Which you use depends on which definition of covariance you wish to employ. Matalb's documentation for the cov() function clearly writes how it computes covariance. The signals' means are removed and there is a bias correction factor (1/N-1).
Covariance is a single quantity, not a matrix. You can compute the covariance matrix, which is contains the variances along the diagonal and the covariances where i~=j.
Where did you read about d*d' being covariance? There is no covariance if there is only one signal (just variance).
