|
| 1 | +use criterion::{black_box, criterion_group, criterion_main, Criterion}; |
| 2 | +use p64::*; |
| 3 | + |
| 4 | +fn make_heels() -> HeelPlanes { |
| 5 | + HeelPlanes::new([ |
| 6 | + 0xAAAA_AAAA_AAAA_AAAA, |
| 7 | + 0x5555_5555_5555_5555, |
| 8 | + 0xFFFF_0000_FFFF_0000, |
| 9 | + 0x0000_FFFF_0000_FFFF, |
| 10 | + 0xFF00_FF00_FF00_FF00, |
| 11 | + 0x00FF_00FF_00FF_00FF, |
| 12 | + 0xF0F0_F0F0_F0F0_F0F0, |
| 13 | + 0x0F0F_0F0F_0F0F_0F0F, |
| 14 | + ]) |
| 15 | +} |
| 16 | + |
| 17 | +fn bench_expand(c: &mut Criterion) { |
| 18 | + let heels = make_heels(); |
| 19 | + c.bench_function("expand_8_to_64", |b| { |
| 20 | + b.iter(|| black_box(black_box(&heels).expand())) |
| 21 | + }); |
| 22 | +} |
| 23 | + |
| 24 | +fn bench_attend(c: &mut Criterion) { |
| 25 | + let palette = make_heels().expand(); |
| 26 | + let query = 0xDEAD_BEEF_CAFE_BABEu64; |
| 27 | + |
| 28 | + c.bench_function("attend_single", |b| { |
| 29 | + b.iter(|| black_box(black_box(&palette).attend(black_box(query), 16))) |
| 30 | + }); |
| 31 | +} |
| 32 | + |
| 33 | +fn bench_attend_batch(c: &mut Criterion) { |
| 34 | + let palette = make_heels().expand(); |
| 35 | + let queries: Vec<u64> = (0..1024).map(|i| i * 0x0123_4567_89AB + 0xDEAD).collect(); |
| 36 | + |
| 37 | + c.bench_function("attend_1024_queries", |b| { |
| 38 | + b.iter(|| { |
| 39 | + for &q in &queries { |
| 40 | + black_box(palette.attend(black_box(q), 16)); |
| 41 | + } |
| 42 | + }) |
| 43 | + }); |
| 44 | +} |
| 45 | + |
| 46 | +fn bench_moe_gate(c: &mut Criterion) { |
| 47 | + let heels = make_heels(); |
| 48 | + let query = 0xDEAD_BEEF_CAFE_BABEu64; |
| 49 | + |
| 50 | + c.bench_function("moe_gate", |b| { |
| 51 | + b.iter(|| black_box(black_box(&heels).moe_gate(black_box(query), 20))) |
| 52 | + }); |
| 53 | +} |
| 54 | + |
| 55 | +fn bench_soft_moe(c: &mut Criterion) { |
| 56 | + let heels = make_heels(); |
| 57 | + let query = 0xDEAD_BEEF_CAFE_BABEu64; |
| 58 | + |
| 59 | + c.bench_function("soft_moe", |b| { |
| 60 | + b.iter(|| black_box(black_box(&heels).soft_moe(black_box(query), 20))) |
| 61 | + }); |
| 62 | +} |
| 63 | + |
| 64 | +fn bench_denoise(c: &mut Criterion) { |
| 65 | + let palette = make_heels().expand(); |
| 66 | + let noisy = 0x1234_5678_9ABC_DEF0u64; |
| 67 | + |
| 68 | + c.bench_function("denoise_10_steps", |b| { |
| 69 | + b.iter(|| black_box(black_box(&palette).denoise(black_box(noisy), 10, 32))) |
| 70 | + }); |
| 71 | +} |
| 72 | + |
| 73 | +fn bench_nearest_k(c: &mut Criterion) { |
| 74 | + let palette = make_heels().expand(); |
| 75 | + let query = 0xDEAD_BEEF_CAFE_BABEu64; |
| 76 | + |
| 77 | + c.bench_function("nearest_k_8", |b| { |
| 78 | + b.iter(|| black_box(black_box(&palette).nearest_k(black_box(query), 8))) |
| 79 | + }); |
| 80 | +} |
| 81 | + |
| 82 | +criterion_group!( |
| 83 | + benches, |
| 84 | + bench_expand, |
| 85 | + bench_attend, |
| 86 | + bench_attend_batch, |
| 87 | + bench_moe_gate, |
| 88 | + bench_soft_moe, |
| 89 | + bench_denoise, |
| 90 | + bench_nearest_k, |
| 91 | +); |
| 92 | +criterion_main!(benches); |
0 commit comments