[c4fd4ef] | 1 | #include <array> |
---|
| 2 | #include <iomanip> |
---|
| 3 | #include <iostream> |
---|
| 4 | #include <locale> |
---|
| 5 | #include <string> |
---|
| 6 | #include <thread> |
---|
| 7 | #include <vector> |
---|
| 8 | |
---|
| 9 | #include <getopt.h> |
---|
| 10 | #include <unistd.h> |
---|
| 11 | #include <sys/sysinfo.h> |
---|
| 12 | |
---|
| 13 | #include "utils.hpp" |
---|
| 14 | |
---|
| 15 | // ================================================================================================ |
---|
| 16 | // UTILS |
---|
| 17 | // ================================================================================================ |
---|
| 18 | |
---|
| 19 | struct local_stat_t { |
---|
| 20 | size_t cnt = 0; |
---|
| 21 | }; |
---|
| 22 | |
---|
| 23 | struct global_stat_t { |
---|
| 24 | std::atomic_size_t cnt = { 0 }; |
---|
| 25 | }; |
---|
| 26 | |
---|
| 27 | void atomic_max(std::atomic_size_t & target, size_t value) { |
---|
| 28 | for(;;) { |
---|
| 29 | size_t expect = target.load(std::memory_order_relaxed); |
---|
| 30 | if(value <= expect) return; |
---|
| 31 | bool success = target.compare_exchange_strong(expect, value); |
---|
| 32 | if(success) return; |
---|
| 33 | } |
---|
| 34 | } |
---|
| 35 | |
---|
| 36 | void atomic_min(std::atomic_size_t & target, size_t value) { |
---|
| 37 | for(;;) { |
---|
| 38 | size_t expect = target.load(std::memory_order_relaxed); |
---|
| 39 | if(value >= expect) return; |
---|
| 40 | bool success = target.compare_exchange_strong(expect, value); |
---|
| 41 | if(success) return; |
---|
| 42 | } |
---|
| 43 | } |
---|
| 44 | |
---|
| 45 | void tally_stats(global_stat_t & global, local_stat_t & local) { |
---|
| 46 | global.cnt += local.cnt; |
---|
| 47 | } |
---|
| 48 | |
---|
| 49 | void waitfor(double & duration, barrier_t & barrier, std::atomic_bool & done) { |
---|
| 50 | std::cout << "Starting" << std::endl; |
---|
| 51 | auto before = Clock::now(); |
---|
| 52 | barrier.wait(0); |
---|
| 53 | |
---|
| 54 | while(true) { |
---|
| 55 | usleep(100000); |
---|
| 56 | auto now = Clock::now(); |
---|
| 57 | duration_t durr = now - before; |
---|
| 58 | if( durr.count() > duration ) { |
---|
| 59 | done = true; |
---|
| 60 | break; |
---|
| 61 | } |
---|
| 62 | std::cout << "\r" << std::setprecision(4) << durr.count(); |
---|
| 63 | std::cout.flush(); |
---|
| 64 | } |
---|
| 65 | |
---|
| 66 | barrier.wait(0); |
---|
| 67 | auto after = Clock::now(); |
---|
| 68 | duration_t durr = after - before; |
---|
| 69 | duration = durr.count(); |
---|
| 70 | std::cout << "\rClosing down" << std::endl; |
---|
| 71 | } |
---|
| 72 | |
---|
| 73 | void waitfor(double & duration, barrier_t & barrier, const std::atomic_size_t & count) { |
---|
| 74 | std::cout << "Starting" << std::endl; |
---|
| 75 | auto before = Clock::now(); |
---|
| 76 | barrier.wait(0); |
---|
| 77 | |
---|
| 78 | while(true) { |
---|
| 79 | usleep(100000); |
---|
| 80 | size_t c = count.load(); |
---|
| 81 | if( c == 0 ) { |
---|
| 82 | break; |
---|
| 83 | } |
---|
| 84 | std::cout << "\r" << c; |
---|
| 85 | std::cout.flush(); |
---|
| 86 | } |
---|
| 87 | |
---|
| 88 | barrier.wait(0); |
---|
| 89 | auto after = Clock::now(); |
---|
| 90 | duration_t durr = after - before; |
---|
| 91 | duration = durr.count(); |
---|
| 92 | std::cout << "\rClosing down" << std::endl; |
---|
| 93 | } |
---|
| 94 | |
---|
| 95 | void print_stats(double duration, unsigned nthread, global_stat_t & global) { |
---|
| 96 | std::cout << "Done" << std::endl; |
---|
| 97 | |
---|
| 98 | size_t ops = global.cnt; |
---|
| 99 | size_t ops_sec = size_t(double(ops) / duration); |
---|
| 100 | size_t ops_thread = ops_sec / nthread; |
---|
| 101 | auto dur_nano = duration_cast<std::nano>(1.0); |
---|
| 102 | |
---|
| 103 | std::cout << "Duration : " << duration << "s\n"; |
---|
| 104 | std::cout << "ns/Op : " << ( dur_nano / ops_thread )<< "\n"; |
---|
| 105 | std::cout << "Ops/sec/thread: " << ops_thread << "\n"; |
---|
| 106 | std::cout << "Ops/sec : " << ops_sec << "\n"; |
---|
| 107 | std::cout << "Total ops : " << ops << "\n"; |
---|
| 108 | } |
---|
| 109 | |
---|
| 110 | static inline bool bts(std::atomic_size_t & target, size_t bit ) { |
---|
| 111 | /* |
---|
| 112 | int result = 0; |
---|
| 113 | asm volatile( |
---|
| 114 | "LOCK btsq %[bit], %[target]\n\t" |
---|
| 115 | :"=@ccc" (result) |
---|
| 116 | : [target] "m" (target), [bit] "r" (bit) |
---|
| 117 | ); |
---|
| 118 | return result != 0; |
---|
| 119 | /*/ |
---|
| 120 | size_t mask = 1ul << bit; |
---|
| 121 | size_t ret = target.fetch_or(mask, std::memory_order_relaxed); |
---|
| 122 | return (ret & mask) != 0; |
---|
| 123 | //*/ |
---|
| 124 | } |
---|
| 125 | |
---|
| 126 | static inline bool btr(std::atomic_size_t & target, size_t bit ) { |
---|
| 127 | /* |
---|
| 128 | int result = 0; |
---|
| 129 | asm volatile( |
---|
| 130 | "LOCK btrq %[bit], %[target]\n\t" |
---|
| 131 | :"=@ccc" (result) |
---|
| 132 | : [target] "m" (target), [bit] "r" (bit) |
---|
| 133 | ); |
---|
| 134 | return result != 0; |
---|
| 135 | /*/ |
---|
| 136 | size_t mask = 1ul << bit; |
---|
| 137 | size_t ret = target.fetch_and(~mask, std::memory_order_relaxed); |
---|
| 138 | return (ret & mask) != 0; |
---|
| 139 | //*/ |
---|
| 140 | } |
---|
| 141 | |
---|
| 142 | // ================================================================================================ |
---|
| 143 | // EXPERIMENTS |
---|
| 144 | // ================================================================================================ |
---|
| 145 | |
---|
| 146 | // ================================================================================================ |
---|
| 147 | __attribute__((noinline)) void runPingPong_body( |
---|
| 148 | std::atomic<bool>& done, |
---|
| 149 | local_stat_t & local, |
---|
| 150 | std::atomic_size_t & target, |
---|
| 151 | size_t id |
---|
| 152 | ) { |
---|
| 153 | while(__builtin_expect(!done.load(std::memory_order_relaxed), true)) { |
---|
| 154 | |
---|
| 155 | bool ret; |
---|
| 156 | ret = bts(target, id); |
---|
| 157 | assert(!ret); |
---|
| 158 | |
---|
| 159 | // ----- |
---|
| 160 | |
---|
| 161 | ret = btr(target, id); |
---|
| 162 | assert(ret); |
---|
| 163 | local.cnt++; |
---|
| 164 | } |
---|
| 165 | } |
---|
| 166 | |
---|
| 167 | void run(unsigned nthread, double duration) { |
---|
| 168 | // Barrier for synchronization |
---|
| 169 | barrier_t barrier(nthread + 1); |
---|
| 170 | |
---|
| 171 | // Data to check everything is OK |
---|
| 172 | global_stat_t global; |
---|
| 173 | |
---|
| 174 | // Flag to signal termination |
---|
| 175 | std::atomic_bool done = { false }; |
---|
| 176 | |
---|
| 177 | std::cout << "Initializing "; |
---|
| 178 | // List being tested |
---|
| 179 | std::atomic_size_t word = { 0 }; |
---|
| 180 | { |
---|
| 181 | std::thread * threads[nthread]; |
---|
| 182 | unsigned i = 1; |
---|
| 183 | for(auto & t : threads) { |
---|
| 184 | t = new std::thread([&done, &word, &barrier, &global](unsigned tid) { |
---|
| 185 | local_stat_t local; |
---|
| 186 | |
---|
| 187 | // affinity(tid); |
---|
| 188 | |
---|
| 189 | barrier.wait(tid); |
---|
| 190 | |
---|
| 191 | // EXPERIMENT START |
---|
| 192 | |
---|
| 193 | runPingPong_body(done, local, word, tid - 1); |
---|
| 194 | |
---|
| 195 | // EXPERIMENT END |
---|
| 196 | |
---|
| 197 | barrier.wait(tid); |
---|
| 198 | |
---|
| 199 | tally_stats(global, local); |
---|
| 200 | }, i++); |
---|
| 201 | } |
---|
| 202 | |
---|
| 203 | waitfor(duration, barrier, done); |
---|
| 204 | |
---|
| 205 | for(auto t : threads) { |
---|
| 206 | t->join(); |
---|
| 207 | delete t; |
---|
| 208 | } |
---|
| 209 | } |
---|
| 210 | |
---|
| 211 | print_stats(duration, nthread, global); |
---|
| 212 | } |
---|
| 213 | |
---|
| 214 | // ================================================================================================ |
---|
| 215 | |
---|
| 216 | int main(int argc, char * argv[]) { |
---|
| 217 | |
---|
| 218 | double duration = 5.0; |
---|
| 219 | unsigned nthreads = 2; |
---|
| 220 | |
---|
| 221 | std::cout.imbue(std::locale("")); |
---|
| 222 | |
---|
| 223 | for(;;) { |
---|
| 224 | static struct option options[] = { |
---|
| 225 | {"duration", required_argument, 0, 'd'}, |
---|
| 226 | {"nthreads", required_argument, 0, 't'}, |
---|
| 227 | {0, 0, 0, 0} |
---|
| 228 | }; |
---|
| 229 | |
---|
| 230 | int idx = 0; |
---|
| 231 | int opt = getopt_long(argc, argv, "d:t:", options, &idx); |
---|
| 232 | |
---|
| 233 | std::string arg = optarg ? optarg : ""; |
---|
| 234 | size_t len = 0; |
---|
| 235 | switch(opt) { |
---|
| 236 | case -1: |
---|
| 237 | if(optind != argc) { |
---|
| 238 | std::cerr << "Too many arguments " << argc << " " << idx << std::endl; |
---|
| 239 | goto usage; |
---|
| 240 | } |
---|
| 241 | goto run; |
---|
| 242 | // Numeric Arguments |
---|
| 243 | case 'd': |
---|
| 244 | try { |
---|
| 245 | duration = std::stod(optarg, &len); |
---|
| 246 | if(len != arg.size()) { throw std::invalid_argument(""); } |
---|
| 247 | } catch(std::invalid_argument &) { |
---|
| 248 | std::cerr << "Duration must be a valid double, was " << arg << std::endl; |
---|
| 249 | goto usage; |
---|
| 250 | } |
---|
| 251 | break; |
---|
| 252 | case 't': |
---|
| 253 | try { |
---|
| 254 | nthreads = std::stoul(optarg, &len); |
---|
| 255 | if(len != arg.size() || nthreads > (8 * sizeof(size_t))) { throw std::invalid_argument(""); } |
---|
| 256 | } catch(std::invalid_argument &) { |
---|
| 257 | std::cerr << "Number of threads must be a positive integer less than or equal to " << sizeof(size_t) * 8 << ", was " << arg << std::endl; |
---|
| 258 | goto usage; |
---|
| 259 | } |
---|
| 260 | break; |
---|
| 261 | // Other cases |
---|
| 262 | default: /* ? */ |
---|
| 263 | std::cerr << opt << std::endl; |
---|
| 264 | usage: |
---|
| 265 | std::cerr << "Usage: " << argv[0] << ": [options]" << std::endl; |
---|
| 266 | std::cerr << std::endl; |
---|
| 267 | std::cerr << " -d, --duration=DURATION Duration of the experiment, in seconds" << std::endl; |
---|
| 268 | std::cerr << " -t, --nthreads=NTHREADS Number of kernel threads" << std::endl; |
---|
| 269 | std::exit(1); |
---|
| 270 | } |
---|
| 271 | } |
---|
| 272 | run: |
---|
| 273 | |
---|
| 274 | check_cache_line_size(); |
---|
| 275 | |
---|
| 276 | std::cout << "Running " << nthreads << " threads for " << duration << " seconds" << std::endl; |
---|
| 277 | run(nthreads, duration); |
---|
| 278 | return 0; |
---|
| 279 | } |
---|