To compute the principal components using the SVD, I think you first want to center the data and compute something at least proportional to the covariance matrix of the data (not just using svd on the data matrix).
If you have the Statistics Toolbox, use pca() that is your best bet.
If you do not and must use svd(), then you really want the eigenvectors of the covariance matrix. In which case you can use eig()
