831ecf030c
# Objective - Systems that use the task pool, either explicitly or implicitly using parallel queries, will often end up executing tasks from different systems. - This can cause random tasks to block the main or render schedule at random, adding frame variance and increasing frame times when CPU bound. - This profile is a common occurrence on `main`. `propagate_parent_transforms` takes more than twice as long as it should, blocking the main schedule for that time, because it uses `task pool.scope`, which has decided to execute tasks from the render schedule on the main schedule.  ## Solution - In task pool scope execution, prefer to check if the current task is complete instead of ticking the executor to find new work. ## Testing - Ran the scene viewer with tracy to look for images like the one in the objective section. - Things look much, much better, and I could not find any occurrences:   |
||
|---|---|---|
| .. | ||
| examples | ||
| src | ||
| Cargo.toml | ||
| LICENSE-APACHE | ||
| LICENSE-MIT | ||
| README.md | ||
Bevy Tasks
A refreshingly simple task executor for bevy. :)
This is a simple threadpool with minimal dependencies. The main usecase is a scoped fork-join, i.e. spawning tasks from
a single thread and having that thread await the completion of those tasks. This is intended specifically for
bevy as a lighter alternative to rayon for this specific usecase. There are also utilities for
generating the tasks from a slice of data. This library is intended for games and makes no attempt to ensure fairness
or ordering of spawned tasks.
It is based on async-executor, a lightweight executor that allows the end user to manage their own threads.
async-executor is based on async-task, a core piece of async-std.
Usage
In order to be able to optimize task execution in multi-threaded environments, bevy provides three different thread pools via which tasks of different kinds can be spawned. (The same API is used in single-threaded environments, even if execution is limited to a single thread. This currently applies to Wasm targets.) The determining factor for what kind of work should go in each pool is latency requirements:
-
For CPU-intensive work (tasks that generally spin until completion) we have a standard [
ComputeTaskPool] and an [AsyncComputeTaskPool]. Work that does not need to be completed to present the next frame should go to the [AsyncComputeTaskPool]. -
For IO-intensive work (tasks that spend very little time in a "woken" state) we have an [
IoTaskPool] whose tasks are expected to complete very quickly. Generally speaking, they should just await receiving data from somewhere (i.e. disk) and signal other systems when the data is ready for consumption. (likely via channels)
no_std Support
To enable no_std support in this crate, you will need to disable default features, and enable the edge_executor and critical-section features. For platforms without full support for Rust atomics, you may also need to enable the portable-atomic feature.