Is it possible to provide a LLVM independent build?

[Neutron3529]

I tried to publish another cuda crate without llvm or specific rust compiler support:

https://github.com/Neutron3529/cuda_min

Such crate works very well, and thus I have some doubt: why a specific llvm version and rust version is needed?

It seems that

1. llvm-bitcode-linker could compile rust code into PTX
2. nvidia driver could load PTX code directly. (no need to install cudart, nvidia's driver provides interfaces such as cuInit.)

I just wonder why a specific llvm version even a specific rustc version is required. We should have the ability to compile a cuda program easier.

[LegNeato]

@brandonros did some experiments with emitting llvm ir and then using llvm-bitcode-linker I believe. Care to comment on if this path is viable?

[brandonros]

@LegNeato I did not feel there was a lot of good warm than reception on that approach.

This repo is really cool because you are taking over somebody's extremely hard to do rustc_codegen approach from LLVMv7

Is there any appetite to consider alternatives?

I get that Blackwell is very early and a very small piece of the market, but if we are forward looking, I would guess that within the next two years, the majority of the market will be using this new LLVM v19 version of NVVM/etc

I would like to sync with you on this, I am sure you were busy, but I basically came up with two different approaches

One of them is the one that I PR'd where I fumbled through making a new rustc_codegen

Another is where I give up a ton of flexibility related to address space but I was able to do LLVM IR -> NVVM which I would really like to show you and get some feedback on whenever you have time

[LegNeato]

I wonder if we can statically link it in, so at least users don't need it installed.

[LegNeato]

@brandonros I took a look at your approach, my main concern is I don't understand the tradeoffs. I also expect that would be harder to get into the compiler proper. But again, I don't really know. I would imagine if we are in rustc we can distribute nvvm or compiled llvm7 in its own component and we'd have to probably do the same for LLVM v19 as I don't think nvidia is going to track LLVM versions with NVVM like rustc does (but who knows!). So this llvm stuff is only a problem because we don't have all the rustup / rustc infra as an external codegen backend.

[LegNeato]

What I dont want to do is reimplement most things that NVIDIA handles. That's a fools errand...we want to swim with them rather than try to reimplement everything (minus some bits that maybe are a pain or need to be rust-native).