AICOS: AI-based Compiler Optimization Skipping

PI Name Johannes Doerfert
PI Institution Argonne National Laboratory
Collaborating ANL Division Mathematics and Computer Science (MCS)
Project Description

During the Google Summer of Code (GSoC) 2020, we developed a prototype for a pass result prediction framework: “Advanced Heuristics for Ordering Compiler Optimization Passes”. The key insight is that we can skip compiler analyses and optimization passes to safe time if they are likely not to change the outcome. Our preliminary results look very promising [0] and our lighting talk at the LLVM Developers meeting [1] was well received. We now want to use the benchmarks (SPEC) and compute power (GPUs) on JLSE to expand our experiments towards a publication.

[0] https://github.com/uenoku/llvm-project/tree/gsoc-2020-submission-pass-p…
[1] https://docs.google.com/presentation/d/1i7l5AYXH5u9pTmYGVFTvZD5fC5MOuyV…

Testbed

Witherspoon, GPU_V100_SMX2