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

0e5a278c · yuyang18 · 40da8c00 · d54ad9f1 · 0e5a278c · 0e5a278c
100 changed file
--- a/Dockerfile
+++ b/Dockerfile
@@ -70,7 +70,7 @@ RUN localedef -i en_US -f UTF-8 en_US.UTF-8
 # specify sphinx version as 1.5.6 and remove -U option for [pip install -U
 # sphinx-rtd-theme] since -U option will cause sphinx being updated to newest
 # version(1.7.1 for now), which causes building documentation failed.
-RUN pip install --upgrade pip==9.0.3 && \
+RUN easy_install -U pip && \
    pip install -U wheel && \
    pip install -U docopt PyYAML sphinx==1.5.6 && \
    pip install sphinx-rtd-theme==0.1.9 recommonmark

--- a/benchmark/fluid/mnist.py
+++ b/benchmark/fluid/mnist.py
@@ -159,6 +159,7 @@ def run_benchmark(model, args):
        paddle.dataset.mnist.train(), batch_size=args.batch_size)

    accuracy = fluid.metrics.Accuracy()
+    train_exe = fluid.ParallelExecutor(use_cuda=True, loss_name=avg_cost.name)
    iters, num_samples, start_time = 0, 0, time.time()
    for pass_id in range(args.pass_num):
        accuracy.reset()
@@ -175,17 +176,20 @@ def run_benchmark(model, args):
            y_data = np.array(map(lambda x: x[1], data)).astype("int64")
            y_data = y_data.reshape([len(y_data), 1])

-            outs = exe.run(
-                fluid.default_main_program(),
+            outs = train_exe.run(
                feed={"pixel": img_data,
                      "label": y_data},
-                fetch_list=[avg_cost, batch_acc, batch_size_tensor]
+                fetch_list=[
+                    avg_cost.name, batch_acc.name, batch_size_tensor.name
+                ]
            )  # The accuracy is the accumulation of batches, but not the current batch.
-            accuracy.update(value=outs[1], weight=outs[2])
+            accuracy.update(
+                value=np.array(np.mean(outs[1])),
+                weight=np.mean(np.array(outs[2])))
            iters += 1
            num_samples += len(y_data)
-            loss = np.array(outs[0])
-            acc = np.array(outs[1])
+            loss = np.mean(np.array(outs[0]))
+            acc = np.mean(np.array(outs[1]))
            train_losses.append(loss)
            train_accs.append(acc)
            print("Pass: %d, Iter: %d, Loss: %f, Accuracy: %f" %

--- a/benchmark/fluid/resnet.py
+++ b/benchmark/fluid/resnet.py
@@ -241,6 +241,7 @@ def run_benchmark(model, args):
    exe = fluid.Executor(place)
    exe.run(fluid.default_startup_program())
    accuracy = fluid.average.WeightedAverage()
+    train_exe = fluid.ParallelExecutor(use_cuda=True, loss_name=avg_cost.name)
    if args.use_fake_data:
        data = train_reader().next()
        image = np.array(map(lambda x: x[0].reshape(dshape), data)).astype(
@@ -264,14 +265,17 @@ def run_benchmark(model, args):
                                     data)).astype('float32')
                label = np.array(map(lambda x: x[1], data)).astype('int64')
                label = label.reshape([-1, 1])
-            loss, acc, weight = exe.run(
-                fluid.default_main_program(),
+            loss, acc, weight = train_exe.run(
                feed={'data': image,
                      'label': label},
-                fetch_list=[avg_cost, batch_acc, batch_size_tensor])
+                fetch_list=[
+                    avg_cost.name, batch_acc.name, batch_size_tensor.name
+                ])
            iters += 1
            num_samples += len(label)
-            accuracy.add(value=acc, weight=weight)
+            accuracy.add(value=np.array(np.mean(acc)), weight=np.mean(weight))
+            loss = np.mean(np.array(loss))
+            acc = np.mean(np.array(acc))
            train_losses.append(loss)
            train_accs.append(acc)
            print("Pass: %d, Iter: %d, Loss: %f, Accuracy: %f" %

--- a/benchmark/fluid/vgg.py
+++ b/benchmark/fluid/vgg.py
@@ -169,6 +169,7 @@ def main():

    iters, num_samples, start_time = 0, 0, time.time()
    accuracy = fluid.average.WeightedAverage()
+    train_exe = fluid.ParallelExecutor(use_cuda=True, loss_name=avg_cost.name)
    for pass_id in range(args.pass_num):
        accuracy.reset()
        train_accs = []
@@ -184,14 +185,17 @@ def main():
            y_data = np.array(map(lambda x: x[1], data)).astype("int64")
            y_data = y_data.reshape([-1, 1])

-            loss, acc, weight = exe.run(
-                fluid.default_main_program(),
+            loss, acc, weight = train_exe.run(
                feed={"pixel": img_data,
                      "label": y_data},
-                fetch_list=[avg_cost, batch_acc, batch_size_tensor])
-            accuracy.add(value=acc, weight=weight)
+                fetch_list=[
+                    avg_cost.name, batch_acc.name, batch_size_tensor.name
+                ])
+            accuracy.add(value=np.array(np.mean(acc)), weight=np.mean(weight))
            iters += 1
            num_samples += len(y_data)
+            loss = np.mean(np.array(loss))
+            acc = np.mean(np.array(acc))
            print(
                "Pass = %d, Iter = %d, Loss = %f, Accuracy = %f" %
                (pass_id, iters, loss, acc)

--- a/cmake/external/boost.cmake
+++ b/cmake/external/boost.cmake
@@ -23,8 +23,12 @@ set(BOOST_PROJECT       "extern_boost")
 # checked that the devtools package of CentOS 6 installs boost 1.41.0.
 # So we use 1.41.0 here.
 set(BOOST_VER           "1.41.0")
-set(BOOST_TAR           "boost_1_41_0")
-set(BOOST_URL           "http://paddlepaddledeps.bj.bcebos.com/${BOOST_TAR}.tar.gz")
+if((NOT DEFINED BOOST_TAR) OR (NOT DEFINED BOOST_URL))
+    message(STATUS "use pre defined download url")
+    set(BOOST_TAR "boost_1_41_0" CACHE STRING "" FORCE)
+    set(BOOST_URL "http://paddlepaddledeps.cdn.bcebos.com/${BOOST_TAR}.tar.gz" CACHE STRING "" FORCE)
+endif()
+MESSAGE(STATUS "BOOST_TAR: ${BOOST_TAR}, BOOST_URL: ${BOOST_URL}")
 set(BOOST_SOURCES_DIR ${THIRD_PARTY_PATH}/boost)
 set(BOOST_DOWNLOAD_DIR  "${BOOST_SOURCES_DIR}/src/${BOOST_PROJECT}")
 set(BOOST_INCLUDE_DIR "${BOOST_DOWNLOAD_DIR}/${BOOST_TAR}" CACHE PATH "boost include directory." FORCE)

--- a/cmake/external/eigen.cmake
+++ b/cmake/external/eigen.cmake
@@ -21,11 +21,12 @@ else()
    ExternalProject_Add(
        extern_eigen3
        ${EXTERNAL_PROJECT_LOG_ARGS}
-        GIT_REPOSITORY  "https://github.com/RLovelett/eigen.git"
+        GIT_REPOSITORY  "https://github.com/eigenteam/eigen-git-mirror"
        # eigen on cuda9.1 missing header of math_funtions.hpp
        # https://stackoverflow.com/questions/43113508/math-functions-hpp-not-found-when-using-cuda-with-eigen
        GIT_TAG         917060c364181f33a735dc023818d5a54f60e54c
        PREFIX          ${EIGEN_SOURCE_DIR}
+        DOWNLOAD_NAME   "eigen"
        UPDATE_COMMAND  ""
        CONFIGURE_COMMAND ""
        BUILD_COMMAND     ""

--- a/cmake/external/mkldnn.cmake
+++ b/cmake/external/mkldnn.cmake
@@ -53,11 +53,9 @@ ExternalProject_Add(
    ${EXTERNAL_PROJECT_LOG_ARGS}
    DEPENDS             ${MKLDNN_DEPENDS}
    GIT_REPOSITORY      "https://github.com/01org/mkl-dnn.git"
-    GIT_TAG             "v0.14"
+    GIT_TAG             "db3424ad44901513c03a1ea31ccaacdf633fbe9f"
    PREFIX              ${MKLDNN_SOURCES_DIR}
    UPDATE_COMMAND      ""
-    # Patch MKLDNN to compile with gcc 4.8, the related issue is in intel/mkl-dnn#237.
-    PATCH_COMMAND       ${CMAKE_COMMAND} -E copy_if_different ${CMAKE_CURRENT_SOURCE_DIR}/patches/mkldnn.hpp ${MKLDNN_SOURCES_DIR}/src/extern_mkldnn/include/mkldnn.hpp
    CMAKE_ARGS          -DCMAKE_INSTALL_PREFIX=${MKLDNN_INSTALL_DIR}
    CMAKE_ARGS          -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE} 
    CMAKE_ARGS          -DMKLROOT=${MKLML_ROOT}

--- a/cmake/external/mklml.cmake
+++ b/cmake/external/mklml.cmake
@@ -27,8 +27,12 @@ ENDIF()
 INCLUDE(ExternalProject)

 SET(MKLML_PROJECT       "extern_mklml")
-SET(MKLML_VER           "mklml_lnx_2018.0.3.20180406")
-SET(MKLML_URL           "http://paddlepaddledeps.bj.bcebos.com/${MKLML_VER}.tgz")
+IF((NOT DEFINED MKLML_VER) OR (NOT DEFINED MKLML_URL))
+  MESSAGE(STATUS "use pre defined download url")
+  SET(MKLML_VER "mklml_lnx_2018.0.3.20180406" CACHE STRING "" FORCE)
+  SET(MKLML_URL "http://paddlepaddledeps.cdn.bcebos.com/${MKLML_VER}.tgz" CACHE STRING "" FORCE)
+ENDIF()
+MESSAGE(STATUS "MKLML_VER: ${MKLML_VER}, MKLML_URL: ${MKLML_URL}")
 SET(MKLML_SOURCE_DIR    "${THIRD_PARTY_PATH}/mklml")
 SET(MKLML_DOWNLOAD_DIR  "${MKLML_SOURCE_DIR}/src/${MKLML_PROJECT}")
 SET(MKLML_DST_DIR       "mklml")

--- a/cmake/external/snappy.cmake
+++ b/cmake/external/snappy.cmake
@@ -47,8 +47,6 @@ ExternalProject_Add(
                     -DCMAKE_INSTALL_LIBDIR:PATH=${SNAPPY_INSTALL_DIR}/lib
                     -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
                     -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}
-    BUILD_COMMAND   make -j8
-    INSTALL_COMMAND make install
 )

 add_library(snappy STATIC IMPORTED GLOBAL)

--- a/cmake/external/snappystream.cmake
+++ b/cmake/external/snappystream.cmake
@@ -46,8 +46,6 @@ ExternalProject_Add(
                        -DCMAKE_INSTALL_PREFIX:PATH=${SNAPPYSTREAM_INSTALL_DIR}
                        -DCMAKE_INSTALL_LIBDIR:PATH=${SNAPPYSTREAM_INSTALL_DIR}/lib
                        -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}
-        BUILD_COMMAND   make -j8
-        INSTALL_COMMAND make install
        DEPENDS snappy
 )


--- a/cmake/inference_lib.cmake
+++ b/cmake/inference_lib.cmake
@@ -98,6 +98,14 @@ elseif (WITH_MKLML)
    )
 endif()

+if(WITH_MKLDNN)
+  set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/mkldnn")
+  copy(mkldnn_lib
+    SRCS ${MKLDNN_INC_DIR} ${MKLDNN_SHARED_LIB}
+    DSTS ${dst_dir} ${dst_dir}/lib
+  )
+endif()
+
 if(NOT MOBILE_INFERENCE AND NOT RPI)
  set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/snappy")
  copy(snappy_lib
@@ -148,4 +156,30 @@ copy(string_lib
  DSTS ${dst_dir}/${module} ${dst_dir}/${module}/tinyformat
 )

+set(module "pybind")
+copy(pybind_lib
+  SRCS ${CMAKE_CURRENT_BINARY_DIR}/paddle/fluid/${module}/pybind.h
+  DSTS ${dst_dir}/${module}
+)
+
+# CMakeCache Info
+copy(cmake_cache
+  SRCS ${CMAKE_CURRENT_BINARY_DIR}/CMakeCache.txt
+  DSTS ${CMAKE_INSTALL_PREFIX})
+
 add_custom_target(inference_lib_dist DEPENDS ${inference_lib_dist_dep}) 
+
+# paddle fluid version
+execute_process(
+  COMMAND ${GIT_EXECUTABLE} log --pretty=format:%H -1
+  OUTPUT_VARIABLE PADDLE_GIT_COMMIT)
+set(version_file ${CMAKE_INSTALL_PREFIX}/version.txt)
+file(WRITE ${version_file}
+  "GIT COMMIT ID: ${PADDLE_GIT_COMMIT}\n"
+  "WITH_MKL: ${WITH_MKL}\n"
+  "WITH_GPU: ${WITH_GPU}\n")
+if(WITH_GPU)
+  file(APPEND ${version_file}
+    "CUDA version: ${CUDA_VERSION}\n"
+    "CUDNN version: v${CUDNN_MAJOR_VERSION}\n")
+endif()
--- a/doc/fluid/design/concepts/functions_operators_layers.md
+++ b/doc/fluid/design/concepts/functions_operators_layers.md
@@ -40,7 +40,7 @@ template <typename T>
 class FCOp : public OperatorBase {
 public:
  void Run(...) {
-    add(mul(Input<T>("X"), Input<T>("W")), Input<T>("b");
+    add(mul(Input<T>("X"), Input<T>("W")), Input<T>("b"));
  }
 };
 REGISTER_OP(FCOp, "fc");

--- a/doc/fluid/howto/optimization/cpu_profiling_cn.md
+++ b/doc/fluid/howto/optimization/cpu_profiling_cn.md
+# CPU性能调优
+
 此教程会介绍如何使用Python的cProfile包、Python库yep、Google perftools来进行性能分析 (profiling) 与调优（performance tuning）。

 Profling 指发现性能瓶颈。系统中的瓶颈可能和程序员开发过程中想象的瓶颈相去甚远。Tuning 指消除瓶颈。性能优化的过程通常是不断重复地 profiling 和 tuning。
@@ -8,7 +10,7 @@ PaddlePaddle 用户一般通过调用 Python API 编写深度学习程序。大
 * Python 与 C++ 混合代码的性能分析


-# Python代码的性能分析
+## Python代码的性能分析

 ### 生成性能分析文件


--- a/doc/fluid/howto/optimization/cpu_profiling_en.md
+++ b/doc/fluid/howto/optimization/cpu_profiling_en.md
+# Tune CPU performance
+
 This tutorial introduces techniques we use to profile and tune the
 CPU performance of PaddlePaddle.  We will use Python packages
 `cProfile` and `yep`, and Google's `perftools`.
@@ -14,7 +16,7 @@ the profiling and tuning of
 1. the Python code and
 1. the mixture of Python and C++ code.

-# Profiling the Python Code
+## Profiling the Python Code

 ### Generate the Performance Profiling File


--- a/doc/v2/build_and_install/pip_install_cn.rst
+++ b/doc/v2/build_and_install/pip_install_cn.rst
@@ -10,20 +10,38 @@ PaddlePaddle可以使用常用的Python包管理工具
 使用pip安装
 ------------------------------

-
-执行下面的命令即可在当前机器上安装PaddlePaddle的运行时环境，并自动下载安装依赖软件，版本为cpu_avx_openblas。
+执行下面的命令即可在当前机器上安装PaddlePaddle的运行时环境，并自动下载安装依赖软件。

  .. code-block:: bash

     pip install paddlepaddle

+当前的默认版本为0.12.0，cpu_avx_openblas，您可以通过指定版本号来安装其它版本，例如:
+
+  .. code-block:: bash
+
+      pip install paddlepaddle==0.11.0
+

-如果需要安装支持GPU的版本（cuda7.5_cudnn5_avx_openblas），需要执行：
+如果需要安装支持GPU的版本（cuda8.0_cudnn5_avx_openblas），需要执行：

  .. code-block:: bash

     pip install paddlepaddle-gpu

+当前的默认版本也是0.12.0，PaddlePaddle针对不同需求提供了更多版本的安装包，部分列表如下：
+
+=================================   ========================================
+版本号                               版本说明
+=================================   ========================================
+paddlepaddle-gpu==0.12.0            使用CUDA 8.0和cuDNN 5编译的0.12.0版本
+paddlepaddle-gpu==0.11.0.post87     使用CUDA 8.0和cuDNN 7编译的0.11.0版本
+paddlepaddle-gpu==0.11.0.post8      使用CUDA 8.0和cuDNN 5编译的0.11.0版本
+paddlepaddle-gpu==0.11.0            使用CUDA 7.5和cuDNN 5编译的0.11.0版本
+=================================   ========================================
+
+您可以在 `Release History <https://pypi.org/project/paddlepaddle-gpu/#history>`_ 中找到paddlepaddle-gpu的各个发行版本。
+
 如果需要获取并安装最新的（开发分支）PaddlePaddle，可以从我们的CI系统中下载最新的whl安装包和c-api开发包并安装，
 您可以从下面的表格中找到需要的版本：

@@ -37,12 +55,11 @@ PaddlePaddle可以使用常用的Python包管理工具
    :header: "版本说明", "cp27-cp27mu", "cp27-cp27m"
    :widths: 1, 3, 3

-    "cpu_avx_mkl", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl>`_"
-    "cpu_avx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl>`_"
-    "cpu_noavx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl>`_"
-    "cuda7.5_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl>`_"
-    "cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl>`_"
-    "cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl>`_"
+    "cpu_avx_mkl", "`paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl>`_"
+    "cpu_avx_openblas", "`paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/.lastSuccessful/paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/.lastSuccessful/paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl>`_"
+    "cpu_noavx_openblas", "`paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl>`_"
+    "cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl>`_"
+    "cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl>`_"

 .. _pip_dependency:

@@ -69,7 +86,7 @@ PaddlePaddle发布的安装包会尽量对齐 `manylinux1 <https://www.python.or
 ------------------------------

 - paddlepaddle*.whl is not a supported wheel on this platform.
-  
+
  出现这个问题的主要原因是，没有找到和当前系统匹配的paddlepaddle安装包。请检查Python版本是否为2.7系列。另外最新的pip官方源中的安装包默认是manylinux1标准，需要使用最新的pip (>9.0.0) 才可以安装。可以使用下面的命令更新您的pip：

    .. code-block:: bash

--- a/doc/v2/build_and_install/pip_install_en.rst
+++ b/doc/v2/build_and_install/pip_install_en.rst
@@ -12,20 +12,38 @@ Install using pip
 ------------------------------

 Run the following command to install PaddlePaddle on the current
-machine, it will also download requirements, the version is cpu_avx_openblas.
+machine, it will also download requirements.

  .. code-block:: bash

     pip install paddlepaddle

+the default version is 0.12.0, cpu_avx_openblas, you can specify the versions to satisfy your demands, like:

-If you wish to install GPU version (cuda7.5_cudnn5_avx_openblas), just run:
+  .. code-block:: bash
+
+      pip install paddlepaddle==0.11.0
+
+If you need to install a GPU-enabled version (cuda8.0_cudnn5_avx_openblas), you need to run:

  .. code-block:: bash

     pip install paddlepaddle-gpu

-If you wish to install the latest develop branch PaddlePaddle, 
+The default version is also 0.12.0, PaddlePaddle provides several versions of packages for different needs, as shown in the table:
+
+=================================   ========================================
+版本号                               版本说明
+=================================   ========================================
+paddlepaddle-gpu==0.12.0            0.12.0 built with CUDA 8.0 and cuDNN 5
+paddlepaddle-gpu==0.11.0.post87     0.11.0 built with CUDA 8.0 and cuDNN 7
+paddlepaddle-gpu==0.11.0.post8      0.11.0 built with CUDA 8.0 and cuDNN 5
+paddlepaddle-gpu==0.11.0            0.11.0 built with CUDA 7.5 and cuDNN 5
+=================================   ========================================
+
+You can find all versions released of paddlepaddle-gpu in `Release History <https://pypi.org/project/paddlepaddle-gpu/#history>`_ .
+
+If you wish to install the latest develop branch PaddlePaddle,
 you can download the latest whl package from our CI system. Access
 the below links, log in as guest, then click at the "Artifact"
 tab, you'll find the download link of whl packages.
@@ -40,12 +58,11 @@ If the links below shows up the login form, just click "Log in as guest" to star
    :header: "version", "cp27-cp27mu", "cp27-cp27m"
    :widths: 1, 3, 3

-    "cpu_avx_mkl", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl>`_"
-    "cpu_avx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl>`_"
-    "cpu_noavx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl>`_"
-    "cuda7.5_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl>`_"
-    "cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl>`_"
-    "cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl>`_"
+    "cpu_avx_mkl", "`paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl>`_"
+    "cpu_avx_openblas", "`paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/.lastSuccessful/paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/.lastSuccessful/paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl>`_"
+    "cpu_noavx_openblas", "`paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl>`_"
+    "cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl>`_"
+    "cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl>`_"

 .. _pip_dependency:

@@ -79,7 +96,7 @@ FAQ
 ------------------------------

 - paddlepaddle*.whl is not a supported wheel on this platform.
-  
+
  The main cause of this issue is that your current platform is
  not supported. Please check that you are using Python 2.7 series.
  Besides, pypi only supports manylinux1 standard, you'll need to

--- a/paddle/CMakeLists.txt
+++ b/paddle/CMakeLists.txt
@@ -24,6 +24,6 @@ if(NOT WITH_FLUID_ONLY)
 endif()

 add_subdirectory(testing)
-if(NOT MOBILE_INFERENCE AND NOT RPI)
+if(NOT MOBILE_INFERENCE AND NOT RPI AND NOT WITH_C_API)
  add_subdirectory(fluid)
 endif()
--- a/contrib/float16/.gitignore
+++ b/contrib/float16/.gitignore
--- a/contrib/float16/README.md
+++ b/contrib/float16/README.md
--- a/contrib/float16/float16_benchmark.md
+++ b/contrib/float16/float16_benchmark.md
--- a/contrib/float16/float16_inference_demo.py
+++ b/contrib/float16/float16_inference_demo.py
--- a/contrib/float16/float16_transpiler.py
+++ b/contrib/float16/float16_transpiler.py
--- a/contrib/float16/run_float16_demo.sh
+++ b/contrib/float16/run_float16_demo.sh
--- a/contrib/inference/README.md
+++ b/contrib/inference/README.md
--- a/contrib/inference/paddle_inference_api.h
+++ b/contrib/inference/paddle_inference_api.h
--- a/paddle/fluid/framework/data_device_transform.cc
+++ b/paddle/fluid/framework/data_device_transform.cc
@@ -36,9 +36,11 @@ void TransDataDevice(const Tensor& in, const platform::Place& dst_place,
  VLOG(3) << "DeviceTransform in, src_place " << in.place()
          << " dst_place: " << dst_place;
  auto* dev_ctx = GetDeviceContext(in.place(), dst_place);
-  dev_ctx->Wait();
+
  TensorCopy(in, dst_place, *dev_ctx, out);
-  dev_ctx->Wait();
+  if (platform::is_gpu_place(in.place()) && platform::is_cpu_place(dst_place)) {
+    dev_ctx->Wait();
+  }
 }

 }  // namespace framework

--- a/paddle/fluid/framework/data_type.cc
+++ b/paddle/fluid/framework/data_type.cc
@@ -58,6 +58,7 @@ static DataTypeMap* InitDataTypeMap() {
  RegType(bool, proto::VarType::BOOL);
  RegType(size_t, proto::VarType::SIZE_T);
  RegType(int16_t, proto::VarType::INT16);
+  RegType(uint8_t, proto::VarType::UINT8);

 #undef RegType
  return retv;

--- a/paddle/fluid/framework/data_type.h
+++ b/paddle/fluid/framework/data_type.h
@@ -47,8 +47,14 @@ inline void VisitDataType(proto::VarType::Type type, Visitor visitor) {
    case proto::VarType::BOOL:
      visitor.template operator()<bool>();
      break;
+    case proto::VarType::UINT8:
+      visitor.template operator()<uint8_t>();
+      break;
+    case proto::VarType::INT16:
+      visitor.template operator()<int16_t>();
+      break;
    default:
-      PADDLE_THROW("Not supported");
+      PADDLE_THROW("Not supported %d", type);
  }
 }


--- a/paddle/fluid/framework/data_type_transform.cc
+++ b/paddle/fluid/framework/data_type_transform.cc
@@ -91,6 +91,12 @@ void TransDataType(const OpKernelType& kernel_type_for_var,
    case proto::VarType::BOOL:
      framework::VisitDataType(dst_type, CastDataType<bool>(in, out, ctx));
      break;
+    case proto::VarType::INT16:
+      framework::VisitDataType(dst_type, CastDataType<bool>(in, out, ctx));
+      break;
+    case proto::VarType::UINT8:
+      framework::VisitDataType(dst_type, CastDataType<bool>(in, out, ctx));
+      break;
    default:
      PADDLE_THROW("Not support type %d", src_type);
  }

--- a/paddle/fluid/framework/details/fetch_op_handle.cc
+++ b/paddle/fluid/framework/details/fetch_op_handle.cc
@@ -48,17 +48,18 @@ void FetchOpHandle::RunImpl() {
  WaitInputVarGenerated(platform::CPUPlace());

  tensors_.resize(inputs_.size());
-  auto *var_handle = static_cast<VarHandle *>(inputs_[0]);
-  auto &var_name = var_handle->name_;
  platform::CPUPlace cpu;
  auto &scopes = *local_scopes_;

-  for (size_t i = 0; i < scopes.size(); ++i) {
-    auto &scope = scopes[i];
-    auto *var =
-        scope->FindVar(kLocalExecScopeName)->Get<Scope *>()->FindVar(var_name);
+  for (size_t i = 0; i < inputs_.size(); ++i) {
+    auto *var_handle = static_cast<VarHandle *>(inputs_[i]);
+    auto &scope = scopes.at(var_handle->scope_idx_);
+    auto *var = scope->FindVar(kLocalExecScopeName)
+                    ->Get<Scope *>()
+                    ->FindVar(var_handle->name_);
    PADDLE_ENFORCE_NOT_NULL(var, "Cannot find variable %s in execution scope",
-                            var_name);
+                            var_handle->name_);
+
    auto &t = var->Get<framework::LoDTensor>();
    if (platform::is_gpu_place(t.place())) {
 #ifdef PADDLE_WITH_CUDA

--- a/paddle/fluid/framework/details/multi_devices_graph_builder.cc
+++ b/paddle/fluid/framework/details/multi_devices_graph_builder.cc
@@ -98,7 +98,7 @@ bool MultiDevSSAGraphBuilder::IsDistTrainOp(const OpDesc &op,
    return false;
  };

-  if (op.Type() == "split") {
+  if (op.Type() == "split" || op.Type() == "split_byref") {
    return checker(op.OutputArgumentNames(), send_op->InputArgumentNames());
  } else if (op.Type() == "concat") {
    return checker(op.InputArgumentNames(), send_op->OutputArgumentNames());

--- a/paddle/fluid/framework/details/op_handle_base.h
+++ b/paddle/fluid/framework/details/op_handle_base.h
@@ -70,6 +70,14 @@ class OpHandleBase {

  const std::vector<VarHandleBase *> &Inputs() const { return inputs_; }

+  size_t NoDupInputSize() const {
+    std::unordered_set<VarHandleBase *> res;
+    for (auto *var : inputs_) {
+      res.emplace(var);
+    }
+    return res.size();
+  }
+
  const std::vector<VarHandleBase *> &Outputs() const { return outputs_; }

 protected:

--- a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc
+++ b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc
@@ -174,7 +174,7 @@ void ThreadedSSAGraphExecutor::InsertFetchOps(
 void ThreadedSSAGraphExecutor::InsertPendingOp(
    std::unordered_map<OpHandleBase *, size_t> *pending_ops,
    OpHandleBase *op_instance) const {
-  pending_ops->insert({op_instance, op_instance->Inputs().size()});
+  pending_ops->insert({op_instance, op_instance->NoDupInputSize()});
 }

 void ThreadedSSAGraphExecutor::InsertPendingVar(

--- a/paddle/fluid/framework/executor.cc
+++ b/paddle/fluid/framework/executor.cc
@@ -228,7 +228,8 @@ static bool has_fetch_operators(
 void Executor::Run(const ProgramDesc& program, Scope* scope,
                   std::map<std::string, const LoDTensor*>* feed_targets,
                   std::map<std::string, LoDTensor*>* fetch_targets,
-                   bool create_vars, const std::string& feed_holder_name,
+                   bool create_local_scope, bool create_vars,
+                   const std::string& feed_holder_name,
                   const std::string& fetch_holder_name) {
  platform::RecordBlock b(kProgramId);
  bool has_feed_ops =
@@ -290,8 +291,9 @@ void Executor::Run(const ProgramDesc& program, Scope* scope,
  }

  auto ctx = Prepare(*copy_program, 0);
-  RunPreparedContext(ctx.get(), scope, feed_targets, fetch_targets, create_vars,
-                     feed_holder_name, fetch_holder_name);
+  RunPreparedContext(ctx.get(), scope, feed_targets, fetch_targets,
+                     create_local_scope, create_vars, feed_holder_name,
+                     fetch_holder_name);
 }

 std::unique_ptr<ExecutorPrepareContext> Executor::Prepare(
@@ -366,8 +368,9 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
 void Executor::RunPreparedContext(
    ExecutorPrepareContext* ctx, Scope* scope,
    std::map<std::string, const LoDTensor*>* feed_targets,
-    std::map<std::string, LoDTensor*>* fetch_targets, bool create_vars,
-    const std::string& feed_holder_name, const std::string& fetch_holder_name) {
+    std::map<std::string, LoDTensor*>* fetch_targets, bool create_local_scope,
+    bool create_vars, const std::string& feed_holder_name,
+    const std::string& fetch_holder_name) {
  auto& global_block = ctx->prog_.Block(ctx->block_id_);

  PADDLE_ENFORCE(
@@ -387,7 +390,7 @@ void Executor::RunPreparedContext(
    }
  }

-  RunPreparedContext(ctx, scope, create_vars, create_vars);
+  RunPreparedContext(ctx, scope, create_local_scope, create_vars);

  // obtain the data of fetch_targets from fetch_holder
  for (auto* op : global_block.AllOps()) {

--- a/paddle/fluid/framework/executor.h
+++ b/paddle/fluid/framework/executor.h
@@ -57,7 +57,7 @@ class Executor {
  void Run(const ProgramDesc& program, Scope* scope,
           std::map<std::string, const LoDTensor*>* feed_targets,
           std::map<std::string, LoDTensor*>* fetch_targets,
-           bool create_vars = true,
+           bool create_local_scope = true, bool create_vars = true,
           const std::string& feed_holder_name = "feed",
           const std::string& fetch_holder_name = "fetch");

@@ -76,6 +76,7 @@ class Executor {
  void RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
                          std::map<std::string, const LoDTensor*>* feed_targets,
                          std::map<std::string, LoDTensor*>* fetch_targets,
+                          bool create_local_scope = true,
                          bool create_vars = true,
                          const std::string& feed_holder_name = "feed",
                          const std::string& fetch_holder_name = "fetch");

--- a/paddle/fluid/framework/framework.proto
+++ b/paddle/fluid/framework/framework.proto
@@ -103,6 +103,7 @@ message VarType {
    FP64 = 6;
    // Tensor<size_t> is used in C++.
    SIZE_T = 19;
+    UINT8 = 20;

    // Other types that may need additional descriptions
    LOD_TENSOR = 7;

--- a/paddle/fluid/framework/lod_tensor_test.cc
+++ b/paddle/fluid/framework/lod_tensor_test.cc
@@ -228,11 +228,12 @@ TEST(LoD, CheckAbsLoD) {
  ASSERT_FALSE(CheckAbsLoD(abs_lod0));
 }

-TEST(LoDTensor, RecordIO) {
+template <typename T>
+static void TestRecordIO() {
  LoDTensor tensor;
-  int* tmp = tensor.mutable_data<int>(make_ddim({4, 5}), platform::CPUPlace());
+  T* tmp = tensor.mutable_data<T>(make_ddim({4, 5}), platform::CPUPlace());
  for (int i = 0; i < 20; ++i) {
-    tmp[i] = i;
+    tmp[i] = static_cast<T>(i);
  }

  std::stringstream* stream = new std::stringstream();
@@ -247,7 +248,7 @@ TEST(LoDTensor, RecordIO) {

  auto assert_tensor_ok = [](const LoDTensor& tensor) {
    for (int i = 0; i < 20; ++i) {
-      ASSERT_EQ(tensor.data<int>()[i], i);
+      ASSERT_EQ(tensor.data<T>()[i], static_cast<T>(i));
    }
  };

@@ -265,5 +266,13 @@ TEST(LoDTensor, RecordIO) {
  }
 }

+TEST(LoDTensor, RecordIO) {
+  TestRecordIO<int>();
+  TestRecordIO<int16_t>();
+  TestRecordIO<uint8_t>();
+  TestRecordIO<float>();
+  TestRecordIO<double>();
+}
+
 }  // namespace framework
 }  // namespace paddle
--- a/paddle/fluid/framework/operator.h
+++ b/paddle/fluid/framework/operator.h
@@ -33,7 +33,6 @@ limitations under the License. */
 #include "paddle/fluid/framework/tensor.h"
 #include "paddle/fluid/platform/device_context.h"
 #include "paddle/fluid/platform/variant.h"
-#include "paddle/utils/Error.h"

 namespace paddle {
 namespace framework {

--- a/paddle/fluid/inference/tensorrt/convert/op_converter.h
+++ b/paddle/fluid/inference/tensorrt/convert/op_converter.h
@@ -49,7 +49,7 @@ class OpConverter {
  // convert fluid block to tensorrt network
  void ConvertBlock(const framework::proto::BlockDesc& block,
                    TensorRTEngine* engine) {
-    for (size_t i = 0; i < block.ops_size(); i++) {
+    for (int i = 0; i < block.ops_size(); i++) {
      const auto& op = block.ops(i);
      OpConverter::Run(op, engine);
    }

--- a/paddle/fluid/inference/tests/test_helper.h
+++ b/paddle/fluid/inference/tests/test_helper.h
@@ -149,7 +149,7 @@ void TestInference(const std::string& dirname,
    state = paddle::platform::ProfilerState::kCPU;
  } else {
 #ifdef PADDLE_WITH_CUDA
-    state = paddle::platform::ProfilerState::kCUDA;
+    state = paddle::platform::ProfilerState::kAll;
    // The default device_id of paddle::platform::CUDAPlace is 0.
    // Users can get the device_id using:
    //   int device_id = place.GetDeviceId();
@@ -172,7 +172,7 @@ void TestInference(const std::string& dirname,
  }
  // Disable the profiler and print the timing information
  paddle::platform::DisableProfiler(paddle::platform::EventSortingKey::kDefault,
-                                    "load_program_profiler.txt");
+                                    "load_program_profiler");
  paddle::platform::ResetProfiler();

  // 3. Get the feed_target_names and fetch_target_names
@@ -208,10 +208,10 @@ void TestInference(const std::string& dirname,
    if (PrepareContext) {
      ctx = executor.Prepare(*inference_program, 0);
      executor.RunPreparedContext(ctx.get(), scope, &feed_targets,
-                                  &fetch_targets, CreateVars);
+                                  &fetch_targets, true, CreateVars);
    } else {
      executor.Run(*inference_program, scope, &feed_targets, &fetch_targets,
-                   CreateVars);
+                   true, CreateVars);
    }

    // Enable the profiler
@@ -236,8 +236,7 @@ void TestInference(const std::string& dirname,

    // Disable the profiler and print the timing information
    paddle::platform::DisableProfiler(
-        paddle::platform::EventSortingKey::kDefault,
-        "run_inference_profiler.txt");
+        paddle::platform::EventSortingKey::kDefault, "run_inference_profiler");
    paddle::platform::ResetProfiler();
  }


--- a/paddle/fluid/operators/CMakeLists.txt
+++ b/paddle/fluid/operators/CMakeLists.txt
@@ -186,11 +186,7 @@ endif()

 add_subdirectory(detail)
 if(WITH_DISTRIBUTE)
-    if(WITH_GPU)
-        op_library(gen_nccl_id_op DEPS nccl_common)
-    else()
-        set(DEPS_OPS ${DEPS_OPS} gen_nccl_id_op)
-    endif()
+    
    set(DISTRIBUTE_DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf)
    set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
    op_library(send_op DEPS ${DISTRIBUTE_DEPS})
@@ -207,7 +203,14 @@ if(WITH_DISTRIBUTE)
    set_source_files_properties(send_barrier_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
    set_source_files_properties(send_recv_op_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
    cc_test(test_send_recv SRCS send_recv_op_test.cc DEPS prefetch_op send_op listen_and_serv_op sum_op executor)
-    cc_test(test_send_nccl_id SRCS test_send_nccl_id.cc DEPS send_op listen_and_serv_op executor)
+    if(WITH_GPU)
+        set_source_files_properties(test_send_nccl_id.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
+        cc_test(test_send_nccl_id SRCS test_send_nccl_id.cc DEPS send_op listen_and_serv_op executor)
+        op_library(gen_nccl_id_op DEPS nccl_common sendrecvop_grpc)
+        set_source_files_properties(gen_nccl_id_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
+    else()
+        set(DEPS_OPS ${DEPS_OPS} gen_nccl_id_op)
+    endif()
 else()
    set(DEPS_OPS ${DEPS_OPS} send_op prefetch_op recv_op listen_and_serv_op send_vars_op send_barrier_op gen_nccl_id_op)
 endif()

--- a/paddle/fluid/operators/activation_mkldnn_op.cc
+++ b/paddle/fluid/operators/activation_mkldnn_op.cc
@@ -15,6 +15,7 @@
 #include "mkldnn.hpp"
 #include "paddle/fluid/operators/activation_op.h"
 #include "paddle/fluid/operators/mkldnn_activation_op.h"
+#include "paddle/fluid/platform/mkldnn_helper.h"

 namespace paddle {
 namespace operators {
@@ -23,6 +24,18 @@ using paddle::framework::Tensor;
 using paddle::platform::MKLDNNDeviceContext;

 namespace {
+std::string gethash(const mkldnn::memory::dims &operand_dims,
+                    const mkldnn::algorithm algorithm) {
+  auto dim2str = [](const mkldnn::memory::dims &operand_dims) {
+    std::string dstr = "";
+    for (size_t i = 0; i < operand_dims.size(); ++i) {
+      dstr += std::to_string(operand_dims[i]) + "-";
+    }
+    return dstr;
+  };
+  return dim2str(operand_dims) + std::to_string(algorithm);
+}
+
 template <typename T, typename ExecContext>
 void eltwise_forward(const ExecContext &ctx, mkldnn::algorithm algorithm,
                     const T alpha = 0, const T beta = 0) {
@@ -37,42 +50,70 @@ void eltwise_forward(const ExecContext &ctx, mkldnn::algorithm algorithm,
  const auto *src_data = src->template data<T>();

  auto *dst = ctx.template Output<Tensor>("Out");
-  const T *dst_data = dst->template mutable_data<T>(ctx.GetPlace());
+  T *dst_data = dst->template mutable_data<T>(ctx.GetPlace());

  // get memory dim
  PADDLE_ENFORCE(src->dims().size() == 2 || src->dims().size() == 4,
                 "Input dim must be with 2 or 4");
  std::vector<int> src_tz = framework::vectorize2int(src->dims());

-  // create memory description
-  auto data_md = src_tz.size() == 2
-                     ? platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32,
-                                               mkldnn::memory::format::nc)
-                     : platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32,
-                                               mkldnn::memory::format::nchw);
-
-  // create memory primitives
-  auto src_memory =
-      mkldnn::memory({data_md, mkldnn_engine},
-                     static_cast<void *>(const_cast<float *>(src_data)));
-  auto dst_memory =
-      mkldnn::memory({data_md, mkldnn_engine},
-                     static_cast<void *>(const_cast<float *>(dst_data)));
-
-  auto forward_desc = mkldnn::eltwise_forward::desc(
-      mkldnn::prop_kind::forward_training, algorithm, data_md, alpha, beta);
-
-  // save prim desc into global device context to be referred in backward path
-  const std::string key = ctx.op().Output("Out");
-  const std::string key_eltwise_pd = key + "@eltwise_pd";
-  auto forward_pd = std::make_shared<mkldnn::eltwise_forward::primitive_desc>(
-      forward_desc, mkldnn_engine);
-  dev_ctx.SetBlob(key_eltwise_pd, forward_pd);
-
-  auto eltwise = mkldnn::eltwise_forward(*forward_pd, src_memory, dst_memory);
+  const std::string key = gethash(src_tz, algorithm);
+  const std::string key_src_data =
+      key + ctx.op().Output("Out") + "@eltwise_fwd_src_data";
+  const std::string key_src_mem = key + "@eltwise_fwd_src_mem";
+  const std::string key_dst_mem = key + "@eltwise_fwd_dst_mem";
+  const std::string key_fwd = key + "@eltwise_fwd";
+
+  auto p_fwd = std::static_pointer_cast<mkldnn::eltwise_forward>(
+      dev_ctx.GetBlob(key_fwd));
+
+  // save input data to be referred in backward path
+  auto p_src_data = std::make_shared<const T *>(src_data);
+  dev_ctx.SetBlob(key_src_data, p_src_data);
+
+  if (p_fwd == nullptr) {
+    // create memory description
+    auto data_md = src_tz.size() == 2
+                       ? platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32,
+                                                 mkldnn::memory::format::nc)
+                       : platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32,
+                                                 mkldnn::memory::format::nchw);
+
+    // create memory primitives
+    auto p_src_mem = std::make_shared<mkldnn::memory>(mkldnn::memory(
+        {data_md, mkldnn_engine}, platform::to_void_cast(src_data)));
+    dev_ctx.SetBlob(key_src_mem, p_src_mem);
+
+    auto p_dst_mem = std::make_shared<mkldnn::memory>(mkldnn::memory(
+        {data_md, mkldnn_engine}, platform::to_void_cast(dst_data)));
+    dev_ctx.SetBlob(key_dst_mem, p_dst_mem);
+
+    auto fwd_desc = mkldnn::eltwise_forward::desc(
+        mkldnn::prop_kind::forward_training, algorithm, data_md, alpha, beta);
+    auto p_fwd_pd = std::make_shared<mkldnn::eltwise_forward::primitive_desc>(
+        fwd_desc, mkldnn_engine);
+    const std::string key_fwd_pd = key + "eltwise_fwd_pd";
+    dev_ctx.SetBlob(key_fwd_pd, p_fwd_pd);
+    p_fwd = std::make_shared<mkldnn::eltwise_forward>(
+        *p_fwd_pd, *(p_src_mem.get()), *(p_dst_mem.get()));
+    dev_ctx.SetBlob(key_fwd, p_fwd);
+  } else {
+    // primitives already exist
+    auto p_src_mem =
+        std::static_pointer_cast<mkldnn::memory>(dev_ctx.GetBlob(key_src_mem));
+    PADDLE_ENFORCE(p_src_mem != nullptr,
+                   "Fail to find eltwise p_src_mem in device context.");
+    auto p_dst_mem =
+        std::static_pointer_cast<mkldnn::memory>(dev_ctx.GetBlob(key_dst_mem));
+    PADDLE_ENFORCE(p_dst_mem != nullptr,
+                   "Fail to find eltwise p_src_mem in device context.");
+
+    p_src_mem->set_data_handle(platform::to_void_reinterpret_cast(src_data));
+    p_dst_mem->set_data_handle(dst_data);
+  }

  // push primitive to stream and wait until it's executed
-  std::vector<mkldnn::primitive> pipeline = {eltwise};
+  std::vector<mkldnn::primitive> pipeline = {*(p_fwd.get())};
  mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
 }

@@ -83,8 +124,7 @@ void eltwise_grad(const ExecContext &ctx, mkldnn::algorithm algorithm,
  const auto &mkldnn_engine = dev_ctx.GetEngine();

  // get buffers
-  const auto *x = ctx.template Input<Tensor>("X");
-  const auto *src = x->template data<T>();
+  const auto *out = ctx.template Input<Tensor>("Out");

  auto *dout = ctx.template Input<Tensor>(framework::GradVarName("Out"));
  const auto *diff_dst = dout->template data<T>();
@@ -94,45 +134,73 @@ void eltwise_grad(const ExecContext &ctx, mkldnn::algorithm algorithm,
  const T *diff_src = dx->template mutable_data<T>(ctx.GetPlace());

  // get memory dim
-  std::vector<int> src_tz = framework::vectorize2int(x->dims());
-
-  // create memory description
-  auto data_md = src_tz.size() == 2
-                     ? platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32,
-                                               mkldnn::memory::format::nc)
-                     : platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32,
-                                               mkldnn::memory::format::nchw);
-
-  // create memory primitives
-  auto src_memory = mkldnn::memory(
-      {data_md, mkldnn_engine}, static_cast<void *>(const_cast<float *>(src)));
-  auto diff_src_memory =
-      mkldnn::memory({data_md, mkldnn_engine},
-                     static_cast<void *>(const_cast<float *>(diff_src)));
-  auto diff_dst_memory =
-      mkldnn::memory({data_md, mkldnn_engine},
-                     static_cast<void *>(const_cast<float *>(diff_dst)));
-
-  auto backward_desc =
-      mkldnn::eltwise_backward::desc(algorithm, data_md, data_md, alpha, beta);
-
-  // retrieve eltwise primitive desc from device context
-  const std::string key = ctx.op().Input("Out");
-  const std::string key_eltwise_pd = key + "@eltwise_pd";
-  const std::shared_ptr<void> forward_pd = dev_ctx.GetBlob(key_eltwise_pd);
-  PADDLE_ENFORCE(forward_pd != nullptr,
-                 "Fail to find eltwise_pd in device context");
-  auto *p_forward_pd =
-      static_cast<mkldnn::eltwise_forward::primitive_desc *>(forward_pd.get());
-
-  auto eltwise_bwd_prim_desc = mkldnn::eltwise_backward::primitive_desc(
-      backward_desc, mkldnn_engine, *p_forward_pd);
-
-  auto eltwise_bwd = mkldnn::eltwise_backward(eltwise_bwd_prim_desc, src_memory,
-                                              diff_dst_memory, diff_src_memory);
+  std::vector<int> src_tz = framework::vectorize2int(out->dims());
+
+  const std::string key = gethash(src_tz, algorithm);
+  const std::string key_diff_src_mem = key + "@eltwise_diff_src_mem";
+  const std::string key_diff_dst_mem = key + "@eltwise_diff_dst_mem";
+  const std::string key_grad = key + "@eltwise_grad";
+
+  const std::string key_src_data =
+      key + ctx.op().Input("Out") + "@eltwise_fwd_src_data";
+  const auto p_src_data =
+      std::static_pointer_cast<T *>(dev_ctx.GetBlob(key_src_data));
+
+  const std::string key_src_mem = key + "@eltwise_fwd_src_mem";
+  auto p_src_mem =
+      std::static_pointer_cast<mkldnn::memory>(dev_ctx.GetBlob(key_src_mem));
+  p_src_mem->set_data_handle(*p_src_data.get());
+
+  auto p_grad = std::static_pointer_cast<mkldnn::eltwise_forward::primitive>(
+      dev_ctx.GetBlob(key_grad));
+
+  if (p_grad == nullptr) {
+    // create memory description
+    auto data_md = src_tz.size() == 2
+                       ? platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32,
+                                                 mkldnn::memory::format::nc)
+                       : platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32,
+                                                 mkldnn::memory::format::nchw);
+
+    // create memory primitives
+    std::shared_ptr<void> p_diff_src_mem =
+        std::make_shared<mkldnn::memory>(mkldnn::memory(
+            {data_md, mkldnn_engine}, platform::to_void_cast(diff_src)));
+    dev_ctx.SetBlob(key_diff_src_mem, p_diff_src_mem);
+    std::shared_ptr<void> p_diff_dst_mem =
+        std::make_shared<mkldnn::memory>(mkldnn::memory(
+            {data_md, mkldnn_engine}, platform::to_void_cast(diff_dst)));
+    dev_ctx.SetBlob(key_diff_dst_mem, p_diff_dst_mem);
+
+    auto bwd_desc = mkldnn::eltwise_backward::desc(algorithm, data_md, data_md,
+                                                   alpha, beta);
+
+    const std::string key_fwd_pd = key + "eltwise_fwd_pd";
+    auto *p_fwd_pd = static_cast<mkldnn::eltwise_forward::primitive_desc *>(
+        dev_ctx.GetBlob(key_fwd_pd).get());
+
+    auto eltwise_bwd_prim_desc = mkldnn::eltwise_backward::primitive_desc(
+        bwd_desc, mkldnn_engine, *p_fwd_pd);
+
+    p_grad = std::make_shared<mkldnn::eltwise_backward>(
+        eltwise_bwd_prim_desc, *static_cast<mkldnn::memory *>(p_src_mem.get()),
+        *(static_cast<mkldnn::memory *>(p_diff_dst_mem.get())),
+        *(static_cast<mkldnn::memory *>(p_diff_src_mem.get())));
+  } else {
+    // primitives already exist
+    auto p_diff_src_mem = std::static_pointer_cast<mkldnn::memory>(
+        dev_ctx.GetBlob(key_diff_src_mem));
+    auto p_diff_dst_mem = std::static_pointer_cast<mkldnn::memory>(
+        dev_ctx.GetBlob(key_diff_dst_mem));
+
+    p_diff_src_mem->set_data_handle(
+        platform::to_void_reinterpret_cast(diff_src));
+    p_diff_dst_mem->set_data_handle(
+        platform::to_void_reinterpret_cast(diff_dst));
+  }

  // push primitive to stream and wait until it's executed
-  std::vector<mkldnn::primitive> pipeline = {eltwise_bwd};
+  std::vector<mkldnn::primitive> pipeline = {*(p_grad.get())};
  mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
 }
 }  // anonymous namespace

--- a/paddle/fluid/operators/activation_op.cc
+++ b/paddle/fluid/operators/activation_op.cc
@@ -41,7 +41,7 @@ namespace operators {
                                                                             \
   protected:                                                                \
    std::unique_ptr<::paddle::framework::OpDesc> Apply() const override {    \
-      auto *op = new ::paddle::framework::OpDesc();                          \
+      auto* op = new ::paddle::framework::OpDesc();                          \
      op->SetType(#KERNEL_TYPE "_grad");                                     \
      op->SetInput("Out", Output("Out"));                                    \
      op->SetInput(::paddle::framework::GradVarName("Out"),                  \
@@ -54,23 +54,50 @@ namespace operators {
    }                                                                        \
  }

+framework::OpKernelType GetKernelType(const framework::ExecutionContext& ctx,
+                                      const framework::OperatorWithKernel& oper,
+                                      const std::string& name) {
+  framework::LibraryType library{framework::LibraryType::kPlain};
+#ifdef PADDLE_WITH_MKLDNN
+  auto it = oper.Attrs().find("use_mkldnn");
+  if (library == framework::LibraryType::kPlain && it != oper.Attrs().end() &&
+      platform::CanMKLDNNBeUsed(ctx)) {
+    library = framework::LibraryType::kMKLDNN;
+  }
+#endif
+  framework::DataLayout layout = framework::DataLayout::kAnyLayout;
+  return framework::OpKernelType(
+      framework::ToDataType(ctx.Input<framework::Tensor>(name)->type()),
+      ctx.GetPlace(), layout, library);
+}
+
 class ActivationOp : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

-  void InferShape(framework::InferShapeContext *ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
    ctx->SetOutputDim("Out", ctx->GetInputDim("X"));
    ctx->ShareLoD("X", /*->*/ "Out");
  }
+
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return GetKernelType(ctx, *this, "X");
+  }
 };

 class ActivationOpGrad : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

-  void InferShape(framework::InferShapeContext *ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
    ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("Out"));
  }
+
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return GetKernelType(ctx, *this, "Out");
+  }
 };

 __attribute__((unused)) constexpr char SigmoidDoc[] = R"DOC(

--- a/paddle/fluid/operators/beam_search_op.h
+++ b/paddle/fluid/operators/beam_search_op.h
@@ -14,10 +14,6 @@ limitations under the License. */

 #pragma once

-#ifdef PADDLE_WITH_TESTING
-#include "gtest/gtest.h"
-#endif
-
 #include <string>
 #include <vector>
 #include "paddle/fluid/framework/lod_tensor.h"

--- a/paddle/fluid/operators/detail/grpc_server.cc
+++ b/paddle/fluid/operators/detail/grpc_server.cc
@@ -184,7 +184,7 @@ class RequestPrefetch final : public RequestBase {
    framework::Scope* local_scope = &scope_->NewScope();
    auto* var = local_scope->FindVar(var_name);
    InitializeVariable(var, var_desc->GetType());
-    executor_->RunPreparedContext(prefetch_ctx_, scope_, false, false);
+    executor_->RunPreparedContext(prefetch_ctx_, scope_);

    SerializeToByteBuffer(var_name, var, *dev_ctx_, &reply);


--- a/paddle/fluid/operators/detection_map_op.cc
+++ b/paddle/fluid/operators/detection_map_op.cc
@@ -51,7 +51,8 @@ class DetectionMAPOp : public framework::OperatorWithKernel {
    PADDLE_ENFORCE_EQ(label_dims.size(), 2,
                      "The rank of Input(Label) must be 2, "
                      "the shape is [N, 6].");
-    PADDLE_ENFORCE_EQ(label_dims[1], 6, "The shape is of Input(Label) [N, 6].");
+    PADDLE_ENFORCE(label_dims[1] == 6 || label_dims[1] == 5,
+                   "The shape of Input(Label) is [N, 6] or [N, 5].");

    if (ctx->HasInput("PosCount")) {
      PADDLE_ENFORCE(ctx->HasInput("TruePos"),
@@ -88,9 +89,10 @@ class DetectionMAPOpMaker : public framework::OpProtoAndCheckerMaker {
             "offset is N + 1, if LoD[i + 1] - LoD[i] == 0, means there is "
             "no detected data.");
    AddInput("Label",
-             "(LoDTensor) A 2-D LoDTensor with shape[N, 6] represents the"
+             "(LoDTensor) A 2-D LoDTensor represents the"
             "Labeled ground-truth data. Each row has 6 values: "
-             "[label, is_difficult, xmin, ymin, xmax, ymax], N is the total "
+             "[label, xmin, ymin, xmax, ymax, is_difficult] or 5 values: "
+             "[label, xmin, ymin, xmax, ymax], where N is the total "
             "number of ground-truth data in this mini-batch. For each "
             "instance, the offsets in first dimension are called LoD, "
             "the number of offset is N + 1, if LoD[i + 1] - LoD[i] == 0, "

--- a/paddle/fluid/operators/detection_map_op.h
+++ b/paddle/fluid/operators/detection_map_op.h
@@ -72,7 +72,7 @@ class DetectionMAPOpKernel : public framework::OpKernel<T> {
    auto* out_false_pos = ctx.Output<framework::LoDTensor>("AccumFalsePos");

    float overlap_threshold = ctx.Attr<float>("overlap_threshold");
-    float evaluate_difficult = ctx.Attr<bool>("evaluate_difficult");
+    bool evaluate_difficult = ctx.Attr<bool>("evaluate_difficult");
    auto ap_type = GetAPType(ctx.Attr<std::string>("ap_type"));
    int class_num = ctx.Attr<int>("class_num");

@@ -175,14 +175,20 @@ class DetectionMAPOpKernel : public framework::OpKernel<T> {
    for (int n = 0; n < batch_size; ++n) {
      std::map<int, std::vector<Box>> boxes;
      for (size_t i = label_index[n]; i < label_index[n + 1]; ++i) {
-        Box box(labels(i, 2), labels(i, 3), labels(i, 4), labels(i, 5));
        int label = labels(i, 0);
-        auto is_difficult = labels(i, 1);
-        if (std::abs(is_difficult - 0.0) < 1e-6)
-          box.is_difficult = false;
-        else
-          box.is_difficult = true;
-        boxes[label].push_back(box);
+        if (input_label.dims()[1] == 6) {
+          Box box(labels(i, 2), labels(i, 3), labels(i, 4), labels(i, 5));
+          auto is_difficult = labels(i, 1);
+          if (std::abs(is_difficult - 0.0) < 1e-6)
+            box.is_difficult = false;
+          else
+            box.is_difficult = true;
+          boxes[label].push_back(box);
+        } else {
+          PADDLE_ENFORCE_EQ(input_label.dims()[1], 5);
+          Box box(labels(i, 1), labels(i, 2), labels(i, 3), labels(i, 4));
+          boxes[label].push_back(box);
+        }
      }
      gt_boxes->push_back(boxes);
    }

--- a/paddle/fluid/operators/is_empty_op.cc
+++ b/paddle/fluid/operators/is_empty_op.cc
@@ -12,45 +12,41 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */

+#include "paddle/fluid/operators/is_empty_op.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/operator.h"

 namespace paddle {
 namespace operators {

-constexpr char kInput[] = "X";
-constexpr char kOutput[] = "Out";
-
-class IsEmptyOp : public framework::OperatorBase {
+class IsEmptyOp : public framework::OperatorWithKernel {
 public:
-  IsEmptyOp(const std::string &type, const framework::VariableNameMap &inputs,
-            const framework::VariableNameMap &outputs,
-            const framework::AttributeMap &attrs)
-      : OperatorBase(type, inputs, outputs, attrs) {}
+  using framework::OperatorWithKernel::OperatorWithKernel;

- private:
-  void RunImpl(const framework::Scope &scope,
-               const platform::Place &place) const override {
-    // get input
-    auto *var = scope.FindVar(Input(kInput));
-    PADDLE_ENFORCE_NOT_NULL(var);
-    auto &tensor = var->Get<framework::LoDTensor>();
-    // get output
-    auto *out = scope.FindVar(Output(kOutput));
-    PADDLE_ENFORCE_NOT_NULL(out);
-    auto *out_tensor = out->GetMutable<framework::LoDTensor>();
+ protected:
+  void InferShape(framework::InferShapeContext *ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"),
+                   "Input(X) of IsEmptyOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(Out) of IsEmptyOp should not be null.");
+    ctx->SetOutputDim("Out", {1});
+  }

-    out_tensor->Resize({1});
-    out_tensor->mutable_data<bool>(platform::CPUPlace())[0] =
-        framework::product(tensor.dims()) == 0;
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext &ctx) const override {
+    framework::OpKernelType kt = framework::OpKernelType(
+        framework::ToDataType(ctx.Input<framework::LoDTensor>("X")->type()),
+        platform::CPUPlace());
+    return kt;
  }
 };

-class IsEmptyOpProtoMaker : public framework::OpProtoAndCheckerMaker {
+class IsEmptyOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
  void Make() override {
-    AddInput(kInput, "(Tensor) Tensor which is to be checked.");
-    AddOutput(kOutput, "(Tensor) a boolean Tensor that indicate empty or not.");
+    AddInput("X", "(LoDTensor) Tensor which is to be checked.");
+    AddOutput("Out",
+              "(LoDTensor) a boolean Tensor that indicate empty or not.");
    AddComment(R"DOC(
 IsEmpty Operator which checks whether a tensor is empty.

@@ -62,5 +58,12 @@ It will just return product(tensor.ddims()) > 0;
 }  // namespace operators
 }  // namespace paddle

-REGISTER_OP_WITHOUT_GRADIENT(is_empty, paddle::operators::IsEmptyOp,
-                             paddle::operators::IsEmptyOpProtoMaker);
+namespace ops = paddle::operators;
+
+REGISTER_OPERATOR(is_empty, ops::IsEmptyOp, ops::IsEmptyOpMaker,
+                  paddle::framework::EmptyGradOpMaker);
+REGISTER_OP_CPU_KERNEL(
+    is_empty, ops::IsEmptyOpKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::IsEmptyOpKernel<paddle::platform::CPUDeviceContext, double>,
+    ops::IsEmptyOpKernel<paddle::platform::CPUDeviceContext, int>,
+    ops::IsEmptyOpKernel<paddle::platform::CPUDeviceContext, int64_t>);
--- a/paddle/fluid/operators/is_empty_op.h
+++ b/paddle/fluid/operators/is_empty_op.h
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/operator.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename DeviceContext, typename T>
+class IsEmptyOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    // get input
+    auto* input_tensor = context.Input<framework::LoDTensor>("X");
+    // get output
+    auto* output_tensor = context.Output<framework::LoDTensor>("Out");
+
+    output_tensor->mutable_data<bool>(platform::CPUPlace())[0] =
+        framework::product(input_tensor->dims()) == 0;
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
--- a/paddle/fluid/operators/listen_and_serv_op.cc
+++ b/paddle/fluid/operators/listen_and_serv_op.cc
@@ -57,8 +57,7 @@ static void ParallelExecuteBlocks(
        framework::Async([&executor, &prepared, &program, &scope, idx]() {
          int run_block = idx;  // thread local
          try {
-            executor->RunPreparedContext(prepared[run_block].get(), scope,
-                                         false, false);
+            executor->RunPreparedContext(prepared[run_block].get(), scope);
          } catch (std::exception &e) {
            LOG(ERROR) << "run sub program error " << e.what();
          }
@@ -211,8 +210,8 @@ static void AsyncUpdateThread(
    }
    auto fs = framework::Async([var_name, &executor, &v, prepared] {
      try {
-        executor->RunPreparedContext(prepared, v.second->GetMutableLocalScope(),
-                                     false, false);
+        executor->RunPreparedContext(prepared,
+                                     v.second->GetMutableLocalScope());
      } catch (std::exception &e) {
        LOG(ERROR) << "run sub program error " << e.what();
      }

--- a/paddle/fluid/operators/math/math_function.cc
+++ b/paddle/fluid/operators/math/math_function.cc
@@ -38,7 +38,9 @@ template struct SetConstant<platform::CPUDeviceContext, bool>;
  template struct Transpose<platform::CPUDeviceContext, double, RANK>;     \
  template struct Transpose<platform::CPUDeviceContext, int, RANK>;        \
  template struct Transpose<platform::CPUDeviceContext, int64_t, RANK>;    \
-  template struct Transpose<platform::CPUDeviceContext, bool, RANK>;
+  template struct Transpose<platform::CPUDeviceContext, bool, RANK>;       \
+  template struct Transpose<platform::CPUDeviceContext, int16_t, RANK>;    \
+  template struct Transpose<platform::CPUDeviceContext, uint8_t, RANK>;

 DEFINE_CPU_TRANS(1);
 DEFINE_CPU_TRANS(2);

--- a/paddle/fluid/operators/mkldnn_activation_op.h
+++ b/paddle/fluid/operators/mkldnn_activation_op.h
@@ -13,6 +13,8 @@ See the License for the specific language governing permissions and
 limitations under the License. */

 #pragma once
+#include <string>
+
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/detail/safe_ref.h"
@@ -60,52 +62,5 @@ class MKLDNNActivationGradKernel
  }
 };

-namespace {  // NOLINT
-framework::OpKernelType GetKernelType(
-    const framework::ExecutionContext& ctx,
-    const framework::OperatorWithKernel& oper) {
-  framework::LibraryType library{framework::LibraryType::kPlain};
-#ifdef PADDLE_WITH_MKLDNN
-  if (library == framework::LibraryType::kPlain &&
-      platform::CanMKLDNNBeUsed(ctx)) {
-    library = framework::LibraryType::kMKLDNN;
-  }
-#endif
-  framework::DataLayout layout = framework::DataLayout::kAnyLayout;
-  return framework::OpKernelType(
-      framework::ToDataType(ctx.Input<framework::Tensor>("X")->type()),
-      ctx.GetPlace(), layout, library);
-}
-}  // anonymous namespace
-
-class ActivationWithMKLDNNOp : public framework::OperatorWithKernel {
- public:
-  using framework::OperatorWithKernel::OperatorWithKernel;
-
-  void InferShape(framework::InferShapeContext* ctx) const override {
-    ctx->SetOutputDim("Out", ctx->GetInputDim("X"));
-    ctx->ShareLoD("X", /*->*/ "Out");
-  }
-
-  framework::OpKernelType GetExpectedKernelType(
-      const framework::ExecutionContext& ctx) const override {
-    return GetKernelType(ctx, *this);
-  }
-};
-
-class ActivationWithMKLDNNOpGrad : public framework::OperatorWithKernel {
- public:
-  using framework::OperatorWithKernel::OperatorWithKernel;
-
-  void InferShape(framework::InferShapeContext* ctx) const override {
-    ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("Out"));
-  }
-
-  framework::OpKernelType GetExpectedKernelType(
-      const framework::ExecutionContext& ctx) const override {
-    return GetKernelType(ctx, *this);
-  }
-};
-
 }  // namespace operators
 }  // namespace paddle
--- a/paddle/fluid/operators/pool_mkldnn_op.cc
+++ b/paddle/fluid/operators/pool_mkldnn_op.cc
@@ -18,6 +18,26 @@ limitations under the License. */
 namespace paddle {
 namespace operators {

+using mkldnn::memory;  // Note: paddle has also "memory" namespace
+using mkldnn::pooling_forward;
+using mkldnn::pooling_backward;
+
+// Generate keys for storing/retriving primitives for this operator
+// TODO(jczaja): Make hashing function more optimial
+static std::string gethash(memory::dims& input_dims, std::string& pooling_type,
+                           std::vector<int>& ksize, std::vector<int>& strides,
+                           std::vector<int>& paddings, std::string suffix) {
+  auto dims2str = [](memory::dims& operand_dims) {
+    std::string dstr = "";
+    for (size_t i = 0; i < operand_dims.size(); ++i) {
+      dstr += std::to_string(operand_dims[i]) + "-";
+    }
+    return dstr;
+  };
+  return dims2str(input_dims) + dims2str(ksize) + dims2str(strides) +
+         dims2str(paddings) + pooling_type + suffix;
+}
+
 template <typename T>
 class PoolMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
 public:
@@ -34,10 +54,6 @@ class PoolMKLDNNOpKernel : public paddle::framework::OpKernel<T> {

    // Get an unique name from "argument" name of "Out" variable
    // This name will be used as key when saving info into device context
-    const std::string key = ctx.op().Output("Out");
-    const std::string key_pool_pd = key + "@pool_pd";
-    const std::string key_pool_workspace_memory =
-        key + "@pool_workspace_memory";

    std::string pooling_type = ctx.Attr<std::string>("pooling_type");
    std::vector<int> ksize = ctx.Attr<std::vector<int>>("ksize");
@@ -63,37 +79,71 @@ class PoolMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
    std::vector<int> src_tz = paddle::framework::vectorize2int(input->dims());
    std::vector<int> dst_tz = paddle::framework::vectorize2int(output->dims());

-    // TODO(pzelazko-intel): support more formats
-    auto src_md = platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32,
-                                          mkldnn::memory::format::nchw);
-    auto dst_md = platform::MKLDNNMemDesc(dst_tz, mkldnn::memory::f32,
-                                          mkldnn::memory::format::nchw);
-
-    std::shared_ptr<mkldnn::pooling_forward::primitive_desc> pool_pd =
-        CreatePrimitiveDesc(src_md, dst_md, strides, paddings, ksize,
-                            pooling_type, mkldnn_engine);
-
-    // save pool_pd into global device context to be referred in backward path
-    dev_ctx.SetBlob(key_pool_pd, pool_pd);
-
-    std::shared_ptr<mkldnn::memory> workspace_memory =
-        CreateWorkspaceMemory(pool_pd, pooling_type, mkldnn_engine);
-
-    // save pool_workspace_memory to be referred in backward path
-    dev_ctx.SetBlob(key_pool_workspace_memory, workspace_memory);
-
-    auto src_memory =
-        mkldnn::memory({src_md, mkldnn_engine},
-                       static_cast<void*>(const_cast<T*>(input_data)));
-    auto dst_memory =
-        mkldnn::memory({dst_md, mkldnn_engine},
-                       static_cast<void*>(const_cast<T*>(output_data)));
+    const std::string key = gethash(src_tz, pooling_type, ksize, strides,
+                                    paddings, ctx.op().Output("Out"));
+    const std::string key_pool_p = key + "@pool_p";
+    const std::string key_pool_pd = key + "@pool_pd";
+    const std::string key_pool_src_mem_p = key + "@pool_src_mem_p";
+    const std::string key_pool_dst_mem_p = key + "@pool_dst_mem_p";
+    const std::string key_pool_workspace_memory =
+        key + "@pool_workspace_memory";

-    auto pool_prim = mkldnn::pooling_forward(*pool_pd, src_memory, dst_memory,
-                                             *workspace_memory);
+    auto pool_p =
+        std::static_pointer_cast<pooling_forward>(dev_ctx.GetBlob(key_pool_p));
+    if (pool_p == nullptr) {
+      // TODO(pzelazko-intel): support more formats
+
+      auto src_md =
+          platform::MKLDNNMemDesc(src_tz, platform::MKLDNNGetDataType<T>(),
+                                  mkldnn::memory::format::nchw);
+      auto dst_md =
+          platform::MKLDNNMemDesc(dst_tz, platform::MKLDNNGetDataType<T>(),
+                                  mkldnn::memory::format::nchw);
+
+      std::shared_ptr<pooling_forward::primitive_desc> pool_pd =
+          CreatePrimitiveDesc(src_md, dst_md, strides, paddings, ksize,
+                              pooling_type, mkldnn_engine);
+
+      // save pool_pd into global device context to be referred in backward path
+      dev_ctx.SetBlob(key_pool_pd, pool_pd);
+
+      std::shared_ptr<mkldnn::memory> workspace_memory =
+          CreateWorkspaceMemory(pool_pd, pooling_type, mkldnn_engine);
+
+      // save pool_workspace_memory to be referred in backward path
+      dev_ctx.SetBlob(key_pool_workspace_memory, workspace_memory);
+
+      auto pool_src_memory_p = std::make_shared<memory>(
+          memory::primitive_desc{src_md, mkldnn_engine},
+          static_cast<void*>(const_cast<T*>(input_data)));
+      dev_ctx.SetBlob(key_pool_src_mem_p, pool_src_memory_p);
+
+      auto pool_dst_memory_p = std::make_shared<memory>(
+          memory::primitive_desc{dst_md, mkldnn_engine},
+          static_cast<void*>(output_data));
+      dev_ctx.SetBlob(key_pool_dst_mem_p, pool_dst_memory_p);
+
+      pool_p = std::make_shared<pooling_forward>(
+          *pool_pd, *(pool_src_memory_p.get()), *(pool_dst_memory_p.get()),
+          *workspace_memory);
+      dev_ctx.SetBlob(key_pool_p, pool_p);
+    } else {
+      // Primitives already exist
+      auto pool_src_memory_p =
+          std::static_pointer_cast<memory>(dev_ctx.GetBlob(key_pool_src_mem_p));
+      PADDLE_ENFORCE(pool_src_memory_p != nullptr,
+                     "Fail to find pooling src mem_p in device context");
+      auto pool_dst_memory_p =
+          std::static_pointer_cast<memory>(dev_ctx.GetBlob(key_pool_dst_mem_p));
+      PADDLE_ENFORCE(pool_dst_memory_p != nullptr,
+                     "Fail to find pooling dst mem_p in device context");
+      pool_src_memory_p->set_data_handle(
+          reinterpret_cast<void*>(const_cast<T*>(input_data)));
+      pool_dst_memory_p->set_data_handle(output_data);
+    }

    // push primitive to stream and wait until it's executed
-    std::vector<mkldnn::primitive> pipeline{pool_prim};
+    std::vector<mkldnn::primitive> pipeline{*(pool_p.get())};
    mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
  }

@@ -120,9 +170,10 @@ class PoolMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
    mkldnn::memory::primitive_desc workspace_md =
        pooling_type == "max"
            ? pool_pd->workspace_primitive_desc()
-            : mkldnn::memory::primitive_desc(
-                  {{}, mkldnn::memory::f32, mkldnn::memory::format::nchw},
-                  engine);
+            : mkldnn::memory::primitive_desc({{},
+                                              platform::MKLDNNGetDataType<T>(),
+                                              mkldnn::memory::format::nchw},
+                                             engine);

    auto p_workspace_memory = new mkldnn::memory(workspace_md);
    return std::unique_ptr<mkldnn::memory>(p_workspace_memory);
@@ -140,13 +191,6 @@ class PoolMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
    const Tensor* out_grad = ctx.Input<Tensor>(framework::GradVarName("Out"));
    Tensor* in_x_grad = ctx.Output<Tensor>(framework::GradVarName("X"));

-    // Get an unique name from "argument" name of "Out" variable
-    // This name will be used as key when referring info from device context
-    const std::string key = ctx.op().Input("Out");
-    const std::string key_pool_pd = key + "@pool_pd";
-    const std::string key_pool_workspace_memory =
-        key + "@pool_workspace_memory";
-
    std::string pooling_type = ctx.Attr<std::string>("pooling_type");
    std::vector<int> ksize = ctx.Attr<std::vector<int>>("ksize");
    std::vector<int> strides = ctx.Attr<std::vector<int>>("strides");
@@ -171,43 +215,76 @@ class PoolMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
    std::vector<int> diff_dst_tz =
        paddle::framework::vectorize2int(out_grad->dims());

-    auto diff_src_md = platform::MKLDNNMemDesc(diff_src_tz, mkldnn::memory::f32,
-                                               mkldnn::memory::format::nchw);
-    auto diff_dst_md = platform::MKLDNNMemDesc(diff_dst_tz, mkldnn::memory::f32,
-                                               mkldnn::memory::format::nchw);
-
-    // Retrieve pool_pd/pool_workspace_memory from device context
-    auto pool_pd =
-        std::static_pointer_cast<mkldnn::pooling_forward::primitive_desc>(
-            dev_ctx.GetBlob(key_pool_pd));
-    PADDLE_ENFORCE(pool_pd != nullptr,
-                   "Fail to find pool_pd in device context");
-
-    auto workspace_memory = std::static_pointer_cast<mkldnn::memory>(
-        dev_ctx.GetBlob(key_pool_workspace_memory));
-    PADDLE_ENFORCE(workspace_memory != nullptr,
-                   "Fail to find workspace_memory in device context");
-
-    auto pool_bwd_desc = mkldnn::pooling_backward::desc(
-        pooling_type == "max" ? mkldnn::algorithm::pooling_max
-                              : mkldnn::algorithm::pooling_avg,
-        diff_src_md, diff_dst_md, strides, ksize, paddings, paddings,
-        mkldnn::padding_kind::zero);
-    auto pool_bwd_pd = mkldnn::pooling_backward::primitive_desc(
-        pool_bwd_desc, mkldnn_engine, *pool_pd);
-
-    auto diff_src_memory =
-        mkldnn::memory({diff_src_md, mkldnn_engine},
-                       static_cast<void*>(const_cast<T*>(in_x_grad_data)));
-    auto diff_dst_memory =
-        mkldnn::memory({diff_dst_md, mkldnn_engine},
-                       static_cast<void*>(const_cast<T*>(out_grad_data)));
+    // Get an unique name from "argument" name of "Out" variable
+    // This name will be used as key when referring info from device context
+    const std::string key = gethash(diff_src_tz, pooling_type, ksize, strides,
+                                    paddings, ctx.op().Input("Out"));
+    const std::string key_pool_bwd_p = key + "@pool_bwd_p";
+    const std::string key_pool_diff_src_mem_p = key + "@pool_diff_src_mem_p";
+    const std::string key_pool_diff_dst_mem_p = key + "@pool_diff_dst_mem_p";
+    const std::string key_pool_pd = key + "@pool_pd";
+    const std::string key_pool_workspace_memory =
+        key + "@pool_workspace_memory";

-    auto bwd_prim = mkldnn::pooling_backward(
-        pool_bwd_pd, diff_dst_memory, *workspace_memory, diff_src_memory);
+    auto pool_bwd_p = std::static_pointer_cast<pooling_backward>(
+        dev_ctx.GetBlob(key_pool_bwd_p));
+    if (pool_bwd_p == nullptr) {
+      auto diff_src_md =
+          platform::MKLDNNMemDesc(diff_src_tz, platform::MKLDNNGetDataType<T>(),
+                                  mkldnn::memory::format::nchw);
+      auto diff_dst_md =
+          platform::MKLDNNMemDesc(diff_dst_tz, platform::MKLDNNGetDataType<T>(),
+                                  mkldnn::memory::format::nchw);
+      // Retrieve pool_pd/pool_workspace_memory from device context
+      auto pool_pd =
+          std::static_pointer_cast<mkldnn::pooling_forward::primitive_desc>(
+              dev_ctx.GetBlob(key_pool_pd));
+      PADDLE_ENFORCE(pool_pd != nullptr,
+                     "Fail to find pool_pd in device context");
+
+      auto workspace_memory = std::static_pointer_cast<mkldnn::memory>(
+          dev_ctx.GetBlob(key_pool_workspace_memory));
+      PADDLE_ENFORCE(workspace_memory != nullptr,
+                     "Fail to find workspace_memory in device context");
+
+      auto pool_diff_src_memory_p = std::make_shared<memory>(memory(
+          {diff_src_md, mkldnn_engine}, static_cast<void*>(in_x_grad_data)));
+      dev_ctx.SetBlob(key_pool_diff_src_mem_p, pool_diff_src_memory_p);
+
+      auto pool_diff_dst_memory_p = std::make_shared<memory>(
+          memory({diff_dst_md, mkldnn_engine},
+                 static_cast<void*>(const_cast<T*>(out_grad_data))));
+      dev_ctx.SetBlob(key_pool_diff_dst_mem_p, pool_diff_dst_memory_p);
+
+      auto pool_bwd_desc = mkldnn::pooling_backward::desc(
+          pooling_type == "max" ? mkldnn::algorithm::pooling_max
+                                : mkldnn::algorithm::pooling_avg,
+          diff_src_md, diff_dst_md, strides, ksize, paddings, paddings,
+          mkldnn::padding_kind::zero);
+      auto pool_bwd_pd = mkldnn::pooling_backward::primitive_desc(
+          pool_bwd_desc, mkldnn_engine, *pool_pd);
+
+      pool_bwd_p = std::make_shared<pooling_backward>(
+          pool_bwd_pd, *(pool_diff_dst_memory_p.get()), *workspace_memory,
+          *(pool_diff_src_memory_p));
+      dev_ctx.SetBlob(key_pool_bwd_p, pool_bwd_p);
+    } else {
+      // Primitives already exist
+      auto pool_diff_src_memory_p = std::static_pointer_cast<memory>(
+          dev_ctx.GetBlob(key_pool_diff_src_mem_p));
+      PADDLE_ENFORCE(pool_diff_src_memory_p != nullptr,
+                     "Fail to find pooling src mem_p in device context");
+      auto pool_diff_dst_memory_p = std::static_pointer_cast<memory>(
+          dev_ctx.GetBlob(key_pool_diff_dst_mem_p));
+      PADDLE_ENFORCE(pool_diff_dst_memory_p != nullptr,
+                     "Fail to find pooling dst mem_p in device context");
+      pool_diff_src_memory_p->set_data_handle(
+          reinterpret_cast<void*>(in_x_grad_data));
+      pool_diff_dst_memory_p->set_data_handle(const_cast<T*>(out_grad_data));
+    }

    // push primitive to stream and wait until it's executed
-    std::vector<mkldnn::primitive> pipeline{bwd_prim};
+    std::vector<mkldnn::primitive> pipeline{*(pool_bwd_p.get())};
    mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
  }  // Compute()
 };

--- a/paddle/fluid/operators/roi_pool_op.cu
+++ b/paddle/fluid/operators/roi_pool_op.cu
@@ -38,10 +38,10 @@ __global__ void GPUROIPoolForward(
  int index = blockIdx.x * blockDim.x + threadIdx.x;
  int offset = blockDim.x * gridDim.x;
  for (size_t i = index; i < nthreads; i += offset) {
-    int pw = index % pooled_width;
-    int ph = (index / pooled_width) % pooled_height;
-    int c = (index / pooled_width / pooled_height) % channels;
-    int n = index / pooled_width / pooled_height / channels;
+    int pw = i % pooled_width;
+    int ph = (i / pooled_width) % pooled_height;
+    int c = (i / pooled_width / pooled_height) % channels;
+    int n = i / pooled_width / pooled_height / channels;

    const int64_t* offset_input_rois = input_rois + n * kROISize;
    int roi_batch_ind = roi_batch_id_data[n];
@@ -52,14 +52,19 @@ __global__ void GPUROIPoolForward(

    int roi_width = max(roi_end_w - roi_start_w + 1, 1);
    int roi_height = max(roi_end_h - roi_start_h + 1, 1);
-    T bin_size_h = static_cast<T>(roi_height) / static_cast<T>(pooled_height);
-    T bin_size_w = static_cast<T>(roi_width) / static_cast<T>(pooled_width);
-
-    int hstart = static_cast<int>(floor(static_cast<T>(ph) * bin_size_h));
-    int wstart = static_cast<int>(floor(static_cast<T>(pw) * bin_size_w));
-    int hend = static_cast<int>(ceil(static_cast<T>(ph + 1) * bin_size_h));
-    int wend = static_cast<int>(ceil(static_cast<T>(pw + 1) * bin_size_w));

+    int hstart = static_cast<int>(floor(static_cast<double>(ph) *
+                                        static_cast<double>(roi_height) /
+                                        static_cast<double>(pooled_height)));
+    int wstart = static_cast<int>(floor(static_cast<double>(pw) *
+                                        static_cast<double>(roi_width) /
+                                        static_cast<double>(pooled_width)));
+    int hend = static_cast<int>(ceil(static_cast<double>(ph + 1) *
+                                     static_cast<double>(roi_height) /
+                                     static_cast<double>(pooled_height)));
+    int wend = static_cast<int>(ceil(static_cast<double>(pw + 1) *
+                                     static_cast<double>(roi_width) /
+                                     static_cast<double>(pooled_width)));
    hstart = min(max(hstart + roi_start_h, 0), height);
    hend = min(max(hend + roi_start_h, 0), height);
    wstart = min(max(wstart + roi_start_w, 0), width);
@@ -79,9 +84,9 @@ __global__ void GPUROIPoolForward(
        }
      }
    }
-    output_data[index] = maxval;
+    output_data[i] = maxval;
    if (argmax_data) {
-      argmax_data[index] = maxidx;
+      argmax_data[i] = maxidx;
    }
  }
 }
@@ -96,10 +101,10 @@ __global__ void GPUROIPoolBackward(
  int index = blockIdx.x * blockDim.x + threadIdx.x;
  int offset = blockDim.x * gridDim.x;
  for (int i = index; i < nthreads; i += offset) {
-    int pw = index % pooled_width;
-    int ph = (index / pooled_width) % pooled_height;
-    int c = (index / pooled_width / pooled_height) % channels;
-    int n = index / pooled_width / pooled_height / channels;
+    int pw = i % pooled_width;
+    int ph = (i / pooled_width) % pooled_height;
+    int c = (i / pooled_width / pooled_height) % channels;
+    int n = i / pooled_width / pooled_height / channels;

    int roi_batch_ind = roi_batch_id_data[n];
    int input_offset = (roi_batch_ind * channels + c) * height * width;
@@ -138,6 +143,7 @@ class GPUROIPoolOpKernel : public framework::OpKernel<T> {
    int width = in_dims[3];

    int rois_num = rois->dims()[0];
+
    if (rois_num == 0) return;

    int output_size = out->numel();

--- a/paddle/fluid/operators/send_recv_op_test.cc
+++ b/paddle/fluid/operators/send_recv_op_test.cc
@@ -92,12 +92,16 @@ void InitSelectedRowsInScope(const p::CPUPlace &place, f::Scope *scope) {

 void AddOp(const std::string &type, const f::VariableNameMap &inputs,
           const f::VariableNameMap &outputs, f::AttributeMap attrs,
-           f::BlockDesc *block) {
+           f::BlockDesc *block, bool is_sparse) {
  // insert output
  for (auto kv : outputs) {
    for (auto v : kv.second) {
      auto var = block->Var(v);
      var->SetDataType(f::proto::VarType::FP32);
+      var->SetPersistable(true);
+      if (is_sparse) {
+        var->SetType(f::proto::VarType::SELECTED_ROWS);
+      }
    }
  }

@@ -128,7 +132,8 @@ void StartServerNet(bool is_sparse, std::atomic<bool> *initialized) {
  auto *optimize_block = program.AppendBlock(root_block);
  auto *prefetch_block = program.AppendBlock(root_block);
  // X for server side tensors, RX for received tensors, must be of same shape.
-  AddOp("sum", {{"X", {"x0", "x1"}}}, {{"Out", {"Out"}}}, {}, optimize_block);
+  AddOp("sum", {{"X", {"x0", "x1"}}}, {{"Out", {"Out"}}}, {}, optimize_block,
+        is_sparse);
  f::AttributeMap attrs;
  attrs.insert({"endpoint", std::string("127.0.0.1:0")});
  attrs.insert({"Fanin", 1});

--- a/paddle/fluid/operators/smooth_l1_loss_op.cc
+++ b/paddle/fluid/operators/smooth_l1_loss_op.cc
@@ -105,7 +105,7 @@ class SmoothL1LossGradOp : public framework::OperatorWithKernel {
  using framework::OperatorWithKernel::OperatorWithKernel;

  void InferShape(framework::InferShapeContext* ctx) const override {
-    auto in_dims = ctx->GetInputDim("X");
+    auto in_dims = ctx->GetInputDim("Diff");
    auto out_dims = ctx->GetInputDim(framework::GradVarName("Out"));

    PADDLE_ENFORCE_GE(out_dims.size(), 2,
@@ -127,12 +127,33 @@ class SmoothL1LossGradOp : public framework::OperatorWithKernel {
  }
 };

+class SmoothL1LossGradMaker : public framework::SingleGradOpDescMaker {
+ public:
+  using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
+
+ protected:
+  std::unique_ptr<framework::OpDesc> Apply() const override {
+    auto* op = new framework::OpDesc();
+    op->SetType("smooth_l1_loss_grad");
+    op->SetInput("InsideWeight", Input("InsideWeight"));
+    op->SetInput("OutsideWeight", Input("OutsideWeight"));
+    op->SetInput("Diff", Output("Diff"));
+    op->SetInput(framework::GradVarName("Out"), OutputGrad("Out"));
+
+    op->SetAttrMap(Attrs());
+
+    op->SetOutput(framework::GradVarName("X"), InputGrad("X"));
+    op->SetOutput(framework::GradVarName("Y"), InputGrad("Y"));
+    return std::unique_ptr<framework::OpDesc>(op);
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle

 namespace ops = paddle::operators;
 REGISTER_OPERATOR(smooth_l1_loss, ops::SmoothL1LossOp, ops::SmoothL1LossOpMaker,
-                  paddle::framework::DefaultGradOpDescMaker<true>);
+                  ops::SmoothL1LossGradMaker);
 REGISTER_OPERATOR(smooth_l1_loss_grad, ops::SmoothL1LossGradOp);
 REGISTER_OP_CPU_KERNEL(
    smooth_l1_loss,

--- a/paddle/fluid/platform/CMakeLists.txt
+++ b/paddle/fluid/platform/CMakeLists.txt
-proto_library(profiler_proto SRCS profiler.proto)
+proto_library(profiler_proto SRCS profiler.proto DEPS framework_proto)
 py_proto_compile(profiler_py_proto SRCS profiler.proto)

 add_custom_target(profiler_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E touch __init__.py)
@@ -49,7 +49,7 @@ nv_test(device_context_test SRCS device_context_test.cu DEPS device_context gpu_
 nv_test(cudnn_helper_test SRCS cudnn_helper_test.cc DEPS dynload_cuda)
 nv_test(transform_test SRCS transform_test.cu DEPS memory place device_context)

-cc_library(device_tracer SRCS device_tracer.cc DEPS boost profiler_proto ${GPU_CTX_DEPS})
+cc_library(device_tracer SRCS device_tracer.cc DEPS boost profiler_proto framework_proto ${GPU_CTX_DEPS})
 cc_library(profiler SRCS profiler.cc DEPS device_context device_tracer)
 cc_test(profiler_test SRCS profiler_test.cc DEPS profiler)


--- a/paddle/fluid/platform/mkldnn_helper.h
+++ b/paddle/fluid/platform/mkldnn_helper.h
@@ -38,6 +38,11 @@ void* to_void_cast(const Type* t) {
  return static_cast<void*>(const_cast<Type*>(t));
 }

+template <typename Type>
+void* to_void_reinterpret_cast(const Type* t) {
+  return reinterpret_cast<void*>(const_cast<Type*>(t));
+}
+
 template <class Type>
 using tf_desc = typename Type::desc;

@@ -71,5 +76,15 @@ inline bool CanMKLDNNBeUsed(const framework::ExecutionContext& ctx) {
  return use_mkldnn && platform::is_cpu_place(ctx.GetPlace());
 }

+template <typename Type>
+mkldnn::memory::data_type MKLDNNGetDataType() {
+  return mkldnn::memory::data_undef;
+}
+
+template <>
+inline mkldnn::memory::data_type MKLDNNGetDataType<float>() {
+  return mkldnn::memory::f32;
+}
+
 }  // namespace platform
 }  // namespace paddle
--- a/paddle/fluid/platform/profiler.cc
+++ b/paddle/fluid/platform/profiler.cc
@@ -173,8 +173,9 @@ void PopEvent(const std::string& name, const DeviceContext* dev_ctx) {
 }

 RecordEvent::RecordEvent(const std::string& name, const DeviceContext* dev_ctx)
-    : start_ns_(PosixInNsec()) {
+    : is_enabled_(false), start_ns_(PosixInNsec()) {
  if (g_state == ProfilerState::kDisabled) return;
+  is_enabled_ = true;
  dev_ctx_ = dev_ctx;
  name_ = name;
  PushEvent(name_, dev_ctx_);
@@ -183,7 +184,7 @@ RecordEvent::RecordEvent(const std::string& name, const DeviceContext* dev_ctx)
 }

 RecordEvent::~RecordEvent() {
-  if (g_state == ProfilerState::kDisabled) return;
+  if (g_state == ProfilerState::kDisabled || !is_enabled_) return;
  DeviceTracer* tracer = GetDeviceTracer();
  if (tracer) {
    tracer->AddCPURecords(CurAnnotation(), start_ns_, PosixInNsec(),
@@ -193,14 +194,16 @@ RecordEvent::~RecordEvent() {
  PopEvent(name_, dev_ctx_);
 }

-RecordBlock::RecordBlock(int block_id) : start_ns_(PosixInNsec()) {
+RecordBlock::RecordBlock(int block_id)
+    : is_enabled_(false), start_ns_(PosixInNsec()) {
  if (g_state == ProfilerState::kDisabled) return;
+  is_enabled_ = true;
  SetCurBlock(block_id);
  name_ = string::Sprintf("block_%d", block_id);
 }

 RecordBlock::~RecordBlock() {
-  if (g_state == ProfilerState::kDisabled) return;
+  if (g_state == ProfilerState::kDisabled || !is_enabled_) return;
  DeviceTracer* tracer = GetDeviceTracer();
  if (tracer) {
    // We try to put all blocks at the same nested depth in the

--- a/paddle/fluid/platform/profiler.h
+++ b/paddle/fluid/platform/profiler.h
@@ -74,6 +74,7 @@ struct RecordEvent {

  ~RecordEvent();

+  bool is_enabled_;
  uint64_t start_ns_;
  // The device context is used by Event to get the current cuda stream.
  const DeviceContext* dev_ctx_;
@@ -89,6 +90,7 @@ struct RecordBlock {
  ~RecordBlock();

 private:
+  bool is_enabled_;
  std::string name_;
  uint64_t start_ns_;
 };

--- a/paddle/fluid/pybind/protobuf.cc
+++ b/paddle/fluid/pybind/protobuf.cc
@@ -238,6 +238,7 @@ void BindVarDsec(pybind11::module *m) {

  pybind11::enum_<pd::proto::VarType::Type>(var_desc, "VarType", "")
      .value("BOOL", pd::proto::VarType::BOOL)
+      .value("UINT8", pd::proto::VarType::UINT8)
      .value("INT16", pd::proto::VarType::INT16)
      .value("INT32", pd::proto::VarType::INT32)
      .value("INT64", pd::proto::VarType::INT64)

--- a/paddle/fluid/train/demo/CMakeLists.txt
+++ b/paddle/fluid/train/demo/CMakeLists.txt
+cmake_minimum_required(VERSION 3.0)
+
+project(cpp_train_demo CXX C)
+
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11")
+
+if(NOT DEFINED PADDLE_LIB)
+  message(FATAL_ERROR "please set PADDLE_LIB with -DPADDLE_LIB=/paddle/lib/dir")
+endif()
+
+option(WITH_MKLDNN     "Compile PaddlePaddle with MKLDNN"                                   OFF)
+option(WITH_MKL        "Compile PaddlePaddle with MKL support, default use openblas."       OFF)
+
+include_directories("${PADDLE_LIB}")
+include_directories("${PADDLE_LIB}/third_party/install/protobuf/include")
+include_directories("${PADDLE_LIB}/third_party/install/glog/include")
+include_directories("${PADDLE_LIB}/third_party/install/gflags/include")
+include_directories("${PADDLE_LIB}/third_party/install/snappy/include")
+include_directories("${PADDLE_LIB}/third_party/install/snappystream/include")
+include_directories("${PADDLE_LIB}/third_party/install/zlib/include")
+
+include_directories("${PADDLE_LIB}/third_party/boost")
+include_directories("${PADDLE_LIB}/third_party/eigen3")
+
+link_directories("${PADDLE_LIB}/third_party/install/snappy/lib")
+link_directories("${PADDLE_LIB}/third_party/install/snappystream/lib")
+link_directories("${PADDLE_LIB}/third_party/install/protobuf/lib")
+link_directories("${PADDLE_LIB}/third_party/install/glog/lib")
+link_directories("${PADDLE_LIB}/third_party/install/gflags/lib")
+link_directories("${PADDLE_LIB}/third_party/install/zlib/lib")
+
+add_executable(demo_trainer demo_trainer.cc)
+
+if(WITH_MKLDNN)
+  include_directories("${PADDLE_LIB}/third_party/install/mkldnn/include")
+  set(MKLDNN_LIB ${PADDLE_LIB}/third_party/install/mkldnn/lib/libmkldnn.so.0)
+endif()
+
+if(WITH_MKL)
+  include_directories("${PADDLE_LIB}/third_party/install/mklml/include")
+  set(MATH_LIB ${PADDLE_LIB}/third_party/install/mklml/lib/libmklml_intel.so)
+else()
+  if(APPLE)
+    set(MATH_LIB cblas)
+  else(APPLE)
+    set(MATH_LIB ${PADDLE_LIB}/third_party/install/openblas/lib/libopenblas.a)
+  endif(APPLE)
+endif()
+
+if(APPLE)
+  set(MACOS_LD_FLAGS "-undefined dynamic_lookup -Wl,-all_load -framework CoreFoundation -framework Security")
+else(APPLE)
+  set(ARCHIVE_START "-Wl,--whole-archive")
+  set(ARCHIVE_END "-Wl,--no-whole-archive")
+  set(EXTERNAL_LIB "-lrt -ldl -lpthread")
+endif(APPLE)
+
+target_link_libraries(demo_trainer
+        ${MACOS_LD_FLAGS}
+        ${ARCHIVE_START}
+        ${PADDLE_LIB}/paddle/fluid/inference/libpaddle_fluid.a
+        ${ARCHIVE_END}
+        ${MATH_LIB}
+        ${MKLDNN_LIB}
+        glog gflags protobuf snappystream snappy z
+        ${EXTERNAL_LIB})
--- a/paddle/fluid/train/demo/README.md
+++ b/paddle/fluid/train/demo/README.md
+
+### step 1. build paddle lib
+
+```
+
+# WITH_MKL=ON|OFF
+# WITH_MKLDNN=ON|OFF
+
+PADDLE_LIB=/paddle/lib/dir
+cmake .. -DCMAKE_INSTALL_PREFIX=$PADDLE_LIB \
+         -DCMAKE_BUILD_TYPE=Release \
+         -DWITH_FLUID_ONLY=ON \
+         -DWITH_GPU=OFF \
+         -DWITH_STYLE_CHECK=OFF \
+         -DWITH_MKL=OFF \
+         -DWITH_MKLDNN=OFF
+make -j8
+make -j8 inference_lib_dist
+```
+
+### step 2. generate program desc
+```
+# please install paddle before run this scripe
+pip install --upgrade paddlepaddle-*.whl
+python demo_network.py
+```
+
+This will generate two program desc files:
+  - startup_program: used to init all parameters
+  - main_program: main logic of the network
+
+### step 3. build demo_trainer and run it.
+
+
+```
+# Make a build dir at the same dir of this README.md document.
+# The demo dir can be put anywhere.
+mkdir build
+cd build
+
+# WITH_MKL=ON|OFF
+# WITH_MKLDNN=ON|OFF
+PADDLE_LIB=/paddle/lib/dir
+
+# PADDLE_LIB is the same with CMAKE_INSTALL_PREFIX when building the lib
+cmake .. -DPADDLE_LIB=$PADDLE_LIB \
+         -DWITH_MKLDNN=OFF \
+         -DWITH_MKL=OFF
+make
+
+# copy startup_program and main_program to this dir
+cp ../startup_program .
+cp ../main_program .
+
+# run demo cpp trainer
+./demo_trainer
+
+```
+
+The output will be:
+```
+step: 0 loss: 1069.02
+step: 1 loss: 1069.02
+step: 2 loss: 1069.02
+....
+```
--- a/paddle/fluid/train/demo/demo_network.py
+++ b/paddle/fluid/train/demo/demo_network.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 paddle.fluid as fluid
+import paddle.fluid.framework as framework
+
+
+def train_network(with_optimize):
+    x = fluid.layers.data(name='x', shape=[13], dtype='float32')
+    y_predict = fluid.layers.fc(input=x, size=1, act=None)
+
+    y = fluid.layers.data(name='y', shape=[1], dtype='float32')
+    cost = fluid.layers.square_error_cost(input=y_predict, label=y)
+    avg_cost = fluid.layers.mean(cost)
+
+    if with_optimize:
+        sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.00001)
+        sgd_optimizer.minimize(avg_cost)
+    else:
+        fluid.backward.append_backward(avg_cost)
+
+
+def save_program_desc(network_func):
+    startup_program = framework.Program()
+    train_program = framework.Program()
+
+    with framework.program_guard(train_program, startup_program):
+        network_func(with_optimize=False)
+
+    with open("startup_program", "w") as f:
+        f.write(startup_program.desc.serialize_to_string())
+    with open("main_program", "w") as f:
+        f.write(train_program.desc.serialize_to_string())
+
+
+save_program_desc(train_network)
--- a/paddle/fluid/train/demo/demo_trainer.cc
+++ b/paddle/fluid/train/demo/demo_trainer.cc
+//   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.
+
+#include <fstream>
+
+#include "paddle/fluid/framework/executor.h"
+#include "paddle/fluid/framework/init.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/program_desc.h"
+#include "paddle/fluid/framework/tensor_util.h"
+#include "paddle/fluid/platform/device_context.h"
+#include "paddle/fluid/platform/place.h"
+
+namespace paddle {
+namespace train {
+
+void ReadBinaryFile(const std::string& filename, std::string* contents) {
+  std::ifstream fin(filename, std::ios::in | std::ios::binary);
+  PADDLE_ENFORCE(static_cast<bool>(fin), "Cannot open file %s", filename);
+  fin.seekg(0, std::ios::end);
+  contents->clear();
+  contents->resize(fin.tellg());
+  fin.seekg(0, std::ios::beg);
+  fin.read(&(contents->at(0)), contents->size());
+  fin.close();
+}
+
+std::unique_ptr<paddle::framework::ProgramDesc> Load(
+    paddle::framework::Executor* executor, const std::string& model_filename) {
+  VLOG(3) << "loading model from " << model_filename;
+  std::string program_desc_str;
+  ReadBinaryFile(model_filename, &program_desc_str);
+
+  std::unique_ptr<paddle::framework::ProgramDesc> main_program(
+      new paddle::framework::ProgramDesc(program_desc_str));
+  return main_program;
+}
+
+}  // namespace train
+}  // namespace paddle
+
+int main() {
+  paddle::framework::InitDevices(false);
+
+  const auto cpu_place = paddle::platform::CPUPlace();
+
+  paddle::framework::Executor executor(cpu_place);
+  paddle::framework::Scope scope;
+  auto startup_program = paddle::train::Load(&executor, "startup_program");
+  auto train_program = paddle::train::Load(&executor, "main_program");
+
+  std::string loss_name = "";
+  for (auto op_desc : train_program->Block(0).AllOps()) {
+    if (op_desc->Type() == "mean") {
+      loss_name = op_desc->Output("Out")[0];
+      break;
+    }
+  }
+
+  PADDLE_ENFORCE_NE(loss_name, "", "loss not found");
+
+  // init all parameters
+  executor.Run(*startup_program.get(), &scope, 0);
+
+  // prepare data
+  auto x_var = scope.Var("x");
+  auto x_tensor = x_var->GetMutable<paddle::framework::LoDTensor>();
+  x_tensor->Resize({2, 13});
+
+  auto x_data = x_tensor->mutable_data<float>(cpu_place);
+  for (int i = 0; i < 2 * 13; ++i) {
+    x_data[i] = static_cast<float>(i);
+  }
+
+  auto y_var = scope.Var("y");
+  auto y_tensor = y_var->GetMutable<paddle::framework::LoDTensor>();
+  y_tensor->Resize({2, 1});
+  auto y_data = y_tensor->mutable_data<float>(cpu_place);
+  for (int i = 0; i < 2 * 1; ++i) {
+    y_data[i] = static_cast<float>(i);
+  }
+
+  auto loss_var = scope.Var(loss_name);
+
+  for (int i = 0; i < 10; ++i) {
+    executor.Run(*train_program.get(), &scope, 0, false, true);
+    std::cout << "step: " << i << " loss: "
+              << loss_var->Get<paddle::framework::LoDTensor>().data<float>()[0]
+              << std::endl;
+  }
+  return 0;
+}
--- a/paddle/scripts/docker/build.sh
+++ b/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==9.0.3 && \
+        apt-get install -y wget python-pip dmidecode python-tk && easy_install -U pip && \
        pip install /*.whl; apt-get install -f -y && \
        apt-get clean -y && \
        rm -f /*.whl && \

--- a/paddle/scripts/paddle_build.sh
+++ b/paddle/scripts/paddle_build.sh
@@ -20,19 +20,15 @@
 #=================================================

 function print_usage() {
-    RED='\033[0;31m'
-    BLUE='\033[0;34m'
-    BOLD='\033[1m'
-    NONE='\033[0m'
-    
    echo -e "\n${RED}Usage${NONE}:
-    ${BOLD}$0${NONE} [OPTION]"
+    ${BOLD}${SCRIPT_NAME}${NONE} [OPTION]"
    
    echo -e "\n${RED}Options${NONE}:
    ${BLUE}build${NONE}: run build for x86 platform
    ${BLUE}build_android${NONE}: run build for android platform
    ${BLUE}build_ios${NONE}: run build for ios platform
    ${BLUE}test${NONE}: run all unit tests
+    ${BLUE}single_test${NONE}: run a single unit test
    ${BLUE}bind_test${NONE}: parallel tests bind to different GPU
    ${BLUE}doc${NONE}: generate paddle documents
    ${BLUE}html${NONE}: convert C++ source code into HTML
@@ -45,7 +41,15 @@ function print_usage() {
 }

 function init() {
+    RED='\033[0;31m'
+    BLUE='\033[0;34m'
+    BOLD='\033[1m'
+    NONE='\033[0m'
+
    PADDLE_ROOT="$( cd "$( dirname "${BASH_SOURCE[0]}")/../../" && pwd )"
+    if [ -z "${SCRIPT_NAME}" ]; then
+        SCRIPT_NAME=$0
+    fi
 }

 function cmake_gen() {
@@ -91,7 +95,6 @@ function cmake_gen() {
        -DWITH_AVX=${WITH_AVX:-OFF}
        -DWITH_GOLANG=${WITH_GOLANG:-OFF}
        -DCUDA_ARCH_NAME=${CUDA_ARCH_NAME:-All}
-        -DWITH_SWIG_PY=ON
        -DWITH_C_API=${WITH_C_API:-OFF}
        -DWITH_PYTHON=${WITH_PYTHON:-ON}
        -DWITH_SWIG_PY=${WITH_SWIG_PY:-ON}
@@ -309,6 +312,25 @@ EOF
    fi
 }

+function single_test() {
+    TEST_NAME=$1
+    if [ -z "${TEST_NAME}" ]; then
+        echo -e "${RED}Usage:${NONE}"
+        echo -e "${BOLD}${SCRIPT_NAME}${NONE} ${BLUE}single_test${NONE} [test_name]"
+        exit 1
+    fi
+    mkdir -p ${PADDLE_ROOT}/build
+    cd ${PADDLE_ROOT}/build
+    if [ ${WITH_TESTING:-ON} == "ON" ] ; then
+    cat <<EOF
+    ========================================
+    Running ${TEST_NAME} ...
+    ========================================
+EOF
+        ctest --output-on-failure -R ${TEST_NAME}
+    fi
+}
+
 function bind_test() {
    # the number of process to run tests
    NUM_PROC=6
@@ -383,17 +405,19 @@ EOF

 function gen_dockerfile() {
    # Set BASE_IMAGE according to env variables
+    CUDA_MAJOR="$(echo $CUDA_VERSION | cut -d '.' -f 1).$(echo $CUDA_VERSION | cut -d '.' -f 2)"
+    CUDNN_MAJOR=$(echo $CUDNN_VERSION | cut -d '.' -f 1)
    if [[ ${WITH_GPU} == "ON" ]]; then
-    BASE_IMAGE="nvidia/cuda:8.0-cudnn5-runtime-ubuntu16.04"
+        BASE_IMAGE="nvidia/cuda:${CUDA_MAJOR}-cudnn${CUDNN_MAJOR}-runtime-ubuntu16.04"
    else
-    BASE_IMAGE="ubuntu:16.04"
+        BASE_IMAGE="ubuntu:16.04"
    fi

    DOCKERFILE_GPU_ENV=""
    DOCKERFILE_CUDNN_DSO=""
    if [[ ${WITH_GPU:-OFF} == 'ON' ]]; then
        DOCKERFILE_GPU_ENV="ENV LD_LIBRARY_PATH /usr/lib/x86_64-linux-gnu:\${LD_LIBRARY_PATH}"
-        DOCKERFILE_CUDNN_DSO="RUN ln -s /usr/lib/x86_64-linux-gnu/libcudnn.so.5 /usr/lib/x86_64-linux-gnu/libcudnn.so"
+        DOCKERFILE_CUDNN_DSO="RUN ln -s /usr/lib/x86_64-linux-gnu/libcudnn.so.${CUDNN_MAJOR} /usr/lib/x86_64-linux-gnu/libcudnn.so"
    fi

    cat <<EOF
@@ -427,7 +451,7 @@ 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==9.0.3 && \
+        apt-get install -y wget python-pip dmidecode python-tk && easy_install -U pip && \
        pip install /*.whl; apt-get install -f -y && \
        apt-get clean -y && \
        rm -f /*.whl && \
@@ -468,7 +492,7 @@ function gen_fluid_inference_lib() {
    Deploying fluid inference library ...
    ========================================
 EOF
-        make inference_lib_dist
+        make -j `nproc` inference_lib_dist
    fi
 }

@@ -480,6 +504,7 @@ function main() {
      build)
        cmake_gen ${PYTHON_ABI:-""}
        build
+        gen_dockerfile
        ;;
      build_android)
        build_android
@@ -490,6 +515,9 @@ function main() {
      test)
        run_test
        ;;
+      single_test)
+        single_test $2
+        ;;
      bind_test)
        bind_test
        ;;
@@ -504,6 +532,7 @@ function main() {
        ;;
      capi)
        cmake_gen ${PYTHON_ABI:-""}
+        build
        gen_capi_package
        ;;
      fluid_inference_lib)

--- a/paddle/scripts/paddle_docker_build.sh
+++ b/paddle/scripts/paddle_docker_build.sh
@@ -14,20 +14,9 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

-function container_running() {
-    name=$1
-    docker ps -a --format "{{.Names}}" | grep "${name}" > /dev/null
-    return $?
-}
-
 function start_build_docker() {
    docker pull $IMG

-    if container_running "${CONTAINER_ID}"; then
-        docker stop "${CONTAINER_ID}" 1>/dev/null
-        docker rm -f "${CONTAINER_ID}" 1>/dev/null
-    fi
-
    apt_mirror='s#http://archive.ubuntu.com/ubuntu#mirror://mirrors.ubuntu.com/mirrors.txt#g'
    DOCKER_ENV=$(cat <<EOL
        -e FLAGS_fraction_of_gpu_memory_to_use=0.15 \
@@ -61,8 +50,8 @@ EOL
    fi
    set -ex
    ${DOCKER_CMD} run -it \
-        --name $CONTAINER_ID \
        ${DOCKER_ENV} \
+        -e SCRIPT_NAME=$0 \
        -v $PADDLE_ROOT:/paddle \
        -v ${HOME}/.ccache:/root/.ccache \
        -w /paddle \
@@ -74,10 +63,8 @@ EOL
 function main() {
    DOCKER_REPO="paddlepaddle/paddle"
    VERSION="latest-dev"
-    CONTAINER_ID="${USER}_paddle_dev"
    PADDLE_ROOT="$( cd "$( dirname "${BASH_SOURCE[0]}")/../../" && pwd )"
    if [ "$1" == "build_android" ]; then
-        CONTAINER_ID="${USER}_paddle_dev_android"
        VERSION="latest-dev-android"
    fi
    IMG=${DOCKER_REPO}:${VERSION}

--- a/patches/mkldnn.hpp
+++ b/patches/mkldnn.hpp
--- a/python/paddle/fluid/data_feeder.py
+++ b/python/paddle/fluid/data_feeder.py
@@ -54,9 +54,9 @@ class DataToLoDTensorConverter(object):
            self.data.append(data)
        else:
            cur_lod_len = len(data)
-            lod[-1].append(lod[-1][-1] + cur_lod_len)
+            lod[0].append(lod[0][-1] + cur_lod_len)
            for each_data in data:
-                self._feed_impl_(each_data, lod[:-1], lod_level - 1)
+                self._feed_impl_(each_data, lod[1:], lod_level - 1)

    def done(self):
        arr = numpy.array(self.data, dtype=self.dtype).reshape(self.shape)

--- a/python/paddle/fluid/evaluator.py
+++ b/python/paddle/fluid/evaluator.py
@@ -273,10 +273,11 @@ class DetectionMAP(Evaluator):
            [M, 6]. The layout is [label, confidence, xmin, ymin, xmax, ymax].
        gt_label (Variable): The ground truth label index, which is a LoDTensor
            with shape [N, 1].
-        gt_difficult (Variable): Whether this ground truth is a difficult
-            bounding box (bbox), which is a LoDTensor [N, 1].
        gt_box (Variable): The ground truth bounding box (bbox), which is a
            LoDTensor with shape [N, 6]. The layout is [xmin, ymin, xmax, ymax].
+        gt_difficult (Variable|None): Whether this ground truth is a difficult
+            bounding bbox, which can be a LoDTensor [N, 1] or not set. If None,
+            it means all the ground truth labels are not difficult bbox.
        class_num (int): The class number.
        background_label (int): The index of background label, the background
            label will be ignored. If set to -1, then all categories will be
@@ -284,7 +285,8 @@ class DetectionMAP(Evaluator):
        overlap_threshold (float): The threshold for deciding true/false
            positive, 0.5 by defalut.
        evaluate_difficult (bool): Whether to consider difficult ground truth
-            for evaluation, True by defalut.
+            for evaluation, True by defalut. This argument does not work when
+            gt_difficult is None.
        ap_version (string): The average precision calculation ways, it must be
            'integral' or '11point'. Please check
            https://sanchom.wordpress.com/tag/average-precision/ for details.
@@ -295,7 +297,7 @@ class DetectionMAP(Evaluator):

        exe = fluid.executor(place)
        map_evaluator = fluid.Evaluator.DetectionMAP(input,
-            gt_label, gt_difficult, gt_box)
+            gt_label, gt_box, gt_difficult)
        cur_map, accum_map = map_evaluator.get_map_var()
        fetch = [cost, cur_map, accum_map]
        for epoch in PASS_NUM:
@@ -313,8 +315,8 @@ class DetectionMAP(Evaluator):
                 input,
                 gt_label,
                 gt_box,
-                 gt_difficult,
-                 class_num,
+                 gt_difficult=None,
+                 class_num=None,
                 background_label=0,
                 overlap_threshold=0.5,
                 evaluate_difficult=True,
@@ -322,8 +324,11 @@ class DetectionMAP(Evaluator):
        super(DetectionMAP, self).__init__("map_eval")

        gt_label = layers.cast(x=gt_label, dtype=gt_box.dtype)
-        gt_difficult = layers.cast(x=gt_difficult, dtype=gt_box.dtype)
-        label = layers.concat([gt_label, gt_difficult, gt_box], axis=1)
+        if gt_difficult:
+            gt_difficult = layers.cast(x=gt_difficult, dtype=gt_box.dtype)
+            label = layers.concat([gt_label, gt_difficult, gt_box], axis=1)
+        else:
+            label = layers.concat([gt_label, gt_box], axis=1)

        # calculate mean average precision (mAP) of current mini-batch
        map = layers.detection_map(

--- a/python/paddle/fluid/framework.py
+++ b/python/paddle/fluid/framework.py
@@ -72,6 +72,8 @@ def convert_np_dtype_to_dtype_(np_dtype):
        return core.VarDesc.VarType.INT64
    elif dtype == np.bool:
        return core.VarDesc.VarType.BOOL
+    elif dtype == np.uint8:
+        return core.VarDesc.VarType.UINT8
    else:
        raise ValueError("Not supported numpy dtype " + str(dtype))


--- a/python/paddle/fluid/inferencer.py
+++ b/python/paddle/fluid/inferencer.py
@@ -12,11 +12,14 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

+import contextlib
+
 import core

 import executor
 import framework
 import io
+import parallel_executor
 import unique_name
 from trainer import check_and_get_place

@@ -24,40 +27,53 @@ __all__ = ['Inferencer', ]


 class Inferencer(object):
-    def __init__(self, infer_func, param_path, place=None):
+    def __init__(self, infer_func, param_path, place=None, parallel=False):
        """
        :param infer_func: a function that will return predict Variable
        :param param_path: the path where the inference model is saved by fluid.io.save_params
        :param place: place to do the inference
+        :param parallel: use parallel_executor to run the inference, it will use multi CPU/GPU.
        """
        self.param_path = param_path
        self.scope = core.Scope()
+        self.parallel = parallel
+        self.place = check_and_get_place(place)

        self.inference_program = framework.Program()
        with framework.program_guard(self.inference_program):
            with unique_name.guard():
                self.predict_var = infer_func()

-        self.exe = executor.Executor(check_and_get_place(place))
-        with executor.scope_guard(self.scope):
+        with self._prog_and_scope_guard():
            # load params from param_path into scope
-            io.load_params(self.exe, param_path, self.inference_program)
+            io.load_params(executor.Executor(self.place), param_path)
+
+        if parallel:
+            with self._prog_and_scope_guard():
+                self.exe = parallel_executor.ParallelExecutor(
+                    use_cuda=isinstance(self.place, core.CUDAPlace),
+                    loss_name=self.predict_var.name)
+        else:
+            self.exe = executor.Executor(self.place)

-    def infer(self, inputs, return_numpy=True):
+    def infer(self, inputs):
        """
        :param inputs: a map of {"input_name": input_var} that will be feed into the inference program
        to get the predict value
-        :param return_numpy: if return numpy value for row tensor
        :return: the predict value of the inference model
        """
        if not isinstance(inputs, dict):
            raise ValueError(
                "inputs should be a map of {'input_name': input_var}")

-        with executor.scope_guard(self.scope):
-            results = self.exe.run(self.inference_program,
-                                   feed=inputs,
-                                   fetch_list=[self.predict_var],
-                                   return_numpy=return_numpy)
+        with self._prog_and_scope_guard():
+            results = self.exe.run(feed=inputs,
+                                   fetch_list=[self.predict_var.name])

        return results
+
+    @contextlib.contextmanager
+    def _prog_and_scope_guard(self):
+        with framework.program_guard(main_program=self.inference_program):
+            with executor.scope_guard(self.scope):
+                yield
--- a/python/paddle/fluid/layers/control_flow.py
+++ b/python/paddle/fluid/layers/control_flow.py
@@ -49,6 +49,7 @@ __all__ = [
    'reorder_lod_tensor_by_rank',
    'ParallelDo',
    'Print',
+    'is_empty',
 ]


@@ -1562,3 +1563,40 @@ def reorder_lod_tensor_by_rank(x, rank_table):
                'RankTable': [rank_table]},
        outputs={'Out': [out]})
    return out
+
+
+def is_empty(x, cond=None, **ignored):
+    """
+    **Is Empty**
+
+    This layer returns the truth value of whether the variable is empty.
+
+    Args:
+        x(Variable): Operand of *is_empty*
+        cond(Variable|None): Optional output variable to store the result
+                             of *is_empty*
+
+    Returns:
+        Variable: The tensor variable storing the output of *is_empty*.
+
+    Raises:
+        TypeError: If input cond is not a variable, or cond's dtype is
+                   not bool
+
+    Examples:
+        .. code-block:: python
+
+          less = fluid.layers.is_empty(x=input)
+    """
+    helper = LayerHelper("is_empty", **locals())
+    if cond is None:
+        cond = helper.create_tmp_variable(dtype='bool')
+        cond.stop_gradient = True
+    elif not isinstance(cond, Variable):
+        raise TypeError("cond takes a variable")
+    elif cond.dtype != 'bool':
+        raise TypeError("The data type of cond must be bool")
+
+    helper.append_op(
+        type='is_empty', inputs={'X': [x]}, outputs={'Out': [cond]})
+    return cond
--- a/python/paddle/fluid/layers/detection.py
+++ b/python/paddle/fluid/layers/detection.py
@@ -23,6 +23,7 @@ import nn
 import math

 __all__ = [
+    'prior_box',
    'multi_box_head',
    'bipartite_match',
    'target_assign',
@@ -564,6 +565,115 @@ def ssd_loss(location,
    return loss


+def prior_box(input,
+              image,
+              min_sizes,
+              max_sizes=None,
+              aspect_ratios=[1.],
+              variance=[0.1, 0.1, 0.2, 0.2],
+              flip=False,
+              clip=False,
+              steps=[0.0, 0.0],
+              offset=0.5,
+              name=None):
+    """
+    **Prior box operator**
+
+    Generate prior boxes for SSD(Single Shot MultiBox Detector) algorithm.
+    Each position of the input produce N prior boxes, N is determined by
+    the count of min_sizes, max_sizes and aspect_ratios, The size of the
+    box is in range(min_size, max_size) interval, which is generated in
+    sequence according to the aspect_ratios.
+
+    Args:
+       input(Variable): The Input Variables, the format is NCHW.
+       image(Variable): The input image data of PriorBoxOp,
+            the layout is NCHW.
+       min_sizes(list|tuple|float value): min sizes of generated prior boxes.
+       max_sizes(list|tuple|None): max sizes of generated prior boxes.
+            Default: None.
+       aspect_ratios(list|tuple|float value): the aspect ratios of generated
+            prior boxes. Default: [1.].
+       variance(list|tuple): the variances to be encoded in prior boxes.
+            Default:[0.1, 0.1, 0.2, 0.2].
+       flip(bool): Whether to flip aspect ratios. Default:False.
+       clip(bool): Whether to clip out-of-boundary boxes. Default: False.
+       step(list|turple): Prior boxes step across width and height, If
+            step[0] == 0.0/step[1] == 0.0, the prior boxes step across
+            height/weight of the input will be automatically calculated.
+            Default: [0., 0.]
+       offset(float): Prior boxes center offset. Default: 0.5
+       name(str): Name of the prior box op. Default: None.
+
+    Returns:
+        boxes(Variable): the output prior boxes of PriorBox.
+             The layout is [H, W, num_priors, 4].
+             H is the height of input, W is the width of input,
+             num_priors is the total
+             box count of each position of input.
+        Variances(Variable): the expanded variances of PriorBox.
+             The layout is [H, W, num_priors, 4].
+             H is the height of input, W is the width of input
+             num_priors is the total
+             box count of each position of input
+
+
+    Examples:
+        .. code-block:: python
+            box, var = prior_box(
+            input=conv1,
+            image=images,
+            min_sizes=[100.],
+            flip=True,
+            clip=True)
+    """
+    helper = LayerHelper("prior_box", **locals())
+    dtype = helper.input_dtype()
+
+    def _is_list_or_tuple_(data):
+        return (isinstance(data, list) or isinstance(data, tuple))
+
+    if not _is_list_or_tuple_(min_sizes):
+        min_sizes = [min_sizes]
+    if not _is_list_or_tuple_(aspect_ratios):
+        aspect_ratios = [aspect_ratios]
+    if not (_is_list_or_tuple_(steps) and len(steps) == 2):
+        raise ValueError('steps should be a list or tuple ',
+                         'with length 2, (step_width, step_height).')
+
+    min_sizes = list(map(float, min_sizes))
+    aspect_ratios = list(map(float, aspect_ratios))
+    steps = list(map(float, steps))
+
+    attrs = {
+        'min_sizes': min_sizes,
+        'aspect_ratios': aspect_ratios,
+        'variances': variance,
+        'flip': flip,
+        'clip': clip,
+        'step_w': steps[0],
+        'step_h': steps[1],
+        'offset': offset
+    }
+    if max_sizes is not None and len(max_sizes) > 0 and max_sizes[0] > 0:
+        if not _is_list_or_tuple_(max_sizes):
+            max_sizes = [max_sizes]
+        attrs['max_sizes'] = max_sizes
+
+    box = helper.create_tmp_variable(dtype)
+    var = helper.create_tmp_variable(dtype)
+    helper.append_op(
+        type="prior_box",
+        inputs={"Input": input,
+                "Image": image},
+        outputs={"Boxes": box,
+                 "Variances": var},
+        attrs=attrs, )
+    box.stop_gradient = True
+    var.stop_gradient = True
+    return box, var
+
+
 def multi_box_head(inputs,
                   image,
                   base_size,
@@ -660,47 +770,6 @@ def multi_box_head(inputs,
            clip=True)
    """

-    def _prior_box_(input,
-                    image,
-                    min_sizes,
-                    max_sizes,
-                    aspect_ratios,
-                    variance,
-                    flip=False,
-                    clip=False,
-                    step_w=0.0,
-                    step_h=0.0,
-                    offset=0.5,
-                    name=None):
-        helper = LayerHelper("prior_box", **locals())
-        dtype = helper.input_dtype()
-
-        attrs = {
-            'min_sizes': min_sizes,
-            'aspect_ratios': aspect_ratios,
-            'variances': variance,
-            'flip': flip,
-            'clip': clip,
-            'step_w': step_w,
-            'step_h': step_h,
-            'offset': offset
-        }
-        if len(max_sizes) > 0 and max_sizes[0] > 0:
-            attrs['max_sizes'] = max_sizes
-
-        box = helper.create_tmp_variable(dtype)
-        var = helper.create_tmp_variable(dtype)
-        helper.append_op(
-            type="prior_box",
-            inputs={"Input": input,
-                    "Image": image},
-            outputs={"Boxes": box,
-                     "Variances": var},
-            attrs=attrs, )
-        box.stop_gradient = True
-        var.stop_gradient = True
-        return box, var
-
    def _reshape_with_axis_(input, axis=1):
        if not (axis > 0 and axis < len(input.shape)):
            raise ValueError("The axis should be smaller than "
@@ -777,11 +846,10 @@ def multi_box_head(inputs,
            aspect_ratio = aspect_ratios[i]
            if not _is_list_or_tuple_(aspect_ratio):
                aspect_ratio = [aspect_ratio]
+        step = [step_w[i] if step_w else 0.0, step_h[i] if step_w else 0.0]

-        box, var = _prior_box_(input, image, min_size, max_size, aspect_ratio,
-                               variance, flip, clip, step_w[i]
-                               if step_w else 0.0, step_h[i]
-                               if step_w else 0.0, offset)
+        box, var = prior_box(input, image, min_size, max_size, aspect_ratio,
+                             variance, flip, clip, step, offset)

        box_results.append(box)
        var_results.append(var)

--- a/python/paddle/fluid/layers/nn.py
+++ b/python/paddle/fluid/layers/nn.py
@@ -1329,6 +1329,8 @@ def sequence_pool(input, pool_type):
         sqrt   : out.data = [2.82, 6.93, 4.24], where 2.82=(1+3)/sqrt(2),
                    6.93=(2+4+6)/sqrt(3), 4.24=(5+1)/sqrt(2)
         max    : out.data = [3, 6, 5], where 3=max(1,3), 6=max(2,4,6), 5=max(5,1)
+         last   : out.data = [3, 6, 1], where 3=last(1,3), 6=last(2,4,6), 1=last(5,1)
+         first  : out.data = [1, 2, 5], where 1=first(1,3), 2=first(2,4,6), 5=first(5,1)

    Args:
        input(variable): The input variable which is a LoDTensor.
@@ -1348,6 +1350,8 @@ def sequence_pool(input, pool_type):
             sum_x = fluid.layers.sequence_pool(input=x, pool_type='sum')
             sqrt_x = fluid.layers.sequence_pool(input=x, pool_type='sqrt')
             max_x = fluid.layers.sequence_pool(input=x, pool_type='max')
+             last_x = fluid.layers.sequence_pool(input=x, pool_type='last')
+             first_x = fluid.layers.sequence_pool(input=x, pool_type='first')
    """
    helper = LayerHelper('sequence_pool', **locals())
    dtype = helper.input_dtype()
@@ -3263,35 +3267,35 @@ def smooth_l1(x, y, inside_weight=None, outside_weight=None, sigma=None):
    """
    **Smooth L1 Loss Operator. **

-    This operator computes the smooth l1 loss for X and Y.
+    This operator computes the smooth L1 loss for X and Y.
    The operator takes the first dimension of X and Y as batch size.
-    For each instance, it computes the smooth l1 loss element by element first
+    For each instance, it computes the smooth L1 loss element by element first
    and then sums all the losses. So the shape of Out is [batch_size, 1].

    Args:
        x (Variable): A tensor with rank at least 2. The input value of smooth
-            l1 loss op with shape [batch_size, dim1, ..., dimN].
+            L1 loss op with shape [batch_size, dim1, ..., dimN].
        y (Variable): A tensor with rank at least 2. The target value of smooth
-            l1 loss op with same shape as x.
+            L1 loss op with same shape as x.
        inside_weight (Variable|None):  A tensor with rank at least 2. This
            input is optional and should have same shape with x. If provided,
            the result of (x - y) will be multiplied by this tensor element by
            element.
        outside_weight (Variable|None): A tensor with rank at least 2. This
            input is optional and should have same shape with x. If provided,
-            the out smooth l1 loss will be multiplied by this tensor element
+            the out smooth L1 loss will be multiplied by this tensor element
            by element.
-        sigma (float|None): Hyper parameter of smooth l1 loss op. A float scalar
+        sigma (float|None): Hyper parameter of smooth L1 loss op. A float scalar
            with default value 1.0.
    Returns:
-        Variable: A tensor with rank be 2. The output smooth l1 loss with
+        Variable: A tensor with rank be 2. The output smooth L1 loss with
            shape [batch_size, 1].

    Examples:
        .. code-block:: python

            data = fluid.layers.data(name='data', shape=[128], dtype='float32')
-            label = fluid.layers.data(name='label', shape=[100], dtype='int64')
+            label = fluid.layers.data(name='label', shape=[100], dtype='float32')
            fc = fluid.layers.fc(input=data, size=100)
            out = fluid.layers.smooth_l1(x=fc, y=label)
    """
@@ -3769,13 +3773,13 @@ def label_smooth(label,

 def roi_pool(input, rois, pooled_height=1, pooled_width=1, spatial_scale=1.0):
    """
-    Region of interest pooling (also known as RoI pooling) is to perform 
+    Region of interest pooling (also known as RoI pooling) is to perform
        is to perform max pooling on inputs of nonuniform sizes to obtain
        fixed-size feature maps (e.g. 7*7).
-    The operator has three steps: 
-        1. Dividing each region proposal into equal-sized sections with 
-           the pooled_width and pooled_height 
-        2. Finding the largest value in each section 
+    The operator has three steps:
+        1. Dividing each region proposal into equal-sized sections with
+           the pooled_width and pooled_height
+        2. Finding the largest value in each section
        3. Copying these max values to the output buffer

    Args:
@@ -3783,8 +3787,8 @@ def roi_pool(input, rois, pooled_height=1, pooled_width=1, spatial_scale=1.0):
        rois (Variable): ROIs (Regions of Interest) to pool over. It should
                         be a 2-D one level LoTensor of shape [num_rois, 4].
                         The layout is [x1, y1, x2, y2], where (x1, y1)
-                         is the top left coordinates, and (x2, y2) is the 
-                         bottom right coordinates. The num_rois is the 
+                         is the top left coordinates, and (x2, y2) is the
+                         bottom right coordinates. The num_rois is the
                         total number of ROIs in this batch data.
        pooled_height (integer): The pooled output height. Default: 1
        pooled_width (integer): The pooled output width. Default: 1
@@ -3793,11 +3797,11 @@ def roi_pool(input, rois, pooled_height=1, pooled_width=1, spatial_scale=1.0):
                               to the scale used when pooling. Default: 1.0

    Returns:
-        pool_out (Variable): The output is a 4-D tensor of the shape 
+        pool_out (Variable): The output is a 4-D tensor of the shape
                             (num_rois, channels, pooled_h, pooled_w).

    Examples:
-            pool_out = fluid.layers.roi_pool(input=x, rois=rois, 7, 7, 1.0) 
+            pool_out = fluid.layers.roi_pool(input=x, rois=rois, 7, 7, 1.0)
    """
    helper = LayerHelper('roi_pool', **locals())
    dtype = helper.input_dtype()

--- a/python/paddle/fluid/tests/book/high-level-api/CMakeLists.txt
+++ b/python/paddle/fluid/tests/book/high-level-api/CMakeLists.txt
@@ -8,3 +8,5 @@ endforeach()

 add_subdirectory(fit_a_line)
 add_subdirectory(recognize_digits)
+add_subdirectory(image_classification)
+add_subdirectory(understand_sentiment)
--- a/python/paddle/fluid/tests/book/high-level-api/fit_a_line/test_fit_a_line.py
+++ b/python/paddle/fluid/tests/book/high-level-api/fit_a_line/test_fit_a_line.py
@@ -57,22 +57,20 @@ def train(use_cuda, train_program, save_dirname):
        optimizer=fluid.optimizer.SGD(learning_rate=0.001))

    def event_handler(event):
-        if isinstance(event, fluid.EndEpochEvent):
-            test_metrics = trainer.test(
-                reader=test_reader, feed_order=['x', 'y'])
-            print test_metrics
-            '''
-            
-            ...
-            ['25.768919467926025']
-            ['15.343549569447836']
-            ...
-            
-            '''
-            if float(test_metrics[0]) < 20.0:
+        if isinstance(event, fluid.EndStepEvent):
+            if event.step == 10:
+                test_metrics = trainer.test(
+                    reader=test_reader, feed_order=['x', 'y'])
+                print test_metrics
+                '''
+                ...
+                ['25.768919467926025']
+                ['15.343549569447836']
+                ...
+                '''
                if save_dirname is not None:
                    trainer.save_params(save_dirname)
-            return
+                trainer.stop()

    trainer.train(
        reader=train_reader,
@@ -94,7 +92,7 @@ def infer(use_cuda, inference_program, save_dirname=None):
    tensor_x = numpy.random.uniform(0, 10, [batch_size, 13]).astype("float32")

    results = inferencer.infer({'x': tensor_x})
-    print("infer results: ", results[0])
+    print("infer results: ", numpy.array(results[0]))


 def main(use_cuda):

--- a/python/paddle/fluid/tests/book/high-level-api/image_classification/CMakeLists.txt
+++ b/python/paddle/fluid/tests/book/high-level-api/image_classification/CMakeLists.txt
+file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
+string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
+
+# default test
+foreach(src ${TEST_OPS})
+    py_test(${src} SRCS ${src}.py)
+endforeach()
--- a/python/paddle/fluid/tests/book/high-level-api/image_classification/cifar10_small_test_set.py
+++ b/python/paddle/fluid/tests/book/high-level-api/image_classification/cifar10_small_test_set.py
+# Copyright (c) 2016 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.
+"""
+CIFAR dataset.
+
+This module will download dataset from
+https://www.cs.toronto.edu/~kriz/cifar.html and parse train/test set into
+paddle reader creators.
+
+The CIFAR-10 dataset consists of 60000 32x32 colour images in 10 classes,
+with 6000 images per class. There are 50000 training images and 10000 test
+images.
+
+The CIFAR-100 dataset is just like the CIFAR-10, except it has 100 classes
+containing 600 images each. There are 500 training images and 100 testing
+images per class.
+
+"""
+
+import cPickle
+import itertools
+import numpy
+import paddle.v2.dataset.common
+import tarfile
+
+__all__ = ['train10']
+
+URL_PREFIX = 'https://www.cs.toronto.edu/~kriz/'
+CIFAR10_URL = URL_PREFIX + 'cifar-10-python.tar.gz'
+CIFAR10_MD5 = 'c58f30108f718f92721af3b95e74349a'
+
+
+def reader_creator(filename, sub_name, batch_size=None):
+    def read_batch(batch):
+        data = batch['data']
+        labels = batch.get('labels', batch.get('fine_labels', None))
+        assert labels is not None
+        for sample, label in itertools.izip(data, labels):
+            yield (sample / 255.0).astype(numpy.float32), int(label)
+
+    def reader():
+        with tarfile.open(filename, mode='r') as f:
+            names = (each_item.name for each_item in f
+                     if sub_name in each_item.name)
+
+            batch_count = 0
+            for name in names:
+                batch = cPickle.load(f.extractfile(name))
+                for item in read_batch(batch):
+                    if isinstance(batch_size, int) and batch_count > batch_size:
+                        break
+                    batch_count += 1
+                    yield item
+
+    return reader
+
+
+def train10(batch_size=None):
+    """
+    CIFAR-10 training set creator.
+
+    It returns a reader creator, each sample in the reader is image pixels in
+    [0, 1] and label in [0, 9].
+
+    :return: Training reader creator
+    :rtype: callable
+    """
+    return reader_creator(
+        paddle.v2.dataset.common.download(CIFAR10_URL, 'cifar', CIFAR10_MD5),
+        'data_batch',
+        batch_size=batch_size)
--- a/python/paddle/fluid/tests/book/high-level-api/image_classification/notest_image_classification_resnet.py
+++ b/python/paddle/fluid/tests/book/high-level-api/image_classification/notest_image_classification_resnet.py
@@ -17,6 +17,7 @@ from __future__ import print_function
 import paddle
 import paddle.fluid as fluid
 import numpy
+import cifar10_small_test_set


 def resnet_cifar10(input, depth=32):
@@ -81,46 +82,50 @@ def train_network():
    cost = fluid.layers.cross_entropy(input=predict, label=label)
    avg_cost = fluid.layers.mean(cost)
    accuracy = fluid.layers.accuracy(input=predict, label=label)
-    return avg_cost, accuracy
+    return [avg_cost, accuracy]


-def train(use_cuda, save_path):
+def train(use_cuda, train_program, save_dirname):
    BATCH_SIZE = 128
    EPOCH_NUM = 1

    train_reader = paddle.batch(
        paddle.reader.shuffle(
-            paddle.dataset.cifar.train10(), buf_size=128 * 10),
+            cifar10_small_test_set.train10(batch_size=10), buf_size=128 * 10),
        batch_size=BATCH_SIZE)

    test_reader = paddle.batch(
        paddle.dataset.cifar.test10(), batch_size=BATCH_SIZE)

    def event_handler(event):
-        if isinstance(event, fluid.EndIteration):
-            if (event.batch_id % 10) == 0:
-                avg_cost, accuracy = trainer.test(reader=test_reader)
+        if isinstance(event, fluid.EndStepEvent):
+            avg_cost, accuracy = trainer.test(
+                reader=test_reader, feed_order=['pixel', 'label'])

-                print('BatchID {1:04}, Loss {2:2.2}, Acc {3:2.2}'.format(
-                    event.batch_id + 1, avg_cost, accuracy))
+            print('Loss {0:2.2}, Acc {1:2.2}'.format(avg_cost, accuracy))

-                if accuracy > 0.01:  # Low threshold for speeding up CI
-                    trainer.params.save(save_path)
-                    return
+            if accuracy > 0.01:  # Low threshold for speeding up CI
+                if save_dirname is not None:
+                    trainer.save_params(save_dirname)
+                return

    place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
    trainer = fluid.Trainer(
-        train_network,
+        train_func=train_program,
        optimizer=fluid.optimizer.Adam(learning_rate=0.001),
-        place=place,
-        event_handler=event_handler)
-    trainer.train(train_reader, EPOCH_NUM, event_handler=event_handler)
+        place=place)

+    trainer.train(
+        reader=train_reader,
+        num_epochs=EPOCH_NUM,
+        event_handler=event_handler,
+        feed_order=['pixel', 'label'])

-def infer(use_cuda, save_path):
-    params = fluid.Params(save_path)
+
+def infer(use_cuda, inference_program, save_dirname=None):
    place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
-    inferencer = fluid.Inferencer(inference_network, params, place=place)
+    inferencer = fluid.Inferencer(
+        infer_func=inference_program, param_path=save_dirname, place=place)

    # The input's dimension of conv should be 4-D or 5-D.
    # Use normilized image pixels as input data, which should be in the range
@@ -135,8 +140,14 @@ def main(use_cuda):
    if use_cuda and not fluid.core.is_compiled_with_cuda():
        return
    save_path = "image_classification_resnet.inference.model"
-    train(use_cuda, save_path)
-    infer(use_cuda, save_path)
+
+    train(
+        use_cuda=use_cuda, train_program=train_network, save_dirname=save_path)
+
+    infer(
+        use_cuda=use_cuda,
+        inference_program=inference_network,
+        save_dirname=save_path)


 if __name__ == '__main__':

--- a/python/paddle/fluid/tests/book/high-level-api/image_classification/notest_image_classification_vgg.py
+++ b/python/paddle/fluid/tests/book/high-level-api/image_classification/notest_image_classification_vgg.py
@@ -17,6 +17,7 @@ from __future__ import print_function
 import paddle
 import paddle.fluid as fluid
 import numpy
+import cifar10_small_test_set


 def vgg16_bn_drop(input):
@@ -60,46 +61,48 @@ def train_network():
    cost = fluid.layers.cross_entropy(input=predict, label=label)
    avg_cost = fluid.layers.mean(cost)
    accuracy = fluid.layers.accuracy(input=predict, label=label)
-    return avg_cost, accuracy
+    return [avg_cost, accuracy]


-def train(use_cuda, save_path):
+def train(use_cuda, train_program, save_dirname):
    BATCH_SIZE = 128
-    EPOCH_NUM = 1
-
    train_reader = paddle.batch(
        paddle.reader.shuffle(
-            paddle.dataset.cifar.train10(), buf_size=128 * 10),
+            cifar10_small_test_set.train10(batch_size=10), buf_size=128 * 10),
        batch_size=BATCH_SIZE)

    test_reader = paddle.batch(
        paddle.dataset.cifar.test10(), batch_size=BATCH_SIZE)

    def event_handler(event):
-        if isinstance(event, fluid.EndIteration):
-            if (event.batch_id % 10) == 0:
-                avg_cost, accuracy = trainer.test(reader=test_reader)
+        if isinstance(event, fluid.EndStepEvent):
+            avg_cost, accuracy = trainer.test(
+                reader=test_reader, feed_order=['pixel', 'label'])

-                print('BatchID {1:04}, Loss {2:2.2}, Acc {3:2.2}'.format(
-                    event.batch_id + 1, avg_cost, accuracy))
+            print('Loss {0:2.2}, Acc {1:2.2}'.format(avg_cost, accuracy))

-                if accuracy > 0.01:  # Low threshold for speeding up CI
-                    trainer.params.save(save_path)
-                    return
+            if accuracy > 0.01:  # Low threshold for speeding up CI
+                if save_dirname is not None:
+                    trainer.save_params(save_dirname)
+                return

    place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
    trainer = fluid.Trainer(
-        train_network,
-        optimizer=fluid.optimizer.Adam(learning_rate=0.001),
+        train_func=train_program,
        place=place,
-        event_handler=event_handler)
-    trainer.train(train_reader, EPOCH_NUM, event_handler=event_handler)
+        optimizer=fluid.optimizer.Adam(learning_rate=0.001))
+
+    trainer.train(
+        reader=train_reader,
+        num_epochs=1,
+        event_handler=event_handler,
+        feed_order=['pixel', 'label'])


-def infer(use_cuda, save_path):
-    params = fluid.Params(save_path)
+def infer(use_cuda, inference_program, save_dirname=None):
    place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
-    inferencer = fluid.Inferencer(inference_network, params, place=place)
+    inferencer = fluid.Inferencer(
+        infer_func=inference_program, param_path=save_dirname, place=place)

    # The input's dimension of conv should be 4-D or 5-D.
    # Use normilized image pixels as input data, which should be in the range
@@ -114,8 +117,14 @@ def main(use_cuda):
    if use_cuda and not fluid.core.is_compiled_with_cuda():
        return
    save_path = "image_classification_vgg.inference.model"
-    train(use_cuda, save_path)
-    infer(use_cuda, save_path)
+
+    train(
+        use_cuda=use_cuda, train_program=train_network, save_dirname=save_path)
+
+    infer(
+        use_cuda=use_cuda,
+        inference_program=inference_network,
+        save_dirname=save_path)


 if __name__ == '__main__':

--- a/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_conv.py
+++ b/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_conv.py
@@ -62,31 +62,31 @@ def train(use_cuda, train_program, save_dirname):
    optimizer = fluid.optimizer.Adam(learning_rate=0.001)

    trainer = fluid.Trainer(
-        train_func=train_program, place=place, optimizer=optimizer)
+        train_func=train_program,
+        place=place,
+        optimizer=optimizer,
+        parallel=True)

    def event_handler(event):
        if isinstance(event, fluid.EndEpochEvent):
            test_reader = paddle.batch(
                paddle.dataset.mnist.test(), batch_size=BATCH_SIZE)
-            test_metrics = trainer.test(
+            avg_cost, acc = trainer.test(
                reader=test_reader, feed_order=['img', 'label'])
-            avg_cost_set = test_metrics[0]
-            acc_set = test_metrics[1]
-
-            # get test acc and loss
-            acc = numpy.array(acc_set).mean()
-            avg_cost = numpy.array(avg_cost_set).mean()

            print("avg_cost: %s" % avg_cost)
            print("acc     : %s" % acc)

-            if float(acc) > 0.2:  # Smaller value to increase CI speed
+            if acc > 0.2:  # Smaller value to increase CI speed
                trainer.save_params(save_dirname)
            else:
                print('BatchID {0}, Test Loss {1:0.2}, Acc {2:0.2}'.format(
-                    event.epoch + 1, float(avg_cost), float(acc)))
-                if math.isnan(float(avg_cost)):
+                    event.epoch + 1, avg_cost, acc))
+                if math.isnan(avg_cost):
                    sys.exit("got NaN loss, training failed.")
+        elif isinstance(event, fluid.EndStepEvent):
+            print("Step {0}, Epoch {1} Metrics {2}".format(
+                event.step, event.epoch, map(numpy.array, event.metrics)))

    train_reader = paddle.batch(
        paddle.reader.shuffle(
@@ -112,7 +112,7 @@ def infer(use_cuda, inference_program, save_dirname=None):

    results = inferencer.infer({'img': tensor_img})

-    print("infer results: ", results[0])
+    print("infer results: ", numpy.array(results[0]))


 def main(use_cuda):
@@ -131,4 +131,4 @@ def main(use_cuda):

 if __name__ == '__main__':
    # for use_cuda in (False, True):
-    main(use_cuda=False)
+    main(use_cuda=True)
--- a/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_mlp.py
+++ b/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_mlp.py
@@ -55,24 +55,18 @@ def train(use_cuda, train_program, save_dirname):
        if isinstance(event, fluid.EndEpochEvent):
            test_reader = paddle.batch(
                paddle.dataset.mnist.test(), batch_size=BATCH_SIZE)
-            test_metrics = trainer.test(
+            avg_cost, acc = trainer.test(
                reader=test_reader, feed_order=['img', 'label'])
-            avg_cost_set = test_metrics[0]
-            acc_set = test_metrics[1]
-
-            # get test acc and loss
-            acc = numpy.array(acc_set).mean()
-            avg_cost = numpy.array(avg_cost_set).mean()

            print("avg_cost: %s" % avg_cost)
            print("acc     : %s" % acc)

-            if float(acc) > 0.2:  # Smaller value to increase CI speed
+            if acc > 0.2:  # Smaller value to increase CI speed
                trainer.save_params(save_dirname)
            else:
                print('BatchID {0}, Test Loss {1:0.2}, Acc {2:0.2}'.format(
-                    event.epoch + 1, float(avg_cost), float(acc)))
-                if math.isnan(float(avg_cost)):
+                    event.epoch + 1, avg_cost, acc))
+                if math.isnan(avg_cost):
                    sys.exit("got NaN loss, training failed.")

    train_reader = paddle.batch(
@@ -99,7 +93,7 @@ def infer(use_cuda, inference_program, save_dirname=None):

    results = inferencer.infer({'img': tensor_img})

-    print("infer results: ", results[0])
+    print("infer results: ", numpy.array(results[0]))


 def main(use_cuda):

--- a/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/CMakeLists.txt
+++ b/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/CMakeLists.txt
+file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
+string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
+
+# default test
+foreach(src ${TEST_OPS})
+    py_test(${src} SRCS ${src}.py)
+endforeach()
--- a/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/notest_understand_sentiment_stacked_lstm.py
+++ b/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/notest_understand_sentiment_stacked_lstm.py
@@ -17,11 +17,13 @@ from __future__ import print_function
 import paddle
 import paddle.fluid as fluid
 from functools import partial
+import numpy as np

 CLASS_DIM = 2
 EMB_DIM = 128
 HID_DIM = 512
 STACKED_NUM = 3
+BATCH_SIZE = 128


 def stacked_lstm_net(data, input_dim, class_dim, emb_dim, hid_dim, stacked_num):
@@ -50,7 +52,7 @@ def stacked_lstm_net(data, input_dim, class_dim, emb_dim, hid_dim, stacked_num):
    return prediction


-def inference_network(word_dict):
+def inference_program(word_dict):
    data = fluid.layers.data(
        name="words", shape=[1], dtype="int64", lod_level=1)

@@ -60,57 +62,71 @@ def inference_network(word_dict):
    return net


-def train_network(word_dict):
-    prediction = inference_network(word_dict)
+def train_program(word_dict):
+    prediction = inference_program(word_dict)
    label = fluid.layers.data(name="label", shape=[1], dtype="int64")
    cost = fluid.layers.cross_entropy(input=prediction, label=label)
    avg_cost = fluid.layers.mean(cost)
    accuracy = fluid.layers.accuracy(input=prediction, label=label)
-    return avg_cost, accuracy
+    return [avg_cost, accuracy]


-def train(use_cuda, save_path):
-    BATCH_SIZE = 128
-    EPOCH_NUM = 5
+def train(use_cuda, train_program, save_dirname):
+    place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
+    optimizer = fluid.optimizer.Adagrad(learning_rate=0.002)

    word_dict = paddle.dataset.imdb.word_dict()
+    trainer = fluid.Trainer(
+        train_func=partial(train_program, word_dict),
+        place=place,
+        optimizer=optimizer)

-    train_data = paddle.batch(
+    def event_handler(event):
+        if isinstance(event, fluid.EndEpochEvent):
+            test_reader = paddle.batch(
+                paddle.dataset.imdb.test(word_dict), batch_size=BATCH_SIZE)
+            avg_cost, acc = trainer.test(
+                reader=test_reader, feed_order=['words', 'label'])
+
+            print("avg_cost: %s" % avg_cost)
+            print("acc     : %s" % acc)
+
+            if acc > 0.2:  # Smaller value to increase CI speed
+                trainer.save_params(save_dirname)
+                trainer.stop()
+
+            else:
+                print('BatchID {0}, Test Loss {1:0.2}, Acc {2:0.2}'.format(
+                    event.epoch + 1, avg_cost, acc))
+                if math.isnan(avg_cost):
+                    sys.exit("got NaN loss, training failed.")
+        elif isinstance(event, fluid.EndStepEvent):
+            print("Step {0}, Epoch {1} Metrics {2}".format(
+                event.step, event.epoch, map(np.array, event.metrics)))
+            if event.step == 1:  # Run 2 iterations to speed CI
+                trainer.save_params(save_dirname)
+                trainer.stop()
+
+    train_reader = paddle.batch(
        paddle.reader.shuffle(
-            paddle.dataset.imdb.train(word_dict), buf_size=1000),
+            paddle.dataset.imdb.train(word_dict), buf_size=25000),
        batch_size=BATCH_SIZE)

-    test_data = paddle.batch(
-        paddle.dataset.imdb.test(word_dict), batch_size=BATCH_SIZE)
-
-    def event_handler(event):
-        if isinstance(event, fluid.EndIteration):
-            if (event.batch_id % 10) == 0:
-                avg_cost, accuracy = trainer.test(reader=test_data)
-
-                print('BatchID {1:04}, Loss {2:2.2}, Acc {3:2.2}'.format(
-                    event.batch_id + 1, avg_cost, accuracy))
+    trainer.train(
+        num_epochs=1,
+        event_handler=event_handler,
+        reader=train_reader,
+        feed_order=['words', 'label'])

-                if accuracy > 0.01:  # Low threshold for speeding up CI
-                    trainer.params.save(save_path)
-                    return

-    place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
-    trainer = fluid.Trainer(
-        partial(train_network, word_dict),
-        optimizer=fluid.optimizer.Adagrad(learning_rate=0.002),
-        place=place,
-        event_handler=event_handler)
-
-    trainer.train(train_data, EPOCH_NUM, event_handler=event_handler)
-
-
-def infer(use_cuda, save_path):
-    params = fluid.Params(save_path)
+def infer(use_cuda, inference_program, save_dirname=None):
    place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
    word_dict = paddle.dataset.imdb.word_dict()
+
    inferencer = fluid.Inferencer(
-        partial(inference_network, word_dict), params, place=place)
+        infer_func=partial(inference_program, word_dict),
+        param_path=save_dirname,
+        place=place)

    def create_random_lodtensor(lod, place, low, high):
        data = np.random.random_integers(low, high,
@@ -131,8 +147,8 @@ def main(use_cuda):
    if use_cuda and not fluid.core.is_compiled_with_cuda():
        return
    save_path = "understand_sentiment_stacked_lstm.inference.model"
-    train(use_cuda, save_path)
-    infer(use_cuda, save_path)
+    train(use_cuda, train_program, save_path)
+    infer(use_cuda, inference_program, save_path)


 if __name__ == '__main__':

--- a/python/paddle/fluid/tests/book/high-level-api/word2vec/no_test_word2vec_new_api.py
+++ b/python/paddle/fluid/tests/book/high-level-api/word2vec/no_test_word2vec_new_api.py
@@ -90,7 +90,7 @@ def train_program(is_sparse):
    return avg_cost


-def train(use_cuda, train_program, save_path):
+def train(use_cuda, train_program, save_dirname):
    train_reader = paddle.batch(
        paddle.dataset.imikolov.train(word_dict, N), BATCH_SIZE)
    test_reader = paddle.batch(
@@ -99,27 +99,36 @@ def train(use_cuda, train_program, save_path):
    place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()

    def event_handler(event):
-        if isinstance(event, fluid.EndEpochEvent):
-            outs = trainer.test(reader=test_reader)
+        if isinstance(event, fluid.EndStepEvent):
+            outs = trainer.test(
+                reader=test_reader,
+                feed_order=['firstw', 'secondw', 'thirdw', 'forthw', 'nextw'])
            avg_cost = outs[0]
            print("loss= ", avg_cost)

-            if avg_cost < 5.0:
-                trainer.save_params(save_path)
-                return
+            if avg_cost < 10.0:
+                trainer.save_params(save_dirname)
+                trainer.stop()
+
            if math.isnan(avg_cost):
                sys.exit("got NaN loss, training failed.")

    trainer = fluid.Trainer(
-        train_program, fluid.optimizer.SGD(learning_rate=0.001), place=place)
+        train_func=train_program,
+        optimizer=fluid.optimizer.SGD(learning_rate=0.001),
+        place=place)
+
    trainer.train(
-        reader=train_reader, num_epochs=1, event_handler=event_handler)
+        reader=train_reader,
+        num_epochs=1,
+        event_handler=event_handler,
+        feed_order=['firstw', 'secondw', 'thirdw', 'forthw', 'nextw'])


-def infer(use_cuda, inference_program, save_path):
+def infer(use_cuda, inference_program, save_dirname=None):
    place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
    inferencer = fluid.Inferencer(
-        infer_func=inference_program, param_path=save_path, place=place)
+        infer_func=inference_program, param_path=save_dirname, place=place)

    lod = [0, 1]
    first_word = create_random_lodtensor(lod, place, low=0, high=dict_size - 1)
@@ -142,9 +151,17 @@ def main(use_cuda, is_sparse):
    if use_cuda and not fluid.core.is_compiled_with_cuda():
        return

-    save_path = "word2vec.params"
-    train(use_cuda, partial(train_program, is_sparse), save_path)
-    infer(use_cuda, partial(inference_program, is_sparse), save_path)
+    save_path = "word2vec.inference.model"
+
+    train(
+        use_cuda=use_cuda,
+        train_program=partial(train_program, is_sparse),
+        save_dirname=save_path)
+
+    infer(
+        use_cuda=use_cuda,
+        inference_program=partial(inference_program, is_sparse),
+        save_dirname=save_path)


 if __name__ == '__main__':

--- a/python/paddle/fluid/tests/book/test_label_semantic_roles.py
+++ b/python/paddle/fluid/tests/book/test_label_semantic_roles.py
@@ -182,12 +182,6 @@ def train(use_cuda, save_dirname=None, is_local=True):
    crf_decode = fluid.layers.crf_decoding(
        input=feature_out, param_attr=fluid.ParamAttr(name='crfw'))

-    chunk_evaluator = fluid.evaluator.ChunkEvaluator(
-        input=crf_decode,
-        label=target,
-        chunk_scheme="IOB",
-        num_chunk_types=int(math.ceil((label_dict_len - 1) / 2.0)))
-
    train_data = paddle.batch(
        paddle.reader.shuffle(
            paddle.dataset.conll05.test(), buf_size=8192),
@@ -203,7 +197,6 @@ def train(use_cuda, save_dirname=None, is_local=True):

    def train_loop(main_program):
        exe.run(fluid.default_startup_program())
-
        embedding_param = fluid.global_scope().find_var(
            embedding_name).get_tensor()
        embedding_param.set(
@@ -213,27 +206,19 @@ def train(use_cuda, save_dirname=None, is_local=True):
        start_time = time.time()
        batch_id = 0
        for pass_id in xrange(PASS_NUM):
-            chunk_evaluator.reset(exe)
            for data in train_data():
-                cost, precision, recall, f1_score = exe.run(
-                    main_program,
-                    feed=feeder.feed(data),
-                    fetch_list=[avg_cost] + chunk_evaluator.metrics)
-                pass_precision, pass_recall, pass_f1_score = chunk_evaluator.eval(
-                    exe)
+                cost = exe.run(main_program,
+                               feed=feeder.feed(data),
+                               fetch_list=[avg_cost])
+                cost = cost[0]

                if batch_id % 10 == 0:
-                    print("avg_cost:" + str(cost) + " precision:" + str(
-                        precision) + " recall:" + str(recall) + " f1_score:" +
-                          str(f1_score) + " pass_precision:" + str(
-                              pass_precision) + " pass_recall:" + str(
-                                  pass_recall) + " pass_f1_score:" + str(
-                                      pass_f1_score))
+                    print("avg_cost:" + str(cost))
                    if batch_id != 0:
                        print("second per batch: " + str((time.time(
                        ) - start_time) / batch_id))
                    # Set the threshold low to speed up the CI test
-                    if float(pass_precision) > 0.01:
+                    if float(cost) < 60.0:
                        if save_dirname is not None:
                            # TODO(liuyiqun): Change the target to crf_decode
                            fluid.io.save_inference_model(save_dirname, [

--- a/python/paddle/fluid/tests/test_data_feeder.py
+++ b/python/paddle/fluid/tests/test_data_feeder.py
@@ -13,15 +13,62 @@
 # limitations under the License.

 import paddle.fluid as fluid
+import unittest


-def test_converter():
-    img = fluid.layers.data(name='image', shape=[1, 28, 28])
-    label = fluid.layers.data(name='label', shape=[1], dtype='int64')
-    feeder = fluid.DataFeeder([img, label], fluid.CPUPlace())
-    result = feeder.feed([[[0] * 784, [9]], [[1] * 784, [1]]])
-    print(result)
+class TestDataFeeder(unittest.TestCase):
+    def test_lod_level_0_converter(self):
+        img = fluid.layers.data(name='image', shape=[1, 28, 28])
+        label = fluid.layers.data(name='label', shape=[1], dtype='int64')
+        feeder = fluid.DataFeeder([img, label], fluid.CPUPlace())
+        result = feeder.feed([([0] * 784, [9]), ([1] * 784, [1])])
+        print(result)
+
+        self.assertEqual(result['image'].shape(), [2, 1, 28, 28])
+        self.assertEqual(result['label'].shape(), [2, 1])
+        self.assertEqual(result['image'].lod(), [])
+        self.assertEqual(result['label'].lod(), [])
+
+    def test_lod_level_1_converter(self):
+        # lod_level = 1
+        # each sentence has a different number of words
+        sentences = fluid.layers.data(
+            name='sentences', shape=[1], dtype='int64', lod_level=1)
+        label = fluid.layers.data(name='label', shape=[1], dtype='int64')
+        feeder = fluid.DataFeeder([sentences, label], fluid.CPUPlace())
+
+        # lod = [[0, 3, 5, 9]]
+        # data = [[1, 2, 3], [4, 5], [6, 7, 8, 9]]
+        # label = [1] * len(data)
+        result = feeder.feed(
+            [([1, 2, 3], [1]), ([4, 5], [1]), ([6, 7, 8, 9], [1])])
+        print(result)
+
+        self.assertEqual(result['sentences'].shape(), [9, 1])
+        self.assertEqual(result['label'].shape(), [3, 1])
+        self.assertEqual(result['sentences'].lod(), [[0, 3, 5, 9]])
+        self.assertEqual(result['label'].lod(), [])
+
+    def test_lod_level_2_converter(self):
+        # lod_level = 2
+        # paragraphs -> sentences -> words
+        paragraphs = fluid.layers.data(
+            name='paragraphs', shape=[1], dtype='int64', lod_level=2)
+        label = fluid.layers.data(name='label', shape=[1], dtype='int64')
+        feeder = fluid.DataFeeder([paragraphs, label], fluid.CPUPlace())
+
+        # lod = [[0, 2, 3], [0, 3, 5, 9]]
+        # data = [[[1, 2, 3], [4, 5]], [[6, 7, 8, 9]]]
+        # label = [1] * len(data)
+        result = feeder.feed(
+            [([[1, 2, 3], [4, 5]], [1]), ([[6, 7, 8, 9]], [1])])
+        print(result)
+
+        self.assertEqual(result['paragraphs'].shape(), [9, 1])
+        self.assertEqual(result['label'].shape(), [2, 1])
+        self.assertEqual(result['paragraphs'].lod(), [[0, 2, 3], [0, 3, 5, 9]])
+        self.assertEqual(result['label'].lod(), [])


 if __name__ == '__main__':
-    test_converter()
+    unittest.main()
--- a/python/paddle/fluid/tests/test_detection.py
+++ b/python/paddle/fluid/tests/test_detection.py
@@ -109,6 +109,24 @@ class TestDetection(unittest.TestCase):
        print(str(program))


+class TestPriorBox(unittest.TestCase):
+    def test_prior_box(self):
+        data_shape = [3, 224, 224]
+        images = fluid.layers.data(
+            name='pixel', shape=data_shape, dtype='float32')
+        conv1 = fluid.layers.conv2d(images, 3, 3, 2)
+        box, var = layers.prior_box(
+            input=conv1,
+            image=images,
+            min_sizes=[100.0],
+            aspect_ratios=[1.],
+            flip=True,
+            clip=True)
+        assert len(box.shape) == 4
+        assert box.shape == var.shape
+        assert box.shape[3] == 4
+
+
 class TestMultiBoxHead(unittest.TestCase):
    def test_multi_box_head(self):
        data_shape = [3, 224, 224]

--- a/python/paddle/fluid/tests/unittests/CMakeLists.txt
+++ b/python/paddle/fluid/tests/unittests/CMakeLists.txt
@@ -28,11 +28,11 @@ function(py_test_modules TARGET_NAME)
  if(WITH_TESTING)
    set(options "")
    set(oneValueArgs "")
-    set(multiValueArgs MODULES DEPS ARGS ENVS)
+    set(multiValueArgs MODULES DEPS ENVS)
    cmake_parse_arguments(py_test_modules "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
    add_test(NAME ${TARGET_NAME}
             COMMAND env PYTHONPATH=${PADDLE_BINARY_DIR}/python ${py_test_modules_ENVS}
-             ${PYTHON_EXECUTABLE} -u -m unittest --verbose ${py_test_modules_MODULES} ${py_test_modules_ARGS}
+             ${PYTHON_EXECUTABLE} ${PADDLE_SOURCE_DIR}/tools/test_runner.py ${py_test_modules_MODULES}
             WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
  endif()
 endfunction()
@@ -66,6 +66,7 @@ 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)
+list(REMOVE_ITEM TEST_OPS test_network_with_dtype)

 # tests that can be bundled together in one python process for speed.
 if(WITH_FAST_BUNDLE_TEST)
@@ -83,6 +84,7 @@ py_test_modules(test_parallel_executor MODULES test_parallel_executor)
 py_test_modules(test_warpctc_op MODULES test_warpctc_op ENVS FLAGS_warpctc_dir=${WARPCTC_LIB_DIR})
 py_test_modules(test_train_dyn_rnn MODULES test_dyn_rnn)
 py_test_modules(test_mul_op MODULES test_mul_op)
+py_test_modules(test_network_with_dtype MODULES test_network_with_dtype)

 # tests that need to be run in separate process.
 py_test_modules(test_multihead_attention MODULES test_multihead_attention)

--- a/python/paddle/fluid/tests/unittests/test_detection_map_op.py
+++ b/python/paddle/fluid/tests/unittests/test_detection_map_op.py
@@ -160,7 +160,9 @@ class TestDetectionMAPOp(OpTest):
        label_count, true_pos, false_pos = get_input_pos(
            self.class_pos_count, self.true_pos, self.true_pos_lod,
            self.false_pos, self.false_pos_lod)
-        for (label, difficult, xmin, ymin, xmax, ymax) in self.label:
+        for v in self.label:
+            label = v[0]
+            difficult = False if len(v) == 5 else v[1]
            if self.evaluate_difficult:
                label_count[label] += 1
            elif not difficult:
@@ -245,6 +247,15 @@ class TestDetectionMAPOpSkipDiff(TestDetectionMAPOp):
                       [2, 0.8, 0, 1], [2, 0.1, 1, 0], [3, 0.2, 0, 1]]


+class TestDetectionMAPOpWithoutDiff(TestDetectionMAPOp):
+    def init_test_case(self):
+        super(TestDetectionMAPOpWithoutDiff, self).init_test_case()
+
+        # label xmin ymin xmax ymax
+        self.label = [[1, 0.1, 0.1, 0.3, 0.3], [1, 0.6, 0.6, 0.8, 0.8],
+                      [2, 0.3, 0.3, 0.6, 0.5], [1, 0.7, 0.1, 0.9, 0.3]]
+
+
 class TestDetectionMAPOp11Point(TestDetectionMAPOp):
    def init_test_case(self):
        super(TestDetectionMAPOp11Point, self).init_test_case()

--- a/python/paddle/fluid/tests/unittests/test_dist_train.py
+++ b/python/paddle/fluid/tests/unittests/test_dist_train.py
@@ -52,15 +52,18 @@ class TestSendOp(unittest.TestCase):
            serv = layers.ListenAndServ(
                "127.0.0.1:0", ["X"], optimizer_mode=False)
            with serv.do():
+                out_var = main.global_block().create_var(
+                    name="scale_0.tmp_0",
+                    psersistable=True,
+                    dtype="float32",
+                    shape=[32, 32])
                x = layers.data(
                    shape=[32, 32],
                    dtype='float32',
                    name="X",
                    append_batch_size=False)
                fluid.initializer.Constant(value=1.0)(x, main.global_block())
-                o = layers.scale(x=x, scale=10.0)
-            main.global_block().create_var(
-                name=o.name, psersistable=False, dtype=o.dtype, shape=o.shape)
+                layers.scale(x=x, scale=10.0, out=out_var)

        self.server_exe = fluid.Executor(place)
        self.server_exe.run(main)

--- a/python/paddle/fluid/tests/unittests/test_is_empty_op.py
+++ b/python/paddle/fluid/tests/unittests/test_is_empty_op.py
@@ -14,42 +14,24 @@

 import unittest
 import numpy as np
-from paddle.fluid.op import Operator
-import paddle.fluid.core as core
+from op_test import OpTest


-def create_tensor(scope, name, np_data):
-    tensor = scope.var(name).get_tensor()
-    tensor.set_dims(np_data.shape)
-    tensor.set(np_data, core.CPUPlace())
-    return tensor
-
-
-class TestIsEmptyOp(unittest.TestCase):
+class TestEmpty(OpTest):
    def setUp(self):
-        self.scope = core.Scope()
-        # create input variables
-        np_data0 = np.array([0, 1, 2])
-        create_tensor(self.scope, "X0", np_data0)
-
-        np_data1 = np.array([1])
-        t = create_tensor(self.scope, "X1", np_data1)
-        t.set_dims([0])
+        self.op_type = "is_empty"
+        self.inputs = {'X': np.array([1, 2, 3])}
+        self.outputs = {'Out': np.array([False])}

-        # create output variables
-        self.scope.var("out")
+    def test_check_output(self):
+        self.check_output()

-    def test_no_empty(self):
-        self.one_case("X0", False)

-    def test_empty(self):
-        self.one_case("X1", True)
-
-    def one_case(self, input, target):
-        op = Operator(type="is_empty", X=input, Out="out")
-        op.run(self.scope, core.CPUPlace())
-        out = self.scope.var("out").get_tensor()
-        self.assertEqual(np.array(out)[0], target)
+class TestNotEmpty(TestEmpty):
+    def setUp(self):
+        self.op_type = "is_empty"
+        self.inputs = {'X': np.array([])}
+        self.outputs = {'Out': np.array([True])}


 if __name__ == "__main__":

--- a/python/paddle/fluid/tests/unittests/test_network_with_dtype.py
+++ b/python/paddle/fluid/tests/unittests/test_network_with_dtype.py
@@ -24,33 +24,30 @@ BATCH_SIZE = 20


 class TestNetWithDtype(unittest.TestCase):
-    def set_network(self):
+    def setUp(self):
        self.dtype = "float64"
        self.init_dtype()
-        main = fluid.Program()
-        with fluid.program_guard(main):
-            self.x = fluid.layers.data(name='x', shape=[13], dtype=self.dtype)
-            self.y = fluid.layers.data(name='y', shape=[1], dtype=self.dtype)
-            y_predict = fluid.layers.fc(input=self.x, size=1, act=None)

-            cost = fluid.layers.square_error_cost(input=y_predict, label=self.y)
+    def run_net_on_place(self, place):
+        main = fluid.Program()
+        startup = fluid.Program()
+        with fluid.program_guard(main, startup):
+            x = fluid.layers.data(name='x', shape=[13], dtype=self.dtype)
+            y = fluid.layers.data(name='y', shape=[1], dtype=self.dtype)
+            y_predict = fluid.layers.fc(input=x, size=1, act=None)
+            cost = fluid.layers.square_error_cost(input=y_predict, label=y)
            avg_cost = fluid.layers.mean(cost)
-        self.program = main
-        self.fetch_list = [avg_cost]
+            sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001)
+            sgd_optimizer.minimize(avg_cost)

-        sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001)
-        sgd_optimizer.minimize(avg_cost)
-
-    def run_net_on_place(self, place):
+        fetch_list = [avg_cost]
        train_reader = paddle.batch(
            paddle.dataset.uci_housing.train(), batch_size=BATCH_SIZE)
-        feeder = fluid.DataFeeder(place=place, feed_list=[self.x, self.y])
+        feeder = fluid.DataFeeder(place=place, feed_list=[x, y])
        exe = fluid.Executor(place)
-        exe.run(fluid.default_startup_program())
+        exe.run(startup)
        for data in train_reader():
-            exe.run(self.program,
-                    feed=feeder.feed(data),
-                    fetch_list=self.fetch_list)
+            exe.run(main, feed=feeder.feed(data), fetch_list=fetch_list)
            # the main program is runable, the datatype is fully supported
            break

@@ -58,14 +55,12 @@ class TestNetWithDtype(unittest.TestCase):
        pass

    def test_cpu(self):
-        self.set_network()
        place = fluid.CPUPlace()
        self.run_net_on_place(place)

    def test_gpu(self):
        if not core.is_compiled_with_cuda():
            return
-        self.set_network()
        place = fluid.CUDAPlace(0)
        self.run_net_on_place(place)


--- a/python/paddle/fluid/tests/unittests/test_parallel_executor.py
+++ b/python/paddle/fluid/tests/unittests/test_parallel_executor.py
@@ -775,7 +775,7 @@ class TestCRFModel(unittest.TestCase):
        build_strategy = fluid.BuildStrategy()
        build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce
        self.check_network_convergence(
-            is_sparse=False, build_strategy=build_strategy)
+            is_sparse=True, build_strategy=build_strategy)

    def test_update_dense_parameter_reduce(self):
        build_strategy = fluid.BuildStrategy()
@@ -849,8 +849,7 @@ class TestFetchOp(unittest.TestCase):
                    assert not math.isnan(np.sum(ret[i])) and \
                           not math.isinf(np.sum(ret[i]))

-    @unittest.skip("this test is buggy")
-    def test_feed(self):
+    def test_fetch_op(self):
        tst_reader = paddle.batch(flowers.test(use_xmap=False), batch_size=16)
        tst_reader_iter = tst_reader()


--- a/python/paddle/fluid/trainer.py
+++ b/python/paddle/fluid/trainer.py
@@ -12,17 +12,18 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

+import contextlib
 import os
+
 import core
-import framework
-import executor
+
 import data_feeder
-import contextlib
+import executor
+import framework
 import io
-import unique_name
-
 # optimizer is same as the parameter of Trainer.__init__. Rename it to opt_module
 import optimizer as opt_module
+import parallel_executor
 from transpiler import distribute_transpiler

 __all__ = [
@@ -48,12 +49,14 @@ class BeginStepEvent(object):
    def __init__(self, epoch_id, step_id):
        self.epoch = epoch_id
        self.step = step_id
+        self.fetch_metrics = True


 class EndStepEvent(object):
-    def __init__(self, epoch_id, step_id):
+    def __init__(self, epoch_id, step_id, metrics):
        self.epoch = epoch_id
        self.step = step_id
+        self.metrics = metrics


 def check_and_get_place(place):
@@ -87,12 +90,18 @@ class Trainer(object):

    Args:
        train_func(callable): A function which will return loss. The loss must be a scalar.
-        infer_func(callable): A function which will return predict, used to save inference model
        optimizer(optimizer.Optimizer): The optimizer should be an instance of Optimizer
        place: The device place of this trainer.
    """

-    def __init__(self, train_func, optimizer, param_path=None, place=None):
+    def __init__(self,
+                 train_func,
+                 optimizer,
+                 param_path=None,
+                 place=None,
+                 parallel=False):
+        self.__stop = False
+        self.parallel = parallel
        # 1. we need to generate a framework.Program by calling
        # program_func. Reference: fluid.program_guard in
        # test_word2vec.py
@@ -106,14 +115,14 @@ class Trainer(object):

        with framework.program_guard(self.train_program, self.startup_program):
            program_func_outs = train_func()
-            self.test_outputs = program_func_outs if isinstance(
+            self.train_func_outputs = program_func_outs if isinstance(
                program_func_outs, list) else [program_func_outs]
            self.test_program = self.train_program.clone()
            if not isinstance(optimizer, opt_module.Optimizer):
                raise TypeError(
                    "The optimizer should be an instance of Optimizer")
            # The fisrt element of program_func_outs is loss.
-            loss = self.test_outputs[0]
+            loss = self.train_func_outputs[0]
            optimize_ops, params_grads = optimizer.minimize(loss)

        self.place = check_and_get_place(place)
@@ -131,7 +140,40 @@ class Trainer(object):
            # load params from param_path into scope
            io.load_persistables(exe, dirname=param_path)

+    def _transpile_nccl2_dist(self):
+        # PADDLE_TRAINER_IPS
+        if "PADDLE_TRAINER_IPS" not in os.environ:
+            self.nccl_id_var = None
+        else:
+            self.trainer_id = int(os.getenv("PADDLE_TRAINER_ID"))
+            port = os.getenv("PADDLE_PSERVER_PORT")
+            worker_ips = os.getenv("PADDLE_TRAINER_IPS")
+            worker_endpoints = []
+            for ip in worker_ips.split(","):
+                worker_endpoints.append(':'.join([ip, port]))
+            self.num_trainers = len(worker_endpoints)
+            current_endpoint = os.getenv("POD_IP") + ":" + port
+            worker_endpoints.remove(current_endpoint)
+            # TODO(wuyi): use self.nccl_id_var, self.num_trainers and self.trainer_id
+            # in ParallelExecutor to start
+            # distributed training using NCCL2
+            self.nccl_id_var = self.startup_program.global_block().create_var(
+                name="NCCLID", persistable=True, type=core.VarDesc.VarType.RAW)
+            self.startup_program.global_block().append_op(
+                type="gen_nccl_id",
+                inputs={},
+                outputs={"NCCLID": self.nccl_id_var},
+                attrs={
+                    "endpoint": current_endpoint,
+                    "endpoint_list": worker_endpoints,
+                    "trainer_id": self.trainer_id
+                })
+
    def _dist_transpile_if_necessary(self, optimize_ops, params_grads):
+        self._transpile_nccl2_dist()
+        if self.nccl_id_var != None:
+            return
+
        if "PADDLE_TRAINING_ROLE" not in os.environ:
            return

@@ -169,12 +211,13 @@ class Trainer(object):
                    'TRAINING_ROLE environment variable must be either TRAINER or PSERVER'
                )

-    def train(self,
-              num_epochs,
-              event_handler,
-              reader,
-              feed_order,
-              parallel=False):
+    def stop(self):
+        """
+        stop training
+        """
+        self.__stop = True
+
+    def train(self, num_epochs, event_handler, reader=None, feed_order=None):
        """
        Train the model.

@@ -182,25 +225,24 @@ class Trainer(object):
            num_epochs: The number of epoch. An epoch will process all data in reader
            event_handler: The event handler. A function with type (ev:Event)->void
            reader:
-            parallel: True if use multi-CPUs or multi-GPUs
            feed_order: Feeding order of reader. None will following the defining
                order in program

        Returns:

        """
-        if parallel:
-            raise NotImplementedError(
-                "Parallel Executor version of trainer is not implemented")
-
        training_role = os.getenv("PADDLE_TRAINING_ROLE", "")
        if training_role == "PSERVER":
            with self._prog_and_scope_guard():
                exe = executor.Executor(self.place)
                exe.run()
                return
-
-        self._train_by_executor(num_epochs, event_handler, reader, feed_order)
+        if self.parallel:
+            self._train_by_parallel_executor(num_epochs, event_handler, reader,
+                                             feed_order)
+        else:
+            self._train_by_executor(num_epochs, event_handler, reader,
+                                    feed_order)

    def test(self, reader, feed_order):
        """
@@ -212,7 +254,8 @@ class Trainer(object):
                order in program
        """

-        return self._test_by_executor(reader, feed_order, self.test_outputs)
+        return self._test_by_executor(reader, feed_order,
+                                      self.train_func_outputs)

    def save_params(self, param_path):
        # reference: save_persistables in io.py
@@ -246,13 +289,27 @@ class Trainer(object):
            feeder = data_feeder.DataFeeder(
                feed_list=feed_var_list, place=self.place)
            exe = executor.Executor(self.place)
-            for epoch_id in range(num_epochs):
-                event_handler(BeginEpochEvent(epoch_id))
-                for step_id, data in enumerate(reader()):
-                    event_handler(BeginStepEvent(epoch_id, step_id))
-                    exe.run(feed=feeder.feed(data), fetch_list=[])
-                    event_handler(EndStepEvent(epoch_id, step_id))
-                event_handler(EndEpochEvent(epoch_id))
+            reader = feeder.decorate_reader(reader, multi_devices=False)
+            self._train_by_any_executor(event_handler, exe, num_epochs, reader)
+
+    def _train_by_any_executor(self, event_handler, exe, num_epochs, reader):
+        for epoch_id in range(num_epochs):
+            event_handler(BeginEpochEvent(epoch_id))
+            for step_id, data in enumerate(reader()):
+                if self.__stop:
+                    return
+                begin_event = BeginStepEvent(epoch_id, step_id)
+                event_handler(begin_event)
+                if begin_event.fetch_metrics:
+                    metrics = exe.run(feed=data,
+                                      fetch_list=[
+                                          var.name
+                                          for var in self.train_func_outputs
+                                      ])
+                else:
+                    metrics = exe.run(feed=data, fetch_list=[])
+                event_handler(EndStepEvent(epoch_id, step_id, metrics))
+            event_handler(EndEpochEvent(epoch_id))

    def _test_by_executor(self, reader, feed_order, fetch_list):
        with executor.scope_guard(self.scope):
@@ -271,6 +328,26 @@ class Trainer(object):

            return [x / count for x in accumulated]

+    def _train_by_parallel_executor(self, num_epochs, event_handler, reader,
+                                    feed_order):
+        with self._prog_and_scope_guard():
+            pe = self._get_or_create_parallel_executor()
+            feed_var_list = build_feed_var_list(self.train_program, feed_order)
+            feeder = data_feeder.DataFeeder(
+                feed_list=feed_var_list, place=self.place)
+            reader = feeder.decorate_reader(reader, multi_devices=True)
+            self._train_by_any_executor(event_handler, pe, num_epochs, reader)
+
+    def _get_parallel_executor(self):
+        return getattr(self, 'parallel_executor', None)
+
+    def _get_or_create_parallel_executor(self):
+        if self._get_parallel_executor() is None:
+            self.parallel_executor = parallel_executor.ParallelExecutor(
+                use_cuda=isinstance(self.place, core.CUDAPlace),
+                loss_name=self.train_func_outputs[0].name)
+        return self._get_parallel_executor()
+

 def build_feed_var_list(program, feed_order):
    if not isinstance(program, framework.Program):

--- a/python/paddle/fluid/transpiler/memory_optimization_transpiler.py
+++ b/python/paddle/fluid/transpiler/memory_optimization_transpiler.py
@@ -24,7 +24,8 @@ dtype_to_size = {
    core.VarDesc.VarType.INT16: 2,
    core.VarDesc.VarType.INT32: 4,
    core.VarDesc.VarType.INT64: 8,
-    core.VarDesc.VarType.BOOL: 1
+    core.VarDesc.VarType.BOOL: 1,
+    core.VarDesc.VarType.UINT8: 1,
 }

 SUB_BLOCK_OPS = [

--- a/tools/test_runner.py
+++ b/tools/test_runner.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 os
+import sys
+import paddle.fluid as fluid
+import importlib
+import cStringIO
+
+
+def main():
+    sys.path.append(os.getcwd())
+    some_test_failed = False
+    for module_name in sys.argv[1:]:
+        buffer = cStringIO.StringIO()
+        main = fluid.Program()
+        startup = fluid.Program()
+        scope = fluid.core.Scope()
+        with fluid.program_guard(main, startup):
+            with fluid.scope_guard(scope):
+                with fluid.unique_name.guard():
+                    test_loader = unittest.TestLoader()
+                    module = importlib.import_module(module_name)
+                    tests = test_loader.loadTestsFromModule(module)
+                    res = unittest.TextTestRunner(stream=buffer).run(tests)
+                    if not res.wasSuccessful():
+                        some_test_failed = True
+                        print >> sys.stderr, module_name, 'failed\n', buffer.getvalue(
+                        )
+
+    if some_test_failed:
+        exit(1)
+
+
+if __name__ == '__main__':
+    main()
--- a/tools/timeline.py
+++ b/tools/timeline.py
@@ -171,7 +171,7 @@ if args.timeline_path:

 profile_paths = profile_path.split(',')
 profile_dict = dict()
-if len(profile_path) == 1:
+if len(profile_paths) == 1:
    with open(profile_path, 'r') as f:
        profile_s = f.read()
        profile_pb = profiler_pb2.Profile()