DirectX® 9.0 and SmartShader™ 2.0 White Paper

Car Paint

This demonstration exemplifies some of the more innovative uses for DirectX 9.0 pixel shaders. In particular, it uses an interesting two-tone car paint effect, as well as a new normal mapping technique that allows 3D objects with low polygon counts to appear to have much more detail.

The car's paint is made up of three layers. The base layer consists of a range of colors, which are stored in a one-dimensional texture. The color used for a given pixel is sampled from this texture according to the viewing angle. In the default example, parts of the car viewed head-on appear yellow, while those viewed at sharp angles or nearly edge-on appear red. Pixels viewed at intermediate angles appear to have a blend of these two basic tones.

The second layer consists of tiny metal flakes or "sparkles" that give some car paints a metallic look. Since these flakes are flat and two-dimensional, they only appear to reflect light when viewed head-on, and seem to disappear when viewed edge-on. To recreate this effect, a subtle noise texture is used. This texture is blended with the base layer to create small variations in the brightness of each pixel, and is also used to perturb the base texture to make the transitions from yellow to red appear fuzzier and less distinct. The result is that the parts of the car viewed directly appear more "sparkly" than those viewed at a sharp angle, which convincingly reproduces the light scattering behavior of real-life car paint.

The third layer is a glossy, transparent clear-coat layer. This layer is simulated using a cube environment map, which is projected on to the surface of the car as a reflection. With the advanced pixel shader capabilities of the Radeon 9700 & 9500 series GPUs, all three layers of the car paint can be generated in a single rendering pass, meaning this complex effect can be handled effortlessly and applied to multiple objects in a scene.

This demo makes extensive use of normal maps (also known as bump maps) on the car and the turntable. A normal is a three-dimensional value that determines the facing of a surface at any given point, which in turn determines the direction that light will be reflected. Since a standard polygon is a flat surface, all of its normals by default point in the same direction. By varying the normals across a flat surface, a normal map can create the illusion of fine details and smooth curvature without requiring additional polygons.

The tricky part about using this technique is the generation of high quality normal maps. There are two primary difficulties: authoring the detailed normal map, and having the necessary precision to faithfully reproduce subtle details. An innovative way to address the first issue is to start with a very high polygon count model, and then create a second version of the same model with a much lower polygon count. For example, the original car model used in the demo was comprised of over one million polygons, while the version actually displayed has only a few thousand polygons. Most 3D game artists are already accustomed to this process, since game engines tend to severely restrict the polygon counts that can be used in order to maintain reasonable performance across a wide range of system configurations.

Once the two versions of the model are completed, they are run through a ray-casting algorithm that calculates the differences between them. These difference values are then filtered and stored as a high precision normal map, which can be applied to the low polygon model while maintaining all of the fine surface detail from the high polygon model. The car paint demo uses normal maps with 16 bits per component. DirectX 8.1 and earlier hardware are limited to just 8 bits per component for normal maps, which can result in image artifacts on subtly curved surfaces.

High precision normal maps can get quite large, and they cannot be easily compressed using standard texture compression methods like DXTC. The car paint demo, however, makes use of a new square root operation introduced in the 2.0 Pixel Shader specification to implement a simple yet effective normal map compression scheme. Each location in a normal map requires three pieces of data, representing the x, y & z components of a 3D vector. A pixel shader using the Pythagorean theorem ( ) to derive the z components from the x & y components makes it possible to compress the normal map to just 2/3 of its original size, with no loss of data.