nvlink error when linking CUDA code against CUDA static library - CMake

Question

I create a static library, which contains some CUDA code and some regular C++ code. The CMakeLists.txt for this static library looks like this:

SET(TARGET_H some_header.h)
SET(CUDA_SRC cudaclass1.cu cudaclass2.cu)
SET(TARGET_SRC cppclass1.cpp cppclass2.cpp)
SET(CUDA_NVCC_FLAGS "")
SET(CUDA_SEPARABLE_COMPILATION ON)
CUDA_ADD_LIBRARY(somestatic ${TARGET_H} ${TARGET_SRC} ${CUDA_SRC} OPTIONS -arch sm_20)

This will produce libsomestatic.a.

Now I want to link this static library against an executable, which itself consists of both CUDA code and C++ code.

In the CUDA code of this executable, I need to instantiate and use CudaClass1 from libsomestatic.a.

This is the CMakeLists.txt file for the executable:

SET(TARGET_H some_header.h)
SET(CUDA_SRC cudafile.cu)
SET(TARGET_SRC main.cpp)
SET(CUDA_NVCC_FLAGS "")
SET(CUDA_SEPARABLE_COMPILATION ON)
CUDA_ADD_EXECUTABLE(some_exe ${TARGET_H} ${TARGET_SRC} ${CUDA_SRC} OPTIONS -arch sm_20)
TARGET_LINK_LIBRARIES(some_exe somestatic)

But I get the following error when building:

nvlink error: Undefined reference to '_ZN10test7CudaClass1C1EPKNS_11intE' in ...

I have two questions:

1. Is it possible at all to accomplish this? Linking CUDA code against a library which contains CUDA code?
2. If yes, how would I do this using CMake?

The CUDA documentation states (http://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html#examples)

"It is possible to do multiple device links within a single host executable, as long as each device link is independent of the other (they cannot share code across device executables)"

This sounds like I cannot do what I want to do, right?

I use CUDA 5.5 and GCC 4.8.