Unresolved external symbol `__cudaRegisterLinkedBinary` when building as a static library

[StonerLing]

Description: When PopSift is integrated into another project via add_subdirectory() and built as a static library (BUILD_SHARED_LIBS=OFF), the linker fails with the following errors:

popsift.lib(sift_conf.obj) : error LNK2019: unresolved external symbol __cudaRegisterLinkedBinary_... referenced in function "void __cdecl __sti____cudaRegisterAll(void)"
popsift.lib(popsift.obj) : error LNK2019: unresolved external symbol __cudaRegisterLinkedBinary_... referenced in function "void __cdecl __sti____cudaRegisterAll(void)"
...

Cause: The root cause seems to be the hardcoded set(CMAKE_CUDA_SEPARABLE_COMPILATION ON) in the main CMakeLists.txt (line 114).

When RDC (Relocatable Device Code) is enabled for a static library, the host project (the executable) is responsible for the final device linking stage. However, most host projects do not enable RDC by default, leading to these unresolved symbols.

Environment:

• OS: Windows 10
• Compiler: MSVC 14.38.33130
• CUDA: 11.3.58
• CMake: 3.31.0

[simogasp]

@griwodz Shall that be enabled only for shared lib build?

[griwodz]

I don't know if that would solve a problem or create a bigger one. I suspect that it creates a bigger one.

The root cause of the problem is that PopSift does contains a few functions ("host code") that call kernels whose code is in a different .cu file. If that weren't the case, we wouldn't need the CMAKE_CUDA_SEPARABLE_COMPILATION at all.
I think that the only functions where this is the case are

• Pyramid::horiz_from_input_image
• Pyramid::horiz_level_from_input_image
• Pyramid::horiz_all_from_input_image
• Pyramid::horiz_from_prev_level
• Pyramid::vert_from_interm
• Pyramid::vert_all_from_interm
  in s_pyramid_build.cu

If someone moves each of these 6 functions to the respective s_pyramid_build_*.cu file where the kernel it calls is located, the CMAKE_CUDA_SEPARABLE_COMPILATION can be avoided altogether.