Compiling dataflow graphs into hardware

Department Head: L. Darrell Whitley.2005 Fall.Includes bibliographical references (pages 121-126).Conventional computers are programmed by supplying a sequence of instructions that perform the desired task. A reconfigurable processor is "programmed" by specifying the interconnections between hardware components, thereby creating a "hardwired" system to do the particular task. For some applications such as image processing, reconfigurable processors can produce dramatic execution speedups. However, programming a reconfigurable processor is essentially a hardware design discipline, making programming difficult for application programmers who are only familiar with software design techniques. To bridge this gap, a programming language, called SA-C (Single Assignment C, pronounced "sassy"), has been designed for programming reconfigurable processors. The process involves two main steps - first, the SA-C compiler analyzes the input source code and produces a hardware-independent intermediate representation of the program, called a dataflow graph (DFG). Secondly, this DFG is combined with hardware-specific information to create the final configuration. This dissertation describes the design and implementation of a system that performs the DFG to hardware translation. The DFG is broken up into three sections: the data generators, the inner loop body, and the data collectors. The second of these, the inner loop body, is used to create a computational structure that is unique for each program. The other two sections are implemented by using prebuilt modules, parameterized for the particular problem. Finally, a "glue module" is created to connect the various pieces into a complete interconnection specification. The dissertation also explores optimizations that can be applied while processing the DFG, to improve performance. A technique for pipelining the inner loop body is described that uses an estimation tool for the propagation delay of the nodes within the dataflow graph. A scheme is also described that identifies subgraphs with the dataflow graph that can be replaced with lookup tables. The lookup tables provide a faster implementation than random logic in some instances