profiler C++ API.

PiperOrigin-RevId: 181397308

4 files changed, 367 insertions, 0 deletions

diff --git a/tensorflow/cc/profiler/BUILD b/tensorflow/cc/profiler/BUILD
new file mode 100644
index 0000000000..00799526fc
--- /dev/null
+++ b/tensorflow/cc/profiler/BUILD
@@ -0,0 +1,36 @@
+package(
+    default_visibility = ["//visibility:public"],
+)
+
+licenses(["notice"])  # Apache 2.0
+
+load("//tensorflow:tensorflow.bzl", "tf_cuda_cc_test")
+
+tf_cuda_cc_test(
+    name = "profiler_test",
+    srcs = ["profiler_test.cc"],
+    deps = [
+        ":profiler",
+        "//tensorflow/cc:cc_ops",
+        "//tensorflow/core:core_cpu",
+        "//tensorflow/core:framework",
+        "//tensorflow/core:lib",
+        "//tensorflow/core:protos_all_cc",
+        "//tensorflow/core:tensorflow",
+        "//tensorflow/core:test",
+        "//tensorflow/core:test_main",
+    ],
+)
+
+cc_library(
+    name = "profiler",
+    srcs = ["profiler.cc"],
+    hdrs = ["profiler.h"],
+    deps = [
+        "//tensorflow/core:lib",
+        "//tensorflow/core:protos_all_cc",
+        "//tensorflow/core/profiler:protos_all_cc",
+        "//tensorflow/core/profiler:tfprof_options",
+        "//tensorflow/core/profiler/internal:tfprof_stats",
+    ],
+)
diff --git a/tensorflow/cc/profiler/profiler.cc b/tensorflow/cc/profiler/profiler.cc
new file mode 100644
index 0000000000..3e55bac73e
--- /dev/null
+++ b/tensorflow/cc/profiler/profiler.cc
@@ -0,0 +1,57 @@
+/* Copyright 2017 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/cc/profiler/profiler.h"
+
+namespace tensorflow {
+namespace tfprof {
+
+Profiler::Profiler(const GraphDef& graph) {
+  std::unique_ptr<GraphDef> graph_ptr(new GraphDef());
+  *graph_ptr = graph;
+  stats_.reset(new TFStats(std::move(graph_ptr), nullptr, nullptr, nullptr));
+}
+
+void Profiler::AddStep(int64 step, const RunMetadata& run_meta) {
+  std::unique_ptr<RunMetadata> run_meta_ptr(new RunMetadata());
+  *run_meta_ptr = run_meta;
+  stats_->AddRunMeta(step, std::move(run_meta_ptr));
+}
+
+GraphNodeProto Profiler::ProfileGraph(const Options& options) {
+  stats_->BuildView(kCmds[1]);
+  return stats_->ShowGraphNode(kCmds[1], options);
+}
+
+GraphNodeProto Profiler::ProfileNameScope(const Options& options) {
+  stats_->BuildView(kCmds[0]);
+  return stats_->ShowGraphNode(kCmds[0], options);
+}
+
+MultiGraphNodeProto Profiler::ProfileOperations(const Options& options) {
+  stats_->BuildView(kCmds[3]);
+  return stats_->ShowMultiGraphNode(kCmds[3], options);
+}
+
+Status Profiler::SerializeToString(string* content) {
+  if (!content) {
+    return Status(error::Code::INVALID_ARGUMENT,
+                  "Cannot use null string pointer for SerializeToString.");
+  }
+  stats_->SerializeToString(content);
+  return Status::OK();
+}
+
+}  // namespace tfprof
+}  // namespace tensorflow
diff --git a/tensorflow/cc/profiler/profiler.h b/tensorflow/cc/profiler/profiler.h
new file mode 100644
index 0000000000..e1ce315d3c
--- /dev/null
+++ b/tensorflow/cc/profiler/profiler.h
@@ -0,0 +1,97 @@
+/* Copyright 2017 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 THIRD_PARTY_TENSORFLOW_CC_PROFILER_PROFILER_H_
+#define THIRD_PARTY_TENSORFLOW_CC_PROFILER_PROFILER_H_
+
+#include "tensorflow/core/framework/graph.pb.h"
+#include "tensorflow/core/lib/core/status.h"
+#include "tensorflow/core/profiler/internal/tfprof_stats.h"
+#include "tensorflow/core/profiler/tfprof_options.h"
+#include "tensorflow/core/profiler/tfprof_output.pb.h"
+
+namespace tensorflow {
+namespace tfprof {
+
+/// @addtogroup core
+/// @{
+
+/// A `Profiler` object lets the caller profile the execution of a graph.
+///
+/// Example:
+///     // First build a graph and run tracing.
+///     Scope root = Scope::NewRootScope();
+///     auto a = Placeholder(root, DT_INT32);
+///     auto c = Add(root, a, {41});
+///
+///     ClientSession session(root);
+///     std::vector<Tensor> outputs;
+///     RunOptions run_options;
+///     run_options.set_trace_level(RunOptions::FULL_TRACE);
+///     RunMetadata run_meta;
+///     Status s = session.Run(run_options, { {a, {1}} }, {c}, &outputs,
+///                            &run_meta);
+///     if (!s.ok()) { ... }
+///
+///     // Then create profiler to do profiling.
+///     GraphDef graph;
+///     root.ToGraphDef(&graph);
+///     Profiler profiler(graph);
+///     profiler.AddStep(0, run_meta);
+///     Options opts = ...  // TODO(xpan): Support option building API.
+///     MultiGraphNodeProto r = profiler.ProfileOperations(opts);
+///
+class Profiler {
+ public:
+  /// `graph` is the model's GraphDef.
+  Profiler(const GraphDef& graph);
+
+  /// Adds tracing information `run_meta` to profiler. A `run_meta` is
+  /// generated by a TensorFlow session run call. `step` is the key
+  /// to the `run_meta`. When calling ProfileXXX methods, caller can specify
+  /// `step` in `options` to seletively profile the corresponding `run_meta`.
+  /// Multiple different `run_meta` can be keyed by the same `step` in order
+  /// to group them together.
+  void AddStep(int64 step, const RunMetadata& run_meta);
+
+  /// Profiles the model by organizing nodes in graph structure.
+  /// Each node is an op and the nodes are contected by the op inputs/outputs.
+  GraphNodeProto ProfileGraph(const Options& options);
+
+  /// Profiles the model by organizing nodes in name scope structure.
+  /// Each node is an op, and nodes are organized by the ops' name
+  /// scope, similar to a filesystem tree.
+  /// E.g. /foo is the root of operation /foo/matmul_1 and foo/conv_2.
+  GraphNodeProto ProfileNameScope(const Options& options);
+
+  /// Profiles the model by organizing nodes by operation types.
+  /// Each node is an operation type (e.g. Conv2D or MatMul), containing all
+  /// ops belonging to that type in the model.
+  MultiGraphNodeProto ProfileOperations(const Options& options);
+
+  /// Serialize the profile content (ProfileProto) into a binary string,
+  /// User can write the string to file for offline analysis by
+  /// tfprof command-line tools or graphical user interface.
+  Status SerializeToString(string* content);
+
+ private:
+  std::unique_ptr<TFStats> stats_;
+};
+/// @}
+
+}  // namespace tfprof
+}  // namespace tensorflow
+
+#endif  // THIRD_PARTY_TENSORFLOW_CC_PROFILER_PROFILER_H_
diff --git a/tensorflow/cc/profiler/profiler_test.cc b/tensorflow/cc/profiler/profiler_test.cc
new file mode 100644
index 0000000000..280cd74827
--- /dev/null
+++ b/tensorflow/cc/profiler/profiler_test.cc
@@ -0,0 +1,177 @@
+/* Copyright 2017 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/platform/test.h"
+
+#include "tensorflow/cc/ops/standard_ops.h"
+#include "tensorflow/cc/profiler/profiler.h"
+#include "tensorflow/core/framework/graph.pb.h"
+#include "tensorflow/core/framework/tensor.h"
+#include "tensorflow/core/graph/default_device.h"
+#include "tensorflow/core/lib/io/path.h"
+#include "tensorflow/core/platform/env.h"
+#include "tensorflow/core/public/session.h"
+
+namespace tensorflow {
+namespace tfprof {
+
+class ProfilerTest : public ::testing::Test {
+ protected:
+  ProfilerTest() {}
+};
+
+GraphDef CreateGraphDef() {
+  Scope root = Scope::NewRootScope();
+
+  auto a = ops::Const<float>(root, {{3, 2}, {-1, 0}});
+
+  auto x = ops::Const(root.WithOpName("x"), {{1.f}, {1.f}});
+
+  auto y = ops::MatMul(root.WithOpName("y"), a, x);
+
+  auto y2 = ops::Square(root, y);
+
+  auto y2_sum = ops::Sum(root, y2, 0);
+
+  auto y_norm = ops::Sqrt(root, y2_sum);
+
+  auto y_div = ops::Div(root.WithOpName("y_normalized"), y, y_norm);
+
+  GraphDef def;
+  TF_CHECK_OK(root.ToGraphDef(&def));
+
+  return def;
+}
+
+Options Default() {
+  Options opts(1000,       /* max_depth */
+               0,          /* min_bytes */
+               0,          /* min_peak_bytes */
+               0,          /* min_residual_bytes */
+               0,          /* min_output_bytes */
+               0,          /* min_micros */
+               0,          /* min_accelerator_micros */
+               0,          /* min_cpu_micros */
+               0,          /* min_params */
+               0,          /* min_float_ops */
+               0,          /* min_occurrence */
+               0,          /* step */
+               "name",     /* order_by */
+               {".*"},     /* account_type_regexes */
+               {".*"},     /* start_name_regexes */
+               {},         /* trim_name_regexes */
+               {".*"}, {}, /* hide_name_regexes */
+               false,      /* account_displayed_op_only */
+               {"micros"}, /* select */
+               {"none"},   /* output_type */
+               {});
+  return opts;
+}
+
+template <typename T>
+const T* ExtractNode(const T& pb, const string& name) {
+  if (pb.name() == name) {
+    return &pb;
+  }
+  for (const T& c : pb.children()) {
+    const T* ret = ExtractNode(c, name);
+    if (ret) return ret;
+  }
+  return nullptr;
+}
+
+TEST_F(ProfilerTest, Basics) {
+  SessionOptions options;
+  options.config.set_allow_soft_placement(true);
+  std::unique_ptr<Session> session(NewSession(options));
+  GraphDef def = CreateGraphDef();
+  if (options.target.empty()) {
+    graph::SetDefaultDevice("/gpu:0", &def);
+  }
+
+  TF_CHECK_OK(session->Create(def));
+
+  Tensor x(DT_FLOAT, TensorShape({2, 1}));
+  auto x_flat = x.flat<float>();
+  x_flat.setRandom();
+  Eigen::Tensor<float, 0, Eigen::RowMajor> inv_norm =
+      x_flat.square().sum().sqrt().inverse();
+  x_flat = x_flat * inv_norm();
+
+  std::vector<Tensor> outputs;
+  RunOptions run_options;
+  run_options.set_trace_level(RunOptions::FULL_TRACE);
+  RunMetadata run_metadata;
+  outputs.clear();
+
+  Profiler profiler(def);
+  for (int i = 0; i < 2; ++i) {
+    TF_CHECK_OK(session->Run(run_options, {{"x", x}}, {"y:0", "y_normalized:0"},
+                             {}, &outputs, &run_metadata));
+    profiler.AddStep(i, run_metadata);
+    CHECK_EQ(size_t{2}, outputs.size());
+  }
+
+  std::vector<DeviceAttributes> resp;
+  TF_CHECK_OK(session->ListDevices(&resp));
+  bool has_gpu = false;
+  for (const auto& dev : resp) {
+    if (dev.device_type() == "GPU") {
+      has_gpu = true;
+    }
+  }
+
+  GraphNodeProto ret = profiler.ProfileNameScope(Default());
+  const GraphNodeProto* matmul = ExtractNode(ret, "y");
+  EXPECT_TRUE(matmul);
+  EXPECT_GT(matmul->exec_micros(), 0);
+  if (has_gpu) {
+    EXPECT_GT(matmul->accelerator_exec_micros(), 0);
+  } else {
+    EXPECT_EQ(matmul->accelerator_exec_micros(), 0);
+  }
+  const GraphNodeProto* square = ExtractNode(ret, "Square");
+  EXPECT_TRUE(square);
+  EXPECT_GT(square->exec_micros(), 0);
+  if (has_gpu) {
+    EXPECT_GT(square->accelerator_exec_micros(), 0);
+  } else {
+    EXPECT_EQ(square->accelerator_exec_micros(), 0);
+  }
+
+  Options opts2 = Default();
+  opts2.output_type = "timeline";
+  string timeline_file = io::JoinPath(testing::TmpDir(), "timeline");
+  opts2.output_options["outfile"] = timeline_file;
+  GraphNodeProto ret2 = profiler.ProfileGraph(opts2);
+  string s;
+  TF_CHECK_OK(ReadFileToString(Env::Default(), timeline_file + "_0", &s));
+  EXPECT_TRUE(s.find("Square") != s.npos);
+
+  MultiGraphNodeProto ret3 = profiler.ProfileOperations(Default());
+  const MultiGraphNodeProto* matmul2 = ExtractNode(ret3, "MatMul");
+  EXPECT_TRUE(matmul2);
+  EXPECT_GT(matmul2->exec_micros(), 0);
+  if (has_gpu) {
+    EXPECT_GT(matmul2->accelerator_exec_micros(), 0);
+  } else {
+    EXPECT_EQ(matmul2->accelerator_exec_micros(), 0);
+  }
+
+  TF_CHECK_OK(session->Close());
+}
+
+}  // namespace tfprof
+}  // namespace tensorflow