Pipelined Rendering (#6503)
# Objective
- Implement pipelined rendering
- Fixes #5082
- Fixes #4718
## User Facing Description
Bevy now implements piplelined rendering! Pipelined rendering allows the app logic and rendering logic to run on different threads leading to large gains in performance.
![image](https://user-images.githubusercontent.com/2180432/202049871-3c00b801-58ab-448f-93fd-471e30aba55f.png)
*tracy capture of many_foxes example*
To use pipelined rendering, you just need to add the `PipelinedRenderingPlugin`. If you're using `DefaultPlugins` then it will automatically be added for you on all platforms except wasm. Bevy does not currently support multithreading on wasm which is needed for this feature to work. If you aren't using `DefaultPlugins` you can add the plugin manually.
```rust
use bevy::prelude::*;
use bevy::render::pipelined_rendering::PipelinedRenderingPlugin;
fn main() {
App::new()
// whatever other plugins you need
.add_plugin(RenderPlugin)
// needs to be added after RenderPlugin
.add_plugin(PipelinedRenderingPlugin)
.run();
}
```
If for some reason pipelined rendering needs to be removed. You can also disable the plugin the normal way.
```rust
use bevy::prelude::*;
use bevy::render::pipelined_rendering::PipelinedRenderingPlugin;
fn main() {
App::new.add_plugins(DefaultPlugins.build().disable::<PipelinedRenderingPlugin>());
}
```
### A setup function was added to plugins
A optional plugin lifecycle function was added to the `Plugin trait`. This function is called after all plugins have been built, but before the app runner is called. This allows for some final setup to be done. In the case of pipelined rendering, the function removes the sub app from the main app and sends it to the render thread.
```rust
struct MyPlugin;
impl Plugin for MyPlugin {
fn build(&self, app: &mut App) {
}
// optional function
fn setup(&self, app: &mut App) {
// do some final setup before runner is called
}
}
```
### A Stage for Frame Pacing
In the `RenderExtractApp` there is a stage labelled `BeforeIoAfterRenderStart` that systems can be added to. The specific use case for this stage is for a frame pacing system that can delay the start of main app processing in render bound apps to reduce input latency i.e. "frame pacing". This is not currently built into bevy, but exists as `bevy`
```text
|-------------------------------------------------------------------|
| | BeforeIoAfterRenderStart | winit events | main schedule |
| extract |---------------------------------------------------------|
| | extract commands | rendering schedule |
|-------------------------------------------------------------------|
```
### Small API additions
* `Schedule::remove_stage`
* `App::insert_sub_app`
* `App::remove_sub_app`
* `TaskPool::scope_with_executor`
## Problems and Solutions
### Moving render app to another thread
Most of the hard bits for this were done with the render redo. This PR just sends the render app back and forth through channels which seems to work ok. I originally experimented with using a scope to run the render task. It was cuter, but that approach didn't allow render to start before i/o processing. So I switched to using channels. There is much complexity in the coordination that needs to be done, but it's worth it. By moving rendering during i/o processing the frame times should be much more consistent in render bound apps. See https://github.com/bevyengine/bevy/issues/4691.
### Unsoundness with Sending World with NonSend resources
Dropping !Send things on threads other than the thread they were spawned on is considered unsound. The render world doesn't have any nonsend resources. So if we tell the users to "pretty please don't spawn nonsend resource on the render world", we can avoid this problem.
More seriously there is this https://github.com/bevyengine/bevy/pull/6534 pr, which patches the unsoundness by aborting the app if a nonsend resource is dropped on the wrong thread. ~~That PR should probably be merged before this one.~~ For a longer term solution we have this discussion going https://github.com/bevyengine/bevy/discussions/6552.
### NonSend Systems in render world
The render world doesn't have any !Send resources, but it does have a non send system. While Window is Send, winit does have some API's that can only be accessed on the main thread. `prepare_windows` in the render schedule thus needs to be scheduled on the main thread. Currently we run nonsend systems by running them on the thread the TaskPool::scope runs on. When we move render to another thread this no longer works.
To fix this, a new `scope_with_executor` method was added that takes a optional `TheadExecutor` that can only be ticked on the thread it was initialized on. The render world then holds a `MainThreadExecutor` resource which can be passed to the scope in the parallel executor that it uses to spawn it's non send systems on.
### Scopes executors between render and main should not share tasks
Since the render world and the app world share the `ComputeTaskPool`. Because `scope` has executors for the ComputeTaskPool a system from the main world could run on the render thread or a render system could run on the main thread. This can cause performance problems because it can delay a stage from finishing. See https://github.com/bevyengine/bevy/pull/6503#issuecomment-1309791442 for more details.
To avoid this problem, `TaskPool::scope` has been changed to not tick the ComputeTaskPool when it's used by the parallel executor. In the future when we move closer to the 1 thread to 1 logical core model we may want to overprovide threads, because the render and main app threads don't do much when executing the schedule.
## Performance
My machine is Windows 11, AMD Ryzen 5600x, RX 6600
### Examples
#### This PR with pipelining vs Main
> Note that these were run on an older version of main and the performance profile has probably changed due to optimizations
Seeing a perf gain from 29% on many lights to 7% on many sprites.
<html>
<body>
<!--StartFragment--><google-sheets-html-origin>
| percent | | | Diff | | | Main | | | PR | |
-- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | --
tracy frame time | mean | median | sigma | mean | median | sigma | mean | median | sigma | mean | median | sigma
many foxes | 27.01% | 27.34% | -47.09% | 1.58 | 1.55 | -1.78 | 5.85 | 5.67 | 3.78 | 4.27 | 4.12 | 5.56
many lights | 29.35% | 29.94% | -10.84% | 3.02 | 3.03 | -0.57 | 10.29 | 10.12 | 5.26 | 7.27 | 7.09 | 5.83
many animated sprites | 13.97% | 15.69% | 14.20% | 3.79 | 4.17 | 1.41 | 27.12 | 26.57 | 9.93 | 23.33 | 22.4 | 8.52
3d scene | 25.79% | 26.78% | 7.46% | 0.49 | 0.49 | 0.15 | 1.9 | 1.83 | 2.01 | 1.41 | 1.34 | 1.86
many cubes | 11.97% | 11.28% | 14.51% | 1.93 | 1.78 | 1.31 | 16.13 | 15.78 | 9.03 | 14.2 | 14 | 7.72
many sprites | 7.14% | 9.42% | -85.42% | 1.72 | 2.23 | -6.15 | 24.09 | 23.68 | 7.2 | 22.37 | 21.45 | 13.35
<!--EndFragment-->
</body>
</html>
#### This PR with pipelining disabled vs Main
Mostly regressions here. I don't think this should be a problem as users that are disabling pipelined rendering are probably running single threaded and not using the parallel executor. The regression is probably mostly due to the switch to use `async_executor::run` instead of `try_tick` and also having one less thread to run systems on. I'll do a writeup on why switching to `run` causes regressions, so we can try to eventually fix it. Using try_tick causes issues when pipeline rendering is enable as seen [here](https://github.com/bevyengine/bevy/pull/6503#issuecomment-1380803518)
<html>
<body>
<!--StartFragment--><google-sheets-html-origin>
| percent | | | Diff | | | Main | | | PR no pipelining | |
-- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | --
tracy frame time | mean | median | sigma | mean | median | sigma | mean | median | sigma | mean | median | sigma
many foxes | -3.72% | -4.42% | -1.07% | -0.21 | -0.24 | -0.04 | 5.64 | 5.43 | 3.74 | 5.85 | 5.67 | 3.78
many lights | 0.29% | -0.30% | 4.75% | 0.03 | -0.03 | 0.25 | 10.29 | 10.12 | 5.26 | 10.26 | 10.15 | 5.01
many animated sprites | 0.22% | 1.81% | -2.72% | 0.06 | 0.48 | -0.27 | 27.12 | 26.57 | 9.93 | 27.06 | 26.09 | 10.2
3d scene | -15.79% | -14.75% | -31.34% | -0.3 | -0.27 | -0.63 | 1.9 | 1.83 | 2.01 | 2.2 | 2.1 | 2.64
many cubes | -2.85% | -3.30% | 0.00% | -0.46 | -0.52 | 0 | 16.13 | 15.78 | 9.03 | 16.59 | 16.3 | 9.03
many sprites | 2.49% | 2.41% | 0.69% | 0.6 | 0.57 | 0.05 | 24.09 | 23.68 | 7.2 | 23.49 | 23.11 | 7.15
<!--EndFragment-->
</body>
</html>
### Benchmarks
Mostly the same except empty_systems has got a touch slower. The maybe_pipelining+1 column has the compute task pool with an extra thread over default added. This is because pipelining loses one thread over main to execute systems on, since the main thread no longer runs normal systems.
<details>
<summary>Click Me</summary>
```text
group main maybe-pipelining+1
----- ------------------------- ------------------
busy_systems/01x_entities_03_systems 1.07 30.7±1.32µs ? ?/sec 1.00 28.6±1.35µs ? ?/sec
busy_systems/01x_entities_06_systems 1.10 52.1±1.10µs ? ?/sec 1.00 47.2±1.08µs ? ?/sec
busy_systems/01x_entities_09_systems 1.00 74.6±1.36µs ? ?/sec 1.00 75.0±1.93µs ? ?/sec
busy_systems/01x_entities_12_systems 1.03 100.6±6.68µs ? ?/sec 1.00 98.0±1.46µs ? ?/sec
busy_systems/01x_entities_15_systems 1.11 128.5±3.53µs ? ?/sec 1.00 115.5±1.02µs ? ?/sec
busy_systems/02x_entities_03_systems 1.16 50.4±2.56µs ? ?/sec 1.00 43.5±3.00µs ? ?/sec
busy_systems/02x_entities_06_systems 1.00 87.1±1.27µs ? ?/sec 1.05 91.5±7.15µs ? ?/sec
busy_systems/02x_entities_09_systems 1.04 139.9±6.37µs ? ?/sec 1.00 134.0±1.06µs ? ?/sec
busy_systems/02x_entities_12_systems 1.05 179.2±3.47µs ? ?/sec 1.00 170.1±3.17µs ? ?/sec
busy_systems/02x_entities_15_systems 1.01 219.6±3.75µs ? ?/sec 1.00 218.1±2.55µs ? ?/sec
busy_systems/03x_entities_03_systems 1.10 70.6±2.33µs ? ?/sec 1.00 64.3±0.69µs ? ?/sec
busy_systems/03x_entities_06_systems 1.02 130.2±3.11µs ? ?/sec 1.00 128.0±1.34µs ? ?/sec
busy_systems/03x_entities_09_systems 1.00 195.0±10.11µs ? ?/sec 1.00 194.8±1.41µs ? ?/sec
busy_systems/03x_entities_12_systems 1.01 261.7±4.05µs ? ?/sec 1.00 259.8±4.11µs ? ?/sec
busy_systems/03x_entities_15_systems 1.00 318.0±3.04µs ? ?/sec 1.06 338.3±20.25µs ? ?/sec
busy_systems/04x_entities_03_systems 1.00 82.9±0.63µs ? ?/sec 1.02 84.3±0.63µs ? ?/sec
busy_systems/04x_entities_06_systems 1.01 181.7±3.65µs ? ?/sec 1.00 179.8±1.76µs ? ?/sec
busy_systems/04x_entities_09_systems 1.04 265.0±4.68µs ? ?/sec 1.00 255.3±1.98µs ? ?/sec
busy_systems/04x_entities_12_systems 1.00 335.9±3.00µs ? ?/sec 1.05 352.6±15.84µs ? ?/sec
busy_systems/04x_entities_15_systems 1.00 418.6±10.26µs ? ?/sec 1.08 450.2±39.58µs ? ?/sec
busy_systems/05x_entities_03_systems 1.07 114.3±0.95µs ? ?/sec 1.00 106.9±1.52µs ? ?/sec
busy_systems/05x_entities_06_systems 1.08 229.8±2.90µs ? ?/sec 1.00 212.3±4.18µs ? ?/sec
busy_systems/05x_entities_09_systems 1.03 329.3±1.99µs ? ?/sec 1.00 319.2±2.43µs ? ?/sec
busy_systems/05x_entities_12_systems 1.06 454.7±6.77µs ? ?/sec 1.00 430.1±3.58µs ? ?/sec
busy_systems/05x_entities_15_systems 1.03 554.6±6.15µs ? ?/sec 1.00 538.4±23.87µs ? ?/sec
contrived/01x_entities_03_systems 1.00 14.0±0.15µs ? ?/sec 1.08 15.1±0.21µs ? ?/sec
contrived/01x_entities_06_systems 1.04 28.5±0.37µs ? ?/sec 1.00 27.4±0.44µs ? ?/sec
contrived/01x_entities_09_systems 1.00 41.5±4.38µs ? ?/sec 1.02 42.2±2.24µs ? ?/sec
contrived/01x_entities_12_systems 1.06 55.9±1.49µs ? ?/sec 1.00 52.6±1.36µs ? ?/sec
contrived/01x_entities_15_systems 1.02 68.0±2.00µs ? ?/sec 1.00 66.5±0.78µs ? ?/sec
contrived/02x_entities_03_systems 1.03 25.2±0.38µs ? ?/sec 1.00 24.6±0.52µs ? ?/sec
contrived/02x_entities_06_systems 1.00 46.3±0.49µs ? ?/sec 1.04 48.1±4.13µs ? ?/sec
contrived/02x_entities_09_systems 1.02 70.4±0.99µs ? ?/sec 1.00 68.8±1.04µs ? ?/sec
contrived/02x_entities_12_systems 1.06 96.8±1.49µs ? ?/sec 1.00 91.5±0.93µs ? ?/sec
contrived/02x_entities_15_systems 1.02 116.2±0.95µs ? ?/sec 1.00 114.2±1.42µs ? ?/sec
contrived/03x_entities_03_systems 1.00 33.2±0.38µs ? ?/sec 1.01 33.6±0.45µs ? ?/sec
contrived/03x_entities_06_systems 1.00 62.4±0.73µs ? ?/sec 1.01 63.3±1.05µs ? ?/sec
contrived/03x_entities_09_systems 1.02 96.4±0.85µs ? ?/sec 1.00 94.8±3.02µs ? ?/sec
contrived/03x_entities_12_systems 1.01 126.3±4.67µs ? ?/sec 1.00 125.6±2.27µs ? ?/sec
contrived/03x_entities_15_systems 1.03 160.2±9.37µs ? ?/sec 1.00 156.0±1.53µs ? ?/sec
contrived/04x_entities_03_systems 1.02 41.4±3.39µs ? ?/sec 1.00 40.5±0.52µs ? ?/sec
contrived/04x_entities_06_systems 1.00 78.9±1.61µs ? ?/sec 1.02 80.3±1.06µs ? ?/sec
contrived/04x_entities_09_systems 1.02 121.8±3.97µs ? ?/sec 1.00 119.2±1.46µs ? ?/sec
contrived/04x_entities_12_systems 1.00 157.8±1.48µs ? ?/sec 1.01 160.1±1.72µs ? ?/sec
contrived/04x_entities_15_systems 1.00 197.9±1.47µs ? ?/sec 1.08 214.2±34.61µs ? ?/sec
contrived/05x_entities_03_systems 1.00 49.1±0.33µs ? ?/sec 1.01 49.7±0.75µs ? ?/sec
contrived/05x_entities_06_systems 1.00 95.0±0.93µs ? ?/sec 1.00 94.6±0.94µs ? ?/sec
contrived/05x_entities_09_systems 1.01 143.2±1.68µs ? ?/sec 1.00 142.2±2.00µs ? ?/sec
contrived/05x_entities_12_systems 1.00 191.8±2.03µs ? ?/sec 1.01 192.7±7.88µs ? ?/sec
contrived/05x_entities_15_systems 1.02 239.7±3.71µs ? ?/sec 1.00 235.8±4.11µs ? ?/sec
empty_systems/000_systems 1.01 47.8±0.67ns ? ?/sec 1.00 47.5±2.02ns ? ?/sec
empty_systems/001_systems 1.00 1743.2±126.14ns ? ?/sec 1.01 1761.1±70.10ns ? ?/sec
empty_systems/002_systems 1.01 2.2±0.04µs ? ?/sec 1.00 2.2±0.02µs ? ?/sec
empty_systems/003_systems 1.02 2.7±0.09µs ? ?/sec 1.00 2.7±0.16µs ? ?/sec
empty_systems/004_systems 1.00 3.1±0.11µs ? ?/sec 1.00 3.1±0.24µs ? ?/sec
empty_systems/005_systems 1.00 3.5±0.05µs ? ?/sec 1.11 3.9±0.70µs ? ?/sec
empty_systems/010_systems 1.00 5.5±0.12µs ? ?/sec 1.03 5.7±0.17µs ? ?/sec
empty_systems/015_systems 1.00 7.9±0.19µs ? ?/sec 1.06 8.4±0.16µs ? ?/sec
empty_systems/020_systems 1.00 10.4±1.25µs ? ?/sec 1.02 10.6±0.18µs ? ?/sec
empty_systems/025_systems 1.00 12.4±0.39µs ? ?/sec 1.14 14.1±1.07µs ? ?/sec
empty_systems/030_systems 1.00 15.1±0.39µs ? ?/sec 1.05 15.8±0.62µs ? ?/sec
empty_systems/035_systems 1.00 16.9±0.47µs ? ?/sec 1.07 18.0±0.37µs ? ?/sec
empty_systems/040_systems 1.00 19.3±0.41µs ? ?/sec 1.05 20.3±0.39µs ? ?/sec
empty_systems/045_systems 1.00 22.4±1.67µs ? ?/sec 1.02 22.9±0.51µs ? ?/sec
empty_systems/050_systems 1.00 24.4±1.67µs ? ?/sec 1.01 24.7±0.40µs ? ?/sec
empty_systems/055_systems 1.05 28.6±5.27µs ? ?/sec 1.00 27.2±0.70µs ? ?/sec
empty_systems/060_systems 1.02 29.9±1.64µs ? ?/sec 1.00 29.3±0.66µs ? ?/sec
empty_systems/065_systems 1.02 32.7±3.15µs ? ?/sec 1.00 32.1±0.98µs ? ?/sec
empty_systems/070_systems 1.00 33.0±1.42µs ? ?/sec 1.03 34.1±1.44µs ? ?/sec
empty_systems/075_systems 1.00 34.8±0.89µs ? ?/sec 1.04 36.2±0.70µs ? ?/sec
empty_systems/080_systems 1.00 37.0±1.82µs ? ?/sec 1.05 38.7±1.37µs ? ?/sec
empty_systems/085_systems 1.00 38.7±0.76µs ? ?/sec 1.05 40.8±0.83µs ? ?/sec
empty_systems/090_systems 1.00 41.5±1.09µs ? ?/sec 1.04 43.2±0.82µs ? ?/sec
empty_systems/095_systems 1.00 43.6±1.10µs ? ?/sec 1.04 45.2±0.99µs ? ?/sec
empty_systems/100_systems 1.00 46.7±2.27µs ? ?/sec 1.03 48.1±1.25µs ? ?/sec
```
</details>
## Migration Guide
### App `runner` and SubApp `extract` functions are now required to be Send
This was changed to enable pipelined rendering. If this breaks your use case please report it as these new bounds might be able to be relaxed.
## ToDo
* [x] redo benchmarking
* [x] reinvestigate the perf of the try_tick -> run change for task pool scope
2023-01-19 23:45:46 +00:00
|
|
|
use async_channel::{Receiver, Sender};
|
|
|
|
|
|
|
|
use bevy_app::{App, AppLabel, Plugin, SubApp};
|
|
|
|
use bevy_ecs::{
|
|
|
|
schedule::{MainThreadExecutor, StageLabel, SystemStage},
|
|
|
|
system::Resource,
|
|
|
|
world::{Mut, World},
|
|
|
|
};
|
|
|
|
use bevy_tasks::ComputeTaskPool;
|
|
|
|
|
|
|
|
use crate::RenderApp;
|
|
|
|
|
|
|
|
/// A Label for the sub app that runs the parts of pipelined rendering that need to run on the main thread.
|
|
|
|
#[derive(Debug, Clone, Copy, Hash, PartialEq, Eq, AppLabel)]
|
|
|
|
pub struct RenderExtractApp;
|
|
|
|
|
|
|
|
/// Labels for stages in the [`RenderExtractApp`] sub app. These will run after rendering has started.
|
|
|
|
#[derive(Debug, Hash, PartialEq, Eq, Clone, StageLabel)]
|
|
|
|
pub enum RenderExtractStage {
|
|
|
|
/// When pipelined rendering is enabled this stage runs after the render schedule starts, but
|
|
|
|
/// before I/O processing and the main app schedule. This can be useful for something like
|
|
|
|
/// frame pacing.
|
|
|
|
BeforeIoAfterRenderStart,
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Channel to send the render app from the main thread to the rendering thread
|
|
|
|
#[derive(Resource)]
|
|
|
|
pub struct MainToRenderAppSender(pub Sender<SubApp>);
|
|
|
|
|
|
|
|
/// Channel to send the render app from the render thread to the main thread
|
|
|
|
#[derive(Resource)]
|
|
|
|
pub struct RenderToMainAppReceiver(pub Receiver<SubApp>);
|
|
|
|
|
|
|
|
/// The [`PipelinedRenderingPlugin`] can be added to your application to enable pipelined rendering.
|
|
|
|
/// This moves rendering into a different thread, so that the Nth frame's rendering can
|
|
|
|
/// be run at the same time as the N + 1 frame's simulation.
|
|
|
|
///
|
|
|
|
/// ```text
|
|
|
|
/// |--------------------|--------------------|--------------------|--------------------|
|
|
|
|
/// | simulation thread | frame 1 simulation | frame 2 simulation | frame 3 simulation |
|
|
|
|
/// |--------------------|--------------------|--------------------|--------------------|
|
|
|
|
/// | rendering thread | | frame 1 rendering | frame 2 rendering |
|
|
|
|
/// |--------------------|--------------------|--------------------|--------------------|
|
|
|
|
/// ```
|
|
|
|
///
|
|
|
|
/// The plugin is dependent on the [`crate::RenderApp`] added by [`crate::RenderPlugin`] and so must
|
|
|
|
/// be added after that plugin. If it is not added after, the plugin will do nothing.
|
|
|
|
///
|
|
|
|
/// A single frame of execution looks something like below
|
|
|
|
///
|
|
|
|
/// ```text
|
|
|
|
/// |-------------------------------------------------------------------|
|
|
|
|
/// | | BeforeIoAfterRenderStart | winit events | main schedule |
|
|
|
|
/// | extract |---------------------------------------------------------|
|
|
|
|
/// | | extract commands | rendering schedule |
|
|
|
|
/// |-------------------------------------------------------------------|
|
|
|
|
/// ```
|
|
|
|
///
|
|
|
|
/// - `extract` is the stage where data is copied from the main world to the render world.
|
|
|
|
/// This is run on the main app's thread.
|
|
|
|
/// - On the render thread, we first apply the `extract commands`. This is not run during extract, so the
|
|
|
|
/// main schedule can start sooner.
|
|
|
|
/// - Then the `rendering schedule` is run. See [`crate::RenderStage`] for the available stages.
|
|
|
|
/// - In parallel to the rendering thread we first run the [`RenderExtractStage::BeforeIoAfterRenderStart`] stage. By
|
|
|
|
/// default this stage is empty. But is useful if you need something to run before I/O processing.
|
|
|
|
/// - Next all the `winit events` are processed.
|
|
|
|
/// - And finally the `main app schedule` is run.
|
|
|
|
/// - Once both the `main app schedule` and the `render schedule` are finished running, `extract` is run again.
|
|
|
|
#[derive(Default)]
|
|
|
|
pub struct PipelinedRenderingPlugin;
|
|
|
|
|
|
|
|
impl Plugin for PipelinedRenderingPlugin {
|
|
|
|
fn build(&self, app: &mut App) {
|
|
|
|
// Don't add RenderExtractApp if RenderApp isn't initialized.
|
|
|
|
if app.get_sub_app(RenderApp).is_err() {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
app.insert_resource(MainThreadExecutor::new());
|
|
|
|
|
|
|
|
let mut sub_app = App::empty();
|
|
|
|
sub_app.add_stage(
|
|
|
|
RenderExtractStage::BeforeIoAfterRenderStart,
|
|
|
|
SystemStage::parallel(),
|
|
|
|
);
|
2023-01-24 21:24:25 +00:00
|
|
|
app.insert_sub_app(RenderExtractApp, SubApp::new(sub_app, update_rendering));
|
Pipelined Rendering (#6503)
# Objective
- Implement pipelined rendering
- Fixes #5082
- Fixes #4718
## User Facing Description
Bevy now implements piplelined rendering! Pipelined rendering allows the app logic and rendering logic to run on different threads leading to large gains in performance.
![image](https://user-images.githubusercontent.com/2180432/202049871-3c00b801-58ab-448f-93fd-471e30aba55f.png)
*tracy capture of many_foxes example*
To use pipelined rendering, you just need to add the `PipelinedRenderingPlugin`. If you're using `DefaultPlugins` then it will automatically be added for you on all platforms except wasm. Bevy does not currently support multithreading on wasm which is needed for this feature to work. If you aren't using `DefaultPlugins` you can add the plugin manually.
```rust
use bevy::prelude::*;
use bevy::render::pipelined_rendering::PipelinedRenderingPlugin;
fn main() {
App::new()
// whatever other plugins you need
.add_plugin(RenderPlugin)
// needs to be added after RenderPlugin
.add_plugin(PipelinedRenderingPlugin)
.run();
}
```
If for some reason pipelined rendering needs to be removed. You can also disable the plugin the normal way.
```rust
use bevy::prelude::*;
use bevy::render::pipelined_rendering::PipelinedRenderingPlugin;
fn main() {
App::new.add_plugins(DefaultPlugins.build().disable::<PipelinedRenderingPlugin>());
}
```
### A setup function was added to plugins
A optional plugin lifecycle function was added to the `Plugin trait`. This function is called after all plugins have been built, but before the app runner is called. This allows for some final setup to be done. In the case of pipelined rendering, the function removes the sub app from the main app and sends it to the render thread.
```rust
struct MyPlugin;
impl Plugin for MyPlugin {
fn build(&self, app: &mut App) {
}
// optional function
fn setup(&self, app: &mut App) {
// do some final setup before runner is called
}
}
```
### A Stage for Frame Pacing
In the `RenderExtractApp` there is a stage labelled `BeforeIoAfterRenderStart` that systems can be added to. The specific use case for this stage is for a frame pacing system that can delay the start of main app processing in render bound apps to reduce input latency i.e. "frame pacing". This is not currently built into bevy, but exists as `bevy`
```text
|-------------------------------------------------------------------|
| | BeforeIoAfterRenderStart | winit events | main schedule |
| extract |---------------------------------------------------------|
| | extract commands | rendering schedule |
|-------------------------------------------------------------------|
```
### Small API additions
* `Schedule::remove_stage`
* `App::insert_sub_app`
* `App::remove_sub_app`
* `TaskPool::scope_with_executor`
## Problems and Solutions
### Moving render app to another thread
Most of the hard bits for this were done with the render redo. This PR just sends the render app back and forth through channels which seems to work ok. I originally experimented with using a scope to run the render task. It was cuter, but that approach didn't allow render to start before i/o processing. So I switched to using channels. There is much complexity in the coordination that needs to be done, but it's worth it. By moving rendering during i/o processing the frame times should be much more consistent in render bound apps. See https://github.com/bevyengine/bevy/issues/4691.
### Unsoundness with Sending World with NonSend resources
Dropping !Send things on threads other than the thread they were spawned on is considered unsound. The render world doesn't have any nonsend resources. So if we tell the users to "pretty please don't spawn nonsend resource on the render world", we can avoid this problem.
More seriously there is this https://github.com/bevyengine/bevy/pull/6534 pr, which patches the unsoundness by aborting the app if a nonsend resource is dropped on the wrong thread. ~~That PR should probably be merged before this one.~~ For a longer term solution we have this discussion going https://github.com/bevyengine/bevy/discussions/6552.
### NonSend Systems in render world
The render world doesn't have any !Send resources, but it does have a non send system. While Window is Send, winit does have some API's that can only be accessed on the main thread. `prepare_windows` in the render schedule thus needs to be scheduled on the main thread. Currently we run nonsend systems by running them on the thread the TaskPool::scope runs on. When we move render to another thread this no longer works.
To fix this, a new `scope_with_executor` method was added that takes a optional `TheadExecutor` that can only be ticked on the thread it was initialized on. The render world then holds a `MainThreadExecutor` resource which can be passed to the scope in the parallel executor that it uses to spawn it's non send systems on.
### Scopes executors between render and main should not share tasks
Since the render world and the app world share the `ComputeTaskPool`. Because `scope` has executors for the ComputeTaskPool a system from the main world could run on the render thread or a render system could run on the main thread. This can cause performance problems because it can delay a stage from finishing. See https://github.com/bevyengine/bevy/pull/6503#issuecomment-1309791442 for more details.
To avoid this problem, `TaskPool::scope` has been changed to not tick the ComputeTaskPool when it's used by the parallel executor. In the future when we move closer to the 1 thread to 1 logical core model we may want to overprovide threads, because the render and main app threads don't do much when executing the schedule.
## Performance
My machine is Windows 11, AMD Ryzen 5600x, RX 6600
### Examples
#### This PR with pipelining vs Main
> Note that these were run on an older version of main and the performance profile has probably changed due to optimizations
Seeing a perf gain from 29% on many lights to 7% on many sprites.
<html>
<body>
<!--StartFragment--><google-sheets-html-origin>
| percent | | | Diff | | | Main | | | PR | |
-- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | --
tracy frame time | mean | median | sigma | mean | median | sigma | mean | median | sigma | mean | median | sigma
many foxes | 27.01% | 27.34% | -47.09% | 1.58 | 1.55 | -1.78 | 5.85 | 5.67 | 3.78 | 4.27 | 4.12 | 5.56
many lights | 29.35% | 29.94% | -10.84% | 3.02 | 3.03 | -0.57 | 10.29 | 10.12 | 5.26 | 7.27 | 7.09 | 5.83
many animated sprites | 13.97% | 15.69% | 14.20% | 3.79 | 4.17 | 1.41 | 27.12 | 26.57 | 9.93 | 23.33 | 22.4 | 8.52
3d scene | 25.79% | 26.78% | 7.46% | 0.49 | 0.49 | 0.15 | 1.9 | 1.83 | 2.01 | 1.41 | 1.34 | 1.86
many cubes | 11.97% | 11.28% | 14.51% | 1.93 | 1.78 | 1.31 | 16.13 | 15.78 | 9.03 | 14.2 | 14 | 7.72
many sprites | 7.14% | 9.42% | -85.42% | 1.72 | 2.23 | -6.15 | 24.09 | 23.68 | 7.2 | 22.37 | 21.45 | 13.35
<!--EndFragment-->
</body>
</html>
#### This PR with pipelining disabled vs Main
Mostly regressions here. I don't think this should be a problem as users that are disabling pipelined rendering are probably running single threaded and not using the parallel executor. The regression is probably mostly due to the switch to use `async_executor::run` instead of `try_tick` and also having one less thread to run systems on. I'll do a writeup on why switching to `run` causes regressions, so we can try to eventually fix it. Using try_tick causes issues when pipeline rendering is enable as seen [here](https://github.com/bevyengine/bevy/pull/6503#issuecomment-1380803518)
<html>
<body>
<!--StartFragment--><google-sheets-html-origin>
| percent | | | Diff | | | Main | | | PR no pipelining | |
-- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | --
tracy frame time | mean | median | sigma | mean | median | sigma | mean | median | sigma | mean | median | sigma
many foxes | -3.72% | -4.42% | -1.07% | -0.21 | -0.24 | -0.04 | 5.64 | 5.43 | 3.74 | 5.85 | 5.67 | 3.78
many lights | 0.29% | -0.30% | 4.75% | 0.03 | -0.03 | 0.25 | 10.29 | 10.12 | 5.26 | 10.26 | 10.15 | 5.01
many animated sprites | 0.22% | 1.81% | -2.72% | 0.06 | 0.48 | -0.27 | 27.12 | 26.57 | 9.93 | 27.06 | 26.09 | 10.2
3d scene | -15.79% | -14.75% | -31.34% | -0.3 | -0.27 | -0.63 | 1.9 | 1.83 | 2.01 | 2.2 | 2.1 | 2.64
many cubes | -2.85% | -3.30% | 0.00% | -0.46 | -0.52 | 0 | 16.13 | 15.78 | 9.03 | 16.59 | 16.3 | 9.03
many sprites | 2.49% | 2.41% | 0.69% | 0.6 | 0.57 | 0.05 | 24.09 | 23.68 | 7.2 | 23.49 | 23.11 | 7.15
<!--EndFragment-->
</body>
</html>
### Benchmarks
Mostly the same except empty_systems has got a touch slower. The maybe_pipelining+1 column has the compute task pool with an extra thread over default added. This is because pipelining loses one thread over main to execute systems on, since the main thread no longer runs normal systems.
<details>
<summary>Click Me</summary>
```text
group main maybe-pipelining+1
----- ------------------------- ------------------
busy_systems/01x_entities_03_systems 1.07 30.7±1.32µs ? ?/sec 1.00 28.6±1.35µs ? ?/sec
busy_systems/01x_entities_06_systems 1.10 52.1±1.10µs ? ?/sec 1.00 47.2±1.08µs ? ?/sec
busy_systems/01x_entities_09_systems 1.00 74.6±1.36µs ? ?/sec 1.00 75.0±1.93µs ? ?/sec
busy_systems/01x_entities_12_systems 1.03 100.6±6.68µs ? ?/sec 1.00 98.0±1.46µs ? ?/sec
busy_systems/01x_entities_15_systems 1.11 128.5±3.53µs ? ?/sec 1.00 115.5±1.02µs ? ?/sec
busy_systems/02x_entities_03_systems 1.16 50.4±2.56µs ? ?/sec 1.00 43.5±3.00µs ? ?/sec
busy_systems/02x_entities_06_systems 1.00 87.1±1.27µs ? ?/sec 1.05 91.5±7.15µs ? ?/sec
busy_systems/02x_entities_09_systems 1.04 139.9±6.37µs ? ?/sec 1.00 134.0±1.06µs ? ?/sec
busy_systems/02x_entities_12_systems 1.05 179.2±3.47µs ? ?/sec 1.00 170.1±3.17µs ? ?/sec
busy_systems/02x_entities_15_systems 1.01 219.6±3.75µs ? ?/sec 1.00 218.1±2.55µs ? ?/sec
busy_systems/03x_entities_03_systems 1.10 70.6±2.33µs ? ?/sec 1.00 64.3±0.69µs ? ?/sec
busy_systems/03x_entities_06_systems 1.02 130.2±3.11µs ? ?/sec 1.00 128.0±1.34µs ? ?/sec
busy_systems/03x_entities_09_systems 1.00 195.0±10.11µs ? ?/sec 1.00 194.8±1.41µs ? ?/sec
busy_systems/03x_entities_12_systems 1.01 261.7±4.05µs ? ?/sec 1.00 259.8±4.11µs ? ?/sec
busy_systems/03x_entities_15_systems 1.00 318.0±3.04µs ? ?/sec 1.06 338.3±20.25µs ? ?/sec
busy_systems/04x_entities_03_systems 1.00 82.9±0.63µs ? ?/sec 1.02 84.3±0.63µs ? ?/sec
busy_systems/04x_entities_06_systems 1.01 181.7±3.65µs ? ?/sec 1.00 179.8±1.76µs ? ?/sec
busy_systems/04x_entities_09_systems 1.04 265.0±4.68µs ? ?/sec 1.00 255.3±1.98µs ? ?/sec
busy_systems/04x_entities_12_systems 1.00 335.9±3.00µs ? ?/sec 1.05 352.6±15.84µs ? ?/sec
busy_systems/04x_entities_15_systems 1.00 418.6±10.26µs ? ?/sec 1.08 450.2±39.58µs ? ?/sec
busy_systems/05x_entities_03_systems 1.07 114.3±0.95µs ? ?/sec 1.00 106.9±1.52µs ? ?/sec
busy_systems/05x_entities_06_systems 1.08 229.8±2.90µs ? ?/sec 1.00 212.3±4.18µs ? ?/sec
busy_systems/05x_entities_09_systems 1.03 329.3±1.99µs ? ?/sec 1.00 319.2±2.43µs ? ?/sec
busy_systems/05x_entities_12_systems 1.06 454.7±6.77µs ? ?/sec 1.00 430.1±3.58µs ? ?/sec
busy_systems/05x_entities_15_systems 1.03 554.6±6.15µs ? ?/sec 1.00 538.4±23.87µs ? ?/sec
contrived/01x_entities_03_systems 1.00 14.0±0.15µs ? ?/sec 1.08 15.1±0.21µs ? ?/sec
contrived/01x_entities_06_systems 1.04 28.5±0.37µs ? ?/sec 1.00 27.4±0.44µs ? ?/sec
contrived/01x_entities_09_systems 1.00 41.5±4.38µs ? ?/sec 1.02 42.2±2.24µs ? ?/sec
contrived/01x_entities_12_systems 1.06 55.9±1.49µs ? ?/sec 1.00 52.6±1.36µs ? ?/sec
contrived/01x_entities_15_systems 1.02 68.0±2.00µs ? ?/sec 1.00 66.5±0.78µs ? ?/sec
contrived/02x_entities_03_systems 1.03 25.2±0.38µs ? ?/sec 1.00 24.6±0.52µs ? ?/sec
contrived/02x_entities_06_systems 1.00 46.3±0.49µs ? ?/sec 1.04 48.1±4.13µs ? ?/sec
contrived/02x_entities_09_systems 1.02 70.4±0.99µs ? ?/sec 1.00 68.8±1.04µs ? ?/sec
contrived/02x_entities_12_systems 1.06 96.8±1.49µs ? ?/sec 1.00 91.5±0.93µs ? ?/sec
contrived/02x_entities_15_systems 1.02 116.2±0.95µs ? ?/sec 1.00 114.2±1.42µs ? ?/sec
contrived/03x_entities_03_systems 1.00 33.2±0.38µs ? ?/sec 1.01 33.6±0.45µs ? ?/sec
contrived/03x_entities_06_systems 1.00 62.4±0.73µs ? ?/sec 1.01 63.3±1.05µs ? ?/sec
contrived/03x_entities_09_systems 1.02 96.4±0.85µs ? ?/sec 1.00 94.8±3.02µs ? ?/sec
contrived/03x_entities_12_systems 1.01 126.3±4.67µs ? ?/sec 1.00 125.6±2.27µs ? ?/sec
contrived/03x_entities_15_systems 1.03 160.2±9.37µs ? ?/sec 1.00 156.0±1.53µs ? ?/sec
contrived/04x_entities_03_systems 1.02 41.4±3.39µs ? ?/sec 1.00 40.5±0.52µs ? ?/sec
contrived/04x_entities_06_systems 1.00 78.9±1.61µs ? ?/sec 1.02 80.3±1.06µs ? ?/sec
contrived/04x_entities_09_systems 1.02 121.8±3.97µs ? ?/sec 1.00 119.2±1.46µs ? ?/sec
contrived/04x_entities_12_systems 1.00 157.8±1.48µs ? ?/sec 1.01 160.1±1.72µs ? ?/sec
contrived/04x_entities_15_systems 1.00 197.9±1.47µs ? ?/sec 1.08 214.2±34.61µs ? ?/sec
contrived/05x_entities_03_systems 1.00 49.1±0.33µs ? ?/sec 1.01 49.7±0.75µs ? ?/sec
contrived/05x_entities_06_systems 1.00 95.0±0.93µs ? ?/sec 1.00 94.6±0.94µs ? ?/sec
contrived/05x_entities_09_systems 1.01 143.2±1.68µs ? ?/sec 1.00 142.2±2.00µs ? ?/sec
contrived/05x_entities_12_systems 1.00 191.8±2.03µs ? ?/sec 1.01 192.7±7.88µs ? ?/sec
contrived/05x_entities_15_systems 1.02 239.7±3.71µs ? ?/sec 1.00 235.8±4.11µs ? ?/sec
empty_systems/000_systems 1.01 47.8±0.67ns ? ?/sec 1.00 47.5±2.02ns ? ?/sec
empty_systems/001_systems 1.00 1743.2±126.14ns ? ?/sec 1.01 1761.1±70.10ns ? ?/sec
empty_systems/002_systems 1.01 2.2±0.04µs ? ?/sec 1.00 2.2±0.02µs ? ?/sec
empty_systems/003_systems 1.02 2.7±0.09µs ? ?/sec 1.00 2.7±0.16µs ? ?/sec
empty_systems/004_systems 1.00 3.1±0.11µs ? ?/sec 1.00 3.1±0.24µs ? ?/sec
empty_systems/005_systems 1.00 3.5±0.05µs ? ?/sec 1.11 3.9±0.70µs ? ?/sec
empty_systems/010_systems 1.00 5.5±0.12µs ? ?/sec 1.03 5.7±0.17µs ? ?/sec
empty_systems/015_systems 1.00 7.9±0.19µs ? ?/sec 1.06 8.4±0.16µs ? ?/sec
empty_systems/020_systems 1.00 10.4±1.25µs ? ?/sec 1.02 10.6±0.18µs ? ?/sec
empty_systems/025_systems 1.00 12.4±0.39µs ? ?/sec 1.14 14.1±1.07µs ? ?/sec
empty_systems/030_systems 1.00 15.1±0.39µs ? ?/sec 1.05 15.8±0.62µs ? ?/sec
empty_systems/035_systems 1.00 16.9±0.47µs ? ?/sec 1.07 18.0±0.37µs ? ?/sec
empty_systems/040_systems 1.00 19.3±0.41µs ? ?/sec 1.05 20.3±0.39µs ? ?/sec
empty_systems/045_systems 1.00 22.4±1.67µs ? ?/sec 1.02 22.9±0.51µs ? ?/sec
empty_systems/050_systems 1.00 24.4±1.67µs ? ?/sec 1.01 24.7±0.40µs ? ?/sec
empty_systems/055_systems 1.05 28.6±5.27µs ? ?/sec 1.00 27.2±0.70µs ? ?/sec
empty_systems/060_systems 1.02 29.9±1.64µs ? ?/sec 1.00 29.3±0.66µs ? ?/sec
empty_systems/065_systems 1.02 32.7±3.15µs ? ?/sec 1.00 32.1±0.98µs ? ?/sec
empty_systems/070_systems 1.00 33.0±1.42µs ? ?/sec 1.03 34.1±1.44µs ? ?/sec
empty_systems/075_systems 1.00 34.8±0.89µs ? ?/sec 1.04 36.2±0.70µs ? ?/sec
empty_systems/080_systems 1.00 37.0±1.82µs ? ?/sec 1.05 38.7±1.37µs ? ?/sec
empty_systems/085_systems 1.00 38.7±0.76µs ? ?/sec 1.05 40.8±0.83µs ? ?/sec
empty_systems/090_systems 1.00 41.5±1.09µs ? ?/sec 1.04 43.2±0.82µs ? ?/sec
empty_systems/095_systems 1.00 43.6±1.10µs ? ?/sec 1.04 45.2±0.99µs ? ?/sec
empty_systems/100_systems 1.00 46.7±2.27µs ? ?/sec 1.03 48.1±1.25µs ? ?/sec
```
</details>
## Migration Guide
### App `runner` and SubApp `extract` functions are now required to be Send
This was changed to enable pipelined rendering. If this breaks your use case please report it as these new bounds might be able to be relaxed.
## ToDo
* [x] redo benchmarking
* [x] reinvestigate the perf of the try_tick -> run change for task pool scope
2023-01-19 23:45:46 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// Sets up the render thread and inserts resources into the main app used for controlling the render thread.
|
|
|
|
fn setup(&self, app: &mut App) {
|
|
|
|
// skip setting up when headless
|
|
|
|
if app.get_sub_app(RenderExtractApp).is_err() {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
let (app_to_render_sender, app_to_render_receiver) = async_channel::bounded::<SubApp>(1);
|
|
|
|
let (render_to_app_sender, render_to_app_receiver) = async_channel::bounded::<SubApp>(1);
|
|
|
|
|
|
|
|
let mut render_app = app
|
|
|
|
.remove_sub_app(RenderApp)
|
|
|
|
.expect("Unable to get RenderApp. Another plugin may have removed the RenderApp before PipelinedRenderingPlugin");
|
|
|
|
|
|
|
|
// clone main thread executor to render world
|
|
|
|
let executor = app.world.get_resource::<MainThreadExecutor>().unwrap();
|
|
|
|
render_app.app.world.insert_resource(executor.clone());
|
|
|
|
|
|
|
|
render_to_app_sender.send_blocking(render_app).unwrap();
|
|
|
|
|
|
|
|
app.insert_resource(MainToRenderAppSender(app_to_render_sender));
|
|
|
|
app.insert_resource(RenderToMainAppReceiver(render_to_app_receiver));
|
|
|
|
|
|
|
|
std::thread::spawn(move || {
|
|
|
|
#[cfg(feature = "trace")]
|
|
|
|
let _span = bevy_utils::tracing::info_span!("render thread").entered();
|
|
|
|
|
|
|
|
loop {
|
|
|
|
// run a scope here to allow main world to use this thread while it's waiting for the render app
|
|
|
|
let mut render_app = ComputeTaskPool::get()
|
|
|
|
.scope(|s| {
|
|
|
|
s.spawn(async { app_to_render_receiver.recv().await.unwrap() });
|
|
|
|
})
|
|
|
|
.pop()
|
|
|
|
.unwrap();
|
|
|
|
|
|
|
|
#[cfg(feature = "trace")]
|
|
|
|
let _sub_app_span =
|
|
|
|
bevy_utils::tracing::info_span!("sub app", name = ?RenderApp).entered();
|
|
|
|
render_app.run();
|
|
|
|
render_to_app_sender.send_blocking(render_app).unwrap();
|
|
|
|
}
|
|
|
|
});
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// This function waits for the rendering world to be received,
|
|
|
|
// runs extract, and then sends the rendering world back to the render thread.
|
|
|
|
fn update_rendering(app_world: &mut World, _sub_app: &mut App) {
|
|
|
|
app_world.resource_scope(|world, main_thread_executor: Mut<MainThreadExecutor>| {
|
|
|
|
// we use a scope here to run any main thread tasks that the render world still needs to run
|
|
|
|
// while we wait for the render world to be received.
|
|
|
|
let mut render_app = ComputeTaskPool::get()
|
|
|
|
.scope_with_executor(true, Some(&*main_thread_executor.0), |s| {
|
|
|
|
s.spawn(async {
|
|
|
|
let receiver = world.get_resource::<RenderToMainAppReceiver>().unwrap();
|
|
|
|
receiver.0.recv().await.unwrap()
|
|
|
|
});
|
|
|
|
})
|
|
|
|
.pop()
|
|
|
|
.unwrap();
|
|
|
|
|
|
|
|
render_app.extract(world);
|
|
|
|
|
|
|
|
let sender = world.resource::<MainToRenderAppSender>();
|
|
|
|
sender.0.send_blocking(render_app).unwrap();
|
|
|
|
});
|
|
|
|
}
|