rapidsai / rapids-cmake

https://docs.rapids.ai/api/rapids-cmake/stable/
Apache License 2.0
28 stars 45 forks source link

Add workaround for gcc bug in cuco #685

Closed miscco closed 2 weeks ago

miscco commented 3 weeks ago

We discovered an unfortunate bug in the gcc constexpr evaluator that broke the alignment calculations in cuco.

There is a simple workaround so employ that to unbreak rapids with cccl ToT

bdice commented 2 weeks ago

@miscco asked me offline about the testing procedure we need for this PR. I pointed to the downstream CI testing we typically do here: https://github.com/rapidsai/rapids-cmake/pull/677#issuecomment-2294110382

Here is the comparison between the previous and new commits of cuCollections: https://github.com/NVIDIA/cuCollections/compare/abc5095b011feccd7d3ecb36b369741b96ae6f0e..d3477661d771e0d6fd22259bf6dd6f8c64a7401c

The diff is pretty small and confined, so I would be okay merging this PR without downstream CI testing if it saves time. @miscco verified that cuDF builds with this change. But I'd want to get an okay from @PointKernel and/or @robertmaynard before merging.

PointKernel commented 2 weeks ago

Thanks for bringing this up.

I reviewed the changes yesterday, and as @bdice mentioned, they are minimal. I don't see any need to retest the RAPIDS suite for such a minor update.

robertmaynard commented 2 weeks ago

Will merge early tomorrow :+1:

robertmaynard commented 2 weeks ago

/merge