Why can I not include my cu-file header in my main.cpp file despite enabling CUDA in CMake?

Question

I have a small c++ and cuda test project where I am trying to compile a simple setup containing a main.cpp file along with a test.cu and test.h file with some gpu functionality. My project setup looks like this:

project
│   CMakeLists.txt
│   main.cpp
|   test.h
|   test.cu
|
└───build

CMakeLists.txt

cmake_minimum_required(VERSION 3.24)
project(test LANGUAGES CXX CUDA)

set(CMAKE_CXX_STANDARD 14)
set(CMAKE_CUDA_STANDARD 14)

set(SOURCES ${SOURCES}
  test.cu
)

add_executable(test main.cpp ${SOURCES})
set_target_properties(${PROJECT_NAME} PROPERTIES CUDA_ARCHITECTURE "75")

test.h

#ifndef TEST_H_ 
#define TEST_H_ 

#include <cuda_runtime.h>

void test(int *arr, int size);

#endif // TEST_H_ 

test.cu

#include <cuda_runtime.h>

#include "test.h"

__global__ void test_kernel(int *arr, int size) {
  arr[threadIdx.x] += 1;
}

void test(int *arr, int size) {
  int *temp;
  cudaMalloc(&temp, sizeof(int)*size);
  cudaMemcpy(temp, arr, sizeof(int)*size, cudaMemcpyHostToDevice);

  test_kernel<<<1, 256>>>(temp, size);

  cudaMemcpy(arr, temp, sizeof(int)*size, cudaMemcpyDeviceToHost);

  cudaDeviceSynchronize();
  cudaFree(temp);
}

main.cpp

#include <iostream>

#include "test.h"

int main(int argc, char **argv) {
  int *arr = new int[10];
  test(arr, 10);

  delete[] arr;
  return 0;
}

I am creating the build files and compiling by doing

cd build
cmake ..
make

however, when running make I get the following error

In file included from /home/bickit/Documents/test/main.cpp:3:
/home/bickit/Documents/test/test.h:4:10: fatal error: cuda_runtime.h: No such file or directory
    4 | #include <cuda_runtime.h>
      |          ^~~~~~~~~~~~~~~~
compilation terminated.
make[2]: *** [CMakeFiles/test.dir/build.make:76: CMakeFiles/test.dir/main.cpp.o] Error 1
make[1]: *** [CMakeFiles/Makefile2:83: CMakeFiles/test.dir/all] Error 2
make: *** [Makefile:91: all] Error 2

Why can I not compile my project like this when I have enabled CUDA as a language in my CMakeLists.txt?

My NVCC version is 11.7 and my cmake version is 3.24.1

Answer 1

You can definitely include your test.h header. The problem is that test.h includes, in turn, <cuda_runtime_api.h>; so the C++ compiler must have the directory containing this CUDA header - the CUDA distribution include directory - in its include path; and for that to happen, CMake must make sure and add that directory to the include path via the compiler's command-line.

Note that for the NVCC compiler, the CUDA include directory is automatically searched without requiring specifying it on the command line; and NVCC is what's used by default to compile the test.cu file - so you would not get the error message when including test.h in test.cu.

To have CMake add the appropriate include directory when compiling C++, do this:

find_package(CUDAToolkit REQUIRED)

# ... etc. 

target_link_libraries(test CUDA::cudart)

or directly add the relevant include directory with target_include_directories() (not recommended).