[Improvement]: add a first GPU-targetting algorithm

[imrn99]

Scope

honeycomb-kernels (algorithms)

Side effects

unsure

Type of change

new feature

Affected item(s)

• new module in the kernel crate, probably
• new dependency (see below)

Changes / proposals

I came across a few promising crates that allow us to write GPU kernels in Rust. Given a simple algorithm, we can probably write a first GPU kernel using those.

GPGPU in Rust can be the object of a proper SOTA study later, starting from @dssgabriel's work and working our way up to today, especially since a lot of things changed since the initial work. As of now, I identified two fitting dependencies for us:

• cudarc - call CUDA kernels from a safe rust interface; we can handle data transform & transfer to the device in our crate easily
• cubecl - a rust DSL based solution; I need to look into it further, but this seems like a very promising implementation (cross-platform too!)

I kept those because they both:

• add no/minimal unsafe to the rust codebase
• do not require writing the kernels in shading languages

[imrn99]

Proof of concept repository: https://github.com/imrn99/poc-honeycomb-gpu

[imrn99]

Other kernel repositories:

• https://github.com/baldraven/overlay-grid-honeycomb
• https://github.com/baldraven/polyhedral-parallel-mesher