GitLab

* * Added support for common operations on FP16 (`half` or `__half`) for older GPU architectures
* Fixed an issue with previous check for FP16 support

* * Removing FP16 type checks, since they should no longer be needed

* * Fixed AtomicAdd to be atomic for `float` and `double` for old GPU architectures.  Unfortunately, it seems that atomicCAS for unsigned short seems to be unavailable until architecture 70, so half will have to stay non-atomic on old GPUs.

* * Fixed non-atomic version of `AtomicAdd<half>` for older GPUs to return old value instead value of new