Benjamin R. Willis Email & Phone Number

My research improves the accessibility of domain-specific computers (like AI hardware and ML accelerators) to the broader software engineering community by automatically structuring computer programs for portable performance.Portable performance allows computer programs to achieve high performance and efficiency on any platform by exploiting the architectural customizations (like GPUs and TPUs) of that platform.Portable performance requires a computer program to be structured as a "produce-consume task graph". Produce-consume task graphs work like an assembly line. Each stage in the assembly line is clearly delineated, and its communication with the next assembly line stage (the "consumer") is explicit. Unfortunately, produce-consume task graphs are not widely available in the open-source software world (called "Code From the Wild", or CFTW).Structuring computer programs for portable performance is difficult. Current solutions require application and architecture experts to structure new applications or refactor existing ones manually. Domain experts are in short supply, and structuring applications is time-consuming. Further, manual design efforts don't scale to novel algorithms and architectures. An automated solution to the portable performance problem is desperately needed. To address this challenge, I built Cyclebite - an open-source project that automatically extracts the produce-consume task graph from CFTW. Cyclebite uses dynamic profiles to localize the tasks of the application and observe their communication with each other. Then, the extracted task graph is exported to Cyclebite-Template, which finds optimization opportunities in the task graph. Finally, Cyclebite-Template exports an optimized implementation of the target application to domain-specific hardware.To develop and test Cyclebite and Cyclebite-Template, I built Dash-Corpus: a collection of over 2,500 C/C++ applications from 25 different open-source libraries. Using Dash-Corpus, I show that Cyclebite extracts the structure from CFTW automatically. Cyclebite-Template's exported task graphs meet or exceed the performance of expert-written applications.Cyclebite has the potential to alleviate major engineering bottlenecks in the pursuit of faster and more efficient programs for AI, scientific computing, signal and image processing, and many other forms of computation. If it achieves its goals, it will change how programmers write programs, and unlock a large swath of information that, along with generative AI, will further automate the architectural design process.