What are the real C++ language constructs supported by CUDA device code?

Question

Appendix D of the 3.2 version of the CUDA documentation refers to C++ support in CUDA device code.
It is clearly mentioned that CUDA supports "Classes for devices of compute capability 2.x". However, I'm working with devices of compute capability 1.1 and 1.3 and I can use this feature!

For instance, this code works:

// class definition voluntary simplified
class Foo {
  private:
    int x_;

  public:
    __device__ Foo() { x_ = 42; }
    __device__ void bar() { return x_; }
};


//kernel using the previous class
__global__ void testKernel(uint32_t* ddata) {
    Foo f;
    ddata[threadIdx.x] = f.bar(); 
}

I'm also able to use widespread libraries such as Thrust::random random generation classes. My only guess is that I'm able to do so thanks to the automatic inlining of __device__ marked function, but this does not explain the handling of member variables withal.

Have you ever used such features in the same conditions, or can you explain to me why my CUDA code behaves this way? Is there something wrong in the reference guide?

Answer 1

Oficially, CUDA has no support for classes on devices prior to 2.0.

Practically, from my experience, you can use all C++ features on all devices as long as the functionality can be resolved at compile-time. Devices prior to 2.0 do not support function calls (all functions are inlined) and no program jumps to a variable address (only jumps at constant address).

This means, you can use the following C++ constructs:

• Visibility (public/protected/private)
• non-virtual inheritance
• whole template programming and metaprogramming (until you stuble on nvcc bugs; there are quite a few of them as of version 3.2)
• constructors (except when object is declared in __ shared __ memory)
• namespaces

You cannot use the following:

• new & delete operators (I believe devices >=2.0 can do that)
• virtual methods (requires jumps at variable address)
• function recursion (requires function calls)
• exceptions

Actually, all examples in chapter D.6 of the CUDA Programming Guide can compile for devices <2.0