Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into attention

.gitignore

@@ -28,4 +28,3 @@ cmake_install.cmake
 paddle/.timestamp
 python/paddlepaddle.egg-info/
 paddle/pybind/pybind.h
-python/paddle/v2/framework/tests/tmp/*

CMakeLists.txt

@@ -36,8 +36,7 @@ include(simd)
 ################################ Configurations #######################################
 option(WITH_GPU         "Compile PaddlePaddle with NVIDIA GPU"          ${CUDA_FOUND})
 option(WITH_AVX         "Compile PaddlePaddle with AVX intrinsics"      ${AVX_FOUND})
-option(WITH_MKLDNN      "Compile PaddlePaddle with mkl-dnn support."    ${AVX_FOUND})
-option(WITH_MKLML       "Compile PaddlePaddle with mklml package."      ${AVX_FOUND})
+option(WITH_MKL         "Compile PaddlePaddle with MKL support."        ${AVX_FOUND})
 option(WITH_DSO         "Compile PaddlePaddle with dynamic linked CUDA" ON)
 option(WITH_TESTING     "Compile PaddlePaddle with unit testing"        ON)
 option(WITH_SWIG_PY     "Compile PaddlePaddle with inference api"       ON)
@@ -82,10 +81,8 @@ if(ANDROID OR IOS)
         "Disable PYTHON when cross-compiling for Android and iOS" FORCE)
     set(WITH_RDMA OFF CACHE STRING
         "Disable RDMA when cross-compiling for Android and iOS" FORCE)
-    set(WITH_MKLDNN OFF CACHE STRING
-        "Disable MKLDNN when cross-compiling for Android and iOS" FORCE)
-    set(WITH_MKLML OFF CACHE STRING
-        "Disable MKLML package when cross-compiling for Android and iOS" FORCE)
+    set(WITH_MKL OFF CACHE STRING
+        "Disable MKL when cross-compiling for Android and iOS" FORCE)
 
     # Compile PaddlePaddle mobile inference library
     if (NOT WITH_C_API)
@@ -111,6 +108,14 @@ else()
     set(THIRD_PARTY_BUILD_TYPE Release)
 endif()
 
+set(WITH_MKLML ${WITH_MKL})
+if (WITH_MKL AND AVX2_FOUND)
+    set(WITH_MKLDNN ON)
+else()
+    message(STATUS "Do not have AVX2 intrinsics and disabled MKL-DNN")
+    set(WITH_MKLDNN OFF)
+endif()
+
 ########################################################################################
 
 include(external/mklml)     # download mklml package
@@ -158,14 +163,15 @@ set(EXTERNAL_LIBS
 )
 
 if(WITH_GPU)
-    list(APPEND EXTERNAL_LIBS ${CUDA_LIBRARIES} ${CUDA_rt_LIBRARY})
-    if(NOT WITH_DSO)
-        list(APPEND EXTERNAL_LIBS ${CUDNN_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} ${CUDA_curand_LIBRARY} ${NCCL_LIBRARY})
-    endif(NOT WITH_DSO)
+  include(cuda)
 endif(WITH_GPU)
 
+if(WITH_MKLML)
+    list(APPEND EXTERNAL_LIBS ${MKLML_IOMP_LIB})
+endif()
+
 if(WITH_MKLDNN)
-    list(APPEND EXTERNAL_LIBS ${MKLDNN_LIB} ${MKLDNN_IOMP_LIB})
+    list(APPEND EXTERNAL_LIBS ${MKLDNN_LIB})
 endif()
 
 if(USE_NNPACK)

benchmark/paddle/image/run_mkldnn.sh

 set -e
 
 function train() {
-  unset OMP_NUM_THREADS MKL_NUM_THREADS
-  export OMP_DYNAMIC="FALSE"
-  export KMP_AFFINITY="granularity=fine,compact,0,0"
+  unset OMP_NUM_THREADS MKL_NUM_THREADS OMP_DYNAMIC KMP_AFFINITY
   topology=$1
   layer_num=$2
   bs=$3
@@ -14,8 +12,6 @@ function train() {
   elif [ $4 == "False" ]; then
     thread=`nproc`
     # each trainer_count use only 1 core to avoid conflict
-    export OMP_NUM_THREADS=1
-    export MKL_NUM_THREADS=1
     log="logs/${topology}-${layer_num}-${thread}mklml-${bs}.log"
   else
     echo "Wrong input $3, use True or False."

cmake/configure.cmake

@@ -76,27 +76,14 @@ else()
     include_directories(${CUDA_TOOLKIT_INCLUDE})
 endif(NOT WITH_GPU)
 
-if(WITH_MKLDNN)
-    add_definitions(-DPADDLE_USE_MKLDNN)
-    if (WITH_MKLML AND MKLDNN_IOMP_DIR)
-        message(STATUS "Enable Intel OpenMP at ${MKLDNN_IOMP_DIR}")
-        set(OPENMP_FLAGS "-fopenmp")
-        set(CMAKE_C_CREATE_SHARED_LIBRARY_FORBIDDEN_FLAGS ${OPENMP_FLAGS})
-        set(CMAKE_CXX_CREATE_SHARED_LIBRARY_FORBIDDEN_FLAGS ${OPENMP_FLAGS})
-        set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${OPENMP_FLAGS}")
-        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OPENMP_FLAGS}")
-    else()
-        find_package(OpenMP)
-        if(OPENMP_FOUND)
-            set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${OpenMP_C_FLAGS}")
-            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OpenMP_CXX_FLAGS}")
-        else()
-            message(WARNING "Can not find OpenMP."
-                 "Some performance features in MKLDNN may not be available")
-        endif()
-    endif()
-
-endif(WITH_MKLDNN)
+if (WITH_MKLML AND MKLML_IOMP_LIB)
+    message(STATUS "Enable Intel OpenMP with ${MKLML_IOMP_LIB}")
+    set(OPENMP_FLAGS "-fopenmp")
+    set(CMAKE_C_CREATE_SHARED_LIBRARY_FORBIDDEN_FLAGS ${OPENMP_FLAGS})
+    set(CMAKE_CXX_CREATE_SHARED_LIBRARY_FORBIDDEN_FLAGS ${OPENMP_FLAGS})
+    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${OPENMP_FLAGS}")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OPENMP_FLAGS}")
+endif()
 
 set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${SIMD_FLAG}")
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${SIMD_FLAG}")

cmake/cross_compiling/ios.cmake

@@ -76,11 +76,9 @@ set(IOS_PLATFORM ${IOS_PLATFORM} CACHE STRING "Type of iOS Platform")
 # Set the architecture for iOS
 if(NOT DEFINED IOS_ARCH)
   if(IOS_PLATFORM STREQUAL "OS")
-    # FIXME(liuyiqun): support "armv7;armv7s;arm64" future
-    set(IOS_ARCH "arm64")
+    set(IOS_ARCH "armv7;armv7s;arm64")
   elseif(IOS_PLATFORM STREQUAL "SIMULATOR")
-    # FIXME(liuyiqun): support "i386;x86_64" future
-    set(IOS_ARCH "x86_64")
+    set(IOS_ARCH "i386;x86_64")
   endif()
 endif()
 set(CMAKE_OSX_ARCHITECTURES ${IOS_ARCH} CACHE string  "Build architecture for iOS")
@@ -248,7 +246,7 @@ set(IOS_COMPILER_FLAGS "${XCODE_IOS_PLATFORM_VERSION_FLAGS} ${XCODE_IOS_BITCODE_
 
 # Hidden visibilty is required for cxx on iOS 
 set(CMAKE_C_FLAGS "${IOS_COMPILER_FLAGS} ${CMAKE_C_FLAGS}" CACHE STRING "C flags")
-set(CMAKE_CXX_FLAGS "${IOS_COMPILER_FLAGS} -fvisibility-inlines-hidden ${CMAKE_CXX_FLAGS}" CACHE STRING "CXX flags")
+set(CMAKE_CXX_FLAGS "${IOS_COMPILER_FLAGS} -fvisibility=hidden -fvisibility-inlines-hidden ${CMAKE_CXX_FLAGS}" CACHE STRING "CXX flags")
 
 set(IOS_LINK_FLAGS "${XCODE_IOS_PLATFORM_VERSION_FLAGS} -Wl,-search_paths_first")
 

cmake/cuda.cmake

+if(NOT WITH_GPU)
+    return()
+endif()
+
+set(paddle_known_gpu_archs "30 35 50 52 60 61 70")
+set(paddle_known_gpu_archs7 "30 35 50 52")
+set(paddle_known_gpu_archs8 "30 35 50 52 60 61")
+
+######################################################################################
+# A function for automatic detection of GPUs installed  (if autodetection is enabled)
+# Usage:
+#   detect_installed_gpus(out_variable)
+function(detect_installed_gpus out_variable)
+  if(NOT CUDA_gpu_detect_output)
+    set(cufile ${PROJECT_BINARY_DIR}/detect_cuda_archs.cu)
+
+    file(WRITE ${cufile} ""
+      "#include <cstdio>\n"
+      "int main() {\n"
+      "  int count = 0;\n"
+      "  if (cudaSuccess != cudaGetDeviceCount(&count)) return -1;\n"
+      "  if (count == 0) return -1;\n"
+      "  for (int device = 0; device < count; ++device) {\n"
+      "    cudaDeviceProp prop;\n"
+      "    if (cudaSuccess == cudaGetDeviceProperties(&prop, device))\n"
+      "      std::printf(\"%d.%d \", prop.major, prop.minor);\n"
+      "  }\n"
+      "  return 0;\n"
+      "}\n")
+
+    execute_process(COMMAND "${CUDA_NVCC_EXECUTABLE}" "-ccbin=${CUDA_HOST_COMPILER}"
+                    "--run" "${cufile}"
+                    WORKING_DIRECTORY "${PROJECT_BINARY_DIR}/CMakeFiles/"
+                    RESULT_VARIABLE nvcc_res OUTPUT_VARIABLE nvcc_out
+                    ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
+
+    if(nvcc_res EQUAL 0)
+      # only keep the last line of nvcc_out
+      STRING(REGEX REPLACE ";" "\\\\;" nvcc_out "${nvcc_out}")
+      STRING(REGEX REPLACE "\n" ";" nvcc_out "${nvcc_out}")
+      list(GET nvcc_out -1 nvcc_out)
+      string(REPLACE "2.1" "2.1(2.0)" nvcc_out "${nvcc_out}")
+      set(CUDA_gpu_detect_output ${nvcc_out} CACHE INTERNAL "Returned GPU architetures from detect_installed_gpus tool" FORCE)
+    endif()
+  endif()
+
+  if(NOT CUDA_gpu_detect_output)
+    message(STATUS "Automatic GPU detection failed. Building for all known architectures.")
+    set(${out_variable} ${paddle_known_gpu_archs} PARENT_SCOPE)
+  else()
+    set(${out_variable} ${CUDA_gpu_detect_output} PARENT_SCOPE)
+  endif()
+endfunction()
+
+
+########################################################################
+# Function for selecting GPU arch flags for nvcc based on CUDA_ARCH_NAME
+# Usage:
+#   select_nvcc_arch_flags(out_variable)
+function(select_nvcc_arch_flags out_variable)
+  # List of arch names
+  set(archs_names "Kepler" "Maxwell" "Pascal" "All" "Manual")
+  set(archs_name_default "All")
+  if(NOT CMAKE_CROSSCOMPILING)
+    list(APPEND archs_names "Auto")
+  endif()
+
+  # set CUDA_ARCH_NAME strings (so it will be seen as dropbox in CMake-Gui)
+  set(CUDA_ARCH_NAME ${archs_name_default} CACHE STRING "Select target NVIDIA GPU achitecture.")
+  set_property( CACHE CUDA_ARCH_NAME PROPERTY STRINGS "" ${archs_names} )
+  mark_as_advanced(CUDA_ARCH_NAME)
+
+  # verify CUDA_ARCH_NAME value
+  if(NOT ";${archs_names};" MATCHES ";${CUDA_ARCH_NAME};")
+    string(REPLACE ";" ", " archs_names "${archs_names}")
+    message(FATAL_ERROR "Only ${archs_names} architeture names are supported.")
+  endif()
+
+  if(${CUDA_ARCH_NAME} STREQUAL "Manual")
+    set(CUDA_ARCH_BIN ${paddle_known_gpu_archs} CACHE STRING "Specify 'real' GPU architectures to build binaries for, BIN(PTX) format is supported")
+    set(CUDA_ARCH_PTX "50"                     CACHE STRING "Specify 'virtual' PTX architectures to build PTX intermediate code for")
+    mark_as_advanced(CUDA_ARCH_BIN CUDA_ARCH_PTX)
+  else()
+    unset(CUDA_ARCH_BIN CACHE)
+    unset(CUDA_ARCH_PTX CACHE)
+  endif()
+
+  if(${CUDA_ARCH_NAME} STREQUAL "Kepler")
+    set(cuda_arch_bin "30 35")
+  elseif(${CUDA_ARCH_NAME} STREQUAL "Maxwell")
+    set(cuda_arch_bin "50")
+  elseif(${CUDA_ARCH_NAME} STREQUAL "Pascal")
+    set(cuda_arch_bin "60 61")
+  elseif(${CUDA_ARCH_NAME} STREQUAL "Volta")
+    set(cuda_arch_bin "70")
+  elseif(${CUDA_ARCH_NAME} STREQUAL "All")
+    set(cuda_arch_bin ${paddle_known_gpu_archs})
+  elseif(${CUDA_ARCH_NAME} STREQUAL "Auto")
+    detect_installed_gpus(cuda_arch_bin)
+  else()  # (${CUDA_ARCH_NAME} STREQUAL "Manual")
+    set(cuda_arch_bin ${CUDA_ARCH_BIN})
+  endif()
+
+  # remove dots and convert to lists
+  string(REGEX REPLACE "\\." "" cuda_arch_bin "${cuda_arch_bin}")
+  string(REGEX REPLACE "\\." "" cuda_arch_ptx "${CUDA_ARCH_PTX}")
+  string(REGEX MATCHALL "[0-9()]+" cuda_arch_bin "${cuda_arch_bin}")
+  string(REGEX MATCHALL "[0-9]+"   cuda_arch_ptx "${cuda_arch_ptx}")
+  list(REMOVE_DUPLICATES cuda_arch_bin)
+  list(REMOVE_DUPLICATES cuda_arch_ptx)
+
+  set(nvcc_flags "")
+  set(nvcc_archs_readable "")
+
+  # Tell NVCC to add binaries for the specified GPUs
+  foreach(arch ${cuda_arch_bin})
+    if(arch MATCHES "([0-9]+)\\(([0-9]+)\\)")
+      # User explicitly specified PTX for the concrete BIN
+      list(APPEND nvcc_flags -gencode arch=compute_${CMAKE_MATCH_2},code=sm_${CMAKE_MATCH_1})
+      list(APPEND nvcc_archs_readable sm_${CMAKE_MATCH_1})
+    else()
+      # User didn't explicitly specify PTX for the concrete BIN, we assume PTX=BIN
+      list(APPEND nvcc_flags -gencode arch=compute_${arch},code=sm_${arch})
+      list(APPEND nvcc_archs_readable sm_${arch})
+    endif()
+  endforeach()
+
+  # Tell NVCC to add PTX intermediate code for the specified architectures
+  foreach(arch ${cuda_arch_ptx})
+    list(APPEND nvcc_flags -gencode arch=compute_${arch},code=compute_${arch})
+    list(APPEND nvcc_archs_readable compute_${arch})
+  endforeach()
+
+  string(REPLACE ";" " " nvcc_archs_readable "${nvcc_archs_readable}")
+  set(${out_variable}          ${nvcc_flags}          PARENT_SCOPE)
+  set(${out_variable}_readable ${nvcc_archs_readable} PARENT_SCOPE)
+endfunction()
+
+message(STATUS "CUDA detected: " ${CUDA_VERSION})
+if (${CUDA_VERSION} LESS 7.0)
+  set(paddle_known_gpu_archs ${paddle_known_gpu_archs})
+elseif (${CUDA_VERSION} LESS 8.0) # CUDA 7.x
+  set(paddle_known_gpu_archs ${paddle_known_gpu_archs7})
+  list(APPEND CUDA_NVCC_FLAGS "-D_MWAITXINTRIN_H_INCLUDED")
+  list(APPEND CUDA_NVCC_FLAGS "-D__STRICT_ANSI__")
+elseif (${CUDA_VERSION} LESS 9.0) # CUDA 8.x
+  set(paddle_known_gpu_archs ${paddle_known_gpu_archs8})
+  list(APPEND CUDA_NVCC_FLAGS "-D_MWAITXINTRIN_H_INCLUDED")
+  list(APPEND CUDA_NVCC_FLAGS "-D__STRICT_ANSI__")
+  # CUDA 8 may complain that sm_20 is no longer supported. Suppress the
+  # warning for now.
+  list(APPEND CUDA_NVCC_FLAGS "-Wno-deprecated-gpu-targets")
+endif()
+
+include_directories(${CUDA_INCLUDE_DIRS})
+list(APPEND EXTERNAL_LIBS ${CUDA_LIBRARIES} ${CUDA_rt_LIBRARY})
+if(NOT WITH_DSO)
+    list(APPEND EXTERNAL_LIBS ${CUDNN_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} ${CUDA_curand_LIBRARY} ${NCCL_LIBRARY})
+endif(NOT WITH_DSO)
+
+# setting nvcc arch flags
+select_nvcc_arch_flags(NVCC_FLAGS_EXTRA)
+list(APPEND CUDA_NVCC_FLAGS ${NVCC_FLAGS_EXTRA})
+message(STATUS "Added CUDA NVCC flags for: ${NVCC_FLAGS_EXTRA_readable}")
+
+# Set C++11 support
+set(CUDA_PROPAGATE_HOST_FLAGS OFF)
+
+# Release/Debug flags set by cmake. Such as -O3 -g -DNDEBUG etc.
+# So, don't set these flags here.
+list(APPEND CUDA_NVCC_FLAGS "-std=c++11")
+list(APPEND CUDA_NVCC_FLAGS "--use_fast_math")
+list(APPEND CUDA_NVCC_FLAGS "-Xcompiler -fPIC")
+# Set :expt-relaxed-constexpr to suppress Eigen warnings
+list(APPEND CUDA_NVCC_FLAGS "--expt-relaxed-constexpr")
+
+if(CMAKE_BUILD_TYPE  STREQUAL "Debug")
+    list(APPEND CUDA_NVCC_FLAGS  ${CMAKE_CXX_FLAGS_DEBUG})
+elseif(CMAKE_BUILD_TYPE  STREQUAL "Release")
+    list(APPEND CUDA_NVCC_FLAGS  ${CMAKE_CXX_FLAGS_RELEASE})
+elseif(CMAKE_BUILD_TYPE  STREQUAL "RelWithDebInfo")
+    list(APPEND CUDA_NVCC_FLAGS  ${CMAKE_CXX_FLAGS_RELWITHDEBINFO})
+elseif(CMAKE_BUILD_TYPE  STREQUAL "MinSizeRel")
+    list(APPEND CUDA_NVCC_FLAGS  ${CMAKE_CXX_FLAGS_MINSIZEREL})
+endif()
+
+mark_as_advanced(CUDA_BUILD_CUBIN CUDA_BUILD_EMULATION CUDA_VERBOSE_BUILD)
+mark_as_advanced(CUDA_SDK_ROOT_DIR CUDA_SEPARABLE_COMPILATION)

cmake/external/mkldnn.cmake

@@ -40,10 +40,9 @@ INCLUDE_DIRECTORIES(${MKLDNN_INC_DIR})
 
 IF(${CBLAS_PROVIDER} STREQUAL "MKLML")
     SET(MKLDNN_DEPENDS   ${MKLML_PROJECT})
-    SET(MKLDNN_MKLROOT   ${MKLML_ROOT})
-    SET(MKLDNN_IOMP_LIB  ${MKLML_IOMP_LIB})
-    SET(MKLDNN_IOMP_DIR  ${MKLML_LIB_DIR})
-    MESSAGE(STATUS "Build MKLDNN with ${MKLDNN_MKLROOT}")
+    MESSAGE(STATUS "Build MKLDNN with MKLML ${MKLML_ROOT}")
+ELSE()
+    MESSAGE(FATAL_ERROR "Should enable MKLML when build MKLDNN")
 ENDIF()
 
 SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} -Wno-error=strict-overflow")
@@ -57,15 +56,16 @@ ExternalProject_Add(
     PREFIX              ${MKLDNN_SOURCES_DIR}
     UPDATE_COMMAND      ""
     CMAKE_ARGS          -DCMAKE_INSTALL_PREFIX=${MKLDNN_INSTALL_DIR}
-    CMAKE_ARGS          -DMKLROOT=${MKLDNN_MKLROOT}
+    CMAKE_ARGS          -DMKLROOT=${MKLML_ROOT}
     CMAKE_ARGS          -DCMAKE_C_FLAGS=${MKLDNN_CFLAG}
     CMAKE_ARGS          -DCMAKE_CXX_FLAGS=${MKLDNN_CXXFLAG}
     CMAKE_CACHE_ARGS    -DCMAKE_INSTALL_PREFIX:PATH=${MKLDNN_INSTALL_DIR}
-                        -DMKLROOT:PATH=${MKLDNN_MKLROOT}
+                        -DMKLROOT:PATH=${MKLML_ROOT}
 )
 
 ADD_LIBRARY(mkldnn SHARED IMPORTED GLOBAL)
 SET_PROPERTY(TARGET mkldnn PROPERTY IMPORTED_LOCATION ${MKLDNN_LIB})
 ADD_DEPENDENCIES(mkldnn ${MKLDNN_PROJECT})
-MESSAGE(STATUS "Mkldnn library: ${MKLDNN_LIB}")
+MESSAGE(STATUS "MKLDNN library: ${MKLDNN_LIB}")
+add_definitions(-DPADDLE_USE_MKLDNN)
 LIST(APPEND external_project_dependencies mkldnn)

cmake/external/openblas.cmake

@@ -29,7 +29,7 @@ IF(NOT ${CBLAS_FOUND})
         "${CBLAS_INSTALL_DIR}/lib/${CMAKE_STATIC_LIBRARY_PREFIX}openblas${CMAKE_STATIC_LIBRARY_SUFFIX}"
         CACHE FILEPATH "openblas library." FORCE)
 
-    SET(OPENBLAS_CC "${CMAKE_C_COMPILER}")
+    SET(OPENBLAS_CC "${CMAKE_C_COMPILER} -Wno-unused-but-set-variable -Wno-unused-variable")
 
     IF(CMAKE_CROSSCOMPILING)
         SET(OPTIONAL_ARGS HOSTCC=${HOST_C_COMPILER})
@@ -45,15 +45,14 @@ IF(NOT ${CBLAS_FOUND})
                 SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} TARGET=ARMV8 BINARY=64 USE_THREAD=0)
             ENDIF()
         ELSEIF(IOS)
-            # FIXME(liuyiqun): support multiple architectures
-            SET(OPENBLAS_COMMIT "b5c96fcfcdc82945502a2303116a64d89985daf5")
-            SET(OPENBLAS_CC "${OPENBLAS_CC} ${CMAKE_C_FLAGS} -isysroot ${CMAKE_OSX_SYSROOT}")
-            IF(CMAKE_OSX_ARCHITECTURES MATCHES "armv7")
-                SET(OPENBLAS_CC "${OPENBLAS_CC} -arch armv7")
-                SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} TARGET=ARMV7 ARM_SOFTFP_ABI=1 USE_THREAD=0)
-            ELSEIF(CMAKE_OSX_ARCHITECTURES MATCHES "arm64")
+            IF(CMAKE_OSX_ARCHITECTURES MATCHES "arm64")
+                SET(OPENBLAS_COMMIT "b5c96fcfcdc82945502a2303116a64d89985daf5")
+                SET(OPENBLAS_CC "${OPENBLAS_CC} ${CMAKE_C_FLAGS} -isysroot ${CMAKE_OSX_SYSROOT}")
                 SET(OPENBLAS_CC "${OPENBLAS_CC} -arch arm64")
                 SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} TARGET=ARMV8 BINARY=64 USE_THREAD=0 CROSS_SUFFIX=${CROSS_SUFFIX})
+            ELSE()
+                MESSAGE(FATAL_ERROR "OpenBLAS only support arm64 architectures on iOS. "
+                       "You can set IOS_USE_VECLIB_FOR_BLAS=ON or USE_EIGEN_FOR_BLAS=ON to use other blas library instead.")
             ENDIF()
         ELSEIF(RPI)
             # use hardfp
@@ -98,7 +97,7 @@ IF(NOT ${CBLAS_FOUND})
         ENDIF()
         INSTALL(CODE "execute_process(
             COMMAND ${CMAKE_COMMAND} -E copy_directory ${CBLAS_INSTALL_DIR}/lib
-                    destination ${CMAKE_INSTALL_PREFIX}/${TMP_INSTALL_DIR}
+                    ${CMAKE_INSTALL_PREFIX}/${TMP_INSTALL_DIR}
             )"
         )
         INSTALL(CODE "MESSAGE(STATUS \"Installing: \"

cmake/external/warpctc.cmake

@@ -12,6 +12,10 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+IF(MOBILE_INFERENCE)
+    return()
+ENDIF()
+
 INCLUDE(ExternalProject)
 
 SET(WARPCTC_SOURCES_DIR ${THIRD_PARTY_PATH}/warpctc)

cmake/flags.cmake

@@ -149,58 +149,3 @@ endforeach()
 foreach(flag ${GPU_COMMON_FLAGS})
     safe_set_nvflag(${flag})
 endforeach()
-
-
-set(CUDA_PROPAGATE_HOST_FLAGS OFF)
-
-# Release/Debug flags set by cmake. Such as -O3 -g -DNDEBUG etc.
-# So, don't set these flags here.
-LIST(APPEND CUDA_NVCC_FLAGS -std=c++11)
-LIST(APPEND CUDA_NVCC_FLAGS --use_fast_math)
-
-if(CMAKE_BUILD_TYPE  STREQUAL "Debug")
-    LIST(APPEND CUDA_NVCC_FLAGS  ${CMAKE_CXX_FLAGS_DEBUG})
-elseif(CMAKE_BUILD_TYPE  STREQUAL "Release")
-    LIST(APPEND CUDA_NVCC_FLAGS  ${CMAKE_CXX_FLAGS_RELEASE})
-elseif(CMAKE_BUILD_TYPE  STREQUAL "RelWithDebInfo")
-    LIST(APPEND CUDA_NVCC_FLAGS  ${CMAKE_CXX_FLAGS_RELWITHDEBINFO})
-elseif(CMAKE_BUILD_TYPE  STREQUAL "MinSizeRel")
-    LIST(APPEND CUDA_NVCC_FLAGS  ${CMAKE_CXX_FLAGS_MINSIZEREL})
-endif()
-
-function(specify_cuda_arch cuda_version cuda_arch)
-    if(${cuda_version} VERSION_GREATER "8.0")
-        foreach(capability 61 62)
-          if(${cuda_arch} STREQUAL ${capability})
-            list(APPEND __arch_flags " -gencode arch=compute_${cuda_arch},code=sm_${cuda_arch}")
-          endif()
-        endforeach()
-    elseif(${cuda_version} VERSION_GREATER "7.0" and ${cuda_arch} STREQUAL "53")
-        list(APPEND __arch_flags " -gencode arch=compute_${cuda_arch},code=sm_${cuda_arch}")
-    endif()
-endfunction()
-
-# Common gpu architectures: Kepler, Maxwell
-foreach(capability 30 35 50)
-      list(APPEND __arch_flags " -gencode arch=compute_${capability},code=sm_${capability}")
-endforeach()
-
-if (CUDA_VERSION VERSION_GREATER "7.0" OR CUDA_VERSION VERSION_EQUAL "7.0")
-      list(APPEND __arch_flags " -gencode arch=compute_52,code=sm_52")
-endif()
-
-# Modern gpu architectures: Pascal
-if (CUDA_VERSION VERSION_GREATER "8.0" OR CUDA_VERSION VERSION_EQUAL "8.0")
-      list(APPEND __arch_flags " -gencode arch=compute_60,code=sm_60")
-      list(APPEND CUDA_NVCC_FLAGS --expt-relaxed-constexpr)
-endif()
-
-# Custom gpu architecture
-set(CUDA_ARCH)
-
-if(CUDA_ARCH)
-  specify_cuda_arch(${CUDA_VERSION} ${CUDA_ARCH})
-endif()
-
-set(CUDA_NVCC_FLAGS ${__arch_flags} ${CUDA_NVCC_FLAGS})
-

cmake/util.cmake

@@ -115,8 +115,8 @@ function(link_paddle_exe TARGET_NAME)
         target_link_libraries(${TARGET_NAME} log)
     endif(ANDROID)
 
-    if(WITH_MKLDNN AND WITH_MKLML AND MKLDNN_IOMP_DIR)
-      target_link_libraries(${TARGET_NAME} "-L${MKLDNN_IOMP_DIR} -liomp5 -Wl,--as-needed")
+    if(WITH_MKLML AND MKLML_LIB_DIR AND MKLML_IOMP_LIB)
+      target_link_libraries(${TARGET_NAME} "-L${MKLML_LIB_DIR} -liomp5 -Wl,--as-needed")
     endif()
 
     add_dependencies(${TARGET_NAME} ${external_project_dependencies})

doc/api/v2/config/layer.rst

@@ -335,6 +335,16 @@ bilinear_interp
 ..  autoclass:: paddle.v2.layer.bilinear_interp
     :noindex:
 
+dot_prod
+---------
+.. autoclass:: paddle.v2.layer.dot_prod
+    :noindex:
+
+out_prod
+--------
+.. autoclass:: paddle.v2.layer.out_prod
+    :noindex:
+
 power
 -----
 ..  autoclass:: paddle.v2.layer.power

doc/design/mkldnn/README.MD

@@ -36,13 +36,13 @@ Figure 1. PaddlePaddle on IA.
 我们把集成方案大致分为了如下几个方面。
 
 ### CMake
-我们会在`CMakeLists.txt`中会添加`WITH_MKLDNN`的选项，当设置这个值为`ON`的时候会启用编译MKL-DNN功能。同时会自动开启OpenMP用于提高MKL-DNN的性能。
+我们会在`CMakeLists.txt`中会给用户添加一个`WITH_MKL`的开关，他是负责`WITH_MKLML`和`WITH_MKLDNN`的总开关。
 
-同时，我们会引入`WITH_MKLML`选项，用于选择是否使用MKL-DNN自带的MKLML安装包。这个安装包可以独立于MKL-DNN使用，但是建议在开启MKL-DNN的同时也打开MKLML的开关，这样才能发挥最好的性能。
+当打开`WITH_MKL`时，会开启MKLML的功能，作为PaddlePaddle的CBLAS和LAPACK库，同时会开启Intel OpenMP用于提高MKLML的性能。 如果系统支持AVX2指令集及以上，同时会开启MKL-DNN功能。
 
-所以，我们会在`cmake/external`目录新建`mkldnn.cmake`和`mklml.cmake`文件，它们会在编译PaddlePaddle的时候下载对应的软件包，并放到PaddlePaddle的third party目录中。
+当关闭`WITH_MKL`时，MKLML和MKL-DNN功能会同时关闭。
 
-**备注**：当`WITH_MKLML=ON`的时候，会优先使用这个包作为PaddlePaddle的CBLAS和LAPACK库，所以会稍微改动`cmake/cblas.cmake`中的逻辑。
+所以，我们会在`cmake/external`目录新建`mkldnn.cmake`和`mklml.cmake`文件，它们会在编译PaddlePaddle的时候下载对应的软件包，并放到PaddlePaddle的third party目录中。
 
 ### Layers
 所有MKL-DNN相关的C++ layers，都会按照PaddlePaddle的目录结构存放在

doc/design/ops/images/2_level_rnn.dot

 digraph G {
 
-  rnn [label="1-th level RNN" shape=box]
+  rnn [label="1st level RNN" shape=box]
 
   subgraph cluster0 {
     label = "time step 0"
@@ -8,7 +8,7 @@ digraph G {
     sent0 [label="sentence"]
     sent1 [label="sentence"]
 
-    rnn1 [label="2-th level RNN" shape=box]
+    rnn1 [label="2nd level RNN" shape=box]
 
     sent0 -> rnn1
     sent1 -> rnn1
@@ -20,7 +20,7 @@ digraph G {
     sent2 [label="sentence"]
     sent3 [label="sentence"]
 
-    rnn2 [label="2-th level RNN" shape=box]
+    rnn2 [label="2nd level RNN" shape=box]
 
     sent2 -> rnn2
     sent3 -> rnn2
@@ -32,7 +32,7 @@ digraph G {
     sent4 [label="sentence"]
     sent5 [label="sentence"]
 
-    rnn3 [label="2-th level RNN" shape=box]
+    rnn3 [label="2nd level RNN" shape=box]
 
     sent4 -> rnn3
     sent5 -> rnn3

doc/design/ops/rnn.md

 # RNNOp design
 
-This document is about an RNN operator which requires that instances in a mini-batch have the same length.  We will have a more flexible RNN operator.
+This document describes the RNN (Recurrent Neural Network) operator and how it is implemented in PaddlePaddle. The RNN op requires that all instances in a mini-batch have the same length. We will have a more flexible dynamic RNN operator in the future.
 
 ## RNN Algorithm Implementation
 
-<p aligh="center">
+<p align="center">
 <img src="./images/rnn.jpg"/>
 </p>
 
 The above diagram shows an RNN unrolled into a full network.
 
-There are several important concepts:
+There are several important concepts here:
 
-- *step-net*: the sub-graph to run at each step,
-- *memory*, $h_t$, the state of the current step,
-- *ex-memory*, $h_{t-1}$, the state of the previous step,
-- *initial memory value*, the ex-memory of the first step.
+- *step-net*: the sub-graph that runs at each step.
+- *memory*, $h_t$, the state of the current step.
+- *ex-memory*, $h_{t-1}$, the state of the previous step.
+- *initial memory value*, the memory of the first (initial) step.
 
 ### Step-scope
 
-There could be local variables defined in step-nets.  PaddlePaddle runtime realizes these variables in *step-scopes* -- scopes created for each step.
+There could be local variables defined in each step-net.  PaddlePaddle runtime realizes these variables in *step-scopes* which are created for each step.
 
-<p aligh="center">
+<p align="center">
 <img src="./images/rnn.png"/><br/>
-Figure 2 the RNN's data flow
+Figure 2 illustrates the RNN's data flow
 </p>
 
-Please be aware that all steps run the same step-net.  Each step
+Please be aware that every step runs the same step-net.  Each step does the following:
 
-1. creates the step-scope,
-2. realizes local variables, including step-outputs, in the step-scope, and
-3. runs the step-net, which could use these variables.
+1. Creates the step-scope.
+2. Initializes the local variables including step-outputs, in the step-scope.
+3. Runs the step-net, which uses the above mentioned variables.
 
-The RNN operator will compose its output from step outputs in step scopes.
+The RNN operator will compose its output from step outputs in each of the step scopes.
 
 ### Memory and Ex-memory
 
-Let's give more details about memory and ex-memory via a simply example:
+Let's give more details about memory and ex-memory using a simple example:
 
 $$
 h_t = U h_{t-1} + W x_t
 $$,
 
-where $h_t$ and $h_{t-1}$ are the memory and ex-memory of step $t$'s respectively.
+where $h_t$ and $h_{t-1}$ are the memory and ex-memory (previous memory) of step $t$ respectively.
 
-In the implementation, we can make an ex-memory variable either "refers to" the memory variable of the previous step,
-or copy the value of the previous memory value to the current ex-memory variable.
+In the implementation, we can make an ex-memory variable either "refer to" the memory variable of the previous step,
+or copy the memory value of the previous step to the current ex-memory variable.
 
 ### Usage in Python
 
 For more information on Block, please refer to the [design doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/block.md).
 
-We can define an RNN's step-net using Block:
+We can define an RNN's step-net using a Block:
 
 ```python
 import paddle as pd
 
-X = some_op() # x is some operator's output, and is a LoDTensor
+X = some_op() # x is some operator's output and is a LoDTensor
 a = some_op()
 
 # declare parameters
@@ -68,7 +68,7 @@ with rnn.stepnet():
     x = rnn.add_input(X)
     # declare a memory (rnn's step)
     h = rnn.add_memory(init=a)
-    # h.pre_state() means previous memory of rnn
+    # h.pre_state(), the previous memory of rnn
     new_state = pd.add_two( pd.matmul(W, x) + pd.matmul(U, h.pre_state()))
     # update current memory
     h.update(new_state)
@@ -80,19 +80,19 @@ out = rnn()
 
 Python API functions in above example:
 
-- `rnn.add_input` indicates the parameter is a variable that will be segmented into step-inputs.
-- `rnn.add_memory` creates a variable used as the memory.
-- `rnn.add_outputs` mark the variables that will be concatenated across steps into the RNN output.
+- `rnn.add_input`: indicates that the parameter is a variable that will be segmented into step-inputs.
+- `rnn.add_memory`: creates a variable used as the memory.
+- `rnn.add_outputs`: marks the variables that will be concatenated across steps into the RNN output.
 
 ### Nested RNN and LoDTensor
 
 An RNN whose step-net includes other RNN operators is known as an *nested RNN*.
 
-For example, we could have a 2-level RNN, where the top level corresponds to paragraphs, and the lower level corresponds to sentences.
+For example, we could have a 2-level RNN, where the top level corresponds to paragraphs, and the lower level corresponds to sentences. Each step of the higher level RNN also receives an input from the corresponding step of the lower level, and additionally the output from the previous time step at the same level.
 
-The following figure illustrates the feeding of text into the lower level, one sentence each step, and the feeding of step outputs to the top level. The final top level output is about the whole text.
+The following figure illustrates feeding in text into the lower level, one sentence at a step, and the feeding in step outputs to the top level. The final top level output is about the whole text.
 
-<p aligh="center">
+<p align="center">
 <img src="./images/2_level_rnn.png"/>
 </p>
 
@@ -110,7 +110,7 @@ a = some_op()
 
 # chapter_data is a set of 128-dim word vectors
 # the first level of LoD is sentence
-# the second level of LoD is chapter
+# the second level of LoD is a chapter
 chapter_data = pd.Variable(shape=[None, 128], type=pd.lod_tensor, level=2)
 
 def lower_level_rnn(paragraph):
@@ -138,14 +138,14 @@ with top_level_rnn.stepnet():
         pd.matmul(W0, paragraph_data) + pd.matmul(U0, h.pre_state()))
     top_level_rnn.add_outputs(h)
 
-# just output the last step
+# output the last step
 chapter_out = top_level_rnn(output_all_steps=False)
 ```
 
-in above example, the construction of the `top_level_rnn` calls  `lower_level_rnn`.  The input is a LoD Tensor. The top level RNN segments input text data into paragraphs, and the lower level RNN segments each paragraph into sentences.
+In the above example, the construction of the `top_level_rnn` calls  `lower_level_rnn`.  The input is an LoD Tensor. The top level RNN segments input text data into paragraphs, and the lower level RNN segments each paragraph into sentences.
 
-By default, the `RNNOp` will concatenate the outputs from all the time steps,
-if the `output_all_steps` set to False, it will only output the final time step.
+By default, the `RNNOp` will concatenate the outputs from all the time steps.
+If the `output_all_steps` is set to False, it will only output the final time step.
 
 
 <p align="center">

doc/design/ops/sequence_decoder.md

 # Design: Sequence Decoder Generating LoDTensors
-In tasks such as machine translation and image to text, 
-a [sequence decoder](https://github.com/PaddlePaddle/book/blob/develop/08.machine_translation/README.md) is necessary to generate sequences.
+In tasks such as machine translation and visual captioning,
+a [sequence decoder](https://github.com/PaddlePaddle/book/blob/develop/08.machine_translation/README.md) is necessary to generate sequences, one word at a time.
 
 This documentation describes how to implement the sequence decoder as an operator.
 
 ## Beam Search based Decoder
-The [beam search algorithm](https://en.wikipedia.org/wiki/Beam_search) is necessary when generating sequences, 
-it is a heuristic search algorithm that explores the paths by expanding the most promising node in a limited set.
+The [beam search algorithm](https://en.wikipedia.org/wiki/Beam_search) is necessary when generating sequences. It is a heuristic search algorithm that explores the paths by expanding the most promising node in a limited set.
 
-In the old version of PaddlePaddle, a C++ class `RecurrentGradientMachine` implements the general sequence decoder based on beam search, 
-due to the complexity, the implementation relays on a lot of special data structures, 
-quite trivial and hard to be customized by users.
+In the old version of PaddlePaddle, the C++ class `RecurrentGradientMachine` implements the general sequence decoder based on beam search, due to the complexity involved, the implementation relies on a lot of special data structures that are quite trivial and hard to be customized by users.
 
-There are a lot of heuristic tricks in the sequence generation tasks, 
-so the flexibility of sequence decoder is very important to users.
+There are a lot of heuristic tricks in the sequence generation tasks, so the flexibility of sequence decoder is very important to users.
 
-During PaddlePaddle's refactoring work,
-some new concept is proposed such as [LoDTensor](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/lod_tensor.md) and [TensorArray](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/tensor_array.md) that can better support sequence usage,
-and they can help to make the implementation of beam search based sequence decoder **more transparent and modular** .
+During the refactoring of PaddlePaddle, some new concepts are proposed such as:  [LoDTensor](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/lod_tensor.md) and [TensorArray](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/tensor_array.md) that can better support the sequence usage, and they can also help make the implementation of beam search based sequence decoder **more transparent and modular** .
 
-For example, the RNN sates, candidates IDs and probabilities of beam search can be represented as `LoDTensors`;
+For example, the RNN states, candidates IDs and probabilities of beam search can be represented all as `LoDTensors`;
 the selected candidate's IDs in each time step can be stored in a `TensorArray`, and `Packed` to the sentences translated.
 
 ## Changing LoD's absolute offset to relative offsets
-The current `LoDTensor` is designed to store levels of variable-length sequences,
-it stores several arrays of integers each represents a level.
+The current `LoDTensor` is designed to store levels of variable-length sequences. It stores several arrays of integers where each represents a level.
 
-The integers in each level represents the begin and end (not inclusive) offset of a sequence **in the underlying tensor**, 
-let's call this format the **absolute-offset LoD** for clear.
+The integers in each level represent the begin and end (not inclusive) offset of a sequence **in the underlying tensor**,
+let's call this format the **absolute-offset LoD** for clarity.
 
-The relative-offset LoD can fast retrieve any sequence but fails to represent empty sequences, for example, a two-level LoD is as follows
+The relative-offset LoD can retrieve any sequence very quickly but fails to represent empty sequences, for example, a two-level LoD is as follows
 ```python
 [[0, 3, 9]
  [0, 2, 3, 3, 3, 9]]
@@ -41,10 +34,9 @@ The first level tells that there are two sequences:
 while on the second level, there are several empty sequences that both begin and end at `3`.
 It is impossible to tell how many empty second-level sequences exist in the first-level sequences.
 
-There are many scenarios that relay on empty sequence representation,
-such as machine translation or image to text, one instance has no translations or the empty candidate set for a prefix.
+There are many scenarios that rely on empty sequence representation, for example in machine translation or visual captioning, one instance has no translation or the empty candidate set for a prefix.
 
-So let's introduce another format of LoD, 
+So let's introduce another format of LoD,
 it stores **the offsets of the lower level sequences** and is called **relative-offset** LoD.
 
 For example, to represent the same sequences of the above data
@@ -54,19 +46,18 @@ For example, to represent the same sequences of the above data
  [0, 2, 3, 3, 3, 9]]
 ```
 
-the first level represents that there are two sequences, 
+the first level represents that there are two sequences,
 their offsets in the second-level LoD is `[0, 3)` and `[3, 5)`.
 
 The second level is the same with the relative offset example because the lower level is a tensor.
 It is easy to find out the second sequence in the first-level LoD has two empty sequences.
 
-The following demos are based on relative-offset LoD.
+The following examples are based on relative-offset LoD.
 
 ## Usage in a simple machine translation model
-Let's start from a simple machine translation model that is simplified from [machine translation chapter](https://github.com/PaddlePaddle/book/tree/develop/08.machine_translation) to draw a simple blueprint of what a sequence decoder can do and how to use it.
+Let's start from a simple machine translation model that is simplified from the [machine translation chapter](https://github.com/PaddlePaddle/book/tree/develop/08.machine_translation) to draw a blueprint of what a sequence decoder can do and how to use it.
 
-The model has an encoder that learns the semantic vector from a sequence,
-and a decoder which uses the sequence decoder to generate new sentences.
+The model has an encoder that learns the semantic vector from a sequence, and a decoder which uses the sequence encoder to generate new sentences.
 
 **Encoder**
 ```python
@@ -117,7 +108,7 @@ def generate():
         # which means there are 2 sentences to translate
         #   - the first sentence has 1 translation prefixes, the offsets are [0, 1)
         #   - the second sentence has 2 translation prefixes, the offsets are [1, 3) and [3, 6)
-        # the target_word.lod is 
+        # the target_word.lod is
         # [[0, 1, 6]
         #  [0, 2, 4, 7, 9 12]]
         # which means 2 sentences to translate, each has 1 and 5 prefixes
@@ -154,37 +145,36 @@ def generate():
 
 translation_ids, translation_scores = decoder()
 ```
-The `decoder.beam_search` is a operator that given the candidates and the scores of translations including the candidates,
-return the result of the beam search algorithm.
+The `decoder.beam_search` is an operator that, given the candidates and the scores of translations including the candidates,
+returns the result of the beam search algorithm.
 
-In this way, users can customize anything on the inputs or outputs of beam search, for example, two ways to prune some translation prefixes
+In this way, users can customize anything on the input or output of beam search, for example:
 
-1. meke the correspondind elements in `topk_generated_scores` zero or some small values, beam_search will discard this candidate.
-2. remove some specific candidate in `selected_ids`
-3. get the final `translation_ids`, remove the translation sequence in it.
+1. Make the corresponding elements in `topk_generated_scores` zero or some small values, beam_search will discard this candidate.
+2. Remove some specific candidate in `selected_ids`.
+3. Get the final `translation_ids`, remove the translation sequence in it.
 
-The implementation of sequence decoder can reuse the C++ class [RNNAlgorithm](https://github.com/Superjom/Paddle/blob/68cac3c0f8451fe62a4cdf156747d6dc0ee000b3/paddle/operators/dynamic_recurrent_op.h#L30),
-so the python syntax is quite similar to a [RNN](https://github.com/Superjom/Paddle/blob/68cac3c0f8451fe62a4cdf156747d6dc0ee000b3/doc/design/block.md#blocks-with-for-and-rnnop).
+The implementation of sequence decoder can reuse the C++ class:  [RNNAlgorithm](https://github.com/Superjom/Paddle/blob/68cac3c0f8451fe62a4cdf156747d6dc0ee000b3/paddle/operators/dynamic_recurrent_op.h#L30),
+so the python syntax is quite similar to that of an  [RNN](https://github.com/Superjom/Paddle/blob/68cac3c0f8451fe62a4cdf156747d6dc0ee000b3/doc/design/block.md#blocks-with-for-and-rnnop).
 
-Both of them are two-level `LoDTensors`
+Both of them are two-level `LoDTensors`:
 
-- the first level represents `batch_size` of (source) sentences;
-- the second level represents the candidate ID sets for translation prefix.
+- The first level represents `batch_size` of (source) sentences.
+- The second level represents the candidate ID sets for translation prefix.
 
-for example, 3 source sentences to translate, and has 2, 3, 1 candidates.
+For example, 3 source sentences to translate, and has 2, 3, 1 candidates.
 
-Unlike an RNN, in sequence decoder, the previous state and the current state have different LoD and shape,
-a `lod_expand` operator is used to expand the LoD of the previous state to fit the current state.
+Unlike an RNN, in sequence decoder, the previous state and the current state have different LoD and shape, and an `lod_expand` operator is used to expand the LoD of the previous state to fit the current state.
 
-For example, the previous state
+For example, the previous state:
 
 * LoD is `[0, 1, 3][0, 2, 5, 6]`
 * content of tensor is `a1 a2 b1 b2 b3 c1`
 
-the current state stored in `encoder_ctx_expanded`
+the current state is stored in `encoder_ctx_expanded`:
 
 * LoD is `[0, 2, 7][0 3 5 8 9 11 11]`
-* the content is 
+* the content is
   - a1 a1 a1 (a1 has 3 candidates, so the state should be copied 3 times for each candidates)
   - a2 a2
   - b1 b1 b1
@@ -192,54 +182,48 @@ the current state stored in `encoder_ctx_expanded`
   - b3 b3
   - None (c1 has 0 candidates, so c1 is dropped)
 
-Benefit from the relative offset LoD, empty candidate set can be represented naturally.
+The benefit from the relative offset LoD is that the empty candidate set can be represented naturally.
 
-the status in each time step can be stored in `TensorArray`, and `Pack`ed to a final LoDTensor, the corresponding syntax is 
+The status in each time step can be stored in `TensorArray`, and `Pack`ed to a final LoDTensor. The corresponding syntax is:
 
 ```python
 decoder.output(selected_ids)
 decoder.output(selected_generation_scores)
 ```
 
-the `selected_ids` is the candidate ids for the prefixes, 
-it will be `Packed` by `TensorArray` to a two-level `LoDTensor`,
-the first level represents the source sequences,
-the second level represents generated sequences.
+The `selected_ids` are the candidate ids for the prefixes, and will be `Packed` by `TensorArray` to a two-level `LoDTensor`, where the first level represents the source sequences and the second level represents generated sequences.
 
-Pack the `selected_scores` will get a `LoDTensor` that stores scores of each candidate of translations.
+Packing the `selected_scores` will get a `LoDTensor` that stores scores of each translation candidate.
 
-Pack the `selected_generation_scores` will get a `LoDTensor`, and each tail is the probability of the translation.
+Packing the `selected_generation_scores` will get a `LoDTensor`, and each tail is the probability of the translation.
 
 ## LoD and shape changes during decoding
 <p align="center">
   <img src="./images/LOD-and-shape-changes-during-decoding.jpg"/>
 </p>
 
-According the image above, the only phrase to change LoD is beam search.
+According to the image above, the only phase that changes the LoD is beam search.
 
 ## Beam search design
-The beam search algorthm will be implemented as one method of the sequence decoder, it has 3 inputs
+The beam search algorithm will be implemented as one method of the sequence decoder and has 3 inputs:
 
-1. `topk_ids`, top K candidate ids for each prefix.
+1. `topk_ids`, the top K candidate ids for each prefix.
 2. `topk_scores`, the corresponding scores for `topk_ids`
 3. `generated_scores`, the score of the prefixes.
 
-All of the are LoDTensors, so that the sequence affilication is clear.
-Beam search will keep a beam for each prefix and select a smaller candidate set for each prefix.
+All of these are LoDTensors, so that the sequence affiliation is clear. Beam search will keep a beam for each prefix and select a smaller candidate set for each prefix.
 
-It will return three variables
+It will return three variables:
 
 1. `selected_ids`, the final candidate beam search function selected for the next step.
 2. `selected_scores`, the scores for the candidates.
-3. `generated_scores`, the updated scores for each prefixes (with the new candidates appended).
+3. `generated_scores`, the updated scores for each prefix (with the new candidates appended).
 
 ## Introducing the LoD-based `Pack` and `Unpack` methods in `TensorArray`
-The `selected_ids`, `selected_scores` and `generated_scores` are LoDTensors,
-and they exist in each time step,
+The `selected_ids`, `selected_scores` and `generated_scores` are LoDTensors that exist at each time step,
 so it is natural to store them in arrays.
 
-Currently, PaddlePaddle has a module called `TensorArray` which can store an array of tensors,
-the results of beam search are better to store in a `TensorArray`.
+Currently, PaddlePaddle has a module called `TensorArray` which can store an array of tensors. It is better to store the results of beam search in a `TensorArray`.
 
-The `Pack` and `UnPack` in `TensorArray` are used to package tensors in the array to a `LoDTensor` or split the `LoDTensor` to an array of tensors. 
-It needs some extensions to support pack or unpack an array of `LoDTensors`.
+The `Pack` and `UnPack` in `TensorArray` are used to pack tensors in the array to an `LoDTensor` or split the `LoDTensor` to an array of tensors.
+It needs some extensions to support the packing or unpacking an array of `LoDTensors`.

doc/howto/dev/write_docs_cn.rst

@@ -34,7 +34,7 @@ PaddlePaddle的文档构建有两种方式。
     cd TO_YOUR_PADDLE_CLONE_PATH
     mkdir -p build
     cd build
-    cmake .. -DCMAKE_BUILD_TYPE=Debug -DWITH_GPU=OFF -DWITH_MKLDNN=OFF -DWITH_MKLML=OFF -DWITH_DOC=ON
+    cmake .. -DCMAKE_BUILD_TYPE=Debug -DWITH_GPU=OFF -DWITH_MKL=OFF -DWITH_DOC=ON
     make gen_proto_py
     make paddle_docs paddle_docs_cn
 

doc/mobile/cross_compiling_for_android_cn.md

-# 构建Android平台上的PaddlePaddle库
+# Android平台编译指南
 
 用户可通过如下两种方式，交叉编译Android平台上适用的PaddlePaddle库：
 - 基于Docker容器的编译方式

doc/mobile/cross_compiling_for_ios_cn.md

-# 构建iOS平台上的PaddlePaddle库
+# iOS平台编译指南
 交叉编译iOS平台上适用的PaddlePaddle库，需要在MacOS系统上进行。本文的将介绍在MacOS上，从源码交叉编译iOS平台上适用的PaddlePaddle库。
 
 ## 准备交叉编译环境
@@ -25,7 +25,7 @@ iOS平台可选配置参数：
 - `IOS_PLATFORM`，可设置为`OS/SIMULATOR`，默认值为`OS`。
   - `OS`，构建目标为`arm`架构的iPhone或者iPad等物理设备。
   - `SIMULATOR`，构建目标为`x86`架构的模拟器平台。
-- `IOS_ARCH`，目标架构。针对不同的`IOS_PLATFORM`，可设置的目标架构如下表所示：
+- `IOS_ARCH`，目标架构。针对不同的`IOS_PLATFORM`，可设置的目标架构如下表所示，默认编译所有架构：
 
     <table class="docutils">
     <colgroup>
@@ -41,11 +41,11 @@ iOS平台可选配置参数：
     <tbody valign="top">
       <tr class="row-even">
       <td>OS</td>
-      <td>armv7, armv7s, arm64 (默认)</td>
+      <td>armv7, armv7s, arm64 </td>
     </tr>
     <tr class="row-odd">
       <td>SIMULATOR</td>
-      <td>i386, x86_64 (默认)</td>
+      <td>i386, x86_64 </td>
     </tr>
     </tbody>
     </table>
@@ -66,7 +66,7 @@ iOS平台可选配置参数：
 ```bash
 cmake -DCMAKE_SYSTEM_NAME=iOS \
       -DIOS_PLATFORM=OS \
-      -DIOS_ARCH="arm64" \
+      -DIOS_ARCH="armv7;arm64" \
       -DIOS_ENABLE_BITCODE=ON \
       -DIOS_USE_VECLIB_FOR_BLAS=ON \
       -DCMAKE_INSTALL_PREFIX=your/path/to/install \
@@ -112,6 +112,6 @@ $ make install
 - `lib`目录，其中包含PaddlePaddle的C-API静态库
 - `third_party`目录，其中包含所依赖的所有第三方库
 
-注意，不同架构的PaddlePaddle库建议安装到不同的目录下，然后使用`lipo`工具将多个静态库合并成一个支持多个架构的fat库。
+注意，如果PaddlePaddle库需要同时支持真机和模拟器，则需要分别编译真机和模拟器版本，然后使用`lipo`工具合并fat库。
 
 自此，PaddlePaddle库已经安装完成，用户可将合成的fat库用于深度学习相关的iOS App中，调用方法见C-API文档。

doc/mobile/cross_compiling_for_raspberry_cn.md

-# 构建Raspberry Pi平台上的PaddlePaddle库
+# Raspberry Pi平台编译指南
 
 通常有两个方法来构建基于 Rasspberry Pi 的版本：
 

paddle/capi/Main.cpp

@@ -29,6 +29,9 @@ static void initPaddle(int argc, char** argv) {
 
 extern "C" {
 paddle_error paddle_init(int argc, char** argv) {
+  static bool isInit = false;
+  if (isInit) return kPD_NO_ERROR;
+
   std::vector<char*> realArgv;
   realArgv.reserve(argc + 1);
   realArgv.push_back(strdup(""));
@@ -37,6 +40,7 @@ paddle_error paddle_init(int argc, char** argv) {
   }
   initPaddle(argc + 1, realArgv.data());
   free(realArgv[0]);
+  isInit = true;
   return kPD_NO_ERROR;
 }
 }

paddle/cuda/include/hl_gpu.h

@@ -25,7 +25,9 @@ limitations under the License. */
 #include "hl_matrix.h"
 #include "hl_sequence.h"
 #include "hl_sparse.h"
+#ifndef PADDLE_MOBILE_INFERENCE
 #include "hl_warpctc_wrap.h"
+#endif
 
 #ifdef HPPL_STUB_FUNC
 #include "stub/hl_aggregate_stub.h"

paddle/framework/CMakeLists.txt

@@ -38,9 +38,9 @@ py_proto_compile(framework_py_proto SRCS framework.proto)
 add_custom_target(framework_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E touch __init__.py)
 add_dependencies(framework_py_proto framework_py_proto_init)
 add_custom_command(TARGET framework_py_proto POST_BUILD
-    COMMAND ${CMAKE_COMMAND} -E make_directory ${PADDLE_SOURCE_DIR}/python/paddle/v2/framework/proto
-    COMMAND cp *.py ${PADDLE_SOURCE_DIR}/python/paddle/v2/framework/proto/
-    COMMENT "Copy generated python proto into directory paddle/v2/framework/proto."
+    COMMAND ${CMAKE_COMMAND} -E make_directory ${PADDLE_SOURCE_DIR}/python/paddle/v2/fluid/proto
+    COMMAND cp *.py ${PADDLE_SOURCE_DIR}/python/paddle/v2/fluid/proto/
+    COMMENT "Copy generated python proto into directory paddle/v2/fluid/proto."
     WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
 
 cc_library(backward SRCS backward.cc DEPS net_op)

paddle/framework/backward.cc

@@ -270,6 +270,19 @@ static bool AllGradInSet(const std::vector<std::string>& names,
       return false;
     }
   }
+  if (VLOG_IS_ON(10)) {
+    std::ostringstream sout;
+    sout << "All input {";
+    for (auto& name : names) {
+      sout << name << ",";
+    }
+    sout << "} is in {";
+    for (auto& name : set) {
+      sout << name << ",";
+    }
+    sout << "}";
+    VLOG(10) << sout.str();
+  }
   return true;
 }
 
@@ -290,14 +303,12 @@ static void CreateGradVarInBlock(
   auto ops = block_desc->AllOps();
   for (size_t op_index = grad_op_start_index; op_index < ops.size();
        ++op_index) {
-    bool need_infer_shape = false;
     std::unordered_set<std::string> new_vars;
     ForEachVarName(ops[op_index]->Outputs(),
                    [&](const std::string& grad_var_name) {
                      if (block_desc->HasVar(grad_var_name)) {
                        return false;
                      }
-                     need_infer_shape = true;
                      auto var = block_desc->Var(grad_var_name);
                      new_vars.insert(var->Name());
                      auto it = param_name_map.find(grad_var_name);
@@ -311,23 +322,21 @@ static void CreateGradVarInBlock(
                      grad_record.op_idx_ = static_cast<int>(op_index);
                      return false; /* not break */
                    });
-    if (need_infer_shape) {
-      ops[op_index]->InferVarType(block_desc);
-      for (auto& arg : ops[op_index]->OutputArgumentNames()) {
-        if (new_vars.find(arg) == new_vars.end()) {
-          continue;
-        }
-        auto pname = FwdName(arg);
-        auto* param = block_desc->FindVarRecursive(pname);
-        auto* grad = block_desc->FindVar(arg);
-        if (param == nullptr) {
-          grad->SetDataType(DataType::FP32);
-        } else {
-          grad->SetDataType(param->GetDataType());
-        }
+    ops[op_index]->InferVarType(block_desc);
+    for (auto& arg : ops[op_index]->OutputArgumentNames()) {
+      if (new_vars.find(arg) == new_vars.end()) {
+        continue;
+      }
+      auto pname = FwdName(arg);
+      auto* param = block_desc->FindVarRecursive(pname);
+      auto* grad = block_desc->FindVar(arg);
+      if (param == nullptr) {
+        grad->SetDataType(DataType::FP32);
+      } else {
+        grad->SetDataType(param->GetDataType());
       }
-      ops[op_index]->InferShape(*block_desc);
     }
+    ops[op_index]->InferShape(*block_desc);
   }
 }
 
@@ -387,6 +396,7 @@ std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
     ProgramDescBind& program_desc, int block_idx,
     std::unordered_set<std::string>* no_grad_vars,
     std::unordered_map<std::string, std::string>* grad_to_var) {
+  VLOG(5) << "MakeBlockBackward";
   BlockDescBind* cur_block = program_desc.MutableBlock(block_idx);
   std::vector<OpDescBind*> op_descs = cur_block->AllOps();
   std::unordered_map<std::string, std::vector<size_t>> dup_out_ops;
@@ -394,9 +404,10 @@ std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
   std::vector<std::unique_ptr<OpDescBind>> backward_descs;
 
   for (auto it = op_descs.rbegin(); it != op_descs.rend(); ++it) {
+    VLOG(5) << "Making backward " << (*it)->Type() << " op";
     std::vector<std::unique_ptr<OpDescBind>> op_grads;
 
-    if ((*it)->Type() == "recurrent") {
+    if ((*it)->Type() == "recurrent" || (*it)->Type() == "while") {
       int step_block_idx = (*it)->GetBlockAttr("step_block");
       BlockDescBind* backward_block = CreateStepBlock(
           program_desc, no_grad_vars, grad_to_var, step_block_idx);
@@ -410,6 +421,15 @@ std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
       op_grads = MakeOpGrad(*it, no_grad_vars, grad_to_var);
     }
 
+    if (VLOG_IS_ON(10)) {
+      std::ostringstream sout;
+      sout << "Made ";
+      for (auto& op_grad : op_grads) {
+        sout << op_grad->Type() << " ";
+      }
+      VLOG(10) << sout.str();
+    }
+
     for (const auto& desc : op_grads) {
       for (const std::string& out_name : desc->OutputArgumentNames()) {
         if (out_name.find("@GRAD") == std::string::npos) {
@@ -425,6 +445,8 @@ std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
         op_grads.begin(), op_grads.end(), std::back_inserter(backward_descs),
         [](std::unique_ptr<OpDescBind>& ptr) { return std::move(ptr); });
   }
+
+  VLOG(5) << "Appending Sums";
   // Check whether some variables are written more than once
   std::list<std::pair<size_t, std::unique_ptr<OpDescBind>>> pending_sum_ops;
   for (const auto& dup : dup_out_ops) {
@@ -432,16 +454,22 @@ std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
     const std::vector<size_t> dup_op = dup.second;
     if (out_name != kEmptyVarName && dup_op.size() > 1) {
       std::vector<std::string> sum_op_inputs;
+      std::string next_g_name = out_name;
       for (size_t i = 0; i < dup_op.size(); ++i) {
+        VLOG(10) << backward_descs[dup_op[i]]->Type() << " has " << out_name
+                 << " duplicated";
         std::string new_name = out_name + "@RENAME@" + std::to_string(i);
-        backward_descs[dup_op[i]]->Rename(out_name, new_name);
+        backward_descs[dup_op[i]]->RenameOutput(out_name, new_name);
+        backward_descs[dup_op[i]]->RenameInput(out_name, next_g_name);
         sum_op_inputs.emplace_back(new_name);
+        next_g_name = sum_op_inputs.back();
       }
       std::unique_ptr<OpDescBind> sum_op(new OpDescBind(
           "sum", {{"X", sum_op_inputs}}, {{"Out", {out_name}}}, {}));
       pending_sum_ops.push_back({dup_op.back(), std::move(sum_op)});
     }
   }
+
   pending_sum_ops.sort(
       [](const std::pair<size_t, std::unique_ptr<OpDescBind>>& a,
          const std::pair<size_t, std::unique_ptr<OpDescBind>>& b) {
@@ -452,6 +480,8 @@ std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
                           std::move(p.second));
   }
 
+  VLOG(5) << "MakeBlockBackward Finished";
+
   return backward_descs;
 }
 
@@ -483,19 +513,14 @@ ParamGradInfoMap AppendBackward(
   const int root_block_idx = 0;
   auto root_block = program_desc.MutableBlock(root_block_idx);
 
-  // insert fill one op for target
-  // TODO(qiao) add some check to the target.
   std::string fill_one_op_out = GradVarName(target.Name());
-  std::vector<int64_t> target_shape_desc = target.Shape();
-  std::vector<int> target_shape;
-  std::transform(target_shape_desc.begin(), target_shape_desc.end(),
-                 std::back_inserter(target_shape),
-                 [](int64_t dim) { return static_cast<int>(dim); });
+  bool is_scalar = target.Shape() == std::vector<int64_t>{1};
+  PADDLE_ENFORCE(is_scalar, "target should be scalar");
   VLOG(3) << "backward from loss=" << target.Name()
           << " data_type=" << target.GetDataType();
   std::unique_ptr<OpDescBind> fill_one_op(
       new OpDescBind("fill_constant", {}, {{"Out", {fill_one_op_out}}},
-                     {{"shape", target_shape},
+                     {{"shape", std::vector<int>{1}},
                       {"value", static_cast<float>(1.0)},
                       {"data_type", target.GetDataType()}}));
   // infer var type of fill_one_op

paddle/framework/backward_test.cc

@@ -508,6 +508,7 @@ TEST(Backward, simple_single_op) {
   op->SetOutput("Out", {"out"});
 
   auto target = f::VarDescBind("out");
+  target.SetShape({1});
   auto var_to_grad = AppendBackward(program, target, {});
 
   ASSERT_EQ(block->AllOps().size(), 3UL);
@@ -544,6 +545,7 @@ TEST(Backward, default_attribute) {
   op->CheckAttrs();
 
   auto target = f::VarDescBind("out");
+  target.SetShape({1});
   AppendBackward(program, target, {});
 
   ASSERT_EQ(block->AllOps().size(), 3UL);
@@ -581,6 +583,7 @@ TEST(Backward, simple_mult_op) {
   op3->SetOutput("Out", {"out3"});
 
   auto target = f::VarDescBind("out3");
+  target.SetShape({1});
   size_t forward_len = block->AllOps().size();
   auto var_to_grad = AppendBackward(program, target, {});
 
@@ -670,6 +673,7 @@ TEST(Backward, intermedia_var_no_grad) {
   op4->SetOutput("Out", {"out4"});
 
   auto target = f::VarDescBind("out4");
+  target.SetShape({1});
   size_t forward_len = block->AllOps().size();
   auto var_to_grad = AppendBackward(program, target, {"out3"});
 
@@ -730,6 +734,7 @@ TEST(Backward, var_no_grad) {
   op2->SetOutput("Z", {"z2"});
 
   auto target = f::VarDescBind("z2");
+  target.SetShape({1});
   size_t forward_len = block->AllOps().size();
   auto var_to_grad = AppendBackward(program, target, {"z1"});
 
@@ -810,6 +815,7 @@ TEST(Backward, shared_var) {
   op3->SetOutput("Out", {"out3"});
 
   auto target = f::VarDescBind("out3");
+  target.SetShape({1});
   size_t forward_len = block->AllOps().size();
   auto var_to_grad = AppendBackward(program, target, {});
 
@@ -888,6 +894,7 @@ TEST(Backward, half_backward) {
   op1->SetOutput("Out", {"out"});
 
   auto target = f::VarDescBind("out");
+  target.SetShape({1});
   size_t forward_len = block->AllOps().size();
   auto var_to_grad = AppendBackward(program, target, {"b"});
   f::OpDescBind *fill_op = block->AllOps()[forward_len];

paddle/framework/data_type.h

@@ -29,6 +29,8 @@ inline DataType ToDataType(std::type_index type) {
     return DataType::INT32;
   } else if (typeid(int64_t).hash_code() == type.hash_code()) {
     return DataType::INT64;
+  } else if (typeid(bool).hash_code() == type.hash_code()) {
+    return DataType::BOOL;
   } else {
     PADDLE_THROW("Not supported");
   }
@@ -44,6 +46,8 @@ inline std::type_index ToTypeIndex(DataType type) {
       return typeid(int);
     case DataType::INT64:
       return typeid(int64_t);
+    case DataType::BOOL:
+      return typeid(bool);
     default:
       PADDLE_THROW("Not support type %d", type);
   }
@@ -64,6 +68,9 @@ inline void VisitDataType(DataType type, Visitor visitor) {
     case DataType::INT64:
       visitor.template operator()<int64_t>();
       break;
+    case DataType::BOOL:
+      visitor.template operator()<bool>();
+      break;
     default:
       PADDLE_THROW("Not supported");
   }

paddle/framework/ddim.cc

@@ -60,8 +60,7 @@ void make_ddim(DDim& ddim, const int64_t* dims, int n) {
       ddim = make_dim<9>(dims);
       break;
     default:
-      throw std::invalid_argument(
-          "Dynamic dimensions must have between [1, 9] dimensions.");
+      PADDLE_THROW("Dynamic dimensions must have between [1, 9] dimensions.");
   }
 }
 

paddle/framework/executor.cc

@@ -120,6 +120,7 @@ void Executor::Run(const ProgramDescBind& pdesc, Scope* scope, int block_id,
 
   for (auto& op_desc : block.AllOps()) {
     auto op = paddle::framework::OpRegistry::CreateOp(*op_desc);
+    VLOG(10) << op->DebugString();
     op->Run(*local_scope, *device);
   }
   if (create_local_scope) {

paddle/framework/op_desc.cc

@@ -235,6 +235,23 @@ void OpDescBind::Rename(const std::string &old_name,
   need_update_ = true;
 }
 
+void OpDescBind::RenameOutput(const std::string &old_name,
+                              const std::string &new_name) {
+  for (auto &output : outputs_) {
+    std::replace(output.second.begin(), output.second.end(), old_name,
+                 new_name);
+  }
+  need_update_ = true;
+}
+
+void OpDescBind::RenameInput(const std::string &old_name,
+                             const std::string &new_name) {
+  for (auto &input : inputs_) {
+    std::replace(input.second.begin(), input.second.end(), old_name, new_name);
+  }
+  need_update_ = true;
+}
+
 struct SetAttrDescVisitor : public boost::static_visitor<void> {
   explicit SetAttrDescVisitor(OpDesc::Attr *attr) : attr_(attr) {}
   mutable OpDesc::Attr *attr_;
@@ -448,7 +465,12 @@ const std::vector<std::string> &CompileTimeInferShapeContext::Outputs(
 DDim CompileTimeInferShapeContext::GetDim(const std::string &name) const {
   auto var = block_.FindVarRecursive(name);
   PADDLE_ENFORCE(var != nullptr, "Cannot find variable %s", name);
-  return framework::make_ddim(var->Shape());
+  try {
+    return framework::make_ddim(var->Shape());
+  } catch (...) {
+    VLOG(5) << "GetDim of variable " << name << " error";
+    std::rethrow_exception(std::current_exception());
+  }
 }
 
 void CompileTimeInferShapeContext::SetDim(const std::string &name,

paddle/framework/op_desc.h

@@ -73,6 +73,10 @@ class OpDescBind {
 
   void Rename(const std::string &old_name, const std::string &new_name);
 
+  void RenameOutput(const std::string &old_name, const std::string &new_name);
+
+  void RenameInput(const std::string &old_name, const std::string &new_name);
+
   // Only be used in C++
   const AttributeMap &GetAttrMap() const;
 

paddle/framework/operator.cc

@@ -403,19 +403,6 @@ class RuntimeInferShapeContext : public InferShapeContext {
 
 void OperatorWithKernel::Run(const Scope& scope,
                              const platform::DeviceContext& dev_ctx) const {
-  if (VLOG_IS_ON(1)) {
-    auto inputs = this->InputVars();
-    auto outputs = this->OutputVars(true);
-    std::ostringstream sout;
-    sout << "Run operator " << this->Type() << " From [";
-    std::ostream_iterator<std::string> out_it(sout, ",");
-    std::copy(inputs.begin(), inputs.end(), out_it);
-    sout << "] to [";
-    std::copy(outputs.begin(), outputs.end(), out_it);
-    sout << "]";
-    VLOG(1) << sout.str();
-  }
-
   RuntimeInferShapeContext infer_shape_ctx(*this, scope);
   this->InferShape(&infer_shape_ctx);
 

paddle/framework/scope.cc

@@ -38,11 +38,12 @@ Scope& Scope::NewScope() const {
 Variable* Scope::Var(const std::string& name) {
   auto iter = vars_.find(name);
   if (iter != vars_.end()) {
+    VLOG(3) << "Get existing variable " << name;
     return iter->second;
   }
   Variable* v = new Variable();
   vars_[name] = v;
-  VLOG(3) << "Create variable " << name << " on scope";
+  VLOG(3) << "Create variable " << name;
   v->name_ = &(vars_.find(name)->first);
   return v;
 }

paddle/framework/shape_inference.h

@@ -53,6 +53,10 @@ class InferShapeContext {
 
   virtual bool IsRuntime() const = 0;
 
+  // Note: In while op, we need this to be public
+  void SetDims(const std::vector<std::string> &names,
+               const std::vector<framework::DDim> &dims);
+
  protected:
   virtual framework::DDim GetDim(const std::string &name) const = 0;
   virtual void SetDim(const std::string &name, const framework::DDim &dim) = 0;
@@ -60,9 +64,6 @@ class InferShapeContext {
   std::vector<framework::DDim> GetDims(
       const std::vector<std::string> &names) const;
 
-  void SetDims(const std::vector<std::string> &names,
-               const std::vector<framework::DDim> &dims);
-
   std::vector<VarDesc::VarType> GetVarTypes(
       const std::vector<std::string> &names) const;
 

paddle/gserver/CMakeLists.txt

@@ -73,7 +73,6 @@ if(MOBILE_INFERENCE)
     list(REMOVE_ITEM GSERVER_SOURCES
          dataproviders/DataProvider.cpp
          dataproviders/MultiDataProvider.cpp
-         dataproviders/ProtoDataProvider.cpp
          dataproviders/PyDataProvider2.cpp
          dataproviders/PyDataProvider.cpp)
 

paddle/gserver/dataproviders/DataProvider.cpp

@@ -16,8 +16,8 @@ limitations under the License. */
 
 #include <unistd.h>
 #include <algorithm>
-#include "ProtoDataProvider.h"
 #include "paddle/utils/Logging.h"
+#include "paddle/utils/Stat.h"
 #include "paddle/utils/StringUtil.h"
 #include "paddle/utils/Util.h"
 
@@ -164,8 +164,6 @@ DataProvider* DataProvider::create(const DataConfig& config,
 
 REGISTER_DATA_PROVIDER(simple, SimpleDataProvider);
 REGISTER_DATA_PROVIDER(dummy, DummyDataProvider);
-REGISTER_DATA_PROVIDER(proto, ProtoDataProvider);
-REGISTER_DATA_PROVIDER(proto_sequence, ProtoSequenceDataProvider);
 
 int64_t DataProvider::getNextBatch(int64_t size, DataBatch* batch) {
   int64_t batchSize = doubleBuffer_ ? getNextBatchFromBuffer(size, batch)

paddle/gserver/dataproviders/ProtoDataProvider.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
-
-#include <vector>
-
-#include "DataFormat.pb.h"
-#include "paddle/utils/Stat.h"
-
-#include "DataProvider.h"
-#include "ProtoReader.h"
-
-namespace paddle {
-
-/**
- * @brief Provider data from protobuf data file with each sample
- * specified by proto message
- *
- * DataSample defined in DataFormat.proto.
- *
- * The file format is
- *
- *    header
- *
- *    sample1
- *
- *    sample2
- *
- *    ...
- *
- *    sampleN
- *
- * @note: In the data file, each message is prefixed with its length.
- * The read/write of the protbuf are implemented in ProtoReader.h
- */
-class ProtoDataProvider : public DataProvider {
-public:
-  ProtoDataProvider(const DataConfig& config,
-                    bool useGpu,
-                    bool loadDataAll = true);
-  virtual void reset();
-
-  /**
-   * @note this size includes the sequences which are skipped because they
-   * are longer than the batch size.
-   */
-  virtual int64_t getSize() {
-    int64_t size = sampleNums_;
-    if (usageRatio_ < 1.0f) {
-      size = static_cast<int64_t>(size * usageRatio_);
-    }
-    return size;
-  }
-  virtual void shuffle();
-
-  void loadData(const std::vector<std::string>& fileList);
-
-  virtual int64_t getNextBatchInternal(int64_t size, DataBatch* batch);
-
-protected:
-  /**
-   * @brief load protobuf data from a list of file
-   * @param[in]  fileName  file name of a file which contains
-   * a list of file names
-   */
-  void loadData(const std::string& fileName);
-
-  /**
-   * @brief load protobuf data from file
-   * @param[in]  fileName   data file name
-   */
-  void loadDataFile(const std::string& fileName);
-  /** @brief check data header of each data sample
-   *  @param[in] header     data header read from protobuf data
-   */
-  void checkDataHeader(const DataHeader& header);
-  /**
-   * @brief fill protobuf data into slot_,
-   * slot_ is a vector of ProtoSlot in memory.
-   * @param[in]  sample     data sample read from protobuf data
-   */
-  void fillSlots(const DataSample& sample);
-
-  /**
-   * @brief return true if each sample is one sequence, i.e., independent
-   * of other samples.
-   */
-  inline bool iidData() const { return sequenceStartPositions_.empty(); }
-
-  /**
-   * @brief check that sample is consistent with header_
-   */
-  void checkSample(const DataSample& sample);
-
-  template <class Op>
-  int64_t sequenceLoop(Op op, int64_t size);
-
-  template <class Op>
-  int64_t sampleLoop(Op op, int64_t size);
-
-  template <class Op>
-  int64_t subSampleLoop(Op op, int64_t size, int slot);
-
-  void showDataStats();
-
-protected:
-  struct ProtoVarSlot {
-    std::vector<real> data;
-    std::vector<int> dims;
-  };
-
-  struct ProtoSlot {
-    SlotDef::SlotType type;
-    int dim;
-    std::vector<int> indexData;
-    std::vector<real> denseData;
-    std::vector<sparse_non_value_t> sparseNonValueData;
-    std::vector<sparse_float_value_t> sparseFloatValueData;
-    std::vector<int64_t> indices;
-    std::vector<int64_t> subIndices;
-
-    std::vector<ProtoVarSlot> varDenseData;
-    std::vector<std::vector<int>> varIndices;
-    std::vector<std::string> strData;
-  };
-  DataHeader header_;
-  int numVecSlots_;
-
-  std::vector<ProtoSlot> slots_;
-  size_t sampleNums_;
-
-  /**
-   * The starting position of each sequence in samples.
-   * The last element should be num of samples.
-   * If empty, each sample is one sequence.
-   */
-  std::vector<size_t> sequenceStartPositions_;
-
-  int64_t currentSequenceIndex_;
-
-  // The size should be the number of sequences.
-  std::vector<size_t> shuffledSequenceIds_;
-
-  ThreadLocalD<DataBatch> cpuBatch_;
-  ThreadLocalD<DataBatch> gpuBatch_;
-
-  RWLock lock_;
-  std::vector<StatPtr> nnzStats_;  // stats for number of none-zeros entries
-};
-
-/**
- * @brief Special use for Proto data: instances should contain sparse-non-value
- * slots
- * and label.
- *
- * @note ProtoSequenceDataProvider treats each SPARSE SLOT as a SEQUENCE
- */
-class ProtoSequenceDataProvider : public ProtoDataProvider {
-public:
-  ProtoSequenceDataProvider(const DataConfig& config,
-                            bool useGpu,
-                            bool loadDataAll = true);
-  ~ProtoSequenceDataProvider() {}
-  virtual int64_t getNextBatchInternal(int64_t size, DataBatch* batch);
-};
-
-}  // namespace paddle

paddle/gserver/layers/DotProdLayer.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 "Layer.h"
+#include "paddle/math/Matrix.h"
+#include "paddle/utils/Logging.h"
+#include "paddle/utils/Stat.h"
+
+namespace paddle {
+
+/**
+ * @brief A layer for computing the dot product of two vectors.
+ * Input1: vector (batchSize * dim)
+ * Input2: vector (batchSize * dim)
+ * Output: a matrix: (batchSize * 1)
+ */
+
+class DotProdLayer : public Layer {
+public:
+  explicit DotProdLayer(const LayerConfig& config) : Layer(config) {}
+
+  ~DotProdLayer() {}
+
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
+};
+
+REGISTER_LAYER(dot_prod, DotProdLayer);
+
+bool DotProdLayer::init(const LayerMap& layerMap,
+                        const ParameterMap& parameterMap) {
+  Layer::init(layerMap, parameterMap);
+
+  CHECK_EQ(inputLayers_.size(), 2U);
+  CHECK_EQ(1UL, getSize())
+      << "The output dimensionality of this layer should be fixed to 1.";
+
+  return true;
+}
+
+void DotProdLayer::forward(PassType passType) {
+  Layer::forward(passType);
+
+  MatrixPtr inV0 = getInputValue(0);
+  MatrixPtr inV1 = getInputValue(1);
+
+  size_t batchSize = inV0->getHeight();
+  CHECK_EQ(inV1->getHeight(), batchSize);
+  CHECK_EQ(inV0->getWidth(), inV1->getWidth());
+
+  {
+    REGISTER_TIMER_INFO("FwResetTimer", getName().c_str());
+    reserveOutput(batchSize, 1);
+  }
+
+  MatrixPtr outV = getOutputValue();
+  {
+    REGISTER_TIMER_INFO("FwDotProdTimer", getName().c_str());
+    outV->sumOfProducts(*inV0, *inV1, 1, 0);
+  }
+}
+
+void DotProdLayer::backward(const UpdateCallback& callback) {
+  MatrixPtr inV0 = getInputValue(0);
+  MatrixPtr inV1 = getInputValue(1);
+  MatrixPtr outG = getOutputGrad();
+  MatrixPtr inG0 = getInputGrad(0);
+  MatrixPtr inG1 = getInputGrad(1);
+
+  {
+    REGISTER_TIMER_INFO("BwDotProdTimer", getName().c_str());
+
+    if (inG0) {
+      inG0->addRowScale(0, *inV1, *outG);
+    }
+
+    if (inG1) {
+      inG1->addRowScale(0, *inV0, *outG);
+    }
+  }
+}
+
+}  // namespace paddle

paddle/gserver/layers/MKLDNNConcatLayer.cpp

+/* Copyright (c) 2017 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 "MKLDNNConcatLayer.h"
+
+using namespace mkldnn;  // NOLINT
+typedef memory::format format;
+
+namespace paddle {
+
+REGISTER_LAYER(mkldnn_concat, MKLDNNConcatLayer);
+
+bool MKLDNNConcatLayer::init(const LayerMap& layerMap,
+                             const ParameterMap& parameterMap) {
+  if (!MKLDNNLayer::init(layerMap, parameterMap)) {
+    return false;
+  }
+  CHECK_GT(inputLayers_.size(), 1UL);
+  CHECK(!biasParameter_);
+  return true;
+}
+
+void MKLDNNConcatLayer::reshape(
+    int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) {
+  reshapeInput(bs, ih, iw);
+  ic = inputLayers_[0]->getSize() / ih / iw;
+  CHECK_EQ((size_t)ic * ih * iw, inputLayers_[0]->getSize());
+  CHECK_EQ(inputElemenCnt_, (size_t)bs * ic * ih * iw);
+  CHECK_GT(inputLayers_.size(), 1UL);
+  channels_.resize(inputLayers_.size());
+  channels_[0] = ic;
+  // need change the output channel, so use oc_ instead
+  // TODO(TJ): change API, use &oc
+  oc_ = ic;
+  for (size_t i = 1; i < inputLayers_.size(); i++) {
+    int batchsize, height, witdh;
+    reshapeInput(batchsize, height, witdh, i);
+    CHECK_EQ(bs, batchsize);
+    CHECK_EQ(ih, height);
+    CHECK_EQ(iw, witdh);
+
+    channels_[i] = inputLayers_[i]->getSize() / height / witdh;
+    CHECK_EQ((size_t)channels_[i] * height * witdh, inputLayers_[i]->getSize());
+    oc_ += channels_[i];
+  }
+  oh = ih;
+  ow = iw;
+  reshapeOutput(oh, ow);
+  resizeOutput(bs, oc_ * oh * ow);
+}
+
+void MKLDNNConcatLayer::resetFwd(std::vector<primitive>& pipeline,
+                                 MKLDNNMatrixPtr& in,
+                                 MKLDNNMatrixPtr& wgt,
+                                 MKLDNNMatrixPtr& bias,
+                                 MKLDNNMatrixPtr& out) {
+  resetFwdBuffers(inVals_, out);
+  in = inVals_[0];
+
+  std::shared_ptr<concat::primitive_desc> fwdPD;
+  resetFwdPD(fwdPD, inVals_, out);
+
+  resetFwdPipeline(pipeline, fwdPD, inVals_, out);
+}
+
+void MKLDNNConcatLayer::resetBwd(std::vector<primitive>& pipeline,
+                                 MKLDNNMatrixPtr& in,
+                                 MKLDNNMatrixPtr& wgt,
+                                 MKLDNNMatrixPtr& bias,
+                                 MKLDNNMatrixPtr& out) {
+  resetBwdBuffers(inGrads_, out);
+  in = inGrads_[0];
+
+  resetBwdPipeline(pipeline, bwds_, inGrads_, out);
+}
+
+void MKLDNNConcatLayer::resetFwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
+                                        MKLDNNMatrixPtr& out) {
+  inputs.resize(inputLayers_.size());
+  bool has8c = false, has16c = false, hasnc = false;
+  for (size_t i = 0; i < inputs.size(); i++) {
+    // resetInValue will use ic_ so temporary change as current input's channel
+    // TODO(TJ): change ic_ as vector then can remove channels_
+    ic_ = channels_[i];
+    resetInValue(inputs[i], nullptr, i);
+    CHECK(inputs[i]);
+    auto dm = inputs[i]->getDims();
+    // inputs format can be different, but ndims must equal
+    CHECK(i == 0 || dm.size() == inputs[0]->getDims().size());
+    CHECK_EQ(bs_, dm[0]);
+    CHECK_EQ(channels_[i], dm[1]);
+    if (dm.size() > 2) {
+      CHECK_EQ(ih_, dm[2]);
+      CHECK_EQ(iw_, dm[3]);
+    }
+    if (inputs[i]->getFormat() == format::nc) {
+      hasnc = true;
+    }
+    if (inputs[i]->getFormat() == format::nChw8c) {
+      has8c = true;
+    }
+    if (inputs[i]->getFormat() == format::nChw16c) {
+      has16c = true;
+    }
+  }
+  // change back, ic_ always save the input 0 size
+  ic_ = channels_[0];
+
+  format outFmt;
+  if (has16c && oc_ % 16 == 0) {
+    outFmt = format::nChw16c;
+  } else if (has8c && oc_ % 8 == 0) {
+    outFmt = format::nChw8c;
+  } else if (hasnc) {
+    CHECK(oh_ == 1 && ow_ == 1);
+    outFmt = format::nc;
+  } else {
+    outFmt = format::nchw;
+  }
+  memory::dims outDims =
+      hasnc ? memory::dims{bs_, oc_} : memory::dims{bs_, oc_, oh_, ow_};
+  auto outPD = MKLDNNMatrix::createPrimitiveDesc(outDims, outFmt, engine_);
+  resetOutValue(out, outPD);
+}
+
+void MKLDNNConcatLayer::resetFwdPD(std::shared_ptr<concat::primitive_desc>& pd,
+                                   std::vector<MKLDNNMatrixPtr>& inputs,
+                                   MKLDNNMatrixPtr out) {
+  std::vector<memory::primitive_desc> srcPDs;
+  for (size_t i = 0; i < inputs.size(); i++) {
+    srcPDs.push_back(inputs[i]->getPrimitiveDesc());
+  }
+  CHECK(out);
+  pd.reset(new concat::primitive_desc(out->getMemoryDesc(), axis_, srcPDs));
+  CHECK_PRIMITIVE_DESC_EQ(out, pd->dst_primitive_desc());
+}
+
+void MKLDNNConcatLayer::resetFwdPipeline(
+    std::vector<primitive>& pipeline,
+    std::shared_ptr<concat::primitive_desc>& pd,
+    std::vector<MKLDNNMatrixPtr>& inputs,
+    MKLDNNMatrixPtr& out) {
+  std::vector<primitive::at> srcs;
+  for (size_t i = 0; i < inputs.size(); i++) {
+    srcs.push_back(*(inputs[i]));
+  }
+  fwd_.reset(new concat(*pd, srcs, *out));
+  pipeline.push_back(*fwd_);
+}
+
+void MKLDNNConcatLayer::resetBwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
+                                        MKLDNNMatrixPtr& out) {
+  CHECK(outVal_);
+  resetOutGrad(out, outVal_->getPrimitiveDesc());
+  CHECK(out);
+
+  inputs.resize(inputLayers_.size());
+  for (size_t i = 0; i < inputs.size(); i++) {
+    CHECK(inVals_[i]);
+    // resetInGrad will use inVal_
+    // TODO(TJ): change move inVals_ to MKLDNNLayer ans remove inVal_
+    inVal_ = inVals_[i];
+    resetInGrad(inputs[i], inVals_[i]->getPrimitiveDesc(), i);
+    CHECK_PRIMITIVE_DESC_EQ(inputs[i], inVals_[i]->getPrimitiveDesc());
+  }
+  // change back, inVal_ always save the input 0
+  inVal_ = inVals_[0];
+}
+
+void MKLDNNConcatLayer::resetBwdPipeline(
+    std::vector<mkldnn::primitive>& pipeline,
+    std::vector<std::shared_ptr<mkldnn::primitive>>& prims,
+    std::vector<MKLDNNMatrixPtr>& inputs,
+    MKLDNNMatrixPtr& out) {
+  // reset the backward primitives
+  memory::dims offsets = {0, 0, 0, 0};
+  prims.resize(inputs.size());
+  CHECK_EQ(inputs.size(), channels_.size());
+  for (size_t i = 0; i < inputs.size(); i++) {
+    auto viewPD = view::primitive_desc(
+        out->getPrimitiveDesc(), inputs[i]->getDims(), offsets);
+    auto bwdPD = reorder::primitive_desc(viewPD.dst_primitive_desc(),
+                                         inputs[i]->getPrimitiveDesc());
+    prims[i].reset(new reorder(bwdPD, *out, *(inputs[i])));
+    offsets[axis_] += channels_[i];
+    // push to pipeline
+    pipeline.push_back(*prims[i]);
+  }
+}
+
+}  // namespace paddle

paddle/gserver/layers/MKLDNNConcatLayer.h

+/* Copyright (c) 2017 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
+
+#include "MKLDNNLayer.h"
+#include "mkldnn.hpp"
+
+namespace paddle {
+
+/**
+ * @brief A subclass of MKLDNNLayer Concatenate layer.
+ *
+ * The config file api is mkldnn_concat
+ */
+class MKLDNNConcatLayer : public MKLDNNLayer {
+protected:
+  std::vector<MKLDNNMatrixPtr> inVals_;
+  std::vector<MKLDNNMatrixPtr> inGrads_;
+  std::vector<std::shared_ptr<mkldnn::primitive>> bwds_;
+  // input channel numbers
+  std::vector<int> channels_;
+
+  // concat_dimension in MKLDNN
+  // if axis_ == 0, concat batchsize
+  // if axis_ == 1, concat channel (default)
+  int axis_;
+
+public:
+  explicit MKLDNNConcatLayer(const LayerConfig& config)
+      : MKLDNNLayer(config), axis_(1) {}
+
+  ~MKLDNNConcatLayer() {}
+
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+
+  void reshape(
+      int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) override;
+
+  void resetFwd(std::vector<mkldnn::primitive>& pipeline,
+                MKLDNNMatrixPtr& in,
+                MKLDNNMatrixPtr& wgt,
+                MKLDNNMatrixPtr& bias,
+                MKLDNNMatrixPtr& out) override;
+
+  void resetBwd(std::vector<mkldnn::primitive>& pipeline,
+                MKLDNNMatrixPtr& in,
+                MKLDNNMatrixPtr& wgt,
+                MKLDNNMatrixPtr& bias,
+                MKLDNNMatrixPtr& out) override;
+
+  void printSizeInfo() override {
+    CHECK_EQ(channels_.size(), inputLayers_.size());
+    for (size_t i = 0; i < channels_.size(); ++i) {
+      VLOG(MKLDNN_SIZES) << "Input " << i << ", " << inputLayers_[i]->getName()
+                         << ": " << bs_ << ", " << channels_[i] << ", " << ih_
+                         << ", " << iw_;
+    }
+    VLOG(MKLDNN_SIZES) << "Output: " << bs_ << ", " << oc_ << ", " << oh_
+                       << ", " << ow_;
+  }
+
+  void printValueFormat() override {
+    for (size_t i = 0; i < inVals_.size(); ++i) {
+      VLOG(MKLDNN_FMTS) << "Input " << i << ", " << inputLayers_[i]->getName()
+                        << ": " << inVals_[i]->getFormat() << " >>>";
+    }
+    if (outVal_) {
+      VLOG(MKLDNN_FMTS) << outVal_->getFormat() << " >>> ";
+    }
+    if (extOutVal_) {
+      VLOG(MKLDNN_FMTS) << extOutVal_->getFormat();
+    }
+  }
+
+  void printGradFormat() override {
+    if (extOutGrad_) {
+      VLOG(MKLDNN_FMTS) << extOutGrad_->getFormat();
+    }
+    if (outGrad_) {
+      VLOG(MKLDNN_FMTS) << outGrad_->getFormat() << " <<< ";
+    }
+    for (size_t i = 0; i < inGrads_.size(); ++i) {
+      VLOG(MKLDNN_FMTS) << "Input " << i << ", " << inputLayers_[i]->getName()
+                        << ": " << inGrads_[i]->getFormat() << "<<<";
+    }
+  }
+
+protected:
+  /**
+   * Forward functions: reset buffers(inputs, output, bias),
+   *                    reset primitive descriptor,
+   *                    reset pipeline.
+   */
+  void resetFwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
+                       MKLDNNMatrixPtr& out);
+  void resetFwdPD(std::shared_ptr<mkldnn::concat::primitive_desc>& pd,
+                  std::vector<MKLDNNMatrixPtr>& inputs,
+                  MKLDNNMatrixPtr out);
+  void resetFwdPipeline(std::vector<mkldnn::primitive>& pipeline,
+                        std::shared_ptr<mkldnn::concat::primitive_desc>& pd,
+                        std::vector<MKLDNNMatrixPtr>& inputs,
+                        MKLDNNMatrixPtr& out);
+
+  /**
+   * Backward functions: reset buffers(inputs, output, bias)
+   *                     reset primitives and pipeline
+   */
+  void resetBwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
+                       MKLDNNMatrixPtr& out);
+  void resetBwdPipeline(std::vector<mkldnn::primitive>& pipeline,
+                        std::vector<std::shared_ptr<mkldnn::primitive>>& prims,
+                        std::vector<MKLDNNMatrixPtr>& inputs,
+                        MKLDNNMatrixPtr& out);
+};
+
+}  // namespace paddle

paddle/gserver/layers/MKLDNNLayer.cpp

@@ -21,8 +21,8 @@ namespace paddle {
 
 bool MKLDNNLayer::init(const LayerMap& layerMap,
                        const ParameterMap& parameterMap) {
-  CHECK(FLAGS_use_mkldnn) << "MkldnnLayers only support use_mkldnn."
-                          << "Please set WITH_MKLDNN=ON "
+  CHECK(FLAGS_use_mkldnn) << "MKLDNNLayers only support use_mkldnn."
+                          << "Please set WITH_MKL=ON "
                           << "and set use_mkldnn=True";
   CHECK(!useGpu_) << "Do not support GPU yet";
 
@@ -138,8 +138,11 @@ void MKLDNNLayer::backward(const UpdateCallback& callback) {
   }
 }
 
-void MKLDNNLayer::reshapeInput(int& batchsize, int& height, int& width) {
-  const Argument& input = inputLayers_[0]->getOutput();
+void MKLDNNLayer::reshapeInput(int& batchsize,
+                               int& height,
+                               int& width,
+                               size_t inputIdx) {
+  const Argument& input = inputLayers_[inputIdx]->getOutput();
   batchsize = input.getBatchSize();
   int h = input.getFrameHeight();
   int w = input.getFrameWidth();

paddle/gserver/layers/MKLDNNLayer.h

@@ -178,7 +178,10 @@ protected:
   /**
    * reshape the input image sizes and input batchsize
    */
-  void reshapeInput(int& batchsize, int& height, int& width);
+  void reshapeInput(int& batchsize,
+                    int& height,
+                    int& width,
+                    size_t inputIdx = 0);
 
   /**
    * reshape output image sizes

paddle/gserver/layers/ROIPoolLayer.cpp

@@ -100,8 +100,9 @@ void ROIPoolLayer::forward(PassType passType) {
     size_t roiEndH = round(bottomROIs[4] * spatialScale_);
     CHECK_GE(roiBatchIdx, 0UL);
     CHECK_LT(roiBatchIdx, batchSize);
-    size_t roiHeight = std::max(roiEndH - roiStartH + 1, 1UL);
-    size_t roiWidth = std::max(roiEndW - roiStartW + 1, 1UL);
+    size_t roiHeight =
+        std::max(roiEndH - roiStartH + 1, static_cast<size_t>(1));
+    size_t roiWidth = std::max(roiEndW - roiStartW + 1, static_cast<size_t>(1));
     real binSizeH =
         static_cast<real>(roiHeight) / static_cast<real>(pooledHeight_);
     real binSizeW =
@@ -114,10 +115,14 @@ void ROIPoolLayer::forward(PassType passType) {
           size_t wstart = static_cast<size_t>(std::floor(pw * binSizeW));
           size_t hend = static_cast<size_t>(std::ceil((ph + 1) * binSizeH));
           size_t wend = static_cast<size_t>(std::ceil((pw + 1) * binSizeW));
-          hstart = std::min(std::max(hstart + roiStartH, 0UL), height_);
-          wstart = std::min(std::max(wstart + roiStartW, 0UL), width_);
-          hend = std::min(std::max(hend + roiStartH, 0UL), height_);
-          wend = std::min(std::max(wend + roiStartW, 0UL), width_);
+          hstart = std::min(
+              std::max(hstart + roiStartH, static_cast<size_t>(0)), height_);
+          wstart = std::min(
+              std::max(wstart + roiStartW, static_cast<size_t>(0)), width_);
+          hend = std::min(std::max(hend + roiStartH, static_cast<size_t>(0)),
+                          height_);
+          wend = std::min(std::max(wend + roiStartW, static_cast<size_t>(0)),
+                          width_);
 
           bool isEmpty = (hend <= hstart) || (wend <= wstart);
           size_t poolIndex = ph * pooledWidth_ + pw;

paddle/gserver/tests/CMakeLists.txt

@@ -29,7 +29,7 @@ gserver_test(test_KmaxSeqScore)
 gserver_test(test_Expand)
 gserver_test(test_MaxPoolingWithMaskOutput)
 
-########## test_Mkldnn layers and activations ##########
+########## test_MKLDNN layers and activations ##########
 if(WITH_MKLDNN)
     add_unittest_without_exec(test_MKLDNN
         test_MKLDNN.cpp
@@ -62,17 +62,6 @@ if(NOT WITH_DOUBLE AND NOT MOBILE_INFERENCE)
 endif()
 
 if(NOT MOBILE_INFERENCE)
-################### test_ProtoDataProvider ############
-    add_unittest_without_exec(test_ProtoDataProvider
-        test_ProtoDataProvider.cpp)
-
-    # test_ProtoDataProvider will mkdir as same name,
-    # so if WORKING_DIRECTORY is default directory, then
-    # mkdir will get error.
-    add_test(NAME test_ProtoDataProvider
-        COMMAND ${CMAKE_CURRENT_BINARY_DIR}/test_ProtoDataProvider
-        WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle)
-
 ################## test_Evaluator #######################
     add_unittest(test_Evaluator
         test_Evaluator.cpp)
@@ -110,3 +99,24 @@ add_test(NAME test_PyDataProvider2
    COMMAND .set_python_path.sh -d ${PADDLE_SOURCE_DIR}/paddle/gserver/tests:${PADDLE_SOURCE_DIR}/python ${CMAKE_CURRENT_BINARY_DIR}/test_PyDataProvider2
         WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle
 )
+
+################# test_CompareSparse ##################
+add_unittest_without_exec(test_CompareSparse
+    test_CompareSparse.cpp)
+if(NOT ON_TRAVIS)
+  add_test(NAME test_CompareSparse
+    COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d
+          ${PADDLE_SOURCE_DIR}/python:${PADDLE_SOURCE_DIR}/paddle/gserver/tests
+              ./.set_port.sh -p port -n 6
+                  ${CMAKE_CURRENT_BINARY_DIR}/test_CompareSparse
+    WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/)
+endif()
+
+################ test_CompareTwoNets ######################
+add_unittest_without_exec(test_CompareTwoNets
+    test_CompareTwoNets.cpp)
+add_test(NAME test_CompareTwoNets
+  COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d
+        ${PADDLE_SOURCE_DIR}/python:${PADDLE_SOURCE_DIR}/paddle/gserver/tests
+        ${CMAKE_CURRENT_BINARY_DIR}/test_CompareTwoNets
+    WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/)

paddle/gserver/tests/MKLDNNTester.h

@@ -23,7 +23,7 @@ limitations under the License. */
 namespace paddle {
 
 /**
- * @brief test the functionality of Mkldnnlayers
+ * @brief test the functionality of MKLDNNlayers and MKLDNNActivations
  * refer to paddle original function
  */
 class MKLDNNTester {

paddle/gserver/tests/proto_files.txt

-./test_ProtoDataProvider/data1.bin
-./test_ProtoDataProvider/data2.bin

paddle/gserver/tests/proto_files_compressed.txt

-./test_ProtoDataProvider/data1.bin.gz
-./test_ProtoDataProvider/data2.bin.gz

paddle/trainer/tests/sample_trainer_config_opt_b.conf → paddle/gserver/tests/sequence_lstm.conf

+#!/usr/bin/env python
 # Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -14,27 +15,50 @@
 
 from paddle.trainer_config_helpers import *
 
-################################### Data Configuration ###################################
-TrainData(ProtoData(files = "trainer/tests/mnist.list"))
-################################### Algorithm Configuration ###################################
-settings(batch_size = 1000,
-         learning_method = MomentumOptimizer(momentum=0.5, sparse=False))
-################################### Network Configuration ###################################
-data = data_layer(name ="input", size=784)
+######################## data source ################################
+dict_path = 'gserver/tests/Sequence/tour_dict_phrase.dict'
+dict_file = dict()
+for line_count, line in enumerate(open(dict_path, "r")):
+    dict_file[line.strip()] = line_count
 
-fc1 = fc_layer(input=data, size=800,
-               bias_attr=True,
-               act=SigmoidActivation())
+define_py_data_sources2(
+    train_list='gserver/tests/Sequence/train.list',
+    test_list=None,
+    module='sequenceGen',
+    obj='process',
+    args={"dict_file": dict_file})
 
-fc2 = fc_layer(input=fc1, size=800,
-               bias_attr=True,
-               act=SigmoidActivation())
+settings(batch_size=5)
+######################## network configure ################################
+dict_dim = len(open(dict_path, 'r').readlines())
+word_dim = 128
+hidden_dim = 256
+label_dim = 3
+sparse_update = get_config_arg("sparse_update", bool, False)
 
-output = fc_layer(input=[fc1, fc2], size=10,
-                  bias_attr=True,
-                  act=SoftmaxActivation())
+data = data_layer(name="word", size=dict_dim)
 
-lbl = data_layer(name ="label", size=1)
+emb = embedding_layer(
+    input=data,
+    size=word_dim,
+    param_attr=ParamAttr(sparse_update=sparse_update))
 
-cost = classification_cost(input=output, label=lbl)
-outputs(cost)
+with mixed_layer(size=hidden_dim * 4) as lstm_input:
+    lstm_input += full_matrix_projection(input=emb)
+
+lstm = lstmemory(
+    input=lstm_input,
+    act=TanhActivation(),
+    gate_act=SigmoidActivation(),
+    state_act=TanhActivation())
+
+lstm_last = last_seq(input=lstm)
+
+with mixed_layer(
+        size=label_dim, act=SoftmaxActivation(), bias_attr=True) as output:
+    output += full_matrix_projection(input=lstm_last)
+
+outputs(
+    classification_cost(
+        input=output, label=data_layer(
+            name="label", size=1)))

paddle/trainer/tests/sample_trainer_config_opt_a.conf → paddle/gserver/tests/sequence_recurrent.py

+#!/usr/bin/env python
 # Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -14,27 +15,42 @@
 
 from paddle.trainer_config_helpers import *
 
-################################### Data Configuration ###################################
-TrainData(ProtoData(files = "trainer/tests/mnist.list"))
-################################### Algorithm Configuration ###################################
-settings(batch_size = 1000,
-         learning_method = MomentumOptimizer(momentum=0.5, sparse=False))
-################################### Network Configuration ###################################
-data = data_layer(name ="input", size=784)
+######################## data source ################################
+dict_path = 'gserver/tests/Sequence/tour_dict_phrase.dict'
+dict_file = dict()
+for line_count, line in enumerate(open(dict_path, "r")):
+    dict_file[line.strip()] = line_count
 
-fc1 = fc_layer(input=data, size=800,
-               bias_attr=True,
-               act=SigmoidActivation())
+define_py_data_sources2(
+    train_list='gserver/tests/Sequence/train.list',
+    test_list=None,
+    module='sequenceGen',
+    obj='process',
+    args={"dict_file": dict_file})
 
-fc2 = fc_layer(input=fc1, size=800,
-               bias_attr=True,
-               act=SigmoidActivation())
+settings(batch_size=5)
+######################## network configure ################################
+dict_dim = len(open(dict_path, 'r').readlines())
+word_dim = 128
+hidden_dim = 128
+label_dim = 3
 
-output = fc_layer(input=[fc1, fc2], size=10,
-                  bias_attr=True,
-                  act=SoftmaxActivation())
+# This config is designed to be equivalent with sequence_recurrent_group.py
 
-lbl = data_layer(name ="label", size=1)
+data = data_layer(name="word", size=dict_dim)
 
-cost = classification_cost(input=output, label=lbl)
-outputs(cost)
+emb = embedding_layer(
+    input=data, size=word_dim, param_attr=ParamAttr(name="emb"))
+
+recurrent = recurrent_layer(input=emb, bias_attr=False, act=SoftmaxActivation())
+
+recurrent_last = last_seq(input=recurrent)
+
+with mixed_layer(
+        size=label_dim, act=SoftmaxActivation(), bias_attr=True) as output:
+    output += full_matrix_projection(input=recurrent_last)
+
+outputs(
+    classification_cost(
+        input=output, label=data_layer(
+            name="label", size=1)))

paddle/gserver/tests/sequence_recurrent_group.py

+#!/usr/bin/env python
+# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from paddle.trainer_config_helpers import *
+
+######################## data source ################################
+dict_path = 'gserver/tests/Sequence/tour_dict_phrase.dict'
+dict_file = dict()
+for line_count, line in enumerate(open(dict_path, "r")):
+    dict_file[line.strip()] = line_count
+
+define_py_data_sources2(
+    train_list='gserver/tests/Sequence/train.list',
+    test_list=None,
+    module='sequenceGen',
+    obj='process',
+    args={"dict_file": dict_file})
+
+settings(batch_size=5)
+######################## network configure ################################
+dict_dim = len(open(dict_path, 'r').readlines())
+word_dim = 128
+hidden_dim = 128
+label_dim = 3
+
+# This config is designed to be equivalent with sequence_recurrent.py
+
+data = data_layer(name="word", size=dict_dim)
+
+emb = embedding_layer(
+    input=data, size=word_dim, param_attr=ParamAttr(name="emb"))
+
+
+def step(y):
+    mem = memory(name="rnn_state", size=hidden_dim)
+    with mixed_layer(
+            name="rnn_state",
+            size=hidden_dim,
+            bias_attr=False,
+            act=SoftmaxActivation()) as out:
+        out += identity_projection(input=y)
+        out += full_matrix_projection(
+            input=mem, param_attr=ParamAttr(name="___recurrent_layer_0__"))
+    return out
+
+
+recurrent = recurrent_group(name="rnn", step=step, input=emb)
+
+recurrent_last = last_seq(input=recurrent)
+
+with mixed_layer(
+        size=label_dim, act=SoftmaxActivation(), bias_attr=True) as output:
+    output += full_matrix_projection(input=recurrent_last)
+
+outputs(
+    classification_cost(
+        input=output, label=data_layer(
+            name="label", size=1)))

paddle/trainer/tests/test_CompareSparse.cpp → paddle/gserver/tests/test_CompareSparse.cpp

@@ -22,8 +22,7 @@ limitations under the License. */
 using namespace paddle;  // NOLINT
 using namespace std;     // NOLINT
 
-static const string& configFile1 =
-    "trainer/tests/sample_trainer_config_compare_sparse.conf";
+static const string& configFile1 = "gserver/tests/sequence_lstm.conf";
 
 DECLARE_bool(use_gpu);
 DECLARE_string(config);

paddle/trainer/tests/test_CompareTwoNets.cpp → paddle/gserver/tests/test_CompareTwoNets.cpp

@@ -30,8 +30,6 @@ DECLARE_bool(use_gpu);
 DECLARE_string(config);
 DECLARE_string(nics);
 
-DEFINE_string(config_file_a, "", "config of one network to compare");
-DEFINE_string(config_file_b, "", "config of another network to compare");
 DEFINE_bool(need_high_accuracy,
             false,
             "whether need to run in double accuracy");
@@ -42,6 +40,10 @@ DEFINE_double(
 DECLARE_bool(thread_local_rand_use_global_seed);
 DECLARE_int32(seed);
 
+static const string& config_file_a = "gserver/tests/sequence_recurrent.py";
+static const string& config_file_b =
+    "gserver/tests/sequence_recurrent_group.py";
+
 struct ComData {
   vector<Argument> outArgs;
   vector<ParameterPtr> parameters;
@@ -66,6 +68,7 @@ void calcGradient(ComData& data, const string configFile) {
   DataBatch dataBatch;
   int32_t batchSize = trainer.getConfig().opt_config().batch_size();
 
+  trainer.getDataProvider()->reset();
   trainer.getDataProvider()->setSkipShuffle();
   trainer.getDataProvider()->getNextBatch(batchSize, &dataBatch);
 
@@ -167,11 +170,11 @@ void compareGradient(ComData& comDataA, ComData& comDataB) {
 
 TEST(Trainer, create) {
   ComData dataA;
-  calcGradient(dataA, FLAGS_config_file_a);
+  calcGradient(dataA, config_file_a);
   LOG(INFO) << "\n\nforwardBackward of Network A is finished\n\n";
 
   ComData dataB;
-  calcGradient(dataB, FLAGS_config_file_b);
+  calcGradient(dataB, config_file_b);
   LOG(INFO) << "\n\nforwardBackward of the Network B is finished\n\n";
 
   compareGradient(dataA, dataB);

paddle/gserver/tests/test_LayerGrad.cpp

@@ -1081,6 +1081,21 @@ TEST(Layer, InterpolationLayer) {
   }
 }
 
+TEST(Layer, DotProdLayer) {
+  TestConfig config;
+  config.layerConfig.set_type("dot_prod");
+  config.layerConfig.set_size(1);
+
+  config.inputDefs.push_back({INPUT_DATA, "layer_0", 10, 0});
+  config.layerConfig.add_inputs();
+  config.inputDefs.push_back({INPUT_DATA, "layer_1", 10, 0});
+  config.layerConfig.add_inputs();
+
+  for (auto useGpu : {false, true}) {
+    testLayerGrad(config, "dot_prod", 10, false, useGpu);
+  }
+}
+
 TEST(Layer, OuterProdLayer) {
   TestConfig config;
   config.layerConfig.set_type("out_prod");

paddle/gserver/tests/test_MKLDNN.cpp

@@ -313,6 +313,47 @@ TEST(MKLDNNLayer, AddtoLayer) {
   testAddtoLayer({4, 12, 1, 1}, 3);
 }
 
+static void getMKLDNNConcatConfig(TestConfig& cfg,
+                                  const std::vector<testImageDesc>& inputs) {
+  CHECK_GE(inputs.size(), 2) << "at least two inputs";
+  int oc = inputs[0].ic;
+  for (size_t i = 1; i < inputs.size(); ++i) {
+    CHECK_EQ(inputs[i].bs, inputs[0].bs);
+    CHECK_EQ(inputs[i].ih, inputs[0].ih);
+    CHECK_EQ(inputs[i].iw, inputs[0].iw);
+    oc += inputs[i].ic;
+  }
+  cfg.biasSize = 0;
+  cfg.layerConfig.set_type("mkldnn_concat");
+  cfg.layerConfig.set_size(oc * inputs[0].ih * inputs[0].iw);
+  cfg.layerConfig.set_active_type("relu");
+  for (size_t i = 0; i < inputs.size(); ++i) {
+    std::stringstream ss;
+    ss << "layer_" << i;
+    cfg.inputDefs.push_back(
+        {INPUT_DATA,
+         ss.str(),
+         (size_t)(inputs[i].ic) * inputs[i].ih * inputs[i].iw,
+         0});
+    LayerInputConfig* input = cfg.layerConfig.add_inputs();
+    ImageConfig* img_conf = input->mutable_image_conf();
+    img_conf->set_channels(inputs[i].ic);
+    img_conf->set_img_size_y(inputs[i].ih);
+    img_conf->set_img_size(inputs[i].iw);
+  }
+}
+
+void testConcatLayer(const std::vector<testImageDesc>& inputs) {
+  TestConfig dnnConfig;
+  getMKLDNNConcatConfig(dnnConfig, inputs);
+  RUN_MKLDNN_TEST_LAYER(dnnConfig, "concat", inputs[0])
+}
+
+TEST(MKLDNNLayer, ConcatLayer) {
+  testConcatLayer({{64, 128, 1, 1}, {64, 32, 1, 1}, {64, 64, 1, 1}});
+  testConcatLayer({{32, 100, 8, 8}, {32, 10, 8, 8}});
+}
+
 void testActivation(std::string actType, const testImageDesc& pm) {
   // TODO(TJ): remove me when paddle support elu activation
   if (actType == "mkldnn_elu") {

paddle/math/Storage.cpp

@@ -17,9 +17,13 @@ limitations under the License. */
 #include "paddle/utils/StringUtil.h"
 #include "paddle/utils/Util.h"
 
+#ifndef PADDLE_MOBILE_INFERENCE
 DEFINE_int32(pool_limit_size,
              536870912,
              "maximum memory size managed by a memory pool, default is 512M");
+#else
+DEFINE_int32(pool_limit_size, 0, "default is 0");
+#endif
 
 namespace paddle {
 

paddle/memory/README.md

 # Region-based Heterogeneous Memory Management
+## Design
 
-Please check out the [design documentation](http://gangliao.me) to find out more details about
-buddy memory allocator for both CPU and GPU.
+### Usage
+
+To allocate 4KB CPU memory:
+
+```cpp
+p = memory::Alloc(platform::CPUPlace(), 4*1024);
+```
+
+To allocate 4KB memory on the 3rd GPU:
+
+```cpp
+p = memory::Alloc(platform::GPUPlace(2), 4*1024);
+```
+
+To free memory and check the so-far used amount of memory on a place:
+
+```cpp
+auto pl = platform::GPUPlace(0);
+p = memory::Alloc(pl, 4*1024);
+cout << memory::Used(pl);
+memory::Free(pl, p);
+```
+
+### API
+
+In `paddle/memory/memory.h` we have:
+
+```cpp
+namespace memory {
+template <typename Place> void* Alloc(Place, size_t);
+template <typename Place> void Free(Place, void*);
+template <typename Place> size_t Used(Place);
+}  // namespace memory
+```
+
+These function templates have specializations on either `platform::CPUPlace` or `platform::GPUPlace`:
+
+```cpp
+template<>
+void* Alloc<CPUPlace>(CPUPlace p, size_t size) {
+  return GetCPUBuddyAllocator()->Alloc(size);
+}
+```
+
+and 
+
+```cpp
+template<>
+void Alloc<GPUPlace>(GPUPlace p, size_t size) {
+  return GetGPUBuddyAllocator(p.id)->Alloc(size);
+}
+```
+
+Similar specializations exist for `Free` and `Used`.
+
+### Implementation
+
+`GetCPUBuddyAllocator` and `GetGPUBuddyAllocator` are singletions.
+
+```cpp
+BuddyAllocator* GetCPUBuddyAllocator() {
+  static BuddyAllocator* a = NULL;
+  if (a == NULL) {
+    a = new BuddyAllocator(new CPUAllocator /*backup allocator*/, ...);
+  }
+  return a;
+}
+
+BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) {
+  static BuddyAllocator* as = NULL;
+  if (as == NULL) {
+    as = new BuddyAllocator*[platform::NumGPUs()];
+    for (int gpu = 0; gpu < platform::NumGPUs(); gpu++) {
+      as[gpu] = new BuddyAllocator(new GPUAllocator(gpu) /* backup allocator */, ...);
+    }
+  }
+  return as[gpu_id);
+```
+
+#### `BuddyAllocator`
+
+`BuddyAllocator` implements the buddy allocation algorithm.  Its constructor takes parameters only related with the algorithm:
+
+```cpp
+BuddyAllocator::BuddyAllocator(initial_pool_size, max_pool_size) {
+  ...
+}
+```
+
+Please be aware that **`BuddyAllocator` always allocate aligned memory**, aligned on 32-bytes, which can hold a `BuddyAllocator::Block` object:
+
+```cpp
+class BuddyAllocator {
+ private:
+  struct Block {
+    size_t size;
+    Block* left, right;
+    size_t index; // allocator id
+  };
+  ...
+};
+```
+
+Because BuddyAllocator has the meta-data of each block, it can trace the used memory -- record the amount returned by `Alloc` freed in `Free`.  Instead, `CPUAllocator` and `GPUAllocator` doesn't know the size of freed memory block and cannot do the trace.
+
+#### System Allocators
+
+The `GPUAllocator` and `CPUAllocator` are calls *system allocators*.  They work as the fallback allocators of `BuddyAllocator`.
+
+## Justification
+
+I got inspiration from Majel and Caffe2, though above design look different from both.
+
+### Caffe2
+
+In Caffe2, `Tensor<Context>::mutable_data()` allocates the memroy.  In particular, [`Tensor<Context>::mutable_data`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/tensor.h#L523) calls [`Tensor<Context>::raw_mutable_data`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/tensor.h#L459), which in turn calls [`Context::New`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/tensor.h#L479).
+
+There are two implementations of `Context`:
+
+1. [`CPUContext`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context.h#L105), whose [`New` method](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context.h#L131) calls [`g_cpu_allocator.get()->New(size_t)`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context.cc#L15) to allocate the memory.
+
+1. [`CUDAContext`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context_gpu.h#L99), which has a data member [`int gpu_id_`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context_gpu.h#L202).  This looks very similar to class `majel::GPUPlace`, who also has an `int id_` data member.   `CUDAContext::New(size_t)` calls [`g_cub_allocator->DeviceAllocate(&ptr, nbytes)`](https://github.com/caffe2/caffe2/blob/v0.7.0/caffe2/core/context_gpu.cu#L355) to allocate the memory.
+
+### Majel
+
+In Majel, there are basically two allocator types:
+
+1. `cpu::SystemAllocator`, which has similar functionality to `caffe2::CPUContext::New/Delete`.
+1. `gpu::SystemAllocator`, which has similar functionality to `caffe2::CUDAContext::New/Delete`.
+
+However, memory allocation is not via these two allocators.  Instead, these two allocators are defined in hidden namespaces.
+
+In Majel there are hidden global variables like:
+
+1. `cpu::SystemAllocator g_cpu_allocator`, and
+1. `vector<gpu::SystemAllocator*> g_gpu_allocators(NUM_GPUS)`.
+
+Programs allocate memory via a BuddyAllocator, which can take the `g_cpu_allocator` or a `g_gpu_allocators[gpu_id]` as its *fallback allocator*, so that if BuddyAllocator cannot find a block in its memory pool, it extends its memory pool by calling the fallback allocator's `New(size_t)`.

paddle/operators/CMakeLists.txt

@@ -9,6 +9,7 @@ function(op_library TARGET)
     set(OP_LIBRARY ${TARGET} ${OP_LIBRARY} PARENT_SCOPE)
     set(cc_srcs)
     set(cu_srcs)
+    set(cu_cc_srcs)
     set(op_common_deps operator op_registry math_function)
     set(options "")
     set(oneValueArgs "")
@@ -22,6 +23,9 @@ function(op_library TARGET)
         if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cc)
             list(APPEND cc_srcs ${TARGET}.cc)
         endif()
+        if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cu.cc)
+            list(APPEND cu_cc_srcs ${TARGET}.cu.cc)
+        endif()
         if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cu)
             list(APPEND cu_srcs ${TARGET}.cu)
         endif()
@@ -29,6 +33,8 @@ function(op_library TARGET)
         foreach(src ${op_library_SRCS})
             if (${src} MATCHES ".*\\.cu$")
                 list(APPEND cu_srcs ${src})
+            elseif(${src} MATCHES ".*\\.cu.cc$")
+                list(APPEND cu_cc_srcs ${src})
             elseif(${src} MATCHES ".*\\.cc$")
                 list(APPEND cc_srcs ${src})
             else()
@@ -43,7 +49,7 @@ function(op_library TARGET)
     endif()
 
     if (WITH_GPU)
-        nv_library(${TARGET} SRCS ${cc_srcs} ${cu_srcs} DEPS ${op_library_DEPS}
+        nv_library(${TARGET} SRCS ${cc_srcs} ${cu_cc_srcs} ${cu_srcs} DEPS ${op_library_DEPS}
                 ${op_common_deps})
     else()
         cc_library(${TARGET} SRCS ${cc_srcs} DEPS ${op_library_DEPS}
@@ -55,6 +61,18 @@ function(op_library TARGET)
         set(pybind_flag 1)
     endif()
 
+    if ("${TARGET}" STREQUAL "compare_op")
+        set(pybind_flag 1)
+        file(APPEND ${pybind_file} "USE_OP(less_than);\nUSE_OP(equal);\n")
+    endif()
+
+    # conv_op contains several operators
+    if ("${TARGET}" STREQUAL "conv_op")
+        set(pybind_flag 1)
+        # It's enough to just adding one operator to pybind
+        file(APPEND ${pybind_file} "USE_OP(conv2d);\n")
+    endif()
+
     # pool_op contains several operators
     if ("${TARGET}" STREQUAL "pool_op")
         set(pybind_flag 1)
@@ -62,23 +80,23 @@ function(op_library TARGET)
         file(APPEND ${pybind_file} "USE_OP(pool2d);\n")
     endif()
 
-    if ("${TARGET}" STREQUAL "compare_op")
+    # pool_cudnn_op contains several operators
+    if ("${TARGET}" STREQUAL "pool_cudnn_op")
         set(pybind_flag 1)
-        file(APPEND ${pybind_file} "USE_OP(less_than);\nUSE_OP(equal);\n")
+        # It's enough to just adding one operator to pybind
+        file(APPEND ${pybind_file} "USE_OP(pool2d_cudnn);\n")
     endif()
 
-    # pool_with_index_op contains several operators
-    if ("${TARGET}" STREQUAL "pool_with_index_op")
+    if ("${TARGET}" STREQUAL "logical_op")
         set(pybind_flag 1)
-        # It's enough to just adding one operator to pybind
-        file(APPEND ${pybind_file} "USE_OP(max_pool2d_with_index);\n")
+        file(APPEND ${pybind_file} "USE_OP(logical_and);\n")
     endif()
 
-    # conv_op contains several operators
-    if ("${TARGET}" STREQUAL "conv_op")
+    # pool_with_index_op contains several operators
+    if ("${TARGET}" STREQUAL "pool_with_index_op")
         set(pybind_flag 1)
         # It's enough to just adding one operator to pybind
-        file(APPEND ${pybind_file} "USE_OP(conv2d);\n")
+        file(APPEND ${pybind_file} "USE_OP(max_pool2d_with_index);\n")
     endif()
 
     # conv_transpose_op contains several operators
@@ -87,12 +105,12 @@ function(op_library TARGET)
         # It's enough to just adding one operator to pybind
         file(APPEND ${pybind_file} "USE_OP(conv2d_transpose);\n")
     endif()
-    
-    # pool_cudnn_op contains several operators
-    if ("${TARGET}" STREQUAL "pool_cudnn_op")
+
+    # conv_transpose_cudnn_op contains two operators
+    if ("${TARGET}" STREQUAL "conv_transpose_cudnn_op")
         set(pybind_flag 1)
         # It's enough to just adding one operator to pybind
-        file(APPEND ${pybind_file} "USE_OP(pool2d_cudnn);\n")
+        file(APPEND ${pybind_file} "USE_OP(conv2d_transpose_cudnn);\n")
     endif()
 
     # save_restore_op contains several operators
@@ -140,7 +158,9 @@ function(op_library TARGET)
 
     # pybind USE_CPU_ONLY_OP
     list(LENGTH cu_srcs cu_srcs_len)
-    if (${pybind_flag} EQUAL 0 AND ${cu_srcs_len} EQUAL 0)
+    list(LENGTH cu_cc_srcs cu_cc_srcs_len)
+
+    if (${pybind_flag} EQUAL 0 AND ${cu_srcs_len} EQUAL 0 AND ${cu_cc_srcs_len} EQUAL 0)
         file(APPEND ${pybind_file} "USE_CPU_ONLY_OP(${TARGET});\n")
         set(pybind_flag 1)
     endif()
@@ -160,11 +180,12 @@ set(DEPS_OPS
     recurrent_op
     dynamic_recurrent_op
     softmax_with_cross_entropy_op
+    softmax_op
+    sequence_softmax_op
     sum_op
     pool_op
     pool_with_index_op
     conv_op
-    lstm_op
     conv_transpose_op
     nccl_op
     sequence_conv_op
@@ -174,13 +195,20 @@ set(DEPS_OPS
     array_to_lod_tensor_op
     lstm_op
     tensor_array_read_write_op
-    gru_op)
+    gru_op
+    adagrad_op
+    sgd_op)
+
 
 op_library(cond_op SRCS cond_op.cc DEPS framework_proto tensor operator net_op)
 op_library(cross_entropy_op DEPS cross_entropy)
 op_library(softmax_with_cross_entropy_op DEPS cross_entropy softmax)
+op_library(softmax_op DEPS softmax)
+op_library(sequence_softmax_op DEPS softmax)
+op_library(sum_op DEPS selected_rows_functor)
+op_library(sgd_op DEPS selected_rows_functor)
+op_library(adagrad_op DEPS selected_rows_functor)
 op_library(conv_op DEPS vol2col)
-op_library(sum_op DEPS net_op selected_rows_functor)
 op_library(pool_op DEPS pooling)
 op_library(pool_with_index_op DEPS pooling)
 op_library(lod_rank_table_op SRCS lod_rank_table_op.cc DEPS lod_rank_table)
@@ -220,6 +248,6 @@ cc_test(dynamic_recurrent_op_test SRCS dynamic_recurrent_op_test.cc
         rnn/recurrent_op_utils.cc
         DEPS dynamic_recurrent_op)
 if(WITH_GPU)
-  nv_test(nccl_op_test SRCS nccl_op_test.cu DEPS nccl_op gpu_info device_context)
+  cc_test(nccl_op_test SRCS nccl_op_test.cu.cc DEPS nccl_op gpu_info device_context)
 endif()
 cc_test(save_load_op_test SRCS save_load_op_test.cc DEPS save_op load_op)

paddle/operators/accuracy_op.cu

@@ -16,6 +16,7 @@ limitations under the License. */
 #include <thrust/reduce.h>
 #include "paddle/operators/accuracy_op.h"
 #include "paddle/platform/cuda_helper.h"
+#include "paddle/platform/gpu_info.h"
 
 namespace paddle {
 namespace operators {
@@ -73,26 +74,28 @@ class AccuracyOpCUDAKernel : public framework::OpKernel<T> {
 
     int num_samples = static_cast<int>(inference->dims()[0]);
     size_t infer_width = inference->dims()[1];
-    PADDLE_ENFORCE(cudaMemset(accuracy_data, 0, sizeof(float)));
-    // cudaMemset((void**)&correct_data, 0, sizeof(float));
+    auto stream = ctx.cuda_device_context().stream();
+    platform::GpuMemsetAsync(accuracy_data, 0, sizeof(float), stream);
 
     if (num_samples == 0) {
       return;
     }
-    cudaMemcpy(total_data, &num_samples, sizeof(int), cudaMemcpyHostToDevice);
+    platform::GpuMemcpyAsync(total_data, &num_samples, sizeof(int),
+                             cudaMemcpyHostToDevice, stream);
 
-    AccuracyCudaKernel<PADDLE_CUDA_NUM_THREADS><<<
-        1, PADDLE_CUDA_NUM_THREADS, 0, ctx.cuda_device_context().stream()>>>(
+    AccuracyCudaKernel<
+        PADDLE_CUDA_NUM_THREADS><<<1, PADDLE_CUDA_NUM_THREADS, 0, stream>>>(
         num_samples, infer_width, indices_data, label_data, correct_data,
         accuracy_data);
 
     int d_num_samples, d_num_correct;
     float d_accuracy;
-    cudaMemcpy(&d_num_correct, correct_data, sizeof(int),
-               cudaMemcpyDeviceToHost);
-    cudaMemcpy(&d_num_samples, total_data, sizeof(int), cudaMemcpyDeviceToHost);
-    cudaMemcpy(&d_accuracy, accuracy_data, sizeof(float),
-               cudaMemcpyDeviceToHost);
+    platform::GpuMemcpyAsync(&d_num_correct, correct_data, sizeof(int),
+                             cudaMemcpyDeviceToHost, stream);
+    platform::GpuMemcpyAsync(&d_num_samples, total_data, sizeof(int),
+                             cudaMemcpyDeviceToHost, stream);
+    platform::GpuMemcpyAsync(&d_accuracy, accuracy_data, sizeof(float),
+                             cudaMemcpyDeviceToHost, stream);
   }
 };
 

paddle/operators/adagrad_op.cc

@@ -14,6 +14,11 @@ limitations under the License. */
 
 #include "paddle/operators/adagrad_op.h"
 
+#include <cmath>
+
+#include "paddle/operators/math/math_function.h"
+#include "paddle/operators/math/selected_rows_functor.h"
+
 namespace paddle {
 namespace operators {
 
@@ -21,7 +26,7 @@ class AdagradOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
 
-  void InferShape(framework::InferShapeContext *ctx) const override {
+  void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("Param"),
                    "Input(Param) of AdagradOp should not be null.");
     PADDLE_ENFORCE(ctx->HasInput("Grad"),
@@ -54,8 +59,8 @@ class AdagradOp : public framework::OperatorWithKernel {
 
 class AdagradOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  AdagradOpMaker(framework::OpProto *proto,
-                 framework::OpAttrChecker *op_checker)
+  AdagradOpMaker(framework::OpProto* proto,
+                 framework::OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("Param", "(Tensor) Input parameter");
     AddInput("Grad", "(Tensor) Input gradient");
@@ -87,10 +92,85 @@ for numerical stability to avoid the division by zero error.
 )DOC");
   }
 };
+
+namespace {
+size_t FindPos(const std::vector<int64_t>& rows, int64_t value) {
+  return std::find(rows.begin(), rows.end(), value) - rows.begin();
+}
+}  // namespace
+
+template <typename T>
+struct SparseAdagradFunctor<platform::CPUPlace, T> {
+  void operator()(const platform::DeviceContext& context,
+                  const framework::SelectedRows& grad,
+                  const framework::Tensor& learning_rate, T epsilon,
+                  framework::Tensor* moment, framework::Tensor* param) {
+    // 1. g_m.rows = set(g.rows)
+    auto grad_rows = grad.rows();
+    std::set<int64_t> row_set(grad_rows.begin(), grad_rows.end());
+    std::vector<int64_t> merge_rows(row_set.begin(), row_set.end());
+
+    auto grad_width = grad.value().dims()[1];
+    std::unique_ptr<framework::SelectedRows> grad_merge{
+        new framework::SelectedRows()};
+    grad_merge->set_rows(merge_rows);
+    grad_merge->set_height(grad.height());
+    grad_merge->mutable_value()->mutable_data<T>(
+        framework::make_ddim(
+            {static_cast<int64_t>(merge_rows.size()), grad_width}),
+        context.GetPlace());
+
+    math::SetConstant<platform::CPUPlace, T> constant_functor;
+    constant_functor(context, grad_merge->mutable_value(), 0.0);
+
+    auto* grad_merge_data = grad_merge->mutable_value()->data<T>();
+    auto* grad_data = grad.value().data<T>();
+
+    for (size_t i = 0; i < grad_rows.size(); i++) {
+      size_t grad_merge_i = FindPos(merge_rows, grad_rows[i]);
+      for (int64_t j = 0; j < grad_width; j++) {
+        grad_merge_data[grad_merge_i * grad_width + j] +=
+            grad_data[i * grad_width + j];
+      }
+    }
+
+    // 2. m += g_m * g_m
+    std::unique_ptr<framework::SelectedRows> grad_square{
+        new framework::SelectedRows()};
+    grad_square->set_rows(grad_merge->rows());
+    grad_square->set_height(grad_merge->height());
+    grad_square->mutable_value()->mutable_data<T>(grad_merge->value().dims(),
+                                                  context.GetPlace());
+    auto gs =
+        framework::EigenVector<T>::Flatten(*(grad_square->mutable_value()));
+    auto gm = framework::EigenVector<T>::Flatten(grad_merge->value());
+    gs.device(*context.GetEigenDevice<platform::CPUPlace>()) = gm * gm;
+
+    math::SelectedRowsAddToTensor<platform::CPUPlace, T> functor;
+    functor(context, *grad_square, moment);
+
+    // 3. update parameter
+    auto* lr = learning_rate.data<T>();
+    auto* param_data = param->data<T>();
+    auto* moment_data = moment->data<T>();
+
+    for (size_t i = 0; i < merge_rows.size(); i++) {
+      for (int64_t j = 0; j < grad_width; j++) {
+        param_data[merge_rows[i] * grad_width + j] -=
+            lr[0] * grad_merge_data[i * grad_width + j] /
+            (std::sqrt(moment_data[merge_rows[i] * grad_width + j]) + epsilon);
+      }
+    }
+  }
+};
+
+template struct SparseAdagradFunctor<platform::CPUPlace, float>;
+template struct SparseAdagradFunctor<platform::CPUPlace, double>;
 }  // namespace operators
 }  // namespace paddle
 
 namespace ops = paddle::operators;
 REGISTER_OP_WITHOUT_GRADIENT(adagrad, ops::AdagradOp, ops::AdagradOpMaker);
-REGISTER_OP_CPU_KERNEL(adagrad,
-                       ops::AdagradOpKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    adagrad, ops::AdagradOpKernel<paddle::platform::CPUPlace, float>,
+    ops::AdagradOpKernel<paddle::platform::CPUPlace, double>);

paddle/operators/adagrad_op.cu

@@ -14,7 +14,138 @@
 
 #define EIGEN_USE_GPU
 #include "paddle/operators/adagrad_op.h"
+#include "paddle/operators/math/selected_rows_functor.h"
+#include "paddle/operators/math/math_function.h"
+#include "paddle/platform/cuda_helper.h"
+
+namespace paddle {
+namespace operators {
+
+namespace {
+
+template <typename T, int block_size>
+__global__ void MergeGradKernel(const T* grad, const int64_t* grad_rows,
+                                T* grad_merge, const int64_t* grad_merge_rows,
+                                size_t grad_merge_rows_size,
+                                int64_t row_numel) {
+  const int ty = blockIdx.y;
+  int tid = threadIdx.x;
+  __shared__ size_t grad_merge_idx;
+
+  if (tid == 0) {
+    for (size_t i = 0; i < grad_merge_rows_size; i++) {
+      if (grad_rows[ty] == grad_merge_rows[i]) {
+        grad_merge_idx = i;
+      }
+    }
+  }
+
+  __syncthreads();
+
+  grad += ty * row_numel;
+  grad_merge += grad_merge_idx * row_numel;
+  for (int index = tid; index < row_numel; index += block_size) {
+    paddle::platform::CudaAtomicAdd(grad_merge + index, grad[index]);
+  }
+}
+
+template <typename T, int block_size>
+__global__ void SparseAdagradFunctorKernel(const T* grad, const int64_t* rows,
+                                           const T* learning_rate, T* param,
+                                           T* moment, int64_t row_numel,
+                                           T epsilon) {
+  const int ty = blockIdx.y;
+  int tid = threadIdx.x;
+
+  grad += ty * row_numel;
+  param += rows[ty] * row_numel;
+  moment += rows[ty] * row_numel;
+
+  for (int index = tid; index < row_numel; index += block_size) {
+    // Since index in rows of SelectedRows can be duplicate, we have to use
+    // Atomic Operation to avoid concurrent write error.
+    paddle::platform::CudaAtomicAdd(param + index,
+                                    -1.0 * learning_rate[0] * grad[index] /
+                                        (sqrt(moment[index]) + epsilon));
+  }
+}
+}  // namespace
+
+template <typename T>
+struct SparseAdagradFunctor<platform::GPUPlace, T> {
+  void operator()(const platform::DeviceContext& context,
+                  const framework::SelectedRows& grad,
+                  const framework::Tensor& learning_rate, T epsilon,
+                  framework::Tensor* moment, framework::Tensor* param) {
+    // 1. g_m.rows = set(g.rows)
+    auto grad_rows = grad.rows();
+    std::set<int64_t> row_set(grad_rows.begin(), grad_rows.end());
+    std::vector<int64_t> merge_rows(row_set.begin(), row_set.end());
+
+    auto grad_width = grad.value().dims()[1];
+    std::unique_ptr<framework::SelectedRows> grad_merge{
+        new framework::SelectedRows()};
+    grad_merge->set_rows(merge_rows);
+    grad_merge->set_height(grad.height());
+    grad_merge->mutable_value()->mutable_data<T>(
+        framework::make_ddim(
+            {static_cast<int64_t>(merge_rows.size()), grad_width}),
+        context.GetPlace());
+
+    math::SetConstant<platform::GPUPlace, T> constant_functor;
+    constant_functor(context, grad_merge->mutable_value(), 0.0);
+
+    auto* grad_merge_data = grad_merge->mutable_value()->data<T>();
+    auto* grad_data = grad.value().data<T>();
+
+    const int block_size = 256;
+    dim3 threads(block_size, 1);
+    dim3 grid1(1, grad_rows.size());
+
+    MergeGradKernel<
+        T, 256><<<grid1, threads, 0,
+                  reinterpret_cast<const platform::CUDADeviceContext&>(context)
+                      .stream()>>>(grad_data, grad.rows().data(),
+                                   grad_merge_data, grad_merge->rows().data(),
+                                   grad_merge->rows().size(), grad_width);
+
+    // 2. m += g_m * g_m
+    std::unique_ptr<framework::SelectedRows> grad_square{
+        new framework::SelectedRows()};
+    grad_square->set_rows(grad_merge->rows());
+    grad_square->set_height(grad_merge->height());
+    grad_square->mutable_value()->mutable_data<T>(grad_merge->value().dims(),
+                                                  context.GetPlace());
+    auto gs =
+        framework::EigenVector<T>::Flatten(*(grad_square->mutable_value()));
+    auto gm = framework::EigenVector<T>::Flatten(grad_merge->value());
+    gs.device(*context.GetEigenDevice<platform::GPUPlace>()) = gm * gm;
+
+    math::SelectedRowsAddToTensor<platform::GPUPlace, T> functor;
+    functor(context, *grad_square, moment);
+
+    // 3. update parameter
+    auto* lr = learning_rate.data<T>();
+    auto* param_data = param->data<T>();
+    auto* moment_data = moment->data<T>();
+
+    dim3 grid2(1, merge_rows.size());
+    SparseAdagradFunctorKernel<
+        T, 256><<<grid2, threads, 0,
+                  reinterpret_cast<const platform::CUDADeviceContext&>(context)
+                      .stream()>>>(grad_merge_data, grad_merge->rows().data(),
+                                   lr, param_data,
+                                   moment_data, grad_width, epsilon);
+  }
+};
+
+template struct SparseAdagradFunctor<platform::GPUPlace, float>;
+template struct SparseAdagradFunctor<platform::GPUPlace, double>;
+
+}  // namespace operators
+}  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(adagrad,
-                       ops::AdagradOpKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(
+    adagrad, ops::AdagradOpKernel<paddle::platform::GPUPlace, float>,
+    ops::AdagradOpKernel<paddle::platform::GPUPlace, double>);

paddle/operators/adagrad_op.h

@@ -19,35 +19,59 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
+template <typename Place, typename T>
+struct SparseAdagradFunctor {
+  void operator()(const platform::DeviceContext& context,
+                  const framework::SelectedRows& grad,
+                  const framework::Tensor& learning_rate, T epsilon,
+                  framework::Tensor* moment, framework::Tensor* param);
+};
+
 template <typename Place, typename T>
 class AdagradOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-    auto param_out_tensor = ctx.Output<framework::Tensor>("ParamOut");
-    auto moment_out_tensor = ctx.Output<framework::Tensor>("MomentOut");
+    auto* param_out_tensor = ctx.Output<framework::Tensor>("ParamOut");
+    auto* moment_out_tensor = ctx.Output<framework::Tensor>("MomentOut");
 
     param_out_tensor->mutable_data<T>(ctx.GetPlace());
     moment_out_tensor->mutable_data<T>(ctx.GetPlace());
 
-    float epsilon = ctx.Attr<float>("epsilon");
-
-    auto param = framework::EigenVector<T>::Flatten(
-        *ctx.Input<framework::Tensor>("Param"));
-    auto grad = framework::EigenVector<T>::Flatten(
-        *ctx.Input<framework::Tensor>("Grad"));
-    auto moment = framework::EigenVector<T>::Flatten(
-        *ctx.Input<framework::Tensor>("Moment"));
-    auto lr = framework::EigenVector<T>::Flatten(
-        *ctx.Input<framework::Tensor>("LearningRate"));
-
-    auto param_out = framework::EigenVector<T>::Flatten(*param_out_tensor);
-    auto moment_out = framework::EigenVector<T>::Flatten(*moment_out_tensor);
-    auto place = ctx.GetEigenDevice<Place>();
-
-    moment_out.device(place) = moment + grad * grad;
-    Eigen::DSizes<int, 1> m_dsize(moment_out_tensor->numel());
-    param_out.device(place) =
-        param - lr.broadcast(m_dsize) * grad / (moment_out.sqrt() + epsilon);
+    T epsilon = static_cast<T>(ctx.Attr<float>("epsilon"));
+
+    auto* grad_var = ctx.InputVar("Grad");
+    if (grad_var->IsType<framework::LoDTensor>()) {
+      auto param = framework::EigenVector<T>::Flatten(
+          *ctx.Input<framework::Tensor>("Param"));
+      auto grad = framework::EigenVector<T>::Flatten(
+          *ctx.Input<framework::Tensor>("Grad"));
+      auto moment = framework::EigenVector<T>::Flatten(
+          *ctx.Input<framework::Tensor>("Moment"));
+      auto lr = framework::EigenVector<T>::Flatten(
+          *ctx.Input<framework::Tensor>("LearningRate"));
+
+      auto param_out = framework::EigenVector<T>::Flatten(*param_out_tensor);
+      auto moment_out = framework::EigenVector<T>::Flatten(*moment_out_tensor);
+      auto place = ctx.GetEigenDevice<Place>();
+
+      moment_out.device(place) = moment + grad * grad;
+      Eigen::DSizes<int, 1> m_dsize(moment_out_tensor->numel());
+      param_out.device(place) =
+          param - lr.broadcast(m_dsize) * grad / (moment_out.sqrt() + epsilon);
+    } else if (grad_var->IsType<framework::SelectedRows>()) {
+      auto* param_tensor = ctx.Input<framework::Tensor>("Param");
+      PADDLE_ENFORCE_EQ(param_tensor, param_out_tensor);
+
+      auto* moment_tensor = ctx.Input<framework::Tensor>("Moment");
+      PADDLE_ENFORCE_EQ(moment_tensor, moment_out_tensor);
+
+      SparseAdagradFunctor<Place, T> functor;
+      functor(ctx.device_context(), *ctx.Input<framework::SelectedRows>("Grad"),
+              *ctx.Input<framework::Tensor>("LearningRate"), epsilon,
+              moment_out_tensor, param_out_tensor);
+    } else {
+      PADDLE_THROW("Unsupported Variable Type of Grad");
+    }
   }
 };
 

paddle/operators/array_operator.h

@@ -42,6 +42,7 @@ class ArrayOp : public framework::OperatorBase {
     } else {
       offset = static_cast<size_t>(*i_tensor.data<int64_t>());
     }
+    VLOG(10) << " Offset = " << offset;
     return offset;
   }
 };

paddle/operators/beam_search_op.cc

+/* 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 "paddle/operators/beam_search_op.h"
+
+#include <map>
+#include "paddle/framework/lod_tensor.h"
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+void BeamSearch::operator()(const framework::LoDTensor &pre_ids,
+                            framework::LoDTensor *selected_ids,
+                            framework::LoDTensor *selected_scores) {
+  auto items = SelectTopBeamSizeItems();
+  auto selected_items = ToMap(items);
+  PruneEndidCandidates(pre_ids, &selected_items);
+  // calculate the output tensor's height
+  size_t num_instances = std::accumulate(
+      std::begin(items), std::end(items), 0,
+      [](size_t a, std::vector<Item> &b) { return a + b.size(); });
+  // the output tensor shape should be [num_instances, 1]
+  auto dims = framework::make_ddim(
+      std::vector<int64_t>({static_cast<int>(num_instances), 1}));
+  selected_ids->Resize(dims);
+  selected_scores->Resize(dims);
+
+  std::map<size_t /*offset*/, std::vector<Item>> hash;
+  framework::LoD new_lod;
+  auto *ids_data = selected_ids->mutable_data<int>(platform::CPUPlace());
+  auto *scores_data =
+      selected_scores->mutable_data<float>(platform::CPUPlace());
+
+  // fill in data
+  std::vector<size_t> low_level;
+  size_t low_offset = 0;
+  for (auto &items : selected_items) {
+    low_level.push_back(low_offset);
+    for (auto &item : items) {
+      ids_data[low_offset] = item.id;
+      scores_data[low_offset] = item.score;
+      low_offset++;
+    }
+  }
+  // fill lod
+  auto abs_lod = framework::ToAbsOffset(ids_->lod());
+  auto &high_level = abs_lod[lod_level_];
+  framework::LoD lod(2);
+  lod[0].assign(high_level.begin(), high_level.end());
+  lod[1].assign(low_level.begin(), low_level.end());
+  selected_ids->set_lod(lod);
+  selected_scores->set_lod(lod);
+}
+
+void BeamSearch::PruneEndidCandidates(const framework::LoDTensor &pre_ids,
+                                      std::vector<std::vector<Item>> *items) {
+  auto *pre_ids_data = pre_ids.data<int>();
+
+  for (size_t offset = 0; offset < items->size(); offset++) {
+    auto prefix_id = pre_ids_data[offset];
+    if (prefix_id == end_id_) {
+      items->at(offset).clear();
+    }
+  }
+}
+
+std::vector<std::vector<BeamSearch::Item>> BeamSearch::ToMap(
+    const std::vector<std::vector<Item>> &items) {
+  std::vector<std::vector<Item>> result;
+  for (auto &entries : items) {
+    for (const auto &item : entries) {
+      if (item.offset >= result.size()) {
+        result.resize(item.offset + 1);
+      }
+      result[item.offset].push_back(item);
+    }
+  }
+  return result;
+}
+
+std::vector<std::vector<BeamSearch::Item>>
+BeamSearch::SelectTopBeamSizeItems() {
+  std::vector<std::vector<Item>> result;
+  std::vector<Item> items;
+  // for each source sentence, select the top beam_size items across all
+  // candidate sets.
+  while (NextItemSet(&items)) {
+    std::nth_element(std::begin(items), std::begin(items) + beam_size_,
+                     std::end(items), [](const Item &a, const Item &b) {
+                       // TODO(superjom) make score's comparation customizable.
+                       // partial sort in descending order
+                       return a.score > b.score;
+                     });
+    // prune the top beam_size items.
+    if (items.size() > beam_size_) {
+      items.resize(beam_size_);
+    }
+    result.emplace_back(items);
+  }
+  return result;
+}
+
+// the candidates of a source
+bool BeamSearch::NextItemSet(std::vector<BeamSearch::Item> *items) {
+  if (sent_offset_ >= ids_->NumElements(lod_level_)) {
+    return false;
+  }
+  // find the current candidates
+  auto ids = *ids_;
+  auto scores = *scores_;
+
+  auto source_abs_two_level_lod = framework::SliceInLevel(
+      ids.lod(), lod_level_, sent_offset_, sent_offset_ + 1);
+  source_abs_two_level_lod = framework::ToAbsOffset(source_abs_two_level_lod);
+  auto abs_lod = framework::ToAbsOffset(ids.lod());
+  PADDLE_ENFORCE_GE(source_abs_two_level_lod.size(), 2UL);
+
+  auto *ids_data = ids.data<int>();
+  auto *scores_data = scores.data<float>();
+
+  size_t instance_dim = 1;
+  for (int i = 1; i < ids.dims().size(); i++) {
+    instance_dim *= ids.dims()[i];
+  }
+
+  items->clear();
+  items->reserve(framework::product(ids.dims()));
+  for (size_t offset = abs_lod[lod_level_][sent_offset_];
+       offset < abs_lod[lod_level_][sent_offset_ + 1]; offset++) {
+    for (int d = 0; d < instance_dim; d++) {
+      const size_t dim_offset = offset * instance_dim + d;
+      items->emplace_back(offset, ids_data[dim_offset],
+                          scores_data[dim_offset]);
+    }
+  }
+
+  sent_offset_++;
+  return true;
+}
+
+class BeamSearchProtoAndCheckerMaker
+    : public framework::OpProtoAndCheckerMaker {
+ public:
+  BeamSearchProtoAndCheckerMaker(framework::OpProto *proto,
+                                 framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    // inputs and outputs stored in proto
+    AddInput("pre_ids", "ids in previous step");
+    AddInput("ids", "a LoDTensor of shape of [None,k]");
+    AddInput("scores",
+             "a LoDTensor that has the same shape and LoD with `ids`");
+    AddOutput("selected_ids",
+              "a LoDTensor that stores the IDs selected by beam search");
+    AddOutput(
+        "selected_scores",
+        "a LoDTensor that has the same shape and LoD with `selected_ids`");
+
+    // Attributes stored in AttributeMap
+    AddAttr<int>("level", "the level of LoDTensor");
+    AddAttr<int>("beam_size", "beam size for beam search");
+    AddAttr<int>("end_id",
+                 "the token id which indicates the end of a sequence");
+
+    AddComment(
+        "This is a beam search operator that help to generate sequences.");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+REGISTER_OP_WITHOUT_GRADIENT(beam_search, paddle::operators::BeamSearchOp,
+                             paddle::operators::BeamSearchProtoAndCheckerMaker);

paddle/operators/bilinear_tensor_product_op.h

@@ -174,7 +174,7 @@ class BilinearTensorProductGradKernel : public framework::OpKernel<T> {
     // Caculate the gradient of Input(Bias).
     if (d_bias) {
       d_bias->mutable_data<T>(ctx.GetPlace());
-      auto d_bias_mat = EigenMatrix<T>::From(*d_bias);
+      auto d_bias_mat = framework::EigenVector<T>::Flatten(*d_bias);
       d_bias_mat.device(place) = d_out_mat.sum(Eigen::DSizes<int, 1>(0));
     }
   }

paddle/operators/conv_cudnn_op.cc

@@ -22,8 +22,6 @@ class CudnnConvOpMaker : public Conv2DOpMaker {
   CudnnConvOpMaker(framework::OpProto* proto,
                    framework::OpAttrChecker* op_checker)
       : Conv2DOpMaker(proto, op_checker) {
-    AddAttr<std::vector<int>>("dilations", "dilations of convolution operator.")
-        .SetDefault(std::vector<int>{1, 1});
     AddAttr<int>("workspace_size_MB",
                  "workspace size for cudnn, in MB, "
                  "workspace is a section of GPU memory which will be "

paddle/operators/conv_cudnn_op.cu → paddle/operators/conv_cudnn_op.cu.cc

@@ -226,9 +226,8 @@ class CudnnConvGradOpKernel : public framework::OpKernel<T> {
     T alpha = 1.0f, beta = 0.0f;
     if (input_grad) {
       T* input_grad_data = input_grad->mutable_data<T>(ctx.GetPlace());
-      auto t = framework::EigenVector<T>::Flatten(*input_grad);
-      t.device(ctx.GetEigenDevice<platform::GPUPlace>()) =
-          t.constant(static_cast<T>(0));
+      // Because beta is zero, it is unnecessary to reset input_grad.
+
       for (int i = 0; i < groups; i++) {
         PADDLE_ENFORCE(platform::dynload::cudnnConvolutionBackwardData(
             handle, &alpha, cudnn_filter_desc,
@@ -241,9 +240,8 @@ class CudnnConvGradOpKernel : public framework::OpKernel<T> {
     // ------------------- cudnn conv backward filter ---------------------
     if (filter_grad) {
       T* filter_grad_data = filter_grad->mutable_data<T>(ctx.GetPlace());
-      auto t = framework::EigenVector<T>::Flatten(*filter_grad);
-      t.device(ctx.GetEigenDevice<platform::GPUPlace>()) =
-          t.constant(static_cast<T>(0));
+      // Because beta is zero, it is unnecessary to reset filter_grad.
+
       for (int i = 0; i < groups; i++) {
         PADDLE_ENFORCE(platform::dynload::cudnnConvolutionBackwardFilter(
             handle, &alpha, cudnn_input_desc, input_data + i * group_offset_in,

paddle/operators/cos_sim_op.h

@@ -132,7 +132,7 @@ class CosSimGradKernel : public framework::OpKernel<T> {
       // compute dy
       if (out_grad_y) {
         out_grad_y->mutable_data<T>(context.GetPlace());
-        auto dy = EigenMatrix<T>::Reshape(*out_grad_y, 1);
+        auto dy = EigenVector<T>::Flatten(*out_grad_y);
         auto grad = x / norm_prod_bcast - z_bcast * y_bcast / y_snorm_bcast;
         dy.device(place) = (dz_bcast * grad).sum(Eigen::array<int, 1>({{0}}));
       }