PerfoGraph+: Dynamic Program Graphs for Code Performance Optimization using Heterogeneous GNNS

Due to the surge of heterogeneous High-Performance Computing (HPC) platforms, there has been a growing need to develop performant programs. More specifically, there has been a growing demand for identifying the right device-type (e.g., CPU or GPU) and also the optimal configurations (e.g., Vectorization Factor, Interleave Count) for a specific program. Low-level program representation (e.g., LLVM-IR) has been used to tackle some code optimization tasks; however, there has not been much effort to integrate dynamic analysis information with LLVM-IR for such cases. In this work, we propose PerfoGraph+: which leverages Dynamic analysis information to construct a flow-aware Program Graph for identifying optimal configurations within a source program. The program graph is constructed using fine-grained LLVM-based Intermediate Representation (IR) along with dynamic analysis information. We evaluate PerfoGraph+ on code performance optimization tasks, including Vectorization and Interleave Factor prediction and CPU/GPU based parallelism detection. Our empirical findings suggest that PerfoGraph+ achieves around 12.3% and 50.4% better performance gain compared to other state-of-the-art models on these code optimization tasks, respectively.

Committee: Ali Jannesari (major professor), Liyi Li, and Myra Cohen