Radiance provides a practical means for the user to move such secondary sources into the direct calculation. For example, the user may specify that a certain window is to be treated as a light source, and a separate calculation will collect samples of the transmitted radiation over all points on the window over all directions, a 4-dimensional function. This distribution is then automatically applied to the window, which is treated as a secondary light source in the final calculation. This method was used in Figure 9a not only for the windows, but also for light reflected by the ceiling. Bright solar patches on interior surfaces can make important contributions to interior illumination. Since this was the desired result of our mirrored light shelf design, we knew in advance that treating the ceiling as a secondary light source might improve the efficiency of our calculation. Using secondary light sources in this scene reduced simulation time to approximately one fifth of what it would have been to reach the same accuracy using the default sampling techniques.
Figure 9b shows a Monte Carlo path tracing calculation of the same scene as 9a, and took roughly the same amount of time to compute. The usual optimizations of sending rays to light sources (the sun in this case) and in specular directions were used. Nevertheless, the image is very noisy due to the difficulty of computing interreflection independently at each pixel. Also, locating the sun reflected in the mirrored light shelf is hopeless with naive sampling; thus the ceiling is extremely noisy and the room is not as well lit as it should be.
An important aspect of secondary light sources in Radiance is that they have a dual nature. When treated in the direct component calculation, they are merely surfaces with precalculated output distributions. Thus, they can be treated efficiently as light sources and the actual variation that may take place over their extent (e.g. the bright and dark slats of venetian blinds) will not translate into excessive variance in the calculated illumination. However, when viewed directly, they revert to their original form, showing all the appropriate detail. In our office scene example, we can still see through the window despite its treatment as a secondary light source. This is because we treat a ray coming from the eye differently, allowing it to interact with the actual window rather than seeing only a surface with a smoothed output distribution. In fact, only shadow rays see the simplified representation. Specular rays and other sampling will be carried out as if the window was not a light source at all. As is true with the computation of indirect irradiance described in section 3.2, extreme care must be exercised to avoid double-counting of light sources and other inconsistencies in the calculation.
