|
| 1 | +use criterion::{black_box, criterion_group, criterion_main, BatchSize, BenchmarkId, Criterion}; |
| 2 | +use tract_linalg::frame::activations::{definitions, reference, ActivationKer, Program}; |
| 3 | + |
| 4 | +fn crit(c: &mut Criterion, name: &str, r: impl Fn(f32) -> f32, prog: &Program<f32>) { |
| 5 | + let mut group = c.benchmark_group(name); |
| 6 | + for size in [1i32, 32, 256, 1024, 8192].iter() { |
| 7 | + group.throughput(criterion::Throughput::Elements(*size as u64)); |
| 8 | + group.bench_with_input(BenchmarkId::new("Reference", size), size, |b, size| { |
| 9 | + b.iter_batched( |
| 10 | + || vec![1.0f32; *size as usize], |
| 11 | + |v| { |
| 12 | + for x in v { |
| 13 | + r(black_box(x)); |
| 14 | + } |
| 15 | + }, |
| 16 | + BatchSize::LargeInput, |
| 17 | + ) |
| 18 | + }); |
| 19 | + #[allow(unused_mut)] |
| 20 | + let mut vms = vec!(tract_linalg::generic::activations::SActivations::act()); |
| 21 | + #[cfg(target_arch="aarch64")] |
| 22 | + { |
| 23 | + vms.push(tract_linalg::arm64::arm64simd_act_f32_32n::act()); |
| 24 | + } |
| 25 | + |
| 26 | + for vm in vms { |
| 27 | + group.bench_with_input(BenchmarkId::new(vm.name(), size), size, |b, size| { |
| 28 | + b.iter_batched( |
| 29 | + || vec![1.0f32; *size as usize], |
| 30 | + |mut v| vm.run(prog, &mut v), |
| 31 | + BatchSize::LargeInput, |
| 32 | + ) |
| 33 | + }); |
| 34 | + } |
| 35 | + } |
| 36 | +} |
| 37 | + |
| 38 | +fn criterion_benchmark(c: &mut Criterion) { |
| 39 | + crit(c, "relu", reference::relu, &definitions::relu()); |
| 40 | + crit(c, "hardswish", reference::hardswish, &definitions::hard_swish()); |
| 41 | + /* |
| 42 | + crit(c, "exp2f", reference::exp2f, &definitions::exp2f()); |
| 43 | + crit(c, "sigmoid", reference::sigmoid, &definitions::sigmoid()); |
| 44 | + */ |
| 45 | +} |
| 46 | + |
| 47 | +criterion_group!(benches, criterion_benchmark); |
| 48 | +criterion_main!(benches); |
0 commit comments