Refactor StatSummarizer: use Stats for more things, remove redundant methods, combine mem/time data structures, and overhaul sorting / output formatting.

Change: 141981032

1 files changed, 39 insertions, 64 deletions

diff --git a/tensorflow/core/util/stat_summarizer.h b/tensorflow/core/util/stat_summarizer.h
index 57b17bd208..f766f0f281 100644
--- a/tensorflow/core/util/stat_summarizer.h
+++ b/tensorflow/core/util/stat_summarizer.h
@@ -38,6 +38,10 @@ 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_);
@@ -50,6 +54,8 @@ class Stat {
 
   bool empty() const { return count_ == 0; }
 
+  ValueType first() const { return first_; }
+
   ValueType newest() const { return newest_; }
 
   ValueType max() const { return max_; }
@@ -77,12 +83,12 @@ class Stat {
     if (empty()) {
       *stream << "count=0";
     } else if (all_same()) {
-      *stream << "curr=" << newest_ << " count=" << count_;
+      *stream << "count=" << count_ << " curr=" << newest_;
       if (count_ > 1) *stream << "(all same)";
     } else {
-      *stream << "curr=" << newest_ << " count=" << count_ << " min=" << min_
-              << " max=" << max_ << " avg=" << avg()
-              << " std=" << std_deviation();
+      *stream << "count=" << count_ << " first=" << first_
+              << " curr=" << newest_ << " min=" << min_ << " max=" << max_
+              << " avg=" << avg() << " std=" << std_deviation();
     }
   }
 
@@ -93,6 +99,7 @@ class Stat {
   }
 
  private:
+  ValueType first_ = 0;
   ValueType newest_ = 0;
   ValueType max_ = std::numeric_limits<ValueType>::min();
   ValueType min_ = std::numeric_limits<ValueType>::max();
@@ -121,93 +128,61 @@ class StatSummarizer {
   // Prints the output tensor sizes and types for each node.
   void PrintOutputs() const;
 
-  // Summarizes all nodes' stat in the order of top durations.
-  // Will stop printing if either cdf_cutoff_ratio or num_max_nodes_to_print
-  // is hit.
-  std::string GetTimingStatsByTopDurations(
-      double cdf_cutoff_ratio = 1.0,
-      int num_max_nodes_to_print = std::numeric_limits<int>::max()) const;
-
-  // Summarizes all nodes' timing stat in the order of nodes getting executed.
-  std::string GetTimingStatsByRunOrder() const;
-
-  // Summarizes all timing stats in the order of node definitions in the graph.
-  std::string GetTimingStatsByOrderOfNodeDefinitions() const;
-
-  // Summarizes all nodes' stat in the order of memory used (high -> low).
-  // Will stop printing if either cdf_cutoff_ratio or num_max_nodes_to_print
-  // is hit.
-  std::string GetMemoryStatsByUsage(
-      double cdf_cutoff_ratio = 1.0,
-      int num_max_nodes_to_print = std::numeric_limits<int>::max()) const;
-
-  // Summarizes all nodes' memory stat in the order of nodes getting executed.
-  std::string GetMemoryStatsByRunOrder() const;
-
-  // Summarizes all memory stats in the order of node definitions in the graph.
-  std::string GetMemoryStatsByOrderOfNodeDefinitions() const;
-
-  // Deprecated, use GetTimingStatsByOrderOfNodeDefinitions instead
-  std::string GetStatsByOrderOfNodeDefinitions() const;
-
-  // Deprecated, use GetTimingStatsByRunOrder instead
-  std::string GetStatsByRunOrder() const;
-
-  // Deprecated, use GetTimingStatsByTopDurations instead
-  std::string GetStatsByTopDurations(
-      double cdf_cutoff_ratio = 1.0,
-      int num_max_nodes_to_print = std::numeric_limits<int>::max()) const;
-
   std::string GetStatsByNodeType() const;
 
   void Reset() {
-    run_total_micros_.Reset();
+    run_total_us_.Reset();
     memory_.Reset();
-    timing_details_.clear();
-    memory_details_.clear();
+    details_.clear();
   }
 
   // Returns number of runs.
-  int num_runs() const { return run_total_micros_.count(); }
+  int num_runs() const { return run_total_us_.count(); }
 
   // Returns stats of total microseconds spent by all nodes in each run.
-  const Stat<int64>& run_total_us() const { return run_total_micros_; }
+  const Stat<int64>& run_total_us() const { return run_total_us_; }
 
  private:
   struct Detail {
-    int64 first_start;
-    int64 first_rel_end;
-    int64 total;
+    string name;
+    string type;
+    int64 run_order;
+    Stat<int64> start_us;
+    Stat<int64> rel_end_us;
+    Stat<int64> mem_used;
     std::vector<TensorDescription> outputs;
   };
 
   enum SortingMetric {
-    BY_TOTAL,
+    BY_NAME,
+    BY_DEFINITION_ORDER,
     BY_RUN_ORDER,
+    BY_TIME,
+    BY_MEMORY,
+    BY_TYPE,
   };
 
   void Validate(const Detail* detail, const NodeExecStats& ns) const;
 
-  // Summarizes all nodes' stat in the order of node names defined in the graph.
-  std::string GetStatsByOrderOfNodeDefinitions(bool use_memory) const;
+  void OrderNodesByMetric(SortingMetric sorting_metric,
+                          std::vector<const Detail*>* details) const;
+
+  std::string GetStatsByMetric(const string& title,
+                               SortingMetric sorting_metric,
+                               int num_stats) const;
 
-  std::string GetStatsBySorting(SortingMetric sorting_metric,
-                                double cdf_cutoff_ratio,
-                                int num_max_nodes_to_print,
-                                bool use_memory) 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;
 
-  std::string HeaderString() const;
-  std::string ColumnString(const std::string& name, const Detail& detail,
-                           int64 cumulative_stat_on_node,
-                           Stat<int64> stat) const;
   std::string ShortSummary() const;
 
-  int64 first_node_start_micros_;
-  Stat<int64> run_total_micros_;
+  Stat<int64> run_total_us_;
   Stat<int64> memory_;
+
   std::vector<string> nodes_in_def_order_;
-  std::map<std::string, Detail> timing_details_;
-  std::map<std::string, Detail> memory_details_;
+  std::map<std::string, Detail> details_;
   std::map<string, string> node_types_;
 };