diff --git a/paddle/platform/CMakeLists.txt b/paddle/platform/CMakeLists.txt
index f0a0ea70a0aa14e1db959e4e6ace2a44363d0c35..8c4803b9739bb54cae89de62468a47631a5dde94 100644
--- a/paddle/platform/CMakeLists.txt
+++ b/paddle/platform/CMakeLists.txt
@@ -30,3 +30,6 @@ nv_test(device_context_test SRCS device_context_test.cu DEPS device_context gpu_
 nv_test(cudnn_helper_test SRCS cudnn_helper_test.cc DEPS dynload_cuda)
 nv_test(transform_test SRCS transform_test.cu DEPS paddle_memory place device_context)
 nv_test(nccl_test SRCS nccl_test.cu DEPS dynload_cuda gpu_info device_context)
+
+cc_library(profiler SRCS profiler.cc DEPS device_context)
+cc_test(profiler_test SRCS profiler_test.cc DEPS profiler)
diff --git a/paddle/platform/profiler.cc b/paddle/platform/profiler.cc
new file mode 100644
index 0000000000000000000000000000000000000000..4e89e5c600bf7f6e23faf8d62bc7c72f7ee8ca7a
--- /dev/null
+++ b/paddle/platform/profiler.cc
@@ -0,0 +1,173 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+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 "paddle/platform/profiler.h"
+
+namespace paddle {
+namespace platform {
+
+// The profiler state, the initial value is ProfilerState::kDisabled
+static ProfilerState g_state = ProfilerState::kDisabled;
+// The thread local event list only can be accessed by the specific thread
+// The thread index of each thread
+static thread_local int32_t g_thread_id;
+// The g_next_thread_id is a global counter for threads, by the g_thread_id and
+// g_next_thread_id, we can know how many threads have created EventList.
+static uint32_t g_next_thread_id = 0;
+// The global mutex
+static std::mutex g_all_event_lists_mutex;
+// The total event lists of all threads
+static std::list<std::shared_ptr<EventList>> g_all_event_lists;
+// The thread local event list only can be accessed by the specific thread
+static thread_local std::shared_ptr<EventList> g_event_list;
+
+inline uint64_t GetTimeInNsec() {
+  using clock = std::conditional<std::chrono::high_resolution_clock::is_steady,
+                                 std::chrono::high_resolution_clock,
+                                 std::chrono::steady_clock>::type;
+  return std::chrono::duration_cast<std::chrono::nanoseconds>(
+             clock::now().time_since_epoch())
+      .count();
+}
+
+Event::Event(EventKind kind, std::string name, uint32_t thread_id,
+             DeviceContext* dev_ctx)
+    : kind_(kind),
+      name_(std::move(name)),
+      thread_id_(thread_id),
+      has_cuda_(false) {
+#ifdef PADDLE_WITH_CUDA
+  auto* cuda_dev_ctx = static_cast<const CUDADeviceContext*>(dev_ctx);
+  if (cuda_dev_ctx) {
+    PADDLE_ENFORCE(cudaGetDevice(&device_));
+    PADDLE_ENFORCE(cudaEventCreate(&event_));
+    auto stream = cuda_dev_ctx->stream();
+    PADDLE_ENFORCE(cudaEventRecord(event_, stream));
+    has_cuda_ = true;
+  }
+#endif
+  cpu_ns_ = GetTimeInNsec();
+}
+
+std::string Event::kind() const {
+  switch (kind_) {
+    case EventKind::kMark:
+      return "mark";
+    case EventKind::kPushRange:
+      return "push";
+    case EventKind::kPopRange:
+      return "pop";
+  }
+  PADDLE_THROW("Unknown EventKind.");
+}
+
+double Event::CpuElapsedUs(const Event& e) const {
+  return (e.cpu_ns_ - cpu_ns_) / (1000.0);
+}
+
+double Event::CudaElapsedUs(const Event& e) const {
+#ifdef PADDLE_WITH_CUDA
+  PADDLE_ENFORCE(e.has_cuda() && has_cuda());
+  PADDLE_ENFORCE(e.device() == device());
+  PADDLE_ENFORCE(cudaEventSynchronize(event_));
+  PADDLE_ENFORCE(cudaEventSynchronize(e.event()));
+  float ms;
+  PADDLE_ENFORCE(cudaEventElapsedTime(&ms, event_, e.event()));
+  return ms * 1000.0;
+#else
+  PADDLE_THROW("CUDA is not enabled");
+#endif
+}
+
+#ifdef PADDLE_WITH_CUDA
+static void ForEachDevice(std::function<void(int)> func) {
+  auto original_device = GetCurrentDeviceId();
+  int count = GetCUDADeviceCount();
+  for (int i = 0; i < count; i++) {
+    SetDeviceId(i);
+    func(i);
+  }
+  SetDeviceId(original_device);
+}
+#endif
+
+inline EventList& GetEventList() {
+  if (!g_event_list) {
+    std::lock_guard<std::mutex> guard(g_all_event_lists_mutex);
+    g_event_list = std::make_shared<EventList>();
+    g_thread_id = g_next_thread_id++;
+    g_all_event_lists.emplace_front(g_event_list);
+  }
+  return *g_event_list;
+}
+
+void Mark(const std::string& name, DeviceContext* dev_ctx) {
+  GetEventList().Record(EventKind::kMark, std::move(name), g_thread_id,
+                        dev_ctx);
+}
+
+RecordEvent::RecordEvent(const std::string& name, DeviceContext* dev_ctx) {
+  if (g_state == ProfilerState::kDisabled) return;
+  dev_ctx_ = dev_ctx;
+  GetEventList().Record(EventKind::kPushRange, std::move(name), g_thread_id,
+                        dev_ctx_);
+}
+
+RecordEvent::~RecordEvent() {
+  if (g_state == ProfilerState::kDisabled) return;
+  GetEventList().Record(EventKind::kPopRange, std::string(), g_thread_id,
+                        dev_ctx_);
+}
+
+void EnableProfiler(ProfilerState state) {
+  PADDLE_ENFORCE(state != ProfilerState::kDisabled,
+                 "Can't enbale profling, since the input state is ",
+                 "ProfilerState::kDisabled");
+  PADDLE_ENFORCE(g_state == ProfilerState::kDisabled,
+                 "The profiling state should be disabled when calling ",
+                 "EnableProfiler.");
+  g_state = state;
+#ifdef PADDLE_WITH_CUDA
+  if (g_state == ProfilerState::kCUDA) {
+    // Generate some dummy evenets first to reduce the startup overhead.
+    for (int i = 0; i < 5; i++) {
+      ForEachDevice([](int d) {
+        DeviceContext* dev_ctx = new CUDADeviceContext(CUDAPlace(d));
+        Mark("_cuda_startup_", dev_ctx);
+        dev_ctx->Wait();
+      });
+    }
+  }
+#endif
+  // Mark the profiling start.
+  Mark("_start_profiler_", nullptr);
+}
+
+std::vector<std::vector<Event>> DisableProfiler() {
+  PADDLE_ENFORCE(g_state != ProfilerState::kDisabled,
+                 "Can't disable profiling, since it's not starting.");
+  // Mark the profiling stop.
+  Mark("_stop_profiler_", nullptr);
+  g_state = ProfilerState::kDisabled;
+  std::vector<std::vector<Event>> result;
+  std::lock_guard<std::mutex> guard(g_all_event_lists_mutex);
+  for (auto it = g_all_event_lists.begin(); it != g_all_event_lists.end();
+       ++it) {
+    result.emplace_back((*it)->Reduce());
+  }
+  return result;
+}
+
+}  // namespace platform
+}  // namespace paddle
diff --git a/paddle/platform/profiler.h b/paddle/platform/profiler.h
new file mode 100644
index 0000000000000000000000000000000000000000..47104ea9d08dad373888c0af463d2b1dbe72a269
--- /dev/null
+++ b/paddle/platform/profiler.h
@@ -0,0 +1,114 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+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 <forward_list>
+#include <list>
+#include <mutex>
+#include <vector>
+#include "paddle/platform/device_context.h"
+
+namespace paddle {
+namespace platform {
+
+enum EventKind { kMark, kPushRange, kPopRange };
+
+class Event {
+ public:
+  // The DeviceContext is used to get the cuda stream.
+  // If CPU profiling mode, can pass nullptr.
+  Event(EventKind kind, std::string name, uint32_t thread_id,
+        DeviceContext* dev_ctx);
+
+  std::string kind() const;
+  std::string name() const { return name_; }
+  bool has_cuda() const { return has_cuda_; }
+
+#ifdef PADDLE_WITH_CUDA
+  cudaEvent_t event() const { return event_; }
+  int device() const { return device_; }
+#endif
+
+  double CpuElapsedUs(const Event& e) const;
+  double CudaElapsedUs(const Event& e) const;
+
+ private:
+  EventKind kind_;
+  std::string name_;
+  uint32_t thread_id_;
+  int64_t cpu_ns_;
+  bool has_cuda_;
+#ifdef PADDLE_WITH_CUDA
+  cudaEvent_t event_ = nullptr;
+  int device_ = -1;
+#endif
+};
+
+struct EventList {
+  constexpr static size_t kMB = 1024 * 1024;
+  constexpr static size_t kEventBlockSize = 16 * kMB;
+  constexpr static size_t kEventSize = sizeof(Event);
+  constexpr static size_t kEventAlign = alignof(Event);
+  constexpr static size_t kNumBlock =
+      kEventBlockSize /
+      ((kEventSize + kEventAlign - 1) / kEventAlign * kEventAlign);
+
+  template <typename... Args>
+  void Record(Args&&... args) {
+    if (event_blocks.empty() || event_blocks.front().size() == kNumBlock) {
+      event_blocks.emplace_front();
+      event_blocks.front().reserve(kNumBlock);
+    }
+    event_blocks.front().emplace_back(std::forward<Args>(args)...);
+  }
+
+  std::vector<Event> Reduce() {
+    std::vector<Event> result;
+    for (auto& block : event_blocks) {
+      result.insert(result.begin(), std::make_move_iterator(block.begin()),
+                    std::make_move_iterator(block.end()));
+    }
+    event_blocks.clear();
+    return result;
+  }
+
+  std::forward_list<std::vector<Event>> event_blocks;
+};
+
+enum ProfilerState {
+  kDisabled,  // disabled state
+  kCPU,       // CPU profiling state
+  kCUDA,      // GPU profiling state
+};
+
+void Mark(const std::string& name, DeviceContext* dev_ctx);
+
+struct RecordEvent {
+  explicit RecordEvent(const std::string& name, DeviceContext* dev_ctx);
+
+  ~RecordEvent();
+
+  // The device context is used by Event to get the current cuda stream.
+  DeviceContext* dev_ctx_;
+};
+
+// Enable the profiling function.
+void EnableProfiler(ProfilerState state);
+
+// Return the event list of all threads. Asummed the returned value calls
+// event_lists, event_lists[i][j] represents the j-th Event of i-th thread.
+std::vector<std::vector<Event>> DisableProfiler();
+
+}  // namespace platform
+}  // namespace paddle
diff --git a/paddle/platform/profiler_test.cc b/paddle/platform/profiler_test.cc
new file mode 100644
index 0000000000000000000000000000000000000000..47cf7be146121c54300df928fe329268ed975373
--- /dev/null
+++ b/paddle/platform/profiler_test.cc
@@ -0,0 +1,98 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+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 "paddle/platform/profiler.h"
+#include "gtest/gtest.h"
+
+TEST(Event, CpuElapsedTime) {
+  using paddle::platform::Event;
+  using paddle::platform::EventKind;
+
+  Event start_event(EventKind::kPushRange, "test", 0, nullptr);
+  EXPECT_TRUE(start_event.has_cuda() == false);
+  int counter = 0;
+  while (counter != 1000) {
+    counter++;
+  }
+  Event stop_event(EventKind::kPopRange, "test", 0, nullptr);
+  EXPECT_GT(start_event.CpuElapsedUs(stop_event), 0);
+}
+
+#ifdef PADDLE_WITH_CUDA
+TEST(Event, CudaElapsedTime) {
+  using paddle::platform::DeviceContext;
+  using paddle::platform::CUDADeviceContext;
+  using paddle::platform::CUDAPlace;
+  using paddle::platform::Event;
+  using paddle::platform::EventKind;
+
+  DeviceContext* dev_ctx = new CUDADeviceContext(CUDAPlace(0));
+  Event start_event(EventKind::kPushRange, "test", 0, dev_ctx);
+  EXPECT_TRUE(start_event.has_cuda() == true);
+  int counter = 0;
+  while (counter != 1000) {
+    counter++;
+  }
+  Event stop_event(EventKind::kPopRange, "test", 0, dev_ctx);
+  EXPECT_GT(start_event.CudaElapsedUs(stop_event), 0);
+}
+#endif
+
+TEST(RecordEvent, RecordEvent) {
+  using paddle::platform::DeviceContext;
+  using paddle::platform::Event;
+  using paddle::platform::EventKind;
+  using paddle::platform::RecordEvent;
+  using paddle::platform::ProfilerState;
+
+  ProfilerState state = ProfilerState::kCPU;
+  DeviceContext* dev_ctx = nullptr;
+#ifdef PADDLE_WITH_CUDA
+  using paddle::platform::CUDADeviceContext;
+  using paddle::platform::CUDAPlace;
+  state = ProfilerState::kCUDA;
+  dev_ctx =
+      new paddle::platform::CUDADeviceContext(paddle::platform::CUDAPlace(0));
+#endif
+  EnableProfiler(state);
+
+  for (int i = 1; i < 5; ++i) {
+    std::string name = "op_" + std::to_string(i);
+    RecordEvent record_event(name, dev_ctx);
+    int counter = 1;
+    while (counter != i * 1000) counter++;
+  }
+  std::vector<std::vector<Event>> events = paddle::platform::DisableProfiler();
+  int cuda_startup_count = 0;
+  int start_profiler_count = 0;
+  int stop_profiler_count = 0;
+  for (size_t i = 0; i < events.size(); ++i) {
+    for (size_t j = 0; j < events[i].size(); ++j) {
+      if (events[i][j].name() == "_cuda_startup_") ++cuda_startup_count;
+      if (events[i][j].name() == "_start_profiler_") ++start_profiler_count;
+      if (events[i][j].name() == "_stop_profiler_") ++stop_profiler_count;
+      if (events[i][j].name() == "push") {
+        EXPECT_EQ(events[i][j + 1].name(), "pop");
+#ifdef PADDLE_WITH_CUDA
+        EXPECT_GT(events[i][j].CudaElapsedUs(events[i][j + 1]), 0);
+#else
+        EXPECT_GT(events[i][j].CpuElapsedUs(events[i][j + 1]), 0);
+#endif
+      }
+    }
+  }
+  EXPECT_EQ(cuda_startup_count % 5, 0);
+  EXPECT_EQ(start_profiler_count, 1);
+  EXPECT_EQ(stop_profiler_count, 1);
+}