CMake and CUDA: Separate Compilation of Class Constructor on Device Fail During Linking

When working with CUDA and C++, it is often necessary to separately compile the class constructor for the device. This can be done using the __global__ keyword. However, if the constructor is not properly declared, it can lead to linking errors.

The Problem

The problem occurs when the class constructor is declared as follows:

__global__ MyClass::MyClass(int a, int b) {
  // ...
}

This declaration is incorrect because the __global__ keyword should only be used for functions that are executed on the device. The class constructor is not a function, so it should not be declared as __global__.

The Solution

To fix the problem, the class constructor should be declared as follows:

MyClass::MyClass(int a, int b) {
  // ...
}

This declaration is correct because it does not use the __global__ keyword.

Conclusion

By following these steps, you can avoid linking errors when separately compiling the class constructor for the device.