Merge branch 'develop' of github.com:PaddlePaddle/Paddle into improve_concat_split_op

Author: Yancey0623

.gitignore

@@ -1,3 +1,9 @@
+paddle/operators/check_t.save
+paddle/operators/check_tensor.ls
+paddle/operators/tensor.save
+python/paddle/v2/fluid/tests/book/image_classification_resnet.inference.model/
+python/paddle/v2/fluid/tests/book/image_classification_vgg.inference.model/
+python/paddle/v2/fluid/tests/book/label_semantic_roles.inference.model/
 *.DS_Store
 build/
 build_doc/

cmake/cuda.cmake

@@ -181,7 +181,8 @@ elseif(CMAKE_BUILD_TYPE  STREQUAL "Release")
 elseif(CMAKE_BUILD_TYPE  STREQUAL "RelWithDebInfo")
     list(APPEND CUDA_NVCC_FLAGS  ${CMAKE_CXX_FLAGS_RELWITHDEBINFO})
 elseif(CMAKE_BUILD_TYPE  STREQUAL "MinSizeRel")
-    list(APPEND CUDA_NVCC_FLAGS  ${CMAKE_CXX_FLAGS_MINSIZEREL})
+    # nvcc 9 does not support -Os. Use Release flags instead
+    list(APPEND CUDA_NVCC_FLAGS  ${CMAKE_CXX_FLAGS_RELEASE})
 endif()
 
 mark_as_advanced(CUDA_BUILD_CUBIN CUDA_BUILD_EMULATION CUDA_VERBOSE_BUILD)

cmake/generic.cmake

@@ -179,20 +179,24 @@ function(cc_library TARGET_NAME)
   set(oneValueArgs "")
   set(multiValueArgs SRCS DEPS)
   cmake_parse_arguments(cc_library "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
-  if (cc_library_SRCS)
-    if (cc_library_SHARED OR cc_library_shared) # build *.so
+  if(cc_library_SRCS)
+    if(cc_library_SHARED OR cc_library_shared) # build *.so
       add_library(${TARGET_NAME} SHARED ${cc_library_SRCS})
     else()
       add_library(${TARGET_NAME} STATIC ${cc_library_SRCS})
     endif()
-    if (cc_library_DEPS)
+    if(cc_library_DEPS)
       # Don't need link libwarpctc.so
-      if ("${cc_library_DEPS};" MATCHES "warpctc;")
+      if("${cc_library_DEPS};" MATCHES "warpctc;")
         list(REMOVE_ITEM cc_library_DEPS warpctc)
         add_dependencies(${TARGET_NAME} warpctc)
       endif()
+      # Support linking flags: --whole-archive (Linux) / -force_load (MacOS)
+      target_circle_link_libraries(${TARGET_NAME} ${cc_library_DEPS})
+      if("${cc_library_DEPS}" MATCHES "ARCHIVE_START")
+        list(REMOVE_ITEM cc_library_DEPS ARCHIVE_START ARCHIVE_END)
+      endif()
       add_dependencies(${TARGET_NAME} ${cc_library_DEPS})
-      target_link_libraries(${TARGET_NAME} ${cc_library_DEPS})
     endif()
 
     # cpplint code style

doc/api/v2/fluid/layers.rst

@@ -323,6 +323,12 @@ batch_norm
 ..  autofunction:: paddle.v2.fluid.layers.batch_norm
     :noindex:
 
+layer_norm
+----------
+
+..  autofunction:: paddle.v2.fluid.layers.layer_norm
+    :noindex:
+
 beam_search_decode
 ------------------
 

doc/howto/rnn/index_cn.rst

@@ -1,4 +1,4 @@
-RNN相关模型
+RNN模型
 ===========
 
 ..  toctree::

doc/index_cn.rst

@@ -8,5 +8,4 @@ PaddlePaddle 文档
   build_and_install/index_cn.rst
   howto/index_cn.rst
   dev/index_cn.rst
-  api/index_cn.rst
   faq/index_cn.rst

doc/index_en.rst

@@ -8,4 +8,3 @@ PaddlePaddle Documentation
   build_and_install/index_en.rst
   howto/index_en.rst
   dev/index_en.rst
-  api/index_en.rst

paddle/framework/block_desc.cc

@@ -162,9 +162,8 @@ BlockDesc::BlockDesc(const BlockDesc &other, proto::BlockDesc *desc,
     : prog_(prog), desc_(desc) {
   need_update_ = true;
   for (auto &op : other.ops_) {
-    ops_.emplace_back(new OpDesc(*op, this));
+    ops_.emplace_back(new OpDesc(*op->Proto(), prog, this));
   }
-
   for (auto &it : other.vars_) {
     auto *var = new VarDesc(*it.second);
     vars_[it.first].reset(var);

paddle/framework/channel.h

@@ -1,4 +1,4 @@
-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+/* 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.

paddle/framework/channel_test.cc

@@ -1,4 +1,4 @@
-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+/* 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.
@@ -25,6 +25,26 @@ using paddle::framework::CloseChannel;
 using paddle::framework::details::Buffered;
 using paddle::framework::details::UnBuffered;
 
+void RecevingOrderEqualToSendingOrder(Channel<int> *ch) {
+  unsigned sum_send = 0;
+  std::thread t([&]() {
+    for (int i = 0; i < 5; i++) {
+      EXPECT_EQ(ch->Send(&i), true);
+      sum_send += i;
+    }
+  });
+  for (int i = 0; i < 5; i++) {
+    int recv;
+    EXPECT_EQ(ch->Receive(&recv), true);
+    EXPECT_EQ(recv, i);
+  }
+
+  CloseChannel(ch);
+  t.join();
+  EXPECT_EQ(sum_send, 10U);
+  delete ch;
+}
+
 TEST(Channel, MakeAndClose) {
   using paddle::framework::details::Buffered;
   using paddle::framework::details::UnBuffered;
@@ -137,9 +157,7 @@ TEST(Channel, ReceiveFromBufferedChannelReturnResidualValuesTest) {
 
   for (size_t i = 0; i < buffer_size; ++i) {
     EXPECT_EQ(ch->Receive(&out),
-              false);  // after receiving residual values, return zeros.
-    // Note: we cannot check EXPECT_EQ(out, 0), because C++ doesn't
-    // define zero values like Go does.
+              false);  // receiving on closed channel should return false
   }
   delete ch;
 }
@@ -158,39 +176,25 @@ TEST(Channel, ConcurrentSendNonConcurrentReceiveWithSufficientBufferSize) {
       sum += i;
     }
   });
-  std::this_thread::sleep_for(std::chrono::milliseconds(100));  // wait 0.5 sec
+  std::this_thread::sleep_for(std::chrono::milliseconds(100));  // wait 0.1 sec
   EXPECT_EQ(sum, 45U);
 
   CloseChannel(ch);
   t.join();
   delete ch;
 }
 
-TEST(Channel, SimpleUnbufferedChannelTest) {
+TEST(Channel, RecevingOrderEqualToSendingOrderWithUnBufferedChannel) {
   auto ch = MakeChannel<int>(0);
-  unsigned sum_send = 0;
-  std::thread t([&]() {
-    for (int i = 0; i < 5; i++) {
-      EXPECT_EQ(ch->Send(&i), true);
-      sum_send += i;
-    }
-  });
-  for (int i = 0; i < 5; i++) {
-    int recv;
-    EXPECT_EQ(ch->Receive(&recv), true);
-    EXPECT_EQ(recv, i);
-  }
+  RecevingOrderEqualToSendingOrder(ch);
+}
 
-  CloseChannel(ch);
-  t.join();
-  EXPECT_EQ(sum_send, 10U);
-  delete ch;
+TEST(Channel, RecevingOrderEqualToSendingOrderWithBufferedChannel) {
+  auto ch = MakeChannel<int>(10);
+  RecevingOrderEqualToSendingOrder(ch);
 }
 
-// This tests that closing a buffered channel also unblocks
-//  any receivers waiting on the channel
-TEST(Channel, BufferedChannelCloseUnblocksReceiversTest) {
-  auto ch = MakeChannel<int>(1);
+void ChannelCloseUnblocksReceiversTest(Channel<int> *ch) {
   size_t num_threads = 5;
   std::thread t[num_threads];
   bool thread_ended[num_threads];
@@ -201,15 +205,14 @@ TEST(Channel, BufferedChannelCloseUnblocksReceiversTest) {
     t[i] = std::thread(
         [&](bool *p) {
           int data;
-          // All reads should return false
           EXPECT_EQ(ch->Receive(&data), false);
           *p = true;
         },
         &thread_ended[i]);
   }
-  std::this_thread::sleep_for(std::chrono::milliseconds(100));  // wait
+  std::this_thread::sleep_for(std::chrono::milliseconds(100));  // wait 0.1 sec
 
-  // Verify that all threads are blocked
+  // Verify that all the threads are blocked
   for (size_t i = 0; i < num_threads; i++) {
     EXPECT_EQ(thread_ended[i], false);
   }
@@ -218,21 +221,20 @@ TEST(Channel, BufferedChannelCloseUnblocksReceiversTest) {
   // This should unblock all receivers
   CloseChannel(ch);
 
-  std::this_thread::sleep_for(std::chrono::milliseconds(200));  // wait
+  std::this_thread::sleep_for(std::chrono::milliseconds(100));  // wait 0.1 sec
 
   // Verify that all threads got unblocked
   for (size_t i = 0; i < num_threads; i++) {
     EXPECT_EQ(thread_ended[i], true);
   }
 
   for (size_t i = 0; i < num_threads; i++) t[i].join();
-  delete ch;
 }
 
-// This tests that closing a buffered channel also unblocks
-//  any senders waiting for channel to have write space
-TEST(Channel, BufferedChannelCloseUnblocksSendersTest) {
-  auto ch = MakeChannel<int>(1);
+void ChannelCloseUnblocksSendersTest(Channel<int> *ch) {
+  using paddle::framework::details::Buffered;
+  using paddle::framework::details::UnBuffered;
+
   size_t num_threads = 5;
   std::thread t[num_threads];
   bool thread_ended[num_threads];
@@ -252,116 +254,72 @@ TEST(Channel, BufferedChannelCloseUnblocksSendersTest) {
   }
   std::this_thread::sleep_for(std::chrono::milliseconds(100));  // wait
 
-  // Verify that atleast 4 threads are blocked
-  int ct = 0;
-  for (size_t i = 0; i < num_threads; i++) {
-    if (thread_ended[i] == false) ct++;
+  if (dynamic_cast<Buffered<int> *>(ch)) {
+    // If ch is Buffered, atleast 4 threads must be blocked.
+    int ct = 0;
+    for (size_t i = 0; i < num_threads; i++) {
+      if (!thread_ended[i]) ct++;
+    }
+    EXPECT_GE(ct, 4);
+  } else {
+    // If ch is UnBuffered, all the threads should be blocked.
+    for (size_t i = 0; i < num_threads; i++) {
+      EXPECT_EQ(thread_ended[i], false);
+    }
   }
-  // Atleast 4 threads must be blocked
-  EXPECT_GE(ct, 4);
-
   // Explicitly close the thread
   // This should unblock all senders
   CloseChannel(ch);
 
-  std::this_thread::sleep_for(std::chrono::milliseconds(200));  // wait
+  std::this_thread::sleep_for(std::chrono::milliseconds(100));  // wait
 
   // Verify that all threads got unblocked
   for (size_t i = 0; i < num_threads; i++) {
     EXPECT_EQ(thread_ended[i], true);
   }
 
-  // Verify that only 1 send was successful
-  ct = 0;
-  for (size_t i = 0; i < num_threads; i++) {
-    if (send_success[i]) ct++;
+  if (dynamic_cast<Buffered<int> *>(ch)) {
+    // Verify that only 1 send was successful
+    int ct = 0;
+    for (size_t i = 0; i < num_threads; i++) {
+      if (send_success[i]) ct++;
+    }
+    // Only 1 send must be successful
+    EXPECT_EQ(ct, 1);
   }
-  // Only 1 send must be successful
-  EXPECT_EQ(ct, 1);
 
   for (size_t i = 0; i < num_threads; i++) t[i].join();
+}
+
+// This tests that closing a buffered channel also unblocks
+//  any receivers waiting on the channel
+TEST(Channel, BufferedChannelCloseUnblocksReceiversTest) {
+  auto ch = MakeChannel<int>(1);
+  ChannelCloseUnblocksReceiversTest(ch);
+  delete ch;
+}
+
+// This tests that closing a buffered channel also unblocks
+//  any senders waiting for channel to have write space
+TEST(Channel, BufferedChannelCloseUnblocksSendersTest) {
+  auto ch = MakeChannel<int>(1);
+  ChannelCloseUnblocksSendersTest(ch);
   delete ch;
 }
 
 // This tests that closing an unbuffered channel also unblocks
 //  unblocks any receivers waiting for senders
 TEST(Channel, UnbufferedChannelCloseUnblocksReceiversTest) {
   auto ch = MakeChannel<int>(0);
-  size_t num_threads = 5;
-  std::thread t[num_threads];
-  bool thread_ended[num_threads];
-
-  // Launches threads that try to read and are blocked becausew of no writers
-  for (size_t i = 0; i < num_threads; i++) {
-    thread_ended[i] = false;
-    t[i] = std::thread(
-        [&](bool *p) {
-          int data;
-          EXPECT_EQ(ch->Receive(&data), false);
-          *p = true;
-        },
-        &thread_ended[i]);
-  }
-  std::this_thread::sleep_for(std::chrono::milliseconds(500));  // wait 0.5 sec
-
-  // Verify that all the threads are blocked
-  for (size_t i = 0; i < num_threads; i++) {
-    EXPECT_EQ(thread_ended[i], false);
-  }
-
-  // Explicitly close the thread
-  // This should unblock all receivers
-  CloseChannel(ch);
-
-  std::this_thread::sleep_for(std::chrono::milliseconds(500));  // wait 0.5 sec
-
-  // Verify that all threads got unblocked
-  for (size_t i = 0; i < num_threads; i++) {
-    EXPECT_EQ(thread_ended[i], true);
-  }
-
-  for (size_t i = 0; i < num_threads; i++) t[i].join();
+  ChannelCloseUnblocksReceiversTest(ch);
   delete ch;
 }
 
 // This tests that closing an unbuffered channel also unblocks
 //  unblocks any senders waiting for senders
 TEST(Channel, UnbufferedChannelCloseUnblocksSendersTest) {
   auto ch = MakeChannel<int>(0);
-  size_t num_threads = 5;
-  std::thread t[num_threads];
-  bool thread_ended[num_threads];
-
-  // Launches threads that try to read and are blocked becausew of no writers
-  for (size_t i = 0; i < num_threads; i++) {
-    thread_ended[i] = false;
-    t[i] = std::thread(
-        [&](bool *p) {
-          int data = 10;
-          EXPECT_EQ(ch->Send(&data), false);
-          *p = true;
-        },
-        &thread_ended[i]);
-  }
-  std::this_thread::sleep_for(std::chrono::milliseconds(500));  // wait 0.5 sec
-
-  // Verify that all the threads are blocked
-  for (size_t i = 0; i < num_threads; i++) {
-    EXPECT_EQ(thread_ended[i], false);
-  }
-
-  // Explicitly close the thread
-  // This should unblock all receivers
-  CloseChannel(ch);
-
-  std::this_thread::sleep_for(std::chrono::milliseconds(500));  // wait 0.5 sec
-
-  // Verify that all threads got unblocked
-  for (size_t i = 0; i < num_threads; i++) {
-    EXPECT_EQ(thread_ended[i], true);
-  }
-
-  for (size_t i = 0; i < num_threads; i++) t[i].join();
+  ChannelCloseUnblocksReceiversTest(ch);
   delete ch;
 }