A real-time procedural shading system for programmable graphics hardware

Reviewer: Howard E. Whitston

The authors describe their shading system, which allows the user to control the underlying hardware for generating shading without having to directly manipulate OpenGL. OpenGL allows individuals absolute control over the rendered image, but also requires users to understand complex and unforgiving assembler language, where minor mistakes may have substantial consequences in the final image. RenderMan Shading Language is a popular way to avoid directly coding OpenGL, and the authors compare their system to this. The system consists of the following components: A shading language and compiler front end, which allows the user to describe the final image in terms of lights and textures in an HLL-like manner Programmable Pipeline abstraction, which hides the hardware component from the user A retargetable compiler back end, which allows the system to be extended as new hardware becomes available Shadow Object code, which provides a stage between the compiler front end and the rendering parts of the system A shadow execution engine, which controls rendering of primitives Graphics hardware, consisting of the graphics and host CPUs, which can each perform computations as determined by the system The pipeline abstraction description covers operations such as vertices or fragment rendering, built-in datatypes handled, and determination of which operators are permitted, which simplifies the compiler front end and allows for better definition of pipelined/parallel graphics hardware. The shading language section starts by discussing the main differences between RenderMan and the authors’ system, including the reasons why the authors selected (or didn’t select) certain features. The language analysis section points out the features of the authors’ system that aid in optimizing the resulting code: the underlying graph relationships, how surface/light shaders are combined in the system and when (programmable OpenGL APIs don’t directly support separate surfaces and lights), and inclusion of support for computation frequencies being able to infer rules (constant, primitive group, vertex, and fragment) in terms of lights and shaders. The retargetable compiler back end section discusses the modular compiler and the desired interactions in terms of number of allowable passes through the code; data flow; resource constraints; host-side compilers (two in C, and two in assembler with primitive and vertex computations); vertex programming architecture with proposed designs; fragment back end and its associated tree-matching and graphs problems/solutions (some of this is still in the research stage); and programmable fragment hardware and its register combiner architecture, mapping the graph operations to a single rendering pass (efficiency of more complex mappings requires additional research). The final section shows the results of the authors’ work. The authors demonstrate their system using a variety of standard images, and discuss where the system is better than RenderMan, and where it needs additional work. The shading language has a tradeoff between vertex and fragment (to be expected) and the effect of combining surface and light shaders. The comparison of vertex-program back end output and hand-written analysis showed that this computer-generated method could be as efficient, although more examples would be needed to prove this point. The fragment back end was compared to hand-generated output, and did not do as well as the authors would have liked (although the register architecture proved useful). The discussion concludes with the actual images and animations rendered. The last part discusses future work. Three things about this paper stand out. The overall enthusiasm of the authors for their system comes through clearly in their writing. The figures and diagrams included by the authors also clarify and enhance the content. However, some terms used by the authors are unfortunately not defined. This negative is the only downside of an otherwise excellent paper. Online Computing Reviews Service