libcxx

libcxx mirror with random patches
git clone https://git.neptards.moe/neptards/libcxx.git
Log | Files | Refs

benchmark_register.cc (14497B)


      1 // Copyright 2015 Google Inc. All rights reserved.
      2 //
      3 // Licensed under the Apache License, Version 2.0 (the "License");
      4 // you may not use this file except in compliance with the License.
      5 // You may obtain a copy of the License at
      6 //
      7 //     http://www.apache.org/licenses/LICENSE-2.0
      8 //
      9 // Unless required by applicable law or agreed to in writing, software
     10 // distributed under the License is distributed on an "AS IS" BASIS,
     11 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
     12 // See the License for the specific language governing permissions and
     13 // limitations under the License.
     14 
     15 #include "benchmark_register.h"
     16 
     17 #ifndef BENCHMARK_OS_WINDOWS
     18 #ifndef BENCHMARK_OS_FUCHSIA
     19 #include <sys/resource.h>
     20 #endif
     21 #include <sys/time.h>
     22 #include <unistd.h>
     23 #endif
     24 
     25 #include <algorithm>
     26 #include <atomic>
     27 #include <condition_variable>
     28 #include <cstdio>
     29 #include <cstdlib>
     30 #include <cstring>
     31 #include <fstream>
     32 #include <iostream>
     33 #include <memory>
     34 #include <sstream>
     35 #include <thread>
     36 
     37 #include "benchmark/benchmark.h"
     38 #include "benchmark_api_internal.h"
     39 #include "check.h"
     40 #include "commandlineflags.h"
     41 #include "complexity.h"
     42 #include "internal_macros.h"
     43 #include "log.h"
     44 #include "mutex.h"
     45 #include "re.h"
     46 #include "statistics.h"
     47 #include "string_util.h"
     48 #include "timers.h"
     49 
     50 namespace benchmark {
     51 
     52 namespace {
     53 // For non-dense Range, intermediate values are powers of kRangeMultiplier.
     54 static const int kRangeMultiplier = 8;
     55 // The size of a benchmark family determines is the number of inputs to repeat
     56 // the benchmark on. If this is "large" then warn the user during configuration.
     57 static const size_t kMaxFamilySize = 100;
     58 }  // end namespace
     59 
     60 namespace internal {
     61 
     62 //=============================================================================//
     63 //                         BenchmarkFamilies
     64 //=============================================================================//
     65 
     66 // Class for managing registered benchmarks.  Note that each registered
     67 // benchmark identifies a family of related benchmarks to run.
     68 class BenchmarkFamilies {
     69  public:
     70   static BenchmarkFamilies* GetInstance();
     71 
     72   // Registers a benchmark family and returns the index assigned to it.
     73   size_t AddBenchmark(std::unique_ptr<Benchmark> family);
     74 
     75   // Clear all registered benchmark families.
     76   void ClearBenchmarks();
     77 
     78   // Extract the list of benchmark instances that match the specified
     79   // regular expression.
     80   bool FindBenchmarks(std::string re,
     81                       std::vector<BenchmarkInstance>* benchmarks,
     82                       std::ostream* Err);
     83 
     84  private:
     85   BenchmarkFamilies() {}
     86 
     87   std::vector<std::unique_ptr<Benchmark>> families_;
     88   Mutex mutex_;
     89 };
     90 
     91 BenchmarkFamilies* BenchmarkFamilies::GetInstance() {
     92   static BenchmarkFamilies instance;
     93   return &instance;
     94 }
     95 
     96 size_t BenchmarkFamilies::AddBenchmark(std::unique_ptr<Benchmark> family) {
     97   MutexLock l(mutex_);
     98   size_t index = families_.size();
     99   families_.push_back(std::move(family));
    100   return index;
    101 }
    102 
    103 void BenchmarkFamilies::ClearBenchmarks() {
    104   MutexLock l(mutex_);
    105   families_.clear();
    106   families_.shrink_to_fit();
    107 }
    108 
    109 bool BenchmarkFamilies::FindBenchmarks(
    110     std::string spec, std::vector<BenchmarkInstance>* benchmarks,
    111     std::ostream* ErrStream) {
    112   CHECK(ErrStream);
    113   auto& Err = *ErrStream;
    114   // Make regular expression out of command-line flag
    115   std::string error_msg;
    116   Regex re;
    117   bool isNegativeFilter = false;
    118   if (spec[0] == '-') {
    119     spec.replace(0, 1, "");
    120     isNegativeFilter = true;
    121   }
    122   if (!re.Init(spec, &error_msg)) {
    123     Err << "Could not compile benchmark re: " << error_msg << std::endl;
    124     return false;
    125   }
    126 
    127   // Special list of thread counts to use when none are specified
    128   const std::vector<int> one_thread = {1};
    129 
    130   MutexLock l(mutex_);
    131   for (std::unique_ptr<Benchmark>& family : families_) {
    132     // Family was deleted or benchmark doesn't match
    133     if (!family) continue;
    134 
    135     if (family->ArgsCnt() == -1) {
    136       family->Args({});
    137     }
    138     const std::vector<int>* thread_counts =
    139         (family->thread_counts_.empty()
    140              ? &one_thread
    141              : &static_cast<const std::vector<int>&>(family->thread_counts_));
    142     const size_t family_size = family->args_.size() * thread_counts->size();
    143     // The benchmark will be run at least 'family_size' different inputs.
    144     // If 'family_size' is very large warn the user.
    145     if (family_size > kMaxFamilySize) {
    146       Err << "The number of inputs is very large. " << family->name_
    147           << " will be repeated at least " << family_size << " times.\n";
    148     }
    149     // reserve in the special case the regex ".", since we know the final
    150     // family size.
    151     if (spec == ".") benchmarks->reserve(family_size);
    152 
    153     for (auto const& args : family->args_) {
    154       for (int num_threads : *thread_counts) {
    155         BenchmarkInstance instance;
    156         instance.name = family->name_;
    157         instance.benchmark = family.get();
    158         instance.aggregation_report_mode = family->aggregation_report_mode_;
    159         instance.arg = args;
    160         instance.time_unit = family->time_unit_;
    161         instance.range_multiplier = family->range_multiplier_;
    162         instance.min_time = family->min_time_;
    163         instance.iterations = family->iterations_;
    164         instance.repetitions = family->repetitions_;
    165         instance.use_real_time = family->use_real_time_;
    166         instance.use_manual_time = family->use_manual_time_;
    167         instance.complexity = family->complexity_;
    168         instance.complexity_lambda = family->complexity_lambda_;
    169         instance.statistics = &family->statistics_;
    170         instance.threads = num_threads;
    171 
    172         // Add arguments to instance name
    173         size_t arg_i = 0;
    174         for (auto const& arg : args) {
    175           instance.name += "/";
    176 
    177           if (arg_i < family->arg_names_.size()) {
    178             const auto& arg_name = family->arg_names_[arg_i];
    179             if (!arg_name.empty()) {
    180               instance.name +=
    181                   StrFormat("%s:", family->arg_names_[arg_i].c_str());
    182             }
    183           }
    184 
    185           // we know that the args are always non-negative (see 'AddRange()'),
    186           // thus print as 'unsigned'. BUT, do a cast due to the 32-bit builds.
    187           instance.name += StrFormat("%lu", static_cast<unsigned long>(arg));
    188           ++arg_i;
    189         }
    190 
    191         if (!IsZero(family->min_time_))
    192           instance.name += StrFormat("/min_time:%0.3f", family->min_time_);
    193         if (family->iterations_ != 0) {
    194           instance.name +=
    195               StrFormat("/iterations:%lu",
    196                         static_cast<unsigned long>(family->iterations_));
    197         }
    198         if (family->repetitions_ != 0)
    199           instance.name += StrFormat("/repeats:%d", family->repetitions_);
    200 
    201         if (family->use_manual_time_) {
    202           instance.name += "/manual_time";
    203         } else if (family->use_real_time_) {
    204           instance.name += "/real_time";
    205         }
    206 
    207         // Add the number of threads used to the name
    208         if (!family->thread_counts_.empty()) {
    209           instance.name += StrFormat("/threads:%d", instance.threads);
    210         }
    211 
    212         if ((re.Match(instance.name) && !isNegativeFilter) ||
    213             (!re.Match(instance.name) && isNegativeFilter)) {
    214           instance.last_benchmark_instance = (&args == &family->args_.back());
    215           benchmarks->push_back(std::move(instance));
    216         }
    217       }
    218     }
    219   }
    220   return true;
    221 }
    222 
    223 Benchmark* RegisterBenchmarkInternal(Benchmark* bench) {
    224   std::unique_ptr<Benchmark> bench_ptr(bench);
    225   BenchmarkFamilies* families = BenchmarkFamilies::GetInstance();
    226   families->AddBenchmark(std::move(bench_ptr));
    227   return bench;
    228 }
    229 
    230 // FIXME: This function is a hack so that benchmark.cc can access
    231 // `BenchmarkFamilies`
    232 bool FindBenchmarksInternal(const std::string& re,
    233                             std::vector<BenchmarkInstance>* benchmarks,
    234                             std::ostream* Err) {
    235   return BenchmarkFamilies::GetInstance()->FindBenchmarks(re, benchmarks, Err);
    236 }
    237 
    238 //=============================================================================//
    239 //                               Benchmark
    240 //=============================================================================//
    241 
    242 Benchmark::Benchmark(const char* name)
    243     : name_(name),
    244       aggregation_report_mode_(ARM_Unspecified),
    245       time_unit_(kNanosecond),
    246       range_multiplier_(kRangeMultiplier),
    247       min_time_(0),
    248       iterations_(0),
    249       repetitions_(0),
    250       use_real_time_(false),
    251       use_manual_time_(false),
    252       complexity_(oNone),
    253       complexity_lambda_(nullptr) {
    254   ComputeStatistics("mean", StatisticsMean);
    255   ComputeStatistics("median", StatisticsMedian);
    256   ComputeStatistics("stddev", StatisticsStdDev);
    257 }
    258 
    259 Benchmark::~Benchmark() {}
    260 
    261 Benchmark* Benchmark::Arg(int64_t x) {
    262   CHECK(ArgsCnt() == -1 || ArgsCnt() == 1);
    263   args_.push_back({x});
    264   return this;
    265 }
    266 
    267 Benchmark* Benchmark::Unit(TimeUnit unit) {
    268   time_unit_ = unit;
    269   return this;
    270 }
    271 
    272 Benchmark* Benchmark::Range(int64_t start, int64_t limit) {
    273   CHECK(ArgsCnt() == -1 || ArgsCnt() == 1);
    274   std::vector<int64_t> arglist;
    275   AddRange(&arglist, start, limit, range_multiplier_);
    276 
    277   for (int64_t i : arglist) {
    278     args_.push_back({i});
    279   }
    280   return this;
    281 }
    282 
    283 Benchmark* Benchmark::Ranges(
    284     const std::vector<std::pair<int64_t, int64_t>>& ranges) {
    285   CHECK(ArgsCnt() == -1 || ArgsCnt() == static_cast<int>(ranges.size()));
    286   std::vector<std::vector<int64_t>> arglists(ranges.size());
    287   std::size_t total = 1;
    288   for (std::size_t i = 0; i < ranges.size(); i++) {
    289     AddRange(&arglists[i], ranges[i].first, ranges[i].second,
    290              range_multiplier_);
    291     total *= arglists[i].size();
    292   }
    293 
    294   std::vector<std::size_t> ctr(arglists.size(), 0);
    295 
    296   for (std::size_t i = 0; i < total; i++) {
    297     std::vector<int64_t> tmp;
    298     tmp.reserve(arglists.size());
    299 
    300     for (std::size_t j = 0; j < arglists.size(); j++) {
    301       tmp.push_back(arglists[j].at(ctr[j]));
    302     }
    303 
    304     args_.push_back(std::move(tmp));
    305 
    306     for (std::size_t j = 0; j < arglists.size(); j++) {
    307       if (ctr[j] + 1 < arglists[j].size()) {
    308         ++ctr[j];
    309         break;
    310       }
    311       ctr[j] = 0;
    312     }
    313   }
    314   return this;
    315 }
    316 
    317 Benchmark* Benchmark::ArgName(const std::string& name) {
    318   CHECK(ArgsCnt() == -1 || ArgsCnt() == 1);
    319   arg_names_ = {name};
    320   return this;
    321 }
    322 
    323 Benchmark* Benchmark::ArgNames(const std::vector<std::string>& names) {
    324   CHECK(ArgsCnt() == -1 || ArgsCnt() == static_cast<int>(names.size()));
    325   arg_names_ = names;
    326   return this;
    327 }
    328 
    329 Benchmark* Benchmark::DenseRange(int64_t start, int64_t limit, int step) {
    330   CHECK(ArgsCnt() == -1 || ArgsCnt() == 1);
    331   CHECK_GE(start, 0);
    332   CHECK_LE(start, limit);
    333   for (int64_t arg = start; arg <= limit; arg += step) {
    334     args_.push_back({arg});
    335   }
    336   return this;
    337 }
    338 
    339 Benchmark* Benchmark::Args(const std::vector<int64_t>& args) {
    340   CHECK(ArgsCnt() == -1 || ArgsCnt() == static_cast<int>(args.size()));
    341   args_.push_back(args);
    342   return this;
    343 }
    344 
    345 Benchmark* Benchmark::Apply(void (*custom_arguments)(Benchmark* benchmark)) {
    346   custom_arguments(this);
    347   return this;
    348 }
    349 
    350 Benchmark* Benchmark::RangeMultiplier(int multiplier) {
    351   CHECK(multiplier > 1);
    352   range_multiplier_ = multiplier;
    353   return this;
    354 }
    355 
    356 Benchmark* Benchmark::MinTime(double t) {
    357   CHECK(t > 0.0);
    358   CHECK(iterations_ == 0);
    359   min_time_ = t;
    360   return this;
    361 }
    362 
    363 Benchmark* Benchmark::Iterations(size_t n) {
    364   CHECK(n > 0);
    365   CHECK(IsZero(min_time_));
    366   iterations_ = n;
    367   return this;
    368 }
    369 
    370 Benchmark* Benchmark::Repetitions(int n) {
    371   CHECK(n > 0);
    372   repetitions_ = n;
    373   return this;
    374 }
    375 
    376 Benchmark* Benchmark::ReportAggregatesOnly(bool value) {
    377   aggregation_report_mode_ = value ? ARM_ReportAggregatesOnly : ARM_Default;
    378   return this;
    379 }
    380 
    381 Benchmark* Benchmark::DisplayAggregatesOnly(bool value) {
    382   // If we were called, the report mode is no longer 'unspecified', in any case.
    383   aggregation_report_mode_ = static_cast<AggregationReportMode>(
    384       aggregation_report_mode_ | ARM_Default);
    385 
    386   if (value) {
    387     aggregation_report_mode_ = static_cast<AggregationReportMode>(
    388         aggregation_report_mode_ | ARM_DisplayReportAggregatesOnly);
    389   } else {
    390     aggregation_report_mode_ = static_cast<AggregationReportMode>(
    391         aggregation_report_mode_ & ~ARM_DisplayReportAggregatesOnly);
    392   }
    393 
    394   return this;
    395 }
    396 
    397 Benchmark* Benchmark::UseRealTime() {
    398   CHECK(!use_manual_time_)
    399       << "Cannot set UseRealTime and UseManualTime simultaneously.";
    400   use_real_time_ = true;
    401   return this;
    402 }
    403 
    404 Benchmark* Benchmark::UseManualTime() {
    405   CHECK(!use_real_time_)
    406       << "Cannot set UseRealTime and UseManualTime simultaneously.";
    407   use_manual_time_ = true;
    408   return this;
    409 }
    410 
    411 Benchmark* Benchmark::Complexity(BigO complexity) {
    412   complexity_ = complexity;
    413   return this;
    414 }
    415 
    416 Benchmark* Benchmark::Complexity(BigOFunc* complexity) {
    417   complexity_lambda_ = complexity;
    418   complexity_ = oLambda;
    419   return this;
    420 }
    421 
    422 Benchmark* Benchmark::ComputeStatistics(std::string name,
    423                                         StatisticsFunc* statistics) {
    424   statistics_.emplace_back(name, statistics);
    425   return this;
    426 }
    427 
    428 Benchmark* Benchmark::Threads(int t) {
    429   CHECK_GT(t, 0);
    430   thread_counts_.push_back(t);
    431   return this;
    432 }
    433 
    434 Benchmark* Benchmark::ThreadRange(int min_threads, int max_threads) {
    435   CHECK_GT(min_threads, 0);
    436   CHECK_GE(max_threads, min_threads);
    437 
    438   AddRange(&thread_counts_, min_threads, max_threads, 2);
    439   return this;
    440 }
    441 
    442 Benchmark* Benchmark::DenseThreadRange(int min_threads, int max_threads,
    443                                        int stride) {
    444   CHECK_GT(min_threads, 0);
    445   CHECK_GE(max_threads, min_threads);
    446   CHECK_GE(stride, 1);
    447 
    448   for (auto i = min_threads; i < max_threads; i += stride) {
    449     thread_counts_.push_back(i);
    450   }
    451   thread_counts_.push_back(max_threads);
    452   return this;
    453 }
    454 
    455 Benchmark* Benchmark::ThreadPerCpu() {
    456   thread_counts_.push_back(CPUInfo::Get().num_cpus);
    457   return this;
    458 }
    459 
    460 void Benchmark::SetName(const char* name) { name_ = name; }
    461 
    462 int Benchmark::ArgsCnt() const {
    463   if (args_.empty()) {
    464     if (arg_names_.empty()) return -1;
    465     return static_cast<int>(arg_names_.size());
    466   }
    467   return static_cast<int>(args_.front().size());
    468 }
    469 
    470 //=============================================================================//
    471 //                            FunctionBenchmark
    472 //=============================================================================//
    473 
    474 void FunctionBenchmark::Run(State& st) { func_(st); }
    475 
    476 }  // end namespace internal
    477 
    478 void ClearRegisteredBenchmarks() {
    479   internal::BenchmarkFamilies::GetInstance()->ClearBenchmarks();
    480 }
    481 
    482 }  // end namespace benchmark