Compile CUDA code with cmake and 3.5 compute capability

Question

I need to compile a CUDA code that use a dynamic parallelism with cmake. The code is:

#include <stdio.h>

__global__ void childKernel() {
    printf("Hello ");    
}

__global__ void parentKernel() {    
    childKernel<<<1,1>>>();
    cudaDeviceSynchronize();    
    printf("World!\n");    
}

int main(int argc, char **argv){

    parentKernel<<<1, 1>>>();

    return 0;
}

and the cmake is the following:

cmake_minimum_required(VERSION 2.8)

find_package(CUDA QUIET REQUIRED)

include_directories(/usr/include)
include_directories(/usr/local/cuda/lib)
include_directories(/usr/local/cuda-8.0/lib)
include_directories(/usr/local/cuda/include)
include_directories(/usr/local/cuda-8.0/include)

set(CUDA_SEPARABLE_COMPILATION ON)
find_package(CUDA QUIET REQUIRED)
set(CUDA_PROPAGATE_HOST_FLAGS OFF)

set(
    CUDA_NVCC_FLAGS 
    ${CUDA_NVCC_FLAGS}; 
    -arch=compute_35 -rdc=true -lcudadevrt
)

cuda_add_executable(
    prova 
    test.cu
)

I have tried to compile the code directly with nvcc passing the -arch=compute_35 -rdc=true -lcudadevrt and it compiles perfectly, but when I try to compile with cmake it returns me the following error:

CMakeFiles/prova.dir/prova_intermediate_link.o: In function `__cudaRegisterLinkedBinary_66_tmpxft_00001101_00000000_13_cuda_device_runtime_compute_62_cpp1_ii_8b1a5d37':
link.stub:(.text+0xcc): undefined reference to `__fatbinwrap_66_tmpxft_00001101_00000000_13_cuda_device_runtime_compute_62_cpp1_ii_8b1a5d37'
link.stub:(.text+0xd0): undefined reference to `__fatbinwrap_66_tmpxft_00001101_00000000_13_cuda_device_runtime_compute_62_cpp1_ii_8b1a5d37'
collect2: error: ld returned 1 exit status
CMakeFiles/prova.dir/build.make:200: recipe for target 'prova' failed
make[2]: *** [prova] Error 1
CMakeFiles/Makefile2:67: recipe for target 'CMakeFiles/prova.dir/all' failed
make[1]: *** [CMakeFiles/prova.dir/all] Error 2
Makefile:83: recipe for target 'all' failed
make: *** [all] Error 2

Answer 1

Thanks for the answer, I have tried to use the correct compute capability of the Tegra X2 (compute_62), and I have inspected the output of verbose make and I found the following output:

CMakeFiles/prova.dir/prova_generated_test.cu.o CMakeFiles/prova.dir/prova_intermediate_link.o  -o prova -rdynamic /usr/local/cuda-8.0/lib64/libcudart_static.a -lpthread -ldl -lrt

It seems that the linker try to search /usr/local/cuda-8.0/lib64/libcudart_static.a but not libcudadevrt.a

Answer 2

The undefined symbols you're seeing are defined in libcudadevrt.a. I see that you're telling CMake to link to it, but it appears the compiler can't find it. Try this on the command line:

VERBOSE=1 make

and inspect the output to see if you're searching /usr/local/cuda/lib64 for libraries.

This may be unrelated, but I also notice that you're instructing CMake to search /usr/local/cuda/lib and /usr/local/cuda-8.0/lib for header files. This is incorrect, as these directories contain only libraries.