NVTX should ship CMake package config so that find_package works

[bdice]

Currently it requires a lot of hacks and non-obvious knowledge to use packages of NVTX with CMake, such as nvtx-c on conda-forge, because the library does not ship a proper package config file.

conda-forge/pytorch-cpu-feedstock#357 is a good example of the kind of problem this creates -- the nvtx-c packages contain NVTX headers but they are not findable by the build system. Several packaging experts and maintainers (including me) were unable to point to a correct solution.

Presently, conda packages that want to get NVTX as a dependency must install nvtx-c and tell their build system something like "trust me, it's already on the include path." This also makes it very hard to support CPM-based workflows that try to find the package locally and fetch it from upstream if not found locally (the find-local step always fails).

As a workaround for this problem, RAPIDS repackages NVTX with package exports and other configuration so that it includes the following files:

lib/cmake/nvtx3/nvtx3-config-version.cmake
lib/cmake/nvtx3/nvtx3-config.cmake
lib/cmake/nvtx3/nvtx3-targets.cmake

Currently librmm conda packages vendor nvtx3, but only because I can't make nvtx-c work without additional package configuration.

I think the solution is that we need to ship lib/cmake/nvtx3/nvtx3-config.cmake as a part of the NVTX tarballs published to GitHub, but I would like confirmation from others.

cc: @robertmaynard @KyleFromNVIDIA @vyasr @evanramos-nvidia for awareness

[bdice]

conda-forge packages of NVTX 3.2.0 are now available. With the new CMake install rules, you can now

conda install -c conda-forge nvtx-c

and then in CMake, find_package(nvtx3) just works!