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 | #define __STDC_FORMAT_MACROS |
38 | #include <inttypes.h> |
39 | |
40 | #include "benchmark/benchmark.h" |
41 | #include "benchmark_api_internal.h" |
42 | #include "check.h" |
43 | #include "commandlineflags.h" |
44 | #include "complexity.h" |
45 | #include "internal_macros.h" |
46 | #include "log.h" |
47 | #include "mutex.h" |
48 | #include "re.h" |
49 | #include "statistics.h" |
50 | #include "string_util.h" |
51 | #include "timers.h" |
52 | |
53 | namespace benchmark { |
54 | |
55 | namespace { |
56 | // For non-dense Range, intermediate values are powers of kRangeMultiplier. |
57 | static const int kRangeMultiplier = 8; |
58 | // The size of a benchmark family determines is the number of inputs to repeat |
59 | // the benchmark on. If this is "large" then warn the user during configuration. |
60 | static const size_t kMaxFamilySize = 100; |
61 | } // end namespace |
62 | |
63 | namespace internal { |
64 | |
65 | //=============================================================================// |
66 | // BenchmarkFamilies |
67 | //=============================================================================// |
68 | |
69 | // Class for managing registered benchmarks. Note that each registered |
70 | // benchmark identifies a family of related benchmarks to run. |
71 | class BenchmarkFamilies { |
72 | public: |
73 | static BenchmarkFamilies* GetInstance(); |
74 | |
75 | // Registers a benchmark family and returns the index assigned to it. |
76 | size_t AddBenchmark(std::unique_ptr<Benchmark> family); |
77 | |
78 | // Clear all registered benchmark families. |
79 | void ClearBenchmarks(); |
80 | |
81 | // Extract the list of benchmark instances that match the specified |
82 | // regular expression. |
83 | bool FindBenchmarks(std::string re, |
84 | std::vector<BenchmarkInstance>* benchmarks, |
85 | std::ostream* Err); |
86 | |
87 | private: |
88 | BenchmarkFamilies() {} |
89 | |
90 | std::vector<std::unique_ptr<Benchmark>> families_; |
91 | Mutex mutex_; |
92 | }; |
93 | |
94 | BenchmarkFamilies* BenchmarkFamilies::GetInstance() { |
95 | static BenchmarkFamilies instance; |
96 | return &instance; |
97 | } |
98 | |
99 | size_t BenchmarkFamilies::AddBenchmark(std::unique_ptr<Benchmark> family) { |
100 | MutexLock l(mutex_); |
101 | size_t index = families_.size(); |
102 | families_.push_back(std::move(family)); |
103 | return index; |
104 | } |
105 | |
106 | void BenchmarkFamilies::ClearBenchmarks() { |
107 | MutexLock l(mutex_); |
108 | families_.clear(); |
109 | families_.shrink_to_fit(); |
110 | } |
111 | |
112 | bool BenchmarkFamilies::FindBenchmarks( |
113 | std::string spec, std::vector<BenchmarkInstance>* benchmarks, |
114 | std::ostream* ErrStream) { |
115 | CHECK(ErrStream); |
116 | auto& Err = *ErrStream; |
117 | // Make regular expression out of command-line flag |
118 | std::string error_msg; |
119 | Regex re; |
120 | bool isNegativeFilter = false; |
121 | if (spec[0] == '-') { |
122 | spec.replace(0, 1, "" ); |
123 | isNegativeFilter = true; |
124 | } |
125 | if (!re.Init(spec, &error_msg)) { |
126 | Err << "Could not compile benchmark re: " << error_msg << std::endl; |
127 | return false; |
128 | } |
129 | |
130 | // Special list of thread counts to use when none are specified |
131 | const std::vector<int> one_thread = {1}; |
132 | |
133 | MutexLock l(mutex_); |
134 | for (std::unique_ptr<Benchmark>& family : families_) { |
135 | // Family was deleted or benchmark doesn't match |
136 | if (!family) continue; |
137 | |
138 | if (family->ArgsCnt() == -1) { |
139 | family->Args({}); |
140 | } |
141 | const std::vector<int>* thread_counts = |
142 | (family->thread_counts_.empty() |
143 | ? &one_thread |
144 | : &static_cast<const std::vector<int>&>(family->thread_counts_)); |
145 | const size_t family_size = family->args_.size() * thread_counts->size(); |
146 | // The benchmark will be run at least 'family_size' different inputs. |
147 | // If 'family_size' is very large warn the user. |
148 | if (family_size > kMaxFamilySize) { |
149 | Err << "The number of inputs is very large. " << family->name_ |
150 | << " will be repeated at least " << family_size << " times.\n" ; |
151 | } |
152 | // reserve in the special case the regex ".", since we know the final |
153 | // family size. |
154 | if (spec == "." ) benchmarks->reserve(family_size); |
155 | |
156 | for (auto const& args : family->args_) { |
157 | for (int num_threads : *thread_counts) { |
158 | BenchmarkInstance instance; |
159 | instance.name.function_name = family->name_; |
160 | instance.benchmark = family.get(); |
161 | instance.aggregation_report_mode = family->aggregation_report_mode_; |
162 | instance.arg = args; |
163 | instance.time_unit = family->time_unit_; |
164 | instance.range_multiplier = family->range_multiplier_; |
165 | instance.min_time = family->min_time_; |
166 | instance.iterations = family->iterations_; |
167 | instance.repetitions = family->repetitions_; |
168 | instance.measure_process_cpu_time = family->measure_process_cpu_time_; |
169 | instance.use_real_time = family->use_real_time_; |
170 | instance.use_manual_time = family->use_manual_time_; |
171 | instance.complexity = family->complexity_; |
172 | instance.complexity_lambda = family->complexity_lambda_; |
173 | instance.statistics = &family->statistics_; |
174 | instance.threads = num_threads; |
175 | |
176 | // Add arguments to instance name |
177 | size_t arg_i = 0; |
178 | for (auto const& arg : args) { |
179 | if (!instance.name.args.empty()) { |
180 | instance.name.args += '/'; |
181 | } |
182 | |
183 | if (arg_i < family->arg_names_.size()) { |
184 | const auto& arg_name = family->arg_names_[arg_i]; |
185 | if (!arg_name.empty()) { |
186 | instance.name.args += StrFormat("%s:" , arg_name.c_str()); |
187 | } |
188 | } |
189 | |
190 | instance.name.args += StrFormat("%" PRId64, arg); |
191 | ++arg_i; |
192 | } |
193 | |
194 | if (!IsZero(family->min_time_)) |
195 | instance.name.min_time = |
196 | StrFormat("min_time:%0.3f" , family->min_time_); |
197 | if (family->iterations_ != 0) { |
198 | instance.name.iterations = |
199 | StrFormat("iterations:%lu" , |
200 | static_cast<unsigned long>(family->iterations_)); |
201 | } |
202 | if (family->repetitions_ != 0) |
203 | instance.name.repetitions = |
204 | StrFormat("repeats:%d" , family->repetitions_); |
205 | |
206 | if (family->measure_process_cpu_time_) { |
207 | instance.name.time_type = "process_time" ; |
208 | } |
209 | |
210 | if (family->use_manual_time_) { |
211 | if (!instance.name.time_type.empty()) { |
212 | instance.name.time_type += '/'; |
213 | } |
214 | instance.name.time_type += "manual_time" ; |
215 | } else if (family->use_real_time_) { |
216 | if (!instance.name.time_type.empty()) { |
217 | instance.name.time_type += '/'; |
218 | } |
219 | instance.name.time_type += "real_time" ; |
220 | } |
221 | |
222 | // Add the number of threads used to the name |
223 | if (!family->thread_counts_.empty()) { |
224 | instance.name.threads = StrFormat("threads:%d" , instance.threads); |
225 | } |
226 | |
227 | const auto full_name = instance.name.str(); |
228 | if ((re.Match(full_name) && !isNegativeFilter) || |
229 | (!re.Match(full_name) && isNegativeFilter)) { |
230 | instance.last_benchmark_instance = (&args == &family->args_.back()); |
231 | benchmarks->push_back(std::move(instance)); |
232 | } |
233 | } |
234 | } |
235 | } |
236 | return true; |
237 | } |
238 | |
239 | Benchmark* RegisterBenchmarkInternal(Benchmark* bench) { |
240 | std::unique_ptr<Benchmark> bench_ptr(bench); |
241 | BenchmarkFamilies* families = BenchmarkFamilies::GetInstance(); |
242 | families->AddBenchmark(std::move(bench_ptr)); |
243 | return bench; |
244 | } |
245 | |
246 | // FIXME: This function is a hack so that benchmark.cc can access |
247 | // `BenchmarkFamilies` |
248 | bool FindBenchmarksInternal(const std::string& re, |
249 | std::vector<BenchmarkInstance>* benchmarks, |
250 | std::ostream* Err) { |
251 | return BenchmarkFamilies::GetInstance()->FindBenchmarks(re, benchmarks, Err); |
252 | } |
253 | |
254 | //=============================================================================// |
255 | // Benchmark |
256 | //=============================================================================// |
257 | |
258 | Benchmark::Benchmark(const char* name) |
259 | : name_(name), |
260 | aggregation_report_mode_(ARM_Unspecified), |
261 | time_unit_(kNanosecond), |
262 | range_multiplier_(kRangeMultiplier), |
263 | min_time_(0), |
264 | iterations_(0), |
265 | repetitions_(0), |
266 | measure_process_cpu_time_(false), |
267 | use_real_time_(false), |
268 | use_manual_time_(false), |
269 | complexity_(oNone), |
270 | complexity_lambda_(nullptr) { |
271 | ComputeStatistics("mean" , StatisticsMean); |
272 | ComputeStatistics("median" , StatisticsMedian); |
273 | ComputeStatistics("stddev" , StatisticsStdDev); |
274 | } |
275 | |
276 | Benchmark::~Benchmark() {} |
277 | |
278 | Benchmark* Benchmark::Arg(int64_t x) { |
279 | CHECK(ArgsCnt() == -1 || ArgsCnt() == 1); |
280 | args_.push_back({x}); |
281 | return this; |
282 | } |
283 | |
284 | Benchmark* Benchmark::Unit(TimeUnit unit) { |
285 | time_unit_ = unit; |
286 | return this; |
287 | } |
288 | |
289 | Benchmark* Benchmark::Range(int64_t start, int64_t limit) { |
290 | CHECK(ArgsCnt() == -1 || ArgsCnt() == 1); |
291 | std::vector<int64_t> arglist; |
292 | AddRange(&arglist, start, limit, range_multiplier_); |
293 | |
294 | for (int64_t i : arglist) { |
295 | args_.push_back({i}); |
296 | } |
297 | return this; |
298 | } |
299 | |
300 | Benchmark* Benchmark::Ranges( |
301 | const std::vector<std::pair<int64_t, int64_t>>& ranges) { |
302 | CHECK(ArgsCnt() == -1 || ArgsCnt() == static_cast<int>(ranges.size())); |
303 | std::vector<std::vector<int64_t>> arglists(ranges.size()); |
304 | std::size_t total = 1; |
305 | for (std::size_t i = 0; i < ranges.size(); i++) { |
306 | AddRange(&arglists[i], ranges[i].first, ranges[i].second, |
307 | range_multiplier_); |
308 | total *= arglists[i].size(); |
309 | } |
310 | |
311 | std::vector<std::size_t> ctr(arglists.size(), 0); |
312 | |
313 | for (std::size_t i = 0; i < total; i++) { |
314 | std::vector<int64_t> tmp; |
315 | tmp.reserve(arglists.size()); |
316 | |
317 | for (std::size_t j = 0; j < arglists.size(); j++) { |
318 | tmp.push_back(arglists[j].at(ctr[j])); |
319 | } |
320 | |
321 | args_.push_back(std::move(tmp)); |
322 | |
323 | for (std::size_t j = 0; j < arglists.size(); j++) { |
324 | if (ctr[j] + 1 < arglists[j].size()) { |
325 | ++ctr[j]; |
326 | break; |
327 | } |
328 | ctr[j] = 0; |
329 | } |
330 | } |
331 | return this; |
332 | } |
333 | |
334 | Benchmark* Benchmark::ArgName(const std::string& name) { |
335 | CHECK(ArgsCnt() == -1 || ArgsCnt() == 1); |
336 | arg_names_ = {name}; |
337 | return this; |
338 | } |
339 | |
340 | Benchmark* Benchmark::ArgNames(const std::vector<std::string>& names) { |
341 | CHECK(ArgsCnt() == -1 || ArgsCnt() == static_cast<int>(names.size())); |
342 | arg_names_ = names; |
343 | return this; |
344 | } |
345 | |
346 | Benchmark* Benchmark::DenseRange(int64_t start, int64_t limit, int step) { |
347 | CHECK(ArgsCnt() == -1 || ArgsCnt() == 1); |
348 | CHECK_LE(start, limit); |
349 | for (int64_t arg = start; arg <= limit; arg += step) { |
350 | args_.push_back({arg}); |
351 | } |
352 | return this; |
353 | } |
354 | |
355 | Benchmark* Benchmark::Args(const std::vector<int64_t>& args) { |
356 | CHECK(ArgsCnt() == -1 || ArgsCnt() == static_cast<int>(args.size())); |
357 | args_.push_back(args); |
358 | return this; |
359 | } |
360 | |
361 | Benchmark* Benchmark::Apply(void (*custom_arguments)(Benchmark* benchmark)) { |
362 | custom_arguments(this); |
363 | return this; |
364 | } |
365 | |
366 | Benchmark* Benchmark::RangeMultiplier(int multiplier) { |
367 | CHECK(multiplier > 1); |
368 | range_multiplier_ = multiplier; |
369 | return this; |
370 | } |
371 | |
372 | Benchmark* Benchmark::MinTime(double t) { |
373 | CHECK(t > 0.0); |
374 | CHECK(iterations_ == 0); |
375 | min_time_ = t; |
376 | return this; |
377 | } |
378 | |
379 | Benchmark* Benchmark::Iterations(IterationCount n) { |
380 | CHECK(n > 0); |
381 | CHECK(IsZero(min_time_)); |
382 | iterations_ = n; |
383 | return this; |
384 | } |
385 | |
386 | Benchmark* Benchmark::Repetitions(int n) { |
387 | CHECK(n > 0); |
388 | repetitions_ = n; |
389 | return this; |
390 | } |
391 | |
392 | Benchmark* Benchmark::ReportAggregatesOnly(bool value) { |
393 | aggregation_report_mode_ = value ? ARM_ReportAggregatesOnly : ARM_Default; |
394 | return this; |
395 | } |
396 | |
397 | Benchmark* Benchmark::DisplayAggregatesOnly(bool value) { |
398 | // If we were called, the report mode is no longer 'unspecified', in any case. |
399 | aggregation_report_mode_ = static_cast<AggregationReportMode>( |
400 | aggregation_report_mode_ | ARM_Default); |
401 | |
402 | if (value) { |
403 | aggregation_report_mode_ = static_cast<AggregationReportMode>( |
404 | aggregation_report_mode_ | ARM_DisplayReportAggregatesOnly); |
405 | } else { |
406 | aggregation_report_mode_ = static_cast<AggregationReportMode>( |
407 | aggregation_report_mode_ & ~ARM_DisplayReportAggregatesOnly); |
408 | } |
409 | |
410 | return this; |
411 | } |
412 | |
413 | Benchmark* Benchmark::MeasureProcessCPUTime() { |
414 | // Can be used together with UseRealTime() / UseManualTime(). |
415 | measure_process_cpu_time_ = true; |
416 | return this; |
417 | } |
418 | |
419 | Benchmark* Benchmark::UseRealTime() { |
420 | CHECK(!use_manual_time_) |
421 | << "Cannot set UseRealTime and UseManualTime simultaneously." ; |
422 | use_real_time_ = true; |
423 | return this; |
424 | } |
425 | |
426 | Benchmark* Benchmark::UseManualTime() { |
427 | CHECK(!use_real_time_) |
428 | << "Cannot set UseRealTime and UseManualTime simultaneously." ; |
429 | use_manual_time_ = true; |
430 | return this; |
431 | } |
432 | |
433 | Benchmark* Benchmark::Complexity(BigO complexity) { |
434 | complexity_ = complexity; |
435 | return this; |
436 | } |
437 | |
438 | Benchmark* Benchmark::Complexity(BigOFunc* complexity) { |
439 | complexity_lambda_ = complexity; |
440 | complexity_ = oLambda; |
441 | return this; |
442 | } |
443 | |
444 | Benchmark* Benchmark::ComputeStatistics(std::string name, |
445 | StatisticsFunc* statistics) { |
446 | statistics_.emplace_back(name, statistics); |
447 | return this; |
448 | } |
449 | |
450 | Benchmark* Benchmark::Threads(int t) { |
451 | CHECK_GT(t, 0); |
452 | thread_counts_.push_back(t); |
453 | return this; |
454 | } |
455 | |
456 | Benchmark* Benchmark::ThreadRange(int min_threads, int max_threads) { |
457 | CHECK_GT(min_threads, 0); |
458 | CHECK_GE(max_threads, min_threads); |
459 | |
460 | AddRange(&thread_counts_, min_threads, max_threads, 2); |
461 | return this; |
462 | } |
463 | |
464 | Benchmark* Benchmark::DenseThreadRange(int min_threads, int max_threads, |
465 | int stride) { |
466 | CHECK_GT(min_threads, 0); |
467 | CHECK_GE(max_threads, min_threads); |
468 | CHECK_GE(stride, 1); |
469 | |
470 | for (auto i = min_threads; i < max_threads; i += stride) { |
471 | thread_counts_.push_back(i); |
472 | } |
473 | thread_counts_.push_back(max_threads); |
474 | return this; |
475 | } |
476 | |
477 | Benchmark* Benchmark::ThreadPerCpu() { |
478 | thread_counts_.push_back(CPUInfo::Get().num_cpus); |
479 | return this; |
480 | } |
481 | |
482 | void Benchmark::SetName(const char* name) { name_ = name; } |
483 | |
484 | int Benchmark::ArgsCnt() const { |
485 | if (args_.empty()) { |
486 | if (arg_names_.empty()) return -1; |
487 | return static_cast<int>(arg_names_.size()); |
488 | } |
489 | return static_cast<int>(args_.front().size()); |
490 | } |
491 | |
492 | //=============================================================================// |
493 | // FunctionBenchmark |
494 | //=============================================================================// |
495 | |
496 | void FunctionBenchmark::Run(State& st) { func_(st); } |
497 | |
498 | } // end namespace internal |
499 | |
500 | void ClearRegisteredBenchmarks() { |
501 | internal::BenchmarkFamilies::GetInstance()->ClearBenchmarks(); |
502 | } |
503 | |
504 | } // end namespace benchmark |
505 | |