improvements to PAPI support

- major (multithreaded) GC is measured separately from minor GC
- events to measure can now be specified on the command line, e.g
     prog +RTS -a+PAPI_TOT_CYC

6 files changed, 121 insertions, 91 deletions

diff --git a/rts/Papi.c b/rts/Papi.c
index d442718fa9..d95e26c8e7 100644
--- a/rts/Papi.c
+++ b/rts/Papi.c
@@ -30,7 +30,7 @@ static Mutex papi_counter_mutex;
 
 struct _papi_events {
   int event_code;
-  char * event_name;
+  const char * event_name;
 };
 
 /* Beware, these counters are Opteron specific
@@ -75,45 +75,54 @@ int papi_error;
 /* Arbitrary, to avoid using malloc */
 #define MAX_PAPI_EVENTS 10
 
-int n_papi_events = 0;
+static nat n_papi_events = 0;
 
 
 /* Events counted during GC and Mutator execution */
 /* There's a trailing comma, do all C compilers accept that? */
 static struct _papi_events papi_events[MAX_PAPI_EVENTS];
 long_long MutatorCounters[MAX_PAPI_EVENTS];
-long_long GCCounters[MAX_PAPI_EVENTS];
+long_long GC0Counters[MAX_PAPI_EVENTS];
+long_long GC1Counters[MAX_PAPI_EVENTS];
 
 long_long start_mutator_cycles;
-long_long start_gc_cycles;
 long_long mutator_cycles;
-long_long gc_cycles;
+long_long start_gc_cycles;
+long_long gc0_cycles;
+long_long gc1_cycles;
 
 
 
 static long_long papi_counter(long_long values[],int event);
 static void papi_add_events(int EventSet);
 
+static nat max_hardware_counters = 2;
+
 /* If you want to add events to count, extend the
  * init_countable_events and the papi_report function.
  * Be aware that your processor can count a limited number
  * of events simultaneously, you can turn on multiplexing
  * to increase that number, though.
  */
+static void papi_add_event(const char *name, int code)
+{
+    if (n_papi_events >= max_hardware_counters) {
+        errorBelch("too many PAPI events for this CPU (max: %d)", 
+                   max_hardware_counters);
+        stg_exit(EXIT_FAILURE);
+    }
+    papi_events[n_papi_events].event_code = code;
+    papi_events[n_papi_events].event_name = name;
+    n_papi_events++;
+}    
+
 static void
 init_countable_events(void) 
 {
-#define PAPI_ADD_EVENT(EVENT)                           \
-    {                                                   \
-        if (n_papi_events >= MAX_PAPI_EVENTS) {         \
-           barf("too many PAPI events");                \
-        }                                               \
-	papi_events[n_papi_events].event_code = EVENT;	\
-	papi_events[n_papi_events].event_name = #EVENT; \
-	n_papi_events++;				\
-    }
+    max_hardware_counters = PAPI_num_counters();
+
+#define PAPI_ADD_EVENT(EVENT) papi_add_event(#EVENT,EVENT)
 
-    PAPI_ADD_EVENT(PAPI_TOT_INS);
     if (RtsFlags.PapiFlags.eventType==PAPI_FLAG_BRANCH) {
 	PAPI_ADD_EVENT(FR_BR);
 	PAPI_ADD_EVENT(FR_BR_MIS);
@@ -133,8 +142,35 @@ init_countable_events(void)
 	PAPI_ADD_EVENT(DC_L2_REFILL_MOES);
 	PAPI_ADD_EVENT(DC_SYS_REFILL_MOES);
 	PAPI_ADD_EVENT(FR_BR_MIS);
+    } else if (RtsFlags.PapiFlags.eventType==PAPI_USER_EVENTS) {
+        nat i;
+        char *name;
+        int code;
+        for (i = 0; i < RtsFlags.PapiFlags.numUserEvents; i++) {
+            name = RtsFlags.PapiFlags.userEvents[i];
+            PAPI_CHECK(PAPI_event_name_to_code(name, &code))
+            papi_add_event(name, code);
+        }
     } else {
-	PAPI_ADD_EVENT(PAPI_STL_ICY);
+	// PAPI_ADD_EVENT(PAPI_L1_DCA); // L1 data cache accesses
+	// PAPI_ADD_EVENT(PAPI_L1_ICR); // L1 instruction cache reads
+	// PAPI_ADD_EVENT(PAPI_L1_ICM); // L1 instruction cache misses
+	// PAPI_ADD_EVENT(PAPI_L1_STM); // L1 store misses
+	// PAPI_ADD_EVENT(PAPI_L1_DCM); // L1 data cache misses
+	// PAPI_ADD_EVENT(PAPI_L1_LDM); // L1 load misses
+	// PAPI_ADD_EVENT(PAPI_L2_TCM); // L2 cache misses
+	// PAPI_ADD_EVENT(PAPI_L2_STM); // L2 store misses
+	// PAPI_ADD_EVENT(PAPI_L2_DCW); // L2 data cache writes
+	// PAPI_ADD_EVENT(PAPI_L2_DCR); // L2 data cache reads
+	// PAPI_ADD_EVENT(PAPI_L2_TCW); // L2 cache writes
+	// PAPI_ADD_EVENT(PAPI_L2_TCR); // L2 cache reads
+	// PAPI_ADD_EVENT(PAPI_CA_CLN); // exclusive access to clean cache line
+	// PAPI_ADD_EVENT(PAPI_TLB_DM); // TLB misses
+        PAPI_ADD_EVENT(PAPI_TOT_INS); // Total instructions
+        PAPI_ADD_EVENT(PAPI_TOT_CYC); // Total instructions
+	// PAPI_ADD_EVENT(PAPI_CA_SHR); // exclusive access to shared cache line
+	// PAPI_ADD_EVENT(PAPI_RES_STL); // Cycles stalled on any resource
+        
     }
 
     // We might also consider:
@@ -143,90 +179,58 @@ init_countable_events(void)
 };
 
 
-static char temp[BIG_STRING_LEN];
-
-static void
-papi_mut_cycles(void)
-{
-    ullong_format_string(mutator_cycles,temp,rtsTrue/*commas*/); 
-    statsPrintf("  (MUT_CYCLES)  : %s\n",temp);
-}
-
 static void
-papi_gc_cycles(void)
+papi_report_event(const char *name,  ullong value)
 {
-    ullong_format_string(gc_cycles,temp,rtsTrue/*commas*/); 
-    statsPrintf("  (GC_CYCLES)  : %s\n",temp);
+    static char temp[BIG_STRING_LEN];
+    ullong_format_string(value,temp,rtsTrue/*commas*/); 
+    statsPrintf("  %15s  %15s\n", name, temp);
 }
 
 /* This function reports counters for GC and mutator */
 static void
-papi_report(long_long PapiCounters[])
+papi_report(long_long counters[])
 {
-
-/* Report the value of a counter */
-#define PAPI_REPORT(EVENTSET,EVENT) \
-  { \
-    ullong_format_string(papi_counter(EVENTSET,EVENT),temp,rtsTrue/*commas*/); \
-    statsPrintf("  (" #EVENT ")  : %s\n",temp);				\
-  }
+    nat i;
 
 /* Report the value of a counter as a percentage of another counter */
 #define PAPI_REPORT_PCT(EVENTSET,EVENT,EVENTTOT) \
-  statsPrintf("  (" #EVENT ") %% of (" #EVENTTOT ") : %.1f%%\n", \
-	      papi_counter(EVENTSET,EVENT)*100.0/papi_counter(EVENTSET,EVENTTOT))
+    statsPrintf("   " #EVENT " %% of " #EVENTTOT " : %.1f%%\n",      \
+         papi_counter(EVENTSET,EVENT)*100.0/papi_counter(EVENTSET,EVENTTOT))
 
-  /* I need to improve formatting aesthetics */
-    PAPI_REPORT(PapiCounters,PAPI_TOT_INS);
-
-    if (RtsFlags.PapiFlags.eventType==PAPI_FLAG_BRANCH) {
-	PAPI_REPORT(PapiCounters,FR_BR);
-	PAPI_REPORT(PapiCounters,FR_BR_MIS);
-	PAPI_REPORT_PCT(PapiCounters,FR_BR_MIS,FR_BR);
-	PAPI_REPORT_PCT(PapiCounters,FR_BR_MISCOMPARE,FR_BR);
+    for (i = 0; i < n_papi_events; i++)
+    {
+        papi_report_event(papi_events[i].event_name, counters[i]);
     }
 
-    else if(RtsFlags.PapiFlags.eventType==PAPI_FLAG_STALLS) {
-	PAPI_REPORT(PapiCounters,FR_DISPATCH_STALLS);
-	PAPI_REPORT(PapiCounters,FR_DISPATCH_STALLS_BR);
-	//PAPI_REPORT_PCT(PapiCounters,FR_DISPATCH_STALLS_BR,PAPI_TOT_CYC);
-	PAPI_REPORT(PapiCounters,FR_DISPATCH_STALLS_FULL_LS);
-	//PAPI_REPORT_PCT(PapiCounters,FR_DISPATCH_STALLS_FULL_LS,PAPI_TOT_CYC);
+    if (RtsFlags.PapiFlags.eventType==PAPI_FLAG_BRANCH) {
+	PAPI_REPORT_PCT(counters,FR_BR_MIS,FR_BR);
+	PAPI_REPORT_PCT(counters,FR_BR_MISCOMPARE,FR_BR);
     }
 
     else if (RtsFlags.PapiFlags.eventType==PAPI_FLAG_CACHE_L1) {
-	PAPI_REPORT(PapiCounters,PAPI_L1_DCA);
-	PAPI_REPORT(PapiCounters,PAPI_L1_DCM);
-	PAPI_REPORT_PCT(PapiCounters,PAPI_L1_DCM,PAPI_L1_DCA);
+	PAPI_REPORT_PCT(counters,PAPI_L1_DCM,PAPI_L1_DCA);
     }
 
     else if (RtsFlags.PapiFlags.eventType==PAPI_FLAG_CACHE_L2) {
-	PAPI_REPORT(PapiCounters,PAPI_L2_DCA);
-	PAPI_REPORT(PapiCounters,PAPI_L2_DCM);
-	PAPI_REPORT_PCT(PapiCounters,PAPI_L2_DCM,PAPI_L2_DCA);
+	PAPI_REPORT_PCT(counters,PAPI_L2_DCM,PAPI_L2_DCA);
     }
-
-    else if (RtsFlags.PapiFlags.eventType==PAPI_FLAG_CB_EVENTS) {
-	PAPI_REPORT(PapiCounters,DC_L2_REFILL_MOES);
-	PAPI_REPORT(PapiCounters,DC_SYS_REFILL_MOES);
-	PAPI_REPORT(PapiCounters,FR_BR_MIS);
-    }
-
-    else {
-	PAPI_REPORT(PapiCounters,PAPI_STL_ICY);
-    }        
 }
 
 void
 papi_stats_report (void)
 {
-    statsPrintf("  -- CPU Mutator counters --\n");
-    papi_mut_cycles();
+    statsPrintf("  Mutator CPU counters\n");
+    papi_report_event("CYCLES", mutator_cycles);
     papi_report(MutatorCounters);
     
-    statsPrintf("\n  -- CPU GC counters --\n");
-    papi_gc_cycles();
-    papi_report(GCCounters);
+    statsPrintf("\n  GC(0) CPU counters\n");
+    papi_report_event("CYCLES", gc0_cycles);
+    papi_report(GC0Counters);
+
+    statsPrintf("\n  GC(1) CPU counters\n");
+    papi_report_event("CYCLES", gc1_cycles);
+    papi_report(GC1Counters);
 }
     
 void
@@ -273,7 +277,7 @@ papi_init (void)
 static long_long
 papi_counter(long_long values[],int event)
 {
-  int i;
+  nat i;
   for(i=0;i<n_papi_events;i++) {
     if(papi_events[i].event_code==event) {
       return values[i];
@@ -288,7 +292,7 @@ papi_counter(long_long values[],int event)
 static void
 papi_add_events(int EventSet)
 {
-  int i;
+  nat i;
   for(i=0;i<n_papi_events;i++) {
     if((papi_error=PAPI_add_event(EventSet,
 				  papi_events[i].event_code))
@@ -334,18 +338,29 @@ papi_start_gc_count(void)
 }
 
 void
-papi_stop_gc_count(void)
+papi_stop_gc0_count(void)
+{
+    ACQUIRE_LOCK(&papi_counter_mutex);
+    PAPI_CHECK( PAPI_accum(GCEvents,GC0Counters));
+    PAPI_CHECK( PAPI_stop(GCEvents,NULL));
+    gc0_cycles += PAPI_cycles() - start_gc_cycles;
+    RELEASE_LOCK(&papi_counter_mutex);
+}
+
+
+void
+papi_stop_gc1_count(void)
 {
     ACQUIRE_LOCK(&papi_counter_mutex);
-    PAPI_CHECK( PAPI_accum(GCEvents,GCCounters));
+    PAPI_CHECK( PAPI_accum(GCEvents,GC1Counters));
     PAPI_CHECK( PAPI_stop(GCEvents,NULL));
-    gc_cycles += PAPI_cycles() - start_gc_cycles;
+    gc1_cycles += PAPI_cycles() - start_gc_cycles;
     RELEASE_LOCK(&papi_counter_mutex);
 }
 
 
 void
-papi_thread_start_gc_count(int event_set)
+papi_thread_start_gc1_count(int event_set)
 {
     ACQUIRE_LOCK(&papi_counter_mutex);
     PAPI_CHECK( PAPI_start(event_set));
@@ -353,10 +368,10 @@ papi_thread_start_gc_count(int event_set)
 }
 
 void
-papi_thread_stop_gc_count(int event_set)
+papi_thread_stop_gc1_count(int event_set)
 {
     ACQUIRE_LOCK(&papi_counter_mutex);
-    PAPI_CHECK( PAPI_accum(event_set,GCCounters));
+    PAPI_CHECK( PAPI_accum(event_set,GC1Counters));
     PAPI_CHECK( PAPI_stop(event_set,NULL));
     RELEASE_LOCK(&papi_counter_mutex);
 }
diff --git a/rts/Papi.h b/rts/Papi.h
index eaabdf5495..fc92a913d5 100644
--- a/rts/Papi.h
+++ b/rts/Papi.h
@@ -16,10 +16,12 @@ void papi_init_eventset(int * event_set);
 void papi_init(void);
 void papi_start_mutator_count(void);
 void papi_stop_mutator_count(void);
+
 void papi_start_gc_count(void);
-void papi_stop_gc_count(void);
+void papi_stop_gc0_count(void);
+void papi_stop_gc1_count(void);
 
 // for multithreaded GC, each sub-thread uses these functions to count
 // events and aggregate them into the main GC counters.
-void papi_thread_start_gc_count(int event_set);
-void papi_thread_stop_gc_count(int event_set);
+void papi_thread_start_gc1_count(int event_set);
+void papi_thread_stop_gc1_count(int event_set);
diff --git a/rts/RtsFlags.c b/rts/RtsFlags.c
index 69064ea743..439dfd7897 100644
--- a/rts/RtsFlags.c
+++ b/rts/RtsFlags.c
@@ -321,6 +321,7 @@ void initRtsFlagsDefaults(void)
 #ifdef USE_PAPI
     /* By default no special measurements taken */
     RtsFlags.PapiFlags.eventType        = 0;
+    RtsFlags.PapiFlags.numUserEvents    = 0;
 #endif
 }
 
@@ -721,6 +722,14 @@ error = rtsTrue;
 		case 'e':
 		  RtsFlags.PapiFlags.eventType = PAPI_FLAG_CB_EVENTS;
 		  break;
+                case '+':
+                  if (RtsFlags.PapiFlags.numUserEvents >= MAX_PAPI_USER_EVENTS) {
+                      errorBelch("maximum number of PAPI events reached");
+                      stg_exit(EXIT_FAILURE);
+                  }
+                  RtsFlags.PapiFlags.eventType = PAPI_USER_EVENTS;
+                  RtsFlags.PapiFlags.userEvents[RtsFlags.PapiFlags.numUserEvents++] = rts_argv[arg] + 3;
+                  break;
 		default:
 		  bad_option( rts_argv[arg] );
 		}
diff --git a/rts/RtsStartup.c b/rts/RtsStartup.c
index d1025a3a8b..070275b534 100644
--- a/rts/RtsStartup.c
+++ b/rts/RtsStartup.c
@@ -167,10 +167,6 @@ hs_init(int *argc, char **argv[])
     argv++; argc--;
 #endif
 
-#ifdef USE_PAPI
-    papi_init();
-#endif
-
     /* Set the RTS flags to default values. */
 
     initRtsFlagsDefaults();
@@ -185,6 +181,10 @@ hs_init(int *argc, char **argv[])
 	setProgArgv(*argc,*argv);
     }
 
+#ifdef USE_PAPI
+    papi_init();
+#endif
+
     /* initTracing must be after setupRtsFlags() */
     initTracing();
 
diff --git a/rts/Stats.c b/rts/Stats.c
index 1127b81566..4bb9c8237e 100644
--- a/rts/Stats.c
+++ b/rts/Stats.c
@@ -347,7 +347,11 @@ stat_endGC (lnat alloc, lnat live, lnat copied, lnat gen)
 #if USE_PAPI
     if(papi_is_reporting) {
       /* Switch to counting mutator events */
-      papi_stop_gc_count();
+      if (gen == 0) {
+          papi_stop_gc0_count();
+      } else {
+          papi_stop_gc1_count();
+      }
       papi_start_mutator_count();
     }
 #endif
diff --git a/rts/sm/GC.c b/rts/sm/GC.c
index 0dac20c7d8..a07086e65f 100644
--- a/rts/sm/GC.c
+++ b/rts/sm/GC.c
@@ -1024,14 +1024,14 @@ gc_thread_mainloop (void)
         if (gct->papi_events == -1) {
             papi_init_eventset(&gct->papi_events);
         }
-        papi_thread_start_gc_count(gct->papi_events);
+        papi_thread_start_gc1_count(gct->papi_events);
 #endif
 
 	gc_thread_work();
 
 #ifdef USE_PAPI
         // count events in this thread towards the GC totals
-        papi_thread_stop_gc_count(gct->papi_events);
+        papi_thread_stop_gc1_count(gct->papi_events);
 #endif
     }
 }