How to edit rpath flag format for linking library

Code
1

I’m building a project that uses the CUDA Runtime library, as well as some tests for the project using CTest. Here is an example to illustrate my problem.

Directory structure:

Dir/
├── CMakeLists.txt
├── main.c
├── main.h
├── kernel.cu
└── kernel.h
├──tests/
    ├──CMakeLists.txt
    ├──test1.c

Environment:

NVCC: 12.1

CMake: 3.25

kernel.h

void f()

kernel.cu

> extern "C" {

#include "kernel.h"
}

extern "C" void f() {
    //implementation
    //use cuda functions
}

main.c

> #include "kernel.h"

void fmain(){
    f();
}

int main() {
return 0;
}

main.h

void fmain();

CMakeLists.txt

> cmake_minimum_required(VERSION 3.18)

if(NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
  set(CMAKE_CUDA_ARCHITECTURES 70)
endif()

project(test LANGUAGES C CXX CUDA)

add_library(cuda_comp SHARED kernel.h kernel.cu)

set_target_properties(cuda_comp
    PROPERTIES
            CUDA_RUNTIME_LIBRARY Shared
)

add_executable(main main.c)

target_link_libraries(main cuda_comp)

mark_as_advanced( BUILD_TESTS )
set( BUILD_TESTS true CACHE BOOL "Tests build target available if true" )
if( BUILD_TESTS )
    enable_testing()
    add_subdirectory( tests )
endif()

This main target builds with no problem, however, when linking the tests I get an unresolved reference to f().

tests/test1.c

> #include "main.h"

int main() {
    fmain(); //fmain() calls f()
    return 0;
}

tests/CMakeLists.txt

> add_executable(test1 test1.c)

add_test( NAME "_test1" COMMAND "test1" )

After adding target_link_libraries(test1 PRIVATE cuda_comp) , therefore having:

tests/CMakeLists.txt

> add_executable(test1 test1.c)

target_link_libraries(test1 PRIVATE cuda_comp)
add_test( NAME "_test1" COMMAND "test1" )

I get a compiler flag error, this is the relevant output:

*Linking C executable test1*

  • test1 -Wl,-rpath,/pathToMainDirectory -lm …/libcuda_comp.so *

  • nvcc fatal : Unknown option ‘-Wl,-rpath,/pathToMainDirectory’*

I know that the -WL,-rpath format is incorrect for NVCC. How can I correct the format of the rpath?

2

Cc: @robert.maynard

3

Can you please post the verbose build output.

Locally and in our CI, CMake passes the correct flags for rpath addition for nvcc.

4

The test1 executable is a C program which means that your C compiler should be doing the linking, and not nvcc which is only a CUDA compiler.