Hi folks, I was really excited to start working with libcu++ as I always wished there was a more ‘modern c++’ style way to write cuda. Ive found that there are sections of the standard library where it would be useful to extend the existing libcxx functionality. For a simple example it would be great if
cuda::std::numeric_limits had definitions for the types intrinsic to the device like half, TF32, and BF16. Another one I repeatedly look for is a templated version of the intrinsic vector_types like uchar4 and float2.
The longer reach feature Ive been looking for is a way to detect if a symbol is defined for the host, device or both. Something like
nvstd::is_device_function_v<Functor::operator()>. Correct me if I am wrong, but there is no way to do this currently?