Ludovic Räss

> BTW, I'm not opposed to the rename: it shouldn't break anything, I just wanted to figure out what exactly is going on. We guess the issue could be addressed...

> > But we also got the above mentioned issue with e.g. CUDA as deps of ExtensionsConflict. > > This is a bit of a weird case: apparently if a...

> > But we also got the above mentioned issue within a project using both CUDA and ExtensionsConflict. > > Do you have an example of that? @utkinis could you...

Indeed, in this setting ☝️ no error occurs. Seems the issue occurs when within the ExtensionsConflict package, and using CUDA from the global environment (e.g. `@v1.9`).

> After login you get the following message: This message you are getting after loading the specific ROCm module combination to have 5.3.3 instead of 5.2 exposed. From LUMI support,...

With @albert-de-montserrat we worked on a MWE comparing AMDGPU, ROCBackend() and HIP C++ perf on the following micro kernel which may be somewhat representative of what happens in @albert-de-montserrat more...

> Okay so on LUMI the C++ code is already doing something quite different from Julia even without `unsafe-fp-atomics` > > Can you get the LLVM IR from hipcc on...

Still getting the error ``` ERROR: LoadError: could not load library "./atomic.so" ./atomic.so: invalid ELF header ``` in ```julia run(`hipcc -emit-llvm -S --offload-device-only -c -xhip -munsafe-fp-atomics -O3 -o $libname --shared...

> You don't want `--shared -fIC` since your `-o` should be a `.ll`, `hipcc -xhip -munsafe-fp-atomics -O3 -o out.ll atomic.cu` If I remove it and change the `-o` from `.so`...

> But now you are missing `-emit-llvm -S --offload-device-only -c`? Ok, got it now. Was confused. The generated out.ll is [out.ll.zip](https://github.com/JuliaGPU/AMDGPU.jl/files/13686776/out.ll.zip)