CUDA support

Warning

This is an experimental feature, available starting with release 2.3.0. It is still incomplete and has only been tested on Linux on x86_64.

CUDA is supported by passing all JITed code through two pipelines: one for the CPU and one for the GPU. Use of the __CUDA__ pre-processor macro enables more fine-grained control over which pipeline sees what, which is used e.g. in the pre-compiled header: the GPU pipeline has the CUDA headers included, the CPU pipeline does not. Building the pre-compiled header will also pick up common CUDA libraries such as cuBLAS, if installed.

Each version of CUDA requires specific versions of Clang and the system compiler (e.g. gcc) for proper functioning; it’s therefore best to build the backend (cppyy-cling) from source for the specific combination of interest. The 3.x series of cppyy uses Clang13, the 2.x series Clang9, and this may limit the CUDA versions supported (especially since CUDA has changed the APIs for launching kernels in v11).

There are three environment variables to control Cling’s handling of CUDA:

• CLING_ENABLE_CUDA (required): set to 1 to enable the CUDA backend.

• CLING_CUDA_PATH (optional): set to the local CUDA installation if not in a standard location.

• CLING_CUDA_ARCH (optional): set the architecture to target; default is sm_35 (Clang9 is limited to sm_75).

After enabling CUDA with CLING_ENABLE_CUDA=1 CUDA code can be used and kernels can be launched from JITed code by in cppyy.cppdef(). There is currently no syntax or helpers yet to launch kernels from Python.