| 1 | /**
|
|---|
| 2 | * polybench.c: This file is part of the PolyBench/C 3.2 test suite.
|
|---|
| 3 | *
|
|---|
| 4 | *
|
|---|
| 5 | * Contact: Louis-Noel Pouchet <pouchet@cse.ohio-state.edu>
|
|---|
| 6 | * Web address: http://polybench.sourceforge.net
|
|---|
| 7 | * License: /LICENSE.OSU.txt
|
|---|
| 8 | */
|
|---|
| 9 | #include <stdio.h>
|
|---|
| 10 | #include <string.h>
|
|---|
| 11 | #include <stdlib.h>
|
|---|
| 12 | #include <unistd.h>
|
|---|
| 13 | #include <assert.h>
|
|---|
| 14 | #include <time.h>
|
|---|
| 15 | #include <sys/time.h>
|
|---|
| 16 | #include <sys/resource.h>
|
|---|
| 17 | #include <sched.h>
|
|---|
| 18 | #include <math.h>
|
|---|
| 19 | #ifdef _OPENMP
|
|---|
| 20 | # include <omp.h>
|
|---|
| 21 | #endif
|
|---|
| 22 |
|
|---|
| 23 | /* By default, collect PAPI counters on thread 0. */
|
|---|
| 24 | #ifndef POLYBENCH_THREAD_MONITOR
|
|---|
| 25 | # define POLYBENCH_THREAD_MONITOR 0
|
|---|
| 26 | #endif
|
|---|
| 27 |
|
|---|
| 28 | /* Total LLC cache size. By default 32+MB.. */
|
|---|
| 29 | #ifndef POLYBENCH_CACHE_SIZE_KB
|
|---|
| 30 | # define POLYBENCH_CACHE_SIZE_KB 32770
|
|---|
| 31 | #endif
|
|---|
| 32 |
|
|---|
| 33 |
|
|---|
| 34 | int polybench_papi_counters_threadid = POLYBENCH_THREAD_MONITOR;
|
|---|
| 35 | double polybench_program_total_flops = 0;
|
|---|
| 36 |
|
|---|
| 37 | #ifdef POLYBENCH_PAPI
|
|---|
| 38 | # include <papi.h>
|
|---|
| 39 | # define POLYBENCH_MAX_NB_PAPI_COUNTERS 96
|
|---|
| 40 | char* _polybench_papi_eventlist[] = {
|
|---|
| 41 | #include "papi_counters.list"
|
|---|
| 42 | NULL
|
|---|
| 43 | };
|
|---|
| 44 | int polybench_papi_eventset;
|
|---|
| 45 | int polybench_papi_eventlist[POLYBENCH_MAX_NB_PAPI_COUNTERS];
|
|---|
| 46 | long_long polybench_papi_values[POLYBENCH_MAX_NB_PAPI_COUNTERS];
|
|---|
| 47 |
|
|---|
| 48 | #endif
|
|---|
| 49 |
|
|---|
| 50 |
|
|---|
| 51 | /* Timer code (gettimeofday). */
|
|---|
| 52 | double polybench_t_start, polybench_t_end;
|
|---|
| 53 | /* Timer code (RDTSC). */
|
|---|
| 54 | unsigned long long int polybench_c_start, polybench_c_end;
|
|---|
| 55 |
|
|---|
| 56 | static
|
|---|
| 57 | double rtclock()
|
|---|
| 58 | {
|
|---|
| 59 | #ifdef POLYBENCH_TIME
|
|---|
| 60 | struct timeval Tp;
|
|---|
| 61 | int stat;
|
|---|
| 62 | stat = gettimeofday (&Tp, NULL);
|
|---|
| 63 | if (stat != 0)
|
|---|
| 64 | printf ("Error return from gettimeofday: %d", stat);
|
|---|
| 65 | return (Tp.tv_sec + Tp.tv_usec * 1.0e-6);
|
|---|
| 66 | #else
|
|---|
| 67 | return 0;
|
|---|
| 68 | #endif
|
|---|
| 69 | }
|
|---|
| 70 |
|
|---|
| 71 |
|
|---|
| 72 | #ifdef POLYBENCH_CYCLE_ACCURATE_TIMER
|
|---|
| 73 | static
|
|---|
| 74 | unsigned long long int rdtsc()
|
|---|
| 75 | {
|
|---|
| 76 | unsigned long long int ret = 0;
|
|---|
| 77 | unsigned int cycles_lo;
|
|---|
| 78 | unsigned int cycles_hi;
|
|---|
| 79 | __asm__ volatile ("RDTSC" : "=a" (cycles_lo), "=d" (cycles_hi));
|
|---|
| 80 | ret = (unsigned long long int)cycles_hi << 32 | cycles_lo;
|
|---|
| 81 |
|
|---|
| 82 | return ret;
|
|---|
| 83 | }
|
|---|
| 84 | #endif
|
|---|
| 85 |
|
|---|
| 86 | void polybench_flush_cache()
|
|---|
| 87 | {
|
|---|
| 88 | int cs = POLYBENCH_CACHE_SIZE_KB * 1024 / sizeof(double);
|
|---|
| 89 | double* flush = (double*) calloc (cs, sizeof(double));
|
|---|
| 90 | int i;
|
|---|
| 91 | double tmp = 0.0;
|
|---|
| 92 | #ifdef _OPENMP
|
|---|
| 93 | #pragma omp parallel for reduction(+:tmp)
|
|---|
| 94 | #endif
|
|---|
| 95 | for (i = 0; i < cs; i++)
|
|---|
| 96 | tmp += flush[i];
|
|---|
| 97 | assert (tmp <= 10.0);
|
|---|
| 98 | free (flush);
|
|---|
| 99 | }
|
|---|
| 100 |
|
|---|
| 101 |
|
|---|
| 102 | #ifdef POLYBENCH_LINUX_FIFO_SCHEDULER
|
|---|
| 103 | void polybench_linux_fifo_scheduler()
|
|---|
| 104 | {
|
|---|
| 105 | /* Use FIFO scheduler to limit OS interference. Program must be run
|
|---|
| 106 | as root, and this works only for Linux kernels. */
|
|---|
| 107 | struct sched_param schedParam;
|
|---|
| 108 | schedParam.sched_priority = sched_get_priority_max (SCHED_FIFO);
|
|---|
| 109 | sched_setscheduler (0, SCHED_FIFO, &schedParam);
|
|---|
| 110 | }
|
|---|
| 111 |
|
|---|
| 112 |
|
|---|
| 113 | void polybench_linux_standard_scheduler()
|
|---|
| 114 | {
|
|---|
| 115 | /* Restore to standard scheduler policy. */
|
|---|
| 116 | struct sched_param schedParam;
|
|---|
| 117 | schedParam.sched_priority = sched_get_priority_max (SCHED_OTHER);
|
|---|
| 118 | sched_setscheduler (0, SCHED_OTHER, &schedParam);
|
|---|
| 119 | }
|
|---|
| 120 | #endif
|
|---|
| 121 |
|
|---|
| 122 | #ifdef POLYBENCH_PAPI
|
|---|
| 123 |
|
|---|
| 124 | static
|
|---|
| 125 | void test_fail(char *file, int line, char *call, int retval)
|
|---|
| 126 | {
|
|---|
| 127 | char buf[128];
|
|---|
| 128 |
|
|---|
| 129 | memset(buf, '\0', sizeof(buf));
|
|---|
| 130 | if (retval != 0)
|
|---|
| 131 | fprintf (stdout,"%-40s FAILED\nLine # %d\n", file, line);
|
|---|
| 132 | else
|
|---|
| 133 | {
|
|---|
| 134 | fprintf (stdout,"%-40s SKIPPED\n", file);
|
|---|
| 135 | fprintf (stdout,"Line # %d\n", line);
|
|---|
| 136 | }
|
|---|
| 137 | if (retval == PAPI_ESYS)
|
|---|
| 138 | {
|
|---|
| 139 | sprintf (buf, "System error in %s", call);
|
|---|
| 140 | perror (buf);
|
|---|
| 141 | }
|
|---|
| 142 | else if (retval > 0)
|
|---|
| 143 | fprintf (stdout,"Error: %s\n", call);
|
|---|
| 144 | else if (retval == 0)
|
|---|
| 145 | fprintf (stdout,"Error: %s\n", call);
|
|---|
| 146 | else
|
|---|
| 147 | {
|
|---|
| 148 | char errstring[PAPI_MAX_STR_LEN];
|
|---|
| 149 | PAPI_perror (retval, errstring, PAPI_MAX_STR_LEN);
|
|---|
| 150 | fprintf (stdout,"Error in %s: %s\n", call, errstring);
|
|---|
| 151 | }
|
|---|
| 152 | fprintf (stdout,"\n");
|
|---|
| 153 | if (PAPI_is_initialized ())
|
|---|
| 154 | PAPI_shutdown ();
|
|---|
| 155 | exit (1);
|
|---|
| 156 | }
|
|---|
| 157 |
|
|---|
| 158 |
|
|---|
| 159 | void polybench_papi_init()
|
|---|
| 160 | {
|
|---|
| 161 | # ifdef _OPENMP
|
|---|
| 162 | #pragma omp parallel
|
|---|
| 163 | {
|
|---|
| 164 | #pragma omp master
|
|---|
| 165 | {
|
|---|
| 166 | if (omp_get_max_threads () < polybench_papi_counters_threadid)
|
|---|
| 167 | polybench_papi_counters_threadid = omp_get_max_threads () - 1;
|
|---|
| 168 | }
|
|---|
| 169 | #pragma omp barrier
|
|---|
| 170 |
|
|---|
| 171 | if (omp_get_thread_num () == polybench_papi_counters_threadid)
|
|---|
| 172 | {
|
|---|
| 173 | # endif
|
|---|
| 174 | int retval;
|
|---|
| 175 | polybench_papi_eventset = PAPI_NULL;
|
|---|
| 176 | if ((retval = PAPI_library_init (PAPI_VER_CURRENT)) != PAPI_VER_CURRENT)
|
|---|
| 177 | test_fail (__FILE__, __LINE__, "PAPI_library_init", retval);
|
|---|
| 178 | if ((retval = PAPI_create_eventset (&polybench_papi_eventset))
|
|---|
| 179 | != PAPI_OK)
|
|---|
| 180 | test_fail (__FILE__, __LINE__, "PAPI_create_eventset", retval);
|
|---|
| 181 | int k;
|
|---|
| 182 | for (k = 0; _polybench_papi_eventlist[k]; ++k)
|
|---|
| 183 | {
|
|---|
| 184 | if ((retval =
|
|---|
| 185 | PAPI_event_name_to_code (_polybench_papi_eventlist[k],
|
|---|
| 186 | &(polybench_papi_eventlist[k])))
|
|---|
| 187 | != PAPI_OK)
|
|---|
| 188 | test_fail (__FILE__, __LINE__, "PAPI_event_name_to_code", retval);
|
|---|
| 189 | }
|
|---|
| 190 | polybench_papi_eventlist[k] = 0;
|
|---|
| 191 |
|
|---|
| 192 |
|
|---|
| 193 | # ifdef _OPENMP
|
|---|
| 194 | }
|
|---|
| 195 | }
|
|---|
| 196 | #pragma omp barrier
|
|---|
| 197 | # endif
|
|---|
| 198 | }
|
|---|
| 199 |
|
|---|
| 200 |
|
|---|
| 201 | void polybench_papi_close()
|
|---|
| 202 | {
|
|---|
| 203 | # ifdef _OPENMP
|
|---|
| 204 | #pragma omp parallel
|
|---|
| 205 | {
|
|---|
| 206 | if (omp_get_thread_num () == polybench_papi_counters_threadid)
|
|---|
| 207 | {
|
|---|
| 208 | # endif
|
|---|
| 209 | int retval;
|
|---|
| 210 | if ((retval = PAPI_destroy_eventset (&polybench_papi_eventset))
|
|---|
| 211 | != PAPI_OK)
|
|---|
| 212 | test_fail (__FILE__, __LINE__, "PAPI_destroy_eventset", retval);
|
|---|
| 213 | if (PAPI_is_initialized ())
|
|---|
| 214 | PAPI_shutdown ();
|
|---|
| 215 | # ifdef _OPENMP
|
|---|
| 216 | }
|
|---|
| 217 | }
|
|---|
| 218 | #pragma omp barrier
|
|---|
| 219 | # endif
|
|---|
| 220 | }
|
|---|
| 221 |
|
|---|
| 222 | int polybench_papi_start_counter(int evid)
|
|---|
| 223 | {
|
|---|
| 224 | # ifndef POLYBENCH_NO_FLUSH_CACHE
|
|---|
| 225 | polybench_flush_cache();
|
|---|
| 226 | # endif
|
|---|
| 227 |
|
|---|
| 228 | # ifdef _OPENMP
|
|---|
| 229 | # pragma omp parallel
|
|---|
| 230 | {
|
|---|
| 231 | if (omp_get_thread_num () == polybench_papi_counters_threadid)
|
|---|
| 232 | {
|
|---|
| 233 | # endif
|
|---|
| 234 |
|
|---|
| 235 | int retval = 1;
|
|---|
| 236 | char descr[PAPI_MAX_STR_LEN];
|
|---|
| 237 | PAPI_event_info_t evinfo;
|
|---|
| 238 | PAPI_event_code_to_name (polybench_papi_eventlist[evid], descr);
|
|---|
| 239 | if (PAPI_add_event (polybench_papi_eventset,
|
|---|
| 240 | polybench_papi_eventlist[evid]) != PAPI_OK)
|
|---|
| 241 | test_fail (__FILE__, __LINE__, "PAPI_add_event", 1);
|
|---|
| 242 | if (PAPI_get_event_info (polybench_papi_eventlist[evid], &evinfo)
|
|---|
| 243 | != PAPI_OK)
|
|---|
| 244 | test_fail (__FILE__, __LINE__, "PAPI_get_event_info", retval);
|
|---|
| 245 | if ((retval = PAPI_start (polybench_papi_eventset)) != PAPI_OK)
|
|---|
| 246 | test_fail (__FILE__, __LINE__, "PAPI_start", retval);
|
|---|
| 247 | # ifdef _OPENMP
|
|---|
| 248 | }
|
|---|
| 249 | }
|
|---|
| 250 | #pragma omp barrier
|
|---|
| 251 | # endif
|
|---|
| 252 | return 0;
|
|---|
| 253 | }
|
|---|
| 254 |
|
|---|
| 255 |
|
|---|
| 256 | void polybench_papi_stop_counter(int evid)
|
|---|
| 257 | {
|
|---|
| 258 | # ifdef _OPENMP
|
|---|
| 259 | # pragma omp parallel
|
|---|
| 260 | {
|
|---|
| 261 | if (omp_get_thread_num () == polybench_papi_counters_threadid)
|
|---|
| 262 | {
|
|---|
| 263 | # endif
|
|---|
| 264 | int retval;
|
|---|
| 265 | long_long values[1];
|
|---|
| 266 | values[0] = 0;
|
|---|
| 267 | if ((retval = PAPI_read (polybench_papi_eventset, &values[0]))
|
|---|
| 268 | != PAPI_OK)
|
|---|
| 269 | test_fail (__FILE__, __LINE__, "PAPI_read", retval);
|
|---|
| 270 |
|
|---|
| 271 | if ((retval = PAPI_stop (polybench_papi_eventset, NULL)) != PAPI_OK)
|
|---|
| 272 | test_fail (__FILE__, __LINE__, "PAPI_stop", retval);
|
|---|
| 273 |
|
|---|
| 274 | polybench_papi_values[evid] = values[0];
|
|---|
| 275 |
|
|---|
| 276 | if ((retval = PAPI_remove_event
|
|---|
| 277 | (polybench_papi_eventset,
|
|---|
| 278 | polybench_papi_eventlist[evid])) != PAPI_OK)
|
|---|
| 279 | test_fail (__FILE__, __LINE__, "PAPI_remove_event", retval);
|
|---|
| 280 | # ifdef _OPENMP
|
|---|
| 281 | }
|
|---|
| 282 | }
|
|---|
| 283 | #pragma omp barrier
|
|---|
| 284 | # endif
|
|---|
| 285 | }
|
|---|
| 286 |
|
|---|
| 287 |
|
|---|
| 288 | void polybench_papi_print()
|
|---|
| 289 | {
|
|---|
| 290 | int verbose = 0;
|
|---|
| 291 | # ifdef _OPENMP
|
|---|
| 292 | # pragma omp parallel
|
|---|
| 293 | {
|
|---|
| 294 | if (omp_get_thread_num() == polybench_papi_counters_threadid)
|
|---|
| 295 | {
|
|---|
| 296 | #ifdef POLYBENCH_PAPI_VERBOSE
|
|---|
| 297 | verbose = 1;
|
|---|
| 298 | #endif
|
|---|
| 299 | if (verbose)
|
|---|
| 300 | printf ("On thread %d:\n", polybench_papi_counters_threadid);
|
|---|
| 301 | #endif
|
|---|
| 302 | int evid;
|
|---|
| 303 | for (evid = 0; polybench_papi_eventlist[evid] != 0; ++evid)
|
|---|
| 304 | {
|
|---|
| 305 | if (verbose)
|
|---|
| 306 | printf ("%s=", _polybench_papi_eventlist[evid]);
|
|---|
| 307 | printf ("%llu ", polybench_papi_values[evid]);
|
|---|
| 308 | if (verbose)
|
|---|
| 309 | printf ("\n");
|
|---|
| 310 | }
|
|---|
| 311 | printf ("\n");
|
|---|
| 312 | # ifdef _OPENMP
|
|---|
| 313 | }
|
|---|
| 314 | }
|
|---|
| 315 | #pragma omp barrier
|
|---|
| 316 | # endif
|
|---|
| 317 | }
|
|---|
| 318 |
|
|---|
| 319 | #endif
|
|---|
| 320 | /* ! POLYBENCH_PAPI */
|
|---|
| 321 |
|
|---|
| 322 | void polybench_prepare_instruments()
|
|---|
| 323 | {
|
|---|
| 324 | #ifndef POLYBENCH_NO_FLUSH_CACHE
|
|---|
| 325 | polybench_flush_cache ();
|
|---|
| 326 | #endif
|
|---|
| 327 | #ifdef POLYBENCH_LINUX_FIFO_SCHEDULER
|
|---|
| 328 | polybench_linux_fifo_scheduler ();
|
|---|
| 329 | #endif
|
|---|
| 330 | }
|
|---|
| 331 |
|
|---|
| 332 |
|
|---|
| 333 | void polybench_timer_start()
|
|---|
| 334 | {
|
|---|
| 335 | polybench_prepare_instruments ();
|
|---|
| 336 | #ifndef POLYBENCH_CYCLE_ACCURATE_TIMER
|
|---|
| 337 | polybench_t_start = rtclock ();
|
|---|
| 338 | #else
|
|---|
| 339 | polybench_c_start = rdtsc ();
|
|---|
| 340 | #endif
|
|---|
| 341 | }
|
|---|
| 342 |
|
|---|
| 343 |
|
|---|
| 344 | void polybench_timer_stop()
|
|---|
| 345 | {
|
|---|
| 346 | #ifndef POLYBENCH_CYCLE_ACCURATE_TIMER
|
|---|
| 347 | polybench_t_end = rtclock ();
|
|---|
| 348 | #else
|
|---|
| 349 | polybench_c_end = rdtsc ();
|
|---|
| 350 | #endif
|
|---|
| 351 | #ifdef POLYBENCH_LINUX_FIFO_SCHEDULER
|
|---|
| 352 | polybench_linux_standard_scheduler ();
|
|---|
| 353 | #endif
|
|---|
| 354 | }
|
|---|
| 355 |
|
|---|
| 356 |
|
|---|
| 357 | void polybench_timer_print()
|
|---|
| 358 | {
|
|---|
| 359 | #ifdef POLYBENCH_GFLOPS
|
|---|
| 360 | if (__polybench_program_total_flops == 0)
|
|---|
| 361 | {
|
|---|
| 362 | printf ("[PolyBench][WARNING] Program flops not defined, use polybench_set_program_flops(value)\n");
|
|---|
| 363 | printf ("%0.6lf\n", polybench_t_end - polybench_t_start);
|
|---|
| 364 | }
|
|---|
| 365 | else
|
|---|
| 366 | printf ("%0.2lf\n",
|
|---|
| 367 | (__polybench_program_total_flops /
|
|---|
| 368 | (double)(polybench_t_end - polybench_t_start)) / 1000000000);
|
|---|
| 369 | #else
|
|---|
| 370 | # ifndef POLYBENCH_CYCLE_ACCURATE_TIMER
|
|---|
| 371 | printf ("%0.6f\n", polybench_t_end - polybench_t_start);
|
|---|
| 372 | # else
|
|---|
| 373 | printf ("%Ld\n", polybench_c_end - polybench_c_start);
|
|---|
| 374 | # endif
|
|---|
| 375 | #endif
|
|---|
| 376 | }
|
|---|
| 377 |
|
|---|
| 378 |
|
|---|
| 379 |
|
|---|
| 380 | static
|
|---|
| 381 | void *
|
|---|
| 382 | xmalloc (size_t num)
|
|---|
| 383 | {
|
|---|
| 384 | void* nnew = NULL;
|
|---|
| 385 | int ret = posix_memalign (&nnew, 32, num);
|
|---|
| 386 | if (! nnew || ret)
|
|---|
| 387 | {
|
|---|
| 388 | fprintf (stderr, "[PolyBench] posix_memalign: cannot allocate memory");
|
|---|
| 389 | exit (1);
|
|---|
| 390 | }
|
|---|
| 391 | return nnew;
|
|---|
| 392 | }
|
|---|
| 393 |
|
|---|
| 394 |
|
|---|
| 395 | void* polybench_alloc_data(unsigned long long int n, int elt_size)
|
|---|
| 396 | {
|
|---|
| 397 | /// FIXME: detect overflow!
|
|---|
| 398 | size_t val = n;
|
|---|
| 399 | val *= elt_size;
|
|---|
| 400 | void* ret = xmalloc (val);
|
|---|
| 401 |
|
|---|
| 402 | return ret;
|
|---|
| 403 | }
|
|---|