FindCUDAToolkit doesn't detect CUDA toolkit shipped inside NVIDIA HPC SDK
So far FindCUDAToolkit cannot fully accept the CUDA toolkit shipped inside NVIDIA HPC SDK (call PGI in short) because of its non-usual layout. nvcc can be found but not header files when PGI modules are loaded. PGI only support 3 versions of CUDA toolkit, for example 20.11 only support 10.2, 11.0, 11.1. Once CUDA 11.2 is installed in the system, FindCUDAToolkit founds CUDA 11.2. PGI cannot link the final executable due to PGI internal 11.1 and external 11.2 conflict. I'm wondering if there are ways to teach FindCUDAToolkit using PGI internal CUDA toolkit. Or if I can query PGI in cmake about its underlying CUDA version then I can at least add a safeguard for mismatched versions.