When #1336 added usage of new intrinsics to <cmath>, we forgot about CUDA. 🙀 (Long ago, we broke CUDA by adding new type traits intrinsics in C++14 mode, but these math functions were just different enough that I didn't remember the interaction. Oops!)

I'm not exactly sure why our CUDA unit test didn't catch this, given that the affected overloads aren't templates:

We are including <cmath>, so there's probably some detail of the CUDA compilation process that I don't understand:

In any event, fixing this should be easy, we just need to backport it to VS 2019 16.9.

Originally encountered in pytorch/pytorch#54382 and tracked by Microsoft-internal VSO-1314894 / AB#1314894 .