Rasterizer

mental ray 3.2 and later support a rendering algorithm called Rapid Motion, which was rewritten in mental ray 3.4 and renamed rasterizer. Its primary difference to regular scanline rendering is its separation of sampling and sample compositing (also called sample collection). Without the rasterizer, mental ray selects spatial and temporal sample points (eye rays) on the image plane in the shutter interval. If an object moves, it will be shaded in multiple samples at different points in time. Since rendering time is roughly proportional to the number of shaded samples, it rises quickly if an object moves quickly.

The rasterizer works by sampling all objects at a fixed time, and caching the shaded samples for re-use. If the object moves, these sample results are re-used at every point the object passes over. The cache is tied to the geometry:

a number of sample points is selected on each triangle based on the visibility, size, and orientation of the triangle. These points are then individually shaded by calling the material shaders as usual. The results are stored in front-to-back order for later combination. Care is taken to minimize the shading of points hidden behind other geometry, but although rendering proceeds in a roughly front-to-back order, there is no guarantee of the exact order, unlike for the regular scanline algorithm or raytracing. For this reason, only the surface shading and transparency are stored initially, and volume and environment shading are calculated later on.
The tile is scanned, and all front-to-back stored shading results are composited to individual screen samples, using their opacities to combine their colors, and the volume and environment shaders are called and combined with the surface shading.

The late compositing of shading samples to form screen samples, and re-using of shading results has several important consequences:

If the material shader traces rays with shader API functions like mi_trace_reflection, the result is re-used at all points the object moves across. This has the effect that the object appears to "drag" reflections and refractions with it. For example, if a mirror that is coplanar to the image plane moves sideways, its edges are always blurred, but the objects being reflected would be blurred only with the rasterizer.
Transparency (mi_trace_transparent) can be calculated by the regular scanline algorithm without tracing rays, by following the chain of depth-sorted triangles behind the current point on the image plane. Since the rasterizer shades points on triangles one by one, and combines the results according to depth at the later compositing stage, mi_trace_transparent will always return false and transparent black. As long as the shader does not do nonstandard linear compositing, this gives the same results, but if the shader makes decisions such as casting environment rays based on the value returned by mi_trace_transparent, the results may differ from expectations.
In particular, shaders that implement matte objects will not work without modification. Matte objects are placeholders for later compositing outside of mental ray; like transparent cut-outs where live action will be added later. Since the rasterizer ties its shading sample combining to the alpha component of the RGBA color returned by the material shader, it will fill in such cut-outs. To avoid this, a shader may use the new mi_opacity_set function to explicitly set the opacity for the compositing stage independently of the returned alpha value. In other words, if an explicit opacity value is set, the alpha channel of the shading result color is ignored for calculating transparency, and is just retained for writing to the final output buffer. Instead, the opacity color is used to combine shading values front-to-back, whereas in the absence of the opacity color, alpha is used to combine shading samples front-to-back. A matte object could have alpha of 0 but set an opacity of 1. In this manner one can render solid objects with transparency built in, for correct results during later, external compositing. There is also an mi_opacity_get function to support cooperating shaders in Phenomena.

The rasterizer is enabled with the command-line option -scanline rast or the statement scanline rast in the options block of the scene file. mental ray 3.4 controls the pixel sampling rate with -samples_collect, which gives the number of samples per pixel-dimension. For example, the default value of 3 gives 9 samples per pixel. The rate of shading is controlled by -shading_samples, and defaults to 1.0, or 1 shading call per pixel. This drives the internal tesselation depth, and takes effect after the geometry's own approximation has been calculated. It is possible to override the shading samples either per object or per instance. Finally, the shading frequency may be reduced for fast-moving objects with the -rast_motion_factor commandline option, or the corresponding string option "rast motion factor". Both versions take a positive floating point value which divides the shading samples taken, proportional to the speed of the motion of the object in screen space. The default value is 0.0, which disables this feature. A good value for many scenes would be 1.0.

Due to the different sampling patterns, it is not a good idea to use multipass rendering with passes that use the rasterizer together with passes that do not use the rasterizer.