It’s after midnight when the carnage begins. Inside a castle, soldiers chase Nazis through the halls. A flame-thrower unfurls a hideous tongue of fire. This is Return to Castle Wolfenstein, a computer game that’s as much a scientific marvel as it is a visceral adventure. It’s also the latest product of Id Software (Mesquite, Texas). Through its technologically innovative games, Id has had a huge influence on everyday computing, from the high-speed, high-color, and high-resolution graphics cards common in today’s PCs to the marshalling of an army of on-line game programmers and players who have helped shape popular culture.
Id shot to prominence 10 years ago with the release of its original kill-the-Nazis-and-escape game, Wolfenstein 3D. It and its successors, Doom and Quake, cast players as endangered foot soldiers, racing through mazes while fighting monsters or, if they so chose, each other. To bring these games to the consumer PC and establish Id as the market leader required skill at simplifying difficult graphics problems and cunning in exploiting on-going improvements in computer graphics cards, processing power, and memory size. To date, their games have earned over US $150 million in sales, according to The NPD Group, a New York City market research firm.
It all began with a guy named Mario
The company owes much of its success to advances made by John Carmack, its 31-year-old lead programmer and cofounder who has been programming games since he was a teenager.
Back in the late 1980s, the electronic gaming industry was dominated by dedicated video game consoles. Most game software was distributed in cartridges, which slotted into the consoles, and as a consequence, writing games required expensive development systems and corporate backing.
The only alternative was home computer game programming, an underworld in which amateurs could develop and distribute software. Writing games for the low-powered machines required only programming skill and a love of gaming.
Four guys with that passion were artist Adrian Carmack; programmer John Carmack (no relation); game designer Tom Hall; and programmer John Romero. While working together at Softdisk (Shreveport, La.), a small software publisher, these inveterate gamers began moonlighting on their own titles.
At the time, the PC was still largely viewed as being for business only. It had, after all, only a handful of screen colors and squeaked out sounds through a tiny tinny speaker. Nonetheless, the Softdisk gamers figured this was enough to start using the PC as a games platform.
First, hey decided to see if they could recreate on a PC the gaming industry’s biggest hit at the time, Super Mario Brothers 3. This two-dimensional game ran on the Super Nintendo Entertainment System, which drove a regular television screen. The object was to make a mustached plumber, named Mario, leap over platforms and dodge hazards while running across a landscape below a blue sky strewn with puffy clouds. As Mario ran, the terrain scrolled from side to side to keep him more or less in the middle of the screen. To get the graphics performance required, the Nintendo console resorted to dedicated hardware. “We had clear examples of console games [like Mario] that did smooth scrolling,” John Carmack says, “but [in 1990] no one had done it on an IBM PC.”
After a few nights of experimentation, Carmack figured out how to emulate the side-scrolling action on a PC. In the game, the screen image was drawn, or rendered, by assembling an array of 16-by-16-pixel tiles. Usually the on-screen background took over 200 of these square tiles, a blue sky tile here, a cloud tile there, and so on. Graphics for active elements, such as Mario, were then drawn on top of the background.
Any attempt to redraw the entire background every frame resulted in a game that ran too slowly, so Carmack figured out how to have to redraw only a handful of tiles every frame, speeding the game up immensely. His technique relied on a new type of graphics card that had become available, and the observation that the player’s movement occurred incrementally, so most of the next frame’s scenery had already been drawn.
The new graphics cards were known as Enhanced Graphics Adapter (EGA) cards. They had more on-board video memory than the earlier Color Graphics Adapter (CGA) cards and could display 16 colors at once, instead of four. For Carmack, the extra memory had two important consequences. First, while intended for a single relatively high-resolution screen image, the card’s memory could hold several video screens’ worth of low-resolution images, typically 300 by 200 pixels, simultaneously, good enough for video games. By pointing to different video memory addresses, the card could switch which image was being sent to the screen at around 60 times a second, allowing smooth animation without annoying flicker. Second, the card could move data around in its video memory much faster than image data could be copied from the PC’s main memory to the card, eliminating a major graphics performance bottleneck.
Carmack wrote a so-called graphics display engine that exploited both properties to the full by using a technique that had been originally developed in the 1970s for scrolling over large images, such as satellite photographs. First, he assembled a complete screen in video memory, tile by tile—plus a border one tile wide [see illustration, “Scrolling With the Action” ]. If the player moved one pixel in any direction, the display engine moved the origin of the image it sent to the screen by one pixel in the corresponding direction. No new tiles had to be drawn. When the player’s movements finally pushed the screen image to the outer edge of a border, the engine still did not redraw most of the screen. Instead, it copied most of the existing image—the part that would remain constant—into another portion of video memory. Then it added the new tiles and moved the origin of the screen display so that it pointed to the new image.
In short, rather than having the PC redraw tens of thousands of pixels every time the player moved, the engine usually had to change only a single memory address—the one that indicated the origin of the screen image—or, at worst, draw a relatively thin strip of pixels for the new tiles. So the PC’s CPU was left with plenty of time for other tasks, such as drawing and animating the game’s moving platforms, hostile characters, and the other active elements with which the player interacted.
Hall and Carmack knocked up a Mario clone for the PC, which they dubbed Dangerous Dave in Copyright Infringement. But Softdisk, their employer, had no interest in publishing what were then high-end EGA games, preferring to stick with the market for CGA applications. So the nascent Id Software company went into moonlight overdrive, using the technology to create its own side-scrolling PC game called Commander Keen. When it came time to release the game, they hooked up with game publisher Scott Miller, who urged them to go with a distribution plan that was as novel as their technology: shareware.
In the 1980s, hackers started making their programs available through shareware, which relied on an honor code: try it and if you like it, pay me. But it had been used only for utilitarian programs like file tools or word processors. The next frontier, Miller suggested, was games. Instead of giving away the entire game, he said, why not give out only the first portion, then make the player buy the rest? Id agreed to let Miller’s company, Apogee, release the game. Prior to Commander Keen, Apogee’s most popular shareware game had sold a few thousand copies. Within months of Keen’s release in December 1990, the game had sold 30 000 copies. For the burgeoning world of PC games, Miller recalls, “it was a little atom bomb.”
Going for depth
Meanwhile programmer Carmack was again pushing the graphics envelope. He had been experimenting with 3-D graphics ever since junior high school, when he produced wire-frame MTV logos on his Apple II. Since then, several game creators had experimented with first-person 3-D points of view, where the flat tiles of 2-D games are replaced by polygons forming the surfaces of the player’s surrounding environment. The player no longer felt outside, looking in on the game’s world, but saw it as if from the inside.
The results had been mixed, though. The PC was simply too slow to redraw detailed 3-D scenes as the player’s position shifted. It had to draw lots of surfaces for each and every frame sent to the screen, including many that would be obscured by other surfaces closer to the player.
Carmack had an idea that would let the computer draw only those surfaces that were seen by the player. “If you’re willing to restrict the flexibility of your approach,” he says, “you can almost always do something better.”
So he chose not to address the general problem of drawing arbitrary polygons that could be positioned anywhere in space, but designed a program that would draw only trapezoids. His concern at this time was with walls (which are shaped like trapezoids in 3-D), not ceilings or floors.
For his program, Carmack simplified a technique for rendering realistic images on then high-end systems. In raycasting, as it is called, the computer draws scenes by extending lines from the player’s position in the direction he or she is facing. When it strikes a surface, the pixel corresponding to that line on the player’s screen is painted the appropriate color. None of the computer’s time is wasted on drawing surfaces that would never be seen anyway. By only drawing walls, Carmack could raycast scenes very quickly.
Carmack’s final challenge was to furnish his 3-D world with treasure chests, hostile characters, and other objects. Once again, he simplified the task, this time by using 2-D graphical icons, known as sprites. He got the computer to scale the size of the sprite, depending on the player’s location, so that he did not have to model the objects as 3-D figures, a task that would have slowed the game painfully. By combining sprites with raycasting, Carmack was able to place players in a fast-moving 3-D world. The upshot was Hovertank, released in April 1991. It was the first fast-action 3-D first-person action shooter for the PC.
Around this time, fellow programmer Romero heard about a new graphics technique called texture mapping. In this technique, realistic textures are applied to surfaces in place of their formerly flat, solid colors. in green slime in its next game, Catacombs 3D. While running through a maze, the player shot fireballs at enemy figures using another novelty—a hand drawn in the lower center of the screen. It was as if the player were looking down on his or her own hand, reaching into the computer screen. By including the hand in Catacombs 3D, Id Software was making a subtle, but strong, psychological point to its audience: you are not just playing the game—you’re part of it.
For Id’s next game, Wolfenstein 3D, Carmack refined his code. A key decision ensured the graphics engine had as little work to do as possible: to make the walls even easier to draw, they would all be the same height.
This speeded up raycasting immensely. In normal raycasting, one line is projected through space for every pixel displayed. A 320-by-200-pixel screen image of the type common at the time required 64 000 lines. But because Carmack’s walls were uniform from top to bottom, he had to raycast along only one horizontal plane, just 320 lines [see diagram, “Raycasting 3-D Rooms”].
With Carmack’s graphics engine now blazingly fast, Romero, Adrian Carmack, and Hall set about creating a brutal game in which an American G.I. had to mow down Nazis while negotiating a series of maze-based levels. Upon its release in May 1992, Wolfenstein 3D was an instant sensation and became something of a benchmark for PCs. When Intel wanted to demonstrate the performance of its new Pentium chip to reporters, it showed them a system running Wolfenstein.
Wolfenstein also empowered gamers in unexpected ways—they could modify the game with their own levels and graphics. Instead of a Nazi officer, players could, for example, substitute Barney, the purple dinosaur star of U.S. children’s television. Carmack and Romero made no attempt to sue the creators of these mutated versions of Wolfenstein, for, as hackers themselves, they couldn’t have been more pleased.
Their next game, Doom, incorporated two important effects Carmack had experimented with in working on another game, Shadowcaster, for a company called Raven in 1992. One was to apply texture mapping to floors and ceilings, as well as to walls. Another was to add diminished lighting. Diminished lighting meant that, as in real life, distant vistas would recede into shadows, whereas in Wolfenstein, every room was brightly lit, with no variation in hue.
By this time, Carmack was programming for the Video Graphics Adapter (VGA) cards that had supplanted the EGA cards. VGA allowed 256 colors—a big step up from EGA’s 16, but still a limited range that made it a challenge to incorporate all the shading needed for diminished lighting effects.
The solution was to restrict the palette used for the game’s graphics, so that 16 shades of each of 16 colors could be accommodated. Carmack then programmed the computer to display different shades based on the player’s location within a room. The darkest hues of a color were applied to far sections of a room; nearer surfaces would always be brighter than those farther away. This added to the moody atmosphere of the game.
Both Carmack and Romero were eager to break away from the simple designs used in the levels of their earlier games. “My whole thing was—let’s not do anything that Wolfenstein does,” Romero says. “Let’s not have the same light levels, let’s not have the same ceiling heights, let’s not have walls that are 90 degrees [to each other]. Let’s show off Carmack’s new technology by making everything look different.”
Profiting from improvements in computer speed and memory, Carmack began working on how to draw polygons with more arbitrary shapes than Wolfenstein’s trapezoids. “It was looking like [the graphics engine] wouldn’t be fast enough,” he recalls, “so we had to come up with a new approach....I knew that to be fast, we still had to have strictly horizontal floors and vertical walls.” The answer was a technique known as binary space partitioning (BSP). Henry Fuchs, Zvi Kedem, and Bruce Naylor had popularized BSP techniques in 1980 while at Bell Labs to render 3-D models of objects on screen.
A fundamental problem in converting a 3-D model of an object into an on-screen image is determining which surfaces are actually visible, which boils down to calculating: is surface Y in front of, or behind, surface X? Traditionally, this calculation was done any time the model changed orientation.
The BSP approach depended on the observation that the model itself is static, and although different views give rise to different images, there is no change in the relationships between its surfaces. BSP allowed the relationships to be determined once and then stored in such a way that determining which surfaces hid other surfaces from any arbitrary viewpoint was a matter of looking up the information, not calculating it anew.
BSP takes the space occupied by the model and partitions it into two sections. If either section contains more than one surface of the model, it is divided again, until the space is completely broken up into sections each containing one surface. The branching hierarchy that results is called a BSP tree and extends all the way from the initial partition of the space down to the individual elements. By following a particular path through the nodes of the stored tree, it is possible to generate key information about the relationships between surfaces in a specific view of the model.
What if, Carmack wondered, you could use a BSP to create not just one 3-D model of an object, but an entire virtual world? Again, he made the problem simpler by imposing a constraint: walls had to be vertical and floors and ceilings horizontal. BSP could then be used to divide up not the 3-D space itself, but a much simpler 2-D plan view of that space and still provide all the important information about which surfaces were in front of which [see diagram, “Divide and Conquer”].
Doom was also designed to make it easy for hackers to extend the game by adding their own graphics and game-level designs. Networking was added to Doom, allowing play between multiple players over a local-area network and modem-to-modem competition.
The game was released in December 1993. Between the multiplayer option, the extensibility, the riveting 3-D graphics, and the cleverly designed levels, which cast the player as a futuristic space marine fighting against the legions of hell, it became a phenomenon. Doom II, the sequel, featured more weapons and new levels but used the same graphics engine. It was released in October 1994 and eventually sold more than 1 500 000 copies at about $50 each; according to the NPD Group, it remains the third best-selling computer game in history.
The finish line
In the mid-1990s, Carmack felt that PC technology had advanced far enough for him to finally achieve two specific goals for his next game, Quake. He wanted to create an arbitrary 3-D world in which true 3-D objects could be viewed from any angle, unlike the flat sprites in Doom and Wolfenstein. The solution was to use the power of the latest generation of PCs to use BSP to chop up the volume of a true 3-D space, rather than just areas of a 2-D plan view. He also wanted to make a game that could be played over the Internet.
For Internet play, a client-server architecture was used. The server—which could be run on any PC—would handle the game environment consisting of rooms, the physics of moving objects, player positions, and so on. Meanwhile, the client PC would be responsible for both the input, through the player’s keyboard and mouse, and the output, in the form of graphics and sound. Being online, however, the game was liable to lags and lapses in network packet deliveries—just the thing to screw up a fast action game. To reduce the problem, Id limited the packet delivery method to only the most necessary information, such as a player’s position.
“The key point was use of an unreliable transport for all communication,” Carmack says, “taking advantage of continuous packet communication and [relaxing] the normal requirements for reliable delivery,” such as handshaking and error correction. A variety of data compression methods were also used to reduce the bandwidth. The multiplayer friendliness of the game that emerged—Quake—was rewarded by the emergence of a huge online community when it was released in June 1996.
Games in general drove the evolution of video cards. But multiplayer games in particular created an insatiable demand for better graphics systems, providing a market for even the most incremental advance. Business users are not concerned if the graphics card they are using to view their e-mail updates the screen 8 times a second while their neighbor’s card allows 10 updates a second. But a gamer playing Quake, in which the difference between killing or being killed is measured in tenths of a second, very much cares.
Quake soon became the de facto benchmark for the consumer graphics card industry. Says David Kirk, chief scientist of NVIDIA, a leading graphics processor manufacturer in Santa Clara, Calif., “Id Software’s games always push the envelope.”
Quake II improved on its predecessor by taking advantage of hardware acceleration that might be present in a PC, allowing much of the work of rendering 3-D scenes to be moved from the CPU to the video card. Quake III, released in December 1999, went a step further and became the first high-profile game to require hardware acceleration, much as Id had been willing to burn its boats in 1990 by insisting on EGA over CGA with Commander Keen.
Carmack himself feels that his real innovations peaked with Quake in 1996. Everything since, he says, is essentially refining a theme. Return to Castle Wolfenstein, in fact, was based on the Quake III engine, with much of the level and game logic development work being done by an outside company.
“There were critical points in the evolution of this stuff,” Carmack says, “getting into first person at all, then getting into arbitrary 3-D, and then getting into hardware acceleration....But the critical goals have been met. There’s still infinite refinement that we can do on all these different things, but...we can build an arbitrary representational world at some level of fidelity. We can be improving our fidelity and our special effects and all that. But we have the fundamental tools necessary to be doing games that are a simulation of the world.”