bevy/benches/Cargo.toml

74 lines
1.5 KiB
TOML
Raw Normal View History

[package]
name = "benches"
edition = "2021"
description = "Benchmarks that test Bevy's performance"
publish = false
license = "MIT OR Apache-2.0"
[dev-dependencies]
glam = "0.27"
rand = "0.8"
rand_chacha = "0.3"
criterion = { version = "0.3", features = ["html_reports"] }
bevy_app = { path = "../crates/bevy_app" }
bevy_ecs = { path = "../crates/bevy_ecs", features = ["multi_threaded"] }
bevy_reflect = { path = "../crates/bevy_reflect" }
bevy_tasks = { path = "../crates/bevy_tasks" }
bevy_utils = { path = "../crates/bevy_utils" }
Add Beziers to `bevy_math` (#7653) # Objective - Adds foundational math for Bezier curves, useful for UI/2D/3D animation and smooth paths. https://user-images.githubusercontent.com/2632925/218883143-e138f994-1795-40da-8c59-21d779666991.mp4 ## Solution - Adds the generic `Bezier` type, and a `Point` trait. The `Point` trait allows us to use control points of any dimension, as long as they support vector math. I've implemented it for `f32`(1D), `Vec2`(2D), and `Vec3`/`Vec3A`(3D). - Adds `CubicBezierEasing` on top of `Bezier` with the addition of an implementation of cubic Bezier easing, which is a foundational tool for UI animation. - This involves solving for $t$ in the parametric Bezier function $B(t)$ using the Newton-Raphson method to find a value with error $\leq$ 1e-7, capped at 8 iterations. - Added type aliases for common Bezier curves: `CubicBezier2d`, `CubicBezier3d`, `QuadraticBezier2d`, and `QuadraticBezier3d`. These types use `Vec3A` to represent control points, as this was found to have an 80-90% speedup over using `Vec3`. - Benchmarking shows quadratic/cubic Bezier evaluations $B(t)$ take \~1.8/2.4ns respectively. Easing, which requires an iterative solve takes \~50ns for cubic Beziers. --- ## Changelog - Added `CubicBezier2d`, `CubicBezier3d`, `QuadraticBezier2d`, and `QuadraticBezier3d` types with methods for sampling position, velocity, and acceleration. The generic `Bezier` type is also available, and generic over any degree of Bezier curve. - Added `CubicBezierEasing`, with additional methods to allow for smooth easing animations.
2023-02-20 18:34:52 +00:00
bevy_math = { path = "../crates/bevy_math" }
bevy_render = { path = "../crates/bevy_render" }
Remove unnecesary branches/panics from Query accesses (#6461) # Objective Supercedes #6452. Upon inspection of the [generated assembly](https://gist.github.com/james7132/c2740c6941b80d7912f1e8888e223cbb#file-original-s) of a [simple Bevy binary](https://gist.github.com/james7132/c2740c6941b80d7912f1e8888e223cbb#file-source-rs) compiled with `cargo rustc --release -- --emit asm`, it's apparent that there are multiple unnecessary branches in the generated assembly: ```assembly .LBB5_5: cmpq %r10, %r11 je .LBB5_15 movq (%r11), %rcx movq 328(%r15), %rdx cmpq %rdx, %rcx jae .LBB5_14 movq 312(%r15), %rdi leaq (%rcx,%rcx,2), %rcx shlq $5, %rcx movq 336(%r12), %rdx movq 64(%rdi,%rcx), %rax cmpq %rdx, %rax jbe .LBB5_4 leaq (%rdi,%rcx), %rsi movq 48(%rsi), %rbp shlq $4, %rdx cmpq $0, (%rbp,%rdx) je .LBB5_4 movq 344(%r12), %rbx cmpq %rbx, %rax jbe .LBB5_4 shlq $4, %rbx cmpq $0, (%rbp,%rbx) je .LBB5_4 addq $8, %r11 movq 88(%rdi,%rcx), %rcx testq %rcx, %rcx je .LBB5_5 movq (%rsi), %rax movq 8(%rbp,%rdx), %rdx leaq (%rdx,%rdx,4), %rdi shlq $4, %rdi movq 32(%rax,%rdi), %rdx movq 56(%rax,%rdi), %r8 movq 8(%rbp,%rbx), %rbp leaq (%rbp,%rbp,4), %rbp shlq $4, %rbp movq 32(%rax,%rbp), %r9 xorl %ebp, %ebp jmp .LBB5_13 .p2align 4, 0x90 ``` Almost every one of the instructions starting with `j` is a potential branch, which can significantly slow down accesses. Of these, two labels are both common and never used: ```asm .LBB5_14: leaq __unnamed_2(%rip), %r8 callq _ZN4core9panicking18panic_bounds_check17h70367088e72af65aE ud2 .LBB5_4: callq _ZN8bevy_ecs5query25debug_checked_unreachable17h0855ff520ceaea77E ud2 .seh_endproc ``` These correpsond to subprocedure calls to panicking due to out of bounds from indexing `Tables` and `debug_checked_unreadable`. Both of which should be inlined and optimized out, but are not. ## Solution Make `debug_checked_unreachable` a macro to forcibly inline either `unreachable!()` in debug builds, and `std::hint::unreachable_unchecked()` in release mode. Replace the `Tables` and `Archetype` index access with `get(id).unwrap_or_else(|| debug_checked_unreachable!())` to assume that the table or archetype provided exists. This has no external breaking change of any kind. The equivalent section of code with these changes removes most of the conditional jump instructions: ```asm .LBB5_5: movss (%rbx,%rbp,4), %xmm0 movl %r14d, 4(%r8,%rbp,8) addss (%rdi,%rbp,4), %xmm0 movss %xmm0, (%rdi,%rbp,4) incq %rbp .LBB5_1: cmpq %rdx, %rbp jne .LBB5_5 .p2align 4, 0x90 .LBB5_2: cmpq %rcx, %rax je .LBB5_6 movq (%rax), %rdx addq $8, %rax movq 312(%rsi), %rbp leaq (%rdx,%rdx,2), %rbx shlq $5, %rbx movq 88(%rbp,%rbx), %rdx testq %rdx, %rdx je .LBB5_2 leaq (%rbx,%rbp), %r8 movq 336(%r15), %rdi movq 344(%r15), %r9 movq 48(%rbp,%rbx), %r10 shlq $4, %rdi movq (%r8), %rbx movq 8(%r10,%rdi), %rdi leaq (%rdi,%rdi,4), %rbp shlq $4, %rbp movq 32(%rbx,%rbp), %rdi movq 56(%rbx,%rbp), %r8 shlq $4, %r9 movq 8(%r10,%r9), %rbp leaq (%rbp,%rbp,4), %rbp shlq $4, %rbp movq 32(%rbx,%rbp), %rbx xorl %ebp, %ebp jmp .LBB5_5 .LBB5_6: addq $40, %rsp popq %rbx popq %rbp popq %rdi popq %rsi popq %r14 popq %r15 retq .seh_endproc ``` ## Performance Microbenchmarks results: <details> ``` group main no-panic-query ----- ---- -------------- busy_systems/01x_entities_03_systems 1.20 42.4±2.66µs ? ?/sec 1.00 35.3±1.68µs ? ?/sec busy_systems/01x_entities_06_systems 1.32 83.8±3.50µs ? ?/sec 1.00 63.6±1.72µs ? ?/sec busy_systems/01x_entities_09_systems 1.15 113.3±8.90µs ? ?/sec 1.00 98.2±6.15µs ? ?/sec busy_systems/01x_entities_12_systems 1.27 160.8±32.44µs ? ?/sec 1.00 126.6±4.70µs ? ?/sec busy_systems/01x_entities_15_systems 1.12 179.6±3.71µs ? ?/sec 1.00 160.3±11.03µs ? ?/sec busy_systems/02x_entities_03_systems 1.18 76.8±3.14µs ? ?/sec 1.00 65.2±3.17µs ? ?/sec busy_systems/02x_entities_06_systems 1.16 144.6±6.10µs ? ?/sec 1.00 124.5±5.14µs ? ?/sec busy_systems/02x_entities_09_systems 1.19 215.3±9.18µs ? ?/sec 1.00 181.5±5.67µs ? ?/sec busy_systems/02x_entities_12_systems 1.20 266.7±8.33µs ? ?/sec 1.00 222.0±9.53µs ? ?/sec busy_systems/02x_entities_15_systems 1.23 338.8±10.53µs ? ?/sec 1.00 276.3±6.94µs ? ?/sec busy_systems/03x_entities_03_systems 1.43 113.5±5.06µs ? ?/sec 1.00 79.6±1.49µs ? ?/sec busy_systems/03x_entities_06_systems 1.38 217.3±12.67µs ? ?/sec 1.00 157.5±3.07µs ? ?/sec busy_systems/03x_entities_09_systems 1.23 308.8±24.75µs ? ?/sec 1.00 251.6±8.93µs ? ?/sec busy_systems/03x_entities_12_systems 1.05 347.7±12.43µs ? ?/sec 1.00 330.6±11.43µs ? ?/sec busy_systems/03x_entities_15_systems 1.13 455.5±13.88µs ? ?/sec 1.00 401.7±17.29µs ? ?/sec busy_systems/04x_entities_03_systems 1.24 144.7±5.89µs ? ?/sec 1.00 116.9±6.29µs ? ?/sec busy_systems/04x_entities_06_systems 1.24 282.8±21.40µs ? ?/sec 1.00 228.6±21.31µs ? ?/sec busy_systems/04x_entities_09_systems 1.35 431.8±14.10µs ? ?/sec 1.00 319.6±9.83µs ? ?/sec busy_systems/04x_entities_12_systems 1.16 493.8±22.87µs ? ?/sec 1.00 424.9±15.24µs ? ?/sec busy_systems/04x_entities_15_systems 1.10 587.5±23.25µs ? ?/sec 1.00 531.7±16.32µs ? ?/sec busy_systems/05x_entities_03_systems 1.14 148.2±9.61µs ? ?/sec 1.00 129.5±4.32µs ? ?/sec busy_systems/05x_entities_06_systems 1.31 359.7±17.46µs ? ?/sec 1.00 273.6±10.55µs ? ?/sec busy_systems/05x_entities_09_systems 1.22 473.5±23.11µs ? ?/sec 1.00 389.3±13.62µs ? ?/sec busy_systems/05x_entities_12_systems 1.05 562.9±20.76µs ? ?/sec 1.00 536.5±24.35µs ? ?/sec busy_systems/05x_entities_15_systems 1.23 818.5±28.70µs ? ?/sec 1.00 666.6±45.87µs ? ?/sec contrived/01x_entities_03_systems 1.27 27.5±0.49µs ? ?/sec 1.00 21.6±1.71µs ? ?/sec contrived/01x_entities_06_systems 1.22 49.9±1.18µs ? ?/sec 1.00 40.7±2.62µs ? ?/sec contrived/01x_entities_09_systems 1.30 72.3±2.39µs ? ?/sec 1.00 55.4±2.60µs ? ?/sec contrived/01x_entities_12_systems 1.28 94.3±9.44µs ? ?/sec 1.00 73.7±3.62µs ? ?/sec contrived/01x_entities_15_systems 1.25 118.0±2.43µs ? ?/sec 1.00 94.1±3.99µs ? ?/sec contrived/02x_entities_03_systems 1.23 41.6±1.71µs ? ?/sec 1.00 33.7±2.30µs ? ?/sec contrived/02x_entities_06_systems 1.19 78.6±2.63µs ? ?/sec 1.00 65.9±2.35µs ? ?/sec contrived/02x_entities_09_systems 1.28 113.6±3.60µs ? ?/sec 1.00 88.6±3.60µs ? ?/sec contrived/02x_entities_12_systems 1.20 146.4±5.75µs ? ?/sec 1.00 121.7±3.35µs ? ?/sec contrived/02x_entities_15_systems 1.23 178.5±4.86µs ? ?/sec 1.00 145.7±4.00µs ? ?/sec contrived/03x_entities_03_systems 1.42 58.3±2.77µs ? ?/sec 1.00 41.1±1.54µs ? ?/sec contrived/03x_entities_06_systems 1.32 108.5±7.30µs ? ?/sec 1.00 82.4±4.86µs ? ?/sec contrived/03x_entities_09_systems 1.23 153.7±4.61µs ? ?/sec 1.00 125.0±4.76µs ? ?/sec contrived/03x_entities_12_systems 1.18 197.5±5.12µs ? ?/sec 1.00 166.8±8.14µs ? ?/sec contrived/03x_entities_15_systems 1.23 238.8±6.38µs ? ?/sec 1.00 194.6±4.55µs ? ?/sec contrived/04x_entities_03_systems 1.34 66.4±3.42µs ? ?/sec 1.00 49.5±1.98µs ? ?/sec contrived/04x_entities_06_systems 1.27 134.3±4.86µs ? ?/sec 1.00 105.8±3.58µs ? ?/sec contrived/04x_entities_09_systems 1.26 193.2±3.83µs ? ?/sec 1.00 153.0±5.60µs ? ?/sec contrived/04x_entities_12_systems 1.16 237.1±5.78µs ? ?/sec 1.00 204.9±18.77µs ? ?/sec contrived/04x_entities_15_systems 1.17 289.2±4.76µs ? ?/sec 1.00 246.3±8.57µs ? ?/sec contrived/05x_entities_03_systems 1.26 80.4±2.90µs ? ?/sec 1.00 63.7±3.07µs ? ?/sec contrived/05x_entities_06_systems 1.27 161.6±13.47µs ? ?/sec 1.00 127.2±5.59µs ? ?/sec contrived/05x_entities_09_systems 1.22 228.0±7.76µs ? ?/sec 1.00 186.2±7.68µs ? ?/sec contrived/05x_entities_12_systems 1.20 289.5±6.21µs ? ?/sec 1.00 241.8±7.52µs ? ?/sec contrived/05x_entities_15_systems 1.18 357.3±11.24µs ? ?/sec 1.00 302.7±7.21µs ? ?/sec heavy_compute/base 1.01 302.4±3.52µs ? ?/sec 1.00 300.2±3.40µs ? ?/sec iter_fragmented/base 1.00 348.1±7.51ns ? ?/sec 1.01 351.9±8.32ns ? ?/sec iter_fragmented/foreach 1.03 239.8±23.78ns ? ?/sec 1.00 233.8±18.12ns ? ?/sec iter_fragmented/foreach_wide 1.00 3.9±0.13µs ? ?/sec 1.02 4.0±0.22µs ? ?/sec iter_fragmented/wide 1.18 4.6±0.15µs ? ?/sec 1.00 3.9±0.10µs ? ?/sec iter_fragmented_sparse/base 1.02 8.1±0.15ns ? ?/sec 1.00 7.9±0.56ns ? ?/sec iter_fragmented_sparse/foreach 1.00 7.8±0.22ns ? ?/sec 1.01 7.9±0.62ns ? ?/sec iter_fragmented_sparse/foreach_wide 1.00 37.2±1.17ns ? ?/sec 1.10 40.9±0.95ns ? ?/sec iter_fragmented_sparse/wide 1.09 48.4±2.13ns ? ?/sec 1.00 44.5±18.34ns ? ?/sec iter_simple/base 1.02 8.4±0.10µs ? ?/sec 1.00 8.2±0.14µs ? ?/sec iter_simple/foreach 1.01 8.3±0.07µs ? ?/sec 1.00 8.2±0.09µs ? ?/sec iter_simple/foreach_sparse_set 1.00 25.3±0.32µs ? ?/sec 1.02 25.7±0.42µs ? ?/sec iter_simple/foreach_wide 1.03 41.1±0.94µs ? ?/sec 1.00 39.9±0.41µs ? ?/sec iter_simple/foreach_wide_sparse_set 1.05 123.6±2.05µs ? ?/sec 1.00 118.1±2.78µs ? ?/sec iter_simple/sparse_set 1.14 30.5±1.40µs ? ?/sec 1.00 26.9±0.64µs ? ?/sec iter_simple/system 1.01 8.4±0.25µs ? ?/sec 1.00 8.4±0.11µs ? ?/sec iter_simple/wide 1.18 48.2±0.62µs ? ?/sec 1.00 40.7±0.38µs ? ?/sec iter_simple/wide_sparse_set 1.12 140.8±21.56µs ? ?/sec 1.00 126.0±2.30µs ? ?/sec query_get/50000_entities_sparse 1.17 378.6±7.60µs ? ?/sec 1.00 324.1±23.17µs ? ?/sec query_get/50000_entities_table 1.08 330.9±10.90µs ? ?/sec 1.00 306.8±4.98µs ? ?/sec query_get_component/50000_entities_sparse 1.00 976.7±19.55µs ? ?/sec 1.00 979.8±35.87µs ? ?/sec query_get_component/50000_entities_table 1.00 1029.0±15.11µs ? ?/sec 1.05 1080.0±59.18µs ? ?/sec query_get_component_simple/system 1.13 839.7±14.18µs ? ?/sec 1.00 742.8±10.72µs ? ?/sec query_get_component_simple/unchecked 1.01 909.0±15.17µs ? ?/sec 1.00 898.0±13.56µs ? ?/sec query_get_many_10/50000_calls_sparse 1.04 5.5±0.54ms ? ?/sec 1.00 5.3±0.67ms ? ?/sec query_get_many_10/50000_calls_table 1.01 4.9±0.49ms ? ?/sec 1.00 4.8±0.45ms ? ?/sec query_get_many_2/50000_calls_sparse 1.28 848.4±210.89µs ? ?/sec 1.00 664.8±47.69µs ? ?/sec query_get_many_2/50000_calls_table 1.05 779.0±73.85µs ? ?/sec 1.00 739.2±83.02µs ? ?/sec query_get_many_5/50000_calls_sparse 1.05 2.4±0.37ms ? ?/sec 1.00 2.3±0.33ms ? ?/sec query_get_many_5/50000_calls_table 1.00 1939.9±75.22µs ? ?/sec 1.04 2.0±0.19ms ? ?/sec run_criteria/yes_using_query/001_systems 1.00 3.7±0.38µs ? ?/sec 1.30 4.9±0.14µs ? ?/sec run_criteria/yes_using_query/006_systems 1.00 8.9±0.40µs ? ?/sec 1.17 10.3±0.57µs ? ?/sec run_criteria/yes_using_query/011_systems 1.00 13.9±0.49µs ? ?/sec 1.08 15.0±0.89µs ? ?/sec run_criteria/yes_using_query/016_systems 1.00 18.8±0.74µs ? ?/sec 1.00 18.8±1.43µs ? ?/sec run_criteria/yes_using_query/021_systems 1.07 24.1±0.87µs ? ?/sec 1.00 22.6±1.58µs ? ?/sec run_criteria/yes_using_query/026_systems 1.04 27.9±0.62µs ? ?/sec 1.00 26.8±1.71µs ? ?/sec run_criteria/yes_using_query/031_systems 1.09 33.3±1.03µs ? ?/sec 1.00 30.5±2.18µs ? ?/sec run_criteria/yes_using_query/036_systems 1.14 38.7±0.80µs ? ?/sec 1.00 33.9±1.75µs ? ?/sec run_criteria/yes_using_query/041_systems 1.18 43.7±1.07µs ? ?/sec 1.00 37.0±2.39µs ? ?/sec run_criteria/yes_using_query/046_systems 1.14 47.6±1.16µs ? ?/sec 1.00 41.9±2.09µs ? ?/sec run_criteria/yes_using_query/051_systems 1.17 52.9±2.04µs ? ?/sec 1.00 45.3±1.75µs ? ?/sec run_criteria/yes_using_query/056_systems 1.25 59.2±2.38µs ? ?/sec 1.00 47.2±2.01µs ? ?/sec run_criteria/yes_using_query/061_systems 1.28 66.1±15.84µs ? ?/sec 1.00 51.5±2.47µs ? ?/sec run_criteria/yes_using_query/066_systems 1.28 70.2±2.57µs ? ?/sec 1.00 54.7±2.58µs ? ?/sec run_criteria/yes_using_query/071_systems 1.30 75.5±2.27µs ? ?/sec 1.00 58.2±3.31µs ? ?/sec run_criteria/yes_using_query/076_systems 1.26 81.5±2.66µs ? ?/sec 1.00 64.5±3.13µs ? ?/sec run_criteria/yes_using_query/081_systems 1.29 89.7±2.58µs ? ?/sec 1.00 69.3±3.47µs ? ?/sec run_criteria/yes_using_query/086_systems 1.33 95.6±3.39µs ? ?/sec 1.00 71.8±3.48µs ? ?/sec run_criteria/yes_using_query/091_systems 1.25 102.0±3.67µs ? ?/sec 1.00 81.4±4.82µs ? ?/sec run_criteria/yes_using_query/096_systems 1.33 111.7±3.29µs ? ?/sec 1.00 83.8±4.15µs ? ?/sec run_criteria/yes_using_query/101_systems 1.29 113.2±12.04µs ? ?/sec 1.00 87.7±5.15µs ? ?/sec world_query_for_each/50000_entities_sparse 1.00 47.4±0.51µs ? ?/sec 1.00 47.3±0.33µs ? ?/sec world_query_for_each/50000_entities_table 1.00 27.2±0.50µs ? ?/sec 1.00 27.2±0.17µs ? ?/sec world_query_get/50000_entities_sparse_wide 1.09 210.5±1.78µs ? ?/sec 1.00 192.5±2.61µs ? ?/sec world_query_get/50000_entities_table 1.00 127.7±2.09µs ? ?/sec 1.07 136.2±5.95µs ? ?/sec world_query_get/50000_entities_table_wide 1.00 209.8±2.37µs ? ?/sec 1.15 240.6±2.04µs ? ?/sec world_query_iter/50000_entities_sparse 1.00 54.2±0.36µs ? ?/sec 1.01 54.7±0.61µs ? ?/sec world_query_iter/50000_entities_table 1.00 27.2±0.31µs ? ?/sec 1.00 27.3±0.64µs ? ?/sec ``` </details> NOTE: This PR includes a change to enable LTO on our benchmarks to get a "fully optimized" baseline for our benchmarks. Both the main and the current PR's results were with LTO enabled.
2022-11-04 06:04:55 +00:00
[profile.release]
opt-level = 3
lto = true
[[bench]]
name = "change_detection"
path = "benches/bevy_ecs/change_detection.rs"
harness = false
[[bench]]
name = "ecs"
path = "benches/bevy_ecs/benches.rs"
harness = false
[[bench]]
name = "reflect_list"
path = "benches/bevy_reflect/list.rs"
harness = false
[[bench]]
name = "reflect_map"
path = "benches/bevy_reflect/map.rs"
harness = false
[[bench]]
name = "reflect_struct"
path = "benches/bevy_reflect/struct.rs"
harness = false
[[bench]]
name = "parse_reflect_path"
path = "benches/bevy_reflect/path.rs"
harness = false
[[bench]]
name = "iter"
path = "benches/bevy_tasks/iter.rs"
2020-09-10 19:54:24 +00:00
harness = false
Add Beziers to `bevy_math` (#7653) # Objective - Adds foundational math for Bezier curves, useful for UI/2D/3D animation and smooth paths. https://user-images.githubusercontent.com/2632925/218883143-e138f994-1795-40da-8c59-21d779666991.mp4 ## Solution - Adds the generic `Bezier` type, and a `Point` trait. The `Point` trait allows us to use control points of any dimension, as long as they support vector math. I've implemented it for `f32`(1D), `Vec2`(2D), and `Vec3`/`Vec3A`(3D). - Adds `CubicBezierEasing` on top of `Bezier` with the addition of an implementation of cubic Bezier easing, which is a foundational tool for UI animation. - This involves solving for $t$ in the parametric Bezier function $B(t)$ using the Newton-Raphson method to find a value with error $\leq$ 1e-7, capped at 8 iterations. - Added type aliases for common Bezier curves: `CubicBezier2d`, `CubicBezier3d`, `QuadraticBezier2d`, and `QuadraticBezier3d`. These types use `Vec3A` to represent control points, as this was found to have an 80-90% speedup over using `Vec3`. - Benchmarking shows quadratic/cubic Bezier evaluations $B(t)$ take \~1.8/2.4ns respectively. Easing, which requires an iterative solve takes \~50ns for cubic Beziers. --- ## Changelog - Added `CubicBezier2d`, `CubicBezier3d`, `QuadraticBezier2d`, and `QuadraticBezier3d` types with methods for sampling position, velocity, and acceleration. The generic `Bezier` type is also available, and generic over any degree of Bezier curve. - Added `CubicBezierEasing`, with additional methods to allow for smooth easing animations.
2023-02-20 18:34:52 +00:00
[[bench]]
name = "bezier"
path = "benches/bevy_math/bezier.rs"
harness = false
Optimize `Entity::eq` (#10519) (This is my first PR here, so I've probably missed some things. Please let me know what else I should do to help you as a reviewer!) # Objective Due to https://github.com/rust-lang/rust/issues/117800, the `derive`'d `PartialEq::eq` on `Entity` isn't as good as it could be. Since that's used in hashtable lookup, let's improve it. ## Solution The derived `PartialEq::eq` short-circuits if the generation doesn't match. However, having a branch there is sub-optimal, especially on 64-bit systems like x64 that could just load the whole `Entity` in one load anyway. Due to complications around `poison` in LLVM and the exact details of what unsafe code is allowed to do with reference in Rust (https://github.com/rust-lang/unsafe-code-guidelines/issues/346), LLVM isn't allowed to completely remove the short-circuiting. `&Entity` is marked `dereferencable(8)` so LLVM knows it's allowed to *load* all 8 bytes -- and does so -- but it has to assume that the `index` might be undef/poison if the `generation` doesn't match, and thus while it finds a way to do it without needing a branch, it has to do something slightly more complicated than optimal to combine the results. (LLVM is allowed to change non-short-circuiting code to use branches, but not the other way around.) Here's a link showing the codegen today: <https://rust.godbolt.org/z/9WzjxrY7c> ```rust #[no_mangle] pub fn demo_eq_ref(a: &Entity, b: &Entity) -> bool { a == b } ``` ends up generating the following assembly: ```asm demo_eq_ref: movq xmm0, qword ptr [rdi] movq xmm1, qword ptr [rsi] pcmpeqd xmm1, xmm0 pshufd xmm0, xmm1, 80 movmskpd eax, xmm0 cmp eax, 3 sete al ret ``` (It's usually not this bad in real uses after inlining and LTO, but it makes a strong demo.) This PR manually implements `PartialEq::eq` *without* short-circuiting, and because that tells LLVM that neither the generations nor the index can be poison, it doesn't need to be so careful and can generate the "just compare the two 64-bit values" code you'd have probably already expected: ```asm demo_eq_ref: mov rax, qword ptr [rsi] cmp qword ptr [rdi], rax sete al ret ``` Since this doesn't change the representation of `Entity`, if it's instead passed by *value*, then each `Entity` is two `u32` registers, and the old and the new code do exactly the same thing. (Other approaches, like changing `Entity` to be `[u32; 2]` or `u64`, affect this case.) This should hopefully merge easily with changes like https://github.com/bevyengine/bevy/pull/9907 that also want to change `Entity`. ## Benchmarks I'm not super-confident that I got my machine fully consistent for benchmarking, but whether I run the old or the new one first I get reasonably consistent results. Here's a fairly typical example of the benchmarks I added in this PR: ![image](https://github.com/bevyengine/bevy/assets/18526288/24226308-4616-4082-b0ff-88fc06285ef1) Building the sets seems to be basically the same. It's usually reported as noise, but sometimes I see a few percent slower or faster. But lookup hits in particular -- since a hit checks that the key is equal -- consistently shows around 10% improvement. `cargo run --example many_cubes --features bevy/trace_tracy --release -- --benchmark` showed as slightly faster with this change, though if I had to bet I'd probably say it's more noise than meaningful (but at least it's not worse either): ![image](https://github.com/bevyengine/bevy/assets/18526288/58bb8c96-9c45-487f-a5ab-544bbfe9fba0) This is my first PR here -- and my first time running Tracy -- so please let me know what else I should run, or run things on your own more reliable machines to double-check. --- ## Changelog (probably not worth including) Changed: micro-optimized `Entity::eq` to help LLVM slightly. ## Migration Guide (I really hope nobody was using this on uninitialized entities where sufficiently tortured `unsafe` could could technically notice that this has changed.)
2023-11-14 02:06:21 +00:00
[[bench]]
name = "torus"
path = "benches/bevy_render/torus.rs"
harness = false
Optimize `Entity::eq` (#10519) (This is my first PR here, so I've probably missed some things. Please let me know what else I should do to help you as a reviewer!) # Objective Due to https://github.com/rust-lang/rust/issues/117800, the `derive`'d `PartialEq::eq` on `Entity` isn't as good as it could be. Since that's used in hashtable lookup, let's improve it. ## Solution The derived `PartialEq::eq` short-circuits if the generation doesn't match. However, having a branch there is sub-optimal, especially on 64-bit systems like x64 that could just load the whole `Entity` in one load anyway. Due to complications around `poison` in LLVM and the exact details of what unsafe code is allowed to do with reference in Rust (https://github.com/rust-lang/unsafe-code-guidelines/issues/346), LLVM isn't allowed to completely remove the short-circuiting. `&Entity` is marked `dereferencable(8)` so LLVM knows it's allowed to *load* all 8 bytes -- and does so -- but it has to assume that the `index` might be undef/poison if the `generation` doesn't match, and thus while it finds a way to do it without needing a branch, it has to do something slightly more complicated than optimal to combine the results. (LLVM is allowed to change non-short-circuiting code to use branches, but not the other way around.) Here's a link showing the codegen today: <https://rust.godbolt.org/z/9WzjxrY7c> ```rust #[no_mangle] pub fn demo_eq_ref(a: &Entity, b: &Entity) -> bool { a == b } ``` ends up generating the following assembly: ```asm demo_eq_ref: movq xmm0, qword ptr [rdi] movq xmm1, qword ptr [rsi] pcmpeqd xmm1, xmm0 pshufd xmm0, xmm1, 80 movmskpd eax, xmm0 cmp eax, 3 sete al ret ``` (It's usually not this bad in real uses after inlining and LTO, but it makes a strong demo.) This PR manually implements `PartialEq::eq` *without* short-circuiting, and because that tells LLVM that neither the generations nor the index can be poison, it doesn't need to be so careful and can generate the "just compare the two 64-bit values" code you'd have probably already expected: ```asm demo_eq_ref: mov rax, qword ptr [rsi] cmp qword ptr [rdi], rax sete al ret ``` Since this doesn't change the representation of `Entity`, if it's instead passed by *value*, then each `Entity` is two `u32` registers, and the old and the new code do exactly the same thing. (Other approaches, like changing `Entity` to be `[u32; 2]` or `u64`, affect this case.) This should hopefully merge easily with changes like https://github.com/bevyengine/bevy/pull/9907 that also want to change `Entity`. ## Benchmarks I'm not super-confident that I got my machine fully consistent for benchmarking, but whether I run the old or the new one first I get reasonably consistent results. Here's a fairly typical example of the benchmarks I added in this PR: ![image](https://github.com/bevyengine/bevy/assets/18526288/24226308-4616-4082-b0ff-88fc06285ef1) Building the sets seems to be basically the same. It's usually reported as noise, but sometimes I see a few percent slower or faster. But lookup hits in particular -- since a hit checks that the key is equal -- consistently shows around 10% improvement. `cargo run --example many_cubes --features bevy/trace_tracy --release -- --benchmark` showed as slightly faster with this change, though if I had to bet I'd probably say it's more noise than meaningful (but at least it's not worse either): ![image](https://github.com/bevyengine/bevy/assets/18526288/58bb8c96-9c45-487f-a5ab-544bbfe9fba0) This is my first PR here -- and my first time running Tracy -- so please let me know what else I should run, or run things on your own more reliable machines to double-check. --- ## Changelog (probably not worth including) Changed: micro-optimized `Entity::eq` to help LLVM slightly. ## Migration Guide (I really hope nobody was using this on uninitialized entities where sufficiently tortured `unsafe` could could technically notice that this has changed.)
2023-11-14 02:06:21 +00:00
[[bench]]
name = "entity_hash"
path = "benches/bevy_ecs/world/entity_hash.rs"
Optimize `Entity::eq` (#10519) (This is my first PR here, so I've probably missed some things. Please let me know what else I should do to help you as a reviewer!) # Objective Due to https://github.com/rust-lang/rust/issues/117800, the `derive`'d `PartialEq::eq` on `Entity` isn't as good as it could be. Since that's used in hashtable lookup, let's improve it. ## Solution The derived `PartialEq::eq` short-circuits if the generation doesn't match. However, having a branch there is sub-optimal, especially on 64-bit systems like x64 that could just load the whole `Entity` in one load anyway. Due to complications around `poison` in LLVM and the exact details of what unsafe code is allowed to do with reference in Rust (https://github.com/rust-lang/unsafe-code-guidelines/issues/346), LLVM isn't allowed to completely remove the short-circuiting. `&Entity` is marked `dereferencable(8)` so LLVM knows it's allowed to *load* all 8 bytes -- and does so -- but it has to assume that the `index` might be undef/poison if the `generation` doesn't match, and thus while it finds a way to do it without needing a branch, it has to do something slightly more complicated than optimal to combine the results. (LLVM is allowed to change non-short-circuiting code to use branches, but not the other way around.) Here's a link showing the codegen today: <https://rust.godbolt.org/z/9WzjxrY7c> ```rust #[no_mangle] pub fn demo_eq_ref(a: &Entity, b: &Entity) -> bool { a == b } ``` ends up generating the following assembly: ```asm demo_eq_ref: movq xmm0, qword ptr [rdi] movq xmm1, qword ptr [rsi] pcmpeqd xmm1, xmm0 pshufd xmm0, xmm1, 80 movmskpd eax, xmm0 cmp eax, 3 sete al ret ``` (It's usually not this bad in real uses after inlining and LTO, but it makes a strong demo.) This PR manually implements `PartialEq::eq` *without* short-circuiting, and because that tells LLVM that neither the generations nor the index can be poison, it doesn't need to be so careful and can generate the "just compare the two 64-bit values" code you'd have probably already expected: ```asm demo_eq_ref: mov rax, qword ptr [rsi] cmp qword ptr [rdi], rax sete al ret ``` Since this doesn't change the representation of `Entity`, if it's instead passed by *value*, then each `Entity` is two `u32` registers, and the old and the new code do exactly the same thing. (Other approaches, like changing `Entity` to be `[u32; 2]` or `u64`, affect this case.) This should hopefully merge easily with changes like https://github.com/bevyengine/bevy/pull/9907 that also want to change `Entity`. ## Benchmarks I'm not super-confident that I got my machine fully consistent for benchmarking, but whether I run the old or the new one first I get reasonably consistent results. Here's a fairly typical example of the benchmarks I added in this PR: ![image](https://github.com/bevyengine/bevy/assets/18526288/24226308-4616-4082-b0ff-88fc06285ef1) Building the sets seems to be basically the same. It's usually reported as noise, but sometimes I see a few percent slower or faster. But lookup hits in particular -- since a hit checks that the key is equal -- consistently shows around 10% improvement. `cargo run --example many_cubes --features bevy/trace_tracy --release -- --benchmark` showed as slightly faster with this change, though if I had to bet I'd probably say it's more noise than meaningful (but at least it's not worse either): ![image](https://github.com/bevyengine/bevy/assets/18526288/58bb8c96-9c45-487f-a5ab-544bbfe9fba0) This is my first PR here -- and my first time running Tracy -- so please let me know what else I should run, or run things on your own more reliable machines to double-check. --- ## Changelog (probably not worth including) Changed: micro-optimized `Entity::eq` to help LLVM slightly. ## Migration Guide (I really hope nobody was using this on uninitialized entities where sufficiently tortured `unsafe` could could technically notice that this has changed.)
2023-11-14 02:06:21 +00:00
harness = false