1 use crate::benchmark::BenchmarkConfig; 2 use crate::connection::OutgoingMessage; 3 use crate::measurement::Measurement; 4 use crate::report::{BenchmarkId, Report, ReportContext}; 5 use crate::{ActualSamplingMode, Bencher, Criterion, DurationExt}; 6 use std::marker::PhantomData; 7 use std::time::Duration; 8 9 /// PRIVATE 10 pub(crate) trait Routine<M: Measurement, T: ?Sized> { 11 /// PRIVATE bench(&mut self, m: &M, iters: &[u64], parameter: &T) -> Vec<f64>12 fn bench(&mut self, m: &M, iters: &[u64], parameter: &T) -> Vec<f64>; 13 /// PRIVATE warm_up(&mut self, m: &M, how_long: Duration, parameter: &T) -> (u64, u64)14 fn warm_up(&mut self, m: &M, how_long: Duration, parameter: &T) -> (u64, u64); 15 16 /// PRIVATE test(&mut self, m: &M, parameter: &T)17 fn test(&mut self, m: &M, parameter: &T) { 18 self.bench(m, &[1u64], parameter); 19 } 20 21 /// Iterates the benchmarked function for a fixed length of time, but takes no measurements. 22 /// This keeps the overall benchmark suite runtime constant-ish even when running under a 23 /// profiler with an unknown amount of overhead. Since no measurements are taken, it also 24 /// reduces the amount of time the execution spends in Criterion.rs code, which should help 25 /// show the performance of the benchmarked code more clearly as well. profile( &mut self, measurement: &M, id: &BenchmarkId, criterion: &Criterion<M>, report_context: &ReportContext, time: Duration, parameter: &T, )26 fn profile( 27 &mut self, 28 measurement: &M, 29 id: &BenchmarkId, 30 criterion: &Criterion<M>, 31 report_context: &ReportContext, 32 time: Duration, 33 parameter: &T, 34 ) { 35 criterion 36 .report 37 .profile(id, report_context, time.to_nanos() as f64); 38 39 let mut profile_path = report_context.output_directory.clone(); 40 if (*crate::CARGO_CRITERION_CONNECTION).is_some() { 41 // If connected to cargo-criterion, generate a cargo-criterion-style path. 42 // This is kind of a hack. 43 profile_path.push("profile"); 44 profile_path.push(id.as_directory_name()); 45 } else { 46 profile_path.push(id.as_directory_name()); 47 profile_path.push("profile"); 48 } 49 criterion 50 .profiler 51 .borrow_mut() 52 .start_profiling(id.id(), &profile_path); 53 54 let time = time.to_nanos(); 55 56 // TODO: Some profilers will show the two batches of iterations as 57 // being different code-paths even though they aren't really. 58 59 // Get the warmup time for one second 60 let (wu_elapsed, wu_iters) = self.warm_up(measurement, Duration::from_secs(1), parameter); 61 if wu_elapsed < time { 62 // Initial guess for the mean execution time 63 let met = wu_elapsed as f64 / wu_iters as f64; 64 65 // Guess how many iterations will be required for the remaining time 66 let remaining = (time - wu_elapsed) as f64; 67 68 let iters = remaining / met; 69 let iters = iters as u64; 70 71 self.bench(measurement, &[iters], parameter); 72 } 73 74 criterion 75 .profiler 76 .borrow_mut() 77 .stop_profiling(id.id(), &profile_path); 78 79 criterion.report.terminated(id, report_context); 80 } 81 sample( &mut self, measurement: &M, id: &BenchmarkId, config: &BenchmarkConfig, criterion: &Criterion<M>, report_context: &ReportContext, parameter: &T, ) -> (ActualSamplingMode, Box<[f64]>, Box<[f64]>)82 fn sample( 83 &mut self, 84 measurement: &M, 85 id: &BenchmarkId, 86 config: &BenchmarkConfig, 87 criterion: &Criterion<M>, 88 report_context: &ReportContext, 89 parameter: &T, 90 ) -> (ActualSamplingMode, Box<[f64]>, Box<[f64]>) { 91 let wu = config.warm_up_time; 92 let m_ns = config.measurement_time.to_nanos(); 93 94 criterion 95 .report 96 .warmup(id, report_context, wu.to_nanos() as f64); 97 98 if let Some(conn) = &criterion.connection { 99 conn.send(&OutgoingMessage::Warmup { 100 id: id.into(), 101 nanos: wu.to_nanos() as f64, 102 }) 103 .unwrap(); 104 } 105 106 let (wu_elapsed, wu_iters) = self.warm_up(measurement, wu, parameter); 107 if crate::debug_enabled() { 108 println!( 109 "\nCompleted {} iterations in {} nanoseconds, estimated execution time is {} ns", 110 wu_iters, 111 wu_elapsed, 112 wu_elapsed as f64 / wu_iters as f64 113 ); 114 } 115 116 // Initial guess for the mean execution time 117 let met = wu_elapsed as f64 / wu_iters as f64; 118 119 let n = config.sample_size as u64; 120 121 let actual_sampling_mode = config 122 .sampling_mode 123 .choose_sampling_mode(met, n, m_ns as f64); 124 125 let m_iters = actual_sampling_mode.iteration_counts(met, n, &config.measurement_time); 126 127 let expected_ns = m_iters 128 .iter() 129 .copied() 130 .map(|count| count as f64 * met) 131 .sum(); 132 133 // Use saturating_add to handle overflow. 134 let mut total_iters = 0u64; 135 for count in m_iters.iter().copied() { 136 total_iters = total_iters.saturating_add(count); 137 } 138 139 criterion 140 .report 141 .measurement_start(id, report_context, n, expected_ns, total_iters); 142 143 if let Some(conn) = &criterion.connection { 144 conn.send(&OutgoingMessage::MeasurementStart { 145 id: id.into(), 146 sample_count: n, 147 estimate_ns: expected_ns, 148 iter_count: total_iters, 149 }) 150 .unwrap(); 151 } 152 153 let m_elapsed = self.bench(measurement, &m_iters, parameter); 154 155 let m_iters_f: Vec<f64> = m_iters.iter().map(|&x| x as f64).collect(); 156 157 ( 158 actual_sampling_mode, 159 m_iters_f.into_boxed_slice(), 160 m_elapsed.into_boxed_slice(), 161 ) 162 } 163 } 164 165 pub struct Function<M: Measurement, F, T> 166 where 167 F: FnMut(&mut Bencher<'_, M>, &T), 168 T: ?Sized, 169 { 170 f: F, 171 // TODO: Is there some way to remove these? 172 _phantom: PhantomData<T>, 173 _phamtom2: PhantomData<M>, 174 } 175 impl<M: Measurement, F, T> Function<M, F, T> 176 where 177 F: FnMut(&mut Bencher<'_, M>, &T), 178 T: ?Sized, 179 { new(f: F) -> Function<M, F, T>180 pub fn new(f: F) -> Function<M, F, T> { 181 Function { 182 f, 183 _phantom: PhantomData, 184 _phamtom2: PhantomData, 185 } 186 } 187 } 188 189 impl<M: Measurement, F, T> Routine<M, T> for Function<M, F, T> 190 where 191 F: FnMut(&mut Bencher<'_, M>, &T), 192 T: ?Sized, 193 { bench(&mut self, m: &M, iters: &[u64], parameter: &T) -> Vec<f64>194 fn bench(&mut self, m: &M, iters: &[u64], parameter: &T) -> Vec<f64> { 195 let f = &mut self.f; 196 197 let mut b = Bencher { 198 iterated: false, 199 iters: 0, 200 value: m.zero(), 201 measurement: m, 202 elapsed_time: Duration::from_millis(0), 203 }; 204 205 iters 206 .iter() 207 .map(|iters| { 208 b.iters = *iters; 209 (*f)(&mut b, parameter); 210 b.assert_iterated(); 211 m.to_f64(&b.value) 212 }) 213 .collect() 214 } 215 warm_up(&mut self, m: &M, how_long: Duration, parameter: &T) -> (u64, u64)216 fn warm_up(&mut self, m: &M, how_long: Duration, parameter: &T) -> (u64, u64) { 217 let f = &mut self.f; 218 let mut b = Bencher { 219 iterated: false, 220 iters: 1, 221 value: m.zero(), 222 measurement: m, 223 elapsed_time: Duration::from_millis(0), 224 }; 225 226 let mut total_iters = 0; 227 let mut elapsed_time = Duration::from_millis(0); 228 loop { 229 (*f)(&mut b, parameter); 230 231 b.assert_iterated(); 232 233 total_iters += b.iters; 234 elapsed_time += b.elapsed_time; 235 if elapsed_time > how_long { 236 return (elapsed_time.to_nanos(), total_iters); 237 } 238 239 b.iters = b.iters.wrapping_mul(2); 240 } 241 } 242 } 243