1 //! Benchmarking module.
2 pub use std::hint::black_box;
5 event::CompletedTest, options::BenchMode, test_result::TestResult, types::TestDesc, Sender,
11 use std::panic::{catch_unwind, AssertUnwindSafe};
12 use std::sync::{Arc, Mutex};
13 use std::time::{Duration, Instant};
15 /// Manager of the benchmarking runs.
17 /// This is fed into functions marked with `#[bench]` to allow for
18 /// set-up & tear-down before running a piece of code repeatedly via a
23 summary: Option<stats::Summary>,
28 /// Callback for benchmark functions to run in their body.
29 pub fn iter<T, F>(&mut self, mut inner: F)
33 if self.mode == BenchMode::Single {
34 ns_iter_inner(&mut inner, 1);
38 self.summary = Some(iter(&mut inner));
41 pub fn bench<F>(&mut self, mut f: F) -> Option<stats::Summary>
43 F: FnMut(&mut Bencher),
50 #[derive(Debug, Clone, PartialEq)]
51 pub struct BenchSamples {
52 pub ns_iter_summ: stats::Summary,
56 pub fn fmt_bench_samples(bs: &BenchSamples) -> String {
58 let mut output = String::new();
60 let median = bs.ns_iter_summ.median as usize;
61 let deviation = (bs.ns_iter_summ.max - bs.ns_iter_summ.min) as usize;
65 "{:>11} ns/iter (+/- {})",
66 fmt_thousands_sep(median, ','),
67 fmt_thousands_sep(deviation, ',')
71 write!(output, " = {} MB/s", bs.mb_s).unwrap();
76 // Format a number with thousands separators
77 fn fmt_thousands_sep(mut n: usize, sep: char) -> String {
79 let mut output = String::new();
80 let mut trailing = false;
81 for &pow in &[9, 6, 3, 0] {
82 let base = 10_usize.pow(pow);
83 if pow == 0 || trailing || n / base != 0 {
85 write!(output, "{}", n / base).unwrap();
87 write!(output, "{:03}", n / base).unwrap();
100 fn ns_iter_inner<T, F>(inner: &mut F, k: u64) -> u64
104 let start = Instant::now();
108 start.elapsed().as_nanos() as u64
111 pub fn iter<T, F>(inner: &mut F) -> stats::Summary
115 // Initial bench run to get ballpark figure.
116 let ns_single = ns_iter_inner(inner, 1);
118 // Try to estimate iter count for 1ms falling back to 1m
119 // iterations if first run took < 1ns.
120 let ns_target_total = 1_000_000; // 1ms
121 let mut n = ns_target_total / cmp::max(1, ns_single);
123 // if the first run took more than 1ms we don't want to just
124 // be left doing 0 iterations on every loop. The unfortunate
125 // side effect of not being able to do as many runs is
126 // automatically handled by the statistical analysis below
127 // (i.e., larger error bars).
130 let mut total_run = Duration::new(0, 0);
131 let samples: &mut [f64] = &mut [0.0_f64; 50];
133 let loop_start = Instant::now();
135 for p in &mut *samples {
136 *p = ns_iter_inner(inner, n) as f64 / n as f64;
139 stats::winsorize(samples, 5.0);
140 let summ = stats::Summary::new(samples);
142 for p in &mut *samples {
143 let ns = ns_iter_inner(inner, 5 * n);
144 *p = ns as f64 / (5 * n) as f64;
147 stats::winsorize(samples, 5.0);
148 let summ5 = stats::Summary::new(samples);
150 let loop_run = loop_start.elapsed();
152 // If we've run for 100ms and seem to have converged to a
154 if loop_run > Duration::from_millis(100)
155 && summ.median_abs_dev_pct < 1.0
156 && summ.median - summ5.median < summ5.median_abs_dev
161 total_run += loop_run;
162 // Longest we ever run for is 3s.
163 if total_run > Duration::from_secs(3) {
167 // If we overflow here just return the results so far. We check a
168 // multiplier of 10 because we're about to multiply by 2 and the
169 // next iteration of the loop will also multiply by 5 (to calculate
171 n = match n.checked_mul(10) {
180 pub fn benchmark<F>(desc: TestDesc, monitor_ch: Sender<CompletedTest>, nocapture: bool, f: F)
182 F: FnMut(&mut Bencher),
184 let mut bs = Bencher { mode: BenchMode::Auto, summary: None, bytes: 0 };
186 let data = Arc::new(Mutex::new(Vec::new()));
189 io::set_output_capture(Some(data.clone()));
192 let result = catch_unwind(AssertUnwindSafe(|| bs.bench(f)));
194 io::set_output_capture(None);
196 let test_result = match result {
198 Ok(Some(ns_iter_summ)) => {
199 let ns_iter = cmp::max(ns_iter_summ.median as u64, 1);
200 let mb_s = bs.bytes * 1000 / ns_iter;
202 let bs = BenchSamples { ns_iter_summ, mb_s: mb_s as usize };
203 TestResult::TrBench(bs)
206 // iter not called, so no data.
207 // FIXME: error in this case?
208 let samples: &mut [f64] = &mut [0.0_f64; 1];
209 let bs = BenchSamples { ns_iter_summ: stats::Summary::new(samples), mb_s: 0 };
210 TestResult::TrBench(bs)
212 Err(_) => TestResult::TrFailed,
215 let stdout = data.lock().unwrap().to_vec();
216 let message = CompletedTest::new(desc, test_result, None, stdout);
217 monitor_ch.send(message).unwrap();
220 pub fn run_once<F>(f: F)
222 F: FnMut(&mut Bencher),
224 let mut bs = Bencher { mode: BenchMode::Single, summary: None, bytes: 0 };