Issue with numpy.interp

[herbelj]

I found an issue when using the function numpy.interp within a Python function compiled using HOPE. The issue is that interpolation produces different results depending on whether HOPE is used or not. The following example illustrates this.

    import numpy as np
    import hope
    import pickle

    def interpolation(x, xp, fp):
        intp = np.interp(x, xp, fp)
        return intp

    @hope.jit
    def interpolation_hope(x, xp, fp):
        intp = np.interp(x, xp, fp)
        return intp

    with open('np.interp_github_issue.pkl', 'r') as f:
        x, xp, fp = pickle.load(f)
    x = x.astype(np.float32)
    xp = xp.astype(np.float32)
    fp = fp.astype(np.float32)

    intp = interpolation(x, xp, fp)
    intp_hope = interpolation_hope(x, xp, fp)
    np.allclose(intp, intp_hope)

np.interp_github_issue.pkl.zip

[cosmo-ethz]

Thanks for reporting this. Seems like HOPE is not handling the boundaries correctly

import numpy as np
import hope

def interpolation(x, xp, fp):
    intp = np.interp(x, xp, fp)
    return intp
interpolation_hope = hope.jit(interpolation)

xp = np.arange(5,15)
fp = -xp
x = np.arange(0, 20)

intp = interpolation(x, xp, fp)
intp_hope = interpolation_hope(x, xp, fp)
np.allclose(intp, intp_hope)

[cosmo-ethz]

@herbelj I modified the bounds checking in the HOPE interp implementation. May I ask you to checkout the interp_bounds branch and quickly check if this resolves your problem? Thanks

[herbelj]

@cosmo-ethz The problem is fixed now, thanks! :)