This crate consists of a number of benchmarks intended for testing out the effects of data layouts and batching strategies on Bevy systems and queries.
To run:
cargo bench
The results will land in your target/criterion/report directory when finished.
These benchmarks were created in support of Bevy issue #1990
The effect of AoS, SoA, and AoSoA layouts on Bevy queries, and different batching methods.
The benchmarks are divided into 6 categories:
- naive: AoS layout without alignment (e.g., Vec3)
- naive_aligned: AoS layout with alignment (e.g., Vec3 with an extra component to get 16 bytes alignment)
- naive_batched_swizzle: Operates on batches of 4 Vec3s, converts their layouts to SoA, does the processing, and swizzles the results back
- aosoa_sse4: uses an AoSoA layout with 4 lanes
- simd_batch_sse4: uses SoA layout. the "optimal" and "suboptimal" benchmarks show differences between different iteration patterns.
For SoA and AoSoA layouts, I implemented the tests using upstream Bevy by having each "entity" manage 4 "virtual entities". This provides the same data layout as if Bevy itself supported AoSoA and SoA by default. Then, each system operates on batches of 4 "virtual entities" using SIMD to simulate the results. The total number of entities across all benchmarks (within the same size category) are the same.
The actual tests involve a loop that updates an Entity's position using a time and velocity. Inputs are randomized to prevent the compiler from optimizing them out, and provide a more realistic distribution.
So far, AoSoA seems to be winning on my system, followed by the SoA layout, followed by the AoS layouts.
I expect, based on the numbers provided, that they do, and a significant speedup is on the table for systems using dense components and linear access patterns.
- This crate relies on the plain mathematical types provided by Bevy
- No fancy packed_simd support, etc
- These benchmarks do not test the effect of
-march=native, fast math optimizations, etc. - Only
f32s tested. No integral types.