| Line |
Branch |
Exec |
Source |
| 1 |
|
|
/*************************************** |
| 2 |
|
|
Auteur : Pierre Aubert |
| 3 |
|
|
Mail : pierre.aubert@lapp.in2p3.fr |
| 4 |
|
|
Licence : CeCILL-C |
| 5 |
|
|
****************************************/ |
| 6 |
|
|
|
| 7 |
|
|
#ifndef __MICRO_BENCHMARK_NS_IMPL_H__ |
| 8 |
|
|
#define __MICRO_BENCHMARK_NS_IMPL_H__ |
| 9 |
|
|
|
| 10 |
|
|
#include <cmath> |
| 11 |
|
|
#include "micro_benchmark_ns.h" |
| 12 |
|
|
|
| 13 |
|
|
///Minimum time for performance test in ns |
| 14 |
|
|
#define PHOENIX_MINIMUM_TIME_NS 1000000000.0 |
| 15 |
|
|
|
| 16 |
|
|
///Prevent compiler from optimising a loop |
| 17 |
|
|
/** @param value : reference to a value |
| 18 |
|
|
*/ |
| 19 |
|
|
template <class T> |
| 20 |
|
|
[[gnu::always_inline]] inline void phoenix_doNotOptimize(T const& value) { |
| 21 |
|
|
asm volatile("" : : "r,m"(value) : "memory"); |
| 22 |
|
|
} |
| 23 |
|
|
|
| 24 |
|
|
///Prevent compiler from optimising a loop |
| 25 |
|
|
/** @param value : reference to a value |
| 26 |
|
|
*/ |
| 27 |
|
|
template <class T> |
| 28 |
|
✗ |
[[gnu::always_inline]] inline void phoenix_doNotOptimize(T & value) { |
| 29 |
|
|
#if defined(__clang__) |
| 30 |
|
|
asm volatile("" : "+r,m"(value) : : "memory"); |
| 31 |
|
|
#else |
| 32 |
|
✗ |
asm volatile("" : "+m,r"(value) : : "memory"); |
| 33 |
|
|
#endif |
| 34 |
|
|
} |
| 35 |
|
|
|
| 36 |
|
|
///Do the micro benchmarking of a given function and gives performance results in ns |
| 37 |
|
|
/** @param[out] ellapsedTimeNs : ellapsed time in ns |
| 38 |
|
|
* @param[out] ellapsedTimeErrorNs : error on the ellapsed time in ns |
| 39 |
|
|
* @param nbTestPerf : number of performance test |
| 40 |
|
|
* @param nbCallPerTest : number of calls per performance test |
| 41 |
|
|
* @param __f : function to be called and benchmarked |
| 42 |
|
|
* @param __args : parameter of the function to be benchmarked |
| 43 |
|
|
*/ |
| 44 |
|
|
template<typename _Callable, typename... _Args> |
| 45 |
|
✗ |
void micro_benchmarkNs(double & ellapsedTimeNs, double & ellapsedTimeErrorNs, size_t nbTestPerf, size_t nbCallPerTest, |
| 46 |
|
|
_Callable&& __f, _Args&&... __args) |
| 47 |
|
|
{ |
| 48 |
|
✗ |
VecEllapsedTime vecTimeNs; |
| 49 |
|
✗ |
int res = 0; |
| 50 |
|
✗ |
::phoenix_doNotOptimize(res); |
| 51 |
|
✗ |
for(size_t i(0lu); i < nbTestPerf; ++i){ |
| 52 |
|
|
//Stating the timer |
| 53 |
|
✗ |
HiPeTime beginTime = phoenix_getTime(); |
| 54 |
|
✗ |
for(size_t j(0lu); j < nbCallPerTest; ++j){ |
| 55 |
|
✗ |
::phoenix_doNotOptimize(res); |
| 56 |
|
✗ |
__f(__args...); |
| 57 |
|
|
} |
| 58 |
|
|
//Get the time of the nbCallPerTest calls |
| 59 |
|
✗ |
NanoSecs elapsedTime(phoenix_getTime() - beginTime); |
| 60 |
|
✗ |
double fullNs(elapsedTime.count()/((double)nbCallPerTest)); |
| 61 |
|
✗ |
vecTimeNs.push_back(fullNs); |
| 62 |
|
|
} |
| 63 |
|
✗ |
MapOrderedTime mapOrderTime; |
| 64 |
|
✗ |
micro_benchmarkVecToMap(mapOrderTime, vecTimeNs); |
| 65 |
|
✗ |
size_t nbValueUsed(vecTimeNs.size()*0.7 + 1lu); |
| 66 |
|
✗ |
if(nbValueUsed > vecTimeNs.size()){ |
| 67 |
|
✗ |
nbValueUsed = vecTimeNs.size(); |
| 68 |
|
|
} |
| 69 |
|
✗ |
micro_benchmarkComputeTime(ellapsedTimeNs, ellapsedTimeErrorNs, mapOrderTime, nbValueUsed); |
| 70 |
|
|
} |
| 71 |
|
|
|
| 72 |
|
|
|
| 73 |
|
|
///Do the micro benchmarking of a given function and gives performance results in ns |
| 74 |
|
|
/** @param[out] ellapsedTimeNs : ellapsed time in ns |
| 75 |
|
|
* @param[out] ellapsedTimeErrorNs : error on the ellapsed time in ns |
| 76 |
|
|
* @param __f : function to be called and benchmarked |
| 77 |
|
|
* @param __args : parameter of the function to be benchmarked |
| 78 |
|
|
* This function tries to find automatically a relevant performance measurement |
| 79 |
|
|
*/ |
| 80 |
|
|
template<typename _Callable, typename... _Args> |
| 81 |
|
✗ |
void micro_benchmarkAutoNs(double & ellapsedTimeNs, double & ellapsedTimeErrorNs, _Callable&& __f, _Args&&... __args){ |
| 82 |
|
✗ |
size_t nbTestPerf(100lu), nbCallPerTest(10lu); |
| 83 |
|
|
//Let's try with default values |
| 84 |
|
✗ |
micro_benchmarkNs(ellapsedTimeNs, ellapsedTimeErrorNs, nbTestPerf, nbCallPerTest, __f, __args...); |
| 85 |
|
|
// while(std::isnan(ellapsedTimeNs)){ |
| 86 |
|
|
// std::cout << "micro_benchmarkAutoNs : ellapsedTimeNs is NaN ! Adjusting nbTestPerf = " << nbTestPerf << std::endl; |
| 87 |
|
|
// nbCallPerTest += 5lu; |
| 88 |
|
|
// micro_benchmarkNs(ellapsedTimeNs, ellapsedTimeErrorNs, nbTestPerf, nbCallPerTest, __f, __args...); |
| 89 |
|
|
// std::cout << "micro_benchmarkAutoNs = nbTestPerf = "<<nbTestPerf<<", nbCallPerTest" << nbCallPerTest << std::endl; |
| 90 |
|
|
// } |
| 91 |
|
✗ |
double fullEllapsedTime(ellapsedTimeNs*((double)nbTestPerf)*((double)nbCallPerTest)); |
| 92 |
|
|
//Check if the total time is more than one second |
| 93 |
|
✗ |
while(fullEllapsedTime < PHOENIX_MINIMUM_TIME_NS && nbCallPerTest < 1000000000lu && nbCallPerTest > 0lu){ //Let's try again if the total time is less than one second |
| 94 |
|
|
//If the total time is less than one second, we try to change nbTestPerf and nbCallPerTest to make it about one second |
| 95 |
|
✗ |
double ratioTime((1.3*PHOENIX_MINIMUM_TIME_NS)/fullEllapsedTime); |
| 96 |
|
✗ |
if(ratioTime < 1.2){ |
| 97 |
|
✗ |
ratioTime = 1.2; |
| 98 |
|
|
} |
| 99 |
|
|
// double ratioSqrt(std::sqrt(ratioTime) + 1.0); |
| 100 |
|
|
//Let's modify the call |
| 101 |
|
|
// nbTestPerf *= ratioSqrt; |
| 102 |
|
|
// nbCallPerTest *= ratioSqrt; |
| 103 |
|
|
|
| 104 |
|
✗ |
nbCallPerTest *= ratioTime; |
| 105 |
|
|
|
| 106 |
|
|
//Let's try again |
| 107 |
|
✗ |
micro_benchmarkNs(ellapsedTimeNs, ellapsedTimeErrorNs, nbTestPerf, nbCallPerTest, __f, __args...); |
| 108 |
|
✗ |
fullEllapsedTime = ellapsedTimeNs*((double)nbTestPerf)*((double)nbCallPerTest); |
| 109 |
|
|
//We loop until we have one second of full time measurement |
| 110 |
|
|
} |
| 111 |
|
✗ |
std::cout << "micro_benchmarkAutoNs : nbCallPerTest = " << nbCallPerTest << std::endl; |
| 112 |
|
✗ |
if(nbCallPerTest > 1000000000lu || nbCallPerTest == 0lu){ |
| 113 |
|
✗ |
std::cout << "micro_benchmarkAutoNs : Warning invalid number of calls per test should lead to unrelevant results!!!" << std::endl; |
| 114 |
|
|
} |
| 115 |
|
|
} |
| 116 |
|
|
|
| 117 |
|
|
///Do the micro benchmarking of a given function and gives performance results in ns |
| 118 |
|
|
/** @param[out] ellapsedTimeNs : ellapsed time in ns |
| 119 |
|
|
* @param[out] ellapsedTimeErrorNs : error on the ellapsed time in ns |
| 120 |
|
|
* @param[out] timePerElement : time per element in ns |
| 121 |
|
|
* @param[out] timeErrorPerElement : error of time per element in ns |
| 122 |
|
|
* @param nbTestPerf : number of performance test |
| 123 |
|
|
* @param nbCallPerTest : number of calls per performance test |
| 124 |
|
|
* @param nbElement : number of element treated by the function __f |
| 125 |
|
|
* @param __f : function to be called and benchmarked |
| 126 |
|
|
* @param __args : parameter of the function to be benchmarked |
| 127 |
|
|
*/ |
| 128 |
|
|
template<typename _Callable, typename... _Args> |
| 129 |
|
✗ |
void micro_benchmarkNs(double & ellapsedTimeNs, double & ellapsedTimeErrorNs, |
| 130 |
|
|
double & timePerElement, double & timeErrorPerElement, size_t nbTestPerf, size_t nbCallPerTest, size_t nbElement, |
| 131 |
|
|
_Callable&& __f, _Args&&... __args) |
| 132 |
|
|
{ |
| 133 |
|
✗ |
micro_benchmarkNs(ellapsedTimeNs, ellapsedTimeErrorNs, nbTestPerf, nbCallPerTest, __f, __args...); |
| 134 |
|
✗ |
timePerElement = ellapsedTimeNs/((double)nbElement); |
| 135 |
|
✗ |
timeErrorPerElement = ellapsedTimeErrorNs/((double)nbElement); |
| 136 |
|
|
} |
| 137 |
|
|
|
| 138 |
|
|
///Do the micro benchmarking of a given function and gives performance results in ns |
| 139 |
|
|
/** @param[out] ellapsedTimeNs : ellapsed time in ns |
| 140 |
|
|
* @param[out] ellapsedTimeErrorNs : error on the ellapsed time in ns |
| 141 |
|
|
* @param[out] timePerElement : time per element in ns |
| 142 |
|
|
* @param[out] timeErrorPerElement : error of time per element in ns |
| 143 |
|
|
* @param nbElement : number of element treated by the function __f |
| 144 |
|
|
* @param __f : function to be called and benchmarked |
| 145 |
|
|
* @param __args : parameter of the function to be benchmarked |
| 146 |
|
|
*/ |
| 147 |
|
|
template<typename _Callable, typename... _Args> |
| 148 |
|
✗ |
void micro_benchmarkAutoNs(double & ellapsedTimeNs, double & ellapsedTimeErrorNs, |
| 149 |
|
|
double & timePerElement, double & timeErrorPerElement, size_t nbElement, |
| 150 |
|
|
_Callable&& __f, _Args&&... __args) |
| 151 |
|
|
{ |
| 152 |
|
✗ |
micro_benchmarkAutoNs(ellapsedTimeNs, ellapsedTimeErrorNs, __f, __args...); |
| 153 |
|
✗ |
timePerElement = ellapsedTimeNs/((double)nbElement); |
| 154 |
|
✗ |
timeErrorPerElement = ellapsedTimeErrorNs/((double)nbElement); |
| 155 |
|
|
} |
| 156 |
|
|
|
| 157 |
|
|
///Do the micro benchmarking of a given function and gives performance results in ns and print the result |
| 158 |
|
|
/** @param testName : name of the performance test |
| 159 |
|
|
* @param nbTestPerf : number of performance test |
| 160 |
|
|
* @param nbCallPerTest : number of calls per performance test |
| 161 |
|
|
* @param nbElement : number of element treated by the function __f |
| 162 |
|
|
* @param __f : function to be called and benchmarked |
| 163 |
|
|
* @param __args : parameter of the function to be benchmarked |
| 164 |
|
|
*/ |
| 165 |
|
|
template<typename _Callable, typename... _Args> |
| 166 |
|
✗ |
void micro_benchmarkNsPrint(const std::string & testName, size_t nbTestPerf, size_t nbCallPerTest, size_t nbElement, _Callable&& __f, _Args&&... __args){ |
| 167 |
|
✗ |
double ellapsedTimeNs(0.0), ellapsedTimeErrorNs(0.0), timePerElement(0.0), timeErrorPerElement(0.0); |
| 168 |
|
✗ |
micro_benchmarkNs(ellapsedTimeNs, ellapsedTimeErrorNs, timePerElement, timeErrorPerElement, nbTestPerf, nbCallPerTest, nbElement, __f, __args...); |
| 169 |
|
|
|
| 170 |
|
✗ |
std::cout << testName << " : nbElement = "<<nbElement<<", timePerElement = " << timePerElement << " ns/el ± "<<timeErrorPerElement<<", elapsedTime = " << ellapsedTimeNs << " ns ± "<<ellapsedTimeErrorNs << std::endl; |
| 171 |
|
✗ |
std::cerr << nbElement << "\t" << timePerElement << "\t" << ellapsedTimeNs << "\t" << timeErrorPerElement << "\t" << ellapsedTimeErrorNs << std::endl; |
| 172 |
|
|
} |
| 173 |
|
|
|
| 174 |
|
|
///Do the micro benchmarking of a given function and gives performance results in ns and print the result |
| 175 |
|
|
/** @param testName : name of the performance test |
| 176 |
|
|
* @param nbElement : number of element treated by the function __f |
| 177 |
|
|
* @param __f : function to be called and benchmarked |
| 178 |
|
|
* @param __args : parameter of the function to be benchmarked |
| 179 |
|
|
*/ |
| 180 |
|
|
template<typename _Callable, typename... _Args> |
| 181 |
|
✗ |
void micro_benchmarkAutoNsPrint(const std::string & testName, size_t nbElement, _Callable&& __f, _Args&&... __args){ |
| 182 |
|
✗ |
double ellapsedTimeNs(0.0), ellapsedTimeErrorNs(0.0), timePerElement(0.0), timeErrorPerElement(0.0); |
| 183 |
|
✗ |
micro_benchmarkAutoNs(ellapsedTimeNs, ellapsedTimeErrorNs, timePerElement, timeErrorPerElement, nbElement, __f, __args...); |
| 184 |
|
|
|
| 185 |
|
✗ |
std::cout << testName << " : nbElement = "<<nbElement<<", timePerElement = " << timePerElement << " ns/el ± "<<timeErrorPerElement<<", elapsedTime = " << ellapsedTimeNs << " ns ± "<<ellapsedTimeErrorNs << std::endl; |
| 186 |
|
✗ |
std::cerr << nbElement << "\t" << timePerElement << "\t" << ellapsedTimeNs << "\t" << timeErrorPerElement << "\t" << ellapsedTimeErrorNs << std::endl; |
| 187 |
|
|
} |
| 188 |
|
|
|
| 189 |
|
|
#endif |
| 190 |
|
|
|