Merge pull request #10206 from JiayiFeng/blocking_queue_for_reader

Blocking queue for reader

Author: JiayiFeng

paddle/fluid/operators/reader/CMakeLists.txt

@@ -23,5 +23,7 @@ reader_library(create_recordio_file_reader_op SRCS create_recordio_file_reader_o
 reader_library(create_double_buffer_reader_op SRCS create_double_buffer_reader_op.cc)
 reader_library(create_multi_pass_reader_op SRCS create_multi_pass_reader_op.cc)
 reader_library(create_threaded_reader_op SRCS create_threaded_reader_op.cc)
+
+cc_test(reader_blocking_queue_test SRCS reader_blocking_queue_test.cc)
 # Export local libraries to parent
 set(READER_LIBRARY ${LOCAL_READER_LIBS} PARENT_SCOPE)

paddle/fluid/operators/reader/blocking_queue.h

@@ -0,0 +1,112 @@
+//   Copyright (c) 2018 PaddlePaddle 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.
+
+#pragma once
+
+#include <condition_variable>  // NOLINT
+#include <deque>
+
+#include "paddle/fluid/platform/enforce.h"
+
+namespace paddle {
+namespace operators {
+namespace reader {
+
+template <typename T>
+class BlockingQueue {
+  // BlockingQueue is for buffered reading and is supposed to use only the
+  // reader package. It is true that we could and we should have been using
+  // framework::Channel, but which has currently a deadlock bug. BlockingQueue
+  // is a workaround and a simplified version of framework::Channel as it
+  // doesn't support GPU and it implements on buffered blocking queue.
+ public:
+  explicit BlockingQueue(size_t capacity)
+      : capacity_(capacity), closed_(false) {
+    PADDLE_ENFORCE_GT(
+        capacity_, 0,
+        "The capacity of a reader::BlockingQueue must be greater than 0.");
+  }
+
+  bool Send(const T& elem) {
+    std::unique_lock<std::mutex> lock(mutex_);
+    send_cv_.wait(lock, [&] { return queue_.size() < capacity_ || closed_; });
+    if (closed_) {
+      VLOG(5)
+          << "WARNING: Sending an element to a closed reader::BlokcingQueue.";
+      return false;
+    }
+    PADDLE_ENFORCE_LT(queue_.size(), capacity_);
+    queue_.push_back(elem);
+    receive_cv_.notify_one();
+    return true;
+  }
+
+  bool Send(T&& elem) {
+    std::unique_lock<std::mutex> lock(mutex_);
+    send_cv_.wait(lock, [&] { return queue_.size() < capacity_ || closed_; });
+    if (closed_) {
+      VLOG(5)
+          << "WARNING: Sending an element to a closed reader::BlokcingQueue.";
+      return false;
+    }
+    PADDLE_ENFORCE_LT(queue_.size(), capacity_);
+    queue_.emplace_back(std::move(elem));
+    receive_cv_.notify_one();
+    return true;
+  }
+
+  bool Receive(T* elem) {
+    std::unique_lock<std::mutex> lock(mutex_);
+    receive_cv_.wait(lock, [&] { return !queue_.empty() || closed_; });
+    if (!queue_.empty()) {
+      PADDLE_ENFORCE_NOT_NULL(elem);
+      *elem = queue_.front();
+      queue_.pop_front();
+      send_cv_.notify_one();
+      return true;
+    } else {
+      PADDLE_ENFORCE(closed_);
+      return false;
+    }
+  }
+
+  void Close() {
+    std::lock_guard<std::mutex> lock(mutex_);
+    closed_ = true;
+    send_cv_.notify_all();
+    receive_cv_.notify_all();
+  }
+
+  bool IsClosed() {
+    std::lock_guard<std::mutex> lock(mutex_);
+    return closed_;
+  }
+
+  size_t Cap() {
+    std::lock_guard<std::mutex> lock(mutex_);
+    return capacity_;
+  }
+
+ private:
+  size_t capacity_;
+  bool closed_;
+  std::deque<T> queue_;
+
+  std::mutex mutex_;
+  std::condition_variable receive_cv_;
+  std::condition_variable send_cv_;
+};
+}  // namespace reader
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/reader/create_double_buffer_reader_op.cc

@@ -14,7 +14,7 @@
 
 #include <thread>  // NOLINT
 
-#include "paddle/fluid/framework/channel.h"
+#include "paddle/fluid/operators/reader/blocking_queue.h"
 #include "paddle/fluid/operators/reader/reader_op_registry.h"
 
 namespace paddle {
@@ -23,13 +23,13 @@ namespace reader {
 
 // 'Double buffer' means we shall maintain two batches of input data at the same
 // time. So the kCacheSize shoul be at least 2.
-static constexpr size_t kCacheSize = 2;
+static constexpr size_t kCacheSize = 3;
 // There will be two bacthes out of the channel during training:
 // 1. the one waiting to be sent to the channel
 // 2. the one just be received from the channel, which is also being used by
 // subsequent operators.
 // So the channel size should be kChacheSize - 2
-static constexpr size_t kChannelSize = 0;  // kCacheSize - 2
+static constexpr size_t kChannelSize = 1;  // kCacheSize - 2
 
 class DoubleBufferReader : public framework::DecoratedReader {
  public:
@@ -55,10 +55,8 @@ class DoubleBufferReader : public framework::DecoratedReader {
   ~DoubleBufferReader() { EndPrefetcher(); }
 
  private:
-  bool HasNext() const;
-
   void StartPrefetcher() {
-    channel_ = framework::MakeChannel<size_t>(kChannelSize);
+    channel_ = new reader::BlockingQueue<size_t>(kChannelSize);
     prefetcher_ = std::thread([this] { PrefetchThreadFunc(); });
   }
 
@@ -74,7 +72,7 @@ class DoubleBufferReader : public framework::DecoratedReader {
   void PrefetchThreadFunc();
 
   std::thread prefetcher_;
-  framework::Channel<size_t>* channel_;
+  reader::BlockingQueue<size_t>* channel_;
   platform::Place place_;
   std::vector<std::vector<framework::LoDTensor>> cpu_tensor_cache_;
   std::vector<std::vector<framework::LoDTensor>> gpu_tensor_cache_;
@@ -139,17 +137,16 @@ class CreateDoubleBufferReaderOpMaker : public DecoratedReaderMakerBase {
 };
 
 void DoubleBufferReader::ReadNext(std::vector<framework::LoDTensor>* out) {
-  out->clear();
-  if (HasNext()) {
-    size_t cached_tensor_id;
-    channel_->Receive(&cached_tensor_id);
+  size_t cached_tensor_id;
+  if (channel_->Receive(&cached_tensor_id)) {
     if (platform::is_gpu_place(place_)) {
       *out = gpu_tensor_cache_[cached_tensor_id];
-      ctxs_[cached_tensor_id]->Wait();
     } else {
       // CPU place
       *out = cpu_tensor_cache_[cached_tensor_id];
     }
+  } else {
+    out->clear();
   }
 }
 
@@ -159,12 +156,6 @@ void DoubleBufferReader::ReInit() {
   StartPrefetcher();
 }
 
-bool DoubleBufferReader::HasNext() const {
-  while (!channel_->IsClosed() && !channel_->CanReceive()) {
-  }
-  return channel_->CanReceive();
-}
-
 void DoubleBufferReader::PrefetchThreadFunc() {
   VLOG(5) << "A new prefetch thread starts.";
   size_t cached_tensor_id = 0;
@@ -185,10 +176,7 @@ void DoubleBufferReader::PrefetchThreadFunc() {
         gpu_batch[i].set_lod(cpu_batch[i].lod());
       }
     }
-    try {
-      size_t tmp = cached_tensor_id;
-      channel_->Send(&tmp);
-    } catch (paddle::platform::EnforceNotMet e) {
+    if (!channel_->Send(cached_tensor_id)) {
       VLOG(5) << "WARNING: The double buffer channel has been closed. The "
                  "prefetch thread will terminate.";
       break;

paddle/fluid/operators/reader/open_files_op.cc

@@ -14,7 +14,7 @@
 
 #include <thread>  // NOLINT
 
-#include "paddle/fluid/framework/channel.h"
+#include "paddle/fluid/operators/reader/blocking_queue.h"
 #include "paddle/fluid/operators/reader/reader_op_registry.h"
 
 namespace paddle {
@@ -37,7 +37,6 @@ class MultiFileReader : public framework::ReaderBase {
   ~MultiFileReader() { EndScheduler(); }
 
  private:
-  bool HasNext();
   void StartNewScheduler();
   void EndScheduler();
   void ScheduleThreadFunc();
@@ -48,15 +47,14 @@ class MultiFileReader : public framework::ReaderBase {
   std::thread scheduler_;
   std::vector<std::thread> prefetchers_;
   size_t buffer_size_;
-  framework::Channel<size_t>* waiting_file_idx_;
-  framework::Channel<size_t>* available_thread_idx_;
-  framework::Channel<std::vector<framework::LoDTensor>>* buffer_;
+  reader::BlockingQueue<size_t>* waiting_file_idx_;
+  reader::BlockingQueue<size_t>* available_thread_idx_;
+  reader::BlockingQueue<std::vector<framework::LoDTensor>>* buffer_;
 };
 
 void MultiFileReader::ReadNext(std::vector<framework::LoDTensor>* out) {
-  out->clear();
-  if (HasNext()) {
-    buffer_->Receive(out);
+  if (!buffer_->Receive(out)) {
+    out->clear();
   }
 }
 
@@ -65,25 +63,19 @@ void MultiFileReader::ReInit() {
   StartNewScheduler();
 }
 
-bool MultiFileReader::HasNext() {
-  while (!buffer_->IsClosed() && !buffer_->CanReceive()) {
-  }
-  return buffer_->CanReceive();
-}
-
 void MultiFileReader::StartNewScheduler() {
   size_t thread_num = prefetchers_.size();
-  waiting_file_idx_ = framework::MakeChannel<size_t>(file_names_.size());
-  available_thread_idx_ = framework::MakeChannel<size_t>(thread_num);
-  buffer_ =
-      framework::MakeChannel<std::vector<framework::LoDTensor>>(buffer_size_);
+  waiting_file_idx_ = new reader::BlockingQueue<size_t>(file_names_.size());
+  available_thread_idx_ = new reader::BlockingQueue<size_t>(thread_num);
+  buffer_ = new reader::BlockingQueue<std::vector<framework::LoDTensor>>(
+      buffer_size_);
 
   for (size_t i = 0; i < file_names_.size(); ++i) {
-    waiting_file_idx_->Send(&i);
+    waiting_file_idx_->Send(i);
   }
   waiting_file_idx_->Close();
   for (size_t i = 0; i < thread_num; ++i) {
-    available_thread_idx_->Send(&i);
+    available_thread_idx_->Send(i);
   }
 
   scheduler_ = std::thread([this] { ScheduleThreadFunc(); });
@@ -149,7 +141,7 @@ void MultiFileReader::PrefetchThreadFunc(std::string file_name,
       break;
     }
     try {
-      buffer_->Send(&ins);
+      buffer_->Send(std::move(ins));
     } catch (paddle::platform::EnforceNotMet e) {
       VLOG(5) << "WARNING: The buffer channel has been closed. The prefetch "
                  "thread of file '"
@@ -158,9 +150,7 @@ void MultiFileReader::PrefetchThreadFunc(std::string file_name,
     }
   }
 
-  try {
-    available_thread_idx_->Send(&thread_idx);
-  } catch (paddle::platform::EnforceNotMet e) {
+  if (!available_thread_idx_->Send(thread_idx)) {
     VLOG(5) << "WARNING: The available_thread_idx_ channel has been closed. "
                "Fail to send thread_idx.";
   }