This is needed to ensure the cublas functions are executed in the same
stream as the rest of the CUDA kernels and thus are synchronous to the
other operations.