2023 LLVM Dev Mtg - Mojo 🔥: A system programming language for heterogenous computing

I was watching a recent talk about the new Mojo language [1]. There is a section on SIMD and how they treat scalar values as a special case of vector operations (around the 33 min time). It does seem that tensors are one of the core abstractions for modern ML systems. I've heard people joke that AI is just matrix multiplication. Python is such a flexible language that creating abstractions around its syntax was super easy. I believe that was part of the reason Python has come to dominate this space (not the only reason obviously). I too felt the same as you, but as a distant admirer of Zig. I totally understand that operator overloading can be misused. But I have an intuition that at least providing operators for linear algebra (and probably complex/quaternion) is a really important feature for any languages from this point in history going forward. 1. [VIDEO]

When mojo was initially released I didn’t think much of it because: 1. It was announced prior to being available, even in closed source form. 2. Emoji filenames meant to me that it wasn’t taking itself too seriously. 3. Many people have tried to make Python fast and failed, and that was my naive take on it, and despite the fact that LLVM allstars were working on it, it seemed like a lofty goal. 4. Closed source programming languages usually suck. They presented at the LLVM conference in 2023 and this video totally changed my view: [VIDEO] There are some fascinating ideas here, I’m especially excited about what could come from their templated IR. Recommend others give it a watch if they’re confused on why they should pay attention to this project.

They've given a talk abut this on the LLVM dev meeting: [VIDEO] They have some fancy tech around compile-time interpreter, mojopkg, and their code generator. They also have a graph compiler (not mentioned in the talk) that can fuse kernels written in Mojo.

I think they had a good dive here [VIDEO] The gist is that MLIR (unlike LLVM) was designed to be multithreaded so the whole stack above LLVM IR is parallel, and then LLVM is used as a per-function codegen tool in parallel, so although LLVM itself is single threaded there are multiple seperate contexts of it. also I think they wrote their own linker.