merge develop branch

CMakeLists.txt

@@ -36,6 +36,7 @@ include(simd)
 
 ################################ Configurations #######################################
 option(WITH_GPU         "Compile PaddlePaddle with NVIDIA GPU"          ${CUDA_FOUND})
+option(WITH_AMD_GPU     "Compile PaddlePaddle with AMD GPU"             OFF)
 option(WITH_AVX         "Compile PaddlePaddle with AVX intrinsics"      ${AVX_FOUND})
 option(WITH_MKL         "Compile PaddlePaddle with MKL support."        ${AVX_FOUND})
 option(WITH_DSO         "Compile PaddlePaddle with dynamic linked CUDA" ON)
@@ -180,6 +181,11 @@ if(WITH_GPU)
     include(cuda)
 endif(WITH_GPU)
 
+if(WITH_AMD_GPU)
+    find_package(HIP)
+    include(hip)
+endif(WITH_AMD_GPU)
+
 if(WITH_MKLML)
     list(APPEND EXTERNAL_LIBS ${MKLML_IOMP_LIB})
 endif()

benchmark/cluster/vgg16/vgg16_fluid.py

@@ -18,12 +18,13 @@ import sys
 import time
 import numpy as np
 import paddle.v2 as paddle
-import paddle.v2.fluid as fluid
-import paddle.v2.fluid.core as core
-import paddle.v2.fluid.profiler as profiler
+import paddle.fluid as fluid
+import paddle.fluid.core as core
+import paddle.fluid.profiler as profiler
 import argparse
 import functools
 import os
+from paddle.fluid import debuger
 
 
 def str2bool(v):
@@ -182,7 +183,6 @@ def main():
             start_time = time.time()
             num_samples = 0
             train_pass_acc.reset()
-            with profiler.profiler("CPU", 'total') as prof:
             for batch_id, data in enumerate(train_reader()):
                 ts = time.time()
                 img_data = np.array(
@@ -254,9 +254,7 @@ def main():
             pserver_prog = t.get_pserver_program(current_endpoint)
             pserver_startup = t.get_startup_program(current_endpoint,
                                                     pserver_prog)
-            print("starting server side startup")
             exe.run(pserver_startup)
-            print("starting parameter server...")
             exe.run(pserver_prog)
         elif training_role == "TRAINER":
             # Parameter initialization

benchmark/cluster/vgg16/vgg16_tf.py

@@ -292,14 +292,18 @@ def run_benchmark(cluster_spec, server):
         return np.mean(test_accs)
 
     config = tf.ConfigProto(
-        intra_op_parallelism_threads=1, inter_op_parallelism_threads=1)
+        intra_op_parallelism_threads=1,
+        inter_op_parallelism_threads=1,
+        log_device_placement=True)
     config.gpu_options.allow_growth = True
 
     hooks = [tf.train.StopAtStepHook(last_step=1000000)]
 
     with tf.train.MonitoredTrainingSession(
-            master=server.target, is_chief=(args.task_index == 0),
-            hooks=hooks) as sess:
+            master=server.target,
+            is_chief=(args.task_index == 0),
+            hooks=hooks,
+            config=config) as sess:
         iters, num_samples, start_time = 0, 0, 0.0
         for pass_id in range(args.num_passes):
             # train

cmake/configure.cmake

@@ -57,11 +57,7 @@ if(NOT WITH_GOLANG)
     add_definitions(-DPADDLE_WITHOUT_GOLANG)
 endif(NOT WITH_GOLANG)
 
-if(NOT WITH_GPU)
-    add_definitions(-DHPPL_STUB_FUNC)
-
-    list(APPEND CMAKE_CXX_SOURCE_FILE_EXTENSIONS cu)
-else()
+if(WITH_GPU)
     add_definitions(-DPADDLE_WITH_CUDA)
 
     FIND_PACKAGE(CUDA REQUIRED)
@@ -84,7 +80,14 @@ else()
     # Include cuda and cudnn
     include_directories(${CUDNN_INCLUDE_DIR})
     include_directories(${CUDA_TOOLKIT_INCLUDE})
-endif(NOT WITH_GPU)
+elseif(WITH_AMD_GPU)
+    add_definitions(-DPADDLE_WITH_HIP)
+    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -D__HIP_PLATFORM_HCC__")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -D__HIP_PLATFORM_HCC__")
+else()
+    add_definitions(-DHPPL_STUB_FUNC)
+    list(APPEND CMAKE_CXX_SOURCE_FILE_EXTENSIONS cu)
+endif()
 
 if (WITH_MKLML AND MKLML_IOMP_LIB)
     message(STATUS "Enable Intel OpenMP with ${MKLML_IOMP_LIB}")

cmake/external/boost.cmake

@@ -24,7 +24,7 @@ set(BOOST_PROJECT       "extern_boost")
 # So we use 1.41.0 here.
 set(BOOST_VER           "1.41.0")
 set(BOOST_TAR           "boost_1_41_0")
-set(BOOST_URL           "http://paddlepaddledeps.s3-website-us-west-1.amazonaws.com/${BOOST_TAR}.tar.gz")
+set(BOOST_URL           "http://paddlepaddledeps.bj.bcebos.com/${BOOST_TAR}.tar.gz")
 set(BOOST_SOURCES_DIR ${THIRD_PARTY_PATH}/boost)
 set(BOOST_DOWNLOAD_DIR  "${BOOST_SOURCES_DIR}/src/${BOOST_PROJECT}")
 set(BOOST_INCLUDE_DIR "${BOOST_DOWNLOAD_DIR}/${BOOST_TAR}" CACHE PATH "boost include directory." FORCE)

cmake/external/eigen.cmake

@@ -4,7 +4,21 @@ SET(EIGEN_SOURCE_DIR ${THIRD_PARTY_PATH}/eigen3)
 SET(EIGEN_INCLUDE_DIR ${EIGEN_SOURCE_DIR}/src/extern_eigen3)
 INCLUDE_DIRECTORIES(${EIGEN_INCLUDE_DIR})
 
-ExternalProject_Add(
+if(WITH_AMD_GPU)
+    ExternalProject_Add(
+        extern_eigen3
+        ${EXTERNAL_PROJECT_LOG_ARGS}
+        GIT_REPOSITORY  "https://github.com/sabreshao/hipeigen.git"
+        GIT_TAG         0cba03ff9f8f9f70bbd92ac5857b031aa8fed6f9
+        PREFIX          ${EIGEN_SOURCE_DIR}
+        UPDATE_COMMAND  ""
+        CONFIGURE_COMMAND ""
+        BUILD_COMMAND     ""
+        INSTALL_COMMAND   ""
+        TEST_COMMAND      ""
+    )
+else()
+    ExternalProject_Add(
         extern_eigen3
         ${EXTERNAL_PROJECT_LOG_ARGS}
         GIT_REPOSITORY  "https://github.com/RLovelett/eigen.git"
@@ -15,7 +29,8 @@ ExternalProject_Add(
         BUILD_COMMAND     ""
         INSTALL_COMMAND   ""
         TEST_COMMAND      ""
-)
+    )
+endif()
 
 if (${CMAKE_VERSION} VERSION_LESS "3.3.0")
     set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/eigen3_dummy.c)

cmake/external/mklml.cmake

@@ -34,7 +34,7 @@ SET(MKLML_DOWNLOAD_DIR  "${MKLML_SOURCE_DIR}/src/${MKLML_PROJECT}")
 SET(MKLML_DST_DIR       "mklml")
 SET(MKLML_INSTALL_ROOT  "${THIRD_PARTY_PATH}/install")
 SET(MKLML_INSTALL_DIR   ${MKLML_INSTALL_ROOT}/${MKLML_DST_DIR})
-SET(MKLML_ROOT          ${MKLML_INSTALL_DIR}/${MKLML_VER})
+SET(MKLML_ROOT          ${MKLML_INSTALL_DIR})
 SET(MKLML_INC_DIR       ${MKLML_ROOT}/include)
 SET(MKLML_LIB_DIR       ${MKLML_ROOT}/lib)
 SET(MKLML_LIB           ${MKLML_LIB_DIR}/libmklml_intel.so)
@@ -46,7 +46,7 @@ INCLUDE_DIRECTORIES(${MKLML_INC_DIR})
 FILE(WRITE ${MKLML_DOWNLOAD_DIR}/CMakeLists.txt
   "PROJECT(MKLML)\n"
   "cmake_minimum_required(VERSION 3.0)\n"
-  "install(DIRECTORY ${MKLML_VER}\n"
+  "install(DIRECTORY ${MKLML_VER}/include ${MKLML_VER}/lib \n"
   "        DESTINATION ${MKLML_DST_DIR})\n")
 
 ExternalProject_Add(

cmake/generic.cmake

@@ -317,6 +317,82 @@ function(nv_test TARGET_NAME)
   endif()
 endfunction(nv_test)
 
+function(hip_library TARGET_NAME)
+  if (WITH_AMD_GPU)
+    set(options STATIC static SHARED shared)
+    set(oneValueArgs "")
+    set(multiValueArgs SRCS DEPS)
+    cmake_parse_arguments(hip_library "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+    set(_sources ${hip_library_SRCS})
+    HIP_PREPARE_TARGET_COMMANDS(${TARGET_NAME} OBJ _generated_files _source_files ${_sources} HIPCC_OPTIONS ${_hipcc_options} HCC_OPTIONS ${_hcc_options} NVCC_OPTIONS ${_nvcc_options})
+    if(_source_files)
+      list(REMOVE_ITEM _sources ${_source_files})
+    endif()
+    if(hip_library_SRCS)
+      if (hip_library_SHARED OR hip_library_shared) # build *.so
+        add_library(${TARGET_NAME} SHARED ${_cmake_options} ${_generated_files} ${_sources})
+        set_target_properties(${TARGET_NAME} PROPERTIES LINKER_LANGUAGE HIP)
+      else()
+        add_library(${TARGET_NAME} STATIC ${_cmake_options} ${_generated_files} ${_sources})
+        set_target_properties(${TARGET_NAME} PROPERTIES LINKER_LANGUAGE CXX)
+        target_link_libraries(${TARGET_NAME} /opt/rocm/hip/lib/libhip_hcc.so /opt/rocm/hip/lib/libhip_device.a)
+	find_fluid_modules(${TARGET_NAME})
+      endif()
+      if (hip_library_DEPS)
+	add_dependencies(${TARGET_NAME} ${hip_library_DEPS})
+	target_link_libraries(${TARGET_NAME} ${hip_library_DEPS})
+      endif()
+      # cpplint code style
+      foreach(source_file ${hip_library_SRCS})
+	string(REGEX REPLACE "\\.[^.]*$" "" source ${source_file})
+	if(EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h)
+	  list(APPEND hip_library_HEADERS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h)
+	endif()
+      endforeach()
+      add_style_check_target(${TARGET_NAME} ${hip_library_SRCS} ${hip_library_HEADERS})
+    else(hip_library_SRCS)
+      if (hip_library_DEPS)
+	merge_static_libs(${TARGET_NAME} ${hip_library_DEPS})
+      else()
+	message(FATAL "Please specify source file or library in nv_library.")
+      endif()
+    endif(hip_library_SRCS)
+  endif()
+endfunction(hip_library)
+
+function(hip_binary TARGET_NAME)
+  if (WITH_AMD_GPU)
+    set(options "")
+    set(oneValueArgs "")
+    set(multiValueArgs SRCS DEPS)
+    cmake_parse_arguments(hip_binary "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+    hip_add_executable(${TARGET_NAME} ${hip_binary_SRCS})
+    if(hip_binary_DEPS)
+      target_link_libraries(${TARGET_NAME} ${hip_binary_DEPS})
+      add_dependencies(${TARGET_NAME} ${hip_binary_DEPS})
+    endif()
+  endif()
+endfunction(hip_binary)
+
+function(hip_test TARGET_NAME)
+  if (WITH_AMD_GPU AND WITH_TESTING)
+    set(options "")
+    set(oneValueArgs "")
+    set(multiValueArgs SRCS DEPS)
+    cmake_parse_arguments(hip_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+    set(_sources ${hip_test_SRCS})
+    HIP_PREPARE_TARGET_COMMANDS(${TARGET_NAME} OBJ _generated_files _source_files ${_sources} HIPCC_OPTIONS ${_hipcc_options} HCC_OPTIONS ${_hcc_options} NVCC_OPTIONS ${_nvcc_options})
+    if(_source_files)
+      list(REMOVE_ITEM _sources ${_source_files})
+    endif()
+    add_executable(${TARGET_NAME} ${_cmake_options} ${_generated_files} ${_sources})
+    set_target_properties(${TARGET_NAME} PROPERTIES LINKER_LANGUAGE HIP)
+    target_link_libraries(${TARGET_NAME} ${hip_test_DEPS} paddle_gtest_main paddle_memory gtest gflags)
+    add_dependencies(${TARGET_NAME} ${hip_test_DEPS} paddle_gtest_main paddle_memory gtest gflags)
+    add_test(${TARGET_NAME} ${TARGET_NAME})
+  endif()
+endfunction(hip_test)
+
 function(go_library TARGET_NAME)
   set(options STATIC static SHARED shared)
   set(oneValueArgs "")

cmake/hip.cmake

+if(NOT WITH_AMD_GPU)
+    return()
+endif()
+
+include_directories("/opt/rocm/include")
+include_directories("/opt/rocm/hipblas/include")
+include_directories("/opt/rocm/hiprand/include")
+include_directories("/opt/rocm/rocrand/include")
+include_directories("/opt/rocm/rccl/include")
+include_directories("/opt/rocm/thrust")
+
+list(APPEND EXTERNAL_LIBS "-L/opt/rocm/lib/ -lhip_hcc")
+
+set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -fPIC -DPADDLE_WITH_HIP -std=c++14" )
+
+if(WITH_DSO)
+  set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_USE_DSO")
+endif(WITH_DSO)
+
+if(WITH_DOUBLE)
+  set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_TYPE_DOUBLE")
+endif(WITH_DOUBLE)
+
+if(WITH_TESTING)
+  set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_WITH_TESTING")
+endif(WITH_TESTING)
+
+if(CMAKE_BUILD_TYPE  STREQUAL "Debug")
+    list(APPEND HIP_HCC_FLAGS  ${CMAKE_CXX_FLAGS_DEBUG})
+elseif(CMAKE_BUILD_TYPE  STREQUAL "RelWithDebInfo")
+    list(APPEND HIP_HCC_FLAGS  ${CMAKE_CXX_FLAGS_RELWITHDEBINFO})
+elseif(CMAKE_BUILD_TYPE  STREQUAL "MinSizeRel")
+    list(APPEND HIP_HCC_FLAGS  ${CMAKE_CXX_FLAGS_MINSIZEREL})
+endif()
+
+if("x${HCC_HOME}" STREQUAL "x")
+  set(HCC_HOME "/opt/rocm/hcc")
+endif()
+
+set(CMAKE_HIP_LINK_EXECUTABLE "${HIP_HIPCC_CMAKE_LINKER_HELPER} ${HCC_HOME} <FLAGS> <CMAKE_CXX_LINK_FLAGS> <LINK_FLAGS> <OBJECTS> -o <TARGET> <LINK_LIBRARIES>")
+set(CMAKE_HIP_CREATE_SHARED_LIBRARY "${HIP_HIPCC_CMAKE_LINKER_HELPER} ${HCC_HOME} <CMAKE_CXX_LINK_FLAGS> <LINK_FLAGS> <OBJECTS> -o <TARGET> <LINK_LIBRARIES> -shared")
+set(CMAKE_HIP_CREATE_SHARED_MODULE "${HIP_HIPCC_CMAKE_LINKER_HELPER} ${HCC_HOME} <CMAKE_CXX_LINK_FLAGS> <LINK_FLAGS> <OBJECTS> -o <TARGET> <LINK_LIBRARIES> -shared")
+

cmake/inference_lib.cmake

@@ -69,6 +69,12 @@ if(NOT CBLAS_FOUND)
       SRCS ${CBLAS_INSTALL_DIR}/lib ${CBLAS_INSTALL_DIR}/include
       DSTS ${dst_dir} ${dst_dir}
     )
+elseif (WITH_MKLML)
+    set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/mklml")
+    copy(mklml_lib
+      SRCS ${MKLML_LIB} ${MKLML_IOMP_LIB} ${MKLML_INC_DIR}
+      DSTS ${dst_dir}/lib ${dst_dir}/lib ${dst_dir}
+    )
 endif()
 
 # paddle fluid module

doc/CMakeLists.txt

 add_subdirectory(v2)
+add_subdirectory(fluid)

doc/fluid/CMakeLists.txt

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

doc/fluid/build_and_install/index_cn.rst

+安装与使用
+------------

doc/fluid/build_and_install/index_en.rst

+Build and Install
+------------

doc/fluid/design/concurrent/go_op.md

+# go_op Design
+
+## Introduction
+
+The **go_op** allows user's of PaddlePaddle to run program blocks on a detached
+thread.  It works in conjuction with CSP operators (channel_send, 
+channel_receive, channel_open, channel_close, and select) to allow users to
+concurrently process data and communicate easily between different threads.
+
+## How to use it
+
+```
+channel = fluid.make_channel(dtype=core.VarDesc.VarType.LOD_TENSOR)
+
+with fluid.Go():
+    # Send a tensor of value 99 to "channel" on a detached thread
+    tensor = fill_constant(shape=[1], dtype='int', value=99)
+    tensor.stop_gradient = True
+    fluid.channel_send(channel, tensor)
+    
+# Receive sent tensor from "channel" on the main thread
+result = fill_constant(shape=[1], dtype='int', value=-1)    
+fluid.channel_recv(ch, result)  
+```
+
+The go operator can be accessed by using the fluid.Go() control flow.  This
+will create a new sub block, where the user can add additional operators
+to be ran on the thread.
+
+**Note:** Since back propegation is currently not support in the go_op, users
+should ensure that operators in the go block does not require gradient 
+calculations.
+
+## How it Works
+
+Similar to other control blocks, go_op will create a sub block and add it
+as a child to the current block.  Operators and variables defined in this
+block will be added to the go sub_block.
+
+In addition, the go operator will create a new child scope whose parent is
+the global scope.  Please refer to [block captures](#block-captures) for more
+information.
+
+When Paddle executor runs go_op, go_op will take the sub_block and pass it to
+the executor.run method (along with a newly created local scope) on a detached
+thread.
+
+An example of the generated program description is shown below.  Take note of
+the **go_op** in particular.  It is added as an operator in the current 
+block (in this example, block0).  The **go_op** contains a `sub_block`
+attribute, which points to the id of the block that will be executed in a 
+detached thread.
+
+```
+blocks {
+  idx: 0
+  parent_idx: -1
+  vars {
+    name: "return_value"
+    type {
+      type: LOD_TENSOR
+      lod_tensor {
+        tensor {
+          data_type: INT64
+        }
+      }
+    }
+  }
+  vars {
+    name: "status_recv"
+    type {
+      type: LOD_TENSOR
+      lod_tensor {
+        tensor {
+          data_type: BOOL
+        }
+      }
+    }
+  }
+  ...
+  ops {
+    outputs {
+      parameter: "Out"
+      arguments: "channel"
+    }
+    type: "channel_create"
+    attrs {
+      name: "data_type"
+      type: INT
+      i: 7
+    }
+    attrs {
+      name: "capacity"
+      type: INT
+      i: 0
+    }
+  }
+  ops {
+    inputs {
+      parameter: "X"
+      arguments: "channel"
+    }
+    type: "go"
+    attrs {
+      name: "sub_block"
+      type: BLOCK
+      block_idx: 1
+    }
+  }
+  ops {
+    inputs {
+      parameter: "Channel"
+      arguments: "channel"
+    }
+    outputs {
+      parameter: "Out"
+      arguments: "return_value"
+    }
+    outputs {
+      parameter: "Status"
+      arguments: "status_recv"
+    }
+    type: "channel_recv"
+  }
+  ...
+}
+
+blocks {
+  idx: 1
+  parent_idx: 0
+  vars {
+    name: "status"
+    type {
+      type: LOD_TENSOR
+      lod_tensor {
+        tensor {
+          data_type: BOOL
+        }
+      }
+    }
+  }
+  ...
+  
+  ops {
+    outputs {
+      parameter: "Out"
+      arguments: "fill_constant_1.tmp_0"
+    }
+    type: "fill_constant"
+    attrs {
+      name: "force_cpu"
+      type: BOOLEAN
+      b: false
+    }
+    attrs {
+      name: "value"
+      type: FLOAT
+      f: 99.0
+    }
+    attrs {
+      name: "shape"
+      type: INTS
+      ints: 1
+    }
+    attrs {
+      name: "dtype"
+      type: INT
+      i: 3
+    }
+  }
+  ops {
+    inputs {
+      parameter: "Channel"
+      arguments: "channel"
+    }
+    inputs {
+      parameter: "X"
+      arguments: "fill_constant_1.tmp_0"
+    }
+    outputs {
+      parameter: "Status"
+      arguments: "status"
+    }
+    type: "channel_send"
+    attrs {
+      name: "copy"
+      type: BOOLEAN
+      b: false
+    }
+  }
+```
+
+## Current Limitations
+
+#### <a name="block-captures"></a>Scopes and block captures:
+
+Paddle utilizes [scopes](./../concepts/scope.md) to store variables used in a
+block.  When a block is executed, a new local scope is created from the parent
+scope (ie: scope derived from the parent block) and associated with the new 
+child block.  After the block finishes executing, then the local scope and
+all associated variables in the scope is deleted.
+
+This works well in a single threaded scenario, however with introduction of
+go_op, a child block may continue to execute even after the parent block has
+exited.  If the go_op tries to access variables located in the parent block's
+scope, it may receive a segmentation fault because the parent scope may have
+been deleted.
+
+We need to implement block closures in order to prevent access to parent
+scope variables from causing a segmentation fault.  As a temporary workaround,
+please ensure that all variables accessed in the go block is not destructed
+before it is being accessed.  Currently, the go_op will explicitly enforce 
+this requirement and raise an exception if a variable could not be found in 
+the scope.
+
+Please refer to [Closure issue](https://github.com/PaddlePaddle/Paddle/issues/8502)
+for more details.
+
+#### Green Threads
+
+Golang utilizes `green threads`, which is a mechnism for the runtime library to 
+manage multiple threads (instead of natively by the OS).  Green threads usually
+allows for faster thread creation and switching, as there is less overhead
+when spawning these threads.  For the first version of CSP, we only support
+OS threads.
+
+
+#### Backward Propegation:
+
+go_op currently does not support backwards propagation.  Please use go_op with
+non training operators.

doc/fluid/design/concurrent/select_op.md

+# select_op Design
+
+## Introduction
+
+In golang, the [**select**](https://golang.org/ref/spec#Select_statements) 
+statement lets a goroutine wait on multiple communication operations at the 
+same time. The **select** blocks until one of its cases can run, then 
+executes the case.  If multiple cases are ready to run, then one case is 
+choosen at random to be executed.
+
+With the introduction of CSP for Paddle, we mimic this behavior by 
+creating a ***select_op***.
+
+## How to use it
+
+The **select_op** is available as a c++ operator.  However most users
+will prefer to use the much simplier Python API.
+
+- **fluid.Select()**: Creates a select operator and adds it to the current
+block within the main program.  Also creates a sub block and adds it to the 
+main program.  This sub block is used to hold all variables and operators 
+used by the case statements.
+ 
+Within the select block, users can add cases by 
+calling **select.case** or **select.default** method.
+
+- **fluid.Select.case(channel_action, channel, result_variable)**: Represents
+a fluid channel send/recv case.  This method creates a SelectCase block
+guard and adds it to the Select block.  The arguments into this method tells
+the select which channel operation to listen to.
+
+- **fluid.Select.default()**: Represents the fluid default case.  This default
+case is executed if none of the channel send/recv cases are available to
+execute.
+
+**Example:**
+```
+ch1 = fluid.make_channel(dtype=core.VarDesc.VarType.LOD_TENSOR)
+quit_ch = fluid.make_channel(dtype=core.VarDesc.VarType.LOD_TENSOR)
+            
+x = fill_constant(shape=[1], dtype=core.VarDesc.VarType.INT32, value=0)
+y = fill_constant(shape=[1], dtype=core.VarDesc.VarType.INT32, value=1)
+ 
+while_cond = fill_constant(shape=[1], dtype=core.VarDesc.VarType.BOOL, value=True)
+while_op = While(cond=while_cond)    
+ 
+with while_op.block():
+    with fluid.Select() as select:
+        with select.case(fluid.channel_send, channel, x):
+            # Send x, then perform Fibonacci calculation on x and y
+            x_tmp = fill_constant(shape=[1], dtype=core.VarDesc.VarType.INT32, value=0)
+            assign(input=x, output=x_tmp)
+            assign(input=y, output=x)
+            assign(elementwise_add(x=x_tmp, y=y), output=y)
+        with select.case(fluid.channel_recv, quit_channel, result2):
+            # Exit out of While loop
+            while_false = fill_constant(shape=[1], dtype=core.VarDesc.VarType.BOOL, value=False)
+            helper = layer_helper.LayerHelper('assign')
+            helper.append_op(
+                type='assign',
+                inputs={'X': [while_false]},
+                outputs={'Out': [while_cond]})
+```
+
+## How it Works
+
+### Program Description
+
+```
+blocks {
+  idx: 0
+  ...
+  // Create "case_to_execute" variable
+  ops {
+    outputs {
+      parameter: "Out"
+      arguments: "fill_constant_110.tmp_0"
+    }
+    type: "fill_constant"
+    attrs {
+      name: "force_cpu"
+      type: BOOLEAN
+      b: false
+    }
+    attrs {
+      name: "value"
+      type: FLOAT
+      f: -1.0
+    }
+    attrs {
+      name: "shape"
+      type: INTS
+      ints: 1
+    }
+    attrs {
+      name: "dtype"
+      type: INT
+      i: 2
+    }
+  }
+  // Create "select" operator.
+  // inputs: 
+  //   X: All input variables used by operators within the select block
+  //   case_to_execute: Variable filled in by select_op when it determines
+  //     which case to execute.
+  //  
+  // outputs:
+  //   Out: All output variables referenced by operators within select block. 
+  // 
+  // attrs:
+  //   sub_block: The block id containing the select "cases"
+  //   cases:  Serialized list of all cases in the select op. 
+  //     Each case is serialized as: '<index>,<type>,<channel>,<value>'
+  //     where type is 0 for default, 1 for send, and 2 for receive.
+  //     No channel and values are needed for default cases.
+  ops {
+    inputs {
+      parameter: "X"
+      arguments: "fill_constant_103.tmp_0"
+      arguments: "fill_constant_104.tmp_0"
+    }
+    inputs {
+      parameter: "case_to_execute"
+      arguments: "fill_constant_110.tmp_0"
+    }
+    outputs {
+      parameter: "Out"
+      arguments: "fill_constant_110.tmp_0"
+    }    
+    type: "select"
+    attrs {
+      name: "sub_block"
+      type: BLOCK
+      block_idx: 1
+    }
+    attrs {
+      name: "cases"
+      type: STRINGS
+      strings: "0,1,channel_101,fill_constant_109.tmp_0"
+      strings: "1,2,channel_102,fill_constant_108.tmp_0"
+    }
+  }
+  ...
+}
+```
+
+The python select API will add the **select_op** to the current block.  In addition, it will
+iterate through all it's case statements and add any input variables required by case statements
+into **X**.  It will also create a temp variable called **case_to_execute**.  This variable is
+filled in by the select_op after it has completed processing the case statements.
+
+If there are no available cases to execute (ie: all cases are blocked on channel operations, and
+there is no default statement), then the select_op will block the current thread.  The thread will 
+unblock once there is a channel operation affecting one of the case statements, at which point, the
+**select_op** will set the **case_to_execute** variable to the index of the case to execute.
+
+Finally the select_op will call executor.run on the **sub_block**.
+
+```
+blocks {
+  idx: 1
+  parent_idx: 0
+  ...
+  // Fill a tensor with the case index (ie: 0,1,2,3,ect.)
+  ops {
+    outputs {
+      parameter: "Out"
+      arguments: "fill_constant_111.tmp_0"
+    }
+    type: "fill_constant"
+    attrs {
+      name: "force_cpu"
+      type: BOOLEAN
+      b: false
+    }
+    attrs {
+      name: "value"
+      type: FLOAT
+      f: 0.0
+    }
+    attrs {
+      name: "shape"
+      type: INTS
+      ints: 1
+    }
+    attrs {
+      name: "dtype"
+      type: INT
+      i: 2
+    }
+  }
+  // Create an "equal" operator to compare the case index with the "case_to_execute"
+  // tensor (which was filled in by the select op).
+  ops {
+    inputs {
+      parameter: "X"
+      arguments: "fill_constant_111.tmp_0"  // case 0
+    }
+    inputs {
+      parameter: "Y"
+      arguments: "fill_constant_110.tmp_0"  // case_to_execute
+    }
+    outputs {
+      parameter: "Out"
+      arguments: "equal_0.tmp_0"
+    }
+    type: "equal"
+    attrs {
+      name: "axis"
+      type: INT
+      i: -1
+    }
+  }
+  // Use the output of the "equal" operator as a condition for the "conditional_block".
+  // If the condition evaluates to true, then execute the "sub_block" (which represents
+  // the select case's body)
+  ops {
+    inputs {
+      parameter: "Params"
+    }
+    inputs {
+      parameter: "X"
+      arguments: "equal_0.tmp_0"
+    }
+    outputs {
+      parameter: "Out"
+    }
+    outputs {
+      parameter: "Scope"
+      arguments: "_generated_var_0"
+    }
+    type: "conditional_block"
+    attrs {
+      name: "is_scalar_condition"
+      type: BOOLEAN
+      b: true
+    }
+    attrs {
+      name: "sub_block"
+      type: BLOCK
+      block_idx: 4
+    }
+  }
+  ...
+  // Repeat the above operators for each case statements inside the select body
+}
+
+```
+
+Cases are represented by a **conditional_block operator**, whose's condition is set as the output of 
+equal(**case_to_execute**, **case_index**).  Since each case index is unique in this sub-block, 
+only one case will be executed.
+
+### select_op flow
+
+<p align="center">
+<img src="./images/select_op_workflow.png"/><br/>
+</p>
+
+The select algorithm is inspired by golang's select routine.  Please refer to 
+http://www.tapirgames.com/blog/golang-concurrent-select-implementation for more information.
+
+## Backward Pass
+
+TODO

doc/design/distributed_lookup_table_design.md → doc/fluid/design/dist_train/distributed_lookup_table_design.md

@@ -26,7 +26,7 @@ lookup of rows.
 The following figure illustrates the multiplication of x with two
 non-zero elements, or say, two symbols, and a lookup table W:
 
-![lookup table](./lookup_table.png)
+![lookup table](./src/lookup_table.png)
 
 ### The Backward Algorithm
 
@@ -42,7 +42,7 @@ or some more sophisticated algorithms that rely on both W' and W:
 $$W = f(W, W')$$
 
 The following figure illustrates the backward pass of the lookup
-operator: ![lookup table training](./lookup_table_training.png)
+operator: ![lookup table training](./src/lookup_table_training.png)
 
 ## Distributed Storage Service
 

doc/fluid/design/dynamic_rnn/rnn.md

@@ -5,7 +5,7 @@ This document describes the RNN (Recurrent Neural Network) operator and how it i
 ## RNN Algorithm Implementation
 
 <p align="center">
-<img src="./images/rnn.jpg"/>
+<img src="./rnn.jpg"/>
 </p>
 
 The above diagram shows an RNN unrolled into a full network.
@@ -22,7 +22,7 @@ There are several important concepts here:
 There could be local variables defined in each step-net.  PaddlePaddle runtime realizes these variables in *step-scopes* which are created for each step.
 
 <p align="center">
-<img src="./images/rnn.png"/><br/>
+<img src="./rnn.png"/><br/>
 Figure 2 illustrates the RNN's data flow
 </p>
 
@@ -49,7 +49,7 @@ or copy the memory value of the previous step to the current ex-memory variable.
 
 ### Usage in Python
 
-For more information on Block, please refer to the [design doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/block.md).
+For more information on Block, please refer to the [design doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/block.md).
 
 We can define an RNN's step-net using a Block:
 
@@ -93,7 +93,7 @@ For example, we could have a 2-level RNN, where the top level corresponds to par
 The following figure illustrates feeding in text into the lower level, one sentence at a step, and the feeding in step outputs to the top level. The final top level output is about the whole text.
 
 <p align="center">
-<img src="./images/2_level_rnn.png"/>
+<img src="./2_level_rnn.png"/>
 </p>
 
 ```python
@@ -149,5 +149,5 @@ If the `output_all_steps` is set to False, it will only output the final time st
 
 
 <p align="center">
-<img src="images/rnn_2level_data.png"/>
+<img src="./rnn_2level_data.png"/>
 </p>

doc/fluid/design/index_cn.rst

+设计思想
+------------

doc/fluid/design/index_en.rst

+Design
+------------

doc/fluid/design/motivation/fluid.md

@@ -103,7 +103,7 @@ In computability theory, a system of data-manipulation rules, such as a programm
 
 There are two ways to execute a Fluid program.  When a program is executed, it creates a protobuf message [`ProgramDesc`](https://github.com/PaddlePaddle/Paddle/blob/a91efdde6910ce92a78e3aa7157412c4c88d9ee8/paddle/framework/framework.proto#L145) that describes the process and is conceptually like an [abstract syntax tree](https://en.wikipedia.org/wiki/Abstract_syntax_tree).
 
-There is a C++ class [`Executor`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/executor.h), which runs a `ProgramDesc`, similar to how an interpreter runs a Python program.
+There is a C++ class [`Executor`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/executor.h), which runs a `ProgramDesc`, similar to how an interpreter runs a Python program.
 
 Fluid is moving towards the direction of a compiler, which is explain in [fluid_compiler.md](fluid_compiler.md).
 

doc/fluid/dev/index_cn.rst

+开发标准
+------------

doc/fluid/dev/index_en.rst

+Development
+------------
+
+This is Development page

doc/fluid/faq/index_cn.rst

+FAQ
+------------

doc/fluid/faq/index_en.rst

+FAQ
+------------

doc/fluid/getstarted/index_cn.rst

+新手入门
+------------
+
+新手入门

doc/fluid/getstarted/index_en.rst

+GET STARTED
+------------
+
+This is get started page

doc/fluid/howto/cluster/fluid_cluster_train_cn.md

+# Fluid 分布式版本使用指南
+本篇文章将说明如何在PaddlePaddle Fluid版本下进行分布式训练的配置和执行，以及将单机训练脚本改造成支持集群训练的版本
+
+## 准备工作
+* 可用的集群
+
+    包含一个或多个计算节点的集群，每一个节点都能够执行PaddlePaddle的训练任务且拥有唯一的IP地址，集群内的所有计算节点可以通过网络相互通信。
+* 安装PaddlePaddle Fluid with Distribution版本
+
+    所有的计算节点上均需要按照分布式版本的PaddlePaddle, 在用于GPU等设备的机器上还需要额外安装好相应的驱动程序和CUDA的库。
+
+    **注意：**当前对外提供的PaddlePaddle版本并不支持分布式，需要通过源码重新编译。编译和安装方法参见[编译和安装指南](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/build_and_install/index_en.html)。
+    cmake编译命令中需要将WITH_DISTRIBUTE设置为ON，下面是一个cmake编译指令示例：
+``` bash
+cmake .. -DWITH_DOC=OFF -DWITH_GPU=OFF -DWITH_DISTRIBUTE=ON -DWITH_SWIG_PY=ON -DWITH_PYTHON=ON
+```
+
+## 更新训练脚本
+这里，我们以[Deep Learing 101](http://www.paddlepaddle.org/docs/develop/book/01.fit_a_line/index.html)课程中的第一章 fit a line 为例，描述如何将单机训练脚本改造成支持集群训练的版本。
+### 单机训练脚本示例
+```python
+import paddle.v2 as paddle
+import paddle.fluid as fluid
+
+x = fluid.layers.data(name='x', shape=[13], dtype='float32')
+y_predict = fluid.layers.fc(input=x, size=1, act=None)
+y = fluid.layers.data(name='y', shape=[1], dtype='float32')
+
+cost = fluid.layers.square_error_cost(input=y_predict, label=y)
+avg_cost = fluid.layers.mean(x=cost)
+
+sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001)
+sgd_optimizer.minimize(avg_cost)
+
+BATCH_SIZE = 20
+
+train_reader = paddle.batch(
+    paddle.reader.shuffle(
+        paddle.dataset.uci_housing.train(), buf_size=500),
+    batch_size=BATCH_SIZE)
+
+place = fluid.CPUPlace()
+feeder = fluid.DataFeeder(place=place, feed_list=[x, y])
+exe = fluid.Executor(place)
+
+exe.run(fluid.default_startup_program())
+
+PASS_NUM = 100
+for pass_id in range(PASS_NUM):
+    fluid.io.save_persistables(exe, "./fit_a_line.model/")
+    fluid.io.load_persistables(exe, "./fit_a_line.model/")
+    for data in train_reader():
+        avg_loss_value, = exe.run(fluid.default_main_program(),
+                                  feed=feeder.feed(data),
+                                  fetch_list=[avg_cost])
+
+        if avg_loss_value[0] < 10.0:
+            exit(0)  # if avg cost less than 10.0, we think our code is good.
+exit(1)
+```
+
+我们创建了一个简单的全连接神经网络程序，并且通过Fluid的Executor执行了100次迭代,现在我们需要将该单机版本的程序更新为分布式版本的程序。
+### 介绍Parameter Server
+在非分布式版本的训练脚本中，只存在Trainer一种角色，它不仅处理常规的计算任务，也处理参数相关的计算、保存和优化任务。在分布式版本的训练过程中，由于存在多个Trainer节点进行同样的数据计算任务，因此需要有一个中心化的节点来统一处理参数相关的保存和分配。在PaddlePaddle中，我们称这样的节点为[Parameter Server](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/dist_train/parameter_server.md)
+
+**因此，在分布式的Fluid环境中，我们有两个角色需要创建，分别是Parameter Server和Trainer。**
+
+### 分布式训练 
+Fliud专门提供了工具[Distributed Transpiler](https://github.com/PaddlePaddle/Paddle/blob/ba65d54d9d3b41cd3c5171b00f476d4e60133ddb/doc/fluid/design/dist_train/distributed_architecture.md#distributed-transpiler)用于将单机版的训练程序转换为分布式版本的训练程序。工具背后的理念是找出程序的优化算子和梯度参数，将他们分隔为两部分，通过send/recv 操作算子进行连接,优化算子和梯度参数可以在优化器的minimize函数的返回值中获取到。
+```python
+optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost) 
+```
+将Distributed Transpiler、优化算子和梯度函数放在一个代码中如下：
+```python
+... #define the program, cost, and create sgd optimizer
+
+optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost) #get optimize OPs and gradient parameters
+
+t = fluid.DistributeTranspiler() # create the transpiler instance
+# slice the program into 2 pieces with optimizer_ops and gradient parameters list, as well as pserver_endpoints, which is a comma separated list of [IP:PORT] and number of trainers
+t.transpile(optimize_ops, params_grads, pservers=pserver_endpoints, trainers=2)
+
+... #create executor
+
+# in pserver, run this
+#current_endpoint here means current pserver IP:PORT you wish to run on
+pserver_prog = t.get_pserver_program(current_endpoint)
+pserver_startup = t.get_startup_program(current_endpoint, pserver_prog)
+exe.run(pserver_startup)
+exe.run(pserver_prog)
+
+# in trainer, run this
+... # define data reader
+exe.run(fluid.default_startup_program())
+for pass_id in range(100):
+    for data in train_reader():
+        exe.run(t.get_trainer_program())
+```
+### 分布式训练脚本运行说明
+分布式任务的运行需要将表格中说明的多个参数进行赋值:
+
+| 参数名 | 值类型 | 说明 | 示例 |
+|:-------------|:------|:---------------------------------------|:-------------|
+| trainer_id | int | 当前训练节点的ID，训练节点ID编号为0 - n-1， n为trainers的值 | 0/1/2/3 |
+| pservers | str | parameter server 列表 | 127.0.0.1:6710,127.0.0.1:6711 |
+| trainers | int | 训练节点的总个数，>0的数字 | 4 |
+| server_endpoint | str | 当前所起的服务节点的IP:PORT | 127.0.0.1:8789 |
+| training_role | str | 节点角色， TRAINER/PSERVER | PSERVER |
+
+**注意：** ```training_role```是用来区分当前所起服务的角色的，用于训练程序中，用户可根据需要自行定义，其他参数为fluid.DistributeTranspiler的transpile函数所需要，需要在调用函数前进行定义，样例如下： 
+
+```python
+t = fluid.DistributeTranspiler()
+t.transpile(
+    optimize_ops,
+    params_grads,
+    trainer_id,
+    pservers=pserver,
+    trainers=trainers)
+if training_role == "PSERVER":
+    pserver_prog = t.get_pserver_program(server_endpoint)
+    pserver_startup = t.get_startup_program(server_endpoint, pserver_prog)
+```
+
+### Demo
+完整的demo代码位于Fluid的test目录下的[book](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/book/test_fit_a_line.py)中。
+
+第一步，进入demo代码所在目录：
+```bash
+cd /paddle/python/paddle/fluid/tests/book
+```
+
+第二步，启动Parameter Server：
+```bash
+PADDLE_INIT_PORT=6174 PADDLE_INIT_PSERVERS=192.168.1.2 TRAINERS=2 POD_IP=192.168.1.2 PADDLE_INIT_TRAINER_ID=1 TRAINING_ROLE=PSERVER python test_fit_a_line.py
+```
+执行命令后请等待出现提示： ```Server listening on 192.168.1.2:6174 ```, 表示Paramter Server已经正常启动。
+
+第三步，启动Trainer：
+```bash
+PADDLE_INIT_PORT=6174 PADDLE_INIT_PSERVERS=192.168.1.3 TRAINERS=2 POD_IP=192.168.1.3 PADDLE_INIT_TRAINER_ID=1 TRAINING_ROLE=TRAINER python test_fit_a_line.py
+```
+由于我们定义的Trainer的数量是2个，因此需要在另外一个计算节点上再启动一个Trainer。
+
+现在我们就启动了一个包含一个Parameter Server和两个Trainer的分布式训练任务。

doc/fluid/howto/index_cn.rst

+进阶使用
+------------

doc/fluid/howto/index_en.rst

+HOW TO
+------------
+
+This is how to page

doc/fluid/index_cn.rst

+ PaddlePaddle Fluid
+==========================
+
+..  toctree::
+  :maxdepth: 1
+
+  getstarted/index_cn.rst
+  design/index_cn.rst
+  build_and_install/index_cn.rst
+  howto/index_cn.rst
+  dev/index_cn.rst
+  faq/index_cn.rst

doc/fluid/index_en.rst

+ PaddlePaddle Fluid
+==========================
+
+..  toctree::
+  :maxdepth: 1
+
+  getstarted/index_en.rst
+  design/index_en.rst
+  build_and_install/index_en.rst
+  howto/index_en.rst
+  dev/index_en.rst
+  faq/index_en.rst

doc/v2/api/fluid/optimizer.rst

@@ -47,3 +47,10 @@ DecayedAdagrad
     :members:
     :noindex:
 
+Adadelta
+--------------
+
+..  autoclass:: paddle.fluid.optimizer.AdadeltaOptimizer
+    :members:
+    :noindex:
+

doc/v2/dev/index_en.rst

 Development
 ------------
 
+
+PaddlePaddle adheres to the following three sections of code and document specifications.
+
+
+PaddlePaddle uses git for version control and Docker is used for building and testing environment. The code includes Cuda, C++, Python, Shell and other programming languages，which comply with Google C++ Style, Pep-8, and the code base includes style checking by an automatic inspection tool. Code comments need to follow the Doxygen specification. The code that does not meet the style requirements will fail to compile. We provide the following guidelines for the use of Git, build tests and code development.
 ..  toctree::
   :maxdepth: 1
 
   contribute_to_paddle_en.md
+
+
+PaddlePaddle is well documented in English and Chinese. We recommend using the English version of the documents and problem description. The design documents focus on problem descriptions, backgrounds, and are followed by solutions. As documents are generated by Sphinx, code comments should comply with the Sphinx documentation standard. We recommend to use the paddlepaddle.org tool to compile and generate and preview documents locally. Please refer to:
+
+..  toctree::
+  :maxdepth: 1
+
   write_docs_en.rst
+
+PaddlePaddle V2 defines new operations by adding new Layers. You can implement various complex layers by combining basic APIs to satisfy most applications. If you want to customize layer, please refer to the following, and welcome to propose patch.
+
+..  toctree::
+  :maxdepth: 1
+
   new_layer_en.rst

doc/v2/dev/write_docs_cn.rst

@@ -2,13 +2,14 @@
 如何贡献文档
 #############
 
-PaddlePaddle的文档包括中英文两个部分。文档都是通过 ``cmake`` 驱动 ``sphinx`` 编译生成，也可以利用paddlepaddle.org工具来编译和预览文档。
+PaddlePaddle的文档包括中英文两个部分。文档都是通过 ``cmake`` 驱动 ``sphinx`` 编译生成的，PaddlePaddle.org工具可以帮助我们实现这一编译过程，并提供更好的预览效果。
 
 如何构建文档
 ============
 
 PaddlePaddle的文档构建有两种方式，分别为使用paddlepaddle.org工具和不使用paddlepaddle.org工具，两种方式都有各自的优点，前者方便预览，后者方便开发者进行调试。这两种方式中又分别有使用docker和不使用docker的两种构建方法。
 
+我们建议使用PaddlePaddle.org工具来构建文档。
 
 使用PaddlePaddle.org工具
 ------------------------
@@ -31,7 +32,7 @@ PaddlePaddle.org工具可以配合Docker使用，需要在系统里先安装好D
     docker run -it -p 8000:8000 -v `pwd`:/var/content paddlepaddle/paddlepaddle.org:latest
 
 注意: PaddlePaddle.org 会在 -v (volume) 指定的内容存储库运行命令
-之后再用网页连到http://localhost:8000就可以在网页上生成需要的文档
+之后再用网页连到 http://localhost:8000 就可以在网页上生成需要的文档
 编译后的文件将被存储在工作目录 <paddlepaddle working directory>/.ppo_workspace/content。
 
 如果不想使用Docker，你还可以通过运行Django框架直接激活工具的服务器。使用下面的命令来运行它。
@@ -56,7 +57,7 @@ PaddlePaddle.org工具可以配合Docker使用，需要在系统里先安装好D
     python manage.py runserver
 
 工具服务器将读取环境变量 CONTENT_DIR 搜索代码库。请指定的PaddlePaddle工作目录给环境变量 CONTENT_DIR。
-之后再用网页连到http://localhost:8000就可以在网页上生成需要的文档。
+之后再用网页连到 http://localhost:8000 就可以在网页上生成需要的文档。
 编译后的文件将被存储在工作目录 <paddlepaddle working directory>/.ppo_workspace/content。
 
 想了解更多PaddlePaddle.org工具的详细信息，可以 `点击这里 <https://github.com/PaddlePaddle/PaddlePaddle.org/blob/develop/README.cn.md>`_ 。
@@ -96,7 +97,7 @@ PaddlePaddle.org工具可以配合Docker使用，需要在系统里先安装好D
 
    python -m SimpleHTTPServer 8088
 
-在浏览器中输入http://localhost:8088就可以看到编译生成的中/英文的文档页面和英文的API页面,下图为生成的英文文档首页示例。注意，示例中由于使用了sphinx的原始主题，所以页面的风格与官网并不一致，但这并不影响开发者进行调试。
+在浏览器中输入 http://localhost:8088 就可以看到编译生成的中/英文的文档页面和英文的API页面,下图为生成的英文文档首页示例。注意，示例中由于使用了sphinx的原始主题，所以页面的风格与官网并不一致，但这并不影响开发者进行调试。
 
 ..  image:: src/doc_en.png
     :align: center

doc/v2/dev/write_docs_en.rst

@@ -2,21 +2,20 @@
 Contribute Documentation
 ########################
 
-PaddlePaddle supports English documentation ``doc`` and Chinese documentation ``doc_cn``.
-Both are compiled by `cmake`_ and `sphinx`_ , the compiled documentations will be stored under ``doc`` and ``doc_cn`` directories.
-When using the PaddlePaddle.org to compile documentations, the compiled documentations will be stored under a consolidated directory: .ppo_workspace/content
+PaddlePaddle's documentation includes both Chinese and English versions. The documentation is built using the ``cmake`` command to drive the ``sphinx`` compiler. The PaddlePaddle.org tool helps us to implement this compilation process and provides better preview results.
 
-How to Build Documentations
-============
+How to build Documentation
+===========================
 
-We recommend using PaddlePaddle.org tool to build documentation
+PaddlePaddle's documentation is built in two ways: using the PaddlePaddle.org tool and without using it. Both methods have their own advantages. The former facilitates previewing, while the latter facilitates debugging by the developer. We could choose to build the documentation with Docker or without it in each of the above ways.
 
+We recommend using PaddlePaddle.org tool to build documentation.
 
-Use PaddlePaddle.org tool
---------------
-This is the recommended method to build documentation. It can compile documentation and preview the documentation in a web browser.
+Using PaddlePaddle.org tool
+-----------------------------
+This is the recommended method to build documentation, because it can automatically compile the documentation and preview the documentation directly in a web page. Note that, although you can preview the documentation in other ways, its style may not be consistent with the official website. Compiling with the PaddlePaddle.org tool produces a preview that will be consistent with the official website documentation style.
 
-The tool uses Docker, please install it on your system. Please check Docker official website on how to install Docker. You may use the following commands to activate the tool
+The PaddlePaddle.org tool can be used with Docker and Docker needs to be installed first. Please refer to `Docker's official website <https://docs.docker.com/>`_ on how to install Docker. After installing Docker, you may use the following commands to activate the tool
 
 ..  code-block:: bash
 
@@ -32,8 +31,8 @@ The tool uses Docker, please install it on your system. Please check Docker offi
     # Please specify the working directory through -v
     docker run -it -p 8000:8000 -v `pwd`:/var/content paddlepaddle/paddlepaddle.org:latest
 
-Note: PaddlePaddle.org will read the content repos specified in the -v (volume) flag of the docker run command
-Use a web browser and navigate to http://localhost:8000, click the buttons to compile the documentation
+Note: PaddlePaddle.org will read the content repos specified in the -v (volume) flag of the docker run commands
+Use a web browser and navigate to http://localhost:8000. Click the buttons to compile the documentation.
 The compiled documentations will be stored in <paddlepaddle working directory>/.ppo_workspace/content
 
 
@@ -58,19 +57,62 @@ If you don't wish to use Docker, you can also activate the tool through Django.
     pip install -r requirements.txt
     python manage.py runserver
 
-Use a web browser and navigate to http://localhost:8000, click the buttons to compile the documentation
+Specify the PaddlePaddle working directory for the environment variable CONTENT_DIR so that the tool could find where the working directory is.
+
+Use a web browser and navigate to http://localhost:8000. Click the buttons to compile the documentation
 The compiled documentations will be stored in <paddlepaddle working directory>/.ppo_workspace/content
 
-If you want to learn more on the PaddlePaddle.org, please `click here <https://github.com/PaddlePaddle/PaddlePaddle.org/blob/develop/README.md>`_ 。
+Please `click here <https://github.com/PaddlePaddle/PaddlePaddle.org/blob/develop/README.md>`_ for more information about the PaddlePaddle.org tool.
+
+
+Manually Building the Documentation
+-------------------------------------
+
+Build PaddlePaddle's documentation with Docker，you need to install Docker first. Please refer to `Docker's official website <https://docs.docker.com/>`_ on how to install Docker. After Docker is installed, you could use the scripts in the source directory to build the documentation.
+
+[TBD]
+
+If you do not wish to use Docker, you can also use the following commands to directly build the PaddlePaddle documentation.
+
+.. code-block:: bash
+
+   mkdir paddle
+   cd paddle
+   git clone https://github.com/PaddlePaddle/Paddle.git
+   mkdir -p build
+   cd build
+   cmake .. -DCMAKE_BUILD_TYPE=Release -DWITH_GPU=OFF -DWITH_MKL=OFF -DWITH_DOC=ON
+
+   # If you only need to build documents, use the following commands
+   make -j $processors gen_proto_py
+   make -j $processors paddle_docs paddle_docs_cn
+
+   # If you only need to build APIs, use the following commands
+   make -j $processors gen_proto_py framework_py_proto
+   make -j $processors copy_paddle_pybind
+   make -j $processors paddle_api_docs
+
+$processors indicates that as many processes as the CPU cores are started to compile in parallel. It should be set according to the number of CPU cores of your machine.
+
+After the compilation is complete, enter the ``doc/v2`` directory. If you chose to build documents, it will generate ``cn/html/`` and ``en/html`` subdirectories under this directory. If you chose to build APIs，it will generate``api/en/html`` subdirectory. Please enter these directories respectively and execute the following commands:
+
+.. code-block:: bash
+
+   python -m SimpleHTTPServer 8088
+
+Use a web browser and navigate to http://localhost:8000, you could see the compiled Chinese/English documents page and the English APIs page. The following figure is an example of the built English documents home page. Note that due to the sphinx's original theme used in the example, the style of the page is not consistent with the official website, but this does not affect the developer's debugging.
 
-How to write Documentations
-============
+..  image:: src/doc_en.png
+    :align: center
+    :scale: 60 %
 
-PaddlePaddle uses `sphinx`_ to compile documentations，Please check sphinx official website for more detail.
+How to write Documentation
+===========================
 
+PaddlePaddle uses `sphinx`_ to compile documentation，Please check sphinx official website for more detail.
 
 How to update www.paddlepaddle.org
-============================
+===================================
 
 Please create PRs and submit them to github, please check `Contribute Code <http://www.paddlepaddle.org/docs/develop/documentation/en/howto/dev/contribute_to_paddle_en.html>`_ 。
 PaddlePaddle develop branch will update the documentation once the PR is merged. User may check latest `Chinese Docs <http://www.paddlepaddle.org/docs/develop/documentation/zh/getstarted/index_cn.html>`_ and

doc/v2/faq/index_en.rst

 FAQ
 ====
 
+This document provides answers to some of the frequently asked questions about PaddlePaddle. If you have a question that is not covered here, please go to `PaddlePaddle Community <https://github.com/PaddlePaddle/Paddle/issues>`_ , to find an answer or submit new `issue <https://github.com/PaddlePaddle/Paddle/issues/new>`_  , we will reply in time.
 
 ..  toctree::
   :maxdepth: 1

doc/v2/getstarted/index_en.rst

 GET STARTED
 ============
 
+If you want to quickly know how to use PaddlePaddle, please refer to the following guide:
+
 ..  toctree::
   :maxdepth: 1
 
   quickstart_en.rst
+  
+  
+While using PaddlePaddle to build applications, please understand some basic concepts.
+
+Here is an example of linear regression. It introduces workflow of PaddlePaddle, including data format, model configuration and training, etc.
+  
+..  toctree::
+  :maxdepth: 1
+  
   concepts/use_concepts_en.rst

doc/v2/howto/capi/index_en.rst

-C-API Prediction Library
+C-API Inference Library
 ========================
 
+After we train a neural network, we use it to do inference. Inference is the process of preparing input data and propagating it through the model to produce the result.
+
+Compared with model training, prediction has the following features:
+
+#. Inference does not require backpropagation and parameter updates, as required during training.
+#. Labels are not needed in prediction.
+#. Most of the time, predictions need to be integrated with the user system.
+
+Therefore, the model prediction SDK needs to be designed separately and has the following features:
+
+#. The predictive SDK does not include backpropagation and parameter updates to reduce the size of the SDK.
+#. The predictive SDK needs a simple user interface for ease of use.
+#. Since the input data may have a variety of structures, the format of the input data is clearly and compactly packaged.
+#. In order to be compatible with user's system, the SDK's interface must conform to the C-standard interface.
+
+PaddlePaddle provides C-API to solve the above problem. Following are the guidelines to use the C-API:
+
 ..  toctree::
   :maxdepth: 1
 

doc/v2/howto/cluster/index_en.rst

@@ -2,6 +2,9 @@ Distributed Training
 ====================
 
 The effectiveness of the deep learning model is often directly related to the scale of the data: it can generally achieve better results after increasing the size of the dataset on the same model. However, it can not fit in one single computer when the amount of data increases to a certain extent. At this point, using multiple computers for distributed training is a natural solution. In distributed training, the training data is divided into multiple copies (sharding), and multiple machines participating in the training read their own data for training and collaboratively update the parameters of the overall model.
+
+Distributed training generally has framwork as shown below:
+
 .. image:: src/ps_en.png
    :width: 500
 

doc/v2/howto/index_en.rst

 HOW TO
-=======
+========
+
+PaddlePaddle provides the users the ability to flexibly set various command line parameters to control the model training and inference process. Please refer to the following instructions on using PaddlePaddle:
 
 ..  toctree::
   :maxdepth: 1
 
   cmd_parameter/index_en.rst
+
+PaddlePaddle supports distributed training tasks on fabric clusters, MPI clusters, and Kubernetes clusters. For detailed configuration and usage instructions, refer to:
+
+..  toctree::
+  :maxdepth: 1
+
   cluster/index_en.rst
+
+PaddlePaddle provides a C-API for inference. We provide the following guidelines  for using the C-API:
+
+..  toctree::
+  :maxdepth: 1
+
   capi/index_en.rst
+
+PaddlePaddle supports a variety of flexible and efficient recurrent neural networks. For details, please refer to：
+
+..  toctree::
+  :maxdepth: 1
+
   rnn/index_en.rst
+
+How to use the built-in timing tool, nvprof, or nvvp to run performance analysis and tuning, please refer to：
+
+..  toctree::
+  :maxdepth: 1
+
   optimization/gpu_profiling_en.rst

paddle/fluid/framework/channel.h

@@ -34,7 +34,7 @@ class Channel {
  public:
   virtual bool CanSend() = 0;
   virtual bool CanReceive() = 0;
-  virtual bool Send(T*) = 0;
+  virtual void Send(T*) = 0;
   virtual bool Receive(T*) = 0;
   virtual size_t Cap() = 0;
   virtual void Lock() = 0;
@@ -84,96 +84,113 @@ class ChannelHolder {
   }
 
   template <typename T>
-  bool Send(T* data) {
-    if (!IsInitialized()) return false;
-    PADDLE_ENFORCE_EQ(holder_->Type(), std::type_index(typeid(T)));
+  void Send(T* data) {
+    PADDLE_ENFORCE_EQ(IsInitialized(), true,
+                      "The Channel hasn't been initialized");
+    PADDLE_ENFORCE_EQ(
+        holder_->Type(), std::type_index(typeid(T)),
+        "Channel type is not same as the type of the data being sent");
     // Static cast should be safe because we have ensured that types are same
     Channel<T>* channel = static_cast<Channel<T>*>(holder_->Ptr());
-    return channel != nullptr ? channel->Send(data) : false;
+    PADDLE_ENFORCE_EQ(channel != nullptr, true, "Channel should not be null.");
+    channel->Send(data);
   }
 
   template <typename T>
   bool Receive(T* data) {
-    if (!IsInitialized()) return false;
-    PADDLE_ENFORCE_EQ(holder_->Type(), std::type_index(typeid(T)));
+    PADDLE_ENFORCE_EQ(IsInitialized(), true,
+                      "The Channel hasn't been initialized");
+    PADDLE_ENFORCE_EQ(
+        holder_->Type(), std::type_index(typeid(T)),
+        "Channel type is not same as the type of the data being sent");
     Channel<T>* channel = static_cast<Channel<T>*>(holder_->Ptr());
-    return channel != nullptr ? channel->Receive(data) : false;
+    PADDLE_ENFORCE_EQ(channel != nullptr, true, "Channel should not be null.");
+    return channel->Receive(data);
   }
 
   bool IsClosed() {
-    if (IsInitialized()) {
+    PADDLE_ENFORCE_EQ(IsInitialized(), true,
+                      "The Channel hasn't been initialized");
     return holder_->IsClosed();
   }
-    return false;
-  }
 
   bool CanSend() {
-    if (IsInitialized()) {
+    PADDLE_ENFORCE_EQ(IsInitialized(), true,
+                      "The Channel hasn't been initialized");
     return holder_->CanSend();
   }
-    return false;
-  }
 
   bool CanReceive() {
-    if (IsInitialized()) {
+    PADDLE_ENFORCE_EQ(IsInitialized(), true,
+                      "The Channel hasn't been initialized");
     return holder_->CanReceive();
   }
-    return false;
-  }
 
   void close() {
-    if (IsInitialized()) holder_->Close();
+    PADDLE_ENFORCE_EQ(IsInitialized(), true,
+                      "The Channel hasn't been initialized");
+    holder_->Close();
   }
 
   size_t Cap() {
-    if (IsInitialized()) return holder_->Cap();
-    return -1;
+    PADDLE_ENFORCE_EQ(IsInitialized(), true,
+                      "The Channel hasn't been initialized");
+    return holder_->Cap();
   }
 
   void Lock() {
-    if (IsInitialized()) holder_->Lock();
+    PADDLE_ENFORCE_EQ(IsInitialized(), true,
+                      "The Channel hasn't been initialized");
+    holder_->Lock();
   }
 
   void Unlock() {
-    if (IsInitialized()) holder_->Unlock();
+    PADDLE_ENFORCE_EQ(IsInitialized(), true,
+                      "The Channel hasn't been initialized");
+    holder_->Unlock();
   }
 
   template <typename T>
   void AddToSendQ(const void* referrer, T* data,
                   std::shared_ptr<std::condition_variable_any> cond,
                   std::function<bool(ChannelAction)> cb) {
-    if (IsInitialized()) {
+    PADDLE_ENFORCE_EQ(IsInitialized(), true,
+                      "The Channel hasn't been initialized");
     Channel<T>* channel = static_cast<Channel<T>*>(holder_->Ptr());
     if (channel != nullptr) {
       channel->AddToSendQ(referrer, data, cond, cb);
     }
   }
-  }
 
   template <typename T>
   void AddToReceiveQ(const void* referrer, T* data,
                      std::shared_ptr<std::condition_variable_any> cond,
                      std::function<bool(ChannelAction)> cb) {
-    if (IsInitialized()) {
+    PADDLE_ENFORCE_EQ(IsInitialized(), true,
+                      "The Channel hasn't been initialized");
     Channel<T>* channel = static_cast<Channel<T>*>(holder_->Ptr());
     if (channel != nullptr) {
       channel->AddToReceiveQ(referrer, data, cond, cb);
     }
   }
-  }
 
   void RemoveFromSendQ(const void* referrer) {
-    if (IsInitialized()) holder_->RemoveFromSendQ(referrer);
+    PADDLE_ENFORCE_EQ(IsInitialized(), true,
+                      "The Channel hasn't been initialized");
+    holder_->RemoveFromSendQ(referrer);
   }
 
   void RemoveFromReceiveQ(const void* referrer) {
-    if (IsInitialized()) holder_->RemoveFromReceiveQ(referrer);
+    PADDLE_ENFORCE_EQ(IsInitialized(), true,
+                      "The Channel hasn't been initialized");
+    holder_->RemoveFromReceiveQ(referrer);
   }
 
   inline bool IsInitialized() const { return holder_ != nullptr; }
 
   inline const std::type_index Type() {
-    PADDLE_ENFORCE_EQ(IsInitialized(), true);
+    PADDLE_ENFORCE_EQ(IsInitialized(), true,
+                      "The Channel hasn't been initialized");
     return holder_->Type();
   }
 

paddle/fluid/framework/channel_impl.h

@@ -31,7 +31,7 @@ class ChannelImpl : public paddle::framework::Channel<T> {
  public:
   virtual bool CanSend();
   virtual bool CanReceive();
-  virtual bool Send(T *);
+  virtual void Send(T *);
   virtual bool Receive(T *);
   virtual size_t Cap() { return cap_; }
   virtual void Lock();
@@ -76,10 +76,9 @@ class ChannelImpl : public paddle::framework::Channel<T> {
     }
   };
 
-  bool send_return(bool value) {
+  void send_return() {
     send_ctr--;
     destructor_cond_.notify_all();
-    return value;
   }
 
   bool recv_return(bool value) {
@@ -118,15 +117,15 @@ bool ChannelImpl<T>::CanReceive() {
 }
 
 template <typename T>
-bool ChannelImpl<T>::Send(T *item) {
+void ChannelImpl<T>::Send(T *item) {
   send_ctr++;
   std::unique_lock<std::recursive_mutex> lock{mu_};
 
-  // If channel is closed, do nothing
+  // If channel is closed, throw exception
   if (closed_) {
     lock.unlock();
-    // TODO(abhinavarora) Should panic on closed channel
-    return send_return(false);
+    send_return();
+    PADDLE_THROW("Cannot send on closed channel");
   }
 
   // If there is a receiver, directly pass the value we want
@@ -143,7 +142,7 @@ bool ChannelImpl<T>::Send(T *item) {
     if (m->callback != nullptr) do_send = m->callback(ChannelAction::SEND);
     if (do_send)
       *(m->data) = std::move(*item);
-    else
+    else {
       // We cannot do the data transfer because
       // this QueueMessage was added by Select
       // and some other case was executed.
@@ -151,12 +150,17 @@ bool ChannelImpl<T>::Send(T *item) {
       // We do not care about notifying other
       // because they would have been notified
       // by the executed select case.
-      return send_return(Send(item));
+      lock.unlock();
+      Send(item);
+      send_return();
+      return;
+    }
 
     // Wake up the blocked process and unlock
     m->Notify();
     lock.unlock();
-    return send_return(true);
+    send_return();
+    return;
   }
 
   // Unbuffered channel will always bypass this
@@ -167,7 +171,8 @@ bool ChannelImpl<T>::Send(T *item) {
     buf_.push_back(std::move(*item));
     // Release lock and return true
     lock.unlock();
-    return send_return(true);
+    send_return();
+    return;
   }
 
   // Block on channel, because some receiver will complete
@@ -175,8 +180,12 @@ bool ChannelImpl<T>::Send(T *item) {
   auto m = std::make_shared<QueueMessage>(item);
   sendq.push_back(m);
   m->Wait(lock);
-  // TODO(abhinavarora) Should panic on closed channel
-  return send_return(!m->chan_closed);
+  if (m->chan_closed) {
+    lock.unlock();
+    send_return();
+    PADDLE_THROW("Cannot send on closed channel");
+  }
+  send_return();
 }
 
 template <typename T>

paddle/fluid/framework/channel_test.cc

@@ -16,7 +16,6 @@ limitations under the License. */
 
 #include <chrono>
 #include <thread>
-
 #include "gtest/gtest.h"
 
 using paddle::framework::Channel;
@@ -41,7 +40,7 @@ void RecevingOrderEqualToSendingOrder(Channel<int> *ch) {
   unsigned sum_send = 0;
   std::thread t([&]() {
     for (int i = 0; i < 5; i++) {
-      EXPECT_EQ(ch->Send(&i), true);
+      ch->Send(&i);
       sum_send += i;
     }
   });
@@ -61,7 +60,7 @@ TEST(Channel, SufficientBufferSizeDoesntBlock) {
   const size_t buffer_size = 10;
   auto ch = MakeChannel<size_t>(buffer_size);
   for (size_t i = 0; i < buffer_size; ++i) {
-    EXPECT_EQ(ch->Send(&i), true);  // should not block
+    ch->Send(&i);
   }
 
   size_t out;
@@ -82,7 +81,7 @@ void SendReceiveWithACloseChannelShouldPanic(Channel<size_t> *ch) {
   const size_t data = 5;
   std::thread send_thread{[&]() {
     size_t i = data;
-    EXPECT_EQ(ch->Send(&i), true);  // should not block
+    ch->Send(&i);  // should not block
   }};
 
   std::thread recv_thread{[&]() {
@@ -94,12 +93,18 @@ void SendReceiveWithACloseChannelShouldPanic(Channel<size_t> *ch) {
   send_thread.join();
   recv_thread.join();
 
-  // After closing send should return false. Receive should
-  // also return false as there is no data in queue.
+  // After closing send should panic. Receive should
+  // also  false as there is no data in queue.
   CloseChannel(ch);
   send_thread = std::thread{[&]() {
     size_t i = data;
-    EXPECT_EQ(ch->Send(&i), false);  // should return false
+    bool is_exception = false;
+    try {
+      ch->Send(&i);
+    } catch (paddle::platform::EnforceNotMet e) {
+      is_exception = true;
+    }
+    EXPECT_EQ(is_exception, true);
   }};
   recv_thread = std::thread{[&]() {
     size_t i;
@@ -129,7 +134,7 @@ TEST(Channel, ReceiveFromBufferedChannelReturnResidualValuesTest) {
   auto ch = MakeChannel<size_t>(buffer_size);
 
   for (size_t i = 0; i < buffer_size; ++i) {
-    EXPECT_EQ(ch->Send(&i), true);  // sending should not block
+    ch->Send(&i);  // sending should not block
   }
 
   size_t out;
@@ -160,9 +165,16 @@ TEST(Channel, ConcurrentSendNonConcurrentReceiveWithSufficientBufferSize) {
     // Try to write more than buffer size.
     for (size_t i = 0; i < 2 * buffer_size; ++i) {
       if (i < buffer_size)
-        EXPECT_EQ(ch->Send(&i), true);  // should block after 10 iterations
-      else
-        EXPECT_EQ(ch->Send(&i), false);
+        ch->Send(&i);  // should block after 10 iterations
+      else {
+        bool is_exception = false;
+        try {
+          ch->Send(&i);
+        } catch (paddle::platform::EnforceNotMet e) {
+          is_exception = true;
+        }
+        EXPECT_EQ(is_exception, true);
+      }
     }
   });
   std::this_thread::sleep_for(std::chrono::milliseconds(200));  // wait 0.2 sec
@@ -231,7 +243,13 @@ void ChannelCloseUnblocksSendersTest(Channel<int> *ch, bool isBuffered) {
     t[i] = std::thread(
         [&](bool *ended, bool *success) {
           int data = 10;
-          *success = ch->Send(&data);
+          bool is_exception = false;
+          try {
+            ch->Send(&data);
+          } catch (paddle::platform::EnforceNotMet e) {
+            is_exception = true;
+          }
+          *success = !is_exception;
           *ended = true;
         },
         &thread_ended[i], &send_success[i]);
@@ -316,8 +334,11 @@ TEST(Channel, UnbufferedLessReceiveMoreSendTest) {
     // Try to send more number of times
     // than receivers
     for (int i = 0; i < 4; i++) {
+      try {
         ch->Send(&i);
         sum_send += i;
+      } catch (paddle::platform::EnforceNotMet e) {
+      }
     }
   });
   for (int i = 0; i < 3; i++) {
@@ -382,7 +403,13 @@ void ChannelDestroyUnblockSenders(Channel<int> *ch, bool isBuffered) {
     t[i] = std::thread(
         [&](bool *ended, bool *success) {
           int data = 10;
-          *success = ch->Send(&data);
+          bool is_exception = false;
+          try {
+            ch->Send(&data);
+          } catch (paddle::platform::EnforceNotMet e) {
+            is_exception = true;
+          }
+          *success = !is_exception;
           *ended = true;
         },
         &thread_ended[i], &send_success[i]);
@@ -508,7 +535,7 @@ void ChannelHolderSendReceive(ChannelHolder *ch) {
   unsigned sum_send = 0;
   std::thread t([&]() {
     for (int i = 0; i < 5; i++) {
-      EXPECT_EQ(ch->Send(&i), true);
+      ch->Send(&i);
       sum_send += i;
     }
   });
@@ -541,8 +568,22 @@ TEST(ChannelHolder, ChannelUninitializedTest) {
   ChannelHolder *ch = new ChannelHolder();
   EXPECT_EQ(ch->IsInitialized(), false);
   int i = 10;
-  EXPECT_EQ(ch->Send(&i), false);
-  EXPECT_EQ(ch->Receive(&i), false);
+  bool send_exception = false;
+  try {
+    ch->Send(&i);
+  } catch (paddle::platform::EnforceNotMet e) {
+    send_exception = true;
+  }
+  EXPECT_EQ(send_exception, true);
+
+  bool recv_exception = false;
+  try {
+    ch->Receive(&i);
+  } catch (paddle::platform::EnforceNotMet e) {
+    recv_exception = true;
+  }
+  EXPECT_EQ(recv_exception, true);
+
   bool is_exception = false;
   try {
     ch->Type();
@@ -669,7 +710,13 @@ void ChannelHolderCloseUnblocksSendersTest(ChannelHolder *ch, bool isBuffered) {
     t[i] = std::thread(
         [&](bool *ended, bool *success) {
           int data = 10;
-          *success = ch->Send(&data);
+          bool is_exception = false;
+          try {
+            ch->Send(&data);
+          } catch (paddle::platform::EnforceNotMet e) {
+            is_exception = true;
+          }
+          *success = !is_exception;
           *ended = true;
         },
         &thread_ended[i], &send_success[i]);
@@ -760,7 +807,13 @@ void ChannelHolderDestroyUnblockSenders(ChannelHolder *ch, bool isBuffered) {
     t[i] = std::thread(
         [&](bool *ended, bool *success) {
           int data = 10;
-          *success = ch->Send(&data);
+          bool is_exception = false;
+          try {
+            ch->Send(&data);
+          } catch (paddle::platform::EnforceNotMet e) {
+            is_exception = true;
+          }
+          *success = !is_exception;
           *ended = true;
         },
         &thread_ended[i], &send_success[i]);

paddle/fluid/framework/concurrency_test.cc

@@ -150,7 +150,8 @@ void AddFibonacciSelect(Scope *scope, p::CPUPlace *place, ProgramDesc *program,
   // Select block
   AddOp("select", {{"X", {dataChanName, quitChanName}},
                    {"case_to_execute", {"caseToExecute"}}},
-        {}, {{"sub_block", casesBlock},
+        {{"Out", {}}},
+        {{"sub_block", casesBlock},
          {"cases", std::vector<std::string>{case0Config, case1Config}}},
         whileBlock);
 
@@ -209,9 +210,8 @@ TEST(Concurrency, Go_Op) {
 
   executor.Run(program, &scope, 0, true, true);
 
-  // After we call executor.run, the Go operator should do a channel_send to set
-  // the
-  // "result" variable to 99
+  // After we call executor.run, the Go operator should do a channel_send to
+  // set the "result" variable to 99.
   auto *finalData = tensor.data<int>();
   EXPECT_EQ(finalData[0], 99);
 }

paddle/fluid/framework/executor.cc

@@ -14,12 +14,8 @@ limitations under the License. */
 
 #include "paddle/fluid/framework/executor.h"
 
-#include <set>
-
-#include "gflags/gflags.h"
 #include "paddle/fluid/framework/channel.h"
 #include "paddle/fluid/framework/feed_fetch_method.h"
-#include "paddle/fluid/framework/feed_fetch_type.h"
 #include "paddle/fluid/framework/lod_rank_table.h"
 #include "paddle/fluid/framework/lod_tensor_array.h"
 #include "paddle/fluid/framework/op_registry.h"
@@ -40,14 +36,13 @@ namespace {
 int kProgramId = -1;
 }  // namespace
 
-struct ExecutorPrepareContext {
-  ExecutorPrepareContext(const framework::ProgramDesc& prog, size_t block_id)
+ExecutorPrepareContext::ExecutorPrepareContext(
+    const framework::ProgramDesc& prog, size_t block_id)
     : prog_(prog), block_id_(block_id) {}
 
-  const framework::ProgramDesc prog_;
-  size_t block_id_;
-  std::vector<std::unique_ptr<OperatorBase>> ops_;
-};
+ExecutorPrepareContext::~ExecutorPrepareContext() {
+  VLOG(5) << "destroy ExecutorPrepareContext";
+}
 
 Executor::Executor(const platform::Place& place) : place_(place) {}
 
@@ -101,9 +96,8 @@ static void CheckTensorNANOrInf(const std::string& name,
 void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
                    bool create_local_scope, bool create_vars) {
   platform::RecordBlock b(block_id);
-  auto* ctx = Prepare(pdesc, block_id);
-  RunPreparedContext(ctx, scope, create_local_scope, create_vars);
-  delete ctx;
+  auto ctx = Prepare(pdesc, block_id);
+  RunPreparedContext(ctx.get(), scope, create_local_scope, create_vars);
 }
 
 // Check whether the block already has feed operators and feed_holder.
@@ -191,7 +185,7 @@ void Executor::Run(const ProgramDesc& program, Scope* scope,
                    std::map<std::string, const LoDTensor*>& feed_targets,
                    std::map<std::string, LoDTensor*>& fetch_targets,
                    const std::string& feed_holder_name,
-                   const std::string& fetch_holder_name) {
+                   const std::string& fetch_holder_name, bool create_vars) {
   platform::RecordBlock b(kProgramId);
   bool has_feed_ops =
       has_feed_operators(program.Block(0), feed_targets, feed_holder_name);
@@ -261,7 +255,7 @@ void Executor::Run(const ProgramDesc& program, Scope* scope,
     }
   }
 
-  Run(*copy_program, scope, 0, true, true);
+  Run(*copy_program, scope, 0, create_vars, create_vars);
 
   // obtain the data of fetch_targets from fetch_holder
   for (auto* op : global_block->AllOps()) {
@@ -274,15 +268,15 @@ void Executor::Run(const ProgramDesc& program, Scope* scope,
   }
 }
 
-ExecutorPrepareContext* Executor::Prepare(const ProgramDesc& program,
-                                          int block_id) {
+std::unique_ptr<ExecutorPrepareContext> Executor::Prepare(
+    const ProgramDesc& program, int block_id) {
   auto* ctx = new ExecutorPrepareContext(program, block_id);
   PADDLE_ENFORCE_LT(static_cast<size_t>(block_id), program.Size());
   auto& block = program.Block(block_id);
   for (auto& op_desc : block.AllOps()) {
     ctx->ops_.push_back(OpRegistry::CreateOp(*op_desc));
   }
-  return ctx;
+  return std::unique_ptr<ExecutorPrepareContext>(ctx);
 }
 
 void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,

paddle/fluid/framework/executor.h

@@ -22,7 +22,16 @@ limitations under the License. */
 
 namespace paddle {
 namespace framework {
-struct ExecutorPrepareContext;
+
+struct ExecutorPrepareContext {
+  ExecutorPrepareContext(const framework::ProgramDesc& prog, size_t block_id);
+  ~ExecutorPrepareContext();
+
+  const framework::ProgramDesc& prog_;
+  size_t block_id_;
+  std::vector<std::unique_ptr<OperatorBase>> ops_;
+};
+
 class Executor {
  public:
   // TODO(dzhwinter) : Do not rely on this function, it will be removed
@@ -45,10 +54,11 @@ class Executor {
            std::map<std::string, const LoDTensor*>& feed_targets,
            std::map<std::string, LoDTensor*>& fetch_targets,
            const std::string& feed_holder_name = "feed",
-           const std::string& fetch_holder_name = "fetch");
+           const std::string& fetch_holder_name = "fetch",
+           bool create_vars = true);
 
-  static ExecutorPrepareContext* Prepare(const ProgramDesc& program,
-                                         int block_id);
+  static std::unique_ptr<ExecutorPrepareContext> Prepare(
+      const ProgramDesc& program, int block_id);
 
   void RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
                           bool create_local_scope = true,

paddle/fluid/framework/mixed_vector.h

@@ -176,7 +176,7 @@ class Vector {
 
   // resize the vector
   void resize(size_t size) {
-    if (size + 1 < capacity()) {
+    if (size + 1 <= capacity()) {
       size_ = size;
     } else {
       MutableCPU();

paddle/fluid/framework/mixed_vector_test.cu

@@ -104,3 +104,11 @@ TEST(mixed_vector, ForEach) {
   for (auto& v : tmp) {
   }
 }
+
+TEST(mixed_vector, Reserve) {
+  paddle::framework::Vector<int> vec;
+  vec.reserve(1);
+  vec.push_back(0);
+  vec.push_back(0);
+  vec.push_back(0);
+}

paddle/fluid/framework/tensor.h

@@ -45,10 +45,11 @@ class Tensor {
   friend struct EigenVector;
 
  public:
-  Tensor() : offset_(0) {}
+  Tensor() : offset_(0), is_pinned_(false) {}
 
   /*! Constructor with place should only be used in pybind. */
-  explicit Tensor(const platform::Place& place) : offset_(0) {
+  explicit Tensor(const platform::Place& place)
+      : offset_(0), is_pinned_(false) {
     holder_->set_place(place);
   }
 
@@ -69,11 +70,12 @@ class Tensor {
    * @note    If not exist, then allocation.
    */
   template <typename T>
-  inline T* mutable_data(platform::Place place);
+  inline T* mutable_data(platform::Place place, bool is_pinned = false);
 
-  inline void* mutable_data(platform::Place place, std::type_index type);
+  inline void* mutable_data(platform::Place place, std::type_index type,
+                            bool is_pinned = false);
 
-  inline void* mutable_data(platform::Place place);
+  inline void* mutable_data(platform::Place place, bool is_pinned = false);
 
   /**
    * @brief     Return a pointer to mutable memory block.
@@ -84,7 +86,8 @@ class Tensor {
    * @note      If not exist, then allocation.
    */
   template <typename T>
-  inline T* mutable_data(DDim dims, platform::Place place);
+  inline T* mutable_data(DDim dims, platform::Place place,
+                         bool is_pinned = false);
 
   /*! Return the dimensions of the memory block. */
   inline const DDim& dims() const;
@@ -92,6 +95,9 @@ class Tensor {
   /*! Return the numel of the memory block. */
   inline int64_t numel() const;
 
+  /*! Return the numel of the memory block. */
+  inline bool isPinned() const;
+
   /*! Resize the dimensions of the memory block. */
   inline Tensor& Resize(const DDim& dims);
 
@@ -146,12 +152,14 @@ class Tensor {
 
   template <typename Place>
   struct PlaceholderImpl : public Placeholder {
-    PlaceholderImpl(Place place, size_t size, std::type_index type)
-        : ptr_(static_cast<uint8_t*>(memory::Alloc(place, size)),
-               memory::PODDeleter<uint8_t, Place>(place)),
+    PlaceholderImpl(Place place, size_t size, std::type_index type,
+                    bool is_pinned = false)
+        : ptr_(static_cast<uint8_t*>(memory::Alloc(place, size, is_pinned)),
+               memory::PODDeleter<uint8_t, Place>(place, is_pinned)),
           place_(place),
           size_(size),
-          type_(type) {
+          type_(type),
+          is_pinned_(is_pinned) {
       PADDLE_ENFORCE_NOT_NULL(ptr_, "Insufficient %s memory to allocation.",
                               (is_cpu_place(place_) ? "CPU" : "GPU"));
     }
@@ -174,6 +182,9 @@ class Tensor {
 
     /* the current type of memory */
     std::type_index type_;
+
+    /*! use pinned memory or not. */
+    bool is_pinned_;
   };
 
   /*! holds the memory block if allocated. */
@@ -208,6 +219,7 @@ class Tensor {
    *          PlaceHolder::ptr_ and where the tensor data really begins.
    */
   size_t offset_;
+  bool is_pinned_;
 };
 
 inline void Tensor::switch_place(platform::Place new_place) {

paddle/fluid/framework/tensor_impl.h

@@ -101,25 +101,27 @@ inline T* Tensor::data() {
 }
 
 template <typename T>
-inline T* Tensor::mutable_data(DDim dims, platform::Place place) {
+inline T* Tensor::mutable_data(DDim dims, platform::Place place,
+                               bool is_pinned) {
   static_assert(std::is_pod<T>::value, "T must be POD");
   Resize(dims);
-  return mutable_data<T>(place);
+  return mutable_data<T>(place, is_pinned);
 }
 
 template <typename T>
-inline T* Tensor::mutable_data(platform::Place place) {
+inline T* Tensor::mutable_data(platform::Place place, bool is_pinned) {
   static_assert(std::is_pod<T>::value, "T must be POD");
-  return reinterpret_cast<T*>(mutable_data(place, typeid(T)));
+  return reinterpret_cast<T*>(mutable_data(place, typeid(T), is_pinned));
 }
 
-inline void* Tensor::mutable_data(platform::Place place, std::type_index type) {
+inline void* Tensor::mutable_data(platform::Place place, std::type_index type,
+                                  bool is_pinned) {
   if (holder_ != nullptr) {
     holder_->set_type(type);
   }
-  PADDLE_ENFORCE_GT(
-      numel(), 0,
-      "When calling this method, the Tensor's numel must be larger than zero. "
+  PADDLE_ENFORCE_GE(numel(), 0,
+                    "When calling this method, the Tensor's numel must be "
+                    "equal or larger than zero. "
                     "Please check Tensor::Resize has been called first.");
   int64_t size = numel() * SizeOfType(type);
   /* some versions of boost::variant don't have operator!= */
@@ -127,26 +129,27 @@ inline void* Tensor::mutable_data(platform::Place place, std::type_index type) {
       holder_->size() < size + offset_) {
     if (platform::is_cpu_place(place)) {
       holder_.reset(new PlaceholderImpl<platform::CPUPlace>(
-          boost::get<platform::CPUPlace>(place), size, type));
+          boost::get<platform::CPUPlace>(place), size, type, is_pinned));
     } else if (platform::is_gpu_place(place)) {
 #ifndef PADDLE_WITH_CUDA
       PADDLE_THROW("'CUDAPlace' is not supported in CPU only device.");
     }
 #else
       holder_.reset(new PlaceholderImpl<platform::CUDAPlace>(
-          boost::get<platform::CUDAPlace>(place), size, type));
+          boost::get<platform::CUDAPlace>(place), size, type, is_pinned));
     }
 #endif
     offset_ = 0;
+    is_pinned_ = is_pinned;
   }
   return reinterpret_cast<void*>(reinterpret_cast<uintptr_t>(holder_->ptr()) +
                                  offset_);
 }
 
-inline void* Tensor::mutable_data(platform::Place place) {
+inline void* Tensor::mutable_data(platform::Place place, bool is_pinned) {
   PADDLE_ENFORCE(this->holder_ != nullptr,
                  "Cannot invoke mutable data if current hold nothing");
-  return mutable_data(place, holder_->type());
+  return mutable_data(place, holder_->type(), is_pinned);
 }
 
 inline Tensor& Tensor::ShareDataWith(const Tensor& src) {
@@ -188,6 +191,8 @@ inline const DDim& Tensor::dims() const { return dims_; }
 
 inline int64_t Tensor::numel() const { return product(dims_); }
 
+inline bool Tensor::isPinned() const { return is_pinned_; }
+
 inline Tensor ReshapeToMatrix(const Tensor& src, int num_col_dims) {
   Tensor res;
   res.ShareDataWith(src);

paddle/fluid/memory/detail/system_allocator.cc

@@ -119,6 +119,50 @@ void GPUAllocator::Free(void* p, size_t size, size_t index) {
 
 bool GPUAllocator::UseGpu() const { return true; }
 
+// PINNED memory allows direct DMA transfers by the GPU to and from system
+// memory. It’s locked to a physical address.
+void* CUDAPinnedAllocator::Alloc(size_t& index, size_t size) {
+  if (size <= 0) return nullptr;
+  void* p;
+  // NOTE: here, we use GpuMaxAllocSize() as the maximum memory size
+  // of host pinned allocation. Allocates too much would reduce
+  // the amount of memory available to the underlying system for paging.
+
+  size_t usable = paddle::platform::GpuMaxAllocSize() - fallback_alloc_size_;
+
+  if (size > usable) return nullptr;
+
+  // PINNED memory is visible to all CUDA contexts.
+  cudaError_t result = cudaMallocHost(&p, size);
+  if (result == cudaSuccess) {
+    index = 1;
+    fallback_alloc_size_ += size;
+    return p;
+  }
+
+  return nullptr;
+}
+
+void CUDAPinnedAllocator::Free(void* p, size_t size, size_t index) {
+  cudaError_t err;
+  PADDLE_ASSERT(index == 1);
+
+  PADDLE_ASSERT(fallback_alloc_size_ >= size);
+  fallback_alloc_size_ -= size;
+  err = cudaFreeHost(p);
+
+  // Purposefully allow cudaErrorCudartUnloading, because
+  // that is returned if you ever call cudaFreeHost after the
+  // driver has already shutdown. This happens only if the
+  // process is terminating, in which case we don't care if
+  // cudaFreeHost succeeds.
+  if (err != cudaErrorCudartUnloading) {
+    PADDLE_ENFORCE(err, "cudaFreeHost failed in GPUPinnedAllocator::Free.");
+  }
+}
+
+bool CUDAPinnedAllocator::UseGpu() const { return true; }
+
 #endif
 
 }  // namespace detail

paddle/fluid/memory/detail/system_allocator.h

@@ -51,6 +51,18 @@ class GPUAllocator : public SystemAllocator {
   size_t gpu_alloc_size_ = 0;
   size_t fallback_alloc_size_ = 0;
 };
+
+class CUDAPinnedAllocator : public SystemAllocator {
+ public:
+  virtual void* Alloc(size_t& index, size_t size);
+  virtual void Free(void* p, size_t size, size_t index);
+  virtual bool UseGpu() const;
+
+ private:
+  size_t gpu_alloc_size_ =
+      0;  // TODO(zcd): how to define the upper limit of CUDAPinnedMemory?
+  size_t fallback_alloc_size_ = 0;
+};
 #endif
 
 }  // namespace detail

paddle/fluid/memory/memory.cc

@@ -38,7 +38,8 @@ BuddyAllocator* GetCPUBuddyAllocator() {
 }
 
 template <>
-void* Alloc<platform::CPUPlace>(platform::CPUPlace place, size_t size) {
+void* Alloc<platform::CPUPlace>(platform::CPUPlace place, size_t size,
+                                bool is_pinned) {
   VLOG(10) << "Allocate " << size << " bytes on " << platform::Place(place);
   void* p = GetCPUBuddyAllocator()->Alloc(size);
   VLOG(10) << "  pointer=" << p;
@@ -46,7 +47,8 @@ void* Alloc<platform::CPUPlace>(platform::CPUPlace place, size_t size) {
 }
 
 template <>
-void Free<platform::CPUPlace>(platform::CPUPlace place, void* p) {
+void Free<platform::CPUPlace>(platform::CPUPlace place, void* p,
+                              bool is_pinned) {
   VLOG(10) << "Free pointer=" << p << " on " << platform::Place(place);
   GetCPUBuddyAllocator()->Free(p);
 }
@@ -82,15 +84,47 @@ BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) {
   return as[gpu_id];
 }
 
+BuddyAllocator* GetCUDAPinnedBuddyAllocator(int gpu_id) {
+  static BuddyAllocator** as = NULL;
+  if (as == NULL) {
+    int gpu_num = platform::GetCUDADeviceCount();
+    as = new BuddyAllocator*[gpu_num];
+    for (int gpu = 0; gpu < gpu_num; gpu++) {
+      as[gpu] = nullptr;
+    }
+  }
+  platform::SetDeviceId(gpu_id);
+  if (!as[gpu_id]) {
+    as[gpu_id] = new BuddyAllocator(new detail::CUDAPinnedAllocator,
+                                    platform::GpuMinChunkSize(),
+                                    platform::GpuMaxChunkSize());
+    VLOG(10) << "\n\nNOTE: each GPU device use "
+             << FLAGS_fraction_of_gpu_memory_to_use * 100
+             << "% of GPU memory.\n"
+             << "You can set GFlags environment variable '"
+             << "FLAGS_fraction_of_gpu_memory_to_use"
+             << "' to change the fraction of GPU usage.\n\n";
+  }
+  return as[gpu_id];
+}
+
 template <>
 size_t Used<platform::CUDAPlace>(platform::CUDAPlace place) {
   return GetGPUBuddyAllocator(place.device)->Used();
 }
 
 template <>
-void* Alloc<platform::CUDAPlace>(platform::CUDAPlace place, size_t size) {
+void* Alloc<platform::CUDAPlace>(platform::CUDAPlace place, size_t size,
+                                 bool is_pinned) {
+  void* ptr;
+  if (is_pinned) {
+    auto* buddy_allocator = GetCUDAPinnedBuddyAllocator(place.device);
+    ptr = buddy_allocator->Alloc(size);
+  } else {
     auto* buddy_allocator = GetGPUBuddyAllocator(place.device);
-  auto* ptr = buddy_allocator->Alloc(size);
+    ptr = buddy_allocator->Alloc(size);
+  }
+
   if (ptr == nullptr) {
     int cur_dev = platform::GetCurrentDeviceId();
     platform::SetDeviceId(place.device);
@@ -108,8 +142,13 @@ void* Alloc<platform::CUDAPlace>(platform::CUDAPlace place, size_t size) {
 }
 
 template <>
-void Free<platform::CUDAPlace>(platform::CUDAPlace place, void* p) {
+void Free<platform::CUDAPlace>(platform::CUDAPlace place, void* p,
+                               bool is_pinned) {
+  if (is_pinned) {
+    GetCUDAPinnedBuddyAllocator(place.device)->Free(p);
+  } else {
     GetGPUBuddyAllocator(place.device)->Free(p);
+  }
 }
 
 #endif

paddle/fluid/memory/memory.h

@@ -33,7 +33,7 @@ namespace memory {
  *          address is valid or not.
  */
 template <typename Place>
-void* Alloc(Place place, size_t size);
+void* Alloc(Place place, size_t size, bool is_pinned = false);
 
 /**
  * \brief   Free memory block in one place.
@@ -43,7 +43,7 @@ void* Alloc(Place place, size_t size);
  *
  */
 template <typename Place>
-void Free(Place place, void* ptr);
+void Free(Place place, void* ptr, bool is_pinned = false);
 
 /**
  * \brief   Total size of used memory in one place.
@@ -74,11 +74,13 @@ class PODDeleter {
   static_assert(std::is_pod<T>::value, "T must be POD");
 
  public:
-  explicit PODDeleter(Place place) : place_(place) {}
-  void operator()(T* ptr) { Free(place_, static_cast<void*>(ptr)); }
+  explicit PODDeleter(Place place, bool is_pinned = false)
+      : place_(place), is_pinned_(is_pinned) {}
+  void operator()(T* ptr) { Free(place_, static_cast<void*>(ptr), is_pinned_); }
 
  private:
   Place place_;
+  bool is_pinned_;
 };
 
 /**

paddle/fluid/memory/memory_test.cc

@@ -59,7 +59,7 @@ TEST(BuddyAllocator, CPUMultAlloc) {
   EXPECT_EQ(total_size, 0UL);
 
   for (auto size :
-       {128, 256, 1024, 4096, 16384, 65536, 262144, 1048576, 4194304}) {
+       {0, 128, 256, 1024, 4096, 16384, 65536, 262144, 1048576, 4194304}) {
     ps[paddle::memory::Alloc(cpu, size)] = size;
 
     // Buddy Allocator doesn't manage too large memory chunk
@@ -117,7 +117,7 @@ TEST(BuddyAllocator, GPUMultAlloc) {
   EXPECT_EQ(total_size, 0UL);
 
   for (auto size :
-       {128, 256, 1024, 4096, 16384, 65536, 262144, 1048576, 4194304}) {
+       {0, 128, 256, 1024, 4096, 16384, 65536, 262144, 1048576, 4194304}) {
     ps[paddle::memory::Alloc(gpu, size)] = size;
 
     // Buddy Allocator doesn't manage too large memory chunk

paddle/fluid/operators/CMakeLists.txt

@@ -12,6 +12,8 @@ function(op_library TARGET)
     set(OP_LIBRARY ${TARGET} ${OP_LIBRARY} PARENT_SCOPE)
     set(cc_srcs)
     set(cu_srcs)
+    set(hip_cu_srcs)
+    set(miopen_hip_cc_srcs)
     set(cu_cc_srcs)
     set(cudnn_cu_cc_srcs)
     set(CUDNN_FILE)
@@ -36,10 +38,19 @@ function(op_library TARGET)
         if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cu)
             list(APPEND cu_srcs ${TARGET}.cu)
         endif()
+        if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.hip.cu)
+            list(APPEND hip_cu_srcs ${TARGET}.hip.cu)
+        endif()
         string(REPLACE "_op" "_cudnn_op" CUDNN_FILE "${TARGET}")
         if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${CUDNN_FILE}.cu.cc)
             list(APPEND cudnn_cu_cc_srcs ${CUDNN_FILE}.cu.cc)
         endif()
+        if(WITH_AMD_GPU)
+            string(REPLACE "_op" "_miopen_op" MIOPEN_FILE "${TARGET}")
+            if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${MIOPEN_FILE}.hip.cc)
+                list(APPEND miopen_hip_cc_srcs ${MIOPEN_FILE}.hip.cc)
+            endif()
+        endif()
         if(WITH_MKLDNN)
             string(REPLACE "_op" "_mkldnn_op" MKLDNN_FILE "${TARGET}")
             if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${MKLDNN_FILE}.cc)
@@ -48,10 +59,14 @@ function(op_library TARGET)
         endif()
     else()
         foreach(src ${op_library_SRCS})
-            if (${src} MATCHES ".*\\.cu$")
+            if (${src} MATCHES ".*\\.hip.cu$")
+                list(APPEND hip_cu_srcs ${src})
+            elseif (${src} MATCHES ".*\\.cu$")
                 list(APPEND cu_srcs ${src})
             elseif(${src} MATCHES ".*_cudnn_op.cu.cc$")
                 list(APPEND cudnn_cu_cc_srcs ${src})
+            elseif(WITH_AMD_GPU AND ${src} MATCHES ".*_miopen_op.hip.cc$")
+                list(APPEND miopen_hip_cc_srcs ${src})
             elseif(WITH_MKLDNN AND ${src} MATCHES ".*_mkldnn_op.cc$")
                 list(APPEND mkldnn_cc_srcs ${src})
             elseif(${src} MATCHES ".*\\.cu.cc$")
@@ -76,6 +91,9 @@ function(op_library TARGET)
     if (WITH_GPU)
         nv_library(${TARGET} SRCS ${cc_srcs} ${cu_cc_srcs} ${cudnn_cu_cc_srcs} ${mkldnn_cc_srcs} ${cu_srcs} DEPS ${op_library_DEPS}
                 ${op_common_deps})
+    elseif (WITH_AMD_GPU)
+        hip_library(${TARGET} SRCS ${cc_srcs} ${hip_cu_srcs} ${miopen_hip_cc_srcs} ${mkldnn_cc_srcs} DEPS ${op_library_DEPS}
+                ${op_common_deps})
     else()
         cc_library(${TARGET} SRCS ${cc_srcs} ${mkldnn_cc_srcs} DEPS ${op_library_DEPS}
             ${op_common_deps})
@@ -88,7 +106,7 @@ function(op_library TARGET)
         endif()
     endforeach()
 
-    # The registration of USE_OP, please refer to paddle/framework/op_registry.h.
+    # The registration of USE_OP, please refer to paddle/fluid/framework/op_registry.h.
     # Note that it's enough to just adding one operator to pybind in a *_op.cc file.
     # And for detail pybind information, please see generated paddle/pybind/pybind.h.
     file(READ ${TARGET}.cc TARGET_CONTENT)
@@ -114,7 +132,10 @@ function(op_library TARGET)
     list(LENGTH cu_srcs cu_srcs_len)
     list(LENGTH cu_cc_srcs cu_cc_srcs_len)
     list(LENGTH mkldnn_cc_srcs mkldnn_cc_srcs_len)
-    if (${pybind_flag} EQUAL 0 AND ${mkldnn_cc_srcs_len} EQUAL 0 AND ${cu_srcs_len} EQUAL 0 AND ${cu_cc_srcs_len} EQUAL 0)
+    list(LENGTH hip_cu_srcs hip_cu_srcs_len)
+    list(LENGTH miopen_hip_cc_srcs miopen_hip_cc_srcs_len)
+    if (${pybind_flag} EQUAL 0 AND ${mkldnn_cc_srcs_len} EQUAL 0 AND ${cu_srcs_len} EQUAL 0 AND ${cu_cc_srcs_len} EQUAL 0 AND
+        ${hip_cu_srcs_len} EQUAL 0 AND ${miopen_hip_cc_srcs_len} EQUAL 0)
         file(APPEND ${pybind_file} "USE_CPU_ONLY_OP(${TARGET});\n")
         set(pybind_flag 1)
     endif()
@@ -125,10 +146,20 @@ function(op_library TARGET)
         file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, CUDNN);\n")
     endif()
 
+    # pybind USE_OP_DEVICE_KERNEL for MIOPEN
+    if (WITH_AMD_GPU AND ${miopen_hip_cc_srcs_len} GREATER 0)
+        file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, MIOPEN);\n")
+    endif()
+
     # pybind USE_OP_DEVICE_KERNEL for MKLDNN
     if (WITH_MKLDNN AND ${mkldnn_cc_srcs_len} GREATER 0)
+      # Append first implemented MKLDNN activation operator
+      if (${MKLDNN_FILE} STREQUAL "activation_mkldnn_op")
+        file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(relu, MKLDNN);\n")
+      else()
         file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, MKLDNN);\n")
       endif()
+    endif()
 
     # pybind USE_OP
     if (${pybind_flag} EQUAL 0)
@@ -156,9 +187,13 @@ if(WITH_DISTRIBUTE)
     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_barrier_op DEPS ${DISTRIBUTE_DEPS})
+    set_source_files_properties(send_barrier_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
     cc_test(test_send_recv SRCS send_recv_op_test.cc DEPS send_op listen_and_serv_op sum_op executor)
 else()
-    set(DEPS_OPS ${DEPS_OPS} send_op recv_op listen_and_serv_op)
+    set(DEPS_OPS ${DEPS_OPS} send_op recv_op listen_and_serv_op send_vars_op send_barrier_op)
 endif()
 
 op_library(cond_op DEPS framework_proto tensor net_op)
@@ -229,3 +264,4 @@ cc_test(strided_memcpy_test SRCS strided_memcpy_test.cc DEPS tensor paddle_memor
 cc_test(save_load_op_test SRCS save_load_op_test.cc DEPS save_op load_op)
 cc_test(save_load_combine_op_test SRCS save_load_combine_op_test.cc DEPS save_combine_op load_combine_op)
 nv_test(nccl_op_test SRCS nccl_op_test.cu.cc DEPS nccl_op gpu_info device_context)
+nv_test(dropout_op_test SRCS dropout_op_test.cc DEPS dropout_op tensor)

paddle/fluid/operators/activation_mkldnn_op.cc

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
+
+   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 "mkldnn.hpp"
+#include "mkldnn_activation_op.h"
+#include "paddle/fluid/operators/activation_op.h"
+
+namespace paddle {
+namespace operators {
+
+using paddle::framework::Tensor;
+using paddle::platform::MKLDNNDeviceContext;
+
+namespace {
+template <typename T, typename ExecContext>
+void eltwise_forward(const ExecContext &ctx, mkldnn::algorithm algorithm,
+                     const T alpha = 0, const T beta = 0) {
+  PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()),
+                 "It must use CPUPlace.");
+
+  auto &dev_ctx = ctx.template device_context<MKLDNNDeviceContext>();
+  const auto &mkldnn_engine = dev_ctx.GetEngine();
+
+  // get buffers
+  const auto *src = ctx.template Input<Tensor>("X");
+  const auto *src_data = src->template data<T>();
+
+  auto *dst = ctx.template Output<Tensor>("Out");
+  const T *dst_data = dst->template mutable_data<T>(ctx.GetPlace());
+
+  // get memory dim
+  PADDLE_ENFORCE(src->dims().size() == 4,
+                 "Input dim must be with 4, i.e. NCHW");
+  std::vector<int> src_tz = framework::vectorize2int(src->dims());
+
+  // create memory description
+  // TODO(kbinias-intel): support more formats
+  auto data_md = platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32,
+                                         mkldnn::memory::format::nchw);
+
+  // create memory primitives
+  auto src_memory = mkldnn::memory({data_md, mkldnn_engine}, (void *)src_data);
+  auto dst_memory = mkldnn::memory({data_md, mkldnn_engine}, (void *)dst_data);
+
+  auto forward_desc = mkldnn::eltwise_forward::desc(
+      mkldnn::prop_kind::forward_training, algorithm, data_md, alpha, beta);
+
+  // save prim desc into global device context to be referred in backward path
+  const std::string key = ctx.op().Output("Out");
+  const std::string key_eltwise_pd = key + "@eltwise_pd";
+  auto forward_pd = std::make_shared<mkldnn::eltwise_forward::primitive_desc>(
+      forward_desc, mkldnn_engine);
+  dev_ctx.SetBlob(key_eltwise_pd, forward_pd);
+
+  auto eltwise = mkldnn::eltwise_forward(*forward_pd, src_memory, dst_memory);
+
+  // push primitive to stream and wait until it's executed
+  std::vector<mkldnn::primitive> pipeline = {eltwise};
+  mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
+}
+
+template <typename T, typename ExecContext>
+void eltwise_grad(const ExecContext &ctx, mkldnn::algorithm algorithm,
+                  const T alpha = 0, const T beta = 0) {
+  auto &dev_ctx = ctx.template device_context<MKLDNNDeviceContext>();
+  const auto &mkldnn_engine = dev_ctx.GetEngine();
+
+  // get buffers
+  const auto *x = ctx.template Input<Tensor>("X");
+  const auto *src = x->template data<T>();
+
+  auto *dout = ctx.template Input<Tensor>(framework::GradVarName("Out"));
+  const auto *diff_dst = dout->template data<T>();
+
+  auto *dx =
+      ctx.template Output<framework::Tensor>(framework::GradVarName("X"));
+  const T *diff_src = dx->template mutable_data<T>(ctx.GetPlace());
+
+  // get memory dim
+  std::vector<int> src_tz = framework::vectorize2int(x->dims());
+
+  // create memory description
+  auto data_md = platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32,
+                                         mkldnn::memory::format::nchw);
+
+  // create memory primitives
+  auto src_memory = mkldnn::memory({data_md, mkldnn_engine}, (void *)src);
+  auto diff_src_memory =
+      mkldnn::memory({data_md, mkldnn_engine}, (void *)diff_src);
+  auto diff_dst_memory =
+      mkldnn::memory({data_md, mkldnn_engine}, (void *)diff_dst);
+
+  auto backward_desc =
+      mkldnn::eltwise_backward::desc(algorithm, data_md, data_md, alpha, beta);
+
+  // retrieve eltwise primitive desc from device context
+  const std::string key = ctx.op().Input("Out");
+  const std::string key_eltwise_pd = key + "@eltwise_pd";
+  const std::shared_ptr<void> forward_pd = dev_ctx.GetBlob(key_eltwise_pd);
+  PADDLE_ENFORCE(forward_pd != nullptr,
+                 "Fail to find eltwise_pd in device context");
+  auto *p_forward_pd =
+      static_cast<mkldnn::eltwise_forward::primitive_desc *>(forward_pd.get());
+
+  auto eltwise_bwd_prim_desc = mkldnn::eltwise_backward::primitive_desc(
+      backward_desc, mkldnn_engine, *p_forward_pd);
+
+  auto eltwise_bwd = mkldnn::eltwise_backward(eltwise_bwd_prim_desc, src_memory,
+                                              diff_dst_memory, diff_src_memory);
+
+  // push primitive to stream and wait until it's executed
+  std::vector<mkldnn::primitive> pipeline = {eltwise_bwd};
+  mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
+}
+}  // anonymous namespace
+
+template <typename T, mkldnn::algorithm algorithm>
+struct MKLDNNActivationFunc : public BaseActivationFunctor<T> {
+  template <typename ExecContext>
+  void operator()(const ExecContext &ctx) const {
+    eltwise_forward<T>(ctx, algorithm);
+  }
+};
+
+template <typename T, mkldnn::algorithm algorithm>
+struct MKLDNNActivationGradFunc : public BaseActivationFunctor<T> {
+  template <typename ExecContext>
+  void operator()(const ExecContext &ctx) const {
+    eltwise_grad<T>(ctx, algorithm);
+  }
+};
+
+template <typename T>
+using ReluMkldnnFunctor =
+    MKLDNNActivationFunc<T, mkldnn::algorithm::eltwise_relu>;
+
+template <typename T>
+using TanhMkldnnFunctor =
+    MKLDNNActivationFunc<T, mkldnn::algorithm::eltwise_tanh>;
+
+template <typename T>
+using SqrtMkldnnFunctor =
+    MKLDNNActivationFunc<T, mkldnn::algorithm::eltwise_sqrt>;
+
+template <typename T>
+using AbsMkldnnFunctor =
+    MKLDNNActivationFunc<T, mkldnn::algorithm::eltwise_abs>;
+
+template <typename T>
+using ReluMkldnnGradFunctor =
+    MKLDNNActivationGradFunc<T, mkldnn::algorithm::eltwise_relu>;
+
+template <typename T>
+using TanhMkldnnGradFunctor =
+    MKLDNNActivationGradFunc<T, mkldnn::algorithm::eltwise_tanh>;
+
+template <typename T>
+using SqrtMkldnnGradFunctor =
+    MKLDNNActivationGradFunc<T, mkldnn::algorithm::eltwise_sqrt>;
+
+template <typename T>
+using AbsMkldnnGradFunctor =
+    MKLDNNActivationGradFunc<T, mkldnn::algorithm::eltwise_abs>;
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+#define REGISTER_ACTIVATION_MKLDNN_KERNEL(act_type, functor, grad_functor) \
+  REGISTER_OP_KERNEL(act_type, MKLDNN, ::paddle::platform::CPUPlace,       \
+                     ops::MKLDNNActivationKernel<ops::functor<float>>);    \
+  REGISTER_OP_KERNEL(                                                      \
+      act_type##_grad, MKLDNN, ::paddle::platform::CPUPlace,               \
+      ops::MKLDNNActivationGradKernel<ops::grad_functor<float>>);
+
+#define FOR_EACH_MKLDNN_KERNEL_FUNCTOR(__macro)            \
+  __macro(relu, ReluMkldnnFunctor, ReluMkldnnGradFunctor); \
+  __macro(tanh, TanhMkldnnFunctor, TanhMkldnnGradFunctor); \
+  __macro(sqrt, SqrtMkldnnFunctor, SqrtMkldnnGradFunctor); \
+  __macro(abs, AbsMkldnnFunctor, AbsMkldnnGradFunctor);
+
+FOR_EACH_MKLDNN_KERNEL_FUNCTOR(REGISTER_ACTIVATION_MKLDNN_KERNEL);

paddle/fluid/operators/activation_op.cc

-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+/* 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.
@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/activation_op.h"
+#include "paddle/fluid/operators/mkldnn_activation_op.h"
 
 namespace paddle {
 namespace operators {
@@ -87,6 +88,9 @@ class ReluOpMaker : public framework::OpProtoAndCheckerMaker {
       : framework::OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "Input of Relu operator");
     AddOutput("Out", "Output of Relu operator");
+    AddAttr<bool>("use_mkldnn",
+                  "(bool, default false) Only used in mkldnn kernel")
+        .SetDefault(false);
     AddComment(R"DOC(
 Relu Activation Operator.
 
@@ -140,6 +144,9 @@ class TanhOpMaker : public framework::OpProtoAndCheckerMaker {
       : framework::OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "Input of Tanh operator");
     AddOutput("Out", "Output of Tanh operator");
+    AddAttr<bool>("use_mkldnn",
+                  "(bool, default false) Only used in mkldnn kernel")
+        .SetDefault(false);
     AddComment(R"DOC(
 Tanh Activation Operator.
 
@@ -193,6 +200,9 @@ class SqrtOpMaker : public framework::OpProtoAndCheckerMaker {
       : framework::OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "Input of Sqrt operator");
     AddOutput("Out", "Output of Sqrt operator");
+    AddAttr<bool>("use_mkldnn",
+                  "(bool, default false) Only used in mkldnn kernel")
+        .SetDefault(false);
     AddComment(R"DOC(
 Sqrt Activation Operator.
 
@@ -208,6 +218,9 @@ class AbsOpMaker : public framework::OpProtoAndCheckerMaker {
       : framework::OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "Input of Abs operator");
     AddOutput("Out", "Output of Abs operator");
+    AddAttr<bool>("use_mkldnn",
+                  "(bool, default false) Only used in mkldnn kernel")
+        .SetDefault(false);
     AddComment(R"DOC(
 Abs Activation Operator.
 
@@ -524,11 +537,11 @@ REGISTER_OP(logsigmoid, ops::ActivationOp, ops::LogSigmoidOpMaker,
 REGISTER_OP(exp, ops::ActivationOp, ops::ExpOpMaker, exp_grad,
             ops::ActivationOpGrad);
 
-REGISTER_OP(relu, ops::ActivationOp, ops::ReluOpMaker, relu_grad,
-            ops::ActivationOpGrad);
+REGISTER_OP(relu, ops::ActivationWithMKLDNNOp, ops::ReluOpMaker, relu_grad,
+            ops::ActivationWithMKLDNNOpGrad);
 
-REGISTER_OP(tanh, ops::ActivationOp, ops::TanhOpMaker, tanh_grad,
-            ops::ActivationOpGrad);
+REGISTER_OP(tanh, ops::ActivationWithMKLDNNOp, ops::TanhOpMaker, tanh_grad,
+            ops::ActivationWithMKLDNNOpGrad);
 
 REGISTER_OP(tanh_shrink, ops::ActivationOp, ops::TanhShrinkOpMaker,
             tanh_shrink_grad, ops::ActivationOpGrad);
@@ -536,11 +549,11 @@ REGISTER_OP(tanh_shrink, ops::ActivationOp, ops::TanhShrinkOpMaker,
 REGISTER_OP(softshrink, ops::ActivationOp, ops::SoftShrinkOpMaker,
             softshrink_grad, ops::ActivationOpGrad);
 
-REGISTER_OP(sqrt, ops::ActivationOp, ops::SqrtOpMaker, sqrt_grad,
-            ops::ActivationOpGrad);
+REGISTER_OP(sqrt, ops::ActivationWithMKLDNNOp, ops::SqrtOpMaker, sqrt_grad,
+            ops::ActivationWithMKLDNNOpGrad);
 
-REGISTER_OP(abs, ops::ActivationOp, ops::AbsOpMaker, abs_grad,
-            ops::ActivationOpGrad);
+REGISTER_OP(abs, ops::ActivationWithMKLDNNOp, ops::AbsOpMaker, abs_grad,
+            ops::ActivationWithMKLDNNOpGrad);
 
 REGISTER_OP(ceil, ops::ActivationOp, ops::CeilOpMaker, ceil_grad,
             ops::ActivationOpGrad);
@@ -613,3 +626,14 @@ REGISTER_OP(swish, ops::ActivationOp, ops::SwishOpMaker, swish_grad,
                                 ops::grad_functor<double>>);
 
 FOR_EACH_KERNEL_FUNCTOR(REGISTER_ACTIVATION_CPU_KERNEL);
+
+REGISTER_OP_CPU_KERNEL(relu,
+                       ops::ActivationKernel<paddle::platform::CPUDeviceContext,
+                                             ops::ReluFunctor<float>>,
+                       ops::ActivationKernel<paddle::platform::CPUDeviceContext,
+                                             ops::ReluFunctor<double>>);
+REGISTER_OP_CPU_KERNEL(
+    relu_grad, ops::ActivationGradKernel<paddle::platform::CPUDeviceContext,
+                                         ops::ReluGradFunctor<float>>,
+    ops::ActivationGradKernel<paddle::platform::CPUDeviceContext,
+                              ops::ReluGradFunctor<double>>);

paddle/fluid/operators/activation_op.cu

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #define EIGEN_USE_GPU
 #include "paddle/fluid/operators/activation_op.h"
+#include "paddle/fluid/platform/float16.h"
 
 namespace ops = paddle::operators;
 
@@ -31,3 +32,16 @@ namespace ops = paddle::operators;
                                 ops::grad_functor<double>>);
 
 FOR_EACH_KERNEL_FUNCTOR(REGISTER_ACTIVATION_CUDA_KERNEL);
+
+REGISTER_OP_CUDA_KERNEL(
+    relu, ops::ActivationKernel<paddle::platform::CUDADeviceContext,
+                                ops::ReluFunctor<float>>,
+    ops::ActivationKernel<paddle::platform::CUDADeviceContext,
+                          ops::ReluFunctor<double>>,
+    ops::ActivationKernel<paddle::platform::CUDADeviceContext,
+                          ops::ReluFunctor<paddle::platform::float16>>);
+REGISTER_OP_CUDA_KERNEL(
+    relu_grad, ops::ActivationGradKernel<paddle::platform::CUDADeviceContext,
+                                         ops::ReluGradFunctor<float>>,
+    ops::ActivationGradKernel<paddle::platform::CUDADeviceContext,
+                              ops::ReluGradFunctor<double>>);

paddle/fluid/operators/activation_op.h

-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+/* 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.
@@ -17,6 +17,10 @@ limitations under the License. */
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/detail/safe_ref.h"
 
+#ifdef PADDLE_WITH_MKLDNN
+#include "paddle/fluid/platform/mkldnn_helper.h"
+#endif
+
 namespace paddle {
 namespace operators {
 
@@ -772,7 +776,6 @@ struct SwishGradFunctor : public BaseActivationFunctor<T> {
   __macro(sigmoid, SigmoidFunctor, SigmoidGradFunctor);              \
   __macro(logsigmoid, LogSigmoidFunctor, LogSigmoidGradFunctor);     \
   __macro(exp, ExpFunctor, ExpGradFunctor);                          \
-  __macro(relu, ReluFunctor, ReluGradFunctor);                       \
   __macro(tanh, TanhFunctor, TanhGradFunctor);                       \
   __macro(softshrink, SoftShrinkFunctor, SoftShrinkGradFunctor);     \
   __macro(sqrt, SqrtFunctor, SqrtGradFunctor);                       \

paddle/fluid/operators/average_accumulates_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/average_accumulates_op.h"
+
+namespace paddle {
+namespace operators {
+
+template <>
+void GetAccumulators<paddle::platform::CPUDeviceContext>(
+    const framework::ExecutionContext& ctx, int64_t& num_updates_,
+    int64_t& num_accumulates_, int64_t& old_num_accumulates_) {
+  auto* in_old_num_accumulates = ctx.Input<Tensor>("in_old_num_accumulates");
+  auto* in_num_accumulates = ctx.Input<Tensor>("in_num_accumulates");
+  auto* in_num_updates = ctx.Input<Tensor>("in_num_updates");
+
+  old_num_accumulates_ = in_old_num_accumulates->data<int64_t>()[0];
+  num_accumulates_ = in_num_accumulates->data<int64_t>()[0];
+  num_updates_ = in_num_updates->data<int64_t>()[0];
+}
+
+template <>
+void SetAccumulators<paddle::platform::CPUDeviceContext>(
+    const framework::ExecutionContext& ctx, int64_t num_updates_,
+    int64_t num_accumulates_, int64_t old_num_accumulates_) {
+  auto* out_old_num_accumulates = ctx.Output<Tensor>("out_old_num_accumulates");
+  auto* out_num_accumulates = ctx.Output<Tensor>("out_num_accumulates");
+  auto* out_num_updates = ctx.Output<Tensor>("out_num_updates");
+
+  out_old_num_accumulates->data<int64_t>()[0] = old_num_accumulates_;
+  out_num_accumulates->data<int64_t>()[0] = num_accumulates_;
+  out_num_updates->data<int64_t>()[0] = num_updates_;
+}
+
+class AverageAccumulatesOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(
+        ctx->HasInput("param"),
+        "Input (param) of average_accumulates op should not be null.");
+    PADDLE_ENFORCE(
+        ctx->HasInput("in_sum_1"),
+        "Input (sum_1) of average_accumulates op should not be null.");
+    PADDLE_ENFORCE(
+        ctx->HasInput("in_sum_2"),
+        "Input (sum_2) of average_accumulates op should not be null.");
+    PADDLE_ENFORCE(
+        ctx->HasInput("in_sum_3"),
+        "Input (sum_3) of average_accumulates op should not be null.");
+    PADDLE_ENFORCE(
+        ctx->HasInput("in_num_accumulates"),
+        "Input (in_num_accumulates) of average_accumulates op should "
+        "not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("in_old_num_accumulates"),
+                   "Input (old_num_accumulates) of average_accumulates op "
+                   "should not be null.");
+    PADDLE_ENFORCE(
+        ctx->HasInput("in_num_updates"),
+        "Input (num_updates) of average_accumulates op should not be null.");
+
+    PADDLE_ENFORCE(
+        ctx->HasOutput("out_sum_1"),
+        "Output (sum_1) of average_accumulates op should not be null.");
+    PADDLE_ENFORCE(
+        ctx->HasOutput("out_sum_2"),
+        "Output (sum_2) of average_accumulates op should not be null.");
+    PADDLE_ENFORCE(
+        ctx->HasOutput("out_sum_3"),
+        "Output (sum_3) of average_accumulates op should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("out_num_accumulates"),
+                   "Output (num_accumulates) of average_accumulates op should "
+                   "not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("out_old_num_accumulates"),
+                   "Output (old_num_accumulates) of average_accumulates op "
+                   "should not be null.");
+    PADDLE_ENFORCE(
+        ctx->HasOutput("out_num_updates"),
+        "Output (num_updates) of average_accumulates op should not be null.");
+
+    auto in_dim = ctx->GetInputDim("param");
+
+    ctx->SetOutputDim("out_sum_1", in_dim);
+    ctx->SetOutputDim("out_sum_2", in_dim);
+    ctx->SetOutputDim("out_sum_3", in_dim);
+    ctx->SetOutputDim("out_num_accumulates", {1});
+    ctx->SetOutputDim("out_old_num_accumulates", {1});
+    ctx->SetOutputDim("out_num_updates", {1});
+  }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(
+        framework::ToDataType(ctx.Input<Tensor>("param")->type()),
+        ctx.GetPlace());
+  }
+};
+
+class AverageAccumulatesOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  AverageAccumulatesOpMaker(OpProto* proto, OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("param", "(Tensor), The parameter to be accumulated.");
+    AddInput("in_sum_1",
+             "(Tensor), A tensor used to store the parameter "
+             "sums with the same shape as input(param).");
+    AddInput("in_sum_2",
+             "(Tensor), A auxiliary tensor to help "
+             "accumulating sums of parameter values with the same shape as "
+             "input(param). It is used to avoid loss of precision due to too "
+             "many sums.");
+    AddInput("in_sum_3",
+             "(Tensor), A auxiliary tensor to help "
+             "accumulating sums of parameter values with the same shape as "
+             "input(param).");
+    AddInput("in_num_accumulates",
+             "(Tensor<int64_t>), The accumulating times of current window with "
+             "shape [1].");
+    AddInput(
+        "in_old_num_accumulates",
+        "(Tensor<int64_t>), The accumulating times of previous window with "
+        "shape [1].");
+    AddInput("in_num_updates",
+             "(Tensor<int64_t>), The total number of batches used by trainning "
+             "before this batch with shape [1].");
+
+    AddOutput("out_sum_1",
+              "(Tensor), A tensor used to store the "
+              "parameter sums with the same shape as input(param).");
+    AddOutput("out_sum_2",
+              "(Tensor), A auxiliary tensor to help "
+              "accumulating sums of parameter values with the same shape as "
+              "input(param). It is used to avoid loss of precision due to too "
+              "many sums.");
+    AddOutput("out_sum_3",
+              "(Tensor), A auxiliary tensor to help "
+              "accumulating sums of parameter values with the same shape as "
+              "input(param).");
+    AddOutput(
+        "out_num_accumulates",
+        "(Tensor<int64_t>), The accumulating times of current window with "
+        "shape [1].");
+    AddOutput(
+        "out_old_num_accumulates",
+        "(Tensor<int64_t>) The accumulating times of previous window with "
+        "shape [1].");
+    AddOutput(
+        "out_num_updates",
+        "(Tensor<int64_t>), The total number of batches used by trainning "
+        "before this batch with shape [1].");
+
+    AddAttr<float>("average_window",
+                   "(float, default 0) "
+                   "The rate of average window size relative to num_updates.")
+        .SetDefault(0);
+    AddAttr<int64_t>("max_average_window",
+                     "(int64_t) "
+                     "Maximum size of average window. It suggests that the "
+                     "number of mini-batches "
+                     "in one pass is appropriate value to set.");
+    AddAttr<int64_t>("min_average_window",
+                     "(int64_t, default 10000L) "
+                     "Minimu size of average window.")
+        .SetDefault(10000L);
+
+    AddComment(R"DOC(
+AverageAccumulates Operator.
+Accumulate the sum of parameter whtin sliding window. The size of sliding window is
+determined by 'average_window', 'max_average_window' and 'min_average_window'.
+Memory was shared by Input(in_sum_1) and Output(out_sum_1) which acts as an accumulator 'sum_1'.
+'sum_2', 'sum_3', 'num_accumulates', 'old_num_accumulates' and 'num_updates' were the same as 'sum_1'.
+
+All the accumulators were inited to zero before training.
+
+And for a mini-batch in training, accumulators were computed as below steps:
+    num_updates += 1
+    num_accumulates += 1
+    sum_1 += param
+    if num_updates % kMaxNumAccumulates == 0:
+        sum_2 += sum_1
+        sum_1 = 0
+    if num_accumulates >= min_average_window && num_accumulates >= min(max_average_window, num_updates * average_window):
+        sum_3 = sum_1 + sum_2
+        sum_1 = 0
+        sum_2 = 0
+        old_num_accumulates = num_accumulates
+        num_accumulates = 0
+
+)DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(average_accumulates, ops::AverageAccumulatesOp,
+                  ops::AverageAccumulatesOpMaker,
+                  paddle::framework::EmptyGradOpMaker);
+REGISTER_OP_CPU_KERNEL(
+    average_accumulates,
+    ops::AverageAccumulatesKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::AverageAccumulatesKernel<paddle::platform::CPUDeviceContext, double>);

paddle/fluid/operators/average_accumulates_op.cu

+/* 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 "paddle/fluid/operators/average_accumulates_op.h"
+#include "paddle/fluid/platform/gpu_info.h"
+
+namespace paddle {
+namespace operators {
+template <>
+void GetAccumulators<paddle::platform::CUDADeviceContext>(
+    const framework::ExecutionContext& ctx, int64_t& num_updates_,
+    int64_t& num_accumulates_, int64_t& old_num_accumulates_) {
+  auto* in_old_num_accumulates = ctx.Input<Tensor>("in_old_num_accumulates");
+  auto* in_num_accumulates = ctx.Input<Tensor>("in_num_accumulates");
+  auto* in_num_updates = ctx.Input<Tensor>("in_num_updates");
+  auto stream = ctx.cuda_device_context().stream();
+  memory::Copy(platform::CPUPlace(), &old_num_accumulates_,
+               platform::CUDAPlace(), in_old_num_accumulates->data<int64_t>(),
+               sizeof(int64_t), stream);
+  memory::Copy(platform::CPUPlace(), &num_accumulates_, platform::CUDAPlace(),
+               in_num_accumulates->data<int64_t>(), sizeof(int64_t), stream);
+  memory::Copy(platform::CPUPlace(), &num_updates_, platform::CUDAPlace(),
+               in_num_updates->data<int64_t>(), sizeof(int64_t), stream);
+}
+
+template <>
+void SetAccumulators<paddle::platform::CUDADeviceContext>(
+    const framework::ExecutionContext& ctx, int64_t num_updates_,
+    int64_t num_accumulates_, int64_t old_num_accumulates_) {
+  auto stream = ctx.cuda_device_context().stream();
+  auto* out_old_num_accumulates = ctx.Output<Tensor>("out_old_num_accumulates");
+  auto* out_num_accumulates = ctx.Output<Tensor>("out_num_accumulates");
+  auto* out_num_updates = ctx.Output<Tensor>("out_num_updates");
+
+  memory::Copy(platform::CUDAPlace(), out_old_num_accumulates->data<int64_t>(),
+               platform::CPUPlace(), &old_num_accumulates_, sizeof(int64_t),
+               stream);
+  memory::Copy(platform::CUDAPlace(), out_num_accumulates->data<int64_t>(),
+               platform::CPUPlace(), &num_accumulates_, sizeof(int64_t),
+               stream);
+  memory::Copy(platform::CUDAPlace(), out_num_updates->data<int64_t>(),
+               platform::CPUPlace(), &num_updates_, sizeof(int64_t), stream);
+}
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_CUDA_KERNEL(
+    average_accumulates,
+    ops::AverageAccumulatesKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::AverageAccumulatesKernel<paddle::platform::CUDADeviceContext, double>);

paddle/fluid/operators/average_accumulates_op.h

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include <algorithm>
+#include "paddle/fluid/framework/eigen.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/math/math_function.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
+
+template <typename DeviceContext>
+void GetAccumulators(const framework::ExecutionContext& ctx,
+                     int64_t& num_updates, int64_t& num_accumulates,
+                     int64_t& old_num_accumulates);
+
+template <typename DeviceContext>
+void SetAccumulators(const framework::ExecutionContext& ctx,
+                     int64_t num_updates, int64_t num_accumulates,
+                     int64_t old_num_accumulates);
+
+template <typename DeviceContext, typename T>
+class AverageAccumulatesKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    // It is used to avoid loss of precision
+    static const int64_t kMaxNumAccumulates = 16384;
+    // Get accumulators from input
+    int64_t num_updates = 0;
+    int64_t num_accumulates = 0;
+    int64_t old_num_accumulates = 0;
+    GetAccumulators<DeviceContext>(ctx, num_updates, num_accumulates,
+                                   old_num_accumulates);
+
+    // Get attrs
+    float average_window = ctx.Attr<float>("average_window");
+    int64_t max_average_window = ctx.Attr<int64_t>("max_average_window");
+    int64_t min_average_window = ctx.Attr<int64_t>("min_average_window");
+    min_average_window =
+        std::min<int64_t>(min_average_window, max_average_window);
+
+    // Get inputs
+    auto* param = ctx.Input<Tensor>("param");
+    auto* in_sum_1 = ctx.Input<Tensor>("in_sum_1");
+    auto* in_sum_2 = ctx.Input<Tensor>("in_sum_2");
+    auto* in_sum_3 = ctx.Input<Tensor>("in_sum_3");
+    auto param_tensor = EigenVector<T>::Flatten(*param);
+    auto in_sum_1_tensor = EigenVector<T>::Flatten(*in_sum_1);
+    auto in_sum_2_tensor = EigenVector<T>::Flatten(*in_sum_2);
+    auto in_sum_3_tensor = EigenVector<T>::Flatten(*in_sum_3);
+
+    // Get outputs
+    auto* out_sum_1 = ctx.Output<Tensor>("out_sum_1");
+    auto* out_sum_2 = ctx.Output<Tensor>("out_sum_2");
+    auto* out_sum_3 = ctx.Output<Tensor>("out_sum_3");
+    auto out_sum_1_tensor = EigenVector<T>::Flatten(*out_sum_1);
+    auto out_sum_2_tensor = EigenVector<T>::Flatten(*out_sum_2);
+    auto out_sum_3_tensor = EigenVector<T>::Flatten(*out_sum_3);
+
+    // Compute
+    auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
+    math::SetConstant<DeviceContext, T> constant_functor;
+    ++num_updates;
+    ++num_accumulates;
+    out_sum_1_tensor.device(place) = in_sum_1_tensor + param_tensor;
+    out_sum_2_tensor.device(place) = in_sum_2_tensor;
+    out_sum_3_tensor.device(place) = in_sum_3_tensor;
+    if (num_updates % kMaxNumAccumulates == 0) {
+      // Move the sum to a different buffer to avoid loss of precision due to
+      // too many sums.
+      out_sum_2_tensor.device(place) = in_sum_2_tensor + in_sum_1_tensor;
+      constant_functor(ctx.template device_context<DeviceContext>(), out_sum_1,
+                       0.0);
+    }
+    if (num_accumulates >= min_average_window &&
+        num_accumulates >= std::min<int64_t>(max_average_window,
+                                             num_updates * average_window)) {
+      //  Now the average window is too long, discard the old sum.
+      out_sum_3_tensor.device(place) = in_sum_1_tensor + in_sum_2_tensor;
+      constant_functor(ctx.template device_context<DeviceContext>(), out_sum_1,
+                       0.0);
+      constant_functor(ctx.template device_context<DeviceContext>(), out_sum_2,
+                       0.0);
+      old_num_accumulates = num_accumulates;
+      num_accumulates = 0;
+    }
+
+    // Set accumulators to output
+    SetAccumulators<DeviceContext>(ctx, num_updates, num_accumulates,
+                                   old_num_accumulates);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/batch_norm_op.cc

@@ -80,6 +80,29 @@ class BatchNormOp : public framework::OperatorWithKernel {
     ctx->SetOutputDim("SavedVariance", {C});
     ctx->ShareLoD("X", "Y");
   }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext &ctx) const override {
+    auto input_data_type =
+        framework::ToDataType(ctx.Input<Tensor>("X")->type());
+    // For float or float16 input tensor, the type of the scale, bias, mean,
+    // and var tensors should both be float.
+    auto bn_param_type = framework::proto::VarType::FP32;
+    PADDLE_ENFORCE_EQ(bn_param_type,
+                      framework::ToDataType(ctx.Input<Tensor>("Scale")->type()),
+                      "Scale input should be of float type");
+    PADDLE_ENFORCE_EQ(bn_param_type,
+                      framework::ToDataType(ctx.Input<Tensor>("Bias")->type()),
+                      "Bias input should be of float type");
+    PADDLE_ENFORCE_EQ(bn_param_type,
+                      framework::ToDataType(ctx.Input<Tensor>("Mean")->type()),
+                      "Mean input should be of float type");
+    PADDLE_ENFORCE_EQ(bn_param_type, framework::ToDataType(
+                                         ctx.Input<Tensor>("Variance")->type()),
+                      "Variance input should be of float type");
+    return framework::OpKernelType(input_data_type, ctx.GetPlace());
+  }
 };
 
 class BatchNormOpMaker : public framework::OpProtoAndCheckerMaker {
@@ -434,12 +457,39 @@ class BatchNormGradKernel<platform::CPUDeviceContext, T>
   }
 };
 
+class BatchNormGradMaker : public framework::SingleGradOpDescMaker {
+ public:
+  using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
+
+ protected:
+  std::unique_ptr<framework::OpDesc> Apply() const override {
+    auto *op = new framework::OpDesc();
+    op->SetType("batch_norm_grad");
+    op->SetInput("X", Input("X"));
+    op->SetInput(framework::GradVarName("Y"), OutputGrad("Y"));
+
+    op->SetInput("Scale", Input("Scale"));
+    op->SetInput("SavedMean", Output("SavedMean"));
+    op->SetInput("SavedVariance", Output("SavedVariance"));
+
+    op->SetAttrMap(Attrs());
+
+    op->SetOutput(framework::GradVarName("X"), InputGrad("X"));
+    op->SetOutput(framework::GradVarName("Scale"), InputGrad("Scale"));
+    op->SetOutput(framework::GradVarName("Bias"), InputGrad("Bias"));
+
+    return std::unique_ptr<framework::OpDesc>(op);
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP(batch_norm, ops::BatchNormOp, ops::BatchNormOpMaker,
-            batch_norm_grad, ops::BatchNormGradOp);
+REGISTER_OPERATOR(batch_norm, ops::BatchNormOp, ops::BatchNormOpMaker,
+                  ops::BatchNormGradMaker);
+REGISTER_OPERATOR(batch_norm_grad, ops::BatchNormGradOp);
+
 REGISTER_OP_CPU_KERNEL(
     batch_norm,
     ops::BatchNormKernel<paddle::platform::CPUDeviceContext, float>);

paddle/fluid/operators/batch_norm_op.cu.cc

@@ -18,6 +18,7 @@ limitations under the License. */
 #include <cfloat>
 #include "paddle/fluid/operators/math/math_function.h"
 #include "paddle/fluid/platform/cudnn_helper.h"
+#include "paddle/fluid/platform/float16.h"
 
 namespace paddle {
 namespace operators {
@@ -26,6 +27,8 @@ using Tensor = framework::Tensor;
 using DataLayout = framework::DataLayout;
 template <typename T>
 using CudnnDataType = platform::CudnnDataType<T>;
+template <typename T>
+using BatchNormParamType = typename CudnnDataType<T>::BatchNormParamType;
 
 void ExtractNCWHD(const framework::DDim &dims, const DataLayout &data_layout,
                   int *N, int *C, int *H, int *W, int *D) {
@@ -104,8 +107,9 @@ class BatchNormKernel<platform::CUDADeviceContext, T>
     CUDNN_ENFORCE(platform::dynload::cudnnSetTensorNdDescriptor(
         data_desc_, CudnnDataType<T>::type,
         x_dims.size() > 3 ? x_dims.size() : 4, dims.data(), strides.data()));
+    // Note: PERSISTENT not implemented for inference
     CUDNN_ENFORCE(platform::dynload::cudnnDeriveBNTensorDescriptor(
-        bn_param_desc_, data_desc_, mode_));
+        bn_param_desc_, data_desc_, is_test ? CUDNN_BATCHNORM_SPATIAL : mode_));
 
     const auto *scale = ctx.Input<Tensor>("Scale");
     const auto *bias = ctx.Input<Tensor>("Bias");
@@ -118,15 +122,16 @@ class BatchNormKernel<platform::CUDADeviceContext, T>
 
     // alloc memory
     y->mutable_data<T>(ctx.GetPlace());
-    mean_out->mutable_data<T>(ctx.GetPlace());
-    variance_out->mutable_data<T>(ctx.GetPlace());
-    saved_mean->mutable_data<T>(ctx.GetPlace());
-    saved_variance->mutable_data<T>(ctx.GetPlace());
+    mean_out->mutable_data<BatchNormParamType<T>>(ctx.GetPlace());
+    variance_out->mutable_data<BatchNormParamType<T>>(ctx.GetPlace());
+    saved_mean->mutable_data<BatchNormParamType<T>>(ctx.GetPlace());
+    saved_variance->mutable_data<BatchNormParamType<T>>(ctx.GetPlace());
 
     auto &dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
-    math::SetConstant<platform::CUDADeviceContext, T> functor;
-    functor(dev_ctx, saved_mean, 0);
-    functor(dev_ctx, saved_variance, 0);
+    math::SetConstant<platform::CUDADeviceContext, BatchNormParamType<T>>
+        functor;
+    functor(dev_ctx, saved_mean, static_cast<BatchNormParamType<T>>(0));
+    functor(dev_ctx, saved_variance, static_cast<BatchNormParamType<T>>(0));
 
     auto handle = dev_ctx.cudnn_handle();
 
@@ -147,8 +152,10 @@ class BatchNormKernel<platform::CUDADeviceContext, T>
           CUDNN_BATCHNORM_SPATIAL, CudnnDataType<T>::kOne(),
           CudnnDataType<T>::kZero(), data_desc_, x->template data<T>(),
           data_desc_, y->template mutable_data<T>(ctx.GetPlace()),
-          bn_param_desc_, scale->template data<T>(), bias->template data<T>(),
-          est_mean->template data<T>(), est_var->template data<T>(), epsilon));
+          bn_param_desc_, scale->template data<BatchNormParamType<T>>(),
+          bias->template data<BatchNormParamType<T>>(),
+          est_mean->template data<BatchNormParamType<T>>(),
+          est_var->template data<BatchNormParamType<T>>(), epsilon));
     } else {
       // Run training mode.
       // obtain running mean and running inv var, and see if we need to
@@ -159,11 +166,16 @@ class BatchNormKernel<platform::CUDADeviceContext, T>
           handle, mode_, CudnnDataType<T>::kOne(), CudnnDataType<T>::kZero(),
           data_desc_, x->template data<T>(), data_desc_,
           y->template mutable_data<T>(ctx.GetPlace()), bn_param_desc_,
-          scale->template data<T>(), bias->template data<T>(), this_factor,
-          mean_out->template mutable_data<T>(ctx.GetPlace()),
-          variance_out->template mutable_data<T>(ctx.GetPlace()), epsilon,
-          saved_mean->template mutable_data<T>(ctx.GetPlace()),
-          saved_variance->template mutable_data<T>(ctx.GetPlace())));
+          scale->template data<BatchNormParamType<T>>(),
+          bias->template data<BatchNormParamType<T>>(), this_factor,
+          mean_out->template mutable_data<BatchNormParamType<T>>(
+              ctx.GetPlace()),
+          variance_out->template mutable_data<BatchNormParamType<T>>(
+              ctx.GetPlace()),
+          epsilon, saved_mean->template mutable_data<BatchNormParamType<T>>(
+                       ctx.GetPlace()),
+          saved_variance->template mutable_data<BatchNormParamType<T>>(
+              ctx.GetPlace())));
     }
 
     // clean when exit.
@@ -270,9 +282,9 @@ class BatchNormGradKernel<platform::CUDADeviceContext, T>
 }  // namespace paddle
 
 namespace ops = paddle::operators;
+namespace plat = paddle::platform;
 REGISTER_OP_CUDA_KERNEL(
-    batch_norm,
-    ops::BatchNormKernel<paddle::platform::CUDADeviceContext, float>);
+    batch_norm, ops::BatchNormKernel<plat::CUDADeviceContext, float>,
+    ops::BatchNormKernel<plat::CUDADeviceContext, plat::float16>);
 REGISTER_OP_CUDA_KERNEL(
-    batch_norm_grad,
-    ops::BatchNormGradKernel<paddle::platform::CUDADeviceContext, float>);
+    batch_norm_grad, ops::BatchNormGradKernel<plat::CUDADeviceContext, float>);

paddle/fluid/operators/box_coder_op.cc

@@ -126,6 +126,7 @@ width and height.
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_WITHOUT_GRADIENT(box_coder, ops::BoxCoderOp, ops::BoxCoderOpMaker);
+REGISTER_OPERATOR(box_coder, ops::BoxCoderOp, ops::BoxCoderOpMaker,
+                  paddle::framework::EmptyGradOpMaker);
 REGISTER_OP_CPU_KERNEL(box_coder, ops::BoxCoderKernel<float>,
                        ops::BoxCoderKernel<double>);

paddle/fluid/operators/channel_send_op.cc

@@ -23,21 +23,10 @@ limitations under the License. */
 
 static constexpr char Channel[] = "Channel";
 static constexpr char X[] = "X";
-static constexpr char Status[] = "Status";
-static constexpr char copy[] = "copy";
 
 namespace paddle {
 namespace operators {
 
-void SetSendStatus(const platform::Place &dev_place,
-                   framework::Variable &status_var, bool status) {
-  auto cpu = platform::CPUPlace();
-  auto status_tensor =
-      status_var.GetMutable<framework::LoDTensor>()->mutable_data<bool>({1},
-                                                                        cpu);
-  status_tensor[0] = status;
-}
-
 class ChannelSendOp : public framework::OperatorBase {
  public:
   ChannelSendOp(const std::string &type,
@@ -51,9 +40,6 @@ class ChannelSendOp : public framework::OperatorBase {
                    "Input(Channel) of ChannelSendOp should not be null.");
     PADDLE_ENFORCE(ctx->HasInput(X),
                    "Input(X) of ChannelSendOp should not be null.");
-    PADDLE_ENFORCE(ctx->HasOutput(Status),
-                   "Output(Status) of ChannelSendOp should not be null.");
-    ctx->SetOutputDim("Status", {1});
   }
 
  private:
@@ -65,10 +51,7 @@ class ChannelSendOp : public framework::OperatorBase {
     auto input_var = scope.FindVar(Input(X));
 
     // Send the input data through the channel.
-    bool ok = concurrency::ChannelSend(ch, input_var);
-
-    // Set the status output of the `ChannelSend` call.
-    SetSendStatus(dev_place, *scope.FindVar(Output(Status)), ok);
+    concurrency::ChannelSend(ch, input_var);
   }
 };
 
@@ -82,12 +65,6 @@ class ChannelSendOpMaker : public framework::OpProtoAndCheckerMaker {
         .AsDuplicable();
     AddInput(X, "(Variable) The value which gets sent by the channel.")
         .AsDuplicable();
-    AddOutput(Status,
-              "(Tensor) An LoD Tensor that returns a boolean status of the"
-              "result of the send operation.")
-        .AsDuplicable();
-    AddAttr<bool>(copy, "(bool, default false) Should copy before send")
-        .SetDefault(false);
     AddComment(R"DOC(
 )DOC");
   }

paddle/fluid/operators/concurrency/channel_util.cc

@@ -17,20 +17,20 @@ limitations under the License. */
 
 namespace poc = paddle::operators::concurrency;
 
-bool poc::ChannelSend(framework::ChannelHolder *ch, framework::Variable *var) {
+void poc::ChannelSend(framework::ChannelHolder *ch, framework::Variable *var) {
   auto type = framework::ToVarType(var->Type());
   if (type == framework::proto::VarType_Type_LOD_TENSOR)
-    return ch->Send(var->GetMutable<framework::LoDTensor>());
+    ch->Send(var->GetMutable<framework::LoDTensor>());
   else if (type == framework::proto::VarType_Type_LOD_RANK_TABLE)
-    return ch->Send(var->GetMutable<framework::LoDRankTable>());
+    ch->Send(var->GetMutable<framework::LoDRankTable>());
   else if (type == framework::proto::VarType_Type_LOD_TENSOR_ARRAY)
-    return ch->Send(var->GetMutable<framework::LoDTensorArray>());
+    ch->Send(var->GetMutable<framework::LoDTensorArray>());
   else if (type == framework::proto::VarType_Type_SELECTED_ROWS)
-    return ch->Send(var->GetMutable<framework::SelectedRows>());
+    ch->Send(var->GetMutable<framework::SelectedRows>());
   else if (type == framework::proto::VarType_Type_READER)
-    return ch->Send(var->GetMutable<framework::ReaderHolder>());
+    ch->Send(var->GetMutable<framework::ReaderHolder>());
   else if (type == framework::proto::VarType_Type_CHANNEL)
-    return ch->Send(var->GetMutable<framework::ChannelHolder>());
+    ch->Send(var->GetMutable<framework::ChannelHolder>());
   else
     PADDLE_THROW("ChannelSend:Unsupported type");
 }

paddle/fluid/operators/concurrency/channel_util.h

@@ -21,7 +21,7 @@ namespace paddle {
 namespace operators {
 namespace concurrency {
 
-bool ChannelSend(framework::ChannelHolder *ch, framework::Variable *var);
+void ChannelSend(framework::ChannelHolder *ch, framework::Variable *var);
 bool ChannelReceive(framework::ChannelHolder *ch, framework::Variable *var);
 
 void ChannelAddToSendQ(framework::ChannelHolder *ch, const void *referrer,

paddle/fluid/operators/cross_entropy_op.h

@@ -78,7 +78,7 @@ class CrossEntropyGradientOpKernel : public framework::OpKernel<T> {
       for (int64_t i = 0; i < batch_size; ++i) {
         PADDLE_ASSERT(label_data[i] >= 0 || label_data[i] < class_num);
         int64_t index = i * class_num + label_data[i];
-        dx_data[index] = -dy_data[i] / x_data[index];
+        dx_data[index] = math::TolerableValue<T>()(-dy_data[i] / x_data[index]);
       }
     }
   }

paddle/fluid/operators/detail/CMakeLists.txt

 if(WITH_DISTRIBUTE)
-  grpc_library(sendrecvop_grpc SRCS bytebuffer_stream.cc sendrecvop_utils.cc grpc_client.cc grpc_server.cc PROTO send_recv.proto DEPS lod_tensor selected_rows)
+  grpc_library(sendrecvop_grpc SRCS bytebuffer_stream.cc sendrecvop_utils.cc grpc_client.cc
+      grpc_server.cc variable_response.cc PROTO send_recv.proto DEPS lod_tensor selected_rows)
   set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
   set_source_files_properties(test_serde.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
-  cc_test(serde_test SRCS test_serde.cc DEPS grpc++_unsecure grpc_unsecure gpr cares zlib protobuf sendrecvop_grpc)
+  cc_test(serde_test SRCS test_serde.cc variable_response.cc DEPS grpc++_unsecure grpc_unsecure gpr
+      cares zlib protobuf sendrecvop_grpc)
 endif()

paddle/fluid/operators/detail/bytebuffer_stream.h

@@ -23,9 +23,107 @@ limitations under the License. */
 #include "google/protobuf/io/coded_stream.h"
 #include "google/protobuf/io/zero_copy_stream.h"
 
+namespace grpc {
+// A ZeroCopyInputStream that reads from grpc_byte_buffer
+class GrpcBufferReader final
+    : public ::google::protobuf::io::ZeroCopyInputStream {
+  typedef void (CoreCodegenInterface::*OldReaderInitAPI)(
+      grpc_byte_buffer_reader* reader, grpc_byte_buffer* buffer);
+  typedef int (CoreCodegenInterface::*NewReaderInitAPI)(
+      grpc_byte_buffer_reader* reader, grpc_byte_buffer* buffer);
+  void ReaderInit(OldReaderInitAPI ptr, grpc_byte_buffer_reader* reader,
+                  grpc_byte_buffer* buffer) {
+    (g_core_codegen_interface->*ptr)(reader, buffer);
+  }
+  void ReaderInit(NewReaderInitAPI ptr, grpc_byte_buffer_reader* reader,
+                  grpc_byte_buffer* buffer) {
+    int result = (g_core_codegen_interface->*ptr)(reader, buffer);
+    (void)result;
+  }
+
+ public:
+  explicit GrpcBufferReader(grpc_byte_buffer* buffer)
+      : byte_count_(0), backup_count_(0) {
+    ReaderInit(&CoreCodegenInterface::grpc_byte_buffer_reader_init, &reader_,
+               buffer);
+  }
+  ~GrpcBufferReader() override {
+    g_core_codegen_interface->grpc_byte_buffer_reader_destroy(&reader_);
+  }
+
+  bool Next(const void** data, int* size) override {
+    if (backup_count_ > 0) {
+      *data = GRPC_SLICE_START_PTR(slice_) + GRPC_SLICE_LENGTH(slice_) -
+              backup_count_;
+      GPR_CODEGEN_ASSERT(backup_count_ <= INT_MAX);
+      *size = (int)backup_count_;
+      backup_count_ = 0;
+      return true;
+    }
+    if (!g_core_codegen_interface->grpc_byte_buffer_reader_next(&reader_,
+                                                                &slice_)) {
+      return false;
+    }
+    g_core_codegen_interface->grpc_slice_unref(slice_);
+    *data = GRPC_SLICE_START_PTR(slice_);
+    // On win x64, int is only 32bit
+    GPR_CODEGEN_ASSERT(GRPC_SLICE_LENGTH(slice_) <= INT_MAX);
+    byte_count_ += * size = (int)GRPC_SLICE_LENGTH(slice_);
+    return true;
+  }
+
+  void BackUp(int count) override { backup_count_ = count; }
+
+  bool Skip(int count) override {
+    const void* data;
+    int size;
+    while (Next(&data, &size)) {
+      if (size >= count) {
+        BackUp(size - count);
+        return true;
+      }
+      // size < count;
+      count -= size;
+    }
+    // error or we have too large count;
+    return false;
+  }
+
+  ::google::protobuf::int64 ByteCount() const override {
+    return byte_count_ - backup_count_;
+  }
+
+ private:
+  int64_t byte_count_;
+  int64_t backup_count_;
+  grpc_byte_buffer_reader reader_;
+  grpc_slice slice_;
+};
+
+};  // namespace grpc
+
 namespace paddle {
 namespace operators {
 namespace detail {
+// Source provides a way for a particular RPC implementation to provide
+// received data to ParseFrom.
+class Source {
+ public:
+  virtual ~Source() {}
+
+  // Return the stream that contains the data to be parsed.
+  // Note that this method might be invoked more than once if
+  // ParseFrom needs to fall back to a more expensive parsing method.
+  // Every call must return a stream pointing at the beginning of
+  // the serialized RecvTensorResponse.
+  //
+  // Note that a subsequent call to contents() invalidates previous
+  // results of contents().
+  //
+  // Ownership of the returned stream is retained by the Source and
+  // should not be deleted by the caller.
+  virtual ::google::protobuf::io::ZeroCopyInputStream* contents() = 0;
+};
 
 // A ZeroCopyInputStream that reads from a grpc::ByteBuffer.
 class GrpcByteBufferSource
@@ -46,6 +144,43 @@ class GrpcByteBufferSource
   ::google::protobuf::int64 byte_count_;
 };
 
+class GrpcByteBufferSourceWrapper : public Source {
+ public:
+  explicit GrpcByteBufferSourceWrapper(GrpcByteBufferSource* source)
+      : source_(source) {}
+  ::google::protobuf::io::ZeroCopyInputStream* contents() override {
+    return source_;
+  }
+
+ private:
+  GrpcByteBufferSource* source_;
+};
+
+class GrpcByteSource : public Source {
+ public:
+  explicit GrpcByteSource(grpc_byte_buffer* buffer) : buffer_(buffer) {}
+  ~GrpcByteSource() override { DeleteStream(); }
+
+  typedef ::grpc::GrpcBufferReader Reader;
+
+  ::google::protobuf::io::ZeroCopyInputStream* contents() override {
+    DeleteStream();
+    stream_ = new (&space_) Reader(buffer_);
+    return stream_;
+  }
+
+ private:
+  void DeleteStream() {
+    if (stream_) {
+      stream_->~Reader();
+    }
+  }
+
+  grpc_byte_buffer* buffer_;  // Not owned
+  Reader* stream_ = nullptr;  // Points into space_ if non-nullptr
+  char space_[sizeof(Reader)];
+};
+
 }  // namespace detail
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/detail/grpc_client.cc

@@ -13,7 +13,9 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "grpc_client.h"
+#include <sys/time.h>
 #include "paddle/fluid/framework/threadpool.h"
+
 namespace paddle {
 namespace operators {
 namespace detail {
@@ -31,8 +33,9 @@ bool RPCClient::AsyncSendVariable(const std::string& ep,
 
   framework::Async([var_name_val, p_ctx, ep_val, p_scope, time_out, ch, this] {
     auto* var = p_scope->FindVar(var_name_val);
-    sendrecv::VariableMessage req;
-    SerializeToMessage(var_name_val, var, *p_ctx, &req);
+
+    ::grpc::ByteBuffer req;
+    SerializeToByteBuffer(var_name_val, var, *p_ctx, &req);
 
     // varhandle
     VarHandle var_h;
@@ -46,8 +49,10 @@ bool RPCClient::AsyncSendVariable(const std::string& ep,
     s->Prepare(var_h, time_out);
     s->response_call_back_ = NULL;
 
-    auto rpc = s->stub_->AsyncSendVariable(s->context_.get(), req, &cq_);
-    rpc->Finish(&s->reply_, &s->status_, (void*)s);
+    auto call = s->stub_g_.PrepareUnaryCall(
+        s->context_.get(), "/sendrecv.SendRecvService/SendVariable", req, &cq_);
+    call->StartCall();
+    call->Finish(&s->reply_, &s->status_, (void*)s);
   });
 
   req_count_++;
@@ -56,9 +61,19 @@ bool RPCClient::AsyncSendVariable(const std::string& ep,
 }
 
 void ProcGetResponse(const VarHandle& var_h,
-                     const sendrecv::VariableMessage& ret_msg) {
-  auto* outvar = var_h.scope->FindVar(var_h.name);
-  DeserializeFromMessage(ret_msg, *var_h.ctx, outvar);
+                     // const sendrecv::VariableMessage& ret_msg) {
+                     const ::grpc::ByteBuffer& ret_msg) {
+  framework::Variable* outvar = NULL;
+  DeserializeFromByteBuffer(ret_msg, *var_h.ctx, var_h.scope, outvar);
+}
+
+template <typename T>
+void RequestToByteBuffer(const T& proto, ::grpc::ByteBuffer* result) {
+  ::grpc::Slice slice(proto.ByteSizeLong());
+  proto.SerializeWithCachedSizesToArray(
+      const_cast<uint8_t*>(reinterpret_cast<const uint8_t*>(slice.begin())));
+  ::grpc::ByteBuffer tmp(&slice, 1);
+  result->Swap(&tmp);
 }
 
 bool RPCClient::AsyncGetVariable(const std::string& ep,
@@ -88,8 +103,13 @@ bool RPCClient::AsyncGetVariable(const std::string& ep,
     s->Prepare(var_h, time_out);
     s->response_call_back_ = ProcGetResponse;
 
-    auto rpc = s->stub_->AsyncGetVariable(s->context_.get(), req, &cq_);
-    rpc->Finish(&s->reply_, &s->status_, (void*)s);
+    ::grpc::ByteBuffer buf;
+    RequestToByteBuffer<sendrecv::VariableMessage>(req, &buf);
+
+    auto call = s->stub_g_.PrepareUnaryCall(
+        s->context_.get(), "/sendrecv.SendRecvService/GetVariable", buf, &cq_);
+    call->StartCall();
+    call->Finish(&s->reply_, &s->status_, (void*)s);
   });
 
   req_count_++;

paddle/fluid/operators/detail/grpc_client.h

@@ -25,6 +25,11 @@ limitations under the License. */
 #include <string>
 #include <vector>
 
+#include <grpc++/generic/generic_stub.h>
+#include <grpc++/grpc++.h>
+#include <grpc++/support/byte_buffer.h>
+#include <grpc++/support/slice.h>
+
 #include "paddle/fluid/framework/data_type.h"
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/scope.h"
@@ -49,15 +54,11 @@ struct VarHandle {
   }
 };
 
-void ProcGetResponse(const VarHandle& var_h,
-                     const sendrecv::VariableMessage& msg);
+void ProcGetResponse(const VarHandle& var_h, const grpc::ByteBuffer& msg);
 
 class BaseProcessor {
  public:
-  explicit BaseProcessor(std::shared_ptr<grpc::Channel> ch) {
-    stub_ = sendrecv::SendRecvService::NewStub(ch);
-    context_ = NULL;
-  }
+  explicit BaseProcessor(std::shared_ptr<grpc::Channel> ch) { context_ = NULL; }
 
   virtual ~BaseProcessor() {}
 
@@ -82,19 +83,18 @@ class BaseProcessor {
 
   virtual void Process() = 0;
 
-  std::unique_ptr<sendrecv::SendRecvService::Stub> stub_;
   std::unique_ptr<grpc::ClientContext> context_;
   grpc::Status status_;
   VarHandle var_h_;
 };
 
-typedef std::function<void(const VarHandle&, const sendrecv::VoidMessage&)>
+typedef std::function<void(const VarHandle&, const ::grpc::ByteBuffer&)>
     RequestSendCallBack;
 
 class SendProcessor : public BaseProcessor {
  public:
   explicit SendProcessor(std::shared_ptr<grpc::Channel> ch)
-      : BaseProcessor(ch) {}
+      : BaseProcessor(ch), stub_g_(ch) {}
 
   virtual ~SendProcessor() {}
 
@@ -104,17 +104,18 @@ class SendProcessor : public BaseProcessor {
     }
   }
 
-  sendrecv::VoidMessage reply_;
+  ::grpc::GenericStub stub_g_;
+  ::grpc::ByteBuffer reply_;
   RequestSendCallBack response_call_back_ = NULL;
 };
 
-typedef std::function<void(const VarHandle&, const sendrecv::VariableMessage&)>
+typedef std::function<void(const VarHandle&, const ::grpc::ByteBuffer&)>
     RequestGetCallBack;
 
 class GetProcessor : public BaseProcessor {
  public:
   explicit GetProcessor(std::shared_ptr<grpc::Channel> ch)
-      : BaseProcessor(ch) {}
+      : BaseProcessor(ch), stub_g_(ch) {}
 
   virtual ~GetProcessor() {}
 
@@ -124,30 +125,37 @@ class GetProcessor : public BaseProcessor {
     }
   }
 
-  sendrecv::VariableMessage reply_;
+  ::grpc::ByteBuffer reply_;
+  ::grpc::GenericStub stub_g_;
   RequestGetCallBack response_call_back_ = ProcGetResponse;
 };
 
 class BatchBarrierProcessor : public BaseProcessor {
  public:
   explicit BatchBarrierProcessor(std::shared_ptr<grpc::Channel> ch)
-      : BaseProcessor(ch) {}
+      : BaseProcessor(ch) {
+    stub_ = sendrecv::SendRecvService::NewStub(ch);
+  }
 
   virtual ~BatchBarrierProcessor() {}
 
   virtual void Process() {}
   sendrecv::VoidMessage reply_;
+  std::unique_ptr<sendrecv::SendRecvService::Stub> stub_;
 };
 
 class FetchBarrierProcessor : public BaseProcessor {
  public:
   explicit FetchBarrierProcessor(std::shared_ptr<grpc::Channel> ch)
-      : BaseProcessor(ch) {}
+      : BaseProcessor(ch) {
+    stub_ = sendrecv::SendRecvService::NewStub(ch);
+  }
 
   virtual ~FetchBarrierProcessor() {}
 
   virtual void Process() {}
   sendrecv::VariableMessage reply_;
+  std::unique_ptr<sendrecv::SendRecvService::Stub> stub_;
 };
 
 class RPCClient {

paddle/fluid/operators/detail/grpc_server.cc

@@ -14,7 +14,7 @@ limitations under the License. */
 
 #include "paddle/fluid/operators/detail/grpc_server.h"
 
-using grpc::ServerAsyncResponseWriter;
+using ::grpc::ServerAsyncResponseWriter;
 
 namespace paddle {
 namespace operators {
@@ -26,9 +26,10 @@ enum CallStatus { PROCESS = 0, FINISH };
 // https://stackoverflow.com/questions/41732884/grpc-multiple-services-in-cpp-async-server
 class RequestBase {
  public:
-  explicit RequestBase(sendrecv::SendRecvService::AsyncService* service,
-                       grpc::ServerCompletionQueue* cq)
-      : service_(service), cq_(cq), status_(PROCESS) {
+  explicit RequestBase(GrpcService::AsyncService* service,
+                       ::grpc::ServerCompletionQueue* cq,
+                       const platform::DeviceContext* dev_ctx)
+      : service_(service), cq_(cq), status_(PROCESS), dev_ctx_(dev_ctx) {
     PADDLE_ENFORCE(cq_);
   }
   virtual ~RequestBase() {}
@@ -42,55 +43,58 @@ class RequestBase {
   }
 
  protected:
-  grpc::ServerContext ctx_;
-  sendrecv::SendRecvService::AsyncService* service_;
-  grpc::ServerCompletionQueue* cq_;
+  ::grpc::ServerContext ctx_;
+  GrpcService::AsyncService* service_;
+  ::grpc::ServerCompletionQueue* cq_;
   CallStatus status_;
+  const platform::DeviceContext* dev_ctx_;
 };
 
-typedef std::pair<std::string, sendrecv::VariableMessage> MessageWithName;
-
 class RequestSend final : public RequestBase {
  public:
-  explicit RequestSend(sendrecv::SendRecvService::AsyncService* service,
-                       grpc::ServerCompletionQueue* cq,
-                       SimpleBlockQueue<MessageWithName>* queue)
-      : RequestBase(service, cq), queue_(queue), responder_(&ctx_) {
-    service_->RequestSendVariable(&ctx_, &request_, &responder_, cq_, cq_,
-                                  this);
+  explicit RequestSend(GrpcService::AsyncService* service,
+                       ::grpc::ServerCompletionQueue* cq,
+                       framework::Scope* scope, ReceivedQueue* queue,
+                       const platform::DeviceContext* dev_ctx)
+      : RequestBase(service, cq, dev_ctx), queue_(queue), responder_(&ctx_) {
+    request_.reset(new VariableResponse(scope, dev_ctx_));
+    int method_id = static_cast<int>(detail::GrpcMethod::kSendVariable);
+    service_->RequestAsyncUnary(method_id, &ctx_, request_.get(), &responder_,
+                                cq_, cq_, this);
   }
 
   virtual ~RequestSend() {}
 
-  virtual std::string GetReqName() { return request_.varname(); }
+  virtual std::string GetReqName() { return request_->Varname(); }
 
   virtual void Process() {
-    MessageWithName msg_with_name =
-        std::make_pair(request_.varname(), std::move(request_));
-    queue_->Push(std::move(msg_with_name));
-    responder_.Finish(reply_, grpc::Status::OK, this);
+    queue_->Push(std::make_pair(request_->Varname(), request_));
+
+    sendrecv::VoidMessage reply;
+    responder_.Finish(reply, ::grpc::Status::OK, this);
     status_ = FINISH;
   }
 
  protected:
-  sendrecv::VariableMessage request_;
-  sendrecv::VoidMessage reply_;
-  SimpleBlockQueue<MessageWithName>* queue_;
+  std::shared_ptr<VariableResponse> request_;
+  ReceivedQueue* queue_;
   ServerAsyncResponseWriter<sendrecv::VoidMessage> responder_;
 };
 
 class RequestGet final : public RequestBase {
  public:
-  explicit RequestGet(sendrecv::SendRecvService::AsyncService* service,
-                      grpc::ServerCompletionQueue* cq, framework::Scope* scope,
+  explicit RequestGet(GrpcService::AsyncService* service,
+                      ::grpc::ServerCompletionQueue* cq,
+                      framework::Scope* scope,
                       const platform::DeviceContext* dev_ctx,
                       SimpleBlockQueue<MessageWithName>* queue)
-      : RequestBase(service, cq),
+      : RequestBase(service, cq, dev_ctx),
         responder_(&ctx_),
         scope_(scope),
-        dev_ctx_(dev_ctx),
         queue_(queue) {
-    service_->RequestGetVariable(&ctx_, &request_, &responder_, cq_, cq_, this);
+    int method_id = static_cast<int>(detail::GrpcMethod::kGetVariable);
+    service_->RequestAsyncUnary(method_id, &ctx_, &request_, &responder_, cq_,
+                                cq_, this);
   }
 
   virtual ~RequestGet() {}
@@ -101,24 +105,26 @@ class RequestGet final : public RequestBase {
     // proc request.
     std::string var_name = request_.varname();
     auto* var = scope_->FindVar(var_name);
+
+    ::grpc::ByteBuffer reply;
     if (var_name != FETCH_BARRIER_MESSAGE) {
-      SerializeToMessage(var_name, var, *dev_ctx_, &reply_);
+      SerializeToByteBuffer(var_name, var, *dev_ctx_, &reply);
     }
-    // TODO(gongwb): check var's info.
-    responder_.Finish(reply_, grpc::Status::OK, this);
+
+    responder_.Finish(reply, ::grpc::Status::OK, this);
     status_ = FINISH;
-    MessageWithName msg_with_name =
-        //          request name    reply
-        std::make_pair(var_name, std::move(reply_));
+
+    if (var_name == FETCH_BARRIER_MESSAGE) {
+      sendrecv::VariableMessage msg;
+      MessageWithName msg_with_name = std::make_pair(var_name, msg);
       queue_->Push(msg_with_name);
     }
+  }
 
  protected:
   sendrecv::VariableMessage request_;
-  sendrecv::VariableMessage reply_;
-  ServerAsyncResponseWriter<sendrecv::VariableMessage> responder_;
+  ServerAsyncResponseWriter<::grpc::ByteBuffer> responder_;
   framework::Scope* scope_;
-  const platform::DeviceContext* dev_ctx_;
   SimpleBlockQueue<MessageWithName>* queue_;
 };
 
@@ -133,8 +139,8 @@ void AsyncGRPCServer::WaitClientGet(int count) {
 }
 
 void AsyncGRPCServer::RunSyncUpdate() {
-  grpc::ServerBuilder builder;
-  builder.AddListeningPort(address_, grpc::InsecureServerCredentials());
+  ::grpc::ServerBuilder builder;
+  builder.AddListeningPort(address_, ::grpc::InsecureServerCredentials());
   builder.SetMaxSendMessageSize(std::numeric_limits<int>::max());
   builder.SetMaxReceiveMessageSize(std::numeric_limits<int>::max());
   builder.RegisterService(&service_);
@@ -182,8 +188,8 @@ void AsyncGRPCServer::TryToRegisterNewSendOne() {
   if (is_shut_down_) {
     return;
   }
-  RequestSend* send =
-      new RequestSend(&service_, cq_send_.get(), &var_recv_queue_);
+  RequestSend* send = new RequestSend(&service_, cq_send_.get(), scope_,
+                                      &var_recv_queue_, dev_ctx_);
   VLOG(4) << "Create RequestSend status:" << send->Status();
 }
 
@@ -198,7 +204,7 @@ void AsyncGRPCServer::TryToRegisterNewGetOne() {
 }
 
 // FIXME(typhoonzero): change cq_name to enum.
-void AsyncGRPCServer::HandleRequest(grpc::ServerCompletionQueue* cq,
+void AsyncGRPCServer::HandleRequest(::grpc::ServerCompletionQueue* cq,
                                     std::string cq_name,
                                     std::function<void()> TryToRegisterNewOne) {
   TryToRegisterNewOne();

paddle/fluid/operators/detail/grpc_server.h

@@ -14,28 +14,31 @@ limitations under the License. */
 
 #pragma once
 
+#include <grpc++/grpc++.h>
+#include <thread>
+
 #include "paddle/fluid/framework/lod_tensor.h"
 #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/simple_block_queue.h"
-
+#include "paddle/fluid/operators/detail/grpc_service.h"
 #include "paddle/fluid/operators/detail/send_recv.grpc.pb.h"
 #include "paddle/fluid/operators/detail/send_recv.pb.h"
-
-#include <grpc++/grpc++.h>
-#include <grpc/support/log.h>
-#include <thread>
 #include "paddle/fluid/operators/detail/sendrecvop_utils.h"
+#include "paddle/fluid/operators/detail/simple_block_queue.h"
 
 namespace paddle {
 namespace operators {
 namespace detail {
 
+typedef std::pair<std::string, std::shared_ptr<VariableResponse>>
+    ReceivedMessage;
+typedef SimpleBlockQueue<ReceivedMessage> ReceivedQueue;
+
 typedef std::pair<std::string, sendrecv::VariableMessage> MessageWithName;
 class RequestBase;
 
-class AsyncGRPCServer final : public sendrecv::SendRecvService::Service {
+class AsyncGRPCServer final {
  public:
   explicit AsyncGRPCServer(const std::string &address) : address_(address) {}
 
@@ -50,14 +53,16 @@ class AsyncGRPCServer final : public sendrecv::SendRecvService::Service {
 
   void SetDevCtx(const platform::DeviceContext *dev_ctx) { dev_ctx_ = dev_ctx; }
 
-  const MessageWithName Get() { return this->var_recv_queue_.Pop(); }
+  const ReceivedMessage Get() { return this->var_recv_queue_.Pop(); }
 
-  void Push(const MessageWithName &msg) { this->var_recv_queue_.Push(msg); }
+  void Push(const std::string &msg_name) {
+    this->var_recv_queue_.Push(std::make_pair(msg_name, nullptr));
+  }
 
   void ShutDown();
 
  protected:
-  void HandleRequest(grpc::ServerCompletionQueue *cq, std::string cq_name,
+  void HandleRequest(::grpc::ServerCompletionQueue *cq, std::string cq_name,
                      std::function<void()> TryToRegisterNewOne);
   void TryToRegisterNewSendOne();
   void TryToRegisterNewGetOne();
@@ -66,18 +71,19 @@ class AsyncGRPCServer final : public sendrecv::SendRecvService::Service {
  private:
   std::mutex cq_mutex_;
   volatile bool is_shut_down_ = false;
-  std::unique_ptr<grpc::ServerCompletionQueue> cq_send_;
-  std::unique_ptr<grpc::ServerCompletionQueue> cq_get_;
+  std::unique_ptr<::grpc::ServerCompletionQueue> cq_send_;
+  std::unique_ptr<::grpc::ServerCompletionQueue> cq_get_;
 
-  sendrecv::SendRecvService::AsyncService service_;
-  std::unique_ptr<grpc::Server> server_;
+  GrpcService::AsyncService service_;
+  std::unique_ptr<::grpc::Server> server_;
 
   std::string address_;
   framework::Scope *scope_;
   const platform::DeviceContext *dev_ctx_;
+
   // received variable from RPC, operators fetch variable from this queue.
-  SimpleBlockQueue<MessageWithName> var_recv_queue_;
   SimpleBlockQueue<MessageWithName> var_get_queue_;
+  ReceivedQueue var_recv_queue_;
 
   // condition of the sub program
   std::mutex barrier_mutex_;

paddle/fluid/operators/detail/grpc_service.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 <grpc++/impl/codegen/async_stream.h>
+#include <grpc++/impl/codegen/async_unary_call.h>
+#include <grpc++/impl/codegen/proto_utils.h>
+#include <grpc++/impl/codegen/rpc_method.h>
+#include <grpc++/impl/codegen/service_type.h>
+#include <grpc++/impl/codegen/status.h>
+#include <grpc++/impl/codegen/stub_options.h>
+#include <grpc++/impl/codegen/sync_stream.h>
+#include <grpc++/support/byte_buffer.h>
+#include "paddle/fluid/operators/detail/variable_response.h"
+
+// NOTE: This method was originally created by tensorflow
+//       (https://github.com/tensorflow/tensorflow/) we borrow this
+//       method and did some modifications so that we can parse gRPC
+//       requests without too much copying of the tensor data.
+
+namespace grpc {
+class CompletionQueue;
+class Channel;
+class RpcService;
+class ServerCompletionQueue;
+class ServerContext;
+
+// Support parsing/unparsing of tensorflow::VariableResponse.
+// Wire-format is identical to RecvVariableResponse.
+template <>
+class SerializationTraits<paddle::operators::detail::VariableResponse> {
+ public:
+  static Status Serialize(
+      const paddle::operators::detail::VariableResponse& msg,
+      grpc_byte_buffer** bp, bool* own_buffer) {
+    PADDLE_ENFORCE(false, "SerializationTraits::Serialize not implemented!");
+    return Status();
+  }
+  static Status Deserialize(grpc_byte_buffer* buffer,
+                            paddle::operators::detail::VariableResponse* msg,
+                            int max_message_size = INT_MAX) {
+    if (buffer == nullptr) {
+      return Status(StatusCode::INTERNAL, "No payload");
+    }
+
+    Status result = g_core_codegen_interface->ok();
+    if (result.ok()) {
+      paddle::operators::detail::GrpcByteSource source(buffer);
+      int ret = msg->Parse(&source);
+      if (ret != 0) {
+        result = Status(StatusCode::INTERNAL, "VariableResponse parse error");
+      }
+    }
+    g_core_codegen_interface->grpc_byte_buffer_destroy(buffer);
+    return result;
+  }
+};
+}  // namespace grpc
+
+namespace paddle {
+namespace operators {
+namespace detail {
+
+enum class GrpcMethod {
+  kSendVariable,
+  kGetVariable,
+};
+
+static const int kGrpcNumMethods =
+    static_cast<int>(GrpcMethod::kGetVariable) + 1;
+
+inline const char* GrpcMethodName(GrpcMethod id) {
+  switch (id) {
+    case GrpcMethod::kSendVariable:
+      return "/sendrecv.SendRecvService/SendVariable";
+    case GrpcMethod::kGetVariable:
+      return "/sendrecv.SendRecvService/GetVariable";
+  }
+
+  // Shouldn't be reached.
+  PADDLE_ENFORCE(false, "Invalid id: not found valid method name");
+  return nullptr;
+}
+
+class GrpcService final {
+ public:
+  class AsyncService : public ::grpc::Service {
+   public:
+    AsyncService() {
+      for (int i = 0; i < kGrpcNumMethods; ++i) {
+        AddMethod(new ::grpc::internal::RpcServiceMethod(
+            GrpcMethodName(static_cast<GrpcMethod>(i)),
+            ::grpc::internal::RpcMethod::NORMAL_RPC, nullptr));
+        ::grpc::Service::MarkMethodAsync(i);
+      }
+    }
+    virtual ~AsyncService() {}
+
+    // Make RequestAsyncUnary public for grpc_call.h
+    using ::grpc::Service::RequestAsyncUnary;
+  };
+};
+
+}  // namespace detail
+}  // namespace operator
+}  // namespace paddle

paddle/fluid/operators/detail/send_recv.proto

@@ -32,6 +32,9 @@ enum VarType {
   SELECTED_ROWS = 1;
 }
 
+// NOTICE(gongwb):don't modify this proto if you are not
+//   not familar with how we serialize in sendrecvop_utils.h
+//   and deserilize it in  variable_response.h.
 message VariableMessage {
   enum Type {
     // Pod Types
@@ -45,7 +48,6 @@ message VariableMessage {
   }
 
   message LodData { repeated int64 lod_data = 1; }
-
   string varname = 1;
   // TODO(Yancey1989): reference framework::proto::VarDesc::VarType
   VarType type = 2;
@@ -64,3 +66,5 @@ message VariableMessage {
 }
 
 message VoidMessage {}
+
+message TestMessage { int64 test_1 = 1; }

paddle/fluid/operators/detail/sendrecvop_utils.cc

@@ -13,61 +13,19 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/detail/sendrecvop_utils.h"
+#include <sys/time.h>
+#include <thread>
 #include "google/protobuf/io/coded_stream.h"
 #include "google/protobuf/io/zero_copy_stream.h"
 #include "paddle/fluid/framework/data_type.h"
 #include "paddle/fluid/operators/detail/bytebuffer_stream.h"
 #include "paddle/fluid/operators/detail/proto_encoder_helper.h"
+#include "paddle/fluid/operators/detail/variable_response.h"
 
 namespace paddle {
 namespace operators {
 namespace detail {
 
-void SerializeToMessage(const std::string& name, const framework::Variable* var,
-                        const platform::DeviceContext& ctx,
-                        sendrecv::VariableMessage* msg) {
-  msg->set_varname(name);
-  std::ostringstream oss;
-  switch (framework::ToVarType(var->Type())) {
-    case framework::proto::VarType_Type_LOD_TENSOR:
-      msg->set_type(sendrecv::VarType::LOD_TENSOR);
-      framework::SerializeToStream(oss, var->Get<framework::LoDTensor>(), ctx);
-      break;
-    case framework::proto::VarType_Type_SELECTED_ROWS:
-      msg->set_type(sendrecv::VarType::SELECTED_ROWS);
-      framework::SerializeToStream(oss, var->Get<framework::SelectedRows>(),
-                                   ctx);
-      break;
-    default: {
-      PADDLE_THROW("Serialize does not support type: %s",
-                   typeid(var->Type()).name());
-      break;
-    }
-  }
-  msg->set_serialized(oss.str());
-}
-
-void DeserializeFromMessage(const sendrecv::VariableMessage& msg,
-                            const platform::DeviceContext& ctx,
-                            framework::Variable* var) {
-  std::istringstream iss(msg.serialized());
-  switch (msg.type()) {
-    case sendrecv::VarType::LOD_TENSOR:
-      DeserializeFromStream(iss, var->GetMutable<framework::LoDTensor>(), ctx);
-      break;
-    case sendrecv::VarType::SELECTED_ROWS: {
-      DeserializeFromStream(iss, var->GetMutable<framework::SelectedRows>(),
-                            ctx);
-      break;
-    }
-    default: {
-      PADDLE_THROW("Deserialize does not support type: %s",
-                   typeid(var->Type()).name());
-      break;
-    }
-  }
-}
-
 void SerializeToByteBuffer(const std::string& name, framework::Variable* var,
                            const platform::DeviceContext& ctx,
                            ::grpc::ByteBuffer* msg) {
@@ -123,6 +81,7 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var,
             static_cast<const platform::CUDADeviceContext&>(ctx);
         auto copy_size = tensor.memory_size();
         payload = memory::Alloc(cpu, copy_size);
+
         memory::Copy(cpu, payload,
                      boost::get<platform::CUDAPlace>(tensor.place()),
                      reinterpret_cast<const void*>(tensor.data<void>()),
@@ -132,6 +91,7 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var,
           platform::CPUPlace cpu;
           memory::Free(cpu, backing);
         };
+
 #endif
       } else {
         payload = tensor.data<void>();
@@ -219,80 +179,11 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var,
 
 void DeserializeFromByteBuffer(const ::grpc::ByteBuffer& msg,
                                const platform::DeviceContext& ctx,
-                               framework::Variable* var) {
-  sendrecv::VariableMessage meta;
-  GrpcByteBufferSource source;
-  source.Init(msg);
-  ::google::protobuf::io::CodedInputStream input(&source);
-  // do zerocopy parsing
-  PADDLE_ENFORCE(meta.ParseFromCodedStream(&input));
-  PADDLE_ENFORCE(input.ConsumedEntireMessage());
-  // dims is needed by both tensor and selectedrows
-  std::vector<int> vecdims;
-  for (auto& d : meta.dims()) {
-    vecdims.push_back(d);
-  }
-  framework::DDim dims = framework::make_ddim(vecdims);
-
-  if (meta.type() == sendrecv::LOD_TENSOR) {
-    auto* tensor = var->GetMutable<framework::LoDTensor>();
-    tensor->Resize(dims);
-    void* tensor_data = tensor->mutable_data(
-        ctx.GetPlace(),
-        paddle::operators::detail::ToTypeIndex(meta.data_type()));
-    framework::LoD lod;
-    for (int i = 0; i < meta.lod_level(); ++i) {
-      framework::Vector<size_t> v;
-      for (int j = 0; j < meta.lod(i).lod_data_size(); ++j) {
-        v.push_back(meta.lod(i).lod_data(j));
-      }
-      lod.push_back(v);
-    }
-    tensor->set_lod(lod);
-    // How to avoid copying and use the message buffer directly?
-    // Maybe need to find a way to release all memory except tensor content.
-    if (platform::is_gpu_place(ctx.GetPlace())) {
-#ifdef PADDLE_WITH_CUDA
-      platform::CPUPlace cpu;
-      auto& gpu_dev_ctx = static_cast<const platform::CUDADeviceContext&>(ctx);
-      memory::Copy(boost::get<platform::CUDAPlace>(tensor->place()),
-                   tensor_data, cpu,
-                   reinterpret_cast<const void*>(meta.serialized().data()),
-                   meta.serialized().size(), gpu_dev_ctx.stream());
-      ctx.Wait();
-#endif
-    } else {
-      memcpy(tensor_data,
-             reinterpret_cast<const void*>(meta.serialized().data()),
-             meta.serialized().size());
-    }
-  } else if (meta.type() == sendrecv::SELECTED_ROWS) {
-    auto* slr = var->GetMutable<framework::SelectedRows>();
-    auto* tensor = slr->mutable_value();
-    int64_t* rows_data = slr->mutable_rows()->data();
-    tensor->Resize(dims);
-    void* tensor_data = tensor->mutable_data(
-        ctx.GetPlace(),
-        paddle::operators::detail::ToTypeIndex(meta.data_type()));
-    if (platform::is_gpu_place(ctx.GetPlace())) {
-#ifdef PADDLE_WITH_CUDA
-      platform::CPUPlace cpu;
-      auto& gpu_dev_ctx = static_cast<const platform::CUDADeviceContext&>(ctx);
-      memory::Copy(boost::get<platform::CUDAPlace>(tensor->place()),
-                   tensor_data, cpu,
-                   reinterpret_cast<const void*>(meta.serialized().data()),
-                   meta.serialized().size(), gpu_dev_ctx.stream());
-      ctx.Wait();
-#endif
-    } else {
-      memcpy(tensor_data,
-             reinterpret_cast<const void*>(meta.serialized().data()),
-             meta.serialized().size());
-    }
-    // copy rows CPU data, GPU data will be copied lazly
-    memcpy(rows_data, reinterpret_cast<const void*>(meta.rows().data()),
-           meta.rows().size());
-  }
+                               const framework::Scope* scope,
+                               framework::Variable*& var) {
+  operators::detail::VariableResponse resp(scope, &ctx);
+  PADDLE_ENFORCE(resp.Parse(msg) == 0, "parse bytebuffer to tensor error!");
+  var = resp.GetVar();
 }
 
 }  // namespace detail

paddle/fluid/operators/detail/sendrecvop_utils.h

@@ -21,6 +21,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/framework/selected_rows.h"
+#include "paddle/fluid/framework/tensor_util.h"
 #include "paddle/fluid/framework/var_type.h"
 
 #include "paddle/fluid/operators/detail/send_recv.grpc.pb.h"
@@ -36,21 +37,14 @@ namespace detail {
 
 typedef void (*DestroyCallback)(void*);
 
-void SerializeToMessage(const std::string& name, const framework::Variable* var,
-                        const platform::DeviceContext& ctx,
-                        sendrecv::VariableMessage* msg);
-
-void DeserializeFromMessage(const sendrecv::VariableMessage& msg,
-                            const platform::DeviceContext& ctx,
-                            framework::Variable* var);
-
 void SerializeToByteBuffer(const std::string& name, framework::Variable* var,
                            const platform::DeviceContext& ctx,
                            ::grpc::ByteBuffer* msg);
 
 void DeserializeFromByteBuffer(const ::grpc::ByteBuffer& msg,
                                const platform::DeviceContext& ctx,
-                               framework::Variable* var);
+                               const framework::Scope* scope,
+                               framework::Variable*& var);
 
 inline std::type_index ToTypeIndex(sendrecv::VariableMessage::Type type) {
   switch (type) {

paddle/fluid/operators/detail/test_serde.cc

@@ -16,11 +16,13 @@ limitations under the License. */
 #include <string>
 #include <thread>
 
+#include <google/protobuf/text_format.h>
 #include "gtest/gtest.h"
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/tensor_util.h"
 #include "paddle/fluid/framework/variable.h"
 #include "paddle/fluid/operators/detail/sendrecvop_utils.h"
+#include "paddle/fluid/operators/detail/variable_response.h"
 #include "paddle/fluid/operators/math/math_function.h"
 #include "paddle/fluid/platform/place.h"
 #include "paddle/fluid/string/printf.h"
@@ -31,19 +33,21 @@ namespace operators = paddle::operators;
 namespace math = paddle::operators::math;
 namespace memory = paddle::memory;
 
-void RunSerdeTestTensor(platform::Place place) {
-  // serialize var to ByteBuffer
-  framework::Variable var;
-  auto* tensor = var.GetMutable<framework::LoDTensor>();
-  tensor->Resize(framework::make_ddim({4, 8, 4, 2}));
-  framework::LoD lod;
-  lod.push_back(framework::Vector<size_t>({1, 3, 8}));
-  tensor->set_lod(lod);
-  int tensor_numel = 4 * 8 * 4 * 2;
+void RunSerdeTestSelectedRows(platform::Place place) {
   platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
   auto& ctx = *pool.Get(place);
+
+  // serialize var to ByteBuffer
+  framework::Variable var;
+  auto* slr = var.GetMutable<framework::SelectedRows>();
+  auto* tensor = slr->mutable_value();
+  auto* rows = slr->mutable_rows();
+  tensor->Resize(framework::make_ddim({2, 10}));
   tensor->mutable_data<float>(place);
-  math::set_constant(ctx, tensor, 31.9);
+  int tensor_numel = 2 * 10;
+  math::set_constant(ctx, tensor, 32.7);
+  rows->push_back(3);
+  rows->push_back(10);
 
   ::grpc::ByteBuffer msg;
   operators::detail::SerializeToByteBuffer("myvar", &var, ctx, &msg);
@@ -56,62 +60,67 @@ void RunSerdeTestTensor(platform::Place place) {
   for (const auto& s : slices) {
     tmp.append(reinterpret_cast<const char*>(s.begin()), s.size());
   }
+
   sendrecv::VariableMessage varmsg;
   EXPECT_TRUE(varmsg.ParseFromString(tmp));
+
   EXPECT_EQ(varmsg.varname(), "myvar");
-  EXPECT_EQ(varmsg.type(), 0);
-  EXPECT_EQ(varmsg.dims()[0], 4);
-  EXPECT_EQ(varmsg.dims()[1], 8);
-  EXPECT_EQ(varmsg.dims()[2], 4);
-  EXPECT_EQ(varmsg.dims()[3], 2);
-  EXPECT_EQ(varmsg.lod_level(), 1);
-  EXPECT_EQ(varmsg.lod(0).lod_data(0), 1);
-  EXPECT_EQ(varmsg.lod(0).lod_data(1), 3);
-  EXPECT_EQ(varmsg.lod(0).lod_data(2), 8);
+  EXPECT_EQ(varmsg.type(), 1);
 
   const float* tensor_data =
       reinterpret_cast<const float*>(varmsg.serialized().data());
+  const int64_t* rows_data =
+      reinterpret_cast<const int64_t*>(varmsg.rows().data());
   for (int i = 0; i < tensor_numel; ++i) {
-    EXPECT_FLOAT_EQ(tensor_data[i], 31.9);
+    EXPECT_FLOAT_EQ(tensor_data[i], 32.7);
   }
-
+  EXPECT_EQ(rows_data[0], 3);
+  EXPECT_EQ(rows_data[1], 10);
   // deserialize zero-copy
-  framework::Variable var2;
-  operators::detail::DeserializeFromByteBuffer(msg, ctx, &var2);
-  auto tensor2 = var2.Get<framework::LoDTensor>();
+  // framework::Variable var2;
+  // operators::detail::DeserializeFromByteBuffer(msg, ctx, &var2);
+  framework::Scope scope;
+  scope.Var("myvar");
+  operators::detail::VariableResponse resp(&scope, &ctx);
+  EXPECT_EQ(resp.Parse(msg), 0);
+
+  framework::Variable* var2 = resp.GetVar();
+
+  auto* slr2 = var2->GetMutable<framework::SelectedRows>();
+  auto* tensor2 = slr2->mutable_value();
+  auto* rows2 = slr2->mutable_rows();
   float* tensor_data2 = nullptr;
   framework::Tensor tmp_tensor;
 
   if (platform::is_gpu_place(ctx.GetPlace())) {
     platform::CPUPlace cpu;
-    framework::TensorCopy(tensor2, cpu, &tmp_tensor);
+    framework::TensorCopy(*tensor2, cpu, &tmp_tensor);
     tensor_data2 = tmp_tensor.data<float>();
   } else {
-    tensor_data2 = const_cast<float*>(tensor2.data<float>());
+    tensor_data2 = const_cast<float*>(tensor2->data<float>());
   }
+  const int64_t* rows_data2 = rows2->data();
 
-  EXPECT_EQ(varmsg.lod_level(), 1);
-  EXPECT_EQ(varmsg.lod(0).lod_data(0), 1);
-  EXPECT_EQ(varmsg.lod(0).lod_data(1), 3);
-  EXPECT_EQ(varmsg.lod(0).lod_data(2), 8);
-  for (int i = 0; i < tensor_numel; ++i) EXPECT_FLOAT_EQ(tensor_data2[i], 31.9);
+  for (int i = 0; i < tensor_numel; ++i) {
+    EXPECT_FLOAT_EQ(tensor_data2[i], 32.7);
+  }
+  EXPECT_EQ(rows_data2[0], 3);
+  EXPECT_EQ(rows_data2[1], 10);
 }
 
-void RunSerdeTestSelectedRows(platform::Place place) {
-  platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
-  auto& ctx = *pool.Get(place);
-
+void RunTestLodTensor(platform::Place place, int from_type = 0) {
   // serialize var to ByteBuffer
   framework::Variable var;
-  auto* slr = var.GetMutable<framework::SelectedRows>();
-  auto* tensor = slr->mutable_value();
-  auto* rows = slr->mutable_rows();
-  tensor->Resize(framework::make_ddim({2, 10}));
+  auto* tensor = var.GetMutable<framework::LoDTensor>();
+  tensor->Resize(framework::make_ddim({4, 8, 4, 2}));
+  framework::LoD lod;
+  lod.push_back(framework::Vector<size_t>({1, 3, 8}));
+  tensor->set_lod(lod);
+  int tensor_numel = 4 * 8 * 4 * 2;
+  platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
+  auto& ctx = *pool.Get(place);
   tensor->mutable_data<float>(place);
-  int tensor_numel = 2 * 10;
-  math::set_constant(ctx, tensor, 32.7);
-  rows->push_back(3);
-  rows->push_back(10);
+  math::set_constant(ctx, tensor, 31.9);
 
   ::grpc::ByteBuffer msg;
   operators::detail::SerializeToByteBuffer("myvar", &var, ctx, &msg);
@@ -126,61 +135,82 @@ void RunSerdeTestSelectedRows(platform::Place place) {
   }
   sendrecv::VariableMessage varmsg;
   EXPECT_TRUE(varmsg.ParseFromString(tmp));
-
   EXPECT_EQ(varmsg.varname(), "myvar");
-  EXPECT_EQ(varmsg.type(), 1);
+  EXPECT_EQ(varmsg.type(), 0);
+  EXPECT_EQ(varmsg.dims()[0], 4);
+  EXPECT_EQ(varmsg.dims()[1], 8);
+  EXPECT_EQ(varmsg.dims()[2], 4);
+  EXPECT_EQ(varmsg.dims()[3], 2);
+  EXPECT_EQ(varmsg.lod_level(), 1);
+  EXPECT_EQ(varmsg.lod(0).lod_data(0), 1);
+  EXPECT_EQ(varmsg.lod(0).lod_data(1), 3);
+  EXPECT_EQ(varmsg.lod(0).lod_data(2), 8);
 
   const float* tensor_data =
       reinterpret_cast<const float*>(varmsg.serialized().data());
-  const int64_t* rows_data =
-      reinterpret_cast<const int64_t*>(varmsg.rows().data());
   for (int i = 0; i < tensor_numel; ++i) {
-    EXPECT_FLOAT_EQ(tensor_data[i], 32.7);
+    EXPECT_FLOAT_EQ(tensor_data[i], 31.9);
   }
-  EXPECT_EQ(rows_data[0], 3);
-  EXPECT_EQ(rows_data[1], 10);
+
+  // message binary
+  std::string str;
+  varmsg.SerializeToString(&str);
+
+  // message bytebuffer
+  ::grpc::Slice slices_2[1];
+  int num_slices = 1;
+  slices_2[0] = ::grpc::Slice(str.length());
+  memcpy(const_cast<uint8_t*>(slices_2[0].begin()), str.c_str(), str.length());
+  ::grpc::ByteBuffer bytebuffer2(&slices_2[0], num_slices);
+
   // deserialize zero-copy
-  framework::Variable var2;
-  operators::detail::DeserializeFromByteBuffer(msg, ctx, &var2);
+  framework::Scope scope;
+  scope.Var("myvar");
+  operators::detail::VariableResponse resp(&scope, &ctx);
+  if (from_type == 0) {
+    EXPECT_EQ(resp.Parse(msg), 0);
+  } else {
+    EXPECT_EQ(resp.Parse(bytebuffer2), 0);
+  }
 
-  auto* slr2 = var2.GetMutable<framework::SelectedRows>();
-  auto* tensor2 = slr2->mutable_value();
-  auto* rows2 = slr2->mutable_rows();
+  framework::Variable* var2 = resp.GetVar();
+
+  auto tensor2 = var2->Get<framework::LoDTensor>();
   float* tensor_data2 = nullptr;
   framework::Tensor tmp_tensor;
 
   if (platform::is_gpu_place(ctx.GetPlace())) {
     platform::CPUPlace cpu;
-    framework::TensorCopy(*tensor2, cpu, &tmp_tensor);
+    framework::TensorCopy(tensor2, cpu, &tmp_tensor);
     tensor_data2 = tmp_tensor.data<float>();
   } else {
-    tensor_data2 = const_cast<float*>(tensor2->data<float>());
+    tensor_data2 = const_cast<float*>(tensor2.data<float>());
   }
-  const int64_t* rows_data2 = rows2->data();
 
-  for (int i = 0; i < tensor_numel; ++i) {
-    EXPECT_FLOAT_EQ(tensor_data2[i], 32.7);
-  }
-  EXPECT_EQ(rows_data2[0], 3);
-  EXPECT_EQ(rows_data2[1], 10);
+  EXPECT_EQ(varmsg.lod_level(), 1);
+  EXPECT_EQ(varmsg.lod(0).lod_data(0), 1);
+  EXPECT_EQ(varmsg.lod(0).lod_data(1), 3);
+  EXPECT_EQ(varmsg.lod(0).lod_data(2), 8);
+  for (int i = 0; i < tensor_numel; ++i) EXPECT_FLOAT_EQ(tensor_data2[i], 31.9);
 }
 
-TEST(SelectedRows, CPU) {
+TEST(LodTensor, Run) {
   platform::CPUPlace place;
-  RunSerdeTestSelectedRows(place);
+  RunTestLodTensor(place);
+  RunTestLodTensor(place, 1);
+#ifdef PADDLE_WITH_CUDA
+  platform::CUDAPlace gpu(0);
+  RunTestLodTensor(gpu);
+  RunTestLodTensor(gpu, 1);
+#endif
 }
 
-TEST(SelectedRows, GPU) {
-  platform::CUDAPlace place;
-  RunSerdeTestSelectedRows(place);
-}
-
-TEST(Tensor, CPU) {
+TEST(SelectedRows, Run) {
   platform::CPUPlace place;
-  RunSerdeTestTensor(place);
-}
+  RunSerdeTestSelectedRows(place);
 
-TEST(Tensor, GPU) {
-  platform::CUDAPlace place;
-  RunSerdeTestTensor(place);
+#ifdef PADDLE_WITH_CUDA
+  platform::CUDAPlace gpu;
+  RunSerdeTestSelectedRows(gpu);
+#endif
 }

paddle/fluid/operators/detail/variable_response.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/variable_response.h"
+#include <string.h>
+#include "paddle/fluid/operators/detail/send_recv.pb.h"
+#include "paddle/fluid/operators/detail/sendrecvop_utils.h"
+
+namespace paddle {
+namespace operators {
+namespace detail {
+
+enum WireType {
+  WIRETYPE_VARINT = 0,
+  WIRETYPE_LENGTH_DELIMITED = 2,
+};
+
+inline int GetTagFieldNumber(uint32_t tag) { return tag >> 3; }
+
+inline WireType GetTagWireType(uint32_t tag) {
+  return static_cast<WireType>(tag & 0x7);
+}
+
+bool ReadVarintSizeAsInt(::google::protobuf::io::CodedInputStream* input,
+                         int* result) {
+  uint64_t v;
+  if (input->ReadVarint64(&v) && v <= static_cast<uint64_t>(INT_MAX)) {
+    *result = static_cast<int>(v);
+    return true;
+  } else {
+    return false;
+  }
+}
+
+bool ReadRaw(::google::protobuf::io::CodedInputStream* input,
+             const platform::DeviceContext& dev_ctx, platform::Place place,
+             void* dest, int size) {
+  const void* data = NULL;
+  int size_to_write = 0;
+
+  if (platform::is_gpu_place(place)) {
+#ifdef PADDLE_WITH_CUDA
+    auto& gpu_dev_ctx =
+        static_cast<const platform::CUDADeviceContext&>(dev_ctx);
+    platform::CPUPlace cpu;
+
+    char* p = reinterpret_cast<char*>(dest);
+    while (size > 0) {
+      if (!input->GetDirectBufferPointer(&data, &size_to_write)) {
+        return false;
+      }
+
+      memory::Copy(boost::get<platform::CUDAPlace>(place),
+                   reinterpret_cast<void*>(p), cpu, data, size_to_write,
+                   gpu_dev_ctx.stream());
+      p += size_to_write;
+      size -= size_to_write;
+
+      input->Skip(size_to_write);
+    }
+    gpu_dev_ctx.Wait();
+#else
+    PADDLE_THROW("Unexpected branch");
+#endif
+    return true;
+  }
+
+  char* p = reinterpret_cast<char*>(dest);
+  while (size > 0) {
+    if (!input->GetDirectBufferPointer(&data, &size_to_write)) {
+      return false;
+    }
+    // TODO(gongwb): can we avoid copy?
+    platform::CPUPlace cpu;
+    memory::Copy(cpu, reinterpret_cast<void*>(p), cpu, data, size_to_write);
+
+    p += size_to_write;
+    size -= size_to_write;
+
+    input->Skip(size_to_write);
+  }
+
+  return true;
+}
+
+bool VariableResponse::CopyLodTensorData(
+    ::google::protobuf::io::CodedInputStream* input,
+    const platform::DeviceContext& ctx, framework::DDim& dims, int length) {
+  auto var = scope_->FindVar(meta_.varname());
+  auto* tensor = var->GetMutable<framework::LoDTensor>();
+  tensor->Resize(dims);
+
+  framework::LoD lod;
+  for (int i = 0; i < meta_.lod_level(); ++i) {
+    framework::Vector<size_t> v;
+    for (int j = 0; j < meta_.lod(i).lod_data_size(); ++j) {
+      v.push_back(meta_.lod(i).lod_data(j));
+    }
+    lod.push_back(v);
+  }
+  tensor->set_lod(lod);
+
+  void* tensor_data =
+      tensor->mutable_data(ctx.GetPlace(), ToTypeIndex(meta_.data_type()));
+
+  if (!ReadRaw(input, ctx, tensor->place(), tensor_data, length)) {
+    return false;
+  }
+
+  return true;
+}
+
+inline framework::DDim GetDims(
+    const ::google::protobuf::RepeatedField<::google::protobuf::int64>& dims) {
+  std::vector<int> vecdims;
+  for (auto& d : dims) {
+    vecdims.push_back(d);
+  }
+  return framework::make_ddim(vecdims);
+}
+
+bool VariableResponse::CopySelectRowsTensorData(
+    ::google::protobuf::io::CodedInputStream* input,
+    const platform::DeviceContext& ctx, framework::DDim& dims, int length) {
+  auto var = scope_->FindVar(meta_.varname());
+  auto* slr = var->GetMutable<framework::SelectedRows>();
+  auto* tensor = slr->mutable_value();
+  tensor->Resize(dims);
+  void* tensor_data = tensor->mutable_data(
+      ctx.GetPlace(),
+      paddle::operators::detail::ToTypeIndex(meta_.data_type()));
+
+  if (!ReadRaw(input, ctx, tensor->place(), tensor_data, length)) {
+    return false;
+  }
+
+  return true;
+}
+
+bool VariableResponse::CopySelectRowsData(
+    ::google::protobuf::io::CodedInputStream* input,
+    const platform::DeviceContext& ctx, int length) {
+  auto var = scope_->FindVar(meta_.varname());
+  auto* slr = var->GetMutable<framework::SelectedRows>();
+  int64_t* rows_data = slr->mutable_rows()->data();
+
+  // copy rows CPU data, GPU data will be copied lazily.
+  platform::CPUPlace cpu;
+  if (!ReadRaw(input, ctx, cpu, rows_data, length)) {
+    return false;
+  }
+
+  return true;
+}
+
+bool ParseLodData(::google::protobuf::io::CodedInputStream* input,
+                  std::vector<int64_t>* lod) {
+  while (true) {
+    auto p = input->ReadTagWithCutoff(127);
+    int tag = GetTagFieldNumber(p.first);
+    WireType wt = GetTagWireType(p.first);
+
+    if (!p.second) {
+      return (tag == 0);
+    }
+
+    switch (tag) {
+      case sendrecv::VariableMessage_LodData::kLodDataFieldNumber: {
+        uint64_t v;
+        if (wt == WIRETYPE_VARINT) {
+          if (!input->ReadVarint64(&v)) {
+            return false;
+          }
+          lod->push_back(v);
+          break;
+        }
+
+        if (wt == WIRETYPE_LENGTH_DELIMITED) {
+          int length = 0;
+          if (!input->ReadVarintSizeAsInt(&length)) {
+            return tag;
+          }
+
+          for (int i = 0; i < length; i++) {
+            uint64_t v;
+            if (!input->ReadVarint64(&v)) {
+              return false;
+            }
+            lod->push_back(v);
+          }
+          break;
+        }
+
+        return false;
+      }
+      default: { return false; }
+    }
+  }
+
+  return true;
+}
+
+int VariableResponse::Parse(const ::grpc::ByteBuffer& byte_buffer) {
+  GrpcByteBufferSource source;
+  source.Init(byte_buffer);
+  GrpcByteBufferSourceWrapper r(&source);
+
+  return Parse(&r);
+}
+
+int VariableResponse::Parse(Source* source) {
+  ::google::protobuf::io::ZeroCopyInputStream* input_stream =
+      source->contents();
+  ::google::protobuf::io::CodedInputStream input(input_stream);
+  input.SetTotalBytesLimit(INT_MAX, INT_MAX);
+
+  while (true) {
+    auto p = input.ReadTagWithCutoff(127);
+    int tag = GetTagFieldNumber(p.first);
+    WireType wt = GetTagWireType(p.first);
+    if (!p.second) {
+      if (tag != 0) {
+        return -1;
+      }
+
+      return 0;
+    }
+
+    switch (tag) {
+      case sendrecv::VariableMessage::kVarnameFieldNumber: {
+        uint32_t length;
+        if ((wt != WIRETYPE_LENGTH_DELIMITED) || !input.ReadVarint32(&length)) {
+          return tag;
+        }
+
+        std::string temp;
+        if (!input.ReadString(&temp, length)) {
+          return tag;
+        }
+
+        meta_.set_varname(temp);
+        break;
+      }
+      case sendrecv::VariableMessage::kTypeFieldNumber: {
+        uint64_t v;
+        if ((wt != WIRETYPE_VARINT) || !input.ReadVarint64(&v)) {
+          return tag;
+        }
+
+        meta_.set_type(static_cast<::sendrecv::VarType>(v));
+        break;
+      }
+      case sendrecv::VariableMessage::kDataTypeFieldNumber: {
+        uint64_t v = 0;
+        if ((wt != WIRETYPE_VARINT) || !input.ReadVarint64(&v)) {
+          return tag;
+        }
+
+        meta_.set_data_type(static_cast<::sendrecv::VariableMessage_Type>(v));
+        break;
+      }
+      case sendrecv::VariableMessage::kDimsFieldNumber: {
+        // not packed
+        if (wt == WIRETYPE_VARINT) {
+          uint64_t v;
+          if (!input.ReadVarint64(&v)) {
+            return tag;
+          }
+          meta_.add_dims(v);
+          break;
+        }
+
+        // packed
+        if (wt == WIRETYPE_LENGTH_DELIMITED) {
+          int length = 0;
+          if (!input.ReadVarintSizeAsInt(&length)) {
+            return tag;
+          }
+          for (int i = 0; i < length; i++) {
+            uint64_t v;
+            if (!input.ReadVarint64(&v)) {
+              return tag;
+            }
+            meta_.add_dims(v);
+          }
+          break;
+        }
+
+        return tag;
+      }
+      case sendrecv::VariableMessage::kLodLevelFieldNumber: {
+        uint64_t v = 0;
+        if ((wt != WIRETYPE_VARINT) || !input.ReadVarint64(&v)) {
+          return tag;
+        }
+        meta_.set_lod_level(static_cast<int64_t>(v));
+        break;
+      }
+      case sendrecv::VariableMessage::kLodFieldNumber: {
+        int length = 0;
+        if (wt != WIRETYPE_LENGTH_DELIMITED ||
+            !ReadVarintSizeAsInt(&input, &length)) {
+          return tag;
+        }
+
+        std::pair<::google::protobuf::io::CodedInputStream::Limit, int> p =
+            input.IncrementRecursionDepthAndPushLimit(length);
+
+        std::vector<int64_t> lod_data;
+        if (p.second < 0 || !ParseLodData(&input, &lod_data)) {
+          return tag;
+        }
+
+        if (!input.DecrementRecursionDepthAndPopLimit(p.first)) {
+          return false;
+        }
+
+        if (lod_data.size() == 0) {
+          break;
+        }
+
+        auto lod = meta_.add_lod();
+        for (uint32_t i = 0; i < lod_data.size(); i++) {
+          lod->add_lod_data(lod_data[i]);
+        }
+        break;
+      }
+      case sendrecv::VariableMessage::kSerializedFieldNumber: {
+        PADDLE_ENFORCE((meta_.type() == sendrecv::SELECTED_ROWS ||
+                        meta_.type() == sendrecv::LOD_TENSOR) &&
+                           meta_.varname() != "",
+                       "meta info should be got first!");
+
+        int length = 0;
+        if (wt != WIRETYPE_LENGTH_DELIMITED ||
+            !ReadVarintSizeAsInt(&input, &length)) {
+          return tag;
+        }
+
+        framework::DDim dims = GetDims(meta_.dims());
+        if (meta_.type() == sendrecv::LOD_TENSOR) {
+          PADDLE_ENFORCE(meta_.lod_size() >= 0,
+                         "lod info should be got first!");
+          if (!CopyLodTensorData(&input, *dev_ctx_, dims, length)) {
+            return tag;
+          }
+          break;
+        }
+
+        if (meta_.type() == sendrecv::SELECTED_ROWS) {
+          if (!CopySelectRowsTensorData(&input, *dev_ctx_, dims, length)) {
+            return tag;
+          }
+          break;
+        }
+
+        return tag;
+      }
+      case sendrecv::VariableMessage::kRowsFieldNumber: {
+        PADDLE_ENFORCE((meta_.type() == sendrecv::SELECTED_ROWS ||
+                        meta_.type() == sendrecv::LOD_TENSOR) &&
+                           meta_.varname() != "",
+                       "meta info should be got first!");
+
+        int length = 0;
+        if (wt != WIRETYPE_LENGTH_DELIMITED ||
+            !ReadVarintSizeAsInt(&input, &length)) {
+          return tag;
+        }
+
+        if (!CopySelectRowsData(&input, *dev_ctx_, length)) {
+          return tag;
+        }
+        break;
+      }
+
+      default: {
+        // Unknown tag, return unknown error.
+        return -1;
+      }
+    }
+  }
+
+  return 0;
+}
+
+};  // namespace detail
+};  // namespace operators
+};  // namespace paddle

paddle/fluid/operators/detail/variable_response.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/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/framework/var_type.h"
+
+#include "paddle/fluid/operators/detail/send_recv.grpc.pb.h"
+#include "paddle/fluid/operators/detail/send_recv.pb.h"
+
+#include "google/protobuf/io/coded_stream.h"
+#include "google/protobuf/io/zero_copy_stream.h"
+#include "paddle/fluid/framework/tensor.h"
+#include "paddle/fluid/operators/detail/bytebuffer_stream.h"
+
+namespace paddle {
+namespace operators {
+namespace detail {
+
+class VariableResponse {
+ public:
+  VariableResponse(const framework::Scope* scope,
+                   const platform::DeviceContext* dev_ctx)
+      : scope_(scope), dev_ctx_(dev_ctx) {}
+
+  virtual ~VariableResponse() {}
+
+  // return:
+  // 0:ok.
+  // -1: unkown error.
+  // other: number of error field.
+  int Parse(Source* source);
+
+  // return:
+  // 0:ok.
+  // -1: unkown error.
+  // other: number of error field.
+  int Parse(const ::grpc::ByteBuffer& byte_buffer);
+
+  inline std::string Varname() { return meta_.varname(); }
+
+  // should call parse first.
+  framework::Variable* GetVar() { return scope_->FindVar(meta_.varname()); }
+
+ private:
+  bool CopySelectRowsTensorData(::google::protobuf::io::CodedInputStream* input,
+                                const platform::DeviceContext& ctx,
+                                framework::DDim& dims, int length);
+
+  bool CopySelectRowsData(::google::protobuf::io::CodedInputStream* input,
+                          const platform::DeviceContext& ctx, int length);
+
+  bool CopyLodTensorData(::google::protobuf::io::CodedInputStream* input,
+                         const platform::DeviceContext& ctx,
+                         framework::DDim& dims, int length);
+
+ private:
+  const framework::Scope* scope_;
+  const platform::DeviceContext* dev_ctx_;
+  // only Skeleton
+  sendrecv::VariableMessage meta_;
+};
+
+};  // namespace detail
+};  // namespace operators
+};  // namespace paddle

paddle/fluid/operators/detection_map_op.cc

@@ -188,8 +188,8 @@ The general steps are as follows. First, calculate the true positive and
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_WITHOUT_GRADIENT(detection_map, ops::DetectionMAPOp,
-                             ops::DetectionMAPOpMaker);
+REGISTER_OPERATOR(detection_map, ops::DetectionMAPOp, ops::DetectionMAPOpMaker,
+                  paddle::framework::EmptyGradOpMaker);
 REGISTER_OP_CPU_KERNEL(
     detection_map, ops::DetectionMAPOpKernel<paddle::platform::CPUPlace, float>,
     ops::DetectionMAPOpKernel<paddle::platform::CPUPlace, double>);

paddle/fluid/operators/dropout_op.cc

@@ -35,7 +35,6 @@ class DropoutOp : public framework::OperatorWithKernel {
   }
 };
 
-template <typename AttrType>
 class DropoutOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   DropoutOpMaker(OpProto* proto, OpAttrChecker* op_checker)
@@ -73,7 +72,6 @@ are set equal to their corresponding inputs.
   }
 };
 
-template <typename AttrType>
 class DropoutOpGrad : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
@@ -103,11 +101,10 @@ class DropoutOpGrad : public framework::OperatorWithKernel {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP(dropout, ops::DropoutOp, ops::DropoutOpMaker<float>, dropout_grad,
-            ops::DropoutOpGrad<float>);
+REGISTER_OP(dropout, ops::DropoutOp, ops::DropoutOpMaker, dropout_grad,
+            ops::DropoutOpGrad);
 REGISTER_OP_CPU_KERNEL(
-    dropout,
-    ops::CPUDropoutKernel<paddle::platform::CPUDeviceContext, float, float>);
+    dropout, ops::CPUDropoutKernel<paddle::platform::CPUDeviceContext, float>);
 REGISTER_OP_CPU_KERNEL(
     dropout_grad,
     ops::DropoutGradKernel<paddle::platform::CPUDeviceContext, float>);

paddle/fluid/operators/dropout_op.cu

@@ -18,20 +18,22 @@ limitations under the License. */
 #include <thrust/random.h>
 #include <thrust/transform.h>
 #include "paddle/fluid/operators/dropout_op.h"
+#include "paddle/fluid/platform/float16.h"
 
 namespace paddle {
 namespace operators {
 
-template <typename T, typename AttrType>
+template <typename T>
 __global__ void RandomGenerator(const size_t n, const int seed,
-                                const AttrType dropout_prob, const T* src,
+                                const float dropout_prob, const T* src,
                                 T* mask_data, T* dst) {
   thrust::minstd_rand rng;
   rng.seed(seed);
-  thrust::uniform_real_distribution<AttrType> dist(0, 1);
+  thrust::uniform_real_distribution<float> dist(0, 1);
 
   int idx = blockDim.x * blockIdx.x + threadIdx.x;
   for (; idx < n; idx += blockDim.x * gridDim.x) {
+    rng.discard(idx);
     if (dist(rng) < dropout_prob) {
       mask_data[idx] = static_cast<T>(0);
     } else {
@@ -44,17 +46,14 @@ __global__ void RandomGenerator(const size_t n, const int seed,
 // It seems that Eigen::Tensor::setRandom in GPU will SEGFAULT.
 // Use std::random and thrust::random(thrust is a std library in CUDA) to
 // implement uniform random.
-template <typename Place, typename T, typename AttrType>
+template <typename Place, typename T>
 class GPUDropoutKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
     auto* x = context.Input<Tensor>("X");
     auto* y = context.Output<Tensor>("Out");
     y->mutable_data<T>(context.GetPlace());
-    AttrType dropout_prob = context.Attr<AttrType>("dropout_prob");
-
-    auto X = EigenMatrix<T>::Reshape(*x, 1);
-    auto Y = EigenMatrix<T>::Reshape(*y, 1);
+    float dropout_prob = context.Attr<float>("dropout_prob");
 
     auto& place = *context.template device_context<Place>().eigen_device();
     if (!context.Attr<bool>("is_test")) {
@@ -70,11 +69,13 @@ class GPUDropoutKernel : public framework::OpKernel<T> {
 
       int threads = 512;
       int grid = (x->numel() + threads - 1) / threads;
-      RandomGenerator<T, AttrType><<<grid, threads, 0,
-                                     context.cuda_device_context().stream()>>>(
+      RandomGenerator<
+          T><<<grid, threads, 0, context.cuda_device_context().stream()>>>(
           size, seed, dropout_prob, x_data, mask_data, y_data);
     } else {
-      Y.device(place) = X * (1.0f - dropout_prob);
+      auto X = EigenMatrix<T>::Reshape(*x, 1);
+      auto Y = EigenMatrix<T>::Reshape(*y, 1);
+      Y.device(place) = X * static_cast<T>(1.0f - dropout_prob);
     }
   }
 };
@@ -83,9 +84,9 @@ class GPUDropoutKernel : public framework::OpKernel<T> {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
+namespace plat = paddle::platform;
 REGISTER_OP_CUDA_KERNEL(
-    dropout,
-    ops::GPUDropoutKernel<paddle::platform::CUDADeviceContext, float, float>);
-REGISTER_OP_CUDA_KERNEL(
-    dropout_grad,
-    ops::DropoutGradKernel<paddle::platform::CUDADeviceContext, float>);
+    dropout, ops::GPUDropoutKernel<plat::CUDADeviceContext, float>,
+    ops::GPUDropoutKernel<plat::CUDADeviceContext, plat::float16>);
+REGISTER_OP_CUDA_KERNEL(dropout_grad,
+                        ops::DropoutGradKernel<plat::CUDADeviceContext, float>);

paddle/fluid/operators/dropout_op.h

@@ -25,7 +25,7 @@ template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
 
-template <typename DeviceContext, typename T, typename AttrType>
+template <typename DeviceContext, typename T>
 class CPUDropoutKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {

paddle/fluid/operators/dropout_op_test.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 <unistd.h>
+#include <string>
+#include <thread>
+
+#include "gtest/gtest.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/operator.h"
+#include "paddle/fluid/framework/program_desc.h"
+#include "paddle/fluid/operators/dropout_op.h"
+#include "paddle/fluid/operators/math/math_function.h"
+#include "paddle/fluid/string/printf.h"
+
+namespace f = paddle::framework;
+namespace p = paddle::platform;
+namespace m = paddle::operators::math;
+
+USE_OP(dropout);
+
+void Compare(f::Scope& scope, p::DeviceContext& ctx) {
+  // init
+  auto var = scope.Var("X");
+  auto tensor = var->GetMutable<f::LoDTensor>();
+  tensor->Resize({10, 10});
+
+  std::vector<float> init;
+  for (int64_t i = 0; i < 10 * 10; ++i) {
+    init.push_back(1.0);
+  }
+
+  TensorFromVector(init, ctx, tensor);
+
+  auto place = ctx.GetPlace();
+  auto out_var = scope.Var("Out");
+  auto out_tensor = out_var->GetMutable<f::LoDTensor>();
+  out_tensor->Resize({10, 10});
+  out_tensor->mutable_data<float>(place);  // allocate
+
+  auto mask_var = scope.Var("Mask");
+  auto mask_tensor = mask_var->GetMutable<f::LoDTensor>();
+  mask_tensor->Resize({10, 10});
+  mask_tensor->mutable_data<float>(place);  // allocate
+
+  // run
+  f::AttributeMap attrs;
+  float dropout_prob = 0.5;
+  attrs.insert({"fix_seed", 1});
+  attrs.insert({"seed", 3});
+  attrs.insert({"dropout_prob", dropout_prob});
+  auto dropout_op = f::OpRegistry::CreateOp(
+      "dropout", {{"X", {"X"}}}, {{"Out", {"Out"}}, {"Mask", {"Mask"}}}, attrs);
+
+  dropout_op->Run(scope, place);
+
+  std::vector<float> out_vec;
+  TensorToVector(*out_tensor, ctx, &out_vec);
+
+  std::vector<float> std_out = {
+      0, 0, 1, 1, 1, 1, 1, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 1, 0, 1,
+      1, 0, 1, 1, 1, 1, 0, 1, 1, 1, 1, 0, 1, 1, 0, 0, 0, 1, 1, 0,
+      1, 0, 1, 1, 0, 0, 0, 1, 1, 0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1,
+      1, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 1, 0,
+      1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 0, 1, 1, 1, 1, 1, 0, 0, 1, 1};
+
+  EXPECT_EQ(out_vec.size(), std_out.size());
+  for (uint32_t i = 0; i < out_vec.size(); i++) {
+    EXPECT_EQ(out_vec[i], std_out[i]);
+  }
+}
+
+TEST(Dropout, CPUDense) {
+  f::Scope scope;
+  p::CPUPlace place;
+  p::CPUDeviceContext ctx(place);
+  Compare(scope, ctx);
+}
+
+TEST(Dropout, GPUDense) {
+  f::Scope scope;
+  p::CUDAPlace place;
+  p::CUDADeviceContext ctx(place);
+  Compare(scope, ctx);
+}

paddle/fluid/operators/elementwise_add_op.cu

@@ -14,19 +14,20 @@ limitations under the License. */
 
 #define EIGEN_USE_GPU
 #include "paddle/fluid/operators/elementwise_add_op.h"
+#include "paddle/fluid/platform/float16.h"
 
 namespace ops = paddle::operators;
+namespace plat = paddle::platform;
 
 REGISTER_OP_CUDA_KERNEL(
-    elementwise_add,
-    ops::ElementwiseAddKernel<paddle::platform::CUDADeviceContext, float>,
-    ops::ElementwiseAddKernel<paddle::platform::CUDADeviceContext, double>,
-    ops::ElementwiseAddKernel<paddle::platform::CUDADeviceContext, int>,
-    ops::ElementwiseAddKernel<paddle::platform::CUDADeviceContext, int64_t>);
+    elementwise_add, ops::ElementwiseAddKernel<plat::CUDADeviceContext, float>,
+    ops::ElementwiseAddKernel<plat::CUDADeviceContext, double>,
+    ops::ElementwiseAddKernel<plat::CUDADeviceContext, int>,
+    ops::ElementwiseAddKernel<plat::CUDADeviceContext, int64_t>,
+    ops::ElementwiseAddKernel<plat::CUDADeviceContext, plat::float16>);
 REGISTER_OP_CUDA_KERNEL(
     elementwise_add_grad,
-    ops::ElementwiseAddGradKernel<paddle::platform::CUDADeviceContext, float>,
-    ops::ElementwiseAddGradKernel<paddle::platform::CUDADeviceContext, double>,
-    ops::ElementwiseAddGradKernel<paddle::platform::CUDADeviceContext, int>,
-    ops::ElementwiseAddGradKernel<paddle::platform::CUDADeviceContext,
-                                  int64_t>);
+    ops::ElementwiseAddGradKernel<plat::CUDADeviceContext, float>,
+    ops::ElementwiseAddGradKernel<plat::CUDADeviceContext, double>,
+    ops::ElementwiseAddGradKernel<plat::CUDADeviceContext, int>,
+    ops::ElementwiseAddGradKernel<plat::CUDADeviceContext, int64_t>);

paddle/fluid/operators/iou_similarity_op.cc

@@ -87,8 +87,9 @@ $$
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_WITHOUT_GRADIENT(iou_similarity, ops::IOUSimilarityOp,
-                             ops::IOUSimilarityOpMaker);
+REGISTER_OPERATOR(iou_similarity, ops::IOUSimilarityOp,
+                  ops::IOUSimilarityOpMaker,
+                  paddle::framework::EmptyGradOpMaker);
 
 REGISTER_OP_CPU_KERNEL(
     iou_similarity,

paddle/fluid/operators/layer_norm_op.h

@@ -22,6 +22,103 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
+// Wrap RowwiseMean and ColwiseMean.
+// Reuse the cpu codes and replace the gpu codes with cublas_gemv, which is
+// significantly faster. Unlike the RowwiseMean and ColwiseMean, the
+// implementation only considers 2D.
+template <typename DeviceContext, typename T>
+struct RowwiseMean2D {
+  RowwiseMean2D(int left, int right, const platform::DeviceContext& dev_ctx);
+
+  void operator()(const platform::DeviceContext& context,
+                  const framework::Tensor& input, framework::Tensor* vec);
+};
+
+#ifdef PADDLE_WITH_CUDA
+template <typename T>
+class RowwiseMean2D<platform::CUDADeviceContext, T> {
+ public:
+  RowwiseMean2D(int left, int right, const platform::DeviceContext& dev_ctx)
+      : left_(left), right_(right) {
+    framework::DDim ones_dim({right_});
+    divisor_.mutable_data<T>(ones_dim, dev_ctx.GetPlace());
+    math::set_constant(dev_ctx, &divisor_, 1.0 / right);
+  }
+  void operator()(const platform::CUDADeviceContext& context,
+                  const framework::Tensor& input, framework::Tensor* out) {
+    math::gemv<platform::CUDADeviceContext, T>(
+        context, false, left_, right_, 1., input.data<T>(), divisor_.data<T>(),
+        0., out->data<T>());
+  }
+
+ private:
+  int left_;
+  int right_;
+  framework::Tensor divisor_;
+};
+#endif
+
+template <typename T>
+class RowwiseMean2D<platform::CPUDeviceContext, T> {
+ public:
+  RowwiseMean2D(int left, int right, const platform::DeviceContext& dev_ctx) {}
+
+  void operator()(const platform::CPUDeviceContext& context,
+                  const framework::Tensor& input, framework::Tensor* out) {
+    row_mean_(context, input, out);
+  }
+
+ private:
+  math::RowwiseMean<platform::CPUDeviceContext, T> row_mean_;
+};
+
+template <typename DeviceContext, typename T>
+struct ColwiseSum2D {
+  ColwiseSum2D(int left, int right, const platform::DeviceContext& dev_ctx);
+
+  void operator()(const platform::DeviceContext& context,
+                  const framework::Tensor& input, framework::Tensor* vec);
+};
+
+#ifdef PADDLE_WITH_CUDA
+template <typename T>
+class ColwiseSum2D<platform::CUDADeviceContext, T> {
+ public:
+  ColwiseSum2D(int left, int right, const platform::DeviceContext& dev_ctx)
+      : left_(left), right_(right) {
+    framework::DDim ones_dim({left_});
+    divisor_.mutable_data<T>(ones_dim, dev_ctx.GetPlace());
+    math::set_constant(dev_ctx, &divisor_, 1.0);
+  }
+
+  void operator()(const platform::CUDADeviceContext& context,
+                  const framework::Tensor& input, framework::Tensor* out) {
+    math::gemv<platform::CUDADeviceContext, T>(
+        context, true, left_, right_, 1., input.data<T>(), divisor_.data<T>(),
+        0., out->data<T>());
+  }
+
+ private:
+  int left_;
+  int right_;
+  framework::Tensor divisor_;
+};
+#endif
+
+template <typename T>
+class ColwiseSum2D<platform::CPUDeviceContext, T> {
+ public:
+  ColwiseSum2D(int left, int right, const platform::DeviceContext& dev_ctx) {}
+
+  void operator()(const platform::CPUDeviceContext& context,
+                  const framework::Tensor& input, framework::Tensor* out) {
+    col_wise_(context, input, out);
+  }
+
+ private:
+  math::ColwiseSum<platform::CPUDeviceContext, T> col_wise_;
+};
+
 template <typename T>
 struct SubAndSquareFunctor {
   inline HOSTDEVICE T operator()(T a, T b) const { return (a - b) * (a - b); }
@@ -67,15 +164,15 @@ using DataLayout = framework::DataLayout;
 template <typename DeviceContext, typename T>
 class LayerNormKernel : public framework::OpKernel<T> {
  public:
-  void Compute(const framework::ExecutionContext &ctx) const override {
+  void Compute(const framework::ExecutionContext& ctx) const override {
     const float epsilon = ctx.Attr<float>("epsilon");
-    auto *scale = ctx.Input<Tensor>("Scale");
-    auto *bias = ctx.Input<Tensor>("Bias");
+    auto* scale = ctx.Input<Tensor>("Scale");
+    auto* bias = ctx.Input<Tensor>("Bias");
     auto x = *ctx.Input<Tensor>("X");
 
-    auto *y = ctx.Output<Tensor>("Y");
-    auto *mean = ctx.Output<Tensor>("Mean");
-    auto *var = ctx.Output<Tensor>("Variance");
+    auto* y = ctx.Output<Tensor>("Y");
+    auto* mean = ctx.Output<Tensor>("Mean");
+    auto* var = ctx.Output<Tensor>("Variance");
     const auto begin_norm_axis = ctx.Attr<int>("begin_norm_axis");
 
     const auto x_dims = x.dims();
@@ -94,8 +191,8 @@ class LayerNormKernel : public framework::OpKernel<T> {
     out.ShareDataWith(*y);
     out.Resize(matrix_shape);
 
-    auto &dev_ctx = ctx.template device_context<DeviceContext>();
-    math::RowwiseMean<DeviceContext, T> row_mean;
+    auto& dev_ctx = ctx.template device_context<DeviceContext>();
+    RowwiseMean2D<DeviceContext, T> row_mean(left, right, ctx.device_context());
 
     // get mean
     row_mean(dev_ctx, x, mean);
@@ -126,31 +223,32 @@ class LayerNormKernel : public framework::OpKernel<T> {
 template <typename DeviceContext, typename T>
 class LayerNormGradKernel : public framework::OpKernel<T> {
  public:
-  void Compute(const framework::ExecutionContext &ctx) const override {
+  void Compute(const framework::ExecutionContext& ctx) const override {
     const float epsilon = ctx.Attr<float>("epsilon");
     auto x = *ctx.Input<Tensor>("X");
-    auto *y = ctx.Input<Tensor>("Y");
-    auto *mean = ctx.Input<Tensor>("Mean");
-    auto *var = ctx.Input<Tensor>("Variance");
-    auto *scale = ctx.Input<Tensor>("Scale");
-    auto *bias = ctx.Input<Tensor>("Bias");
+    auto* y = ctx.Input<Tensor>("Y");
+    auto* mean = ctx.Input<Tensor>("Mean");
+    auto* var = ctx.Input<Tensor>("Variance");
+    auto* scale = ctx.Input<Tensor>("Scale");
+    auto* bias = ctx.Input<Tensor>("Bias");
     auto d_y = *ctx.Input<Tensor>(framework::GradVarName("Y"));
     const auto begin_norm_axis = ctx.Attr<int>("begin_norm_axis");
 
     // init output
-    auto *d_x = ctx.Output<Tensor>(framework::GradVarName("X"));
-    auto *d_scale = ctx.Output<Tensor>(framework::GradVarName("Scale"));
-    auto *d_bias = ctx.Output<Tensor>(framework::GradVarName("Bias"));
+    auto* d_x = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto* d_scale = ctx.Output<Tensor>(framework::GradVarName("Scale"));
+    auto* d_bias = ctx.Output<Tensor>(framework::GradVarName("Bias"));
 
-    const auto &x_dims = x.dims();
+    const auto& x_dims = x.dims();
     auto matrix_dim = framework::flatten_to_2d(x_dims, begin_norm_axis);
     int left = static_cast<int>(matrix_dim[0]);
     int right = static_cast<int>(matrix_dim[1]);
     framework::DDim matrix_shape({left, right});
 
     d_y.Resize(matrix_shape);
-    auto &dev_ctx = ctx.template device_context<DeviceContext>();
-    math::ColwiseSum<DeviceContext, T> colwise_sum;
+    auto& dev_ctx = ctx.template device_context<DeviceContext>();
+    ColwiseSum2D<DeviceContext, T> colwise_sum(left, right,
+                                               ctx.device_context());
 
     Tensor temp;
     Tensor temp_norm;
@@ -190,7 +288,8 @@ class LayerNormGradKernel : public framework::OpKernel<T> {
       Tensor temp_vec;
       temp_vec.mutable_data<T>(vec_shape, ctx.GetPlace());
 
-      math::RowwiseMean<DeviceContext, T> row_mean;
+      RowwiseMean2D<DeviceContext, T> row_mean(left, right,
+                                               ctx.device_context());
 
       if (d_scale) {
         // dy_dx

paddle/fluid/operators/listen_and_serv_op.cc

@@ -24,6 +24,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/proto_desc.h"
+#include "paddle/fluid/framework/threadpool.h"
 #include "paddle/fluid/operators/detail/grpc_server.h"
 #include "paddle/fluid/operators/detail/sendrecvop_utils.h"
 #include "paddle/fluid/operators/detail/simple_block_queue.h"
@@ -68,9 +69,7 @@ class ListenAndServOp : public framework::OperatorBase {
   }
 
   void Stop() override {
-    detail::MessageWithName term_msg;
-    term_msg.first = LISTEN_TERMINATE_MESSAGE;
-    rpc_service_->Push(term_msg);
+    rpc_service_->Push(LISTEN_TERMINATE_MESSAGE);
     rpc_service_->ShutDown();
     server_thread_->join();
   }
@@ -89,6 +88,10 @@ class ListenAndServOp : public framework::OperatorBase {
 
     auto *block = Attr<framework::BlockDesc *>(kOptimizeBlock);
     auto *program = block->Program();
+    int num_blocks = program->Size();
+    PADDLE_ENFORCE_GE(num_blocks, 2,
+                      "server program should have at least 2 blocks");
+
     framework::Executor executor(dev_place);
 
     // TODO(typhoonzero): change this to a while_op for every cluster-batch.
@@ -103,7 +106,7 @@ class ListenAndServOp : public framework::OperatorBase {
       size_t recv_var_cnt = 0;
       int batch_barrier = 0;
       while (batch_barrier != fan_in) {
-        const detail::MessageWithName &v = rpc_service_->Get();
+        const detail::ReceivedMessage v = rpc_service_->Get();
         auto recv_var_name = v.first;
         if (recv_var_name == LISTEN_TERMINATE_MESSAGE) {
           LOG(INFO) << "received terminate message and exit";
@@ -116,12 +119,11 @@ class ListenAndServOp : public framework::OperatorBase {
         } else {
           VLOG(3) << "received grad: " << recv_var_name;
           recv_var_cnt++;
-          auto *var = recv_scope.FindVar(recv_var_name);
+          auto var = v.second->GetVar();
           if (var == nullptr) {
             LOG(ERROR) << "Can not find server side var: " << recv_var_name;
             PADDLE_THROW("Can not find server side var");
           }
-          detail::DeserializeFromMessage(v.second, dev_ctx, var);
           if (var->IsType<framework::SelectedRows>()) {
             sparse_vars.push_back(var);
           }
@@ -132,12 +134,35 @@ class ListenAndServOp : public framework::OperatorBase {
         rpc_service_->ShutDown();
         break;
       }
+
+      // put optimize blocks in the thread pool to start run, the last block
+      // should be global ops.
+      // NOTE: if is_gpu_place, CUDA kernels are laugched by multiple threads
+      // and this will still work.
+
+      std::vector<std::future<void>> fs;
+      // block0 contains only listen_and_serv op, start run from block1.
+      for (int blkid = 1; blkid < num_blocks - 1; ++blkid) {
+        fs.push_back(
+            framework::Async([&executor, &program, &recv_scope, blkid]() {
+              int run_block = blkid;  // thread local
+              try {
+                executor.Run(*program, &recv_scope, run_block, false, false);
+              } catch (std::exception &e) {
+                LOG(ERROR) << "run sub program error " << e.what();
+              }
+            }));
+      }
+      for (int i = 0; i < num_blocks - 2; ++i) fs[i].wait();
+      // Run global block at final step, or block1 if there are only 2 blocks
+      if (num_blocks >= 2) {
         try {
-        executor.Run(*program, &recv_scope, block->ID(), /*global_block*/
-                     false /*create_local_scope*/, false /*create_vars*/);
+          executor.Run(*program, &recv_scope, num_blocks - 1, false, false);
         } catch (std::exception &e) {
           LOG(ERROR) << "run sub program error " << e.what();
         }
+      }
+
       // Reset the received sparse variables, the sum operator would not
       // sum the input sparse variables which rows is empty at the next
       // mini-batch.
@@ -151,6 +176,10 @@ class ListenAndServOp : public framework::OperatorBase {
       rpc_service_->WaitClientGet(fan_in);
       sparse_vars.clear();
     }  // while(true)
+
+    // for (int i = 0; i < num_blocks; ++i) {
+    //   delete blk_ctx_list[i];
+    // }
   }
 
  protected:

paddle/fluid/operators/lrn_mkldnn_op.cc

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
+
+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/tensor.h"
+#include "paddle/fluid/operators/lrn_op.h"
+#include "paddle/fluid/platform/mkldnn_helper.h"
+
+namespace paddle {
+namespace operators {
+
+using paddle::framework::Tensor;
+using paddle::platform::MKLDNNDeviceContext;
+
+namespace {
+template <typename T, typename... Args>
+std::shared_ptr<T> insert_to_context(const std::string& key,
+                                     const MKLDNNDeviceContext& dev_ctx,
+                                     Args&&... args) {
+  auto p = std::static_pointer_cast<T, void>(dev_ctx.GetBlob(key));
+
+  if (!p) {
+    p = std::make_shared<T>(args...);
+    dev_ctx.SetBlob(key, std::static_pointer_cast<void, T>(p));
+  }
+
+  return p;
+}
+
+template <typename... Args>
+void run_primitive(Args&&... args) {
+  auto forward_op = mkldnn::lrn_forward{args...};
+
+  std::vector<mkldnn::primitive> pipeline = {forward_op};
+  mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
+}
+}  // namespace
+
+template <typename T>
+class LRNMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
+ public:
+  void Compute(const paddle::framework::ExecutionContext& ctx) const override {
+    PADDLE_ENFORCE(std::is_same<T, float>::value,
+                   "MKLDNN LRN must use float data.");
+    PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()),
+                   "MKLDNN LRN must use CPUPlace.");
+
+    auto& dev_ctx = ctx.template device_context<MKLDNNDeviceContext>();
+    const auto& mkldnn_engine = dev_ctx.GetEngine();
+
+    auto x = ctx.Input<Tensor>("X");
+    auto out = ctx.Output<Tensor>("Out");
+    auto mid = ctx.Output<Tensor>("MidOut");
+
+    auto input_data = x->data<T>();
+    auto output_data = out->mutable_data<T>(ctx.GetPlace());
+    mid->mutable_data<T>(ctx.GetPlace());
+
+    const int n = ctx.Attr<int>("n");
+    const float alpha = ctx.Attr<float>("alpha");
+    const float beta = ctx.Attr<float>("beta");
+    const float k = ctx.Attr<float>("k");
+    const bool is_test = ctx.Attr<bool>("is_test");
+
+    auto e_mid = framework::EigenTensor<T, 4>::From(*mid);
+    e_mid = e_mid.constant(k);
+
+    auto dims = paddle::framework::vectorize2int(x->dims());
+
+    auto src_md = paddle::platform::MKLDNNMemDesc(
+        dims, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw);
+
+    auto dst_md = paddle::platform::MKLDNNMemDesc(
+        dims, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw);
+
+    auto forward_desc = mkldnn::lrn_forward::desc{mkldnn::prop_kind::forward,
+                                                  mkldnn::lrn_across_channels,
+                                                  src_md,
+                                                  n,
+                                                  alpha,
+                                                  beta,
+                                                  k};
+
+    auto src_memory_pd = mkldnn::memory::primitive_desc{src_md, mkldnn_engine};
+    auto dst_memory = mkldnn::memory{{dst_md, mkldnn_engine},
+                                     static_cast<void*>(output_data)};
+
+    if (!is_test) {
+      const std::string key = ctx.op().Output("Out");
+      const std::string key_src_memory = key + "@lrn_src_memory";
+      const std::string key_pd = key + "@lrn_pd";
+      const std::string key_workspace_memory = key + "@lrn_workspace_memory";
+
+      auto forward_pd = insert_to_context<mkldnn::lrn_forward::primitive_desc>(
+          key_pd, dev_ctx, forward_desc, mkldnn_engine);
+
+      auto src_memory = insert_to_context<mkldnn::memory>(
+          key_src_memory, dev_ctx, src_memory_pd);
+
+      src_memory->set_data_handle(
+          static_cast<void*>(const_cast<T*>(input_data)));
+
+      auto workspace_memory = insert_to_context<mkldnn::memory>(
+          key_workspace_memory, dev_ctx,
+          forward_pd->workspace_primitive_desc());
+
+      run_primitive(*forward_pd, *src_memory, *workspace_memory, dst_memory);
+    } else {
+      auto forward_pd =
+          mkldnn::lrn_forward::primitive_desc{forward_desc, mkldnn_engine};
+      auto src_memory = mkldnn::memory{
+          src_memory_pd, static_cast<void*>(const_cast<T*>(input_data))};
+      auto workspace_memory =
+          mkldnn::memory{forward_pd.workspace_primitive_desc()};
+
+      run_primitive(forward_pd, src_memory, workspace_memory, dst_memory);
+    }
+  }
+};
+
+template <typename T>
+class LRNMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
+ public:
+  void Compute(const paddle::framework::ExecutionContext& ctx) const override {
+    PADDLE_ENFORCE(std::is_same<T, float>::value,
+                   "MKLDNN LRN must use float data.");
+    PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()),
+                   "MKLDNN LRN must use CPUPlace.");
+    PADDLE_ENFORCE(
+        !ctx.Attr<bool>("is_test"),
+        "is_test attribute should be set to False in training phase.");
+
+    auto x = ctx.Input<Tensor>("X");
+
+    auto out_grad = ctx.Input<Tensor>(framework::GradVarName("Out"));
+    auto x_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
+
+    const std::string key = ctx.op().Input("Out");
+    const std::string key_src_memory = key + "@lrn_src_memory";
+    const std::string key_pd = key + "@lrn_pd";
+    const std::string key_workspace_memory = key + "@lrn_workspace_memory";
+
+    const int n = ctx.Attr<int>("n");
+    const float alpha = ctx.Attr<float>("alpha");
+    const float beta = ctx.Attr<float>("beta");
+    const float k = ctx.Attr<float>("k");
+
+    auto& dev_ctx = ctx.template device_context<MKLDNNDeviceContext>();
+    const auto& mkldnn_engine = dev_ctx.GetEngine();
+
+    auto x_grad_data = x_grad->mutable_data<T>(ctx.GetPlace());
+    auto out_grad_data = out_grad->data<T>();
+
+    auto dims = paddle::framework::vectorize2int(x->dims());
+
+    auto src_md = paddle::platform::MKLDNNMemDesc(
+        dims, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw);
+
+    auto diff_src_md = paddle::platform::MKLDNNMemDesc(
+        dims, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw);
+
+    auto diff_dst_md = paddle::platform::MKLDNNMemDesc(
+        dims, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw);
+
+    auto diff_dst_memory =
+        mkldnn::memory{{diff_dst_md, mkldnn_engine},
+                       static_cast<void*>(const_cast<float*>(out_grad_data))};
+
+    auto diff_src_memory = mkldnn::memory{{diff_src_md, mkldnn_engine},
+                                          static_cast<void*>(x_grad_data)};
+
+    auto backward_desc = mkldnn::lrn_backward::desc{
+        mkldnn::lrn_across_channels, src_md, diff_src_md, n, alpha, beta, k};
+
+    auto forward_pd = dev_ctx.GetBlob(key_pd);
+
+    auto backward_pd = mkldnn::lrn_backward::primitive_desc{
+        backward_desc, mkldnn_engine,
+        *static_cast<mkldnn::lrn_forward::primitive_desc*>(forward_pd.get())};
+
+    std::shared_ptr<void> workspace_memory =
+        dev_ctx.GetBlob(key_workspace_memory);
+
+    auto src_memory = dev_ctx.GetBlob(key_src_memory);
+    auto backward_op = mkldnn::lrn_backward{
+        backward_pd, *static_cast<mkldnn::memory*>(src_memory.get()),
+        diff_dst_memory, *static_cast<mkldnn::memory*>(workspace_memory.get()),
+        diff_src_memory};
+
+    std::vector<mkldnn::primitive> pipeline = {backward_op};
+    mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+REGISTER_OP_KERNEL(lrn, MKLDNN, paddle::platform::CPUPlace,
+                   ops::LRNMKLDNNOpKernel<float>);
+REGISTER_OP_KERNEL(lrn_grad, MKLDNN, paddle::platform::CPUPlace,
+                   ops::LRNMKLDNNGradOpKernel<float>);

paddle/fluid/operators/lrn_op.cc

@@ -13,6 +13,9 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/lrn_op.h"
+#ifdef PADDLE_WITH_MKLDNN
+#include "paddle/fluid/platform/mkldnn_helper.h"
+#endif
 
 namespace paddle {
 namespace operators {
@@ -116,6 +119,26 @@ struct LRNGradFunctor<platform::CPUDeviceContext, T> {
 template struct LRNGradFunctor<platform::CPUDeviceContext, float>;
 template struct LRNGradFunctor<platform::CPUDeviceContext, double>;
 
+namespace {
+framework::OpKernelType GetExpectedLRNKernel(
+    const framework::ExecutionContext& ctx) {
+  framework::LibraryType library_{framework::LibraryType::kPlain};
+#ifdef PADDLE_WITH_MKLDNN
+  if (library_ == framework::LibraryType::kPlain &&
+      platform::CanMKLDNNBeUsed(ctx)) {
+    library_ = framework::LibraryType::kMKLDNN;
+  }
+#endif
+
+  std::string data_format = ctx.Attr<std::string>("data_format");
+  // TODO(pzelazko-intel): enable MKLDNN layout when it's ready
+  framework::DataLayout layout_ = framework::StringToDataLayout(data_format);
+  return framework::OpKernelType(
+      framework::ToDataType(ctx.Input<Tensor>("X")->type()), ctx.GetPlace(),
+      layout_, library_);
+}
+}  // namespace
+
 class LRNOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
@@ -132,8 +155,13 @@ class LRNOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(x_dim.size(), 4, "Input(X)'rank of LRNOp should be 4.");
 
     ctx->SetOutputDim("Out", x_dim);
-    ctx->SetOutputDim("MidOut", x_dim);
     ctx->ShareLoD("X", /*->*/ "Out");
+    ctx->SetOutputDim("MidOut", x_dim);
+  }
+
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return GetExpectedLRNKernel(ctx);
   }
 };
 
@@ -176,6 +204,17 @@ class LRNOpMaker : public framework::OpProtoAndCheckerMaker {
                "beta is the power number.")
         .SetDefault(0.75)
         .GreaterThan(0.0);
+    AddAttr<bool>("use_mkldnn",
+                  "(bool, default false) Only used in mkldnn kernel")
+        .SetDefault(false);
+    AddAttr<std::string>(
+        "data_format",
+        "(string, default NCHW) Only used in "
+        "An optional string from: \"NHWC\", \"NCHW\". "
+        "Defaults to \"NHWC\". Specify the data format of the output data, "
+        "the input will be transformed automatically. ")
+        .SetDefault("AnyLayout");
+    AddAttr<bool>("is_test", "").SetDefault(false);
 
     AddComment(R"DOC(
 Local Response Normalization Operator.
@@ -223,8 +262,12 @@ class LRNOpGrad : public framework::OperatorWithKernel {
     auto x_dims = ctx->GetInputDim("X");
     ctx->SetOutputDim(framework::GradVarName("X"), x_dims);
   }
-};
 
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return GetExpectedLRNKernel(ctx);
+  }
+};
 }  // namespace operators
 }  // namespace paddle
 

paddle/fluid/operators/math/CMakeLists.txt

@@ -6,6 +6,7 @@ function(math_library TARGET)
     # But it handle split GPU/CPU code and link some common library.
     set(cc_srcs)
     set(cu_srcs)
+    set(hip_srcs)
     set(math_common_deps device_context framework_proto)
     set(multiValueArgs DEPS)
     cmake_parse_arguments(math_library "${options}" "${oneValueArgs}"
@@ -17,10 +18,15 @@ function(math_library TARGET)
     if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cu)
         list(APPEND cu_srcs ${TARGET}.cu)
     endif()
+    if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.hip.cu)
+        list(APPEND hip_srcs ${TARGET}.hip.cu)
+    endif()
 
     list(LENGTH cc_srcs cc_srcs_len)
     if (WITH_GPU)
         nv_library(${TARGET} SRCS ${cc_srcs} ${cu_srcs} DEPS ${math_library_DEPS} ${math_common_deps})
+    elseif (WITH_AMD_GPU)
+        hip_library(${TARGET} SRCS ${cc_srcs} ${hip_srcs} DEPS ${math_library_DEPS} ${math_common_deps})
     elseif(${cc_srcs_len} GREATER 0)
         cc_library(${TARGET} SRCS ${cc_srcs} DEPS ${math_library_DEPS} ${math_common_deps})
     endif()
@@ -43,7 +49,7 @@ math_library(sequence2batch)
 math_library(sequence_padding)
 math_library(sequence_pooling DEPS math_function)
 math_library(sequence_scale)
-math_library(softmax)
+math_library(softmax DEPS math_function)
 math_library(unpooling)
 math_library(vol2col)
 

paddle/fluid/operators/math/concat.cc

@@ -20,7 +20,7 @@ namespace math {
 
 /*
  * All tensors' dimension should be the same and the values of
- * each dimension are the same, except the axis dimension.
+ * each dimension must be the same, except the axis dimension.
  */
 template <typename T>
 class ConcatFunctor<platform::CPUDeviceContext, T> {
@@ -63,7 +63,7 @@ class ConcatFunctor<platform::CPUDeviceContext, T> {
 
 /*
  * All tensors' dimension should be the same and the values of
- * each dimension are the same, except the axis dimension.
+ * each dimension must be the same, except the axis dimension.
  */
 template <typename T>
 class ConcatGradFunctor<platform::CPUDeviceContext, T> {

paddle/fluid/operators/math/concat.cu

@@ -66,68 +66,66 @@ __global__ void KernelConcat(T** inputs, const int* input_cols, int col_size,
 }
 
 template <typename T>
-__global__ void KernelConcat(T** inputs, const int input_col,
-                             const int output_rows, const int output_cols,
-                             T* output) {
+__global__ void KernelConcat(T** inputs_data, const int fixed_in_col,
+                             const int out_rows, const int out_cols,
+                             T* output_data) {
   int tid_x = blockIdx.x * blockDim.x + threadIdx.x;
-  double inv_input_col = 1.0 / input_col;
-  for (; tid_x < output_cols; tid_x += blockDim.x * gridDim.x) {
-    int split = tid_x * inv_input_col;
-    int in_offset = tid_x - split * input_col;
-    T* input_ptr = inputs[split];
+  for (; tid_x < out_cols; tid_x += blockDim.x * gridDim.x) {
+    int split = tid_x * 1.0 / fixed_in_col;
+    int in_offset = tid_x - split * fixed_in_col;
+    T* input_ptr = inputs_data[split];
     int tid_y = blockIdx.y * blockDim.y + threadIdx.y;
-    for (; tid_y < output_rows; tid_y += blockDim.y * gridDim.y) {
-      output[tid_y * output_cols + tid_x] =
-          input_ptr[tid_y * input_col + in_offset];
+    for (; tid_y < out_rows; tid_y += blockDim.y * gridDim.y) {
+      output_data[tid_y * out_cols + tid_x] =
+          input_ptr[tid_y * fixed_in_col + in_offset];
     }
   }
 }
 
 template <typename T>
-__global__ void KernelConcatGrad(const T* input, const int input_row,
-                                 const int input_col, const int* output_cols,
-                                 int col_size, T** outputs) {
+__global__ void KernelConcatGrad(const T* input_data, const int in_row,
+                                 const int in_col, const int* out_cols,
+                                 int out_cols_size, T** outputs_data) {
   int tid_x = blockIdx.x * blockDim.x + threadIdx.x;
-  int segment = upper_bound<int>(output_cols, col_size, tid_x) - 1;
-  int curr_offset = output_cols[segment];
+  int segment = upper_bound<int>(out_cols, out_cols_size, tid_x) - 1;
+  int curr_offset = out_cols[segment];
   int curr_segment = segment;
-  for (; tid_x < input_col; tid_x += blockDim.x * gridDim.x) {
+  for (; tid_x < in_col; tid_x += blockDim.x * gridDim.x) {
     T curr_col_offset;
-    while ((curr_col_offset = output_cols[curr_segment + 1]) <= tid_x) {
+    while ((curr_col_offset = out_cols[curr_segment + 1]) <= tid_x) {
       curr_offset = curr_col_offset;
       ++curr_segment;
     }
 
     int local_col = tid_x - curr_offset;
     int segment_width = curr_col_offset - curr_offset;
-    T* output_ptr = outputs[curr_segment];
+    T* output_ptr = outputs_data[curr_segment];
     int tid_y = blockIdx.y * blockDim.y + threadIdx.y;
-    for (; tid_y < input_row; tid_y += blockDim.y * gridDim.y)
+    for (; tid_y < in_row; tid_y += blockDim.y * gridDim.y)
       output_ptr[tid_y * segment_width + local_col] =
-          input[tid_y * input_col + tid_x];
+          input_data[tid_y * in_col + tid_x];
   }
 }
 
 template <typename T>
-__global__ void KernelConcatGrad(const T* input, const int input_row,
-                                 const int input_col, const int output_cols,
-                                 T** outputs) {
+__global__ void KernelConcatGrad(const T* input_data, const int in_row,
+                                 const int in_col, const int fixed_out_col,
+                                 T** outputs_data) {
   int tid_x = blockIdx.x * blockDim.x + threadIdx.x;
-  double inv_input_col = 1.0 / input_col;
-  for (; tid_x < input_col; tid_x += blockDim.x * gridDim.x) {
-    int split = tid_x * inv_input_col;
-    int in_offset = tid_x - split * input_col;
-    T* output_ptr = outputs[split];
+  for (; tid_x < in_col; tid_x += blockDim.x * gridDim.x) {
+    int split = tid_x / fixed_out_col;
+    int in_offset = tid_x - split * fixed_out_col;
+    T* output_ptr = outputs_data[split];
     int tid_y = blockIdx.y * blockDim.y + threadIdx.y;
-    for (; tid_y < input_row; tid_y += blockDim.y * gridDim.y)
-      output_ptr[tid_y * output_cols + in_offset] =
-          input[tid_y * input_col + tid_x];
+    for (; tid_y < in_row; tid_y += blockDim.y * gridDim.y)
+      output_ptr[tid_y * fixed_out_col + in_offset] =
+          input_data[tid_y * in_col + tid_x];
   }
 }
 
 /*
  * All tensors' dimension should be the same and the values of
- * each dimension are the same, except the axis dimension.
+ * each dimension must be the same, except the axis dimension.
  */
 template <typename T>
 class ConcatFunctor<platform::CUDADeviceContext, T> {
@@ -136,41 +134,40 @@ class ConcatFunctor<platform::CUDADeviceContext, T> {
                   const std::vector<framework::Tensor>& input, const int axis,
                   framework::Tensor* output) {
     // TODO(zcd): Add input data validity checking
-    int num = input.size();
-    int rows = 1;
+    int in_num = input.size();
+    int in_row = 1;
     auto dim_0 = input[0].dims();
     for (int i = 0; i < axis; ++i) {
-      rows *= dim_0[i];
+      in_row *= dim_0[i];
     }
-    int cols = input[0].numel() / rows;
-    int out_rows = rows, out_cols = 0;
+    int in_col = input[0].numel() / in_row;
+    int out_row = in_row, out_col = 0;
 
-    framework::Vector<int16_t> inputs_data(num * sizeof(T*) / 2);
-    framework::Vector<int> inputs_cols(num + 1);
-    inputs_cols[0] = 0;
+    framework::Vector<int16_t> inputs_data(in_num * sizeof(T*) / 2);
+    framework::Vector<int> inputs_col(in_num + 1);
     T** inputs_ptr = reinterpret_cast<T**>(inputs_data.data());
 
+    inputs_col[0] = 0;
     bool sameShape = true;
-    for (int i = 0; i < num; ++i) {
-      int t_cols = input[i].numel() / rows;
+    for (int i = 0; i < in_num; ++i) {
+      int t_cols = input[i].numel() / in_row;
       if (sameShape) {
-        if (t_cols != cols) sameShape = false;
+        if (t_cols != in_col) sameShape = false;
       }
-      out_cols += t_cols;
-      inputs_cols[i + 1] = out_cols;
+      out_col += t_cols;
+      inputs_col[i + 1] = out_col;
       inputs_ptr[i] = const_cast<T*>(input[i].data<T>());
     }
 
-    T** ins_gpu =
+    T** dev_ins_data =
         reinterpret_cast<T**>(inputs_data.CUDAMutableData(context.GetPlace()));
-    const int* ins_col_gpu = inputs_cols.CUDAData(context.GetPlace());
 
     // computation
     // set the thread block and grid according to CurrentDeviceId
     const int kThreadsPerBlock = 1024;
     int block_cols = kThreadsPerBlock;
-    if (out_cols < kThreadsPerBlock) {  // block_cols is aligned by 32.
-      block_cols = ((out_cols + 31) >> 5) << 5;
+    if (out_col < kThreadsPerBlock) {  // block_cols is aligned by 32.
+      block_cols = ((out_col + 31) >> 5) << 5;
     }
     int block_rows = kThreadsPerBlock / block_cols;
     dim3 block_size = dim3(block_cols, block_rows, 1);
@@ -179,25 +176,26 @@ class ConcatFunctor<platform::CUDADeviceContext, T> {
     int max_blocks = std::max(max_threads / kThreadsPerBlock, 1);
 
     int grid_cols =
-        std::min((out_cols + block_cols - 1) / block_cols, max_blocks);
+        std::min((out_col + block_cols - 1) / block_cols, max_blocks);
     int grid_rows =
-        std::min(max_blocks / grid_cols, std::max(out_rows / block_rows, 1));
+        std::min(max_blocks / grid_cols, std::max(out_row / block_rows, 1));
     dim3 grid_size = dim3(grid_cols, grid_rows, 1);
 
     if (sameShape) {
       KernelConcat<<<grid_size, block_size, 0, context.stream()>>>(
-          ins_gpu, cols, out_rows, out_cols, output->data<T>());
+          dev_ins_data, in_col, out_row, out_col, output->data<T>());
     } else {
+      const int* dev_ins_col_data = inputs_col.CUDAData(context.GetPlace());
       KernelConcat<<<grid_size, block_size, 0, context.stream()>>>(
-          ins_gpu, ins_col_gpu, static_cast<int>(inputs_cols.size()), out_rows,
-          out_cols, output->data<T>());
+          dev_ins_data, dev_ins_col_data, static_cast<int>(inputs_col.size()),
+          out_row, out_col, output->data<T>());
     }
   }
 };
 
 /*
  * All tensors' dimension should be the same and the values of
- * each dimension are the same, except the axis dimension.
+ * each dimension must be the same, except the axis dimension.
  */
 template <typename T>
 class ConcatGradFunctor<platform::CUDADeviceContext, T> {
@@ -206,41 +204,40 @@ class ConcatGradFunctor<platform::CUDADeviceContext, T> {
                   const framework::Tensor& input, const int axis,
                   std::vector<framework::Tensor>& outputs) {
     // TODO(zcd): Add input data validity checking
-    int num = outputs.size();
-    int input_row = 1;
+    int o_num = outputs.size();
+    int out_row = 1;
     auto dim_0 = outputs[0].dims();
     for (int i = 0; i < axis; ++i) {
-      input_row *= dim_0[i];
+      out_row *= dim_0[i];
     }
 
-    int output_col_0 = outputs[0].numel() / input_row;
-    int input_col = 0;
+    int out_col = outputs[0].numel() / out_row;
+    int in_col = 0, in_row = out_row;
     bool sameShape = true;
 
-    framework::Vector<int16_t> outputs_data(num * sizeof(T*) / 2);
-    framework::Vector<int> outputs_cols(num + 1);
-    outputs_cols[0] = 0;
+    framework::Vector<int16_t> outputs_data(o_num * sizeof(T*) / 2);
+    framework::Vector<int> outputs_cols(o_num + 1);
     T** outputs_ptr = reinterpret_cast<T**>(outputs_data.data());
 
-    for (int i = 0; i < num; ++i) {
-      int t_col = outputs[i].numel() / input_row;
+    outputs_cols[0] = 0;
+    for (int i = 0; i < o_num; ++i) {
+      int t_col = outputs[i].numel() / out_row;
       if (sameShape) {
-        if (t_col != output_col_0) sameShape = false;
+        if (t_col != out_col) sameShape = false;
       }
-      input_col += t_col;
-      outputs_cols[i + 1] = input_col;
+      in_col += t_col;
+      outputs_cols[i + 1] = in_col;
       outputs_ptr[i] = outputs[i].data<T>();
     }
 
-    T** outs_gpu =
+    T** dev_out_gpu_data =
         reinterpret_cast<T**>(outputs_data.CUDAMutableData(context.GetPlace()));
-    const int* outs_col_gpu = outputs_cols.CUDAData(context.GetPlace());
 
     // computation
     const int kThreadsPerBlock = 1024;
     int block_cols = kThreadsPerBlock;
-    if (input_col < kThreadsPerBlock) {  // block_cols is aligned by 32.
-      block_cols = ((input_col + 31) >> 5) << 5;
+    if (in_col < kThreadsPerBlock) {  // block_cols is aligned by 32.
+      block_cols = ((in_col + 31) >> 5) << 5;
     }
     int block_rows = kThreadsPerBlock / block_cols;
     dim3 block_size = dim3(block_cols, block_rows, 1);
@@ -249,18 +246,19 @@ class ConcatGradFunctor<platform::CUDADeviceContext, T> {
     int max_blocks = std::max(max_threads / kThreadsPerBlock, 1);
 
     int grid_cols =
-        std::min((input_col + block_cols - 1) / block_cols, max_blocks);
+        std::min((in_col + block_cols - 1) / block_cols, max_blocks);
     int grid_rows =
-        std::min(max_blocks / grid_cols, std::max(input_row / block_rows, 1));
+        std::min(max_blocks / grid_cols, std::max(out_row / block_rows, 1));
     dim3 grid_size = dim3(grid_cols, grid_rows, 1);
 
     if (sameShape) {
       KernelConcatGrad<<<grid_size, block_size, 0, context.stream()>>>(
-          input.data<T>(), input_row, input_col, output_col_0, outs_gpu);
+          input.data<T>(), in_row, in_col, out_col, dev_out_gpu_data);
     } else {
+      const int* dev_outs_col_data = outputs_cols.CUDAData(context.GetPlace());
       KernelConcatGrad<<<grid_size, block_size, 0, context.stream()>>>(
-          input.data<T>(), input_row, input_col, outs_col_gpu,
-          static_cast<int>(outputs_cols.size()), outs_gpu);
+          input.data<T>(), in_row, in_col, dev_outs_col_data,
+          static_cast<int>(outputs_cols.size()), dev_out_gpu_data);
     }
   }
 };

paddle/fluid/operators/math/concat.hip.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 <hip/hip_runtime.h>

paddle/fluid/operators/math/math_function.cc

@@ -278,6 +278,7 @@ void axpy<platform::CPUDeviceContext, double>(
   cblas_daxpy(n, alpha, x, 1, y, 1);
 }
 
+template struct SetConstant<platform::CPUDeviceContext, platform::float16>;
 template struct SetConstant<platform::CPUDeviceContext, float>;
 template struct SetConstant<platform::CPUDeviceContext, double>;
 template struct SetConstant<platform::CPUDeviceContext, int>;
@@ -371,6 +372,8 @@ template struct RowwiseAdd<platform::CPUDeviceContext, double>;
 
 template struct ColwiseSum<platform::CPUDeviceContext, float>;
 template struct ColwiseSum<platform::CPUDeviceContext, double>;
+template struct ColwiseSum<platform::CPUDeviceContext, int>;
+template struct ColwiseSum<platform::CPUDeviceContext, int64_t>;
 
 template struct RowwiseSum<platform::CPUDeviceContext, float>;
 template struct RowwiseSum<platform::CPUDeviceContext, double>;

paddle/fluid/operators/math/math_function.cu

@@ -348,6 +348,7 @@ void axpy<platform::CUDADeviceContext, double>(
                                                 &alpha, x, 1, y, 1));
 }
 
+template struct SetConstant<platform::CUDADeviceContext, platform::float16>;
 template struct SetConstant<platform::CUDADeviceContext, float>;
 template struct SetConstant<platform::CUDADeviceContext, double>;
 template struct SetConstant<platform::CUDADeviceContext, int>;
@@ -422,6 +423,8 @@ struct RowwiseAdd<platform::CUDADeviceContext, T> {
 template struct RowwiseAdd<platform::CUDADeviceContext, float>;
 template struct RowwiseAdd<platform::CUDADeviceContext, double>;
 template struct ColwiseSum<platform::CUDADeviceContext, float>;
+template struct ColwiseSum<platform::CUDADeviceContext, int>;
+template struct ColwiseSum<platform::CUDADeviceContext, int64_t>;
 // template struct ColwiseSum<platform::CUDADeviceContext, double>;
 // The ColwiseSum<platform::CUDADeviceContext, double> failed in debug mode,
 // and only failed for this case. So reimplemented it.

paddle/fluid/operators/math/math_function.h

@@ -19,13 +19,6 @@ limitations under the License. */
 #include <mkl_vml_functions.h>
 #endif
 
-#ifdef PADDLE_USE_ATLAS
-extern "C" {
-#include <cblas.h>
-#include <clapack.h>
-}
-#endif
-
 #ifdef PADDLE_USE_OPENBLAS
 #include <cblas.h>
 #include <lapacke.h>

paddle/fluid/operators/math/softmax.cu

@@ -89,6 +89,7 @@ void SoftmaxGradCUDNNFunctor<T>::operator()(
       XGrad->mutable_data<T>(context.GetPlace())));
 }
 
+template class SoftmaxCUDNNFunctor<platform::float16>;
 template class SoftmaxCUDNNFunctor<float>;
 template class SoftmaxCUDNNFunctor<double>;
 template class SoftmaxGradCUDNNFunctor<float>;

paddle/fluid/operators/mine_hard_examples_op.cc

@@ -324,8 +324,9 @@ MatchIndices elements with value -1.
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_WITHOUT_GRADIENT(mine_hard_examples, ops::MineHardExamplesOp,
-                             ops::MineHardExamplesOpMaker);
+REGISTER_OPERATOR(mine_hard_examples, ops::MineHardExamplesOp,
+                  ops::MineHardExamplesOpMaker,
+                  paddle::framework::EmptyGradOpMaker);
 
 REGISTER_OP_CPU_KERNEL(
     mine_hard_examples,

paddle/fluid/operators/mkldnn_activation_op.h

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include "paddle/fluid/framework/eigen.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/detail/safe_ref.h"
+
+#ifdef PADDLE_WITH_MKLDNN
+#include "paddle/fluid/platform/mkldnn_helper.h"
+#endif
+
+namespace paddle {
+namespace operators {
+
+template <typename Functor>
+class MKLDNNActivationKernel
+    : public framework::OpKernel<typename Functor::ELEMENT_TYPE> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    PADDLE_ENFORCE(context.Input<framework::Tensor>("X") != nullptr,
+                   "Cannot get input tensor X, variable name = %s",
+                   context.op().Input("X"));
+    PADDLE_ENFORCE(context.Output<framework::Tensor>("Out") != nullptr,
+                   "Cannot find output tensor Out, variable name = %s",
+                   context.op().Output("Out"));
+    Functor functor;
+
+    auto attrs = functor.GetAttrs();
+    for (auto& attr : attrs) {
+      *attr.second = context.Attr<float>(attr.first);
+    }
+    functor(context);
+  }
+};
+
+template <typename Functor>
+class MKLDNNActivationGradKernel
+    : public framework::OpKernel<typename Functor::ELEMENT_TYPE> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    Functor functor;
+
+    auto attrs = functor.GetAttrs();
+    for (auto& attr : attrs) {
+      *attr.second = context.Attr<float>(attr.first);
+    }
+    functor(context);
+  }
+};
+
+namespace {
+framework::OpKernelType GetKernelType(
+    const framework::ExecutionContext& ctx,
+    const framework::OperatorWithKernel& oper) {
+  framework::LibraryType library{framework::LibraryType::kPlain};
+#ifdef PADDLE_WITH_MKLDNN
+  if (library == framework::LibraryType::kPlain &&
+      platform::CanMKLDNNBeUsed(ctx)) {
+    library = framework::LibraryType::kMKLDNN;
+  }
+#endif
+  framework::DataLayout layout = framework::DataLayout::kAnyLayout;
+  return framework::OpKernelType(
+      framework::ToDataType(ctx.Input<framework::Tensor>("X")->type()),
+      ctx.GetPlace(), layout, library);
+}
+}  // anonymous namespace
+
+class ActivationWithMKLDNNOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    ctx->SetOutputDim("Out", ctx->GetInputDim("X"));
+    ctx->ShareLoD("X", /*->*/ "Out");
+  }
+
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return GetKernelType(ctx, *this);
+  }
+};
+
+class ActivationWithMKLDNNOpGrad : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("Out"));
+  }
+
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return GetKernelType(ctx, *this);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/prior_box_op.cc

@@ -168,7 +168,9 @@ https://arxiv.org/abs/1512.02325.
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_WITHOUT_GRADIENT(prior_box, ops::PriorBoxOp, ops::PriorBoxOpMaker);
+REGISTER_OPERATOR(prior_box, ops::PriorBoxOp, ops::PriorBoxOpMaker,
+                  paddle::framework::EmptyGradOpMaker);
+
 REGISTER_OP_CPU_KERNEL(
     prior_box, ops::PriorBoxOpKernel<paddle::platform::CPUPlace, float>,
     ops::PriorBoxOpKernel<paddle::platform::CPUPlace, double>);

paddle/fluid/operators/reader/CMakeLists.txt

@@ -15,10 +15,12 @@ function(reader_library TARGET_NAME)
         PARENT_SCOPE)
 endfunction()
 
+reader_library(open_files_op SRCS open_files_op.cc)
 reader_library(create_random_data_generator_op SRCS create_random_data_generator_op.cc)
 reader_library(create_shuffle_reader_op SRCS create_shuffle_reader_op.cc)
 reader_library(create_batch_reader_op SRCS create_batch_reader_op.cc)
 reader_library(create_recordio_file_reader_op SRCS create_recordio_file_reader_op.cc)
 reader_library(create_double_buffer_reader_op SRCS create_double_buffer_reader_op.cc)
+reader_library(create_multi_pass_reader_op SRCS create_multi_pass_reader_op.cc)
 # Export local libraries to parent
 set(READER_LIBRARY ${LOCAL_READER_LIBS} PARENT_SCOPE)

paddle/fluid/operators/reader/create_double_buffer_reader_op.cc

@@ -48,20 +48,24 @@ class DoubleBufferReader : public framework::DecoratedReader {
 
   void start_thread() {
     buffer_ = framework::MakeChannel<Item>(kDoubleBufferSize);
-    std::thread prefetch([this] { PrefetchThreadFunc(); });
-    prefetch.detach();
+    prefetcher_ = std::thread([this] { PrefetchThreadFunc(); });
   }
 
   void ReadNext(std::vector<framework::LoDTensor>* out) override;
   void ReInit() override;
 
-  ~DoubleBufferReader() { buffer_->Close(); }
+  ~DoubleBufferReader() {
+    buffer_->Close();
+    prefetcher_.join();
+    delete buffer_;
+  }
 
   bool HasNext() const override;
 
  private:
   void PrefetchThreadFunc();
 
+  std::thread prefetcher_;
   framework::Channel<Item>* buffer_;
   platform::Place place_;
   std::vector<std::unique_ptr<platform::DeviceContext>> ctxs_;
@@ -120,10 +124,13 @@ class CreateDoubleBufferReaderOpMaker : public DecoratedReaderMakerBase {
 };
 
 void DoubleBufferReader::ReadNext(std::vector<framework::LoDTensor>* out) {
+  if (!HasNext()) {
+    PADDLE_THROW("There is no next data!");
+  }
+
   if (local_buffer_.payloads_.empty()) {
     buffer_->Receive(&local_buffer_);
   }
-
   *out = local_buffer_.payloads_;
   local_buffer_.payloads_.clear();
   if (local_buffer_.ctx_) {
@@ -134,6 +141,8 @@ void DoubleBufferReader::ReadNext(std::vector<framework::LoDTensor>* out) {
 void DoubleBufferReader::ReInit() {
   reader_->ReInit();
   buffer_->Close();
+  prefetcher_.join();
+  delete buffer_;
   start_thread();
 }
 
@@ -157,13 +166,16 @@ void DoubleBufferReader::PrefetchThreadFunc() {
       std::swap(gpu_batch, batch.payloads_);
     }
 
-    if (!buffer_->Send(&batch)) {
+    try {
+      buffer_->Send(&batch);
+    } catch (paddle::platform::EnforceNotMet e) {
       VLOG(5) << "WARNING: The double buffer channel has been closed. The "
-                 "prefetch thread terminates.";
+                 "prefetch thread will terminate.";
       break;
     }
   }
   buffer_->Close();
+  VLOG(5) << "Prefetch thread terminates.";
 }
 
 bool DoubleBufferReader::HasNext() const {

paddle/fluid/operators/reader/create_multi_pass_reader_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/detail/safe_ref.h"
+#include "paddle/fluid/operators/reader/reader_op_registry.h"
+
+namespace paddle {
+namespace operators {
+namespace reader {
+
+class MultiPassReader : public framework::DecoratedReader {
+ public:
+  MultiPassReader(ReaderBase* reader, int pass_num)
+      : DecoratedReader(reader), pass_num_(pass_num), pass_count_(0) {}
+
+  void ReadNext(std::vector<framework::LoDTensor>* out) override {
+    if (!HasNext()) {
+      PADDLE_THROW("There is no next data!");
+    }
+    reader_->ReadNext(out);
+  }
+
+  bool HasNext() const override {
+    if (reader_->HasNext()) {
+      return true;
+    } else {
+      ++pass_count_;
+      if (pass_count_ >= pass_num_) {
+        return false;
+      } else {
+        reader_->ReInit();
+        return true;
+      }
+    }
+  }
+
+  void ReInit() override {
+    pass_count_ = 0;
+    reader_->ReInit();
+  }
+
+ private:
+  int pass_num_;
+  mutable int pass_count_;
+};
+
+class CreateMultiPassReaderOp : public framework::OperatorBase {
+ public:
+  using framework::OperatorBase::OperatorBase;
+
+ private:
+  void RunImpl(const framework::Scope& scope,
+               const platform::Place& dev_place) const override {
+    const auto& underlying_reader = scope.FindVar(Input("UnderlyingReader"))
+                                        ->Get<framework::ReaderHolder>();
+    auto& out = detail::Ref(scope.FindVar(Output("Out")));
+    int pass_num = Attr<int>("pass_num");
+    out.GetMutable<framework::ReaderHolder>()->Reset(
+        new MultiPassReader(underlying_reader.Get(), pass_num));
+  }
+};
+
+class CreateMultiPassReaderOpMaker : public DecoratedReaderMakerBase {
+ public:
+  CreateMultiPassReaderOpMaker(OpProto* op_proto, OpAttrChecker* op_checker)
+      : DecoratedReaderMakerBase(op_proto, op_checker) {
+    AddAttr<int>("pass_num", "The number of pass to run.").GreaterThan(0);
+    AddComment(R"DOC(
+      CreateMultiPassReader Operator
+
+      This operator creates a multi-pass reader. A multi-pass reader 
+      is used to yield data for several pass training continuously. 
+      It takes the the number of pass to run as one of its attributes
+      ('pass_num'), and maintains a pass counter to record how many 
+      passes it has completed. When the underlying reader reach the EOF, 
+      the multi-pass reader checks whether it has completed training 
+      of the given number of pass. If not, the underlying reader will 
+      be re-initialized and starts a new pass automatically.
+    )DOC");
+  }
+};
+
+}  // namespace reader
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators::reader;
+REGISTER_DECORATED_READER_OPERATOR(create_multi_pass_reader,
+                                   ops::CreateMultiPassReaderOp,
+                                   ops::CreateMultiPassReaderOpMaker);

paddle/fluid/operators/reader/create_shuffle_reader_op.cc

@@ -34,6 +34,9 @@ class ShuffleReader : public framework::DecoratedReader {
   }
 
   void ReadNext(std::vector<framework::LoDTensor>* out) override {
+    if (!HasNext()) {
+      PADDLE_THROW("There is no next data!");
+    }
     if (iteration_pos_ >= buffer_.size()) {
       VLOG(10) << "Resetting shuffle buffer";
       ReadIntoBuffers();
@@ -50,7 +53,6 @@ class ShuffleReader : public framework::DecoratedReader {
     buffer_.clear();
     buffer_.reserve(buffer_size_);
     iteration_pos_ = 0;
-    PADDLE_ENFORCE(reader_->HasNext());
     for (size_t i = 0; i < buffer_size_; ++i) {
       if (!reader_->HasNext()) {
         break;

paddle/fluid/operators/reader/open_files_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/framework/channel.h"
+#include "paddle/fluid/operators/reader/reader_op_registry.h"
+
+namespace paddle {
+namespace operators {
+namespace reader {
+
+class MultipleReader : public framework::ReaderBase {
+ public:
+  MultipleReader(const std::vector<std::string>& file_names,
+                 const std::vector<framework::DDim>& dims, size_t thread_num)
+      : file_names_(file_names), dims_(dims) {
+    prefetchers_.resize(thread_num);
+    StartNewScheduler();
+  }
+
+  void ReadNext(std::vector<framework::LoDTensor>* out) override;
+  bool HasNext() const override;
+  void ReInit() override;
+
+  ~MultipleReader() { EndScheduler(); }
+
+ private:
+  void StartNewScheduler();
+  void EndScheduler();
+  void ScheduleThreadFunc();
+  void PrefetchThreadFunc(std::string file_name, size_t thread_idx);
+
+  std::vector<std::string> file_names_;
+  std::vector<framework::DDim> dims_;
+  std::thread scheduler_;
+  std::vector<std::thread> prefetchers_;
+  framework::Channel<size_t>* waiting_file_idx_;
+  framework::Channel<size_t>* available_thread_idx_;
+  framework::Channel<std::vector<framework::LoDTensor>>* buffer_;
+  mutable std::vector<framework::LoDTensor> local_buffer_;
+};
+
+void MultipleReader::ReadNext(std::vector<framework::LoDTensor>* out) {
+  if (!HasNext()) {
+    PADDLE_THROW("There is no next data!");
+  }
+
+  if (local_buffer_.empty()) {
+    buffer_->Receive(&local_buffer_);
+  }
+  *out = local_buffer_;
+  local_buffer_.clear();
+}
+
+bool MultipleReader::HasNext() const {
+  return local_buffer_.empty() ? buffer_->Receive(&local_buffer_) : true;
+}
+
+void MultipleReader::ReInit() {
+  EndScheduler();
+  local_buffer_.clear();
+  StartNewScheduler();
+}
+
+void MultipleReader::StartNewScheduler() {
+  size_t thread_num = prefetchers_.size();
+  waiting_file_idx_ = framework::MakeChannel<size_t>(file_names_.size());
+  available_thread_idx_ = framework::MakeChannel<size_t>(thread_num);
+  buffer_ =
+      framework::MakeChannel<std::vector<framework::LoDTensor>>(thread_num);
+
+  for (size_t i = 0; i < file_names_.size(); ++i) {
+    waiting_file_idx_->Send(&i);
+  }
+  waiting_file_idx_->Close();
+  for (size_t i = 0; i < thread_num; ++i) {
+    available_thread_idx_->Send(&i);
+  }
+
+  scheduler_ = std::thread([this] { ScheduleThreadFunc(); });
+}
+
+void MultipleReader::EndScheduler() {
+  available_thread_idx_->Close();
+  buffer_->Close();
+  waiting_file_idx_->Close();
+  if (scheduler_.joinable()) {
+    scheduler_.join();
+  }
+  delete buffer_;
+  delete available_thread_idx_;
+  delete waiting_file_idx_;
+}
+
+void MultipleReader::ScheduleThreadFunc() {
+  VLOG(5) << "MultipleReader schedule thread starts.";
+  size_t completed_thread_num = 0;
+  size_t thread_idx;
+  while (available_thread_idx_->Receive(&thread_idx)) {
+    std::thread& prefetcher = prefetchers_[thread_idx];
+    if (prefetcher.joinable()) {
+      prefetcher.join();
+    }
+    size_t file_idx;
+    if (waiting_file_idx_->Receive(&file_idx)) {
+      // Still have files to read. Start a new prefetch thread.
+      std::string file_name = file_names_[file_idx];
+      prefetcher = std::thread([this, file_name, thread_idx] {
+        PrefetchThreadFunc(file_name, thread_idx);
+      });
+    } else {
+      // No more file to read.
+      ++completed_thread_num;
+      if (completed_thread_num == prefetchers_.size()) {
+        buffer_->Close();
+        break;
+      }
+    }
+  }
+  // If users invoke ReInit() when scheduler is running, it will close the
+  // 'avaiable_thread_idx_' and prefecther threads have no way to tell scheduler
+  // to release their resource. So a check is needed before scheduler ends.
+  for (auto& p : prefetchers_) {
+    if (p.joinable()) {
+      p.join();
+    }
+  }
+  VLOG(5) << "MultipleReader schedule thread terminates.";
+}
+
+void MultipleReader::PrefetchThreadFunc(std::string file_name,
+                                        size_t thread_idx) {
+  VLOG(5) << "The prefetch thread of file '" << file_name << "' starts.";
+  std::unique_ptr<framework::ReaderBase> reader =
+      CreateReaderByFileName(file_name, dims_);
+  while (reader->HasNext()) {
+    std::vector<framework::LoDTensor> ins;
+    reader->ReadNext(&ins);
+    try {
+      buffer_->Send(&ins);
+    } catch (paddle::platform::EnforceNotMet e) {
+      VLOG(5) << "WARNING: The buffer channel has been closed. The prefetch "
+                 "thread of file '"
+              << file_name << "' will terminate.";
+      break;
+    }
+  }
+
+  try {
+    available_thread_idx_->Send(&thread_idx);
+  } catch (paddle::platform::EnforceNotMet e) {
+    VLOG(5) << "WARNING: The available_thread_idx_ channel has been closed. "
+               "Fail to send thread_idx.";
+  }
+  VLOG(5) << "The prefetch thread of file '" << file_name << "' terminates.";
+}
+
+class OpenFilesOp : public framework::OperatorBase {
+ public:
+  using framework::OperatorBase::OperatorBase;
+
+ private:
+  void RunImpl(const framework::Scope& scope,
+               const platform::Place& dev_place) const override {
+    const auto& shape_concat = Attr<std::vector<int>>("shape_concat");
+    const auto& ranks = Attr<std::vector<int>>("ranks");
+    PADDLE_ENFORCE(!shape_concat.empty() && !ranks.empty());
+    PADDLE_ENFORCE_EQ(std::accumulate(ranks.begin(), ranks.end(), 0),
+                      int(shape_concat.size()),
+                      "The accumulate of all ranks should be equal to the "
+                      "shape concat's length.");
+    const auto& file_names = Attr<std::vector<std::string>>("file_names");
+    PADDLE_ENFORCE(!file_names.empty(), "No file to be read!");
+    const size_t thread_num = Attr<int>("thread_num");
+
+    auto* out = scope.FindVar(Output("Out"))
+                    ->template GetMutable<framework::ReaderHolder>();
+    out->Reset(new MultipleReader(
+        file_names, RestoreShapes(shape_concat, ranks), thread_num));
+  }
+};
+
+class OpenFilesOpMaker : public FileReaderMakerBase {
+ public:
+  OpenFilesOpMaker(OpProto* op_proto, OpAttrChecker* op_checker)
+      : FileReaderMakerBase(op_proto, op_checker) {
+    AddAttr<std::vector<std::string>>("file_names", "Files to be read.");
+    AddAttr<int>("thread_num", "The maximal concurrent prefetch thread number.")
+        .GreaterThan(0);
+
+    AddComment(R"DOC(
+      OpenFiles Operator
+
+      An OpenFilesOp creates a MultipleReader, which is able to 
+      read data multi-threaded from multiple files.
+    )DOC");
+  }
+};
+
+}  // namespace reader
+}  // namespace operators
+}  // namespace paddle
+
+namespace reader = paddle::operators::reader;
+
+REGISTER_FILE_READER_OPERATOR(open_files, reader::OpenFilesOp,
+                              reader::OpenFilesOpMaker);

paddle/fluid/operators/reader/reader_op_registry.cc

@@ -36,6 +36,21 @@ std::unordered_map<std::string, FileReaderCreator>& FileReaderRegistry() {
   return regs;
 }
 
+std::unique_ptr<framework::ReaderBase> CreateReaderByFileName(
+    const std::string& file_name, const std::vector<framework::DDim>& dims) {
+  size_t separator_pos = file_name.find_last_of(kFileFormatSeparator);
+  PADDLE_ENFORCE_NE(separator_pos, std::string::npos,
+                    "File name illegal! A legal file name should be like: "
+                    "[file_name].[file_format] (e.g., 'data_file.recordio').");
+  std::string filetype = file_name.substr(separator_pos + 1);
+
+  auto itor = FileReaderRegistry().find(filetype);
+  PADDLE_ENFORCE(itor != FileReaderRegistry().end(),
+                 "No file reader registered for '%s' format.", filetype);
+  framework::ReaderBase* reader = (itor->second)(file_name, dims);
+  return std::unique_ptr<framework::ReaderBase>(reader);
+}
+
 FileReaderMakerBase::FileReaderMakerBase(
     framework::OpProtoAndCheckerMaker::OpProto* op_proto,
     framework::OpAttrChecker* op_checker)

paddle/fluid/operators/reader/reader_op_registry.h

@@ -21,6 +21,8 @@ namespace paddle {
 namespace operators {
 namespace reader {
 
+static constexpr char kFileFormatSeparator[] = ".";
+
 using FileReaderCreator = std::function<framework::ReaderBase*(
     const std::string&, const std::vector<framework::DDim>&)>;
 
@@ -29,12 +31,15 @@ std::unordered_map<std::string, FileReaderCreator>& FileReaderRegistry();
 template <typename Reader>
 int RegisterFileReader(const std::string& filetype) {
   FileReaderRegistry()[filetype] = [](
-      const std::string& fn, const std::vector<paddle::framework::DDim>& dim) {
-    return new Reader(fn, dim);
+      const std::string& fn, const std::vector<framework::DDim>& dims) {
+    return new Reader(fn, dims);
   };
   return 0;
 }
 
+std::unique_ptr<framework::ReaderBase> CreateReaderByFileName(
+    const std::string& file_name, const std::vector<framework::DDim>& dims);
+
 extern std::vector<framework::DDim> RestoreShapes(
     const std::vector<int>& shape_concat, const std::vector<int>& ranks);
 

paddle/fluid/operators/select_op.cc

@@ -27,6 +27,7 @@ namespace operators {
 
 static constexpr char kX[] = "X";
 static constexpr char kCaseToExecute[] = "case_to_execute";
+static constexpr char kOutputs[] = "Out";
 
 static constexpr char kCases[] = "cases";
 static constexpr char kCasesBlock[] = "sub_block";
@@ -388,6 +389,10 @@ class SelectOpMaker : public framework::OpProtoAndCheckerMaker {
              "(Int) The variable the sets the index of the case to execute, "
              "after evaluating the channels being sent to and received from")
         .AsDuplicable();
+    AddOutput(kOutputs,
+              "A set of variables, which will be assigned with values "
+              "generated by the operators inside the cases of Select Op.")
+        .AsDuplicable();
     AddAttr<std::vector<std::string>>(kCases,
                                       "(String vector) Serialized list of"
                                       "all cases in the select op. Each"

paddle/fluid/operators/send_barrier_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 <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 <future>
+#include "paddle/fluid/operators/detail/grpc_client.h"
+
+namespace paddle {
+namespace operators {
+
+class SendBarrierOp : public framework::OperatorBase {
+ public:
+  SendBarrierOp(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 {
+    std::vector<std::string> eps = Attr<std::vector<std::string>>("endpoints");
+
+    auto client_var_name = Output("RPCClient");
+    PADDLE_ENFORCE_NOT_NULL(scope.FindVar(client_var_name),
+                            "Can not find variable '%s' in the scope.",
+                            client_var_name);
+    auto* client_var = scope.FindVar(client_var_name);
+    detail::RPCClient* rpc_client = client_var->GetMutable<detail::RPCClient>();
+
+    // need to wait before sending send_barrier message
+    PADDLE_ENFORCE(rpc_client->Wait());
+
+    for (auto& ep : eps) {
+      VLOG(3) << "send barrier, ep: " << ep;
+      rpc_client->AsyncSendBatchBarrier(ep);
+    }
+    PADDLE_ENFORCE(rpc_client->Wait());
+  }
+};
+
+class SendBarrierOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  SendBarrierOpMaker(OpProto* proto, OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddOutput("RPCClient",
+              "(RPCClient) The RPC client object which is"
+              "initialized at most once.");
+    AddComment(R"DOC(
+SendBarrier operator
+
+This operator will send a send barrier signal to list_and_serv op, so that
+the Parameter Server would knew all variables have been sent.
+)DOC");
+
+    AddAttr<std::vector<std::string>>("endpoints",
+                                      "(string vector, default 127.0.0.1:6164)"
+                                      "Server endpoints to send variables to.")
+        .SetDefault({"127.0.0.1:6164"});
+  }
+};
+
+class SendBarrierOpVarTypeInference : public framework::VarTypeInference {
+ public:
+  void operator()(const framework::OpDesc& op_desc,
+                  framework::BlockDesc* block) const override {
+    auto out_var_name = op_desc.Output("RPCClient").front();
+    auto& out_var = block->FindRecursiveOrCreateVar(out_var_name);
+    auto var_type = framework::proto::VarType::RAW;
+    out_var.SetType(var_type);
+  }
+};
+
+class SendBarrierOpShapeInference : public framework::InferShapeBase {
+ public:
+  void operator()(framework::InferShapeContext* ctx) const override {}
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+REGISTER_OPERATOR(send_barrier, ops::SendBarrierOp,
+                  paddle::framework::EmptyGradOpMaker, ops::SendBarrierOpMaker,
+                  ops::SendBarrierOpVarTypeInference,
+                  ops::SendBarrierOpShapeInference);

paddle/fluid/operators/send_op.cc

@@ -21,6 +21,7 @@ limitations under the License. */
 
 #include <future>
 #include "paddle/fluid/operators/detail/grpc_client.h"
+#include "paddle/fluid/platform/profiler.h"
 
 namespace paddle {
 namespace operators {
@@ -59,6 +60,9 @@ class SendOp : public framework::OperatorBase {
     platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
     auto& ctx = *pool.Get(place);
 
+    // For profiling
+    platform::RecordEvent record_event(Type(), &ctx);
+
     auto client_var_name = Output("RPCClient");
     PADDLE_ENFORCE_NOT_NULL(scope.FindVar(client_var_name),
                             "Can not find variable '%s' in the scope.",
@@ -68,7 +72,7 @@ class SendOp : public framework::OperatorBase {
 
     for (size_t i = 0; i < ins.size(); i++) {
       if (NeedSend(scope, ins[i])) {
-        VLOG(3) << "sending " << ins[i] << " to " << epmap[i];
+        VLOG(2) << "sending " << ins[i] << " to " << epmap[i];
         rpc_client->AsyncSendVariable(epmap[i], ctx, scope, ins[i]);
       } else {
         VLOG(3) << "don't send no-initialied variable: " << ins[i];
@@ -77,20 +81,20 @@ class SendOp : public framework::OperatorBase {
     PADDLE_ENFORCE(rpc_client->Wait());
 
     for (auto& ep : endpoints) {
-      VLOG(3) << "batch barrier, ep: " << ep;
+      VLOG(2) << "batch barrier, ep: " << ep;
       rpc_client->AsyncSendBatchBarrier(ep);
     }
     PADDLE_ENFORCE(rpc_client->Wait());
 
     if (outs.size() > 0) {
       for (size_t i = 0; i < outs.size(); i++) {
-        VLOG(3) << "getting " << outs[i] << " from " << epmap[i];
+        VLOG(2) << "getting " << outs[i] << " from " << epmap[i];
         rpc_client->AsyncGetVariable(epmap[i], ctx, scope, outs[i]);
       }
       PADDLE_ENFORCE(rpc_client->Wait());
       // tell pservers that current trainer have called fetch
       for (auto& ep : endpoints) {
-        VLOG(3) << "send fetch barrier, ep: " << ep;
+        VLOG(2) << "send fetch barrier, ep: " << ep;
         rpc_client->AsyncSendFetchBarrier(ep);
       }
       PADDLE_ENFORCE(rpc_client->Wait());

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 <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 <future>
+#include "paddle/fluid/operators/detail/grpc_client.h"
+
+namespace paddle {
+namespace operators {
+static bool NeedSend(const framework::Scope& scope,
+                     const std::string& varname) {
+  auto* var = scope.FindVar(varname);
+  PADDLE_ENFORCE_NOT_NULL(var, "Can not find variable '%s' in the send side.",
+                          varname);
+  if (var->IsType<framework::LoDTensor>()) {
+    return var->Get<framework::LoDTensor>().IsInitialized();
+  } else if (var->IsType<framework::SelectedRows>()) {
+    return var->Get<framework::SelectedRows>().rows().size() > 0UL;
+  } else {
+    PADDLE_THROW(
+        "Variable type in send side should be in "
+        "[LodTensor, SelectedRows]");
+  }
+  return false;
+}
+
+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_sent");
+
+    platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
+    auto& ctx = *pool.Get(place);
+
+    auto client_var_name = Output("RPCClient");
+    PADDLE_ENFORCE_NOT_NULL(scope.FindVar(client_var_name),
+                            "Can not find variable '%s' in the scope.",
+                            client_var_name);
+    auto* client_var = scope.FindVar(client_var_name);
+    detail::RPCClient* rpc_client = client_var->GetMutable<detail::RPCClient>();
+
+    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->AsyncSendVariable(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:
+  SendVarsOpMaker(OpProto* proto, OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "(Tensor, SelectedRows) Input variables to be sent")
+        .AsDuplicable();
+    AddOutput("RPCClient",
+              "(RPCClient) The RPC client object which will be"
+              "initialized at most once.");
+    AddComment(R"DOC(
+Send operator
+
+This operator will send variables to listen_and_serve op at the parameter server.
+)DOC");
+    AddAttr<int>("ync_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 SendVarsOpVarTypeInference : public framework::VarTypeInference {
+ public:
+  void operator()(const framework::OpDesc& op_desc,
+                  framework::BlockDesc* block) const override {
+    auto out_var_name = op_desc.Output("RPCClient").front();
+    auto& out_var = block->FindRecursiveOrCreateVar(out_var_name);
+    auto var_type = framework::proto::VarType::RAW;
+    out_var.SetType(var_type);
+  }
+};
+
+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::SendVarsOpVarTypeInference,
+                  ops::SendVarsOpShapeInference);

paddle/fluid/operators/sequence_expand_op.cc

@@ -17,7 +17,7 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
-using framework::Tensor;
+using framework::LoDTensor;
 
 class SequenceExpandOp : public framework::OperatorWithKernel {
  public:
@@ -25,15 +25,71 @@ class SequenceExpandOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(framework::InferShapeContext* ctx) const override {
-    PADDLE_ENFORCE(ctx->HasInput("X"));
-    PADDLE_ENFORCE(ctx->HasOutput("Out"));
-    PADDLE_ENFORCE(ctx->HasInput("Y"));
-    framework::DDim out_dim;
-    auto y_dim = ctx->GetInputDim("Y");
-    out_dim = ctx->GetInputDim("X");
-    out_dim[0] = y_dim[0];
-    ctx->ShareLoD("Y", "Out");
-    ctx->SetOutputDim("Out", out_dim);
+    PADDLE_ENFORCE(ctx->HasInput("X"),
+                   "Input(X) of SequenceExpandOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Y"),
+                   "Input(Y) of SequenceExpandOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(Out) of SequenceExpandOp should not be null.");
+
+    auto x_dims = ctx->GetInputDim("X");
+    auto out_dims = x_dims;
+    int ref_level = ctx->Attrs().Get<int>("ref_level");
+
+    PADDLE_ENFORCE_GE(x_dims.size(), 2,
+                      "Dimension number of Input(X) should be at least 2.");
+
+    if (ctx->IsRuntime()) {
+      framework::Variable* x_var =
+          boost::get<framework::Variable*>(ctx->GetInputVarPtrs("X")[0]);
+      framework::Variable* y_var =
+          boost::get<framework::Variable*>(ctx->GetInputVarPtrs("Y")[0]);
+
+      auto& x_lod = x_var->Get<LoDTensor>().lod();
+      auto& y_lod = y_var->Get<LoDTensor>().lod();
+
+      PADDLE_ENFORCE_LE(x_lod.size(), 1,
+                        "Level number of Input(X)'s lod should not be "
+                        "greater than 1.");
+      PADDLE_ENFORCE_GT(y_lod.size(), 0,
+                        "Level number of Input(Y)'s lod should be "
+                        "greater than 0.");
+      PADDLE_ENFORCE(
+          ref_level == -1 ||
+              (ref_level >= 0 && ref_level < static_cast<int>(y_lod.size())),
+          "Invlid `ref_level`, which should be either equal to -1 "
+          "or in [0, %d)",
+          y_lod.size());
+
+      if (ref_level == -1) ref_level = y_lod.size() - 1;
+
+      if (x_lod.size() > 0) {
+        PADDLE_ENFORCE(x_lod[0].size() == y_lod[ref_level].size(),
+                       "Level number of Input(X)'s lod could be 0. Otherwise "
+                       "size of Input(X)'s first level lod should be equal to "
+                       "size of Input(Y)'s referred level lod.");
+      }
+
+      int64_t out_first_dim = 0;
+      if (y_lod[ref_level].size() <= 1) {
+        out_first_dim = x_dims[0];
+      } else {
+        for (size_t i = 1; i < y_lod[ref_level].size(); ++i) {
+          int x_seq_len = 1;
+          if (x_lod.size() == 1) {
+            x_seq_len = x_lod[0][i] - x_lod[0][i - 1];
+          }
+          out_first_dim +=
+              (y_lod[ref_level][i] - y_lod[ref_level][i - 1]) * x_seq_len;
+        }
+      }
+      out_dims[0] = out_first_dim;
+      ctx->SetOutputDim("Out", out_dims);
+    } else {
+      out_dims[0] = -1;
+      ctx->SetOutputDim("Out", out_dims);
+      ctx->ShareLoD("X", /*->*/ "Out");
+    }
   }
 };
 
@@ -42,83 +98,81 @@ class SequenceExpandOpMaker : public framework::OpProtoAndCheckerMaker {
   SequenceExpandOpMaker(OpProto* proto, OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X",
-             "(Tensor or LoDTensor) The input(X) of this operator can be a "
-             "LoDTensor or a base Tensor.");
+             "(LoDTensor, default LoDTensor<float>) A 2-D LoDTensor whose lod "
+             "level is at most 1.");
     AddInput("Y",
-             "(LoDTensor)The reference input(Y) of sequence_expand op."
-             "It must be a LoDTensor with k-level(k>0)."
-             "The input(X) will be expanded according to LOD of input(Y)."
-             "The element numbers of last level in input(Y) "
-             "must be equal to dims[0] of input(X).");
+             "(LoDTensor, default LoDTensor<float>) Referred LoDTensor whose "
+             "lod (specified level) is referred by Input(X).");
     AddOutput("Out",
-              "(LodTensor)The output of sequence_expand op."
-              "The lod of output will be as same as input(Y)'s lod.");
+              "(LodTensor, default LoDTensor<float>) Output LoDTensor which is "
+              "generated from Input(X) by referring lod of Input(Y).");
+    AddAttr<int>("ref_level", "Specify lod level of Input(Y).").SetDefault(-1);
     AddComment(R"DOC(
 Sequence Expand Operator.
 
-This operator expands input(X) according to LOD of input(Y).
+This operator expands `X` according to specified level lod of `Y`. Current
+implementation constaints that lod level of `X` should be at most 1. Attribute
+`ref_level` is used to specify which level lod of `Y` is referred to expand `X`.
+If set `ref_level` to -1, then last level lod of `Y` would be referred.
+Please note, rank of `X` should be at least 2, when the rank exceeds 2, `X`
+would be viewed as a 2-D tensor.
+
 Following are cases to better explain how this works:
+
 Case 1:
 
-Given a 2-level LoDTensor input(X)
-    X.lod = [[0,       2, 3],
-             [0, 1,    3, 4]]
-    X.data = [a, b, c, d]
+Given a 1-level LoDTensor input(X)
+    X.lod =  [[0,   2,        4]]
+    X.data = [[a], [b], [c], [d]]
     X.dims = [4, 1]
 and input(Y)
     Y.lod = [[0,    2,    4],
              [0, 3, 6, 7, 8]]
-with condition len(Y.lod[-1]) -1 == X.dims[0]
-then we get 2-level LoDTensor
-    Out.lod = [[0,                2,    4],
-               [0,       3,       6, 7, 8]]
-    Out.data = [a, a, a, b, b, b, c, d]
+ref_level: 0
+then we get 1-level LoDTensor
+    Out.lod =  [[0,   2,        4,        6,        8]]
+    Out.data = [[a], [b], [a], [b], [c], [d], [c], [d]]
     Out.dims = [8, 1]
 
 Case 2:
 
+Given 1-level LoDTensor input(X)
+    X.lod =  [[0,   1,        4]]
+    X.data = [[a], [b], [c], [d]]
+    X.dims = [4, 1]
+and input(Y)
+    Y.lod = [[0,    2,    4],
+             [0, 3, 6, 6, 8]]
+ref_level: 0
+then we get 1-level LoDTensor
+    Out.lod =  [[0,   1,   2,        5,             8]]
+    Out.data = [[a], [a], [b], [c], [d], [b], [c], [d]]
+    Out.dims = [8, 1]
+
+Case 3:
+
 Given a common Tensor input(X)
-    X.data = [a, b, c]
+    X.data = [[a], [b], [c]]
     X.dims = [3, 1]
 and input(Y)
     Y.lod = [[0, 2, 3, 6]]
-with condition len(Y.lod[-1]) -1 == X.dims[0]
-then we get 1-level LoDTensor
-    Out.lod = [[0,    2, 3,      6]]
-    Out.data = [a, a, b, c, c, c]
+ref_level: -1
+then we get a common Tensor
+    Out.data = [[a], [a], [b], [c], [c], [c]]
     Out.dims = [6, 1]
 
-Case 3:
+Case 4:
 
 Given a common Tensor input(X)
     X.data = [[a, b], [c, d], [e, f]]
     X.dims = [3, 2]
 and input(Y)
     Y.lod = [[0, 2, 3, 6]]
-with condition len(Y.lod[-1]) -1 == X.dims[0]
-then we get 1-level LoDTensor
-    Out.lod = [[0,           2,     3,                     6]]
-    Out.data = [[a,b], [a,b] [c,d], [e, f], [e, f], [e, f]]
+ref_level: 0
+then we get a common LoDTensor
+    Out.data = [[a, b], [a, b] [c, d], [e, f], [e, f], [e, f]]
     Out.dims = [6, 2]
 
-Case 4:
-
-Given 2-level a LoDTensor input(X)
-    X.lod = [[0,       2, 3],
-             [0, 1,    3, 4]]
-    X.data = [a, b, c, d]
-    X.dims = [4, 1]
-and input(Y)
-    Y.lod = [[0,    2,    4],
-             [0, 3, 6, 6, 8]]
-with condition len(Y.lod[-1]) -1 == X.dims[0]
-then we get 2-level LoDTensor
-    Out.lod = [[0,                2,    4],
-               [0,       3,       6, 6, 8]]
-    Out.data = [a, a, a, b, b, b, d, d]
-    Out.dims = [8, 1]
-
-
 )DOC");
   }
 };
@@ -129,12 +183,14 @@ class SequenceExpandOpGrad : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(framework::InferShapeContext* ctx) const override {
-    PADDLE_ENFORCE(ctx->HasInput("X"));
-    PADDLE_ENFORCE(ctx->HasInput("Out"));
+    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Out"), "Input(Out) should not be null.");
     PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
-                   "The input(Out@GRAD) should not be null");
+                   "Input(Out@GRAD) should not be null.");
+
     auto x_dims = ctx->GetInputDim("X");
     auto x_grad_name = framework::GradVarName("X");
+
     if (ctx->HasOutput(x_grad_name)) {
       ctx->SetOutputDim(x_grad_name, x_dims);
     }
@@ -149,7 +205,13 @@ REGISTER_OP(sequence_expand, ops::SequenceExpandOp, ops::SequenceExpandOpMaker,
             sequence_expand_grad, ops::SequenceExpandOpGrad);
 REGISTER_OP_CPU_KERNEL(
     sequence_expand,
-    ops::SequenceExpandKernel<paddle::platform::CPUDeviceContext, float>);
+    ops::SequenceExpandKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::SequenceExpandKernel<paddle::platform::CPUDeviceContext, double>,
+    ops::SequenceExpandKernel<paddle::platform::CPUDeviceContext, int>,
+    ops::SequenceExpandKernel<paddle::platform::CPUDeviceContext, int64_t>);
 REGISTER_OP_CPU_KERNEL(
     sequence_expand_grad,
-    ops::SequenceExpandGradKernel<paddle::platform::CPUDeviceContext, float>);
+    ops::SequenceExpandGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::SequenceExpandGradKernel<paddle::platform::CPUDeviceContext, double>,
+    ops::SequenceExpandGradKernel<paddle::platform::CPUDeviceContext, int>,
+    ops::SequenceExpandGradKernel<paddle::platform::CPUDeviceContext, int64_t>);

paddle/fluid/operators/sequence_expand_op.cu

@@ -150,7 +150,14 @@ struct SequenceExpandGradFunctor<platform::CUDADeviceContext, T> {
 namespace ops = paddle::operators;
 REGISTER_OP_CUDA_KERNEL(
     sequence_expand,
-    ops::SequenceExpandKernel<paddle::platform::CUDADeviceContext, float>);
+    ops::SequenceExpandKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::SequenceExpandKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::SequenceExpandKernel<paddle::platform::CUDADeviceContext, int>,
+    ops::SequenceExpandKernel<paddle::platform::CUDADeviceContext, int64_t>);
 REGISTER_OP_CUDA_KERNEL(
     sequence_expand_grad,
-    ops::SequenceExpandGradKernel<paddle::platform::CUDADeviceContext, float>);
+    ops::SequenceExpandGradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::SequenceExpandGradKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::SequenceExpandGradKernel<paddle::platform::CUDADeviceContext, int>,
+    ops::SequenceExpandGradKernel<paddle::platform::CUDADeviceContext,
+                                  int64_t>);

paddle/fluid/operators/softmax_cudnn_op.cu.cc

@@ -56,7 +56,9 @@ class SoftmaxGradCUDNNKernel : public framework::OpKernel<T> {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_KERNEL(softmax, CUDNN, ::paddle::platform::CUDAPlace,
-                   ops::SoftmaxCUDNNKernel<float>);
-REGISTER_OP_KERNEL(softmax_grad, CUDNN, ::paddle::platform::CUDAPlace,
+namespace plat = paddle::platform;
+REGISTER_OP_KERNEL(softmax, CUDNN, plat::CUDAPlace,
+                   ops::SoftmaxCUDNNKernel<float>,
+                   ops::SoftmaxCUDNNKernel<plat::float16>);
+REGISTER_OP_KERNEL(softmax_grad, CUDNN, plat::CUDAPlace,
                    ops::SoftmaxGradCUDNNKernel<float>);

paddle/fluid/operators/softmax_mkldnn_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 "mkldnn.hpp"
+#include "paddle/fluid/operators/softmax_op.h"
+#include "paddle/fluid/platform/mkldnn_helper.h"
+
+#include <iostream>
+
+namespace paddle {
+namespace operators {
+
+using paddle::framework::Tensor;
+using paddle::platform::MKLDNNDeviceContext;
+using paddle::platform::MKLDNNMemDesc;
+
+using mkldnn::memory;  // Note: paddle has also "memory" namespace
+using mkldnn::primitive;
+using mkldnn::softmax_forward;
+using mkldnn::prop_kind;
+using mkldnn::stream;
+
+template <typename T>
+class SoftmaxMKLDNNKernel : public paddle::framework::OpKernel<T> {
+ public:
+  void Compute(const paddle::framework::ExecutionContext& ctx) const override {
+    PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()),
+                   "It must use CPUPlace.");
+    auto& dev_ctx = ctx.template device_context<MKLDNNDeviceContext>();
+    auto mkldnn_engine = dev_ctx.GetEngine();
+    const Tensor* input = ctx.Input<Tensor>("X");
+    Tensor* output = ctx.Output<Tensor>("Out");
+    PADDLE_ENFORCE(input->dims().size() == 2UL,
+                   "The input of softmax op must be a 2D matrix.");
+    const T* input_data = input->data<T>();
+    // allocate memory for output
+    T* output_data = output->mutable_data<T>(ctx.GetPlace());
+    std::vector<int> src_tz = paddle::framework::vectorize2int(input->dims());
+    std::vector<int> dst_tz = paddle::framework::vectorize2int(output->dims());
+    // MKL-DNN does support softmax over selected axis. Having 2D Tensor,
+    // we will make normalization after final eg. axis: 1
+    PADDLE_ENFORCE(((src_tz[0] == dst_tz[0]) && (src_tz[1] == dst_tz[1])),
+                   "Softmax input and output dimensions should match");
+    // Same memory descriptor to be used for input and output
+    memory::dims softmax_tz = {src_tz[0], src_tz[1]};
+    // Currently only supports NC data format
+    // TODO(jczaja-intel): support more formats
+    auto softmax_md =
+        MKLDNNMemDesc({softmax_tz}, memory::f32, memory::format::nc);
+    // Normalization is made after innermost dimension eg. C out of NC
+    auto softmax_desc = softmax_forward::desc(prop_kind::forward_scoring,
+                                              softmax_md, 1 /*dim: C*/);
+    // create memory primitives
+    auto softmax_src_memory =
+        memory({softmax_md, mkldnn_engine}, (void*)input_data);
+    auto softmax_dst_memory =
+        memory({softmax_md, mkldnn_engine}, (void*)output_data);
+    auto softmax_prim_desc =
+        softmax_forward::primitive_desc(softmax_desc, mkldnn_engine);
+    auto softmax = softmax_forward(softmax_prim_desc, softmax_src_memory,
+                                   softmax_dst_memory);
+    std::vector<primitive> pipeline{softmax};
+    stream(stream::kind::eager).submit(pipeline).wait();
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+REGISTER_OP_KERNEL(softmax, MKLDNN, ::paddle::platform::CPUPlace,
+                   ops::SoftmaxMKLDNNKernel<float>);

paddle/fluid/operators/softmax_op.cc

@@ -13,7 +13,13 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/softmax_op.h"
+#ifdef PADDLE_WITH_CUDA
+#include "paddle/fluid/platform/cudnn_helper.h"
+#endif
 
+#ifdef PADDLE_WITH_MKLDNN
+#include "paddle/fluid/platform/mkldnn_helper.h"
+#endif
 namespace paddle {
 namespace operators {
 
@@ -38,26 +44,32 @@ class SoftmaxOp : public framework::OperatorWithKernel {
   framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext& ctx) const override {
     // choose cudnn kernel if the runtime supported.
-    bool use_cudnn = ctx.Attr<bool>("use_cudnn");
-    bool runtime_cudnn_support = false;
+    framework::LibraryType library_{framework::LibraryType::kPlain};
 #ifdef PADDLE_WITH_CUDA
-    if (platform::is_gpu_place(ctx.GetPlace())) {
-      auto& dev_ctx =
-          ctx.template device_context<platform::CUDADeviceContext>();
-      runtime_cudnn_support = dev_ctx.cudnn_handle() != nullptr ? true : false;
+    if (platform::CanCUDNNBeUsed(ctx)) {
+      library_ = framework::LibraryType::kCUDNN;
     }
 #endif
-    framework::LibraryType library_ = framework::LibraryType::kPlain;
-    if (use_cudnn && runtime_cudnn_support) {
-      library_ = framework::LibraryType::kCUDNN;
+#ifdef PADDLE_WITH_MKLDNN
+    if (library_ == framework::LibraryType::kPlain &&
+        platform::CanMKLDNNBeUsed(ctx)) {
+      library_ = framework::LibraryType::kMKLDNN;
     }
+#endif
+
+    auto input_data_type =
+        framework::ToDataType(ctx.Input<Tensor>("X")->type());
+    if (input_data_type == framework::proto::VarType::FP16) {
+      PADDLE_ENFORCE_EQ(library_, framework::LibraryType::kCUDNN,
+                        "float16 can only be used when CUDNN is used");
+    }
+
     std::string data_format = ctx.Attr<std::string>("data_format");
-    return framework::OpKernelType(
-        framework::ToDataType(ctx.Input<Tensor>("X")->type()), ctx.GetPlace(),
-        framework::StringToDataLayout(data_format), library_);
+    return framework::OpKernelType(input_data_type, ctx.GetPlace(),
+                                   framework::StringToDataLayout(data_format),
+                                   library_);
   }
 };
-
 class SoftmaxOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   SoftmaxOpMaker(OpProto* proto, OpAttrChecker* op_checker)
@@ -77,6 +89,9 @@ class SoftmaxOpMaker : public framework::OpProtoAndCheckerMaker {
         "Defaults to \"NHWC\". Specify the data format of the output data, "
         "the input will be transformed automatically. ")
         .SetDefault("AnyLayout");
+    AddAttr<bool>("use_mkldnn",
+                  "(bool, default false) Only used in mkldnn kernel")
+        .SetDefault(false);
     AddComment(R"DOC(
 Softmax Operator.
 
@@ -119,19 +134,12 @@ class SoftmaxOpGrad : public framework::OperatorWithKernel {
   framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext& ctx) const override {
     // choose cudnn kernel if the runtime supported.
-    bool use_cudnn = ctx.Attr<bool>("use_cudnn");
-    bool runtime_cudnn_support = false;
+    framework::LibraryType library_{framework::LibraryType::kPlain};
 #ifdef PADDLE_WITH_CUDA
-    if (platform::is_gpu_place(ctx.GetPlace())) {
-      auto& dev_ctx =
-          ctx.template device_context<platform::CUDADeviceContext>();
-      runtime_cudnn_support = dev_ctx.cudnn_handle() != nullptr ? true : false;
-    }
-#endif
-    framework::LibraryType library_ = framework::LibraryType::kPlain;
-    if (use_cudnn && runtime_cudnn_support) {
+    if (platform::CanCUDNNBeUsed(ctx)) {
       library_ = framework::LibraryType::kCUDNN;
     }
+#endif
     std::string data_format = ctx.Attr<std::string>("data_format");
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("X")->type()), ctx.GetPlace(),

paddle/fluid/operators/split_ids_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/split_ids_op.h"
+
+namespace paddle {
+namespace operators {
+
+class SplitIdsOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  SplitIdsOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("Ids", "(LoDTensor) the input ids with shape{batch_num, 1}");
+    AddOutput("Out", "(LoDTensor) The outputs of the input Ids.")
+        .AsDuplicable();
+
+    AddComment(R"DOC(
+Split a LoDTensor of Ids into multi LoDTensors, the number is pserver's number
+Example:
+  Input:
+    X = [1,2,3,4,5,6]
+
+  Out(3 output):
+    out0 = [3, 6]
+    out1 = [1, 4]
+    out2 = [2, 5]
+)DOC");
+  }
+};
+
+class SplitIdsOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext *ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("Ids"), "SplitIdsOp must has input Ids.");
+    PADDLE_ENFORCE(ctx->HasOutputs("Out"), "SplitIdsOp must has output Out.");
+
+    auto ids_var_type = ctx->GetInputsVarType("Ids").front();
+    PADDLE_ENFORCE_EQ(ids_var_type, framework::proto::VarType::LOD_TENSOR);
+
+    auto ids_dims = ctx->GetInputDim("Ids");
+    PADDLE_ENFORCE_EQ(ids_dims.size(), 2);
+    PADDLE_ENFORCE_EQ(ids_dims[1], 1);
+  }
+};
+
+class SplitIdsOpInferVarType : public framework::VarTypeInference {
+ public:
+  void operator()(const framework::OpDesc &op_desc,
+                  framework::BlockDesc *block) const override {
+    for (auto &out_var : op_desc.Output("Out")) {
+      block->Var(out_var)->SetType(framework::proto::VarType::LOD_TENSOR);
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(split_ids, ops::SplitIdsOp, ops::SplitIdsOpMaker,
+                  ops::SplitIdsOpInferVarType);
+REGISTER_OP_CPU_KERNEL(
+    split_ids, ops::SplitIdsOpKernel<paddle::platform::CPUPlace, int64_t>);

paddle/fluid/operators/split_ids_op.h

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include <vector>
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/math/selected_rows_functor.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename DeviceContext, typename T>
+class SplitIdsOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto place = ctx.GetPlace();
+    if (!platform::is_cpu_place(place)) {
+      PADDLE_THROW("SplitIds do not support GPU kernel");
+    }
+
+    const auto* ids_t = ctx.Input<framework::LoDTensor>("Ids");
+    auto& ids_dims = ids_t->dims();
+    auto outs = ctx.MultiOutput<framework::LoDTensor>("Out");
+
+    const T* ids = ids_t->data<T>();
+
+    const size_t shard_num = outs.size();
+
+    std::vector<std::vector<T>> out_ids;
+    out_ids.resize(outs.size());
+
+    // split id by their shard_num.
+    for (size_t i = 0; i < ids_dims[0]; ++i) {
+      T id = ids[i];
+      size_t shard_id = static_cast<size_t>(id) % shard_num;
+      out_ids[shard_id].push_back(id);
+    }
+
+    // create tensor for each shard and send to parameter server
+    for (size_t i = 0; i < out_ids.size(); ++i) {
+      auto* shard_t = outs[i];
+      std::vector<T> ids = out_ids[i];
+      auto* shard_data = shard_t->mutable_data<T>(
+          framework::make_ddim({static_cast<int64_t>(ids.size()), 1}), place);
+      for (size_t i = 0; i < ids.size(); ++i) {
+        shard_data[i] = ids[i];
+      }
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/target_assign_op.cc

@@ -153,8 +153,8 @@ template struct NegTargetAssignFunctor<platform::CPUDeviceContext, float,
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_WITHOUT_GRADIENT(target_assign, ops::TargetAssignOp,
-                             ops::TargetAssignOpMaker);
+REGISTER_OPERATOR(target_assign, ops::TargetAssignOp, ops::TargetAssignOpMaker,
+                  paddle::framework::EmptyGradOpMaker);
 REGISTER_OP_CPU_KERNEL(
     target_assign,
     ops::TargetAssignKernel<paddle::platform::CPUDeviceContext, int, float>,

paddle/fluid/platform/CMakeLists.txt

@@ -49,7 +49,7 @@ nv_test(device_context_test SRCS device_context_test.cu DEPS device_context gpu_
 nv_test(cudnn_helper_test SRCS cudnn_helper_test.cc DEPS dynload_cuda)
 nv_test(transform_test SRCS transform_test.cu DEPS paddle_memory place device_context)
 
-cc_library(device_tracer SRCS device_tracer.cc DEPS profiler_proto ${GPU_CTX_DEPS})
+cc_library(device_tracer SRCS device_tracer.cc DEPS boost profiler_proto ${GPU_CTX_DEPS})
 cc_library(profiler SRCS profiler.cc DEPS device_context device_tracer)
 cc_test(profiler_test SRCS profiler_test.cc DEPS profiler)
 

paddle/fluid/platform/cudnn_helper.h

@@ -86,7 +86,8 @@ class CudnnDataType<float16> {
  public:
   static const cudnnDataType_t type = CUDNN_DATA_HALF;
   // The scaling param type is float for HALF and FLOAT tensors
-  typedef const float ScalingParamType;
+  using ScalingParamType = const float;
+  using BatchNormParamType = float;
   static ScalingParamType* kOne() {
     static ScalingParamType v = 1.0;
     return &v;
@@ -101,7 +102,8 @@ template <>
 class CudnnDataType<float> {
  public:
   static const cudnnDataType_t type = CUDNN_DATA_FLOAT;
-  typedef const float ScalingParamType;
+  using ScalingParamType = const float;
+  using BatchNormParamType = float;
   static ScalingParamType* kOne() {
     static ScalingParamType v = 1.0;
     return &v;
@@ -116,7 +118,8 @@ template <>
 class CudnnDataType<double> {
  public:
   static const cudnnDataType_t type = CUDNN_DATA_DOUBLE;
-  typedef const double ScalingParamType;
+  using ScalingParamType = const double;
+  using BatchNormParamType = double;
   static ScalingParamType* kOne() {
     static ScalingParamType v = 1.0;
     return &v;

paddle/fluid/platform/device_context.cc

@@ -10,43 +10,45 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/platform/device_context.h"
+#include <unordered_set>
 #include "paddle/fluid/memory/memory.h"
-
 namespace paddle {
 namespace platform {
 
 DeviceContextPool* DeviceContextPool::pool = nullptr;
 
-const platform::DeviceContext* DeviceContextPool::Get(
-    const platform::Place& place) {
+platform::DeviceContext* DeviceContextPool::Get(const platform::Place& place) {
   auto it = device_contexts_.find(place);
   if (it == device_contexts_.end()) {
     PADDLE_THROW(
         "'Place' is not supported, Please re-compile with WITH_GPU "
         "option");
   }
-  return it->second;
+  return it->second.get();
 }
 
 DeviceContextPool::DeviceContextPool(
     const std::vector<platform::Place>& places) {
   PADDLE_ENFORCE_GT(places.size(), 0);
-  for (size_t i = 0; i < places.size(); i++) {
-    if (platform::is_cpu_place(places[i])) {
+  using PtrType = std::unique_ptr<DeviceContext>;
+  std::unordered_set<Place, PlaceHash> set;
+  for (auto& p : places) {
+    set.insert(p);
+  }
+
+  for (auto& p : set) {
+    if (platform::is_cpu_place(p)) {
 #ifdef PADDLE_WITH_MKLDNN
-      device_contexts_.emplace(places[i],
-                               new platform::MKLDNNDeviceContext(
-                                   boost::get<platform::CPUPlace>(places[i])));
+      device_contexts_.emplace(
+          p, PtrType(new MKLDNNDeviceContext(boost::get<CPUPlace>(p))));
 #else
-      device_contexts_.emplace(places[i],
-                               new platform::CPUDeviceContext(
-                                   boost::get<platform::CPUPlace>(places[i])));
+      device_contexts_.emplace(
+          p, PtrType(new CPUDeviceContext(boost::get<CPUPlace>(p))));
 #endif
-    } else if (platform::is_gpu_place(places[i])) {
+    } else if (platform::is_gpu_place(p)) {
 #ifdef PADDLE_WITH_CUDA
-      device_contexts_.emplace(places[i],
-                               new platform::CUDADeviceContext(
-                                   boost::get<platform::CUDAPlace>(places[i])));
+      device_contexts_.emplace(
+          p, PtrType(new CUDADeviceContext(boost::get<CUDAPlace>(p))));
 #else
       PADDLE_THROW(
           "'CUDAPlace' is not supported, Please re-compile with WITH_GPU "
@@ -159,6 +161,7 @@ CUDADeviceContext::~CUDADeviceContext() {
 Place CUDADeviceContext::GetPlace() const { return place_; }
 
 void CUDADeviceContext::Wait() const {
+  std::lock_guard<std::mutex> guard(mutex_);
   PADDLE_ENFORCE(cudaStreamSynchronize(stream_));
   PADDLE_ENFORCE(cudaGetLastError());
 }

paddle/fluid/platform/device_context.h

@@ -103,6 +103,7 @@ class CUDADeviceContext : public DeviceContext {
   std::unique_ptr<Eigen::GpuDevice> eigen_device_;
   std::unique_ptr<EigenCudaStreamDevice> eigen_stream_;
 
+  mutable std::mutex mutex_;
   cudaStream_t stream_;
   cudnnHandle_t cudnn_handle_;
   cublasHandle_t cublas_handle_;
@@ -159,7 +160,7 @@ class DeviceContextPool {
   }
 
   /*! \brief  Return handle of single device context. */
-  const platform::DeviceContext* Get(const platform::Place& place);
+  platform::DeviceContext* Get(const platform::Place& place);
 
   template <typename Place>
   const typename DefaultDeviceContextType<Place>::TYPE* GetByPlace(
@@ -172,19 +173,8 @@ class DeviceContextPool {
 
  private:
   static DeviceContextPool* pool;
-  constexpr static int LEFT_SHIFT = 8;
-  struct Hash {
-    std::hash<int> hash_;
-    size_t operator()(const platform::Place& place) const {
-      int pre_hash = place.which() << LEFT_SHIFT;
-      if (platform::is_gpu_place(place)) {
-        pre_hash += boost::get<platform::CUDAPlace>(place).GetDeviceId();
-      }
-      return hash_(pre_hash);
-    }
-  };
-  std::unordered_map<const platform::Place, const platform::DeviceContext*,
-                     Hash>
+  std::unordered_map<const platform::Place,
+                     std::unique_ptr<platform::DeviceContext>, PlaceHash>
       device_contexts_;
   DISABLE_COPY_AND_ASSIGN(DeviceContextPool);
 };

paddle/fluid/platform/float16.h

@@ -483,9 +483,124 @@ DEVICE inline bool operator>=(const half& a, const half& b) {
 
 #endif  // PADDLE_CUDA_FP16
 
-// Arithmetic operators on ARMv8.2-A CPU
-#if defined(PADDLE_WITH_NATIVE_FP16)
-HOST inline float16 operator+(const float16& a, const float16& b) {
+// Arithmetic operators for float16 on GPU
+#if defined(PADDLE_CUDA_FP16)
+HOSTDEVICE inline float16 operator+(const float16& a, const float16& b) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
+  return float16(__hadd(half(a), half(b)));
+#else
+  return float16(float(a) + float(b));
+#endif
+}
+
+HOSTDEVICE inline float16 operator-(const float16& a, const float16& b) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
+  return float16(__hsub(half(a), half(b)));
+#else
+  return float16(float(a) - float(b));
+#endif
+}
+
+HOSTDEVICE inline float16 operator*(const float16& a, const float16& b) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
+  return float16(__hmul(half(a), half(b)));
+#else
+  return float16(float(a) * float(b));
+#endif
+}
+
+HOSTDEVICE inline float16 operator/(const float16& a, const float16& b) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
+  // TODO(kexinzhao): check which cuda version starts to support __hdiv
+  float num = __half2float(half(a));
+  float denom = __half2float(half(b));
+  return float16(num / denom);
+#else
+  return float16(float(a) / float(b));
+#endif
+}
+
+HOSTDEVICE inline float16 operator-(const float16& a) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
+  return float16(__hneg(half(a)));
+#else
+  float16 res;
+  res.x = a.x ^ 0x8000;
+  return res;
+#endif
+}
+
+HOSTDEVICE inline float16& operator+=(float16& a, const float16& b) {
+  a = a + b;
+  return a;
+}
+
+HOSTDEVICE inline float16& operator-=(float16& a, const float16& b) {
+  a = a - b;
+  return a;
+}
+
+HOSTDEVICE inline float16& operator*=(float16& a, const float16& b) {
+  a = a * b;
+  return a;
+}
+
+HOSTDEVICE inline float16& operator/=(float16& a, const float16& b) {
+  a = a / b;
+  return a;
+}
+
+HOSTDEVICE inline bool operator==(const float16& a, const float16& b) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
+  return __heq(half(a), half(b));
+#else
+  return float(a) == float(b);
+#endif
+}
+
+HOSTDEVICE inline bool operator!=(const float16& a, const float16& b) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
+  return __hne(half(a), half(b));
+#else
+  return float(a) != float(b);
+#endif
+}
+
+HOSTDEVICE inline bool operator<(const float16& a, const float16& b) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
+  return __hlt(half(a), half(b));
+#else
+  return float(a) < float(b);
+#endif
+}
+
+HOSTDEVICE inline bool operator<=(const float16& a, const float16& b) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
+  return __hle(half(a), half(b));
+#else
+  return float(a) <= float(b);
+#endif
+}
+
+HOSTDEVICE inline bool operator>(const float16& a, const float16& b) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
+  return __hgt(half(a), half(b));
+#else
+  return float(a) > float(b);
+#endif
+}
+
+HOSTDEVICE inline bool operator>=(const float16& a, const float16& b) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
+  return __hge(half(a), half(b));
+#else
+  return float(a) >= float(b);
+#endif
+}
+
+// Arithmetic operators for float16 on ARMv8.2-A CPU
+#elif defined(PADDLE_WITH_NATIVE_FP16)
+inline float16 operator+(const float16& a, const float16& b) {
   float16 res;
   asm volatile(
       "ld1 {v0.h}[0], [%[a_ptr]]\n"
@@ -501,7 +616,7 @@ HOST inline float16 operator+(const float16& a, const float16& b) {
   return res;
 }
 
-HOST inline float16 operator-(const float16& a, const float16& b) {
+inline float16 operator-(const float16& a, const float16& b) {
   float16 res;
   asm volatile(
       "ld1 {v0.h}[0], [%[a_ptr]]\n"
@@ -517,7 +632,7 @@ HOST inline float16 operator-(const float16& a, const float16& b) {
   return res;
 }
 
-HOST inline float16 operator*(const float16& a, const float16& b) {
+inline float16 operator*(const float16& a, const float16& b) {
   float16 res;
   asm volatile(
       "ld1 {v0.h}[0], [%[a_ptr]]\n"
@@ -533,7 +648,7 @@ HOST inline float16 operator*(const float16& a, const float16& b) {
   return res;
 }
 
-HOST inline float16 operator/(const float16& a, const float16& b) {
+inline float16 operator/(const float16& a, const float16& b) {
   float16 res;
   asm volatile(
       "ld1 {v0.h}[0], [%[a_ptr]]\n"
@@ -549,7 +664,7 @@ HOST inline float16 operator/(const float16& a, const float16& b) {
   return res;
 }
 
-HOST inline float16 operator-(const float16& a) {
+inline float16 operator-(const float16& a) {
   float16 res;
   asm volatile(
       "ld1 {v0.h}[0], [%[a_ptr]]\n"
@@ -564,27 +679,27 @@ HOST inline float16 operator-(const float16& a) {
   return res;
 }
 
-HOST inline float16& operator+=(float16& a, const float16& b) {
+inline float16& operator+=(float16& a, const float16& b) {
   a = a + b;
   return a;
 }
 
-HOST inline float16& operator-=(float16& a, const float16& b) {
+inline float16& operator-=(float16& a, const float16& b) {
   a = a - b;
   return a;
 }
 
-HOST inline float16& operator*=(float16& a, const float16& b) {
+inline float16& operator*=(float16& a, const float16& b) {
   a = a * b;
   return a;
 }
 
-HOST inline float16& operator/=(float16& a, const float16& b) {
+inline float16& operator/=(float16& a, const float16& b) {
   a = a / b;
   return a;
 }
 
-HOST inline bool operator==(const float16& a, const float16& b) {
+inline bool operator==(const float16& a, const float16& b) {
   uint16_t res;
   asm volatile(
       "ld1 {v0.h}[0], [%[a_ptr]]\n"
@@ -600,11 +715,9 @@ HOST inline bool operator==(const float16& a, const float16& b) {
   return (res & 0xffff) != 0;
 }
 
-HOST inline bool operator!=(const float16& a, const float16& b) {
-  return !(a == b);
-}
+inline bool operator!=(const float16& a, const float16& b) { return !(a == b); }
 
-HOST inline bool operator<(const float16& a, const float16& b) {
+inline bool operator<(const float16& a, const float16& b) {
   uint16_t res;
   asm volatile(
       "ld1 {v1.h}[0], [%[a_ptr]]\n"
@@ -620,7 +733,7 @@ HOST inline bool operator<(const float16& a, const float16& b) {
   return (res & 0xffff) != 0;
 }
 
-HOST inline bool operator<=(const float16& a, const float16& b) {
+inline bool operator<=(const float16& a, const float16& b) {
   uint16_t res;
   asm volatile(
       "ld1 {v1.h}[0], [%[a_ptr]]\n"
@@ -636,7 +749,7 @@ HOST inline bool operator<=(const float16& a, const float16& b) {
   return (res & 0xffff) != 0;
 }
 
-HOST inline bool operator>(const float16& a, const float16& b) {
+inline bool operator>(const float16& a, const float16& b) {
   uint16_t res;
   asm volatile(
       "ld1 {v0.h}[0], [%[a_ptr]]\n"
@@ -652,7 +765,7 @@ HOST inline bool operator>(const float16& a, const float16& b) {
   return (res & 0xffff) != 0;
 }
 
-HOST inline bool operator>=(const float16& a, const float16& b) {
+inline bool operator>=(const float16& a, const float16& b) {
   uint16_t res;
   asm volatile(
       "ld1 {v0.h}[0], [%[a_ptr]]\n"
@@ -668,71 +781,71 @@ HOST inline bool operator>=(const float16& a, const float16& b) {
   return (res & 0xffff) != 0;
 }
 
-// Arithmetic operators, software emulated on other CPU
+// Arithmetic operators for float16, software emulated on other CPU
 #else
-HOSTDEVICE inline float16 operator+(const float16& a, const float16& b) {
+inline float16 operator+(const float16& a, const float16& b) {
   return float16(float(a) + float(b));
 }
 
-HOSTDEVICE inline float16 operator-(const float16& a, const float16& b) {
+inline float16 operator-(const float16& a, const float16& b) {
   return float16(float(a) - float(b));
 }
 
-HOSTDEVICE inline float16 operator*(const float16& a, const float16& b) {
+inline float16 operator*(const float16& a, const float16& b) {
   return float16(float(a) * float(b));
 }
 
-HOSTDEVICE inline float16 operator/(const float16& a, const float16& b) {
+inline float16 operator/(const float16& a, const float16& b) {
   return float16(float(a) / float(b));
 }
 
-HOSTDEVICE inline float16 operator-(const float16& a) {
+inline float16 operator-(const float16& a) {
   float16 res;
   res.x = a.x ^ 0x8000;
   return res;
 }
 
-HOSTDEVICE inline float16& operator+=(float16& a, const float16& b) {
+inline float16& operator+=(float16& a, const float16& b) {
   a = float16(float(a) + float(b));
   return a;
 }
 
-HOSTDEVICE inline float16& operator-=(float16& a, const float16& b) {
+inline float16& operator-=(float16& a, const float16& b) {
   a = float16(float(a) - float(b));
   return a;
 }
 
-HOSTDEVICE inline float16& operator*=(float16& a, const float16& b) {
+inline float16& operator*=(float16& a, const float16& b) {
   a = float16(float(a) * float(b));
   return a;
 }
 
-HOSTDEVICE inline float16& operator/=(float16& a, const float16& b) {
+inline float16& operator/=(float16& a, const float16& b) {
   a = float16(float(a) / float(b));
   return a;
 }
 
-HOSTDEVICE inline bool operator==(const float16& a, const float16& b) {
+inline bool operator==(const float16& a, const float16& b) {
   return float(a) == float(b);
 }
 
-HOSTDEVICE inline bool operator!=(const float16& a, const float16& b) {
+inline bool operator!=(const float16& a, const float16& b) {
   return float(a) != float(b);
 }
 
-HOSTDEVICE inline bool operator<(const float16& a, const float16& b) {
+inline bool operator<(const float16& a, const float16& b) {
   return float(a) < float(b);
 }
 
-HOSTDEVICE inline bool operator<=(const float16& a, const float16& b) {
+inline bool operator<=(const float16& a, const float16& b) {
   return float(a) <= float(b);
 }
 
-HOSTDEVICE inline bool operator>(const float16& a, const float16& b) {
+inline bool operator>(const float16& a, const float16& b) {
   return float(a) > float(b);
 }
 
-HOSTDEVICE inline bool operator>=(const float16& a, const float16& b) {
+inline bool operator>=(const float16& a, const float16& b) {
   return float(a) >= float(b);
 }
 #endif

paddle/fluid/platform/place.h

@@ -65,6 +65,18 @@ bool is_cpu_place(const Place &);
 bool places_are_same_class(const Place &, const Place &);
 bool is_same_place(const Place &, const Place &);
 
+struct PlaceHash {
+  std::size_t operator()(const Place &p) const {
+    constexpr size_t num_dev_bits = 4;
+    std::hash<int> ihash;
+    size_t dev_id = 0;
+    if (is_gpu_place(p)) {
+      dev_id = boost::get<CUDAPlace>(p).device;
+    }
+    return ihash(dev_id << num_dev_bits | p.which());
+  }
+};
+
 std::ostream &operator<<(std::ostream &, const Place &);
 
 template <typename Visitor>

paddle/fluid/pybind/CMakeLists.txt

 if(WITH_PYTHON)
+  if(WITH_AMD_GPU)
+    hip_library(paddle_pybind SHARED
+      SRCS pybind.cc exception.cc protobuf.cc const_value.cc recordio.cc
+      DEPS pybind python backward proto_desc paddle_memory executor prune init profiler feed_fetch_method
+      ${GLOB_OP_LIB})
+  else()
     cc_library(paddle_pybind SHARED
       SRCS pybind.cc exception.cc protobuf.cc const_value.cc recordio.cc
       DEPS pybind python backward proto_desc paddle_memory executor prune init profiler feed_fetch_method
@@ -6,4 +12,5 @@ if(WITH_PYTHON)
     if(NOT APPLE AND NOT ANDROID)
       target_link_libraries(paddle_pybind rt)
     endif(NOT APPLE AND NOT ANDROID)
+  endif(WITH_AMD_GPU)
 endif(WITH_PYTHON)

paddle/fluid/recordio/header.cc

@@ -29,8 +29,8 @@ Header::Header(uint32_t num, uint32_t sum, Compressor c, uint32_t cs)
 
 bool Header::Parse(std::istream& is) {
   uint32_t magic;
-  size_t read_size =
-      is.readsome(reinterpret_cast<char*>(&magic), sizeof(uint32_t));
+  is.read(reinterpret_cast<char*>(&magic), sizeof(uint32_t));
+  size_t read_size = is.gcount();
   if (read_size < sizeof(uint32_t)) {
     return false;
   }

paddle/fluid/recordio/scanner.cc

@@ -28,6 +28,7 @@ Scanner::Scanner(const std::string &filename) {
 }
 
 void Scanner::Reset() {
+  stream_->clear();
   stream_->seekg(0, std::ios::beg);
   ParseNextChunk();
 }

paddle/math/MathFunctions.cpp

@@ -59,17 +59,10 @@ void* lapack_dso_handle = nullptr;
   } __name;  // struct DynLoad__##__name
 #endif
 
-#ifdef PADDLE_USE_ATLAS
-  #define  PADDLE_SGETRF  clapack_sgetrf
-  #define  PADDLE_DGETRF  clapack_dgetrf
-  #define  PADDLE_SGETRI  clapack_sgetri
-  #define  PADDLE_DGETRI  clapack_dgetri
-#else
-  #define  PADDLE_SGETRF  LAPACKE_sgetrf
-  #define  PADDLE_DGETRF  LAPACKE_dgetrf
-  #define  PADDLE_SGETRI  LAPACKE_sgetri
-  #define  PADDLE_DGETRI  LAPACKE_dgetri
-#endif
+#define  PADDLE_SGETRF  LAPACKE_sgetrf
+#define  PADDLE_DGETRF  LAPACKE_dgetrf
+#define  PADDLE_SGETRI  LAPACKE_sgetri
+#define  PADDLE_DGETRI  LAPACKE_dgetri
 
 #define LAPACK_ROUTINE_EACH(__macro)       \
   __macro(PADDLE_SGETRF)                   \

paddle/math/MathFunctions.h

@@ -21,7 +21,7 @@ limitations under the License. */
 #include <mkl_vml_functions.h>
 #endif
 
-#if defined(PADDLE_USE_ATLAS) || defined(PADDLE_USE_VECLIB)
+#if defined(PADDLE_USE_VECLIB)
 extern "C" {
 #include <cblas.h>
 #include <clapack.h>

paddle/scripts/docker/build.sh

@@ -35,8 +35,9 @@ function cmake_gen() {
         -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE:-Release}
         ${PYTHON_FLAGS}
         -DWITH_DSO=ON
-        -DWITH_DOC=OFF
+        -DWITH_DOC=${WITH_DOC:-OFF}
         -DWITH_GPU=${WITH_GPU:-OFF}
+        -DWITH_AMD_GPU=${WITH_AMD_GPU:-OFF}
         -DWITH_DISTRIBUTE=${WITH_DISTRIBUTE:-OFF}
         -DWITH_MKL=${WITH_MKL:-ON}
         -DWITH_AVX=${WITH_AVX:-OFF}
@@ -50,6 +51,7 @@ function cmake_gen() {
         -DWITH_STYLE_CHECK=${WITH_STYLE_CHECK:-ON}
         -DWITH_TESTING=${WITH_TESTING:-ON}
         -DWITH_FAST_BUNDLE_TEST=ON
+        -DCMAKE_MODULE_PATH=/opt/rocm/hip/cmake
         -DCMAKE_EXPORT_COMPILE_COMMANDS=ON
     ========================================
 EOF
@@ -60,8 +62,9 @@ EOF
         -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE:-Release} \
         ${PYTHON_FLAGS} \
         -DWITH_DSO=ON \
-        -DWITH_DOC=OFF \
+        -DWITH_DOC=${WITH_DOC:-OFF} \
         -DWITH_GPU=${WITH_GPU:-OFF} \
+        -DWITH_AMD_GPU=${WITH_AMD_GPU:-OFF} \
         -DWITH_DISTRIBUTE=${WITH_DISTRIBUTE:-OFF} \
         -DWITH_MKL=${WITH_MKL:-ON} \
         -DWITH_AVX=${WITH_AVX:-OFF} \
@@ -74,6 +77,7 @@ EOF
         -DWITH_STYLE_CHECK=${WITH_STYLE_CHECK:-ON} \
         -DWITH_TESTING=${WITH_TESTING:-ON} \
         -DWITH_FAST_BUNDLE_TEST=ON \
+        -DCMAKE_MODULE_PATH=/opt/rocm/hip/cmake \
         -DCMAKE_EXPORT_COMPILE_COMMANDS=ON
 }
 

paddle/scripts/tools/build_docs/.gitignore

-doc
-doc_cn

paddle/scripts/tools/build_docs/build_docs.sh

-#!/bin/bash
-docker run --rm \
-       -v $(git rev-parse --show-toplevel):/paddle \
-       -e "WITH_GPU=OFF" \
-       -e "WITH_AVX=ON" \
-       -e "WITH_DOC=ON" \
-       -e "WOBOQ=ON" \
-       ${1:-"paddlepaddle/paddle:latest-dev"}

python/paddle/fluid/concurrency.py

@@ -12,7 +12,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from layers.control_flow import BlockGuard, Select
+from layers.control_flow import BlockGuard, equal
+from .framework import Operator
 from layer_helper import LayerHelper, unique_name
 from layers import fill_constant
 import core
@@ -75,6 +76,206 @@ class Go(BlockGuard):
             attrs={'sub_block': go_block})
 
 
+class SelectCase(object):
+    DEFAULT = 0
+    SEND = 1
+    RECEIVE = 2
+
+    def __init__(self,
+                 select,
+                 case_idx,
+                 case_to_execute,
+                 channel_action_fn=None,
+                 channel=None,
+                 value=None,
+                 is_copy=False):
+        self.select = select
+        self.helper = LayerHelper('conditional_block')
+        self.main_program = self.helper.main_program
+        self.is_scalar_condition = True
+
+        self.case_to_execute = case_to_execute
+        self.idx = case_idx
+
+        # Since we aren't going to use the `channel_send` or `channel_recv`
+        # functions directly, we just need to capture the name.
+        self.action = (self.SEND
+                       if channel_action_fn.__name__ == ('channel_send') else
+                       self.RECEIVE) if channel_action_fn else self.DEFAULT
+
+        X = value
+        if self.action == self.SEND and is_copy:
+            # We create of copy of the data we want to send
+            copied_X = self.select.parent_block.create_var(
+                name=unique_name.generate(value.name + '_copy'),
+                type=value.type,
+                dtype=value.dtype,
+                shape=value.shape,
+                lod_level=value.lod_level,
+                capacity=value.capacity
+                if hasattr(value, 'capacity') else None, )
+
+            self.select.parent_block.append_op(
+                type="assign", inputs={"X": value}, outputs={"Out": copied_X})
+            X = copied_X
+
+        self.value = X
+        self.channel = channel
+
+    def __enter__(self):
+        self.block = self.main_program.create_block()
+
+    def construct_op(self):
+        main_program = self.helper.main_program
+        cases_block = main_program.current_block()
+
+        inner_outputs = set()
+        input_set = set()
+        params = set()
+
+        for op in self.block.ops:
+            # Iterate over all operators, get all the inputs
+            # and add as input to the SelectCase operator.
+            for iname in op.input_names:
+                for in_var_name in op.input(iname):
+                    if in_var_name not in inner_outputs:
+                        input_set.add(in_var_name)
+
+            for oname in op.output_names:
+                for out_var_name in op.output(oname):
+                    inner_outputs.add(out_var_name)
+
+        param_list = [
+            cases_block.var(each_name) for each_name in params
+            if each_name not in input_set
+        ]
+
+        # Iterate over all operators, get all the outputs
+        # add to the output list of SelectCase operator only if
+        # they exist in the parent block.
+        out_vars = []
+        for inner_out_name in inner_outputs:
+            if inner_out_name in cases_block.vars:
+                out_vars.append(cases_block.var(inner_out_name))
+
+        # First, create an op that will determine whether or not this is the
+        # conditional variable to execute.
+        should_execute_block = equal(
+            fill_constant(
+                shape=[1], dtype=core.VarDesc.VarType.INT32, value=self.idx),
+            self.case_to_execute)
+
+        step_scope = cases_block.create_var(
+            type=core.VarDesc.VarType.STEP_SCOPES)
+
+        cases_block.append_op(
+            type='conditional_block',
+            inputs={'X': [should_execute_block],
+                    'Params': param_list},
+            outputs={'Out': out_vars,
+                     'Scope': [step_scope]},
+            attrs={
+                'sub_block': self.block,
+                'is_scalar_condition': self.is_scalar_condition
+            })
+
+        return '%s,%s,%s,%s' % (self.idx, self.action, self.channel.name
+                                if self.channel else '', self.value.name
+                                if self.value else '')
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        self.main_program.rollback()
+        if exc_type is not None:
+            return False  # re-raise exception
+        return True
+
+
+class Select(BlockGuard):
+    def __init__(self, name=None):
+        self.helper = LayerHelper('select', name=name)
+        self.parent_block = self.helper.main_program.current_block()
+        self.cases = []
+
+        super(Select, self).__init__(self.helper.main_program)
+        self.case_to_execute = fill_constant(
+            shape=[1], dtype=core.VarDesc.VarType.INT32, value=-1)
+
+    def __enter__(self):
+        super(Select, self).__enter__()
+        return self
+
+    def case(self, channel_action_fn, channel, value, is_copy=False):
+        """Create a new block for this condition.
+        """
+        select_case = SelectCase(self,
+                                 len(self.cases), self.case_to_execute,
+                                 channel_action_fn, channel, value, is_copy)
+
+        self.cases.append(select_case)
+
+        return select_case
+
+    def default(self):
+        """Create a default case block for this condition.
+        """
+        default_case = SelectCase(self, len(self.cases), self.case_to_execute)
+
+        self.cases.append(default_case)
+
+        return default_case
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        if exc_type is not None:
+            return False
+
+        # Create a select op and another block to wrap its
+        # case blocks.
+        select_block = self.helper.main_program.current_block()
+        parent_block = self.helper.main_program.block(select_block.parent_idx)
+
+        # Construct each case op, inside the newly created select block.
+        serialized_cases = []
+        for case in self.cases:
+            serialized_cases.append(case.construct_op())
+
+        intermediate = set()
+        params = set()
+
+        for case_block in select_block.ops:
+            if case_block.attrs and 'sub_block' in case_block.attrs:
+                for each_op in case_block.attrs['sub_block'].ops:
+                    assert isinstance(each_op, Operator)
+                    for iname in each_op.input_names:
+                        for in_var_name in each_op.input(iname):
+                            if in_var_name not in intermediate:
+                                params.add(in_var_name)
+
+                    for oname in each_op.output_names:
+                        for out_var_name in each_op.output(oname):
+                            intermediate.add(out_var_name)
+
+        out_list = [
+            parent_block.var(var_name) for var_name in parent_block.vars
+            if var_name in intermediate
+        ]
+
+        X = [select_block.var_recursive(x_name) for x_name in params]
+
+        # Needs to be used by `equal` inside the cases block.
+        X.append(self.case_to_execute)
+
+        # Construct the select op.
+        parent_block.append_op(
+            type='select',
+            inputs={'X': X,
+                    'case_to_execute': self.case_to_execute},
+            attrs={'sub_block': select_block,
+                   'cases': serialized_cases},
+            outputs={'Out': out_list})
+
+        return super(Select, self).__exit__(exc_type, exc_val, exc_tb)
+
+
 def make_channel(dtype, capacity=0):
     """
     Helps implementation of a concurrent program by creating a "channel" of
@@ -131,7 +332,7 @@ def make_channel(dtype, capacity=0):
     return channel
 
 
-def channel_send(channel, value, copy=False):
+def channel_send(channel, value, is_copy=False):
     """
     Sends a value through a channel variable. Used by an unbuffered or buffered
     channel to pass data from within or to a concurrent Go block, where
@@ -141,8 +342,8 @@ def channel_send(channel, value, copy=False):
         channel (Variable|Channel): Channel variable created using
         `make_channel`.
         value (Variable): Value to send to channel
-        copy (bool): Copy data while channel send. If False, then data
-        is moved. The input cannot be used after move.
+        is_copy (bool): Copy data while channel send. If False, then data
+        is moved. The input cannot be used after move. (default False)
     Returns:
         Variable: The boolean status on whether or not the channel
                   successfully sent the passed value.
@@ -159,35 +360,26 @@ def channel_send(channel, value, copy=False):
     main_program = helper.main_program
     channel_send_block = main_program.current_block()
 
-    status = helper.create_variable(
-        name=unique_name.generate('status'),
-        type=core.VarDesc.VarType.LOD_TENSOR,
-        dtype=core.VarDesc.VarType.BOOL)
-
     X = value
 
-    if copy is True:
+    if is_copy:
         copied_X = helper.create_variable(
             name=unique_name.generate(value.name + '_copy'),
             type=value.type,
             dtype=value.dtype,
             shape=value.shape,
             lod_level=value.lod_level,
-            capacity=value.capacity)
+            capacity=value.capacity if hasattr(value, 'capacity') else None)
 
         assign_op = channel_send_block.append_op(
-            type="assign_op", inputs={"X": value}, outputs={"Out": copied_X})
+            type="assign", inputs={"X": value}, outputs={"Out": copied_X})
         X = copied_X
 
-    channel_send_op = channel_send_block.append_op(
-        type="channel_send",
-        inputs={
+    channel_send_block.append_op(
+        type="channel_send", inputs={
             "Channel": channel,
             "X": X,
-        },
-        outputs={"Status": status})
-
-    return status
+        })
 
 
 def channel_recv(channel, return_value):

python/paddle/fluid/debuger.py

@@ -16,7 +16,6 @@ import sys
 import re
 from graphviz import GraphPreviewGenerator
 import proto.framework_pb2 as framework_pb2
-import paddle.fluid.core as core
 
 _vartype2str_ = [
     "UNK",
@@ -126,7 +125,6 @@ def pprint_block_codes(block_desc, show_backward=False):
     def is_var_backward(var_desc):
         return "@GRAD" in var_desc.name
 
-    #print(type(block_desc))
     if type(block_desc) is not framework_pb2.BlockDesc:
         block_desc = framework_pb2.BlockDesc.FromString(
             block_desc.serialize_to_string())

python/paddle/fluid/distribute_transpiler.py

@@ -20,6 +20,7 @@ from layer_helper import LayerHelper
 from distributed_spliter import *
 import math
 from . import core
+import debuger
 
 
 class VarBlock:
@@ -289,6 +290,7 @@ class DistributeTranspiler:
                     dtype=v.dtype,
                     shape=v.shape)
                 recv_inputs.append(var)
+
         # step3
         optimize_block = pserver_program.create_block(0)
         # step 4
@@ -307,15 +309,57 @@ class DistributeTranspiler:
         # Iterate through the ops, and if an op and the optimize ops
         # which located on current pserver are in one set, then 
         # append it into the sub program.
-        for _, op in enumerate(self.optimize_ops):
-            for _, opt_op in enumerate(opt_op_on_pserver):
-                if ufind.is_connected(op, opt_op):
+
+        # We try to put optimization program run parallelly, assume
+        # optimization program always looks like:
+        #
+        # prevop -> prevop -> opt op -> following op -> following op; ->
+        # prevop -> prevop -> opt op -> following op -> following op; ->
+        # global op -> global op
+        #
+        # we put operators that can run parallelly to many program blocks.
+        # in above example, we seperate ops by the ";". Global ops must run
+        # after all the optimize ops finished.
+
+        global_ops = []
+        # HACK: optimization global ops only used to scale beta1 and beta2
+        # replace it with dependency engine.
+        for op in self.optimize_ops:
+            if op.type == "scale":
+                for in_name in op.input_arg_names:
+                    if in_name.startswith("beta1_pow_acc") or\
+                        in_name.startswith("beta2_pow_acc"):
+                        global_ops.append(op)
+
+        def __append_optimize_op__(op, block):
             if self._is_opt_op(op):
-                        self._append_pserver_ops(optimize_block, op, endpoint,
+                self._append_pserver_ops(block, op, endpoint,
                                          default_main_program())
             else:
-                        self._append_pserver_non_opt_ops(optimize_block, op)
-                    break
+                self._append_pserver_non_opt_ops(block, op)
+
+        # append op to the current block
+        per_opt_block = optimize_block
+        for _, opt_op in enumerate(opt_op_on_pserver):
+            for _, op in enumerate(self.optimize_ops):
+                # optimizer is connected to itself
+                if ufind.is_connected(op, opt_op) and \
+                    op not in global_ops:
+                    __append_optimize_op__(op, per_opt_block)
+            per_opt_block = pserver_program.create_block(0)
+
+        # append global ops
+        for glb_op in global_ops:
+            __append_optimize_op__(glb_op, per_opt_block)
+
+        # NOT USED: single block version:
+        #
+        # for _, op in enumerate(self.optimize_ops):
+        #     for _, opt_op in enumerate(opt_op_on_pserver):
+        #         if ufind.is_connected(op, opt_op):
+        #             __append_optimize_op__(glb_op, optimize_block)
+        #             break
+
         # step5 append the listen_and_serv op
         pserver_program.global_block().append_op(
             type="listen_and_serv",
@@ -521,6 +565,8 @@ class DistributeTranspiler:
         orig_var_name = ""
         if suff_idx >= 0:
             orig_var_name = varname[:suff_idx]
+        else:
+            orig_var_name = varname
         return orig_var_name
 
     def _append_pserver_ops(self, optimize_block, opt_op, endpoint,
@@ -535,7 +581,8 @@ class DistributeTranspiler:
                 grad_block = None
                 for g in self.param_grad_ep_mapping[endpoint]["grads"]:
                     if same_or_split_var(
-                            self._orig_varname(g.name), opt_op.input(key)[0]):
+                            self._orig_varname(g.name),
+                            self._orig_varname(opt_op.input(key)[0])):
                         grad_block = g
                         break
                 if not grad_block:
@@ -660,10 +707,22 @@ class DistributeTranspiler:
         # If one op's input is another op's output or
         # one op's output is another op's input, we say
         # the two operator is connected.
-        op1_input_names = op1.desc.input_arg_names()
+        def _append_inname_remove_beta(varname_list):
+            op_input_names = []
+            for in_name in varname_list:
+                # HACK: remove beta1 and beta2 to avoid let all
+                # ops connected.
+                if in_name.startswith("beta2_pow_acc") or \
+                    in_name.startswith("beta1_pow_acc"):
+                    continue
+                else:
+                    op_input_names.append(in_name)
+            return op_input_names
+
+        op1_input_names = _append_inname_remove_beta(op1.desc.input_arg_names())
         op1_output_names = op1.desc.output_arg_names()
 
-        op2_input_names = op2.desc.input_arg_names()
+        op2_input_names = _append_inname_remove_beta(op2.desc.input_arg_names())
         op2_output_names = op2.desc.output_arg_names()
 
         if set(op1_output_names) & set(op2_input_names) or \
@@ -694,7 +753,7 @@ class DistributeTranspiler:
         param_names = [
             p.name for p in self.param_grad_ep_mapping[endpoint]["params"]
         ]
-        if op.input("Param") in param_names:
+        if op.input("Param")[0] in param_names:
             return True
         else:
             for n in param_names:

python/paddle/fluid/executor.py

@@ -235,66 +235,17 @@ class Executor(object):
             tensor.set_lod(lod)
             return tensor
 
-    def run(self,
-            program=None,
-            feed=None,
-            fetch_list=None,
-            feed_var_name='feed',
-            fetch_var_name='fetch',
-            scope=None,
-            return_numpy=True,
-            use_program_cache=False):
-        """ Run program by this Executor. Feed data by feed map, fetch result by fetch_list.
-
-        Python executor takes a program, add feed operators and fetch operators to this program according
-        to feed map and fetch_list. Feed map provides input data for the program. fetch_list provides
-        the variables(or names) that user want to get after program run. Note: the executor will run all
-        operators in the program but not only the operators dependent by the fetch_list
-
-        :param program: the program that need to run, if not provied, then default_main_program will be used.
-        :param feed: feed variable map, e.g. {"image": ImageData, "label": LableData}
-        :param fetch_list: a list of variable or variable names that user want to get, run will return them according
-        to this list.
-        :param feed_var_name: the name for the input variable of feed Operator.
-        :param fetch_var_name: the name for the output variable of feed Operator.
-        :param scope: the scope used to run this program, you can switch it to different scope. default is global_scope
-        :param return_numpy: if convert the fetched tensor to numpy
-        :param use_program_cache: set use_program_cache to true if program not changed compare to the last step.
-        :return: result according to fetch_list.
-        """
-        if feed is None:
-            feed = {}
-        if not isinstance(feed, dict):
-            raise TypeError("feed should be a map")
-        if fetch_list is None:
-            fetch_list = []
-
-        if program is None:
-            program = default_main_program()
+    def _get_program_cache(self, program_cache_key):
+        return self.program_caches.get(program_cache_key, None)
 
-        if not isinstance(program, Program):
-            raise TypeError()
-
-        if scope is None:
-            scope = global_scope()
-
-        program_cache = None
-        program_cache_key = get_program_cache_key(feed, fetch_list)
+    def _add_program_cache(self, program_cache_key, program):
+        self.program_caches[program_cache_key] = program
 
-        if use_program_cache:
-            # find program cache by cache_key
-            program_cache = self.program_caches.get(program_cache_key, None)
-            # TODO(qiao): Should check program_cache and program are exactly the same.
-        else:
-            self.program_caches.pop(program_cache_key, None)
-
-        if program_cache is None:
-            program_cache = program.clone()
-
-            if use_program_cache:
-                self.program_caches[program_cache_key] = program_cache
+    def _add_feed_fetch_ops(self, program, feed, fetch_list, feed_var_name,
+                            fetch_var_name):
+        tmp_program = program.clone()
 
-            global_block = program_cache.global_block()
+        global_block = tmp_program.global_block()
 
         if feed_var_name in global_block.vars:
             feed_var = global_block.var(feed_var_name)
@@ -323,8 +274,7 @@ class Executor(object):
                     attrs={'col': i})
 
         # append fetch_operators
-            if not has_fetch_operators(global_block, fetch_list,
-                                       fetch_var_name):
+        if not has_fetch_operators(global_block, fetch_list, fetch_var_name):
             for i, var in enumerate(fetch_list):
                 assert isinstance(var, Variable) or isinstance(var, str), (
                     "Wrong type for fetch_list[%s]: %s" % (i, type(var)))
@@ -334,8 +284,11 @@ class Executor(object):
                     outputs={'Out': [fetch_var]},
                     attrs={'col': i})
 
+        return tmp_program
+
+    def _feed_data(self, program, feed, feed_var_name, scope):
         # feed var to framework
-        for op in program_cache.global_block().ops:
+        for op in program.global_block().ops:
             if op.desc.type() == 'feed':
                 feed_target_name = op.desc.output('Out')[0]
                 cur_feed = feed[feed_target_name]
@@ -346,11 +299,79 @@ class Executor(object):
             else:
                 break
 
-        self.executor.run(program_cache.desc, scope, 0, True, True)
+    def _fetch_data(self, fetch_list, fetch_var_name, scope):
         outs = [
             core.get_fetch_variable(scope, fetch_var_name, i)
             for i in xrange(len(fetch_list))
         ]
+        return outs
+
+    def run(self,
+            program=None,
+            feed=None,
+            fetch_list=None,
+            feed_var_name='feed',
+            fetch_var_name='fetch',
+            scope=None,
+            return_numpy=True,
+            use_program_cache=False):
+        """ Run program by this Executor. Feed data by feed map, fetch result by fetch_list.
+
+        Python executor takes a program, add feed operators and fetch operators to this program according
+        to feed map and fetch_list. Feed map provides input data for the program. fetch_list provides
+        the variables(or names) that user want to get after program run. Note: the executor will run all
+        operators in the program but not only the operators dependent by the fetch_list
+
+        :param program: the program that need to run, if not provied, then default_main_program will be used.
+        :param feed: feed variable map, e.g. {"image": ImageData, "label": LableData}
+        :param fetch_list: a list of variable or variable names that user want to get, run will return them according
+        to this list.
+        :param feed_var_name: the name for the input variable of feed Operator.
+        :param fetch_var_name: the name for the output variable of feed Operator.
+        :param scope: the scope used to run this program, you can switch it to different scope. default is global_scope
+        :param return_numpy: if convert the fetched tensor to numpy
+        :param use_program_cache: set use_program_cache to true if program not changed compare to the last step.
+        :return: result according to fetch_list.
+        """
+        if feed is None:
+            feed = {}
+        if not isinstance(feed, dict):
+            raise TypeError("feed should be a map")
+        if fetch_list is None:
+            fetch_list = []
+        if program is None:
+            program = default_main_program()
+
+        if not isinstance(program, Program):
+            raise TypeError()
+
+        if scope is None:
+            scope = global_scope()
+
+        cache_key = get_program_cache_key(feed, fetch_list)
+        if use_program_cache:
+            cached_program = self._get_program_cache(cache_key)
+            if cached_program is None:
+                cached_program = self._add_feed_fetch_ops(
+                    program=program,
+                    feed=feed,
+                    fetch_list=fetch_list,
+                    feed_var_name=feed_var_name,
+                    fetch_var_name=fetch_var_name)
+                self._add_program_cache(cache_key, cached_program)
+            program = cached_program
+        else:
+            self.program_caches.pop(cache_key, None)
+            program = self._add_feed_fetch_ops(
+                program=program,
+                feed=feed,
+                fetch_list=fetch_list,
+                feed_var_name=feed_var_name,
+                fetch_var_name=fetch_var_name)
+
+        self._feed_data(program, feed, feed_var_name, scope)
+        self.executor.run(program.desc, scope, 0, True, True)
+        outs = self._fetch_data(fetch_list, fetch_var_name, scope)
         if return_numpy:
             outs = as_numpy(outs)
         return outs

python/paddle/fluid/framework.py

@@ -918,6 +918,24 @@ class Block(object):
                 name=v.name)
             self.vars[new_p.name] = new_p
 
+    def clone_variable(self, var):
+        """
+        Clone a variable into current block.
+        Args:
+            var: the variable to be cloned.
+
+        Returns:
+            The new  variable cloned from 'var' in current block.
+        """
+        assert isinstance(var, Variable)
+        return self.create_var(
+            name=var.name,
+            shape=var.shape,
+            dtype=var.dtype,
+            type=var.type,
+            lod_level=var.lod_level,
+            persistable=True)
+
 
 class Program(object):
     def __init__(self):

python/paddle/fluid/layer_helper.py

@@ -399,7 +399,12 @@ class LayerHelper(object):
         if isinstance(act, basestring):
             act = {'type': act}
         tmp = self.create_tmp_variable(dtype=input_var.dtype)
+
+        if 'use_mkldnn' in self.kwargs:
+            act['use_mkldnn'] = self.kwargs.get('use_mkldnn')
         act_type = act.pop('type')
+        if 'use_mkldnn' in self.kwargs:
+            act['use_mkldnn'] = self.kwargs.get('use_mkldnn')
         self.append_op(
             type=act_type,
             inputs={"X": [input_var]},

python/paddle/fluid/layers/control_flow.py

@@ -16,7 +16,7 @@ import contextlib
 from layer_function_generator import autodoc
 from tensor import assign, fill_constant
 from .. import core
-from ..framework import Program, Variable, Operator, Block
+from ..framework import Program, Variable, Operator
 from ..layer_helper import LayerHelper, unique_name
 from ops import logical_and, logical_not, logical_or
 
@@ -29,7 +29,6 @@ __all__ = [
     'WhileGuard',
     'While',
     'Switch',
-    'Select',
     'lod_rank_table',
     'max_sequence_len',
     'topk',
@@ -1212,186 +1211,6 @@ class Switch(object):
         return True
 
 
-class SelectCase(object):
-    DEFAULT = 0
-    SEND = 1
-    RECEIVE = 2
-
-    def __init__(self,
-                 case_idx,
-                 case_to_execute,
-                 channel_action_fn=None,
-                 channel=None,
-                 value=None):
-        self.helper = LayerHelper('conditional_block')
-        self.main_program = self.helper.main_program
-        self.is_scalar_condition = True
-
-        self.case_to_execute = case_to_execute
-        self.idx = case_idx
-
-        # Since we aren't going to use the `channel_send` or `channel_recv`
-        # functions directly, we just need to capture the name.
-        self.action = (self.SEND
-                       if channel_action_fn.__name__ == ('channel_send') else
-                       self.RECEIVE) if channel_action_fn else (self.DEFAULT)
-        self.value = value
-        self.channel = channel
-
-    def __enter__(self):
-        self.block = self.main_program.create_block()
-
-    def construct_op(self):
-        main_program = self.helper.main_program
-        cases_block = main_program.current_block()
-
-        inner_outputs = set()
-        input_set = set()
-        params = set()
-
-        for op in self.block.ops:
-            # Iterate over all operators, get all the inputs
-            # and add as input to the SelectCase operator.
-            for iname in op.input_names:
-                for in_var_name in op.input(iname):
-                    if in_var_name not in inner_outputs:
-                        input_set.add(in_var_name)
-
-            for oname in op.output_names:
-                for out_var_name in op.output(oname):
-                    inner_outputs.add(out_var_name)
-
-        param_list = [
-            cases_block.var(each_name) for each_name in params
-            if each_name not in input_set
-        ]
-
-        # Iterate over all operators, get all the outputs
-        # add to the output list of SelectCase operator only if
-        # they exist in the parent block.
-        out_vars = []
-        for inner_out_name in inner_outputs:
-            if inner_out_name in cases_block.vars:
-                out_vars.append(cases_block.var(inner_out_name))
-
-        # First, create an op that will determine whether or not this is the
-        # conditional variable to execute.
-        should_execute_block = equal(
-            fill_constant(
-                shape=[1], dtype=core.VarDesc.VarType.INT32, value=self.idx),
-            self.case_to_execute)
-
-        step_scope = cases_block.create_var(
-            type=core.VarDesc.VarType.STEP_SCOPES)
-
-        cases_block.append_op(
-            type='conditional_block',
-            inputs={'X': [should_execute_block],
-                    'Params': param_list},
-            outputs={'Out': out_vars,
-                     'Scope': [step_scope]},
-            attrs={
-                'sub_block': self.block,
-                'is_scalar_condition': self.is_scalar_condition
-            })
-
-        return '%s,%s,%s,%s' % (self.idx, self.action, self.channel.name
-                                if self.channel else '', self.value.name
-                                if self.value else '')
-
-    def __exit__(self, exc_type, exc_val, exc_tb):
-        self.main_program.rollback()
-        if exc_type is not None:
-            return False  # re-raise exception
-        return True
-
-
-class Select(BlockGuard):
-    def __init__(self, name=None):
-        self.helper = LayerHelper('select', name=name)
-        self.cases = []
-
-        super(Select, self).__init__(self.helper.main_program)
-        self.case_to_execute = fill_constant(
-            shape=[1], dtype=core.VarDesc.VarType.INT32, value=-1)
-
-    def __enter__(self):
-        super(Select, self).__enter__()
-        return self
-
-    def case(self, channel_action_fn, channel, value):
-        """Create a new block for this condition.
-        """
-        select_case = SelectCase(
-            len(self.cases), self.case_to_execute, channel_action_fn, channel,
-            value)
-
-        self.cases.append(select_case)
-
-        return select_case
-
-    def default(self):
-        """Create a default case block for this condition.
-        """
-        default_case = SelectCase(len(self.cases), self.case_to_execute)
-
-        self.cases.append(default_case)
-
-        return default_case
-
-    def __exit__(self, exc_type, exc_val, exc_tb):
-        if exc_type is not None:
-            return False
-
-        # Create a select op and another block to wrap its
-        # case blocks.
-        select_block = self.helper.main_program.current_block()
-        parent_block = self.helper.main_program.block(select_block.parent_idx)
-
-        # Construct each case op, inside the newly created select block.
-        serialized_cases = []
-        for case in self.cases:
-            serialized_cases.append(case.construct_op())
-
-        intermediate = set()
-        params = set()
-
-        for case_block in select_block.ops:
-            if case_block.attrs and 'sub_block' in case_block.attrs:
-                for each_op in case_block.attrs['sub_block'].ops:
-                    assert isinstance(each_op, Operator)
-                    for iname in each_op.input_names:
-                        for in_var_name in each_op.input(iname):
-                            if in_var_name not in intermediate:
-                                params.add(in_var_name)
-
-                    for oname in each_op.output_names:
-                        for out_var_name in each_op.output(oname):
-                            intermediate.add(out_var_name)
-
-        # TODO(varunarora): Figure out if defining output is needed.
-        out_list = [
-            parent_block.var(var_name) for var_name in parent_block.vars
-            if var_name in intermediate
-        ]
-
-        X = [select_block.var_recursive(x_name) for x_name in params]
-
-        # Needs to be used by `equal` inside the cases block.
-        X.append(self.case_to_execute)
-
-        # Construct the select op.
-        parent_block.append_op(
-            type='select',
-            inputs={'X': X,
-                    'case_to_execute': self.case_to_execute},
-            attrs={'sub_block': select_block,
-                   'cases': serialized_cases},
-            outputs={})
-
-        return super(Select, self).__exit__(exc_type, exc_val, exc_tb)
-
-
 class IfElseBlockGuard(object):
     def __init__(self, is_true, ifelse):
         if not isinstance(ifelse, IfElse):

python/paddle/fluid/layers/detection.py

@@ -129,13 +129,12 @@ def detection_output(loc,
         prior_box_var=prior_box_var,
         target_box=loc,
         code_type='decode_center_size')
-
     old_shape = scores.shape
     scores = ops.reshape(x=scores, shape=(-1, old_shape[-1]))
     scores = nn.softmax(input=scores)
     scores = ops.reshape(x=scores, shape=old_shape)
     scores = nn.transpose(scores, perm=[0, 2, 1])
-
+    scores.stop_gradient = True
     nmsed_outs = helper.create_tmp_variable(dtype=decoded_box.dtype)
     helper.append_op(
         type="multiclass_nms",
@@ -150,6 +149,7 @@ def detection_output(loc,
             'score_threshold': score_threshold,
             'nms_eta': 1.0
         })
+    nmsed_outs.stop_gradient = True
     return nmsed_outs
 
 
@@ -475,6 +475,7 @@ def ssd_loss(location,
     # 2. Compute confidence for mining hard examples
     # 2.1. Get the target label based on matched indices
     gt_label = ops.reshape(x=gt_label, shape=gt_label.shape + (1, ))
+    gt_label.stop_gradient = True
     target_label, _ = target_assign(
         gt_label, matched_indices, mismatch_value=background_label)
     # 2.2. Compute confidence loss.
@@ -482,10 +483,12 @@ def ssd_loss(location,
     confidence = __reshape_to_2d(confidence)
     target_label = tensor.cast(x=target_label, dtype='int64')
     target_label = __reshape_to_2d(target_label)
+    target_label.stop_gradient = True
     conf_loss = nn.softmax_with_cross_entropy(confidence, target_label)
 
     # 3. Mining hard examples
     conf_loss = ops.reshape(x=conf_loss, shape=(num, num_prior))
+    conf_loss.stop_gradient = True
     neg_indices = helper.create_tmp_variable(dtype='int32')
     dtype = matched_indices.dtype
     updated_matched_indices = helper.create_tmp_variable(dtype=dtype)
@@ -695,6 +698,8 @@ def multi_box_head(inputs,
             outputs={"Boxes": box,
                      "Variances": var},
             attrs=attrs, )
+        box.stop_gradient = True
+        var.stop_gradient = True
         return box, var
 
     def _reshape_with_axis_(input, axis=1):
@@ -834,4 +839,6 @@ def multi_box_head(inputs,
         mbox_locs_concat = tensor.concat(mbox_locs, axis=1)
         mbox_confs_concat = tensor.concat(mbox_confs, axis=1)
 
+    box.stop_gradient = True
+    var.stop_gradient = True
     return mbox_locs_concat, mbox_confs_concat, box, var

python/paddle/fluid/layers/io.py

@@ -21,7 +21,8 @@ from ..executor import global_scope
 
 __all__ = [
     'data', 'BlockGuardServ', 'ListenAndServ', 'Send', 'open_recordio_file',
-    'read_file', 'create_shuffle_reader', 'create_double_buffer_reader'
+    'open_files', 'read_file', 'create_shuffle_reader',
+    'create_double_buffer_reader', 'create_multi_pass_reader'
 ]
 
 
@@ -112,9 +113,9 @@ class ListenAndServ(object):
     which can receive variables from clients and run a block.
     """
 
-    def __init__(self, endpoint, fan_in=1, optimizer_mode=True):
+    def __init__(self, endpoint, inputs, fan_in=1, optimizer_mode=True):
         self.helper = LayerHelper("listen_and_serv")
-        self.inputs = []
+        self.inputs = inputs
         self.outputs = []
         self.endpoint = endpoint
         self.fan_in = fan_in
@@ -159,18 +160,13 @@ class ListenAndServ(object):
         current_block = main_program.current_block()
         parent_block = self.parent_block()
 
-        params, grads = self.get_params_and_grads()
-        param_names = [p.name for p in params]
-        grad_names = [g.name for g in grads]
         parent_block.append_op(
             type='listen_and_serv',
-            inputs={},
+            inputs={"X": self.inputs},
             outputs={},
             attrs={
                 'endpoint': self.endpoint,
                 'Fanin': self.fan_in,
-                'ParamList': param_names,
-                'GradList': grad_names,
                 'OptimizeBlock': current_block
             })
 
@@ -195,10 +191,14 @@ def Send(endpoints, send_vars, get_vars):
     endpoints = list(set(epmap))
 
     helper = LayerHelper("Send", **locals())
+    rpc_client_var = default_main_program().global_block().create_var(
+        name="RPC_CLIENT_VAR", persistable=True, type=core.VarDesc.VarType.RAW)
+
     helper.append_op(
         type="send",
         inputs={"X": send_vars},
-        outputs={"Out": get_vars},
+        outputs={"Out": get_vars,
+                 "RPCClient": rpc_client_var},
         attrs={"endpoints": endpoints,
                "epmap": epmap})
 
@@ -287,6 +287,36 @@ def open_recordio_file(filename, shapes, lod_levels, dtypes):
                              startup_var)
 
 
+def open_files(filenames, thread_num, shapes, lod_levels, dtypes):
+    dtypes = [convert_np_dtype_to_dtype_(dt) for dt in dtypes]
+    shape_concat = []
+    ranks = []
+
+    for shape in shapes:
+        shape_concat.extend(shape)
+        ranks.append(len(shape))
+
+    var_name = unique_name('multiple_reader')
+
+    startup_blk = default_startup_program().current_block()
+    startup_var = startup_blk.create_var(name=var_name)
+    startup_blk.append_op(
+        type='open_files',
+        outputs={'Out': [startup_var]},
+        attrs={
+            'shape_concat': shape_concat,
+            'lod_levels': lod_levels,
+            'ranks': ranks,
+            'file_names': filenames,
+            'thread_num': thread_num
+        })
+
+    startup_var.desc.set_dtypes(dtypes)
+    startup_var.persistable = True
+    return _copy_reader_var_(default_main_program().current_block(),
+                             startup_var)
+
+
 def __create_decorated_reader__(op_type, reader, attrs):
     var_name = unique_name(op_type)
     startup_blk = default_startup_program().current_block()
@@ -314,6 +344,11 @@ def create_double_buffer_reader(reader, place=None):
                                        attrs)
 
 
+def create_multi_pass_reader(reader, pass_num):
+    return __create_decorated_reader__('create_multi_pass_reader', reader,
+                                       {'pass_num': int(pass_num)})
+
+
 def read_file(file_obj):
     helper = LayerHelper('read_file')
     out = [

python/paddle/fluid/layers/nn.py

@@ -74,6 +74,7 @@ __all__ = [
     'one_hot',
     'autoincreased_step_counter',
     'lod_reset',
+    'lrn',
 ]
 
 
@@ -82,6 +83,7 @@ def fc(input,
        num_flatten_dims=1,
        param_attr=None,
        bias_attr=None,
+       use_mkldnn=False,
        act=None,
        name=None):
     """
@@ -163,8 +165,11 @@ def fc(input,
             inputs={"X": input_var,
                     "Y": w},
             outputs={"Out": tmp},
-            attrs={"x_num_col_dims": num_flatten_dims,
-                   "y_num_col_dims": 1})
+            attrs={
+                "x_num_col_dims": num_flatten_dims,
+                "y_num_col_dims": 1,
+                'use_mkldnn': use_mkldnn
+            })
         mul_results.append(tmp)
 
     # sum
@@ -1117,12 +1122,14 @@ def conv2d(input,
            filter_size,
            stride=1,
            padding=0,
+           dilation=1,
            groups=None,
            param_attr=None,
            bias_attr=None,
            use_cudnn=True,
            use_mkldnn=False,
-           act=None):
+           act=None,
+           name=None):
     """
     **Convlution2D Layer**
 
@@ -1183,6 +1190,9 @@ def conv2d(input,
        padding(int|tuple): The padding size. If padding is a tuple, it must
            contain two integers, (padding_H, padding_W). Otherwise, the
            padding_H = padding_W = padding. Default: padding = 0.
+       dilation(int|tuple): The dilation size. If dilation is a tuple, it must
+           contain two integers, (dilation_H, dilation_W). Otherwise, the
+           dilation_H = dilation_W = dilation. Default: dilation = 1.
        groups(int): The groups number of the Conv2d Layer. According to grouped
            convolution in Alex Krizhevsky's Deep CNN paper: when group=2,
            the first half of the filters is only connected to the first half
@@ -1193,6 +1203,8 @@ def conv2d(input,
        use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn
            library is installed. Default: True
        act(str): Activation type. Default: None
+       name(str|None): A name for this layer(optional). If set None, the layer
+           will be named automatically.
 
     Returns:
         Variable: The tensor variable storing the convolution and \
@@ -1233,6 +1245,7 @@ def conv2d(input,
     filter_size = utils.convert_to_list(filter_size, 2, 'filter_size')
     stride = utils.convert_to_list(stride, 2, 'stride')
     padding = utils.convert_to_list(padding, 2, 'padding')
+    dilation = utils.convert_to_list(dilation, 2, 'dilation')
 
     if not isinstance(use_cudnn, bool):
         raise ValueError("use_cudnn should be True or False")
@@ -1262,6 +1275,7 @@ def conv2d(input,
         attrs={
             'strides': stride,
             'paddings': padding,
+            'dilations': dilation,
             'groups': groups,
             'use_cudnn': use_cudnn,
             'use_mkldnn': use_mkldnn
@@ -1670,7 +1684,9 @@ def conv2d_transpose(input,
                      stride=1,
                      dilation=1,
                      param_attr=None,
+                     bias_attr=None,
                      use_cudnn=True,
+                     act=None,
                      name=None):
     """
     **Convlution2D transpose layer**
@@ -1739,8 +1755,10 @@ def conv2d_transpose(input,
            dilation_H = dilation_W = dilation. Default: dilation = 1.
        param_attr(ParamAttr): The parameters to the Conv2d_transpose Layer.
                               Default: None
+       bias_attr(ParamAttr): Bias parameter for the Conv2d layer. Default: None
        use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn
            library is installed. Default: True
+       act(str): Activation type. Default: None
        name(str|None): A name for this layer(optional). If set None, the layer
            will be named automatically.
 
@@ -1793,12 +1811,12 @@ def conv2d_transpose(input,
     img_filter = helper.create_parameter(
         dtype=input.dtype, shape=filter_shape, attr=helper.param_attr)
 
-    out = helper.create_tmp_variable(dtype=input.dtype)
+    pre_bias = helper.create_tmp_variable(dtype=input.dtype)
     helper.append_op(
         type='conv2d_transpose',
         inputs={'Input': [input],
                 'Filter': [img_filter]},
-        outputs={'Output': out},
+        outputs={'Output': pre_bias},
         attrs={
             'strides': stride,
             'paddings': padding,
@@ -1806,53 +1824,55 @@ def conv2d_transpose(input,
             'use_cudnn': use_cudnn
         })
 
+    pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2)
+    out = helper.append_activation(pre_act)
     return out
 
 
-def sequence_expand(x, y, name=None):
+def sequence_expand(x, y, ref_level=-1, name=None):
     """Sequence Expand Layer. This layer will expand the input variable **x**
-    according to LoD information of **y**. And the following examples will
-    explain how sequence_expand works:
+    according to specified level lod of **y**. Please note that lod level of
+    **x** is at most 1 and rank of **x** is at least 2. When rank of **x**
+    is greater than 2, then it would be viewed as a 2-D tensor.
+    Following examples will explain how sequence_expand works:
 
     .. code-block:: text
 
         * Case 1
             x is a LoDTensor:
-                x.lod = [[0,       2, 3],
-                         [0, 1,    3, 4]]
-                x.data = [a, b, c, d]
+                x.lod  = [[0,   2,        4]]
+                x.data = [[a], [b], [c], [d]]
                 x.dims = [4, 1]
 
             y is a LoDTensor:
                 y.lod = [[0,    2,    4],
                          [0, 3, 6, 7, 8]]
 
-            with condition len(y.lod[-1]) - 1 == x.dims[0]
+            ref_level: 0
 
-            then output is a 2-level LoDTensor:
-                out.lod = [[0,                2,    4],
-                           [0,       3,       6, 7, 8]]
-                out.data = [a, a, a, b, b, b, c, d]
+            then output is a 1-level LoDTensor:
+                out.lod =  [[0,   2,        4,        6,        8]]
+                out.data = [[a], [b], [a], [b], [c], [d], [c], [d]]
                 out.dims = [8, 1]
 
         * Case 2
             x is a Tensor:
-                x.data = [a, b, c]
+                x.data = [[a], [b], [c]]
                 x.dims = [3, 1]
 
             y is a LoDTensor:
-                y.lod = [[0, 2, 3, 6]]
-
-            with condition len(y.lod[-1]) - 1 == x.dims[0]
+                y.lod = [[0, 2, 2, 5]]
 
-            then output is a 1-level LoDTensor:
-                out.lod = [[0,    2, 3,      6]]
-                out.data = [a, a, b, c, c, c]
-                out.dims = [6, 1]
+            ref_level: -1
 
+            then output is a Tensor:
+                out.data = [[a], [a], [c], [c], [c]]
+                out.dims = [5, 1]
     Args:
         x (Variable): The input variable which is a Tensor or LoDTensor.
         y (Variable): The input variable which is a LoDTensor.
+        ref_level (int): Lod level of `y` to be referred by `x`. If set to -1,
+                         refer the last level of lod.
         name(str|None): A name for this layer(optional). If set None, the layer
                         will be named automatically.
 
@@ -1865,14 +1885,17 @@ def sequence_expand(x, y, name=None):
             x = fluid.layers.data(name='x', shape=[10], dtype='float32')
             y = fluid.layers.data(name='y', shape=[10, 20],
                              dtype='float32', lod_level=1)
-            out = layers.sequence_expand(x=x, y=y)
+            out = layers.sequence_expand(x=x, y=y, ref_level=0)
     """
     helper = LayerHelper('sequence_expand', input=x, **locals())
     dtype = helper.input_dtype()
     tmp = helper.create_tmp_variable(dtype)
     helper.append_op(
-        type='sequence_expand', inputs={'X': x,
-                                        'Y': y}, outputs={'Out': tmp})
+        type='sequence_expand',
+        inputs={'X': x,
+                'Y': y},
+        outputs={'Out': tmp},
+        attrs={'ref_level': ref_level})
     return tmp
 
 
@@ -3388,3 +3411,73 @@ def lod_reset(x, y=None, target_lod=None):
         raise ValueError("y and target_lod should not be both None.")
 
     return out
+
+
+def lrn(input, n=5, k=1.0, alpha=1e-4, beta=0.75, name=None):
+    """
+    Local Response Normalization Layer. This layer performs a type of
+    "lateral inhibition" by normalizing over local input regions.
+
+    The formula is as follows:
+
+    .. math::
+
+        Output(i, x, y) = Input(i, x, y) / \left(
+        k + \alpha \sum\limits^{\min(C, c + n/2)}_{j = \max(0, c - n/2)}
+        (Input(j, x, y))^2 \right)^{\beta}
+
+    In the above equation:
+
+    * :math:`n`: The number of channels to sum over.
+    * :math:`k`: The offset (avoid being divided by 0).
+    * :math:`alpha`: The scaling parameter.
+    * :math:`beta`: The exponent parameter.
+
+    Refer to `ImageNet Classification with Deep Convolutional Neural Networks
+    <https://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks.pdf>`_
+
+    Args:
+        input (Variable): The input tensor of this layer, and the dimension of input tensor must be 4.
+        n (int, default 5): The number of channels to sum over.
+        k (float, default 1.0): An offset (usually positive to avoid dividing by 0).
+        alpha (float, default 1e-4): The scaling parameter.
+        beta (float, default 0.75): The exponent.
+        name (str, default None): A name for this operation.
+
+    Raises:
+        ValueError: If rank of the input tensor is not 4.
+
+    Returns:
+        A tensor variable storing the transformation result.
+
+    Examples:
+        .. code-block:: python
+
+          data = fluid.layers.data(name="data", shape=[3, 112, 112], dtype="float32")
+          lrn = fluid.layers.lrn(input=data)
+    """
+    helper = LayerHelper('lrn', **locals())
+    dtype = helper.input_dtype()
+    input_shape = input.shape
+    dims = len(input_shape)
+
+    if dims != 4:
+        raise ValueError(
+            "dims of input must be 4(not %d), and it's order must be NCHW" %
+            (dims))
+
+    mid_out = helper.create_tmp_variable(dtype=dtype, stop_gradient=True)
+    lrn_out = helper.create_tmp_variable(dtype)
+    helper.append_op(
+        type="lrn",
+        inputs={"X": input},
+        outputs={
+            "Out": lrn_out,
+            "MidOut": mid_out,
+        },
+        attrs={"n": n,
+               "k": k,
+               "alpha": alpha,
+               "beta": beta})
+
+    return lrn_out

python/paddle/fluid/layers/ops.py

@@ -69,6 +69,7 @@ __all__ = [
     'gaussian_random_batch_size_like',
     'cumsum',
     'scatter',
+    'sum',
 ] + __activations__
 
 for _OP in set(__all__):

python/paddle/fluid/optimizer.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 from collections import defaultdict
-
+from paddle.fluid.framework import Program
 import framework
 import layers
 from backward import append_backward
@@ -23,8 +23,12 @@ from initializer import Constant
 from layer_helper import LayerHelper
 from regularizer import append_regularization_ops
 from clip import append_gradient_clip_ops, error_clip_callback
+from contextlib import contextmanager
 
-__all__ = ['SGD', 'Momentum', 'Adagrad', 'Adam', 'Adamax', 'DecayedAdagrad']
+__all__ = [
+    'SGD', 'Momentum', 'Adagrad', 'Adam', 'Adamax', 'DecayedAdagrad',
+    'Adadelta', 'ModelAverage'
+]
 
 
 class Optimizer(object):
@@ -119,7 +123,12 @@ class Optimizer(object):
         """
         pass
 
-    def _add_accumulator(self, name, param, dtype=None, fill_value=0.0):
+    def _add_accumulator(self,
+                         name,
+                         param,
+                         dtype=None,
+                         fill_value=0.0,
+                         shape=None):
         """Utility function to add an accumulator for a parameter
 
         Args:
@@ -133,17 +142,19 @@ class Optimizer(object):
                 param.name in self._accumulators[name]):
             raise Exception("Accumulator {} already exists for parameter {}".
                             format(name, param.name))
-
+        if shape == None:
+            shape = param.shape
         assert isinstance(self.helper, LayerHelper)
         var = self.helper.create_global_variable(
             name=unique_name.generate(name),
             persistable=True,
             dtype=dtype or param.dtype,
             type=param.type,
-            shape=param.shape)
+            shape=shape)
         self.helper.set_variable_initializer(
             var, initializer=Constant(value=float(fill_value)))
         self._accumulators[name][param.name] = var
+        return var
 
     def _get_accumulator(self, name, param):
         """Utility function to fetch an accumulator for a parameter
@@ -580,6 +591,205 @@ class DecayedAdagradOptimizer(Optimizer):
         return decayed_adagrad_op
 
 
+class AdadeltaOptimizer(Optimizer):
+    """
+    **Adadelta Optimizer**
+    Simple Adadelta optimizer with average squared grad state and
+    average squared update state.
+    The details of adadelta please refer to this
+    `ADADELTA: AN ADAPTIVE LEARNING RATE METHOD
+    <http://www.matthewzeiler.com/pubs/googleTR2012/googleTR2012.pdf>`_.
+
+    ..  math::
+
+        E(g_t^2) &= \\rho * E(g_{t-1}^2) + (1-\\rho) * g^2 \\\\
+        learning\\_rate &= sqrt( ( E(dx_{t-1}^2) + \\epsilon ) / ( \\
+                          E(g_t^2) + \\epsilon ) ) \\\\
+        E(dx_t^2) &= \\rho * E(dx_{t-1}^2) + (1-\\rho) * (-g*learning\\_rate)^2
+
+    Args:
+        learning_rate(float): global leraning rate
+        rho(float): rho in equation
+        epsilon(float): epsilon in equation
+
+    Examples:
+        .. code-block:: python
+
+            optimizer = fluid.optimizer.Adadelta(
+                learning_rate=0.0003, epsilon=1.0e-6, rho=0.95)
+            _, params_grads = optimizer.minimize(cost)
+    """
+
+    _avg_squared_grad_acc_str = "_avg_squared_grad"
+    _avg_squared_update_acc_str = "_avg_squared_update"
+
+    def __init__(self, learning_rate, epsilon=1.0e-6, rho=0.95, **kwargs):
+        if learning_rate is None:
+            raise ValueError("learning_rate is not set.")
+        if epsilon is None:
+            raise ValueError("epsilon is not set.")
+        if rho is None:
+            raise ValueError("rho is not set.")
+        super(AdadeltaOptimizer, self).__init__(
+            learning_rate=learning_rate, **kwargs)
+        self.type = "adadelta"
+        self._epsilon = epsilon
+        self._rho = rho
+
+    def _create_accumulators(self, block, parameters):
+        if not isinstance(block, framework.Block):
+            raise TypeError("block is not instance of framework.Block.")
+
+        for p in parameters:
+            self._add_accumulator(self._avg_squared_grad_acc_str, p)
+            self._add_accumulator(self._avg_squared_update_acc_str, p)
+
+    def _append_optimize_op(self, block, param_and_grad):
+        if not isinstance(block, framework.Block):
+            raise TypeError("block is not instance of framework.Block.")
+
+        avg_squared_grad_acc = self._get_accumulator(
+            self._avg_squared_grad_acc_str, param_and_grad[0])
+        avg_squared_update_acc = self._get_accumulator(
+            self._avg_squared_update_acc_str, param_and_grad[0])
+
+        # Create the adadelta optimizer op
+        adadelta_op = block.append_op(
+            type=self.type,
+            inputs={
+                "Param": param_and_grad[0],
+                "Grad": param_and_grad[1],
+                "AvgSquaredGrad": avg_squared_grad_acc,
+                "AvgSquaredUpdate": avg_squared_update_acc
+            },
+            outputs={
+                "ParamOut": param_and_grad[0],
+                "AvgSquaredGradOut": avg_squared_grad_acc,
+                "AvgSquaredUpdateOut": avg_squared_update_acc
+            },
+            attrs={"epsilon": self._epsilon,
+                   "rho": self._rho})
+
+        return adadelta_op
+
+
+class RMSPropOptimizer(Optimizer):
+    """
+    Root Mean Squared Propagation (RMSProp) is an unpublished, adaptive learning
+    rate method. The original slides proposed RMSProp: Slide 29 of
+    http://www.cs.toronto.edu/~tijmen/csc321/slides/lecture_slides_lec6.pdf .
+
+    The original equation is as follows:
+
+    ..  math::
+
+        r(w, t) & = \\rho r(w, t-1) + (1 - \\rho)(\\nabla Q_{i}(w))^2 \\\\
+
+        w & = w - \\frac{\\eta} {\\sqrt{r(w,t) + \\epsilon}} \\nabla Q_{i}(w)
+
+    The first equation calculates moving average of the squared gradient for
+    each weight. Then dividing the gradient by :math: `sqrt{v(w,t)}`.
+
+    In some cases, adding a momentum term :math: `\\beta` is beneficial.
+    In our implementation, Nesterov momentum is used:
+
+    ..  math::
+
+        r(w, t) & = \\rho r(w, t-1) + (1 - \\rho)(\\nabla Q_{i}(w))^2 \\\\
+
+        v(w, t) & = \\beta v(w, t-1) + \\frac{\\eta} {\\sqrt{v(w,t) +
+            \\epsilon}} \\nabla Q_{i}(w)
+
+        w & = w - v(w, t)
+
+    where, :math: `\\rho` is a hyperparameter and typical values are 0.9, 0.95
+    and so on. :math: `beta` is the momentum term. :math: `\\epsilon` is a
+    smoothing term to avoid division by zero, usually set somewhere in range
+    from 1e-4 to 1e-8.
+
+
+    Args:
+        learning_rate(float): global leraning rate.
+        rho(float): rho is :math: `\\rho` in equation, set 0.95 by default.
+        epsilon(float): :math: `\\epsilon` in equation is smoothing term to
+            avoid division by zero, set 1e-6 by default.
+        momentum(float): :math: `\\beta` in equation is the momentum term,
+            set 0.0 by default.
+
+    Raises:
+        ValueError: If learning_rate, rho, epsilon, momentum are None.
+
+    Examples:
+          .. code-block:: python
+
+              optimizer = fluid.optimizer.RMSProp(0.0001)
+              _, params_grads = optimizer.minimize(cost)
+    """
+
+    _momentum_acc_str = "momentum"
+    _mean_square_acc_str = "mean_square"
+
+    def __init__(self,
+                 learning_rate,
+                 rho=0.95,
+                 epsilon=1.0e-6,
+                 momentum=0.0,
+                 **kwargs):
+        super(RMSPropOptimizer, self).__init__(
+            learning_rate=learning_rate, **kwargs)
+        if learning_rate is None:
+            raise ValueError("learning_rate is not set.")
+        if rho is None:
+            raise ValueError("rho is not set.")
+        if epsilon is None:
+            raise ValueError("epsilon is not set.")
+        if momentum is None:
+            raise ValueError("momentum is not set.")
+
+        self.type = "rmsprop"
+        self._rho = rho
+        self._epsilon = epsilon
+        self._momentum = momentum
+
+    def _create_accumulators(self, block, parameters):
+        if not isinstance(block, framework.Block):
+            raise TypeError("block is not instance of framework.Block.")
+
+        for p in parameters:
+            self._add_accumulator(self._momentum_acc_str, p)
+            self._add_accumulator(self._mean_square_acc_str, p)
+
+    def _append_optimize_op(self, block, param_and_grad):
+        if not isinstance(block, framework.Block):
+            raise TypeError("block is not instance of framework.Block.")
+
+        momentum_acc = self._get_accumulator(self._momentum_acc_str,
+                                             param_and_grad[0])
+        mean_square_acc = self._get_accumulator(self._mean_square_acc_str,
+                                                param_and_grad[0])
+        rmsprop_op = block.append_op(
+            type=self.type,
+            inputs={
+                "Param": param_and_grad[0],
+                "Grad": param_and_grad[1],
+                "Moment": momentum_acc,
+                "MeanSquare": mean_square_acc,
+                "LearningRate": self._create_param_lr(param_and_grad),
+            },
+            outputs={
+                "ParamOut": param_and_grad[0],
+                "MomentOut": momentum_acc,
+                "MeanSquareOut": mean_square_acc
+            },
+            attrs={
+                "epsilon": self._epsilon,
+                "decay": self._rho,
+                "momentum": self._momentum
+            })
+
+        return rmsprop_op
+
+
 # We short the class name, since users will use the optimizer with the package
 # name. The sample code:
 #
@@ -594,3 +804,145 @@ Adagrad = AdagradOptimizer
 Adam = AdamOptimizer
 Adamax = AdamaxOptimizer
 DecayedAdagrad = DecayedAdagradOptimizer
+Adadelta = AdadeltaOptimizer
+RMSProp = RMSPropOptimizer
+
+
+class ModelAverage(Optimizer):
+    """Accumulate the average of parameters whtin sliding window. The average
+    result will be saved in temporary variables which can be applied to
+    parameter variables of current model by calling 'apply()' method. And the
+    'restore()' method is used to restored the parameter values of current model.
+
+    The size of average window is determined by average_window_rate,
+    min_average_window, max_average_window and current update times.
+
+    Args:
+        params_grads: A list of parameter-grad variable pairs.
+        average_window_rate: The rate of average window.
+        min_average_window: The minimum size of average window.
+        max_average_window: The maximum size of average window.
+
+    Examples:
+        ...
+        optimizer = fluid.optimizer.Momentum()
+        _, params_grads = optimizer.minimize(cost)
+        model_average = fluid.optimizer.ModelAverage(params_grads, 0.15,
+                                                min_average_window=10000,
+                                                max_average_window=20000)
+        for pass_id in range(args.pass_num):
+            for data in train_reader():
+                exe.run(fluid.default_main_program()...)
+
+            with model_average.apply(exe):
+                for data in test_reader():
+                    exe.run(inference_program...)
+    """
+
+    def __init__(self,
+                 params_grads,
+                 average_window_rate,
+                 min_average_window=10000,
+                 max_average_window=10000,
+                 **kwargs):
+        super(ModelAverage, self).__init__(0.0, **kwargs)
+        self.average_window = average_window_rate
+        self.min_average_window = min_average_window
+        self.max_average_window = max_average_window
+        self.params_grads = params_grads
+        for param, grad in self.params_grads:
+            if grad is not None:
+                self._append_average_accumulate_op(param)
+
+        self.apply_program = Program()
+        block = self.apply_program.global_block()
+        with program_guard(main_program=self.apply_program):
+            for param_grad in self.params_grads:
+                if param_grad[1] is not None:
+                    self._add_average_apply_op(block, param_grad)
+
+        self.restore_program = Program()
+        block = self.restore_program.global_block()
+        with program_guard(main_program=self.restore_program):
+            for param_grad in self.params_grads:
+                if param_grad[1] is not None:
+                    self._add_average_restore_op(block, param_grad)
+
+    def _add_average_apply_op(self, block, param_grad):
+        param = block.clone_variable(param_grad[0])
+        grad = block.clone_variable(param_grad[1])
+        sum_1 = block.clone_variable(self._get_accumulator('sum_1', param))
+        sum_2 = block.clone_variable(self._get_accumulator('sum_2', param))
+        sum_3 = block.clone_variable(self._get_accumulator('sum_3', param))
+        num_accumulates = block.clone_variable(
+            self._get_accumulator('num_accumulates', param))
+        old_num_accumulates = block.clone_variable(
+            self._get_accumulator('old_num_accumulates', param))
+        num_updates = block.clone_variable(
+            self._get_accumulator('num_updates', param))
+        # backup param value to grad
+        layers.assign(input=param, output=grad)
+        # param = (sum_1 + sum_2 + sum_3) / (num_accumulates + old_num_accumulates)
+        tmp = layers.sum(x=[num_accumulates, old_num_accumulates])
+        sum = layers.sum(x=[sum_1, sum_2, sum_3])
+        tmp = layers.cast(x=tmp, dtype='float32')
+        sum = layers.cast(x=sum, dtype='float32')
+        layers.elementwise_div(x=sum, y=tmp, out=param)
+
+    def _add_average_restore_op(self, block, param_grad):
+        param = block.clone_variable(param_grad[0])
+        grad = block.clone_variable(param_grad[1])
+        layers.assign(input=grad, output=param)
+
+    def _append_average_accumulate_op(self, param):
+        self.helper = LayerHelper("average_accumulate")
+        sum_1 = self._add_accumulator('sum_1', param)
+        sum_2 = self._add_accumulator('sum_2', param)
+        sum_3 = self._add_accumulator('sum_3', param)
+        num_accumulates = self._add_accumulator(
+            'num_accumulates', param, dtype='int64', shape=[1])
+        old_num_accumulates = self._add_accumulator(
+            'old_num_accumulates', param, dtype='int64', shape=[1])
+        num_updates = self._add_accumulator(
+            'num_updates', param, dtype='int64', shape=[1])
+
+        self.helper.append_op(
+            type='average_accumulates',
+            inputs={
+                "param": param,
+                "in_sum_1": sum_1,
+                "in_sum_2": sum_2,
+                "in_sum_3": sum_3,
+                "in_num_accumulates": num_accumulates,
+                "in_old_num_accumulates": old_num_accumulates,
+                "in_num_updates": num_updates
+            },
+            outputs={
+                "out_sum_1": sum_1,
+                "out_sum_2": sum_2,
+                "out_sum_3": sum_3,
+                "out_num_accumulates": num_accumulates,
+                "out_old_num_accumulates": old_num_accumulates,
+                "out_num_updates": num_updates,
+            },
+            attrs={
+                "average_window": self.average_window,
+                "min_average_window": self.min_average_window,
+                "max_average_window": self.max_average_window,
+            })
+
+    @contextmanager
+    def apply(self, executor, need_restore=True):
+        """Apply average values to parameters of current model.
+        """
+        executor.run(self.apply_program)
+        try:
+            yield
+        finally:
+            if need_restore:
+                self.restore(executor)
+
+    def restore(self, executor):
+        """Restore parameter values of current model.
+        """
+        executor.run(self.restore_program)

python/paddle/fluid/tests/book/test_machine_translation.py

@@ -118,12 +118,12 @@ def decoder_decode(context, is_sparse):
             is_sparse=is_sparse)
 
         # use rnn unit to update rnn
-        current_state = pd.fc(input=[pre_ids_emb, pre_state_expanded],
+        current_state = pd.fc(input=[pre_state_expanded, pre_ids_emb],
                               size=decoder_size,
                               act='tanh')
-
+        current_state_with_lod = pd.lod_reset(x=current_state, y=pre_score)
         # use score to do beam search
-        current_score = pd.fc(input=current_state,
+        current_score = pd.fc(input=current_state_with_lod,
                               size=target_dict_dim,
                               act='softmax')
         topk_scores, topk_indices = pd.topk(current_score, k=50)

python/paddle/fluid/tests/test_concurrency.py

@@ -173,16 +173,10 @@ class TestRoutineOp(unittest.TestCase):
                 with while_op.block():
                     result2 = fill_constant(
                         shape=[1], dtype=core.VarDesc.VarType.INT32, value=0)
-                    x_to_send_tmp = fill_constant(
-                        shape=[1], dtype=core.VarDesc.VarType.INT32, value=0)
-
-                    # TODO(abhinav): Need to perform copy when doing a channel send.
-                    #   Once this is complete, we can remove these lines
-                    assign(input=x, output=x_to_send_tmp)
 
                     with fluid.Select() as select:
-                        with select.case(fluid.channel_send, channel,
-                                         x_to_send_tmp):
+                        with select.case(
+                                fluid.channel_send, channel, x, is_copy=True):
                             assign(input=x, output=x_tmp)
                             assign(input=y, output=x)
                             assign(elementwise_add(x=x_tmp, y=y), output=y)
@@ -230,21 +224,12 @@ class TestRoutineOp(unittest.TestCase):
                                               core.VarDesc.VarType.LOD_TENSOR,
                                               core.VarDesc.VarType.FP64)
 
-            pong_result = self._create_tensor('pong_return_value',
-                                              core.VarDesc.VarType.LOD_TENSOR,
-                                              core.VarDesc.VarType.FP64)
-
             def ping(ch, message):
-                message_to_send_tmp = fill_constant(
-                    shape=[1], dtype=core.VarDesc.VarType.FP64, value=0)
-
-                assign(input=message, output=message_to_send_tmp)
-                fluid.channel_send(ch, message_to_send_tmp)
+                fluid.channel_send(ch, message, is_copy=True)
 
             def pong(ch1, ch2):
                 fluid.channel_recv(ch1, ping_result)
-                assign(input=ping_result, output=pong_result)
-                fluid.channel_send(ch2, pong_result)
+                fluid.channel_send(ch2, ping_result, is_copy=True)
 
             pings = fluid.make_channel(
                 dtype=core.VarDesc.VarType.LOD_TENSOR, capacity=1)

python/paddle/fluid/tests/unittests/.gitignore

 mnist.recordio
+mnist_0.recordio
+mnist_1.recordio
+mnist_2.recordio

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

@@ -14,6 +14,7 @@
 
 import unittest
 import numpy as np
+import paddle.fluid.core as core
 from op_test import OpTest
 from scipy.special import expit
 
@@ -212,18 +213,39 @@ class TestRound(OpTest):
 class TestRelu(OpTest):
     def setUp(self):
         self.op_type = "relu"
-        x = np.random.uniform(-1, 1, [11, 17]).astype("float32")
+        self.dtype = np.float32
+        self.init_dtype()
+
+        x = np.random.uniform(-1, 1, [11, 17]).astype(self.dtype)
         # The same reason with TestAbs
         x[np.abs(x) < 0.005] = 0.02
-        self.inputs = {'X': x}
-        self.outputs = {'Out': np.maximum(self.inputs['X'], 0)}
+        out = np.maximum(x, 0)
+
+        self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
+        self.outputs = {'Out': out}
 
     def test_check_output(self):
         self.check_output()
 
     def test_check_grad(self):
+        if self.dtype == np.float16:
+            return
         self.check_grad(['X'], 'Out', max_relative_error=0.007)
 
+    def init_dtype(self):
+        pass
+
+
+class TestFP16Relu(TestRelu):
+    def init_dtype(self):
+        self.dtype = np.float16
+
+    def test_check_output(self):
+        if core.is_compiled_with_cuda():
+            place = core.CUDAPlace(0)
+            if core.is_float16_supported(place):
+                self.check_output_with_place(place, atol=1e-3)
+
 
 class TestBRelu(OpTest):
     def setUp(self):
@@ -484,5 +506,54 @@ class TestSwish(OpTest):
         self.check_grad(['X'], 'Out', max_relative_error=0.008)
 
 
+#--------------------test MKLDNN--------------------
+class TestMKLDNNRelu(TestRelu):
+    def setUp(self):
+        super(TestMKLDNNRelu, self).setUp()
+
+        x = np.random.uniform(-1, 1, [2, 4, 3, 5]).astype("float32")
+        # The same reason with TestAbs
+        x[np.abs(x) < 0.005] = 0.02
+        out = np.maximum(x, 0)
+
+        self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
+        self.outputs = {'Out': out}
+        self.attrs = {"use_mkldnn": True}
+
+
+class TestMKLDNNTanh(TestTanh):
+    def setUp(self):
+        super(TestMKLDNNTanh, self).setUp()
+
+        self.inputs = {
+            'X': np.random.uniform(0.1, 1, [2, 4, 3, 5]).astype("float32")
+        }
+        self.outputs = {'Out': np.tanh(self.inputs['X'])}
+        self.attrs = {"use_mkldnn": True}
+
+
+class TestMKLDNNSqrt(TestSqrt):
+    def setUp(self):
+        super(TestMKLDNNSqrt, self).setUp()
+
+        self.inputs = {
+            'X': np.random.uniform(0.1, 1, [2, 4, 3, 5]).astype("float32")
+        }
+        self.outputs = {'Out': np.sqrt(self.inputs['X'])}
+        self.attrs = {"use_mkldnn": True}
+
+
+class TestMKLDNNAbs(TestAbs):
+    def setUp(self):
+        super(TestMKLDNNAbs, self).setUp()
+
+        x = np.random.uniform(-1, 1, [2, 4, 3, 5]).astype("float32")
+        # The same reason with TestAbs
+        x[np.abs(x) < 0.005] = 0.02
+        self.inputs = {'X': x}
+        self.outputs = {'Out': np.abs(self.inputs['X'])}
+        self.attrs = {"use_mkldnn": True}
+
+
 if __name__ == "__main__":
     unittest.main()

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

@@ -31,6 +31,37 @@ def get_backward_op(scope, op, no_grad_set):
     return backward_op
 
 
+def _reference_testing(x, scale, offset, mean, var, epsilon, data_format):
+    x_shape = x.shape
+    if len(x_shape) == 2:
+        if data_format == "NCHW":
+            x = np.reshape(x, (x.shape[0], x.shape[1], 1, 1))
+        else:
+            x = np.reshape(x, (x.shape[0], 1, 1, x.shape[1]))
+
+    if data_format == "NCHW":
+        n, c, h, w = x.shape
+        mean_tile = np.reshape(mean, (1, c, 1, 1))
+        mean_tile = np.tile(mean_tile, (n, 1, h, w))
+        var_tile = np.reshape(var, (1, c, 1, 1))
+        var_tile = np.tile(var_tile, (n, 1, h, w))
+        normalized = (x - mean_tile) / np.sqrt(var_tile + epsilon)
+        scale_tile = np.reshape(scale, (1, c, 1, 1))
+        scale_tile = np.tile(scale_tile, (n, 1, h, w))
+        offset_tile = np.reshape(offset, (1, c, 1, 1))
+        offset_tile = np.reshape(offset_tile, (1, c, 1, 1))
+        y = normalized * scale_tile + offset_tile
+    elif data_format == "NHWC":
+        normalized = (x - mean) / np.sqrt(var + epsilon)
+        y = normalized * scale + offset
+    else:
+        raise ValueError("Unknown data order.")
+
+    if len(x_shape) == 2:
+        y = np.reshape(y, x_shape)
+    return y
+
+
 def _reference_training(x, scale, offset, epsilon, data_format):
     x_shape = x.shape
     if len(x_shape) == 2:
@@ -155,11 +186,159 @@ def set_output_grad(scope, outputs, place, feed_dict=None):
         __set_tensor__(output, data)
 
 
-class TestBatchNormOp(OpTest):
+class TestBatchNormOpInference(OpTest):
+    def setUp(self):
+        self.dtype = np.float32
+
     def __assert_close(self, tensor, np_array, msg, atol=1e-4):
         self.assertTrue(np.allclose(np.array(tensor), np_array, atol=atol), msg)
 
-    def test_python(self):
+    def check_with_place(self, place, data_layout, dtype, shape):
+        epsilon = 0.00001
+        if len(shape) == 2:
+            x_shape = shape
+            c = x_shape[1]
+        else:
+            n, h, w, c = shape[0], shape[1], shape[2], shape[3]
+            if data_layout == "NHWC":
+                x_shape = [n, h, w, c]
+            elif data_layout == "NCHW":
+                x_shape = [n, c, h, w]
+            else:
+                raise ValueError("Unknown data layout.")
+        scale_shape = [c]
+
+        x_val = np.random.random_sample(x_shape).astype(dtype)
+        scale_val = np.random.random_sample(scale_shape).astype(np.float32)
+        bias_val = np.random.random_sample(scale_shape).astype(np.float32)
+
+        mean = np.zeros(scale_shape).astype(np.float32)
+        variance = np.ones(scale_shape).astype(np.float32)
+
+        y_out = _reference_testing(x_val, scale_val, bias_val, mean, variance,
+                                   epsilon, data_layout).astype(dtype)
+
+        scope = core.Scope()
+
+        # create input
+        x_tensor = create_or_get_tensor(scope, "x_val",
+                                        OpTest.np_dtype_to_fluid_dtype(x_val),
+                                        place)
+        scale_tensor = create_or_get_tensor(
+            scope, "scale_val",
+            OpTest.np_dtype_to_fluid_dtype(scale_val), place)
+        bias_tensor = create_or_get_tensor(
+            scope, "bias_val", OpTest.np_dtype_to_fluid_dtype(bias_val), place)
+        mean_tensor = create_or_get_tensor(scope, "mean",
+                                           OpTest.np_dtype_to_fluid_dtype(mean),
+                                           place)
+        variance_tensor = create_or_get_tensor(
+            scope, "variance", OpTest.np_dtype_to_fluid_dtype(variance), place)
+
+        # create output
+        y_tensor = create_or_get_tensor(scope, "y_out", None, place)
+        saved_mean_tensor = create_or_get_tensor(scope, "saved_mean", None,
+                                                 place)
+        saved_variance_tensor = create_or_get_tensor(scope, "saved_variance",
+                                                     None, place)
+        mean_out_tensor = mean_tensor
+        variance_out_tensor = variance_tensor
+
+        batch_norm_op = Operator(
+            "batch_norm",
+            # inputs
+            X="x_val",
+            Scale="scale_val",
+            Bias="bias_val",
+            Mean="mean",
+            Variance="variance",
+            # outputs
+            Y="y_out",
+            MeanOut="mean",
+            VarianceOut="variance",
+            SavedMean="saved_mean",
+            SavedVariance="saved_variance",
+            # attrs
+            is_test=True,
+            data_layout=data_layout,
+            epsilon=epsilon)
+
+        batch_norm_op.run(scope, place)
+
+        # check inference result
+        self.__assert_close(
+            y_tensor,
+            y_out,
+            "inference output are different at " + str(place) + ", " +
+            data_layout + ", " + str(np.dtype(dtype)) +
+            str(np.array(y_tensor)) + str(y_out),
+            atol=1e-3)
+
+    def test_check_output(self):
+        places = [core.CPUPlace()]
+        if core.is_compiled_with_cuda() and core.op_support_gpu("batch_norm"):
+            places.append(core.CUDAPlace(0))
+
+        for place in places:
+            for data_format in ["NCHW", "NHWC"]:
+                self.check_with_place(place, data_format, self.dtype,
+                                      [2, 3, 4, 5])
+                self.check_with_place(place, data_format, self.dtype, [2, 3])
+
+
+class TestFP16BatchNormOpInference(TestBatchNormOpInference):
+    def setUp(self):
+        self.dtype = np.float16
+
+    def test_check_output(self):
+        places = []
+        if core.is_compiled_with_cuda() and core.op_support_gpu("batch_norm"):
+            place = core.CUDAPlace(0)
+            if core.is_float16_supported(place):
+                places.append(place)
+
+        for place in places:
+            for data_format in ["NCHW", "NHWC"]:
+                self.check_with_place(place, data_format, self.dtype,
+                                      [2, 3, 4, 5])
+                self.check_with_place(place, data_format, self.dtype, [2, 3])
+
+
+class TestBatchNormOpTraining(OpTest):
+    def __assert_close(self, tensor, np_array, msg, atol=1e-4):
+        self.assertTrue(np.allclose(np.array(tensor), np_array, atol=atol), msg)
+
+    def test_python_testing(self):
+        data_format = "NHWC"
+        epsilon = 0.00001
+
+        n, h, w, c = 2, 3, 4, 5
+        x_shape = [n, h, w, c]
+        scale_shape = [c]
+
+        x_val = np.random.random_sample(x_shape).astype(np.float32)
+        scale_val = np.random.random_sample(scale_shape).astype(np.float32)
+        bias_val = np.random.random_sample(scale_shape).astype(np.float32)
+
+        mean = np.zeros(scale_shape).astype(np.float32)
+        variance = np.ones(scale_shape).astype(np.float32)
+
+        y_out = _reference_testing(x_val, scale_val, bias_val, mean, variance,
+                                   epsilon, "NHWC")
+
+        # running N, C, H, W case
+        # should produce the same results
+        x_shape2 = [n, c, h, w]
+        x_val2 = np.transpose(x_val, (0, 3, 1, 2))
+        y_out2 = _reference_testing(x_val2, scale_val, bias_val, mean, variance,
+                                    epsilon, "NCHW")
+
+        # transfer (N, C, H, W) back to (N, H, W, C)
+        y_out2_trans = np.transpose(y_out2, (0, 2, 3, 1))
+        self.__assert_close(y_out, y_out2_trans, "inference output")
+        print 'python: NHWC, NCHW, inference checking passed'
+
+    def test_python_training(self):
         data_format = "NHWC"
         epsilon = 0.00001
         momentum = 0.9
@@ -197,7 +376,7 @@ class TestBatchNormOp(OpTest):
 
         # transfer (N, C, H, W) back to (N, H, W, C)
         y_out2_trans = np.transpose(y_out2, (0, 2, 3, 1))
-        self.__assert_close(y_out, y_out2_trans, "batch variance")
+        self.__assert_close(y_out, y_out2_trans, "batch output")
         print 'python: NHWC, NCHW, forward checking passed'
 
         # test backward now

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

@@ -20,19 +20,35 @@ from op_test import OpTest
 class TestConcatOp(OpTest):
     def setUp(self):
         self.op_type = "concat"
-        x0 = np.random.random((2, 1, 4, 5)).astype('float32')
-        x1 = np.random.random((2, 2, 4, 5)).astype('float32')
-        x2 = np.random.random((2, 3, 4, 5)).astype('float32')
-        axis = 1
-        self.inputs = {'X': [('x0', x0), ('x1', x1), ('x2', x2)]}
-        self.attrs = {'axis': axis}
-        self.outputs = {'Out': np.concatenate((x0, x1, x2), axis=axis)}
+        self.init_test_data()
+        self.inputs = {'X': [('x0', self.x0), ('x1', self.x1), ('x2', self.x2)]}
+        self.attrs = {'axis': self.axis}
+        self.outputs = {
+            'Out': np.concatenate(
+                (self.x0, self.x1, self.x2), axis=self.axis)
+        }
 
     def test_check_output(self):
         self.check_output()
 
     def test_check_grad(self):
         self.check_grad(['x0'], 'Out')
+        self.check_grad(['x1'], 'Out')
+        self.check_grad(['x2'], 'Out')
+
+    def init_test_data(self):
+        self.x0 = np.random.random((2, 1, 4, 5)).astype('float32')
+        self.x1 = np.random.random((2, 2, 4, 5)).astype('float32')
+        self.x2 = np.random.random((2, 3, 4, 5)).astype('float32')
+        self.axis = 1
+
+
+class TestConcatOp2(OpTest):
+    def init_test_data(self):
+        self.x0 = np.random.random((2, 3, 4, 5)).astype('float32')
+        self.x1 = np.random.random((2, 3, 4, 5)).astype('float32')
+        self.x2 = np.random.random((2, 3, 4, 5)).astype('float32')
+        self.axis = 1
 
 
 if __name__ == '__main__':

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

@@ -14,6 +14,7 @@
 
 import unittest
 import numpy as np
+import paddle.fluid.core as core
 from op_test import OpTest
 
 
@@ -82,5 +83,37 @@ class TestDropoutOp5(OpTest):
         self.check_output()
 
 
+class TestFP16DropoutOp(OpTest):
+    def setUp(self):
+        self.op_type = "dropout"
+        self.init_test_case()
+
+        x = np.random.random(self.input_size).astype("float16")
+        out = x * (1.0 - self.prob)
+        self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
+        self.attrs = {
+            'dropout_prob': self.prob,
+            'fix_seed': self.fix_seed,
+            'is_test': True
+        }
+        self.outputs = {'Out': out}
+
+    def init_test_case(self):
+        self.input_size = [32, 64]
+        self.prob = 0.35
+        self.fix_seed = True
+
+    def test_check_output(self):
+        if core.is_compiled_with_cuda() and core.op_support_gpu("dropout"):
+            self.check_output_with_place(core.CUDAPlace(0), atol=1e-3)
+
+
+class TestFP16DropoutOp2(TestFP16DropoutOp):
+    def init_test_case(self):
+        self.input_size = [32, 64, 3]
+        self.prob = 0.75
+        self.fix_seed = False
+
+
 if __name__ == '__main__':
     unittest.main()

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

@@ -13,158 +13,243 @@
 # limitations under the License.
 import unittest
 import numpy as np
+import paddle.fluid.core as core
 from op_test import OpTest
 
 
-class TestElementwiseOp(OpTest):
+class TestElementwiseAddOp(OpTest):
     def setUp(self):
         self.op_type = "elementwise_add"
+        self.dtype = np.float32
+        self.axis = -1
+        self.init_dtype()
+        self.init_input_output()
+        self.init_axis()
+
         self.inputs = {
-            'X': np.random.uniform(0.1, 1, [13, 17]).astype("float32"),
-            'Y': np.random.uniform(0.1, 1, [13, 17]).astype("float32")
+            'X': OpTest.np_dtype_to_fluid_dtype(self.x),
+            'Y': OpTest.np_dtype_to_fluid_dtype(self.y)
         }
-        self.outputs = {'Out': np.add(self.inputs['X'], self.inputs['Y'])}
+        self.attrs = {'axis': self.axis}
+        self.outputs = {'Out': self.out}
 
     def test_check_output(self):
         self.check_output()
 
     def test_check_grad_normal(self):
+        if self.dtype == np.float16:
+            return
         self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.005)
 
     def test_check_grad_ingore_x(self):
+        if self.dtype == np.float16:
+            return
         self.check_grad(
             ['Y'], 'Out', max_relative_error=0.005, no_grad_set=set("X"))
 
     def test_check_grad_ingore_y(self):
+        if self.dtype == np.float16:
+            return
         self.check_grad(
             ['X'], 'Out', max_relative_error=0.005, no_grad_set=set('Y'))
 
+    def init_input_output(self):
+        self.x = np.random.uniform(0.1, 1, [13, 17]).astype(self.dtype)
+        self.y = np.random.uniform(0.1, 1, [13, 17]).astype(self.dtype)
+        self.out = np.add(self.x, self.y)
 
-class TestElementwiseAddOp_scalar(TestElementwiseOp):
-    def setUp(self):
-        self.op_type = "elementwise_add"
-        self.inputs = {
-            'X': np.random.rand(2, 3, 4).astype(np.float32),
-            'Y': np.random.rand(1).astype(np.float32)
-        }
-        self.outputs = {'Out': self.inputs['X'] + self.inputs['Y']}
+    def init_dtype(self):
+        pass
 
+    def init_axis(self):
+        pass
 
-class TestElementwiseAddOp_scalar2(TestElementwiseOp):
-    def setUp(self):
-        self.op_type = "elementwise_add"
-        self.inputs = {
-            'X': np.random.rand(2, 3, 4).astype(np.float32),
-            'Y': np.random.rand(1, 1).astype(np.float32)
-        }
-        self.outputs = {'Out': self.inputs['X'] + self.inputs['Y']}
 
+class TestFP16ElementwiseAddOp(TestElementwiseAddOp):
+    def init_dtype(self):
+        self.dtype = np.float16
 
-class TestElementwiseAddOp_Vector(TestElementwiseOp):
-    def setUp(self):
-        self.op_type = "elementwise_add"
-        self.inputs = {
-            'X': np.random.random((32, )).astype("float32"),
-            'Y': np.random.random((32, )).astype("float32")
-        }
-        self.outputs = {'Out': np.add(self.inputs['X'], self.inputs['Y'])}
+    def test_check_output(self):
+        if core.is_compiled_with_cuda():
+            place = core.CUDAPlace(0)
+            if core.is_float16_supported(place):
+                self.check_output_with_place(place, atol=1e-3)
 
 
-class TestElementwiseAddOp_broadcast_0(TestElementwiseOp):
-    def setUp(self):
-        self.op_type = "elementwise_add"
-        self.inputs = {
-            'X': np.random.rand(2, 3, 4).astype(np.float32),
-            'Y': np.random.rand(2).astype(np.float32)
-        }
+class TestElementwiseAddOp_scalar(TestElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 3, 4).astype(self.dtype)
+        self.y = np.random.rand(1).astype(self.dtype)
+        self.out = self.x + self.y
 
-        self.attrs = {'axis': 0}
-        self.outputs = {
-            'Out': self.inputs['X'] + self.inputs['Y'].reshape(2, 1, 1)
-        }
 
+class TestFP16ElementwiseAddOp_scalar(TestFP16ElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 3, 4).astype(self.dtype)
+        self.y = np.random.rand(1).astype(self.dtype)
+        self.out = self.x + self.y
 
-class TestElementwiseAddOp_broadcast_1(TestElementwiseOp):
-    def setUp(self):
-        self.op_type = "elementwise_add"
-        self.inputs = {
-            'X': np.random.rand(2, 3, 4).astype(np.float32),
-            'Y': np.random.rand(3).astype(np.float32)
-        }
 
-        self.attrs = {'axis': 1}
-        self.outputs = {
-            'Out': self.inputs['X'] + self.inputs['Y'].reshape(1, 3, 1)
-        }
+class TestElementwiseAddOp_scalar2(TestElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 3, 4).astype(self.dtype)
+        self.y = np.random.rand(1, 1).astype(self.dtype)
+        self.out = self.x + self.y
 
 
-class TestElementwiseAddOp_broadcast_2(TestElementwiseOp):
-    def setUp(self):
-        self.op_type = "elementwise_add"
-        self.inputs = {
-            'X': np.random.rand(2, 3, 4).astype(np.float32),
-            'Y': np.random.rand(4).astype(np.float32)
-        }
+class TestFP16ElementwiseAddOp_scalar2(TestFP16ElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 3, 4).astype(self.dtype)
+        self.y = np.random.rand(1, 1).astype(self.dtype)
+        self.out = self.x + self.y
 
-        self.outputs = {
-            'Out': self.inputs['X'] + self.inputs['Y'].reshape(1, 1, 4)
-        }
 
+class TestElementwiseAddOp_Vector(TestElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.random((32, )).astype(self.dtype)
+        self.y = np.random.random((32, )).astype(self.dtype)
+        self.out = np.add(self.x, self.y)
 
-class TestElementwiseAddOp_broadcast_3(TestElementwiseOp):
-    def setUp(self):
-        self.op_type = "elementwise_add"
-        self.inputs = {
-            'X': np.random.rand(2, 3, 4, 5).astype(np.float32),
-            'Y': np.random.rand(3, 4).astype(np.float32)
-        }
 
-        self.attrs = {'axis': 1}
-        self.outputs = {
-            'Out': self.inputs['X'] + self.inputs['Y'].reshape(1, 3, 4, 1)
-        }
+class TestFP16ElementwiseAddOp_Vector(TestFP16ElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.random((32, )).astype(self.dtype)
+        self.y = np.random.random((32, )).astype(self.dtype)
+        self.out = np.add(self.x, self.y)
 
 
-class TestElementwiseAddOp_broadcast_4(TestElementwiseOp):
-    def setUp(self):
-        self.op_type = "elementwise_add"
-        self.inputs = {
-            'X': np.random.rand(2, 3, 4, 5).astype(np.float32),
-            'Y': np.random.rand(2, 1).astype(np.float32)
-        }
+class TestElementwiseAddOp_broadcast_0(TestElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 3, 4).astype(self.dtype)
+        self.y = np.random.rand(2).astype(self.dtype)
+        self.out = self.x + self.y.reshape(2, 1, 1)
 
-        self.attrs = {'axis': 0}
-        self.outputs = {
-            'Out': self.inputs['X'] + self.inputs['Y'].reshape(2, 1, 1, 1)
-        }
+    def init_axis(self):
+        self.axis = 0
 
 
-class TestElementwiseAddOp_rowwise_add_0(TestElementwiseOp):
-    def setUp(self):
-        self.op_type = "elementwise_add"
-        self.inputs = {
-            'X': np.random.rand(2, 3, 4).astype(np.float32),
-            'Y': np.random.rand(3, 4).astype(np.float32)
-        }
+class TestFP16ElementwiseAddOp_broadcast_0(TestFP16ElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 3, 4).astype(self.dtype)
+        self.y = np.random.rand(2).astype(self.dtype)
+        self.out = self.x + self.y.reshape(2, 1, 1)
 
-        self.attrs = {'axis': 1}
-        self.outputs = {
-            'Out': self.inputs['X'] + self.inputs['Y'].reshape(1, 3, 4)
-        }
+    def init_axis(self):
+        self.axis = 0
 
 
-class TestElementwiseAddOp_rowwise_add_1(TestElementwiseOp):
-    def setUp(self):
-        self.op_type = "elementwise_add"
-        self.inputs = {
-            'X': np.random.rand(2, 1).astype(np.float32),
-            'Y': np.random.rand(1).astype(np.float32)
-        }
+class TestElementwiseAddOp_broadcast_1(TestElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 3, 4).astype(self.dtype)
+        self.y = np.random.rand(3).astype(self.dtype)
+        self.out = self.x + self.y.reshape(1, 3, 1)
 
-        self.attrs = {'axis': 1}
-        self.outputs = {
-            'Out': self.inputs['X'] + self.inputs['Y'].reshape(1, 1)
-        }
+    def init_axis(self):
+        self.axis = 1
+
+
+class TestFP16ElementwiseAddOp_broadcast_1(TestFP16ElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 3, 4).astype(self.dtype)
+        self.y = np.random.rand(3).astype(self.dtype)
+        self.out = self.x + self.y.reshape(1, 3, 1)
+
+    def init_axis(self):
+        self.axis = 1
+
+
+class TestElementwiseAddOp_broadcast_2(TestElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 3, 4).astype(self.dtype)
+        self.y = np.random.rand(4).astype(self.dtype)
+        self.out = self.x + self.y.reshape(1, 1, 4)
+
+
+class TestFP16ElementwiseAddOp_broadcast_2(TestFP16ElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 3, 4).astype(self.dtype)
+        self.y = np.random.rand(4).astype(self.dtype)
+        self.out = self.x + self.y.reshape(1, 1, 4)
+
+
+class TestElementwiseAddOp_broadcast_3(TestElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 3, 4, 5).astype(self.dtype)
+        self.y = np.random.rand(3, 4).astype(self.dtype)
+        self.out = self.x + self.y.reshape(1, 3, 4, 1)
+
+    def init_axis(self):
+        self.axis = 1
+
+
+class TestFP16ElementwiseAddOp_broadcast_3(TestFP16ElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 3, 4, 5).astype(self.dtype)
+        self.y = np.random.rand(3, 4).astype(self.dtype)
+        self.out = self.x + self.y.reshape(1, 3, 4, 1)
+
+    def init_axis(self):
+        self.axis = 1
+
+
+class TestElementwiseAddOp_broadcast_4(TestElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 3, 4, 5).astype(self.dtype)
+        self.y = np.random.rand(2, 1).astype(self.dtype)
+        self.out = self.x + self.y.reshape(2, 1, 1, 1)
+
+    def init_axis(self):
+        self.axis = 0
+
+
+class TestFP16ElementwiseAddOp_broadcast_4(TestFP16ElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 3, 4, 5).astype(self.dtype)
+        self.y = np.random.rand(2, 1).astype(self.dtype)
+        self.out = self.x + self.y.reshape(2, 1, 1, 1)
+
+    def init_axis(self):
+        self.axis = 0
+
+
+class TestElementwiseAddOp_rowwise_add_0(TestElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 3, 4).astype(self.dtype)
+        self.y = np.random.rand(3, 4).astype(self.dtype)
+        self.out = self.x + self.y.reshape(1, 3, 4)
+
+    def init_axis(self):
+        self.axis = 1
+
+
+class TestFP16ElementwiseAddOp_rowwise_add_0(TestFP16ElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 3, 4).astype(self.dtype)
+        self.y = np.random.rand(3, 4).astype(self.dtype)
+        self.out = self.x + self.y.reshape(1, 3, 4)
+
+    def init_axis(self):
+        self.axis = 1
+
+
+class TestElementwiseAddOp_rowwise_add_1(TestElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 1).astype(self.dtype)
+        self.y = np.random.rand(1).astype(self.dtype)
+        self.out = self.x + self.y.reshape(1, 1)
+
+    def init_axis(self):
+        self.axis = 1
+
+
+class TestFP16ElementwiseAddOp_rowwise_add_1(TestFP16ElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 1).astype(self.dtype)
+        self.y = np.random.rand(1).astype(self.dtype)
+        self.out = self.x + self.y.reshape(1, 1)
+
+    def init_axis(self):
+        self.axis = 1
 
 
 if __name__ == '__main__':

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

@@ -16,7 +16,6 @@ import unittest
 
 import numpy
 import paddle.fluid.core as core
-
 from paddle.fluid.executor import Executor
 from paddle.fluid.layers import mul, data
 

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

@@ -181,8 +181,8 @@ class TestBook(unittest.TestCase):
         with program_guard(program):
             x = layers.data(name='x', shape=[10], dtype='float32')
             y = layers.data(
-                name='y', shape=[10, 20], dtype='float32', lod_level=1)
-            self.assertIsNotNone(layers.sequence_expand(x=x, y=y))
+                name='y', shape=[10, 20], dtype='float32', lod_level=2)
+            self.assertIsNotNone(layers.sequence_expand(x=x, y=y, ref_level=1))
         print(str(program))
 
     def test_lstm_unit(self):
@@ -231,6 +231,13 @@ class TestBook(unittest.TestCase):
             self.assertIsNotNone(layers.softmax(hid))
         print(str(program))
 
+    def test_lrn(self):
+        program = Program()
+        with program_guard(program):
+            data = layers.data(name='data', shape=[6, 2, 2], dtype='float32')
+            self.assertIsNotNone(layers.lrn(data))
+        print(str(program))
+
     def test_get_places(self):
         program = Program()
         with program_guard(program):

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

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

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

+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import unittest
+
+import paddle.fluid as fluid
+import paddle.v2 as paddle
+import paddle.v2.dataset.mnist as mnist
+
+
+class TestMultipleReader(unittest.TestCase):
+    def setUp(self):
+        self.batch_size = 64
+        self.pass_num = 3
+        # Convert mnist to recordio file
+        with fluid.program_guard(fluid.Program(), fluid.Program()):
+            data_file = paddle.batch(mnist.train(), batch_size=self.batch_size)
+            feeder = fluid.DataFeeder(
+                feed_list=[
+                    fluid.layers.data(
+                        name='image', shape=[784]),
+                    fluid.layers.data(
+                        name='label', shape=[1], dtype='int64'),
+                ],
+                place=fluid.CPUPlace())
+            self.num_batch = fluid.recordio_writer.convert_reader_to_recordio_file(
+                './mnist.recordio', data_file, feeder)
+
+    def test_main(self):
+        with fluid.program_guard(fluid.Program(), fluid.Program()):
+            data_file = fluid.layers.open_recordio_file(
+                filename='./mnist.recordio',
+                shapes=[(-1, 784), (-1, 1)],
+                lod_levels=[0, 0],
+                dtypes=['float32', 'int64'])
+            data_file = fluid.layers.create_multi_pass_reader(
+                reader=data_file, pass_num=self.pass_num)
+            img, label = fluid.layers.read_file(data_file)
+
+            if fluid.core.is_compiled_with_cuda():
+                place = fluid.CUDAPlace(0)
+            else:
+                place = fluid.CPUPlace()
+
+            exe = fluid.Executor(place)
+            exe.run(fluid.default_startup_program())
+
+            batch_count = 0
+            while not data_file.eof():
+                img_val, = exe.run(fetch_list=[img])
+                batch_count += 1
+                self.assertLessEqual(img_val.shape[0], self.batch_size)
+            data_file.reset()
+            self.assertEqual(batch_count, self.num_batch * self.pass_num)

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

+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import unittest
+
+import paddle.fluid as fluid
+import paddle.v2 as paddle
+import paddle.v2.dataset.mnist as mnist
+from shutil import copyfile
+
+
+class TestMultipleReader(unittest.TestCase):
+    def setUp(self):
+        self.batch_size = 64
+        # Convert mnist to recordio file
+        with fluid.program_guard(fluid.Program(), fluid.Program()):
+            reader = paddle.batch(mnist.train(), batch_size=self.batch_size)
+            feeder = fluid.DataFeeder(
+                feed_list=[  # order is image and label
+                    fluid.layers.data(
+                        name='image', shape=[784]),
+                    fluid.layers.data(
+                        name='label', shape=[1], dtype='int64'),
+                ],
+                place=fluid.CPUPlace())
+            self.num_batch = fluid.recordio_writer.convert_reader_to_recordio_file(
+                './mnist_0.recordio', reader, feeder)
+        copyfile('./mnist_0.recordio', './mnist_1.recordio')
+        copyfile('./mnist_0.recordio', './mnist_2.recordio')
+
+    def main(self, thread_num):
+        file_list = [
+            './mnist_0.recordio', './mnist_1.recordio', './mnist_2.recordio'
+        ]
+        with fluid.program_guard(fluid.Program(), fluid.Program()):
+            data_files = fluid.layers.open_files(
+                filenames=file_list,
+                thread_num=thread_num,
+                shapes=[(-1, 784), (-1, 1)],
+                lod_levels=[0, 0],
+                dtypes=['float32', 'int64'])
+            img, label = fluid.layers.read_file(data_files)
+
+            if fluid.core.is_compiled_with_cuda():
+                place = fluid.CUDAPlace(0)
+            else:
+                place = fluid.CPUPlace()
+
+            exe = fluid.Executor(place)
+            exe.run(fluid.default_startup_program())
+
+            batch_count = 0
+            while not data_files.eof():
+                img_val, = exe.run(fetch_list=[img])
+                batch_count += 1
+                self.assertLessEqual(img_val.shape[0], self.batch_size)
+            data_files.reset()
+            self.assertEqual(batch_count, self.num_batch * 3)
+
+    def test_main(self):
+        self.main(thread_num=3)  # thread number equals to file number
+        self.main(thread_num=10)  # thread number is larger than file number
+        self.main(thread_num=2)  # thread number is less than file number

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

@@ -38,14 +38,15 @@ class TestRecvOp(unittest.TestCase):
     def init_serv(self, place):
         main = fluid.Program()
         with fluid.program_guard(main):
+            serv = layers.ListenAndServ(
+                "127.0.0.1:6174", ["X"], optimizer_mode=False)
+            with serv.do():
                 x = layers.data(
                     shape=[32, 32],
                     dtype='float32',
                     name="X",
                     append_batch_size=False)
                 fluid.initializer.Constant(value=1.0)(x, main.global_block())
-            serv = layers.ListenAndServ("127.0.0.1:6174", optimizer_mode=False)
-            with serv.do():
                 o = layers.scale(x=x, scale=10.0)
             main.global_block().create_var(
                 name=o.name, psersistable=False, dtype=o.dtype, shape=o.shape)

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

@@ -33,12 +33,36 @@ class TestSequenceExpand(OpTest):
     def compute(self):
         x = self.inputs['X']
         x_data, x_lod = x if type(x) == tuple else (x, None)
-        n = 1 + x_data.shape[0] if not x_lod else len(x_lod[0])
         y_data, y_lod = self.inputs['Y']
-        repeats = [((y_lod[-1][i + 1] - y_lod[-1][i]))
-                   for i in range(len(y_lod[-1]) - 1)]
-        out = x_data.repeat(repeats, axis=0)
+
+        if hasattr(self, 'attrs'):
+            ref_level = self.attrs['ref_level']
+        else:
+            ref_level = len(y_lod) - 1
+
+        out = np.zeros(shape=((0, ) + x_data.shape[1:]), dtype=x_data.dtype)
+
+        if x_lod is None:
+            x_idx = [i for i in xrange(x_data.shape[0] + 1)]
+        else:
+            x_idx = x_lod[0]
+            out_lod = [[0]]
+
+        for i in xrange(1, len(y_lod[ref_level])):
+            repeat_num = y_lod[ref_level][i] - y_lod[ref_level][i - 1]
+            x_len = x_idx[i] - x_idx[i - 1]
+            if repeat_num > 0:
+                x_sub = x_data[x_idx[i - 1]:x_idx[i], :]
+                x_sub = np.repeat(x_sub, repeat_num, axis=0)
+                out = np.vstack((out, x_sub))
+                if x_lod is not None:
+                    for j in xrange(repeat_num):
+                        out_lod[0].append(out_lod[0][-1] + x_len)
+
+        if x_lod is None:
             self.outputs = {'Out': out}
+        else:
+            self.outputs = {'Out': (out, out_lod)}
 
     def setUp(self):
         self.op_type = 'sequence_expand'
@@ -51,43 +75,44 @@ class TestSequenceExpand(OpTest):
     def test_check_grad(self):
         self.check_grad(["X"], "Out")
 
-    # class TestSequenceExpandCase1(TestSequenceExpand):
-    #     def set_data(self):
-    #         x_data = np.random.uniform(0.1, 1, [5, 1]).astype('float32')
-    #         x_lod = [[0, 2, 5]]
-    #         y_data = np.random.uniform(0.1, 1, [13, 1]).astype('float32')
-    #         y_lod = [[0, 2, 5], [0, 2, 4, 7, 10, 13]]
-    #         self.inputs = {'X': (x_data, x_lod), 'Y': (y_data, y_lod)}
-
-    # class TestSequenceExpandCase2(TestSequenceExpand):
-    #     def set_data(self):
-    #         x_data = np.random.uniform(0.1, 1, [1, 2, 2]).astype('float32')
-    #         x_lod = [[0, 1]]
-    #         y_data = np.random.uniform(0.1, 1, [2, 2, 2]).astype('float32')
-    #         y_lod = [[0, 2]]
-    #         self.inputs = {'X': (x_data, x_lod), 'Y': (y_data, y_lod)}
-
-    # class TestSequenceExpandCase3(TestSequenceExpand):
-    #     def set_data(self):
-    #         x_data = np.random.uniform(0.1, 1, [4, 1]).astype('float32')
-    #         x_lod = [[0, 1, 2, 3, 4]]
-    #         y_data = np.random.uniform(0.1, 1, [6, 1]).astype('float32')
-    #         y_lod = [[0, 2, 4, 4, 6]]
-    #         self.inputs = {'X': (x_data, x_lod), 'Y': (y_data, y_lod)}
-
-    # class TestSequenceExpandCase4(TestSequenceExpand):
-    #     def set_data(self):
-    #         x_data = np.array(
-    #             [0.1, 0.3, 0.2, 0.15, 0.25, 0.2, 0.15, 0.25, 0.1, 0.3]).reshape(
-    #                 [2, 5]).astype('float32')
-    #         x_lod = [[
-    #             0,
-    #             1,
-    #             2,
-    #         ]]
-    #         y_data = np.random.uniform(0.1, 1, [2, 1]).astype('float32')
-    #         y_lod = [[0, 1, 2], [0, 1, 2]]
-    #         self.inputs = {'X': (x_data, x_lod), 'Y': (y_data, y_lod)}
+
+class TestSequenceExpandCase1(TestSequenceExpand):
+    def set_data(self):
+        x_data = np.random.uniform(0.1, 1, [5, 1]).astype('float32')
+        x_lod = [[0, 2, 5]]
+        y_data = np.random.uniform(0.1, 1, [13, 1]).astype('float32')
+        y_lod = [[0, 2, 5], [0, 2, 4, 7, 10, 13]]
+        self.inputs = {'X': x_data, 'Y': (y_data, y_lod)}
+        self.attrs = {'ref_level': 0}
+
+
+class TestSequenceExpandCase2(TestSequenceExpand):
+    def set_data(self):
+        x_data = np.random.uniform(0.1, 1, [1, 2, 2]).astype('float32')
+        x_lod = [[0, 1]]
+        y_data = np.random.uniform(0.1, 1, [2, 2, 2]).astype('float32')
+        y_lod = [[0, 2], [0, 2]]
+        self.inputs = {'X': (x_data, x_lod), 'Y': (y_data, y_lod)}
+        self.attrs = {'ref_level': 0}
+
+
+class TestSequenceExpandCase3(TestSequenceExpand):
+    def set_data(self):
+        x_data = np.random.uniform(0.1, 1, [4, 1]).astype('float32')
+        x_lod = [[0, 1, 2, 3, 4]]
+        y_data = np.random.uniform(0.1, 1, [6, 1]).astype('float32')
+        y_lod = [[0, 2, 4, 4, 6]]
+        self.inputs = {'X': (x_data, x_lod), 'Y': (y_data, y_lod)}
+
+
+class TestSequenceExpandCase4(TestSequenceExpand):
+    def set_data(self):
+        data = [0.1, 0.3, 0.2, 0.15, 0.25, 0.2, 0.15, 0.25, 0.1, 0.3]
+        x_data = np.array(data).reshape([5, 2]).astype('float32')
+        x_lod = [[0, 2, 5]]
+        y_data = np.random.uniform(0.1, 1, [2, 1]).astype('float32')
+        y_lod = [[0, 1, 2], [0, 1, 2]]
+        self.inputs = {'X': (x_data, x_lod), 'Y': (y_data, y_lod)}
 
 
 if __name__ == '__main__':

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

@@ -29,15 +29,20 @@ class TestSoftmaxOp(OpTest):
     def setUp(self):
         self.op_type = "softmax"
         self.use_cudnn = False
-        self.inputs = {
-            'X': np.random.uniform(0.1, 1, [10, 10]).astype("float32")
-        }
-        self.outputs = {
-            'Out': np.apply_along_axis(stable_softmax, 1, self.inputs['X'])
+        self.use_mkldnn = False
+        self.dtype = np.float32
+        self.init_kernel_type()
+
+        x = np.random.uniform(0.1, 1, [10, 10]).astype(self.dtype)
+        out = np.apply_along_axis(stable_softmax, 1, x)
+        self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
+        self.outputs = {'Out': out}
+        self.attrs = {
+            'use_cudnn': self.use_cudnn,
+            'use_mkldnn': self.use_mkldnn
         }
-        self.attrs = {'use_cudnn': self.use_cudnn, }
 
-    def init_op_type(self):
+    def init_kernel_type(self):
         pass
 
     def test_check_output(self):
@@ -48,6 +53,8 @@ class TestSoftmaxOp(OpTest):
             self.check_output()
 
     def test_check_grad(self):
+        if self.dtype == np.float16:
+            return
         if self.use_cudnn:
             place = core.CUDAPlace(0)
             self.check_grad_with_place(
@@ -57,8 +64,25 @@ class TestSoftmaxOp(OpTest):
 
 
 class TestSoftmaxCUDNNOp(TestSoftmaxOp):
-    def init_op_type(self):
+    def init_kernel_type(self):
+        self.use_cudnn = True
+
+
+class TestSoftmaxFP16CUDNNOp(TestSoftmaxOp):
+    def init_kernel_type(self):
         self.use_cudnn = True
+        self.dtype = np.float16
+
+    def test_check_output(self):
+        if core.is_compiled_with_cuda():
+            place = core.CUDAPlace(0)
+            if core.is_float16_supported(place):
+                self.check_output_with_place(place, atol=1e-3)
+
+
+class TestSoftmaxMKLDNNOp(TestSoftmaxOp):
+    def init_kernel_type(self):
+        self.use_mkldnn = True
 
 
 if __name__ == "__main__":

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

+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestSplitIdsOp(OpTest):
+    def setUp(self):
+        self.op_type = "split_ids"
+        ids = np.array([[0], [2], [2], [3], [5], [5], [6]]).astype('int64')
+        out0 = np.array([[0], [3], [6]]).astype('int64')
+        out1 = np.array([[]]).astype('int64')
+        out2 = np.array([[2], [2], [5], [5]]).astype('int64')
+        self.inputs = {'Ids': ids}
+        self.outputs = {'Out': [('out0', out0), ('out1', out1), ('out2', out2)]}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+if __name__ == '__main__':
+    unittest.main()

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

@@ -126,7 +126,6 @@ class TestTensor(unittest.TestCase):
     def test_lod_tensor_gpu_init(self):
         if not core.is_compiled_with_cuda():
             return
-        scope = core.Scope()
         place = core.CUDAPlace(0)
         lod_py = [[0, 2, 5], [0, 2, 4, 5]]
         lod_tensor = core.LoDTensor()
@@ -144,6 +143,25 @@ class TestTensor(unittest.TestCase):
         self.assertAlmostEqual(2.0, lod_v[0, 0, 0, 1])
         self.assertListEqual(lod_py, lod_tensor.lod())
 
+    def test_empty_tensor(self):
+        place = core.CPUPlace()
+        scope = core.Scope()
+        var = scope.var("test_tensor")
+
+        tensor = var.get_tensor()
+
+        tensor.set_dims([0, 1])
+        tensor.alloc_float(place)
+
+        tensor_array = numpy.array(tensor)
+        self.assertEqual((0, 1), tensor_array.shape)
+
+        if core.is_compiled_with_cuda():
+            gpu_place = core.CUDAPlace(0)
+            tensor.alloc_float(gpu_place)
+            tensor_array = numpy.array(tensor)
+            self.assertEqual((0, 1), tensor_array.shape)
+
 
 if __name__ == '__main__':
     unittest.main()

python/paddle/trainer_config_helpers/activations.py

@@ -77,7 +77,7 @@ class SoftmaxActivation(BaseActivation):
 
     .. math::
 
-       P(y=j|x) = \\frac{e^{x_j}} {\\sum^K_{k=1} e^{x_j} }
+       P(y=j|x) = \\frac{e^{x_j}} {\\sum^K_{k=1} e^{x_k} }
     """
 
     def __init__(self):

python/paddle/trainer_config_helpers/layers.py

@@ -2747,17 +2747,17 @@ def img_pool_layer(input,
 
     ..  math::
 
-        w & = 1 + \\frac{ceil(input\_width + 2 * padding - pool\_size)}{stride}
+        w & = 1 + ceil(\\frac{input\_width + 2 * padding - pool\_size}{stride})
 
-        h & = 1 + \\frac{ceil(input\_height + 2 * padding\_y - pool\_size\_y)}{stride\_y}
+        h & = 1 + ceil(\\frac{input\_height + 2 * padding\_y - pool\_size\_y}{stride\_y})
 
     - ceil_mode=False:
 
     ..  math::
 
-        w & = 1 + \\frac{floor(input\_width + 2 * padding - pool\_size)}{stride}
+        w & = 1 + floor(\\frac{input\_width + 2 * padding - pool\_size}{stride})
 
-        h & = 1 + \\frac{floor(input\_height + 2 * padding\_y - pool\_size\_y)}{stride\_y}
+        h & = 1 + floor(\\frac{input\_height + 2 * padding\_y - pool\_size\_y}{stride\_y})
 
     The example usage is:
 

python/paddle/v2/reader/creator.py

@@ -16,7 +16,7 @@ Creator package contains some simple reader creator, which could
 be used in user program.
 """
 
-__all__ = ['np_array', 'text_file', "cloud_reader"]
+__all__ = ['np_array', 'text_file', 'recordio', 'cloud_reader']
 
 
 def np_array(x):