GPU Offload in Rust: Portable, Safe, and Fast
221 points by linggen 18 hours ago | 45 comments

whateverboat 15 hours ago
> This module is under active development. Once upstream, it should allow Rust developers to run Rust code on GPUs. We aim to develop a rusty GPU programming interface, which is safe, convenient and sufficiently fast by default. This includes automatic data movement to and from the GPU, in a efficient way. We will (later) also offer more advanced, possibly unsafe, interfaces which allow a higher degree of control.

I really appreciate the work and the effort that went into this. However, such an approach has previously not really worked for C++ with LLVM offload. Why would it work for Rust?

reply
ux266478 14 hours ago
> However, such an approach has previously not really worked for C++ with LLVM offload. Why would it work for Rust?

They're very different languages, with different semantics. Without reading more than the synopsis of the paper, they're 100% leveraging the substructural type system and will have a really tight requirement for you to use a certain kind of Rust code at the CPU/GPU boundary.

reply
erupti 5 hours ago
Unfortunately, the Rust description itself is inconsistent. It claims to be "sufficiently fast by default", yet "sufficiently fast" depends entirely on the requirements of a specific user project. And then it also plans to provide options that do not guarantee memory safety when the default speed is insufficient. It is already common for Rust projects to sprinkle memory unsafe code around when performance is needed.
reply
eptcyka 35 minutes ago
I wonder if they're looking to achieve easy speedups. I can see lots of value enabling performance gains where normally people wouldn't bother because it's too much effort. I don't think this will take away work from those who hand-optimise their kernels and scheduling, this is to enable GPU acceleration for those, who otherwise wouldn't.
reply
aw1621107 2 hours ago
> It claims to be "sufficiently fast by default", yet "sufficiently fast" depends entirely on the requirements of a specific user project.

I think that's why the "by default" is there; the goal is to offer a safe/convenient API that performs well enough that by default you don't need to reach beyond said safe/convenient API. And if you happen to be in a situation where the default performance of the safe/convenient APIs is insufficient, more advanced APIs will be provided.

It's a mirror of Rust's general design goals, if anything.

reply
aw1621107 14 hours ago
> However, such an approach has previously not really worked for C++ with LLVM offload. Why would it work for Rust?

I think that will depend on the exact reason(s) C++ with LLVM offload didn't work out? If Rust differs from C++ in a way that addresses pain points/failure modes/etc. from the C++ attempt, for instance, then perhaps it isn't unreasonable to think Rust could succeed where C++ didn't (c.f., Mozilla's pre-Rust attempts to parallelize Firefox's CSS styling engine). Inversely, if Rust doesn't do things differently in the right way perhaps one might expect the effort to also not work out. Or maybe the problems are entirely non-technical and things could work out in either language.

reply
erupti 5 hours ago
> I think that will depend on the exact reason(s) C++ with LLVM offload didn't work out?

Are you making a claim or asking a question? ESL?

reply
boxed 13 hours ago
Seems to work out well for Mojo, so I'd guess it's more an issue with C++.
reply
jcelerier 13 hours ago
> However, such an approach has previously not really worked for C++ with LLVM offload

... isn't Metal shading language just C++17 compiled with LLVM ? working on every Mac and iPhone in the world is not what I would call "not really worked". Likewise, SYCL works just fine.

reply
pjmlp 4 hours ago
It is C++14.
reply
mathisfun123 12 hours ago
bruh lol this is so wrong and so confident i don't want to even attempt to explain how wrong you are.

> just C++17 compiled with LLVM

i invite you to attempt to compile/run absolutely any C++17 codebase on your iphone's GPU lol!

reply
pjmlp 5 hours ago
For starters Metal Shading Language is based on C++14, so naturally any C++17 will fail.
reply
nish__ 11 hours ago
Wouldn't any metal app have to be written in C++17?
reply
jcelerier 11 hours ago
I mean "any C++ codebase" doesn't make sense in general. I run C++23 code on ESP32, that doesn't mean I'm gonna build KDE or chrome for it, and that doesn't make it any less C++.
reply
winningChild 13 hours ago
[dead]
reply
winningChild 13 hours ago
[dead]
reply
bicepjai 13 hours ago
I write all my code in Rust because I am a Rustacean. In many of my custom LLM inference engine projects, the biggest fight has always been bindings. I don’t want to maintain and write bindings; also, if I use an existing project that provides bindings, then I have to wait for the owner to update or fork it and then maintain it on top. It has been a big headache. Running Rust core on GPU sounds like something I will try from day one. Kudos to the team and will watch it closely.
reply
YuechenLi 14 hours ago
So... why go through LLVM at all instead of having the MIR target PTX/HIP C directly then?

If they really wanted a vendor neutral solution for Rust GPU, that already exists: you write the CPU side code, including buffering, allocation, concurrency, etc through Vulkan binding and consume the compute kernel in SPIR-V from HLSL/GLSL/WGSL etc. As it stands, the way they use Rust here feels more like using it like TypeScript types/interfaces than anything else.

Again, the size of most operations that should be done on the GPU is known ahead of time before compilation, so it's very much possible to statically allocate memory at compile time instead of going through all this trouble to write what's essentially a Rust shaped DSL for GPU compute.

reply
minraws 11 hours ago
Because it's convenient? Shader and vulkan semantics can be quite limiting and annoying to write.

Maybe it doesn't matter in a post AI world but perhaps it will allow better abstractions.

No need to yuck someone else's yum.

reply
gmueckl 4 hours ago
The compute shader side of Vulkan is actually fairly smooth compared to graphics. It's not really that different from low level CUDA or OpenCL. The Vulkan complexities overwhelmingly concern rasterization and raytracing.
reply
sanxiyn 13 hours ago
People go through trouble to write Python-shaped DSL for GPU compute. We will go "why o why?", but apparently such things are necessary to succeed in the market.
reply
YuechenLi 13 hours ago
Well, I suppose fake Python is better than fake C++ at least.

But yeah, I think Python's dominance in science/ML will eventually pass, just as FORTRAN and Matlab did before.

reply
Driftbench 3 hours ago
Ownership tracking should map well to GPU memory lifetimes. That's one place Rust has a real edge over C++.
reply
Thomashuet 16 hours ago
That's promising but did they publish any code? I can't find anything in the abstract.
reply
boywitharupee 15 hours ago
is this mainly about making host binaries self-contained for heterogenous workloads?

also, seems like this is mostly targeted towards HPC audience?

reply
maxchisto 16 hours ago
does anyone know Mojo well enough to comment how Rust + gpu-offload compares to it?
reply
giancarlostoro 15 hours ago
Mojo is not fully open sourced yet, but it will eventually be, would be an interesting comparison though.
reply
maxchisto 15 hours ago
Mojo's OSS status doesn't prevent us from evaluating its memory model, writing and benchmarking kernels in it, etc
reply
giancarlostoro 15 hours ago
Sure, and I realized after I posted the std lib is opened up, not sure how much of it will reveal the underlying Mojo specifics though.
reply
jasonjmcghee 16 hours ago
> the rust-gpu project has to emulate pointers[8], which we consider a blocking issue for most HPC benchmarks.

Why is it a blocking issue?

I feel like this is very aligned with the goals of rust-gpu.

reply
minraws 15 hours ago
Pointers are sort of needed for high performance memory management for HPC targets for existing design patterns, maybe we can think of better solutions down the line but it's hard for me to say anything I just use/abuse CUDA pointers as well.
reply
adgjlsfhk1 15 hours ago
Julia has pretty good design heritage for how to deal with this sort of thing. you build the right abstractions and everything works (the main key is making sure the compiler elides bounds checks)
reply
andreypk 2 hours ago
[dead]
reply
Arsen-V 3 hours ago
[flagged]
reply
Alephinitesimal 15 hours ago
The NVIDIA+AMD support is the part I find really interesting. I know OpenMP and SYCL can already target multiple GPU vendors, but doing this while keeping Rust's safety model seems pretty compelling. I'm curious how portable the performance is in practice.
reply
jheriko 16 hours ago
[dead]
reply
iberator 17 hours ago
[flagged]
reply
aabhay 17 hours ago
Genuinely curious what these constant changes are. In fact I don’t feel they’re moving fast enough to give us improvements to core.
reply
frollogaston 16 hours ago
It was changing quickly in the earliest releases, about 10 years ago. Like one day I pulled our repo and there was new "?" syntax, but that was a feature I'd been wanting anyway.

Edit: Oh, async/await was a bigger and more recent one, 2019. I've heard that this wasn't an easy decision for them but was kinda needed.

reply
aabhay 16 hours ago
A few years in between releases with a sane versioning system seems fine to me. If you don’t want new features don’t build with that new toolchain.

It’s not javascript where you need to support all possible browsers.

reply
frollogaston 16 hours ago
Teams will disagree over what toolchain to use, and you will read others' code, so this doesn't dodge the issue. Otherwise there'd be no complaint about C++. I don't think Rust is bloated though, every feature has a very good reason.
reply
speedstyle 11 hours ago
Would they disagree, new toolchains work great with old code? And even add 95% of the features to older editions, just not new keywords or inference defaults. I think in C++ people complain about the effort to migrate --std, rather than having to learn variants or concepts. It does add to the complexity of course, and it's useful to agree on a consistent style, I just mean Rust doesn't have the particular issues with everyone having to migrate in lockstep, or anyone needing to for new tools
reply
frollogaston 11 hours ago
Same with C++, new standards will support old code. But they'll disagree on using the new features in the new toolchains. Or more likely, they'll just use the new features. So there's no option of sticking to old feature sets, you will encounter the new ones at the very least.

Even with a style guide in C++... Google is known for having one. I was forced to change how I use basic strings multiple times there. They also changed the rules around refs (&).

reply
jjice 17 hours ago
That's an opinion. One you're very welcome to hold, but it's not universal by any means. Lots of people enjoy rust.
reply
rfgplk 16 hours ago
Fascinating how many people still overcomplicate offloading to GPUs.
reply
MBCook 15 hours ago
How so? I don’t know anything about this area.
reply
konradha 15 hours ago
What's the easy way here? Linking CUDA into your Rust binary?
reply
binsquare 14 hours ago
:)

This might be an relevant read: https://smolmachines.com/engineering/gpu-over-vsock

reply
dc443 13 hours ago
I just read this tool up and down but still can't conceive of a use case.
reply
foltik 6 hours ago
No, it’s not relevant to simple GPU offload, stop spam linking your slop blog.
reply