|
|
 |
 |
 | | | | |
 |
|||
 |
||||||
          |
 |
 |
||||
|
Tool Postmortem: Climax Brighton's Supertools Curved Surfaces
A bezier patch is a rectangular area defined by sixteen control points. One at each corner, two along each edge, and four in the middle. They are simple, easy to model with, and can be efficiently implemented in hardware, but they are plagued by the problem of surface continuity. In simple terms this means that when you put two patches next to each other to build up a larger surface, they must be carefully arranged to make sure the surface will be smooth across the join. The traditional solution is to use NURBS (Non Uniform Rational B-Splines). NURBS are a superset of the bezier patch, and allow large surfaces with many control points to be described as a single object which will always be perfectly smooth. After it has been constructed, a NURBS surface can be subdivided into multiple bezier patches for rendering, or at the simplest level, a NURBS object with the knot values (0, 0, 1, 1) is identical to a bezier patch. The trouble with NURBS is that you simply cannot build efficient game artwork with them. Because they consist of a single large grid of control points, there is no way to add detail only where needed. For instance if you want an extra little curve in the nose of a character, you have to add a whole row of new control points running the entire way around the back of the head. Far too wasteful! The big secret is that for games, surface continuity actually doesn't matter very much. As long as there are no holes between patches (which is trivial to arrange), and as long as the normals point in more or less the same direction along the join, it makes no difference if they are slightly wrong. Minor visual errors can be corrected by averaging the normals from adjacent patches, and even though this is mathematically bogus, who cares as long as nobody notices? The only thing that really matters is whether you can build good looking and efficient models quickly enough to get the game finished on time. Efficient modelling requires that you use many small patches only in areas where detail is needed, and fewer large patches elsewhere. To connect the two areas, irregular triangular constructions will inevitably be required, but NURBS prevent that for the sake of maintaining an unnecessary level of surface smoothness. Using Maya, we had to work around the system by creating every single patch as an individual NURBS surface, and writing our own plugins to line them up next to each other. This was awkward, and became pathologically slow for large models as Maya was never designed to have so many NURBS objects in one scene. Max does support arbitrary combinations of rectangular and triangular patches, but, again in the interest of maintaining good continuity, it hides access to the internal control points, which hugely limits the range of shapes that can be created. When nothing else can do the job, the only option is to write your own. SuperModel would be useless for many tasks as it completely ignores the entire continuity issue, but for making games it is exactly what we needed. Bezier curves aren't the only option, of course. Max and Maya both also support subdivision surfaces, which are appealing because they gracefully handle arbitrary types of mesh construction. Subdivs are less suitable for realtime tesselation, however, as they need too much information about the surrounding mesh, and they produce high output triangle counts from their recursive passes over the geometry, going up in steps of 4, 16, 64, 256, while patches can tesselate to 4, 9, 16, 25, and so on. The coarse nature of subdiv stepping is no problem for offline tesselation, where you can follow the subdivision with a progressive mesh style collapse back down to any desired polygon count, but one of the constraints on our choice was that the technique be appropriate for realtime implementation in the PS2 vector unit, for which bezier patches made a great deal more sense.
In most art packages, polygonal and curved surface modelling are entirely different modes, using different tools and having to be learned independently. In contrast, SuperModel lets the artists mix and match geometry types within a single object. If it curves along both axis, use a bezier patch. If it only curves in one axis, use a bezier strip. If it doesn't curve at all, use a polygon. It doesn't matter: everything is still made up of vertices and edges, which are manipulated using the same set of tools. If you extrude the edge of a polygon, another polygon will be created, but if you extrude the edge of a patch it will create another patch. Strips behave as either polygons or patches depending on which edge you select, so you can make half an object out of patches and the other half from polygons, while still having it entirely connected and all the edges welded together. The more heavily a bezier model is optimised, the more difficult it becomes to texture, as the texture map will distort in strange ways when it is applied to the curved surface. Max simply ignores this problem (you have to live with the texture warping, or just not build such strangely curved patches in the first place), while Maya works around it by not supporting arbitrary UV mapping on curved surfaces. Not ideal for game use! SuperTed solves this problem by supporting bezier interpolation for UV coordinates as well as for vertex positions. The UV control points can be automatically calculated to make a texture sit "flat" on a curved surface, or they can be edited by hand for particularly troublesome patches. These tools allow us to accurately and efficiently texture geometry in ways that would be impossible with any other software. Multitexturing
Modern gamers expect to see individual blades of grass and bits of gravel along with a decent amount of variety in the world, but although the hardware can easily draw this, consoles lack the memory to hold too many large textures. If you don't have room to store it, you'll have to synthesise it at runtime. Take a low resolution basic texture, add a layer of high frequency but repetitive detail, and over the top of that add some dirt or scratches. Reuse your set of dirt textures in different combinations over various base textures, and you can construct a massive number of variations from only a few building blocks. The nice thing is that this way of working turns out to be intuitive and easily understood by artists, as it is similar to the way they would have traditionally constructed a texture using multiple layers in Photoshop. Most multitexture editors work by creating materials which combine a number of textures and blend modes, and then mapping these composite materials onto faces. In contrast, SuperTed works entirely with individual textures. It supports anywhere from one to sixteen layers, which can be textured independently, and every texture specifies how it is to be blended with the layer below it. This approach encourages the artists to reuse textures in creative ways, for instance adding in a bit of sand to make the side of a building look dusty, or UV mapping a single black circle to approximate the shadows from a huge variety of objects. A few statistics might be in order here. Donington, one of the tracks in MotoGP, uses 352 textures, which are arranged into 602 unique combinations. Although the artists only choose one of seven possible blend modes for each texture, with three layers in use MotoGP ended up requiring a total of 84 different pixel shaders, 62 of which are used on the environment. The downside to this variety of layers is that one patch is unlikely to be using the same set of textures as the ones around it, so our geometry tends to strip rather poorly. But I don't care: we may not be winning prizes for raw triangle throughput, but as long as we can manage a reasonable amount of geometry, I think it is more important to keep every surface looking as rich and interesting as possible. This layer based approach to multitexturing does a good job of scaling across differences in hardware capabilities. Our PS2 projects are using two texture layers, with the gouraud alpha controlling a crossfade between them. On Gamecube we add a third multiply mode detail layer, while on Xbox the flexibility of pixel shaders lets the artists choose any possible combine modes for their three layers, with the fourth generally reserved for the programmers to do dynamic lighting or reflections.
|
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
|
Features
