The symbol argument in cudaMemcpyFrom/ToSymbol

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi, I observed that when using cudaMemcpyFrom/ToSymbol series of API, one pass the symbol by name (reference?) but if the same API is wrapped in some helper function, then one pass by pointer.

__constant__ int C;

namespace helper {
cudaError_t cudaMemcpyToSymbolAsync(const void* symbol,
                                    const void* src,
                                    size_t count,
                                    size_t offset,
                                    cudaMemcpyKind kind,
                                    cudaStream_t stream) {
  return ::cudaMemcpyToSymbolAsync(symbol, src, count, offset, kind, stream);
}
} // namespace helper

int main() {
  cudaStream_t s = 0;
  int h{10};
  assert(cudaSuccess == cudaMemcpyToSymbolAsync(C, &h, sizeof(int), 0, cudaMemcpyHostToDevice, s));
  assert(cudaSuccess == helper::cudaMemcpyToSymbolAsync(&C, &h, sizeof(int), 0, cudaMemcpyHostToDevice));
}

First it otherwise would not compile, because passing an int for const void* is disallowed.
Secondly, it is a interesting that the “target” pointer for the memcpy is a const*.

How come C can pass to a const void* argument in the first case, but not the second case and we need to address-of- C?