Speed up query usage in heuristics by splitting DataQuery into separate query and update_properties passes #4563
Conversation
jleibs
added
📉 performance
Wumpf
force-pushed
the
jleibs/skip_empty_queries
branch
from
3f0cd4a to
f3e06ae
Compare
Wumpf
force-pushed
the
jleibs/optmize_queries
branch
from
80d3841 to
5a0d21e
Compare
Wumpf
left a comment
Wumpf
left a comment
There was a problem hiding this comment.
I think the code needs to clear up better what's fast vs normal. Looks & feels very hacky right now. Let's maybe talk it through a bit
| /// Called at the start of each frame | ||
| pub fn on_frame_start(&mut self, item_retain_condition: impl Fn(&Item) -> bool) { | ||
| re_tracing::profile_function!(); | ||
| re_tracing::profile_scope!("SelectionState::on_frame_start"); |
There was a problem hiding this comment.
isn't that esactly what the profile_function macro prints, no?
There was a problem hiding this comment.
Not quite -- it seems like puffin doesn't track the struct name and so I was seeing two calls to on_frame_start coming from different call-sites resulting in some unexpected aggregation.
There was a problem hiding this comment.
Is that with latest puffin? afaik what scopes are merged when improved a bit, but might have been mostly about traits
jleibs
force-pushed
the
jleibs/optmize_queries
branch
2 times, most recently
from
0329e12 to
18571c5
Compare
jleibs
force-pushed
the
jleibs/optmize_queries
branch
from
18571c5 to
b822d31
Compare
jleibs
changed the title
Wumpf
left a comment
Wumpf
left a comment
There was a problem hiding this comment.
so much better, flow is totally clear now! The only wrinkle now is the thing you already put in the comments of having the same query results object with and without filled out properties
| // in on_frame_start, but on_frame_start also needs the queries. | ||
| let updated_query_results = { | ||
| { | ||
| re_tracing::profile_scope!("updated_query_results"); |
There was a problem hiding this comment.
nit: more like "update_overrides" now
What
After adding some scopes, found some more wins on query processing. This is a mixture of some optimized short-circuit early-outs, and doing the PropertyOverride logic on a second-pass in order to avoid any overhead related to creating the EntityProperty structures and override entity-paths.
This has the added benefit of also removing the second query and replacing it with a much easier-to-follow override processor.
Before:
After:
Checklist
mainbuild: app.rerun.ionightlybuild: app.rerun.io