"Rosetta" for switching to native CMake CUDA support - cmake

Question

I want to switch a CMake-based project of mine to using CMake 3.8 and up's native support for CUDA. I know that, basically, I enable_language(CUDA), then use the regular add_target and add_library instead of the cuda_xxx variants of the same.
However, I'm not sure how to rewrite my code exactly. How do I...:
Set of global CUDA-related CMake variables , e.g. separable compilation, propagation of host flags?
Invoke the CUDA equivalent of CMAKE_CXX_STANDARDand friends?
Add dependencies on CUDA libraries which are not required by default (e.g. NVTX)
... and everything else?
Is there some uniform method to all this, or do I just have to memorize a bunch of new commands?

Answer 1

CMake's native CUDA support works very much in the same way as it does for C and C++ projects. Using CUDA instead of C++ is a similar process to using C instead of C++.
First of all, you should enable CUDA support in your project by adding:
project(myproject LANGUAGES CUDA)
This way, you shouldn't need to run enable_language, and your CMake cache will not be cluttered with the C and C++ compiler configuration generated by the default implied project(myproject LANGUAGES C CXX) you get when you write just project(myproject).
The equivalent of CMAKE_CXX_STANDARD for CUDA is CMAKE_CUDA_STANDARD, much like how there is also CMAKE_C_STANDARD. Additionally, there is CMAKE_CUDA_EXTENSIONS which can be configured to enable / disable compiler specific extensions (i.e. whether to choose -std=c++14 or -std=gnu++14). It's generally the case that if you see CMake variables of the form CMAKE_CXX_..... or CMAKE_C_....., there is an equivalent CMAKE_CUDA_...... And in general, where you see CXX or C in variables and properties, there may be a CUDA equivalent.
set(CMAKE_CUDA_STANDARD 11)
set(CMAKE_CUDA_STANDARD_REQUIRED ON)
set(CMAKE_CUDA_EXTENSIONS OFF)
One thing to note however, is that these particular variables have target-specific equivalents in the form of target properties (both for C++ and CUDA):
set_target_properties(
myprojectlibrary PROPERTIES
CUDA_STANDARD 11
CUDA_STANDARD_REQUIRED ON
CUDA_EXTENSIONS OFF)
As for CUDA-specific features like in the old FindCUDA, you should find these as more target properties just like those in the previous example:
set_target_properties(
myprojectlibrary PROPERTIES
CUDA_SEPARABLE_COMPILATION ON
CUDA_RESOLVE_DEVICE_SYMBOLS ON)
Information about these target properties can be found here, searching on the page for "CUDA".
As for linking to third-party libraries, this can be done the same way as it is done for regular C++ projects:
target_link_libraries(myprojectlib nvtx)
Or if your library isn't defined as a first-class CMake target:
target_link_libraries(myprojectlib ${NVTX_LIBRARY})
target_include_directories(myprojectlib PRIVATE ${NVTX_INCLUDE_DIR})
target_compile_definitions(myprojectlib PRIVATE ${NVTX_DEFINES})
Lastly, if there are any specific nvcc flags required to build your project, these can be set as follows:
target_compile_options(myprojectlib PRIVATE --default-stream=per-thread)
Note that the global user-configurable nvcc flags are also available as the variable CMAKE_CUDA_FLAGS.
Also note that you may have to change PRIVATE to PUBLIC in places where you want particular flags / include directories / compile definitions to also be propagated to targets linking to your library, in the case that myprojectlib is a library and not an executable, which may or may not be desired depending on how your project is structured.