Update compilation targets #319
Open
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Remove and add some `NvvmArch` variants. Remove `compute_35` and `compute_37`, which are no longer needed/supported now that CUDA 11.x support is gone (Rust-GPU#312). - Add `compute_73`, which is supported in CUDA 12.0-12.8 but was never added. - Add `compute_88` and `compute110{,f,a}`, which are new in CUDA 13.0.
CUDA C++ has the `__CUDA_ARCH__` macro for conditional compilation. rust-cuda has a `CUDA_ARCH` environment variable that is similar, and the `from_cuda_arch_env` method parses the environment variable's value to produce a value of type `ComputeCapability`, which can be queried for conditional compilation. But `ComputeCapability` has a big problem. It's missing all the capabilities after 80, including the 'a' and 'f' suffix ones. We could just add them, but it implements `PartialOrd`/`Ord` and uses ordering to determine feature availability. This was valid before the 'a' and 'f' suffixes were added but is no longer, because some pairs of values are incomparable. E.g. `100a` and `101a` -- each one has some features the other doesn't, so neither is clearly larger than the other, and they're also not equal. So, what to do? Well, `CUDA_ARCH` was added in 2022. More recently, another mechanism for conditional compilation was added: `target_feature`, in Rust-GPU#239. This does work with the 'a' and 'f' suffix targets, and it's more Rust-y. So this commit just removes `CUDA_ARCH` and `ComputeCapability` (removing two more places where the default compilation target is specified) and changes the only uses (in `cuda_std/src/atomic/mid.rs`) to use `target_feature` instead. We don't have any tests exercising conditional compilation, alas, but I did some manual checking locally to verify that it works the same.
It includes some now-unsupported targets and is also missing some new targets. The obvious thing to do is update it, but it's simpler and better to recognize that it's identical to the generated `driver_sys::CUjit_target` type and instead use that generated type directly, avoiding the need for manual updating in the future. Especially given that there is a non-trivial encoding for targets with 'a' and 'f' suffixes (which involves adding 2^16 and 2^17, respectively, to the base number). This seems fine because this `ModuleJitOption` type is obscure and has no existing uses in the codebase.
nnethercote
force-pushed
the
update-compilation-targets
branch
from
bba43c3 to
9cd54eb
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
With CUDA 11.x support dropped (#312) and CUDA 13.0 support added (#299), the supported compilation targets need updating, with some additions and removals. There are three different types that encode compilation targets.
NvvmArch: this is the main one. The first commit updates it.ComputeCapability: this is an auxiliary one used only for theCUDA_ARCHenvironment variable, used for conditional compilation. The second commit removes it becausecfgprovides a better way of doing conditional compilation.JitTarget: this is an auxiliary one used only for the obscureModuleJitOptions. It can be removed because it's identical to a generated type incust_raw.Details in the individual commit messages.