Chris Metcalf
Laboratory for Computer Science
Room NE43-635
The parallel prefix routine is a useful one to be able to invoke. Particularly in data-parallel models, inventively-used parallel prefix operations can make up substantial parts of programs. The NuMesh is able to run parallel prefix at high speeds by using a short loop in the communications FSM to handle all the interconnections among the processors during the execution. The FSMs safely but asynchronously switch phase during execution of the program, running a different loop of code. Implementations of various kinds of prefix are given, including Boolean operators, multi-word addition, and user-defined functions.