Biography
Currently, I am a member of NVIDIA’s CUDA compiler team. Before NVIDIA, I was an AI Compiler Engineer at Microsoft, where I am also involved in the MAIA project, working on improving the compiler’s performance for the MAIA 100 and 200 chips.
I earned a Ph.D. in Computer Science from Georgia Institute of Technology, where I conducted research under the guidance of Prof. Vivek Sarkar. Prior to my Ph.D., I completed a Master’s degree in Computer Science at Rice University in 2017.
My PhD research is primarily centered around ensuring the correctness and enhancing the performance of accelerated applications. Most of my work is related to OpenMP and the LLVM compiler. I worked closely with researchers from national laboratories and actively engaged with the LLVM and OpenMP communities, while also collaborating with colleagues from various universities.
- High Performance Computing
- Program Analysis
- Debugging
-
Ph.D. in Computer Science, 2024
Georgia Institute of Technology
-
Master in Computer Science, 2017
Rice University
-
Master in Computer Science, 2015
Xi’an Jiaotong University
-
Bachelor in Software Engineering, 2012
Xi’an Jiaotong University
Selected Publications
Patents
Experience
- Worked on AI compilers for MAIA 100 and 200 chips.
- Led a side project, Triton-San, which explores feasible approches of detecting concurrency issues in Triton programs.
- Contributed a few PRs to the Triton and Triton-Shared projects.
- CS 6241 Compiler Design
- Graduate-level compiler course focusing on SSA IR, control & dataflow analysis, and dependence test
- Redesigned all projects using LLVM 15
- Helped the application team to solve OpenMP-related issues, including performance issues and concurrency bugs
- Host: Annemarie Southwell, Simone Atzeni
- Designed the implementation strategy of OpenMP Tool Interface (OMPT) for NVIDIA’s OpenMP runtime
- Reported a number of OMPT design issues to the OpenMP community
- Host: Lei Tian, Adnan Aziz
- Participated in the development of a performance analysis tool for PyTorch applications
- Designed a number of new features for the tool and improved the tool’s performance by 14.2 times
- The optimized performance analysis tool, including all new features, has been deployed in Meta
- Host: Joachim Protze, Christian Terboven
- Developed ARBALEST, a dynamic data mapping issue detector for OpenMP applications
- Host: Joachim Protze, Christian Terboven
- Designed and implemented DataRaceOnAccelerator (DRACC) micro-benchmark suite
- Examined the implementation of OpenMP target offloading features in LLVM 8.0
- Reported a number of bugs and inconsistencies to the LLVM community
- Host: Oscar Hernandez
- Collaborated with application developers at ORNL to find out potential concurrency bugs in OpenACC applications
- Host: Oscar Hernandez
- Explored concurrency bug detection algorithms and tools for GPU applications
- Analyzed Julia’s compiler to figure out the lowering process from Julia AST to LLVM IR
- Advisor: Prof. Vivek Sarkar
- Master’s Thesis: Data Race Detection for Event-Driven Parallel Runtime Systems