cmake, gcc, cuda and -m32

问题:

I figured out that CUDA does not work in 64bit mode on my mac (or couldn't get it running so far). Therefore I decided to compile everything for 32bit.

I use cmake 2.8 and added the following options

add_definitions(-Wall -m32)
set(CUDA_64_BIT_DEVICE_CODE OFF)
set(CMAKE_MODULE_LINKER_FLAGS -m32)

However when it tries to link it it does something like this:

/usr/bin/c++    -mmacosx-version-min=10.6 -Wl,-search_paths_first -headerpad_max_install_names  CMakeFiles/SimpleTestsCUDA.dir/BlockMatrix.cpp.o CMakeFiles/SimpleTestsCUDA.dir/Matrix.cpp.o ./SimpleTestsCUDA_generated_SimpleTests.cu.o ./SimpleTestsCUDA_generated_BlockMatrix.cu.o  -o SimpleTestsCUDA  /usr/local/cuda/lib/libcudart.dylib /usr/local/cuda/lib/libcuda.dylib 

Which fails with a lot of "file is not of required architecture" warnings from ld. Now if I add manually -m32 to the command above it works. However I have no idea how to teach cmake to add -m32 to every gcc (or ld) invocation. So far it does it for nvcc and gcc, but not for linking..

回答1:

If you set the env var LDFLAGS before you run cmake on the project it will work as well:

export LDFLAGS=-m32
cmake ../source

回答2:

see above

set(CMAKE_C_FLAGS -m32)
set(CMAKE_CXX_FLAGS -m32)

回答3:

Another solution might be to say:

if (Apple)
  set (CMAKE_OSX_ARCHITECTURES i386)
  set (CUDA_64_BIT_DEVICE_CODE OFF) 
endif (Apple)

Hope this helps.