Writing Hardware Shaders

Hardware shaders are closely tied to the graphics hardware available. It is generally not possible to use a Cg shader that uses special NVIDIA hardware features for ATI hardware, and vice versa. The main classes are:

NVIDIA shaders are written in NVIDIA's Cg 1.2 language, which is loosely based on C. It supports shader graphs and is well-suited for mental ray's shader graphs and Phenomena. NVIDIA hardware shaders come in pairs: a vertex shader operates on triangle vertices, and a fragment shader operates on pixel colors. Most of the time, only a fragment shader is supplied; non-default vertex shader are rarely needed.
OpenGL prior to version 2.0 does not support programmable shaders at all. All functions must be mapped to the hardwired shading model. Only very simple shaders can be mapped to OpenGL hardware, but it is still useful for very fast low-quality previews.
OpenGL 2.0 supports a native shading language similar to NVIDIA's Cg, but it has not been finalized at this time and will not be described here.
HLSL support may be added in future versions of mental ray that support DirectX. HLSL is the native shader language of DirectX, which is Microsoft's proprietary alternative to OpenGL It and may become available for special OEM versions of mental ray. Unlike NVIDIA's Cg, HLSL is not portable.

The main difference between software and hardware shaders to keep in mind is that hardware shading always proceeds triangle by triangle, not ray by ray. Each of the three vertices of the triangle pass through a vertex shader which may move vertices, alter normals, or set up other information tied to the vertex.

When all three vertices have been set up, the hardware rasterizes the triangle by picking points on the triangle that correspond to image pixels. For each of them, the vertex information stored by the vertex shader is interpolated, and the fragment shader is called to compute a pixel color value that is then stored or merged into the frame buffer.

For the shader writer, this means that there is no geometrical information except as stored by the vertex shader. There is no way to, for example, trace a ray, find a ray or volume length, or do other operations that involve other geometry. Reflections, shadows, and other features are available only through precomputed map images such as texture maps.

mental ray supports a large number of shader types. Only material and light shaders are directly supported. Since Cg supports shader graphs, subshaders including texture shaders are supported as well.

shader type	support
material	supported

light	supported

subshaders	supported in Cg 1.2

texture	supported

volume	not supported

environment	supported for cameras but not materials

lens	not supported; will never support ray bending

output	not supported

shadow	not applicable; shadow shaders are called during ray tracing only. Hardware rendering uses shadow maps instead.

photon	not applicable; photon mapping is based on ray tracing and must be completed before hardware rendering begins.

geometry	not applicable, geometry is built before rendering begins.

displacement	not applicable for the same reason, although vertex shaders can be written that displace preexisting vertices.

The next sections will show how to write a simple Phong shader and a point light shader.