressed halfword, ie. "[addr AND NOT 1]=data*101h".

Using Invalid Tile Numbers
In Text mode, large tile numbers (combined with a non-zero character base setting in BGnCNT register) may exceed the available 64K of BG VRAM.
On GBA and GBA SP, such invalid tiles are displayed as if the character data is filled by the 16bit BG Map entry value (ie. as vertically striped tiles). Above applies only if there is only one BG layer enabled, with two or more layers, things are getting much more complicated: tile-data is then somehow derived from the other layers, depending on their priority order and scrolling offsets.
On NDS (in GBA mode), such invalid tiles are displayed as if the character data is zero-filled (ie. as invisible/transparent tiles).

Accessing SRAM Area by 16bit/32bit
Reading retrieves 8bit value from specified address, multiplied by 0101h (LDRH) or by 01010101h (LDR). Writing changes the 8bit value at the specified address only, being set to LSB of (source_data ROR (address*8)).

NDS Reference

DS Technical Data

Processors

  1x ARM946E-S 32bit RISC CPU, 66MHz (NDS9 video) (not used in GBA mode)
  1x ARM7TDMI  32bit RISC CPU, 33MHz (NDS7 sound) (16MHz in GBA mode)

Internal Memory

  4096KB Main RAM (8192KB in debug version)
  96KB   WRAM (64K mapped to NDS7, plus 32K mappable to NDS7 or NDS9)
  60KB   TCM/Cache (TCM: 16K Data, 32K Code) (Cache: 4K Data, 8K Code)
  656KB  VRAM (allocateable as BG/OBJ/2D/3D/Palette/Texture/WRAM memory)
  4KB    OAM/PAL (2K OBJ Attribute Memory, 2K Standard Palette RAM)
  248KB  Internal 3D Memory (104K Polygon RAM, 144K Vertex RAM)
  ?KB    Matrix Stack, 48 scanline cache
  8KB    Wifi RAM
  256KB  Firmware FLASH (512KB in iQue variant, with chinese charset)
  36KB   BIOS ROM (4K NDS9, 16K NDS7, 16K GBA)

Video

  2x LCD screens (each 256x192 pixel, 3 inch, 18bit color depth, backlight)
  2x 2D video engines (extended variants of the GBA's video controller)
  1x 3D video engine (can be assigned to upper or lower screen)
  1x video capture (for effects, or for forwarding 3D to the 2nd 2D engine)

Sound

  16 sound channels (16x PCM8/PCM16/IMA-ADPCM, 6x PSG-Wave, 2x PSG-Noise)
  2 sound capture units (for echo effects, etc.)
  Output: Two built-in stereo speakers, and headphones socket
  Input:  One built-in microphone, and microphone socket

Controls

  Gamepad      4 Direction Keys, 8 Buttons
  Touchscreen  (on lower LCD screen)

Communication Ports

  Wifi IEEE802.11b

Specials

  Built-in Real Time Clock
  Power Managment Device
  Hardware divide and square root functions
  CP15 System Control Coprocessor (cache, tcm, pu, bist, etc.)

External Memory

  NDS Slot (for NDS games) (encrypted 8bit data bus, and serial 1bit bus)
  GBA Slot (for NDS expansions, or for GBA games) (but not for DMG/CGB games)

Manufactured Cartridges

  ROM: 16MB, 32MB, or 64MB
  EEPROM/FLASH/FRAM: 0.5KB, 8KB, 64KB, 256KB, or 512KB

Can be booted from

  NDS Cartridge (NDS mode)
  Firmware FLASH (NDS mode) (eg. by patching firmware via ds-xboo cable)
  Wifi (NDS mode)
  GBA Cartridge (GBA mode) (without DMG/CGB support) (without SIO support)

Power Supply

  Built-in rechargeable Lithium ion battery, 3.7V 1000mAh (DS-Lite)
  External Supply: 5.2V DC

NDS-Lite
Slightly smaller than the original NDS, coming in a more decently elegant case. The LCDs are much more colorful (and thus not backwards compatible with any older NDS or GBA games), and the LCDs support wider viewing angles. Slightly different power managment device (with selectable backlight brightness, new external power source flag, lost audio amplifier mute flag). Slightly different Wifi controller (different chip ID, different dirt effects when accessing invalid wifi ports and unused wifi memory regions, different behaviour on GAPDISP registers, RF/BB chips replaced by a single chip). Slightly different touch screen controller (with new unused input, and slightly different powerdown bits).

Notice
NDS9 means the ARM9 processor and its memory and I/O ports in NDS mode
NDS7 means the ARM7 processor and its memory and I/O ports in NDS mode
GBA means the ARM7 processor and its memory and I/O ports in GBA mode

The two Processors
Most game code is usually executed on the ARM9 processor (in fact, Nintendo reportedly doesn't allow developers use the ARM7 processor, except by predefined API functions, anyways, even with the most likely inefficient API code, most of the ARM7's 33MHz horsepower is left unused).
The ARM9's 66MHz "horsepower" is a different tale - it seems Nintendo thought that a 33MHz processor would be too "slow" for 3D games, and so they (tried to) badge an additional CPU to the original GBA hardware.
However, the real 66MHz can be used only with cache and tcm, all other memory and I/O accesses are delayed to the 33MHz bus clock, that'd be still quite fast, but, there seems to be a hardware glitch that adds 3 waitcycles to all nonsequential accesses at the NDS9 side, which effectively drops its bus clock to about 8MHz, making it ways slower than the 33MHz NDS7 processor, it's even slower than the original 16MHz GBA processor.
Altogether, with the bugged 66MHz, and the unused 33MHz, Nintendo could have reached almost the same power when staying with the GBA's 16MHz processor :-)
Although, when properly using cache/tcm, then the 66MHz processor <can> be very fast, still, the NDS should have worked as well with a single processor, though using only an ARM9 might cause a lot of compatibility problems with GBA games, so there's at least one reason for keeping the ARM7 included.

DS I/O Maps

ARM9 I/O Map
ARM9 Display Engine A

  4000000h  4    2D Engine A - DISPCNT - LCD Control (Read/Write)
  4000004h  2    2D Engine A+B - DISPSTAT - General LCD Status (Read/Write)
  4000006h  2    2D Engine A+B - VCOUNT - Vertical Counter (Read only)
  4000008h  50h  2D Engine A (same registers as GBA, some changed bits)
  4000060h  2    DISP3DCNT - 3D Display Control Register (R/W)
  4000064h  4    DISPCAPCNT - Display Capture Control Register (R/W)
  4000068h  4    DISP_MMEM_FIFO - Main Memory Display FIFO (R?/W)
  400006Ch  2    2D Engine A - MASTER_BRIGHT - Master Brightness Up/Down

GBA I/O Map
ARM9 DMA, Timers, and Keypad

  40000B0h  30h  DMA Channel 0..3
  40000E0h  10h  DMA FILL Registers for Channel 0..3
  4000100h  10h  Timers 0..3
  4000130h  2    KEYINPUT
  4000132h  2    KEYCNT

ARM9 IPC/ROM

  4000180h  2  IPCSYNC - IPC Synchronize Register (R/W)
  4000184h  2  IPCFIFOCNT - IPC Fifo Control Register (R/W)
  4000188h  4  IPCFIFOSEND - IPC Send Fifo (W)
  40001A0h  2  AUXSPICNT - Gamecard ROM and SPI Control
  40001A2h  2  AUXSPIDATA - Gamecard SPI Bus Data/Strobe
  40001A4h  4  Gamecard bus timing/control
  40001A8h  8  Gamecard bus 8-byte command out
  40001B0h  4  Gamecard Encryption Seed 0 Lower 32bit
  40001B4h  4  Gamecard Encryption Seed 1 Lower 32bit
  40001B8h  2  Gamecard Encryption Seed 0 Upper 7bit (bit7-15 unused)
  40001BAh  2  Gamecard Encryption Seed 1 Upper 7bit (bit7-15 unused)

ARM9 Memory and IRQ Control

  4000204h  2  EXMEMCNT - External Memory Control (R/W)
  4000208h  2  IME - Interrupt Master Enable (R/W)
  4000210h  4  IE  - Interrupt Enable (R/W)
  4000214h  4  IF  - Interrupt Request Flags (R/W)
  4000240h  1  VRAMCNT_A - VRAM-A (128K) Bank Control (W)
  4000241h  1  VRAMCNT_B - VRAM-B (128K) Bank Control (W)
  4000242h  1  VRAMCNT_C - VRAM-C (128K) Bank Control (W)
  4000243h  1  VRAMCNT_D - VRAM-D (128K) Bank Control (W)
  4000244h  1  VRAMCNT_E - VRAM-E (64K) Bank Control (W)
  4000245h  1  VRAMCNT_F - VRAM-F (16K) Bank Control (W)
  4000246h  1  VRAMCNT_G - VRAM-G (16K) Bank Control (W)
  4000247h  1  WRAMCNT   - WRAM Bank Control (W)
  4000248h  1  VRAMCNT_H - VRAM-H (32K) Bank Control (W)
  4000249h  1  VRAMCNT_I - VRAM-I (16K) Bank Control (W)

ARM9 Maths

  4000280h  2  DIVCNT - Division Control (R/W)
  4000290h  8  DIV_NUMER - Division Numerator (R/W)
  4000298h  8  DIV_DENOM - Division Denominator (R/W)
  40002A0h  8  DIV_RESULT - Division Quotient (=Numer/Denom) (R)
  40002A8h  8  DIVREM_RESULT - Division Remainder (=Numer MOD Denom) (R)
  40002B0h  2  SQRTCNT - Square Root Control (R/W)
  40002B4h  4  SQRT_RESULT - Square Root Result (R)
  40002B8h  8  SQRT_PARAM - Square Root Parameter Input (R/W)
  4000300h  4  POSTFLG - Undoc
  4000304h  2  POWCNT1 - Graphics Power Control Register (R/W)

ARM9 3D Display Engine

  4000320h..6A3h

DS 3D I/O Map
ARM9 Display Engine B

  4001000h  4    2D Engine B - DISPCNT - LCD Control (Read/Write)
  4001008h  50h  2D Engine B (same registers as GBA, some changed bits)
  400106Ch  2    2D Engine B - MASTER_BRIGHT - 16bit - Brightness Up/Down

ARM9 DSi Extra Registers

  40021Axh  ..  DSi Registers
  4004xxxh  ..  DSi Registers

ARM9 IPC/ROM

  4100000h  4    IPCFIFORECV - IPC Receive Fifo (R)
  4100010h  4    Gamecard bus 4-byte data in, for manual or dma read (R) (or W)

ARM9 DS Debug Registers (Emulator/Devkits)

  4FFF0xxh  ..   Ensata Emulator Debug Registers
  4FFFAxxh  ..   No$gba Emulator Debug Registers

ARM9 Hardcoded RAM Addresses for Exception Handling

  27FFD9Ch   ..  NDS9 Debug Stacktop / Debug Vector (0=None)
  DTCM+3FF8h 4   NDS9 IRQ Check Bits (hardcoded RAM address)
  DTCM+3FFCh 4   NDS9 IRQ Handler (hardcoded RAM address)

Main Memory Control

  27FFFFEh  2    Main Memory Control

Further Memory Control Registers
ARM CP15 System Control Coprocessor

ARM7 I/O Map

  4000004h  2   DISPSTAT
  4000006h  2   VCOUNT
  40000B0h  30h DMA Channels 0..3
  4000100h  10h Timers 0..3
  4000120h  4   Debug SIODATA32
  4000128h  4   Debug SIOCNT
  4000130h  2   KEYINPUT
  4000132h  2   KEYCNT
  4000134h  2   Debug RCNT
  4000136h  2   EXTKEYIN
  4000138h  1   RTC Realtime Clock Bus
  4000180h  2   IPCSYNC - IPC Synchronize Register (R/W)
  4000184h  2   IPCFIFOCNT - IPC Fifo Control Register (R/W)
  4000188h  4   IPCFIFOSEND - IPC Send Fifo (W)
  40001A0h  2   AUXSPICNT - Gamecard ROM and SPI Control
  40001A2h  2   AUXSPIDATA - Gamecard SPI Bus Data/Strobe
  40001A4h  4   Gamecard bus timing/control
  40001A8h  8   Gamecard bus 8-byte command out
  40001B0h  4   Gamecard Encryption Seed 0 Lower 32bit
  40001B4h  4   Gamecard Encryption Seed 1 Lower 32bit
  40001B8h  2   Gamecard Encryption Seed 0 Upper 7bit (bit7-15 unused)
  40001BAh  2   Gamecard Encryption Seed 1 Upper 7bit (bit7-15 unused)
  40001C0h  2   SPI bus Control (Firmware, Touchscreen, Powerman)
  40001C2h  2   SPI bus Data

ARM7 Memory and IRQ Control

  4000204h  2   EXMEMSTAT - External Memory Status
  4000206h  2   WIFIWAITCNT
  4000208h  4   IME - Interrupt Master Enable (R/W)
  4000210h  4   IE  - Interrupt Enable (R/W)
  4000214h  4   IF  - Interrupt Request Flags (R/W)
  4000218h  -   IE2  ;\DSi only (additional ARM7 interrupt sources)
  400021Ch  -   IF2  ;/
  4000240h  1   VRAMSTAT - VRAM-C,D Bank Status (R)
  4000241h  1   WRAMSTAT - WRAM Bank Status (R)
  4000300h  1   POSTFLG
  4000301h  1   HALTCNT (different bits than on GBA) (plus NOP delay)
  4000304h  2   POWCNT2  Sound/Wifi Power Control Register (R/W)
  4000308h  4   BIOSPROT - Bios-data-read-protection address

ARM7 Sound Registers

  4000400h 100h Sound Channel 0..15 (10h bytes each)
  40004x0h  4   SOUNDxCNT - Sound Channel X Control Register (R/W)
  40004x4h  4   SOUNDxSAD - Sound Channel X Data Source Register (W)
  40004x8h  2   SOUNDxTMR - Sound Channel X Timer Register (W)
  40004xAh  2   SOUNDxPNT - Sound Channel X Loopstart Register (W)
  40004xCh  4   SOUNDxLEN - Sound Channel X Length Register (W)
  4000500h  2   SOUNDCNT - Sound Control Register (R/W)
  4000504h  2   SOUNDBIAS - Sound Bias Register (R/W)
  4000508h  1   SNDCAP0CNT - Sound Capture 0 Control Register (R/W)
  4000509h  1   SNDCAP1CNT - Sound Capture 1 Control Register (R/W)
  4000510h  4   SNDCAP0DAD - Sound Capture 0 Destination Address (R/W)
  4000514h  2   SNDCAP0LEN - Sound Capture 0 Length (W)
  4000518h  4   SNDCAP1DAD - Sound Capture 1 Destination Address (R/W)
  400051Ch  2   SNDCAP1LEN - Sound Capture 1 Length (W)

ARM7 DSi Extra Registers

  40021Axh  ..  DSi Registers
  4004xxxh  ..  DSi Registers
  4004700h  2   DSi SNDEXCNT Register  ;\mapped even in DS mode
  4004C0xh  ..  DSi GPIO Registers     ;/

ARM7 IPC/ROM

  4100000h  4   IPCFIFORECV - IPC Receive Fifo (R)
  4100010h  4   Gamecard bus 4-byte data in, for manual or dma read (R) (or W)

ARM7 3DS

  4700000h  4   Disable ARM7 bootrom overlay (W) (3DS only)

ARM7 WLAN Registers

  4800000h  ..  Wifi WS0 Region (32K) (Wifi Ports, and 8K Wifi RAM)
  4808000h  ..  Wifi WS1 Region (32K) (mirror of above, other waitstates)

ARM7 Hardcoded RAM Addresses for Exception Handling

  380FFC0h  4   DSi7 IRQ IF2 Check Bits (hardcoded RAM address) (DSi only)
  380FFDCh  ..  NDS7 Debug Stacktop / Debug Vector (0=None)
  380FFF8h  4   NDS7 IRQ IF Check Bits (hardcoded RAM address)
  380FFFCh  4   NDS7 IRQ Handler (hardcoded RAM address)

DS Memory Maps

NDS9 Memory Map

  00000000h  Instruction TCM (32KB) (not moveable) (mirror-able to 1000000h)
  0xxxx000h  Data TCM        (16KB) (moveable)
  02000000h  Main Memory     (4MB)
  03000000h  Shared WRAM     (0KB, 16KB, or 32KB can be allocated to ARM9)
  04000000h  ARM9-I/O Ports
  05000000h  Standard Palettes (2KB) (Engine A BG/OBJ, Engine B BG/OBJ)
  06000000h  VRAM - Engine A, BG VRAM  (max 512KB)
  06200000h  VRAM - Engine B, BG VRAM  (max 128KB)
  06400000h  VRAM - Engine A, OBJ VRAM (max 256KB)
  06600000h  VRAM - Engine B, OBJ VRAM (max 128KB)
  06800000h  VRAM - "LCDC"-allocated (max 656KB)
  07000000h  OAM (2KB) (Engine A, Engine B)
  08000000h  GBA Slot ROM (max 32MB)
  0A000000h  GBA Slot RAM (max 64KB)
  FFFF0000h  ARM9-BIOS (32KB) (only 3K used)

The ARM9 Exception Vectors are located at FFFF0000h. The IRQ handler redirects to [DTCM+3FFCh].

NDS7 Memory Map

  00000000h  ARM7-BIOS (16KB)
  02000000h  Main Memory (4MB)
  03000000h  Shared WRAM (0KB, 16KB, or 32KB can be allocated to ARM7)
  03800000h  ARM7-WRAM (64KB)
  04000000h  ARM7-I/O Ports
  04800000h  Wireless Communications Wait State 0 (8KB RAM at 4804000h)
  04808000h  Wireless Communications Wait State 1 (I/O Ports at 4808000h)
  06000000h  VRAM allocated as Work RAM to ARM7 (max 256K)
  08000000h  GBA Slot ROM (max 32MB)
  0A000000h  GBA Slot RAM (max 64KB)

The ARM7 Exception Vectors are located at 00000000h. The IRQ handler redirects to [3FFFFFCh aka 380FFFCh].

Further Memory (not mapped to ARM9/ARM7 bus)

  3D Engine Polygon RAM (52KBx2)
  3D Engine Vertex RAM (72KBx2)
  Firmware (256KB) (built-in serial flash memory)
  GBA-BIOS (16KB) (not used in NDS mode)
  NDS Slot ROM (serial 8bit-bus, max 4GB with default protocol)
  NDS Slot FLASH/EEPROM/FRAM (serial 1bit-bus)

Shared-RAM
Even though Shared WRAM begins at 3000000h, programs are commonly using mirrors at 37F8000h (both ARM9 and ARM7). At the ARM7-side, this allows to use 32K Shared WRAM and 64K ARM7-WRAM as a continous 96K RAM block.

Undefined I/O Ports
On the NDS (at the ARM9-side at least) undefined I/O ports are always zero.

Undefined Memory Regions
16MB blocks that do not contain any defined memory regions (or that contain only mapped TCM regions) are typically completely undefined.
16MB blocks that do contain valid memory regions are typically containing mirrors of that memory in the unused upper part of the 16MB area (only exceptions are TCM and BIOS which are not mirrored).

DS Memory Control

Memory Control
DS Memory Control - Cache and TCM
DS Memory Control - Cartridges and Main RAM
DS Memory Control - WRAM
DS Memory Control - VRAM
DS Memory Control - BIOS

Memory Access Time
DS Memory Timings

DS Memory Control - Cache and TCM

TCM and Cache are controlled by the System Control Coprocessor,
ARM CP15 System Control Coprocessor

The specifications for the NDS9 are:

Tightly Coupled Memory (TCM)

  ITCM 32K, base=00000000h (fixed, not move-able)
  DTCM 16K, base=moveable  (default base=27C0000h)

Note: Although ITCM is NOT moveable, the NDS Firmware configures the ITCM size to 32MB, and so, produces ITCM mirrors at 0..1FFFFFFh. Furthermore, the PU can be used to lock/unlock memory in that region. That trick allows to move ITCM anywhere within the lower 32MB of memory.

Cache

  Data Cache 4KB, Instruction Cache 8KB
  4-way set associative method
  Cache line 8 words (32 bytes)
  Read-allocate method (ie. writes are not allocating cache lines)
  Round-robin and Pseudo-random replacement algorithms selectable
  Cache Lockdown, Instruction Prefetch, Data Preload
  Data write-through and write-back modes selectable

Protection Unit (PU)
Recommended/default settings are:

  Region  Name            Address   Size   Cache WBuf Code Data
  -       Background      00000000h 4GB    -     -    -    -
  0       I/O and VRAM    04000000h 64MB   -     -    R/W  R/W
  1       Main Memory     02000000h 4MB    On    On   R/W  R/W
  2       ARM7-dedicated  027C0000h 256KB  -     -    -    -
  3       GBA Slot        08000000h 128MB  -     -    -    R/W
  4       DTCM            027C0000h 16KB   -     -    -    R/W
  5       ITCM            01000000h 32KB   -     -    R/W  R/W
  6       BIOS            FFFF0000h 32KB   On    -    R    R
  7       Shared Work     027FF000h 4KB    -     -    -    R/W

Notes: In Nintendo's hardware-debugger, Main Memory is expanded to 8MB (for that reason, some addresses are at 27NN000h instead 23NN000h) (some of the extra memory is reserved for the debugger, some can be used for game development). Region 2 and 7 are not understood? GBA Slot should be max 32MB+64KB, rounded up to 64MB, no idea why it is 128MB? DTCM and ITCM do not use Cache and Write-Buffer because TCM is fast. Above settings do not allow to access Shared Memory at 37F8000h? Do not use cache/wbuf for I/O, doing so might suppress writes, and/or might read outdated values.
The main purpose of the Protection Unit is debugging, a major problem with GBA programs have been faulty accesses to memory address 00000000h and up (due to [base+offset] addressing with uninitialized (zero) base values). This problem has been fixed in the NDS, for the ARM9 processor at least, still there are various leaks: For example, the 64MB I/O and VRAM area contains only ca. 660KB valid addresses, and the ARM7 probably doesn't have a Protection Unit at all. Alltogether, the protection is better than in GBA, but it's still pretty crude compared with software debugging tools.
Region address/size are unified (same for code and data), however, cachabilty and access rights are non-unified (and may be separately defined for code and data).

Note: The NDS7 doesn't have any TCM, Cache, or CP15.

DS Memory Control - Cartridges and Main RAM

4000204h - NDS9 - EXMEMCNT - 16bit - External Memory Control (R/W)
4000204h - NDS7 - EXMEMSTAT - 16bit - External Memory Status (R/W..R)

  0-1   32-pin GBA Slot SRAM Access Time    (0-3 = 10, 8, 6, 18 cycles)
  2-3   32-pin GBA Slot ROM 1st Access Time (0-3 = 10, 8, 6, 18 cycles)
  4     32-pin GBA Slot ROM 2nd Access Time (0-1 = 6, 4 cycles)
  5-6   32-pin GBA Slot PHI-pin out   (0-3 = Low, 4.19MHz, 8.38MHz, 16.76MHz)
  7     32-pin GBA Slot Access Rights     (0=ARM9, 1=ARM7)
  8-10  Not used (always zero)
  11    17-pin NDS Slot Access Rights     (0=ARM9, 1=ARM7)
  12    Not used (always zero)
  13    NDS:Always set?  ;set/tested by DSi bootcode: Main RAM enable, CE2 pin?
  14    Main Memory Interface Mode Switch (0=Async/GBA/Reserved, 1=Synchronous)
  15    Main Memory Access Priority       (0=ARM9 Priority, 1=ARM7 Priority)

Bit0-6 can be changed by both NDS9 and NDS7, changing these bits affects the local EXMEM register only, not that of the other CPU.
Bit7-15 can be changed by NDS9 only, changing these bits affects both EXMEM registers, ie. both NDS9 and NDS7 can read the current NDS9 setting.
Bit14=0 is intended for GBA mode, however, writes to this bit appear to be ignored?
DS Main Memory Control

GBA Slot (8000000h-AFFFFFFh)
The GBA Slot can be mapped to ARM9 or ARM7 via EXMEMCNT.7.
For the selected CPU, memory at 8000000h-9FFFFFFh contains the "GBA ROM" region, and memory at A000000h-AFFFFFFh contains the "GBA SRAM" region (repeated every 64Kbytes). If there is no cartridge in GBA Slot, then the ROM/SRAM regions will contain open-bus values: SRAM region is FFh-filled (High-Z). And ROM region is filled by increasing 16bit values (Addr/2), possibly ORed with garbage depending on the selected ROM Access Time:

  6 clks   --> returns "Addr/2"
  8 clks   --> returns "Addr/2"
  10 clks  --> returns "Addr/2 OR FE08h" (or similar garbage)
  18 clks  --> returns "FFFFh" (High-Z)

For the deselected CPU, all memory at 8000000h-AFFFFFFh becomes 00h-filled, this is required for bugged games like Digimon Story: Super Xros Wars (which is accidently reading deselected GBA SRAM at [main_ram_base+main_ram_addr*4], whereas it does presumably want to read Main RAM at [main_ram_base+index*4]).

DS Memory Control - WRAM

4000247h - NDS9 - WRAMCNT - 8bit - WRAM Bank Control (R/W)
4000241h - NDS7 - WRAMSTAT - 8bit - WRAM Bank Status (R)
Should not be changed when using Nintendo's API.

  0-1   ARM9/ARM7 (0-3 = 32K/0K, 2nd 16K/1st 16K, 1st 16K/2nd 16K, 0K/32K)
  2-7   Not used

The ARM9 WRAM area is 3000000h-3FFFFFFh (16MB range).
The ARM7 WRAM area is 3000000h-37FFFFFh (8MB range).
The allocated 16K or 32K are mirrored everywhere in the above areas.
De-allocation (0K) is a special case: At the ARM9-side, the WRAM area is then empty (containing undefined data). At the ARM7-side, the WRAM area is then containing mirrors of the 64KB ARM7-WRAM (the memory at 3800000h and up).

DS Memory Control - VRAM

4000240h - NDS7 - VRAMSTAT - 8bit - VRAM Bank Status (R)

  0     VRAM C enabled and allocated to NDS7  (0=No, 1=Yes)
  1     VRAM D enabled and allocated to NDS7  (0=No, 1=Yes)
  2-7   Not used (always zero)

The register indicates if VRAM C/D are allocated to NDS7 (as Work RAM), ie. if VRAMCNT_C/D are enabled (Bit7=1), with MST=2 (Bit0-2). However, it does not reflect the OFS value.

4000240h - NDS9 - VRAMCNT_A - 8bit - VRAM-A (128K) Bank Control (W)
4000241h - NDS9 - VRAMCNT_B - 8bit - VRAM-B (128K) Bank Control (W)
4000242h - NDS9 - VRAMCNT_C - 8bit - VRAM-C (128K) Bank Control (W)
4000243h - NDS9 - VRAMCNT_D - 8bit - VRAM-D (128K) Bank Control (W)
4000244h - NDS9 - VRAMCNT_E - 8bit - VRAM-E (64K) Bank Control (W)
4000245h - NDS9 - VRAMCNT_F - 8bit - VRAM-F (16K) Bank Control (W)
4000246h - NDS9 - VRAMCNT_G - 8bit - VRAM-G (16K) Bank Control (W)
4000248h - NDS9 - VRAMCNT_H - 8bit - VRAM-H (32K) Bank Control (W)
4000249h - NDS9 - VRAMCNT_I - 8bit - VRAM-I (16K) Bank Control (W)

  0-2   VRAM MST              ;Bit2 not used by VRAM-A,B,H,I
  3-4   VRAM Offset (0-3)     ;Offset not used by VRAM-E,H,I
  5-6   Not used
  7     VRAM Enable (0=Disable, 1=Enable)

There is a total of 656KB of VRAM in Blocks A-I.
Table below shows the possible configurations.

  VRAM    SIZE  MST  OFS   ARM9, Plain ARM9-CPU Access (so-called LCDC mode)
  A       128K  0    -     6800000h-681FFFFh
  B       128K  0    -     6820000h-683FFFFh
  C       128K  0    -     6840000h-685FFFFh
  D       128K  0    -     6860000h-687FFFFh
  E       64K   0    -     6880000h-688FFFFh
  F       16K   0    -     6890000h-6893FFFh
  G       16K   0    -     6894000h-6897FFFh
  H       32K   0    -     6898000h-689FFFFh
  I       16K   0    -     68A0000h-68A3FFFh
  VRAM    SIZE  MST  OFS   ARM9, 2D Graphics Engine A, BG-VRAM (max 512K)
  A,B,C,D 128K  1    0..3  6000000h+(20000h*OFS)
  E       64K   1    -     6000000h
  F,G     16K   1    0..3  6000000h+(4000h*OFS.0)+(10000h*OFS.1)
  VRAM    SIZE  MST  OFS   ARM9, 2D Graphics Engine A, OBJ-VRAM (max 256K)
  A,B     128K  2    0..1  6400000h+(20000h*OFS.0)  ;(OFS.1 must be zero)
  E       64K   2    -     6400000h
  F,G     16K   2    0..3  6400000h+(4000h*OFS.0)+(10000h*OFS.1)
  VRAM    SIZE  MST  OFS   2D Graphics Engine A, BG Extended Palette
  E       64K   4    -     Slot 0-3  ;only lower 32K used
  F,G     16K   4    0..1  Slot 0-1 (OFS=0), Slot 2-3 (OFS=1)
  VRAM    SIZE  MST  OFS   2D Graphics Engine A, OBJ Extended Palette
  F,G     16K   5    -     Slot 0  ;16K each (only lower 8K used)
  VRAM    SIZE  MST  OFS   Texture/Rear-plane Image
  A,B,C,D 128K  3    0..3  Slot OFS(0-3)   ;(Slot2-3: Texture, or Rear-plane)
  VRAM    SIZE  MST  OFS   Texture Palette
  E       64K   3    -     Slots 0-3                 ;OFS=don't care
  F,G     16K   3    0..3  Slot (OFS.0*1)+(OFS.1*4)  ;ie. Slot 0, 1, 4, or 5
  VRAM    SIZE  MST  OFS   ARM9, 2D Graphics Engine B, BG-VRAM (max 128K)
  C       128K  4    -     6200000h
  H       32K   1    -     6200000h
  I       16K   1    -     6208000h
  VRAM    SIZE  MST  OFS   ARM9, 2D Graphics Engine B, OBJ-VRAM (max 128K)
  D       128K  4    -     6600000h
  I       16K   2    -     6600000h
  VRAM    SIZE  MST  OFS   2D Graphics Engine B, BG Extended Palette
  H       32K   2    -     Slot 0-3
  VRAM    SIZE  MST  OFS   2D Graphics Engine B, OBJ Extended Palette
  I       16K   3    -     Slot 0  ;(only lower 8K used)
  VRAM    SIZE  MST  OFS   <ARM7>, Plain <ARM7>-CPU Access
  C,D     128K  2    0..1  6000000h+(20000h*OFS.0)  ;OFS.1 must be zero

Notes
In Plain-CPU modes, VRAM can be accessed only by the CPU (and by the Capture Unit, and by VRAM Display mode). In "Plain <ARM7>-CPU Access" mode, the VRAM blocks are allocated as Work RAM to the NDS7 CPU.
In BG/OBJ VRAM modes, VRAM can be accessed by the CPU at specified addresses, and by the display controller.
In Extended Palette and Texture Image/Palette modes, VRAM is not mapped to CPU address space, and can be accessed only by the display controller (so, to initialize or change the memory, it should be temporarily switched to Plain-CPU mode).
All VRAM (and Palette, and OAM) can be written to only in 16bit and 32bit units (STRH, STR opcodes), 8bit writes are ignored (by STRB opcode). The only exception is "Plain <ARM7>-CPU Access" mode: The ARM7 CPU can use STRB to write to VRAM (the reason for this special feature is that, in GBA mode, two 128K VRAM blocks are used to emulate the GBA's 256K Work RAM).

Other Video RAM
Aside from the map-able VRAM blocks, there are also some video-related memory regions at fixed addresses:

  5000000h Engine A Standard BG Palette (512 bytes)
  5000200h Engine A Standard OBJ Palette (512 bytes)
  5000400h Engine B Standard BG Palette (512 bytes)
  5000600h Engine B Standard OBJ Palette (512 bytes)
  7000000h Engine A OAM (1024 bytes)
  7000400h Engine B OAM (1024 bytes)

DS Memory Control - BIOS

4000308h - NDS7 - BIOSPROT - Bios-data-read-protection address
Used to double-protect the first some KBytes of the NDS7 BIOS. The BIOS is split into two protection regions, one always active, one controlled by the BIOSPROT register. The overall idea is that only the BIOS can read from itself, any other attempts to read from that regions return FFh-bytes.

  Opcodes at...      Can read from      Expl.
  0..[BIOSPROT]-1    0..3FFFh           Double-protected (when BIOSPROT is set)
  [BIOSPROT]..3FFFh  [BIOSPROT]..3FFFh  Normal-protected (always active)

The initial BIOSPROT setting on power-up is zero (disabled). Before starting the cartridge, the BIOS boot code sets the register to 1204h (actually 1205h, but the mis-aligned low-bit is ignored). Once when initialized, further writes to the register are ignored.

The double-protected region contains the exception vectors, some bytes of code, and the cartridge KEY1 encryption seed (about 4KBytes). As far as I know, it is impossible to unlock the memory once when it is locked, however, with some trickery, it is possible execute code before it gets locked. Also, the two THUMB opcodes at 05ECh can be used to read all memory at 0..3FFFh,

  05ECh  ldrb r3,[r3,12h]      ;requires incoming r3=src-12h
  05EEh  pop  r2,r4,r6,r7,r15  ;requires dummy values & THUMB retadr on stack

Additionally most BIOS functions (eg. CpuSet), include a software-based protection which rejects source addresses in the BIOS area (the only exception is GetCRC16, though it still cannot bypass the BIOSPROT setting).

Note
The NDS9 BIOS doesn't include any software or hardware based read protection.

DS Memory Timings

System Clock

  Bus clock  = 33MHz (33.513982 MHz) (1FF61FEh Hertz)
  NDS7 clock = 33MHz (same as bus clock)
  NDS9 clock = 66MHz (internally twice bus clock; for cache/tcm)

Most timings in this document are specified for 33MHz clock (not for the 66MHz clock). Respectively, NDS9 timings are counted in "half" cycles.

Memory Access Times
Tables below show the different access times for code/data fetches on arm7/arm9 cpus, measured for sequential/nonsequential 32bit/16bit accesses.

  NDS7/CODE             NDS9/CODE
  N32 S32 N16 S16 Bus   N32 S32 N16 S16 Bus
  9   2   8   1   16    9   9   4.5 4.5 16  Main RAM (read) (cache off)
  1   1   1   1   32    4   4   2   2   32  WRAM,BIOS,I/O,OAM
  2   2   1   1   16    5   5   2.5 2.5 16  VRAM,Palette RAM
  16  12  10  6   16    19  19  9.5 9.5 16  GBA ROM (example 10,6 access)
  -   -   -   -   -     0.5 0.5 0.5 0.5 32  TCM, Cache_Hit
  -   -   -   -   -     (--Load 8 words--)  Cache_Miss

  NDS7/DATA             NDS9/DATA
  N32 S32 N16 S16 Bus   N32 S32 N16 S16 Bus
  10  2   9   1   16    10  2   9   1   16  Main RAM (read) (cache off)
  1   1   1   1   32    4   1   4   1   32  WRAM,BIOS,I/O,OAM
  1?  2   1   1   16    5   2   4   1   16  VRAM,Palette RAM
  15  12  9   6   16    19  12  13  6   16  GBA ROM (example 10,6 access)
  9   10  9   10  8     13  10  13  10  8   GBA RAM (example 10 access)
  -   -   -   -   -     0.5 0.5 0.5 -   32  TCM, Cache_Hit
  -   -   -   -   -     (--Load 8 words--)  Cache_Miss
  -   -   -   -   -     11  11  11  -   32  Cache_Miss (BIOS)
  -   -   -   -   -     23  23  23  -   16  Cache_Miss (Main RAM)

All timings are counted in 33MHz units (so "half" cycles can occur on NDS9).
Note: 8bit data accesses have same timings than 16bit data.

*** DS Memory Timing Notes ***

The NDS timings are altogether pretty messed up, with different timings for CODE and DATA fetches, and different timings for NDS7 and NDS9...

NDS7/CODE
Timings for this region can be considered as "should be" timings.

NDS7/DATA
Quite the same as NDS7/CODE. Except that, nonsequential Main RAM accesses are 1 cycle slower, and more strange, nonsequential GBA Slot accesses are 1 cycle faster.

NDS9/CODE
This is the most messiest timing. An infamous PENALTY of 3 cycles is added to all nonsequential accesses (except cache, tcm, and main ram). And, all opcode fetches are forcefully made nonsequential 32bit (the NDS9 simply doesn't support fast sequential opcode fetches). That applies also for THUMB code (two 16bit opcodes are fetched by a single nonsequential 32bit access) (so the time per 16bit opcode is one half of the 32bit fetch) (unless a branch causes only one of the two 16bit opcodes to be executed, then that opcode will have the full 32bit access time).

NDS9/DATA
Allows both sequential and nonsequential access, and both 16bit and 32bit access, so it's faster than NDS9/CODE. Nethertheless, it's still having the 3 cycle PENALTY on nonsequential accesses. And, similar as NDS7/DATA, it's also adding 1 cycle to nonsequential Main RAM accesses.

*** More Timing Notes / Lots of unsorted Info ***

Actual CPU Performance
The 33MHz NDS7 is running more or less nicely at 33MHz. However, the so-called "66MHz" NDS9 is having <much> higher waitstates, and it's effective bus speed is barely about 8..16MHz, the only exception is code/data in cache/tcm, which is eventually reaching real 66MHz (that, assuming cache HITS, otherwise, in case of cache MISSES, the cached memory timing might even drop to 1.4MHz or so?).
*********************
ARM9 opcode fetches are always N32 + 3 waits.

  S16 and N16 do not exist (because thumb-double-fetching) (see there).
  S32 becomes N32 (ie. the ARM9 does NOT support fast sequential timing).

That N32 is having same timing as normal N32 access on NDS7, plus 3 waits.

  Eg. an ARM9 N32 or S32 to 16bit bus will take: N16 + S16 + 3 waits.
  Eg. an ARM9 N32 or S32 to 32bit bus will take: N32 + 3 waits.

Main Memory is ALWAYS having the nonsequential 3 wait PENALTY (even on ARM7).
*********************
ARM9 Data fetches however are allowed to use sequential timing, as well as raw 16bit accesses (which aren't forcefully expanded to slow 32bit accesses).
Nethertheless, the 3 wait PENALTY is added to any NONSEQUENTIAL accesses.
Only exceptions are cache and tcm which do not have that penalty.

 Eg. LDRH on 16bit-data-bus is N16+3waits.
 Eg. LDR  on 16bit-data-bus is N16+S16+3waits.
 Eg. LDM  on 16bit-data-bus is N16+(n*2-1)*S16+3waits.

Eventually, data fetches can take place parallel with opcode fetches.

 That is NOT true for LDM (works only for LDR/LDRB/LDRH).
 That is NOT true for DATA in SAME memory region than CODE.
 That is NOT true for DATA in ITCM (no matter if CODE is in ITCM).

*********************

NDS9 Busses
Unlike ARM7, the ARM9 has separate code and data busses, allowing it to perform code and data fetches simultaneously (provided that both are in different memory regions).
Normally, opcode execution times are calculated as "(codetime+datatime)", with the two busses, it can (ideally) be "MAX(codetime,datatime)", so the data access time may virtually take "NULL" clock cycles.
In practice, DTCM and Data Cache access can take NULL cycles (however, data access to ITCM can't).
When executing code in cache/itcm, data access to non-cache/tcm won't be any faster than with only one bus (as it's best, it could subtract 0.5 cycles from datatime, but, the access must be "aligned" to the bus-clock, so the "datatime-0.5" will be rounded back to the original "datatime").
When executing code in uncached main ram, and accessing data (elsewhere than in main memory, cache/tcm), then execution time is typically "codetime+datatime-2".

NDS9 Internal Cycles
Additionally to codetime+datatime, some opcodes include one or more internal cycles. Compared with ARM7, the behaviour of that internal cycles is slightly different on ARM9. First of, on the NDS9, the internal cycles are of course "half" cycles (ie. counted in 66MHz units, not in 33MHz units) (although they may get rounded to "full" cycles upon next memory access outside tcm/cache). And, the ARM9 is in some cases "skipping" the internal cycles, that often depending on whether or not the next opcode is using the result of the current opcode.
Another big difference is that the ARM9 has lost the fast-multiply feature for small numbers; in some cases that may result in faster execution, but may also result in slower execution (one workaround would be to manually replace MUL opcodes by the new ARM9 halfword multiply opcodes); the slowest case are MUL opcodes that do update flags (eg. MULS, MLAS, SMULLS, etc. in ARM mode, and all ALL multiply opcodes in THUMB mode).

NDS9 Thumb Code
In thumb mode, the NDS9 is fetching two 16bit opcodes by a single 32bit read. In case of 32bit bus, this reduces the amount of memory traffic and may result in faster execution time, of course that works only if the two opcodes are within a word-aligned region (eg. loops at word-aligned addresses will be faster than non-aligned loops). However, the double-opcode-fetching is also done on 16bit bus memory, including for unnecessary fetches, such like opcodes after branch commands, so the feature may cause heavy slowdowns.

Main Memory
Reportedly, the main memory access times would be 5 cycles (nonsequential read), 4 cycles (nonsequential write), and 1 cycle (sequential read or write). Plus whatever termination cycles. Plus 3 cycles on nonsequential access to the last 2-bytes of a 32-byte block.
That's of course all wrong. Reads are much slower than 5 cycles. Not yet tested if writes are faster. And, I haven't been able to reproduce the 3 cycles on last 2-bytes effect, actually, it looks more as if that 3 cycles are accidently added to ALL nonsequential accesses, at ALL main memory addresses, and even to most OTHER memory regions... which might be the source of the PENALTY which occurs on VRAM/WRAM/OAM/Palette and I/O accesses.

DMA
In some cases DMA main memory read cycles are reportedly performed simultaneously with DMA write cycles to other memory.

NDS9
On the NDS9, all external memory access (and I/O) is delayed to bus clock (or actually MUCH slower due to the massive waitstates), so the full 66MHz can be used only internally in the NDS9 CPU core, ie. with cache and TCM.

Bus Clock
The exact bus clock is specified as 33.513982 MHz (1FF61FEh Hertz). However, on my own NDS, measured in relation to the RTC seconds IRQ, it appears more like 1FF6231h, that inaccuary of 1 cycle per 657138 cycles (about one second per week) on either oscillator, isn't too significant though.

GBA Slot
The access time for GBA slot can be configured via EXMEMCNT register.

VRAM Waitstates
Additionally, on NDS9, a one cycle wait can be added to VRAM accesses (when the video controller simultaneously accesses it) (that can be disabled by Forced Blank, see DISPCNT.Bit7). Moreover, additional VRAM waitstates occur when using the video capture function.
Note: VRAM being mapped to NDS7 is always free of additional waits.

DS Video

The NDS has two 2D Video Engines, each basically the same as in GBA, see
GBA LCD Video Controller

NDS Specific 2D Video Features
DS Video Stuff
DS Video BG Modes / Control
DS Video OBJs
DS Video Extended Palettes
DS Video Capture and Main Memory Display Mode
DS Video Display System Block Diagram

NDS/DSi File Formats for 2D video
DS Files - 2D Video

For Display Power Control (and Display Swap), and VRAM Allocation, see
DS Power Control
DS Power Management Device
DS Memory Control - VRAM

DS Video Stuff

DS Display Dimensions / Timings
Dot clock = 5.585664 MHz (=33.513982 MHz / 6)
H-Timing: 256 dots visible, 99 dots blanking, 355 dots total (15.7343KHz)
V-Timing: 192 lines visible, 71 lines blanking, 263 lines total (59.8261 Hz)
The V-Blank cycle for the 3D Engine consists of the 23 lines, 191..213.
Screen size 62.5mm x 47.0mm (each) (256x192 pixels)
Vertical space between screens 22mm (equivalent to 90 pixels)

400006Ch - NDS9 - MASTER_BRIGHT - 16bit - Master Brightness Up/Down

  0-4   Factor used for 6bit R,G,B Intensities (0-16, values >16 same as 16)
          Brightness up:   New = Old + (63-Old) * Factor/16
          Brightness down: New = Old - Old      * Factor/16
  5-13  Not used
  14-15 Mode (0=Disable, 1=Up, 2=Down, 3=Reserved)
  16-31 Not used

DISPSTAT/VCOUNT
The LY and LYC values are in range 0..262, so LY/LYC values have been expanded to 9bit values: LY = VCOUNT Bit 0..8, and LYC=DISPSTAT Bit8..15,7.
VCOUNT register is write-able, allowing to synchronize linked DS consoles.
For proper synchronization:

  write new LY values only in range of 202..212
  write only while old LY values are in range of 202..212

DISPSTAT/VCOUNT supported by NDS9 (Engine A Ports, without separate Engine B Ports), and by NDS7 (allowing to synchronize NDS7 with display timings).
Similar as on GBA, the VBlank flag isn't set in the last line (ie. only in lines 192..261, but not in line 262).
Although the drawing time is only 1536 cycles (256*6), the NDS9 H-Blank flag is "0" for a total of 1606 cycles (and, for whatever reason, a bit longer, 1613 cycles in total, on NDS7).

VRAM Waitstates
The display controller performs VRAM-reads once every 6 clock cycles, a 1 cycle waitstate is generated if the CPU simultaneously accesses VRAM. With capture enabled, additionally VRAM-writes take place once every 6 cycles, so the total VRAM-read/write access rate is then once every 3 cycles.

DS Window Glitches
The DS counts scanlines in range 0..262 (0..106h), of which only the lower 8bit are compared with the WIN0V/WIN1V register settings. Respectively, Y1 coordinates 00h..06h will be triggered in scanlines 100h-106h by mistake. That means, the window gets activated within VBlank period, and will be active in scanline 0 and up (that is no problem with Y1=0, but Y1=1..6 will appear as if if Y1 would be 0). Workaround would be to disable the Window during VBlank, or to change Y1 during VBlank (to a value that does not occur during VBlank period, ie. 7..191).
Also, there's a problem to fit the 256 pixel horizontal screen resolution into 8bit values: X1=00h is treated as 0 (left-most), X2=00h is treated as 100h (right-most). However, the window is not displayed if X1=X2=00h; the window width can be max 255 pixels.

2D Engines
Includes two 2D Engines, called A and B. Both engines are accessed by the ARM9 processor, each using different memory and register addresses:

  Region______Engine A______________Engine B___________
  I/O Ports   4000000h              4001000h
  Palette     5000000h (1K)         5000400h (1K)
  BG VRAM     6000000h (max 512K)   6200000h (max 128K)
  OBJ VRAM    6400000h (max 256K)   6600000h (max 128K)
  OAM         7000000h (1K)         7000400h (1K)

Engine A additionally supports 3D and large-screen 256-color Bitmaps, plus main-memory-display and vram-display modes, plus capture unit.

Viewing Angles
The LCD screens are best viewed at viewing angles of 90 degrees. Colors may appear distorted, and may even become invisible at other viewing angles.
When the console is handheld, both screens can be turned into preferred direction. When the console is settled on a table, only the upper screen can be turned, but the lower screen is stuck into horizontal position - which results in rather bad visibility (unless the user moves his/her head directly above of it).

4000070h - NDS9 - TVOUTCNT - Unknown (W)

  Bit0-3  "COMMAND"  (?)
  Bit4-7  "COMMAND2" (?)
  Bit8-11 "COMMAND3" (?)

This register has been mentioned in an early I/O map from Nintendo, as far as I know, the register isn't used by any games/firmware/bios, not sure if it does really exist on release-version, or if it's been prototype stuff...?

DS-Lite Screens
The screens in the DS-Lite seem to allow a wider range of vertical angles.
The bad news is that the colors of the DS-Lite are (no surprise) not backwards compatible with older NDS and GBA displays. The good news is that Nintendo has finally reached near-CRT-quality (without blurred colors), so one could hope that they won't show up with more displays with other colors in future.
Don't know if there's an official/recommended way to detect DS-Lite displays (?) possible methods would be whatever values in Firmware header, or by functionality of Power Managment device, or (not too LCD-related) by Wifi Chip ID.

DS Video BG Modes / Control

4000000h - NDS9 - DISPCNT

  Bit  Engine Expl.
  0-2   A+B   BG Mode
  3     A     BG0 2D/3D Selection (instead CGB Mode) (0=2D, 1=3D)
  4     A+B   Tile OBJ Mapping        (0=2D; max 32KB, 1=1D; max 32KB..256KB)
  5     A+B   Bitmap OBJ 2D-Dimension (0=128x512 dots, 1=256x256 dots)
  6     A+B   Bitmap OBJ Mapping      (0=2D; max 128KB, 1=1D; max 128KB..256KB)
  7-15  A+B   Same as GBA
  16-17 A+B   Display Mode (Engine A: 0..3, Engine B: 0..1, GBA: Green Swap)
  18-19 A     VRAM block (0..3=VRAM A..D) (For Capture & above Display Mode=2)
  20-21 A+B   Tile OBJ 1D-Boundary   (see Bit4)
  22    A     Bitmap OBJ 1D-Boundary (see Bit5-6)
  23    A+B   OBJ Processing during H-Blank (was located in Bit5 on GBA)
  24-26 A     Character Base (in 64K steps) (merged with 16K step in BGxCNT)
  27-29 A     Screen Base (in 64K steps) (merged with 2K step in BGxCNT)
  30    A+B   BG Extended Palettes   (0=Disable, 1=Enable)
  31    A+B   OBJ Extended Palettes  (0=Disable, 1=Enable)

BG Mode
Engine A BG Mode (DISPCNT LSBs) (0-6, 7=Reserved)

  Mode  BG0      BG1      BG2      BG3
  0     Text/3D  Text     Text     Text
  1     Text/3D  Text     Text     Affine
  2     Text/3D  Text     Affine   Affine
  3     Text/3D  Text     Text     Extended
  4     Text/3D  Text     Affine   Extended
  5     Text/3D  Text     Extended Extended
  6     3D       -        Large    -

Of which, the "Extended" modes are sub-selected by BGxCNT bits:

  BGxCNT.Bit7 BGxCNT.Bit2 Extended Affine Mode Selection
  0           CharBaseLsb rot/scal with 16bit bgmap entries (Text+Affine mixup)
  1           0           rot/scal 256 color bitmap
  1           1           rot/scal direct color bitmap

Engine B: Same as above, except that: Mode 6 is reserved (no Large screen bitmap), and BG0 is always Text (no 3D support).
Affine = formerly Rot/Scal mode (with 8bit BG Map entries)
Large Screen Bitmap = rot/scal 256 color bitmap (using all 512K of 2D VRAM)

Display Mode (DISPCNT.16-17):

  0  Display off (screen becomes white)
  1  Graphics Display (normal BG and OBJ layers)
  2  Engine A only: VRAM Display (Bitmap from block selected in DISPCNT.18-19)
  3  Engine A only: Main Memory Display (Bitmap DMA transfer from Main RAM)

Mode 2-3 display a raw direct color bitmap (15bit RGB values, the upper bit in each halfword is unused), without any further BG,OBJ,3D layers, these modes are completely bypassing the 2D/3D engines as well as any 2D effects, however the Master Brightness effect can be applied to these modes. Mode 2 is particulary useful to display captured 2D/3D images (in that case it can indirectly use the 2D/3D engine).

BGxCNT
character base extended from bit2-3 to bit2-5 (bit4-5 formerly unused)

  engine A screen base: BGxCNT.bits*2K + DISPCNT.bits*64K
  engine B screen base: BGxCNT.bits*2K + 0
  engine A char base: BGxCNT.bits*16K + DISPCNT.bits*64K
  engine B char base: BGxCNT.bits*16K + 0

char base is used only in tile/map modes (not bitmap modes)
screen base is used in tile/map modes,
screen base used in bitmap modes as BGxCNT.bits*16K, without DISPCNT.bits*64K
screen base however NOT used at all for Large screen bitmap mode

  bgcnt size  text     rotscal    bitmap   large bmp
  0           256x256  128x128    128x128  512x1024
  1           512x256  256x256    256x256  1024x512
  2           256x512  512x512    512x256  -
  3           512x512  1024x1024  512x512  -

bitmaps that require more than 128K VRAM are supported on engine A only.

For BGxCNT.Bit7 and BGxCNT.Bit2 in Extended Affine modes, see above BG Mode description (extended affine doesn't include 16-color modes, so color depth bit can be used for mode selection. Also, bitmap modes do not use charbase, so charbase.0 can be used for mode selection as well).

  for BG0CNT, BG1CNT only: bit13 selects extended palette slot
                           (BG0: 0=Slot0, 1=Slot2, BG1: 0=Slot1, 1=Slot3)

Direct Color Bitmap BG, and Direct Color Bitmap OBJ
BG/OBJ Supports 32K colors (15bit RGB value) - so far same as GBAs BG.
However, the upper bit (Bit15) is used as Alpha flag. That is, Alpha=0=Transparent, Alpha=1=Normal (ie. on the NDS, Direct Color values 0..7FFFh are NOT displayed).

Unlike GBA bitmap modes, NDS bitmap modes are supporting the Area Overflow bit (BG2CNT and BG3CNT, Bit 13).

DS Video OBJs

DS OBJ Priority
The GBA has been assigning OBJ priority in respect to the 7bit OAM entry number, regardless of the OBJs 2bit BG-priority attribute (which allowed to specify invalid priority orders). That problem has been fixed in DS mode by combining the above two values into a 9bit priority value.

OBJ Tile Mapping (DISPCNT.4,20-21):

  Bit4  Bit20-21  Dimension Boundary Total ;Notes
  0     x         2D        32       32K   ;Same as GBA 2D Mapping
  1     0         1D        32       32K   ;Same as GBA 1D Mapping
  1     1         1D        64       64K
  1     2         1D        128      128K
  1     3         1D        256      256K  ;Engine B: 128K max

TileVramAddress = TileNumber * BoundaryValue
Even if the boundary gets changed, OBJs are kept composed of 8x8 tiles.

Bitmap OBJ Mapping (DISPCNT.6,5,22):
Bitmap OBJs are 15bit Direct Color data, plus 1bit Alpha flag (in bit15).

  Bit6 Bit5 Bit22 Dimension    Boundary   Total ;Notes
  0    0    x     2D/128 dots  8x8 dots   128K  ;Source Bitmap width 128 dots
  0    1    x     2D/256 dots  8x8 dots   128K  ;Source Bitmap width 256 dots
  1    0    0     1D           128 bytes  128K  ;Source Width = Target Width
  1    0    1     1D           256 bytes  256K  ;Engine A only
  1    1    x     Reserved

In 1D mapping mode, the Tile Number is simply multiplied by the boundary value.

  1D_BitmapVramAddress = TileNumber(0..3FFh) * BoundaryValue(128..256)
  2D_BitmapVramAddress = (TileNo AND MaskX)*10h + (TileNo AND NOT MaskX)*80h

In 2D mode, the Tile Number is split into X and Y indices, the X index is located in the LSBs (ie. MaskX=0Fh, or MaskX=1Fh, depending on DISPCNT.5).

OBJ Attribute 0 and 2
Setting the OBJ Mode bits (Attr 0, Bit10-11) to a value of 3 has been prohibited in GBA, however, in NDS it selects the new Bitmap OBJ mode; in that mode, the Color depth bit (Attr 0, Bit13) should be set to zero; also in that mode, the color bits (Attr 2, Bit 12-15) are used as Alpha-OAM value (instead of as palette setting).

OBJ Vertical Wrap
On the GBA, a large OBJ (with 64pix height, scaled into double-size region of 128pix height) located near the bottom of the screen has been wrapped to the top of the screen (and was NOT displayed at the bottom of the screen).
This problem has been "corrected" in the NDS (except in GBA mode), that is, on the NDS, the OBJ appears BOTH at the top and bottom of the screen. That isn't necessarily better - the advantage is that one can manually enable/disable the OBJ in the desired screen-half on IRQ level; that'd be required only if the wrapped portion is non-transparent.

DS Video Extended Palettes

Extended Palettes
When allocating extended palettes, the allocated memory is not mapped to the CPU bus, so the CPU can access extended palette only when temporarily de-allocating it.

Color 0 of all standard/extended palettes is transparent, color 0 of BG standard palette 0 is used as backdrop. extended palette memory must be allocated to VRAM.

BG Extended Palette enabled in DISPCNT Bit 30, when enabled,

 standard palette --> 16-color tiles (with 16bit bgmap entries) (text)
                      256-color tiles (with 8bit bgmap entries) (rot/scal)
                      256-color bitmaps
                      backdrop-color (color 0)
 extended palette --> 256-color tiles (with 16bit bgmap entries)(text,rot/scal)

Allocated VRAM is split into 4 slots of 8K each (32K used in total), normally BG0..3 are using Slot 0..3, however BG0 and BG1 can be optionally changed to BG0=Slot2, and BG1=Slot3 via BG0CNT and BG1CNT.

OBJ Extended Palette enabled in DISPCNT Bit 31, when enabled,

 16 colors x 16 palettes --> standard palette memory (=256 colors)
 256 colors x 16 palettes --> extended palette memory (=4096 colors)

Extended OBJ palette memory must be allocated to VRAM F, G, or I (which are 16K) of which only the first 8K are used for extended palettes (=1000h 16bit entries).

DS Video Capture and Main Memory Display Mode

4000064h - NDS9 - DISPCAPCNT - 32bit - Display Capture Control Register (R/W)
Capture is supported for Display Engine A only.

  0-4   EVA               (0..16 = Blending Factor for Source A)
  5-7   Not used
  8-12  EVB               (0..16 = Blending Factor for Source B)
  13-15 Not used
  16-17 VRAM Write Block  (0..3 = VRAM A..D) (VRAM must be allocated to LCDC)
  18-19 VRAM Write Offset (0=00000h, 0=08000h, 0=10000h, 0=18000h)
  20-21 Capture Size      (0=128x128, 1=256x64, 2=256x128, 3=256x192 dots)
  22-23 Not used
  24    Source A          (0=Graphics Screen BG+3D+OBJ, 1=3D Screen)
  25    Source B          (0=VRAM, 1=Main Memory Display FIFO)
  26-27 VRAM Read Offset  (0=00000h, 0=08000h, 0=10000h, 0=18000h)
  28    Not used
  29-30 Capture Source    (0=Source A, 1=Source B, 2/3=Sources A+B blended)
  31    Capture Enable    (0=Disable/Ready, 1=Enable/Busy)

Notes:
VRAM Read Block (VRAM A..D) is selected in DISPCNT Bits 18-19.
VRAM Read Block can be (or must be ?) allocated to LCDC (MST=0).
VRAM Read Offset is ignored (zero) in VRAM Display Mode (DISPCNT.16-17).
VRAM Read/Write Offsets wrap to 00000h when exceeding 1FFFFh (max 128K).
Capture Sizes less than 256x192 capture the upper-left portion of the screen.
Blending factors EVA and EVB are used only if "Source A+B blended" selected.
After setting the Capture Enable bit, capture starts at next line 0, and the capture enable/busy bit is then automatically cleared (in line 192, regardless of the capture size).

Capture data is 15bit color depth (even when capturing 18bit 3D-images).
Capture A: Dest_Intensity = SrcA_Intensitity ; Dest_Alpha=SrcA_Alpha.
Capture B: Dest_Intensity = SrcB_Intensitity ; Dest_Alpha=SrcB_Alpha.
Capture A+B (blending):

 Dest_Intensity = (  (SrcA_Intensitity * SrcA_Alpha * EVA)
                   + (SrcB_Intensitity * SrcB_Alpha * EVB) ) / 16
 Dest_Alpha = (SrcA_Alpha AND (EVA>0)) OR (SrcB_Alpha AND EVB>0))

Capture provides a couple of interesting effects.
For example, 3D Engine output can be captured via source A (to LCDC-allocated VRAM), in the next frame, either Graphics Engine A or B can display the captured 3D image in VRAM image as BG2, BG3, or OBJ (from BG/OBJ-allocated VRAM); this method requires to switch between LCDC- and BG/OBJ-allocation.
Another example would be to capture Engine A output, the captured image can be displayed (via VRAM Display mode) in the following frames, simultaneously the new Engine A output can be captured, blended with the old captured image; in that mode moved objects will leave traces on the screen; this method works with a single LCDC-allocated VRAM block.
DS Video Display System Block Diagram

4000068h - NDS9 - DISP_MMEM_FIFO - 32bit - Main Memory Display FIFO (R?/W)
Intended to send 256x192 pixel 32K color bitmaps by DMA directly

 - to Screen A             (set DISPCNT to Main Memory Display mode), or
 - to Display Capture unit (set DISPCAPCNT to Main Memory Source).

The FIFO can receive 4 words (8 pixels) at a time, each pixel is a 15bit RGB value (the upper bit, bit15, is unused).
Set DMA to Main Memory mode, 32bit transfer width, word count set to 4, destination address to DISP_MMEM_FIFO, source address must be in Main Memory.
Transfer starts at next frame.
Main Memory Display/Capture is supported for Display Engine A only.

DS Video Display System Block Diagram

             _____________               __________
  VRAM A -->| 2D Graphics |--------OBJ->|          |
  VRAM B -->| Engine A    |--------BG3->| Layering |
  VRAM C -->|             |--------BG2->| and      |
  VRAM D -->|             |--------BG1->| Special  |
  VRAM E -->|             |   ___       | Effects  |
  VRAM F -->|             |->|SEL|      |          |          ______
  VRAM G -->| - - - - - - |  |BG0|-BG0->|          |----o--->|      |
            | 3D Graphics |->|___|      |__________|    |    |Select|
            | Engine      |                             |    |Video |
            |_____________|--------3D----------------.  |    |Input |
             _______      _______              ___   |  |    |      |
            |       |    |       |<-----------|SEL|<-'  |    |and   |-->
            |       |    |       |    _____   |A  |     |    |      |
  VRAM A <--|Select |    |Select |   |     |<-|___|<----'    |Master|
  VRAM B <--|Capture|<---|Capture|<--|Blend|   ___           |Bright|
  VRAM C <--|Dest.  |    |Source |   |_____|<-|SEL|<----.    |A     |
  VRAM D <--|       |    |       |            |B  |     |    |      |
            |_______|    |_______|<-----------|___|<-.  |    |      |
             _______                                 |  |    |      |
  VRAM A -->|Select |                                |  |    |      |
  VRAM B -->|Display|--------------------------------o------>|      |
  VRAM C -->|VRAM   |                                   |    |      |
  VRAM D -->|_______|   _____________                   |    |      |
                       |Main Memory  |                  |    |      |
  Main   ------DMA---->|Display FIFO |------------------o--->|______|
  Memory               |_____________|
             _____________               __________           ______
  VRAM C -->| 2D Graphics |--------OBJ->| Layering |         |      |
  VRAM D -->| Engine B    |--------BG3->| and      |         |Master|
  VRAM H -->|             |--------BG2->| Special  |-------->|Bright|-->
  VRAM I -->|             |--------BG1->| Effects  |         |B     |
            |_____________|--------BG0->|__________|         |______|

DS Files - 2D Video

eg. used in DSi Launcher "rom:\layout\cmn\launcher_d.szs\.."
http://www.romhacking.net/documents/[469]nds_formats.htm

 ____________________________ Nitro Color Palette _____________________________

Header

  000h 4    Chunk ID "RLCN" (aka NCLR backwards, Nitro Color Resource)
  004h 2    Byte Order    (FEFFh)
  006h 2    Version       (0100h)
  008h 4    Total Filesize
  00Ch 2    Offset to "TTLP" Chunk, aka Size of "RLCN" Chunk (0010h)
  00Eh 2    Total number of following Chunks (1=TTLP) (or 2=TTLP+PMCP ?)

TTLP Chunk

  000h 4    Chunk ID "TTLP" (aka PLTT backwards, Palette data)
  004h 4    Chunk Size (eg. 0218h)
  008h 4    Reportedly Color Depth (ie. "tile usage info") (3=4bpp, 4=8bpp)
  00Ch 4    Zero
  010h 4    Palette Data Size in bytes (eg. 200h) (or 200h-N? no, blah!)
  014h 4    Offset from TTLP+8 to Palette Data? (always 10h)
  018h N*2  Palete Data (16bit colors, 0000h..7FFFh)

Most DSi titles use full 200h-byte palettes (Paper Plane has a smaller one in Graphics.NARC\Seq\pause.zcl). There seem to be no DSi titles with PMCP chunks.

PMCP Chunk (if any) (reportedly exists, but not in DSi Launcher...?)

  000h 4    Chunk ID "PMCP" (aka PCMP backwards, Palette CMP?)
  004h 4    Chunk Size (reportedly always 12h ???)
  008h 2    Number of palettes in file (uh?)
  00Ah 2    Unused (BEEFh=Bullshit)
  00Ch 4    Offset from PMCP+8 to Palette IDs? (always 08h)
  DATA N*2  "Palette ID numbers for each palette (starting from 0)"

 ___________________________ Nitro Character Tiles ____________________________

  eg. DSi Launcher "rom:\debug\DebugFont.NCGR"         -- with SOPC chunk
  eg. DSi Launcher "rom:\layout\cmn\launcher_d.szs\.." -- without SOPC chunk

Header

  000h 4    Chunk ID "RGCN" (aka NCGR backwards, Nitro Char Graphics Resource)
  004h 2    Byte Order    (FEFFh)
  006h 2    Version       (0101h) (unknown if 0100h does also exist?)
  008h 4    Total Filesize
  00Ch 2    Offset to "RAHC" Chunk, aka Size of "RGCN" Chunk (0010h)
  00Eh 2    Total number of following Chunks (1=RAHC, or 2=RAHC+SOPC)

RAHC Chunk

  000h 4    Chunk ID "RAHC" (aka CHAR backwards)
  004h 4    Chunk Size (eg. 1420h)
  008h 2    Tile Data Size in Kilobytes   ;\or both set to FFFFh
  00Ah 2    Unknown (always 20h)          ;/(when size<>N*1024)
  00Ch 4    Color Depth (3=4bpp, 4=8bpp)
  010h 2    Zero   ;or 10h (when SOPC not exists? kbyte size rounded up?)
  012h 2    Zero   ;or 20h (when SOPC not exists?)
  014h 4    Zero
  018h 4    Tile Data Size in Bytes (eg. 1400h)
  01Ch 4    Offset from RAHC+8 to Tile Data?  ;=always 18h
  020h ...  Tile Data (eg. 20h-byte zerofilled for 4bpp SPC char?)

Nonzero [10h,12h] spotted in Paper Plane "rom:\Graphics.NARC\Plane\plane.zcg".

SOPC Chunk (only present if Tile Data size is N*1024 bytes)

  000h 4    Chunk ID "SOPC" (aka CPOS backwards)
  004h 4    Chunk Size (10h)
  008h 4    Zero
  00Ch 2    Same as [00Ah] in RAHC chunk? (always 20h)
  00Eh 2    Same as [008h] in RAHC chunk? (size in kilobytes)

 __________________________ Unknown Character Tiles ___________________________

Apart from above NCGR, there is reportedly another tile format:

  NCGR (Nitro Character Graphic Resource) - Graphical Tiles --> see above
  NBGR (Nitro Basic Graphic Resource)     - Graphical Tiles --> what ???

If it does really exist for real... the header ID be "RGBN" (aka NBGR backwards), and file extension might be NBGR? But even if so, it's unknown if/when/where/why that NBGR format is used. If the "B" is for "Basic" then might have less features than NCGR, or maybe it might be "B" for non-tiled Bitmaps, or whatever?

 ___________________________ Nitro BG Maps Screens ____________________________

Header

  000h 4    Chunk ID "RCSN" (aka NSCR backwards, Nitro Screen Resource)
  004h 2    Byte Order    (FEFFh)
  006h 2    Version       (0100h)
  008h 4    Total Filesize
  00Ch 2    Offset to "NRCS" Chunk, aka Size of "RCSN" Chunk (0010h)
  00Eh 2    Total number of following Chunks (1=NRCS)

NRCS Chunk

  000h 4    Chunk ID "NRCS" (aka SCRN backwards, Screen)
  004h 4    Chunk Size
  008h 4    Screen Width in pixels
  00Ah 2    Screen Height in pixels
  00Ch 4    Zero
  010h 4    Screen Data Size (width/8)*(height/8)*2
  014h N*2  Screen Data (16bit BG Map entries, palette+xyflip+tileno)

 ____________________________ Nitro OBJ Animations ____________________________

Header

  000h 4    Chunk ID "RNAN" (aka NANR backwards, Nitro Animation Resource)
  004h 2    Byte Order    (FEFFh)
  006h 2    Version       (0100h)
  008h 4    Total Filesize
  00Ch 2    Offset to "KNBA" Chunk, aka Size of "RNAN" Chunk (0010h)
  00Eh 2    Total number of following Chunks (1=KNBA, or 3=KNBA+LBAL+TXEU)

One chunk exists in DSi Launcher
Three chunks exist in DSi Flipnote "rom:ManualData.Eu\md2res_narc.blz\data\obj

KNBA Chunk

  000h 4    Chunk ID "KNBA" (aka ABNK backwards, Animation Bank)
  004h 4    Chunk Size (always padded to 4-byte boundary if LABL chunk follows)
  008h 2    Number of 16-byte Animation Blocks ;implies NumLabels in LABL chunk
  00Ah 2    Number of 8-byte Frame Blocks
  00Ch 4    Offset from KNBA+8 to Animation Blocks ;=18h
  010h 4    Offset from KNBA+8 to Frame Blocks     ;=[0Ch]+[08h]*10h
  014h 4    Offset from KNBA+8 to Frame Data       ;=[10h]+[0Ah]*8
  018h 8    Zero
  DATA ..   Animation Blocks (16-byte entries)
   00h 4      Number of Frames
   04h 2      Unknown        (0)
   06h 2      Unknown Always (1)   ;reportedly "always unknown"
   08h 4      Unknown        (1..2)
   0Ch 4      Offset from FrameBlock+0 to First Frame
  DATA ..   Frame Blocks (8-byte entries)
   00h 4      Offset from FrameData+0 to whatever?   (always 4-byte aligned?)
   04h 2      Frame Width   ;3Ch or 01..06h   ;Time in 60Hz units? num meta's?
   06h 2      Unused (usually 0000h, or BEEFh=Bullshit)
  DATA ..   Frame Data (2-byte entries)
   00h 2      Unknown 16bit values? (maybe CELL index or whatever??)  (CCCCh=?)

LBAL Chunk

  000h 4    Chunk ID "LBAL" (aka LABL backwards, Labels)
  004h 4    Chunk Size (not padded to 4-byte size, following TXEU is unaligned)
  008h 4*N  Offsets from LabelArea+0 to Labels (for each Animation Block)
  ...  ..   Label Area (ASCII Strings, terminated by 00h)

The LabelArea starts at LBAL+8+NumLabels*4 (whereas, NumLabels is found in KNBA chunk).

TXEU Chunk
Caution: Not 4-byte aligned (the preceeding LBAL chunk can have odd size).

  000h 4    Chunk ID "TXEU" (aka UEXT backwards, Whatever Extension or so?)
  004h 4    Chunk Size (0Ch)
  008h 4    Unknown (usually 0) (reportedly 0 or 1)

 __________________________ Nitro OBJ Metatile Cells __________________________

Header

  000h 4    Chunk ID "RECN" (aka NCER backwards, Nitro Cell Resource)
  004h 2    Byte Order    (FEFFh)
  006h 2    Version       (0100h)
  008h 4    Total Filesize
  00Ch 2    Offset to "KBEC" Chunk, aka Size of "RECN" Chunk (0010h)
  00Eh 2    Total number of following Chunks (1=KBEC, or 3=KBEC+LBAL+TXEU)

KBEC Chunk

  000h 4    Chunk ID "KBEC" (aka CEBK backwards, Cell Bank)
  004h 4    Chunk Size (always padded to 4-byte boundary if LABL chunk follows)
  008h 2    Number of Metatiles
  00Ah 2    Metatiles Entry Size (0=Normal 8 bytes, 1=Extended 16 bytes)
              (DSi Launcher ..layout\cmn\launcher_u\.. uses 16-byte size)
  00Ch 4    Offset from KBEC+8 to Metatile Table? (18h)
  010h 4    Boundary Size (?)       (but is ZERO in layout\cmn\launcher_u\)
               "Specifies the area in which the image can be drawn,
               multiplied by 64, ie. 2 means that the area is 128x128 pixels."
  014h 0Ch  Zero
  020h ..   Metatile Table (8 bytes each) (or 16 bytes)
  ...  ..   OBJ Attribute Table (6-bytes each)

Metatile Table entries are (8-byte or 16-byte):

  000h 2    Number of OBJs
  002h 2    Unknown
  004h 4    OBJ Data Offset (from begin of OBJ Attr Table)
 (008h 2)   Unknown (can be 02h,10h,48h,74h)
 (00Ah 2)   Unknown (can be 08h)
 (00Ch 2)   Unknown (can be FFA0h..FFF0h) ;\maybe extra coordinate offsets?
 (00Eh 2)   Unknown (can be FFF0h..FFF9h) ;/

OBJ Attribute Table...

  starts at Number of Cells * 8 | each cell is made up of 6 bytes)

The 6-byte OBJ Attributes seem to be in normal OAM format, containing the coordinates, tile number, tile size, and other flags (however, the coordinates contain signed values; ie. one needs to add the current OBJ position to those values). For details on the OBJ Attributes, see:
LCD OBJ - OAM Attributes

LBAL and TEXU Chunks (if any)
Same as in Animation files (see there). In fact, the content seems to be SAME as the corresponding Animation file (for pairs of filename.NANR and filename.NCER), and the number of labels must be obtained from the NANR file's KNBA chunk (as such, it's rather useless to have LBAL in NCER files, except perhaps for error checking that the correct file pair was loaded).

Note: DSi Launcher layout\cmn\logodemo.szs has KBEC Chunk Size 2B6h (although the filesize is padded as if it were 2B8h bytes) (the file has no LBAL chunk, so it's unclear if/how it were aligned if present).

 ____________________________ Nitro Unknown Files  ____________________________

DSi Deep Psyche has two extra file types:

  .NMAR file (with "RAMN" header ID, and "KNBA"+"LBAL" chunks)
  .NMCR file (with "RCMN" header ID, and "KBCM" chunk)

The purpose is unknown, but they are probably also animating something...

  OBJ with 16bit x/y (instead 9bit/8bit)?
  OBJ with fractional x/y-stepping (moving/motion)?
  OBJ rotation/scaling?
  BG scroll offsets?
  BG tile replacement?

Going by the filename (a01_obj01.NMAR) they seem to be OBJ related. Labels include things like "a01_upNN" and "a01_windowNN".
The chunks seem to resemble those in RNAN/RECN files (Animation+Cells). Except, the KBCM has 8-byte entries (unlike the 6-byte ones in KBEC):

  000h 2    Unknown (000xh..007Ah, maybe time or so?)
  002h 2    Unknown (signed 16bit?)
  004h 2    Unknown (signed 16bit?)
  006h 2    Unknown (0x21h, with x=0..8)

 _________________________ Nitro More Unknown Files  __________________________

Some 2D folders contain more unknown files (eg. DSi Camera "rom:layout\cmn\fusion_camera.szs"):

JNBL (whatever, with .bnbl extension)

  000h 4     ID "JNBL"
  004h 2     Zero
  006h 2     Number of 6-byte entries (01h or more)
  008h N*6   Unknown

JNCL (whatever, with .bncl extension)

  000h 4     ID "JNCL"
  004h 2     Zero (0000h)
  006h 2     Number of 8-byte entries (01h or more)
  008h N*8   Unknown (eg. 80h,10h,C0h,20h,00h,00h,00h,00h)

JNLL (whatever, with .bnll extension)

  000h 4     ID "JNLL"
  004h 2     Zero (0000h)
  006h 2     Number of 16-byte entries (01h or more)
  008h N*16  Unknown (eg. 80h,50h,60h,10h,7Ch,29h,FFh,FDh,0Dh,19h,0,0,0,0,0,0)

BNGL (whatever, with .bngl extension)

  000h 4   ID "BNGL"   ;this same as file extension (not JNGL)
  004h 2   Zero (0000h)
  006h 2   Number of ?-byte entries (01h or more)
  008h 2   Unknown (can be 02h,04h,06h,0Ah)
  00Ah 2   Number of ?-byte other entries maybe (01h or more)
  ...
  ...      Entries?
  ...      Other Entries?
  ...      Maybe More Other Entries?

Filesize can range from 18h bytes to 24Ah bytes (or maybe yet smaller/biggger).

 ______________________________ .ntft and .ntfp _______________________________

.ntft file
Probably texture data (maybe for use as extra 2D layer), size is usually/always a power of 2 (ranging from 80h bytes to at least 64Kbytes (or even 512Kbytes?). Color depth can be 16bit or 8bit (and maybe less). Files with less than 16bit are bundled with a .ntfp palette file.

.ntfp file
Probably texture palette, with 16bit color numbers. The files can be quite small (eg. only 6 or 8 bytes).

 ______________________________ .wmif and .wmpf _______________________________

DSi Sudoku rom:\Textures\ has "Wild Magic" .wmif and .wmpf (image+palette) files.

.wmif file

  000h 1Bh  ID "Wild Magic Image File 3.00",00h
  01Bh 4    Palette Filename Length (eg. 0Dh)
  01Fh LEN  Palette Filename        (eg. "BG_Board.wmpf")
  ...  4    Texture Format (6=4bpp, 7=8bpp)
  ...  4    Texture Width in pixels
  ...  4    Texture Height in pixels
  ...  ..   Texture data

.wmpf file

  000h 1Dh  ID "Wild Magic Palette File 1.00",00h
  01Dh 4    Zero?
  021h 4    Number of Colors
  025h ..   Colors, 16bit (0000h..7FFFh)

DS Files - 3D Video (mostly unknown)

Some Nintendo DS games are use the following set of Nitro Studio files for 3D Models:

  .NSBMD (ID="BMD0") - Nitro Polygon Model
  .NSBTX (ID="BTX0") - Nitro Texture and Palette
  .NSBCA (ID="BCA0") - Nitro Skeletal Character Animation
  .NSBTP (ID="BTP0") - Nitro Texture Pattern-swap Animation
  .NSBTA (ID="BTA0") - Nitro Texture UV-change Animation (aka texcoords?)
  .NSBMA (ID="BMA0") - Nitro Material-swap Animation (whut?)
  .NSBVA (ID="BVA0") - Nitro Vis... Animation?

Unknown which games are actually using that format... probably some mid-NDS-era titles from Nintendo (early NDS launch titles didn't use it, 3rd party NDS titles tend to use custom formats, and later DSi mini-games don't support 3D graphics at all).

Below is an attempt to clean up rather confusing specs (hopefully appearing less confusing, but probably containing some misinterpretations in cases where the original specs where too confusing; the much-too-much confusing parts are left intact and marked as unknown what they might be meant to mean).
DS Files - 3D Video BMD0 (Model Data)
DS Files - 3D Video BTX0 (Texture)
DS Files - 3D Video BCA0 (Character Skeletal Animation)
DS Files - 3D Video BTP0/BTA0/BMA0/NVA0 (Unknown Animations)
Note: The format resembles the later CGFX format used on 3DS.

Dict "typedef struct Header" - alike "DICT" on 3DS?

  000h 1   Dummy 0
  001h 1   Amount of "objects"
  002h 2   Size of this Header (that is... what? header up to names?)
  ...  ..  Probably followed by the three sections mentioned below...?

Unknown Section ;maybe Name lookup (Patricia Tree or Hash or so)?

 Sub-Header
  000h 2   Size of this Sub-Header, always = 8
  002h 2   Size of this Unknown Section (that is... N*?+8 ...?)
  004h 4   Constant = 0000017Fh
 Unknown Data (repeats * Amount of "objects")
  008h 4   Unknown     ;1st object
  00Ch 4   ...         ;2nd object? or 2nd word of 1st object?
  etc                  ;etc?

Data Info Section

 The Data content depends on what the Dict is used for. It can contain
 4 or 8 bytes of data stored directly, or it can contain a 2 or 4 byte
 offset to larger data).
  000h 2      Size of each Data entry (siz) (usually 4 or 8)
  002h 2      Size of this Data Info Section (that is... N*siz+8 ...?)
  004h N*siz  Data (siz bytes, for each "object")

Name Section

  000h N*16   Name String (in ASCII maybe?) (for each "object")

DS Files - 3D Video BMD0 (Model Data)

Aka NSBMD

BMD0 Header

  000h 4     ID "BMD0" (Basic Model Data)
  004h 2     Byte Order (FEFFh)
  006h 2     Version? (reportedly 1 or 2, in whatever audio/video files?)
  008h 4     Total Filesize
  00Ch 2     Size of this structure (always 16 ???)
  00Eh 2     Number of chunks (1=MDL0 or 2=MDL0+TEX0)
  010h 4     Offset to MDL0 Chunk
  014h 4     Offset to TEX0 Chunk (if any?) (otherwise TEX0 is in NSBTX file)

MDL0 Chunk

  000h 4     Chunk ID "MDL0" (Model Block)
  004h 4     Chunk Size
  008h ..    Model Dict (with 32bit offsets to Models from "Block_MDL0")
  ...  ..    Models

Model (all offsets are relative to this Model section)

  000h 4    Size of Model (including these 4 bytes)
  004h 4    Offset of Additional Model Data          ;whut?
  008h 4    Offset of Texture & Palette Offset       ;?
  00Ch 4    Offset of Display List Start             ;?
  010h 4    Offset of Display List End               ;?
  014h 1    Unknown
  015h 1    Unknown
  016h 1    Unknown
  017h 1    Amount of Objects           ;\
  018h 1    Amount of Materials         ; what for?
  019h 1    Amount of Polygons          ; is that just some usage comment,
  020h 4    Unknown                     ; for statistical purposes?
  024h 2    Amount of Vertices          ;
  026h 2    Amount of Surfaces          ;
  028h 2    Amount of Triangles         ;
  02Ah 2    Amount of Quads             ;/
  02Ch 2    Bounding box X      (signed fixed point 1.3.12)
  02Eh 2    Bounding box Y      (signed fixed point 1.3.12)
  030h 2    Bounding box Z      (signed fixed point 1.3.12)
  032h 2    Bounding box Width  (signed fixed point 1.3.12)
  034h 2    Bounding box Height (signed fixed point 1.3.12)
  036h 2    Bounding box Depth  (signed fixed point 1.3.12)
  038h 4    Runtime use data
  03Ch 4    Runtime use data
  040h ..   Polygonal Object Dict (with 32bit offsets to Objects from
                                                 "this Object Header")
  ...  ..   Object Definitions

Object Definition (repeats * Amount of objects)

 In each definition:
  000h 1   Transform Flag byte
  001h 1   Unknown
  002h 1   Unknown
  003h 1   Unknown
  ...  4   Translation X (s32 signed fixed point 1.3.12)  ;\
  ...  4   Translation Y (s32 signed fixed point 1.3.12)  ; when Flags.0=0
  ...  4   Translation Z (s32 signed fixed point 1.3.12)  ;/
  ...  4   Scale X       (s32 signed fixed point 1.3.12)  ;\
  ...  4   Scale Y       (s32 signed fixed point 1.3.12)  ; when Flags.2=0
  ...  4   Scale Z       (s32 signed fixed point 1.3.12)  ;/
  ...  4   Rotation A    (s32 whatever, fixed point?)     ;\
  ...  4   Rotation B    (s32 whatever, fixed point?)     ; when Flags.1=0
  ...  4   Rotation C    (s32 whatever, fixed point?)     ; and Flags.3=0
  ...  4   Rotation D    (s32 whatever, fixed point?)     ;/
  ...  2   Pivot Rotation A (signed fixed point 1.3.12)   ;\when Flags.1=0
  ...  2   Pivot Rotation B (signed fixed point 1.3.12)   ;/and Flags.3=1

Unknown how above "1.3.12" can be "s32" (maybe actually 17.3.12 ?)
The Transform Flag bits are:

  0    Translation   (0=Yes, 1=No)
  1    Rotation      (0=Yes, 1=No)
  2    Scaling       (0=Yes, 1=No)
  3    Rotation Type (0=Rotate A,B,C,D, 1=Pivot A,B)
  4-7  Pivot Matrix  (0..8, see below) (used when bit1=0 and bit3=1)

Pivot matrix type (0..8):

  0:  | 1  0  0|  1:  | 0  1  0|  2:  | 0  0  1|
      | 0  A  B|      | A  0  B|      | A  B  0|
      | 0  B -A|      | B  0 -A|      | B -A  0|

  3:  | 0  A  B|  4:  | A  0  B|  5:  | A  B  0|
      | 1  0  0|      | 0  1  0|      | 0  0  1|
      | 0  B -A|      | B  0 -A|      | B -A  0|

  6:  | 0  A  B|  7:  | A  0  B|  8:  | A  B  0|
      | 0  B -A|      | B  0 -A|      | B -A  0|
      | 1  0  0|      | 0  1  0|      | 0  0  1|

Non-Pivot does probably also use some kind of matrix made of A,B,C,D...?

Bone/Skeleton Section <-- what is that? does it follow after above?

  000h ..    Definitions

Each definition consists of 1 Command Byte + Parameters. Each Parameter is 1 byte.

  Cmd  Params  Description
  06h  3       params: Object ID, Parent ID, dummy 0
  26h  4       params: Object ID, Parent ID, dummy 0, Stack ID
  46h  4       params: Object ID, Parent ID, dummy 0, Stack ID
  66h  5       params: Object ID, Parent ID, dummy 0, Stack ID, Restore ID

  00h  0       NOP (empty command)
  01h  0       End of Bone/Skeleton Section
  02h  2       params: Node ID, Visibility
  03h  1       Set Polygon Stack ID?
  04h  3       params: Material ID, 05h, Polygon ID
  05h  1       ??
  06h  3       params: Object ID, Parent ID, Dummy 0
  07h  1       ??
  08h  1       ??
  09h  8       ??

  0Bh  0       BEGIN (indicate begin of Polygon/Material pairing)
  2Bh  0       END (indicate end of Polygon/Material pairing)

  These are Material/Polygon pairing commands, 4 bytes.
  The lower nibble of 2nd Parameter must be 5.
  04h  3       Material ID, 05, Polygon ID
  24h  3       ...
  44h  3       ...

Texture & Palette Offset <-- what is that, addressed from Model [008h] ?

    000h 2   Offset of Texture Section (relative to Texture & Palette Offset)
    002h 2   Offset of Palette Section (relative to Texture & Palette Offset)

Material Section <-- WHAT is THAT ??? and WHERE is it ???

    Dict struct Header;// Material Header: one Header object for each material
    {// The 'Data' for this Header is like so:
      000h 4   Offset of Material Definition, relative to the start of
                 this Material Section
    }

    Texture Section
      Dict struct Header;// Texture Header: one Header object for each texture
      {// The u32 'Data' for this Header is like so:
          000h 2   Offset of Matching Data (relative to Texture & Palette
                     Offset)
          002h 2   Amount of associated Materials (a texture can be in more
                     than one material)
      }

    Palette Section
      Dict struct Header;// Palette Header: one Header object for each palette
      {// The u32 'Data' for this Header is like so:
          000h 2   Offset of Matching Data (relative to Texture & Palette
                     Offset)
          002h 2   Amount of associated Materials (a palette can be in more
                     than one material)
      }

    Material Definition (repeats * Amount of Materials)
      ;// Usually 48 bytes for each material

Polygon Section

    Dict struct Header;// Polygon Header: one Header object for each material
    {// The 'Data' for this Header is like so:
        000h 4   Offset of Polygon Definition, relative to the start of this
                   Polygon Section
    }

    Polygon Definition (repeats * Amount of Polygons) (10h-bytes each)
      000h 4   Unknown
      004h 4   Unknown
      008h 4   Offset of Display List, relative to Polygon Definition
      00Ch 4   Size of Display List

    Display List
      // The Display List is actually packed geometry command.
      // See the DStek specification for more information:
      // http://www.akkit.org/info/gbatek.htm#ds3dvideo    ;uh?

TEX0 Chunk
Reportedly same as in BTX0 files, see there for details.

DS Files - 3D Video BTX0 (Texture)

The NSBTX file format stores texture image and palette information.
The "TEX0" Block can be found in certain NSBMD files, not just in NSBTX files.

BTX0 Header

  000h 4     ID "BTX0" (Basic Texture)
  004h 2     Byte Order (FEFFh)
  006h 2     Version? (reportedly 1 or 2, in whatever audio/video files?)
  008h 4     Total Filesize
  00Ch 2     Size of this structure (always 16 ???)
  00Eh 2     Number of chunks (1=TEX0)
  010h 4     Offset to TEX0 Chunk

TEX0 Chunk

  000h 4     Chunk ID "TEX0" (Texture Block)
  004h 4     Chunk Size
  008h 4     Padding (0)
  00Ch 2     Texture Data Size (bitshift << 3)             ;\
  00Eh 2     Texture Dict Offset (03Ch)                    ; Texture
  010h 4     Padding (0)                                   ;
  014h 4     Texture Data Offset                           ;/
  018h 4     Padding (0)
  01Ch 2     Compressed Texture Data Size (bitshift << 3)  ;\<-- Size
  01Eh 2     Compressed Texture Dict? Offset (03Ch, again?); <-- Dict?
  020h 4     Padding (0)                                   ;
  024h 4     Compressed Texture Data Offset                ; <-- Data
  028h 4     Compressed Texture Info Data Offset           ;/<-- InfoData?
  02Ch 4     Padding (0)
  030h 4     Palette Data Size (bitshift << 3)             ;\
  034h 4     Palette Dict Offset                           ; Palette
  038h 4     Palette Data Offset                           ;/
  03Ch ..    Texture Dict (with 8-byte entries, see below)
  N/A? ?     Compressed Texture Dict? (maybe here? with whatever entries?)
  ...  ..    Palette Dict (with 4-byte entries, see below)
  ...  ..    Texture Data Section     (unknown... maybe bitmap/pixels?)
  ...  ..    Compressed Texture Data Section            (unknown...)
  ...  ..    Compressed Texture Info Data Section       (unknown...)
  ...  ..    Palette Data Section     (unknown... maybe palette/colors?)

Format of 8-byte entries in Texture Dict:

  000h 2     Texture Offset (bitshift << 3), relative to the start
                                             of Texture Data
  002h 2     Parameters   ;<-- probably "upper 16bit of TEXIMAGE_PARAM" ?
                The format is, using knock-out description:
                bit:  15..............0
                   0 b --CFFFHHHWWW-----   <-- uh, is that "0 b" and 17bits???
                where:
                  C = Palette ID
                  F = Format (0-7)
                  H = Height (8 << Height)
                  W = Width (8 << Width)
                To Calculate the Data Size of a Texture:
                Bit Depth = Format: <0, 8, 2, 4, 8, 2, 8, 16>
                Width * Height * BitDepth / 8
  004h 1     Width (should match W << 3)
  005h 1     Unknown (is 00h or 80h)
  006h 1     Height? (can be 0, 1, 2, 4, 8)
  007h 1     Unknown (is 80h)

Compressed Texture Dict? (maybe here? with whatever entries?)
Compressed Texture might have a Dict, too? with whatever entries?

Format of 4-byte entries in Palette Dict:

  000h   2   u16 ;// (bitshift << 3) Palette Offset, relative to the start
                                                     of Palette Data
  000h?  2   u16 ;// Unknown (is 0 or 1)

Texture Data Section:
Compressed Texture Data Section:
Compressed Texture Info Data Section:
Palette Data Section:

  Unknown

DS Files - 3D Video BCA0 (Character Skeletal Animation)

The NSBCA file format stores character skeletal/bone/joint animation data.

BCA0 Header

  000h 4     ID "BCA0" (Basic Character Animation)
  004h 2     Byte Order (FEFFh)
  006h 2     Version? (reportedly 1 or 2, in whatever audio/video files?)
  008h 4     Total Filesize
  00Ch 2     Size of this structure (always 16 ???)
  00Eh 2     Number of chunks (1=JNT0)
  010h 4     Offset to JNT0 Chunk

JNT0 Chunk

  000h 4     Chunk ID "JNT0" (Joint Block)
  004h 4     Chunk Size
  008h ..    Joint Dict (with 32bit offsets to Joints, from "Block_JNT0")

Joints
Reportedly Joints "repeats * Amount of Objects"
Uh, with Amount of Objects stored inside of below?
Maybe rather meant to mean that Joints exist for each Model (not Object)?

  000h 4    ID 'J.AC' (Joint Animation Content ?)
  004h 2    Amount of Frames
  006h 2    Amount of Objects (should be same as in BMD0 file)
  008h 4    Unknown
  00Ch 4    Offset to Unknown1 chunk. Relative to start of this block.
  010h 4    Offset to Unknown2 chunk. Relative to start of this block.
  014h 4    Object Info Offset (repeats * Amount of Objects), relative
               to start of this Joint Section.
               uh, is above meant to be Offset to first Object?
               or an array with Offsets to all Objects?
  ...  ..   Object Info, for each object (supposedly here?)
  ...  ..   Unknown1
  ...  ..   Unknown2
  ...  ..   Reportedly "end of file" (is that possible with multiple joints?)

Object Info - Repeats * Number of Objects

    000h 2   Flag - Indicates what sort of Transformations are applied.
              Bit0  -  Unused?
              Bit1  T  Translate  (0=Yes; data follows, 1=No?)
              Bit2  -  Unused?
              Bit3  X  Affects how WHICH data is stored but dunno WHERE?
              Bit4  Y  Affects how WHICH data is stored but dunno WHERE?
              Bit5  Z  Affects how WHICH data is stored but dunno WHERE?
              Bit6  R  Rotate     (0=Yes; data follows, 1=No?)
              Bit7  -  Unused?
              Bit8  r  Affects how WHICH data is stored but dunno WHERE?
              Bit9  S  Scale      (0=Yes; data follows, 1=No?)
              Bit10 -  Unused?
              Bit11 x  Affects how WHICH data is stored but dunno WHERE?
              Bit12 y  Affects how WHICH data is stored but dunno WHERE?
              Bit13 z  Affects how WHICH data is stored but dunno WHERE?
              Bit14 -  Unused?
              Bit15 -  Unused?
    002h 1   Unknown   (reportedly u32, 1-byte wide ???)
    003h 1   ID Number (reportedly u32, 1-byte wide ???)
  *** (Unknown what below crap means, it does probably refer to multiple
  *** entries, for Translate, Rotate, and Scale; and possibly even to
  *** multiple sub-entries for X,Y,Z or whatever?)
  Transformation Info (Translation XYZ, Rotation, Scale XYZ) when WHAT=1
    000h 4   Actual value (uh, WHICH value?)
  Transformation Info (Translation XYZ, Rotation, Scale XYZ) when WHAT=0
    000h 2   Unknown - typically 0000h
    002h 2   Unknown
    004h 4   Offset to data. Relative to "(Object Info Offset + 4)".

Unknown1 - Repeats until Unknown2

  000h 2   Reportedly 36, 32 & 0.
  002h 2   Unknown (reportedly signed)
  004h 2   Unknown (reportedly signed)

Unknown2 - Repeats until end of file

  "All transformation offsets point somewhere in this section.
  It's clearly broken up into parts (Translation, Rotation & Scale),
  however I'm not totally sure how data is store in here yet."

DS Files - 3D Video BTP0/BTA0/BMA0/NVA0 (Unknown Animations)

 ____________ DS Files - 3D Video BTP0 (Texture Pattern Animation) ____________

BTP0 Header

  000h 4     ID "BTP0" (Texture Pattern Animation)
  ...  ..    XXX

PAT0 Chunk

  000h 4     Chunk ID "PAT0" (Pattern Block)
  004h 4     Chunk Size
  ...  ..    Unknown

 ________________ DS Files - 3D Video BTA0 (Texture Animation) ________________

BTA0 Header

  000h 4     ID "BTA0" (Texture Animation)
  ...  ..    XXX

SRT0 Chunk

  000h 4     Chunk ID "SRT0"   (maybe short for Scale/Rotate/Translate?)
  004h 4     Chunk Size
  ...  ..    Unknown

 _______________ DS Files - 3D Video BMA0 (Material Animation) ________________

BMA0 Header

  000h 4     ID "BMA0" (Material Animation)
  ...  ..    XXX

MAT0 Chunk

  000h 4     Chunk ID "MAT0" (Material Block)
  004h 4     Chunk Size
  ...  ..    Unknown

 ____________________ DS Files - 3D Video BVA0 (Unknown?) _____________________

BVA0 Header

  000h 4     ID "BVA0" (whatever Vis... Animation?)
  ...  ..    XXX

VIS0 Chunk

  000h 4     Chunk ID "VIS0" (Visibility...?)
  004h 4     Chunk Size
  ...  ..    Unknown

DS 3D Video

DS 3D Overview
DS 3D I/O Map
DS 3D Display Control
DS 3D Geometry Commands
DS 3D Matrix Load/Multiply
DS 3D Matrix Types
DS 3D Matrix Stack
DS 3D Matrix Examples (Projection)
DS 3D Matrix Examples (Rotate/Scale/Translate)
DS 3D Matrix Examples (Maths Basics)
DS 3D Polygon Attributes
DS 3D Polygon Definitions by Vertices
DS 3D Polygon Light Parameters
DS 3D Shadow Polygons
DS 3D Texture Attributes
DS 3D Texture Formats
DS 3D Texture Coordinates
DS 3D Texture Blending
DS 3D Toon, Edge, Fog, Alpha-Blending, Anti-Aliasing
DS 3D Status
DS 3D Tests
DS 3D Rear-Plane
DS 3D Final 2D Output

3D is more or less (about 92%) understood and described.

DS 3D Overview

The NDS 3D hardware consists of a Geometry Engine, and a Rendering Engine.

Geometry Engine (Precalculate coordinates & assign polygon attributes)
Geometry commands can be sent via Ports 4000440h and up (or alternately, written directly to Port 4000400h).
The commands include matrix and vector multiplications, the purpose is to rotate/scale/translate coordinates (vertices), the resulting coordinates are stored in Vertex RAM.
Moreover, it allows to assign attributes to the polygons and vertices, that includes vertex colors (or automatically calculated light colors), texture attributes, number of vertices per polygon (three or four), and a number of flags, these attributes are stored in Polygon RAM. Polygon RAM also contains pointers to the corresponding vertices in Vertex RAM.

Swap Buffers (Pass data from the Geometry Engine to the Rendering Engine)
The hardware includes two sets of Vertex/Polygon RAM, one used by the Geometry Engine, one by the Rendering Engine. The SwapBuffers command simply exchanges these buffers (so the new Geometry Data is passed to the Rendering Engine) (and the old buffer is emptied, so the Geometry engine can write new data to it). Additionally, the two parameter bits from the <previous> SwapBuffers command are copied to the Geometry Engine.
Data that is NOT swapped: SwapBuffers obviously can't swap Texture memory (so software must take care that Texture memory is kept mapped throughout rendering). Moreover, the rendering control registers (ports 4000060h, and 4000330h..40003BFh) are not swapped (so that values must be kept intact during rendering, too).

Rendering Engine (Display Output)
The Rendering Engine draws the various Polygons, and outputs them as BG0 layer to the 2D Video controller (which may then output them to the screen, or to the video capture unit). The Rendering part is done automatically by hardware, so the software has little influence on it.
Rendering is done scanline-by-scanline, so there's only a limited number of clock cycles per scanline, which is limiting the maximum number of polygons per scanline. However, due to the 48-line cache (see below), some scanlines are allowed to exceed that maximum.
Rendering starts 48 lines in advance (while still in the Vblank period) (and does then continue throughout the whole display period), the rendered data is written to a small cache that can hold up to 48 scanlines.

Scanline Cache vs Framebuffer
Note: There's only the 48-line cache (not a full 192-line framebuffer to store the whole rendered image). That is perfectly reasonable since animated data is normally drawn only once (so there would be no need to store it). That, assuming that the Geometry Engine presents new data every frame (otherwise, if the Geometry software is too slow, or if the image isn't animated, then the hardware is automatically rendering the same image again, and again).

DS 3D I/O Map

3D I/O Map

  Address  Siz Name            Expl.
  Rendering Engine (per Frame settings)
  4000060h 2   DISP3DCNT       3D Display Control Register (R/W)
  4000320h 1   RDLINES_COUNT   Rendered Line Count Register (R)
  4000330h 10h EDGE_COLOR      Edge Colors 0..7 (W)
  4000340h 1   ALPHA_TEST_REF  Alpha-Test Comparision Value (W)
  4000350h 4   CLEAR_COLOR     Clear Color Attribute Register (W)
  4000354h 2   CLEAR_DEPTH     Clear Depth Register (W)
  4000356h 2   CLRIMAGE_OFFSET Rear-plane Bitmap Scroll Offsets (W)
  4000358h 4   FOG_COLOR       Fog Color (W)
  400035Ch 2   FOG_OFFSET      Fog Depth Offset (W)
  4000360h 20h FOG_TABLE       Fog Density Table, 32 entries (W)
  4000380h 40h TOON_TABLE      Toon Table, 32 colors (W)
  Geometry Engine (per Polygon/Vertex settings)
  4000400h 40h GXFIFO          Geometry Command FIFO (W)
  4000440h ... ...             Geometry Command Ports (see below)
  4000600h 4   GXSTAT          Geometry Engine Status Register (R and R/W)
  4000604h 4   RAM_COUNT       Polygon List & Vertex RAM Count Register (R)
  4000610h 2   DISP_1DOT_DEPTH 1-Dot Polygon Display Boundary Depth (W)
  4000620h 10h POS_RESULT      Position Test Results (R)
  4000630h 6   VEC_RESULT      Vector Test Results (R)
  4000640h 40h CLIPMTX_RESULT  Read Current Clip Coordinates Matrix (R)
  4000680h 24h VECMTX_RESULT   Read Current Directional Vector Matrix (R)

Geometry Commands (can be invoked by Port Address, or by Command ID)
Table shows Port Address, Command ID, Number of Parameters, and Clock Cycles.

  Address  Cmd Pa.Cy.
  N/A      00h -  -   NOP - No Operation (for padding packed GXFIFO commands)
  4000440h 10h 1  1   MTX_MODE - Set Matrix Mode (W)
  4000444h 11h -  17  MTX_PUSH - Push Current Matrix on Stack (W)
  4000448h 12h 1  36  MTX_POP - Pop Current Matrix from Stack (W)
  400044Ch 13h 1  17  MTX_STORE - Store Current Matrix on Stack (W)
  4000450h 14h 1  36  MTX_RESTORE - Restore Current Matrix from Stack (W)
  4000454h 15h -  19  MTX_IDENTITY - Load Unit Matrix to Current Matrix (W)
  4000458h 16h 16 34  MTX_LOAD_4x4 - Load 4x4 Matrix to Current Matrix (W)
  400045Ch 17h 12 30  MTX_LOAD_4x3 - Load 4x3 Matrix to Current Matrix (W)
  4000460h 18h 16 35* MTX_MULT_4x4 - Multiply Current Matrix by 4x4 Matrix (W)
  4000464h 19h 12 31* MTX_MULT_4x3 - Multiply Current Matrix by 4x3 Matrix (W)
  4000468h 1Ah 9  28* MTX_MULT_3x3 - Multiply Current Matrix by 3x3 Matrix (W)
  400046Ch 1Bh 3  22  MTX_SCALE - Multiply Current Matrix by Scale Matrix (W)
  4000470h 1Ch 3  22* MTX_TRANS - Mult. Curr. Matrix by Translation Matrix (W)
  4000480h 20h 1  1   COLOR - Directly Set Vertex Color (W)
  4000484h 21h 1  9*  NORMAL - Set Normal Vector (W)
  4000488h 22h 1  1   TEXCOORD - Set Texture Coordinates (W)
  400048Ch 23h 2  9   VTX_16 - Set Vertex XYZ Coordinates (W)
  4000490h 24h 1  8   VTX_10 - Set Vertex XYZ Coordinates (W)
  4000494h 25h 1  8   VTX_XY - Set Vertex XY Coordinates (W)
  4000498h 26h 1  8   VTX_XZ - Set Vertex XZ Coordinates (W)
  400049Ch 27h 1  8   VTX_YZ - Set Vertex YZ Coordinates (W)
  40004A0h 28h 1  8   VTX_DIFF - Set Relative Vertex Coordinates (W)
  40004A4h 29h 1  1   POLYGON_ATTR - Set Polygon Attributes (W)
  40004A8h 2Ah 1  1   TEXIMAGE_PARAM - Set Texture Parameters (W)
  40004ACh 2Bh 1  1   PLTT_BASE - Set Texture Palette Base Address (W)
  40004C0h 30h 1  4   DIF_AMB - MaterialColor0 - Diffuse/Ambient Reflect. (W)
  40004C4h 31h 1  4   SPE_EMI - MaterialColor1 - Specular Ref. & Emission (W)
  40004C8h 32h 1  6   LIGHT_VECTOR - Set Light's Directional Vector (W)
  40004CCh 33h 1  1   LIGHT_COLOR - Set Light Color (W)
  40004D0h 34h 32 32  SHININESS - Specular Reflection Shininess Table (W)
  4000500h 40h 1  1   BEGIN_VTXS - Start of Vertex List (W)
  4000504h 41h -  1   END_VTXS - End of Vertex List (W)
  4000540h 50h 1  392 SWAP_BUFFERS - Swap Rendering Engine Buffer (W)
  4000580h 60h 1  1   VIEWPORT - Set Viewport (W)
  40005C0h 70h 3  103 BOX_TEST - Test if Cuboid Sits inside View Volume (W)
  40005C4h 71h 2  9   POS_TEST - Set Position Coordinates for Test (W)
  40005C8h 72h 1  5   VEC_TEST - Set Directional Vector for Test (W)

All cycle timings are counted in 33.51MHz units. NORMAL commands takes 9..12 cycles, depending on the number of enabled lights in PolyAttr (Huh, 9..12 (four timings) cycles for 0..4 (five settings) lights?) Total execution time of SwapBuffers is Duration until VBlank, plus 392 cycles.
In MTX_MODE=2 (Simultanous Set), MTX_MULT/TRANS take additional 30 cycles.

DS 3D Display Control

4000060h - DISP3DCNT - 3D Display Control Register (R/W)

  0     Texture Mapping      (0=Disable, 1=Enable)
  1     PolygonAttr Shading  (0=Toon Shading, 1=Highlight Shading)
  2     Alpha-Test           (0=Disable, 1=Enable) (see ALPHA_TEST_REF)
  3     Alpha-Blending       (0=Disable, 1=Enable) (see various Alpha values)
  4     Anti-Aliasing        (0=Disable, 1=Enable)
  5     Edge-Marking         (0=Disable, 1=Enable) (see EDGE_COLOR)
  6     Fog Color/Alpha Mode (0=Alpha and Color, 1=Only Alpha) (see FOG_COLOR)
  7     Fog Master Enable    (0=Disable, 1=Enable)
  8-11  Fog Depth Shift      (FOG_STEP=400h shr FOG_SHIFT) (see FOG_OFFSET)
  12    Color Buffer RDLINES Underflow (0=None, 1=Underflow/Acknowledge)
  13    Polygon/Vertex RAM Overflow    (0=None, 1=Overflow/Acknowledge)
  14    Rear-Plane Mode                (0=Blank, 1=Bitmap)
  15-31 Not used

4000540h - Cmd 50h - SWAP_BUFFERS - Swap Rendering Engine Buffer (W)
SwapBuffers exchanges the two sets of Polygon/Vertex RAM buffers, that is, the newly defined polygons/vertices are passed to the rendering engine (and will be displayed in following frame(s)). The other buffer is emptied, and passed to the Geometry Engine (to be filled with new polygons/vertices by Geometry Commands).

  0     Translucent polygon Y-sorting (0=Auto-sort, 1=Manual-sort)
  1     Depth Buffering  (0=With Z-value, 1=With W-value)
        (mode 1 does not function properly with orthogonal projections)
  2-31  Not used

SwapBuffers isn't executed until next VBlank (Scanline 192) (the Geometry Engine is halted for that duration). SwapBuffers should not be issued within Begin/End. The two parameter bits of the SwapBuffers command are used for the following gxcommands (ie. not for the old gxcommands prior to SwapBuffers).
SwapBuffers does lock-up the 3D hardware if an incomplete polygon list has been defined (eg. a triangle with only 2 vertices). On lock-up, only 2D video is kept working, any wait-loops for GXSTAT.27 will hang the program. Once lock-up has occured, there seems to be no way to recover by software, not by sending the missing veric(es), and not even by pulsing POWCNT1.Bit2-3.

4000580h - Cmd 60h - VIEWPORT - Set Viewport (W)

  0-7   Screen/BG0 Coordinate X1 (0..255) (For Fullscreen: 0=Left-most)
  8-15  Screen/BG0 Coordinate Y1 (0..191) (For Fullscreen: 0=Bottom-most)
  16-23 Screen/BG0 Coordinate X2 (0..255) (For Fullscreen: 255=Right-most)
  24-31 Screen/BG0 Coordinate Y2 (0..191) (For Fullscreen: 191=Top-most)

Coordinate 0,0 is the lower-left (unlike for 2D where it'd be upper-left).
The 3D view-volume (size as defined by the Projection Matrix) is automatically scaled to match into the Viewport area. Although polygon vertices are clipped to the view-volume, some vertices may still exceed to X2,Y1 (lower-right) boundary by one pixel, due to some sort of rounding errors. The Viewport settings don't affect the size or position of the 3D Rear-Plane. Viewport should not be issued within Begin/End.

4000610h - DISP_1DOT_DEPTH - 1-Dot Polygon Display Boundary Depth (W)
1-Dot Polygons are very small, or very distant polygons, which would be rendered as a single pixel on screen. Polygons with a depth value greater (more distant) than DISP_1DOT_DEPTH can be automatically hidden; in order to reduce memory consumption, or to reduce dirt on the screen.

  0-14  W-Coordinate (Unsigned, 12bit integer, 3bit fractional part)
  15-31 Not used                 (0000h=Closest, 7FFFh=Most Distant)

The DISP_1DOT_DEPTH comparision can be enabled/disabled per polygon (via POLYGON_ATTR.Bit13), so "important" polygons can be displayed regardless of their size and distance.
Note: The comparision is always using the W-coordinate of the vertex (not the Z-coordinate) (ie. no matter if using Z-buffering, or W-buffering). The polygon is rendered if at least one of its vertices is having a w-coordinate less or equal than DISP_1DOT_DEPTH. NB. despite of checking the w-coords of ALL vertices, the polygon is rendered using the color/depth/texture of its FIRST vertex.
Note: The hardware does round-up the width and height of all polygons to at least 1, so polygons of 0x0, 1x0, 0x1, and 1x1 dots will be all rounded-up to a size of 1x1. Of which, the so-called "1dot" depth check is applied only to the 0x0 dot variant (so "0dot" depth check would be a better name for it).
Caution: Although DISP_1DOT_DEPTH is a Geometry Engine parameter, it is NOT routed through GXFIFO, ie. changes will take place immediately, and will affect all following polygons, including such that are still in GXFIFO. Workaround: ensure that GXFIFO is empty before changing this parameter.

4000340h - ALPHA_TEST_REF - Alpha-Test Comparision Value (W)
Alpha Test can be enabled in DISP3DCNT.Bit2. When enabled, pixels are rendered only if their Alpha value is GREATER than ALPHA_TEST_REF. Otherwise, when disabled, pixels are rendered only if their Alpha value is GREATER than zero. Alpha Test is performed on the final polygon pixels (ie. after texture blending).

  0-4   Alpha-Test Comparision Value (0..31) (Draw pixels if Alpha>AlphaRef)
  5-31  Not used

Value 00h is effectively the same as when Alpha Test is disabled. Value 1Fh hides all polygons, including opaque ones.

DS 3D Geometry Commands

4000400h - GXFIFO - Geometry Command FIFO (W) (mirrored up to 400043Fh?)
Used to send packed commands, unpacked commands,

  0-7   First  Packed Command (or Unpacked Command)
  8-15  Second Packed Command (or 00h=None)
  16-23 Third  Packed Command (or 00h=None)
  24-31 Fourth Packed Command (or 00h=None)

and parameters,

  0-31  Parameter data for the previously sent (packed) command(s)

to the Geometry engine.

FIFO / PIPE Number of Entries
The FIFO has 256 entries, additionally, there is a PIPE with four entries (giving a total of 260 entries). If the FIFO is empty, and if the PIPE isn't full, then data is moved directly into the PIPE, otherwise it is moved into the FIFO. If the PIPE runs half empty (less than 3 entries) then 2 entries are moved from the FIFO to the PIPE. The state of the FIFO can be obtained in GXSTAT.Bit16-26, observe that there may be still data in the PIPE, even if the FIFO is empty. Check the busy flag in GXSTAT.Bit27 to see if the PIPE or FIFO contains data (or if a command is still executing).
Each PIPE/FIFO entry consists of 40bits of data (8bit command code, plus 32bit parameter value). Commands without parameters occupy 1 entry, and Commands with N parameters occupy N entries.

Sending Commands by Ports 4000440h..40005FFh
Geometry commands can be indirectly sent to the FIFO via ports 4000440h and up.
For a command with N paramters: issue N writes to the port.
For a command without parameters: issue one dummy-write to the port.
That mechanism puts the 8bit command + 32bit parameter into the FIFO/PIPE.
If the FIFO is full, then a wait is generated until data is removed from the FIFO, ie. the STR opcode gets freezed, during the wait, the bus cannot be used even by DMA, interrupts, or by the NDS7 CPU.

GXFIFO Access via DMA
Larger pre-calculated data blocks can be sent directly to the FIFO. This is usually done via DMA (use DMA in Geometry Command Mode, 32bit units, Dest=4000400h/fixed, Length=NumWords, Repeat=0). The timings are handled automatically, ie. the system (should) doesn't freeze when the FIFO is full (see below Overkill note though). DMA starts when the FIFO becomes less than half full, the DMA does then write 112 words to the GXFIFO register (or less, if the remaining DMA transfer length gets zero).

GXFIFO Access via STR,STRD,STM
If desired, STR,STRD,STM opcodes can be used to write to the FIFO.
Opcodes that write more than one 32bit value (ie. STRD and STM) can be used to send ONE UNPACKED command, plus any parameters which belong to that command. After that, there must be a 1 cycle delay before sending the next command (ie. one cannot sent more than one command at once with a single opcode, each command must be invoked by a new opcode). STRD and STM can be used because the GXFIFO register is mirrored to 4000400h..43Fh (16 words).
As with Ports 4000440h and up, the CPU gets stopped if (and as long as) the FIFO is full.

GXFIFO / Unpacked Commands

  - command1 (upper 24bit zero)
  - parameter(s) for command1 (if any)
  - command2 (upper 24bit zero)
  - parameter(s) for command2 (if any)
  - command3 (upper 24bit zero)
  - parameter(s) for command3 (if any)

GXFIFO / Packed Commands

  - command1,2,3,4 packed into one 32bit value (all bits used)
  - parameter(s) for command1 (if any)
  - parameter(s) for command2 (if any)
  - parameter(s) for command3 (if any)
  - parameter(s) for command4 (top-most packed command MUST have parameters)
  - command5,6 packed into one 32bit value (upper 16bit zero)
  - parameter(s) for command5 (if any)
  - parameter(s) for command6 (top-most packed command MUST have parameters)
  - command7,8,9 packed into one 32bit value (upper 8bit zero)
  - parameter(s) for command7 (if any)
  - parameter(s) for command8 (if any)
  - parameter(s) for command9 (top-most packed command MUST have parameters)

Packed commands are first decompressed and then stored in command the FIFO.

GXFIFO DMA Overkill on Packed Commands Without Parameters
Normally, the 112 word limit ensures that the FIFO (256 entries) doesn't get full, however, this limit is much too high for sending a lot of "Packed Commands Without Parameters" (ie. PUSH, IDENTITY, or END) - eg. sending 112 x Packed(00151515h) to GXFIFO would write 336 x Cmd(15h) to the FIFO, which is causing the FIFO to get full, and which is causing the DMA (and CPU) to be paused (for several seconds, in WORST case) until enough FIFO commands have been processed to allow the DMA to finish the 112 word transfer.
Not sure if there's much chance to get Overkills in practice. Normally most commands DO have parameters, and so, usually even LESS than 112 FIFO entries are occupied (since 8bit commands with 32bit parameters are merged into single 40bit FIFO entries).

Invalid GX commands
Invalid commands (anything else than 10h..1Ch, 20h..2Bh, 30h..33h, 40h..41h, 50h, 60h, or 70h..72h) seem to be simply ignored by the hardware (at least, testing has confirmed that they do not fetch any parameters from the gxfifo).

DS 3D Matrix Load/Multiply

4000440h - Cmd 10h - MTX_MODE - Set Matrix Mode (W)

  0-1   Matrix Mode (0..3)
         0  Projection Matrix
         1  Position Matrix (aka Modelview Matrix)
         2  Position & Vector Simultaneous Set mode (used for Light+VEC_TEST)
         3  Texture Matrix (see DS 3D Texture Coordinates chapter)
  2-31  Not used

Selects the current Matrix, all following MTX commands (load, multiply, push, pop, etc.) are applied to that matrix. In Mode 2, all MTX commands are applied to both the Position and Vector matrices. There are two special cases:

  MTX_SCALE in Mode 2:                  uses ONLY Position Matrix
  MTX_PUSH/POP/STORE/RESTORE in Mode 1: uses BOTH Position AND Vector Matrices

Ie. the four stack commands act like mode 2 (even when in mode 1; keeping the two stacks somewhat in sync), and scale acts like mode 1 (even when in mode 2; keeping the light vector length's intact).

 vice-versa for the scale command.

For the above cases, the commands do always act like mode 1, even when they are i

4000454h - Cmd 15h - MTX_IDENTITY - Load Unit Matrix to Current Matrix (W)
Sets C=I. Parameters: None
The Identity Matrix (I), aka Unit Matrix, consists of all zeroes, with a diagonal row of ones. A matrix multiplied by the Unit Matrix is left unchanged.

4000458h - Cmd 16h - MTX_LOAD_4x4 - Load 4x4 Matrix to Current Matrix (W)
Sets C=M. Parameters: 16, m[0..15]

400045Ch - Cmd 17h - MTX_LOAD_4x3 - Load 4x3 Matrix to Current Matrix (W)
Sets C=M. Parameters: 12, m[0..11]

4000460h - Cmd 18h - MTX_MULT_4x4 - Multiply Current Matrix by 4x4 Matrix (W)
Sets C=M*C. Parameters: 16, m[0..15]

4000464h - Cmd 19h - MTX_MULT_4x3 - Multiply Current Matrix by 4x3 Matrix (W)
Sets C=M*C. Parameters: 12, m[0..11]

4000468h - Cmd 1Ah - MTX_MULT_3x3 - Multiply Current Matrix by 3x3 Matrix (W)
Sets C=M*C. Parameters: 9, m[0..8]

400046Ch - Cmd 1Bh - MTX_SCALE - Multiply Current Matrix by Scale Matrix (W)
Sets C=M*C. Parameters: 3, m[0..2]
Note: MTX_SCALE doesn't change Vector Matrix (even when in MTX_MODE=2) (that's done so for keeping the length of the light vector's intact).

4000470h - Cmd 1Ch - MTX_TRANS - Mult. Curr. Matrix by Translation Matrix (W)
Sets C=M*C. Parameters: 3, m[0..2] (x,y,z position)

4000640h..67Fh - CLIPMTX_RESULT - Read Current Clip Coordinates Matrix (R)
This 64-byte region (16 words) contains the m[0..15] values of the Current Clip Coordinates Matrix, arranged in 4x4 Matrix format. Make sure that the Geometry Engine is stopped (GXSTAT.27) before reading from these registers.
The Clip Matrix is internally used to convert vertices to screen coordinates, and is internally re-calculated anytime when changing the Position or Projection matrices:

  ClipMatrix = PositionMatrix * ProjectionMatrix

To read only the Position Matrix, or only the Projection Matrix: Use Load Identity on the OTHER matrix, so the ClipMatrix becomes equal to the DESIRED matrix (multiplied by the Identity Matrix, which has no effect on the result).

4000680h..6A3h - VECMTX_RESULT - Read Current Directional Vector Matrix (R)
This 36-byte region (9 words) contains the m[0..8] values of the Current Directional Vector Matrix, arranged in 3x3 Matrix format (the fourth row/column may contain any values).
Make sure that the Geometry Engine is stopped (GXSTAT.27) before reading from these registers.

DS 3D Matrix Types

Essentially, all matrices in the NDS are 4x4 Matrices, consisting of 16 values, m[0..15]. Each element is a signed fixed-point 32bit number, with a fractional part in the lower 12bits.
The other Matrix Types are used to reduce the number of parameters being transferred, for example, 3x3 Matrix requires only nine parameters, the other seven elements are automatically set to 0 or 1.0 (whereas "1.0" means "1 SHL 12" in 12bit fixed-point notation).

   _      4x4 Matrix       _        _    Identity Matrix    _
  | m[0]  m[1]  m[2]  m[3]  |      |  1.0   0     0     0    |
  | m[4]  m[5]  m[6]  m[7]  |      |  0     1.0   0     0    |
  | m[8]  m[9]  m[10] m[11] |      |  0     0     1.0   0    |
  |_m[12] m[13] m[14] m[15]_|      |_ 0     0     0     1.0 _|

   _      4x3 Matrix       _        _  Translation Matrix   _
  | m[0]  m[1]  m[2]   0    |      |  1.0   0     0     0    |
  | m[3]  m[4]  m[5]   0    |      |  0     1.0   0     0    |
  | m[6]  m[7]  m[8]   0    |      |  0     0     1.0   0    |
  |_m[9]  m[10] m[11]  1.0 _|      |_m[0]  m[1]  m[2]   1.0 _|

   _      3x3 Matrix       _        _     Scale Matrix      _
  | m[0]  m[1]  m[2]   0    |      | m[0]   0     0     0    |
  | m[3]  m[4]  m[5]   0    |      |  0    m[1]   0     0    |
  | m[6]  m[7]  m[8]   0    |      |  0     0    m[2]   0    |
  |_ 0     0     0     1.0 _|      |_ 0     0     0     1.0 _|

DS 3D Matrix Stack

Matrix Stack
The NDS has three Matrix Stacks, and two Matrix Stack Pointers (the Coordinate Matrix stack pointer is also shared for Directional Matrix Stack).

  Matrix Stack________Valid Stack Area____Stack Pointer___________________
  Projection Stack    0..0  (1 entry)     0..1  (1bit) (GXSTAT: 1bit)
  Coordinate Stack    0..30 (31 entries)  0..63 (6bit) (GXSTAT: 5bit only)
  Directional Stack   0..30 (31 entries)  (uses Coordinate Stack Pointer)
  Texture Stack       One..None?          0..1  (1bit) (GXSTAT: N/A)

Which of the stacks/matrices depends on the current Matrix Mode (as usually,
but with one exception; stack operations MTX_PUSH/POP/STORE/RESTORE in Mode 1 are acting same as in Mode 2):

  MTX_MODE = 0      --> Projection Stack
  MTX_MODE = 1 or 2 --> BOTH Coordinate AND Directional Stack
  MTX_MODE = 3      --> Texture Stack

The initial value of the Stack Pointers is zero, the current value of the pointers can be read from GXSTAT (read-only), that register does also indicate stack overflows (errors flag gets set on read/write to invalid entries, ie. entries 1 or 1Fh..3Fh). For all stacks, the upper half (ie. 1 or 20h..3Fh) are mirrors of the lower half (ie. 0 or 0..1Fh).

4000444h - Cmd 11h - MTX_PUSH - Push Current Matrix on Stack (W)
Parameters: None. Sets [S]=C, and then S=S+1.

4000448h - Cmd 12h - MTX_POP - Pop Current Matrix from Stack (W)
Sets S=S-N, and then C=[S].

  Parameter Bit0-5:  Stack Offset (signed value, -30..+31) (usually +1)
  Parameter Bit6-31: Not used

Offset N=(+1) pops the most recently pushed value, larger offsets of N>1 will "deallocate" N values (and load the Nth value into C). Zero or negative values can be used to pop previously "deallocated" values.
The stack has only one level (at address 0) in projection mode, in that mode, the parameter value is ignored, the offset is always +1 in that mode.

400044Ch - Cmd 13h - MTX_STORE - Store Current Matrix on Stack (W)
Sets [N]=C. The stack pointer S is not used, and is left unchanged.

  Parameter Bit0-4:  Stack Address (0..30) (31 causes overflow in GXSTAT.15)
  Parameter Bit5-31: Not used

The stack has only one level (at address 0) in projection mode, in that mode, the parameter value is ignored.

4000450h - Cmd 14h - MTX_RESTORE - Restore Current Matrix from Stack (W)
Sets C=[N]. The stack pointer S is not used, and is left unchanged.

  Parameter Bit0-4:  Stack Address (0..30) (31 causes overflow in GXSTAT.15)
  Parameter Bit5-31: Not used

The stack has only one level (at address 0) in projection mode, in that mode, the parameter value is ignored.

In Projection mode, the parameter for POP, STORE, and RESTORE is unused - not sure if the parameter (ie. a dummy value) is - or is not - to be written to the command FIFO?
There appear to be actually 32 entries in Coordinate & Directional Stacks, entry 31 appears to exist, and appears to be read/write-able (although the stack overflow flag gets set when accessing it).

DS 3D Matrix Examples (Projection)

The most important matrix is the Projection Matrix (to be initialized with MTX_MODE=0 via MTX_LOAD_4x4 command). It does specify the dimensions of the view volume.

With Perspective Projections more distant objects will appear smaller, with Orthogonal Projects the size of the objects is always same regardless of their distance.

  Perspective Projection     Orthogonal Projection
                   __                  __________
       top __..--''  |            top |          |
          |   view   |                |   view   |
  Eye ----|--------->|        Eye ----|--------->|
          |__volume  |                |  volume  |
     bottom  ''--..__|          bottom|__________|
        near        far             near        far

Correctly initializing the projection matrix (as shown in the examples below) can be quite difficult (mind that fixed point multiply/divide requires to adjust the fixed-point width before/after calculation). For beginners, it may be recommended to start with a simple Identity Matrix (MTX_IDENTITY command) used as Projection Matrix (ie. Ortho with t,b,l,r set to +/-1).

Orthogonal Projections (Ortho)

  | (2.0)/(r-l)       0             0            0     |
  |      0       (2.0)/(t-b)        0            0     |
  |      0            0        (2.0)/(n-f)       0     |
  | (l+r)/(l-r)  (b+t)/(b-t)   (n+f)/(n-f)      1.0    |

n,f specify the distance from eye to near and far clip planes. t,b,l,r are the coordinates of near clip plane (top,bottom,left,right). For a symmetrical view (ie. the straight-ahead view line centered in the middle of viewport) t,b,l,r should be usually t=+ysiz/2, b=-ysiz/2, r=+xsiz/2, l=-xsiz/2; the (xsiz/ysiz) ratio should be usually equal to the viewport's (width/heigh) ratio. Examples for a asymmetrical view would be b=0 (frog's view), or t=0 (bird's view).

Left-Right Asymmetrical Perspective Projections (Frustum)

  | (2*n)/(r-l)       0             0            0     |
  |      0       (2*n)/(t-b)        0            0     |
  | (r+l)/(r-l)  (t+b)/(t-b)   (n+f)/(n-f)     -1.0    |
  |      0            0       (2*n*f)/(n-f)      0     |

n,f,t,b,l,r have same meanings as above (Ortho), the difference is that more distant objects will appear smaller with Perspective Projection (unlike Orthogonal Projection where the size isn't affected by the distance).

Left-Right Symmetrical Perspective Projections (Perspective)

  | cos/(asp*sin)     0             0            0     |
  |      0         cos/sin          0            0     |
  |      0            0        (n+f)/(n-f)     -1.0    |
  |      0            0       (2*n*f)/(n-f)      0     |

Quite the same as above (Frustum), but with symmetrical t,b values (which are in this case obtained from a vertical view range specified in degrees), and l,r are matched to the aspect ratio of the viewport (asp=height/width).

Moving the Camera
After initializing the Projection Matrix, you may multiply it with Rotate and/or Translation Matrices to change camera's position and view direction.

DS 3D Matrix Examples (Rotate/Scale/Translate)

Identity Matrix
The MTX_IDENTITY command can be used to initialize the Position Matrix before doing any Translation/Scaling/Rotation, for example:

  Load(Identity)                           ;no rotation/scaling used
  Load(Identity), Mul(Rotate), Mul(Scale)  ;rotation/scaling (not so efficient)
  Load(Rotate), Mul(Scale)                 ;rotation/scaling (more efficient)

Rotation Matrices
Rotation can be performed with MTX_MULT_3x3 command, simple examples are:

  Around X-Axis          Around Y-Axis          Around Z-Axis
  | 1.0  0     0   |     | cos   0    sin |     | cos   sin   0   |
  | 0    cos   sin |     | 0     1.0  0   |     | -sin  cos   0   |
  | 0    -sin  cos |     | -sin  0    cos |     | 0     0     1.0 |

Scale Matrix
The MTX_SCALE command allows to adjust the size of the polygon. The x,y,z parameters should be normally all having the same value, x=y=z (unless if you want to change only the height of the object, for example). Identical results can be obtained with MTX_MULT commands, however, when using lighting (MTX_MODE=2), then scaling should be done ONLY with MTX_SCALE (which keeps the length of the light's directional vector intact).

Translation Matrix
The MTX_TRANS command allows to move polygons to the desired position. The polygon VTX commands are spanning only a small range of coordinates (near zero-coordinate), so translation is required to move the polygons to other locations in the world coordinates. Aside from that, translation is useful for moved objects (at variable coordinates), and for re-using an object at various locations (eg. you can create a forest by translating a tree to different coordinates).

Matrix Multiply Order
The Matrix must be set up BEFORE sending the Vertices (which are then automatically multiplied by the matrix). When using multiple matrices multiplied with each other: Mind that, for matrix maths A*B is NOT the same as B*A. For example, if you combine Rotate and Translate Matrices, the object will be either rotated around it's own zero-coordinate, or around world-space zero-coordinate, depending on the multiply order.

DS 3D Matrix Examples (Maths Basics)

Below is a crash-course on matrix maths. Most of it is carried out automatically by the hardware. So this chapter is relevant only if you are interested in details about what happens inside of the 3D engine.

Matrix-by-Matrix Multiplication
Matrix multiplication, C = A * B, is possible only if the number of columns in A is equal to the number of rows in B, so it works fine with the 4x4 matrices which are used in the NDS. For the multiplication, assume matrix C to consist of elements cyx, and respecitively, matrix A and B to consist of elements ayx and byx. So that C = A * B looks like:

  | c11 c12 c13 c14 |     | a11 a12 a13 a14 |     | b11 b12 b13 b14 |
  | c21 c22 c23 c24 |  =  | a21 a22 a23 a24 |  *  | b21 b22 b23 b24 |
  | c31 c32 c33 c34 |     | a31 a32 a33 a34 |     | b31 b32 b33 b34 |
  | c41 c42 c43 c44 |     | a41 a42 a43 a44 |     | b41 b42 b43 b44 |

Each element in C is calculated by multiplying the elements from one row in A by the elements from the corresponding column in B, and then taking the sum of the products, ie.

  cyx = ay1*b1x + ay2*b2x + ay3*b3x + ay4*b4x

In total, that requires 64 multiplications (four multiplications for each of the 16 cyx elements), and 48 additions (three per cyx element), the hardware carries out that operation at a relative decent speed of 30..35 clock cycles, possibly by performing several multiplications simultaneously with separate multiply units.
Observe that for matrix multiplication, A*B is NOT the same as B*A.

Matrix-by-Vector & Vector-by-Matrix Multiplication
Vectors are Matrices with only one row, or only one column. Multiplication works as for normal matrices; the number of rows/columns must match up, repectively, row-vectors can be multiplied by matrices; and matrices can be multiplied by column-vectors (but not vice-versa). Eg. C = A * B:

                                                  | b11 b12 b13 b14 |
  | c11 c12 c13 c14 |  =  | a11 a12 a13 a14 |  *  | b21 b22 b23 b24 |
                                                  | b31 b32 b33 b34 |
                                                  | b41 b42 b43 b44 |

The formula for calculating the separate elements is same as above,

  cyx = ay1*b1x + ay2*b2x + ay3*b3x + ay4*b4x

Of which, C and A have only one y-index, so one may replace "cyx and ayx" by "c1x and a1x", or completely leave out the y-index, ie. "cx and ax".

Matrix-by-Number Multiplication
Simply multiply all elements of the Matrix by the number, C = A * n:

  cyx = ayx*n

Of course, works also with vectors (matrices with only one row/column).

Matrix-to-Matrix Addition/Subtraction
Both matrices must have the same number of rows & columns, add/subtract all elements with corresponding elements in other matrix, C = A +/- B:

  cyx = ayx +/- byx

Of course, works also with vectors (two matrices with only one row/column).

Vectors
A vector, for example (x,y,z), consists of offsets along x-,y-, and z-axis. The line from origin to origin-plus-offset is having two characteristics: A direction, and a length.
The length (aka magnitude) can be calculated as L=sqrt(x^2+y^2+z^2).

Vector-by-Vector Multiplication
This can be processed as LineVector*RowVector, so the result is a number (aka scalar) (aka a matrix with only 1x1 elements). Multiplying two (normalized) vectors results in: "cos(angle)=vec1*vec2", ie. the cosine of the angle between the two vectors (eg. used for light vectors). Multiplying a vector with itself, and taking the square root of the result obtains its length, ie. "length=sqrt(vec^2)".
That stuff should be done with 3-dimensional vectors (not 4-dimensionals).

Normalized Vectors
Normalized Vectors (aka Unit Vectors) are vectors with length=1.0. To normalize a vector, divide its coordinates by its length, ie. x=x/L, y=y/L, z=z/L, the direction remains the same, but the length is now 1.0.
On the NDS, normalized vectors should have a length of something less than 1.0 (eg. something like 0.99) because several NDS registers are limited to 1bit sign, 0bit integer, Nbit fractional part (so vectors that are parallel to the x,y,z axes, or that become parallel to them after rotation, cannot have a length of 1.0).

Fixed-Point Numbers
The NDS uses fixed-point numbers (rather than floating point numbers). Addition and Subtraction works as with normal integers, provided that the fractional part is the same for both numbers. If it is not the same: Shift-left the value with the smaller fractional part.
For multiplication, the fractional part of result is the sum of the fractional parts (eg. 12bit fraction * 12bit fraction = 24bit fraction; shift-right the result by 12 to convert it 12bit fraction). The NDS matrix multiply unit is maintaining the full 24bit fraction when processing the

  cyx = ay1*b1x + ay2*b2x + ay3*b3x + ay4*b4x

formula, ie. the three additions are using full 24bit fractions (with carry-outs to upper bits), the final result of the additions is then shifted-right by 12.
For division, it's vice versa, the fractions of the operands are substracted, 24bit fraction / 12bit fraction = 12bit fraction. When dividing two 12bit numbers, shift-left the first number by 12 before division to get a result with 12bit fractional part.

Four-Dimensional Matrices
The NDS uses four-dimensional matrices and vectors, ie. matrices with 4x4 elements, and vectors with 4 elements. The first three elements are associated with the X,Y,Z-axes of the three-dimensional space. The fourth element is somewhat a "W-axis".
With 4-dimensional matrices, the Translate matrix can be used to move an object to another position. Ie. once when you've setup a matrix (which may consists of pre-multiplied scaling, rotation, translation matrices), then that matrix can be used on vertices to perform the rotation, scaling, translation all-at-once; by a single Vector*Matrix operation.
With 3-dimensional matrices, translation would require a separate addition, additionally to the multiply operation.

DS 3D Polygon Attributes

40004A4h - Cmd 29h - POLYGON_ATTR - Set Polygon Attributes (W)

  0-3   Light 0..3 Enable Flags (each bit: 0=Disable, 1=Enable)
  4-5   Polygon Mode  (0=Modulation,1=Decal,2=Toon/Highlight Shading,3=Shadow)
  6     Polygon Back Surface   (0=Hide, 1=Render)  ;Line-segments are always
  7     Polygon Front Surface  (0=Hide, 1=Render)  ;rendered (no front/back)
  8-10  Not used
  11    Depth-value for Translucent Pixels    (0=Keep Old, 1=Set New Depth)
  12    Far-plane intersecting polygons       (0=Hide, 1=Render/clipped)
  13    1-Dot polygons behind DISP_1DOT_DEPTH (0=Hide, 1=Render)
  14    Depth Test, Draw Pixels with Depth    (0=Less, 1=Equal) (usually 0)
  15    Fog Enable                            (0=Disable, 1=Enable)
  16-20 Alpha      (0=Wire-Frame, 1..30=Translucent, 31=Solid)
  21-23 Not used
  24-29 Polygon ID (00h..3Fh, used for translucent, shadow, and edge-marking)
  30-31 Not used

Writes to POLYGON_ATTR have no effect until next BEGIN_VTXS command.
Changes to the Light bits have no effect until lighting is re-calculated by Normal command. The interior of Wire-frame polygons is transparent (Alpha=0), and only the lines at the polygon edges are rendered, using a fixed Alpha value of 31.

4000480h - Cmd 20h - COLOR - Directly Set Vertex Color (W)

  Parameter 1, Bit 0-4    Red
  Parameter 1, Bit 5-9    Green
  Parameter 1, Bit 10-14  Blue
  Parameter 1, Bit 15-31  Not used

The 5bit RGB values are internally expanded to 6bit RGB as follows: X=X*2+(X+31)/32, ie. zero remains zero, all other values are X=X*2+1.
Aside from by using the Color command, the color can be also changed by MaterialColor0 command (if MaterialColor0.Bit15 is set, it acts identical as the Color Command), and by the Normal command (which calculates the color based on light/material parameters).

Depth Test
The Depth Test compares the depth of the pixels of the polygon with the depth of previously rendered polygons (or of the rear plane if there have been none rendered yet). The new pixels are drawn if the new depth is Less (closer to the camera), or if it is Equal, as selected by POLYGON_ATTR.Bit14.
Normally, Depth Equal would work only exact matches (ie. if the overlapping polygons have exactly the same coordinates; and thus have the same rounding errors), however, the NDS hardware is allowing "Equal" to have a tolerance of +/-200h (within the 24bit depth range of 0..FFFFFFh), that may bypass rounding errors, but it may also cause nearby polygons to be accidently treated to have equal depth.

DS 3D Polygon Definitions by Vertices

The DS supports polygons with 3 or 4 edges, triangles and quadliterals.
The position of the edges is defined by vertices, each consisting of (x,y,z) values.

For Line Segments, use Triangles with twice the same vertex, Line Segments are rendered always because they do not have any front and back sides.
The Prohibited Quad shapes may produce unintended results, namely, that are Quads with crossed sides, and quads with angles greater than 180 degrees.

  Separate Tri.     Triangle Strips   Line Segment
  v0                 v2___v4____v6
  |\      v3         /|\  |\    /\     v0    v1
  | \     /\      v0( | \ | \  /  \     ------
  |__\   /__\        \|__\|__\/____\         v2
  v1 v2 v4  v5       v1   v3  v5   v7

  Separate Quads          Quadliteral Strips         Prohibited Quads
    v0__v3                 v0__v2____v4     v10__    v0__v3     v4
     /  \   v4____v7        /  \     |\ _____ / /v11   \/       |\
    /    \   |    \        /    \    | |v6 v8| /       /\     v5| \
   /______\  |_____\      /______\___|_|_____|/       /__\     /___\
   v1    v2  v5    v6     v1    v3  v5 v7   v9       v2   v1   v6   v7

The vertices are normally arranged anti-clockwise, except that: in triangle-strips each second polygon uses clockwise arranged vertices, and quad-strips are sorts of "up-down" arranged (whereas "up" and "down" may be anywhere due to rotation). Other arrangements may result in quads with crossed lines, or may swap the front and back sides of the polygon (above examples are showing the front sides).

4000500h - Cmd 40h - BEGIN_VTXS - Start of Vertex List (W)

  Parameter 1, Bit 0-1    Primitive Type (0..3, see below)
  Parameter 1, Bit 2-31   Not used

Indicates the Start of a Vertex List, and its Primitive Type:

  0  Separate Triangle(s)    ;3*N vertices per N triangles
  1  Separate Quadliteral(s) ;4*N vertices per N quads
  2  Triangle Strips         ;3+(N-1) vertices per N triangles
  3  Quadliteral Strips      ;4+(N-1)*2 vertices per N quads

The BEGIN_VTX command should be followed by VTX_-commands to define the Vertices of the list, and should be then terminated by END_VTX command.
BEGIN_VTX additionally applies changes to POLYGON_ATTR.

4000504h - Cmd 41h - END_VTXS - End of Vertex List (W)
Parameters: None. This is a Dummy command for OpenGL compatibility. It should be used to terminate a BEGIN_VTX, VTX_<values> sequence. END_VTXS is possibly required for Nintendo's software emulator? On real NDS consoles (and in no$gba) it does have no effect, it can be left out, or can be issued multiple times inside of a vertex list, without disturbing the display.

400048Ch - Cmd 23h - VTX_16 - Set Vertex XYZ Coordinates (W)

  Parameter 1, Bit 0-15   X-Coordinate (signed, with 12bit fractional part)
  Parameter 1, Bit 16-31  Y-Coordinate (signed, with 12bit fractional part)
  Parameter 2, Bit 0-15   Z-Coordinate (signed, with 12bit fractional part)
  Parameter 2, Bit 16-31  Not used

4000490h - Cmd 24h - VTX_10 - Set Vertex XYZ Coordinates (W)

  Parameter 1, Bit 0-9    X-Coordinate (signed, with 6bit fractional part)
  Parameter 1, Bit 10-19  Y-Coordinate (signed, with 6bit fractional part)
  Parameter 1, Bit 20-29  Z-Coordinate (signed, with 6bit fractional part)
  Parameter 1, Bit 30-31  Not used

Same as VTX_16, with only one parameter, with smaller fractional part.

4000494h - Cmd 25h - VTX_XY - Set Vertex XY Coordinates (W)

  Parameter 1, Bit 0-15   X-Coordinate (signed, with 12bit fractional part)
  Parameter 1, Bit 16-31  Y-Coordinate (signed, with 12bit fractional part)

The Z-Coordinate is kept unchanged, and re-uses the value from previous VTX.

4000498h - Cmd 26h - VTX_XZ - Set Vertex XZ Coordinates (W)

  Parameter 1, Bit 0-15   X-Coordinate (signed, with 12bit fractional part)
  Parameter 1, Bit 16-31  Z-Coordinate (signed, with 12bit fractional part)

The Y-Coordinate is kept unchanged, and re-uses the value from previous VTX.

400049Ch - Cmd 27h - VTX_YZ - Set Vertex YZ Coordinates (W)

  Parameter 1, Bit 0-15   Y-Coordinate (signed, with 12bit fractional part)
  Parameter 1, Bit 16-31  Z-Coordinate (signed, with 12bit fractional part)

The X-Coordinate is kept unchanged, and re-uses the value from previous VTX.

40004A0h - Cmd 28h - VTX_DIFF - Set Relative Vertex Coordinates (W)

  Parameter 1, Bit 0-9    X-Difference (signed, with 9/12bit fractional part)
  Parameter 1, Bit 10-19  Y-Difference (signed, with 9/12bit fractional part)
  Parameter 1, Bit 20-29  Z-Difference (signed, with 9/12bit fractional part)
  Parameter 1, Bit 30-31  Not used

Sets XYZ-Coordinate relative to the XYZ-Coordinates from previous VTX. In detail: The 9bit fractional values are divided by 8 (sign expanded to 12bit fractions, in range +/-0.125), and that 12bit fraction is then added to the old vtx coordinates. The result of the addition should not overflow 16bit vertex coordinate range (1bit sign, 3bit integer, 12bit fraction).

Notes on VTX commands
On each VTX command, the viewport coordinates of the vertex are calculated and stored in Vertex RAM,

  ( xx, yy, zz, ww ) = ( x, y, z, 1.0 ) * ClipMatrix

The actual screen position (in pixels) is then,

  screen_x = (xx+ww)*viewport_width / (2*ww) + viewport_x1
  screen_y = (yy+ww)*viewport_height / (2*ww) + viewport_y1

Each VTX command that completes the definition of a polygon (ie. each 3rd for Separate Trangles) does additionally store data in Polygon List RAM.
VTX commands may be issued only between Begin and End commands.

Clipping
Polygons are clipped to the 6 sides of the view volume (ie. to the left, right, top, bottom, near, and far edges). If one or more vertic(es) exceed one of these sides, then these vertic(es) are replaced by two newly created vertices (which are located on the intersections of the polygon edges and the view volume edge).
Depending on the number of clipped vertic(es), this may increase or decrease the number of entries in Vertex RAM (ie. minus N clipped vertices, plus 2 new vertices). Also, clipped polygons which are part of polygon strips are converted to separate polygons (which does increase number of entries in Vertex RAM). Polygons that are fully outside of the View Volume aren't stored in Vertex RAM, nor in Polygon RAM (the only exception are polygons that are located exactly one pixel below of, or right of lower/right edges, which appear to be accidently stored in memory).

DS 3D Polygon Light Parameters

The lighting operation is performed by executing the Normal command (which sets the VertexColor based on the Light/Material parameters) (to the rest of the hardware it doesn't matter if the VertexColor was set by Color command or by Normal command). Light is calculated only for the Front side of the polygon (assuming that the Normal is matched to that side), so the Back side will be (incorrectly) using the same color.

40004C8h - Cmd 32h - LIGHT_VECTOR - Set Light's Directional Vector (W)
Sets direction of the specified light (ie. the light selected in Bit30-31).

  0-9   Directional Vector's X component (1bit sign + 9bit fractional part)
  10-19 Directional Vector's Y component (1bit sign + 9bit fractional part)
  20-29 Directional Vector's Z component (1bit sign + 9bit fractional part)
  30-31 Light Number                     (0..3)

Upon executing this command, the incoming vector is multiplied by the current Directional Matrix, the result is then applied as LightVector. This allows to rotate the light direction. However, normally, to keep the light unrotated, be sure to use LoadIdentity (in MtxMode=2) before setting the LightVector.

40004CCh - Cmd 33h - LIGHT_COLOR - Set Light Color (W)
Sets the color of the specified light (ie. the light selected in Bit30-31).

  0-4   Red          (0..1Fh)      ;\light color this will be combined with
  5-9   Green        (0..1Fh)      ; diffuse, specular, and ambient colors
  10-14 Blue         (0..1Fh)      ;/upon execution of the normal command
  15-29 Not used
  30-31 Light Number (0..3)

40004C0h - Cmd 30h - DIF_AMB - MaterialColor0 - Diffuse/Ambient Reflect. (W)

  0-4   Diffuse Reflection Red     ;\light(s) that directly hits the polygon,
  5-9   Diffuse Reflection Green   ; ie. max when NormalVector has opposite
  10-14 Diffuse Reflection Blue    ;/direction of LightVector
  15    Set Vertex Color (0=No, 1=Set Diffuse Reflection Color as Vertex Color)
  16-20 Ambient Reflection Red     ;\light(s) that indirectly hits the polygon,
  21-25 Ambient Reflection Green   ; ie. assuming that light is reflected by
  26-30 Ambient Reflection Blue    ;/walls/floor, regardless of LightVector
  31    Not used

With Bit15 set, the lower 15bits are applied as VertexColor (exactly as when when executing the Color command), the purpose is to use it as default color (eg. when outcommenting the Normal command), normally, when using lighting, the color setting gets overwritten (as soon as executing the Normal command).

40004C4h - Cmd 31h - SPE_EMI - MaterialColor1 - Specular Ref. & Emission (W)

  0-4   Specular Reflection Red    ;\light(s) reflected towards the camera,
  5-9   Specular Reflection Green  ; ie. max when NormalVector is in middle of
  10-14 Specular Reflection Blue   ;/LightVector and ViewDirection
  15    Specular Reflection Shininess Table (0=Disable, 1=Enable)
  16-20 Emission Red               ;\light emitted by the polygon itself,
  21-25 Emission Green             ; ie. regardless of light colors/vectors,
  26-30 Emission Blue              ;/and no matter if any lights are enabled
  31    Not used

Caution: Specular Reflection WON'T WORK when the ProjectionMatrix is rotated.

40004D0h - Cmd 34h - SHININESS - Specular Reflection Shininess Table (W)
Write 32 parameter words (each 32bit word containing four 8bit entries), entries 0..3 in the first word, through entries 124..127 in the last word:

  0-7   Shininess 0 (unsigned fixed-point, 0bit integer, 8bit fractional part)
  8-15  Shininess 1 ("")
  16-23 Shininess 2 ("")
  24-31 Shininess 3 ("")

If the table is disabled (by MaterialColor1.Bit15), then reflection will act as if the table would be filled with linear increasing numbers.

4000484h - Cmd 21h - NORMAL - Set Normal Vector (W)
In short, this command does calculate the VertexColor, based on the various light-parameters.
In detail, upon executing this command, the incoming vector is multiplied by the current Directional Matrix, the result is then applied as NormalVector (giving it the same rotation as used for the following polygon vertices).

  0-9   X-Component of Normal Vector (1bit sign + 9bit fractional part)
  10-19 Y-Component of Normal Vector (1bit sign + 9bit fractional part)
  20-29 Z-Component of Normal Vector (1bit sign + 9bit fractional part)
  30-31 Not used

Defines the Polygon's Normal. And, does then update the Vertex Color; by recursing the View Direction, the NormalVector, the LightVector(s), and Light/Material Colors. The execution time of the Normal command varies depending on the number of enabled light(s).

Additional Light Registers
Additionally to above registers, light(s) must be enabled in PolygonAttr (mind that changes to PolygonAttr aren't applied until next Begin command). And, the Directional Matrix must be set up correctly (in MtxMode=2) for the LightVector and NormalVector commands.

Normal Vector
The Normal vector must point "away from the polygon surface" (eg. for the floor, the Normal should point upwards). That direction is implied by the polygon vertices, however, the hardware cannot automatically calculate it, so it must be set manually with the Normal command (prior to the VTX-commands).
When using lighting, the Normal command must be re-executed after switching Lighting on/off, or after changing light/material parameters. And, of course, also before defining polygons with different orientation. Polygons with same orientation (eg. horizontal polygon surfaces) and same material color can use the same Normal. Changing the Normal per polygon gives differently colored polygons with flat surfaces, changing the Normal per vertex gives the illusion of curved surfaces.

Light Vector
Each light consists of parallel beams; similar to sunlight, which appears to us (due to the great distance) to consist of parallel beams, all emmitted into the same direction; towards Earth.
In reality, light is emitted into ALL directions, originated from the light source (eg. a candle), the hardware doesn't support that type of non-parallel light. However, the light vectors can be changed per polygon, so a polygon that is located north of the light source may use different light direction than a polygon that is east of the light source.
And, of course, Light 0..3 may (and should) have different directions.

Normalized Vectors
The Normal Vector and the Light Vectors should be normalized (ie. their length should be 1.0) (in practice: something like 0.99, since the registers have only fractional parts) (a length of 1.0 can cause overflows).

Lighting Limitations
The functionality of the light feature is limited to reflecting light to the camera (light is not reflected to other polygons, nor does it cast shadows on other polygons). However, independently of the lighting feature, the DS hardware does allow to create shadows, see:
DS 3D Shadow Polygons

Internal Operation on Normal Command

  IF TexCoordTransformMode=2 THEN TexCoord=NormalVector*Matrix (see TexCoord)
  NormalVector=NormalVector*DirectionalMatrix
  VertexColor = EmissionColor
  FOR i=0 to 3
   IF PolygonAttrLight[i]=enabled THEN
    DiffuseLevel = max(0,-(LightVector[i]*NormalVector))
    ShininessLevel = max(0,(-HalfVector[i])*(NormalVector))^2
    IF TableEnabled THEN ShininessLevel = ShininessTable[ShininessLevel]
    ;note: below processed separately for the R,G,B color components...
    VertexColor = VertexColor + SpecularColor*LightColor[i]*ShininessLevel
    VertexColor = VertexColor + DiffuseColor*LightColor[i]*DiffuseLevel
    VertexColor = VertexColor + AmbientColor*LightColor[i]
   ENDIF
  NEXT i

Internal Operation on Light_Vector Command (for Light i)

  LightVector[i] = (LightVector*DirectionalMatrix)
  HalfVector[i] = (LightVector[i]+LineOfSightVector)/2

LineOfSightVector (how it SHOULD work)
Ideally, the LineOfSightVector should point from the camera to the vertic(es), however, the vertic(es) are still unknown at time of normal command, so it is just pointing from the camera to the screen, ie.

  LineOfSightVector = (0,0,-1.0)

Moreover, the LineOfSightVector should be multiplied by the Projection Matrix (so the vector would get rotated accordingly when the camera gets rotated), and, after multiplication by a scaled matrix, it'd be required to normalize the resulting vector.

LineOfSightVector (how it DOES actually work)
However, the NDS cannot normalize vectors by hardware, and therefore, it does completely leave out the LineOfSightVector*ProjectionMatrix multiplication. So, the LineOfSightVector is always (0,0,-1.0), no matter of any camera rotation. That means,

  Specular Reflection WON'T WORK when the ProjectionMatrix is rotated (!)

So, if you want to rotate the "camera" (in MTX_MODE=0), then you must instead rotate the "world" in the opposite direction (in MTX_MODE=2).
That problem applies only to Specular Reflection, ie. only if Lighting is used, and only if the Specular Material Color is nonzero.

Maths Notes
Note on Vector*Vector multiplication: Processed as LineVector*RowVector, so the result is a number (aka scalar) (aka a matrix with only 1x1 elements), multiplying two (normalized) vectors results in: "cos(angle)=vec1*vec2", ie. the cosine of the angle between the two vectors.
The various Normal/Light/Half/Sight vectors are only 3-dimensional (x,y,z), ie. only the upper-left 3x3 matrix elements are used on multiplications with the 4x4 DirectionalMatrix.

DS 3D Shadow Polygons

The DS hardware's Light-function allows to reflect light to the camera, it does not reflect light to other polygons, and it does not cast any shadows. For shadows at fixed locations it'd be best to pre-calculate their shape and position, and to change the vertex color of the shaded polygons.
Additionally, the Shadow Polygon feature can be used to create animated shadows, ie. moved objects and variable light sources.

Shadow Polygons and Shadow Volume
The software must define a Shadow Volume (ie. the region which doesn't contain light), the hardware does then automatically draw the shadow on all pixels whose x/y/z-coordinates are inside of that region.
The Shadow Volume must be defined by several Shadow Polygons which are enclosing the shaded region. The 'top' of the shadow volume should be usually translated to the position of the object that casts the shadow, if the light direction changes then the shadow volume should be also rotated to match the light direction. The 'length' of the shadow volume should be (at least) long enough to reach from the object to the walls/floor where the shadow is to be drawn. The shadow volume must be passed TWICE to the hardware:

Step 1 - Shadow Volume for Mask
Set Polygon_Attr Mode=Shadow, PolygonID=00h, Back=Render, Front=Hide, Alpha=01h..1Eh, and pass the shadow volume (ie. the shadow polygons) to the geometry engine.
The Back=Render / Front=Hide setting causes the 'rear-side' of the shadow volume to be rendered, of course only as far as it is in front of other polygons. The Mode=Shadow / ID=00h setting causes the polygon NOT to be drawn to the Color Buffer - instead, flags are set in the Stencil Buffer (to be used in Step 2).

Step 2 - Shadow Volume for Rendering
Simply repeat step 1, but with Polygon_Attr Mode=Shadow, PolygonID=01h..3Fh, Back=Render(what/why?), Front=Render, Alpha=01h..1Eh.
The Front=Render setting causes the 'front-side' of the shadow volume to be rendered, again, only as far as it is in front of other polygons. The Mode=Shadow / ID>00h setting causes the polygon to be drawn to the Color Buffer as usually, but only if the Stencil Buffer bits are zero (ie. the portion from Step 1 is excluded) (additionally, Step 2 resets the stencil bits after checking them). Moreover, the shadow is rendered only if its Polygon ID differs from the ID in the Attribute Buffer.

Shadow Alpha and Shadow Color
The Alpha=Translucent setting in Step 1 and 2 ensures that the Shadow is drawn AFTER the normal (opaque) polygons have been rendered. In Step 2 it does additionally specify the 'intensity' of the shadow. For normal shadows, the Vertex Color should be usually black, however, the shadow volume may be also used as 'spotlight volume' when using other colors.

Rendering Order
The Mask Volume must be rendered prior to the Rendering Volume, ie. Step 1 and 2 must be performed in that order, and, to keep that order intact, Auto-sorting must have been disabled in the previous Swap_Buffers command.
The shadow volume must be rendered after the 'target' polygons have been rendered, for opaque targets this is done automatically (due to the translucent alpha setting; translucent polygons are always rendered last, even with auto-sort disabled).

Translucent Targets
Casting shadows on Translucent Polygons. First draw the translucent target (with update depth buffer enabled, required for the shadow z-coordinates), then draw the Shadow Mask/Rendering volumes.
Due to the updated depth buffer the shadow will be cast only on the translucent target (not on any other polygons underneath of the translucent polygon). If you want the shadow to appear on both: Draw draw the Shadow Mask/Rendering volume TWICE (once before, and once after drawing the translucent target).

Polygon ID and Fog Enable
The "Render only if Polygon ID differs" feature (see Step 2) allows to prevent the shadow to be cast on the object that casts the shadow (ie. the object and shadow should have the same IDs). The feature also allows to select whether overlapping shadows (with same/different IDs) are shaded once or twice.
The old Fog Enable flag in the Attribute Buffer is ANDed with the Fog Enable flag of the Shadow Polygons, this allows to exclude Fog in shaded regions.

Shadow Volume Open/Closed Shapes
Normally, the shadow volume should have a closed shape, ie. should have rear-sides (step 1), and corresponding front-sides (step 2) for all possible viewing angles. That is required for the shadow to be drawn correctly, and also for the Stencil Buffer to be reset to zero (in step 2, so that the stencil bits won't disturb other shadow volumes).
Due to that, drawing errors may occur if the shadow volume's front or rear side gets clipped by near/far clip plane.
One exception is that the volume doesn't need a bottom-side (with a suitable volume length, the bottom may be left open, since it vanishes in the floor/walls anyways).

DS 3D Texture Attributes

4000488h - Cmd 22h - TEXCOORD - Set Texture Coordinates (W)
Specifies the texture source coordinates within the texture bitmap which are to be associated with the next vertex.

  Parameter 1, Bit 0-15   S-Coordinate (X-Coordinate in Texture Source)
  Parameter 1, Bit 16-31  T-Coordinate (Y-Coordinate in Texture Source)
  Both values are 1bit sign + 11bit integer + 4bit fractional part.
  A value of 1.0 (=1 SHL 4) equals to one Texel.

With Position 0.0 , 0.0 drawing starts from upperleft of the Texture.
With positive offsets, drawing origin starts more "within" the texture.
With negative offsets, drawing starts "before" the texture.
"When texture mapping, the Geometry Engine works faster if you issue commands in the order TexCoord -> Normal -> Vertex."

40004A8h - Cmd 2Ah - TEXIMAGE_PARAM - Set Texture Parameters (W)

  0-15  Texture VRAM Offset div 8 (0..FFFFh -> 512K RAM in Slot 0,1,2,3)
        (VRAM must be allocated as Texture data, see Memory Control chapter)
  16    Repeat in S Direction (0=Clamp Texture, 1=Repeat Texture)
  17    Repeat in T Direction (0=Clamp Texture, 1=Repeat Texture)
  18    Flip in S Direction   (0=No, 1=Flip each 2nd Texture) (requires Repeat)
  19    Flip in T Direction   (0=No, 1=Flip each 2nd Texture) (requires Repeat)
  20-22 Texture S-Size        (for N=0..7: Size=(8 SHL N); ie. 8..1024 texels)
  23-25 Texture T-Size        (for N=0..7: Size=(8 SHL N); ie. 8..1024 texels)
  26-28 Texture Format        (0..7, see below)
  29    Color 0 of 4/16/256-Color Palettes (0=Displayed, 1=Made Transparent)
  30-31 Texture Coordinates Transformation Mode (0..3, see below)

Texture Formats:

  0  No Texture
  1  A3I5 Translucent Texture
  2  4-Color Palette Texture
  3  16-Color Palette Texture
  4  256-Color Palette Texture
  5  4x4-Texel Compressed Texture
  6  A5I3 Translucent Texture
  7  Direct Texture

Texture Coordinates Transformation Modes:

  0  Do not Transform texture coordinates
  1  TexCoord source
  2  Normal source
  3  Vertex source

The S-Direction equals to the horizontal direction of the source bitmap.
The T-Direction, T-repeat, and T-flip are the same in vertical direction.
For a "/" shaped texture, the S-clamp, S-repeat, and S-flip look like so:

  Clamp _____  Repeat       Repeat+Flip
  _____/       ///////////  /\/\/\/\/\/

With "Clamp", the texture coordinates are clipped to MinMax(0,Size-1), so the texels at the edges of the texture bitmap are repeated (to avoid that effect, fill the bitmap edges by texels with alpha=0, so they become invisible).

40004ACh - Cmd 2Bh - PLTT_BASE - Set Texture Palette Base Address (W)

  0-12   Palette Base Address (div8 or div10h, see below)
         (Not used for Texture Format 7: Direct Color Texture)
         (0..FFF8h/8 for Texture Format 2: ie. 4-color-palette Texture)
         (0..17FF0h/10h for all other Texture formats)
  13-31  Not used

The palette data occupies 16bit per color, Bit0-4: Red, Bit5-9: Green, Bit10-14: Blue, Bit15: Not used.
(VRAM must be allocated as Texture Palette, there can be up to 6 Slots allocated, ie. the addressable 18000h bytes, see Memory Control chapter)

TexImageParam and TexPlttBase
Can be issued per polygon (except within polygon strips).

DS 3D Texture Formats

Format 2: 4-Color Palette Texture
Each Texel occupies 2bit, the first Texel is located in LSBs of 1st byte.
In this format, the Palette Base is specified in 8-byte steps; all other formats use 16-byte steps (see PLTT_BASE register).

Format 3: 16-Color Palette Texture
Each Texel occupies 4bit, the 1st Texel is located in LSBs of 1st byte.

Format 4: 256-Color Palette Texture
Each Texel occupies 8bit, the 1st Texel is located in 1st byte.

Format 7: Direct Color Texture
Each Texel occupies 16bit, the 1st Texel is located in 1st halfword.
Bit0-4: Red, Bit5-9: Green, Bit10-14: Blue, Bit15: Alpha

Format 1: A3I5 Translucent Texture (3bit Alpha, 5bit Color Index)
Each Texel occupies 8bit, the 1st Texel is located in 1st byte.

  Bit0-4: Color Index (0..31) of a 32-color Palette
  Bit5-7: Alpha       (0..7; 0=Transparent, 7=Solid)

The 3bit Alpha value (0..7) is internally expanded into a 5bit Alpha value (0..31) as follows: Alpha=(Alpha*4)+(Alpha/2).

Format 6: A5I3 Translucent Texture (5bit Alpha, 3bit Color Index)
Each Texel occupies 8bit, the 1st Texel is located in 1st byte.

  Bit0-2: Color Index (0..7) of a 8-color Palette
  Bit3-7: Alpha       (0..31; 0=Transparent, 31=Solid)

Format 5: 4x4-Texel Compressed Texture
Consists of 4x4 Texel blocks in Slot 0 or 2, 32bit per block, 2bit per Texel,

  Bit0-7   Upper 4-Texel row (LSB=first/left-most Texel)
  Bit8-15  Next  4-Texel row ("")
  Bit16-23 Next  4-Texel row ("")
  Bit24-31 Lower 4-Texel row ("")

Additional Palette Index Data for each 4x4 Texel Block is located in Slot 1,

  Bit0-13  Palette Offset in 4-byte steps; Addr=(PLTT_BASE*10h)+(Offset*4)
  Bit14-15 Transparent/Interpolation Mode (0..3, see below)

whereas, the Slot 1 offset is related to above Slot 0 or 2 offset,

  slot1_addr = slot0_addr / 2           ;lower 64K of Slot1 assoc to Slot0
  slot1_addr = slot2_addr / 2 + 10000h  ;upper 64K of Slot1 assoc to Slot2

The 2bit Texel values (0..3) are intepreted depending on the Mode (0..3),

  Texel  Mode 0       Mode 1             Mode 2         Mode 3
  0      Color 0      Color0             Color 0        Color 0
  1      Color 1      Color1             Color 1        Color 1
  2      Color 2      (Color0+Color1)/2  Color 2        (Color0*5+Color1*3)/8
  3      Transparent  Transparent        Color 3        (Color0*3+Color1*5)/8

Mode 1 and 3 are using only 2 Palette Colors (which requires only half as much Palette memory), the 3rd (and 4th) Texel Colors are automatically set to above values (eg. to gray-shades if color 0 and 1 are black and white).
Note: The maximum size for 4x4-Texel Compressed Textures is 1024x512 or 512x1024 (which are both occupying the whole 128K in slot 0 or 2, plus 64K in slot1), a larger size of 1024x1024 cannot be used because of the gap between slot 0 and 2.

DS 3D Texture Coordinates

For textured polygons, a texture coordinate must be associated with each vertex of the polygon. The coordinates (S,T) are defined by TEXCOORD command (typically issued prior to each VTX command), and can be optionally automatically transformed, by the Transformation Mode selected in TEXIMAGE_PARAM register.

Texture Matrix
Although the texture matrix is 4x4, with values m[0..15], only the left two columns of this matrix are actually used. In Mode 2 and 3, the bottom row of the matrix is replaced by S and T values from most recent TEXCOORD command.

Texture Coordinates Transformation Mode 0 - No Transform
The values are set upon executing the TEXCOORD command,

  ( S' T' )  =  ( S  T )

Simple coordinate association, without using the Texture Matrix at all.

Texture Coordinates Transformation Mode 1 - TexCoord source
The values are calculated upon executing the TEXCOORD command,

                                     | m[0]  m[1]  |
  ( S' T' )  =  ( S  T 1/16 1/16 ) * | m[4]  m[5]  |
                                     | m[8]  m[9]  |
                                     | m[12] m[13] |

Can be used to produce a simple texture scrolling, rotation, or scaling, by setting a translate, rotate, or scale matrix for the texture matrix.

Texture Coordinates Transformation Mode 2 - Normal source
The values are calculated upon executing the NORMAL command,

                                     | m[0]  m[1]  |
  ( S' T' )  =  ( Nx  Ny  Nz 1.0 ) * | m[4]  m[5]  |
                                     | m[8]  m[9]  |
                                     | S     T     |

Can be used to produce spherical reflection mapping by setting the texture matrix to the current directional vector matrix, multiplied by a scaling matrix that expands the directional vector space from -1.0..+1.0 to one half of the texture size. For that purpose, translate the origin of the texture coordinate to the center of the spherical texture by using TexCoord command (spherical texture means a bitmap that contains some circle-shaped image).

Texture Coordinates Transformation Mode 3 - Vertex source
The values are calculated upon executing any VTX commands,

                                     | m[0]  m[1]  |
  ( S' T' )  =  ( Vx  Vy  Vz 1.0 ) * | m[4]  m[5]  |
                                     | m[8]  m[9]  |
                                     | S     T     |

Can be used to produce texture scrolls dependent on the View coordinates by copying the current position coordinate matrix into the texture matrix. For example, the PositionMatrix can be obtained via CLIPMTX_RESULT (see there for details), and that values can be then manually copied to the TextureMatrix.

Sign+Integer+Fractional Parts used in above Formulas

  Matrix    m[..]     1+19+12 (32bit)
  Vertex    Vx,Vy,Vz  1+3+12  (16bit)
  Normal    Nx,Ny,Nz  1+0+9   (10bit)
  Constant  1.0       0+1+0   (1bit)
  Constant  1/16      0+0+4   (4bit)
  TexCoord  S,T       1+11+4  (16bit)
  Result    S',T'     1+11+4  (16bit) <-------- clipped to that size !

Observe that the S',T' values are clipped to 16bit size. Ie. after the Vector*Matrix calaction, the result is shifted right (to make it having a 4bit fraction), and the value is then masked to 16bit size.

DS 3D Texture Blending

Polygon pixels consist of a Vertex Color, and of Texture Colors.
These colors can be blended as described below. Or, to use only either one:
To use only the Vertex Color: Select No Texture in TEXIMAGE_PARAM.
To use only the Texture Color: Select Modulation Mode and Alpha=31 in POLYGON_ATTR, and set COLOR to 7FFFh (white), or to gray values (to decrease brightness of the texture color).

Vertex Color (Rv,Gv,Bv,Av)
The Vertex Color (Rv,Gv,Bv) can be changed per Vertex (either by Color, Normal, or Material0 command), pixels between vertices are shaded to medium values of the surrounding vertices. The Vertex Alpha (Av), can be changed only per polygon (by PolygonAttr command).

Texture Colors (Rt,Gt,Bt,At)
The Texture Colors (Rt,Gt,Bt), and Alpha value (At), are defined by the Texture Bitmap. For formats without Alpha value, assume At=31 (solid), and for formats with 1bit Alpha assume At=A*31.

Shading Table Colors (Rs,Gs,Bs)
In Toon/Highlight Shading Mode, the red component of the Vertex Color (Rv) is mis-used as an index in the Shading Table, ie. Rv is used to read Shading Colors (Rs,Gs,Bs) from the table; the green and blue components of the Vertex Color (Gv,Bv) are unused in this mode. The Vertex Alpha (Av) is kept used.
Shading is used in Polygon Mode 2, whether it is Toon or Highlight Shading is selected in DISP3DCNT; this is a per-frame selection, so only either one can be used.

Texture Blending - Modulation Mode (Polygon Attr Mode 0)

  R = ((Rt+1)*(Rv+1)-1)/64
  G = ((Gt+1)*(Gv+1)-1)/64
  B = ((Bt+1)*(Bv+1)-1)/64
  A = ((At+1)*(Av+1)-1)/64

The multiplication result is decreased intensity (unless both factors are 63).

Texture Blending - Decal Mode (Polygon Attr Mode 1)

  R = (Rt*At + Rv*(63-At))/64  ;except, when At=0: R=Rv, when At=31: R=Rt
  G = (Gt*At + Gv*(63-At))/64  ;except, when At=0: G=Gv, when At=31: G=Gt
  B = (Bt*At + Bv*(63-At))/64  ;except, when At=0: B=Bv, when At=31: B=Bt
  A = Av

The At value is used (only) as ratio for Texture color vs Vertex Color.

Texture Blending - Toon Shading (Polygon Mode 2, DISP3DCNT=Toon)
The vertex color Red component (Rv) is used as an index in the toon table.

  R = ((Rt+1)*(Rs+1)-1)/64   ;Rs=ToonTableRed[Rv]
  G = ((Gt+1)*(Gs+1)-1)/64   ;Gs=ToonTableGreen[Rv]
  B = ((Bt+1)*(Bs+1)-1)/64   ;Bs=ToonTableBlue[Rv]
  A = ((At+1)*(Av+1)-1)/64

This is same as Modulation Mode, but using Rs,Gs,Bs instead Rv,Gv,Bv.

Texture Blending - Highlight Shading (Polygon Mode 2, DISP3DCNT=Highlight)

  R = ((Rt+1)*(Rs+1)-1)/64+Rs ;truncated to MAX=63
  G = ((Gt+1)*(Gs+1)-1)/64+Gs ;truncated to MAX=63
  B = ((Bt+1)*(Bs+1)-1)/64+Bs ;truncated to MAX=63
  A = ((At+1)*(Av+1)-1)/64

Same as Toon Shading, with additional addition offset, the addition may increase the intensity, however, it may also change the hue of the color.

Above formulas are for 6bit RGBA values, ie. 5bit values internally expanded to 6bit as such: IF X>0 THEN X=X*2+1.

Uni-Colored Textures
Although textures are normally containing "pictures", in some cases it makes sense to use "blank" textures that are filled with a single color:
Wire-frame polygons are always having Av=31, however, they can be made transparent by using Translucent Textures (ie. A5I3 or A3I5 formats) with At<31.
In Toon/Highlight shading modes, the Vertex Color is mis-used as table index, however, Toon/Highlight shading can be used on uni-colored textures, which is more or less the same as using Toon/Highlight shading on uni-colored Vertex-colors.

DS 3D Toon, Edge, Fog, Alpha-Blending, Anti-Aliasing

4000380h..3BFh - TOON_TABLE - Toon Table (W)
This 64-byte region contains the 32 toon colors (16bit per color), used for both Toon and Highlight Shading. In both modes, the Red (R) component of the RGBA vertex color is mis-used as index to obtain the new RGB value from the toon table, vertex Alpha (A) is kept used as is.

  Bit0-4: Red, Bit5-9: Green, Bit10-14: Blue, Bit15: Not Used

Shading can be enabled (per polygon) in Polygon_Attr, whether it is Toon or Highlight Shading is set (per frame) in DISP3DCNT. For more info on shading, see:
DS 3D Texture Blending

4000330h..33Fh - EDGE_COLOR - Edge Colors 0..7 (W)
This 16-byte region contains the 8 edge colors (16bit per color), Edge Color 0 is used for Polygon ID 00h..07h, Color 1 for ID 08h..0Fh, and so on.

  Bit0-4: Red, Bit5-9: Green, Bit10-14: Blue, Bit15: Not Used

Edge Marking allows to mark the edges of an object (whose polygons all have the same ID) in a wire-frame style. Edge Marking can be enabled (per frame) in DISP3DCNT. When enabled, the polygon edges are drawn at the edge color, but only if the old ID value in the Attribute Buffer is different than the Polygon ID of the new polygon, so no edges are drawn between connected or overlapping polygons with same ID values.
Edge Marking is applied ONLY to opaque polygons (including wire-frames).
Edge Marking increases the size of opaque polygons (see notes below).
Edge Marking doesn't work very well with Anti-Aliasing (see Anti-Aliasing).
Technically, when rendering a polygon, it's edges (ie. the wire-frame region) are flagged as possible-edges (but it's still rendered normally, without using the edge-color). Once when all opaque polygons (*) have been rendered, the edge color is applied to these flagged pixels, under following conditions: At least one of the four surrounding pixels (up, down, left, right) must have different polygon_id than the edge, and, the edge depth must be LESS than the depth of that surrounding pixel (ie. no edges are rendered if the depth is GREATER or EQUAL, even if the polygon_id differs). At the screen borders, edges seem to be rendered in respect to the rear-plane's polygon_id entry (see Port 4000350h).
(*) Actually, edge-marking is reportedly performed not until all opaque AND translucent polygons have been rendered. That brings up some effects/problems when edges are covered by translucent polys: The edge-color is probably drawn as is (ie. it'll overwrite the translucent color, rather than being blended with the translucent color). And, any translucent polygons that do update the depth buffer will cause total edge-marking malfunction (since edge-marking involves the comparision of the current/surrounding pixel's depth values).

4000358h - FOG_COLOR - Fog Color (W)
Fog can be used to let more distant polygons to disappear in foggy grayness (or in darkness, or other color). This is particulary useful to "hide" the far clip plane. Fog can be enabled in DISP3DCNT.Bit7, moreover, when enabled, it can be activated or deactivated per polygon (POLYGON_ATTR.Bit15), and per Rear-plane (see there).

  0-4    Fog Color, Red     ;\
  5-9    Fog Color, Green   ; used only when DISP3DCNT.Bit6 is zero
  10-14  Fog Color, Blue    ;/
  15     Not used
  16-20  Fog Alpha          ;-used no matter of DISP3DCNT.Bit6
  21-31  Not used

Whether or not fog is applied to a pixel depends on the Fog flag in the framebuffer, the initial value of that flag can be defined in the rear-plane. When rendering opaque pixels, the framebuffer's fog flag gets replaced by PolygonAttr.Bit15. When rendering translucent pixels, the old flag in the framebuffer gets ANDed with PolygonAttr.Bit15.

400035Ch - FOG_OFFSET - Fog Depth Offset (W)

  0-14   Fog Offset (Unsigned) (0..7FFFh)
  15-31  Not used

FogDepthBoundary[0..31] (for FogDensity[0..31]) are defined as:

  FogDepthBoundary[n] = FOG_OFFSET + FOG_STEP*(n+1)   ;with n = 0..31

Whereas FOG_STEP is derived from the FOG_SHIFT value in DISP3DCNT.Bit8-11 (FOG_STEP=400h shr FOG_SHIFT) (normally FOG_SHIFT should be 0..10 (bigger shift amounts of 11..15 would cause FOG_STEP to become zero, so only Density[0] and Density[31] would be used).
The meaning of the depth values depends on whether z-values or w-values are stored in the framebuffer (see SwapBuffers.Bit1).
For translucent polygons, the depth value (and therefore: the amount of fog) depends on the depth update bit (see PolygonAttr.Bit11).

4000360h..37Fh - FOG_TABLE - Fog Density Table (W)
This 32-byte region contains FogDensity[0..31] (used at FogDepthBoundary[n]),

  0-6    Fog Density (00h..7Fh = None..Full) (usually increasing values)
  7      Not used

FogDensity[0] is used for all pixels closer than FogDepthBoundary[0], FogDensity[31] is used for all pixels more distant than FogDepthBoundary[0].
Density is linear interpolated for pixels that are between two Density depth boundaries. The formula for Fog Blending is:

  FrameBuffer[R] = (FogColor[R]*Density + FrameBuffer[R]*(128-Density)) / 128
  FrameBuffer[G] = (FogColor[G]*Density + FrameBuffer[G]*(128-Density)) / 128
  FrameBuffer[B] = (FogColor[B]*Density + FrameBuffer[B]*(128-Density)) / 128
  FrameBuffer[A] = (FogColor[A]*Density + FrameBuffer[A]*(128-Density)) / 128

If DISP3DCNT.Bit6 is set (=Alpha Only), then only FrameBuffer[A] is updated, and FrameBuffer[RGB] are kepth unchanged. Density=127 is handled as if Density=128.
Fog Glitch: The fog_alpha value appears to be ignored (treated as fog_alpha=1Fh) in the region up to the first density boundary. However, normally that value will be multiplied by zero (assumung that density[0] is usually zero), so you won't ever notice that hardware glitch.

Alpha-Blending (Polygon vs FrameBuffer)
Alpha-Blending occurs for pixels of translucent polygons,

  FrameBuf[R] = (Poly[R]*(Poly[A]+1) + FrameBuf[R]*(31-(Poly[A])) / 32
  FrameBuf[G] = (Poly[G]*(Poly[A]+1) + FrameBuf[G]*(31-(Poly[A])) / 32
  FrameBuf[B] = (Poly[B]*(Poly[A]+1) + FrameBuf[B]*(31-(Poly[A])) / 32
  FrameBuf[A] = max(Poly[A],FrameBuf[A])

There are three situations in which Alpha-Blending is bypassed (the old Framebuf[R,G,B,A] value is then simply overwritten by Poly[R,G,B,A]):

  1) Alpha-Blending is disabled                       (DISP3DCNT.Bit3=0)
  2) The polygon pixel is opaque                      (Poly[A]=31)
  3) The old framebuffer value is totally transparent (FrameBuf[A]=0)

The third case can happen if the rear-plane was initialized with Alpha=0, which causes the polygon not to be blended with the rear-plane (which may give better results when subsequently blending the 3D layer with the 2D engine).
Note: Totally transparent pixels (with Poly[A]=0) are not rendered (ie. neither FrameBuf[R,G,B,A] nor FrameBuf[Depth,Fog,PolyID,etc.] are updated.

Anti-Aliasing
Anti-Aliasing can be enabled in DISP3DCNT, when enabled, the edges of opaque polygons will be anti-aliased (ie. the pixels at the edges may become translucent).
Anti-Aliasing is not applied on translucent polygons. And, Anti-Aliasing is not applied on the interiors of the poylgons (eg. an 8x8 chessboard texture will be anti-aliased only at the board edges, not at the edges of the 64 fields).
Anti-Aliasing is (accidently) applied to opaque 1dot polygongs, line-segments and wire-frames (which results in dirty lines with missing pixels, 1dot polys become totally invisible), workaround is to use translucent dots, lines and wires (eg. with alpha=30).
Anti-Aliasing is (correctly) not applied to edges of Edge-Marked polygons, in that special case even opaque line-segments and wire-frames are working even if anti-aliasing is enabled (provided that they are edge-marked, ie. if their polygon ID differs from the framebuffer's ID).
Anti-Aliasing is (accidently) making the edges of Edge-Marked polygons translucent (with alpha=16 or so?), that reduces the contrast of the edge colors. Moreover, if two of these translucent edges do overlap, then they are blended twice (even if they have the same polygon_id, and even if the depth_update bit in polygon_attr is set; both should normally prevent double-blending), that scatters the brightness of such edges.

Polygon Size
In some cases, the NDS hardware doesn't render the lower/right edges of certain polygons. That feature reduces rendering load, and, when rendering connected polygons (eg. strips), then it'd be unnecessary to render that edges (since they'd overlap with the upper/left edges of the other polygon). On the contrary, if there's no connected polygon displayed, then the polygon may appear smaller than expected. Small polygons with excluded edges are:

  Opaque polygons (except wire-frames) without Edge-Marking and Anti-Aliasing,
  and, all polygons with vertical right-edges (except line-segments).
  Plus, Translucent Polys when Alpha-Blending is disabled in DISP3DCNT.Bit3.

All other polygons are rendered at full size with all edges included (except vertical right edges). Note: To disable the small-polygon feature, you can enable edge-marking (which does increase the polygon size, even if no edges are drawn, ie. even if all polys do have the same ID).

DS 3D Status

4000600h - GXSTAT - Geometry Engine Status Register (R and R/W)
Bit 30-31 are R/W. Writing "1" to Bit15 does reset the Error Flag (Bit15), and additionally resets the Projection Stack Pointer (Bit13), and probably (?) also the Texture Stack Pointer. All other GXSTAT bits are read-only.

  0     BoxTest,PositionTest,VectorTest Busy (0=Ready, 1=Busy)
  1     BoxTest Result  (0=All Outside View, 1=Parts or Fully Inside View)
  2-7   Not used
  8-12  Position & Vector Matrix Stack Level (0..31) (lower 5bit of 6bit value)
  13    Projection Matrix Stack Level        (0..1)
  14    Matrix Stack Busy (0=No, 1=Yes; Currently executing a Push/Pop command)
  15    Matrix Stack Overflow/Underflow Error (0=No, 1=Error/Acknowledge/Reset)
  16-24 Number of 40bit-entries in Command FIFO  (0..256)
 (24)   Command FIFO Full (MSB of above)  (0=No, 1=Yes; Full)
  25    Command FIFO Less Than Half Full  (0=No, 1=Yes; Less than Half-full)
  26    Command FIFO Empty                (0=No, 1=Yes; Empty)
  27    Geometry Engine Busy (0=No, 1=Yes; Busy; Commands are executing)
  28-29 Not used
  30-31 Command FIFO IRQ (0=Never, 1=Less than half full, 2=Empty, 3=Reserved)

When GXFIFO IRQ is enabled (setting 1 or 2), the IRQ flag (IF.Bit21) is set while and as long as the IRQ condition is true (and attempts to acknowledge the IRQ by writing to IF.Bit21 have no effect). So that, the IRQ handler must either fill the FIFO, or disable the IRQ (setting 0), BEFORE trying to acknowledge the IRQ.

4000604h - RAM_COUNT - Polygon List & Vertex RAM Count Register (R)

  0-11   Number of Polygons currently stored in Polygon List RAM (0..2048)
  12-15  Not used
  16-28  Number of Vertices currently stored in Vertex RAM       (0..6144)
  13-15  Not used

If a SwapBuffers command has been sent, then the counters are reset 10 cycles (at 33.51MHz clock) after next VBlank.

4000320h - RDLINES_COUNT - Rendered Line Count Register (R)
Rendering starts in scanline 214, the rendered lines are stored in a buffer that can hold up to 48 scanlines. The actual screen output begins after scanline 262, the lines are then read from the buffer and sent to the display. Simultaneously, the rendering engine keeps writing new lines to the buffer (ideally at the same speed than display output, so the buffer would always contain 48 pre-calculated lines).

  0-5    Minimum Number (minus 2) of buffered lines in previous frame (0..46)
  6-31   Not used

If rendering becomes slower than the display output, then the number of buffered lines decreases. Smaller values in RDLINES indicate that additional load to the rendering engine may cause buffer underflows in further frames, if so, the program should reduce the number of polygons to avoid display glitches.
Even if RDLINES becomes zero, it doesn't indicate whether actual buffer underflows have occured or not (underflows are indicated in DISP3DCNT Bit12).

DS 3D Tests

40005C0h - Cmd 70h - BOX_TEST - Test if Cuboid Sits inside View Volume (W)
The BoxTest result indicates if one or more of the 6 faces of the box are fully or parts of inside of the view volume. Can be used to reduce unnecessary overload, ie. if the result is false, then the program can skip drawing of objects which are inside of the box.
BoxTest verifies only if the faces of the box are inside view volume, and so, it will return false if the whole view volume is located inside of the box (still objects inside of the box may be inside of view).

  Parameter 1, Bit 0-15   X-Coordinate
  Parameter 1, Bit 16-31  Y-Coordinate
  Parameter 2, Bit 0-15   Z-Coordinate
  Parameter 2, Bit 16-31  Width  (presumably: X-Offset?)
  Parameter 3, Bit 0-15   Height (presumably: Y-Offset?)
  Parameter 3, Bit 16-31  Depth  (presumably: Z-Offset?)
  All values are 1bit sign, 3bit integer, 12bit fractional part

The result of the "coordinate+offset" additions should not overflow 16bit vertex coordinate range (1bit sign, 3bit integer, 12bit fraction).
Before using BoxTest, be sure that far-plane-intersecting & 1-dot polygons are enabled, if they aren't: Send the PolygonAttr command (with bit12,13 set to enable them), followed by dummy Begin and End commands (required to apply the new PolygonAttr settings). BoxTest should not be issued within Begin/End.
After sending the BoxTest command, wait until GXSTAT.Bit0 indicates Ready, then read the result from GXSTAT.Bit1.

40005C4h - Cmd 71h - POS_TEST - Set Position Coordinates for Test (W)

  Parameter 1, Bit 0-15   X-Coordinate
  Parameter 1, Bit 16-31  Y-Coordinate
  Parameter 2, Bit 0-15   Z-Coordinate
  Parameter 2, Bit 16-31  Not used
  All values are 1bit sign, 3bit integer, 12bit fractional part.

Multiplies the specified line-vector (x,y,z,1) by the clip coordinate matrix.
After sending the command, wait until GXSTAT.Bit0 indicates Ready, then read the result from POS_RESULT registers. POS_TEST can be issued anywhere (except within polygon strips, huh?).
Caution: POS_TEST overwrites the internal VTX registers, so the next vertex should be <fully> defined by VTX_10 or VTX_16, otherwise, when using VTX_XY, VTX_XZ, VTX_YZ, or VTX_DIFF, then the new vertex will be relative to the POS_TEST coordinates (rather than to the previous vertex).

4000620h..62Fh - POS_RESULT - Position Test Results (R)
This 16-byte region (4 words) contains the resulting clip coordinates (x,y,z,w) from the POS_TEST command. Each value is 1bit sign, 19bit integer, 12bit fractional part.

40005C8h - Cmd 72h - VEC_TEST - Set Directional Vector for Test (W)

  Parameter 1, Bit 0-9    X-Component
  Parameter 1, Bit 10-19  Y-Component
  Parameter 1, Bit 20-29  Z-Component
  Parameter 1, Bit 30-31  Not used
  All values are 1bit sign, 9bit fractional part.

Multiplies the specified line-vector (x,y,z,0) by the directional vector matrix. Similar as for the NORMAL command, it does require Matrix Mode 2 (ie. Position & Vector Simultaneous Set mode).
After sending the command, wait until GXSTAT.Bit0 indicates Ready, then read the result ("the directional vector in the View coordinate space") from VEC_RESULT registers.

4000630h..635h - VEC_RESULT - Vector Test Results (R)
This 6-byte region (3 halfwords) contains the resulting vector (x,y,z) from the VEC_TEST command. Each value is 4bit sign, 0bit integer, 12bit fractional part. The 4bit sign is either 0000b (positive) or 1111b (negative).
There is no integer part, so values >=1.0 or <-1.0 will cause overflows.
(Eg. +1.0 aka 1000h will be returned as -1.0 aka F000h due to overflow and sign-expansion).

DS 3D Rear-Plane

Other docs seem to refer to this as Clear-plane, rather than Rear-plane, anyways, the plane can be an image, so it isn't always "cleared".
The view order is as such:

  --> 2D Layers --> 3D Polygons --> 3D Rear-plane --> 2D Layers --> 2D Backdrop

The rear-plane can be disabled (by making it transparent; alpha=0), so that the 2D layers become visible as background.
2D layers can be moved in front of, or behind the 3D layer-group (which is represented as BG0 to the 2D Engine), 2D layers behind BG0 can be used instead of, or additionally to the rear-plane.

The rear-plane can be initialized via below two registers (so all pixels in the plane have the same colors and attributes), this method is used when DISP3DCNT.14 is zero:

4000350h - CLEAR_COLOR - Clear Color Attribute Register (W)

  0-4    Clear Color, Red
  5-9    Clear Color, Green
  10-14  Clear Color, Blue
  15     Fog (enables Fog to the rear-plane) (doesn't affect Fog of polygons)
  16-20  Alpha
  21-23  Not used
  24-29  Clear Polygon ID (affects edge-marking, at the screen-edges?)
  30-31  Not used

4000354h - CLEAR_DEPTH - Clear Depth Register (W)

  0-14   Clear Depth (0..7FFFh) (usually 7FFFh = most distant)
  15     Not used
  16-31  See Port 4000356h, CLRIMAGE_OFFSET

The 15bit Depth is expanded to 24bit as "X=(X*200h)+((X+1)/8000h)*1FFh".

Rear Color/Depth Bitmaps
Alternately, the rear-plane can be initialized by bitmap data (allowing to assign different colors & attributes to each pixel), this method is used when DISP3DCNT.14 is set:
Consists of two bitmaps (one with color data, one with depth data), each containing 256x256 16bit entries, and so, each occupying a whole 128K slot,

  Rear Color Bitmap (located in Texture Slot 2)
    0-4    Clear Color, Red
    5-9    Clear Color, Green
    10-14  Clear Color, Blue
    15     Alpha (0=Transparent, 1=Solid) (equivalent to 5bit-alpha 0 and 31)
  Rear Depth Bitmap (located in Texture Slot 3)
    0-14   Clear Depth, expanded to 24bit as X=(X*200h)+((X+1)/8000h)*1FFh
    15     Clear Fog (Initial fog enable value)

This method requires VRAM to be allocated to Texture Slot 2 and 3 (see Memory Control chapter). Of course, in that case the VRAM is used as Rear-plane, and cannot be used for Textures.
The bitmap method is restricted to 1bit alpha values (the register-method allows to use a 5bit alpha value).
The Clear Polygon ID is kept defined in the CLEAR_COLOR register, even in bitmap mode.

4000356h - CLRIMAGE_OFFSET - Rear-plane Bitmap Scroll Offsets (W)
The visible portion of the bitmap is 256x192 pixels (regardless of the viewport setting, which is used only for polygon clipping). Internally, the bitmap is 256x256 pixels, so the bottom-most 64 rows are usually offscreen, unless scrolling is used to move them into view.

  Bit0-7   X-Offset (0..255; 0=upper row of bitmap)
  Bit8-14  Y-Offset (0..255; 0=left column of bitmap)

The bitmap wraps to the upper/left edges when exceeding the lower/right edges.

DS 3D Final 2D Output

The final 3D image (consisting of polygons and rear-plane) is passed to 2D Engine A as BG0 layer (provided that DISPCNT is configured to use 3D as BG0).

Scrolling
The BG0HOFS register (4000010h) can be used the scroll the 3D layer horizontally, the scroll region is 512 pixels, consisting of 256 pixels for the 3D image, followed by 256 transparent pixels, and then wrapped to the 3D image again. Vertical scrolling (and rotation/scaling) cannot be used on the 3D layer.

BG Priority Order
The lower 2bit of the BG0CNT register (4000008h) control the priority relative to other BGs and OBJs, so the 3D layer can be in front of or behind 2D layers. All other bits in BG0CNT have no effect on 3D, namely, mosaic cannot be used on the 3D layer.

Special Effects
Special Effects Registers (4000050h..54h) can be used as such:

  Brightness up/down with BG0 as 1st Target via EVY   (as for 2D)
  Blending with BG0 as 2nd Target via EVA/EVB         (as for 2D)
  Blending with BG0 as 1st Target via 3D Alpha-values (unlike as for 2D)

The latter method probably (?) uses per-pixel 3D alpha values as such: EVA=A/2, and EVB=16-A/2, without using the EVA/EVB settings in 4000052h.

Window Feature
Window Feature (4000040h..4Bh) can be used as for 2D.
"If the 3D screen has highest priority, then alpha-blending is always enabled, regardless of the Window Control register's color effect enable flag [ie. regardless of Bit5 of WIN0IN, WIN1IN, WINOBJ, WINOUT registers]"... not sure if that is true, and if it superseedes the effect selection in Port 4000050h...?

DS Sound

DS Sound Hardware
The DS contains 16 hardware sound channels.
The console contains two speakers, arranged left and right of the upper screen, and so, provides stereo sound even without using the headphone socket.
DS Sound Channels 0..15
DS Sound Control Registers
DS Sound Capture
DS Sound Block Diagrams
DS Sound Notes

DS Sound Files
DS Files - Sound (SDAT etc.)

Power control
When restoring power supply to the sound circuit, do not output any sound during the first 15 milliseconds.

DS Sound Channels 0..15

Each of the 16 sound channels occopies 16 bytes in the I/O region, starting with channel 0 at 4000400h..400040Fh, up to channel 15 at 40004F0h..40004FFh.

40004x0h - NDS7 - SOUNDxCNT - Sound Channel X Control Register (R/W)

  Bit0-6    Volume Mul   (0..127=silent..loud)
  Bit7      Not used     (always zero)
  Bit8-9    Volume Div   (0=Normal, 1=Div2, 2=Div4, 3=Div16)
  Bit10-14  Not used     (always zero)
  Bit15     Hold         (0=Normal, 1=Hold last sample after one-shot sound)
  Bit16-22  Panning      (0..127=left..right) (64=half volume on both speakers)
  Bit23     Not used     (always zero)
  Bit24-26  Wave Duty    (0..7) ;HIGH=(N+1)*12.5%, LOW=(7-N)*12.5% (PSG only)
  Bit27-28  Repeat Mode  (0=Manual, 1=Loop Infinite, 2=One-Shot, 3=Prohibited)
  Bit29-30  Format       (0=PCM8, 1=PCM16, 2=IMA-ADPCM, 3=PSG/Noise)
  Bit31     Start/Status (0=Stop, 1=Start/Busy)

All channels support ADPCM/PCM formats, PSG rectangular wave can be used only on channels 8..13, and white noise only on channels 14..15.

40004x4h - NDS7 - SOUNDxSAD - Sound Channel X Data Source Register (W)

  Bit0-26  Source Address (must be word aligned, bit0-1 are always zero)
  Bit27-31 Not used

40004x8h - NDS7 - SOUNDxTMR - Sound Channel X Timer Register (W)

  Bit0-15  Timer Value, Sample frequency, timerval=-(33513982Hz/2)/freq

The PSG Duty Cycles are composed of eight "samples", and so, the frequency for Rectangular Wave is 1/8th of the selected sample frequency.
For PSG Noise, the noise frequency is equal to the sample frequency.

40004xAh - NDS7 - SOUNDxPNT - Sound Channel X Loopstart Register (W)

  Bit0-15  Loop Start, Sample loop start position
           (counted in words, ie. N*4 bytes)

40004xCh - NDS7 - SOUNDxLEN - Sound Channel X Length Register (W)
The number of samples for N words is 4*N PCM8 samples, 2*N PCM16 samples, or 8*(N-1) ADPCM samples (the first word containing the ADPCM header). The Sound Length is not used in PSG mode.

  Bit0-21  Sound length (counted in words, ie. N*4 bytes)
  Bit22-31 Not used

Minimum length (the sum of PNT+LEN) is 4 words (16 bytes), smaller values (0..3 words) are causing hang-ups (busy bit remains set infinite, but no sound output occurs).

In One-shot mode, the sound length is the sum of (PNT+LEN).
In Looped mode, the length is (1*PNT+Infinite*LEN), ie. the first part (PNT) is played once, the second part (LEN) is repeated infinitely.

DS Sound Control Registers

4000500h - NDS7 - SOUNDCNT - Sound Control Register (R/W)

  Bit0-6   Master Volume       (0..127=silent..loud)
  Bit7     Not used            (always zero)
  Bit8-9   Left Output from    (0=Left Mixer, 1=Ch1, 2=Ch3, 3=Ch1+Ch3)
  Bit10-11 Right Output from   (0=Right Mixer, 1=Ch1, 2=Ch3, 3=Ch1+Ch3)
  Bit12    Output Ch1 to Mixer (0=Yes, 1=No) (both Left/Right)
  Bit13    Output Ch3 to Mixer (0=Yes, 1=No) (both Left/Right)
  Bit14    Not used            (always zero)
  Bit15    Master Enable       (0=Disable, 1=Enable)
  Bit16-31 Not used            (always zero)

4000504h - NDS7 - SOUNDBIAS - Sound Bias Register (R/W)

  Bit0-9   Sound Bias    (0..3FFh, usually 200h)
  Bit10-31 Not used      (always zero)

After applying the master volume, the signed left/right audio signals are in range -200h..+1FFh (with medium level zero), the Bias value is then added to convert the signed numbers into unsigned values (with medium level 200h).
BIAS output is always enabled, even when Master Enable (SOUNDCNT.15) is off.

The sampling frequency of the mixer is 1.04876 MHz with an amplitude resolution of 24 bits, but the sampling frequency after mixing with PWM modulation is 32.768 kHz with an amplitude resolution of 10 bits.

DS Sound Capture

The DS contains 2 built-in sound capture devices that can capture output waveform data to memory.
Sound capture 0 can capture output from left-mixer or output from channel 0.
Sound capture 1 can capture output from right-mixer or output from channel 2.

4000508h - NDS7 - SNDCAP0CNT - Sound Capture 0 Control Register (R/W)
4000509h - NDS7 - SNDCAP1CNT - Sound Capture 1 Control Register (R/W)

  Bit0     Control of Associated Sound Channels (ANDed with Bit7)
            SNDCAP0CNT: Output Sound Channel 1 (0=As such, 1=Add to Channel 0)
            SNDCAP1CNT: Output Sound Channel 3 (0=As such, 1=Add to Channel 2)
            Caution: Addition mode works only if BOTH Bit0 and Bit7 are set.
  Bit1     Capture Source Selection
            SNDCAP0CNT: Capture 0 Source (0=Left Mixer, 1=Channel 0/Bugged)
            SNDCAP1CNT: Capture 1 Source (0=Right Mixer, 1=Channel 2/Bugged)
  Bit2     Capture Repeat        (0=Loop, 1=One-shot)
  Bit3     Capture Format        (0=PCM16, 1=PCM8)
  Bit4-6   Not used              (always zero)
  Bit7     Capture Start/Status  (0=Stop, 1=Start/Busy)

4000510h - NDS7 - SNDCAP0DAD - Sound Capture 0 Destination Address (R/W)
4000518h - NDS7 - SNDCAP1DAD - Sound Capture 1 Destination Address (R/W)

  Bit0-26  Destination address (word aligned, bit0-1 are always zero)
  Bit27-31 Not used (always zero)

Capture start address (also used as re-start address for looped capture).

4000514h - NDS7 - SNDCAP0LEN - Sound Capture 0 Length (W)
400051Ch - NDS7 - SNDCAP1LEN - Sound Capture 1 Length (W)

  Bit0-15  Buffer length (1..FFFFh words) (ie. N*4 bytes)
  Bit16-31 Not used

Minimum length is 1 word (attempts to use 0 words are interpreted as 1 word).

SOUND1TMR - NDS7 - Sound Channel 1 Timer shared as Capture 0 Timer
SOUND3TMR - NDS7 - Sound Channel 3 Timer shared as Capture 1 Timer
There are no separate capture frequency registers, instead, the sample frequency of Channel 1/3 is shared for Capture 0/1. These channels are intended to output the captured data, so it makes sense that both capture and sound output use the same frequency.

For Capture 0, a=0, b=1, x=0.
For Capture 1, a=2, b=3, x=1.

Capture Bugs
The NDS contains two hardware bugs which do occur when capturing data from ch(a) (SNDCAPxCNT.Bit1=1), if so, either bug occurs depending on whether ch(a)+ch(b) addition is enabled or disabled (SNDCAPxCNT.Bit0).

  1) Both Negative Bug - SNDCAPxCNT Bit1=1, Bit0=0 (addition disabled)
   Capture data is accidently set to -8000h if ch(a) and ch(b) are both <0.
   Otherwise the correct capture result is returned, ie. plain ch(a) data,
   not being affected by ch(b) (since addition is disabled).
   Workaround: Ensure that ch(a) and/or ch(b) are >=0 (or disabled).
 2) Overflow Bug - SNDCAPxCNT Bit1=1, Bit0=1 (addition enabled)
   In this mode, Capture data isn't clipped to MinMax(-8000h,+7FFFh),
   instead, it is ANDed with FFFFh, so the sign bit is lost if the
   addition result ch(a)+ch(b) is less/greater than -8000h/+7FFFh.
   Workaround: Reduce ch(a)/ch(b) volume or data to avoid overflows.

These bugs occur only for capture (speaker output remains intact), and they occur only when capturing ch(a) (capturing mixer-output works flawless).

ch(a)+ch(b) Channel Addition
The ch(a)+ch(b) addition unit has 2 outputs, with slightly different results:

 1) Addition Result for Capture(x) when using capture source=ch(a):
  Addition is performed always, no matter of SOUNDCNT.Bit12/13.
  And, no matter of ch(a) enable, result is plain ch(b) if ch(a) is disabled.
  Result is 16bit (plus fraction) with overflow error (see Capture Bugs).
 2) Addition Result for Mixer (towards speakers, and capture source=mixer):
  Ch(b) is muted if ch(a) is disabled.
  Ch(b) is muted if ch(b) SOUNDCNT.Bit12/13 is set to "Ch(b) not to mixer".
  Result is 17bit (plus fraction) without overflow error.

Addition mode can be used only if the <corresponding> capture unit is enabled, ie. if SNDCAPxCNT (Bit0 AND Bit7)=1. If so, addition affects both mixers (and so, may also affect the <other> capture unit if it reads from mixer).

DS Sound Block Diagrams

Left Mixer with Capture 0
(Right Mixer with Capture 1, respectively)

                       _____
  Ch0.L ------------->|     |  .------------------------------> to Capture 0
               ___    |     |  |                  ___
  Ch1.L ---o->|Sel|-->|     |  |       Ch0..Ch15 |   |
           |  |___|   |Left |--o---------------->|   |
  Ch2.L ---|--------->|Mixer|                    |Sel|   ______    ____
           |   ___    |     |                Ch1 |   |  |Master|  |Add |
  Ch3.L -o-|->|Sel|-->|     | .----------------->|   |->|Volume|->|Bias|-> L
         | |  |___|   |     | |                  |   |  |______|  |____|
  Ch4.L -|-|--------->|     | |              Ch3 |   |
  ...   -|-|--------->|     | | .--------------->|   |
  Ch15.L-|-|--------->|_____| | |   ___          |   |
         | '------------------o-|->|Add| Ch1+Ch3 |   |
         '----------------------o->|___|-------->|___|

Channel 0 and 1, Capture 0 with input from Left Mixer
(Channel 2 and 3, Capture 1 with input from Right Mixer, respectively)

  ____     _________     ___     ___      ___
 |FIFO|-->|Channel 0|-->|Vol|-->|Add|-o->|Pan|--> Ch0.L
 |____|   |_________|   |___|   |___| |  |___|--> Ch0.R
  ____     _________     ___      ^   |
 |FIFO|<--|Capture 0|<--|Sel|<----|---'
 |____|   |_ _____ _|   |___|<----|-------------- Left Mixer
  ____     _:Timer:_     ___     _|_      ___
 |FIFO|-->|Channel 1|-->|Vol|-->|Sel|--->|Pan|--> Ch1.L
 |____|   |_________|   |___|   |___|    |___|--> Ch1.R

Channel 4 (Channel 5..15, respectively)

  ____     _________     ___              ___
 |FIFO|-->|Channel 4|-->|Vol|----------->|Pan|--> Ch4.L
 |____|   |_________|   |___|            |___|--> Ch4.R

The FIFO isn't used in PSG/Noise modes (supported on channel 8..15).

DS Sound Notes

Sound delayed Start/Restart (timing glitch)
A sound will be started/restarted when changing its start bit from 0 to 1, however, the sound won't start immediately: PSG/Noise starts after 1 sample, PCM starts after 3 samples, and ADPCM starts after 11 samples (3 dummy samples as for PCM, plus 8 dummy samples for the ADPCM header).

Sound Stop (timing note)
In one-shot mode, the Busy bit gets cleared automatically at the BEGIN of the last sample period, nethertheless (despite of the cleared Busy bit) the last sample is kept output until the END of the last sample period (or, if the Hold flag is set, then the last sample is kept output infinitely, that is, until Hold gets cleared, or until the sound gets restarted).

Hold Flag (appears useless/bugged)
The Hold flag allows to keep the last sample being output infinitely after the end of one-shot sounds. This feature is probably intended to allow to play two continous one-shot sound blocks (without producing any scratch noise upon small delays between both blocks, which would occur if the output level would drop to zero).
However, the feature doesn't work as intended. As described above, PCM8/PCM16 sound starts are delayed by 3 samples. With Hold flag set, old output level is acually kept intact during the 1st sample, but the output level drops to zero during 2nd-3rd sample, before starting the new sound in 4th sample.

7bit Volume and Panning Values

  data.vol   = data*N/128
  pan.left   = data*(128-N)/128
  pan.right  = data*N/128
  master.vol = data*N/128/64

Register settings of 0..126,127 are interpreted as N=0..126,128.

Max Output Levels
When configured to max volume (and left-most or right-most panning), each channel can span the full 10bit output range (-200h..1FFh) on one speaker, as well as the full 16bit input range (-8000h..7FFFh) on one capture unit.
(It needs 2 channels to span the whole range on BOTH speakers/capture units.)
Together, all sixteen channels could thus reach levels up to -1E00h..21F0h (with default BIAS=200h) on one speaker, and -80000h..+7FFF0h on one capture unit. However, to avoid overflows, speaker outputs are clipped to MinMax(0,3FFh), and capture inputs to MinMax(-8000h..+7FFFh).

Channel/Mixer Bit-Widths

  Step                           Bits  Min        Max
  0 Incoming PCM16 Data          16.0  -8000h     +7FFFh
  1 Volume Divider (div 1..16)   16.4  -8000h     +7FFFh
  2 Volume Factor (mul N/128)    16.11 -8000h     +7FFFh
  3 Panning (mul N/128)          16.18 -8000h     +7FFFh
  4 Rounding Down (strip 10bit)  16.8  -8000h     +7FFFh
  5 Mixer (add channel 0..15)    20.8  -80000h    +7FFF0h
  6 Master Volume (mul N/128/64) 14.21 -2000h     +1FF0h
  7 Strip fraction               14.0  -2000h     +1FF0h
  8 Add Bias (0..3FFh, def=200h) 15.0  -2000h+0   +1FF0h+3FFh
  9 Clip (min/max 0h..3FFh)      10.0  0          +3FFh

Table shows integer.fractional bits, and min/max values (without fraction).

Capture Clipping/Rounding
Incoming ch(a) is NOT clipped, ch(a)+ch(b) may overflow (see Capture Bugs).
Incoming mixer data (20.8bits) is clipped to 16.8bits (MinMax -8000h..7FFFh).
For PCM8 capture format, the 16.8 bits are divided by 100h (=8.16 bits).
If the MSB of the fractional part is set, then data is rounded towards zero.
(Positive values are rounded down, negative values are rounded up.)
The fractional part is then discarded, and plain integer data is captured.

PSG Sound
The output volume equals to PCM16 values +7FFFh (HIGH) and -7FFFh (LOW).
PSG sound is always Infinite (the SOUNDxLEN Register, and the SOUNDxCNT Repeat Mode bits have no effect). The PSG hardware doesn't support sound length, sweep, or volume envelopes, however, these effects can be produced by software with little overload (or, more typically, with enormous overload, depending on the programming language used).

PSG Wave Duty (channel 8..13 in PSG mode)
Each duty cycle consists of eight HIGH or LOW samples, so the sound frequency is 1/8th of the selected sample rate. The duty cycle always starts at the begin of the LOW period when the sound gets (re-)started.

  0  12.5% "_______-_______-_______-"
  1  25.0% "______--______--______--"
  2  37.5% "_____---_____---_____---"
  3  50.0% "____----____----____----"
  4  62.5% "___-----___-----___-----"
  5  75.0% "__------__------__------"
  6  87.5% "_-------_-------_-------"
  7   0.0% "________________________"

The Wave Duty bits exist and are read/write-able on all channels (although they are actually used only in PSG mode on channels 8-13).

PSG Noise (channel 14..15 in PSG mode)
Noise randomly switches between HIGH and LOW samples, the output levels are calculated, at the selected sample rate, as such:

  X=X SHR 1, IF carry THEN Out=LOW, X=X XOR 6000h ELSE Out=HIGH

The initial value when (re-)starting the sound is X=7FFFh. The formula is more or less same as "15bit polynomial counter" used on 8bit Gameboy and GBA.

PCM8 and PCM16
Signed samples in range -80h..+7Fh (PCM8), or -8000h..+7FFFh (PCM16).
The output volume of PCM8=NNh is equal to PCM16=NN00h.

IMA-ADPCM Format
IMA-ADPCM is a Adaptive Differential Pulse Code Modulation (ADPCM) variant, designed by International Multimedia Association (IMA), the format is used, among others, in IMA-ADPCM compressed Windows .WAV files.
The NDS data consist of a 32bit header, followed by 4bit values (so each byte contains two values, the first value in the lower 4bits, the second in upper 4 bits). The 32bit header contains initial values:

  Bit0-15   Initial PCM16 Value (Pcm16bit = -7FFFh..+7FFF) (not -8000h)
  Bit16-22  Initial Table Index Value (Index = 0..88)
  Bit23-31  Not used (zero)

In theory, the 4bit values are decoded into PCM16 values, as such:

  Diff = ((Data4bit AND 7)*2+1)*AdpcmTable[Index]/8      ;see rounding-error
  IF (Data4bit AND 8)=0 THEN Pcm16bit = Max(Pcm16bit+Diff,+7FFFh)
  IF (Data4bit AND 8)=8 THEN Pcm16bit = Min(Pcm16bit-Diff,-7FFFh)
  Index = MinMax (Index+IndexTable[Data4bit AND 7],0,88)

In practice, the first line works like so (with rounding-error):

  Diff = AdpcmTable[Index]/8
  IF (data4bit AND 1) THEN Diff = Diff + AdpcmTable[Index]/4
  IF (data4bit AND 2) THEN Diff = Diff + AdpcmTable[Index]/2
  IF (data4bit AND 4) THEN Diff = Diff + AdpcmTable[Index]/1

And, a note on the second/third lines (with clipping-error):

  Max(+7FFFh) leaves -8000h unclipped (can happen if initial PCM16 was -8000h)
  Min(-7FFFh) clips -8000h to -7FFFh (possibly unlike windows .WAV files?)

Whereas, IndexTable[0..7] = -1,-1,-1,-1,2,4,6,8. And AdpcmTable [0..88] =

  0007h,0008h,0009h,000Ah,000Bh,000Ch,000Dh,000Eh,0010h,0011h,0013h,0015h
  0017h,0019h,001Ch,001Fh,0022h,0025h,0029h,002Dh,0032h,0037h,003Ch,0042h
  0049h,0050h,0058h,0061h,006Bh,0076h,0082h,008Fh,009Dh,00ADh,00BEh,00D1h
  00E6h,00FDh,0117h,0133h,0151h,0173h,0198h,01C1h,01EEh,0220h,0256h,0292h
  02D4h,031Ch,036Ch,03C3h,0424h,048Eh,0502h,0583h,0610h,06ABh,0756h,0812h
  08E0h,09C3h,0ABDh,0BD0h,0CFFh,0E4Ch,0FBAh,114Ch,1307h,14EEh,1706h,1954h
  1BDCh,1EA5h,21B6h,2515h,28CAh,2CDFh,315Bh,364Bh,3BB9h,41B2h,4844h,4F7Eh
  5771h,602Fh,69CEh,7462h,7FFFh

The closest way to reproduce the AdpcmTable with 32bit integer maths appears:

  X=000776d2h, FOR I=0 TO 88, Table[I]=X SHR 16, X=X+(X/10), NEXT I
  Table[3]=000Ah, Table[4]=000Bh, Table[88]=7FFFh, Table[89..127]=0000h

When using ADPCM and loops, set the loopstart position to the data part, rather than the header. At the loop end, the SAD value is reloaded to the loop start location, additionally index and pcm16 values are reloaded to the values that have originally appeared at that location. Do not change the ADPCM loop start position during playback.

Microphone Input
For Microphone (and Touchscreen) inputs, see
DS Touch Screen Controller (TSC)

DS Files - Sound (SDAT etc.)

Sound data is often stored in a SDAT file (with SSEQ, SSAR, SBNK, SWAR, STRM blocks inside of the SDAT files). Samples can be stored in SWAV files (or be contained in SWAR's inside of SDAT files).
DS Sound Files - SDAT (Sound Data Archive)
DS Sound Files - SSEQ (Sound Sequence)
DS Sound Files - SSAR (Sound Sequence Archive)
DS Sound Files - SBNK (Sound Bank)
DS Sound Files - SWAR (Sound Wave Archive)
DS Sound Files - SWAV (Sound Wave Data)
DS Sound Files - STRM (Sound Wave Stream)
Thanks: http://www.feshrine.net/hacking/doc/nds-sdat.php - 2007 by kiwi.ds

DS Sound Files - SDAT (Sound Data Archive)

SDAT Header

  000h 4     ID "SDAT"           ;alike "CSAR" on 3DS
  004h 2     Byte Order    (FEFFh)
  006h 2     Version       (0100h)
  008h 4     Total Filesize
  00Ch 2     Header Size (usually 40h)
  00Eh 2     Number of Blocks (usually 4 = SYMB+INFO+FAT+FILE) (or 3=no SYMB)
  010h 4+4   SYMB Block (Offset from SDAT+0, Size) ;=0,0 if above is 3=no SYMB
  018h 4+4   INFO Block (Offset from SDAT+0, Size) ;\
  020h 4+4   FAT  Block (Offset from SDAT+0, Size) ; always present
  028h 4+4   FILE Block (Offset from SDAT+0, Size) ;/
  030h 10h   Padding to 20h-byte boundary (0)

The SYMB block exists in most SDAT files (except in some titles like Downhill Jam and Over the Hedge).

 _________________________________ SYMB Block _________________________________

Symbol Block (if present) (names for the corresponding items in INFO block)

  000h 4     ID "SYMB"
  004h 4     SYMB Block Size (rounded up to 4-byte boundary, unlike as in SDAT)
  008h 4     File List SSEQ    (Offset from SYMB+0) Sequences (songs)
  00Ch 4     Folder List SSAR  (Offset from SYMB+0) Sequence Archives (fx)
  010h 4     File List BANK    (Offset from SYMB+0) Banks
  014h 4     File List SWAR    (Offset from SYMB+0) Wave Archives (samples)
  018h 4     File List Player  (Offset from SYMB+0) Player (Group-related)
  01Ch 4     File List Group   (Offset from SYMB+0) Group (SSEQ+SSAR+BANK+SWAR)
  020h 4     File List Player2 (Offset from SYMB+0) Player2 (Stream-related)
  024h 4     File List STRM    (Offset from SYMB+0) Wave Stream
  028h 18h   Reserved (0)
  040h ..    File/Folder Lists (see below)
  ..   ..    File/Folder Name Strings (ASCII, terminated by 0)
  ..   ..    Padding to 4-byte boundary (0)

File List's are having following format:

  000h 4       Number of entries in this list (can be 0=None)
  004h N*4     File Name (Offset from SYMB+0)

Folder List (for SSAR):

  000h 4       Number of entries in this list (can be 0=None)
  004h N*(4+4) SSAR "Folder Name" and SSEQ "File List" (Offset's from SYMB+0)

 _________________________________ INFO Block _________________________________

Info Block

  000h 4     ID "INFO"
  004h 4     INFO Block Size (same as in SDAT header)
  008h 4     Info List SSEQ    (Offset from INFO+0) Sequences (songs)
  00Ch 4     Info List SSAR    (Offset from INFO+0) Sequence Archives (fx)
  010h 4     Info List BANK    (Offset from INFO+0) Banks
  014h 4     Info List SWAR    (Offset from INFO+0) Wave Archives (samples)
  018h 4     Info List Player  (Offset from INFO+0) Player (Group-related)
  01Ch 4     Info List Group   (Offset from INFO+0) Group (SSEQ+SSAR+BANK+SWAR)
  020h 4     Info List Player2 (Offset from INFO+0) Player2 (Stream-related)
  024h 4     Info List STRM    (Offset from INFO+0) Wave Stream
  028h 18h   Reserved (0)
  ..   ..    Info Lists (see below)
  ..   ..    Info Entries (see below)
  ..   ..    Padding to 4-byte boundary (0)

Info List's are having following format:

  000h 4     Number of entries in this list (can be 0=None)
  004h N*4   Info Entries (Offset from INFO+0)

SSEQ Info Entry (Sequences, eg. songs)

  000h 2   FAT fileID of SSEQ file      ;for accessing this file
  002h 2   Unknown
  004h 2   bnk         ;Associated BANK
  006h 1   vol         ;Volume
  007h 1   cpr
  008h 1   ppr
  009h 1   ply
  00Ah 2   Unknown (0)

SSAR Info Entry (Sequence Archives, eg. collections of sound effects)

  000h 2   FAT fileID of SSAR file
  002h 2   unknown

Note: bnk/vol/cpr/ppr/ply is stored in the SSAR file (instead of in Info). The actual sequences are also in SSAR? Or does the SSAR contain pointers to SSEQ files?

BANK Info Entry (Banks)

  000h 2   FAT fileID of SBNK file
  002h 2   unknown
  004h 2   1st SWAR   ;\
  006h 2   2nd SWAR   ; Associated Wave Archives (FFFFh=Unused entry)
  008h 2   3rd SWAR   ;
  00Ah 2   4th SWAR   ;/

SWAR Info Entry (Wave Archives, eg. collections of instruments or effects)

  000h 2   FAT fileID of SWAR file
  002h 2   unknown

Player Info Entry (whatever)

  000h 1   Unknown
  001h 3   Padding
  004h 4   Unknown

Group Info Entry (whatever, list of files that must be loaded to memory?)

  000h 4       Number of items in this group
  004h N*(4+4) Array (with ID+Index pairs)

ID values: 700h=SSEQ, 803h=SSAR, 601h=BANK, 402h=SWAR.
Index: Entry number in the corresponding SSEQ/SSAR/BANK/SWAR list.

Player2 Info Entry (whatever, Stream related?)

  000h 1   nCount             ;number of USED entries in below array
  001h 16  v[16]              ;unknown array (UNUSED entries are set to FFh
  011h 7   Reserved (0)

STRM Info Entry

  000h 2   FAT fileID of STRM file    ;for accessing the file
  002h 2   Unknown
  004h 1   vol            ;volume
  005h 1   pri            ;priority?
  006h 1   ply            ;play?
  007h 5   Reserved (0)

 ____________________________ FAT and FILE Blocks _____________________________

FAT Block

  000h 4         ID "FAT "
  004h 4         FAT Block Size (same as in SDAT header) (0Ch+N*10h)
  008h 4         Number of files
  00Ch N*(4+4+8) File Entries (Offset from SDAT+0, Size, Zero)

The 8-byte Zero entries can be used for storing data at runtime.

File Block

  000h 4         ID "FILE"
  004h 4         FILE Block Size (same as in SDAT header)
  008h 4         Number of files (same as in FAT block)
  00Ch 4         Reserved (0)
  010h ..        Files (SSEQ,SSAR,SBNK,SWAR,STRM) (at offsets specified in FAT)

DS Sound Files - SSEQ (Sound Sequence)

It is a converted MIDI sequence. Linked to a BANK for instruments.
SSEQ is usually found inside of SDAT files (but also exists as standalone file, eg. 3DS Circle Pad Pro test/calib, RomFS:\extrapad_bcwav_LZ.bin\*.sseq, and 3DS Picture Picker, RomFS:\sound\csnd.LZ\*).

SSEQ Header

  000h 4     ID "SSEQ"                                  ;\
  004h 2     Byte Order    (FEFFh)                      ;
  006h 2     Version       (0100h)                      ; Main header
  008h 4     Total Filesize                             ;
  00Ch 2     Header Size (usually 10h)                  ;
  00Eh 2     Number of Blocks (usually 1 = DATA)        ;/
  010h 4     ID "DATA"                                  ;\
  014h 4     Total Filesize, minus 10h                  ; Sub header
  018h 4     Offset to data (from SSEQ+0) (1Ch)         ;/
  01Ch ..    Arrays of sequence data..                  ;-

NB. For the details of the SSEQ file, please refer to loveemu's sseq2mid

Description
A SSEQ can have at maximum 16 tracks, notes in the range of 0..127 (middle C is 60). Each quartet note has a fixed tick length of 48. Tempo in the range of 1..240 BPM (default is 120). The SSEQ will not be played correctly if tempo higher than 240.

The SEQ player uses ARM7's Timer1 for timing. The ARM7's 4 Timers runs at 33MHz (approximately 2^25). The SEQ player sets Timer1 reload value to 2728, prescaler to F/64. So on about every 0.0052 sec (64*2728/33MHz) the SEQ Player will be notified (1 cycle). As a quartet note has fixed tick value of 48, the highest tempo that SEQ Player can handle is 240 BPM (60/(0.0052*48)).

During each cycle, the SEQ player adds the tempo value to a variable. Then it checks if the value exceeds 240. If it does, the SEQ player subtracts 240 from the variable, and process the SSEQ file. Using this method, the playback is not very precise but the difference is too small to be noticed.

Take an example with tempo = 160 BPM, the SSEQ file is processed twice in 3 notifications.

  cycle  variable action
  1       0       Add 160
  2       160     Add 160
  3       320     Subtract 240, process once, add 160
  4       240     Subtract 240, process once, add 160
  5       160     Add 160
  6       320     Subtract 240, process once, add 160
  7       240     Subtract 240, process once, add 160
  8       160     Add 160

Events

  ID      Parameter             Description
  00h-7Fh Velocity: 1 byte [0..127]
          Duration: Variable Length
                                NOTE-ON. Duration is expressed in tick.
                                48 for quartet note.
                                Usually it is NOT a multiple of 3.
  80h     Duration: Variable Length
                                REST. It tells the SSEQ-sequencer to wait for
                                a certain tick. Usually it is a multiple of 3.
  81h     Bank & Program Number:
          Variable Length
                                bits[0..7] is the program number,
                                bits[8..14] is the bank number.
                                Bank change is seldomly found,
                                so usually bank 0 is used.
  FEh     2 bytes Indicates which tracks are used.
          Bit0 for track 0, ... Bit15 for track 15.
          If the bit is set, the corresponding track is used.
                                Indication begin of multitrack. Must be in the
                                beginning of the first track to work. A series
                                of event 0x93 follows.
  93h     4 bytes 1st byte is track number [0..15]
          The other 3 bytes are the relative adress of track data.
          Add nDataOffset (usually 0x1C) to find out the absolute address.
                                SSEQ is similar to MIDI in that track data are
                                stored one after one track. Unlike mod music.
  94h     JUMP Address: 3 bytes
          (Add nDataOffset (usually 0x1C) to find out the absolute address.)
                                JUMP. A jump must be backward. So that the
                                song will loop forever.
  95h     CALL Address: 3 bytes
          (Add nDataOffset (usually 0x1C) to find out the absolute address.)
  A0h-BFh See loveemu's sseq2mid for more details.
                                Some arithmetic operations / comparions.
                                Affect how SSEQ is to be played.
  C0h     1 byte  PAN                (0..127, middle is 64, uh?)
  C1h     1 byte  VOLUME             (0..127)
  C2h     1 byte  MASTER VOLUME      (0..127)
  C3h     1 byte  TRANSPOSE (Channel Coarse Tuning) (0..64 = 64..128 in MIDI)
  C4h     1 byte  PITCH BEND
  C5h     1 byte  PITCH BEND RANGE
  C6h     1 byte  TRACK PRIORITY
  C7h     1 byte  MONO/POLY          (0=Poly, 1=Mono)
  C8h     1 byte  TIE (unknown)      (0=Off, 1=On)
  C9h     1 byte  PORTAMENTO CONTROL
  CAh     1 byte  MODULATION DEPTH   (0=Off, 1=On)
  CBh     1 byte  MODULATION SPEED
  CCh     1 byte  MODULATION TYPE    (0=Pitch, 1=Volume, 2=Pan)
  CDh     1 byte  MODULATION RANGE
  CEh     1 byte  PORTAMENTO ON/OFF
  CFh     1 byte  PORTAMENTO TIME
  D0h     1 byte  ATTACK RATE
  D1h     1 byte  DECAY RATE
  D2h     1 byte  SUSTAIN RATE
  D3h     1 byte  RELEASE RATE
  D4h     1 byte  LOOP START         (how many times to be looped)
  D5h     1 byte  EXPRESSION
  D6h     1 byte  PRINT VARIABLE     (unknown)
  E0h     2 byte  MODULATION DELAY
  E1h     2 byte  TEMPO
  E3h     2 byte  SWEEP PITCH
  FCh     -       LOOP END           (for LOOP START)
  FDh     -       RETURN from CALL command
  FFh     -       EOT: End Of Track

DS Sound Files - SSAR (Sound Sequence Archive)

It is a collection of SSEQ sequences (mainly for relative short sound effects) (longer sequences like music are usually stored in separate SSEQ files instead of in SSAR archives).

SSAR Header

  000h 4     ID "SSAR"                                  ;\
  004h 2     Byte Order    (FEFFh)                      ;
  006h 2     Version       (0100h)                      ; Main header
  008h 4     Total Filesize                             ;
  00Ch 2     Header Size (usually 10h)                  ;
  00Eh 2     Number of Blocks (usually 1 = DATA)        ;/
  010h 4     ID "DATA"                                  ;\
  014h 4     Total Filesize, minus 10h                  ;
  018h 4     Offset to data (from SSAR+0) (20h+N*0Ch)   ; Sub header
  01Ch 4     Number of records                          ;
  020h N*0Ch Records (12 bytes each)                    ;/
  ..   ..    data... unknown content? alike SSEQ?       ;-

SSAR Records

  000h 4   nOffset    ;relative offset of the archived SEQ file,
                           absolute offset = nOffset + SSAR::nDataOffset
  004h 2   bnk        ;bank
  006h 1   vol        ;volume
  007h 1   cpr        ;channel pressure
  008h 1   ppr        ;polyphonic pressure
  009h 1   ply        ;play
  00Ah 2   reserved (0)

SSAR Data (addressed via relative offset in above 0Ch-byte record)

  data... unknown content? alike SSEQ?

NB. Archived SSEQ files are not stored in sequence (order). So Rec[0].nOffset may point to 0x100 but Rec[1].nOffset points to 0x40.

NB. Archived SSEQ files cannot be readily extracted from SSAR file because data in one SSEQ may 'call' data in other SSEQ.

DS Sound Files - SBNK (Sound Bank)

This seems to assign ADSR patterns to each note of each SWAV instrument...?

"A bank is linked to up to 4 SWAR files which contain the samples. It defines the instruments which a SSEQ sequence can use. You may imagine SSEQ + SBNK + SWAR are similar to module music created by trackers." uh?

SBNK Header

  000h 4   ID "SBNK"                                    ;\
  004h 2   Byte Order    (FEFFh)                        ;
  006h 2   Version       (0100h)                        ; Main header
  008h 4   Total Filesize                               ;
  00Ch 2   Header Size (usually 10h)                    ;
  00Eh 2   Number of Blocks (usually 1 = DATA)          ;/
  010h 4   ID "DATA"                                    ;\
  014h 4   Total Filesize, minus 10h                    ;
  018h 20h Reserved (0) (for use at runtime)            ; Sub header
  038h 4   Number of Instruments (SWAV's)               ;
  03Ch N*4 Instrument Records (1+2+1 bytes per instr.)  ;/
  ...  ..  Instrument Data (depending of above records) ;-

Instrument Record

  000h 1   fRecord     ;can be either 0, 1..4, 16 or 17
  001h 2   nOffset     ;absolute offset of the data in file  ;uh, misaligned?
  003h 1   Reserved (0)

fRecord = 0, it is empty (unknown purpose, dummy? or maybe disable ADSR?)
nOffset will also = 0.

fRecord < 16, the record is a note/wave definition
"I have seen values 1, 2 and 3. But it seems the value does not affect the wave/note definition that follows. Instrument record size is 10 bytes."

  00h 10    SWAV, SWAR, Note, Attack, Decay, Sustain, Release, Pan

fRecord = 16, the record is a range of note/wave definitions

  00h      1   Lower note (0..127)          ;eg. 10 ;\notes 10..20
  01h      1   Upper note (0..127)          ;eg. 20 ;/
  02h+N*12 2   Unknown (usually 0001h)
  04h+N*12 10  SWAV, SWAR, Note, Attack, Decay, Sustain, Release, Pan

fRecord = 17, the record is a regional wave/note definition

  00h      1   End of 1st region (0..127)   ;eg. 25  = notes 0..25
  01h      1   End of 2nd region (0..127)   ;eg. 35  = notes 26..35
  02h      1   End of 3rd region (0..127)   ;eg. 45  = notes 36..45
  03h      1   End of 4th region (0..127)   ;eg. 55  = notes 46..55
  04h      1   End of 5th region (0..127)   ;eg. 65  = notes 56..65
  05h      1   End of 6th region (0..127)   ;eg. 127 = notes 66..last
  06h      1   End of 7th region (0..127)   ;eg. 0   = none
  07h      1   End of 8th region (0..127)   ;eg. 0   = none
  08h+N*12 2   Unknown (usually 0001h)
  08h+N*12 10  SWAV, SWAR, Note, Attack, Decay, Sustain, Release, Pan

REMARKS: Unknown bytes before wave/defnition definition = 5, not 1 in stage_04_bank.sbnk, stage_04.sdat, Rom No.1156, uh?

SWAV, SWAR, Note, Attack, Decay, Sustain, Release, Pan

  00h 2  SWAV Number   the swav used
  02h 2  SWAR Mumber   the swar used (see Info Block --> "BANK Info Entry")
  04h 1  Note Number   (0..127)
  05h 1  Attack Rate   (0..127, 127=fast)
  06h 1  Decay Rate    (0..127, 127=fast)
  07h 1  Sustain Level (0..127, 127=stay at max, no decay)
  08h 1  Release Rate  (0..127, 127=fast)
  09h 1  Pan           (0..127, 64=middle) (uh, what=left, what=right?)

Articulation Data

             .                   <-- max level (127)
            / \
           /   \
          /     '---------.      <-- sustain level (0..127)
         /                 \
        /                   \
  -----'---------------------'-- <-- min level (0)
      Attack Decay Sustain Release

  "The SEQ Player treats 0 as the 100% amplitude value and -92544 (723*128)
  as the 0% amplitude value. The starting ampltitude is 0% (-92544)." uh?

  "During the attack phase, in each cycle, the SSEQ Player calculates the new
  amplitude value: amplitude value = attack rate * amplitude value / 255. The
  attack phase stops when amplitude reaches 0." THAT IS... NON-LINEAR attack?

  "During the decay phase, in each cycle, the SSEQ Player calculates the new
  amplitude value: amplitude value = amplitude value - decay rate.
  Note the starting amplitude value is 0. The decay phase stops when
  amplitude reaches sustain level." THAT IS... LINEAR decay/release?

DS Sound Files - SWAR (Sound Wave Archive)

It is a collection of mono wave (SWAV) samples only (which can be in either PCM8, PCM16 or ADPCM compression).

SWAR Header

  000h 4     ID "SWAR"                                  ;\
  004h 2     Byte Order    (FEFFh)                      ;
  006h 2     Version       (0100h)                      ; Main header
  008h 4     Total Filesize (including SWAV's)          ;
  00Ch 2     Header Size (usually 10h)                  ;
  00Eh 2     Number of Blocks (usually 1 = DATA)        ;/
  010h 4     ID "DATA"                                  ;\
  014h 4     Total Filesize, minus 10h                  ;
  018h 20h   Reserved (0) (for use at runtime)          ; Sub header
  038h 4     Number of SWAV sample blocks               ;
  03Ch N*4   Offsets to Sample blocks (from SWAR+0)     ;/
  ..   ..    Sample blocks... starting with Type (0=PCM8, 1=PCM16, 2=IMA-ADPCM)

When extracting single sample block, one can convert them to SWAV files (by inserting an 18h-byte SWAV header).
The sample blocks are usually (always?) stored at increasing offsets (so one can determine the size by computing the distance to next offset; or to filesize for last entry) (alternately, the size can be computed by looking at the Sound Length entry of sample block).

DS Sound Files - SWAV (Sound Wave Data)

SWAV Header (present only in standalone SWAV files) (not in SWAR archives)

  000h 4     ID "SWAV"                                  ;\
  004h 2     Byte Order    (FEFFh)                      ;
  006h 2     Version       (0100h)                      ; Main header
  008h 4     Total Filesize                             ;
  00Ch 2     Header Size (usually 10h)                  ;
  00Eh 2     Number of Blocks (usually 1 = DATA)        ;/
  010h 4     ID "DATA"                                  ;\Sub header
  014h 4     Total Filesize, minus 10h                  ;/
  018h ..    Sample block (see below)

Note: System Flaw has a lot of SWAV files (instead of using SWAR archives).

Sample Block Format (present in SWAV files and SWAR archives)

  000h 1    WaveType (0=PCM8, 1=PCM16, 2=IMA-ADPCM)
  001h 1    Loop flag = TRUE|FALSE   ;uh?
  002h 2    Sampling Rate
  004h 2    Time (ARM7_CLOCK / nSampleRate)
                [ARM7_CLOCK: 33.513982MHz/2 = 1.6756991 E +7]
  006h 2    Loop Offset, in 4-byte units
  008h 4    Sound Length, in 4-byte units (exluding ADPCM header, if any)
  00Ch ...  Data... (samples) (with 32bit header in case of ADPCM)

DS Sound Files - STRM (Sound Wave Stream)

It is an individual mono/stereo wave file (PCM8, PCM16 or ADPCM) (eg. used in Ultimate Spiderman rom:\sound\sound_stream.sdat).

STRM Header

  000h 4     ID "STRM"                                          ;\
  004h 2     Byte Order    (FEFFh)                              ;
  006h 2     Version       (0100h)                              ; Main header
  008h 4     Total Filesize                                     ;
  00Ch 2     Header Size (usually 10h)                          ;
  00Eh 2     Number of Blocks (usually 2 = HEAD+DATA)           ;/
  010h 4     ID "HEAD"                                          ;\
  014h 4     Size of HEAD structure (uh, this is... 50h?)       ;
  018h 1     Type       (0=PCM8, 1=PCM16, 2=IMA-ADPCM)          ; Sub header
  019h 1     Loop flag  (?=TRUE|FALSE)  ;uh?                    ;
  01Ah 1     Channels   (?=What)        ;mono/stereo?           ;
  01Bh 1     Unknown    (always 0)                              ;
  01Ch 2     Sampling Rate (perhaps resampled from original)    ;
  01Eh 2     Time (1.0 / rate * ARM7_CLOCK / 32)                ;
               [ARM7_CLOCK: 33.513982MHz/2 = 1.6756991e7]       ;
  020h 4     Loop Offset (samples)                              ;
  024h 4     Number of Samples                                  ;
  028h 4     Wave Data Offset (always 68h)                      ;
  02Ch 4     Number of Blocks       (per what?)                 ;
  030h 4     Block Length           (per Channel)               ;
  034h 4     Samples Per Block      (per Channel)               ;
  038h 4     Last Block Length      (per Channel)               ;
  03Ch 4     Samples Per Last Block (per Channel)               ;
  040h 20h   Reserved (always 0)                                ;/
  060h 4     ID "DATA"                                          ;\Data header
  064h 4     Data Size (8+N ?)                                  ;/
  068h N     Wave Data blocks...                                ;-Sample data

Mono blocks are ordered: Block1, Block2, Block3, etc.
Stereo blocks are ordered: LeftBlock1, RightBlock1, LeftBlock2, etc.

DS System and Built-in Peripherals

DS DMA Transfers
DS Timers
DS Interrupts
DS Maths
DS Inter Process Communication (IPC)
DS Keypad
DS Absent Link Port
DS Real-Time Clock (RTC)
DS Serial Peripheral Interface Bus (SPI)
DS Touch Screen Controller (TSC)
DS Power Management Device
DS Power Control
DS Backwards-compatible GBA-Mode
DS Debug Registers (Emulator/Devkits)

DS DMA Transfers

The DS includes four DMA channels for each CPU (ie. eight channels in total), which are working more or less the same as on GBA:
GBA DMA Transfers
All NDS9 and NDS7 DMA Registers are R/W. The gamepak bit (Bit 27) has been removed (on the NDS9 the bit is used to expand the mode setting to 3bits).

NDS9 DMA
Word count of all channels is expanded to 21bits (max 1..1FFFFFh units, or 0=200000h units), and SAD/DAD registers for all channels support ranges of 0..0FFFFFFEh. The transfer modes (DMACNT Bit27-29) are:

  0  Start Immediately
  1  Start at V-Blank
  2  Start at H-Blank (paused during V-Blank)
  3  Synchronize to start of display
  4  Main memory display
  5  DS Cartridge Slot
  6  GBA Cartridge Slot
  7  Geometry Command FIFO

NDS7 DMA
Word Count, SAD, and DAD are R/W, aside from that they do have the same restrictions as on GBA (max 4000h or 10000h units, some addresses limited to 0..07FFFFFEh). DMACNT Bit27 is unused on NDS7. The transfer modes (DMACNT Bit28-29) are:

  0  Start Immediately
  1  Start at V-Blank
  2  DS Cartridge Slot
  3  DMA0/DMA2: Wireless interrupt, DMA1/DMA3: GBA Cartridge Slot

40000E0h - NDS9 only - DMA0FILL - DMA 0 Filldata (R/W)
40000E4h - NDS9 only - DMA1FILL - DMA 1 Filldata (R/W)
40000E8h - NDS9 only - DMA2FILL - DMA 2 Filldata (R/W)
40000ECh - NDS9 only - DMA3FILL - DMA 3 Filldata (R/W)

  Bit0-31 Filldata

The DMA Filldata registers contain 16 bytes of general purpose WRAM, intended to be used as fixed source addresses for DMA memfill operations.
This is useful because DMA cannot read from TCM, and reading from Main RAM would require to recurse cache & write buffer.
The DMA Filldata is used with Src=Fixed and SAD=40000Exh (which isn't optimal because it's doing repeated reads from SAD, and, for that reason, a memfill via STMIA opcodes can be faster than DMA; the DSi's new NDMA channels are providing a faster fill method with Src=Fill and SAD=Unused).

NDS7 Sound DMA
The NDS additionally includes 16 Sound DMA channels, plus 2 Sound Capture DMA channels (see Sound chapter). The priority of these channels is unknown.

NDS9 Cache, Writebuffer, DTCM, and ITCM
Cache and tightly coupled memory are connected directly to the NDS9 CPU, without using the system bus. So that, DMA cannot access DTCM/ITCM, and access to cached memory regions must be handled with care: Drain the writebuffer before DMA-reads, and invalidate the cache after DMA-writes. See,
ARM CP15 System Control Coprocessor
The CPU can be kept running during DMA, provided that it is accessing only TCM (or cached memory), otherwise the CPU is halted until DMA finishes.
Respectively, interrupts executed during DMA will usually halt the CPU (unless the IRQ handler uses only TCM and cache; the IRQ vector at FFFF00xxh must be cached, or relocated to ITCM at 000000xxh, and the IRQ handler may not access IE, IF, or other I/O ports).

NDS Sequential Main Memory DMA
Main RAM has different access time for sequential and non-sequential access. Normally DMA uses sequential access (except for the first word), however, if the source and destination addresses are both in Main RAM, then all accesses become non-sequential. In that case it would be faster to use two DMA transfers, one from Main RAM to a scratch buffer in WRAM, and one from WRAM to Main RAM.

DS Timers

Same as GBA, except F = 33.513982 MHz (for both NDS9 and NDS7).
GBA Timers
Both NDS9 and NDS7 have four Timers each, eight Timers in total.
The NDS sound controller is having its own frequency generators (unlike GBA, which needed to use Timers to drive channel A/B sounds).

DS Interrupts

4000208h - NDS9/NDS7 - IME - Interrupt Master Enable (R/W)

  0     Disable all interrupts  (0=Disable All, 1=See IE register)
  1-31  Not used

4000210h - NDS9/NDS7 - IE - 32bit - Interrupt Enable (R/W)
4000214h - NDS9/NDS7 - IF - 32bit - Interrupt Request Flags (R/W)
Bits in the IE register are 0=Disable, 1=Enable.
Reading IF returns 0=No request, 1=Interrupt Request.
Writing IF acts as 0=No change, 1=Acknowledge (clears that bit).

  0     LCD V-Blank
  1     LCD H-Blank
  2     LCD V-Counter Match
  3     Timer 0 Overflow
  4     Timer 1 Overflow
  5     Timer 2 Overflow
  6     Timer 3 Overflow
  7     NDS7 only: SIO/RCNT/RTC (Real Time Clock)
  8     DMA 0
  9     DMA 1
  10    DMA 2
  11    DMA 3
  12    Keypad
  13    GBA-Slot (external IRQ source) / DSi: None such
  14    Not used                       / DSi9: NDS-Slot Card change?
  15    Not used                       / DSi: dito for 2nd NDS-Slot?
  16    IPC Sync
  17    IPC Send FIFO Empty
  18    IPC Recv FIFO Not Empty
  19    NDS-Slot Game Card Data Transfer Completion
  20    NDS-Slot Game Card IREQ_MC
  21    NDS9 only: Geometry Command FIFO
  22    NDS7 only: Screens unfolding
  23    NDS7 only: SPI bus
  24    NDS7 only: Wifi    / DSi9: XpertTeak DSP
  25    Not used           / DSi9: Camera
  26    Not used           / DSi9: Undoc, IF.26 set on FFh-filling 40021Axh
  27    Not used           / DSi:  Maybe IREQ_MC for 2nd gamecard?
  28    Not used           / DSi: NewDMA0
  29    Not used           / DSi: NewDMA1
  30    Not used           / DSi: NewDMA2
  31    Not used           / DSi: NewDMA3
  ?     DSi7: any further new IRQs on ARM7 side... in bit13-15,21,25-26?

Raw TCM-only IRQs can be processed even during DMA ?
Trying to set all IE bits gives FFFFFFFFh (DSi7) or FFFFFF7Fh (DSi9).

4000218h - DSi7 - IE2 - DSi7 Extra Interrupt Enable Bits
400021Ch - DSi7 - IF2 - DSi7 Extra Interrupt Flags

  0     DSi7: GPIO18[0]   ;\
  1     DSi7: GPIO18[1]   ; maybe 1.8V signals?
  2     DSi7: GPIO18[2]   ;/
  3     DSi7: Unused (0)
  4     DSi7: GPIO33[0] unknown (related to "GPIO330" testpoint on mainboard?)
  5     DSi7: GPIO33[1] Headphone connect (HP#SP) (static state)
  6     DSi7: GPIO33[2] Powerbutton interrupt (short pulse upon key-down)
  7     DSi7: GPIO33[3] sound enable output (ie. not a useful irq-input)
  8     DSi7: SD/MMC Controller   ;-Onboard eMMC and External SD Slot
  9     DSi7: SD Slot Data1 pin   ;-For SDIO hardware in External SD Slot
  10    DSi7: SDIO Controller     ;\Atheros Wifi Unit
  11    DSi7: SDIO Data1 pin      ;/
  12    DSi7: AES interrupt
  13    DSi7: I2C interrupt
  14    DSi7: Microphone Extended interrupt
  15-31 DSi7: Unused (0)

Trying to set all IE2 bits gives 00007FF7h (DSi7) or 00000000h (DSi9).

DTCM+3FFCh - NDS9 - IRQ Handler (hardcoded DTCM address)
380FFFCh - NDS7 - IRQ Handler (hardcoded RAM address)

  Bit 0-31  Pointer to IRQ Handler

NDS7 Handler must use ARM code, NDS9 Handler can be ARM/THUMB (Bit0=Thumb).

DTCM+3FF8h - NDS9 - IRQ Check Bits (hardcoded DTCM address)
380FFF8h - NDS7 - IRQ Check Bits (hardcoded RAM address)

  Bit 0-31  IRQ Flags (same format as IE/IF registers)

When processing & acknowleding interrupts via IF register, the user interrupt handler should also set the corresponding bits of the IRQ Check value (required for BIOS IntrWait and VBlankIntrWait SWI functions).

380FFC0h - DSi7 only - Extra IRQ Check Bits for IE2/IF2 (hardcoded RAM addr)
Same as the above 380FFF8h value, but for new IE2/IF2 registers, intended for use with IntrWait and VBlankIntrWait functions. However, that functions are BUGGED on DSi and won't actually work in practice (they do support only the new 380FFC0h bits, but do accidently ignore the old 380FFF8h bits).

--- Below for other (non-IRQ) exceptions ---

27FFD9Ch - RAM - NDS9 Debug Stacktop / Debug Vector (0=None)
380FFDCh - RAM - NDS7 Debug Stacktop / Debug Vector (0=None)
These addresses contain a 32bit pointer to the Debug Handler, and, memory below of the addresses is used as Debug Stack. The debug handler is called on undefined instruction exceptions, on data/prefetch aborts (caused by the protection unit), on FIQ (possibly caused by hardware debuggers). It is also called by accidental software-jumps to the reset vector, and by unused SWI numbers within range 0..1Fh.

DS Maths

4000280h - NDS9 - DIVCNT - Division Control (R/W)

  0-1   Division Mode    (0-2=See below) (3=Reserved; same as Mode 1)
  2-13  Not used
  14    Division by zero (0=Okay, 1=Division by zero error; 64bit Denom=0)
  15    Busy             (0=Ready, 1=Busy) (Execution time see below)
  16-31 Not used

Division Modes and Busy Execution Times

  Mode  Numer / Denom = Result, Remainder ; Cycles
  0     32bit / 32bit = 32bit , 32bit     ; 18 clks
  1     64bit / 32bit = 64bit , 32bit     ; 34 clks
  2     64bit / 64bit = 64bit , 64bit     ; 34 clks

Division is started when writing to any of the DIVCNT/NUMER/DENOM registers.

4000290h - NDS9 - DIV_NUMER - 64bit Division Numerator (R/W)
4000298h - NDS9 - DIV_DENOM - 64bit Division Denominator (R/W)
Signed 64bit values (or signed 32bit values in 32bit modes, the upper 32bits are then unused, with one exception: the DIV0 flag in DIVCNT is set only if the full 64bit DIV_DENOM value is zero, even in 32bit mode).

40002A0h - NDS9 - DIV_RESULT - 64bit Division Quotient (=Numer/Denom) (R)
40002A8h - NDS9 - DIVREM_RESULT - 64bit Remainder (=Numer MOD Denom) (R)
Signed 64bit values (in 32bit modes, the values are sign-expanded to 64bit).

Division Overflows
Overflows occur on "DIV0" and "-MAX/-1" (eg. -80000000h/-1 in 32bit mode):

  DIV0     -->  REMAIN=NUMER, RESULT=+/-1 (with sign opposite of NUMER)
  -MAX/-1  -->  RESULT=-MAX               (instead +MAX)

On overflows in 32bit/32bit=32bit mode: the upper 32bit of the sign-expanded 32bit result are inverted. This feature produces a correct 64bit (+MAX) result in case of the incorrect 32bit (-MAX) result. The feature also applies on DIV0 errors (which makes the sign-expanded 64bit result even more messed-up than the normal 32bit result).
The DIV0 flag in DIVCNT.14 indicates DENOM=0 errors (it does not indicate "-MAX/-1" errors). The DENOM=0 check relies on the full 64bit value (so, in 32bit mode, the flag works only if the unused upper 32bit of DENOM are zero).

40002B0h - NDS9 - SQRTCNT - Square Root Control (R/W)

  0     Mode (0=32bit input, 1=64bit input)
  1-14  Not used
  15    Busy (0=Ready, 1=Busy) (Execution time is 13 clks, in either Mode)
  16-31 Not used

Calculation is started when writing to any of the SQRTCNT/PARAM registers.

40002B4h - NDS9 - SQRT_RESULT - 32bit - Square Root Result (R)
40002B8h - NDS9 - SQRT_PARAM - 64bit - Square Root Parameter Input (R/W)
Unsigned 64bit parameter, and unsigned 32bit result.

IRQ Notes
Push all DIV/SQRT values (parameters and control registers) when using DIV/SQRT registers on interrupt level, and, after restoring them, be sure to wait until the busy flag goes off, before leaving the IRQ handler.

BIOS Notes
The NDS9 and NDS7 BIOSes additionally contain software based division and square root functions, which are NOT using above hardware registers (even the NDS9 functions are raw software).

Timing Notes
The Div/Sqrt timings are counted in 33.51MHz units. Although the calculations are quite fast, mind that reading/writing the result/parameter registers takes up additional clock cycles (especially due to the PENALTY cycle glitch for non-sequential accesses; parts of that problem can be eventually bypassed by using sequential STMIA/LDMIA opcodes) (nethertheless, in some cases, software may be actually faster than the hardware registers; eg. for small 8bit numbers; that of course NOT by using the BIOS software functions which are endless inefficient).

DS Inter Process Communication (IPC)

Allows to exchange status information between ARM7 and ARM9 CPUs.
The register can be accessed simultaneously by both CPUs (without violating access permissions, and without generating waitstates at either side).

4000180h - NDS9/NDS7 - IPCSYNC - IPC Synchronize Register (R/W)

  Bit   Dir  Expl.
  0-3   R    Data input from IPCSYNC Bit8-11 of remote CPU (00h..0Fh)
  4-7   -    Not used
  8-11  R/W  Data output to IPCSYNC Bit0-3 of remote CPU   (00h..0Fh)
  12    -    Not used
  13    W    Send IRQ to remote CPU      (0=None, 1=Send IRQ)
  14    R/W  Enable IRQ from remote CPU  (0=Disable, 1=Enable)
  15-31 -    Not used

4000184h - NDS9/NDS7 - IPCFIFOCNT - IPC Fifo Control Register (R/W)

  Bit   Dir  Expl.
  0     R    Send Fifo Empty Status      (0=Not Empty, 1=Empty)
  1     R    Send Fifo Full Status       (0=Not Full, 1=Full)
  2     R/W  Send Fifo Empty IRQ         (0=Disable, 1=Enable)
  3     W    Send Fifo Clear             (0=Nothing, 1=Flush Send Fifo)
  4-7   -    Not used
  8     R    Receive Fifo Empty          (0=Not Empty, 1=Empty)
  9     R    Receive Fifo Full           (0=Not Full, 1=Full)
  10    R/W  Receive Fifo Not Empty IRQ  (0=Disable, 1=Enable)
  11-13 -    Not used
  14    R/W  Error, Read Empty/Send Full (0=No Error, 1=Error/Acknowledge)
  15    R/W  Enable Send/Receive Fifo    (0=Disable, 1=Enable)
  16-31 -    Not used

4000188h - NDS9/NDS7 - IPCFIFOSEND - IPC Send Fifo (W)

  Bit0-31  Send Fifo Data (max 16 words; 64bytes)

4100000h - NDS9/NDS7 - IPCFIFORECV - IPC Receive Fifo (R)

  Bit0-31  Receive Fifo Data (max 16 words; 64bytes)

IPCFIFO Notes
When IPCFIFOCNT.15 is disabled: Writes to IPCFIFOSEND are ignored (no data is stored in the FIFO, the error bit doesn't get set though), and reads from IPCFIFORECV return the oldest FIFO word (as usually) (but without removing the word from the FIFO).
When the Receive FIFO is empty: Reading from IPCFIFORECV returns the most recently received word (if any), or ZERO (if there was no data, or if the FIFO was cleared via IPCFIFOCNT.3), and, in either case the error bit gets set.
The Fifo-IRQs are edge triggered, IF.17 gets set when the condition "(IPCFIFOCNT.2 AND IPCFIFOCNT.0)" changes from 0-to-1, and IF.18 gets set when "(IPCFIFOCNT.10 AND NOT IPCFIFOCNT.8)" changes from 0-to-1. The IRQ flags can be acknowledged even while that conditions are true.

DS Keypad

For the GBA-buttons: Same as GBA, both ARM7 and ARM9 have keyboard input registers, and each its own keypad IRQ control register.
GBA Keypad Input

For Touchscreen (and Microphone) inputs, see
DS Touch Screen Controller (TSC)

4000136h - NDS7 - EXTKEYIN - Key X/Y Input (R)

  0      Button X     (0=Pressed, 1=Released)
  1      Button Y     (0=Pressed, 1=Released)
  3      DEBUG button (0=Pressed, 1=Released/None such)
  6      Pen down     (0=Pressed, 1=Released/Disabled) (always 0 in DSi mode)
  7      Hinge/folded (0=Open, 1=Closed)
  2,4,5  Unknown / set
  8..15  Unknown / zero

The Hinge stuff is a magnetic sensor somewhere underneath of the Start/Select buttons (NDS) or between A/B/X/Y buttons (DSi), it will be triggered by the magnet field from the right speaker when the console is closed. The hinge generates an interrupt request (there seems to be no way to disable this, unlike as for all other IRQ sources), however, the interrupt execution can be disabled in IE register (as for other IRQ sources).
The Pen Down is the /PENIRQ signal from the Touch Screen Controller (TSC), if it is enabled in the TSC control register, then it will notify the program when the screen pressed, the program should then read data from the TSC (if there's no /PENIRQ then doing unneccassary TSC reads would just waste CPU power). However, the user may release the screen before the program performs the TSC read, so treat the screen as not pressed if you get invalid TSC values (even if /PENIRQ was LOW).
Not sure if the TSC /PENIRQ is actually triggering an IRQ in the NDS?
The Debug Button should be connected to R03 and GND (on original NDS, R03 is the large soldering point between the SL1 jumper and the VR1 potentiometer) (there is no R03 signal visible on the NDS-Lite board).
Interrupts are reportedly not supported for X,Y buttons.

DS Absent Link Port

The DS doesn't have a Serial Link Port Socket, however, internally, the NDS7 contains the complete set of Serial I/O Ports, as contained in the GBA:
GBA Communication Ports

In GBA mode, the ports are working as on real GBA (as when no cable is connected). In NDS mode, the ports are even containing some additional bits:

NDS7 SIO Bits (according to an early I/O map from Nintendo)

  NDS7 4000128h SIOCNT   Bit15 "CKUP"  New Bit in NORMAL/MULTI/UART mode (R/W)
  NDS7 4000128h SIOCNT   Bit14 "N/A"   Removed IRQ Bit in UART mode (?)
  NDS7 400012Ah SIOCNT_H Bit14 "TFEMP" New Bit (R/W)
  NDS7 400012Ah SIOCNT_H Bit15 "RFFUL" New Bit (always zero?)
  NDS7 400012Ch SIOSEL   Bit0  "SEL"   New Bit (always zero?)
  NDS7 4000140h JOYCNT   Bit7  "MOD"   New Bit (R/W)

The "CKUP" bit duplicates the internal clock transfer rate (selected in SIOCNT.1) (tested in normal mode) (probably works also in multi/uart mode?).

NDS7 DS-Lite 4001080h (W) (?)
DS-Lite Firmware writes FFFFh to this address (prior to accessing SIOCNT), so it's probably SIO or debugging related (might be as well a bug or so). Reading from the port always returns 0000h on both DS and DS-Lite.

NDS9 SIO Bits (according to an early I/O map from Nintendo)

  NDS9 4000120h SIODATA32 Bit0-31 Data            (always zero?)
  NDS9 4000128h SIOCNT    Bit2    "TRECV" New Bit (always zero?)
  NDS9 4000128h SIOCNT    Bit3    "TSEND" New Bit (always zero?)
  NDS9 400012Ch SIOSEL    Bit0    "SEL"   New Bit (always zero?)

Not sure if these ports really exist in the release-version, or if it's been prototype stuff?

RCNT
RCNT (4000134h) should be set to 80xxh (general purpose mode) before accessing EXTKEYIN (4000136h) or RTC (4000138h). No idea why (except when using RTC/SI-interrupt).

DS Serial Port
The SI line is labeled "INT" on the NDS mainboard, it is connected to Pin 1 of the RTC chip (ie. the /INT interrupt pin).
I have no idea where to find SO, SC, and SD. I've written a test proggy that pulsed all four RCNT bits - but all I could find was the SI signal. However, the BIOS contains some code that uses SIO normal mode transfers (for the debug version), so at least SI, SO, SC should exist...?
MAYBE that three signals are somehow replaced by EXTKEYIN bit0,1,3?

DS Real-Time Clock (RTC)

NDS
Seiko Instruments Inc. S-35180 (compatible with S-35190A)
Miniature 8pin RTC with 3-wire serial bus

DSi
Seiko S-35199A01 (12pin BGA, with some extra functions like FOUT and Alarm Date)

4000138h - NDS7 - Real Time Clock Register

  Bit  Expl.
  0    Data I/O   (0=Low, 1=High)
  1    Clock Out  (0=Low, 1=High)
  2    Select Out (0=Low, 1=High/Select)
  4    Data  Direction  (0=Read, 1=Write)
  5    Clock Direction  (should be 1=Write)
  6    Select Direction (should be 1=Write)
  3,8-11   Unused I/O Lines
  7,12-15  Direction for Bit3,8-11 (usually 0)
  16-31    Not used

Serial Transfer Flowchart
Chipselect and Command/Parameter Sequence:

  Init CS=LOW and /SCK=HIGH, and wait at least 1us
  Switch CS=HIGH, and wait at least 1us
  Send the Command byte (see bit-transfer below)
  Send/receive Parameter byte(s) associated with the command (see below)
  Switch CS to LOW

Bit transfer (repeat 8 times per cmd/param byte) (bits transferred LSB first):

  Output /SCK=LOW and SIO=databit (when writing), then wait at least 5us
  Output /SCK=HIGH, wait at least 5us, then read SIO=databit (when reading)
  In either direction, data is output on (or immediately after) falling edge.

Ideally, <both> commands and parameters should be transmitted LSB-first (unlike the original Seiko document, which recommends LSB-first for data, and MSB-first for commands) (actually, later Seiko datasheets are going so far to recommend MSB-first for everything, eg. to use bit-reversed Data=C8h for Year=13h).

Command Register

  Command Register
    Fwd  Rev
    0    7   Fixed Code (must be 0)
    1    6   Fixed Code (must be 1)
    2    5   Fixed Code (must be 1)
    3    4   Fixed Code (must be 0, or, DSi only: 1=Extended Command)
    4-6  3-1 Command
             Fwd Rev Parameter bytes (read/write access)
             0   0   1 byte, status register 1
             4   1   1 byte, status register 2
             2   2   7 bytes, date & time (year,month,day,day_of_week,hh,mm,ss)
             6   3   3 bytes, time (hh,mm,ss)
             1*  4*  1 byte, int1, frequency duty setting
             1*  4*  3 bytes, int1, alarm time 1 (day_of_week, hour, minute)
             5   5   3 bytes, int2, alarm time 2 (day_of_week, hour, minute)
             3   6   1 byte, clock adjustment register
             7   7   1 byte, free register
             Extended command (when above "fourth bit" was set, DSi only)
             Fwd Rev Parameter bytes (read/write access)
             0   0   3 byte, up counter (msw,mid,lsw) (read only)
             4   1   1 byte, FOUT register setting 1
             2   2   1 byte, FOUT register setting 2
             6   3   reserved
             1   4   3 bytes, alarm date 1 (year,month,day)
             5   5   3 bytes, alarm date 2 (year,month,day)
             3   6   reserved
             7   7   reserved
    7    0   Parameter Read/Write Access (0=Write, 1=Read)

* INT1: Type and number of parameters depend on INT1 setting in stat reg2.
The "Fwd" bit numbers and command values for LSB-first command transfers (ie. both commands and parameters use the same bit-order).
The "Rev" numbers/values are for MSB-first command transfers (ie. commands using opposite bit-order than parameters, as being suggested by Seiko).

Control and Status Registers

  Status Register 1
    0   W   Reset                (0=Normal, 1=Reset)
    1   R/W 12/24 hour mode      (0=12 hour, 1=24 hour)
    2-3 R/W General purpose bits
    4   R   Interrupt 1 Flag (1=Yes)                      ;auto-cleared on read
    5   R   Interrupt 2 Flag (1=Yes)                      ;auto-cleared on read
    6   R   Power Low Flag (0=Normal, 1=Power is/was low) ;auto-cleared on read
    7   R   Power Off Flag (0=Normal, 1=Power was off)    ;auto-cleared on read
    Power off indicates that the battery was removed or fully discharged,
    all registers are reset to 00h (or 01h), and must be re-initialized.
  Status Register 2
    0-3 R/W INT1 Mode/Enable
            0000b Disable
            0x01b Selected Frequency steady interrupt
            0x10b Per-minute edge interrupt
            0011b Per-minute steady interrupt 1 (duty 30.0 seconds)
            0100b Alarm 1 interrupt
            0111b Per-minute steady interrupt 2 (duty 0.0079 seconds)
            1xxxb 32kHz output
    4-5 R/W General purpose bits
    6   R/W INT2 Enable
            0b    Disable
            1b    Alarm 2 interrupt
    7   R/W Test Mode (0=Normal, 1=Test, don't use) (cleared on Reset)
  Clock Adjustment Register (to compensate oscillator inaccuracy)
    0-7 R/W Adjustment (00h=Normal, no adjustment)
  Free Register
    0-7 R/W General purpose bits

Date Registers

  Year Register
    0-7 R/W Year     (BCD 00h..99h = 2000..2099)
  Month Register
    0-4 R/W Month    (BCD 01h..12h = January..December)
    5-7 -   Not used (always zero)
  Day Register
    0-5 R/W Day      (BCD 01h..28h,29h,30h,31h, range depending on month/year)
    6-7 -   Not used (always zero)
  Day of Week Register (septenary counter)
    0-2 R/W Day of Week (00h..06h, custom assignment, usually 0=Monday?)
    3-7 -   Not used (always zero)

Time Registers

  Hour Register
    0-5 R/W Hour     (BCD 00h..23h in 24h mode, or 00h..11h in 12h mode)
    6   *   AM/PM    (0=AM before noon, 1=PM after noon)
            * 24h mode: AM/PM flag is read only (PM=1 if hour = 12h..23h)
            * 12h mode: AM/PM flag is read/write-able
            * 12h mode: Observe that 12 o'clock is defined as 00h (not 12h)
    7   -   Not used (always zero)
  Minute Register
    0-6 R/W Minute   (BCD 00h..59h)
    7   -   Not used (always zero)
  Second Register
    0-6 R/W Minute   (BCD 00h..59h)
    7   -   Not used (always zero)

Alarm 1 and Alarm 2 Registers

  Alarm1 and Alarm2 Day of Week Registers (INT1 and INT2 each)
    0-2 R/W Day of Week (00h..06h)
    3-6 -   Not used (always zero)
    7   R/W Compare Enable (0=Alarm every day, 1=Alarm only at specified day)
  Alarm1 and Alarm2 Hour Registers (INT1 and INT2 each)
    0-5 R/W Hour     (BCD 00h..23h in 24h mode, or 00h..11h in 12h mode)
    6   R/W AM/PM    (0=AM, 1=PM) (must be correct even in 24h mode?)
    7   R/W Compare Enable (0=Alarm every hour, 1=Alarm only at specified hour)
  Alarm1 and Alarm2 Minute Registers (INT1 and INT2 each)
    0-6 R/W Minute   (BCD 00h..59h)
    7   R/W Compare Enable (0=Alarm every min, 1=Alarm only at specified min)
  Selected Frequency Steady Interrupt Register (INT1 only) (when Stat2/Bit2=0)
    0   R/W Enable 1Hz Frequency  (0=Disable, 1=Enable)
    1   R/W Enable 2Hz Frequency  (0=Disable, 1=Enable)
    2   R/W Enable 4Hz Frequency  (0=Disable, 1=Enable)
    3   R/W Enable 8Hz Frequency  (0=Disable, 1=Enable)
    4   R/W Enable 16Hz Frequency (0=Disable, 1=Enable)
            The signals are ANDed when two or more frequencies are enabled,
            ie. the /INT signal gets LOW when either of the signals is LOW.
    5-7 R/W General purpose bits

Note: There is only one register shared as "Selected Frequency Steady Interrupt" (accessed as single byte parameter when Stat2/Bit2=0) and as "Alarm1 Minute" (accessed as 3rd byte of 3-byte parameter when Stat2/Bit2=1), changing either value will also change the other value.

Up Counter (DSi only)

  Up Counter Msw
    0-7 R   Up Counter bit16-23 (non-BCD, 00h..FFh)
  Up Counter Mid
    0-7 R   Up Counter bit8-15  (non-BCD, 00h..FFh)
  Up Counter Lsw
    0-7 R   Up Counter bit0-7   (non-BCD, 00h..FFh)

The 24bit Up Counter is incremented when seconds=00h (that is, once per minute; unless the Time is getting changed by write commands, which may cause some stuttering). The Up Counter starts at 000000h upon power-up, and, if the battery lasts that long: wraps from FFFFFFh to 000000h after about 30 years.

Alarm 1 and Alarm 2 Date Registers (DSi only)

  Alarm 1 and Alarm 2 Year Register
    0-7 R/W Year     (BCD 00h..99h = 2000..2099)
  Alarm 1 and Alarm 2 Month Register
    0-4 R/W Month    (BCD 01h..12h = January..December)
    5   -   Not used (always zero)
    6   R/W Year Compare Enable (0=Ignore, 1=Enable)
    7   R/W Month Compare Enable (0=Ignore, 1=Enable)
  Alarm 1 and Alarm 2 Day Register
    0-5 R/W Day      (BCD 01h..28h,29h,30h,31h, range depending on month/year)
    6   -   Not used (always zero)
    7   R/W Day Compare Enable (0=Ignore, 1=Enable)

XXX unspecified if above Alarm Date stuff is really R/W (or write only)

FOUT Register (DSi only)

  FOUT Register Setting 1
    0-7 R/W  Enable bits (bit0=256Hz, bit1=512Hz, ..., bit7=32768Hz)
  FOUT Register Setting 2
    0-7 R/W  Enable bits (bit0=1Hz,   bit1=2Hz, ...,   bit7=128Hz)
  The above sixteen FOUT signals are ANDed when two or more frequencies are
  enabled, ie. the FOUT signal gets LOW when either of the signals is LOW.

Note: The FOUT pin goes to the DSi's wifi daughterboard: FOUT must be set to 32768Hz (FOUT1=80h, FOUT2=00h), that setting is required for exchanging Atheros WMI commands/events, the firmware does take care of initializing that setting (eg. needed after battery removal, or after overwriting the FOUT registers by software).

Interrupt
There's only one /INT signal, shared for both INT1 and INT2.
In the NDS, it is connected to the SI-input of the SIO unit (and so, also shared with SIO interrupts). To enable the interrupt, RCNT should be set to 8144h (Bit14-15=General Purpose mode, Bit8=SI Interrupt Enable, Bit6,2=SI Output/High).
The Output/High settings seems to be used as pullup (giving faster reactions on low-to-high transitions) (nethertheless, in most cases it seems to be also working okay as Input, ie. with RCNT=8100h).
The RCNT interrupt is generated on high-to-low transitions on the SI line (but only if the IRQ is enabled in RCNT.8, and only if RCNT is set to general purpose mode) (note: changing RCNT.8 from off-to-on does NOT generate IRQs, even when SI is LOW).

Pin-Outs

  1 /INT      8 VDD
  2 XOUT      7 SIO
  3 XIN       6 /SCK
  4 GND       5 CS

DS Serial Peripheral Interface Bus (SPI)

Serial Peripheral Interface Bus
SPI Bus is a 4-wire (Data In, Data Out, Clock, and Chipselect) serial bus.
The NDS supports the following SPI devices (each with its own chipselect).
DS Firmware Serial Flash Memory
DS Touch Screen Controller (TSC)
DS Power Management Device

40001C0h - NDS7 - SPICNT - SPI Bus Control/Status Register

  0-1   Baudrate (0=4MHz/Firmware, 1=2MHz/Touchscr, 2=1MHz/Powerman., 3=512KHz)
  2     DSi: Baudrate MSB   (4=8MHz, 5..7=None/0Hz) (when SCFG_EXT7.bit9=1)
  2     NDS: Not used       (Zero)
  3-6   Not used            (Zero)
  7     Busy Flag           (0=Ready, 1=Busy) (presumably Read-only)
  8-9   Device Select       (0=Powerman., 1=Firmware, 2=Touchscr, 3=Reserved)
  10    Transfer Size       (0=8bit/Normal, 1=16bit/Bugged)
  11    Chipselect Hold     (0=Deselect after transfer, 1=Keep selected)
  12-13 Not used            (Zero)
  14    Interrupt Request   (0=Disable, 1=Enable)
  15    SPI Bus Enable      (0=Disable, 1=Enable)

The "Hold" flag should be cleared BEFORE transferring the LAST data unit, the chipselect will be then automatically cleared after the transfer, the program should issue a WaitByLoop(3) manually AFTER the LAST transfer.

40001C2h - NDS7 - SPIDATA - SPI Bus Data/Strobe Register (R/W)
The SPI transfer is started on writing to this register, so one must <write> a dummy value (should be zero) even when intending to <read> from SPI bus.

  0-7   Data
  8-15  Not used (always zero, even in bugged-16bit mode)

During transfer, the Busy flag in SPICNT is set, and the written SPIDATA value is transferred to the device (via output line), simultaneously data is received (via input line). Upon transfer completion, the Busy flag goes off (with optional IRQ), and the received value can be then read from SPIDATA, if desired.

Notes/Glitches
SPICNT Bits 12,13 appear to be unused (always zero), although the BIOS (attempts to) set Bit13=1, and Bit12=Bit11 when accessing the firmware.
The SPIDATA register is restricted to 8bit, so that only each 2nd byte will appear in SPIDATA when attempting to use the bugged-16bit mode.

Cartridge Backup Auxiliar SPI Bus
The NDS Cartridge Slot uses a separate SPI bus (with other I/O Ports), see
DS Cartridge Backup

DS Touch Screen Controller (TSC)

Texas Instruments TSC2046 (NDS)
Asahi Kasei Microsystems AK4148AVT (NDS-Lite)
The Touch Screen Controller (for lower LCD screen) is accessed via SPI bus,
DS Serial Peripheral Interface Bus (SPI)

Control Byte (transferred MSB first)

  0-1  Power Down Mode Select
  2    Reference Select (0=Differential, 1=Single-Ended)
  3    Conversion Mode  (0=12bit, max CLK=2MHz, 1=8bit, max CLK=3MHz)
  4-6  Channel Select   (0-7, see below)
  7    Start Bit (Must be set to access Control Byte)

Channel

  0 Temperature 0 (requires calibration, step 2.1mV per 1'C accuracy)
  1 Touchscreen Y-Position  (somewhat 0B0h..F20h, or FFFh=released)
  2 Battery Voltage         (not used, connected to GND in NDS, always 000h)
  3 Touchscreen Z1-Position (diagonal position for pressure measurement)
  4 Touchscreen Z2-Position (diagonal position for pressure measurement)
  5 Touchscreen X-Position  (somewhat 100h..ED0h, or 000h=released)
  6 AUX Input               (connected to Microphone in the NDS)
  7 Temperature 1 (difference to Temp 0, without calibration, 2'C accuracy)

All channels can be accessed in Single-Ended mode.
In differential mode, only channel 1,3,4,5 (X,Z1,Z2,Y) can be accessed.
On AK4148AVT, channel 6 (AUX) is split into two separate channels, IN1 and IN2, separated by Bit2 (Reference Select). IN1 is selected when Bit2=1, IN2 is selected when Bit2=0 (despite of the Bit2 settings, both IN1 and IN2 are using single ended more). On the NDS-Lite, IN1 connects to the mircrophone (as on original NDS), and the new IN2 input is simply wired to VDD3.3 (which is equal to the external VREF voltage, so IN2 is always FFFh).

Power Down Mode

  Mode /PENIRQ   VREF  ADC   Recommended use
  0    Enabled   Auto  Auto  Differential Mode (Touchscreen, Penirq)
  1    Disabled  Off   On    Single-Ended Mode (Temperature, Microphone)
  2    Enabled   On    Off   Don't use
  3    Disabled  On    On    Don't use

Allows to enable/disable the /PENIRQ output, the internal reference voltage (VREF), and the Analogue-Digital Converter.
For AK4148AVT, Power Down modes are slightly different (among others, /PENIRQ is enabled in Mode 0..2).

Reference Voltage (VREF)
VREF is used as reference voltage in single ended mode, at 12bit resolution one ADC step equals to VREF/4096. The TSC generates an internal VREF of 2.5V (+/-0.05V), however, the NDS uses as external VREF of 3.33V (sinks to 3.31V at low battery charge), the external VREF is always enabled, no matter if internal VREF is on or off. Power Down Mode 1 disables the internal VREF, which may reduce power consumption in single ended mode. After conversion, Power Down Mode 0 should be restored to re-enable the Penirq signal.

Sending the first Command after Chip-Select
Switch chipselect low, then output the command byte (MSB first).

Reply Data
The following reply data is received (via Input line) after the Command byte has been transferred: One dummy bit (zero), followed by the 8bit or 12bit conversion result (MSB first), followed by endless padding (zero).
Note: The returned ADC value may become unreliable if there are longer delays between sending the command, and receiving the reply byte(s).

Sending further Commands during/after receiving Reply Data
In general, the Output line should be LOW during the reply period, however, once when Data bit6 has been received (or anytime later), a new Command can be invoked (started by sending the HIGH-startbit, ie. Command bit7), simultaneously, the remaining reply-data bits (bit5..0) can be received.
In other words, the new command can be output after receiving 3 bits in 8bit mode (the dummy bit, and data bits 7..6), or after receiving 7 bits in 12bit mode (the dummy bit, and data bits 11..6).
In practice, the NDS SPI register always transfers 8 bits at once, so that one would usually receive 8 bits (rather than above 3 or 7 bits), before outputting a new command.

Touchscreen Position
Read the X and Y positions in 12bit differential mode, then convert the touchscreen values (adc) to screen/pixel positions (scr), as such:

  scr.x = (adc.x-adc.x1) * (scr.x2-scr.x1) / (adc.x2-adc.x1) + (scr.x1-1)
  scr.y = (adc.y-adc.y1) * (scr.y2-scr.y1) / (adc.y2-adc.y1) + (scr.y1-1)

The X1,Y1 and X2,Y2 calibration points are found in Firmware User Settings,
DS Firmware User Settings
scr.x1,y1,x2,y2 are originated at 1,1 (converted to 0,0 by above formula).

Touchscreen Pressure (not supported on DSi)
To calculate the pressure resistance, in respect to X/Y/Z positions and X/Y plate resistances, either of below formulas can be used,

  Rtouch = (Rx_plate*Xpos*(Z2pos/Z1pos-1))/4096
  Rtouch = (Rx_plate*Xpos*(4096/Z1pos-1)-Ry_plate*(1-Ypos))/4096

The second formula requires less CPU load (as it doesn't require to measure Z2), the downside is that one must know both X and Y plate resistance (or at least their ratio). The first formula doesn't require that ratio, and so Rx_plate can be set to any value, setting it to 4096 results in

  touchval = Xpos*(Z2pos/Z1pos-1)

Of course, in that case, touchval is just a number, not a resistance in Ohms.

Touchscreen Notes
It may be impossible to press locations close to the screen borders.
When pressing two or more locations the TSC values will be somewhere in the middle of these locations.
The TSC values may be garbage if the screen becomes newly pressed or released, to avoid invalid inputs: read TSC values at least two times, and ignore BOTH positions if ONE position was invalid.

Microphone / AUX Channel
Observe that the microphone amplifier is switched off after power up, see:
DS Power Management Device
DS Power Control

Temperature Calculation (not supported on DSi)
TP0 decreases by circa 2.1mV per degree Kelvin. The voltage difference between TP1 minus TP0 increases by circa 0.39mV (1/2573 V) per degree Kelvin. At VREF=3.33V, one 12bit ADC step equals to circa 0.8mV (VREF/4096).
Temperature can be calculated at best resolution when using the current TP0 value, and two calibration values (an ADC value, and the corresponding temperature in degrees kelvin):

  K = (CAL.TP0-ADC.TP0) * 0.4 + CAL.KELVIN

Alternately, temperature can be calculated at rather bad resolution, but without calibration, by using the difference between TP1 and TP0:

  K = (ADC.TP1-ADC.TP0) * 8568 / 4096

To convert Kelvin to other formats,

  Celsius:     C = (K-273.15)
  Fahrenheit:  F = (K-273.15)*9/5+32
  Reaumur:     R = (K-273.15)*4/5
  Rankine:     X = (K)*9/5

The Temperature Range for the TSC 2046 chip is -40'C..+85'C (for AK4181AVT only -20'C..+70'C). According to Nintendo, the DS should not be exposed to "extreme" heat or cold, the optimal battery charging temperature is specified as +10'C..+40'C.
The original firmware does not support temperature calibration, calibration is supported by nocash firmware (if present). See Extended Settings,
DS Firmware Extended Settings

Pin-Outs

         ________
  VCC  1|o       |16 DCLK
  X+   2|        |15 /CS
  Y+   3|  TSC   |14 DIN
  X-   4|  2046  |13 BUSY
  Y-   5|        |12 DOUT
  GND  6|        |11 /PENIRQ
  VBAT 7|        |10 IOVDD
  AUX  8|________|9  VREF

For AK4181AVT, same pins as above, except that IOVDD replaced by the new IN2 input, the pin is wired to VDD3.3 (so IN2 is always equal to VREF, which is wired to VDD3.3, too) (and AUX is renamed to IN1, and is kept used for MIC input).

DSi Touchscreen Controller (in NDS mode)
DSi in NDS mode does support only X, Y, and MIC (all other channels do return FFFh in 12bit mode, and FFh in 8bit mode, ie. no pressure, no temperature, and no GNDed battery sensor). On DSi, MIC does return data in both single-ended and differential mode (unlike as on real NDS).

DSi Touchscreen Controller (in DSi mode)
The DSi touchscreen controller supports a NDS backwards compatibility mode. But, in DSi mode, it is working entirely different (it's still accessed via SPI bus, but with some new MODE/INDEX values).
DSi Touchscreen/Sound Controller
The NDS Touchscreen controller did additionally allow to read Temperature and Touchscreen Pressure - unknown if the DSi is also supporting such stuff (via whatever DSi-specific registers).
The touchscreen hardware can be switched to NDS compatibility mode (for older games), but unknown how to do that.

DS Power Control

The DS contains several Power Managment functions, some accessed via I/O ports (described below), and some accessed via SPI bus:
DS Power Management Device

4000304h - NDS9 - POWCNT1 - Graphics Power Control Register (R/W)

  0     Enable Flag for both LCDs (0=Disable) (Prohibited, see notes)
  1     2D Graphics Engine A      (0=Disable) (Ports 008h-05Fh, Pal 5000000h)
  2     3D Rendering Engine       (0=Disable) (Ports 320h-3FFh)
  3     3D Geometry Engine        (0=Disable) (Ports 400h-6FFh)
  4-8   Not used
  9     2D Graphics Engine B      (0=Disable) (Ports 1008h-105Fh, Pal 5000400h)
  10-14 Not used
  15    Display Swap (0=Send Display A to Lower Screen, 1=To Upper Screen)
  16-31 Not used

Use SwapBuffers command once after enabling Rendering/Geometry Engine.
Improper use of Bit0 may damage the hardware?
When disabled, corresponding Ports become Read-only, corresponding (palette-) memory becomes read-only-zero-filled.

4000304h - NDS7 - POWCNT2 - Sound/Wifi Power Control Register (R/W)

  Bit   Expl.
  0     Sound Speakers (0=Disable, 1=Enable) (Initial setting = 1)
  1     Wifi           (0=Disable, 1=Enable) (Initial setting = 0)
  2-31  Not used

Note: Bit0 disables the internal Speaker only, headphones are not disabled.
Bit1 disables Port 4000206h, and Ports 4800000h-480FFFFh.

4000206h - NDS7 - WIFIWAITCNT - Wifi Waitstate Control

  Bit   Expl.
  0-1   WS0 nonsequential time (0-3 = 10, 8, 6, 18 cycles) ;\4800000h-4807FFFh
  2     WS0 sequential time    (0-1 = 6, 4 cycles)         ;/   (used for RAM)
  3-4   WS1 nonsequential time (0-3 = 10, 8, 6, 18 cycles) ;\4808000h-480FFFFh
  5     WS1 sequential time    (0-1 = 10, 4 cycles)        ;/   (used for I/O)
  6-15  Not used (zero)

This register is initialized by firmware on power-up (set to 0030h, other software shouldn't change that setting). The timings are per 16bit halfword access.
Note: WIFIWAITCNT can be accessed only when enabled in POWCNT2.

4000301h - NDS7 - HALTCNT - Low Power Mode Control (R/W)
In Halt mode, the CPU is paused as long as (IE AND IF)=0.
In Sleep mode, most of the hardware including sound and video are paused, this very-low-power mode could be used much like a screensaver.

  Bit   Expl.
  0-5   Not used (zero)
  6-7   Power Down Mode  (0=No function, 1=Enter GBA Mode, 2=Halt, 3=Sleep)

The HALTCNT register should not be accessed directly. Instead, the BIOS Halt, Sleep, CustomHalt, IntrWait, or VBlankIntrWait SWI functions should be used.
BIOS Halt Functions
ARM CP15 System Control Coprocessor
The NDS9 does not have a HALTCNT register, instead, the Halt function uses the co-processor opcode "mcr p15,0,r0,c7,c0,4" - this opcode locks up if interrupts are disabled via IME=0 (unlike NDS7 HALTCNT method which doesn't check IME).

4000300h - NDS7/NDS9 - POSTFLG - BYTE - Post Boot Flag (R/W)
The NDS7 and NDS9 post boot flags are usually set upon BIOS/Firmware boot completion, once when set the reset vector is redirected to the debug handler of Nintendo's hardware debugger. That allows the NDS7 debugger to capture accidental jumps to address 0, that appears to be a common problem with HLL-programmers, asm-coders know that (and why) they should not jump to 0.

  Bit   Expl.
  0     Post Boot Flag (0=Boot in progress, 1=Boot completed)
  1     NDS7: Not used (always zero), NDS9: Bit1 is read-writeable
  2-7   Not used (always zero)

There are some write-restrictions: The NDS7 register can be written to only from code executed in BIOS (done by NDS boot ROM, or by DSi firmware, whereas the DSi firmware is using the CpuSet SWI function to issue the POSTFLG write from within ROM). Bit0 of both NDS7 and NDS9 registers cannot be cleared (except by Reset) once when it is set. DSi games seem to run regardless of POSTFLG, whilst NDS games somewhat refuse to run when POSTFLG=0.

Memory Power Down Functions
DS Main Memory Control
DS Firmware Serial Flash Memory

DS Power Management Device

Power Management Device - Mitsumi 3152A (NDS) / Mitsumi 3205B (NDS-LITE)
The Power Management Device is accessed via SPI bus,
DS Serial Peripheral Interface Bus (SPI)
To access the device, write the Index Register, then read or write the data register, and release the chipselect line when finished.

  Index Register
  Bit0-6 Register Select          (0..3) (0..4 for DS-Lite) (0..7Fh for DSi)
  Bit7   Register Direction       (0=Write, 1=Read)
  Register 0 - Powermanagement Control (R/W)
  Bit0   Sound Amplifier Enable   (0=Disable, 1=Enable)
         (Old-DS:  Disabled: Sound is very silent, but still audible)
         (DS-Lite: Disabled: Sound is NOT audible)
         (DSi in NDS Mode: R/W, but effect is unknown yet)
         (DSi in DSi Mode: Not used, Bit0 is always 1)
  Bit1   Sound Amplifier Mute     (0=Normal, 1=Mute) (Old-DS Only, not DS-Lite)
         (Old-DS:  Muted: Sound is NOT audible, that works only if Bit0=1)
         (DS-Lite: Not used, Bit1 is always zero)
         (DSi in NDS Mode: R/W, but effect is unknown yet)
         (DSi in DSi Mode: R/W, but effect is unknown yet)
  Bit2   Lower Backlight          (0=Disable, 1=Enable)
  Bit3   Upper Backlight          (0=Disable, 1=Enable)
  Bit4   Power LED Blink Enable   (0=Always ON, 1=Blinking OFF/ON)
  Bit5   Power LED Blink Speed    (0=Slow, 1=Fast) (only if Blink enabled)
         (DSi: Power LED Blinking isn't supported, neither in NDS nor DSi mode)
  Bit6   DS System Power          (0=Normal, 1=Shut Down)
  Bit7   Not used                 (always 0)
  Register 1 - Battery Status (R)
  Bit0   Battery Power LED Status (0=Power Good/Green, 1=Power Low/Red)
         (DSi: Usually 0, not tested if it changes upon Power=Low)
  Bit1-7 Not used
  Register 2 - Microphone Amplifier Control (R/W)
  Bit0   Amplifier                (0=Disable, 1=Enable)
  Bit1-7 Not used                 (always 0)
  (DSi in NDS Mode: looks same as NDS, ie. only bit0 is R/W)
  (DSi in DSi Mode: Not used, always FFh)
  Register 3 - Microphone Amplifier Gain Control (R/W)
  Bit0-1 Gain                     (0..3=Gain 20, 40, 80, 160)
  Bit2-7 Not used                 (always 0)
  (DSi in NDS Mode: looks same as NDS, ie. only bit0-1 are R/W)
  (DSi in DSi Mode: Not used, always FFh)
  Register 4 - DS-Lite and DSi Only - Backlight Levels/Power Source (R/W)
  Bit0-1 Backlight Brightness (0..3=Low,Med,High,Max)   (R/W)
         (when bit2+3 are both set, then reading bit0-1 always returns 3)
  Bit2   Force Max Brightness when Bit3=1 (0=No, 1=Yes) (R/W)
  Bit3   External Power Present           (0=No, 1=Yes) (Read-Only)
  Bit4-7 Unknown (Always 4) (Read-Only)
  (DSi in NDS Mode: looks same as in DSi mode)
  (DSi in DSi Mode: Bit0-1 are R/W, but ignored, bit2-3 are always 0)
  Register 10h - DSi Only - Backlight Mirrors & Reset (R/W)
  Bit0   Reset (0=No, 1=Reboot DSi) (same/similar as BPTWL reset feature?)
  Bit1   Unknown (R/W) (note: whatever it is, it isn't warmboot flag)
  Bit2-3 Mirror of Register 0, bit2-3 (backlight enable bits) (R/W)
  Bit4-7 Not used (always 0)
  Bit5   Not used (always 0) - but DSi bootrom sets that bit on boot error?
  (This register works in NDS mode and DSi mode, though it's mainly intended
  for NDS mode, eg. DS Download Play uses the Reset bit to return to DSi menu)
  (note: writing bit2 seems to affect BOTH bit1 and bit2 in register 0)
  Register 1Fh and 20h - DSi Only (?)
  DSi bootrom sets register 1Fh and 20h bit0-4 to value 1Fh on boot error,
  unknown purpose, seems to have no effect, maybe prototype backlight level?

On Old-DS, registers 4..7Fh are mirrors of 0..3. On DS-Lite, registers 5,6,7 are mirrors of 4, register 8..7Fh are mirrors of 0-7.
On DSi (in DS mode), index 0,1,2,3,4,10h are used (reads as 0Fh,00h,00h,01h,41h,0Fh - regardless of backlight level, and power source), index 5..0Fh and 11h..7Fh return 00h (ie. unlike DS and DS-Lite, there are no mirrors; aside from the mirrored bits in register 10h).

Backlight Dimming / Backlight caused Shut-Down(s)
The above bits are essentially used to switch Backlights on or off. However, there a number of strange effects. Backlight dimming is possible by pulse width modulation, ie. by using a timer interrupt to issue pulse widths of N% ON, and 100-N% OFF. Too long pulses are certainly resulting in flickering. Too short pulses are ignored, the backlights will remain OFF, even if the ON and OFF pulses are having the same length. Much too short pulses cause the power supply to shut-down; after changing the backlight state, further changes must not occur within the next (circa) 2500 clock cycles. The mainboard can be operated without screens & backlights connected, however, if so, the power supply will shut-down as soon as backlights are enabled.
Pulse width modulated dimming does also work on the DS-Lite, allowing to use smoother fade in/out effects as when using the five "hardware" levels (Off,Low,Med,High,Max).

DS Main Memory Control

Main Memory
The DS Main Memory is 2Mx16bit (4MByte), 1.8V Pseudo SRAM (PSRAM); all Dynamic RAM refresh is handled internally, the chip doesn't require any external refresh signals, and alltogether behaves like Static RAM. Non-sequential access time is 70ns, sequential (burst) access time is 12ns.

Main Memory Control
The memory chips contain built-in Control functions, which can be accessed via Port 27FFFFEh and/or by EXMEMCNT Bit 14. Nintendo is using at least two different types of memory chips in DS consoles, Fujitsu 82DBS02163C-70L, and ST M69AB048BL70ZA8, both appear to have different control mechanisms, other chips (with 8MB size) are used in the semi-professional DS hardware debuggers, and further chips may be used in future, so using the memory control functions may lead into compatibitly problems.

Power Consumption / Power Control
Power Consumption during operation (read/write access) is somewhat 30mA, in standby mode (no read/write access) consumption is reduced to 100uA.
Furthermore, a number of power-down modes are supported: In "Deep" Power Down mode the refresh is fully disabled, consumption is 10uA (and all data will be lost), in "Partial" Power Down modes only fragment of memory is refreshed, for smallest fragments, consumption goes to down to circa 50uA. The chip cannot be accessed while it is in Deep or Partial Power Down mode.

Fujitsu 82DBS02163C-70L
The Configuration Register (CR) can be written to by the following sequence:

  LDRH R0,[27FFFFEh]      ;read one value
  STRH R0,[27FFFFEh]      ;write should be same value as above
  STRH R0,[27FFFFEh]      ;write should be same value as above
  STRH R0,[27FFFFEh]      ;write any value
  STRH R0,[27FFFFEh]      ;write any value
  LDRH R0,[2400000h+CR*2] ;read, address-bits are defining new CR value

Do not access any other Main Memory addresses during above sequence (ie. disable interrupts, and do not execute the sequence by code located in Main Memory). The CR value is write-only. The CR bits are:

  Bit    Expl.
  0-6    Reserved         (Must be 7Fh)
  7      Write Control
           0=WE Single Clock Pulse Control without Write Suspend Function
           1=WE Level Control with Write Suspend Function)
          Burst Read/Single Write is not supported at WE Single Clock Mode.
  8      Reserved         (Must be 1)
  9      Valid Clock Edge (0=Falling Edge, 1=Rising Edge)
  10     Single Write     (0=Burst Read/Burst Write, 1=Burst Read/Single Write)
  11     Burst Sequence   (0=Reserved, 1=Sequential)
  12-14  Read Latency     (1=3 clocks, 2=4 clocks, 3=5 clocks, other=Reserved)
  15     Mode
           0=Synchronous:  Burst Read, Burst Write
           1=Asynchronous: Page Read, Normal Write
          In Mode 1 (Async), only the Partial Size bits are used,
          all other bits, CR bits 0..18, must be "1".
  16-18  Burst Length     (2=8 Words, 3=16Words, 7=Continous, other=Reserved)
  19-20  Partial Size     (0=1MB, 1=512KB, 2=Reserved, 3=Deep/0 bytes)

The Power Down mode is entered by setting CE2=LOW, this can be probably done by setting EXMEMCNT Bit14 to zero.

ST Microelectronics M69AB048BL70ZA8
The chip name decodes as PSRAM (M96), Asynchronous (A), 1.8V Burst (B), 2Mx16 (048), Two Chip Enables (B), Low Leakage (L), 70ns (70), Package (ZA), -30..+85'C (8).
There are three data sheets for different PSRAM chips available at www.st.com (unfortunately none for M69AB048BL70ZA8), each using different memory control mechanisms.

NDS9 BIOS
The NDS9 BIOS contains the following Main Memory initialization code, that method doesn't match up with any ST (nor Fujitsu) data sheets that I've seen. At its best, it looks like a strange (and presumably non-functional) mix-up of different ST control methods.

  STRH 2000h,[4000204h]    ;EXMEMCNT, enable RAM, async mode
  LDRH R0,[27FFFFEh]
  STRH R0,[27FFFFEh]
  STRH R0,[27FFFFEh]
  STRH FFDFh,[27FFFFEh]
  STRH E732h,[27FFFFEh]
  LDRH R0,[27E57FEh]
  STRH 6000h,[4000204h]    ;EXMEMCNT, enable RAM, normal mode

DS Backwards-compatible GBA-Mode

When booting a 32pin GBA cartridge, the NDS is automatically switched into GBA mode, in that mode all NDS related features are disabled, and the console behaves (almost) like a GBA.

GBA Features that are NOT supported on NDS in GBA Mode.
Unlike real GBA, the NDS does not support 8bit DMG/CGB cartridges.
The undocumented Internal Memory Control register (Port 800h) isn't supported, so the NDS doesn't allow to use 'overclocked' RAM.
The NDS doesn't have a link-port, so GBA games can be played only in single player mode, link-port accessories cannot be used, and the NDS cannot run GBA code via multiboot.

GBA Features that are slightly different on NDS in GBA Mode.
The CPU, Timers, and Sound Frequencies are probably clocked at 16.76MHz; 33.51MHz/2; a bit slower than the original GBA's 16.78MHz clock?
In the BIOS, a single byte in a formerly 00h-filled area has been changed from 00h to 01h, resulting in SWI 0Dh returning a different BIOS checksum.
The GBA picture can be shown on upper or lower screen (selectable in boot-menu), the backlight for the selected screen is always on, resulting in different colors & much better visibility than original GBA. Unlike GBA-SP, the NDS doesn't have a backlight-button.

Screen Border in GBA mode
The GBA screen is centered in the middle of the NDS screen. The surrounding pixels are defined by 32K-color bitmap data in VRAM Block A and B. Each frame, the GBA picture is captured into one block, and is displayed in the next frame (while capturing new data to the other block).
To get a flicker-free border, both blocks should be initialized to contain the same image before entering GBA mode (usually both are zero-filled, resulting in a plain black border).
Note: When using two different borders, the flickering will be irregular - so there appears to be a frame inserted or skipped once every some seconds in GBA mode?!

Switching from NDS Mode to GBA Mode

  --- NDS9: ---
  ZEROFILL VRAM A,B     ;init black screen border (or other color/image)
  POWCNT=8003h          ;enable 2D engine A on upper screen (0003h=lower)
  EXMEMCNT=...          ;set Async Main Memory mode (clear bit14)
  IME=0                 ;disable interrupts
  SWI 06h               ;halt with interrupts disabled (lockdown)
  --- NDS7: ---
  POWERMAN.REG0=09h     ;enable sound amplifier & upper backlight (05h=lower)
  IME=0                 ;disable interrupts
  wait for VCOUNT=200   ;wait until VBlank
  SWI 1Fh with R2=40h   ;enter GBA mode, by CustomHalt(40h)

After that, the GBA BIOS will be booted, the GBA Intro will be displayed, and the GBA cartridge (if any) will be started.

DS Debug Registers (Emulator/Devkits)

No$gba Emulator Pseudo I/O Ports (no$gba) (GBA,NDS9,NDS7)

  4FFFA00h..A0Fh R Emulation ID (16 bytes, eg. "no$gba v2.7", padded with 20h)
  4FFFA10h       W String Out (raw)
  4FFFA14h       W String Out (with %param's)
  4FFFA18h       W String Out (with %param's, plus linefeed)
  4FFFA1Ch       W Char Out (nocash)
  4FFFA20h..A27h R Clock Cycles (64bit)
  4FFFA28h..A3Fh - N/A

Note: Above ports can be disabled via the "Debug I/O" option in no$gba setup.

Ensata Emulator Pseudo I/O Ports (NDS9)

  4000640h (32bit) ;aka CLIPMTX_RESULT (mis-used to invoke detection)
  4000006h (16bit) ;aka VCOUNT (mis-used to get detection result)
  4FFF010h (32bit) ;use to initialize/unlock/reset something
  4FFF000h (8bit)  ;debug message character output (used when Ensata detected)

The Ensata detection works by mis-using CLIPMTX_RESULT and VCOUNT registers:

  [4000640h]=2468ACE0h      ;CLIPMTX_RESULT (on real hardware it's read-only)
  if ([4000006h] AND 1FFh)=10Eh ;VCOUNT (on real hardware it's 000h..106h)
    [4FFF010h]=13579BDFh        ;\initialize/reset something
    [4FFF010h]=FDB97531h        ;/
    Ensata=true
  else
    Ensata=false
  endif

Once when a commercial game has detected Ensata, it stops communicating with the ARM7, and instead it does seem to want to communicate with the Ensata executable (which has little to do with real NDS hardware). Ie. aside from "unlocking" port 4FFF000h, it does also "lock" access to the ARM7 hardware (like sound, touchscreen, RTC, etc).

ISD (Intelligent Systems Debugger or so) I/O Ports
The ISD ports seem to be real (non-emulated) debugging ports, mapped to the GBA Slot region at 8000000h-9FFFFFFh, and used to output text messages, and possible also other debugging stuff.
There are appear to be two variants: nitroemu and cgbemu (the latter appears to be dating back to old 8bit CGB hardware; which was apparently still used for the NDS two hardware generations later).

NDS Devkit
In Nintendo's devkit, debug messages are handled in file "os_printf.c", this file detects the available hardware/software based debug I/O ports, and redirects the [OS_PutString] vector to the corresponding string_out function (eg. to OS_PutStringAris for writing a 00h-terminated string to port 4FFF000h). With some minimal efforts, this could be redirected to the corresponding no$gba debug I/O ports.

DS Cartridges, Encryption, Firmware

Cartridges
DS Cartridge Header
DS Cartridge Secure Area
DS Cartridge Icon/Title
DS Cartridge Protocol
DS Cartridge Backup
DS Cartridge NAND
DS Cartridge I/O Ports
DS Cartridge NitroROM and NitroARC File Systems
DS Cartridge Unknown Commands
DS Cartridge PassMe/PassThrough
DS Cartridge GBA Slot

Cartridge File Formats
DS File Formats

Add-Ons
DS Cart Rumble Pak
DS Cart Slider with Rumble
DS Cart Expansion RAM
DS Cart Infrared/Pedometers
DS Cart Unknown Extras

Special Cartridges
DS Cart Cheat Action Replay DS
DS Cart Cheat Codebreaker DS
DS Cart DLDI Driver

Encryption
DS Encryption by Gamecode/Idcode (KEY1)
DS Encryption by Random Seed (KEY2)

Firmware / Wifi Flash
DS Firmware Serial Flash Memory
DS Firmware Header
DS Firmware Wifi Calibration Data
DS Firmware Wifi Internet Access Points
DS Firmware User Settings
DS Firmware Extended Settings

DS Cartridge Header

Header Overview (loaded from ROM Addr 0 to Main RAM 27FFE00h on Power-up)

  Address Bytes Expl.
  000h    12    Game Title  (Uppercase ASCII, padded with 00h)
  00Ch    4     Gamecode    (Uppercase ASCII, NTR-<code>)        (0=homebrew)
  010h    2     Makercode   (Uppercase ASCII, eg. "01"=Nintendo) (0=homebrew)
  012h    1     Unitcode    (00h=NDS, 02h=NDS+DSi, 03h=DSi) (bit1=DSi)
  013h    1     Encryption Seed Select (00..07h, usually 00h)
  014h    1     Devicecapacity         (Chipsize = 128KB SHL nn) (eg. 7 = 16MB)
  015h    7     Reserved    (zero filled)
  01Ch    1     Reserved    (zero)                      (except, used on DSi)
  01Dh    1     NDS Region  (00h=Normal, 80h=China, 40h=Korea) (other on DSi)
  01Eh    1     ROM Version (usually 00h)
  01Fh    1     Autostart (Bit2: Skip "Press Button" after Health and Safety)
                (Also skips bootmenu, even in Manual mode & even Start pressed)
  020h    4     ARM9 rom_offset    (4000h and up, align 1000h)
  024h    4     ARM9 entry_address (2000000h..23BFE00h)
  028h    4     ARM9 ram_address   (2000000h..23BFE00h)
  02Ch    4     ARM9 size          (max 3BFE00h) (3839.5KB)
  030h    4     ARM7 rom_offset    (8000h and up)
  034h    4     ARM7 entry_address (2000000h..23BFE00h, or 37F8000h..3807E00h)
  038h    4     ARM7 ram_address   (2000000h..23BFE00h, or 37F8000h..3807E00h)
  03Ch    4     ARM7 size          (max 3BFE00h, or FE00h) (3839.5KB, 63.5KB)
  040h    4     File Name Table (FNT) offset
  044h    4     File Name Table (FNT) size
  048h    4     File Allocation Table (FAT) offset
  04Ch    4     File Allocation Table (FAT) size
  050h    4     File ARM9 overlay_offset
  054h    4     File ARM9 overlay_size
  058h    4     File ARM7 overlay_offset
  05Ch    4     File ARM7 overlay_size
  060h    4     Port 40001A4h setting for normal commands (usually 00586000h)
  064h    4     Port 40001A4h setting for KEY1 commands   (usually 001808F8h)
  068h    4     Icon/Title offset (0=None) (8000h and up)
  06Ch    2     Secure Area Checksum, CRC-16 of [[020h]..00007FFFh]
  06Eh    2     Secure Area Delay (in 131kHz units) (051Eh=10ms or 0D7Eh=26ms)
  070h    4     ARM9 Auto Load List Hook RAM Address (?) ;\endaddr of auto-load
  074h    4     ARM7 Auto Load List Hook RAM Address (?) ;/functions
  078h    8     Secure Area Disable (by encrypted "NmMdOnly") (usually zero)
  080h    4     Total Used ROM size (remaining/unused bytes usually FFh-padded)
  084h    4     ROM Header Size (4000h)
  088h    4     Unknown, some rom_offset, or zero? (DSi: slightly different)
  08Ch    8     Reserved (zero filled; except, [88h..93h] used on DSi)
  094h    2     NAND end of ROM area  ;\in 20000h-byte units (DSi: 80000h-byte)
  096h    2     NAND start of RW area ;/usually both same address (0=None)
  098h    18h   Reserved (zero filled)
  0B0h    10h   Reserved (zero filled; or "DoNotZeroFillMem"=unlaunch fastboot)
  0C0h    9Ch   Nintendo Logo (compressed bitmap, same as in GBA Headers)
  15Ch    2     Nintendo Logo Checksum, CRC-16 of [0C0h-15Bh], fixed CF56h
  15Eh    2     Header Checksum, CRC-16 of [000h-15Dh]
  160h    4     Debug rom_offset   (0=none) (8000h and up)       ;only if debug
  164h    4     Debug size         (0=none) (max 3BFE00h)        ;version with
  168h    4     Debug ram_address  (0=none) (2400000h..27BFE00h) ;SIO and 8MB
  16Ch    4     Reserved (zero filled) (transferred, and stored, but not used)
  170h    90h   Reserved (zero filled) (transferred, but not stored in RAM)
  200h    E00h  Reserved (zero filled) (usually not transferred)

DSi Cartridges are using an extended cartridge header,
DSi Cartridge Header
Some of that new/changed DSi header entries are important even in NDS mode:
- On DSi, ARM9/ARM7 areas are restricted to 2.75MB (instead 3.8MB on real NDS)
- New NDS titles must have RSA signatures (and old titles must be in whitelist)

For more info about CRC-16, see description of GetCRC16 BIOS function,
BIOS Misc Functions
For the Logo checksum, the BIOS verifies only [15Ch]=CF56h, it does NOT verify the actual data at [0C0h-15Bh] (nor it's checksum), however, the data is verified by the firmware.

Secure Area Delay
The Secure Area Delay at header[06Eh] is counted in 130.912kHz units (which can be clocked via one of the hardware timers with prescaler=F/256 and reload=(10000h-((X AND 3FFFh)+2)); for some weird reason, in case of Header checksum it's ANDed with 1FFFh instead of 3FFFh). Commonly used values are X=051Eh (10ms), and X=0D7Eh (26ms).
The delay is used for all Blowfish encrypted commands, the actual usage/purpose differs depending on bit31 of the ROM Chip ID:
When ChipID.Bit31=0 (commands are sent ONCE): The delay is issued BEFORE sending the command:

  Delay,Cmd

Older/newer games are using delays of 10ms/26ms (although all known existing cartridges with Bit31=0 would actually work WITHOUT delays).
When ChipID.Bit31=1 (commands are repeated MULTIPLE times): The delay is issued AFTER sending the command for the FIRST time:

  Cmd,Delay,Cmd                                ;for 2x repeat
  Cmd,Delay,Cmd,Cmd,Cmd,Cmd,Cmd,Cmd,Cmd,Cmd    ;for 9x repeat

Known games are using delays of 26ms (although all known existing cartridges (=Cooking Coach) with Bit31=1 would actually work with shorter delays of ca. 7ms (but, better use 8ms for safety)).

NDS Gamecodes
This is the same code as the NTR-UTTD (NDS) or TWL-UTTD (DSi) code which is printed on the package and sticker on (commercial) cartridges (excluding the leading "NTR-" or "TWL-" part).

  U  Unique Code          (usually "A", "B", "C", or special meaning)
  TT Short Title          (eg. "PM" for Pac Man)
  D  Destination/Language (usually "J" or "E" or "P" or specific language)

The first character (U) is usually "A" or "B", in detail:

  A NDS common games
  B NDS common games
  C NDS common games
  D DSi-exclusive games
  H DSiWare (system utilities and browser) (eg. HNGP=browser)
  I NDS and DSi-enhanced games with built-in Infrared port
  K DSiWare (dsiware games and flipnote) (eg. KGUV=flipnote)
  N NDS nintendo channel demo's japan (NTR-NTRJ-JPN)
  T NDS many games
  U NDS and DSi uncommon extra hardware (eg. NAND, ram, microSD, TV, azimuth)
  V DSi-enhanced games
  Y NDS many games

The second/third characters (TT) are:

  Usually an abbreviation of the game title (eg. "PM" for "Pac Man") (unless
  that gamecode was already used for another game, then TT is just random)

The fourth character (D) indicates Destination/Language:

  A Asian    E English/USA  I Italian   M Swedish  Q Danish   U Australian
  B N/A      F French       J Japanese  N Nor      R Russian  V EUR+AUS
  C Chinese  G N/A          K Korean    O Int      S Spanish  W..Z Europe #3..5
  D German   H Dutch        L USA #2    P Europe   T USA+AUS

DS Cartridge Secure Area

The Secure Area is located in ROM at 4000h..7FFFh, it can contain normal program code and data, however, it can be used only for ARM9 boot code, it cannot be used for ARM7 boot code, icon/title, filesystem, or other data.

Secure Area Size
The Secure Area exists if the ARM9 boot code ROM source address (src) is located within 4000h..7FFFh, if so, it will be loaded (by BIOS via KEY1 encrypted commands) in 4K portions, starting at src, aligned by 1000h, up to address 7FFFh. The secure area size if thus 8000h-src, regardless of the ARM9 boot code size entry in header.
Note: The BIOS silently skips any NDS9 bootcode at src<4000h.
Cartridges with src>=8000h do not have a secure area.

Secure Area ID
The first 8 bytes of the secure area are containing the Secure Area ID, the ID is required (verified by BIOS boot code), the ID value changes during boot process:

  Value                Expl.
  "encryObj"           raw ID before encryption (raw ROM-image)
  (encrypted)          encrypted ID after encryption (encrypted ROM-image)
  "encryObj"           raw ID after decryption (verified by BIOS boot code)
  E7FFDEFFh,E7FFDEFFh  destroyed ID (overwritten by BIOS after verify)

If the decrypted ID does match, then the BIOS overwrites the first 8 bytes by E7FFDEFFh-values (ie. only the ID is destroyed). If the ID doesn't match, then the first 800h bytes (2K) are overwritten by E7FFDEFFh-values.

Secure Area First 2K Encryption/Content
The first 2K of the Secure Area (if it exists) are KEY1 encrypted. In official games, this 2K region contains data like so (in decrypted form):

  000h..007h  Secure Area ID (see above)
  008h..00Dh  Fixed (FFh,DEh,FFh,E7h,FFh,DEh)
  00Eh..00Fh  CRC16 across following 7E0h bytes, ie. [010h..7FFh]
  010h..7FDh  Unknown/random values, mixed with some THUMB SWI calls
  7FEh..7FFh  Fixed (00h,00h)

Of which, only the ID in the first 8 bytes is verified. Neither BIOS nor (current) firmare versions are verifying the data at 008h..7FFh, so the 7F8h bytes may be also used for normal program code/data.

Avoiding Secure Area Encryption
WLAN files are reportedly same format as cartridges, but without Secure Area, so games with Secure Area cannot be booted via WLAN. No$gba can encrypt and decrypt Secure Areas only if the NDS BIOS-images are present. And, Nintendo's devkit doesn't seem to support Secure Area encryption of unreleased games.
So, unencrypted cartridges are more flexible in use. Ways to avoid encryption (which still work on real hardware) are:
1) Set NDS9 ROM offset to 4000h, and leave the first 800h bytes of the Secure Area 00h-filled, which can be (and will be) safely destroyed during loading; due to the missing "encryObj" ID; that method is used by Nintendo's devkit.
2) Set NDS9 ROM offset to 8000h or higher (cartridge has no Secure Area at all).
3) Set NDS9 ROM offset, RAM address, and size to zero, set NDS7 ROM offset to 200h, and point both NDS9 and NDS7 entrypoints to the loaded NDS7 region. That method avoids waste of unused memory at 200h..3FFFh, and it should be compatible with the NDS console, however, it is not comaptible with commercial cartridges - which do silently redirect address below 4000h to "addr=8000h+(addr AND 1FFh)". Still, it should work with inofficial flashcards, which do not do that redirection. No$gba emulates the redirection for regular official cartridges, but it disables redirection for homebrew carts if NDS7 rom offset<8000h, and NDS7 size>0.
[One possible problem: Newer "anti-passme" firmware versions reportedly check that the entrypoint isn't set to 80000C0h, that firmwares might also reject NDS9 entrypoints within the NDS7 bootcode region?]

DS Cartridge Icon/Title

The ROM offset of the Icon/Title is defined in CartHdr[68h]. The size was originally implied by the size of the original Icon/Title structure rounded to 200h-byte sector boundary (ie. A00h bytes for Version 1 or 2), however, later DSi carts are having a size entry at CartHdr[208h] (usually 23C0h).
If it is present (ie. if CartHdr[68h]=nonzero), then Icon/Title are displayed in the bootmenu.

  0000h 2     Version (0001h, 0002h, 0003h, or 0103h)
  0002h 2     CRC16 across entries 0020h..083Fh (all versions)
  0004h 2     CRC16 across entries 0020h..093Fh (Version 0002h and up)
  0006h 2     CRC16 across entries 0020h..0A3Fh (Version 0003h and up)
  0008h 2     CRC16 across entries 1240h..23BFh (Version 0103h and up)
  000Ah 16h   Reserved (zero-filled)
  0020h 200h  Icon Bitmap  (32x32 pix) (4x4 tiles, 4bit depth) (4x8 bytes/tile)
  0220h 20h   Icon Palette (16 colors, 16bit, range 0000h-7FFFh)
              (Color 0 is transparent, so the 1st palette entry is ignored)
  0240h 100h  Title 0 Japanese  (128 characters, 16bit Unicode)
  0340h 100h  Title 1 English   ("")
  0440h 100h  Title 2 French    ("")
  0540h 100h  Title 3 German    ("")
  0640h 100h  Title 4 Italian   ("")
  0740h 100h  Title 5 Spanish   ("")
  0840h 100h  Title 6 Chinese   ("")                 (Version 0002h and up)
  0940h 100h  Title 7 Korean    ("")                 (Version 0003h and up)
  0A40h 800h  Zerofilled (probably reserved for Title 8..15)

Below for animated DSi icons only (Version 0103h and up):

  1240h 1000h Icon Animation Bitmap 0..7 (200h bytes each, format as above)
  2240h 100h  Icon Animation Palette 0..7 (20h bytes each, format as above)
  2340h 80h   Icon Animation Sequence (16bit tokens)

Unused/padding bytes:

  0840h 1C0h  Unused/padding (FFh-filled) in Version 0001h
  0940h C0h   Unused/padding (FFh-filled) in Version 0002h
  23C0h 40h   Unused/padding (FFh-filled) in Version 0103h

Versions

  0001h = Original
  0002h = With Chinese Title
  0003h = With Chinese+Korean Titles
  0103h = With Chinese+Korean Titles and animated DSi icon

Title Strings
Usually, for non-multilanguage games, the same (english) title is stored in all title entries. The title may consist of ASCII characters 0020h-007Fh, character 000Ah (linefeed), and should be terminated/padded by 0000h.
The whole text should not exceed the dimensions of the DS cart field in the bootmenu (the maximum number of characters differs due to proportional font).
The title is usually split into a primary title, optional sub-title, and manufacturer, each separated by 000Ah character(s). For example: "America", 000Ah, "The Axis of War", 000Ah, "Cynicware", 0000h.

Icon Animation Sequence (DSi)
The sequence is represented by 16bit tokens, in the following format:

  15    Flip Vertically   (0=No, 1=Yes)
  14    Flip Horizontally (0=No, 1=Yes)
  13-11 Palette Index     (0..7)
  10-8  Bitmap Index      (0..7)
  7-0   Frame Duration    (01h..FFh) (in 60Hz units)

Value 0000h indicates the end of the sequence. If the first token is 0000h, then the non-animated default image is shown.
Uh, actually, a non-animated icon uses values 01h,00h,00h,01h, followed by 7Ch zerofilled bytes (ie. 0001h, 0100h, 3Eh x 0000h)?

FAT16:\title\000300tt\4ggggggg\data\banner.sav ;if carthdr[1BFh].bit2=1
Some DSi games are having a separate "banner.sav" file stored in the eMMC filesystem, enabled via carthdr[1BFh].bit2 (allowing to indicate the game progress by overriding the default icon). The banner files are 4000h bytes in size, the animation data is same as above, but without title strings and without non-animated icon.

  0000h 2     Version (0103h)
  0002h 6     Reserved (zero-filled)
  0008h 2     CRC16 across entries 0020h..119Fh (with initial value FFFFh)
  000Ah 16h   Reserved (zero-filled)
  0020h 1000h Icon Animation Bitmap 0..7 (200h bytes each)  ;\same format as
  1020h 100h  Icon Animation Palette 0..7 (20h bytes each)  ; in Icon/Title
  1120h 80h   Icon Animation Sequence (16bit tokens)        ;/
  11A0h 2E60h Garbage (random values, maybe due to eMMC decryption)

The feature is used by some Brain Age Express games (for example, Brain Age Express Sudoku: 'title\00030004\4b4e3945\data\banner.sav').
The feature does probably work only for DSiware titles (unless there are any DSi carts with SD/MMC access enabled; or unless there is a feature for storing similar data in cartridge memory).

DS Cartridge Protocol

Communication with Cartridge ROM relies on sending 8 byte commands to the cartridge, after the sending the command, a data stream can be received from the cartridge (the length of the data stream isn't fixed, below descriptions show the default length in brackets, but one may receive more, or less bytes, if desired).

Cartridge Memory Map

  0000000h-0000FFFh Header (unencrypted)
  0001000h-0003FFFh Not read-able (zero filled in ROM-images)
  0004000h-0007FFFh Secure Area, 16KBytes (first 2Kbytes with extra encryption)
  0008000h-...      Main Data Area

DSi cartridges are split into a NDS area (as above), and a new DSi area:

  XX00000h XX02FFFh DSi Not read-able (XX00000h=first megabyte after NDS area)
  XX03000h-XX06FFFh DSi ARM9i Secure Area (usually with modcrypt encryption)
  XX07000h-...      DSi Main Data Area

Cartridge memory must be copied to RAM (the CPU cannot execute code in ROM).

Command Summary, Cmd/Reply-Encryption Type, Default Length

  Command/Params    Expl.                             Cmd  Reply Len
  -- Unencrypted Load --
  9F00000000000000h Dummy (read HIGH-Z bytes)         RAW  RAW   2000h
  0000000000000000h Get Cartridge Header              RAW  RAW   200h DSi:1000h
  00aaaaaaaa000000h Get Cartridge Header (1T-ROM,NAND)RAW  RAW   200h
  9000000000000000h 1st Get ROM Chip ID               RAW  RAW   4
  A000000000000000h Get 3DS encryption type (3DS)     RAW  RAW   4
  00aaaaaaaa000000h Unencrypted Data (debug ver only) RAW  RAW   200h
  3Ciiijjjxkkkkkxxh Activate KEY1 Encryption (NDS)    RAW  RAW   0
  3Diiijjjxkkkkkxxh Activate KEY1 Encryption (DSi)    RAW  RAW   0
  3E00000000000000h Activate 16-byte commands (3DS)   RAW  RAW   0
  -- Secure Area Load --
  4llllmmmnnnkkkkkh Activate KEY2 Encryption Mode     KEY1 FIX   910h+0
  1lllliiijjjkkkkkh 2nd Get ROM Chip ID               KEY1 KEY2  910h+4
  xxxxxxxxxxxxxxxxh Invalid - Get KEY2 Stream XOR 00h KEY1 KEY2  910h+...
  2bbbbiiijjjkkkkkh Get Secure Area Block (4Kbytes)   KEY1 KEY2  910h+10A8h
  6lllliiijjjkkkkkh Optional KEY2 Disable             KEY1 KEY2  910h+?
  Alllliiijjjkkkkkh Enter Main Data Mode              KEY1 KEY2  910h+0
  -- Main Data Load --
  B7aaaaaaaa000000h Encrypted Data Read               KEY2 KEY2  200h
  B800000000000000h 3rd Get ROM Chip ID               KEY2 KEY2  4
  xxxxxxxxxxxxxxxxh Invalid - Get KEY2 Stream XOR 00h KEY2 KEY2  ...
  B500000000000000h Whatever NAND related? (DSi?)     KEY2 KEY2  0
  D600000000000000h Whatever NAND related? (DSi?)     KEY2 KEY2  4

The parameter digits contained in above commands are:

  aaaaaaaa     32bit ROM address (command B7 can access only 8000h and up)
  bbbb         Secure Area Block number (0004h..0007h for addr 4000h..7000h)
  x,xx         Random, not used in further commands (DSi: always zero)
  iii,jjj,llll Random, must be SAME value in further commands
  kkkkk        Random, must be INCREMENTED after FURTHER commands
  mmm,nnn      Random, used as KEY2-encryption seed

 ____________ Unencrypted Commands (First Part of Boot Procedure) _____________

Cartridge Reset
The /RES Pin switches the cartridge into unencrypted mode. After reset, the first two commands (9Fh and 00h) are transferred at 4MB/s CLK rate.

9F00000000000000h (2000h) - Dummy
Dummy command send after reset, returns endless stream of HIGH-Z bytes (ie. usually receiving FFh, immediately after sending the command, the first 1-2 received bytes may be equal to the last command byte).

0000000000000000h (200h) (DSi:1000h) - Get Header (from address 00000000h)
00aaaaaaaa000000h (200h) - DSi Get Header Snippet (from address aaaaaaaah)
Returns RAW unencrypted cartridge header, usually repeated every 1000h bytes. Some carts allow to read the header with a single 200h-byte or 1000h-byte read, others require eight separate 200h-byte reads.

  NDS/MROM   --> Read 200h bytes from address 000h
  NDS/1T-ROM --> Read 200h bytes from address 000h
  NDS/NAND   --> Read 200h bytes from address 000h
  DSi/MROM   --> Read 1000h bytes from address 000h
  DSi/1T-ROM --> Read 8x200h bytes from address 000h,200h,400h,..,E00h
  DSi/NAND   --> Read 8x200h bytes from address 000h,200h,400h,..,E00h

Note: DSi/1T-ROM usually allows to read 1000h bytes or 8x200h, but DSi/NAND works only when reading 8x200h bytes.
The Gamecode header entry is used later on to initialize the encryption. Also, the ROM Control entries define the length of the KEY1 dummy periods (typically 910h clocks), and the CLK transfer rate for further commands (typically faster than the initial 4MB/s after power up).

9000000000000000h (4) - 1st Get ROM Chip ID
Returns RAW unencrypted Chip ID (eg. C2h,0Fh,00h,00h), repeated every 4 bytes.

  1st byte - Manufacturer (eg. C2h=Macronix) (roughly based on JEDEC IDs)
  2nd byte - Chip size (00h..7Fh: (N+1)Mbytes, F0h..FFh: (100h-N)*256Mbytes?)
  3rd byte - Flags (see below)
  4th byte - Flags (see below)

The Flag Bits in 3th byte can be

  0   Uses Infrared (but via SPI, unrelated to ROM) (also Jam with the Band)
  1   Unknown (set in some 3DS carts)
  2-6 Zero
  7   Unknown (set in Kingdom Hearts - Re-Coded)

The Flag Bits in 4th byte can be

  0-2 Zero
  3   NAND flag (0=ROM, 1=NAND)
  4   3DS Flag  (0=NDS/DSi, 1=3DS)
  5   Unknown   (0=Normal, 1=Support cmd B5h/D6h)
  6   DSi flag  (0=NDS/3DS, 1=DSi) (but also set in NDS Walk with Me)
  7   Cart Protocol Variant (0=old/smaller MROM, 1=new/bigger 1T-ROM or NAND)

Existing/known ROM IDs are:

  C2h,07h,00h,00h NDS Macronix 8MB ROM  (eg. DS Vision, with microSD slot)
  AEh,0Fh,00h,00h NDS Noname   16MB ROM (eg. Meine Tierarztpraxis)
  C2h,0Fh,00h,00h NDS Macronix 16MB ROM (eg. Metroid Demo)
  C2h,1Fh,00h,00h NDS Macronix 32MB ROM (eg. Over the Hedge)
  C2h,1Fh,00h,40h DSi Macronix 32MB ROM (eg. Art Academy, TWL-VAAV, SystemFlaw)
  80h,3Fh,01h,E0h NDS SanDisk  64MB ROM+Infrared (eg. Walk with Me, NTR-IMWP)
  AEh,3Fh,00h,E0h DSi Noname   64MB ROM (eg. de Blob 2, TWL-VD2V)
  C2h,3Fh,00h,00h NDS Macronix 64MB ROM (eg. Ultimate Spiderman)
  C2h,3Fh,00h,40h DSi Macronix 64MB ROM (eg. Crime Lab, NTR-VAOP)
  80h,7Fh,00h,80h NDS SanDisk  128MB ROM (DS Zelda, NTR-AZEP-0)
  80h,7Fh,01h,E0h ?   SanDisk? 128MB ROM+Infrared (P-letter SoulSilver, IPGE)
  C2h,7Fh,00h,80h NDS Macronix 128MB ROM (eg. Spirit Tracks, NTR-BKIP)
  C2h,7Fh,00h,C0h DSi Macronix 128MB ROM (eg. Cooking Coach, TWL-VCKE)
  ECh,7Fh,00h,88h NDS Samsung  128MB NAND (eg. Warioware D.I.Y., NTR-UORE)
  ECh,7Fh,01h,88h NDS Samsung  128MB NAND (eg. Jam with the Band, NTR-UXBP)
  ECh,7Fh,00h,E8h DSi Samsung  128MB NAND (eg. Face Training, TWL-USKV)
  80h,FFh,80h,E0h NDS SanDisk? 256MB ROM (Kingdom Hearts - Re-Coded, NTR-BK9P)
  C2h,FFh,01h,C0h DSi Macronix 256MB ROM+Infrared (eg. P-Letter White)
  C2h,FFh,00h,80h NDS Macronix 256MB ROM (eg. Band Hero, NTR-BGHP)
  C2h,FEh,01h,C0h DSi Macronix 512MB ROM+Infrared (eg. P-Letter White 2)
  C2h,FEh,00h,90h 3DS Macronix probably 512MB? ROM (eg. Sims 3)
  45h,FAh,00h,90h 3DS SanDisk? maybe... 1GB?   ROM (eg. Starfox)
  C2h,F8h,00h,90h 3DS Macronix maybe... 2GB?   ROM (eg. Kid Icarus)
  C2h,7Fh,00h,90h 3DS Macronix 128MB ROM CTR-P-AENJ MMinna no Ennichi
  C2h,FFh,00h,90h 3DS Macronix 256MB ROM CTR-P-AFSJ Pro Yakyuu Famista 2011
  C2h,FEh,00h,90h 3DS Macronix 512MB ROM CTR-P-AFAJ Real 3D Bass FishingFishOn
  C2h,FAh,00h,90h 3DS Macronix 1GB ROM CTR-P-ASUJ Hana to Ikimono Rittai Zukan
  C2h,FAh,02h,90h 3DS Macronix 1GB ROM CTR-P-AGGW Luigis Mansion 2 ASiA CHT
  C2h,F8h,00h,90h 3DS Macronix 2GB ROM CTR-P-ACFJ Castlevania - Lords of Shadow
  C2h,F8h,02h,90h 3DS Macronix 2GB ROM CTR-P-AH4J Monster Hunter 4
  AEh,FAh,00h,90h 3DS Noname?  1GB ROM CTR-P-AGKJ Gyakuten Saiban 5
  AEh,FAh,00h,98h 3DS Noname?  1GB NAND CTR-P-EGDJ Tobidase Doubutsu no Mori
  45h,FAh,00h,90h 3DS SanDisk? 1GB ROM CTR-P-AFLJ Fantasy Life
  45h,F8h,00h,90h 3DS SanDisk? 2GB ROM CTR-P-AVHJ Senran Kagura Burst - Guren
  C2h,F0h,00h,90h 3DS Macronix 4GB ROM CTR-P-ABRJ Biohazard Revelations
  ?,?,?,?         NDS ?        ? (eg. Japanese TV Tuner, NTR-UNSJ)
  00h,00h,00h,00h Cart Reset Busy (Face Training needs 20ms delay after reset)
  FFh,FFh,FFh,FFh None (no cartridge inserted)

The official JEDEC ID for Samsung would be "CEh", but for some reason, Samsung's NDS chip does spit out "ECh" as Maker ID. SanDisk has two IDs (80h on NDS/DSi, and 45h on 3DS).

3Ciiijjjxkkkkkxxh (0) - Activate KEY1 Encryption Mode
The 3Ch command returns endless stream of HIGH-Z bytes, all following commands, and their return values, are encrypted. The random parameters iii,jjj,kkkkk must be re-used in further commands; the 20bit kkkkk value is to be incremented by one after each <further> command (it is <not> incremented after the 3Ch command).

3Diiijjjxkkkkkxxh (0) - Activate KEY1 Encryption Mode and Unlock DSi Mode
Same as command 3Ch (but with different initial 1048h-byte encryption values), and works only on DSi carts. Command 3Dh is unlocking two features on DSi carts:

  1) Command 2bbbbiiijjjkkkkkh loads ARM9i secure area (instead of ARM9 area)
  2) Command B7aaaaaaaa000000h allows to read the 'whole' cartridge space

Without command 3Dh, DSi carts will allow to read only the first some megabytes (for example, the first 11 Mbyte of the System Flaw cartridge), and the remaining memory returns mirrors of "addr=8000h+(addr AND 1FFh)").
Note: After reset, the cartridge protocol allows to send only either one of the 3Ch/3Dh commands (DSi consoles can control the cartridge reset pin, so they can first send 3Ch and read the normal secure area, then issue a reset and 3Dh and read the DSi secure area) (on a NDS one could do the same by ejecting/inserting the cartridge instead of toggling the reset pin).

 ____________ KEY1 Encrypted Commands (2nd Part of Boot procedure) ____________

4llllmmmnnnkkkkkh (910h) - Activate KEY2 Encryption Mode
KEY1 encrypted command, parameter mmmnnn is used to initialize the KEY2 encryption stream. Returns 910h dummy bytes (which are still subject to old KEY2 settings; at pre-initialization time, this is fixed: HIGH-Z, C5h, 3Ah, 81h, etc.). The new KEY2 seeds are then applied, and the first KEY2 byte is then precomputed. The 910h dummy stream is followed by that precomputed byte value endless repeated (this is the same value as that "underneath" of the first HIGH-Z dummy-byte of the next command).
Secure1000h: Returns repeated FFh bytes (instead of the leading C5h, 3Ah, 81h, etc. stuff).
Secure1000h: Returns repeated FFh bytes (instead of the repeated precomputed value).

1lllliiijjjkkkkkh (914h) - 2nd Get ROM Chip ID / Get KEY2 Stream
KEY1 encrypted command. Returns 910h dummy bytes, followed by KEY2 encrypted Chip ID repeated every 4 bytes, which must be identical as for the 1st Get ID command. The BIOS randomly executes this command once or twice. Changing the first command byte to any other value returns an endless KEY2 encrypted stream of 00h bytes, that is the easiest way to retrieve encryption values and to bypass the copyprotection.

2bbbbiiijjjkkkkkh (19B8h) - Get Secure Area Block
KEY1 encrypted command. Used to read a secure area block (bbbb in range 0004h..0007h for addr 4000h..7000h) (or, after sending command 3Dh on a DSi: bbbb in range 0004h..0007h for addr XX03000h..XX06000h).
Each block is 4K, so it requires four Get Secure Area commands to receive the whole Secure Area (ROM locations 4000h-7FFFh), the BIOS is reading these blocks in random order.
Normally (if the upper bit of the Chip ID is set): Returns 910h dummy bytes, followed by 200h KEY2 encrypted Secure Area bytes, followed by 18h KEY2 encrypted 00h bytes, then the next 200h KEY2 encrypted Secure Area bytes, again followed by 18h KEY2 encrypted 00h bytes, and so on. That stream is repeated every 10C0h bytes (8x200h data bytes, plus 8x18h zero bytes).
Alternately (if the upper bit of the Chip ID is zero): Returns 910h dummy bytes, followed by 1000h KEY2 encrypted Secure Area bytes, presumably followed by 18h bytes, too.
Aside from above KEY2 encryption (which is done by hardware), the first 2K of the NDS Secure Area is additionally KEY1 encrypted; which must be resolved after transfer by software (and the DSi Secure Area is usually modcrypted, as specified in the cartridge header).

6lllliiijjjkkkkkh (0) - Optional KEY2 Disable
KEY1 encrypted command. Returns 910h dummy bytes (which are still KEY2 affected), followed by endless stream of RAW 00h bytes. KEY2 encryption is disabled for all following commands.
This command is send only if firmware[18h] matches encrypted string "enPngOFF", and ONLY if firmware get_crypt_keys had completed BEFORE completion of secure area loading, this timing issue may cause unstable results.

Alllliiijjjkkkkkh (910h) - Enter Main Data Mode
KEY1 encrypted command. Returns 910h dummy bytes, followed by endless KEY2 encrypted stream of 00h bytes. All following commands are KEY2 encrypted.

 ________________ KEY2 Encrypted Commands (Main Data Transfer) ________________

B7aaaaaaaa000000h (200h) - Get Data
KEY2 encrypted command. The desired ROM address is specifed, MSB first, in parameter bytes (a). Returned data is KEY2 encrypted.
There is no alignment restriction for the address. However, the datastream wraps to the begin of the current 4K block when address+length crosses a 4K boundary (1000h bytes). Special case: SanDisk ROMs are forcefully 200h-byte aligned, and can merely read max 200h bytes (padded with unencrypted FFh bytes when trying to read more data).
The command can be used only for addresses 8000h and up. Addresses 0..7FFFh are silently redirected to address "8000h+(addr AND 1FFh)". DSi cartridges will also reject XX00000h..XX06FFFh in the same fashion (and also XX07000h and up if the DSi cartridge isn't unlocked via command 3Dh).
Addresses that do exceed the ROM size do mirror to the valid address range (that includes mirroring non-loadable regions like 0..7FFFh to "8000h+(addr AND 1FFh)"; some newer games are using this behaviour for some kind of anti-piracy checks).

B800000000000000h (4) - 3rd Get ROM Chip ID
KEY2 encrypted command. Returns KEY2 encrypted Chip ID repeated every 4 bytes.

xxxxxxxxxxxxxxxxh - Invalid Command
Any other command (anything else than above B7h and B8h) in KEY2 command mode causes communcation failures. The invalid command returns an endless KEY2 encrypted stream of 00h bytes. After the invalid command, the KEY2 stream is NOT advanced for further command bytes, further commands seems to return KEY2 encrypted 00h bytes, of which, the first returned byte appears to be HIGH-Z.
Ie. the cartridge seems to have switched back to a state similar to the KEY1-phase, although it doesn't seem to be possible to send KEY1 commands.

 ___________________________________ Notes ___________________________________

KEY1 Command Encryption / 910h Dummy Bytes
All KEY1 encrypted commands are followed by 910h dummy byte transfers, these 910h clock cycles are probably used to decrypt the command at the cartridge side; communication will fail when transferring less than 910h bytes.
The return values for the dummy transfer are: A single HIGH-Z byte, followed by 90Fh KEY2-encrypted 00h bytes. The KEY2 encryption stream is advanced for all 910h bytes, including for the HIGH-Z byte.
Note: Current cartridges are using 910h bytes, however, other carts might use other amounts of dummy bytes, the 910h value can be calculated based on ROM Control entries in cartridge header. For the KEY1 formulas, see:
DS Encryption by Gamecode/Idcode (KEY1)

KEY2 Command/Data Encryption
DS Encryption by Random Seed (KEY2)

Cart Protocol Variants (Chip ID.Bit31)
There are two protocol variants for NDS carts, indicated by Bit31 of the ROM Chip ID (aka bit7 of the 4th ID byte):

  1) Chip ID.Bit31=0  Used by older/smaller carts with up to 64MB ROM
  2) Chip ID.Bit31=1  Used by newer/bigger carts with 64MB or more ROM

The first variant (for older carts) is described above. The second second variant includes some differences for KEY1 encrypted commands:
GAPS: The commands have the same 910h-cycle gaps, but without outputting CLK pulses during those gaps (ie. used with ROMCTRL.Bit28=0) (the absence of the CLKs implies that there is no dummy data transferred during gaps, and accordingly, that the KEY2 stream isn't advanced during the 910h gap cycles).
REPEATED COMMANDS and SECURE AREA DELAY: All KEY1 encrypted commands must be sent TWICE (or even NINE times). First, send the command with 0-byte Data transfer length. Second, issue the Secure Area Delay (required; use the delay specified in cart header[06Eh]).
Third, send the command once again with 0-byte or 4-byte data transfer length (usually 0 bytes, or 4-bytes for Chip ID command), or sent it eight times with 200h-byte data transfer length (for the 1000h-byte secure area load command).
For those repeats, always resend exactly the same command (namely, kkkkk is NOT incremented during repeats, and there is no extra index needed to select 200h-byte portions within 1000h-byte blocks; the cartridge is automatically outputting the eight portions one after another).

DS Cartridge Backup

SPI Bus Backup Memory

  Type   Total Size  Page Size  Chip/Example      Game/Example
  EEPROM 0.5K bytes   16 bytes  ST M95040-W       (eg. Metroid Demo)
  EEPROM   8K bytes   32 bytes  ST M95640-W       (eg. Super Mario DS)
  EEPROM  64K bytes  128 bytes  ST M95512-W       (eg. Downhill Jam)
  EEPROM 128K bytes    ? bytes  ?                 (eg. Explorers of Sky)
  FLASH  256K bytes  256 bytes  ST M45PE20        (eg. Skateland)
  FLASH  256K bytes             Sanyo LE25FW203T  (eg. Mariokart)
  FLASH  512K bytes  256 bytes  ST M25PE40?       (eg. which/any games?)
  FLASH  512K bytes             ST 45PE40V6       (eg. DS Zelda, NTR-AZEP-0)
  FLASH 1024K bytes             ST 45PE80V6       (eg. Spirit Tracks, NTR-BKIP)
  FLASH 8192K bytes             MX25L6445EZNI-10G (Art Academy only, TWL-VAAV)
  FRAM     8K bytes   No limit  ?                 (eg. which/any games?)
  FRAM    32K bytes   No limit  Ramtron FM25L256? (eg. which/any games?)

Lifetime Stats (might vary per manufacturer)

  Type      Max Writes per Page    Data Retention
  EEPROM    100,000                40 years
  FLASH     100,000                20 years
  FRAM      No limit               10 years

SPI Bus Backup Memory is accessed via Ports 40001A0h and 40001A2h, see
DS Cartridge I/O Ports

Commands
For all EEPROM and FRAM types:

  06h WREN  Write Enable                Cmd, no parameters
  04h WRDI  Write Disable               Cmd, no parameters
  05h RDSR  Read Status Register        Cmd, read repeated status value(s)
  01h WRSR  Write Status Register       Cmd, write one-byte value
  9Fh RDID  Read JEDEC ID (not supported on EEPROM/FLASH, returns FFh-bytes)

For 0.5K EEPROM (8+1bit Address):

  03h RDLO  Read from Memory 000h-0FFh  Cmd, addr lsb, read byte(s)
  0Bh RDHI  Read from Memory 100h-1FFh  Cmd, addr lsb, read byte(s)
  02h WRLO  Write to Memory 000h-0FFh   Cmd, addr lsb, write 1..MAX byte(s)
  0Ah WRHI  Write to Memory 100h-1FFh   Cmd, addr lsb, write 1..MAX byte(s)

For 8K..64K EEPROM and for FRAM (16bit Address):

  03h RD    Read from Memory            Cmd, addr msb,lsb, read byte(s)
  02h WR    Write to Memory             Cmd, addr msb,lsb, write 1..MAX byte(s)

For 128K EEPROM (24bit Address):

  As above, but with 24bit addr msb,mid,lsb ?

Note: MAX = Page Size (see above chip list) (no limit for FRAM).

For FLASH backup, commands should be same as for Firmware FLASH memory:
DS Firmware Serial Flash Memory
A few NDS/DSi carts are sharing the SPI bus for FLASH and Infrared, this requires a 00h-prefix byte for FLASH access, with slower 1MHz SPI clock and delays, see:
DS Cart Infrared/Pedometers

Status Register

  0   WIP  Write in Progress (1=Busy) (Read only) (always 0 for FRAM chips)
  1   WEL  Write Enable Latch (1=Enable) (Read only, except by WREN,WRDI)
  2-3 WP   Write Protect (0=None, 1=Upper quarter, 2=Upper Half, 3=All memory)

For 0.5K EEPROM:

  4-7 ONEs Not used (all four bits are always set to "1" each)

For 8K..64K EEPROM and for FRAM:

  4-6 ZERO Not used (all three bits are always set to "0" each)
  7   SRWD Status Register Write Disable (0=Normal, 1=Lock) (Only if /W=LOW)

WEL gets reset on Power-up, WRDI, WRSR, WRITE/LO/HI, and on /W=LOW.
The WRSR command allows to change ONLY the two WP bits, and the SRWD bit (if any), these bits are non-volatile (remain intact during power-down), respectively, the WIP bit must be checked to sense WRSR completion.

Detection (by examining hardware responses)
The overall memory type and bus-width can be detected by RDSR/RDID commands:

  RDSR  RDID          Type         (bus-width)
  FFh,  FFh,FFh,FFh   None         (none)
  F0h,  FFh,FFh,FFh   EEPROM       (with 8+1bit address bus)
  00h,  FFh,FFh,FFh   EEPROM/FRAM  (with 16bit address bus)
  ?     ?,?,?         EEPROM       (with 24bit address bus)
  00h,  xxh,xxh,xxh   FLASH        (usually with 24bit address bus)

And, the RD commands can be used to detect the memory size/mirrors (though that won't work if the memory is empty).

Detection (in emulators)
Nintendo is using different functions for sending cmd+addr and data. The bus-width can be detected by counting the bytes transferred with same program counter after chip selection. One could also try to examine code/data in the ROM-image (but that may envolve self-decompressing code and other obstacles).
Special cases:
Over the Hedge does initially try to access 8Kbyte EEPROM, but does actually use 0.5Kbyte EEPROM (as workaround: re-detect the bus-width on each transfer).
Rune Factory - A Fantasy Harvest Moon seems to be also difficult to detect (as workaround: force 64Kbyte EEPROM on gamecode ARFx).
FLASH has same 24bit bus-width as 128Kbyte EEPROM, but isn't compatible on writing. EEPROM does use 02h=Write+Erase. FLASH does use 0Ah=Write+Erase (or D8h/DBh=Erase and 02h=Write separatly).

Pin-Outs for EEPROM and FRAM chips

  Pin Name Expl.
  1  /S    Chip Select
  2  Q     Data Out
  3  /W    Write-Protect (not used in NDS, wired to VCC)
  4  VSS   Ground
  5  D     Data In
  6  C     Clock
  7  /HOLD Transfer-pause (not used in NDS, wired to VCC)
  8  VCC   Supply 2.5 to 5.5V for M95xx0-W

FRAM (Ferroelectric Nonvolatile RAM) is fully backwards compatible with normal EEPROMs, but comes up with faster write/erase time (no delays), and with lower power consumption, and unlimited number of write/erase cycles. Unlike as for normal RAM, as far as I understand, the data remains intact without needing any battery.

Other special save memory

  DS Vision (NDS cart with microSD slot... and maybe ALSO with EEPROM?)
  NAND carts can store data in a read/write-able portion of the "ROM" chip
  Typing Adventure does have SPI FLASH (but not directly wired to SPI bus)

DSi Internal eMMC and External SD Card
DSi cartridges are usually (maybe always) having SD/MMC access disabled, so they must stick using EEPROM/FLASH chips inside of the cartridges (which is required for NDS compatibility anyways).
However, DSiware games (downloaded from DSi Shop) are allowed to save data on eMMC, using "private.sav" or "public.sav" files in their data folder. The size of that files is preset in cartridge header.

DS Cartridge NAND

SLC-NAND is used in at least three NDS/DSi games. The ROM Chip ID has bit3 in 4th byte set.

  ECh,7Fh,00h,88h NDS Samsung  128MB NAND (eg. Warioware D.I.Y., NTR-UORE)
  ECh,7Fh,01h,88h NDS Samsung  128MB NAND (eg. Jam with the Band, NTR-UXBP)
  ECh,7Fh,00h,E8h DSi Samsung  128MB NAND (eg. Face Training, TWL-USKV)

In the cart header, the 1st byte of the Gamecode is "U" (that "U" is also used for a few other carts with "uncommon" hardware), and header entries [094h,096h] indicate the end of the ROM region and start of RW region (in 128Kbyte units for NDS, or 512Kbyte units for DSi; exception: the oldest NAND title (japanese version of Jam with the Band) did have [094h,096h] set to all zeroes). The chips are all 128MByte (122MByte usable), with memory map as so:

  00000000h  ROM region (one large region)                                  (R)
  0xxx0000h  RW region (split into several 128KByte blocks)               (R/W)
  07A00000h  Reserved region                                                (R)

The RW space is 8MB for Jam with the Band, 16MB for WarioWare DIY, and 82MB in Face Training. Unknown if some of the cart memory is reserved for broken sector handling.

DS Cartridges with NAND memory
NAND memory contains both the game and save memory (normal NDS games contain separate ROM and SPI FLASH/EEPROM chips for that purposes). The advantage is that NAND allows more storage than the usual FLASH chips. Nintendo also claims SLC-NAND to be very fast, but that's only half true (it's much slower than MROM, but might be slightly faster than SPI FLASH).

Command Summary
For whatever reason, random access is slightly restricted: One must use command 8Bh/B2h to select ROM region or a 128KByte RW window before reading/writing the selected area. Writing is done in 2Kbyte units (4x200h bytes).

 In ROM access mode:
  9400000000000000h Len=200h  NAND Read ID
  B2aaaaaaaa000000h Len=0     NAND Select 128Kbyte RW access mode
  B300000000000000h Len=04h   Unknown (returns 00000000h)
  BB00000000000000h Len=200h  Unknown (returns 1X 04 09 20 04, plus zeroes)
 In RW access mode (on DSi carts, this works ONLY in DSi mode):
  81aaaaaaaa000000h Len=200h  NAND Write to Write Buffer (must be issued 4x)
  8200000000000000h Len=0     NAND Forward Write Buffer to NAND
  8400000000000000h Len=0     NAND Discard Write Buffer
  8500000000000000h Len=0     NAND Write Enable
  8600000000000000h Len=0     Unknown
  8700000000000000h Len=0     NAND Write Disable
  8B00000000000000h Len=0     NAND Select ROM access mode
 In either mode:
  0B00000000000000h Len=200h  Returns cart header[000h..1FFh]
  0C00000000000000h Len=200h  Returns corrupted cart header[1F8h..3F7h] ??
  58h..5Fh          Len=0     Unknown (looks same/similar as in 1T-ROM carts)
  60h..68h          Len=800h  Unknown (looks same/similar as in 1T-ROM carts)
  B000000000000000h Len=04h   Unknown (returns 01010101h)
  B500000000000000h Len=0     Unknown (looks same/similar as in SanDisk carts)
  B7aaaaaaaa000000h Len=200h  NAND Read from ROM or RW area
  B800000000000000h Len=04h   Read Chip ID
  D600000000000000h Len=04h   NAND Read Status
 Further command(s) spotted in Face Training disassembly:
  8800000000000000h Len=0     Unknown (is in disassembly, but fails on HW?)

81aaaaaaaa000000h - NAND Write to Write Buffer + Data[200h]
The command must be issued 4 times with the same address (seems to use the address from the first command), a full write is 800h bytes (4x200h), the chip won't respond to other commands until all 4 commands are sent. Unknown what happens when trying to send more than 4 commands.
This command doesn't directly write to NAND, instead it stores the incoming data in a 800h-byte write buffer (latter committed using command 82h).
The provided address needs to be within the access window specified by command B2h.

8200000000000000h - NAND Forward Write Buffer to NAND
The data in the 800h-byte write buffer is written to the actual NAND, and the write-enable bit in the status register is cleared.

8400000000000000h - NAND Discard Write Buffer
The data in the write buffer is discarded (a subsequent command 82h would have no effect).
Games seem to always use this after command 82h, so maybe it's required even after a write.

8500000000000000h - NAND Write Enable
The write-enable bit in the status register is set.
Only works in RW access mode.

8B00000000000000h - NAND Select ROM access mode
Switch to ROM access mode.

B2aaaaaaaa000000h - NAND Select 128Kbyte RW access mode
The provided address defines an accessible 128K window in RW address space, reads and writes may only take place within that window, the lower 17 bits of the address are ignored.
Addresses below start of RW area are ignored (the cart stays in ROM mode, and the cart does then reportedly get stuck permanently busy).
Addresses that go past the end of the RW address space will just read all FF's.

B7aaaaaaaa000000h - NAND Read from ROM or RW area + Data[200h]
This command is used for reading data under both ROM and RW modes.
In ROM mode: returns data from the ROM space, pretty much like the 'regular' B7 command, trying to read from RW space in this mode will return all FF's.
In RW mode: returns data from the RW space, the provided address needs to be within the access window specified by command B2h.

D600000000000000h - NAND Read Status + Data[4]
Status register bits:

  0-1   Unknown (usually zero)
  2-3   Unknown (usually zero, but tested by DSi Launcher, not NAND related?)
  4     NAND write enable
  5     NAND status (0=busy, 1=ready)
  6     Unknown (usually zero, but set by DeSmuME)
  7     Unknown (possible error flag?)
  8-15  Same as bit0-7
  16-23 Same as bit0-7
  24-31 Same as bit0-7

Value on startup is 20h (aka 20202020h when reading 4 bytes).

BB00000000000000h Len=200h Unknown (returns 1X 04 09 20 04, plus zeroes)

  Values in Jam with the Band (nocash dump): 17 04 09 20 04, plus 1FBh zeroes
  Values in Face Training (nocash dump):     10 04 09 20 04, plus 1FBh zeroes

9400000000000000h - NAND Read ID + Data[200h]
Returns NAND ID data of sorts.

 Values in Jam with the Band (arisotura dump):
  000h  EC F1 00 95 40 00 00 00 00 00 00 00 00 00 00 00  ....@...........
  010h  00 00 00 00 00 00 00 00 EC 00 9E A1 51 65 34 35  ............Qe45
  020h  30 35 30 31 19 19 02 0A 00 00 00 00 00 00 00 00  0501............
  030h  FF FF FF .. (1D0h bytes) (why not 00h's ???)     ................
 Values in Jam with the Band (nocash dump):
  000h  EC F1 00 95 40 00 00 00 00 00 00 00 00 00 00 00  ....@...........
  010h  00 00 00 00 00 00 00 00 EC 00 3B 5A 32 9B 32 30  ..........;Z2.20
  020h  35 35 30 30 19 19 02 0A 00 00 00 00 00 00 00 00  5500............
  030h  00 00 00 .. (1D0h bytes)                         ................
 Values in Face Training (nocash dump):
  000h  EC F1 00 95 40 00 00 00 00 00 00 00 00 00 00 00  ....@...........
  010h  00 00 00 00 00 00 00 00 EC 00 5A 36 5C 14 35 35  ..........Z6\.55
  020h  32 36 30 36 04 04 08 0A 00 00 00 00 00 00 00 00  2606............
  030h  00 00 00 .. (1D0h bytes)                         ................

For Jam with Band, the 16 bytes at 018h are also found in the last 800h bytes of the RW space (see below). The bytes with value ECh might be related to the Maker ID in Chip ID.

Last 128Kbyte in RW region (at 079E0000h..079FFFFFh)
The last 128Kbyte of RW area is normal write-able memory in Face Training without special content. However, in Jam with the Band, the last 128KByte are readonly, they can be read like normal RW blocks (select via command B2h, then read via B7h), but writing isn't supported (although NAND Status bit5 gets cleared for a moment upon write attempts to this 128Kbyte area; unlike as for when trying to write 7A00000h and up).

 Values in Jam with the Band (arisotura dump):
  079E0000h  FF FF FF .. (1F800h bytes)                      ................
  079FF800h  EC 00 9E A1 51 65 34 35 30 35 30 31 19 19 02 0A ....Qe450501....
  079FF810h  00 00 00 00 6D D6 DA 9B B0 24 22 88 79 3B BF EA ....m....$".y;..
  079FF820h  E6 AC 5E FA 69 12 0D 52 5D 5B F5 80 FF FF FF FF ..^.i..R][......
  079FF830h  FF FF FF .. (7D0h bytes)                        ................
 Values in Jam with the Band (nocash dump):
  079E0000h  FF FF FF .. (1F800h bytes)                      ................
  079FF800h  EC 00 3B 5A 32 9B 32 30 35 35 30 30 19 19 02 0A ..;Z2.205500....
  079FF810h  00 00 00 00 DD 58 84 07 F9 72 19 04 96 8C FF 67 .....X...r.....g
  079FF820h  7F 66 B9 E5 FD F7 3F 1A AE 60 60 00 FF FF FF FF .f....?..``.....
  079FF830h  FF FF FF .. (7D0h bytes)                        ................

Jam with the Band checks the bytes at 079FF800h (unknown what it does with them, or what it does when they are missing).

Reserved Area (at 07A00000h and up) (including 08000000h and up)
This memory isn't intended to be used, in Jam with the Band it just returns FFh's. However, in Face Training it does return some interesting internal data (in RW mode):
The DSi Blowfish key (with the gamecode pre-applied), some NDS ARM code (for the secure area or so), and most interestingly some more ARM+THUMB code (apparently containing the firmware running on an ARM CPU inside of the game cartridge itself).
Unfortunately, that internal memory is returned as "raw" data with faulty bits, and it seems to be required to apply error correction to convert it to actual intact data (the "raw" stuff consists of 1E8h-byte data snippets, usually followed by 8-byte ECC info, or sometimes 16-byte or 20-byte ECC, or no ECC at all). There are several "backup" copies of the blowfish/firmware blocks.

Write Example

  B2aaaaaaaa000000h - Select 128Kbyte RW access mode (unlesss already)
  8500000000000000h - NAND Write Enable
  81aaaaaaaa000000h - NAND Write to Write Buffer        + Data[200h]
  81aaaaaaaa000000h - NAND Write to Write Buffer        + Data[200h]
  81aaaaaaaa000000h - NAND Write to Write Buffer        + Data[200h]
  81aaaaaaaa000000h - NAND Write to Write Buffer        + Data[200h]
  8200000000000000h - NAND Forward Write Buffer to NAND
  D600000000000000h - NAND Read Status                  + Data[4]
  (...repeat reading status until bit5=1=ready...)
  8400000000000000h - NAND Discard SRAM write
  8B00000000000000h - NAND Select ROM access mode (if desired)

Chipsets
Warioware D.I.Y., NTR-UORE:

  PCB "DI X-7 C17-01"
  Chip "SAMSUNG 004, KLC2811ANB-P204, NTR-UORE-0"

Jam with the Band, NTR-UXBP:

  PCB (Unknown)
  Chip "SAMSUNG 013, KLC2811UOC-P30A, NTR-UXBP-0, WKA069J2"

Face Training, TWL-USKV:

  PCB "DI X-8 C17-01"
  U1 "SAMSUNG 031, KLC2811UOC-P309, TWL-USKV-0, WKE114(80?)"
  (this chip must be slightly different, for DSi mode support)

There are also a bunch of 3DS games with similar chips (with the RW area being called "CARD2" area on 3DS, and, the 3DS is using 16-byte commands, so the protocol is different).

Notes
Unlike ROM carts, the NAND cart does crash upon invalid commands (and stops to respond to further commands). That is, upon invalid command numbers, upon ROM mode commands in RW mode (or vice-versa), upon any data lengths other listed above.

DS Cartridge I/O Ports

The Gamecard bus registers can be mapped to NDS7 or NDS9 via EXMEMCNT, see
DS Memory Control

40001A0h - NDS7/NDS9 - AUXSPICNT - Gamecard ROM and SPI Control (R/W)

  0-1   SPI Baudrate        (0=4MHz/Default, 1=2MHz, 2=1MHz, 3=512KHz)
  2-5   Not used            (always zero)
  6     SPI Hold Chipselect (0=Deselect after transfer, 1=Keep selected)
  7     SPI Busy            (0=Ready, 1=Busy) (presumably Read-only)
  8-12  Not used            (always zero)
  13    NDS Slot Mode       (0=Parallel/ROM, 1=Serial/SPI-Backup)
  14    Transfer Ready IRQ  (0=Disable, 1=Enable) (for ROM, not for AUXSPI)
  15    NDS Slot Enable     (0=Disable, 1=Enable) (for both ROM and AUXSPI)

The "Hold" flag should be cleared BEFORE transferring the LAST data unit, the chipselect will be then automatically cleared after the transfer, the program should issue a WaitByLoop(12) on NDS7 (or longer on NDS9) manually AFTER the LAST transfer.

40001A2h - NDS7/NDS9 - AUXSPIDATA - Gamecard SPI Bus Data/Strobe (R/W)
The SPI transfer is started on writing to this register, so one must <write> a dummy value (should be zero) even when intending to <read> from SPI bus.

  0-7  Data
  8-15 Not used (always zero)

During transfer, the Busy flag in AUXSPICNT is set, and the written DATA value is transferred to the device (via output line), simultaneously data is received (via input line). Upon transfer completion, the Busy flag goes off, and the received value can be then read from AUXSPIDATA, if desired.

40001A4h - NDS7/NDS9 - ROMCTRL - Gamecard Bus ROMCTRL (R/W)

  0-12  KEY1 gap1 length  (0-1FFFh) (forced min 08F8h by BIOS) (leading gap)
  13    KEY2 encrypt data (0=Disable, 1=Enable KEY2 Encryption for Data)
  14     "SE" Unknown? (usually same as Bit13) (does NOT affect timing?)
  15    KEY2 Apply Seed   (0=No change, 1=Apply Encryption Seed) (Write only)
  16-21 KEY1 gap2 length  (0-3Fh)   (forced min 18h by BIOS) (200h-byte gap)
  22    KEY2 encrypt cmd  (0=Disable, 1=Enable KEY2 Encryption for Commands)
  23    Data-Word Status  (0=Busy, 1=Ready/DRQ) (Read-only)
  24-26 Data Block size   (0=None, 1..6=100h SHL (1..6) bytes, 7=4 bytes)
  27    Transfer CLK rate (0=6.7MHz=33.51MHz/5, 1=4.2MHz=33.51MHz/8)
  28    KEY1 Gap CLKs (0=Hold CLK High during gaps, 1=Output Dummy CLK Pulses)
  29    RESB Release Reset  (0=Reset, 1=Release) (cannot be cleared once set)
  30    Data Direction "WR" (0=Normal/read, 1=Write, for FLASH/NAND carts)
  31    Block Start/Status  (0=Ready, 1=Start/Busy) (IRQ See 40001A0h/Bit14)

The cartridge header is booted at 4.2MHz CLK rate, and following transfers are then using ROMCTRL settings specified in cartridge header entries [060h] and [064h], which are usually using 6.7MHz CLK rate for the main data transfer phase (whereof, older MROM carts can actually transfer 6.7Mbyte/s, but newer 1T-ROM carts default to reading 200h-byte blocks with gap1=657h, thus reaching only 1.6Mbyte/s).
Transfer length of null, four, and 200h..4000h bytes are supported by the console, however, retail cartridges cannot cross 1000h-byte boundaries (and, SanDisk ROM chips and Samsung NAND chips cannot transfer more than 200h bytes).
Default cart header entries

  hdr[60h]   hdr[64h]   hdr[6Eh]
  00586000h  001808F8h  051Eh     ;older/faster MROM
  00416657h  081808F8h  0D7Eh     ;newer/slower 1T-ROM
  00416657h  081808F8h  0D7Eh     ;newer/slower NAND

Older/Faster MROM
The romctrl values in cartheader[60h,64h] are okay, but the secure delay in [6Eh] is nonsense (should be zero).
Misdeclared MROM
Some carts like SystemFlaw and BiggestLoser are actually containing MROM chips despite of having 1T-ROM values in cart header (gap1=657h is making loading insane slow, gap2=01h causes errors on 1000h-byte blocks, and secure.clk=4.2MHz is slowing down secure area loading, combined with even slower secure area delay despite of not needing any delay for MROM).
As the cart header entries are wrong, some other detection is needed: This can be probably done by checking ChipID.bit31 (or otherwise by testing if 1000h-block reading works with gap1=01h, if so, then it's 1T-ROM).
Newer/Slower 1T-ROM
Actual 1T-ROM carts can be very slow, especially when using the insane cart header values and default firmware blocksize of 200h bytes which drops loading speed from 6.7Mbytes/s to 1.6Mbyte/s (as workaround, use gap1=180h, blocksize=1000h, also secure area delay should be 400h, not D7Eh) (tested/working for CookingCoach, unknown if that timings work for all other carts).
Superslow Writeable NAND
This is having the same official insane delays as 1T-ROM, but with less tolerance for faster timings (fastest working values are gap1=380h, blocksize=200h, secure delay 400h, plus 20ms delay after releasing reset).
Cart Reset
Reset flag in bit29 can be set once only (to release reset), the only way to clear the bit is power-off. However, there are some ways to issue resets:
1) On NDS: Manually eject/insert the cart (that won't affect bit29, but the cart will reset itself anyways upon power loss) (eject on DSi will power-off the cart slot).
2) If one of the two ROMCTRL registers (on ARM7 and ARM9) is still zero: Temporarily toggle ARM7/ARM9 cart access via EXMEMCNT on ARM9 side.
3) On DSi: If the 2nd cart slot ROMCTRL register (40021A4h) is still zero: Temporarily swap 1ns/2nd cart slot via SCFG_MC.bit15 on ARM7 side.
4) On DSi: Use SCFG_MC to toggle cart power off/on; this will actually reset bit29, the DSi firmware is actually using that method, but it's very slow (takes about 300ms, for the power-off wait, plus (unneccassary) hardcoded power-on delays).

40001A8h - NDS7/NDS9 - Gamecard bus 8-byte Command Out
The separate commands are described in the Cartridge Protocol chapter, however, once when the BIOS boot procedure has completed, one would usually only need command "B7aaaaaaaa000000h", for reading data (usually 200h bytes) from address aaaaaaaah (which should be usually aligned by 200h).

  0-7   1st Command Byte (at 40001A8h) (eg. B7h) (MSB)
  8-15  2nd Command Byte (at 40001A9h) (eg. addr bit 24-31)
  16-23 3rd Command Byte (at 40001AAh) (eg. addr bit 16-23)
  24-31 4th Command Byte (at 40001ABh) (eg. addr bit 8-15) (when aligned=even)
  32-39 5th Command Byte (at 40001ACh) (eg. addr bit 0-7)  (when aligned=00h)
  40-47 6th Command Byte (at 40001ADh) (eg. 00h)
  48-57 7th Command Byte (at 40001AEh) (eg. 00h)
  56-63 8th Command Byte (at 40001AFh) (eg. 00h) (LSB)

Observe that the command/parameter MSB is located at the smallest memory location (40001A8h), ie. compared with the CPU, the byte-order is reversed.

4100010h - NDS7/NDS9 - Gamecard bus 4-byte Data In (R) (or W)

  0-7   1st received Data Byte (at 4100010h)
  8-15  2nd received Data Byte (at 4100011h)
  16-23 3rd received Data Byte (at 4100012h)
  24-31 4th received Data Byte (at 4100013h)

After sending a command, data can be read from this register manually (when the DRQ bit is set), or by DMA (with DMASAD=4100010h, Fixed Source Address, Length=1, Size=32bit, Repeat=On, Mode=DS Gamecard).

40001B0h - 32bit - NDS7/NDS9 - Encryption Seed 0 Lower 32bit (W)
40001B4h - 32bit - NDS7/NDS9 - Encryption Seed 1 Lower 32bit (W)
40001B8h - 16bit - NDS7/NDS9 - Encryption Seed 0 Upper 7bit (bit7-15 unused)
40001BAh - 16bit - NDS7/NDS9 - Encryption Seed 1 Upper 7bit (bit7-15 unused)
These registers are used by the NDS7 BIOS to initialize KEY2 encryption (and there's normally no need to change that initial settings). Writes to the Seed registers do not have direct effect on the internal encryption registers, until the Seed gets applied by writing "1" to ROMCTRL.Bit15.

 For more info:

DS Encryption by Random Seed (KEY2)
Note: There are <separate> Seed registers for both NDS7 and NDS9, which can be applied by ROMCTRL on NDS7 and NDS9 respectively (however, once when applied to the internal registers, the new internal setting is used for <both> CPUs).

DS Cartridge NitroROM and NitroARC File Systems

The DS hardware, BIOS, and Firmware do NOT contain any built-in filesystem functions. The ARM9/ARM7 boot code (together max 3903KB), and Icon/Title information are automatically loaded on power-up.
Programs that require to load additional data from cartridge ROM may do that either by implementing whatever functions to translate filenames to ROM addresses, or by reading from ROM directly.

NitroROM
The NitroROM Filesystem is used by many NDS games (at least those that have been developed with Nintendo's tools). It's used for ROM Cartridges, and, on the DSi, it's also used for DSiWare games (in the latter case, NitroROM acts as a 2nd virtual filesystem inside of the DSi's FAT16 filesystem).

  FNT = cart_hdr[040h]     ;\origin as defined in ROM cartridge header
  FAT = cart_hdr[048h]     ;/
  IMG = 00000000h          ;-origin at begin of ROM

Aside from using filenames, NitroROM files can be alternately accessed via Overlay IDs (see later on below).

NitroARC (Nitro Archive)
NARC Files are often found inside of NitroROM Filesystems (ie. NARC is a second virtual filesystem, nested inside of the actual filesystem). The NARC Format is very similar to the NitroROM Format, but with additional Chunk Headers (instead of the Cartridge ROM Header).

  ...  ...  Optional Header (eg. compression header, or RSA signature)
  000h 4    Chunk Name "NARC" (Nitro Archive)                   ;\
  004h 2    Byte Order (FFFEh) (unlike usually, not FEFFh)      ;
  006h 2    Version (0100h)                                     ; NARC
  008h 4    File Size (from "NARC" ID to end of file)           ; Header
  00Ch 2    Chunk Size (0010h)                                  ;
  00Eh 2    Number of following chunks (0003h)                  ;/
  010h 4    Chunk Name "BTAF" (File Allocation Table Block)     ;\
  014h 4    Chunk Size (including above chunk name)             ; File
  018h 2    Number of Files                                     ; Allocation
  01Ah 2    Reserved (0000h)                                    ; Table
  01Ch ...  FAT (see below)                                     ;/
  ...  4    Chunk Name "BTNF" (File Name Table Block)           ;\
  ...  4    Chunk Size (including above chunk name)             ; File Name
  ...  ...  FNT (see below)                                     ; Table
  ...  ..   Padding for 4-byte alignment (FFh-filled, if any)   ;/
  ...  4    Chunk Name "GMIF" (File Image Block)                ;\
  ...  4    Chunk Size (including above chunk name)             ; File Data
  ...  ...  IMG (File Data)                                     ;/

NARCless variant
There are a few NARC archives with crippled header, without "NARC" string (eg. rom:\dwc\utility.bin in Over the Hedge, Downhill Jam, and Tony Hawk's Skateland).

  000h 4   FNT Filename Table Offset (always at 10h)
  004h 4   FNT Filename Table Size
  008h 4   FAT Allocaton Table Offset (at above Offset+Size+Padding)
  00Ch 4   FAT Allocaton Table Size
  010h ..  FNT Filename Table Data
  ...  ..  FAT Allocaton Table Data
  ...  ..  IMG File Data

The offsets in FAT are relative to IMG=0 (as if IMG would start at begin of file).

File Allocation Table (FAT) (base/size defined in cart header)
Contains ROM addresses for up to 61440 files (File IDs 0000h and up).

  Addr Size Expl.
  00h  4    Start address (originated at IMG base) (0=Unused Entry)
  04h  4    End address   (Start+Len)              (0=Unused Entry)

For NitroROM, addresses must be after Secure Area (at 8000h and up).
For NitroARC, addresses can be anywhere in the IMG area (at 0 and up).
Directories are fully defined in FNT area, and do not require FAT entries.

File Name Table (FNT) (base/size defined in cart header)
Consists of the FNT Directory Table, followed by one or more FNT Sub-Tables.
To interprete the directory tree: Start at the 1st Main-Table entry, which is referencing to a Sub-Table, any directories in the Sub-Table are referencing to Main-Table entries, which are referencing to further Sub-Tables, and so on.

FNT Directory Main-Table (base=FNT+0, size=[FNT+06h]*8)
Consists of a list of up to 4096 directories (Directory IDs F000h and up).

  Addr Size Expl.
  00h  4    Offset to Sub-table             (originated at FNT base)
  04h  2    ID of first file in Sub-table   (0000h..EFFFh)

For first entry (ID F000h, root directory):

  06h  2    Total Number of directories     (1..4096)

Further entries (ID F001h..FFFFh, sub-directories):

  06h  2    ID of parent directory (F000h..FFFEh)

FNT Sub-tables (base=FNT+offset, ends at Type/Length=00h)
Contains ASCII names for all files and sub-directories within a directory.

  Addr Size Expl.
  00h  1    Type/Length
              01h..7Fh File Entry          (Length=1..127, without ID field)
              81h..FFh Sub-Directory Entry (Length=1..127, plus ID field)
              00h      End of Sub-Table
              80h      Reserved
  01h  LEN  File or Sub-Directory Name, case-sensitive, without any ending
              zero, ASCII 20h..7Eh, except for characters \/?"<>*:;|

Below for Sub-Directory Entries only:

  LEN+1 2    Sub-Directory ID (F001h..FFFFh) ;see FNT+(ID AND FFFh)*8

File Entries do not have above ID field. Instead, File IDs are assigned in incrementing order (starting at the "First ID" value specified in the Directory Table).

ARM9 and ARM7 Overlay Tables (OVT) (base/size defined in cart header)
Somehow related to Nintendo's compiler, allows to assign compiler Overlay IDs to filesystem File IDs, and to define additional information such like load addresses.

  Addr Size Expl.
  00h  4    Overlay ID
  04h  4    RAM Address ;Point at which to load
  08h  4    RAM Size    ;Amount to load
  0Ch  4    BSS Size    ;Size of BSS data region
  10h  4    Static initialiser start address
  14h  4    Static initialiser end address
  18h  4    File ID  (0000h..EFFFh)
  1Ch  4    Reserved (zero)

Cartridge Header
The base/size of FAT, FNT, OVT areas is defined in cartridge header,
DS Cartridge Header

DS Cartridge Unknown Commands

The Main Data transfer mode is normally using only two commands (B7h/B8h), however most cartridges do support one or more undocumented commands (as opposed to invalid commands, which will cause the cart to stop responding).

  Title                    Chip ID   Commands...
  Metroid First Hunt       00000FC2  B7 B8 D8
  Meine Tierarztpraxis     00000FAE  B7 B8 D8
  Meine Tierpension        00000FC2  B7 B8 D8
  Nanostray                00000FC2  B7 B8 D8
  Over the Hedge           00001FC2  B7 B8 D8
  Tony Hawk's Skateland    00003FC2  B7 B8
  Tony Hawk's Downhill Jam 00003FC2  B7 B8
  Ultimate Spiderman       00003FC2  B7 B8
  System Flaw (DSi)        40001FC2  B7 B8 F1
  Biggest Loser (DSi)      40001FC2  B7 B8 F1
  Cooking Coach (DSi)      C0007FC2  58..5F 60..68 B7 B8
  Walk with Me             E0013F80  69..6C B5 B7 B8 D6
  Face Training (DSI NAND) E8007FEC  0x 5x 6x 8x 94 Bx D6 (see NAND chapter)

The presence of those commands was tested on DSi (where one can reset the cart by software to recover from invalid commands), with all parameter bits set to zero.
That testing is more difficult on NAND carts because the commands do only work when transferring the correct number of data bytes (0, 4, 200h, or 800h bytes), and only when being in the correct mode (ROM or RW mode; whereof, on DSi carts, the RW mode works ONLY in DSi mode).

 ______________________________ Command 58h..68h ______________________________

5800000000000000h-5F00000000000000h (..) - Get HighZ
Unknown purpose, returns HighZ.

6000000000000000h-6800000000000000h (..) - Get Zeroes
Unknown purpose, returns at least 4000h encrypted 00h bytes.

 ______________________________ Command 69h..6Ch ______________________________

6900000000000000h-6C00000000000000h (..) - Get Zeroes
Unknown purpose, returns at least 4000h encrypted 00h bytes.

 ______________________________ Command B5h/D6h _______________________________

B500000000000000h (0) - Start/Stop/Reset/Ack something?
Unknown. Returns only HighZ.

D600000000000000h (4) - Get Status Byte (also used on NAND carts)
Returns a status byte (repeated in first four bytes, then followed by HighZ bytes).

Command B5h/D6h are supported in carts that have bit5 set in 4th byte of Chip ID (eg. in Walk with Me). Command D6h (and maybe also B5h) works both on power up (before secure area), and in main data phase (after secure area). The DSi Launcher contains following code to deal with such carts:

 if chip_id AND 20000000h
   get_nand_status(cmd_D6h)
   if (nand_status AND 0Ch)<>0                    ;whatever bits
     whatever(cmd_B5h)                            ;whatever command
    loop:
     get_nand_status(cmd_D6h)
     if (nand_status AND 20h)=0 then goto loop    ;wait for ready flag

 ________________________________ Command D8h _________________________________

D800000000000000h (..) - Get Zeroes
Unknown purpose, returns at least 4000h encrypted 00h bytes.

 ________________________________ Command F1h _________________________________

F100000000000000h (1000h) - Get Serial/Manufacturer Info or so
Returns 1000h bytes of data (repeated each 1000h bytes), the data is mostly FFh-filled, with some kind of Serial Number or Manufacturer Info at offset E00h. Unknown what those values are good for, and if they contain unique IDs.
This command works in main data phase only (returns only FFh/HighZ on power up before secure area loading).
System Flaw cmd F1 response:

  0000..0DFF  FF-filled
  0E00        1E 40 05 5A FF FF 0D 01  32 68 38 7A 23 3F FF FF
  0E10        03 0B 00 00 03 09 FF FF  FF FF FF FF FF FF FF FF
  0E20        1E 40 05 03 0B 00 00 03  09 00 00 FF FF FF FF FF
  0E30        FF FF FF FF FF FF FF FF  FF FF FF FF FF FF FF FF
  0E40        FF FF FF FF FF FF FF FF  FF FF FF FF FF FF FF FF
  0E50        FF FF FF FF FF 5A FF 5E  FF FF FF FF FF FF 5A FF
  0E60        FF FF FF FF FF FF FF FF  FF FF FF FF FF FF FF FF
  0E70        FF FF FF FF FF FF FF FF  FF FF FF FF FF FF FF FF
  0E80..0FFF  FF-filled
  1000..3FFF  mirrors of 0000-0FFF

Biggest Loser cmd F1 response:

  0000..0DFF  FF-filled
  0E00        11 16 08 5A FF FF 0D 0B  39 7C 40 8E 2A 53 FF FF
  0E10        03 0A 07 05 05 04 00 00  07 00 7F FF 00 FF FF FF
  0E20        FF FF FF FF FF FF FF FF  FF FF FF FF FF FF FF FF
  0E30        FF FF FF FF FF FF FF FF  FF FF FF FF FF FF FF FF
  0E40        FF FF FF FF FF FF FF FF  FF FF FF FF FF FF FF FF
  0E50        FF FF FF FF FF 5A FF 5E  FF FF FF FF FF FF 5A FF
  0E60        FF FF FF FF FF FF FF FF  FF FF FF FF FF FF FF FF
  0E70        FF FF FF FF FF FF FF FF  FF FF FF FF FF FF FF FF
  0E80..0FFF  FF-filled
  1000..3FFF  mirrors of 0000-0FFF

DS Cartridge PassMe/PassThrough

PassMe is an adapter connected between the DS and an original NDS cartridge, used to boot unencrypted code from a flash cartridge in the GBA slot, it replaces the following entries in the original NDS cartridge header:

  Addr  Siz Patch
  004h  4   E59FF018h  ;opcode LDR PC,[027FFE24h] at 27FFE04h
  01Fh  1   04h        ;set autostart bit
  022h  1   01h        ;set ARM9 rom offset to nn01nnnnh (above secure area)
  024h  4   027FFE04h  ;patch ARM9 entry address to endless loop
  034h  4   080000C0h  ;patch ARM7 entry address in GBA slot
  15Eh  2   nnnnh      ;adjust header crc16

After having verified the encrypted chip IDs (from the original cartridge), the console thinks that it has successfully loaded a NDS cartridge, and then jumps to the (patched) entrypoints.

GBA Flashcard Format
Although the original PassMe requires only the entrypoint, PassMe programs should additionally contain one (or both) of the ID values below, allowing firmware patches to identify & start PassMe games without real PassMe hardware.

  0A0h  GBA-style Title    ("DSBooter")
  0ACh  GBA-style Gamecode ("PASS")
  0C0h  ARM7 Entrypoint    (32bit ARM code)

Of course, that applies only to early homebrew programs, newer games should use normal NDS cartridge headers.

ARM9 Entrypoint
The GBA-slot access rights in the EXMEMCNT register are initially assigned to the ARM7 CPU, so the ARM9 cannot boot from the flashcard, instead it is switched into an endless loop in Main RAM (which contains a copy of the cartridge header at 27FFE00h and up). The ARM7 must thus copy ARM9 code to Main RAM, and then set the ARM9 entry address by writing to [027FFE24h].

DS Cartridge GBA Slot

Aside from the 17-pin NDS slot, the DS also includes a 32-pin GBA slot. This slot is used for GBA backwards compatibility mode. Additionally, in DS mode, it can be as expansion port, or for importing data from GBA games.

  NDS:     Normal 32pin slot
  DS Lite: Short 32pin slot (GBA cards stick out)
  DSi:     N/A (dropped support for GBA carts, and for DS-expansions)

In DS mode, ROM, SRAM, FLASH backup, and whatever peripherals contained in older GBA cartridges can be accessed (almost) identically as in GBA mode,
GBA Cartridges

Addressing
In DS mode, only one ROM-region is present at 8000000h-9FFFFFFh (ie. without the GBA's mirrored WS1 and WS2 regions at A000000h-DFFFFFFh). The expansion region (for SRAM/FLASH/etc) has been moved from E000000h-E00FFFFh (GBA-mode) to A000000h-A00FFFFh (DS-mode).

Timings
GBA timings are specified as "waitstates" (excluding 1 access cycle), NDS timings are specified as (total) "access time". And, the NDS bus-clock is twice as fast as for GBA. So, for "N" GBA waitstates, the NDS access time should be "(N+1)*2". Timings are controlled via NDS EXMEMCNT instead GBA WAITCNT,
DS Memory Control - Cartridges and Main RAM

GBA EEPROM
EEPROMs in GBA carts cannot be accessed in DS mode. The EEPROMs should be accessed with 8 waits on GBA, ie. 18 cycles on NDS on both 1st/2nd access. But, 2nd access is restricted to max 6 cycles in NDS mode, which is ways too fast.

DS Cart Rumble Pak

DS Rumble Option Pak
The Rumble Pak comes bundled with Metroid Prime Pinball. It contains a small actuator made by ALPS to make it rumble. The original device (NTR-008) is sized like a normal GBA cartridge, and there's also shorter variant for the DS-Lite (USG-006).
The rumble pak is pretty simple internally, it only wires up to a few pins on the GBA Cartridge Port:

  VCC, GND, /WR, AD1, and IRQ (grounded)

AD1 runs into a little 8 pin chip, which is probably just a latch on the rising edge of /WR. A line runs from this chip to a transistor that is directly connected to the actuator. The only other chip on the board is a 5 pin jobber, probably a power component.
For detection, AD1 seems to be pulled low when reading from it, the other AD lines are open bus (containing the halfword address), so one can do:

  for i=0 to 0FFFh
    if halfword[8000000h+i*2]<>(i and FFFDh) then <not_a_ds_rumble_pak>
  next i

The actuator doesn't have an on/off setting like a motor, it rumbles when you switch it between the two settings. Switch frequently for a fast rumble, and fairly slowly for more of a 'tick' effect. That should be done via timer irq:

  rumble_state = rumble_state xor 0002h
  halfword[8000000h]=rumble_state

Unknown if one of the two states has higher power-consumption than the other, ie. if it's a "pull/release" mechanism, if so, then disabling rumble should be done by using the "release" state, which would be AD1=0, or AD1=1...?
Note: The v3 firmware can detect the Rumble Pak as an option pak, but it does not provide an enable/disable rumble option in the alarm menu.

Other DS Rumble device
There's also another DS add-on with rumble. That device uses AD8 (instead AD1) to control rumble, and, it's using a classic motor (ie. it's rumbling while and as long as the latched AD8 value is "1").
DS Cart Slider with Rumble

GBA Rumble Carts
There are also a few GBA games that contain built-in Rumble, and which could be used by NDS games as well. To be user friendly, best support both types.
GBA Cart Rumble

DS Cart Slider with Rumble

Add-on device for the japanese title Magukiddo. The optical sensor is attached underneath of the console (connected to the GBA slot).
The sensor is an Agilent ADNS-2030 Low Power Optical Mouse Sensor (16pin DIP chip with built-in optical sensor, and external LED light source) with two-wire serial bus (CLK and DTA).

ADNS-2030 Registers (write 1 byte index, then read/write 1 byte data)
Index (Bit7=Direction; 0=Read, 1=Write):

  00h Product_ID (R) (03h)
  01h Revision_ID (R) (10h=Rev. 1.0) (20h=Used in DS-option-pak)
  02h Motion/Status Flags (R)
  03h Delta_X (R) (signed 8bit) (automatically reset to 00h after reading)
  04h Delta_Y (R) (signed 8bit) (automatically reset to 00h after reading)
  05h SQUAL (R) (surface quality) (unsigned 8bit)
  06h Average_Pixel (R) (unsigned 6bit, upper 2bit unused)
  07h Maximum_Pixel (R) (unsigned 6bit, upper 2bit unused)
  08h Reserved
  09h Reserved
  0Ah Configuration_bits (R/W)
  0Bh Reserved
  0Ch Data_Out_Lower (R)
  0Dh Data_Out_Upper (R)
  0Eh Shutter_Lower (R)
  0Fh Shutter_Upper (R)
  10h Frame_Period_Lower (R/W)
  11h Frame_Period_Upper (R/W)

Motion/Status Flags:

  7 Motion since last report or PD (0=None, 1=Motion occurred)
  6 Reserved
  5 LED Fault detected (0=No fault,  1=Fault detected)
  4 Delta Y Overflow (0=No overflow, 1=Overflow occured)
  3 Delta X Overflow (0=No overflow, 1=Overflow occured)
  2 Reserved
  1 Reserved
  0 Resolution in counts per inch (0=400, 1=800)

Configuration_bits:

  7 Reset Power up defaults (W) (0=No, 1=Reset)
  6 LED Shutter Mode (0=LED always on, 1=LED only on when shutter is open)
  5 Self Test (W) (0=No, 1=Perform all self tests)
  4 Resolution in counts per inch (0=400, 1=800)
  3 Dump 16x16 Pixel bitmap (0=No, 1=Dump via Data_Out ports)
  2 Reserved
  1 Reserved
  0 Sleep Mode (0=Normal/Sleep after 1 second, 1=Always awake)
                          _______
                         |74273  |
  /WR -----------------> |CLK    |                       _____
  AD1/SIO CLK ---------> |D1   Q1|--------------> CLK   |74125|
  AD2 power control ---> |D2   Q2|--->     ____         |     |
  AD3/SIO DIR ---------> |D3   Q3|------o-|7400\________|/EN  |
  AD8 rumble on/off ---> |D?   Q?|--->  '-|____/        |     |
  AD0/SIO DTA ----o----> |D5   Q5|----------------------|A   Y|--o--DTA
                  |      |_______|                      |- - -|  |
          ____    '-------------------------------------|Y   A|--'
  /RD ---|7400\______ ____                              |     |
  /RD ---|____/      |7400\_____________________________|/EN  |
  A19 _______________|____/                             |_____|

7400 Quad NAND Gate, 74273 8bit Latch

AD0 Optical Sensor Serial Data (0=Low, 1=High)
AD1 Optical Sensor Serial Clock (0=Low, 1=High)
AD2 Optical Sensor Power (0=Off, 1=On)
AD3 Optical Sensor Serial Direction (0=Read, 1=Write)
AD8 Rumble Motor (0=Off, 1=On)

Thanks: Daniel Palmer

DS Cart Expansion RAM

DS Memory Expansions
There are several RAM expansions for the NDS. The RAM cartridge connects to the GBA slot; can can be then accessed from cartridges in the DS slot.

  Opera         (8MB RAM) (official RAM expansion for Opera browser)
  EZ3/4/3-in-1  (8-16MB RAM, plus FLASH, plus rumble)
  Supercard     (32MB)
  M3            (32MB)
  G6            (32MB)

The recommended access time (waitstates) for all memory types is unknown. Unknown which programs do use these expansions for which purposes (aside from the Opera browser).
Thanks to Rick "Lick" Wong for info on detection and unlocking.

Opera / DS Memory Expansion Pak (NTR-011 or USG-007)

  base=9000000h, size=800000h (8MB)
  unlock=1, lock=0
  STRH [8240000h],lock/unlock

  base=8400000h, size=VAR (8MB..16MB)
  locking/unlocking/detection see below

Supercard

  base=8000000h, size=1FFFFFEh (32MB minus last two bytes?)
  unlock=5 (RAM_RW), lock=3 (MEDIA)
  STRH [9FFFFFEh],A55Ah
  STRH [9FFFFFEh],A55Ah
  STRH [9FFFFFEh],lock/unlock
  STRH [9FFFFFEh],lock/unlock

  base=8000000h, size=2000000h (32MB)
  unlock=00400006h, lock=00400003h
  LDRH Rd,[8E00002h]
  LDRH Rd,[800000Eh]
  LDRH Rd,[8801FFCh]
  LDRH Rd,[800104Ah]
  LDRH Rd,[8800612h]
  LDRH Rd,[8000000h]
  LDRH Rd,[8801B66h]
  LDRH Rd,[8000000h+(lock/unlock)*2]
  LDRH Rd,[800080Eh]
  LDRH Rd,[8000000h]
  LDRH Rd,[80001E4h]
  LDRH Rd,[80001E4h]
  LDRH Rd,[8000188h]
  LDRH Rd,[8000188h]

  base=8000000h, size=2000000h (32MB)
  unlock=6, lock=3
  LDRH Rd,[9000000h]
  LDRH Rd,[9FFFFE0h]
  LDRH Rd,[9FFFFECh]
  LDRH Rd,[9FFFFECh]
  LDRH Rd,[9FFFFECh]
  LDRH Rd,[9FFFFFCh]
  LDRH Rd,[9FFFFFCh]
  LDRH Rd,[9FFFFFCh]
  LDRH Rd,[9FFFF4Ah]
  LDRH Rd,[9FFFF4Ah]
  LDRH Rd,[9FFFF4Ah]
  LDRH Rd,[9200000h+(lock/unlock)*2]
  LDRH Rd,[9FFFFF0h]
  LDRH Rd,[9FFFFE8h]

Detection
For EZ, detection works as so:

 ez_ram_test:   ;Based on DSLinux Amadeus' detection
  ez_subfunc(9880000h,8000h) ;-SetRompage (OS mode)
  ez_subfunc(9C40000h,1500h) ;-OpenNorWrite
  [08400000h]=1234h          ;\
  if [08400000h]=1234h       ; test writability at 8400000h
    [8000000h]=4321h         ; and non-writability at 8000000h
    if [8000000h]<>4321h     ;
      return true            ;/
  ez_subfunc(9C40000h,D200h) ;CloseNorWrite
  ez_subfunc(9880000h,0160h) ;SetRompage (0160h)
  ez_subfunc(9C40000h,1500h) ;OpenNorWrite
  [8400000h]=1234h           ;\
  if [8400000h]=1234h        ; test writability at 8400000h
    return true              ;/
  return false               ;-failed
 ez_subfunc(addr,data):
  STRH [9FE0000h],D200h
  STRH [8000000h],1500h
  STRH [8020000h],D200h
  STRH [8040000h],1500h
  STRH [addr],data
  STRH [9FC0000h],1500h

For all other types (everything except EZ), simply verify that you can write (when unlocked), and that you can't (when locked).

DS Cart Infrared/Pedometers

NDS/DSi Cartridges with IR port
DS Cart Infrared Cartridge SPI Commands
DS Cart Infrared Cartridge Memory Map

Activity Meter
Pedometer with two-color LED and button. The step counter results can be transferred to NDS via IrDA.
DS Cart Infrared Activity Meter IR Commands
DS Cart Infrared Activity Meter Memory Map

P-Walker
Pedometer with LCD, speaker, and three buttons. There is no intended way to run custom program code (though it can be tweaked to do so via CPU Memory Write command).
DS Cart Infrared P-Walker IR Commands
DS Cart Infrared P-Walker Memory Map
DS Cart Infrared P-Walker Ports LCD Controller
DS Cart Infrared P-Walker Ports Accelerometer BMA150
The purpose is more or less unknown: Apart from communicating with the NDS, the IrDA can be also used communicate with other P-Walkers (maybe for trading/fighting?). The GUI supports Teams, Routes, Events, Items (maybe for some built-in interactice game engine?).

Component Lists
DS Cart Infrared Component Lists

H8/38602 CPU series with H8/300H instruction set
The NDS cartridges and Activity Meter and P-Walker contain Renesas H8/3860X CPUs with H8/300H instruction set and on-chip firmware. In the cartridge it's merely used for forwading IR messages via SPI bus, in the pedometer it's handling step sensors, step counting, EEPROM logging, buttons, LED/LCD, RTC/time, IR messages, etc.
H8/386 SFRs
H8/386 Exception Vectors
H8/300H Operands
H8/300H Opcodes

References

  H8/300H Series Programming Manual (Hitachi, 257 pages)  ;-Opcodes
  H8/38602R Group Hardware Manual (Renesas, 554 pages)    ;-SFR's
  The addition of H8/38606 Group (Renesas, 6 pages)       ;-FLASH/ROM/RAM

For P-Walker:

  BMA150 Triaxial digital acceleration sensor Data sheet (Bosch, 56 pages)
  SSD1850 Advance Information (Solomon System, 56 pages)  ;-LCD driver
  http://dmitry.gr/?r=05.Projects&proj=28.%20pokewalker   ;-Disassembly/Story
  http://forums.nesdev.com/viewtopic.php?f=23&t=21140#p265388  ;-Forum

DS Cart Infrared Cartridge SPI Commands

There are two NDS cart firmware versions with minor differences:

  OLD was used in Walk with Me (maybe also Active Health?)
  NEW was used in the P-Letter game series

The IR-port is accessed via SPI bus commands; that bus is also shared for accessing FLASH/EEPROM memory (via 00h-prefix).

Infrared NDS cartridge SPI command summary

  04h,04h                  Initial dummy in walk with me (bugged read or wrdi?)
  00h,cmd,params[...]      Savedata access
  01h,00h,00h              Infrared RX (none, len=0, plus dummy data=0)
  01h,len,data[len]        Infrared RX (OLD: max 84h bytes, NEW: max B8h bytes)
  02h,data[...]            Infrared TX (OLD: max 84h bytes, NEW: max B8h bytes)
  02h,F2h,data[...]        OLD: ignored (refuses to TX data starting with F2h)
  03h,msb,lsb,data         Memory Write 8bit   ;\MOV.B
  04h,msb,lsb,data         Memory Read 8bit    ;/
  05h,msb,lsb,data,data    Memory Write 16bit  ;\MOV.W (fails on 8bit SFRs?)
  06h,msb,lsb,data,data    Memory Read 16bit   ;/
  07h,00h,num,num,num,...  Blah, returns num params from previous spi command
  08h..FFh                 OLD: Ignored (keeps awaiting a SPI command byte)
  08h,ver                  NEW: Returns version (ver=AAh)
  09h..FFh                 OLD: Ignored (returns zeropadding)

SPI Transfers
The SPI transfers are working at max 1MHz transfer clock, and they do require a delay after each byte:
Waiting 800h clks at 33MHz seems to work okay (eg. MOV r0,200h // SWI 03h on ARM7). The NEW ROM version disables IR polling when doing the SPI transfers for RX/TX data blocks (so it may work with shorter delays between the data[...] bytes).
The savedata access is directly passed to the FLASH/EEPROM chip and does work at 4MHz without delays (except, the leading 00h prefix must be transferred at 1MHz plus delay, and another delay is needed when releasing chipselect after last byte).
Note: The NDS cart slot IRQ pin is GNDed in Walk with Me (ie. there is no IRQ for SPI/IR status).
SPI access does require cart power on and reset (via DSi SCFG registers), but doesn't require any ROM commands like secure area loading. Confusingly, the ROM Chip ID seems to have an IR flag in bit0 of 3rd byte (although the ROM chip isn't wired to IR hardware at all).

IR Transfers
The IR transfers are using a fixed baudrate: 115200 baud, 8n1, one-directional (RX gets disabled during TX). The RX/TX commands are transferring "packets" (with each "packet" being terminated by a "pause" in the IR transmission; that "packet+pause" mechanism is making it impossible to use streaming for transferring larger blocks that exceed the buffer size of max 84h or B8h bytes).
The RX command will return empty data with len=00h until a WHOLE packet has been received via IR.
The TX command won't start the IR transfer until the WHOLE packet has been written via SPI, with packet end indicated by releasing chip select.
There is no way to detect TX transfer end (other than computing the expected tranfer time and using an ARM timer).
However, Nintendo is sending a Reply for most TX commands, so one can simply wait for RX packets to determine TX completion (if neither Reply nor Checksum Error are received then one will still need timeout handling).

Memory Read/Write
These commands are normally not used. The memory commands do forcefully abort IR transfers, so they can't be used for polling IR transfer status.
However, they can be used for dumping the firmware ROM, and they could be used to upload/execute custom code in RAM, which may allow to overcome some of the above IR transfer restrictions (other baudrates, fewer delays, better streaming, not ignoring byte F2h, etc).

IR Cart Detection
NDS/DSi Carts with IR support are having special game code with "I" in first letter (NTR-Ixxx or TWL-Ixxx). There are reportedly pirate/bootleg versions of the P-Letter games without IR hardware, unknown if they do nethertheless have the "I" in the gamecode.
Emulators can detect the leading 00h prefix for Savedata access, although that detection may go wrong if preceeded by IR access. Emulators can additionally detect the slow 1MHz SPI clock used for IR access (and for 00h prefix).

DS Cart Infrared Cartridge Memory Map

IR Cartridge SFR Usage

  [0FFD6h].0  Port 3 Data bit0 OUT  IrDA PWDOWN (1=disable IrDA RX)
  [0FFD6h].1  Port 3 Data bit1 IN   IrDA RXD  ;\via serial IrDA registers
  [0FFD6h].2  Port 3 Data bit2 OUT  IrDA TXD  ;/
  [0FFDBh].3  Port 8 Data bit3 OUT  Savedata chipselect (0=select) (cmd 00h)
  [0FFDBh].2  Port 8 Data bit2 OUT  LED color    ;\used in UNUSED functions,
  [0FFDBh].3  Port 8 Data bit3 OUT  LED color    ; in OLD ROM only, and
  [0FFDEh].0  Port B Data bit0 IN   Button input ;/conflicting with Savedata
  IrDA  IR Transfers
  SPI   NDS Console (and cmd 00h forwarding to Savedata)

IR Cartridge (OLD Version)

  FB80h 200h  undocumented and unused RAM, is R/W in my 38600R (!)
  FD80h 2     unused           ;-unused
  FD82h 2     ir_callback      ;\main callbacks for ir/spi polling
  FD84h 2     spi_callback     ;/
  FD86h 2     ir_timestamp     ;-last ir access (for timeout)?
  FD88h 2     spi_timestamp    ;-last spi access (for debug or so)?
  FD8Ah 1     initial_blah     ;-initial state of Port 8.bit3 (not really used)
  FD8Bh 1     ir_rxbuf_wrptr   ;-ir_rxbuf_wrptr (for incoming IR data)?
  FD8Ch 1     ir_rxbuf_rdptr   ;-ir_rxbuf_rdptr (for forwarding to spi)?
  FD8Dh 84h   spi_rx_buf       ;-spi_rx_buf  ;(also ir TX buf)
  FE11h 84h   infrared_rx_buf  ;-infrared_rx_buf
  FE95h 1     spi_index        ;-spi_index
  FE96h 1     ir_tx_index      ;-ir_tx_index   (from spi buf to TX infrared)
  FE97h 1     ir_timeout_flag  ;-ir_timeout_flag   (or packet end or so?)
  FE98h 2     button_num_changes   ;\
  FE9Ah 2     button_num_pushes    ; used only in
  FE9Ch 1     button_new_state     ; UNUSED functions
  FE9Dh 1     button_old_state     ;
  FE9Eh 1     button_newly_pushed  ;
  FE9Fh 1     button_offhold       ;/
  FEA0h E0h   stack_area (stacktop at FF80h)

IR Cartridge (NEW Version)

  FB80h 200h  undocumented and unused RAM, is R/W in my 38600R (!)
  FD80h 2     unused           ;-unused
  FD82h 2     ir_callback      ;\main callbacks for ir/spi polling
  FD84h 2     spi_callback     ;/
  FD86h 2     ir_timestamp     ;-last ir access (for timeout)?
  FD88h 1     ir_rxbuf_wrptr   ;-ir_rxbuf_wrptr (for incoming IR data)?
  FD89h 1     ir_rxbuf_rdptr   ;-blah, always set to 0, never used
  FD8Ah 1     spi_index        ;-spi_index
  FD8Bh 1     ir_tx_index      ;-ir_tx_index   (from spi buf to TX infrared)
  FD8Ch B8h   spi_rx_buf       ;-spi_rx_buf  ;(also ir TX buf)
  FE44h B8h+1 infrared_rx_buf  ;-infrared_rx_buf (plus space for appending 00h)
  FEFDh 1     ir_timeout_flag  ;-ir_timeout_flag   (or packet end or so?)
  FEFEh 82h   stack_area (stacktop at FF80h)

DS Cart Infrared Activity Meter IR Commands

Packet Encryption/Checksumming
Nintendo wants all IR packet bytes to be "encrypted" (XORed by AAh), that encryption/decryption must be done on ARM side. The checksums are calculated as so (on decrypted packets):

  sum=0, packet[2,3]=00h,00h                    ;-initial chksum
  for i=0 to size-1
    if (i and 1)=0 then sum=sum+packet[i]*100h  ;\add in big-endian fashion
    if (i and 1)=1 then sum=sum+packet[i]       ;/
  sum=(sum/10000h)+(sum AND FFFFh)              ;\final adjust
  sum=(sum/10000h)+(sum)                        ;/
  packet[2,3]=sum,sum/100h                      ;-store in little-endian

The packets are transferred at 115200 baud, 8n1. End of Packet is indicated by a pause in the IR transmission (that does also indicate the packet size).
Before sending a command packet, one should always wait for incoming data from the Activity Meter (ie. for the FCh byte, or for Reply/ChecksumError responses for previous command).

Activity Meter IR commands (from NDS):

  08,xx,cc,cc,msb,lsb,data[..]  CPU Memory Write    (len=3Eh max)  ;Reply=08
  0A,xx,cc,cc,msb,lsb,len       CPU Memory Read     (len=40h max)  ;Reply=0A
  0A,xx,cc,cc,FB,9C,len         CPU Memory Read FB9Ch with ClrFlag ;Reply=0A
  20,xx,cc,cc,msb,lsb,data[..]  Serial EEPROM Write (len=3Eh max)  ;Reply=20
  22,xx,cc,cc,msb,lsb,len       Serial EEPROM Read  (len=40h max)  ;Reply=22
  24,00,cc,cc,ss,ss,ss,ss       Update Ringbuf_mm    ;\            ;Reply=24
  24,01,cc,cc,ss,ss,ss,ss       Update Ringbuf_hh    ; and set     ;Reply=24
  24,02,cc,cc,ss,ss,ss,ss       Update Ringbuf_dd    ; 32bit       ;Reply=24
  24,03,cc,cc,ss,mm,hh          Set RTC hh:mm:ss     ; seconds     ;Reply=24
  24,04,cc,cc,ss,ss,ss,ss       Raw Set ssssssss ?   ;/            ;Reply=24
  24,xx,cc,cc,ss,ss,ss,ss       Invalid (same as 24,04)            ;Reply=24
  26,xx,cc,cc                   Deadlock   ;\both same (maybe      ;Reply=26
  28,xx,cc,cc                   Deadlock   ;/Watchdog/reboot?)     ;Reply=26
  2A,xx,cc,cc,00,nn             Stepback Ringbuf_hh  ;\go back nn  ;Reply=2A
  2A,xx,cc,cc,01,nn             Stepback Ringbuf_mm  ; entries,    ;Reply=2A
  2A,xx,cc,cc,02,nn             Stepback Ringbuf_dd  ;/see [FCDAh] ;Reply=2A
  2A,xx,cc,cc,xx,..             Invalid                            ;Reply=2A
  2C,cs,cc,cc                   Toggle one LED on/off              ;Reply=2C
  F4,xx,cc,cc                   Disconnect                         ;Reply=None
  F6,xx,cc,cc                   Force "Bad Chksum" reply           ;Reply=FC
  FA,xx,cc,cc                   Connect                            ;Reply=F8
  FE,...                        Noise      ;\ignored, noise        ;Reply=None
  FF,...                        Noise      ;/                      ;Reply=None
  xx,xx,cc,cc                   Invalid    ;-ignored, invalid cmd  ;Reply=None
  xx,xx,xx,xx                   Bad Chksum                         ;Reply=FC

Activity Meter IR replies (to NDS):

  08,sq,cc,cc           Reply to Cmd 08 (CPU Memory Write reply)
  0A,sq,cc,cc,data[..]  Reply to Cmd 0A (CPU Memory Read reply)
  20,sq,cc,cc           Reply to Cmd 20 (Serial EEPROM Write reply)
  22,sq,cc,cc,data[..]  Reply to Cmd 22 (Serial EEPROM Read reply)
  26,xx,cc,cc           Reply to Cmd 26 and 28 (Deadlock reply)
  24,xx,cc,cc           Reply to Cmd 24 (Update, or Set RTC time)
  2A,xx,cc,cc           Reply to Cmd 2A (Stepback, with result at [FCDAh])
  2C,cs,cc,cc           Reply to Cmd 2C (LED reply)
  80,FF,cc,cc           Factory Reset and Hardware Test completed (or failed)
  F8,00,cc,cc           Reply to Cmd FA (Connect reply)
  FC,xx,cc,cc           Reply to Cmd's with Bad Chksum (and Cmd F6)
  FC                    Advertising Msg (after pressing button) (single byte)

Notes

  cc,cc        Checksum (LITTLE-ENDIAN)
  msb,lsb      Memory Address (big-endian)
  ss,ss,ss,ss  Seconds since 2001 (big-endian)
  ss,mm,hh     RTC time HH:MM:SS (BCD) (caution: smashes seconds since 2001)
  sq           Increasing sequence number in Memory Access replies
  cs           LED color/state (c=color red/green, s=state on/off)
  xx           Whatever (don't care?)

There aren't any specific commands for reading things like step counters, one must instead use the Memory Read/Write commands with hardcoded RAM or EEPROM address, see:
DS Cart Infrared Activity Meter Memory Map
The command/reply are intended to be transferred from/to NDS accordingly, but things could go wrong if there are multiple consoles or activity meters (all trying to process the same message, or even mistreating replies as commands).
The infrared range (distance/angle) is unknown. Dumping the whole 64K CPU memory space worked without checksum errors at about 5-10cm distance (and that worked without even using the Connect command).

DS Cart Infrared Activity Meter Memory Map

The most important RAM locations are FCE8h=Total steps, and FB9Ch=Unique ID (when using multiple Activity Meters), FCF0h=Daily Goal (to change LED color after N steps). Nintendo is reading/writing a few more RAM locations. And, there are ring buffers with steps per minute/hour/day in EEPROM.

Activity Meter SFR Usage

  [0FFD4h].0  Port 1 Data bit0 IN   Factory Test (0=Test, 1=Normal)
  [0FFD4h].2  Port 1 Data bit2 OUT  Set for sum of eight A/D conversions
  [0FFD6h].0  Port 3 Data bit0 OUT  IrDA PWDOWN (1=disable IrDA RX)
  [0FFD6h].1  Port 3 Data bit1 IN   IrDA RXD  ;\via serial IrDA registers
  [0FFD6h].2  Port 3 Data bit2 OUT  IrDA TXD  ;/
  [0FFDBh].2  Port 8 Data bit2 OUT  LED color?
  [0FFDBh].3  Port 8 Data bit3 OUT  LED color?
  [0FFDCh].0  Port 9 Data bit0 OUT  SPI EEPROM chipselect (0=select)
  [0FFDEh].0  Port B Data bit0 IN   Button input
  IrDA  IR Transfers
  SPI   SPI 8Kbyte EEPROM
  A/D   Used to read two single-axis sensors (for step counting)?
  A/D   Also used to read sum of eight A/D conversions (for wakeup from sleep)?

The SPI EEPROM uses same commands as NDS cart savedata:
DS Cartridge Backup

Activity Meter RAM (1Kbyte, FB80h-FF7Fh)

  FB80h 1   Button flags (bit7=curr.state, bit6=newly.pressed, bit5=old.state)
  FB81h 1   ... cleared if memread src was unique_id (and other cases)
  FB82h 1   ... sys/power mode ?
  FB83h 1   ... adc_mode, or power_saving?
  FB84h 1   ... clock change request
  FB85h 1   ... led_extra_mask (never CLEARED, except on boot, or maybe via IR)
  FB86h 1   adc_array_index (index in ADC array X/Y, wraps in range 00h..3Fh)
  FB87h 1   ... entrysize of current data in ringbuf (per newest TAG) or so?
  FB88h 1   SPI overrun error (probably nonsense, SPI clk can't outrun itself)
  FB89h 1   Unused
  FB8Ah 1   num_steps_curr_minute (00h..FCh) (no conflict with tag FDh,FEh,FFh)
  FB8Bh 1   rtc_event_flags (bit0=minute, bit1=hour, bit2=day, bit3=also.hour)
  FB8Ch 1   ... timing offhold for various stuff
  FB8Dh 1   some_shift_amount         ;READ via IR
  FB8Eh 1   Daily goal reached flag (aka LED color) (bit0=reached, bit1=???)
  FB8Fh 1   ... timing for LED step pulses?
  FB90h 1   ... timing for LED step pulses?
  FB91h 1   ... flag for LED step pulse state?
  FB92h 1   Hour when new day starts (BCD, usually/always 03h)   ;READ via IR
  FB93h 1   ... some flag for inactivity low-power mode ?
  FB94h 1   LED animation number (1..5, or 0=none) (factory test result)
  FB95h 1   Unused
  FB96h 1   New day flag
  FB97h 1   Fixed LED mask (this is a "fixed" setting from EEPROM)
  FB98h 1   Compare_ctrl_0  ;\for "Compare Control" HW registers (89h,89h)
  FB99h 1   Compare_ctrl_1  ;/
  FB9Ah 1   New Goal flag (apply [FCF4h] as new goal, starting on next day?)
  FB9Bh 1   Unused
  FB9Ch 28h Unique ID ;READ via IR (initally set by NDS via RAM+EEPROM writes?)
  FBC4h 2   adc_current_x
  FBC6h 2   adc_current_y
  FBC8h 80h adc_array_x (40h x 16bit)
  FC48h 80h adc_array_y (40h x 16bit)
  FCC8h 2   adc_scale_factor_x  ;\scale factors
  FCCAh 2   adc_scale_factor_y  ;/
  FCCCh 2   adc_scale_unused_z  ;\semi-unused (written, but never read)
  FCCEh 2   adc_scale_unused_t  ;/
  FCD0h 2   ringbuf_mm_index (0020h..16A0h)   ;READ via IR
  FCD2h 2   ringbuf_hh_index (16A1h..1C42h)   ;READ via IR
  FBD4h 2   Unused
  FCD6h 2   ringbuf_dd_index (1C43h..1CDEh)   ;READ via IR
  FCD8h 2   num_steps_curr_hour    (16bit step counter for current hour)
  FCDAh 2   ringbuf_stepback_index (result from cmd_2Ah, to be read by cmd_0Ah)
  FCDCh 2   adc_inactivity_timer   (time since last pedometer step)
  FCDEh 2   SPI overrun error counter (related to flag at FB88h)
  FCE0h 2   Unused
  FCE2h 2   adc_current_sum (sum of eight A/D conversions)
  FCE4h 4   seconds_counter (seconds since 1st Jan 2001?, initially 0D2B0B80h)
  FCE8h 4   num_steps_lifelong (lifelong TOTAL steps)
  FCECh 4   num_steps_today    (step counter, for current day)
  FCF0h 4   Daily_goal         (WRITTEN via IR, NDS cart default=3000 decimal)
  FCF4h 4   new_goal_steps     (somewhat reload value for daily goal?)
  FCF8h 18h Unused
  FD10h 2   main_callback (main_adc_button_callback, or ir_callback)
  FD12h 40h ir_tx_data    (buffer for Memory & EEPROM reads)
  FD52h 2   clk_callback  (clk_whatever_callback, or 0=none)
  FD54h 2   ir_callback   (ir_active_callback, or ir_dummy_callback)
  FD56h 2   ir_timestamp_last_byte (for sensing SHORT GAPs, aka end-of-packet)
  FD58h 2   RX chksum from hdr[2..3]
  FD5Ah 2   RX chksum from calculation
  FD5Ch 2   ir_timestamp_last_xfer (for sensing LONG GAPs, aka sleep mode)
  FD5Eh 1   Unused
  FD5Fh 1   ir_rx_len
  FD60h 1   ... semi-unused (set to 00h?) (but never read)
  FD61h 44h ir_rxtx_buf, hdr[4]+data[40h]
  FDA5h 1   ir_tx_hdr_len    ;\memorized TX len+hdr[4]
  FDA6h 4   ir_tx_hdr_copy   ;/(never actually used)
  FDAAh 1   bad_chksum_count, give up sending bad_chksum replies after 3 errors
  FDABh 1   bad_chksum_flag, request reply_FCh (bad_chksum)
  FDACh 80h ... array (40h x 16bit)          ;\
  FE2Ch 80h ... array (40h x 16bit)          ; analog sine/cosine
  FEACh 4   ... dword                        ; stuff for converting
  FEB0h 4   ... dword                        ; adc to step counter?
  FEB4h 1   ... byte                         ;
  FEB5h 1   ... byte                         ;/
  FEB6h 2   Incremented in main_adc_button_callback (but not used elsewhere)
  FEB8h 1   Unused                   ;\maybe meant to be 4-byte tx hdr,
  FEB9h 1   TX sequence number       ; but only hdr[1] used (as increasing
  FEBAh 2   Unused                   ;/seq.no for memory read/write replies)
  FEBCh 4   ... array (2 x 16bit)
  FEC0h 4   ... array (2 x 16bit)
  FEC4h 4   ... array (2 x 16bit)
  FEC8h 4   ... array (2 x 16bit)
  FECEh B2h CPU Stack area, initial SP=FF80h

Activity Meter EEPROM (8Kbytes, serial SPI bus)

  EEPROM:0000h 9     ID "nintendo",00h (9 bytes)
  EEPROM:0009h 17h   Unused (FFh-filled)
  EEPROM:0020h 1681h Ringbuf_mm ;steps per MINUTE for 4 days ;(24*60*4-1)*8bit
  EEPROM:16A1h 5A2h  Ringbuf_hh ;steps per HOUR for 30 days  ;(24*30+1)*16bit
  EEPROM:1C43h 9Ch   Ringbuf_dd ;steps per DAY for 52 days   ;(52)*24bit
  EEPROM:1CDFh 1     Unused (FFh) (padding ringbuf's to 20h-byte-boundary)
  EEPROM:1CE0h 200h  Unused (FFh-filled)
  EEPROM:1EE0h 8+1   ADC_scale_values (4x16bit) ;RAM:FCC8h ;\
  EEPROM:1EE9h 2+1   ADC sum_limit              ;RAM:stack ; these EEPROM
  EEPROM:1EECh 3     Unused                                ; settings
  EEPROM:1EEFh 4+1   Num_steps_lifelong         ;RAM:FCE8h ; have 1-byte
  EEPROM:1EF4h 1+1   Fixed LED Mask             ;RAM:FB97h ; checksums
  EEPROM:1EF6h 2     Unused                                ; appended, and
  EEPROM:1EF8h 1+1   Some_shift_amount          ;RAM:FB8Dh ; backups at
  EEPROM:1EFAh 4+1   Daily_goal                 ;RAM:FCF0h ; 1F40h-1F9Fh
  EEPROM:1EFFh 4+1   New_goal_steps             ;RAM:FCF4h ;
  EEPROM:1F04h 28h+1 Unique ID                  ;RAM:FB9Ch ;
  EEPROM:1F2Dh 13h   Unused (00h-filled)                   ;/
  EEPROM:1F40h 60h   Backup copies of above data at 1EE0h..1F3Fh ;-backups
  EEPROM:1FA0h 2     Error code     (initially FFFFh)
  EEPROM:1FA2h 1     Reboot counter (initially 00h or 01h ?)
  EEPROM:1FA3h 5Dh   Unused (FFh-filled)

The MM/HH/DD ring buffers are steps per minute/hour/day accordingly, mixed with special tags:

  00xxh          Zero steps for N minutes (N=max FCh)       ;\in ringbuf_mm
  xxh            N steps per minute (N=01h..FCh)            ;/
  xxxxh          N steps per hour (N=0000h..FFFFh)          ;-in ringbuf_hh
  xxxxxxh        N steps per day (N=000000h..FFFFFFh)       ;-in ringbuf_dd
  FDxxxxxxxxxxh  Timestamp, reversed-BCD-digit-order, seconds since 2001 or so?
  FEh            Newest entry marker?
  FFh            Unused entry marker?

The firmware contains code for searching tags FDh/FEh/FFh forwards and backwards, that works well with BCD values and 8bit counters, but unknown it can work with the 16bit/24bit counters. The BCD digits are stored backwards (eg. 12345 = FD5432100000h). Rinbuf pointers can be read from RAM locations FCD0h, FCD2h, FCD6h, and FCDAh).

DS Cart Infrared P-Walker IR Commands

Commands are usually send from NDS (or from other Walkers)

  EEPROM Commands  (Cmd 02,04,0C,0E,82)     ;\From NDS or Walker
  Connect Commands (Cmd F8,FA,FC)           ;/
  Peer Commands    (Cmd 10...1C)            ;-From Walker
  Unused Commands  (Cmd's with * marking)   ;-From Prototype tests?
  Other Commands   (Cmd's other than above) ;-From NDS

Incoming Commands from NDS (or incoming commands/replies from another Walker)

  00,hi,..,lzss(..)     EEPROM Write [hi00h..hi7Fh] Compressed   ;Reply=04
  80,hi,..,lzss(..)     EEPROM Write [hi80h..hiFFh] Compressed   ;Reply=04
  02,hi,..,data(..)     EEPROM Write [hi00h..hi7Fh] Raw          ;Reply=04
  82,hi,..,data(..)     EEPROM Write [hi80h..hiFFh] Raw          ;Reply=04
  04,xx,..              EEPROM Write Reply                     ;SendMoreCmd(s)
  06,hi,..,lo,data(nn)* CPU Memory Write [hilo+(0..nn-1)]        ;Reply=06
  0A,hi,..,lo,data(nn)* EEPROM Write Random Len [hilo+(0..nn-1)] ;Reply=04
  0C,xx,..,hi,lo,nn     EEPROM Read Request [hilo+(0..nn-1)]     ;Reply=0E
  0E,xx,..,data(nn)     EEPROM Read Reply                      ;SendMoreCmd(s)
  10,xx,..,data(68h)    Peer Step 1 Request                      ;Reply=12
  12,xx,..,data(68h)    Peer Step 1 Reply                      ;SendMoreCmd(s)
  14,xx,..,data(38h)    Peer Step 2 Request                      ;Reply=16
  16,xx,..              Peer Step 2 Reply                ;Reply=16 or None
  1C,xx,..              Peer Refuse                      ;Reply=None+Disconnect
  20,xx,..              Identity Read Request                    ;Reply=22
  24,xx,..            * Ping Request                             ;Reply=26
  2A,xx,..,none?        Unique ID Read Request                   ;Reply=2A
  2C,xx,..,none?      * Unique ID Read Request slightly other    ;Reply=2A
  32,xx,..,data(28h?) * Identity Write Request 1                 ;Reply=34
  36,xx,..            * Connection Error 1                       ;Reply=None
  38,xx,..            * Walk Start Silent                        ;Reply=38
  40,xx,..,data(28h?) * Identity Write Request 2                 ;Reply=42
  44,xx,..            * Connection Error 2                       ;Reply=None
  4E,xx,..              Walk End Request                         ;Reply=50
  52,xx,..,data(28h?)?? Identity Write Request 3                 ;Reply=54
  56,xx,..            * Connection Error 3                       ;Reply=None
  5A,xx,..              Walk Start Nonsilent                     ;Reply=5A
  60,xx,..,data(28h?) * Identity Write Request 4                 ;Reply=62
  64,xx,..            * Connection Error 4                       ;Reply=None
  66,xx,..            * Walk End Request OTHER                   ;Reply=68
  9C,xx,..            * Error Whatever                   ;Reply=9C+Disconnect
  9E,xx,..            * Error Weird Participate          ;Reply=9E+Disconnect
  A0,xx,..            * Weird Participate 1                  ;Reply=A0 or 9E
  A2,xx,..            * Weird Participate 2                  ;Reply=A2 or 9E
  A4,xx,..            * Weird Participate 3                  ;Reply=A4 or 9E
  A6,xx,..            * Weird Participate 4                  ;Reply=A6 or 9E
  A8,xx,..            * Weird Participate 5                  ;Reply=A8 or 9E
  AA,xx,..            * Weird Participate 6                  ;Reply=AA or 9E
  AC,xx,..            * Weird Participate 7                  ;Reply=AC or 9E
  AE,xx,..            * Weird Participate 8                  ;Reply=AE or 9E
  B8,xx,..            * Award Stamp Heart                        ;Reply=D8
  BA,xx,..            * Award Stamp Spade                        ;Reply=DA
  BC,xx,..            * Award Stamp Diamond                      ;Reply=DC
  BE,xx,..            * Award Stamp Club                         ;Reply=DE
  C0,xx,..            * Award Special Map                        ;Reply=C0
  C2,xx,..            * Award Event P-Letter                     ;Reply=C2
  C4,xx,..            * Award Event Item                         ;Reply=C4
  C6,xx,..            * Award Event Route                        ;Reply=C6
  D0,xx,..            * Award All Stamps and Special Map         ;Reply=C0
  D2,xx,..            * Award All Stamps and Event P-Letter      ;Reply=C2
  D4,xx,..            * Award All Stamps and Event Item          ;Reply=C4
  D6,xx,..            * Award All Stamps and Event Route         ;Reply=C6
  D8,xx,..            * Connection Error 5                       ;Reply=None
  F0,xx,..,data(71h) ?? Enroll Data (28h+40h+8+1 bytes)          ;Reply=F0
  F4,xx,..            * Disconnect                       ;Reply=None+Disconnect
  F8,02,..              Connection Reply from Walker             ;SendCmd=1002
  FA,01,..              Connection Request from NDS              ;Reply=F802
  FA,02,..              Connection Request from Walker           ;Reply=F802
  FA,xx,..              Connection Request invalid       ;Reply=None+Disconnect
  FC                    Connection Beacon from Walker            ;SendCmd=FA
  FE,01,..,data(8)    * EEPROM Write [0008h..000Fh]              ;Reply=FE
  xx                  * Ignored (single byte other than FC)      ;Reply=None
  xx,xx,..            * Invalid Cmd                              ;Reply=None
  xx,xx,xxxx          * Ignored (wrong 4-byte ID for Cmd 00-F7)  ;Reply=None
  xx,xx,xxxx          * Bad Checksum (disconnect if too often)   ;Reply=None

Outgoing Replies to NDS (or outgoing commands/replies to another Walker)

  02,hi,..,data(nn)     EEPROM Write ...                         ;Cmd=Peer
  82,hi,..,data(nn)     EEPROM Write ...                         ;Cmd=Peer
  04,hi,..              EEPROM Write Reply                  ;Cmd=00/02/0A/80/82
  06,hi,..            * CPU Memory Write Reply                   ;Cmd=06h
  0C,02,..,hi,lo,nn     EEPROM Read Request  ;Peer, EEPROM Read  ;Cmd=0Eh
  0E,02,..,data(nn)     EEPROM Read Reply                        ;Cmd=0Ch
  10,02,..,data(68h)    Peer Step 1 Request (after Connect Reply);Cmd=F8h
  12,02,..,data(68h)    Peer Step 1 Reply                        ;Cmd=10h
  14,02,..,data(38h)    Peer Step 2 Request                      ;Cmd=0Eh
  16,02,..              Peer Step 2 Reply                        ;Cmd=14h/16h
  1C,02,..              Peer Refuse                              ;Cmd=10h/12h
  22,02,..,data(68h)    Identitiy Read Reply                     ;Cmd=20h
  26,02,..            * Ping Reply                               ;Cmd=24h
  2A,02,..,data(28h)    Unique ID Reply                          ;Cmd=2Ah/2Ch
  34,02,..            * Identitiy Write 1 Reply                  ;Cmd=32h
  38,02,..            * Walk Start silent Reply                  ;Cmd=38h
  42,02,..            * Identitiy Write 2 Reply                  ;Cmd=40h
  50,02,..              Walk End Reply                           ;Cmd=4Eh
  54,02,..           ?? Identitiy Write 3 Reply                  ;Cmd=52h
  5A,02,..              Walk Start nonsilent Reply               ;Cmd=5Ah
  62,02,..            * Identitiy Write 4 Reply                  ;Cmd=60h
  68,02,..            * Walk End OTHER Reply                     ;Cmd=66h
  9C,02,..            * Weird Whatever Reply-to-Reply?           ;Cmd=9Ch
  9E,02,..,data(11h)  * Weird Participated Reply                 ;Cmd=A0h..AEh
  9E,02,..            * Weird Participated Reply-to-Reply?       ;Cmd=9Eh
  A0,02,..,data(11h)  * Weird Participated Reply 1               ;Cmd=A0h
  A2,02,..,data(11h)  * Weird Participated Reply 2               ;Cmd=A2h
  A4,02,..,data(11h)  * Weird Participated Reply 3               ;Cmd=A4h
  A6,02,..,data(11h)  * Weird Participated Reply 4               ;Cmd=A6h
  A8,02,..,data(11h)  * Weird Participated Reply 5               ;Cmd=A8h
  AA,02,..,data(11h)  * Weird Participated Reply 6               ;Cmd=AAh
  AC,02,..,data(11h)  * Weird Participated Reply 7               ;Cmd=ACh
  AE,02,..,data(11h)  * Weird Participated Reply 8               ;Cmd=AEh
  C0,02,..            * Award Special Map Reply                  ;Cmd=C0h/D0h
  C2,02,..            * Award Event P-Letter Reply               ;Cmd=C2h/D2h
  C4,02,..            * Award Event Item Reply                   ;Cmd=C4h/D4h
  C6,02,..            * Award Event Route Reply                  ;Cmd=C6h/D6h
  C8,02,..            * Award Stamp Heart Reply                  ;Cmd=B8h
  CA,02,..            * Award Stamp Spade Reply                  ;Cmd=BAh
  CC,02,..            * Award Stamp Diamond Reply                ;Cmd=BCh
  CE,02,..            * Award Stamp Club Reply                   ;Cmd=BEh
  F0,02,..,data(28h) ?? Enroll Reply                             ;Cmd=F0h
  F8,02,..              Connect Reply                            ;Cmd=FAh
  FA,02,..              Connect Request from walker              ;Cmd=FCh
  FC                    Connection Beacon                        ;Button?
  FE,02,..            * EEPROM Write [0008h..000Fh] Reply        ;Cmd=FEh
  -?-                   Checksum Error... has no reply? or maybe sends Beacons?

Notes

  ..    short for 16bit Checksum at hdr[2..3] and 32bit Session ID at hdr[4..7]
  xx    somewhat don't care (usually 01h=From NDS, or 02h=From Walker)

All 16bit checksum and the IR "encryption" (XOR by AAh), seem to be same as for Activity Meter (see there, except use initial sum=2, not sum=0).
The Connect Request & Reply commands are sending a "random" SessionID each, all following commands/replies (except F8h and up) must use

  SessionID = ConnectRequestRandomID XOR ConnectReplyRandomID.

The Compressed EEPROM commands are writing decompressed 80h-bytes of data to EEPROM (ie. the compression is only used to speedup the IR transfer, not to save memory). The compression format is Nintendo's standard LZSS (including the header value 10h, and little-endian 24bit length; which should be always 80h). Special case: Compressed writes with exactly 80h bytes of incoming data are treated as uncompressed writes (to be used when compression ratio is worse than no compression).

DS Cart Infrared P-Walker Memory Map

P-Walker SFR Usage

  [0FFD4h].0  Port 1 Data bit0 OUT  SPI LCD chipselect (0=select)
  [0FFD4h].1  Port 1 Data bit1 OUT  SPI LCD access mode (0=Cmd, 1=Data)
  [0FFD4h].2  Port 1 Data bit2 OUT  SPI EEPROM chipselect (0=select)
  [0FFD6h].0  Port 3 Data bit0 OUT  IrDA PWDOWN (1=disable IrDA RX)
  [0FFD6h].1  Port 3 Data bit1 IN   IrDA RXD  ;\via serial IrDA registers
  [0FFD6h].2  Port 3 Data bit2 OUT  IrDA TXD  ;/
  [0FFDBh].2  Port 8 Data bit2      ?
  [0FFDBh].3  Port 8 Data bit3      ?
  [0FFDBh].4  Port 8 Data bit4 OUT  A/D related ... whatfor LCD? accel? batt?
  [0FFDCh].0  Port 9 Data bit0 OUT  SPI Accelerometer chipselect (0=select)
  [0FFDEh].0  Port B Data bit0 IN   ?  ;\
  [0FFDEh].2  Port B Data bit2 IN   ?  ; maybe buttons
  [0FFDEh].4  Port B Data bit4 IN   ?  ;/
  [0FFDEh].5  Port B Data bit5 OUT  ?
  Timer W General A/B/C  Audio Frequency/Volume
  IrDA                   IR Transfers
  SPI                    SPI 64Kbyte EEPROM, LCD Cmd/Data, Accelerometer
  A/D                    whatfor LCD? accel? batt?

The SPI EEPROM uses same commands as NDS cart savedata:
DS Cartridge Backup
For the other two SPI-like devices, see:
DS Cart Infrared P-Walker Ports LCD Controller
DS Cart Infrared P-Walker Ports Accelerometer BMA150
Note: SPI LCD chipselect is also used for outputting factory test results via SPI bus (requires special hardware attached instead of the LCD screen).

RAM Map (2Kbytes, F780h..FF7Fh)

  F780h 60h      Misc variables
  F7E0h 2        main_callback                          ;<--
  F7E2h ECh      Misc variables
  F8CEh 8+80h    Infrared RX/TX buffer hdr+data (also misc/heap)
  F956h 62Ah     Temp buffer, free RAM, and stack       ;<--
  FF80h -        Stacktop (end of RAM)

ROM Map
For some quick hacks, Dmitry recommends these ROM addresses (which won't work when reflashing the firmware).

  0772h          Send IR packet (F8D6h=src, r0l=len, r0h=hdr[0], r1l=hdr[1])
  08D6h          Default callback (when in IR transfer mode)
  259Eh          Watchdog refresh

EEPROM Map
The EEPROM contains some important basic data, plus GUI related bitmaps (mostly text strings pre-rendered as bitmaps for the local user name & game language) (and maybe(?) also game specific customizations).

  0000h 8        ID "nintendo" (set after initial power-up eeprom init)
  0008h 8        ID whatever   (set via Cmd F0h and FEh) (never read)
  0010h 62h      ???
  0072h 1        Number of watchdog resets
  0073h 0Dh      ???
  0080h 02h+1    ADC calibration (factory-provided)               ;\
  0083h 28h+1    Unique ID      (set via Cmd F0h)                 ; with 1-byte
  00ACh 40h+1    LCD ConfigCmds (set via Cmd F0h)                 ; checksums
  00EDh 68h+1    Identity Data ("provisioned" at walk start time) ; and backup
  0156h 18h+1    Health Data   ("provisioned" at walk start time) ; copies at
  016Fh 01h+1    Copy Flag (00h=Normal, A5h=copy was interrupted) ; 0180h-027Fh
  0171h 0Fh      Unused                                           ;/
  0180h 100h     Backup copies of entries at 0080h-0017Fh
  0280h ...      Various Bitmaps
  8C70h ...      Various Garbage, Bitmaps, Items, Team, Route
  CE8Ah 2        current watts written to eeprom by cmd 20h before replying
                   (likely so remote can read them directly). u16 BE
  CE8Ch ...      Various stuff
  CEF0h 1Ch      Historic step count per day. u32 each, BE,
                   [0] is yesterday, [1] is day before, etc...
  CF0Ch ...      Various stuff

 __________________ Data Structures (in EEPROM and Packets) ___________________

Unique ID (28h bytes)

  00h 28h Generated by the DS game at pairing time, unique per walker

Identity Data (68h bytes) //stored reliably at 0x00ED/0x01ED
All multi-byte values LE, unless otherwise specified

  00h 4   Unknown (LE, always 1?) ;\written from game packet at walk start
  04h 4   Unknown (LE, always 1?) ;  ;<-- 0 at walk end  ;copied from [0]
  08h 2   Unknown (LE, always 7?) ;
  0Ah 2   Unknown (LE, always 7?) ;/ ;<-- 0 at walk end  ;copied from [8?]
  0Ch 2   TrainerTID
  0Eh 2   TrainerSID
  10h 28h Unique ID
  38h 10h EventBitmap (aka bitfield with 128 event flags?)
  48h 10h Trainer Name (8 chars, using a custom 16bit charset, non-unicode)
  58h 1   Unknown
  59h 1   Unknown
  5Ah 1   Unknown
  5Bh 1   Flags (bit0=PairedToGame, bit1=HasPoke, bit2=PokeJoinedOnAWalk)
  5Ch 1   ProtoVer (02h)    (written by DS, refuse peer's with other values)
  5Dh 1   Unknown
  5Eh 1   ProtoSubver (00h) (written by DS, refuse peer's with other values)
  5Fh 1   Unknown (02h)     (written by DS at walk start)
  60h 4   LastSyncTime ;Big Endian  ;in WHAT... maybe seconds since WHEN?
  64h 4   StepCount    ;Big Endian  ;since WHEN... today? lifetime? lastsync?

Peer Play Data (38h bytes) //aka "PEER PLAY DATA"
All multi-byte values LE, unless otherwise specified

  00h 4   curStepCount  (since WHEN?)
  04h 2   curWatts
  06h 2   Unused
  08h 4   Unknown, copied from IdentityData[00h]
  0Ch 2   Unknown, copied from IdentityData[08h]
  0Eh 2   Species
  10h 16h P-Nickname   (11 chars)  ;\the actual names in bitmap format
  26h 10h Trainer Name (8 chars)   ;/are stored elsewhere in EEPROM?
  36h 1   GenderForm
  37h 1   HasSpecialForms (spinda, arceus, unown, etc.)

LcdConfigCmds (40h bytes)
Stored reliably at 00ACh/01ACh.

  00h 1   u8 contrastAndFlags (if 00h/FFh? commands at ROM:BEB8h will be used)
  01h 3Fh u8 commands[3fh] (Commands, or FDh,NNh=Delay(NNh), FEh=End of list)

Enroll Data (71h bytes)

  00h 28h Unique ID             ;always written
  28h 40h LCD Config Data       ;written or verified depending on byte[70h]
  68h 8   Whatever ID           ;always written to EEPROM:0008h
  70h 1   LCD Action (00h=WriteA, 01h=Compare, 03h=WriteB, 02h/04h-FFh=Nop)

HealthData (18h bytes)
Stored reliably at 0156h/0256h. Cached in RAM at F780h.
Big Endian unless otherwise noted.

  00h 4  u32 lifetimeTotalSteps
  04h 4  u32 todaySteps           //zeroed at midnight
  08h 4  u32 lastSyncTime
  0Ch 2  u16 totalDays
  0Eh 2  u16 curWatts
  10h 2  u16 unk_0
  12h 1  u8 unk_1
  13h 1  u8 unk_2
  14h 3  u8 padding[3]
  17h 1  u8 settings (bit0=isOnSpecialRoute, bit1-2=Volume, bit3-6=Contrast)

See Dmitry's webpage for more "game-specifc" data structures and memory addresses.

DS Cart Infrared P-Walker Ports LCD Controller

LCD Command Table
Commands are 1 or 2 bytes long (with D/C=Low, and, unlike normal SPI, one can release /CS after each byte, even when sending 2-byte commands).

  1st Byte     2nd Byte  Description
  00h+(0..Fh)  -       Set Column Address bit0-3   ;\VRAM xloc in 1-pixel units
  10h+(0..7)   -       Set Column Address bit4-6   ;/
  18h+(0..7)   -       Reserved
  20h+(0..7)   -       Set Internal Regulator Resistor Ratio
                         (0..7 = 2.3, 3.0, 3.7, 4.4, 5.1, 5.8, 6.5, 7.2)
  28h+(0..7)   -       Set Power Control Register
                         bit2: Internal Voltage Booster (0=Off, 1=On)
                         bit1: Internal Regulator       (0=Off, 1=On)
                         bit0: Output Op-amp Buffer     (0=Off, 1=On)
  30h+(0..0Fh) -       Reserved
  40h          00h-xxh Set Display Start Line (0..127?) (ROW) ("scroll yloc")
  41h+(0..2)   00h-xxh Same as above?
  44h          00h-xxh Set Display Offset (0..63) (COM0=ROW0..63) (pinout?)
  45h+(0..2)   00h-xxh Same as above?
  48h          00h-xxh Set Multiplex Ratio (num lines, duty 1/((16..64)+icon))
  49h+(0..2)   00h-xxh Same as above?
  4Ch          00h-3Fh Set N-line Inversion (0=Off, 1..31=Reduce crosstalk?)
  4Ch          20h-FFh Same as above?
  4Dh+(0..2)   00h-FFh Same as above?
  50h+(0..7)   -       Set LCD Bias (0..5=1/(4..9), 6=1/9, too)
  57h+(0..0Ch) -       Reserved
  64h+(0..3)   -       Set DC-DC Converter Factor (0=2x/3x, 1=4x, 2/3=5x)
  68h+(0..18h) -       Reserved
  81h          00h-3Fh Set Contrast (0..3Fh, 3Fh=Darkest)
  82h          OTP     Set VL6 voltage (00h..0Fh = original+Signed4bit(N))
  83h          OTP     OTP Programming?
  84h+(0..7)   -       Reserved
  88h          00h-FFh Set White Mode      (bit0-3=FrameA, bit4-7=FrameB)
  89h          00h-FFh Set White Mode      (bit0-3=FrameC, bit4-7=FrameD)
  8Ah          00h-FFh Set Light Gray Mode (bit0-3=FrameA, bit4-7=FrameB)
  8Bh          00h-FFh Set Light Gray Mode (bit0-3=FrameC, bit4-7=FrameD)
  8Ch          00h-FFh Set Dark Gray Mode  (bit0-3=FrameA, bit4-7=FrameB)
  8Dh          00h-FFh Set Dark Gray Mode  (bit0-3=FrameC, bit4-7=FrameD)
  8Eh          00h-FFh Set Black Mode      (bit0-3=FrameA, bit4-7=FrameB)
  8Fh          00h-FFh Set Black Mode      (bit0-3=FrameC, bit4-7=FrameD)
                         Above defines the grayscale palette for color 0-3,
                         normally all frames should use the same setting.
                         Color 0 is usually white (set to zero), color 3 is
                         usually black (set to number of levels selected via
                         cmd 90h). Color 1 and 2 are usually light/dark gray,
                         set to desired contrast, which may depend on the LCD.
  90h+(0..7)   -       Set PWM and FRC for gray-scale operation
                         bit0-1: Levels (0/1=Nine, 2=Twelve, 3=Fifteen Levels)
                         bit2:   Frames (0=Four, 1=Three Frames)
                         Note: Nintendo uses "9 levels" ranging from "0 to 9"
                         (maybe level 0 is treated as off, thus not counted)
  98h+(0..7)   -       Reserved
  A0h+(0..1)   -       Set Segment Remap (0=Col00h is SEG0, 1=Col7Fh is SEG0)
                         Aka xflip mirror?
  A2h+(0..1)   -       Set Icon Enable (0=Disable, 1=Enable)
  A4h+(0..1)   -       Set Entire Display On/Off (0=Show RAM, 1=All Pixels On)
  A6h+(0..1)   -       Set Inverse Display (0=Normal, 1=Inverse On/Off Pixels)
  A8h+(0..1)   -       Set Power Save Mode (0=Standby, 1=Sleep)
  AAh          -       Reserved
  ABh          -       Start Internal Oscillator (needed after reset)
  ACh+(0..1)   ?       Reserved
  AEh+(0..1)   -       Set Display On/Off (0=Off, 1=On)
  B0h+(0..0Fh) -       Set Page Address (00h..0Fh)  ;VRAM yloc in 8-pixel units
  C0h+(0,8)    -       Set COM Output Scan Direction (0=Normal, 8=Remapped)
                         Remapped: COM[0..(N-1)] becomes COM[(N-1)..0])
                         Aka yflip mirror?
  C1h+(0..6)   -       Same as above (Normal)
  C9h+(0..6)   -       Same as above (Remapped)
  D0h+(0..10h) -       Reserved
  E1h          -       Exit Power-save Mode (return from Sleep/Standby modes)
  E2h          -       Software Reset (initialize some internal registers)
  E3h          -       Reserved
  E4h          -       Exit N-line Inversion mode
  E5h+(0..2)   -       Reserved
  E8h          LEN+DTA Transfer VRAM Display Data (for 3-wire SPI mode only)
                         (LEN=00h-FFh, followed by LEN+1 data bytes)
  E9h+(0..6)   -       Reserved
  F0h+(0..0Fh) ..      Test mode commands and Extended features
  F0h          00h-03h Ext. Set VL6 Noise reduction (0=Enable, 3=Disable)
  F1h          08h-0Fh Ext. Set TC Value per 'C (0=-0.05%, 1=-0.07%, 2..7=?)
  F2h          00h-07h Ext. Oscillator Adjustment
                         (0..7 = -9%, -6%, -3%, +0%, +3%, +6%, +9%, +12%)
  F7h          00h-01h Ext. Oscillator Source (0=Internal, 1=External, 2=UNDOC)
  F6h          00h-1Fh Ext. Frame Frequency Adjust
                         bit0-2: FrameFQ (0..7 = 0..7)
                         bit3-4: Fosc    (0-3 = 59kHz, 75kHz, 94kHz, 113kHz)
  FBh          X2h,X6h Ext. Lock/Unlock Interface (bit2: 0=Unlock, 1=Lock)

VRAM Writing
First set the VRAM address using three commands with D/C=Low:

  00h+(xloc) AND 0Fh     ;Set Column Address bit0-3
  10h+(xloc/10h)         ;Set Column Address bit4-6
  B0h+(yloc/8)           ;Set Page Address (00h..0Ah, other=reserved?)

For the actual data transfer, set D/C=High and write data (each two bytes contain vertical 8-pixel columns, with bit0=top, bit7=bottom):

  1st byte = bitplane 1   ;\color 0..3 are usually white, lgray, dgray, black
  2nd byte = bitplane 0   ;/(palette can be changed via cmd 88h-8Fh though)

Note: Column Address (xloc) is auto-incremented after each byte pair, the Page Address (yloc) is NOT incremented when wrapping from x=7Fh to x=00h.
The last page is the "icon", this has only 1 pixel per xloc, maybe intended to enable special symbols instead of pixels (on displays to combine symbols and pixels).

VRAM Reading and Status Byte (both supported in parallel Non-SPI mode only)
The controller supports four modes (Nintendo uses the 4-wire SPI mode):

  3-wire SPI Serial write-only (/CS, CLK, MOSI, with cmd E8h instead D/C pin)
  4-wire SPI Serial write-only (/CS, CLK, MOSI, D/C=Data/Cmd)
  12-wire 8080 Parallel read/write (/CS, D0-D7, D/C, /RD, /WR)
  12-wire 6800 Parallel read/write (/CS, D0-D7, D/C, E, R/W)

VRAM can be read setting the VRAM address as usually, and then reading with D/C=High. And, an 8bit status byte can be read with D/C=Low:

  7    BUSY    Chip is executing instruction (0=Ready, 1=Busy)
  6    ON      Display is On/Off             (0=Off, 1=On)
  5    RES     Chip is executing reset       (0=Ready, 1=Busy)
  4-0  -       Fixed Chip ID                 (08h=SSD1850)

Note
Nintendo uses a 96x64 pixel grayscale LCD screen with 2bpp, without backlight or frontlight. VRAM is double buffered (VRAM size is at least 96x128 pixels).
The display controller seems to be a Solomon System SSD1850 chip... or similar, there are a number of datasheets that are more or less matching the LCD commands used in the firmware:

  SSD0852  128x128 would allow double-buffer, but extended commands are wrong
  SSD0858  104x65  close, but extended commands are wrong
  SSD0859  128x81  could be correct (almost same as SSD1850)
  SSD1820  128x65  wrong, lacks palette (command 88h-8Fh)
  SSD1820A 128x65  wrong, lacks palette (command 88h-8Fh)
  SSD1821  128x81  wrong, lacks palette (command 88h-8Fh)
  SSD1850  128x65  could be correct (ysiz is good, but no double-buffering)
  SSD1851  128x81  as above, but more lines than needed
  SSD1852  128x128 would allow double-buffer, but extended commands are wrong
  SSD1854  128x160 wrong, uses 2-byte command B0h,YYh lacks extended commands
                   (also cmd 18h,20h,4xh,50h,60h-63h,64h,82h,83h,etc. differ)
  SSD1858  104x65  close, but lacks many extended commands
  SSD1859  128x81  could be correct (almost same as SSD1850)

But, Nintendo's firmware contains extended command F7h,02h (which isn't mentioned in any of the above datasheets).
And, the firmware seems to use two 64-line framebuffers at DisplayStart=0 and DisplayStart=64, ie. needing 128 lines in total, which leaves no space for the "icon" scanline (at least not addressable via cmd B0h+y).

DS Cart Infrared P-Walker Ports Accelerometer BMA150

Might do drilling machine, not good honking out paste pretty screenshot?
Official BMA150 register specs are in a Bosch-Captcha, which appears to be intended to fool humans and bots alike (text on hatched background, undefined color codes, increasingly unpleasant medusa-like details at closer look). Anyways, here's a plaintext-hack of the Bosch-Captcha:

  00h  Chip ID     (bit7-3=Unused, bit2-0=02h)
  01h  Version     (bit7-4=al_version, bit3-0=ml_version) (undefined values)
  02h  Acc X Low   (bit7-6=DataLsb, bit5-1=Unused, bit0=NewDataFlag)
  03h  Acc X High  (bit7-0=DataMsb)
  04h  Acc Y Low   (bit7-6=DataLsb, bit5-1=Unused, bit0=NewDataFlag)
  05h  Acc Y High  (bit7-0=DataMsb)
  06h  Acc Z Low   (bit7-6=DataLsb, bit5-1=Unused, bit0=NewDataFlag)
  07h  Acc Z High  (bit7-0=DataMsb)
  08h  Temperature (bit7-0=DataTempMsb) (Lsb not existing, except in Trimming?)
  09h  Status Flags     (see below)
  0Ah  Control Flags    (see below)
  0Bh  Config Flags     (see below)
  0Ch  LG Threshold     (bit7-0)
  0Dh  LG Duration      (bit7-0)
  0Eh  HG Threshold     (bit7-0)
  0Fh  HG Duration      (bit7-0)
  10h  Any Motion Threshold (bit7-0)
  11h  Misc Stuff       (bit7-6=AnyMotionDur, bit5-3=HG Hyst, bit2-0=LG Hyst)
  12h  Customer Reserved 1 (bit7-0)
  13h  Customer Reserved 2 (bit7-0)
  14h  Range/Bandwidth  (bit7-5=Reserved, bit4-3=Range, bit2-0=Bandwidth)
  15h  Misc Flags       (see below)
  16h  Trimming X Low   (bit7-6=OffsetLsb, bit5-0=Gain)
  17h  Trimming Y Low   (bit7-6=OffsetLsb, bit5-0=Gain)
  18h  Trimming Z Low   (bit7-6=OffsetLsb, bit5-0=Gain)
  19h  Trimming T Low   (bit7-6=OffsetLsb, bit5-0=Gain)
  1Ah  Trimming X High  (bit7-0=OffsetMsb)
  1Bh  Trimming Y High  (bit7-0=OffsetMsb)
  1Ch  Trimming Z High  (bit7-0=OffsetMsb)
  1Dh  Trimming T High  (bit7-0=OffsetMsb)
  1Eh-22h BST reserved  (official blank/green)
  23h     BST reserved  (official blank/white)
  24h-2Ah Not used      (official gray/dither)
  2Bh-3Dh EEPROM Defaults for Registers 0Bh-1Dh
  3Eh-42h BST reserved  (official blank/orange)
  43h-49h Not used      (official gray/dither)
  4Ah-4Fh Not mentioned (official not here)
  50h-7Fh BST reserved  (official blank/cyan)

Obscure Notes: Registers are 00h-06h,21h-22h,43h-7Fh are classifed as NOTHING, 07h-20h as IMAGE, and 23h-42h as EEPROM (whatever that crap means). Also, Registers 00h-15h are OPERATIONAL, 16h-3Dh are DEFAULT SETTING, 3Eh-7Fh are BOSCH SENSORTEC RESERVED.

Register 09h - Status Flags

  7   ST Result
  6-5 Not used (official piss/dither)
  4   Alert Phase
  3   LG_latched
  2   HG_latched
  1   LG_status
  0   HG_status

Register 0Ah - Control Flags

  7   Reserved (official gray/dither)
  6   Reset INT
  5   Update IMAGE
  4   EE_W (uh? maybe eeprom write?)
  3   Self Test 1
  2   Self Test 0
  1   Soft Reset
  0   Sleep

Register 0Bh - Config Flags

  7   Alert
  6   Any Motion
  5-4 Counter HG
  3-2 Counter LG
  1   Enable HG
  0   Enable LG

Register 15h - Misc Flags

  7   SPI4
  6   enable_adv_INT
  5   new_data_INT
  4   latch_INT
  3   shadow_dis
  2-1 wake_up_pause
  0   wake_up

DS Cart Infrared Component Lists

Gamecart for Walk with Me/Laufrhythmus

  Case "Nintendo, NTR-031. PAT. PEND., IMWPN1J12"
  PCB  "DA A-4  IRU01-10" (two layers) plus "IRL01-01 "(brown extra film layer)
  U1 32pin  "S906748-1, SanDisk, 11014-64B, P0A837.00, 0843, NTR-IMWP-1" (ROM)
  U2 32pin  "38600R, A06V, AH00167, 0832" (CPU, ROM 8Kbyte, RAM 0.5KByte)
  U3  5pin  "?" (OR-gate? flipflop?) (for forwarding SPI /CS to FLASH /CS)
  U4  8pin  "45PE80VG, HPAMZ V5, KOR 833X, ST e3" (SPI FLASH 1024 Kbytes)
  U1' 7pin  "5  S.. 9" IR transceiver (on brown film layer)
  X1  6pin  "737Wv"   ;7.37MHz?    /FLASH.CS --|""""|-- GND
  R1,R2,RA1 resistors                          | U3 |-- /SPI.CS (from NDS)
  C1,C2,C3,C4,C5,C6 capacitors         VDD33 --|____|-- U2.pinxxx

Note: The printed part number on the CPU is 38600R, which does officially have ROM 8Kbyte, RAM 0.5KByte (and the software does use only that much memory, but it does actually contain twice as much ROM and RAM, ie. it seems to be a 38602R chip... or, as the part number is wrong, it MIGHT even be a F38602 with FLASH memory?).
IR transceiver wiring: 1=GND, 2=CPU.pin24=TXD, 3=CPU.pin25=RXD, 4=CPU.pin26=PWDOWN, 5=NC, 6=GND, 7=VCC.

Gamecart for P-Letter Black

  PCB  "DI Y-1  IRC02-01" (two layers, without brown extra film layer)
  U1 32pin  "MXIC..." (ROM)
  U2 32pin  "..." (CPU, ROM 8Kbyte, RAM 0.5KByte)
  U3  5pin  "..." (OR-gate? flipflop?) (for forwarding SPI /CS to FLASH /CS)
  U4  8pin  "..." (SPI FLASH)
  U5  7pin  "..." IR transceiver
  X1  6pin  "..."
  R1,R2,RA1 resistors
  C1,C2,C3,C4,C5,C6 capacitors

More or less same as in Walk with Me, but without film layer, and with non-legible part numbers on existing photos.

Activity Meter (Actimeter in german) (Nintendo, 2008-2009)
NTR-IMWJ Aruite Wakaru Seikatsu Rhythm (JP)
NTR-IMWE Personal Trainer: Walking (US)
NTR-IMWP Walk with Me (EU) (Laufrhythmus in german)
NTR-IA8P Active Health with Carol Vorderman (EU)

  Case "Nintendo DS, NTR-027, (C) 2008 Nintendo, NTR-A-HC, Made in Japan"
  Case "CE ./ VCI, ACN 060 566 083, Nintendo"
  PCB  "NTR-DHC-01" (in water resistant case)
  Ux  32pin Side-A "38602R, F22V, AH04731, 0834" (CPU, ROM 16Kbyte, RAM 1KByte)
  U2  8pin  Side-B "564X, 48H3, 30" (SPI EEPROM 8Kbyte, ST M95640-W or similar)
  U3  7pin  Side-B "1  S.  9" IR transceiver
  ??  2pin  Side-A huge smd capacitor shaped thing, maybe analog 1-axis sensor?
  ??  2pin  Side-A huge smd capacitor shaped thing, maybe analog 1-axis sensor?
  Ux/Qx     Side-A many small chips with 3-6 pins and few markings
  Xx  3pin  Side-A "CB825"      ;32.768-kHz or 38.4-kHz Crystal Resonator?
  Yx  6pin  Side-A ":i] 3.68t"  ;3.68MHz (115.2kHz*32)
  C1..C34   Plenty capacitors
  R1..R28   Plenty resistors
  BTI 2pin  Side-B Battery holder (for CR2032 H, 3V)
  Button    Side-A Push button (communication button)
  |<  4pin  Side-A Two color LED

Activity Meter Instruction Booklet, 310 pages: "Do not disassemble or attempt to repair the Activity Meter yourself. Doing so could result in injury or electrocution."
IR transceiver wiring: 1=BAT+, 2=TX-RC=NC?, 3=1K-to-CPU.pin25=RXD, 4=CPU.pin26=PWDOWN, 5=CPU.pin24=TXD, 6=?, 7=GND.

P-Walker (Nintendo, 2009)
NTR-IPKx/NTR-IPGx P-letter HeartGold/SoulSilver
TWL-IRBO/TWL-IRAO P-letter Black/White
TWL-IREO/TWL-IRDO P-letter Black/White 2

  Case "?"
  PCB "NTR-PHC-01" (with green solder stop & unconventional black text layer)
  U1  32pin Side-B "F38606, F04V, AK04052, 0942" (CPU,FLASH 48Kbyte,RAM 2KByte)
  U2  4pin  Side-A "?"
  U3  4pin  Side-A "?"
  U4  4pin  Side-A "M_RA"
  U5  7pin  Side-B IR transceiver
  U6  8pin  Side-A "Sxxxx, xxxx" (maybe SPI EEPROM?)
  U7  12pin Side-B "043, A939, 021" (accelerometer?) (Bosch BMA150 ?)
  U8  5pin  Side-A "?"
  Q1  6pin  Side-A "Z4"
  D1  3pin  Side-A "?" dual diode or so
  X1  3pin  Side-B "EAJJ"       ;32.768-kHz or 38.4-kHz Crystal Resonator?
  Y1  6pin  Side-B "3.68"       ;3.68MHz (115.2kHz*32)
  BZ1 2pin  Side-B wires to piezo speaker (aka buzzer)
  CN1 14pin Side-A LCD connector 14pin? or 2x14pin? (without backlight)
            (with SSD1850 display controller (or similar) inside of LCD screen)
            (96x64 2-bit greyscale screen) (reportedly with SPI bus)
  BT+/-     Side-B Battery contacts for removeable battery (for CR2032, 3V)
  C1..C29   Plenty capacitors
  R1..R22   Plenty resistors
  SW's      Side-A Three buttons (left, center, right)

Fit Meter (for Wii Fit U) (2013?)

  PCB "SAMU-01" (with green solder stop & unconventional black text layer)
  U1  40pin Side-B "R5F101EEA, 1242KE415, SINGAPORE" (RL78 CPU)
  U2  7pin  Side-B "845G2947" IR transceiver, with metal shield
  U3  16pin Side-B (not installed)
  U4  16pin Side-A --UNKNOWN MARKING, BAD PHOTO-- maybe accelerometer
  U5  4pin  Side-A --UNKNOWN MARKING, BAD PHOTO-- maybe SPI EEPROM/FLASH
  U6  4pin  Side-A (not installed)
  U7  7pin  Side-B "I357, U231, 094" whatever, in metal shielded case
  X1  3pin  Side-B "EABL" crystal or so
  BZ1 2pin  Side-B wires to piezo speaker (aka buzzer)
  CN1 14pin Side-A LCD connector 14pin? or 2x14pin? (without backlight)
  BT+/-     Side-B Battery contacts for removeable battery
  Q1  3pin  Side-A Transistor or so
  D1..D2    Side-A Diodes (3pin each)
  C1..C29   Plenty capacitors
  R1..R22   Plenty resistors
  SW's      Side-A Three buttons (left, center, right)

Similar as P-Walker, but uses a RL78 CPU (alike 3DS MCU) instead of H8/386.

The IR transceivers do resemble RPM841-H16 or similar (as used in 3DS). Observe that 7pin transceiver pinouts are NOT same. Walk with Me, Activity Meter, and 3DS all have different 7pin transceiver wirings (and, some have additional metal shields).

H8/386 SFRs

H8/386 Special Function Registers at F000h and up

  F020h FLMCR1  FLASH Memory Control 1
  F021h FLMCR2  FLASH Memory Control 2
  F022h FLPWCR  FLASH Memory Power Control
  F023h EBR1    FLASH Erase Block 1
  F02Bh FENR    FLASH Memory Enable
  F067h RTCFLG  RTC Interrupt Flag
  F068h RSECDR  RTC Seconds / Free running counter
  F069h RMINDR  RTC Minutes
  F06Ah RHRDR   RTC Hours
  F06Bh RWKDR   RTC Day-of-week
  F06Ch RTCCR1  RTC Control 1
  F06Dh RTCCR2  RTC Control 2
  F06Fh RTCCSR  RTC Clock Source Select
  F078h ICCR1   I2C Bus Control 1
  F079h ICCR2   I2C Bus Control 2
  F07Ah ICMR    I2C Bus Mode
  F07Bh ICIER   I2C Bus Interrupt Enable
  F07Ch ICSR    I2C Bus Status
  F07Dh SAR     I2C Slave Address
  F07Eh ICDRT   I2C Bus Transmit Data
  F07Fh ICDRR   I2C Bus Receive Data
  F085h PFCR    System Port Function Control
  F086h PUCR8   Port 8 Pull-up Control
  F087h PUCR9   Port 9 Pull-up Control
  F08Ch PODR9   Port 9 Open-drain Control
  F0D0h TMB1    Timer B1 Mode
  F0D1h TC/LB1  Timer B1 Counter (R) / Load (W)
  F0DCh CMCR0   Compare Control 0
  F0DDh CMCR1   Compare Control 1
  F0DEh CMDR    Compare Data
  F0E0h SSCRH   SPI Synchronous Serial Control H     (AccessState3)
  F0E1h SSCRL   SPI Synchronous Serial Control L     (AccessState3)
  F0E2h SSMR    SPI Synchronous Serial Mode          (AccessState3)
  F0E3h SSER    SPI Synchronous Serial Enable        (AccessState3)
  F0E4h SSSR    SPI Synchronous Serial Status        (AccessState3)
  F0E9h SSRDR   SPI Synchronous Serial Receive Data  (AccessState3)
  F0EBh SSTDR   SPI Synchronous Serial Transmit Data (AccessState3)
  F0F0h TMRW    Timer W Mode
  F0F1h TCRW    Timer W Control
  F0F2h TIERW   Timer W Interrupt Enable
  F0F3h TSRW    Timer W Status
  F0F4h TIOR0   Timer W I/O control 0
  F0F5h TIOR1   Timer W I/O control 1
  F0F6h TCNT    Timer W Counter   (16bit)
  F0F8h GRA     Timer W General A (16bit)
  F0FAh GRB     Timer W General B (16bit)
  F0FCh GRC     Timer W General C (16bit)
  F0FEh GRD     Timer W General D (16bit)

H8/386 Special Function Registers at FF80h and up, with [FFnn] & [FFnn].bitno

  FF8Ch ECPWCR  Async Event Counter PWM Compare (16bit)
  FF8Eh ECPWDR  Async Event Counter PWM Data (16bit)
  FF91h SPCR    IrDA UART Serial 3 Port Control
  FF92h AEGSR   Async Event Input Pin Edge Select
  FF94h ECCR    Async Event Counter Control
  FF95h ECCSR   Async Event Counter Control/Status
  FF96h ECH     Async Event Counter H
  FF97h ECL     Async Event Counter L
  FF98h SMR3    IrDA UART Serial 3 Mode          (AccessState3)
  FF99h BRR3    IrDA UART Serial 3 Bit Rate      (AccessState3)
  FF9Ah SCR3    IrDA UART Serial 3 Control       (AccessState3)
  FF9Bh TDR3    IrDA UART Serial 3 Transmit Data (AccessState3)
  FF9Ch SSR3    IrDA UART Serial 3 Status        (AccessState3)
  FF9Dh RDR3    IrDA UART Serial 3 Receive Data  (AccessState3)
  FFA6h SEMR    IrDA UART Serial 3 Extended Mode (AccessState3)
  FFA7h IrCR    IrDA Control
  FFB0h TMWD    Timer WD Watchdog Mode
  FFB1h TCSRWD1 Timer WD Watchdog Control/Status 1
  FFB2h TCSRWD2 Timer WD Watchdog Control/Status 2
  FFB3h TCWD    Timer WD Watchdog Counter
  FFBCh ADRR    A/D Converter Result (16bit)
  FFBEh AMR     A/D Converter Mode
  FFBFh ADSR    A/D Converter Start
  FFC0h PMR1    Port 1 Mode
  FFC2h PMR3    Port 3 Mode
  FFCAh PMRB    Port B Mode
  FFD4h PDR1    Port 1 Data
  FFD6h PDR3    Port 3 Data
  FFDBh PDR8    Port 8 Data
  FFDCh PDR9    Port 9 Data
  FFDEh PDRB    Port B Data
  FFE0h PUCR1   Port 1 Pull-up Control
  FFE1h PUCR3   Port 3 Pull-up Control
  FFE4h PCR1    Port 1 Control
  FFE6h PCR3    Port 3 Control
  FFEBh PCR8    Port 8 Control
  FFECh PCR9    Port 9 Control
  FFF0h SYSCR1  System Control 1
  FFF1h SYSCR2  System Control 2
  FFF2h IEGR    Interrupt Edge Select
  FFF3h IENR1   Interrupt Enable 1
  FFF4h IENR2   Interrupt Enable 2
  FFF5h OSCCR   System Oscillator Control
  FFF6h IRR1    Interrupt Flag 1
  FFF7h IRR2    Interrupt Flag 2
  FFFAh CKSTPR1 Clock Stop 1
  FFFBh CKSTPR2 Clock Stop 2

H8/386 Exception Vectors

H8/386 Exception Vectors (Vector 0000h has highest priority, 004Eh lowest)
Below are for Normal Mode with 16bit addressing (Extended Mode has 32bit vectors at 0000h..009Fh accordingly).

  0000h Reset/Watchdog
  0002h Reserved
  0004h Reserved
  0006h Reserved
  0008h Reserved
  000Ah Reserved
  000Ch Reserved
  000Eh External NMI interrupt
  0010h Trap 0 opcode
  0012h Trap 1 opcode
  0014h Trap 2 opcode
  0016h Trap 3 opcode
  0018h Reserved
  001Ah CPU Direct transition by executing SLEEP
  001Ch Reserved
  001Eh Reserved
  0020h External IRQ0 interrupt
  0022h External IRQ1 interrupt
  0024h External IRQAEC interrupt
  0026h Reserved
  0028h Reserved
  002Ah Comparator COMP0
  002Ch Comparator COMP1
  002Eh RTC per 0.25 seconds (4Hz)  ;0.25-second overflow
  0030h RTC per 0.5 seconds (2Hz)   ;0.5-second overflow
  0032h RTC per second (1Hz)        ;Second periodic overflow
  0034h RTC per minute              ;Minute periodic overflow
  0036h RTC per hour                ;Hour periodic overflow
  0038h RTC per day                 ;Day-of-week periodic overflow
  003Ah RTC per week (7 days)       ;Week periodic overflow
  003Ch RTC Free-running overflow
  003Eh WDT overflow (interval timer)
  0040h Asynchronous event counter overflow
  0042h Timer B1 Overflow
  0044h Serial SPI (or IIC2) (aka I2C ?)
  0046h Timer W Overflow or Capture/compare A,B,C,D
  0048h Reserved
  004Ah IrDA UART Serial 3
  004Ch A/D Conversion end
  004Eh Reserved

Note: The SSU (aka SPI) and IIC (aka I2C) share the same vector address. When using the IIC, shift the SSU to standby mode using CKSTPR2.

H8/300H Operands

CPU Registers

  R0..R6  32bit General Purpose   ;\can be alternately used as
  R7 (SP) 32bit Stack Pointer     ;/8bit/16bit registers (see below)
  PC      24bit Program Counter
  CCR     8bit  Flags (occupies 16bit when pushed/stored in memory)

Registers R0..R7 can be split into 8bit/16bit registers (alike 80x86 registers):

  .-----------------------.
  | ERx                   | 32bit  (ERx)
  |-----------+-----------|
  | Ex        | Rx        | 16bit  (Rx)
  '-----------+-----+-----|
              ' RxH | RxL | 8bit   (RxB)
              '-----'-----'

There are no opcodes for splitting upper 16bit in Ex into 8bit ExL,ExH (except, the sign/zero-extend opcodes seem to allow to extend ExL to Ex).

Normal and Extended Mode
The CPU can be wired to two different addressing modes:

  Normal Mode   --> 16bit addressing (default)
  Extended Mode --> 24bit addressing

Normal mode is usually used (the CPU has less than 64K ROM/RAM anyways), the opcodes with [er0..er7] memory addressing are then actually accessing [r0..r7], ie. the upper 16bit in e0..e7 are ignored (and can be used for general purpose data).
Extended mode, if it were ever used, uses the lower 24bit of er0..er7 for [er0..er7] addressing, and CALL/RET and indirect JMP opcodes are a bit slower, and exception vectors are 32bit wide.

Memory Addressing
Data is stored in Big-Endian. 16bit/32bit values must be stored at even addresses (with bit0 cleared) (there is no need to clear bit1 for 32bit values, namely, push/pop work regardless of bit1).

  Native        Nocash
  @aa:8         [FFaa]        Memory at FF00h..FFFFh (upper RAM and SFR's)
  @aa:16        [nnnn]
  @aa:24        [nnnnnn]
  @Erm          [Erm]
  @(d:16,ERm)   [ERm+nnnn]
  @(d:24,ERm)   [ERm+nnnnnn]
  @ERm+         [ERm+]          Memory access with post-increment
  @-ERm         [ERm-]          Memory access with pre-decrement
  (implied)     [ER6+],[ER5+]   Memory block transfer (EEPMOV)

Bit Addressing

  Native        Nocash
  #nn:8,@aa:8   [FFaa].n
  RnB,@aa:8     [FFaa].RnB
  #nn:8,RdB     RdB.n
  RnB,RdB       RdB.RnB

Note: The "#nn:8" suggests native syntax to use "#0x80" to select bit7, however, the existing disassembler does instead use "#7".

H8/300H Opcodes

All opcodes are multiples of 2 bytes (2,4,6,8,10 bytes), all opcodes should be always located at even addresses (ie. bit0 of jump address/disp operands should be always 0).

Register Encoding

  0..7  8bit Registers R0H..R7H  (bit8-15)     ;\RxB
  8..F  8bit Registers R0L..R7L  (bit0-7)      ;/
  0..7  16bit Registers R0..R7   (bit0-15)     ;\Rx
  8..F  16bit Registers E0..E7   (bit16-31)    ;/
  0..7  32bit Registers ER0..ER7 (bit0-31)     ;-ERx (in normal opcodes)
  8..F  32bit Registers ER0..ER7 (bit0-31)     ;-ERx (in opcodes marked *m,*s)

H8/300H Main Opcodes

  Opcode             Native                  Nocash             States IxHUNZVC
  0..                --> Misc 0xxx
  1..                --> Misc 1xxx
  2dnn               MOV.B @aa:8,RdB         MOV.B RdB,[FFaa]        4 ----nz0-
  3snn               MOV.B Rs,@aa:8          MOV.B [FFaa],Rs         4 ----nz0-
  4cnn               --> Jumps (relative 8bit range)
  5..                --> Jumps (various) and unsigned mul/div
  6..                --> Misx 6xxx
  7..                --> Misc 7xxx
  8dnn               ADD.B #nn:8,RdB         ADD.B RdB,nn            2 --h-nzvc
  9dnn               ADDX  #nn:8,RdB         ADC.B RdB,nn            2 --h-nzvc
  Adnn               CMP.B #nn:8,RdB         CMP.B RdB,nn            2 --h-nzvc
  Bdnn               SUBX  #nn:8,RdB         SBC.B RdB,nn            2 --h-nzvc
  Cdnn               OR.B  #nn:8,RdB         OR.B  RdB,nn            2 ----nz0-
  Ddnn               XOR.B #nn:8,RdB         XOR.B RdB,nn            2 ----nz0-
  Ednn               AND.B #nn:8,RdB         AND.B RdB,nn            2 ----nz0-
  Fdnn               MOV.B #nn:8,RdB         MOV.B RdB,nn            2 ----nz0-

H8/300H Misc 0xxx

  0000               NOP                     NOP                     2 --------
  01..               --> Misc 01xx ;Memory Load/Store (32bit ERn) etc.
  020d               STC.B CCR,RdB           MOV.B RdB,CCR           2 --------
  030s               LDC.B RsB,CCR           MOV.B CCR,RsB           2 xxxxxxxx
  04nn               ORC   #nn:8,CCR         OR.B  CCR,nn            2 xxxxxxxx
  05nn               XORC  #nn:8,CCR         XOR.B CCR,nn            2 xxxxxxxx
  06nn               ANDC  #nn:8,CCR         AND.B CCR,nn            2 xxxxxxxx
  07nn               LDC.B #nn:8,CCR         MOV.B CCR,nn            2 xxxxxxxx
  08sd               ADD.B RsB,RdB           ADD.B RdB,RsB           2 --h-nzvc
  09sd               ADD.W Rs,Rd             ADD.W Rd,Rs             2 --h-nzvc
  0A..               --> Increment/Add
  0B..               --> Increment/Add
  0Csd               MOV.B RsB,RdB           MOV.B RdB,RsB           2 ----nz0-
  0Dsd               MOV.W Rs,Rd             MOV.W Rd,Rs             2 ----nz0-
  0Esd               ADDX  RsB,RdB           ADC.B RdB,RsB           2 --h-nzvc
  0F0d               DAA   RdB               DAA.B RdB               2 --U-nzUc
  0Fsd            *s MOV.L ERs,ERd           MOV.L ERd,ERs           2 ----nz0-

H8/300H Misc 01xx ;Memory Load/Store (32bit ERn) etc.

  010069md           MOV.L @ERm,ERd          MOV.L ERd,[ERm]         8 ----nz0-
  014069m0           LDC.W @ERm,CCR          MOV.W CCR,[ERm]         6 xxxxxxxx
  010069ms        *m MOV.L ERs,@ERm          MOV.L [ERm],ERs         8 ----nz0-
  014069m0        *m STC.W CCR,@ERm          MOV.W [ERm],CCR         6 --------
  01006B0dnnnn       MOV.L @aa:16,ERd        MOV.L ERd,[nnnn]       10 ----nz0-
  01406B00nnnn       LDC.W @aa:16,CCR        MOV.W CCR,[nnnn]        8 xxxxxxxx
  01006B2d00nnnnnn   MOV.L @aa:24,ERd        MOV.L ERd,[nnnnnn]     12 ----nz0-
  01406B2000nnnnnn   LDC.W @aa:24,CCR        MOV.W CCR,[nnnnnn]     10 xxxxxxxx
  01006B8snnnn       MOV.L ERs,@aa:16        MOV.L [nnnn],ERs       10 ----nz0-
  01406B80nnnn       STC.W CCR,@aa:16        MOV.W [nnnn],CCR        8 --------
  01006BAs00nnnnnn   MOV.L ERs,@aa:24        MOV.L [nnnnnn],ERs     12 ----nz0-
  01406BA000nnnnnn   STC.W CCR,@aa:24        MOV.W [nnnnnn],CCR     10 --------
  01006Dmd           MOV.L @ERm+,ERd         MOV.L ERd,[ERm+]       10 ----nz0-
  01406Dm0           LDC.W @ERm+,CCR         MOV.W CCR,[ERm+]        8 xxxxxxxx
  01006Dms        *m MOV.L ERs,@-ERm         MOV.L [ERm-],ERs       10 ----nz0-
  01406Dm0        *m STC.W CCR,@-ERm         MOV.W [ERm-],CCR        8 --------
  01006Fmdnnnn       MOV.L @(d:16,ERm),ERd   MOV.L ERd,[ERm+nnnn]   10 ----nz0-
  01406Fm0nnnn       LDC.W @(d:16,ERm),CCR   MOV.W CCR,[ERm+nnnn]    8 xxxxxxxx
  01006Fmsnnnn    *m MOV.L ERs,@(d:16,ERm)   MOV.L [ERm+nnnn],ERs   10 ----nz0-
  01406Fm0nnnn    *m STC.W CCR,@(d:16,ERm)   MOV.W [ERm+nnnn],CCR    8 --------
  010078m06B2d00..   MOV.L @(d:24,ERm),ERd   MOV.L ERd,[ERm+nnnnnn] 14 ----nz0-
  014078m06B2000..   LDC.W @(d:24,ERm),CCR   MOV.W CCR,[ERm+nnnnnn] 12 xxxxxxxx
  010078m06BAs00..*? MOV.L ERs,@(d:24,ERm)   MOV.L [ERm+nnnnnn],ERs 14 ----nz0-
  014078m06BA000..   STC.W CCR,@(d:24,ERm)   MOV.W [ERm+nnnnnn],CCR 12 --------
  0180               SLEEP                   HALT                    2 --------
  01C050sd           MULXS.B RsB,Rd          SMUL.B Rd,RsB          16 ----nz--
  01C052sd           MULXS.W Rs,ERd          SMUL.W ERd,Rs          24 ----nz--
  01D051sd           DIVXS.B RsB,Rd          SDIV.B Rd,RsB          16 ----nz--
  01D053sd           DIVXS.W Rs,ERd          SDIV.W ERd,Rs          24 ----nz--
  01F064sd           OR.L    ERs,ERd         OR.L   ERd,ERs          4 ----nz0-
  01F065sd           XOR.L   ERs,ERd         XOR.L  ERd,ERs          4 ----nz0-
  01F066sd           AND.L   E?Rs,ERd        AND.L  ERd,ERs          4 ----nz0-

H8/300H Misc 1xxx

  10..               --> Shift/Rotate (shift left)
  11..               --> Shift/Rotate (shift right)
  12..               --> Shift/Rotate (rotate left)
  13..               --> Shift/Rotate (rotate right)
  14sd               OR.B  RsB,RdB           OR.B  RdB,RsB           2 ----nz0-
  15sd               XOR.B RsB,RdB           XOR.B RdB,RsB           2 ----nz0-
  16sd               AND.B RsB,RdB           AND.B RdB,RsB           2 ----nz0-
  17..               --> Not/Neg/Extend
  18sd               SUB.B RsB,RdB           SUB.B RdB,RsB           2 --h-nzvc
  19sd               SUB.W Rs,Rd             SUB.W Rd,Rs             2 --h-nzvc
  1A..               --> Decrement/Subtract
  1B..               --> Decrement/Subtract
  1Csd               CMP.B RsB,RdB           CMP.B RdB,RsB           2 --h-nzvc
  1Dsd               CMP.W Rs,Rd             CMP.W Rd,Rs             2 --h-nzvc
  1Esd               SUBX  RsB,RdB           SBC.B RdB,RsB           2 --h-nzvc
  1F0d               DAS   RdB               DAS.B RdB               2 --U-nzU?
  1Fsd            *s CMP.L ERs,ERd           CMP.L ERd,ERs           2 --h-nzvc

H8/300H Shift/Rotate (Shift Logical/Arithmetic, Rotate through Carry or not)

  100d               SHLL.B RdB              SHL.B RdB               2 ----nz0c
  101d               SHLL.W Rd               SHL.W Rd                2 ----nz0c
  103d               SHLL.L ERd              SHL.L ERd               2 ----nz0c
  108d               SHAL.B RdB              SAL.B RdB               2 ----nzvc
  109d               SHAL.W Rd               SAL.W Rd                2 ----nzvc
  10Bd               SHAL.L ERd              SAL.L ERd               2 ----nzvc
  110d               SHLR.B RdB              SHR.B RdB               2 ----0z0c
  111d               SHLR.W Rd               SHR.W Rd                2 ----0z0c
  113d               SHLR.L ERd              SHR.L ERd               2 ----0z0c
  118d               SHAR.B RdB              SAR.B RdB               2 ----nz0c
  119d               SHAR.W Rd               SAR.W Rd                2 ----nz0c
  11Bd               SHAR.L ERd              SAR.L ERd               2 ----nz0c
  120d               ROTXL.B RdB             RCL.B RdB               2 ----nz0c
  121d               ROTXL.W Rd              RCL.W Rd                2 ----nz0c
  123d               ROTXL.L ERd             RCL.L ERd               2 ----nz0c
  128d               ROTL.B RdB              ROL.B RdB               2 ----nz0c
  129d               ROTL.W Rd               ROL.W Rd                2 ----nz0c
  12Bd               ROTL.L ERd              ROL.L ERd               2 ----nz0c
  130d               ROTXR.B RdB             RCR.B RdB               2 ----nz0c
  131d               ROTXR.W Rd              RCR.W Rd                2 ----nz0c
  133d               ROTXR.L ERd             RCR.L ERd               2 ----nz0c
  138d               ROTR.B RdB              ROR.B RdB               2 ----nz0c
  139d               ROTR.W Rd               ROR.W Rd                2 ----nz0c
  13Bd               ROTR.L ERd              ROR.L ERd               2 ----nz0c

H8/300H Not/Neg/Extend

  170d               NOT.B RdB               NOT.B RdB               2 ----nz0-
  171d               NOT.W Rd                NOT.W Rd                2 ----nz0-
  173d               NOT.L Rd                NOT.L ERd               2 ----nz0-
  175d               EXTU.W Rd               UMOV Rd,RdL ;or Ed,EdL? 2 ----0z0-
  177d               EXTU.L ERd              UMOV ERd,Rd             2 ----0z0-
  178d               NEG.B RdB               NEG.B RdB               2 --h-nzvc
  179d               NEG.W Rd                NEG.W Rd                2 --h-nzvc
  17Bd               NEG.L Rd                NEG.L ERd               2 --h-nzvc
  17Dd               EXTS.W Rd               SMOV Rd,RdL ;or Ed,EdL? 2 ----nz0-
  17Fd               EXTS.L ERd              SMOV ERd,Rd             2 ----nz0-

H8/300H Increment/Add and Decrement/Subtract

  0A0d               INC.B RdB               INC.B RdB,1             2 ----nzv-
  1A0d               DEC.B RdB               DEC.B RdB,1             2 ----nzv-
  0Asd            *s ADD.L E?Rs,ERd          ADD.L ERd,ERs           2 --h-nzvc
  1Asd            *s SUB.L ERs,ERd           SUB.L ERd,ERs           2 --h-nzvc
  0B0d               ADDS  #1,ERd            INC.S ERd,1             2 --------
  1B0d               SUBS  #1,ERd            DEC.S ERd,1             2 --------
  0B5d               INC.W #1,Rd             INC.W Rd,1              2 ----nzv-
  1B5d               DEC.W #1,Rd             DEC.W Rd,1              2 ----nzv-
  0B7d               INC.L #1,ERd            INC.L ERd,1             2 ----nzv-
  1B7d               DEC.L #1,ERd            DEC.L ERd,1             2 ----nzv-
  0B8d               ADDS  #2,ERd            INC.S ERd,2             2 --------
  1B8d               SUBS  #2,ERd            DEC.S ERd,2             2 --------
  0B9d               ADDS  #4,ERd            INC.S ERd,4             2 --------
  1B9d               SUBS  #4,ERd            DEC.S ERd,4             2 --------
  0BDd               INC.W #2,Rd             INC.W Rd,2              2 ----nzv-
  1BDd               DEC.W #2,Rd             DEC.W Rd,2              2 ----nzv-
  0BFd               INC.L #2,ERd            INC.L ERd,2             2 ----nzv-
  1BFd               DEC.L #2,ERd            DEC.L ERd,2             2 ----nzv-

H8/300H Jump Opcodes and Unsigned Mul/Div

  50sd               MULXU.B RsB,Rd          UMUL.B Rd,RsB          14 ----nz--
  51sd               DIVXU.B RsB,Rd          UDIV.B Rd,RsB          14 ----nz--
  52sd               MULXU.W Rs,ERd          UMUL.W ERd,Rs          22 ----nz--
  53sd               DIVXU.W Rs,ERd          UDIV.W ERd,Rs          22 ----nz--
  5470               RTS                     RET                  8,10 --------
  55nn               BSR d:8                 CALL $+/-nn           6,8 --------
  5670               RTE                     RETI                   10 xxxxxxxx
  57n0               TRAPA #n:2              TRAP 0..3 ;[0010h+n*2] 14 1x------
  58c0nnnn           --> Jumps (relative 16bit range)
  59s0               JMP @ERs                JMP  ERs                4 --------
  5Annnnnn           JMP @aa:24              JMP  nnnnnn             6 --------
  5Baa               JMP @@aa:8              JMP  [FFaa]          8,10 --------
  5C00nnnn           BSR d:16                CALL $+/-nnnn        8,10 --------
  5Ds0               JSR @ERs                CALL ERs              6,8 --------
  5Ennnnnn           JSR @aa:24              CALL nnnnnn          8,10 --------
  5Faa               JSR @@aa:8              CALL [FFaa]          8,12 --------

H8/300H Relative Jump Opcodes (8bit/16bit range)

  4cnn               Bcc d:8                 Jcc   $+/-nn            4 --------
  58c0nnnn           Bcc d:16                Jcc   $+/-nnnn          6 --------

The 4bit condition code can be:

  0  BRA or BT    JMP        ;always/true
  1  BRN or BF    -          ;never/false
  2  BHI          JA         ;unsigned-above
  3  BLS          JBE        ;unsigned-below-equal
  4  BCC or BHS   JNC or JAE ;unsigned-above-equal
  5  BCS or BLO   JC  or JB  ;unsigned-below
  6  BNE          JNZ or JNE ;not equal/zero
  7  BEQ          JZ  or JE  ;equal/zero
  8  BVC          JNO        ;signed-no overflow
  9  BVS          JO         ;signed-n-overflow
  A  BPL          JNS        ;signed-n-plus
  B  BMI          JS         ;signed-n-minus
  C  BGE          JGE        ;signed-n-greater-eq
  D  BLT          JL         ;signed-n-less
  E  BGT          JG         ;signed-n-greater
  F  BLE          JLE        ;signed-n-less-equal

Destination address should be equal (although, the opcodes are weirdly using byte offsets, hence limiting 8bit range to even 7bit range).
The execution time for condition=false is unknown. The time for BRN (always false) is said to be equal to two NOPs (but unknown why one would use that opcode, and if it does refer to the BRN opcode with 8bit and/or 16bit range).

H8/300H Misc 6xxx

  60nd               BSET   RnB,RdB          SET   RdB.RnB           2 --------
  61nd               BNOT   RnB,RdB          NOT   RdB.RnB           2 --------
  62nd               BCLR   RnB,RdB          CLR   RdB.RnB           2 --------
  63nd               BTST   RnB,RdB          TST   RdB.RnB           2 -----z--
  64sd               OR.W   Rs,Rd            OR.W  Rd,Rs             2 ----nz0-
  65sd               XOR.W  Rs,Rd            XOR.W Rd,Rs             2 ----nz0-
  66sd               AND.W  Rs,Rd            AND.W Rd,Rs             2 ----nz0-
  67nd            *i B{I}ST #nn:8,RdB        MOV   RdB.n,{not} C     2 --------
  68md               MOV.B  @ERm,RdB         MOV.B RdB,[ERm]         4 ----nz0-
  68ms            *m MOV.B  RsB,@ERm         MOV.B [ERm],RsB         4 ----nz0-
  69md               MOV.W  @ERm,Rd          MOV.W Rd,[ERm]          4 ----nz0-
  69ms            *m MOV.W  Rs,@ERm          MOV.W [ERm],Rs          4 ----nz0-
  6A0dnnnn           MOV.B  @aa:16,RdB       MOV.B RdB,[aaaa]        6 ----nz0-
  6A2d00nnnnnn       MOV.B  @aa:24,RdB       MOV.B RdB,[aaaaaa]      8 ----nz0-
  6A4dnnnn           MOVFPE @aa:16,RdB       MOV.B RdB,[periph:aaaa] * ----nz0-
  6A8snnnn           MOV.B  RsB,@aa:16       MOV.B [aaaa],RsB        6 ----nz0-
  6AAs00nnnnnn       MOV.B  RsB,@aa:24       MOV.B [aaaaaa],RsB      8 ----nz0-
  6ACsnnnn           MOVTPE RsB,@aa:16       MOV.B [periph:aaaa],RsB * ----nz0-
  6B0dnnnn           MOV.W  @aa:16,Rd        MOV.W Rd,[aaaa]         6 ----nz0-
  6B2d00nnnnnn       MOV.W  @aa:24,Rd        MOV.W Rd,[aaaaaa]       8 ----nz0-
  6B8snnnn           MOV.W  Rs,@aa:16        MOV.W [aaaa],Rs         6 ----nz0-
  6BAs00nnnnnn       MOV.W  Rs,@aa:24        MOV.W [aaaaaa],Rs       8 ----nz0-
  6Cmd               MOV.B  @ERm+,RdB        MOV.B RdB,[ERm+]        6 ----nz0-
  6Cms            *m MOV.B  RsB,@-ERm        MOV.B [ERm-],RsB        6 ----nz0-
  6Dmd               MOV.W  @ERm+,RdB        MOV.W RdB,[ERm+]        6 ----nz0-
  6Dms            *m MOV.W  RsB,@-ERm        MOV.W [ERm-],RsB        6 ----nz0-
  6Emdnnnn           MOV.B  @(d:16,ERm),RdB  MOV.B RdB,[ERm+nnnn]    6 ----nz0-
  6Emsnnnn        *m MOV.B  RsB,@(d:16,ERm)  MOV.B [ERm+nnnn],RsB    6 ----nz0-
  6Fmdnnnn           MOV.W  @(d:16,ERm),Rd   MOV.W Rd,[ERm+nnnn]     6 ----nz0-
  6Fmsnnnn        *m MOV.W  Rs,@(d:16,ERm)   MOV.W [ERm+nnnn],Rs     6 ----nz0-

H8/300H Misc 7xxx

  70nd               BSET  #nn:8,RdB         SET   RdB.n             2 --------
  71nd               BNOT  #nn:8,RdB         NOT   RdB.n             2 --------
  72nd               BCLR  #nn:8,RdB         CLR   RdB.n             2 --------
  73nd               BTST  #nn:8,RdB         TST   RdB.n             2 -----z--
  74nd            *i B{I}OR  #nn:8,RdB       OR    C,{not} RdB.n     2 -------c
  75nd            *i B{I}XOR #nn:8,RdB       XOR   C,{not} RdB.n     2 -------c
  76nd            *i B{I}AND #nn:8,RdB       AND   C,{not} RdB.n     2 -------c
  77nd            *i B{I}LD  #nn:8,RdB       MOV   C,{not} RdB.n     2 -------c
  78m06A2d00nnnnnn   MOV.B @(d:24,ERm),RdB   MOV.B RdB,[ERm+nnnnnn] 10 ----nz0-
  78m06AAs00nnnnnn   MOV.B RsB,@(d:24,ERm)   MOV.B [ERm+nnnnnn],RsB 10 ----nz0-
  78m06B2d00nnnnnn   MOV.W @(d:24,ERm),Rd    MOV.W Rd,[ERm+nnnnnn]  10 ----nz0-
  78m06BAs00nnnnnn*? MOV.W Rs,@(d:24,ERm)    MOV.W [ERm+nnnnnn],Rs  10 ----nz0-
  79..               --> Immediate (16bit)
  7A..               --> Immediate (32bit)
  7B5C498F           EEPMOV.B              MOV [ER6+],[ER5+],R4L- 8+4n --------
  7BD4598F           EEPMOV.W              MOV [ER6+],[ER5+],R4-  8+4n --------
  7C..               --> Bit Operations (Memory at ERm)
  7D..               --> Bit Operations (Memory at ERm)
  7E..               --> Bit Operations (Memory at FFaa)
  7F..               --> Bit Operations (Memory at FFaa)

The EEPMOV opcodes were originally intended to write to an EEPROM, but they can be used as generic memory transfers with number of bytes in R4L or R4. EEPMOV.B is blocking IRQs and NMIs. EEPMOV.W is also blocking IRQs, however, EEPMOV.W gets aborted upon NMI (without resuming the transfer upon return, so software must manually retry if R4=nonzero after executing EEPMOV.W in combination with NMI sources).

H8/300H Immediate 16bit/32bit

  790dnnnn           MOV.W #nnnn:16,Rd       MOV.W Rd,nnnn           4 ----nz0-
  791dnnnn           ADD.W #nnnn:16,Rd       ADD.W Rd,nnnn           4 --h-nzvc
  792dnnnn           CMP.W #nnnn:16,Rd       CMP.W Rd,nnnn           4 --h-nzvc
  793dnnnn           SUB.W #nnnn:16,Rd       SUB.W Rd,nnnn           4 --h-nzvc
  794dnnnn           OR.W  #nnnn:16,Rd       OR.W  Rd,nnnn           4 ----nz0-
  795dnnnn           XOR.W #nnnn:16,Rd       XOR.W Rd,nnnn           4 ----nz0-
  796dnnnn           AND.W #nnnn:16,Rd       AND.W Rd,nnnn           4 ----nz0-
  7A0dnnnnnnnn       MOV.L #nnnnnnnn:32,E?Rd MOV.L E?Rd,nnnnnnnn     6 ----nz0-
  7A1dnnnnnnnn       ADD.L #nnnnnnnn:32,ERd  ADD.L ERd,nnnnnnnn      6 --h-nzvc
  7A2dnnnnnnnn       CMP.L #nnnnnnnn:32,ERd  CMP.L ERd,nnnnnnnn      6 --h-nzvc
  7A3dnnnnnnnn       SUB.L #nnnnnnnn:32,ERd  SUB.L ERd,nnnnnnnn      6 --h-nzvc
  7A4dnnnnnnnn       OR.L  #nnnnnnnn:32,ERd  OR.L  ERd,nnnnnnnn      6 ----nz0-
  7A5dnnnnnnnn       XOR.L #nnnnnnnn:32,ERd  XOR.L ERd,nnnnnnnn      6 ----nz0-
  7A6dnnnnnnnn       AND.L #nnnnnnnn:32,ERd  AND.L ERd,nnnnnnnn      6 ----nz0-

H8/300H Bit Operations

  7Cm074n0        *i B{I}OR  #nn:8,@ERm      OR    C,{not} [ERm].n   6 -------c
  7Cm075n0        *i B{I}XOR #nn:8,@ERm      XOR   C,{not} [ERm].n   6 -------c
  7Cm076n0        *i B{I}AND #nn:8,@ERm      AND   C,{not} [ERm].n   6 -------c
  7Cm077n0        *i B{I}LD  #nn:8,@ERm      MOV   C,{not} [ERm].n   6 -------c
  7Dm060n0           BSET    RnB,@ERm        SET   [ERm].RnB         8 --------
  7Dm061n0           BNOT    RnB,@ERm        NOT   [ERm].RnB         8 --------
  7Dm062n0           BCLR    RnB,@ERm        CLR   [ERm].RnB         8 --------
  7Dm063n0           BTST    RnB,@ERm        TST   [ERm].RnB         8 -----z--
  7Dm067n0        *i B{I}ST  #nn:8,@ERm      MOV   [ERm].n,{not} C   8 --------
  7Dm070n0           BSET    #nn:8,@ERm      SET   [ERm].n           8 --------
  7Dm071n0           BNOT    #nn:8,@ERm      NOT   [ERm].n           8 --------
  7Dm072n0           BCLR    #nn:8,@ERm      CLR   [ERm].n           8 --------
  7Dm073n0           BTST    #nn:8,@ERm      TST   [ERm].n           8 -----z--
  7Eaa74n0        *i B{I}OR  #nn:8,@aa:8     OR    C,{not} [FFaa].n  6 -------c
  7Eaa75n0        *i B{I}XOR #nn:8,@aa:8     XOR   C,{not} [FFaa].n  6 -------c
  7Eaa76n0        *i B{I}AND #nn:8,@aa:8     AND   C,{not} [FFaa].n  6 -------c
  7Eaa77n0        *i B{I}LD  #nn:8,@aa:8     MOV   C,{not} [FFaa].n  6 -------c
  7Faa60n0           BSET    RnB,@aa:8       SET   [FFaa].RnB        8 --------
  7Faa61n0           BNOT    RnB,@aa:8       NOT   [FFaa].RnB        8 --------
  7Faa62n0           BCLR    RnB,@aa:8       CLR   [FFaa].RnB        8 --------
  7Faa63n0           BTST    RnB,@aa:8       TST   [FFaa].RnB        8 -----z--
  7Faa67n0        *i B{I}ST  #nn:8,@aa:8     MOV   [FFaa].n,{not} C  8 --------
  7Faa70n0           BSET    #nn:8,@aa:8     SET   [FFaa].n          8 --------
  7Faa71n0           BNOT    #nn:8,@aa:8     NOT   [FFaa].n          8 --------
  7Faa72n0           BCLR    #nn:8,@aa:8     CLR   [FFaa].n          8 --------
  7Faa73n0           BTST    #nn:8,@aa:8     TST   [FFaa].n          8 -----z--

Nintendo uses undocumented opcode 7Eaa73n0 instead of official 7Faa73n0. Unknown if both are working. Unknown if all other opcodes in 7EaaXXn0 range do also act like 7FaaXXn0, and perhaps also 7Cm0XXn0 like 7Dm0XXn0 (nintendo uses the offical 7D/7F for SET/NOT/CLR, so weirdness may apply for TST only).

H8/300H Encoding Notes

  *i   optional inverted source operand (when setting bit3 in the "n" digit)
  *s   must have bit3 set in "s" digit
  *m   must have bit3 set in "m" digit
  *?   must have bit3 set-or-not-set (has conflicting info in official specs)
  E?Rs meant to be ERs (although official specs omit the E in some cases)
  E?Rd meant to be ERd (although official specs omit the E in some cases)
  xxxS meant to be Silent, no flags affected (although specs say Sign Extend)
  xxxX meant to mean Carry, or meant to mean nothing specific in other cases

H8/300H Pseudo Opcodes
The official "PUSH/POP.W/L Rn" opcodes are normal MOV.W/L opcodes with ER7 (SP) and post-increment or pre-decrement.

  6DFn               PUSH.W Rn    ;MOV.W [ER7-],Rn
  6D7n               POP.W  Rn    ;MOV.W Rn,[ER7+]
  01006DFn           PUSH.L ERn   ;MOV.L [ER7-],ERn
  01006D7n           POP.L  ERn   ;MOV.L ERn,[ER7+]

Although not officially defined, one could also implement "PUSH/POP.W CCR".
There are no PUSH/POP.B opcodes (because that would misalign the stack).

H8/300H N/A

  ---N/A---          MOV.L @aa:8,ERd         MOV.L ERd,[FFaa]        - ----nz0-
  ---N/A---          MOV.L ERs,@aa:8         MOV.L [FFaa],ERs        - ----nz0-
  ---N/A---          MOV.W @aa:8,Rd          MOV.W Rd,[FFaa]         - ----nz0-
  ---N/A---          MOV.W Rs,@aa:8          MOV.W [FFaa],Rs         - ----nz0-
  ---N/A---          SUB.B #nn:8,RdB         SUB.B RdB,nn            - --h-nzvc

DS Cart Unknown Extras

DS Cartridges with built-in MicroSD Card Slot
The DS Vision cartridge contains a built-in microSD card slot. Users can download videos from internet (against a fee), store the videos on microSD cards, and then view them on the NDS via DS Vision cartridge.
Unknown how the microSD is accessed; via parallel 'ROM' bus and/or via serial SPI bus; by which commands? Also unknown if the thing contains built-in video decoder hardware, or if videos are decoded on ARM cpus.

DS Cartridge with Bluetooth Keyboard (Typing Adventure)
Mostly unknown. The components in the cartridge are:

  typical Macronix ROM
  STMicroelectronics M25PE10 SPI FLASH memory, presumably 128K
  Broadcom BCM2070 Bluetooth controller
  26MHz crystal oscillator

the FLASH chip isn't connected directly to the cart-SPI bus, instead it's prolly accessed through the Bluetooth controller.

NTR-UNSJ - Japanese TV Tuner
Some huge cartridge with digital TV receiver.
Unknown if the "U" in the gamecode means that it contains NAND, too.
DSi Launcher sets BPTWL[21h].bit0=0 for this cartridge (unknown what for).

NTR/TWL-Uxxx gamecodes (for carts with uncommon hardware)

  NTR-UNSJ      Japanese TV Tuner, with TV receiver
  NTR-UBRP      Nintendo DS Brower, with RAM cart in GBA slot
  NTR-UAMA      DS Vision Starter Kit, with microSD
  NTR-UEIJ      Starry Sky Navigation, with azimuth
  NTR/TWL-Uxxx  NAND carts (see NAND chapter)

DS Cart Cheat Action Replay DS

The first commercial DS cheat code solution, this device was developed by Datel. It supports swapping out cartridges after loading the AR software. For updating, the user may either manually enter codes or use the included proprietary USB cable that comes with the device. The user has been able to manually update codes since firmware version 1.52.

Action Replay DS Codes

  ABCD-NNNNNNNN       Game ID ;ASCII Gamecode [00Ch] and CRC32 across [0..1FFh]
  00000000 XXXXXXXX   manual hook codes (rarely used) (default is auto hook)
  0XXXXXXX YYYYYYYY   word[XXXXXXX+offset] = YYYYYYYY
  1XXXXXXX 0000YYYY   half[XXXXXXX+offset] = YYYY
  2XXXXXXX 000000YY   byte[XXXXXXX+offset] = YY
  3XXXXXXX YYYYYYYY   IF YYYYYYYY > word[XXXXXXX]   ;unsigned    ;\
  4XXXXXXX YYYYYYYY   IF YYYYYYYY < word[XXXXXXX]   ;unsigned    ; for v1.54,
  5XXXXXXX YYYYYYYY   IF YYYYYYYY = word[XXXXXXX]                ; when X=0,
  6XXXXXXX YYYYYYYY   IF YYYYYYYY <> word[XXXXXXX]               ; uses
  7XXXXXXX ZZZZYYYY   IF YYYY > ((not ZZZZ) AND half[XXXXXXX])   ; [offset]
  8XXXXXXX ZZZZYYYY   IF YYYY < ((not ZZZZ) AND half[XXXXXXX])   ; instead of
  9XXXXXXX ZZZZYYYY   IF YYYY = ((not ZZZZ) AND half[XXXXXXX])   ; [XXXXXXX]
  AXXXXXXX ZZZZYYYY   IF YYYY <> ((not ZZZZ) AND half[XXXXXXX])  ;/
  BXXXXXXX 00000000   offset = word[XXXXXXX+offset]
  C0000000 YYYYYYYY   FOR loopcount=0 to YYYYYYYY  ;execute Y+1 times
  C4000000 00000000   offset = address of the C4000000 code           ;v1.54
  C5000000 XXXXYYYY   counter=counter+1, IF (counter AND YYYY) = XXXX ;v1.54
  C6000000 XXXXXXXX   [XXXXXXXX]=offset                               ;v1.54
  D0000000 00000000   ENDIF
  D1000000 00000000   NEXT loopcount
  D2000000 00000000   NEXT loopcount, and then FLUSH everything
  D3000000 XXXXXXXX   offset = XXXXXXXX
  D4000000 XXXXXXXX   datareg = datareg + XXXXXXXX
  D5000000 XXXXXXXX   datareg = XXXXXXXX
  D6000000 XXXXXXXX   word[XXXXXXXX+offset]=datareg, offset=offset+4
  D7000000 XXXXXXXX   half[XXXXXXXX+offset]=datareg, offset=offset+2
  D8000000 XXXXXXXX   byte[XXXXXXXX+offset]=datareg, offset=offset+1
  D9000000 XXXXXXXX   datareg = word[XXXXXXXX+offset]
  DA000000 XXXXXXXX   datareg = half[XXXXXXXX+offset]
  DB000000 XXXXXXXX   datareg = byte[XXXXXXXX+offset] ;bugged on pre-v1.54
  DC000000 XXXXXXXX   offset = offset + XXXXXXXX
  EXXXXXXX YYYYYYYY   Copy YYYYYYYY parameter bytes to [XXXXXXXX+offset...]
  44332211 88776655   parameter bytes 1..8 for above code  (example)
  0000AA99 00000000   parameter bytes 9..10 for above code (padded with 00s)
  FXXXXXXX YYYYYYYY   Copy YYYYYYYY bytes from [offset..] to [XXXXXXX...]

IF/ENDIF can be nested up to 32 times. FOR/NEXT cannot be nested, any FOR statement does forcefully terminate any prior loop. FOR does backup the current IF condidition flags, and NEXT does restore these flags, so ENDIF(s) aren't required inside of the loop. The NEXT+FLUSH command does (after finishing the loop) reset offset=0, datareg=0, and does clear all condition flags, so further ENDIF(s) aren't required after the loop.
Before v1.54, the DB000000 code did accidently set offset=offset+XXXXXXX after execution of the code. For all word/halfword accesses, the address should be aligned accordingly. For the COPY commands, addresses should be aligned by four (all data is copied with ldr/str, except, on odd lengths, the last 1..3 bytes do use ldrb/strb).
offset, datareg, loopcount, and counter are internal registers in the action replay software.

> The condition register is checked, for all code types
> but the D0, D1 and D2 code type
Makes sense.

> and for the C5 code type it's checked AFTER the counter has
> been incremented (so the counter is always incremented
I love that exceptions ;-)

Hook
The hook codes consist of a series of nine 00000000 XXXXXXXX codes, and must be marked as (M) code (for not being confused with normal 0XXXXXXX YYYYYYYY codes). For all nine codes, the left 32bit are actually don't care (but should be zero), the meaning of the right 32bit varies from 1st to 9th code.

  1st: Address used prior to launching game (eg. 23xxxxxh)
  2nd: Address to write the hook at (inside the ARM7 executable)
  3rd: Hook final address (huh?)
  4th: Hook mode selection (0=auto, 1=mode1, 2=mode2)
  5th: Opcode that replaces the hooked one (eg. E51DE004h)
  6th: Address to store important stuff  (default 23FE000h)
  7th: Address to store the code handler (default 23FE074h)
  8th: Address to store the code list    (default 23FE564h)
  9th: Must be 1 (00000001h)

For most games, the AR does automatically hook code on the ARM7. Doing that automatically is nice, but hooking ARM7 means that there is no access to VRAM, TCM and Cache, which <might> cause problems since efficient games <should> store all important data in TCM or Cache (though, in practice, I'd doubt that any existing NDS games are that efficient).

Thanks
To Kenobi and Dualscreenman from Kodewerx for above ARDS cheat info.

DS Cart Cheat Codebreaker DS

This is Pelican's entry into the DS cheat-device industry. It supports swapping out the cartridges, and alternately, also gives the user the option of connecting another gamecard onto it. For updating, the user may either manually enter codes, or use Wifi to connect to the Codebreaker update site (that updating will overwrite all manually entered codes though).

Codebreaker DS Codes

  ---Initialization---
  0000CR16 GAMECODE                    Specify Game ID, use Encrypted codes
  8000CR16 GAMECODE                    Specify Game ID, use Unencrypted codes
  BEEFC0DE XXXXXXXX                    Change Encryption Keys
  A0XXXXXX YYYYYYYY                    Bootup-Hook 1, X=Address, Y=Value
  A8XXXXXX YYYYYYYY                    Bootup-Hook 2, X=Address, Y=Value
  F0XXXXXX TYYYYYYY         Code-Hook 1 (T=Type,Y=CheatEngineAddr,X=HookAddr)
  F8XXXXXX TPPPPPPP         Code-Hook 2 (T=Type,X=CheatEngineHookAddr,P=Params)
  ---General codes---
  00XXXXXX 000000YY                    [X]=YY
  10XXXXXX 0000YYYY                    [X]=YYYY
  20XXXXXX YYYYYYYY                    [X]=YYYYYYYY
  60XXXXXX 000000YY ZZZZZZZZ 00000000  [[X]+Z]=YY
  60XXXXXX 0000YYYY ZZZZZZZZ 10000000  [[X]+Z]=YYYY
  60XXXXXX YYYYYYYY ZZZZZZZZ 20000000  [[X]+Z]=YYYYYYYY
  30XXXXXX 000000YY                    [X]=[X] + YY
  30XXXXXX 0001YYYY                    [X]=[X] + YYYY
  38XXXXXX YYYYYYYY                    [X]=[X] + YYYYYYYY
  70XXXXXX 000000YY                    [X]=[X] OR  YY
  70XXXXXX 001000YY                    [X]=[X] AND YY
  70XXXXXX 002000YY                    [X]=[X] XOR YY
  70XXXXXX 0001YYYY                    [X]=[X] OR  YYYY
  70XXXXXX 0011YYYY                    [X]=[X] AND YYYY
  70XXXXXX 0021YYYY                    [X]=[X] XOR YYYY
  ---Memory fill/copy---
  40XXXXXX 2NUMSTEP 000000YY 000000ZZ  byte[X+(0..NUM-1)*STEP*1]=Y+(0..NUM-1)*Z
  40XXXXXX 1NUMSTEP 0000YYYY 0000ZZZZ  half[X+(0..NUM-1)*STEP*2]=Y+(0..NUM-1)*Z
  40XXXXXX 0NUMSTEP YYYYYYYY ZZZZZZZZ  word[X+(0..NUM-1)*STEP*4]=Y+(0..NUM-1)*Z
  50XXXXXX YYYYYYYY ZZZZZZZZ 00000000  copy Y bytes from [X] to [Z]
  ---Conditional codes (bugged)---
  60XXXXXX 000000YY ZZZZZZZZ 01c100VV  IF [[X]+Z] .. VV   THEN [[X]+Z]=YY
  60XXXXXX 000000YY ZZZZZZZZ 01c0VVVV  IF [[X]+Z] .. VVVV THEN [[X]+Z]=YY
  60XXXXXX 0000YYYY ZZZZZZZZ 11c100VV  IF [[X]+Z] .. VV   THEN [[X]+Z]=YYYY
  60XXXXXX 0000YYYY ZZZZZZZZ 11c0VVVV  IF [[X]+Z] .. VVVV THEN [[X]+Z]=YYYY
  60XXXXXX YYYYYYYY ZZZZZZZZ 21c100VV  IF [[X]+Z] .. VV   THEN [[X]+Z]=YYYYYYYY
  60XXXXXX YYYYYYYY ZZZZZZZZ 21c0VVVV  IF [[X]+Z] .. VVVV THEN [[X]+Z]=YYYYYYYY
  ---Conditional codes (working)---
  D0XXXXXX NNc100YY                    IF [X] .. YY   THEN exec max(1,NN) lines
  D0XXXXXX NNc0YYYY                    IF [X] .. YYYY THEN exec max(1,NN) lines

The condition digits (c=0..7), have the following functions:

  0 IF [mem] =  imm THEN ...              4 IF ([mem] AND imm) =  0   THEN ...
  1 IF [mem] <> imm THEN ...              5 IF ([mem] AND imm) <> 0   THEN ...
  2 IF [mem] <  imm THEN ... (unsigned)   6 IF ([mem] AND imm) =  imm THEN ...
  3 IF [mem] >  imm THEN ... (unsigned)   7 IF ([mem] AND imm) <> imm THEN ...

Notes

  GAMECODE  Cartridge Header[00Ch] (32bit in reversed byte-order)
  CR16      Cartridge Header[15Eh] (16bit in normal byte-order)
  XXXXXX    27bit addr (actually 7 digits, XXXXXXX, overlaps 5bit code number)

The "bugged" conditional codes (60XXXXXX) are accidently skipping NN lines when the condition is false, where NN is taken from the upper 8bit of the code's last 32bit values (ie. exactly as for the D0XXXXXX codes). For byte-writes, that would be NN=01h, which can be eventually dealt with, although there may be compatibility problems which future versions that might fix that bug. For halfword/word writes, NN would be 11h or 21h, so that codes are about totally unusable.

Codebreaker DS / Encrypted Codes
The overall "address value" decryption works like so:

  for i=4Fh to 00h
    y=77628ECFh
    if i>13h then y=59E5DC8Ah
    if i>27h then y=054A7818h
    if i>3Bh then y=B1BF0855h
    address = (Key0-value) xor address
    value   = value - Key1 - (address ror 1Bh)
    address = (address xor (value + y)) ror 13h
    if (i>13h) then
      if (i<=27h) or (i>3Bh) then x=Key2 xor Key1 xor Key0
      else x=((Key2 xor Key1) and Key0) xor (Key1 and Key2)
      value=value xor (x+y+address)
      x = Secure[((i*4+00h) and FCh)+000h]
      x = Secure[((i*4+34h) and FCh)+100h] xor x
      x = Secure[((i*4+20h) and FCh)+200h] xor x
      x = Secure[((i*4+08h) and FCh)+300h] xor x
      address = address - (x ror 19h)
  next i

Upon startup, the initial key settings are:

  Secure[0..7FFh] = Copy of the ENCRYPTED 1st 2Kbytes of the game's Secure Area
  Key0 = 0C2EAB3Eh, Key1 = E2AE295Dh, Key2 = E1ACC3FFh, Key3 = 70D3AF46h
  scramble_keys

Upon BEEFC0DE XXXXXXXX, the keys get changed like so:

  Key0 = Key0 + (XXXXXXXX ror 1Dh)
  Key1 = Key1 - (XXXXXXXX ror 05h)
  Key2 = Key2 xor (Key3 xor Key0)
  Key3 = Key3 xor (Key2  -  Key1)
  scramble_keys

The above scramble_keys function works like so:

  for i=0 to FFh
    y = byte(xlat_table[i])
    Secure[i*4+000h] = (Secure[i*4+000h] xor Secure[y*4]) + Secure[y*4+100h]
    Secure[i*4+400h] = (Secure[i*4+400h] xor Secure[y*4]) - Secure[y*4+200h]
  next i
  for i=0 to 63h
    Key0 = Key0 xor (Secure[i*4] + Secure[i*4+190h])
    Key1 = Key1 xor (Secure[i*4] + Secure[i*4+320h])
    Key2 = Key2 xor (Secure[i*4] + Secure[i*4+4B0h])
    Key3 = Key3 xor (Secure[i*4] + Secure[i*4+640h])
  next i
  Key0 = Key0  -  Secure[7D0h]
  Key1 = Key1 xor Secure[7E0h]
  Key2 = Key2  +  Secure[7F0h]
  Key3 = Key3 xor Secure[7D0h] xor Secure[7F0h]

the xlat_table consists of 256 fixed 8bit values:

  34h,59h,00h,32h,7Bh,D3h,32h,C9h,9Bh,77h,75h,44h,E0h,73h,46h,06h
  0Bh,88h,B3h,3Eh,ACh,F2h,BAh,FBh,2Bh,56h,FEh,7Ah,90h,F7h,8Dh,BCh
  8Bh,86h,9Ch,89h,00h,19h,CDh,4Ch,54h,30h,01h,93h,30h,01h,FCh,36h
  4Dh,9Fh,FDh,D7h,32h,94h,AEh,BCh,2Bh,61h,DFh,B3h,44h,EAh,8Bh,A3h
  2Bh,53h,33h,54h,42h,27h,21h,DFh,A9h,DDh,C0h,35h,58h,EFh,8Bh,33h
  B4h,D3h,1Bh,C7h,93h,AEh,32h,30h,F1h,CDh,A8h,8Ah,47h,8Ch,70h,0Ch
  17h,4Eh,0Eh,A2h,85h,0Dh,6Eh,37h,4Ch,39h,1Fh,44h,98h,26h,D8h,A1h
  B6h,54h,F3h,AFh,98h,83h,74h,0Eh,13h,6Eh,F4h,F7h,86h,80h,ECh,8Eh
  EEh,4Ah,05h,A1h,F1h,EAh,B4h,D6h,B8h,65h,8Ah,39h,B3h,59h,11h,20h
  B6h,BBh,4Dh,88h,68h,24h,12h,9Bh,59h,38h,06h,FAh,15h,1Dh,40h,F0h
  01h,77h,57h,F5h,5Dh,76h,E5h,F1h,51h,7Dh,B4h,FAh,7Eh,D6h,32h,4Fh
  0Eh,C8h,61h,C1h,EEh,FBh,2Ah,FCh,ABh,EAh,97h,D5h,5Dh,E8h,FAh,2Ch
  06h,CCh,86h,D2h,8Ch,10h,D7h,4Ah,CEh,8Fh,EBh,03h,16h,ADh,84h,98h
  F5h,88h,2Ah,18h,ACh,7Fh,F6h,94h,FBh,3Fh,00h,B6h,32h,A2h,ABh,28h
  64h,5Ch,0Fh,C6h,23h,12h,0Ch,D2h,BAh,4Dh,A3h,F2h,C9h,86h,31h,57h
  0Eh,F8h,ECh,E1h,A0h,9Ah,3Ch,65h,17h,18h,A0h,81h,D0h,DBh,D5h,AEh

all used operations are unsigned 32bit integer.

Thanks
To Kenobi and Dualscreenman from Kodewerx for above CBDS cheat info.

DS Cart DLDI Driver

DLDI (Dynamically Linked Device Interface for libfat) is a popular yet undocumented flashcart driver for homebrew NDS software dating back to 2006. Below was reverse-engineered 11/2018.

file.dldi --> driver file (can be small like 1.5Kbyte, or max 32Kbyte)
file.nds --> ROM image (must contain 32Kbyte space with DLDI ID for patching)

Driver patch file standard header

  00h 4  DLDI ID       (EDh,A5h,8Dh,BFh) (aka BF8DA5EDh)  ;\patching tools will
  04h 8  DLDI String   (20h,"Chishm",00h)                 ; refuse any other
  0Ch 1  DLDI Version  (01h in .dldi, don't care in .nds) ;/values
  0Dh 1  Size of .dldi+BSS (rounded up to 1 SHL N bytes) (max 0Fh=32Kbytes)
  0Eh 1  Sections to fix/destroy (see FIX_xxx)
  0Fh 1  Space in .nds file (1 SHL N) (0Eh..0Fh in .nds, can be 0 in .dldi)
  10h 48 ASCII Full Driver Name (max 47 chars, plus zero padding)
  40h 4  Address of ALL start (text) ;-base address (BF800000h in .dldi)
  44h 4  Address of ALL end   (data) ;-for highly-unstable FIX_ALL addr.adjusts
  48h 4  Address of GLUE start       ;\for semi-stable FIX_GLUE addr.adjusts
  4Ch 4  Address of GLUE end         ;/  ("Interworking glue" for ARM-vs-THUMB)
  50h 4  Address of GOT start        ;\for semi-stable FIX_GOT addr.adjusts
  54h 4  Address of GOT end          ;/  ("Global Offset Table")
  58h 4  Address of BSS start        ;\for zerofilling "BSS" via FIX_BSS
  5Ch 4  Address of BSS end          ;/  ("Block Started by Symbol")
  60h 4  ASCII Short Driver/Device Name (4 chars, eg. "MYHW" for MyHardware)
  64h 4  Flags 2 (see FEATURE_xxx) (usually 13h=GbaSlot, or 23h=NdsSlot)
  68h 4  Address of Function startup() ;<-- must be at offset +80h !! ;\
  6Ch 4  Address of Function isInserted() ;out: 0=no/fail, 1=yes/okay ; all
  70h 4  Address of Function readSectors(sector,numSectors,buf)       ; return
  74h 4  Address of Function writeSectors(sector,numSectors,buf)      ; 0=fail,
  78h 4  Address of Function clearStatus()                            ; 1=okay
  7Ch 4  Address of Function shutdown()                               ;/
  80h .. Driver Code (can/must begin with "startup()")            ;\max 7F80h
  ..  .. Glue section (usually a small snippet within above code) ; bytes (when
  ..  .. GOT section (usually after above code) (pointer table)   ; having 32K
  ..  .. BSS section (usually at end, may exceed .dldi filesize)  ; allocated)
  ..  .. Optional two garbage NOPs at end of default.dldi         ;/

hdr[0Eh] - Sections to fix/destroy (FIX_xxx):

  0    FIX_ALL   ;-installer uses highly-unstable guessing in whole dldi file
  1    FIX_GLUE  ;-installer uses semi-stable address guessing in GLUE area
  2    FIX_GOT   ;-installer uses semi-stable address guessing in GOT area
  3    FIX_BSS   ;-installer will zerofill BSS area
  4-7  Reserved (0)

hdr[64h] - Flags 2 (FEATURE_xxx) (usually 13h=GbaSlot, or 23h=NdsSlot):

  0    FEATURE_MEDIUM_CANREAD          00000001h (usually set)
  1    FEATURE_MEDIUM_CANWRITE         00000002h (a few carts can't write)
  2-3  Reserved (0)
  4    FEATURE_SLOT_GBA                00000010h (need EXMEMCNT bit7 adjusted)
  5    FEATURE_SLOT_NDS                00000020h (need EXMEMCNT bit11 adjusted)
  6-31 Reserved (0)

Note: The allocated driver size in hdr[0Fh] was 0Fh=32Kbytes between 2006 and 2016, however, libnds has changed that to 0Eh=16Kbytes in January 2017 (maybe intending to free more RAM, especially when using ARM7 WRAM).
However, there's at least one driver exceeding 16K (rpg_nand.dldi in AKAIO package; the driver disguises itself as 8K driver in hdr[0Dh], but its BSS area actually needs ways MORE than 16K).

Required entries in .nds file
Officially, dldi could be at any 4-byte aligned location, however, for faster lookup, better use this locations:

  dldi area should be located at a 40h-byte aligned address in ROM image.
  dldi area should be located in ARM9 (or ARM7) bootcode area.

An "empty" driver needs to contain:

  dldi[00h..0Bh] must contain DLDI ID word/string
  dldi[0Fh]      must contain allocated size (0Eh=16Kbyte or 0Fh=32Kbyte)
  dldi[40h..43h] must contain RAM base address of DLDI block
  and other entries should contain valid dummy strings and dummy functions.

An installed driver should contain a copy of the .dldi file, with addresses adjusted to RAM locations, and BSS area zerofilled (if FIX_BSS was set)

  dldi[0Fh] must be kept as in the old .nds file (not as in .dldi file)

Some installers might try to detect homebrew by looking at nds carthdr for deciding whether or not to try to install dldi (unknown if/which ones are doing such things and looking at which carthdr entries).

Functions
startup, isInserted, clearStatus, shutdown can be dummy functions that do nothing (other than returning r0=1=okay).
Alternately startup/shutdown can initialize or power down the hardware, clearStatus is meant to be some sort of soft reset, and isInserted is allowing to test if the SD card is inserted & working.
read/write sectors are reading/writing one or more sectors. Sector size is 200h bytes, sector numbers is 0=First 200h bytes, 1=Next 200h bytes, and so on.
buf should be usually 4-byte aligned, however, some drivers do also support unaligned buffers using slower transfer code (better implement that when making .dldi drivers, but better don't rely on it being supported when making .nds files).
The driver functions can support SD and SDHC (or the flashcart manufacturer might release driver updates if SDHC wasn't supported).
Higher level FAT functions must be contained in the .nds file (so a driver update won't help if the .nds file lacks FAT32) (and ExFAT most unlikely to be supported).
Functions should be ARM7 compatible, ie. don't use BLX or POP r15 for mode switching, so the driver can be used on both ARM9 and ARM7 (or even on GBA).

SLOT_GBA/NDS
SLOT_GBA/NDS seem to relate to GBA and NDS slots, the driver can probably have only one of the SLOT bits set (the functions don't allow to select which slot to use).
Purpose is unclear to me, maybe just telling the .nds file that the flashcart is in the given slot (and thereby shouldn't expect other hardware in that slot). Or maybe telling telling the installer which hardware the driver is supposed for.

FIX_xxx does maybe relate to address adjustments made by the dldi installer. Unknown if/how that's working.
DS Cart DLDI Driver - Guessed Address-Adjustments

Some DLDI flashcarts support extra features like Rumble. However, that extra hardware is accessed via direct I/O, not via DLDI driver. Unknown which I/O ports are used for that stuff - probably something compatible with official GBA/NDS rumble cart(s).

DS Cart DLDI Driver - Guessed Address-Adjustments

The DLDI installer uses some guessing method for address-adjustments (the FIX_xxx flags are supposed to patch addresses, but not opcodes or other data).

Unaligned-Word patching and over-shooting bug
The central mistake in the official DLDI installer is that it is patching all words at [start..end-1], using 1-byte address increments instead of 4-byte increments. This includes patching words at non-word aligned locations, or patching words whose lower bytes were already patched, and over-shooting to words at [end-1..end+2].

ddmem Base BF800000h
Most or all .dldi files are using ddmem base BF800000h (defined in dldi.ld). That value was chosen because it won't conflict with opcodes (as NDS BIOS doesn't use SWI function 80h, BF80xxxxh would be an invalid SWI function in ARM; and BF80h would be an invalid opcode in THUMB).
So far, this would have worked well, but it doesn't work with unaligned-word patching bug (eg. THUMB opcodes 8000h,BF00h, or ARM opcodes xxBF80xxh or the like). And, even if it would have worked for opcodes - it might still fail for data values.

FIX_GOT
GOT does usually contain BF80xxxxh address pointers (plus some 00000000h words). The guessing works quite stable (the maximum for 32K files is xxxxh=7FFFh, so there's no risk to encounter xxxx=BF80h) (one could encounter xxxxh=xxBFh, but the previously patched word is usually in RAM area, eg. 02xxxxxxh, so this would form BF02h, without risk to be seen as BF80h).
BUG: The GOT table is usually located at the end of the .dldi file, meaning that the over-shooting bug will see three uninitialized bytes at [got_end+0..2], and may go amok if they are BF80xxh or xxBF80h. The value of those bytes depends on left-over from previously installed .dldi driver(s) and on the ddmem base used in the .nds file, so the bug may take place randomly depending on several factors.

FIX_GLUE
GLUE does usually contain a handful of opcodes and .pool values for switching between ARM and THUMB code. The intention is to patch the addresses in the .pool, and to leave the opcodes intact. This can be potentially stable, assuming that the used opcodes in the GLUE (and the next three bytes after glue_end) usually won't contain BF80h).

FIX_ALL
This is the mother of all bugs. Fortunately there aren't any .dldi drivers with FIX_ALL flag - and one should never make drivers that do use it.
ALL is covering the whole dldi space, including the 80h-byte DLDI header, the code area, including GLUE area, and GOT area, and probably also the yet uninitialized BSS area, and the next three bytes after end_all.
Patching the whole code area means an increased risk to hit opcodes or data values that contain BF80h. The over-shooting bug may even destroy the next three bytes after the 32Kbyte area.
Patching the DLDI header could destroy the header itself, the header in the .dldi file usually won't contain BF80h at unintended locations, however, the pointers in that header are adjusted before applying FIX_ALL, for example, RAM base 0200BF00h (in .nds file) combined with a function at BF800080h (in .dldi file) would result in 0200BF80h. The nasty thing is that the problem won't occur with other RAM base values (in other .nds files).

Avoiding the bugs
When making .dldi drivers: Never use FIX_ALL. And preferably avoid FIX_GOT and FIX_GLUE as well (ARM CPU can do relative jumps and relative addressing, eg. via ADR and ADRL pseudo opcodes, so there's no point in needing address adjustments). Or otherwise append padding after GOT area, and try to avoid using opcodes/data with BF80h in/after GLUE area.
When making dldi installers: Best patch only word-aligned words (ARM CPU can't access unaligned data, so there's little chance that DLDI drivers would ever contain unaligned words). Or, when maintaining unaligned patching: At the very least skip the 80h-byte header on FIX_ALL, and after patching a word at other locations, skip the next three bytes, and don't do the over-shooting at end.
When making .nds files: There isn't too much that could be done here. One could set ddmem to 64Kbyte aligned addresses (so functions won't end up at xxxxBF80h). Or one could even set ddmem to BF800000h (so patching will leave everything unchanged & intact - so one could then do the address adjustments manually, and hopefully more reliable than other DLDI installers).

DS Encryption by Gamecode/Idcode (KEY1)

KEY1 - Gamecode / Idcode Encryption
The KEY1 encryption relies only on the gamecode (or firmware idcode), it does not contain any random components. The fact that KEY1 encrypted commands appear random is just because the <unencrypted> commands contain random values, so the encryption result looks random.

KEY1 encryption is used for KEY1 encrypted gamecart commands (ie. for loading the secure area). It is also used for resolving the extra decryption of the first 2K of the secure area, and for firmware decryption, and to decode some encrypted values in gamecart/firmware header.

Initial Encryption Values
Below formulas can be used only with a copy of the 1048h-byte key tables from NDS/DSi BIOS. The values can be found at:

  NDS.ARM7 ROM: 00000030h..00001077h (values 99 D5 20 5F ..) Blowfish/NDS-mode
  DSi.ARM9 ROM: FFFF99A0h..FFFFA9E7h (values 99 D5 20 5F ..) ""
  DSi.TCM Copy: 01FFC894h..01FFD8DBh (values 99 D5 20 5F ..) ""
  DSi.ARM7 ROM: 0000C6D0h..0000D717h (values 59 AA 56 8E ..) Blowfish/DSi-mode
  DSi.RAM Copy: 03FFC654h..03FFD69Bh (values 59 AA 56 8E ..) ""
  DSi.Debug:    (stored in launcher) (values 69 63 52 05 ..) Blowfish/DSi-debug

The DSi ROM sections are disabled after booting, but the RAM/TCM copies can be dumped (eg. with some complex main memory hardware mods, or via unlaunch exploit). The DSi.Debug key is stored in launcher, and it's used when SCFG_OP is nonzero (as so on debugging on hardware).

encrypt_64bit(ptr) / decrypt_64bit(ptr)

  Y=[ptr+0]
  X=[ptr+4]
  FOR I=0 TO 0Fh (encrypt), or FOR I=11h TO 02h (decrypt)
    Z=[keybuf+I*4] XOR X
    X=[keybuf+048h+((Z SHR 24) AND FFh)*4]
    X=[keybuf+448h+((Z SHR 16) AND FFh)*4] + X
    X=[keybuf+848h+((Z SHR  8) AND FFh)*4] XOR X
    X=[keybuf+C48h+((Z SHR  0) AND FFh)*4] + X
    X=Y XOR X
    Y=Z
  NEXT I
  [ptr+0]=X XOR [keybuf+40h] (encrypt), or [ptr+0]=X XOR [keybuf+4h] (decrypt)
  [ptr+4]=Y XOR [keybuf+44h] (encrypt), or [ptr+4]=Y XOR [keybuf+0h] (decrypt)

apply_keycode(modulo)

  encrypt_64bit(keycode+4)
  encrypt_64bit(keycode+0)
  [scratch]=0000000000000000h   ;S=0 (64bit)
  FOR I=0 TO 44h STEP 4         ;xor with reversed byte-order (bswap)
    [keybuf+I]=[keybuf+I] XOR bswap_32bit([keycode+(I MOD modulo)])
  NEXT I
  FOR I=0 TO 1040h STEP 8
    encrypt_64bit(scratch)      ;encrypt S (64bit) by keybuf
    [keybuf+I+0]=[scratch+4]    ;write S to keybuf (first upper 32bit)
    [keybuf+I+4]=[scratch+0]    ;write S to keybuf (then lower 32bit)
  NEXT I

init_keycode(idcode,level,modulo,key)

  if key=nds then copy [nds_arm7bios+0030h..1077h] to [keybuf+0..1047h]
  if key=dsi then copy [dsi_arm7bios+C6D0h..D717h] to [keybuf+0..1047h]
  [keycode+0]=[idcode]
  [keycode+4]=[idcode]/2
  [keycode+8]=[idcode]*2
  IF level>=1 THEN apply_keycode(modulo) ;first apply (always)
  IF level>=2 THEN apply_keycode(modulo) ;second apply (optional)
  [keycode+4]=[keycode+4]*2
  [keycode+8]=[keycode+8]/2
  IF level>=3 THEN apply_keycode(modulo) ;third apply (optional)

firmware_decryption

  init_keycode(firmware_header+08h,1,0Ch,nds) ;idcode (usually "MACP"), level 1
  decrypt_64bit(firmware_header+18h)          ;rominfo
  init_keycode(firmware_header+08h,2,0Ch,nds) ;idcode (usually "MACP"), level 2
  decrypt ARM9 and ARM7 bootcode by decrypt_64bit (each 8 bytes)
  decompress ARM9 and ARM7 bootcode by LZ77 function (swi)
  calc CRC16 on decrypted/decompressed ARM9 bootcode followed by ARM7 bootcode

Note: The sizes of the compressed/encrypted bootcode areas are unknown (until they are fully decompressed), one way to solve that problem is to decrypt the next 8 bytes each time when the decompression function requires more data.

gamecart_decryption

  init_keycode(cart_header+0Ch,1,08h,nds)   ;gamecode, level 1, modulo 8
  decrypt_64bit(cart_header+78h)            ;rominfo (secure area disable)
  init_keycode(cart_header+0Ch,2,08h,nds)   ;gamecode, level 2, modulo 8
  encrypt_64bit all NDS KEY1 commands (1st command byte in MSB of 64bit value)
  after loading the secure_area, calculate secure_area crc, then
  decrypt_64bit(secure_area+0)              ;first 8 bytes of secure area
  init_keycode(cart_header+0Ch,3,08h,nds)   ;gamecode, level 3, modulo 8
  decrypt_64bit(secure_area+0..7F8h)        ;each 8 bytes in first 2K of secure
  init_keycode(cart_header+0Ch,1,08h,dsi)   ;gamecode, level 1, modulo 8
  encrypt_64bit all DSi KEY1 commands (1st command byte in MSB of 64bit value)

After secure area decryption, the ID field in the first 8 bytes should be "encryObj", if it matches then first 8 bytes are filled with E7FFDEFFh, otherwise the whole 2K are filled by that value.

Gamecart Command Register
Observe that the byte-order of the command register [40001A8h] is reversed. The way how the CPU stores 64bit data in memory (and the way how the "encrypt_64bit" function for KEY1-encrypted commands expects data in memory) is LSB at [addr+0] and MSB at [addr+7]. This value is to be transferred MSB first. However, the DS hardware transfers [40001A8h+0] first, and [40001A8h+7] last. So, the byte order must be reversed when copying the value from memory to the command register.

Note
The KEY1 encryption is based on Bruce Schneier's "Blowfish Encryption Algorithm".

DS Encryption by Random Seed (KEY2)

KEY2 39bit Seed Values
The pre-initialization settings at cartridge-side (after reset) are:

  Seed0 = 58C56DE0E8h
  Seed1 = 5C879B9B05h

The post-initialization settings (after sending command 4llllmmmnnnkkkkkh to the cartridge, and after writing the Seed values to Port 40001Bxh) are:

  Seed0 = (mmmnnn SHL 15)+6000h+Seedbyte
  Seed1 = 5C879B9B05h

The seedbyte is selected by Cartridge Header [013h].Bit0-2, this index value should be usually in range 0..5, however, possible values for index 0..7 are: E8h,4Dh,5Ah,B1h,17h,8Fh,99h,D5h.
The 24bit random value (mmmnnn) is derived from the real time clock setting, and also scattered by KEY1 encryption, anyways, it's just random and doesn't really matter where it comes from.

KEY2 Encryption
Relies on two 39bit registers (x and y), which are initialized as such:

  x = reversed_bit_order(seed0)  ;ie. LSB(bit0) exchanged with MSB(bit38), etc.
  y = reversed_bit_order(seed1)

During transfer, x, y, and transferred data are modified as such:

  x = (((x shr 5)xor(x shr 17)xor(x shr 18)xor(x shr 31)) and 0FFh)+(x shl 8)
  y = (((y shr 5)xor(y shr 23)xor(y shr 18)xor(y shr 31)) and 0FFh)+(y shl 8)
  data = (data xor x xor y) and 0FFh

DS Firmware Serial Flash Memory

ST Microelectronics SPI Bus Compatible Serial FLASH Memory

 Chips used as wifi-flash:
  ID 20h,40h,12h - ST M45PE20 - 256 KBytes (Nintendo DS) (in my old DS)
  ID 20h,50h,12h - ST M35PE20 - 256 KBytes (Nintendo DS) (in my DS-Lite)
  ID 20h,80h,13h - ST M25PE40 - 512 KBytes (iQue DS, with chinese charset)
  ID 20h,40h,11h - ST 45PE10V6 - 128 Kbytes (Nintendo DSi) (in my DSi)
  ID 20h,58h,0Ch?- 5A32        - 4 Kbytes (Nintendo DSi, newer models)
  ID ?           - 26FV032T    - (Nintendo DSi, J27H020) (this has big package)
  ID ?           - 5K32        - (3DS?)
  ID 62h,62h,0Ch - 32B, 3XH    - 4 Kbytes (New3DS)
 Other similar chips (used in game cartridges):
  ID 20h,40h,13h - ST 45PE40V6 - 512 KBytes (DS Zelda, NTR-AZEP-0)
  ID 20h,40h,14h - ST 45PE80V6 - 1024 Kbytes (eg. Spirit Tracks, NTR-BKIP)
 +ID 62h,11h,00h - Sanyo ?          - 512 Kbytes (P-Letter Diamond, ADAE)
  ID 62h,16h,00h - Sanyo LE25FW203T - 256 KBytes (Mariokart backup)
 +ID 62h,26h,11h - Sanyo ?          - ? Kbytes (3DS: CTR-P-AXXJ)
 +ID 62h,26h,13h - Sanyo ?          - ? Kbytes (3DS: CTR-P-APDJ)
  ID C2h,22h,11h - Macronix MX25L1021E? 128 Kbytes (eg. 3DS Starfox)
  ID C2h,22h,13h - Macronix ...? 512 Kbytes (eg. 3DS Kid Icarus, 3DS Sims 3)
  ID C2h,20h,17h - Macronix MX25L6445EZNI-10G 8192 Kbytes (DSi Art Academy)
  ID 01h,F0h,00h - Garbage/Infrared on SPI-bus? (eg. P-Letter White)
  ID 03h,F8h,00h - Garbage/Infrared on SPI-bus? (eg. P-Letter White 2)

FLASH has more than 100,000 Write Cycles, more than 20 Year Data Retention
The Firmware Flash Memory is accessed via SPI bus,
DS Serial Peripheral Interface Bus (SPI)

Instruction Codes

  06h  WREN Write Enable (No Parameters)
  04h  WRDI Write Disable (No Parameters)
  9Fh  RDID Read JEDEC Identification (Read 1..3 ID Bytes)
             (Manufacturer, Device Type, Capacity)
  05h  RDSR Read Status Register (Read Status Register, endless repeated)
             Bit7-2  Not used (zero)
             Bit1    WEL Write Enable Latch             (0=No, 1=Enable)
             Bit0    WIP Write/Program/Erase in Progess (0=No, 1=Busy)
  03h  READ Read Data Bytes (Write 3-Byte-Address, read endless data stream)
  0Bh  FAST Read Data Bytes at Higher Speed (Write 3-Byte-Address, write 1
             dummy-byte, read endless data stream) (max 25Mbit/s)
  0Ah  PW   Page Write (Write 3-Byte-Address, write 1..256 data bytes)
             (changing bits to 0 or 1) (reads unchanged data, erases the page,
             then writes new & unchanged data) (11ms typ, 25ms max)
  02h  PP   Page Program (Write 3-Byte-Address, write 1..256 data bytes)
             (changing bits from 1 to 0) (1.2ms typ, 5ms max)
  DBh  PE   Page Erase 100h bytes (Write 3-Byte-Address) (10ms typ, 20ms max)
  D8h  SE   Sector Erase 10000h bytes (Write 3-Byte-Address) (1s typ, 5s max)
  B9h  DP   Deep Power-down (No Parameters) (consumption 1uA typ, 10uA max)
             (3us) (ignores all further instructions, except RDP)
  ABh  RDP  Release from Deep Power-down (No Parameters) (30us)

Write/Program may not cross page-boundaries. Write/Program/Erase are rejected during first 1..10ms after power up. The WEL bit is automatically cleared on Power-Up, on /Reset, and on completion of WRDI/PW/PP/PE/SE instructions. WEL is set by WREN instruction (which must be issued before any write/program/erase instructions). Don't know how RDSR behaves when trying to write to the write-protected region?

Communication Protocol

  Set Chip Select LOW to invoke the command
  Transmit the instruction byte
  Transmit any parameter bytes
  Transmit/receive any data bytes
  Set Chip Select HIGH to finish the command

All bytes (and 3-byte addresses) transferred most significant bit/byte first.

DSi "5A32" chip (32Kit aka 4Kbyte)
Early DSi DWM-W015 boards did have 128Kbyte FLASH, but later boards have custum 4Kbyte FLASH chips (these 4K chips are found on later DSi DWM-W015 boards, and DSi DWM-W024 boards, and 3DS DWM-W028 boards). The chips are having a 24bit address bus (needed for NDS compatibility), and, a weird non-writeable gap within a 128Kbyte memory are:

  000000h..0002FFh  Writeable only if /WP=HIGH (otherwise writes are ignored)
  000300h..01F2FFh  Not writeable (FFh-filled, writes are ignored)
  01F300h..01FFFFh  Writeable
  020000h and up    Mirrors of 0..01FFFFh (same read/write-ability as above)

There are several part numbers: "5A32" (DSi), "5K32" (3DS), "32B, 3XH" (3DS), and "26FV032T" (DSi), that chips are probably all the same size & functionally same; most of those 4Kbyte chips have tiny packages (except "26FV032T" which comes in classic large package).

Pin-Outs (Large Package, in NDS, and early DSi boards)

  1   D    Serial Data In (latched at rising clock edge)          _________
  2   C    Serial Clock (max 25MHz)                             /|o        |
  3   /RES Reset                                            1 -| |         |- 8
  4   /S   Chip Select (instructions start at falling edge) 2 -| |         |- 7
  5   /W   Write Protect (makes first 256 pages read-only)  3 -| |_________|- 6
  6   VCC  Supply (2.7V..3.6V typ) (4V max) (DS:VDD3.3)     4 -|/          |- 5
  7   VSS  Ground                                              |___________|
  8   Q    Serial Data Out (changes at falling clock edge)

Pin-Outs (Tiny Package, in newer DSi boards, and 3DS)

  1   /S   Chip Select (instructions start at falling edge)     ___________
  2   Q    Serial Data Out (changes at falling clock edge)  1 -| o         |- 8
  3   /W   Write Protect (makes first pages read-only)      2 -|           |- 7
  4   VSS  Ground                                           3 -|           |- 6
  5   D    Serial Data In (latched at rising clock edge)    4 -|___________|- 5
  6   C    Serial Clock
  7   /RES Reset
  8   VCC  Supply (2.7V..3.6V typ) (DSi: VDD33)

DS Firmware Header

Firmware Memory Map

  00000h-00029h  Firmware Header
  0002Ah-001FFh  Wifi Settings
  00200h-3F9FFh  Firmware Code/Data                       ;-NDS only (not DSi)
  00200h-002FEh  00h-filled                               ;\
  002FFh         80h                                      ;
  00300h-1F2FFh  FFh-filled (not write-able on 4K chips)  ; DSi only (not NDS)
  1F300h-1F3FEh  FFh-filled (write-able)                  ;
  1F3FFh         Whatever Debug/Bootflags                 ;
  1F400h-1F5FFh  Wifi Access Point 4  ;\with WPA/WPA2     ;
  1F600h-1F7FFh  Wifi Access Point 5  ; support           ;
  1F800h-1F9FFh  Wifi Access Point 6  ;/                  ;/
  3FA00h-3FAFFh  Wifi Access Point 1  ;\
  3FB00h-3FBFFh  Wifi Access Point 2  ; Open/WEP only
  3FC00h-3FCFFh  Wifi Access Point 3  ;/
  3FD00h-3FDFFh  Not used
  3FE00h-3FEFFh  User Settings Area 1
  3FF00h-3FFFFh  User Settings Area 2

On iQue DS (with 512K flash memory), user settings are moved to 7FE00h and up, and, there seems to be some unknown stuff at 200h..27Fh.

Firmware Header (00000h-001FFh)

  Addr Size Expl.
  000h 2    part3 romaddr/8 (arm9 gui code) (LZ/huffman compression)
  002h 2    part4 romaddr/8 (arm7 wifi code) (LZ/huffman compression)
  004h 2    part3/4 CRC16 arm9/7 gui/wifi code
  006h 2    part1/2 CRC16 arm9/7 boot code
  008h 4    firmware identifier (usually nintendo "MAC",nn) (or nocash "XBOO")
            the 4th byte (nn) occassionally changes in different versions
  00Ch 2    part1 arm9 boot code romaddr/2^(2+shift1) (LZSS compressed)
  00Eh 2    part1 arm9 boot code 2800000h-ramaddr/2^(2+shift2)
  010h 2    part2 arm7 boot code romaddr/2^(2+shift3) (LZSS compressed)
  012h 2    part2 arm7 boot code 3810000h-ramaddr/2^(2+shift4)
  014h 2    shift amounts, bit0-2=shift1, bit3-5=shift2, bit6-8=shift3,
            bit9-11=shift4, bit12-15=firmware_chipsize/128K
  016h 2    part5 data/gfx romaddr/8 (LZ/huffman compression)
  018h 8    Optional KEY1-encrypted "enPngOFF"=Cartridge KEY2 Disable
            (feature isn't used in any consoles, instead contains timestamp)
  018h 5    Firmware version built timestamp (BCD minute,hour,day,month,year)
  01Dh 1    Console type
              FFh=Nintendo DS
              20h=Nintendo DS-lite
              57h=Nintendo DSi (also iQueDSi)
              43h=iQueDS
              63h=iQueDS-lite
            The entry was unused (FFh) in older NDS, ie. replace FFh by 00h)
              Bit0   seems to be DSi/iQue related
              Bit1   seems to be DSi/iQue related
              Bit2   seems to be DSi related
              Bit3   zero
              Bit4   seems to be DSi related
              Bit5   seems to be DS-Lite related
              Bit6   indicates presence of "extended" user settings (DSi/iQue)
              Bit7   zero
  01Eh 2    Unused (FFh-filled)
  020h 2    User Settings Offset (div8) (usually last 200h flash bytes)
  022h 2    Unknown (7EC0h or 0B51h)
  024h 2    Unknown (7E40h or 0DB3h)
  026h 2    part5 CRC16 data/gfx
  028h 2    unused (FFh-filled)
  02Ah-1FFh Wifi Calibration Data (see next chapter)

DSi

  000h 1Dh   Zerofilled (bootcode is in new eMMC chip, not on old FLASH chip)
  01Dh 6     Same as on DS (header: Console Type and User Settings Offset)
  022h 6     Zerofilled (bootcode is in new eMMC chip, not on old FLASH chip)
  028h..1FCh Same as on DS (wifi calibration)
  1FDh 1     Wifi Board (01h=DWM-W015, 02h=W024, 03h=W028)   ;\this was
  1FEh 1     Wifi Flash (20h=With access point 4/5/6)        ; FFh-filled
  1FFh 1     Same as on DS (FFh)                             ;/on DS
  200h FFh   Zerofilled                                      ;\
  2FFh 1     Unknown (80h)                                   ; this was
  300h   1F000h FFh's (not write-able on 4K chips)           ; bootcode
  1F300h FFh    FFh's (write-able) ;twl-debugger: 00h's      ; on DS
  1F3FFh 1      FFh                ;twl-debugger: 40h        ;/

The bytes [000h..027h] cannot be changed on DSi because they are part of the RSA signature in DSi's Boot Info Block (at eMMC offset 200h..3FFh).

DS Firmware Wifi Calibration Data

Wifi Calibration/Settings (located directly after Firmware Header)

  Addr Size Expl.
  000h-029h Firmware Header (see previous chapter)
  02Ah 2    CRC16 (with initial value 0) of [2Ch..2Ch+config_length-1]
  02Ch 2    config_length (usually 0138h, ie. entries 2Ch..163h)
  02Eh 1    Unused        (00h)
  02Fh 1    Version (0=v1..v4, 3=v5, 5=v6..v7,6=W006,15=W015,24=W024,34=N3DS)
  030h 6    Unused        (00h-filled) (DS-Lite and DSi: FF,FF,FF,FF,FF,00)
  036h 6    48bit MAC address (v1-v5: 0009BFxxxxxx, v6-v7: 001656xxxxxx)
  03Ch 2    list of enabled channels ANDed with 7FFE (Bit1..14 = Channel 1..14)
            (usually 3FFEh, ie. only channel 1..13 enabled)
  03Eh 2    Whatever Flags (usually FFFFh)
  040h 1    RF Chip Type (NDS: usually 02h) (DS-Lite and DSi/3DS: usually 03h)
  041h 1    RF Bits per entry at 0CEh (usually 18h=24bit=3byte) (Bit7=?)
  042h 1    RF Number of entries at 0CEh (usually 0Ch)
  043h 1    Unknown (usually 01h)
  044h 2    Initial Value for [4808146h]  ;W_CONFIG_146h
  046h 2    Initial Value for [4808148h]  ;W_CONFIG_148h
  048h 2    Initial Value for [480814Ah]  ;W_CONFIG_14Ah
  04Ah 2    Initial Value for [480814Ch]  ;W_CONFIG_14Ch
  04Ch 2    Initial Value for [4808120h]  ;W_CONFIG_120h
  04Eh 2    Initial Value for [4808122h]  ;W_CONFIG_122h
  050h 2    Initial Value for [4808154h]  ;W_CONFIG_154h
  052h 2    Initial Value for [4808144h]  ;W_CONFIG_144h
  054h 2    Initial Value for [4808130h]  ;W_CONFIG_130h
  056h 2    Initial Value for [4808132h]  ;W_CONFIG_132h
  058h 2    Initial Value for [4808140h]  ;W_CONFIG_140h ;maybe ACK timeout?
  05Ah 2    Initial Value for [4808142h]  ;W_CONFIG_142h
  05Ch 2    Initial Value for [4808038h]  ;W_POWER_TX
  05Eh 2    Initial Value for [4808124h]  ;W_CONFIG_124h
  060h 2    Initial Value for [4808128h]  ;W_CONFIG_128h
  062h 2    Initial Value for [4808150h]  ;W_CONFIG_150h
  064h 69h  Initial 8bit values for BB[0..68h]
  0CDh 1    Unused (00h)

Below for Type2 (ie. when [040h]=2) (Mitsumi MM3155 and RF9008):

  0CEh 24h  Initial 24bit values for RF[0,4,5,6,7,8,9,0Ah,0Bh,1,2,3]
  0F2h 54h  Channel 1..14 2x24bit values for RF[5,6]
  146h 0Eh  Channel 1..14 8bit values for BB[1Eh] (usually somewhat B1h..B7h)
  154h 0Eh  Channel 1..14 8bit values for RF[9].Bit10..14 (usually 10h-filled)

Below for Type3 (ie. when [040h]=3) (Mitsumi MM3218) (and AR6013G):

  --- Type3 values are originated at 0CEh, following addresses depend on:  ---
  1) number of initial values, found at [042h]        ;usually 29h
  2) number of BB indices,     found at [0CEh+[042h]] ;usually 02h
  3) number of RF indices,     found at [043h]        ;usually 02h
  --- Below example addresses assume above values to be set to 29h,02h,02h ---
  0CEh 29h  Initial 8bit values for RF[0..28h]
  0F7h 1    Number of BB indices per channel
  0F8h 1    1st BB index
  0F9h 14   1st BB data for channel 1..14
  107h 1    2nd BB index
  108h 14   2nd BB data for channel 1..14
  116h 1    1st RF index
  117h 14   1st RF data for channel 1..14
  125h 1    2nd RF index
  126h 14   2nd RF data for channel 1..14
  134h 46   Unused (FFh-filled)

Below for both Type2 and Type3:

  162h 1    Unknown (usually 19h..1Ch)
  163h 1    Unused (FFh) (Inside CRC16 region, with config_length=138h)
  164h 99h  Unused (FFh-filled) (Outside CRC16 region, with config_length=138h)
  1FDh 1    DSi/3DS Wifi Board (01h=W015, 02h=W024, 03h=W028);\this was
  1FEh 1    DSi/3DS Wifi Flash (20h=With access point 4/5/6) ; FFh-filled on DS
  1FFh 1    DSi/3DS Same as on DS (FFh)                      ;/

Most of the Wifi settings seem to be always the same values on all currently existing consoles. Except for:
Values that are (obviously) different are the CRC16, and 4th-6th bytes of the MAC address. Also, initial values for BB[01h] and BB[1Eh], and channel 1..14 values for BB[1Eh], and unknown entry [162h] contain different calibration settings on all consoles.
Firmware v5 is having a new wifi ID [2Fh]=03h, and different RF[9] setting.
Firmware v6 (dslite) has wifi ID [2Fh]=05h, and same RF[9] setting as v5, additionally, v6 and up have different 2nd-3rd bytes of the MAC address.

Moreover, a LOT of values are different with Type3 chips (ie. when [040h]=3).

Note
Unlike for Firmware User Settings, the Firmware Header (and Wifi Settings) aren't stored in RAM upon boot. So the data must be retrieved via SPI bus by software.

DS Firmware Wifi Internet Access Points

Connection data 1 at WifiFlash[00020h]*8-400h (eg. 01FA00h/03FA00h/07FA00h)
Connection data 2 at WifiFlash[00020h]*8-300h (eg. 01FB00h/03FB00h/07FB00h)
Connection data 3 at WifiFlash[00020h]*8-200h (eg. 01FC00h/03FC00h/07FC00h)
These three 100h byte regions are used to memorize known internet access points. The NDS firmware doesn't use these regions, but games that support internet are allowed to read (and configure/write) them. The DSi firmware also supports configuring these entries.

  Addr Siz Expl.
  000h 64  Unknown (usually 00h-filled) (no Proxy supported on NDS)
  040h 32  SSID (ASCII name of the access point) (padded with 00h's)
  060h 32  SSID for WEP64 on AOSS router (each security level has its own SSID)
  080h 16  WEP Key 1 (for type/size, see entry E6h)
  090h 16  WEP Key 2  ;\
  0A0h 16  WEP Key 3  ; (usually 00h-filled)
  0B0h 16  WEP Key 4  ;/
  0C0h 4   IP Address           (0=Auto/DHCP)
  0C4h 4   Gateway              (0=Auto/DHCP)
  0C8h 4   Primary DNS Server   (0=Auto/DHCP)
  0CCh 4   Secondary DNS Server (0=Auto/DHCP)
  0D0h 1   Subnet Mask (0=Auto/DHCP, 1..1Ch=Leading Ones) (eg. 6 = FC.00.00.00)
  0D1h ..  Unknown (usually 00h-filled)
  0E6h 1   WEP Mode (0=None, 1/2/3=5/13/16 byte hex, 5/6/7=5/13/16 byte ascii)
  0E7h 1   Status (00h=Normal, 01h=AOSS, FFh=connection not configured/deleted)
  0E8h 1   Zero (not SSID Length, ie. unlike as entry 4,5,6 on DSi)
  0E9h 1   Unknown (usually 00h)
  0EAh 2   DSi only: MTU (Max transmission unit) (576..1500, usually 1400)
  0ECh 3   Unknown (usually 00h-filled)
  0EFh 1   bit0/1/2 - connection 1/2/3 (1=Configured, 0=Not configured)
  0F0h 6   Nintendo Wifi Connection (WFC) 43bit User ID
           (ID=([F0h] AND 07FFFFFFFFFFh)*1000, shown as decimal string
           NNNN-NNNN-NNNN-N000) (the upper 5bit of the last byte are
           containing additional/unknown nonzero data)
  0F6h 8   Unknown (nonzero stuff !?!)
  0FEh 2   CRC16 for Entries 000h..0FDh (with initial value 0000h)

For connection 3: entries [0EFh..0FDh] - always zero-filled?
The location of the first data block is at the User Settings address minus 400h, ie. Firmware Header [00020h]*8-400h.

Connection data 4 at WifiFlash[00020h]*8-A00h (eg. 01F400h) (DSi only)
Connection data 5 at WifiFlash[00020h]*8-800h (eg. 01F600h) (DSi only)
Connection data 6 at WifiFlash[00020h]*8-600h (eg. 01F800h) (DSi only)
The DSi has three extra 200h-byte regions (for use DSi games, with the new WPA/WPA2 encryption support, and with additional proxy support), these extra regions are found under "Advanced Setup" in the DSi firmware's "Internet" configuration menu.

  Addr Siz Expl.
  000h 32  Proxy Authentication Username (ASCII string, padded with 00's)
  000h 32  Proxy Authentication Password (ASCII string, padded with 00's)
  040h 32  SSID (ASCII string, padded with 00's) (see [0E8h] for length)
  060h ..  Maybe same as NDS
  080h 16  WEP Key (zerofilled for WPA)
  0xxh ..  Maybe same as NDS
  0C0h 4   IP Address           (0=Auto/DHCP)
  0C4h 4   Gateway              (0=Auto/DHCP)
  0C8h 4   Primary DNS Server   (0=Auto/DHCP)
  0CCh 4   Secondary DNS Server (0=Auto/DHCP)
  0D0h 1   Subnet Mask (0=Auto/DHCP, 1..1Ch=Leading Ones) (eg. 6 = FC.00.00.00)
  0D1h ..  Unknown (zerofilled)
  0E6h 1   WEP (00h=None/WPA/WPA2, 01h/02h/03h/05h/06h/07h=WEP, same as NDS)
  0E7h 1   WPA (00h=Normal, 10h=WPA/WPA2, 13h=WPS+WPA/WPA2, FFh=unused/deleted)
  0E8h 1   SSID Length in characters (01h..20h, or 00h=unused)
  0E9h 1   Unknown (usually 00h)
  0EAh 2   MTU Value (Max transmission unit) (576..1500, usually 1400)
  0ECh 3   Unknown (usually 00h-filled)
  0EFh 1   bit0/1/2 - connection 4/5/6 (1=Configured, 0=Not configured)
  0F0h 14  Zerofilled (or maybe ID as on NDS, if any such ID exists for DSi?)
  0FEh 2   CRC16 for Entries 000h..0FDh (with initial value 0000h)
  100h 32  Precomputed PSK (based on WPA/WPA2 password and SSID) ;\all zero
  120h 64  WPA/WPA2 password (ASCII string, padded with 00's)    ;/for WEP
  160h 33  Zerofilled
  181h 1   WPA (0=None/WEP, 4=WPA-TKIP, 5=WPA2-TKIP, 6=WPA-AES, 7=WPA2-AES)
  182h 1   Proxy Enable         (00h=None, 01h=Yes)
  183h 1   Proxy Authentication (00h=None, 01h=Yes)
  184h 48  Proxy Name (ASCII string, max 47 chars, padded with 00's)
  1B4h 52  Zerofilled
  1E8h 2   Proxy Port (16bit)
  1EAh 20  Zerofilled
  1FEh 2   CRC16 for Entries 100h..1FDh (with initial value 0000h) (0=deleted)

The location of the first data block (aka settings number 4) is at the User Settings address minus A00h, ie. Firmware Header [00020h]*8-A00h.
Observe that NDS consoles do have NDS Firmware bootcode/data in that area, so those new regions can exist on DSi only (or on homebrew NDS firmwares). Presence of the new regions is indicated in Firmware Header [001FEh], that byte is usually FFh=NDS or 20h=DSi, the DSi browser does internally replace FFh by 10h, and does then check if byte>=20h (ie. the new areas exist if the byte is 20h..FEh).
Note that the Proxy feature can be used to redirect internet access (when using a custom proxy server, one could redirect commercial games to homebrew servers; as done by the http://pbsds.net/ project) (actually the same should be possible with the DNS server entry, possibly with less traffic).

Note
The location of the user settings & connection data varies (eg. 01Fxxxh=DSi, 03Fxxxh=NDS, 07Fxxxh=iQueDS).

Nintendo Zone Beacons
DSi games and DSi browser can reportedly also connect to Nintendo's public access points (those that are announced via Nintendo Zone Beacons).

Hidden Connection data at WifiFlash[00020h]*8-100h (eg. xxFD00h)
This seems to hold an extra 100h-byte region (same as Connection data 1-3), some (or all) NDS/DSi games seem to be capable of using it (although it isn't shown in config menues). One theory is that it might be used for the Nintendo Wi-Fi USB Connector?

DS Firmware User Settings

Current Settings (RAM 27FFC80h-27FFCEFh, DSi: 2FFFC80h-2FFFCEFh)
User Settings 0 (Firmware 3FE00h-3FEFFh) ;(DSi & iQue use different address,
User Settings 1 (Firmware 3FF00h-3FFFFh) ;see Firmware Header [020h])

  Addr Size Expl.
  000h  2   Version (5) (Always 5, for all NDS/DSi Firmware versions)
  002h  1   Favorite color (0..15) (0=Gray, 1=Brown, etc.)
  003h  1   Birthday month (1..12) (Binary, non-BCD)
  004h  1   Birthday day   (1..31) (Binary, non-BCD)
  005h  1   Not used (zero)
  006h  20  Nickname string in UTF-16 format
  01Ah  2   Nickname length in characters    (0..10)
  01Ch  52  Message string in UTF-16 format
  050h  2   Message length in characters     (0..26)
  052h  1   Alarm hour     (0..23) (Binary, non-BCD)
  053h  1   Alarm minute   (0..59) (Binary, non-BCD)
  054h  2
  056h  1   80h=enable alarm (huh?), bit 0..6=enable?
  057h  1   Zero (1 byte)
  058h  2x2 Touch-screen calibration point (adc.x1,y1) 12bit ADC-position
  05Ch  2x1 Touch-screen calibration point (scr.x1,y1) 8bit pixel-position
  05Eh  2x2 Touch-screen calibration point (adc.x2,y2) 12bit ADC-position
  062h  2x1 Touch-screen calibration point (scr.x2,y2) 8bit pixel-position
  064h  2   Language and Flags (see below)
  066h  1   Year (2000..2255) (when having entered date in the boot menu)
  067h  1   Unknown (usually 00h...08h or 78h..7Fh or so)
  068h  4   RTC Offset (difference in seconds when RTC time/date was changed)
  06Ch  4   Not used (FFh-filled, sometimes 00h-filled) (=MSBs of above?)

Below not stored in RAM (found only in FLASH memory)...

  070h  2   Update counter (used to check latest) (must be 0000h..007Fh)
  072h  2   CRC16 of entries 00h..6Fh (70h bytes)
  074h  8Ch Not used (FFh-filled) (or extended data, see below)

Below extended data was invented for iQue DS (for adding the chinese language setting), and is also included in Nintendo DSi models. Presence of extended data is indicated in Firmware Header entry [1Dh].Bit6.

  074h  1   Unknown (01h) (maybe version?)
  075h  1   Extended Language (0..5=Same as Entry 064h, plus 6=Chinese)
            (for language 6, entry 064h defaults to english; for compatibility)
            (for language 0..5, both entries 064h and 075h have same value)
  076h  2   Bitmask for Supported Languages (Bit0..6)
            (007Eh for iQue DS, ie. with chinese, but without japanese)
            (0042h for iQue DSi, chinese (and english, but only for NDS mode))
            (003Eh for DSi/EUR, ie. without chinese, and without japanese)
  078h  86h Not used (FFh-filled on iQue DS, 00h-filled on DSi)
  0FEh  2   CRC16 of entries 74h..FDh (8Ah bytes)

Note: The DSi does store the user settings in eMMC files (TWLCFGn.dat), that files include NDS-style settings (username etc), plus additional DSi-specific settings (country, parental controls, etc). For backwards compatibilty, the DSi does also store a copy of those settings in NDS-style format in Wifi FLASH and Main RAM at 2FFFC80h.
DSi SD/MMC Firmware System Settings Data Files
DSi Backlight level and DSi sound volume seem to be stored in the BPTWL chip (or possibly in its attached I2C potentiometer).

Language and Flags (Entry 064h)

  0..2 Language (0=Japanese, 1=English, 2=French, 3=German,
       4=Italian, 5=Spanish, 6..7=Reserved) (for Chinese see Entry 075h)
       (the language setting also implies time/data format)
  3    GBA mode screen selection (0=Upper, 1=Lower)
  4-5  Backlight Level    (0..3=Low,Med,High,Max) (DS-Lite only)
  6    Bootmenu Disable   (0=Manual/bootmenu, 1=Autostart Cartridge)
  7-8  ?
  9    Settings Lost (1=Prompt for User Info, and Language, and Calibration)
  10   Settings Okay (0=Prompt for User Info)
  11   Settings Okay (0=Prompt for User Info) (Same as Bit10)
  12   No function
  13   Settings Okay (0=Prompt for User Info, and Language)
  14   Settings Okay (0=Prompt for User Info) (Same as Bit10)
  15   Settings Okay (0=Prompt for User Info) (Same as Bit10)

The Health and Safety message is skipped if Bit9=1, or if one or more of the following bits is zero: Bits 10,11,13,14,15. However, as soon as entering the bootmenu, the Penalty-Prompt occurs.

Note: There are two User Settings areas in the firmware chip, at offset 3FE00h and 3FF00h, if both areas have valid CRCs, then the current/newest area is that whose Update Counter is one bigger than in the other/older area.

  IF count1=((count0+1) AND 7Fh) THEN area1=newer ELSE area0=newer

When changing settings, the older area is overwritten with new data (and incremented Update Counter). The two areas allow to recover previous settings in case of a write-error (eg. on a battery failure during write).

Battery Removal
Even though the battery is required only for the RTC (not for the firmware flash memory), most of the firmware user settings are reset when removing the battery. This appears to be a strange bug-or-feature of the DS bios, at least, fortunately, it still keeps the rest of the firmware intact.

DS Firmware Extended Settings

Extended Settings contain some additional information which is not supported by the original firmware (current century, date/time formats, temperature calibration, etc.), the settings are supported by Nocash Firmware, by the no$gba emulator, and may be eventually also supported by other emulators. If present, the values can be used by games, otherwise games should use either whatever default settings, or contain their own configuration menu.

Extended Settings - loaded to 23FEE00h (aka fragments of NDS9 boot code)

  Addr Siz Expl.
  00h  8  ID "XbooInfo"
  08h  2  CRC16 Value [0Ch..0Ch+Length-1]
  0Ah  2  CRC16 Length (from 0Ch and up)
  0Ch  1  Version (currently 01h)
  0Dh  1  Update Count (newer = (older+1) AND FFh)
  0Eh  1  Bootmenu Flags
            Bit6   Important Info  (0=Disable, 1=Enable)
            Bit7   Bootmenu Screen (0=Upper, 1=Lower)
  0Fh  1  GBA Border (0=Black, 1=Gray Line)
  10h  2  Temperature Calibration TP0 ADC value  (x16) (sum of 16 ADC values)
  12h  2  Temperature Calibration TP1 ADC value  (x16) (sum of 16 ADC values)
  14h  2  Temperature Calibration Degrees Kelvin (x100) (0=none)
  16h  1  Temperature Flags
            Bit0-1 Format (0=Celsius, 1=Fahrenheit, 2=Reaumur, 3=Kelvin)
  17h  1  Backlight Intensity (0=0ff .. FFh=Full)
  18h  4  Date Century Offset       (currently 20, for years 2000..2099)
  1Ch  1  Date Month Recovery Value (1..12)
  1Dh  1  Date Day Recovery Value   (1..31)
  1Eh  1  Date Year Recovery Value  (0..99)
  1Fh  1  Date/Time Flags
            Bit0-1 Date Format   (0=YYYY-MM-DD, 1=MM-DD-YYYY, 2=DD-MM-YYYY)
            Bit2   Friendly Date (0=Raw Numeric, 1=With Day/Month Names)
            Bit5   Time DST      (0=Hide DST, 1=Show DST=On/Off)
            Bit6   Time Seconds  (0=Hide Seconds, 1=Show Seconds)
            Bit7   Time Format   (0=24 hour, 1=12 hour)
  20h  1  Date Separator      (Ascii, usually Slash, or Dot)
  21h  1  Time Separator      (Ascii, usually Colon, or Dot)
  22h  1  Decimal Separator   (Ascii, usually Comma, or Dot)
  23h  1  Thousands Separator (Ascii, usually Comma, or Dot)
  24h  1  Daylight Saving Time (Nth)
             Bit 0-3 Activate on (0..4 = Last,1st,2nd,3rd,4th)
             Bit 4-7 Deactivate on (0..4 = Last,1st,2nd,3rd,4th)
  25h  1  Daylight Saving Time (Day)
             Bit 0-3 Activate on (0..7 = Mon,Tue,Wed,Thu,Fri,Sat,Sun,AnyDay)
             Bit 4-7 Deactivate on (0..7 = Mon,Tue,Wed,Thu,Fri,Sat,Sun,AnyDay)
  26h  1  Daylight Saving Time (of Month)
             Bit 0-3 Activate DST in Month   (1..12)
             Bit 4-7 Deactivate DST in Month (1..12)
  27h  1  Daylight Saving Time (Flags)
             Bit 0   Current DST State (0=Off, 1=On)
             Bit 1   Adjust DST Enable (0=Disable, 1=Enable)

Note: With the original firmware, the memory region at 23FEE00h and up contains un-initialized, non-zero-filled data (fragments of boot code).

DS File Formats

DS Files - 2D Video
DS Files - 3D Video (mostly unknown)
DS Files - Sound (SDAT etc.)
DS Files - Text Messages (MESG)
DS Files - Text Manuals
DS Cartridge Nitro Font Resource Format
DS Cartridge NitroROM and NitroARC File Systems
LZ Decompression Functions
LZ Decompression Functions ASH0
ZIP Decompression
DS Encrypted Arika Archives with ALZ1 compression

DS Files - Text Messages (MESG)

MESG files contain localized strings (like "Okay" and "Cancel" and longer text messages). Different languages are stored separate message files, usually with .bmg extension. MESG files are used in NDS/DSi titles (in little endian), and also on Wii (in big endian).
DSi titles are: DSi Shop, Launcher, System Settings, and DSi Camera (however, the DSi-to-3DS Transfer Tool uses the newer MsgPrjBn/MsgStdBn format; despite of being a DSi program).

MESG Header

  000h 8     ID "MESGbmg1"         ;or "GSEM1gmb" in Super Mario 64 DS
  008h 4     Total Filesize        ;or Filesize+1 in Super Mario 64 DS
  00Ch 4     Number of Chunks (2=INF1+DAT1, 3=INF1+DAT1+MID1)
  010h 1     Encoding (1=CP1252, 2=UTF-16, 3=Shift-JIS, 4=UTF-8)
  011h 15    Padding (0)

Encoding UTF-16 appears to be most common (Super Mario 64 DS uses Shift-JIS).

INF1 Chunk (Message Info)
It comes just after the BMG header. It contains information (like pointers) about the messages.

  000h 4     Chunk ID "INF1"       ;or "1FNI" in Super Mario 64 DS
  004h 4     Chunk Size
  008h 2     Number of messages (N)
  00Ah 2     Size of each INF data in bytes  ;or in BITs in Super Mario 64 DS
  00Ch 4     "BMG file ID = ID for this BMG file (usually 0)"
  010h N*siz Message Info (32bit offset from DAT1+8, and optional attributes)

For each message, there is a INF data about it. At Wii Fit, there is just the message offset. At Wii Fit Plus, there is two another fields.

  000h 4     Offset to the message (after DAT1+8 section header)
  004h siz-4 Attributes/flags (if entrysize is bigger than 4 bytes)

The optional Attribute bits could be used to select different font types or window styles.

DAT1 Chunk (Message Strings)

  000h 4     Chunk ID "DAT1"       ;or "1TAD" in Super Mario 64 DS
  004h 4     Chunk Size            ;or Size+1 in Super Mario 64 DS
  008h ..    Message strings (usually UTF-16, depending on Encoding in header)

The DSi Camera uses both char 001Ah and 0025h for escape codes.
UTF-16 string characters:

  0000                           End of String (except inside Escape sequences)
  000A                           Linebreak
  001A,nn,command,parameters     Escape Sequences (nn=length in bytes)
  001A,08,00,0000,00xx           Set font size (64h=100%=Normal Size)
  001A,08,00,0001,00xx           Set text color to xx
  001A,08,01,0000,24xx           Draw Unicode char U+2460..246E ;"(1)"..("15)"
  001A,08,01,0000,xxxx           Draw Unicode char U+E068..F12B ;custom?
  001A,06,02,0000                Draw Name of current player
  001A,0A,02,0010,000x,000w      Draw Integer from index x with w digits
  001A,08,02,0011,00xx           Unknown (with xx=0..8)
  001A,08,02,0012,0000           Draw Name of a player
  001A,08,02,0013,0000           Unknown
  001A,08,02,0014,0000           Unknown
  001A,08,02,0015,0000           Unknown
  001A,0A,02,0016,0000,0000      Unknown
  001A,08,02,0017,0000           Unknown
  001A,08,02,0020,0000           Draw Name of a Wii friend
  001A,08,03,0010,0000           Unknown
  001A,0C,04,0000,000x,yyyy,zzzz Unknown (x=0..1, y=0524..14A4, and z=y+1)
  0025,00xx,00yy,00zz            Escape codes in form of "%xyz" (or similar)
  00xx                           ASCII Characters 20h..7Eh
  E0xx                           Custom button symbols (eg. in DSi Launcher)

Shift-JIS (or whatever) strings in Super Mario 64 DS:

  0D       Linebreak?
  10..1F   Escape codes?
  xx,xx    Unknown (doesn't really look like english Shift-JIS characters)
  FF       End of String

MID1 Chunk (Message IDs) (if any)

  000h 4     Chunk ID "MID1"
  004h 4     Chunk Size
  008h 2     Number of messages (same as in INF1 block)
  00Ah 2     Unknown (usually 1000h)
  00Ch 4     Padding (0)
  010h 4*N   Message IDs

Messages can be repeated in the different files (such like menu and ingame), if so, all messages with the same message ID of the same language are always having the same text, no differences between the text files can be found.
"Elements with the same table index are attributes (not IDs?) for the same string."

FLW1 and FLT1 Chunks (if any)
Some games are reportedly having additional "FLW1" and "FLT1" chunks. Unknown which games, and unknown what for, and unknown if such chunks exist in any DS games (or only in Wii games or whatever).

Thanks to http://wiki.tockdom.com/wiki/BMG

DS Files - Text Manuals

Manual files are common in DSiware download titles, unknown if DS/DSi cartridges did also use that files (or if they did stick with print manuals).
Used in DSi Flipnote, Sudoku, Paper Plane, Deep Psyche, Dr. Mario, DSi-to-3DS Transfer Tool.
Usually found in "rom:\..\manpages_narc.blz" (with two folders in the .blz file, "arc" for the manual, and "gpArc" with a quick reference on how to read the manual; the latter containing only two "pages", without "content" file).

NTLI File (manual language info) (in "ntli" folder)

  000h 4     ID "NTLI"
  004h 2     Byte Order  (FEFFh)
  006h 2     Version     (can be 0200h)
  008h 4     Total Filesize
  00Ch 2     Header Size (10h)
  00Eh 4     Number of Chunks (usually 1 = mtl1)

mtl1 Chunk

  000h 4     Chunk ID "mtl1"
  004h 4     Chunk Size
  008h 4     Number of supported languages
  00Ch 2*N   Language IDs (two-letter ASCII spelled backwards)

Known IDs are ne=English, rf=French, se=Spanish, ti=Italian, ed=German.
EUR games do usually support 5 languages (Deep Psyche supports less).
IDs for Japanese/Chinese/Korean are unknown (also unknown if US has different IDs for english/spanish etc, and if portugese or so do exist on DSi).
Note: The two-letters do directly translate to the "ntmc\xx\" and "ntpg\xx\" folder names.

NTMC File (manual contents) (in "ntmc\xx\" folders)
The content contains "hyperlinks" to the actual chapters. A chapter can be divided into subsections (with the subsections being shown at the right of the main chapter name).

  000h 4     ID "NTMC"
  004h 2     Byte Order  (FEFFh)
  006h 2     Version     (can be 0200h)
  008h 4     Total Filesize
  00Ch 2     Header Size (10h)
  00Eh 4     Number of Chunks (usually 3 = nap1+txp1+mtc1)

nap1 Chunk

  000h 4     Chunk ID "nap1"
  004h 4     Chunk Size
  008h 4     Number of chapters (aka pages) minus 1?  (eg. 18h=19h)
  00Ch 4*N   Offsets to filenames (from nap+0Ch)
  ...  ..    Filenames (ASCII, terminated by 00h)

txp1 Chunk

  000h 4     Chunk ID "txp1"
  004h 4     Chunk Size
  008h 4     Number of something?? minus 1            (eg. 25h=26h)
  00Ch 4*N   Offsets to something?? (from txp1+0Ch)
  ...  ..    Somewhat corrupt UTF-16 strings (many aborted with char 20xxh)

mtc1 Chunk

  000h 4     Chunk ID "mtc1"
  004h 4     Chunk Size
  008h 4     Number of dunno what                  (eg. 0Dh=Much more?)
  00Ch 2*?   16bit Indices in txp1? (eg. 0000h..0025h)

NTPG File (manual pages) (or rather chapters?) (in "ntpg\xx\" folders)
The pages can contain text, tables, and symbols/images/screenshots (there appears to be no support for hyperlinks inside of the pages; probably because they default to be displayed on upper screen, without touchscreen support).

  000h 4     ID "NTPC"
  004h 2     Byte Order  (FEFFh)
  006h 2     Version     (can be 0200h)
  008h 4     Total Filesize
  00Ch 2     Header Size (10h)
  00Eh 4     Number of Chunks (usually 7 = nap1+txp1+pag1+pan1+pas1+txt1+pae1)

nap1 Chunk

  000h 4     Chunk ID "nap1"
  004h 4     Chunk Size (10h)
  008h 4     Zero (unlike as in NTMC file)
  00Ch 4     Unknown (4)

txp1 Chunk

  000h 4     Chunk ID "txp1"
  004h 4     Chunk Size
  008h 4     Number of something?? minus 1            (eg. 02h=03h)
  00Ch 4*N   Offsets to something?? (from txp1+0Ch)
  ...  ..    UTF-16 strings (Headline, Body, Footer?)

pag1 Chunk

  000h 4     Chunk ID "pag1"
  004h 4     Chunk Size (10h)
  008h 2?    0000h text color black?
  00Ah 2?    0160h link color or so?
  00Ch 2?    7FFFh bg color white?
  00Eh 2?    0000h

pan1 Chunk

  000h 4     Chunk ID "pan1"
  004h 4     Chunk Size (10h)
  008h 2?    0000h
  00Ah 2?    0000h
  00Ch 2?    0100h
  00Eh 2?    0160h link color or so?

pas1 Chunk (start?)

  000h 4     Chunk ID "pan1"
  004h 4     Chunk Size (0Ch)
  008h 4?    00000001h

txt1 Chunk

  000h 4     Chunk ID "txt1"
  004h 4     Chunk Size
  008h 2?    0008h
  00Ah 2?    0008h
  00Ch 2?    00F0h
  00Eh 2?    0150h
  010h 2?    0001h
  012h 2?    0015h  Number of 8-byte entries? start/end line-wrapping list?
  014h 2?    000Dh
  016h 2?    0010h
  018h 4?    00000000h
  01Ch 4?    00001CE7h
  020h 4?    00000000h
  024h 4?    00000000h
  028h N*8   Unknown 8-byte entries? (00xxh,0010h,0000h,00yyh)
             Or maybe positioning for symbols/images/tables?

pae1 Chunk (end?)

  000h 4     Chunk ID "pae1"
  004h 4     Chunk Size (08h)

NTTF File (manual graphics/symbols) (in "nttf" folder)

  000h 2  Bitmap Width in bytes (eg. 0Fh, 10h, 14h, 40h, 60h)
  002h 2  Bitmap Height         (eg. 12h, 0Eh, 14h, 30h, 47h)
  004h 1  Unknown (04h) (maybe color depth)
  005h 1  Unknown (00h or 01h) (often same as [007h], but not always)
  006h 2  Number of Palette entries (usually 10h, 80h, or 100h)
  008h 8  Zerofilled
  010h .. Palette data (with 16bit values in range 0000h..7FFFh)
  ...  .. Bitmap data (seems to be always 8bpp)

Notes
DSi manuals are using the TWLFontTable.dat system font (Flipnote, at least).
3DS manuals consist of CLYT and CLIM files (inside of a manual BCMA darc archive in a NCCH file).

DS Wireless Communications

DS Wifi I/O Map
DS Wifi Control
DS Wifi Interrupts
DS Wifi Power-Down Registers
DS Wifi Receive Control
DS Wifi Receive Buffer
DS Wifi Receive Statistics
DS Wifi Transmit Control
DS Wifi Transmit Buffers
DS Wifi Transmit Errors
DS Wifi Status
DS Wifi Timers
DS Wifi Multiplay Master
DS Wifi Multiplay Slave
DS Wifi Configuration Ports
DS Wifi Baseband Chip (BB)
DS Wifi RF Chip
DS Wifi RF9008 Registers
DS Wifi Unknown Registers
DS Wifi Unused Registers
DS Wifi Initialization
DS Wifi Flowcharts
DS Wifi Hardware Headers
DS Wifi Nintendo Beacons
DS Wifi Nintendo DS Download Play
DS Wifi IEEE802.11 Frames
DS Wifi IEEE802.11 Managment Frames (Type=0)
DS Wifi IEEE802.11 Control and Data Frames (Type=1 and 2)
DS Wifi WPA/WPA2 Handshake Messages (EAPOL)
DS Wifi WPA/WPA2 Keys and MICs
DS Wifi WPA/WPA2 Encryption
DS Wifi FFC ID
DS Wifi Dslink/Wifiboot Protocol
DS Firmware Wifi Calibration Data
DS Firmware Wifi Internet Access Points

2.4GHz band, Wireless LAN (WLAN) IEEE802.11b protocol

Credits
A very large part of the DS Wifi chapters is based on Stephen Stair's great DS Wifi document, thanks there.

DS Wifi I/O Map

Notice
Wifi Registers & RAM cannot be written to by STRB opcodes (ignored).

Registers - NDS7 - 4808000h..4808FFFh

  Address  Dir   Name            r/w  [Init] Description
  4808000h R     W_ID            ---- [1440] Chip ID (1440h=DS, C340h=DS-Lite)
  4808004h R/W   W_MODE_RST      9fff [0000] Mode/Reset
  4808006h R/W   W_MODE_WEP      --7f [0000] Mode/Wep modes
  4808008h R/W   W_TXSTATCNT     ffff [0000] Beacon Status Request
  480800Ah R/W   W_X_00Ah        ffff [0000] [bit7 - ingore rx duplicates]
  4808010h R/W   W_IF            ackk [0000] Wifi Interrupt Request Flags
  4808012h R/W   W_IE            ffff [0000] Wifi Interrupt Enable
  4808018h R/W   W_MACADDR_0     ffff [0000] Hardware MAC Address, 1st 2 bytes
  480801Ah R/W   W_MACADDR_1     ffff [0000] Hardware MAC Address, next 2 bytes
  480801Ch R/W   W_MACADDR_2     ffff [0000] Hardware MAC Address, last 2 bytes
  4808020h R/W   W_BSSID_0       ffff [0000] BSSID (first 2 bytes)
  4808022h R/W   W_BSSID_1       ffff [0000] BSSID (next 2 bytes)
  4808024h R/W   W_BSSID_2       ffff [0000] BSSID (last 2 bytes)
  4808028h R/W   W_AID_LOW       ---f [0000] usually as lower 4bit of AID value
  480802Ah R/W   W_AID_FULL      -7ff [0000] AID value assigned by a BSS.
  480802Ch R/W   W_TX_RETRYLIMIT ffff [0707] Tx Retry Limit (set from 00h-FFh)
  480802Eh R/W   W_INTERNAL      ---1 [0000]
  4808030h R/W   W_RXCNT         ff0e [0000] Receive control
  4808032h R/W   W_WEP_CNT       ffff [0000] WEP engine enable
  4808034h R?    W_INTERNAL      0000 [0000] bit0,1 (see ports 004h,040h,1A0h)

Power-Down Registers (and Random Generator)

  4808036h R/W   W_POWER_US      ---3 [0001]
  4808038h R/W   W_POWER_TX      ---7 [0003]
  480803Ch R/W   W_POWERSTATE    -r-2 [0200]
  4808040h R/W   W_POWERFORCE    8--1 [0000]
  4808044h R     W_RANDOM        0xxx [0xxx]
  4808048h R/W   W_POWER_?       ---3 [0000]

Receive Control/Memory

  4808050h R/W   W_RXBUF_BEGIN   ffff [4000]
  4808052h R/W   W_RXBUF_END     ffff [4800]
  4808054h R     W_RXBUF_WRCSR   0rrr [0000]
  4808056h R/W   W_RXBUF_WR_ADDR -fff [0000]
  4808058h R/W   W_RXBUF_RD_ADDR 1ffe [0000]
  480805Ah R/W   W_RXBUF_READCSR -fff [0000]
  480805Ch R/W   W_RXBUF_COUNT   -fff [0000]
  4808060h R     W_RXBUF_RD_DATA rrrr [xxxx]
  4808062h R/W   W_RXBUF_GAP     1ffe [0000]
  4808064h R/W   W_RXBUF_GAPDISP -fff [0000]

Transmit Control/Memory

  4808068h R/W   W_TXBUF_WR_ADDR 1ffe [0000]
  480806Ch R/W   W_TXBUF_COUNT   -fff [0000]
  4808070h W     W_TXBUF_WR_DATA xxxx [xxxx]
  4808074h R/W   W_TXBUF_GAP     1ffe [0000]
  4808076h R/W   W_TXBUF_GAPDISP 0fff [0000]
  4808078h W     W_INTERNAL      mirr [mirr] Read: Mirror of 068h
  4808080h R/W   W_TXBUF_BEACON  ffff [0000] Beacon Transmit Location
  4808084h R/W   W_TXBUF_TIM     --ff [0000] Beacon TIM Index in Frame Body
  4808088h R/W   W_LISTENCOUNT   --ff [0000] Listen Count
  480808Ch R/W   W_BEACONINT     -3ff [0064] Beacon Interval
  480808Eh R/W   W_LISTENINT     --ff [0000] Listen Interval
  4808090h R/W   W_TXBUF_CMD     ffff [0000] Multiplay Command
  4808094h R/W   W_TXBUF_REPLY1  ffff [0000] Multiplay Next Reply
  4808098h R     W_TXBUF_REPLY2  0000 [0000] Multiplay Current Reply
  480809Ch R/W   W_INTERNAL      ffff [0050] value 4x00h --> preamble+x*12h us?
  48080A0h R/W   W_TXBUF_LOC1    ffff [0000]
  48080A4h R/W   W_TXBUF_LOC2    ffff [0000]
  48080A8h R/W   W_TXBUF_LOC3    ffff [0000]
  48080ACh W     W_TXREQ_RESET   fixx [0050]
  48080AEh W     W_TXREQ_SET     fixx [0050]
  48080B0h R     W_TXREQ_READ    --1f [0010]
  48080B4h W     W_TXBUF_RESET   0000 [0000]    (used by firmware part4)
  48080B6h R     W_TXBUSY        0000 [0000]    (used by firmware part4)
  48080B8h R     W_TXSTAT        0000 [0000]
  48080BAh ?     W_INTERNAL      0000 [0000]
  48080BCh R/W   W_PREAMBLE      ---3 [0001]
  48080C0h R/W x W_CMD_TOTALTIME ffff [0000]    (used by firmware part4)
  48080C4h R/W x W_CMD_REPLYTIME ffff [0000]    (used by firmware part4)

xxx

  48080C8h ?     W_INTERNAL      0000 [0000]
  48080D0h R/W   W_RXFILTER      1fff [0401]
  48080D4h R/W   W_CONFIG_0D4h   ---3 [0001]
  48080D8h R/W   W_CONFIG_0D8h   -fff [0004]
  48080DAh R/W   W_RX_LEN_CROP   ffff [0602]
  48080E0h R/W   W_RXFILTER2     ---f [0008]

Wifi Timers

  48080E8h R/W   W_US_COUNTCNT   ---1 [0000] Microsecond counter enable
  48080EAh R/W   W_US_COMPARECNT ---1 [0000] Microsecond compare enable
  48080ECh R/W   W_CONFIG_0ECh   3f1f [3F03]
  48080EEh R/W   W_CMD_COUNTCNT  ---1 [0001]
  48080F0h R/W   W_US_COMPARE0   fc-- [FC00] Microsecond compare, bits 0-15
  48080F2h R/W   W_US_COMPARE1   ffff [FFFF] Microsecond compare, bits 16-31
  48080F4h R/W   W_US_COMPARE2   ffff [FFFF] Microsecond compare, bits 32-47
  48080F6h R/W   W_US_COMPARE3   ffff [FFFF] Microsecond compare, bits 48-63
  48080F8h R/W   W_US_COUNT0     ffff [0000] Microsecond counter, bits 0-15
  48080FAh R/W   W_US_COUNT1     ffff [0000] Microsecond counter, bits 16-31
  48080FCh R/W   W_US_COUNT2     ffff [0000] Microsecond counter, bits 32-47
  48080FEh R/W   W_US_COUNT3     ffff [0000] Microsecond counter, bits 48-63
  4808100h ?     W_INTERNAL      0000 [0000]
  4808102h ?     W_INTERNAL      0000 [0000]
  4808104h ?     W_INTERNAL      0000 [0000]
  4808106h ?     W_INTERNAL      0000 [0000]
  480810Ch R/W   W_CONTENTFREE   ffff [0000] ...
  4808110h R/W   W_PRE_BEACON    ffff [0000]
  4808118h R/W   W_CMD_COUNT     ffff [0000]
  480811Ch R/W   W_BEACON_COUNT  ffff [0000] reloaded with W_BEACONINT

Configuration Ports (and some other Registers)

  4808120h R/W   W_CONFIG_120h   81ff [0048] init from firmware[04Ch]
  4808122h R/W   W_CONFIG_122h   ffff [4840] init from firmware[04Eh]
  4808124h R/W   W_CONFIG_124h   ffff [0000] init from firmware[05Eh], or 00C8h
  4808126h ?     W_INTERNAL      fixx [ 0080]
  4808128h R/W   W_CONFIG_128h   ffff [0000] init from firmware[060h], or 07D0h
  480812Ah ?     W_INTERNAL      fixx [1000] lower 12bit same as W_CONFIG_128h
  4808130h R/W   W_CONFIG_130h   -fff [0142] init from firmware[054h]
  4808132h R/W   W_CONFIG_132h   8fff [8064] init from firmware[056h]
  4808134h R/W   W_POST_BEACON   ffff [FFFF] ...
  4808140h R/W   W_CONFIG_140h   ffff [0000] init from firmware[058h], or xx
  4808142h R/W   W_CONFIG_142h   ffff [2443] init from firmware[05Ah]
  4808144h R/W   W_CONFIG_144h   --ff [0042] init from firmware[052h]
  4808146h R/W   W_CONFIG_146h   --ff [0016] init from firmware[044h]
  4808148h R/W   W_CONFIG_148h   --ff [0016] init from firmware[046h]
  480814Ah R/W   W_CONFIG_14Ah   --ff [0016] init from firmware[048h]
  480814Ch R/W   W_CONFIG_14Ch   ffff [162C] init from firmware[04Ah]
  4808150h R/W   W_CONFIG_150h   ff3f [0204] init from firmware[062h], or 202h
  4808154h R/W   W_CONFIG_154h   7a7f [0058] init from firmware[050h]

Baseband Chip Ports

  4808158h W     W_BB_CNT        mirr [00B5] BB Access Start/Direction/Index
  480815Ah W     W_BB_WRITE      ???? [0000] BB Access data byte to write
  480815Ch R     W_BB_READ       00rr [00B5] BB Access data byte read
  480815Eh R     W_BB_BUSY       000r [0000] BB Access Busy flag
  4808160h R/W   W_BB_MODE       41-- [0100] BB Access Mode
  4808168h R/W   W_BB_POWER      8--f [800D] BB Access Powerdown

Internal Stuff

  480816Ah ?     W_INTERNAL      0000 [0001] (or 0000h?)
  4808170h ?     W_INTERNAL      0000 [0000]
  4808172h ?     W_INTERNAL      0000 [0000]
  4808174h ?     W_INTERNAL      0000 [0000]
  4808176h ?     W_INTERNAL      0000 [0000]
  4808178h W     W_INTERNAL      fixx [0800] Read: mirror of 17Ch

RF Chip Ports

  480817Ch R/W   W_RF_DATA2      ffff [0800]
  480817Eh R/W   W_RF_DATA1      ffff [C008]
  4808180h R     W_RF_BUSY       000r [0000]
  4808184h R/W   W_RF_CNT        413f [0018]

xxx

  4808190h R/W   W_INTERNAL      ffff [0000]
  4808194h R/W   W_TX_HDR_CNT    ---7 [0000] used by firmware part4 (0 or 6)
  4808198h R/W   W_INTERNAL      ---f [0000]
  480819Ch R     W_RF_PINS       fixx [0004]
  48081A0h R/W   W_X_1A0h        -933 [0000] used by firmware part4 (0 or 823h)
  48081A2h R/W   W_X_1A2h        ---3 [0001] used by firmware part4
  48081A4h R/W   W_X_1A4h        ffff [0000] "Rate used when signal test..."

Wifi Statistics

  48081A8h R     W_RXSTAT_INC_IF rrrr [0000] Stats Increment Flags
  48081AAh R/W   W_RXSTAT_INC_IE ffff [0000] Stats Increment IRQ Enable
  48081ACh R     W_RXSTAT_OVF_IF rrrr [0000] Stats Half-Overflow Flags
  48081AEh R/W   W_RXSTAT_OVF_IE ffff [0000] Stats Half-Overflow IRQ Enable
  48081B0h R/W   W_RXSTAT        --ff [0000]
  48081B2h R/W   W_RXSTAT        ffff [0000] RX_LengthRateErrorCount
  48081B4h R/W   W_RXSTAT        rrff [0000] ... firmware uses also MSB ... ?
  48081B6h R/W   W_RXSTAT        ffff [0000]
  48081B8h R/W   W_RXSTAT        --ff [0000]
  48081BAh R/W   W_RXSTAT        --ff [0000]
  48081BCh R/W   W_RXSTAT        ffff [0000]
  48081BEh R/W   W_RXSTAT        ffff [0000]
  48081C0h R/W   W_TX_ERR_COUNT  --ff [0000] TransmitErrorCount
  48081C4h R     W_RX_COUNT      fixx [0000]

[1D0 - 1DE are 15 entries related to multiplayer response errors]

  48081D0h R/W   W_CMD_STAT      ff-- [0000]
  48081D2h R/W   W_CMD_STAT      ffff [0000]
  48081D4h R/W   W_CMD_STAT      ffff [0000]
  48081D6h R/W   W_CMD_STAT      ffff [0000]
  48081D8h R/W   W_CMD_STAT      ffff [0000]
  48081DAh R/W   W_CMD_STAT      ffff [0000]
  48081DCh R/W   W_CMD_STAT      ffff [0000]
  48081DEh R/W   W_CMD_STAT      ffff [0000]

Internal Diagnostics Registers (usually not used for anything)

  48081F0h R/W   W_INTERNAL      ---3 [0000]
  4808204h ?     W_INTERNAL      fixx [0000]
  4808208h ?     W_INTERNAL      fixx [0000]
  480820Ch W     W_INTERNAL      fixx [0050]
  4808210h R     W_TX_SEQNO      fixx [0000]
  4808214h R     W_RF_STATUS     XXXX [0009]    (used by firmware part4)
  480821Ch W     W_IF_SET        fbff [0000] Force Interrupt (set bits in W_IF)
  4808220h R/W   W_RAM_DISABLE   ffff [0000] WifiRAM control
  4808224h R/W   W_INTERNAL      ---3 [0003]
  4808228h W     W_X_228h        fixx [0000]    (used by firmware part4) (bit3)
  4808230h R/W   W_INTERNAL      --ff [0047]
  4808234h R/W   W_INTERNAL      -eff [0EFF]
  4808238h R/W   W_INTERNAL      ffff [0000] ;rx_seq_no-60h+/-x   ;why that?
                                   ;other day: fixed value, not seq_no related?
  480823Ch ?     W_INTERNAL      fixx [0000] like W_TXSTAT... ONLY for beacons?
  4808244h R/W   W_X_244h        ffff [0000]    (used by firmware part4)
  4808248h R/W   W_INTERNAL      ffff [0000]
  480824Ch R     W_INTERNAL      fixx [0000] ;rx_mac_addr_0 ;\OverTheHedge
  480824Eh R     W_INTERNAL      fixx [0000] ;rx_mac_addr_1 ;/writes FFFFh?
  4808250h R     W_INTERNAL      fixx [0000] ;rx_mac_addr_2
  4808254h ?     W_CONFIG_254h   fixx [0000] (read: FFFFh=DS, EEEEh=DS-Lite)
  4808258h ?     W_INTERNAL      fixx [0000]
  480825Ch ?     W_INTERNAL      fixx [0000]
  4808260h ?     W_INTERNAL      fixx [ 0FEF]
  4808264h R     W_INTERNAL      fixx [0000] ;rx_addr_1 (usually "rxtx_addr-x")
  4808268h R     W_RXTX_ADDR     fixx [0005] ;rxtx_addr
  4808270h R     W_INTERNAL      fixx [0000] ;rx_addr_2 (usually "rx_addr_1-1")
  4808274h ?     W_INTERNAL      fixx [ 0001]
  4808278h R/W   W_INTERNAL      ffff [000F]
  480827Ch ?     W_INTERNAL      fixx [ 000A]
  4808290h (R/W) W_X_290h        fixx [FFFF] bit 0 = ? (used by firmware part4)
  4808298h W     W_INTERNAL      fixx [0000]
  48082A0h R/W   W_INTERNAL      ffff [0000]
  48082A2h R     W_INTERNAL      XXXX [7FFF] 15bit shift reg (used during tx?)
  48082A4h R     W_INTERNAL      fixx [0000] ;rx_rate_1 not ALWAYS same as 2C4h
  48082A8h W     W_INTERNAL      fixx [0000]
  48082ACh ?     W_INTERNAL      fixx [ 0038]
  48082B0h W     W_INTERNAL      fixx [0000]
  48082B4h R/W   W_INTERNAL      -1-3 [0000]
  48082B8h ?     W_INTERNAL      fixx [0000] ;dsi launcher checks if zero
  48082C0h R/W   W_INTERNAL      ---1 [0000]
  48082C4h R     W_INTERNAL      fixx [000A] ;rx_rate_2 (0Ah,14h = 1,2 Mbit/s)
  48082C8h R     W_INTERNAL      fixx [0000] ;rx_duration/length/rate (or so?)
  48082CCh R     W_INTERNAL      fixx [0000] ;rx_framecontrol; from ieee header
  48082D0h DIS   W_INTERNAL                  ;"W_POWERACK" (internal garbage)
                                             ;normally DISABLED (unless FORCE)
  48082F0h R/W   W_INTERNAL      ffff [0000]
  48082F2h R/W   W_INTERNAL      ffff [0000]
  48082F4h R/W   W_INTERNAL      ffff [0000]
  48082F6h R/W   W_INTERNAL      ffff [0000]

All other ports in range 4808000h..4808FFFh are unused.
All registers marked as "W_INTERNAL" aren't used by Firmware part4, and are probably unimportant, except for whatever special diagnostics purposes.
Reading from write-only ports (W) often mirrors to data from other ports.

Additionally, there are 69h Baseband Chip Registers (BB), and 0Fh RF Chip Registers (see BB and RF chapters).

For Wifi Power Managment (POWCNT2), for Wifi Waitstates (WIFIWAITCNT), see:
DS Power Control
For the Power LED Blink Feature (conventionally used to indicate Wifi activity) see:
DS Power Management Device

For Wifi Configuration and Calibration data in Firmware Header, see:
DS Cartridges, Encryption, Firmware

Wifi RAM - NDS7 - Memory (4804000h..4805FFFh)

  4804000h W_MACMEM RX/TX Buffers (2000h bytes) (excluding below specials)
  4805F60h Used for something, not included in the rx circular buffer.
  4805F80h W_WEPKEY_0 (32 bytes)
  4805FA0h W_WEPKEY_1 (32 bytes)
  4805FC0h W_WEPKEY_2 (32 bytes)
  4805FE0h W_WEPKEY_3 (32 bytes)

Unlike all other NDS memory, Wifi RAM is left uninitialized after boot.

4805F60h - Reserved 20h-byte RAM area
Nintendo's software does init/change several values:

  [480xxxxh]=5A5Ah/A5A5h  ;-initial dummy WifiRAM memfill values
  [4805F70h]=FFFFh        ;\
  [4805F72h]=FFFFh        ; set to FFFFh by software
  [4805F76h]=FFFFh        ;
  [4805F7Eh]=FFFFh        ;/

The hardware does update several values (unless disabled in W_RAM_DISABLE.bit5):

  [4805F6Eh]=0F00h (nothing received), or 0F01h (received something)

And, if received packet was accepted (eg. when W_RXFILTER.bit0=1=AcceptAll):

  [4805F70h]=Received MAC Address (6 bytes, looks same as port 480824Ch)
  [4805F76h]=xxx0h (increasing value, Sequence Control from packet header)

And, in multiplay slave mode (when receiving CMD packets?):

  [4805F7Eh]=xxx0h (next higher sequence number? ie. [4805F76h]+10h)

4805F80h - W_WEPKEY_0 thru W_WEPKEY_3 - Wifi WEP keys (R/W)
These WEP key slots store the WEP keys that are used for encryption for 802.11 keys IDs 0-3.

DS Wifi Control

4808000h - W_ID - Wifi Chip ID (R)

  0-15   Chip ID (1440h on NDS, C340h on NDS-lite)

The NDS-lite is more or less backwards compatible with the original NDS (the W_RXBUF_GAPDISP and W_TXBUF_GAPDISP are different, and most of the garbage effects on unused/mirrored ports are different, too).

4808004h - W_MODE_RST - Wifi Hardware mode / reset (R/W)

  0     Adjust some ports (0/1=see lists below) (R/W)
        TX Master Enable for LOC1..3 and Beacon  (0=Disable, 1=Enable)
  1-12  Unknown (R/W)
  13    Reset some ports (0=No change, 1=Reset/see list below) (Write-Only)
  14    Reset some ports (0=No change, 1=Reset/see list below) (Write-Only)
  15    Unknown (R/W)

4808006h - W_MODE_WEP - Wifi Software mode / Wep mode (R/W)

  0-2   Unknown, specify a software mode for wifi operation
        (may be related to hardware but a correlation has not yet been found)
  3-5   WEP Encryption Key Size:
         0=Reserved (acts same as 1)
         1=64bit WEP  (IV=24bit + KEY=40bit)  (aka 3+5 bytes)   ;standard/us
         2=128bit WEP (IV=24bit + KEY=104bit) (aka 3+13 bytes)  ;standard/world
         3=152bit WEP (IV=24bit + KEY=128bit) (aka 3+16 bytes)  ;uncommon
         4=Unknown, mabye 256bit WEP (IV=24bit + KEY=232bit) (aka 3+29 bytes)?
         5=Reserved (acts same as 1)
         6=Reserved (acts same as 1)
         7=Reserved (acts same as 1)
  6     Unknown
  7-15  Always zero

4808018h - W_MACADDR_0 - MAC Address (R/W)
480801Ah - W_MACADDR_1 - MAC Address (R/W)
480801Ch - W_MACADDR_2 - MAC Address (R/W)
48bit MAC Address of the console. Should be initialized from firmware[036h]. The hardware receives only packets that are sent to this address (or to group addresses, like FF:FF:FF:FF:FF:FF).

4808020h - W_BSSID_0 - BSSID (R/W)
4808022h - W_BSSID_1 - BSSID (R/W)
4808024h - W_BSSID_2 - BSSID (R/W)
48bit BSSID stored here. Ie. the MAC address of the host, obtained from Beacon frames (on the host itself, that should be just same as W_MACADDR). See W_RXFILTER.

4808028h - W_AID_LOW (R/W)

  Bit0-3   Multiplay Slave number (1..15, or 0)
  Bit4-15  Not used

Usually set to zero, or equal to the lower 4bit of the W_AID_FULL value.

480802Ah - W_AID_FULL - Association ID (R/W)

  Bit0-10  Association ID (AID) (1..2007, or zero)
  Bit11-15 Not used

4808032h - W_WEP_CNT - WEP Engine Enable (R/W)

  0-14  Unknown (usually zero)
  15    WEP Engine Enable  (0=Disable, 1=Enable)

Normally, bit15 should be always set (but it will only affect WEP-encrypted packets, ie. packets with Frame Control bit14=1 in 802.11 header). When disabled, WEP packets aren't received at all (neither in encrypted nor decrypted form), and sending WEP packets might be also ignored(?).
The firmware contains some code that does toggle the bit off for a moment (apparently to reset something after transfer errors).

4808044h - W_RANDOM - Random Generator (R)

  0-10  Random
  11-15 Not used (zero)

The random generator is updated at 33.51MHz rate, as such:

  X = (X AND 1) XOR (X ROL 1)  ;(rotation within 11bit range)

That random sequence goes through 5FDh different values before it restarts.
When reading from the random register, the old latched value is returned to the CPU, and the new current random value is then latched, so reads always return the older value, timed from the previous read.
Occassionally, about once every some thousand reads, the latching appears to occur 4 cycles earlier than normally, so the value on the next read will be 4 cycles older than expected.
The random register has ACTIVE mirrors.

48080BCh - W_PREAMBLE - Preamble Control (R/W)

  Bit   Dir  Expl.
  0     R/W  Unknown                    (this does NOT affect TX)
  1     R/W  Preamble (0=Long, 1=Short) (this does NOT affect TX)
  2     W    Preamble (0=Long, 1=Short) (this does affect TX) (only at 2Mbit/s)
  3-15  -    Always zero

Short preamble works only with 2Mbit/s transfer rate (ie. when set as so in TX hardware header). 1Mbit/s rate always uses long preamble.

  Type   Carrier Signal  SFD Value     PLCP Header     Data
  Long   128bit, 1Mbit   16bit, 1Mbit  48bit, 1Mbit    N bits, 1Mbit or 2Mbit
  Short  56bit, 1Mbit    16bit, 1Mbit  48bit, 2Mbit    N bits, 2Mbit

Length of the Carrier+SFD+PLCP part is thus 192us (long) or 96us (short).
Note: The Carrier+SFD+PLCP part is sent between IRQ14 and IRQ07 (not between IRQ07 and IRQ01).
Firmware also changes W_CONFIG_140h alongsides with the preamble setting.

Writing "0-then-1" to W_MODE_RST.Bit0 does reset following ports:

  [4808034h]=0002h ;W_INTERNAL
  [480819Ch]=0046h ;W_RF_PINS
  [4808214h]=0009h ;W_RF_STATUS
  [480827Ch]=0005h ;W_INTERNAL
  [48082A2h]=?     ;...unstable?

Writing "1-then-0" to W_MODE_RST.Bit0 does reset following ports:

  [480827Ch]=000Ah ;W_INTERNAL

Writing "1" to W_MODE_RST.Bit13 does reset following ports:

  [4808056h]=0000h ;W_RXBUF_WR_ADDR
  [48080C0h]=0000h ;W_CMD_TOTALTIME
  [48080C4h]=0000h ;W_CMD_REPLYTIME
  [48081A4h]=0000h ;W_X_1A4h
  [4808278h]=000Fh ;W_INTERNAL
  ...Also, following may be affected (results are unstable though)...
  [48080AEh]=?     ;or rather the actual port (which it is an mirror of)
  [48080BAh]=?     ;W_INTERNAL (occassionally unstable)
  [4808204h]=?     ;W_INTERNAL
  [480825Ch]=?     ;W_INTERNAL
  [4808268h]=?     ;W_RXTX_ADDR
  [4808274h]=?     ;W_INTERNAL

Writing "1" to W_MODE_RST.Bit14 does reset following ports:

  [4808006h]=0000h ;W_MODE_WEP
  [4808008h]=0000h ;W_TXSTATCNT
  [480800Ah]=0000h ;W_X_00Ah
  [4808018h]=0000h ;W_MACADDR_0
  [480801Ah]=0000h ;W_MACADDR_1
  [480801Ch]=0000h ;W_MACADDR_2
  [4808020h]=0000h ;W_BSSID_0
  [4808022h]=0000h ;W_BSSID_1
  [4808024h]=0000h ;W_BSSID_2
  [4808028h]=0000h ;W_AID_LOW
  [480802Ah]=0000h ;W_AID_FULL
  [480802Ch]=0707h ;W_TX_RETRYLIMIT
  [480802Eh]=0000h ;W_INTERNAL
  [4808050h]=4000h ;W_RXBUF_BEGIN
  [4808052h]=4800h ;W_RXBUF_END
  [4808084h]=0000h ;W_TXBUF_TIM
  [48080BCh]=0001h ;W_PREAMBLE
  [48080D0h]=0401h ;W_RXFILTER
  [48080D4h]=0001h ;W_CONFIG_0D4h
  [48080E0h]=0008h ;W_RXFILTER2
  [48080ECh]=3F03h ;W_CONFIG_0ECh
  [4808194h]=0000h ;W_TX_HDR_CNT
  [4808198h]=0000h ;W_INTERNAL
  [48081A2h]=0001h ;W_X_1A2h
  [4808224h]=0003h ;W_INTERNAL
  [4808230h]=0047h ;W_INTERNAL

DS Wifi Interrupts

4808010h - W_IF - Wifi Interrupt Request Flags (R/W)

  0   Receive Complete  (packet received and stored in the RX fifo)
  1   Transmit Complete (packet is done being transmitted) (no matter if error)
  2   Receive Event Increment        (IRQ02, see W_RXSTAT_INC_IE)
  3   Transmit Error Increment       (IRQ03, see W_TX_ERR_COUNT)
  4   Receive Event Half-Overflow    (IRQ04, see W_RXSTAT_OVF_IE)
  5   Transmit Error Half-Overflow   (IRQ05, see W_TX_ERR_COUNT.Bit7)
  6   Start Receive     (IRQ06, a packet has just started to be received)
  7   Start Transmit    (IRQ07, a packet has just started to be transmitted)
  8   Txbuf Count Expired            (IRQ08, see W_TXBUF_COUNT)
  9   Rxbuf Count Expired            (IRQ09, see W_RXBUF_COUNT)
  10  Not used (always zero, even when trying to set it with W_IF_SET)
  11  RF Wakeup                      (IRQ11, see W_POWERSTATE)
  12  Multiplay CMD done (or failed) (IRQ12, see W_CMD_COUNT)
  13  Post-Beacon Timeslot           (IRQ13, see W_POST_BEACON)
  14  Beacon Timeslot                (IRQ14, see W_BEACON_COUNT/W_US_COMPARE)
  15  Pre-Beacon Timeslot            (IRQ15, see W_BEACON_COUNT/W_PRE_BEACON)

Write a '1' to a bit to clear it. For the Half-Overflow flags that works ONLY if the counter MSBs are zero (ie. one must first read the counters (to reset them), and THEN acknowledge the corresponding W_IF bit).
The Transmit Start/Complete bits (Bit7,1) are set for EACH packet (including beacons, and including retries).

4808012h - W_IE - Wifi Interrupt Enable Flags (R/W)

  0-15  Enable Flags, same bits as W_IF  (0=Disable, 1=Enable)

In W_IE, Bit10 is R/W, but seems to have no function since IRQ10 doesn't exist.

480821Ch - W_IF_SET (W_INTERNAL) - Force Wifi Interrupt Flags (W)

  0-15  Set corresponding bits in W_IF  (0=No change, 1=Set Bit)

Notes: Bit10 cannot be set since no IRQ10 exists. This register does only set IRQ flags, but without performing special actions (such like W_BEACON_COUNT and W_POST_BEACON reloads that occur on real IRQ14's).

Wifi Primary IRQ Flag (IF.Bit24, Port 4000214h)
IF.Bit24 gets set <only> when (W_IF AND W_IE) changes from 0000h to non-zero.
IF.Bit24 can be reset (ack) <even> when (W_IF AND W_IE) is still non-zero.

  Caution  Caution  Caution  Caution  Caution
  That means, when acknowledging IF.Bit24, then NO FURTHER wifi IRQs
  will be executed whilst and as long as (W_IF AND W_IE) is non-zero.

One work-around is to process/acknowledge ALL wifi IRQs in a loop, including further IRQs that may occur inside of that loop, until (W_IF AND W_IE) becomes 0000h.
Another work-around (for single IRQs) would be to acknowledge IF and W_IF, and then to set W_IE temporarily to 0000h, and then back to the old W_IE setting.

DS Wifi Power-Down Registers

4808036h - W_POWER_US (R/W)

  0     Disable W_US_COUNT and W_BB_ports  (0=Enable, 1=Disable)
  1     Unknown (usually 0)
  2-15  Always zero

Bit0=0 enables RFU by setting RFU.Pin11=HIGH, which activates the 22.000MHz oscillator on the RFU board, the 22MHz clock is then output to RFU.Pin26.

4808038h - W_POWER_TX (R/W)
transmit-related power save or sth
init from firmware[05Ch]

  0     Auto Wakeup (1=Leave Idle Mode a while after Pre-Beacon IRQ15)
  1     Auto Sleep  (0=Enter Idle Mode on Post-Beacon IRQ13)
  2     Unknown
  3     Unknown (Write-only) (used by firmware)
  4-15  Always zero

480803Ch - W_POWERSTATE (R/W)/(R)

  0     Unknown (usually 0)                         (R/W)
  1     Request Power Enable (0=No, 1=Yes/queued)   (R/W, but not always)
  2-7   Always zero
  8     Indicates that Bit9 is about the be cleared (Read only)
  9     Current power state (0=Enabled, 1=Disabled) (Read only)
  10-15 Always zero

[value =1: queue disable power state] ;<-- seems to be incorrect
[value =2: queue enable power state] ;<-- seems to be correct
Enabling causes wakeup interrupt (IRQ11).
Note: That queue stuff seems to work only if W_POWER_US=0 and W_MODE_RST=1.

4808040h - W_POWERFORCE - Force Power State (R/W)

  0     New value for W_POWERSTATE.Bit9  (0=Clear/Delayed, 1=Set/Immediately)
  1-14  Always zero
  15    Apply Bit0 to W_POWERSTATE.Bit9  (0=No, 1=Yes)

Setting W_POWERFORCE=8001h whilst W_POWERSTATE.Bit9=0 acts immediately:

  (Doing this is okay. Switches to power down mode. Similar to IRQ13.)
  [4808034h]=0002h ;W_INTERNAL
  [480803Ch]=02xxh ;W_POWERSTATE
  [48080B0h]=0000h ;W_TXREQ_READ
  [480819Ch]=0046h ;W_RF_PINS
  [4808214h]=0009h ;W_RF_STATUS (idle)

Setting W_POWERFORCE=8000h whilst W_POWERSTATE.Bit9=1 acts delayed:

  (Don't do this. After that sequence, the hardware seems to be messed up)
  W_POWERSTATE.Bit8 gets set to indicate the pending operation,
  while pending, changes to W_POWERFORCE aren't applied to W_POWERSTATE,
  while pending, W_POWERACK becomes Read/Write-able,
  writing 0000h to W_POWERACK does clear W_POWERSTATE.Bit8,
  and does apply POWERFORCE.Bit0 to W_POWERSTATE.Bit9
  and does deactivate Port W_POWERACK again.

4808048h - W_POWER_? (R/W)

  0     Unknown
  1     Unknown
  2-15  Always zero

At whatever time (during transmit or so) it gets set to 0003h by hardware.

See also: POWCNT2, W_BB_POWER.

DS Wifi Receive Control

4808030h - W_RXCNT - Wifi Receive Control (parts R/W and W)

  0     Copy W_RXBUF_WR_ADDR to W_RXBUF_WRCSR (aka force RX buf empty)      (W)
  1-3   Unknown                                                           (R/W)
  4-6   Always zero
  7     Copy W_TXBUF_REPLY1 to W_TXBUF_REPLY2, set W_TXBUF_REPLY1 to 0000h  (W)
  8-14  Unknown                                                           (R/W)
  15    Enable Queuing received data to RX FIFO                           (R/W)

48080D0h - W_RXFILTER - (R/W)

  0     For Broadcasts? (0=Insist on W_BSSID, 1=Accept no matter of W_BSSID)
  1     Unknown (usually zero)
  2     Unknown (usually zero)
  3     Unknown (usually zero)
  4     Unknown (usually zero)
  5     Unknown (usually zero)
  6     Unknown (usually zero)
  7     Unknown (0 or 1)
  8     Empty Packets (0=Ignore, 1=Accept; with RXHDR[0]=801Fh)
  9     Unknown (0 or 1)
  10    Unknown (0 or 1)       (when set, receives beacons, and maybe others)
  11    Unknown (usually zero)     ;reportedly "allow toDS" ?
  12    Update W_RXBUF_WRCSR after IEEE header (instead after full packets?)
         (setting bit12 causes a mess, where new "packets" in RX buf could
         either contain RXHDR+IEEE header, or Data corresponding to that
         headers, which could be useful only if there's a way to distinguish
         between headers and data, and knowing the size of the data blocks).
  13-15 Not used (always zero)

Specifies what packets to allow. Some random notes...

  0000h = Disable receive.
  FFFFh = Enable receive.
  0400h = Receives managment frames (and possibly others, too)

The exact meaning of the bits is unknown. The most import part is the address filtering based on the DA and BSSID values in IEEE header:

  DA=W_MACADDR                    is always received
  DA=Broadcast, and BSSID=W_BSSID is always received
  DA=Broadcast, and BSSID=other   is received only if RXFILTER.bit0=1

Broadcast addresses (aka group addresses) have DA.firstbyte.bit0=1. RXFILTER.bit0 should be set when searching beacons, and cleared after connecting to an access point.

48080E0h - W_RXFILTER2 - (R/W)

  0     Unknown (0=Receive Data Frames, 1=Ignore Data Frames) (?)
  1     Unknown
  2     Unknown
  3     Unknown (usually set)
  4-15  Not used (always zero)

Firmware writes values 08h, 0Bh, 0Dh (aka 1000b, 1011b, 1101b).
Firmware usually has bit0 set, even when receiving data frames, so, in some situations data frames seem to pass-through even when bit0 is set...? Possibly that situation is when W_BSSID matches...?
Control/PS-Poll frames seem to be passed always (even if W_RXFILTER2=0Fh).

DS Wifi Receive Buffer

The dimensions of the circular Buffer are set with BEGIN/END values, hardware automatically wraps to BEGIN when an incremented pointer hits END address.

Write Area
Memory between WRCSR and READCSR is free for receiving data, the hardware writes incoming packets to this region (to WRCSR and up) (but without exceeding READCSR), once when it has successfully received a complete packet, the hardware moves WRCSR after the packet (aligned to a 4-byte boundary).

Read Area
Memory between READCSR and WRCSR contains received data, which can be read by the CPU via RD_ADDR and RD_DATA registers (or directly from memory). Once when having processed that data, the CPU must set READCSR to the end of it.

4808050h - W_RXBUF_BEGIN - Wifi RX Fifo start location (R/W)
4808052h - W_RXBUF_END - Wifi RX Fifo end location (R/W)

  0-15  Byte-offset in Wifi Memory (usually 4000h..5FFEh)

Although the full 16bit are R/W, only the 12bit halfword offset in Bit1-12 is actually used, the other bits seem to have no effect.
Some or all (?) of the below incrementing registers are automatically matched to begin/end, that is, after incrementing, IF adr=end THEN adr=begin.

4808054h - W_RXBUF_WRCSR - Wifi RX Fifo Write or "end" cursor (R)

  0-11  Halfword Address in RAM
  12-15 Always zero

Hardware does set this register pointing to end of the most recently received packet (plus 32bit alignment padding). Hardware will write the next packet to that address. And more importantly, software can read from RX fifo up to that address.

4808056h - W_RXBUF_WR_ADDR - Wifi RX Fifo Write Cursor Latch value (R/W)

  0-11  Halfword Address in RAM
  12-15 Always zero

This is a value that is latched into W_RXBUF_WRCSR, when the W_RXCNT latch bit (W_RXCNT.Bit0) is written.

4808058h - W_RXBUF_RD_ADDR - Wifi CircBuf Read Address (R/W)

  0     Always zero
  1-12  Halfword Address in RAM for reading via W_RXBUF_RD_DATA
  13-15 Always zero

The circular buffer limits are the same as the range specified for the receive FIFO, however the address can be set outside of that range and will only be affected by the FIFO boundary if it crosses the FIFO end location by reading from the circular buffer.

480805Ah - W_RXBUF_READCSR - Wifi RX Fifo Read or "start" cursor (R/W)

  0-11  Halfword Address in RAM
  12-15 Always zero

This value is specified the same as W_RXBUF_WRCSR - it's purely software controlled so it's up to the programmer to move the start cursor after loading a packet. If W_RXBUF_READCSR != W_RXBUF_WRCSR, then one or more packets exist in the FIFO that need to be processed (see the section on HW RX Headers, for information on calculating packet lengths). Once a packet has been processed, the software should advance the read cursor to the beginning of the next packet.

4808060h - W_RXBUF_RD_DATA - Wifi CircBuf Read Data (R)

  0-15  Data

returns the 16bit value at the address specified by W_RXBUF_RD_ADDR, and increments W_RXBUF_RD_ADDR by 2. If the increment causes W_RXBUF_RD_ADDR to equal the address specified in W_RXBUF_END, W_RXBUF_RD_ADDR will be reset to the address specified in W_RXBUF_BEGIN.
Ports 1060h, 6060h, 7060h are PASSIVE mirrors of 0060h, reading from these mirrors returns the old latched value from previous read from 0060h, but without reading a new value from RAM, and without incrementing the address.

4808062h - W_RXBUF_GAP - Wifi RX Gap Address (R/W)

  0     Always zero
  1-12  Halfword Address in RAM
  13-15 Always zero

Seems to be intended to define a "gap" in the circular buffer, done like so:

  Addr=Addr+2 and 1FFEh  ;address increment (by W_RXBUF_RD_DATA read)
  if Addr=RXBUF_END then ;normal begin/end wrapping (done before gap wraps)
     Addr=RXBUF_BEGIN
  if Addr=RXBUF_GAP then ;now gap-wrap (may include further begin/end wrap)
     Addr=RXBUF_GAP+RXBUF_GAPDISP*2
     if Addr>=RXBUF_END then Addr=Addr+RXBUF_BEGIN-RXBUF_END  ;wrap more

To disable the gap stuff, set both W_RXBUF_GAP and W_RXBUF_GAPDISP to zero.

4808064h - W_RXBUF_GAPDISP - Wifi RX Gap Displacement Offset (R/W)

  0-11  Halfword Offset, used with W_RXBUF_GAP (see there)
  12-15 Always zero

Caution: On the DS-Lite, after adding it to W_RXBUF_RD_ADDR, the W_RXBUF_GAPDISP setting is destroyed (reset to 0000h) by hardware. The original DS leaves W_RXBUF_GAPDISP intact.

480805Ch - W_RXBUF_COUNT (R/W)

  0-11  Decremented on reads from W_RXBUF_RD_DATA
  12-15 Always zero

Triggers IRQ09 when it reaches zero, and does then stay at zero (without further decrementing, and without generating further IRQs).
Note: Also decremented on (accidental) writes to read-only W_RXBUF_RD_DATA.

DS Wifi Receive Statistics

48081A8h - W_RXSTAT_INC_IF - Statistics Increment Flags (R)

  0-12   Increment Flags (see Port 48081B0h..1BFh)
  13-15  Always zero

Bitmask for which statistics have been increased at least once.
Unknown how to reset/acknowledge these bits (neither reading/writing 48081A8h, nor reading 48081B0h..1BFh, nor writing 48081ACh seem to work).

48081AAh - W_RXSTAT_INC_IE - Statistics Increment Interrupt Enable (R/W)

  0-12   Counter Increment Interrupt Enable (see 48081B0h..1BFh) (1=Enable)
  13-15  Unknown (usually zero)

Statistic Interrupt Enable Control register for Count Up.
Note: ------> seems to trigger IRQ02 ...?

48081ACh - W_RXSTAT_OVF_IF - Statistics Half-Overflow Flags (R)

  0-12   Half-Overflow Flags (see Port 48081B0h..1BFh)
  13-15  Always zero

The W_RXSTAT_OVF_IF bits are simply containing the current bit7-value of the corresponding counters, setting or clearing that counter bits is directly reflected to W_RXSTAT_OVF_IF.
The recommended way to acknowledge W_RXSTAT_OVF_IF is to read the corresponding counters (which are reset to 00h after reading). For some reason, the firmware is additionally writing FFFFh to W_RXSTAT_OVF_IF (that is possibly a bug, or it does acknowledge something internally?).

48081AEh - W_RXSTAT_OVF_IE - Statistics Half-Overflow Interrupt Enable (R/W)

  0-12   Half-Overflow Interrupt Enable (see Port 48081B0h..1BFh) (1=Enable)
  13-15  Unknown (usually zero)

Statistic Interrupt Enable for Overflow, bits same as in W_RXSTAT_INC_IE
Note: ------> seems to trigger IRQ04 ...?

48081B0h..1BFh - W_RXSTAT - Receive Statistics (R/W, except 1B5h: Read-only)
W_RXSTAT is a collection of 8bit counters, which are incremented upon certain events. These entries are automatically reset to 0000h after reading. Should be accessed with LDRH opcodes (using LDRB to read only 8bits does work, but the read is internally expanded to 16bit, and so, the whole 16bit value will be reset to 0000h).

  Port      Dir  Bit  Expl.
  48081B0h  R/W  0    W_RXSTAT  ?
  48081B1h  -    -    Always 0  -
  48081B2h  R/W  1    W_RXSTAT  ?    "RX_RateErrorCount"
  48081B3h  R/W  2    W_RXSTAT  Length>2348 error
  48081B4h  R/W  3    W_RXSTAT  RXBUF Full error
  48081B5h  R    4?   W_RXSTAT  ?    (R) (but seems to exist; used by firmware)
  48081B6h  R/W  5    W_RXSTAT  Length=0 or Wrong FCS Error
  48081B7h  R/W  6    W_RXSTAT  Packet Received Okay
                             (also increments on W_MACADDR mis-match)
                             (also increments on internal ACK packets)
                             (also increments on invalid IEEE type=3)
                             (also increments TOGETHER with 1BCh and 1BEh)
                             (not incremented on RXBUF_FULL error)
  48081B8h  R/W  7    W_RXSTAT  ?
  48081B9h  -    -    Always 0  -
  48081BAh  R/W  8    W_RXSTAT  ?
  48081BBh  -    -    Always 0  -
  48081BCh  R/W  9    W_RXSTAT  WEP Error (when FC.Bit14 is set)
  48081BDh  R/W  10   W_RXSTAT  ?
  48081BEh  R/W  11   W_RXSTAT  (duplicated sequence control)
  48081BFh  R/W  12   W_RXSTAT  ?

48081C4h - W_RX_COUNT (W_INTERNAL) (R)

  0-?   Receive Okay Count (increments together with ports 48081B4h, 48081B7h)
  8-?   Receive Error Count (increments together with ports 48081B3h, 48081B6h)

Increments when receiving a packet. Automatically reset to zero after reading.

48081D0h..1DFh - W_CMD_STAT - Multiplay Response Error Counters (R/W)
The multiplay error counters are only used when sending a multiplay command (via W_TXBUF_CMD) to any connected slaves (which must be indicated by flags located in the second halfword of the multiplay command's frame body).

  48081D0h        Not used (always zero)
  48081D1h..1DFh  Client 1..15 Response Error (increments on missing replies)

If one or more of those slaves fail to respond, then the corresponding error counters get incremented (at the master side). Automatically reset to zero after reading.
Unknown if these counters do also increment at the slave side?

DS Wifi Transmit Control

48080ACh - W_TXREQ_RESET - Reset Transfer Request Bits (W)

  0-3   Reset corresponding bits in W_TXREQ_READ (0=No change, 1=Reset)
  4-15  Unknown (if any)

Firmware writes values 01h,02h,08h,0Dh, and FFFFh.

48080AEh - W_TXREQ_SET - Set Transfer Request Bits (W)

  0-3   Set corresponding bits in W_TXREQ_READ   (0=No change, 1=Set)
  4-15  Unknown (if any)

Firmware writes values 01h,02h,05h,08h,0Dh.

48080B0h - W_TXREQ_READ - Get Transfer Request Bits (R)

  0     Send W_TXBUF_LOC1  (1=Transfer, if enabled in W_TXBUF_LOC1.Bit15)
  1     Send W_TXBUF_CMD   (1=Transfer, if enabled in W_TXBUF_CMD.Bit15)
  2     Send W_TXBUF_LOC2  (1=Transfer, if enabled in W_TXBUF_LOC2.Bit15)
  3     Send W_TXBUF_LOC3  (1=Transfer, if enabled in W_TXBUF_LOC3.Bit15)
  4     Unknown (Beacon?)  (always 1, except when cleared via W_POWERFORCE)
  5-15  Unknown/Not used

Bit0-3 can be set/reset via W_TXREQ_SET/W_TXREQ_RESET. The setting in W_TXREQ_READ remains intact even after the transfer(s) have completed.
If more than one of the LOC1,2,3 bits is set, then LOC3 is transferred first, LOC1 last.
Beacons are transferred in every Beacon Timeslot (if enabled in W_TXBUF_BEACON.Bit15).
Bit0,2,3 are automatically reset upon IRQ14 (by hardware).

48080B6h - W_TXBUSY (R)

  0     W_TXBUF_LOC1  (1=Requested Transfer busy, or not yet started at all)
  1     W_TXBUF_CMD   (1=Requested Transfer busy, or not yet started at all)
  2     W_TXBUF_LOC2  (1=Requested Transfer busy, or not yet started at all)
  3     W_TXBUF_LOC3  (1=Requested Transfer busy, or not yet started at all)
  4     W_TXBUF_BEACON  (1=Beacon Transfer busy)
  5-15  Unknown (if any)

Busy bits. If all three W_TXBUF_LOC's are sent, then it goes through values 0Dh,05h,01h,00h; ie. LOC3 is transferred first, LOC1 last. The register is updated upon IRQ01 (by hardware).
Bit4 is set only in Beacon Timeslots.

48080B8h - W_TXSTAT - RESULT - Status of transmitted frame (R)
For LOC1-3, this register is updated at the end of a transfer (upon the IRQ01 request), if retries occur then it is updated only after the final retry.
For BEACON, this register is updated only if enabled in W_TXSTATCNT.Bit15, and only after successful transfers (since beacon errors result in infinite retries).
For CMD, this register is updated only if enabled in W_TXSTATCNT.Bit13,14).
Bit0/1 act similar to W_IF Bit1/3, however, the W_IF Bits are set after each transmit (including retries).

  0     One (or more) Packet has Completed (1=Yes)
        (No matter if successful, for that info see Bit1)
        (No matter if ALL packets are done, for that info see Bit12-13)
  1     Packet Failed (1=Error)
  2-7   Unknown/Not used
  8-11  Usually 0, ...but firmware is checking for values 03h,08h,0Bh
        (gets set to 07h when transferred W_TXBUF_LOC1/2/3 did have Bit12=set)
        (gets set to 00h otherwise)
        (gets set to 03h after beacons   ;if enabled in W_TXSTATCNT.Bit15)
        (gets set to 08h after cmd's     ;if enabled in W_TXSTATCNT.Bit14)
        (gets set to 0Bh after cmd ack's ;if enabled in W_TXSTATCNT.Bit13)
        (gets set to 04h after reply's   ;if enabled in W_TXSTATCNT.Bit12)
  12-13 Packet that updated W_TXSTAT (0=LOC1/BEACON/CMD/REPLY, 1=LOC2, 2=LOC3)
  14-15 Unknown/Not used

No idea how to reset bit0/1 once when they are set?

4808008h - W_TXSTATCNT (R/W)

  0-11  Unknown (usually zero) (otherwise disables RXing multiplay REPLY's?)
  12    Update W_TXSTAT=0401h and trigger IRQ01 after REPLY transmits   (1=Yes)
  13    Update W_TXSTAT=0B01h and trigger IRQ01 after CMD ACK transmits (1=Yes)
  14    Update W_TXSTAT=0800h and trigger IRQ01 after CMD DATA transmits(1=Yes)
  15    Update W_TXSTAT=0301h and trigger IRQ01 after BEACON transmits  (1=Yes)

Note: LOC1..3 transmits are always updating W_TXSTAT and triggering IRQ01.

4808194h - W_TX_HDR_CNT - Disable Transmit Header Adjustments (R/W)

  0     IEEE FC.Bit12 and Duration (0=Auto/whatever, 1=Manual/Wifi RAM)
  1     IEEE Frame Check Sequence  (0=Auto/FCS/CRC32, 1=Manual/Wifi RAM)
  2     IEEE Sequence Control      (0=Auto/W_TX_SEQNO, 1=Manual/Wifi RAM)
  3-15  Always zero

Allows to disable automatic adjustments of the IEEE header and checksum.
Note: W_TX_SEQNO can be also disabled by W_TXBUF_LOCn.Bit13 and by TXHDR[04h].

4808210h - W_TX_SEQNO - Transmit Sequence Number (R)

  0-11   Increments on IRQ07 (Transmit Start Interrupt)
  12-15  Always zero

Also incremented shortly after IRQ12.
When enabled in W_TXBUF_LOCn.Bit13, this value replaces the upper 12bit of the IEEE Frame Header's Sequence Control value (otherwise, when disabled, the original value in Wifi RAM is used, and, in that case, W_TX_SEQNO is NOT incremented).
Aside from W_TXBUF_LOCn.Bit13, other ways to disable W_TX_SEQNO are: Transmit Hardware Header entry TXHDR[04h], and W_TX_HDR_CNT.Bit2.

DS Wifi Transmit Buffers

4808068h - W_TXBUF_WR_ADDR - Wifi CircBuf Write Address (R/W)

  0     Always zero
  1-12  Halfword Address in RAM for Writes via W_TXBUF_WR_DATA
  13-15 Always zero

4808070h - W_TXBUF_WR_DATA - Wifi CircBuf Write Data (W)

  0-15  Data to be written to address specified in W_TXBUF_WR_ADDR

After writing to this register, W_TXBUF_WR_ADDR is automatically incremented by 2, and, if it gets equal to W_TXBUF_GAP, then it gets additonally incremented by W_TXBUF_GAPDISP*2.

4808074h - W_TXBUF_GAP - Wifi CircBuf Write Top (R/W)

  0     Always zero
  1-12  Halfword Address
  13-15 Always zero

4808076h - W_TXBUF_GAPDISP - CircBuf Write Offset from Top to Bottom (R/W)

  0-11  Halfword Offset (added to; if equal to W_TXBUF_GAP)
  12-15 Always zero

Should be "0-write_buffer_size" (wrap from end to begin), or zero (no wrapping).
Caution: On the DS-Lite, after adding it to W_TXBUF_WR_ADDR, the W_TXBUF_GAPDISP setting is destroyed (reset to 0000h) by hardware. The original DS leaves W_TXBUF_GAPDISP intact.

Note: W_TXBUF_GAP and W_TXBUF_GAPDISP may be (not TOO probably) also used by transmits via W_TXBUF_LOCn and W_TXBUF_BEACON (not tested).

4808080h - W_TXBUF_BEACON - Beacon Transmit Location (R/W)
4808090h - W_TXBUF_CMD - Multiplay Command Transmit Location (R/W)
48080A0h - W_TXBUF_LOC1 - Transmit location 1 (R/W)
48080A4h - W_TXBUF_LOC2 - Transmit location 2 (R/W)
48080A8h - W_TXBUF_LOC3 - Transmit location 3 (R/W)

  0-11  Halfword Address of TX Frame Header in RAM
  12    For LOC1-3: When set, W_TXSTAT.bit8-10 are set to 07h after transfer
                    And, when set, the transferred frame-body gets messed up?
        For BEACON: Unknown, no effect on W_TXSTAT
        For CMD: Unknown, no effect on W_TXSTAT
  13    IEEE Sequence Control (0=From W_TX_SEQNO, 1=Value in Wifi RAM)
        For BEACON: Unknown (always uses W_TX_SEQNO) (no matter of bit13)
  14    Unknown
  15    Transfer Request (1=Request/Pending)

For LOC1..3 and CMD, Bit15 is automatically cleared after (or rather: during?) transfer (no matter if the transfer was successful). For Beacons, bit15 is kept unchanged since beacons are intended to be transferred repeatedly.
W_TXBUF_CMD.Bit15 can be set ONLY while W_CMD_COUNT is non-zero.

48080B4h - W_TXBUF_RESET (W)

  0     Disable LOC1    (0=No change, 1=Reset W_TXBUF_LOC1.Bit15)
  1     Disable CMD     (0=No change, 1=Reset W_TXBUF_CMD.Bit15)
  2     Disable LOC2    (0=No change, 1=Reset W_TXBUF_LOC2.Bit15)
  3     Disable LOC3    (0=No change, 1=Reset W_TXBUF_LOC3.Bit15)
  4-5   Unknown/Not used
  6     Disable REPLY2  (0=No change, 1=Reset W_TXBUF_REPLY2.Bit15)
  7     Disable REPLY1  (0=No change, 1=Reset W_TXBUF_REPLY1.Bit15)
  8-15  Unknown/Not used

Firmware writes values FFFFh, 40h, 02h, xxxx, 09h, 01h, 02h, C0h.

4808084h - W_TXBUF_TIM - Beacon TIM Index in Frame Body (R/W)

  0-7   Location of TIM parameters within Beacon Frame Body
  8-15  Not used/zero

Usually set to 15h, that assuming that preceding Frame Body content is: Timestamp(8), BeaconInterval(2), Capability(2), SuppRatesTagLenParams(4), ChannelTagLenParam(3), TimTagLen(2); so the value points to TimParams (ie. after TimTagLen).

480806Ch - W_TXBUF_COUNT (R/W)

  0-11  Decremented on writes to W_TXBUF_WR_DATA
  12-15 Always zero

Triggers IRQ08 when it reaches zero, and does then stay at zero (without further decrementing, and without generating further IRQs).
Note: Not affected by (accidental) reads from write-only W_TXBUF_WR_DATA.

DS Wifi Transmit Errors

Automatic ACKs
Transmit errors occur on missing ACKs. The NDS hardware is automatically responding with an ACK when receiving a packet (if it has been addressed to the receipients W_MACADDR setting). And, when sending a packet, the NDS hardware is automatically checking for ACK responses.
The only exception are packets that are sent to group addresses (ie. Bit0 of the 48bit MAC address being set to "1", eg. Beacons sent to FF:FF:FF:FF:FF:FF), the receipient(s) don't need to respond to such packets, and the sender always passes okay without checking for ACKs.

480802Ch - W_TX_RETRYLIMIT (R/W)
Specifies the maximum number of retries on Transmit Errors (eg. 07h means one initial transmit, plus up to 7 retries, ie. max 8 transmits in total).

  0-7   Retry Count (usually 07h)
  8-15  Unknown     (usually 07h)

The Retry Count value is decremented on each Error (unless it is already 00h). There's no automatic reload, so W_TX_RETRYLIMIT should be reinitialized by software prior to each transmit (or, actually, there IS a reload?).
When sending multiple packets (by setting more than one bit with W_TXREQ_SET), then the first packet may eat-up all retries, leaving only a single try to the other packet(s).

48081C0h - W_TX_ERR_COUNT - TransmitErrorCount (R/W)

  0-7   TransmitErrorCount
  8-15  Always zero

Increments on Transmit Errors. Automatically reset to zero after reading.
IRQ03 triggered when W_TX_ERR_COUNT is incremented (for NON-beacons ONLY).
IRQ05 triggered when W_TX_ERR_COUNT > 7Fh (happens INCLUDING for beacons).

Error Notification
Transmit Errors can be sensed via W_TX_ERR_COUNT, IRQ03, IRQ05, TX Hardware Header entry [00h], and W_TXSTAT.Bit1.

W_TXBUF_BEACON Errors
As the name says, W_TXBUF_BEACON is intended for sending Beacons to group addresses (which do not require to respond by ACKs). So, transmit errors would occur only when mis-using W_TXBUF_BEACON to send packets to individual addresses, but the W_TXBUF_BEACON error handling isn't fully implemented:
First of, W_TX_RETRYLIMIT isn't used, instead, W_TXBUF_BEACON errors will result in infinite retries.
Moreover, W_TXBUF_BEACON errors seem to increment W_TX_ERR_COUNT, but without generating IRQ03, however, IRQ05 is generated when W_TX_ERR_COUNT>7Fh.

Other Errors
The NDS transmit hardware seems to do little error checking on the packet headers. The only known error-checked part is byte [04h] in the TX hardware header (which must be 00h, 01h, or 02h). Aside from that, when sent to a group address, it is passing okay even with invalid IEEE type/subtypes, and even with Length/Rate entries set to zero. However, when sending such data to an individual address, the receiving NDS won't respond by ACKs.

Note
Received ACKs aren't stored in WifiRAM (or, possibly, they ARE stored, but without advancing W_RXBUF_WRCSR, so that the software won't see them, and so that they will be overwritten by the next packet).

DS Wifi Status

480819Ch - W_RF_PINS - Status of RF-Chip Control Signals (R)

  0    Reportedly "carrier sense" (maybe 1 during RX.DTA?) (usually 0)
  1    TX.MAIN (RFU.Pin17) Transmit Data Phase          (0=No, 1=Active)
  2    Unknown (RFU.Pin3)  Seems to be always high      (Always 1=high?)
  3-5  Not used                                         (Always zero)
  6    TX.ON   (RFU.Pin14) Transmit Preamble+Data Phase (0=No, 1=Active)
         Uhhh, no that seems to be still wrong...
         Bit6 is often set, even when not transmitting anything...
  7    RX.ON   (RFU.Pin15) Receive Mode                 (0=No, 1=Enable)
  8-15 Not used                                         (Always zero)

Physical state of the RFU board's RX/TX pins. Similar to W_RF_STATUS.
XXX better expl. for bit0-1:

  0    RX.BUSY Receiving Preamble or Data   (0=Idle or TX Busy, 1=RX Busy)
  1    Data Phase (for both RX/TX mode)     (0=Idle or Preamble, 1=Data)

4808214h - W_RF_STATUS - Current Transmit/Receive State (R)

  0-3  Current Transmit/Receive State:
        0 = Initial value on Power-up (before raising W_MODE_RST.Bit0)
        1 = RX Mode enabled (waiting for incoming data)
        2 = Switching from RX to TX (takes a few clock cycles)
        3 = TX Mode active  (sending preamble and data)
        4 = Switching from TX to RX (takes a few clock cycles)
        5 = Multiplay: CMD was sent, waiting for replies (RF_PINS=0084h) (uh?)
            Or rather: CMD was received, preparing reply? (on slave side!)
        6 = RX (processing incoming data?)
        7 = Switching from RX/REPLY to TX/ACK (between STAT=5 and STAT=8)
        8 = Multiplay: Sending REPLY, or CMD-Ack (RF_PINS=0046h)
        9 = Idle (upon IRQ13, and upon raising W_MODE_RST.Bit0)
  4-15 Always zero?

Numeric Status Code. Similar to W_RF_PINS.

4808268h - W_RXTX_ADDR - Current Receive/Transmit Address (R)

  0-11   Halfword address
  12-15  Always zero

Indicates the halfword that is currently transmitted or received. Can be used by Start Receive IRQ06 handler to determine how many halfwords of the packet have been already received (allowing to pre-examine portions of the packet header even when the whole packet isn't fully received). Can be also used in Transmit Start IRQ07 handler to determine which packet is currently transmitted.

DS Wifi Timers

48080E8h - W_US_COUNTCNT - Microsecond counter enable (R/W)

  0     Counter Enable (0=Disable, 1=Enable)
  1-15  Always zero

Activates W_US_COUNT, and also W_BEACON_COUNT and W_POST_BEACON (which are decremented when lower 10bit of W_US_COUNT wrap from 3FFh to 000h). Note: W_POWER_US must be enabled, too.

48080F8h - W_US_COUNT0 - Microsecond counter, bits 0-15 (R/W)
48080FAh - W_US_COUNT1 - Microsecond counter, bits 16-31 (R/W)
48080FCh - W_US_COUNT2 - Microsecond counter, bits 32-47 (R/W)
48080FEh - W_US_COUNT3 - Microsecond counter, bits 48-63 (R/W)

  0-63  Counter Value in microseconds (incrementing)

Clocked by the 22.00MHz oscillator on the RFU board (ie. not by the 33.51MHz system clock). The 22.00MHz are divided by a 22-step prescaler.

48080EAh - W_US_COMPARECNT - Microsecond compare enable (R/W)

  0     Compare Enable (0=Disable, 1=Enable) (IRQ14/IRQ15)
  1     Force IRQ14    (0=No, 1=Force Now)   (Write-only)
  2-15  Always zero

Activates IRQ14 on W_US_COMPARE matches, and IRQ14/IRQ15 on W_BEACON_COUNT.

48080F0h - W_US_COMPARE0 - Microsecond compare, bits 0-15 (R/W)
48080F2h - W_US_COMPARE1 - Microsecond compare, bits 16-31 (R/W)
48080F4h - W_US_COMPARE2 - Microsecond compare, bits 32-47 (R/W)
48080F6h - W_US_COMPARE3 - Microsecond compare, bits 48-63 (R/W)

  0     Block Beacon IRQ14 until W_US_COUNT=W_US_COMPARE (0=No, 1=Block)  (W)
  1-9   Always zero
  10-63 Compare Value in milliseconds (aka microseconds/1024)             (R/W)

Triggers IRQ14 (see IRQ14 notes below) when W_US_COMPARE matches W_US_COUNT.
Usually set to FFFFFFFFFFFFFC00h (ie. almost/practically never). Instead, IRQ14 is usually derived via W_BEACON_COUNT.

480811Ch - W_BEACON_COUNT - Beacon Counter (R/W)
Triggers IRQ14 and IRQ15 (see IRQ14/IRQ15 notes below) when it reaches 0000h (if W_PRE_BEACON is non-zero, then IRQ15 occurs that many microseconds in advance).

  0-15  Decrementing Millisecond Counter (reloaded with W_BEACONINT upon IRQ14)

Set to W_BEACONINT upon IRQ14 events (unlike the other W_US_COMPARE related actions, this is done always, even if W_US_COMPARECNT is zero).
When reaching 0000h, it is immediately reloaded (as for US_COUNT matches), so the counting sequence is ..,3,2,1,BEACONINT,.. (not 3,2,1,ZERO,BEACONINT).

4808134h - W_POST_BEACON - Post-Beacon Counter (R/W)

  0-15  Decrementing Millisecond Counter (reloaded with FFFFh upon IRQ14)

Triggers IRQ13 when it reaches 0000h (no matter of W_US_COMPARECNT), and does then stay fixed at 0000h (without any further decrement/wrapping to FFFFh).
Set to FFFFh upon IRQ14 (by hardware), the IRQ14 handler should then adjust the register (by software) by adding the Tag=DDh Beacon header's Stepping value (usually 000Ah) to it.
Seems to be used to indicate beacon transmission time (possibly including additional time being reserved for responses)?

480808Ch - W_BEACONINT - Beacon Interval (R/W)
Reload value for W_BEACON_COUNT.

  0-9   Frequency in milliseconds of beacon transmission
  10-15 Always zero

Should be initialized randomly to 0CEh..0DEh or so. The random setting reduces risk of repeated overlaps with beacons from other hosts.

4808110h - W_PRE_BEACON - Pre-Beacon Time (R/W)

  0-15  Pre-Beacon Time in microseconds (static value, ie. NOT decrementing)

Allows to define the distance between IRQ15 and IRQ14. The setting doesn't affect the IRQ14 timing (which occurs at the W_BEACON_COUNT'th millisecond boundary), but IRQ15 occurs in advance (at the W_BEACON_COUNT'th millisecond boundary minus W_PRE_BEACON microseconds). If W_PRE_BEACON is zero, then both IRQ14 and IRQ15 occur exactly at the same time.

4808088h - W_LISTENCOUNT - Listen Count (R/W)

  0-7   Decremented by hardware at IRQ14 events (ie. once every beacon)
  8-15  Always zero

Reload occurs immediately BEFORE decrement, ie. with W_LISTENINT=04h, it will go through values 03h,02h,01h,00h,03h,02h,01h,00h,etc.

480808Eh - W_LISTENINT - Listen Interval (R/W)

  0-7   Listen Interval, counted in beacons (usually 02h)
  8-15  Always zero

Reload value for W_LISTENCOUNT.

480810Ch - W_CONTENTFREE (R/W)

  0-15  Decrementing microsecond counter

Operated always (no matter of W_US_COUNTCNT).
Once when it has reached 0000h, it seems to stay fixed at 0000h.
"[Set to the remaining duration of contention-free period when
receiving beacons - only *really* necessary for powersaving mode]"

IRQ13 Notes (Post-Beacon Interrupt)
IRQ13 is generated by W_POST_BEACON. It's simply doing:

  W_IF.Bit13=1      ;interrupt request

If W_POWER_TX.Bit1=0, then additionally enter sleep mode:

  [4808034h]=0002h ;W_INTERNAL   ;(similar to W_POWERFORCE=8001h)
  [480803Ch]=02xxh ;W_POWERSTATE ;(W_TXREQ_READ.Bit4 is kept intact though)
  [480819Ch]=0046h ;W_RF_PINS.7=0;disable receive (enter idle mode) (RX.ON=Low)
  [4808214h]=0009h ;W_RF_STATUS=9;indicate idle mode

Unlike for IRQ14/IRQ15, that's done no matter of W_US_COMPARECNT.

IRQ14 Notes (Beacon Interrupt)
IRQ14 is generated by W_US_COMPARE, and by W_BEACON_COUNT.
Aside from just setting the IRQ flag in W_IF, the hardware does:

  W_BEACON_COUNT=W_BEACONINT                             ;next IRQ15/IRQ14
  (Above is NOT done when IRQ14 was forced via W_US_COMPARECNT.Bit1)

If W_US_COMPARECNT is 1, then the hardware does additionally:

  (Below IS ALSO DONE when IRQ14 was forced via W_US_COMPARECNT.Bit1)
  W_IF.Bit14=1
  W_POST_BEACON=FFFFh ;about 64 secs (ie. almost never) ;next IRQ13 ("never")
  W_TXREQ_READ=W_TXREQ_READ AND FFF2h
  if W_TXBUF_BEACON.15 then W_TXBUSY.Bit4=1
  if W_LISTENCOUNT=00h then W_LISTENCOUNT=W_LISTENINT
  W_LISTENCOUNT=W_LISTENCOUNT-1

If W_TXBUF_BEACON.Bit15=1, then following is done shortly after IRQ14:

  W_RF_PINS.Bit7=0  ;disable receive (RX.ON=Low)
  W_RF_STATUS=2     ;indicate switching from RX to TX mode

If W_TXBUF_BEACON.Bit15=1, then following is done a bit later:

  W_RF_PINS.Bit6=1  ;transmit preamble start (TX.ON=High)
  W_RF_STATUS=3     ;indicate TX mode

The IRQ14 handler should then do the following (by software):

  W_POST_BEACON = W_POST_BEACON + TagDDhSteppingValue  ;next IRQ13

For using only ONE of the two IRQ14 sources: W_BEACON_COUNT can be disabled by setting both W_BEACON_COUNT and W_BEACONINT to zero. W_US_COMPARE can be sorts of "disabled" by setting it to value distant from W_US_COUNT, such like compare=count-400h.

IRQ07 Notes (Transmit Start Data; occurs after preamble)

  W_IF.Bit7=1       ;interrupt request
  W_RF_PINS.Bit1=1  ;start data transfer (preamble finished now) (TX.MAIN=High)

Below only if packet was sent through W_TXBUF_BEACON, or if it was sent via W_TXBUF_LOCn, with W_TXBUF_LOCn.Bit13 being zero:

  [TXBUF...] = W_TX_SEQNO*10h   ;auto-adjust IEEE Sequence Control
  W_TX_SEQNO=W_TX_SEQNO+1       ;increase sequence number

IRQ01 Notes (Transmit Done)
The following happens shortly before IRQ01:

  W_RF_PINS.Bit6=0  ;disable TX (TX.ON=Low)
  W_RF_STATUS=4     ;indicate switching from TX to RX mode

Then, upon IRQ01, the following happens:

  W_IF.Bit1=1       ;interrupt request
  W_RF_PINS.Bit1=0  ;disable TX (TX.MAIN=Low)
  W_RF_PINS.Bit7=1  ;enable RX (RX.ON=High)
  W_RF_STATUS=1     ;indicate RX mode

IRQ15 Notes (Pre-Beacon Interrupt)
IRQ15 is generated via W_BEACON_COUNT and W_PRE_BEACON. It's simply doing:

  if W_US_COMPARECNT=1 then W_IF.Bit15=1

If W_POWER_TX.Bit0=1, then additionally wakeup from sleep mode:

  W_RF_PINS.Bit7=1  ;enable RX (RX.ON=High) ;\gets set like so a good while
  W_RF_STATUS=1     ;indicate RX mode       ;/after IRQ15 (but not immediately)

Beacon IRQ Sequence

  IRQ15  Pre-Beacon  (beacon will be transferred soon)
  IRQ14  Beacon      (beacon will be transferred very soon) (carrier starts)
  IRQ07  Tx Start    (beacon transfer starts) (if enabled in W_TXBUF_BEACON.15)
  IRQ01  Tx End      (beacon transfer done) (if enabled in W_TXSTATCNT.15)
  IRQ13  Post-Beacon (beacon transferred) (unless next IRQ14 occurs earlier)

That, for tranmitting beacons. (For receiving, IRQ07/IRQ01 would be replaced by Rx IRQ's, provided that a remote unit is sending beacons).

DS Wifi Multiplay Master

These registers are used for multiplay host-to-client (aka master to slave) commands.

48080EEh - W_CMD_COUNTCNT (R/W)

  0     Enable W_CMD_COUNT (0=Disable, 1=Enable)
  1-15  Always Zero

4808118h - W_CMD_COUNT (R/W)

  0-15  Decremented once every 10 microseconds (Stopped at 0000h)

Defines the time interval during which W_TXBUF_CMD transfers are possible. That CMD will be automatically transferred again if there are errors (missing REPLY's), but it will abort if there isn't enough time left in W_CMD_COUNT.
IRQ12 is thrown when cmd/reply were successful, or when CMD_COUNT reached zero (failed).
Firmware Beacon IRQ14 handler checks for W_CMD_COUNT<=0Ah.

48080C0h - W_CMD_TOTALTIME - (R/W)

  0-15  Duration per ALL slave response packet(s) in microseconds

Before sending a MASTER packet, this port should be set to the same value as the MASTER packet's IEEE header's Duration/ID entry.

48080C4h - W_CMD_REPLYTIME - (R/W)

  0-15  Duration per SINGLE slave response packet in microseconds

Before sending a MASTER packet, this port should be set to the expected per slave response time.
Note: Nintendo's multiboot/pictochat code is also putting this value in the 1st halfword of the MASTER packet's frame body.

At 2MBit/s transfer rate, the values should be set up sorts of like so:

  master_time = (master_bytes*4)+(60h)     ;60h = 96 decimal = short preamble
  slave_time = (slave_bytes*4)+(0D0h..0D2h)
  all_slave_time = (EAh..F0h)+(slave_time+0Ah)*num_slaves
  txhdr[2]   = slave_bits      ;hardware header (*)
  ieee[2]    = all_slave_time  ;ieee header (duration/id)
  body[0]    = slave_time      ;duration per slave (for multiboot/pictochat)
  body[2]    = slave_bits      ;frame body -- required (*)
  [48080C0h] = all_slave_time  ;W_CMD_TOTALTIME
  [48080C4h] = slave_time      ;W_CMD_REPLYTIME duration per slave
  [4808118h] = (388h+(num_slaves*slave_time)+master_time+32h)/10  ;W_CMD_COUNT
  [4808090h] = 8000h+master_packet_address   ;start transmit      ;W_TXBUF_CMD

With the byte values counting the ieee frame header+body+fcs.
(*) The hardware doesn't actually seem to use the "slave_bits" entry in the hardware header, instead, it is using the "slave_bits" entry in the frame body(!)

Flowchart (at Master side)

 After starting transfer via TXREQ and TXBUF_CMD write:
   TXBUSY=2 (formerly 0) (after TXBUF_CMD write, or sometimes a bit later)
 After about 50-500 microseconds:                               ;\
   RF_STAT=3 (TXing) (formerly 2)                               ;
   RXTX_ADDR=0006h..0008h (TXbuf+0Ch..) (formerly in RXBUF)     ; CMD
   SEQNO+1                                                      ;
 After TX preamble:                                             ;
   IF=80h (TX Start, for CMD)                                   ;
   RXTX_ADDR=0009h..0xxxh (TXbuf..)                             ;
 After TX data:                                                 ;
   optional: IF=02h (TX Done, for CMD) (if enabled in TXSTATCNT);
   optional: TXSTAT=0800h (CMD done)   (if enabled in TXSTATCNT);
   RF_STAT=5 (CMD done, prepare for REPLY)                      ;/
 US=0017h                                                       ;\
   RXTX_ADDR=rxbuf..                                            ;
 After RX preamble:                                             ;
   IF=40h (RX Start, for REPLY)                                 ; REPLY
   RXTX_ADDR=rxbuf..                                            ; (if any)
 After RX data:                                                 ;
   IF=01h (RX Done, for REPLY)                                  ;
   WRCSR+18h (for REPLY)                                        ;/
 After a dozen microseconds:                                    ;\
   RF_STAT=7 ;Switching from REPLY to ACK                       ;
   RF_STAT=8 ;TXing ACK  (shortly after above STAT=7)           ;
   RXTX_ADDR=0FC0h (special dummy addr during TX ACK)           ;
 After TX preamble:                                             ; ACK
   IF=80h (TX Start, for ACK)                                   ;
 After TX data:                                                 ;
   optional: IF=02h (TX Done, for ACK) (if enabled in TXSTATCNT);
   optional: TXSTAT=0B01h (ACK done)   (if enabled in TXSTATCNT);
   TXBUSY=0000h (formerly 0002h)                                ;
   TXBUF_CMD.bit15=0                                            ;
   TXHDR_0=0001h (okay)           (formerly 0000h)              ;
   TXHDR_2=0000h (no error flags) (formerly 0002h)              ;
   SEQNO+1                                                      ;
   RF_STAT=1   ;RX awaiting                                     ;
   IF=1000h (CMD timeslot done) (shortly AFTER above IF=02h)    ;/

Sending a W_TXBUF_CMD frame starts the following process:

 1. MP host sends the CMD frame, as soon as possible. after preamble,
     IRQ7 is triggered
 2. once the transfer is finished: if bit14 in W_TXSTATCNT is set,
     W_TXSTAT is set to 0x0800, and IRQ1 is triggered
     somewhere here: set W_RF_STATUS=5, RFPINS=0x0084
 3. hardware waits for MP clients' replies, duration is:
     16 + ((10 + W_CMD_REPLYTIME) * count_ones(client_mask_from_frame_body))
 4. MP host sends the CMD ack. after preamble, IRQ7 is triggered
     (this is why you get two IRQ7's from a CMD transfer)
 5. during the ack transfer, W_RF_STATUS is 8, and W_RXTXADDR is 0x0FC0
 6. once the transfer is finished: if bit13 in W_TXSTATCNT is set,
     W_TXSTAT is set to 0x0B01, and IRQ1 is triggered.
 7. the TX header of the CMD frame is adjusted: bits in TXheader[02] are
     cleared to indicate that the corresponding clients responded
     successfully. Nintendo software checks this.

I haven't looked a lot into how retries work. they seem to repeat the entire process.
the CMD ack is sent automatically. the packet is what is described here: the 03:09:BF:00:00:03 flow. MP clients will receive it like a regular frame.
IRQ12 is used to signal the end of the CMD transfer process. it will be either when the final ack is done transferring, if everything was successful, or otherwise at the end of W_CMD_COUNT.

DS Wifi Multiplay Slave

The slave mode is activated by setting W_AID_LOW to a nonzero value (the slave number in range 1..15). That settings specifies that (and when) to send replies to incoming commands.

The replies are queued in W_TXBUF_REPLY1 (next packet) and W_TXBUF_REPLY2 (readonly, current packet).
Software does usually write to the REPLY1 register. Upon incomings CMD's, the hardware does automatically forward REPLY1 to REPLY2 (and reset REPLY1 to 0000h).
Because of that queuing, one should use two alternating Wifi RAM locations (ie. don't overwrite the current packet when preparing the next packet).
Alternately, one can manually forward REPLY1 to REPLY2 (via W_RXCNT.bit7, but that's normally not needed, except maybe for things like manually resetting REPLY2).

These registers are used for multiplay client-to-host (aka slave to master) responses.

4808094h - W_TXBUF_REPLY1 - Multiplay Next Reply Transmit Location (R/W)
4808098h - W_TXBUF_REPLY2 - Multiplay Current Reply Transmit Location (R)

  0-11  Halfword address
  12-14 Unknown (the bits can be set, ie. they DO exist)
  15    Enable

Note: W_TXBUF_REPLY2.Bit15 is NOT cleared after SENDING the reply (instead, REPLY2 gets overwritten by REPLY1 when RECEIVING the next CMD).

Flowchart (at Slave side)

  At incoming CMD DATA packet:                                          ;\
    RF_STATUS=6  ;RX processing incoming stuff                          ;
  After RX preamble:                                                    ; CMD
    IRQ6 (RX Start, for CMD DATA)                                       ; DATA
  After RX data:                                                        ;
    IRQ0 (RX Done, for CMD DATA)                                        ;
    WRCSR=WRCSR+(size of CMD DATA)                                      ;
    RF_STATUS=5  ;preparing REPLY                                       ;
    if REPLY2.bit15=1                                                   ;
      TXHDR[1]=TXHDR[0]  ;<-- or sometimes random?  ;\adjust TXHDR[0,1] ;
      TXHDR[0]=01h       ;<-- mark done/discarded   ;/for <old> REPLY2  ;
    REPLY2=REPLY1, REPLY1=0000h                     ;-forward new reply ;
    if REPLY2.bit15=1                                                   ;
      TXHDR[4] incremented (unless already max FFh) ;\adjust TXHDR[4,5] ;
      TXHDR[5]=00h                                  ;/for <new> REPLY2  ;
      TX_SEQNO incremented  ;<-- done here if REPLY2 exists             ;/
  After some moment (at the AID_LOW slot?):                             ;\
    RF_STATUS=8  ;TX sending REPLY                                      ;
  After TX preamble:                                                    ; REPLY
    IRQ7 (TX Start, for REPLY)                                          ;
  After TX data:                                                        ;
    RF_STATUS=1  ;RX awaiting next packet                               ;
    optional: IRQ1 (TX Done) (only if enabled in TXSTATCNT, and REPLY2.bit15=1)
    optional: TXSTAT=0401h   (only if enabled in TXSTATCNT)             ;
    if REPLY2.bit15=0                                                   ;
      SEQNO increased      ;<-- done here when REPLY2 is empty          ;/
  After some moment:                                                    ;\
    RF_STATUS=6  ;RX processing incoming stuff                          ;
  After RX preamble:                                                    ; CMD
    IRQ6 (RX Start, for CMD ACK)                                        ; ACK
  After RX data:                                                        ;
    IRQ0 (RX Done, for CMD ACK)                                         ;
    WRCSR=WRCSR+(size of CMD ACK)                                       ;
    RF_STATUS=1  ;RX awaiting next packet                               ;/
  Thereafter, Nintendo's software seems to require a delay (at least
  100h microseconds) before receiving the next CMD DATA packet.

on client side
The reply transfer is automatically initiated when receiving a MP CMD frame (this seems to be based on the incoming IEEE header's frame control value and/or broadcast address, CMD frames typically have FC=0228h).

When receiving the MP CMD frame, the hardware determines its position using its W_AID_LOW register and the frame's client bitmask, then waits for its turn to reply.

W_TXBUF_REPLY1 is latched into W_TXBUF_REPLY2, and reset to zero.
The byte at TXHDR[04h] is set incremented. Nintendo software checks for this.

when transferring the MP reply itself:
IRQ7 is triggered. W_RF_STATUS is 8.
at the end: if W_TXSTATCNT.bit12=1, W_TXSTAT is set to 0401h, and IRQ1 is triggered.

Reportedly...
It will always send a reply - if W_TXBUF_REPLY1 isn't configured, it will send an empty reply frame (with frame control 0x0158). W_RXTXADDR isn't modified when sending a default empty reply.

DS Wifi Configuration Ports

4808120h - W_CONFIG_120h (R/W) ;init from firmware[04Ch] ;81ff 0048->SAME
4808122h - W_CONFIG_122h (R/W) ;init from firmware[04Eh] ;ffff 4840->SAME
4808124h - W_CONFIG_124h (R/W) ;init from firmware[05Eh] ;ffff 0000->0032
4808128h - W_CONFIG_128h (R/W) ;init from firmware[060h] ;ffff 0000->01F4
4808130h - W_CONFIG_130h (R/W) ;init from firmware[054h] ;0fff 0142->0140
4808132h - W_CONFIG_132h (R/W) ;init from firmware[056h] ;8fff 8064->SAME
4808140h - W_CONFIG_140h (R/W) ;init from firmware[058h] ;ffff 0000->E0E0
4808142h - W_CONFIG_142h (R/W) ;init from firmware[05Ah] ;ffff 2443->SAME
4808144h - W_CONFIG_144h (R/W) ;init from firmware[052h] ;00ff 0042->SAME
4808146h - W_CONFIG_146h (R/W) ;init from firmware[044h] ;00ff 0016->0002
4808148h - W_CONFIG_148h (R/W) ;init from firmware[046h] ;00ff 0016->0017
480814Ah - W_CONFIG_14Ah (R/W) ;init from firmware[048h] ;00ff 0016->0026
480814Ch - W_CONFIG_14Ch (R/W) ;init from firmware[04Ah] ;ffff 162C->1818
4808150h - W_CONFIG_150h (R/W) ;init from firmware[062h] ;ff3f 0204->0101
4808154h - W_CONFIG_154h (R/W) ;init from firmware[050h] ;7a7f 0058->SAME
These ports are to be initialized from firmware settings.
Above comments show the R/W bits (eg. 81FFh means bit15 and bit8-0 are R/W, bit14-9 are always zero), followed by the initial value on Reset (eg. 0048h), followed by new value after initialization from firmware settings (eg. 0032h, or SAME if the Firmware value is equal to the Reset value; these values seem to be identical in all currently existing consoles).
Note: Firmware part4 changes W_CONFIG_124h to C8h, and W_CONFIG_128h to 7D0h, and W_CONFIG_150h to 202h, and W_CONFIG_140h depending on tx rate and preamble:

  W_CONFIG_140h = firmware[058h]+0202h             ;1Mbit/s
  W_CONFIG_140h = firmware[058h]+0202h-6161h       ;2Mbit/s with long preamble
  W_CONFIG_140h = firmware[058h]+0202h-6161h-6060h ;2Mbit/s with short preamble

48080ECh - W_CONFIG_0ECh (R/W) ;firmware writes 3F03h (same as on power-up)
48080D4h - W_CONFIG_0D4h (R/W) ;firmware writes 0003h (affectd by W_MODE_RST)
48080D8h - W_CONFIG_0D8h (R/W) ;firmware writes 0004h (same as on power-up)
4808254h - W_CONFIG_254h (?) ;firmware writes 0000h (read: EEEEh on DS-Lite)
Firmware just initializes these ports with fixed values, without further using them after initialization.

48080DAh - W_RX_LEN_CROP (R/W) ;firmware writes 0602h (same as on power-up)

  0-7   Decrease RX Length by N halfwords for Non-WEP packets (usually 2)
  8-15  Decrease RX Length by N halfwords for WEP packets     (usually 6)

Used to exclude the FCS (and WEP IV+ICV) from the length entry in the Hardware RX Header.

DS Wifi Baseband Chip (BB)

BB-Chip Mitsumi MM3155 (DS), or BB/RF-Chip Mitsumi MM3218 (DS-Lite)

4808158h - W_BB_CNT - Baseband serial transfer control (W)

  0-7   Index     (00h-68h)
  8-11  Not used  (should be zero)
  12-15 Direction (5=Write BB_WRITE to Chip, 6=Read from Chip to BB_READ)

Transfer is started after writing to this register.

480815Ah - W_BB_WRITE - Baseband serial write data (W)

  0-7   Data to be sent to chip (by following W_BB_CNT transfer)
  8-15  Not used (should be zero)

480815Ch - W_BB_READ - Baseband serial read data (R)

  0-7   Data received from chip (from previous W_BB_CNT transfer)
  8-15  Not used (always zero)

480815Eh - W_BB_BUSY - Baseband serial busy flag (R)

  0     Transfer Busy (0=Ready, 1=Busy)
  1-15  Always zero

Used to sense transfer completion after writes to W_BB_CNT.
Not sure if I am doing something wrong... but the busy flag doesn't seem to get set immediately after W_BB_CNT writes, and works only after waiting a good number of clock cycles?

4808160h - W_BB_MODE (R/W)

  0-7   Always zero
  8     Unknown (usually 1) (no effect no matter what setting?)
  9-13  Always zero
  14    Unknown (usually 0) (W_BB_READ gets unstable when set)
  15    Always zero

This register is initialized by firmware bootcode - don't change.

4808168h - W_BB_POWER (R/W)

  0-3   Disable whatever   (usually 0Dh=disable)
  4-14  Always zero
  15    Disable W_BB_ports (usually 1=Disable)

Must be set to 0000h before accessing BB registers.

Read-Write-Ability of the BB-Chip Mitsumi MM3155 registers (DS)

  Index    Num Dir Expl.
  00h        1 R   always 6Dh (R) (Chip ID)
  01h..0Ch  12 R/W 8bit R/W
  0Dh..12h   6 -   always 00h
  13h..15h   3 R/W 8bit R/W
  16h..1Ah   5 -   always 00h
  1Bh..26h  12 R/W 8bit R/W
  27h        1 -   always 00h
  28h..4Ch     R/W 8bit R/W
  4Dh        1 R   always 00h or BFh (depending on other regs)
  4Eh..5Ch     R/W 8bit R/W
  5Dh        1 R   always 01h (R)
  5Eh..61h     -   always 00h
  62h..63h   2 R/W 8bit R/W
  64h        1 R   always FFh or 3Fh (depending on other regs)
  65h        1 R/W 8bit R/W
  66h        1 -   always 00h
  67h..68h   2 R/W 8bit R/W
  69h..FFh     -   always 00h

Read-Write-Ability of the BB/RF-Chip Mitsumi MM3218 (DS-Lite)
Same as above. Except that reading always seems to return [5Dh]=00h. And, for whatever reason, Nintendo initializes DS-Lite registers by writing [00h]=03h and [66h]=12h. Nethertheless, the registers always read as [00h]=6Dh and [66h]=00h, ie. same as on original DS.

Important BB Registers
Registers 0..68h are initialized by firmware bootcode, and (most) of these settings do not need to be changed by other programs, except for:

  Addr Initial Meaning
  01h 0x9E    [unsetting/resetting bit 7 initializes/resets the system?]
  02h         unknown (firmware is messing with this register)
  06h         unknown (firmware is messing with this register, too)
  13h 0x00    CCA operation - criteria for receiving
                    0=only use Carrier Sense (CS)
                    1=only use Energy Detection (ED)
                    2=receive if CS OR ED
                    3=receive only if CS AND ED
  1Eh 0xBB    see change channels flowchart (Ext. Gain when RF[09h].bit16=0)
  35h 0x1F    Energy Detection (ED) criteria
              value 0..61 (representing energy levels of -60dBm to -80dBm)

DS Wifi RF Chip

RF-Chip RF9008 (compatible to RF2958 from RF Micro Devices, Inc.) (Original DS)
BB/RF-Chip Mitsumi MM3218 (DS-Lite)

480817Ch - W_RF_DATA2 - RF chip serial data/transfer enable (R/W)
For Type2 (ie. firmware[040h]<>3):

  0-1   Upper 2bit of 18bit data
  2-6   Index   (00h..1Fh) (firmware uses only 00h..0Bh)
  7     Command (0=Write data, 1=Read data)
  8-15  Should be zero (not used with 24bit transfer)

For Type3 (ie. firmware[040h]=3):

  0-3   Command (5=Write data, 6=Read data)
  4-15  Should be zero (not used with 20bit transfer)

Writing to this register starts the transfer.

480817Eh - W_RF_DATA1 - RF chip serial data (R/W)
For Type2 (ie. firmware[040h]<>3):

  0-15  Lower 16bit of 18bit data

For Type3 (ie. firmware[040h]=3):

  0-7   Data (to be written to chip) (or being received from chip)
  8-15  Index (usually 00h..28h) (index 40h..FFh are mirrors of 00h..3Fh)

This value should be set before setting W_RF_DATA2.

4808180h - W_RF_BUSY - RF chip serial busy flag (R)

  0     Transfer Busy (0=Ready, 1=Busy)
  1-15  Always zero

Used to sense transfer completion after writes to W_RF_DATA2.

4808184h - W_RF_CNT - RF chip serial control (R/W)

  0-5   Transfer length (init from firmware[041h].Bit0-5)
  6-7   Always zero
  8     Unknown         (init from firmware[041h].Bit7)
  9-13  Always zero
  14    Unknown         (usually 0)
  15    Always zero

This register is initialized by firmware bootcode - don't change.
Usually, Type2 has length=24bit and flag=0. Type3 uses length=20bit and flag=1.

Caution For Type2 (ie. firmware[040h]<>3)
Before accessing Type2 RF Registers, first BB[01h] must have been properly initialized (ie. BB[01h].Bit7 must have been toggled from 0-to-1).

DS Wifi RF9008 Registers

RF9008 (RF2958 compatible)
2.4GHz Spread-Spectrum Transceiver - RF Micro Devices, Inc.

RF chip data (Type2) (initial NDS settings from firmware, example)

  Firmware   Index   Data
  (24bit)    (4bit)  (18bit)
  00C007h  =  00h  + 0C007h ;-also set to 0C008h for power-down
  129C03h  =  04h  + 29C03h
  141728h  =  05h  + 01728h ;\these are also written when changing channels
  1AE8BAh  =  06h  + 2E8BAh ;/
  1D456Fh  =  07h  + 1456Fh
  23FFFAh  =  08h  + 3FFFAh
  241D30h  =  09h  + 01D30h ;-bit10..14 should be also changed per channel?
  """"50h  =  """  + """50h ;firmware v5 and up uses narrower tx filter
  280001h  =  0Ah  + 00001h
  2C0000h  =  0Bh  + 00000h
  069C03h  =  01h  + 29C03h
  080022h  =  02h  + 00022h
  0DFF6Fh  =  03h  + 1FF6Fh

RF[00h] - Configuration Register 1 (CFG1) (Power on: 00007h)

  17-16 Reserved, program to zero (0)
  15-14 Reference Divider Value (0=Div2, 1=Div3, 2=Div44, 3=Div1)
  3     Sleep Mode Current      (0=Normal, 1=Very Low)
  2     RF VCO Regulator Enable (0=Disable, 1=Enable)
  1     IF VCO Regulator Enable (0=Disable, 1=Enable)
  0     IF VGA Regulator Enable (0=Disable, 1=Enable)

RF[01h] - IF PLL Register 1 (IFPLL1) (Power on: 09003h)

  17    IF PLL Enable                      (0=Disable, 1=Enable)
  16    IF PLL KV Calibration Enable       (0=Disable, 1=Enable)
  15    IF PLL Coarse Tuning Enable        (0=Disable, 1=Enable)
  14    IF PLL Loop Filter Select          (0=Internal, 1=External)
  13    IF PLL Charge Pump Leakage Current (0=Minimum value, 1=2*Minimum value)
  12    IF PLL Phase Detector Polarity     (0=Positive, 1=Negative)
  11    IF PLL Auto Calibration Enable     (0=Disable, 1=Enable)
  10    IF PLL Lock Detect Enable          (0=Disable, 1=Enable)
  9     IF PLL Prescaler Modulus           (0=4/5 Mode, 1=8/9 Mode)
  8-4   Reserved, program to zero (0)
  3-0   IF VCO Coarse Tuning Voltage       (N=Voltage*16/VDD)

RF[02h] - IF PLL Register 2 (IFPLL2) (Power on: 00022h)

  17-16 Reserved, program to zero (0)
  15-0  IF PLL divide-by-N value

RF[03h] - IF PLL Register 3 (IFPLL3) (Power on: 1FF78h)

  17    Reserved, program to zero (0)
  16-8  IF VCO KV Calibration, delta N value (signed)  ;DeltaF=(DN/Fr)
  7-4   IF VCO Coarse Tuning Default Value
  3-0   IF VCO KV Calibration Default Value

RF[04h] - RF PLL Register 1 (RFPLL1) (Power on: 09003h)

  17-10 Same as for RF[01h] (but for RF, not for IF)
  9     RF PLL Prescaler Modulus (0=8/9 Mode, 1=8/10 Mode)
  8-0   Same as for RF[01h] (but for RF, not for IF)

RF[05h] - RF PLL Register 2 (RFPLL2) (Power on: 01780h)

  17-6  RF PLL Divide By N Value
  5-0   RF PLL Numerator Value (Bits 23-18)

RF[06h] - RF PLL Register 3 (RFPLL3) (Power on: 00000h)

  17-0  RF PLL Numerator Value (Bits 17-0)

RF[07h] - RF PLL Register 4 (RFPLL4) (Power on: 14578h)

  17-10 Same as for RF[03h] (but for RF, not for IF) ;and, DN=(deltaF/Fr)*256

RF[08h] - Calibration Register 1 (CAL1) (Power on: 1E742h)

  17-13  VCO1 Warm-up Time  ;TVCO1=(approximate warm-up time)*(Fr/32)
  12-8   VCO1 Tuning Gain Calibration ;TLOCK1=(approximate lock time)*(Fr/128)
  7-3    VCO1 Coarse Tune Calibration Reference  ;VALUE=(average time)*(Fr/32)
  2-0    Lock Detect Resolution (0..7)

RF[09h] - TXRX Register 1 (TXRX1) (Power on: 00120h)

  17    Receiver DC Removal Loop          (0=Enable DC Removal Loop, 1=Disable)
  16    Internal Variable Gain for VGA  (0=Disable/External, 1=Enable/Internal)
  15    Internal Variable Gain Source (0=From TXVGC Bits, 1=From Power Control)
  14-10 Transmit Variable Gain Select (TXVGC)   (0..1Fh = High..low gain)
  9-7   Receive Baseband Low Pass Filter     (0=Wide Bandwidth, 7=Narrow)
  6-4   Transmit Baseband Low Pass Filter    (0=Wide Bandwidth, 7=Narrow)
  3     Mode Switch            (0=Single-ended mode, 1=Differential mode)
  2     Input Buffer Enable TX (0=Input Buffer Controlled by TXEN, 1=By BBEN)
  1     Internal Bias Enable   (0=Disable/External, 1=Enable/Internal)
  0     TX Baseband Filters Bypass        (0=Not Bypassed, 1=Bypassed)

RF[0Ah] - Power Control Register 1 (PCNT1) (Power on: 00000h)

  17-15 Select MID_BIAS Level                          (1.6V through 2.6V)
  14-9  Desired output power at antenna                (N*0.5dBm)
  8-3   Power Control loop-variation-adjustment Offset (signed, N*0.5dB)
  2-0   Desired delay for using a single TX_PE line    (N*0.5us)

RF[0Bh] - Power Control Register 2 (PCNT2) (Power on: 00000h)

  17-12 Desired MAX output power when PABIAS=MAX=2.6V (N*0.5dBm)
  11-6  Desired MAX output power when PABIAS=MID_BIAS (N*0.5dBm)
  5-0   Desired MAX output power when PABIAS=MIN=1.6V (N*0.5dBm)

RF[0Ch] - VCOT Register 1 (VCOT1) (Power on: 00000h)

  17    IF VCO Band Current Compensation (0=Disable, 1=Enable)
  16    RF VCO Band Current Compensation (0=Disable, 1=Enable)
  15-0  Reserved, program to zero (0)

RF[0Dh..1Ah] - N/A (Power on: 00000h)

  Not used.

RF[1Bh] - Test Register 1 (TEST) (Power on: 0000Fh)

  17-0  This is a test register for internal use only.

RF[1Ch..1Eh] - N/A (Power on: 00000h)

  Not used.

RF[1Fh] - Reset Register (Power on: 00001h)

  17-0  Don't care (writing any value resets the chip)

DS Wifi Unknown Registers

480800Ah - W_X_00Ah (R/W)

  0-15  Unknown (usually zero)

"[bit7 - ingore rx duplicates]" <--- that is NOT correct (no effect).
Firmware writes 0000h to it. That, done many times. So, eventually some bits in this register are automatically set by hardware in whatever situations, otherwise repeatedly writing 0000h to it would be kinda useless...?

---

Below Ports W_X_244h and W_X_228h might be related to deciding when to send multiplay replies...?

4808244h - W_X_244h (R/W) x ffff [0000] (used by firmware part4)
Unknown. Seems to be W_IF/W_IE related. Firmware sets Port 4808244h bits 6,7,12 to 1-then-0 upon IRQ06,IRQ07,IRQ12 respectively.

4808228h - W_X_228h (W) fixx [0000] (used by firmware part4) (bit3)
Unknown. Firmware writes 8-then-0 (done in IRQ06 handler, after Port 4808244h access).

---

Below Ports 48081A0h, 48081A2h, 48081A4h are somehow related to BB[02h]...

48081A0h - W_X_1A0h - (R/W) -933 [0000]

  0-1   Unknown
  2-3   Always zero
  4-5   Unknown
  6-7   Always zero
  8     Unknown
  9-10  Always zero
  11    Unknown
  12-15 Always zero

Firmware writes values 000h, 823h. Seems to be power-related. The following experimental code toggles RXTX.ON (RFU.Pin4): "x=0 / @@lop: / [48081A0h]=x / [4808036h]=0 / x=x XOR 3 / wait_by_loop(1000h) / b @@lop".
Also, writing to port 48081A0h affects ports 4808034h, 480819Ch, 480821Ch, and 48082A2h.

48081A2h - W_X_1A2h - (R/W) ---3 [0001] (used by firmware part4)

  0-1   Unknown. Firmware writes values 03h, 01h, and VAR.
  2-15  Always zero

Used in combination with Port 48081A0h, so it's probably power-related, too.

48081A4h - W_X_1A4h - (R/W) ffff [0000]
"Rate used when signal test is enabled (0x0A or 0x14 for 1 or 2 mbit)"
(Not too sure if that's correct, there is no visible relation to any rate.)
(This register seems to be R/W only on certain Port 48081A0h settings.)
Unknown. Firmware writes whatever.

---

4808220h - W_RAM_DISABLE - RAM Control (R/W)

  0-1   Disable WifiRAM      (0=Normal, other=locks memory at 4804000h-5FFFh)
  2-4   Unknown              (0=Normal, other=prevents/affects RX to ram?)
  5     Disable Special Log? (0=Normal, 1=Prevent 4805F6Eh..5F77h updates)
  6-15  Unknown              (0=Normal, other=?)

4808290h - W_X_290h - (R/W or Disabled)
Reportedly, this is the "antenna" register, which should exist on official devkits, allowing to switch between wired Ethernet, and wireless Wifi mode.

  0     Unknown (R/W) (if present)
  1-15  Not used

On normal NDS release versions, this register seems to be disabled (if it is implemented at all), and trying to read from it acts as for unused registers, ie. reads return FFFFh (or probably 0000h on NDS-lite). The NDS firmware contains code for accessing this port, even in release versions.

W_INTERNAL
All registers marked as "W_INTERNAL" aren't used by Firmware part4, and are probably unimportant, except for whatever special diagnostics purposes.

Wifi DMA
Wifi RAM can be accessed with normal "Start Immediately" DMA transfers (typically by reading through W_RXBUF_RD_DATA, so the DMA automatically wraps from END to BEGIN).
Additionally, DMA0 and DMA2 can be reportedly synchronized to "Wireless Interrupt" (rather than using "Start Immediately" timing), no idea if/how that's working though... and if it gets started on any Wifi IRQ, or only on specific IRQs...?
Possibly some of the above unknown registers, or some unknown bits in other registers, are DMA related...?
Reportedly, early firmwares did use "Wireless Interrupt" DMAs (that'd be firmware v1/v2... or, only earlier unreleased prototype versions?).

DS Wifi Unused Registers

Wifi WS0 and WS1 Regions in NDS7 I/O Space
Wifi hardware occupies two 32K slots, but most of it is filled with unused or duplicated regions. The timings (waitstates) for WS0 and WS1 are initialized in WIFIWAITCNT (by firmware).

  4800000h-4807FFFh Wifi WS0 Region (32K)  ;used for RAM at 4804000h
  4808000h-4808000h Wifi WS1 Region (32K)  ;used for registers at 4808000h
  4810000h-4FFFFFFh Not used (00h-filled)

Structure of the 32K Wifi Regions (WS0 and WS1)

  Wifi-WS0-Region    Wifi-WS1-Region    Content
  4800000h-4800FFFh  4808000h-4808FFFh  Registers
  4801000h-4801FFFh  4809000h-4809FFFh  Registers (mirror)
  4802000h-4803FFFh  480A000h-480BFFFh  Unused
  4804000h-4805FFFh  480C000h-480DFFFh  Wifi RAM  (8K)
  4806000h-4806FFFh  480E000h-480EFFFh  Registers (mirror)
  4807000h-4807FFFh  480F000h-480FFFFh  Registers (mirror)

Wifi Registers (recommended 4808000h-4808FFFh) appear more stable in WS1?
Wifi RAM (recommended 4804000h-4805FFFh) appears more stable in WS0?

Unused Ports (Original NDS)
Aside from those ports listed in the Wifi I/O Map, all other ports in range 4808000h..4808FFFh are unused. On the original DS, reading from these ports returns FFFFh.

Unused Ports (NDS-Lite)
Reading from unused I/O ports acts as PASSIVE mirror of W_RXBUF_RD_DATA. Exceptions are: Ports 4808188h, and 48082D8h..48082E6h; which always return 0000h.

Unused Memory (Original NDS)
Unused Wifi Memory is at 2000h..3FFFh. On the original DS, reading from that region returns FFFFh.

Unused Memory (NDS-Lite)
Reading from unused memory acts as PASSIVE mirror of WifiRAM (ie. reading from it returns the value being most recently read from 4000h..5FFFh) (that not affected by indirect WifiRAM reads via W_RXBUF_RD_DATA) (and, that not affected by writes to wifi memory, including writes that do overwrite the most recent read value) (and, that only if WifiRAM is properly enabled, ie. Port 4808220h.Bits0-1 should be 0).
Moreover, certain addresses are additionally ORed with mirrored I/O Ports. That addresses are:

  2030h, 2044h, 2056h, 2080h, 2090h, 2094h, 2098h, 209Ch, 20A0h, 20A4h,
  20A8h, 20AAh, 20B0h, 20B6h, 20BAh, 21C0h, 2208h, 2210h, 2244h, 31D0h,
  31D2h, 31D4h, 31D6h, 31D8h, 31DAh, 31DCh, 31DEh.

For example, 2044h is a PASSIVE mirror of WifiRAM, ORed with an ACTIVE mirror of W_RANDOM (Port 044h). Note that some mirrors are at 2000h-2FFFh, and some at 3000h-3FFFh. The W_CMD_STAT mirrors are PASSIVE (that, in unused memory region only) (in normal port-mirror regions like 1000h-1FFF, W_CMD_STAT mirrors are ACTIVE).

Known (W) Mirrors (when reading from Write-only ports)

  Read from (W)           Mirrors to (NDS)       Or to (NDS-Lite)
  070h W_TXBUF_WR_DATA    060h W_RXBUF_RD_DATA   074h W_TXBUF_GAP
  078h W_INTERNAL         068h W_TXBUF_WR_ADDR   074h W_TXBUF_GAP
  0ACh W_TXREQ_RESET      09Ch W_INTERNAL        ? (zero)
  0AEh W_TXREQ_SET        09Ch W_INTERNAL        ? (zero)
  0B4h W_TXBUF_RESET      0B6h W_TXBUSY          ? (zero)
  158h W_BB_CNT           15Ch W_BB_READ         ? (zero)
  15Ah W_BB_WRITE         ? (zero)               ? (zero)
  178h W_INTERNAL         17Ch W_RF_DATA2        ? (zero)
  20Ch W_INTERNAL         09Ch W_INTERNAL        ? (zero)
  21Ch W_IF_SET           010h W_IF              010h