Reptile : a distributed ILP compiler

Abstract

The past few years witnessed a dramatic shift in computer microprocessor design. Rather than continue with the traditional pursuit of increased sequential program performance, industry and academia alike chose to focus on distributed, multi-core designs. If multi-core designs are to maintain the decades-long trend of increased single threaded performance, compiler technology capable of converting a single threaded program into multiple programs must be developed. In this thesis I present the Raw Explicitly Parallel Tile Compiler (Reptile), a compiler targeting the RAW computer architecture capable of converting a single threaded program into multiple threads communicating at the instruction operand granularity. On applications with sufficient amounts of parallelism Reptile has generated code which, on the Raw processor, achieves a speedup of as much as 2.3x (cycle to cycle) over an Athlon64.