Pre-LLVM-DEV'23 - ML-Guided Compiler Optimization Workshop

Trying to gauge if there’s interest in having a workshop on ML-Guided Optimizations (see Tanya’s post about workshops).

The format would be roughly the same as before’s BoF / Panel, i.e. presentations on what folks have been working on, followed by discussion. I’d hope that the workshop format would give us more time for both!

Off the top of my head, some topics of interest:

  • latency modeling, ideally for industrial workloads (aided/not aided by ML, but definitely interesting to ML training)
  • using ML to speed up compilation - not only pass level (ongoing GSoC work), but maybe also within a pass (like guide where deeper analytical, costly analysis should even happen)
  • new optimizations using ML
  • diagnosability / maintainability (when living with learned policies)
  • compiler support / infrastructure

If interested, please reply to this thread, and please feel free to also add or nuance areas you may be interested in either presenting or seeing discussed!



Hi Mircea-

I am interested though I have no new work to show case. So I can help with the workshop in general I suppose.



I would like to learn more about ML-guided optimization.

1 Like

+1, I’ll join :slight_smile:

1 Like

I’d be interested in joining. Whether I’ll be able to travel is a different question though :slight_smile:

1 Like

Same for me. Happy to help organize it, nothing new to show atm.


1 Like

This is a great initiative! I’m fully supportive of course.

This workshop would be a great addition! I’d be happy to join and present our work.


would love to join! I’d like to see any ongoing work on latency modeling, for statically comparing two IRs/Object files, to avoid having to run the app on each training epoch. Thanks.

This year I don’t have the capacity to join… but just throwing out a proposal:
maybe invite the authors of the recent DeepMind Nature paper that updated the LLVM sorting library: Faster sorting algorithms discovered using deep reinforcement learning | Nature

1 Like