| 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/benchmark.h" |
| 16 | #include "benchmark_api_internal.h" |
| 17 | #include "internal_macros.h" |
| 18 | |
| 19 | #ifndef BENCHMARK_OS_WINDOWS |
| 20 | #ifndef __Fuchsia__ |
| 21 | #include <sys/resource.h> |
| 22 | #endif |
| 23 | #include <sys/time.h> |
| 24 | #include <unistd.h> |
| 25 | #endif |
| 26 | |
| 27 | #include <algorithm> |
| 28 | #include <atomic> |
| 29 | #include <condition_variable> |
| 30 | #include <cstdio> |
| 31 | #include <cstdlib> |
| 32 | #include <cstring> |
| 33 | #include <fstream> |
| 34 | #include <iostream> |
| 35 | #include <memory> |
| 36 | #include <thread> |
| 37 | |
| 38 | #include "check.h" |
| 39 | #include "commandlineflags.h" |
| 40 | #include "complexity.h" |
| 41 | #include "log.h" |
| 42 | #include "mutex.h" |
| 43 | #include "re.h" |
| 44 | #include "stat.h" |
| 45 | #include "string_util.h" |
| 46 | #include "sysinfo.h" |
| 47 | #include "timers.h" |
| 48 | |
| 49 | namespace benchmark { |
| 50 | |
| 51 | namespace { |
| 52 | // For non-dense Range, intermediate values are powers of kRangeMultiplier. |
| 53 | static const int kRangeMultiplier = 8; |
| 54 | // The size of a benchmark family determines is the number of inputs to repeat |
| 55 | // the benchmark on. If this is "large" then warn the user during configuration. |
| 56 | static const size_t kMaxFamilySize = 100; |
| 57 | } // end namespace |
| 58 | |
| 59 | namespace internal { |
| 60 | |
| 61 | //=============================================================================// |
| 62 | // BenchmarkFamilies |
| 63 | //=============================================================================// |
| 64 | |
| 65 | // Class for managing registered benchmarks. Note that each registered |
| 66 | // benchmark identifies a family of related benchmarks to run. |
| 67 | class BenchmarkFamilies { |
| 68 | public: |
| 69 | static BenchmarkFamilies* GetInstance(); |
| 70 | |
| 71 | // Registers a benchmark family and returns the index assigned to it. |
| 72 | size_t AddBenchmark(std::unique_ptr<Benchmark> family); |
| 73 | |
| 74 | // Extract the list of benchmark instances that match the specified |
| 75 | // regular expression. |
| 76 | bool FindBenchmarks(const std::string& re, |
| 77 | std::vector<Benchmark::Instance>* benchmarks, |
| 78 | std::ostream* Err); |
| 79 | |
| 80 | private: |
| 81 | BenchmarkFamilies() {} |
| 82 | |
| 83 | std::vector<std::unique_ptr<Benchmark>> families_; |
| 84 | Mutex mutex_; |
| 85 | }; |
| 86 | |
| 87 | BenchmarkFamilies* BenchmarkFamilies::GetInstance() { |
| 88 | static BenchmarkFamilies instance; |
| 89 | return &instance; |
| 90 | } |
| 91 | |
| 92 | size_t BenchmarkFamilies::AddBenchmark(std::unique_ptr<Benchmark> family) { |
| 93 | MutexLock l(mutex_); |
| 94 | size_t index = families_.size(); |
| 95 | families_.push_back(std::move(family)); |
| 96 | return index; |
| 97 | } |
| 98 | |
| 99 | bool BenchmarkFamilies::FindBenchmarks( |
| 100 | const std::string& spec, std::vector<Benchmark::Instance>* benchmarks, |
| 101 | std::ostream* ErrStream) { |
| 102 | CHECK(ErrStream); |
| 103 | auto& Err = *ErrStream; |
| 104 | // Make regular expression out of command-line flag |
| 105 | std::string error_msg; |
| 106 | Regex re; |
| 107 | if (!re.Init(spec, &error_msg)) { |
| 108 | Err << "Could not compile benchmark re: " << error_msg << std::endl; |
| 109 | return false; |
| 110 | } |
| 111 | |
| 112 | // Special list of thread counts to use when none are specified |
| 113 | const std::vector<int> one_thread = {1}; |
| 114 | |
| 115 | MutexLock l(mutex_); |
| 116 | for (std::unique_ptr<Benchmark>& family : families_) { |
| 117 | // Family was deleted or benchmark doesn't match |
| 118 | if (!family) continue; |
| 119 | |
| 120 | if (family->ArgsCnt() == -1) { |
| 121 | family->Args({}); |
| 122 | } |
| 123 | const std::vector<int>* thread_counts = |
| 124 | (family->thread_counts_.empty() |
| 125 | ? &one_thread |
| 126 | : &static_cast<const std::vector<int>&>(family->thread_counts_)); |
| 127 | const size_t family_size = family->args_.size() * thread_counts->size(); |
| 128 | // The benchmark will be run at least 'family_size' different inputs. |
| 129 | // If 'family_size' is very large warn the user. |
| 130 | if (family_size > kMaxFamilySize) { |
| 131 | Err << "The number of inputs is very large. " << family->name_ |
| 132 | << " will be repeated at least " << family_size << " times.\n" ; |
| 133 | } |
| 134 | // reserve in the special case the regex ".", since we know the final |
| 135 | // family size. |
| 136 | if (spec == "." ) benchmarks->reserve(family_size); |
| 137 | |
| 138 | for (auto const& args : family->args_) { |
| 139 | for (int num_threads : *thread_counts) { |
| 140 | Benchmark::Instance instance; |
| 141 | instance.name = family->name_; |
| 142 | instance.benchmark = family.get(); |
| 143 | instance.report_mode = family->report_mode_; |
| 144 | instance.arg = args; |
| 145 | instance.time_unit = family->time_unit_; |
| 146 | instance.range_multiplier = family->range_multiplier_; |
| 147 | instance.min_time = family->min_time_; |
| 148 | instance.repetitions = family->repetitions_; |
| 149 | instance.use_real_time = family->use_real_time_; |
| 150 | instance.use_manual_time = family->use_manual_time_; |
| 151 | instance.complexity = family->complexity_; |
| 152 | instance.complexity_lambda = family->complexity_lambda_; |
| 153 | instance.threads = num_threads; |
| 154 | |
| 155 | // Add arguments to instance name |
| 156 | size_t arg_i = 0; |
| 157 | for (auto const& arg : args) { |
| 158 | instance.name += "/" ; |
| 159 | |
| 160 | if (arg_i < family->arg_names_.size()) { |
| 161 | const auto& arg_name = family->arg_names_[arg_i]; |
| 162 | if (!arg_name.empty()) { |
| 163 | instance.name += |
| 164 | StringPrintF("%s:" , family->arg_names_[arg_i].c_str()); |
| 165 | } |
| 166 | } |
| 167 | |
| 168 | AppendHumanReadable(arg, &instance.name); |
| 169 | ++arg_i; |
| 170 | } |
| 171 | |
| 172 | if (!IsZero(family->min_time_)) { |
| 173 | instance.name += StringPrintF("/min_time:%0.3f" , family->min_time_); |
| 174 | } |
| 175 | if (family->repetitions_ != 0) { |
| 176 | instance.name += StringPrintF("/repeats:%d" , family->repetitions_); |
| 177 | } |
| 178 | if (family->use_manual_time_) { |
| 179 | instance.name += "/manual_time" ; |
| 180 | } else if (family->use_real_time_) { |
| 181 | instance.name += "/real_time" ; |
| 182 | } |
| 183 | |
| 184 | // Add the number of threads used to the name |
| 185 | if (!family->thread_counts_.empty()) { |
| 186 | instance.name += StringPrintF("/threads:%d" , instance.threads); |
| 187 | } |
| 188 | |
| 189 | if (re.Match(instance.name)) { |
| 190 | instance.last_benchmark_instance = (&args == &family->args_.back()); |
| 191 | benchmarks->push_back(std::move(instance)); |
| 192 | } |
| 193 | } |
| 194 | } |
| 195 | } |
| 196 | return true; |
| 197 | } |
| 198 | |
| 199 | Benchmark* RegisterBenchmarkInternal(Benchmark* bench) { |
| 200 | std::unique_ptr<Benchmark> bench_ptr(bench); |
| 201 | BenchmarkFamilies* families = BenchmarkFamilies::GetInstance(); |
| 202 | families->AddBenchmark(std::move(bench_ptr)); |
| 203 | return bench; |
| 204 | } |
| 205 | |
| 206 | // FIXME: This function is a hack so that benchmark.cc can access |
| 207 | // `BenchmarkFamilies` |
| 208 | bool FindBenchmarksInternal(const std::string& re, |
| 209 | std::vector<Benchmark::Instance>* benchmarks, |
| 210 | std::ostream* Err) { |
| 211 | return BenchmarkFamilies::GetInstance()->FindBenchmarks(re, benchmarks, Err); |
| 212 | } |
| 213 | |
| 214 | //=============================================================================// |
| 215 | // Benchmark |
| 216 | //=============================================================================// |
| 217 | |
| 218 | Benchmark::Benchmark(const char* name) |
| 219 | : name_(name), |
| 220 | report_mode_(RM_Unspecified), |
| 221 | time_unit_(kNanosecond), |
| 222 | range_multiplier_(kRangeMultiplier), |
| 223 | min_time_(0), |
| 224 | repetitions_(0), |
| 225 | use_real_time_(false), |
| 226 | use_manual_time_(false), |
| 227 | complexity_(oNone), |
| 228 | complexity_lambda_(nullptr) {} |
| 229 | |
| 230 | Benchmark::~Benchmark() {} |
| 231 | |
| 232 | void Benchmark::AddRange(std::vector<int>* dst, int lo, int hi, int mult) { |
| 233 | CHECK_GE(lo, 0); |
| 234 | CHECK_GE(hi, lo); |
| 235 | CHECK_GE(mult, 2); |
| 236 | |
| 237 | // Add "lo" |
| 238 | dst->push_back(lo); |
| 239 | |
| 240 | static const int kint32max = std::numeric_limits<int32_t>::max(); |
| 241 | |
| 242 | // Now space out the benchmarks in multiples of "mult" |
| 243 | for (int32_t i = 1; i < kint32max / mult; i *= mult) { |
| 244 | if (i >= hi) break; |
| 245 | if (i > lo) { |
| 246 | dst->push_back(i); |
| 247 | } |
| 248 | } |
| 249 | // Add "hi" (if different from "lo") |
| 250 | if (hi != lo) { |
| 251 | dst->push_back(hi); |
| 252 | } |
| 253 | } |
| 254 | |
| 255 | Benchmark* Benchmark::Arg(int x) { |
| 256 | CHECK(ArgsCnt() == -1 || ArgsCnt() == 1); |
| 257 | args_.push_back({x}); |
| 258 | return this; |
| 259 | } |
| 260 | |
| 261 | Benchmark* Benchmark::Unit(TimeUnit unit) { |
| 262 | time_unit_ = unit; |
| 263 | return this; |
| 264 | } |
| 265 | |
| 266 | Benchmark* Benchmark::Range(int start, int limit) { |
| 267 | CHECK(ArgsCnt() == -1 || ArgsCnt() == 1); |
| 268 | std::vector<int> arglist; |
| 269 | AddRange(&arglist, start, limit, range_multiplier_); |
| 270 | |
| 271 | for (int i : arglist) { |
| 272 | args_.push_back({i}); |
| 273 | } |
| 274 | return this; |
| 275 | } |
| 276 | |
| 277 | Benchmark* Benchmark::Ranges(const std::vector<std::pair<int, int>>& ranges) { |
| 278 | CHECK(ArgsCnt() == -1 || ArgsCnt() == static_cast<int>(ranges.size())); |
| 279 | std::vector<std::vector<int>> arglists(ranges.size()); |
| 280 | std::size_t total = 1; |
| 281 | for (std::size_t i = 0; i < ranges.size(); i++) { |
| 282 | AddRange(&arglists[i], ranges[i].first, ranges[i].second, |
| 283 | range_multiplier_); |
| 284 | total *= arglists[i].size(); |
| 285 | } |
| 286 | |
| 287 | std::vector<std::size_t> ctr(arglists.size(), 0); |
| 288 | |
| 289 | for (std::size_t i = 0; i < total; i++) { |
| 290 | std::vector<int> tmp; |
| 291 | tmp.reserve(arglists.size()); |
| 292 | |
| 293 | for (std::size_t j = 0; j < arglists.size(); j++) { |
| 294 | tmp.push_back(arglists[j].at(ctr[j])); |
| 295 | } |
| 296 | |
| 297 | args_.push_back(std::move(tmp)); |
| 298 | |
| 299 | for (std::size_t j = 0; j < arglists.size(); j++) { |
| 300 | if (ctr[j] + 1 < arglists[j].size()) { |
| 301 | ++ctr[j]; |
| 302 | break; |
| 303 | } |
| 304 | ctr[j] = 0; |
| 305 | } |
| 306 | } |
| 307 | return this; |
| 308 | } |
| 309 | |
| 310 | Benchmark* Benchmark::ArgName(const std::string& name) { |
| 311 | CHECK(ArgsCnt() == -1 || ArgsCnt() == 1); |
| 312 | arg_names_ = {name}; |
| 313 | return this; |
| 314 | } |
| 315 | |
| 316 | Benchmark* Benchmark::ArgNames(const std::vector<std::string>& names) { |
| 317 | CHECK(ArgsCnt() == -1 || ArgsCnt() == static_cast<int>(names.size())); |
| 318 | arg_names_ = names; |
| 319 | return this; |
| 320 | } |
| 321 | |
| 322 | Benchmark* Benchmark::DenseRange(int start, int limit, int step) { |
| 323 | CHECK(ArgsCnt() == -1 || ArgsCnt() == 1); |
| 324 | CHECK_GE(start, 0); |
| 325 | CHECK_LE(start, limit); |
| 326 | for (int arg = start; arg <= limit; arg += step) { |
| 327 | args_.push_back({arg}); |
| 328 | } |
| 329 | return this; |
| 330 | } |
| 331 | |
| 332 | Benchmark* Benchmark::Args(const std::vector<int>& args) { |
| 333 | CHECK(ArgsCnt() == -1 || ArgsCnt() == static_cast<int>(args.size())); |
| 334 | args_.push_back(args); |
| 335 | return this; |
| 336 | } |
| 337 | |
| 338 | Benchmark* Benchmark::Apply(void (*custom_arguments)(Benchmark* benchmark)) { |
| 339 | custom_arguments(this); |
| 340 | return this; |
| 341 | } |
| 342 | |
| 343 | Benchmark* Benchmark::RangeMultiplier(int multiplier) { |
| 344 | CHECK(multiplier > 1); |
| 345 | range_multiplier_ = multiplier; |
| 346 | return this; |
| 347 | } |
| 348 | |
| 349 | Benchmark* Benchmark::Repetitions(int n) { |
| 350 | CHECK(n > 0); |
| 351 | repetitions_ = n; |
| 352 | return this; |
| 353 | } |
| 354 | |
| 355 | Benchmark* Benchmark::ReportAggregatesOnly(bool value) { |
| 356 | report_mode_ = value ? RM_ReportAggregatesOnly : RM_Default; |
| 357 | return this; |
| 358 | } |
| 359 | |
| 360 | Benchmark* Benchmark::MinTime(double t) { |
| 361 | CHECK(t > 0.0); |
| 362 | min_time_ = t; |
| 363 | return this; |
| 364 | } |
| 365 | |
| 366 | Benchmark* Benchmark::UseRealTime() { |
| 367 | CHECK(!use_manual_time_) |
| 368 | << "Cannot set UseRealTime and UseManualTime simultaneously." ; |
| 369 | use_real_time_ = true; |
| 370 | return this; |
| 371 | } |
| 372 | |
| 373 | Benchmark* Benchmark::UseManualTime() { |
| 374 | CHECK(!use_real_time_) |
| 375 | << "Cannot set UseRealTime and UseManualTime simultaneously." ; |
| 376 | use_manual_time_ = true; |
| 377 | return this; |
| 378 | } |
| 379 | |
| 380 | Benchmark* Benchmark::Complexity(BigO complexity) { |
| 381 | complexity_ = complexity; |
| 382 | return this; |
| 383 | } |
| 384 | |
| 385 | Benchmark* Benchmark::Complexity(BigOFunc* complexity) { |
| 386 | complexity_lambda_ = complexity; |
| 387 | complexity_ = oLambda; |
| 388 | return this; |
| 389 | } |
| 390 | |
| 391 | Benchmark* Benchmark::Threads(int t) { |
| 392 | CHECK_GT(t, 0); |
| 393 | thread_counts_.push_back(t); |
| 394 | return this; |
| 395 | } |
| 396 | |
| 397 | Benchmark* Benchmark::ThreadRange(int min_threads, int max_threads) { |
| 398 | CHECK_GT(min_threads, 0); |
| 399 | CHECK_GE(max_threads, min_threads); |
| 400 | |
| 401 | AddRange(&thread_counts_, min_threads, max_threads, 2); |
| 402 | return this; |
| 403 | } |
| 404 | |
| 405 | Benchmark* Benchmark::DenseThreadRange(int min_threads, int max_threads, |
| 406 | int stride) { |
| 407 | CHECK_GT(min_threads, 0); |
| 408 | CHECK_GE(max_threads, min_threads); |
| 409 | CHECK_GE(stride, 1); |
| 410 | |
| 411 | for (auto i = min_threads; i < max_threads; i += stride) { |
| 412 | thread_counts_.push_back(i); |
| 413 | } |
| 414 | thread_counts_.push_back(max_threads); |
| 415 | return this; |
| 416 | } |
| 417 | |
| 418 | Benchmark* Benchmark::ThreadPerCpu() { |
| 419 | static int num_cpus = NumCPUs(); |
| 420 | thread_counts_.push_back(num_cpus); |
| 421 | return this; |
| 422 | } |
| 423 | |
| 424 | void Benchmark::SetName(const char* name) { name_ = name; } |
| 425 | |
| 426 | int Benchmark::ArgsCnt() const { |
| 427 | if (args_.empty()) { |
| 428 | if (arg_names_.empty()) return -1; |
| 429 | return static_cast<int>(arg_names_.size()); |
| 430 | } |
| 431 | return static_cast<int>(args_.front().size()); |
| 432 | } |
| 433 | |
| 434 | //=============================================================================// |
| 435 | // FunctionBenchmark |
| 436 | //=============================================================================// |
| 437 | |
| 438 | void FunctionBenchmark::Run(State& st) { func_(st); } |
| 439 | |
| 440 | } // end namespace internal |
| 441 | } // end namespace benchmark |
| 442 | |