SVD on a matrix that does not fit into memory

[aasdelat]

Who knows …
So, as conclusion, I think that your first answer (without the implementation of the mul_by_cols function) is the best one.

And the whole and definitive code (if you accept it) is the one I posted above:

@aasdelat

using HDF5, KrylovKit

fid = h5open("nc_file.nc", "r")
array = HDF5.readmmap(fid["nc_var_name"])
# Make lons and lats a single dimension
array = reshape(array, (size(array)[1] * size(array)[2], size(array)[3]))
array = transpose(array) # (time x space). It was (space x time)
vals, lvecs, rvecs, info = KrylovKit.svdsolve(array, 100, krylovdim=200)