Speed up Point3D #1064
Speed up Point3D #1064
Conversation
| ) -> Result<(Vec<ResolvedAnnotationInfo>, Keypoints), QueryError> { | ||
| crate::profile_function!(); | ||
|
|
||
| let mut keypoints: Keypoints = HashMap::new(); | ||
|
|
||
| // No need to process annotations if we don't have keypoints or class-ids | ||
| if !entity_view.has_component::<KeypointId>() && !entity_view.has_component::<ClassId>() { |
There was a problem hiding this comment.
that makes soooo much sense, we should have that sort of thing everywhere!
Looking forward to making this a reusable pattern of sorts
| pub fn add_points( | ||
| &mut self, | ||
| count: usize, | ||
| positions: impl Iterator<Item = glam::Vec3>, | ||
| ) -> PointsBuilder<'_, PerPointUserData> { |
There was a problem hiding this comment.
This needs a docstring now. Does it only add the first count number of positions?
Can we use ExactSizeIterator instead? When we can just call positions.len()!
There was a problem hiding this comment.
Looks like using size_hint() from the standard iterator is a better option. This won't preclude us from using in-exact iterators, but will get us effectively the same benefit we have here.
There was a problem hiding this comment.
sigh looks like arrow2-convert doesn't implement size-hints, and adding a wrapper and doing the book-keeping seems like it potentially adds more overhead than we're saving. Going to add a docstring for now and came back to it later.
crates/re_viewer/src/ui/view_spatial/scene/scene_part/points3d.rs
Outdated
Show resolved
Hide resolved
I realized we were doing a bunch of processing related to annotations even when annotations weren't being used. This short-circuits that case which gives us a decent 33% speed-bump on NYU dataset (and any large point-clouds that don't use class-ids).
I did a bit of refactoring to remove the temporary Point-array. As far as I can tell iterating the point array and iterating the struct are "close enough" that it's not worth making the copy and we're just as good re-iterating the structure.
I also did a bit of refactoring to avoid unnecessary default initialization for points and user-data. These are now lazily added at the end if a user didn't add them.
I also patched NYUD demo to include class-ids to make sure this wasn't regressing our nominal case and it looks fine.
Before
After
Before w/ Class-ids
After w/ Class-Ids
Checklist
CHANGELOG.md(if this is a big enough change to warrant it)