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

CMakeLists.txt

@@ -39,7 +39,6 @@ option(WITH_GPU         "Compile PaddlePaddle with NVIDIA GPU"          ${CUDA_F
 option(WITH_AMD_GPU     "Compile PaddlePaddle with AMD GPU"             OFF)
 option(WITH_AVX         "Compile PaddlePaddle with AVX intrinsics"      ${AVX_FOUND})
 option(WITH_MKL         "Compile PaddlePaddle with MKL support."        ${AVX_FOUND})
-option(WITH_TENSORRT    "Compile PaddlePaddle with TensorRT support."   OFF)
 option(WITH_DSO         "Compile PaddlePaddle with dynamic linked CUDA" ON)
 option(WITH_TESTING     "Compile PaddlePaddle with unit testing"        OFF)
 option(WITH_SWIG_PY     "Compile PaddlePaddle with inference api"       ON)
@@ -180,13 +179,9 @@ set(EXTERNAL_LIBS
 
 if(WITH_GPU)
     include(cuda)
+    include(tensorrt)
 endif(WITH_GPU)
 
-# TensorRT depends on GPU.
-if (NOT WITH_GPU)
-  set(WITH_TENSORRT OFF)
-endif()
-
 if(WITH_AMD_GPU)
     find_package(HIP)
     include(hip)

Dockerfile

@@ -46,7 +46,7 @@ ENV PATH=${PATH}:${GOROOT}/bin:${GOPATH}/bin
 RUN curl -s -q https://glide.sh/get | sh
 
 # Install TensorRT
-# The unnecessary files has been removed to make the library small.
+# The unnecessary files has been removed to make the library small. It only contains include and lib now.
 RUN wget -qO- http://paddlepaddledeps.bj.bcebos.com/TensorRT-4.0.0.3.Ubuntu-16.04.4.x86_64-gnu.cuda-8.0.cudnn7.0.tar.gz | \
     tar -xz -C /usr/local && \
     cp -rf /usr/local/TensorRT/include /usr && \

Dockerfile.android

@@ -27,7 +27,7 @@ RUN git config --global credential.helper store
 # Fix locales to en_US.UTF-8
 RUN localedef -i en_US -f UTF-8 en_US.UTF-8
 
-RUN pip install --upgrade pip && \
+RUN pip install --upgrade pip==9.0.3 && \
     pip install -U 'protobuf==3.1.0' && \
     pip install -U wheel sphinx && \
     pip install pre-commit

cmake/configure.cmake

@@ -80,6 +80,16 @@ if(WITH_GPU)
     # Include cuda and cudnn
     include_directories(${CUDNN_INCLUDE_DIR})
     include_directories(${CUDA_TOOLKIT_INCLUDE})
+
+    if(TENSORRT_FOUND)
+        if(${CUDA_VERSION_MAJOR} VERSION_LESS 8)
+            message(FATAL_ERROR "TensorRT needs CUDA >= 8.0 to compile")
+        endif()
+        if(${CUDNN_MAJOR_VERSION} VERSION_LESS 7)
+            message(FATAL_ERROR "TensorRT needs CUDNN >= 7.0 to compile")
+        endif()
+        include_directories(${TENSORRT_INCLUDE_DIR})
+    endif()
 elseif(WITH_AMD_GPU)
     add_definitions(-DPADDLE_WITH_HIP)
     set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -D__HIP_PLATFORM_HCC__")

cmake/tensorrt.cmake

+if(NOT WITH_GPU)
+    return()
+endif()
+
+set(TENSORRT_ROOT "/usr" CACHE PATH "TENSORRT ROOT")
+find_path(TENSORRT_INCLUDE_DIR NvInfer.h
+    PATHS ${TENSORRT_ROOT} ${TENSORRT_ROOT}/include
+    $ENV{TENSORRT_ROOT} $ENV{TENSORRT_ROOT}/include
+    NO_DEFAULT_PATH
+)
+
+find_library(TENSORRT_LIBRARY NAMES libnvinfer.so libnvinfer.a
+    PATHS ${TENSORRT_ROOT} ${TENSORRT_ROOT}/lib
+    $ENV{TENSORRT_ROOT} $ENV{TENSORRT_ROOT}/lib
+    NO_DEFAULT_PATH
+    DOC "Path to TensorRT library.")
+
+if(TENSORRT_INCLUDE_DIR AND TENSORRT_LIBRARY)
+    set(TENSORRT_FOUND ON)
+else()
+    set(TENSORRT_FOUND OFF)
+endif()
+
+if(TENSORRT_FOUND)
+    file(READ ${TENSORRT_INCLUDE_DIR}/NvInfer.h TENSORRT_VERSION_FILE_CONTENTS)
+    string(REGEX MATCH "define NV_TENSORRT_MAJOR +([0-9]+)" TENSORRT_MAJOR_VERSION
+        "${TENSORRT_VERSION_FILE_CONTENTS}")
+    string(REGEX REPLACE "define NV_TENSORRT_MAJOR +([0-9]+)" "\\1"
+        TENSORRT_MAJOR_VERSION "${TENSORRT_MAJOR_VERSION}")
+
+    message(STATUS "Current TensorRT header is ${TENSORRT_INCLUDE_DIR}/NvInfer.h. "
+        "Current TensorRT version is v${TENSORRT_MAJOR_VERSION}. ")
+endif()

doc/CMakeLists.txt

@@ -3,7 +3,9 @@ add_custom_target(paddle_apis ALL
 
 add_custom_target(paddle_docs ALL
                   DEPENDS paddle_v2_docs paddle_v2_docs_cn
-                  paddle_fluid_docs paddle_fluid_docs_cn)
+                  paddle_fluid_docs paddle_fluid_docs_cn
+                  paddle_mobile_docs paddle_mobile_docs_cn)
 
 add_subdirectory(v2)
 add_subdirectory(fluid)
+add_subdirectory(mobile)

doc/fluid/api/initializer.rst

@@ -33,3 +33,45 @@ Xavier
     :members:
     :noindex:
 
+MSRA
+------
+
+..  autoclass:: paddle.fluid.initializer.MSRA
+    :members:
+    :noindex:
+
+ConstantInitializer
+-------------------
+
+..  autoclass:: paddle.fluid.initializer.ConstantInitializer
+    :members:
+    :noindex:
+
+UniformInitializer
+------------------
+
+..  autoclass:: paddle.fluid.initializer.UniformInitializer
+    :members:
+    :noindex:
+
+NormalInitializer
+-----------------
+
+..  autoclass:: paddle.fluid.initializer.NormalInitializer
+    :members:
+    :noindex:
+
+XavierInitializer
+-----------------
+
+..  autoclass:: paddle.fluid.initializer.XavierInitializer
+    :members:
+    :noindex:
+    MSRA
+    ------
+
+MSRAInitializer
+-----------------
+..  autoclass:: paddle.fluid.initializer.MSRAInitializer
+    :members:
+    :noindex:

doc/fluid/design/concepts/parallel_executor.md

@@ -84,7 +84,7 @@ Running an operator can be asynchronized. There is a thread pool to execute an `
 
 ## Synchronize GPU Kernels
 
-The GPU is a non-blocking device. The different streams need be synchronized when switing streams. In current implementation, the synchronization based on the following algorithm:
+The GPU is a non-blocking device. The different streams need be synchronized when switching streams. In current implementation, the synchronization based on the following algorithm:
 
 1. `OpHandle` will record `DeviceContext` that it is used.
 2. In `OpHandle::Run`, if the `DeviceContext` of current operator is different from `DeviceContext` of any input variable, just wait the generate operator of this input variable.

doc/fluid/design/dist_train/README.md

+## Distributed training overview doc
+
+Currently Paddle Fluid use parameter server architecture to support distributed training.
+
+For synchronous and asynchronous training, the differences are mostly in the logic of parameter server. Now we have already support synchronous training.
+
+### Synchronous training
+
+The training process of synchronous training is:
+
+![synchronous distributed training](./src/sync_distributed_training.png)
+
+1. Pserver
+	1. set `barrier_condition_` to 0 and waits for trainers to send gradient.
+1. Trainer
+	1. Trainer read minibatch of data, run forward-backward with local parameter copy and get the gradients for parameters.
+	1. Trainer use split op to split all the gradient into blocks. The split method is determined at compile time.
+	1. Trainer use send_op to send all the split gradients to corresponding parameter server.
+	1. After trainer send all the gradients, it will send a `BATCH_BARRIER_MESSAGE` to all pservers.
+	1. Trainer call GetVariable to pserver and wait for `barrier_condition_` on pserver to be 1.
+1. Pserver
+   1. Pserver will count the number of `BATCH_BARRIER_MESSAGE`.
+	1. When the count of `BATCH_BARRIER_MESSAGE` is equal to the number of Trainer. Pserver thinks it received all gradient from all trainers.
+	1. Pserver will run the optimization block to optimize the parameters.
+	1. After optimization, pserver set `barrier_condition_` to 1.
+	1. Pserver wait for `FETCH_BARRIER_MESSAGE`.
+1. Trainer.
+	1. The trainer uses GetVariable to get all the parameters from pserver.
+	1. Trainer sends a `FETCH_BARRIER_MESSAGE` to each pserver.
+1. Pserver.
+	1. when the number of `FETCH_BARRIER_MESSAGE` reach the number of all trainers. Pserver think all the parameters have been got. it will go back to 1. to set `barrier_condition_` to 0.
+
+### Asynchronous training
+In the above process. There are two barriers for all trainers to synchronize with each other. In asynchronous training, these two barriers are not needed. The trainer can just send gradients to pserver and then get parameters back.
+
+The training process of asynchronous training can be:
+
+![asynchronous distributed training](./src/async_distributed_training.png)
+
+1. Pserver:
+	1. Each parameter has a queue to receive its gradient from trainers.
+	1. Each parameter has a thread to read data from the queue and run optimize block, using the gradient to optimize the parameter.
+	1. Using an independent thread to handle RPC call `GetVariable` for trainers to get parameters back.(Maybe here we should use a thread pool to speed up fetching the parameters.)
+
+1. Trainer:
+	1. Trainer read a batch of data. Run forward and backward with local parameter copy and get the gradients for parameters.
+	1. Trainer split all gradients to blocks and then send these gradient blocks to pservers(pserver will put them into the queue).
+	2. Trainer gets all parameters back from pserver.
+
+### Note:
+There are also some conditions that need to consider. For exmaple:
+
+1. If trainer needs to wait for the pserver to apply it's gradient and then get back the parameters back.
+1. If we need a lock between parameter update and parameter fetch.
+1. If one parameter must be on one server, or it can also be split and send to multiple parameter servers.
+
+The above architecture of asynchronous training can support different mode, we can have a detailed test in the future for these problems.

doc/fluid/design/dist_train/async_update.md

+# Design Doc: Asynchronous Update With Distributed Training
+
+## Background
+
+For the typical synchronous distributed training, some significant steps are as follows:
+
+1. A Trainer will compute the gradients and SEND them to the Parameter Server(PServer) nodes.
+1. After the PServer node received gradients came from all the Trainers, It will aggregate the
+gradient variables for the same parameter into one gradient variable and then apply the aggregated
+gradient to the respective parameter, finally using an optimize algorithms(SGD, Monument...)
+to update the parameters.
+1. The Trainer would wait for the PServers finished the optimize stage, and GET the parameters from PServer,
+so all the Trainers would get the same parameters.
+
+In the synchronously distributed training, there should be a `Barrier` to synchronise the
+parameters after the optimizing stage. The performance of a distributed training job would
+depend on the slowest node if there were hundreds or thousands of training nodes in a
+Job, the performance of synchronously distributed training might be very poor because of
+the slow node. So this design doc would introduce an approach to implement
+*asynchronously* distributed training in PaddlePaddle Fluid.
+
+## Design
+
+<img src="./src/async_update.png" width="600"/>
+
+As the figure above, we describe a global view of asynchronously update process and use
+the parameter `w1` as an example to introduce the steps:
+1. For each gradient variables, they may distribute on different GPU card and aggregate
+them while they are all calculated.
+1. Split the gradient variable into multiple blocks according to the number of PServer
+instances and then send them.
+1. PServer would run an `Optimize Block` using a specified optimize algorithm to update
+the specified parameter.
+1. The trainer will fetch latest parameter from PServer before running forward Op which depends
+on the specified parameter.
+1. Broadcast the received variable into multiple GPU cards and continue to run the next
+mini-batch.
+
+### Trainer
+
+- For the multiple devices distributed training, we need to aggregate the gradient
+variables which placed on different devices firstly and then schedule a `SendVars` Operator to
+send the gradient variables to the multiple PServer instances.
+- Schedule `FetchVars` operator to fetch the latest parameter from PServer before running
+the forward ops.
+- There could be a large number of gradient variables to be sent, so we need to use another
+thread pool(IO Threadpool) whose a number of the schedulable threads is larger than the
+computing thread pool to avoid competitive the thread resources with computing.
+
+### Parameter Server
+
+<img src="./src/async_pserver.png" width="750"/>
+
+- There should be multiple trainer instances want to optimize the same parameter at
+the same time, to avoid the racing, we need one `BlockingQueue` for each gradient
+variable to process them one by one.
+- We need a `Map` structure to map a gradient variable name to the `OptimizeBlock` which
+can optimize the respective parameter.

doc/fluid/design/dist_train/mpi_enabled_design.md

+# MPI-enabled PaddlePaddle Design doc
+
+# Background
+When we do distribute multi GPU training, the communication overhead between servers become the major bottleneck, because of the following reasons:
+1. Must copy at least once from GPU to CPU memory so that the data can be ready to transfer. And for the pserver side, copy data from CPU to GPU introduce more overhead.
+2. GPU->CPU data transfer is 10 times slower than data transfer between GPUs or between PCIe devices.
+3. TCP connections can not make full use of RDMA 100Gb devices.
+
+We will use OpenMPI API to PaddlePaddle, which can bring two benefits to PaddlePaddle:
+1. Enable RDMA with PaddlePaddle, which bring high-performance low latency networks.
+2. Enable GPUDriect with PaddlePaddle, which bring the highest throughput and lowest latency GPU read and write.
+
+# Change list
+* Compile args: Need add compile args to enable MPI support.
+* Execute args:  Need add execute args to assign when and how to use MPI operations.
+* New ops:  Need new op  ```mpi_send_op``` and ```mpi_listenandserve_op``` to support MPI send and receive.
+* Transpiler optimized: Which can add   ```mpi_send_op``` and ```mpi_listenandserve_op```  to the running graph.
+* MPI utils package: Need MPI utils package as the low-level API supported.
+
+## Compile args
+Because MPI or CUDA need hardware supported, so we will add compile args to enable MPI support and control compiling.Add ```WITH_MPI```  compile args to control MPI to use or not. If the  ```WITH_MPI``` is ```ON```, compile system will find openMPI codes in configuration. We should prepare openMPI environment before compiling.
+
+## Execute args
+Launch the script using the ```mpirun``` launcher, For example: ```mpirun -np 3 -hosts node1,node2,node3 python train.py```. By doing this, We can number the actors (trainer/pserver/master) with o .. (n-1). The node's number is the Rank of the calling process in a group of comm (integer),  The MPI processes identify each other using a Rank ID. We have to create a mapping between PaddlePaddle's nodes and their Rank ID so that we can communicate with the correct destinations when using MPI operations.
+
+## New ops
+We won't replace all the gRPC requests to MPI requests,  the standard gRPC library is used for all administrative operations and the MPI API will be used to transfer tensor or selectRows to Pservers. The base of this idea, we create two new operators to handle requests and receives,  the two operators are ```mpi_send_op``` and ```mpi_listenandserve_op```. They are a little similar to [send_op](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/operators/send_op.cc) and [listen_and_serv_op](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/operators/listen_and_serv_op.cc), also, We will build a new module to package MPI send and receive process.
+
+### mpi_send_op
+Very similar with ```send_op```, we will replace gRPC code which used to send gradient with ```mpi_module```, at the same time, we will wrap it with ```framework::Async```.
+
+### mpi_listenandserve_op
+Very similar with ```listen_and_serv_op```, we will replace gRPC code which used to receive gradient with ```mpi_module```, at the same time, we will wrap it with ```framework::Async```.
+
+## Transpiler optimized
+**We can get env ```OMPI_COMM_WORLD_SIZE``` and ```OMPI_COMM_WORLD_RANK``` to distinguish use MPI or not, If we use openMPI, the variable in env must exist.**
+ if  confirm to use MPI, we will modify  ```send_op``` to ```mpi_send_op``` in distribute_transpiler, and modify ```listenandserve_op``` to ```mpi_listenandserve_op``` also.
+
+## MPI utils package
+In this package, We will write openMPI low-level API to use MPI.
+The API included in this package are:
+* MPI send and receive module, We will build a new module to package MPI send and receive process. MPI send and receive are different to gRPC, the MPI [recvice](https://www.open-mpi.org/doc/v1.8/man3/MPI_Irecv.3.php) must know receive buffer size and receive buffer element. For this reason, We have to make communications twice, the first one is to send metadata about gradient through gRPC, the second one is the real communication through MPI which send gradient data to mpi_listenandserve_op.
+The detailed flow is below:
+![](https://github.com/seiriosPlus/Paddle/blob/mpi_enabled/doc/fluid/design/dist_train/src/mpi_module.png)
+* MPI global configurations, which store the Rank ID and the mapping in global variables, for example:
+gRPC client : MPI nodes :``` 127.0.0.1:32004 : 3 ```

doc/mobile/CMakeLists.txt

+if(NOT DEFINED SPHINX_THEME)
+    set(SPHINX_THEME default)
+endif()
+
+if(NOT DEFINED SPHINX_THEME_DIR)
+    set(SPHINX_THEME_DIR)
+endif()
+
+# configured documentation tools and intermediate build results
+set(BINARY_BUILD_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/_build")
+
+# Sphinx cache with pickled ReST documents
+set(SPHINX_CACHE_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/_doctrees")
+
+# HTML output director
+set(SPHINX_HTML_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/html")
+
+configure_file(
+        "${CMAKE_CURRENT_SOURCE_DIR}/../templates/conf.py.en.in"
+        "${BINARY_BUILD_DIR_EN}/conf.py"
+        @ONLY)
+
+sphinx_add_target(paddle_mobile_docs
+        html
+        ${BINARY_BUILD_DIR_EN}
+        ${SPHINX_CACHE_DIR_EN}
+        ${CMAKE_CURRENT_SOURCE_DIR}
+        ${SPHINX_HTML_DIR_EN})
+
+add_dependencies(paddle_mobile_docs gen_proto_py paddle_python)
+
+# configured documentation tools and intermediate build results
+set(BINARY_BUILD_DIR_CN "${CMAKE_CURRENT_BINARY_DIR}/cn/_build")
+
+# Sphinx cache with pickled ReST documents
+set(SPHINX_CACHE_DIR_CN "${CMAKE_CURRENT_BINARY_DIR}/cn/_doctrees")
+
+# HTML output director
+set(SPHINX_HTML_DIR_CN "${CMAKE_CURRENT_BINARY_DIR}/cn/html")
+
+configure_file(
+        "${CMAKE_CURRENT_SOURCE_DIR}/../templates/conf.py.cn.in"
+        "${BINARY_BUILD_DIR_CN}/conf.py"
+        @ONLY)
+
+sphinx_add_target(paddle_mobile_docs_cn
+        html
+        ${BINARY_BUILD_DIR_CN}
+        ${SPHINX_CACHE_DIR_CN}
+        ${CMAKE_CURRENT_SOURCE_DIR}
+        ${SPHINX_HTML_DIR_CN})
+
+add_dependencies(paddle_mobile_docs_cn gen_proto_py paddle_python)

doc/mobile/index_cn.rst

+移动端
+=====
+
+..  toctree::
+  :maxdepth: 1
+
+  cross_compiling_for_android_cn.md
+  cross_compiling_for_ios_cn.md
+  cross_compiling_for_raspberry_cn.md
\ No newline at end of file

doc/mobile/index_en.rst

+Mobile
+======
+
+..  toctree::
+  :maxdepth: 1
+
+  cross_compiling_for_android_en.md
+  cross_compiling_for_ios_en.md
+  cross_compiling_for_raspberry_en.md

paddle/fluid/framework/details/broadcast_op_handle_test.cc

@@ -77,14 +77,9 @@ struct TestBroadcastOpHandle {
     local_scopes_[input_scope_idx]->Var("input");
 
     op_handle_.reset(new BroadcastOpHandle(local_scopes_, gpu_list_));
-
-    vars_.emplace_back(new VarHandle());
-    VarHandle* in_var_handle = static_cast<VarHandle*>(vars_.back().get());
-    in_var_handle->place_ = gpu_list_[input_scope_idx];
-    in_var_handle->name_ = "input";
-    in_var_handle->version_ = 1;
-    in_var_handle->scope_idx_ = input_scope_idx;
-    in_var_handle->generated_op_ = nullptr;
+    auto* in_var_handle =
+        new VarHandle(1, input_scope_idx, "input", gpu_list_[input_scope_idx]);
+    vars_.emplace_back(in_var_handle);
     op_handle_->AddInput(in_var_handle);
 
     // add dummy var
@@ -96,12 +91,8 @@ struct TestBroadcastOpHandle {
 
     for (size_t j = 0; j < gpu_list_.size(); ++j) {
       op_handle_->dev_ctxes_[gpu_list_[j]] = ctxs_[j].get();
-      vars_.emplace_back(new VarHandle());
-      VarHandle* out_var_handle = static_cast<VarHandle*>(vars_.back().get());
-      out_var_handle->place_ = gpu_list_[j];
-      out_var_handle->name_ = "out";
-      out_var_handle->version_ = 2;
-      out_var_handle->scope_idx_ = j;
+      VarHandle* out_var_handle = new VarHandle(2, j, "out", gpu_list_[j]);
+      vars_.emplace_back(out_var_handle);
       op_handle_->AddOutput(out_var_handle);
     }
 

paddle/fluid/framework/details/gather_op_handle_test.cc

@@ -79,13 +79,8 @@ struct TestGatherOpHandle {
     // add input
     for (size_t j = 0; j < gpu_list_.size(); ++j) {
       op_handle_->dev_ctxes_[gpu_list_[j]] = ctxs_[j].get();
-      vars_.emplace_back(new VarHandle());
-      VarHandle* in_var_handle = static_cast<VarHandle*>(vars_.back().get());
-      in_var_handle->place_ = gpu_list_[j];
-      in_var_handle->name_ = "input";
-      in_var_handle->version_ = 1;
-      in_var_handle->scope_idx_ = j;
-      in_var_handle->generated_op_ = nullptr;
+      auto* in_var_handle = new VarHandle(1, j, "input", gpu_list_[j]);
+      vars_.emplace_back(in_var_handle);
       op_handle_->AddInput(in_var_handle);
     }
 
@@ -97,12 +92,9 @@ struct TestGatherOpHandle {
     op_handle_->AddInput(in_dummy_var_handle);
 
     // add output
-    vars_.emplace_back(new VarHandle());
-    VarHandle* out_var_handle = static_cast<VarHandle*>(vars_.back().get());
-    out_var_handle->place_ = gpu_list_[input_scope_idx];
-    out_var_handle->name_ = "out";
-    out_var_handle->version_ = 2;
-    out_var_handle->scope_idx_ = input_scope_idx;
+    auto* out_var_handle =
+        new VarHandle(2, input_scope_idx, "out", gpu_list_[input_scope_idx]);
+    vars_.emplace_back(out_var_handle);
     op_handle_->AddOutput(out_var_handle);
 
     // add dummy var

paddle/fluid/framework/details/multi_devices_graph_builder.cc

@@ -177,13 +177,9 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
             auto &prev_grad = vars[vars.size() - 1];
             op_handle->AddInput(prev_grad.get());
 
-            vars.emplace_back(new VarHandle);
-            auto &var = vars.back();
-            var->place_ = p;
-            var->name_ = og;
-            var->version_ = vars.size() - 1;
-
-            op_handle->AddOutput(var.get());
+            auto var = new VarHandle(vars.size() - 1, i, og, p);
+            vars.emplace_back(var);
+            op_handle->AddOutput(var);
           }
 #else
           PADDLE_ENFORCE("Not implemented");

paddle/fluid/framework/details/reduce_op_handle_test.cc

@@ -111,13 +111,9 @@ struct TestReduceOpHandle {
       if (!use_gpu_) {
         op_handle_->dev_ctxes_[gpu_list_[j]] = ctxs_[j].get();
       }
-      vars_.emplace_back(new VarHandle());
-      VarHandle *in_var_handle = static_cast<VarHandle *>(vars_.back().get());
-      in_var_handle->place_ = gpu_list_[j];
-      in_var_handle->name_ = "input";
-      in_var_handle->version_ = 1;
-      in_var_handle->scope_idx_ = j;
+      auto *in_var_handle = new VarHandle(1, j, "input", gpu_list_[j]);
       in_var_handle->generated_op_ = nullptr;
+      vars_.emplace_back(in_var_handle);
       op_handle_->AddInput(in_var_handle);
     }
 
@@ -129,12 +125,9 @@ struct TestReduceOpHandle {
     op_handle_->AddInput(in_dummy_var_handle);
 
     // add output
-    vars_.emplace_back(new VarHandle());
-    VarHandle *out_var_handle = static_cast<VarHandle *>(vars_.back().get());
-    out_var_handle->place_ = gpu_list_[input_scope_idx];
-    out_var_handle->name_ = "out";
-    out_var_handle->version_ = 2;
-    out_var_handle->scope_idx_ = input_scope_idx;
+    auto *out_var_handle =
+        new VarHandle(2, input_scope_idx, "out", gpu_list_[input_scope_idx]);
+    vars_.emplace_back(out_var_handle);
     op_handle_->AddOutput(out_var_handle);
 
     // add dummy var

paddle/fluid/framework/details/ssa_graph_builder.cc

@@ -54,13 +54,8 @@ VarHandle *SSAGraphBuilder::CreateOrGetLatestVarHandle(
   auto &var_holder = var_holders[each_var_name];
   VarHandle *var = nullptr;
   if (var_holder.empty()) {
-    var_holder.emplace_back(new VarHandle);
-    auto &init_var = var_holder[0];
-    init_var->place_ = place;
-    init_var->name_ = each_var_name;
-    init_var->generated_op_ = nullptr;
-    init_var->version_ = 0;
-    var = init_var.get();
+    var = new VarHandle(0, place_offset, each_var_name, place);
+    var_holder.emplace_back(var);
   } else {
     var = var_holder.rbegin()->get();
   }
@@ -73,12 +68,9 @@ void SSAGraphBuilder::CreateOpOutput(SSAGraph *graph, OpHandleBase *op_handle,
                                      size_t place_offset) {
   auto &vars = graph->vars_[place_offset][each_var_name];
   size_t version = vars.size();
-  vars.emplace_back(new VarHandle());
-  auto &var = vars.back();
-  var->version_ = version;
-  var->name_ = each_var_name;
-  var->place_ = place;
-  op_handle->AddOutput(var.get());
+  auto var = new VarHandle(version, place_offset, each_var_name, place);
+  vars.emplace_back(var);
+  op_handle->AddOutput(var);
 }
 
 template <typename Callback>

paddle/fluid/framework/details/threaded_ssa_graph_executor.cc

@@ -33,13 +33,6 @@ ThreadedSSAGraphExecutor::ThreadedSSAGraphExecutor(
       running_ops_(0),
       allow_op_delay_(allow_op_delay) {}
 
-void ThreadedSSAGraphExecutor::RunDelayedOps(
-    const std::unordered_set<OpHandleBase *> &delayed_ops) {
-  for (auto op : delayed_ops) {
-    op->Run(use_event_);
-  }
-}
-
 FeedFetchList ThreadedSSAGraphExecutor::Run(
     const std::vector<std::string> &fetch_tensors) {
   std::unordered_map<OpHandleBase *, size_t> pending_ops;
@@ -51,8 +44,6 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
   // together since we currently cannot overlap computation and memcpy streams.
   // Should revisit it if overlapping is available.
   std::unordered_set<OpHandleBase *> delayed_ops;
-  std::unordered_set<OpHandleBase *> blocked_by_delayed_ops;
-  std::unordered_set<VarHandleBase *> delayed_vars;
 
   auto InsertPendingVar = [&pending_vars, &ready_vars](VarHandleBase &var) {
     pending_vars.insert(&var);
@@ -122,24 +113,26 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
     InsertPendingOp(*op);
   }
 
-  auto run_all_ready_ops = [&] {
-    for (auto *op : ready_ops) {
-      if (op->IsMultiDeviceTransfer() && allow_op_delay_) {
-        delayed_ops.insert(op);
-        delayed_vars.insert(op->outputs_.begin(), op->outputs_.end());
-        ready_vars.Extend(op->outputs_);
-        continue;
-      }
+  auto run_all_ops = [&](std::unordered_set<OpHandleBase *> &set) {
+    for (auto *op : set) {
       running_ops_++;
       RunOp(&ready_vars, op);
     }
-    ready_ops.clear();
+    set.clear();
   };
 
   // Step 3. Execution
-  while (!pending_vars.empty() || !ready_ops.empty() || !delayed_ops.empty()) {
+  while (!pending_vars.empty()) {
     // 1. Run All Ready ops
-    run_all_ready_ops();
+    // Keep loop until all vars are ready.
+    //
+    // NOTE: DelayedOps have a lower priority. It will be scheduled after all
+    // ready_ops have been performed.
+    if (ready_ops.empty() && allow_op_delay_) {
+      run_all_ops(delayed_ops);
+    } else {
+      run_all_ops(ready_ops);
+    }
 
     // 2. Find ready variable
     bool timeout;
@@ -160,29 +153,16 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
         auto &deps = pending_ops[op];
         --deps;
         if (deps == 0) {
-          if (delayed_vars.find(ready_var) != delayed_vars.end()) {
-            blocked_by_delayed_ops.insert(op);
+          if (op->IsMultiDeviceTransfer() && allow_op_delay_) {
+            delayed_ops.insert(op);
           } else {
             ready_ops.insert(op);
           }
         }
       }
     }
-    // When there are no other ops to schedule, schedule buffered delayed
-    // ops and unblock other ops.
-    if (ready_ops.empty() && !delayed_ops.empty() && running_ops_ == 0) {
-      RunDelayedOps(delayed_ops);
-      delayed_ops.clear();
-      for (auto *op : blocked_by_delayed_ops) {
-        ready_ops.insert(op);
-      }
-      blocked_by_delayed_ops.clear();
-    }
-    // Keep loop until all vars are ready.
   }
   PADDLE_ENFORCE(ready_ops.empty());
-  PADDLE_ENFORCE(delayed_ops.empty());
-  PADDLE_ENFORCE(blocked_by_delayed_ops.empty());
 
   // Wait FetchOps.
   if (!fetch_ops.empty()) {

paddle/fluid/framework/details/threaded_ssa_graph_executor.h

@@ -88,8 +88,6 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
   void RunOp(BlockingQueue<VarHandleBase *> *ready_var_q,
              details::OpHandleBase *op);
 
-  void RunDelayedOps(const std::unordered_set<OpHandleBase *> &delayed_ops);
-
  private:
   std::unique_ptr<::ThreadPool> pool_;
   std::vector<Scope *> local_scopes_;

paddle/fluid/framework/details/var_handle.h

@@ -16,6 +16,7 @@
 #include <sstream>
 #include <string>
 #include <unordered_set>
+#include <utility>
 
 #include "paddle/fluid/platform/place.h"
 
@@ -33,10 +34,10 @@ struct VarHandleBase {
 
   // The operator who generate this variable. nullptr if the variable
   // is a root node.
-  OpHandleBase *generated_op_;
+  OpHandleBase* generated_op_{nullptr};
 
   // Operators which depend on this variable ready.
-  std::unordered_set<OpHandleBase *> pending_ops_;
+  std::unordered_set<OpHandleBase*> pending_ops_;
 };
 
 // VarHandle is actually a single version of Runtime Variable.
@@ -47,6 +48,13 @@ struct VarHandleBase {
 struct VarHandle : public VarHandleBase {
   std::string DebugString() const override;
 
+  VarHandle(size_t version, size_t scope_index, std::string name,
+            platform::Place place)
+      : version_(version),
+        scope_idx_(scope_index),
+        name_(std::move(name)),
+        place_(std::move(place)) {}
+
   // version field currently is not used, however, just store the version to
   // debug easily.
   size_t version_;

paddle/fluid/framework/parallel_executor.cc

@@ -155,13 +155,9 @@ void ParallelExecutor::BCastParamsToGPUs(
 #endif
 }
 
-void ParallelExecutor::Run(
-    const std::vector<std::string> &fetch_tensors,
-    const std::string &fetched_var_name,
-    const std::unordered_map<std::string, LoDTensor> &feed_tensors) {
+void ParallelExecutor::Run(const std::vector<std::string> &fetch_tensors,
+                           const std::string &fetched_var_name) {
   platform::RecordBlock b(0);
-  SplitTensorToPlaces(feed_tensors);
-
   // Create local scopes.
   for (auto &scope : member_->local_scopes_) {
     Scope &local_scope = scope->NewScope();
@@ -195,14 +191,28 @@ void ParallelExecutor::Run(
     auto &local_scope =
         *scope->Var(details::kLocalExecScopeName)->GetMutable<Scope *>();
     scope->DeleteScope(local_scope);
-    local_scope = nullptr;
   }
 }
 
-void ParallelExecutor::SplitTensorToPlaces(
-    const std::unordered_map<std::string, LoDTensor> &feed_tensors) {
-  for (auto it : feed_tensors) {
-    auto lod_tensors = it.second.SplitLoDTensor(member_->places_);
+void ParallelExecutor::FeedTensorsIntoLocalScopes(
+    const std::vector<std::unordered_map<std::string, LoDTensor>> &tensors) {
+  PADDLE_ENFORCE_EQ(member_->local_scopes_.size(), tensors.size());
+
+  for (size_t i = 0; i < tensors.size(); ++i) {
+    auto &map = tensors[i];
+    auto *scope = member_->local_scopes_[i];
+    for (auto &pair : map) {
+      auto *trg = scope->Var(pair.first)->GetMutable<LoDTensor>();
+      trg->ShareDataWith(pair.second);
+      trg->set_lod(pair.second.lod());
+    }
+  }
+}
+
+void ParallelExecutor::FeedAndSplitTensorIntoLocalScopes(
+    const std::unordered_map<std::string, LoDTensor> &tensors) {
+  for (auto pair : tensors) {
+    auto lod_tensors = pair.second.SplitLoDTensor(member_->places_);
     PADDLE_ENFORCE_EQ(
         member_->places_.size(), lod_tensors.size(),
         "The number of samples of current batch is less than the count of "
@@ -211,7 +221,7 @@ void ParallelExecutor::SplitTensorToPlaces(
     for (size_t j = 0; j < member_->places_.size(); ++j) {
       // TODO(panxy0718): Do I need to delete this var?
       auto t =
-          member_->local_scopes_[j]->Var(it.first)->GetMutable<LoDTensor>();
+          member_->local_scopes_[j]->Var(pair.first)->GetMutable<LoDTensor>();
       t->ShareDataWith(lod_tensors[j]);
       t->set_lod(lod_tensors[j].lod());
     }

paddle/fluid/framework/parallel_executor.h

@@ -44,16 +44,22 @@ class ParallelExecutor {
 
   std::vector<Scope*>& GetLocalScopes();
 
+  /**
+   * Feed tensors to local scopes. The size of tensors should be equal to the
+   * size of local scopes.
+   */
+  void FeedTensorsIntoLocalScopes(
+      const std::vector<std::unordered_map<std::string, LoDTensor>>& tensors);
+
+  void FeedAndSplitTensorIntoLocalScopes(
+      const std::unordered_map<std::string, LoDTensor>& tensors);
+
   void Run(const std::vector<std::string>& fetch_tensors,
-           const std::string& fetched_var_name,
-           const std::unordered_map<std::string, LoDTensor>& feed_tensors);
+           const std::string& fetched_var_name);
 
   void BCastParamsToGPUs(const std::unordered_set<std::string>& vars) const;
 
  private:
-  void SplitTensorToPlaces(
-      const std::unordered_map<std::string, LoDTensor>& feed_tensors);
-
   ParallelExecutorPrivate* member_;
 };
 

paddle/fluid/inference/CMakeLists.txt

@@ -21,7 +21,7 @@ endif()
 
 if(WITH_TESTING)
   add_subdirectory(tests/book)
-  if (WITH_TENSORRT)
+  if (TENSORRT_FOUND)
     add_subdirectory(tensorrt)
   endif()
 endif()

paddle/fluid/operators/beam_search_decode_op.cc

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/beam_search_decode_op.h"
+#include <string>
 #include "paddle/fluid/platform/device_context.h"
 
 namespace paddle {

paddle/fluid/operators/beam_search_decode_op.h

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include <vector>
 #include "paddle/fluid/framework/lod_tensor_array.h"
 #include "paddle/fluid/framework/op_registry.h"
 
@@ -87,7 +88,7 @@ struct BeamSearchDecoder {
    */
   std::vector<BeamNodeVector<T>> PackTwoSteps(
       const LoDTensor& cur_ids, const LoDTensor& cur_scores,
-      std::vector<BeamNodeVector<T>>& prefixes_list,
+      std::vector<BeamNodeVector<T>>* prefixes_list,
       std::vector<SentenceVector<T>>* sentence_vector_list) const;
 
   /**
@@ -140,7 +141,7 @@ Sentence<T> BeamSearchDecoder<T>::MakeSentence(const BeamNode<T>* node) const {
 template <typename T>
 std::vector<BeamNodeVector<T>> BeamSearchDecoder<T>::PackTwoSteps(
     const LoDTensor& cur_ids, const LoDTensor& cur_scores,
-    std::vector<BeamNodeVector<T>>& prefixes_list,
+    std::vector<BeamNodeVector<T>>* prefixes_list,
     std::vector<SentenceVector<T>>* sentence_vector_list) const {
   std::vector<BeamNodeVector<T>> result;
 
@@ -153,7 +154,7 @@ std::vector<BeamNodeVector<T>> BeamSearchDecoder<T>::PackTwoSteps(
 
     // if prefixes size is 0, it means this is the first step. In this step,
     // all candidate id is the start of candidate sentences.
-    if (prefixes_list.empty()) {
+    if (prefixes_list->empty()) {
       PADDLE_ENFORCE_EQ(cur_ids.lod().at(kSourceLevel).back(),
                         cur_ids.lod().at(kSentenceLevel).back(),
                         "in the first step");
@@ -162,7 +163,7 @@ std::vector<BeamNodeVector<T>> BeamSearchDecoder<T>::PackTwoSteps(
             cur_ids.data<int64_t>()[id_idx], cur_scores.data<T>()[id_idx])));
       }
     } else {
-      BeamNodeVector<T>& prefixes = prefixes_list[src_idx];
+      BeamNodeVector<T>& prefixes = prefixes_list->at(src_idx);
       SentenceVector<T>& sentence_vector = (*sentence_vector_list)[src_idx];
 
       PADDLE_ENFORCE_EQ(src_end - src_start, prefixes.size(),
@@ -262,7 +263,7 @@ void BeamSearchDecoder<T>::PackAllSteps(const LoDTensorArray& step_ids,
   for (size_t step_id = 0; step_id < step_num; ++step_id) {
     beamnode_vector_list =
         PackTwoSteps(step_ids.at(step_id), step_scores.at(step_id),
-                     beamnode_vector_list, &sentence_vector_list);
+                     &beamnode_vector_list, &sentence_vector_list);
   }
   // append last beam_node to result
   for (size_t src_idx = 0; src_idx < src_num; ++src_idx) {

paddle/fluid/operators/beam_search_decode_op_test.cc

@@ -125,7 +125,7 @@ TEST(BeamSearchDecodeOp, PackTwoStepsFistStep) {
 
   BeamSearchDecoder<float> helper;
   beamnode_vector_list = helper.PackTwoSteps(
-      ids[0], scores[0], beamnode_vector_list, &sentence_vector_list);
+      ids[0], scores[0], &beamnode_vector_list, &sentence_vector_list);
   ASSERT_EQ(beamnode_vector_list.size(), 2UL);
   ASSERT_EQ(beamnode_vector_list[0].size(), 2UL);
   ASSERT_EQ(beamnode_vector_list[1].size(), 4UL);
@@ -167,7 +167,7 @@ TEST(BeamSearchDecodeOp, PackTwoSteps) {
 
   BeamSearchDecoder<float> helper1;
   beamnode_vector_list = helper1.PackTwoSteps(
-      ids[0], scores[0], beamnode_vector_list, &sentence_vector_list);
+      ids[0], scores[0], &beamnode_vector_list, &sentence_vector_list);
 
   ASSERT_EQ(sentence_vector_list[0].size(), 1UL);
   ASSERT_EQ(sentence_vector_list[1].size(), 0UL);

paddle/fluid/operators/beam_search_op.cc

@@ -14,7 +14,10 @@ limitations under the License. */
 
 #include "paddle/fluid/operators/beam_search_op.h"
 
+#include <algorithm>
 #include <map>
+#include <string>
+#include <vector>
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/op_registry.h"
 

paddle/fluid/operators/beam_search_op.h

@@ -18,6 +18,8 @@ limitations under the License. */
 #include "gtest/gtest.h"
 #endif
 
+#include <string>
+#include <vector>
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/operator.h"
 

paddle/fluid/operators/chunk_eval_op.cc

@@ -13,6 +13,8 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/chunk_eval_op.h"
+#include <string>
+#include <vector>
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/chunk_eval_op.h

@@ -14,6 +14,9 @@ limitations under the License. */
 
 #pragma once
 #include <set>
+#include <string>
+#include <vector>
+
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/op_registry.h"
 
@@ -36,11 +39,11 @@ class ChunkEvalKernel : public framework::OpKernel<T> {
   };
 
   void GetSegments(const int64_t* label, int length,
-                   std::vector<Segment>& segments, int num_chunk_types,
+                   std::vector<Segment>* segments, int num_chunk_types,
                    int num_tag_types, int other_chunk_type, int tag_begin,
                    int tag_inside, int tag_end, int tag_single) const {
-    segments.clear();
-    segments.reserve(length);
+    segments->clear();
+    segments->reserve(length);
     int chunk_start = 0;
     bool in_chunk = false;
     int tag = -1;
@@ -58,7 +61,7 @@ class ChunkEvalKernel : public framework::OpKernel<T> {
             i - 1,        // end
             prev_type,
         };
-        segments.push_back(segment);
+        segments->push_back(segment);
         in_chunk = false;
       }
       if (ChunkBegin(prev_tag, prev_type, tag, type, other_chunk_type,
@@ -73,7 +76,7 @@ class ChunkEvalKernel : public framework::OpKernel<T> {
           length - 1,   // end
           type,
       };
-      segments.push_back(segment);
+      segments->push_back(segment);
     }
   }
 
@@ -177,8 +180,8 @@ class ChunkEvalKernel : public framework::OpKernel<T> {
     for (int i = 0; i < num_sequences; ++i) {
       int seq_length = lod[0][i + 1] - lod[0][i];
       EvalOneSeq(inference_data + lod[0][i], label_data + lod[0][i], seq_length,
-                 output_segments, label_segments, *num_infer_chunks_data,
-                 *num_label_chunks_data, *num_correct_chunks_data,
+                 &output_segments, &label_segments, num_infer_chunks_data,
+                 num_label_chunks_data, num_correct_chunks_data,
                  num_chunk_types, num_tag_types, other_chunk_type, tag_begin,
                  tag_inside, tag_end, tag_single, excluded_chunk_types);
     }
@@ -197,10 +200,10 @@ class ChunkEvalKernel : public framework::OpKernel<T> {
   }
 
   void EvalOneSeq(const int64_t* output, const int64_t* label, int length,
-                  std::vector<Segment>& output_segments,
-                  std::vector<Segment>& label_segments,
-                  int64_t& num_output_segments, int64_t& num_label_segments,
-                  int64_t& num_correct, int num_chunk_types, int num_tag_types,
+                  std::vector<Segment>* output_segments,
+                  std::vector<Segment>* label_segments,
+                  int64_t* num_output_segments, int64_t* num_label_segments,
+                  int64_t* num_correct, int num_chunk_types, int num_tag_types,
                   int other_chunk_type, int tag_begin, int tag_inside,
                   int tag_end, int tag_single,
                   const std::set<int>& excluded_chunk_types) const {
@@ -209,25 +212,29 @@ class ChunkEvalKernel : public framework::OpKernel<T> {
     GetSegments(label, length, label_segments, num_chunk_types, num_tag_types,
                 other_chunk_type, tag_begin, tag_inside, tag_end, tag_single);
     size_t i = 0, j = 0;
-    while (i < output_segments.size() && j < label_segments.size()) {
-      if (output_segments[i] == label_segments[j] &&
-          excluded_chunk_types.count(output_segments[i].type) != 1) {
-        ++num_correct;
+    while (i < output_segments->size() && j < label_segments->size()) {
+      if (output_segments->at(i) == label_segments->at(j) &&
+          excluded_chunk_types.count(output_segments->at(i).type) != 1) {
+        ++(*num_correct);
       }
-      if (output_segments[i].end < label_segments[j].end) {
+      if (output_segments->at(i).end < label_segments->at(j).end) {
         ++i;
-      } else if (output_segments[i].end > label_segments[j].end) {
+      } else if (output_segments->at(i).end > label_segments->at(j).end) {
         ++j;
       } else {
         ++i;
         ++j;
       }
     }
-    for (auto& segment : label_segments) {
-      if (excluded_chunk_types.count(segment.type) != 1) ++num_label_segments;
+    for (auto& segment : (*label_segments)) {
+      if (excluded_chunk_types.count(segment.type) != 1) {
+        ++(*num_label_segments);
+      }
     }
-    for (auto& segment : output_segments) {
-      if (excluded_chunk_types.count(segment.type) != 1) ++num_output_segments;
+    for (auto& segment : (*output_segments)) {
+      if (excluded_chunk_types.count(segment.type) != 1) {
+        ++(*num_output_segments);
+      }
     }
   }
 };

paddle/fluid/operators/conv_mkldnn_op.cc

@@ -72,10 +72,10 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     auto dst_md = platform::MKLDNNMemDesc(
         dst_tz, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw);
 
-    auto src_memory =
-        mkldnn::memory({src_md, mkldnn_engine}, (void*)input_data);
-    auto weights_memory =
-        mkldnn::memory({weights_md, mkldnn_engine}, (void*)filter_data);
+    auto src_memory = mkldnn::memory({src_md, mkldnn_engine},
+                                     reinterpret_cast<void*>(input_data));
+    auto weights_memory = mkldnn::memory({weights_md, mkldnn_engine},
+                                         reinterpret_cast<void*>(filter_data));
     auto dst_memory = mkldnn::memory({dst_md, mkldnn_engine}, output_data);
 
     std::shared_ptr<mkldnn::convolution_forward::primitive_desc> conv_pd =
@@ -180,8 +180,9 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
         dst_tz, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw);
 
     // create memory
-    auto diff_dst_memory = mkldnn::memory({diff_weights_md, mkldnn_engine},
-                                          (void*)output_grad_data);
+    auto diff_dst_memory =
+        mkldnn::memory({diff_weights_md, mkldnn_engine},
+                       reinterpret_cast<void*>(output_grad_data));
     // Retrieve conv_pd from device context
     auto conv_pd =
         std::static_pointer_cast<mkldnn::convolution_forward::primitive_desc>(
@@ -198,10 +199,11 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
                                       mkldnn_engine);
 
       // create memory
-      auto diff_weights_memory = mkldnn::memory(
-          {diff_weights_md, mkldnn_engine}, (void*)filter_grad_data);
-      auto src_memory =
-          mkldnn::memory({src_md, mkldnn_engine}, (void*)input_data);
+      auto diff_weights_memory =
+          mkldnn::memory({diff_weights_md, mkldnn_engine},
+                         reinterpret_cast<void*>(filter_grad_data));
+      auto src_memory = mkldnn::memory({src_md, mkldnn_engine},
+                                       reinterpret_cast<void*>(input_data));
 
       // create backward conv primitive for weights
       auto conv_bwd_weights_prim = mkldnn::convolution_backward_weights(
@@ -221,9 +223,10 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
 
       // create memory
       auto diff_src_memory =
-          mkldnn::memory({diff_src_md, mkldnn_engine}, (void*)input_grad_data);
-      auto weights_memory =
-          mkldnn::memory({weights_md, mkldnn_engine}, (void*)filter_data);
+          mkldnn::memory({diff_src_md, mkldnn_engine},
+                         reinterpret_cast<void*>(input_grad_data));
+      auto weights_memory = mkldnn::memory(
+          {weights_md, mkldnn_engine}, reinterpret_cast<void*>(filter_data));
 
       // create backward conv primitive for data
       auto conv_bwd_data_prim = mkldnn::convolution_backward_data(

paddle/fluid/operators/conv_op.h

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include <vector>
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/math/depthwise_conv.h"
@@ -41,9 +42,10 @@ inline int ConvOutputSize(int input_size, int filter_size, int dilation,
 
   return output_size;
 }
-inline bool IsExpand(std::vector<int64_t>& filter_dim,
-                     std::vector<int>& strides, std::vector<int>& paddings,
-                     std::vector<int>& dilations) {
+inline bool IsExpand(const std::vector<int64_t>& filter_dim,
+                     const std::vector<int>& strides,
+                     const std::vector<int>& paddings,
+                     const std::vector<int>& dilations) {
   bool filter_1 = true, strides_1 = true, padding_0 = true, dilation_1 = true;
   for (size_t j = 0; j < strides.size(); ++j) {
     filter_1 = filter_1 && (static_cast<int>(filter_dim[j + 2]) == 1);

paddle/fluid/operators/detection_map_op.cc

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/detection_map_op.h"
+#include <string>
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/detection_map_op.h

@@ -13,6 +13,11 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+#include <algorithm>
+#include <map>
+#include <string>
+#include <utility>
+#include <vector>
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/op_registry.h"
 
@@ -82,7 +87,7 @@ class DetectionMAPOpKernel : public framework::OpKernel<T> {
     std::vector<std::map<int, std::vector<Box>>> gt_boxes;
     std::vector<std::map<int, std::vector<std::pair<T, Box>>>> detect_boxes;
 
-    GetBoxes(*in_label, *in_detect, gt_boxes, detect_boxes);
+    GetBoxes(*in_label, *in_detect, &gt_boxes, detect_boxes);
 
     std::map<int, int> label_pos_count;
     std::map<int, std::vector<std::pair<T, int>>> true_pos;
@@ -95,20 +100,20 @@ class DetectionMAPOpKernel : public framework::OpKernel<T> {
     }
 
     if (in_pos_count != nullptr && state) {
-      GetInputPos(*in_pos_count, *in_true_pos, *in_false_pos, label_pos_count,
-                  true_pos, false_pos, class_num);
+      GetInputPos(*in_pos_count, *in_true_pos, *in_false_pos, &label_pos_count,
+                  &true_pos, &false_pos, class_num);
     }
 
     CalcTrueAndFalsePositive(gt_boxes, detect_boxes, evaluate_difficult,
-                             overlap_threshold, label_pos_count, true_pos,
-                             false_pos);
+                             overlap_threshold, &label_pos_count, &true_pos,
+                             &false_pos);
 
     int background_label = ctx.Attr<int>("background_label");
     T map = CalcMAP(ap_type, label_pos_count, true_pos, false_pos,
                     background_label);
 
-    GetOutputPos(ctx, label_pos_count, true_pos, false_pos, *out_pos_count,
-                 *out_true_pos, *out_false_pos, class_num);
+    GetOutputPos(ctx, label_pos_count, true_pos, false_pos, out_pos_count,
+                 out_true_pos, out_false_pos, class_num);
 
     T* map_data = out_map->mutable_data<T>(ctx.GetPlace());
     map_data[0] = map;
@@ -155,7 +160,7 @@ class DetectionMAPOpKernel : public framework::OpKernel<T> {
 
   void GetBoxes(const framework::LoDTensor& input_label,
                 const framework::LoDTensor& input_detect,
-                std::vector<std::map<int, std::vector<Box>>>& gt_boxes,
+                std::vector<std::map<int, std::vector<Box>>>* gt_boxes,
                 std::vector<std::map<int, std::vector<std::pair<T, Box>>>>&
                     detect_boxes) const {
     auto labels = framework::EigenTensor<T, 2>::From(input_label);
@@ -179,7 +184,7 @@ class DetectionMAPOpKernel : public framework::OpKernel<T> {
           box.is_difficult = true;
         boxes[label].push_back(box);
       }
-      gt_boxes.push_back(boxes);
+      gt_boxes->push_back(boxes);
     }
 
     auto detect_index = detect_lod[0];
@@ -200,9 +205,9 @@ class DetectionMAPOpKernel : public framework::OpKernel<T> {
       const std::map<int, int>& label_pos_count,
       const std::map<int, std::vector<std::pair<T, int>>>& true_pos,
       const std::map<int, std::vector<std::pair<T, int>>>& false_pos,
-      framework::Tensor& output_pos_count,
-      framework::LoDTensor& output_true_pos,
-      framework::LoDTensor& output_false_pos, const int class_num) const {
+      framework::Tensor* output_pos_count,
+      framework::LoDTensor* output_true_pos,
+      framework::LoDTensor* output_false_pos, const int class_num) const {
     int true_pos_count = 0;
     int false_pos_count = 0;
     for (auto it = true_pos.begin(); it != true_pos.end(); ++it) {
@@ -214,12 +219,12 @@ class DetectionMAPOpKernel : public framework::OpKernel<T> {
       false_pos_count += fp.size();
     }
 
-    int* pos_count_data = output_pos_count.mutable_data<int>(
+    int* pos_count_data = output_pos_count->mutable_data<int>(
         framework::make_ddim({class_num, 1}), ctx.GetPlace());
 
-    T* true_pos_data = output_true_pos.mutable_data<T>(
+    T* true_pos_data = output_true_pos->mutable_data<T>(
         framework::make_ddim({true_pos_count, 2}), ctx.GetPlace());
-    T* false_pos_data = output_false_pos.mutable_data<T>(
+    T* false_pos_data = output_false_pos->mutable_data<T>(
         framework::make_ddim({false_pos_count, 2}), ctx.GetPlace());
     true_pos_count = 0;
     false_pos_count = 0;
@@ -261,21 +266,21 @@ class DetectionMAPOpKernel : public framework::OpKernel<T> {
     framework::LoD false_pos_lod;
     false_pos_lod.emplace_back(false_pos_starts);
 
-    output_true_pos.set_lod(true_pos_lod);
-    output_false_pos.set_lod(false_pos_lod);
+    output_true_pos->set_lod(true_pos_lod);
+    output_false_pos->set_lod(false_pos_lod);
     return;
   }
 
   void GetInputPos(const framework::Tensor& input_pos_count,
                    const framework::LoDTensor& input_true_pos,
                    const framework::LoDTensor& input_false_pos,
-                   std::map<int, int>& label_pos_count,
-                   std::map<int, std::vector<std::pair<T, int>>>& true_pos,
-                   std::map<int, std::vector<std::pair<T, int>>>& false_pos,
+                   std::map<int, int>* label_pos_count,
+                   std::map<int, std::vector<std::pair<T, int>>>* true_pos,
+                   std::map<int, std::vector<std::pair<T, int>>>* false_pos,
                    const int class_num) const {
     const int* pos_count_data = input_pos_count.data<int>();
     for (int i = 0; i < class_num; ++i) {
-      label_pos_count[i] = pos_count_data[i];
+      (*label_pos_count)[i] = pos_count_data[i];
     }
 
     auto SetData = [](const framework::LoDTensor& pos_tensor,
@@ -291,8 +296,8 @@ class DetectionMAPOpKernel : public framework::OpKernel<T> {
       }
     };
 
-    SetData(input_true_pos, true_pos);
-    SetData(input_false_pos, false_pos);
+    SetData(input_true_pos, *true_pos);
+    SetData(input_false_pos, *false_pos);
     return;
   }
 
@@ -301,9 +306,9 @@ class DetectionMAPOpKernel : public framework::OpKernel<T> {
       const std::vector<std::map<int, std::vector<std::pair<T, Box>>>>&
           detect_boxes,
       bool evaluate_difficult, float overlap_threshold,
-      std::map<int, int>& label_pos_count,
-      std::map<int, std::vector<std::pair<T, int>>>& true_pos,
-      std::map<int, std::vector<std::pair<T, int>>>& false_pos) const {
+      std::map<int, int>* label_pos_count,
+      std::map<int, std::vector<std::pair<T, int>>>* true_pos,
+      std::map<int, std::vector<std::pair<T, int>>>* false_pos) const {
     int batch_size = gt_boxes.size();
     for (int n = 0; n < batch_size; ++n) {
       auto image_gt_boxes = gt_boxes[n];
@@ -320,10 +325,10 @@ class DetectionMAPOpKernel : public framework::OpKernel<T> {
           continue;
         }
         int label = it->first;
-        if (label_pos_count.find(label) == label_pos_count.end()) {
-          label_pos_count[label] = count;
+        if (label_pos_count->find(label) == label_pos_count->end()) {
+          (*label_pos_count)[label] = count;
         } else {
-          label_pos_count[label] += count;
+          (*label_pos_count)[label] += count;
         }
       }
     }
@@ -338,8 +343,8 @@ class DetectionMAPOpKernel : public framework::OpKernel<T> {
           int label = it->first;
           for (size_t i = 0; i < pred_boxes.size(); ++i) {
             auto score = pred_boxes[i].first;
-            true_pos[label].push_back(std::make_pair(score, 0));
-            false_pos[label].push_back(std::make_pair(score, 1));
+            (*true_pos)[label].push_back(std::make_pair(score, 0));
+            (*false_pos)[label].push_back(std::make_pair(score, 1));
           }
         }
         continue;
@@ -351,8 +356,8 @@ class DetectionMAPOpKernel : public framework::OpKernel<T> {
         if (image_gt_boxes.find(label) == image_gt_boxes.end()) {
           for (size_t i = 0; i < pred_boxes.size(); ++i) {
             auto score = pred_boxes[i].first;
-            true_pos[label].push_back(std::make_pair(score, 0));
-            false_pos[label].push_back(std::make_pair(score, 1));
+            (*true_pos)[label].push_back(std::make_pair(score, 0));
+            (*false_pos)[label].push_back(std::make_pair(score, 1));
           }
           continue;
         }
@@ -381,17 +386,17 @@ class DetectionMAPOpKernel : public framework::OpKernel<T> {
                 (!evaluate_difficult && !matched_bboxes[max_idx].is_difficult);
             if (match_evaluate_difficult) {
               if (!visited[max_idx]) {
-                true_pos[label].push_back(std::make_pair(score, 1));
-                false_pos[label].push_back(std::make_pair(score, 0));
+                (*true_pos)[label].push_back(std::make_pair(score, 1));
+                (*false_pos)[label].push_back(std::make_pair(score, 0));
                 visited[max_idx] = true;
               } else {
-                true_pos[label].push_back(std::make_pair(score, 0));
-                false_pos[label].push_back(std::make_pair(score, 1));
+                (*true_pos)[label].push_back(std::make_pair(score, 0));
+                (*false_pos)[label].push_back(std::make_pair(score, 1));
               }
             }
           } else {
-            true_pos[label].push_back(std::make_pair(score, 0));
-            false_pos[label].push_back(std::make_pair(score, 1));
+            (*true_pos)[label].push_back(std::make_pair(score, 0));
+            (*false_pos)[label].push_back(std::make_pair(score, 1));
           }
         }
       }

paddle/fluid/operators/edit_distance_op.cu

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #include <algorithm>
 #include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/edit_distance_op.h"
 #include "paddle/fluid/operators/math/math_function.h"
 #include "paddle/fluid/platform/cuda_helper.h"
 #include "paddle/fluid/platform/gpu_info.h"

paddle/fluid/operators/listen_and_serv_op.cc

@@ -12,6 +12,7 @@ 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 <fstream>
 #include <ostream>
 #include <thread>  // NOLINT
 #include <vector>
@@ -67,7 +68,7 @@ ListenAndServOp::ListenAndServOp(const std::string &type,
                                  const framework::AttributeMap &attrs)
     : OperatorBase(type, inputs, outputs, attrs) {}
 
-int ListenAndServOp::GetSelectedPort() {
+int ListenAndServOp::GetSelectedPort() const {
   return rpc_service_->GetSelectedPort();
 }
 
@@ -99,7 +100,7 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope,
   framework::Executor executor(dev_place);
   std::vector<int> block_list;
   for (size_t blkid = 1; blkid < num_blocks; ++blkid) {
-    if (blkid != prefetch_block->ID()) {
+    if (blkid != static_cast<size_t>(prefetch_block->ID())) {
       block_list.push_back(blkid);
     }
   }
@@ -121,10 +122,14 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope,
   rpc_service_->SetProgram(program);
   // start the server listening after all member initialized.
   server_thread_.reset(new std::thread(RunServer, rpc_service_));
-  // FIXME(typhoonzero): do we need to wait until the server port is ready?
+  VLOG(3) << "wait server thread to become ready...";
   sleep(5);
+  // Write to a file of server selected port for python use.
+  std::ofstream port_file;
+  port_file.open("/tmp/paddle.selected_port");
+  port_file << rpc_service_->GetSelectedPort();
+  port_file.close();
 
-  // TODO(typhoonzero): change this to a while_op for every cluster-batch.
   bool exit_flag = false;
   // Record received sparse variables, so that
   // we could reset those after execute optimize program
@@ -175,7 +180,7 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope,
     parallel_blkids.push_back(1);
     double ts = detail::GetTimestamp();
     for (size_t blkid = 2; blkid < num_blocks; ++blkid) {
-      if (blkid != prefetch_block->ID()) {
+      if (blkid != static_cast<size_t>(prefetch_block->ID())) {
         if (program->Block(blkid).Parent() != last_parent_blkid) {
           ParallelExecuteBlocks(parallel_blkids, &executor, optimize_prepared,
                                 program, &recv_scope);

paddle/fluid/operators/listen_and_serv_op.h

@@ -39,7 +39,7 @@ class ListenAndServOp : public framework::OperatorBase {
                   const framework::VariableNameMap &outputs,
                   const framework::AttributeMap &attrs);
 
-  int GetSelectedPort();
+  int GetSelectedPort() const;
 
   void Stop() override;
 

paddle/fluid/operators/send_recv_op_test.cc

@@ -139,7 +139,6 @@ void StartServerNet(bool is_sparse) {
   attrs.insert({"PrefetchBlock", prefetch_block});
   listen_and_serv_op =
       f::OpRegistry::CreateOp("listen_and_serv", {{"X", {"x1"}}}, {}, attrs);
-  LOG(INFO) << "selected port before run " << selected_port;
   listen_and_serv_op->Run(scope, place);
   LOG(INFO) << "server exit";
 }
@@ -158,16 +157,13 @@ TEST(SendRecvOp, CPUDense) {
   selected_port = static_cast<paddle::operators::ListenAndServOp *>(
                       listen_and_serv_op.get())
                       ->GetSelectedPort();
-  LOG(INFO) << "selected port " << selected_port;
   std::string endpoint = paddle::string::Sprintf("127.0.0.1:%d", selected_port);
   attrs.insert({"endpoints", std::vector<std::string>({endpoint})});
   attrs.insert({"epmap", std::vector<std::string>({endpoint})});
   auto send_op = f::OpRegistry::CreateOp(
       "send", {{"X", {"x1"}}},
       {{"Out", {"Out"}}, {"RPCClient", {"RPC_CLIENT_VAR"}}}, attrs);
-  LOG(INFO) << "before run " << endpoint;
   send_op->Run(scope, place);
-  LOG(INFO) << "end run";
 
   auto in_var = scope.Var("x1");
   auto tensor = in_var->GetMutable<f::LoDTensor>();
@@ -180,7 +176,6 @@ TEST(SendRecvOp, CPUDense) {
   for (int64_t i = 0; i < target->numel(); ++i) {
     EXPECT_EQ(expected[i] * 2, actual[i]);
   }
-  LOG(INFO) << "before stop";
   listen_and_serv_op->Stop();
   server_thread.join();
   listen_and_serv_op.reset(nullptr);
@@ -199,7 +194,6 @@ TEST(SendRecvOp, CPUSparse) {
   selected_port = static_cast<paddle::operators::ListenAndServOp *>(
                       listen_and_serv_op.get())
                       ->GetSelectedPort();
-  LOG(INFO) << "selected port " << selected_port;
   std::string endpoint = paddle::string::Sprintf("127.0.0.1:%d", selected_port);
   attrs.insert({"endpoints", std::vector<std::string>({endpoint})});
   attrs.insert({"epmap", std::vector<std::string>({endpoint})});

paddle/fluid/platform/dynload/CMakeLists.txt

 cc_library(dynamic_loader SRCS dynamic_loader.cc DEPS glog gflags enforce)
 
 list(APPEND CUDA_SRCS cublas.cc cudnn.cc curand.cc nccl.cc)
-if (WITH_TENSORRT)
+if (TENSORRT_FOUND)
   list(APPEND CUDA_SRCS tensorrt.cc)
 endif()
 

paddle/fluid/platform/float16.h

@@ -873,6 +873,11 @@ HOSTDEVICE inline bool(isfinite)(const float16& a) {
   return !((isnan)(a)) && !((isinf)(a));
 }
 
+inline std::ostream& operator<<(std::ostream& os, const float16& a) {
+  os << static_cast<float>(a);
+  return os;
+}
+
 }  // namespace platform
 }  // namespace paddle
 

paddle/fluid/platform/float16_test.cc

@@ -141,5 +141,10 @@ TEST(float16, lod_tensor_cpu) {
   }
 }
 
+TEST(float16, print) {
+  float16 a = float16(1.0f);
+  std::cout << a << std::endl;
+}
+
 }  // namespace platform
 }  // namespace paddle

paddle/fluid/pybind/pybind.cc

@@ -505,11 +505,19 @@ All parameter, weight, gradient are variables in Paddle.
                                   scope, local_scopes, allow_op_delay);
            })
       .def("bcast_params", &ParallelExecutor::BCastParamsToGPUs)
+      // NOTE: even we return a vec<Scope*>* to Python use reference policy.
+      // We still cannot get local_scope from this vector, since the element
+      // of vec<Scope*> will be freed by Python GC. We can only return Scope*
+      // one by one and mark them as reference.
       .def("local_scopes",
            [](ParallelExecutor &self) -> std::vector<Scope *> * {
              return &self.GetLocalScopes();
            },
            py::return_value_policy::reference)
+      .def("feed_tensors_into_local_scopes",
+           &ParallelExecutor::FeedTensorsIntoLocalScopes)
+      .def("feed_and_split_tensor_into_local_scopes",
+           &ParallelExecutor::FeedAndSplitTensorIntoLocalScopes)
       .def("run", &ParallelExecutor::Run);
 
   BindRecordIOWriter(&m);

paddle/fluid/pybind/tensor_py.h

@@ -190,6 +190,11 @@ void PyCUDATensorSetFromArray(
       static_cast<const platform::CUDADeviceContext *>(pool.Get(place));
   paddle::platform::GpuMemcpyAsync(dst, array.data(), sizeof(T) * array.size(),
                                    cudaMemcpyHostToDevice, dev_ctx->stream());
+  // NOTE: For safety, here wait the copy complete.
+  // It because the CPU array.data() could be destroyed after this method.
+  // If we make this method async, it could be copied data from a memory buffer
+  // that has been freed.
+  dev_ctx->Wait();
 }
 
 template <>
@@ -216,6 +221,11 @@ void PyCUDATensorSetFromArray(
   paddle::platform::GpuMemcpyAsync(dst, array.data(),
                                    sizeof(uint16_t) * array.size(),
                                    cudaMemcpyHostToDevice, dev_ctx->stream());
+  // NOTE: For safety, here wait the copy complete.
+  // It because the CPU array.data() could be destroyed after this method.
+  // If we make this method async, it could be copied data from a memory buffer
+  // that has been freed.
+  dev_ctx->Wait();
 }
 
 template <typename T>

paddle/scripts/docker/build.sh

@@ -198,7 +198,7 @@ EOF
     # run paddle version to install python packages first
     RUN apt-get update &&\
         ${NCCL_DEPS}\
-        apt-get install -y wget python-pip dmidecode python-tk && pip install -U pip && \
+        apt-get install -y wget python-pip dmidecode python-tk && pip install -U pip==9.0.3 && \
         pip install /*.whl; apt-get install -f -y && \
         apt-get clean -y && \
         rm -f /*.whl && \

python/paddle/fluid/layers/io.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 from .. import core
-from ..framework import convert_np_dtype_to_dtype_, default_main_program, default_startup_program
+from ..framework import convert_np_dtype_to_dtype_, default_main_program, default_startup_program, Program
 from ..unique_name import generate as unique_name
 from control_flow import BlockGuard
 from ..layer_helper import LayerHelper
@@ -158,6 +158,7 @@ class ListenAndServ(object):
         main_program = self.helper.main_program
         current_block = main_program.current_block()
         parent_block = self.parent_block()
+        empty_block = Program().global_block()
 
         parent_block.append_op(
             type='listen_and_serv',
@@ -166,11 +167,12 @@ class ListenAndServ(object):
             attrs={
                 'endpoint': self.endpoint,
                 'Fanin': self.fan_in,
-                'OptimizeBlock': current_block
+                'OptimizeBlock': current_block,
+                'PrefetchBlock': empty_block
             })
 
 
-def Send(endpoints, send_vars, get_vars):
+def Send(endpoints, send_vars, get_vars=None):
     """
     Send layer
 
@@ -184,7 +186,6 @@ def Send(endpoints, send_vars, get_vars):
     side when server have finished running server side program.
     """
     assert (type(send_vars) == list)
-    assert (type(get_vars) == list)
 
     epmap = endpoints.split(",")
     endpoints = list(set(epmap))
@@ -192,6 +193,11 @@ def Send(endpoints, send_vars, get_vars):
     helper = LayerHelper("Send", **locals())
     rpc_client_var = default_main_program().global_block().create_var(
         name="RPC_CLIENT_VAR", persistable=True, type=core.VarDesc.VarType.RAW)
+    if not get_vars:
+        get_vars = []
+        for s in send_vars:
+            v = helper.create_tmp_variable(dtype=s.dtype, stop_gradient=True)
+            get_vars.append(v)
 
     helper.append_op(
         type="send",
@@ -200,6 +206,7 @@ def Send(endpoints, send_vars, get_vars):
                  "RPCClient": rpc_client_var},
         attrs={"endpoints": endpoints,
                "epmap": epmap})
+    return get_vars
 
 
 def Recv(endpoints, get_vars):

python/paddle/fluid/parallel_executor.py

@@ -16,6 +16,7 @@ import core
 import multiprocessing
 import framework
 import executor
+import sys
 
 __all__ = ['ParallelExecutor']
 
@@ -123,28 +124,93 @@ class ParallelExecutor(object):
             allow_op_delay)
         self.scope = scope
 
-    def run(self, fetch_list, feed_dict={}):
+    def run(self, fetch_list, feed=None, feed_dict=None):
         """
-        :param fetch_list: A list of variable names that will be fetched.
-        :param feed_dict: A dict mapping for feed variable name to LoDTensor
-          or numpy array.
-        :return: fetched value list.
-        """
-        if not isinstance(feed_dict, dict):
-            raise TypeError("feed_dict should be a dict")
+        Run a parallel executor with fetch_list.
+
+        The feed parameter can be a dict or a list. If feed is a dict, the
+        feed data will be split into multiple devices. If feed is a list, we
+        assume the data has been splitted into multiple devices, the each
+        element in the list will be copied to each device directly.
+
+        For example, if the feed is a dict:
+        >>> exe = ParallelExecutor()
+        >>> # the image will be splitted into devices. If there is two devices
+        >>> # each device will process an image with shape (24, 1, 28, 28)
+        >>> exe.run(feed={'image': numpy.random.random(size=(48, 1, 28, 28))})
+
+        For example, if the feed is a list:
+        >>> exe = ParallelExecutor()
+        >>> # each device will process each element in the list.
+        >>> # the 1st device will process an image with shape (48, 1, 28, 28)
+        >>> # the 2nd device will process an image with shape (32, 1, 28, 28)
+        >>> #
+        >>> # you can use exe.device_count to get the device number.
+        >>> exe.run(feed=[{"image": numpy.random.random(size=(48, 1, 28, 28))},
+        >>>               {"image": numpy.random.random(size=(32, 1, 28, 28))},
+        >>>              ])
+
+
+        Args:
+            fetch_list(list): The fetched variable names
+            feed(list|dict|None): The feed variables. If the feed is a dict,
+                tensors in that dict will be splitted into each devices. If
+                the feed is a list, each element of the list will be copied
+                to each device.
+            feed_dict: Alias for feed parameter, for backward compatibility.
+                This parameter is deprecated.
 
-        feed_tensor_dict = {}
-        for i, feed_name in enumerate(feed_dict):
-            feed_tensor = feed_dict[feed_name]
-            if not isinstance(feed_tensor, core.LoDTensor):
-                feed_tensor = core.LoDTensor()
-                feed_tensor.set(feed_dict[feed_name], self._act_places[0])
-            feed_tensor_dict[feed_name] = feed_tensor
+        Returns: fetched result list.
+
+        """
+        if feed is None:
+            feed = feed_dict
+            print >> sys.stderr, "`feed_dict` is deprecated. Please use `feed=`"
+
+        if isinstance(feed, dict):
+            feed_tensor_dict = dict()
+            for feed_name in feed:
+                feed_tensor = feed[feed_name]
+                if not isinstance(feed_tensor, core.LoDTensor):
+                    feed_tensor = core.LoDTensor()
+                    # always set to CPU place, since the tensor need to be splitted
+                    # it is fast in CPU
+                    feed_tensor.set(feed[feed_name], core.CPUPlace())
+                feed_tensor_dict[feed_name] = feed_tensor
+
+            self.executor.feed_and_split_tensor_into_local_scopes(
+                feed_tensor_dict)
+        elif isinstance(feed, list) or isinstance(feed, tuple):
+            if len(feed) != len(self._act_places):
+                raise ValueError(
+                    "Feed a list of tensor, the list should be the same size as places"
+                )
+
+            res = list()
+
+            for i, each in enumerate(feed):
+                if not isinstance(each, dict):
+                    raise TypeError(
+                        "Each element of feed list should be a dict")
+                res_dict = dict()
+                for feed_name in each:
+                    tensor = each[feed_name]
+                    if not isinstance(tensor, core.LoDTensor):
+                        tmp = core.LoDTensor()
+                        tmp.set(tensor, self._act_places[i])
+                        tensor = tmp
+                    res_dict[feed_name] = tensor
+                res.append(res_dict)
+            self.executor.feed_tensors_into_local_scopes(res)
 
         fetch_var_name = '@FETCHED_VAR_NAME@'
-        self.executor.run(fetch_list, fetch_var_name, feed_tensor_dict)
+        self.executor.run(fetch_list, fetch_var_name)
         arr = self.scope.find_var(fetch_var_name).get_lod_tensor_array()
         return [arr[i] for i in range(len(arr))]
 
     def bcast_params(self):
         self.executor.bcast_params(set(self.persistable_vars))
+
+    @property
+    def device_count(self):
+        return len(self._act_places)

python/paddle/fluid/tests/unittests/CMakeLists.txt

 file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
 string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
 
-# The fully connected test is removed whe the WITH_MKLDNN flag is OFF
-# Because the fully connected layer has only one kernel (MKLDNN)
+# The MKLDNN tests are skiped when the MKLDNN flag is OFF
 if(NOT WITH_MKLDNN)
-    list(REMOVE_ITEM TEST_OPS test_fc_op)
+    foreach(src ${TEST_OPS})
+        if(${src} MATCHES ".*_mkldnn_op$")
+            list(REMOVE_ITEM TEST_OPS ${src})
+        endif()
+    endforeach()
 endif(NOT WITH_MKLDNN)
 
 if(NOT WITH_DISTRIBUTE)
@@ -62,6 +65,7 @@ list(REMOVE_ITEM TEST_OPS test_registry)
 list(REMOVE_ITEM TEST_OPS test_fetch_var)
 list(REMOVE_ITEM TEST_OPS test_parallel_op)
 list(REMOVE_ITEM TEST_OPS test_dynrnn_static_input)
+list(REMOVE_ITEM TEST_OPS test_dist_train)
 
 # tests that can be bundled together in one python process for speed.
 if(WITH_FAST_BUNDLE_TEST)
@@ -100,3 +104,4 @@ py_test_modules(test_registry MODULES test_registry)
 py_test_modules(test_fetch_var MODULES test_fetch_var)
 py_test_modules(test_dynrnn_static_input MODULES test_dynrnn_static_input)
 py_test_modules(test_parallel_op MODULES test_parallel_op)
+py_test_modules(test_dist_train MODULES test_dist_train)

python/paddle/fluid/tests/unittests/test_activation_mkldnn_op.py

+#   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.
+
+import unittest
+import numpy as np
+import paddle.fluid.core as core
+from op_test import OpTest
+from scipy.special import expit
+from test_activation_op import TestRelu, TestTanh, TestSqrt, TestAbs
+
+
+class TestMKLDNNReluDim2(TestRelu):
+    def setUp(self):
+        super(TestMKLDNNReluDim2, self).setUp()
+
+        self.attrs = {"use_mkldnn": True}
+
+
+class TestMKLDNNTanhDim2(TestTanh):
+    def setUp(self):
+        super(TestMKLDNNTanhDim2, self).setUp()
+
+        self.attrs = {"use_mkldnn": True}
+
+
+class TestMKLDNNSqrtDim2(TestSqrt):
+    def setUp(self):
+        super(TestMKLDNNSqrtDim2, self).setUp()
+
+        self.attrs = {"use_mkldnn": True}
+
+
+class TestMKLDNNAbsDim2(TestAbs):
+    def setUp(self):
+        super(TestMKLDNNAbsDim2, self).setUp()
+        self.attrs = {"use_mkldnn": True}
+
+
+class TestMKLDNNReluDim4(TestRelu):
+    def setUp(self):
+        super(TestMKLDNNReluDim4, self).setUp()
+
+        x = np.random.uniform(-1, 1, [2, 4, 3, 5]).astype("float32")
+        # The same reason with TestAbs
+        x[np.abs(x) < 0.005] = 0.02
+        out = np.maximum(x, 0)
+
+        self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
+        self.outputs = {'Out': out}
+        self.attrs = {"use_mkldnn": True}
+
+
+class TestMKLDNNTanhDim4(TestTanh):
+    def setUp(self):
+        super(TestMKLDNNTanhDim4, self).setUp()
+
+        self.inputs = {
+            'X': np.random.uniform(0.1, 1, [2, 4, 3, 5]).astype("float32")
+        }
+        self.outputs = {'Out': np.tanh(self.inputs['X'])}
+        self.attrs = {"use_mkldnn": True}
+
+
+class TestMKLDNNSqrtDim4(TestSqrt):
+    def setUp(self):
+        super(TestMKLDNNSqrtDim4, self).setUp()
+
+        self.inputs = {
+            'X': np.random.uniform(0.1, 1, [2, 4, 3, 5]).astype("float32")
+        }
+        self.outputs = {'Out': np.sqrt(self.inputs['X'])}
+        self.attrs = {"use_mkldnn": True}
+
+
+class TestMKLDNNAbsDim4(TestAbs):
+    def setUp(self):
+        super(TestMKLDNNAbsDim4, self).setUp()
+
+        x = np.random.uniform(-1, 1, [2, 4, 3, 5]).astype("float32")
+        # The same reason with TestAbs
+        x[np.abs(x) < 0.005] = 0.02
+        self.inputs = {'X': x}
+        self.outputs = {'Out': np.abs(self.inputs['X'])}
+        self.attrs = {"use_mkldnn": True}
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/tests/unittests/test_activation_op.py

@@ -1098,82 +1098,5 @@ class TestFP16Swish(TestSwish):
                 self.check_output_with_place(place, atol=1e-3)
 
 
-#--------------------test MKLDNN--------------------
-class TestMKLDNNReluDim2(TestRelu):
-    def setUp(self):
-        super(TestMKLDNNReluDim2, self).setUp()
-
-        self.attrs = {"use_mkldnn": True}
-
-
-class TestMKLDNNTanhDim2(TestTanh):
-    def setUp(self):
-        super(TestMKLDNNTanhDim2, self).setUp()
-
-        self.attrs = {"use_mkldnn": True}
-
-
-class TestMKLDNNSqrtDim2(TestSqrt):
-    def setUp(self):
-        super(TestMKLDNNSqrtDim2, self).setUp()
-
-        self.attrs = {"use_mkldnn": True}
-
-
-class TestMKLDNNAbsDim2(TestAbs):
-    def setUp(self):
-        super(TestMKLDNNAbsDim2, self).setUp()
-
-        self.attrs = {"use_mkldnn": True}
-
-
-class TestMKLDNNReluDim4(TestRelu):
-    def setUp(self):
-        super(TestMKLDNNReluDim4, self).setUp()
-
-        x = np.random.uniform(-1, 1, [2, 4, 3, 5]).astype("float32")
-        # The same reason with TestAbs
-        x[np.abs(x) < 0.005] = 0.02
-        out = np.maximum(x, 0)
-
-        self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
-        self.outputs = {'Out': out}
-        self.attrs = {"use_mkldnn": True}
-
-
-class TestMKLDNNTanhDim4(TestTanh):
-    def setUp(self):
-        super(TestMKLDNNTanhDim4, self).setUp()
-
-        self.inputs = {
-            'X': np.random.uniform(0.1, 1, [2, 4, 3, 5]).astype("float32")
-        }
-        self.outputs = {'Out': np.tanh(self.inputs['X'])}
-        self.attrs = {"use_mkldnn": True}
-
-
-class TestMKLDNNSqrtDim4(TestSqrt):
-    def setUp(self):
-        super(TestMKLDNNSqrtDim4, self).setUp()
-
-        self.inputs = {
-            'X': np.random.uniform(0.1, 1, [2, 4, 3, 5]).astype("float32")
-        }
-        self.outputs = {'Out': np.sqrt(self.inputs['X'])}
-        self.attrs = {"use_mkldnn": True}
-
-
-class TestMKLDNNAbsDim4(TestAbs):
-    def setUp(self):
-        super(TestMKLDNNAbsDim4, self).setUp()
-
-        x = np.random.uniform(-1, 1, [2, 4, 3, 5]).astype("float32")
-        # The same reason with TestAbs
-        x[np.abs(x) < 0.005] = 0.02
-        self.inputs = {'X': x}
-        self.outputs = {'Out': np.abs(self.inputs['X'])}
-        self.attrs = {"use_mkldnn": True}
-
-
 if __name__ == "__main__":
     unittest.main()

python/paddle/fluid/tests/unittests/test_conv2d_mkldnn_op.py

+# 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.
+
+import unittest
+
+from test_conv2d_op import TestConv2dOp, TestWithPad, TestWithStride
+
+
+class TestMKLDNN(TestConv2dOp):
+    def init_kernel_type(self):
+        self.use_mkldnn = True
+
+
+class TestMKLDNNWithPad(TestWithPad):
+    def init_kernel_type(self):
+        self.use_mkldnn = True
+
+
+class TestMKLDNNWithStride(TestWithStride):
+    def init_kernel_type(self):
+        self.use_mkldnn = True
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/tests/unittests/test_conv2d_op.py

@@ -373,22 +373,5 @@ class TestDepthwiseConv2(TestConv2dOp):
 #     def init_op_type(self):
 #         self.op_type = "conv_cudnn"
 
-
-#----------------Conv2dMKLDNN----------------
-class TestMKLDNN(TestConv2dOp):
-    def init_kernel_type(self):
-        self.use_mkldnn = True
-
-
-class TestMKLDNNWithPad(TestWithPad):
-    def init_kernel_type(self):
-        self.use_mkldnn = True
-
-
-class TestMKLDNNWithStride(TestWithStride):
-    def init_kernel_type(self):
-        self.use_mkldnn = True
-
-
 if __name__ == '__main__':
     unittest.main()

python/paddle/fluid/tests/unittests/test_recv_op.py → python/paddle/fluid/tests/unittests/test_dist_train.py

@@ -15,31 +15,42 @@
 import unittest
 
 import paddle.fluid as fluid
+import paddle.fluid.core as core
 import paddle.fluid.layers as layers
 import numpy
 from multiprocessing import Process
+from threading import Thread
 import os, sys
 import time
 
 
-class TestRecvOp(unittest.TestCase):
-    def no_test_send(self):
+class TestSendOp(unittest.TestCase):
+    def test_send(self):
         # Run init_serv in a thread
         place = fluid.CPUPlace()
+        # NOTE: python thread will not work here due to GIL.
         p = Process(target=self.init_serv, args=(place, ))
         p.daemon = True
         p.start()
-        time.sleep(1)
-        self.init_client(place)
+
+        time.sleep(10)
+        with open("/tmp/paddle.selected_port", "r") as fn:
+            selected_port = int(fn.readlines()[0])
+        self.init_client(place, selected_port)
+
+        self.run_local(place)
+        self.assertTrue(numpy.allclose(self.local_out, self.dist_out))
+
         # FIXME(typhoonzero): find a way to gracefully shutdown the server.
         os.system("kill -9 %d" % p.pid)
         p.join()
 
     def init_serv(self, place):
         main = fluid.Program()
+
         with fluid.program_guard(main):
             serv = layers.ListenAndServ(
-                "127.0.0.1:6174", ["X"], optimizer_mode=False)
+                "127.0.0.1:0", ["X"], optimizer_mode=False)
             with serv.do():
                 x = layers.data(
                     shape=[32, 32],
@@ -50,10 +61,29 @@ class TestRecvOp(unittest.TestCase):
                 o = layers.scale(x=x, scale=10.0)
             main.global_block().create_var(
                 name=o.name, psersistable=False, dtype=o.dtype, shape=o.shape)
+
+        self.server_exe = fluid.Executor(place)
+        self.server_exe.run(main)
+
+    def init_client(self, place, port):
+        main = fluid.Program()
+        with fluid.program_guard(main):
+            x = layers.data(
+                shape=[32, 32],
+                dtype='float32',
+                name='X',
+                append_batch_size=False)
+            fluid.initializer.Constant(value=2.3)(x, main.global_block())
+            get_var = main.global_block().create_var(
+                name="scale_0.tmp_0",  # server side var
+                dtype="float32",
+                persistable=False,
+                shape=[32, 32])
+            o = layers.Send("127.0.0.1:%d" % port, [x], [get_var])
         exe = fluid.Executor(place)
-        exe.run(main)
+        self.dist_out = exe.run(main, fetch_list=o)  # o is a list
 
-    def init_client(self, place):
+    def run_local(self, place):
         main = fluid.Program()
         with fluid.program_guard(main):
             x = layers.data(
@@ -61,10 +91,10 @@ class TestRecvOp(unittest.TestCase):
                 dtype='float32',
                 name='X',
                 append_batch_size=False)
-            fluid.initializer.Constant(value=1.0)(x, main.global_block())
-            layers.Send("127.0.0.1:6174", [x], [x])
+            fluid.initializer.Constant(value=2.3)(x, main.global_block())
+            o = layers.scale(x=x, scale=10.0)
         exe = fluid.Executor(place)
-        exe.run(main)
+        self.local_out = exe.run(main, fetch_list=[o])
 
 
 if __name__ == "__main__":

python/paddle/fluid/tests/unittests/test_lrn_mkldnn_op.py

+#   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.
+
+import unittest
+from test_lrn_op import TestLRNOp
+
+
+class TestLRNMKLDNNOp(TestLRNOp):
+    def get_attrs(self):
+        attrs = TestLRNOp.get_attrs(self)
+        attrs['use_mkldnn'] = True
+        return attrs
+
+    def test_check_output(self):
+        self.check_output(atol=0.002)
+
+
+class TestLRNMKLDNNOpWithIsTest(TestLRNMKLDNNOp):
+    def get_attrs(self):
+        attrs = TestLRNMKLDNNOp.get_attrs(self)
+        attrs['is_test'] = True
+        return attrs
+
+    def test_check_grad_normal(self):
+        def check_raise_is_test():
+            try:
+                self.check_grad(['X'], 'Out', max_relative_error=0.01)
+            except Exception as e:
+                t = \
+                "is_test attribute should be set to False in training phase."
+                if t in str(e):
+                    raise AttributeError
+
+        self.assertRaises(AttributeError, check_raise_is_test)
+
+
+if __name__ == "__main__":
+    unittest.main()

python/paddle/fluid/tests/unittests/test_lrn_op.py

@@ -87,34 +87,5 @@ class TestLRNOp(OpTest):
         self.check_grad(['X'], 'Out', max_relative_error=0.01)
 
 
-class TestLRNMKLDNNOp(TestLRNOp):
-    def get_attrs(self):
-        attrs = TestLRNOp.get_attrs(self)
-        attrs['use_mkldnn'] = True
-        return attrs
-
-    def test_check_output(self):
-        self.check_output(atol=0.002)
-
-
-class TestLRNMKLDNNOpWithIsTest(TestLRNMKLDNNOp):
-    def get_attrs(self):
-        attrs = TestLRNMKLDNNOp.get_attrs(self)
-        attrs['is_test'] = True
-        return attrs
-
-    def test_check_grad_normal(self):
-        def check_raise_is_test():
-            try:
-                self.check_grad(['X'], 'Out', max_relative_error=0.01)
-            except Exception as e:
-                t = \
-                "is_test attribute should be set to False in training phase."
-                if t in str(e):
-                    raise AttributeError
-
-        self.assertRaises(AttributeError, check_raise_is_test)
-
-
 if __name__ == "__main__":
     unittest.main()

python/paddle/fluid/tests/unittests/test_parallel_executor.py

@@ -203,31 +203,32 @@ class TestParallelExecutorBase(unittest.TestCase):
                                   iter=10,
                                   batch_size=None,
                                   allow_op_delay=False,
-                                  feed_dict={}):
+                                  feed_dict=None):
         main = fluid.Program()
         startup = fluid.Program()
+        startup.random_seed = 1  # Fix random seed
         with fluid.program_guard(main, startup):
-            loss = method(use_feed=len(feed_dict) > 0)
+            loss = method(use_feed=feed_dict is not None)
             adam = fluid.optimizer.Adam()
             adam.minimize(loss)
             if memory_opt:
                 fluid.memory_optimize(main)
-
             place = fluid.CUDAPlace(0)
             startup_exe = fluid.Executor(place)
             startup_exe.run(startup)
 
-            exe = fluid.ParallelExecutor(True, loss_name=loss.name)
+            exe = fluid.ParallelExecutor(
+                True, loss_name=loss.name, allow_op_delay=allow_op_delay)
             if batch_size is not None:
                 batch_size *= fluid.core.get_cuda_device_count()
             begin = time.time()
-            first_loss, = exe.run([loss.name], feed_dict=feed_dict)
+            first_loss, = exe.run([loss.name], feed=feed_dict)
             first_loss = numpy.array(first_loss)
 
             for i in xrange(iter):
-                exe.run([], feed_dict=feed_dict)
+                exe.run([], feed=feed_dict)
 
-            last_loss, = exe.run([loss.name], feed_dict=feed_dict)
+            last_loss, = exe.run([loss.name], feed=feed_dict)
             end = time.time()
 
             if batch_size is not None:
@@ -648,5 +649,5 @@ class TestCRFModel(unittest.TestCase):
             for i in xrange(10):
                 cur_batch = next(data)
                 print map(numpy.array,
-                          pe.run(feed_dict=feeder.feed(cur_batch),
+                          pe.run(feed=feeder.feed(cur_batch),
                                  fetch_list=[avg_cost.name]))[0]

python/paddle/fluid/tests/unittests/test_pool2d_mkldnn_op.py

+#   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.
+
+import unittest
+from test_pool2d_op import TestPool2d_Op, TestCase1, TestCase2, TestCase3, TestCase4, TestCase5
+
+
+class TestMKLDNNCase1(TestPool2d_Op):
+    def init_kernel_type(self):
+        self.use_mkldnn = True
+
+
+class TestMKLDNNCase2(TestCase1):
+    def init_kernel_type(self):
+        self.use_mkldnn = True
+
+
+class TestMKLDNNCase3(TestCase2):
+    def init_kernel_type(self):
+        self.use_mkldnn = True
+
+
+class TestMKLDNNCase4(TestCase3):
+    def init_kernel_type(self):
+        self.use_mkldnn = True
+
+
+class TestMKLDNNCase5(TestCase4):
+    def init_kernel_type(self):
+        self.use_mkldnn = True
+
+
+class TestMKLDNNCase6(TestCase5):
+    def init_kernel_type(self):
+        self.use_mkldnn = True
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/tests/unittests/test_pool2d_op.py

@@ -317,36 +317,5 @@ class TestCeilModeCase4(TestCase2):
         self.ceil_mode = True
 
 
-#--------------------test pool2d MKLDNN--------------------
-class TestMKLDNNCase1(TestPool2d_Op):
-    def init_kernel_type(self):
-        self.use_mkldnn = True
-
-
-class TestMKLDNNCase2(TestCase1):
-    def init_kernel_type(self):
-        self.use_mkldnn = True
-
-
-class TestMKLDNNCase3(TestCase2):
-    def init_kernel_type(self):
-        self.use_mkldnn = True
-
-
-class TestMKLDNNCase4(TestCase3):
-    def init_kernel_type(self):
-        self.use_mkldnn = True
-
-
-class TestMKLDNNCase5(TestCase4):
-    def init_kernel_type(self):
-        self.use_mkldnn = True
-
-
-class TestMKLDNNCase6(TestCase5):
-    def init_kernel_type(self):
-        self.use_mkldnn = True
-
-
 if __name__ == '__main__':
     unittest.main()

python/paddle/v2/reader/__init__.py

@@ -50,7 +50,7 @@ An example implementation for single item data reader creator:
         def reader():
             while True:
                 yield numpy.random.uniform(-1, 1, size=width*height)
-    return reader
+        return reader
 
 An example implementation for multiple item data reader creator:
 
@@ -60,7 +60,7 @@ An example implementation for multiple item data reader creator:
         def reader():
             while True:
                 yield numpy.random.uniform(-1, 1, size=width*height), label
-    return reader
+        return reader
 
 
 TODO(yuyang18): Should we add whole design doc here?

tools/aws_benchmarking/README.md

+# AWS benchmark testing tool
+This is an automation tool for deploying paddlepaddle benchmark testing to AWS.
+
+## Features
+
+ - subnet creation to fit just the amount of ec2 instances required.
+ - pserver and trainer ec2 instances allocation, and instance state verification
+ - nvidia-docker ready for GPU training
+ - Instances and network element garbage collection when a task is accomplished or an error occurred
+ - Test log is collected in realtime
+ - Web service for checking log or tearing down the testing setup
+ - No testing code change needed
+ - Lots of optional configuration options
+
+ ## Usages
+
+ ### Prerequisites
+
+ - You have a working AWS account
+ - You have [AWS Command Line Interface](https://aws.amazon.com/cli/) installed
+ - Your AWS cli is bind with a account which has `AmazonEC2FullAccess` permission, and it's set as default credential.
+ - You have key pair created and pem file downloaded.
+ - You have a default VPC in the region you want to run the test.
+ - You have a Security Group created for the VPC mentioned above, which allows port 22 and the port you want to expose your control web service (5436 by default)
+ - If your test is supposed to run in a GPU machine, especially a multi card GPU machine (p2, p3 series), you might need to contact amazon to raise the limit which allows no more than 1 GPU instance at a time.
+
+ ### Start a benchmark test
+
+#### Create training image
+
+*What to expect in this step:*
+
+*You will have your training logic packed with paddle runtime in a docker image, and be able to be picked up by AWS instance for training.*
+
+Training python script and PaddlePaddle runtime are supposed to be packed into one docker image. Use PaddlePaddle production images as base image and create the training images with the docker file as follows:
+
+```Dockerfile
+FROM paddlepaddle/paddle:latest-gpu
+
+ENV HOME /root
+COPY ./ /root/
+WORKDIR /root
+RUN pip install -r /root/requirements.txt
+ENTRYPOINT ["python", "my_training.py"]
+```
+
+***Please Note***
+Training nodes will run your `ENTRYPOINT` script with the following environment variables:
+
+ - `TASK_NAME`: unique name to identify this training process.
+ - `TRAINING_ROLE`: current node's role in this training process, either "PSERVER" or "TRAINER"
+ - `PSERVER_HOSTS`: comma separated value of pserver end points, I.E. "192.168.1.2:5436,192.168.1.3:5436"
+ - `PSERVERS`: same as above
+ - `TRAINERS`: trainer count
+ - `SERVER_ENDPOINT`: current server end point if the node role is a pserver
+ - `TRAINER_INDEX`: an integer to identify the index of current trainer if the node role is a trainer.
+ - `PADDLE_INIT_TRAINER_ID`: same as above
+
+ Now we have a working distributed training script which takes advantage of node environment variables and docker file to generate the training image. Run the following command:
+
+ ```bash
+ docker build -t myreponname/paddle_benchmark .
+ ```
+
+ Now you have the image built and tagged with `myreponame/paddle_benchmark`, let's push it to dockerhub so that it can be picked up by out AWS instance.
+
+ ```bash
+ docker push myreponame/paddle_benchmark
+ ```
+
+#### Create instances and start training
+
+*What to expect in this step*
+
+*you will be asked to provide some basic settings to config your training, and this tool will have your training started and monitored*
+
+Now let's start the training process:
+
+```bash
+docker run -i -v $HOME/.aws:/root/.aws -v <full path to your pem file>:/root/<key pare name>.pem \
+putcn/paddle_aws_client \
+--action create \
+--key_name <your key pare name> \
+--security_group_id <your security group id> \
+--docker_image myreponame/paddle_benchmark \
+--pserver_count 2 \
+--trainer_count 2
+```
+
+Now just wait until you see this:
+```
+master server finished init process, visit http://XXX:XXX/status to check master log
+```
+That means you can turn off your laptop and your cluster is creating instances, starting training process, collecting logs and eventually shut all pservers and trainers down when training is finished.
+
+#### Post creation operations
+
+To access the master log:
+
+```bash
+docker run -i -v $HOME/.aws:/root/.aws \
+putcn/paddle_aws_client \
+--action status \
+--master_server_public_ip <master ip> \
+--master_server_port <master port>
+```
+
+To tear down the training setup:
+
+```bash
+docker run -i -v $HOME/.aws:/root/.aws \
+putcn/paddle_aws_client \
+--action cleanup \
+--master_server_public_ip <master ip> \
+--master_server_port <master port>
+```
+
+To retrieve training logs
+TBD
+
+### Tech details
+
+*What to expect in this step*
+
+*You will understand what is happening behind the scene, and how to check the training log, how to tear down the training on the fly, etc.*
+
+Let's understand what is happening under the hood when you run above command in your laptop
+
+![alt](diagram.png)
+
+There are 4 roles in the figure above:
+ - client: your laptop
+ - master: who tasks to aws api server to create/tear down instances, and monitor training process
+ - AWS api server: the one who actually creates and manages instances
+ - pservers and trainers: training instances
+
+When you run the `docker run` command above, what it actually does is to ask aws api service to create a subnet (step 1) and a master instance (step 2), and pass all the parameters the client collected or generated (step 3). The master is kept as minimum hardware config to keep the running cost low.
+
+Then when the master is up and running, it will ask the aws api server to create the heavy lifting training instances who are expensive to run (step 4). And the master will start training process as soon as they are done initializing (step 5).
+
+Meanwhile, the master will expose a web service for client to check training log or even tear the training setup down by a web service call.
+
+if you are creating the training with client docker container, and also monitoring your aws dashboard, you will initially see a instance tagged with `ROLE=MASTER` and `TASK_NAME=<yourtask name>_master` starts, then you will see several instances tagged with `ROLE=PSERVER` and `ROLE=TRAINER` starts.
+When the training is finished, pservers and trainers will be terminated. All their logs are kept in master node's docker env.
+
+Master exposes 4 major services:
+
+ - GET `/status`: return master log
+ - GET `/logs`: return list of log file names
+ - GET `/log/<logfile name>`: return a particular log by log file name
+ - POST `/cleanup`: teardown the whole setup
+
+
+### Parameters
+
+TBD, please refer to client/cluster_launcher.py for now
+
+### Trouble shooting
+
+TBD

tools/aws_benchmarking/client/Dockerfile

+FROM python:2.7.14-stretch
+
+ENV HOME /root
+COPY ./ /root/
+WORKDIR /root
+RUN pip install -r /root/requirements.txt
+ENTRYPOINT ["python", "cluster_launcher.py"]
\ No newline at end of file

tools/aws_benchmarking/client/cluster_launcher.py

+#   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.
+
+import argparse
+import os
+import time
+import math
+import logging
+import copy
+
+import netaddr
+import boto3
+import namesgenerator
+import paramiko
+from scp import SCPClient
+import requests
+
+
+def str2bool(v):
+    if v.lower() in ('yes', 'true', 't', 'y', '1'):
+        return True
+    elif v.lower() in ('no', 'false', 'f', 'n', '0'):
+        return False
+    else:
+        raise argparse.ArgumentTypeError('Boolean value expected.')
+
+
+parser = argparse.ArgumentParser(description=__doc__)
+parser.add_argument(
+    '--key_name', type=str, default="", help="required, key pair name")
+parser.add_argument(
+    '--security_group_id',
+    type=str,
+    default="",
+    help="required, the security group id associated with your VPC")
+
+parser.add_argument(
+    '--vpc_id',
+    type=str,
+    default="",
+    help="The VPC in which you wish to run test")
+parser.add_argument(
+    '--subnet_id',
+    type=str,
+    default="",
+    help="The Subnet_id in which you wish to run test")
+
+parser.add_argument(
+    '--pserver_instance_type',
+    type=str,
+    default="c5.2xlarge",
+    help="your pserver instance type, c5.2xlarge by default")
+parser.add_argument(
+    '--trainer_instance_type',
+    type=str,
+    default="p2.8xlarge",
+    help="your trainer instance type, p2.8xlarge by default")
+
+parser.add_argument(
+    '--task_name',
+    type=str,
+    default="",
+    help="the name you want to identify your job")
+parser.add_argument(
+    '--pserver_image_id',
+    type=str,
+    default="ami-da2c1cbf",
+    help="ami id for system image, default one has nvidia-docker ready, \
+    use ami-1ae93962 for us-east-2")
+
+parser.add_argument(
+    '--pserver_command', type=str, default="", help="pserver start command")
+
+parser.add_argument(
+    '--trainer_image_id',
+    type=str,
+    default="ami-da2c1cbf",
+    help="ami id for system image, default one has nvidia-docker ready, \
+    use ami-1ae93962 for us-west-2")
+
+parser.add_argument(
+    '--trainer_command', type=str, default="", help="trainer start command")
+
+parser.add_argument(
+    '--availability_zone',
+    type=str,
+    default="us-east-2a",
+    help="aws zone id to place ec2 instances")
+
+parser.add_argument(
+    '--trainer_count', type=int, default=1, help="Trainer count")
+
+parser.add_argument(
+    '--pserver_count', type=int, default=1, help="Pserver count")
+
+parser.add_argument(
+    '--action', type=str, default="create", help="create|cleanup|status")
+
+parser.add_argument('--pem_path', type=str, help="private key file")
+
+parser.add_argument(
+    '--pserver_port', type=str, default="5436", help="pserver port")
+
+parser.add_argument(
+    '--docker_image', type=str, default="busybox", help="training docker image")
+
+parser.add_argument(
+    '--master_server_port', type=int, default=5436, help="master server port")
+
+parser.add_argument(
+    '--master_server_public_ip', type=str, help="master server public ip")
+
+parser.add_argument(
+    '--master_docker_image',
+    type=str,
+    default="putcn/paddle_aws_master:latest",
+    help="master docker image id")
+
+parser.add_argument(
+    '--no_clean_up',
+    type=str2bool,
+    default=False,
+    help="whether to clean up after training")
+
+args = parser.parse_args()
+
+logging.basicConfig(level=logging.INFO, format='%(asctime)s %(message)s')
+
+ec2client = boto3.client('ec2')
+
+
+def print_arguments():
+    print('-----------  Configuration Arguments -----------')
+    for arg, value in sorted(vars(args).iteritems()):
+        print('%s: %s' % (arg, value))
+    print('------------------------------------------------')
+
+
+def create_subnet():
+    # if no vpc id provided, list vpcs
+    logging.info("start creating subnet")
+    if not args.vpc_id:
+        logging.info("no vpc provided, trying to find the default one")
+        vpcs_desc = ec2client.describe_vpcs(
+            Filters=[{
+                "Name": "isDefault",
+                "Values": ["true", ]
+            }], )
+        if len(vpcs_desc["Vpcs"]) == 0:
+            raise ValueError('No default VPC')
+        args.vpc_id = vpcs_desc["Vpcs"][0]["VpcId"]
+        vpc_cidrBlock = vpcs_desc["Vpcs"][0]["CidrBlock"]
+
+        logging.info("default vpc fount with id %s and CidrBlock %s" %
+                     (args.vpc_id, vpc_cidrBlock))
+
+    if not vpc_cidrBlock:
+        logging.info("trying to find cidrblock for vpc")
+        vpcs_desc = ec2client.describe_vpcs(
+            Filters=[{
+                "Name": "vpc-id",
+                "Values": [args.vpc_id, ],
+            }], )
+        if len(vpcs_desc["Vpcs"]) == 0:
+            raise ValueError('No VPC found')
+        vpc_cidrBlock = vpcs_desc["Vpcs"][0]["CidrBlock"]
+        logging.info("cidrblock for vpc is %s" % vpc_cidrBlock)
+
+    # list subnets in vpc in order to create a new one
+
+    logging.info("trying to find ip blocks for new subnet")
+    subnets_desc = ec2client.describe_subnets(
+        Filters=[{
+            "Name": "vpc-id",
+            "Values": [args.vpc_id, ],
+        }], )
+
+    ips_taken = []
+    for subnet_dec in subnets_desc["Subnets"]:
+        ips_taken.append(subnet_dec["CidrBlock"])
+
+    ip_blocks_avaliable = netaddr.IPSet(
+        [vpc_cidrBlock]) ^ netaddr.IPSet(ips_taken)
+    # adding 10 addresses as buffer
+    cidr_prefix = 32 - math.ceil(
+        math.log(args.pserver_count + args.trainer_count + 10, 2))
+    if cidr_prefix <= 16:
+        raise ValueError('Too many nodes to fit in current VPC')
+
+    for ipnetwork in ip_blocks_avaliable.iter_cidrs():
+        try:
+            subnet_cidr = ipnetwork.subnet(int(cidr_prefix)).next()
+            logging.info("subnet ip block found %s" % (subnet_cidr))
+            break
+        except Exception:
+            pass
+
+    if not subnet_cidr:
+        raise ValueError(
+            'No avaliable subnet to fit required nodes in current VPC')
+
+    logging.info("trying to create subnet")
+    subnet_desc = ec2client.create_subnet(
+        CidrBlock=str(subnet_cidr),
+        VpcId=args.vpc_id,
+        AvailabilityZone=args.availability_zone)
+
+    subnet_id = subnet_desc["Subnet"]["SubnetId"]
+
+    subnet_waiter = ec2client.get_waiter('subnet_available')
+    # sleep for 1s before checking its state
+    time.sleep(1)
+    subnet_waiter.wait(SubnetIds=[subnet_id, ])
+
+    logging.info("subnet created")
+
+    logging.info("adding tags to newly created subnet")
+    ec2client.create_tags(
+        Resources=[subnet_id, ],
+        Tags=[{
+            "Key": "Task_name",
+            'Value': args.task_name
+        }])
+    return subnet_id
+
+
+def run_instances(image_id, instance_type, count=1, role="MASTER", cmd=""):
+    response = ec2client.run_instances(
+        ImageId=image_id,
+        InstanceType=instance_type,
+        MaxCount=count,
+        MinCount=count,
+        UserData=cmd,
+        DryRun=False,
+        InstanceInitiatedShutdownBehavior="stop",
+        KeyName=args.key_name,
+        Placement={'AvailabilityZone': args.availability_zone},
+        NetworkInterfaces=[{
+            'DeviceIndex': 0,
+            'SubnetId': args.subnet_id,
+            "AssociatePublicIpAddress": True,
+            'Groups': args.security_group_ids
+        }],
+        TagSpecifications=[{
+            'ResourceType': "instance",
+            'Tags': [{
+                "Key": 'Task_name',
+                "Value": args.task_name + "_master"
+            }, {
+                "Key": 'Role',
+                "Value": role
+            }]
+        }])
+
+    instance_ids = []
+    for instance in response["Instances"]:
+        instance_ids.append(instance["InstanceId"])
+
+    if len(instance_ids) > 0:
+        logging.info(str(len(instance_ids)) + " instance(s) created")
+    else:
+        logging.info("no instance created")
+    #create waiter to make sure it's running
+
+    logging.info("waiting for instance to become accessible")
+    waiter = ec2client.get_waiter('instance_status_ok')
+    waiter.wait(
+        Filters=[{
+            "Name": "instance-status.status",
+            "Values": ["ok"]
+        }, {
+            "Name": "instance-status.reachability",
+            "Values": ["passed"]
+        }, {
+            "Name": "instance-state-name",
+            "Values": ["running"]
+        }],
+        InstanceIds=instance_ids)
+
+    instances_response = ec2client.describe_instances(InstanceIds=instance_ids)
+
+    return instances_response["Reservations"][0]["Instances"]
+
+
+def generate_task_name():
+    return namesgenerator.get_random_name()
+
+
+def init_args():
+
+    if not args.task_name:
+        args.task_name = generate_task_name()
+        logging.info("task name generated %s" % (args.task_name))
+
+    if not args.pem_path:
+        args.pem_path = os.path.expanduser("~") + "/" + args.key_name + ".pem"
+    if args.security_group_id:
+        args.security_group_ids = (args.security_group_id, )
+
+
+def create():
+
+    init_args()
+
+    # create subnet
+    if not args.subnet_id:
+        args.subnet_id = create_subnet()
+
+    # create master node
+
+    master_instance_response = run_instances(
+        image_id="ami-7a05351f", instance_type="t2.nano")
+
+    logging.info("master server started")
+
+    args.master_server_public_ip = master_instance_response[0][
+        "PublicIpAddress"]
+    args.master_server_ip = master_instance_response[0]["PrivateIpAddress"]
+
+    logging.info("master server started, master_ip=%s, task_name=%s" %
+                 (args.master_server_public_ip, args.task_name))
+
+    # cp config file and pems to master node
+
+    ssh_key = paramiko.RSAKey.from_private_key_file(args.pem_path)
+    ssh_client = paramiko.SSHClient()
+    ssh_client.set_missing_host_key_policy(paramiko.AutoAddPolicy())
+    ssh_client.connect(
+        hostname=args.master_server_public_ip, username="ubuntu", pkey=ssh_key)
+
+    with SCPClient(ssh_client.get_transport()) as scp:
+        scp.put(os.path.expanduser("~") + "/" + ".aws",
+                recursive=True,
+                remote_path='/home/ubuntu/')
+        scp.put(args.pem_path,
+                remote_path='/home/ubuntu/' + args.key_name + ".pem")
+
+    logging.info("credentials and pem copied to master")
+
+    # set arguments and start docker
+    kick_off_cmd = "docker run -d -v /home/ubuntu/.aws:/root/.aws/"
+    kick_off_cmd += " -v /home/ubuntu/" + args.key_name + ".pem:/root/" + args.key_name + ".pem"
+    kick_off_cmd += " -v /home/ubuntu/logs/:/root/logs/"
+    kick_off_cmd += " -p " + str(args.master_server_port) + ":" + str(
+        args.master_server_port)
+    kick_off_cmd += " " + args.master_docker_image
+
+    args_to_pass = copy.copy(args)
+    args_to_pass.action = "serve"
+    del args_to_pass.pem_path
+    del args_to_pass.security_group_ids
+    del args_to_pass.master_docker_image
+    del args_to_pass.master_server_public_ip
+    for arg, value in sorted(vars(args_to_pass).iteritems()):
+        if value:
+            kick_off_cmd += ' --%s %s' % (arg, value)
+
+    logging.info(kick_off_cmd)
+    stdin, stdout, stderr = ssh_client.exec_command(command=kick_off_cmd)
+    return_code = stdout.channel.recv_exit_status()
+    logging.info(return_code)
+    if return_code != 0:
+        raise Exception("Error while kicking off master")
+
+    logging.info(
+        "master server finished init process, visit %s to check master log" %
+        (get_master_web_url("/status")))
+
+
+def cleanup():
+    print requests.post(get_master_web_url("/cleanup")).text
+
+
+def status():
+    print requests.post(get_master_web_url("/status")).text
+
+
+def get_master_web_url(path):
+    return "http://" + args.master_server_public_ip + ":" + str(
+        args.master_server_port) + path
+
+
+if __name__ == "__main__":
+    print_arguments()
+    if args.action == "create":
+        if not args.key_name or not args.security_group_id:
+            raise ValueError("key_name and security_group_id are required")
+        create()
+    elif args.action == "cleanup":
+        if not args.master_server_public_ip:
+            raise ValueError("master_server_public_ip is required")
+        cleanup()
+    elif args.action == "status":
+        if not args.master_server_public_ip:
+            raise ValueError("master_server_public_ip is required")
+        status()

tools/aws_benchmarking/client/requirements.txt

+netaddr==0.7.19
+boto3==1.6.21
+namesgenerator==0.3
+paramiko==2.4.1
+scp
+requests

tools/aws_benchmarking/server/Dockerfile

+FROM python:2.7.14-stretch
+
+ENV HOME /root
+COPY ./ /root/
+WORKDIR /root
+RUN pip install -r /root/requirements.txt
+ENTRYPOINT ["python", "cluster_master.py"]
\ No newline at end of file

tools/aws_benchmarking/server/pserver.sh.template

+#!/bin/bash
+docker run --network="host" -i -e "SERVER_ENDPOINT={SERVER_ENDPOINT}" -e "MASTER_ENDPOINT={MASTER_ENDPOINT}" -e "TASK_NAME={TASK_NAME}" -e "TRAINER_INDEX={TRAINER_INDEX}" -e "TRAINING_ROLE=PSERVER" -e "TRAINER_COUNT={TRAINER_COUNT}" -e "TRAINERS={TRAINER_COUNT}" -e "PSERVER_HOSTS={PSERVER_HOSTS}" -e "PSERVERS={PSERVER_HOSTS}" {DOCKER_IMAGE} {COMMAND} --device CPU
\ No newline at end of file

tools/aws_benchmarking/server/requirements.txt

+netaddr==0.7.19
+boto3==1.6.21
+namesgenerator==0.3
+paramiko==2.4.1

tools/aws_benchmarking/server/trainer.sh.template

+#!/bin/bash 
+nvidia-docker run --network="host" -i  -e "MASTER_ENDPOINT={MASTER_ENDPOINT}" -e "TASK_NAME={TASK_NAME}" -e "TRAINER_COUNT={TRAINER_COUNT}" -e "TRAINERS={TRAINER_COUNT}" -e "TRAINER_INDEX={TRAINER_INDEX}"  -e "PADDLE_INIT_TRAINER_ID={TRAINER_INDEX}" -e "TRAINING_ROLE=TRAINER"  -e "PSERVER_HOSTS={PSERVER_HOSTS}"  -e "PSERVERS={PSERVER_HOSTS}" {DOCKER_IMAGE} {COMMAND} --device GPU
\ No newline at end of file

tools/manylinux1/Dockerfile.android

@@ -37,7 +37,7 @@ RUN git config --global credential.helper store
 # Fix locales to en_US.UTF-8
 RUN localedef -i en_US -f UTF-8 en_US.UTF-8
 
-RUN pip install --upgrade pip && \
+RUN pip install --upgrade pip==9.0.3 && \
     pip install -U 'protobuf==3.1.0' && \
     pip install -U wheel sphinx && \
     pip install pre-commit