The Fifth Elephant 2024 Annual Conference (12th &13th July)

Abstract:

Compilers are workhorses of performance behind all AI algorithms. Making algorithms work effectively on GPUs is especially hard - called kernel programming. Compiler ecosystem around GPUs is especially messed up. Compilers are supposed to allow for performance portability of different hardwares but this is usually not the case. See below infographic for current state of AI compilers.

I am designing a compiler framework called llama.lisp. As you can tell from the name, the framework is highly inspired by Lisp. Limitations of LLVM are examined critically to improve the framework. A multi layered approach is adopted to tame the complexity of writing such an compiler. No single language is chosen as implementation language to allow for the right tools can be used for right problem. A message oriented architecture is used for foreign function interfacing instead of traditional C based API.

In this talk, I will go through design choices of this compiler framework.

Additional Information:

Audience for the session

Folks who are interested in GPU programming, systems programming. Their roles might currently be called as Data engineer, AI engineer or systems engineer. Experience with high performance computing or array programming would make you appreciate the talk better.

Scope of the session

We’ll go through high level design decisions without digging too much into details. We might look at some scary assembly code but don’t expect you to read it.

Takeaways from the sessions

By the end of this talk, you will have a good idea of computer architectures of AI and how programming languages are implemented.

Format of the session

30 mins talk with 10/15 mins QnA