[CMake] Passing -x=cu to FindCUDA, and specifying libraries to nvcc

[Irwin Zaid]

Hi James,

Thanks for the quick reply. Some thoughts...

> FindCUDA only looks for .cu files. In CUDA_WRAP_SRCS:
>
> get_source_file_property(_is_header ${file} HEADER_FILE_ONLY)
> if(${file} MATCHES "\\.cu$" AND NOT _is_header)
>
> What benefit is there in keeping cuda code in .cpp files?

It allows us to write code that supports both a host library when there 
is no GPU and a CUDA library. We've gotten very good at this, and keep 
all our code in .cpp files.

Would it be possible to add a flag or option that disables that 
checking? I'd even be happy with it as a global "everything is CUDA" 
flag. NVCC does support this with '-x=cu'.

> You can't use nvcc to link. CMake makes this difficult, and in VS it is
> impossible. Instead what I've done for device linking is to create an
> intermediate object file that is the product of linking. FindCUDA
> supports separable compilation for a single executable module, but not
> for multiple ones. I had an example laying around for doing this, but
> I'm not sure where it ended up. Basically you add a prelink custom
> command that takes the objects and libraries that contain device code
> for your target executable module and link them using nvcc producing
> another object file that is then linked into the final executable module
> using the host linker.

I follow you here. I do this as well in a different part. I agree it's 
hard, so things may be as good as they can be.

> I'm not sure what is going wrong. If you do this:
>
> set(CUDA_NVCC_FLAGS -arch sm_20)
> add_subdirectory(tests)
>
> then CUDA_NVCC_FLAGS should have "-arch;sm_20" defined in
> tests/CMakeLists.txt.

Ah, I meant that the subproject itself can't modify CUDA_NVCC_FLAGS. The 
flags itself are passed from the superproject fine. I think this isn't 
really an issue though, the device linking is probably good enough.

Irwin