Stream compaction for deferred shading

The GPU leverages SIMD efficiency when shading because it rasterizes a triangle at a time, running the same shader on all of its fragments. Ray tracing sacrifices this shader coherence, and the result is that SIMD units often must run different shaders simultaneously resulting in serialization. We study this problem and define a new measure called heterogeneous efficiency to measure SIMD divergence among multiple shaders of different complexities in a ray tracing application. We devise seven different algorithms for scheduling shaders onto SIMD processors to avoid divergence. In all but simply shaded scenes, we show the expense of sorting shaders pays off with better overall shading performance.