| 21 | |
| 22 | template<class T> |
| 23 | __global__ void wrapper(T* in, T* out) { |
| 24 | // kernels can be called in device code from other kernels |
| 25 | // (without launching a new kernel, so no dynamic parallelism!) |
| 26 | copy_kernel(in, out); |
| 27 | } |
| 28 | |
| 29 | |
| 30 | struct complex_struct { |