[Bug middle-end/121664] [Nvptx][OpenMP] 'omp target ... simd' with 'collapse' – leads to illegal memory access

[schulz.benjamin at googlemail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121664

--- Comment #4 from Benjamin Schulz <schulz.benjamin at googlemail dot com> ---
What I also want to note is that with the new nvidia-driver, the system
supports cuda-13 and cuda 13 has changed its abi. 

They threw out support for old cards. This includes, i think, cards for which
the offload gcc compiler is compiled if one does not specify the arch.

Here is the recent fix in clang:


[LLVM] Fix offload and update CUDA ABI for all SM values #159354

https://github.com/llvm/llvm-project/pull/159354


Here is the statement from nvidia where they list all the deprecations (which
does not only include gpu chips):

https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html

To me, its somewhat not clear whether one must then also use an old driver with
offload gcc or if this just concerns nvidia-cuda-toolkit.

What I notice is that code which worked in the past, suddenly got tons of 
libgomp: cuCtxSynchronize error

after system upgrades (kernel, gpu, nvidia-drivers, nvidia-cuda-toolkit).
Unfortunately, reverting these did not really resolve this. 

So I really don't know what this is. 


I just want to note that clang had to react to abi changes in cuda recently. I
would assume that gcc has then react to that, too,