Best way to improve the speed of a Python script

Scripting
, , ,
1

Dear all,

I am all new to Rhino / Grasshopper and would like to know how to improve the speed of a Python script.

I am particularly interested in 2 options:

  • using Numba and its @jit operator for parallelization
  • replacing all the for loops by matrix operations with Numpy

I haven’t tested the first option yet since I can’t seem to install “GH_CPython” on Rhino 6. I don’t know whether it is a compatibility issue or not, and if by chance @Mahmoud_Mohamed_Ouf reads this message, I would really appreciate to have some information regarding that matter.

I should also mention that the “GH_PythonRemote” component that I am currently testing instead doesn’t work with jit operators.

Now, the problem that I have with Numpy is when I need to convert numpy arrays of vectors to Rhino Geometry objects like Point3d. It takes forever.

In the example below I have a simple line growth algorithm where positions, distances and forces computation is entirely vectorized with numpy. Without converting the vectors array to Point3d the script runs faster than the regular for loop version and computes the position of about 2000 points under 100 ms per iteration (without RTrees or KDTrees). However, when I iterate over the numpy array to convert the vectors to Point3D the script slows down drastically (23 times slower in this example) and becomes much slower than the for loop version.

ss210%20(2)
ss210%20(2)1573×635 53 KB

definition: Line Growth (numpy).gh (8.7 KB)

script 
import Rhino.Geometry as rg
import ghpythonremote
import scriptcontext

np = scriptcontext.sticky['numpy']
sp = scriptcontext.sticky['scipy.spatial']
distance = sp.distance


def setup():
    
    global arr

    #Take every consecutive pair of points on a curve and insert a new one in between
    pts = np.array([[p.X, p.Y, p.Z] for p in iPoints])
    avg = (pts[:-1] + pts[1:]) * .5
    arr = np.insert(avg, range(pts.shape[0]), pts, 0)
    
    print "loaded"



if iDraw:
    push()
    subdivide()
    
    #Convert numpy vectors array to Point3d
    oPoints = [rg.Point3d(p[0], p[1], p[2]) for p in arr]
    
else:
    setup()



def subdivide():
    global arr
    
    #Check distances between each consecutive pair of points
    #insert a new point if distance is below the specified threshold
    d = np.sqrt(np.sum(np.diff(arr, axis=0)**2,1)) 
    idc = np.nonzero(d>iDiameter)[0]
    means = (arr[idc] + arr[idc+1]) * .5
    arr = np.insert(arr, idc+1, means, 0) 



def push():
    
    global arr

    #Computes distances matrix and set self-comparisons to NaN
    d = distance.cdist(arr, arr)
    np.fill_diagonal(d, None)
    
    #Find pairs of vectors whose separation distance is < iDiameter
    id1, id2 = np.nonzero(d<iDiameter)
    
    #Count (number of "too close" vectors for each vector)
    count = np.vstack(np.bincount(id1))
    
    #Difference between pairs of selected vectors
    dif = arr[id2] - arr[id1]
    
    #Distances between pairs of selected vectors
    dis = np.vstack(d[id1, id2])
    
    #Normalized differences 
    dif /= dis
    
    #Scaled differences
    dif *= .5 * (iDiameter - dis)
    
    #Average of differences (0 at start)
    avg = np.zeros([arr.shape[0], 3])
    
    #Indicates where to sum (indices)
    idx = np.flatnonzero(np.concatenate(([1], np.diff(id1))))
    
    #Average sums of differences
    avg[np.unique(id1)] = np.add.reduceat(dif, idx, axis = 0) / count

    #Push
    arr -= avg

Questions:

  • Is there a faster way to convert the numpy array to Point3d ?
  • If not, how would you use numpy to speed-up this script ?
  • Do you know another / better way to improve the speed of this script ?

Any help, guidelines or hints, would be greatly appreciated.

If something isn’t clear about my question, please let me know.

2

Hello,

Is it possible to turn them into a point cloud? That should allow you to pass all the points in one go… in theory…?

3

Hi Graham, thank you for the reply.

I am not sure to understand, do you mean converting the numpy array to a Rhino Geometry PointCloud ? If so, I don’t know how to do that.

I see a PointCloud Constructor but have no idea how to transfer the floats of the numpy array to it.

4

Ahh yes - looks like you need point3ds to go into it so that’s no help

5

Are you using something like ghpythonremote.obtain(arr.tolist())? I don’t see it in the script you posted…perhaps I missed it or it is somewhere else in the definition. The usage notes indicate creating remote array-like objects will be slow unless you use .deliver(). Maybe the inverse is also true hence obtain(). Quick look at the ghpythonremote source shows it is probably using rpyc to connect to the cpython instance, which won’t be fast if it has to make a call for every element in the numpy array. I think making oPoints is the only time you visit each member of the array. Have you tried something simple on p like sum() to see if that is slow too?

1 Like
6

You could try Grasshopper.Kernel.Types.GH_Point(Rhino.Geometery.Point3d()) instead of Rhino.Geometery.Point3d() (cf. doc). As I understand, this should be at least a little faster, since it casts directly to a Grasshopper type point, instead of doing that conversion at the component output.

Maybe multithreading?

What’s your script about? Sphere/circle relaxation?

1 Like
8

Thanks so much for all your replies and suggestions.

Nathan was right, I needed to first send the list to the remote interpreter with ghpythonremote.deliver(), then recover it with ghpythonremote.obtain().

The script runs much faster now, around 20 times faster than the non-vectorized version.

Thanks again.

1 Like