[b232745] | 1 | #if !defined(LIST_VARIANT_HPP) |
---|
| 2 | #define LIST_VARIANT_HPP "relaxed_list.hpp" |
---|
| 3 | #endif |
---|
| 4 | |
---|
| 5 | #include LIST_VARIANT_HPP |
---|
| 6 | #if !defined(LIST_VARIANT) |
---|
| 7 | #error not variant selected |
---|
| 8 | #endif |
---|
[b2a37b0] | 9 | |
---|
| 10 | #include <array> |
---|
[c921712] | 11 | #include <iomanip> |
---|
[b2a37b0] | 12 | #include <iostream> |
---|
| 13 | #include <locale> |
---|
| 14 | #include <string> |
---|
| 15 | #include <thread> |
---|
| 16 | #include <vector> |
---|
| 17 | |
---|
[9421f3d8] | 18 | #include <getopt.h> |
---|
[b2a37b0] | 19 | #include <unistd.h> |
---|
| 20 | #include <sys/sysinfo.h> |
---|
| 21 | |
---|
| 22 | #include "utils.hpp" |
---|
| 23 | |
---|
[50aeb6f] | 24 | struct __attribute__((aligned(64))) Node { |
---|
[b2a37b0] | 25 | static std::atomic_size_t creates; |
---|
| 26 | static std::atomic_size_t destroys; |
---|
| 27 | |
---|
| 28 | _LinksFields_t<Node> _links; |
---|
| 29 | |
---|
| 30 | int value; |
---|
[807a632] | 31 | int id; |
---|
[b2a37b0] | 32 | |
---|
[9421f3d8] | 33 | Node() { creates++; } |
---|
| 34 | Node(int value): value(value) { creates++; } |
---|
| 35 | ~Node() { destroys++; } |
---|
[b2a37b0] | 36 | }; |
---|
| 37 | |
---|
| 38 | std::atomic_size_t Node::creates = { 0 }; |
---|
| 39 | std::atomic_size_t Node::destroys = { 0 }; |
---|
| 40 | |
---|
| 41 | bool enable_stats = false; |
---|
| 42 | |
---|
[9421f3d8] | 43 | template<> |
---|
[b232745] | 44 | thread_local LIST_VARIANT<Node>::TLS LIST_VARIANT<Node>::tls = {}; |
---|
[9421f3d8] | 45 | |
---|
| 46 | template<> |
---|
[b232745] | 47 | std::atomic_uint32_t LIST_VARIANT<Node>::ticket = { 0 }; |
---|
[807a632] | 48 | |
---|
| 49 | #ifndef NO_STATS |
---|
| 50 | template<> |
---|
[b232745] | 51 | LIST_VARIANT<Node>::GlobalStats LIST_VARIANT<Node>::global_stats = {}; |
---|
[807a632] | 52 | #endif |
---|
[9421f3d8] | 53 | |
---|
| 54 | // ================================================================================================ |
---|
| 55 | // UTILS |
---|
| 56 | // ================================================================================================ |
---|
| 57 | |
---|
[50aeb6f] | 58 | struct local_stat_t { |
---|
| 59 | size_t in = 0; |
---|
| 60 | size_t out = 0; |
---|
| 61 | size_t empty = 0; |
---|
| 62 | size_t crc_in = 0; |
---|
| 63 | size_t crc_out = 0; |
---|
[807a632] | 64 | size_t valmax = 0; |
---|
| 65 | size_t valmin = 100000000ul; |
---|
[95cb63b] | 66 | struct { |
---|
| 67 | size_t val = 0; |
---|
| 68 | size_t cnt = 0; |
---|
| 69 | } comp; |
---|
| 70 | struct { |
---|
| 71 | size_t val = 0; |
---|
| 72 | size_t cnt = 0; |
---|
| 73 | } subm; |
---|
[50aeb6f] | 74 | }; |
---|
| 75 | |
---|
[9421f3d8] | 76 | struct global_stat_t { |
---|
| 77 | std::atomic_size_t in = { 0 }; |
---|
| 78 | std::atomic_size_t out = { 0 }; |
---|
| 79 | std::atomic_size_t empty = { 0 }; |
---|
| 80 | std::atomic_size_t crc_in = { 0 }; |
---|
| 81 | std::atomic_size_t crc_out = { 0 }; |
---|
[807a632] | 82 | std::atomic_size_t valmax = { 0 }; |
---|
| 83 | std::atomic_size_t valmin = { 100000000ul }; |
---|
[95cb63b] | 84 | struct { |
---|
| 85 | std::atomic_size_t val = { 0 }; |
---|
| 86 | std::atomic_size_t cnt = { 0 }; |
---|
| 87 | } comp; |
---|
| 88 | struct { |
---|
| 89 | std::atomic_size_t val = { 0 }; |
---|
| 90 | std::atomic_size_t cnt = { 0 }; |
---|
| 91 | } subm; |
---|
[9421f3d8] | 92 | }; |
---|
| 93 | |
---|
[807a632] | 94 | void atomic_max(std::atomic_size_t & target, size_t value) { |
---|
| 95 | for(;;) { |
---|
| 96 | size_t expect = target.load(std::memory_order_relaxed); |
---|
| 97 | if(value <= expect) return; |
---|
| 98 | bool success = target.compare_exchange_strong(expect, value); |
---|
| 99 | if(success) return; |
---|
| 100 | } |
---|
| 101 | } |
---|
| 102 | |
---|
| 103 | void atomic_min(std::atomic_size_t & target, size_t value) { |
---|
| 104 | for(;;) { |
---|
| 105 | size_t expect = target.load(std::memory_order_relaxed); |
---|
| 106 | if(value >= expect) return; |
---|
| 107 | bool success = target.compare_exchange_strong(expect, value); |
---|
| 108 | if(success) return; |
---|
| 109 | } |
---|
| 110 | } |
---|
| 111 | |
---|
[9421f3d8] | 112 | void tally_stats(global_stat_t & global, local_stat_t & local) { |
---|
[807a632] | 113 | |
---|
[9421f3d8] | 114 | global.in += local.in; |
---|
| 115 | global.out += local.out; |
---|
| 116 | global.empty += local.empty; |
---|
| 117 | |
---|
| 118 | global.crc_in += local.crc_in; |
---|
| 119 | global.crc_out += local.crc_out; |
---|
| 120 | |
---|
[95cb63b] | 121 | global.comp.val += local.comp.val; |
---|
| 122 | global.comp.cnt += local.comp.cnt; |
---|
| 123 | global.subm.val += local.subm.val; |
---|
| 124 | global.subm.cnt += local.subm.cnt; |
---|
| 125 | |
---|
[807a632] | 126 | atomic_max(global.valmax, local.valmax); |
---|
| 127 | atomic_min(global.valmin, local.valmin); |
---|
[9421f3d8] | 128 | |
---|
[b232745] | 129 | LIST_VARIANT<Node>::stats_tls_tally(); |
---|
[9421f3d8] | 130 | } |
---|
| 131 | |
---|
| 132 | void waitfor(double & duration, barrier_t & barrier, std::atomic_bool & done) { |
---|
| 133 | std::cout << "Starting" << std::endl; |
---|
| 134 | auto before = Clock::now(); |
---|
| 135 | barrier.wait(0); |
---|
[16ee228] | 136 | bool is_tty = isatty(STDOUT_FILENO); |
---|
[9421f3d8] | 137 | |
---|
| 138 | while(true) { |
---|
| 139 | usleep(100000); |
---|
| 140 | auto now = Clock::now(); |
---|
| 141 | duration_t durr = now - before; |
---|
| 142 | if( durr.count() > duration ) { |
---|
| 143 | done = true; |
---|
| 144 | break; |
---|
| 145 | } |
---|
[16ee228] | 146 | if(is_tty) { |
---|
| 147 | std::cout << "\r" << std::setprecision(4) << durr.count(); |
---|
| 148 | std::cout.flush(); |
---|
| 149 | } |
---|
[9421f3d8] | 150 | } |
---|
| 151 | |
---|
| 152 | barrier.wait(0); |
---|
| 153 | auto after = Clock::now(); |
---|
| 154 | duration_t durr = after - before; |
---|
| 155 | duration = durr.count(); |
---|
| 156 | std::cout << "\rClosing down" << std::endl; |
---|
| 157 | } |
---|
| 158 | |
---|
[807a632] | 159 | void waitfor(double & duration, barrier_t & barrier, const std::atomic_size_t & count) { |
---|
| 160 | std::cout << "Starting" << std::endl; |
---|
| 161 | auto before = Clock::now(); |
---|
| 162 | barrier.wait(0); |
---|
| 163 | |
---|
| 164 | while(true) { |
---|
| 165 | usleep(100000); |
---|
| 166 | size_t c = count.load(); |
---|
| 167 | if( c == 0 ) { |
---|
| 168 | break; |
---|
| 169 | } |
---|
| 170 | std::cout << "\r" << c; |
---|
| 171 | std::cout.flush(); |
---|
| 172 | } |
---|
| 173 | |
---|
| 174 | barrier.wait(0); |
---|
| 175 | auto after = Clock::now(); |
---|
| 176 | duration_t durr = after - before; |
---|
| 177 | duration = durr.count(); |
---|
| 178 | std::cout << "\rClosing down" << std::endl; |
---|
| 179 | } |
---|
| 180 | |
---|
[9421f3d8] | 181 | void print_stats(double duration, unsigned nthread, global_stat_t & global) { |
---|
| 182 | assert(Node::creates == Node::destroys); |
---|
| 183 | assert(global.crc_in == global.crc_out); |
---|
| 184 | |
---|
| 185 | std::cout << "Done" << std::endl; |
---|
| 186 | |
---|
| 187 | size_t ops = global.in + global.out; |
---|
| 188 | size_t ops_sec = size_t(double(ops) / duration); |
---|
| 189 | size_t ops_thread = ops_sec / nthread; |
---|
| 190 | auto dur_nano = duration_cast<std::nano>(1.0); |
---|
| 191 | |
---|
[95cb63b] | 192 | if(global.valmax != 0) { |
---|
| 193 | std::cout << "Max runs : " << global.valmax << "\n"; |
---|
| 194 | std::cout << "Min runs : " << global.valmin << "\n"; |
---|
| 195 | } |
---|
| 196 | if(global.comp.cnt != 0) { |
---|
| 197 | std::cout << "Submit count : " << global.subm.cnt << "\n"; |
---|
| 198 | std::cout << "Submit average: " << ((double(global.subm.val)) / global.subm.cnt) << "\n"; |
---|
| 199 | std::cout << "Complete count: " << global.comp.cnt << "\n"; |
---|
| 200 | std::cout << "Complete avg : " << ((double(global.comp.val)) / global.comp.cnt) << "\n"; |
---|
| 201 | } |
---|
[9421f3d8] | 202 | std::cout << "Duration : " << duration << "s\n"; |
---|
| 203 | std::cout << "ns/Op : " << ( dur_nano / ops_thread )<< "\n"; |
---|
| 204 | std::cout << "Ops/sec/thread: " << ops_thread << "\n"; |
---|
| 205 | std::cout << "Ops/sec : " << ops_sec << "\n"; |
---|
| 206 | std::cout << "Total ops : " << ops << "(" << global.in << "i, " << global.out << "o, " << global.empty << "e)\n"; |
---|
| 207 | #ifndef NO_STATS |
---|
[a1b9bc3] | 208 | LIST_VARIANT<Node>::stats_print(std::cout, duration); |
---|
[9421f3d8] | 209 | #endif |
---|
| 210 | } |
---|
| 211 | |
---|
[807a632] | 212 | void save_fairness(const int data[], int factor, unsigned nthreads, size_t columns, size_t rows, const std::string & output); |
---|
| 213 | |
---|
[9421f3d8] | 214 | // ================================================================================================ |
---|
| 215 | // EXPERIMENTS |
---|
| 216 | // ================================================================================================ |
---|
| 217 | |
---|
| 218 | // ================================================================================================ |
---|
| 219 | __attribute__((noinline)) void runChurn_body( |
---|
[50aeb6f] | 220 | std::atomic<bool>& done, |
---|
| 221 | Random & rand, |
---|
[9421f3d8] | 222 | Node * my_nodes[], |
---|
| 223 | unsigned nslots, |
---|
[50aeb6f] | 224 | local_stat_t & local, |
---|
[b232745] | 225 | LIST_VARIANT<Node> & list |
---|
[50aeb6f] | 226 | ) { |
---|
| 227 | while(__builtin_expect(!done.load(std::memory_order_relaxed), true)) { |
---|
[9421f3d8] | 228 | int idx = rand.next() % nslots; |
---|
[50aeb6f] | 229 | if (auto node = my_nodes[idx]) { |
---|
| 230 | local.crc_in += node->value; |
---|
| 231 | list.push(node); |
---|
| 232 | my_nodes[idx] = nullptr; |
---|
| 233 | local.in++; |
---|
| 234 | } |
---|
| 235 | else if(auto node = list.pop()) { |
---|
| 236 | local.crc_out += node->value; |
---|
| 237 | my_nodes[idx] = node; |
---|
| 238 | local.out++; |
---|
| 239 | } |
---|
| 240 | else { |
---|
| 241 | local.empty++; |
---|
| 242 | } |
---|
| 243 | } |
---|
| 244 | } |
---|
[b2a37b0] | 245 | |
---|
[9421f3d8] | 246 | void runChurn(unsigned nthread, unsigned nqueues, double duration, unsigned nnodes, const unsigned nslots) { |
---|
| 247 | std::cout << "Churn Benchmark" << std::endl; |
---|
| 248 | assert(nnodes <= nslots); |
---|
[b2a37b0] | 249 | // List being tested |
---|
| 250 | |
---|
| 251 | // Barrier for synchronization |
---|
| 252 | barrier_t barrier(nthread + 1); |
---|
| 253 | |
---|
| 254 | // Data to check everything is OK |
---|
[9421f3d8] | 255 | global_stat_t global; |
---|
[b2a37b0] | 256 | |
---|
| 257 | // Flag to signal termination |
---|
| 258 | std::atomic_bool done = { false }; |
---|
| 259 | |
---|
| 260 | // Prep nodes |
---|
[c921712] | 261 | std::cout << "Initializing "; |
---|
| 262 | size_t npushed = 0; |
---|
[b232745] | 263 | LIST_VARIANT<Node> list = { nthread, nqueues }; |
---|
[807a632] | 264 | { |
---|
| 265 | Node** all_nodes[nthread]; |
---|
| 266 | for(auto & nodes : all_nodes) { |
---|
| 267 | nodes = new __attribute__((aligned(64))) Node*[nslots + 8]; |
---|
| 268 | Random rand(rdtscl()); |
---|
| 269 | for(unsigned i = 0; i < nnodes; i++) { |
---|
| 270 | nodes[i] = new Node(rand.next() % 100); |
---|
| 271 | } |
---|
[9421f3d8] | 272 | |
---|
[807a632] | 273 | for(unsigned i = nnodes; i < nslots; i++) { |
---|
| 274 | nodes[i] = nullptr; |
---|
| 275 | } |
---|
[b2a37b0] | 276 | |
---|
[807a632] | 277 | for(int i = 0; i < 10 && i < (int)nslots; i++) { |
---|
| 278 | int idx = rand.next() % nslots; |
---|
| 279 | if (auto node = nodes[idx]) { |
---|
| 280 | global.crc_in += node->value; |
---|
| 281 | list.push(node); |
---|
| 282 | npushed++; |
---|
| 283 | nodes[idx] = nullptr; |
---|
| 284 | } |
---|
[b2a37b0] | 285 | } |
---|
| 286 | } |
---|
| 287 | |
---|
[807a632] | 288 | std::cout << nnodes << " nodes (" << nslots << " slots)" << std::endl; |
---|
[c921712] | 289 | |
---|
[807a632] | 290 | enable_stats = true; |
---|
[b2a37b0] | 291 | |
---|
[807a632] | 292 | std::thread * threads[nthread]; |
---|
| 293 | unsigned i = 1; |
---|
| 294 | for(auto & t : threads) { |
---|
| 295 | auto & my_nodes = all_nodes[i - 1]; |
---|
| 296 | t = new std::thread([&done, &list, &barrier, &global, &my_nodes, nslots](unsigned tid) { |
---|
| 297 | Random rand(tid + rdtscl()); |
---|
[b2a37b0] | 298 | |
---|
[807a632] | 299 | local_stat_t local; |
---|
[b2a37b0] | 300 | |
---|
[807a632] | 301 | // affinity(tid); |
---|
[b2a37b0] | 302 | |
---|
[807a632] | 303 | barrier.wait(tid); |
---|
[b2a37b0] | 304 | |
---|
[807a632] | 305 | // EXPERIMENT START |
---|
[b2a37b0] | 306 | |
---|
[807a632] | 307 | runChurn_body(done, rand, my_nodes, nslots, local, list); |
---|
[b2a37b0] | 308 | |
---|
[807a632] | 309 | // EXPERIMENT END |
---|
[b2a37b0] | 310 | |
---|
[807a632] | 311 | barrier.wait(tid); |
---|
[b2a37b0] | 312 | |
---|
[807a632] | 313 | tally_stats(global, local); |
---|
[b2a37b0] | 314 | |
---|
[807a632] | 315 | for(unsigned i = 0; i < nslots; i++) { |
---|
| 316 | delete my_nodes[i]; |
---|
| 317 | } |
---|
| 318 | }, i++); |
---|
| 319 | } |
---|
[b2a37b0] | 320 | |
---|
[807a632] | 321 | waitfor(duration, barrier, done); |
---|
[b2a37b0] | 322 | |
---|
[807a632] | 323 | for(auto t : threads) { |
---|
| 324 | t->join(); |
---|
| 325 | delete t; |
---|
| 326 | } |
---|
[b2a37b0] | 327 | |
---|
[807a632] | 328 | enable_stats = false; |
---|
[b2a37b0] | 329 | |
---|
[807a632] | 330 | while(auto node = list.pop()) { |
---|
| 331 | global.crc_out += node->value; |
---|
| 332 | delete node; |
---|
| 333 | } |
---|
[b2a37b0] | 334 | |
---|
[807a632] | 335 | for(auto nodes : all_nodes) { |
---|
| 336 | delete[] nodes; |
---|
| 337 | } |
---|
[9421f3d8] | 338 | } |
---|
[b2a37b0] | 339 | |
---|
[9421f3d8] | 340 | print_stats(duration, nthread, global); |
---|
| 341 | } |
---|
[b2a37b0] | 342 | |
---|
[9421f3d8] | 343 | // ================================================================================================ |
---|
| 344 | __attribute__((noinline)) void runPingPong_body( |
---|
| 345 | std::atomic<bool>& done, |
---|
| 346 | Node initial_nodes[], |
---|
| 347 | unsigned nnodes, |
---|
| 348 | local_stat_t & local, |
---|
[b232745] | 349 | LIST_VARIANT<Node> & list |
---|
[9421f3d8] | 350 | ) { |
---|
| 351 | Node * nodes[nnodes]; |
---|
| 352 | { |
---|
| 353 | unsigned i = 0; |
---|
| 354 | for(auto & n : nodes) { |
---|
| 355 | n = &initial_nodes[i++]; |
---|
| 356 | } |
---|
| 357 | } |
---|
[b2a37b0] | 358 | |
---|
[9421f3d8] | 359 | while(__builtin_expect(!done.load(std::memory_order_relaxed), true)) { |
---|
| 360 | |
---|
| 361 | for(Node * & node : nodes) { |
---|
| 362 | local.crc_in += node->value; |
---|
| 363 | list.push(node); |
---|
| 364 | local.in++; |
---|
| 365 | } |
---|
| 366 | |
---|
| 367 | // ----- |
---|
| 368 | |
---|
| 369 | for(Node * & node : nodes) { |
---|
[a1b9bc3] | 370 | node = nullptr; |
---|
| 371 | while(!node) { |
---|
| 372 | node = list.pop(); |
---|
| 373 | } |
---|
[9421f3d8] | 374 | local.crc_out += node->value; |
---|
| 375 | local.out++; |
---|
| 376 | } |
---|
| 377 | } |
---|
| 378 | } |
---|
| 379 | |
---|
| 380 | void runPingPong(unsigned nthread, unsigned nqueues, double duration, unsigned nnodes) { |
---|
| 381 | std::cout << "PingPong Benchmark" << std::endl; |
---|
| 382 | |
---|
| 383 | |
---|
| 384 | // Barrier for synchronization |
---|
| 385 | barrier_t barrier(nthread + 1); |
---|
| 386 | |
---|
| 387 | // Data to check everything is OK |
---|
| 388 | global_stat_t global; |
---|
| 389 | |
---|
| 390 | // Flag to signal termination |
---|
| 391 | std::atomic_bool done = { false }; |
---|
| 392 | |
---|
| 393 | std::cout << "Initializing "; |
---|
[807a632] | 394 | // List being tested |
---|
[b232745] | 395 | LIST_VARIANT<Node> list = { nthread, nqueues }; |
---|
[807a632] | 396 | { |
---|
| 397 | enable_stats = true; |
---|
[9421f3d8] | 398 | |
---|
[807a632] | 399 | std::thread * threads[nthread]; |
---|
| 400 | unsigned i = 1; |
---|
| 401 | for(auto & t : threads) { |
---|
| 402 | t = new std::thread([&done, &list, &barrier, &global, nnodes](unsigned tid) { |
---|
| 403 | Random rand(tid + rdtscl()); |
---|
[9421f3d8] | 404 | |
---|
[807a632] | 405 | Node nodes[nnodes]; |
---|
| 406 | for(auto & n : nodes) { |
---|
| 407 | n.value = (int)rand.next() % 100; |
---|
| 408 | } |
---|
[9421f3d8] | 409 | |
---|
[807a632] | 410 | local_stat_t local; |
---|
[9421f3d8] | 411 | |
---|
[807a632] | 412 | // affinity(tid); |
---|
[9421f3d8] | 413 | |
---|
[807a632] | 414 | barrier.wait(tid); |
---|
[9421f3d8] | 415 | |
---|
[807a632] | 416 | // EXPERIMENT START |
---|
[9421f3d8] | 417 | |
---|
[807a632] | 418 | runPingPong_body(done, nodes, nnodes, local, list); |
---|
[9421f3d8] | 419 | |
---|
[807a632] | 420 | // EXPERIMENT END |
---|
[9421f3d8] | 421 | |
---|
[807a632] | 422 | barrier.wait(tid); |
---|
[9421f3d8] | 423 | |
---|
[807a632] | 424 | tally_stats(global, local); |
---|
| 425 | }, i++); |
---|
| 426 | } |
---|
| 427 | |
---|
| 428 | waitfor(duration, barrier, done); |
---|
| 429 | |
---|
| 430 | for(auto t : threads) { |
---|
| 431 | t->join(); |
---|
| 432 | delete t; |
---|
| 433 | } |
---|
| 434 | |
---|
| 435 | enable_stats = false; |
---|
[9421f3d8] | 436 | } |
---|
| 437 | |
---|
[807a632] | 438 | print_stats(duration, nthread, global); |
---|
| 439 | } |
---|
[9421f3d8] | 440 | |
---|
[95cb63b] | 441 | // ================================================================================================ |
---|
| 442 | struct __attribute__((aligned(64))) Slot { |
---|
| 443 | Node * volatile node; |
---|
| 444 | }; |
---|
| 445 | |
---|
| 446 | __attribute__((noinline)) void runProducer_body( |
---|
| 447 | std::atomic<bool>& done, |
---|
| 448 | Random & rand, |
---|
| 449 | Slot * slots, |
---|
| 450 | int nslots, |
---|
| 451 | local_stat_t & local, |
---|
[b232745] | 452 | LIST_VARIANT<Node> & list |
---|
[95cb63b] | 453 | ) { |
---|
| 454 | while(__builtin_expect(!done.load(std::memory_order_relaxed), true)) { |
---|
| 455 | |
---|
| 456 | Node * node = list.pop(); |
---|
| 457 | if(!node) { |
---|
| 458 | local.empty ++; |
---|
| 459 | continue; |
---|
| 460 | } |
---|
| 461 | |
---|
| 462 | local.crc_out += node->value; |
---|
| 463 | local.out++; |
---|
| 464 | |
---|
| 465 | if(node->id == 0) { |
---|
| 466 | unsigned cnt = 0; |
---|
| 467 | for(int i = 0; i < nslots; i++) { |
---|
| 468 | Node * found = __atomic_exchange_n( &slots[i].node, nullptr, __ATOMIC_SEQ_CST ); |
---|
| 469 | if( found ) { |
---|
| 470 | local.crc_in += found->value; |
---|
| 471 | local.in++; |
---|
| 472 | cnt++; |
---|
| 473 | list.push( found ); |
---|
| 474 | } |
---|
| 475 | } |
---|
| 476 | |
---|
| 477 | local.crc_in += node->value; |
---|
| 478 | local.in++; |
---|
| 479 | list.push( node ); |
---|
| 480 | |
---|
| 481 | local.comp.cnt++; |
---|
| 482 | local.comp.val += cnt; |
---|
| 483 | } |
---|
| 484 | else { |
---|
| 485 | unsigned len = 0; |
---|
| 486 | while(true) { |
---|
| 487 | auto off = rand.next(); |
---|
| 488 | for(int i = 0; i < nslots; i++) { |
---|
| 489 | Node * expected = nullptr; |
---|
| 490 | int idx = (i + off) % nslots; |
---|
| 491 | Slot & slot = slots[ idx ]; |
---|
| 492 | if( |
---|
| 493 | slot.node == nullptr && |
---|
| 494 | __atomic_compare_exchange_n( &slot.node, &expected, node, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST ) |
---|
| 495 | ) { |
---|
| 496 | local.subm.cnt++; |
---|
| 497 | local.subm.val += len; |
---|
| 498 | goto LOOP; |
---|
| 499 | } |
---|
| 500 | assert( expected != node ); |
---|
| 501 | len++; |
---|
| 502 | } |
---|
| 503 | } |
---|
| 504 | } |
---|
| 505 | |
---|
| 506 | LOOP:; |
---|
| 507 | } |
---|
| 508 | } |
---|
| 509 | |
---|
| 510 | void runProducer(unsigned nthread, unsigned nqueues, double duration, unsigned nnodes) { |
---|
| 511 | std::cout << "Producer Benchmark" << std::endl; |
---|
| 512 | |
---|
| 513 | // Barrier for synchronization |
---|
| 514 | barrier_t barrier(nthread + 1); |
---|
| 515 | |
---|
| 516 | // Data to check everything is OK |
---|
| 517 | global_stat_t global; |
---|
| 518 | |
---|
| 519 | // Flag to signal termination |
---|
| 520 | std::atomic_bool done = { false }; |
---|
| 521 | |
---|
| 522 | std::cout << "Initializing "; |
---|
| 523 | |
---|
| 524 | int nslots = nnodes * 4; |
---|
| 525 | Slot * slots = new Slot[nslots]; |
---|
| 526 | std::cout << nnodes << " nodes (" << nslots << " slots)" << std::endl; |
---|
| 527 | |
---|
| 528 | // List being tested |
---|
[b232745] | 529 | LIST_VARIANT<Node> list = { nthread, nqueues }; |
---|
[95cb63b] | 530 | { |
---|
| 531 | Random rand(rdtscl()); |
---|
| 532 | for(unsigned i = 0; i < nnodes; i++) { |
---|
| 533 | Node * node = new Node(rand.next() % 100); |
---|
| 534 | node->id = i; |
---|
| 535 | global.crc_in += node->value; |
---|
| 536 | list.push(node); |
---|
| 537 | } |
---|
| 538 | |
---|
| 539 | for(int i = 0; i < nslots; i++) { |
---|
| 540 | slots[i].node = nullptr; |
---|
| 541 | } |
---|
| 542 | } |
---|
| 543 | |
---|
| 544 | { |
---|
| 545 | enable_stats = true; |
---|
| 546 | |
---|
| 547 | std::thread * threads[nthread]; |
---|
| 548 | unsigned i = 1; |
---|
| 549 | for(auto & t : threads) { |
---|
| 550 | t = new std::thread([&done, &list, &barrier, &global, slots, nslots](unsigned tid) { |
---|
| 551 | Random rand(tid + rdtscl()); |
---|
| 552 | |
---|
| 553 | local_stat_t local; |
---|
| 554 | barrier.wait(tid); |
---|
| 555 | |
---|
| 556 | // EXPERIMENT START |
---|
| 557 | |
---|
| 558 | runProducer_body(done, rand, slots, nslots, local, list); |
---|
| 559 | |
---|
| 560 | // EXPERIMENT END |
---|
| 561 | |
---|
| 562 | barrier.wait(tid); |
---|
| 563 | |
---|
| 564 | tally_stats(global, local); |
---|
| 565 | }, i++); |
---|
| 566 | } |
---|
| 567 | |
---|
| 568 | waitfor(duration, barrier, done); |
---|
| 569 | |
---|
| 570 | for(auto t : threads) { |
---|
| 571 | t->join(); |
---|
| 572 | delete t; |
---|
| 573 | } |
---|
| 574 | |
---|
| 575 | enable_stats = false; |
---|
| 576 | } |
---|
| 577 | |
---|
| 578 | { |
---|
| 579 | while(Node * node = list.pop()) { |
---|
| 580 | global.crc_out += node->value; |
---|
| 581 | delete node; |
---|
| 582 | } |
---|
| 583 | |
---|
| 584 | for(int i = 0; i < nslots; i++) { |
---|
| 585 | delete slots[i].node; |
---|
| 586 | } |
---|
| 587 | |
---|
| 588 | delete [] slots; |
---|
| 589 | } |
---|
| 590 | |
---|
| 591 | print_stats(duration, nthread, global); |
---|
| 592 | } |
---|
| 593 | |
---|
[807a632] | 594 | // ================================================================================================ |
---|
| 595 | __attribute__((noinline)) void runFairness_body( |
---|
| 596 | unsigned tid, |
---|
| 597 | size_t width, |
---|
| 598 | size_t length, |
---|
| 599 | int output[], |
---|
| 600 | std::atomic_size_t & count, |
---|
| 601 | Node initial_nodes[], |
---|
| 602 | unsigned nnodes, |
---|
| 603 | local_stat_t & local, |
---|
[b232745] | 604 | LIST_VARIANT<Node> & list |
---|
[807a632] | 605 | ) { |
---|
| 606 | Node * nodes[nnodes]; |
---|
| 607 | { |
---|
| 608 | unsigned i = 0; |
---|
| 609 | for(auto & n : nodes) { |
---|
| 610 | n = &initial_nodes[i++]; |
---|
| 611 | } |
---|
[9421f3d8] | 612 | } |
---|
| 613 | |
---|
[807a632] | 614 | while(__builtin_expect(0 != count.load(std::memory_order_relaxed), true)) { |
---|
| 615 | |
---|
| 616 | for(Node * & node : nodes) { |
---|
| 617 | local.crc_in += node->id; |
---|
| 618 | list.push(node); |
---|
| 619 | local.in++; |
---|
| 620 | } |
---|
| 621 | |
---|
| 622 | // ----- |
---|
| 623 | |
---|
| 624 | for(Node * & node : nodes) { |
---|
| 625 | node = list.pop(); |
---|
| 626 | assert(node); |
---|
| 627 | |
---|
| 628 | if (unsigned(node->value) < length) { |
---|
| 629 | size_t idx = (node->value * width) + node->id; |
---|
| 630 | assert(idx < (width * length)); |
---|
| 631 | output[idx] = tid; |
---|
| 632 | } |
---|
| 633 | |
---|
| 634 | node->value++; |
---|
| 635 | if(unsigned(node->value) == length) count--; |
---|
| 636 | |
---|
| 637 | local.crc_out += node->id; |
---|
| 638 | local.out++; |
---|
| 639 | } |
---|
| 640 | } |
---|
| 641 | } |
---|
| 642 | |
---|
| 643 | void runFairness(unsigned nthread, unsigned nqueues, double duration, unsigned nnodes, const std::string & output) { |
---|
| 644 | std::cout << "Fairness Benchmark, outputing to : " << output << std::endl; |
---|
| 645 | |
---|
| 646 | // Barrier for synchronization |
---|
| 647 | barrier_t barrier(nthread + 1); |
---|
| 648 | |
---|
| 649 | // Data to check everything is OK |
---|
| 650 | global_stat_t global; |
---|
| 651 | |
---|
| 652 | std::cout << "Initializing "; |
---|
| 653 | |
---|
| 654 | // Check fairness by creating a png of where the threads ran |
---|
| 655 | size_t width = nthread * nnodes; |
---|
| 656 | size_t length = 100000; |
---|
| 657 | |
---|
| 658 | std::unique_ptr<int[]> data_out { new int[width * length] }; |
---|
| 659 | |
---|
| 660 | // Flag to signal termination |
---|
| 661 | std::atomic_size_t count = width; |
---|
| 662 | |
---|
| 663 | // List being tested |
---|
[b232745] | 664 | LIST_VARIANT<Node> list = { nthread, nqueues }; |
---|
[807a632] | 665 | { |
---|
| 666 | enable_stats = true; |
---|
| 667 | |
---|
| 668 | std::thread * threads[nthread]; |
---|
| 669 | unsigned i = 1; |
---|
| 670 | for(auto & t : threads) { |
---|
| 671 | t = new std::thread([&count, &list, &barrier, &global, nnodes, width, length, data_out = data_out.get()](unsigned tid) { |
---|
| 672 | unsigned int start = (tid - 1) * nnodes; |
---|
| 673 | Node nodes[nnodes]; |
---|
| 674 | for(auto & n : nodes) { |
---|
| 675 | n.id = start; |
---|
| 676 | n.value = 0; |
---|
| 677 | start++; |
---|
| 678 | } |
---|
| 679 | |
---|
| 680 | local_stat_t local; |
---|
| 681 | |
---|
| 682 | // affinity(tid); |
---|
| 683 | |
---|
| 684 | barrier.wait(tid); |
---|
| 685 | |
---|
| 686 | // EXPERIMENT START |
---|
| 687 | |
---|
| 688 | runFairness_body(tid, width, length, data_out, count, nodes, nnodes, local, list); |
---|
| 689 | |
---|
| 690 | // EXPERIMENT END |
---|
| 691 | |
---|
| 692 | barrier.wait(tid); |
---|
| 693 | |
---|
| 694 | for(const auto & n : nodes) { |
---|
[7530049d] | 695 | local.valmax = std::max(local.valmax, size_t(n.value)); |
---|
| 696 | local.valmin = std::min(local.valmin, size_t(n.value)); |
---|
[807a632] | 697 | } |
---|
| 698 | |
---|
| 699 | tally_stats(global, local); |
---|
| 700 | }, i++); |
---|
| 701 | } |
---|
| 702 | |
---|
| 703 | waitfor(duration, barrier, count); |
---|
| 704 | |
---|
| 705 | for(auto t : threads) { |
---|
| 706 | t->join(); |
---|
| 707 | delete t; |
---|
| 708 | } |
---|
| 709 | |
---|
| 710 | enable_stats = false; |
---|
| 711 | } |
---|
[9421f3d8] | 712 | |
---|
| 713 | print_stats(duration, nthread, global); |
---|
[807a632] | 714 | |
---|
[95cb63b] | 715 | // save_fairness(data_out.get(), 100, nthread, width, length, output); |
---|
[b2a37b0] | 716 | } |
---|
| 717 | |
---|
[807a632] | 718 | // ================================================================================================ |
---|
| 719 | |
---|
[9421f3d8] | 720 | bool iequals(const std::string& a, const std::string& b) |
---|
| 721 | { |
---|
| 722 | return std::equal(a.begin(), a.end(), |
---|
| 723 | b.begin(), b.end(), |
---|
| 724 | [](char a, char b) { |
---|
| 725 | return std::tolower(a) == std::tolower(b); |
---|
| 726 | }); |
---|
[b2a37b0] | 727 | } |
---|
| 728 | |
---|
| 729 | int main(int argc, char * argv[]) { |
---|
| 730 | |
---|
| 731 | double duration = 5.0; |
---|
| 732 | unsigned nthreads = 2; |
---|
[9421f3d8] | 733 | unsigned nqueues = 4; |
---|
| 734 | unsigned nnodes = 100; |
---|
| 735 | unsigned nslots = 100; |
---|
[807a632] | 736 | std::string out = "fairness.png"; |
---|
[9421f3d8] | 737 | |
---|
| 738 | enum { |
---|
| 739 | Churn, |
---|
| 740 | PingPong, |
---|
[95cb63b] | 741 | Producer, |
---|
[807a632] | 742 | Fairness, |
---|
[9421f3d8] | 743 | NONE |
---|
| 744 | } benchmark = NONE; |
---|
[b2a37b0] | 745 | |
---|
| 746 | std::cout.imbue(std::locale("")); |
---|
| 747 | |
---|
[9421f3d8] | 748 | for(;;) { |
---|
| 749 | static struct option options[] = { |
---|
| 750 | {"duration", required_argument, 0, 'd'}, |
---|
| 751 | {"nthreads", required_argument, 0, 't'}, |
---|
| 752 | {"nqueues", required_argument, 0, 'q'}, |
---|
| 753 | {"benchmark", required_argument, 0, 'b'}, |
---|
| 754 | {0, 0, 0, 0} |
---|
| 755 | }; |
---|
| 756 | |
---|
| 757 | int idx = 0; |
---|
| 758 | int opt = getopt_long(argc, argv, "d:t:q:b:", options, &idx); |
---|
| 759 | |
---|
| 760 | std::string arg = optarg ? optarg : ""; |
---|
| 761 | size_t len = 0; |
---|
| 762 | switch(opt) { |
---|
| 763 | // Exit Case |
---|
| 764 | case -1: |
---|
| 765 | /* paranoid */ assert(optind <= argc); |
---|
| 766 | switch(benchmark) { |
---|
| 767 | case NONE: |
---|
| 768 | std::cerr << "Must specify a benchmark" << std::endl; |
---|
| 769 | goto usage; |
---|
| 770 | case PingPong: |
---|
| 771 | nnodes = 1; |
---|
| 772 | switch(argc - optind) { |
---|
| 773 | case 0: break; |
---|
| 774 | case 1: |
---|
| 775 | try { |
---|
| 776 | arg = optarg = argv[optind]; |
---|
[7530049d] | 777 | nnodes = std::stoul(optarg, &len); |
---|
[9421f3d8] | 778 | if(len != arg.size()) { throw std::invalid_argument(""); } |
---|
| 779 | } catch(std::invalid_argument &) { |
---|
| 780 | std::cerr << "Number of nodes must be a positive integer, was " << arg << std::endl; |
---|
| 781 | goto usage; |
---|
| 782 | } |
---|
| 783 | break; |
---|
| 784 | default: |
---|
[95cb63b] | 785 | std::cerr << "'PingPong' benchmark doesn't accept more than 1 extra arguments" << std::endl; |
---|
| 786 | goto usage; |
---|
| 787 | } |
---|
| 788 | break; |
---|
| 789 | case Producer: |
---|
| 790 | nnodes = 32; |
---|
| 791 | switch(argc - optind) { |
---|
| 792 | case 0: break; |
---|
| 793 | case 1: |
---|
| 794 | try { |
---|
| 795 | arg = optarg = argv[optind]; |
---|
[7530049d] | 796 | nnodes = std::stoul(optarg, &len); |
---|
[95cb63b] | 797 | if(len != arg.size()) { throw std::invalid_argument(""); } |
---|
| 798 | } catch(std::invalid_argument &) { |
---|
| 799 | std::cerr << "Number of nodes must be a positive integer, was " << arg << std::endl; |
---|
| 800 | goto usage; |
---|
| 801 | } |
---|
| 802 | break; |
---|
| 803 | default: |
---|
| 804 | std::cerr << "'Producer' benchmark doesn't accept more than 1 extra arguments" << std::endl; |
---|
[9421f3d8] | 805 | goto usage; |
---|
| 806 | } |
---|
| 807 | break; |
---|
| 808 | case Churn: |
---|
| 809 | nnodes = 100; |
---|
| 810 | nslots = 100; |
---|
| 811 | switch(argc - optind) { |
---|
| 812 | case 0: break; |
---|
| 813 | case 1: |
---|
| 814 | try { |
---|
| 815 | arg = optarg = argv[optind]; |
---|
[7530049d] | 816 | nnodes = std::stoul(optarg, &len); |
---|
[9421f3d8] | 817 | if(len != arg.size()) { throw std::invalid_argument(""); } |
---|
| 818 | nslots = nnodes; |
---|
| 819 | } catch(std::invalid_argument &) { |
---|
| 820 | std::cerr << "Number of nodes must be a positive integer, was " << arg << std::endl; |
---|
| 821 | goto usage; |
---|
| 822 | } |
---|
| 823 | break; |
---|
| 824 | case 2: |
---|
| 825 | try { |
---|
| 826 | arg = optarg = argv[optind]; |
---|
[7530049d] | 827 | nnodes = std::stoul(optarg, &len); |
---|
[9421f3d8] | 828 | if(len != arg.size()) { throw std::invalid_argument(""); } |
---|
| 829 | } catch(std::invalid_argument &) { |
---|
| 830 | std::cerr << "Number of nodes must be a positive integer, was " << arg << std::endl; |
---|
| 831 | goto usage; |
---|
| 832 | } |
---|
| 833 | try { |
---|
| 834 | arg = optarg = argv[optind + 1]; |
---|
[7530049d] | 835 | nslots = std::stoul(optarg, &len); |
---|
[9421f3d8] | 836 | if(len != arg.size()) { throw std::invalid_argument(""); } |
---|
| 837 | } catch(std::invalid_argument &) { |
---|
| 838 | std::cerr << "Number of slots must be a positive integer, was " << arg << std::endl; |
---|
| 839 | goto usage; |
---|
| 840 | } |
---|
| 841 | break; |
---|
| 842 | default: |
---|
| 843 | std::cerr << "'Churn' benchmark doesn't accept more than 2 extra arguments" << std::endl; |
---|
| 844 | goto usage; |
---|
| 845 | } |
---|
| 846 | break; |
---|
[807a632] | 847 | case Fairness: |
---|
| 848 | nnodes = 1; |
---|
| 849 | switch(argc - optind) { |
---|
| 850 | case 0: break; |
---|
| 851 | case 1: |
---|
| 852 | arg = optarg = argv[optind]; |
---|
| 853 | out = arg; |
---|
| 854 | break; |
---|
| 855 | default: |
---|
| 856 | std::cerr << "'Churn' benchmark doesn't accept more than 2 extra arguments" << std::endl; |
---|
| 857 | goto usage; |
---|
| 858 | } |
---|
[9421f3d8] | 859 | } |
---|
| 860 | goto run; |
---|
| 861 | // Benchmarks |
---|
| 862 | case 'b': |
---|
| 863 | if(benchmark != NONE) { |
---|
| 864 | std::cerr << "Only when benchmark can be run" << std::endl; |
---|
| 865 | goto usage; |
---|
| 866 | } |
---|
| 867 | if(iequals(arg, "churn")) { |
---|
| 868 | benchmark = Churn; |
---|
| 869 | break; |
---|
| 870 | } |
---|
| 871 | if(iequals(arg, "pingpong")) { |
---|
| 872 | benchmark = PingPong; |
---|
| 873 | break; |
---|
| 874 | } |
---|
[95cb63b] | 875 | if(iequals(arg, "producer")) { |
---|
| 876 | benchmark = Producer; |
---|
| 877 | break; |
---|
| 878 | } |
---|
[807a632] | 879 | if(iequals(arg, "fairness")) { |
---|
| 880 | benchmark = Fairness; |
---|
| 881 | break; |
---|
| 882 | } |
---|
[9421f3d8] | 883 | std::cerr << "Unkown benchmark " << arg << std::endl; |
---|
| 884 | goto usage; |
---|
| 885 | // Numeric Arguments |
---|
| 886 | case 'd': |
---|
| 887 | try { |
---|
[7530049d] | 888 | duration = std::stod(optarg, &len); |
---|
[9421f3d8] | 889 | if(len != arg.size()) { throw std::invalid_argument(""); } |
---|
| 890 | } catch(std::invalid_argument &) { |
---|
| 891 | std::cerr << "Duration must be a valid double, was " << arg << std::endl; |
---|
| 892 | goto usage; |
---|
| 893 | } |
---|
| 894 | break; |
---|
| 895 | case 't': |
---|
| 896 | try { |
---|
[7530049d] | 897 | nthreads = std::stoul(optarg, &len); |
---|
[9421f3d8] | 898 | if(len != arg.size()) { throw std::invalid_argument(""); } |
---|
| 899 | } catch(std::invalid_argument &) { |
---|
| 900 | std::cerr << "Number of threads must be a positive integer, was " << arg << std::endl; |
---|
| 901 | goto usage; |
---|
| 902 | } |
---|
| 903 | break; |
---|
| 904 | case 'q': |
---|
| 905 | try { |
---|
[7530049d] | 906 | nqueues = std::stoul(optarg, &len); |
---|
[9421f3d8] | 907 | if(len != arg.size()) { throw std::invalid_argument(""); } |
---|
| 908 | } catch(std::invalid_argument &) { |
---|
| 909 | std::cerr << "Number of queues must be a positive integer, was " << arg << std::endl; |
---|
| 910 | goto usage; |
---|
| 911 | } |
---|
| 912 | break; |
---|
| 913 | // Other cases |
---|
| 914 | default: /* ? */ |
---|
| 915 | std::cerr << opt << std::endl; |
---|
| 916 | usage: |
---|
| 917 | std::cerr << "Usage: " << argv[0] << ": [options] -b churn [NNODES] [NSLOTS = NNODES]" << std::endl; |
---|
| 918 | std::cerr << " or: " << argv[0] << ": [options] -b pingpong [NNODES]" << std::endl; |
---|
[95cb63b] | 919 | std::cerr << " or: " << argv[0] << ": [options] -b producer [NNODES]" << std::endl; |
---|
[9421f3d8] | 920 | std::cerr << std::endl; |
---|
| 921 | std::cerr << " -d, --duration=DURATION Duration of the experiment, in seconds" << std::endl; |
---|
| 922 | std::cerr << " -t, --nthreads=NTHREADS Number of kernel threads" << std::endl; |
---|
| 923 | std::cerr << " -q, --nqueues=NQUEUES Number of queues per threads" << std::endl; |
---|
| 924 | std::exit(1); |
---|
[b2a37b0] | 925 | } |
---|
| 926 | } |
---|
[9421f3d8] | 927 | run: |
---|
[b2a37b0] | 928 | |
---|
| 929 | check_cache_line_size(); |
---|
| 930 | |
---|
[50aeb6f] | 931 | std::cout << "Running " << nthreads << " threads (" << (nthreads * nqueues) << " queues) for " << duration << " seconds" << std::endl; |
---|
[b232745] | 932 | std::cout << "Relaxed list variant: " << LIST_VARIANT<Node>::name() << std::endl; |
---|
[9421f3d8] | 933 | switch(benchmark) { |
---|
| 934 | case Churn: |
---|
| 935 | runChurn(nthreads, nqueues, duration, nnodes, nslots); |
---|
| 936 | break; |
---|
| 937 | case PingPong: |
---|
| 938 | runPingPong(nthreads, nqueues, duration, nnodes); |
---|
| 939 | break; |
---|
[95cb63b] | 940 | case Producer: |
---|
| 941 | runProducer(nthreads, nqueues, duration, nnodes); |
---|
| 942 | break; |
---|
[807a632] | 943 | case Fairness: |
---|
| 944 | runFairness(nthreads, nqueues, duration, nnodes, out); |
---|
| 945 | break; |
---|
[9421f3d8] | 946 | default: |
---|
| 947 | abort(); |
---|
| 948 | } |
---|
[b2a37b0] | 949 | return 0; |
---|
| 950 | } |
---|
| 951 | |
---|
[807a632] | 952 | const char * __my_progname = "Relaxed List"; |
---|
| 953 | |
---|
| 954 | struct rgb_t { |
---|
| 955 | double r; // a fraction between 0 and 1 |
---|
| 956 | double g; // a fraction between 0 and 1 |
---|
| 957 | double b; // a fraction between 0 and 1 |
---|
| 958 | }; |
---|
| 959 | |
---|
| 960 | struct hsv_t { |
---|
| 961 | double h; // angle in degrees |
---|
| 962 | double s; // a fraction between 0 and 1 |
---|
| 963 | double v; // a fraction between 0 and 1 |
---|
| 964 | }; |
---|
| 965 | |
---|
| 966 | rgb_t hsv2rgb(hsv_t in) { |
---|
| 967 | double hh, p, q, t, ff; |
---|
| 968 | long i; |
---|
| 969 | rgb_t out; |
---|
| 970 | |
---|
| 971 | if(in.s <= 0.0) { // < is bogus, just shuts up warnings |
---|
| 972 | out.r = in.v; |
---|
| 973 | out.g = in.v; |
---|
| 974 | out.b = in.v; |
---|
| 975 | return out; |
---|
| 976 | } |
---|
| 977 | hh = in.h; |
---|
| 978 | if(hh >= 360.0) hh = 0.0; |
---|
| 979 | hh /= 60.0; |
---|
| 980 | i = (long)hh; |
---|
| 981 | ff = hh - i; |
---|
| 982 | p = in.v * (1.0 - in.s); |
---|
| 983 | q = in.v * (1.0 - (in.s * ff)); |
---|
| 984 | t = in.v * (1.0 - (in.s * (1.0 - ff))); |
---|
| 985 | |
---|
| 986 | switch(i) { |
---|
| 987 | case 0: |
---|
| 988 | out.r = in.v; |
---|
| 989 | out.g = t; |
---|
| 990 | out.b = p; |
---|
| 991 | break; |
---|
| 992 | case 1: |
---|
| 993 | out.r = q; |
---|
| 994 | out.g = in.v; |
---|
| 995 | out.b = p; |
---|
| 996 | break; |
---|
| 997 | case 2: |
---|
| 998 | out.r = p; |
---|
| 999 | out.g = in.v; |
---|
| 1000 | out.b = t; |
---|
| 1001 | break; |
---|
| 1002 | |
---|
| 1003 | case 3: |
---|
| 1004 | out.r = p; |
---|
| 1005 | out.g = q; |
---|
| 1006 | out.b = in.v; |
---|
| 1007 | break; |
---|
| 1008 | case 4: |
---|
| 1009 | out.r = t; |
---|
| 1010 | out.g = p; |
---|
| 1011 | out.b = in.v; |
---|
| 1012 | break; |
---|
| 1013 | case 5: |
---|
| 1014 | default: |
---|
| 1015 | out.r = in.v; |
---|
| 1016 | out.g = p; |
---|
| 1017 | out.b = q; |
---|
| 1018 | break; |
---|
| 1019 | } |
---|
| 1020 | return out; |
---|
| 1021 | } |
---|
| 1022 | |
---|
[95cb63b] | 1023 | // void save_fairness(const int data[], int factor, unsigned nthreads, size_t columns, size_t rows, const std::string & output) { |
---|
| 1024 | // std::ofstream os(output); |
---|
| 1025 | // os << "<html>\n"; |
---|
| 1026 | // os << "<head>\n"; |
---|
| 1027 | // os << "<style>\n"; |
---|
| 1028 | // os << "</style>\n"; |
---|
| 1029 | // os << "</head>\n"; |
---|
| 1030 | // os << "<body>\n"; |
---|
| 1031 | // os << "<table style=\"width=100%\">\n"; |
---|
| 1032 | |
---|
| 1033 | // size_t idx = 0; |
---|
| 1034 | // for(size_t r = 0ul; r < rows; r++) { |
---|
| 1035 | // os << "<tr>\n"; |
---|
| 1036 | // for(size_t c = 0ul; c < columns; c++) { |
---|
| 1037 | // os << "<td class=\"custom custom" << data[idx] << "\"></td>\n"; |
---|
| 1038 | // idx++; |
---|
| 1039 | // } |
---|
| 1040 | // os << "</tr>\n"; |
---|
| 1041 | // } |
---|
| 1042 | |
---|
| 1043 | // os << "</table>\n"; |
---|
| 1044 | // os << "</body>\n"; |
---|
| 1045 | // os << "</html>\n"; |
---|
| 1046 | // os << std::endl; |
---|
| 1047 | // } |
---|
| 1048 | |
---|
| 1049 | // #include <png.h> |
---|
| 1050 | // #include <setjmp.h> |
---|
[807a632] | 1051 | |
---|
| 1052 | /* |
---|
| 1053 | void save_fairness(const int data[], int factor, unsigned nthreads, size_t columns, size_t rows, const std::string & output) { |
---|
| 1054 | int width = columns * factor; |
---|
| 1055 | int height = rows / factor; |
---|
| 1056 | |
---|
| 1057 | int code = 0; |
---|
| 1058 | int idx = 0; |
---|
| 1059 | FILE *fp = NULL; |
---|
| 1060 | png_structp png_ptr = NULL; |
---|
| 1061 | png_infop info_ptr = NULL; |
---|
| 1062 | png_bytep row = NULL; |
---|
| 1063 | |
---|
| 1064 | // Open file for writing (binary mode) |
---|
| 1065 | fp = fopen(output.c_str(), "wb"); |
---|
| 1066 | if (fp == NULL) { |
---|
| 1067 | fprintf(stderr, "Could not open file %s for writing\n", output.c_str()); |
---|
| 1068 | code = 1; |
---|
| 1069 | goto finalise; |
---|
| 1070 | } |
---|
| 1071 | |
---|
| 1072 | // Initialize write structure |
---|
| 1073 | png_ptr = png_create_write_struct(PNG_LIBPNG_VER_STRING, NULL, NULL, NULL); |
---|
| 1074 | if (png_ptr == NULL) { |
---|
| 1075 | fprintf(stderr, "Could not allocate write struct\n"); |
---|
| 1076 | code = 1; |
---|
| 1077 | goto finalise; |
---|
| 1078 | } |
---|
| 1079 | |
---|
| 1080 | // Initialize info structure |
---|
| 1081 | info_ptr = png_create_info_struct(png_ptr); |
---|
| 1082 | if (info_ptr == NULL) { |
---|
| 1083 | fprintf(stderr, "Could not allocate info struct\n"); |
---|
| 1084 | code = 1; |
---|
| 1085 | goto finalise; |
---|
| 1086 | } |
---|
| 1087 | |
---|
| 1088 | // Setup Exception handling |
---|
| 1089 | if (setjmp(png_jmpbuf(png_ptr))) { |
---|
| 1090 | fprintf(stderr, "Error during png creation\n"); |
---|
| 1091 | code = 1; |
---|
| 1092 | goto finalise; |
---|
| 1093 | } |
---|
| 1094 | |
---|
| 1095 | png_init_io(png_ptr, fp); |
---|
| 1096 | |
---|
| 1097 | // Write header (8 bit colour depth) |
---|
| 1098 | png_set_IHDR(png_ptr, info_ptr, width, height, |
---|
| 1099 | 8, PNG_COLOR_TYPE_RGB, PNG_INTERLACE_NONE, |
---|
| 1100 | PNG_COMPRESSION_TYPE_BASE, PNG_FILTER_TYPE_BASE); |
---|
| 1101 | |
---|
| 1102 | png_write_info(png_ptr, info_ptr); |
---|
| 1103 | |
---|
| 1104 | // Allocate memory for one row (3 bytes per pixel - RGB) |
---|
| 1105 | row = (png_bytep) malloc(3 * width * sizeof(png_byte)); |
---|
| 1106 | |
---|
| 1107 | // Write image data |
---|
| 1108 | int x, y; |
---|
| 1109 | for (y=0 ; y<height ; y++) { |
---|
| 1110 | for (x=0 ; x<width ; x++) { |
---|
| 1111 | auto & r = row[(x * 3) + 0]; |
---|
| 1112 | auto & g = row[(x * 3) + 1]; |
---|
| 1113 | auto & b = row[(x * 3) + 2]; |
---|
| 1114 | assert(idx < (rows * columns)); |
---|
| 1115 | int color = data[idx] - 1; |
---|
| 1116 | assert(color < nthreads); |
---|
| 1117 | assert(color >= 0); |
---|
| 1118 | idx++; |
---|
| 1119 | |
---|
| 1120 | double angle = double(color) / double(nthreads); |
---|
| 1121 | |
---|
| 1122 | auto c = hsv2rgb({ 360.0 * angle, 0.8, 0.8 }); |
---|
| 1123 | |
---|
| 1124 | r = char(c.r * 255.0); |
---|
| 1125 | g = char(c.g * 255.0); |
---|
| 1126 | b = char(c.b * 255.0); |
---|
| 1127 | |
---|
| 1128 | } |
---|
| 1129 | png_write_row(png_ptr, row); |
---|
| 1130 | } |
---|
| 1131 | |
---|
| 1132 | assert(idx == (rows * columns)); |
---|
| 1133 | |
---|
| 1134 | // End write |
---|
| 1135 | png_write_end(png_ptr, NULL); |
---|
| 1136 | |
---|
| 1137 | finalise: |
---|
| 1138 | if (fp != NULL) fclose(fp); |
---|
| 1139 | if (info_ptr != NULL) png_free_data(png_ptr, info_ptr, PNG_FREE_ALL, -1); |
---|
| 1140 | if (png_ptr != NULL) png_destroy_write_struct(&png_ptr, (png_infopp)NULL); |
---|
| 1141 | if (row != NULL) free(row); |
---|
| 1142 | } |
---|
| 1143 | */ |
---|