Got exactly the same problem on CUDA 1.0. However the 1.1 documentation says that the symbol could either be a name or a variable residing in global memory. So it seems we cannot get adresses of constant memory variables. Indeed
Thanks for your help. Using memcpytosymbol seems do have done the trick, gave a 40% speedup, and worked correctly. Still only manages to encrypt 650mbit/s tho.
Am I right that cudaGetSymbolAddress() cannot fetch the address of a constant variable and thus returns a cudaErrorAddressOfConstant error, and that, in contrast, cudaGetSymbolSize() is able to return the size of a constant variable?