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 | // Support for registering benchmarks for functions. |
16 | |
17 | /* Example usage: |
18 | // Define a function that executes the code to be measured a |
19 | // specified number of times: |
20 | static void BM_StringCreation(benchmark::State& state) { |
21 | for (auto _ : state) |
22 | std::string empty_string; |
23 | } |
24 | |
25 | // Register the function as a benchmark |
26 | BENCHMARK(BM_StringCreation); |
27 | |
28 | // Define another benchmark |
29 | static void BM_StringCopy(benchmark::State& state) { |
30 | std::string x = "hello"; |
31 | for (auto _ : state) |
32 | std::string copy(x); |
33 | } |
34 | BENCHMARK(BM_StringCopy); |
35 | |
36 | // Augment the main() program to invoke benchmarks if specified |
37 | // via the --benchmarks command line flag. E.g., |
38 | // my_unittest --benchmark_filter=all |
39 | // my_unittest --benchmark_filter=BM_StringCreation |
40 | // my_unittest --benchmark_filter=String |
41 | // my_unittest --benchmark_filter='Copy|Creation' |
42 | int main(int argc, char** argv) { |
43 | benchmark::Initialize(&argc, argv); |
44 | benchmark::RunSpecifiedBenchmarks(); |
45 | return 0; |
46 | } |
47 | |
48 | // Sometimes a family of microbenchmarks can be implemented with |
49 | // just one routine that takes an extra argument to specify which |
50 | // one of the family of benchmarks to run. For example, the following |
51 | // code defines a family of microbenchmarks for measuring the speed |
52 | // of memcpy() calls of different lengths: |
53 | |
54 | static void BM_memcpy(benchmark::State& state) { |
55 | char* src = new char[state.range(0)]; char* dst = new char[state.range(0)]; |
56 | memset(src, 'x', state.range(0)); |
57 | for (auto _ : state) |
58 | memcpy(dst, src, state.range(0)); |
59 | state.SetBytesProcessed(state.iterations() * state.range(0)); |
60 | delete[] src; delete[] dst; |
61 | } |
62 | BENCHMARK(BM_memcpy)->Arg(8)->Arg(64)->Arg(512)->Arg(1<<10)->Arg(8<<10); |
63 | |
64 | // The preceding code is quite repetitive, and can be replaced with the |
65 | // following short-hand. The following invocation will pick a few |
66 | // appropriate arguments in the specified range and will generate a |
67 | // microbenchmark for each such argument. |
68 | BENCHMARK(BM_memcpy)->Range(8, 8<<10); |
69 | |
70 | // You might have a microbenchmark that depends on two inputs. For |
71 | // example, the following code defines a family of microbenchmarks for |
72 | // measuring the speed of set insertion. |
73 | static void BM_SetInsert(benchmark::State& state) { |
74 | set<int> data; |
75 | for (auto _ : state) { |
76 | state.PauseTiming(); |
77 | data = ConstructRandomSet(state.range(0)); |
78 | state.ResumeTiming(); |
79 | for (int j = 0; j < state.range(1); ++j) |
80 | data.insert(RandomNumber()); |
81 | } |
82 | } |
83 | BENCHMARK(BM_SetInsert) |
84 | ->Args({1<<10, 128}) |
85 | ->Args({2<<10, 128}) |
86 | ->Args({4<<10, 128}) |
87 | ->Args({8<<10, 128}) |
88 | ->Args({1<<10, 512}) |
89 | ->Args({2<<10, 512}) |
90 | ->Args({4<<10, 512}) |
91 | ->Args({8<<10, 512}); |
92 | |
93 | // The preceding code is quite repetitive, and can be replaced with |
94 | // the following short-hand. The following macro will pick a few |
95 | // appropriate arguments in the product of the two specified ranges |
96 | // and will generate a microbenchmark for each such pair. |
97 | BENCHMARK(BM_SetInsert)->Ranges({{1<<10, 8<<10}, {128, 512}}); |
98 | |
99 | // For more complex patterns of inputs, passing a custom function |
100 | // to Apply allows programmatic specification of an |
101 | // arbitrary set of arguments to run the microbenchmark on. |
102 | // The following example enumerates a dense range on |
103 | // one parameter, and a sparse range on the second. |
104 | static void CustomArguments(benchmark::internal::Benchmark* b) { |
105 | for (int i = 0; i <= 10; ++i) |
106 | for (int j = 32; j <= 1024*1024; j *= 8) |
107 | b->Args({i, j}); |
108 | } |
109 | BENCHMARK(BM_SetInsert)->Apply(CustomArguments); |
110 | |
111 | // Templated microbenchmarks work the same way: |
112 | // Produce then consume 'size' messages 'iters' times |
113 | // Measures throughput in the absence of multiprogramming. |
114 | template <class Q> int BM_Sequential(benchmark::State& state) { |
115 | Q q; |
116 | typename Q::value_type v; |
117 | for (auto _ : state) { |
118 | for (int i = state.range(0); i--; ) |
119 | q.push(v); |
120 | for (int e = state.range(0); e--; ) |
121 | q.Wait(&v); |
122 | } |
123 | // actually messages, not bytes: |
124 | state.SetBytesProcessed(state.iterations() * state.range(0)); |
125 | } |
126 | BENCHMARK_TEMPLATE(BM_Sequential, WaitQueue<int>)->Range(1<<0, 1<<10); |
127 | |
128 | Use `Benchmark::MinTime(double t)` to set the minimum time used to run the |
129 | benchmark. This option overrides the `benchmark_min_time` flag. |
130 | |
131 | void BM_test(benchmark::State& state) { |
132 | ... body ... |
133 | } |
134 | BENCHMARK(BM_test)->MinTime(2.0); // Run for at least 2 seconds. |
135 | |
136 | In a multithreaded test, it is guaranteed that none of the threads will start |
137 | until all have reached the loop start, and all will have finished before any |
138 | thread exits the loop body. As such, any global setup or teardown you want to |
139 | do can be wrapped in a check against the thread index: |
140 | |
141 | static void BM_MultiThreaded(benchmark::State& state) { |
142 | if (state.thread_index == 0) { |
143 | // Setup code here. |
144 | } |
145 | for (auto _ : state) { |
146 | // Run the test as normal. |
147 | } |
148 | if (state.thread_index == 0) { |
149 | // Teardown code here. |
150 | } |
151 | } |
152 | BENCHMARK(BM_MultiThreaded)->Threads(4); |
153 | |
154 | |
155 | If a benchmark runs a few milliseconds it may be hard to visually compare the |
156 | measured times, since the output data is given in nanoseconds per default. In |
157 | order to manually set the time unit, you can specify it manually: |
158 | |
159 | BENCHMARK(BM_test)->Unit(benchmark::kMillisecond); |
160 | */ |
161 | |
162 | #ifndef BENCHMARK_BENCHMARK_H_ |
163 | #define BENCHMARK_BENCHMARK_H_ |
164 | |
165 | // The _MSVC_LANG check should detect Visual Studio 2015 Update 3 and newer. |
166 | #if __cplusplus >= 201103L || (defined(_MSVC_LANG) && _MSVC_LANG >= 201103L) |
167 | #define BENCHMARK_HAS_CXX11 |
168 | #endif |
169 | |
170 | #include <stdint.h> |
171 | |
172 | #include <algorithm> |
173 | #include <cassert> |
174 | #include <cstddef> |
175 | #include <iosfwd> |
176 | #include <map> |
177 | #include <set> |
178 | #include <string> |
179 | #include <utility> |
180 | #include <vector> |
181 | |
182 | #if defined(BENCHMARK_HAS_CXX11) |
183 | #include <initializer_list> |
184 | #include <type_traits> |
185 | #include <utility> |
186 | #endif |
187 | |
188 | #if defined(_MSC_VER) |
189 | #include <intrin.h> // for _ReadWriteBarrier |
190 | #endif |
191 | |
192 | #ifndef BENCHMARK_HAS_CXX11 |
193 | #define BENCHMARK_DISALLOW_COPY_AND_ASSIGN(TypeName) \ |
194 | TypeName(const TypeName&); \ |
195 | TypeName& operator=(const TypeName&) |
196 | #else |
197 | #define BENCHMARK_DISALLOW_COPY_AND_ASSIGN(TypeName) \ |
198 | TypeName(const TypeName&) = delete; \ |
199 | TypeName& operator=(const TypeName&) = delete |
200 | #endif |
201 | |
202 | #if defined(__GNUC__) |
203 | #define BENCHMARK_UNUSED __attribute__((unused)) |
204 | #define BENCHMARK_ALWAYS_INLINE __attribute__((always_inline)) |
205 | #define BENCHMARK_NOEXCEPT noexcept |
206 | #define BENCHMARK_NOEXCEPT_OP(x) noexcept(x) |
207 | #elif defined(_MSC_VER) && !defined(__clang__) |
208 | #define BENCHMARK_UNUSED |
209 | #define BENCHMARK_ALWAYS_INLINE __forceinline |
210 | #if _MSC_VER >= 1900 |
211 | #define BENCHMARK_NOEXCEPT noexcept |
212 | #define BENCHMARK_NOEXCEPT_OP(x) noexcept(x) |
213 | #else |
214 | #define BENCHMARK_NOEXCEPT |
215 | #define BENCHMARK_NOEXCEPT_OP(x) |
216 | #endif |
217 | #define __func__ __FUNCTION__ |
218 | #else |
219 | #define BENCHMARK_UNUSED |
220 | #define BENCHMARK_ALWAYS_INLINE |
221 | #define BENCHMARK_NOEXCEPT |
222 | #define BENCHMARK_NOEXCEPT_OP(x) |
223 | #endif |
224 | |
225 | #define BENCHMARK_INTERNAL_TOSTRING2(x) #x |
226 | #define BENCHMARK_INTERNAL_TOSTRING(x) BENCHMARK_INTERNAL_TOSTRING2(x) |
227 | |
228 | #if defined(__GNUC__) || defined(__clang__) |
229 | #define BENCHMARK_BUILTIN_EXPECT(x, y) __builtin_expect(x, y) |
230 | #define BENCHMARK_DEPRECATED_MSG(msg) __attribute__((deprecated(msg))) |
231 | #else |
232 | #define BENCHMARK_BUILTIN_EXPECT(x, y) x |
233 | #define BENCHMARK_DEPRECATED_MSG(msg) |
234 | #define BENCHMARK_WARNING_MSG(msg) \ |
235 | __pragma(message(__FILE__ "(" BENCHMARK_INTERNAL_TOSTRING( \ |
236 | __LINE__) ") : warning note: " msg)) |
237 | #endif |
238 | |
239 | #if defined(__GNUC__) && !defined(__clang__) |
240 | #define BENCHMARK_GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__) |
241 | #endif |
242 | |
243 | #ifndef __has_builtin |
244 | #define __has_builtin(x) 0 |
245 | #endif |
246 | |
247 | #if defined(__GNUC__) || __has_builtin(__builtin_unreachable) |
248 | #define BENCHMARK_UNREACHABLE() __builtin_unreachable() |
249 | #elif defined(_MSC_VER) |
250 | #define BENCHMARK_UNREACHABLE() __assume(false) |
251 | #else |
252 | #define BENCHMARK_UNREACHABLE() ((void)0) |
253 | #endif |
254 | |
255 | namespace benchmark { |
256 | class BenchmarkReporter; |
257 | class MemoryManager; |
258 | |
259 | void Initialize(int* argc, char** argv); |
260 | |
261 | // Report to stdout all arguments in 'argv' as unrecognized except the first. |
262 | // Returns true there is at least on unrecognized argument (i.e. 'argc' > 1). |
263 | bool ReportUnrecognizedArguments(int argc, char** argv); |
264 | |
265 | // Generate a list of benchmarks matching the specified --benchmark_filter flag |
266 | // and if --benchmark_list_tests is specified return after printing the name |
267 | // of each matching benchmark. Otherwise run each matching benchmark and |
268 | // report the results. |
269 | // |
270 | // The second and third overload use the specified 'display_reporter' and |
271 | // 'file_reporter' respectively. 'file_reporter' will write to the file |
272 | // specified |
273 | // by '--benchmark_output'. If '--benchmark_output' is not given the |
274 | // 'file_reporter' is ignored. |
275 | // |
276 | // RETURNS: The number of matching benchmarks. |
277 | size_t RunSpecifiedBenchmarks(); |
278 | size_t RunSpecifiedBenchmarks(BenchmarkReporter* display_reporter); |
279 | size_t RunSpecifiedBenchmarks(BenchmarkReporter* display_reporter, |
280 | BenchmarkReporter* file_reporter); |
281 | |
282 | // Register a MemoryManager instance that will be used to collect and report |
283 | // allocation measurements for benchmark runs. |
284 | void RegisterMemoryManager(MemoryManager* memory_manager); |
285 | |
286 | namespace internal { |
287 | class Benchmark; |
288 | class BenchmarkImp; |
289 | class BenchmarkFamilies; |
290 | |
291 | void UseCharPointer(char const volatile*); |
292 | |
293 | // Take ownership of the pointer and register the benchmark. Return the |
294 | // registered benchmark. |
295 | Benchmark* RegisterBenchmarkInternal(Benchmark*); |
296 | |
297 | // Ensure that the standard streams are properly initialized in every TU. |
298 | int InitializeStreams(); |
299 | BENCHMARK_UNUSED static int stream_init_anchor = InitializeStreams(); |
300 | |
301 | } // namespace internal |
302 | |
303 | #if (!defined(__GNUC__) && !defined(__clang__)) || defined(__pnacl__) || \ |
304 | defined(__EMSCRIPTEN__) |
305 | #define BENCHMARK_HAS_NO_INLINE_ASSEMBLY |
306 | #endif |
307 | |
308 | // The DoNotOptimize(...) function can be used to prevent a value or |
309 | // expression from being optimized away by the compiler. This function is |
310 | // intended to add little to no overhead. |
311 | // See: https://youtu.be/nXaxk27zwlk?t=2441 |
312 | #ifndef BENCHMARK_HAS_NO_INLINE_ASSEMBLY |
313 | template <class Tp> |
314 | inline BENCHMARK_ALWAYS_INLINE void DoNotOptimize(Tp const& value) { |
315 | asm volatile("" : : "r,m" (value) : "memory" ); |
316 | } |
317 | |
318 | template <class Tp> |
319 | inline BENCHMARK_ALWAYS_INLINE void DoNotOptimize(Tp& value) { |
320 | #if defined(__clang__) |
321 | asm volatile("" : "+r,m" (value) : : "memory" ); |
322 | #else |
323 | asm volatile("" : "+m,r" (value) : : "memory" ); |
324 | #endif |
325 | } |
326 | |
327 | // Force the compiler to flush pending writes to global memory. Acts as an |
328 | // effective read/write barrier |
329 | inline BENCHMARK_ALWAYS_INLINE void ClobberMemory() { |
330 | asm volatile("" : : : "memory" ); |
331 | } |
332 | #elif defined(_MSC_VER) |
333 | template <class Tp> |
334 | inline BENCHMARK_ALWAYS_INLINE void DoNotOptimize(Tp const& value) { |
335 | internal::UseCharPointer(&reinterpret_cast<char const volatile&>(value)); |
336 | _ReadWriteBarrier(); |
337 | } |
338 | |
339 | inline BENCHMARK_ALWAYS_INLINE void ClobberMemory() { _ReadWriteBarrier(); } |
340 | #else |
341 | template <class Tp> |
342 | inline BENCHMARK_ALWAYS_INLINE void DoNotOptimize(Tp const& value) { |
343 | internal::UseCharPointer(&reinterpret_cast<char const volatile&>(value)); |
344 | } |
345 | // FIXME Add ClobberMemory() for non-gnu and non-msvc compilers |
346 | #endif |
347 | |
348 | // This class is used for user-defined counters. |
349 | class Counter { |
350 | public: |
351 | enum Flags { |
352 | kDefaults = 0, |
353 | // Mark the counter as a rate. It will be presented divided |
354 | // by the duration of the benchmark. |
355 | kIsRate = 1U << 0U, |
356 | // Mark the counter as a thread-average quantity. It will be |
357 | // presented divided by the number of threads. |
358 | kAvgThreads = 1U << 1U, |
359 | // Mark the counter as a thread-average rate. See above. |
360 | kAvgThreadsRate = kIsRate | kAvgThreads, |
361 | // Mark the counter as a constant value, valid/same for *every* iteration. |
362 | // When reporting, it will be *multiplied* by the iteration count. |
363 | kIsIterationInvariant = 1U << 2U, |
364 | // Mark the counter as a constant rate. |
365 | // When reporting, it will be *multiplied* by the iteration count |
366 | // and then divided by the duration of the benchmark. |
367 | kIsIterationInvariantRate = kIsRate | kIsIterationInvariant, |
368 | // Mark the counter as a iteration-average quantity. |
369 | // It will be presented divided by the number of iterations. |
370 | kAvgIterations = 1U << 3U, |
371 | // Mark the counter as a iteration-average rate. See above. |
372 | kAvgIterationsRate = kIsRate | kAvgIterations, |
373 | |
374 | // In the end, invert the result. This is always done last! |
375 | kInvert = 1U << 31U |
376 | }; |
377 | |
378 | enum OneK { |
379 | // 1'000 items per 1k |
380 | kIs1000 = 1000, |
381 | // 1'024 items per 1k |
382 | kIs1024 = 1024 |
383 | }; |
384 | |
385 | double value; |
386 | Flags flags; |
387 | OneK oneK; |
388 | |
389 | BENCHMARK_ALWAYS_INLINE |
390 | Counter(double v = 0., Flags f = kDefaults, OneK k = kIs1000) |
391 | : value(v), flags(f), oneK(k) {} |
392 | |
393 | BENCHMARK_ALWAYS_INLINE operator double const&() const { return value; } |
394 | BENCHMARK_ALWAYS_INLINE operator double&() { return value; } |
395 | }; |
396 | |
397 | // A helper for user code to create unforeseen combinations of Flags, without |
398 | // having to do this cast manually each time, or providing this operator. |
399 | Counter::Flags inline operator|(const Counter::Flags& LHS, |
400 | const Counter::Flags& RHS) { |
401 | return static_cast<Counter::Flags>(static_cast<int>(LHS) | |
402 | static_cast<int>(RHS)); |
403 | } |
404 | |
405 | // This is the container for the user-defined counters. |
406 | typedef std::map<std::string, Counter> UserCounters; |
407 | |
408 | // TimeUnit is passed to a benchmark in order to specify the order of magnitude |
409 | // for the measured time. |
410 | enum TimeUnit { kNanosecond, kMicrosecond, kMillisecond }; |
411 | |
412 | // BigO is passed to a benchmark in order to specify the asymptotic |
413 | // computational |
414 | // complexity for the benchmark. In case oAuto is selected, complexity will be |
415 | // calculated automatically to the best fit. |
416 | enum BigO { oNone, o1, oN, oNSquared, oNCubed, oLogN, oNLogN, oAuto, oLambda }; |
417 | |
418 | typedef uint64_t IterationCount; |
419 | |
420 | // BigOFunc is passed to a benchmark in order to specify the asymptotic |
421 | // computational complexity for the benchmark. |
422 | typedef double(BigOFunc)(IterationCount); |
423 | |
424 | // StatisticsFunc is passed to a benchmark in order to compute some descriptive |
425 | // statistics over all the measurements of some type |
426 | typedef double(StatisticsFunc)(const std::vector<double>&); |
427 | |
428 | namespace internal { |
429 | struct Statistics { |
430 | std::string name_; |
431 | StatisticsFunc* compute_; |
432 | |
433 | Statistics(const std::string& name, StatisticsFunc* compute) |
434 | : name_(name), compute_(compute) {} |
435 | }; |
436 | |
437 | struct BenchmarkInstance; |
438 | class ThreadTimer; |
439 | class ThreadManager; |
440 | |
441 | enum AggregationReportMode |
442 | #if defined(BENCHMARK_HAS_CXX11) |
443 | : unsigned |
444 | #else |
445 | #endif |
446 | { |
447 | // The mode has not been manually specified |
448 | ARM_Unspecified = 0, |
449 | // The mode is user-specified. |
450 | // This may or may not be set when the following bit-flags are set. |
451 | ARM_Default = 1U << 0U, |
452 | // File reporter should only output aggregates. |
453 | ARM_FileReportAggregatesOnly = 1U << 1U, |
454 | // Display reporter should only output aggregates |
455 | ARM_DisplayReportAggregatesOnly = 1U << 2U, |
456 | // Both reporters should only display aggregates. |
457 | ARM_ReportAggregatesOnly = |
458 | ARM_FileReportAggregatesOnly | ARM_DisplayReportAggregatesOnly |
459 | }; |
460 | |
461 | } // namespace internal |
462 | |
463 | // State is passed to a running Benchmark and contains state for the |
464 | // benchmark to use. |
465 | class State { |
466 | public: |
467 | struct StateIterator; |
468 | friend struct StateIterator; |
469 | |
470 | // Returns iterators used to run each iteration of a benchmark using a |
471 | // C++11 ranged-based for loop. These functions should not be called directly. |
472 | // |
473 | // REQUIRES: The benchmark has not started running yet. Neither begin nor end |
474 | // have been called previously. |
475 | // |
476 | // NOTE: KeepRunning may not be used after calling either of these functions. |
477 | BENCHMARK_ALWAYS_INLINE StateIterator begin(); |
478 | BENCHMARK_ALWAYS_INLINE StateIterator end(); |
479 | |
480 | // Returns true if the benchmark should continue through another iteration. |
481 | // NOTE: A benchmark may not return from the test until KeepRunning() has |
482 | // returned false. |
483 | bool KeepRunning(); |
484 | |
485 | // Returns true iff the benchmark should run n more iterations. |
486 | // REQUIRES: 'n' > 0. |
487 | // NOTE: A benchmark must not return from the test until KeepRunningBatch() |
488 | // has returned false. |
489 | // NOTE: KeepRunningBatch() may overshoot by up to 'n' iterations. |
490 | // |
491 | // Intended usage: |
492 | // while (state.KeepRunningBatch(1000)) { |
493 | // // process 1000 elements |
494 | // } |
495 | bool KeepRunningBatch(IterationCount n); |
496 | |
497 | // REQUIRES: timer is running and 'SkipWithError(...)' has not been called |
498 | // by the current thread. |
499 | // Stop the benchmark timer. If not called, the timer will be |
500 | // automatically stopped after the last iteration of the benchmark loop. |
501 | // |
502 | // For threaded benchmarks the PauseTiming() function only pauses the timing |
503 | // for the current thread. |
504 | // |
505 | // NOTE: The "real time" measurement is per-thread. If different threads |
506 | // report different measurements the largest one is reported. |
507 | // |
508 | // NOTE: PauseTiming()/ResumeTiming() are relatively |
509 | // heavyweight, and so their use should generally be avoided |
510 | // within each benchmark iteration, if possible. |
511 | void PauseTiming(); |
512 | |
513 | // REQUIRES: timer is not running and 'SkipWithError(...)' has not been called |
514 | // by the current thread. |
515 | // Start the benchmark timer. The timer is NOT running on entrance to the |
516 | // benchmark function. It begins running after control flow enters the |
517 | // benchmark loop. |
518 | // |
519 | // NOTE: PauseTiming()/ResumeTiming() are relatively |
520 | // heavyweight, and so their use should generally be avoided |
521 | // within each benchmark iteration, if possible. |
522 | void ResumeTiming(); |
523 | |
524 | // REQUIRES: 'SkipWithError(...)' has not been called previously by the |
525 | // current thread. |
526 | // Report the benchmark as resulting in an error with the specified 'msg'. |
527 | // After this call the user may explicitly 'return' from the benchmark. |
528 | // |
529 | // If the ranged-for style of benchmark loop is used, the user must explicitly |
530 | // break from the loop, otherwise all future iterations will be run. |
531 | // If the 'KeepRunning()' loop is used the current thread will automatically |
532 | // exit the loop at the end of the current iteration. |
533 | // |
534 | // For threaded benchmarks only the current thread stops executing and future |
535 | // calls to `KeepRunning()` will block until all threads have completed |
536 | // the `KeepRunning()` loop. If multiple threads report an error only the |
537 | // first error message is used. |
538 | // |
539 | // NOTE: Calling 'SkipWithError(...)' does not cause the benchmark to exit |
540 | // the current scope immediately. If the function is called from within |
541 | // the 'KeepRunning()' loop the current iteration will finish. It is the users |
542 | // responsibility to exit the scope as needed. |
543 | void SkipWithError(const char* msg); |
544 | |
545 | // Returns true if an error has been reported with 'SkipWithError(...)'. |
546 | bool error_occurred() const { return error_occurred_; } |
547 | |
548 | // REQUIRES: called exactly once per iteration of the benchmarking loop. |
549 | // Set the manually measured time for this benchmark iteration, which |
550 | // is used instead of automatically measured time if UseManualTime() was |
551 | // specified. |
552 | // |
553 | // For threaded benchmarks the final value will be set to the largest |
554 | // reported values. |
555 | void SetIterationTime(double seconds); |
556 | |
557 | // Set the number of bytes processed by the current benchmark |
558 | // execution. This routine is typically called once at the end of a |
559 | // throughput oriented benchmark. |
560 | // |
561 | // REQUIRES: a benchmark has exited its benchmarking loop. |
562 | BENCHMARK_ALWAYS_INLINE |
563 | void SetBytesProcessed(int64_t bytes) { |
564 | counters["bytes_per_second" ] = |
565 | Counter(static_cast<double>(bytes), Counter::kIsRate, Counter::kIs1024); |
566 | } |
567 | |
568 | BENCHMARK_ALWAYS_INLINE |
569 | int64_t bytes_processed() const { |
570 | if (counters.find("bytes_per_second" ) != counters.end()) |
571 | return static_cast<int64_t>(counters.at("bytes_per_second" )); |
572 | return 0; |
573 | } |
574 | |
575 | // If this routine is called with complexity_n > 0 and complexity report is |
576 | // requested for the |
577 | // family benchmark, then current benchmark will be part of the computation |
578 | // and complexity_n will |
579 | // represent the length of N. |
580 | BENCHMARK_ALWAYS_INLINE |
581 | void SetComplexityN(int64_t complexity_n) { complexity_n_ = complexity_n; } |
582 | |
583 | BENCHMARK_ALWAYS_INLINE |
584 | int64_t complexity_length_n() const { return complexity_n_; } |
585 | |
586 | // If this routine is called with items > 0, then an items/s |
587 | // label is printed on the benchmark report line for the currently |
588 | // executing benchmark. It is typically called at the end of a processing |
589 | // benchmark where a processing items/second output is desired. |
590 | // |
591 | // REQUIRES: a benchmark has exited its benchmarking loop. |
592 | BENCHMARK_ALWAYS_INLINE |
593 | void SetItemsProcessed(int64_t items) { |
594 | counters["items_per_second" ] = |
595 | Counter(static_cast<double>(items), benchmark::Counter::kIsRate); |
596 | } |
597 | |
598 | BENCHMARK_ALWAYS_INLINE |
599 | int64_t items_processed() const { |
600 | if (counters.find("items_per_second" ) != counters.end()) |
601 | return static_cast<int64_t>(counters.at("items_per_second" )); |
602 | return 0; |
603 | } |
604 | |
605 | // If this routine is called, the specified label is printed at the |
606 | // end of the benchmark report line for the currently executing |
607 | // benchmark. Example: |
608 | // static void BM_Compress(benchmark::State& state) { |
609 | // ... |
610 | // double compress = input_size / output_size; |
611 | // state.SetLabel(StrFormat("compress:%.1f%%", 100.0*compression)); |
612 | // } |
613 | // Produces output that looks like: |
614 | // BM_Compress 50 50 14115038 compress:27.3% |
615 | // |
616 | // REQUIRES: a benchmark has exited its benchmarking loop. |
617 | void SetLabel(const char* label); |
618 | |
619 | void BENCHMARK_ALWAYS_INLINE SetLabel(const std::string& str) { |
620 | this->SetLabel(str.c_str()); |
621 | } |
622 | |
623 | // Range arguments for this run. CHECKs if the argument has been set. |
624 | BENCHMARK_ALWAYS_INLINE |
625 | int64_t range(std::size_t pos = 0) const { |
626 | assert(range_.size() > pos); |
627 | return range_[pos]; |
628 | } |
629 | |
630 | BENCHMARK_DEPRECATED_MSG("use 'range(0)' instead" ) |
631 | int64_t range_x() const { return range(0); } |
632 | |
633 | BENCHMARK_DEPRECATED_MSG("use 'range(1)' instead" ) |
634 | int64_t range_y() const { return range(1); } |
635 | |
636 | BENCHMARK_ALWAYS_INLINE |
637 | IterationCount iterations() const { |
638 | if (BENCHMARK_BUILTIN_EXPECT(!started_, false)) { |
639 | return 0; |
640 | } |
641 | return max_iterations - total_iterations_ + batch_leftover_; |
642 | } |
643 | |
644 | private |
645 | : // items we expect on the first cache line (ie 64 bytes of the struct) |
646 | // When total_iterations_ is 0, KeepRunning() and friends will return false. |
647 | // May be larger than max_iterations. |
648 | IterationCount total_iterations_; |
649 | |
650 | // When using KeepRunningBatch(), batch_leftover_ holds the number of |
651 | // iterations beyond max_iters that were run. Used to track |
652 | // completed_iterations_ accurately. |
653 | IterationCount batch_leftover_; |
654 | |
655 | public: |
656 | const IterationCount max_iterations; |
657 | |
658 | private: |
659 | bool started_; |
660 | bool finished_; |
661 | bool error_occurred_; |
662 | |
663 | private: // items we don't need on the first cache line |
664 | std::vector<int64_t> range_; |
665 | |
666 | int64_t complexity_n_; |
667 | |
668 | public: |
669 | // Container for user-defined counters. |
670 | UserCounters counters; |
671 | // Index of the executing thread. Values from [0, threads). |
672 | const int thread_index; |
673 | // Number of threads concurrently executing the benchmark. |
674 | const int threads; |
675 | |
676 | private: |
677 | State(IterationCount max_iters, const std::vector<int64_t>& ranges, |
678 | int thread_i, int n_threads, internal::ThreadTimer* timer, |
679 | internal::ThreadManager* manager); |
680 | |
681 | void StartKeepRunning(); |
682 | // Implementation of KeepRunning() and KeepRunningBatch(). |
683 | // is_batch must be true unless n is 1. |
684 | bool KeepRunningInternal(IterationCount n, bool is_batch); |
685 | void FinishKeepRunning(); |
686 | internal::ThreadTimer* timer_; |
687 | internal::ThreadManager* manager_; |
688 | |
689 | friend struct internal::BenchmarkInstance; |
690 | }; |
691 | |
692 | inline BENCHMARK_ALWAYS_INLINE bool State::KeepRunning() { |
693 | return KeepRunningInternal(1, /*is_batch=*/false); |
694 | } |
695 | |
696 | inline BENCHMARK_ALWAYS_INLINE bool State::KeepRunningBatch(IterationCount n) { |
697 | return KeepRunningInternal(n, /*is_batch=*/true); |
698 | } |
699 | |
700 | inline BENCHMARK_ALWAYS_INLINE bool State::KeepRunningInternal(IterationCount n, |
701 | bool is_batch) { |
702 | // total_iterations_ is set to 0 by the constructor, and always set to a |
703 | // nonzero value by StartKepRunning(). |
704 | assert(n > 0); |
705 | // n must be 1 unless is_batch is true. |
706 | assert(is_batch || n == 1); |
707 | if (BENCHMARK_BUILTIN_EXPECT(total_iterations_ >= n, true)) { |
708 | total_iterations_ -= n; |
709 | return true; |
710 | } |
711 | if (!started_) { |
712 | StartKeepRunning(); |
713 | if (!error_occurred_ && total_iterations_ >= n) { |
714 | total_iterations_ -= n; |
715 | return true; |
716 | } |
717 | } |
718 | // For non-batch runs, total_iterations_ must be 0 by now. |
719 | if (is_batch && total_iterations_ != 0) { |
720 | batch_leftover_ = n - total_iterations_; |
721 | total_iterations_ = 0; |
722 | return true; |
723 | } |
724 | FinishKeepRunning(); |
725 | return false; |
726 | } |
727 | |
728 | struct State::StateIterator { |
729 | struct BENCHMARK_UNUSED Value {}; |
730 | typedef std::forward_iterator_tag iterator_category; |
731 | typedef Value value_type; |
732 | typedef Value reference; |
733 | typedef Value pointer; |
734 | typedef std::ptrdiff_t difference_type; |
735 | |
736 | private: |
737 | friend class State; |
738 | BENCHMARK_ALWAYS_INLINE |
739 | StateIterator() : cached_(0), parent_() {} |
740 | |
741 | BENCHMARK_ALWAYS_INLINE |
742 | explicit StateIterator(State* st) |
743 | : cached_(st->error_occurred_ ? 0 : st->max_iterations), parent_(st) {} |
744 | |
745 | public: |
746 | BENCHMARK_ALWAYS_INLINE |
747 | Value operator*() const { return Value(); } |
748 | |
749 | BENCHMARK_ALWAYS_INLINE |
750 | StateIterator& operator++() { |
751 | assert(cached_ > 0); |
752 | --cached_; |
753 | return *this; |
754 | } |
755 | |
756 | BENCHMARK_ALWAYS_INLINE |
757 | bool operator!=(StateIterator const&) const { |
758 | if (BENCHMARK_BUILTIN_EXPECT(cached_ != 0, true)) return true; |
759 | parent_->FinishKeepRunning(); |
760 | return false; |
761 | } |
762 | |
763 | private: |
764 | IterationCount cached_; |
765 | State* const parent_; |
766 | }; |
767 | |
768 | inline BENCHMARK_ALWAYS_INLINE State::StateIterator State::begin() { |
769 | return StateIterator(this); |
770 | } |
771 | inline BENCHMARK_ALWAYS_INLINE State::StateIterator State::end() { |
772 | StartKeepRunning(); |
773 | return StateIterator(); |
774 | } |
775 | |
776 | namespace internal { |
777 | |
778 | typedef void(Function)(State&); |
779 | |
780 | // ------------------------------------------------------ |
781 | // Benchmark registration object. The BENCHMARK() macro expands |
782 | // into an internal::Benchmark* object. Various methods can |
783 | // be called on this object to change the properties of the benchmark. |
784 | // Each method returns "this" so that multiple method calls can |
785 | // chained into one expression. |
786 | class Benchmark { |
787 | public: |
788 | virtual ~Benchmark(); |
789 | |
790 | // Note: the following methods all return "this" so that multiple |
791 | // method calls can be chained together in one expression. |
792 | |
793 | // Run this benchmark once with "x" as the extra argument passed |
794 | // to the function. |
795 | // REQUIRES: The function passed to the constructor must accept an arg1. |
796 | Benchmark* Arg(int64_t x); |
797 | |
798 | // Run this benchmark with the given time unit for the generated output report |
799 | Benchmark* Unit(TimeUnit unit); |
800 | |
801 | // Run this benchmark once for a number of values picked from the |
802 | // range [start..limit]. (start and limit are always picked.) |
803 | // REQUIRES: The function passed to the constructor must accept an arg1. |
804 | Benchmark* Range(int64_t start, int64_t limit); |
805 | |
806 | // Run this benchmark once for all values in the range [start..limit] with |
807 | // specific step |
808 | // REQUIRES: The function passed to the constructor must accept an arg1. |
809 | Benchmark* DenseRange(int64_t start, int64_t limit, int step = 1); |
810 | |
811 | // Run this benchmark once with "args" as the extra arguments passed |
812 | // to the function. |
813 | // REQUIRES: The function passed to the constructor must accept arg1, arg2 ... |
814 | Benchmark* Args(const std::vector<int64_t>& args); |
815 | |
816 | // Equivalent to Args({x, y}) |
817 | // NOTE: This is a legacy C++03 interface provided for compatibility only. |
818 | // New code should use 'Args'. |
819 | Benchmark* ArgPair(int64_t x, int64_t y) { |
820 | std::vector<int64_t> args; |
821 | args.push_back(x); |
822 | args.push_back(y); |
823 | return Args(args); |
824 | } |
825 | |
826 | // Run this benchmark once for a number of values picked from the |
827 | // ranges [start..limit]. (starts and limits are always picked.) |
828 | // REQUIRES: The function passed to the constructor must accept arg1, arg2 ... |
829 | Benchmark* Ranges(const std::vector<std::pair<int64_t, int64_t> >& ranges); |
830 | |
831 | // Run this benchmark once for each combination of values in the (cartesian) |
832 | // product of the supplied argument lists. |
833 | // REQUIRES: The function passed to the constructor must accept arg1, arg2 ... |
834 | Benchmark* ArgsProduct(const std::vector<std::vector<int64_t> >& arglists); |
835 | |
836 | // Equivalent to ArgNames({name}) |
837 | Benchmark* ArgName(const std::string& name); |
838 | |
839 | // Set the argument names to display in the benchmark name. If not called, |
840 | // only argument values will be shown. |
841 | Benchmark* ArgNames(const std::vector<std::string>& names); |
842 | |
843 | // Equivalent to Ranges({{lo1, hi1}, {lo2, hi2}}). |
844 | // NOTE: This is a legacy C++03 interface provided for compatibility only. |
845 | // New code should use 'Ranges'. |
846 | Benchmark* RangePair(int64_t lo1, int64_t hi1, int64_t lo2, int64_t hi2) { |
847 | std::vector<std::pair<int64_t, int64_t> > ranges; |
848 | ranges.push_back(std::make_pair(lo1, hi1)); |
849 | ranges.push_back(std::make_pair(lo2, hi2)); |
850 | return Ranges(ranges); |
851 | } |
852 | |
853 | // Pass this benchmark object to *func, which can customize |
854 | // the benchmark by calling various methods like Arg, Args, |
855 | // Threads, etc. |
856 | Benchmark* Apply(void (*func)(Benchmark* benchmark)); |
857 | |
858 | // Set the range multiplier for non-dense range. If not called, the range |
859 | // multiplier kRangeMultiplier will be used. |
860 | Benchmark* RangeMultiplier(int multiplier); |
861 | |
862 | // Set the minimum amount of time to use when running this benchmark. This |
863 | // option overrides the `benchmark_min_time` flag. |
864 | // REQUIRES: `t > 0` and `Iterations` has not been called on this benchmark. |
865 | Benchmark* MinTime(double t); |
866 | |
867 | // Specify the amount of iterations that should be run by this benchmark. |
868 | // REQUIRES: 'n > 0' and `MinTime` has not been called on this benchmark. |
869 | // |
870 | // NOTE: This function should only be used when *exact* iteration control is |
871 | // needed and never to control or limit how long a benchmark runs, where |
872 | // `--benchmark_min_time=N` or `MinTime(...)` should be used instead. |
873 | Benchmark* Iterations(IterationCount n); |
874 | |
875 | // Specify the amount of times to repeat this benchmark. This option overrides |
876 | // the `benchmark_repetitions` flag. |
877 | // REQUIRES: `n > 0` |
878 | Benchmark* Repetitions(int n); |
879 | |
880 | // Specify if each repetition of the benchmark should be reported separately |
881 | // or if only the final statistics should be reported. If the benchmark |
882 | // is not repeated then the single result is always reported. |
883 | // Applies to *ALL* reporters (display and file). |
884 | Benchmark* ReportAggregatesOnly(bool value = true); |
885 | |
886 | // Same as ReportAggregatesOnly(), but applies to display reporter only. |
887 | Benchmark* DisplayAggregatesOnly(bool value = true); |
888 | |
889 | // By default, the CPU time is measured only for the main thread, which may |
890 | // be unrepresentative if the benchmark uses threads internally. If called, |
891 | // the total CPU time spent by all the threads will be measured instead. |
892 | // By default, the only the main thread CPU time will be measured. |
893 | Benchmark* MeasureProcessCPUTime(); |
894 | |
895 | // If a particular benchmark should use the Wall clock instead of the CPU time |
896 | // (be it either the CPU time of the main thread only (default), or the |
897 | // total CPU usage of the benchmark), call this method. If called, the elapsed |
898 | // (wall) time will be used to control how many iterations are run, and in the |
899 | // printing of items/second or MB/seconds values. |
900 | // If not called, the CPU time used by the benchmark will be used. |
901 | Benchmark* UseRealTime(); |
902 | |
903 | // If a benchmark must measure time manually (e.g. if GPU execution time is |
904 | // being |
905 | // measured), call this method. If called, each benchmark iteration should |
906 | // call |
907 | // SetIterationTime(seconds) to report the measured time, which will be used |
908 | // to control how many iterations are run, and in the printing of items/second |
909 | // or MB/second values. |
910 | Benchmark* UseManualTime(); |
911 | |
912 | // Set the asymptotic computational complexity for the benchmark. If called |
913 | // the asymptotic computational complexity will be shown on the output. |
914 | Benchmark* Complexity(BigO complexity = benchmark::oAuto); |
915 | |
916 | // Set the asymptotic computational complexity for the benchmark. If called |
917 | // the asymptotic computational complexity will be shown on the output. |
918 | Benchmark* Complexity(BigOFunc* complexity); |
919 | |
920 | // Add this statistics to be computed over all the values of benchmark run |
921 | Benchmark* ComputeStatistics(std::string name, StatisticsFunc* statistics); |
922 | |
923 | // Support for running multiple copies of the same benchmark concurrently |
924 | // in multiple threads. This may be useful when measuring the scaling |
925 | // of some piece of code. |
926 | |
927 | // Run one instance of this benchmark concurrently in t threads. |
928 | Benchmark* Threads(int t); |
929 | |
930 | // Pick a set of values T from [min_threads,max_threads]. |
931 | // min_threads and max_threads are always included in T. Run this |
932 | // benchmark once for each value in T. The benchmark run for a |
933 | // particular value t consists of t threads running the benchmark |
934 | // function concurrently. For example, consider: |
935 | // BENCHMARK(Foo)->ThreadRange(1,16); |
936 | // This will run the following benchmarks: |
937 | // Foo in 1 thread |
938 | // Foo in 2 threads |
939 | // Foo in 4 threads |
940 | // Foo in 8 threads |
941 | // Foo in 16 threads |
942 | Benchmark* ThreadRange(int min_threads, int max_threads); |
943 | |
944 | // For each value n in the range, run this benchmark once using n threads. |
945 | // min_threads and max_threads are always included in the range. |
946 | // stride specifies the increment. E.g. DenseThreadRange(1, 8, 3) starts |
947 | // a benchmark with 1, 4, 7 and 8 threads. |
948 | Benchmark* DenseThreadRange(int min_threads, int max_threads, int stride = 1); |
949 | |
950 | // Equivalent to ThreadRange(NumCPUs(), NumCPUs()) |
951 | Benchmark* ThreadPerCpu(); |
952 | |
953 | virtual void Run(State& state) = 0; |
954 | |
955 | protected: |
956 | explicit Benchmark(const char* name); |
957 | Benchmark(Benchmark const&); |
958 | void SetName(const char* name); |
959 | |
960 | int ArgsCnt() const; |
961 | |
962 | private: |
963 | friend class BenchmarkFamilies; |
964 | |
965 | std::string name_; |
966 | AggregationReportMode aggregation_report_mode_; |
967 | std::vector<std::string> arg_names_; // Args for all benchmark runs |
968 | std::vector<std::vector<int64_t> > args_; // Args for all benchmark runs |
969 | TimeUnit time_unit_; |
970 | int range_multiplier_; |
971 | double min_time_; |
972 | IterationCount iterations_; |
973 | int repetitions_; |
974 | bool measure_process_cpu_time_; |
975 | bool use_real_time_; |
976 | bool use_manual_time_; |
977 | BigO complexity_; |
978 | BigOFunc* complexity_lambda_; |
979 | std::vector<Statistics> statistics_; |
980 | std::vector<int> thread_counts_; |
981 | |
982 | Benchmark& operator=(Benchmark const&); |
983 | }; |
984 | |
985 | } // namespace internal |
986 | |
987 | // Create and register a benchmark with the specified 'name' that invokes |
988 | // the specified functor 'fn'. |
989 | // |
990 | // RETURNS: A pointer to the registered benchmark. |
991 | internal::Benchmark* RegisterBenchmark(const char* name, |
992 | internal::Function* fn); |
993 | |
994 | #if defined(BENCHMARK_HAS_CXX11) |
995 | template <class Lambda> |
996 | internal::Benchmark* RegisterBenchmark(const char* name, Lambda&& fn); |
997 | #endif |
998 | |
999 | // Remove all registered benchmarks. All pointers to previously registered |
1000 | // benchmarks are invalidated. |
1001 | void ClearRegisteredBenchmarks(); |
1002 | |
1003 | namespace internal { |
1004 | // The class used to hold all Benchmarks created from static function. |
1005 | // (ie those created using the BENCHMARK(...) macros. |
1006 | class FunctionBenchmark : public Benchmark { |
1007 | public: |
1008 | FunctionBenchmark(const char* name, Function* func) |
1009 | : Benchmark(name), func_(func) {} |
1010 | |
1011 | virtual void Run(State& st); |
1012 | |
1013 | private: |
1014 | Function* func_; |
1015 | }; |
1016 | |
1017 | #ifdef BENCHMARK_HAS_CXX11 |
1018 | template <class Lambda> |
1019 | class LambdaBenchmark : public Benchmark { |
1020 | public: |
1021 | virtual void Run(State& st) { lambda_(st); } |
1022 | |
1023 | private: |
1024 | template <class OLambda> |
1025 | LambdaBenchmark(const char* name, OLambda&& lam) |
1026 | : Benchmark(name), lambda_(std::forward<OLambda>(lam)) {} |
1027 | |
1028 | LambdaBenchmark(LambdaBenchmark const&) = delete; |
1029 | |
1030 | private: |
1031 | template <class Lam> |
1032 | friend Benchmark* ::benchmark::RegisterBenchmark(const char*, Lam&&); |
1033 | |
1034 | Lambda lambda_; |
1035 | }; |
1036 | #endif |
1037 | |
1038 | } // namespace internal |
1039 | |
1040 | inline internal::Benchmark* RegisterBenchmark(const char* name, |
1041 | internal::Function* fn) { |
1042 | return internal::RegisterBenchmarkInternal( |
1043 | ::new internal::FunctionBenchmark(name, fn)); |
1044 | } |
1045 | |
1046 | #ifdef BENCHMARK_HAS_CXX11 |
1047 | template <class Lambda> |
1048 | internal::Benchmark* RegisterBenchmark(const char* name, Lambda&& fn) { |
1049 | using BenchType = |
1050 | internal::LambdaBenchmark<typename std::decay<Lambda>::type>; |
1051 | return internal::RegisterBenchmarkInternal( |
1052 | ::new BenchType(name, std::forward<Lambda>(fn))); |
1053 | } |
1054 | #endif |
1055 | |
1056 | #if defined(BENCHMARK_HAS_CXX11) && \ |
1057 | (!defined(BENCHMARK_GCC_VERSION) || BENCHMARK_GCC_VERSION >= 409) |
1058 | template <class Lambda, class... Args> |
1059 | internal::Benchmark* RegisterBenchmark(const char* name, Lambda&& fn, |
1060 | Args&&... args) { |
1061 | return benchmark::RegisterBenchmark( |
1062 | name, [=](benchmark::State& st) { fn(st, args...); }); |
1063 | } |
1064 | #else |
1065 | #define BENCHMARK_HAS_NO_VARIADIC_REGISTER_BENCHMARK |
1066 | #endif |
1067 | |
1068 | // The base class for all fixture tests. |
1069 | class Fixture : public internal::Benchmark { |
1070 | public: |
1071 | Fixture() : internal::Benchmark("" ) {} |
1072 | |
1073 | virtual void Run(State& st) { |
1074 | this->SetUp(st); |
1075 | this->BenchmarkCase(st); |
1076 | this->TearDown(st); |
1077 | } |
1078 | |
1079 | // These will be deprecated ... |
1080 | virtual void SetUp(const State&) {} |
1081 | virtual void TearDown(const State&) {} |
1082 | // ... In favor of these. |
1083 | virtual void SetUp(State& st) { SetUp(const_cast<const State&>(st)); } |
1084 | virtual void TearDown(State& st) { TearDown(const_cast<const State&>(st)); } |
1085 | |
1086 | protected: |
1087 | virtual void BenchmarkCase(State&) = 0; |
1088 | }; |
1089 | |
1090 | } // namespace benchmark |
1091 | |
1092 | // ------------------------------------------------------ |
1093 | // Macro to register benchmarks |
1094 | |
1095 | // Check that __COUNTER__ is defined and that __COUNTER__ increases by 1 |
1096 | // every time it is expanded. X + 1 == X + 0 is used in case X is defined to be |
1097 | // empty. If X is empty the expression becomes (+1 == +0). |
1098 | #if defined(__COUNTER__) && (__COUNTER__ + 1 == __COUNTER__ + 0) |
1099 | #define BENCHMARK_PRIVATE_UNIQUE_ID __COUNTER__ |
1100 | #else |
1101 | #define BENCHMARK_PRIVATE_UNIQUE_ID __LINE__ |
1102 | #endif |
1103 | |
1104 | // Helpers for generating unique variable names |
1105 | #define BENCHMARK_PRIVATE_NAME(n) \ |
1106 | BENCHMARK_PRIVATE_CONCAT(_benchmark_, BENCHMARK_PRIVATE_UNIQUE_ID, n) |
1107 | #define BENCHMARK_PRIVATE_CONCAT(a, b, c) BENCHMARK_PRIVATE_CONCAT2(a, b, c) |
1108 | #define BENCHMARK_PRIVATE_CONCAT2(a, b, c) a##b##c |
1109 | // Helper for concatenation with macro name expansion |
1110 | #define BENCHMARK_PRIVATE_CONCAT_NAME(BaseClass, Method) \ |
1111 | BaseClass##_##Method##_Benchmark |
1112 | |
1113 | #define BENCHMARK_PRIVATE_DECLARE(n) \ |
1114 | static ::benchmark::internal::Benchmark* BENCHMARK_PRIVATE_NAME(n) \ |
1115 | BENCHMARK_UNUSED |
1116 | |
1117 | #define BENCHMARK(n) \ |
1118 | BENCHMARK_PRIVATE_DECLARE(n) = \ |
1119 | (::benchmark::internal::RegisterBenchmarkInternal( \ |
1120 | new ::benchmark::internal::FunctionBenchmark(#n, n))) |
1121 | |
1122 | // Old-style macros |
1123 | #define BENCHMARK_WITH_ARG(n, a) BENCHMARK(n)->Arg((a)) |
1124 | #define BENCHMARK_WITH_ARG2(n, a1, a2) BENCHMARK(n)->Args({(a1), (a2)}) |
1125 | #define BENCHMARK_WITH_UNIT(n, t) BENCHMARK(n)->Unit((t)) |
1126 | #define BENCHMARK_RANGE(n, lo, hi) BENCHMARK(n)->Range((lo), (hi)) |
1127 | #define BENCHMARK_RANGE2(n, l1, h1, l2, h2) \ |
1128 | BENCHMARK(n)->RangePair({{(l1), (h1)}, {(l2), (h2)}}) |
1129 | |
1130 | #ifdef BENCHMARK_HAS_CXX11 |
1131 | |
1132 | // Register a benchmark which invokes the function specified by `func` |
1133 | // with the additional arguments specified by `...`. |
1134 | // |
1135 | // For example: |
1136 | // |
1137 | // template <class ...ExtraArgs>` |
1138 | // void BM_takes_args(benchmark::State& state, ExtraArgs&&... extra_args) { |
1139 | // [...] |
1140 | //} |
1141 | // /* Registers a benchmark named "BM_takes_args/int_string_test` */ |
1142 | // BENCHMARK_CAPTURE(BM_takes_args, int_string_test, 42, std::string("abc")); |
1143 | #define BENCHMARK_CAPTURE(func, test_case_name, ...) \ |
1144 | BENCHMARK_PRIVATE_DECLARE(func) = \ |
1145 | (::benchmark::internal::RegisterBenchmarkInternal( \ |
1146 | new ::benchmark::internal::FunctionBenchmark( \ |
1147 | #func "/" #test_case_name, \ |
1148 | [](::benchmark::State& st) { func(st, __VA_ARGS__); }))) |
1149 | |
1150 | #endif // BENCHMARK_HAS_CXX11 |
1151 | |
1152 | // This will register a benchmark for a templatized function. For example: |
1153 | // |
1154 | // template<int arg> |
1155 | // void BM_Foo(int iters); |
1156 | // |
1157 | // BENCHMARK_TEMPLATE(BM_Foo, 1); |
1158 | // |
1159 | // will register BM_Foo<1> as a benchmark. |
1160 | #define BENCHMARK_TEMPLATE1(n, a) \ |
1161 | BENCHMARK_PRIVATE_DECLARE(n) = \ |
1162 | (::benchmark::internal::RegisterBenchmarkInternal( \ |
1163 | new ::benchmark::internal::FunctionBenchmark(#n "<" #a ">", n<a>))) |
1164 | |
1165 | #define BENCHMARK_TEMPLATE2(n, a, b) \ |
1166 | BENCHMARK_PRIVATE_DECLARE(n) = \ |
1167 | (::benchmark::internal::RegisterBenchmarkInternal( \ |
1168 | new ::benchmark::internal::FunctionBenchmark(#n "<" #a "," #b ">", \ |
1169 | n<a, b>))) |
1170 | |
1171 | #ifdef BENCHMARK_HAS_CXX11 |
1172 | #define BENCHMARK_TEMPLATE(n, ...) \ |
1173 | BENCHMARK_PRIVATE_DECLARE(n) = \ |
1174 | (::benchmark::internal::RegisterBenchmarkInternal( \ |
1175 | new ::benchmark::internal::FunctionBenchmark( \ |
1176 | #n "<" #__VA_ARGS__ ">", n<__VA_ARGS__>))) |
1177 | #else |
1178 | #define BENCHMARK_TEMPLATE(n, a) BENCHMARK_TEMPLATE1(n, a) |
1179 | #endif |
1180 | |
1181 | #define BENCHMARK_PRIVATE_DECLARE_F(BaseClass, Method) \ |
1182 | class BaseClass##_##Method##_Benchmark : public BaseClass { \ |
1183 | public: \ |
1184 | BaseClass##_##Method##_Benchmark() : BaseClass() { \ |
1185 | this->SetName(#BaseClass "/" #Method); \ |
1186 | } \ |
1187 | \ |
1188 | protected: \ |
1189 | virtual void BenchmarkCase(::benchmark::State&); \ |
1190 | }; |
1191 | |
1192 | #define BENCHMARK_TEMPLATE1_PRIVATE_DECLARE_F(BaseClass, Method, a) \ |
1193 | class BaseClass##_##Method##_Benchmark : public BaseClass<a> { \ |
1194 | public: \ |
1195 | BaseClass##_##Method##_Benchmark() : BaseClass<a>() { \ |
1196 | this->SetName(#BaseClass "<" #a ">/" #Method); \ |
1197 | } \ |
1198 | \ |
1199 | protected: \ |
1200 | virtual void BenchmarkCase(::benchmark::State&); \ |
1201 | }; |
1202 | |
1203 | #define BENCHMARK_TEMPLATE2_PRIVATE_DECLARE_F(BaseClass, Method, a, b) \ |
1204 | class BaseClass##_##Method##_Benchmark : public BaseClass<a, b> { \ |
1205 | public: \ |
1206 | BaseClass##_##Method##_Benchmark() : BaseClass<a, b>() { \ |
1207 | this->SetName(#BaseClass "<" #a "," #b ">/" #Method); \ |
1208 | } \ |
1209 | \ |
1210 | protected: \ |
1211 | virtual void BenchmarkCase(::benchmark::State&); \ |
1212 | }; |
1213 | |
1214 | #ifdef BENCHMARK_HAS_CXX11 |
1215 | #define BENCHMARK_TEMPLATE_PRIVATE_DECLARE_F(BaseClass, Method, ...) \ |
1216 | class BaseClass##_##Method##_Benchmark : public BaseClass<__VA_ARGS__> { \ |
1217 | public: \ |
1218 | BaseClass##_##Method##_Benchmark() : BaseClass<__VA_ARGS__>() { \ |
1219 | this->SetName(#BaseClass "<" #__VA_ARGS__ ">/" #Method); \ |
1220 | } \ |
1221 | \ |
1222 | protected: \ |
1223 | virtual void BenchmarkCase(::benchmark::State&); \ |
1224 | }; |
1225 | #else |
1226 | #define BENCHMARK_TEMPLATE_PRIVATE_DECLARE_F(n, a) \ |
1227 | BENCHMARK_TEMPLATE1_PRIVATE_DECLARE_F(n, a) |
1228 | #endif |
1229 | |
1230 | #define BENCHMARK_DEFINE_F(BaseClass, Method) \ |
1231 | BENCHMARK_PRIVATE_DECLARE_F(BaseClass, Method) \ |
1232 | void BENCHMARK_PRIVATE_CONCAT_NAME(BaseClass, Method)::BenchmarkCase |
1233 | |
1234 | #define BENCHMARK_TEMPLATE1_DEFINE_F(BaseClass, Method, a) \ |
1235 | BENCHMARK_TEMPLATE1_PRIVATE_DECLARE_F(BaseClass, Method, a) \ |
1236 | void BENCHMARK_PRIVATE_CONCAT_NAME(BaseClass, Method)::BenchmarkCase |
1237 | |
1238 | #define BENCHMARK_TEMPLATE2_DEFINE_F(BaseClass, Method, a, b) \ |
1239 | BENCHMARK_TEMPLATE2_PRIVATE_DECLARE_F(BaseClass, Method, a, b) \ |
1240 | void BENCHMARK_PRIVATE_CONCAT_NAME(BaseClass, Method)::BenchmarkCase |
1241 | |
1242 | #ifdef BENCHMARK_HAS_CXX11 |
1243 | #define BENCHMARK_TEMPLATE_DEFINE_F(BaseClass, Method, ...) \ |
1244 | BENCHMARK_TEMPLATE_PRIVATE_DECLARE_F(BaseClass, Method, __VA_ARGS__) \ |
1245 | void BENCHMARK_PRIVATE_CONCAT_NAME(BaseClass, Method)::BenchmarkCase |
1246 | #else |
1247 | #define BENCHMARK_TEMPLATE_DEFINE_F(BaseClass, Method, a) \ |
1248 | BENCHMARK_TEMPLATE1_DEFINE_F(BaseClass, Method, a) |
1249 | #endif |
1250 | |
1251 | #define BENCHMARK_REGISTER_F(BaseClass, Method) \ |
1252 | BENCHMARK_PRIVATE_REGISTER_F(BENCHMARK_PRIVATE_CONCAT_NAME(BaseClass, Method)) |
1253 | |
1254 | #define BENCHMARK_PRIVATE_REGISTER_F(TestName) \ |
1255 | BENCHMARK_PRIVATE_DECLARE(TestName) = \ |
1256 | (::benchmark::internal::RegisterBenchmarkInternal(new TestName())) |
1257 | |
1258 | // This macro will define and register a benchmark within a fixture class. |
1259 | #define BENCHMARK_F(BaseClass, Method) \ |
1260 | BENCHMARK_PRIVATE_DECLARE_F(BaseClass, Method) \ |
1261 | BENCHMARK_REGISTER_F(BaseClass, Method); \ |
1262 | void BENCHMARK_PRIVATE_CONCAT_NAME(BaseClass, Method)::BenchmarkCase |
1263 | |
1264 | #define BENCHMARK_TEMPLATE1_F(BaseClass, Method, a) \ |
1265 | BENCHMARK_TEMPLATE1_PRIVATE_DECLARE_F(BaseClass, Method, a) \ |
1266 | BENCHMARK_REGISTER_F(BaseClass, Method); \ |
1267 | void BENCHMARK_PRIVATE_CONCAT_NAME(BaseClass, Method)::BenchmarkCase |
1268 | |
1269 | #define BENCHMARK_TEMPLATE2_F(BaseClass, Method, a, b) \ |
1270 | BENCHMARK_TEMPLATE2_PRIVATE_DECLARE_F(BaseClass, Method, a, b) \ |
1271 | BENCHMARK_REGISTER_F(BaseClass, Method); \ |
1272 | void BENCHMARK_PRIVATE_CONCAT_NAME(BaseClass, Method)::BenchmarkCase |
1273 | |
1274 | #ifdef BENCHMARK_HAS_CXX11 |
1275 | #define BENCHMARK_TEMPLATE_F(BaseClass, Method, ...) \ |
1276 | BENCHMARK_TEMPLATE_PRIVATE_DECLARE_F(BaseClass, Method, __VA_ARGS__) \ |
1277 | BENCHMARK_REGISTER_F(BaseClass, Method); \ |
1278 | void BENCHMARK_PRIVATE_CONCAT_NAME(BaseClass, Method)::BenchmarkCase |
1279 | #else |
1280 | #define BENCHMARK_TEMPLATE_F(BaseClass, Method, a) \ |
1281 | BENCHMARK_TEMPLATE1_F(BaseClass, Method, a) |
1282 | #endif |
1283 | |
1284 | // Helper macro to create a main routine in a test that runs the benchmarks |
1285 | #define BENCHMARK_MAIN() \ |
1286 | int main(int argc, char** argv) { \ |
1287 | ::benchmark::Initialize(&argc, argv); \ |
1288 | if (::benchmark::ReportUnrecognizedArguments(argc, argv)) return 1; \ |
1289 | ::benchmark::RunSpecifiedBenchmarks(); \ |
1290 | } \ |
1291 | int main(int, char**) |
1292 | |
1293 | // ------------------------------------------------------ |
1294 | // Benchmark Reporters |
1295 | |
1296 | namespace benchmark { |
1297 | |
1298 | struct CPUInfo { |
1299 | struct CacheInfo { |
1300 | std::string type; |
1301 | int level; |
1302 | int size; |
1303 | int num_sharing; |
1304 | }; |
1305 | |
1306 | enum Scaling { |
1307 | UNKNOWN, |
1308 | ENABLED, |
1309 | DISABLED |
1310 | }; |
1311 | |
1312 | int num_cpus; |
1313 | double cycles_per_second; |
1314 | std::vector<CacheInfo> caches; |
1315 | Scaling scaling; |
1316 | std::vector<double> load_avg; |
1317 | |
1318 | static const CPUInfo& Get(); |
1319 | |
1320 | private: |
1321 | CPUInfo(); |
1322 | BENCHMARK_DISALLOW_COPY_AND_ASSIGN(CPUInfo); |
1323 | }; |
1324 | |
1325 | // Adding Struct for System Information |
1326 | struct SystemInfo { |
1327 | std::string name; |
1328 | static const SystemInfo& Get(); |
1329 | |
1330 | private: |
1331 | SystemInfo(); |
1332 | BENCHMARK_DISALLOW_COPY_AND_ASSIGN(SystemInfo); |
1333 | }; |
1334 | |
1335 | // BenchmarkName contains the components of the Benchmark's name |
1336 | // which allows individual fields to be modified or cleared before |
1337 | // building the final name using 'str()'. |
1338 | struct BenchmarkName { |
1339 | std::string function_name; |
1340 | std::string args; |
1341 | std::string min_time; |
1342 | std::string iterations; |
1343 | std::string repetitions; |
1344 | std::string time_type; |
1345 | std::string threads; |
1346 | |
1347 | // Return the full name of the benchmark with each non-empty |
1348 | // field separated by a '/' |
1349 | std::string str() const; |
1350 | }; |
1351 | |
1352 | // Interface for custom benchmark result printers. |
1353 | // By default, benchmark reports are printed to stdout. However an application |
1354 | // can control the destination of the reports by calling |
1355 | // RunSpecifiedBenchmarks and passing it a custom reporter object. |
1356 | // The reporter object must implement the following interface. |
1357 | class BenchmarkReporter { |
1358 | public: |
1359 | struct Context { |
1360 | CPUInfo const& cpu_info; |
1361 | SystemInfo const& sys_info; |
1362 | // The number of chars in the longest benchmark name. |
1363 | size_t name_field_width; |
1364 | static const char* executable_name; |
1365 | Context(); |
1366 | }; |
1367 | |
1368 | struct Run { |
1369 | static const int64_t no_repetition_index = -1; |
1370 | enum RunType { RT_Iteration, RT_Aggregate }; |
1371 | |
1372 | Run() |
1373 | : run_type(RT_Iteration), |
1374 | error_occurred(false), |
1375 | iterations(1), |
1376 | threads(1), |
1377 | time_unit(kNanosecond), |
1378 | real_accumulated_time(0), |
1379 | cpu_accumulated_time(0), |
1380 | max_heapbytes_used(0), |
1381 | complexity(oNone), |
1382 | complexity_lambda(), |
1383 | complexity_n(0), |
1384 | report_big_o(false), |
1385 | report_rms(false), |
1386 | counters(), |
1387 | has_memory_result(false), |
1388 | allocs_per_iter(0.0), |
1389 | max_bytes_used(0) {} |
1390 | |
1391 | std::string benchmark_name() const; |
1392 | BenchmarkName run_name; |
1393 | RunType run_type; |
1394 | std::string aggregate_name; |
1395 | std::string report_label; // Empty if not set by benchmark. |
1396 | bool error_occurred; |
1397 | std::string error_message; |
1398 | |
1399 | IterationCount iterations; |
1400 | int64_t threads; |
1401 | int64_t repetition_index; |
1402 | int64_t repetitions; |
1403 | TimeUnit time_unit; |
1404 | double real_accumulated_time; |
1405 | double cpu_accumulated_time; |
1406 | |
1407 | // Return a value representing the real time per iteration in the unit |
1408 | // specified by 'time_unit'. |
1409 | // NOTE: If 'iterations' is zero the returned value represents the |
1410 | // accumulated time. |
1411 | double GetAdjustedRealTime() const; |
1412 | |
1413 | // Return a value representing the cpu time per iteration in the unit |
1414 | // specified by 'time_unit'. |
1415 | // NOTE: If 'iterations' is zero the returned value represents the |
1416 | // accumulated time. |
1417 | double GetAdjustedCPUTime() const; |
1418 | |
1419 | // This is set to 0.0 if memory tracing is not enabled. |
1420 | double max_heapbytes_used; |
1421 | |
1422 | // Keep track of arguments to compute asymptotic complexity |
1423 | BigO complexity; |
1424 | BigOFunc* complexity_lambda; |
1425 | int64_t complexity_n; |
1426 | |
1427 | // what statistics to compute from the measurements |
1428 | const std::vector<internal::Statistics>* statistics; |
1429 | |
1430 | // Inform print function whether the current run is a complexity report |
1431 | bool report_big_o; |
1432 | bool report_rms; |
1433 | |
1434 | UserCounters counters; |
1435 | |
1436 | // Memory metrics. |
1437 | bool has_memory_result; |
1438 | double allocs_per_iter; |
1439 | int64_t max_bytes_used; |
1440 | }; |
1441 | |
1442 | // Construct a BenchmarkReporter with the output stream set to 'std::cout' |
1443 | // and the error stream set to 'std::cerr' |
1444 | BenchmarkReporter(); |
1445 | |
1446 | // Called once for every suite of benchmarks run. |
1447 | // The parameter "context" contains information that the |
1448 | // reporter may wish to use when generating its report, for example the |
1449 | // platform under which the benchmarks are running. The benchmark run is |
1450 | // never started if this function returns false, allowing the reporter |
1451 | // to skip runs based on the context information. |
1452 | virtual bool ReportContext(const Context& context) = 0; |
1453 | |
1454 | // Called once for each group of benchmark runs, gives information about |
1455 | // cpu-time and heap memory usage during the benchmark run. If the group |
1456 | // of runs contained more than two entries then 'report' contains additional |
1457 | // elements representing the mean and standard deviation of those runs. |
1458 | // Additionally if this group of runs was the last in a family of benchmarks |
1459 | // 'reports' contains additional entries representing the asymptotic |
1460 | // complexity and RMS of that benchmark family. |
1461 | virtual void ReportRuns(const std::vector<Run>& report) = 0; |
1462 | |
1463 | // Called once and only once after ever group of benchmarks is run and |
1464 | // reported. |
1465 | virtual void Finalize() {} |
1466 | |
1467 | // REQUIRES: The object referenced by 'out' is valid for the lifetime |
1468 | // of the reporter. |
1469 | void SetOutputStream(std::ostream* out) { |
1470 | assert(out); |
1471 | output_stream_ = out; |
1472 | } |
1473 | |
1474 | // REQUIRES: The object referenced by 'err' is valid for the lifetime |
1475 | // of the reporter. |
1476 | void SetErrorStream(std::ostream* err) { |
1477 | assert(err); |
1478 | error_stream_ = err; |
1479 | } |
1480 | |
1481 | std::ostream& GetOutputStream() const { return *output_stream_; } |
1482 | |
1483 | std::ostream& GetErrorStream() const { return *error_stream_; } |
1484 | |
1485 | virtual ~BenchmarkReporter(); |
1486 | |
1487 | // Write a human readable string to 'out' representing the specified |
1488 | // 'context'. |
1489 | // REQUIRES: 'out' is non-null. |
1490 | static void PrintBasicContext(std::ostream* out, Context const& context); |
1491 | |
1492 | private: |
1493 | std::ostream* output_stream_; |
1494 | std::ostream* error_stream_; |
1495 | }; |
1496 | |
1497 | // Simple reporter that outputs benchmark data to the console. This is the |
1498 | // default reporter used by RunSpecifiedBenchmarks(). |
1499 | class ConsoleReporter : public BenchmarkReporter { |
1500 | public: |
1501 | enum OutputOptions { |
1502 | OO_None = 0, |
1503 | OO_Color = 1, |
1504 | OO_Tabular = 2, |
1505 | OO_ColorTabular = OO_Color | OO_Tabular, |
1506 | OO_Defaults = OO_ColorTabular |
1507 | }; |
1508 | explicit ConsoleReporter(OutputOptions opts_ = OO_Defaults) |
1509 | : output_options_(opts_), |
1510 | name_field_width_(0), |
1511 | prev_counters_(), |
1512 | printed_header_(false) {} |
1513 | |
1514 | virtual bool ReportContext(const Context& context); |
1515 | virtual void ReportRuns(const std::vector<Run>& reports); |
1516 | |
1517 | protected: |
1518 | virtual void PrintRunData(const Run& report); |
1519 | virtual void (const Run& report); |
1520 | |
1521 | OutputOptions output_options_; |
1522 | size_t name_field_width_; |
1523 | UserCounters prev_counters_; |
1524 | bool ; |
1525 | }; |
1526 | |
1527 | class JSONReporter : public BenchmarkReporter { |
1528 | public: |
1529 | JSONReporter() : first_report_(true) {} |
1530 | virtual bool ReportContext(const Context& context); |
1531 | virtual void ReportRuns(const std::vector<Run>& reports); |
1532 | virtual void Finalize(); |
1533 | |
1534 | private: |
1535 | void PrintRunData(const Run& report); |
1536 | |
1537 | bool first_report_; |
1538 | }; |
1539 | |
1540 | class BENCHMARK_DEPRECATED_MSG( |
1541 | "The CSV Reporter will be removed in a future release" ) CSVReporter |
1542 | : public BenchmarkReporter { |
1543 | public: |
1544 | CSVReporter() : printed_header_(false) {} |
1545 | virtual bool ReportContext(const Context& context); |
1546 | virtual void ReportRuns(const std::vector<Run>& reports); |
1547 | |
1548 | private: |
1549 | void PrintRunData(const Run& report); |
1550 | |
1551 | bool ; |
1552 | std::set<std::string> user_counter_names_; |
1553 | }; |
1554 | |
1555 | // If a MemoryManager is registered, it can be used to collect and report |
1556 | // allocation metrics for a run of the benchmark. |
1557 | class MemoryManager { |
1558 | public: |
1559 | struct Result { |
1560 | Result() : num_allocs(0), max_bytes_used(0) {} |
1561 | |
1562 | // The number of allocations made in total between Start and Stop. |
1563 | int64_t num_allocs; |
1564 | |
1565 | // The peak memory use between Start and Stop. |
1566 | int64_t max_bytes_used; |
1567 | }; |
1568 | |
1569 | virtual ~MemoryManager() {} |
1570 | |
1571 | // Implement this to start recording allocation information. |
1572 | virtual void Start() = 0; |
1573 | |
1574 | // Implement this to stop recording and fill out the given Result structure. |
1575 | virtual void Stop(Result* result) = 0; |
1576 | }; |
1577 | |
1578 | inline const char* GetTimeUnitString(TimeUnit unit) { |
1579 | switch (unit) { |
1580 | case kMillisecond: |
1581 | return "ms" ; |
1582 | case kMicrosecond: |
1583 | return "us" ; |
1584 | case kNanosecond: |
1585 | return "ns" ; |
1586 | } |
1587 | BENCHMARK_UNREACHABLE(); |
1588 | } |
1589 | |
1590 | inline double GetTimeUnitMultiplier(TimeUnit unit) { |
1591 | switch (unit) { |
1592 | case kMillisecond: |
1593 | return 1e3; |
1594 | case kMicrosecond: |
1595 | return 1e6; |
1596 | case kNanosecond: |
1597 | return 1e9; |
1598 | } |
1599 | BENCHMARK_UNREACHABLE(); |
1600 | } |
1601 | |
1602 | } // namespace benchmark |
1603 | |
1604 | #endif // BENCHMARK_BENCHMARK_H_ |
1605 | |