1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227
|
// Usage: benchmark_branchfree [u32] [u64] [s32] [s64] [branchfree] [branchfull] [sys|system]
//
// The branchfree benchmark iterates over an array of dividers and computes
// divisions. This is the use case where the branchfree divider generally
// shines and where the default branchfull divider performs poorly because
// the CPU is not able to correctly predict the branches of the many different
// dividers.
//
#include "libdivide.h"
#include <algorithm>
#include <cstdlib>
#include <iostream>
#include <chrono>
#include <vector>
#include <string>
#include <typeinfo>
#if defined(__GNUC__)
#define NOINLINE __attribute__((__noinline__))
#elif defined(_MSC_VER)
#define NOINLINE __declspec(noinline)
#else
#define NOINLINE
#endif
// Generate primes using the sieve of Eratosthenes.
// These primes will later be used as dividers in the benchmark.
template <typename divider_type, typename T>
std::vector<divider_type> get_primes(T max)
{
uint64_t n = (uint64_t) max;
std::vector<divider_type> primes;
std::vector<char> sieve(n + 1, true);
for (uint64_t i = 2; i * i <= n; i++)
if (sieve[i])
for (uint64_t j = i * i; j <= n; j += i)
sieve[j] = false;
for (uint64_t i = 2; i <= n; i++)
if (sieve[i])
primes.push_back((T) i);
return primes;
}
// Here we iterate over an array of dividers and compute divisions.
// libdivide's branchfull divider will not perform well as the
// CPU will not be able to correctly predict the branches.
// The branchfree divider is perfectly suited for this use case
// and will perform much better.
//
template<typename N, typename T>
NOINLINE size_t sum_dividers(N numerator, const T& dividers)
{
size_t sum = 0;
for (const auto& divider: dividers)
sum += numerator / divider;
return sum;
}
struct result_t {
double duration;
size_t sum;
};
template<typename T, typename D>
NOINLINE result_t benchmark_sum_dividers(const D& dividers, size_t iters) {
auto t1 = std::chrono::system_clock::now();
size_t sum = 0;
for (; iters > 0; iters--) {
// Unsigned branchfree divider cannot be 1
T numerator = std::max((T) 2, (T) iters);
sum += sum_dividers(numerator, dividers);
}
auto t2 = std::chrono::system_clock::now();
std::chrono::duration<double> seconds = t2 - t1;
return result_t{seconds.count(), sum};
}
enum {
TEST_U32 = 1 << 0,
TEST_U64 = 1 << 1,
TEST_S32 = 1 << 2,
TEST_S64 = 1 << 3,
TEST_ALL_TYPES = (TEST_U32 |
TEST_U64 |
TEST_S32 |
TEST_S64),
TEST_SYSTEM = 1 << 4,
TEST_BRANCHFREE = 1 << 5,
TEST_BRANCHFULL = 1 << 6,
TEST_ALL_ALGOS = (TEST_SYSTEM |
TEST_BRANCHFREE |
TEST_BRANCHFULL),
};
using tasks_t = unsigned int;
template<typename T>
void benchmark(tasks_t tasks, size_t max, size_t iters) {
bool test_system = !!(tasks & TEST_SYSTEM);
bool test_branchfull = !!(tasks & TEST_BRANCHFULL);
bool test_branchfree = !!(tasks & TEST_BRANCHFREE);
result_t sys = {0, 0};
result_t branchfull = {0, 0};
result_t branchfree = {0, 0};
if (test_system) {
using divider_type = T;
auto dividers = get_primes<divider_type>((T) max);
sys = benchmark_sum_dividers<T>(dividers, iters);
std::cout << '.' << std::flush;
}
if (test_branchfull) {
using divider_type = libdivide::divider<T>;
auto dividers = get_primes<divider_type>((T) max);
branchfull = benchmark_sum_dividers<T>(dividers, iters);
std::cout << '.' << std::flush;
}
if (test_branchfree) {
using divider_type = libdivide::branchfree_divider<T>;
auto dividers = get_primes<divider_type>((T) max);
branchfree = benchmark_sum_dividers<T>(dividers, iters);
std::cout << '.' << std::endl;
}
if (test_system && test_branchfull && branchfull.sum != sys.sum) {
std::cerr << "Error: branchfull_divider<" << typeid(T).name() << "> sum: " << branchfull.sum << ", but system sum: " << sys.sum << std::endl;
std::exit(1);
}
if (test_system && test_branchfree && branchfree.sum != sys.sum) {
std::cerr << "Error: branchfree_divider<" << typeid(T).name() << "> sum: " << branchfree.sum << ", but system sum: " << sys.sum << std::endl;
std::exit(1);
}
if (test_system)
std::cout << " system: " << sys.duration << " seconds" << std::endl;
if (test_branchfull)
std::cout << "branchfull: " << branchfull.duration << " seconds" << std::endl;
if (test_branchfree)
std::cout << "branchfree: " << branchfree.duration << " seconds" << std::endl;
std::cout << std::endl;
}
void usage() {
std::cout << "Usage: benchmark_branchfree [u32] [u64] [s32] [s64] [branchfree] [branchfull] [sys|system]\n"
"\n"
"The branchfree benchmark iterates over an array of dividers and computes\n"
"divisions. This is the use case where the branchfree divider generally\n"
"shines and where the default branchfull divider performs poorly because\n"
"the CPU is not able to correctly predict the branches of the many different\n"
"dividers." << std::endl;
}
int main(int argc, const char *argv[]) {
tasks_t tasks = 0;
for (int i = 1; i < argc; i++) {
std::string arg(argv[i]);
if (arg == "u32") {
tasks |= TEST_U32;
} else if (arg == "s32") {
tasks |= TEST_S32;
} else if (arg == "u64") {
tasks |= TEST_U64;
} else if (arg == "s64") {
tasks |= TEST_S64;
} else if (arg == "branchfree") {
tasks |= TEST_BRANCHFREE;
} else if (arg == "branchfull") {
tasks |= TEST_BRANCHFULL;
} else if (arg == "sys" || arg == "system") {
tasks |= TEST_SYSTEM;
} else {
usage();
return 1;
}
}
// Set default tasks
if (!(tasks & TEST_ALL_TYPES)) {
tasks |= TEST_ALL_TYPES;
}
if (!(tasks & TEST_ALL_ALGOS)) {
tasks |= TEST_ALL_ALGOS;
}
size_t iters = 3000;
size_t max_divider = 1 << 22;
if (tasks & TEST_U32) {
std::cout << "----- u32 -----" << std::endl;
benchmark<uint32_t>(tasks, max_divider, iters);
}
if (tasks & TEST_S32) {
std::cout << "----- s32 -----" << std::endl;
benchmark<int32_t>(tasks, max_divider, iters);
}
if (tasks & TEST_U64) {
std::cout << "----- u64 -----" << std::endl;
benchmark<uint64_t>(tasks, max_divider, iters);
}
if (tasks & TEST_S64) {
std::cout << "----- s64 -----" << std::endl;
benchmark<int64_t>(tasks, max_divider, iters);
}
std::cout << "All tests passed successfully!" << std::endl;
return 0;
}
|