# HG changeset patch
# User Merten Sach <msach@mailbox.tu-berlin.de>
# Date 1324398087 -3600
# Node ID c3561dbac1dcd2156032a1cecca7ac65a25055ab
# Parent  85f55731f6cd919fef6816481b17076531f75ae0
added cache misses counter

diff -r 85f55731f6cd -r c3561dbac1dc src/Application/main.c
--- a/src/Application/main.c	Tue Dec 20 15:00:07 2011 +0100
+++ b/src/Application/main.c	Tue Dec 20 17:21:27 2011 +0100
@@ -60,6 +60,21 @@
 //#error "I don't know how to measure time on your platform"
 #endif
 
+//======================== Globals =========================
+char __ProgrammName[] = "overhead_test";
+char __DataSet[255];
+
+int outer_iters, inner_iters, num_threads;
+size_t chunk_size = 0;
+
+int cycles_counter_main_fd;
+int misses_counter_fd;
+
+uint64_t cache_misses;
+
+int cycles_counter_fd[NUM_CORES];
+struct perf_event_attr* hw_event;
+
 //======================== Defines =========================
 typedef struct perfData measurement_t;
 struct perfData{
@@ -95,7 +110,6 @@
    barr->cond    = VPThread__make_cond(barr->mutex, animatingPr);
  }
 
-int cycles_counter_main_fd;
 void inline barrier_wait(barrier *barr, VirtProcr *animatingPr)
  { int i;
 
@@ -133,24 +147,14 @@
      struct WorkerParams_t data;
      char padding[CACHELINE_SIZE];
  } WorkerParams __align_to_cacheline__;
+ 
+WorkerParams *workerParamsArray;
 
 typedef struct
  { measurement_t *startExeCycles;
    measurement_t *endExeCycles;
  } BenchParams __align_to_cacheline__;
 
-//======================== Globals =========================
-char __ProgrammName[] = "overhead_test";
-char __DataSet[255];
-
-int outer_iters, inner_iters, num_threads;
-size_t chunk_size = 0;
-
-int cycles_counter_fd[NUM_CORES];
-struct perf_event_attr* hw_event;
-
-WorkerParams *workerParamsArray;
-
 //======================== App Code =========================
 /*
  p* Workload
@@ -166,6 +170,16 @@
        cycles = 0;                                      \
    }                                                    \
 } while (0) //macro magic for scoping
+ 
+#define saveMisses(misses) do{     \
+   int nread;                                           \
+                                                        \
+   nread = read(misses_counter_fd,&(misses),sizeof(misses));    \
+   if(nread<0){                                         \
+       perror("Error reading misses counter");          \
+       misses = 0;                                      \
+   }                                                    \
+} while (0) //macro magic for scoping
 
 
 double
@@ -264,7 +278,9 @@
     { 
        workerParamsArray[i].data.barrier = &barr;
     }
-     
+    
+   uint64_t cache_misses_at_start, cache_misses_at_end;
+   saveMisses(cache_misses_at_start);
    //save cycles before execution of threads, to get total exe cycles
    int nread = read(cycles_counter_main_fd, &(params->startExeCycles->cycles),
                 sizeof(params->startExeCycles->cycles));
@@ -279,8 +295,8 @@
   
    //endBarrierCycles read in barrier_wait()!  Merten, email me if want to chg
    params->endExeCycles->cycles = barr.endBarrierCycles.cycles;
-   
-
+   saveMisses(cache_misses_at_end);
+   cache_misses = cache_misses_at_end-cache_misses_at_start;
 /*
    uint64_t overallWorkCycles = 0;
    for(i=0; i<num_threads; i++){ 
@@ -420,6 +436,28 @@
       perror("Failed to open cycles counter");
     }
    
+   //Set up counters to count cache misses
+    hw_event->type = PERF_TYPE_HARDWARE;
+    hw_event->config = PERF_COUNT_HW_CACHE_MISSES; //misses
+    
+   retries = 0;
+   do
+    { retries += 1;
+      misses_counter_fd = 
+       syscall(__NR_perf_event_open, hw_event,
+               0,//pid_t: 0 is "pid of calling process" 
+               -1,//int: cpu, -1 means accumulate from all cores
+               -1,//int: group_fd, -1 is "leader" == independent
+               0//unsigned long: flags
+              );
+    }
+   while(misses_counter_fd<0 && retries < 100);
+   if(retries >= 100)
+    {
+      fprintf(stderr,"in main ");
+      perror("Failed to misses counter");
+    }
+   
    measurement_t startExeCycles, endExeCycles;
    BenchParams *benchParams;
    
@@ -453,6 +491,7 @@
    uint64 totalOverhead = totalExeCycles - totalWorkCyclesAcrossCores;
    int32  numSyncs = outer_iters * num_threads * 2;
    printf("Total Execution Cycles: %lu\n", totalExeCycles);
+   printf("Total number of cache misses: %lu\n", cache_misses);
    printf("Sum across threads of work cycles: %lu\n", totalWorkCyclesAcrossCores);
    printf("Sum across threads of bad work cycles: %lu\n", totalBadCyclesAcrossCores);
 //   printf("Sum across threads of Bad Sync cycles: %lu\n", totalBadSyncCyclesAcrossCores);