In Numeric the row and columns are implemented asymmetrically. In MatPy we try to provide both types of operations by transposing both the operands and the results. I do not know the cost of these overheads. In the future it might be worthwhile to implement both versions in C.
Because the types of NumPy and Mat do not match exactly, there are many conditional statements in the initializers. Such costs will be eliminated when these classes are implemented natively.
The matrix functions like sqrtm, expm, logm, etc, are implemented by eigenvalue decomposition through an intermediate mfunc. It looks worthwhile to re-implement this in C.