perf stat: Fix IPC and other formulas with -A
authorAndi Kleen <ak@linux.intel.com>
Wed, 11 Mar 2015 15:28:01 +0000 (08:28 -0700)
committerArnaldo Carvalho de Melo <acme@redhat.com>
Fri, 13 Mar 2015 10:46:10 +0000 (07:46 -0300)
perf stat didn't compute the IPC and other formulas for individual CPUs
with -A. Fix this for the easy -A case. As before, --per-core and
--per-socket do not handle it, they simply print nothing.

Signed-off-by: Andi Kleen <ak@linux.intel.com>
Link: http://lkml.kernel.org/r/1426087682-22765-2-git-send-email-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
tools/perf/builtin-stat.c

index d58e50c..c95dbda 100644 (file)
@@ -353,39 +353,40 @@ static struct perf_evsel *nth_evsel(int n)
  * more semantic information such as miss/hit ratios,
  * instruction rates, etc:
  */
-static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
+static void update_shadow_stats(struct perf_evsel *counter, u64 *count,
+                               int cpu)
 {
        if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
-               update_stats(&runtime_nsecs_stats[0], count[0]);
+               update_stats(&runtime_nsecs_stats[cpu], count[0]);
        else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
-               update_stats(&runtime_cycles_stats[0], count[0]);
+               update_stats(&runtime_cycles_stats[cpu], count[0]);
        else if (transaction_run &&
                 perf_evsel__cmp(counter, nth_evsel(T_CYCLES_IN_TX)))
-               update_stats(&runtime_cycles_in_tx_stats[0], count[0]);
+               update_stats(&runtime_cycles_in_tx_stats[cpu], count[0]);
        else if (transaction_run &&
                 perf_evsel__cmp(counter, nth_evsel(T_TRANSACTION_START)))
-               update_stats(&runtime_transaction_stats[0], count[0]);
+               update_stats(&runtime_transaction_stats[cpu], count[0]);
        else if (transaction_run &&
                 perf_evsel__cmp(counter, nth_evsel(T_ELISION_START)))
-               update_stats(&runtime_elision_stats[0], count[0]);
+               update_stats(&runtime_elision_stats[cpu], count[0]);
        else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
-               update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
+               update_stats(&runtime_stalled_cycles_front_stats[cpu], count[0]);
        else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
-               update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
+               update_stats(&runtime_stalled_cycles_back_stats[cpu], count[0]);
        else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
-               update_stats(&runtime_branches_stats[0], count[0]);
+               update_stats(&runtime_branches_stats[cpu], count[0]);
        else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
-               update_stats(&runtime_cacherefs_stats[0], count[0]);
+               update_stats(&runtime_cacherefs_stats[cpu], count[0]);
        else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
-               update_stats(&runtime_l1_dcache_stats[0], count[0]);
+               update_stats(&runtime_l1_dcache_stats[cpu], count[0]);
        else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
-               update_stats(&runtime_l1_icache_stats[0], count[0]);
+               update_stats(&runtime_l1_icache_stats[cpu], count[0]);
        else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
-               update_stats(&runtime_ll_cache_stats[0], count[0]);
+               update_stats(&runtime_ll_cache_stats[cpu], count[0]);
        else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
-               update_stats(&runtime_dtlb_cache_stats[0], count[0]);
+               update_stats(&runtime_dtlb_cache_stats[cpu], count[0]);
        else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
-               update_stats(&runtime_itlb_cache_stats[0], count[0]);
+               update_stats(&runtime_itlb_cache_stats[cpu], count[0]);
 }
 
 static void zero_per_pkg(struct perf_evsel *counter)
@@ -447,7 +448,8 @@ static int read_cb(struct perf_evsel *evsel, int cpu, int thread __maybe_unused,
                        perf_evsel__compute_deltas(evsel, cpu, count);
                perf_counts_values__scale(count, scale, NULL);
                evsel->counts->cpu[cpu] = *count;
-               update_shadow_stats(evsel, count->values);
+               if (aggr_mode == AGGR_NONE)
+                       update_shadow_stats(evsel, count->values, cpu);
                break;
        case AGGR_GLOBAL:
                aggr->val += count->val;
@@ -495,7 +497,7 @@ static int read_counter_aggr(struct perf_evsel *counter)
        /*
         * Save the full runtime - to allow normalization during printout:
         */
-       update_shadow_stats(counter, count);
+       update_shadow_stats(counter, count, 0);
 
        return 0;
 }