Merge remote-tracking branch 'origin/develop' into fix/sign_op

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

@@ -0,0 +1,50 @@
+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

@@ -64,11 +64,11 @@ 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))
-paddle.fluid.layers.pool3d 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))
+paddle.fluid.layers.pool2d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name', 'exclusive'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None, True))
+paddle.fluid.layers.pool3d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name', 'exclusive'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None, True))
 paddle.fluid.layers.batch_norm ArgSpec(args=['input', 'act', 'is_test', 'momentum', 'epsilon', 'param_attr', 'bias_attr', 'data_layout', 'in_place', 'name', 'moving_mean_name', 'moving_variance_name', 'do_model_average_for_mean_and_var', 'fuse_with_relu'], varargs=None, keywords=None, defaults=(None, False, 0.9, 1e-05, None, None, 'NCHW', False, None, None, None, False, False))
 paddle.fluid.layers.beam_search_decode ArgSpec(args=['ids', 'scores', 'beam_size', 'end_id', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.conv2d_transpose ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, 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))
@@ -103,11 +103,11 @@ paddle.fluid.layers.beam_search ArgSpec(args=['pre_ids', 'pre_scores', 'ids', 's
 paddle.fluid.layers.row_conv ArgSpec(args=['input', 'future_context_size', 'param_attr', 'act'], varargs=None, keywords=None, defaults=(None, None))
 paddle.fluid.layers.multiplex ArgSpec(args=['inputs', 'index'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.layer_norm ArgSpec(args=['input', 'scale', 'shift', 'begin_norm_axis', 'epsilon', 'param_attr', 'bias_attr', 'act', 'name'], varargs=None, keywords=None, defaults=(True, True, 1, 1e-05, None, None, None, None))
-paddle.fluid.layers.softmax_with_cross_entropy ArgSpec(args=['logits', 'label', 'soft_label', 'ignore_index'], varargs=None, keywords=None, defaults=(False, -100))
+paddle.fluid.layers.softmax_with_cross_entropy ArgSpec(args=['logits', 'label', 'soft_label', 'ignore_index', 'numeric_stable_mode'], varargs=None, keywords=None, defaults=(False, -100, False))
 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 +174,13 @@ 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.grid_sampler ArgSpec(args=['x', 'grid', 'name'], varargs=None, keywords=None, defaults=(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 +359,8 @@ 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)
@@ -33,13 +35,15 @@ if(WITH_GPU)
           all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle graph graph_helper pass)
 endif()
 
+cc_library(sequential_execution_pass SRCS sequential_execution_pass.cc DEPS graph graph_helper pass)
+
 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)
+  cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS graph framework_proto reference_count_pass sequential_execution_pass)
 else()
-  cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS graph framework_proto)
+  cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS graph framework_proto sequential_execution_pass)
 endif()
 
 cc_library(threaded_ssa_graph_executor SRCS threaded_ssa_graph_executor.cc DEPS fetch_op_handle ssa_graph_executor scope
@@ -54,8 +58,9 @@ 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/build_strategy.cc

@@ -16,6 +16,7 @@ limitations under the License. */
 
 #include "paddle/fluid/framework/details/multi_devices_graph_check_pass.h"
 #include "paddle/fluid/framework/details/multi_devices_graph_print_pass.h"
+#include "paddle/fluid/framework/details/sequential_execution_pass.h"
 #include "paddle/fluid/framework/ir/graph.h"
 #include "paddle/fluid/framework/ir/graph_viz_pass.h"
 
@@ -27,6 +28,10 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
  public:
   explicit ParallelExecutorPassBuilder(const BuildStrategy &strategy)
       : ir::PassBuilder(), strategy_(strategy) {
+    if (strategy_.enable_sequential_execution_) {
+      AppendPass("sequential_execution_pass");
+    }
+
     // Add a graph viz pass to record a graph.
     if (!strategy_.debug_graphviz_path_.empty()) {
       auto viz_pass = AppendPass("graph_viz_pass");
@@ -110,6 +115,11 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
       pass->Erase("nccl_ctxs");
       pass->SetNotOwned<platform::NCCLContextMap>("nccl_ctxs", nctx);
 #endif
+    } else if (pass->Type() == "sequential_execution_pass") {
+      pass->Erase(kAllOpDescs);
+      pass->Set<const std::vector<OpDesc *>>(
+          kAllOpDescs,
+          new std::vector<OpDesc *>(main_program.Block(0).AllOps()));
     }
     graph = pass->Apply(std::move(graph));
   }
@@ -121,6 +131,8 @@ 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);
+USE_PASS(sequential_execution_pass);

paddle/fluid/framework/details/build_strategy.h

@@ -69,6 +69,10 @@ struct BuildStrategy {
 
   bool enable_data_balance_{false};
 
+  bool enable_sequential_execution_{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();