unsupported gpu architecture?

[clarkedavida]

It seems that the half-precision stuff doesn't compile for GTX Titan X cards, architecture=52. In particular it doesn't seem to like the half precision stuff.

One option could be to prevent half-precision compilation for this architecture. Another could be to simply make a list of unsupported architectures somewhere.

What do you all think?

[dbollweg]

Half precision is only available from Pascal (sm=60) and up.

[clarkedavida]

I see. Would it be sufficient to add some logic to the CMakeLists.txt to get it not to try to compile for half precision when the architecture is lower than 60? Otherwise I think we need to specify an architecture requirement in our documentation