1 | // Support for registering benchmarks for functions. |
2 | |
3 | /* Example usage: |
4 | // Define a function that executes the code to be measured a |
5 | // specified number of times: |
6 | static void BM_StringCreation(benchmark::State& state) { |
7 | while (state.KeepRunning()) |
8 | std::string empty_string; |
9 | } |
10 | |
11 | // Register the function as a benchmark |
12 | BENCHMARK(BM_StringCreation); |
13 | |
14 | // Define another benchmark |
15 | static void BM_StringCopy(benchmark::State& state) { |
16 | std::string x = "hello"; |
17 | while (state.KeepRunning()) |
18 | std::string copy(x); |
19 | } |
20 | BENCHMARK(BM_StringCopy); |
21 | |
22 | // Augment the main() program to invoke benchmarks if specified |
23 | // via the --benchmarks command line flag. E.g., |
24 | // my_unittest --benchmark_filter=all |
25 | // my_unittest --benchmark_filter=BM_StringCreation |
26 | // my_unittest --benchmark_filter=String |
27 | // my_unittest --benchmark_filter='Copy|Creation' |
28 | int main(int argc, char** argv) { |
29 | benchmark::Initialize(&argc, argv); |
30 | benchmark::RunSpecifiedBenchmarks(); |
31 | return 0; |
32 | } |
33 | |
34 | // Sometimes a family of microbenchmarks can be implemented with |
35 | // just one routine that takes an extra argument to specify which |
36 | // one of the family of benchmarks to run. For example, the following |
37 | // code defines a family of microbenchmarks for measuring the speed |
38 | // of memcpy() calls of different lengths: |
39 | |
40 | static void BM_memcpy(benchmark::State& state) { |
41 | char* src = new char[state.range(0)]; char* dst = new char[state.range(0)]; |
42 | memset(src, 'x', state.range(0)); |
43 | while (state.KeepRunning()) |
44 | memcpy(dst, src, state.range(0)); |
45 | state.SetBytesProcessed(int64_t(state.iterations()) * |
46 | int64_t(state.range(0))); |
47 | delete[] src; delete[] dst; |
48 | } |
49 | BENCHMARK(BM_memcpy)->Arg(8)->Arg(64)->Arg(512)->Arg(1<<10)->Arg(8<<10); |
50 | |
51 | // The preceding code is quite repetitive, and can be replaced with the |
52 | // following short-hand. The following invocation will pick a few |
53 | // appropriate arguments in the specified range and will generate a |
54 | // microbenchmark for each such argument. |
55 | BENCHMARK(BM_memcpy)->Range(8, 8<<10); |
56 | |
57 | // You might have a microbenchmark that depends on two inputs. For |
58 | // example, the following code defines a family of microbenchmarks for |
59 | // measuring the speed of set insertion. |
60 | static void BM_SetInsert(benchmark::State& state) { |
61 | while (state.KeepRunning()) { |
62 | state.PauseTiming(); |
63 | set<int> data = ConstructRandomSet(state.range(0)); |
64 | state.ResumeTiming(); |
65 | for (int j = 0; j < state.range(1); ++j) |
66 | data.insert(RandomNumber()); |
67 | } |
68 | } |
69 | BENCHMARK(BM_SetInsert) |
70 | ->Args({1<<10, 1}) |
71 | ->Args({1<<10, 8}) |
72 | ->Args({1<<10, 64}) |
73 | ->Args({1<<10, 512}) |
74 | ->Args({8<<10, 1}) |
75 | ->Args({8<<10, 8}) |
76 | ->Args({8<<10, 64}) |
77 | ->Args({8<<10, 512}); |
78 | |
79 | // The preceding code is quite repetitive, and can be replaced with |
80 | // the following short-hand. The following macro will pick a few |
81 | // appropriate arguments in the product of the two specified ranges |
82 | // and will generate a microbenchmark for each such pair. |
83 | BENCHMARK(BM_SetInsert)->Ranges({{1<<10, 8<<10}, {1, 512}}); |
84 | |
85 | // For more complex patterns of inputs, passing a custom function |
86 | // to Apply allows programmatic specification of an |
87 | // arbitrary set of arguments to run the microbenchmark on. |
88 | // The following example enumerates a dense range on |
89 | // one parameter, and a sparse range on the second. |
90 | static void CustomArguments(benchmark::internal::Benchmark* b) { |
91 | for (int i = 0; i <= 10; ++i) |
92 | for (int j = 32; j <= 1024*1024; j *= 8) |
93 | b->Args({i, j}); |
94 | } |
95 | BENCHMARK(BM_SetInsert)->Apply(CustomArguments); |
96 | |
97 | // Templated microbenchmarks work the same way: |
98 | // Produce then consume 'size' messages 'iters' times |
99 | // Measures throughput in the absence of multiprogramming. |
100 | template <class Q> int BM_Sequential(benchmark::State& state) { |
101 | Q q; |
102 | typename Q::value_type v; |
103 | while (state.KeepRunning()) { |
104 | for (int i = state.range(0); i--; ) |
105 | q.push(v); |
106 | for (int e = state.range(0); e--; ) |
107 | q.Wait(&v); |
108 | } |
109 | // actually messages, not bytes: |
110 | state.SetBytesProcessed( |
111 | static_cast<int64_t>(state.iterations())*state.range(0)); |
112 | } |
113 | BENCHMARK_TEMPLATE(BM_Sequential, WaitQueue<int>)->Range(1<<0, 1<<10); |
114 | |
115 | Use `Benchmark::MinTime(double t)` to set the minimum time used to run the |
116 | benchmark. This option overrides the `benchmark_min_time` flag. |
117 | |
118 | void BM_test(benchmark::State& state) { |
119 | ... body ... |
120 | } |
121 | BENCHMARK(BM_test)->MinTime(2.0); // Run for at least 2 seconds. |
122 | |
123 | In a multithreaded test, it is guaranteed that none of the threads will start |
124 | until all have called KeepRunning, and all will have finished before KeepRunning |
125 | returns false. As such, any global setup or teardown you want to do can be |
126 | wrapped in a check against the thread index: |
127 | |
128 | static void BM_MultiThreaded(benchmark::State& state) { |
129 | if (state.thread_index == 0) { |
130 | // Setup code here. |
131 | } |
132 | while (state.KeepRunning()) { |
133 | // Run the test as normal. |
134 | } |
135 | if (state.thread_index == 0) { |
136 | // Teardown code here. |
137 | } |
138 | } |
139 | BENCHMARK(BM_MultiThreaded)->Threads(4); |
140 | |
141 | |
142 | If a benchmark runs a few milliseconds it may be hard to visually compare the |
143 | measured times, since the output data is given in nanoseconds per default. In |
144 | order to manually set the time unit, you can specify it manually: |
145 | |
146 | BENCHMARK(BM_test)->Unit(benchmark::kMillisecond); |
147 | */ |
148 | |
149 | #ifndef BENCHMARK_BENCHMARK_API_H_ |
150 | #define BENCHMARK_BENCHMARK_API_H_ |
151 | |
152 | #include <assert.h> |
153 | #include <stddef.h> |
154 | #include <stdint.h> |
155 | |
156 | #include <string> |
157 | #include <vector> |
158 | |
159 | #include "macros.h" |
160 | |
161 | #if defined(BENCHMARK_HAS_CXX11) |
162 | #include <type_traits> |
163 | #include <utility> |
164 | #endif |
165 | |
166 | namespace benchmark { |
167 | class BenchmarkReporter; |
168 | |
169 | void Initialize(int* argc, char** argv); |
170 | |
171 | // Generate a list of benchmarks matching the specified --benchmark_filter flag |
172 | // and if --benchmark_list_tests is specified return after printing the name |
173 | // of each matching benchmark. Otherwise run each matching benchmark and |
174 | // report the results. |
175 | // |
176 | // The second and third overload use the specified 'console_reporter' and |
177 | // 'file_reporter' respectively. 'file_reporter' will write to the file |
178 | // specified |
179 | // by '--benchmark_output'. If '--benchmark_output' is not given the |
180 | // 'file_reporter' is ignored. |
181 | // |
182 | // RETURNS: The number of matching benchmarks. |
183 | size_t RunSpecifiedBenchmarks(); |
184 | size_t RunSpecifiedBenchmarks(BenchmarkReporter* console_reporter); |
185 | size_t RunSpecifiedBenchmarks(BenchmarkReporter* console_reporter, |
186 | BenchmarkReporter* file_reporter); |
187 | |
188 | // If this routine is called, peak memory allocation past this point in the |
189 | // benchmark is reported at the end of the benchmark report line. (It is |
190 | // computed by running the benchmark once with a single iteration and a memory |
191 | // tracer.) |
192 | // TODO(dominic) |
193 | // void MemoryUsage(); |
194 | |
195 | namespace internal { |
196 | class Benchmark; |
197 | class BenchmarkImp; |
198 | class BenchmarkFamilies; |
199 | |
200 | template <class T> |
201 | struct Voider { |
202 | typedef void type; |
203 | }; |
204 | |
205 | template <class T, class = void> |
206 | struct EnableIfString {}; |
207 | |
208 | template <class T> |
209 | struct EnableIfString<T, typename Voider<typename T::basic_string>::type> { |
210 | typedef int type; |
211 | }; |
212 | |
213 | void UseCharPointer(char const volatile*); |
214 | |
215 | // Take ownership of the pointer and register the benchmark. Return the |
216 | // registered benchmark. |
217 | Benchmark* RegisterBenchmarkInternal(Benchmark*); |
218 | |
219 | // Ensure that the standard streams are properly initialized in every TU. |
220 | int InitializeStreams(); |
221 | BENCHMARK_UNUSED static int stream_init_anchor = InitializeStreams(); |
222 | |
223 | } // end namespace internal |
224 | |
225 | // The DoNotOptimize(...) function can be used to prevent a value or |
226 | // expression from being optimized away by the compiler. This function is |
227 | // intended to add little to no overhead. |
228 | // See: https://youtu.be/nXaxk27zwlk?t=2441 |
229 | #if defined(__GNUC__) |
230 | template <class Tp> |
231 | inline BENCHMARK_ALWAYS_INLINE void DoNotOptimize(Tp const& value) { |
232 | asm volatile("" : : "g" (value) : "memory" ); |
233 | } |
234 | // Force the compiler to flush pending writes to global memory. Acts as an |
235 | // effective read/write barrier |
236 | inline BENCHMARK_ALWAYS_INLINE void ClobberMemory() { |
237 | asm volatile("" : : : "memory" ); |
238 | } |
239 | #else |
240 | template <class Tp> |
241 | inline BENCHMARK_ALWAYS_INLINE void DoNotOptimize(Tp const& value) { |
242 | internal::UseCharPointer(&reinterpret_cast<char const volatile&>(value)); |
243 | } |
244 | // FIXME Add ClobberMemory() for non-gnu compilers |
245 | #endif |
246 | |
247 | // TimeUnit is passed to a benchmark in order to specify the order of magnitude |
248 | // for the measured time. |
249 | enum TimeUnit { kNanosecond, kMicrosecond, kMillisecond }; |
250 | |
251 | // BigO is passed to a benchmark in order to specify the asymptotic |
252 | // computational |
253 | // complexity for the benchmark. In case oAuto is selected, complexity will be |
254 | // calculated automatically to the best fit. |
255 | enum BigO { oNone, o1, oN, oNSquared, oNCubed, oLogN, oNLogN, oAuto, oLambda }; |
256 | |
257 | // BigOFunc is passed to a benchmark in order to specify the asymptotic |
258 | // computational complexity for the benchmark. |
259 | typedef double(BigOFunc)(int); |
260 | |
261 | namespace internal { |
262 | class ThreadTimer; |
263 | class ThreadManager; |
264 | |
265 | #if defined(BENCHMARK_HAS_CXX11) |
266 | enum ReportMode : unsigned { |
267 | #else |
268 | enum ReportMode { |
269 | #endif |
270 | RM_Unspecified, // The mode has not been manually specified |
271 | RM_Default, // The mode is user-specified as default. |
272 | RM_ReportAggregatesOnly |
273 | }; |
274 | } |
275 | |
276 | // State is passed to a running Benchmark and contains state for the |
277 | // benchmark to use. |
278 | class State { |
279 | public: |
280 | // Returns true if the benchmark should continue through another iteration. |
281 | // NOTE: A benchmark may not return from the test until KeepRunning() has |
282 | // returned false. |
283 | bool KeepRunning() { |
284 | if (BENCHMARK_BUILTIN_EXPECT(!started_, false)) { |
285 | StartKeepRunning(); |
286 | } |
287 | bool const res = total_iterations_++ < max_iterations; |
288 | if (BENCHMARK_BUILTIN_EXPECT(!res, false)) { |
289 | FinishKeepRunning(); |
290 | } |
291 | return res; |
292 | } |
293 | |
294 | // REQUIRES: timer is running and 'SkipWithError(...)' has not been called |
295 | // by the current thread. |
296 | // Stop the benchmark timer. If not called, the timer will be |
297 | // automatically stopped after KeepRunning() returns false for the first time. |
298 | // |
299 | // For threaded benchmarks the PauseTiming() function only pauses the timing |
300 | // for the current thread. |
301 | // |
302 | // NOTE: The "real time" measurement is per-thread. If different threads |
303 | // report different measurements the largest one is reported. |
304 | // |
305 | // NOTE: PauseTiming()/ResumeTiming() are relatively |
306 | // heavyweight, and so their use should generally be avoided |
307 | // within each benchmark iteration, if possible. |
308 | void PauseTiming(); |
309 | |
310 | // REQUIRES: timer is not running and 'SkipWithError(...)' has not been called |
311 | // by the current thread. |
312 | // Start the benchmark timer. The timer is NOT running on entrance to the |
313 | // benchmark function. It begins running after the first call to KeepRunning() |
314 | // |
315 | // NOTE: PauseTiming()/ResumeTiming() are relatively |
316 | // heavyweight, and so their use should generally be avoided |
317 | // within each benchmark iteration, if possible. |
318 | void ResumeTiming(); |
319 | |
320 | // REQUIRES: 'SkipWithError(...)' has not been called previously by the |
321 | // current thread. |
322 | // Skip any future iterations of the 'KeepRunning()' loop in the current |
323 | // thread and report an error with the specified 'msg'. After this call |
324 | // the user may explicitly 'return' from the benchmark. |
325 | // |
326 | // For threaded benchmarks only the current thread stops executing and future |
327 | // calls to `KeepRunning()` will block until all threads have completed |
328 | // the `KeepRunning()` loop. If multiple threads report an error only the |
329 | // first error message is used. |
330 | // |
331 | // NOTE: Calling 'SkipWithError(...)' does not cause the benchmark to exit |
332 | // the current scope immediately. If the function is called from within |
333 | // the 'KeepRunning()' loop the current iteration will finish. It is the users |
334 | // responsibility to exit the scope as needed. |
335 | void SkipWithError(const char* msg); |
336 | |
337 | // REQUIRES: called exactly once per iteration of the KeepRunning loop. |
338 | // Set the manually measured time for this benchmark iteration, which |
339 | // is used instead of automatically measured time if UseManualTime() was |
340 | // specified. |
341 | // |
342 | // For threaded benchmarks the final value will be set to the largest |
343 | // reported values. |
344 | void SetIterationTime(double seconds); |
345 | |
346 | // Set the number of bytes processed by the current benchmark |
347 | // execution. This routine is typically called once at the end of a |
348 | // throughput oriented benchmark. If this routine is called with a |
349 | // value > 0, the report is printed in MB/sec instead of nanoseconds |
350 | // per iteration. |
351 | // |
352 | // REQUIRES: a benchmark has exited its KeepRunning loop. |
353 | BENCHMARK_ALWAYS_INLINE |
354 | void SetBytesProcessed(size_t bytes) { bytes_processed_ = bytes; } |
355 | |
356 | BENCHMARK_ALWAYS_INLINE |
357 | size_t bytes_processed() const { return bytes_processed_; } |
358 | |
359 | // If this routine is called with complexity_n > 0 and complexity report is |
360 | // requested for the |
361 | // family benchmark, then current benchmark will be part of the computation |
362 | // and complexity_n will |
363 | // represent the length of N. |
364 | BENCHMARK_ALWAYS_INLINE |
365 | void SetComplexityN(int complexity_n) { complexity_n_ = complexity_n; } |
366 | |
367 | BENCHMARK_ALWAYS_INLINE |
368 | int complexity_length_n() { return complexity_n_; } |
369 | |
370 | // If this routine is called with items > 0, then an items/s |
371 | // label is printed on the benchmark report line for the currently |
372 | // executing benchmark. It is typically called at the end of a processing |
373 | // benchmark where a processing items/second output is desired. |
374 | // |
375 | // REQUIRES: a benchmark has exited its KeepRunning loop. |
376 | BENCHMARK_ALWAYS_INLINE |
377 | void SetItemsProcessed(size_t items) { items_processed_ = items; } |
378 | |
379 | BENCHMARK_ALWAYS_INLINE |
380 | size_t items_processed() const { return items_processed_; } |
381 | |
382 | // If this routine is called, the specified label is printed at the |
383 | // end of the benchmark report line for the currently executing |
384 | // benchmark. Example: |
385 | // static void BM_Compress(benchmark::State& state) { |
386 | // ... |
387 | // double compress = input_size / output_size; |
388 | // state.SetLabel(StringPrintf("compress:%.1f%%", 100.0*compression)); |
389 | // } |
390 | // Produces output that looks like: |
391 | // BM_Compress 50 50 14115038 compress:27.3% |
392 | // |
393 | // REQUIRES: a benchmark has exited its KeepRunning loop. |
394 | void SetLabel(const char* label); |
395 | |
396 | // Allow the use of std::string without actually including <string>. |
397 | // This function does not participate in overload resolution unless StringType |
398 | // has the nested typename `basic_string`. This typename should be provided |
399 | // as an injected class name in the case of std::string. |
400 | template <class StringType> |
401 | void SetLabel(StringType const& str, |
402 | typename internal::EnableIfString<StringType>::type = 1) { |
403 | this->SetLabel(str.c_str()); |
404 | } |
405 | |
406 | // Range arguments for this run. CHECKs if the argument has been set. |
407 | BENCHMARK_ALWAYS_INLINE |
408 | int range(std::size_t pos = 0) const { |
409 | assert(range_.size() > pos); |
410 | return range_[pos]; |
411 | } |
412 | |
413 | BENCHMARK_DEPRECATED_MSG("use 'range(0)' instead" ) |
414 | int range_x() const { return range(0); } |
415 | |
416 | BENCHMARK_DEPRECATED_MSG("use 'range(1)' instead" ) |
417 | int range_y() const { return range(1); } |
418 | |
419 | BENCHMARK_ALWAYS_INLINE |
420 | size_t iterations() const { return total_iterations_; } |
421 | |
422 | private: |
423 | bool started_; |
424 | bool finished_; |
425 | size_t total_iterations_; |
426 | |
427 | std::vector<int> range_; |
428 | |
429 | size_t bytes_processed_; |
430 | size_t items_processed_; |
431 | |
432 | int complexity_n_; |
433 | |
434 | bool error_occurred_; |
435 | |
436 | public: |
437 | // Index of the executing thread. Values from [0, threads). |
438 | const int thread_index; |
439 | // Number of threads concurrently executing the benchmark. |
440 | const int threads; |
441 | const size_t max_iterations; |
442 | |
443 | // TODO make me private |
444 | State(size_t max_iters, const std::vector<int>& ranges, int thread_i, |
445 | int n_threads, internal::ThreadTimer* timer, |
446 | internal::ThreadManager* manager); |
447 | |
448 | private: |
449 | void StartKeepRunning(); |
450 | void FinishKeepRunning(); |
451 | internal::ThreadTimer* timer_; |
452 | internal::ThreadManager* manager_; |
453 | BENCHMARK_DISALLOW_COPY_AND_ASSIGN(State); |
454 | }; |
455 | |
456 | namespace internal { |
457 | |
458 | typedef void(Function)(State&); |
459 | |
460 | // ------------------------------------------------------ |
461 | // Benchmark registration object. The BENCHMARK() macro expands |
462 | // into an internal::Benchmark* object. Various methods can |
463 | // be called on this object to change the properties of the benchmark. |
464 | // Each method returns "this" so that multiple method calls can |
465 | // chained into one expression. |
466 | class Benchmark { |
467 | public: |
468 | virtual ~Benchmark(); |
469 | |
470 | // Note: the following methods all return "this" so that multiple |
471 | // method calls can be chained together in one expression. |
472 | |
473 | // Run this benchmark once with "x" as the extra argument passed |
474 | // to the function. |
475 | // REQUIRES: The function passed to the constructor must accept an arg1. |
476 | Benchmark* Arg(int x); |
477 | |
478 | // Run this benchmark with the given time unit for the generated output report |
479 | Benchmark* Unit(TimeUnit unit); |
480 | |
481 | // Run this benchmark once for a number of values picked from the |
482 | // range [start..limit]. (start and limit are always picked.) |
483 | // REQUIRES: The function passed to the constructor must accept an arg1. |
484 | Benchmark* Range(int start, int limit); |
485 | |
486 | // Run this benchmark once for all values in the range [start..limit] with |
487 | // specific step |
488 | // REQUIRES: The function passed to the constructor must accept an arg1. |
489 | Benchmark* DenseRange(int start, int limit, int step = 1); |
490 | |
491 | // Run this benchmark once with "args" as the extra arguments passed |
492 | // to the function. |
493 | // REQUIRES: The function passed to the constructor must accept arg1, arg2 ... |
494 | Benchmark* Args(const std::vector<int>& args); |
495 | |
496 | // Equivalent to Args({x, y}) |
497 | // NOTE: This is a legacy C++03 interface provided for compatibility only. |
498 | // New code should use 'Args'. |
499 | Benchmark* ArgPair(int x, int y) { |
500 | std::vector<int> args; |
501 | args.push_back(x); |
502 | args.push_back(y); |
503 | return Args(args); |
504 | } |
505 | |
506 | // Run this benchmark once for a number of values picked from the |
507 | // ranges [start..limit]. (starts and limits are always picked.) |
508 | // REQUIRES: The function passed to the constructor must accept arg1, arg2 ... |
509 | Benchmark* Ranges(const std::vector<std::pair<int, int> >& ranges); |
510 | |
511 | // Equivalent to ArgNames({name}) |
512 | Benchmark* ArgName(const std::string& name); |
513 | |
514 | // Set the argument names to display in the benchmark name. If not called, |
515 | // only argument values will be shown. |
516 | Benchmark* ArgNames(const std::vector<std::string>& names); |
517 | |
518 | // Equivalent to Ranges({{lo1, hi1}, {lo2, hi2}}). |
519 | // NOTE: This is a legacy C++03 interface provided for compatibility only. |
520 | // New code should use 'Ranges'. |
521 | Benchmark* RangePair(int lo1, int hi1, int lo2, int hi2) { |
522 | std::vector<std::pair<int, int> > ranges; |
523 | ranges.push_back(std::make_pair(lo1, hi1)); |
524 | ranges.push_back(std::make_pair(lo2, hi2)); |
525 | return Ranges(ranges); |
526 | } |
527 | |
528 | // Pass this benchmark object to *func, which can customize |
529 | // the benchmark by calling various methods like Arg, Args, |
530 | // Threads, etc. |
531 | Benchmark* Apply(void (*func)(Benchmark* benchmark)); |
532 | |
533 | // Set the range multiplier for non-dense range. If not called, the range |
534 | // multiplier kRangeMultiplier will be used. |
535 | Benchmark* RangeMultiplier(int multiplier); |
536 | |
537 | // Set the minimum amount of time to use when running this benchmark. This |
538 | // option overrides the `benchmark_min_time` flag. |
539 | // REQUIRES: `t > 0` |
540 | Benchmark* MinTime(double t); |
541 | |
542 | // Specify the amount of times to repeat this benchmark. This option overrides |
543 | // the `benchmark_repetitions` flag. |
544 | // REQUIRES: `n > 0` |
545 | Benchmark* Repetitions(int n); |
546 | |
547 | // Specify if each repetition of the benchmark should be reported separately |
548 | // or if only the final statistics should be reported. If the benchmark |
549 | // is not repeated then the single result is always reported. |
550 | Benchmark* ReportAggregatesOnly(bool v = true); |
551 | |
552 | // If a particular benchmark is I/O bound, runs multiple threads internally or |
553 | // if for some reason CPU timings are not representative, call this method. If |
554 | // called, the elapsed time will be used to control how many iterations are |
555 | // run, and in the printing of items/second or MB/seconds values. If not |
556 | // called, the cpu time used by the benchmark will be used. |
557 | Benchmark* UseRealTime(); |
558 | |
559 | // If a benchmark must measure time manually (e.g. if GPU execution time is |
560 | // being |
561 | // measured), call this method. If called, each benchmark iteration should |
562 | // call |
563 | // SetIterationTime(seconds) to report the measured time, which will be used |
564 | // to control how many iterations are run, and in the printing of items/second |
565 | // or MB/second values. |
566 | Benchmark* UseManualTime(); |
567 | |
568 | // Set the asymptotic computational complexity for the benchmark. If called |
569 | // the asymptotic computational complexity will be shown on the output. |
570 | Benchmark* Complexity(BigO complexity = benchmark::oAuto); |
571 | |
572 | // Set the asymptotic computational complexity for the benchmark. If called |
573 | // the asymptotic computational complexity will be shown on the output. |
574 | Benchmark* Complexity(BigOFunc* complexity); |
575 | |
576 | // Support for running multiple copies of the same benchmark concurrently |
577 | // in multiple threads. This may be useful when measuring the scaling |
578 | // of some piece of code. |
579 | |
580 | // Run one instance of this benchmark concurrently in t threads. |
581 | Benchmark* Threads(int t); |
582 | |
583 | // Pick a set of values T from [min_threads,max_threads]. |
584 | // min_threads and max_threads are always included in T. Run this |
585 | // benchmark once for each value in T. The benchmark run for a |
586 | // particular value t consists of t threads running the benchmark |
587 | // function concurrently. For example, consider: |
588 | // BENCHMARK(Foo)->ThreadRange(1,16); |
589 | // This will run the following benchmarks: |
590 | // Foo in 1 thread |
591 | // Foo in 2 threads |
592 | // Foo in 4 threads |
593 | // Foo in 8 threads |
594 | // Foo in 16 threads |
595 | Benchmark* ThreadRange(int min_threads, int max_threads); |
596 | |
597 | // For each value n in the range, run this benchmark once using n threads. |
598 | // min_threads and max_threads are always included in the range. |
599 | // stride specifies the increment. E.g. DenseThreadRange(1, 8, 3) starts |
600 | // a benchmark with 1, 4, 7 and 8 threads. |
601 | Benchmark* DenseThreadRange(int min_threads, int max_threads, int stride = 1); |
602 | |
603 | // Equivalent to ThreadRange(NumCPUs(), NumCPUs()) |
604 | Benchmark* ThreadPerCpu(); |
605 | |
606 | virtual void Run(State& state) = 0; |
607 | |
608 | // Used inside the benchmark implementation |
609 | struct Instance; |
610 | |
611 | protected: |
612 | explicit Benchmark(const char* name); |
613 | Benchmark(Benchmark const&); |
614 | void SetName(const char* name); |
615 | |
616 | int ArgsCnt() const; |
617 | |
618 | static void AddRange(std::vector<int>* dst, int lo, int hi, int mult); |
619 | |
620 | private: |
621 | friend class BenchmarkFamilies; |
622 | |
623 | std::string name_; |
624 | ReportMode report_mode_; |
625 | std::vector<std::string> arg_names_; // Args for all benchmark runs |
626 | std::vector<std::vector<int> > args_; // Args for all benchmark runs |
627 | TimeUnit time_unit_; |
628 | int range_multiplier_; |
629 | double min_time_; |
630 | int repetitions_; |
631 | bool use_real_time_; |
632 | bool use_manual_time_; |
633 | BigO complexity_; |
634 | BigOFunc* complexity_lambda_; |
635 | std::vector<int> thread_counts_; |
636 | |
637 | Benchmark& operator=(Benchmark const&); |
638 | }; |
639 | |
640 | } // namespace internal |
641 | |
642 | // Create and register a benchmark with the specified 'name' that invokes |
643 | // the specified functor 'fn'. |
644 | // |
645 | // RETURNS: A pointer to the registered benchmark. |
646 | internal::Benchmark* RegisterBenchmark(const char* name, |
647 | internal::Function* fn); |
648 | |
649 | #if defined(BENCHMARK_HAS_CXX11) |
650 | template <class Lambda> |
651 | internal::Benchmark* RegisterBenchmark(const char* name, Lambda&& fn); |
652 | #endif |
653 | |
654 | namespace internal { |
655 | // The class used to hold all Benchmarks created from static function. |
656 | // (ie those created using the BENCHMARK(...) macros. |
657 | class FunctionBenchmark : public Benchmark { |
658 | public: |
659 | FunctionBenchmark(const char* name, Function* func) |
660 | : Benchmark(name), func_(func) {} |
661 | |
662 | virtual void Run(State& st); |
663 | |
664 | private: |
665 | Function* func_; |
666 | }; |
667 | |
668 | #ifdef BENCHMARK_HAS_CXX11 |
669 | template <class Lambda> |
670 | class LambdaBenchmark : public Benchmark { |
671 | public: |
672 | virtual void Run(State& st) { lambda_(st); } |
673 | |
674 | private: |
675 | template <class OLambda> |
676 | LambdaBenchmark(const char* name, OLambda&& lam) |
677 | : Benchmark(name), lambda_(std::forward<OLambda>(lam)) {} |
678 | |
679 | LambdaBenchmark(LambdaBenchmark const&) = delete; |
680 | |
681 | private: |
682 | template <class Lam> |
683 | friend Benchmark* ::benchmark::RegisterBenchmark(const char*, Lam&&); |
684 | |
685 | Lambda lambda_; |
686 | }; |
687 | #endif |
688 | |
689 | } // end namespace internal |
690 | |
691 | inline internal::Benchmark* RegisterBenchmark(const char* name, |
692 | internal::Function* fn) { |
693 | return internal::RegisterBenchmarkInternal( |
694 | ::new internal::FunctionBenchmark(name, fn)); |
695 | } |
696 | |
697 | #ifdef BENCHMARK_HAS_CXX11 |
698 | template <class Lambda> |
699 | internal::Benchmark* RegisterBenchmark(const char* name, Lambda&& fn) { |
700 | using BenchType = |
701 | internal::LambdaBenchmark<typename std::decay<Lambda>::type>; |
702 | return internal::RegisterBenchmarkInternal( |
703 | ::new BenchType(name, std::forward<Lambda>(fn))); |
704 | } |
705 | #endif |
706 | |
707 | #if defined(BENCHMARK_HAS_CXX11) && \ |
708 | (!defined(BENCHMARK_GCC_VERSION) || BENCHMARK_GCC_VERSION >= 409) |
709 | template <class Lambda, class... Args> |
710 | internal::Benchmark* RegisterBenchmark(const char* name, Lambda&& fn, |
711 | Args&&... args) { |
712 | return benchmark::RegisterBenchmark( |
713 | name, [=](benchmark::State& st) { fn(st, args...); }); |
714 | } |
715 | #else |
716 | #define BENCHMARK_HAS_NO_VARIADIC_REGISTER_BENCHMARK |
717 | #endif |
718 | |
719 | // The base class for all fixture tests. |
720 | class Fixture : public internal::Benchmark { |
721 | public: |
722 | Fixture() : internal::Benchmark("" ) {} |
723 | |
724 | virtual void Run(State& st) { |
725 | this->SetUp(st); |
726 | this->BenchmarkCase(st); |
727 | this->TearDown(st); |
728 | } |
729 | |
730 | // These will be deprecated ... |
731 | virtual void SetUp(const State&) {} |
732 | virtual void TearDown(const State&) {} |
733 | // ... In favor of these. |
734 | virtual void SetUp(State& st) { SetUp(const_cast<const State&>(st)); } |
735 | virtual void TearDown(State& st) { TearDown(const_cast<const State&>(st)); } |
736 | |
737 | protected: |
738 | virtual void BenchmarkCase(State&) = 0; |
739 | }; |
740 | |
741 | } // end namespace benchmark |
742 | |
743 | // ------------------------------------------------------ |
744 | // Macro to register benchmarks |
745 | |
746 | // Check that __COUNTER__ is defined and that __COUNTER__ increases by 1 |
747 | // every time it is expanded. X + 1 == X + 0 is used in case X is defined to be |
748 | // empty. If X is empty the expression becomes (+1 == +0). |
749 | #if defined(__COUNTER__) && (__COUNTER__ + 1 == __COUNTER__ + 0) |
750 | #define BENCHMARK_PRIVATE_UNIQUE_ID __COUNTER__ |
751 | #else |
752 | #define BENCHMARK_PRIVATE_UNIQUE_ID __LINE__ |
753 | #endif |
754 | |
755 | // Helpers for generating unique variable names |
756 | #define BENCHMARK_PRIVATE_NAME(n) \ |
757 | BENCHMARK_PRIVATE_CONCAT(_benchmark_, BENCHMARK_PRIVATE_UNIQUE_ID, n) |
758 | #define BENCHMARK_PRIVATE_CONCAT(a, b, c) BENCHMARK_PRIVATE_CONCAT2(a, b, c) |
759 | #define BENCHMARK_PRIVATE_CONCAT2(a, b, c) a##b##c |
760 | |
761 | #define BENCHMARK_PRIVATE_DECLARE(n) \ |
762 | static ::benchmark::internal::Benchmark* BENCHMARK_PRIVATE_NAME(n) \ |
763 | BENCHMARK_UNUSED |
764 | |
765 | #define BENCHMARK(n) \ |
766 | BENCHMARK_PRIVATE_DECLARE(n) = \ |
767 | (::benchmark::internal::RegisterBenchmarkInternal( \ |
768 | new ::benchmark::internal::FunctionBenchmark(#n, n))) |
769 | |
770 | // Old-style macros |
771 | #define BENCHMARK_WITH_ARG(n, a) BENCHMARK(n)->Arg((a)) |
772 | #define BENCHMARK_WITH_ARG2(n, a1, a2) BENCHMARK(n)->Args({(a1), (a2)}) |
773 | #define BENCHMARK_WITH_UNIT(n, t) BENCHMARK(n)->Unit((t)) |
774 | #define BENCHMARK_RANGE(n, lo, hi) BENCHMARK(n)->Range((lo), (hi)) |
775 | #define BENCHMARK_RANGE2(n, l1, h1, l2, h2) \ |
776 | BENCHMARK(n)->RangePair({{(l1), (h1)}, {(l2), (h2)}}) |
777 | |
778 | #if __cplusplus >= 201103L |
779 | |
780 | // Register a benchmark which invokes the function specified by `func` |
781 | // with the additional arguments specified by `...`. |
782 | // |
783 | // For example: |
784 | // |
785 | // template <class ...ExtraArgs>` |
786 | // void BM_takes_args(benchmark::State& state, ExtraArgs&&... extra_args) { |
787 | // [...] |
788 | //} |
789 | // /* Registers a benchmark named "BM_takes_args/int_string_test` */ |
790 | // BENCHMARK_CAPTURE(BM_takes_args, int_string_test, 42, std::string("abc")); |
791 | #define BENCHMARK_CAPTURE(func, test_case_name, ...) \ |
792 | BENCHMARK_PRIVATE_DECLARE(func) = \ |
793 | (::benchmark::internal::RegisterBenchmarkInternal( \ |
794 | new ::benchmark::internal::FunctionBenchmark( \ |
795 | #func "/" #test_case_name, \ |
796 | [](::benchmark::State& st) { func(st, __VA_ARGS__); }))) |
797 | |
798 | #endif // __cplusplus >= 11 |
799 | |
800 | // This will register a benchmark for a templatized function. For example: |
801 | // |
802 | // template<int arg> |
803 | // void BM_Foo(int iters); |
804 | // |
805 | // BENCHMARK_TEMPLATE(BM_Foo, 1); |
806 | // |
807 | // will register BM_Foo<1> as a benchmark. |
808 | #define BENCHMARK_TEMPLATE1(n, a) \ |
809 | BENCHMARK_PRIVATE_DECLARE(n) = \ |
810 | (::benchmark::internal::RegisterBenchmarkInternal( \ |
811 | new ::benchmark::internal::FunctionBenchmark(#n "<" #a ">", n<a>))) |
812 | |
813 | #define BENCHMARK_TEMPLATE2(n, a, b) \ |
814 | BENCHMARK_PRIVATE_DECLARE(n) = \ |
815 | (::benchmark::internal::RegisterBenchmarkInternal( \ |
816 | new ::benchmark::internal::FunctionBenchmark(#n "<" #a "," #b ">", \ |
817 | n<a, b>))) |
818 | |
819 | #if __cplusplus >= 201103L |
820 | #define BENCHMARK_TEMPLATE(n, ...) \ |
821 | BENCHMARK_PRIVATE_DECLARE(n) = \ |
822 | (::benchmark::internal::RegisterBenchmarkInternal( \ |
823 | new ::benchmark::internal::FunctionBenchmark( \ |
824 | #n "<" #__VA_ARGS__ ">", n<__VA_ARGS__>))) |
825 | #else |
826 | #define BENCHMARK_TEMPLATE(n, a) BENCHMARK_TEMPLATE1(n, a) |
827 | #endif |
828 | |
829 | #define BENCHMARK_PRIVATE_DECLARE_F(BaseClass, Method) \ |
830 | class BaseClass##_##Method##_Benchmark : public BaseClass { \ |
831 | public: \ |
832 | BaseClass##_##Method##_Benchmark() : BaseClass() { \ |
833 | this->SetName(#BaseClass "/" #Method); \ |
834 | } \ |
835 | \ |
836 | protected: \ |
837 | virtual void BenchmarkCase(::benchmark::State&); \ |
838 | }; |
839 | |
840 | #define BENCHMARK_DEFINE_F(BaseClass, Method) \ |
841 | BENCHMARK_PRIVATE_DECLARE_F(BaseClass, Method) \ |
842 | void BaseClass##_##Method##_Benchmark::BenchmarkCase |
843 | |
844 | #define BENCHMARK_REGISTER_F(BaseClass, Method) \ |
845 | BENCHMARK_PRIVATE_REGISTER_F(BaseClass##_##Method##_Benchmark) |
846 | |
847 | #define BENCHMARK_PRIVATE_REGISTER_F(TestName) \ |
848 | BENCHMARK_PRIVATE_DECLARE(TestName) = \ |
849 | (::benchmark::internal::RegisterBenchmarkInternal(new TestName())) |
850 | |
851 | // This macro will define and register a benchmark within a fixture class. |
852 | #define BENCHMARK_F(BaseClass, Method) \ |
853 | BENCHMARK_PRIVATE_DECLARE_F(BaseClass, Method) \ |
854 | BENCHMARK_REGISTER_F(BaseClass, Method); \ |
855 | void BaseClass##_##Method##_Benchmark::BenchmarkCase |
856 | |
857 | // Helper macro to create a main routine in a test that runs the benchmarks |
858 | #define BENCHMARK_MAIN() \ |
859 | int main(int argc, char** argv) { \ |
860 | ::benchmark::Initialize(&argc, argv); \ |
861 | ::benchmark::RunSpecifiedBenchmarks(); \ |
862 | } |
863 | |
864 | #endif // BENCHMARK_BENCHMARK_API_H_ |
865 | |