Home | History | Annotate | Download | only in src
      1 // Copyright 2016 Ismael Jimenez Martinez. All rights reserved.
      2 // Copyright 2017 Roman Lebedev. All rights reserved.
      3 //
      4 // Licensed under the Apache License, Version 2.0 (the "License");
      5 // you may not use this file except in compliance with the License.
      6 // You may obtain a copy of the License at
      7 //
      8 //     http://www.apache.org/licenses/LICENSE-2.0
      9 //
     10 // Unless required by applicable law or agreed to in writing, software
     11 // distributed under the License is distributed on an "AS IS" BASIS,
     12 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
     13 // See the License for the specific language governing permissions and
     14 // limitations under the License.
     15 
     16 #include "benchmark/benchmark.h"
     17 
     18 #include <algorithm>
     19 #include <cmath>
     20 #include <string>
     21 #include <vector>
     22 #include <numeric>
     23 #include "check.h"
     24 #include "statistics.h"
     25 
     26 namespace benchmark {
     27 
     28 auto StatisticsSum = [](const std::vector<double>& v) {
     29   return std::accumulate(v.begin(), v.end(), 0.0);
     30 };
     31 
     32 double StatisticsMean(const std::vector<double>& v) {
     33   if (v.size() == 0) return 0.0;
     34   return StatisticsSum(v) * (1.0 / v.size());
     35 }
     36 
     37 double StatisticsMedian(const std::vector<double>& v) {
     38   if (v.size() < 3) return StatisticsMean(v);
     39   std::vector<double> partial;
     40   // we need roundDown(count/2)+1 slots
     41   partial.resize(1 + (v.size() / 2));
     42   std::partial_sort_copy(v.begin(), v.end(), partial.begin(), partial.end());
     43   // did we have odd number of samples?
     44   // if yes, then the last element of partially-sorted vector is the median
     45   // it no, then the average of the last two elements is the median
     46   if(v.size() % 2 == 1)
     47     return partial.back();
     48   return (partial[partial.size() - 2] + partial[partial.size() - 1]) / 2.0;
     49 }
     50 
     51 // Return the sum of the squares of this sample set
     52 auto SumSquares = [](const std::vector<double>& v) {
     53   return std::inner_product(v.begin(), v.end(), v.begin(), 0.0);
     54 };
     55 
     56 auto Sqr = [](const double dat) { return dat * dat; };
     57 auto Sqrt = [](const double dat) {
     58   // Avoid NaN due to imprecision in the calculations
     59   if (dat < 0.0) return 0.0;
     60   return std::sqrt(dat);
     61 };
     62 
     63 double StatisticsStdDev(const std::vector<double>& v) {
     64   const auto mean = StatisticsMean(v);
     65   if (v.size() == 0) return mean;
     66 
     67   // Sample standard deviation is undefined for n = 1
     68   if (v.size() == 1)
     69     return 0.0;
     70 
     71   const double avg_squares = SumSquares(v) * (1.0 / v.size());
     72   return Sqrt(v.size() / (v.size() - 1.0) * (avg_squares - Sqr(mean)));
     73 }
     74 
     75 std::vector<BenchmarkReporter::Run> ComputeStats(
     76     const std::vector<BenchmarkReporter::Run>& reports) {
     77   typedef BenchmarkReporter::Run Run;
     78   std::vector<Run> results;
     79 
     80   auto error_count =
     81       std::count_if(reports.begin(), reports.end(),
     82                     [](Run const& run) { return run.error_occurred; });
     83 
     84   if (reports.size() - error_count < 2) {
     85     // We don't report aggregated data if there was a single run.
     86     return results;
     87   }
     88 
     89   // Accumulators.
     90   std::vector<double> real_accumulated_time_stat;
     91   std::vector<double> cpu_accumulated_time_stat;
     92   std::vector<double> bytes_per_second_stat;
     93   std::vector<double> items_per_second_stat;
     94 
     95   real_accumulated_time_stat.reserve(reports.size());
     96   cpu_accumulated_time_stat.reserve(reports.size());
     97   bytes_per_second_stat.reserve(reports.size());
     98   items_per_second_stat.reserve(reports.size());
     99 
    100   // All repetitions should be run with the same number of iterations so we
    101   // can take this information from the first benchmark.
    102   int64_t const run_iterations = reports.front().iterations;
    103   // create stats for user counters
    104   struct CounterStat {
    105     Counter c;
    106     std::vector<double> s;
    107   };
    108   std::map< std::string, CounterStat > counter_stats;
    109   for(Run const& r : reports) {
    110     for(auto const& cnt : r.counters) {
    111       auto it = counter_stats.find(cnt.first);
    112       if(it == counter_stats.end()) {
    113         counter_stats.insert({cnt.first, {cnt.second, std::vector<double>{}}});
    114         it = counter_stats.find(cnt.first);
    115         it->second.s.reserve(reports.size());
    116       } else {
    117         CHECK_EQ(counter_stats[cnt.first].c.flags, cnt.second.flags);
    118       }
    119     }
    120   }
    121 
    122   // Populate the accumulators.
    123   for (Run const& run : reports) {
    124     CHECK_EQ(reports[0].benchmark_name, run.benchmark_name);
    125     CHECK_EQ(run_iterations, run.iterations);
    126     if (run.error_occurred) continue;
    127     real_accumulated_time_stat.emplace_back(run.real_accumulated_time);
    128     cpu_accumulated_time_stat.emplace_back(run.cpu_accumulated_time);
    129     items_per_second_stat.emplace_back(run.items_per_second);
    130     bytes_per_second_stat.emplace_back(run.bytes_per_second);
    131     // user counters
    132     for(auto const& cnt : run.counters) {
    133       auto it = counter_stats.find(cnt.first);
    134       CHECK_NE(it, counter_stats.end());
    135       it->second.s.emplace_back(cnt.second);
    136     }
    137   }
    138 
    139   // Only add label if it is same for all runs
    140   std::string report_label = reports[0].report_label;
    141   for (std::size_t i = 1; i < reports.size(); i++) {
    142     if (reports[i].report_label != report_label) {
    143       report_label = "";
    144       break;
    145     }
    146   }
    147 
    148   for(const auto& Stat : *reports[0].statistics) {
    149     // Get the data from the accumulator to BenchmarkReporter::Run's.
    150     Run data;
    151     data.benchmark_name = reports[0].benchmark_name + "_" + Stat.name_;
    152     data.report_label = report_label;
    153     data.iterations = run_iterations;
    154 
    155     data.real_accumulated_time = Stat.compute_(real_accumulated_time_stat);
    156     data.cpu_accumulated_time = Stat.compute_(cpu_accumulated_time_stat);
    157     data.bytes_per_second = Stat.compute_(bytes_per_second_stat);
    158     data.items_per_second = Stat.compute_(items_per_second_stat);
    159 
    160     data.time_unit = reports[0].time_unit;
    161 
    162     // user counters
    163     for(auto const& kv : counter_stats) {
    164       const auto uc_stat = Stat.compute_(kv.second.s);
    165       auto c = Counter(uc_stat, counter_stats[kv.first].c.flags);
    166       data.counters[kv.first] = c;
    167     }
    168 
    169     results.push_back(data);
    170   }
    171 
    172   return results;
    173 }
    174 
    175 }  // end namespace benchmark
    176