famicom party

10. Sprite Graphics

Now that we have talked about the PPU at a high level, we are ready to dive into the details of drawing sprites. By the end of this chapter, you will know how to create sprite tiles and draw them to the screen.

Sprite Data

Internally, the PPU uses 256 bytes of memory to store sprite information. It takes four bytes of data to describe a sprite, which is where the limit of 64 sprites at a time comes from. Those four bytes of data encode the following information:

  1. Y position of the top left corner of the sprite (0-255)
  2. Tile number from the sprite pattern table (0-255)
  3. Special attribute flags (horizontal/vertical flipping, palette number, etc.)
  4. X position of the top left corner of the sprite (0-255)

The third byte (attribute flags) uses its eight bits to store multiple pieces of information in a compact format. The eight bits control the following properties of a sprite:

Sprite Attribute Flags
Bit # Purpose
7 Flips sprite vertically (if "1")
6 Flips sprite horizontally (if "1")
5 Sprite priority (behind background if "1")
4-2 Not used
1-0 Palette for sprite

Object Attribute Memory (OAM)

The region in PPU memory where sprite data is stored is called "Object Attribute Memory", or "OAM". This region of memory is different in that there are special MMIO addresses that the CPU can use to update the contents of OAM all at once, at high speed. Being able to update OAM quickly is a necessity for fast-paced games, so that all 64 sprites can move smoothly every frame.

To use this high-speed copying, the CPU needs to have all of the sprite data ready to go in a contiguous page of memory. (A page is a block of 256 bytes.) Generally, this "sprite buffer" is placed in CPU memory addresses $0200-$02ff. A few things to note about the memory map image here. First, "RAM" is listed as extending from $0300-$0800. In reality, the entire range from $0000-$0800 is RAM (2KB); the region from $0300-$0800 is just the portion of RAM that is not commonly allocated for a specific purpose. The space from $0800-$2000 in the memory map is empty space - writes to that region of memory will silently fail, and reads from that range have undefined behavior. Finally, note that the entire area from $8000-$FFFF comes from the PRG-ROM chip on the cartridge, including the six bytes that define the locations of interrupt handlers.

A common CPU memory layout for the NES. While it is possible to change where in memory some items like the stack or sprite buffer are located, this is the most common arrangement.

Within the sprite buffer (and in OAM itself), every four bytes defines one sprite. So, the first eight bytes of the sprite buffer look like this:

Memory address Purpose
$0200 Y position of sprite 0 (first sprite)
$0201 Tile number of sprite 0
$0202 Attribute flags for sprite 0
$0203 X position of sprite 0
$0204 Y position of sprite 1 (second sprite)
$0205 Tile number of sprite 1
$0206 Attribute flags for sprite 1
$0207 X position of sprite 1

$2003: OAMADDR and $4014: OAMDMA

Once we have set up all of the sprite data we want to transfer, we use two new MMIO addresses in our code to send all of the sprite data to the PPU. OAMADDR is used to set where in OAM we want to write to; for all of our projects (and for most commercial games), this will always be $00, the beginning of the OAM block. OAMDMA initiates the transfer of an entire page of memory into OAM. Writing the high byte of a memory address to OAMDMA will transfer that page.

While it might seem like we only need to write to OAM when something has changed, the OAM portion of PPU memory is implemented with "dynamic RAM", which means that it is highly unstable and needs to be continuously refreshed, even if nothing has changed. In practice, this means that we want to write to OAM once per frame of graphics (60 times per second).

Non-Maskable Interrupts (NMI)

Fortunately, the NES has an easy-to-use system for running code once per frame: the Non-Maskable Interrupt (NMI). NMI is one of the three interrupt vectors the 6502 knows how to handle. The NMI event is triggered each time the PPU enters "vblank", which occurs at the end of each frame of graphics. "Vblank" stands for "vertical blank"; there is a similar "Hblank" or "horizontal blank" as well. To understand what these terms mean, we need to look at how the CRT televisions and monitors of the era work.

Until new technologies like plasma and LCD took over in the mid-2000's, most televisions used a technology called "cathode-ray tube" or CRT. CRTs work by shooting a beam of electrons through an "electron gun" to hit the inside of a phosphorescent screen, which absorbs the energy of the electrons and converts it into light. The electron gun sweeps across the screen continuously in horizontal lines from top to bottom, starting at the top-left corner and ending at the bottom-right corner before beginning again. The speed of these sweeps is determined by the video signal the television is designed to display. The NTSC standard used in the US and Japan calls for 60 frames per second, while the competing PAL standard used in Europe uses 50 frames per second.

When the electron gun is resetting itself — either to start a new horizontal line from the left edge or when moving from bottom right to top left to start a new frame — the stream of electrons is temporarily stopped, so as not to inadvertently cause graphical issues. These "blanking periods" are the only times when the display on screen is not changing. "Hblank" occurs at the end of each horizontal line and is incredibly brief, lasting only 10.9 microseconds for NTSC. "Vblank" is comparatively much longer, though still short: about 1250 microseconds, or 0.00125 seconds.

Since Vblank is one of the only times that nothing is being output to the screen, and since Hblank is far too short to do meaningful work, it is common practice to perform most graphical updates during Vblank, i.e. as part of the NMI handler. Currently, the NMI handler in the test project looks like this:

.proc nmi_handler
  RTI
.endproc

RTI, as discussed previously, is the opcode for Return from Interrupt. Let's update the NMI handler to copy the memory from $0200-$02ff into OAM each time it runs:

.proc nmi_handler
  LDA #$00
  STA OAMADDR
  LDA #$02
  STA OAMDMA
  RTI
.endproc

A quick review of the assembly we learned back in Chapter 5 is in order. On line 2, we load the literal value zero into the accumulator. In case you've forgotten: a number given to an instruction like LDA is, by default, a memory address. LDA $00 means "load the accumulator with the value stored at memory address zero". Adding a "#" tells the assembler that this is a literal value, not a memory address. On line 3, we store (write) this zero to the OAMADDR address. This tells the PPU to prepare for a transfer to OAM starting at byte zero. Next, we load the literal value two into the accumulator, and write it to OAMDMA. This tells the PPU to initiate a high-speed transfer of the 256 bytes from $0200-$02ff into OAM.

In order to use OAMADDR and OAMDMA in our code, we need to update our constants file to include these new constants. Here is the updated constants.inc:

PPUCTRL   = $2000
PPUMASK   = $2001
PPUSTATUS = $2002
PPUADDR   = $2006
PPUDATA   = $2007
OAMADDR   = $2003
OAMDMA    = $2004

We now have a reliable and automated way to keep OAM up to date. But where do our sprite graphics come from? As mentioned in the last chapter, the CHR-ROM chip in a cartridge holds two pattern tables, one for sprites and one for backgrounds. We will need to create our own pattern tables to display sprites on screen. This is where NES Screen Tool comes in handy.

Using NES Screen Tool

Open NES Screen Tool. You should see a screen similar to this:

The main screen of NES Screen Tool.

Before we dive into creating sprites, let's take a quick tour of how NES Screen Tool is organized. The large area on the left two-thirds of the screen is a sort of canvas that can be used to draw backgrounds using tiles from a pattern table. We will not use this area until we start discussing background graphics. The right one-third of the interface is broadly divided into pattern tables ("Tileset") and palettes.

The "Tileset" area.

Within the "Tileset" area, the main element is a display of a pattern table. Just below the pattern table display is a toggle switch for "Pattern table A/B". As mentioned before, generally one pattern table is used for sprites, and the other is used for backgrounds. The A/B switch lets you flip between the two pattern tables. Also note that this panel has, in the bottom right, a "Grid" selector, with options for "All", "1x", "2x", and "4x". These grid options primarily affect the large canvas to the left, but turning on any of the grid options will also turn on the grid for the pattern table view, making it more clear what the boundaries of the tiles in the table are.

The "Palettes" area.

The "Palettes" area lets you preview how different palette choices will affect your tiles. The four numbered palettes at the top let you adjust the colors contained in the four available sprite palettes by clicking one of the four colors in the palette, then choosing a color from the color chart below. Changing the first color of any palette will change the first color of the other palettes automatically, as well as changing the background color of the large canvas area and the pattern table display. Hovering over the colors in a palette will display the color's hex value and the memory address for that palette index in the status bar along the bottom of the application window. Finally, the "A/B/C/D" selector to the right of the color table allows you to create four different "presets" of color palettes, so that you can quickly switch between color palette combinations to see the result on your tiles.

Making Your Own Tiles

To help you get started, I have created a starter .chr file that features some basic sprites and a full font in the background table. Download graphics.chr and open it in NES Screen Tool ("All" → "Open" → select graphics.chr).

To edit or create tiles, first click on the space in the pattern table display for the tile you wish to alter, then click "CHR Editor" in the menu bar. This will open a separate tile-editing window.

NES Screen Tool's "CHR Editor" window.

To edit a tile (or create new tiles!), select a palette index from any of the four palettes in the lower right of the main window, then click pixels in the CHR Editor panel to set them to that palette index. Note that clicking on different palettes in the main window changes all displayed colors to the colors from that palette. This is extremely useful for testing out what your tiles will look like in the palettes used by your game. Once you have created a set of tiles to work with, save your work by choosing "Patterns" → "Save CHR" → "Latest (8K)".

Displaying Sprites In Your Game

Download the full source code for this example: 10-spritegraphics.zip

In order to display our tiles in a game, we first need to make sure we load the .chr file that contains them. In our last project, we simply reserved 8,192 bytes of empty space for the CHR-ROM chip. Now that we have actual tiles, we will load them from a .chr file directly. Change the .segment "CHR" section to this:

.segment "CHR"
.incbin "graphics.chr"

.incbin, as you might expect, is a new assembler directive that instructs ca65 to include raw binary data (as opposed to .include, which is processed by the assembler). With our tiles in place, it's time to draw something. Here, we will use the four "spaceship" tiles from graphics.chr, but feel free to use your own tiles instead.

Next, we will need to fill out an entire palette, instead of just setting the first palette color to green ($29). We'll extend .proc main as follows:

.proc main
  ; write a palette
  LDX #$3f
  STX PPUADDR
  LDX #$00
  STX PPUADDR
  LDA #$29
  STA PPUDATA
  LDA #$19
  STA PPUDATA
  LDA #$09
  STA PPUDATA
  LDA #$0f
  STA PPUDATA

Notice that we only have to set an address with PPUADDR once; each time we write to PPUDATA, the PPU memory address is automatically increased by one.

Next, we need to store the data for our sprites. We are going to start by drawing one sprite, the top-left "corner" of the spaceship. As discussed earlier, we will store all our sprite information in memory between $0200-$02ff, and copy it to the PPU using DMA transfer (in our NMI handler). Let's continue modifying .proc main to copy sprite data into $0200-$02ff:


  ; write sprite data
  LDA #$70
  STA $0200 ; Y-coord of first sprite
  LDA #$05
  STA $0201 ; tile number of first sprite
  LDA #$00
  STA $0202 ; attributes of first sprite
  LDA #$80
  STA $0203 ; X-coord of first sprite

Finally, we need to make one more change to .proc main. In our previous examples, after writing palette data, we turn on the screen by writing to PPUMASK. However, now that we are using the NMI handler, we need to tell the CPU to generate NMI events. We can do that by writing to PPUCTRL, like this: For full details of the possible options that can be set via PPUCTRL and PPUMASK, see the NESDev wiki.


vblankwait:       ; wait for another vblank before continuing
  BIT PPUSTATUS
  BPL vblankwait

  LDA #%10010000  ; turn on NMIs, sprites use first pattern table
  STA PPUCTRL
  LDA #%00011110  ; turn on screen
  STA PPUMASK

At this point, our code will draw a single sprite (the top-left of the spaceship) to the screen, at X $80 and Y $70 (near the middle of the screen). The sprite data that we wrote to CPU memory between $0200-$02ff will be copied to PPU memory once per frame by our NMI handler. Assembling, linking, and running this code gives the output seen below.

While this works, our code is very inefficient. To draw all 64 sprites this way, we would need 128 lines of code. To make our sprite code more manageable, we will store the sprite data separately from the code that writes it to memory, and use a loop to iterate over the data. This requires a few new assembly opcodes, which we will learn next chapter.