Merge branch 'develop' of github.com:baidu/Paddle into feature/c_api

CMakeLists.txt

@@ -12,19 +12,26 @@
 # See the License for the specific language governing permissions and
 # limitations under the License
 
-cmake_minimum_required(VERSION 3.0)
-
-project(paddle CXX C)
-
 set(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH} "${CMAKE_SOURCE_DIR}/cmake")
 set(PROJ_ROOT ${CMAKE_SOURCE_DIR})
 
+include(system)
+
+if(ANDROID)
+    cmake_minimum_required(VERSION 3.7)
+else()
+    cmake_minimum_required(VERSION 3.0)
+endif()
+
+project(paddle CXX C)
+
 find_package(Sphinx)
-find_package(CUDA QUIET)
+if(NOT CMAKE_CROSSCOMPILING)
+    find_package(CUDA QUIET)
+endif(NOT CMAKE_CROSSCOMPILING)
 find_package(Git REQUIRED)
 find_package(Threads REQUIRED)
 
-include(system)
 include(simd)
 
 ################################ Configurations #######################################
@@ -52,6 +59,21 @@ if(NOT CMAKE_BUILD_TYPE)
       FORCE)
 endif()
 
+if(ANDROID)
+    if(${CMAKE_SYSTEM_VERSION} VERSION_LESS "21")
+        message(FATAL_ERROR "Unsupport standalone toolchains with Android API level lower than 21")
+    endif()
+
+    set(WITH_GPU OFF CACHE STRING
+        "Disable GPU when cross-compiling for Android" FORCE)
+    set(WITH_AVX OFF CACHE STRING
+        "Disable AVX when cross-compiling for Android" FORCE)
+    set(WITH_PYTHON OFF CACHE STRING
+        "Disable PYTHON when cross-compiling for Android" FORCE)
+    set(WITH_RDMA OFF CACHE STRING
+        "Disable RDMA when cross-compiling for Android" FORCE)
+endif(ANDROID)
+
 set(THIRD_PARTY_PATH "${PROJ_ROOT}/third_party" CACHE STRING
   "A path setting third party libraries download & build directories.")
 
@@ -72,6 +94,7 @@ include(external/python)    # download, build, install python
 include(external/openblas)  # download, build, install openblas
 include(external/swig)      # download, build, install swig
 include(external/warpctc)   # download, build, install warpctc
+include(external/any)       # download libn::any
 
 include(package)            # set paddle packages
 include(cpplint)            # set paddle c++ style
@@ -82,7 +105,6 @@ include(flags)              # set paddle compile flags
 include(cudnn)              # set cudnn libraries
 include(version)            # set PADDLE_VERSION
 include(coveralls)          # set code coverage
-
 include(configure)          # add paddle env configuration
 
 include_directories("${PROJ_ROOT}")

Dockerfile

 # A image for building paddle binaries
 # Use cuda devel base image for both cpu and gpu environment
-FROM nvidia/cuda:7.5-cudnn5-devel-ubuntu14.04
+FROM nvidia/cuda:8.0-cudnn5-devel-ubuntu14.04
 MAINTAINER PaddlePaddle Authors <paddle-dev@baidu.com>
 
 ARG UBUNTU_MIRROR
 RUN /bin/bash -c 'if [[ -n ${UBUNTU_MIRROR} ]]; then sed -i 's#http://archive.ubuntu.com/ubuntu#${UBUNTU_MIRROR}#g' /etc/apt/sources.list; fi'
 
 # ENV variables
-ARG BUILD_WOBOQ
 ARG WITH_GPU
 ARG WITH_AVX
 ARG WITH_DOC
 ARG WITH_STYLE_CHECK
 
-ENV BUILD_WOBOQ=${BUILD_WOBOQ:-OFF}
+ENV WOBOQ OFF
 ENV WITH_GPU=${WITH_AVX:-OFF}
 ENV WITH_AVX=${WITH_AVX:-ON}
 ENV WITH_DOC=${WITH_DOC:-OFF}
@@ -37,18 +36,20 @@ RUN git config --global credential.helper store
 # Fix locales to en_US.UTF-8
 RUN localedef -i en_US -f UTF-8 en_US.UTF-8
 
+# FIXME: due to temporary ipykernel dependency issue, specify ipykernel jupyter
+# version util jupyter fixes this issue.
 RUN pip install --upgrade pip && \
     pip install -U 'protobuf==3.1.0' && \
     pip install -U wheel pillow BeautifulSoup && \
     pip install -U docopt PyYAML sphinx && \
     pip install -U sphinx-rtd-theme==0.1.9 recommonmark && \
-    pip install -U pre-commit 'requests==2.9.2' jupyter
+    pip install pre-commit 'requests==2.9.2' 'ipykernel==4.6.0' 'jupyter==1.0.0'
 
 RUN curl -sSL https://cmake.org/files/v3.4/cmake-3.4.1.tar.gz | tar -xz && \
     cd cmake-3.4.1 && ./bootstrap && make -j `nproc` && make install && \
     cd .. && rm -rf cmake-3.4.1
 
-VOLUME ["/usr/share/nginx/html/data", "/usr/share/nginx/html/paddle"]
+VOLUME ["/woboq_out"]
 
 # Configure OpenSSH server. c.f. https://docs.docker.com/engine/examples/running_ssh_service
 RUN mkdir /var/run/sshd

cmake/cblas.cmake

@@ -19,9 +19,9 @@ set(CBLAS_FOUND OFF)
 set(INTEL_ROOT "/opt/intel" CACHE PATH "Folder contains intel libs")
 set(MKL_ROOT ${INTEL_ROOT}/mkl CACHE PATH "Folder contains MKL")
 
-find_path(MKL_INCLUDE_DIR mkl.h PATHS
+find_path(MKL_INC_DIR mkl.h PATHS
   ${MKL_ROOT}/include)
-find_path(MKL_INCLUDE_DIR mkl_lapacke.h PATHS
+find_path(MKL_LAPACK_INC_DIR mkl_lapacke.h PATHS
   ${MKL_ROOT}/include)
 find_library(MKL_CORE_LIB NAMES mkl_core PATHS
   ${MKL_ROOT}/lib
@@ -34,15 +34,19 @@ find_library(MKL_INTEL_LP64 NAMES mkl_intel_lp64 PATHS
   ${MKL_ROOT}/lib/intel64)
 
 
-if(MKL_INCLUDE_DIR AND MKL_CORE_LIB AND MKL_SEQUENTIAL_LIB AND MKL_INTEL_LP64)
+if(MKL_INC_DIR AND MKL_CORE_LIB AND MKL_SEQUENTIAL_LIB AND MKL_INTEL_LP64)
   set(CBLAS_PROVIDER MKL)
-  set(CBLAS_INC_DIR ${MKL_INCLUDE_DIR})
+  set(CBLAS_INC_DIR ${MKL_INC_DIR})
   set(CBLAS_LIBRARIES ${MKL_INTEL_LP64}
           ${MKL_SEQUENTIAL_LIB}
           ${MKL_CORE_LIB})
   add_definitions(-DPADDLE_USE_MKL)
   message(STATUS "Found MKL (include: ${CBLAS_INC_DIR}, library: ${CBLAS_LIBRARIES})")
   set(CBLAS_FOUND ON)
+  if(${MKL_LAPACK_INC_DIR})
+    add_definitions(-DPADDLE_USE_LAPACK)
+    message(STATUS "Found lapack in MKL (include: ${MKL_LAPACK_INC_DIR})")
+  endif()
   return() # return file.
 endif()
 
@@ -68,13 +72,17 @@ find_library(ATLAS_CBLAS_LIB NAMES cblas libcblas.so.3
 find_library(ATLAS_LIB NAMES lapack_atlas liblapack_atlas.so.3
   PATHS ${ATLAS_LIB_SEARCH_PATHS})
 
-if(ATLAS_INC_DIR AND ATLAS_CBLAS_LIB AND ATLAS_LIB)
+if(ATLAS_INC_DIR AND ATLAS_CBLAS_LIB AND ATLAS_LIB AND NOT CBLAS_FOUND)
   set(CBLAS_PROVIDER ATLAS)
-  set(CBLAS_INC_DIR ${ATLAS_INC_DIR} ${ATLAS_CLAPACK_INC_DIR})
+  set(CBLAS_INC_DIR ${ATLAS_INC_DIR})
   set(CBLAS_LIBRARIES ${ATLAS_LIB} ${ATLAS_CBLAS_LIB})
   add_definitions(-DPADDLE_USE_ATLAS)  
-  message(STATUS "Found Atlas (include: ${CBLAS_INC_DIR}, library: ${CBLAS_LIBRARIES})")
+  message(STATUS "Found ATLAS (include: ${CBLAS_INC_DIR}, library: ${CBLAS_LIBRARIES})")
   set(CBLAS_FOUND ON)
+  if(ATLAS_CLAPACK_INC_DIR)
+    add_definitions(-DPADDLE_USE_LAPACK)
+    message(STATUS "Found lapack in ATLAS (include: ${ATLAS_CLAPACK_INC_DIR})")
+  endif()
   return()
 endif()
 
@@ -103,8 +111,12 @@ if(OPENBLAS_INC_DIR AND OPENBLAS_LIB)
   set(CBLAS_PROVIDER OPENBLAS)
   set(CBLAS_INC_DIR ${OPENBLAS_INC_DIR})
   set(CBLAS_LIBRARIES ${OPENBLAS_LIB})
-  message(STATUS "Found OpenBlas (include: ${CBLAS_INC_DIR}, library: ${CBLAS_LIBRARIES})")
+  message(STATUS "Found OpenBLAS (include: ${CBLAS_INC_DIR}, library: ${CBLAS_LIBRARIES})")
   set(CBLAS_FOUND ON)
+  if(OPENBLAS_LAPACKE_INC_DIR)
+    add_definitions(-DPADDLE_USE_LAPACK)
+    message(STATUS "Found lapack in OpenBLAS (include: ${OPENBLAS_LAPACKE_INC_DIR})")
+  endif()
   return()
 endif()
 

cmake/configure.cmake

@@ -32,6 +32,14 @@ if(NOT WITH_PROFILER)
     add_definitions(-DPADDLE_DISABLE_PROFILER)
 endif(NOT WITH_PROFILER)
 
+if(NOT CMAKE_CROSSCOMPILING)
+    if(WITH_AVX AND AVX_FOUND)
+        set(SIMD_FLAG ${AVX_FLAG})
+    elseif(SSE3_FOUND)
+        set(SIMD_FLAG ${SSE3_FLAG})
+    endif()
+endif()
+
 if(NOT WITH_GPU)
     add_definitions(-DPADDLE_ONLY_CPU)
     add_definitions(-DHPPL_STUB_FUNC)
@@ -48,21 +56,12 @@ else()
         message(FATAL_ERROR "Paddle need cudnn to compile")
     endif()
 
-    if(WITH_AVX)
-        set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} "-Xcompiler ${AVX_FLAG}")
-    else(WITH_AVX)
-        set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} "-Xcompiler ${SSE3_FLAG}")
-    endif(WITH_AVX)
+    set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} "-Xcompiler ${SIMD_FLAG}")
 
     # Include cuda and cudnn
     include_directories(${CUDNN_INCLUDE_DIR})
     include_directories(${CUDA_TOOLKIT_INCLUDE})
 endif(NOT WITH_GPU)
 
-if(WITH_AVX)
-    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${AVX_FLAG}")
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${AVX_FLAG}")
-else(WITH_AVX)
-    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${SSE3_FLAG}")
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${SSE3_FLAG}")
-endif(WITH_AVX)
+set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${SIMD_FLAG}")
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${SIMD_FLAG}")

cmake/cudnn.cmake

+if(NOT WITH_GPU)
+    return()
+endif()
+
 set(CUDNN_ROOT "" CACHE PATH "CUDNN ROOT")
 find_path(CUDNN_INCLUDE_DIR cudnn.h
     PATHS ${CUDNN_ROOT} ${CUDNN_ROOT}/include
@@ -11,6 +15,7 @@ list(APPEND CUDNN_CHECK_LIBRARY_DIRS
     ${CUDNN_ROOT}
     ${CUDNN_ROOT}/lib64
     ${CUDNN_ROOT}/lib
+    ${CUDNN_ROOT}/lib/x86_64-linux-gnu
     $ENV{CUDNN_ROOT}
     $ENV{CUDNN_ROOT}/lib64
     $ENV{CUDNN_ROOT}/lib

cmake/external/any.cmake

+INCLUDE(ExternalProject)
+
+SET(ANY_SOURCE_DIR ${THIRD_PARTY_PATH}/any)
+
+INCLUDE_DIRECTORIES(${ANY_SOURCE_DIR}/src/linb_any)
+
+ExternalProject_Add(
+    linb_any
+    ${EXTERNAL_PROJECT_LOG_ARGS}
+    GIT_REPOSITORY  "https://github.com/thelink2012/any.git"
+    GIT_TAG         "8fef1e93710a0edf8d7658999e284a1142c4c020"
+    PREFIX          ${ANY_SOURCE_DIR}
+    UPDATE_COMMAND  ""
+    CONFIGURE_COMMAND ""
+    BUILD_COMMAND     ""
+    INSTALL_COMMAND   ""
+    TEST_COMMAND      ""
+)
+
+add_definitions(-DANY_IMPL_ANY_CAST_MOVEABLE)

cmake/external/gflags.cmake

@@ -31,9 +31,17 @@ ExternalProject_Add(
     GIT_REPOSITORY  "https://github.com/gflags/gflags.git"
     PREFIX          ${GFLAGS_SOURCES_DIR}
     UPDATE_COMMAND  ""
+    CMAKE_ARGS      -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
+    CMAKE_ARGS      -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
+    CMAKE_ARGS      -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
+    CMAKE_ARGS      -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
     CMAKE_ARGS      -DCMAKE_INSTALL_PREFIX=${GFLAGS_INSTALL_DIR}
     CMAKE_ARGS      -DCMAKE_POSITION_INDEPENDENT_CODE=ON
     CMAKE_ARGS      -DBUILD_TESTING=OFF
+    CMAKE_ARGS      -DCMAKE_BUILD_TYPE=Release
+    CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${GFLAGS_INSTALL_DIR}
+                     -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
+                     -DCMAKE_BUILD_TYPE:STRING=Release
 )
 
 LIST(APPEND external_project_dependencies gflags)

cmake/external/glog.cmake

@@ -33,11 +33,19 @@ ExternalProject_Add(
     GIT_REPOSITORY  "https://github.com/google/glog.git"
     PREFIX          ${GLOG_SOURCES_DIR}
     UPDATE_COMMAND  ""
+    CMAKE_ARGS      -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
+    CMAKE_ARGS      -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
+    CMAKE_ARGS      -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
+    CMAKE_ARGS      -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
     CMAKE_ARGS      -DCMAKE_INSTALL_PREFIX=${GLOG_INSTALL_DIR}
     CMAKE_ARGS      -DCMAKE_POSITION_INDEPENDENT_CODE=ON
     CMAKE_ARGS      -DWITH_GFLAGS=ON
     CMAKE_ARGS      -Dgflags_DIR=${GFLAGS_INSTALL_DIR}/lib/cmake/gflags
     CMAKE_ARGS      -DBUILD_TESTING=OFF
+    CMAKE_ARGS      -DCMAKE_BUILD_TYPE=Release
+    CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${GLOG_INSTALL_DIR}
+                     -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
+                     -DCMAKE_BUILD_TYPE:STRING=Release
 )
 
 LIST(APPEND external_project_dependencies glog)

cmake/external/gtest.cmake

@@ -41,11 +41,19 @@ IF(WITH_TESTING)
         GIT_TAG         "release-1.8.0"
         PREFIX          ${GTEST_SOURCES_DIR}
         UPDATE_COMMAND  ""
-        CMAKE_ARGS      -DCMAKE_INSTALL_PREFIX:PATH=${GTEST_INSTALL_DIR}
+        CMAKE_ARGS      -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
+        CMAKE_ARGS      -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
+        CMAKE_ARGS      -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
+        CMAKE_ARGS      -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
+        CMAKE_ARGS      -DCMAKE_INSTALL_PREFIX=${GTEST_INSTALL_DIR}
         CMAKE_ARGS      -DCMAKE_POSITION_INDEPENDENT_CODE=ON
         CMAKE_ARGS      -DBUILD_GMOCK=ON
         CMAKE_ARGS      -Dgtest_disable_pthreads=ON
         CMAKE_ARGS      -Dgtest_force_shared_crt=ON
+        CMAKE_ARGS      -DCMAKE_BUILD_TYPE=Release
+        CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${GTEST_INSTALL_DIR}
+                         -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
+                         -DCMAKE_BUILD_TYPE:STRING=Release
     )
     LIST(APPEND external_project_dependencies gtest)
 ENDIF(WITH_TESTING)

cmake/external/openblas.cmake

@@ -29,7 +29,24 @@ IF(NOT ${CBLAS_FOUND})
 
     IF(CMAKE_COMPILER_IS_GNUCC)
         ENABLE_LANGUAGE(Fortran)
-        LIST(APPEND CBLAS_LIBRARIES gfortran pthread)
+        if (NOT CMAKE_Fortran_COMPILER_VERSION)
+          # cmake < 3.4 cannot get CMAKE_Fortran_COMPILER_VERSION directly.
+          execute_process(COMMAND ${CMAKE_Fortran_COMPILER} -dumpversion
+                    OUTPUT_VARIABLE CMAKE_Fortran_COMPILER_VERSION)
+        endif()
+        string(REGEX MATCHALL "[0-9]+" Fortran_VERSION ${CMAKE_Fortran_COMPILER_VERSION})
+        list(GET Fortran_VERSION 0 Fortran_MAJOR)
+        list(GET Fortran_VERSION 1 Fortran_MINOR)
+        find_library(GFORTRAN_LIBRARY NAMES gfortran PATHS 
+                     /lib
+                     /usr/lib
+                     /usr/lib/gcc/x86_64-linux-gnu/${Fortran_MAJOR}.${Fortran_MINOR}/
+                     /usr/lib/gcc/x86_64-linux-gnu/${Fortran_MAJOR}/)
+        if (NOT GFORTRAN_LIBRARY)
+            message(FATAL_ERROR "Cannot found gfortran library which it is used by openblas")
+        endif()
+        find_package(Threads REQUIRED)
+        LIST(APPEND CBLAS_LIBRARIES ${GFORTRAN_LIBRARY} ${CMAKE_THREAD_LIBS_INIT})
     ENDIF(CMAKE_COMPILER_IS_GNUCC)
 
     IF(NOT CMAKE_Fortran_COMPILER)
@@ -37,6 +54,8 @@ IF(NOT ${CBLAS_FOUND})
                 "you need to set gfortran compiler: cmake .. -DCMAKE_Fortran_COMPILER=...")
     ENDIF(NOT CMAKE_Fortran_COMPILER)
 
+    ADD_DEFINITIONS(-DPADDLE_USE_LAPACK)
+
     ExternalProject_Add(
         openblas
         ${EXTERNAL_PROJECT_LOG_ARGS}

cmake/external/protobuf.cmake

@@ -58,12 +58,20 @@ IF(NOT PROTOBUF_FOUND)
         GIT_TAG         "9f75c5aa851cd877fb0d93ccc31b8567a6706546"
         CONFIGURE_COMMAND
         ${CMAKE_COMMAND} ${PROTOBUF_SOURCES_DIR}/src/protobuf/cmake
-        -Dprotobuf_BUILD_TESTS=OFF
-        -DZLIB_ROOT:FILEPATH=${ZLIB_ROOT}
-        -DCMAKE_POSITION_INDEPENDENT_CODE=ON
-        -DCMAKE_BUILD_TYPE=Release
-        -DCMAKE_INSTALL_PREFIX=${PROTOBUF_INSTALL_DIR}
-        -DCMAKE_INSTALL_LIBDIR=lib
+            -Dprotobuf_BUILD_TESTS=OFF
+            -DZLIB_ROOT:FILEPATH=${ZLIB_ROOT}
+            -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
+            -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
+            -DCMAKE_POSITION_INDEPENDENT_CODE=ON
+            -DCMAKE_BUILD_TYPE=Release
+            -DCMAKE_INSTALL_PREFIX=${PROTOBUF_INSTALL_DIR}
+            -DCMAKE_INSTALL_LIBDIR=lib
+        CMAKE_CACHE_ARGS
+            -DCMAKE_INSTALL_PREFIX:PATH=${PROTOBUF_INSTALL_DIR}
+            -DCMAKE_BUILD_TYPE:STRING=Release
+            -DCMAKE_VERBOSE_MAKEFILE:BOOL=OFF
+            -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
+            -DZLIB_ROOT:STRING=${ZLIB_ROOT}
     )
 
     LIST(APPEND external_project_dependencies protobuf)

cmake/external/python.cmake

@@ -219,9 +219,9 @@ ELSE(PYTHONLIBS_FOUND AND PYTHONINTERP_FOUND)
 
 ENDIF(PYTHONLIBS_FOUND AND PYTHONINTERP_FOUND)
 
-INCLUDE_DIRECTORIES(${PYTHON_INCLUDE_DIR})
-INCLUDE_DIRECTORIES(${PYTHON_NUMPY_INCLUDE_DIR})
-
-IF(NOT WITH_PYTHON)
+IF(WITH_PYTHON)
+    INCLUDE_DIRECTORIES(${PYTHON_INCLUDE_DIR})
+    INCLUDE_DIRECTORIES(${PYTHON_NUMPY_INCLUDE_DIR})
+ELSE()
     SET(PYTHON_LIBRARIES "")
 ENDIF()

cmake/external/warpctc.cmake

@@ -50,12 +50,19 @@ ExternalProject_Add(
     UPDATE_COMMAND  ""
     CMAKE_ARGS      -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
     CMAKE_ARGS      -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
+    CMAKE_ARGS      -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
+    CMAKE_ARGS      -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
     CMAKE_ARGS      -DCMAKE_INSTALL_PREFIX=${WARPCTC_INSTALL_DIR}
     CMAKE_ARGS      -DWITH_GPU=${WITH_GPU}
     CMAKE_ARGS      -DWITH_OMP=${USE_OMP}
     CMAKE_ARGS      -DWITH_TORCH=OFF
-    CMAKE_ARGS      -DCMAKE_DISABLE_FIND_PACKAGE_Torch=TRUE
+    CMAKE_ARGS      -DCMAKE_DISABLE_FIND_PACKAGE_Torch=ON
     CMAKE_ARGS      -DBUILD_SHARED=ON
+    CMAKE_ARGS      -DCMAKE_POSITION_INDEPENDENT_CODE=ON
+    CMAKE_ARGS      -DCMAKE_BUILD_TYPE=Release
+    CMAKE_CACHE_ARGS -DCMAKE_BUILD_TYPE:STRING=Release
+                     -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
+                     -DCMAKE_INSTALL_PREFIX:PATH=${WARPCTC_INSTALL_DIR}
 )
 
 LIST(APPEND external_project_dependencies warpctc)

cmake/external/zlib.cmake

@@ -22,7 +22,7 @@ SET(ZLIB_INCLUDE_DIR "${ZLIB_INSTALL_DIR}/include" CACHE PATH "zlib include dire
 IF(WIN32)
   SET(ZLIB_LIBRARIES "${ZLIB_INSTALL_DIR}/lib/zlibstatic.lib" CACHE FILEPATH "zlib library." FORCE)
 ELSE(WIN32)
-  set(ZLIB_LIBRARIES "${ZLIB_INSTALL_DIR}/lib/libz.a" CACHE FILEPATH "zlib library." FORCE)
+  SET(ZLIB_LIBRARIES "${ZLIB_INSTALL_DIR}/lib/libz.a" CACHE FILEPATH "zlib library." FORCE)
 ENDIF(WIN32)
 
 INCLUDE_DIRECTORIES(${ZLIB_INCLUDE_DIR})
@@ -34,10 +34,18 @@ ExternalProject_Add(
     GIT_TAG         "v1.2.8"
     PREFIX          ${ZLIB_SOURCES_DIR}
     UPDATE_COMMAND  ""
+    CMAKE_ARGS      -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
+    CMAKE_ARGS      -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
+    CMAKE_ARGS      -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
+    CMAKE_ARGS      -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
     CMAKE_ARGS      -DCMAKE_INSTALL_PREFIX=${ZLIB_INSTALL_DIR}
     CMAKE_ARGS      -DBUILD_SHARED_LIBS=OFF
     CMAKE_ARGS      -DCMAKE_POSITION_INDEPENDENT_CODE=ON
     CMAKE_ARGS      -DCMAKE_MACOSX_RPATH=ON
+    CMAKE_ARGS      -DCMAKE_BUILD_TYPE=Release
+    CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${ZLIB_INSTALL_DIR}
+                     -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
+                     -DCMAKE_BUILD_TYPE:STRING=Release
 )
 
 LIST(APPEND external_project_dependencies zlib)

cmake/flags.cmake

@@ -2,6 +2,7 @@
 include(CheckCXXCompilerFlag)
 include(CheckCCompilerFlag)
 include(CheckCXXSymbolExists)
+include(CheckTypeSize)
 
 function(CheckCompilerCXX11Flag)
     if(CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
@@ -25,7 +26,7 @@ function(CheckCompilerCXX11Flag)
 endfunction()
 
 CheckCompilerCXX11Flag()
-LIST(APPEND CMAKE_CXX_FLAGS -std=c++11)
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11")
 
 # safe_set_flag
 #
@@ -83,6 +84,17 @@ if(NOT UINT64_MAX_EXISTS)
   endif()
 endif()
 
+SET(CMAKE_EXTRA_INCLUDE_FILES "pthread.h")
+CHECK_TYPE_SIZE(pthread_spinlock_t SPINLOCK_FOUND)
+CHECK_TYPE_SIZE(pthread_barrier_t BARRIER_FOUND)
+if(SPINLOCK_FOUND)
+  add_definitions(-DPADDLE_USE_PTHREAD_SPINLOCK)
+endif(SPINLOCK_FOUND)
+if(BARRIER_FOUND)
+  add_definitions(-DPADDLE_USE_PTHREAD_BARRIER)
+endif(BARRIER_FOUND)
+SET(CMAKE_EXTRA_INCLUDE_FILES "")
+
 # Common flags. the compiler flag used for C/C++ sources whenever release or debug
 # Do not care if this flag is support for gcc.
 set(COMMON_FLAGS

cmake/simd.cmake

@@ -2,6 +2,7 @@
 # so that PaddlePaddle can unleash the vectorization power of muticore.
 
 INCLUDE(CheckCXXSourceRuns)
+INCLUDE(CheckCXXSourceCompiles)
 
 IF(CMAKE_COMPILER_IS_GNUCC OR CMAKE_COMPILER_IS_GNUCXX OR CMAKE_CXX_COMPILER_ID MATCHES "Clang")
     set(MMX_FLAG "-mmmx")
@@ -17,6 +18,8 @@ ELSEIF(MSVC)
     SET(AVX2_FLAG "/arch:AVX2")
 ENDIF()
 
+set(CMAKE_REQUIRED_FLAGS_RETAINED ${CMAKE_REQUIRED_FLAGS})
+
 # Check  MMX
 set(CMAKE_REQUIRED_FLAGS ${MMX_FLAG})
 CHECK_CXX_SOURCE_RUNS("
@@ -73,4 +76,5 @@ int main()
     return 0;
 }" AVX2_FOUND)
 
+set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_RETAINED})
 mark_as_advanced(MMX_FOUND SSE2_FOUND SSE3_FOUND AVX_FOUND AVX2_FOUND)

cmake/system.cmake

@@ -67,6 +67,12 @@ MARK_AS_ADVANCED(HOST_SYSTEM CPU_CORES)
 MESSAGE(STATUS "Found Paddle host system: ${HOST_SYSTEM}")
 MESSAGE(STATUS "Found Paddle host system's CPU: ${CPU_CORES} cores")
 
+IF(DEFINED CMAKE_SYSTEM_NAME)
+    IF(${CMAKE_SYSTEM_NAME} STREQUAL "Android")
+        SET(ANDROID TRUE)
+    ENDIF()
+ENDIF()
+
 # external dependencies log output
 SET(EXTERNAL_PROJECT_LOG_ARGS
     LOG_DOWNLOAD    0     # Wrap download in script to log output

cmake/util.cmake

@@ -90,6 +90,10 @@ function(link_paddle_exe TARGET_NAME)
         ${RDMA_LD_FLAGS}
         ${RDMA_LIBS})
 
+    if(ANDROID)
+        target_link_libraries(${TARGET_NAME} log)
+    endif(ANDROID)
+
     add_dependencies(${TARGET_NAME} ${external_project_dependencies})
 endfunction()
 

demo/seqToseq/api_train_v2.py

 import sys
+
 import paddle.v2 as paddle
 
 
-def seqToseq_net(source_dict_dim, target_dict_dim):
+def seqToseq_net(source_dict_dim, target_dict_dim, is_generating=False):
     ### Network Architecture
     word_vector_dim = 512  # dimension of word vector
     decoder_size = 512  # dimension of hidden unit in GRU Decoder network
     encoder_size = 512  # dimension of hidden unit in GRU Encoder network
 
+    beam_size = 3
+    max_length = 250
+
     #### Encoder
     src_word_id = paddle.layer.data(
         name='source_language_word',
@@ -67,79 +71,143 @@ def seqToseq_net(source_dict_dim, target_dict_dim):
     group_input2 = paddle.layer.StaticInputV2(input=encoded_proj, is_seq=True)
     group_inputs = [group_input1, group_input2]
 
-    trg_embedding = paddle.layer.embedding(
-        input=paddle.layer.data(
-            name='target_language_word',
-            type=paddle.data_type.integer_value_sequence(target_dict_dim)),
-        size=word_vector_dim,
-        param_attr=paddle.attr.ParamAttr(name='_target_language_embedding'))
-    group_inputs.append(trg_embedding)
-
-    # For decoder equipped with attention mechanism, in training,
-    # target embeding (the groudtruth) is the data input,
-    # while encoded source sequence is accessed to as an unbounded memory.
-    # Here, the StaticInput defines a read-only memory
-    # for the recurrent_group.
-    decoder = paddle.layer.recurrent_group(
-        name=decoder_group_name,
-        step=gru_decoder_with_attention,
-        input=group_inputs)
-
-    lbl = paddle.layer.data(
-        name='target_language_next_word',
-        type=paddle.data_type.integer_value_sequence(target_dict_dim))
-    cost = paddle.layer.classification_cost(input=decoder, label=lbl)
-
-    return cost
+    if not is_generating:
+        trg_embedding = paddle.layer.embedding(
+            input=paddle.layer.data(
+                name='target_language_word',
+                type=paddle.data_type.integer_value_sequence(target_dict_dim)),
+            size=word_vector_dim,
+            param_attr=paddle.attr.ParamAttr(name='_target_language_embedding'))
+        group_inputs.append(trg_embedding)
+
+        # For decoder equipped with attention mechanism, in training,
+        # target embeding (the groudtruth) is the data input,
+        # while encoded source sequence is accessed to as an unbounded memory.
+        # Here, the StaticInput defines a read-only memory
+        # for the recurrent_group.
+        decoder = paddle.layer.recurrent_group(
+            name=decoder_group_name,
+            step=gru_decoder_with_attention,
+            input=group_inputs)
+
+        lbl = paddle.layer.data(
+            name='target_language_next_word',
+            type=paddle.data_type.integer_value_sequence(target_dict_dim))
+        cost = paddle.layer.classification_cost(input=decoder, label=lbl)
+
+        return cost
+    else:
+        # In generation, the decoder predicts a next target word based on
+        # the encoded source sequence and the last generated target word.
+
+        # The encoded source sequence (encoder's output) must be specified by
+        # StaticInput, which is a read-only memory.
+        # Embedding of the last generated word is automatically gotten by
+        # GeneratedInputs, which is initialized by a start mark, such as <s>,
+        # and must be included in generation.
+
+        trg_embedding = paddle.layer.GeneratedInputV2(
+            size=target_dict_dim,
+            embedding_name='_target_language_embedding',
+            embedding_size=word_vector_dim)
+        group_inputs.append(trg_embedding)
+
+        beam_gen = paddle.layer.beam_search(
+            name=decoder_group_name,
+            step=gru_decoder_with_attention,
+            input=group_inputs,
+            bos_id=0,
+            eos_id=1,
+            beam_size=beam_size,
+            max_length=max_length)
+
+        return beam_gen
 
 
 def main():
     paddle.init(use_gpu=False, trainer_count=1)
+    is_generating = False
 
     # source and target dict dim.
     dict_size = 30000
     source_dict_dim = target_dict_dim = dict_size
 
-    # define network topology
-    cost = seqToseq_net(source_dict_dim, target_dict_dim)
-    parameters = paddle.parameters.create(cost)
-
-    # define optimize method and trainer
-    optimizer = paddle.optimizer.Adam(
-        learning_rate=5e-5,
-        regularization=paddle.optimizer.L2Regularization(rate=1e-3))
-    trainer = paddle.trainer.SGD(cost=cost,
-                                 parameters=parameters,
-                                 update_equation=optimizer)
-
-    # define data reader
-    feeding = {
-        'source_language_word': 0,
-        'target_language_word': 1,
-        'target_language_next_word': 2
-    }
-
-    wmt14_reader = paddle.batch(
-        paddle.reader.shuffle(
-            paddle.dataset.wmt14.train(dict_size=dict_size), buf_size=8192),
-        batch_size=5)
-
-    # define event_handler callback
-    def event_handler(event):
-        if isinstance(event, paddle.event.EndIteration):
-            if event.batch_id % 10 == 0:
-                print "\nPass %d, Batch %d, Cost %f, %s" % (
-                    event.pass_id, event.batch_id, event.cost, event.metrics)
+    # train the network
+    if not is_generating:
+        cost = seqToseq_net(source_dict_dim, target_dict_dim)
+        parameters = paddle.parameters.create(cost)
+
+        # define optimize method and trainer
+        optimizer = paddle.optimizer.Adam(
+            learning_rate=5e-5,
+            regularization=paddle.optimizer.L2Regularization(rate=8e-4))
+        trainer = paddle.trainer.SGD(cost=cost,
+                                     parameters=parameters,
+                                     update_equation=optimizer)
+        # define data reader
+        wmt14_reader = paddle.batch(
+            paddle.reader.shuffle(
+                paddle.dataset.wmt14.train(dict_size), buf_size=8192),
+            batch_size=5)
+
+        # define event_handler callback
+        def event_handler(event):
+            if isinstance(event, paddle.event.EndIteration):
+                if event.batch_id % 10 == 0:
+                    print "\nPass %d, Batch %d, Cost %f, %s" % (
+                        event.pass_id, event.batch_id, event.cost,
+                        event.metrics)
+                else:
+                    sys.stdout.write('.')
+                    sys.stdout.flush()
+
+        # start to train
+        trainer.train(
+            reader=wmt14_reader, event_handler=event_handler, num_passes=2)
+
+    # generate a english sequence to french
+    else:
+        # use the first 3 samples for generation
+        gen_creator = paddle.dataset.wmt14.gen(dict_size)
+        gen_data = []
+        gen_num = 3
+        for item in gen_creator():
+            gen_data.append((item[0], ))
+            if len(gen_data) == gen_num:
+                break
+
+        beam_gen = seqToseq_net(source_dict_dim, target_dict_dim, is_generating)
+        # get the pretrained model, whose bleu = 26.92
+        parameters = paddle.dataset.wmt14.model()
+        # prob is the prediction probabilities, and id is the prediction word. 
+        beam_result = paddle.infer(
+            output_layer=beam_gen,
+            parameters=parameters,
+            input=gen_data,
+            field=['prob', 'id'])
+
+        # get the dictionary
+        src_dict, trg_dict = paddle.dataset.wmt14.get_dict(dict_size)
+
+        # the delimited element of generated sequences is -1,
+        # the first element of each generated sequence is the sequence length
+        seq_list = []
+        seq = []
+        for w in beam_result[1]:
+            if w != -1:
+                seq.append(w)
             else:
-                sys.stdout.write('.')
-                sys.stdout.flush()
-
-    # start to train
-    trainer.train(
-        reader=wmt14_reader,
-        event_handler=event_handler,
-        num_passes=10000,
-        feeding=feeding)
+                seq_list.append(' '.join([trg_dict.get(w) for w in seq[1:]]))
+                seq = []
+
+        prob = beam_result[0]
+        beam_size = 3
+        for i in xrange(gen_num):
+            print "\n*******************************************************\n"
+            print "src:", ' '.join(
+                [src_dict.get(w) for w in gen_data[i][0]]), "\n"
+            for j in xrange(beam_size):
+                print "prob = %f:" % (prob[i][j]), seq_list[i * beam_size + j]
 
 
 if __name__ == '__main__':

doc/design/dist/README.md

+# Design Doc: Distributed Training
+
+## Objective
+
+In [this slides](https://www.slideshare.net/cxwangyi/paddlepaddle-a-complete-solution-for-businesses), we explained that we'd like PaddlePaddle running on general-purpose clusters like those managed by Kubernetes, so to address demands for AI from both Internet and non-Internet industries.
+
+This poses technical challenges to PaddlePaddle:
+
+1. Support fault-recovery.
+1. Support both offline and online training.
+1. [Serverless computing](https://en.wikipedia.org/wiki/Serverless_computing) of distributed training.
+
+
+## Training Job
+
+A training job will be created once user asks Paddle cloud to train a model. The training job is made up of different processes that collaboratively consume data and produce a trained model. There are three kinds of processes:
+
+1. the *master process*, which dispatches tasks to
+1. one or more *trainer processes*, which run distributed training and synchronize gradients/models via
+1. one or more *parameter server processes*, where each holds a shard of the global model.
+
+Their relation is illustrated in the following graph:
+
+<img src="src/paddle-model-sharding.png"/>
+
+### Master Process
+
+The master process will:
+
+- Partition a dataset into [tasks](#task) and dispatch tasks to trainers.
+- Keep track of training progress on the dataset with [task queue](#task-queue). A training job will iterate on the dataset for a full pass until it goes into next pass.
+
+
+#### Task 
+
+A task is a data shard to be trained. The total number of tasks will be much bigger than the total number of trainers. The number of data instances inside a task will be much bigger than the mini-batch size.
+
+#### Task Queue
+
+The master process has three task queues to track training progress. As illustrated in the graph below, Job A and Job B both have one master process. Each master process has three task queues.
+
+<img src="src/paddle-task-queues.png"/>
+
+- The todo queue holds tasks to be dispatched. When a job starts, the master process fills in the todo queue with all tasks.
+- The pending queue holds tasks that are currently training by trainers.
+- the done queue holds tasks that are already trained.
+
+The life cycle of a single task is illustrated below:
+
+<img src="src/paddle-task-states.png"/>
+
+1. When a new pass of training starts, all tasks will be placed in the todo queue.
+1. The master process will dispatch few tasks to each trainer at a time, puts them in the pending queue and waits for completion.
+1. The trainer will work on its tasks and tell the master process once a task is completed. The master process will dispatch a new task to that trainer.
+1. If a task timeout. the master process will move it back to the todo queue. The timeout count will increase by one. If the timeout count is above a threshold, the task is likely to cause a trainer to crash, so it will be discarded.
+1. The master process will move completed task to the done queue. When the todo queue is empty, the master process will start a new pass by moving all tasks in the done queue to todo queue and reset the timeout counter of all tasks to zero.
+
+### Trainer Process
+
+The trainer process will:
+
+- Receive tasks from the master.
+- Work on the tasks: calculate and upload gradient to parameter servers, and update local model by downloading new parameters from parameter servers.
+
+### Parameter Server Process
+
+Parameter server processes hold the parameters collaboratively. The parameters are partitioned on different parameter servers.
+
+The parameter server will:
+
+- Receive gradient from the trainers, update its parameters, and give the trainers the latest parameters.
+- Periodically save its parameters to distributed file system by overriding the previous save.
+
+### Optimization Algorithms
+
+The communication pattern between the trainers and the parameter servers depends on the category of optimization algorithm:
+
+- Synchronous Stochastic Gradient Descent (sync-SGD)
+
+	Parameter server will wait for all trainer finish n-th mini-batch calculation and send their gradients before broadcasting new parameters to every trainer. Every trainer will wait for the new parameters before starting n+1-th mini-batch.
+  
+- Asynchronous Stochastic Gradient Descent (async-SGD)
+
+	There will no synchronization between different trainers, and parameter server updates its parameter as soon as it receives new gradient:
+
+	- Each trainer uploads its accumulated gradient every n mini-batches.
+	- Every m mini-batches, the trainer downloads new parameters from parameter server.
+	- n and m do not have to be equal.
+
+## Fault Tolerant
+
+The training job will pause if the master processes is dead, or any of the parameter server process is dead. They will be started by [Kubernetes](https://kubernetes.io/) and recover in few minutes. Please refer to [fault recovery](#fault-recovery).
+
+The training job will continue to make progress if there is at least one training process running. The strategy depends on the type of optimization algorithm:
+
+- sync-SGD
+
+	TODO
+
+- async-SGD
+
+	Since async-SGD does not require synchronization between mini-batches, the system will by definition make process if at least one trainer is running.
+
+## Fault Recovery
+
+PaddlePaddle uses [etcd](https://github.com/coreos/etcd) to keep track of the states of processes. Because etcd is a distributed reliable key-value store, the restarted process can recover its states from etcd. The model parameters are periodically saved into distributed file system, so a restarted parameter server can recover its parameters from the saved file.
+
+Now we will introduce how each process recovers from a failure, the graph below shows how etcd is used:
+
+<img src="src/paddle-etcd.png"/>
+
+### Master Process
+
+When the master is started by the Kubernetes, it executes the following steps at startup:
+
+1. Grabs a unique *master* lock in etcd, which prevents concurrent master instantiations.
+1. Recovers the task queues from etcd if they already exist, otherwise, the master will create them.
+1. Watches the trainer prefix keys `/trainer/` on etcd to find the live trainers.
+1. Starts dispatching the tasks to the trainers, and updates task queue using an etcd transaction to ensure lock is held during the update.
+
+The master process will kill itself if its etcd lease expires.
+
+When the master process is dead for any reason, Kubernetes will restart it. It will be online again with all states recovered from etcd in few minutes.
+
+### Trainer Process
+
+When the trainer is started by the Kubernetes, it executes the following steps at startup:
+
+1. Watches the available parameter server prefix keys `/ps/` on etcd and waits until the count of parameter servers reaches the desired count.
+1. Generates a unique ID, and sets key `/trainer/<unique ID>` with its contact address as value. The key will be deleted when the lease expires, so the master will be aware of the trainer being online and offline.
+1. Waits for tasks from the master to start training.
+
+If trainer's etcd lease expires, it will try set key `/trainer/<unique ID>` again so that the master process can discover the trainer again.
+
+### Parameter Server Process
+
+When the parameter server is started by Kubernetes, it executes the following steps at startup:
+
+1. Read desired total number of parameter servers from etcd `/ps_desired`
+1. Search through etcd keys `/ps/<index>` (`/ps/0`, `/ps/1`, ...) to find the first non-existant key whose index is smaller than the total number of parameter servers. Set the key using a transaction to avoid concurrent writes. The parameter server's index is inferred from the key name.
+
+	The desired number of parameter servers is 3:
+	
+	<img src="src/paddle-ps-0.png"/>
+	
+	The third parameter server joined:
+	
+	<img src="src/paddle-ps-1.png"/>
+
+1. The parameter server can load parameters if there are already saved parameters in the save path (inferred from its index).
+1. Now the parameter server is ready for the trainers' requests.
+
+If the parameter server's etcd lease expires, the parameter server will kill itself.
+
+
+## Dynamic Scaling
+
+### Trainer Scaling
+
+TODO
+
+### Parameter Server Scaling
+
+Not planned for v1.
+
+## Training Dataset Format
+
+TODO
+
+## User Interface
+
+TODO

doc/getstarted/build_and_install/docker_install_cn.rst

@@ -4,119 +4,139 @@ PaddlePaddle的Docker容器使用方式
 PaddlePaddle目前唯一官方支持的运行的方式是Docker容器。因为Docker能在所有主要操作系统（包括Linux，Mac OS X和Windows）上运行。 请注意，您需要更改 `Dockers设置 <https://github.com/PaddlePaddle/Paddle/issues/627>`_ 才能充分利用Mac OS X和Windows上的硬件资源。
 
 
-纯CPU和GPU的docker镜像使用说明
+PaddlePaddle发布的docker镜像使用说明
 ------------------------------
 
-对于每一个PaddlePaddle版本，我们都会发布两个Docker镜像：纯CPU的和GPU的。
-我们通过设置 `dockerhub.com <https://hub.docker.com/r/paddledev/paddle/>`_ 自动生成最新的docker镜像：
-`paddledev/paddle:0.10.0rc1-cpu` 和 `paddledev/paddle:0.10.0rc1-gpu`。
+对于每一个PaddlePaddle版本，我们都会发布两种Docker镜像：开发镜像、运行镜像。运行镜像包括纯CPU版本和GPU版本以及其对应的非AVX版本。
+我们会在 `dockerhub.com <https://hub.docker.com/r/paddledev/paddle/>`_ 提供最新的docker镜像，可以在"tags"标签下找到最新的Paddle镜像版本。
+1. 开发镜像：:code:`paddlepaddle/paddle:<version>-dev`
 
-以交互容器方式运行纯CPU的镜像：
+    这个镜像包含了Paddle相关的开发工具以及编译和运行环境。用户可以使用开发镜像代替配置本地环境，完成开发，编译，发布，
+    文档编写等工作。由于不同的Paddle的版本可能需要不同的依赖和工具，所以如果需要自行配置开发环境需要考虑版本的因素。
+    开发镜像包含了以下工具：
+    - gcc/clang
+    - nvcc
+    - Python
+    - sphinx
+    - woboq
+    - sshd
+    很多开发者会使用远程的安装有GPU的服务器工作，用户可以使用ssh登录到这台服务器上并执行 :code:`docker exec`进入开发镜像并开始工作，
+    也可以在开发镜像中启动一个SSHD服务，方便开发者直接登录到镜像中进行开发:
 
-.. code-block:: bash
+    以交互容器方式运行开发镜像：
 
-    docker run -it --rm paddledev/paddle:0.10.0rc1-cpu /bin/bash
+    .. code-block:: bash
 
-或者，可以以后台进程方式运行容器：
+        docker run -it --rm paddledev/paddle:<version>-dev /bin/bash
 
-.. code-block:: bash
+    或者，可以以后台进程方式运行容器：
 
-    docker run -d -p 2202:22 -p 8888:8888 paddledev/paddle:0.10.0rc1-cpu
+    .. code-block:: bash
 
-然后用密码 :code:`root` SSH进入容器：
+        docker run -d -p 2202:22 -p 8888:8888 paddledev/paddle:<version>-dev
 
-.. code-block:: bash
+    然后用密码 :code:`root` SSH进入容器：
 
-    ssh -p 2202 root@localhost
+    .. code-block:: bash
 
-SSH方式的一个优点是我们可以从多个终端进入容器。比如，一个终端运行vi，另一个终端运行Python。另一个好处是我们可以把PaddlePaddle容器运行在远程服务器上，并在笔记本上通过SSH与其连接。
+        ssh -p 2202 root@localhost
 
+    SSH方式的一个优点是我们可以从多个终端进入容器。比如，一个终端运行vi，另一个终端运行Python。另一个好处是我们可以把PaddlePaddle容器运行在远程服务器上，并在笔记本上通过SSH与其连接。
 
-以上方法在GPU镜像里也能用－只是请不要忘记按装CUDA驱动，以及告诉Docker：
+2. 运行镜像：根据CPU、GPU和非AVX区分了如下4个镜像：
+    - GPU/AVX：:code:`paddlepaddle/paddle:<version>-gpu`
+    - GPU/no-AVX：:code:`paddlepaddle/paddle:<version>-gpu-noavx`
+    - CPU/AVX：:code:`paddlepaddle/paddle:<version>`
+    - CPU/no-AVX：:code:`paddlepaddle/paddle:<version>-noavx`
 
-.. code-block:: bash
+    纯CPU镜像以及GPU镜像都会用到AVX指令集，但是2008年之前生产的旧电脑不支持AVX。以下指令能检查Linux电脑是否支持AVX：
 
-    export CUDA_SO="$(\ls /usr/lib64/libcuda* | xargs -I{} echo '-v {}:{}') $(\ls /usr/lib64/libnvidia* | xargs -I{} echo '-v {}:{}')"
-    export DEVICES=$(\ls /dev/nvidia* | xargs -I{} echo '--device {}:{}')
-    docker run ${CUDA_SO} ${DEVICES} -it paddledev/paddle:0.10.0rc1-gpu
+    .. code-block:: bash
 
+       if cat /proc/cpuinfo | grep -i avx; then echo Yes; else echo No; fi
 
-运行PaddlePaddle书籍
----------------------
+    如果输出是No，就需要选择使用no-AVX的镜像
 
-Jupyter Notebook是一个开源的web程序，大家可以通过它制作和分享带有代码、公式、图表、文字的交互式文档。用户可以通过网页浏览文档。
+    以上方法在GPU镜像里也能用，只是请不要忘记提前在物理机上安装GPU最新驱动。
+    为了保证GPU驱动能够在镜像里面正常运行，我们推荐使用[nvidia-docker](https://github.com/NVIDIA/nvidia-docker)来运行镜像。
 
-PaddlePaddle书籍是为用户和开发者制作的一个交互式的Jupyter Nodebook。
-如果您想要更深入了解deep learning，PaddlePaddle书籍一定是您最好的选择。
+    .. code-block:: bash
 
-当您进入容器内之后，只用运行以下命令：
+        nvidia-docker run -it --rm paddledev/paddle:0.10.0rc1-gpu /bin/bash
 
-.. code-block:: bash
-        
-    jupyter notebook
+    注意: 如果使用nvidia-docker存在问题，你也许可以尝试更老的方法，具体如下，但是我们并不推荐这种方法。：
 
-然后在浏览器中输入以下网址：
-    
-.. code-block:: text
+    .. code-block:: bash
 
-    http://localhost:8888/
+        export CUDA_SO="$(\ls /usr/lib64/libcuda* | xargs -I{} echo '-v {}:{}') $(\ls /usr/lib64/libnvidia* | xargs -I{} echo '-v {}:{}')"
+        export DEVICES=$(\ls /dev/nvidia* | xargs -I{} echo '--device {}:{}')
+        docker run ${CUDA_SO} ${DEVICES} -it paddledev/paddle:<version>-gpu
 
-就这么简单，享受您的旅程！
+3. 使用运行镜像发布你的AI程序
+    假设您已经完成了一个AI训练的python程序 :code:`a.py`，这个程序是您在开发机上使用开发镜像完成开发。此时您可以运行这个命令在开发机上进行测试运行：
 
+    .. code-block:: bash
 
-非AVX镜像
----------
+        docker run -it -v $PWD:/work paddle /work/a.py
 
-纯CPU镜像以及GPU镜像都会用到AVX指令集，但是2008年之前生产的旧电脑不支持AVX。以下指令能检查Linux电脑是否支持AVX：
+    这里`a.py`包含的所有依赖假设都可以在Paddle的运行容器中。如果需要包含更多的依赖、或者需要发布您的应用的镜像，可以编写`Dockerfile`使用`FROM paddledev/paddle:<version>`
+    创建和发布自己的AI程序镜像。
 
-.. code-block:: bash
+运行PaddlePaddle书籍
+---------------------
 
-   if cat /proc/cpuinfo | grep -i avx; then echo Yes; else echo No; fi
+Jupyter Notebook是一个开源的web程序，大家可以通过它制作和分享带有代码、公式、图表、文字的交互式文档。用户可以通过网页浏览文档。
 
-如果输出是No，我们就需要手动编译一个非AVX版本的镜像：
+PaddlePaddle书籍是为用户和开发者制作的一个交互式的Jupyter Nodebook。
+如果您想要更深入了解deep learning，PaddlePaddle书籍一定是您最好的选择。
+
+我们提供可以直接运行PaddlePaddle书籍的docker镜像，直接运行：
 
 .. code-block:: bash
 
-   cd ~
-   git clone https://github.com/PaddlePaddle/Paddle.git
-   cd Paddle
-   docker build --build-arg WITH_AVX=OFF -t paddle:cpu-noavx -f paddle/scripts/docker/Dockerfile .
-   docker build --build-arg WITH_AVX=OFF -t paddle:gpu-noavx -f paddle/scripts/docker/Dockerfile.gpu .
+    docker run -p 8888:8888 paddlepaddle/book
 
+然后在浏览器中输入以下网址：
+
+.. code-block:: text
+
+    http://localhost:8888/
+
+就这么简单，享受您的旅程！
 
 通过Docker容器开发PaddlePaddle
 ------------------------------
 
-开发人员可以在Docker中开发PaddlePaddle。这样开发人员可以以一致的方式在不同的平台上工作 - Linux，Mac OS X和Windows。
+开发人员可以在Docker开发镜像中开发PaddlePaddle。这样开发人员可以以一致的方式在不同的平台上工作 - Linux，Mac OS X和Windows。
+
+1. 构建开发镜像
 
-1. 将开发环境构建为Docker镜像
-   
    .. code-block:: bash
 
       git clone --recursive https://github.com/PaddlePaddle/Paddle
       cd Paddle
-      docker build -t paddle:dev -f paddle/scripts/docker/Dockerfile .
+      docker build -t paddle:dev .
 
 
-   请注意，默认情况下，:code:`docker build` 不会将源码导入到镜像中并编译它。如果我们想这样做，需要设置一个参数：
+   请注意，默认情况下，:code:`docker build` 不会将源码导入到镜像中并编译它。如果我们想这样做，需要构建完开发镜像，然后执行：
 
    .. code-block:: bash
 
-      docker build -t paddle:dev -f paddle/scripts/docker/Dockerfile --build-arg BUILD_AND_INSTALL=ON .
+      docker run -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_AVX=ON" -e "TEST=OFF" paddle:dev
 
 
 2. 运行开发环境
 
    当我们编译好了 :code:`paddle:dev`， 我们可以在docker容器里做开发，源代码可以通过挂载本地文件来被载入Docker的开发环境里面：
-   
+
    .. code-block:: bash
 
-      docker run -d -p 2202:22 -v $PWD:/paddle paddle:dev
+      docker run -d -p 2202:22 -v $PWD:/paddle paddle:dev sshd
 
    以上代码会启动一个带有PaddlePaddle开发环境的docker容器，源代码会被挂载到 :code:`/paddle` 。
 
-   请注意， :code:`paddle:dev` 的默认入口是 :code:`sshd` 。以上的 :code:`docker run` 命令其实会启动一个在2202端口监听的SSHD服务器。这样，我们就能SSH进入我们的开发容器了：
-   
+   以上的 :code:`docker run` 命令其实会启动一个在2202端口监听的SSHD服务器。这样，我们就能SSH进入我们的开发容器了：
+
    .. code-block:: bash
 
       ssh root@localhost -p 2202
@@ -124,13 +144,13 @@ PaddlePaddle书籍是为用户和开发者制作的一个交互式的Jupyter Nod
 3. 在Docker开发环境中编译与安装PaddlPaddle代码
 
    当在容器里面的时候，可以用脚本 :code:`paddle/scripts/docker/build.sh` 来编译、安装与测试PaddlePaddle：
-   
+
    .. code-block:: bash
-		      
+
       /paddle/paddle/scripts/docker/build.sh
 
    以上指令会在 :code:`/paddle/build` 中编译PaddlePaddle。通过以下指令可以运行单元测试：
-   
+
    .. code-block:: bash
 
       cd /paddle/build
@@ -140,14 +160,14 @@ PaddlePaddle书籍是为用户和开发者制作的一个交互式的Jupyter Nod
 文档
 ----
 
-Paddle的Docker镜像带有一个通过 `woboq code browser
+Paddle的Docker开发镜像带有一个通过 `woboq code browser
 <https://github.com/woboq/woboq_codebrowser>`_ 生成的HTML版本的C++源代码，便于用户浏览C++源码。
 
 只要在Docker里启动PaddlePaddle的时候给它一个名字，就可以再运行另一个Nginx Docker镜像来服务HTML代码：
 
 .. code-block:: bash
 
-   docker run -d --name paddle-cpu-doc paddle:0.10.0rc1-cpu
+   docker run -d --name paddle-cpu-doc paddle:<version>-dev
    docker run -d --volumes-from paddle-cpu-doc -p 8088:80 nginx
 
 接着我们就能够打开浏览器在 http://localhost:8088/paddle/ 浏览代码。

doc/getstarted/build_and_install/docker_install_en.rst

@@ -8,173 +8,255 @@ Please be aware that you will need to change `Dockers settings
 <https://github.com/PaddlePaddle/Paddle/issues/627>`_ to make full use
 of your hardware resource on Mac OS X and Windows.
 
+Working With Docker
+-------------------
+
+Docker is simple as long as we understand a few basic concepts:
+
+- *image*: A Docker image is a pack of software. It could contain one or more programs and all their dependencies. For example, the PaddlePaddle's Docker image includes pre-built PaddlePaddle and Python and many Python packages. We can run a Docker image directly, other than installing all these software. We can type
+
+  .. code-block:: bash
+
+     docker images
+
+  to list all images in the system. We can also run
+
+  .. code-block:: bash
+		  
+     docker pull paddlepaddle/paddle:0.10.0rc2
+
+  to download a Docker image, paddlepaddle/paddle in this example,
+  from Dockerhub.com.
+
+- *container*: considering a Docker image a program, a container is a
+  "process" that runs the image. Indeed, a container is exactly an
+  operating system process, but with a virtualized filesystem, network
+  port space, and other virtualized environment. We can type
+
+  .. code-block:: bash
+
+     docker run paddlepaddle/paddle:0.10.0rc2
+
+  to start a container to run a Docker image, paddlepaddle/paddle in this example.
+
+- By default docker container have an isolated file system namespace,
+  we can not see the files in the host file system. By using *volume*,
+  mounted files in host will be visible inside docker container.
+  Following command will mount current dirctory into /data inside
+  docker container, run docker container from debian image with
+  command :code:`ls /data`.
+
+  .. code-block:: bash
+
+     docker run --rm -v $(pwd):/data debian ls /data
 
 Usage of CPU-only and GPU Images
 ----------------------------------
 
-For each version of PaddlePaddle, we release 2 Docker images, a
-CPU-only one and a CUDA GPU one.  We do so by configuring
-`dockerhub.com <https://hub.docker.com/r/paddledev/paddle/>`_
-automatically generate the latest docker images `paddledev/paddle:0.10.0rc1-cpu`
-and `paddledev/paddle:0.10.0rc1-gpu`.
+For each version of PaddlePaddle, we release two types of Docker images:
+development image and production image. Production image includes
+CPU-only version and a CUDA GPU version and their no-AVX versions. We
+put the docker images on `dockerhub.com
+<https://hub.docker.com/r/paddledev/paddle/>`_. You can find the
+latest versions under "tags" tab at dockerhub.com
 
-To run the CPU-only image as an interactive container:
+1. Production images, this image might have multiple variants:
 
-.. code-block:: bash
+   - GPU/AVX：:code:`paddlepaddle/paddle:<version>-gpu`
+   - GPU/no-AVX：:code:`paddlepaddle/paddle:<version>-gpu-noavx`
+   - CPU/AVX：:code:`paddlepaddle/paddle:<version>`
+   - CPU/no-AVX：:code:`paddlepaddle/paddle:<version>-noavx`
 
-    docker run -it --rm paddledev/paddle:0.10.0rc1-cpu /bin/bash
+   Please be aware that the CPU-only and the GPU images both use the
+   AVX instruction set, but old computers produced before 2008 do not
+   support AVX.  The following command checks if your Linux computer
+   supports AVX:
 
-or, we can run it as a daemon container
+   .. code-block:: bash
 
-.. code-block:: bash
+      if cat /proc/cpuinfo | grep -i avx; then echo Yes; else echo No; fi
 
-    docker run -d -p 2202:22 -p 8888:8888 paddledev/paddle:0.10.0rc1-cpu
+   
+   To run the CPU-only image as an interactive container:
 
-and SSH to this container using password :code:`root`:
+   .. code-block:: bash
 
-.. code-block:: bash
+      docker run -it --rm paddlepaddle/paddle:0.10.0rc2 /bin/bash
 
-    ssh -p 2202 root@localhost
+   Above method work with the GPU image too -- the recommended way is
+   using `nvidia-docker <https://github.com/NVIDIA/nvidia-docker>`_.
 
-An advantage of using SSH is that we can connect to PaddlePaddle from
-more than one terminals.  For example, one terminal running vi and
-another one running Python interpreter.  Another advantage is that we
-can run the PaddlePaddle container on a remote server and SSH to it
-from a laptop.
+   Please install nvidia-docker first following this `tutorial
+   <https://github.com/NVIDIA/nvidia-docker#quick-start>`_.
 
-Above methods work with the GPU image too -- just please don't forget
-to install CUDA driver and let Docker knows about it:
+   Now you can run a GPU image:
 
-.. code-block:: bash
+   .. code-block:: bash
 
-    export CUDA_SO="$(\ls /usr/lib64/libcuda* | xargs -I{} echo '-v {}:{}') $(\ls /usr/lib64/libnvidia* | xargs -I{} echo '-v {}:{}')"
-    export DEVICES=$(\ls /dev/nvidia* | xargs -I{} echo '--device {}:{}')
-    docker run ${CUDA_SO} ${DEVICES} -it paddledev/paddle:0.10.0rc1-gpu
+      nvidia-docker run -it --rm paddlepaddle/paddle:0.10.0rc2-gpu /bin/bash
 
+2. development image :code:`paddlepaddle/paddle:<version>-dev`
 
-PaddlePaddle Book
-------------------
+   This image has packed related develop tools and runtime
+   environment. Users and developers can use this image instead of
+   their own local computer to accomplish development, build,
+   releasing, document writing etc. While different version of paddle
+   may depends on different version of libraries and tools, if you
+   want to setup a local environment, you must pay attention to the
+   versions.  The development image contains:
+   
+   - gcc/clang
+   - nvcc
+   - Python
+   - sphinx
+   - woboq
+   - sshd
+     
+   Many developers use servers with GPUs, they can use ssh to login to
+   the server and run :code:`docker exec` to enter the docker
+   container and start their work.  Also they can start a development
+   docker image with SSHD service, so they can login to the container
+   and start work.
 
-The Jupyter Notebook is an open-source web application that allows
-you to create and share documents that contain live code, equations,
-visualizations and explanatory text in a single browser.
 
-PaddlePaddle Book is an interactive Jupyter Notebook for users and developers. 
-We already exposed port 8888 for this book. If you want to
-dig deeper into deep learning, PaddlePaddle Book definitely is your best choice.
+Train Model Using Python API
+----------------------------
 
-Once you are inside the container, simply issue the command:
+Our official docker image provides a runtime for PaddlePaddle
+programs. The typical workflow will be as follows:
+
+Create a directory as workspace:
 
 .. code-block:: bash
-        
-    jupyter notebook
 
-Then, you would back and paste the address into the local browser:
-    
-.. code-block:: text
+   mkdir ~/workspace
 
-    http://localhost:8888/
+Edit a PaddlePaddle python program using your favourite editor
 
-That's all. Enjoy your journey!
+.. code-block:: bash
 
+   emacs ~/workspace/example.py
 
-Non-AVX Images
---------------
+Run the program using docker:
 
-Please be aware that the CPU-only and the GPU images both use the AVX
-instruction set, but old computers produced before 2008 do not support
-AVX.  The following command checks if your Linux computer supports
-AVX:
+.. code-block:: bash
+
+   docker run --rm -v ~/workspace:/workspace paddlepaddle/paddle:0.10.0rc2 python /workspace/example.py
+
+Or if you are using GPU for training:
 
 .. code-block:: bash
 
-   if cat /proc/cpuinfo | grep -i avx; then echo Yes; else echo No; fi
+   nvidia-docker run --rm -v ~/workspace:/workspace paddlepaddle/paddle:0.10.0rc2-gpu python /workspace/example.py
 
+Above commands will start a docker container by running :code:`python
+/workspace/example.py`. It will stop once :code:`python
+/workspace/example.py` finishes.
 
-If it doesn't, we will need to build non-AVX images manually from
-source code:
+Another way is to tell docker to start a :code:`/bin/bash` session and
+run PaddlePaddle program interactively:
 
 .. code-block:: bash
 
-   cd ~
-   git clone https://github.com/PaddlePaddle/Paddle.git
-   cd Paddle
-   docker build --build-arg WITH_AVX=OFF -t paddle:cpu-noavx -f paddle/scripts/docker/Dockerfile .
-   docker build --build-arg WITH_AVX=OFF -t paddle:gpu-noavx -f paddle/scripts/docker/Dockerfile.gpu .
+   docker run -it -v ~/workspace:/workspace paddlepaddle/paddle:0.10.0rc2 /bin/bash
+   # now we are inside docker container
+   cd /workspace
+   python example.py
 
+Running with GPU is identical:
 
-Development Using Docker
-------------------------
+.. code-block:: bash
 
-Developers can work on PaddlePaddle using Docker.  This allows
-developers to work on different platforms -- Linux, Mac OS X, and
-Windows -- in a consistent way.
+   nvidia-docker run -it -v ~/workspace:/workspace paddlepaddle/paddle:0.10.0rc2-gpu /bin/bash
+   # now we are inside docker container
+   cd /workspace
+   python example.py
 
-1. Build the Development Environment as a Docker Image
 
-   .. code-block:: bash
+Develop PaddlePaddle or Train Model Using C++ API
+---------------------------------------------------
 
-      git clone --recursive https://github.com/PaddlePaddle/Paddle
-      cd Paddle
-      docker build -t paddle:dev -f paddle/scripts/docker/Dockerfile .
+We will be using PaddlePaddle development image since it contains all
+compiling tools and dependencies.
 
+Let's clone PaddlePaddle repo first:
 
-   Note that by default :code:`docker build` wouldn't import source
-   tree into the image and build it.  If we want to do that, we need
-   to set a build arg:
+.. code-block:: bash
 
-   .. code-block:: bash
+   git clone https://github.com/PaddlePaddle/Paddle.git && cd Paddle
 
-      docker build -t paddle:dev -f paddle/scripts/docker/Dockerfile --build-arg BUILD_AND_INSTALL=ON .
+Mount both workspace folder and paddle code folder into docker
+container, so we can access them inside docker container. There are
+two ways of using PaddlePaddle development docker image:
 
+- run interactive bash directly
 
-2. Run the Development Environment
+  .. code-block:: bash
 
-   Once we got the image :code:`paddle:dev`, we can use it to develop
-   Paddle by mounting the local source code tree into a container that
-   runs the image:
+     # use nvidia-docker instead of docker if you need to use GPU
+     docker run -it -v ~/workspace:/workspace -v $(pwd):/paddle paddlepaddle/paddle:0.10.0rc2-dev /bin/bash
+     # now we are inside docker container
 
-   .. code-block:: bash
+- or, we can run it as a daemon container
 
-      docker run -d -p 2202:22 -p 8888:8888 -v $PWD:/paddle paddle:dev
+  .. code-block:: bash
 
-   This runs a container of the development environment Docker image
-   with the local source tree mounted to :code:`/paddle` of the
-   container.
+     # use nvidia-docker instead of docker if you need to use GPU
+     docker run -d -p 2202:22 -p 8888:8888 -v ~/workspace:/workspace -v $(pwd):/paddle paddlepaddle/paddle:0.10.0rc2-dev /usr/sbin/sshd -D
 
-   Note that the default entry-point of :code:`paddle:dev` is
-   :code:`sshd`, and above :code:`docker run` commands actually starts
-   an SSHD server listening on port 2202.  This allows us to log into
-   this container with:
+  and SSH to this container using password :code:`root`:
 
-   .. code-block:: bash
+  .. code-block:: bash
 
-      ssh root@localhost -p 2202
+     ssh -p 2202 root@localhost
 
-   Usually, I run above commands on my Mac.  I can also run them on a
-   GPU server :code:`xxx.yyy.zzz.www` and ssh from my Mac to it:
+  An advantage is that we can run the PaddlePaddle container on a
+  remote server and SSH to it from a laptop.
 
-   .. code-block:: bash
+When developing PaddlePaddle, you can edit PaddlePaddle source code
+from outside of docker container using your favoriate editor. To
+compile PaddlePaddle, run inside container:
 
-      my-mac$ ssh root@xxx.yyy.zzz.www -p 2202
+.. code-block:: bash
 
-3. Build and Install Using the Development Environment
+   WITH_GPU=OFF WITH_AVX=ON WITH_TEST=ON bash /paddle/paddle/scripts/docker/build.sh
 
-   Once I am in the container, I can use
-   :code:`paddle/scripts/docker/build.sh` to build, install, and test
-   Paddle:
+This builds everything about Paddle in :code:`/paddle/build`.  And we
+can run unit tests there:
 
-   .. code-block:: bash
+.. code-block:: bash
 
-      /paddle/paddle/scripts/docker/build.sh
+   cd /paddle/build
+   ctest
 
-   This builds everything about Paddle in :code:`/paddle/build`.  And
-   we can run unit tests there:
+When training model using C++ API, we can edit paddle program in
+~/workspace outside of docker. And build from /workspace inside of
+docker.
 
-   .. code-block:: bash
+PaddlePaddle Book
+------------------
+
+The Jupyter Notebook is an open-source web application that allows
+you to create and share documents that contain live code, equations,
+visualizations and explanatory text in a single browser.
+
+PaddlePaddle Book is an interactive Jupyter Notebook for users and developers.
+We already exposed port 8888 for this book. If you want to
+dig deeper into deep learning, PaddlePaddle Book definitely is your best choice.
+
+We provide a packaged book image, simply issue the command:
+
+.. code-block:: bash
 
-      cd /paddle/build
-      ctest
+    docker run -p 8888:8888 paddlepaddle/book
+
+Then, you would back and paste the address into the local browser:
+
+.. code-block:: text
+
+    http://localhost:8888/
+
+That's all. Enjoy your journey!
 
 
 Documentation
@@ -191,7 +273,7 @@ container:
 
 .. code-block:: bash
 
-   docker run -d --name paddle-cpu-doc paddle:0.10.0rc1-cpu
+   docker run -d --name paddle-cpu-doc paddle:<version>
    docker run -d --volumes-from paddle-cpu-doc -p 8088:80 nginx
 
 

doc/getstarted/build_and_install/ubuntu_install_cn.rst

@@ -46,7 +46,6 @@ PaddlePaddle提供了ubuntu 14.04 deb安装包。
         with_double: OFF
         with_python: ON
         with_rdma: OFF
-        with_metric_learning:
         with_timer: OFF
         with_predict_sdk:
 

doc/templates/conf.py.cn.in

@@ -55,6 +55,7 @@ extensions = [
     'sphinx.ext.napoleon',
     'sphinx.ext.graphviz'
 ]
+mathjax_path="https://cdn.bootcss.com/mathjax/2.7.0/MathJax.js"
 table_styling_embed_css = True
 
 autodoc_member_order = 'bysource'

paddle/CMakeLists.txt

@@ -8,16 +8,13 @@ add_subdirectory(gserver)
 add_subdirectory(pserver)
 add_subdirectory(trainer)
 add_subdirectory(scripts)
+
 if(WITH_C_API)
     add_subdirectory(capi)
 endif()
-configure_file(${CMAKE_CURRENT_SOURCE_DIR}/setup.py.in
-    ${CMAKE_CURRENT_SOURCE_DIR}/setup.py)
-
-if(WITH_PREDICT_SDK)
-    add_subdirectory(predict)
-endif()
 
 if(WITH_SWIG_PY)
+  configure_file(${CMAKE_CURRENT_SOURCE_DIR}/setup.py.in
+          ${CMAKE_CURRENT_SOURCE_DIR}/setup.py)
   add_subdirectory(api)
 endif()

paddle/api/CMakeLists.txt

@@ -76,8 +76,6 @@ SWIG_LINK_LIBRARIES(swig_paddle
     ${CMAKE_DL_LIBS}
     ${EXTERNAL_LIBS}
     ${CMAKE_THREAD_LIBS_INIT}
-    ${RDMA_LD_FLAGS}
-    ${RDMA_LIBS}
     ${START_END}
 )
 

paddle/cuda/include/hl_cpu_matrix_kernel.cuh

@@ -17,7 +17,11 @@ limitations under the License. */
 
 #include <stdio.h>
 #include "hl_base.h"
+#if defined(__ARM_NEON__) || defined(__ARM_NEON)
+#include "hl_neon_matrix_kernel.cuh"
+#else
 #include "hl_sse_matrix_kernel.cuh"
+#endif
 
 /**
  * @brief   cpu element wise unary operator.

paddle/cuda/include/hl_matrix_base.cuh

@@ -66,6 +66,8 @@ typedef BaseOp SSESquaredDiff;
 typedef BaseOp SSEFirst;
 typedef BaseOp SSESecond;
 typedef BaseOp SSEClassificationError;
+#elif defined(__ARM__NEON__) || defined(__ARM_NEON)
+#include "hl_matrix_base_neon.cuh"
 #else
 #include "hl_matrix_base_sse.cuh"
 #endif

paddle/cuda/include/hl_matrix_base_neon.cuh

+/* Copyright (c) 2016 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. */
+
+
+#ifndef HL_MATRIX_BASE_NEON_CUH_
+#define HL_MATRIX_BASE_NEON_CUH_
+
+namespace aggregate {
+class SSESum {
+public:
+  static const bool sse = true;
+  INLINE float32x4_t vecOp(const float32x4_t a, const float32x4_t b) const {
+    return vaddq_f32(a, b);
+  }
+};
+
+class SSEMax {
+public:
+  static const bool sse = true;
+  INLINE float32x4_t vecOp(const float32x4_t a, const float32x4_t b) const {
+    return vmaxq_f32(a, b);
+  }
+};
+
+class SSEMin {
+public:
+  static const bool sse = true;
+  INLINE float32x4_t vecOp(const float32x4_t a, const float32x4_t b) const {
+    return vminq_f32(a, b);
+  }
+};
+}  // namespace aggregate
+
+namespace base {
+namespace unary {
+class SSEIdentity {
+public:
+  static const bool sse = true;
+  INLINE float32x4_t vecOp(const float32x4_t a) const {
+    return a;
+  }
+};
+}  // namespace unary
+
+namespace binary {
+class SSEAdd {
+public:
+  static const bool sse = true;
+  INLINE float32x4_t vecOp(const float32x4_t a, const float32x4_t b) const {
+    return vaddq_f32(a, b);
+  }
+};
+
+class SSEAdd2 {
+public:
+  static const bool sse = true;
+  const real p1;
+  const real p2;
+  float32x4_t mp1;
+  float32x4_t mp2;
+
+public:
+  SSEAdd2(const real s1, const real s2) : p1(s1), p2(s2) {
+    mp1 = vdupq_n_f32(p1);
+    mp2 = vdupq_n_f32(p2);
+  }
+  INLINE float32x4_t vecOp(const float32x4_t a, const float32x4_t b) const {
+    float32x4_t tmp1, tmp2;
+    tmp1 = vmulq_f32(mp1, a);
+    tmp2 = vmulq_f32(mp2, b);
+    return vaddq_f32(tmp1, tmp2);
+  }
+};
+
+class SSESub {
+public:
+  static const bool sse = true;
+  INLINE float32x4_t vecOp(const float32x4_t a, const float32x4_t b) const {
+    return vsubq_f32(a, b);
+  }
+};
+
+class SSEMul {
+public:
+  static const bool sse = true;
+  INLINE float32x4_t vecOp(const float32x4_t a, const float32x4_t b) const {
+    return vmulq_f32(a, b);
+  }
+};
+
+class SSEDiv {
+public:
+  static const bool sse = true;
+  INLINE float32x4_t vecOp(const float32x4_t a, const float32x4_t b) const {
+    float32x4_t tmp;
+    tmp = vrecpeq_f32(b);
+    return vmulq_f32(a, tmp);
+  }
+};
+
+class SSESquaredDiff {
+public:
+  static const bool sse = true;
+  INLINE float32x4_t vecOp(const float32x4_t a, const float32x4_t b) const {
+    float32x4_t tmp;
+    tmp = vsubq_f32(a, b);
+    return vmulq_f32(tmp, tmp);
+  }
+};
+
+class SSEFirst {
+public:
+  static const bool sse = true;
+  INLINE float32x4_t vecOp(const float32x4_t a, const float32x4_t b) const {
+    return a;
+  }
+};
+
+class SSESecond {
+public:
+  static const bool sse = true;
+  INLINE float32x4_t vecOp(const float32x4_t a, const float32x4_t b) const {
+    return b;
+  }
+};
+
+class SSEClassificationError {
+public:
+  static const bool sse = true;
+  const real p;
+  float32x4_t mp;
+  uint32x4_t result;
+
+public:
+  explicit SSEClassificationError(const real s) : p(s) {
+    mp = vdupq_n_f32(p);
+    result = vdupq_n_u32(1);
+  }
+  // TODO: to be check
+  INLINE float32x4_t vecOp(const float32x4_t a, const float32x4_t b) const {
+    uint32x4_t tmp1 = vcgtq_f32(a, mp);
+    uint32x4_t tmp2 = vcgtq_f32(b, mp);
+    uint32x4_t tmp3 = veorq_u32(tmp1, tmp2);
+    return vcvtq_f32_u32(vandq_u32(tmp3, result));
+  }
+};
+}  // namespace binary
+}  // namespace base
+
+#endif /* HL_MATRIX_BASE_NEON_CUH_ */

paddle/cuda/include/hl_matrix_type.cuh

@@ -17,13 +17,20 @@ limitations under the License. */
 
 #include "hl_base.h"
 
-#ifdef __CUDA_ARCH__
+#if defined(__CUDA_ARCH__)
 #include <vector_types.h>
 #ifndef PADDLE_TYPE_DOUBLE
 typedef float4 vecType;
 #else
 typedef double2 vecType;
 #endif
+#elif (defined  __ARM_NEON) || (defined __ARM_NEON__)
+#include <arm_neon.h>
+#ifndef PADDLE_TYPE_DOUBLE
+typedef float32x4_t  vecType;
+#else
+#error NEON instructions does not support double precision
+#endif
 #else
 #include <mmintrin.h>
 #include <xmmintrin.h>

paddle/cuda/include/hl_neon_matrix_kernel.cuh

+/* Copyright (c) 2016 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. */
+
+
+#ifndef HL_NEON_MATRIX_KERNEL_CUH_
+#define HL_NEON_MATRIX_KERNEL_CUH_
+
+#include "hl_matrix_type.cuh"
+
+#define VECTOR_SIZE     16
+
+/* number of float in vector */
+#define     VECTOR_LEN      4
+#define     VECTOR_SET      vdupq_n_f32
+
+inline bool hl_check_align(size_t size) {
+  return !(size & (VECTOR_SIZE - 1));
+}
+
+inline bool hl_check_align(void *ptr) {
+  return hl_check_align(reinterpret_cast<size_t>(ptr));
+}
+
+template <class Agg>
+inline real hl_agg_op(Agg agg, vecType mm) {
+  float32x4_t rev = vrev64q_f32(mm);
+  float32x4_t tmp1 = agg.vecOp(rev, rev);
+  float32x2_t lo = vget_high_f32(rev);
+  float32x2_t hi = vget_low_f32(rev);
+  float32x4_t tmp2 = vcombine_f32(hi, lo);
+  float32x4_t ret = agg.vecOp(tmp1, tmp2);
+
+  return vgetq_lane_f32(ret, 0);
+}
+
+template <class Agg, class Op, class Saver>
+void hl_sse_matrix_row_op(Agg agg, Op op, Saver sv,
+                          int dimM, int dimN,
+                          real *dst, int ld,
+                          real *A, int lda) {
+  for (int i = 0; i < dimM; i++, A += lda) {
+    vecType mm = VECTOR_SET(agg.init());
+    vecType *a = (vecType*)(A);
+    for (int j = 0; j < dimN / VECTOR_LEN; j++, a++) {
+      mm = agg.vecOp(mm, op.vecOp(*a));
+    }
+
+    int rem = dimN % VECTOR_LEN;
+    if (rem) {
+      real tmp = hl_agg_op(agg, mm);
+      real *a = A + (dimN / VECTOR_LEN) * VECTOR_LEN;
+      for (int j = 0; j < rem; j++) {
+        tmp = agg(tmp, op(a[j]));
+      }
+      dst[i*ld] = sv(dst[i*ld], tmp);
+    } else {
+      dst[i*ld] = sv(dst[i*ld], hl_agg_op(agg, mm));
+    }
+  }
+}
+
+template <class Agg, class Op, class Saver>
+void hl_sse_matrix_row_op(Agg agg, Op op, Saver sv,
+                          int dimM, int dimN,
+                          real *dst, int ld,
+                          real *A, int lda,
+                          real *B, int ldb) {
+  for (int i = 0; i < dimM; i++, A += lda, B += ldb) {
+    vecType mm = VECTOR_SET(agg.init());
+    vecType *a = (vecType*)(A);
+    vecType *b = (vecType*)(B);
+    for (int j = 0; j < dimN / VECTOR_LEN; j++, a++, b++) {
+        mm = agg.vecOp(mm, op.vecOp(*a, *b));
+    }
+
+    int rem = dimN % VECTOR_LEN;
+    if (rem) {
+      real tmp = hl_agg_op(agg, mm);
+      real *a = A + (dimN / VECTOR_LEN) * VECTOR_LEN;
+      real *b = B + (dimN / VECTOR_LEN) * VECTOR_LEN;
+      for (int j = 0; j < rem; j++) {
+          tmp = agg(tmp, op(a[j], b[j]));
+      }
+      dst[i*ld] = sv(dst[i*ld], tmp);
+    } else {
+        dst[i*ld] = sv(dst[i*ld], hl_agg_op(agg, mm));
+    }
+  }
+}
+
+template <class Agg, class Op, class Saver>
+void hl_matrix_column_op(Agg agg, Op op, Saver sv,
+                         int dimM, int dimN,
+                         real *dst,
+                         real *A, int lda) {
+  for (int j = 0; j < dimN; j++) {
+    real tmp = agg.init();
+    for (int i = 0; i < dimM; i++) {
+        tmp = agg(tmp, op(A[i * lda + j]));
+    }
+    dst[j] = sv(dst[j], tmp);
+  }
+}
+
+template <class Agg, class Op, class Saver>
+void hl_matrix_column_op(Agg agg, Op op, Saver sv,
+                         int dimM, int dimN,
+                         real *dst,
+                         real *A, int lda,
+                         real *B, int ldb) {
+  for (int j = 0; j < dimN; j++) {
+    real tmp = agg.init();
+    for (int i = 0; i < dimM; i++) {
+        tmp = agg(tmp, op(A[i * lda + j], B[i * ldb + j]));
+    }
+    dst[j] = sv(dst[j], tmp);
+  }
+}
+
+/*
+ * MaxRow greater than or equal dimN
+ * dimN is multiples of VECTOR_LEN
+ * so rem <= MaxRow / VECTOR_LEN
+ */
+template <int MaxRow, class Agg, class Op, class Saver>
+void hl_sse_column_op_with_rem(Agg agg, Op op, Saver sv,
+                               int dimM, int dimN,
+                               real *dst,
+                               real *A, int lda) {
+  vecType mm[MaxRow / VECTOR_LEN];
+  for (int n = 0; n < MaxRow / VECTOR_LEN; n++) {
+    mm[n] = VECTOR_SET(agg.init());
+  }
+
+  for (int i = 0; i < dimM; i++) {
+    vecType *a = (vecType*)(A + i * lda);
+    for (int n = 0; n < dimN / VECTOR_LEN; n++) {
+      mm[n] = agg.vecOp(mm[n], op.vecOp(a[n]));
+    }
+  }
+
+  vecType *result = (vecType*)(dst);
+  for (int n = 0; n < dimN / VECTOR_LEN; n++) {
+    result[n] = sv.vecOp(result[n], mm[n]);
+  }
+
+  int rem = dimN % VECTOR_LEN;
+  if (rem) {
+    A += (dimN / VECTOR_LEN) * VECTOR_LEN;
+    dst += (dimN / VECTOR_LEN) * VECTOR_LEN;
+    hl_matrix_column_op(agg, op, sv, dimM, rem, dst, A, lda);
+  }
+}
+
+/*
+ * dimN is multiples of VECTOR_LEN
+ * dimN greater than Step
+ */
+template <int Step, class Agg, class Op, class Saver>
+void hl_sse_matrix_column_op(Agg agg, Op op, Saver sv,
+                             int dimM, int dimN,
+                             real *dst,
+                             real *A, int lda) {
+  for (int j = 0; j < dimN / Step; j++, dst += Step, A += Step) {
+    vecType mm[Step / VECTOR_LEN];
+    for (int n = 0; n < Step / VECTOR_LEN; n++) {
+      mm[n] = VECTOR_SET(agg.init());
+    }
+
+    for (int i = 0; i < dimM; i++) {
+      vecType *a = (vecType*)(A + i * lda);
+      for (int n = 0; n < Step / VECTOR_LEN; n++) {
+        mm[n] = agg.vecOp(mm[n], op.vecOp(a[n]));
+      }
+    }
+
+    vecType *result = (vecType*)(dst);
+    for (int n = 0; n < Step / VECTOR_LEN; n++) {
+      result[n] = sv.vecOp(result[n], mm[n]);
+    }
+  }
+
+  int remRow = dimN % Step;
+  if (remRow) {
+    hl_sse_column_op_with_rem<Step>(agg, op, sv, dimM, remRow, dst, A, lda);
+  }
+}
+
+template <class Agg, class Op, class Saver>
+void hl_sse_matrix_column_op(Agg agg, Op op, Saver sv,
+                             int dimM, int dimN,
+                             real *dst,
+                             real *A, int lda) {
+  if (dimN <= 16) {
+    hl_sse_matrix_column_op<16>(agg, op, sv, dimM, dimN, dst, A, lda);
+  } else if (dimN <= 32) {
+    hl_sse_matrix_column_op<32>(agg, op, sv, dimM, dimN, dst, A, lda);
+  } else if (dimN <= 1024 || dimM <= 512) {
+    hl_sse_matrix_column_op<64>(agg, op, sv, dimM, dimN, dst, A, lda);
+  } else {
+    hl_sse_matrix_column_op<1024>(agg, op, sv, dimM, dimN, dst, A, lda);
+  }
+}
+
+template <int MaxRow, class Agg, class Op, class Saver>
+void hl_sse_column_op_with_rem(Agg agg, Op op, Saver sv,
+                               int dimM, int dimN,
+                               real *dst,
+                               real *A, int lda,
+                               real *B, int ldb) {
+  vecType mm[MaxRow / VECTOR_LEN];
+  for (int n = 0; n < MaxRow / VECTOR_LEN; n++) {
+    mm[n] = VECTOR_SET(agg.init());
+  }
+
+  for (int i = 0; i < dimM; i++) {
+    vecType *a = (vecType*)(A + i * lda);
+    vecType *b = (vecType*)(B + i * ldb);
+    for (int n = 0; n < dimN / VECTOR_LEN; n++) {
+      mm[n] = agg.vecOp(mm[n], op.vecOp(a[n], b[n]));
+    }
+  }
+
+  vecType *result = (vecType*)(dst);
+  for (int n = 0; n < dimN / VECTOR_LEN; n++) {
+    result[n] = sv.vecOp(result[n], mm[n]);
+  }
+
+  int rem = dimN % VECTOR_LEN;
+  if (rem) {
+    A += (dimN / VECTOR_LEN) * VECTOR_LEN;
+    B += (dimN / VECTOR_LEN) * VECTOR_LEN;
+    dst += (dimN / VECTOR_LEN) * VECTOR_LEN;
+    hl_matrix_column_op(agg, op, sv, dimM, rem, dst, A, lda, B, ldb);
+  }
+}
+
+template <int Step, class Agg, class Op, class Saver>
+void hl_sse_matrix_column_op(Agg agg, Op op, Saver sv,
+                             int dimM, int dimN,
+                             real *dst,
+                             real *A, int lda,
+                             real *B, int ldb) {
+  for (int j = 0; j < dimN / Step; j++, dst += Step, A += Step, B += Step) {
+    vecType mm[Step / VECTOR_LEN];
+    for (int n = 0; n < Step / VECTOR_LEN; n++) {
+      mm[n] = VECTOR_SET(agg.init());
+    }
+
+    for (int i = 0; i < dimM; i++) {
+      vecType *a = (vecType*)(A + i * lda);
+      vecType *b = (vecType*)(B + i * ldb);
+      for (int n = 0; n < Step / VECTOR_LEN; n++) {
+        mm[n] = agg.vecOp(mm[n], op.vecOp(a[n], b[n]));
+      }
+    }
+
+    vecType *result = (vecType*)(dst);
+    for (int n = 0; n < Step / VECTOR_LEN; n++) {
+      result[n] = sv.vecOp(result[n], mm[n]);
+    }
+  }
+
+  int remRow = dimN % Step;
+  if (remRow) {
+    hl_sse_column_op_with_rem<Step>(
+        agg, op, sv, dimM, remRow, dst, A, lda, B, ldb);
+  }
+}
+
+template <class Agg, class Op, class Saver>
+void hl_sse_matrix_column_op(Agg agg, Op op, Saver sv,
+                             int dimM, int dimN,
+                             real *dst,
+                             real *A, int lda,
+                             real *B, int ldb) {
+  if (dimN <= 16) {
+    hl_sse_matrix_column_op<16>(agg, op, sv, dimM, dimN, dst, A, lda, B, ldb);
+  } else if (dimN <= 32) {
+    hl_sse_matrix_column_op<32>(agg, op, sv, dimM, dimN, dst, A, lda, B, ldb);
+  } else if (dimN <= 1024 || dimM <= 512) {
+    hl_sse_matrix_column_op<64>(agg, op, sv, dimM, dimN, dst, A, lda, B, ldb);
+  } else {
+    hl_sse_matrix_column_op<1024>(agg, op, sv, dimM, dimN, dst, A, lda, B, ldb);
+  }
+}
+
+#endif /* HL_NEON_MATRIX_KERNEL_CUH_ */

paddle/cuda/include/hl_sequence.h

@@ -159,4 +159,10 @@ extern void hl_sequence_avg_forward(real* dst,
                                     int width,
                                     const int mode);
 
+extern void hl_sequence_avg_backward(real* dst,
+                                     real* src,
+                                     const int* starts,
+                                     int height,
+                                     int width,
+                                     const int mode);
 #endif /* HL_SEQUENCE_H_ */

paddle/cuda/include/stub/hl_sequence_stub.h

@@ -57,4 +57,10 @@ inline void hl_sequence_avg_forward(real* dst,
                                     int width,
                                     const int mode) {}
 
+inline void hl_sequence_avg_backward(real* dst,
+                                     real* src,
+                                     const int* starts,
+                                     int height,
+                                     int width,
+                                     const int mode) {}
 #endif  // HL_SEQUENCE_STUB_H_

paddle/cuda/src/hl_cuda_sequence.cu

@@ -325,12 +325,12 @@ __global__ void KeSequenceAvgForward(real* dst,
     int seqLength = end - start;
     if (seqLength == 0) return;
     real sum = 0.0;
-    for (int i = 0; i < seqLength; i++) {
-      sum += src[(start + i) * width + col];
+    for (int i = start; i < end; i++) {
+      sum += src[i * width + col];
     }
     sum = mode == 1 ? sum :
         (mode == 0 ? sum / seqLength : sum * my_rsqrt((real)seqLength));
-    dst[row * width + col] = sum;
+    dst[gid] = sum;
   }
 }
 
@@ -354,3 +354,48 @@ void hl_sequence_avg_forward(real* dst,
            (dst, src, starts, height, width, mode);
   CHECK_SYNC("hl_sequence_avg_forward failed");
 }
+
+__global__ void KeSequenceAvgBackward(real* dst,
+                                      real* src,
+                                      const int* starts,
+                                      int height,
+                                      int width,
+                                      const int mode) {
+  int gid = blockIdx.x * blockDim.x + threadIdx.x;
+  int row = gid / width;
+  int col = gid % width;
+
+  if (gid < height * width) {
+    int start = starts[row];
+    int end = starts[row + 1];
+    int seqLength = end - start;
+    if (seqLength == 0) return;
+    real grad = src[gid];
+    grad = mode == 1 ? grad :
+        (mode == 0 ? grad / seqLength : grad * my_rsqrt((real)seqLength));
+    for (int i = start; i < end; i++) {
+      dst[i * width + col] += grad;
+    }
+  }
+}
+
+void hl_sequence_avg_backward(real* dst,
+                              real* src,
+                              const int* starts,
+                              int height,
+                              int width,
+                              const int mode) {
+  CHECK_NOTNULL(dst);
+  CHECK_NOTNULL(src);
+  CHECK_NOTNULL(starts);
+
+  int block = 512;
+  int grid = DIVUP(width * height, 512);
+
+  CHECK(mode == 0 || mode == 1 || mode == 2)
+    << "mode error in hl_sequence_avg_backward!";
+
+  KeSequenceAvgBackward<<< grid, block, 0, STREAM_DEFAULT >>>
+           (dst, src, starts, height, width, mode);
+  CHECK_SYNC("hl_sequence_avg_backward failed");
+}

paddle/function/Function.cpp

@@ -16,66 +16,6 @@ limitations under the License. */
 
 namespace paddle {
 
-template <>
-size_t FuncConfig::get<size_t>(const std::string& key) const {
-  auto it = valueMap_.find(key);
-  CHECK(it != valueMap_.end()) << "Cannot find value: '" << key << "'";
-  return it->second.s;
-}
-
-template <>
-real FuncConfig::get<real>(const std::string& key) const {
-  auto it = valueMap_.find(key);
-  CHECK(it != valueMap_.end()) << "Cannot find value: '" << key << "'";
-  return it->second.r;
-}
-
-template <>
-int FuncConfig::get<int>(const std::string& key) const {
-  auto it = valueMap_.find(key);
-  CHECK(it != valueMap_.end()) << "Cannot find value: '" << key << "'";
-  return it->second.i;
-}
-
-template <>
-bool FuncConfig::get<bool>(const std::string& key) const {
-  auto it = valueMap_.find(key);
-  CHECK(it != valueMap_.end()) << "Cannot find value: '" << key << "'";
-  return it->second.b;
-}
-
-template <>
-FuncConfig& FuncConfig::set<size_t>(const std::string& key, size_t v) {
-  CHECK_EQ(static_cast<int>(valueMap_.count(key)), 0) << "Duplicated value: "
-                                                      << key;
-  valueMap_[key].s = v;
-  return *this;
-}
-
-template <>
-FuncConfig& FuncConfig::set<real>(const std::string& key, real v) {
-  CHECK_EQ(static_cast<int>(valueMap_.count(key)), 0) << "Duplicated value: "
-                                                      << key;
-  valueMap_[key].r = v;
-  return *this;
-}
-
-template <>
-FuncConfig& FuncConfig::set<int>(const std::string& key, int v) {
-  CHECK_EQ(static_cast<int>(valueMap_.count(key)), 0) << "Duplicated value: "
-                                                      << key;
-  valueMap_[key].i = v;
-  return *this;
-}
-
-template <>
-FuncConfig& FuncConfig::set<bool>(const std::string& key, bool v) {
-  CHECK_EQ(static_cast<int>(valueMap_.count(key)), 0) << "Duplicated value: "
-                                                      << key;
-  valueMap_[key].b = v;
-  return *this;
-}
-
 void BufferArgs::addArg(const Matrix& arg,
                         const TensorShape& shape,
                         ArgType argType) {

paddle/function/Function.h

@@ -18,32 +18,49 @@ limitations under the License. */
 #include <vector>
 #include "BufferArg.h"
 #include "paddle/math/Matrix.h"
+#include "paddle/utils/Any.h"
 #include "paddle/utils/ClassRegistrar.h"
+#include "paddle/utils/Error.h"
 
 namespace paddle {
 
 /**
  * Function Configuration.
  * The argument type of Function::init.
- * Follow-up will consider moving this data structure to Proto inside.
  */
 class FuncConfig {
 public:
-  union value {
-    size_t s;
-    real r;
-    int i;
-    bool b;
-  };
-
   template <typename T>
-  T get(const std::string& key) const;
+  T get(const std::string& key, Error* err = nullptr) const {
+    try {
+      return any_cast<T>(valueMap_.at(key));
+    } catch (std::exception& e) {  // could be cast or out of range exception.
+      if (err) {
+        *err = Error(e.what());
+      } else {
+        LOG(FATAL) << "Cannot get key " << key << " with error " << e.what();
+      }
+      return T();
+    }
+  }
 
   template <typename T>
-  FuncConfig& set(const std::string& key, T v);
+  FuncConfig& set(const std::string& key, T v, Error* err = nullptr) {
+    auto it = valueMap_.find(key);
+    if (it != valueMap_.end()) {  // already contains key.
+      if (err) {
+        *err = Error("Key %s is already set in FuncConfig", key.c_str());
+      } else {
+        LOG(FATAL) << "Key " << key << " is already set in FuncConfig.";
+      }
+      return *this;
+    }
+    valueMap_[key] = any(v);
+    return *this;
+  }
 
 protected:
-  std::map<std::string, value> valueMap_;
+  mutable std::unordered_map<std::string, any> valueMap_;
 };
 
 /**

paddle/function/PadOp.cpp

@@ -25,9 +25,9 @@ void Pad<DEVICE_TYPE_CPU>(real* outputs,
                           const int inH,
                           const int inW,
                           const PadConf& pad) {
-  int cstart = pad.channelStart, cend = pad.channelEnd;
-  int hstart = pad.heightStart, hend = pad.heightEnd;
-  int wstart = pad.widthStart, wend = pad.widthEnd;
+  int cstart = pad.channel[0], cend = pad.channel[1];
+  int hstart = pad.height[0], hend = pad.height[1];
+  int wstart = pad.width[0], wend = pad.width[1];
   int outC = inC + cstart + cend;
   int outH = inH + hstart + hend;
   int outW = inW + wstart + wend;
@@ -51,9 +51,9 @@ void PadGrad<DEVICE_TYPE_CPU>(real* inGrad,
                               const int inH,
                               const int inW,
                               const PadConf& pad) {
-  int cstart = pad.channelStart, cend = pad.channelEnd;
-  int hstart = pad.heightStart, hend = pad.heightEnd;
-  int wstart = pad.widthStart, wend = pad.widthEnd;
+  int cstart = pad.channel[0], cend = pad.channel[1];
+  int hstart = pad.height[0], hend = pad.height[1];
+  int wstart = pad.width[0], wend = pad.width[1];
   int outC = inC + cstart + cend;
   int outH = inH + hstart + hend;
   int outW = inW + wstart + wend;
@@ -71,6 +71,12 @@ void PadGrad<DEVICE_TYPE_CPU>(real* inGrad,
   }
 }
 
+static inline PadConf castToPadConf(const FuncConfig& conf) {
+  return {conf.get<std::vector<uint32_t>>("channel"),
+          conf.get<std::vector<uint32_t>>("height"),
+          conf.get<std::vector<uint32_t>>("width")};
+}
+
 /**
  * \brief Padding zeros to input according to the specify dimension.
  *        The struct pad_ contains the padding size in each dimension.
@@ -127,14 +133,7 @@ void PadGrad<DEVICE_TYPE_CPU>(real* inGrad,
 template <DeviceType Device>
 class PadFunc : public FunctionBase {
 public:
-  void init(const FuncConfig& config) override {
-    pad_.channelStart = config.get<int>("cstart");
-    pad_.channelEnd = config.get<int>("cend");
-    pad_.heightStart = config.get<int>("hstart");
-    pad_.heightEnd = config.get<int>("hend");
-    pad_.widthStart = config.get<int>("wstart");
-    pad_.widthEnd = config.get<int>("wend");
-  }
+  void init(const FuncConfig& config) override { pad_ = castToPadConf(config); }
 
   void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
     CHECK_EQ(1UL, inputs.size());
@@ -175,14 +174,7 @@ private:
 template <DeviceType Device>
 class PadGradFunc : public FunctionBase {
 public:
-  void init(const FuncConfig& config) override {
-    pad_.channelStart = config.get<int>("cstart");
-    pad_.channelEnd = config.get<int>("cend");
-    pad_.heightStart = config.get<int>("hstart");
-    pad_.heightEnd = config.get<int>("hend");
-    pad_.widthStart = config.get<int>("wstart");
-    pad_.widthEnd = config.get<int>("wend");
-  }
+  void init(const FuncConfig& config) override { pad_ = castToPadConf(config); }
 
   void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
     CHECK_EQ(1UL, inputs.size());

paddle/function/PadOp.h

@@ -19,18 +19,12 @@ limitations under the License. */
 namespace paddle {
 
 struct PadConf {
-  /// how many values to add before the data along channel dimension.
-  int channelStart;
-  /// how many values to add after the data along channel dimension.
-  int channelEnd;
-  /// how many values to add before the data along height dimension.
-  int heightStart;
-  /// how many values to add after the data along height dimension.
-  int heightEnd;
-  /// how many values to add before the data along width dimension.
-  int widthStart;
-  /// how many values to add after the data along width dimension.
-  int widthEnd;
+  /// how many values to add before/after the data along channel dimension.
+  std::vector<uint32_t> channel;
+  /// how many values to add before/after the data along height dimension.
+  std::vector<uint32_t> height;
+  /// how many values to add before/after the data along width dimension.
+  std::vector<uint32_t> width;
 };
 
 /**

paddle/function/PadOpGpu.cu

@@ -44,9 +44,9 @@ void Pad<DEVICE_TYPE_GPU>(real* outputs,
   size_t nth = num * inC * inH * inW;
   int blockSize = 1024;
   int gridSize = (nth + 1024 - 1) / 1024;
-  int cstart = pad.channelStart, cend = pad.channelEnd;
-  int hstart = pad.heightStart, hend = pad.heightEnd;
-  int wstart = pad.widthStart, wend = pad.widthEnd;
+  int cstart = pad.channel[0], cend = pad.channel[1];
+  int hstart = pad.height[0], hend = pad.height[1];
+  int wstart = pad.width[0], wend = pad.width[1];
   int outC = inC + cstart + cend;
   int outH = inH + hstart + hend;
   int outW = inW + wstart + wend;
@@ -83,9 +83,9 @@ void PadGrad<DEVICE_TYPE_GPU>(real* inGrad,
   int nth = num * inC * inH * inW;
   int blockSize = 1024;
   int gridSize = (nth + 1024 - 1) / 1024;
-  int cstart = pad.channelStart, cend = pad.channelEnd;
-  int hstart = pad.heightStart, hend = pad.heightEnd;
-  int wstart = pad.widthStart, wend = pad.widthEnd;
+  int cstart = pad.channel[0], cend = pad.channel[1];
+  int hstart = pad.height[0], hend = pad.height[1];
+  int wstart = pad.width[0], wend = pad.width[1];
   int outC = inC + cstart + cend;
   int outH = inH + hstart + hend;
   int outW = inW + wstart + wend;

paddle/function/PadOpTest.cpp

@@ -24,48 +24,22 @@ TEST(Pad, real) {
         for (size_t imgSizeW : {5, 32, 96}) {
           VLOG(3) << " numSamples=" << numSamples << " channels=" << channels
                   << " imgSizeH=" << imgSizeH << " imgSizeW=" << imgSizeW;
-
-          FunctionCompare compare("Pad",
-                                  FuncConfig()
-                                      .set("cstart", 2)
-                                      .set("cend", 3)
-                                      .set("hstart", 1)
-                                      .set("hend", 2)
-                                      .set("wstart", 3)
-                                      .set("wend", 2));
-          TensorShape inDims{numSamples, channels, imgSizeH, imgSizeW};
-          TensorShape outDims{
-              numSamples, channels + 5, imgSizeH + 3, imgSizeW + 5};
-          compare.addInputs(BufferArg(VALUE_TYPE_FLOAT, inDims));
-          compare.addOutputs(BufferArg(VALUE_TYPE_FLOAT, outDims, ASSIGN_TO));
-          compare.run();
-        }
-      }
-    }
-  }
-}
-
-TEST(PadGrad, real) {
-  for (size_t numSamples : {5, 32}) {
-    for (size_t channels : {1, 5, 32}) {
-      for (size_t imgSizeH : {5, 33, 100}) {
-        for (size_t imgSizeW : {5, 32, 96}) {
-          VLOG(3) << " numSamples=" << numSamples << " channels=" << channels
-                  << " imgSizeH=" << imgSizeH << " imgSizeW=" << imgSizeW;
-          FunctionCompare compare("PadGrad",
-                                  FuncConfig()
-                                      .set("cstart", 2)
-                                      .set("cend", 3)
-                                      .set("hstart", 1)
-                                      .set("hend", 2)
-                                      .set("wstart", 3)
-                                      .set("wend", 2));
-          TensorShape inDims{numSamples, channels, imgSizeH, imgSizeW};
-          TensorShape outDims{
-              numSamples, channels + 5, imgSizeH + 3, imgSizeW + 5};
-          compare.addInputs(BufferArg(VALUE_TYPE_FLOAT, outDims));
-          compare.addOutputs(BufferArg(VALUE_TYPE_FLOAT, inDims, ASSIGN_TO));
-          compare.run();
+          for (bool test_grad : {false, true}) {
+            FunctionCompare compare(
+                test_grad ? "PadGrad" : "Pad",
+                FuncConfig()
+                    .set<std::vector<uint32_t>>("channel", {2, 3})
+                    .set<std::vector<uint32_t>>("height", {1, 2})
+                    .set<std::vector<uint32_t>>("width", {3, 2}));
+            TensorShape inDims{numSamples, channels, imgSizeH, imgSizeW};
+            TensorShape outDims{
+                numSamples, channels + 5, imgSizeH + 3, imgSizeW + 5};
+            compare.addInputs(
+                BufferArg(VALUE_TYPE_FLOAT, test_grad ? outDims : inDims));
+            compare.addOutputs(BufferArg(
+                VALUE_TYPE_FLOAT, test_grad ? inDims : outDims, ASSIGN_TO));
+            compare.run();
+          }
         }
       }
     }

paddle/gserver/layers/AgentLayer.cpp

@@ -42,7 +42,8 @@ void AgentLayer::forward(PassType passType) {
   // get Arguments from real layers
   if (numSamples_ > 0 && numSamples_ < realHeight) {
     if (realOutput.ids) {
-      output_.ids->subVecFrom(*realOutput.ids, 0, numSamples_);
+      output_.ids =
+          IVector::create(realOutput.ids->getData(), numSamples_, useGpu_);
     } else {
       output_.subArgFrom(
           realOutput, /* offset */ 0, numSamples_, getSize(), useGpu_);

paddle/gserver/layers/AverageLayer.cpp

@@ -26,8 +26,6 @@ bool AverageLayer::init(const LayerMap& layerMap,
                         const ParameterMap& parameterMap) {
   SequencePoolLayer::init(layerMap, parameterMap);
 
-  dataMtx_ = Matrix::create(nullptr, 1, 1, false, useGpu_);
-  outMtx_ = Matrix::create(nullptr, 1, getSize(), false, useGpu_);
   // average strategy
   if (config_.average_strategy() == "average") {
     mode_ = kAverage;
@@ -60,43 +58,9 @@ void AverageLayer::forward(PassType passType) {
 void AverageLayer::backward(const UpdateCallback& callback) {
   SequencePoolLayer::backward(callback);
 
-  const int* starts = startPositions_->getData(false);
-  MatrixPtr grad = getInputGrad(0);
-
-  if (grad) {
-    size_t dim = getSize();
-    real* gradientData = getInputGrad(0)->getData();
-    real* gradient = getOutputGrad()->getData();
-    size_t numSequences = startPositions_->getSize() - 1;
-    for (size_t sequenceId = 0; sequenceId < numSequences; ++sequenceId) {
-      // TODO(Dangqingqing) optimization for GPU
-      int sequenceLength = starts[sequenceId + 1] - starts[sequenceId];
-      if (0 == sequenceLength) {
-        // empty sequence
-        continue;
-      }
-      dataMtx_->setData(
-          gradientData + starts[sequenceId] * dim, sequenceLength, dim);
-      outMtx_->setData(gradient + sequenceId * dim);
-      switch (mode_) {
-        case kAverage: {
-          // plain average
-          dataMtx_->addBias(*outMtx_, 1.0f / sequenceLength);
-          break;
-        }
-        case kSum: {
-          // sum instead of average
-          dataMtx_->addBias(*outMtx_, 1.0f);
-          break;
-        }
-        case kAverageSquareRootN: {
-          // divide by square root of sequenceLength
-          dataMtx_->addBias(*outMtx_, 1.0f / sqrt(sequenceLength));
-          break;
-        }
-        default: { LOG(FATAL) << "should not reach here"; }
-      }
-    }
+  if (getInputGrad(0)) {
+    getInputGrad(0)->sequenceAvgBackward(
+        *getOutputGrad(), *startPositions_->getVector(useGpu_), mode_);
   }
 }
 

paddle/gserver/layers/AverageLayer.h

@@ -45,8 +45,6 @@ public:
   void backward(const UpdateCallback& callback = nullptr) override;
 
 protected:
-  MatrixPtr outMtx_;
-  MatrixPtr dataMtx_;
   int mode_;
 };
 }  // namespace paddle

paddle/gserver/layers/ExpandConvBaseLayer.cpp

@@ -107,6 +107,10 @@ void ExpandConvBaseLayer::expandOneFrame(MatrixPtr image,
   int channel = isDeconv_ ? numFilters_ : channels_[inIdx];
 
   resetExpandInput(subK_[inIdx] * groups_[inIdx], subN_[inIdx]);
+
+  CHECK_EQ(image->getWidth(),
+           static_cast<size_t>(imgSizeH_[inIdx] * imgSizeW_[inIdx] * channel));
+
   real *imgData = image->getData() + startIdx * image->getWidth();
   MatrixPtr imageTmp =
       Matrix::create(imgData,

paddle/gserver/layers/HierarchicalSigmoidLayer.h

@@ -36,7 +36,7 @@ namespace paddle {
  * |   |- 5
  * |
  * |-*- 0
- * |- 1
+ *   |- 1
  * @endcode
  *
  * where * indicates an internal node, and each leaf node represents a class.

paddle/gserver/layers/PadLayer.cpp

@@ -36,12 +36,9 @@ bool PadLayer::init(const LayerMap& layerMap,
   CHECK_EQ(2, pad_conf.pad_c_size());
   CHECK_EQ(2, pad_conf.pad_h_size());
   CHECK_EQ(2, pad_conf.pad_w_size());
-  padc_.push_back(pad_conf.pad_c(0));
-  padc_.push_back(pad_conf.pad_c(1));
-  padh_.push_back(pad_conf.pad_h(0));
-  padh_.push_back(pad_conf.pad_h(1));
-  padw_.push_back(pad_conf.pad_w(0));
-  padw_.push_back(pad_conf.pad_w(1));
+  padc_ = {pad_conf.pad_c(0), pad_conf.pad_c(1)};
+  padh_ = {pad_conf.pad_h(0), pad_conf.pad_h(1)};
+  padw_ = {pad_conf.pad_w(0), pad_conf.pad_w(1)};
 
   outDims_ = TensorShape(4);
   setOutDims(0);
@@ -49,21 +46,15 @@ bool PadLayer::init(const LayerMap& layerMap,
   createFunction(forward_,
                  "Pad",
                  FuncConfig()
-                     .set("cstart", padc_[0])
-                     .set("cend", padc_[1])
-                     .set("hstart", padh_[0])
-                     .set("hend", padh_[1])
-                     .set("wstart", padw_[0])
-                     .set("wend", padw_[1]));
+                     .set("channel", padc_)
+                     .set("height", padh_)
+                     .set("width", padw_));
   createFunction(backward_,
                  "PadGrad",
                  FuncConfig()
-                     .set("cstart", padc_[0])
-                     .set("cend", padc_[1])
-                     .set("hstart", padh_[0])
-                     .set("hend", padh_[1])
-                     .set("wstart", padw_[0])
-                     .set("wend", padw_[1]));
+                     .set("channel", padc_)
+                     .set("height", padh_)
+                     .set("width", padw_));
 
   return true;
 }

paddle/gserver/layers/PadLayer.h

@@ -38,9 +38,9 @@ protected:
   void setOutDims(const size_t batchSize);
   void setTensorDim(const size_t batchSize);
 
-  std::vector<int> padc_;
-  std::vector<int> padh_;
-  std::vector<int> padw_;
+  std::vector<uint32_t> padc_;
+  std::vector<uint32_t> padh_;
+  std::vector<uint32_t> padw_;
   TensorShape inDims_;
   TensorShape outDims_;
 };

paddle/gserver/layers/SequenceLastInstanceLayer.cpp

@@ -25,6 +25,11 @@ namespace paddle {
  * Input: a sequence
  * If SequenceLevel = kNonseq:
  *   Output: a sequence containing only the last instance of the input sequence
+ *   If stride_ > 0:
+ *      Output: a shorten sequence. The operation of getting last instance of a
+ *              sequence is independently performed on every slice of the input
+ *              sequence, which is obtained by sliding a window with the window
+ *              size set to stride_.
  * If SequenceLevel = kSeq:
  *   Check input sequence must has sub-sequence
  *   Output: a sequence containing only the last instance of each sub-sequence
@@ -37,6 +42,7 @@ class SequenceLastInstanceLayer : public SequencePoolLayer {
 protected:
   MatrixPtr tmpSrc_;
   MatrixPtr tmpDest_;
+  std::vector<int> instanceIds_;
 
 public:
   explicit SequenceLastInstanceLayer(const LayerConfig& config)
@@ -54,6 +60,7 @@ REGISTER_LAYER(seqlastins, SequenceLastInstanceLayer);
 bool SequenceLastInstanceLayer::init(const LayerMap& layerMap,
                                      const ParameterMap& parameterMap) {
   SequencePoolLayer::init(layerMap, parameterMap);
+  reversed_ = config_.select_first();
 
   tmpSrc_ =
       Matrix::create(nullptr, /* height= */ 1, 1, /* trans= */ false, useGpu_);
@@ -66,7 +73,8 @@ bool SequenceLastInstanceLayer::init(const LayerMap& layerMap,
 void SequenceLastInstanceLayer::forward(PassType passType) {
   SequencePoolLayer::forward(passType);
 
-  const int* starts = startPositions_->getData(false);
+  auto starts = (stride_ > 0) ? stridePositions_->getData()
+                              : startPositions_->getData(false);
   MatrixPtr inputValue = getInputValue(0);
   MatrixPtr outputValue = getOutputValue();
 
@@ -74,9 +82,10 @@ void SequenceLastInstanceLayer::forward(PassType passType) {
     AsyncGpuBlock asyncGpuBlock;
     REGISTER_TIMER_INFO("SequenceLastInstanceLayerForward", getName().c_str());
 
+    instanceIds_.clear();
     for (size_t seqId = 0; seqId < newBatchSize_; ++seqId) {
-      int insId =
-          config_.select_first() ? starts[seqId] : starts[seqId + 1] - 1;
+      int insId = reversed_ ? starts[seqId] : starts[seqId + 1] - 1;
+      instanceIds_.push_back(insId);
 
       outputValue->subMatrix(seqId, 1, tmpDest_)
           ->assign(*(inputValue->subMatrix(insId, 1, tmpSrc_)));
@@ -96,18 +105,13 @@ void SequenceLastInstanceLayer::backward(const UpdateCallback& callback) {
 
   MatrixPtr inputGrad = getInputGrad(0);
   MatrixPtr outputGrad = getOutputGrad();
-  const int* starts = startPositions_->getData(false);
-  size_t numSequences = startPositions_->getSize() - 1;
 
   if (inputGrad) {
     AsyncGpuBlock asyncGpuBlock;
     REGISTER_TIMER_INFO("SequenceLastInstanceLayerBackward", getName().c_str());
 
-    for (size_t seqId = 0; seqId < numSequences; ++seqId) {
-      int insId =
-          config_.select_first() ? starts[seqId] : starts[seqId + 1] - 1;
-
-      inputGrad->subMatrix(insId, 1, tmpDest_)
+    for (size_t seqId = 0; seqId < newBatchSize_; ++seqId) {
+      inputGrad->subMatrix(instanceIds_[seqId], 1, tmpDest_)
           ->add(*(outputGrad->subMatrix(seqId, 1, tmpSrc_)));
     }
   }

paddle/gserver/layers/SequencePoolLayer.cpp

@@ -37,6 +37,7 @@ bool SequencePoolLayer::init(const LayerMap& layerMap,
   } else {
     LOG(FATAL) << "Unknown trans_type: " << config_.trans_type();
   }
+  stride_ = config_.seq_pool_stride();
   setNeedSequenceInfo(false);
   return true;
 }
@@ -55,8 +56,6 @@ void SequencePoolLayer::forward(PassType passType) {
   CHECK_EQ(starts->getData()[newBatchSize_], input.getBatchSize());
   CHECK_EQ(newBatchSize_, starts->getSize() - 1);
 
-  resetOutput(newBatchSize_, dim);
-
   /* If type_ = kNonSeq, both seq has or not has sub-seq degrade to a non-seq,
    * thus, in this case, output_ has no sequenceStartPositions.
    * If type_ = kSeq, seq has sub-seq degrades to a seq, thus, only in this
@@ -67,6 +66,15 @@ void SequencePoolLayer::forward(PassType passType) {
         << "when trans_type = seq, input must hasSubseq";
     output_.degradeSequence(input);
   }
+  if (stride_ > 0) {
+    CHECK_EQ(input.hasSubseq(), 0UL)
+        << "sequence stride pooling is invalid for hasSubseq now";
+    output_.poolSequenceWithStride(
+        input, stride_, &stridePositions_, reversed_);
+    newBatchSize_ = stridePositions_->getSize() - 1;
+  }
+
+  resetOutput(newBatchSize_, dim);
 }
 
 void SequencePoolLayer::backward(const UpdateCallback& callback) {

paddle/gserver/layers/SequencePoolLayer.h

@@ -26,6 +26,10 @@ namespace paddle {
  *    Output: output size is the number of input sequences (NOT input instances)
  *    output[i] = seqlastin/average/max_{for each instance in this
  * sequence}{input[i]}
+ *    If stride_ > 0:
+ *        Check input sequence must not have sub-sequence
+ *        Output: a shorten sequence, pooling is performed upon a small local
+ *                area
  * If SequenceLevel = kSeq:
  *    Check input sequence must has sub-sequence
  *    Output: output size is the number of input sub-sequences
@@ -42,6 +46,11 @@ protected:
   enum SequenceLevel { kNonSeq = 0, kSeq = 1 };
   size_t newBatchSize_;
   ICpuGpuVectorPtr startPositions_;
+  int stride_;
+  // Store the start position of each window.
+  IVectorPtr stridePositions_;
+  // Whether the input sequence is reversed or not.
+  bool reversed_ = false;
 
 public:
   explicit SequencePoolLayer(const LayerConfig& config) : Layer(config) {}

paddle/gserver/tests/LayerGradUtil.cpp

@@ -778,8 +778,10 @@ void testProjectionGrad(ProjectionConfig conf,
   config.biasSize = biasSize == 0 ? config.layerConfig.size() : biasSize;
   config.layerConfig.set_bias_size(config.biasSize);
   config.layerConfig.set_shared_biases(sharedBias);
-  config.inputDefs.push_back(
-      {inputType, "layer_0", conf.input_size(), parameterSize});
+  config.inputDefs.push_back({inputType,
+                              "layer_0",
+                              static_cast<size_t>(conf.input_size()),
+                              parameterSize});
   *config.layerConfig.add_inputs()->mutable_proj_conf() = conf;
   config.testState = testState;
   testLayerGrad(config, "mixed", batchSize, false, useGpu);

paddle/gserver/tests/test_LayerGrad.cpp

@@ -804,10 +804,14 @@ TEST(Layer, ExpandLayer) {
   testExpandLayer("seq", true);       // seq expand to hasSubseq
 }
 
-void testDegradeLayer(bool hasSubseq, string layer_type, string trans_type) {
+void testDegradeLayer(bool hasSubseq,
+                      string layer_type,
+                      string trans_type,
+                      int stride) {
   TestConfig config;
   config.layerConfig.set_type(layer_type);
   config.layerConfig.set_size(10);
+  config.layerConfig.set_seq_pool_stride(stride);
   config.biasSize = 0;
 
   config.inputDefs.push_back(
@@ -827,36 +831,46 @@ void testDegradeLayer(bool hasSubseq, string layer_type, string trans_type) {
   if (layer_type == "average") {
     for (auto strategy : {"average", "sum", "squarerootn"}) {
       LOG(INFO) << " hasSubseq=" << hasSubseq << " trans_type=" << trans_type
-                << " average_strategy=" << strategy;
+                << " average_strategy=" << strategy
+                << " seq_pool_stride=" << stride;
       config.layerConfig.set_average_strategy(strategy);
       testDegradeLayerGrad(config, layer_type);
     }
   } else {
-    LOG(INFO) << " hasSubseq=" << hasSubseq << " trans_type=" << trans_type;
+    LOG(INFO) << " hasSubseq=" << hasSubseq << " trans_type=" << trans_type
+              << " seq_pool_stride=" << stride;
     testDegradeLayerGrad(config, layer_type);
   }
 }
 
 TEST(Layer, MaxLayer) {
-  testDegradeLayer(false, "max", "non-seq");  // seq max to non-seq
-  testDegradeLayer(true, "max", "non-seq");   // hasSubseq max to non-seq
-  testDegradeLayer(true, "max", "seq");       // hasSubseq max to seq
+  testDegradeLayer(false, "max", "non-seq", -1);  // seq max to non-seq
+  testDegradeLayer(true, "max", "non-seq", -1);   // hasSubseq max to non-seq
+  testDegradeLayer(true, "max", "seq", -1);       // hasSubseq max to seq
 }
 
 TEST(Layer, SequenceLastInstanceLayer) {
   testDegradeLayer(false,
                    "seqlastins",
-                   "non-seq");  // seq seqlastins to non-seq
+                   "non-seq",
+                   -1);  // seq seqlastins to non-seq
+  testDegradeLayer(false,
+                   "seqlastins",
+                   "non-seq",
+                   5);  // seq seqlastins to a shorten seq, stride window = 5
   testDegradeLayer(true,
                    "seqlastins",
-                   "non-seq");  // hasSubseq seqlastins to non-seq
-  testDegradeLayer(true, "seqlastins", "seq");  // hasSubseq seqlastins to seq
+                   "non-seq",
+                   -1);  // hasSubseq seqlastins to non-seq
+  testDegradeLayer(
+      true, "seqlastins", "seq", -1);  // hasSubseq seqlastins to seq
 }
 
 TEST(Layer, AverageLayer) {
-  testDegradeLayer(false, "average", "non-seq");  // seq average to non-seq
-  testDegradeLayer(true, "average", "non-seq");  // hasSubseq average to non-seq
-  testDegradeLayer(true, "average", "seq");      // hasSubseq average to seq
+  testDegradeLayer(false, "average", "non-seq", -1);  // seq average to non-seq
+  testDegradeLayer(
+      true, "average", "non-seq", -1);           // hasSubseq average to non-seq
+  testDegradeLayer(true, "average", "seq", -1);  // hasSubseq average to seq
 }
 
 TEST(Layer, SequenceConcatLayer) {

paddle/math/MathFunctions.cpp

@@ -85,11 +85,16 @@ int getrf<float>(const CBLAS_ORDER order,
                  float* A,
                  const int lda,
                  int* ipiv) {
+#ifdef PADDLE_USE_LAPACK
 #ifdef PADDLE_USE_ATLAS
   return clapack_sgetrf(order, M, N, A, lda, ipiv);
 #else
   return LAPACKE_sgetrf(order, M, N, A, lda, ipiv);
 #endif
+#else
+  LOG(FATAL) << "Not implemented";
+#endif
+  return 0;
 }
 
 template <>
@@ -99,11 +104,16 @@ int getrf<double>(const CBLAS_ORDER order,
                   double* A,
                   const int lda,
                   int* ipiv) {
+#ifdef PADDLE_USE_LAPACK
 #ifdef PADDLE_USE_ATLAS
   return clapack_dgetrf(order, M, N, A, lda, ipiv);
 #else
   return LAPACKE_dgetrf(order, M, N, A, lda, ipiv);
 #endif
+#else
+  LOG(FATAL) << "Not implemented";
+#endif
+  return 0;
 }
 
 template <>
@@ -112,11 +122,16 @@ int getri<float>(const CBLAS_ORDER order,
                  float* A,
                  const int lda,
                  const int* ipiv) {
+#ifdef PADDLE_USE_LAPACK
 #ifdef PADDLE_USE_ATLAS
   return clapack_sgetri(order, N, A, lda, ipiv);
 #else
   return LAPACKE_sgetri(order, N, A, lda, ipiv);
 #endif
+#else
+  LOG(FATAL) << "Not implemented";
+#endif
+  return 0;
 }
 
 template <>
@@ -125,11 +140,16 @@ int getri<double>(const CBLAS_ORDER order,
                   double* A,
                   const int lda,
                   const int* ipiv) {
+#ifdef PADDLE_USE_LAPACK
 #ifdef PADDLE_USE_ATLAS
   return clapack_dgetri(order, N, A, lda, ipiv);
 #else
   return LAPACKE_dgetri(order, N, A, lda, ipiv);
 #endif
+#else
+  LOG(FATAL) << "Not implemented";
+#endif
+  return 0;
 }
 
 template <>

paddle/math/MathFunctions.h

@@ -17,11 +17,14 @@ limitations under the License. */
 
 #ifdef PADDLE_USE_MKL
 #include <mkl.h>
+#ifdef PADDLE_USE_LAPACK
 #include <mkl_lapacke.h>
+#endif
 #else
 extern "C" {
 #include <cblas.h>
 }
+#ifdef PADDLE_USE_LAPACK
 #ifdef PADDLE_USE_ATLAS
 extern "C" {
 #include <clapack.h>
@@ -30,6 +33,7 @@ extern "C" {
 #include <lapacke.h>
 #endif
 #endif
+#endif
 
 #include <cmath>
 

paddle/math/Matrix.cpp

@@ -483,6 +483,20 @@ void GpuMatrix::sequenceAvgForward(Matrix& a,
   hl_sequence_avg_forward(dst, src, starts, height, width, mode);
 }
 
+void GpuMatrix::sequenceAvgBackward(Matrix& a,
+                                    const IVector& startsPos,
+                                    int mode) {
+  size_t height = a.getHeight();
+  size_t width = getWidth();
+  CHECK_EQ(height, startsPos.getSize() - 1);
+  CHECK_EQ(width, a.getWidth());
+  real* dst = getData();
+  real* src = a.getData();
+  const int* starts = startsPos.getData();
+
+  hl_sequence_avg_backward(dst, src, starts, height, width, mode);
+}
+
 /* this = scaleAB*(a*b) +  scaleT*this */
 void GpuMatrix::mul(const GpuMatrix& a,
                     const GpuMatrix& b,
@@ -2304,6 +2318,41 @@ void CpuMatrix::sequenceAvgForward(Matrix& a,
   }
 }
 
+void CpuMatrix::sequenceAvgBackward(Matrix& a,
+                                    const IVector& startsPos,
+                                    int mode) {
+  size_t height = a.getHeight();
+  size_t width = getWidth();
+  CHECK_EQ(height, startsPos.getSize() - 1);
+  CHECK_EQ(width, a.getWidth());
+  real* dst = getData();
+  real* src = a.getData();
+  const int* starts = startsPos.getData();
+  MatrixPtr outMtx = Matrix::create(nullptr, 1, width, false, false);
+  MatrixPtr dataMtx = Matrix::create(nullptr, 1, width, false, false);
+  for (size_t i = 0; i < height; ++i) {
+    int sequenceLength = starts[i + 1] - starts[i];
+    if (0 == sequenceLength) {
+      // empty sequence
+      continue;
+    }
+    outMtx->setData(dst + starts[i] * width, sequenceLength, width);
+    dataMtx->setData(src + i * width);
+    if (mode == 0) {
+      // plain average
+      outMtx->addBias(*dataMtx, 1.0f / sequenceLength);
+    } else if (mode == 1) {
+      // sum instead of average
+      outMtx->addBias(*dataMtx, 1.0f);
+    } else if (mode == 2) {
+      // divide by square root of sequenceLength
+      outMtx->addBias(*dataMtx, 1.0f / std::sqrt(sequenceLength));
+    } else {
+      LOG(FATAL) << "should not reach here";
+    }
+  }
+}
+
 /* this = scaleAB*(a*b) + scaleT*this*/
 void CpuMatrix::mul(const Matrix& a,
                     const Matrix& b,
@@ -2377,41 +2426,8 @@ void CpuMatrix::mul(CpuMatrix* a, CpuMatrix* b, real scaleAB, real scaleT) {
   int lda = a->getStride();
   int ldb = b->getStride();
   int ldc = getStride();
-#ifndef PADDLE_TYPE_DOUBLE
-  cblas_sgemm(CblasRowMajor,
-              a_trans,
-              b_trans,
-              M,
-              N,
-              K,
-              scaleAB,
-              A,
-              lda,
-              B,
-              ldb,
-              scaleT,
-              C,
-              ldc);
-#else
-  cblas_dgemm(CblasRowMajor,
-              a_trans,
-              b_trans,
-              M,
-              N,
-              K,
-              scaleAB,
-              A,
-              lda,
-              B,
-              ldb,
-              scaleT,
-              C,
-              ldc);
-// TODO(yuyang18): Is gemm defined other place?
-#endif
-
-  VLOG(2) << " A[0]=" << A[0] << " A[1]=" << A[1] << " B[0]=" << B[0]
-          << " B[1]=" << B[1] << " C[0]=" << C[0] << " C[1]=" << C[1];
+  gemm<real>(
+      a_trans, b_trans, M, N, K, scaleAB, A, lda, B, ldb, scaleT, C, ldc);
 }
 
 void CpuMatrix::mul(

paddle/math/Matrix.h

@@ -461,6 +461,12 @@ public:
     LOG(FATAL) << "Not implemented";
   }
 
+  virtual void sequenceAvgBackward(Matrix& a,
+                                   const IVector& startsPos,
+                                   int mode) {
+    LOG(FATAL) << "Not implemented";
+  }
+
   /**
    * @code
    * this = scaleAB*(a*b) + scaleT*this
@@ -1203,6 +1209,7 @@ public:
   void collectSharedBias(Matrix& a, real scale);
 
   void sequenceAvgForward(Matrix& a, const IVector& startsPos, int mode);
+  void sequenceAvgBackward(Matrix& a, const IVector& startsPos, int mode);
 
   /**
    * @code
@@ -1619,6 +1626,7 @@ public:
   void collectSharedBias(Matrix& a, real scale);
 
   void sequenceAvgForward(Matrix& a, const IVector& startsPos, int mode);
+  void sequenceAvgBackward(Matrix& a, const IVector& startsPos, int mode);
 
   /**
    * @code

paddle/math/SIMDFunctions.cpp

@@ -13,119 +13,12 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "SIMDFunctions.h"
+#ifdef __SSE3__
 #include <immintrin.h>
+#endif
 #include <algorithm>
 
-#ifndef __AVX__
-static void addto_sse(float* a, const float* b, size_t len) {
-  int offset = len % 16;
-  __m128 ma0, ma1, ma2, ma3;
-  __m128 mb0, mb1, mb2, mb3;
-
-  for (unsigned int k = 0; k < len / 16; k++, a += 16, b += 16) {
-    ma0 = _mm_load_ps(a);
-    ma1 = _mm_load_ps(a + 4);
-    ma2 = _mm_load_ps(a + 8);
-    ma3 = _mm_load_ps(a + 12);
-
-    mb0 = _mm_load_ps(b);
-    mb1 = _mm_load_ps(b + 4);
-    mb2 = _mm_load_ps(b + 8);
-    mb3 = _mm_load_ps(b + 12);
-
-    ma0 = _mm_add_ps(ma0, mb0);
-    ma1 = _mm_add_ps(ma1, mb1);
-    ma2 = _mm_add_ps(ma2, mb2);
-    ma3 = _mm_add_ps(ma3, mb3);
-
-    _mm_store_ps(a, ma0);
-    _mm_store_ps(a + 4, ma1);
-    _mm_store_ps(a + 8, ma2);
-    _mm_store_ps(a + 12, ma3);
-  }
-
-  for (int i = 0; i < offset; i++) a[i] += b[i];
-}
-
-static void batch_addto_sse(float* a, const float* b[], int batch, size_t len) {
-  int offset = len % 16;
-
-  __m128 ma0, ma1, ma2, ma3;
-  __m128 mb0, mb1, mb2, mb3;
-
-  for (unsigned int k = 0; k < len / 16; k++, a += 16) {
-    ma0 = _mm_load_ps(a);
-    ma1 = _mm_load_ps(a + 4);
-    ma2 = _mm_load_ps(a + 8);
-    ma3 = _mm_load_ps(a + 12);
-
-    for (int i = 0; i < batch; i++) {
-      mb0 = _mm_load_ps(b[i]);
-      mb1 = _mm_load_ps(b[i] + 4);
-      mb2 = _mm_load_ps(b[i] + 8);
-      mb3 = _mm_load_ps(b[i] + 12);
-      ma0 = _mm_add_ps(ma0, mb0);
-      ma1 = _mm_add_ps(ma1, mb1);
-      ma2 = _mm_add_ps(ma2, mb2);
-      ma3 = _mm_add_ps(ma3, mb3);
-      b[i] += 16;
-    }
-
-    _mm_store_ps(a, ma0);
-    _mm_store_ps(a + 4, ma1);
-    _mm_store_ps(a + 8, ma2);
-    _mm_store_ps(a + 12, ma3);
-  }
-
-  for (int i = 0; i < offset; i++) {
-    for (int k = 0; k < batch; k++) a[i] += b[k][i];
-  }
-  return;
-}
-
-static void col_max_sse(float* result,
-                        const float* data,
-                        int dim,
-                        int numSamples) {
-  // first sample, direct copy
-  for (int d = 0; d < dim; ++d) {
-    result[d] = data[d];
-  }
-  int offset = dim % 16;
-  __m128 ma0, ma1, ma2, ma3;
-  __m128 mb0, mb1, mb2, mb3;
-  // first 16n dims
-  for (int k = 0; k < dim / 16; k++, result += 16, data += 16) {
-    ma0 = _mm_load_ps(result);
-    ma1 = _mm_load_ps(result + 4);
-    ma2 = _mm_load_ps(result + 8);
-    ma3 = _mm_load_ps(result + 12);
-    for (int i = 1; i < numSamples; i++) {
-      mb0 = _mm_load_ps(data + i * dim);
-      mb1 = _mm_load_ps(data + i * dim + 4);
-      mb2 = _mm_load_ps(data + i * dim + 8);
-      mb3 = _mm_load_ps(data + i * dim + 12);
-      ma0 = _mm_max_ps(ma0, mb0);
-      ma1 = _mm_max_ps(ma1, mb1);
-      ma2 = _mm_max_ps(ma2, mb2);
-      ma3 = _mm_max_ps(ma3, mb3);
-    }
-    _mm_store_ps(result, ma0);
-    _mm_store_ps(result + 4, ma1);
-    _mm_store_ps(result + 8, ma2);
-    _mm_store_ps(result + 12, ma3);
-  }
-  // last dims
-  for (int d = 0; d < offset; ++d) {
-    float sm = data[d];
-    for (int i = 1; i < numSamples; ++i) {
-      sm = std::max(sm, data[i * dim + d]);
-    }
-    result[d] = sm;
-  }
-}
-
-#else
+#ifdef __AVX__
 static void addto_avx(float* a, const float* b, size_t len) {
   int offset = len % 32;
 
@@ -355,17 +248,128 @@ static void decayL1_avx(
   }
 }
 
+#elif defined(__SSE3__)
+
+static void addto_sse(float* a, const float* b, size_t len) {
+  int offset = len % 16;
+  __m128 ma0, ma1, ma2, ma3;
+  __m128 mb0, mb1, mb2, mb3;
+
+  for (unsigned int k = 0; k < len / 16; k++, a += 16, b += 16) {
+    ma0 = _mm_load_ps(a);
+    ma1 = _mm_load_ps(a + 4);
+    ma2 = _mm_load_ps(a + 8);
+    ma3 = _mm_load_ps(a + 12);
+
+    mb0 = _mm_load_ps(b);
+    mb1 = _mm_load_ps(b + 4);
+    mb2 = _mm_load_ps(b + 8);
+    mb3 = _mm_load_ps(b + 12);
+
+    ma0 = _mm_add_ps(ma0, mb0);
+    ma1 = _mm_add_ps(ma1, mb1);
+    ma2 = _mm_add_ps(ma2, mb2);
+    ma3 = _mm_add_ps(ma3, mb3);
+
+    _mm_store_ps(a, ma0);
+    _mm_store_ps(a + 4, ma1);
+    _mm_store_ps(a + 8, ma2);
+    _mm_store_ps(a + 12, ma3);
+  }
+
+  for (int i = 0; i < offset; i++) a[i] += b[i];
+}
+
+static void batch_addto_sse(float* a, const float* b[], int batch, size_t len) {
+  int offset = len % 16;
+
+  __m128 ma0, ma1, ma2, ma3;
+  __m128 mb0, mb1, mb2, mb3;
+
+  for (unsigned int k = 0; k < len / 16; k++, a += 16) {
+    ma0 = _mm_load_ps(a);
+    ma1 = _mm_load_ps(a + 4);
+    ma2 = _mm_load_ps(a + 8);
+    ma3 = _mm_load_ps(a + 12);
+
+    for (int i = 0; i < batch; i++) {
+      mb0 = _mm_load_ps(b[i]);
+      mb1 = _mm_load_ps(b[i] + 4);
+      mb2 = _mm_load_ps(b[i] + 8);
+      mb3 = _mm_load_ps(b[i] + 12);
+      ma0 = _mm_add_ps(ma0, mb0);
+      ma1 = _mm_add_ps(ma1, mb1);
+      ma2 = _mm_add_ps(ma2, mb2);
+      ma3 = _mm_add_ps(ma3, mb3);
+      b[i] += 16;
+    }
+
+    _mm_store_ps(a, ma0);
+    _mm_store_ps(a + 4, ma1);
+    _mm_store_ps(a + 8, ma2);
+    _mm_store_ps(a + 12, ma3);
+  }
+
+  for (int i = 0; i < offset; i++) {
+    for (int k = 0; k < batch; k++) a[i] += b[k][i];
+  }
+  return;
+}
+
+static void col_max_sse(float* result,
+                        const float* data,
+                        int dim,
+                        int numSamples) {
+  // first sample, direct copy
+  for (int d = 0; d < dim; ++d) {
+    result[d] = data[d];
+  }
+  int offset = dim % 16;
+  __m128 ma0, ma1, ma2, ma3;
+  __m128 mb0, mb1, mb2, mb3;
+  // first 16n dims
+  for (int k = 0; k < dim / 16; k++, result += 16, data += 16) {
+    ma0 = _mm_load_ps(result);
+    ma1 = _mm_load_ps(result + 4);
+    ma2 = _mm_load_ps(result + 8);
+    ma3 = _mm_load_ps(result + 12);
+    for (int i = 1; i < numSamples; i++) {
+      mb0 = _mm_load_ps(data + i * dim);
+      mb1 = _mm_load_ps(data + i * dim + 4);
+      mb2 = _mm_load_ps(data + i * dim + 8);
+      mb3 = _mm_load_ps(data + i * dim + 12);
+      ma0 = _mm_max_ps(ma0, mb0);
+      ma1 = _mm_max_ps(ma1, mb1);
+      ma2 = _mm_max_ps(ma2, mb2);
+      ma3 = _mm_max_ps(ma3, mb3);
+    }
+    _mm_store_ps(result, ma0);
+    _mm_store_ps(result + 4, ma1);
+    _mm_store_ps(result + 8, ma2);
+    _mm_store_ps(result + 12, ma3);
+  }
+  // last dims
+  for (int d = 0; d < offset; ++d) {
+    float sm = data[d];
+    for (int i = 1; i < numSamples; ++i) {
+      sm = std::max(sm, data[i * dim + d]);
+    }
+    result[d] = sm;
+  }
+}
+
 #endif
 
-#ifndef __AVX__
-#define SIMD_INVOKE(func, ...) func##_sse(__VA_ARGS__)
-#else
+#if defined(__AVX__)
 #define SIMD_INVOKE(func, ...) func##_avx(__VA_ARGS__)
+#elif defined(__SSE3__)
+#define SIMD_INVOKE(func, ...) func##_sse(__VA_ARGS__)
 #endif
 
 namespace paddle {
 namespace simd {
 namespace internal {
+#ifdef __SSE3__
 void addToImpl(float* a, const float* b, size_t len) {
   SIMD_INVOKE(addto, a, b, len);
 }
@@ -376,6 +380,7 @@ void batchAddToImpl(float* a, const float* b[], int batch, size_t len) {
 void colMaxImpl(float* result, const float* data, int dim, int numSamples) {
   SIMD_INVOKE(col_max, result, data, dim, numSamples);
 }
+#endif
 
 #ifdef __AVX__
 void decayL1AvxImpl(float* dst, float* src, float lambda, size_t len) {
@@ -385,8 +390,8 @@ void decayL1AvxImpl(
     float* dst, float* src, float* lr, float lambda, size_t len) {
   decayL1_avx(dst, src, lr, lambda, len);
 }
-
 #endif
+
 }  // namespace internal
 }  // namespace simd
 }  // namespace paddle

paddle/math/SIMDFunctions.h

@@ -128,17 +128,29 @@ void decayL1AvxImpl(
 
 template <>
 inline void addTo(float* a, const float* b, size_t len) {
+#ifdef __SSE3__
   internal::addToImpl(a, b, len);
+#else
+  naive::addTo(a, b, len);
+#endif
 }
 
 template <>
 inline void batchAddTo(float* a, const float* b[], int batch, size_t len) {
+#ifdef __SSE3__
   internal::batchAddToImpl(a, b, batch, len);
+#else
+  naive::batchAddTo(a, b, batch, len);
+#endif
 }
 
 template <>
 inline void colMax(float* result, const float* data, int dim, int numSamples) {
+#ifdef __SSE3__
   internal::colMaxImpl(result, data, dim, numSamples);
+#else
+  naive::colMax(result, data, dim, numSamples);
+#endif
 }
 
 template <>

paddle/math/Storage.cpp

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #include "Storage.h"
 #include "Allocator.h"
+#include "paddle/utils/StringUtil.h"
 #include "paddle/utils/Util.h"
 
 DEFINE_int32(pool_limit_size,
@@ -62,7 +63,7 @@ PoolAllocator* StorageEngine::getGpuAllocator(int deviceId) {
     }
     if (gpuAllocator_[deviceId] == nullptr) {
       std::string name =
-          "gpu" + std::to_string(deviceId) + std::string("_pool");
+          "gpu" + str::to_string(deviceId) + std::string("_pool");
       gpuAllocator_[deviceId] =
           new PoolAllocator(new GpuAllocator(), FLAGS_pool_limit_size, name);
     }

paddle/math/tests/test_matrixCompare.cpp

@@ -685,7 +685,7 @@ TEST(SMatrix, topK) {
   }
 }
 
-void testMatrixSequenceAvgForward(int batchSize, int inputDim, int mode) {
+void testMatrixSequenceAvg(int batchSize, int inputDim, int mode) {
   MatrixPtr cpuInput = std::make_shared<CpuMatrix>(batchSize, inputDim);
   MatrixPtr gpuInput = std::make_shared<GpuMatrix>(batchSize, inputDim);
   cpuInput->randomizeUniform();
@@ -706,15 +706,25 @@ void testMatrixSequenceAvgForward(int batchSize, int inputDim, int mode) {
   gpuOutput->sequenceAvgForward(*gpuInput, *gpuSequence, mode);
 
   TensorCheckErr(*cpuOutput, *gpuOutput);
+
+  MatrixPtr cpuInGrad = std::make_shared<CpuMatrix>(batchSize, inputDim);
+  MatrixPtr gpuInGrad = std::make_shared<GpuMatrix>(batchSize, inputDim);
+  cpuInGrad->randomizeUniform();
+  gpuInGrad->copyFrom(*cpuInGrad);
+
+  cpuInGrad->sequenceAvgBackward(*cpuOutput, *cpuSequence, mode);
+  gpuInGrad->sequenceAvgBackward(*gpuOutput, *gpuSequence, mode);
+
+  TensorCheckErr(*cpuInGrad, *gpuInGrad);
 }
 
-TEST(Matrix, sequenceAvgForward) {
+TEST(Matrix, sequenceAvg) {
   for (auto batchSize : {10, 128, 6000}) {
     for (auto inputDim : {32, 100, 512}) {
       for (auto mode : {0, 1, 2}) {
         VLOG(3) << " batchSize=" << batchSize << " inputDim=" << inputDim
                 << " mode=" << mode;
-        testMatrixSequenceAvgForward(batchSize, inputDim, mode);
+        testMatrixSequenceAvg(batchSize, inputDim, mode);
       }
     }
   }

paddle/parameter/Argument.cpp

@@ -559,6 +559,49 @@ void Argument::degradeSequence(const Argument& input) {
   tgtBuf[numSequences] = numSubSequences;
 }
 
+void Argument::poolSequenceWithStride(const Argument& input,
+                                      size_t stride,
+                                      IVectorPtr* stridePostions,
+                                      bool reversed) {
+  // If input.sequenceStartPositions = [0, 9, 14, 17, 30] and stride = 5,
+  // then sequenceStartPositions = [0, 2, 3, 4, 7].
+  // If reversed = false, stridePostions = [0, 5, 9, 14, 17, 22, 27, 30];
+  // else reversed = true, stridePostions = [0, 4, 9, 14, 17, 20, 25, 30]
+
+  CHECK(input.sequenceStartPositions);
+  CHECK_EQ(input.hasSubseq(), 0UL);
+  CHECK_GT(stride, 0) << "stride must larger than 0";
+  size_t numSequences = input.getNumSequences();
+  ICpuGpuVector::resizeOrCreate(
+      sequenceStartPositions, numSequences + 1, false);
+  const int* starts = input.sequenceStartPositions->getData(false);
+  int* tgtBuf = sequenceStartPositions->getMutableData(false);
+  // first index of target sequence and stride positions are both 0
+  tgtBuf[0] = 0;
+  std::vector<int> stridePos;
+  for (size_t seqId = 0; seqId < numSequences; ++seqId) {
+    size_t seqLength = starts[seqId + 1] - starts[seqId];
+    stridePos.emplace_back(starts[seqId]);
+    if (seqLength == 0) {
+      // empty sequence
+      tgtBuf[seqId + 1] = tgtBuf[seqId];
+    } else {
+      int size = ceil((float)seqLength / stride);
+      tgtBuf[seqId + 1] = tgtBuf[seqId] + size;
+      for (int i = 0; i < size - 1; ++i) {
+        int cur = reversed ? starts[seqId + 1] - (size - 1 - i) * stride
+                           : stridePos.back() + stride;
+        stridePos.emplace_back(cur);
+      }
+    }
+  }
+  stridePos.emplace_back(starts[numSequences]);
+  int size = stridePos.size();
+  CHECK_EQ(size - 1, tgtBuf[numSequences]);
+  IVector::resizeOrCreate(*stridePostions, size, false);
+  (*stridePostions)->copyFrom(stridePos.data(), size);
+}
+
 void Argument::getValueString(
     std::unordered_map<std::string, std::string>* out) const {
   if (value) {

paddle/parameter/Argument.h

@@ -291,6 +291,15 @@ struct Argument {
    */
   void degradeSequence(const Argument& input);
 
+  /*
+   After pooling with stride n (n is smaller than sequence length),
+   a long sequence will be shorten.
+   This function is invalid for sequence having sub-sequence.
+   */
+  void poolSequenceWithStride(const Argument& input,
+                              size_t stride,
+                              IVectorPtr* stridePositions,
+                              bool reversed = false);
   /**
    * @brief getValueString will return the argument's output in string. There
    * are several kinds of output. The keys of output dictionary are 'value',

paddle/parameter/tests/CMakeLists.txt

 add_simple_unittest(test_common)
+add_simple_unittest(test_argument)

paddle/parameter/tests/test_argument.cpp

+/* Copyright (c) 2016 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 <gtest/gtest.h>
+#include <paddle/parameter/Argument.h>
+
+using namespace paddle;  // NOLINT
+
+TEST(Argument, poolSequenceWithStride) {
+  Argument input, output;
+  ICpuGpuVector::resizeOrCreate(input.sequenceStartPositions, 5, false);
+  int* inStart = input.sequenceStartPositions->getMutableData(false);
+  inStart[0] = 0;
+  inStart[1] = 9;
+  inStart[2] = 14;
+  inStart[3] = 17;
+  inStart[4] = 30;
+
+  int strideResult[] = {0, 5, 9, 14, 17, 22, 27, 30};
+  int strideResultReversed[] = {0, 4, 9, 14, 17, 20, 25, 30};
+
+  for (auto reversed : {false, true}) {
+    IVectorPtr stridePositions;
+    output.poolSequenceWithStride(
+        input, 5 /* stride */, &stridePositions, reversed);
+
+    const int* outStart = output.sequenceStartPositions->getData(false);
+    CHECK_EQ(outStart[0], 0);
+    CHECK_EQ(outStart[1], 2);
+    CHECK_EQ(outStart[2], 3);
+    CHECK_EQ(outStart[3], 4);
+    CHECK_EQ(outStart[4], 7);
+
+    CHECK_EQ(stridePositions->getSize(), 8);
+    auto result = reversed ? strideResultReversed : strideResult;
+    for (int i = 0; i < 8; i++) {
+      CHECK_EQ(stridePositions->getData()[i], result[i]);
+    }
+  }
+}
+
+int main(int argc, char** argv) {
+  testing::InitGoogleTest(&argc, argv);
+  initMain(argc, argv);
+  return RUN_ALL_TESTS();
+}

paddle/pserver/ParameterServer2.cpp

@@ -29,6 +29,7 @@ limitations under the License. */
 #include "paddle/utils/Flags.h"
 #include "paddle/utils/GlobalConstants.h"
 #include "paddle/utils/Stat.h"
+#include "paddle/utils/StringUtil.h"
 
 DEFINE_int32(pserver_num_threads, 1, "number of threads for sync op exec");
 DEFINE_double(async_lagged_ratio_min,
@@ -218,7 +219,8 @@ void ParameterServer2::setConfig(const SetConfigRequest& request,
   callback(response);
 
   /// always defined, barrier slowest node function need it.
-  statSet_.reset(new StatSet("ParameterServer" + std::to_string(serverId_)));
+  statSet_.reset(new StatSet("ParameterServer" +
+                             str::to_string(static_cast<int>(serverId_))));
 }
 
 real bufferSum(const std::vector<ParameterServer2::Buffer>& buffers) {

paddle/py_paddle/dataprovider_converter.py

@@ -160,10 +160,19 @@ class SparseFloatScanner(SparseBinaryScanner):
 class IndexScanner(IScanner):
     def __init__(self, input_type, pos):
         IScanner.__init__(self, input_type, pos)
-        self.__ids__ = []
+        self.__ids__ = None
+        self.__idx__ = 0
+
+    def pre_scan(self, dat):
+        self.__idx__ += 1
+
+    def finish_pre_scan(self, argument):
+        self.__ids__ = [0] * self.__idx__
+        self.__idx__ = 0
 
     def scan(self, dat):
-        self.__ids__.append(dat)
+        self.__ids__[self.__idx__] = dat
+        self.__idx__ += 1
 
     def finish_scan(self, argument):
         ids = swig_paddle.IVector.create(self.__ids__, self.data_in_gpu)
@@ -178,6 +187,13 @@ class SequenceScanner(IScanner):
         self.__inner_scanner__ = inner_scanner
         self.__setter__ = setter
 
+    def pre_scan(self, dat):
+        for each in dat:
+            self.__inner_scanner__.pre_scan(each)
+
+    def finish_pre_scan(self, argument):
+        self.__inner_scanner__.finish_pre_scan(argument)
+
     def scan(self, dat):
         self.__seq__.append(self.__seq__[-1] + self.get_size(dat))
         for each in dat:

paddle/py_paddle/util.py

@@ -83,13 +83,17 @@ def __arguments_to_numpy__(i, arg):
     assert isinstance(arg, swig_paddle.Arguments)
     value = arg.getSlotValue(i)
     ids = arg.getSlotIds(i)
+    prob = arg.getSlotIn(i)
     if value is not None:
         assert isinstance(value, swig_paddle.Matrix)
         value = value.copyToNumpyMat()
     if ids is not None:
         assert isinstance(ids, swig_paddle.IVector)
         ids = ids.copyToNumpyArray()
-    return {"value": value, "id": ids}
+    if prob is not None:
+        assert isinstance(prob, swig_paddle.Matrix)
+        prob = prob.copyToNumpyMat()
+    return {"value": value, "id": ids, "prob": prob}
 
 
 def __monkeypatch_gradient_machine__():

paddle/scripts/docker/README.md

@@ -94,7 +94,7 @@ docker build -t paddle:dev --build-arg UBUNTU_MIRROR=mirror://mirrors.ubuntu.com
 Given the development image `paddle:dev`, the following command builds PaddlePaddle from the source tree on the development computer (host):
 
 ```bash
-docker run -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_AVX=ON" -e "WITH_TEST=OFF" -e "RUN_TEST=OFF" paddle:dev
+docker run --rm -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_AVX=ON" -e "WITH_TEST=OFF" -e "RUN_TEST=OFF" paddle:dev
 ```
 
 This command mounts the source directory on the host into `/paddle` in the container, so the default entry point of `paddle:dev`, `build.sh`, could build the source code with possible local changes.  When it writes to `/paddle/build` in the container, it writes to `$PWD/build` on the host indeed.
@@ -110,7 +110,7 @@ Users can specify the following Docker build arguments with either "ON" or "OFF"
 - `WITH_AVX`: ***Required***. Set to "OFF" prevents from generating AVX instructions. If you don't know what is AVX, you might want to set "ON".
 - `WITH_TEST`: ***Optional, default OFF***. Build unit tests binaries. Once you've built the unit tests, you can run these test manually by the following command:
   ```bash
-    docker run -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_AVX=ON" paddle:dev sh -c "cd /paddle/build; make coverall"
+    docker run --rm -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_AVX=ON" paddle:dev sh -c "cd /paddle/build; make coverall"
   ```
 - `RUN_TEST`: ***Optional, default OFF***. Run unit tests after building. You can't run unit tests without building it.
 
@@ -129,7 +129,7 @@ This production image is minimal -- it includes binary `paddle`, the shared libr
 Again the development happens on the host.  Suppose that we have a simple application program in `a.py`, we can test and run it using the production image:
 
 ```bash
-docker run -it -v $PWD:/work paddle /work/a.py
+docker run --rm -it -v $PWD:/work paddle /work/a.py
 ```
 
 But this works only if all dependencies of `a.py` are in the production image. If this is not the case, we need to build a new Docker image from the production image and with more dependencies installs.
@@ -166,3 +166,18 @@ docker tag myapp me/myapp
 docker push
 kubectl ...
 ```
+
+### Reading source code with woboq codebrowser
+For developers who are interested in the C++ source code, please use -e "WOBOQ=ON" to enable the building of C++ source code into HTML pages using [Woboq codebrowser](https://github.com/woboq/woboq_codebrowser).
+
+- The following command builds PaddlePaddle, generates HTML pages from C++ source code, and writes HTML pages into `$HOME/woboq_out` on the host:
+
+```bash
+docker run -v $PWD:/paddle -v $HOME/woboq_out:/woboq_out -e "WITH_GPU=OFF" -e "WITH_AVX=ON" -e "WITH_TEST=ON" -e "WOBOQ=ON" paddle:dev
+```
+
+- You can open the generated HTML files in your Web browser. Or, if you want to run a Nginx container to serve them for a wider audience, you can run:
+
+```
+docker run -v $HOME/woboq_out:/usr/share/nginx/html -d -p 8080:80 nginx
+```

paddle/scripts/docker/build.sh

@@ -4,7 +4,7 @@ set -e
 
 # Set BASE_IMAGE according to env variables
 if [ ${WITH_GPU} == "ON" ]; then
-  BASE_IMAGE="nvidia/cuda:7.5-cudnn5-runtime-ubuntu14.04"
+  BASE_IMAGE="nvidia/cuda:8.0-cudnn5-runtime-ubuntu14.04"
   # additional packages to install when building gpu images
   GPU_DOCKER_PKG="python-pip python-dev"
 else
@@ -12,11 +12,10 @@ else
 fi
 
 DOCKERFILE_GPU_ENV=""
+DOCKERFILE_CUDNN_DSO=""
 if [[ ${WITH_GPU:-OFF} == 'ON' ]]; then
     DOCKERFILE_GPU_ENV="ENV LD_LIBRARY_PATH /usr/lib/x86_64-linux-gnu:${LD_LIBRARY_PATH}"
-
-    # for cmake to find cudnn
-    ln -s /usr/lib/x86_64-linux-gnu/libcudnn.so /usr/lib/libcudnn.so
+    DOCKERFILE_CUDNN_DSO="RUN ln -s /usr/lib/x86_64-linux-gnu/libcudnn.so.5 /usr/lib/x86_64-linux-gnu/libcudnn.so"
 fi
 
 mkdir -p /paddle/build
@@ -47,7 +46,7 @@ make install
 # install them in docker
 cpack -D CPACK_GENERATOR='DEB' -D CPACK_DEBIAN_PACKAGE_DEPENDS="" ..
 
-if [[ ${BUILD_WOBOQ:-OFF} == 'ON' ]]; then
+if [[ ${WOBOQ:-OFF} == 'ON' ]]; then
     apt-get install -y clang-3.8 llvm-3.8 libclang-3.8-dev
     # Install woboq_codebrowser.
     git clone https://github.com/woboq/woboq_codebrowser /woboq
@@ -57,7 +56,7 @@ if [[ ${BUILD_WOBOQ:-OFF} == 'ON' ]]; then
           .
     make
 
-    export WOBOQ_OUT=/usr/share/nginx/html/paddle
+    export WOBOQ_OUT=/woboq_out/paddle
     export BUILD_DIR=/paddle/build
     mkdir -p $WOBOQ_OUT
     cp -rv /woboq/data $WOBOQ_OUT/../data
@@ -95,7 +94,10 @@ RUN ${MIRROR_UPDATE}
 # Use different deb file when building different type of images
 ADD build/*.deb /usr/local/opt/paddle/deb/
 # run paddle version to install python packages first
-RUN dpkg -i /usr/local/opt/paddle/deb/*.deb && rm -f /usr/local/opt/paddle/deb/*.deb && paddle version
+RUN dpkg -i /usr/local/opt/paddle/deb/*.deb && \
+    rm -f /usr/local/opt/paddle/deb/*.deb && \
+    paddle version
+${DOCKERFILE_CUDNN_DSO} 
 ${DOCKERFILE_GPU_ENV}
 # default command shows the paddle version and exit
 CMD ["paddle", "version"]

paddle/scripts/submit_local.sh.in

@@ -21,9 +21,7 @@ function version(){
         echo "    with_double: @WITH_DOUBLE@"
         echo "    with_python: @WITH_PYTHON@"
         echo "    with_rdma: @WITH_RDMA@"
-        echo "    with_metric_learning: @WITH_METRIC@"
         echo "    with_timer: @WITH_TIMER@"
-        echo "    with_predict_sdk: @WITH_PREDICT_SDK@"
 }
 
 function ver2num() {

paddle/utils/Any.h

+/* Copyright (c) 2016 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. */
+
+#pragma once
+#if __cplusplus > 201402L
+#include <any>
+
+namespace paddle {
+// using std::any for C++ 17
+using std::any;
+using std::any_cast;
+using std::bad_any_cast;
+}  // namespace paddle
+
+#else
+#include <any.hpp>
+
+namespace paddle {
+// use linb::any for C++ 11
+using linb::any;
+using linb::any_cast;
+using linb::bad_any_cast;
+}  // namespace paddle
+#endif

paddle/utils/CpuId.cpp

@@ -19,7 +19,7 @@ limitations under the License. */
 /// for MSVC
 #define CPUID(info, x) __cpuidex(info, x, 0)
 
-#else
+#elif !defined(__ANDROID__)
 
 #include <cpuid.h>
 
@@ -31,6 +31,7 @@ limitations under the License. */
 namespace paddle {
 
 SIMDFlags::SIMDFlags() {
+#if !defined(__ANDROID__)
   unsigned int cpuInfo[4];
   // CPUID: https://en.wikipedia.org/wiki/CPUID
   // clang-format off
@@ -51,6 +52,9 @@ SIMDFlags::SIMDFlags() {
   CPUID(cpuInfo, 0x80000001);
   simd_flags_ |= cpuInfo[2] & (1 << 16) ? SIMD_FMA4  : SIMD_NONE;
   // clang-fotmat on
+#else
+  simd_flags_ = SIMD_NEON;
+#endif
 }
 
 SIMDFlags const* SIMDFlags::instance() {

paddle/utils/CpuId.h

@@ -30,6 +30,7 @@ enum simd_t {
   SIMD_AVX    = 1 << 8,     ///< AVX
   SIMD_AVX2   = 1 << 9,     ///< AVX 2
   SIMD_AVX512 = 1 << 10,    ///< AVX 512
+  SIMD_NEON   = 1 << 11,    ///  NEON
 };
 // clang-format on
 
@@ -96,6 +97,7 @@ private:
 #define HAS_AVX     HAS_SIMD(SIMD_AVX)
 #define HAS_AVX2    HAS_SIMD(SIMD_AVX2)
 #define HAS_AVX512  HAS_SIMD(SIMD_AVX512)
+#define HAS_NEON    HAS_SIMD(SIMD_NEON)
 // clang-format on
 
 /**

paddle/utils/Logging.cpp

@@ -18,6 +18,7 @@ limitations under the License. */
  */
 
 #include "Logging.h"
+#include <cstdlib>
 
 namespace paddle {
 

paddle/utils/StringUtil.h

@@ -54,6 +54,25 @@ inline T toWithStatus(const std::string& s, bool* ok = nullptr) {
   return v;
 }
 
+/**
+ * Cast type T to string with status.
+ *
+ * @param [in] v input value of type T.
+ * @param [out] ok status, return true if there is no error in casting. Set
+ *              nullptr if user don't care error at all.
+ * @return result of casting. If error occurred, a empty string will be
+ *              returned.
+ */
+template <class T>
+inline std::string toWithStatus(const T v, bool* ok = nullptr) {
+  std::ostringstream sout;
+  sout << v;
+  if (ok) {
+    *ok = !sout.fail();
+  }
+  return sout.str();
+}
+
 /// Convert string to type T. It makes sure all the characters in s are used.
 /// Otherwise it will abort.
 ///
@@ -67,6 +86,18 @@ inline T to(const std::string& s) {
   return v;
 }
 
+/// Convert type T to string.
+///
+/// @tparam T type of input value
+/// @param v input value of type T
+template <class T>
+std::string to_string(T v) {
+  bool ok;
+  std::string s = toWithStatus<T>(v, &ok);
+  CHECK(ok) << "Cannot convert v(" << v << ") to type std::string";
+  return s;
+}
+
 }  // namespace str
 
 #undef DEFINE_STRING_CONVERSION

paddle/utils/Util.cpp

@@ -15,11 +15,16 @@ limitations under the License. */
 #include "Util.h"
 
 #include <dirent.h>
-#include <pmmintrin.h>
 #include <signal.h>
 #include <sys/stat.h>
 #include <sys/types.h>
+
+#ifdef __SSE__
 #include <xmmintrin.h>
+#endif
+#ifdef __SSE3__
+#include <pmmintrin.h>
+#endif
 
 #include <fstream>
 #include <mutex>
@@ -163,8 +168,12 @@ void initMain(int argc, char** argv) {
 
   installProfilerSwitch();
 
+#ifdef __SSE__
   _MM_SET_FLUSH_ZERO_MODE(_MM_FLUSH_ZERO_ON);
+#endif
+#ifdef __SSE3__
   _MM_SET_DENORMALS_ZERO_MODE(_MM_DENORMALS_ZERO_ON);
+#endif
 
   if (FLAGS_seed == 0) {
     unsigned int t = time(NULL);

paddle/utils/arch/linux/Locks.cpp

@@ -15,6 +15,7 @@ limitations under the License. */
 #include "paddle/utils/Locks.h"
 #include <semaphore.h>
 #include <unistd.h>
+#include "paddle/utils/Logging.h"
 
 namespace paddle {
 class SemaphorePrivate {
@@ -26,7 +27,10 @@ Semaphore::Semaphore(int initValue) : m(new SemaphorePrivate()) {
   sem_init(&m->sem, 0, initValue);
 }
 
-Semaphore::~Semaphore() { sem_destroy(&m->sem); }
+Semaphore::~Semaphore() {
+  sem_destroy(&m->sem);
+  delete m;
+}
 
 bool Semaphore::timeWait(struct timespec* ts) {
   return (0 == sem_timedwait(&m->sem, ts));
@@ -36,36 +40,101 @@ void Semaphore::wait() { sem_wait(&m->sem); }
 
 void Semaphore::post() { sem_post(&m->sem); }
 
+#ifdef PADDLE_USE_PTHREAD_SPINLOCK
+
 class SpinLockPrivate {
 public:
   inline SpinLockPrivate() { pthread_spin_init(&lock_, 0); }
   inline ~SpinLockPrivate() { pthread_spin_destroy(&lock_); }
+
+  inline void lock() { pthread_spin_lock(&lock_); }
+  inline void unlock() { pthread_spin_unlock(&lock_); }
+
   pthread_spinlock_t lock_;
   char padding_[64 - sizeof(pthread_spinlock_t)];
 };
 
-SpinLock::SpinLock() : m(new SpinLockPrivate()) {}
+#else
 
-SpinLock::~SpinLock() { delete m; }
+#include <atomic>
+class SpinLockPrivate {
+public:
+  inline void lock() {
+    while (lock_.test_and_set(std::memory_order_acquire)) {
+    }
+  }
+  inline void unlock() { lock_.clear(std::memory_order_release); }
+
+  std::atomic_flag lock_ = ATOMIC_FLAG_INIT;
+  char padding_[64 - sizeof(lock_)];  // Padding to cache line size
+};
 
-void SpinLock::lock() { pthread_spin_lock(&m->lock_); }
+#endif
 
-void SpinLock::unlock() { pthread_spin_unlock(&m->lock_); }
+SpinLock::SpinLock() : m(new SpinLockPrivate()) {}
+SpinLock::~SpinLock() { delete m; }
+void SpinLock::lock() { m->lock(); }
+void SpinLock::unlock() { m->unlock(); }
+
+#ifdef PADDLE_USE_PTHREAD_BARRIER
 
 class ThreadBarrierPrivate {
 public:
   pthread_barrier_t barrier_;
+
+  inline explicit ThreadBarrierPrivate(int count) {
+    pthread_barrier_init(&barrier_, nullptr, count);
+  }
+
+  inline ~ThreadBarrierPrivate() { pthread_barrier_destroy(&barrier_); }
+
+  inline void wait() { pthread_barrier_wait(&barrier_); }
 };
 
-ThreadBarrier::ThreadBarrier(int count) : m(new ThreadBarrierPrivate()) {
-  pthread_barrier_init(&m->barrier_, nullptr, count);
-}
+#else
 
-ThreadBarrier::~ThreadBarrier() {
-  pthread_barrier_destroy(&m->barrier_);
-  delete m;
-}
+class ThreadBarrierPrivate {
+public:
+  pthread_mutex_t mutex_;
+  pthread_cond_t cond_;
+  int count_;
+  int tripCount_;
+
+  inline explicit ThreadBarrierPrivate(int cnt) : count_(0), tripCount_(cnt) {
+    CHECK_NE(cnt, 0);
+    CHECK_GE(pthread_mutex_init(&mutex_, 0), 0);
+    CHECK_GE(pthread_cond_init(&cond_, 0), 0);
+  }
+
+  inline ~ThreadBarrierPrivate() {
+    pthread_cond_destroy(&cond_);
+    pthread_mutex_destroy(&mutex_);
+  }
+
+  /**
+   * @brief wait
+   * @return true if the last wait
+   */
+  inline bool wait() {
+    pthread_mutex_lock(&mutex_);
+    ++count_;
+    if (count_ >= tripCount_) {
+      count_ = 0;
+      pthread_cond_broadcast(&cond_);
+      pthread_mutex_unlock(&mutex_);
+      return true;
+    } else {
+      pthread_cond_wait(&cond_, &mutex_);
+      pthread_mutex_unlock(&mutex_);
+      return false;
+    }
+  }
+};
+
+#endif
 
-void ThreadBarrier::wait() { pthread_barrier_wait(&m->barrier_); }
+ThreadBarrier::ThreadBarrier(int count) : m(new ThreadBarrierPrivate(count)) {}
+ThreadBarrier::~ThreadBarrier() { delete m; }
+void ThreadBarrier::wait() { m->wait(); }
 
 }  // namespace paddle

paddle/utils/tests/test_CustomStackTrace.cpp

@@ -19,6 +19,7 @@ limitations under the License. */
 
 #include "paddle/utils/CustomStackTrace.h"
 #include "paddle/utils/Locks.h"
+#include "paddle/utils/StringUtil.h"
 #include "paddle/utils/Util.h"
 
 DEFINE_int32(test_thread_num, 10, "testing thread number");
@@ -69,11 +70,11 @@ TEST(CustomStackTrace, normalTrain) {
     while (countDown-- > 0) {
       start.wait();
       for (size_t i = 0; i < layerSize; ++i) {
-        tracer.push("layer_" + std::to_string(i));
+        tracer.push("layer_" + paddle::str::to_string(i));
       }
       tracer.pop("");
       for (size_t i = 0; i < layerSize; ++i) {
-        tracer.pop("layer_" + std::to_string(layerSize - 1 - i));
+        tracer.pop("layer_" + paddle::str::to_string(layerSize - 1 - i));
       }
       finish.wait();
     }
@@ -89,7 +90,7 @@ TEST(CustomStackTrace, normalTest) {
     while (countDown-- > 0) {
       start.wait();
       for (size_t i = 0; i < layerSize; ++i) {
-        tracer.push("layer_" + std::to_string(i));
+        tracer.push("layer_" + paddle::str::to_string(i));
       }
       tracer.clear();  // in forward test, tracer will clear after forward.
       finish.wait();

paddle/utils/tests/test_CustomStackTracePrint.cpp

@@ -13,13 +13,14 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/utils/CustomStackTrace.h"
+#include "paddle/utils/StringUtil.h"
 #include "paddle/utils/Util.h"
 
 int main(int argc, char** argv) {
   paddle::initMain(argc, argv);
 
   for (size_t i = 0; i < 1000; ++i) {
-    paddle::gLayerStackTrace.push("layer_" + std::to_string(i));
+    paddle::gLayerStackTrace.push("layer_" + paddle::str::to_string(i));
     if (i == 998) {
       throw "Unhandle exception";
     }

paddle/utils/tests/test_SIMDFlags.cpp

@@ -18,7 +18,8 @@ limitations under the License. */
 using namespace paddle;  // NOLINT
 
 TEST(SIMDFlags, gccTest) {
-#if (defined(__GNUC__) || defined(__GNUG__)) && !(defined(__clang__))
+#if (defined(__GNUC__) || defined(__GNUG__)) && !(defined(__clang__)) && \
+    !defined(__arm__)
   // clang-format off
   CHECK(!__builtin_cpu_supports("sse")    != HAS_SSE);
   CHECK(!__builtin_cpu_supports("sse2")   != HAS_SSE2);
@@ -43,4 +44,5 @@ TEST(SIMDFlags, normalPrint) {
   LOG(INFO) << "Has AVX:     " << std::boolalpha << HAS_AVX;
   LOG(INFO) << "Has AVX2:    " << std::boolalpha << HAS_AVX2;
   LOG(INFO) << "Has AVX512:  " << std::boolalpha << HAS_AVX512;
+  LOG(INFO) << "Has NEON:    " << std::boolalpha << HAS_NEON;
 }

proto/ModelConfig.proto

@@ -441,6 +441,11 @@ message LayerConfig {
 
   // blank label used in ctc loss
   optional uint32 blank = 52 [default = 0];
+
+  // stride parameter for seqlastins layer, AverageLayer, MaxLayer, which 
+  // controls the scope of pooling operation. can be set > 0.
+  // leave empty or set to -1 to disable this stride pooling.
+  optional int32 seq_pool_stride = 53 [default = -1];
 }
 
 message EvaluatorConfig {

python/CMakeLists.txt

@@ -24,8 +24,9 @@ add_custom_target(paddle_python ALL DEPENDS
     ${OUTPUT_DIR}/.timestamp)
 
 add_subdirectory(paddle/trainer_config_helpers/tests)
-add_subdirectory(paddle/v2/reader/tests)
 add_subdirectory(paddle/v2/tests)
+add_subdirectory(paddle/v2/reader/tests)
+add_subdirectory(paddle/v2/plot/tests)
 
 install(DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/dist/
     DESTINATION opt/paddle/share/wheels

python/paddle/trainer/config_parser.py

@@ -2485,6 +2485,7 @@ class SequenceLastInstanceLayer(LayerBase):
                  active_type='linear',
                  trans_type='non-seq',
                  bias=False,
+                 stride=-1,
                  **xargs):
         super(SequenceLastInstanceLayer, self).__init__(
             name,
@@ -2495,10 +2496,11 @@ class SequenceLastInstanceLayer(LayerBase):
             **xargs)
         config_assert(
             len(inputs) == 1, 'SequenceLastInstanceLayer must have 1 input')
+        if trans_type == 'seq':
+            config_assert(stride == -1, 'subseq does not support stride window')
         self.config.trans_type = trans_type
-        for input_index in xrange(len(self.inputs)):
-            input_layer = self.get_input_layer(input_index)
-            self.set_layer_size(input_layer.size)
+        self.config.seq_pool_stride = stride
+        self.set_layer_size(self.get_input_layer(0).size)
         self.create_bias_parameter(bias, self.config.size)
 
 
@@ -2510,10 +2512,16 @@ class SequenceFirstInstanceLayer(SequenceLastInstanceLayer):
                  active_type='linear',
                  trans_type='non-seq',
                  bias=False,
+                 stride=-1,
                  **xargs):
         super(SequenceFirstInstanceLayer, self).__init__(
-            name, inputs=inputs, active_type=active_type, bias=bias, **xargs)
-        self.config.trans_type = trans_type
+            name,
+            inputs=inputs,
+            active_type=active_type,
+            trans_type=trans_type,
+            bias=bias,
+            stride=stride,
+            **xargs)
         self.config.select_first = True
 
 

python/paddle/trainer_config_helpers/layers.py

@@ -18,7 +18,7 @@ import inspect
 
 from paddle.trainer.config_parser import *
 from .activations import LinearActivation, SigmoidActivation, TanhActivation, \
-    ReluActivation, IdentityActivation, SoftmaxActivation
+    ReluActivation, IdentityActivation, SoftmaxActivation, BaseActivation
 from .evaluators import *
 from .poolings import MaxPooling, AvgPooling, BasePoolingType
 from .attrs import *
@@ -1342,10 +1342,16 @@ def grumemory(input,
 def last_seq(input,
              name=None,
              agg_level=AggregateLevel.EACH_TIMESTEP,
+             stride=-1,
              layer_attr=None):
     """
     Get Last Timestamp Activation of a sequence.
 
+    If stride > 0, this layer slides a window whose size is determined by stride, 
+    and return the last value of the window as the output. Thus, a long sequence 
+    will be shorten. Note that for sequence with sub-sequence, the default value 
+    of stride is -1.
+
     The simple usage is:
 
     .. code-block:: python
@@ -1357,6 +1363,8 @@ def last_seq(input,
     :type name: basestring
     :param input: Input layer name.
     :type input: LayerOutput
+    :param stride: window size.  
+    :type stride: Int
     :param layer_attr: extra layer attributes.
     :type layer_attr: ExtraLayerAttribute.
     :return: LayerOutput object.
@@ -1368,11 +1376,15 @@ def last_seq(input,
                        " series information at all. Maybe you want to use"
                        " first_seq instead.")
 
+    if agg_level == AggregateLevel.EACH_SEQUENCE:
+        assert stride == -1
+
     Layer(
         name=name,
         type=LayerType.SEQUENCE_LAST_INSTANCE,
         inputs=[input.name],
         trans_type=agg_level,
+        stride=stride,
         **ExtraLayerAttribute.to_kwargs(layer_attr))
     return LayerOutput(
         name,
@@ -1386,10 +1398,16 @@ def last_seq(input,
 def first_seq(input,
               name=None,
               agg_level=AggregateLevel.EACH_TIMESTEP,
+              stride=-1,
               layer_attr=None):
     """
     Get First Timestamp Activation of a sequence.
 
+    If stride > 0, this layer slides a window whose size is determined by stride, 
+    and return the first value of the window as the output. Thus, a long sequence 
+    will be shorten. Note that for sequence with sub-sequence, the default value 
+    of stride is -1.
+
     The simple usage is:
 
     .. code-block:: python
@@ -1401,6 +1419,8 @@ def first_seq(input,
     :type name: basestring
     :param input: Input layer name.
     :type input: LayerOutput
+    :param stride: window size.  
+    :type stride: Int
     :param layer_attr: extra layer attributes.
     :type layer_attr: ExtraLayerAttribute.
     :return: LayerOutput object.
@@ -1413,11 +1433,15 @@ def first_seq(input,
                        ' time series information at all. Maybe you want to use'
                        ' last_seq instead.')
 
+    if agg_level == AggregateLevel.EACH_SEQUENCE:
+        assert stride == -1
+
     Layer(
         name=name,
         type=LayerType.SEQUENCE_FIRST_INSTANCE,
         inputs=[input.name],
         trans_type=agg_level,
+        stride=stride,
         **ExtraLayerAttribute.to_kwargs(layer_attr))
     return LayerOutput(
         name,
@@ -1916,7 +1940,7 @@ def cos_sim(a, b, scale=1, size=1, name=None, layer_attr=None):
 @layer_support()
 def hsigmoid(input,
              label,
-             num_classes,
+             num_classes=None,
              name=None,
              bias_attr=None,
              param_attr=None,
@@ -1932,8 +1956,7 @@ def hsigmoid(input,
     ..  code-block:: python
 
         cost = hsigmoid(input=[layer1, layer2],
-                        label=data_layer,
-                        num_classes=3)
+                        label=data_layer)
 
     :param input: Input layers. It could be a LayerOutput or list/tuple of
                  LayerOutput.
@@ -1941,12 +1964,14 @@ def hsigmoid(input,
     :param label: Label layer.
     :type label: LayerOutput
     :param num_classes: number of classes.
-    :type num_classes: int
+    :type num_classes: int|None
     :param name: layer name
     :type name: basestring
     :param bias_attr: Bias attribute. None means default bias.
                       False means no bias.
     :type bias_attr: ParameterAttribute|False
+    :param param_attr: Parameter Attribute. None means default parameter.
+    :type param_attr: ParameterAttribute|None
     :param layer_attr: Extra Layer Attribute.
     :type layer_attr: ExtraLayerAttribute
     :return: LayerOutput object.
@@ -1966,6 +1991,11 @@ def hsigmoid(input,
     assert isinstance(label, LayerOutput)
     assert label.layer_type == LayerType.DATA
 
+    if num_classes is None:
+        num_classes = label.size
+    if num_classes is None or num_classes <= 2:
+        raise ValueError("hsigmoid label size must larger than 2.")
+
     ipts_for_layer = []
     parents = []
     for each_input, each_param_attr in zip(input, param_attr):
@@ -2253,8 +2283,9 @@ def img_pool_layer(input,
         pool_type.name = 'avg'
 
     type_name = pool_type.name + '-projection' \
-      if (isinstance(pool_type, AvgPooling) or isinstance(pool_type, MaxPooling)) \
-      else pool_type.name
+        if (
+    isinstance(pool_type, AvgPooling) or isinstance(pool_type, MaxPooling)) \
+        else pool_type.name
 
     pool_size_y = pool_size if pool_size_y is None else pool_size_y
     stride_y = stride if stride_y is None else stride_y
@@ -3294,8 +3325,8 @@ def recurrent_group(step,
 
     assert (targetInlink == None or targetInlink_in_inlinks())
     targetInlinkName = None if targetInlink == None \
-                            else targetInlink.name if isinstance(targetInlink, LayerOutput) \
-                                                   else targetInlink.input.name
+        else targetInlink.name if isinstance(targetInlink, LayerOutput) \
+        else targetInlink.input.name
 
     contains_sub_seq = [False]
 
@@ -4807,12 +4838,14 @@ def crf_decoding_layer(input,
     return LayerOutput(name, LayerType.CRF_DECODING_LAYER, parents, size=1)
 
 
+@wrap_act_default(act=SigmoidActivation())
 @wrap_bias_attr_default(has_bias=True)
 @wrap_name_default()
 @layer_support()
 def nce_layer(input,
               label,
               num_classes,
+              act=None,
               weight=None,
               num_neg_samples=10,
               neg_distribution=None,
@@ -4841,6 +4874,8 @@ def nce_layer(input,
     :type weight: LayerOutput
     :param num_classes: number of classes.
     :type num_classes: int
+    :param act: Activation, default is Sigmoid.
+    :type act: BaseActivation
     :param num_neg_samples: number of negative samples. Default is 10.
     :type num_neg_samples: int
     :param neg_distribution: The distribution for generating the random negative labels.
@@ -4862,7 +4897,9 @@ def nce_layer(input,
     if neg_distribution is not None:
         assert isinstance(neg_distribution, collections.Sequence)
         assert len(neg_distribution) == num_classes
-        assert sum(neg_distribution) == 1
+        assert abs(sum(neg_distribution) - 1.0) < 1e-5
+    if not isinstance(act, BaseActivation):
+        raise TypeError()
 
     ipts_for_layer = []
     parents = []
@@ -4884,12 +4921,17 @@ def nce_layer(input,
         type=LayerType.NCE_LAYER,
         num_classes=num_classes,
         neg_sampling_dist=neg_distribution,
+        active_type=act.name,
         num_neg_samples=num_neg_samples,
         inputs=ipts_for_layer,
         bias=ParamAttr.to_bias(bias_attr),
         **ExtraLayerAttribute.to_kwargs(layer_attr))
     return LayerOutput(
-        name, LayerType.NCE_LAYER, parents=parents, size=l.config.size)
+        name,
+        LayerType.NCE_LAYER,
+        parents=parents,
+        size=l.config.size,
+        activation=act)
 
 
 """

python/paddle/trainer_config_helpers/tests/configs/last_first_seq.py

@@ -14,4 +14,7 @@ for op in seq_op:
     for al in agg_level:
         opts.append(op(input=din, agg_level=al))
 
+for op in seq_op:
+    opts.append(op(input=din, agg_level=AggregateLevel.EACH_TIMESTEP, stride=5))
+
 outputs(opts)

python/paddle/trainer_config_helpers/tests/configs/protostr/last_first_seq.protostr

@@ -15,6 +15,7 @@ layers {
   }
   select_first: true
   trans_type: "seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__first_seq_1__"
@@ -26,6 +27,7 @@ layers {
   }
   select_first: true
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__last_seq_0__"
@@ -36,6 +38,7 @@ layers {
     input_layer_name: "data"
   }
   trans_type: "seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__last_seq_1__"
@@ -46,12 +49,38 @@ layers {
     input_layer_name: "data"
   }
   trans_type: "non-seq"
+  seq_pool_stride: -1
+}
+layers {
+  name: "__first_seq_2__"
+  type: "seqlastins"
+  size: 30
+  active_type: "linear"
+  inputs {
+    input_layer_name: "data"
+  }
+  select_first: true
+  trans_type: "non-seq"
+  seq_pool_stride: 5
+}
+layers {
+  name: "__last_seq_2__"
+  type: "seqlastins"
+  size: 30
+  active_type: "linear"
+  inputs {
+    input_layer_name: "data"
+  }
+  trans_type: "non-seq"
+  seq_pool_stride: 5
 }
 input_layer_names: "data"
 output_layer_names: "__first_seq_0__"
 output_layer_names: "__first_seq_1__"
 output_layer_names: "__last_seq_0__"
 output_layer_names: "__last_seq_1__"
+output_layer_names: "__first_seq_2__"
+output_layer_names: "__last_seq_2__"
 sub_models {
   name: "root"
   layer_names: "data"
@@ -59,11 +88,15 @@ sub_models {
   layer_names: "__first_seq_1__"
   layer_names: "__last_seq_0__"
   layer_names: "__last_seq_1__"
+  layer_names: "__first_seq_2__"
+  layer_names: "__last_seq_2__"
   input_layer_names: "data"
   output_layer_names: "__first_seq_0__"
   output_layer_names: "__first_seq_1__"
   output_layer_names: "__last_seq_0__"
   output_layer_names: "__last_seq_1__"
+  output_layer_names: "__first_seq_2__"
+  output_layer_names: "__last_seq_2__"
   is_recurrent_layer_group: false
 }
 

python/paddle/trainer_config_helpers/tests/configs/protostr/shared_gru.protostr

@@ -128,6 +128,7 @@ layers {
     input_layer_name: "__simple_gru_0__"
   }
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__last_seq_1__"
@@ -138,6 +139,7 @@ layers {
     input_layer_name: "__simple_gru_1__"
   }
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__fc_layer_0__"

python/paddle/trainer_config_helpers/tests/configs/protostr/shared_lstm.protostr

@@ -210,6 +210,7 @@ layers {
     input_layer_name: "__lstm_group_0__"
   }
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__last_seq_1__"
@@ -220,6 +221,7 @@ layers {
     input_layer_name: "__lstm_group_1__"
   }
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__fc_layer_0__"

python/paddle/trainer_config_helpers/tests/configs/protostr/simple_rnn_layers.protostr

@@ -143,6 +143,7 @@ layers {
     input_layer_name: "__recurrent_layer_0__"
   }
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__first_seq_0__"
@@ -154,6 +155,7 @@ layers {
   }
   select_first: true
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__last_seq_1__"
@@ -164,6 +166,7 @@ layers {
     input_layer_name: "__lstmemory_0__"
   }
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__first_seq_1__"
@@ -175,6 +178,7 @@ layers {
   }
   select_first: true
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__last_seq_2__"
@@ -185,6 +189,7 @@ layers {
     input_layer_name: "__gru_0__"
   }
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__first_seq_2__"
@@ -196,6 +201,7 @@ layers {
   }
   select_first: true
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 parameters {
   name: "___fc_layer_0__.w0"