Re: New src:rocm-compiler for hipcc, comgr and device-libs
Hi Xuanteng,
On 2024-09-06 09:55, Xuanteng Huang wrote:
> I’ve prepared a draft of src:rocm-compiler [1] (the name still needs to be discussed) to accommodate the upstream reorganization of hipcc, comgr (compilersupport) and device-libs which are now located in the forked llvm repository.
Excellent, thanks for taking a stab at this!
> And here are some steps before we can upload this new source package:
>
> 1. Determine the name of source package
>
> There is already a removed package named rocm-llvm in Salsa [2], and I think rocm-llvm is not precise to describe the content as this package doesn’t provide LLVM. Not sure whether rocm-compiler is concrete enough.
Side note: from the looks of it, [2] should be removed. It's last change
is 4 years ago, it's never been uploaded, and someone stated in the
project information that it should be removed.
Unless someone objects within a week, I'll make a backup copy, and
remove it.
I think your reasoning for naming the package rocm-compiler is sound,
but let's see if there are other proposals.
> 2. File concflicts of binary packages
>
> Intuitively, there will be some file conflicts between the binary packages generated from the holistic new source and the original separated source packages (e.g., /usr/bin/hipcc). At least we should “rename” the binary packages and declare the conflict relations between them.
The conflict relations are indeed necessary. However, I think it's
possible to do this without needing package renames, though I can't
find an authoritative reference at the moment.
(For example, src:libfoo1 and src:libfoo2 usually both have a
bin:libfoo-dev.)
> 3. (Optional) Migrate tests
>
> There are some tests for the compilers [3], but they depend on rocminfo (and possibly other packages). We should package them with 6.1.2 version as well.
Agreed, these are pretty useful (thanks, Cory, for creating them).
In d/copyright, Files-Excluded removes some assemblies and binaries. Can
you assess the impact of this?
If this is an issue, then we might be able to keep at least the
assemblies. Some looked hand-written to me (IOW, the original source) so
they would be DFSG-compliant, though the legacy/* assemblies look generated.
I'd like to double-check some more things and run some tests, but
generally speaking, things look pretty good at first glance and the
build went fine. Hope to report back more, tomorrow.
Best,
Christian
> [1]: https://salsa.debian.org/xuantengh/rocm-compiler
> [2]: https://salsa.debian.org/rocm-team/rocm-llvm
> [3]: https://salsa.debian.org/rocm-team/rocm-hipamd/-/tree/master/debian/tests
>
Reply to: