@luispedro I did not see your reply, hence the duplication. I see that Milk isnt doing too well on pca (11 times slower than the best performer). From the code it seems you are importing numpy.linalg. I think if you import scipy.linalg your results will be better without any change to the algorithm (unless you have built numpy with ATLAS or MKL).
Scipy.linalg links with underlying BLAS if its avaliable whereas the standard build of numpy implements linalg on its own (but it is possible to override that). Second point, I think for large data sets you are better of giving an api for computing an user specified number of principal components rather than all of them. Thirdly, if you find limiting yourself to gcc not too restrictive, you can stick with stl style algos in place of c++ looping. The advantage is that you will get parallelism for free http://gcc.gnu.org/onlinedocs/libstdc++/manual/parallel_mode...
Thanks for the tips. I know that the PCA implementation is very naïve, but I never put as much work into it as into some of the other parts of the library.
https://github.com/luispedro/milk
The main documentation for the package is actually at: