test=develop

.gitignore

@@ -28,3 +28,4 @@ third_party/
 build_*
 # clion workspace.
 cmake-build-*
+model_test

CMakeLists.txt

@@ -62,13 +62,12 @@ option(WITH_DISTRIBUTE  "Compile with distributed support"              OFF)
 option(USE_EIGEN_FOR_BLAS   "Use matrix multiplication in Eigen"        OFF)
 option(EIGEN_USE_THREADS "Compile with multi-threaded Eigen"            OFF)
 option(WITH_ARM_FP16    "Use half precision support on armv8.2-a cpu"   OFF)
-option(WITH_FAST_BUNDLE_TEST    "Bundle tests that can be run in a single process together to reduce launch overhead"   OFF)
 option(WITH_CONTRIB     "Compile the third-party contributation"        OFF)
 option(REPLACE_ENFORCE_GLOG "Replace PADDLE_ENFORCE with glog/CHECK for better debug." OFF)
 option(WITH_ANAKIN      "Compile with Anakin library"                   OFF)
 option(WITH_GRPC     "Use grpc as the default rpc framework"            ${WITH_DISTRIBUTE})
 option(WITH_BRPC_RDMA     "Use brpc rdma as the rpc protocal"           OFF)
-option(WITH_INFERENCE    "Compile fluid inference library"              ON)
+option(ON_INFER         "Turn on inference optimization."               OFF)
 option(WITH_INFERENCE_API_TEST   "Test fluid inference high-level api interface"  OFF)
 option(WITH_SYSTEM_BLAS   "Use system blas library"           OFF)
 option(PY_VERSION       "Compile PaddlePaddle with python3 support"     ${PY_VERSION})
@@ -179,6 +178,7 @@ include(external/eigen)     # download eigen3
 include(external/pybind11)  # download pybind11
 include(external/cares)
 include(external/cub)
+include(external/xxhash)    # download xxhash
 
 if (NOT WIN32)
 # there is no official support of snappystream, warpctc, nccl, cupti in windows
@@ -301,3 +301,11 @@ if(WITH_DOC)
     find_python_module(recommonmark REQUIRED)
     add_subdirectory(doc)
 endif()
+
+if (ON_INFER)
+    message(STATUS "On inference mode, will take place some specific optimization.")
+    add_definitions(-DPADDLE_ON_INFERENCE)
+else()
+    #TODO(luotao), combine this warning with `make inference_lib_dist` command.
+    message(WARNING "On inference mode, will take place some specific optimization. Turn on the ON_INFER flag when building inference_lib only.")
+endif()

Dockerfile

@@ -75,14 +75,14 @@ RUN pip3 install -U wheel && \
     pip3 install -U docopt PyYAML sphinx==1.5.6 && \
     pip3 install sphinx-rtd-theme==0.1.9 recommonmark && \
     easy_install -U pip && \
-    pip install -U wheel && \
+    pip install -U pip setuptools wheel && \
     pip install -U docopt PyYAML sphinx==1.5.6 && \
     pip install sphinx-rtd-theme==0.1.9 recommonmark
 
-RUN pip3 install pre-commit 'ipython==5.3.0' && \
+RUN pip3 install 'pre-commit==1.10.4' 'ipython==5.3.0' && \
     pip3 install 'ipykernel==4.6.0' 'jupyter==1.0.0' && \
     pip3 install opencv-python && \
-    pip install pre-commit 'ipython==5.3.0' && \
+    pip install 'pre-commit==1.10.4' 'ipython==5.3.0' && \
     pip install 'ipykernel==4.6.0' 'jupyter==1.0.0' && \
     pip install opencv-python
 

README.md

@@ -2,8 +2,8 @@
 
 
 [![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle)
-[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.0/getstarted/index_en.html)
-[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.0/beginners_guide/index.html)
+[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.1/getstarted/index_en.html)
+[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.1/beginners_guide/index.html)
 [![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle.svg)](https://github.com/PaddlePaddle/Paddle/releases)
 [![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE)
 
@@ -19,7 +19,7 @@ Our vision is to enable deep learning for everyone via PaddlePaddle.
 Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddle/releases) to track the latest feature of PaddlePaddle.
 
 
-### Latest PaddlePaddle Release: [Fluid 1.0.1](https://github.com/PaddlePaddle/Paddle/tree/release/1.0.0)
+### Latest PaddlePaddle Release: [Fluid 1.1.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.1)
 ### Install Latest Stable Release:
 ```
 # Linux CPU
@@ -27,9 +27,9 @@ pip install paddlepaddle
 # Linux GPU cuda9cudnn7
 pip install paddlepaddle-gpu
 # Linux GPU cuda8cudnn7
-pip install paddlepaddle-gpu==1.0.1.post87
+pip install paddlepaddle-gpu==1.1.0.post87
 # Linux GPU cuda8cudnn5
-pip install paddlepaddle-gpu==1.0.1.post85
+pip install paddlepaddle-gpu==1.1.0.post85
 
 # For installation on other platform, refer to http://paddlepaddle.org/
 ```
@@ -76,26 +76,26 @@ pip install paddlepaddle-gpu==1.0.1.post85
 
 ## Installation
 
-It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/1.0/beginners_guide/index.html) on our website.
+It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/1.1/beginners_guide/index.html) on our website.
 
 ## Documentation
 
-We provide [English](http://paddlepaddle.org/documentation/docs/en/1.0.0/getstarted/index_en.html) and
-[Chinese](http://paddlepaddle.org/documentation/docs/zh/1.0/beginners_guide/index.html) documentation.
+We provide [English](http://paddlepaddle.org/documentation/docs/en/1.1/getstarted/index_en.html) and
+[Chinese](http://paddlepaddle.org/documentation/docs/zh/1.1/beginners_guide/index.html) documentation.
 
 - [Deep Learning 101](https://github.com/PaddlePaddle/book)
 
   You might want to start from this online interactive book that can run in a Jupyter Notebook.
 
-- [Distributed Training](http://paddlepaddle.org/documentation/docs/zh/1.0/user_guides/howto/training/cluster_howto.html)
+- [Distributed Training](http://paddlepaddle.org/documentation/docs/zh/1.1/user_guides/howto/training/cluster_howto.html)
 
   You can run distributed training jobs on MPI clusters.
 
-- [Python API](http://paddlepaddle.org/documentation/api/zh/1.0/fluid.html)
+- [Python API](http://paddlepaddle.org/documentation/api/zh/1.1/fluid.html)
 
    Our new API enables much shorter programs.
 
-- [How to Contribute](http://paddlepaddle.org/documentation/docs/zh/1.0/advanced_usage/development/contribute_to_paddle.html)
+- [How to Contribute](http://paddlepaddle.org/documentation/docs/zh/1.1/advanced_usage/development/contribute_to_paddle.html)
 
    We appreciate your contributions!
 

benchmark/fluid/args.py

@@ -142,5 +142,10 @@ def parse_args():
         choices=['reduce', 'all_reduce'],
         default='all_reduce',
         help='Specify the reduce strategy, can be reduce, all_reduce')
+    parser.add_argument(
+        '--fuse_broadcast_op',
+        action='store_true',
+        help='If set, would fuse multiple broadcast operators into one fused_broadcast operator.'
+    )
     args = parser.parse_args()
     return args

benchmark/fluid/fluid_benchmark.py

@@ -177,6 +177,7 @@ def train_parallel(train_args, test_args, args, train_prog, test_prog,
     else:
         build_strategy.reduce_strategy = fluid.BuildStrategy(
         ).ReduceStrategy.AllReduce
+    build_strategy.fuse_broadcast_op = args.fuse_broadcast_op
 
     avg_loss = train_args[0]
 
@@ -240,7 +241,6 @@ def train_parallel(train_args, test_args, args, train_prog, test_prog,
 
             if args.use_fake_data or args.use_reader_op:
                 try:
-
                     fetch_ret = exe.run(fetch_list)
                 except fluid.core.EOFException as eof:
                     break

cmake/external/xxhash.cmake

+INCLUDE(ExternalProject)
+
+set(XXHASH_SOURCE_DIR ${THIRD_PARTY_PATH}/xxhash)
+set(XXHASH_INSTALL_DIR ${THIRD_PARTY_PATH}/install/xxhash)
+set(XXHASH_INCLUDE_DIR "${XXHASH_INSTALL_DIR}/include")
+
+IF(WITH_STATIC_LIB)
+  SET(BUILD_CMD make lib)
+ELSE()
+  IF(APPLE)
+    SET(BUILD_CMD sed -i \"\" "s/-Wstrict-prototypes -Wundef/-Wstrict-prototypes -Wundef -fPIC/g" ${XXHASH_SOURCE_DIR}/src/extern_xxhash/Makefile && make lib)
+  ELSE(APPLE)
+    SET(BUILD_CMD sed -i "s/-Wstrict-prototypes -Wundef/-Wstrict-prototypes -Wundef -fPIC/g" ${XXHASH_SOURCE_DIR}/src/extern_xxhash/Makefile && make lib)
+  ENDIF(APPLE)
+ENDIF()
+
+ExternalProject_Add(
+    extern_xxhash
+    ${EXTERNAL_PROJECT_LOG_ARGS}
+    GIT_REPOSITORY  "https://github.com/Cyan4973/xxHash"
+    GIT_TAG         "v0.6.5"
+    PREFIX          ${XXHASH_SOURCE_DIR}
+    DOWNLOAD_NAME   "xxhash"
+    UPDATE_COMMAND  ""
+    CONFIGURE_COMMAND ""
+    BUILD_IN_SOURCE 1
+    PATCH_COMMAND
+    BUILD_COMMAND     ${BUILD_CMD}
+    INSTALL_COMMAND   export PREFIX=${XXHASH_INSTALL_DIR}/ && make install
+    TEST_COMMAND      ""
+)
+
+set(XXHASH_LIBRARIES "${XXHASH_INSTALL_DIR}/lib/libxxhash.a")
+INCLUDE_DIRECTORIES(${XXHASH_INCLUDE_DIR})
+
+add_library(xxhash STATIC IMPORTED GLOBAL)
+set_property(TARGET xxhash PROPERTY IMPORTED_LOCATION ${XXHASH_LIBRARIES})
+include_directories(${XXHASH_INCLUDE_DIR})
+add_dependencies(xxhash extern_xxhash)
+
+LIST(APPEND external_project_dependencies xxhash)
+
+IF(WITH_C_API)
+  INSTALL(DIRECTORY ${XXHASH_INCLUDE_DIR} DESTINATION third_party/xxhash)
+  IF(ANDROID)
+    INSTALL(FILES ${XXHASH_LIBRARIES} DESTINATION third_party/xxhash/lib/${ANDROID_ABI})
+  ELSE()
+    INSTALL(FILES ${XXHASH_LIBRARIES} DESTINATION third_party/xxhash/lib)
+  ENDIF()
+ENDIF()

cmake/inference_lib.cmake

@@ -31,7 +31,7 @@ function(copy TARGET)
     foreach(index RANGE ${len})
         list(GET copy_lib_SRCS ${index} src)
         list(GET copy_lib_DSTS ${index} dst)
-        add_custom_command(TARGET ${TARGET} PRE_BUILD 
+        add_custom_command(TARGET ${TARGET} PRE_BUILD
           COMMAND mkdir -p "${dst}"
           COMMAND cp -r "${src}" "${dst}"
           COMMENT "copying ${src} -> ${dst}")
@@ -67,6 +67,13 @@ copy(boost_lib
   DEPS boost
 )
 
+set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/xxhash")
+copy(xxhash_lib
+  SRCS ${XXHASH_INCLUDE_DIR} ${XXHASH_LIBRARIES}
+  DSTS ${dst_dir} ${dst_dir}/lib
+  DEPS xxhash
+)
+
 if(NOT PROTOBUF_FOUND)
     set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/protobuf")
     copy(protobuf_lib
@@ -186,7 +193,7 @@ copy(cmake_cache
   DSTS ${FLUID_INSTALL_DIR})
 
 # This command generates a complete fluid library for both train and inference
-add_custom_target(fluid_lib_dist DEPENDS ${fluid_lib_dist_dep}) 
+add_custom_target(fluid_lib_dist DEPENDS ${fluid_lib_dist_dep})
 
 # Following commands generate a inference-only fluid library
 # third_party, version.txt and CMakeCache.txt are the same position with ${FLUID_INSTALL_DIR}

paddle/CMakeLists.txt

@@ -24,6 +24,7 @@ if(NOT WITH_FLUID_ONLY)
 endif()
 
 add_subdirectory(testing)
+set(PYTHON_TESTS_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/tests CACHE INTERNAL "python tests directory")
 if(NOT MOBILE_INFERENCE AND NOT RPI AND NOT WITH_C_API)
   add_subdirectory(fluid)
 endif()

paddle/fluid/API.spec

@@ -64,7 +64,7 @@ paddle.fluid.layers.chunk_eval ArgSpec(args=['input', 'label', 'chunk_scheme', '
 paddle.fluid.layers.sequence_conv ArgSpec(args=['input', 'num_filters', 'filter_size', 'filter_stride', 'padding', 'bias_attr', 'param_attr', 'act', 'name'], varargs=None, keywords=None, defaults=(3, 1, None, None, None, None, None))
 paddle.fluid.layers.conv2d ArgSpec(args=['input', 'num_filters', 'filter_size', 'stride', 'padding', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(1, 0, 1, None, None, None, True, None, None))
 paddle.fluid.layers.conv3d ArgSpec(args=['input', 'num_filters', 'filter_size', 'stride', 'padding', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(1, 0, 1, None, None, None, True, None, None))
-paddle.fluid.layers.sequence_pool ArgSpec(args=['input', 'pool_type'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.layers.sequence_pool ArgSpec(args=['input', 'pool_type', 'is_test'], varargs=None, keywords=None, defaults=(False,))
 paddle.fluid.layers.sequence_softmax ArgSpec(args=['input', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(False, None))
 paddle.fluid.layers.softmax ArgSpec(args=['input', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(True, None))
 paddle.fluid.layers.pool2d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None))
@@ -86,7 +86,7 @@ paddle.fluid.layers.reduce_prod ArgSpec(args=['input', 'dim', 'keep_dim', 'name'
 paddle.fluid.layers.sequence_first_step ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.sequence_last_step ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.sequence_slice ArgSpec(args=['input', 'offset', 'length', 'name'], varargs=None, keywords=None, defaults=(None,))
-paddle.fluid.layers.dropout ArgSpec(args=['x', 'dropout_prob', 'is_test', 'seed', 'name'], varargs=None, keywords=None, defaults=(False, None, None))
+paddle.fluid.layers.dropout ArgSpec(args=['x', 'dropout_prob', 'is_test', 'seed', 'name', 'dropout_implementation'], varargs=None, keywords=None, defaults=(False, None, None, 'downgrade_in_infer'))
 paddle.fluid.layers.split ArgSpec(args=['input', 'num_or_sections', 'dim', 'name'], varargs=None, keywords=None, defaults=(-1, None))
 paddle.fluid.layers.ctc_greedy_decoder ArgSpec(args=['input', 'blank', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.edit_distance ArgSpec(args=['input', 'label', 'normalized', 'ignored_tokens'], varargs=None, keywords=None, defaults=(True, None))
@@ -107,7 +107,7 @@ paddle.fluid.layers.softmax_with_cross_entropy ArgSpec(args=['logits', 'label',
 paddle.fluid.layers.smooth_l1 ArgSpec(args=['x', 'y', 'inside_weight', 'outside_weight', 'sigma'], varargs=None, keywords=None, defaults=(None, None, None))
 paddle.fluid.layers.one_hot ArgSpec(args=['input', 'depth'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.autoincreased_step_counter ArgSpec(args=['counter_name', 'begin', 'step'], varargs=None, keywords=None, defaults=(None, 1, 1))
-paddle.fluid.layers.reshape ArgSpec(args=['x', 'shape', 'actual_shape', 'act', 'inplace', 'name'], varargs=None, keywords=None, defaults=(None, None, True, None))
+paddle.fluid.layers.reshape ArgSpec(args=['x', 'shape', 'actual_shape', 'act', 'inplace', 'name'], varargs=None, keywords=None, defaults=(None, None, False, None))
 paddle.fluid.layers.squeeze ArgSpec(args=['input', 'axes', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.unsqueeze ArgSpec(args=['input', 'axes', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.lod_reset ArgSpec(args=['x', 'y', 'target_lod'], varargs=None, keywords=None, defaults=(None, None))
@@ -174,7 +174,12 @@ paddle.fluid.layers.mean ArgSpec(args=['x', 'name'], varargs=None, keywords=None
 paddle.fluid.layers.mul ArgSpec(args=['x', 'y', 'x_num_col_dims', 'y_num_col_dims', 'name'], varargs=None, keywords=None, defaults=(1, 1, None))
 paddle.fluid.layers.sigmoid_cross_entropy_with_logits ArgSpec(args=['x', 'label', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.maxout ArgSpec(args=['x', 'groups', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.affine_grid ArgSpec(args=['theta', 'out_shape', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.sequence_reverse ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.affine_channel ArgSpec(args=['x', 'scale', 'bias', 'data_layout', 'name'], varargs=None, keywords=None, defaults=(None, None, 'NCHW', None))
+paddle.fluid.layers.hash ArgSpec(args=['input', 'hash_size', 'num_hash', 'name'], varargs=None, keywords=None, defaults=(1, None))
+paddle.fluid.layers.log_loss ArgSpec(args=['input', 'label', 'epsilon', 'name'], varargs=None, keywords=None, defaults=(0.0001, None))
+paddle.fluid.layers.add_position_encoding ArgSpec(args=['input', 'alpha', 'beta', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.data ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True))
 paddle.fluid.layers.open_files ArgSpec(args=['filenames', 'shapes', 'lod_levels', 'dtypes', 'thread_num', 'buffer_size', 'pass_num', 'is_test'], varargs=None, keywords=None, defaults=(None, None, 1, None))
 paddle.fluid.layers.read_file ArgSpec(args=['reader'], varargs=None, keywords=None, defaults=None)
@@ -353,6 +358,8 @@ paddle.fluid.optimizer.ModelAverage.__init__ ArgSpec(args=['self', 'average_wind
 paddle.fluid.optimizer.ModelAverage.apply ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None)
 paddle.fluid.optimizer.ModelAverage.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))
 paddle.fluid.optimizer.ModelAverage.restore ArgSpec(args=['self', 'executor'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.optimizer.LarsMomentumOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'momentum', 'lars_coeff', 'lars_weight_decay', 'regularization', 'name'], varargs=None, keywords=None, defaults=(0.001, 0.0005, None, None))
+paddle.fluid.optimizer.LarsMomentumOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))
 paddle.fluid.backward.append_backward ArgSpec(args=['loss', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None))
 paddle.fluid.regularizer.L1DecayRegularizer.__init__ ArgSpec(args=['self', 'regularization_coeff'], varargs=None, keywords=None, defaults=(0.0,))
 paddle.fluid.regularizer.L2DecayRegularizer.__init__ ArgSpec(args=['self', 'regularization_coeff'], varargs=None, keywords=None, defaults=(0.0,))

paddle/fluid/CMakeLists.txt

@@ -9,8 +9,6 @@ add_subdirectory(pybind)
 add_subdirectory(recordio)
 endif(NOT WIN32)
 
-if(WITH_INFERENCE)
-  # NOTE: please add subdirectory inference at last.
-  add_subdirectory(inference)
-  add_subdirectory(train)
-endif()
+# NOTE: please add subdirectory inference at last.
+add_subdirectory(inference)
+add_subdirectory(train)

paddle/fluid/framework/attribute.cc

@@ -64,6 +64,13 @@ Attribute GetAttrValue(const proto::OpDesc::Attr& attr_desc) {
     case proto::AttrType::LONG: {
       return attr_desc.l();
     }
+    case proto::AttrType::LONGS: {
+      std::vector<int64_t> val(attr_desc.longs_size());
+      for (int i = 0; i < attr_desc.longs_size(); ++i) {
+        val[i] = attr_desc.longs(i);
+      }
+      return val;
+    }
     default:
       PADDLE_THROW("Unsupport attr type %d", attr_desc.type());
   }

paddle/fluid/framework/attribute.h

@@ -26,6 +26,113 @@ limitations under the License. */
 
 namespace paddle {
 namespace framework {
+
+template <typename T>
+struct ExtractAttribute {
+  explicit ExtractAttribute(const std::string& attr_name)
+      : attr_name_(attr_name) {}
+
+  T* operator()(Attribute& attr) const {
+    T* attr_value = nullptr;
+    try {
+      attr_value = &boost::get<T>(attr);
+    } catch (boost::bad_get& bad_get) {
+      PADDLE_THROW("Cannot get attribute %s by type %s, its type is %s",
+                   attr_name_, paddle::platform::demangle(typeid(T).name()),
+                   paddle::platform::demangle(attr.type().name()));
+    }
+    return attr_value;
+  }
+
+  const std::string& attr_name_;
+};
+
+// special handle bool
+// FIXME(yuyang18): Currently we cast bool into int in python binding. It is
+// hard to change the logic there. In another way, we should correct handle
+// if the user set `some_flag=1`.
+//
+// FIX ME anytime if there is a better solution.
+template <>
+struct ExtractAttribute<bool> {
+  explicit ExtractAttribute(const std::string& attr_name)
+      : attr_name_(attr_name) {}
+
+  bool* operator()(Attribute& attr) const {
+    if (attr.type() == typeid(int)) {  // NOLINT
+      int val = boost::get<int>(attr);
+      attr = static_cast<bool>(val);
+    } else if (attr.type() == typeid(float)) {  // NOLINT
+      float val = boost::get<float>(attr);
+      attr = static_cast<bool>(val);
+    }
+    bool* attr_value = nullptr;
+    try {
+      attr_value = &boost::get<bool>(attr);
+    } catch (boost::bad_get& bad_get) {
+      PADDLE_THROW("Cannot get attribute %s by type bool, its type is %s",
+                   attr_name_, paddle::platform::demangle(attr.type().name()));
+    }
+    return attr_value;
+  }
+
+  const std::string& attr_name_;
+};
+
+template <>
+struct ExtractAttribute<int64_t> {
+  explicit ExtractAttribute(const std::string& attr_name)
+      : attr_name_(attr_name) {}
+
+  int64_t* operator()(Attribute& attr) const {
+    if (attr.type() == typeid(int)) {  // NOLINT
+      int val = boost::get<int>(attr);
+      attr = static_cast<int64_t>(val);
+    } else if (attr.type() == typeid(float)) {  // NOLINT
+      int val = boost::get<float>(attr);
+      attr = static_cast<int64_t>(val);
+    }
+    int64_t* attr_value = nullptr;
+    try {
+      attr_value = &boost::get<int64_t>(attr);
+    } catch (boost::bad_get& bad_get) {
+      PADDLE_THROW("Cannot get attribute %s by type int64_t, its type is %s",
+                   attr_name_, paddle::platform::demangle(attr.type().name()));
+    }
+    return attr_value;
+  }
+
+  const std::string& attr_name_;
+};
+
+template <>
+struct ExtractAttribute<std::vector<int64_t>> {
+  explicit ExtractAttribute(const std::string& attr_name)
+      : attr_name_(attr_name) {}
+
+  std::vector<int64_t>* operator()(Attribute& attr) const {
+    if (attr.type() == typeid(std::vector<int>)) {  // NOLINT
+      std::vector<int> val = boost::get<std::vector<int>>(attr);
+      std::vector<int64_t> vec(val.begin(), val.end());
+      attr = vec;
+    } else if (attr.type() == typeid(std::vector<float>)) {  // NOLINT
+      std::vector<float> val = boost::get<std::vector<float>>(attr);
+      std::vector<int64_t> vec(val.begin(), val.end());
+      attr = vec;
+    }
+    std::vector<int64_t>* attr_value = nullptr;
+    try {
+      attr_value = &boost::get<std::vector<int64_t>>(attr);
+    } catch (boost::bad_get& bad_get) {
+      PADDLE_THROW("Cannot get attribute %s by type int64_t, its type is %s",
+                   attr_name_, paddle::platform::demangle(attr.type().name()));
+    }
+    return attr_value;
+  }
+
+  const std::string& attr_name_;
+};
+
 template <typename T>
 inline proto::AttrType AttrTypeID() {
   Attribute tmp = T();
@@ -42,7 +149,11 @@ class AttrReader {
   inline const T& Get(const std::string& name) const {
     PADDLE_ENFORCE(attrs_.count(name) != 0, "%s should be in AttributeMap",
                    name);
-    return boost::get<T>(attrs_.at(name));
+
+    Attribute& attr = const_cast<Attribute&>(attrs_.at(name));
+    ExtractAttribute<T> extract_attr(name);
+    T* attr_value = extract_attr(attr);
+    return *attr_value;
   }
 
  private:
@@ -82,7 +193,7 @@ class DefaultValueSetter {
  public:
   explicit DefaultValueSetter(T default_value)
       : default_value_(default_value) {}
-  void operator()(T& value) const { value = default_value_; }
+  void operator()(T& value) const { value = default_value_; }  // NOLINT
 
  private:
   T default_value_;
@@ -117,84 +228,6 @@ class EnumInContainer {
   std::unordered_set<T> container_;
 };
 
-template <typename T>
-struct ExtractAttribute {
-  explicit ExtractAttribute(const std::string& attr_name)
-      : attr_name_(attr_name) {}
-
-  T* operator()(Attribute& attr) const {
-    T* attr_value = nullptr;
-    try {
-      attr_value = &boost::get<T>(attr);
-    } catch (boost::bad_get& bad_get) {
-      PADDLE_THROW("Cannot get attribute %s by type %s, its type is %s",
-                   attr_name_, paddle::platform::demangle(typeid(T).name()),
-                   paddle::platform::demangle(attr.type().name()));
-    }
-    return attr_value;
-  }
-
-  const std::string& attr_name_;
-};
-
-// special handle bool
-// FIXME(yuyang18): Currently we cast bool into int in python binding. It is
-// hard to change the logic there. In another way, we should correct handle
-// if the user set `some_flag=1`.
-//
-// FIX ME anytime if there is a better solution.
-template <>
-struct ExtractAttribute<bool> {
-  explicit ExtractAttribute(const std::string& attr_name)
-      : attr_name_(attr_name) {}
-
-  bool* operator()(Attribute& attr) const {
-    if (attr.type() == typeid(int)) {  // NOLINT
-      int val = boost::get<int>(attr);
-      attr = static_cast<bool>(val);
-    } else if (attr.type() == typeid(float)) {  // NOLINT
-      float val = boost::get<float>(attr);
-      attr = static_cast<bool>(val);
-    }
-    bool* attr_value = nullptr;
-    try {
-      attr_value = &boost::get<bool>(attr);
-    } catch (boost::bad_get& bad_get) {
-      PADDLE_THROW("Cannot get attribute %s by type bool, its type is %s",
-                   attr_name_, paddle::platform::demangle(attr.type().name()));
-    }
-    return attr_value;
-  }
-
-  const std::string& attr_name_;
-};
-
-template <>
-struct ExtractAttribute<int64_t> {
-  explicit ExtractAttribute(const std::string& attr_name)
-      : attr_name_(attr_name) {}
-
-  int64_t* operator()(Attribute& attr) const {
-    if (attr.type() == typeid(int)) {  // NOLINT
-      int val = boost::get<int>(attr);
-      attr = static_cast<int64_t>(val);
-    } else if (attr.type() == typeid(float)) {  // NOLINT
-      int val = boost::get<float>(attr);
-      attr = static_cast<int64_t>(val);
-    }
-    int64_t* attr_value = nullptr;
-    try {
-      attr_value = &boost::get<int64_t>(attr);
-    } catch (boost::bad_get& bad_get) {
-      PADDLE_THROW("Cannot get attribute %s by type int64_t, its type is %s",
-                   attr_name_, paddle::platform::demangle(attr.type().name()));
-    }
-    return attr_value;
-  }
-
-  const std::string& attr_name_;
-};
-
 // check whether a certain attribute fit its limits
 // an attribute can have more than one limits
 template <typename T>
@@ -235,7 +268,7 @@ class TypedAttrChecker {
     return *this;
   }
 
-  void operator()(AttributeMap& attr_map) const {
+  void operator()(AttributeMap& attr_map) const {  // NOLINT
     if (!attr_map.count(attr_name_)) {
       // user do not set this attr
       PADDLE_ENFORCE(!default_value_setter_.empty(),
@@ -271,7 +304,7 @@ class OpAttrChecker {
     return *(checker.target<TypedAttrChecker<T>>());
   }
 
-  void Check(AttributeMap& attr_map) const {
+  void Check(AttributeMap& attr_map) const {  // NOLINT
     for (const auto& checker : attr_checkers_) {
       checker(attr_map);
     }

paddle/fluid/framework/details/CMakeLists.txt

@@ -16,12 +16,14 @@ if(WITH_GPU)
             dynload_cuda variable_visitor)
     nv_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope ddim dynload_cuda)
     nv_library(broadcast_op_handle SRCS broadcast_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor dynload_cuda)
+    nv_library(fused_broadcast_op_handle SRCS fused_broadcast_op_handle.cc DEPS broadcast_op_handle)
 
 else()
     cc_library(all_reduce_op_handle SRCS all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory
              variable_visitor)
     cc_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope ddim)
     cc_library(broadcast_op_handle SRCS broadcast_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor)
+    cc_library(fused_broadcast_op_handle SRCS fused_broadcast_op_handle.cc DEPS broadcast_op_handle)
 endif()
 
 cc_library(data_balance_op_handle SRCS data_balance_op_handle.cc DEPS op_handle_base scope lod_tensor)
@@ -34,7 +36,7 @@ if(WITH_GPU)
 endif()
 
 cc_library(multi_devices_graph_pass SRCS multi_devices_graph_pass.cc DEPS multi_devices_helper computation_op_handle
-        scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle)
+        scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle fused_broadcast_op_handle)
 
 if(WITH_GPU)
   cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS graph framework_proto reference_count_pass)
@@ -54,8 +56,9 @@ cc_library(scope_buffered_ssa_graph_executor SRCS scope_buffered_ssa_graph_execu
 #        device_context reduce_op_handle )
 cc_library(fast_threaded_ssa_graph_executor SRCS fast_threaded_ssa_graph_executor.cc
         DEPS fetch_op_handle ssa_graph_executor scope simple_threadpool device_context)
+cc_test(fused_broadcast_op_test SRCS fused_broadcast_op_handle_test.cc DEPS fused_broadcast_op_handle)
 
 cc_library(build_strategy SRCS build_strategy.cc DEPS
         graph_viz_pass multi_devices_graph_pass
         multi_devices_graph_print_pass multi_devices_graph_check_pass
-        fuse_elewise_add_act_pass)
+        fuse_elewise_add_act_pass multi_batch_merge_pass)

paddle/fluid/framework/details/all_reduce_op_handle.cc

@@ -34,7 +34,7 @@ AllReduceOpHandle::AllReduceOpHandle(ir::Node *node,
       nccl_ctxs_(ctxs) {
   if (nccl_ctxs_) {
     for (auto &p : places_) {
-      this->dev_ctxes_[p] = nccl_ctxs_->DevCtx(p);
+      this->SetDeviceContext(p, nccl_ctxs_->DevCtx(p));
     }
   }
 }
@@ -46,7 +46,7 @@ AllReduceOpHandle::AllReduceOpHandle(ir::Node *node,
 #endif
 
 void AllReduceOpHandle::RunImpl() {
-  platform::RecordEvent record_event(Name(), dev_ctxes_.begin()->second);
+  platform::RecordEvent record_event(Name(), dev_ctxes_.cbegin()->second);
 
   if (NoDummyInputSize() == 1) {
     return;  // No need to all reduce when GPU count = 1;
@@ -127,7 +127,7 @@ void AllReduceOpHandle::RunImpl() {
             *local_scopes_[i]->FindVar(kLocalExecScopeName)->Get<Scope *>();
         auto &p = places_[i];
         auto *var = scope.FindVar(out_var_handles[i]->name_);
-        auto *dev_ctx = dev_ctxes_[p];
+        auto *dev_ctx = dev_ctxes_.at(p);
 
         RunAndRecordEvent(p, [&trg, var, dev_ctx, p] {
           auto &tensor_gpu = *var->GetMutable<framework::LoDTensor>();

paddle/fluid/framework/details/broadcast_op_handle.cc

@@ -48,16 +48,27 @@ void BroadcastOpHandle::RunImpl() {
     var_scopes.emplace_back(s->FindVar(kLocalExecScopeName)->Get<Scope *>());
   }
 
+  BroadcastOneVar(*in_var_handle, out_var_handles, var_scopes);
+}
+
+void BroadcastOpHandle::BroadcastOneVar(
+    const VarHandle &in_var_handle,
+    const std::vector<VarHandle *> &out_var_handles,
+    const std::vector<const Scope *> &var_scopes) {
   auto *in_var =
-      var_scopes.at(in_var_handle->scope_idx_)->FindVar(in_var_handle->name_);
+      var_scopes.at(in_var_handle.scope_idx_)->FindVar(in_var_handle.name_);
   PADDLE_ENFORCE_NOT_NULL(in_var);
   Tensor &in_tensor = VariableVisitor::GetMutableTensor(in_var);
+  if (UNLIKELY(!in_tensor.IsInitialized())) {
+    VLOG(3) << "in var " << in_var_handle.name_ << "not inited, return!";
+    return;
+  }
 
-  InitOutputValue(*in_var_handle, out_var_handles);
+  InitOutputValue(in_var_handle, out_var_handles);
 
   if (platform::is_cpu_place(in_tensor.place())) {
     for (auto *out_var_handle : out_var_handles) {
-      if (out_var_handle->IsTheSameVar(*in_var_handle)) {
+      if (out_var_handle->IsTheSameVar(in_var_handle)) {
         continue;
       }
       auto &out_p = out_var_handle->place_;
@@ -114,12 +125,12 @@ void BroadcastOpHandle::RunImpl() {
         }
       }
 
-      if (!out_handle->IsTheSameVar(*in_var_handle)) {
-        auto out_var = var_scopes.at(in_var_handle->scope_idx_)
+      if (!out_handle->IsTheSameVar(in_var_handle)) {
+        auto out_var = var_scopes.at(in_var_handle.scope_idx_)
                            ->FindVar(out_var_handles[0]->name_);
         paddle::framework::TensorCopy(
-            in_tensor, in_var_handle->place_,
-            *(dev_ctxes_.at(in_var_handle->place_)),
+            in_tensor, in_var_handle.place_,
+            *(dev_ctxes_.at(in_var_handle.place_)),
             &VariableVisitor::GetMutableTensor(out_var));
       }
     });

paddle/fluid/framework/details/broadcast_op_handle.h

@@ -44,7 +44,8 @@ struct BroadcastOpHandle : public OpHandleBase {
         nccl_ctxs_(nccl_ctxs) {
     if (nccl_ctxs_) {
       for (auto &p_ctx : nccl_ctxs_->contexts_) {
-        dev_ctxes_[platform::CUDAPlace(p_ctx.first)] = p_ctx.second.ctx_.get();
+        this->SetDeviceContext(platform::CUDAPlace(p_ctx.first),
+                               p_ctx.second.ctx_.get());
       }
     }
   }
@@ -61,7 +62,10 @@ struct BroadcastOpHandle : public OpHandleBase {
  protected:
   void RunImpl() override;
 
- private:
+  void BroadcastOneVar(const VarHandle &in_var_handle,
+                       const std::vector<VarHandle *> &out_var_handles,
+                       const std::vector<const Scope *> &var_scopes);
+
   std::vector<Scope *> local_scopes_;
   std::vector<platform::Place> places_;
 #ifdef PADDLE_WITH_CUDA

paddle/fluid/framework/details/broadcast_op_handle_test.cc

@@ -12,232 +12,12 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/fluid/framework/details/broadcast_op_handle.h"
-#include "gtest/gtest.h"
-
-#include "paddle/fluid/platform/device_context.h"
+#include "paddle/fluid/framework/details/broadcast_op_handle_test.h"
 
 namespace paddle {
 namespace framework {
 namespace details {
 
-namespace f = paddle::framework;
-namespace p = paddle::platform;
-
-// test data amount
-const f::DDim kDims = {20, 20};
-
-struct TestBroadcastOpHandle {
-  std::vector<std::unique_ptr<p::DeviceContext>> ctxs_;
-  std::vector<Scope*> local_scopes_;
-  std::vector<Scope*> param_scopes_;
-  Scope g_scope_;
-  std::unique_ptr<OpHandleBase> op_handle_;
-  std::vector<std::unique_ptr<VarHandleBase>> vars_;
-  std::vector<p::Place> gpu_list_;
-  bool use_gpu_;
-#ifdef PADDLE_WITH_CUDA
-  std::unique_ptr<platform::NCCLContextMap> nccl_ctxs_;
-#endif
-
-  void WaitAll() {
-    for (size_t j = 0; j < ctxs_.size(); ++j) {
-      ctxs_[j]->Wait();
-    }
-#ifdef PADDLE_WITH_CUDA
-    if (nccl_ctxs_) {
-      nccl_ctxs_->WaitAll();
-    }
-#endif
-  }
-
-  void InitCtxOnGpu(bool use_gpu) {
-    use_gpu_ = use_gpu;
-    if (use_gpu_) {
-#ifdef PADDLE_WITH_CUDA
-      int count = p::GetCUDADeviceCount();
-      if (count <= 1) {
-        LOG(WARNING) << "Cannot test multi-gpu Broadcast, because the CUDA "
-                        "device count is "
-                     << count;
-        exit(0);
-      }
-      for (int i = 0; i < count; ++i) {
-        auto p = p::CUDAPlace(i);
-        gpu_list_.push_back(p);
-        ctxs_.emplace_back(new p::CUDADeviceContext(p));
-      }
-      nccl_ctxs_.reset(new platform::NCCLContextMap(gpu_list_));
-#else
-      PADDLE_THROW("CUDA is not support.");
-#endif
-    } else {
-      int count = 8;
-      for (int i = 0; i < count; ++i) {
-        auto p = p::CPUPlace();
-        gpu_list_.push_back(p);
-        ctxs_.emplace_back(new p::CPUDeviceContext(p));
-      }
-#ifdef PADDLE_WITH_CUDA
-      nccl_ctxs_.reset(nullptr);
-#endif
-    }
-  }
-
-  void InitBroadcastOp(size_t input_scope_idx) {
-    for (size_t j = 0; j < gpu_list_.size(); ++j) {
-      local_scopes_.push_back(&(g_scope_.NewScope()));
-      Scope& local_scope = local_scopes_.back()->NewScope();
-      *local_scopes_.back()
-           ->Var(details::kLocalExecScopeName)
-           ->GetMutable<Scope*>() = &local_scope;
-      local_scope.Var("out");
-      param_scopes_.emplace_back(&local_scope);
-    }
-    param_scopes_[input_scope_idx]->Var("input");
-
-    std::unique_ptr<ir::Node> n =
-        ir::CreateNodeForTest("node0", ir::Node::Type::kOperation);
-    if (use_gpu_) {
-#ifdef PADDLE_WITH_CUDA
-      op_handle_.reset(new BroadcastOpHandle(n.get(), local_scopes_, gpu_list_,
-                                             nccl_ctxs_.get()));
-#else
-      PADDLE_THROW("CUDA is not support.");
-#endif
-    } else {
-#ifdef PADDLE_WITH_CUDA
-      op_handle_.reset(new BroadcastOpHandle(n.get(), local_scopes_, gpu_list_,
-                                             nccl_ctxs_.get()));
-#else
-      op_handle_.reset(
-          new BroadcastOpHandle(n.get(), local_scopes_, gpu_list_));
-#endif
-    }
-
-    std::unique_ptr<ir::Node> v =
-        ir::CreateNodeForTest("node1", ir::Node::Type::kVariable);
-    auto* in_var_handle = new VarHandle(v.get(), 1, input_scope_idx, "input",
-                                        gpu_list_[input_scope_idx]);
-    vars_.emplace_back(in_var_handle);
-    op_handle_->AddInput(in_var_handle);
-
-    // add dummy var
-
-    std::unique_ptr<ir::Node> v2 =
-        ir::CreateNodeForTest("node2", ir::Node::Type::kVariable);
-    vars_.emplace_back(new DummyVarHandle(v2.get()));
-    DummyVarHandle* dummy_var_handle =
-        static_cast<DummyVarHandle*>(vars_.back().get());
-    dummy_var_handle->ClearGeneratedOp();
-    op_handle_->AddInput(dummy_var_handle);
-
-    for (size_t j = 0; j < gpu_list_.size(); ++j) {
-      if (!use_gpu_) {
-        op_handle_->SetDeviceContext(gpu_list_[j], ctxs_[j].get());
-      }
-      std::unique_ptr<ir::Node> v3 =
-          ir::CreateNodeForTest("node3", ir::Node::Type::kVariable);
-      VarHandle* out_var_handle =
-          new VarHandle(v3.get(), 2, j, "out", gpu_list_[j]);
-      vars_.emplace_back(out_var_handle);
-      op_handle_->AddOutput(out_var_handle);
-    }
-
-    // add dummy var
-    std::unique_ptr<ir::Node> v4 =
-        ir::CreateNodeForTest("node4", ir::Node::Type::kVariable);
-    vars_.emplace_back(new DummyVarHandle(v4.get()));
-    DummyVarHandle* out_dummy_var_handle =
-        static_cast<DummyVarHandle*>(vars_.back().get());
-    out_dummy_var_handle->ClearGeneratedOp();
-    op_handle_->AddOutput(out_dummy_var_handle);
-  }
-
-  void TestBroadcastLodTensor(size_t input_scope_idx) {
-    auto in_var = param_scopes_[input_scope_idx]->FindVar("input");
-    PADDLE_ENFORCE_NOT_NULL(in_var);
-    auto in_lod_tensor = in_var->GetMutable<f::LoDTensor>();
-    in_lod_tensor->mutable_data<float>(kDims, gpu_list_[input_scope_idx]);
-
-    std::vector<float> send_vector(static_cast<size_t>(f::product(kDims)));
-    for (size_t k = 0; k < send_vector.size(); ++k) {
-      send_vector[k] = k;
-    }
-    f::LoD lod{{0, 10, 20}};
-    paddle::framework::TensorFromVector<float>(
-        send_vector, *(ctxs_[input_scope_idx]), in_lod_tensor);
-    in_lod_tensor->set_lod(lod);
-    in_lod_tensor->Resize(kDims);
-
-    op_handle_->Run(false);
-
-    WaitAll();
-
-    p::CPUPlace cpu_place;
-    for (size_t j = 0; j < gpu_list_.size(); ++j) {
-      auto out_var = param_scopes_[j]->FindVar("out");
-      PADDLE_ENFORCE_NOT_NULL(out_var);
-      auto out_tensor = out_var->Get<f::LoDTensor>();
-      PADDLE_ENFORCE_EQ(out_tensor.lod(), lod, "lod is not equal.");
-
-      f::Tensor result_tensor;
-      f::TensorCopySync(out_tensor, cpu_place, &result_tensor);
-      float* ct = result_tensor.mutable_data<float>(cpu_place);
-
-      for (int64_t i = 0; i < f::product(kDims); ++i) {
-        ASSERT_NEAR(ct[i], send_vector[i], 1e-5);
-      }
-    }
-  }
-
-  void TestBroadcastSelectedRows(size_t input_scope_idx) {
-    auto in_var = param_scopes_[input_scope_idx]->FindVar("input");
-    PADDLE_ENFORCE_NOT_NULL(in_var);
-    auto in_selected_rows = in_var->GetMutable<f::SelectedRows>();
-    auto value = in_selected_rows->mutable_value();
-    value->mutable_data<float>(kDims, gpu_list_[input_scope_idx]);
-    int height = static_cast<int>(kDims[0]) * 2;
-    std::vector<int64_t> rows{0, 1, 2, 3, 3, 0, 14, 7, 3, 1,
-                              2, 4, 6, 3, 1, 1, 1,  1, 3, 7};
-    in_selected_rows->set_height(height);
-    in_selected_rows->set_rows(rows);
-
-    std::vector<float> send_vector(static_cast<size_t>(f::product(kDims)));
-    for (size_t k = 0; k < send_vector.size(); ++k) {
-      send_vector[k] = k;
-    }
-    paddle::framework::TensorFromVector<float>(
-        send_vector, *(ctxs_[input_scope_idx]), value);
-
-    op_handle_->Run(false);
-
-    WaitAll();
-
-    p::CPUPlace cpu_place;
-    for (size_t j = 0; j < gpu_list_.size(); ++j) {
-      auto out_var = param_scopes_[j]->FindVar("out");
-      PADDLE_ENFORCE_NOT_NULL(out_var);
-      auto& out_select_rows = out_var->Get<f::SelectedRows>();
-      auto rt = out_select_rows.value();
-
-      PADDLE_ENFORCE_EQ(out_select_rows.height(), height,
-                        "height is not equal.");
-      for (size_t k = 0; k < out_select_rows.rows().size(); ++k) {
-        PADDLE_ENFORCE_EQ(out_select_rows.rows()[k], rows[k]);
-      }
-
-      f::Tensor result_tensor;
-      f::TensorCopySync(rt, cpu_place, &result_tensor);
-      float* ct = result_tensor.data<float>();
-
-      for (int64_t i = 0; i < f::product(kDims); ++i) {
-        ASSERT_NEAR(ct[i], send_vector[i], 1e-5);
-      }
-    }
-  }
-};
-
 TEST(BroadcastTester, TestCPUBroadcastTestLodTensor) {
   TestBroadcastOpHandle test_op;
   size_t input_scope_idx = 0;

paddle/fluid/framework/details/broadcast_op_handle_test.h

+//   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include <string>
+#include <vector>
+
+#include "gtest/gtest.h"
+#include "paddle/fluid/framework/details/broadcast_op_handle.h"
+
+#include "paddle/fluid/platform/device_context.h"
+
+namespace paddle {
+namespace framework {
+namespace details {
+
+namespace f = paddle::framework;
+namespace p = paddle::platform;
+
+// test data amount
+const f::DDim kDims = {20, 20};
+
+struct TestBroadcastOpHandle {
+  std::vector<std::unique_ptr<p::DeviceContext>> ctxs_;
+  std::vector<Scope*> local_scopes_;
+  std::vector<Scope*> param_scopes_;
+  Scope g_scope_;
+  std::unique_ptr<OpHandleBase> op_handle_;
+  std::vector<std::unique_ptr<VarHandleBase>> vars_;
+  std::vector<p::Place> place_list_;
+  bool use_gpu_;
+#ifdef PADDLE_WITH_CUDA
+  std::unique_ptr<platform::NCCLContextMap> nccl_ctxs_;
+#endif
+
+  void WaitAll() {
+    for (size_t j = 0; j < ctxs_.size(); ++j) {
+      ctxs_[j]->Wait();
+    }
+#ifdef PADDLE_WITH_CUDA
+    if (nccl_ctxs_) {
+      nccl_ctxs_->WaitAll();
+    }
+#endif
+  }
+
+  void InitCtxOnGpu(bool use_gpu) {
+    use_gpu_ = use_gpu;
+    if (use_gpu_) {
+#ifdef PADDLE_WITH_CUDA
+      int count = p::GetCUDADeviceCount();
+      if (count <= 1) {
+        LOG(WARNING) << "Cannot test multi-gpu Broadcast, because the CUDA "
+                        "device count is "
+                     << count;
+        exit(0);
+      }
+      for (int i = 0; i < count; ++i) {
+        auto p = p::CUDAPlace(i);
+        place_list_.push_back(p);
+        ctxs_.emplace_back(new p::CUDADeviceContext(p));
+      }
+      nccl_ctxs_.reset(new platform::NCCLContextMap(place_list_));
+#else
+      PADDLE_THROW("CUDA is not support.");
+#endif
+    } else {
+      int count = 8;
+      for (int i = 0; i < count; ++i) {
+        auto p = p::CPUPlace();
+        place_list_.push_back(p);
+        ctxs_.emplace_back(new p::CPUDeviceContext(p));
+      }
+#ifdef PADDLE_WITH_CUDA
+      nccl_ctxs_.reset(nullptr);
+#endif
+    }
+  }
+
+  void InitBroadcastOp(size_t input_scope_idx) {
+    for (size_t j = 0; j < place_list_.size(); ++j) {
+      local_scopes_.push_back(&(g_scope_.NewScope()));
+      Scope& local_scope = local_scopes_.back()->NewScope();
+      *local_scopes_.back()
+           ->Var(details::kLocalExecScopeName)
+           ->GetMutable<Scope*>() = &local_scope;
+      local_scope.Var("out");
+      param_scopes_.emplace_back(&local_scope);
+    }
+    param_scopes_[input_scope_idx]->Var("input");
+
+    std::unique_ptr<ir::Node> n =
+        ir::CreateNodeForTest("node0", ir::Node::Type::kOperation);
+    if (use_gpu_) {
+#ifdef PADDLE_WITH_CUDA
+      op_handle_.reset(new BroadcastOpHandle(n.get(), local_scopes_,
+                                             place_list_, nccl_ctxs_.get()));
+#else
+      PADDLE_THROW("CUDA is not support.");
+#endif
+    } else {
+#ifdef PADDLE_WITH_CUDA
+      op_handle_.reset(new BroadcastOpHandle(n.get(), local_scopes_,
+                                             place_list_, nccl_ctxs_.get()));
+#else
+      op_handle_.reset(
+          new BroadcastOpHandle(n.get(), local_scopes_, place_list_));
+#endif
+    }
+
+    std::unique_ptr<ir::Node> v =
+        ir::CreateNodeForTest("node1", ir::Node::Type::kVariable);
+    auto* in_var_handle = new VarHandle(v.get(), 1, input_scope_idx, "input",
+                                        place_list_[input_scope_idx]);
+    vars_.emplace_back(in_var_handle);
+    op_handle_->AddInput(in_var_handle);
+
+    // add dummy var
+
+    std::unique_ptr<ir::Node> v2 =
+        ir::CreateNodeForTest("node2", ir::Node::Type::kVariable);
+    vars_.emplace_back(new DummyVarHandle(v2.get()));
+    DummyVarHandle* dummy_var_handle =
+        static_cast<DummyVarHandle*>(vars_.back().get());
+    dummy_var_handle->ClearGeneratedOp();
+    op_handle_->AddInput(dummy_var_handle);
+
+    for (size_t j = 0; j < place_list_.size(); ++j) {
+      if (!use_gpu_) {
+        op_handle_->SetDeviceContext(place_list_[j], ctxs_[j].get());
+      }
+      std::unique_ptr<ir::Node> v3 =
+          ir::CreateNodeForTest("node3", ir::Node::Type::kVariable);
+      VarHandle* out_var_handle =
+          new VarHandle(v3.get(), 2, j, "out", place_list_[j]);
+      vars_.emplace_back(out_var_handle);
+      op_handle_->AddOutput(out_var_handle);
+    }
+
+    // add dummy var
+    std::unique_ptr<ir::Node> v4 =
+        ir::CreateNodeForTest("node4", ir::Node::Type::kVariable);
+    vars_.emplace_back(new DummyVarHandle(v4.get()));
+    DummyVarHandle* out_dummy_var_handle =
+        static_cast<DummyVarHandle*>(vars_.back().get());
+    out_dummy_var_handle->ClearGeneratedOp();
+    op_handle_->AddOutput(out_dummy_var_handle);
+  }
+
+  std::vector<float> InitLoDTensor(const std::string& varname,
+                                   size_t input_scope_idx, const f::LoD& lod,
+                                   float val_scalar = 0.0) {
+    auto var = param_scopes_[input_scope_idx]->FindVar(varname);
+
+    PADDLE_ENFORCE_NOT_NULL(var);
+    auto lod_tensor = var->GetMutable<f::LoDTensor>();
+    std::vector<float> send_vector(static_cast<size_t>(f::product(kDims)));
+    for (size_t k = 0; k < send_vector.size(); ++k) {
+      send_vector[k] = k + val_scalar;
+    }
+    paddle::framework::TensorFromVector<float>(
+        send_vector, *(ctxs_[input_scope_idx]), lod_tensor);
+    lod_tensor->set_lod(lod);
+    lod_tensor->Resize(kDims);
+    return send_vector;
+  }
+
+  std::vector<float> InitSelectedRows(const std::string& varname,
+                                      size_t input_scope_idx,
+                                      const std::vector<int64_t>& rows,
+                                      int height, float value_scalar = 0.0) {
+    std::vector<float> send_vector(static_cast<size_t>(f::product(kDims)));
+    for (size_t k = 0; k < send_vector.size(); ++k) {
+      send_vector[k] = k + value_scalar;
+    }
+
+    auto var = param_scopes_[input_scope_idx]->FindVar(varname);
+    PADDLE_ENFORCE_NOT_NULL(var);
+    auto selected_rows = var->GetMutable<f::SelectedRows>();
+    auto value = selected_rows->mutable_value();
+    value->mutable_data<float>(kDims, place_list_[input_scope_idx]);
+    selected_rows->set_height(height);
+    selected_rows->set_rows(rows);
+
+    paddle::framework::TensorFromVector<float>(
+        send_vector, *(ctxs_[input_scope_idx]), value);
+
+    return send_vector;
+  }
+
+  void SelectedRowsEqual(const std::string& varname, int input_scope_idx,
+                         const std::vector<float>& send_vector,
+                         const std::vector<int64_t>& rows, int height) {
+    auto var = param_scopes_[input_scope_idx]->FindVar(varname);
+    PADDLE_ENFORCE_NOT_NULL(var);
+    auto& selected_rows = var->Get<f::SelectedRows>();
+    auto rt = selected_rows.value();
+    PADDLE_ENFORCE_EQ(selected_rows.height(), height, "height is not equal.");
+
+    for (size_t k = 0; k < selected_rows.rows().size(); ++k) {
+      PADDLE_ENFORCE_EQ(selected_rows.rows()[k], rows[k]);
+    }
+
+    p::CPUPlace cpu_place;
+    f::Tensor result_tensor;
+    f::TensorCopySync(rt, cpu_place, &result_tensor);
+    float* ct = result_tensor.data<float>();
+
+    for (int64_t i = 0; i < f::product(kDims); ++i) {
+      ASSERT_NEAR(ct[i], send_vector[i], 1e-5);
+    }
+  }
+
+  void LoDTensorEqual(const std::string& varname,
+                      const std::vector<float>& send_vec, const f::LoD& lod,
+                      framework::Scope* scope) {
+    p::CPUPlace cpu_place;
+    auto var = scope->FindVar(varname);
+    PADDLE_ENFORCE_NOT_NULL(var);
+    auto tensor = var->Get<f::LoDTensor>();
+    PADDLE_ENFORCE_EQ(tensor.lod(), lod, "lod is not equal.");
+    f::Tensor result_tensor;
+    f::TensorCopySync(tensor, cpu_place, &result_tensor);
+    float* ct = result_tensor.mutable_data<float>(cpu_place);
+    for (int64_t k = 0; k < f::product(kDims); ++k) {
+      ASSERT_NEAR(ct[k], send_vec[k], 1e-5);
+    }
+  }
+
+  void TestBroadcastLodTensor(size_t input_scope_idx) {
+    f::LoD lod{{0, 10, 20}};
+    auto send_vector = InitLoDTensor("input", input_scope_idx, lod);
+
+    op_handle_->Run(false);
+
+    WaitAll();
+    for (size_t j = 0; j < place_list_.size(); ++j) {
+      LoDTensorEqual("out", send_vector, lod, param_scopes_[j]);
+    }
+  }
+
+  void TestBroadcastSelectedRows(size_t input_scope_idx) {
+    std::vector<int64_t> rows{0, 1, 2, 3, 3, 0, 14, 7, 3, 1,
+                              2, 4, 6, 3, 1, 1, 1,  1, 3, 7};
+    int height = static_cast<int>(kDims[0] * 2);
+    auto send_vector = InitSelectedRows("input", input_scope_idx, rows, height);
+
+    op_handle_->Run(false);
+
+    WaitAll();
+    for (size_t j = 0; j < place_list_.size(); ++j) {
+      SelectedRowsEqual("out", input_scope_idx, send_vector, rows, height);
+    }
+  }
+};
+
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/details/build_strategy.cc

@@ -121,6 +121,7 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
 
 USE_PASS(fuse_elewise_add_act_pass);
 USE_PASS(graph_viz_pass);
+USE_PASS(multi_batch_merge_pass);
 USE_PASS(multi_devices_pass);
 USE_PASS(multi_devices_check_pass);
 USE_PASS(multi_devices_print_pass);

paddle/fluid/framework/details/build_strategy.h

@@ -69,6 +69,8 @@ struct BuildStrategy {
 
   bool enable_data_balance_{false};
 
+  bool fuse_broadcast_op_{false};
+
   // User normally doesn't need to call this API.
   // The PassBuilder allows for more customized insert, remove of passes
   // from python side.

paddle/fluid/framework/details/computation_op_handle.cc

@@ -37,7 +37,7 @@ void ComputationOpHandle::RunImpl() {
 bool ComputationOpHandle::NeedWait(VarHandleBase *in_var) {
   bool need_wait =
       in_var && in_var->GeneratedOp() &&
-      in_var->GeneratedOp()->DeviceContext(place_) != dev_ctxes_[place_];
+      in_var->GeneratedOp()->DeviceContext(place_) != dev_ctxes_.at(place_);
   return need_wait;
 }
 

paddle/fluid/framework/details/data_balance_op_handle.cc

@@ -28,7 +28,7 @@ DataBalanceOpHandle::DataBalanceOpHandle(
     : OpHandleBase(node), local_scopes_(local_scopes), places_(places) {
   if (ctxs) {
     for (auto &p : places_) {
-      this->dev_ctxes_[p] = ctxs->DevCtx(p);
+      this->SetDeviceContext(p, ctxs->DevCtx(p));
     }
   }
 }
@@ -89,8 +89,8 @@ void DataBalanceOpHandle::RunImpl() {
   PADDLE_ENFORCE_GT(places_.size(), 1,
                     "Data balance can only be enabled when the number of "
                     "places to run larger than 1.");
-  auto in_var_handles = DynamicCast<VarHandle>(inputs_);
-  auto out_var_handles = DynamicCast<VarHandle>(outputs_);
+  auto in_var_handles = DynamicCast<VarHandle>(this->Inputs());
+  auto out_var_handles = DynamicCast<VarHandle>(this->Outputs());
   PADDLE_ENFORCE(in_var_handles.size() % places_.size() == 0);
   PADDLE_ENFORCE_EQ(
       in_var_handles.size(), out_var_handles.size(),

paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc

@@ -92,13 +92,13 @@ FeedFetchList FastThreadedSSAGraphExecutor::Run(
 
   size_t num_complete = 0;
   remaining_ = 0;
-  BlockingQueue<size_t> complete_q;
+  auto complete_q = std::make_shared<BlockingQueue<size_t>>();
   for (auto op : bootstrap_ops_) {
-    RunOpAsync(op_deps.get(), op, &complete_q);
+    RunOpAsync(op_deps.get(), op, complete_q);
   }
 
   while (num_complete != op_deps->size()) {
-    size_t num_comp = complete_q.Pop();
+    size_t num_comp = complete_q->Pop();
     if (num_comp == -1UL) {
       int remaining = 0;
       while (true) {
@@ -107,7 +107,7 @@ FeedFetchList FastThreadedSSAGraphExecutor::Run(
           break;
         }
         for (int i = 0; i < remaining; ++i) {
-          complete_q.Pop();
+          complete_q->Pop();
         }
       }
       exception_.ReThrow();
@@ -120,7 +120,8 @@ FeedFetchList FastThreadedSSAGraphExecutor::Run(
 }
 void FastThreadedSSAGraphExecutor::RunOpAsync(
     std::unordered_map<OpHandleBase *, std::atomic<int>> *op_deps,
-    OpHandleBase *op, BlockingQueue<size_t> *complete_q) {
+    OpHandleBase *op,
+    const std::shared_ptr<BlockingQueue<size_t>> &complete_q) {
   ++remaining_;
   this->pool_.enqueue([=] {
     OpHandleBase *op_to_run = op;
@@ -144,7 +145,7 @@ void FastThreadedSSAGraphExecutor::RunOpAsync(
             if (op_to_run == nullptr) {
               op_to_run = pending_op;
             } else {
-              this->RunOpAsync(op_deps, pending_op, complete_q);
+              RunOpAsync(op_deps, pending_op, complete_q);
             }
           }
         }
@@ -156,8 +157,7 @@ void FastThreadedSSAGraphExecutor::RunOpAsync(
 }
 void FastThreadedSSAGraphExecutor::PrepareAtomicOpDeps() {
   atomic_op_deps_ = pool_.enqueue([&] {
-    std::unordered_map<OpHandleBase *, std::atomic<int>> *op_deps =
-        new std::unordered_map<OpHandleBase *, std::atomic<int>>;
+    auto *op_deps = new std::unordered_map<OpHandleBase *, std::atomic<int>>;
     for (auto &pair : op_deps_) {
       (*op_deps)[pair.first] = pair.second;
     }

paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h

@@ -50,7 +50,8 @@ class FastThreadedSSAGraphExecutor : public SSAGraphExecutor {
   std::atomic<int> remaining_;
 
   void RunOpAsync(std::unordered_map<OpHandleBase *, std::atomic<int>> *op_deps,
-                  OpHandleBase *op, BlockingQueue<size_t> *complete_q);
+                  OpHandleBase *op,
+                  const std::shared_ptr<BlockingQueue<size_t>> &complete_q);
 
   void PrepareAtomicOpDeps();
 

paddle/fluid/framework/details/fused_broadcast_op_handle.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 "paddle/fluid/framework/details/fused_broadcast_op_handle.h"
+#include "paddle/fluid/framework/details/container_cast.h"
+#include "paddle/fluid/framework/details/variable_visitor.h"
+#include "paddle/fluid/platform/profiler.h"
+
+namespace paddle {
+namespace framework {
+namespace details {
+
+void FusedBroadcastOpHandle::RunImpl() {
+  platform::RecordEvent record_event(Name(), dev_ctxes_.begin()->second);
+
+  if (places_.size() == 1UL) return;
+
+  auto in_var_handles = DynamicCast<VarHandle>(inputs_);
+  auto out_var_handles = DynamicCast<VarHandle>(outputs_);
+
+  WaitInputVarGenerated();
+
+  std::vector<const Scope *> var_scopes;
+  for (auto *s : local_scopes_) {
+    var_scopes.emplace_back(s->FindVar(kLocalExecScopeName)->Get<Scope *>());
+  }
+
+  size_t place_num = places_.size();
+  PADDLE_ENFORCE_EQ(in_var_handles.size() * place_num, out_var_handles.size());
+
+  for (size_t i = 0; i < in_var_handles.size(); ++i) {
+    BroadcastOneVar(
+        *in_var_handles[i],
+        std::vector<VarHandle *>(out_var_handles.begin() + i * place_num,
+                                 out_var_handles.begin() + (i + 1) * place_num),
+        var_scopes);
+  }
+}
+
+std::string FusedBroadcastOpHandle::Name() const { return "fused_broadcast"; }
+
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/details/fused_broadcast_op_handle.h

+//   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include <map>
+#include <string>
+#include <vector>
+
+#include "paddle/fluid/framework/details/broadcast_op_handle.h"
+#include "paddle/fluid/framework/details/multi_devices_helper.h"
+#include "paddle/fluid/framework/lod_tensor.h"
+#include "paddle/fluid/framework/scope.h"
+#include "paddle/fluid/framework/selected_rows.h"
+#include "paddle/fluid/platform/device_context.h"
+
+#ifdef PADDLE_WITH_CUDA
+#include "paddle/fluid/platform/nccl_helper.h"
+#endif
+
+namespace paddle {
+namespace framework {
+namespace details {
+
+struct FusedBroadcastOpHandle : public BroadcastOpHandle {
+ public:
+#ifdef PADDLE_WITH_CUDA
+  FusedBroadcastOpHandle(ir::Node *node,
+                         const std::vector<Scope *> local_scopes,
+                         const std::vector<platform::Place> &places,
+                         const platform::NCCLContextMap *nccl_ctx)
+      : BroadcastOpHandle(node, local_scopes, places, nccl_ctx) {}
+#else
+  FusedBroadcastOpHandle(ir::Node* node, const std::vector<Scope*> local_scopes,
+                         const std::vector<platform::Place>& places)
+      : BroadcastOpHandle(node, local_scopes, places) {}
+#endif
+  std::string Name() const override;
+
+ protected:
+  void RunImpl() override;
+};
+
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/details/fused_broadcast_op_handle_test.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 "paddle/fluid/framework/details/fused_broadcast_op_handle.h"
+#include "gtest/gtest.h"
+#include "paddle/fluid/framework/details/broadcast_op_handle_test.h"
+
+namespace paddle {
+namespace framework {
+namespace details {
+
+struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle {
+  std::vector<std::string> out_varnames_;
+
+  void InitFusedBroadcastOp(std::vector<size_t> input_scope_idxes) {
+    // initialize scope and var
+    for (size_t i = 0; i < place_list_.size(); ++i) {
+      local_scopes_.push_back(&(g_scope_.NewScope()));
+      Scope& local_scope = local_scopes_.back()->NewScope();
+      *local_scopes_.back()
+           ->Var(details::kLocalExecScopeName)
+           ->GetMutable<Scope*>() = &local_scope;
+      for (size_t j = 0; j < input_scope_idxes.size(); ++j) {
+        local_scope.Var("out_var" + j);
+        if (i == j) local_scope.Var("in_var" + j);
+      }
+      param_scopes_.emplace_back(&local_scope);
+    }
+
+    // create op handle node
+    std::unique_ptr<ir::Node> n =
+        ir::CreateNodeForTest("fused_broadcast", ir::Node::Type::kOperation);
+    if (use_gpu_) {
+#ifdef PADDLE_WITH_CUDA
+      op_handle_.reset(new FusedBroadcastOpHandle(
+          n.get(), local_scopes_, place_list_, nccl_ctxs_.get()));
+#else
+      PADDLE_THROW("CUDA is not supported.");
+#endif
+    } else {
+#ifdef PADDLE_WITH_CUDA
+      op_handle_.reset(new FusedBroadcastOpHandle(
+          n.get(), local_scopes_, place_list_, nccl_ctxs_.get()));
+#else
+      op_handle_.reset(
+          new FusedBroadcastOpHandle(n.get(), local_scopes_, place_list_));
+#endif
+    }
+
+    for (size_t i = 0; i < input_scope_idxes.size(); ++i) {
+      // add input var handle
+      std::unique_ptr<ir::Node> in_node =
+          ir::CreateNodeForTest("in_node" + i, ir::Node::Type::kVariable);
+      VarHandle* in_var_handle =
+          new VarHandle(in_node.get(), 1, input_scope_idxes[i], "in_var" + i,
+                        place_list_[input_scope_idxes[i]]);
+      vars_.emplace_back(in_var_handle);
+      op_handle_->AddInput(in_var_handle);
+
+      // add output var handle
+      for (size_t j = 0; j < place_list_.size(); ++j) {
+        std::unique_ptr<ir::Node> out_node =
+            ir::CreateNodeForTest("out_node" + i, ir::Node::Type::kVariable);
+        VarHandle* out_var_handle =
+            new VarHandle(out_node.get(), 2, j, "out_var" + i, place_list_[j]);
+        vars_.emplace_back(out_var_handle);
+        op_handle_->AddOutput(out_var_handle);
+      }
+    }
+  }
+
+  void TestFusedBroadcastLoDTensor(std::vector<size_t> input_scope_idxes) {
+    std::vector<std::vector<float>> send_vec;
+    f::LoD lod{{0, 10, 20}};
+    for (size_t i = 0; i < input_scope_idxes.size(); ++i) {
+      const std::string varname("in_var" + i);
+      float val_scalar = static_cast<float>(i);
+      send_vec.push_back(
+          InitLoDTensor(varname, input_scope_idxes[i], lod, val_scalar));
+    }
+
+    op_handle_->Run(false);
+
+    WaitAll();
+    for (size_t i = 0; i < input_scope_idxes.size(); ++i) {
+      const std::string& varname("out_var" + i);
+      for (size_t j = 0; j < place_list_.size(); ++j) {
+        LoDTensorEqual(varname, send_vec[i], lod, param_scopes_[j]);
+      }
+    }
+  }
+
+  void TestFusedBroadcastSelectedRows(std::vector<size_t> input_scope_idxes) {
+    std::vector<std::vector<float>> send_vector;
+    std::vector<int64_t> rows{0, 1, 2, 3, 3, 0, 14, 7, 3, 1,
+                              2, 4, 6, 3, 1, 1, 1,  1, 3, 7};
+    int height = static_cast<int>(kDims[0] * 2);
+    for (size_t i = 0; i < input_scope_idxes.size(); ++i) {
+      const std::string varname("in_var" + i);
+      float val_scalar = static_cast<float>(i);
+      send_vector.push_back(InitSelectedRows(varname, input_scope_idxes[i],
+                                             rows, height, val_scalar));
+    }
+
+    op_handle_->Run(false);
+
+    WaitAll();
+    for (size_t i = 0; i < input_scope_idxes.size(); ++i) {
+      const std::string& varname("out_var" + i);
+      for (size_t j = 0; j < place_list_.size(); ++j) {
+        SelectedRowsEqual(varname, input_scope_idxes[i], send_vector[i], rows,
+                          height);
+      }
+    }
+  }
+};
+
+TEST(FusedBroadcastTester, CPULodTensor) {
+  TestFusedBroadcastOpHandle test_op;
+  std::vector<size_t> input_scope_idxes = {0, 1};
+  test_op.InitCtxOnGpu(false);
+  test_op.InitFusedBroadcastOp(input_scope_idxes);
+  test_op.TestFusedBroadcastLoDTensor(input_scope_idxes);
+}
+
+TEST(FusedBroadcastTester, CPUSelectedRows) {
+  TestFusedBroadcastOpHandle test_op;
+  std::vector<size_t> input_scope_idxes = {0, 1};
+  test_op.InitCtxOnGpu(false);
+  test_op.InitFusedBroadcastOp(input_scope_idxes);
+  test_op.TestFusedBroadcastSelectedRows(input_scope_idxes);
+}
+
+#ifdef PADDLE_WITH_CUDA
+TEST(FusedBroadcastTester, GPULodTensor) {
+  TestFusedBroadcastOpHandle test_op;
+  std::vector<size_t> input_scope_idxes = {0, 1};
+  test_op.InitCtxOnGpu(true);
+  test_op.InitFusedBroadcastOp(input_scope_idxes);
+  test_op.TestFusedBroadcastLoDTensor(input_scope_idxes);
+}
+
+TEST(FusedBroadcastTester, GPUSelectedRows) {
+  TestFusedBroadcastOpHandle test_op;
+  std::vector<size_t> input_scope_idxes = {0, 1};
+  test_op.InitCtxOnGpu(true);
+  test_op.InitFusedBroadcastOp(input_scope_idxes);
+  test_op.TestFusedBroadcastSelectedRows(input_scope_idxes);
+}
+#endif
+
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/details/gather_op_handle.cc

@@ -36,7 +36,7 @@ void GatherOpHandle::RunImpl() {
 
   VarHandle *out_var_handle;
   {
-    auto out_var_handles = DynamicCast<VarHandle>(outputs_);
+    auto out_var_handles = DynamicCast<VarHandle>(this->Outputs());
     PADDLE_ENFORCE_EQ(out_var_handles.size(), 1,
                       "The number of output should be one.");
     out_var_handle = out_var_handles.front();
@@ -99,7 +99,7 @@ void GatherOpHandle::RunImpl() {
   Tensor *out_tensor = out_value->mutable_value();
 
   // copy
-  auto dev_ctx = dev_ctxes_[out_var_handle->place_];
+  auto dev_ctx = dev_ctxes_.at(out_var_handle->place_);
   RunAndRecordEvent(out_var_handle->place_, [in_tensors, out_tensor, &dev_ctx,
                                              t_out_p] {
     int s = 0, e = 0;

paddle/fluid/framework/details/multi_devices_graph_pass.cc

@@ -21,6 +21,7 @@
 #include "paddle/fluid/framework/details/broadcast_op_handle.h"
 #include "paddle/fluid/framework/details/computation_op_handle.h"
 #include "paddle/fluid/framework/details/data_balance_op_handle.h"
+#include "paddle/fluid/framework/details/fused_broadcast_op_handle.h"
 #include "paddle/fluid/framework/details/multi_devices_graph_pass.h"
 #include "paddle/fluid/framework/details/reduce_op_handle.h"
 #include "paddle/fluid/framework/details/rpc_op_handle.h"
@@ -347,7 +348,7 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
           BuildStrategy::GradientScaleStrategy::kCustomized) {
         // TODO(paddle-dev): Why is there no input for this op_handle?
         auto loss_grad_name = node->Op()->OutputArgumentNames()[0];
-        CreateScaleLossGradOp(&result, loss_grad_name);
+        CreateScaleLossGradOp(&result, loss_grad_name, node->outputs[0]);
       }
       // This assumes the backward generating code will ensure IsScaleLossOp
       // is true only for the op that scale the final scalar loss.
@@ -436,10 +437,14 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
   if ((use_gpu &&
        strategy_.reduce_ == BuildStrategy::ReduceStrategy::kReduce) ||
       is_dist_train) {
-    for (size_t dev_id = 0; dev_id < bcast_var_name_set.size(); ++dev_id) {
-      auto &to_bcast_set = bcast_var_name_set[dev_id];
-      for (auto &bcast_name : to_bcast_set) {
-        CreateBroadcastOp(&result, bcast_name, dev_id);
+    if (strategy_.fuse_broadcast_op_) {
+      CreateFusedBroadcastOp(&result, bcast_var_name_set);
+    } else {
+      for (size_t dev_id = 0; dev_id < bcast_var_name_set.size(); ++dev_id) {
+        auto &to_bcast_set = bcast_var_name_set[dev_id];
+        for (auto &bcast_name : to_bcast_set) {
+          CreateBroadcastOp(&result, bcast_name, dev_id);
+        }
       }
     }
   }
@@ -508,6 +513,44 @@ void MultiDevSSAGraphBuilder::CreateBroadcastOp(ir::Graph *result,
   }
 }
 
+void MultiDevSSAGraphBuilder::CreateFusedBroadcastOp(
+    ir::Graph *result,
+    const std::vector<std::unordered_set<std::string>> &bcast_varnames) const {
+#ifdef PADDLE_WITH_CUDA
+  auto *op_handle = new FusedBroadcastOpHandle(
+      result->CreateEmptyNode("fused_broadcast", ir::Node::Type::kOperation),
+      local_scopes_, places_, nccl_ctxs_);
+#else
+  auto *op_handle = new FusedBroadcastOpHandle(
+      result->CreateEmptyNode("fused_broadcast", ir::Node::Type::kOperation),
+      local_scopes_, places_);
+#endif
+  result->Get<GraphOps>(kGraphOps).emplace_back(op_handle);
+
+  for (size_t i = 0; i < places_.size(); ++i) {
+    auto &p = places_[i];
+    SetCommunicationContext(op_handle, p);
+  }
+
+  for (size_t dev_id = 0; dev_id < bcast_varnames.size(); ++dev_id) {
+    for (auto &p_name : bcast_varnames[dev_id]) {
+      auto *in =
+          result->Get<GraphVars>(kGraphVars).at(dev_id).at(p_name).back().get();
+      op_handle->AddInput(in);
+      for (size_t out_dev_id = 0; out_dev_id < places_.size(); ++out_dev_id) {
+        auto &p = places_[out_dev_id];
+        auto &vars =
+            result->Get<GraphVars>(kGraphVars).at(out_dev_id).at(p_name);
+        auto *out_var = new VarHandle(
+            result->CreateEmptyNode(p_name, ir::Node::Type::kVariable),
+            vars.size(), out_dev_id, p_name, p);
+        vars.emplace_back(out_var);
+        op_handle->AddOutput(out_var);
+      }
+    }
+  }
+}
+
 void MultiDevSSAGraphBuilder::CreateComputationalOp(ir::Graph *result,
                                                     ir::Node *node,
                                                     int dev_id) const {
@@ -602,7 +645,8 @@ int MultiDevSSAGraphBuilder::GetVarDeviceID(const ir::Graph &graph,
 }
 
 void MultiDevSSAGraphBuilder::CreateScaleLossGradOp(
-    ir::Graph *result, const std::string &loss_grad_name) const {
+    ir::Graph *result, const std::string &loss_grad_name,
+    ir::Node *out_var_node) const {
   for (size_t i = 0; i < places_.size(); ++i) {
     // Insert ScaleCost OpHandle
     auto *dev_ctx = platform::DeviceContextPool::Instance().Get(places_[i]);
@@ -617,10 +661,8 @@ void MultiDevSSAGraphBuilder::CreateScaleLossGradOp(
     // loss->pending_ops_.emplace_back(op_handle);
     // op_handle->inputs_.emplace_back(loss);
 
-    CreateOpOutput(
-        result, op_handle,
-        result->CreateEmptyNode(loss_grad_name, ir::Node::Type::kVariable),
-        places_[i], i);
+    CreateOpOutput(result, op_handle,
+                   result->CreateVarNode(out_var_node->Var()), places_[i], i);
   }
 }
 
@@ -680,7 +722,8 @@ int MultiDevSSAGraphBuilder::CreateDistTrainOp(ir::Graph *result,
   }
 
   if (node->Op()->Type() == "split_byref" ||
-      node->Op()->Type() == "split_selected_rows") {
+      node->Op()->Type() == "split_selected_rows" ||
+      node->Op()->Type() == "split_ids") {
     // TODO(paddle-dev): getting the first var is not safe.
     op_dev_id = GetVarDeviceID(*result, input_var_names[0]);
     if (strategy_.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce) {

paddle/fluid/framework/details/multi_devices_graph_pass.h

@@ -61,7 +61,8 @@ class MultiDevSSAGraphBuilder : public ir::Pass {
                               size_t num_places) const;
 
   void CreateScaleLossGradOp(ir::Graph *result,
-                             const std::string &loss_grad_name) const;
+                             const std::string &loss_grad_name,
+                             ir::Node *out_var_node) const;
 
   VarHandle *CreateReduceOp(ir::Graph *result, const std::string &og,
                             int dst_dev_id) const;
@@ -78,6 +79,10 @@ class MultiDevSSAGraphBuilder : public ir::Pass {
   void CreateBroadcastOp(ir::Graph *result, const std::string &p_name,
                          size_t src_dev_id) const;
 
+  void CreateFusedBroadcastOp(
+      ir::Graph *result,
+      const std::vector<std::unordered_set<std::string>> &bcast_varnames) const;
+
   bool IsSparseGradient(const std::string &og) const;
 
   size_t GetAppropriateDeviceID(

paddle/fluid/framework/details/op_handle_base.cc

@@ -103,7 +103,7 @@ void OpHandleBase::WaitInputVarGenerated() {
 void OpHandleBase::WaitInputVarGenerated(const platform::Place &place) {
   for (auto *in : inputs_) {
     if (NeedWait(in)) {
-      in->GeneratedOp()->RecordWaitEventOnCtx(dev_ctxes_[place]);
+      in->GeneratedOp()->RecordWaitEventOnCtx(dev_ctxes_.at(place));
     }
   }
 }

paddle/fluid/framework/details/reduce_op_handle.cc

@@ -27,7 +27,7 @@ namespace framework {
 namespace details {
 
 void ReduceOpHandle::RunImpl() {
-  platform::RecordEvent record_event(Name(), dev_ctxes_.begin()->second);
+  platform::RecordEvent record_event(Name(), dev_ctxes_.cbegin()->second);
 
   if (places_.size() == 1) return;
   // the input and output may have dummy var.

paddle/fluid/framework/details/reduce_op_handle.h

@@ -46,7 +46,8 @@ struct ReduceOpHandle : public OpHandleBase {
         nccl_ctxs_(nccl_ctxs) {
     if (nccl_ctxs_) {
       for (auto &p_ctx : nccl_ctxs_->contexts_) {
-        dev_ctxes_[platform::CUDAPlace(p_ctx.first)] = p_ctx.second.ctx_.get();
+        this->SetDeviceContext(platform::CUDAPlace(p_ctx.first),
+                               p_ctx.second.ctx_.get());
       }
     }
   }

paddle/fluid/framework/details/rpc_op_handle.cc

@@ -38,7 +38,7 @@ void RPCOpHandle::RunImpl() {
       continue;
     }
     if (in->GeneratedOp()) {
-      in->GeneratedOp()->RecordWaitEventOnCtx(dev_ctxes_[p]);
+      in->GeneratedOp()->RecordWaitEventOnCtx(dev_ctxes_.at(p));
     }
   }
   auto &tmp_scope = local_scope_->FindVar(kLocalExecScopeName)->Get<Scope *>();

paddle/fluid/framework/details/scale_loss_grad_op_handle.cc

@@ -27,7 +27,7 @@ ScaleLossGradOpHandle::ScaleLossGradOpHandle(ir::Node *node, size_t num_dev,
       coeff_(static_cast<float>(1.0 / num_dev)),
       scope_(scope),
       place_(place) {
-  dev_ctxes_[place_] = dev_ctx;
+  this->SetDeviceContext(place_, dev_ctx);
 }
 
 ScaleLossGradOpHandle::~ScaleLossGradOpHandle() {}
@@ -46,9 +46,9 @@ void ScaleLossGradOpHandle::RunImpl() {
   } else {
 #ifdef PADDLE_WITH_CUDA
     this->RunAndRecordEvent([&] {
-      auto stream =
-          static_cast<platform::CUDADeviceContext *>(this->dev_ctxes_[place_])
-              ->stream();
+      auto stream = static_cast<platform::CUDADeviceContext *>(
+                        this->dev_ctxes_.at(place_))
+                        ->stream();
       memory::Copy(boost::get<platform::CUDAPlace>(place_), tmp,
                    platform::CPUPlace(), &coeff_, sizeof(float), stream);
       VLOG(10) << place_ << "RUN Scale loss grad op";

paddle/fluid/framework/details/threaded_ssa_graph_executor.cc

@@ -39,7 +39,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
       new platform::RecordEvent("ThreadedSSAGraphExecutorPrepare", nullptr));
   std::unordered_map<OpHandleBase *, size_t> pending_ops;
   std::unordered_set<VarHandleBase *> pending_vars;
-  BlockingQueue<VarHandleBase *> ready_vars;
+  auto ready_vars = std::make_shared<BlockingQueue<VarHandleBase *>>();
   std::unordered_set<OpHandleBase *> ready_ops;
   // For ops (e.g. nccl_all_reduce) that need to coordinate multiple
   // streams from multiple GPUs, it's faster to buffer them and schedule
@@ -51,12 +51,12 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
   for (auto &var_map : graph_->Get<details::GraphVars>(details::kGraphVars)) {
     for (auto &name_pair : var_map) {
       for (auto &version_pair : name_pair.second) {
-        InsertPendingVar(&pending_vars, &ready_vars, version_pair.get());
+        InsertPendingVar(&pending_vars, ready_vars.get(), version_pair.get());
       }
     }
   }
   for (auto &var : graph_->Get<details::GraphDepVars>(details::kGraphDepVars)) {
-    InsertPendingVar(&pending_vars, &ready_vars, var.get());
+    InsertPendingVar(&pending_vars, ready_vars.get(), var.get());
   }
 
   for (auto &op : graph_->Get<details::GraphOps>(details::kGraphOps)) {
@@ -73,12 +73,12 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
   FeedFetchList fetch_data(fetch_tensors.size());
 
   InsertFetchOps(fetch_tensors, &fetch_ops, &fetch_dependencies, &pending_ops,
-                 &pending_vars, &ready_vars, &fetch_data);
+                 &pending_vars, ready_vars.get(), &fetch_data);
 
   auto run_all_ops = [&](std::unordered_set<OpHandleBase *> &set) {
     for (auto *op : set) {
       running_ops_++;
-      RunOp(&ready_vars, op);
+      RunOp(ready_vars, op);
     }
     set.clear();
   };
@@ -87,7 +87,6 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
   run_op_futures_.clear();
   exception_holder_.Clear();
   event.reset(nullptr);
-
   // Step 3. Execution
   while (!pending_vars.empty()) {
     // 1. Run All Ready ops
@@ -103,7 +102,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
 
     // 2. Find ready variable
     bool timeout;
-    auto cur_ready_vars = ready_vars.PopAll(1, &timeout);
+    auto cur_ready_vars = ready_vars->PopAll(1, &timeout);
 
     if (timeout) {
       if (exception_holder_.IsCaught()) {
@@ -133,7 +132,6 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
     }
   }
   PADDLE_ENFORCE(ready_ops.empty());
-
   // Wait FetchOps.
   ClearFetchOp(graph_.get(), &fetch_ops);
 
@@ -206,7 +204,8 @@ void ThreadedSSAGraphExecutor::InsertPendingVar(
 }
 
 void ThreadedSSAGraphExecutor::RunOp(
-    BlockingQueue<VarHandleBase *> *ready_var_q, details::OpHandleBase *op) {
+    const std::shared_ptr<BlockingQueue<VarHandleBase *>> &ready_var_q,
+    details::OpHandleBase *op) {
   auto op_run = [ready_var_q, op, this] {
     try {
       if (VLOG_IS_ON(10)) {

paddle/fluid/framework/details/threaded_ssa_graph_executor.h

@@ -51,7 +51,7 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
   ~ThreadedSSAGraphExecutor() {}
 
  private:
-  void RunOp(BlockingQueue<VarHandleBase *> *ready_var_q,
+  void RunOp(const std::shared_ptr<BlockingQueue<VarHandleBase *>> &ready_var_q,
              details::OpHandleBase *op);
 
  private:

paddle/fluid/framework/framework.proto

@@ -35,6 +35,7 @@ enum AttrType {
   BLOCK = 8;
   LONG = 9;
   BLOCKS = 10;
+  LONGS = 11;
 }
 
 // OpDesc describes an instance of a C++ framework::OperatorBase
@@ -55,6 +56,7 @@ message OpDesc {
     optional int32 block_idx = 12;
     optional int64 l = 13;
     repeated int32 blocks_idx = 14;
+    repeated int64 longs = 15;
   };
 
   message Var {

paddle/fluid/framework/ir/CMakeLists.txt

@@ -36,10 +36,12 @@ pass_library(fc_lstm_fuse_pass inference)
 pass_library(embedding_fc_lstm_fuse_pass inference)
 pass_library(fc_gru_fuse_pass inference)
 pass_library(seq_concat_fc_fuse_pass inference)
+pass_library(multi_batch_merge_pass base)
 pass_library(conv_bn_fuse_pass inference)
 pass_library(seqconv_eltadd_relu_fuse_pass inference)
 if(WITH_MKLDNN)
     pass_library(mkldnn_placement_pass base)
+    pass_library(depthwise_conv_mkldnn_pass base)
     pass_library(conv_bias_mkldnn_fuse_pass inference)
     pass_library(conv_relu_mkldnn_fuse_pass inference)
     pass_library(conv_elementwise_add_mkldnn_fuse_pass inference)
@@ -58,6 +60,7 @@ cc_test(graph_to_program_pass_test SRCS graph_to_program_pass_test.cc DEPS graph
 cc_test(test_graph_pattern_detector SRCS graph_pattern_detector_tester.cc DEPS graph_pattern_detector)
 cc_test(test_fc_fuse_pass SRCS fc_fuse_pass_tester.cc DEPS fc_fuse_pass framework_proto)
 if (WITH_MKLDNN)
+    cc_test(test_depthwise_conv_mkldnn_pass SRCS depthwise_conv_mkldnn_pass_tester.cc DEPS depthwise_conv_mkldnn_pass)
     cc_test(test_conv_relu_mkldnn_fuse_pass SRCS conv_relu_mkldnn_fuse_pass_tester.cc DEPS conv_relu_mkldnn_fuse_pass)
     cc_test(test_conv_elementwise_add_mkldnn_fuse_pass SRCS conv_elementwise_add_mkldnn_fuse_pass_tester.cc DEPS conv_elementwise_add_mkldnn_fuse_pass)
 endif ()

paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.h

@@ -31,7 +31,8 @@ class ConvReLUFusePass : public FusePassBase {
   virtual ~ConvReLUFusePass() {}
 
  protected:
-  std::unique_ptr<ir::Graph> ApplyImpl(std::unique_ptr<ir::Graph> graph) const;
+  std::unique_ptr<ir::Graph> ApplyImpl(
+      std::unique_ptr<ir::Graph> graph) const override;
 };
 
 }  // namespace ir

paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass_tester.cc

@@ -15,6 +15,7 @@
 #include "paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.h"
 
 #include <gtest/gtest.h>
+#include "paddle/fluid/framework/op_proto_maker.h"
 
 namespace paddle {
 namespace framework {
@@ -36,6 +37,8 @@ void SetOp(ProgramDesc* prog, const std::string& type, const std::string& name,
     op->SetInput("X", inputs);
   }
   op->SetOutput("Out", outputs);
+  op->SetAttr(OpProtoAndCheckerMaker::OpRoleAttrName(),
+              static_cast<int>(OpRole::kForward));
 }
 
 // a->OP0->b

paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.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 "paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.h"
+#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+#define GET_NODE(id, pattern)                               \
+  PADDLE_ENFORCE(subgraph.count(pattern.RetrieveNode(#id)), \
+                 "pattern has no Node called %s", #id);     \
+  auto* id = subgraph.at(pattern.RetrieveNode(#id));        \
+  PADDLE_ENFORCE_NOT_NULL(id, "subgraph has no node %s", #id);
+
+std::unique_ptr<ir::Graph> DepthwiseConvMKLDNNPass::ApplyImpl(
+    std::unique_ptr<ir::Graph> graph) const {
+  PADDLE_ENFORCE(graph.get());
+  FusePassBase::Init("depthwise_conv_mkldnn_pass", graph.get());
+  GraphPatternDetector gpd;
+
+  auto* pattern = gpd.mutable_pattern();
+  pattern->NewNode("depthwise_conv")
+      ->assert_is_op("depthwise_conv2d")
+      ->assert_op_attr("use_mkldnn", true);
+
+  int found_depthwise_conv_mkldnn_count = 0;
+  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
+                     Graph* g) {
+    VLOG(3) << "handle DepthwiseConvMKLDNN fuse";
+    GET_NODE(depthwise_conv, (*pattern));
+    depthwise_conv->Op()->SetType("conv2d");
+    found_depthwise_conv_mkldnn_count++;
+  };
+
+  gpd(graph.get(), handler);
+  AddStatis(found_depthwise_conv_mkldnn_count);
+  return graph;
+}
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_PASS(depthwise_conv_mkldnn_pass,
+              paddle::framework::ir::DepthwiseConvMKLDNNPass);

paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.h

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include "paddle/fluid/framework/ir/fuse_pass_base.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+class DepthwiseConvMKLDNNPass : public FusePassBase {
+ public:
+  virtual ~DepthwiseConvMKLDNNPass() {}
+
+ protected:
+  std::unique_ptr<ir::Graph> ApplyImpl(
+      std::unique_ptr<ir::Graph> graph) const override;
+};
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass_tester.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 "paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.h"
+
+#include <gtest/gtest.h>
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+void SetOp(ProgramDesc* prog, const std::string& type, const std::string& name,
+           const std::vector<std::string>& inputs,
+           const std::vector<std::string>& outputs, bool use_mkldnn = false) {
+  auto* op = prog->MutableBlock(0)->AppendOp();
+  op->SetType(type);
+  op->SetAttr("use_mkldnn", use_mkldnn);
+  op->SetAttr("name", name);
+  op->SetInput("Input", {inputs[0]});
+  op->SetInput("Filter", {inputs[1]});
+  op->SetInput("Bias", {inputs[2]});
+  op->SetOutput("Out", outputs);
+}
+
+// (a, weights, bias)->depthwise conv mkldnn->b
+// (b, weights2, bias2)->depthwise conv no mkldnn->c
+// (c, weights3, bias3)->conv mkldnn->d
+// (d, weights3, bias3)->conv no mkldnn->e
+ProgramDesc BuildProgramDesc() {
+  ProgramDesc prog;
+  for (auto& v : std::vector<std::string>(
+           {"a", "b", "c", "d", "e", "weights", "bias", "weights2", "bias2",
+            "weights3", "bias3", "weights4", "bias4"})) {
+    auto* var = prog.MutableBlock(0)->Var(v);
+    var->SetType(proto::VarType::SELECTED_ROWS);
+    if (v == "weights" || v == "bias" || v == "weights2" || v == "bias2" ||
+        v == "weights3" || v == "bias3" || v == "weights4" || v == "bias4") {
+      var->SetPersistable(true);
+    }
+  }
+
+  // depthwise conv with MKL-DNN
+  SetOp(&prog, "depthwise_conv2d", "conv1",
+        std::vector<std::string>({"a", "weights", "bias"}),
+        std::vector<std::string>({"b"}), true);
+  // depthwise conv without MKL-DNN
+  SetOp(&prog, "depthwise_conv2d", "conv2",
+        std::vector<std::string>({"b", "weights2", "bias2"}),
+        std::vector<std::string>({"c"}), false);
+  // conv with MKL-DNN
+  SetOp(&prog, "conv2d", "conv3",
+        std::vector<std::string>({"c", "weights3", "bias3"}),
+        std::vector<std::string>({"d"}), true);
+  // conv without MKL-dNN
+  SetOp(&prog, "conv2d", "conv4",
+        std::vector<std::string>({"d", "weights4", "bias4"}),
+        std::vector<std::string>({"e"}), false);
+
+  return prog;
+}
+
+TEST(DepthwiseConvMKLDNNPass, basic) {
+  auto prog = BuildProgramDesc();
+
+  std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
+
+  auto pass = PassRegistry::Instance().Get("depthwise_conv_mkldnn_pass");
+
+  struct counters {
+    int mkldnn_depthwise_conv_nodes;
+    int other_depthwise_conv_nodes;
+    int mkldnn_conv_nodes;
+    int other_conv_nodes;
+  };
+
+  counters before{1, 1, 1, 1};
+
+  graph = pass->Apply(std::move(graph));
+
+  // initialize counters before loop
+  counters after{0, 0, 0, 0};
+
+  for (auto* node : graph->Nodes()) {
+    if (node->IsOp()) {
+      auto* op = node->Op();
+      if (op->Type() == "conv2d") {
+        if (boost::get<bool>(op->GetAttr("use_mkldnn")))
+          after.mkldnn_conv_nodes++;
+        else
+          after.other_conv_nodes++;
+      } else if (op->Type() == "depthwise_conv2d") {
+        if (boost::get<bool>(op->GetAttr("use_mkldnn")))
+          after.mkldnn_depthwise_conv_nodes++;
+        else
+          after.other_depthwise_conv_nodes++;
+      }
+    }
+  }
+
+  EXPECT_EQ(after.other_depthwise_conv_nodes,
+            before.other_depthwise_conv_nodes);
+  EXPECT_EQ(after.other_conv_nodes, before.other_conv_nodes);
+  EXPECT_EQ(after.mkldnn_depthwise_conv_nodes,
+            before.mkldnn_depthwise_conv_nodes - 1);
+  EXPECT_EQ(after.mkldnn_conv_nodes, before.mkldnn_conv_nodes + 1);
+}
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+USE_PASS(depthwise_conv_mkldnn_pass);

paddle/fluid/framework/ir/fc_fuse_pass_tester.cc

@@ -15,6 +15,7 @@
 #include "paddle/fluid/framework/ir/fc_fuse_pass.h"
 
 #include <gtest/gtest.h>
+#include "paddle/fluid/framework/op_proto_maker.h"
 
 namespace paddle {
 namespace framework {
@@ -32,6 +33,8 @@ void SetOp(ProgramDesc* prog, const std::string& type,
     op->SetInput("X", inputs);
   }
   op->SetOutput("Out", outputs);
+  op->SetAttr(OpProtoAndCheckerMaker::OpRoleAttrName(),
+              static_cast<int>(OpRole::kForward));
 }
 
 // a->OP0->b

paddle/fluid/framework/ir/graph.cc

@@ -23,80 +23,78 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 namespace ir {
-
-std::vector<std::string> FindDistTrainSendVars(
-    const std::vector<ir::Node *> &nodes) {
-  std::vector<std::string> send_vars;
-  // since parameters are all in block 0,
-  // it's enough to only scan send ops in block 0
-  for (auto &node : nodes) {
-    auto op_vars = node->Op()->InputArgumentNames();
-    send_vars.reserve(send_vars.size() +
-                      std::distance(op_vars.begin(), op_vars.end()));
-    send_vars.insert(send_vars.end(), op_vars.begin(), op_vars.end());
-  }
-  return send_vars;
-}
-
-std::vector<std::string> FindDistTrainRecvVars(
-    const std::vector<ir::Node *> &nodes) {
-  std::vector<std::string> recv_vars;
-  for (auto &node : nodes) {
-    auto op_vars = node->Op()->OutputArgumentNames();
-    recv_vars.reserve(recv_vars.size() +
-                      std::distance(op_vars.begin(), op_vars.end()));
-    recv_vars.insert(recv_vars.end(), op_vars.begin(), op_vars.end());
-  }
-  return recv_vars;
-}
-
-bool IsDistTrainOp(ir::Node *node, const std::vector<std::string> &send_vars,
-                   const std::vector<std::string> &recv_vars) {
-  if (send_vars.size() == 0 || recv_vars.size() == 0) {
-    return false;
-  }
-
-  /**
-   * Check any of opvars contains `.block` and in sendvars
-   */
-  auto checker = [](const std::vector<std::string> &opvars,
-                    const std::vector<std::string> &rpc_vars) -> bool {
-    for (auto &var : opvars) {
-      // a variable name with the suffix `.block` means it's a splited
-      // variable by (DistributeTranspiler)
-      // [python/paddle/fluid/transpiler/distribute_transpiler.py]
-      if (var.find(".block") != std::string::npos &&
-          std::find(rpc_vars.begin(), rpc_vars.end(), var) != rpc_vars.end()) {
-        return true;
+namespace {
+
+void CheckProgram(const ProgramDesc &program) {
+  std::map<int, bool> visit;
+#define _INT(role) static_cast<int>(role)
+
+  for (size_t i = 0; i < program.Size(); ++i) {
+    for (OpDesc *op : program.Block(i).AllOps()) {
+      // For backward compatibility, some program doesn't have role added.
+      if (!op->HasAttr(OpProtoAndCheckerMaker::OpRoleAttrName())) continue;
+      int role_id = boost::get<int>(
+          op->GetAttr(OpProtoAndCheckerMaker::OpRoleAttrName()));
+      visit[role_id] = true;
+      switch (role_id) {
+        case _INT(OpRole::kForward):
+          PADDLE_ENFORCE(
+              visit.find(_INT(OpRole::kBackward)) == visit.end(),
+              "Cannot add forward operator before backward operator.");
+          break;
+        case _INT(OpRole::kBackward):
+        case _INT(OpRole::kBackward) | _INT(OpRole::kLoss):
+          PADDLE_ENFORCE(
+              visit.find(_INT(OpRole::kOptimize)) == visit.end(),
+              "Cannot add backward operator before optimize operator.");
+          break;
+        case _INT(OpRole::kForward) | _INT(OpRole::kLoss):
+          PADDLE_ENFORCE(visit.find(_INT(OpRole::kBackward) |
+                                    _INT(OpRole::kLoss)) == visit.end(),
+                         "Cannot add backward|loss operator before "
+                         "forward|loss operator.");
+          PADDLE_ENFORCE(
+              visit.find(_INT(OpRole::kOptimize)) == visit.end(),
+              "Cannot add backward operator before optimize operator.");
+          break;
+        case _INT(OpRole::kOptimize):
+        case _INT(OpRole::kOptimize) | _INT(OpRole::kLRSched):
+          PADDLE_ENFORCE(visit.find(_INT(OpRole::kBackward)) != visit.end(),
+                         "Optimize operators must follow backward operator.");
+          break;
+        case _INT(OpRole::kLRSched):
+        case _INT(OpRole::kDist):
+        case _INT(OpRole::kRPC):
+        case _INT(OpRole::kNotSpecified):
+          break;
+        default:
+          LOG(FATAL) << "Unknown operator role. Don't add new role because "
+                        "you don't know what you are doing.";
       }
     }
-    return false;
-  };
-
-  std::vector<std::string> input_var_names;
-  std::vector<std::string> output_var_names;
-  for (ir::Node *input : node->inputs) {
-    input_var_names.push_back(input->Name());
   }
-  for (ir::Node *output : node->outputs) {
-    output_var_names.push_back(output->Name());
-  }
-
-  return checker(output_var_names, send_vars) ||
-         checker(input_var_names, recv_vars);
+#undef _INT
 }
+}  // namespace
 
 Graph::Graph(const ProgramDesc &program) : program_(program) {
+  CheckProgram(program_);
   // Make the nodes id start from 0.
   Node::ResetId();
+  auto var_nodes = InitFromProgram(program_);
+  ResolveHazard(var_nodes);
+}
 
+std::map<std::string, std::vector<ir::Node *>> Graph::InitFromProgram(
+    const ProgramDesc &program) {
   VLOG(3) << "block in program:" << program_.Size();
   std::unordered_map<std::string, VarDesc *> all_vars;
+  // var nodes for each var name, will have multiple versions in SSA
+  std::map<std::string, std::vector<ir::Node *>> var_nodes;
   for (auto *var : program.Block(0).AllVars()) {
     all_vars.emplace(var->Name(), var);
   }
 
-  std::map<std::string, std::vector<ir::Node *>> var_nodes;
   for (auto *op : program.Block(0).AllOps()) {
     ir::Node *node = CreateOpNode(op);
     // For input args, reuse the same var name if it was created before.
@@ -134,7 +132,11 @@ Graph::Graph(const ProgramDesc &program) : program_(program) {
       var->inputs.push_back(node);
     }
   }
+  return std::move(var_nodes);
+}
 
+void Graph::ResolveHazard(
+    const std::map<std::string, std::vector<ir::Node *>> &var_nodes) {
   /**
    * We should handle write after read(WAR) and write after write(WAW) here.
    * Because some of the operators of the program can be executed parallelly.
@@ -153,6 +155,7 @@ Graph::Graph(const ProgramDesc &program) : program_(program) {
     auto it_old = versions.rbegin();
     ++it_old;
     for (; it_old != versions.rend(); it_new = it_old, ++it_old) {
+      VLOG(3) << "deal with var: " << (*it_new)->Name();
       ir::Node *write_op =
           (*it_new)->inputs.empty() ? nullptr : (*it_new)->inputs[0];
       const auto &read_ops = (*it_old)->outputs;

paddle/fluid/framework/ir/graph.h

@@ -160,6 +160,12 @@ class Graph {
     return nullptr;
   }
 
+  std::map<std::string, std::vector<ir::Node *>> InitFromProgram(
+      const ProgramDesc &program);
+
+  void ResolveHazard(
+      const std::map<std::string, std::vector<ir::Node *>> &var_nodes);
+
  private:
   // This method takes ownership of `node`.
   ir::Node *AddNode(ir::Node *node) {

paddle/fluid/framework/ir/graph_helper.cc

@@ -120,19 +120,25 @@ size_t GraphNum(const Graph &graph) {
   std::deque<ir::Node *> q_nodes;
   std::vector<std::unordered_set<ir::Node *>> graph_nodes;
   std::unordered_set<ir::Node *> g_nodes;
+  // q_set used to record records in the queue.
+  std::unordered_set<ir::Node *> q_set;
   size_t graph_count = 0;
 
-  auto traverse_nodes = [&visited_nodes,
-                         &q_nodes](const std::vector<ir::Node *> &nodes) {
-    std::copy_if(
-        nodes.begin(), nodes.end(), std::back_inserter(q_nodes),
-        [&visited_nodes](Node *node) { return !visited_nodes.count(node); });
+  auto traverse_nodes = [&visited_nodes, &q_nodes,
+                         &q_set](const std::vector<ir::Node *> &nodes) {
+    for (auto n : nodes) {
+      if (visited_nodes.count(n) == 0 && q_set.count(n) == 0) {
+        q_nodes.push_back(n);
+        q_set.insert(n);
+      }
+    }
   };
 
   while (visited_nodes.size() != nodes.size()) {
     if (!q_nodes.empty()) {
       auto cur_node = q_nodes.front();
       q_nodes.pop_front();
+      q_set.erase(cur_node);
       visited_nodes.insert(cur_node);
       g_nodes.insert(cur_node);
       traverse_nodes(cur_node->inputs);
@@ -146,6 +152,7 @@ size_t GraphNum(const Graph &graph) {
       for (auto &n : nodes) {
         if (visited_nodes.count(n) == 0) {
           q_nodes.push_back(n);
+          q_set.insert(n);
           break;
         }
       }

paddle/fluid/framework/ir/graph_pattern_detector.cc

@@ -259,6 +259,15 @@ GraphPatternDetector::DetectPatterns() {
   return result;
 }
 
+bool GraphItemCMP(const std::pair<PDNode *, Node *> &a,
+                  const std::pair<PDNode *, Node *> &b) {
+  if (a.first != b.first) {
+    return a.first < b.first;
+  } else {
+    return a.second < b.second;
+  }
+}
+
 // TODO(Superjomn) enhance the function as it marks unique unique as duplicates
 // see https://github.com/PaddlePaddle/Paddle/issues/13550
 void GraphPatternDetector::UniquePatterns(
@@ -267,12 +276,16 @@ void GraphPatternDetector::UniquePatterns(
   std::vector<GraphPatternDetector::subgraph_t> result;
 
   std::unordered_set<size_t> set;
+  std::hash<std::string> hasher;
   for (auto &g : *subgraphs) {
-    size_t key = 0;
-    for (auto &item : g) {
-      key ^= std::hash<void *>{}(item.first);
-      key ^= std::hash<void *>{}(item.second);
+    // Sort the items in the sub-graph, and transform to a string key.
+    std::vector<std::pair<PDNode *, Node *>> sorted_keys(g.begin(), g.end());
+    std::sort(sorted_keys.begin(), sorted_keys.end(), GraphItemCMP);
+    std::stringstream ss;
+    for (auto &item : sorted_keys) {
+      ss << item.first << ":" << item.second;
     }
+    auto key = hasher(ss.str());
     if (!set.count(key)) {
       result.emplace_back(g);
       set.insert(key);

paddle/fluid/framework/ir/multi_batch_merge_pass.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 "paddle/fluid/framework/ir/multi_batch_merge_pass.h"
+
+#include <map>
+#include <string>
+#include <vector>
+
+#include "paddle/fluid/framework/ir/graph_helper.h"
+#include "paddle/fluid/framework/op_proto_maker.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+static const char kNumRepeats[] = "num_repeats";
+typedef std::unordered_map<std::string, std::vector<ir::Node*>> SSAVarList;
+
+ir::Node* SameNameVar(std::unordered_set<ir::Node*> all, ir::Node* target) {
+  for (auto n : all) {
+    if (target->IsVar() && target->Name() == n->Name()) {
+      return n;
+    }
+  }
+  return nullptr;
+}
+
+VarDesc CopyVarDesc(VarDesc* var_desc) {
+  VarDesc repeated_var(var_desc->Name());
+  // copy other variable attributes
+  if (var_desc->GetType() != proto::VarType::READER) {
+    repeated_var.SetType(var_desc->GetType());
+    repeated_var.SetShape(var_desc->GetShape());
+    repeated_var.SetDataType(var_desc->GetDataType());
+    repeated_var.SetLoDLevel(var_desc->GetLoDLevel());
+    repeated_var.SetPersistable(var_desc->Persistable());
+  } else {
+    // TODO(typhoonzero): copy reader var
+  }
+  return repeated_var;
+}
+
+VarDesc UpdateGradVarDesc(
+    VarDesc* var_desc, int repeat,
+    const std::unordered_set<std::string>& grad_names,
+    const std::unordered_set<std::string>& bn_vars_need_rename) {
+  if (grad_names.find(var_desc->Name()) != grad_names.end() ||
+      bn_vars_need_rename.find(var_desc->Name()) != bn_vars_need_rename.end()) {
+    std::string new_gname =
+        string::Sprintf("%s.repeat.%d", var_desc->Name(), repeat);
+    VarDesc repeated_var = CopyVarDesc(var_desc);
+    repeated_var.SetName(new_gname);
+    VLOG(3) << "update " << var_desc->Name() << " to repeat " << repeat;
+    return repeated_var;
+  }
+  return *var_desc;
+}
+
+std::unique_ptr<Graph> BatchMergePass::ApplyImpl(
+    std::unique_ptr<Graph> graph) const {
+  int num_repeats = Get<const int>(kNumRepeats);
+  std::vector<Node*> forward_backward_ops;
+  std::vector<Node*> optimize_ops;
+  std::vector<Node*> lr_ops;  // ops other than forward/backward/optimize
+  std::unordered_set<std::string> grad_names;
+
+  std::vector<ir::Node*> nodes = TopologySortOperations(*graph);
+  auto origin_nodes = graph->ReleaseNodes();
+  VLOG(3) << "origin nodes count: " << origin_nodes.size();
+  ir::Graph& result = *graph;
+
+  // 1. record op nodes of different roles
+  for (auto node : nodes) {
+    if (node->IsVar()) continue;
+    int op_role = boost::get<int>(node->Op()->GetAttr(
+        framework::OpProtoAndCheckerMaker::OpRoleAttrName()));
+    if ((op_role == static_cast<int>(framework::OpRole::kForward)) ||
+        (op_role & static_cast<int>(framework::OpRole::kBackward)) ||
+        (op_role & static_cast<int>(framework::OpRole::kLoss))) {
+      forward_backward_ops.push_back(node);
+    } else if ((op_role & static_cast<int>(framework::OpRole::kOptimize)) ||
+               (op_role & static_cast<int>(framework::OpRole::kDist)) ||
+               (op_role & static_cast<int>(framework::OpRole::kRPC))) {
+      optimize_ops.push_back(node);
+      auto op_role_var = node->Op()->GetNullableAttr(
+          OpProtoAndCheckerMaker::OpRoleVarAttrName());
+      auto op_role_vars = boost::get<std::vector<std::string>>(op_role_var);
+      for (size_t i = 0; i < op_role_vars.size(); i += 2) {
+        grad_names.insert(op_role_vars[i + 1]);
+      }
+    } else if (op_role & static_cast<int>(framework::OpRole::kLRSched)) {
+      lr_ops.push_back(node);
+    } else {  // NOLINT
+      PADDLE_THROW("Invalid op_role: %d", static_cast<int>(op_role));
+    }
+  }
+
+  // 2. copy forward backward
+  ir::Node* prev_repeat_last_op_node = nullptr;
+  // record origin_grad -> repeated grad list map.
+  std::map<ir::Node*, std::vector<ir::Node*>> grad_repeated_map;
+  std::map<std::string, std::vector<ir::Node*>> created;
+  std::unordered_set<std::string> bn_vars_need_rename;
+  for (int i = 0; i < num_repeats; ++i) {
+    std::unordered_set<ir::Node*> copied;
+    for (size_t node_idx = 0; node_idx < forward_backward_ops.size();
+         ++node_idx) {
+      auto node = forward_backward_ops[node_idx];
+      OpDesc repeated_op(*(node->Op()), node->Op()->Block());
+      // 3. rename grad outputs to current repeat.
+      for (auto outname : repeated_op.OutputArgumentNames()) {
+        if (grad_names.find(outname) != grad_names.end()) {
+          std::string new_gname = string::Sprintf("%s.repeat.%d", outname, i);
+          repeated_op.RenameOutput(outname, new_gname);
+        }
+      }
+      // 3.5 let batch_norm ops use independent vars, note batch_norm_grad do
+      // not need this update
+      if (node->Name() == "batch_norm") {
+        // NOTE: assume bn op created by layers use save var as output mean and
+        // variance
+        std::string new_mean_name =
+            string::Sprintf("%s.repeat.%d", repeated_op.Input("Mean")[0], i);
+        std::string new_var_name = string::Sprintf(
+            "%s.repeat.%d", repeated_op.Input("Variance")[0], i);
+        bn_vars_need_rename.insert(repeated_op.Input("Mean")[0]);
+        bn_vars_need_rename.insert(repeated_op.Input("Variance")[0]);
+        VLOG(3) << "renaming " << repeated_op.Input("Mean")[0] << " to "
+                << new_mean_name;
+        repeated_op.RenameInput(repeated_op.Input("Mean")[0], new_mean_name);
+        repeated_op.RenameInput(repeated_op.Input("Variance")[0], new_var_name);
+        repeated_op.RenameOutput(repeated_op.Output("MeanOut")[0],
+                                 new_mean_name);
+        repeated_op.RenameOutput(repeated_op.Output("VarianceOut")[0],
+                                 new_var_name);
+      }
+
+      // 3.9 do copy
+      auto repeated_node = result.CreateOpNode(&repeated_op);
+      copied.insert(node);
+
+      // 4. add deps between repeats
+      if (node_idx == forward_backward_ops.size() - 1) {
+        prev_repeat_last_op_node = repeated_node;
+      }
+      if (node_idx == 0 && prev_repeat_last_op_node) {
+        auto* depvar = result.CreateControlDepVar();
+        prev_repeat_last_op_node->outputs.push_back(depvar);
+        depvar->inputs.push_back(prev_repeat_last_op_node);
+        repeated_node->inputs.push_back(depvar);
+        depvar->outputs.push_back(repeated_node);
+      }
+
+      for (auto in_node : node->inputs) {
+        if (in_node->IsCtrlVar()) {
+          continue;
+        }
+        ir::Node* var = nullptr;
+        auto updated_var = UpdateGradVarDesc(in_node->Var(), i, grad_names,
+                                             bn_vars_need_rename);
+        // should be initialized by startup, how to initilize tensor in the
+        // scope?
+        if (node->Name() == "batch_norm" &&
+            bn_vars_need_rename.find(in_node->Name()) !=
+                bn_vars_need_rename.end()) {
+          // Create bn mean/variance for each repeat
+          var = result.CreateVarNode(&updated_var);
+          created[updated_var.Name()].push_back(var);
+          copied.insert(in_node);
+          repeated_node->inputs.push_back(var);
+          var->outputs.push_back(repeated_node);
+          continue;
+        }
+
+        // for other ops
+        if (in_node->inputs.empty() && i > 0) {
+          // do not copy head vars (inputs, params) in repeats > 0
+          var = created.at(in_node->Name()).back();
+        } else {
+          if (copied.find(in_node) == copied.end()) {
+            var = result.CreateVarNode(&updated_var);
+            if (grad_names.find(in_node->Var()->Name()) != grad_names.end()) {
+              grad_repeated_map[in_node].push_back(var);
+            }
+            copied.insert(in_node);
+            created[updated_var.Name()].push_back(var);
+          } else {
+            var = created.at(updated_var.Name()).back();
+          }
+        }
+        repeated_node->inputs.push_back(var);
+        var->outputs.push_back(repeated_node);
+      }
+      for (auto out_node : node->outputs) {
+        if (out_node->IsCtrlVar()) {
+          continue;
+        }
+        ir::Node* var = nullptr;
+        auto updated_var = UpdateGradVarDesc(out_node->Var(), i, grad_names,
+                                             bn_vars_need_rename);
+        if (copied.find(out_node) == copied.end()) {
+          var = result.CreateVarNode(&updated_var);
+          if (grad_names.find(out_node->Var()->Name()) != grad_names.end()) {
+            grad_repeated_map[out_node].push_back(var);
+          }
+          copied.insert(out_node);
+          created[updated_var.Name()].push_back(var);
+        } else {
+          var = created.at(updated_var.Name()).back();
+        }
+        repeated_node->outputs.push_back(var);
+        var->inputs.push_back(repeated_node);
+      }
+    }
+  }
+
+  // 5. create GRAD merge op node
+  for (auto kv : grad_repeated_map) {
+    OpDesc sum_op;
+    sum_op.SetType("sum");
+    std::vector<std::string> repeated_grad_names;
+    for (auto r : kv.second) {
+      repeated_grad_names.push_back(r->Var()->Name());
+    }
+    sum_op.SetInput("X", repeated_grad_names);
+    sum_op.SetOutput("Out", {kv.first->Var()->Name()});
+    sum_op.SetAttr(OpProtoAndCheckerMaker::OpRoleAttrName(),
+                   static_cast<int>(OpRole::kBackward));
+    auto sum_op_node = result.CreateOpNode(&sum_op);
+    for (auto r : kv.second) {
+      sum_op_node->inputs.push_back(r);
+      r->outputs.push_back(sum_op_node);
+    }
+    auto sum_out_var_node = result.CreateVarNode(kv.first->Var());
+    sum_op_node->outputs.push_back(sum_out_var_node);
+    sum_out_var_node->inputs.push_back(sum_op_node);
+    created[sum_out_var_node->Name()].push_back(sum_out_var_node);
+
+    OpDesc scale_op;
+    scale_op.SetType("scale");
+    scale_op.SetInput("X", {sum_out_var_node->Var()->Name()});
+    // NOTE: inplace scale.
+    scale_op.SetOutput("Out", {sum_out_var_node->Var()->Name()});
+    scale_op.SetAttr("scale", static_cast<float>(1.0f / num_repeats));
+    scale_op.SetAttr(OpProtoAndCheckerMaker::OpRoleAttrName(),
+                     static_cast<int>(OpRole::kBackward));
+    auto scale_op_node = result.CreateOpNode(&scale_op);
+    scale_op_node->inputs.push_back(sum_out_var_node);
+    sum_out_var_node->outputs.push_back(scale_op_node);
+    auto scale_out_var_node = result.CreateVarNode(sum_out_var_node->Var());
+    scale_op_node->outputs.push_back(scale_out_var_node);
+    scale_out_var_node->inputs.push_back(scale_op_node);
+    created[scale_out_var_node->Name()].push_back(scale_out_var_node);
+  }
+  // 6. add optimize ops
+  {
+    auto copy_node = [&result, &created](ir::Node* node) {
+      auto op_node = result.CreateOpNode(node->Op());
+      // copy op ins/outs
+      // NOTE: for send/recv ops, the OpDesc uses ctrldepvar to describe
+      // dependencies, so create those depvars if OpDesc have in/outs.
+      for (auto in_node : node->inputs) {
+        if (in_node->IsCtrlVar() && !in_node->Var()) {
+          continue;
+        }
+        ir::Node* var = nullptr;
+        if (created.find(in_node->Name()) == created.end()) {
+          var = result.CreateVarNode(in_node->Var());
+          created[in_node->Name()].push_back(var);
+        } else {
+          var = created.at(in_node->Name()).back();
+        }
+        op_node->inputs.push_back(var);
+        var->outputs.push_back(op_node);
+      }
+      for (auto out_node : node->outputs) {
+        if (out_node->IsCtrlVar() && !out_node->Var()) {
+          continue;
+        }
+        auto var = result.CreateVarNode(out_node->Var());
+        created[out_node->Name()].push_back(var);
+        op_node->outputs.push_back(var);
+        var->inputs.push_back(op_node);
+      }
+    };
+    for (auto node : lr_ops) {
+      copy_node(node);
+    }
+    for (auto node : optimize_ops) {
+      copy_node(node);
+    }
+  }
+
+  result.ResolveHazard(created);
+  return graph;
+}
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_PASS(multi_batch_merge_pass, paddle::framework::ir::BatchMergePass)
+    .RequirePassAttr(paddle::framework::ir::kNumRepeats);

paddle/fluid/framework/ir/multi_batch_merge_pass.h

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include "paddle/fluid/framework/ir/fuse_pass_base.h"
+#include "paddle/fluid/framework/ir/graph.h"
+#include "paddle/fluid/framework/ir/pass.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+// BatchMergePass is used to copy forward and backward ops for several
+// times to run several batches to simulate large batch size training
+// as if we have more than 1 GPUs.
+// User can define how many batches to run, gradients will be merged
+// through those repeats, and then do optimization using merged gradients.
+// This pass is extremely useful when doing large batch-size distributed
+// sync training, we can simulate even large batch size as if we have more
+// GPUs.
+
+class BatchMergePass : public Pass {
+ public:
+  virtual ~BatchMergePass() {}
+
+ protected:
+  std::unique_ptr<Graph> ApplyImpl(std::unique_ptr<Graph> graph) const override;
+};
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/ir/node.h

@@ -44,6 +44,7 @@ class Node {
     return op_desc_.get();
   }
 
+  // Please don't use this API!
   int id() const { return id_; }
 
   bool IsOp() const { return type_ == Type::kOperation; }
@@ -92,6 +93,7 @@ class Node {
   Node() = delete;
 
   static int count_;
+  // Please don't use this API or make this public.
   static void ResetId() { count_ = 0; }
   DISABLE_COPY_AND_ASSIGN(Node);
 };

paddle/fluid/framework/lod_tensor.cc

@@ -418,7 +418,7 @@ void LoDTensor::MergeLoDTensor(
     PADDLE_ENFORCE_EQ(new_lod.size(), lod.size());
     for (size_t j = 0; j < lod.size(); ++j) {
       auto &sub_lod = new_lod[j];
-      auto &offset = sub_lod.back();
+      size_t offset = sub_lod.back();
       for (size_t k = 1; k < lod[j].size(); ++k) {
         sub_lod.push_back(lod[j][k] + offset);
       }

paddle/fluid/framework/lod_tensor_array.h

@@ -18,6 +18,8 @@ limitations under the License. */
 
 namespace paddle {
 namespace framework {
+
 using LoDTensorArray = std::vector<LoDTensor>;
-}
+
+}  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/mixed_vector.h

@@ -542,6 +542,33 @@ class CPUVector : public std::vector<T, std::allocator<T>> {
     this->reserve(this->size() + size_t(end - begin));
     this->insert(this->end(), begin, end);
   }
+
+  const T *CUDAData(platform::Place place) const {
+    PADDLE_THROW(
+        "Vector::CUDAData() method is not supported in CPU-only version");
+  }
+
+  T *CUDAMutableData(platform::Place place) {
+    PADDLE_THROW(
+        "Vector::CUDAMutableData() method is not supported in CPU-only "
+        "version");
+  }
+
+  const T *Data(platform::Place place) const {
+    PADDLE_ENFORCE(
+        platform::is_cpu_place(place),
+        "Vector::Data() method is not supported when not in CPUPlace");
+    return this->data();
+  }
+
+  T *MutableData(platform::Place place) {
+    PADDLE_ENFORCE(
+        platform::is_cpu_place(place),
+        "Vector::MutableData() method is not supported when not in CPUPlace");
+    return this->data();
+  }
+
+  const void *Handle() const { return static_cast<const void *>(this); }
 };
 
 template <typename T>

paddle/fluid/framework/naive_executor.cc

@@ -146,22 +146,5 @@ void NaiveExecutor::CleanFeedFetchOps() {
   ops_.swap(ops);
 }
 
-void NaiveExecutor::EnableMKLDNN(const ProgramDesc &program) {
-#ifdef PADDLE_WITH_MKLDNN
-  VLOG(3) << "use_mkldnn=True";
-  for (size_t block_id = 0; block_id < program.Size(); ++block_id) {
-    auto *block = const_cast<ProgramDesc &>(program).MutableBlock(block_id);
-    for (auto *op : block->AllOps()) {
-      if (op->HasAttr("use_mkldnn")) {
-        op->SetAttr("use_mkldnn", true);
-      }
-    }
-  }
-#else
-  LOG(WARNING)
-      << "'MKLDNN' is not supported, Please re-compile with WITH_MKLDNN option";
-#endif
-}
-
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/naive_executor.h

@@ -48,8 +48,6 @@ class NaiveExecutor {
 
   void CleanFeedFetchOps();
 
-  void EnableMKLDNN(const ProgramDesc& program);
-
  protected:
   void CreateVariables(const ProgramDesc& desc, Scope* scope, int block_id);
 

paddle/fluid/framework/op_desc.cc

@@ -419,8 +419,15 @@ struct SetAttrDescVisitor : public boost::static_visitor<void> {
     }
     VectorToRepeated(blocks_idx, attr_->mutable_blocks_idx());
   }
+
   void operator()(BlockDesc *desc) const { attr_->set_block_idx(desc->ID()); }
+
   void operator()(int64_t v) const { attr_->set_l(v); }
+
+  void operator()(const std::vector<int64_t> &v) const {
+    VectorToRepeated(v, attr_->mutable_longs());
+  }
+
   void operator()(boost::blank) const { PADDLE_THROW("Unexpected branch"); }
 };
 

paddle/fluid/framework/op_desc.h

@@ -121,10 +121,6 @@ class OpDesc {
 
   BlockDesc *Block() { return this->block_; }
 
-  const BlockDesc &BlockRef() const { return *this->block_; }
-
-  void SetBlock(BlockDesc *block) { this->block_ = block; }
-
  private:
   template <typename MapType>
   static std::vector<typename MapType::key_type> MapKeys(const MapType &map) {

paddle/fluid/framework/op_proto_maker.h

@@ -28,12 +28,12 @@ enum class OpRole {
   kBackward = 0x0001,
   kOptimize = 0x0002,
   // RPC role is for send/recv releated op
-  kRPC = 0x0003,
+  kRPC = 0x0004,
   // Dist role is for split_byref/split_selected_rows/concat
   // used for distributed training.
-  kDist = 0x0004,
+  kDist = 0x0008,
   // Tag all learning rate scheduler operators.
-  kLRSched = 0x0005,
+  kLRSched = 0x0010,
 
   kLoss = 0x0100,
   // The default value of op's role. This should be only used for unittests and

paddle/fluid/framework/operator.cc

@@ -354,18 +354,18 @@ void OperatorBase::GenerateTemporaryNames() {
   }
 }
 
-static bool VarIsTensor(const Variable* var) {
-  return var->IsType<LoDTensor>() || var->IsType<SelectedRows>();
+static bool VarIsTensor(const Variable& var) {
+  return var.IsType<LoDTensor>() || var.IsType<SelectedRows>();
 }
 
-static const Tensor* GetTensorFromVar(Variable* var) {
-  if (var->IsType<LoDTensor>()) {
-    return var->GetMutable<LoDTensor>();
-  } else if (var->IsType<SelectedRows>()) {
-    return var->GetMutable<SelectedRows>()->mutable_value();
+const Tensor* GetTensorFromVar(const Variable& var) {
+  if (var.IsType<LoDTensor>()) {
+    return static_cast<const Tensor*>(&(var.Get<LoDTensor>()));
+  } else if (var.IsType<SelectedRows>()) {
+    return &(var.Get<SelectedRows>().value());
   } else {
     PADDLE_THROW("Variable type_id %s, expect LoDTensor/SelectedRows.",
-                 var->Type().name());
+                 var.Type().name());
   }
 }
 
@@ -415,8 +415,7 @@ bool ExecutionContext::HasOutput(const std::string& name) const {
 template <>
 const Tensor* ExecutionContext::Input<Tensor>(const std::string& name) const {
   auto* var = InputVar(name);
-  return var == nullptr ? nullptr
-                        : GetTensorFromVar(const_cast<Variable*>(var));
+  return var == nullptr ? nullptr : GetTensorFromVar(*var);
 }
 
 template <>
@@ -428,7 +427,7 @@ const std::vector<const Tensor*> ExecutionContext::MultiInput<Tensor>(
   std::transform(names.begin(), names.end(), std::back_inserter(res),
                  [&](const std::string& sub_name) {
                    auto var = scope_.FindVar(sub_name);
-                   return var == nullptr ? nullptr : GetTensorFromVar(var);
+                   return var == nullptr ? nullptr : GetTensorFromVar(*var);
                  });
   return res;
 }
@@ -770,8 +769,10 @@ void OperatorWithKernel::TransferInplaceVarsBack(
   for (auto& var_name : inplace_vars) {
     VLOG(3) << "share inplace var " + var_name + " back to it's original scope";
     auto* original_tensor = GetMutableTensorFromVar(scope.FindVar(var_name));
-    auto* transformed_tensor =
-        GetTensorFromVar(transfer_scope.FindVar(var_name));
+    auto* var = transfer_scope.FindVar(var_name);
+    PADDLE_ENFORCE(var != nullptr, "The var[%s] should not be nullptr",
+                   var_name);
+    auto* transformed_tensor = GetTensorFromVar(*var);
     original_tensor->ShareDataWith(*transformed_tensor);
   }
 }
@@ -784,11 +785,11 @@ Scope* OperatorWithKernel::TryTransferData(
     for (auto& var_name : var_name_item.second) {
       auto* var = scope.FindVar(var_name);
       // Only tensor can be tranfer to another device.
-      if (var == nullptr || !VarIsTensor(var)) {
+      if (var == nullptr || !VarIsTensor(*var)) {
         continue;
       }
 
-      auto* tensor_in = GetTensorFromVar(var);
+      auto* tensor_in = GetTensorFromVar(*var);
       if (!tensor_in->IsInitialized()) {
         continue;
       }

paddle/fluid/framework/operator.h

@@ -63,6 +63,7 @@ inline std::string GradVarName(const std::string& var_name) {
 }
 
 proto::VarType::Type GetDataTypeOfVar(const Variable* var);
+const Tensor* GetTensorFromVar(const Variable& var);
 
 class OperatorBase;
 class ExecutionContext;

paddle/fluid/framework/parallel_executor.cc

@@ -109,18 +109,9 @@ ParallelExecutor::ParallelExecutor(
   if (member_->local_scopes_.size() != 1 && local_scopes.empty()) {
     BCastParamsToDevices(bcast_vars);
   }
-  // Startup Program has been run. All local scopes has correct parameters.
+// Startup Program has been run. All local scopes has correct parameters.
 
-  // Step 2. Create vars in each scope;
-  std::vector<details::VariableInfo> var_infos;
-  for (auto *var : main_program.Block(0).AllVars()) {
-    var_infos.emplace_back();
-    var_infos.back().name_ = var->Name();
-    var_infos.back().type_ = var->GetType();
-    var_infos.back().persistable_ = var->Persistable();
-  }
-
-// Step 3. Convert main_program to SSA form and dependency graph. Also, insert
+// Step 2. Convert main_program to SSA form and dependency graph. Also, insert
 // ncclOp
 #ifdef PADDLE_WITH_CUDA
   std::unique_ptr<ir::Graph> graph = build_strategy.Apply(
@@ -156,6 +147,23 @@ ParallelExecutor::ParallelExecutor(
                            params, member_->local_scopes_, member_->use_cuda_);
 #endif
 
+  // Step 3. Create vars in each scope. Passes may also create new vars.
+  //         skip control vars and empty vars
+  std::vector<details::VariableInfo> var_infos;
+  for (auto &node : graph->Nodes()) {
+    if (node->IsVar() && !node->IsCtrlVar() && node->Var()) {
+      var_infos.emplace_back();
+      var_infos.back().name_ = node->Var()->Name();
+      var_infos.back().type_ = node->Var()->GetType();
+      var_infos.back().persistable_ = node->Var()->Persistable();
+    }
+  }
+  // If the loss_var_name is given, the number of graph should be only one.
+  if (loss_var_name.size()) {
+    PADDLE_ENFORCE_EQ(ir::GraphNum(*graph), 1,
+                      "The number of graph should be only one");
+  }
+
   if (exec_strategy.type_ == ExecutionStrategy::kDefault) {
     member_->executor_.reset(new details::ThreadedSSAGraphExecutor(
         exec_strategy, member_->local_scopes_, places, std::move(graph)));
@@ -179,6 +187,10 @@ void ParallelExecutor::BCastParamsToDevices(
     }
 
     auto &main_tensor = main_var->Get<LoDTensor>();
+    if (!main_tensor.IsInitialized()) {
+      VLOG(3) << "one in var not inited, return!";
+      continue;
+    }
     auto &dims = main_tensor.dims();
     if (paddle::platform::is_gpu_place(main_tensor.place())) {
 #ifdef PADDLE_WITH_CUDA
@@ -291,10 +303,8 @@ void ParallelExecutor::FeedAndSplitTensorIntoLocalScopes(
 }
 
 ParallelExecutor::~ParallelExecutor() {
-  const auto dev_ctxs =
-      platform::DeviceContextPool::Instance().GetAllDeviceContexts();
-  for (auto &dev_ctx : dev_ctxs) {
-    dev_ctx->Wait();
+  for (auto &p : member_->places_) {
+    platform::DeviceContextPool::Instance().Get(p)->Wait();
   }
 
   if (member_->own_local_scope_) {

paddle/fluid/framework/scope.h

@@ -78,6 +78,8 @@ class Scope {
   /// Drop all kids scopes belonged to this scope.
   void DropKids();
 
+  std::list<Scope*>& kids() const { return kids_; }
+
   /// Find if a scope exists in the kid scopes
   bool HasKid(const Scope* scope) const;
 

paddle/fluid/framework/tensor_test.cc

@@ -75,6 +75,19 @@ TEST(Tensor, MutableData) {
                                         platform::CPUPlace());
     EXPECT_EQ(p1, p2);
   }
+  // Not sure if it's desired, but currently, Tensor type can be changed.
+  {
+    framework::Tensor src_tensor;
+    int8_t* p1 = src_tensor.mutable_data<int8_t>(framework::make_ddim({1}),
+                                                 platform::CPUPlace());
+    EXPECT_NE(p1, nullptr);
+    *p1 = 1;
+
+    uint8_t* p2 = src_tensor.mutable_data<uint8_t>(framework::make_ddim({1}),
+                                                   platform::CPUPlace());
+    EXPECT_NE(p2, nullptr);
+    EXPECT_EQ(static_cast<int>(p2[0]), 1);
+  }
 
 #ifdef PADDLE_WITH_CUDA
   {

paddle/fluid/framework/tensor_util.cc

@@ -153,6 +153,12 @@ void TensorCopySync(const Tensor& src, const platform::Place& dst_place,
     auto src_gpu_place = boost::get<platform::CUDAPlace>(src_place);
     auto dst_gpu_place = boost::get<platform::CUDAPlace>(dst_place);
     memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size, nullptr);
+  } else if (platform::is_cuda_pinned_place(src_place) &&
+             platform::is_gpu_place(dst_place)) {
+    auto src_pinned_place = boost::get<platform::CUDAPinnedPlace>(src_place);
+    auto dst_gpu_place = boost::get<platform::CUDAPlace>(dst_place);
+    memory::Copy(dst_gpu_place, dst_ptr, src_pinned_place, src_ptr, size,
+                 nullptr);
   }
 #endif
 }

paddle/fluid/framework/threadpool.cc

@@ -25,7 +25,6 @@ DEFINE_int32(dist_threadpool_size, 0,
 
 namespace paddle {
 namespace framework {
-
 std::unique_ptr<ThreadPool> ThreadPool::threadpool_(nullptr);
 std::once_flag ThreadPool::init_flag_;
 
@@ -47,8 +46,7 @@ void ThreadPool::Init() {
   }
 }
 
-ThreadPool::ThreadPool(int num_threads)
-    : total_threads_(num_threads), idle_threads_(num_threads), running_(true) {
+ThreadPool::ThreadPool(int num_threads) : running_(true) {
   threads_.resize(num_threads);
   for (auto& thread : threads_) {
     // TODO(Yancey1989): binding the thread on the specify CPU number
@@ -59,6 +57,7 @@ ThreadPool::ThreadPool(int num_threads)
 ThreadPool::~ThreadPool() {
   {
     // notify all threads to stop running
+    std::lock_guard<std::mutex> l(mutex_);
     running_ = false;
     scheduled_.notify_all();
   }
@@ -69,36 +68,24 @@ ThreadPool::~ThreadPool() {
   }
 }
 
-void ThreadPool::Wait() {
-  std::unique_lock<std::mutex> lock(mutex_);
-  completed_.wait(lock, [=] { return Done() == true; });
-}
-
 void ThreadPool::TaskLoop() {
-  while (running_) {
+  while (true) {
     std::unique_lock<std::mutex> lock(mutex_);
-    scheduled_.wait(lock, [=] { return !tasks_.empty() || !running_; });
 
-    if (!running_) {
-      break;
+    scheduled_.wait(
+        lock, [this] { return !this->tasks_.empty() || !this->running_; });
+
+    if (!running_ || tasks_.empty()) {
+      return;
     }
+
     // pop a task from the task queue
     auto task = std::move(tasks_.front());
     tasks_.pop();
-
-    --idle_threads_;
     lock.unlock();
 
     // run the task
     task();
-
-    {
-      std::unique_lock<std::mutex> lock(mutex_);
-      ++idle_threads_;
-      if (Done()) {
-        completed_.notify_all();
-      }
-    }
   }
 }
 

paddle/fluid/framework/threadpool.h

@@ -57,15 +57,6 @@ class ThreadPool {
 
   ~ThreadPool();
 
-  // Returns the number of threads created by the constructor.
-  size_t Threads() const { return total_threads_; }
-
-  // Returns the number of currently idle threads.
-  size_t IdleThreads() {
-    std::unique_lock<std::mutex> lock(mutex_);
-    return idle_threads_;
-  }
-
   // Run pushes a function to the task queue and returns a std::future
   // object.  To wait for the completion of the task, call
   // std::future::wait().
@@ -94,25 +85,13 @@ class ThreadPool {
     });
     std::future<std::unique_ptr<platform::EnforceNotMet>> f = task.get_future();
     tasks_.push(std::move(task));
-    lock.unlock();
     scheduled_.notify_one();
     return f;
   }
 
-  // Wait until all the tasks are completed.
-  void Wait();
-
  private:
   DISABLE_COPY_AND_ASSIGN(ThreadPool);
 
-  // If the task queue is empty and avaialbe is equal to the number of
-  // threads, means that all tasks are completed.  Note: this function
-  // is not thread-safe.  Returns true if all tasks are completed.
-  // Note: don't delete the data member total_threads_ and use
-  // threads_.size() instead; because you'd need to lock the mutex
-  // before accessing threads_.
-  bool Done() { return tasks_.empty() && idle_threads_ == total_threads_; }
-
   // The constructor starts threads to run TaskLoop, which retrieves
   // and runs tasks from the queue.
   void TaskLoop();
@@ -125,14 +104,11 @@ class ThreadPool {
   static std::once_flag init_flag_;
 
   std::vector<std::unique_ptr<std::thread>> threads_;
-  const size_t total_threads_;
-  size_t idle_threads_;
 
   std::queue<Task> tasks_;
   std::mutex mutex_;
   bool running_;
   std::condition_variable scheduled_;
-  std::condition_variable completed_;
 };
 
 class ThreadPoolIO : ThreadPool {

paddle/fluid/framework/threadpool_test.cc

@@ -19,10 +19,11 @@ limitations under the License. */
 
 namespace framework = paddle::framework;
 
-void do_sum(framework::ThreadPool* pool, std::atomic<int>* sum, int cnt) {
-  std::vector<std::future<void>> fs;
+void do_sum(std::vector<std::future<void>>* fs, std::mutex* mu,
+            std::atomic<int>* sum, int cnt) {
   for (int i = 0; i < cnt; ++i) {
-    fs.push_back(framework::Async([sum]() { sum->fetch_add(1); }));
+    std::lock_guard<std::mutex> l(*mu);
+    fs->push_back(framework::Async([sum]() { sum->fetch_add(1); }));
   }
 }
 
@@ -40,18 +41,21 @@ TEST(ThreadPool, ConcurrentInit) {
 }
 
 TEST(ThreadPool, ConcurrentRun) {
-  framework::ThreadPool* pool = framework::ThreadPool::GetInstance();
   std::atomic<int> sum(0);
   std::vector<std::thread> threads;
+  std::vector<std::future<void>> fs;
+  std::mutex fs_mu;
   int n = 50;
   // sum = (n * (n + 1)) / 2
   for (int i = 1; i <= n; ++i) {
-    std::thread t(do_sum, pool, &sum, i);
+    std::thread t(do_sum, &fs, &fs_mu, &sum, i);
     threads.push_back(std::move(t));
   }
   for (auto& t : threads) {
     t.join();
   }
-  pool->Wait();
+  for (auto& t : fs) {
+    t.wait();
+  }
   EXPECT_EQ(sum, ((n + 1) * n) / 2);
 }

paddle/fluid/framework/type_defs.h

@@ -36,7 +36,7 @@ using Attribute =
     boost::variant<boost::blank, int, float, std::string, std::vector<int>,
                    std::vector<float>, std::vector<std::string>, bool,
                    std::vector<bool>, BlockDesc*, int64_t,
-                   std::vector<BlockDesc*>>;
+                   std::vector<BlockDesc*>, std::vector<int64_t>>;
 
 using AttributeMap = std::unordered_map<std::string, Attribute>;
 

paddle/fluid/inference/CMakeLists.txt

+if(WITH_TESTING)
+  include(test.cmake) # some generic cmake funtion for inference
+endif()
 # analysis and tensorrt must be added before creating static library,
 # otherwise, there would be undefined reference to them in static library.
 add_subdirectory(analysis)
@@ -30,7 +33,7 @@ if (WITH_GPU AND TENSORRT_FOUND)
 endif()
 
 # Create static library
-cc_library(paddle_fluid DEPS ${fluid_modules} ${STATIC_INFERENCE_APIS} zero_copy_tensor)
+cc_library(paddle_fluid DEPS ${fluid_modules} ${STATIC_INFERENCE_APIS} zero_copy_tensor reset_tensor_array)
 
 if(NOT APPLE)
   # TODO(liuyiqu: Temporarily disable the link flag because it is not support on Mac.
@@ -40,7 +43,7 @@ endif()
 
 # Create shared library
 cc_library(paddle_fluid_shared SHARED SRCS ${SHARED_INFERENCE_SRCS}
-    DEPS ${fluid_modules} paddle_fluid_api)
+    DEPS ${fluid_modules} paddle_fluid_api reset_tensor_array)
 
 set_target_properties(paddle_fluid_shared PROPERTIES OUTPUT_NAME paddle_fluid)
 if(NOT APPLE)

paddle/fluid/inference/analysis/CMakeLists.txt

@@ -20,22 +20,17 @@ cc_test(test_node SRCS node_tester.cc DEPS analysis)
 cc_test(test_dot SRCS dot_tester.cc DEPS analysis)
 cc_binary(inference_analyzer SRCS analyzer_main.cc DEPS analysis paddle_fluid)
 
-function (inference_analysis_test TARGET)
-    if(WITH_TESTING)
-        set(options "")
-        set(oneValueArgs "")
-        set(multiValueArgs SRCS ARGS EXTRA_DEPS)
-        cmake_parse_arguments(analysis_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
-        set(mem_opt "")
-        if(WITH_GPU)
-            set(mem_opt "--fraction_of_gpu_memory_to_use=0.5")
-        endif()
-        cc_test(${TARGET}
-                SRCS "${analysis_test_SRCS}"
-                DEPS analysis pass ${GLOB_PASS_LIB} ${analysis_test_EXTRA_DEPS}
-                ARGS --inference_model_dir=${PYTHON_TESTS_DIR}/book/word2vec.inference.model ${mem_opt} ${analysis_test_ARGS})
-        set_tests_properties(${TARGET} PROPERTIES DEPENDS test_word2vec)
-    endif(WITH_TESTING)
+function(inference_analysis_test TARGET)
+  if(WITH_TESTING)
+     set(options "")
+     set(oneValueArgs "")
+     set(multiValueArgs SRCS ARGS EXTRA_DEPS)
+     cmake_parse_arguments(analysis_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+     inference_base_test(${TARGET}
+             SRCS ${analysis_test_SRCS}
+             DEPS analysis pass ${GLOB_PASS_LIB} ${analysis_test_EXTRA_DEPS}
+             ARGS --inference_model_dir=${WORD2VEC_MODEL_DIR} ${analysis_test_ARGS})
+  endif()
 endfunction(inference_analysis_test)
 
 inference_analysis_test(test_analyzer SRCS analyzer_tester.cc EXTRA_DEPS paddle_inference_api)

paddle/fluid/inference/analysis/analyzer.cc

@@ -107,6 +107,9 @@ void Analyzer::Run(Argument* argument) {
     passes.push_back("mkldnn_placement_pass");
   }
 #endif
+  // infer_clean_graph_pass should be the first default pass
+  // after mkldnn_placement_pass.
+  passes.push_back("infer_clean_graph_pass");
   for (auto& pass : ir_passes_) {
     if (!disabled_ir_passes_.count(pass)) {
       passes.push_back(pass);

paddle/fluid/inference/analysis/analyzer.h

@@ -67,7 +67,6 @@ class Analyzer : public OrderedRegistry<PassManager> {
   // larger fusion.
   const std::vector<std::string> all_ir_passes_{{
       // Manual update the passes here.
-      "infer_clean_graph_pass",         //
       "attention_lstm_fuse_pass",       //
       "seqconv_eltadd_relu_fuse_pass",  //
       "embedding_fc_lstm_fuse_pass",    //
@@ -80,6 +79,7 @@ class Analyzer : public OrderedRegistry<PassManager> {
       "conv_bn_fuse_pass",              //
       "conv_eltwiseadd_bn_fuse_pass",   //
 #ifdef PADDLE_WITH_MKLDNN
+      "depthwise_conv_mkldnn_pass",             //
       "conv_bias_mkldnn_fuse_pass",             //
       "conv_relu_mkldnn_fuse_pass",             //
       "conv_elementwise_add_mkldnn_fuse_pass",  //

paddle/fluid/inference/analysis/data_flow_graph_tester.cc

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/inference/analysis/data_flow_graph.h"
+#include "paddle/fluid/framework/op_proto_maker.h"
 #include "paddle/fluid/framework/program_desc.h"
 #include "paddle/fluid/inference/analysis/ut_helper.h"
 
@@ -130,6 +131,8 @@ void SetOp(framework::ProgramDesc* prog, const std::string& type,
   op->SetType(type);
   op->SetInput("Xs", inputs);
   op->SetOutput("Xs", outputs);
+  op->SetAttr(framework::OpProtoAndCheckerMaker::OpRoleAttrName(),
+              static_cast<int>(framework::OpRole::kForward));
 }
 
 TEST(DataFlowGraph, Build_IR_Graph) {

paddle/fluid/inference/api/CMakeLists.txt

@@ -17,32 +17,14 @@ if(APPLE)
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-error=pessimizing-move")
 endif(APPLE)
 
-
 set(inference_deps paddle_inference_api paddle_fluid_api analysis pass ir_pass_manager naive_executor ${GLOB_PASS_LIB})
 
 if(WITH_GPU AND TENSORRT_FOUND)
     set(inference_deps ${inference_deps} paddle_inference_tensorrt_subgraph_engine analysis_predictor)
 endif()
 
-function(inference_api_test TARGET_NAME)
-    if (WITH_TESTING)
-        set(options "")
-        set(oneValueArgs SRC)
-        set(multiValueArgs ARGS)
-        cmake_parse_arguments(inference_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
-
-        cc_test(${TARGET_NAME}
-                SRCS ${inference_test_SRC}
-                DEPS "${inference_deps}"
-                ARGS --dirname=${PYTHON_TESTS_DIR}/book/)
-        if(inference_test_ARGS)
-            set_tests_properties(${TARGET_NAME}
-                    PROPERTIES DEPENDS "${inference_test_ARGS}")
-        endif()
-    endif(WITH_TESTING)
-endfunction(inference_api_test)
-
-cc_library(paddle_inference_api SRCS api.cc api_impl.cc helper.cc DEPS lod_tensor scope)
+cc_library(reset_tensor_array SRCS details/reset_tensor_array.cc DEPS lod_tensor scope)
+cc_library(paddle_inference_api SRCS api.cc api_impl.cc helper.cc DEPS reset_tensor_array lod_tensor scope)
 cc_library(analysis_predictor SRCS analysis_predictor.cc DEPS paddle_inference_api analysis naive_executor zero_copy_tensor)
 cc_library(zero_copy_tensor SRCS details/zero_copy_tensor.cc DEPS paddle_inference_api)
 cc_library(zero_copy_tensor_dummy SRCS details/zero_copy_tensor_dummy.cc DEPS paddle_inference_api)
@@ -50,10 +32,11 @@ cc_test(test_paddle_inference_api
         SRCS api_tester.cc
         DEPS paddle_inference_api)
 
-inference_api_test(test_api_impl SRC api_impl_tester.cc
-                    ARGS test_word2vec test_image_classification)
-
-set(PYTHON_TESTS_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/tests)
+if(WITH_TESTING)
+  inference_base_test(test_api_impl SRCS api_impl_tester.cc DEPS ${inference_deps}
+                      ARGS --word2vec_dirname=${WORD2VEC_MODEL_DIR} --book_dirname=${PYTHON_TESTS_DIR}/book)
+  set_tests_properties(test_api_impl PROPERTIES DEPENDS test_image_classification)
+endif()
 cc_test(test_analysis_predictor SRCS analysis_predictor_tester.cc DEPS analysis_predictor ${inference_deps} paddle_inference_api
         ARGS --dirname=${PYTHON_TESTS_DIR}/book)
 
@@ -61,8 +44,10 @@ if(WITH_GPU AND TENSORRT_FOUND)
 cc_library(paddle_inference_tensorrt_subgraph_engine
         SRCS api_tensorrt_subgraph_engine.cc
         DEPS paddle_inference_api analysis tensorrt_engine paddle_inference_api paddle_fluid_api tensorrt_converter zero_copy_tensor_dummy)
-
-inference_api_test(test_api_tensorrt_subgraph_engine SRC api_tensorrt_subgraph_engine_tester.cc ARGS test_word2vec)
+  if(WITH_TESTING)
+    inference_base_test(test_api_tensorrt_subgraph_engine SRCS api_tensorrt_subgraph_engine_tester.cc DEPS ${inference_deps}
+                      ARGS --dirname=${WORD2VEC_MODEL_DIR})
+  endif()
 endif()
 
 if (WITH_ANAKIN AND WITH_MKL) # only needed in CI

paddle/fluid/inference/api/analysis_predictor.cc

@@ -82,6 +82,7 @@ bool AnalysisPredictor::Init(
 
   // Get the feed_target_names and fetch_target_names
   PrepareFeedFetch();
+
   return true;
 }
 
@@ -109,6 +110,10 @@ bool AnalysisPredictor::Run(const std::vector<PaddleTensor> &inputs,
     return false;
   }
   VLOG(3) << "predict cost: " << timer.toc() << "ms";
+
+  // Fix TensorArray reuse not cleaned bug.
+  tensor_array_batch_cleaner_.CollectTensorArrays(scope_.get());
+  tensor_array_batch_cleaner_.ResetTensorArray();
   return true;
 }
 
@@ -322,6 +327,9 @@ std::unique_ptr<ZeroCopyTensor> AnalysisPredictor::GetOutputTensor(
 
 bool AnalysisPredictor::ZeroCopyRun() {
   executor_->Run();
+  // Fix TensorArray reuse not cleaned bug.
+  tensor_array_batch_cleaner_.CollectTensorArrays(scope_.get());
+  tensor_array_batch_cleaner_.ResetTensorArray();
   return true;
 }
 

paddle/fluid/inference/api/analysis_predictor.h

@@ -18,6 +18,7 @@
 #include "paddle/fluid/framework/naive_executor.h"
 #include "paddle/fluid/inference/analysis/analyzer.h"
 #include "paddle/fluid/inference/api/api_impl.h"
+#include "paddle/fluid/inference/api/details/reset_tensor_array.h"
 #include "paddle/fluid/inference/api/paddle_inference_api.h"
 #include "paddle/fluid/string/printf.h"
 
@@ -88,6 +89,7 @@ class AnalysisPredictor : public PaddlePredictor {
   // Memory buffer for feed inputs. The temporary LoDTensor will cause serious
   // concurrency problems, so cache them.
   std::vector<framework::LoDTensor> feed_tensors_;
+  details::TensorArrayBatchCleaner tensor_array_batch_cleaner_;
 };
 
 }  // namespace paddle

paddle/fluid/inference/api/api_impl.cc

@@ -22,6 +22,7 @@ limitations under the License. */
 
 #include "paddle/fluid/framework/feed_fetch_method.h"
 #include "paddle/fluid/inference/api/api_impl.h"
+#include "paddle/fluid/inference/api/details/reset_tensor_array.h"
 #include "paddle/fluid/inference/api/helper.h"
 #include "paddle/fluid/platform/cpu_helper.h"
 #include "paddle/fluid/platform/profiler.h"
@@ -157,6 +158,10 @@ bool NativePaddlePredictor::Run(const std::vector<PaddleTensor> &inputs,
     return false;
   }
   VLOG(3) << "predict cost: " << timer.toc() << "ms";
+
+  // Fix TensorArray reuse not cleaned bug.
+  tensor_array_batch_cleaner_.CollectTensorArrays(scope_.get());
+  tensor_array_batch_cleaner_.ResetTensorArray();
   return true;
 }
 

paddle/fluid/inference/api/api_impl.h

@@ -26,11 +26,11 @@ limitations under the License. */
 #include <string>
 #include <vector>
 
-#include "paddle/fluid/inference/api/paddle_inference_api.h"
-
 #include "paddle/fluid/framework/ddim.h"
 #include "paddle/fluid/framework/lod_tensor.h"
+#include "paddle/fluid/framework/lod_tensor_array.h"
 #include "paddle/fluid/framework/naive_executor.h"
+#include "paddle/fluid/inference/api/details/reset_tensor_array.h"
 #include "paddle/fluid/inference/api/paddle_inference_api.h"
 #include "paddle/fluid/inference/io.h"
 #include "paddle/fluid/platform/init.h"
@@ -77,6 +77,7 @@ class NativePaddlePredictor : public PaddlePredictor {
   std::vector<framework::OpDesc *> fetchs_;
   // Do not use unique_ptr, use parent scope to delete
   framework::Scope *sub_scope_{nullptr};
+  details::TensorArrayBatchCleaner tensor_array_batch_cleaner_;
 };
 
 }  // namespace paddle

paddle/fluid/inference/api/api_impl_tester.cc

@@ -27,7 +27,9 @@ limitations under the License. */
 #define ACC_DIFF 1e-3
 #endif
 
-DEFINE_string(dirname, "", "Directory of the inference model.");
+DEFINE_string(word2vec_dirname, "",
+              "Directory of the word2vec inference model.");
+DEFINE_string(book_dirname, "", "Directory of the book inference model.");
 
 namespace paddle {
 
@@ -49,7 +51,7 @@ PaddleTensor LodTensorToPaddleTensor(framework::LoDTensor* t) {
 
 NativeConfig GetConfig() {
   NativeConfig config;
-  config.model_dir = FLAGS_dirname + "/word2vec.inference.model";
+  config.model_dir = FLAGS_word2vec_dirname;
   LOG(INFO) << "dirname  " << config.model_dir;
   config.fraction_of_gpu_memory = 0.15;
 #ifdef PADDLE_WITH_CUDA
@@ -116,7 +118,7 @@ void MainImageClassification(bool use_gpu) {
   NativeConfig config = GetConfig();
   config.use_gpu = use_gpu;
   config.model_dir =
-      FLAGS_dirname + "/image_classification_resnet.inference.model";
+      FLAGS_book_dirname + "/image_classification_resnet.inference.model";
 
   const bool is_combined = false;
   std::vector<std::vector<int64_t>> feed_target_shapes =
@@ -187,7 +189,7 @@ void MainThreadsWord2Vec(bool use_gpu) {
   std::vector<std::thread> threads;
   for (int tid = 0; tid < num_jobs; ++tid) {
     threads.emplace_back([&, tid]() {
-      auto predictor = main_predictor->Clone();
+      auto predictor = CreatePaddlePredictor(config);
       auto& local_inputs = paddle_tensor_feeds[tid];
       std::vector<PaddleTensor> local_outputs;
       ASSERT_TRUE(predictor->Run(local_inputs, &local_outputs));
@@ -220,7 +222,7 @@ void MainThreadsImageClassification(bool use_gpu) {
   NativeConfig config = GetConfig();
   config.use_gpu = use_gpu;
   config.model_dir =
-      FLAGS_dirname + "/image_classification_resnet.inference.model";
+      FLAGS_book_dirname + "/image_classification_resnet.inference.model";
 
   auto main_predictor = CreatePaddlePredictor<NativeConfig>(config);
   std::vector<framework::LoDTensor> jobs(num_jobs);
@@ -245,7 +247,7 @@ void MainThreadsImageClassification(bool use_gpu) {
   std::vector<std::thread> threads;
   for (int tid = 0; tid < num_jobs; ++tid) {
     threads.emplace_back([&, tid]() {
-      auto predictor = main_predictor->Clone();
+      auto predictor = CreatePaddlePredictor(config);
       auto& local_inputs = paddle_tensor_feeds[tid];
       std::vector<PaddleTensor> local_outputs;
       ASSERT_TRUE(predictor->Run(local_inputs, &local_outputs));
@@ -271,7 +273,7 @@ TEST(inference_api_native, word2vec_cpu_threads) {
   MainThreadsWord2Vec(false /*use_gpu*/);
 }
 TEST(inference_api_native, image_classification_cpu) {
-  MainThreadsImageClassification(false /*use_gpu*/);
+  MainImageClassification(false /*use_gpu*/);
 }
 TEST(inference_api_native, image_classification_cpu_threads) {
   MainThreadsImageClassification(false /*use_gpu*/);
@@ -279,15 +281,17 @@ TEST(inference_api_native, image_classification_cpu_threads) {
 
 #ifdef PADDLE_WITH_CUDA
 TEST(inference_api_native, word2vec_gpu) { MainWord2Vec(true /*use_gpu*/); }
-TEST(inference_api_native, word2vec_gpu_threads) {
-  MainThreadsWord2Vec(true /*use_gpu*/);
-}
+// Turn off temporarily for the unstable result.
+// TEST(inference_api_native, word2vec_gpu_threads) {
+//   MainThreadsWord2Vec(true /*use_gpu*/);
+// }
 TEST(inference_api_native, image_classification_gpu) {
-  MainThreadsImageClassification(true /*use_gpu*/);
-}
-TEST(inference_api_native, image_classification_gpu_threads) {
-  MainThreadsImageClassification(true /*use_gpu*/);
+  MainImageClassification(true /*use_gpu*/);
 }
+// Turn off temporarily for the unstable result.
+// TEST(inference_api_native, image_classification_gpu_threads) {
+//   MainThreadsImageClassification(true /*use_gpu*/);
+// }
 
 #endif
 

paddle/fluid/inference/api/api_tensorrt_subgraph_engine_tester.cc

@@ -29,13 +29,13 @@ void CompareTensorRTWithFluid(bool enable_tensorrt) {
 
   //# 1. Create PaddlePredictor with a config.
   NativeConfig config0;
-  config0.model_dir = FLAGS_dirname + "word2vec.inference.model";
+  config0.model_dir = FLAGS_dirname;
   config0.use_gpu = true;
   config0.fraction_of_gpu_memory = 0.3;
   config0.device = 0;
 
   MixedRTConfig config1;
-  config1.model_dir = FLAGS_dirname + "word2vec.inference.model";
+  config1.model_dir = FLAGS_dirname;
   config1.use_gpu = true;
   config1.fraction_of_gpu_memory = 0.3;
   config1.device = 0;

paddle/fluid/inference/api/demo_ci/CMakeLists.txt

@@ -52,6 +52,7 @@ 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("${PADDLE_LIB}/third_party/install/xxhash/include")
 if (NOT WIN32)
 include_directories("${PADDLE_LIB}/third_party/install/snappy/include")
 include_directories("${PADDLE_LIB}/third_party/install/snappystream/include")
@@ -61,8 +62,8 @@ endif(NOT WIN32)
 include_directories("${PADDLE_LIB}/third_party/boost")
 include_directories("${PADDLE_LIB}/third_party/eigen3")
 
-if (NOT WIN32) 
-  if (USE_TENSORRT AND WITH_GPU) 
+if (NOT WIN32)
+  if (USE_TENSORRT AND WITH_GPU)
       include_directories("${TENSORRT_INCLUDE_DIR}")
       link_directories("${TENSORRT_LIB_DIR}")
   endif()
@@ -77,13 +78,14 @@ endif(NOT WIN32)
 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/xxhash/lib")
 link_directories("${PADDLE_LIB}/paddle/lib")
 
 add_executable(${DEMO_NAME} ${DEMO_NAME}.cc)
 
 if(WITH_MKL)
   include_directories("${PADDLE_LIB}/third_party/install/mklml/include")
-  set(MATH_LIB ${PADDLE_LIB}/third_party/install/mklml/lib/libmklml_intel${CMAKE_SHARED_LIBRARY_SUFFIX} 
+  set(MATH_LIB ${PADDLE_LIB}/third_party/install/mklml/lib/libmklml_intel${CMAKE_SHARED_LIBRARY_SUFFIX}
                ${PADDLE_LIB}/third_party/install/mklml/lib/libiomp5${CMAKE_SHARED_LIBRARY_SUFFIX})
   set(MKLDNN_PATH "${PADDLE_LIB}/third_party/install/mkldnn")
   if(EXISTS ${MKLDNN_PATH})
@@ -107,7 +109,7 @@ if (NOT WIN32)
 set(EXTERNAL_LIB "-lrt -ldl -lpthread")
 set(DEPS ${DEPS}
     ${MATH_LIB} ${MKLDNN_LIB}
-    glog gflags protobuf snappystream snappy z
+    glog gflags protobuf snappystream snappy z xxhash
     ${EXTERNAL_LIB})
 else()
 set(DEPS ${DEPS}
@@ -120,7 +122,7 @@ endif(NOT WIN32)
 
 if(WITH_GPU)
   if(NOT WIN32)
-    if (USE_TENSORRT) 
+    if (USE_TENSORRT)
       set(DEPS ${DEPS} ${TENSORRT_LIB_DIR}/libnvinfer${CMAKE_STATIC_LIBRARY_SUFFIX})
       set(DEPS ${DEPS} ${TENSORRT_LIB_DIR}/libnvinfer_plugin${CMAKE_STATIC_LIBRARY_SUFFIX})
     endif()

paddle/fluid/inference/api/demo_ci/run.sh

@@ -16,7 +16,7 @@ if [ $2 == ON ]; then
 fi
 if [ $3 == ON ]; then
   use_gpu_list='true false'
-else    
+else
   use_gpu_list='false'
 fi
 
@@ -62,7 +62,7 @@ for WITH_STATIC_LIB in ON OFF; do
     -DWITH_GPU=$TEST_GPU_CPU \
     -DWITH_STATIC_LIB=$WITH_STATIC_LIB
   make -j
-  word2vec_model=${PADDLE_ROOT}'/build/python/paddle/fluid/tests/book/word2vec.inference.model'
+  word2vec_model=$DATA_DIR'/word2vec/word2vec.inference.model'
   if [ -d $word2vec_model ]; then
     for use_gpu in $use_gpu_list; do
       ./simple_on_word2vec \
@@ -83,7 +83,7 @@ for WITH_STATIC_LIB in ON OFF; do
     -DWITH_STATIC_LIB=$WITH_STATIC_LIB
   make -j
   for use_gpu in $use_gpu_list; do
-    for vis_demo_name in $vis_demo_list; do 
+    for vis_demo_name in $vis_demo_list; do
       ./vis_demo \
         --modeldir=$DATA_DIR/$vis_demo_name/model \
         --data=$DATA_DIR/$vis_demo_name/data.txt \
@@ -95,7 +95,7 @@ for WITH_STATIC_LIB in ON OFF; do
       fi
     done
   done
-  
+
   # --------tensorrt mobilenet------
   if [ $USE_TENSORRT == ON -a $TEST_GPU_CPU == ON ]; then
     rm -rf *
@@ -107,7 +107,7 @@ for WITH_STATIC_LIB in ON OFF; do
       -DUSE_TENSORRT=$USE_TENSORRT \
       -DTENSORRT_INCLUDE_DIR=$TENSORRT_INCLUDE_DIR \
       -DTENSORRT_LIB_DIR=$TENSORRT_LIB_DIR
-    make -j 
+    make -j
     ./trt_mobilenet_demo \
       --modeldir=$DATA_DIR/mobilenet/model \
       --data=$DATA_DIR/mobilenet/data.txt \

paddle/fluid/inference/api/details/reset_tensor_array.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 "paddle/fluid/inference/api/details/reset_tensor_array.h"
+
+namespace paddle {
+namespace details {
+
+// Should be called after the parameters are loaded.
+void TensorArrayBatchCleaner::CollectTensorArrays(framework::Scope *scope) {
+  if (flag_) {
+    for (auto &var_name : scope->LocalVarNames()) {
+      auto *var = scope->FindVar(var_name);
+      // TODO(Superjomn) should avoid the case when a TensorArray is a
+      // parameter.
+      if (var_name == "feed" || var_name == "fetch") continue;
+      if (var->Type() == typeid(framework::LoDTensorArray)) {
+        VLOG(4) << "collect " << var_name;
+        arrays_.push_back(var->GetMutable<framework::LoDTensorArray>());
+      }
+    }
+    for (auto *kid : scope->kids()) {
+      CollectTensorArrays(kid);
+    }
+
+    VLOG(3) << "Collect " << arrays_.size() << " arrays";
+    flag_ = false;
+  }
+}
+
+// Should be called when `Run` finished.
+void TensorArrayBatchCleaner::ResetTensorArray() {
+  for (auto *arr : arrays_) {
+    arr->clear();
+  }
+}
+
+}  // namespace details
+}  // namespace paddle

paddle/fluid/inference/api/details/reset_tensor_array.h

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include <vector>
+#include "paddle/fluid/framework/lod_tensor_array.h"
+#include "paddle/fluid/framework/scope.h"
+
+namespace paddle {
+namespace details {
+
+// Clean the TensorArray each batch to make the behavior the same with the
+// training phase.
+struct TensorArrayBatchCleaner {
+  // Fix the tensor array not clear in the inference scenarios.
+  void CollectTensorArrays(framework::Scope *scope);
+  void ResetTensorArray();
+
+ private:
+  bool flag_{true};
+  std::vector<framework::LoDTensorArray *> arrays_;
+};
+
+}  // namespace details
+}  // namespace paddle

paddle/fluid/inference/api/helper.h

@@ -160,7 +160,8 @@ static void PrintTime(int batch_size, int repeat, int num_threads, int tid,
                       double latency, int epoch = 1) {
   LOG(INFO) << "====== batch_size: " << batch_size << ", repeat: " << repeat
             << ", threads: " << num_threads << ", thread id: " << tid
-            << ", latency: " << latency << "ms ======";
+            << ", latency: " << latency << "ms, fps: " << 1 / (latency / 1000.f)
+            << " ======";
   if (epoch > 1) {
     int samples = batch_size * epoch;
     LOG(INFO) << "====== sample number: " << samples

paddle/fluid/inference/api/paddle_inference_api.h

@@ -124,7 +124,7 @@ class ZeroCopyTensor {
   std::vector<std::vector<size_t>> lod() const;
 
  protected:
-  ZeroCopyTensor(void* scope) : scope_{scope} {}
+  explicit ZeroCopyTensor(void* scope) : scope_{scope} {}
   void SetName(const std::string& name) { name_ = name; }
   void* FindTensor() const;
 
@@ -259,12 +259,6 @@ struct AnalysisConfig : public NativeConfig {
     kExclude   // Specify the disabled passes in `ir_passes`.
   };
 
-  void SetIncludeMode() {
-    ir_mode = IrPassMode::kInclude;
-    // this pass has to be run at the beginning of all fuse passes
-    ir_passes = {"infer_clean_graph_pass"};
-  }
-
   // Determine whether to perform graph optimization.
   bool enable_ir_optim = true;
   // Manually determine the IR passes to run.

paddle/fluid/inference/test.cmake

+set(INFERENCE_URL "http://paddle-inference-dist.cdn.bcebos.com" CACHE STRING "inference download url")
+set(INFERENCE_DEMO_INSTALL_DIR "${THIRD_PARTY_PATH}/inference_demo" CACHE STRING
+    "A path setting inference demo download directories.")
+function (inference_download install_dir url filename)
+    message(STATUS "Download inference test stuff from ${url}/${filename}")
+    execute_process(COMMAND bash -c "mkdir -p ${install_dir}")
+    execute_process(COMMAND bash -c "cd ${install_dir} && wget -q ${url}/${filename}")
+    message(STATUS "finish downloading ${filename}")
+endfunction()
+
+function (inference_download_and_uncompress install_dir url filename)
+    inference_download(${install_dir} ${url} ${filename})
+    execute_process(COMMAND bash -c "cd ${install_dir} && tar xzf ${filename}")
+endfunction()
+
+set(WORD2VEC_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/word2vec")
+if (NOT EXISTS ${WORD2VEC_INSTALL_DIR})
+    inference_download_and_uncompress(${WORD2VEC_INSTALL_DIR} ${INFERENCE_URL} "word2vec.inference.model.tar.gz")
+endif()
+set(WORD2VEC_MODEL_DIR "${WORD2VEC_INSTALL_DIR}/word2vec.inference.model")
+
+function (inference_base_test TARGET)
+   set(options "")
+   set(oneValueArgs "")
+   set(multiValueArgs SRCS ARGS DEPS)
+   cmake_parse_arguments(base_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+   if(WITH_GPU)
+       set(mem_opt "--fraction_of_gpu_memory_to_use=0.5")
+   endif()
+   cc_test(${TARGET} SRCS ${base_test_SRCS} DEPS ${base_test_DEPS} ARGS ${mem_opt} ${base_test_ARGS})
+endfunction()

paddle/fluid/inference/tests/api/CMakeLists.txt

-set(INFERENCE_URL "http://paddle-inference-dist.cdn.bcebos.com")
-set(INFERENCE_DEMO_INSTALL_DIR "${THIRD_PARTY_PATH}/inference_demo" CACHE STRING
-    "A path setting inference demo download directories.")
 set(INFERENCE_EXTRA_DEPS paddle_inference_api paddle_fluid_api ir_pass_manager analysis_predictor)
-function (inference_download install_dir url filename)
-    message(STATUS "Download inference test stuff from ${url}/${filename}")
-    execute_process(COMMAND bash -c "mkdir -p ${install_dir}")
-    execute_process(COMMAND bash -c "cd ${install_dir} && wget -q ${url}/${filename}")
-    message(STATUS "finish downloading ${filename}")
-endfunction()
-
-function (inference_download_and_uncompress install_dir url filename)
-    inference_download(${install_dir} ${url} ${filename})
-    execute_process(COMMAND bash -c "cd ${install_dir} && tar xzf ${filename}")
-endfunction()
 
 function(download_model_and_data install_dir model_name data_name)
     if (NOT EXISTS ${install_dir})

paddle/fluid/inference/tests/api/analyzer_rnn1_tester.cc

@@ -228,6 +228,7 @@ void SetInput(std::vector<std::vector<PaddleTensor>> *inputs) {
 TEST(Analyzer_rnn1, profile) {
   contrib::AnalysisConfig cfg;
   SetConfig(&cfg);
+  cfg.use_gpu = false;
   std::vector<PaddleTensor> outputs;
 
   std::vector<std::vector<PaddleTensor>> input_slots_all;

paddle/fluid/inference/tests/api/tester_helper.h

@@ -139,6 +139,9 @@ void TestMultiThreadPrediction(
   }
   for (int tid = 0; tid < num_threads; ++tid) {
     threads.emplace_back([&, tid]() {
+#ifdef PADDLE_WITH_MKLDNN
+      platform::set_cur_thread_id(static_cast<int>(tid) + 1);
+#endif
       // Each thread should have local inputs and outputs.
       // The inputs of each thread are all the same.
       std::vector<std::vector<PaddleTensor>> inputs_tid = inputs;

paddle/fluid/operators/CMakeLists.txt

@@ -268,6 +268,7 @@ if (WITH_GPU AND TENSORRT_FOUND)
 else()
     set(DEPS_OPS ${DEPS_OPS} tensorrt_engine_op)
 endif()
+op_library(hash_op DEPS xxhash)
 op_library(clip_by_norm_op DEPS selected_rows_functor selected_rows)
 op_library(sum_op DEPS selected_rows_functor)
 op_library(sgd_op DEPS selected_rows_functor)
@@ -300,6 +301,7 @@ op_library(flatten_op DEPS reshape_op)
 op_library(sequence_pad_op DEPS sequence_padding)
 op_library(unstack_op DEPS stack_op)
 op_library(fake_quantize_op DEPS memory)
+op_library(crf_decoding_op DEPS jit_kernel)
 op_library(fusion_lstm_op DEPS jit_kernel)
 if (WITH_GPU)
     op_library(conv_op DEPS vol2col depthwise_conv im2col)

paddle/fluid/operators/add_position_encoding_op.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 "paddle/fluid/operators/add_position_encoding_op.h"
+
+namespace paddle {
+namespace operators {
+
+class AddPositionEncodingOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"),
+                   "X(Input) of add_position_encoding_op should not be null.");
+    PADDLE_ENFORCE(
+        ctx->HasOutput("Out"),
+        "Out(Output) of add_position_encoding_op should not be null.");
+
+    auto x_dims = ctx->GetInputDim("X");
+    ctx->SetOutputDim("Out", x_dims);
+    ctx->ShareLoD("X", /*->*/ "Out");
+  }
+};
+
+class AddPositionEncodingOpGrad : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"), "X(Input) must not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Out"), "Out must not be null.");
+    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
+                   "Out@GRAD must not be null.");
+
+    auto out_dims = ctx->GetInputDim("Out");
+    if (ctx->HasOutput(framework::GradVarName("X"))) {
+      ctx->SetOutputDim(framework::GradVarName("X"), out_dims);
+    }
+  }
+};
+
+class AddPositionEncodingOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput("X", "Input of AddPositionEncoding operator");
+    AddOutput("Out", "Output of AddPositionEncoding operator");
+    AddAttr<float>("alpha", "The scale of Original Embedding.")
+        .SetDefault(1.0f)
+        .AddCustomChecker([](const float& alpha) {
+          PADDLE_ENFORCE(alpha >= 0.0f, "'alpha' must be above 0.0.");
+        });
+    AddAttr<float>("beta", "The scale of Position Embedding.")
+        .SetDefault(1.0f)
+        .AddCustomChecker([](const float& beta) {
+          PADDLE_ENFORCE(beta >= 0.0f, "'beta' must be between 0.0.");
+        });
+    AddComment(R"DOC(
+    Add Position Encoding Operator.
+    
+    The add position encoding calculates the output based on the input, alpha, beta.
+    The size of each dimension of the parameters checked in the infer-shape.
+  )DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+namespace plt = paddle::platform;
+
+REGISTER_OPERATOR(add_position_encoding, ops::AddPositionEncodingOp,
+                  ops::AddPositionEncodingOpMaker,
+                  paddle::framework::DefaultGradOpDescMaker<true>);
+REGISTER_OPERATOR(add_position_encoding_grad, ops::AddPositionEncodingOpGrad);
+
+REGISTER_OP_CPU_KERNEL(
+    add_position_encoding,
+    ops::AddPositionEncodingKernel<plt::CPUDeviceContext, float>,
+    ops::AddPositionEncodingKernel<plt::CPUDeviceContext, double>);
+
+REGISTER_OP_CPU_KERNEL(
+    add_position_encoding_grad,
+    ops::AddPositionEncodingGradKernel<plt::CPUDeviceContext, float>,
+    ops::AddPositionEncodingGradKernel<plt::CPUDeviceContext, double>);

paddle/fluid/operators/add_position_encoding_op.h

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include "paddle/fluid/framework/eigen.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/detail/safe_ref.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename DeviceContext, typename T>
+class AddPositionEncodingKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* X = context.Input<framework::LoDTensor>("X");
+    auto& x_lod = X->lod();
+    auto* src_ptr = X->data<T>();
+
+    auto* Out = context.Output<framework::LoDTensor>("Out");
+    auto* dst_ptr = Out->mutable_data<T>(context.GetPlace());
+
+    float alpha = context.Attr<float>("alpha");
+    float beta = context.Attr<float>("beta");
+
+    auto x_dim = X->dims();
+    int batch_size = 0;
+    int max_seq_len = 0;
+    int enc_size = 0;
+
+    if (x_lod.empty()) {
+      PADDLE_ENFORCE(
+          x_dim.size() == 3UL,
+          "The input X of Add Position Encoding should be 3-D Tensor!");
+      batch_size = x_dim[0];
+      max_seq_len = x_dim[1];
+      enc_size = x_dim[2];
+    } else {
+      PADDLE_ENFORCE(
+          x_dim.size() == 2UL,
+          "The input X of Add Position Encoding should be 2-D LoDTensor!");
+      PADDLE_ENFORCE(
+          x_lod.size() == 1UL,
+          "The Add Position Encoding Op only supports lod_level == 1!");
+      batch_size = x_lod[0].size() - 1;
+      max_seq_len = -1;
+      enc_size = x_dim[1];
+    }
+
+    PADDLE_ENFORCE(enc_size % 2 == 0, "Only support even encode size!");
+
+    const int half_size = enc_size / 2;
+    for (int i = 0; i < batch_size; ++i) {
+      const int max_length =
+          x_lod.empty() ? max_seq_len : x_lod[0][i + 1] - x_lod[0][i];
+      for (int j = 0; j < max_length; ++j) {
+        for (int k = 0; k < half_size; ++k) {
+          const double val = (half_size > 1)
+                                 ? j / pow(10000.0, double(k) / (half_size - 1))
+                                 : j / 10000.0;
+          dst_ptr[k] = src_ptr[k] * alpha + sin(val) * beta;
+          dst_ptr[half_size + k] =
+              src_ptr[half_size + k] * alpha + cos(val) * beta;
+        }
+        src_ptr += enc_size;
+        dst_ptr += enc_size;
+      }
+    }
+  }
+};
+
+template <typename DeviceContext, typename T>
+class AddPositionEncodingGradKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* dOut =
+        context.Input<framework::LoDTensor>(framework::GradVarName("Out"));
+    auto dout = framework::EigenVector<T>::Flatten(*dOut);
+
+    auto* dX =
+        context.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    dX->mutable_data<T>(context.GetPlace());
+    auto dx = framework::EigenVector<T>::Flatten(*dX);
+
+    float alpha = context.Attr<float>("alpha");
+
+    auto* place =
+        context.template device_context<DeviceContext>().eigen_device();
+    dx.device(*place) = dout * static_cast<T>(alpha);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/affine_grid_cudnn_op.cu.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 "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/platform/cudnn_helper.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+using ScopedSpatialTransformerDescriptor =
+    platform::ScopedSpatialTransformerDescriptor;
+
+template <typename T>
+class CUDNNAffineGridOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
+                   "It must use CUDAPlace.");
+    auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
+    auto handle = dev_ctx.cudnn_handle();
+    auto* theta = ctx.Input<Tensor>("Theta");
+    auto* output = ctx.Output<Tensor>("Output");
+    const T* theta_data = theta->data<T>();
+
+    int n = theta->dims()[0];
+    auto size_attr = ctx.Attr<std::vector<int>>("output_shape");
+    Tensor h_sizes;
+    int* h_size_data;
+    if (size_attr.size() == 0) {
+      auto* output_shape = ctx.Input<Tensor>("OutputShape");
+      framework::TensorCopy(*output_shape, platform::CPUPlace(), &h_sizes);
+      h_size_data = h_sizes.data<int>();
+    } else {
+      h_size_data = h_sizes.mutable_data<int>({4}, platform::CPUPlace());
+      h_size_data[0] = n;
+      h_size_data[1] = size_attr[1];
+      h_size_data[2] = size_attr[2];
+      h_size_data[3] = size_attr[3];
+    }
+
+    T* output_data = output->mutable_data<T>(
+        {n, h_size_data[2], h_size_data[3], 2}, ctx.GetPlace());
+    ScopedSpatialTransformerDescriptor st_desc;
+    cudnnSpatialTransformerDescriptor_t cudnn_st_desc =
+        st_desc.descriptor<T>(4, h_size_data);
+
+    PADDLE_ENFORCE(platform::dynload::cudnnSpatialTfGridGeneratorForward(
+        handle, cudnn_st_desc, theta_data, output_data));
+  }
+};
+
+template <typename T>
+class CUDNNAffineGridGradOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
+                   "It must use CUDAPlace.");
+    auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
+    auto handle = dev_ctx.cudnn_handle();
+    auto output_grad = ctx.Input<Tensor>(framework::GradVarName("Output"));
+    auto theta_grad = ctx.Output<Tensor>(framework::GradVarName("Theta"));
+
+    int n = output_grad->dims()[0];
+    auto size_attr = ctx.Attr<std::vector<int>>("output_shape");
+    Tensor h_sizes;
+    int* h_size_data;
+    if (size_attr.size() == 0) {
+      auto* output_shape = ctx.Input<Tensor>("OutputShape");
+      framework::TensorCopy(*output_shape, platform::CPUPlace(), &h_sizes);
+      h_size_data = h_sizes.data<int>();
+    } else {
+      h_size_data = h_sizes.mutable_data<int>({4}, platform::CPUPlace());
+      h_size_data[0] = n;
+      h_size_data[1] = size_attr[1];
+      h_size_data[2] = size_attr[2];
+      h_size_data[3] = size_attr[3];
+    }
+
+    ScopedSpatialTransformerDescriptor st_desc;
+    cudnnSpatialTransformerDescriptor_t cudnn_st_desc =
+        st_desc.descriptor<T>(4, h_size_data);
+
+    const T* output_grad_data = output_grad->data<T>();
+    T* theta_grad_data = theta_grad->mutable_data<T>(ctx.GetPlace());
+
+    PADDLE_ENFORCE(platform::dynload::cudnnSpatialTfGridGeneratorBackward(
+        handle, cudnn_st_desc, output_grad_data, theta_grad_data));
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace plat = paddle::platform;
+REGISTER_OP_KERNEL(affine_grid, CUDNN, plat::CUDAPlace,
+                   paddle::operators::CUDNNAffineGridOpKernel<float>,
+                   paddle::operators::CUDNNAffineGridOpKernel<double>);
+REGISTER_OP_KERNEL(affine_grid_grad, CUDNN, plat::CUDAPlace,
+                   paddle::operators::CUDNNAffineGridGradOpKernel<float>,
+                   paddle::operators::CUDNNAffineGridGradOpKernel<double>);

paddle/fluid/operators/affine_grid_op.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 "paddle/fluid/operators/affine_grid_op.h"
+#include <string>
+#include "paddle/fluid/framework/op_registry.h"
+#ifdef PADDLE_WITH_CUDA
+#include "paddle/fluid/platform/cudnn_helper.h"
+#endif
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+template <typename T>
+struct Linspace<paddle::platform::CPUDeviceContext, T> {
+  framework::Tensor operator()(T start, T end, int count,
+                               const framework::ExecutionContext& ctx) {
+    Tensor numbers;
+    T* number_data = numbers.mutable_data<T>({count}, platform::CPUPlace());
+    T slice = (end - start) / (T)(count - 1);
+    for (int i = 0; i < count; ++i) {
+      number_data[i] = start + (T)i * slice;
+    }
+    return numbers;
+  }
+};
+
+class AffineGridOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("Theta"),
+                   "Input(Theta) of AffineGridOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Output"),
+                   "Output(Output) of AffineGridOp should not be null.");
+    auto theta_dims = ctx->GetInputDim("Theta");
+    PADDLE_ENFORCE(theta_dims.size() == 3,
+                   "AffineGrid's Input(Theta) should be 3-D tensor.");
+
+    auto output_shape = ctx->Attrs().Get<std::vector<int>>("output_shape");
+    if (output_shape.size() == 0) {
+      PADDLE_ENFORCE(ctx->HasInput("OutputShape"),
+                     "Input(OutputShape) of AffineGridOp should not be null if "
+                     "attr(output_shape) is not configured.");
+      auto output_shape_dims = ctx->GetInputDim("OutputShape");
+      PADDLE_ENFORCE(output_shape_dims.size() == 1,
+                     "AffineGrid's Input(OutputShape) should be 1-D tensor.");
+    } else {
+      PADDLE_ENFORCE(output_shape.size() == 4,
+                     "The size of attr(output_shape) should be 4.");
+    }
+
+    PADDLE_ENFORCE(theta_dims[1] == 2, "Input(theta) dims[1] should be 2.");
+    PADDLE_ENFORCE(theta_dims[2] == 3, "Input(theta) dims[2] should be 3.");
+    // N * H * W * 2
+    ctx->SetOutputDim("Output",
+                      framework::make_ddim({theta_dims[0], -1, -1, 2}));
+    ctx->ShareLoD("Theta", "Output");
+  }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    framework::LibraryType library{framework::LibraryType::kPlain};
+#ifdef PADDLE_WITH_CUDA
+    if (platform::CanCUDNNBeUsed(ctx)) {
+      library = framework::LibraryType::kCUDNN;
+    }
+#endif
+    auto data_type = framework::ToDataType(ctx.Input<Tensor>("Theta")->type());
+    return framework::OpKernelType(data_type, ctx.GetPlace(),
+                                   framework::DataLayout::kAnyLayout, library);
+  }
+};
+
+class AffineGridOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput(
+        "Theta",
+        "(Tensor) A batch of affine transform parameters with shape [N, 2, 3]. "
+        "It is used to transform coordinate (x_0, y_0) to coordinate (x_1, "
+        "y_1).");
+    AddInput("OutputShape",
+             "(Tensor) The shape of target image with format [N, C, H, W].")
+        .AsDispensable();
+    AddOutput("Output", "(Tensor) Output Tensor with shape [N, H, W, 2].");
+    AddAttr<bool>(
+        "use_cudnn",
+        "(bool, default false) Only used in cudnn kernel, need install cudnn")
+        .SetDefault(true);
+    AddAttr<std::vector<int>>(
+        "output_shape",
+        "The target output image shape with format [N, C, H, W].")
+        .SetDefault(std::vector<int>());
+
+    AddComment(R"DOC(
+    It generates a grid of (x,y) coordinates using the parameters of the
+    affine transformation that correspond to a set of points where the input
+    feature map should be sampled to produce the transformed output feature map.
+
+    Given:
+        Theta = [[[x_11, x_12, x_13]
+                  [x_14, x_15, x_16]]
+                 [[x_21, x_22, x_23]
+                  [x_24, x_25, x_26]]]
+    
+        OutputShape = [2, 3, 5, 5]
+
+    Step 1:
+
+        Generate relative coordinates according to OutputShape.
+        The values of relative coordinates are in the interval between -1 and 1.
+        The shape of the relative coordinates is [2, H, W] as below:
+    
+        C = [[[-1.  -1.  -1.  -1.  -1. ]
+              [-0.5 -0.5 -0.5 -0.5 -0.5]
+              [ 0.   0.   0.   0.   0. ]
+              [ 0.5  0.5  0.5  0.5  0.5]
+              [ 1.   1.   1.   1.   1. ]] 
+             [[-1.  -0.5  0.   0.5  1. ]
+              [-1.  -0.5  0.   0.5  1. ]
+              [-1.  -0.5  0.   0.5  1. ]
+              [-1.  -0.5  0.   0.5  1. ]
+              [-1.  -0.5  0.   0.5  1. ]]]
+        C[0] is the coordinates in height axis and  C[1] is the coordinates in width axis.
+    
+    Step2:
+        Tanspose and reshape C to shape [H * W, 2] and append ones to last dimension. The we get:
+        C_ = [[-1.  -1.   1. ]
+              [-0.5 -1.   1. ]
+              [ 0.  -1.   1. ]
+              [ 0.5 -1.   1. ]
+              [ 1.  -1.   1. ]
+              [-1.  -0.5  1. ]
+              [-0.5 -0.5  1. ]
+              [ 0.  -0.5  1. ]
+              [ 0.5 -0.5  1. ]
+              [ 1.  -0.5  1. ]
+              [-1.   0.   1. ]
+              [-0.5  0.   1. ]
+              [ 0.   0.   1. ]
+              [ 0.5  0.   1. ]
+              [ 1.   0.   1. ]
+              [-1.   0.5  1. ]
+              [-0.5  0.5  1. ]
+              [ 0.   0.5  1. ]
+              [ 0.5  0.5  1. ]
+              [ 1.   0.5  1. ]
+              [-1.   1.   1. ]
+              [-0.5  1.   1. ]
+              [ 0.   1.   1. ]
+              [ 0.5  1.   1. ]
+              [ 1.   1.   1. ]]
+    Step3:
+        Compute output by equation $$Output[i] = C_ * Theta[i]^T$$
+    )DOC");
+  }
+};
+
+class AffineGridOpGrad : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    auto theta_dims = ctx->GetInputDim("Theta");
+    if (ctx->HasOutput(framework::GradVarName("Theta"))) {
+      ctx->SetOutputDim(framework::GradVarName("Theta"), theta_dims);
+    }
+  }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    framework::LibraryType library_{framework::LibraryType::kPlain};
+#ifdef PADDLE_WITH_CUDA
+    if (platform::CanCUDNNBeUsed(ctx)) {
+      library_ = framework::LibraryType::kCUDNN;
+    }
+#endif
+    return framework::OpKernelType(
+        framework::ToDataType(ctx.Input<Tensor>("Theta")->type()),
+        ctx.GetPlace(), framework::DataLayout::kAnyLayout, library_);
+  }
+};
+
+class AffineGridGradMaker : public framework::SingleGradOpDescMaker {
+ public:
+  using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
+
+ protected:
+  std::unique_ptr<framework::OpDesc> Apply() const override {
+    auto* op = new framework::OpDesc();
+    op->SetType("affine_grid_grad");
+    op->SetInput("Theta", Input("Theta"));
+    op->SetInput("OutputShape", Input("OutputShape"));
+    op->SetInput(framework::GradVarName("Output"), OutputGrad("Output"));
+
+    op->SetAttrMap(Attrs());
+
+    op->SetOutput(framework::GradVarName("Theta"), InputGrad("Theta"));
+    return std::unique_ptr<framework::OpDesc>(op);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(affine_grid, ops::AffineGridOp, ops::AffineGridOpMaker,
+                  ops::AffineGridGradMaker);
+REGISTER_OPERATOR(affine_grid_grad, ops::AffineGridOpGrad);
+
+REGISTER_OP_CPU_KERNEL(
+    affine_grid,
+    ops::AffineGridOpKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::AffineGridOpKernel<paddle::platform::CPUDeviceContext, double>);
+REGISTER_OP_CPU_KERNEL(
+    affine_grid_grad,
+    ops::AffineGridGradOpKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::AffineGridGradOpKernel<paddle::platform::CPUDeviceContext, double>);

paddle/fluid/operators/affine_grid_op.h

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include <vector>
+#include "paddle/fluid/framework/eigen.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/math/blas.h"
+#include "paddle/fluid/operators/math/math_function.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+template <typename T, size_t D, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenTensor = framework::EigenTensor<T, D, MajorType, IndexType>;
+
+using Array1 = Eigen::DSizes<int64_t, 1>;
+using Array2 = Eigen::DSizes<int64_t, 2>;
+using Array3 = Eigen::DSizes<int64_t, 3>;
+using Array4 = Eigen::DSizes<int64_t, 4>;
+
+/**
+ *Return a tensor with evenly spaced numbers over a specified interval.
+ */
+template <typename DeviceContext, typename T>
+struct Linspace {
+  framework::Tensor operator()(T start, T end, int count,
+                               const framework::ExecutionContext& ctx);
+};
+
+template <typename DeviceContext, typename T>
+class AffineGridOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
+    auto* theta = ctx.Input<Tensor>("Theta");
+    int n = theta->dims()[0];
+
+    auto size_attr = ctx.Attr<std::vector<int>>("output_shape");
+    int h = 0;
+    int w = 0;
+    if (size_attr.size() == 0) {
+      auto* output_shape = ctx.Input<Tensor>("OutputShape");
+      Tensor h_sizes;
+      framework::TensorCopy(*output_shape, platform::CPUPlace(), &h_sizes);
+      const int* h_size_data = h_sizes.data<int>();
+      h = h_size_data[2];
+      w = h_size_data[3];
+    } else {
+      h = size_attr[2];
+      w = size_attr[3];
+    }
+
+    auto* output = ctx.Output<Tensor>("Output");
+    output->mutable_data<T>({n, h, w, 2}, ctx.GetPlace());
+
+    math::SetConstant<DeviceContext, T>()(
+        ctx.template device_context<DeviceContext>(), output,
+        static_cast<T>(0));
+
+    Linspace<DeviceContext, T> linspace;
+    // Get indexes of height with shape [height, width, 1]
+    auto h_idx = linspace((T)-1, (T)1, h, ctx);
+    auto h_idx_t = EigenTensor<T, 1>::From(h_idx);
+    // Get indexes of width with shape [height, width, 1]
+    auto w_idx = linspace((T)-1, (T)1, w, ctx);
+    auto w_idx_t = EigenTensor<T, 1>::From(w_idx);
+    // Get constant ones tensor with shape [height, width, 1]
+    Tensor ones;
+    ones.mutable_data<T>({h, w, 1}, ctx.GetPlace());
+    auto ones_t = EigenTensor<T, 3>::From(ones).setConstant((T)1);
+    // Get grid tensor with shape [n, h, w, 3] by concatenating h_idx, w_idx and
+    // ones
+    Tensor grid;
+    grid.mutable_data<T>({n, h, w, 3}, ctx.GetPlace());
+    auto grid_t = EigenTensor<T, 4>::From(grid);
+
+    grid_t.device(place) = w_idx_t.reshape(Array2(1, w))
+                               .broadcast(Array2(h, 1))
+                               .reshape(Array3(h, w, 1))
+                               .concatenate(h_idx_t.reshape(Array2(1, h))
+                                                .broadcast(Array2(w, 1))
+                                                .shuffle(Array2(1, 0))
+                                                .reshape(Array3(h, w, 1)),
+                                            2)
+                               .eval()
+                               .concatenate(ones_t, 2)
+                               .reshape(Array4(1, h, w, 3))
+                               .broadcast(Array4(n, 1, 1, 1));
+
+    // output = grid * theta.T
+    // TODO(wanghaoshuang): Refine batched matrix multiply
+    auto blas = math::GetBlas<DeviceContext, T>(ctx);
+    for (int i = 0; i < n; ++i) {
+      Tensor sliced_grid = grid.Slice(i, i + 1).Resize({h * w, 3});
+      Tensor sliced_theta = theta->Slice(i, i + 1).Resize({2, 3});
+      Tensor sliced_out = output->Slice(i, i + 1).Resize({h * w, 2});
+      blas.MatMul(sliced_grid, false, sliced_theta, true, T(1), &sliced_out,
+                  T(0));
+    }
+  }
+};
+
+template <typename DeviceContext, typename T>
+class AffineGridGradOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
+    auto output_grad = ctx.Input<Tensor>(framework::GradVarName("Output"));
+    auto theta_grad = ctx.Output<Tensor>(framework::GradVarName("Theta"));
+
+    int n = output_grad->dims()[0];
+    auto size_attr = ctx.Attr<std::vector<int>>("output_shape");
+    int h = 0;
+    int w = 0;
+    if (size_attr.size() == 0) {
+      auto* output_shape = ctx.Input<Tensor>("OutputShape");
+      Tensor h_sizes;
+      framework::TensorCopy(*output_shape, platform::CPUPlace(), &h_sizes);
+      const int* h_size_data = h_sizes.data<int>();
+      h = h_size_data[2];
+      w = h_size_data[3];
+    } else {
+      h = size_attr[2];
+      w = size_attr[3];
+    }
+
+    theta_grad->mutable_data<T>({n, 2, 3}, ctx.GetPlace());
+
+    math::SetConstant<DeviceContext, T>()(
+        ctx.template device_context<DeviceContext>(), theta_grad,
+        static_cast<T>(0));
+
+    Linspace<DeviceContext, T> linspace;
+
+    // Get indexes of height with shape [height, width, 1]
+    auto h_idx = linspace((T)-1, (T)1, h, ctx);
+    auto h_idx_t = EigenTensor<T, 1>::From(h_idx);
+    // Get indexes of width with shape [height, width, 1]
+    auto w_idx = linspace((T)-1, (T)1, w, ctx);
+    auto w_idx_t = EigenTensor<T, 1>::From(w_idx);
+    // Get constant ones tensor with shape [height, width, 1]
+    Tensor ones;
+    ones.mutable_data<T>({h, w, 1}, ctx.GetPlace());
+    auto ones_t = EigenTensor<T, 3>::From(ones).setConstant((T)1);
+    // Get grid tensor with shape [n, h, w, 3] by concatenating h_idx, w_idx and
+    // ones
+    Tensor grid;
+    grid.mutable_data<T>({n, h, w, 3}, ctx.GetPlace());
+    auto grid_t = EigenTensor<T, 4>::From(grid);
+    grid_t.device(place) = w_idx_t.reshape(Array2(1, w))
+                               .broadcast(Array2(h, 1))
+                               .reshape(Array3(h, w, 1))
+                               .concatenate(h_idx_t.reshape(Array2(1, h))
+                                                .broadcast(Array2(w, 1))
+                                                .shuffle(Array2(1, 0))
+                                                .reshape(Array3(h, w, 1)),
+                                            2)
+                               .eval()
+                               .concatenate(ones_t, 2)
+                               .reshape(Array4(1, h, w, 3))
+                               .broadcast(Array4(n, 1, 1, 1));
+    // output = grid * theta.T
+    // TODO(wanghaoshuang): Refine batched matrix multiply
+    auto blas = math::GetBlas<DeviceContext, T>(ctx);
+    for (int i = 0; i < n; ++i) {
+      Tensor sliced_grid = grid.Slice(i, i + 1).Resize({h * w, 3});
+      Tensor sliced_out_grad = output_grad->Slice(i, i + 1).Resize({h * w, 2});
+      Tensor sliced_theta_grad = theta_grad->Slice(i, i + 1).Resize({2, 3});
+      blas.MatMul(sliced_out_grad, true, sliced_grid, false, T(1),
+                  &sliced_theta_grad, T(0));
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/beam_search_decode_op.cc

@@ -79,6 +79,9 @@ struct BeamSearchDecodeFunctor {
   bool tensor_on_gpu_;
   size_t beam_size_;
   int end_id_;
+  // TODO(Superjomn) Here might result serious performance issue in the
+  // concurrency
+  // scenarios.
   const LoDTensorArray& step_ids_origin_;
   const LoDTensorArray& step_scores_origin_;
   LoDTensorArray step_ids_ = LoDTensorArray();