* Remove references to device-only intrinsics when compiling for host.
Currently, we attempt to use the `__device__`-only functions
`__cvta_generic_to_shared` and `__nvvm_get_smem_pointer` when compiling
`cute::cast_smem_ptr_to_uint` for the host on Clang. This results in a
compilation error, as expected. This commit changes the definition of
the `*_ACTIVATED` macros so that they are only true when `__CUDA_ARCH__`
is defined; that is, when compiling for the device.
Additionally, the declaration of `__nvvm_get_smem_pointer`
is currently only visible during the device compilation pass when
compiling with NVCC; this commit makes the declaration visible during
host compilation with the `__device__` annotation.
* Annotate cute::cast_smem_ptr_to_uint as device-only.
The implementation of `cute::cast_smem_ptr_to_uint` is currently an
unchecked failure on host code, and the only host implementation I can
think of -- casting a probably-64-bit pointer to 32 bits somehow --
doesn't make sense to implement. This commit marks this function as
device-only so that it can't be accidentally used on host code.
* small change
---------
Co-authored-by: Haicheng Wu <haichengw@nvidia.com>
fcfbd23e26