[Feature Request]: cuda 13.1 support

[johnfuller360]

I would like to use chapel, but have installed 13.1 since it was the latest distribution.
It is difficult to manage multiple drivers, so it would be ideal if I could use 13.1

[e-kayrakli]

Just to capture my observations from Discourse here:

The current limitation is coming from Clang compiling CUDA. Chapel supports Clang up to version 20, which can't compile CUDA 13, so we can't build the runtime with CUDA 13.

I haven't checked closely whether Clang 21 can compile CUDA 13. But in any case, we already have plans to extend the supported Clang versions. And once we do that, we can take a look at CUDA 13. Things might "just work" at that point, or there may be other CUDA-specific fixes needed.

Discourse thread

[jabraham17]

Supporting this is not likely to be possible with LLVM 21. I think we will need to wait for Chapel to support LLVM 22 at least. This is because you cannot seem to use LLVM 21 with CUDA 13

main...jabraham17:chapel:cuda13 has the results of my experiments. In more detail, that branch has the following changes

• CUB now requires C++ 17
• sm_60 has been dropped, we now need to default to sm_75
• CUB now lives with a bunch of other libraries called CCCL. These paths are not easy to find, for now I just included everything nvcc does. This works, but is over-aggressive and imo may cause issues

I also found that CUDA 13 support in LLVM is blocked until llvm/llvm-project#151766 is released. That patch is included in LLVM 22 (unreleased as I write this). I patched my local copy and kept going.

CUB has deprecated a bunch of interfaces we use, I haven't fixed those but we will need to address them in the future. For now its fine.

gpu-nvidia-cub.cc:92:46: warning: 'ArgMax<double *, cub::KeyValuePair<int, double> *>' is deprecated: CUB has superseded this
      interface in favor of the ArgMax interface that takes two separate iterators: one iterator to which the extremum is written and another
      iterator to which the index of the found extremum is written. [-Wdeprecated-declarations]
   92 | GPU_DEV_CUB_WRAP(DEF_ONE_REDUCE_RET_VAL_IDX, ArgMax, maxloc)
      |                                              ^
cuda/13.0/bin/../targets/sbsa-linux/include/cccl/cub/device/device_reduce.cuh:1024:3: note:
      'ArgMax<double *, cub::KeyValuePair<int, double> *>' has been explicitly marked deprecated here
 1024 |   CCCL_DEPRECATED_BECAUSE("CUB has superseded this interface in favor of the ArgMax interface that takes two separate "
      |   ^
cuda/13.0/bin/../targets/sbsa-linux/include/cccl/cuda/std/__cccl/deprecated.h:66:42: note:
      expanded from macro 'CCCL_DEPRECATED_BECAUSE'
   66 | #  define CCCL_DEPRECATED_BECAUSE(MSG) [[deprecated(MSG)]]

The final issues I hit was that in CUDA 12.9 the fatbinary executable remove the flag --image. This is problematic for two reasons.

• clang uses it internally when we build the runtime
• we invoke fatbinary directly when compiling Chapel code
  The fix is [CUDA] Use --image3 to construct fat binary llvm/llvm-project#151760, but its only in LLVM 22 (unreleased as I write this). This is where I stopped, as that patch is harder to make. I never adjusted the Chapel compiler as I did not get a successful runtime build yet