Collaborative Filtering: the new state-of-the-art for compiler autotuning

Share this post

Collaborative Filtering: the new state-of-the-art for compiler autotuning

Selecting the right compiler optimizations is known to have huge impacts on application performance. However, optimization options constantly increase and their effect is highly dependent on the specific program, preventing a manual approach.

This paper by Akamas Research and Politecnico di Milano was presented at the 2020 ACM International Conference on Languages, Compilers, and Tools for Embedded Systems (LCTES).

Highlights:

  • How traditional characterization techniques based on workload metrics may mislead the compiler auto-tuning task;
  • A novel methodology and algorithm to optimize compiler flags by using Collaborative Filtering techniques already validated in recommender;
  • This new autotuning algorithm outperforms previously proposed techniques, thus representing the new state-of-the-art.