|Previous||Table of Contents||Next|
This rasterizer, which we call the subdivision rasterizer, first draws all the vertices in the model. Then it takes each front-facing triangle, and determines if it has a side thats at least two pixels long. If it does, we split that side into two pieces at the pixel nearest to the middle (using adds and shifts to average the endpoints of that side), draw the vertex at the split point, and process each of the two split triangles recursively, until we get down to triangles that have only one-pixel sides and hence have nothing left to draw. This approach is hideously slow and quite ugly (due to inaccuracies from integer quantization) for 100-pixel trianglesbut its very fast for, say, five-pixel triangles, and is indistinguishable from more accurate rasterization when a model is 25 or 50 feet away. Better yet, the subdivider is ridiculously simplea few dozen lines of code, far simpler than the affine rasterizerand was implemented in an evening, immediately making the drawing of distant models about three times as fast, a very good return for a bit of conceptual work. The affine rasterizer got fairly close to the same performance with further optimizationin the range of 10% to 50% slowerbut that took weeks of difficult programming.
We switch between the two rasterizers based on the models distance and average triangle size, and in almost any scene, most models are far enough away so subdivision rasterization is used. There are undoubtedly faster ways yet to rasterize distant models adequately well, but the subdivider was clearly a win, and is a good example of how thinking in a radically different direction can pay off handsomely.
We had hoped to be able to eliminate sprites completely, making Quake 100% 3-D, but spritesalthough sometimes very visibly 2-Dwere used for a few purposes, most noticeably the cores of explosions. As of CGDC last year, explosions consisted of an exploding spray of particles (discussed below), but there just wasnt enough visual punch with that representation; adding a series of sprites animating an explosion did the trick. (In hindsight, we probably should have made the explosions polygon models rather than sprites; it would have looked about as good, and the few sprites we used didnt justify the considerable amount of code and programming time required to support them.) Drawing a sprite is similar to drawing a normal polygon, complete with perspective correction, although of course the inner loop must detect and skip over transparent pixels, and must also perform z-buffering.
The last drawing entity type is particles. Each particle is a solid-colored rectangle, scaled by distance from the viewer and drawn with z-buffering. There can be up to 2,000 particles in a scene, and they are used for rocket trails, explosions, and the like. In one sense, particles are very primitive technology, but they allow effects that would be extremely difficult to do well with the other types of entities, and they work well in tandem with other entities, as, for example, providing a trail of fire behind a polygon-model lava ball that flies into the air, or generating an expanding cloud around a sprite explosion core.
Since shipping Quake in the summer of 1996, weve extended it in several ways: Weve worked with Rendition to port it to the Verite accelerator chip, weve ported it to OpenGL, weve ported it to Win32, weve done QuakeWorld, and weve added features for Quake 2. Ill discuss each of these briefly.
Verite Quake (VQuake) was the first hardware-accelerated version of Quake. It looks extremely good, due to bilinear texture filtering, which eliminates most pixel aliasing, and because it provides good performance at higher resolutions such as 512×384 and 640×480. Implementing VQuake proved to be an interesting task, for two reasons: The Verite chips fill rate was marginal for Quakes needs, and Verite contains a programmable RISC chip, enabling more sophisticated processing than most 3-D accelerators. The need to squeeze as much performance as possible out of Verite ruled out the use of a standard API such as Direct 3D or OpenGL; instead, VQuake uses Renditions proprietary API, Speedy3D, with the addition of some special calls and custom Verite code.
Interestingly, VQuake is very similar to software Quake; in order to allow Verite to handle the high pixel processing loads of high-res, VQuake uses an edge list and builds span lists on the CPU, just as in software Quake, then Verite DMAs the span descriptors to onboard memory and draws them. (This was only possible because Verite is fully programmable; most accelerators wouldnt be able to support this architecture.) Similarly, the CPU builds lit, tiled surfaces in system RAM, then Verite DMAs them to an onboard surface cache, from which they are texture-mapped. In short, VQuake is very much like normal Quake, except that the drawing of the spans is done by a specialized processor.
This approach works well, but some of the drawbacks of a surface cache become more noticeable when hardware is involved. First, the DMAing is an extra step thats not necessary in software, slowing things down. Second, onboard memory is a relatively limited resource (4 MB total), and textures must be 16-bpp (because hardware can only do filtering in RGB modes), thus eating up twice as much memory as the software versions 8-bpp texturesand memory becomes progressively scarcer at higher resolutions, especially given the need for a z-buffer and two 16-bpp pages. (Note that using the edge list helps here, because it filters out spans from polygons that are in the PVS but fully occluded, reducing the number of surfaces that have to be downloaded.) Surface caching in VQuake usually works just fine, but response when coming around corners into complex scenes or when spinning can be more sluggish than in software Quake.
An alternative to surface caching would have been to do two passes across each span, one tiling the texture, and the other doing an alpha blend using the light map as a texture, to light the texture (two-pass alpha lighting). This approach produces exactly the same results as the surface cache, without requiring downloading and caching of large surfaces, and has the advantage of very level performance. However, this approach requires at least twice the fill rate of the surface cache approach, and Verite didnt have enough fill rate for that at higher resolutions. Its also worth noting that two-pass alpha lighting doesnt have the same potential for procedural texturing that surface caching does. In fact, given MMX and ever-faster CPUs, and the ability of the CPU and the accelerator to process in parallel, it will become increasingly tempting to use the CPU to build surfaces with procedural texturing such as bump mapping, shimmers, and warps; this sort of procedural texturing has the potential to give accelerated games highly distinctive visuals. So the choice between surface caching and two-pass alpha lighting for hardware accelerators depends on a games needs, and it seems most likely that the two approaches will be mixed together, with surface caching used for special surfaces, and two-pass alpha lighting used for most drawing.
|Previous||Table of Contents||Next|