2 * Copyright 2016 Facebook, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
17 #ifndef FOLLY_BENCHMARK_H_
18 #define FOLLY_BENCHMARK_H_
20 #include <folly/ScopeGuard.h>
21 #include <folly/Portability.h>
22 #include <folly/Preprocessor.h> // for FB_ANONYMOUS_VARIABLE
25 #include <boost/function_types/function_arity.hpp>
27 #include <glog/logging.h>
28 #include <gflags/gflags.h>
30 #include <type_traits>
32 DECLARE_bool(benchmark);
37 * Runs all benchmarks defined. Usually put in main().
42 * Runs all benchmarks defined if and only if the --benchmark flag has
43 * been passed to the program. Usually put in main().
45 inline bool runBenchmarksOnFlag() {
46 if (FLAGS_benchmark) {
49 return FLAGS_benchmark;
55 * This is the clock ID used for measuring time. On older kernels, the
56 * resolution of this clock will be very coarse, which will cause the
59 enum Clock { DEFAULT_CLOCK_ID = CLOCK_REALTIME };
61 typedef std::pair<uint64_t, unsigned int> TimeIterPair;
64 * Adds a benchmark wrapped in a std::function. Only used
65 * internally. Pass by value is intentional.
67 void addBenchmarkImpl(const char* file,
69 std::function<TimeIterPair(unsigned int)>);
72 * Takes the difference between two timespec values. end is assumed to
75 inline uint64_t timespecDiff(timespec end, timespec start) {
76 if (end.tv_sec == start.tv_sec) {
77 assert(end.tv_nsec >= start.tv_nsec);
78 return end.tv_nsec - start.tv_nsec;
80 assert(end.tv_sec > start.tv_sec);
81 auto diff = uint64_t(end.tv_sec - start.tv_sec);
83 std::numeric_limits<uint64_t>::max() / 1000000000UL);
84 return diff * 1000000000UL
85 + end.tv_nsec - start.tv_nsec;
89 * Takes the difference between two sets of timespec values. The first
90 * two come from a high-resolution clock whereas the other two come
91 * from a low-resolution clock. The crux of the matter is that
92 * high-res values may be bogus as documented in
93 * http://linux.die.net/man/3/clock_gettime. The trouble is when the
94 * running process migrates from one CPU to another, which is more
95 * likely for long-running processes. Therefore we watch for high
96 * differences between the two timings.
98 * This function is subject to further improvements.
100 inline uint64_t timespecDiff(timespec end, timespec start,
101 timespec endCoarse, timespec startCoarse) {
102 auto fine = timespecDiff(end, start);
103 auto coarse = timespecDiff(endCoarse, startCoarse);
104 if (coarse - fine >= 1000000) {
105 // The fine time is in all likelihood bogus
111 } // namespace detail
114 * Supporting type for BENCHMARK_SUSPEND defined below.
116 struct BenchmarkSuspender {
117 BenchmarkSuspender() {
118 CHECK_EQ(0, clock_gettime(detail::DEFAULT_CLOCK_ID, &start));
121 BenchmarkSuspender(const BenchmarkSuspender &) = delete;
122 BenchmarkSuspender(BenchmarkSuspender && rhs) noexcept {
124 rhs.start.tv_nsec = rhs.start.tv_sec = 0;
127 BenchmarkSuspender& operator=(const BenchmarkSuspender &) = delete;
128 BenchmarkSuspender& operator=(BenchmarkSuspender && rhs) {
129 if (start.tv_nsec > 0 || start.tv_sec > 0) {
133 rhs.start.tv_nsec = rhs.start.tv_sec = 0;
137 ~BenchmarkSuspender() {
138 if (start.tv_nsec > 0 || start.tv_sec > 0) {
144 assert(start.tv_nsec > 0 || start.tv_sec > 0);
146 start.tv_nsec = start.tv_sec = 0;
150 assert(start.tv_nsec == 0 || start.tv_sec == 0);
151 CHECK_EQ(0, clock_gettime(detail::DEFAULT_CLOCK_ID, &start));
155 auto dismissing(F f) -> typename std::result_of<F()>::type {
156 SCOPE_EXIT { rehire(); };
162 * This is for use inside of if-conditions, used in BENCHMARK macros.
163 * If-conditions bypass the explicit on operator bool.
165 explicit operator bool() const {
170 * Accumulates nanoseconds spent outside benchmark.
172 typedef uint64_t NanosecondsSpent;
173 static NanosecondsSpent nsSpent;
178 CHECK_EQ(0, clock_gettime(detail::DEFAULT_CLOCK_ID, &end));
179 nsSpent += detail::timespecDiff(end, start);
187 * Adds a benchmark. Usually not called directly but instead through
188 * the macro BENCHMARK defined below. The lambda function involved
189 * must take exactly one parameter of type unsigned, and the benchmark
190 * uses it with counter semantics (iteration occurs inside the
193 template <typename Lambda>
194 typename std::enable_if<
195 boost::function_types::function_arity<decltype(&Lambda::operator())>::value
198 addBenchmark(const char* file, const char* name, Lambda&& lambda) {
199 auto execute = [=](unsigned int times) {
200 BenchmarkSuspender::nsSpent = 0;
204 // CORE MEASUREMENT STARTS
205 auto const r1 = clock_gettime(detail::DEFAULT_CLOCK_ID, &start);
206 niter = lambda(times);
207 auto const r2 = clock_gettime(detail::DEFAULT_CLOCK_ID, &end);
208 // CORE MEASUREMENT ENDS
213 return detail::TimeIterPair(
214 detail::timespecDiff(end, start) - BenchmarkSuspender::nsSpent,
218 detail::addBenchmarkImpl(file, name,
219 std::function<detail::TimeIterPair(unsigned int)>(execute));
223 * Adds a benchmark. Usually not called directly but instead through
224 * the macro BENCHMARK defined below. The lambda function involved
225 * must take zero parameters, and the benchmark calls it repeatedly
226 * (iteration occurs outside the function).
228 template <typename Lambda>
229 typename std::enable_if<
230 boost::function_types::function_arity<decltype(&Lambda::operator())>::value
233 addBenchmark(const char* file, const char* name, Lambda&& lambda) {
234 addBenchmark(file, name, [=](unsigned int times) {
235 unsigned int niter = 0;
236 while (times-- > 0) {
244 * Call doNotOptimizeAway(var) against variables that you use for
245 * benchmarking but otherwise are useless. The compiler tends to do a
246 * good job at eliminating unused variables, and this function fools
247 * it into thinking var is in fact needed.
251 #pragma optimize("", off)
254 void doNotOptimizeAway(T&& datum) {
258 #pragma optimize("", on)
260 #elif defined(__clang__)
263 __attribute__((__optnone__)) void doNotOptimizeAway(T&& /* datum */) {}
268 void doNotOptimizeAway(T&& datum) {
269 asm volatile("" : "+r" (datum));
277 * Introduces a benchmark function. Used internally, see BENCHMARK and
280 #define BENCHMARK_IMPL(funName, stringName, rv, paramType, paramName) \
281 static void funName(paramType); \
282 static bool FB_ANONYMOUS_VARIABLE(follyBenchmarkUnused) = ( \
283 ::folly::addBenchmark(__FILE__, stringName, \
284 [](paramType paramName) -> unsigned { funName(paramName); \
287 static void funName(paramType paramName)
290 * Introduces a benchmark function with support for returning the actual
291 * number of iterations. Used internally, see BENCHMARK_MULTI and friends
294 #define BENCHMARK_MULTI_IMPL(funName, stringName, paramType, paramName) \
295 static unsigned funName(paramType); \
296 static bool FB_ANONYMOUS_VARIABLE(follyBenchmarkUnused) = ( \
297 ::folly::addBenchmark(__FILE__, stringName, \
298 [](paramType paramName) { return funName(paramName); }), \
300 static unsigned funName(paramType paramName)
303 * Introduces a benchmark function. Use with either one or two arguments.
304 * The first is the name of the benchmark. Use something descriptive, such
305 * as insertVectorBegin. The second argument may be missing, or could be a
306 * symbolic counter. The counter dictates how many internal iteration the
307 * benchmark does. Example:
309 * BENCHMARK(vectorPushBack) {
314 * BENCHMARK(insertVectorBegin, n) {
316 * FOR_EACH_RANGE (i, 0, n) {
317 * v.insert(v.begin(), 42);
321 #define BENCHMARK(name, ...) \
324 FB_STRINGIZE(name), \
325 FB_ARG_2_OR_1(1, ## __VA_ARGS__), \
326 FB_ONE_OR_NONE(unsigned, ## __VA_ARGS__), \
330 * Like BENCHMARK above, but allows the user to return the actual
331 * number of iterations executed in the function body. This can be
332 * useful if the benchmark function doesn't know upfront how many
333 * iterations it's going to run or if it runs through a certain
334 * number of test cases, e.g.:
336 * BENCHMARK_MULTI(benchmarkSomething) {
337 * std::vector<int> testCases { 0, 1, 1, 2, 3, 5 };
338 * for (int c : testCases) {
341 * return testCases.size();
344 #define BENCHMARK_MULTI(name, ...) \
345 BENCHMARK_MULTI_IMPL( \
347 FB_STRINGIZE(name), \
348 FB_ONE_OR_NONE(unsigned, ## __VA_ARGS__), \
352 * Defines a benchmark that passes a parameter to another one. This is
353 * common for benchmarks that need a "problem size" in addition to
354 * "number of iterations". Consider:
356 * void pushBack(uint n, size_t initialSize) {
358 * BENCHMARK_SUSPEND {
359 * v.resize(initialSize);
361 * FOR_EACH_RANGE (i, 0, n) {
365 * BENCHMARK_PARAM(pushBack, 0)
366 * BENCHMARK_PARAM(pushBack, 1000)
367 * BENCHMARK_PARAM(pushBack, 1000000)
369 * The benchmark above estimates the speed of push_back at different
370 * initial sizes of the vector. The framework will pass 0, 1000, and
371 * 1000000 for initialSize, and the iteration count for n.
373 #define BENCHMARK_PARAM(name, param) \
374 BENCHMARK_NAMED_PARAM(name, param, param)
377 * Same as BENCHMARK_PARAM, but allows to return the actual number of
378 * iterations that have been run.
380 #define BENCHMARK_PARAM_MULTI(name, param) \
381 BENCHMARK_NAMED_PARAM_MULTI(name, param, param)
384 * Like BENCHMARK_PARAM(), but allows a custom name to be specified for each
385 * parameter, rather than using the parameter value.
387 * Useful when the parameter value is not a valid token for string pasting,
388 * of when you want to specify multiple parameter arguments.
392 * void addValue(uint n, int64_t bucketSize, int64_t min, int64_t max) {
393 * Histogram<int64_t> hist(bucketSize, min, max);
395 * FOR_EACH_RANGE (i, 0, n) {
396 * hist.addValue(num);
398 * if (num > max) { num = min; }
402 * BENCHMARK_NAMED_PARAM(addValue, 0_to_100, 1, 0, 100)
403 * BENCHMARK_NAMED_PARAM(addValue, 0_to_1000, 10, 0, 1000)
404 * BENCHMARK_NAMED_PARAM(addValue, 5k_to_20k, 250, 5000, 20000)
406 #define BENCHMARK_NAMED_PARAM(name, param_name, ...) \
408 FB_CONCATENATE(name, FB_CONCATENATE(_, param_name)), \
409 FB_STRINGIZE(name) "(" FB_STRINGIZE(param_name) ")", \
413 name(iters, ## __VA_ARGS__); \
417 * Same as BENCHMARK_NAMED_PARAM, but allows to return the actual number
418 * of iterations that have been run.
420 #define BENCHMARK_NAMED_PARAM_MULTI(name, param_name, ...) \
421 BENCHMARK_MULTI_IMPL( \
422 FB_CONCATENATE(name, FB_CONCATENATE(_, param_name)), \
423 FB_STRINGIZE(name) "(" FB_STRINGIZE(param_name) ")", \
426 return name(iters, ## __VA_ARGS__); \
430 * Just like BENCHMARK, but prints the time relative to a
431 * baseline. The baseline is the most recent BENCHMARK() seen in
432 * lexical order. Example:
434 * // This is the baseline
435 * BENCHMARK(insertVectorBegin, n) {
437 * FOR_EACH_RANGE (i, 0, n) {
438 * v.insert(v.begin(), 42);
442 * BENCHMARK_RELATIVE(insertListBegin, n) {
444 * FOR_EACH_RANGE (i, 0, n) {
445 * s.insert(s.begin(), 42);
449 * Any number of relative benchmark can be associated with a
450 * baseline. Another BENCHMARK() occurrence effectively establishes a
453 #define BENCHMARK_RELATIVE(name, ...) \
456 "%" FB_STRINGIZE(name), \
457 FB_ARG_2_OR_1(1, ## __VA_ARGS__), \
458 FB_ONE_OR_NONE(unsigned, ## __VA_ARGS__), \
462 * Same as BENCHMARK_RELATIVE, but allows to return the actual number
463 * of iterations that have been run.
465 #define BENCHMARK_RELATIVE_MULTI(name, ...) \
466 BENCHMARK_MULTI_IMPL( \
468 "%" FB_STRINGIZE(name), \
469 FB_ONE_OR_NONE(unsigned, ## __VA_ARGS__), \
473 * A combination of BENCHMARK_RELATIVE and BENCHMARK_PARAM.
475 #define BENCHMARK_RELATIVE_PARAM(name, param) \
476 BENCHMARK_RELATIVE_NAMED_PARAM(name, param, param)
479 * Same as BENCHMARK_RELATIVE_PARAM, but allows to return the actual
480 * number of iterations that have been run.
482 #define BENCHMARK_RELATIVE_PARAM_MULTI(name, param) \
483 BENCHMARK_RELATIVE_NAMED_PARAM_MULTI(name, param, param)
486 * A combination of BENCHMARK_RELATIVE and BENCHMARK_NAMED_PARAM.
488 #define BENCHMARK_RELATIVE_NAMED_PARAM(name, param_name, ...) \
490 FB_CONCATENATE(name, FB_CONCATENATE(_, param_name)), \
491 "%" FB_STRINGIZE(name) "(" FB_STRINGIZE(param_name) ")", \
495 name(iters, ## __VA_ARGS__); \
499 * Same as BENCHMARK_RELATIVE_NAMED_PARAM, but allows to return the
500 * actual number of iterations that have been run.
502 #define BENCHMARK_RELATIVE_NAMED_PARAM_MULTI(name, param_name, ...) \
503 BENCHMARK_MULTI_IMPL( \
504 FB_CONCATENATE(name, FB_CONCATENATE(_, param_name)), \
505 "%" FB_STRINGIZE(name) "(" FB_STRINGIZE(param_name) ")", \
508 return name(iters, ## __VA_ARGS__); \
512 * Draws a line of dashes.
514 #define BENCHMARK_DRAW_LINE() \
515 static bool FB_ANONYMOUS_VARIABLE(follyBenchmarkUnused) = ( \
516 ::folly::addBenchmark(__FILE__, "-", []() -> unsigned { return 0; }), \
520 * Allows execution of code that doesn't count torward the benchmark's
521 * time budget. Example:
523 * BENCHMARK_START_GROUP(insertVectorBegin, n) {
525 * BENCHMARK_SUSPEND {
528 * FOR_EACH_RANGE (i, 0, n) {
529 * v.insert(v.begin(), 42);
533 #define BENCHMARK_SUSPEND \
534 if (auto FB_ANONYMOUS_VARIABLE(BENCHMARK_SUSPEND) = \
535 ::folly::BenchmarkSuspender()) {} \
538 #endif // FOLLY_BENCHMARK_H_