From: Shashi Shekhar <shashishekhar@google.com>
Date: Thu, 24 May 2018 00:14:39 +0000 (-0700)
Subject: Refactor StatSummarizer extract common functionality without proto dependencies.
X-Git-Tag: upstream/v1.9.0_rc1~38^2~4^2~147
X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=2307db76a2a07c7af6581e0ef4c6a5a0b83921f4;p=platform%2Fupstream%2Ftensorflow.git

Refactor StatSummarizer extract common functionality without proto dependencies.

PiperOrigin-RevId: 197816405
---

diff --git a/tensorflow/contrib/android/jni/run_stats_jni.cc b/tensorflow/contrib/android/jni/run_stats_jni.cc
index 707853b..30de7b5 100644
--- a/tensorflow/contrib/android/jni/run_stats_jni.cc
+++ b/tensorflow/contrib/android/jni/run_stats_jni.cc
@@ -16,6 +16,7 @@ limitations under the License.
 #include "tensorflow/contrib/android/jni/run_stats_jni.h"
 
 #include <jni.h>
+
 #include <sstream>
 
 #include "tensorflow/core/protobuf/config.pb.h"
@@ -73,7 +74,8 @@ JNIEXPORT jstring RUN_STATS_METHOD(summary)(JNIEnv* env, jclass clazz,
   StatSummarizer* s = requireHandle(env, handle);
   if (s == nullptr) return nullptr;
   std::stringstream ret;
-  ret << s->GetStatsByMetric("Top 10 CPU", StatSummarizer::BY_TIME, 10)
+  ret << s->GetStatsByMetric("Top 10 CPU", tensorflow::StatsCalculator::BY_TIME,
+                             10)
       << s->GetStatsByNodeType() << s->ShortSummary();
   return env->NewStringUTF(ret.str().c_str());
 }
diff --git a/tensorflow/contrib/lite/profiling/BUILD b/tensorflow/contrib/lite/profiling/BUILD
index 15999e5..c86be65 100644
--- a/tensorflow/contrib/lite/profiling/BUILD
+++ b/tensorflow/contrib/lite/profiling/BUILD
@@ -31,6 +31,33 @@ cc_library(
     copts = common_copts,
 )
 
+cc_library(
+    name = "profile_summarizer",
+    srcs = ["profile_summarizer.cc"],
+    hdrs = ["profile_summarizer.h"],
+    deps = [
+        ":profiler",
+        "//tensorflow/contrib/lite:framework",
+        "//tensorflow/contrib/lite/schema:schema_fbs",
+        "//tensorflow/core:stats_calculator_portable",
+    ],
+)
+
+cc_test(
+    name = "profile_summarizer_test",
+    srcs = ["profile_summarizer_test.cc"],
+    deps = [
+        ":profile_summarizer",
+        "//tensorflow/contrib/lite:framework",
+        "//tensorflow/contrib/lite:schema_fbs_version",
+        "//tensorflow/contrib/lite/kernels:builtin_ops",
+        "//tensorflow/contrib/lite/kernels:kernel_util",
+        "//tensorflow/contrib/lite/kernels:test_util",
+        "//tensorflow/contrib/lite/testing:util",
+        "@com_google_googletest//:gtest",
+    ],
+)
+
 cc_test(
     name = "profile_buffer_test",
     srcs = ["profile_buffer_test.cc"],
diff --git a/tensorflow/contrib/lite/profiling/profile_summarizer.cc b/tensorflow/contrib/lite/profiling/profile_summarizer.cc
new file mode 100644
index 0000000..788f692
--- /dev/null
+++ b/tensorflow/contrib/lite/profiling/profile_summarizer.cc
@@ -0,0 +1,140 @@
+/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+
+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.
+==============================================================================*/
+
+#include "tensorflow/contrib/lite/profiling/profile_summarizer.h"
+
+#include <sstream>
+
+#include "tensorflow/contrib/lite/schema/schema_generated.h"
+
+namespace tflite {
+namespace profiling {
+namespace {
+
+using Detail = tensorflow::StatsCalculator::Detail;
+
+struct OperatorDetails {
+  string name;
+  std::vector<string> inputs;
+  std::vector<string> outputs;
+};
+
+string GetTensorName(const tflite::Interpreter& interpreter, int tensor_index) {
+  const auto tensor = interpreter.tensor(tensor_index);
+  if (tensor == nullptr || tensor->name == nullptr) {
+    return "Unknown";
+  }
+  return tensor->name;
+}
+std::vector<string> GetTensorNames(const tflite::Interpreter& interpreter,
+                                   const TfLiteIntArray* tensor_indices) {
+  std::vector<string> tensors;
+  tensors.reserve(tensor_indices->size);
+  for (int i = 0; i < tensor_indices->size; i++) {
+    tensors.push_back(GetTensorName(interpreter, tensor_indices->data[i]));
+  }
+  return tensors;
+}
+
+string ToString(const std::vector<string>& str_vector) {
+  std::stringstream stream;
+  stream << "[";
+  bool first = true;
+  for (const auto& s : str_vector) {
+    if (!first) {
+      stream << ", ";
+    } else {
+      first = false;
+    }
+    stream << s;
+  }
+  stream << "]";
+  return stream.str();
+}
+
+OperatorDetails GetOperatorDetails(const tflite::Interpreter& interpreter,
+                                   int node_index) {
+  auto node_reg = interpreter.node_and_registration(node_index);
+  auto inputs = node_reg->first.inputs;
+  auto outputs = node_reg->first.outputs;
+  int code = node_reg->second.builtin_code;
+  const char* op_name = nullptr;
+  if (code == tflite::BuiltinOperator_CUSTOM) {
+    const char* custom_name = node_reg->second.custom_name;
+    op_name = custom_name ? custom_name : "UnknownCustomOp";
+  } else {
+    op_name = tflite::EnumNamesBuiltinOperator()[code];
+  }
+  OperatorDetails details;
+  details.name = op_name;
+  details.inputs = GetTensorNames(interpreter, inputs);
+  details.outputs = GetTensorNames(interpreter, outputs);
+  return details;
+}
+
+}  // namespace
+
+ProfileSummarizer::ProfileSummarizer()
+    : stats_calculator_(new ::tensorflow::StatsCalculator(
+          tensorflow::StatSummarizerOptions())) {}
+
+void ProfileSummarizer::ProcessProfiles(
+    const std::vector<const ProfileEvent*>& profile_stats,
+    const tflite::Interpreter& interpreter) {
+  std::vector<const ProfileEvent*> events;
+  std::copy_if(profile_stats.begin(), profile_stats.end(),
+               std::back_inserter(events), [](const ProfileEvent* e) {
+                 return e->event_type ==
+                            ProfileEvent::EventType::OPERATOR_INVOKE_EVENT &&
+                        e->end_timestamp_us >= e->begin_timestamp_us;
+               });
+  // Sort with begin_time.
+  std::sort(events.begin(), events.end(),
+            [](const ProfileEvent* const& a, const ProfileEvent* const& b) {
+              return a->begin_timestamp_us < b->begin_timestamp_us;
+            });
+  if (events.empty()) {
+    return;
+  }
+
+  int64_t base_start_us = events[0]->begin_timestamp_us;
+  int node_num = 0;
+  int64_t curr_total_us = 0;
+  std::map<std::string, Detail> details;
+  for (auto event : events) {
+    auto op_details = GetOperatorDetails(interpreter, event->event_metadata);
+    auto node_name = ToString(op_details.outputs);
+    auto result = details.emplace(node_name, Detail());
+    Detail* detail = &(result.first->second);
+    detail->start_us.UpdateStat(event->begin_timestamp_us - base_start_us);
+    int64_t node_exec_time =
+        event->end_timestamp_us - event->begin_timestamp_us;
+    detail->rel_end_us.UpdateStat(node_exec_time);
+    curr_total_us += node_exec_time;
+    ++node_num;
+
+    if (result.second) {
+      detail->name = node_name;
+      detail->type = op_details.name;
+      detail->run_order = node_num;
+      detail->times_called = 0;
+    }
+    ++detail->times_called;
+  }
+  stats_calculator_->UpdateDetails(details);
+  stats_calculator_->UpdateRunTotalUs(curr_total_us);
+}
+}  // namespace profiling
+}  // namespace tflite
diff --git a/tensorflow/contrib/lite/profiling/profile_summarizer.h b/tensorflow/contrib/lite/profiling/profile_summarizer.h
new file mode 100644
index 0000000..6fe6ca0
--- /dev/null
+++ b/tensorflow/contrib/lite/profiling/profile_summarizer.h
@@ -0,0 +1,58 @@
+/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+
+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.
+==============================================================================*/
+
+#ifndef TENSORFLOW_CONTRIB_LITE_PROFILING_PROFILE_SUMMARIZER_H_
+#define TENSORFLOW_CONTRIB_LITE_PROFILING_PROFILE_SUMMARIZER_H_
+
+#include <vector>
+
+#include "tensorflow/contrib/lite/interpreter.h"
+#include "tensorflow/contrib/lite/profiling/profiler.h"
+#include "tensorflow/core/util/stats_calculator.h"
+
+namespace tflite {
+namespace profiling {
+
+// Creates a summary of operator invocations in the interpreter.
+class ProfileSummarizer {
+ public:
+  ProfileSummarizer();
+  virtual ~ProfileSummarizer() {}
+
+  // Process profile events to update statistics for operator invocations.
+  void ProcessProfiles(const std::vector<const ProfileEvent*>& profile_stats,
+                       const tflite::Interpreter& interpreter);
+
+  // Returns a string detailing the accumulated runtime stats in a tab-separated
+  // format which can be pasted into a spreadsheet for further analysis.
+  std::string GetOutputString() const {
+    return stats_calculator_->GetOutputString();
+  }
+
+  std::string GetShortSummary() const {
+    return stats_calculator_->GetShortSummary();
+  }
+
+  // Prints the string returned by GetOutputString().
+  void PrintStepStats() const { stats_calculator_->PrintStepStats(); }
+
+ private:
+  std::unique_ptr<tensorflow::StatsCalculator> stats_calculator_;
+};
+
+}  // namespace profiling
+}  // namespace tflite
+
+#endif  // TENSORFLOW_CONTRIB_LITE_PROFILING_PROFILE_SUMMARIZER_H_
diff --git a/tensorflow/contrib/lite/profiling/profile_summarizer_test.cc b/tensorflow/contrib/lite/profiling/profile_summarizer_test.cc
new file mode 100644
index 0000000..35cf780
--- /dev/null
+++ b/tensorflow/contrib/lite/profiling/profile_summarizer_test.cc
@@ -0,0 +1,116 @@
+/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+
+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.
+==============================================================================*/
+
+#include <string>
+#include <vector>
+
+#include <gmock/gmock.h>
+#include <gtest/gtest.h>
+#include "tensorflow/contrib/lite/context.h"
+#include "tensorflow/contrib/lite/kernels/kernel_util.h"
+#include "tensorflow/contrib/lite/kernels/test_util.h"
+#include "tensorflow/contrib/lite/model.h"
+#include "tensorflow/contrib/lite/profiling/profile_summarizer.h"
+#include "tensorflow/contrib/lite/testing/util.h"
+#include "tensorflow/contrib/lite/version.h"
+
+namespace tflite {
+namespace profiling {
+
+namespace {
+
+TfLiteStatus SimpleOpEval(TfLiteContext* context, TfLiteNode* node) {
+  const TfLiteTensor* input1 = tflite::GetInput(context, node, /*index=*/0);
+  const TfLiteTensor* input2 = tflite::GetInput(context, node, /*index=*/1);
+
+  TfLiteTensor* output = GetOutput(context, node, /*index=*/0);
+
+  int32_t* output_data = output->data.i32;
+  *output_data = *(input1->data.i32) + *(input2->data.i32);
+  return kTfLiteOk;
+}
+
+TfLiteRegistration* RegisterSimpleOp() {
+  static TfLiteRegistration registration = {nullptr,
+                                            nullptr,
+                                            nullptr,
+                                            SimpleOpEval,
+                                            tflite::BuiltinOperator_CUSTOM,
+                                            "SimpleOpEval",
+                                            1};
+  return &registration;
+}
+
+class SimpleOpModel : public SingleOpModel {
+ public:
+  void Init();
+  tflite::Interpreter* GetInterpreter() { return interpreter_.get(); }
+  void SetInputs(int32_t x, int32_t y) {
+    PopulateTensor(inputs_[0], {x});
+    PopulateTensor(inputs_[1], {y});
+  }
+  int32_t GetOutput() { return ExtractVector<int32_t>(output_)[0]; }
+
+ private:
+  int inputs_[2];
+  int output_;
+};
+
+void SimpleOpModel::Init() {
+  inputs_[0] = AddInput({TensorType_INT32, {1}});
+  inputs_[1] = AddInput({TensorType_INT32, {1}});
+  output_ = AddOutput({TensorType_INT32, {}});
+  SetCustomOp("SimpleAdd", {}, RegisterSimpleOp);
+  BuildInterpreter({GetShape(inputs_[0]), GetShape(inputs_[1])});
+}
+
+TEST(ProfileSummarizerTest, Empty) {
+  ProfileSummarizer summarizer;
+  std::string output = summarizer.GetOutputString();
+  EXPECT_GT(output.size(), 0);
+}
+
+#ifdef TFLITE_PROFILING_ENABLED
+TEST(ProfileSummarizerTest, Interpreter) {
+  Profiler profiler;
+  SimpleOpModel m;
+  m.Init();
+  auto interpreter = m.GetInterpreter();
+  interpreter->SetProfiler(&profiler);
+  profiler.StartProfiling();
+  m.SetInputs(1, 2);
+  m.Invoke();
+  // 3 = 1 + 2
+  EXPECT_EQ(m.GetOutput(), 3);
+  profiler.StopProfiling();
+  ProfileSummarizer summarizer;
+  auto events = profiler.GetProfileEvents();
+  EXPECT_EQ(1, events.size());
+  summarizer.ProcessProfiles(profiler.GetProfileEvents(), *interpreter);
+  auto output = summarizer.GetOutputString();
+  // TODO(shashishekhar): Add a better test here.
+  ASSERT_TRUE(output.find("SimpleOp") != std::string::npos) << output;
+}
+#endif
+
+}  // namespace
+}  // namespace profiling
+}  // namespace tflite
+
+int main(int argc, char** argv) {
+  ::tflite::LogToStderr();
+  ::testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
diff --git a/tensorflow/core/BUILD b/tensorflow/core/BUILD
index 19e88d6..ce68ee1 100644
--- a/tensorflow/core/BUILD
+++ b/tensorflow/core/BUILD
@@ -827,6 +827,8 @@ tf_cuda_library(
         "util/sparse/group_iterator.h",
         "util/sparse/sparse_tensor.h",
         "util/stat_summarizer.h",
+        "util/stat_summarizer_options.h",
+        "util/stats_calculator.h",
         "util/stream_executor_util.h",
         "util/strided_slice_op.h",
         "util/tensor_format.h",
@@ -852,6 +854,16 @@ tf_cuda_library(
 )
 
 cc_library(
+    name = "stats_calculator_portable",
+    srcs = ["util/stats_calculator.cc"],
+    hdrs = [
+        "util/stat_summarizer_options.h",
+        "util/stats_calculator.h",
+    ],
+    deps = [":platform_base"],
+)
+
+cc_library(
     name = "overflow",
     hdrs = ["util/overflow.h"],
     deps = [
diff --git a/tensorflow/core/util/stat_summarizer.cc b/tensorflow/core/util/stat_summarizer.cc
index 8447028..42a4801 100644
--- a/tensorflow/core/util/stat_summarizer.cc
+++ b/tensorflow/core/util/stat_summarizer.cc
@@ -31,26 +31,22 @@ limitations under the License.
 
 namespace tensorflow {
 
+using Detail = StatsCalculator::Detail;
+
 StatSummarizer::StatSummarizer(const StatSummarizerOptions& options)
-    : options_(options) {}
+    : stats_calculator_(new StatsCalculator(options)) {}
 
 StatSummarizer::StatSummarizer(const tensorflow::GraphDef& tensorflow_graph)
-    : StatSummarizer(StatSummarizerOptions()) {}
+    : stats_calculator_(new StatsCalculator(StatSummarizerOptions())) {}
 
 StatSummarizer::~StatSummarizer() {}
 
-void StatSummarizer::Reset() {
-  run_total_us_.Reset();
-  memory_.Reset();
-  details_.clear();
-}
-
-void StatSummarizer::Validate(const Detail* detail,
+void StatSummarizer::Validate(const std::vector<TensorDescription>* outputs,
                               const NodeExecStats& ns) const {
-  if (detail->outputs.size() != ns.output_size()) {
+  if (outputs->size() != ns.output_size()) {
     LOG(WARNING) << "Number of outputs changed between runs for '"
-                 << ns.node_name() << "' - was " << detail->outputs.size()
-                 << ", now " << ns.output_size();
+                 << ns.node_name() << "' - was " << outputs->size() << ", now "
+                 << ns.output_size();
   } else {
     for (const auto& output : ns.output()) {
       const int32 slot = output.slot();
@@ -58,7 +54,7 @@ void StatSummarizer::Validate(const Detail* detail,
         // This is not a hard error for Switch ops, so just pass.
         continue;
       }
-      const auto& stored = detail->outputs[slot];
+      const auto& stored = (*outputs)[slot];
       const auto& current = output.tensor_description();
 
       bool do_tensors_match =
@@ -129,6 +125,7 @@ void StatSummarizer::ProcessStepStats(const StepStats& step_stats) {
 
   int64 first_node_start_us =
       step_stats.dev_stats(0).node_stats(0).all_start_micros();
+  std::map<std::string, Detail> details;
 
   int node_num = 0;
   for (const auto& ds : step_stats.dev_stats()) {
@@ -172,7 +169,10 @@ void StatSummarizer::ProcessStepStats(const StepStats& step_stats) {
       ++node_num;
       const int64 curr_time = ns.all_end_rel_micros();
       curr_total_us += curr_time;
-      auto result = details_.emplace(name, Detail());
+      auto result = details.emplace(name, Detail());
+      auto output_result =
+          outputs_.emplace(name, std::vector<TensorDescription>());
+      std::vector<TensorDescription>* outputs = &(output_result.first->second);
       Detail* detail = &(result.first->second);
 
       detail->start_us.UpdateStat(ns.all_start_micros() - first_node_start_us);
@@ -185,16 +185,15 @@ void StatSummarizer::ProcessStepStats(const StepStats& step_stats) {
 
         detail->run_order = node_num;
 
-        detail->outputs.resize(ns.output_size());
+        outputs->resize(ns.output_size());
         for (const auto& output : ns.output()) {
           const int32 slot = output.slot();
           if ((slot < 0) || (slot >= ns.output_size())) {
             // This is not a hard error for Switch ops, so just pass.
             continue;
           }
-          detail->outputs[slot] = output.tensor_description();
+          (*outputs)[slot] = output.tensor_description();
         }
-
         detail->times_called = 0;
       }
 
@@ -207,273 +206,22 @@ void StatSummarizer::ProcessStepStats(const StepStats& step_stats) {
       mem_total += curr_node_mem;
 
       ++detail->times_called;
+      stats_calculator_->UpdateDetails(details);
 
-      Validate(detail, ns);
-    }
-  }
-
-  run_total_us_.UpdateStat(curr_total_us);
-  memory_.UpdateStat(mem_total);
-}
-
-std::string StatSummarizer::ShortSummary() const {
-  std::stringstream stream;
-  stream << "Timings (microseconds): ";
-  run_total_us_.OutputToStream(&stream);
-  stream << std::endl;
-
-  stream << "Memory (bytes): ";
-  memory_.OutputToStream(&stream);
-  stream << std::endl;
-
-  stream << details_.size() << " nodes observed" << std::endl;
-  return stream.str();
-}
-
-std::ostream& InitField(std::ostream& stream, int width) {
-  stream << "\t" << std::right << std::setw(width) << std::fixed
-         << std::setprecision(3);
-  return stream;
-}
-
-std::string StatSummarizer::HeaderString(const string& title) const {
-  std::stringstream stream;
-
-  stream << "============================== " << title
-         << " ==============================" << std::endl;
-
-  InitField(stream, 24) << "[node type]";
-  InitField(stream, 9) << "[start]";
-  InitField(stream, 9) << "[first]";
-  InitField(stream, 9) << "[avg ms]";
-  InitField(stream, 8) << "[%]";
-  InitField(stream, 8) << "[cdf%]";
-  InitField(stream, 10) << "[mem KB]";
-  InitField(stream, 9) << "[times called]";
-  stream << "\t"
-         << "[Name]";
-  return stream.str();
-}
-
-std::string StatSummarizer::ColumnString(const Detail& detail,
-                                         const int64 cumulative_stat_on_node,
-                                         const Stat<int64>& stat) const {
-  const double start_ms = detail.start_us.avg() / 1000.0;
-  const double first_time_ms = detail.rel_end_us.first() / 1000.0;
-  const double avg_time_ms = detail.rel_end_us.avg() / 1000.0;
-  const double percentage = detail.rel_end_us.sum() * 100.0 / stat.sum();
-  const double cdf_percentage = (cumulative_stat_on_node * 100.0f) / stat.sum();
-  const int64 times_called = detail.times_called / num_runs();
-
-  std::stringstream stream;
-  InitField(stream, 24) << detail.type;
-  InitField(stream, 9) << start_ms;
-  InitField(stream, 9) << first_time_ms;
-  InitField(stream, 9) << avg_time_ms;
-  InitField(stream, 7) << percentage << "%";
-  InitField(stream, 7) << cdf_percentage << "%";
-  InitField(stream, 10) << detail.mem_used.newest() / 1000.0;
-  InitField(stream, 9) << times_called;
-  stream << "\t" << detail.name;
-
-  return stream.str();
-}
-
-void StatSummarizer::OrderNodesByMetric(
-    SortingMetric metric, std::vector<const Detail*>* details) const {
-  std::priority_queue<std::pair<string, const Detail*>> sorted_list;
-  const int num_nodes = details_.size();
-
-  for (const auto& det : details_) {
-    const Detail* detail = &(det.second);
-    std::stringstream stream;
-    stream << std::setw(20) << std::right << std::setprecision(10)
-           << std::fixed;
-
-    switch (metric) {
-      case BY_NAME:
-        stream << detail->name;
-        break;
-      case BY_RUN_ORDER:
-        stream << num_nodes - detail->run_order;
-        break;
-      case BY_TIME:
-        stream << detail->rel_end_us.avg();
-        break;
-      case BY_MEMORY:
-        stream << detail->mem_used.avg();
-        break;
-      case BY_TYPE:
-        stream << detail->type;
-        break;
-      default:
-        stream << "";
-        break;
+      Validate(outputs, ns);
     }
-
-    sorted_list.emplace(stream.str(), detail);
-  }
-
-  while (!sorted_list.empty()) {
-    auto entry = sorted_list.top();
-    sorted_list.pop();
-    details->push_back(entry.second);
   }
-}
-
-void StatSummarizer::ComputeStatsByType(
-    std::map<string, int64>* node_type_map_count,
-    std::map<string, int64>* node_type_map_time,
-    std::map<string, int64>* node_type_map_memory,
-    std::map<string, int64>* node_type_map_times_called,
-    int64* accumulated_us) const {
-  int64 run_count = run_total_us_.count();
-
-  for (const auto& det : details_) {
-    const string node_name = det.first;
-    const Detail& detail = det.second;
-
-    int64 curr_time_val =
-        static_cast<int64>(detail.rel_end_us.sum() / run_count);
-    *accumulated_us += curr_time_val;
 
-    int64 curr_memory_val = detail.mem_used.newest();
-
-    const string& node_type = detail.type;
-
-    (*node_type_map_count)[node_type] += 1;
-    (*node_type_map_time)[node_type] += curr_time_val;
-    (*node_type_map_memory)[node_type] += curr_memory_val;
-    (*node_type_map_times_called)[node_type] += detail.times_called / run_count;
-  }
+  stats_calculator_->UpdateRunTotalUs(curr_total_us);
+  stats_calculator_->UpdateMemoryUsed(mem_total);
 }
 
-std::string StatSummarizer::GetStatsByNodeType() const {
-  std::stringstream stream;
-
-  stream << "============================== Summary by node type "
-            "=============================="
-         << std::endl;
-
-  LOG(INFO) << "Number of nodes executed: " << details_.size();
-
-  std::map<string, int64> node_type_map_count;
-  std::map<string, int64> node_type_map_time;
-  std::map<string, int64> node_type_map_memory;
-  std::map<string, int64> node_type_map_times_called;
-  int64 accumulated_us = 0;
-
-  ComputeStatsByType(&node_type_map_count, &node_type_map_time,
-                     &node_type_map_memory, &node_type_map_times_called,
-                     &accumulated_us);
-
-  // Sort them.
-  std::priority_queue<std::pair<int64, std::pair<string, int64>>> timings;
-  for (const auto& node_type : node_type_map_time) {
-    const int64 mem_used = node_type_map_memory[node_type.first];
-    timings.emplace(node_type.second,
-                    std::pair<string, int64>(node_type.first, mem_used));
-  }
-
-  InitField(stream, 24) << "[Node type]";
-  InitField(stream, 9) << "[count]";
-  InitField(stream, 10) << "[avg ms]";
-  InitField(stream, 11) << "[avg %]";
-  InitField(stream, 11) << "[cdf %]";
-  InitField(stream, 10) << "[mem KB]";
-  InitField(stream, 10) << "[times called]";
-  stream << std::endl;
-
-  float cdf = 0.0f;
-  while (!timings.empty()) {
-    auto entry = timings.top();
-    timings.pop();
-
-    const string node_type = entry.second.first;
-    const float memory = entry.second.second / 1000.0f;
-
-    const int64 node_type_total_us = entry.first;
-    const float time_per_run_ms = node_type_total_us / 1000.0f;
-
-    const float percentage =
-        ((entry.first / static_cast<float>(accumulated_us)) * 100.0f);
-    cdf += percentage;
-
-    InitField(stream, 24) << node_type;
-    InitField(stream, 9) << node_type_map_count[node_type];
-    InitField(stream, 10) << time_per_run_ms;
-    InitField(stream, 10) << percentage << "%";
-    InitField(stream, 10) << cdf << "%";
-    InitField(stream, 10) << memory;
-    InitField(stream, 9) << node_type_map_times_called[node_type];
-    stream << std::endl;
-  }
-  stream << std::endl;
-  return stream.str();
-}
-
-std::string StatSummarizer::GetStatsByMetric(const string& title,
-                                             SortingMetric sorting_metric,
-                                             int num_stats) const {
-  std::vector<const Detail*> details;
-  OrderNodesByMetric(sorting_metric, &details);
-
-  double cumulative_stat_on_node = 0;
-
-  std::stringstream stream;
-  stream << HeaderString(title) << std::endl;
-  int stat_num = 0;
-  for (auto detail : details) {
-    ++stat_num;
-    if (num_stats > 0 && stat_num > num_stats) {
-      break;
-    }
-
-    // TODO(andrewharp): Make this keep track of the particular metric for cdf.
-    cumulative_stat_on_node += detail->rel_end_us.sum();
-    stream << ColumnString(*detail, cumulative_stat_on_node, run_total_us_)
-           << std::endl;
-  }
-  stream << std::endl;
-  return stream.str();
-}
-
-std::string StatSummarizer::GetOutputString() const {
-  std::stringstream stream;
-  if (options_.show_run_order) {
-    stream << GetStatsByMetric("Run Order", BY_RUN_ORDER,
-                               options_.run_order_limit);
-  }
-  if (options_.show_time) {
-    stream << GetStatsByMetric("Top by Computation Time", BY_TIME,
-                               options_.time_limit);
-  }
-  if (options_.show_memory) {
-    stream << GetStatsByMetric("Top by Memory Use", BY_MEMORY,
-                               options_.memory_limit);
-  }
-  if (options_.show_type) {
-    stream << GetStatsByNodeType();
-  }
-  if (options_.show_summary) {
-    stream << ShortSummary() << std::endl;
-  }
-  return stream.str();
-}
-
-void StatSummarizer::PrintStepStats() const {
-  string output = GetOutputString();
-  std::istringstream iss(output);
-  for (std::string line; std::getline(iss, line);) {
-    LOG(INFO) << line;
-  }
-}
 
 void StatSummarizer::PrintOutputs() const {
   std::priority_queue<
       std::pair<int64, const std::pair<const std::string, Detail>*>>
       timings;
-  for (const auto& entry : details_) {
+  for (const auto& entry : stats_calculator_->GetDetails()) {
     timings.emplace(-entry.second.start_us.avg(), &entry);
   }
 
@@ -481,10 +229,10 @@ void StatSummarizer::PrintOutputs() const {
   while (!timings.empty()) {
     auto entry = timings.top();
     timings.pop();
-    const Detail& detail = entry.second->second;
     std::stringstream stream;
-    stream << entry.second->first << "\t" << detail.outputs.size();
-    for (const auto& tensor : detail.outputs) {
+    const auto detail_outputs = outputs_.at(entry.second->first);
+    stream << entry.second->first << "\t" << detail_outputs.size();
+    for (const auto& tensor : detail_outputs) {
       stream << "\t" << DataTypeString(tensor.dtype());
       stream << "\t" << tensor.shape().dim_size();
       for (const auto& d : tensor.shape().dim()) {
diff --git a/tensorflow/core/util/stat_summarizer.h b/tensorflow/core/util/stat_summarizer.h
index 79fa637..39cd948 100644
--- a/tensorflow/core/util/stat_summarizer.h
+++ b/tensorflow/core/util/stat_summarizer.h
@@ -13,20 +13,23 @@ See the License for the specific language governing permissions and
 limitations under the License.
 ==============================================================================*/
 
-#ifndef TENSORFLOW_UTIL_STAT_SUMMARIZER_H_
-#define TENSORFLOW_UTIL_STAT_SUMMARIZER_H_
+#ifndef TENSORFLOW_CORE_UTIL_STAT_SUMMARIZER_H_
+#define TENSORFLOW_CORE_UTIL_STAT_SUMMARIZER_H_
 
 #include <stdlib.h>
 
 #include <cmath>
 #include <limits>
 #include <map>
+#include <memory>
 #include <sstream>
 #include <string>
 
 #include "tensorflow/core/framework/tensor.h"
 #include "tensorflow/core/framework/types.pb.h"
 #include "tensorflow/core/platform/types.h"
+#include "tensorflow/core/util/stat_summarizer_options.h"
+#include "tensorflow/core/util/stats_calculator.h"
 
 namespace tensorflow {
 
@@ -34,103 +37,6 @@ class GraphDef;
 class StepStats;
 class NodeExecStats;
 
-template <typename ValueType, typename HighPrecisionValueType = double>
-class Stat {
- public:
-  void UpdateStat(ValueType v) {
-    if (count_ == 0) {
-      first_ = v;
-    }
-
-    newest_ = v;
-    max_ = std::max(v, max_);
-    min_ = std::min(v, min_);
-    ++count_;
-    sum_ += v;
-    squared_sum_ += static_cast<HighPrecisionValueType>(v) * v;
-  }
-
-  void Reset() { new (this) Stat<ValueType, HighPrecisionValueType>(); }
-
-  bool empty() const { return count_ == 0; }
-
-  ValueType first() const { return first_; }
-
-  ValueType newest() const { return newest_; }
-
-  ValueType max() const { return max_; }
-
-  ValueType min() const { return min_; }
-
-  int64 count() const { return count_; }
-
-  ValueType sum() const { return sum_; }
-
-  HighPrecisionValueType squared_sum() const { return squared_sum_; }
-
-  bool all_same() const { return (count_ == 0 || min_ == max_); }
-
-  HighPrecisionValueType avg() const {
-    return empty() ? std::numeric_limits<ValueType>::quiet_NaN()
-                   : static_cast<HighPrecisionValueType>(sum_) / count_;
-  }
-
-  ValueType std_deviation() const {
-    return all_same() ? 0 : sqrt(squared_sum_ / count_ - avg() * avg());
-  }
-
-  void OutputToStream(std::ostream* stream) const {
-    if (empty()) {
-      *stream << "count=0";
-    } else if (all_same()) {
-      *stream << "count=" << count_ << " curr=" << newest_;
-      if (count_ > 1) *stream << "(all same)";
-    } else {
-      *stream << "count=" << count_ << " first=" << first_
-              << " curr=" << newest_ << " min=" << min_ << " max=" << max_
-              << " avg=" << avg() << " std=" << std_deviation();
-    }
-  }
-
-  friend std::ostream& operator<<(std::ostream& stream,
-                                  const Stat<ValueType>& stat) {
-    stat.OutputToStream(&stream);
-    return stream;
-  }
-
- private:
-  ValueType first_ = 0;
-  ValueType newest_ = 0;
-  ValueType max_ = std::numeric_limits<ValueType>::min();
-  ValueType min_ = std::numeric_limits<ValueType>::max();
-  int64 count_ = 0;
-  ValueType sum_ = 0;
-  HighPrecisionValueType squared_sum_ = 0;
-};
-
-// Used to control the output of the statistics summarizer;
-class StatSummarizerOptions {
- public:
-  StatSummarizerOptions()
-      : show_run_order(true),
-        run_order_limit(0),
-        show_time(true),
-        time_limit(10),
-        show_memory(true),
-        memory_limit(10),
-        show_type(true),
-        show_summary(true) {}
-
-  bool show_run_order;
-  int run_order_limit;
-  bool show_time;
-  int time_limit;
-  bool show_memory;
-  int memory_limit;
-  bool show_type;
-  bool show_summary;
-};
-
 // A StatSummarizer assists in performance analysis of Graph executions.
 //
 // It summarizes time spent executing (on GPU/CPU), memory used etc. across
@@ -140,14 +46,6 @@ class StatSummarizerOptions {
 // See tensorflow/tools/benchmark/benchmark_model.cc for an example usage.
 class StatSummarizer {
  public:
-  enum SortingMetric {
-    BY_NAME,
-    BY_RUN_ORDER,
-    BY_TIME,
-    BY_MEMORY,
-    BY_TYPE,
-  };
-
   explicit StatSummarizer(const StatSummarizerOptions& options);
 
   // Deprecated: Use StatSummarizer(const StatSummarizerOptions&) instead. The
@@ -161,65 +59,51 @@ class StatSummarizer {
 
   // Returns a string detailing the accumulated runtime stats in a tab-separated
   // format which can be pasted into a spreadsheet for further analysis.
-  std::string GetOutputString() const;
+  std::string GetOutputString() const {
+    return stats_calculator_->GetOutputString();
+  }
 
-  std::string ShortSummary() const;
+  std::string ShortSummary() const {
+    return stats_calculator_->GetShortSummary();
+  }
 
   // Prints the string returned by GetOutputString().
-  void PrintStepStats() const;
+  void PrintStepStats() const { stats_calculator_->PrintStepStats(); }
 
   // Prints the output tensor sizes and types for each node.
   void PrintOutputs() const;
 
-  void ComputeStatsByType(std::map<string, int64>* node_type_map_count,
-                          std::map<string, int64>* node_type_map_time,
-                          std::map<string, int64>* node_type_map_memory,
-                          std::map<string, int64>* node_type_map_times_called,
-                          int64* accumulated_us) const;
+  void ComputeStatsByType(
+      std::map<std::string, int64_t>* node_type_map_count,
+      std::map<std::string, int64_t>* node_type_map_time,
+      std::map<std::string, int64_t>* node_type_map_memory,
+      std::map<std::string, int64_t>* node_type_map_times_called,
+      int64_t* accumulated_us) const {
+    stats_calculator_->ComputeStatsByType(
+        node_type_map_count, node_type_map_time, node_type_map_memory,
+        node_type_map_times_called, accumulated_us);
+  }
 
-  std::string GetStatsByNodeType() const;
+  std::string GetStatsByNodeType() const {
+    return stats_calculator_->GetStatsByNodeType();
+  }
 
   std::string GetStatsByMetric(const string& title,
-                               SortingMetric sorting_metric,
-                               int num_stats) const;
-
-  void Reset();
+                               StatsCalculator::SortingMetric sorting_metric,
+                               int num_stats) const {
+    return stats_calculator_->GetStatsByMetric(title, sorting_metric,
+                                               num_stats);
+  }
 
-  // Returns number of runs.
-  int num_runs() const { return static_cast<int>(run_total_us_.count()); }
+ private:
+  void Validate(const std::vector<TensorDescription>* outputs,
+                const NodeExecStats& ns) const;
 
-  // Returns stats of total microseconds spent by all nodes in each run.
-  const Stat<int64>& run_total_us() const { return run_total_us_; }
+  std::map<std::string, std::vector<TensorDescription> > outputs_;
 
- private:
-  struct Detail {
-    string name;
-    string type;
-    int64 run_order;
-    Stat<int64> start_us;
-    Stat<int64> rel_end_us;
-    Stat<int64> mem_used;
-    std::vector<TensorDescription> outputs;
-    int64 times_called;
-  };
-
-  void Validate(const Detail* detail, const NodeExecStats& ns) const;
-
-  void OrderNodesByMetric(SortingMetric sorting_metric,
-                          std::vector<const Detail*>* details) const;
-
-  std::string HeaderString(const string& title) const;
-  std::string ColumnString(const Detail& detail,
-                           const int64 cumulative_stat_on_node,
-                           const Stat<int64>& stat) const;
-
-  Stat<int64> run_total_us_;
-  Stat<int64> memory_;
-
-  std::map<std::string, Detail> details_;
-  StatSummarizerOptions options_;
+  std::unique_ptr<StatsCalculator> stats_calculator_;
 };
 
 }  // namespace tensorflow
 
-#endif  // TENSORFLOW_UTIL_STAT_SUMMARIZER_H_
+#endif  // TENSORFLOW_CORE_UTIL_STAT_SUMMARIZER_H_
diff --git a/tensorflow/core/util/stat_summarizer_options.h b/tensorflow/core/util/stat_summarizer_options.h
new file mode 100644
index 0000000..5780206
--- /dev/null
+++ b/tensorflow/core/util/stat_summarizer_options.h
@@ -0,0 +1,43 @@
+/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+
+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.
+==============================================================================*/
+
+#ifndef TENSORFLOW_CORE_UTIL_STAT_SUMMARIZER_OPTIONS_H_
+#define TENSORFLOW_CORE_UTIL_STAT_SUMMARIZER_OPTIONS_H_
+namespace tensorflow {
+// Used to control the output of the statistics summarizer;
+class StatSummarizerOptions {
+ public:
+  StatSummarizerOptions()
+      : show_run_order(true),
+        run_order_limit(0),
+        show_time(true),
+        time_limit(10),
+        show_memory(true),
+        memory_limit(10),
+        show_type(true),
+        show_summary(true) {}
+
+  bool show_run_order;
+  int run_order_limit;
+  bool show_time;
+  int time_limit;
+  bool show_memory;
+  int memory_limit;
+  bool show_type;
+  bool show_summary;
+};
+}  // namespace tensorflow
+
+#endif  // TENSORFLOW_CORE_UTIL_STAT_SUMMARIZER_OPTIONS_H_
diff --git a/tensorflow/core/util/stats_calculator.cc b/tensorflow/core/util/stats_calculator.cc
new file mode 100644
index 0000000..20353ec
--- /dev/null
+++ b/tensorflow/core/util/stats_calculator.cc
@@ -0,0 +1,289 @@
+/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+
+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.
+==============================================================================*/
+
+#include "tensorflow/core/util/stats_calculator.h"
+
+#include <iomanip>
+#include <map>
+#include <queue>
+#include <sstream>
+#include <string>
+
+#include "tensorflow/core/platform/logging.h"
+
+namespace tensorflow {
+
+StatsCalculator::StatsCalculator(const StatSummarizerOptions& options)
+    : options_(options) {}
+
+std::string StatsCalculator::GetShortSummary() const {
+  std::stringstream stream;
+  stream << "Timings (microseconds): ";
+  run_total_us_.OutputToStream(&stream);
+  stream << std::endl;
+
+  stream << "Memory (bytes): ";
+  memory_.OutputToStream(&stream);
+  stream << std::endl;
+
+  stream << details_.size() << " nodes observed" << std::endl;
+  return stream.str();
+}
+
+std::ostream& InitField(std::ostream& stream, int width) {
+  stream << "\t" << std::right << std::setw(width) << std::fixed
+         << std::setprecision(3);
+  return stream;
+}
+
+std::string StatsCalculator::HeaderString(const std::string& title) const {
+  std::stringstream stream;
+
+  stream << "============================== " << title
+         << " ==============================" << std::endl;
+
+  InitField(stream, 24) << "[node type]";
+  InitField(stream, 9) << "[start]";
+  InitField(stream, 9) << "[first]";
+  InitField(stream, 9) << "[avg ms]";
+  InitField(stream, 8) << "[%]";
+  InitField(stream, 8) << "[cdf%]";
+  InitField(stream, 10) << "[mem KB]";
+  InitField(stream, 9) << "[times called]";
+  stream << "\t"
+         << "[Name]";
+  return stream.str();
+}
+
+std::string StatsCalculator::ColumnString(const Detail& detail,
+                                          const int64_t cumulative_stat_on_node,
+                                          const Stat<int64_t>& stat) const {
+  const double start_ms = detail.start_us.avg() / 1000.0;
+  const double first_time_ms = detail.rel_end_us.first() / 1000.0;
+  const double avg_time_ms = detail.rel_end_us.avg() / 1000.0;
+  const double percentage = detail.rel_end_us.sum() * 100.0 / stat.sum();
+  const double cdf_percentage = (cumulative_stat_on_node * 100.0f) / stat.sum();
+  const int64_t times_called = detail.times_called / num_runs();
+
+  std::stringstream stream;
+  InitField(stream, 24) << detail.type;
+  InitField(stream, 9) << start_ms;
+  InitField(stream, 9) << first_time_ms;
+  InitField(stream, 9) << avg_time_ms;
+  InitField(stream, 7) << percentage << "%";
+  InitField(stream, 7) << cdf_percentage << "%";
+  InitField(stream, 10) << detail.mem_used.newest() / 1000.0;
+  InitField(stream, 9) << times_called;
+  stream << "\t" << detail.name;
+
+  return stream.str();
+}
+
+void StatsCalculator::OrderNodesByMetric(
+    SortingMetric metric, std::vector<const Detail*>* details) const {
+  std::priority_queue<std::pair<string, const Detail*>> sorted_list;
+  const int num_nodes = details_.size();
+
+  for (const auto& det : details_) {
+    const Detail* detail = &(det.second);
+    std::stringstream stream;
+    stream << std::setw(20) << std::right << std::setprecision(10)
+           << std::fixed;
+
+    switch (metric) {
+      case BY_NAME:
+        stream << detail->name;
+        break;
+      case BY_RUN_ORDER:
+        stream << num_nodes - detail->run_order;
+        break;
+      case BY_TIME:
+        stream << detail->rel_end_us.avg();
+        break;
+      case BY_MEMORY:
+        stream << detail->mem_used.avg();
+        break;
+      case BY_TYPE:
+        stream << detail->type;
+        break;
+      default:
+        stream << "";
+        break;
+    }
+
+    sorted_list.emplace(stream.str(), detail);
+  }
+
+  while (!sorted_list.empty()) {
+    auto entry = sorted_list.top();
+    sorted_list.pop();
+    details->push_back(entry.second);
+  }
+}
+
+void StatsCalculator::ComputeStatsByType(
+    std::map<std::string, int64_t>* node_type_map_count,
+    std::map<std::string, int64_t>* node_type_map_time,
+    std::map<std::string, int64_t>* node_type_map_memory,
+    std::map<std::string, int64_t>* node_type_map_times_called,
+    int64_t* accumulated_us) const {
+  int64_t run_count = run_total_us_.count();
+
+  for (const auto& det : details_) {
+    const string node_name = det.first;
+    const Detail& detail = det.second;
+
+    int64_t curr_time_val =
+        static_cast<int64_t>(detail.rel_end_us.sum() / run_count);
+    *accumulated_us += curr_time_val;
+
+    int64_t curr_memory_val = detail.mem_used.newest();
+
+    const string& node_type = detail.type;
+
+    (*node_type_map_count)[node_type] += 1;
+    (*node_type_map_time)[node_type] += curr_time_val;
+    (*node_type_map_memory)[node_type] += curr_memory_val;
+    (*node_type_map_times_called)[node_type] += detail.times_called / run_count;
+  }
+}
+
+std::string StatsCalculator::GetStatsByNodeType() const {
+  std::stringstream stream;
+
+  stream << "============================== Summary by node type "
+            "=============================="
+         << std::endl;
+
+  LOG(INFO) << "Number of nodes executed: " << details_.size();
+
+  std::map<std::string, int64_t> node_type_map_count;
+  std::map<std::string, int64_t> node_type_map_time;
+  std::map<std::string, int64_t> node_type_map_memory;
+  std::map<std::string, int64_t> node_type_map_times_called;
+  int64_t accumulated_us = 0;
+
+  ComputeStatsByType(&node_type_map_count, &node_type_map_time,
+                     &node_type_map_memory, &node_type_map_times_called,
+                     &accumulated_us);
+
+  // Sort them.
+  std::priority_queue<std::pair<int64_t, std::pair<string, int64_t>>> timings;
+  for (const auto& node_type : node_type_map_time) {
+    const int64_t mem_used = node_type_map_memory[node_type.first];
+    timings.emplace(node_type.second,
+                    std::pair<string, int64_t>(node_type.first, mem_used));
+  }
+
+  InitField(stream, 24) << "[Node type]";
+  InitField(stream, 9) << "[count]";
+  InitField(stream, 10) << "[avg ms]";
+  InitField(stream, 11) << "[avg %]";
+  InitField(stream, 11) << "[cdf %]";
+  InitField(stream, 10) << "[mem KB]";
+  InitField(stream, 10) << "[times called]";
+  stream << std::endl;
+
+  float cdf = 0.0f;
+  while (!timings.empty()) {
+    auto entry = timings.top();
+    timings.pop();
+
+    const string node_type = entry.second.first;
+    const float memory = entry.second.second / 1000.0f;
+
+    const int64_t node_type_total_us = entry.first;
+    const float time_per_run_ms = node_type_total_us / 1000.0f;
+
+    const float percentage =
+        ((entry.first / static_cast<float>(accumulated_us)) * 100.0f);
+    cdf += percentage;
+
+    InitField(stream, 24) << node_type;
+    InitField(stream, 9) << node_type_map_count[node_type];
+    InitField(stream, 10) << time_per_run_ms;
+    InitField(stream, 10) << percentage << "%";
+    InitField(stream, 10) << cdf << "%";
+    InitField(stream, 10) << memory;
+    InitField(stream, 9) << node_type_map_times_called[node_type];
+    stream << std::endl;
+  }
+  stream << std::endl;
+  return stream.str();
+}
+
+std::string StatsCalculator::GetStatsByMetric(const std::string& title,
+                                              SortingMetric sorting_metric,
+                                              int num_stats) const {
+  std::vector<const Detail*> details;
+  OrderNodesByMetric(sorting_metric, &details);
+
+  double cumulative_stat_on_node = 0;
+
+  std::stringstream stream;
+  stream << HeaderString(title) << std::endl;
+  int stat_num = 0;
+  for (auto detail : details) {
+    ++stat_num;
+    if (num_stats > 0 && stat_num > num_stats) {
+      break;
+    }
+
+    // TODO(andrewharp): Make this keep track of the particular metric for cdf.
+    cumulative_stat_on_node += detail->rel_end_us.sum();
+    stream << ColumnString(*detail, cumulative_stat_on_node, run_total_us_)
+           << std::endl;
+  }
+  stream << std::endl;
+  return stream.str();
+}
+
+std::string StatsCalculator::GetOutputString() const {
+  std::stringstream stream;
+  if (options_.show_run_order) {
+    stream << GetStatsByMetric("Run Order", BY_RUN_ORDER,
+                               options_.run_order_limit);
+  }
+  if (options_.show_time) {
+    stream << GetStatsByMetric("Top by Computation Time", BY_TIME,
+                               options_.time_limit);
+  }
+  if (options_.show_memory) {
+    stream << GetStatsByMetric("Top by Memory Use", BY_MEMORY,
+                               options_.memory_limit);
+  }
+  if (options_.show_type) {
+    stream << GetStatsByNodeType();
+  }
+  if (options_.show_summary) {
+    stream << GetShortSummary() << std::endl;
+  }
+  return stream.str();
+}
+
+void StatsCalculator::PrintStepStats() const {
+  string output = GetOutputString();
+  std::istringstream iss(output);
+  for (std::string line; std::getline(iss, line);) {
+    LOG(INFO) << line;
+  }
+}
+
+void StatsCalculator::UpdateDetails(
+    const std::map<std::string, Detail>& details) {
+  details_.insert(details.begin(), details.end());
+}
+
+}  // namespace tensorflow
diff --git a/tensorflow/core/util/stats_calculator.h b/tensorflow/core/util/stats_calculator.h
new file mode 100644
index 0000000..a103346
--- /dev/null
+++ b/tensorflow/core/util/stats_calculator.h
@@ -0,0 +1,189 @@
+/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+
+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.
+==============================================================================*/
+
+#ifndef TENSORFLOW_CORE_UTIL_STATS_CALCULATOR_H_
+#define TENSORFLOW_CORE_UTIL_STATS_CALCULATOR_H_
+
+#include <stdlib.h>
+
+#include <cmath>
+#include <limits>
+#include <map>
+#include <sstream>
+#include <string>
+#include <vector>
+
+#include "tensorflow/core/util/stat_summarizer_options.h"
+
+namespace tensorflow {
+
+template <typename ValueType, typename HighPrecisionValueType = double>
+class Stat {
+ public:
+  void UpdateStat(ValueType v) {
+    if (count_ == 0) {
+      first_ = v;
+    }
+
+    newest_ = v;
+    max_ = std::max(v, max_);
+    min_ = std::min(v, min_);
+    ++count_;
+    sum_ += v;
+    squared_sum_ += static_cast<HighPrecisionValueType>(v) * v;
+  }
+
+  void Reset() { new (this) Stat<ValueType, HighPrecisionValueType>(); }
+
+  bool empty() const { return count_ == 0; }
+
+  ValueType first() const { return first_; }
+
+  ValueType newest() const { return newest_; }
+
+  ValueType max() const { return max_; }
+
+  ValueType min() const { return min_; }
+
+  int64_t count() const { return count_; }
+
+  ValueType sum() const { return sum_; }
+
+  HighPrecisionValueType squared_sum() const { return squared_sum_; }
+
+  bool all_same() const { return (count_ == 0 || min_ == max_); }
+
+  HighPrecisionValueType avg() const {
+    return empty() ? std::numeric_limits<ValueType>::quiet_NaN()
+                   : static_cast<HighPrecisionValueType>(sum_) / count_;
+  }
+
+  ValueType std_deviation() const {
+    return all_same() ? 0 : sqrt(squared_sum_ / count_ - avg() * avg());
+  }
+
+  void OutputToStream(std::ostream* stream) const {
+    if (empty()) {
+      *stream << "count=0";
+    } else if (all_same()) {
+      *stream << "count=" << count_ << " curr=" << newest_;
+      if (count_ > 1) *stream << "(all same)";
+    } else {
+      *stream << "count=" << count_ << " first=" << first_
+              << " curr=" << newest_ << " min=" << min_ << " max=" << max_
+              << " avg=" << avg() << " std=" << std_deviation();
+    }
+  }
+
+  friend std::ostream& operator<<(std::ostream& stream,
+                                  const Stat<ValueType>& stat) {
+    stat.OutputToStream(&stream);
+    return stream;
+  }
+
+ private:
+  ValueType first_ = 0;
+  ValueType newest_ = 0;
+  ValueType max_ = std::numeric_limits<ValueType>::min();
+  ValueType min_ = std::numeric_limits<ValueType>::max();
+  int64_t count_ = 0;
+  ValueType sum_ = 0;
+  HighPrecisionValueType squared_sum_ = 0;
+};
+
+// A StatsCalculator assists in performance analysis of Graph executions.
+//
+// It summarizes time spent executing (on GPU/CPU), memory used etc for
+// graph execution.
+//
+// For example usage see StatsSummarizer.
+class StatsCalculator {
+ public:
+  enum SortingMetric {
+    BY_NAME,
+    BY_RUN_ORDER,
+    BY_TIME,
+    BY_MEMORY,
+    BY_TYPE,
+  };
+
+  explicit StatsCalculator(const StatSummarizerOptions& options);
+
+  // Returns a string detailing the accumulated runtime stats in a tab-separated
+  // format which can be pasted into a spreadsheet for further analysis.
+  std::string GetOutputString() const;
+
+  std::string GetShortSummary() const;
+
+  // Prints the string returned by GetOutputString().
+  void PrintStepStats() const;
+
+  void ComputeStatsByType(
+      std::map<std::string, int64_t>* node_type_map_count,
+      std::map<std::string, int64_t>* node_type_map_time,
+      std::map<std::string, int64_t>* node_type_map_memory,
+      std::map<std::string, int64_t>* node_type_map_times_called,
+      int64_t* accumulated_us) const;
+
+  std::string GetStatsByNodeType() const;
+
+  std::string GetStatsByMetric(const std::string& title,
+                               SortingMetric sorting_metric,
+                               int num_stats) const;
+
+  // Returns number of runs.
+  int num_runs() const { return static_cast<int>(run_total_us_.count()); }
+
+  // Returns stats of total microseconds spent by all nodes in each run.
+  const Stat<int64_t>& run_total_us() const { return run_total_us_; }
+
+  void UpdateRunTotalUs(int64_t run_total_us) {
+    run_total_us_.UpdateStat(run_total_us);
+  }
+
+  void UpdateMemoryUsed(int64_t memory) { memory_.UpdateStat(memory); }
+
+  struct Detail {
+    std::string name;
+    std::string type;
+    int64_t run_order;
+    Stat<int64_t> start_us;
+    Stat<int64_t> rel_end_us;
+    Stat<int64_t> mem_used;
+    int64_t times_called;
+  };
+
+  const std::map<std::string, Detail>& GetDetails() const { return details_; }
+  void UpdateDetails(const std::map<std::string, Detail>& details);
+
+ private:
+  void OrderNodesByMetric(SortingMetric sorting_metric,
+                          std::vector<const Detail*>* details) const;
+
+  std::string HeaderString(const std::string& title) const;
+  std::string ColumnString(const Detail& detail,
+                           const int64_t cumulative_stat_on_node,
+                           const Stat<int64_t>& stat) const;
+
+  Stat<int64_t> run_total_us_;
+  Stat<int64_t> memory_;
+
+  std::map<std::string, Detail> details_;
+  StatSummarizerOptions options_;
+};
+
+}  // namespace tensorflow
+
+#endif  // TENSORFLOW_CORE_UTIL_STATS_CALCULATOR_H_
diff --git a/tensorflow/python/util/stat_summarizer.i b/tensorflow/python/util/stat_summarizer.i
index 6aeaa0e..f423553 100644
--- a/tensorflow/python/util/stat_summarizer.i
+++ b/tensorflow/python/util/stat_summarizer.i
@@ -73,7 +73,7 @@ void _DeleteStatSummarizer(tensorflow::StatSummarizer* ss);
     return ss;
 }
 }
-
+%include "tensorflow/core/util/stat_summarizer_options.h"
 %include "tensorflow/core/util/stat_summarizer.h"
 %unignoreall
 
diff --git a/tensorflow/tools/benchmark/benchmark_model.cc b/tensorflow/tools/benchmark/benchmark_model.cc
index eeb1fab..de93b12 100644
--- a/tensorflow/tools/benchmark/benchmark_model.cc
+++ b/tensorflow/tools/benchmark/benchmark_model.cc
@@ -667,12 +667,12 @@ int Main(int argc, char** argv) {
         output_prefix, benchmark_name, "meta-init-plus-first-inference", 1,
         initialization_time_s + (warmup_time_us / 1000000.0) / warmup_runs);
 
-    std::map<string, int64> node_type_map_count;
-    std::map<string, int64> node_type_map_time;
-    std::map<string, int64> node_type_map_memory;
-    std::map<string, int64> node_type_map_times_called;
+    std::map<std::string, int64_t> node_type_map_count;
+    std::map<std::string, int64_t> node_type_map_time;
+    std::map<std::string, int64_t> node_type_map_memory;
+    std::map<std::string, int64_t> node_type_map_times_called;
 
-    int64 accumulated_us;
+    int64_t accumulated_us;
     stats->ComputeStatsByType(&node_type_map_count, &node_type_map_time,
                               &node_type_map_memory,
                               &node_type_map_times_called, &accumulated_us);