Mojo's impact on Numba

[pauljurczak]

I just took a quick read of Numba roadmap 2023 and I wonder how Numba could benefit from architectural design decisions of Mojo: GitHub - modularml/mojo: The Mojo Programming Language. Specifically, I have MLIR (https://mlir.llvm.org/) in mind. Apologies for a wishy-washy post, but I’m a new Numba user with a very limited knowledge of its internals.

[nelson2005]

This doesn’t answer your question but I found this graphic to have an interesting perspective of such projects. It looks like it doesn’t have mojo included yet.

[DrTodd13]

There is work ongoing that is attempting to use MLIR in Numba for the parallel regions of the code as identified when you have the parallel=True njit flag.

[diptorupd]

There is an exploratory Numba to MLIR integration: GitHub - numba/numba-mlir: POC work on MLIR backend. The work mostly explores replacing the post-typing lowering phase in Numba.

And as @DrTodd13 mentioned, there is an exploratory work to reimplement Numba’s loop parallelizer using MLIR.

I feel at this time Mojo sits parallel to both Cpython and Numba.

[esc]

There will also be a presentation about numba-mlir next week in the public dev meeting:

Numba Meeting June 6 2023: Presentation on Numba-MLIR

On June 6, Ivan Butygin will represent his work on Numba-MLIR: https://github.com/numba/numba-mlir Here’s a description of numba-mlir from the project README: The goal of this project is to provide efficient code generation for CPUs and GPUs using Multi-Level Intermediate Representation (MLIR) infrastructure. It uses Numba infrastructure as a frontend but have completely separate codepaths through MLIR infrastructure for low level code generation. Abstract: TBD

[sseibert]

Hi, apologies to revive this thread, but I have now posted a longer blog post on this topic that I think helps frame the comparison and give some thoughts on the relation the two projects could have:

Numba / Mojo blog post

Hi all, This has sat in my backlog for a while, but the recent announcement of a new investment round in Modular.ai (the creators of Mojo) prompted to finish off a blog post over at the Anaconda Engineering Blog talking about Numba, Mojo, and generally how I think about the goals when compiling Python: If you have any questions, feel free to ask them here!