tango_bench/

cli.rs

//! Contains functionality of a `cargo bench` harness
use crate::{
    dylib::{FunctionIdx, Spi, SpiModeKind},
    CacheFirewall, Error, FlatSampleLength, LinearSampleLength, MeasurementSettings,
    RandomSampleLength, SampleLength, SampleLengthKind,
};
use anyhow::{bail, Context};
use clap::{ArgAction, Parser};
use colorz::mode::{self, Mode};
use core::fmt;
use glob_match::glob_match;
use std::{
    env::{self, args, temp_dir},
    fmt::Display,
    fs,
    io::{stderr, Write},
    num::NonZeroUsize,
    path::{Path, PathBuf},
    process::{Command, ExitCode, Stdio},
    str::FromStr,
    time::Duration,
};

pub type Result<T> = anyhow::Result<T>;
pub(crate) type StdResult<T, E> = std::result::Result<T, E>;

#[derive(Parser, Debug)]
enum BenchmarkMode {
    List {
        #[command(flatten)]
        bench_flags: CargoBenchFlags,
    },
    Compare(PairedOpts),
    Solo(SoloOpts),
}

#[derive(Parser, Debug)]
struct PairedOpts {
    #[command(flatten)]
    bench_flags: CargoBenchFlags,

    /// Path to the executable to test against. Tango will test against itself if no executable given
    path: Option<PathBuf>,

    /// write CSV dumps of all the measurements in a given location
    #[arg(short = 'd', long = "dump")]
    path_to_dump: Option<PathBuf>,

    /// generate gnuplot graphs for each test (requires --dump [path] to be specified)
    #[arg(long = "gnuplot")]
    gnuplot: bool,

    /// seed for the random number generator or omit to use a random seed
    #[arg(long = "seed")]
    seed: Option<u64>,

    /// Number of samples to take for each test
    #[arg(short = 's', long = "samples")]
    samples: Option<NonZeroUsize>,

    /// The strategy to decide the number of iterations to run for each sample (values: flat, linear, random)
    #[arg(long = "sampler")]
    sampler: Option<SampleLengthKind>,

    /// Duration of each sample in seconds
    #[arg(short = 't', long = "time")]
    time: Option<f64>,

    /// Fail if the difference between the two measurements is greater than the given threshold in percent
    #[arg(long = "fail-threshold")]
    fail_threshold: Option<f64>,

    /// Should we terminate early if --fail-threshold is exceed
    #[arg(long = "fail-fast")]
    fail_fast: bool,

    /// Perform a read of a dummy data between samsples to minimize the effect of cache on the performance
    /// (size in Kbytes)
    #[arg(long = "cache-firewall")]
    cache_firewall: Option<usize>,

    /// Perform a randomized offset to the stack frame for each sample.
    /// (size in bytes)
    #[arg(long = "randomize-stack")]
    randomize_stack: Option<usize>,

    /// Delegate control back to the OS before each sample
    #[arg(long = "yield-before-sample")]
    yield_before_sample: Option<bool>,

    /// Filter tests by name (eg. '*/{sorted,unsorted}/[0-9]*')
    #[arg(short = 'f', long = "filter")]
    filter: Option<String>,

    /// Report only statistically significant results
    #[arg(short = 'g', long = "significant-only", default_value_t = false)]
    significant_only: bool,

    /// Enable outlier detection
    #[arg(short = 'o', long = "filter-outliers")]
    filter_outliers: bool,

    /// Perform warmup iterations before taking measurements (1/10 of sample iterations)
    #[arg(long = "warmup")]
    warmup_enabled: Option<bool>,

    #[arg(short = 'p', long = "parallel")]
    parallel: bool,

    /// Quiet mode
    #[arg(short = 'q')]
    quiet: bool,

    #[arg(short = 'v', long = "verbose", default_value_t = false)]
    verbose: bool,

    /// Disables checking proportion of the time spent in a system/kernel mode
    #[arg(long = "no-system-time-check", default_value_t = true, action = ArgAction::SetFalse)]
    system_time_check: bool,
}

#[derive(Parser, Debug)]
struct SoloOpts {
    #[command(flatten)]
    bench_flags: CargoBenchFlags,

    /// seed for the random number generator or omit to use a random seed
    #[arg(long = "seed")]
    seed: Option<u64>,

    /// Number of samples to take for each test
    #[arg(short = 's', long = "samples")]
    samples: Option<NonZeroUsize>,

    /// The strategy to decide the number of iterations to run for each sample (values: flat, linear, random)
    #[arg(long = "sampler")]
    sampler: Option<SampleLengthKind>,

    /// Duration of each sample in seconds
    #[arg(short = 't', long = "time")]
    time: Option<f64>,

    /// Perform a read of a dummy data between samsples to minimize the effect of cache on the performance
    /// (size in Kbytes)
    #[arg(long = "cache-firewall")]
    cache_firewall: Option<usize>,

    /// Perform a randomized offset to the stack frame for each sample.
    /// (size in bytes)
    #[arg(long = "randomize-stack")]
    randomize_stack: Option<usize>,

    /// Delegate control back to the OS before each sample
    #[arg(long = "yield-before-sample")]
    yield_before_sample: Option<bool>,

    /// Filter tests by name (eg. '*/{sorted,unsorted}/[0-9]*')
    #[arg(short = 'f', long = "filter")]
    filter: Option<String>,

    /// Perform warmup iterations before taking measurements (1/10 of sample iterations)
    #[arg(long = "warmup")]
    warmup_enabled: Option<bool>,
}

#[derive(Parser, Debug)]
#[command(author, version, about, long_about = None)]
struct Opts {
    #[command(subcommand)]
    subcommand: Option<BenchmarkMode>,

    #[command(flatten)]
    bench_flags: CargoBenchFlags,

    #[arg(long = "color", default_value = "detect")]
    coloring_mode: String,
}

impl FromStr for SampleLengthKind {
    type Err = Error;

    fn from_str(s: &str) -> StdResult<Self, Self::Err> {
        match s {
            "flat" => Ok(SampleLengthKind::Flat),
            "linear" => Ok(SampleLengthKind::Linear),
            "random" => Ok(SampleLengthKind::Random),
            _ => Err(Error::UnknownSamplerType),
        }
    }
}

/// Definition of the flags required to comply with `cargo bench` calling conventions.
#[derive(Parser, Debug, Clone)]
struct CargoBenchFlags {
    #[arg(long = "bench", default_value_t = true)]
    bench: bool,
}

pub fn run(settings: MeasurementSettings) -> Result<ExitCode> {
    let opts = Opts::parse();

    match Mode::from_str(&opts.coloring_mode) {
        Ok(coloring_mode) => mode::set_coloring_mode(coloring_mode),
        Err(_) => eprintln!("[WARN] Invalid coloring mode: {}", opts.coloring_mode),
    }

    let subcommand = opts.subcommand.unwrap_or(BenchmarkMode::List {
        bench_flags: opts.bench_flags,
    });

    match subcommand {
        BenchmarkMode::List { bench_flags: _ } => {
            let spi = Spi::for_self(SpiModeKind::Synchronous).ok_or(Error::SpiSelfWasMoved)?;
            for func in spi.tests() {
                println!("{}", func.name);
            }
            Ok(ExitCode::SUCCESS)
        }
        BenchmarkMode::Compare(opts) => paired_test::run_test(opts, settings),
        BenchmarkMode::Solo(opts) => solo_test::run_test(opts, settings),
    }
}

// Automatically removes a file when goes out of scope
struct AutoDelete(PathBuf);

impl std::ops::Deref for AutoDelete {
    type Target = PathBuf;

    fn deref(&self) -> &Self::Target {
        &self.0
    }
}

impl Drop for AutoDelete {
    fn drop(&mut self) {
        if let Err(e) = fs::remove_file(&self.0) {
            eprintln!("Failed to delete file {}: {}", self.0.display(), e);
        }
    }
}

fn create_loop_mode(samples: Option<NonZeroUsize>, time: Option<f64>) -> Result<LoopMode> {
    let loop_mode = match (samples, time) {
        (Some(samples), None) => LoopMode::Samples(samples.into()),
        (None, Some(time)) => LoopMode::Time(Duration::from_millis((time * 1000.) as u64)),
        (None, None) => LoopMode::Time(Duration::from_millis(100)),
        (Some(_), Some(_)) => bail!("-t and -s are mutually exclusive"),
    };
    Ok(loop_mode)
}

#[derive(Clone, Copy)]
enum LoopMode {
    Samples(usize),
    Time(Duration),
}

impl LoopMode {
    fn should_continue(&self, iter_no: usize, loop_time: Duration) -> bool {
        match self {
            LoopMode::Samples(samples) => iter_no < *samples,
            LoopMode::Time(duration) => loop_time < *duration,
        }
    }
}

mod solo_test {
    use super::*;
    use crate::{dylib::Spi, CacheFirewall, Summary};
    use alloca::with_alloca;
    use rand::{distributions, rngs::SmallRng, Rng, SeedableRng};
    use std::thread;

    pub(super) fn run_test(opts: SoloOpts, mut settings: MeasurementSettings) -> Result<ExitCode> {
        let SoloOpts {
            bench_flags: _,
            filter,
            samples,
            time,
            seed,
            sampler,
            cache_firewall,
            yield_before_sample,
            warmup_enabled,
            randomize_stack,
        } = opts;

        let mut spi_self = Spi::for_self(SpiModeKind::Synchronous).ok_or(Error::SpiSelfWasMoved)?;

        settings.cache_firewall = cache_firewall;
        settings.randomize_stack = randomize_stack;

        if let Some(warmup_enabled) = warmup_enabled {
            settings.warmup_enabled = warmup_enabled;
        }
        if let Some(yield_before_sample) = yield_before_sample {
            settings.yield_before_sample = yield_before_sample;
        }
        if let Some(sampler) = sampler {
            settings.sampler_type = sampler;
        }

        let filter = filter.as_deref().unwrap_or("");
        let loop_mode = create_loop_mode(samples, time)?;

        let test_names = spi_self
            .tests()
            .iter()
            .map(|t| &t.name)
            .cloned()
            .collect::<Vec<_>>();
        for func_name in test_names {
            if !filter.is_empty() && !glob_match(filter, &func_name) {
                continue;
            }

            let result = run_solo_test(&mut spi_self, &func_name, settings, seed, loop_mode)?;

            reporting::default_reporter_solo(&func_name, &result);
        }

        Ok(ExitCode::SUCCESS)
    }

    fn run_solo_test(
        spi: &mut Spi,
        test_name: &str,
        settings: MeasurementSettings,
        seed: Option<u64>,
        loop_mode: LoopMode,
    ) -> Result<Summary<f64>> {
        const TIME_SLICE_MS: u32 = 10;

        let firewall = settings
            .cache_firewall
            .map(|s| s * 1024)
            .map(CacheFirewall::new);
        let baseline_func = spi.lookup(test_name).ok_or(Error::InvalidTestName)?;

        let mut spi_func = TestedFunction::new(spi, baseline_func.idx);

        let seed = seed.unwrap_or_else(rand::random);

        spi_func.spi.prepare_state(seed)?;
        let iters = spi_func.spi.estimate_iterations(TIME_SLICE_MS)?;
        let mut iterations_per_sample = (iters / 2).max(1);
        let mut sampler = create_sampler(&settings, seed);

        let mut rng = SmallRng::seed_from_u64(seed);
        let stack_offset_distr = settings
            .randomize_stack
            .map(|offset| distributions::Uniform::new(0, offset));

        let mut i = 0;

        let mut sample_iterations = vec![];

        if let LoopMode::Samples(samples) = loop_mode {
            sample_iterations.reserve(samples);
            spi_func.samples.reserve(samples);
        }

        let mut loop_time = Duration::from_secs(0);
        let mut loop_iterations = 0;
        while loop_mode.should_continue(i, loop_time) {
            if loop_time > Duration::from_millis(100) {
                // correcting time slice estimates
                iterations_per_sample =
                    loop_iterations * TIME_SLICE_MS as usize / loop_time.as_millis() as usize;
            }
            let iterations = sampler.next_sample_iterations(i, iterations_per_sample);
            loop_iterations += iterations;
            let warmup_iterations = settings.warmup_enabled.then(|| (iterations / 10).max(1));

            if settings.yield_before_sample {
                thread::yield_now();
            }

            let prepare_state_seed = (i % settings.samples_per_haystack == 0).then_some(seed);

            prepare_func(
                prepare_state_seed,
                &mut spi_func,
                warmup_iterations,
                firewall.as_ref(),
            )?;

            // Allocate a custom stack frame during runtime, to try to offset alignment of the stack.
            if let Some(distr) = stack_offset_distr {
                with_alloca(rng.sample(distr), |_| {
                    spi_func.spi.measure(iterations).unwrap();
                });
            } else {
                spi_func.spi.measure(iterations)?;
            }

            loop_time += Duration::from_nanos(spi_func.read_sample()?);
            sample_iterations.push(iterations);
            i += 1;
        }

        let samples = spi_func
            .samples
            .iter()
            .zip(sample_iterations.iter())
            .map(|(sample, iterations)| *sample as f64 / *iterations as f64)
            .collect::<Vec<_>>();
        Ok(Summary::from(&samples).unwrap())
    }
}

mod paired_test {
    use super::*;
    use crate::{
        calculate_run_result,
        platform::{self, RUsage},
        CacheFirewall, RunResult,
    };
    use alloca::with_alloca;
    use fs::File;
    use rand::{distributions, rngs::SmallRng, Rng, SeedableRng};
    use std::{
        io::{self, BufWriter},
        mem, thread,
    };

    pub(super) fn run_test(
        opts: PairedOpts,
        mut settings: MeasurementSettings,
    ) -> Result<ExitCode> {
        let PairedOpts {
            bench_flags: _,
            path,
            verbose,
            filter,
            samples,
            time,
            filter_outliers,
            path_to_dump,
            gnuplot,
            fail_threshold,
            fail_fast,
            significant_only,
            seed,
            sampler,
            cache_firewall,
            yield_before_sample,
            warmup_enabled,
            parallel,
            quiet,
            randomize_stack,
            system_time_check,
        } = opts;
        let mut path = path
            .or_else(|| args().next().map(PathBuf::from))
            .expect("No path given");
        if path.is_relative() {
            // Resolving paths relative to PWD if given
            if let Ok(pwd) = env::var("PWD") {
                path = PathBuf::from(pwd).join(path)
            }
        };

        #[cfg(target_os = "linux")]
        let path = crate::linux::patch_pie_binary_if_needed(&path)?.unwrap_or(path);

        let mode = if parallel {
            SpiModeKind::Asynchronous
        } else {
            SpiModeKind::Synchronous
        };

        let mut spi_self = Spi::for_self(mode).ok_or(Error::SpiSelfWasMoved)?;
        let mut spi_lib = Spi::for_library(&path, mode).with_context(|| {
            format!(
                "Unable to load benchmark: {}. Make sure it exists and it is valid tango benchmark.",
                path.display()
            )
        })?;

        settings.filter_outliers = filter_outliers;
        settings.cache_firewall = cache_firewall;
        settings.randomize_stack = randomize_stack;

        if let Some(warmup_enabled) = warmup_enabled {
            settings.warmup_enabled = warmup_enabled;
        }
        if let Some(yield_before_sample) = yield_before_sample {
            settings.yield_before_sample = yield_before_sample;
        }
        if let Some(sampler) = sampler {
            settings.sampler_type = sampler;
        }

        let filter = filter.as_deref().unwrap_or("");
        let loop_mode = create_loop_mode(samples, time)?;

        let mut exit_code = ExitCode::SUCCESS;

        if let Some(path) = &path_to_dump {
            if !path.exists() {
                fs::create_dir_all(path)?;
            }
        }
        if gnuplot && path_to_dump.is_none() {
            eprintln!("warn: --gnuplot requires -d to be specified. No plots will be generated")
        }

        let mut sample_dumps = vec![];

        let test_names = spi_self
            .tests()
            .iter()
            .map(|t| &t.name)
            .cloned()
            .collect::<Vec<_>>();
        for func_name in test_names {
            if !filter.is_empty() && !glob_match(filter, &func_name) {
                continue;
            }

            if spi_lib.lookup(&func_name).is_none() {
                if !quiet {
                    writeln!(stderr(), "{} skipped...", &func_name)?;
                }
                continue;
            }

            let rusage_before = system_time_check.then(platform::rusage);
            let (result, sample_dump) = run_paired_test(
                &mut spi_lib,
                &mut spi_self,
                &func_name,
                settings,
                seed,
                loop_mode,
                path_to_dump.as_ref(),
            )?;
            if let Some(usage_before) = rusage_before {
                let rusage = platform::rusage() - usage_before;
                if detect_system_time_bias(&rusage) {
                    reporting::report_system_time_bias(&result, &rusage);
                }
            }

            if let Some(dump) = sample_dump {
                sample_dumps.push(dump);
            }

            if result.diff_estimate.significant || !significant_only {
                if verbose {
                    reporting::verbose_reporter(&result);
                } else {
                    reporting::default_reporter(&result);
                }
            }

            if result.diff_estimate.significant {
                if let Some(threshold) = fail_threshold {
                    if result.diff_estimate.pct >= threshold {
                        eprintln!(
                            "[ERROR] Performance regressed {:+.1}% >= {:.1}%  -  test: {}",
                            result.diff_estimate.pct, threshold, func_name
                        );
                        if fail_fast {
                            return Ok(ExitCode::FAILURE);
                        } else {
                            exit_code = ExitCode::FAILURE;
                        }
                    }
                }
            }
        }

        if gnuplot && !sample_dumps.is_empty() {
            generate_plots(sample_dumps.as_slice())?;
        }

        Ok(exit_code)
    }

    /// Checking if test spent too much time in a system/kernel mode.
    ///
    /// OS doesn't provide fairness guarantees, this can influence result
    fn detect_system_time_bias(rusage: &RUsage) -> bool {
        // system time is at least 5% of CPU time overall
        let system = rusage.system_time.as_secs_f64();
        let overall = (rusage.user_time + rusage.system_time).as_secs_f64();
        system / overall > 0.05
    }

    /// Measure the difference in performance of two functions
    ///
    /// Provides a way to save a raw dump of measurements into directory
    ///
    /// The format is as follows
    /// ```txt
    /// b_1,c_1
    /// b_2,c_2
    /// ...
    /// b_n,c_n
    /// ```
    /// where `b_1..b_n` are baseline absolute time (in nanoseconds) measurements
    /// and `c_1..c_n` are candidate time measurements
    ///
    /// Returns a statistical results of a test run and path to raw samples of sample dump was requested
    fn run_paired_test(
        baseline: &mut Spi,
        candidate: &mut Spi,
        test_name: &str,
        settings: MeasurementSettings,
        seed: Option<u64>,
        loop_mode: LoopMode,
        samples_dump_path: Option<&PathBuf>,
    ) -> Result<(RunResult, Option<PathBuf>)> {
        const TIME_SLICE_MS: u32 = 10;

        let firewall = settings
            .cache_firewall
            .map(|s| s * 1024)
            .map(CacheFirewall::new);
        let baseline_func = baseline.lookup(test_name).ok_or(Error::InvalidTestName)?;
        let candidate_func = candidate.lookup(test_name).ok_or(Error::InvalidTestName)?;

        let mut baseline = TestedFunction::new(baseline, baseline_func.idx);
        let mut candidate = TestedFunction::new(candidate, candidate_func.idx);

        let mut a_func = &mut baseline;
        let mut b_func = &mut candidate;

        let seed = seed.unwrap_or_else(rand::random);

        a_func
            .spi
            .prepare_state(seed)
            .context("Unable to prepare benchmark state")?;
        let a_iters = a_func
            .spi
            .estimate_iterations(TIME_SLICE_MS)
            .context("Failed to estimate required iterations number")?;
        let a_estimate = (a_iters / 2).max(1);

        b_func
            .spi
            .prepare_state(seed)
            .context("Unable to prepare benchmark state")?;
        let b_iters = b_func
            .spi
            .estimate_iterations(TIME_SLICE_MS)
            .context("Failed to estimate required iterations number")?;
        let b_estimate = (b_iters / 2).max(1);

        let mut iterations_per_sample = a_estimate.min(b_estimate);
        let mut sampler = create_sampler(&settings, seed);

        let mut rng = SmallRng::seed_from_u64(seed);
        let stack_offset_distr = settings
            .randomize_stack
            .map(|offset| distributions::Uniform::new(0, offset));

        let mut i = 0;
        let mut switch_counter = 0;

        let mut sample_iterations = vec![];

        if let LoopMode::Samples(samples) = loop_mode {
            sample_iterations.reserve(samples);
            a_func.samples.reserve(samples);
            b_func.samples.reserve(samples);
        }

        let mut loop_time = Duration::from_secs(0);
        let mut loop_iterations = 0;
        while loop_mode.should_continue(i, loop_time) {
            if loop_time > Duration::from_millis(100) {
                // correcting time slice estimates
                iterations_per_sample =
                    loop_iterations * TIME_SLICE_MS as usize / loop_time.as_millis() as usize;
            }
            let iterations = sampler.next_sample_iterations(i, iterations_per_sample);
            loop_iterations += iterations;
            let warmup_iterations = settings.warmup_enabled.then(|| (iterations / 10).max(1));

            // !!! IMPORTANT !!!
            // Algorithms should be called in different order on each new iteration.
            // This equalize the probability of facing unfortunate circumstances like cache misses or page faults
            // for both functions. Although both algorithms are from distinct shared objects and therefore
            // must be fully self-contained in terms of virtual address space (each shared object has its own
            // generator instances, static variables, memory mappings, etc.) it might be the case that
            // on the level of physical memory both of them rely on the same memory-mapped test data, for example.
            // In that case first function will experience the larger amount of major page faults.
            {
                mem::swap(&mut a_func, &mut b_func);
                switch_counter += 1;
            }

            if settings.yield_before_sample {
                thread::yield_now();
            }

            let prepare_state_seed = (i % settings.samples_per_haystack == 0).then_some(seed);
            let mut sample_time = 0;

            prepare_func(
                prepare_state_seed,
                a_func,
                warmup_iterations,
                firewall.as_ref(),
            )?;
            prepare_func(
                prepare_state_seed,
                b_func,
                warmup_iterations,
                firewall.as_ref(),
            )?;

            // Allocate a custom stack frame during runtime, to try to offset alignment of the stack.
            if let Some(distr) = stack_offset_distr {
                with_alloca(rng.sample(distr), |_| {
                    a_func.spi.measure(iterations).unwrap();
                    b_func.spi.measure(iterations).unwrap();
                });
            } else {
                a_func.spi.measure(iterations)?;
                b_func.spi.measure(iterations)?;
            }

            let a_sample_time = a_func.read_sample()?;
            let b_sample_time = b_func.read_sample()?;
            sample_time += a_sample_time.max(b_sample_time);

            loop_time += Duration::from_nanos(sample_time);
            sample_iterations.push(iterations);
            i += 1;
        }

        // If we switched functions odd number of times then we need to swap them back so that
        // the first function is always the baseline.
        if switch_counter % 2 != 0 {
            mem::swap(&mut a_func, &mut b_func);
        }

        let run_result = calculate_run_result(
            test_name,
            &a_func.samples,
            &b_func.samples,
            &sample_iterations,
            settings.filter_outliers,
        )
        .ok_or(Error::NoMeasurements)?;

        let samples_path = if let Some(path) = samples_dump_path {
            let file_path = write_samples(path, test_name, a_func, b_func, sample_iterations)?;
            Some(file_path)
        } else {
            None
        };

        Ok((run_result, samples_path))
    }

    fn write_samples(
        path: &Path,
        test_name: &str,
        a_func: &TestedFunction,
        b_func: &TestedFunction,
        iterations: Vec<usize>,
    ) -> Result<PathBuf> {
        let file_name = format!("{}.csv", test_name.replace('/', "-"));
        let file_path = path.join(file_name);
        let s_samples = a_func.samples.iter().copied();
        let b_samples = b_func.samples.iter().copied();
        let values = s_samples
            .zip(b_samples)
            .zip(iterations.iter().copied())
            .map(|((a, b), c)| (a, b, c));
        write_csv(&file_path, values).context("Unable to write raw measurements")?;
        Ok(file_path)
    }

    fn write_csv<A: Display, B: Display, C: Display>(
        path: impl AsRef<Path>,
        values: impl IntoIterator<Item = (A, B, C)>,
    ) -> io::Result<()> {
        let mut file = BufWriter::new(File::create(path)?);
        for (a, b, c) in values {
            writeln!(&mut file, "{},{},{}", a, b, c)?;
        }
        Ok(())
    }

    fn generate_plots(sample_dumps: &[PathBuf]) -> Result<()> {
        let gnuplot_file = AutoDelete(temp_dir().join("tango-plot.gnuplot"));
        fs::write(&*gnuplot_file, include_bytes!("plot.gnuplot"))?;
        let gnuplot_file_str = gnuplot_file.to_str().unwrap();

        for input in sample_dumps {
            let csv_input = input.to_str().unwrap();
            let svg_path = input.with_extension("svg");
            let cmd = Command::new("gnuplot")
                .args([
                    "-c",
                    gnuplot_file_str,
                    csv_input,
                    svg_path.to_str().unwrap(),
                ])
                .stdin(Stdio::null())
                .stdout(Stdio::inherit())
                .stderr(Stdio::inherit())
                .status()
                .context("Failed to execute gnuplot")?;

            if !cmd.success() {
                bail!("gnuplot command failed");
            }
        }
        Ok(())
    }
}

mod reporting {
    use crate::{
        cli::{colorize, HumanTime},
        platform::RUsage,
        RunResult, Summary,
    };
    use colorz::{ansi, mode::Stream, Colorize, Style};

    pub(super) fn verbose_reporter(results: &RunResult) {
        let base = results.baseline;
        let candidate = results.candidate;

        let significant = results.diff_estimate.significant;

        println!(
            "{}  (n: {}, outliers: {})",
            results.name.bold().stream(Stream::Stdout),
            results.diff.n,
            results.outliers
        );

        println!(
            "    {:12}   {:>15} {:>15} {:>15}",
            "",
            "baseline".bold().stream(Stream::Stdout),
            "candidate".bold().stream(Stream::Stdout),
            "∆".bold().stream(Stream::Stdout),
        );
        println!(
            "    {:12} ╭────────────────────────────────────────────────",
            ""
        );
        println!(
            "    {:12} │ {:>15} {:>15} {:>15}  {:+4.2}{}{}",
            "mean",
            HumanTime(base.mean),
            HumanTime(candidate.mean),
            colorize(
                HumanTime(results.diff.mean),
                significant,
                results.diff.mean < 0.
            ),
            colorize(
                results.diff_estimate.pct,
                significant,
                results.diff.mean < 0.
            ),
            colorize("%", significant, results.diff.mean < 0.),
            if significant { "*" } else { "" },
        );
        println!(
            "    {:12} │ {:>15} {:>15} {:>15}",
            "min",
            HumanTime(base.min),
            HumanTime(candidate.min),
            HumanTime(candidate.min - base.min)
        );
        println!(
            "    {:12} │ {:>15} {:>15} {:>15}",
            "max",
            HumanTime(base.max),
            HumanTime(candidate.max),
            HumanTime(candidate.max - base.max),
        );
        println!(
            "    {:12} │ {:>15} {:>15} {:>15}",
            "std. dev.",
            HumanTime(base.variance.sqrt()),
            HumanTime(candidate.variance.sqrt()),
            HumanTime(results.diff.variance.sqrt()),
        );
        println!();
    }

    pub(super) fn default_reporter(results: &RunResult) {
        let base = results.baseline;
        let candidate = results.candidate;
        let diff = results.diff;

        let significant = results.diff_estimate.significant;

        let speedup = results.diff_estimate.pct;
        let candidate_faster = diff.mean < 0.;
        println!(
            "{:50} [ {:>8} ... {:>8} ]    {:>+7.2}{}{}",
            colorize(&results.name, significant, candidate_faster),
            HumanTime(base.mean),
            colorize(HumanTime(candidate.mean), significant, candidate_faster),
            colorize(speedup, significant, candidate_faster),
            colorize("%", significant, candidate_faster),
            if significant { "*" } else { "" },
        )
    }

    pub(super) fn default_reporter_solo(name: &str, results: &Summary<f64>) {
        println!(
            "{:50}  [ {:>8} ... {:>8} ... {:>8} ]  stddev: {:>8}",
            name,
            HumanTime(results.min),
            HumanTime(results.mean),
            HumanTime(results.max),
            HumanTime(results.variance.sqrt()),
        )
    }

    pub(super) fn report_system_time_bias(result: &RunResult, rusage: &RUsage) {
        const RED: Style = Style::new().fg(ansi::Red).const_into_runtime_style();

        eprintln!(
            "{}: {} benchmark spent too much time in system mode (sys: {:?}, usr: {:?}). Results may be inaccurate",
            "WARN".into_style_with(RED).stream(Stream::Stderr),
            &result.name,
            rusage.system_time,
            rusage.user_time
        );
    }
}

struct TestedFunction<'a> {
    pub(crate) spi: &'a mut Spi,
    pub(crate) samples: Vec<u64>,
}

impl<'a> TestedFunction<'a> {
    pub(crate) fn new(spi: &'a mut Spi, func: FunctionIdx) -> Self {
        spi.select(func);
        TestedFunction {
            spi,
            samples: Vec::new(),
        }
    }

    pub(crate) fn read_sample(&mut self) -> Result<u64> {
        let sample = self.spi.read_sample().context("Unable to read sample")?;
        self.samples.push(sample);
        Ok(sample)
    }

    pub(crate) fn run(&mut self, iterations: usize) -> Result<u64> {
        self.spi
            .run(iterations)
            .context("Unable to run measurement")
    }
}

fn prepare_func(
    prepare_state_seed: Option<u64>,
    f: &mut TestedFunction,
    warmup_iterations: Option<usize>,
    firewall: Option<&CacheFirewall>,
) -> Result<()> {
    if let Some(seed) = prepare_state_seed {
        f.spi.prepare_state(seed)?;
        if let Some(firewall) = firewall {
            firewall.issue_read();
        }
    }
    if let Some(warmup_iterations) = warmup_iterations {
        f.run(warmup_iterations)?;
    }
    Ok(())
}

fn create_sampler(settings: &MeasurementSettings, seed: u64) -> Box<dyn SampleLength> {
    match settings.sampler_type {
        SampleLengthKind::Flat => Box::new(FlatSampleLength::new(settings)),
        SampleLengthKind::Linear => Box::new(LinearSampleLength::new(settings)),
        SampleLengthKind::Random => Box::new(RandomSampleLength::new(settings, seed)),
    }
}

fn colorize<T: Display>(value: T, do_paint: bool, is_improved: bool) -> impl Display {
    use colorz::{ansi, mode::Stream::Stdout, Colorize, Style};

    const RED: Style = Style::new().fg(ansi::Red).const_into_runtime_style();
    const GREEN: Style = Style::new().fg(ansi::Green).const_into_runtime_style();
    const DEFAULT: Style = Style::new().const_into_runtime_style();

    if do_paint {
        if is_improved {
            value.into_style_with(GREEN).stream(Stdout)
        } else {
            value.into_style_with(RED).stream(Stdout)
        }
    } else {
        value.into_style_with(DEFAULT).stream(Stdout)
    }
}

struct HumanTime(f64);

impl fmt::Display for HumanTime {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        const USEC: f64 = 1_000.;
        const MSEC: f64 = USEC * 1_000.;
        const SEC: f64 = MSEC * 1_000.;

        if self.0.abs() > SEC {
            f.pad(&format!("{:.1} s", self.0 / SEC))
        } else if self.0.abs() > MSEC {
            f.pad(&format!("{:.1} ms", self.0 / MSEC))
        } else if self.0.abs() > USEC {
            f.pad(&format!("{:.1} us", self.0 / USEC))
        } else if self.0 == 0. {
            f.pad("0 ns")
        } else {
            f.pad(&format!("{:.1} ns", self.0))
        }
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn check_human_time() {
        assert_eq!(format!("{}", HumanTime(0.1)), "0.1 ns");
        assert_eq!(format!("{:>5}", HumanTime(0.)), " 0 ns");

        assert_eq!(format!("{}", HumanTime(120.)), "120.0 ns");

        assert_eq!(format!("{}", HumanTime(1200.)), "1.2 us");

        assert_eq!(format!("{}", HumanTime(1200000.)), "1.2 ms");

        assert_eq!(format!("{}", HumanTime(1200000000.)), "1.2 s");

        assert_eq!(format!("{}", HumanTime(-1200000.)), "-1.2 ms");
    }

    // Sane checking some simple patterns
    #[test]
    fn check_glob() {
        let patterns = vec!["a/*/*", "a/**", "*/32/*", "**/b", "a/{32,64}/*"];
        let input = "a/32/b";
        for pattern in patterns {
            assert!(
                glob_match(pattern, input),
                "failed to match {} against {}",
                pattern,
                input
            );
        }
    }
}