lgarrison

/*

This is a "minimal" reproducer for an apparent GCC-7 bug related to AVX-512.
It has a 100% reproduction rate on my machine (Cascade Lake).

A simple loop that copies the values from a double array to a float array
sometimes gives the wrong answer in the first few elements (zero, nan, or garbage).

The bug seems correlated with alignment values of the input and output
arrays, but it's not a one-to-one relationship.

lgarrison

lgarrison

lgarrison

lgarrison

# Note: this function uses the term 'radius', but the histogramming is completely general

import numpy as np

def RadialBin(radii, values, bin_edges, bin_indices=None):
    """
    Radial histogramming that supports fast re-evaluation
    for the same `radii` with new `values` by passing back `bin_indices`.
    
    Parameters

lgarrison

lgarrison

swot: https://github.com/jcoupon/swot/
corrfunc: https://github.com/manodeep/Corrfunc

To test the speed and agreement of swot vs corrfuc, we generated random
datasets of increasing size and measured the w(theta) autocorrelation function.

Swot has an approximation parameter 'OA' (opening angle) that critically impacts its runtime.
We tried OA=0.03 and OA='no'.  However, even the conservative value of OA=0.03
caused >10% errors in wtheta for this randomized configuration.  Thus, it should
be used with caution.