program in python, using numpy is disappointingly slow.

rdx · 11-17-2015, 08:05 AM

Is there some trick to telling python (2.7) about ATLAS? I wrote a function to find the intersection of a line and a plane, first in C++, then in Python and then in PHP. Even PHP, which is mathematically challenged, run 3x faster than Python.

Code:

def ray_test( ray, plane ) :
    d = ray.get_delta()
    if np.dot( d, plane.normal() ) > -1e-9 :
        return 0
    p0 = plane.get_p0()
    A = np.transpose( [-d, plane.get_p1() - p0, plane.get_p2() - p0] )
    return np.linalg.solve( A, ray.get_origin() - p0 )

As you can see, I call dot and cross products, transpose and solve, which should be passed on to the ATLAS library but it doesn't seem like it is. What am I doing wrong?

genss · 11-18-2015, 04:53 PM

idk python but isn't numpy made for such things ?

dugan · 11-19-2015, 09:35 AM

numpy needs to have ATLAS support compiled into it. Check that.

rdx · 11-19-2015, 11:39 AM

Quote:

Originally Posted by dugan

numpy needs to have ATLAS support compiled into it. Check that.

Okay, that sounds promising. I'm not sure how to do that but I'll look into it. The same function in C++:

Code:

#include    <armadillo>

vec ray_test( Ray3D &ray, Plane3D &plane )
{
    vec     d, p0, p1, p2, vi, vo ;
    mat     A ;

    d = ray.delta() ;
    if( ( dot( d, plane.normal() )) > -1e-9 )
        return vo ;
    p0 = plane.point( 0 ) ;
    p1 = plane.point( 1 ) - p0 ;
    p2 = plane.point( 2 ) - p0 ;

    A = join_rows( -d, p1 ) ;
    A = join_rows( A, p2 ) ;
    return solve( A, ray.origin() - p0 ) ;
}

This is more than 300x faster than python.

dugan · 11-19-2015, 12:56 PM

Quote:

Originally Posted by rdx

Okay, that sounds promising. I'm not sure how to do that but I'll look into it.

If you're building numpy from source, then you simply have ATLAS installed when you build numpy. While I haven't confirmed this personally, you would likely see in the console output that ATLAS was detected at compile time.

http://docs.scipy.org/doc/numpy/user/install.html

If you're downloading prebuilt numpy packages from your distribution's repositories, then they should have ATLAS as a dependency.

rdx · 11-20-2015, 08:53 AM

What I did was grab Numpy 1.9.1 from Slackbuilds.org and process it as usual ( bash numpy.Slackbuild) which compiles th source and creates a Slackware package, and then installpkg on the package it created. Atlas is installed as lapack-atlas-3.3.1-x86-64-1_SBo according to my package list ( in /var/log/packages ). I saw no indication on an option to include ATLAS nor any mention that it has done so. I'm pretty sure the Armadillo package and C++ are using it though.

dugan · 11-20-2015, 11:42 AM

I'm out of ideas. I'd recommend contacting the person who maintains those SlackBuilds and pointing the maintainer to this thread.

rdx · 11-21-2015, 12:50 PM

I suppose what I need to do is link the C++ version into a Python lib so I can import it and use it. I'll put that on my list of things to do and maybe, in a year ot two, ... lol

dugan · 11-22-2015, 03:47 AM

Quote:

Originally Posted by rdx

I suppose what I need to do is link the C++ version into a Python lib so I can import it and use it. I'll put that on my list of things to do and maybe, in a year ot two, ... lol

Or you could contact the SlackBuild maintainer like I recommended?