Dynamic Parallelism - undefined reference to __cudaRegisterLinkedBinary linking error while compiling - separate compilation

Question

I got a problem when I try to compile a simple code there are C++ and Cuda code compile in a separated way.

Here\'s my code

main.cpp:

#includ         


        

Answer 1

This question is pretty much a duplicate of this recent question.

Dynamic parallelism requires relocatable device code linking, in addition to compiling.

Your nvcc command line specifies a compile-only operation (-rdc=true -c).

g++ does not do any device code linking. So in a scenario like this, when doing the final link operation using g++ an extra device code link step is required.

Something like this:

nvcc -arch=sm_35 -rdc=true -c file.cu
nvcc -arch=sm_35 -dlink -o file_link.o file.o -lcudadevrt -lcudart
g++ file.o file_link.o main.cpp -L<path> -lcudart -lcudadevrt

Answer 2

When using CMake, setting CUDA_SEPARABLE_COMPILATION before find_package() enables both relocatable device code compiling and linking:

SET(CUDA_SEPARABLE_COMPILATION ON)
find_package(CUDA QUIET REQUIRED)