CUDA support is available in two flavors. The new method, introduced in CMake 3.8 (3.9 for Windows), should be strongly preferred over the old, hacky method - I only mention the old method due to the high chances of an old package somewhere having it. Unlike the older languages, CUDA support has been rapidly evolving, and building CUDA is hard, so I would recommend you _require a very recent version_ of CMake! CMake 3.17 and 3.18 have a lot of improvements directly targeting CUDA.
A good resource for CUDA and Modern CMake is [this talk](http://on-demand.gputechconf.com/gtc/2017/presentation/S7438-robert-maynard-build-systems-combining-cuda-and-machine-learning.pdf) by CMake developer Robert Maynard at GTC 2017.
When you build CUDA code, you generally should be targeting an architecture. If you don't, you compile 'ptx', which provide the basic instructions but is compiled at runtime, making it potentially much slower to load.
All cards have an architecture level, like "7.2". You have two choices; the first is the code level; this will report to the code being compiled a version, like "5.0", and it will take advantage of all the features up to 5.0 but not past (assuming well written code / standard libraries). Then there's a target architecture, which must be equal or greater to the code architecture. This needs to have the same major number as your target card, and be equal to or less than the target card. So 7.0 would be a common choice for our 7.2 card. Finally, you can also generate PTX; this will work on all future cards, but will compile just in time.
In CMake 3.18, it became very easy to target architectures. If you have a version range that includes 3.18 or newer, you will be using `CMAKE_CUDA_ARCHITECTURES` variable and the `CUDA_ARCHITECTURES` property on targets. You can list values (without the `.`), like 50 for arch 5.0. If set to OFF, it will not pass architectures.
Using targets should work similarly to CXX, but there's a problem. If you include a target that includes compiler options (flags), most of the time, the options will not be protected by the correct includes (and the chances of them having the correct CUDA wrapper is even smaller). Here's what a correct compiler options line should look like:
However, if you using almost any find_package, and using the Modern CMake methods of targets and inheritance, everything will break. I've learned that the hard way.
For now, here's a pretty reasonable solution, _as long as you know the un-aliased target name_. It's a function that will fix a C++ only target by wrapping the flags if using a CUDA compiler:
If you want to support an older version of CMake, I recommend at least including the FindCUDA from CMake version 3.9 in your cmake folder (see the CLIUtils github organization for a [git repository](https://github.com/CLIUtils/cuda_support)). You'll want two features that were added: `CUDA_LINK_LIBRARIES_KEYWORD` and `cuda_select_nvcc_arch_flags`, along with the newer architectures and CUDA versions.
To use the old CUDA support, you use `find_package`:
```cmake
find_package(CUDA 7.0 REQUIRED)
message(STATUS "Found CUDA ${CUDA_VERSION_STRING} at ${CUDA_TOOLKIT_ROOT_DIR}")
```
You can control the CUDA flags with `CUDA_NVCC_FLAGS` (list append) and you can control separable compilation with `CUDA_SEPARABLE_COMPILATION`. You'll also want to make sure CUDA plays nice and adds keywords to the targets (CMake 3.9+):
```cmake
set(CUDA_LINK_LIBRARIES_KEYWORD PUBLIC)
```
You'll also might want to allow a user to check for the arch flags of their current hardware:
```cmake
cuda_select_nvcc_arch_flags(ARCH_FLAGS) # optional argument for arch to add