Fix hardcoded CUDA versions to folder name autodetection? about cudaapi.jl HOT 2 CLOSED

This isn't as easy as it sounds. Instead of doing Libdl.find_library on a couple of library names (i.e. appending possible version strings) we'd need to redo all that logic and actively scan the system library paths for candidate libraries and extract a version tag.

Alternatively, we could have some functionality that scans the system and returns candidate version numbers for use with the existing infrastructure in CUDAapi, but then we'd be duplicating functionality (that hard codes, e.g., the paths where CUDA is installed, or where libcudnn could reside).

In summary, I think it's much easier to just maintain a version list and bump them when necessary.
If you think otherwise, please go ahead and propose a change 🙂

from cudaapi.jl.

Besides, there's other compatibility information to be extracted from a new CUDA release which is even harder to automate: #67

from cudaapi.jl.