google
diff --git a/‎src/benchmark.cc
Lines changed: 203 additions & 118 deletions b/‎src/benchmark.cc
Lines changed: 203 additions & 118 deletions
@@ -198,140 +198,225 @@ struct RunResults {
   bool file_report_aggregates_only = false;
 };
 
-RunResults RunBenchmark(
-    const benchmark::internal::Benchmark::Instance& b,
-    std::vector<BenchmarkReporter::Run>* complexity_reports) {
+class BenchmarkRunner {
   RunResults run_results;
 
-  const bool has_explicit_iteration_count = b.iterations != 0;
-  size_t iters = has_explicit_iteration_count ? b.iterations : 1;
-  std::unique_ptr<internal::ThreadManager> manager;
-  std::vector<std::thread> pool(b.threads - 1);
-  const int repeats =
-      b.repetitions != 0 ? b.repetitions : FLAGS_benchmark_repetitions;
-  if (repeats != 1) {
-    run_results.display_report_aggregates_only =
-        (FLAGS_benchmark_report_aggregates_only ||
-         FLAGS_benchmark_display_aggregates_only);
-    run_results.file_report_aggregates_only =
-        FLAGS_benchmark_report_aggregates_only;
-    if (b.aggregation_report_mode != internal::ARM_Unspecified) {
-      run_results.display_report_aggregates_only =
-          (b.aggregation_report_mode &
-           internal::ARM_DisplayReportAggregatesOnly);
-      run_results.file_report_aggregates_only =
-          (b.aggregation_report_mode & internal::ARM_FileReportAggregatesOnly);
+  const benchmark::internal::Benchmark::Instance& b;
+  std::vector<BenchmarkReporter::Run>& complexity_reports;
+
+  const double min_time;
+  const int repeats;
+  const bool has_explicit_iteration_count;
+
+  std::vector<std::thread> pool;
+
+  size_t iters;  // preserved between repetitions!
+  // So only the first repetition has to find/calculate it,
+  // the other repetitions will just use that precomputed iteration count.
+
+  struct IterationResults {
+    internal::ThreadManager::Result results;
+    size_t iters;
+    double seconds;
+  };
+  IterationResults doNIterations() {
+    VLOG(2) << "Running " << b.name << " for " << iters << "\n";
+
+    std::unique_ptr<internal::ThreadManager> manager;
+    manager.reset(new internal::ThreadManager(b.threads));
+
+    // Run all but one thread in separate threads
+    for (std::size_t ti = 0; ti < pool.size(); ++ti) {
+      pool[ti] = std::thread(&RunInThread, &b, iters, static_cast<int>(ti + 1),
+                             manager.get());
     }
+    // And run one thread here directly.
+    // (If we were asked to run just one thread, we don't create new threads.)
+    // Yes, we need to do this here *after* we start the separate threads.
+    RunInThread(&b, iters, 0, manager.get());
+
+    // The main thread has finished. Now let's wait for the other threads.
+    manager->WaitForAllThreads();
+    for (std::thread& thread : pool) thread.join();
+
+    IterationResults i;
+    // Acquire the measurements/counters from the manager, UNDER THE LOCK!
+    {
+      MutexLock l(manager->GetBenchmarkMutex());
+      i.results = manager->results;
+    }
+
+    // And get rid of the manager.
+    manager.reset();
+
+    // Adjust real/manual time stats since they were reported per thread.
+    i.results.real_time_used /= b.threads;
+    i.results.manual_time_used /= b.threads;
+
+    VLOG(2) << "Ran in " << i.results.cpu_time_used << "/"
+            << i.results.real_time_used << "\n";
+
+    // So for how long were we running?
+    i.iters = iters;
     // Base decisions off of real time if requested by this benchmark.
+    i.seconds = i.results.cpu_time_used;
+    if (b.use_manual_time) {
+      i.seconds = i.results.manual_time_used;
+    } else if (b.use_real_time) {
+      i.seconds = i.results.real_time_used;
+    }
+
+    return i;
+  }
+
+  size_t predictNumItersNeeded(const IterationResults& i) const {
+    // See how much iterations should be increased by.
+    // Note: Avoid division by zero with max(seconds, 1ns).
+    double multiplier = min_time * 1.4 / std::max(i.seconds, 1e-9);
+    // If our last run was at least 10% of FLAGS_benchmark_min_time then we
+    // use the multiplier directly.
+    // Otherwise we use at most 10 times expansion.
+    // NOTE: When the last run was at least 10% of the min time the max
+    // expansion should be 14x.
+    bool is_significant = (i.seconds / min_time) > 0.1;
+    multiplier = is_significant ? multiplier : std::min(10.0, multiplier);
+    if (multiplier <= 1.0) multiplier = 2.0;
+
+    // So what seems to be the sufficiently-large iteration count? Round up.
+    const size_t max_next_iters =
+        0.5 + std::max(multiplier * i.iters, i.iters + 1.0);
+    // But we do have *some* sanity limits though..
+    const size_t next_iters = std::min(max_next_iters, kMaxIterations);
+
+    VLOG(3) << "Next iters: " << next_iters << ", " << multiplier << "\n";
+    return next_iters;  // round up before conversion to integer.
+  }
+
+  bool shouldReportIterationResults(const IterationResults& i) const {
+    // Determine if this run should be reported;
+    // Either it has run for a sufficient amount of time
+    // or because an error was reported.
+    return i.results.has_error_ ||
+           i.iters >= kMaxIterations ||  // Too many iterations already.
+           i.seconds >= min_time ||      // The elapsed time is large enough.
+           // CPU time is specified but the elapsed real time greatly exceeds
+           // the minimum time.
+           // Note that user provided timers are except from this sanity check.
+           ((i.results.real_time_used >= 5 * min_time) && !b.use_manual_time);
   }
-  for (int repetition_num = 0; repetition_num < repeats; repetition_num++) {
+ 10000
+  void doOneRepetition(bool not_in_the_first_repetition) {
+    IterationResults i;
+
+    // We *may* be gradually increasing the length (iteration count)
+    // of the benchmark until we decide the results are significant.
+    // And once we do, we report those last results and exit.
+    // Please do note that the if there are repetitions, the iteration count
+    // is *only* calculated for the *first* repetition, and other repetitions
+    // simply use that precomputed iteration count.
     for (;;) {
-      // Try benchmark
-      VLOG(2) << "Running " << b.name << " for " << iters << "\n";
+      i = doNIterations();
+
+      // Do we consider the results to be significant?
+      // If we are doing repetitions, and the first repetition was already done,
+      // it has calculated the correct iteration time, so we have run that very
+      // iteration count just now. No need to calculate anything. Just report.
+      // Else, the normal rules apply.
+      const bool results_are_significant = not_in_the_first_repetition ||
+                                           has_explicit_iteration_count ||
+                                           shouldReportIterationResults(i);
+
+      if (results_are_significant) break;  // Good, let's report them!
+
+      // Nope, bad iteration. Let's re-estimate the hopefully-sufficient
+      // iteration count, and run the benchmark again...
+
+      iters = predictNumItersNeeded(i);
+      assert(iters > i.iters &&
+             "if we did more iterations than we want to do the next time, "
+             "then we should have accepted the current iteration run.");
+    }
 
-      manager.reset(new internal::ThreadManager(b.threads));
-      for (std::size_t ti = 0; ti < pool.size(); ++ti) {
-        pool[ti] = std::thread(&RunInThread, &b, iters,
-                               static_cast<int>(ti + 1), manager.get());
-      }
-      RunInThread(&b, iters, 0, manager.get());
+    // Oh, one last thing, we need to also produce the 'memory measurements'..
+    MemoryManager::Result memory_result;
+    size_t memory_iterations = 0;
+    if (memory_manager != nullptr) {
+      // Only run a few iterations to reduce the impact of one-time
+      // allocations in benchmarks that are not properly managed.
+      memory_iterations = std::min<size_t>(16, iters);
+      memory_manager->Start();
+      std::unique_ptr<internal::ThreadManager> manager;
+      manager.reset(new internal::ThreadManager(1));
+      RunInThread(&b, memory_iterations, 0, manager.get());
       manager->WaitForAllThreads();
-      for (std::thread& thread : pool) thread.join();
-      internal::ThreadManager::Result results;
-      {
-        MutexLock l(manager->GetBenchmarkMutex());
-        results = manager->results;
-      }
       manager.reset();
-      // Adjust real/manual time stats since they were reported per thread.
-      results.real_time_used /= b.threads;
-      results.manual_time_used /= b.threads;
-
-      VLOG(2) << "Ran in " << results.cpu_time_used << "/"
-              << results.real_time_used << "\n";
-
-      // Base decisions off of real time if requested by this benchmark.
-      double seconds = results.cpu_time_used;
-      if (b.use_manual_time) {
-        seconds = results.manual_time_used;
-      } else if (b.use_real_time) {
-        seconds = results.real_time_used;
-      }
 
-      const double min_time =
-          !IsZero(b.min_time) ? b.min_time : FLAGS_benchmark_min_time;
-
-      // clang-format off
-      // turn off clang-format since it mangles prettiness here
-      // Determine if this run should be reported; Either it has
-      // run for a sufficient amount of time or because an error was reported.
-      const bool should_report =  repetition_num > 0
-        || has_explicit_iteration_count  // An exact iteration count was requested
-        || results.has_error_
-        || iters >= kMaxIterations  // No chance to try again, we hit the limit.
-        || seconds >= min_time  // the elapsed time is large enough
-        // CPU time is specified but the elapsed real time greatly exceeds the
-        // minimum time. Note that user provided timers are except from this
-        // sanity check.
-        || ((results.real_time_used >= 5 * min_time) && !b.use_manual_time);
-      // clang-format on
-
-      if (should_report) {
-        MemoryManager::Result memory_result;
-        size_t memory_iterations = 0;
-        if (memory_manager != nullptr) {
-          // Only run a few iterations to reduce the impact of one-time
-          // allocations in benchmarks that are not properly managed.
-          memory_iterations = std::min<size_t>(16, iters);
-          memory_manager->Start();
-          manager.reset(new internal::ThreadManager(1));
-          RunInThread(&b, memory_iterations, 0, manager.get());
-          manager->WaitForAllThreads();
-          manager.reset();
-
-          memory_manager->Stop(&memory_result);
-        }
-
-        BenchmarkReporter::Run report = CreateRunReport(
-            b, results, memory_iterations, memory_result, seconds);
-        if (!report.error_occurred && b.complexity != oNone)
-          complexity_reports->push_back(report);
-        run_results.non_aggregates.push_back(report);
-        break;
-      }
+      memory_manager->Stop(&memory_result);
+    }
+
+    // Ok, now actualy report.
+    BenchmarkReporter::Run report = CreateRunReport(
+        b, i.results, memory_iterations, memory_result, i.seconds);
+
+    if (!report.error_occurred && b.complexity != oNone)
+      complexity_reports.push_back(report);
+
+    run_results.non_aggregates.push_back(report);
+  }
 
-      // See how much iterations should be increased by
-      // Note: Avoid division by zero with max(seconds, 1ns).
-      double multiplier = min_time * 1.4 / std::max(seconds, 1e-9);
-      // If our last run was at least 10% of FLAGS_benchmark_min_time then we
-      // use the multiplier directly. Otherwise we use at most 10 times
-      // expansion.
-      // expansion should be 14x.
-      bool is_significant = (seconds / min_time) > 0.1;
-      multiplier = is_significant ? multiplier : std::min(10.0, multiplier);
-      if (multiplier <= 1.0) multiplier = 2.0;
-      double next_iters = std::max(multiplier * iters, iters + 1.0);
-      if (next_iters > kMaxIterations) {
-        next_iters = kMaxIterations;
+ public:
+  BenchmarkRunner(const benchmark::internal::Benchmark::Instance& b_,
+                  std::vector<BenchmarkReporter::Run>* complexity_reports_)
+      : b(b_),
+        complexity_reports(*complexity_reports_),
+        min_time(!IsZero(b.min_time) ? b.min_time : FLAGS_benchmark_min_time),
+        repeats(b.repetitions != 0 ? b.repetitions
+                                   : FLAGS_benchmark_repetitions),
+        has_explicit_iteration_count(b.iterations != 0),
+        pool(b.threads - 1),
+        iters(has_explicit_iteration_count ? b.iterations : 1) {
+    if (repeats != 1) {
+      run_results.display_report_aggregates_only =
+          (FLAGS_benchmark_report_aggregates_only ||
+           FLAGS_benchmark_display_aggregates_only);
+      run_results.file_report_aggregates_only =
+          FLAGS_benchmark_report_aggregates_only;
+      if (b.aggregation_report_mode != internal::ARM_Unspecified) {
+        run_results.display_report_aggregates_only =
+            (b.aggregation_report_mode &
+             internal::ARM_DisplayReportAggregatesOnly);
+        run_results.file_report_aggregates_only =
+            (b.aggregation_report_mode &
+             internal::ARM_FileReportAggregatesOnly);
       }
-      VLOG(3) << "Next iters: " << next_iters << ", " << multiplier << "\n";
-      iters = static_cast<int>(next_iters + 0.5);
     }
-  }
 
-  // Calculate additional statistics
-  run_results.aggregates_only = ComputeStats(run_results.non_aggregates);
+    for (int repetition_num = 0; repetition_num < repeats; repetition_num++) {
+      const bool not_in_the_first_repetition = repetition_num != 0;
+      doOneRepetition(not_in_the_first_repetition);
+    }
+
+    // Calculate additional statistics
+    run_results.aggregates_only = ComputeStats(run_results.non_aggregates);
 
-  // Maybe calculate complexity report
-  if ((b.complexity != oNone) && b.last_benchmark_instance) {
-    auto additional_run_stats = ComputeBigO(*complexity_reports);
-    run_results.aggregates_only.insert(run_results.aggregates_only.end(),
-                                       additional_run_stats.begin(),
-                                       additional_run_stats.end());
-    complexity_reports->clear();
+    // Maybe calculate complexity report
+    if ((b.complexity != oNone) && b.last_benchmark_instance) {
+      auto additional_run_stats = ComputeBigO(complexity_reports);
+      run_results.aggregates_only.insert(run_results.aggregates_only.end(),
+                                         additional_run_stats.begin(),
+                                         additional_run_stats.end());
+      complexity_reports.clear();
+    }
   }
 
-  return run_results;
+  RunResults getResults() { return run_results; }
+};
+
+RunResults RunBenchmark(
+    const benchmark::internal::Benchmark::Instance& b,
+    std::vector<BenchmarkReporter::Run>* complexity_reports) {
+  BenchmarkRunner r(b, complexity_reports);
+  return r.getResults();
 }
 
 }  // namespace