Rerun is extremely slow when updating a root pose

[Danvil]

Describe the bug
rerun app is very slow and "laggy" when updating a root transform frame at 100 Hz.

To Reproduce

1. Use Rust 1.80, Linux and rerun 0.18.2
2. Publish a pose at 100 Hz: rr.log("root", rerun::Transform3D::from_translation_mat3x3(...))?;
3. Publish some other data attached to the pose at 10 Hz: rr.log("root/points", rerun::Points3D::new({...}))?;
4. It takes about 130 ms to render a single frame on a fairly powerful desktop.

Expected behavior
100 Hz pose update should be blazingly fast.

Screenshots
Some examples from the profiler widget:

[emilk]

Thanks for reporting!

Do you have a sample .rrd for us to take a look at? Or a full sample code to reproduce?

How many entities are there?

Are you compiling with --release?

[Danvil]

Are you compiling with --release?

All my programs are compiled with --release.
For rerun I use cargo install and have this:

$ rerun --version
rerun-cli 0.18.2 [rustc 1.80.0 (051478957 2024-07-21), LLVM 18.1.7] x86_64-unknown-linux-gnu

Do you have a sample .rrd for us to take a look at?

Please find a a RRD file here
Note also that it took about 5 minutes to ingest this file into rerun ..

[Danvil]

@emilk Any thoughts on this? Happy to provide more information as necessary

[Wumpf]

I'm working a bit on transform performance right now and the rrd is indeed proving useful as a a case for slow transforms. That said it's by far not the only thing that is slow in that scene (and I plan to work on that as well! :)).

What is however a bit confusing to me is this bit

Note also that it took about 5 minutes to ingest this file into rerun ..

I'm running on an M1 Max (which tbf is a very powerful machine) but even though I'm getting nowhere near that slow ingestion time on either latest Rerun or 0.18.2 as described, the entire loading is way under a second.
I used the official python build, installed via pip install rerun-sdk==0.18.2.

@Danvil could you please share what exact build (or how you built it) of Rerun you used and on which machine you ran this?

---

For posterity - mostly to keep better track myself:

On M1 Max @ rerun 0.20.2 (!!, that's not the originally posted version) I get about 23ms cpu time both while playing & at stills. Profiler shows about 8ms of that being in the transform context, but there's a ton of profiling overhead.

On a WIP branch (f170086, as of writing that's tagged as 0.21-alpha.0; not known yet if it lands in 0.21!) I'm down to 14ms while playing & at stills so far. Profiler now with much less reduced overhead (but still significant in places) tells me that transform context is now 0.07ms