Windows CUDART link

[yhmtsai]

In #351 , how do we link cudart in Windows?
We set CUDA_RUNTIME_LIBS as the cudart shared libraries, but nvcc will use -cudart=static as its default.
It works fine in Linux, but it can not be resolved in Windows.
The problem is in benchmark/spmv which links to static and shared library.
The following are workaround solutions

1. Do not link the cuda library in benchmark/spmv in Windows

   ginkgo/benchmark/spmv/CMakeLists.txt

   Lines 5 to 9 in a983181

   	if (NOT MSVC)
   	target_link_libraries(spmv ginkgo ${CUDA_RUNTIME_LIBS}
   	${CUBLAS} ${CUSPARSE})
   	target_include_directories(spmv SYSTEM PRIVATE ${CUDA_INCLUDE_DIRS})
   	endif()

2. Force nvcc to use -cudart=shared in Windows.
3. Use cudart shared library when giving -cudart=shared, otherwise use cudart static library.
   cudart_static needs rt, dl, and pthread, so it needs to add them into the libraries.
   

   ginkgo/cuda/CMakeLists.txt

   Lines 45 to 55 in 01ee57c

   	if("${CMAKE_CUDA_FLAGS}" MATCHES "-cudart(=| )shared" OR "${GINKGO_CUDA_COMPILER_FLAGS}" MATCHES "-cudart(=| )shared")
   	set(CUDA_RUNTIME_LIBS "${CUDA_RUNTIME_LIBS_DYNAMIC}" CACHE STRING "Path to a library" FORCE)
   	else()
   	set(CUDA_RUNTIME_LIBS "${CUDA_RUNTIME_LIBS_STATIC}" CACHE STRING "Path to a library" FORCE)
   	if(NOT MSVC)
   	set(THREADS_PREFER_PTHREAD_FLAG ON)
   	find_package(Threads REQUIRED)
   	# link cudart_static need rt, pthread, and dl
   	set(CUDA_RUNTIME_LIBS "${CUDA_RUNTIME_LIBS_STATIC};rt;Threads::Threads;-Wl,--no-as-needed;dl" CACHE STRING "Path to a library" FORCE)
   	endif()
   	endif()

   
   

   ginkgo/benchmark/spmv/CMakeLists.txt

   Lines 5 to 9 in 01ee57c

   	if("${CUDA_RUNTIME_LIBS}" MATCHES "Threads::Threads")
   	# need to import Threads
   	set(THREADS_PREFER_PTHREAD_FLAG ON)
   	find_package(Threads REQUIRED)
   	endif()

it seems to be failed with Threads::Threads but successful with -pthread.

I prefer to choose 2 or 3

I think option 1 is weird.
option 2 does not change too many things in CMake, but it maybe reduces usability in Windows.
option 3 works fine but benchmark/spmv needs to include threads to use Threads::threads correctly.

Which option do you think better way? Alternatively, is there any idea to solve this problem?

edit:
delete the fail with Threads::Threads

[yhmtsai]

Conclusion:
Keep the same behavior in linux/MacOS.