A pipelined code mapping scheme for static data flow computers

Computers built on data Ilow principles promise efficient parallel computation limited in speed only by data dependencies in the calculation being performed. We demonstrate how the massive parallelism of array operations in numerical scientific computation programs can be effectively exploited by the fine-grain parallelism of static data flow architecture. The power of such fine-grain parallelism derives from machine-level programs that form large pipelines in which thousands of actors in hundreds of stages are executed concurrently. Each actor in the pipe is activated in a totally data-driven manner, and no explicit sequential control is needed. This thesis studies the principles of program mapping techniques that can be made to achieve high performance for numerical programs when executed on a computer based on data flow principles. A simple value-oriented language is specified as the source language to express user programs. The key of the program mapping techniques is to organize the data flow machine program graph such that array operations can be effectively pipelined. Program transformation can be performed on the basis of both the global and local data flow analysis to generate efficient pipelined data flow machine code. A pipelined code