Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into add-merge-splited-ids

CMakeLists.txt

@@ -55,12 +55,13 @@ option(WITH_FLUID_ONLY  "Compile PaddlePaddle fluid only"               OFF)
 option(WITH_GOLANG      "Compile PaddlePaddle with GOLANG"              OFF)
 option(GLIDE_INSTALL    "Download and install go dependencies "         ON)
 option(USE_NNPACK       "Compile PaddlePaddle with NNPACK library"      OFF)
-option(WITH_DISTRIBUTE  "Compile with grpc distributed support"         OFF)
+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(WITH_GRPC     "Use grpc as the default rpc framework"            ${WITH_DISTRIBUTE})
 
 # CMAKE_BUILD_TYPE
 if(NOT CMAKE_BUILD_TYPE)
@@ -147,7 +148,16 @@ include(external/any)       # download libn::any
 include(external/eigen)     # download eigen3
 include(external/pybind11)  # download pybind11
 include(external/cares)
-include(external/grpc)
+
+if(WITH_DISTRIBUTE)
+    if(WITH_GRPC)
+        include(external/grpc)
+    else()
+        include(external/leveldb)
+        include(external/brpc)
+    endif()
+endif()
+
 include(external/snappy)    # download snappy
 include(external/snappystream)
 include(external/threadpool)

benchmark/fluid/README.md

@@ -24,10 +24,12 @@ Currently supported `--model` argument include:
 
 * Run the following command to start a benchmark job locally:
     ```bash
-      python fluid_benchmark.py --model mnist  --device GPU
+      python fluid_benchmark.py --model mnist --device GPU
     ```
     You can choose to use GPU/CPU training. With GPU training, you can specify
     `--gpus <gpu_num>` to run multi GPU training.
+    You can set async mode parameter server. With async mode, you can specify
+    `--async_mode` to train model asynchronous.
 * Run distributed training with parameter servers:
     * see [run_fluid_benchmark.sh](https://github.com/PaddlePaddle/Paddle/blob/develop/benchmark/fluid/run_fluid_benchmark.sh) as an example.
     * start parameter servers:

benchmark/fluid/args.py

+# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+
+__all__ = ['parse_args', ]
+
+BENCHMARK_MODELS = [
+    "machine_translation", "resnet", "vgg", "mnist", "stacked_dynamic_lstm"
+]
+
+
+def parse_args():
+    parser = argparse.ArgumentParser('Fluid model benchmarks.')
+    parser.add_argument(
+        '--model',
+        type=str,
+        choices=BENCHMARK_MODELS,
+        default='resnet',
+        help='The model to run benchmark with.')
+    parser.add_argument(
+        '--batch_size', type=int, default=32, help='The minibatch size.')
+    #  args related to learning rate
+    parser.add_argument(
+        '--learning_rate', type=float, default=0.001, help='The learning rate.')
+    # TODO(wuyi): add "--use_fake_data" option back.
+    parser.add_argument(
+        '--skip_batch_num',
+        type=int,
+        default=5,
+        help='The first num of minibatch num to skip, for better performance test'
+    )
+    parser.add_argument(
+        '--iterations', type=int, default=80, help='The number of minibatches.')
+    parser.add_argument(
+        '--pass_num', type=int, default=100, help='The number of passes.')
+    parser.add_argument(
+        '--data_format',
+        type=str,
+        default='NCHW',
+        choices=['NCHW', 'NHWC'],
+        help='The data data_format, now only support NCHW.')
+    parser.add_argument(
+        '--device',
+        type=str,
+        default='GPU',
+        choices=['CPU', 'GPU'],
+        help='The device type.')
+    parser.add_argument(
+        '--gpus',
+        type=int,
+        default=1,
+        help='If gpus > 1, will use ParallelExecutor to run, else use Executor.')
+    # this option is available only for vgg and resnet.
+    parser.add_argument(
+        '--cpus',
+        type=int,
+        default=1,
+        help='If cpus > 1, will use ParallelDo to run, else use Executor.')
+    parser.add_argument(
+        '--data_set',
+        type=str,
+        default='flowers',
+        choices=['cifar10', 'flowers'],
+        help='Optional dataset for benchmark.')
+    parser.add_argument(
+        '--infer_only', action='store_true', help='If set, run forward only.')
+    parser.add_argument(
+        '--use_cprof', action='store_true', help='If set, use cProfile.')
+    parser.add_argument(
+        '--use_nvprof',
+        action='store_true',
+        help='If set, use nvprof for CUDA.')
+    parser.add_argument(
+        '--no_test',
+        action='store_true',
+        help='If set, do not test the testset during training.')
+    parser.add_argument(
+        '--memory_optimize',
+        action='store_true',
+        help='If set, optimize runtime memory before start.')
+    parser.add_argument(
+        '--use_fake_data',
+        action='store_true',
+        help='If set ommit the actual read data operators.')
+    parser.add_argument(
+        '--profile', action='store_true', help='If set, profile a few steps.')
+    parser.add_argument(
+        '--update_method',
+        type=str,
+        default='local',
+        choices=['local', 'pserver', 'nccl2'],
+        help='Choose parameter update method, can be local, pserver, nccl2.')
+    parser.add_argument(
+        '--no_split_var',
+        action='store_true',
+        default=False,
+        help='Whether split variables into blocks when update_method is pserver')
+    parser.add_argument(
+        '--async_mode',
+        action='store_true',
+        default=False,
+        help='Whether start pserver in async mode to support ASGD')
+    parser.add_argument(
+        '--use_reader_op',
+        action='store_true',
+        help='Whether to use reader op, and must specify the data path if set this to true.'
+    )
+    parser.add_argument(
+        '--data_path',
+        type=str,
+        default="",
+        help='Directory that contains all the training recordio files.')
+    args = parser.parse_args()
+    return args

benchmark/fluid/fluid_benchmark.py

@@ -24,108 +24,7 @@ import paddle.fluid.core as core
 import paddle.fluid.profiler as profiler
 import paddle.fluid.transpiler.distribute_transpiler as distribute_transpiler
 
-BENCHMARK_MODELS = [
-    "machine_translation", "resnet", "vgg", "mnist", "stacked_dynamic_lstm"
-]
-
-
-def parse_args():
-    parser = argparse.ArgumentParser('Fluid model benchmarks.')
-    parser.add_argument(
-        '--model',
-        type=str,
-        choices=BENCHMARK_MODELS,
-        default='resnet',
-        help='The model to run benchmark with.')
-    parser.add_argument(
-        '--batch_size',
-        type=int,
-        default=32,
-        help='The batch size on each gpu.')
-    parser.add_argument(
-        '--learning_rate', type=float, default=0.001, help='The learning rate.')
-    parser.add_argument(
-        '--skip_batch_num',
-        type=int,
-        default=5,
-        help='The first num of minibatch num to skip, for better performance test'
-    )
-    parser.add_argument(
-        '--iterations',
-        type=int,
-        default=80,
-        help='The number of minibatches, set to -1 to run all batches.')
-    parser.add_argument(
-        '--pass_num', type=int, default=100, help='The number of passes.')
-    parser.add_argument(
-        '--data_format',
-        type=str,
-        default='NCHW',
-        choices=['NCHW', 'NHWC'],
-        help='The data data_format, now only support NCHW.')
-    parser.add_argument(
-        '--device',
-        type=str,
-        default='GPU',
-        choices=['CPU', 'GPU'],
-        help='The device type.')
-    parser.add_argument(
-        '--gpus',
-        type=int,
-        default=1,
-        help='If gpus > 1, will use ParallelExecutor to run, else use Executor.')
-    # this option is available only for vgg and resnet.
-    parser.add_argument(
-        '--cpus',
-        type=int,
-        default=1,
-        help='If cpus > 1, will use ParallelDo to run, else use Executor.')
-    parser.add_argument(
-        '--data_set',
-        type=str,
-        default='flowers',
-        choices=['cifar10', 'flowers', 'imagenet'],
-        help='Optional dataset for benchmark.')
-    parser.add_argument(
-        '--infer_only', action='store_true', help='If set, run forward only.')
-    parser.add_argument(
-        '--use_cprof', action='store_true', help='If set, use cProfile.')
-    parser.add_argument(
-        '--use_nvprof',
-        action='store_true',
-        help='If set, use nvprof for CUDA.')
-    parser.add_argument(
-        '--no_test',
-        action='store_true',
-        help='If set, do not test the testset during training.')
-    parser.add_argument(
-        '--memory_optimize',
-        action='store_true',
-        help='If set, optimize runtime memory before start.')
-    parser.add_argument(
-        '--use_fake_data',
-        action='store_true',
-        help='If set ommit the actual read data operators.')
-    parser.add_argument(
-        '--profile', action='store_true', help='If set, profile a few steps.')
-    parser.add_argument(
-        '--update_method',
-        type=str,
-        default='local',
-        choices=['local', 'pserver', 'nccl2'],
-        help='Choose parameter update method, can be local, pserver, nccl2.')
-    parser.add_argument(
-        '--use_reader_op',
-        action='store_true',
-        help='Whether to use reader op, and must specify the data path if set this to true.'
-    )
-    parser.add_argument(
-        '--data_path',
-        type=str,
-        default="",
-        help='Directory that contains all the training recordio files.')
-    args = parser.parse_args()
-    return args
+from args import *
 
 
 def append_nccl2_prepare(trainer_id):
@@ -160,7 +59,7 @@ def append_nccl2_prepare(trainer_id):
                         "nccl-based dist train.")
 
 
-def dist_transpile(trainer_id):
+def dist_transpile(trainer_id, args):
     if trainer_id < 0:
         return None, None
 
@@ -182,7 +81,12 @@ def dist_transpile(trainer_id):
     training_role = os.getenv("PADDLE_TRAINING_ROLE")
 
     t = distribute_transpiler.DistributeTranspiler()
-    t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers)
+    t.transpile(
+        trainer_id,
+        pservers=pserver_endpoints,
+        trainers=trainers,
+        sync_mode=not args.async_mode,
+        slice_var_up=not args.no_split_var)
     if training_role == "PSERVER":
         pserver_program = t.get_pserver_program(current_endpoint)
         pserver_startup_program = t.get_startup_program(current_endpoint,
@@ -417,7 +321,7 @@ def main():
         fluid.memory_optimize(fluid.default_main_program())
 
     if args.update_method == "pserver":
-        train_prog, startup_prog = dist_transpile(trainer_id)
+        train_prog, startup_prog = dist_transpile(trainer_id, args)
         if not train_prog:
             raise Exception(
                 "Must configure correct environments to run dist train.")

benchmark/fluid/models/stacked_dynamic_lstm.py

@@ -104,8 +104,9 @@ def get_model(args):
     loss = fluid.layers.mean(x=loss)
 
     # add acc
+    batch_size_tensor = fluid.layers.create_tensor(dtype='int64')
     batch_acc = fluid.layers.accuracy(input=logit, label=fluid.layers.data(name='label', \
-                shape=[1], dtype='int64'))
+                shape=[1], dtype='int64'), total=batch_size_tensor)
 
     inference_program = fluid.default_main_program().clone()
     with fluid.program_guard(inference_program):

benchmark/fluid/models/vgg.py

@@ -82,7 +82,8 @@ def get_model(args):
                 data_file, batch_size=args.batch_size))
         images, label = fluid.layers.read_file(data_file)
     else:
-        images = fluid.layers.data(name='data', shape=dshape, dtype='float32')
+        images = fluid.layers.data(
+            name='data', shape=data_shape, dtype='float32')
         label = fluid.layers.data(name='label', shape=[1], dtype='int64')
 
     # Train program

cmake/configure.cmake

@@ -166,3 +166,7 @@ if(WITH_GOLANG)
   endif()
 
 endif(WITH_GOLANG)
+
+if(WITH_GRPC)
+    add_definitions(-DPADDLE_WITH_GRPC)
+endif(WITH_GRPC)

cmake/external/brpc.cmake

+# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+INCLUDE(ExternalProject)
+
+SET(BRPC_SOURCES_DIR ${THIRD_PARTY_PATH}/brpc)
+SET(BRPC_INSTALL_DIR ${THIRD_PARTY_PATH}/install/brpc)
+SET(BRPC_INCLUDE_DIR "${BRPC_INSTALL_DIR}/include" CACHE PATH "brpc include directory." FORCE)
+SET(BRPC_LIBRARIES "${BRPC_INSTALL_DIR}/lib/libbrpc.a" CACHE FILEPATH "brpc library." FORCE)
+
+INCLUDE_DIRECTORIES(${BRPC_INCLUDE_DIR})
+
+# Reference https://stackoverflow.com/questions/45414507/pass-a-list-of-prefix-paths-to-externalproject-add-in-cmake-args
+set(prefix_path "${THIRD_PARTY_PATH}/install/gflags|${THIRD_PARTY_PATH}/install/leveldb|${THIRD_PARTY_PATH}/install/snappy|${THIRD_PARTY_PATH}/install/gtest|${THIRD_PARTY_PATH}/install/protobuf")
+
+# If minimal .a is need, you can set  WITH_DEBUG_SYMBOLS=OFF
+ExternalProject_Add(
+    extern_brpc
+    ${EXTERNAL_PROJECT_LOG_ARGS}
+    GIT_REPOSITORY  "https://github.com/brpc/brpc"
+    GIT_TAG         "6d153dd7ff00f960ae6895c9c5fff0ce9f07aff2"
+    PREFIX          ${BRPC_SOURCES_DIR}
+    UPDATE_COMMAND  ""
+    CMAKE_ARGS      -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
+                    -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
+                    -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
+                    -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
+                    -DCMAKE_INSTALL_PREFIX=${BRPC_INSTALL_DIR}
+                    -DCMAKE_INSTALL_LIBDIR=${BRPC_INSTALL_DIR}/lib
+                    -DCMAKE_POSITION_INDEPENDENT_CODE=ON
+                    -DCMAKE_BUILD_TYPE=${THIRD_PARTY_BUILD_TYPE}
+                    -DCMAKE_PREFIX_PATH=${prefix_path}
+                    -DBRPC_WITH_GLOG=ON
+                    ${EXTERNAL_OPTIONAL_ARGS}
+    LIST_SEPARATOR |
+    CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${BRPC_INSTALL_DIR}
+                     -DCMAKE_INSTALL_LIBDIR:PATH=${BRPC_INSTALL_DIR}/lib
+                     -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
+                     -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}
+)
+ADD_DEPENDENCIES(extern_brpc protobuf leveldb gflags glog gtest snappy)
+ADD_LIBRARY(brpc STATIC IMPORTED GLOBAL)
+SET_PROPERTY(TARGET brpc PROPERTY IMPORTED_LOCATION ${BRPC_LIBRARIES})
+ADD_DEPENDENCIES(brpc extern_brpc)
+
+
+LIST(APPEND external_project_dependencies brpc)

cmake/external/leveldb.cmake

+# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+INCLUDE(ExternalProject)
+
+SET(LEVELDB_SOURCES_DIR ${THIRD_PARTY_PATH}/leveldb)
+SET(LEVELDB_INSTALL_DIR ${THIRD_PARTY_PATH}/install/leveldb)
+SET(LEVELDB_INCLUDE_DIR "${LEVELDB_INSTALL_DIR}/include" CACHE PATH "leveldb include directory." FORCE)
+SET(LEVELDB_LIBRARIES "${LEVELDB_INSTALL_DIR}/lib/libleveldb.a" CACHE FILEPATH "leveldb library." FORCE)
+INCLUDE_DIRECTORIES(${LEVELDB_INCLUDE_DIR})
+
+ExternalProject_Add(
+    extern_leveldb
+    ${EXTERNAL_PROJECT_LOG_ARGS}
+    PREFIX ${LEVELDB_SOURCES_DIR}
+    URL "https://github.com/google/leveldb/archive/v1.18.tar.gz"
+    URL_MD5 "73770de34a2a5ab34498d2e05b2b7fa0"
+    CONFIGURE_COMMAND ""
+    BUILD_COMMAND CXXFLAGS=-fPIC make -j ${NUM_OF_PROCESSOR} libleveldb.a
+    INSTALL_COMMAND mkdir -p ${LEVELDB_INSTALL_DIR}/lib/ 
+        && cp ${LEVELDB_SOURCES_DIR}/src/extern_leveldb/libleveldb.a ${LEVELDB_LIBRARIES}
+        && cp -r ${LEVELDB_SOURCES_DIR}/src/extern_leveldb/include ${LEVELDB_INSTALL_DIR}/
+    BUILD_IN_SOURCE 1
+)
+
+ADD_DEPENDENCIES(extern_leveldb snappy)
+
+ADD_LIBRARY(leveldb STATIC IMPORTED GLOBAL)
+SET_PROPERTY(TARGET leveldb PROPERTY IMPORTED_LOCATION ${LEVELDB_LIBRARIES})
+ADD_DEPENDENCIES(leveldb extern_leveldb)
+
+LIST(APPEND external_project_dependencies leveldb)
+

cmake/generic.cmake

@@ -610,3 +610,21 @@ function(grpc_library TARGET_NAME)
     COMPILE_FLAGS  "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
   cc_library("${TARGET_NAME}" SRCS "${grpc_library_SRCS}" DEPS "${TARGET_NAME}_grpc" "${TARGET_NAME}_proto" "${grpc_library_DEPS}")
 endfunction()
+
+
+function(brpc_library TARGET_NAME)
+  set(oneValueArgs PROTO)
+  set(multiValueArgs SRCS DEPS)
+  set(options "")
+  cmake_parse_arguments(brpc_library "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+
+  message(STATUS "generating brpc ${brpc_library_PROTO}")
+
+  get_filename_component(ABS_PROTO ${brpc_library_PROTO} ABSOLUTE)
+  get_filename_component(PROTO_WE ${brpc_library_PROTO} NAME_WE)
+  get_filename_component(PROTO_PATH ${ABS_PROTO} PATH)
+
+  protobuf_generate_cpp(brpc_proto_srcs brpc_proto_hdrs "${ABS_PROTO}")
+  cc_library("${TARGET_NAME}_proto" SRCS "${brpc_proto_srcs}")
+  cc_library("${TARGET_NAME}" SRCS "${brpc_library_SRCS}" DEPS "${TARGET_NAME}_proto" "${brpc_library_DEPS}")
+endfunction()

doc/fluid/dev/api_doc_std_cn.md

 # API注释撰写标准
 
-- [API注释模块](#API注释模块)
-- [格式及示例](#格式及示例)
-- [完整示例](#完整示例)
+- [API注释撰写标准](#api)
+    - [API注释模块](#api)
+    - [格式及示例](#)
+    - [完整示例](#)
 
 
 ## API注释模块
@@ -217,4 +218,4 @@ API文档须使用reStructuredText格式撰写，该格式详情请参考[链接
 
 ## 完整示例
 
-fc 的完整注释见[示例](src/fc.py)。
+fc 的完整注释见[示例](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/dev/src/fc.py)。

doc/fluid/dev/api_doc_std_en.md

 # API Doc Standard
 
-- [API Doc Structure](#API Doc Structure)
-- [Format and Examples](#Format and Examples)
-- [Complete Example](#Complete Example)
+- [API Doc Standard](#api-doc-standard)
+    - [API Doc Structure](#api-doc-structure)
+    - [Format and Examples](#format-and-examples)
+    - [Complete Example](#complete-example)
 
 
 ## API Doc Structure
@@ -223,4 +224,4 @@ Format and examples of each part of API documantation are as follows: (take fc f
 
 ## Complete Example
 
-Complete Example of fc please see [here](src/fc.py)。
+Complete Example of fc please see [here](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/dev/src/fc.py)。

paddle/contrib/inference/test_paddle_inference_api_impl.cc

@@ -109,7 +109,6 @@ void MainWord2Vec(bool use_gpu) {
 
 void MainImageClassification(bool use_gpu) {
   int batch_size = 2;
-  bool use_mkldnn = false;
   bool repeat = false;
   NativeConfig config = GetConfig();
   config.use_gpu = use_gpu;
@@ -134,12 +133,8 @@ void MainImageClassification(bool use_gpu) {
   std::vector<framework::LoDTensor*> cpu_fetchs1;
   cpu_fetchs1.push_back(&output1);
 
-  TestInference<platform::CPUPlace, false, true>(config.model_dir,
-                                                 cpu_feeds,
-                                                 cpu_fetchs1,
-                                                 repeat,
-                                                 is_combined,
-                                                 use_mkldnn);
+  TestInference<platform::CPUPlace, false, true>(
+      config.model_dir, cpu_feeds, cpu_fetchs1, repeat, is_combined);
 
   auto predictor = CreatePaddlePredictor(config);
   std::vector<PaddleTensor> paddle_tensor_feeds;

paddle/fluid/framework/CMakeLists.txt

@@ -87,7 +87,7 @@ cc_library(executor SRCS executor.cc DEPS op_registry device_context scope
 framework_proto glog lod_rank_table feed_fetch_method)
 
 
-cc_library(parallel_executor SRCS parallel_executor.cc DEPS graph_builder_factory threaded_ssa_graph_executor scope_buffered_ssa_graph_executor)
+cc_library(parallel_executor SRCS parallel_executor.cc DEPS ssa_graph_builder_factory threaded_ssa_graph_executor scope_buffered_ssa_graph_executor)
 
 cc_library(prune SRCS prune.cc DEPS framework_proto)
 cc_test(prune_test SRCS prune_test.cc DEPS op_info prune recurrent_op device_context)

paddle/fluid/framework/data_type.cc

@@ -28,6 +28,9 @@ struct DataTypeMap {
 };
 
 static DataTypeMap* InitDataTypeMap();
+// C++11 removes the need for manual locking. Concurrent execution shall wait if
+// a static local variable is already being initialized.
+// https://stackoverflow.com/questions/11711920/how-to-implement-multithread-safe-singleton-in-c11-without-using-mutex
 static DataTypeMap& gDataTypeMap() {
   static DataTypeMap* g_data_type_map_ = InitDataTypeMap();
   return *g_data_type_map_;

paddle/fluid/framework/details/CMakeLists.txt

@@ -8,18 +8,19 @@ cc_library(rpc_op_handle SRCS rpc_op_handle.cc DEPS framework_proto scope place
 cc_library(ssa_graph SRCS ssa_graph.cc DEPS var_handle op_handle_base)
 cc_library(ssa_graph_builder SRCS ssa_graph_builder.cc DEPS ssa_graph)
 cc_library(ssa_graph_printer SRCS ssa_graph_printer.cc DEPS ssa_graph_builder)
+cc_library(ssa_graph_checker SRCS ssa_graph_checker.cc DEPS ssa_graph_builder)
 
 cc_library(variable_visitor SRCS variable_visitor.cc DEPS lod_tensor selected_rows)
 
 if(WITH_GPU)
-    nv_library(nccl_all_reduce_op_handle SRCS nccl_all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory
+    nv_library(all_reduce_op_handle SRCS all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory
             dynload_cuda variable_visitor)
-    set(multi_devices_graph_builder_deps nccl_all_reduce_op_handle)
     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)
 
 else()
-    set(multi_devices_graph_builder_deps)
+    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)
 endif()
@@ -28,10 +29,10 @@ cc_library(gather_op_handle SRCS gather_op_handle.cc DEPS op_handle_base scope d
 cc_library(fuse_vars_op_handle SRCS fuse_vars_op_handle.cc DEPS op_handle_base scope)
 
 cc_library(multi_devices_graph_builder SRCS multi_devices_graph_builder.cc DEPS ssa_graph_builder computation_op_handle
-        scale_loss_grad_op_handle rpc_op_handle ${multi_devices_graph_builder_deps} reduce_op_handle broadcast_op_handle)
+        scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle)
 
 
-cc_library(graph_builder_factory SRCS graph_builder_factory.cc DEPS multi_devices_graph_builder ssa_graph_printer)
+cc_library(ssa_graph_builder_factory SRCS ssa_graph_builder_factory.cc DEPS multi_devices_graph_builder ssa_graph_printer ssa_graph_checker)
 
 cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS ssa_graph framework_proto)
 cc_library(threaded_ssa_graph_executor SRCS threaded_ssa_graph_executor.cc DEPS fetch_op_handle ssa_graph_executor scope

paddle/fluid/framework/details/nccl_all_reduce_op_handle.cc → paddle/fluid/framework/details/all_reduce_op_handle.cc

@@ -13,25 +13,33 @@
 // limitations under the License.
 #include <algorithm>
 
+#include "paddle/fluid/framework/details/all_reduce_op_handle.h"
 #include "paddle/fluid/framework/details/container_cast.h"
-#include "paddle/fluid/framework/details/nccl_all_reduce_op_handle.h"
 #include "paddle/fluid/framework/details/reduce_and_gather.h"
 #include "paddle/fluid/framework/details/variable_visitor.h"
 
 namespace paddle {
 namespace framework {
 namespace details {
-NCCLAllReduceOpHandle::NCCLAllReduceOpHandle(
-    const std::vector<Scope *> &local_scopes,
-    const std::vector<platform::Place> &places,
-    const platform::NCCLContextMap &ctxs)
+
+#ifdef PADDLE_WITH_CUDA
+AllReduceOpHandle::AllReduceOpHandle(const std::vector<Scope *> &local_scopes,
+                                     const std::vector<platform::Place> &places,
+                                     const platform::NCCLContextMap *ctxs)
     : local_scopes_(local_scopes), places_(places), nccl_ctxs_(ctxs) {
-  for (auto &p : places_) {
-    this->dev_ctxes_[p] = nccl_ctxs_.DevCtx(p);
+  if (nccl_ctxs_) {
+    for (auto &p : places_) {
+      this->dev_ctxes_[p] = nccl_ctxs_->DevCtx(p);
+    }
   }
 }
+#else
+AllReduceOpHandle::AllReduceOpHandle(const std::vector<Scope *> &local_scopes,
+                                     const std::vector<platform::Place> &places)
+    : local_scopes_(local_scopes), places_(places) {}
+#endif
 
-void NCCLAllReduceOpHandle::RunImpl() {
+void AllReduceOpHandle::RunImpl() {
   if (NoDummyInputSize() == 1) {
     return;  // No need to all reduce when GPU count = 1;
   } else {
@@ -58,6 +66,8 @@ void NCCLAllReduceOpHandle::RunImpl() {
     }
 
     if (platform::is_gpu_place(lod_tensors[0]->place())) {
+#ifdef PADDLE_WITH_CUDA
+      PADDLE_ENFORCE(nccl_ctxs_, "nccl_ctxs should not be nullptr.");
       int dtype = -1;
       size_t numel = 0;
       std::vector<std::function<void()>> all_reduce_calls;
@@ -75,7 +85,7 @@ void NCCLAllReduceOpHandle::RunImpl() {
         }
 
         int dev_id = boost::get<platform::CUDAPlace>(p).device;
-        auto &nccl_ctx = nccl_ctxs_.at(dev_id);
+        auto &nccl_ctx = nccl_ctxs_->at(dev_id);
         auto stream = nccl_ctx.stream();
         auto comm = nccl_ctx.comm_;
         all_reduce_calls.emplace_back([=] {
@@ -90,22 +100,25 @@ void NCCLAllReduceOpHandle::RunImpl() {
           call();
         }
       });
+#else
+      PADDLE_THROW("Not compiled with CUDA");
+#endif
     } else {  // Special handle CPU only Operator's gradient. Like CRF
       auto &trg = *this->local_scopes_[0]
                        ->FindVar(kLocalExecScopeName)
                        ->Get<Scope *>()
-                       ->Var()
+                       ->FindVar(out_var_handles[0]->name_)
                        ->GetMutable<framework::LoDTensor>();
 
       // Reduce All Tensor to trg in CPU
       ReduceLoDTensor func(lod_tensors, &trg);
       VisitDataType(ToDataType(lod_tensors[0]->type()), func);
 
-      for (size_t i = 0; i < local_scopes_.size(); ++i) {
+      for (size_t i = 1; i < local_scopes_.size(); ++i) {
         auto &scope =
             *local_scopes_[i]->FindVar(kLocalExecScopeName)->Get<Scope *>();
         auto &p = places_[i];
-        auto *var = scope.FindVar(in_var_handles[i]->name_);
+        auto *var = scope.FindVar(out_var_handles[i]->name_);
         auto *dev_ctx = dev_ctxes_[p];
 
         RunAndRecordEvent(p, [&trg, var, dev_ctx, p] {
@@ -118,7 +131,7 @@ void NCCLAllReduceOpHandle::RunImpl() {
   }
 }
 
-std::string NCCLAllReduceOpHandle::Name() const { return "nccl_all_reduce"; }
+std::string AllReduceOpHandle::Name() const { return "all_reduce"; }
 }  // namespace details
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/details/nccl_all_reduce_op_handle.h → paddle/fluid/framework/details/all_reduce_op_handle.h

@@ -20,17 +20,23 @@
 #include "paddle/fluid/framework/details/op_handle_base.h"
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/scope.h"
+#ifdef PADDLE_WITH_CUDA
 #include "paddle/fluid/platform/nccl_helper.h"
+#endif
 
 namespace paddle {
 namespace framework {
 namespace details {
 
-struct NCCLAllReduceOpHandle : public OpHandleBase {
-  NCCLAllReduceOpHandle(const std::vector<Scope *> &local_scopes,
-                        const std::vector<platform::Place> &places,
-                        const platform::NCCLContextMap &ctxs);
-
+struct AllReduceOpHandle : public OpHandleBase {
+#ifdef PADDLE_WITH_CUDA
+  AllReduceOpHandle(const std::vector<Scope *> &local_scopes,
+                    const std::vector<platform::Place> &places,
+                    const platform::NCCLContextMap *ctxs);
+#else
+  AllReduceOpHandle(const std::vector<Scope *> &local_scopes,
+                    const std::vector<platform::Place> &places);
+#endif
   std::string Name() const override;
 
   // Delay and buffer nccl_all_reduce together can significantly increase
@@ -43,7 +49,9 @@ struct NCCLAllReduceOpHandle : public OpHandleBase {
  private:
   std::vector<Scope *> local_scopes_;
   std::vector<platform::Place> places_;
-  const platform::NCCLContextMap &nccl_ctxs_;
+#ifdef PADDLE_WITH_CUDA
+  const platform::NCCLContextMap *nccl_ctxs_;
+#endif
 };
 
 }  // namespace details

paddle/fluid/framework/details/execution_strategy.h

@@ -20,7 +20,7 @@ namespace details {
 
 struct ExecutionStrategy {
   size_t num_threads_{0};
-  bool use_event_{true};
+  bool use_cuda_{true};
   bool allow_op_delay_{false};
   size_t num_iteration_per_drop_scope_{100};
 };

paddle/fluid/framework/details/multi_devices_graph_builder.cc

@@ -17,6 +17,7 @@
 #include <utility>
 #include <vector>
 
+#include "paddle/fluid/framework/details/all_reduce_op_handle.h"
 #include "paddle/fluid/framework/details/broadcast_op_handle.h"
 #include "paddle/fluid/framework/details/computation_op_handle.h"
 #include "paddle/fluid/framework/details/multi_devices_graph_builder.h"
@@ -26,10 +27,6 @@
 #include "paddle/fluid/framework/op_info.h"
 #include "paddle/fluid/framework/scope.h"
 
-#ifdef PADDLE_WITH_CUDA
-#include "paddle/fluid/framework/details/nccl_all_reduce_op_handle.h"
-#endif
-
 namespace paddle {
 namespace framework {
 namespace details {
@@ -89,7 +86,7 @@ std::vector<std::string> MultiDevSSAGraphBuilder::FindDistTrainSendVars(
   for (auto *op : program.Block(0).AllOps()) {
     // TODO(Yancey1989): use a graceful method to find send op,
     // instead of the the hard code string
-    if (op->Type() == "send_vars") {
+    if (op->Type() == "send") {
       auto op_vars = op->InputArgumentNames();
       send_vars.reserve(send_vars.size() +
                         std::distance(op_vars.begin(), op_vars.end()));
@@ -243,7 +240,7 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
                     CreateReduceOp(&result, g_name, 0);
                     CreateBroadcastOp(&result, g_name, 0);
                   } else {
-                    InsertNCCLAllReduceOp(&result, g_name);
+                    InsertAllReduceOp(&result, g_name);
                   }
                   break;
               }
@@ -286,6 +283,19 @@ bool MultiDevSSAGraphBuilder::IsSparseGradient(
   return false;
 }
 
+void MultiDevSSAGraphBuilder::SetCommunicationContext(
+    OpHandleBase *op_handle, const platform::Place &p) const {
+#ifdef PADDLE_WITH_CUDA
+  if (nccl_ctxs_ == nullptr) {
+    op_handle->SetDeviceContext(p,
+                                platform::DeviceContextPool::Instance().Get(p));
+  }
+#else
+  op_handle->SetDeviceContext(p,
+                              platform::DeviceContextPool::Instance().Get(p));
+#endif
+}
+
 void MultiDevSSAGraphBuilder::CreateBroadcastOp(SSAGraph *result,
                                                 const std::string &p_name,
                                                 size_t src_dev_id) const {
@@ -300,15 +310,12 @@ void MultiDevSSAGraphBuilder::CreateBroadcastOp(SSAGraph *result,
   op_handle->AddInput(in);
 
   for (size_t i = 0; i < places_.size(); ++i) {
-    auto &vars = result->vars_.at(i).at(p_name);
     auto &p = places_[i];
+    SetCommunicationContext(op_handle, p);
+    auto &vars = result->vars_.at(i).at(p_name);
     auto *out_var = new VarHandle(vars.size(), i, p_name, p);
     vars.emplace_back(out_var);
     op_handle->AddOutput(out_var);
-#ifndef ADDLE_WITH_CUDA
-    op_handle->SetDeviceContext(p,
-                                platform::DeviceContextPool::Instance().Get(p));
-#endif
   }
 }
 
@@ -320,15 +327,19 @@ void MultiDevSSAGraphBuilder::CreateComputationalOp(SSAGraph *result,
   CreateOpHandleIOs(result, op, dev_id);
 }
 
-void MultiDevSSAGraphBuilder::InsertNCCLAllReduceOp(
-    SSAGraph *result, const std::string &og) const {
+void MultiDevSSAGraphBuilder::InsertAllReduceOp(SSAGraph *result,
+                                                const std::string &og) const {
 #ifdef PADDLE_WITH_CUDA
   result->ops_.emplace_back(
-      new NCCLAllReduceOpHandle(local_scopes_, places_, *nccl_ctxs_));
+      new AllReduceOpHandle(local_scopes_, places_, nccl_ctxs_));
+#else
+  result->ops_.emplace_back(new AllReduceOpHandle(local_scopes_, places_));
+#endif
   auto *op_handle = result->ops_.back().get();
 
   for (size_t i = 0; i < places_.size(); ++i) {
     auto &p = places_[i];
+    SetCommunicationContext(op_handle, p);
     auto &vars = result->vars_[i][og];
     PADDLE_ENFORCE(!vars.empty());
     auto &prev_grad = vars.back();
@@ -338,9 +349,6 @@ void MultiDevSSAGraphBuilder::InsertNCCLAllReduceOp(
     vars.emplace_back(var);
     op_handle->AddOutput(var);
   }
-#else
-  PADDLE_ENFORCE("Not implemented");
-#endif
 }
 
 bool MultiDevSSAGraphBuilder::IsParameterGradientOnce(
@@ -379,7 +387,9 @@ void MultiDevSSAGraphBuilder::CreateScaleLossGradOp(SSAGraph *result) const {
   for (size_t i = 0; i < places_.size(); ++i) {
 // Insert ScaleCost OpHandle
 #ifdef PADDLE_WITH_CUDA
-    auto *communication_dev_ctx = nccl_ctxs_->DevCtx(places_[i]);
+    auto *communication_dev_ctx =
+        nccl_ctxs_ ? nccl_ctxs_->DevCtx(places_[i])
+                   : platform::DeviceContextPool::Instance().Get(places_[i]);
 #else
     auto *communication_dev_ctx =
         platform::DeviceContextPool::Instance().Get(platform::CPUPlace());
@@ -424,12 +434,9 @@ VarHandle *MultiDevSSAGraphBuilder::CreateReduceOp(SSAGraph *result,
   auto *op_handle = result->ops_.back().get();
 
   for (size_t i = 0; i < places_.size(); ++i) {
-    auto &vars = result->vars_[i][og];
-#ifndef PADDLE_WITH_CUDA
     auto &p = places_[i];
-    op_handle->SetDeviceContext(p,
-                                platform::DeviceContextPool::Instance().Get(p));
-#endif
+    SetCommunicationContext(op_handle, p);
+    auto &vars = result->vars_[i][og];
     PADDLE_ENFORCE(!vars.empty());
     auto &prev_grad = vars.back();
     op_handle->AddInput(prev_grad.get());
@@ -468,17 +475,17 @@ void MultiDevSSAGraphBuilder::CreateRPCOp(SSAGraph *result,
       new RPCOpHandle(op, local_scopes_[0], op.Type(), places_[0]));
 
   if (op.Type() == "send_barrier") {
-    ConnectOp(result, result->ops_.back().get(), "send_vars");
+    ConnectOp(result, result->ops_.back().get(), "send");
   } else if (op.Type() == "recv") {
     ConnectOp(result, result->ops_.back().get(), "send_barrier");
   } else if (op.Type() == "fetch_barrier") {
     ConnectOp(result, result->ops_.back().get(), "recv");
-  } else if (op.Type() == "send_vars") {
+  } else if (op.Type() == "send") {
     // do nothing
   } else {
     PADDLE_THROW(
         "rpc op should be in ["
-        "send_vars, send_barrier. recv, fetch_barrier]");
+        "send, send_barrier. recv, fetch_barrier]");
   }
 
   // TODO(Yancey1989): schedule rpc op on different place may

paddle/fluid/framework/details/multi_devices_graph_builder.h

@@ -100,7 +100,7 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
       const std::vector<std::unordered_set<std::string>> &var_name_on_devices,
       const OpDesc &op) const;
 
-  void InsertNCCLAllReduceOp(SSAGraph *result, const std::string &og) const;
+  void InsertAllReduceOp(SSAGraph *result, const std::string &og) const;
 
   void CreateBroadcastOp(SSAGraph *result, const std::string &p_name,
                          size_t src_dev_id) const;
@@ -111,6 +111,9 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
 
  private:
   BuildStrategy strategy_;
+
+  void SetCommunicationContext(OpHandleBase *op_handle,
+                               const platform::Place &p) const;
 };
 }  // namespace details
 }  // namespace framework

paddle/fluid/framework/details/op_handle_base.cc

@@ -39,9 +39,9 @@ OpHandleBase::~OpHandleBase() {
 #endif
 }
 
-void OpHandleBase::Run(bool use_event) {
+void OpHandleBase::Run(bool use_cuda) {
 #ifdef PADDLE_WITH_CUDA
-  if (events_.empty() && use_event) {
+  if (events_.empty() && use_cuda) {
     for (auto &p : dev_ctxes_) {
       int dev_id = boost::get<platform::CUDAPlace>(p.first).device;
       PADDLE_ENFORCE(cudaSetDevice(dev_id));
@@ -50,7 +50,7 @@ void OpHandleBase::Run(bool use_event) {
     }
   }
 #else
-  PADDLE_ENFORCE(!use_event);
+  PADDLE_ENFORCE(!use_cuda);
 #endif
 
   RunImpl();

paddle/fluid/framework/details/op_handle_base.h

@@ -36,7 +36,7 @@ class OpHandleBase {
 
   virtual std::string Name() const = 0;
 
-  void Run(bool use_event);
+  void Run(bool use_cuda);
 
   virtual void RecordWaitEventOnCtx(platform::DeviceContext *waited_ctx);
 

paddle/fluid/framework/details/reduce_and_gather.h

@@ -37,7 +37,9 @@ struct ReduceLoDTensor {
     PADDLE_ENFORCE_NE(t0.numel(), 0);
     dst_tensor_.Resize(t0.dims());
     T *dst = dst_tensor_.mutable_data<T>(platform::CPUPlace());
-    std::copy(t0.data<T>(), t0.data<T>() + t0.numel(), dst);
+    if (dst != t0.data<T>()) {
+      std::copy(t0.data<T>(), t0.data<T>() + t0.numel(), dst);
+    }
 
     for (size_t i = 1; i < src_tensors_.size(); ++i) {
       auto &t = *src_tensors_[i];

paddle/fluid/framework/details/ssa_graph_builder.cc

@@ -11,8 +11,8 @@
 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 // See the License for the specific language governing permissions and
 // limitations under the License.
-
 #include "paddle/fluid/framework/details/ssa_graph_builder.h"
+#include <utility>
 
 namespace paddle {
 namespace framework {

paddle/fluid/framework/details/graph_builder_factory.cc → paddle/fluid/framework/details/ssa_graph_builder_factory.cc

@@ -12,9 +12,10 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/fluid/framework/details/graph_builder_factory.h"
+#include "paddle/fluid/framework/details/ssa_graph_builder_factory.h"
 #include <fstream>
 #include "paddle/fluid/framework/details/multi_devices_graph_builder.h"
+#include "paddle/fluid/framework/details/ssa_graph_checker.h"
 #include "paddle/fluid/framework/details/ssa_graph_printer.h"
 
 namespace paddle {
@@ -40,6 +41,8 @@ std::unique_ptr<SSAGraphBuilder> SSAGraphBuilderFactory::Create() {
     res.reset(new SSAGraghBuilderWithPrinter(
         std::move(fout), std::move(graphviz_printer), std::move(res)));
   }
+  res.reset(new SSAGraghBuilderWithChecker(std::move(res)));
+
   return res;
 }
 }  // namespace details

paddle/fluid/framework/details/graph_builder_factory.h → paddle/fluid/framework/details/ssa_graph_builder_factory.h

@@ -40,7 +40,11 @@ class SSAGraphBuilderFactory {
         loss_var_name_(loss_var_name),
         param_names_(param_names),
         local_scopes_(local_scopes),
-        strategy_(strategy) {}
+        strategy_(strategy) {
+#ifdef PADDLE_WITH_CUDA
+    nccl_ctxs_ = nullptr;
+#endif
+  }
 
 #ifdef PADDLE_WITH_CUDA
   void SetNCCLContextMap(platform::NCCLContextMap* nccl_ctxs) {

paddle/fluid/framework/details/ssa_graph_checker.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/ssa_graph.h"
+#include <string>
+#include "paddle/fluid/framework/details/ssa_graph_checker.h"
+
+namespace paddle {
+namespace framework {
+namespace details {
+
+bool SSAGraghBuilderWithChecker::IsValidGraph(const SSAGraph *graph) const {
+  std::unordered_map<OpHandleBase *, size_t> pending_ops;
+  std::unordered_set<VarHandleBase *> pending_vars;
+  std::unordered_set<VarHandleBase *> ready_vars;
+  std::unordered_set<OpHandleBase *> ready_ops;
+
+  auto insert_pending_var = [&](VarHandleBase *var) {
+    pending_vars.insert(var);
+    if (var->generated_op_ == nullptr) {
+      ready_vars.emplace(var);
+    }
+  };
+
+  for (auto &var_map : graph->vars_) {
+    for (auto &name_pair : var_map) {
+      for (auto &version_pair : name_pair.second) {
+        insert_pending_var(version_pair.get());
+      }
+    }
+  }
+
+  for (auto &var : graph->dep_vars_) {
+    insert_pending_var(var.get());
+  }
+
+  for (auto &op : graph->ops_) {
+    if (op->Inputs().empty()) {
+      ready_ops.insert(op.get());
+    } else {
+      pending_ops.insert({op.get(), op.get()->NoDupInputSize()});
+    }
+  }
+
+  auto run_all_ops = [&](std::unordered_set<OpHandleBase *> &set) {
+    for (auto *op : set) {
+      for (auto out : op->Outputs()) {
+        ready_vars.emplace(out);
+      }
+    }
+    set.clear();
+  };
+
+  while (!pending_vars.empty()) {
+    run_all_ops(ready_ops);
+
+    if (ready_vars.empty()) {
+      return false;
+    }
+
+    for (auto ready_var : ready_vars) {
+      pending_vars.erase(ready_var);
+      for (auto *op : ready_var->pending_ops_) {
+        auto &deps = --pending_ops[op];
+        if (deps == 0) {
+          ready_ops.insert(op);
+        }
+      }
+    }
+    ready_vars.clear();
+  }
+  return true;
+}
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/details/ssa_graph_checker.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/details/ssa_graph_builder.h"
+
+namespace paddle {
+namespace framework {
+namespace details {
+class SSAGraph;
+
+class SSAGraghBuilderWithChecker : public SSAGraphBuilder {
+ public:
+  explicit SSAGraghBuilderWithChecker(
+      std::unique_ptr<SSAGraphBuilder>&& builder)
+      : builder_(std::move(builder)) {}
+
+  std::unique_ptr<SSAGraph> Build(const ProgramDesc& program) const override {
+    auto graph = builder_->Build(program);
+    PADDLE_ENFORCE(IsValidGraph(graph.get()));
+    return graph;
+  }
+
+  bool IsValidGraph(const SSAGraph* graph) const;
+
+ private:
+  std::unique_ptr<SSAGraphBuilder> builder_;
+};
+
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/details/threaded_ssa_graph_executor.cc

@@ -185,6 +185,7 @@ void ThreadedSSAGraphExecutor::InsertPendingVar(
     ready_vars->Push(var);
   }
 }
+
 void ThreadedSSAGraphExecutor::RunOp(
     BlockingQueue<VarHandleBase *> *ready_var_q, details::OpHandleBase *op) {
   auto op_run = [ready_var_q, op, this] {
@@ -192,7 +193,7 @@ void ThreadedSSAGraphExecutor::RunOp(
       if (VLOG_IS_ON(10)) {
         VLOG(10) << op << " " << op->Name() << " : " << op->DebugString();
       }
-      op->Run(strategy_.use_event_);
+      op->Run(strategy_.use_cuda_);
       VLOG(10) << op << " " << op->Name() << " Done ";
       running_ops_--;
       ready_var_q->Extend(op->Outputs());

paddle/fluid/framework/executor.cc

@@ -24,6 +24,7 @@ limitations under the License. */
 #include "paddle/fluid/platform/profiler.h"
 
 DECLARE_bool(benchmark);
+DEFINE_bool(use_mkldnn, false, "Use MKLDNN to run");
 
 namespace paddle {
 namespace framework {
@@ -115,6 +116,7 @@ void Executor::CreateVariables(const ProgramDesc& pdesc, Scope* scope,
 void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
                    bool create_local_scope, bool create_vars) {
   platform::RecordBlock b(block_id);
+  if (FLAGS_use_mkldnn) EnableMKLDNN(pdesc);
   auto ctx = Prepare(pdesc, block_id);
   RunPreparedContext(ctx.get(), scope, create_local_scope, create_vars);
 }
@@ -214,6 +216,7 @@ void Executor::Run(const ProgramDesc& program, Scope* scope,
                    const std::string& feed_holder_name,
                    const std::string& fetch_holder_name) {
   platform::RecordBlock b(kProgramId);
+  if (FLAGS_use_mkldnn) EnableMKLDNN(program);
   bool has_feed_ops =
       has_feed_operators(program.Block(0), *feed_targets, feed_holder_name);
   bool has_fetch_ops =
@@ -225,7 +228,6 @@ void Executor::Run(const ProgramDesc& program, Scope* scope,
     unique_ptr_of_copy_program.reset(new ProgramDesc(program));
     copy_program = unique_ptr_of_copy_program.get();
   }
-
   auto* global_block = copy_program->MutableBlock(0);
 
   if (!has_feed_ops) {
@@ -382,5 +384,19 @@ void Executor::RunPreparedContext(
   }
 }
 
+void Executor::EnableMKLDNN(const ProgramDesc& program) {
+#ifdef PADDLE_WITH_MKLDNN
+  VLOG(3) << "use_mkldnn=True";
+  for (size_t bid = 0; bid < program.Size(); ++bid) {
+    auto* block = const_cast<ProgramDesc&>(program).MutableBlock(bid);
+    for (auto* op : block->AllOps()) {
+      if (op->HasAttr("use_mkldnn")) {
+        op->SetAttr("use_mkldnn", true);
+      }
+    }
+  }
+#endif
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/executor.h

@@ -81,6 +81,8 @@ class Executor {
                           const std::string& feed_holder_name = "feed",
                           const std::string& fetch_holder_name = "fetch");
 
+  void EnableMKLDNN(const ProgramDesc& program);
+
  private:
   const platform::Place place_;
 };

paddle/fluid/framework/framework.proto

@@ -71,6 +71,7 @@ message OpProto {
     optional bool duplicable = 3 [ default = false ];
     optional bool intermediate = 4 [ default = false ];
     optional bool dispensable = 5 [ default = false ];
+    optional string reuse = 6;
   }
 
   // AttrProto describes the C++ type Attribute.

paddle/fluid/framework/op_info.cc

@@ -17,12 +17,11 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-static OpInfoMap* g_op_info_map = nullptr;
-
+// C++11 removes the need for manual locking. Concurrent execution shall wait if
+// a static local variable is already being initialized.
+// https://stackoverflow.com/questions/11711920/how-to-implement-multithread-safe-singleton-in-c11-without-using-mutex
 OpInfoMap& OpInfoMap::Instance() {
-  if (g_op_info_map == nullptr) {
-    g_op_info_map = new OpInfoMap();
-  }
+  static OpInfoMap* g_op_info_map = new OpInfoMap();
   return *g_op_info_map;
 }
 }  // namespace framework

paddle/fluid/framework/op_proto_maker.cc

@@ -21,6 +21,7 @@ namespace framework {
 void OpProtoAndCheckerMaker::Validate() {
   validated_ = true;
   CheckNoDuplicatedInOutAttrs();
+  CheckReuseVars();
 }
 
 OpProtoAndCheckerMaker::VariableBuilder OpProtoAndCheckerMaker::AddInput(
@@ -56,6 +57,24 @@ void OpProtoAndCheckerMaker::CheckNoDuplicatedInOutAttrs() {
   }
 }
 
+void OpProtoAndCheckerMaker::CheckReuseVars() {
+  std::unordered_set<std::string> names;
+  for (auto& input : proto_->inputs()) {
+    names.insert(input.name());
+  }
+  auto checker = [&](const std::string& name, const std::string& reused) {
+    PADDLE_ENFORCE(
+        names.count(reused),
+        "Output [%s] reuse Input [%s], but the input is not registered.", name,
+        reused);
+  };
+  for (auto& output : proto_->outputs()) {
+    if (output.has_reuse()) {
+      checker(output.name(), output.reuse());
+    }
+  }
+}
+
 void OpProtoAndCheckerMaker::operator()(proto::OpProto* proto,
                                         OpAttrChecker* attr_checker) {
   proto_ = proto;

paddle/fluid/framework/op_proto_maker.h

@@ -14,6 +14,8 @@ limitations under the License. */
 #pragma once
 
 #include <string>
+#include <unordered_set>
+
 #include "glog/logging.h"
 #include "paddle/fluid/framework/attribute.h"
 #include "paddle/fluid/framework/framework.pb.h"
@@ -64,6 +66,11 @@ class OpProtoAndCheckerMaker {
       var_->set_dispensable(true);
       return *this;
     }
+
+    VariableBuilder &Reuse(const std::string &name) {
+      var_->set_reuse(name);
+      return *this;
+    }
   };
 
   VariableBuilder AddInput(const std::string &name, const std::string &comment);
@@ -89,6 +96,8 @@ class OpProtoAndCheckerMaker {
   void CheckNoDuplicatedInOutAttrs();
   void Validate();
 
+  void CheckReuseVars();
+
   proto::OpProto *proto_;
   OpAttrChecker *op_checker_;
   bool validated_{false};

paddle/fluid/framework/op_proto_maker_test.cc

@@ -47,3 +47,23 @@ TEST(ProtoMaker, DuplicatedInOut) {
   ASSERT_THROW(proto_maker(&op_proto, &op_checker),
                paddle::platform::EnforceNotMet);
 }
+
+class TestInplaceProtoMaker : public paddle::framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() {
+    AddInput("X", "input of test op");
+    AddOutput("XOut", "output of test op").Reuse("X");
+    AddOutput("NoOut", "output of test op").Reuse("NotExists");
+  }
+};
+
+TEST(ProtoMaker, InplaceOutput) {
+  paddle::framework::proto::OpProto op_proto;
+  paddle::framework::OpAttrChecker op_checker;
+  TestInplaceProtoMaker proto_maker;
+  ASSERT_THROW(proto_maker(&op_proto, &op_checker),
+               paddle::platform::EnforceNotMet);
+  // proto_maker(&op_proto, &op_checker);
+  // proto_maker.Make();
+  // ASSERT_THROW(proto_maker.Validate(), paddle::platform::EnforceNotMet);
+}

paddle/fluid/framework/parallel_executor.cc

@@ -22,8 +22,8 @@ limitations under the License. */
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif
 
-#include "paddle/fluid/framework/details/graph_builder_factory.h"
 #include "paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h"
+#include "paddle/fluid/framework/details/ssa_graph_builder_factory.h"
 #include "paddle/fluid/framework/details/threaded_ssa_graph_executor.h"
 #include "paddle/fluid/platform/profiler.h"
 
@@ -43,7 +43,8 @@ class ParallelExecutorPrivate {
 #ifdef PADDLE_WITH_CUDA
   std::unique_ptr<platform::NCCLContextMap> nccl_ctxs_;
 #endif
-  bool own_local_scope;
+  bool own_local_scope_;
+  bool use_cuda_;
 };
 
 std::vector<Scope *> &ParallelExecutor::GetLocalScopes() {
@@ -60,35 +61,40 @@ ParallelExecutor::ParallelExecutor(
     size_t num_trainers, size_t trainer_id)
     : member_(new ParallelExecutorPrivate(places)) {
   member_->global_scope_ = scope;
+  member_->use_cuda_ = exec_strategy.use_cuda_;
 
   // Step 1. Bcast the params to devs.
   // Create local scopes
   if (local_scopes.empty()) {
-    member_->own_local_scope = true;
+    member_->own_local_scope_ = true;
     member_->local_scopes_.emplace_back(member_->global_scope_);
     for (size_t i = 1; i < member_->places_.size(); ++i) {
       member_->local_scopes_.emplace_back(&scope->NewScope());
     }
   } else {
-    member_->own_local_scope = false;
+    member_->own_local_scope_ = false;
     PADDLE_ENFORCE_EQ(member_->places_.size(), local_scopes.size());
     for (size_t i = 0; i < member_->places_.size(); ++i) {
       member_->local_scopes_.emplace_back(&local_scopes[i]->NewScope());
     }
   }
 
+  if (member_->use_cuda_) {
 // Bcast Parameters to all GPUs
 #ifdef PADDLE_WITH_CUDA
-  auto *nccl_id_var = scope->FindVar(NCCL_ID_VARNAME);
-  ncclUniqueId *nccl_id = nullptr;
-  if (nccl_id_var != nullptr) {
-    nccl_id = nccl_id_var->GetMutable<ncclUniqueId>();
-  }
-  member_->nccl_ctxs_.reset(new platform::NCCLContextMap(
-      member_->places_, nccl_id, num_trainers, trainer_id));
+    auto *nccl_id_var = scope->FindVar(NCCL_ID_VARNAME);
+    ncclUniqueId *nccl_id = nullptr;
+    if (nccl_id_var != nullptr) {
+      nccl_id = nccl_id_var->GetMutable<ncclUniqueId>();
+    }
+    member_->nccl_ctxs_.reset(new platform::NCCLContextMap(
+        member_->places_, nccl_id, num_trainers, trainer_id));
+#else
+    PADDLE_THROW("Not compiled with CUDA");
 #endif
-  if (platform::is_gpu_place(places[0]) && member_->local_scopes_.size() != 1 &&
-      local_scopes.empty()) {  // Is CUDA
+  }
+
+  if (member_->local_scopes_.size() != 1 && local_scopes.empty()) {
     BCastParamsToGPUs(bcast_vars);
   }
   // Startup Program has been run. All local scopes has correct parameters.
@@ -108,9 +114,13 @@ ParallelExecutor::ParallelExecutor(
   details::SSAGraphBuilderFactory builder_factory(
       member_->places_, loss_var_name, params, member_->local_scopes_,
       build_strategy);
+  if (member_->use_cuda_) {
 #ifdef PADDLE_WITH_CUDA
-  builder_factory.SetNCCLContextMap(member_->nccl_ctxs_.get());
+    builder_factory.SetNCCLContextMap(member_->nccl_ctxs_.get());
+#else
+    PADDLE_THROW("Not compiled with CUDA");
 #endif
+  }
 
   member_->executor_.reset(new details::ThreadedSSAGraphExecutor(
       exec_strategy, member_->local_scopes_, places,
@@ -123,7 +133,6 @@ ParallelExecutor::ParallelExecutor(
 
 void ParallelExecutor::BCastParamsToGPUs(
     const std::unordered_set<std::string> &vars) const {
-#ifdef PADDLE_WITH_CUDA
   auto *main_scope = member_->local_scopes_[0];
 
   for (auto &var : vars) {
@@ -135,6 +144,7 @@ void ParallelExecutor::BCastParamsToGPUs(
     auto &main_tensor = main_var->Get<LoDTensor>();
     auto &dims = main_tensor.dims();
     if (paddle::platform::is_gpu_place(main_tensor.place())) {
+#ifdef PADDLE_WITH_CUDA
       size_t numel = main_tensor.numel();
       ncclDataType_t data_type = platform::ToNCCLDataType(main_tensor.type());
       platform::NCCLGroupGuard guard;
@@ -153,6 +163,10 @@ void ParallelExecutor::BCastParamsToGPUs(
         platform::dynload::ncclBcast(buffer, numel, data_type, 0,
                                      nccl_ctx.comm_, nccl_ctx.stream());
       }
+      member_->nccl_ctxs_->WaitAll();
+#else
+      PADDLE_THROW("Not compiled with CUDA");
+#endif
     } else {
       platform::CPUPlace cpu;
       for (size_t i = 1; i < member_->places_.size(); ++i) {
@@ -163,11 +177,7 @@ void ParallelExecutor::BCastParamsToGPUs(
         paddle::framework::TensorCopy(main_tensor, cpu, t);
       }
     }
-    member_->nccl_ctxs_->WaitAll();
   }
-#else
-  PADDLE_THROW("Not compiled with CUDA");
-#endif
 }
 
 void ParallelExecutor::Run(const std::vector<std::string> &fetch_tensors,
@@ -213,7 +223,7 @@ void ParallelExecutor::FeedAndSplitTensorIntoLocalScopes(
 }
 
 ParallelExecutor::~ParallelExecutor() {
-  if (member_->own_local_scope) {
+  if (member_->own_local_scope_) {
     for (size_t i = 1; i < member_->local_scopes_.size(); ++i) {
       member_->global_scope_->DeleteScope(member_->local_scopes_[i]);
     }

paddle/fluid/inference/tensorrt/convert/ut_helper.h

@@ -64,7 +64,8 @@ class TRTConvertValidation {
 
   TRTConvertValidation(int batch_size,
                        const std::unordered_set<std::string>& parameters,
-                       framework::Scope& scope, int workspace_size = 1 << 10)
+                       framework::Scope& scope,  // NOLINT
+                       int workspace_size = 1 << 10)
       : parameters_(parameters), scope_(scope) {
     // create engine.
     engine_.reset(new TensorRTEngine(10, 1 << 10, &stream_));

paddle/fluid/inference/tests/book/test_inference_image_classification.cc

@@ -21,7 +21,6 @@ DEFINE_string(fp16_dirname, "", "Directory of the float16 inference model.");
 DEFINE_int32(batch_size, 1, "Batch size of input data");
 DEFINE_int32(repeat, 1, "Running the inference program repeat times");
 DEFINE_bool(skip_cpu, false, "Skip the cpu test");
-DEFINE_bool(use_mkldnn, false, "Use MKLDNN to run inference");
 
 TEST(inference, image_classification) {
   if (FLAGS_dirname.empty() || FLAGS_batch_size < 1 || FLAGS_repeat < 1) {
@@ -59,10 +58,8 @@ TEST(inference, image_classification) {
     // Run inference on CPU
     LOG(INFO) << "--- CPU Runs: ---";
     LOG(INFO) << "Batch size is " << FLAGS_batch_size;
-    LOG(INFO) << "FLAGS_use_mkldnn: " << FLAGS_use_mkldnn;
     TestInference<paddle::platform::CPUPlace, false, true>(
-        dirname, cpu_feeds, cpu_fetchs1, FLAGS_repeat, is_combined,
-        FLAGS_use_mkldnn);
+        dirname, cpu_feeds, cpu_fetchs1, FLAGS_repeat, is_combined);
     LOG(INFO) << output1.dims();
   }
 

paddle/fluid/inference/tests/book/test_inference_nlp.cc

@@ -27,7 +27,6 @@ limitations under the License. */
 DEFINE_string(model_path, "", "Directory of the inference model.");
 DEFINE_string(data_file, "", "File of input index data.");
 DEFINE_int32(repeat, 100, "Running the inference program repeat times");
-DEFINE_bool(use_mkldnn, false, "Use MKLDNN to run inference");
 DEFINE_bool(prepare_vars, true, "Prepare variables before executor");
 DEFINE_int32(num_threads, 1, "Number of threads should be used");
 
@@ -190,9 +189,6 @@ TEST(inference, nlp) {
     std::unique_ptr<paddle::framework::ProgramDesc> inference_program;
     inference_program = InitProgram(&executor, scope.get(), FLAGS_model_path,
                                     /*model combined*/ false);
-    if (FLAGS_use_mkldnn) {
-      EnableMKLDNN(inference_program);
-    }
     // always prepare context
     std::unique_ptr<paddle::framework::ExecutorPrepareContext> ctx;
     ctx = executor.Prepare(*inference_program, 0);

paddle/fluid/inference/tests/test_helper.h

@@ -22,6 +22,8 @@ limitations under the License. */
 #include "paddle/fluid/inference/io.h"
 #include "paddle/fluid/platform/profiler.h"
 
+DECLARE_bool(use_mkldnn);
+
 template <typename T>
 void SetupTensor(paddle::framework::LoDTensor* input,
                  paddle::framework::DDim dims, T lower, T upper) {
@@ -133,24 +135,11 @@ std::vector<std::vector<int64_t>> GetFeedTargetShapes(
   return feed_target_shapes;
 }
 
-void EnableMKLDNN(
-    const std::unique_ptr<paddle::framework::ProgramDesc>& program) {
-  for (size_t bid = 0; bid < program->Size(); ++bid) {
-    auto* block = program->MutableBlock(bid);
-    for (auto* op : block->AllOps()) {
-      if (op->HasAttr("use_mkldnn")) {
-        op->SetAttr("use_mkldnn", true);
-      }
-    }
-  }
-}
-
 template <typename Place, bool CreateVars = true, bool PrepareContext = false>
 void TestInference(const std::string& dirname,
                    const std::vector<paddle::framework::LoDTensor*>& cpu_feeds,
                    const std::vector<paddle::framework::LoDTensor*>& cpu_fetchs,
-                   const int repeat = 1, const bool is_combined = false,
-                   const bool use_mkldnn = false) {
+                   const int repeat = 1, const bool is_combined = false) {
   // 1. Define place, executor, scope
   auto place = Place();
   auto executor = paddle::framework::Executor(place);
@@ -182,9 +171,6 @@ void TestInference(const std::string& dirname,
         "init_program",
         paddle::platform::DeviceContextPool::Instance().Get(place));
     inference_program = InitProgram(&executor, scope, dirname, is_combined);
-    if (use_mkldnn) {
-      EnableMKLDNN(inference_program);
-    }
   }
   // Disable the profiler and print the timing information
   paddle::platform::DisableProfiler(paddle::platform::EventSortingKey::kDefault,
@@ -210,7 +196,10 @@ void TestInference(const std::string& dirname,
     fetch_targets[fetch_target_names[i]] = cpu_fetchs[i];
   }
 
-  // 6. Run the inference program
+  // 6. If export Flags_use_mkldnn=True, use mkldnn related ops.
+  if (FLAGS_use_mkldnn) executor.EnableMKLDNN(*inference_program);
+
+  // 7. Run the inference program
   {
     if (!CreateVars) {
       // If users don't want to create and destroy variables every time they

paddle/fluid/operators/CMakeLists.txt

@@ -186,19 +186,23 @@ endif()
 
 add_subdirectory(detail)
 if(WITH_DISTRIBUTE)
-    
-    set(DISTRIBUTE_DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf)
+
+    set(DISTRIBUTE_DEPS "")
+    if(WITH_GRPC)
+        set(DISTRIBUTE_DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf)
+    else()
+        set(DISTRIBUTE_DEPS sendrecvop_brpc brpc leveldb snappystream snappy protobuf ssl crypto zlib)
+    endif()
+
     set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
-    op_library(send_op DEPS ${DISTRIBUTE_DEPS})
-    set_source_files_properties(send_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
     op_library(prefetch_op DEPS ${DISTRIBUTE_DEPS})
     set_source_files_properties(prefetch_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
     op_library(recv_op DEPS ${DISTRIBUTE_DEPS})
     set_source_files_properties(recv_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
     op_library(listen_and_serv_op DEPS ${DISTRIBUTE_DEPS})
     set_source_files_properties(listen_and_serv_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
-    op_library(send_vars_op DEPS ${DISTRIBUTE_DEPS})
-    set_source_files_properties(send_vars_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
+    op_library(send_op DEPS ${DISTRIBUTE_DEPS})
+    set_source_files_properties(send_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
     op_library(send_barrier_op DEPS ${DISTRIBUTE_DEPS})
     op_library(fetch_barrier_op DEPS ${DISTRIBUTE_DEPS})
     set_source_files_properties(send_barrier_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
@@ -208,15 +212,18 @@ if(WITH_DISTRIBUTE)
     #        listen_and_serv_op sum_op executor SERIAL)
     if(WITH_GPU)
         set_source_files_properties(test_send_nccl_id.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
-        cc_test(test_send_nccl_id SRCS test_send_nccl_id.cc DEPS send_op
-                listen_and_serv_op executor SERIAL)
-        op_library(gen_nccl_id_op DEPS nccl_common sendrecvop_grpc)
+        cc_test(test_send_nccl_id SRCS test_send_nccl_id.cc DEPS listen_and_serv_op executor SERIAL)
+        if(WITH_GRPC)
+            op_library(gen_nccl_id_op DEPS nccl_common sendrecvop_grpc)
+        else()
+            op_library(gen_nccl_id_op DEPS nccl_common sendrecvop_brpc)
+        endif()
         set_source_files_properties(gen_nccl_id_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
     else()
         set(DEPS_OPS ${DEPS_OPS} gen_nccl_id_op)
     endif()
 else()
-    set(DEPS_OPS ${DEPS_OPS} send_op prefetch_op recv_op listen_and_serv_op send_vars_op send_barrier_op fetch_barrier_op gen_nccl_id_op)
+    set(DEPS_OPS ${DEPS_OPS}  prefetch_op recv_op listen_and_serv_op send_op send_barrier_op fetch_barrier_op gen_nccl_id_op)
 endif()
 
 op_library(cross_entropy_op DEPS cross_entropy)

paddle/fluid/operators/activation_op.cc

@@ -25,7 +25,7 @@ namespace operators {
    public:                                                              \
     void Make() override {                                              \
       AddInput("X", "Input of " #OP_NAME " operator");                  \
-      AddOutput("Out", "Output of " #OP_NAME " operator");              \
+      AddOutput("Out", "Output of " #OP_NAME " operator").Reuse("X");   \
       AddAttr<bool>("use_mkldnn",                                       \
                     "(bool, default false) Only used in mkldnn kernel") \
           .SetDefault(false);                                           \

paddle/fluid/operators/adam_op.cc

@@ -89,9 +89,9 @@ class AdamOpMaker : public framework::OpProtoAndCheckerMaker {
     AddInput("Beta1Pow", "(Tensor) Input beta1 power accumulator");
     AddInput("Beta2Pow", "(Tensor) Input beta2 power accumulator");
 
-    AddOutput("ParamOut", "(Tensor) Output parameter");
-    AddOutput("Moment1Out", "(Tensor) Output first moment");
-    AddOutput("Moment2Out", "(Tensor) Output second moment");
+    AddOutput("ParamOut", "(Tensor) Output parameter").Reuse("Param");
+    AddOutput("Moment1Out", "(Tensor) Output first moment").Reuse("Moment1");
+    AddOutput("Moment2Out", "(Tensor) Output second moment").Reuse("Moment2");
 
     AddAttr<float>("beta1",
                    "(float, default 0.9) "

paddle/fluid/operators/arg_max_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/arg_min_max_op_base.h"
+
+REGISTER_OPERATOR(arg_max, paddle::operators::ArgMinMaxOp,
+                  paddle::operators::ArgMaxOpMaker,
+                  paddle::framework::EmptyGradOpMaker);
+
+REGISTER_OP_CPU_KERNEL(
+    arg_max,
+    paddle::operators::ArgMaxKernel<paddle::platform::CPUDeviceContext, float>,
+    paddle::operators::ArgMaxKernel<paddle::platform::CPUDeviceContext, double>,
+    paddle::operators::ArgMaxKernel<paddle::platform::CPUDeviceContext,
+                                    int64_t>,
+    paddle::operators::ArgMaxKernel<paddle::platform::CPUDeviceContext,
+                                    int32_t>,
+    paddle::operators::ArgMaxKernel<paddle::platform::CPUDeviceContext,
+                                    int16_t>,
+    paddle::operators::ArgMaxKernel<paddle::platform::CPUDeviceContext, size_t>,
+    paddle::operators::ArgMaxKernel<paddle::platform::CPUDeviceContext,
+                                    uint8_t>);

paddle/fluid/operators/arg_max_op.cu

+/* 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/arg_min_max_op_base.h"
+
+REGISTER_OP_CUDA_KERNEL(
+    arg_max,
+    paddle::operators::ArgMaxKernel<paddle::platform::CUDADeviceContext, float>,
+    paddle::operators::ArgMaxKernel<paddle::platform::CUDADeviceContext,
+                                    double>,
+    paddle::operators::ArgMaxKernel<paddle::platform::CUDADeviceContext,
+                                    int64_t>,
+    paddle::operators::ArgMaxKernel<paddle::platform::CUDADeviceContext,
+                                    int32_t>,
+    paddle::operators::ArgMaxKernel<paddle::platform::CUDADeviceContext,
+                                    int16_t>,
+    paddle::operators::ArgMaxKernel<paddle::platform::CUDADeviceContext,
+                                    size_t>,
+    paddle::operators::ArgMaxKernel<paddle::platform::CUDADeviceContext,
+                                    uint8_t>);

paddle/fluid/operators/arg_min_max_op_base.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 <type_traits>
+#include <vector>
+#include "paddle/fluid/framework/ddim.h"
+#include "paddle/fluid/framework/eigen.h"
+#include "paddle/fluid/framework/lod_tensor.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/operator.h"
+#include "paddle/fluid/platform/enforce.h"
+#include "paddle/fluid/string/printf.h"
+
+namespace paddle {
+namespace operators {
+
+enum ArgMinMaxType { kArgMin, kArgMax };
+
+template <typename DeviceContext, typename T, typename Tout, int64_t Rank,
+          ArgMinMaxType argMinMaxValue>
+struct ArgMinMaxFunctor {};
+
+#define DECLARE_ARG_MIN_MAX_FUNCTOR(eigen_op_type, enum_argminmax_value)      \
+  template <typename DeviceContext, typename T, typename Tout, int64_t Rank>  \
+  struct ArgMinMaxFunctor<DeviceContext, T, Tout, Rank,                       \
+                          enum_argminmax_value> {                             \
+    void operator()(const DeviceContext& ctx, const framework::LoDTensor& in, \
+                    framework::LoDTensor* out, int64_t axis) {                \
+      auto in_eigen = framework::EigenTensor<T, Rank>::From(in);              \
+      auto out_eigen = framework::EigenTensor<Tout, Rank - 1>::From(*out);    \
+      out_eigen.device(*(ctx.eigen_device())) =                               \
+          in_eigen.eigen_op_type(axis).template cast<Tout>();                 \
+    }                                                                         \
+  }
+
+DECLARE_ARG_MIN_MAX_FUNCTOR(argmin, ArgMinMaxType::kArgMin);
+DECLARE_ARG_MIN_MAX_FUNCTOR(argmax, ArgMinMaxType::kArgMax);
+
+template <typename DeviceContext, typename T, typename Tout,
+          ArgMinMaxType EnumArgMinMaxValue>
+class ArgMinMaxKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto& x = *(ctx.Input<framework::LoDTensor>("X"));
+    auto& out = *(ctx.Output<framework::LoDTensor>("Out"));
+    out.mutable_data<Tout>(ctx.GetPlace());
+    auto axis = ctx.Attr<int64_t>("axis");
+    auto& dev_ctx = ctx.template device_context<DeviceContext>();
+
+#define CALL_ARG_MINMAX_FUNCTOR(rank)                                \
+  ArgMinMaxFunctor<DeviceContext, T, Tout, rank, EnumArgMinMaxValue> \
+      functor##rank;                                                 \
+  functor##rank(dev_ctx, x, &out, axis)
+
+    switch (x.dims().size()) {
+      case 1:
+        CALL_ARG_MINMAX_FUNCTOR(1);
+        break;
+      case 2:
+        CALL_ARG_MINMAX_FUNCTOR(2);
+        break;
+      case 3:
+        CALL_ARG_MINMAX_FUNCTOR(3);
+        break;
+      case 4:
+        CALL_ARG_MINMAX_FUNCTOR(4);
+        break;
+      case 5:
+        CALL_ARG_MINMAX_FUNCTOR(5);
+        break;
+      case 6:
+        CALL_ARG_MINMAX_FUNCTOR(6);
+        break;
+      default:
+        PADDLE_THROW(
+            "%s operator doesn't supports tensors whose ranks are greater "
+            "than 6.",
+            (EnumArgMinMaxValue == kArgMin ? "argmin" : "argmax"));
+        break;
+#undef CALL_ARG_MINMAX_FUNCTOR
+    }
+  }
+};
+
+template <typename DeviceContext, typename T>
+using ArgMinKernel =
+    ArgMinMaxKernel<DeviceContext, T, int64_t, ArgMinMaxType::kArgMin>;
+
+template <typename DeviceContext, typename T>
+using ArgMaxKernel =
+    ArgMinMaxKernel<DeviceContext, T, int64_t, ArgMinMaxType::kArgMax>;
+
+class ArgMinMaxOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) should not be null");
+    const auto& x_dims = ctx->GetInputDim("X");
+    int64_t axis = ctx->Attrs().Get<int64_t>("axis");
+    PADDLE_ENFORCE(axis >= -x_dims.size() && axis < x_dims.size(),
+                   "'axis' must be inside [-Rank(X), Rank(X))");
+
+    auto x_rank = x_dims.size();
+    if (axis < 0) axis += x_rank;
+
+    std::vector<int64_t> vec;
+    for (int64_t i = 0; i < axis; i++) vec.push_back(x_dims[i]);
+    for (int64_t i = axis + 1; i < x_rank; i++) vec.push_back(x_dims[i]);
+    ctx->SetOutputDim("Out", framework::make_ddim(vec));
+  }
+};
+
+class BaseArgMinMaxOpMaker : public framework::OpProtoAndCheckerMaker {
+ protected:
+  virtual const char* OpName() const = 0;
+  virtual const char* Name() const = 0;
+
+ public:
+  void Make() override {
+    AddInput("X", "Input tensor.");
+    AddOutput("Out", "Output tensor.");
+    AddAttr<int64_t>("axis", "The axis in which to compute the arg indics.");
+    AddComment(string::Sprintf(R"DOC(
+      %s Operator.
+
+      Computes the indices of the %s elements of the input tensor's element
+      along the provided axis.
+)DOC",
+                               OpName(), Name()));
+  }
+};
+
+class ArgMinOpMaker : public BaseArgMinMaxOpMaker {
+ protected:
+  const char* OpName() const override { return "ArgMin"; }
+  const char* Name() const override { return "min"; }
+};
+
+class ArgMaxOpMaker : public BaseArgMinMaxOpMaker {
+ protected:
+  const char* OpName() const override { return "ArgMax"; }
+  const char* Name() const override { return "max"; }
+};
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/arg_min_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/arg_min_max_op_base.h"
+
+REGISTER_OPERATOR(arg_min, paddle::operators::ArgMinMaxOp,
+                  paddle::operators::ArgMinOpMaker,
+                  paddle::framework::EmptyGradOpMaker);
+
+REGISTER_OP_CPU_KERNEL(
+    arg_min,
+    paddle::operators::ArgMinKernel<paddle::platform::CPUDeviceContext, float>,
+    paddle::operators::ArgMinKernel<paddle::platform::CPUDeviceContext, double>,
+    paddle::operators::ArgMinKernel<paddle::platform::CPUDeviceContext,
+                                    int64_t>,
+    paddle::operators::ArgMinKernel<paddle::platform::CPUDeviceContext,
+                                    int32_t>,
+    paddle::operators::ArgMinKernel<paddle::platform::CPUDeviceContext,
+                                    int16_t>,
+    paddle::operators::ArgMinKernel<paddle::platform::CPUDeviceContext, size_t>,
+    paddle::operators::ArgMinKernel<paddle::platform::CPUDeviceContext,
+                                    uint8_t>);

paddle/fluid/operators/arg_min_op.cu

+/* 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/arg_min_max_op_base.h"
+
+REGISTER_OP_CUDA_KERNEL(
+    arg_min,
+    paddle::operators::ArgMinKernel<paddle::platform::CUDADeviceContext, float>,
+    paddle::operators::ArgMinKernel<paddle::platform::CUDADeviceContext,
+                                    double>,
+    paddle::operators::ArgMinKernel<paddle::platform::CUDADeviceContext,
+                                    int64_t>,
+    paddle::operators::ArgMinKernel<paddle::platform::CUDADeviceContext,
+                                    int32_t>,
+    paddle::operators::ArgMinKernel<paddle::platform::CUDADeviceContext,
+                                    int16_t>,
+    paddle::operators::ArgMinKernel<paddle::platform::CUDADeviceContext,
+                                    size_t>,
+    paddle::operators::ArgMinKernel<paddle::platform::CUDADeviceContext,
+                                    uint8_t>);

paddle/fluid/operators/batch_norm_op.cc

@@ -151,13 +151,15 @@ class BatchNormOpMaker : public framework::OpProtoAndCheckerMaker {
     AddInput("Variance",
              "The global variance (for training) "
              "or estimated Variance (for testing)");
-    AddOutput("Y", "result after normalization");
+    AddOutput("Y", "result after normalization").Reuse("X");
     AddOutput("MeanOut",
               "Share memory with Mean. "
-              "Store the global mean when training");
+              "Store the global mean when training")
+        .Reuse("Mean");
     AddOutput("VarianceOut",
               "Share memory with Variance. "
-              "Store the global Variance when training");
+              "Store the global Variance when training")
+        .Reuse("Variance");
     AddOutput("SavedMean",
               "Mean of the current mini batch, "
               "will apply to output when training")

paddle/fluid/operators/batch_size_like.h

@@ -54,18 +54,18 @@ class BatchSizeLikeOp : public framework::OperatorWithKernel {
 class BatchSizeLikeOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() final {
-    AddInput("Input",
-             "(Tensor) Tensor "
-             "whose input_dim_idx'th dimension specifies the batch_size");
+    AddInput(
+        "Input",
+        "Tensor whose input_dim_idx'th dimension specifies the batch_size");
     AddOutput("Out",
-              "(Tensor) Tensor of specified shape will be filled "
+              "Tensor of specified shape will be filled "
               "with the specified value");
-    AddAttr<std::vector<int>>("shape", "(vector<int>) The shape of the output");
+    AddAttr<std::vector<int>>("shape", "The shape of the output");
     AddAttr<int>("input_dim_idx",
-                 "(int, default 0) The index of input's batch size dimension")
+                 "default 0. The index of input's batch size dimension")
         .SetDefault(0);
     AddAttr<int>("output_dim_idx",
-                 "(int, default 0) The index of output's batch size dimension")
+                 "default 0. The index of output's batch size dimension")
         .SetDefault(0);
     Apply();
   }

paddle/fluid/operators/bilinear_interp_op.cc

@@ -56,17 +56,16 @@ class BilinearInterpOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() override {
     AddInput("X",
-             "(Tensor) The input tensor of bilinear interpolation, "
+             "The input tensor of bilinear interpolation, "
              "This is a 4-D tensor with shape of (N x C x h x w)");
     AddInput("OutSize",
-             "(Tensor) This is a 1-D tensor with two number. "
+             "This is a 1-D tensor with two number. "
              "The first number is height and the second number is width.")
         .AsDispensable();
-    AddOutput("Out",
-              "(Tensor) The dimension of output is (N x C x out_h x out_w]");
+    AddOutput("Out", "The dimension of output is (N x C x out_h x out_w)");
 
-    AddAttr<int>("out_h", "(int) output height of bilinear interpolation op.");
-    AddAttr<int>("out_w", "(int) output width of bilinear interpolation op.");
+    AddAttr<int>("out_h", "output height of bilinear interpolation op.");
+    AddAttr<int>("out_w", "output width of bilinear interpolation op.");
     AddComment(R"DOC(
           Bilinear interpolation is an extension of linear interpolation for 
           interpolating functions of two variables (e.g. H-direction and 

paddle/fluid/operators/conv_op.cc

@@ -125,7 +125,8 @@ void Conv2DOpMaker::Make() {
            "input image channels divided by the groups.");
   AddOutput("Output",
             "(Tensor) The output tensor of convolution operator. "
-            "The format of output tensor is also NCHW.");
+            "The format of output tensor is also NCHW.")
+      .Reuse("Input");
   AddAttr<std::vector<int>>("strides",
                             "(vector<int> default:{1, 1}), the "
                             "strides(h_stride, w_stride) of "
@@ -220,7 +221,8 @@ void Conv3DOpMaker::Make() {
            "input image channels divided by the groups.");
   AddOutput("Output",
             "(Tensor) The output tensor of convolution operator."
-            "The format of output tensor is also NCDHW.");
+            "The format of output tensor is also NCDHW.")
+      .Reuse("Input");
   AddAttr<std::vector<int>>("strides",
                             "(vector<int>, default:{1, 1, 1}), the "
                             "strides(d_stride, h_stride, w_stride) of "

paddle/fluid/operators/cross_entropy_op.cc

@@ -124,7 +124,8 @@ class CrossEntropyOpMaker : public framework::OpProtoAndCheckerMaker {
              "Tensor<float/double> with shape [N x D].");
     AddOutput("Y",
               "(Tensor, default Tensor<float>), a 2-D tensor with shape "
-              "[N x 1]. The cross entropy loss.");
+              "[N x 1]. The cross entropy loss.")
+        .Reuse("X");
     AddAttr<bool>("soft_label",
                   "(bool, default false), a flag indicating whether to "
                   "interpretate the given labels as soft labels.")

paddle/fluid/operators/detail/CMakeLists.txt

-if(WITH_DISTRIBUTE)
+if(NOT WITH_DISTRIBUTE)
+    return()
+endif()
+
+
+if(WITH_GRPC)
   grpc_library(sendrecvop_grpc SRCS bytebuffer_stream.cc sendrecvop_utils.cc grpc_client.cc
       request_handler_impl.cc rpc_client.cc rpc_server.cc grpc_server.cc variable_response.cc PROTO send_recv.proto DEPS lod_tensor
       selected_rows memory)
   set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
-  set_source_files_properties(serde_test.cc grpc_server_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
-  cc_test(serde_test SRCS serde_test.cc variable_response.cc DEPS grpc++_unsecure grpc_unsecure gpr
+  set_source_files_properties(grpc_serde_test.cc rpc_server_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
+  cc_test(serde_test SRCS grpc_serde_test.cc variable_response.cc DEPS grpc++_unsecure grpc_unsecure gpr
           cares zlib protobuf sendrecvop_grpc SERIAL)
-  cc_test(grpc_server_test SRCS grpc_server_test.cc DEPS sendrecvop_grpc
+  cc_test(grpc_server_test SRCS rpc_server_test.cc DEPS sendrecvop_grpc
           grpc++_unsecure grpc_unsecure gpr cares zlib protobuf executor
           proto_desc lookup_table_op SERIAL)
+  return()
 endif()
+
+
+set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
+set_source_files_properties(brpc_server.cc brpc_client.cc rpc_server_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
+brpc_library(sendrecvop_brpc SRCS brpc_client.cc brpc_server.cc rpc_server.cc  rpc_client.cc request_handler_impl.cc
+  PROTO send_recv.proto
+  DEPS lod_tensor selected_rows memory)
+
+find_library(OPENSSL_CRYPTO_LIBRARY_STATIC NAMES libcrypto.so)
+ADD_LIBRARY(crypto SHARED IMPORTED GLOBAL)
+SET_PROPERTY(TARGET crypto PROPERTY IMPORTED_LOCATION ${OPENSSL_CRYPTO_LIBRARY_STATIC})
+
+
+find_library(OPENSSL_SSL_LIBRARY_STATIC NAMES libssl.so)
+ADD_LIBRARY(ssl SHARED IMPORTED GLOBAL)
+SET_PROPERTY(TARGET ssl PROPERTY IMPORTED_LOCATION ${OPENSSL_SSL_LIBRARY_STATIC})
+
+cc_test(brpc_server_test SRCS rpc_server_test.cc DEPS sendrecvop_brpc 
+       brpc protobuf leveldb gflags glog
+       protobuf executor proto_desc lookup_table_op snappystream snappy ssl crypto SERIAL)

paddle/fluid/operators/detail/brpc_client.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/detail/brpc_client.h"
+#include "paddle/fluid/framework/threadpool.h"
+
+namespace paddle {
+namespace operators {
+namespace detail {
+
+DEFINE_int32(brpc_channel_num, 24,
+             "Number of channels to send requests connected to one server");
+DEFINE_int32(timeout_ms, 30000, "RPC timeout in milliseconds");
+DEFINE_int32(max_retry, 3, "Max retries(not including the first RPC)");
+
+BRPCClient::~BRPCClient() { Wait(); }
+
+void HandleSendResponse(brpc::Controller* cntl,
+                        sendrecv::VoidMessage* response) {
+  // std::unique_ptr makes sure cntl/response will be deleted before returning.
+  std::unique_ptr<brpc::Controller> cntl_guard(cntl);
+  std::unique_ptr<sendrecv::VoidMessage> response_guard(response);
+
+  if (cntl->Failed()) {
+    LOG(WARNING) << "Fail to send EchoRequest, " << cntl->ErrorText();
+    return;
+  }
+  LOG(INFO) << "Received response from " << cntl->remote_side()
+            << " latency=" << cntl->latency_us() << "us";
+}
+
+bool BRPCClient::AsyncSendVar(const std::string& ep,
+                              const platform::DeviceContext& ctx,
+                              const framework::Scope& scope,
+                              const std::string& var_name, int64_t time_out) {
+  const platform::DeviceContext* p_ctx = &ctx;
+  const std::string ep_val = ep;
+  const std::string var_name_val = var_name;
+  const framework::Scope* p_scope = &scope;
+  const auto ch_ptr = GetChannel(ep_val);
+
+  framework::AsyncIO(
+      [var_name_val, p_ctx, ep_val, p_scope, time_out, ch_ptr, this] {
+        auto ch_ctx = ch_ptr->Pop();
+        brpc::Controller* cntl = new brpc::Controller();
+        sendrecv::VoidMessage* response = new sendrecv::VoidMessage();
+        cntl->set_timeout_ms(time_out);
+
+        google::protobuf::Closure* done =
+            brpc::NewCallback(&HandleSendResponse, cntl, response);
+
+        sendrecv::VariableMessage request;
+        ch_ctx->stub->SendVariable(cntl, &request, response, done);
+      });
+  req_count_++;
+
+  return true;
+}
+
+void HandleGetResponse(brpc::Controller* cntl,
+                       sendrecv::VariableMessage* response) {
+  // std::unique_ptr makes sure cntl/response will be deleted before returning.
+  std::unique_ptr<brpc::Controller> cntl_guard(cntl);
+  std::unique_ptr<sendrecv::VariableMessage> response_guard(response);
+
+  if (cntl->Failed()) {
+    LOG(WARNING) << "Fail to send EchoRequest, " << cntl->ErrorText();
+    return;
+  }
+  LOG(INFO) << "Received response from " << cntl->remote_side()
+            << " latency=" << cntl->latency_us() << "us";
+
+  // framework::Variable* outvar = nullptr;
+  // DeserializeFromByteBuffer(ret_msg, *var_h.ctx, var_h.scope, &outvar);
+}
+
+bool BRPCClient::AsyncGetVar(const std::string& ep,
+                             const platform::DeviceContext& ctx,
+                             const framework::Scope& scope,
+                             const std::string& var_name, int64_t time_out) {
+  const platform::DeviceContext* p_ctx = &ctx;
+  const std::string ep_val = ep;
+  const std::string var_name_val = var_name;
+  const framework::Scope* p_scope = &scope;
+  const auto ch = GetChannel(ep_val);
+
+  framework::AsyncIO(
+      [var_name_val, ep_val, p_scope, p_ctx, time_out, ch, this] {});
+
+  req_count_++;
+
+  return true;
+}
+
+bool BRPCClient::AsyncPrefetchVar(const std::string& ep,
+                                  const platform::DeviceContext& ctx,
+                                  const framework::Scope& scope,
+                                  const std::string& in_var_name,
+                                  const std::string& out_var_name,
+                                  int64_t time_out) {
+  const platform::DeviceContext* p_ctx = &ctx;
+  const std::string ep_val = ep;
+  const std::string in_var_name_val = in_var_name;
+  const std::string out_var_name_val = out_var_name;
+  const framework::Scope* p_scope = &scope;
+  const auto ch = GetChannel(ep_val);
+
+  framework::AsyncIO([in_var_name_val, out_var_name_val, ep_val, p_scope, p_ctx,
+                      time_out, ch, this] {});
+
+  req_count_++;
+  return true;
+}
+
+void BRPCClient::AsyncSendBatchBarrier(const std::string& ep,
+                                       int64_t time_out) {
+  req_count_++;
+}
+
+void BRPCClient::AsyncSendFetchBarrier(const std::string& ep,
+                                       int64_t time_out) {
+  req_count_++;
+}
+
+void BRPCClient::Wait() {
+  std::unique_lock<std::mutex> lk(sync_mutex_);
+  sync_cond_.wait(lk, [this] { return req_count_ == 0; });
+}
+
+ChannelQueuePtr BRPCClient::GetChannel(const std::string& ep) {
+  {
+    std::lock_guard<std::mutex> guard(chan_mutex_);
+    auto it = channels_.find(ep);
+    if (it != channels_.end()) {
+      return it->second;
+    }
+  }
+
+  ChannelQueuePtr q(new framework::BlockingQueue<ChannelContextPtr>());
+
+  brpc::ChannelOptions options;
+  options.protocol = "baidu_std";
+  options.connection_type = "pooled";
+  options.connect_timeout_ms = 100;
+  options.timeout_ms = FLAGS_timeout_ms /*milliseconds*/;
+  options.max_retry = FLAGS_max_retry;
+  for (int i = 0; i < FLAGS_brpc_channel_num; ++i) {
+    std::shared_ptr<ChannelContext> c(new ChannelContext());
+    if (c->channel.Init(ep.c_str(), &options) != 0) {
+      LOG(ERROR) << "Fail to initialize channel";
+      return nullptr;
+    }
+
+    c->stub.reset(new sendrecv::SendRecvService_Stub(
+        static_cast<google::protobuf::RpcChannel*>(&c->channel)));
+    q->Push(c);
+  }
+
+  {
+    std::lock_guard<std::mutex> guard(chan_mutex_);
+    channels_[ep] = q;
+  }
+
+  return q;
+}
+
+}  // namespace detail
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/detail/brpc_client.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include <time.h>
+
+#include <chrono>  // NOLINT
+#include <ctime>
+#include <functional>
+#include <iostream>
+#include <map>
+#include <mutex>  // NOLINT
+#include <string>
+#include <vector>
+
+#include "brpc/channel.h"
+#include "paddle/fluid/framework/blocking_queue.h"
+#include "paddle/fluid/framework/data_type.h"
+#include "paddle/fluid/framework/lod_tensor.h"
+#include "paddle/fluid/framework/scope.h"
+#include "paddle/fluid/framework/selected_rows.h"
+#include "paddle/fluid/operators/detail/rpc_client.h"
+#include "paddle/fluid/operators/detail/send_recv.pb.h"
+#include "paddle/fluid/platform/macros.h"  // for DISABLE_COPY_AND_ASSIGN
+
+namespace paddle {
+namespace operators {
+namespace detail {
+
+struct ChannelContext {
+  brpc::Channel channel;
+  std::shared_ptr<sendrecv::SendRecvService_Stub> stub;
+};
+
+typedef std::shared_ptr<ChannelContext> ChannelContextPtr;
+typedef std::shared_ptr<framework::BlockingQueue<ChannelContextPtr>>
+    ChannelQueuePtr;
+
+class BRPCClient : public RPCClient {
+ public:
+  BRPCClient() {}
+  virtual ~BRPCClient();
+
+  bool AsyncSendVar(const std::string& ep, const platform::DeviceContext& ctx,
+                    const framework::Scope& scope, const std::string& var_name,
+                    int64_t time_out = RPCClient::rpc_time_out) override;
+
+  bool AsyncGetVar(const std::string& ep, const platform::DeviceContext& ctx,
+                   const framework::Scope& scope, const std::string& var_name,
+                   int64_t time_out = RPCClient::rpc_time_out) override;
+
+  bool AsyncPrefetchVar(const std::string& ep,
+                        const platform::DeviceContext& ctx,
+                        const framework::Scope& scope,
+                        const std::string& in_var_name,
+                        const std::string& out_var_name,
+                        int64_t time_out = RPCClient::rpc_time_out) override;
+
+  void AsyncSendBatchBarrier(
+      const std::string& ep,
+      int64_t time_out = RPCClient::rpc_time_out) override;
+
+  void AsyncSendFetchBarrier(
+      const std::string& ep,
+      int64_t time_out = RPCClient::rpc_time_out) override;
+
+  void Wait() override;
+
+ private:
+  void Proceed();
+  ChannelQueuePtr GetChannel(const std::string& ep);
+
+ private:
+  std::unordered_map<std::string, ChannelQueuePtr> channels_;
+
+  // mutex for Wait client sync
+  std::mutex sync_mutex_;
+  std::condition_variable sync_cond_;
+  std::atomic<int64_t> req_count_{0};
+
+  // mutex for GetChannel thread safety
+  std::mutex chan_mutex_;
+  DISABLE_COPY_AND_ASSIGN(BRPCClient);
+};
+
+}  // namespace detail
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/detail/brpc_server.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/detail/brpc_server.h"
+#include "paddle/fluid/operators/detail/request_handler.h"
+
+namespace sendrecv {
+
+typedef std::unordered_map<std::string,
+                           paddle::operators::detail::RequestHandler*>
+    HandlerMap;
+
+class BRPCServiceImpl : public SendRecvService {
+ public:
+  explicit BRPCServiceImpl(const HandlerMap& rpc_call_map)
+      : request_send_h_(nullptr),
+        request_get_h_(nullptr),
+        request_prefetch_h_(nullptr) {
+    auto it = rpc_call_map.find(paddle::operators::detail::kRequestSend);
+    if (it != rpc_call_map.end()) {
+      request_send_h_ = it->second;
+    }
+
+    it = rpc_call_map.find(paddle::operators::detail::kRequestSend);
+    if (it != rpc_call_map.end()) {
+      request_get_h_ = it->second;
+    }
+
+    it = rpc_call_map.find(paddle::operators::detail::kRequestPrefetch);
+    if (it != rpc_call_map.end()) {
+      request_prefetch_h_ = it->second;
+    }
+  }
+
+  virtual ~BRPCServiceImpl() {}
+
+  void SendVariable(google::protobuf::RpcController* cntl_butil,
+                    const VariableMessage* request, VoidMessage* response,
+                    google::protobuf::Closure* done) override {
+    PADDLE_ENFORCE(request_send_h_ != nullptr,
+                   "RequestSend handler should be registed first!");
+    brpc::ClosureGuard done_guard(done);
+
+    paddle::framework::Scope* local_scope = request_send_h_->scope();
+    paddle::framework::Variable* outvar = nullptr;
+    paddle::framework::Variable* invar = nullptr;
+
+    std::string varname = request->varname();
+
+    if (!request_send_h_->sync_mode()) {
+      local_scope = &request_send_h_->scope()->NewScope();
+      invar = local_scope->Var(varname);
+    } else {
+      invar = local_scope->FindVar(varname);
+    }
+
+    request_send_h_->Handle(varname, local_scope, invar, &outvar);
+
+    if (!request_send_h_->sync_mode()) {
+      request_send_h_->scope()->DeleteScope(local_scope);
+    }
+  }
+
+  void GetVariable(google::protobuf::RpcController* cntl_butil,
+                   const VariableMessage* request, VariableMessage* response,
+                   google::protobuf::Closure* done) override {
+    PADDLE_ENFORCE(request_get_h_ != nullptr,
+                   "RequestGet handler should be registed first!");
+  }
+
+  void PrefetchVariable(google::protobuf::RpcController* cntl_butil,
+                        const VariableMessage* request,
+                        VariableMessage* response,
+                        google::protobuf::Closure* done) override {
+    PADDLE_ENFORCE(request_prefetch_h_ != nullptr,
+                   "kRequestPrefetch handler should be registed first!");
+  }
+
+ private:
+  paddle::operators::detail::RequestHandler* request_send_h_;
+  paddle::operators::detail::RequestHandler* request_get_h_;
+  paddle::operators::detail::RequestHandler* request_prefetch_h_;
+};
+}  // namespace sendrecv
+
+namespace paddle {
+namespace operators {
+namespace detail {
+
+void AsyncBRPCServer::StartServer() {
+  // Instance of your service.
+  sendrecv::BRPCServiceImpl service_impl(rpc_call_map_);
+
+  // Add the service into server. Notice the second parameter, because the
+  // service is put on stack, we don't want server to delete it, otherwise
+  // use brpc::SERVER_OWNS_SERVICE.
+  if (server_.AddService(&service_impl, brpc::SERVER_DOESNT_OWN_SERVICE) != 0) {
+    LOG(FATAL) << "Fail to add service";
+    return;
+  }
+
+  brpc::ServerOptions options;
+  options.idle_timeout_sec = idle_timeout_s_;
+  options.max_concurrency = max_concurrency_;
+  if (server_.Start(bind_address_.c_str(), &options) != 0) {
+    LOG(FATAL) << "Fail to start EchoServer" << bind_address_;
+    return;
+  }
+
+  butil::EndPoint ep = server_.listen_address();
+  selected_port_ = ep.port;
+
+  {
+    std::lock_guard<std::mutex> lock(this->mutex_ready_);
+    ready_ = 1;
+  }
+  condition_ready_.notify_all();
+
+  server_.Join();
+}
+
+void AsyncBRPCServer::ShutDownImpl() { server_.Stop(1000); }
+
+void AsyncBRPCServer::WaitServerReady() {
+  VLOG(3) << "AsyncGRPCServer is wait server ready";
+  std::unique_lock<std::mutex> lock(this->mutex_ready_);
+  condition_ready_.wait(lock, [=] { return this->ready_ == 1; });
+  VLOG(3) << "AsyncGRPCServer WaitSeverReady";
+}
+
+};  // namespace detail
+};  // namespace operators
+};  // namespace paddle

paddle/fluid/operators/detail/brpc_server.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include <condition_variable>  // NOLINT
+#include <mutex>               // NOLINT
+#include <string>
+
+#include "brpc/server.h"
+#include "paddle/fluid/operators/detail/rpc_server.h"
+#include "paddle/fluid/operators/detail/send_recv.pb.h"
+
+namespace paddle {
+namespace operators {
+namespace detail {
+
+class AsyncBRPCServer final : public RPCServer {
+ public:
+  explicit AsyncBRPCServer(const std::string& address, int client_num)
+      : RPCServer(address, client_num), ready_(0) {}
+
+  virtual ~AsyncBRPCServer() {}
+  void StartServer() override;
+  void WaitServerReady() override;
+
+ private:
+  void ShutDownImpl() override;
+
+  brpc::Server server_;
+
+  static constexpr int idle_timeout_s_ = -1;
+  static constexpr int max_concurrency_ = 0;
+
+  std::mutex mutex_ready_;
+  std::condition_variable condition_ready_;
+  int ready_;
+};
+
+};  // namespace detail
+};  // namespace operators
+};  // namespace paddle

paddle/fluid/operators/detail/grpc_client.cc

@@ -19,6 +19,7 @@ limitations under the License. */
 #include <limits>
 
 #include "paddle/fluid/framework/threadpool.h"
+#include "paddle/fluid/operators/detail/request_handler.h"
 #include "paddle/fluid/platform/profiler.h"
 
 namespace paddle {

paddle/fluid/operators/detail/grpc_server.cc

@@ -41,11 +41,22 @@ class RequestBase {
   virtual ~RequestBase() {}
   virtual void Process() = 0;
 
-  CallStatus Status() { return status_; }
-  void SetStatus(CallStatus status) { status_ = status; }
+  CallStatus Status() const {
+    std::lock_guard<std::mutex> l(status_mu_);
+    return status_;
+  }
+
+  template <typename T>
+  void Finish(const T& reply, ServerAsyncResponseWriter<T>* responder) {
+    std::lock_guard<std::mutex> l(status_mu_);
+    status_ = FINISH;
+    responder->Finish(reply, ::grpc::Status::OK,
+                      reinterpret_cast<void*>(static_cast<intptr_t>(req_id_)));
+  }
   virtual std::string GetReqName() = 0;
 
  protected:
+  mutable std::mutex status_mu_;
   ::grpc::ServerContext ctx_;
   GrpcService::AsyncService* service_;
   ::grpc::ServerCompletionQueue* cq_;
@@ -80,9 +91,7 @@ class RequestSend final : public RequestBase {
     framework::Variable* outvar = nullptr;
 
     request_handler_->Handle(varname, scope, invar, &outvar);
-    status_ = FINISH;
-    responder_.Finish(reply_, ::grpc::Status::OK,
-                      reinterpret_cast<void*>(static_cast<intptr_t>(req_id_)));
+    Finish(reply_, &responder_);
   }
 
  protected:
@@ -122,9 +131,7 @@ class RequestGet final : public RequestBase {
       SerializeToByteBuffer(varname, outvar, *request_handler_->dev_ctx(),
                             &reply_);
     }
-    status_ = FINISH;
-    responder_.Finish(reply_, ::grpc::Status::OK,
-                      reinterpret_cast<void*>(static_cast<intptr_t>(req_id_)));
+    Finish(reply_, &responder_);
   }
 
  protected:
@@ -166,9 +173,7 @@ class RequestPrefetch final : public RequestBase {
 
     SerializeToByteBuffer(varname, outvar, *request_handler_->dev_ctx(),
                           &reply_);
-    responder_.Finish(reply_, ::grpc::Status::OK,
-                      reinterpret_cast<void*>(static_cast<intptr_t>(req_id_)));
-    status_ = FINISH;
+    Finish(reply_, &responder_);
   }
 
  protected:

paddle/fluid/operators/detail/grpc_server.h

@@ -53,6 +53,7 @@ class AsyncGRPCServer final : public RPCServer {
   void StartServer() override;
 
  private:
+  // HandleRequest needs to be thread-safe.
   void HandleRequest(
       ::grpc::ServerCompletionQueue* cq, const std::string& rpc_name,
       std::function<void(const std::string&, int)> TryToRegisterNewOne);

paddle/fluid/operators/detail/macros.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
+
+#ifdef PADDLE_WITH_GRPC
+#include "paddle/fluid/operators/detail/grpc_client.h"
+#include "paddle/fluid/operators/detail/grpc_server.h"
+#define RPCSERVER_T detail::AsyncGRPCServer
+#define RPCCLIENT_T detail::GRPCClient
+#else
+#include "paddle/fluid/operators/detail/brpc_client.h"
+#include "paddle/fluid/operators/detail/brpc_server.h"
+#define RPCSERVER_T detail::AsyncBRPCServer
+#define RPCCLIENT_T detail::BRPCClient
+#endif

paddle/fluid/operators/detail/request_handler.h

@@ -28,7 +28,6 @@
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/framework/selected_rows.h"
 #include "paddle/fluid/framework/var_type.h"
-#include "paddle/fluid/operators/detail/sendrecvop_utils.h"
 
 namespace paddle {
 namespace operators {
@@ -38,6 +37,10 @@ constexpr char kRequestSend[] = "RequestSend";
 constexpr char kRequestGet[] = "RequestGet";
 constexpr char kRequestPrefetch[] = "RequestPrefetch";
 
+#define LISTEN_TERMINATE_MESSAGE "TERMINATE@RECV"
+#define BATCH_BARRIER_MESSAGE "BATCH_BARRIER@RECV"
+#define FETCH_BARRIER_MESSAGE "FETCH_BARRIER@RECV"
+
 class RPCServer;
 
 class RequestHandler {

paddle/fluid/operators/detail/request_handler_impl.cc

@@ -16,15 +16,12 @@
 #include <string>
 #include <vector>
 
-#include "paddle/fluid/framework/blocking_queue.h"
 #include "paddle/fluid/framework/data_type.h"
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/framework/selected_rows.h"
 #include "paddle/fluid/operators/detail/request_handler_impl.h"
 #include "paddle/fluid/operators/detail/rpc_server.h"
-#include "paddle/fluid/operators/detail/sendrecvop_utils.h"
-#include "paddle/fluid/operators/detail/variable_response.h"
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/detail/request_handler_impl.h

@@ -29,7 +29,6 @@
 #include "paddle/fluid/framework/selected_rows.h"
 #include "paddle/fluid/framework/var_type.h"
 #include "paddle/fluid/operators/detail/request_handler.h"
-#include "paddle/fluid/operators/detail/sendrecvop_utils.h"
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/detail/rpc_client.h

@@ -26,6 +26,8 @@ namespace detail {
 
 class RPCClient {
  public:
+  RPCClient() {}
+  virtual ~RPCClient() {}
   virtual bool AsyncSendVar(const std::string& ep,
                             const platform::DeviceContext& ctx,
                             const framework::Scope& scope,

paddle/fluid/operators/detail/grpc_server_test.cc → paddle/fluid/operators/detail/rpc_server_test.cc

@@ -17,15 +17,14 @@ limitations under the License. */
 #include <thread>  // NOLINT
 
 #include "gtest/gtest.h"
-#include "paddle/fluid/operators/detail/grpc_client.h"
-#include "paddle/fluid/operators/detail/grpc_server.h"
-#include "paddle/fluid/operators/detail/rpc_client.h"
-
 #include "paddle/fluid/framework/block_desc.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/operator.h"
 
+#include "paddle/fluid/operators/detail/macros.h"
 #include "paddle/fluid/operators/detail/request_handler_impl.h"
+#include "paddle/fluid/operators/detail/rpc_client.h"
+#include "paddle/fluid/operators/detail/rpc_server.h"
 
 namespace framework = paddle::framework;
 namespace platform = paddle::platform;
@@ -33,7 +32,7 @@ namespace detail = paddle::operators::detail;
 
 USE_OP(lookup_table);
 
-std::unique_ptr<detail::AsyncGRPCServer> g_rpc_service;
+std::unique_ptr<detail::RPCServer> g_rpc_service;
 std::unique_ptr<detail::RequestHandler> g_req_handler;
 
 framework::BlockDesc* AppendPrefetchBlcok(framework::ProgramDesc* program) {
@@ -112,20 +111,19 @@ void StartServer() {
   g_req_handler->SetRPCServer(g_rpc_service.get());
 
   std::thread server_thread(
-      std::bind(&detail::AsyncGRPCServer::StartServer, g_rpc_service.get()));
+      std::bind(&detail::RPCServer::StartServer, g_rpc_service.get()));
 
   server_thread.join();
 }
 
 TEST(PREFETCH, CPU) {
   g_req_handler.reset(new detail::RequestPrefetchHandler(true));
-  g_rpc_service.reset(new detail::AsyncGRPCServer("127.0.0.1:0", 1));
+  g_rpc_service.reset(new RPCSERVER_T("127.0.0.1:0", 1));
+  detail::RPCClient* client = detail::RPCClient::GetInstance<RPCCLIENT_T>();
 
   std::thread server_thread(StartServer);
   g_rpc_service->WaitServerReady();
 
-  detail::RPCClient* client =
-      detail::RPCClient::GetInstance<detail::GRPCClient>();
   int port = g_rpc_service->GetSelectedPort();
   std::string ep = paddle::string::Sprintf("127.0.0.1:%d", port);
 

paddle/fluid/operators/detail/send_recv.proto

@@ -14,6 +14,8 @@ limitations under the License. */
 syntax = "proto3";
 package sendrecv;
 
+// option cc_generic_services = true;
+
 service SendRecvService {
   // For parameter server round-robin like hashing, do not split tensors.
   // Send and recv only one tensor

paddle/fluid/operators/detail/sendrecvop_utils.h

@@ -32,16 +32,6 @@ namespace paddle {
 namespace operators {
 namespace detail {
 
-#define LISTEN_TERMINATE_MESSAGE "TERMINATE@RECV"
-#define BATCH_BARRIER_MESSAGE "BATCH_BARRIER@RECV"
-#define FETCH_BARRIER_MESSAGE "FETCH_BARRIER@RECV"
-
-static int64_t GetTimestamp() {
-  struct timeval tp;
-  gettimeofday(&tp, NULL);
-  return tp.tv_sec * 1000 + tp.tv_usec / 1000;
-}
-
 typedef void (*DestroyCallback)(void*);
 
 void SerializeToByteBuffer(const std::string& name, framework::Variable* var,

paddle/fluid/operators/elementwise_op.h

@@ -59,7 +59,7 @@ class ElementwiseOpMaker : public framework::OpProtoAndCheckerMaker {
   void Make() final {
     AddInput("X", "(Tensor), The first input tensor of elementwise op.");
     AddInput("Y", "(Tensor), The second input tensor of elementwise op.");
-    AddOutput("Out", "The output of elementwise op.");
+    AddOutput("Out", "The output of elementwise op.").Reuse("X");
     AddAttr<int>("axis",
                  "(int, default -1). The start dimension index "
                  "for broadcasting Y onto X.")

paddle/fluid/operators/fetch_barrier_op.cc

@@ -19,9 +19,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/framework.pb.h"
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/op_registry.h"
-
-#include "paddle/fluid/operators/detail/grpc_client.h"
-#include "paddle/fluid/operators/detail/rpc_client.h"
+#include "paddle/fluid/operators/detail/macros.h"
 #include "paddle/fluid/platform/profiler.h"
 
 namespace paddle {
@@ -45,7 +43,7 @@ class FetchBarrierOp : public framework::OperatorBase {
     platform::RecordEvent record_event(Type(), &ctx);
 
     detail::RPCClient* rpc_client =
-        detail::RPCClient::GetInstance<detail::GRPCClient>();
+        detail::RPCClient::GetInstance<RPCCLIENT_T>();
 
     rpc_client->Wait();
 

paddle/fluid/operators/fill_constant_batch_size_like_op.cc

@@ -32,16 +32,16 @@ class FillConstantBatchSizeLikeOp : public BatchSizeLikeOp {
 class FillConstantBatchSizeLikeOpMaker : public BatchSizeLikeOpMaker {
  protected:
   void Apply() override {
-    AddAttr<int>("dtype",
-                 "(int, default 5 (FP32)) "
-                 "Output data type")
+    AddAttr<int>(
+        "dtype",
+        "It could be numpy.dtype. Output data type. Default is float32")
         .SetDefault(framework::proto::VarType::FP32);
-    AddAttr<float>("value", "(float, default 0) The value to be filled")
+    AddAttr<float>("value", "default 0. The value to be filled")
         .SetDefault(0.0f);
     AddComment(R"DOC(
-FillConstantBatchSizeLike Operator.
-
-Fill up a variable with specified constant value.
+This function creates a tensor of specified *shape*, *dtype* and batch size,
+and initializes this with a constant supplied in *value*. The batch size is
+obtained from the `input` tensor.
 
 )DOC");
   }

paddle/fluid/operators/gen_nccl_id_op.cc

@@ -21,8 +21,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/threadpool.h"
-#include "paddle/fluid/operators/detail/grpc_client.h"
-#include "paddle/fluid/operators/detail/grpc_server.h"
+#include "paddle/fluid/operators/detail/macros.h"
 #include "paddle/fluid/operators/detail/request_handler_impl.h"
 #include "paddle/fluid/platform/nccl_helper.h"
 
@@ -61,8 +60,8 @@ class GenNCCLIdOp : public framework::OperatorBase {
 
     std::vector<std::string> endpoint_list =
         Attr<std::vector<std::string>>("endpoint_list");
-    detail::RPCClient* client =
-        detail::RPCClient::GetInstance<detail::GRPCClient>();
+    detail::RPCClient* client = detail::RPCClient::GetInstance<RPCCLIENT_T>();
+
     for (auto& ep : endpoint_list) {
       VLOG(3) << "sending nccl id to " << ep;
       client->AsyncSendVar(ep, dev_ctx, *scope, NCCL_ID_VARNAME);
@@ -78,9 +77,11 @@ class GenNCCLIdOp : public framework::OperatorBase {
     // deleter will call GRPC Server's base class's dtor and
     // that will cause a wired crash.
     detail::RequestSendHandler rpc_h(true);
-    detail::AsyncGRPCServer rpc_service(endpoint, 1);
-    rpc_service.RegisterRPC(detail::kRequestSend, &rpc_h);
-    rpc_h.SetRPCServer(&rpc_service);
+    std::unique_ptr<detail::RPCServer> rpc_service(
+        new RPCSERVER_T(endpoint, 1));
+
+    rpc_service->RegisterRPC(detail::kRequestSend, &rpc_h);
+    rpc_h.SetRPCServer(rpc_service.get());
 
     framework::ProgramDesc empty_program;
     framework::Executor executor(dev_ctx.GetPlace());
@@ -90,12 +91,13 @@ class GenNCCLIdOp : public framework::OperatorBase {
     rpc_h.SetExecutor(&executor);
 
     std::thread server_thread(
-        std::bind(&detail::AsyncGRPCServer::StartServer, &rpc_service));
-    rpc_service.SetCond(detail::kRequestSend);
+        std::bind(&detail::RPCServer::StartServer, rpc_service.get()));
+
+    rpc_service->SetCond(detail::kRequestSend);
     VLOG(3) << "start getting nccl id from trainer 0...";
-    rpc_service.WaitBarrier(detail::kRequestSend);
+    rpc_service->WaitBarrier(detail::kRequestSend);
     VLOG(3) << "got nccl id and stop server...";
-    rpc_service.ShutDown();
+    rpc_service->ShutDown();
     VLOG(3) << "rpc server stopped";
     server_thread.join();
   }

paddle/fluid/operators/linear_chain_crf_op.cc

@@ -67,8 +67,6 @@ class LinearChainCRFOpMaker : public framework::OpProtoAndCheckerMaker {
         "mini-batch. Note: S is equal to the sequence number in a mini-batch. "
         "The output is no longer a LoDTensor.");
     AddComment(R"DOC(
-LinearChainCRF Operator.
-
 Conditional Random Field defines an undirected probabilistic graph with nodes
 denoting random variables and edges denoting dependencies between these
 variables. CRF learns the conditional probability $P(Y|X)$, where

paddle/fluid/operators/listen_and_serv_op.cc

@@ -19,7 +19,8 @@ limitations under the License. */
 #include <thread>  // NOLINT
 #include <vector>
 
-#include "paddle/fluid/operators/detail/grpc_server.h"
+#include "paddle/fluid/operators/detail/macros.h"
+
 #include "paddle/fluid/operators/detail/request_handler_impl.h"
 #include "paddle/fluid/operators/listen_and_serv_op.h"
 #include "paddle/fluid/platform/profiler.h"
@@ -89,6 +90,12 @@ void ListenAndServOp::SavePort() const {
   rpc_service_->SavePort();
 }
 
+static int64_t GetTimestamp() {
+  struct timeval tp;
+  gettimeofday(&tp, NULL);
+  return tp.tv_sec * 1000 + tp.tv_usec / 1000;
+}
+
 void ListenAndServOp::RunSyncLoop(framework::Executor *executor,
                                   framework::ProgramDesc *program,
                                   framework::Scope *recv_scope,
@@ -127,7 +134,7 @@ void ListenAndServOp::RunSyncLoop(framework::Executor *executor,
     int32_t last_parent_blkid = program->Block(1).Parent();
     std::vector<size_t> parallel_blkids;
     parallel_blkids.push_back(1);
-    double ts = detail::GetTimestamp();
+    double ts = GetTimestamp();
     for (size_t blkid = 2; blkid < num_blocks; ++blkid) {
       if (blkid != static_cast<size_t>(prefetch_block->ID())) {
         if (program->Block(blkid).Parent() != last_parent_blkid) {
@@ -141,7 +148,7 @@ void ListenAndServOp::RunSyncLoop(framework::Executor *executor,
     }
     ParallelExecuteBlocks(parallel_blkids, executor, optimize_prepared, program,
                           recv_scope);
-    VLOG(2) << "run all blocks spent " << detail::GetTimestamp() - ts << "(ms)";
+    VLOG(2) << "run all blocks spent " << GetTimestamp() - ts << "(ms)";
 
     rpc_service_->SetCond(detail::kRequestGet);
     rpc_service_->WaitBarrier(detail::kRequestGet);
@@ -235,8 +242,8 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope,
   LOG(INFO) << "sync_mode:" << sync_mode << ", fan_in:" << fan_in
             << ", end_point:" << endpoint;
 
-  // request_handler_.reset(new detail::GRPCRequestSendHandler(sync_mode));
-  rpc_service_.reset(new detail::AsyncGRPCServer(endpoint, fan_in));
+  rpc_service_.reset(new RPCSERVER_T(endpoint, fan_in));
+
   request_send_handler_.reset(new detail::RequestSendHandler(sync_mode));
   request_get_handler_.reset(new detail::RequestGetHandler(sync_mode));
   request_prefetch_handler_.reset(

paddle/fluid/operators/load_op.cc

@@ -74,25 +74,18 @@ class LoadOp : public framework::OperatorBase {
 class LoadOpProtoMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() override {
-    AddOutput("Out", "(Tensor) The tensor need to be loaded");
+    AddOutput("Out", "The tensor need to be loaded");
     AddAttr<bool>(
         "load_as_fp16",
-        "(boolean, default false)"
         "If true, the tensor will be first loaded and then "
         "converted to float16 data type. Otherwise, the tensor will be "
-        "directly loaded without data type conversion.")
+        "directly loaded without data type conversion. Default is false.")
         .SetDefault(false);
     AddAttr<std::string>("file_path",
-                         "(string) "
-                         "Variable will be loaded from \"file_path\".")
+                         R"(Variable will be loaded from "file_path")")
         .AddCustomChecker(
             [](const std::string &path) { return !path.empty(); });
-    AddComment(R"DOC(
-Load Operator.
-
-Load operator will load a tensor variable from disk file.
-
-)DOC");
+    AddComment("Load operator will load a tensor variable from disk file.");
   }
 };
 }  // namespace operators

paddle/fluid/operators/max_sequence_len_op.cc

@@ -42,10 +42,15 @@ class MaxSeqenceLenOp : public framework::OperatorBase {
 class MaxSeqenceLenOpProtoMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() override {
-    AddInput("RankTable", "The lod_rank_table.");
-    AddOutput("Out", "The max sequence length.");
-    AddComment(
-        R"DOC(Calculate the max sequence length through lod_rank_table.)DOC");
+    AddInput("RankTable", "Input variable which is a LoDRankTable object");
+    AddOutput("Out", "The max sequence length");
+    AddComment(R"DOC(
+    Given a LoDRankTable object, this layer returns the max length of
+    a batch of sequences. In fact, a LoDRankTable object contains a list of
+    tuples(<sequence index, sequence length>) and the list is already sorted by
+    sequence length in descending order, so the operator just returns the
+    sequence length of the first tuple element
+)DOC");
   }
 };
 

paddle/fluid/operators/mean_op.cc

@@ -34,7 +34,7 @@ class MeanOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() override {
     AddInput("X", "The input of mean op");
-    AddOutput("Out", "The output of mean op");
+    AddOutput("Out", "The output of mean op").Reuse("X");
     AddComment(R"DOC(
 Mean Operator.
 

paddle/fluid/operators/pool_mkldnn_op.cc

@@ -18,9 +18,14 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
-using mkldnn::memory;  // Note: paddle has also "memory" namespace
-using mkldnn::pooling_forward;
+using framework::DataLayout;
+using mkldnn::memory;
 using mkldnn::pooling_backward;
+using mkldnn::pooling_forward;
+using mkldnn::primitive;
+using mkldnn::reorder;
+using mkldnn::stream;
+using platform::to_void_cast;
 
 // Generate keys for storing/retriving primitives for this operator
 // TODO(jczaja): Make hashing function more optimial
@@ -55,8 +60,9 @@ class PoolMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     const Tensor* input = ctx.Input<Tensor>("X");
     Tensor* output = ctx.Output<Tensor>("Out");
 
-    // Get an unique name from "argument" name of "Out" variable
-    // This name will be used as key when saving info into device context
+    PADDLE_ENFORCE(input->layout() == DataLayout::kMKLDNN &&
+                       input->format() != memory::format::format_undef,
+                   "Wrong layout/format set for Input tensor");
 
     std::string pooling_type = ctx.Attr<std::string>("pooling_type");
     std::vector<int> ksize = ctx.Attr<std::vector<int>>("ksize");
@@ -82,6 +88,9 @@ class PoolMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     std::vector<int> src_tz = paddle::framework::vectorize2int(input->dims());
     std::vector<int> dst_tz = paddle::framework::vectorize2int(output->dims());
 
+    auto input_format = input->format();
+    memory::format output_format{memory::format::format_undef};
+
     const std::string key = gethash(src_tz, pooling_type, ksize, strides,
                                     paddings, ctx.op().Output("Out"));
     const std::string key_pool_p = key + "@pool_p";
@@ -94,16 +103,17 @@ class PoolMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     auto pool_p =
         std::static_pointer_cast<pooling_forward>(dev_ctx.GetBlob(key_pool_p));
     if (pool_p == nullptr) {
-      // TODO(pzelazko-intel): support more formats
+      auto src_md = platform::MKLDNNMemDesc(
+          src_tz, platform::MKLDNNGetDataType<T>(), input_format);
 
-      auto src_md =
-          platform::MKLDNNMemDesc(src_tz, platform::MKLDNNGetDataType<T>(),
-                                  mkldnn::memory::format::nchw);
-      auto dst_md =
-          platform::MKLDNNMemDesc(dst_tz, platform::MKLDNNGetDataType<T>(),
-                                  mkldnn::memory::format::nchw);
+      /* create memory descriptor for pooling without specified format
+       * ('any') which lets a primitive (pooling in this case) choose
+       * the memory format preferred for best performance
+       */
+      auto dst_md = platform::MKLDNNMemDesc(dst_tz, mkldnn::memory::f32,
+                                            mkldnn::memory::format::any);
 
-      std::shared_ptr<pooling_forward::primitive_desc> pool_pd =
+      std::shared_ptr<mkldnn::pooling_forward::primitive_desc> pool_pd =
           CreatePrimitiveDesc(src_md, dst_md, strides, paddings, ksize,
                               pooling_type, mkldnn_engine);
 
@@ -116,20 +126,22 @@ class PoolMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
       // save pool_workspace_memory to be referred in backward path
       dev_ctx.SetBlob(key_pool_workspace_memory, workspace_memory);
 
-      auto pool_src_memory_p = std::make_shared<memory>(
-          memory::primitive_desc{src_md, mkldnn_engine},
-          static_cast<void*>(const_cast<T*>(input_data)));
-      dev_ctx.SetBlob(key_pool_src_mem_p, pool_src_memory_p);
+      auto src_memory = std::make_shared<memory>(pool_pd->src_primitive_desc(),
+                                                 to_void_cast<T>(input_data));
+      auto dst_memory =
+          std::make_shared<memory>(pool_pd->dst_primitive_desc(), output_data);
 
-      auto pool_dst_memory_p = std::make_shared<memory>(
-          memory::primitive_desc{dst_md, mkldnn_engine},
-          static_cast<void*>(output_data));
-      dev_ctx.SetBlob(key_pool_dst_mem_p, pool_dst_memory_p);
+      dev_ctx.SetBlob(key_pool_src_mem_p, src_memory);
+      dev_ctx.SetBlob(key_pool_dst_mem_p, dst_memory);
+
+      pool_p = std::make_shared<pooling_forward>(*pool_pd, *(src_memory.get()),
+                                                 *(dst_memory.get()),
+                                                 *workspace_memory);
 
-      pool_p = std::make_shared<pooling_forward>(
-          *pool_pd, *(pool_src_memory_p.get()), *(pool_dst_memory_p.get()),
-          *workspace_memory);
       dev_ctx.SetBlob(key_pool_p, pool_p);
+
+      output_format =
+          (memory::format)dst_memory->get_primitive_desc().desc().data.format;
     } else {
       // Primitives already exist
       auto pool_src_memory_p =
@@ -140,14 +152,20 @@ class PoolMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
           std::static_pointer_cast<memory>(dev_ctx.GetBlob(key_pool_dst_mem_p));
       PADDLE_ENFORCE(pool_dst_memory_p != nullptr,
                      "Fail to find pooling dst mem_p in device context");
-      pool_src_memory_p->set_data_handle(
-          reinterpret_cast<void*>(const_cast<T*>(input_data)));
+      pool_src_memory_p->set_data_handle(to_void_cast<T>(input_data));
       pool_dst_memory_p->set_data_handle(output_data);
+
+      output_format = (memory::format)pool_dst_memory_p->get_primitive_desc()
+                          .desc()
+                          .data.format;
     }
 
     // push primitive to stream and wait until it's executed
     std::vector<mkldnn::primitive> pipeline{*(pool_p.get())};
-    mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
+    stream(stream::kind::eager).submit(pipeline).wait();
+
+    output->set_layout(DataLayout::kMKLDNN);
+    output->set_format(output_format);
   }
 
  private:
@@ -194,6 +212,13 @@ class PoolMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
     const Tensor* out_grad = ctx.Input<Tensor>(framework::GradVarName("Out"));
     Tensor* in_x_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
 
+    PADDLE_ENFORCE(in_x->layout() == DataLayout::kMKLDNN &&
+                       in_x->format() != memory::format::format_undef,
+                   "Wrong layout/format set for Input X tensor");
+    PADDLE_ENFORCE(out_grad->layout() == DataLayout::kMKLDNN &&
+                       out_grad->format() != memory::format::format_undef,
+                   "Wrong layout/format set for Input output_grad tensor");
+
     std::string pooling_type = ctx.Attr<std::string>("pooling_type");
     std::vector<int> ksize = ctx.Attr<std::vector<int>>("ksize");
     std::vector<int> strides = ctx.Attr<std::vector<int>>("strides");
@@ -212,6 +237,7 @@ class PoolMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
 
     const T* out_grad_data = out_grad->data<T>();
     T* in_x_grad_data = in_x_grad->mutable_data<T>(ctx.GetPlace());
+    memory::format in_x_grad_format{memory::format::format_undef};
 
     std::vector<int> diff_src_tz =
         paddle::framework::vectorize2int(in_x_grad->dims());
@@ -225,39 +251,48 @@ class PoolMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
     const std::string key_pool_bwd_p = key + "@pool_bwd_p";
     const std::string key_pool_diff_src_mem_p = key + "@pool_diff_src_mem_p";
     const std::string key_pool_diff_dst_mem_p = key + "@pool_diff_dst_mem_p";
+    const std::string key_pool_src_mem_p = key + "@pool_src_mem_p";
+    const std::string key_pool_dst_mem_p = key + "@pool_dst_mem_p";
     const std::string key_pool_pd = key + "@pool_pd";
     const std::string key_pool_workspace_memory =
         key + "@pool_workspace_memory";
 
+    auto user_diff_dst_memory =
+        memory({{{diff_dst_tz}, memory::data_type::f32, out_grad->format()},
+                mkldnn_engine},
+               to_void_cast<T>(out_grad_data));
+
+    std::shared_ptr<memory> diff_src_memory;
+    std::shared_ptr<memory> diff_dst_memory;
+    auto dst_memory =
+        std::static_pointer_cast<memory>(dev_ctx.GetBlob(key_pool_dst_mem_p));
+    PADDLE_ENFORCE(dst_memory != nullptr,
+                   "Fail to find dst_memory in device context");
+
+    primitive reorder_diff_dst;
+    bool is_diff_dst_reordered = false;
     auto pool_bwd_p = std::static_pointer_cast<pooling_backward>(
         dev_ctx.GetBlob(key_pool_bwd_p));
     if (pool_bwd_p == nullptr) {
-      auto diff_src_md =
-          platform::MKLDNNMemDesc(diff_src_tz, platform::MKLDNNGetDataType<T>(),
-                                  mkldnn::memory::format::nchw);
-      auto diff_dst_md =
-          platform::MKLDNNMemDesc(diff_dst_tz, platform::MKLDNNGetDataType<T>(),
-                                  mkldnn::memory::format::nchw);
+      // Retrieve src_memory/dst_memory saved in forward pass
+      auto src_memory =
+          std::static_pointer_cast<memory>(dev_ctx.GetBlob(key_pool_src_mem_p));
+      PADDLE_ENFORCE(src_memory != nullptr,
+                     "Fail to find src_memory in device context");
       // Retrieve pool_pd/pool_workspace_memory from device context
       auto pool_pd =
           std::static_pointer_cast<mkldnn::pooling_forward::primitive_desc>(
               dev_ctx.GetBlob(key_pool_pd));
       PADDLE_ENFORCE(pool_pd != nullptr,
                      "Fail to find pool_pd in device context");
-
-      auto workspace_memory = std::static_pointer_cast<mkldnn::memory>(
+      auto workspace_memory = std::static_pointer_cast<memory>(
           dev_ctx.GetBlob(key_pool_workspace_memory));
       PADDLE_ENFORCE(workspace_memory != nullptr,
                      "Fail to find workspace_memory in device context");
 
-      auto pool_diff_src_memory_p = std::make_shared<memory>(memory(
-          {diff_src_md, mkldnn_engine}, static_cast<void*>(in_x_grad_data)));
-      dev_ctx.SetBlob(key_pool_diff_src_mem_p, pool_diff_src_memory_p);
-
-      auto pool_diff_dst_memory_p = std::make_shared<memory>(
-          memory({diff_dst_md, mkldnn_engine},
-                 static_cast<void*>(const_cast<T*>(out_grad_data))));
-      dev_ctx.SetBlob(key_pool_diff_dst_mem_p, pool_diff_dst_memory_p);
+      // create memory descriptors for pooling
+      auto diff_src_md = src_memory.get()->get_primitive_desc().desc();
+      auto diff_dst_md = dst_memory.get()->get_primitive_desc().desc();
 
       auto pool_bwd_desc = mkldnn::pooling_backward::desc(
           pooling_type == "max" ? mkldnn::algorithm::pooling_max
@@ -267,35 +302,74 @@ class PoolMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
       auto pool_bwd_pd = mkldnn::pooling_backward::primitive_desc(
           pool_bwd_desc, mkldnn_engine, *pool_pd);
 
+      // reorder between user_diff_dst and pool diff_dst if needed
+      diff_dst_memory = std::make_shared<memory>(user_diff_dst_memory);
+      if (memory::primitive_desc(dst_memory->get_primitive_desc()) !=
+          user_diff_dst_memory.get_primitive_desc()) {
+        diff_dst_memory =
+            std::make_shared<memory>(dst_memory.get()->get_primitive_desc());
+        reorder_diff_dst = reorder(user_diff_dst_memory, *diff_dst_memory);
+        is_diff_dst_reordered = true;
+      }
+
+      diff_src_memory = std::make_shared<memory>(
+          pool_bwd_pd.diff_src_primitive_desc(), in_x_grad_data);
+
+      dev_ctx.SetBlob(key_pool_diff_src_mem_p, diff_src_memory);
+      dev_ctx.SetBlob(key_pool_diff_dst_mem_p, diff_dst_memory);
+
       pool_bwd_p = std::make_shared<pooling_backward>(
-          pool_bwd_pd, *(pool_diff_dst_memory_p.get()), *workspace_memory,
-          *(pool_diff_src_memory_p));
+          pool_bwd_pd, *(diff_dst_memory.get()), *workspace_memory,
+          *(diff_src_memory));
       dev_ctx.SetBlob(key_pool_bwd_p, pool_bwd_p);
+
     } else {
       // Primitives already exist
-      auto pool_diff_src_memory_p = std::static_pointer_cast<memory>(
+      diff_src_memory = std::static_pointer_cast<memory>(
           dev_ctx.GetBlob(key_pool_diff_src_mem_p));
-      PADDLE_ENFORCE(pool_diff_src_memory_p != nullptr,
+      PADDLE_ENFORCE(diff_src_memory != nullptr,
                      "Fail to find pooling src mem_p in device context");
-      auto pool_diff_dst_memory_p = std::static_pointer_cast<memory>(
+      diff_dst_memory = std::static_pointer_cast<memory>(
           dev_ctx.GetBlob(key_pool_diff_dst_mem_p));
-      PADDLE_ENFORCE(pool_diff_dst_memory_p != nullptr,
+      PADDLE_ENFORCE(diff_dst_memory != nullptr,
                      "Fail to find pooling dst mem_p in device context");
-      pool_diff_src_memory_p->set_data_handle(
-          reinterpret_cast<void*>(in_x_grad_data));
-      pool_diff_dst_memory_p->set_data_handle(const_cast<T*>(out_grad_data));
+
+      diff_src_memory->set_data_handle(reinterpret_cast<void*>(in_x_grad_data));
+      diff_dst_memory->set_data_handle(const_cast<T*>(out_grad_data));
+
+      // reorder between user_diff_dst and pool diff_dst if needed
+      if (memory::primitive_desc(dst_memory->get_primitive_desc()) !=
+          user_diff_dst_memory.get_primitive_desc()) {
+        diff_dst_memory =
+            std::make_shared<memory>(dst_memory.get()->get_primitive_desc());
+        reorder_diff_dst = reorder(user_diff_dst_memory, *diff_dst_memory);
+        is_diff_dst_reordered = true;
+      }
     }
 
+    in_x_grad_format = (memory::format)diff_src_memory->get_primitive_desc()
+                           .desc()
+                           .data.format;
+
     // push primitive to stream and wait until it's executed
-    std::vector<mkldnn::primitive> pipeline{*(pool_bwd_p.get())};
+    std::vector<mkldnn::primitive> pipeline;
+    if (is_diff_dst_reordered) {
+      pipeline.push_back(reorder_diff_dst);
+    }
+    pipeline.push_back(*(pool_bwd_p.get()));
     mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
+
+    in_x_grad->set_layout(DataLayout::kMKLDNN);
+    in_x_grad->set_format(in_x_grad_format);
   }  // Compute()
 };
 
 }  // namespace operators
 }  // namespace paddle
 
+namespace ops = paddle::operators;
+
 REGISTER_OP_KERNEL(pool2d, MKLDNN, ::paddle::platform::CPUPlace,
-                   paddle::operators::PoolMKLDNNOpKernel<float>);
+                   ops::PoolMKLDNNOpKernel<float>);
 REGISTER_OP_KERNEL(pool2d_grad, MKLDNN, ::paddle::platform::CPUPlace,
-                   paddle::operators::PoolMKLDNNGradOpKernel<float>);
+                   ops::PoolMKLDNNGradOpKernel<float>);

paddle/fluid/operators/pool_op.cc

@@ -151,7 +151,8 @@ void Pool2dOpMaker::Make() {
             "The format of output tensor is also NCHW, "
             "where N is batch size, C is the number of channels, "
             "H is the height of the feature, "
-            "and W is the width of the feature.");
+            "and W is the width of the feature.")
+      .Reuse("X");
 
   AddAttr<std::string>("pooling_type",
                        "(string), pooling type, can be \"max\" for max-pooling "
@@ -244,7 +245,8 @@ void Pool3dOpMaker::Make() {
             "The format of output tensor is also NCDHW, "
             "where N is batch size, C is "
             "the number of channels, and D, H and W is the depth, height and "
-            "width of the feature, respectively.");
+            "width of the feature, respectively.")
+      .Reuse("X");
 
   AddAttr<std::string>("pooling_type",
                        "(string) Pooling type, can be \"max\" for max-pooling "

paddle/fluid/operators/prefetch_op.cc

@@ -18,7 +18,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/data_type.h"
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/op_registry.h"
-#include "paddle/fluid/operators/detail/grpc_client.h"
+#include "paddle/fluid/operators/detail/macros.h"
 #include "paddle/fluid/operators/send_recv_util.h"
 
 namespace paddle {
@@ -42,7 +42,7 @@ class PrefetchOp : public framework::OperatorBase {
     auto& ctx = *pool.Get(place);
 
     detail::RPCClient* rpc_client =
-        detail::RPCClient::GetInstance<detail::GRPCClient>();
+        detail::RPCClient::GetInstance<RPCCLIENT_T>();
 
     for (size_t i = 0; i < ins.size(); i++) {
       if (NeedSend(scope, ins[i])) {

paddle/fluid/operators/recv_op.cc

@@ -19,8 +19,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/framework.pb.h"
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/op_registry.h"
-
-#include "paddle/fluid/operators/detail/grpc_client.h"
+#include "paddle/fluid/operators/detail/macros.h"
 #include "paddle/fluid/platform/profiler.h"
 
 namespace paddle {
@@ -45,7 +44,7 @@ class RecvOp : public framework::OperatorBase {
     platform::RecordEvent record_event(Type(), &ctx);
 
     detail::RPCClient* rpc_client =
-        detail::RPCClient::GetInstance<detail::GRPCClient>();
+        detail::RPCClient::GetInstance<RPCCLIENT_T>();
 
     for (size_t i = 0; i < outs.size(); i++) {
       VLOG(3) << "getting " << outs[i] << " from " << epmap[i];
@@ -78,9 +77,15 @@ This operator can get variables from server side.
   }
 };
 
+class RecvOpShapeInference : public framework::InferShapeBase {
+ public:
+  void operator()(framework::InferShapeContext* ctx) const override {}
+};
+
 }  // namespace operators
 }  // namespace paddle
 
 namespace ops = paddle::operators;
 
-REGISTER_OPERATOR(recv, ops::RecvOp, ops::RecvOpMaker);
+REGISTER_OPERATOR(recv, ops::RecvOp, paddle::framework::EmptyGradOpMaker,
+                  ops::RecvOpMaker, ops::RecvOpShapeInference);

paddle/fluid/operators/send_barrier_op.cc

@@ -19,8 +19,8 @@ limitations under the License. */
 #include "paddle/fluid/framework/framework.pb.h"
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/detail/macros.h"
 
-#include "paddle/fluid/operators/detail/grpc_client.h"
 #include "paddle/fluid/platform/profiler.h"
 
 namespace paddle {
@@ -45,7 +45,7 @@ class SendBarrierOp : public framework::OperatorBase {
     platform::RecordEvent record_event(Type(), &ctx);
 
     detail::RPCClient* rpc_client =
-        detail::RPCClient::GetInstance<detail::GRPCClient>();
+        detail::RPCClient::GetInstance<RPCCLIENT_T>();
 
     VLOG(3) << "SendBarrierOp sync_mode:" << sync_mode;
 

paddle/fluid/operators/send_op.cc

@@ -16,10 +16,9 @@ limitations under the License. */
 #include <ostream>
 
 #include "paddle/fluid/framework/data_type.h"
-#include "paddle/fluid/framework/framework.pb.h"
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/op_registry.h"
-#include "paddle/fluid/operators/detail/grpc_client.h"
+#include "paddle/fluid/operators/detail/macros.h"
 #include "paddle/fluid/operators/send_recv_util.h"
 #include "paddle/fluid/platform/profiler.h"
 
@@ -36,12 +35,9 @@ class SendOp : public framework::OperatorBase {
   void RunImpl(const framework::Scope& scope,
                const platform::Place& place) const override {
     auto ins = Inputs("X");
-    auto outs = Outputs("Out");
-    std::vector<std::string> epmap = Attr<std::vector<std::string>>("epmap");
-    std::vector<std::string> endpoints =
-        Attr<std::vector<std::string>>("endpoints");
 
-    bool sync_mode = Attr<bool>("sync_mode");
+    std::vector<std::string> epmap = Attr<std::vector<std::string>>("epmap");
+    int sync_send = Attr<int>("sync_mode");
 
     platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
     auto& ctx = *pool.Get(place);
@@ -50,37 +46,19 @@ class SendOp : public framework::OperatorBase {
     platform::RecordEvent record_event(Type(), &ctx);
 
     detail::RPCClient* rpc_client =
-        detail::RPCClient::GetInstance<detail::GRPCClient>();
+        detail::RPCClient::GetInstance<RPCCLIENT_T>();
 
     for (size_t i = 0; i < ins.size(); i++) {
       if (NeedSend(scope, ins[i])) {
         VLOG(3) << "sending " << ins[i] << " to " << epmap[i];
+        // TODO(Yancey1989): we need to use an IO threadpool which has
+        // a larger number of threads than the computing threadpool.
         rpc_client->AsyncSendVar(epmap[i], ctx, scope, ins[i]);
       } else {
         VLOG(3) << "don't send no-initialied variable: " << ins[i];
       }
     }
-    rpc_client->Wait();
-
-    if (sync_mode) {
-      for (auto& ep : endpoints) {
-        VLOG(3) << "batch barrier, ep: " << ep;
-        rpc_client->AsyncSendBatchBarrier(ep);
-      }
-      rpc_client->Wait();
-    }
-
-    if (outs.size() > 0) {
-      for (size_t i = 0; i < outs.size(); i++) {
-        VLOG(2) << "getting " << outs[i] << " from " << epmap[i];
-        rpc_client->AsyncGetVar(epmap[i], ctx, scope, outs[i]);
-      }
-      rpc_client->Wait();
-      // tell pservers that current trainer have called fetch
-      for (auto& ep : endpoints) {
-        VLOG(2) << "send fetch barrier, ep: " << ep;
-        rpc_client->AsyncSendFetchBarrier(ep);
-      }
+    if (sync_send) {
       rpc_client->Wait();
     }
   }
@@ -89,26 +67,22 @@ class SendOp : public framework::OperatorBase {
 class SendOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() {
-    AddInput("X", "(Tensor) Input tensor to be sent").AsDuplicable();
-    AddOutput("Out", "(Tensor) Output tensor to be received from server")
+    AddInput("X", "(Tensor, SelectedRows) Input variables to be sent")
         .AsDuplicable();
     AddComment(R"DOC(
 Send operator
 
-This operator will send tensor to recv_op at the parameter server.
+This operator will send variables to listen_and_serve op at the parameter server.
 )DOC");
-    // TODO(typhoonzero): remove this attr generate de-duplicated vector from
-    // epmap when initializing.
-    AddAttr<std::vector<std::string>>("endpoints",
-                                      "(string vector, default 127.0.0.1:6164)"
-                                      "Server endpoints to send variables to.")
-        .SetDefault({});
+    AddAttr<int>("sync_mode",
+                 "(int, default 0)"
+                 "sync send or async send.")
+        .SetDefault(0);
     AddAttr<std::vector<std::string>>("epmap",
                                       "(string vector, default 127.0.0.1:6164)"
                                       "Server endpoints in the order of input "
                                       "variables for mapping")
-        .SetDefault({});
-    AddAttr<bool>("sync_mode", "work in sync_mode or not").SetDefault(true);
+        .SetDefault({"127.0.0.1:6164"});
   }
 };
 

paddle/fluid/operators/send_vars_op.cc

-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License. */
-
-#include <future>  // NOLINT
-#include <ostream>
-
-#include "paddle/fluid/framework/data_type.h"
-#include "paddle/fluid/framework/lod_tensor.h"
-#include "paddle/fluid/framework/op_registry.h"
-#include "paddle/fluid/operators/detail/grpc_client.h"
-#include "paddle/fluid/operators/send_recv_util.h"
-#include "paddle/fluid/platform/profiler.h"
-
-namespace paddle {
-namespace operators {
-
-class SendVarsOp : public framework::OperatorBase {
- public:
-  SendVarsOp(const std::string& type, const framework::VariableNameMap& inputs,
-             const framework::VariableNameMap& outputs,
-             const framework::AttributeMap& attrs)
-      : OperatorBase(type, inputs, outputs, attrs) {}
-
-  void RunImpl(const framework::Scope& scope,
-               const platform::Place& place) const override {
-    auto ins = Inputs("X");
-
-    std::vector<std::string> epmap = Attr<std::vector<std::string>>("epmap");
-    int sync_send = Attr<int>("sync_send");
-
-    platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
-    auto& ctx = *pool.Get(place);
-
-    // For profiling
-    platform::RecordEvent record_event(Type(), &ctx);
-
-    detail::RPCClient* rpc_client =
-        detail::RPCClient::GetInstance<detail::GRPCClient>();
-
-    for (size_t i = 0; i < ins.size(); i++) {
-      if (NeedSend(scope, ins[i])) {
-        VLOG(3) << "sending " << ins[i] << " to " << epmap[i];
-        // TODO(Yancey1989): we need to use an IO threadpool which has
-        // a larger number of threads than the computing threadpool.
-        rpc_client->AsyncSendVar(epmap[i], ctx, scope, ins[i]);
-      } else {
-        VLOG(3) << "don't send no-initialied variable: " << ins[i];
-      }
-    }
-    if (sync_send) {
-      rpc_client->Wait();
-    }
-  }
-};
-
-class SendVarsOpMaker : public framework::OpProtoAndCheckerMaker {
- public:
-  void Make() {
-    AddInput("X", "(Tensor, SelectedRows) Input variables to be sent")
-        .AsDuplicable();
-    AddComment(R"DOC(
-Send operator
-
-This operator will send variables to listen_and_serve op at the parameter server.
-)DOC");
-    AddAttr<int>("sync_send",
-                 "(int, default 0)"
-                 "sync send or async send.")
-        .SetDefault(0);
-    AddAttr<std::vector<std::string>>("epmap",
-                                      "(string vector, default 127.0.0.1:6164)"
-                                      "Server endpoints in the order of input "
-                                      "variables for mapping")
-        .SetDefault({"127.0.0.1:6164"});
-  }
-};
-
-class SendVarsOpShapeInference : public framework::InferShapeBase {
- public:
-  void operator()(framework::InferShapeContext* ctx) const override {}
-};
-
-}  // namespace operators
-}  // namespace paddle
-
-namespace ops = paddle::operators;
-
-REGISTER_OPERATOR(send_vars, ops::SendVarsOp,
-                  paddle::framework::EmptyGradOpMaker, ops::SendVarsOpMaker,
-                  ops::SendVarsOpShapeInference);

paddle/fluid/operators/sgd_op.cc

@@ -74,7 +74,8 @@ class SGDOpMaker : public framework::OpProtoAndCheckerMaker {
     AddInput("Grad", "(Tensor or SelectedRows) Input gradient");
     AddOutput("ParamOut",
               "(Tensor or SelectedRows, same with Param) "
-              "Output parameter, should share the same memory with Param");
+              "Output parameter, should share the same memory with Param")
+        .Reuse("Param");
     AddComment(R"DOC(
 
 SGD operator

paddle/fluid/operators/softmax_op.cc

@@ -83,7 +83,8 @@ class SoftmaxOpMaker : public framework::OpProtoAndCheckerMaker {
     AddInput("X",
              "The input tensor of softmax. "
              "2-D with shape [batch_size, input_feature_dimensions].");
-    AddOutput("Out", "The normalized values with the same shape as X.");
+    AddOutput("Out", "The normalized values with the same shape as X.")
+        .Reuse("X");
     AddAttr<bool>(
         "use_cudnn",
         "(bool, default false) Only used in cudnn kernel, need install cudnn")

paddle/fluid/operators/sum_op.cc

@@ -115,7 +115,7 @@ class SumOpMaker : public framework::OpProtoAndCheckerMaker {
   void Make() override {
     AddInput("X", "(vector<Tensor>) The input tensors of sum operator.")
         .AsDuplicable();
-    AddOutput("Out", "(Tensor) The output tensor of sum operator.");
+    AddOutput("Out", "(Tensor) The output tensor of sum operator.").Reuse("X");
     AddComment(R"DOC(
 Sum operator.
 

paddle/fluid/operators/test_send_nccl_id.cc

@@ -20,8 +20,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/operator.h"
 #include "paddle/fluid/framework/program_desc.h"
-#include "paddle/fluid/operators/detail/grpc_client.h"
-#include "paddle/fluid/operators/detail/grpc_server.h"
+#include "paddle/fluid/operators/detail/macros.h"
 #include "paddle/fluid/operators/detail/request_handler_impl.h"
 #include "paddle/fluid/operators/listen_and_serv_op.h"
 #include "paddle/fluid/operators/math/math_function.h"
@@ -29,6 +28,10 @@ limitations under the License. */
 #include "paddle/fluid/platform/nccl_helper.h"
 #include "paddle/fluid/string/printf.h"
 
+#ifdef PADDLE_WITH_GRPC
+#include "paddle/fluid/operators/send_recv_util.h"
+#endif
+
 USE_NO_KERNEL_OP(listen_and_serv);
 
 namespace f = paddle::framework;
@@ -37,7 +40,7 @@ namespace m = paddle::operators::math;
 namespace detail = paddle::operators::detail;
 namespace string = paddle::string;
 
-std::unique_ptr<detail::AsyncGRPCServer> g_rpc_service;
+std::unique_ptr<detail::RPCServer> g_rpc_service;
 std::unique_ptr<detail::RequestHandler> g_req_handler;
 
 void StartServer() {
@@ -58,7 +61,7 @@ void StartServer() {
   g_req_handler->SetRPCServer(g_rpc_service.get());
 
   std::thread server_thread(
-      std::bind(&detail::AsyncGRPCServer::StartServer, g_rpc_service.get()));
+      std::bind(&detail::RPCServer::StartServer, g_rpc_service.get()));
 
   g_rpc_service->SetCond(detail::kRequestSend);
   g_rpc_service->WaitBarrier(detail::kRequestSend);
@@ -68,9 +71,9 @@ void StartServer() {
   server_thread.join();
 }
 
-TEST(SendNcclId, GrpcServer) {
+TEST(SendNcclId, RPCServer) {
   g_req_handler.reset(new detail::RequestSendHandler(true));
-  g_rpc_service.reset(new detail::AsyncGRPCServer("127.0.0.1:0", 1));
+  g_rpc_service.reset(new RPCSERVER_T("127.0.0.1:0", 1));
 
   std::thread server_thread(StartServer);
   g_rpc_service->WaitServerReady();
@@ -87,8 +90,9 @@ TEST(SendNcclId, GrpcServer) {
   int port = g_rpc_service->GetSelectedPort();
 
   std::string ep = string::Sprintf("127.0.0.1:%d", port);
-  detail::RPCClient* client =
-      detail::RPCClient::GetInstance<detail::GRPCClient>();
+
+  detail::RPCClient* client = detail::RPCClient::GetInstance<RPCCLIENT_T>();
+
   LOG(INFO) << "connect to server" << ep;
   client->AsyncSendVar(ep, dev_ctx, scope, NCCL_ID_VARNAME);
   client->Wait();

paddle/fluid/operators/top_k_op.cc

@@ -50,7 +50,7 @@ class TopkOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() override {
     AddInput("X", "(Tensor) The input of Topk op");
-    AddOutput("Out", "(Tensor) The output tensor of Topk op");
+    AddOutput("Out", "(Tensor) The output tensor of Topk op").Reuse("X");
     AddOutput("Indices", "(Tensor) The indices of Topk elements of input");
     AddComment(R"DOC(
 Top K operator

paddle/fluid/platform/device_context.h

@@ -11,6 +11,7 @@ limitations under the License. */
 #pragma once
 
 #include <memory>
+#include <mutex>  // NOLINT
 #include <string>
 #include <unordered_map>
 #include <vector>
@@ -100,6 +101,7 @@ class CUDADeviceContext : public DeviceContext {
 
   template <typename Callback>
   void RecordEvent(cudaEvent_t ev, Callback callback) {
+    std::lock_guard<std::mutex> guard(mtx_);
     callback();
     PADDLE_ENFORCE(cudaEventRecord(ev, stream_));
   }
@@ -116,6 +118,8 @@ class CUDADeviceContext : public DeviceContext {
   int compute_capability;
   int multi_process;
   int max_threads_per_mp;
+
+  std::mutex mtx_;
 };
 
 template <>

paddle/fluid/pybind/pybind.cc

@@ -509,10 +509,10 @@ All parameter, weight, gradient are variables in Paddle.
             self.num_threads_ = num_threads;
           })
       .def_property(
-          "use_event",
-          [](const ExecutionStrategy &self) { return self.use_event_; },
-          [](ExecutionStrategy &self, bool use_event) {
-            self.use_event_ = use_event;
+          "use_cuda",
+          [](const ExecutionStrategy &self) { return self.use_cuda_; },
+          [](ExecutionStrategy &self, bool use_cuda) {
+            self.use_cuda_ = use_cuda;
           })
       .def_property(
           "allow_op_delay",

paddle/testing/paddle_gtest_main.cc

@@ -30,7 +30,7 @@ int main(int argc, char** argv) {
   new_argv.push_back(
       strdup("--tryfromenv=fraction_of_gpu_memory_to_use,use_pinned_memory"));
 #else
-  new_argv.push_back(strdup("--tryfromenv=use_pinned_memory"));
+  new_argv.push_back(strdup("--tryfromenv=use_pinned_memory,use_mkldnn"));
 #endif
   int new_argc = static_cast<int>(new_argv.size());
   char** new_argv_address = new_argv.data();

python/paddle/dataset/flowers.py

@@ -119,7 +119,8 @@ def reader_creator(data_file,
                 yield sample, int(label) - 1
 
     if use_xmap:
-        return xmap_readers(mapper, reader, cpu_count(), buffered_size)
+        cpu_num = int(os.environ.get('CPU_NUM', cpu_count()))
+        return xmap_readers(mapper, reader, cpu_num, buffered_size)
     else:
         return map_readers(mapper, reader)
 

python/paddle/fluid/__init__.py

@@ -117,7 +117,7 @@ def __bootstrap__():
 
     read_env_flags = [
         'use_pinned_memory', 'check_nan_inf', 'benchmark', 'warpctc_dir',
-        'eager_delete_scope'
+        'eager_delete_scope', 'use_mkldnn'
     ]
     if core.is_compiled_with_cuda():
         read_env_flags += [

python/paddle/fluid/data_feeder.py

@@ -15,6 +15,7 @@
 from __future__ import print_function
 import core
 import numpy
+import os
 import six.moves as six
 import multiprocessing
 
@@ -150,7 +151,9 @@ class DataFeeder(object):
         elif isinstance(self.place, core.CUDAPlace):
             return core.get_cuda_device_count()
         else:
-            return multiprocessing.cpu_count()
+            cpu_num = int(
+                os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
+            return cpu_num
 
     def decorate_reader(self,
                         reader,

python/paddle/fluid/layers/control_flow.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 import contextlib
 
-from layer_function_generator import autodoc
+from layer_function_generator import autodoc, templatedoc
 from tensor import assign, fill_constant
 from .. import core
 from ..framework import Program, Variable, Operator
@@ -721,26 +721,22 @@ def lod_rank_table(x, level=0):
     return table
 
 
+@templatedoc()
 def max_sequence_len(rank_table):
-    """Max Sequence Len Operator. Given a LoDRankTable object, this layer
-    returns the max length of a batch of sequences. In fact, a LoDRankTable
-    object contains a list of tuples(<sequence index, sequence length>) and
-    the list is already sorted by sequence length in descending order, so the
-    operator just returns the sequence length of the first tuple element.
+    """
+    ${comment}
+
+    >>> import paddle.fluid as fluid
+    >>> x = fluid.layers.data(name='x', shape=[10], dtype='float32',
+    >>>                       lod_level=1)
+    >>> rank_table = layers.lod_rank_table(x=x, level=0)
+    >>> max_seq_len = layers.max_sequence_len(rank_table)
 
     Args:
-        rank_table (Variable): Input variable which is a LoDRankTable object.
+        rank_table(${rank_table_type}): ${rank_table_comment}.
 
     Returns:
-        Variable: The max length of sequence.
-
-    Examples:
-        .. code-block:: python
-
-            x = fluid.layers.data(name='x', shape=[10],
-                            dtype='float32', lod_level=1)
-            rank_table = layers.lod_rank_table(x=x, level=0)
-            max_seq_len = layers.max_sequence_len(rank_table)
+        ${out_comment}.
     """
     helper = LayerHelper("max_seqence_len", **locals())
     res = helper.create_tmp_variable(dtype="int64")