// Copyright 2015 Google Inc. All rights reserved. // // Licensed under the Apache License, Version 2.0 (the "License"); // you may not use this file except in compliance with the License. // You may obtain a copy of the License at // // http://www.apache.org/licenses/LICENSE-2.0 // // Unless required by applicable law or agreed to in writing, software // distributed under the License is distributed on an "AS IS" BASIS, // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. // See the License for the specific language governing permissions and // limitations under the License. #include "benchmark/benchmark.h" #include "benchmark_api_internal.h" #include "internal_macros.h" #ifndef BENCHMARK_OS_WINDOWS #include <sys/resource.h> #include <sys/time.h> #include <unistd.h> #endif #include <algorithm> #include <atomic> #include <condition_variable> #include <cstdio> #include <cstdlib> #include <cstring> #include <fstream> #include <iostream> #include <memory> #include <thread> #include "check.h" #include "commandlineflags.h" #include "complexity.h" #include "log.h" #include "mutex.h" #include "re.h" #include "stat.h" #include "string_util.h" #include "sysinfo.h" #include "timers.h" namespace benchmark { namespace { // For non-dense Range, intermediate values are powers of kRangeMultiplier. static const int kRangeMultiplier = 8; // The size of a benchmark family determines is the number of inputs to repeat // the benchmark on. If this is "large" then warn the user during configuration. static const size_t kMaxFamilySize = 100; } // end namespace namespace internal { //=============================================================================// // BenchmarkFamilies //=============================================================================// // Class for managing registered benchmarks. Note that each registered // benchmark identifies a family of related benchmarks to run. class BenchmarkFamilies { public: static BenchmarkFamilies* GetInstance(); // Registers a benchmark family and returns the index assigned to it. size_t AddBenchmark(std::unique_ptr<Benchmark> family); // Extract the list of benchmark instances that match the specified // regular expression. bool FindBenchmarks(const std::string& re, std::vector<Benchmark::Instance>* benchmarks, std::ostream* Err); private: BenchmarkFamilies() {} std::vector<std::unique_ptr<Benchmark>> families_; Mutex mutex_; }; BenchmarkFamilies* BenchmarkFamilies::GetInstance() { static BenchmarkFamilies instance; return &instance; } size_t BenchmarkFamilies::AddBenchmark(std::unique_ptr<Benchmark> family) { MutexLock l(mutex_); size_t index = families_.size(); families_.push_back(std::move(family)); return index; } bool BenchmarkFamilies::FindBenchmarks( const std::string& spec, std::vector<Benchmark::Instance>* benchmarks, std::ostream* ErrStream) { CHECK(ErrStream); auto& Err = *ErrStream; // Make regular expression out of command-line flag std::string error_msg; Regex re; if (!re.Init(spec, &error_msg)) { Err << "Could not compile benchmark re: " << error_msg << std::endl; return false; } // Special list of thread counts to use when none are specified const std::vector<int> one_thread = {1}; MutexLock l(mutex_); for (std::unique_ptr<Benchmark>& family : families_) { // Family was deleted or benchmark doesn't match if (!family) continue; if (family->ArgsCnt() == -1) { family->Args({}); } const std::vector<int>* thread_counts = (family->thread_counts_.empty() ? &one_thread : &static_cast<const std::vector<int>&>(family->thread_counts_)); const size_t family_size = family->args_.size() * thread_counts->size(); // The benchmark will be run at least 'family_size' different inputs. // If 'family_size' is very large warn the user. if (family_size > kMaxFamilySize) { Err << "The number of inputs is very large. " << family->name_ << " will be repeated at least " << family_size << " times.\n"; } // reserve in the special case the regex ".", since we know the final // family size. if (spec == ".") benchmarks->reserve(family_size); for (auto const& args : family->args_) { for (int num_threads : *thread_counts) { Benchmark::Instance instance; instance.name = family->name_; instance.benchmark = family.get(); instance.report_mode = family->report_mode_; instance.arg = args; instance.time_unit = family->time_unit_; instance.range_multiplier = family->range_multiplier_; instance.min_time = family->min_time_; instance.repetitions = family->repetitions_; instance.use_real_time = family->use_real_time_; instance.use_manual_time = family->use_manual_time_; instance.complexity = family->complexity_; instance.complexity_lambda = family->complexity_lambda_; instance.threads = num_threads; // Add arguments to instance name size_t arg_i = 0; for (auto const& arg : args) { instance.name += "/"; if (arg_i < family->arg_names_.size()) { const auto& arg_name = family->arg_names_[arg_i]; if (!arg_name.empty()) { instance.name += StringPrintF("%s:", family->arg_names_[arg_i].c_str()); } } AppendHumanReadable(arg, &instance.name); ++arg_i; } if (!IsZero(family->min_time_)) { instance.name += StringPrintF("/min_time:%0.3f", family->min_time_); } if (family->repetitions_ != 0) { instance.name += StringPrintF("/repeats:%d", family->repetitions_); } if (family->use_manual_time_) { instance.name += "/manual_time"; } else if (family->use_real_time_) { instance.name += "/real_time"; } // Add the number of threads used to the name if (!family->thread_counts_.empty()) { instance.name += StringPrintF("/threads:%d", instance.threads); } if (re.Match(instance.name)) { instance.last_benchmark_instance = (&args == &family->args_.back()); benchmarks->push_back(std::move(instance)); } } } } return true; } Benchmark* RegisterBenchmarkInternal(Benchmark* bench) { std::unique_ptr<Benchmark> bench_ptr(bench); BenchmarkFamilies* families = BenchmarkFamilies::GetInstance(); families->AddBenchmark(std::move(bench_ptr)); return bench; } // FIXME: This function is a hack so that benchmark.cc can access // `BenchmarkFamilies` bool FindBenchmarksInternal(const std::string& re, std::vector<Benchmark::Instance>* benchmarks, std::ostream* Err) { return BenchmarkFamilies::GetInstance()->FindBenchmarks(re, benchmarks, Err); } //=============================================================================// // Benchmark //=============================================================================// Benchmark::Benchmark(const char* name) : name_(name), report_mode_(RM_Unspecified), time_unit_(kNanosecond), range_multiplier_(kRangeMultiplier), min_time_(0), repetitions_(0), use_real_time_(false), use_manual_time_(false), complexity_(oNone), complexity_lambda_(nullptr) {} Benchmark::~Benchmark() {} void Benchmark::AddRange(std::vector<int>* dst, int lo, int hi, int mult) { CHECK_GE(lo, 0); CHECK_GE(hi, lo); CHECK_GE(mult, 2); // Add "lo" dst->push_back(lo); static const int kint32max = std::numeric_limits<int32_t>::max(); // Now space out the benchmarks in multiples of "mult" for (int32_t i = 1; i < kint32max / mult; i *= mult) { if (i >= hi) break; if (i > lo) { dst->push_back(i); } } // Add "hi" (if different from "lo") if (hi != lo) { dst->push_back(hi); } } Benchmark* Benchmark::Arg(int x) { CHECK(ArgsCnt() == -1 || ArgsCnt() == 1); args_.push_back({x}); return this; } Benchmark* Benchmark::Unit(TimeUnit unit) { time_unit_ = unit; return this; } Benchmark* Benchmark::Range(int start, int limit) { CHECK(ArgsCnt() == -1 || ArgsCnt() == 1); std::vector<int> arglist; AddRange(&arglist, start, limit, range_multiplier_); for (int i : arglist) { args_.push_back({i}); } return this; } Benchmark* Benchmark::Ranges(const std::vector<std::pair<int, int>>& ranges) { CHECK(ArgsCnt() == -1 || ArgsCnt() == static_cast<int>(ranges.size())); std::vector<std::vector<int>> arglists(ranges.size()); std::size_t total = 1; for (std::size_t i = 0; i < ranges.size(); i++) { AddRange(&arglists[i], ranges[i].first, ranges[i].second, range_multiplier_); total *= arglists[i].size(); } std::vector<std::size_t> ctr(arglists.size(), 0); for (std::size_t i = 0; i < total; i++) { std::vector<int> tmp; tmp.reserve(arglists.size()); for (std::size_t j = 0; j < arglists.size(); j++) { tmp.push_back(arglists[j].at(ctr[j])); } args_.push_back(std::move(tmp)); for (std::size_t j = 0; j < arglists.size(); j++) { if (ctr[j] + 1 < arglists[j].size()) { ++ctr[j]; break; } ctr[j] = 0; } } return this; } Benchmark* Benchmark::ArgName(const std::string& name) { CHECK(ArgsCnt() == -1 || ArgsCnt() == 1); arg_names_ = {name}; return this; } Benchmark* Benchmark::ArgNames(const std::vector<std::string>& names) { CHECK(ArgsCnt() == -1 || ArgsCnt() == static_cast<int>(names.size())); arg_names_ = names; return this; } Benchmark* Benchmark::DenseRange(int start, int limit, int step) { CHECK(ArgsCnt() == -1 || ArgsCnt() == 1); CHECK_GE(start, 0); CHECK_LE(start, limit); for (int arg = start; arg <= limit; arg += step) { args_.push_back({arg}); } return this; } Benchmark* Benchmark::Args(const std::vector<int>& args) { CHECK(ArgsCnt() == -1 || ArgsCnt() == static_cast<int>(args.size())); args_.push_back(args); return this; } Benchmark* Benchmark::Apply(void (*custom_arguments)(Benchmark* benchmark)) { custom_arguments(this); return this; } Benchmark* Benchmark::RangeMultiplier(int multiplier) { CHECK(multiplier > 1); range_multiplier_ = multiplier; return this; } Benchmark* Benchmark::Repetitions(int n) { CHECK(n > 0); repetitions_ = n; return this; } Benchmark* Benchmark::ReportAggregatesOnly(bool value) { report_mode_ = value ? RM_ReportAggregatesOnly : RM_Default; return this; } Benchmark* Benchmark::MinTime(double t) { CHECK(t > 0.0); min_time_ = t; return this; } Benchmark* Benchmark::UseRealTime() { CHECK(!use_manual_time_) << "Cannot set UseRealTime and UseManualTime simultaneously."; use_real_time_ = true; return this; } Benchmark* Benchmark::UseManualTime() { CHECK(!use_real_time_) << "Cannot set UseRealTime and UseManualTime simultaneously."; use_manual_time_ = true; return this; } Benchmark* Benchmark::Complexity(BigO complexity) { complexity_ = complexity; return this; } Benchmark* Benchmark::Complexity(BigOFunc* complexity) { complexity_lambda_ = complexity; complexity_ = oLambda; return this; } Benchmark* Benchmark::Threads(int t) { CHECK_GT(t, 0); thread_counts_.push_back(t); return this; } Benchmark* Benchmark::ThreadRange(int min_threads, int max_threads) { CHECK_GT(min_threads, 0); CHECK_GE(max_threads, min_threads); AddRange(&thread_counts_, min_threads, max_threads, 2); return this; } Benchmark* Benchmark::DenseThreadRange(int min_threads, int max_threads, int stride) { CHECK_GT(min_threads, 0); CHECK_GE(max_threads, min_threads); CHECK_GE(stride, 1); for (auto i = min_threads; i < max_threads; i += stride) { thread_counts_.push_back(i); } thread_counts_.push_back(max_threads); return this; } Benchmark* Benchmark::ThreadPerCpu() { static int num_cpus = NumCPUs(); thread_counts_.push_back(num_cpus); return this; } void Benchmark::SetName(const char* name) { name_ = name; } int Benchmark::ArgsCnt() const { if (args_.empty()) { if (arg_names_.empty()) return -1; return static_cast<int>(arg_names_.size()); } return static_cast<int>(args_.front().size()); } //=============================================================================// // FunctionBenchmark //=============================================================================// void FunctionBenchmark::Run(State& st) { func_(st); } } // end namespace internal } // end namespace benchmark