[Merged by Bors] - Add a method `iter_combinations` on query to iterate over combinations of query results #1763

Frizi · 2021-03-26T00:26:51Z

With const generics, it is now possible to write generic iterator over multiple entities at once.

This enables patterns of query iterations like

for [e1, e2, e3] in query.iter_combinations() {
   // do something with relation of all three entities
}

The compiler is able to infer the correct iterator for given size of array, so either of those work

for [e1, e2] in query.iter_combinations()  { ... }
for [e1, e2, e3] in query.iter_combinations()  { ... }

This feature can be very useful for systems like collision detection.

When you ask for permutations of size K of N entities:

if K == N, you get one result of all entities
if K < N, you get all possible subsets of N with size K, without repetition
if K > N, the result set is empty (no permutation of size K exist)

mockersf · 2021-03-26T00:30:50Z

I'm guessing this would need Rust 1.51 - just a note to keep track of the minimum supported rust version

Also, this seems like a nice generalisation of #1263

Frizi · 2021-03-26T00:32:43Z

crates/bevy_ecs/src/query/iter.rs

+            'outer: for i in (0..K).rev() {
+                match self.cursors[i].next(&self.tables, &self.archetypes, &self.query_state) {
+                    Some(_) => {
+                        // walk forward up to last element, propagating cursor state forward
+                        for j in (i + 1)..K {
+                            self.cursors[j] = self.cursors[j - 1].clone();
+                            match self.cursors[j].next(
+                                &self.tables,
+                                &self.archetypes,
+                                &self.query_state,
+                            ) {
+                                Some(_) => {}
+                                None if i > 0 => continue 'outer,
+                                None => return None,
+                            }
+                        }
+                        break;
+                    }
+                    None if i > 0 => continue,
+                    None => return None,
                }


I'm not a fan of this complex double loop, and I believe this can be written simpler. Just wasn't able to simplify it more for now. If you have any idea how to do that, I'm all ears.

alice-i-cecile · 2021-03-26T00:36:45Z

I'd love to have a short user facing example for this. It's very useful non-obvious functionality that will be hard to discover otherwise.

Frizi · 2021-03-26T22:43:23Z

Also, this seems like a nice generalisation of #1263

I don't believe it's a generalization, but those could work together.

The linked PR is about getting a singular entity from the query. The permutation iterator doesn't allow you to do so. But it does allow getting a (mutable) reference to multiple entities from a single query at once. You can do k_iter::<5>().next() to get references to all of them at once, but there is no guarantee there aren't more.

If we would generalize #1263 to support this, we would have something like this:

fn exactly_k::<const K: usize>(&self) -> Result<[<Q::Fetch as Fetch<'_>>::Item; K], QueryExactlyError>

Which would basically do k_iter::<K>().next() and check if there are no more or less entities than expected.

Is this worth adding to this PR? The implementation of that wouldn't be terribly complicated, given that k_iter is working. But it is a separate bit of functionality, and I would rather avoid another bikeshedding over the function name :) We can always add that separately.

alice-i-cecile · 2021-03-27T01:19:31Z

I don't mind the idea of exactly_k, but a) don't see any common use case for it b) think it's definitely out of scope for this PR. We can consider it in the future if we get user requests for it.

Frizi · 2021-03-27T15:18:24Z

added a simple example with n-body simulation

bjorn3 · 2021-03-27T16:40:42Z

I think the name iter_permutations would be more descriptive than k_iter.

alice-i-cecile · 2021-03-27T17:06:12Z

The n_bodies example is cool! It definitely doesn't belong in the 3D folder though: that's for 3D rendering / assets mostly.

In its current form, it would be a good fit for the ecs folder, renamed to iter_permutations. My preference would be to flesh it out a little more (create some basic 3D meshes, a light source and add a flycam) and put it in the games folder though :) We could use a couple more complete examples, especially in 3D.

Wherever it ends up, make sure to update the examples/README.md to make sure it's discoverable.

Frizi · 2021-03-27T17:47:11Z

I think the name iter_permutations would be more descriptive than k_iter.

Agree, will change it that way.

My preference would be to flesh it out a little more (create some basic 3D meshes, a light source and add a flycam) and put it in the games folder though :)

I don't exactly want to write the full example game just yet :) Maybe after this PR lands the example could be extended.

But I agree that in current form it fits into ECS. That was also my initial idea, but I changed my mind after seeing no other examples with 3d rendering in ecs folder. I will move it now.

TheRawMeatball · 2021-03-27T17:59:22Z

Hmm, should we use iter_combinations to have used the correct word mathematics-wise, or is permutations fine?

alice-i-cecile · 2021-03-27T18:12:57Z

Hmm, should we use iter_combinations to have used the correct word mathematics-wise, or is permutations fine?

Very much agree with iter_combinations, since order doesn't matter. That's the correct behavior for 99% of the use cases.

Using the right name is much less confusing and lets us use iter_permutations for the order-matters behavior later if we need it.

Frizi · 2021-03-27T18:13:19Z

Is there anything that could be done about that Android build failing? It is clearly caused by too old rust version. All other builders are already updated and passing.

alice-i-cecile · 2021-03-27T18:16:59Z

For consistency with iter_tools, do we want this to be a method on QueryIter or the like? I think that's the more idiomatic design and would be easier to discover.

Frizi · 2021-03-27T18:22:03Z

Hmm, should we use iter_combinations to have used the correct word mathematics-wise, or is permutations fine?

Agree again, iter_combinations is more accurate to what's happening here. Doc comments even talk about combinations.

Frizi · 2021-03-27T18:30:25Z

For consistency with iter_tools, do we want this to be a method on QueryIter or the like? I think that's the more idiomatic design and would be easier to discover.

I'm not sure how itertools come to the picture here, but QueryIter is a separate type from QueryCombinationIter. Those share some common implementations details, but those are fundamentally separate iterator types. In theory QueryIter could be a special case of QueryCombinationIter with K = 1, but that is likely to cause unnecessary performance regression. It's easier to optimize the base case.

alice-i-cecile · 2021-03-27T18:40:41Z

I'm not sure how itertools come to the picture here,

It's just the common ecosystem crate for this sort of thing.

QueryIter is a separate type from QueryCombinationIter. Those share some common implementations details, but those are fundamentally separate iterator types. In theory QueryIter could be a special case of QueryCombinationIter with K = 1, but that is likely to cause unnecessary performance regression. It's easier to optimize the base case.

Okay, good to know. Should be fine to leave as is then :)

Frizi · 2021-04-24T23:03:15Z

Fixed the doc comments. Should be ready to go.

cart

Alrighty this looks good to me! It definitely fills a painful functionality gap. I did my two regular benchmarks for ECS changes:

A clean build of Bevy looking for significant compile time regressions. Nothing measurable to report here, which is in line with my expectations for a change like this.
A run of my fork of ecs_bench_suite. QueryIter is notoriously subject to the whims of the compiler and sometimes even minor structural changes can cause regressions. Unfortunately this regresses a few benchmarks enough to merit some investigation:

This is surprising given the relatively superficial QueryIterationCursor refactor. Structurally and functionally it seems pretty much identical. I think its worth experimenting with adding/removing inlines here and there to see if we can regain whatever optimization magic the compiler is currently giving us.

crates/bevy_ecs/src/query/iter.rs

cart · 2021-04-30T23:58:36Z

crates/bevy_ecs/src/query/state.rs

+    }
+
+    #[inline]
+    pub fn iter_combinations_mut<'w, 's, const K: usize>(


I think this could use docs explaining the quirks of how to use it and the current rust limitations that force it to work that way (ex: how to write the while loop that consumes the iterator, why we can't safely implement Iterator for mutable combinations, and why for_each (which should be the favored implementation) can't work due to rustc errors).

I expect this to be a source of confusion, so we should try to nip it in the bud.

This is thoroughly documented on Query::iter_combinations_mut. I wonder if I should repeat those docs here as well. I specifically documented in on Query, as this is how other docs are done as well. All other QueryState methods are left undocumented except for safety, so I was just following an established pattern.

Hmm yeah thats reasonable. QueryState is the api people doing direct World queries will use, so it should be documented, but its also unfair to hold you to a standard that I didn't hold myself to :)

I created a new issue to track this work #2090

Frizi · 2021-05-03T16:22:47Z

I tried looking at the benchmarks and finding the right combinations of inlines, but it turns out for some reason the benchmarks are not very repeatable for me. After running literally the same benchmark multiple times on main branch, I'm getting regressions and improvements as high as 40%. Without any changes at all (not rebuild). Maybe my machine is particularily noisy at this time. I'm running Windows, who knows what kind of crap is being run in background. If that's the case and the benchmark is predictible for others, then I would need help with determining the cause of this regression (if it is indeed there).

For reference, here is one of the runs testing bevy main branch. It is comparing against previous run of the exact same code.

cart · 2021-05-03T19:02:34Z

The "insert" benchmarks do vary wildly for me as well. I think this is because its largely a benchmark of the os memory allocator, which I would expect to swing around based on the current memory usage / cached allocations.

The iterator benchmarks (edit: don't) swing around for me in the same way. Usually they stay "within noise threshold" for me.

alice-i-cecile · 2021-05-03T21:28:28Z

Is there a way to convince the benchmarking library to intersperse the two versions? That would help control for the random noise, especially when combined with increasing the sample size.

cart · 2021-05-04T19:13:51Z

@alice-i-cecile : not that I know of. Criterion is a "compiled in" orchestrator, so it would need to be a post-processing step that somehow consolidates two sets of results.

Can someone that isn't me try to validate the regression here? Ex: run the benchmark suite once on main, then again on this branch (rebased on main to control for recent changes).

cart · 2021-05-04T19:15:13Z

I can try to find time to play around with optimizations, but that won't be for at least a couple of weeks while we sort out rendering.

cart · 2021-05-04T19:17:10Z

I think "worst case scenario" we can just use the old impl for normal iterators and just eat the code-duplication complexity from the added "cursor" impl.

… regression

Frizi · 2021-05-11T20:28:08Z

I ended up focusing on a few benchmarks that I was able to reliably test against

fragmented_iter/bevy
fragmented_iter/bevy_foreach
sparse_fragmented_iter/bevy
sparse_fragmented_iter/bevy_foreach

The foreach variants don't actually excersize changed code, so I was kinda using them as a control group. I've only looked at benchmarks where those had no performance change detected. And I indeed have seen the regression at around 70% worse performance in fragmented_iter/bevy.

Unfortunately, no amount of adding or removing inline or inline(force) was able to remove that regression. I ended up just inlining it manually and it totally brought it back to previous performance. We are probably just on the verge of some very useful compiler heurestic right now, and this leaves me somewhat uneasy about that whole situation. I wonder if we can do better, to allow abstracting this code away and keep the performance. But for now, it is just manually duplicated.

cart · 2021-05-17T23:26:36Z

I think its worth considering using a macro to cut down on code duplication here, but I don't see that as a prerequisite for merging. I added comments to remind devs to update iterators in multiple places (and a reference to this issue explaining why the code isn't unified).

std uses a macro for cases like this:

cart · 2021-05-17T23:26:49Z

bors r+

Frizi · 2021-05-17T23:27:34Z

I see the added comments about duplicated code, and while this is the only exact copy, those aren't really the only similar ones. There is also QueryState::for_each_unchecked_manual and QueryState::par_for_each_unchecked_manual which share a lot of the same logic.

cart · 2021-05-17T23:27:55Z

bors r-

…s of query results (#1763) Related to [discussion on discord](https://discord.com/channels/691052431525675048/742569353878437978/824731187724681289) With const generics, it is now possible to write generic iterator over multiple entities at once. This enables patterns of query iterations like ```rust for [e1, e2, e3] in query.iter_combinations() { // do something with relation of all three entities } ``` The compiler is able to infer the correct iterator for given size of array, so either of those work ```rust for [e1, e2] in query.iter_combinations() { ... } for [e1, e2, e3] in query.iter_combinations() { ... } ``` This feature can be very useful for systems like collision detection. When you ask for permutations of size K of N entities: - if K == N, you get one result of all entities - if K < N, you get all possible subsets of N with size K, without repetition - if K > N, the result set is empty (no permutation of size K exist) Co-authored-by: Carter Anderson <[email protected]>

bors · 2021-05-17T23:28:08Z

Canceled.

cart · 2021-05-17T23:33:06Z

Updated!

cart · 2021-05-17T23:33:32Z

bors r+

…s of query results (#1763) Related to [discussion on discord](https://discord.com/channels/691052431525675048/742569353878437978/824731187724681289) With const generics, it is now possible to write generic iterator over multiple entities at once. This enables patterns of query iterations like ```rust for [e1, e2, e3] in query.iter_combinations() { // do something with relation of all three entities } ``` The compiler is able to infer the correct iterator for given size of array, so either of those work ```rust for [e1, e2] in query.iter_combinations() { ... } for [e1, e2, e3] in query.iter_combinations() { ... } ``` This feature can be very useful for systems like collision detection. When you ask for permutations of size K of N entities: - if K == N, you get one result of all entities - if K < N, you get all possible subsets of N with size K, without repetition - if K > N, the result set is empty (no permutation of size K exist) Co-authored-by: Carter Anderson <[email protected]>

bors · 2021-05-18T00:10:11Z

Pull request successfully merged into main.

Build succeeded:

…s of query results (bevyengine#1763) Related to [discussion on discord](https://discord.com/channels/691052431525675048/742569353878437978/824731187724681289) With const generics, it is now possible to write generic iterator over multiple entities at once. This enables patterns of query iterations like ```rust for [e1, e2, e3] in query.iter_combinations() { // do something with relation of all three entities } ``` The compiler is able to infer the correct iterator for given size of array, so either of those work ```rust for [e1, e2] in query.iter_combinations() { ... } for [e1, e2, e3] in query.iter_combinations() { ... } ``` This feature can be very useful for systems like collision detection. When you ask for permutations of size K of N entities: - if K == N, you get one result of all entities - if K < N, you get all possible subsets of N with size K, without repetition - if K > N, the result set is empty (no permutation of size K exist) Co-authored-by: Carter Anderson <[email protected]>

alice-i-cecile added A-ECS Entities, components, systems, and events C-Feature A new feature, making something new possible C-Usability A targeted quality-of-life change that makes Bevy easier to use labels Mar 26, 2021

Frizi commented Mar 26, 2021

View reviewed changes

Frizi force-pushed the query-k-iter branch 3 times, most recently from 58b332a to fd71bef Compare March 26, 2021 22:31

Frizi force-pushed the query-k-iter branch 4 times, most recently from 40785c3 to 08f100c Compare March 27, 2021 16:36

Frizi changed the title ~~Add a method k_iter on query to iterate over permutations of query results~~ Add a method iter_permutations on query to iterate over permutations of query results Mar 27, 2021

Frizi changed the title ~~Add a method iter_permutations on query to iterate over permutations of query results~~ Add a method iter_combinations on query to iterate over combinations of query results Mar 27, 2021

Frizi changed the title ~~Add a method iter_combinations on query to iterate over combinations of query results~~ Add a method iter_combinations on query to iterate over combinations of query results Mar 27, 2021

alice-i-cecile mentioned this pull request Mar 29, 2021

[Question/PR] Should We Increase the Max Query Set Size? #1784

Closed

finish sentence in the QueryCombinationIter comment

901c8d0

alice-i-cecile mentioned this pull request Apr 29, 2021

Safe method for multiple mutable references to components from Query #2042

Closed

cart reviewed May 1, 2021

View reviewed changes

label comments about unstable methods as todo

93d7482

manually inline iteratrion cursor into QueryIter to avoid performance…

06ba263

… regression

alice-i-cecile mentioned this pull request May 15, 2021

RFC: Add a standard trait for getting many &mut to places rust-lang/rfcs#3100

Open

Add notes reminding devs to update iter code in both places

08c8ebd

add reminder comments to all query iterator impls

575fe00

bors bot changed the title ~~Add a method iter_combinations on query to iterate over combinations of query results~~ [Merged by Bors] - Add a method iter_combinations on query to iterate over combinations of query results May 18, 2021

bors bot closed this May 18, 2021

CleanCut mentioned this pull request Jan 9, 2022

permutations -> combinations bevyengine/bevy-website#250

Merged

[Merged by Bors] - Add a method iter_combinations on query to iterate over combinations of query results #1763

[Merged by Bors] - Add a method iter_combinations on query to iterate over combinations of query results #1763

Conversation

Frizi commented Mar 26, 2021 • edited Loading

mockersf commented Mar 26, 2021

Frizi Mar 26, 2021

Choose a reason for hiding this comment

alice-i-cecile commented Mar 26, 2021

Frizi commented Mar 26, 2021 • edited Loading

alice-i-cecile commented Mar 27, 2021

Frizi commented Mar 27, 2021

bjorn3 commented Mar 27, 2021

alice-i-cecile commented Mar 27, 2021 • edited Loading

Frizi commented Mar 27, 2021

TheRawMeatball commented Mar 27, 2021

alice-i-cecile commented Mar 27, 2021

Frizi commented Mar 27, 2021

alice-i-cecile commented Mar 27, 2021

Frizi commented Mar 27, 2021

Frizi commented Mar 27, 2021

alice-i-cecile commented Mar 27, 2021

Frizi commented Apr 24, 2021

cart left a comment • edited Loading

Choose a reason for hiding this comment

cart Apr 30, 2021

Choose a reason for hiding this comment

Frizi May 2, 2021

Choose a reason for hiding this comment

cart May 3, 2021

Choose a reason for hiding this comment

Frizi commented May 3, 2021

cart commented May 3, 2021 • edited Loading

alice-i-cecile commented May 3, 2021

cart commented May 4, 2021

cart commented May 4, 2021

cart commented May 4, 2021

Frizi commented May 11, 2021 • edited Loading

cart commented May 17, 2021

cart commented May 17, 2021

Frizi commented May 17, 2021 • edited Loading

cart commented May 17, 2021

bors bot commented May 17, 2021

cart commented May 17, 2021

cart commented May 17, 2021

bors bot commented May 18, 2021

[Merged by Bors] - Add a method `iter_combinations` on query to iterate over combinations of query results #1763

[Merged by Bors] - Add a method `iter_combinations` on query to iterate over combinations of query results #1763

Frizi commented Mar 26, 2021 •

edited

Loading

Frizi commented Mar 26, 2021 •

edited

Loading

alice-i-cecile commented Mar 27, 2021 •

edited

Loading

cart left a comment •

edited

Loading

cart commented May 3, 2021 •

edited

Loading

Frizi commented May 11, 2021 •

edited

Loading

Frizi commented May 17, 2021 •

edited

Loading