A few profiler improvements.

1. Op view proto copy uses too much memory and time, optimized.
2. Add a hint to use "bazel-bin" instead of "bazel run"
3. Make proto string parsing explicit (seems no longer throwing error)

PiperOrigin-RevId: 175745677

3 files changed, 19 insertions, 12 deletions

diff --git a/tensorflow/core/profiler/g3doc/profiler_ui.jpg b/tensorflow/core/profiler/g3doc/profiler_ui.jpg
index 36aa94502a..77346e61ae 100644
--- a/tensorflow/core/profiler/g3doc/profiler_ui.jpg
+++ b/tensorflow/core/profiler/g3doc/profiler_ui.jpg
diff --git a/tensorflow/core/profiler/internal/tfprof_op.cc b/tensorflow/core/profiler/internal/tfprof_op.cc
index c04b0ea0c6..5a8429d489 100644
--- a/tensorflow/core/profiler/internal/tfprof_op.cc
+++ b/tensorflow/core/profiler/internal/tfprof_op.cc
@@ -109,7 +109,6 @@ const ShowMultiNode* TFOp::ShowInternal(const Options& opts,
     fprintf(stderr, "Only 'code' view supports pprof output now.\n");
     return root_.get();
   }
-
   if (opts.output_type == kOutput[1] || opts.output_type == kOutput[2]) {
     root_->formatted_str = FormatNode(root_.get(), root_.get(), opts);
   }
@@ -130,7 +129,6 @@ const ShowMultiNode* TFOp::ShowInternal(const Options& opts,
     nodes.push_back(n.second.get());
   }
   nodes = SortNodes(nodes, opts);
-
   // pre keeps track of previous visited node.
   OpNode* pre = nullptr;
   std::vector<OpNode*> account_nodes;
@@ -166,10 +164,6 @@ const ShowMultiNode* TFOp::ShowInternal(const Options& opts,
       (*it)->AddSelfToTotalStats();
       if (pre) (*it)->AggregateTotalStats(pre);
     }
-    if (pre) {
-      (*it)->mutable_proto()->add_children()->MergeFrom(pre->proto());
-      pre->mutable_proto()->clear_children();
-    }
     pre = *it;
   }
   if (opts.account_displayed_op_only) {
@@ -178,11 +172,6 @@ const ShowMultiNode* TFOp::ShowInternal(const Options& opts,
       root_->AggregateTotalStats(pre);
     }
   }
-  if (pre) {
-    root_->mutable_proto()->add_children()->MergeFrom(pre->proto());
-    pre->mutable_proto()->clear_children();
-  }
-
   if (opts.output_type == kOutput[1] || opts.output_type == kOutput[2]) {
     string display_str = FormatLegend(opts);
     for (OpNode* node : show_nodes) {
@@ -192,6 +181,13 @@ const ShowMultiNode* TFOp::ShowInternal(const Options& opts,
     // TODO(xpan): Is it the right choice?
     root_->formatted_str = display_str;
   }
+  // Populate the chidren field.
+  auto* pre_pb = root_->mutable_proto();
+  for (auto& show_node : show_nodes) {
+    pre_pb->clear_children();
+    pre_pb->add_children()->Swap(show_node->mutable_proto());
+    pre_pb = pre_pb->mutable_children(0);
+  }
   return root_.get();
 }
 
diff --git a/tensorflow/core/profiler/profiler.cc b/tensorflow/core/profiler/profiler.cc
index a5e513aa21..b280242df1 100644
--- a/tensorflow/core/profiler/profiler.cc
+++ b/tensorflow/core/profiler/profiler.cc
@@ -266,7 +266,18 @@ int Run(int argc, char** argv) {
   linenoiseSetCompletionCallback(completion);
   linenoiseHistoryLoad(".tfprof_history.txt");
 
-  for (char* line = nullptr; (line = linenoise("tfprof> ")) != nullptr;) {
+  bool looped = false;
+  while (true) {
+    char* line = linenoise("tfprof> ");
+    if (line == nullptr) {
+      if (!looped) {
+        fprintf(stderr,
+                "Cannot start interative shell, "
+                "use 'bazel-bin' instead of 'bazel run'.\n");
+      }
+      break;
+    }
+    looped = true;
     string line_s = line;
     free(line);