Re: ROCm, arm64, and glibc 2.38 (Was: ci-dev.rocm.debian.net, with Ubuntu environments)
Hey Cory,
On 2024-04-11 10:01, Cordell Bloor wrote:
> Awesome! I expect ROCm on Ubuntu to generally be extremely similar to
> Debian, but those rare differences might be of great interest. Speaking
> of which, I finally tracked down why the ROCm libraries FTBFS on ubuntu
> noble for arm64.
>
> The problem began with glibc 2.38, which added SVE vector types in its
> math headers. This causes failures when compiling HIP code, as the SVE
> vector types are not defined when building for the GPU. I've filed a
> merge request on glibc in noble to block them from being defined in
> device code [1], but I'm not expecting my suggestion to be merged in
> time for release. So, I'll be filing a request to remove the ROCm arm64
> packages from the noble archives.
Great catch. I can image that it was tricky to get to the root cause.
> What does this have to do with Debian? Well, sid is still on glibc 2.37.
> The problems on Noble are an early warning for upcoming problems on
> Debian. I see that glibc 2.38-6 is available in experimental, so I
> suppose I should be filing a bug report.
One of the many drafts I have lying around is of a rebuilding service
(similar to Lucas'). It's not that complicated and that should have
caught this, too. I want to finalize Ubuntu support for workers, and
will then look at that.
Best,
Christian
Reply to: