cuda 12.5

[cekees]

I saw during configuration that chrono does not support cuda 12.5 yet. Is that a hard constraint, or should I modify the cmakefile and try it?

[rserban]

Unfortunately, it's not as simple as modifying the CMakeLists file.

There are several issues with using CUDA 12.5 in Chrono:

• the latest version of Thrust (2.4.0) included with the CUDA 12.5 toolkit does not allow thrust/ headers in host-only TUs when THRUST_DEVICE_SYSTEM=CUDA and only allows thrust/ headers in host-only TUs when THRUST_DEVICE_SYSTEM={CPP, OMP, TBB}.
  This will require a refactoring of the internal multicore collision detection algorithms and of the Chrono::Multicore module, both of which rely on Thrust with the OpenMP backend.
• the API for cusparse and cublas has changed, with some functionality (e.g. incomplete decomposition preconditioners) completely removed.
  This will require a refactoring of the Chrono::FSI module (and possibly the Chrono::GPU module).

All of the above will have to happen at some point, but I do not have the bandwidth to tackle any of it for quite a while.

[cekees]

Thanks, @rserban. OK, it likes like I may be able top install CUDA 12.3 alongside CUDA 12.5.

[salhus]

@cekees If you don't mind answering. Did it work with CUDA 12.3? The documentation suggests that the GPU and FSI module were tested with Pascal architecture with CUDA 11.4.
I have Ada with 12.3, but I am having errors about a mismatch between the static and dynamic library configuration settings. If you know a configuration that works, it will be a massive help.

Cheers,
sal