entity-trait-system

An alternative to ECS. Here is a video summary.

Setup

Requires the nightly compiler and specialization feature - it's only used in a sound way.

rustup override set nightly # enable nightly compiler
cargo add entity-trait-system # get lib

// explicitly opt in to language feature
#![allow(incomplete_features)]
#![feature(specialization)]

// declare world, entities, and traits which the entities could have
entity_trait_system::world!(
    MyWorld, Enemy, Player; TestTrait, SecondTestTrait);

let mut world = MyWorld::default();
// directly access arena member
let player_id = world.player.insert(Player { id: 1 });
// compile time type accessor of arena member (similar)
world.arena_mut::<Enemy>().insert(Enemy { hp: 10 });

// visit all arenas with types that implement trait
// (likely static dispatch)
#[cfg(feature = "rayon")]
world.par_visit_test_trait(|e| e.do_something());
#[cfg(not(feature = "rayon"))]
world.visit_test_trait(|e| e.do_something());

// runtime type API - access type-erased (Any) arena
let arena_id = MyWorld::arena_id::<Player>();
let player_arena = world.any_arena_mut(arena_id);
// unwrap: I know that this is a player
// and that the reference is valid
let player = player_arena
    .get_mut(player_id).unwrap()
    .downcast_mut::<Player>().unwrap();
player.do_something_else();

API Overview

Per-Trait Methods (generated for each trait)

• visit_<trait> - Iter over implementing entities
• visit_mut_<trait> - Mutable iter over implementing entities
• visit_key_<trait> - Iter (Key, &Value) tuples
• visit_key_mut_<trait> - Mutable iter (Key, &mut Value) tuples
• retain_<trait> - Keep entities matching predicate
• diff_<trait> - Gather diff vector from immutable view, to apply later
• diff_mut_<trait> - Same as previous, but from mutable view
• diff_apply_<trait> - Apply diff vector
• clear_<trait> - Clear all arenas implementing trait
• len_<trait> - Count entities implementing trait
• any_arenas_<trait> - Array of type-erased arenas implementing trait
• any_arenas_mut_<trait> - Mutable version of above

World Methods

• arena<T>() / arena_mut<T>() - Compile-time typed arena access
• any_arena() / any_arena_mut() - Runtime type-erased access
• arena_id<T>() - Get stable arena identifier - serializes to type name.
• clear() - Clear all arenas
• len() - Total entity count across all arenas

Rayon Support

Parallel operations exposed via par_* variants.

Serde Support

Both map (serde_json) and seq (bincode) style ser/deserialization.

Performance Note

Benchmarks

Here's a benchmark test suite comparing ETS to other popular libraries.

Simple Insert

ETS performed the best. It's the same speed as the underlying dense slot map insertion.

Simple Iter

The results for this benchmark form two distinct groups. The fastest group, including hecs and legion, arrange the components of the data as a structure of arrays. The second group (including ETS) iterate through an array of structures. A structure of arrays performs better since only the relevant data is loaded into the cache.

Since ETS doesn't use components, it is inside the second group.

Fragmented Iter

ETS arrived in second place, just behind shipyard.

System Scheduling

ETS performed the best. But, disjoint systems (outer parallelism) must be stated explicitly.

Heavy Compute

ETS overlapped with the other libraries - not much of a difference.

Add/Remove Component

ETS is omitted from this benchmark since it doesn't use components; it's not applicable. For sparse components, an auxiliary structure can store entity keys. For dense components, a trait in the ETS can implement the desire behaviour.

Serialize

Only three libraries implemented serialization. ETS arrived in second.

Optimization

This can be found in the implementation of visit_*:

fn v_if_applicable<F>(
    arena: &DenseSlotMap<DefaultKey, T>,
    mut handler: F) where F: FnMut(&dyn $trait_name)
{
    arena.values_as_slice().iter()
        .for_each(|entity| {
            // implicit type erase T -> &dyn $trait_name
            handler(entity)
        });
}

The handler is typically inlined and devirtualized to erase dynamic dispatch, since the type is known at compile time and is type erased just before use. This means that static dispatch is reliant on compiler optimization; likely but not guaranteed.

License: MIT