unique: Designing a safe buffer writing API is hard

[Firestar99]

Blog link: https://github.com/Rust-GPU/rust-gpu.github.io/blob/unique_index/blog/2026-02-23-safe-buffer-writing.mdx

I just wanted to dump this out of my head, may make for a decent blog post.

May need some more refining towards the end too.

[fluffysquirrels]

I really like this topic (data-race-free access) and I see the way you're trying to approach it "bottom-up" from initially guaranteed unique values, proven using the type system. It's a shame you found a flaw in this, but I think some ingredients from this idea will still be useful.

As I said on the other thread, I was inspired by rayon, its scope API, and its parallel collections API. I think writing up some of my thoughts as a blog post (as you did) might clarify this. OK to make a PR on this repo?

[Firestar99]

My concern is that a rayon-like API is better suited at a graphics API level on top of wgpu / vulkan, so I'm not quite sure what you want to do. Always feel free to PR stuff (or to send links to commits)

[fluffysquirrels]

Ah, I haven't seen enough graphics code to propose something there. I can do more research on that.

OK. I think it would help if I wrote up a post and some examples I was thinking of for compute.

Until then, I was inspired by some existing NVIDIA on-GPU utility libraries for CUDA C++. I assume they are mainly for compute applications, but I suspect some modern non-fixed function graphics may also use them too. It would be great for rust-gpu and Rust to have something similar available (in-tree or elsewhere).

CUB (docs) seems like the leading option. Under "API documentation" they have utilities that work at sub-group (CUDA warp), workgroup (CUDA block), and device levels of abstraction.

The algorithms available are classic parallel collection primitives: load, shuffle, store, merge, sort, reduce, scan, histogram, etc.

moderngpu (src) is similar to CUB but has some different algorithms and a nice approach to composing its algorithms. It also has an excellent tutorial book explaining its implementation as of its v1.0 release.

Thrust (docs) seems older, supports multiple backends (several on CPU, or CUDA), has similar algorithms to CUB.

cuCollections (docs) has a few data structures -- various hash tables, some bitmaps -- but seems young and less capable for now.