Timo Röhling

I submitted a similar PR last year, where I replaced the regex with proper unicode character class queries: #124

AFAICT the C module relies on the libyaml implementation, which does not use regular expressions but a simple [range check](https://github.com/yaml/libyaml/blob/690a781ef6af70ce6749d6e2be91743345123998/src/reader.c#L411)

Appendix D of the [HEALPix Introduction](https://healpix.sourceforge.io/pdf/intro.pdf) states: > Note that, due to finite precision of floating-point arithmetics and differences between > numerical libraries, the result of HEALPix functions like ang2pix...

That's why I proposed the forward-backward round trip through both implementations, which effectively shows that they behave equivalently (up to FP precision).

@mreineck Given that the functions are numerically unstable if provided with vectors which are on (or very close to) a pixel border, I see two options: 1. Use your proposed...

Unfortunately, I was not involved in the decision to drop Open3D from Ubuntu Noble. I suspect they have problems with PyTorch, and I recently added PyTorch support to the Debian...

I already [have a PPA](https://launchpad.net/~roehling/+archive/ubuntu/open3d) for Ubuntu 20.04, so yes, I am planning to host Open3D binaries for 24.04 there, as a fallback solution.

I'm currently working on a PPA release. I have to fix Filament first (which was broken by a new glslang release), but I expect to be done by the end...

Open3D 0.18 for Ubuntu Noble will be available from my [PPA](https://launchpad.net/~roehling/+archive/ubuntu/open3d) as soon as Launchpad has finished building the binary packages.

> The main point of these tests is to ensure that the SIMD is actually bit-exact with the non-SIMD code on a specific CPU, rather than prove out any kind...