The decline of Moore's law and an increasing reliance on computation has led to an explosion of specialized software packages and hardware architectures. While this diversity enables unprecedented flexibility, it also requires domain-experts to learn how to customize programs to efficiently leverage the latest platform-specific API's and data structures, instead of working on their intended problem. Rather than forcing each user to bear this burden, I propose building high-level abstractions within general-purpose compilers that enable fast, portable, and composable programs to be automatically generated.
This talk will demonstrate this approach through compilers that I built for two domains: automatic differentiation and parallelism. These domains are critical to both scientific computing and machine learning, forming the basis of neural network training, uncertainty quantification, and high-performance computing. For example, a researcher hoping to incorporate their climate simulation into a machine learning model must also provide a corresponding derivative simulation. My compiler, Enzyme, automatically generates these derivatives from existing computer programs, without modifying the original application. Moreover, operating within the compiler enables Enzyme to combine differentiation with program optimization, resulting in asymptotically and empirically faster code. Looking forward, this talk will also touch on how this domain-agnostic compiler approach can be applied to new directions, including probabilistic programming.
William Moses is a Ph.D. Candidate at MIT, where he also received his M.Eng in electrical engineering and computer science (EECS) and B.S. in EECS and physics. William's research involves creating compilers and program representations that enable performance and use-case portability, thus enabling non-experts to leverage the latest in high-performance computing and ML. He is known as the lead developer of Enzyme (NeurIPS '20, SC '21, best student paper at SC '22), an automatic differentiation tool for LLVM capable of differentiating code in a variety of languages, after optimization, and for a variety of architectures and the lead developer of Polygeist (PACT '21, PPoPP '23), a polyhedral compiler and C++ frontend for MLIR. He has also worked on the Tensor Comprehensions framework for synthesizing high-performance GPU kernels of ML code, the Tapir compiler for parallel programs (best paper at PPoPP '17), and compilers that use machine learning to better optimize (AutoPhase, TransformLLVM, ProTuner). He is a recipient of the U.S. Department of Energy Computational Science Graduate Fellowship and the Karl Taylor Compton Prize, MIT's highest student award.
Faculty Host: Vikram Adve
Meeting ID: 854 1000 2462; Password: csillinois