Addition of PAPI to RTS

This patch still requires the addition of the USE_PAPI
define to compile with PAPI. Also, programs must be
compiled and linked with the appropriate library flags
for papi.

rts/Papi.c

+
+
+#include "Papi.h"
+#include "Rts.h"
+#include "RtsUtils.h"
+#include "Stats.h"
+
+
+/* These constants specify which events to keep track of.
+ * Probably it is better to count one set of events at a time.
+ * The reason is that processors have limited counters and
+ * multiplexing is not enabled (yet).
+ */
+#define PAPI_COUNT_BRANCHES 0
+/* The one below is Opteron specific.
+ */
+#define PAPI_COUNT_STALLS 0
+#define PAPI_COUNT_DCACHE1_MISSES 1
+#define PAPI_COUNT_DCACHE2_MISSES 0
+
+struct _papi_events {
+  int event_code;
+  char * event_name;
+};
+
+#define PAPI_ADD_EVENT(EVENT) { EVENT, #EVENT }
+
+/* Beware, these counters are Opteron specific */
+#define FR_BR 0x40000040
+#define FR_BR_MIS 0x40000041
+#define FR_BR_MISCOMPARE 0x40000048
+#define DC_ACCESS 0x40000019
+#define DC_MISS 0x4000001a
+#define FR_DISPATCH_STALLS_BR 0x40000055
+#define FR_DISPATCH_STALLS_FULL_LS 0x4000005b
+
+/* 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);				\
+  }
+
+/* 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))
+
+/* Number of counted events, computed from size of papi_events */
+#define N_PAPI_EVENTS ((int)(sizeof(papi_events)/sizeof(struct _papi_events)))
+
+/* This is bad, it should be in a header */
+#define BIG_STRING_LEN 512
+
+/* While PAPI reporting is going on this flag is on */
+int papi_is_reporting;
+
+/* Event sets and counter arrays for GC and mutator */
+
+int MutatorEvents = PAPI_NULL;
+int GCEvents = PAPI_NULL;
+
+int papi_error;
+
+
+/* If you want to add events to count, extend the
+ * papi_events array and the papi_report function.
+ */
+
+/* Events counted during GC and Mutator execution */
+/* There's a trailing comma, do all C compilers accept that? */
+static struct _papi_events papi_events[] = {
+  PAPI_ADD_EVENT(PAPI_TOT_CYC),
+#if PAPI_COUNT_BRANCHES
+  PAPI_ADD_EVENT(FR_BR),
+  PAPI_ADD_EVENT(FR_BR_MIS),
+  /* Docs are wrong? Opteron does not count indirect branch misses apparently */
+  PAPI_ADD_EVENT(FR_BR_MISCOMPARE),
+#endif
+#if PAPI_COUNT_STALLS
+  PAPI_ADD_EVENT(FR_DISPATCH_STALLS_BR),
+  PAPI_ADD_EVENT(FR_DISPATCH_STALLS_FULL_LS),
+#endif
+#if PAPI_COUNT_DCACHE1_MISSES
+  PAPI_ADD_EVENT(PAPI_L1_DCA),
+  PAPI_ADD_EVENT(PAPI_L1_DCM),
+#endif
+#if PAPI_COUNT_DCACHE2_MISSES
+  PAPI_ADD_EVENT(PAPI_L2_DCA),
+  PAPI_ADD_EVENT(PAPI_L2_DCM),
+#endif
+};
+
+long_long MutatorCounters[N_PAPI_EVENTS];
+long_long GCCounters[N_PAPI_EVENTS];
+
+
+/* Extract the value corresponding to an event */
+long_long
+papi_counter(long_long values[],int event)
+{
+  int i;
+  for(i=0;i<N_PAPI_EVENTS;i++) {
+    if(papi_events[i].event_code==event) {
+      return values[i];
+    }
+  }
+  /* Passed a wrong event? */
+  debugBelch("Event %d is not part of event set\n",event);
+  return 0;
+}
+
+/* This function reports counters for GC and mutator */
+void
+papi_report(long_long PapiCounters[])
+{
+    char temp[BIG_STRING_LEN];
+
+    /* I need to improve formatting aesthetics */
+    PAPI_REPORT(PapiCounters,PAPI_TOT_CYC);
+#if PAPI_COUNT_BRANCHES
+    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);
+#endif
+#if PAPI_COUNT_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);
+#endif
+#if PAPI_COUNT_DCACHE1_MISSES
+    PAPI_REPORT(PapiCounters,PAPI_L1_DCA);
+    PAPI_REPORT(PapiCounters,PAPI_L1_DCM);
+    PAPI_REPORT_PCT(PapiCounters,PAPI_L1_DCM,PAPI_L1_DCA);
+#endif
+#if PAPI_COUNT_DCACHE2_MISSES
+    PAPI_REPORT(PapiCounters,PAPI_L2_DCA);
+    PAPI_REPORT(PapiCounters,PAPI_L2_DCM);
+    PAPI_REPORT_PCT(PapiCounters,PAPI_L2_DCM,PAPI_L2_DCA);
+#endif
+}
+
+/* Add the events of papi_events into an event set */
+void
+papi_add_events(int EventSet)
+{
+  int i;
+  for(i=0;i<N_PAPI_EVENTS;i++) {
+    if((papi_error=PAPI_add_event(EventSet,
+				  papi_events[i].event_code))
+       != PAPI_OK)
+      debugBelch("Failed adding %s to event set with error code %d\n",
+		 papi_events[i].event_name,papi_error);
+  }
+}
+
+void
+papi_init_eventsets(void)
+{
+
+    /* One event set for the mutator and another for the GC */
+    PAPI_CHECK( PAPI_create_eventset(&MutatorEvents));
+    PAPI_CHECK( PAPI_create_eventset(&GCEvents));
+
+    /* Both sets contain the same events */
+    papi_add_events(MutatorEvents);
+    papi_add_events(GCEvents);
+
+}
+
+void
+papi_start_mutator_count(void)
+{
+    PAPI_CHECK( PAPI_start(MutatorEvents));
+}
+
+void
+papi_stop_mutator_count(void)
+{
+    PAPI_CHECK( PAPI_accum(MutatorEvents,MutatorCounters));
+    PAPI_CHECK( PAPI_stop(MutatorEvents,NULL));
+}
+
+void
+papi_start_gc_count(void)
+{
+      PAPI_CHECK( PAPI_start(GCEvents));
+}
+
+void
+papi_stop_gc_count(void)
+{
+      PAPI_CHECK( PAPI_accum(GCEvents,GCCounters));
+      PAPI_CHECK( PAPI_stop(GCEvents,NULL));
+}

rts/Papi.h

+
+
+#include <papi.h>
+
+
+
+#define PAPI_CHECK(CALL) \
+  if((papi_error=(CALL)) != PAPI_OK) { \
+   debugBelch("PAPI function failed in module %s at line %d with error code %d\n", \
+	      __FILE__,__LINE__,papi_error);				\
+  }
+
+/* Check the error value of a PAPI call, reporting an error, if needed */
+extern int papi_error;
+
+/* While PAPI reporting is going on this flag is on */
+extern int papi_is_reporting;
+
+/* Event sets and counter arrays for GC and mutator */
+
+extern int MutatorEvents;
+extern int GCEvents;
+
+extern long_long MutatorCounters[];
+extern long_long GCCounters[];
+
+long_long papi_counter(long_long values[],int event);
+void papi_report(long_long PapiCounters[]);
+void papi_add_events(int EventSet);
+
+void papi_init_eventsets(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);
+
+

rts/RtsStartup.c

@@ -6,7 +6,9 @@
  *
  * ---------------------------------------------------------------------------*/
 
-#include "PosixSource.h"
+// PAPI uses caddr_t, which is not POSIX
+// #include "PosixSource.h"
+
 #include "Rts.h"
 #include "RtsAPI.h"
 #include "RtsUtils.h"
@@ -67,6 +69,10 @@
 #include <signal.h>
 #endif
 
+#if USE_PAPI
+#include "Papi.h"
+#endif
+
 // Count of how many outstanding hs_init()s there have been.
 static int hs_init_count = 0;
 
@@ -152,7 +158,23 @@ hs_init(int *argc, char **argv[])
     argv++; argc--;
 #endif
 
+    /* Initialise the performance tracking library */
+#ifdef USE_PAPI
+    /* Must fix to abort gracefully */
+    if(PAPI_library_init(PAPI_VER_CURRENT) != PAPI_VER_CURRENT)
+      exit(1);
+#ifdef THREADED_RTS
+    {
+	int err;
+	if ((err = PAPI_thread_init(osThreadId)) < 0) {
+	    barf("PAPI_thread_init: %d",err);
+	}
+    }
+#endif
+#endif
+
     /* Set the RTS flags to default values. */
+
     initRtsFlagsDefaults();
 
     /* Call the user hook to reset defaults, if present */

rts/Stats.c

@@ -17,6 +17,10 @@
 #include "Profiling.h"
 #include "GetTime.h"
 
+#if USE_PAPI
+#include "Papi.h"
+#endif
+
 /* huh? */
 #define BIG_STRING_LEN              512
 
@@ -64,9 +68,6 @@ static lnat GC_start_faults = 0, GC_end_faults = 0;
 
 static Ticks *GC_coll_times;
 
-static void statsPrintf( char *s, ... ) 
-    GNUC3_ATTRIBUTE(format (printf, 1, 2));
-
 static void statsFlush( void );
 static void statsClose( void );
 
@@ -170,6 +171,18 @@ stat_endInit(void)
     } else {
 	InitElapsedTime = elapsed - ElapsedTimeStart;
     }
+#if USE_PAPI
+    papi_init_eventsets();
+
+    /* We start counting events for the mutator
+     * when garbage collection starts
+     * we switch to the GC event set. */
+    papi_start_mutator_count();
+
+    /* This flag is needed to avoid counting the last GC */
+    papi_is_reporting = 1;
+
+#endif
 }
 
 /* -----------------------------------------------------------------------------
@@ -192,6 +205,16 @@ stat_startExit(void)
 
     MutUserTime = user - GC_tot_time - PROF_VAL(RP_tot_time + HC_tot_time) - InitUserTime;
     if (MutUserTime < 0) { MutUserTime = 0; }
+
+#if USE_PAPI
+    /* We stop counting mutator events
+     * GC events are not being counted at this point */
+    papi_stop_mutator_count();
+
+    /* This flag is needed, because GC is run once more after this function */
+    papi_is_reporting = 0;
+
+#endif
 }
 
 void
@@ -249,6 +272,15 @@ stat_startGC(void)
 	    GC_start_faults = getPageFaults();
 	}
     }
+
+#if USE_PAPI
+    if(papi_is_reporting) {
+      /* Switch to counting GC events */
+      papi_stop_mutator_count();
+      papi_start_gc_count();
+    }
+#endif
+
 }
 
 /* -----------------------------------------------------------------------------
@@ -316,6 +348,14 @@ stat_endGC (lnat alloc, lnat live, lnat copied,
 	debugBelch("\b\b\b  \b\b\b");
 	rub_bell = 0;
     }
+
+#if USE_PAPI
+    if(papi_is_reporting) {
+      /* Switch to counting mutator events */
+      papi_stop_gc_count();
+      papi_start_mutator_count();
+    }
+#endif
 }
 
 /* -----------------------------------------------------------------------------
@@ -517,6 +557,16 @@ stat_exit(int alloc)
 		    TICK_TO_DBL(time - GC_tot_time - 
 				PROF_VAL(RP_tot_time + HC_tot_time) - InitUserTime) * 100 
 		    / TICK_TO_DBL(etime));
+#if USE_PAPI
+	    /* PAPI reporting, should put somewhere else?
+	     * Note that the cycles are counted _after_ the initialization of the RTS -- AR */
+
+	    statsPrintf("  -- CPU Mutator counters --\n");
+	    papi_report(MutatorCounters);
+
+	    statsPrintf("\n  -- CPU GC counters --\n");
+	    papi_report(GCCounters);
+#endif
 	}
 
 	if (RtsFlags.GcFlags.giveStats == ONELINE_GC_STATS) {
@@ -606,7 +656,7 @@ extern HsInt64 getAllocations( void )
    Dumping stuff in the stats file, or via the debug message interface
    -------------------------------------------------------------------------- */
 
-static void
+void
 statsPrintf( char *s, ... )
 {
     FILE *sf = RtsFlags.GcFlags.statsFile;

rts/Stats.h

@@ -54,4 +54,9 @@ HsInt64   getAllocations( void );
 Ticks stat_getElapsedGCTime(void);
 Ticks stat_getElapsedTime(void);
 
+/* Only exported for Papi.c */
+void statsPrintf( char *s, ... ) 
+    GNUC3_ATTRIBUTE(format (printf, 1, 2));
+
+
 #endif /* STATS_H */

rts/package.conf.in

@@ -53,6 +53,9 @@ extra-libraries:		"m"		/* for ldexp() */
                               ,"mingwex"
 # endif
 #endif
+#if USE_PAPI
+			     , "papi"
+#endif
 
 #ifdef INSTALLING
 include-dirs:		INCLUDE_DIR

rts/posix/GetTime.c

@@ -32,6 +32,10 @@
 # include <sys/times.h>
 #endif
 
+#ifdef USE_PAPI
+# include <papi.h>
+#endif
+
 #if ! ((defined(HAVE_GETRUSAGE) && !irix_HOST_OS) || defined(HAVE_TIMES))
 #error No implementation for getProcessCPUTime() available.
 #endif
@@ -68,9 +72,17 @@ void getProcessTimes(Ticks *user, Ticks *elapsed)
 
 Ticks getProcessCPUTime(void)
 {
+#if !defined(THREADED_RTS) && USE_PAPI
+    long long usec;
+    if ((usec = PAPI_get_virt_usec()) < 0) {
+	barf("PAPI_get_virt_usec: %lld", usec);
+    }
+    return ((usec * TICKS_PER_SECOND) / 1000000);
+#else
     Ticks user, elapsed;
     getProcessTimes(&user,&elapsed);
     return user;
+#endif
 }
 
 Ticks getProcessElapsedTime(void)
@@ -115,7 +127,14 @@ void getProcessTimes(Ticks *user, Ticks *elapsed)
 
 Ticks getThreadCPUTime(void)
 {
-#if defined(HAVE_CLOCK_GETTIME) && defined(CLOCK_THREAD_CPUTIME_ID)
+#if USE_PAPI
+    long long usec;
+    if ((usec = PAPI_get_virt_usec()) < 0) {
+	barf("PAPI_get_virt_usec: %lld", usec);
+    }
+    return ((usec * TICKS_PER_SECOND) / 1000000);
+
+#elif defined(HAVE_CLOCK_GETTIME) && defined(CLOCK_THREAD_CPUTIME_ID)
     // clock_gettime() gives us per-thread CPU time.  It isn't
     // reliable on Linux, but it's the best we have.
     struct timespec ts;