[LLVM-DEV'23] ML-Guided Compiler Optimization in LLVM Workshop

Community US Developer Meeting
3

Here are the summaries - thanks @AmirAshouri @boomanaiden154-1 @ChrisCummins @jdoerfert @ondrasej !

LLM Table

  1. How powerful LLMs can be in the domain of compiler optimization, i.e., code optimization, phase ordering of passes, etc.

  2. Potential ideas for future LLMs applications:

  • Construction of a better heuristics for Register Allocation?
  • Leveraging Meta’s pretrained LLM model in compiler domain

  1. LLVM and the vast number of available knobs under optimization O3

  2. Inherent shortcomings of LLVM with Phase Ordering of passes on breaking the code

  3. […] ideas for how LLMs could be useful for compiler writers, such as as a large scale pattern matching tool to identify possible optimization opportunities (supplementing the role of the profiter), and we talked about the ingredients needed to replicate and expand on LLM work: a large dataset, base models, and compute.

Latency Prediction Table

  • Dynamic Causal Modelling & modeling a “generic”, parameterized CPU, and learning the parameters. PMEvo was mentioned; also using bayesian inference
  • discovering internal resource usage by adding one instruction at a time and measuring
  • predicting a probability distribution instead of a single number
  • the scalability (or not) of creating simulators by hand
  • these are it’s in my notes, but don’t remember the specifics, if someone does, please update:
    • cost modeling “per app”
    • cost modeling for future chips

Datasets Table
We discussed mostly input capturing and generating to make IR datasets executable, with different granularity

There was one more round table, but I haven’t gotten any signal - please do, and I’ll update.

Thanks!

1 Like