Will NVTX3 ship the C++ V1 API for eternity?

[bernhardmgruber]

We have included the NVTX3 C++ wrapper into the CCCL/CUB and hit a couple of challenges on how NVTX and CCCL can evolve with minimal maintenance effort. I am tempted to think the best way is to use NVTX3 like:

Inside CCCL/CUB:

#if __has_include(<nvtx3/nvtx3.hpp>)
#  include <nvtx3/nvtx3.hpp>
#  ifndef NVTX3_CPP_DEFINITIONS_V1_0
#    error 
#  endif
...
nvtx::v1::scoped_range
#endif

This way, we are independent of whether the user defines NVTX3_CPP_REQUIRE_EXPLICIT_VERSION or not (see discussion in #93) and whether they use a newer NVTX3 version, since we use the oldest API. We can happily ship CCCL and as long as we don't need any new NVTX features, this code would work if and for as long as nvtx3/nvtx3.hpp will exist.

So, does the NVTX3 C++ wrapper guarantee that if you release a new version, e.g. NVTX3 V2 or V66, all the V1 symbols will still be available, forever? An NVTX4 library can of course break this, and would also require new work on cmake, etc.

There is a comment section in nvtx3.hpp that suggests this, but I want to be sure: https://github.com/NVIDIA/NVTX/blob/release-v3/c/include/nvtx3/nvtx3.hpp#L2835-L2841