RoSt RoSt - 5 months ago 44
Python Question

Memory usage doubled when passing matrix to shared object

I have a linear set of equations, where
A x = b and A is a large matrix and b is known as well.
The matrix A is set up with python.
Now I want to invert matrix A to get x.

A and b are passed to a Fortran 90 program via a shared object. I compiled the Fortran program using numpy.f2py:

import numpy.f2py.f2py2e as f2py2e
import sys, os
sys.argv += "-lmkl_rt -c -m MKL_MODULE MKL_WRAPPER.f90".split()

Finally, I call the f90 subroutine:

MKL_MODULE.mkl_wrapper.call_dgelsd(A, b, np.shape(A)[0], np.shape(A)[1])

When calling the fortran program, the memory usage doubles, apparently due to an internal copy of the matrix A and b.
However, once I have the vector x I'm not interested in A or b anymore.
Is there any way to avoid that internal copy and passing A to the fortran program?

I already had the idea of saving A and b to the HD and reading it from the Fortran program, but this takes very long time and is not really an option for matrices of the size I'm dealing with.


No internal copy will be made if the arrays are in F-order

[How to force numpy array order to fortran style? to-fortran-style][1]