º£½ÇÉçÇø

[Seminar] Simplifying High Performance Computing with Python and Compilers

Friday, February 14, 2025

11:00 am - 12:00 pm

Speaker

Tong Zhou
Georgia Institute of Technology

Location

PGH 232

Abstract

There is a growing reliance on computational methods in scientific and engineering fields to solve complex problems in domains such as chemistry, physics, and biology. Additionally, machine learning and data science demand more computing power than ever before. However, writing high-performance code has traditionally required low-level programming, such as C++ and CUDA, combined with extensive manual optimizations. This process not only demands specialized programming skills and significant effort but also results in complex, hard-to-read code that is difficult to port across different hardware platforms.

At the same time, Python has become the dominant programming language for general software development, scientific computing, data analysis, and machine learning. Its simplicity and extensive ecosystem make it highly attractive to domain scientists, who often prioritize productivity over low-level performance tuning. However, Python’s inherent performance limitations can become a bottleneck for computationally intensive tasks.

My research aims to bridge this gap by developing compiler techniques that translate high-level programs into efficient native code, reducing the complexity of writing high-performance computing (HPC) code. This enables domain scientists and engineers to write programs in an easy-to-use high-level language, such as Python, while still achieving high performance. The compiler takes responsibility for generating optimized code for various hardware platforms. In this talk, I will present three projects: 1) APPy – A compiler that transforms annotated Python loops and tensor expressions into efficient GPU code; 2) Intrepydd – A compiler that translates Python loops and NumPy-style array operations with type annotations into optimized CPU code; 3) ReACT – A redundancy-aware compiler for efficient CPU code generation of sparse and dense tensor expressions written in index notation.

About the Speaker

Tong Zhou received his PhD in Computer Science from the Georgia Institute of Technology in 2024 under the supervision of Dr. Vivek Sarkar and Dr. Jun Shirako. His research interests include compiler optimizations and parallel computing, with a recent focus on generating efficient CPU and GPU code for matrix-oriented programs written in high-level programming models. His work has been published in well-established conferences in compilation and parallel computing, including PACT and CC (Compiler Construction). In addition to his research, he has developed several web-based tools for his compiler projects, allowing users to experiment with his techniques online.

Full Seminar Schedule