[🐛 bug report] warp::square_to_von_mises_fisher([0.5, 0.5], ...) returns nan.

[LukasRuppert]

Summary

warp::square_to_von_mises_fisher returns nan, when the sample is exactly [0.5, 0.5].

System configuration

• Platform: openSUSE Tumbleweed (same issue on Ubuntu 20.04)
• Compiler: GCC 10.2.1
• Python version: 3.8.5
• Mitsuba 2 version: 2.2.1 (b87c11ace78e110b094b23b83ac5902a4773202d)
• Compiled variants:
  • scalar_rgb
  • spectral_rgb

Description

For the sample [0.5, 0.5], p will be exactly at the origin. Therefore, r2will be 0 and dividing 0/0 produces nan in p *= safe_sqrt((1.f - sqr(cos_theta)) / r2);. Instead, masked(p, r2 > 0.0f) *= safe_sqrt((1.f - sqr(cos_theta)) / r2); works fine, but there may be more elegant alternatives. Neat trick though to use the concentric disc sampling for vMFs :+1: .

Steps to reproduce

import mitsuba
mitsuba.set_variant('scalar_rgb')
mitsuba.core.warp.square_to_von_mises_fisher([0.5, 0.5], 1.0)

returns [-nan, -nan, 1] for any kappa != 0.0.