• edit the Makefile to point CUDA_BASE to a CUDA installation (e.g. `CUDA_BASE = /usr/local/cuda')
• run make to build all binaries:

  $ make
  /usr/local/cuda/bin/nvcc --generate-line-info --source-in-ptx --expt-relaxed-constexpr --expt-extended-lambda -std=c++14 -O2 --cudart=shared -gencode arch=compute_50,code=sm_50 --compiler-options "-O2 -pthread -fPIC" -dc foo.cu -o foo.o
  /usr/local/cuda/bin/nvcc --generate-line-info --source-in-ptx --expt-relaxed-constexpr --expt-extended-lambda -std=c++14 -O2 --cudart=shared -gencode arch=compute_50,code=sm_50 --compiler-options "-O2 -pthread -fPIC" -dlink foo.o -o foo_dlink.o
  g++ -O2 -pthread -fPIC --shared foo.o foo_dlink.o -L/usr/local/cuda/lib64 -lcudart -lcudadevrt -Wl,-rpath -Wl,/usr/local/cuda/lib64 -o libfoo.so
  /usr/local/cuda/bin/nvcc --generate-line-info --source-in-ptx --expt-relaxed-constexpr --expt-extended-lambda -std=c++14 -O2 --cudart=shared -gencode arch=compute_50,code=sm_50 --compiler-options "-O2 -pthread -fPIC" -dc bar.cu -o bar.o
  /usr/local/cuda/bin/nvcc --generate-line-info --source-in-ptx --expt-relaxed-constexpr --expt-extended-lambda -std=c++14 -O2 --cudart=shared -gencode arch=compute_50,code=sm_50 --compiler-options "-O2 -pthread -fPIC" -dlink bar.o -o bar_dlink.o
  g++ -O2 -pthread -fPIC --shared bar.o bar_dlink.o -L/usr/local/cuda/lib64 -lcudart -lcudadevrt -Wl,-rpath -Wl,/usr/local/cuda/lib64 -o libbar.so
  g++ -O2 -pthread -fPIC main.cc -L. -lfoo -ldl -Wl,-rpath -Wl,. -L/usr/local/cuda/lib64 -lcudart -lcudadevrt -Wl,-rpath -Wl,/usr/local/cuda/lib64 -o test
  /usr/local/cuda/bin/nvcc --generate-line-info --source-in-ptx --expt-relaxed-constexpr --expt-extended-lambda -std=c++14 -O2 --cudart=shared -gencode arch=compute_50,code=sm_50 --compiler-options "-O2 -pthread -fPIC" -DMAY_CRASH -dc bar.cu -o bar_crash.o
  /usr/local/cuda/bin/nvcc --generate-line-info --source-in-ptx --expt-relaxed-constexpr --expt-extended-lambda -std=c++14 -O2 --cudart=shared -gencode arch=compute_50,code=sm_50 --compiler-options "-O2 -pthread -fPIC" -dlink bar_crash.o -o bar_crash_dlink.o
  g++ -O2 -pthread -fPIC --shared bar_crash.o bar_crash_dlink.o -L/usr/local/cuda/lib64 -lcudart -lcudadevrt -Wl,-rpath -Wl,/usr/local/cuda/lib64 -o libbar_crash.so
  g++ -DMAY_CRASH -O2 -pthread -fPIC main.cc -L. -lfoo -ldl -Wl,-rpath -Wl,. -L/usr/local/cuda/lib64 -lcudart -lcudadevrt -Wl,-rpath -Wl,/usr/local/cuda/lib64 -o crash

• run test:

  $ ./test
  Hello
  Loaded
  bar
  

• run crash:

  $ ./crash
  Hello
  Loaded
  bar.cu, line 24: cudaErrorInvalidDeviceFunction: invalid device function
  

main() does

• call fooWrapper() from foo.cc, which in turns launches a CUDA kernel foo<<<1,1>>>()
• dynamically loads a shared library libbar.so
• upon loading, libbar.so calls wrapper() which in turn calls a CUDA kernel bar<<<1,1>>>()
• if, inside libbar.so, there is another kernel that makes use of dynamic parallelism (e.g. it calls bar<<<1,1>>>() and cudaDeviceSynchronize()), the call to bar<<<1,1>>>() will fail even if the kernel with dynamic parallelism is never called