Rust ML IG Meeting 00033

Meeting Info

Date: 20211201

Start time: 1100ET

Zoom: https://fau.zoom.us/j/67809173119

Agenda

https://github.com/Rust-GPU/Rust-CUDA
- Implicit memory hierarchy
Enzyme working on moving into rustc
linfa
- mixed typed datasets
- Naive Bayes with custom distribution parametrizations (Multinomial/Bernoulli Naive Bayes)

Participants

Manuel
Lorenz
Riccardo
Qing
Yuhan

Minutes

Rust-CUDA
- codegen for compiling Rust to PTX and runner
- wrapping existing CUDA librarys for Rust
- cuDNN offers architecture independent performance
- PTX uses unbounded number of virtual register -> can be optimized on-the-fly for specific architecture (AMD assembly is more architecture specific)
custom codegen driver
- communication overhead when establishing dynamic library support to cg_llvm
automatic GPU offloading
- very hard, even harder for asynchronous accelerator architecutres
- working for specific problems, but not applicable to real-world problems
- Polly kind supports that, but interesting for better optimization passes
Further questions about Rust-CUDA
- memory architecture -> globaly, shared, local memory how does the compiler differentiate
- launching kernels requires mutable pointers (because of ownership management of Rust) and is therefore unsafe
Linfa
- once GAT are stabilized, integrate space crate into Linfa
- approximate DBSCAN still slower than standard DBSCAN
- extended Naive Bayes depends on mixed typed datasets and generic distribution parametrization

Actions

look at rustc llvm codegen
create dummy integration of Rust-CUDA into ndarray
mixed type datasets RFC for Linfa