simpleperf: fix percentage comment in stat cmd.

Fix a bug of aggregating time_enabled,time_running introduced
in
https://android-review.googlesource.com/c/platform/system/extras/+/1220206/1/simpleperf/cmd_stat.cpp#1012.

Move some code to cmd_stat_impl.h to add tests.

Bug: 152518722
Test: run simpleperf_unit_test.
Change-Id: I50dcaf9aa83de838cc00554420d232f1a8d54601
(cherry picked from commit b3ed39b2ce227407cd828c458233e9dc222641e8)

3 files changed, 415 insertions, 176 deletions

diff --git a/simpleperf/cmd_stat.cpp b/simpleperf/cmd_stat.cpp
index 0f8f6150..c8712df7 100644
--- a/simpleperf/cmd_stat.cpp
+++ b/simpleperf/cmd_stat.cpp
@@ -32,6 +32,7 @@
 #include <android-base/strings.h>
 #include <android-base/unique_fd.h>
 
+#include "cmd_stat_impl.h"
 #include "command.h"
 #include "environment.h"
 #include "event_attr.h"
@@ -42,6 +43,8 @@
 #include "utils.h"
 #include "workload.h"
 
+using namespace simpleperf;
+
 namespace {
 
 static std::vector<std::string> default_measured_event_types{
@@ -50,95 +53,6 @@ static std::vector<std::string> default_measured_event_types{
     "task-clock",   "context-switches",        "page-faults",
 };
 
-struct CounterSum {
-  uint64_t value = 0;
-  uint64_t time_enabled = 0;
-  uint64_t time_running = 0;
-};
-
-struct ThreadInfo {
-  pid_t tid;
-  pid_t pid;
-  std::string name;
-};
-
-struct CounterSummary {
-  std::string type_name;
-  std::string modifier;
-  uint32_t group_id;
-  const ThreadInfo* thread;
-  int cpu;  // -1 represents all cpus
-  uint64_t count;
-  double scale;
-  std::string readable_count;
-  std::string comment;
-  bool auto_generated;
-
-  CounterSummary(const std::string& type_name, const std::string& modifier, uint32_t group_id,
-                 const ThreadInfo* thread, int cpu, uint64_t count, double scale,
-                 bool auto_generated, bool csv)
-      : type_name(type_name),
-        modifier(modifier),
-        group_id(group_id),
-        thread(thread),
-        cpu(cpu),
-        count(count),
-        scale(scale),
-        auto_generated(auto_generated) {
-    readable_count = ReadableCountValue(csv);
-  }
-
-  bool IsMonitoredAtTheSameTime(const CounterSummary& other) const {
-    // Two summaries are monitored at the same time if they are in the same
-    // group or are monitored all the time.
-    if (group_id == other.group_id) {
-      return true;
-    }
-    return IsMonitoredAllTheTime() && other.IsMonitoredAllTheTime();
-  }
-
-  std::string Name() const {
-    if (modifier.empty()) {
-      return type_name;
-    }
-    return type_name + ":" + modifier;
-  }
-
-  bool IsMonitoredAllTheTime() const {
-    // If an event runs all the time it is enabled (by not sharing hardware
-    // counters with other events), the scale of its summary is usually within
-    // [1, 1 + 1e-5]. By setting SCALE_ERROR_LIMIT to 1e-5, We can identify
-    // events monitored all the time in most cases while keeping the report
-    // error rate <= 1e-5.
-    constexpr double SCALE_ERROR_LIMIT = 1e-5;
-    return (fabs(scale - 1.0) < SCALE_ERROR_LIMIT);
-  }
-
- private:
-  std::string ReadableCountValue(bool csv) {
-    if (type_name == "cpu-clock" || type_name == "task-clock") {
-      // Convert nanoseconds to milliseconds.
-      double value = count / 1e6;
-      return android::base::StringPrintf("%lf(ms)", value);
-    } else {
-      // Convert big numbers to human friendly mode. For example,
-      // 1000000 will be converted to 1,000,000.
-      std::string s = android::base::StringPrintf("%" PRIu64, count);
-      if (csv) {
-        return s;
-      } else {
-        for (size_t i = s.size() - 1, j = 1; i > 0; --i, ++j) {
-          if (j == 3) {
-            s.insert(s.begin() + i, ',');
-            j = 0;
-          }
-        }
-        return s;
-      }
-    }
-  }
-};
-
 static const std::unordered_map<std::string_view, std::pair<std::string_view, std::string_view>>
     COMMON_EVENT_RATE_MAP = {
         {"cache-misses", {"cache-references", "miss rate"}},
@@ -178,8 +92,9 @@ static const std::unordered_map<std::string_view, std::pair<std::string_view, st
 
 class CounterSummaries {
  public:
-  explicit CounterSummaries(bool csv) : csv_(csv) {}
-  std::vector<CounterSummary>& Summaries() { return summaries_; }
+  explicit CounterSummaries(std::vector<CounterSummary>&& summaries, bool csv)
+      : summaries_(std::move(summaries)), csv_(csv) {}
+  const std::vector<CounterSummary>& Summaries() { return summaries_; }
 
   const CounterSummary* FindSummary(const std::string& type_name, const std::string& modifier,
                                     const ThreadInfo* thread, int cpu) {
@@ -204,8 +119,8 @@ class CounterSummaries {
         const CounterSummary* other = FindSummary(s.type_name, "k", s.thread, s.cpu);
         if (other != nullptr && other->IsMonitoredAtTheSameTime(s)) {
           if (FindSummary(s.type_name, "", s.thread, s.cpu) == nullptr) {
-            Summaries().emplace_back(s.type_name, "", s.group_id, s.thread, s.cpu,
-                                     s.count + other->count, s.scale, true, csv_);
+            summaries_.emplace_back(s.type_name, "", s.group_id, s.thread, s.cpu,
+                                    s.count + other->count, s.scale, true, csv_);
           }
         }
       }
@@ -904,16 +819,11 @@ void StatCommand::AdjustToIntervalOnlyValues(std::vector<CountersInfo>& counters
     }
     for (size_t j = 0; j < counters_per_event.size(); j++) {
       PerfCounter& counter = counters_per_event[j].counter;
-      CounterSum& sum = last_sum[j];
-      uint64_t tmp = counter.value;
-      counter.value -= sum.value;
-      sum.value = tmp;
-      tmp = counter.time_enabled;
-      counter.time_enabled -= sum.time_enabled;
-      sum.time_enabled = tmp;
-      tmp = counter.time_running;
-      counter.time_running -= sum.time_running;
-      sum.time_running = tmp;
+      CounterSum new_sum;
+      new_sum.FromCounter(counter);
+      CounterSum delta = new_sum - last_sum[j];
+      delta.ToCounter(counter);
+      last_sum[j] = new_sum;
     }
   }
 }
@@ -949,81 +859,11 @@ bool StatCommand::ShowCounters(const std::vector<CountersInfo>& counters,
     }
   }
 
-  bool counters_always_available = true;
-  CounterSummaries summaries(csv_);
-
-  auto add_summary = [&](const CountersInfo& info, pid_t tid, int cpu, const CounterSum& sum) {
-    double scale = 1.0;
-    if (sum.time_running < sum.time_enabled && sum.time_running != 0) {
-      scale = static_cast<double>(sum.time_enabled) / sum.time_running;
-    }
-    if ((report_per_thread_ || report_per_core_) && sum.time_running == 0) {
-      // No need to report threads or cpus not running.
-      return;
-    }
-    ThreadInfo* thread = nullptr;
-    if (report_per_thread_) {
-      auto it = thread_info_.find(tid);
-      CHECK(it != thread_info_.end());
-      thread = &it->second;
-    }
-    summaries.Summaries().emplace_back(info.event_name, info.event_modifier, info.group_id,
-                                       thread, cpu, sum.value, scale, false, csv_);
-    counters_always_available &= summaries.Summaries().back().IsMonitoredAllTheTime();
-  };
-
-  auto sort_summaries = [&](std::vector<CounterSummary>::iterator begin,
-                            std::vector<CounterSummary>::iterator end) {
-    if (report_per_thread_ && report_per_core_) {
-      // First sort by event count for all cpus in a thread, then sort by event count of each cpu.
-      std::unordered_map<pid_t, uint64_t> count_per_thread;
-      for (auto it = begin; it != end; ++it) {
-        count_per_thread[it->thread->tid] += it->count;
-      }
-      std::sort(begin, end, [&](const CounterSummary& s1, const CounterSummary& s2) {
-        pid_t tid1 = s1.thread->tid;
-        pid_t tid2 = s2.thread->tid;
-        if (tid1 != tid2) {
-          if (count_per_thread[tid1] != count_per_thread[tid2]) {
-            return count_per_thread[tid1] > count_per_thread[tid2];
-          }
-          return tid1 < tid2;
-        }
-        return s1.count > s2.count;
-      });
-    } else {
-      std::sort(begin, end, [](const CounterSummary& s1, const CounterSummary& s2) {
-        return s1.count > s2.count;
-      });
-    }
-  };
-
+  CounterSummaryBuilder builder(report_per_thread_, report_per_core_, csv_, thread_info_);
   for (const auto& info : counters) {
-    std::unordered_map<uint64_t, CounterSum> sum_map;
-    for (auto& counter : info.counters) {
-      uint64_t key = 0;
-      if (report_per_thread_) {
-        key |= counter.tid;
-      }
-      if (report_per_core_) {
-        key |= static_cast<uint64_t>(counter.cpu) << 32;
-      }
-      CounterSum& sum = sum_map[key];
-      sum.value += counter.counter.value;
-      sum.time_enabled = counter.counter.time_enabled;
-      sum.time_running = counter.counter.time_running;
-    }
-    size_t pre_sum_count = summaries.Summaries().size();
-    for (const auto& pair : sum_map) {
-      pid_t tid = report_per_thread_ ? static_cast<pid_t>(pair.first & UINT32_MAX) : 0;
-      int cpu = report_per_core_ ? static_cast<int>(pair.first >> 32) : -1;
-      const CounterSum& sum = pair.second;
-      add_summary(info, tid, cpu, sum);
-    }
-    if (report_per_thread_ || report_per_core_) {
-      sort_summaries(summaries.Summaries().begin() + pre_sum_count, summaries.Summaries().end());
-    }
+    builder.AddCountersForOneEventType(info);
   }
+  CounterSummaries summaries(builder.Build(), csv_);
   summaries.AutoGenerateSummaries();
   summaries.GenerateComments(duration_in_sec);
   summaries.Show(fp);
@@ -1041,6 +881,13 @@ bool StatCommand::ShowCounters(const std::vector<CountersInfo>& counters,
       event_selection_set_.GetMonitoredThreads() == std::set<pid_t>({-1})) {
     // We either monitor a thread on all cpus, or monitor all threads on a cpu. In both cases,
     // if percentages < 100%, probably it is caused by hardware counter multiplexing.
+    bool counters_always_available = true;
+    for (const auto& summary : summaries.Summaries()) {
+      if (!summary.IsMonitoredAllTheTime()) {
+        counters_always_available = false;
+        break;
+      }
+    }
     if (!counters_always_available) {
       LOG(WARNING) << "Percentages < 100% means some events only run a subset of enabled time,\n"
                    << COUNTER_MULTIPLEX_INFO;
diff --git a/simpleperf/cmd_stat_impl.h b/simpleperf/cmd_stat_impl.h
new file mode 100644
index 00000000..16dd091d
--- /dev/null
+++ b/simpleperf/cmd_stat_impl.h
@@ -0,0 +1,246 @@
+/*
+ * Copyright (C) 2020 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#pragma once
+
+#include <math.h>
+#include <sys/types.h>
+
+#include <algorithm>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+#include <android-base/stringprintf.h>
+
+#include "event_selection_set.h"
+
+namespace simpleperf {
+
+struct CounterSum {
+  uint64_t value = 0;
+  uint64_t time_enabled = 0;
+  uint64_t time_running = 0;
+
+  void FromCounter(const PerfCounter& counter) {
+    value = counter.value;
+    time_enabled = counter.time_enabled;
+    time_running = counter.time_running;
+  }
+
+  void ToCounter(PerfCounter& counter) const {
+    counter.value = value;
+    counter.time_enabled = time_enabled;
+    counter.time_running = time_running;
+  }
+
+  CounterSum operator+(const CounterSum& other) const {
+    CounterSum res;
+    res.value = value + other.value;
+    res.time_enabled = time_enabled + other.time_enabled;
+    res.time_running = time_running + other.time_running;
+    return res;
+  }
+
+  CounterSum operator-(const CounterSum& other) const {
+    CounterSum res;
+    res.value = value - other.value;
+    res.time_enabled = time_enabled - other.time_enabled;
+    res.time_running = time_running - other.time_running;
+    return res;
+  }
+};
+
+struct ThreadInfo {
+  pid_t tid;
+  pid_t pid;
+  std::string name;
+};
+
+struct CounterSummary {
+  std::string type_name;
+  std::string modifier;
+  uint32_t group_id;
+  const ThreadInfo* thread;
+  int cpu;  // -1 represents all cpus
+  uint64_t count;
+  double scale;
+  std::string readable_count;
+  std::string comment;
+  bool auto_generated;
+
+  CounterSummary(const std::string& type_name, const std::string& modifier, uint32_t group_id,
+                 const ThreadInfo* thread, int cpu, uint64_t count, double scale,
+                 bool auto_generated, bool csv)
+      : type_name(type_name),
+        modifier(modifier),
+        group_id(group_id),
+        thread(thread),
+        cpu(cpu),
+        count(count),
+        scale(scale),
+        auto_generated(auto_generated) {
+    readable_count = ReadableCountValue(csv);
+  }
+
+  bool IsMonitoredAtTheSameTime(const CounterSummary& other) const {
+    // Two summaries are monitored at the same time if they are in the same
+    // group or are monitored all the time.
+    if (group_id == other.group_id) {
+      return true;
+    }
+    return IsMonitoredAllTheTime() && other.IsMonitoredAllTheTime();
+  }
+
+  std::string Name() const {
+    if (modifier.empty()) {
+      return type_name;
+    }
+    return type_name + ":" + modifier;
+  }
+
+  bool IsMonitoredAllTheTime() const {
+    // If an event runs all the time it is enabled (by not sharing hardware
+    // counters with other events), the scale of its summary is usually within
+    // [1, 1 + 1e-5]. By setting SCALE_ERROR_LIMIT to 1e-5, We can identify
+    // events monitored all the time in most cases while keeping the report
+    // error rate <= 1e-5.
+    constexpr double SCALE_ERROR_LIMIT = 1e-5;
+    return (fabs(scale - 1.0) < SCALE_ERROR_LIMIT);
+  }
+
+ private:
+  std::string ReadableCountValue(bool csv) {
+    if (type_name == "cpu-clock" || type_name == "task-clock") {
+      // Convert nanoseconds to milliseconds.
+      double value = count / 1e6;
+      return android::base::StringPrintf("%lf(ms)", value);
+    } else {
+      // Convert big numbers to human friendly mode. For example,
+      // 1000000 will be converted to 1,000,000.
+      std::string s = android::base::StringPrintf("%" PRIu64, count);
+      if (csv) {
+        return s;
+      } else {
+        for (size_t i = s.size() - 1, j = 1; i > 0; --i, ++j) {
+          if (j == 3) {
+            s.insert(s.begin() + i, ',');
+            j = 0;
+          }
+        }
+        return s;
+      }
+    }
+  }
+};
+
+// Build a vector of CounterSummary.
+class CounterSummaryBuilder {
+ public:
+  CounterSummaryBuilder(bool report_per_thread, bool report_per_core, bool csv,
+                        const std::unordered_map<pid_t, ThreadInfo>& thread_map)
+      : report_per_thread_(report_per_thread),
+        report_per_core_(report_per_core),
+        csv_(csv),
+        thread_map_(thread_map) {}
+
+  void AddCountersForOneEventType(const CountersInfo& info) {
+    std::unordered_map<uint64_t, CounterSum> sum_map;
+    for (const auto& counter : info.counters) {
+      uint64_t key = 0;
+      if (report_per_thread_) {
+        key |= counter.tid;
+      }
+      if (report_per_core_) {
+        key |= static_cast<uint64_t>(counter.cpu) << 32;
+      }
+      CounterSum& sum = sum_map[key];
+      CounterSum add;
+      add.FromCounter(counter.counter);
+      sum = sum + add;
+    }
+    size_t pre_sum_count = summaries_.size();
+    for (const auto& pair : sum_map) {
+      pid_t tid = report_per_thread_ ? static_cast<pid_t>(pair.first & UINT32_MAX) : 0;
+      int cpu = report_per_core_ ? static_cast<int>(pair.first >> 32) : -1;
+      const CounterSum& sum = pair.second;
+      AddSummary(info, tid, cpu, sum);
+    }
+    if (report_per_thread_ || report_per_core_) {
+      SortSummaries(summaries_.begin() + pre_sum_count, summaries_.end());
+    }
+  }
+
+  std::vector<CounterSummary> Build() {
+    std::vector<CounterSummary> res = std::move(summaries_);
+    summaries_.clear();
+    return res;
+  }
+
+ private:
+  void AddSummary(const CountersInfo& info, pid_t tid, int cpu, const CounterSum& sum) {
+    double scale = 1.0;
+    if (sum.time_running < sum.time_enabled && sum.time_running != 0) {
+      scale = static_cast<double>(sum.time_enabled) / sum.time_running;
+    }
+    if ((report_per_thread_ || report_per_core_) && sum.time_running == 0) {
+      // No need to report threads or cpus not running.
+      return;
+    }
+    const ThreadInfo* thread = nullptr;
+    if (report_per_thread_) {
+      auto it = thread_map_.find(tid);
+      CHECK(it != thread_map_.end());
+      thread = &it->second;
+    }
+    summaries_.emplace_back(info.event_name, info.event_modifier, info.group_id, thread, cpu,
+                            sum.value, scale, false, csv_);
+  }
+
+  void SortSummaries(std::vector<CounterSummary>::iterator begin,
+                     std::vector<CounterSummary>::iterator end) {
+    if (report_per_thread_ && report_per_core_) {
+      // First sort by event count for all cpus in a thread, then sort by event count of each cpu.
+      std::unordered_map<pid_t, uint64_t> count_per_thread;
+      for (auto it = begin; it != end; ++it) {
+        count_per_thread[it->thread->tid] += it->count;
+      }
+      std::sort(begin, end, [&](const CounterSummary& s1, const CounterSummary& s2) {
+        pid_t tid1 = s1.thread->tid;
+        pid_t tid2 = s2.thread->tid;
+        if (tid1 != tid2) {
+          if (count_per_thread[tid1] != count_per_thread[tid2]) {
+            return count_per_thread[tid1] > count_per_thread[tid2];
+          }
+          return tid1 < tid2;
+        }
+        return s1.count > s2.count;
+      });
+    } else {
+      std::sort(begin, end, [](const CounterSummary& s1, const CounterSummary& s2) {
+        return s1.count > s2.count;
+      });
+    }
+  };
+
+  const bool report_per_thread_;
+  const bool report_per_core_;
+  const bool csv_;
+  const std::unordered_map<pid_t, ThreadInfo>& thread_map_;
+  std::vector<CounterSummary> summaries_;
+};
+
+}  // namespace simpleperf
\ No newline at end of file
diff --git a/simpleperf/cmd_stat_test.cpp b/simpleperf/cmd_stat_test.cpp
index bbeb4b66..8bc4e4ff 100644
--- a/simpleperf/cmd_stat_test.cpp
+++ b/simpleperf/cmd_stat_test.cpp
@@ -22,12 +22,15 @@
 
 #include <thread>
 
+#include "cmd_stat_impl.h"
 #include "command.h"
 #include "environment.h"
 #include "event_selection_set.h"
 #include "get_test_data.h"
 #include "test_util.h"
 
+using namespace simpleperf;
+
 static std::unique_ptr<Command> StatCmd() {
   return CreateCommandInstance("stat");
 }
@@ -341,3 +344,146 @@ TEST(stat_cmd, per_core_option) {
   ASSERT_TRUE(StatCmd()->Run({"--per-core", "sleep", "0.1"}));
   TEST_IN_ROOT(StatCmd()->Run({"--per-core", "-a", "--duration", "0.1"}));
 }
+
+TEST(stat_cmd, counter_sum) {
+  PerfCounter counter;
+  counter.value = 1;
+  counter.time_enabled = 2;
+  counter.time_running = 3;
+  CounterSum a;
+  a.FromCounter(counter);
+  ASSERT_EQ(a.value, 1);
+  ASSERT_EQ(a.time_enabled, 2);
+  ASSERT_EQ(a.time_running, 3);
+  CounterSum b = a + a;
+  ASSERT_EQ(b.value, 2);
+  ASSERT_EQ(b.time_enabled, 4);
+  ASSERT_EQ(b.time_running, 6);
+  CounterSum c = a - a;
+  ASSERT_EQ(c.value, 0);
+  ASSERT_EQ(c.time_enabled, 0);
+  ASSERT_EQ(c.time_running, 0);
+  b.ToCounter(counter);
+  ASSERT_EQ(counter.value, 2);
+  ASSERT_EQ(counter.time_enabled, 4);
+  ASSERT_EQ(counter.time_running, 6);
+}
+
+class StatCmdSummaryBuilderTest : public ::testing::Test {
+ protected:
+  void AddCounter(int event_id, pid_t tid, int cpu, int value, int time_enabled, int time_running) {
+    if (thread_map_.count(tid) == 0) {
+      ThreadInfo& thread = thread_map_[tid];
+      thread.pid = thread.tid = tid;
+      thread.name = "thread" + std::to_string(tid);
+    }
+    if (event_id >= counters_.size()) {
+      counters_.resize(event_id + 1);
+      counters_[event_id].group_id = 0;
+      counters_[event_id].event_name = "event" + std::to_string(event_id);
+    }
+    CountersInfo& info = counters_[event_id];
+    info.counters.resize(info.counters.size() + 1);
+    CounterInfo& counter = info.counters.back();
+    counter.tid = tid;
+    counter.cpu = cpu;
+    counter.counter.id = 0;
+    counter.counter.value = value;
+    counter.counter.time_enabled = time_enabled;
+    counter.counter.time_running = time_running;
+  }
+
+  std::vector<CounterSummary> BuildSummary(bool report_per_thread, bool report_per_core) {
+    CounterSummaryBuilder builder(report_per_thread, report_per_core, false, thread_map_);
+    for (auto& info : counters_) {
+      builder.AddCountersForOneEventType(info);
+    }
+    return builder.Build();
+  }
+
+  std::unordered_map<pid_t, ThreadInfo> thread_map_;
+  std::vector<CountersInfo> counters_;
+};
+
+TEST_F(StatCmdSummaryBuilderTest, multiple_events) {
+  AddCounter(0, 0, 0, 1, 1, 1);
+  AddCounter(1, 0, 0, 2, 2, 2);
+  std::vector<CounterSummary> summaries = BuildSummary(false, false);
+  ASSERT_EQ(summaries.size(), 2);
+  ASSERT_EQ(summaries[0].type_name, "event0");
+  ASSERT_EQ(summaries[0].count, 1);
+  ASSERT_NEAR(summaries[0].scale, 1.0, 1e-5);
+  ASSERT_EQ(summaries[1].type_name, "event1");
+  ASSERT_EQ(summaries[1].count, 2);
+  ASSERT_NEAR(summaries[1].scale, 1.0, 1e-5);
+}
+
+TEST_F(StatCmdSummaryBuilderTest, default_aggregate) {
+  AddCounter(0, 0, 0, 1, 1, 1);
+  AddCounter(0, 0, 1, 1, 1, 1);
+  AddCounter(0, 1, 0, 1, 1, 1);
+  AddCounter(0, 1, 1, 2, 2, 1);
+  std::vector<CounterSummary> summaries = BuildSummary(false, false);
+  ASSERT_EQ(summaries.size(), 1);
+  ASSERT_EQ(summaries[0].count, 5);
+  ASSERT_NEAR(summaries[0].scale, 1.25, 1e-5);
+}
+
+TEST_F(StatCmdSummaryBuilderTest, per_thread_aggregate) {
+  AddCounter(0, 0, 0, 1, 1, 1);
+  AddCounter(0, 0, 1, 1, 1, 1);
+  AddCounter(0, 1, 0, 1, 1, 1);
+  AddCounter(0, 1, 1, 2, 2, 1);
+  std::vector<CounterSummary> summaries = BuildSummary(true, false);
+  ASSERT_EQ(summaries.size(), 2);
+  ASSERT_EQ(summaries[0].thread->tid, 1);
+  ASSERT_EQ(summaries[0].cpu, -1);
+  ASSERT_EQ(summaries[0].count, 3);
+  ASSERT_NEAR(summaries[0].scale, 1.5, 1e-5);
+  ASSERT_EQ(summaries[1].thread->tid, 0);
+  ASSERT_EQ(summaries[0].cpu, -1);
+  ASSERT_EQ(summaries[1].count, 2);
+  ASSERT_NEAR(summaries[1].scale, 1.0, 1e-5);
+}
+
+TEST_F(StatCmdSummaryBuilderTest, per_core_aggregate) {
+  AddCounter(0, 0, 0, 1, 1, 1);
+  AddCounter(0, 0, 1, 1, 1, 1);
+  AddCounter(0, 1, 0, 1, 1, 1);
+  AddCounter(0, 1, 1, 2, 2, 1);
+  std::vector<CounterSummary> summaries = BuildSummary(false, true);
+  ASSERT_EQ(summaries.size(), 2);
+  ASSERT_TRUE(summaries[0].thread == nullptr);
+  ASSERT_EQ(summaries[0].cpu, 1);
+  ASSERT_EQ(summaries[0].count, 3);
+  ASSERT_NEAR(summaries[0].scale, 1.5, 1e-5);
+  ASSERT_TRUE(summaries[1].thread == nullptr);
+  ASSERT_EQ(summaries[1].cpu, 0);
+  ASSERT_EQ(summaries[1].count, 2);
+  ASSERT_NEAR(summaries[1].scale, 1.0, 1e-5);
+}
+
+TEST_F(StatCmdSummaryBuilderTest, per_thread_core_aggregate) {
+  AddCounter(0, 0, 0, 1, 1, 1);
+  AddCounter(0, 0, 1, 2, 1, 1);
+  AddCounter(0, 1, 0, 3, 1, 1);
+  AddCounter(0, 1, 1, 4, 2, 1);
+  std::vector<CounterSummary> summaries = BuildSummary(true, true);
+  ASSERT_EQ(summaries.size(), 4);
+  ASSERT_EQ(summaries[0].thread->tid, 1);
+  ASSERT_EQ(summaries[0].cpu, 1);
+  ASSERT_EQ(summaries[0].count, 4);
+  ASSERT_NEAR(summaries[0].scale, 2.0, 1e-5);
+  ASSERT_EQ(summaries[1].thread->tid, 1);
+  ASSERT_EQ(summaries[1].cpu, 0);
+  ASSERT_EQ(summaries[1].count, 3);
+  ASSERT_NEAR(summaries[1].scale, 1.0, 1e-5);
+  ASSERT_EQ(summaries[2].thread->tid, 0);
+  ASSERT_EQ(summaries[2].cpu, 1);
+  ASSERT_EQ(summaries[2].count, 2);
+  ASSERT_NEAR(summaries[2].scale, 1.0, 1e-5);
+  ASSERT_EQ(summaries[3].thread->tid, 0);
+  ASSERT_EQ(summaries[3].cpu, 0);
+  ASSERT_EQ(summaries[3].count, 1);
+  ASSERT_NEAR(summaries[3].scale, 1.0, 1e-5);
+}