rts/Papi.c

   1 /* -----------------------------------------------------------------------------
   2  * (c) The GHC Team 2006
   3  *
   4  * Initialization and use of the PAPI performance monitoring library
   5  *
   6  *
   7  * For adding events or add your processor counters modify
   8  *
   9  *   init_countable_events
  10  *   papi_report
  11  *
  12  * ---------------------------------------------------------------------------*/
  13
  14
  15 #ifdef USE_PAPI /* ugly */
  16
  17 #include <papi.h>
  18
  19 #include "Papi.h"
  20 #include "Rts.h"
  21 #include "RtsUtils.h"
  22 #include "Stats.h"
  23 #include "RtsFlags.h"
  24 #include "OSThreads.h"
  25
  26 // used to protect the aggregated counters
  27 #ifdef THREADED_RTS
  28 static Mutex papi_counter_mutex;
  29 #endif
  30
  31 struct _papi_events {
  32   int event_code;
  33   char * event_name;
  34 };
  35
  36 /* Beware, these counters are Opteron specific
  37  * I obtained the numbers using the papi_avail
  38  * and papi_native_avail utilities.
  39  * This is certainly not the official PAPI way
  40  * of doing things.
  41  */
  42 #define FR_BR 0x40000040
  43 #define FR_BR_MIS 0x40000041
  44 #define FR_BR_MISCOMPARE 0x40000048
  45 #define DC_ACCESS 0x40000019
  46 #define DC_MISS 0x4000001a
  47 #define FR_DISPATCH_STALLS 0x40000054
  48 #define FR_DISPATCH_STALLS_BR 0x40000055
  49 #define FR_DISPATCH_STALLS_FULL_REORDER 0x40000058
  50 #define FR_DISPATCH_STALLS_FULL_RESERVATION 0x40000059
  51 #define FR_DISPATCH_STALLS_FULL_LS 0x4000005b
  52 #define DC_L2_REFILL_MOES 0x40001e1b
  53 #define DC_SYS_REFILL_MOES 0x40001e1c
  54
  55 /* This is bad, it should be in a header */
  56 #define BIG_STRING_LEN 512
  57
  58
  59 #define PAPI_CHECK(CALL) \
  60   if((papi_error=(CALL)) != PAPI_OK) { \
  61    debugBelch("PAPI function failed in module %s at line %d with error code %d\n", \
  62               __FILE__,__LINE__,papi_error);                            \
  63   }
  64
  65 /* While PAPI reporting is going on this flag is on */
  66 int papi_is_reporting;
  67
  68 /* Event sets and counter arrays for GC and mutator */
  69
  70 int MutatorEvents = PAPI_NULL;
  71 int GCEvents = PAPI_NULL;
  72
  73 int papi_error;
  74
  75 /* Arbitrary, to avoid using malloc */
  76 #define MAX_PAPI_EVENTS 10
  77
  78 int n_papi_events = 0;
  79
  80
  81 /* Events counted during GC and Mutator execution */
  82 /* There's a trailing comma, do all C compilers accept that? */
  83 static struct _papi_events papi_events[MAX_PAPI_EVENTS];
  84 long_long MutatorCounters[MAX_PAPI_EVENTS];
  85 long_long GCCounters[MAX_PAPI_EVENTS];
  86
  87 long_long start_mutator_cycles;
  88 long_long start_gc_cycles;
  89 long_long mutator_cycles;
  90 long_long gc_cycles;
  91
  92
  93
  94 static long_long papi_counter(long_long values[],int event);
  95 static void papi_add_events(int EventSet);
  96
  97 /* If you want to add events to count, extend the
  98  * init_countable_events and the papi_report function.
  99  * Be aware that your processor can count a limited number
 100  * of events simultaneously, you can turn on multiplexing
 101  * to increase that number, though.
 102  */
 103 static void
 104 init_countable_events(void)
 105 {
 106 #define PAPI_ADD_EVENT(EVENT)                           \
 107     {                                                   \
 108         if (n_papi_events >= MAX_PAPI_EVENTS) {         \
 109            barf("too many PAPI events");                \
 110         }                                               \
 111         papi_events[n_papi_events].event_code = EVENT;  \
 112         papi_events[n_papi_events].event_name = #EVENT; \
 113         n_papi_events++;                                \
 114     }
 115
 116     PAPI_ADD_EVENT(PAPI_TOT_INS);
 117     if (RtsFlags.PapiFlags.eventType==PAPI_FLAG_BRANCH) {
 118         PAPI_ADD_EVENT(FR_BR);
 119         PAPI_ADD_EVENT(FR_BR_MIS);
 120         /* Docs are wrong? Opteron does not count indirect branch misses exclusively */
 121         PAPI_ADD_EVENT(FR_BR_MISCOMPARE);
 122     } else if (RtsFlags.PapiFlags.eventType==PAPI_FLAG_STALLS) {
 123         PAPI_ADD_EVENT(FR_DISPATCH_STALLS);
 124         PAPI_ADD_EVENT(FR_DISPATCH_STALLS_BR);
 125         PAPI_ADD_EVENT(FR_DISPATCH_STALLS_FULL_LS);
 126     } else if (RtsFlags.PapiFlags.eventType==PAPI_FLAG_CACHE_L1) {
 127         PAPI_ADD_EVENT(PAPI_L1_DCA);
 128         PAPI_ADD_EVENT(PAPI_L1_DCM);
 129     } else if (RtsFlags.PapiFlags.eventType==PAPI_FLAG_CACHE_L2) {
 130         PAPI_ADD_EVENT(PAPI_L2_DCA);
 131         PAPI_ADD_EVENT(PAPI_L2_DCM);
 132     } else if (RtsFlags.PapiFlags.eventType==PAPI_FLAG_CB_EVENTS) {
 133         PAPI_ADD_EVENT(DC_L2_REFILL_MOES);
 134         PAPI_ADD_EVENT(DC_SYS_REFILL_MOES);
 135         PAPI_ADD_EVENT(FR_BR_MIS);
 136     } else {
 137         PAPI_ADD_EVENT(PAPI_STL_ICY);
 138     }
 139
 140     // We might also consider:
 141     //  PAPI_BR_MSP     Conditional branch instructions mispredicted
 142     //  PAPI_RES_STL    Cycles stalled on any resource
 143 };
 144
 145
 146 static char temp[BIG_STRING_LEN];
 147
 148 void
 149 papi_mut_cycles()
 150 {
 151     ullong_format_string(mutator_cycles,temp,rtsTrue/*commas*/);
 152     statsPrintf("  (MUT_CYCLES)  : %s\n",temp);
 153 }
 154
 155 void
 156 papi_gc_cycles()
 157 {
 158     ullong_format_string(gc_cycles,temp,rtsTrue/*commas*/);
 159     statsPrintf("  (GC_CYCLES)  : %s\n",temp);
 160 }
 161
 162 /* This function reports counters for GC and mutator */
 163 static void
 164 papi_report(long_long PapiCounters[])
 165 {
 166
 167 /* Report the value of a counter */
 168 #define PAPI_REPORT(EVENTSET,EVENT) \
 169   { \
 170     ullong_format_string(papi_counter(EVENTSET,EVENT),temp,rtsTrue/*commas*/); \
 171     statsPrintf("  (" #EVENT ")  : %s\n",temp);                         \
 172   }
 173
 174 /* Report the value of a counter as a percentage of another counter */
 175 #define PAPI_REPORT_PCT(EVENTSET,EVENT,EVENTTOT) \
 176   statsPrintf("  (" #EVENT ") %% of (" #EVENTTOT ") : %.1f%%\n", \
 177               papi_counter(EVENTSET,EVENT)*100.0/papi_counter(EVENTSET,EVENTTOT))
 178
 179   /* I need to improve formatting aesthetics */
 180     PAPI_REPORT(PapiCounters,PAPI_TOT_INS);
 181
 182     if (RtsFlags.PapiFlags.eventType==PAPI_FLAG_BRANCH) {
 183         PAPI_REPORT(PapiCounters,FR_BR);
 184         PAPI_REPORT(PapiCounters,FR_BR_MIS);
 185         PAPI_REPORT_PCT(PapiCounters,FR_BR_MIS,FR_BR);
 186         PAPI_REPORT_PCT(PapiCounters,FR_BR_MISCOMPARE,FR_BR);
 187     }
 188
 189     else if(RtsFlags.PapiFlags.eventType==PAPI_FLAG_STALLS) {
 190         PAPI_REPORT(PapiCounters,FR_DISPATCH_STALLS);
 191         PAPI_REPORT(PapiCounters,FR_DISPATCH_STALLS_BR);
 192         //PAPI_REPORT_PCT(PapiCounters,FR_DISPATCH_STALLS_BR,PAPI_TOT_CYC);
 193         PAPI_REPORT(PapiCounters,FR_DISPATCH_STALLS_FULL_LS);
 194         //PAPI_REPORT_PCT(PapiCounters,FR_DISPATCH_STALLS_FULL_LS,PAPI_TOT_CYC);
 195     }
 196
 197     else if (RtsFlags.PapiFlags.eventType==PAPI_FLAG_CACHE_L1) {
 198         PAPI_REPORT(PapiCounters,PAPI_L1_DCA);
 199         PAPI_REPORT(PapiCounters,PAPI_L1_DCM);
 200         PAPI_REPORT_PCT(PapiCounters,PAPI_L1_DCM,PAPI_L1_DCA);
 201     }
 202
 203     else if (RtsFlags.PapiFlags.eventType==PAPI_FLAG_CACHE_L2) {
 204         PAPI_REPORT(PapiCounters,PAPI_L2_DCA);
 205         PAPI_REPORT(PapiCounters,PAPI_L2_DCM);
 206         PAPI_REPORT_PCT(PapiCounters,PAPI_L2_DCM,PAPI_L2_DCA);
 207     }
 208
 209     else if (RtsFlags.PapiFlags.eventType==PAPI_FLAG_CB_EVENTS) {
 210         PAPI_REPORT(PapiCounters,DC_L2_REFILL_MOES);
 211         PAPI_REPORT(PapiCounters,DC_SYS_REFILL_MOES);
 212         PAPI_REPORT(PapiCounters,FR_BR_MIS);
 213     }
 214
 215     else {
 216         PAPI_REPORT(PapiCounters,PAPI_STL_ICY);
 217     }
 218 }
 219
 220 void
 221 papi_stats_report (void)
 222 {
 223     statsPrintf("  -- CPU Mutator counters --\n");
 224     papi_mut_cycles();
 225     papi_report(MutatorCounters);
 226
 227     statsPrintf("\n  -- CPU GC counters --\n");
 228     papi_gc_cycles();
 229     papi_report(GCCounters);
 230 }
 231
 232 void
 233 papi_init_eventset (int *event_set)
 234 {
 235     PAPI_register_thread();
 236     PAPI_CHECK( PAPI_create_eventset(event_set));
 237     papi_add_events(*event_set);
 238 }
 239
 240 void
 241 papi_init (void)
 242 {
 243     /* Initialise the performance tracking library */
 244     int ver;
 245     if ((ver = PAPI_library_init(PAPI_VER_CURRENT)) != PAPI_VER_CURRENT) {
 246         if (ver > 0) {
 247             errorBelch("PAPI_library_init: wrong version: %x", ver);
 248             stg_exit(EXIT_FAILURE);
 249         } else {
 250             sysErrorBelch("PAPI_library_init");
 251             stg_exit(EXIT_FAILURE);
 252         }
 253     }
 254
 255 #ifdef THREADED_RTS
 256     {
 257         int err;
 258         if ((err = PAPI_thread_init(osThreadId)) < 0) {
 259             barf("PAPI_thread_init: %d",err);
 260         }
 261
 262         initMutex(&papi_counter_mutex);
 263     }
 264 #endif
 265
 266     init_countable_events();
 267
 268     papi_init_eventset(&MutatorEvents);
 269     papi_init_eventset(&GCEvents);
 270 }
 271
 272 /* Extract the value corresponding to an event */
 273 static long_long
 274 papi_counter(long_long values[],int event)
 275 {
 276   int i;
 277   for(i=0;i<n_papi_events;i++) {
 278     if(papi_events[i].event_code==event) {
 279       return values[i];
 280     }
 281   }
 282   /* Passed a wrong event? */
 283   debugBelch("Event %d is not part of event set\n",event);
 284   return 0;
 285 }
 286
 287 /* Add the events of papi_events into an event set */
 288 static void
 289 papi_add_events(int EventSet)
 290 {
 291   int i;
 292   for(i=0;i<n_papi_events;i++) {
 293     if((papi_error=PAPI_add_event(EventSet,
 294                                   papi_events[i].event_code))
 295        != PAPI_OK)
 296       debugBelch("Failed adding %s to event set with error code %d\n",
 297                  papi_events[i].event_name,papi_error);
 298   }
 299 }
 300
 301 /* We should be using elapsed cycles
 302  * to be consistent with time metric chosen in Stats.c (Elapsed time).
 303  * This is an approximation to the cycles that the program spends.
 304  * Note that the counters, in contrast, are virtual and user space.
 305  */
 306 #define PAPI_cycles PAPI_get_virt_cyc
 307
 308 void
 309 papi_start_mutator_count(void)
 310 {
 311     ACQUIRE_LOCK(&papi_counter_mutex);
 312     PAPI_CHECK( PAPI_start(MutatorEvents));
 313     start_mutator_cycles = PAPI_cycles();
 314     RELEASE_LOCK(&papi_counter_mutex);
 315 }
 316
 317 void
 318 papi_stop_mutator_count(void)
 319 {
 320     ACQUIRE_LOCK(&papi_counter_mutex);
 321     mutator_cycles += PAPI_cycles() - start_mutator_cycles;
 322     PAPI_CHECK( PAPI_accum(MutatorEvents,MutatorCounters));
 323     PAPI_CHECK( PAPI_stop(MutatorEvents,NULL));
 324     RELEASE_LOCK(&papi_counter_mutex);
 325 }
 326
 327 void
 328 papi_start_gc_count(void)
 329 {
 330     ACQUIRE_LOCK(&papi_counter_mutex);
 331     PAPI_CHECK( PAPI_start(GCEvents));
 332     start_gc_cycles = PAPI_cycles();
 333     RELEASE_LOCK(&papi_counter_mutex);
 334 }
 335
 336 void
 337 papi_stop_gc_count(void)
 338 {
 339     ACQUIRE_LOCK(&papi_counter_mutex);
 340     PAPI_CHECK( PAPI_accum(GCEvents,GCCounters));
 341     PAPI_CHECK( PAPI_stop(GCEvents,NULL));
 342     gc_cycles += PAPI_cycles() - start_gc_cycles;
 343     RELEASE_LOCK(&papi_counter_mutex);
 344 }
 345
 346
 347 void
 348 papi_thread_start_gc_count(int event_set)
 349 {
 350     ACQUIRE_LOCK(&papi_counter_mutex);
 351     PAPI_CHECK( PAPI_start(event_set));
 352     RELEASE_LOCK(&papi_counter_mutex);
 353 }
 354
 355 void
 356 papi_thread_stop_gc_count(int event_set)
 357 {
 358     ACQUIRE_LOCK(&papi_counter_mutex);
 359     PAPI_CHECK( PAPI_accum(event_set,GCCounters));
 360     PAPI_CHECK( PAPI_stop(event_set,NULL));
 361     RELEASE_LOCK(&papi_counter_mutex);
 362 }
 363
 364 #endif /* USE_PAPI */