Hello, Le 19/10/2025 à 14:01, Christian Kastner a écrit :
Hi Cory, sounds reasonable to me, with one comment (see below). How do you plan to test this? In case we both missed something. Breaking up circular dependencies and renames always have the potential to break something. With testing, I mean install-time, as in user A has a trixie or forky system, and runs apt-get upgrade, possibly not even on a ROCm component but something that uses one. On 2025-10-17 21:52, Cordell Bloor wrote:5. Propose src:llvm-toolchain-rocm with the full AMD LLVM fork (i.e., amdclang++). It would provide hipcc-rocm, amd-device-libs-rocm, clang- rocm, etc. If we decide to switch to it, we can update src:rocm-llvm- defaults to make the change.I find the -rocm suffix is pretty confusing; hipcc-rocm sounds tautological. Sadly, I don't have any better ideas.
amdhipcc would maybe have some logic with amdclang ? I have also some questions on the good way to declare compilers I used export CXX=clang++ export HIPCXX=clang++ export CC=clang export PATH:=/usr/lib/llvm-21/bin:$(PATH) for composable-kernel If I didn't set HIPCXX=clang++ There was an error on cmake not finding hipcc in /usr that disapeared what is the recommended/target practice ? I was also wondering if I should call src:composable-kernelsrc:composable-kernel-rocm which is some kind of coherent with rocm native tarballs when you start a new package [1]
tar tvzf composable_kernel-rocm-7.0.2.tar.gz | head -5drwxrwxr-x root/root 0 2025-08-07 06:10 composable_kernel-rocm-7.0.2/ drwxrwxr-x root/root 0 2025-08-07 06:10 composable_kernel-rocm-7.0.2/.azuredevops/ -rw-rw-r-- root/root 496 2025-08-07 06:10 composable_kernel-rocm-7.0.2/.azuredevops/rocm-ci.yml -rw-rw-r-- root/root 2628 2025-08-07 06:10 composable_kernel-rocm-7.0.2/.clang-format -rw-rw-r-- root/root 141 2025-08-07 06:10 composable_kernel-rocm-7.0.2/.clang-tidy
But it's not what is done with other packages [1] https://github.com/ROCm/composable_kernel/releases/tag/rocm-7.0.2