Support CUDA 12.4 and 12.5

[ichinii]

after updating to cuda 12.5 i get following error message when running my alpaka program:

what(): /home/ich/projects/rayx/Extern/alpaka/include/alpaka/mem/buf/uniformCudaHip/Copy.hpp(143) 'TApi::setDevice(m_iDstDevice)' A previous API call (not this one) set the error : 'cudaErrorUnsupportedPtxVersion': 'the provided PTX was compiled with an unsupported toolchain.'!

It is not clear to me where this error comes from and if this might be a issue that a future alpaka version would fix. Do you maybe have a guess?

At this point I am asking myself if it is expected, that new cuda versions will likely make alpaka applications unable to function correctly? In other words: Should one be aware of the alpaka cuda support, before updating cuda on a system?

[fwyzard]

what GPU are you compiling for ?

[psychocoderHPC]

@ichinii, please post the cmake command used to configure and compile.

Do you set -DCMAKE_CUDA_ARCHITECTURES=52 where 52 is the sm level of your target GPU.

[psychocoderHPC]

https://forums.developer.nvidia.com/t/provided-ptx-was-compiled-with-an-unsupported-toolchain-error-using-cub/168292 is saying you have a driver and toolchain mismatch.

• Do you compiled and executed the app on the same system?
• Do you have more than one CUDA SDK on the system?

Most most likely you missed to update you system driver.

[ichinii]

Oh I am sorry. Turns out the driver version does not catch up to the cuda version, although both were up to date on my manjaro system. Downgrading cuda to 12.3. works.

[psychocoderHPC]

I closed the issue, looks like it is solved

[psychocoderHPC]

I reponed it again, missed that the original issue is about updating the CI

[SimeonEhrig]

At this point I am asking myself if it is expected, that new cuda versions will likely make alpaka applications unable to function correctly? In other words: Should one be aware of the alpaka cuda support, before updating cuda on a system?

@ichinii Yes, you should check the official supported CUDA version before you update. Otherwise you code could broke. Unfortunately it happens, that some new CUDA SDK versions introduces bugs, which we trigger in alpaka because we use complicated meta template programming. Normally this bugs are fixed in one of the next releases. But this means, we as alpaka developers needs to implement workarounds for specific CUDA versions.

But most of the time new CUDA SDK versions works out-of-the box. But be aware, if you use an unsupported CUDA version, maybe you need to implement workarounds by yourself.