Comparison to instant-rm

[ogencoglu]

Thanks for the interesting work!

This is not an issue as such but can you shortly summarize the differences to NVlabs/instant-rm?

[jeertmans]

Hi! Thanks for your question!

NVlabs/instant-rm is a very interesting work that sits between Sionna v0.19 and Sionna-RT v1.0, with the aim to generate radio map very fast.

DiffeRT is rather is more general-purpose ray tracer for radio propagation. Speed is also an important factor, but I don't think it can match Instant RM when it comes to generating radio maps. The main reason is that Instant RM (and also Sionna-RT) uses some kind of Ray Launching technique to generate the radio maps. DiffeRT also offers a way to perform Ray Launching (via the method="sbr" in TriangleScene.compute_paths, but I did not spend much time optimizing it. E.g., DiffeRT currently uses ray-sphere intersection to check if a ray is close-enough to a receiver, whereas Sionna-RT (and Instant RM) will use ray-plane intersection and later tile the plane into a grid of receivers.

In conclusion, I would say that DiffeRT is closer to Sionna-RT than it is to Instant RM. I think Instant RM can be seen as a specialized implementation of Sionna-RT for radio maps.

Regarding speed, both DiffeRT and Sionna-RT rely on just-in-time compilation to produce fast and GPU-compatible code. Both solutions should be fast, but one may observe differences (if you do, please raise an issue!). Both solutions are also fully differentiable, and DiffeRT might even be more differentiable as it exposes the internal mesh better, whereas Sionna-RT relies on Mitsuba, where scene primitives (i.e., triangles position) are more complex to get.

Let me know if you have any other question :-)

[ogencoglu]

Thank you for the swift reply!