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

Author: JiayiFeng

Dockerfile

@@ -76,7 +76,8 @@ RUN easy_install -U pip && \
     pip install sphinx-rtd-theme==0.1.9 recommonmark
 
 RUN pip install pre-commit 'ipython==5.3.0' && \
-    pip install 'ipykernel==4.6.0' 'jupyter==1.0.0'
+    pip install 'ipykernel==4.6.0' 'jupyter==1.0.0' && \
+    pip install opencv-python
 
 #For docstring checker
 RUN pip install pylint pytest astroid isort

cmake/external/mkldnn.cmake

@@ -54,7 +54,7 @@ ExternalProject_Add(
     ${EXTERNAL_PROJECT_LOG_ARGS}
     DEPENDS             ${MKLDNN_DEPENDS}
     GIT_REPOSITORY      "https://github.com/01org/mkl-dnn.git"
-    GIT_TAG             "db3424ad44901513c03a1ea31ccaacdf633fbe9f"
+    GIT_TAG             "a29d8487a63afca3d5b8c5bbdbb473cf8ccc6e51"
     PREFIX              ${MKLDNN_SOURCES_DIR}
     UPDATE_COMMAND      ""
     CMAKE_ARGS          -DCMAKE_INSTALL_PREFIX=${MKLDNN_INSTALL_DIR}

doc/v2/faq/build_and_install/index_cn.rst

@@ -213,3 +213,12 @@ virtualenv本身也是Python的一个包，可以用pip进行安装：
 保存并关闭文件。
 
 这样，每次打开终端时就会自动启动名为‘paddle’的Python环境了。
+
+10. 通过pip安装的PaddlePaddle在  :code:`import paddle.fluid` 报找不到 :code:`libmkldnn.so` 或 :code:`libmklml_intel.so`
+------------------------------------------------------------------------------------------
+出现这种问题的原因是在导入 :code:`paddle.fluid` 时需要加载 :code:`libmkldnn.so` 和 :code:`libmklml_intel.so`，
+但是系统没有找到该文件。一般通过pip安装PaddlePaddle时会将 :code:`libmkldnn.so` 和 :code:`libmklml_intel.so`
+拷贝到 :code:`/usr/local/lib` 路径下，所以解决办法是将该路径加到 :code:`LD_LIBRARY_PATH` 环境变量下，
+即： :code:`export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH` 。
+
+**注意**：如果是在虚拟环境中安装PaddlePaddle， :code:`libmkldnn.so` 和 :code:`libmklml_intel.so` 可能不在 :code:`/usr/local/lib` 路径下。

paddle/fluid/framework/details/broadcast_op_handle.cc

@@ -73,6 +73,9 @@ void BroadcastOpHandle::RunImpl() {
     int root_id = boost::get<platform::CUDAPlace>(in_tensor.place()).device;
     std::vector<std::function<void()>> broadcast_calls;
 
+    int type = platform::ToNCCLDataType(in_tensor.type());
+    size_t numel = static_cast<size_t>(in_tensor.numel());
+
     for (auto out_var_handle : out_var_handles) {
       Variable *out_var = var_scopes.at(out_var_handle->scope_idx_)
                               ->FindVar(out_var_handle->name_);
@@ -87,13 +90,11 @@ void BroadcastOpHandle::RunImpl() {
         send_recv_buffer = const_cast<void *>(in_tensor.data<void>());
         out_handle = out_var_handle;
       } else {
-        send_recv_buffer =
-            VariableVisitor::GetMutableTensor(out_var).mutable_data(
-                out_var_handle->place_);
+        send_recv_buffer = VariableVisitor::GetMutableTensor(out_var)
+                               .Resize(in_tensor.dims())
+                               .mutable_data(out_var_handle->place_);
       }
 
-      int type = platform::ToNCCLDataType(in_tensor.type());
-      size_t numel = static_cast<size_t>(in_tensor.numel());
       broadcast_calls.emplace_back(
           [send_recv_buffer, numel, type, root_id, &nccl_ctx] {
             PADDLE_ENFORCE(platform::dynload::ncclBcast(

paddle/fluid/framework/details/multi_devices_graph_builder.cc

@@ -351,7 +351,7 @@ void MultiDevSSAGraphBuilder::InsertAllReduceOp(SSAGraph *result,
     auto &prev_grad = vars.back();
     op_handle->AddInput(prev_grad.get());
 
-    auto var = new VarHandle(vars.size() - 1, i, og, p);
+    auto var = new VarHandle(vars.size(), i, og, p);
     vars.emplace_back(var);
     op_handle->AddOutput(var);
   }
@@ -447,8 +447,7 @@ VarHandle *MultiDevSSAGraphBuilder::CreateReduceOp(SSAGraph *result,
     op_handle->AddInput(prev_grad.get());
   }
   auto &vars = result->vars_[dst_dev_id][og];
-  auto var =
-      new VarHandle(vars.size() - 1, dst_dev_id, og, places_[dst_dev_id]);
+  auto var = new VarHandle(vars.size(), dst_dev_id, og, places_[dst_dev_id]);
   vars.emplace_back(var);
   op_handle->AddOutput(var);
   return var;

paddle/fluid/framework/details/multi_devices_graph_builder.h

@@ -47,7 +47,7 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
 #endif
 
   std::unique_ptr<SSAGraph> Build(const ProgramDesc &program) const override;
-  int GetVarDeviceID(const std::string &varname) const;
+  int GetVarDeviceID(const std::string &varname) const override;
 
  private:
   void CreateOpHandleIOs(SSAGraph *result, const OpDesc &op,

paddle/fluid/framework/details/op_handle_base.cc

@@ -11,8 +11,8 @@
 // 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/fluid/framework/details/op_handle_base.h"
+#include <map>
 
 namespace paddle {
 namespace framework {
@@ -122,11 +122,16 @@ void OpHandleBase::RunAndRecordEvent(const std::function<void()> &callback) {
 #ifdef PADDLE_WITH_CUDA
   if (!events_.empty()) {  // Use event
     std::function<void()> method = callback;
-
+    // NOTE(zcd): device context must be ordered here because RecordEvent
+    // will use a mutex to ensure the safe of multi-threads.
+    std::map<platform::DeviceContext *, platform::Place> ordered_ctxes;
     for (auto &p : dev_ctxes_) {
+      ordered_ctxes.emplace(p.second, p.first);
+    }
+    for (auto &p : ordered_ctxes) {
       method = [method, p, this]() {
-        static_cast<platform::CUDADeviceContext *>(p.second)->RecordEvent(
-            events_.at(boost::get<platform::CUDAPlace>(p.first).device),
+        static_cast<platform::CUDADeviceContext *>(p.first)->RecordEvent(
+            events_.at(boost::get<platform::CUDAPlace>(p.second).device),
             method);
       };
     }

paddle/fluid/framework/framework.proto

@@ -27,6 +27,7 @@ enum AttrType {
   BOOLEANS = 7;
   BLOCK = 8;
   LONG = 9;
+  BLOCKS = 10;
 }
 
 // OpDesc describes an instance of a C++ framework::OperatorBase
@@ -46,6 +47,7 @@ message OpDesc {
     repeated bool bools = 11;
     optional int32 block_idx = 12;
     optional int64 l = 13;
+    repeated int32 blocks_idx = 14;
   };
 
   message Var {

paddle/fluid/framework/lod_tensor.cc

@@ -51,8 +51,6 @@ std::ostream &operator<<(std::ostream &os, const LoD &lod) {
 }
 
 std::ostream &operator<<(std::ostream &os, const LoDTensor &t) {
-  PADDLE_ENFORCE(t.type().hash_code() == typeid(float).hash_code());
-
   if (!platform::is_cpu_place(t.place())) {
     LoDTensor tt;
     framework::TensorCopy(t, platform::CPUPlace(), &tt);
@@ -70,7 +68,13 @@ std::ostream &operator<<(std::ostream &os, const LoDTensor &t) {
   // only print first ten elements
   int64_t size = t.numel() < 10 ? t.numel() : 10;
   for (int64_t i = 0; i < size; ++i) {
-    os << t.data<float>()[i] << " ";
+    if (t.type().hash_code() == typeid(float).hash_code()) {
+      os << t.data<float>()[i] << " ";
+    } else if (t.type().hash_code() == typeid(int64_t).hash_code()) {
+      os << t.data<int64_t>()[i] << " ";
+    } else {
+      PADDLE_THROW("LoDTensor data type not in [float, int64_t]");
+    }
   }
 
   return os;

paddle/fluid/framework/lod_tensor_test.cc

@@ -26,6 +26,20 @@
 namespace paddle {
 namespace framework {
 
+TEST(LoD, PrintLoDTensor) {
+  LoDTensor tensor1;
+  tensor1.mutable_data<float>(platform::CPUPlace());
+  tensor1.data<float>()[0] = 0.2;
+  tensor1.data<float>()[1] = 0.5;
+  LOG(INFO) << tensor1;
+
+  LoDTensor tensor2;
+  tensor2.mutable_data<int64_t>(platform::CPUPlace());
+  tensor2.data<int64_t>()[0] = 1;
+  tensor2.data<int64_t>()[1] = 2;
+  LOG(INFO) << tensor2;
+}
+
 TEST(LoD, data) {
   LoD lod{{0, 1, 2}};
   lod.push_back({0, 2, 4, 5});
@@ -37,7 +51,7 @@ TEST(LoD, data) {
   }
 }
 
-TEST(LodExpand, test) {
+TEST(LoD, ExpandLoD) {
   LoD lod{{0, 2}};
   LoDTensor tensor;
   tensor.set_lod(lod);