|
|
|
|
|
|
|
|
|
cmake_minimum_required(VERSION 3.7)
|
|
|
|
|
project (CudaOnlySeparateCompilation CUDA)
|
|
|
|
|
|
|
|
|
|
#Goal for this example:
|
|
|
|
|
#Build a static library that defines multiple methods and kernels that
|
|
|
|
|
#use each other.
|
|
|
|
|
#After that confirm that we can call those methods from dynamic libraries
|
|
|
|
|
#and executables.
|
|
|
|
|
#We complicate the matter by also testing that multiple static libraries
|
|
|
|
|
#all containing cuda separable compilation code links properly
|
|
|
|
|
string(APPEND CMAKE_CUDA_FLAGS " -gencode arch=compute_30,code=compute_30")
|
|
|
|
|
set(CMAKE_CXX_STANDARD 11)
|
|
|
|
|
set(CMAKE_CUDA_STANDARD 11)
|
|
|
|
|
|
|
|
|
|
add_library(CUDASeparateLibA STATIC file1.cu file2.cu file3.cu)
|
|
|
|
|
|
|
|
|
|
if(CMAKE_CUDA_SIMULATE_ID STREQUAL "MSVC")
|
|
|
|
|
# Test adding a flag that is not in our CUDA flag table for VS.
|
|
|
|
|
if(NOT CMAKE_CUDA_COMPILER_VERSION VERSION_LESS 8)
|
|
|
|
|
string(APPEND CMAKE_CUDA_FLAGS " --ftemplate-depth 50")
|
|
|
|
|
endif()
|
|
|
|
|
# Test adding a flag that nvcc should pass to the host compiler.
|
|
|
|
|
target_compile_options(CUDASeparateLibA PRIVATE -Xcompiler=-bigobj)
|
|
|
|
|
endif()
|
|
|
|
|
|
|
|
|
|
#Having file4/file5 in a shared library causes serious problems
|
|
|
|
|
#with the nvcc linker and it will generate bad entries that will
|
|
|
|
|
#cause a segv when trying to run the executable
|
|
|
|
|
#
|
|
|
|
|
add_library(CUDASeparateLibB STATIC file4.cu file5.cu)
|
|
|
|
|
target_link_libraries(CUDASeparateLibB PRIVATE CUDASeparateLibA)
|
|
|
|
|
|
|
|
|
|
add_executable(CudaOnlySeparateCompilation main.cu)
|
|
|
|
|
target_link_libraries(CudaOnlySeparateCompilation
|
|
|
|
|
PRIVATE CUDASeparateLibB)
|
|
|
|
|
|
|
|
|
|
set_target_properties(CUDASeparateLibA
|
|
|
|
|
CUDASeparateLibB
|
|
|
|
|
PROPERTIES CUDA_SEPARABLE_COMPILATION ON
|
|
|
|
|
POSITION_INDEPENDENT_CODE ON)
|
|
|
|
|
|
|
|
|
|
if (CMAKE_GENERATOR MATCHES "^Visual Studio")
|
|
|
|
|
#Visual Studio CUDA integration will not perform device linking
|
|
|
|
|
#on a target that itself does not have GenerateRelocatableDeviceCode
|
|
|
|
|
#enabled.
|
|
|
|
|
set_target_properties(CudaOnlySeparateCompilation
|
|
|
|
|
PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
|
|
|
|
|
endif()
|
|
|
|
|
|
|
|
|
|
if (APPLE)
|
|
|
|
|
# We need to add the default path to the driver (libcuda.dylib) as an rpath, so that
|
|
|
|
|
# the static cuda runtime can find it at runtime.
|
|
|
|
|
target_link_libraries(CudaOnlySeparateCompilation PRIVATE -Wl,-rpath,/usr/local/cuda/lib)
|
|
|
|
|
endif()
|
