Basic cross-compilation misses CUDA drivers

Accelerated Computing HPC Compilers nvc, nvc++ and nvfortran
1

I’m trying to cross-compile Fortran code on a system without any GPU and CUDA installation to run on another system with GPUs. As far as I understand, this should be easily done e.g. by

nvfortran -mp=gpu -gpu=cc80 test.f90

or -acc=gpu for OpenACC, and so on. I am using the nvhpc module of the Nvidia HPC SDK, version 24.9. I have also set CUDA_HOME to point to the cuda directory therein. However, the compilation fails with

nvfortran-Error-A CUDA toolkit matching the current driver version (0) or a supported older version (11.8) was not installed with this HPC SDK.

It makes sense that no CUDA driver is found, since none is installed. But this shouldn’t matter for cross-compilation, shouldn’t it?

What am I missing out on here? As far as I understand, there is no additional flag required for cross compilation.

Regards,
Christian

2

You’re missing the target CUDA version, i.e. “-gpu=cudaXX.YY”.

Like the target device, the compiler auto-detects the CUDA driver version on the system to know which version to target. If you don’t have a driver installed, then you need to specify it.

Note that we ignore CUDA_HOME. If you want to use a different CUDA install than the ones that ship with the NVHPC SDK, use NVHPC_CUDA_HOME. Though this shouldn’t be necessary since all the CUDA components are included in the SDK.

3

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.