1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
//! Benchmarking module.

use super::{
    event::CompletedTest,
    options::BenchMode,
    test_result::TestResult,
    types::{TestDesc, TestId},
    Sender,
};

use crate::stats;
use std::cmp;
use std::io;
use std::panic::{catch_unwind, AssertUnwindSafe};
use std::sync::{Arc, Mutex};
use std::time::{Duration, Instant};

/// An identity function that *__hints__* to the compiler to be maximally pessimistic about what
/// `black_box` could do.
///
/// See [`std::hint::black_box`] for details.
#[inline(always)]
pub fn black_box<T>(dummy: T) -> T {
    std::hint::black_box(dummy)
}

/// Manager of the benchmarking runs.
///
/// This is fed into functions marked with `#[bench]` to allow for
/// set-up & tear-down before running a piece of code repeatedly via a
/// call to `iter`.
#[derive(Clone)]
pub struct Bencher {
    mode: BenchMode,
    summary: Option<stats::Summary>,
    pub bytes: u64,
}

impl Bencher {
    /// Callback for benchmark functions to run in their body.
    pub fn iter<T, F>(&mut self, mut inner: F)
    where
        F: FnMut() -> T,
    {
        if self.mode == BenchMode::Single {
            ns_iter_inner(&mut inner, 1);
            return;
        }

        self.summary = Some(iter(&mut inner));
    }

    pub fn bench<F>(&mut self, mut f: F) -> Result<Option<stats::Summary>, String>
    where
        F: FnMut(&mut Bencher) -> Result<(), String>,
    {
        let result = f(self);
        result.map(|_| self.summary)
    }
}

#[derive(Debug, Clone, PartialEq)]
pub struct BenchSamples {
    pub ns_iter_summ: stats::Summary,
    pub mb_s: usize,
}

pub fn fmt_bench_samples(bs: &BenchSamples) -> String {
    use std::fmt::Write;
    let mut output = String::new();

    let median = bs.ns_iter_summ.median;
    let deviation = bs.ns_iter_summ.max - bs.ns_iter_summ.min;

    write!(
        output,
        "{:>14} ns/iter (+/- {})",
        fmt_thousands_sep(median, ','),
        fmt_thousands_sep(deviation, ',')
    )
    .unwrap();
    if bs.mb_s != 0 {
        write!(output, " = {} MB/s", bs.mb_s).unwrap();
    }
    output
}

// Format a number with thousands separators
fn fmt_thousands_sep(mut n: f64, sep: char) -> String {
    use std::fmt::Write;
    let mut output = String::new();
    let mut trailing = false;
    for &pow in &[9, 6, 3, 0] {
        let base = 10_usize.pow(pow);
        if pow == 0 || trailing || n / base as f64 >= 1.0 {
            match (pow, trailing) {
                // modern CPUs can execute multiple instructions per nanosecond
                // e.g. benching an ADD takes about 0.25ns.
                (0, true) => write!(output, "{:06.2}", n / base as f64).unwrap(),
                (0, false) => write!(output, "{:.2}", n / base as f64).unwrap(),
                (_, true) => write!(output, "{:03}", n as usize / base).unwrap(),
                _ => write!(output, "{}", n as usize / base).unwrap(),
            }
            if pow != 0 {
                output.push(sep);
            }
            trailing = true;
        }
        n %= base as f64;
    }

    output
}

fn ns_iter_inner<T, F>(inner: &mut F, k: u64) -> u64
where
    F: FnMut() -> T,
{
    let start = Instant::now();
    for _ in 0..k {
        black_box(inner());
    }
    start.elapsed().as_nanos() as u64
}

pub fn iter<T, F>(inner: &mut F) -> stats::Summary
where
    F: FnMut() -> T,
{
    // Initial bench run to get ballpark figure.
    let ns_single = ns_iter_inner(inner, 1);

    // Try to estimate iter count for 1ms falling back to 1m
    // iterations if first run took < 1ns.
    let ns_target_total = 1_000_000; // 1ms
    let mut n = ns_target_total / cmp::max(1, ns_single);

    // if the first run took more than 1ms we don't want to just
    // be left doing 0 iterations on every loop. The unfortunate
    // side effect of not being able to do as many runs is
    // automatically handled by the statistical analysis below
    // (i.e., larger error bars).
    n = cmp::max(1, n);

    let mut total_run = Duration::new(0, 0);
    let samples: &mut [f64] = &mut [0.0_f64; 50];
    loop {
        let loop_start = Instant::now();

        for p in &mut *samples {
            *p = ns_iter_inner(inner, n) as f64 / n as f64;
        }

        stats::winsorize(samples, 5.0);
        let summ = stats::Summary::new(samples);

        for p in &mut *samples {
            let ns = ns_iter_inner(inner, 5 * n);
            *p = ns as f64 / (5 * n) as f64;
        }

        stats::winsorize(samples, 5.0);
        let summ5 = stats::Summary::new(samples);

        let loop_run = loop_start.elapsed();

        // If we've run for 100ms and seem to have converged to a
        // stable median.
        if loop_run > Duration::from_millis(100)
            && summ.median_abs_dev_pct < 1.0
            && summ.median - summ5.median < summ5.median_abs_dev
        {
            return summ5;
        }

        total_run += loop_run;
        // Longest we ever run for is 3s.
        if total_run > Duration::from_secs(3) {
            return summ5;
        }

        // If we overflow here just return the results so far. We check a
        // multiplier of 10 because we're about to multiply by 2 and the
        // next iteration of the loop will also multiply by 5 (to calculate
        // the summ5 result)
        n = match n.checked_mul(10) {
            Some(_) => n * 2,
            None => {
                return summ5;
            }
        };
    }
}

pub fn benchmark<F>(
    id: TestId,
    desc: TestDesc,
    monitor_ch: Sender<CompletedTest>,
    nocapture: bool,
    f: F,
) where
    F: FnMut(&mut Bencher) -> Result<(), String>,
{
    let mut bs = Bencher { mode: BenchMode::Auto, summary: None, bytes: 0 };

    let data = Arc::new(Mutex::new(Vec::new()));

    if !nocapture {
        io::set_output_capture(Some(data.clone()));
    }

    let result = catch_unwind(AssertUnwindSafe(|| bs.bench(f)));

    io::set_output_capture(None);

    let test_result = match result {
        //bs.bench(f) {
        Ok(Ok(Some(ns_iter_summ))) => {
            let ns_iter = cmp::max(ns_iter_summ.median as u64, 1);
            let mb_s = bs.bytes * 1000 / ns_iter;

            let bs = BenchSamples { ns_iter_summ, mb_s: mb_s as usize };
            TestResult::TrBench(bs)
        }
        Ok(Ok(None)) => {
            // iter not called, so no data.
            // FIXME: error in this case?
            let samples: &mut [f64] = &mut [0.0_f64; 1];
            let bs = BenchSamples { ns_iter_summ: stats::Summary::new(samples), mb_s: 0 };
            TestResult::TrBench(bs)
        }
        Err(_) => TestResult::TrFailed,
        Ok(Err(_)) => TestResult::TrFailed,
    };

    let stdout = data.lock().unwrap().to_vec();
    let message = CompletedTest::new(id, desc, test_result, None, stdout);
    monitor_ch.send(message).unwrap();
}

pub fn run_once<F>(f: F) -> Result<(), String>
where
    F: FnMut(&mut Bencher) -> Result<(), String>,
{
    let mut bs = Bencher { mode: BenchMode::Single, summary: None, bytes: 0 };
    bs.bench(f).map(|_| ())
}