SIFt: A Compiler for Streaming Applications

Due to the increasing popularity of multimedia content and wireless computing, streaming applications have become an important part of modern computing workloads. Several hardware architectures that are geared towards such applications have been proposed, but compiler support for streams has not kept pace. This thesis presents an intermediate format and set of compilation techniques for a stream-aware compiler. The prototype implementation of this compiler targets the MIT Raw architecture. Performance results suggest that certain applications perform better when compiled using these techniques than when compiled using ILP-extraction techniques alone. However, these same results show that further optimizations are necessary to realize the maximum throughput of a streaming application. Three such optimizations are outlined. Thesis Supervisor: Saman Amarasinghe Title: Assistant Professor