High level compilation for gate reconfigurable architectures

Abstract

A continuing exponential increase in the number of programmable elements is turning management of gate-reconfigurable architectures as "glue logic" into an intractable problem; it is past time to raise this abstraction level. The physical hardware in gate-reconfigurable architectures is all low level - individual wires, bit-level functions, and single bit registers - hence one should look to the fetch-decode-execute machinery of traditional computers for higher level abstractions. Ordinary computers have machine-level architectural mechanisms that interpret instructions - instructions that are generated by a high-level compiler. Efficiently moving up to the next abstraction level requires leveraging these mechanisms without introducing the overhead of machine-level interpretation. In this dissertation, I solve this fundamental problem by specializing architectural mechanisms with respect to input programs. This solution is the key to efficient compilation of high-level programs to gate reconfigurable architectures. My approach to specialization includes several novel techniques. I develop, with others, extensive bitwidth analyses that apply to registers, pointers, and arrays. I use pointer analysis and memory disambiguation to target devices with blocks of embedded memory. My approach to memory parallelization generates a spatial hierarchy that enables easier-to-synthesize logic state machines with smaller circuits and no long wires.

(cont.) My space-time scheduling approach integrates the techniques of high-level synthesis with the static routing concepts developed for single-chip multiprocessors. Using DeepC, a prototype compiler demonstrating my thesis, I compile a new benchmark suite to Xilinx Virtex FPGAs. Resulting performance is comparable to a custom MIPS processor, with smaller area (40 percent on average), higher evaluation speeds (2.4x), and lower energy (18x) and energy-delay (45x). Specialization of advanced mechanisms results in additional speedup, scaling with hardware area, at the expense of power. For comparison, I also target IBM's standard cell SA-27E process and the RAW microprocessor. Results include sensitivity analysis to the different mechanisms specialized and a grand comparison between alternate targets.