Although you can think of these specialized functions as wrappers around cudaMemcpy, they are well-optimized to transfer both strided and unstrided data.