[FEA] Support latest NVCC to compile with cuda10

[Xseventh]

Is your feature request related to a problem? Please describe. recently, I try to compile cutlass with cuda10 and nvcc11.6, because nvcc is not bind with cuda lib. If use nvcc11, we can support more cuda_arch_gencode, and can get more compiler optimize opportunity.

Describe the solution you'd like I found many macro CUDACC_VER_MAJOR for check cuda version in cutlass, but CUDA_VERSION macro can more exactly express current used cuda version. If not necessary required compiler version, we should use CUDA_VERSION to instead of CUDACC_VER_MAJOR macro.

[hwu36]

We actually tried CUDA_VERSION, but it causes us issues.

We do not recommend to mix different versions of cuda and nvcc, especially the major version number is different. They are not tested in this way.

[Xseventh]

Thanks for your reply, I want to known what issue caused by macro CUDA_VERSION, do you have some comment?

[hwu36]

I honestly don't remember. We need to support the big matrix of different version {nvcc, nvrtc, clang, msvc, nvc++} x different gpus. One or a few of the combination has the issue.

[github-actions[bot]]

This issue has been labeled inactive-30d due to no recent activity in the past 30 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be labeled inactive-90d if there is no activity in the next 60 days.