Merge pull request #1 from PaddlePaddle/develop

update from origin

benchmark/paddle/image/resnet.py

+#!/usr/bin/env python
+from paddle.trainer_config_helpers import *
+
+height = 224
+width = 224
+num_class = 1000
+batch_size = get_config_arg('batch_size', int, 64)
+layer_num = get_config_arg("layer_num", int, 50)
+is_test = get_config_arg("is_test", bool, False)
+
+args = {'height': height, 'width': width, 'color': True, 'num_class': num_class}
+define_py_data_sources2(
+    "train.list", None, module="provider", obj="process", args=args)
+
+settings(
+    batch_size=batch_size,
+    learning_rate=0.01 / batch_size,
+    learning_method=MomentumOptimizer(0.9),
+    regularization=L2Regularization(0.0005 * batch_size))
+
+
+#######################Network Configuration #############
+def conv_bn_layer(name,
+                  input,
+                  filter_size,
+                  num_filters,
+                  stride,
+                  padding,
+                  channels=None,
+                  active_type=ReluActivation()):
+    """
+    A wrapper for conv layer with batch normalization layers.
+    Note:
+    conv layer has no activation.
+    """
+
+    tmp = img_conv_layer(
+        name=name + "_conv",
+        input=input,
+        filter_size=filter_size,
+        num_channels=channels,
+        num_filters=num_filters,
+        stride=stride,
+        padding=padding,
+        act=LinearActivation(),
+        bias_attr=False)
+    return batch_norm_layer(
+        name=name + "_bn", input=tmp, act=active_type, use_global_stats=is_test)
+
+
+def bottleneck_block(name, input, num_filters1, num_filters2):
+    """
+    A wrapper for bottlenect building block in ResNet.
+    Last conv_bn_layer has no activation.
+    Addto layer has activation of relu.
+    """
+    last_name = conv_bn_layer(
+        name=name + '_branch2a',
+        input=input,
+        filter_size=1,
+        num_filters=num_filters1,
+        stride=1,
+        padding=0)
+    last_name = conv_bn_layer(
+        name=name + '_branch2b',
+        input=last_name,
+        filter_size=3,
+        num_filters=num_filters1,
+        stride=1,
+        padding=1)
+    last_name = conv_bn_layer(
+        name=name + '_branch2c',
+        input=last_name,
+        filter_size=1,
+        num_filters=num_filters2,
+        stride=1,
+        padding=0,
+        active_type=LinearActivation())
+
+    return addto_layer(
+        name=name + "_addto", input=[input, last_name], act=ReluActivation())
+
+
+def mid_projection(name, input, num_filters1, num_filters2, stride=2):
+    """
+    A wrapper for middile projection in ResNet.
+    projection shortcuts are used for increasing dimensions,
+    and other shortcuts are identity
+    branch1: projection shortcuts are used for increasing
+    dimensions, has no activation.
+    branch2x: bottleneck building block, shortcuts are identity.
+    """
+    # stride = 2
+    branch1 = conv_bn_layer(
+        name=name + '_branch1',
+        input=input,
+        filter_size=1,
+        num_filters=num_filters2,
+        stride=stride,
+        padding=0,
+        active_type=LinearActivation())
+
+    last_name = conv_bn_layer(
+        name=name + '_branch2a',
+        input=input,
+        filter_size=1,
+        num_filters=num_filters1,
+        stride=stride,
+        padding=0)
+    last_name = conv_bn_layer(
+        name=name + '_branch2b',
+        input=last_name,
+        filter_size=3,
+        num_filters=num_filters1,
+        stride=1,
+        padding=1)
+
+    last_name = conv_bn_layer(
+        name=name + '_branch2c',
+        input=last_name,
+        filter_size=1,
+        num_filters=num_filters2,
+        stride=1,
+        padding=0,
+        active_type=LinearActivation())
+
+    return addto_layer(
+        name=name + "_addto", input=[branch1, last_name], act=ReluActivation())
+
+
+img = data_layer(name='image', size=height * width * 3)
+
+
+def deep_res_net(res2_num=3, res3_num=4, res4_num=6, res5_num=3):
+    """
+    A wrapper for 50,101,152 layers of ResNet.
+    res2_num: number of blocks stacked in conv2_x
+    res3_num: number of blocks stacked in conv3_x
+    res4_num: number of blocks stacked in conv4_x
+    res5_num: number of blocks stacked in conv5_x
+    """
+    # For ImageNet
+    # conv1: 112x112
+    tmp = conv_bn_layer(
+        "conv1",
+        input=img,
+        filter_size=7,
+        channels=3,
+        num_filters=64,
+        stride=2,
+        padding=3)
+    tmp = img_pool_layer(name="pool1", input=tmp, pool_size=3, stride=2)
+
+    # conv2_x: 56x56
+    tmp = mid_projection(
+        name="res2_1", input=tmp, num_filters1=64, num_filters2=256, stride=1)
+    for i in xrange(2, res2_num + 1, 1):
+        tmp = bottleneck_block(
+            name="res2_" + str(i), input=tmp, num_filters1=64, num_filters2=256)
+
+    # conv3_x: 28x28
+    tmp = mid_projection(
+        name="res3_1", input=tmp, num_filters1=128, num_filters2=512)
+    for i in xrange(2, res3_num + 1, 1):
+        tmp = bottleneck_block(
+            name="res3_" + str(i),
+            input=tmp,
+            num_filters1=128,
+            num_filters2=512)
+
+    # conv4_x: 14x14
+    tmp = mid_projection(
+        name="res4_1", input=tmp, num_filters1=256, num_filters2=1024)
+    for i in xrange(2, res4_num + 1, 1):
+        tmp = bottleneck_block(
+            name="res4_" + str(i),
+            input=tmp,
+            num_filters1=256,
+            num_filters2=1024)
+
+    # conv5_x: 7x7
+    tmp = mid_projection(
+        name="res5_1", input=tmp, num_filters1=512, num_filters2=2048)
+    for i in xrange(2, res5_num + 1, 1):
+        tmp = bottleneck_block(
+            name="res5_" + str(i),
+            input=tmp,
+            num_filters1=512,
+            num_filters2=2048)
+
+    tmp = img_pool_layer(
+        name='avgpool',
+        input=tmp,
+        pool_size=7,
+        stride=1,
+        pool_type=AvgPooling())
+
+    return fc_layer(input=tmp, size=num_class, act=SoftmaxActivation())
+
+
+if layer_num == 50:
+    resnet = deep_res_net(3, 4, 6, 3)
+elif layer_num == 101:
+    resnet = deep_res_net(3, 4, 23, 3)
+elif layer_num == 152:
+    resnet = deep_res_net(3, 8, 36, 3)
+else:
+    print("Wrong layer number.")
+
+lbl = data_layer(name="label", size=num_class)
+loss = cross_entropy(name='loss', input=resnet, label=lbl)
+inputs(img, lbl)
+outputs(loss)

benchmark/paddle/image/run_mkldnn.sh

@@ -5,22 +5,23 @@ function train() {
   export OMP_DYNAMIC="FALSE"
   export KMP_AFFINITY="granularity=fine,compact,0,0"
   topology=$1
-  bs=$2
-  use_mkldnn=$3
-  if [ $3 == "True" ]; then
+  layer_num=$2
+  bs=$3
+  use_mkldnn=$4
+  if [ $4 == "True" ]; then
     thread=1
-    log="logs/${topology}-mkldnn-${bs}.log"
-  elif [ $3 == "False" ]; then
+    log="logs/${topology}-${layer_num}-mkldnn-${bs}.log"
+  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}-${thread}mklml-${bs}.log"
+    log="logs/${topology}-${layer_num}-${thread}mklml-${bs}.log"
   else
     echo "Wrong input $3, use True or False."
     exit 0
   fi
-  args="batch_size=${bs}"
+  args="batch_size=${bs},layer_num=${layer_num}"
   config="${topology}.py"
   paddle train --job=time \
     --config=$config \
@@ -40,12 +41,9 @@ if [ ! -d "logs" ]; then
   mkdir logs
 fi
 
-#========== mkldnn ==========#
-train vgg 64 True
-train vgg 128 True
-train vgg 256 True
-
-#========== mklml ===========#
-train vgg 64 False
-train vgg 128 False
-train vgg 256 False
+for use_mkldnn in True False; do
+  for batchsize in 64 128 256; do
+    train vgg 19 $batchsize $use_mkldnn
+    train resnet 50  $batchsize $use_mkldnn
+  done
+done

benchmark/paddle/image/vgg.py

@@ -13,7 +13,7 @@ define_py_data_sources2(
 
 settings(
     batch_size=batch_size,
-    learning_rate=0.01 / batch_size,
+    learning_rate=0.001 / batch_size,
     learning_method=MomentumOptimizer(0.9),
     regularization=L2Regularization(0.0005 * batch_size))
 

cmake/cblas.cmake

 # Find the CBlas and lapack libraries
 #
-# It will search MKL, atlas, OpenBlas, reference-cblas in order.
+# It will search MKLML, atlas, OpenBlas, reference-cblas in order.
 #
 # If any cblas implementation found, the following variable will be set.
-#    CBLAS_PROVIDER  # one of MKL, ATLAS, OPENBLAS, REFERENCE
+#    CBLAS_PROVIDER  # one of MKLML, ATLAS, OPENBLAS, REFERENCE
 #    CBLAS_INC_DIR   # the include directory for cblas.
 #    CBLAS_LIBS      # a list of libraries should be linked by paddle.
 #                    # Each library should be full path to object file.
-#
-# User should set one of MKL_ROOT, ATLAS_ROOT, OPENBLAS_ROOT, REFERENCE_CBLAS_ROOT
-# during cmake. If none of them set, it will try to find cblas implementation in
-# system paths.
-#
 
 set(CBLAS_FOUND OFF)
 
@@ -30,44 +25,6 @@ if(WITH_MKLML AND MKLML_INC_DIR AND MKLML_LIB)
   return()
 endif()
 
-## Then find MKL.
-set(INTEL_MKL_ROOT "/opt/intel/mkl" CACHE PATH "Folder contains intel mkl libs")
-set(MKL_ROOT $ENV{MKL_ROOT} CACHE PATH "Folder contains env MKL")
-
-set(MKL_INCLUDE_SEARCH_PATHS
-  ${MKL_ROOT}/include
-  ${INTEL_MKL_ROOT}/include)
-set(MKL_LIB_SEARCH_PATHS
-  ${MKL_ROOT}/lib
-  ${MKL_ROOT}/lib/intel64
-  ${INTEL_MKL_ROOT}/lib
-  ${INTEL_MKL_ROOT}/lib/intel64)
-
-find_path(MKL_INC_DIR mkl.h PATHS
-  ${MKL_INCLUDE_SEARCH_PATHS})
-find_path(MKL_LAPACK_INC_DIR mkl_lapacke.h PATHS
-  ${MKL_INCLUDE_SEARCH_PATHS})
-find_library(MKL_CORE_LIB NAMES mkl_core PATHS
-  ${MKL_LIB_SEARCH_PATHS})
-find_library(MKL_SEQUENTIAL_LIB NAMES mkl_sequential PATHS
-  ${MKL_LIB_SEARCH_PATHS})
-find_library(MKL_INTEL_LP64 NAMES mkl_intel_lp64 PATHS
-  ${MKL_LIB_SEARCH_PATHS})
-
-if(MKL_LAPACK_INC_DIR AND MKL_INC_DIR AND MKL_CORE_LIB AND MKL_SEQUENTIAL_LIB AND MKL_INTEL_LP64)
-  set(CBLAS_FOUND ON)
-  set(CBLAS_PROVIDER MKL)
-  set(CBLAS_INC_DIR ${MKL_INC_DIR} ${MKL_LAPACK_INC_DIR})
-  set(CBLAS_LIBRARIES ${MKL_INTEL_LP64} ${MKL_SEQUENTIAL_LIB} ${MKL_CORE_LIB})
-
-  add_definitions(-DPADDLE_USE_MKL)
-  add_definitions(-DLAPACK_FOUND)
-
-  message(STATUS "Found MKL (include: ${MKL_INC_DIR}, library: ${CBLAS_LIBRARIES})")
-  message(STATUS "Found lapack in MKL (include: ${MKL_LAPACK_INC_DIR})")
-  return()
-endif()
-
 ## Then find atlas.
 set(ATLAS_ROOT $ENV{ATLAS_ROOT} CACHE PATH "Folder contains Atlas")
 set(ATLAS_INCLUDE_SEARCH_PATHS

cmake/cross_compiling/ios.cmake

@@ -79,9 +79,8 @@ if(NOT DEFINED IOS_ARCH)
     # FIXME(liuyiqun): support "armv7;armv7s;arm64" future
     set(IOS_ARCH "arm64")
   elseif(IOS_PLATFORM STREQUAL "SIMULATOR")
-    set(IOS_ARCH "i386;x86_64")
-  elseif(IOS_PLATFORM STREQUAL "WATCHOS")
-    set(IOS_ARCH armv7k)
+    # FIXME(liuyiqun): support "i386;x86_64" future
+    set(IOS_ARCH "x86_64")
   endif()
 endif()
 set(CMAKE_OSX_ARCHITECTURES ${IOS_ARCH} CACHE string  "Build architecture for iOS")

cmake/external/mkldnn.cmake

@@ -46,16 +46,20 @@ IF(${CBLAS_PROVIDER} STREQUAL "MKLML")
     MESSAGE(STATUS "Build MKLDNN with ${MKLDNN_MKLROOT}")
 ENDIF()
 
+SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} -Wno-error=strict-overflow")
+SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} -Wno-error=strict-overflow")
 ExternalProject_Add(
     ${MKLDNN_PROJECT}
     ${EXTERNAL_PROJECT_LOG_ARGS}
     DEPENDS             ${MKLDNN_DEPENDS}
     GIT_REPOSITORY      "https://github.com/01org/mkl-dnn.git"
-    GIT_TAG             "v0.10"
+    GIT_TAG             "v0.11"
     PREFIX              ${MKLDNN_SOURCES_DIR}
     UPDATE_COMMAND      ""
     CMAKE_ARGS          -DCMAKE_INSTALL_PREFIX=${MKLDNN_INSTALL_DIR}
     CMAKE_ARGS          -DMKLROOT=${MKLDNN_MKLROOT}
+    CMAKE_ARGS          -DCMAKE_C_FLAGS=${MKLDNN_CFLAG}
+    CMAKE_ARGS          -DCMAKE_CXX_FLAGS=${MKLDNN_CXXFLAG}
     CMAKE_CACHE_ARGS    -DCMAKE_INSTALL_PREFIX:PATH=${MKLDNN_INSTALL_DIR}
                         -DMKLROOT:PATH=${MKLDNN_MKLROOT}
 )

cmake/external/mklml.cmake

@@ -27,8 +27,8 @@ ENDIF()
 INCLUDE(ExternalProject)
 
 SET(MKLML_PROJECT       "extern_mklml")
-SET(MKLML_VER           "mklml_lnx_2018.0.20170720")
-SET(MKLML_URL           "https://github.com/01org/mkl-dnn/releases/download/v0.10/${MKLML_VER}.tgz")
+SET(MKLML_VER           "mklml_lnx_2018.0.1.20171007")
+SET(MKLML_URL           "https://github.com/01org/mkl-dnn/releases/download/v0.11/${MKLML_VER}.tgz")
 SET(MKLML_SOURCE_DIR    "${THIRD_PARTY_PATH}/mklml")
 SET(MKLML_DOWNLOAD_DIR  "${MKLML_SOURCE_DIR}/src/${MKLML_PROJECT}")
 SET(MKLML_DST_DIR       "mklml")

cmake/external/nccl.cmake

+# 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.
+
+if(NOT WITH_GPU)
+  return()
+endif()
+
 include(ExternalProject)
 
 set(NCCL_SOURCE_DIR ${THIRD_PARTY_PATH}/nccl)

cmake/external/openblas.cmake

@@ -86,7 +86,7 @@ IF(NOT ${CBLAS_FOUND})
         UPDATE_COMMAND      ""
         CONFIGURE_COMMAND   ""
     )
-
+    SET(CBLAS_PROVIDER openblas)
     IF(WITH_C_API)
         INSTALL(DIRECTORY ${CBLAS_INC_DIR} DESTINATION third_party/openblas)
         # Because libopenblas.a is a symbolic link of another library, thus need to
@@ -115,7 +115,7 @@ INCLUDE_DIRECTORIES(${CBLAS_INC_DIR})
 # linear algebra libraries for cc_library(xxx SRCS xxx.c DEPS cblas)
 SET(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/cblas_dummy.c)
 FILE(WRITE ${dummyfile} "const char * dummy = \"${dummyfile}\";")
-IF(${CBLAS_PROVIDER} MATCHES MKL)
+IF("${CBLAS_PROVIDER}" STREQUAL "MKLML")
     ADD_LIBRARY(cblas SHARED ${dummyfile})
 ELSE()
     ADD_LIBRARY(cblas STATIC ${dummyfile})

cmake/external/pybind11.cmake

-INCLUDE(ExternalProject)
+# 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.
 
-SET(PYBIND_SOURCE_DIR ${THIRD_PARTY_PATH}/pybind)
+if(NOT WITH_PYTHON)
+    return()
+endif()
+
+include(ExternalProject)
 
-INCLUDE_DIRECTORIES(${PYBIND_SOURCE_DIR}/src/extern_pybind/include)
+set(PYBIND_SOURCE_DIR ${THIRD_PARTY_PATH}/pybind)
+
+include_directories(${PYBIND_SOURCE_DIR}/src/extern_pybind/include)
 
 ExternalProject_Add(
         extern_pybind
@@ -17,14 +35,12 @@ ExternalProject_Add(
         TEST_COMMAND      ""
 )
 
-if (${CMAKE_VERSION} VERSION_LESS "3.3.0")
+if(${CMAKE_VERSION} VERSION_LESS "3.3.0")
     set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/pybind_dummy.c)
-    file(WRITE ${dummyfile} "const char * dummy_any = \"${dummyfile}\";")
+    file(WRITE ${dummyfile} "const char * dummy_pybind = \"${dummyfile}\";")
     add_library(pybind STATIC ${dummyfile})
 else()
     add_library(pybind INTERFACE)
 endif()
 
 add_dependencies(pybind extern_pybind)
-
-LIST(APPEND external_project_dependencies pybind)

cmake/generic.cmake

@@ -93,7 +93,7 @@ include_directories(${CMAKE_CURRENT_BINARY_DIR})
 if(NOT APPLE AND NOT ANDROID)
     find_package(Threads REQUIRED)
     link_libraries(${CMAKE_THREAD_LIBS_INIT})
-    set(CMAKE_CXX_LINK_EXECUTABLE "${CMAKE_CXX_LINK_EXECUTABLE} -ldl -lrt")
+    set(CMAKE_CXX_LINK_EXECUTABLE "${CMAKE_CXX_LINK_EXECUTABLE} -pthread -ldl -lrt")
 endif(NOT APPLE AND NOT ANDROID)
 
 function(merge_static_libs TARGET_NAME)

cmake/simd.cmake

 # This file is use to check all support level of AVX on your machine
 # so that PaddlePaddle can unleash the vectorization power of muticore.
 
-INCLUDE(CheckCXXSourceRuns)
-INCLUDE(CheckCXXSourceCompiles)
+include(CheckCXXSourceRuns)
+include(CheckCXXSourceCompiles)
 
-IF(CMAKE_COMPILER_IS_GNUCC OR CMAKE_COMPILER_IS_GNUCXX OR CMAKE_CXX_COMPILER_ID MATCHES "Clang")
+if(CMAKE_COMPILER_IS_GNUCC OR CMAKE_COMPILER_IS_GNUCXX OR CMAKE_CXX_COMPILER_ID MATCHES "Clang")
     set(MMX_FLAG "-mmmx")
     set(SSE2_FLAG "-msse2")
     set(SSE3_FLAG "-msse3")
-    SET(AVX_FLAG "-mavx")
-    SET(AVX2_FLAG "-mavx2")
-ELSEIF(MSVC)
+    set(AVX_FLAG "-mavx")
+    set(AVX2_FLAG "-mavx2")
+elseif(MSVC)
     set(MMX_FLAG "/arch:MMX")
     set(SSE2_FLAG "/arch:SSE2")
     set(SSE3_FLAG "/arch:SSE3")
     SET(AVX_FLAG "/arch:AVX")
     SET(AVX2_FLAG "/arch:AVX2")
-ENDIF()
+endif()
 
 set(CMAKE_REQUIRED_FLAGS_RETAINED ${CMAKE_REQUIRED_FLAGS})
 
 # Check  MMX
 set(CMAKE_REQUIRED_FLAGS ${MMX_FLAG})
+set(MMX_FOUND_EXITCODE 1 CACHE STRING "Result from TRY_RUN" FORCE)
 CHECK_CXX_SOURCE_RUNS("
 #include <mmintrin.h>
 int main()
@@ -32,6 +33,7 @@ int main()
 
 # Check SSE2
 set(CMAKE_REQUIRED_FLAGS ${SSE2_FLAG})
+set(SSE2_FOUND_EXITCODE 1 CACHE STRING "Result from TRY_RUN" FORCE)
 CHECK_CXX_SOURCE_RUNS("
 #include <emmintrin.h>
 int main()
@@ -42,6 +44,7 @@ int main()
 
 # Check SSE3
 set(CMAKE_REQUIRED_FLAGS ${SSE3_FLAG})
+set(SSE3_FOUND_EXITCODE 1 CACHE STRING "Result from TRY_RUN" FORCE)
 CHECK_CXX_SOURCE_RUNS("
 #include <pmmintrin.h>
 int main()
@@ -55,6 +58,7 @@ int main()
 
 # Check AVX
 set(CMAKE_REQUIRED_FLAGS ${AVX_FLAG})
+set(AVX_FOUND_EXITCODE 1 CACHE STRING "Result from TRY_RUN" FORCE)
 CHECK_CXX_SOURCE_RUNS("
 #include <immintrin.h>
 int main()
@@ -67,6 +71,7 @@ int main()
 
 # Check AVX 2
 set(CMAKE_REQUIRED_FLAGS ${AVX2_FLAG})
+set(AVX2_FOUND_EXITCODE 1 CACHE STRING "Result from TRY_RUN" FORCE)
 CHECK_CXX_SOURCE_RUNS("
 #include <immintrin.h>
 int main()

doc/api/v2/config/layer.rst

@@ -82,6 +82,11 @@ maxout
 ..  autoclass:: paddle.v2.layer.maxout
     :noindex:
 
+roi_pool
+--------
+..  autoclass:: paddle.v2.layer.roi_pool
+    :noindex:
+
 Norm Layer
 ==========
 

doc/api/v2/data.rst

@@ -2,112 +2,9 @@
 Data Reader Interface and DataSets
 ==================================
 
+..  toctree::
+    :maxdepth: 1
 
-DataTypes
-=========
-
-..  automodule:: paddle.v2.data_type
-    :members:
-    :noindex:
-
-DataFeeder
-==========
-
-..  automodule:: paddle.v2.data_feeder
-    :members:
-    :noindex:
-
-Reader
-======
-
-..  automodule:: paddle.v2.reader
-    :members:
-    :noindex:
-
-..  automodule:: paddle.v2.reader.creator
-    :members:
-    :noindex:
-
-minibatch
-=========
-
-..  automodule:: paddle.v2.minibatch
-    :members:
-    :noindex:
-
-Dataset
-=======
-
-..  automodule:: paddle.v2.dataset
-    :members:
-    :noindex:
-
-mnist
-+++++
-
-..  automodule:: paddle.v2.dataset.mnist
-    :members:
-    :noindex:
-
-cifar
-+++++
-
-..  automodule:: paddle.v2.dataset.cifar
-    :members:
-    :noindex:
-
-conll05
-+++++++
-
-..  automodule:: paddle.v2.dataset.conll05
-    :members: get_dict,get_embedding,test
-    :noindex:
-
-imdb
-++++
-
-..  automodule:: paddle.v2.dataset.imdb
-    :members:
-    :noindex:
-
-imikolov
-++++++++
-
-..  automodule:: paddle.v2.dataset.imikolov
-    :members:
-    :noindex:
-
-movielens
-+++++++++
-
-..  automodule:: paddle.v2.dataset.movielens
-    :members:
-    :noindex:
-
-..  autoclass:: paddle.v2.dataset.movielens.MovieInfo
-    :noindex:
-    
-..  autoclass:: paddle.v2.dataset.movielens.UserInfo
-    :noindex:
-
-sentiment
-+++++++++
-
-..  automodule:: paddle.v2.dataset.sentiment
-    :members:
-    :noindex:
-
-uci_housing
-+++++++++++
-
-..  automodule:: paddle.v2.dataset.uci_housing
-    :members:
-    :noindex:
-
-wmt14
-+++++
-
-..  automodule:: paddle.v2.dataset.wmt14
-    :members:
-    :noindex:
-
+    data/data_reader.rst
+    data/image.rst
+    data/dataset.rst

doc/api/v2/data/data_reader.rst

+=====================
+Data Reader Interface
+=====================
+
+
+DataTypes
+=========
+
+..  automodule:: paddle.v2.data_type
+    :members:
+    :noindex:
+
+DataFeeder
+==========
+
+..  automodule:: paddle.v2.data_feeder
+    :members:
+    :noindex:
+
+Reader
+======
+
+..  automodule:: paddle.v2.reader
+    :members:
+    :noindex:
+
+..  automodule:: paddle.v2.reader.creator
+    :members:
+    :noindex:
+
+minibatch
+=========
+
+..  automodule:: paddle.v2.minibatch
+    :members:
+    :noindex:

doc/api/v2/data/dataset.rst

+Dataset
+=======
+
+..  automodule:: paddle.v2.dataset
+    :members:
+    :noindex:
+
+mnist
++++++
+
+..  automodule:: paddle.v2.dataset.mnist
+    :members:
+    :noindex:
+
+cifar
++++++
+
+..  automodule:: paddle.v2.dataset.cifar
+    :members:
+    :noindex:
+
+conll05
++++++++
+
+..  automodule:: paddle.v2.dataset.conll05
+    :members: get_dict,get_embedding,test
+    :noindex:
+
+imdb
+++++
+
+..  automodule:: paddle.v2.dataset.imdb
+    :members:
+    :noindex:
+
+imikolov
+++++++++
+
+..  automodule:: paddle.v2.dataset.imikolov
+    :members:
+    :noindex:
+
+movielens
++++++++++
+
+..  automodule:: paddle.v2.dataset.movielens
+    :members:
+    :noindex:
+
+..  autoclass:: paddle.v2.dataset.movielens.MovieInfo
+    :noindex:
+    
+..  autoclass:: paddle.v2.dataset.movielens.UserInfo
+    :noindex:
+
+sentiment
++++++++++
+
+..  automodule:: paddle.v2.dataset.sentiment
+    :members:
+    :noindex:
+
+uci_housing
++++++++++++
+
+..  automodule:: paddle.v2.dataset.uci_housing
+    :members:
+    :noindex:
+
+wmt14
++++++
+
+..  automodule:: paddle.v2.dataset.wmt14
+    :members:
+    :noindex:

doc/api/v2/data/image.rst

+Image Interface
+===============
+
+..  automodule:: paddle.v2.image
+    :members:

doc/design/float16.md

+# Design Doc: float16
+
+## Why float16
+Half precision (float16) is a binary floating-point format that occupies 16 bits in memory. float16 is half the size of traditional 32-bit single precision format (float) and has lower precision and smaller range. 
+
+When high precision computation is not required, using float16 data type could potentially 
+
+- reduce storage space, memory bandwidth, and power usages; 
+- increase the chance of data fitting into a smaller cache of lower latency; 
+- provide arithmetic speed up if supported by hardware. 
+
+## Survey of current float16 support
+A brief survey of float16 support on different compilers, hardwares, and libraries can be found below. Interested readers can refer to [link1](https://github.com/PaddlePaddle/Paddle/issues/4853) and [link2](https://github.com/Xreki/Xreki.github.io/blob/master/multi_data_types_in_dl_framework/ppt/float16_and_quantized_type.md) for more info.
+
+The goal of float16 is to serve as a key for the executor to find and run the correct version of compute method specialized for float16 in operator kernel. It should be compatible with various natively supported float16 implementations including `__half` for cuda, `float16_t` for ARM, and `Eigen::half` for Eigen to make writing customized float16 kernels easier. 
+
+### Compiler
+- nvcc supports `__half` data type after CUDA 7.5.
+- `__fp16` or `float16_t` is supported as storage type for gcc >= 6.1 and clang >= 3.4.
+- `__fp16` or `float16_t` is supported as arithmetic type for gcc >= 7.1 and clang >= 3.9.
+
+### Hardware
+- `__half` is supported on GPU with compute capability >= 5.3.
+- `__fp16` is supported as storage type for ARMv7-A, ARMv8-A, and above.
+- `__fp16` is supported as arithmetic type after ARMv8.2-A (currently, the only microarchitecture implementing ARMv8.2-A is ARM Cortex-A75, which is announced in May 2017. There seems to be no application processors currently available on market that adopts this architecture. It is reported that Qualcomm Snapdragon 845 uses Cortex-A75 design and will be available in mobile devices in early 2018).
+
+### Libraries
+- [Eigen](https://github.com/RLovelett/eigen) >= 3.3 supports float16 calculation on both GPU and CPU using the `Eigen::half` class. It is mostly useful for Nvidia GPUs because of the overloaded arithmetic operators using cuda intrinsics. It falls back to using software emulation on CPU for calculation and there is no special treatment to ARM processors.
+- [ARM compute library](https://github.com/ARM-software/ComputeLibrary) >= 17.02.01 supports NEON FP16 kernels (requires ARMv8.2-A CPU).
+
+
+## Implementation
+The float16 class holds a 16-bit `uint16_t` data internally.
+```
+struct float16 {
+  uint16_t x;
+};
+``` 
+
+float16 supports the following features:
+  - constructors / assignment operators that take input from primitive data types including bool, integers of various length, float, and double. 
+  - constructors / assignment operators that take input from `__half` on cuda, `float16_t` on ARM, and `Eigen::half` on Eigen.
+  - conversion operators to primitive data types and half precision data types on cuda, ARM and Eigen. 
+  - overloaded arithmetic operators for cuda, arm, and non-arm cpu, respectively. These operators will take advantage of the cuda and ARM intrinsics on the corresponding hardware. 
+  
+To support the above features, two fundamental conversion functions are provided:
+```
+float16 float_to_half_rn(float f);  // convert to half precision in round-to-nearest-even mode
+float half_to_float(float16 h);
+```
+which provides one-to-one conversion between float32 and float16. These twos functions will do different conversion routines based on the current hardware. CUDA/ARM instrinsics will be used when the corresonding hardware is available. If the hardware or compiler level does not support float32 to float16 conversion, software emulation will be performed to do the conversion.
+
+## To do
+After float16 class is available, some of the future items are below:
+
+- Update pybind/tensor_py.h to bind c++ float16 with numpy float16. 
+
+- Modify `GetKernelType()` method in `framework/operator.h` to make it compatible with float16.
+
+- Create a type-casting operator that can convert the data type in tensor between float16 and other types.

doc/design/mkldnn/README.MD

@@ -15,6 +15,7 @@
  	- [CMake](#cmake)
 	- [Layers](#layers)
 	- [Activations](#activations)
+	- [Weights](#weights)
 	- [Unit Tests](#unit-tests)
 	- [Protobuf Messages](#protobuf-messages)
 	- [Python API](#python-api)
@@ -45,17 +46,23 @@ Figure 1. PaddlePaddle on IA.
 
 ### Layers
 所有MKL-DNN相关的C++ layers，都会按照PaddlePaddle的目录结构存放在
-`paddle/gserver/layers`中，并且文件名都会一以*Mkldnn*开头。
+`paddle/gserver/layers`中，并且文件名都会一以*MKLDNN*开头。
 
-所有MKL-DNN的layers都会继承于一个叫做`MkldnnLayer`的父类，该父类继承于PaddlePaddle的基类`Layer`。
+所有MKL-DNN的layers都会继承于一个叫做`MKLDNNLayer`的父类，该父类继承于PaddlePaddle的基类`Layer`。
+
+在`MKLDNNLayer`中会提供一些必要的接口和函数，并且会写好`forward`和`backward`的基本逻辑。部分函数定义为纯虚函数，子类只需要实现这些函数即可。
 
 ### Activations
-由于在PaddlePaddle中，激活函数是独立于layer概念的，所以会在`paddle/gserver/activations`目录下添加一个`MkldnnActivation.h`文件定义一些用于MKL-DNN的接口，实现方法还是会在`ActivationFunction.cpp`文件。
+由于在PaddlePaddle中，激活函数是独立于layer概念的，所以会在`paddle/gserver/activations`目录下添加`MKLDNNActivation.h`和`MKLDNNActivation.cpp`文件用于定义和使用MKL-DNN的接口。
 
-### Unit Tests
-会在`paddle/gserver/test`目录下添加`test_Mkldnn.cpp`和`MkldnnTester.*`用于MKL-DNN的测试。
+### Weights
+由于有些layer是含有参数的，我们会尽量让MKL-DNN的参数与PaddlePaddle中`parameter`共享一块内存。
+同时，由于MKL-DNN在训练时使用的参数layout可能与PaddlePaddle默认的`nchw`不一致，我们会在网络训练的开始和结束时分别转换这个layout，使得最终保存的参数格式与PaddlePaddle一致。
 
-Activation的测试，计划在PaddlePaddle原有的测试文件上直接添加新的测试type。
+### Unit Tests
+会在`paddle/gserver/test`目录下添加`test_MKLDNN.cpp`和`MKLDNNTester.*`用于MKL-DNN的测试。
+测试分为每个layer(或activation)的单元测试和简单网络的整体测试。
+每个测试会对比PaddlePaddle中CPU算出的结果与MKL-DNN的结果，小于某个比较小的阈值认为通过。
 
 ### Protobuf Messages
 根据具体layer的需求可能会在`proto/ModelConfig.proto`里面添加必要的选项。
@@ -82,7 +89,7 @@ if use_mkldnn
 会在`v1_api_demo`目录下添加一个`mkldnn`的文件夹，里面放入一些用于MKL-DNN测试的demo脚本。
 
 ### Benchmarking
-会考虑添加部分逻辑在`benchmark/paddle/image/run.sh`，添加使用MKL-DNN的测试。
+会添加`benchmark/paddle/image/run_mkldnn.sh`，用于测试使用MKL-DNN之后的性能。
 
 ### Others
 1. 如果在使用MKL-DNN的情况下，会把CPU的Buffer对齐为64。
@@ -94,14 +101,16 @@ if use_mkldnn
 
 我们总结出一些特别需要注意的点：
 
-1. 使用**deviceId_**。为了尽可能少的在父类Layer中添加变量或者函数，我们决定使用已有的`deviceId_`变量来区分layer的属性，定义`-2`为`MkldnnLayer`特有的设备ID。
+1. 使用**deviceId_**。为了尽可能少的在父类Layer中添加变量或者函数，我们决定使用已有的`deviceId_`变量来区分layer的属性，定义`-2`为`MKLDNNLayer`特有的设备ID。
 2. 重写父类Layer的**init**函数，修改`deviceId_`为`-2`，代表这个layer是用于跑在MKL-DNN的环境下。
-3. 创建`MkldnnMatrix`，用于管理MKL-DNN会用到的相关memory函数、接口以及会用的到格式信息。
-4. 创建`MkldnnBase`，定义一些除了layer和memory相关的类和函数。包括MKL-DNN会用到`MkldnnStream`和`CpuEngine`，和未来可能还会用到`FPGAEngine`等。
-5. 在**Argument**里添加两个`MkldnnMatrixPtr`，取名为`mkldnnValue`和`mkldnnGrad`，用于存放`MkldnnLayer`会用到的memory buffer。 并且添加函数cvt(会修改为一个更加合适的函数名)，用于处理"CPU device"和"MKL-DNN device"之间memory的相互转化。
-6. 在父类`Layer`中的`getOutput`函数中添加一段逻辑，用于判断`deviceId`，并针对device在MKL-DNN和CPU之间不统一的情况，做一个前期转换。 也就是调用`Argument`的cvt函数把output统一到需要的device上。
-7. 在原来的`FLAGS`中添加一个`use_mkldnn`的flag，用于选择是否使用MKL-DNN的相关功能。
-8. 关于MKLDNN参数的保存。由于MKLDNN参数的格式与PaddlePaddle原有的格式存在不一样的情况，所以需要在保存参数时同时保存该格式信息。目前准备扩展[Header](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/parameter/Parameter.h#L247)里面的`int32_t version`。这个值不管是在v1还是在v2里面，一直保存的是0，所以可以充分利用这个信息，定义一个枚举处理所有MKLDNN的参数格式，从而`MKLDNNLayer`就可以从输入的参数中获取需要的格式信息。
+3. 创建`MKLDNNMatrix`，同时继承`CpuMatrix`和`mkldnn::memory`。用于管理MKL-DNN会用到的相关memory函数、接口以及会用的到格式信息。
+4. 创建`MKLDNNBase`，定义一些除了layer和memory相关的类和函数。包括MKL-DNN会用到`MKLDNNStream`和`CPUEngine`，和未来可能还会用到`FPGAEngine`等。
+5. 每个`MKLDNNlayer`都会有`inVal_`,`inGrad_`,`outVal_`和`outGrad_`，分别代表input value， input gradient，output value和output gradient。他们会存放MKL-DNN用到的internal memory。同时还会定义以*ext*开头的`MKLDNNMatrix`(表示external的memory)，主要是在格式与PaddlePaddle默认的`nchw`格式不匹配时，用于转换内存的工作。必要的转换函数也会在`MKLDNNLayer`中提前定义好，每个子类只需要调用定义好的reset buffer函数即可。
+6. 每个`MKLDNNlayer`的resetbuffer相关的函数（包括reset input、output的Value和grad），他们会根据输入参数reset internal和external的memory，当然这两者也可以相等，即表示不需要转换。只需要把握一个原则，每个`MKLDNNlayer`的子类，只需要使用internal的memory就可以了，所有external的转换工作在父类的reset函数中都提前准备好了。
+7. 一般来说，external的memory会尽量与PaddlePaddle中的`value`和`grad`共享内存。同时每个`MKLDNNLayer`中的external output value和gradient(也就是`extOutVal_`和`extOutGrad_`)必须分别与`output_.value`和`output_.grad`共享内存，因为PaddlePaddle的activation会直接使用`output_.value`和`output_.grad`。如果不需要external的buffer用于转换，那么internal的buffer也会与他们共享内存。
+8. 如果MKL-DNN layer的后面接有cpu device，那么就会使`output_.value`与`extOutVal_`共享内存，同时数据格式就是`nchw`，这样下一个cpu device就能拿到正确的数据。在有cpu device的时候，external的memory的格式始终是`nchw`或者`nc`。
+9. 由于MKL-DNN的输出操作都是覆盖data的，不是在原来的数据上累加，所以当网络出现分支时，在`backward`时会需要merge不同layer的梯度。`MKLDNNlayer`中会实现merge的方法，此时每个小分支的input gradient会先临时保存在一个`MKLDNNMatrix`中，由分支处的layer负责求和，并把结果放到这个layer的`output_.grad`中。所以整体上，每个子类并不会需要关心分支的事情，也是在父类都实现好了。
+10. 在原来的`FLAGS`中添加一个`use_mkldnn`的flag，用于选择是否使用MKL-DNN的相关功能。
 
 ## References
 

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.
+
+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.
+
+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.
+
+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** .
+
+For example, the RNN sates, candidates IDs and probabilities of beam search can be represented 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 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 relative-offset LoD can fast retrieve any sequence 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]]
+```
+The first level tells that there are two sequences:
+- the first's offset is `[0, 3)`
+- the second's offset is `[3, 9)`
+
+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.
+
+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
+
+```python
+[[0, 3, 6]
+ [0, 2, 3, 3, 3, 9]]
+```
+
+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.
+
+## 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.
+
+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.
+
+**Encoder**
+```python
+import paddle as pd
+
+dict_size = 8000
+source_dict_size = dict_size
+target_dict_size = dict_size
+word_vector_dim = 128
+encoder_dim = 128
+decoder_dim = 128
+beam_size = 5
+max_length = 120
+
+# encoder
+src_word_id = pd.data(
+    name='source_language_word',
+    type=pd.data.integer_value_sequence(source_dict_dim))
+src_embedding = pd.embedding(size=source_dict_size, size=word_vector_dim)
+
+src_word_vec = pd.lookup(src_embedding, src_word_id)
+
+encoder_out_seq = pd.gru(input=src_word_vec, size=encoder_dim)
+
+encoder_ctx = pd.last_seq(encoder_out_seq)
+# encoder_ctx_proj is the learned semantic vector
+encoder_ctx_proj = pd.fc(
+    encoder_ctx, size=decoder_dim, act=pd.activation.Tanh(), bias=None)
+```
+
+**Decoder**
+
+```python
+def generate():
+    decoder = pd.while_loop()
+    with decoder.step():
+        decoder_mem = decoder.memory(init=encoder_ctx)  # mark the memory
+        generated_ids = decoder.memory() # TODO init to batch_size <s>s
+        generated_scores = decoder.memory() # TODO init to batch_size 1s or 0s
+
+        target_word = pd.lookup(trg_embedding, gendrated_ids)
+        # expand encoder_ctx's batch to fit target_word's lod
+        # for example
+        # decoder_mem.lod is
+        # [[0 1 3],
+        #  [0 1 3 6]]
+        # its tensor content is [a1 a2 a3 a4 a5]
+        # 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 
+        # [[0, 1, 6]
+        #  [0, 2, 4, 7, 9 12]]
+        # which means 2 sentences to translate, each has 1 and 5 prefixes
+        # the first prefix has 2 candidates
+        # the following has 2, 3, 2, 3 candidates
+        # the encoder_ctx_expanded's content will be
+        # [a1 a1 a2 a2 a3 a3 a3 a4 a4 a5 a5 a5]
+        encoder_ctx_expanded = pd.lod_expand(encoder_ctx, target_word)
+        decoder_input = pd.fc(
+            act=pd.activation.Linear(),
+            input=[target_word, encoder_ctx],
+            size=3 * decoder_dim)
+        gru_out, cur_mem = pd.gru_step(
+            decoder_input, mem=decoder_mem, size=decoder_dim)
+        scores = pd.fc(
+            gru_out,
+            size=trg_dic_size,
+            bias=None,
+            act=pd.activation.Softmax())
+        # K is an config
+        topk_scores, topk_ids = pd.top_k(scores, K)
+        topk_generated_scores = pd.add_scalar(topk_scores, generated_scores)
+
+        selected_ids, selected_generation_scores = decoder.beam_search(
+            topk_ids, topk_generated_scores)
+
+        # update the states
+        decoder_mem.update(cur_mem)  # tells how to update state
+        generated_ids.update(selected_ids)
+        generated_scores.update(selected_generation_scores)
+
+        decoder.output(selected_ids)
+        decoder.output(selected_generation_scores)
+
+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.
+
+In this way, users can customize anything on the inputs or outputs of beam search, for example, two ways to prune some translation prefixes
+
+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.
+
+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).
+
+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.
+
+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.
+
+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`
+
+* LoD is `[0, 2, 7][0 3 5 8 9 11 11]`
+* 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
+  - b2
+  - 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 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.
+
+Pack the `selected_scores` will get a `LoDTensor` that stores scores of each candidate of translations.
+
+Pack 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.
+
+## Beam search design
+The beam search algorthm will be implemented as one method of the sequence decoder, it has 3 inputs
+
+1. `topk_ids`, 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.
+
+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).
+
+## 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,
+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`.
+
+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`.

doc/faq/local/index_cn.rst

@@ -99,7 +99,7 @@ PaddlePaddle支持Sparse的训练，sparse训练需要训练特征是 :code:`spa
 利用更多的计算资源
 ++++++++++++++++++
 
-利用更多的计算资源可以分为一下几个方式来进行\:
+利用更多的计算资源可以分为以下几个方式来进行\:
 
 * 单机CPU训练
 

doc/getstarted/build_and_install/docker_install_cn.rst

@@ -145,7 +145,7 @@ PaddlePaddle发布新版本的时候都会发布对应版本的生产镜像以
 
 Jupyter Notebook是一个开源的web程序，大家可以通过它制作和分享带有代码、公式、图表、文字的交互式文档。用户可以通过网页浏览文档。
 
-PaddlePaddle Book是为用户和开发者制作的一个交互式的Jupyter Nodebook。
+PaddlePaddle Book是为用户和开发者制作的一个交互式的Jupyter Notebook。
 如果您想要更深入了解deep learning，PaddlePaddle Book一定是您最好的选择。
 
 我们提供可以直接运行PaddlePaddle Book的Docker镜像，直接运行：

doc/howto/dev/new_op_cn.md

@@ -214,7 +214,7 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs,
 
     ```cpp
     // if use Eigen unsupported module before include head files
-    #define EIGEN_USE_GPU
+    // #define EIGEN_USE_GPU
 
     namespace ops = paddle::operators;
     REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel<paddle::platform::GPUPlace, float>);

doc/howto/usage/cmd_parameter/arguments_cn.md

@@ -63,7 +63,7 @@
 </tr>
 
 <tr>
-<td class="left" rowspan="15">训练</td><td class="left">dot_period</td>
+<td class="left" rowspan="14">训练</td><td class="left">dot_period</td>
 <td class="left">√</td><td class="left">√</td><td class="left"></td><td class="left"></td>
 </tr>
 

doc/index_cn.rst

@@ -8,3 +8,4 @@ PaddlePaddle 文档
   howto/index_cn.rst
   api/index_cn.rst
   faq/index_cn.rst
+  mobile/index_cn.rst

doc/index_en.rst

@@ -7,3 +7,4 @@ PaddlePaddle Documentation
   getstarted/index_en.rst
   howto/index_en.rst
   api/index_en.rst
+  mobile/index_en.rst

doc/howto/cross_compiling/cross_compiling_for_android_cn.md → doc/mobile/cross_compiling_for_android_cn.md

@@ -20,10 +20,32 @@ $ docker build -t username/paddle-android:dev . -f Dockerfile.android
 构建好开发镜像后，即可使用开发镜像来编译Android版PaddlePaddle C-API库。
 Android的Docker开发镜像向用户提供两个可配置的参数：
 
-| Argument        | Optional Values         | Default |
-|-----------------|-------------------------|---------|
-|`ANDROID_ABI`    |`armeabi-v7a, arm64-v8a` | `armeabi-v7a` |
-|`ANDROID_API`    |`>= 21` | `21` |
+<table class="docutils">
+<colgroup>
+  <col width="25%" />
+  <col width="50%" />
+  <col width="25%" />
+</colgroup>
+<thead valign="bottom">
+  <tr class="row-odd">
+  <th class="head">Argument</th>
+  <th class="head">Optional Values</th>
+  <th class="head">Default</th>
+</tr>
+</thead>
+<tbody valign="top">
+  <tr class="row-even">
+  <td>ANDROID_ABI</td>
+  <td>armeabi-v7a, arm64-v8a</td>
+  <td>armeabi-v7a</td>
+</tr>
+<tr class="row-odd">
+  <td>ANDROID_API</td>
+  <td>>= 21</td>
+  <td>21</td>
+</tr>
+</tbody>
+</table>
 
 - 编译`armeabi-v7a`，`Android API 21`的PaddlePaddle库
   ```bash

doc/howto/cross_compiling/cross_compiling_for_android.md → doc/mobile/cross_compiling_for_android_en.md

@@ -26,10 +26,32 @@ $ docker run -it --rm -v $PWD:/paddle -e "ANDROID_ABI=armeabi-v7a" -e "ANDROID_A
 
 The Docker image accepts two arguments `ANDROID_ABI` and `ANDROID_API`:
 
-| Argument        | Optional Values         | Default |
-|-----------------|-------------------------|---------|
-|`ANDROID_ABI`    |`armeabi-v7a, arm64-v8a` | `armeabi-v7a` |
-|`ANDROID_API`    |`>= 21` | `21` |
+<table class="docutils">
+<colgroup>
+  <col width="25%" />
+  <col width="50%" />
+  <col width="25%" />
+</colgroup>
+<thead valign="bottom">
+  <tr class="row-odd">
+  <th class="head">Argument</th>
+  <th class="head">Optional Values</th>
+  <th class="head">Default</th>
+</tr>
+</thead>
+<tbody valign="top">
+  <tr class="row-even">
+  <td>ANDROID_ABI</td>
+  <td>armeabi-v7a, arm64-v8a</td>
+  <td>armeabi-v7a</td>
+</tr>
+<tr class="row-odd">
+  <td>ANDROID_API</td>
+  <td>>= 21</td>
+  <td>21</td>
+</tr>
+</tbody>
+</table>
 
 The ARM-64 architecture (`arm64-v8a`) requires at least level 21 of Android API.
 

doc/howto/cross_compiling/cross_compiling_for_ios_cn.md → doc/mobile/cross_compiling_for_ios_cn.md

@@ -27,10 +27,28 @@ iOS平台可选配置参数：
   - `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_PLATFORM | IOS_ARCH             |
-   |--------------|----------------------|
-   |   OS         | armv7, armv7s, arm64 (默认) |
-   | SIMULATOR    | i386, x86_64 (默认)         |   
+    <table class="docutils">
+    <colgroup>
+      <col width="35%" />
+      <col width="65%" />
+    </colgroup>
+    <thead valign="bottom">
+      <tr class="row-odd">
+      <th class="head">IOS_PLATFORM</th>
+      <th class="head">IOS_ARCH</th>
+    </tr>
+    </thead>
+    <tbody valign="top">
+      <tr class="row-even">
+      <td>OS</td>
+      <td>armv7, armv7s, arm64 (默认)</td>
+    </tr>
+    <tr class="row-odd">
+      <td>SIMULATOR</td>
+      <td>i386, x86_64 (默认)</td>
+    </tr>
+    </tbody>
+    </table>
 
 - `IOS_DEPLOYMENT_TARGET`，最小的iOS部署版本，默认值为`7.0`。
 - `IOS_ENABLE_BITCODE`，是否使能[Bitcode](https://developer.apple.com/library/content/documentation/IDEs/Conceptual/AppDistributionGuide/AppThinning/AppThinning.html#//apple_ref/doc/uid/TP40012582-CH35-SW3)，可设置`ON/OFF`，默认值为`ON`。

doc/mobile/index_cn.rst

+MOBILE
+======
+
+..  toctree::
+  :maxdepth: 1
+
+  cross_compiling_for_android_cn.md
+  cross_compiling_for_ios_cn.md
+  cross_compiling_for_raspberry_cn.md

doc/mobile/index_en.rst

+MOBILE
+======
+
+..  toctree::
+  :maxdepth: 1
+
+  cross_compiling_for_android_en.md
+  cross_compiling_for_raspberry_en.md

paddle/capi/CMakeLists.txt

@@ -29,32 +29,32 @@ add_style_check_target(paddle_capi ${CAPI_SOURCES} ${CAPI_HEADER}
 add_dependencies(paddle_capi paddle_proto)
 
 # TODO: paddle_capi_whole will be removed.
+set(PADDLE_CAPI_LAYERS_LIBS
+    paddle_function
+    paddle_gserver)
 if(MOBILE_INFERENCE)
-    set(PADDLE_CAPI_INFER_LIBS
+  set(PADDLE_CAPI_ENGINE_LIBS
       paddle_utils
       paddle_parameter
       paddle_math
       paddle_cuda
-        paddle_function
-        paddle_gserver
       paddle_proto)
 else()
-    set(PADDLE_CAPI_INFER_LIBS
+  set(PADDLE_CAPI_ENGINE_LIBS
       paddle_utils
       paddle_parameter
       paddle_math
       paddle_cuda
-        paddle_function
-        paddle_gserver
       paddle_proto
       paddle_pserver
       paddle_network)
 endif()
+set(PADDLE_CAPI_INFER_LIBS ${PADDLE_CAPI_LAYERS_LIBS} ${PADDLE_CAPI_ENGINE_LIBS})
 cc_library(paddle_capi_whole DEPS paddle_capi ${PADDLE_CAPI_INFER_LIBS})
 
 # Link the static library for inference
-cc_library(paddle_capi_engine DEPS paddle_capi paddle_utils paddle_parameter paddle_math paddle_cuda paddle_proto)
-cc_library(paddle_capi_layers DEPS paddle_function paddle_gserver)
+cc_library(paddle_capi_engine DEPS paddle_capi ${PADDLE_CAPI_ENGINE_LIBS})
+cc_library(paddle_capi_layers DEPS ${PADDLE_CAPI_LAYERS_LIBS})
 
 # Link the shared library for inference
 if(NOT IOS)

paddle/capi/Matrix.cpp

@@ -54,6 +54,46 @@ paddle_error paddle_matrix_set_row(paddle_matrix mat,
   return kPD_NO_ERROR;
 }
 
+PD_API paddle_error paddle_matrix_set_value(paddle_matrix mat,
+                                          paddle_real* value) {
+  if (mat == nullptr || value == nullptr) return kPD_NULLPTR;
+  auto ptr = cast(mat);
+  if (ptr->mat == nullptr) return kPD_NULLPTR;
+  paddle::real* buf = ptr->mat->getRowBuf(0);
+  size_t width = ptr->mat->getWidth();
+  size_t height = ptr->mat->getHeight();
+  if (ptr->mat->useGpu()) {
+#ifdef PADDLE_WITH_CUDA
+    hl_memcpy(buf, value, sizeof(paddle::real) * width * height);
+#else
+    return kPD_NOT_SUPPORTED;
+#endif
+  } else {
+    std::copy(value, value + width * height, buf);
+  }
+  return kPD_NO_ERROR;
+}
+
+PD_API paddle_error paddle_matrix_get_value(paddle_matrix mat,
+                                          paddle_real* result) {
+  if (mat == nullptr || result == nullptr) return kPD_NULLPTR;
+  auto ptr = cast(mat);
+  if (ptr->mat == nullptr) return kPD_NULLPTR;
+  paddle::real* buf = ptr->mat->getRowBuf(0);
+  size_t width = ptr->mat->getWidth();
+  size_t height = ptr->mat->getHeight();
+  if (ptr->mat->useGpu()) {
+#ifdef PADDLE_WITH_CUDA
+    hl_memcpy(result, buf, width * height * sizeof(paddle::real));
+#else
+    return kPD_NOT_SUPPORTED;
+#endif
+  } else {
+    std::copy(buf, buf + width * height, result);
+  }
+  return kPD_NO_ERROR;
+}
+
 paddle_error paddle_matrix_get_row(paddle_matrix mat,
                                    uint64_t rowID,
                                    paddle_real** rawRowBuffer) {

paddle/capi/examples/model_inference/dense/main.c

@@ -27,19 +27,21 @@ int main() {
   CHECK(paddle_arguments_resize(in_args, 1));
 
   // Create input matrix.
-  paddle_matrix mat = paddle_matrix_create(/* sample_num */ 1,
+  paddle_matrix mat = paddle_matrix_create(/* sample_num */ 10,
                                            /* size */ 784,
                                            /* useGPU */ false);
   srand(time(0));
-  paddle_real* array;
 
-  // Get First row.
-  CHECK(paddle_matrix_get_row(mat, 0, &array));
+  std::vector<paddle_real> input;
+  input.resize(784 * 10);
 
-  for (int i = 0; i < 784; ++i) {
-    array[i] = rand() / ((float)RAND_MAX);
+  for (int i = 0; i < input.size(); ++i) {
+    input[i] = rand() / ((float)RAND_MAX);
   }
   
+  // Set value for the input matrix
+  CHECK(paddle_matrix_set_value(mat, input.data()));
+
   CHECK(paddle_arguments_set_value(in_args, 0, mat));
 
   paddle_arguments out_args = paddle_arguments_create_none();
@@ -51,11 +53,17 @@ int main() {
 
   CHECK(paddle_arguments_get_value(out_args, 0, prob));
 
-  CHECK(paddle_matrix_get_row(prob, 0, &array));
+  std::std::vector<paddle_real> result;
+  int height;
+  int width;
+
+  CHECK(paddle_matrix_get_shape(prob, &height, &width);
+  result.resize(height * width);
+  CHECK(paddle_matrix_get_value(prob, result.data()));
 
   printf("Prob: ");
-  for (int i = 0; i < 10; ++i) {
-    printf("%.2f ", array[i]);
+  for (int i = 0; i < height * width; ++i) {
+    printf("%.2f ", result[i]);
   }
   printf("\n");
 

paddle/capi/matrix.h

@@ -70,6 +70,16 @@ PD_API paddle_error paddle_matrix_set_row(paddle_matrix mat,
                                           uint64_t rowID,
                                           paddle_real* rowArray);
 
+/**
+ * @brief paddle_matrix_set_value Set value to matrix.
+ * @param mat Target Matrix
+ * @param value Row data.
+ * @return paddle_error
+ * @note  value should contain enough element of data to init the mat
+ */
+PD_API paddle_error paddle_matrix_set_value(paddle_matrix mat,
+                                          paddle_real* value);
+
 /**
  * @brief PDMatGetRow Get raw row buffer from matrix
  * @param [in] mat Target matrix
@@ -81,6 +91,15 @@ PD_API paddle_error paddle_matrix_get_row(paddle_matrix mat,
                                           uint64_t rowID,
                                           paddle_real** rawRowBuffer);
 
+/**
+ * @brief copy data from the matrix 
+ * @param [in] mat Target matrix
+ * @param [out] result pointer to store the matrix data 
+ * @return paddle_error
+ * @note the space of the result should allocated before invoke this API
+ */
+PD_API paddle_error paddle_matrix_get_value(paddle_matrix mat,
+                                          paddle_real* result);
 /**
  * @brief PDMatCreateNone Create None Matrix
  * @return

paddle/capi/tests/test_Matrix.cpp

@@ -45,3 +45,49 @@ TEST(CAPIMatrix, createNone) {
   paddle_matrix mat = paddle_matrix_create_none();
   ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_destroy(mat));
 }
+
+TEST(CAPIMatrix, cpu_get_set_value) {
+  paddle_matrix mat = paddle_matrix_create(128, 32, false);
+  std::vector<paddle_real> sample;
+  std::vector<paddle_real> result;
+  sample.resize(128 * 32);
+  result.resize(128 * 32);
+  for (size_t i = 0; i < sample.size(); ++i) {
+    sample[i] = 1.0 / (i + 1.0);
+  }
+  ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_set_value(mat, sample.data()));
+  ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_get_value(mat, result.data()));
+  for (size_t i = 0; i < sample.size(); ++i) {
+    ASSERT_NEAR(sample[i], result[i], 1e-5);
+  }
+
+  uint64_t height, width;
+  ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_get_shape(mat, &height, &width));
+  ASSERT_EQ(128UL, height);
+  ASSERT_EQ(32UL, width);
+  ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_destroy(mat));
+}
+
+#ifdef PADDLE_WITH_CUDA
+TEST(CAPIMatrix, gpu_get_set_value) {
+  paddle_matrix mat = paddle_matrix_create(128, 32, true);
+  std::vector<paddle_real> sample;
+  std::vector<paddle_real> result;
+  sample.resize(128 * 32);
+  result.resize(128 * 32);
+  for (size_t i = 0; i < sample.size(); ++i) {
+    sample[i] = 1.0 / (i + 1.0);
+  }
+  ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_set_value(mat, sample.data()));
+  ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_get_value(mat, result.data()));
+  for (size_t i = 0; i < sample.size(); ++i) {
+    ASSERT_NEAR(sample[i], result[i], 1e-5);
+  }
+
+  uint64_t height, width;
+  ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_get_shape(mat, &height, &width));
+  ASSERT_EQ(128UL, height);
+  ASSERT_EQ(32UL, width);
+  ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_destroy(mat));
+}
+#endif

paddle/framework/backward.cc

@@ -321,8 +321,6 @@ static void CreateGradVarInBlock(
         auto* param = block_desc->FindVarRecursive(pname);
         auto* grad = block_desc->FindVar(arg);
         if (param == nullptr) {
-          LOG(WARNING) << "Cannot find forward variable of " << arg
-                       << ". Set its gradient to FP32";
           grad->SetDataType(DataType::FP32);
         } else {
           grad->SetDataType(param->GetDataType());
@@ -408,6 +406,11 @@ std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
 
     for (const auto& desc : op_grads) {
       for (const std::string& out_name : desc->OutputArgumentNames()) {
+        if (out_name.find("@GRAD") == std::string::npos) {
+          // Not all outputs of a backward operator is a gradient. Only gradient
+          // need to be sum. Skip variables are not gradient.
+          continue;
+        }
         dup_out_ops[out_name].emplace_back(grad_desc_idx);
       }
       ++grad_desc_idx;

paddle/framework/backward_test.cc

@@ -21,7 +21,7 @@
 #include "paddle/framework/var_desc.h"
 #include "paddle/operators/net_op.h"
 
-USE_OP(fill_constant);
+USE_NO_KERNEL_OP(fill_constant);
 
 namespace paddle {
 namespace framework {

paddle/framework/block_desc.cc

@@ -50,6 +50,15 @@ VarDescBind *BlockDescBind::FindVarRecursive(const std::string &name) const {
   return it->second.get();
 }
 
+VarDescBind *BlockDescBind::FindRecursiveOrCreateVar(
+    const std::string &name_bytes) {
+  VarDescBind *res = FindVarRecursive(name_bytes);
+  if (res == nullptr) {
+    res = Var(name_bytes);
+  }
+  return res;
+}
+
 bool BlockDescBind::HasVarRecursive(const std::string &name) const {
   return FindVarRecursive(name) != nullptr;
 }

paddle/framework/block_desc.h

@@ -58,6 +58,8 @@ class BlockDescBind {
 
   VarDescBind *FindVarRecursive(const std::string &name_bytes) const;
 
+  VarDescBind *FindRecursiveOrCreateVar(const std::string &name_bytes);
+
   bool HasVarRecursive(const std::string &var_name) const;
 
   std::set<std::string> LocalVarNames() const {

paddle/framework/data_type.h

@@ -34,6 +34,21 @@ inline DataType ToDataType(std::type_index type) {
   }
 }
 
+inline std::type_index ToTypeIndex(DataType type) {
+  switch (type) {
+    case DataType::FP32:
+      return typeid(float);
+    case DataType::FP64:
+      return typeid(double);
+    case DataType::INT32:
+      return typeid(int);
+    case DataType::INT64:
+      return typeid(int64_t);
+    default:
+      PADDLE_THROW("Not support type %d", type);
+  }
+}
+
 template <typename Visitor>
 inline void VisitDataType(DataType type, Visitor visitor) {
   switch (type) {

paddle/framework/ddim.cc

@@ -79,6 +79,13 @@ DDim make_ddim(const std::vector<int64_t>& dims) {
   return result;
 }
 
+DDim make_ddim(const std::vector<int>& dims) {
+  std::vector<int64_t> res(dims.size());
+  std::transform(dims.begin(), dims.end(), res.begin(),
+                 [](int d) { return static_cast<int64_t>(d); });
+  return make_ddim(res);
+}
+
 /// @cond HIDDEN
 // XXX For some reason, putting this in an anonymous namespace causes errors
 class DynamicMutableIndexer : public boost::static_visitor<int64_t&> {
@@ -117,7 +124,7 @@ int64_t DDim::operator[](int idx) const {
   return boost::apply_visitor(DynamicConstIndexer(idx), var);
 }
 
-int64_t DDim::size() const { return arity(*this); }
+int DDim::size() const { return arity(*this); }
 
 bool DDim::operator==(DDim d) const {
   if (var.which() != d.getVar().which()) {

paddle/framework/ddim.h

@@ -71,7 +71,7 @@ struct DDim {
 
   DDim operator*(DDim d) const;
 
-  int64_t size() const;
+  int size() const;
 };
 
 /**
@@ -81,6 +81,8 @@ struct DDim {
  */
 DDim make_ddim(const std::vector<int64_t>& dims);
 
+DDim make_ddim(const std::vector<int>& dims);
+
 /**
  * \brief Make a DDim from an initializer list
  *

paddle/framework/executor.cc

@@ -23,6 +23,7 @@ limitations under the License. */
 #include "paddle/framework/feed_fetch_type.h"
 #include "paddle/framework/lod_rank_table.h"
 #include "paddle/framework/lod_tensor.h"
+#include "paddle/framework/lod_tensor_array.h"
 #include "paddle/framework/op_registry.h"
 #include "paddle/framework/scope.h"
 
@@ -73,6 +74,8 @@ static void CreateTensor(Variable* var, VarDesc::VarType var_type) {
     var->GetMutable<std::vector<framework::Scope>>();
   } else if (var_type == VarDesc::LOD_RANK_TABLE) {
     var->GetMutable<LoDRankTable>();
+  } else if (var_type == VarDesc::LOD_TENSOR_ARRAY) {
+    var->GetMutable<LoDTensorArray>();
   } else {
     PADDLE_THROW(
         "Variable type %d is not in "

paddle/framework/framework.proto

@@ -109,6 +109,11 @@ message LoDTensorDesc {
   optional int32 lod_level = 2 [ default = 0 ];
 }
 
+message LoDTensorArrayDesc {
+  required TensorDesc tensor = 1;
+  optional int32 lod_level = 2 [ default = 0 ];
+}
+
 message VarDesc {
   enum VarType {
     LOD_TENSOR = 1;
@@ -117,11 +122,13 @@ message VarDesc {
     FETCH_LIST = 4;
     STEP_SCOPES = 5;
     LOD_RANK_TABLE = 6;
+    LOD_TENSOR_ARRAY = 7;
   }
   required string name = 1;
   required VarType type = 2;
   optional LoDTensorDesc lod_tensor = 3;
   optional TensorDesc selected_rows = 4;
+  optional LoDTensorArrayDesc tensor_array = 6;
   optional bool persistable = 5 [ default = false ];
 }
 

paddle/framework/lod_rank_table.cc

@@ -31,9 +31,15 @@ void LoDRankTable::Reset(const LoD& lod, size_t level) {
     TableItem item;
     item.index = i;
     item.length = vec[i + 1] - vec[i];
+    VLOG(10) << "Add item to rank table " << item.index << " " << item.length;
     items_.emplace_back(item);
   }
-  std::sort(items_.begin(), items_.end(),
+  // NOTE(yuyang18):
+  //
+  // The time complexity of stable_sort is O(N*log(N)) if additional memory is
+  // available. It is easy to debug and unit test when using `stable_sort`
+  // instead of `sort`. Also, the items of a rank table will not be too large.
+  std::stable_sort(items_.begin(), items_.end(),
                    [](const TableItem& a, const TableItem& b) {
                      return a.length > b.length;
                    });

paddle/framework/lod_tensor.cc

@@ -27,6 +27,20 @@
 namespace paddle {
 namespace framework {
 
+std::ostream& operator<<(std::ostream& os, const LoD& lod) {
+  os << "{";
+  for (auto& v : lod) {
+    os << "{";
+    for (auto& i : v) {
+      os << i << ",";
+    }
+    os << "}";
+  }
+  os << "}";
+
+  return os;
+}
+
 LoD SliceLevels(const LoD& in, size_t level_begin, size_t level_end) {
   LoD new_lod;
   new_lod.reserve(level_end - level_begin);
@@ -135,5 +149,41 @@ void LoDTensor::ShrinkInLevel(size_t level, size_t elem_begin,
   PADDLE_ENFORCE_LT(begin, end, "Cannot shrink, the result tensor is empty.");
   ShareDataWith(Slice(begin, end));
 }
+
+using LoDAndOffset = std::pair<LoD, std::pair<size_t, size_t>>;
+LoDAndOffset GetSubLoDAndAbsoluteOffset(const LoD& lod, size_t start_idx,
+                                        size_t end_idx, size_t start_level) {
+  LoD sub_lod;
+
+  for (size_t level_idx = start_level; level_idx < lod.size(); ++level_idx) {
+    PADDLE_ENFORCE_LE(start_idx, end_idx);
+    PADDLE_ENFORCE_LT(end_idx, lod[level_idx].size());
+    std::vector<size_t> level_lens;
+    for (size_t i = start_idx; i < end_idx; ++i) {
+      level_lens.push_back(lod[level_idx][i + 1] - lod[level_idx][i]);
+    }
+    sub_lod.emplace_back(level_lens);
+    start_idx = lod[level_idx][start_idx];
+    end_idx = lod[level_idx][end_idx];
+  }
+
+  return LoDAndOffset{sub_lod, {start_idx, end_idx}};
+}
+
+void AppendLoD(LoD* lod, const LoD& lod_length) {
+  PADDLE_ENFORCE(
+      lod->empty() || lod->size() == lod_length.size(),
+      "The lod_length should has the same size with the appended lod.");
+  if (lod->empty()) {
+    *lod = LoD(lod_length.size(), std::vector<size_t>({0}));
+  }
+  for (size_t i = 0; i < lod->size(); ++i) {
+    auto& level = (*lod)[i];
+    for (size_t len : lod_length[i]) {
+      level.push_back(level.back() + len);
+    }
+  }
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/lod_tensor.h

@@ -56,6 +56,8 @@ using Vector = thrust::host_vector<
  */
 using LoD = std::vector<Vector<size_t>>;
 
+std::ostream& operator<<(std::ostream& os, const LoD& lod);
+
 /*
  * Slice levels from a LoD.
  * NOTE the lowest level should always be the absolute offsets of the underlying
@@ -181,5 +183,10 @@ LoDTensor LodExpand(const LoDTensor& source, const LoD& lod, size_t level,
   return tensor;
 }
 
+std::pair<LoD, std::pair<size_t, size_t>> GetSubLoDAndAbsoluteOffset(
+    const LoD& lod, size_t start_idx, size_t end_idx, size_t start_level);
+
+void AppendLoD(LoD* lod, const LoD& lod_length);
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/lod_tensor.md

@@ -140,19 +140,9 @@ Similarly, the lengths in the top level LoD
 are transformed into offsets of elements/words as follows:
 
 ```
-0 9     10  15
+0 3 4   6
   = =   =
-  3+2+4 1+9 2+3+10
-```
-
-so we can tell that the first article is from word 0 to word 9, and the second article is from word 9 to word 10.
-
-The complete offset representation is as follows:
-
-```
-0           9 10       15
-0   3  5    9 10  12   15
- ||| || |||| |  ||  |||
+  3 3+1 4+2
 ```
 
 ## Slicing of LoD Tensors

paddle/framework/lod_tensor_array.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#pragma once
+#include <vector>
+#include "paddle/framework/lod_tensor.h"
+
+namespace paddle {
+namespace framework {
+using LoDTensorArray = std::vector<LoDTensor>;
+}
+}  // namespace paddle

paddle/framework/lod_tensor_test.cc

@@ -144,5 +144,48 @@ TEST(LodExpand, test) {
   }
 }
 
+TEST(LoD, GetFineGrainedLoDLength) {
+  LoD lod;
+  lod.push_back(std::vector<size_t>({0, 2, 4, 5}));
+  lod.push_back(std::vector<size_t>({0, 1, 6, 8, 10, 11}));
+  lod.push_back(
+      std::vector<size_t>({0, 2, 5, 7, 10, 12, 15, 17, 20, 24, 26, 29}));
+
+  auto lod_and_offset =
+      paddle::framework::GetSubLoDAndAbsoluteOffset(lod, 1, 2, 0);
+  LoD lod_length = lod_and_offset.first;
+  size_t start_offset = lod_and_offset.second.first;
+  size_t end_offset = lod_and_offset.second.second;
+
+  LoD expected;
+  expected.push_back(std::vector<size_t>{2});
+  expected.push_back(std::vector<size_t>{2, 2});
+  expected.push_back(std::vector<size_t>{2, 3, 4, 2});
+  EXPECT_EQ(lod_length, expected);
+  EXPECT_EQ(start_offset, 15UL);
+  EXPECT_EQ(end_offset, 26UL);
+}
+
+TEST(LoD, AppendLoD) {
+  LoD lod_lens;
+  lod_lens.push_back(std::vector<size_t>({2}));
+  lod_lens.push_back(std::vector<size_t>({2, 2}));
+  lod_lens.push_back(std::vector<size_t>({2, 3, 4, 2}));
+
+  LoD origin;
+  origin.push_back(std::vector<size_t>({0, 2}));
+  origin.push_back(std::vector<size_t>({0, 1, 6}));
+  origin.push_back(std::vector<size_t>({0, 2, 5, 7, 10, 12, 15}));
+
+  paddle::framework::AppendLoD(&origin, lod_lens);
+
+  LoD expected;
+  expected.push_back(std::vector<size_t>({0, 2, 4}));
+  expected.push_back(std::vector<size_t>({0, 1, 6, 8, 10}));
+  expected.push_back(
+      std::vector<size_t>({0, 2, 5, 7, 10, 12, 15, 17, 20, 24, 26}));
+  EXPECT_EQ(origin, expected);
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/op_desc.cc

@@ -67,8 +67,11 @@ class CompileTimeInferShapeContext : public InferShapeContext {
                       out);
     in_var->SetLoDLevel(out_var->GetLodLevel());
   }
+  bool IsRuntime() const override;
+
+ protected:
+  VarDesc::VarType GetVarType(const std::string &name) const override;
 
- private:
   DDim GetDim(const std::string &name) const override;
 
   void SetDim(const std::string &name, const DDim &dim) override;
@@ -349,9 +352,13 @@ void OpDescBind::InferVarType(BlockDescBind *block) const {
     info.infer_var_type_(*this, block);
   } else {
     // all output type is LoDTensor by default
+    VLOG(10) << this->Type()
+             << " has not registered InferVarType. Set output variables to "
+                "LOD_TENSOR";
     for (auto &out_pair : this->outputs_) {
       for (auto &out_var_name : out_pair.second) {
-        block->Var(out_var_name)->SetType(VarDesc::LOD_TENSOR);
+        block->FindRecursiveOrCreateVar(out_var_name)
+            ->SetType(VarDesc::LOD_TENSOR);
       }
     }
   }
@@ -448,6 +455,12 @@ void CompileTimeInferShapeContext::SetDim(const std::string &name,
                                           const DDim &dim) {
   block_.FindVarRecursive(name)->SetShape(framework::vectorize(dim));
 }
+bool CompileTimeInferShapeContext::IsRuntime() const { return false; }
+
+VarDesc::VarType CompileTimeInferShapeContext::GetVarType(
+    const std::string &name) const {
+  return block_.FindVarRecursive(name)->GetType();
+}
 
 }  // namespace framework
 }  // namespace paddle

paddle/framework/op_registry.h

@@ -92,8 +92,7 @@ struct OpKernelRegistrarFunctor<PlaceType, false, I, KernelTypes...> {
 
   void operator()(const char* op_type) const {
     using T = typename KERNEL_TYPE::ELEMENT_TYPE;
-    OperatorWithKernel::OpKernelKey key(ToDataType(std::type_index(typeid(T))),
-                                        PlaceType());
+    OpKernelType key(ToDataType(std::type_index(typeid(T))), PlaceType());
     OperatorWithKernel::AllOpKernels()[op_type][key].reset(new KERNEL_TYPE);
 
     constexpr auto size = std::tuple_size<std::tuple<KernelTypes...>>::value;

paddle/framework/operator.cc

@@ -15,7 +15,9 @@ limitations under the License. */
 #include "paddle/framework/operator.h"
 #include <algorithm>
 #include <atomic>
+#include "paddle/framework/lod_tensor_array.h"
 #include "paddle/framework/shape_inference.h"
+#include "paddle/framework/var_type.h"
 
 namespace paddle {
 namespace framework {
@@ -252,8 +254,7 @@ std::vector<Tensor*> ExecutionContext::MultiOutput<Tensor>(
   return res;
 }
 
-std::ostream& operator<<(std::ostream& os,
-                         const OperatorWithKernel::OpKernelKey& kernel_key) {
+std::ostream& operator<<(std::ostream& os, const OpKernelType& kernel_key) {
   os << "place[" << kernel_key.place_ << "]:data_type[" << kernel_key.data_type_
      << "]";
   return os;
@@ -365,7 +366,9 @@ class RuntimeInferShapeContext : public InferShapeContext {
     out_tensor->set_lod(in_tensor.lod());
   }
 
- private:
+  bool IsRuntime() const override { return true; }
+
+ protected:
   DDim GetDim(const std::string& name) const override {
     Variable* var = scope_.FindVar(name);
     if (var->IsType<LoDTensor>()) {
@@ -388,6 +391,12 @@ class RuntimeInferShapeContext : public InferShapeContext {
     }
   }
 
+  VarDesc::VarType GetVarType(const std::string& name) const override {
+    auto* var = scope_.FindVar(name);
+    return ToVarType(var->Type());
+  }
+
+ private:
   const OperatorBase& op_;
   const Scope& scope_;
 };
@@ -422,7 +431,7 @@ void OperatorWithKernel::Run(const Scope& scope,
 
   // check if op[type] have kernel for kernel_key
   OpKernelMap& kernels = kernels_iter->second;
-  auto kernel_key = OpKernelKey(IndicateDataType(ctx), dev_ctx);
+  auto kernel_key = GetKernelType(ctx);
   auto kernel_iter = kernels.find(kernel_key);
 
   if (kernel_iter == kernels.end()) {
@@ -430,6 +439,41 @@ void OperatorWithKernel::Run(const Scope& scope,
   }
 
   kernel_iter->second->Compute(ctx);
+
+  // throws errors if have.
+  dev_ctx.Finish();
+}
+OpKernelType OperatorWithKernel::GetKernelType(
+    const ExecutionContext& ctx) const {
+  return OpKernelType(IndicateDataType(ctx), ctx.device_context());
+}
+DataType OperatorWithKernel::IndicateDataType(
+    const ExecutionContext& ctx) const {
+  auto& scope = ctx.scope();
+  int data_type = -1;
+  for (auto& input : this->inputs_) {
+    for (auto& ipt_name : input.second) {
+      auto* var = scope.FindVar(ipt_name);
+      if (var != nullptr) {
+        const Tensor* t = nullptr;
+        if (var->IsType<Tensor>()) {
+          t = &var->Get<Tensor>();
+        } else if (var->IsType<LoDTensor>()) {
+          t = &var->Get<LoDTensor>();
+        } else if (var->IsType<SelectedRows>()) {
+          t = &(var->Get<SelectedRows>().value());
+        }
+        if (t != nullptr) {
+          int tmp = static_cast<int>(ToDataType(t->type()));
+          PADDLE_ENFORCE(tmp == data_type || data_type == -1,
+                         "DataType of Paddle Op %s must be the same.", Type());
+          data_type = tmp;
+        }
+      }
+    }
+  }
+  PADDLE_ENFORCE(data_type != -1, "DataType should be indicated by input");
+  return static_cast<DataType>(data_type);
 }
 
 }  // namespace framework

paddle/framework/operator.h

@@ -298,11 +298,10 @@ class ExecutionContext {
   }
 
 #ifdef PADDLE_WITH_CUDA
-  const platform::CUDADeviceContext& cuda_device_context() const {
+  const inline platform::CUDADeviceContext& cuda_device_context() const {
     PADDLE_ENFORCE(platform::is_gpu_place(device_context_.GetPlace()));
-    auto cuda_ctx =
-        reinterpret_cast<const platform::CUDADeviceContext*>(&device_context_);
-    return *cuda_ctx;
+    return *reinterpret_cast<const platform::CUDADeviceContext*>(
+        &device_context_);
   }
 #endif
 
@@ -346,38 +345,38 @@ class OpKernel : public OpKernelBase {
   using ELEMENT_TYPE = T;
 };
 
-class OperatorWithKernel : public OperatorBase {
- public:
-  struct OpKernelKey {
+struct OpKernelType {
+  struct Hash {
+    std::hash<int> hash_;
+    size_t operator()(const OpKernelType& key) const {
+      int place = key.place_.which();
+      int data_type = static_cast<int>(key.data_type_);
+      int pre_hash = data_type << NUM_PLACE_TYPE_LIMIT_IN_BIT |
+                     (place & ((1 << NUM_PLACE_TYPE_LIMIT_IN_BIT) - 1));
+      return hash_(pre_hash);
+    }
+  };
+
   platform::Place place_;
   DataType data_type_;
 
-    OpKernelKey(DataType data_type, platform::Place place)
+  OpKernelType(DataType data_type, platform::Place place)
       : place_(place), data_type_(data_type) {}
 
-    OpKernelKey(DataType data_type, const platform::DeviceContext& dev_ctx)
+  OpKernelType(DataType data_type, const platform::DeviceContext& dev_ctx)
       : place_(dev_ctx.GetPlace()), data_type_(data_type) {}
 
-    bool operator==(const OpKernelKey& o) const {
+  bool operator==(const OpKernelType& o) const {
     return platform::places_are_same_class(place_, o.place_) &&
            data_type_ == o.data_type_;
   }
-  };
-
-  struct OpKernelHash {
-    std::hash<int> hash_;
-    size_t operator()(const OpKernelKey& key) const {
-      int place = key.place_.which();
-      int data_type = static_cast<int>(key.data_type_);
-      int pre_hash = data_type << NUM_PLACE_TYPE_LIMIT_IN_BIT |
-                     (place & ((1 << NUM_PLACE_TYPE_LIMIT_IN_BIT) - 1));
-      return hash_(pre_hash);
-    }
-  };
+};
 
+class OperatorWithKernel : public OperatorBase {
+ public:
   using OpKernelMap =
-      std::unordered_map<OpKernelKey, std::unique_ptr<OpKernelBase>,
-                         OpKernelHash>;
+      std::unordered_map<OpKernelType, std::unique_ptr<OpKernelBase>,
+                         OpKernelType::Hash>;
 
   OperatorWithKernel(const std::string& type, const VariableNameMap& inputs,
                      const VariableNameMap& outputs, const AttributeMap& attrs)
@@ -405,40 +404,15 @@ class OperatorWithKernel : public OperatorBase {
   }
 
  protected:
+  virtual OpKernelType GetKernelType(const ExecutionContext& ctx) const;
+
+ private:
   // indicate kernel DataType by input data. Defaultly all input data must be
   // same.
-  virtual DataType IndicateDataType(const ExecutionContext& ctx) const {
-    auto& scope = ctx.scope();
-    int data_type = -1;
-    for (auto& input : this->inputs_) {
-      for (auto& ipt_name : input.second) {
-        auto* var = scope.FindVar(ipt_name);
-        if (var != nullptr) {
-          const Tensor* t = nullptr;
-          if (var->IsType<Tensor>()) {
-            t = &var->Get<Tensor>();
-          } else if (var->IsType<LoDTensor>()) {
-            t = &var->Get<LoDTensor>();
-          } else if (var->IsType<SelectedRows>()) {
-            t = &(var->Get<SelectedRows>().value());
-          }
-          if (t != nullptr) {
-            int tmp = static_cast<int>(ToDataType(t->type()));
-            PADDLE_ENFORCE(tmp == data_type || data_type == -1,
-                           "DataType of Paddle Op %s must be the same.",
-                           Type());
-            data_type = tmp;
-          }
-        }
-      }
-    }
-    PADDLE_ENFORCE(data_type != -1, "DataType should be indicated by input");
-    return static_cast<DataType>(data_type);
-  }
+  DataType IndicateDataType(const ExecutionContext& ctx) const;
 };
 
-std::ostream& operator<<(std::ostream& os,
-                         const OperatorWithKernel::OpKernelKey& kernel_key);
+std::ostream& operator<<(std::ostream& os, const OpKernelType& kernel_key);
 
 extern bool OpSupportGPU(const std::string& op_type);
 

paddle/framework/operator_test.cc

@@ -114,8 +114,8 @@ class OpWithKernelTest : public OperatorWithKernel {
 
  protected:
   void InferShape(framework::InferShapeContext* ctx) const override {}
-  DataType IndicateDataType(const ExecutionContext& ctx) const override {
-    return DataType::FP32;
+  OpKernelType GetKernelType(const ExecutionContext& ctx) const override {
+    return OpKernelType(DataType::FP32, ctx.device_context());
   }
 };
 

paddle/framework/scope.cc

@@ -98,5 +98,23 @@ void Scope::DeleteScope(Scope* scope) {
   delete scope;
 }
 
+void Scope::Rename(const std::string& origin_name,
+                   const std::string& new_name) const {
+  auto origin_it = vars_.find(origin_name);
+  PADDLE_ENFORCE(origin_it != vars_.end(),
+                 "Cannot find original variable with name %s", origin_name);
+  auto new_it = vars_.find(new_name);
+  PADDLE_ENFORCE(new_it == vars_.end(),
+                 "The variable with name %s is already in the scope", new_name);
+  vars_[new_name] = origin_it->second;
+  vars_.erase(origin_it);
+}
+
+std::string Scope::Rename(const std::string& origin_name) const {
+  auto var_name = string::Sprintf("%p.%d", this, vars_.size());
+  Rename(origin_name, var_name);
+  return var_name;
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/scope.h

@@ -68,11 +68,18 @@ class Scope {
   // enumerate all the variables current contains.
   std::vector<std::string> GetAllNames(bool recursive = false) const;
 
+  // Rename variable to a new name
+  void Rename(const std::string& origin_name,
+              const std::string& new_name) const;
+
+  // Rename variable to a new name and return the new name
+  std::string Rename(const std::string& origin_name) const;
+
  private:
   // Call Scope::NewScope for a sub-scope.
   explicit Scope(Scope const* parent) : parent_(parent) {}
 
-  std::unordered_map<std::string, Variable*> vars_;
+  mutable std::unordered_map<std::string, Variable*> vars_;
   mutable std::list<Scope*> kids_;
   Scope const* parent_{nullptr};
 

paddle/framework/shape_inference.cc

@@ -46,6 +46,23 @@ void InferShapeContext::SetDims(const std::vector<std::string> &names,
     SetDim(names[i], dims[i]);
   }
 }
+std::vector<VarDesc::VarType> InferShapeContext::GetInputsVarType(
+    const std::string &name) const {
+  return GetVarTypes(Inputs(name));
+}
+std::vector<VarDesc::VarType> InferShapeContext::GetOutputsVarType(
+    const std::string &name) const {
+  return GetVarTypes(Outputs(name));
+}
+std::vector<VarDesc::VarType> InferShapeContext::GetVarTypes(
+    const std::vector<std::string> &names) const {
+  std::vector<VarDesc::VarType> retv;
+  retv.resize(names.size());
+  std::transform(names.begin(), names.end(), retv.begin(),
+                 std::bind(std::mem_fn(&InferShapeContext::GetVarType), this,
+                           std::placeholders::_1));
+  return retv;
+}
 
 }  // namespace framework
 }  // namespace paddle

paddle/framework/shape_inference.h

@@ -16,6 +16,7 @@ limitations under the License. */
 
 #include "paddle/framework/attribute.h"
 #include "paddle/framework/ddim.h"
+#include "paddle/framework/framework.pb.h"
 
 namespace paddle {
 namespace framework {
@@ -26,6 +27,10 @@ class InferShapeContext {
   virtual bool HasInput(const std::string &name) const = 0;
   virtual bool HasOutput(const std::string &name) const = 0;
 
+  std::vector<VarDesc::VarType> GetInputsVarType(const std::string &name) const;
+  std::vector<VarDesc::VarType> GetOutputsVarType(
+      const std::string &name) const;
+
   virtual bool HasInputs(const std::string &name) const = 0;
   virtual bool HasOutputs(const std::string &name) const = 0;
 
@@ -46,6 +51,8 @@ class InferShapeContext {
   virtual void ShareLoD(const std::string &in, const std::string &out,
                         size_t i = 0, size_t j = 0) const = 0;
 
+  virtual bool IsRuntime() const = 0;
+
  protected:
   virtual framework::DDim GetDim(const std::string &name) const = 0;
   virtual void SetDim(const std::string &name, const framework::DDim &dim) = 0;
@@ -55,6 +62,11 @@ class InferShapeContext {
 
   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;
+
+  virtual VarDesc::VarType GetVarType(const std::string &name) const = 0;
 };
 
 }  // namespace framework

paddle/framework/tensor_impl.h

@@ -52,7 +52,7 @@ struct SizeOfTypeFunctor<HEAD, TAIL...> {
 };
 
 static inline size_t SizeOfType(std::type_index type) {
-  SizeOfTypeFunctor<int, float, double, int16_t, int64_t> functor;
+  SizeOfTypeFunctor<int, float, double, int16_t, int64_t, bool> functor;
   size_t size = functor(type);
   PADDLE_ENFORCE(size != 0UL, "Cannot get size of type %s", type.name());
   return size;

paddle/framework/var_desc.cc

@@ -37,13 +37,29 @@ std::vector<int64_t> VarDescBind::Shape() const {
 DataType VarDescBind::GetDataType() const { return tensor_desc().data_type(); }
 
 void VarDescBind::SetLoDLevel(int32_t lod_level) {
-  PADDLE_ENFORCE(desc_.type() == VarDesc::LOD_TENSOR);
+  switch (desc_.type()) {
+    case VarDesc::LOD_TENSOR:
       desc_.mutable_lod_tensor()->set_lod_level(lod_level);
+      break;
+    case VarDesc::LOD_TENSOR_ARRAY:
+      desc_.mutable_tensor_array()->set_lod_level(lod_level);
+      break;
+    default:
+      PADDLE_THROW("Tensor type=%d does not support LoDLevel",
+                   desc_.tensor_array().lod_level());
+  }
 }
 
 int32_t VarDescBind::GetLodLevel() const {
-  PADDLE_ENFORCE(desc_.type() == VarDesc::LOD_TENSOR);
+  switch (desc_.type()) {
+    case VarDesc::LOD_TENSOR:
       return desc_.lod_tensor().lod_level();
+    case VarDesc::LOD_TENSOR_ARRAY:
+      return desc_.tensor_array().lod_level();
+    default:
+      PADDLE_THROW("Tensor type=%d does not support LoDLevel",
+                   desc_.tensor_array().lod_level());
+  }
 }
 
 const TensorDesc &VarDescBind::tensor_desc() const {
@@ -53,6 +69,8 @@ const TensorDesc &VarDescBind::tensor_desc() const {
       return desc_.selected_rows();
     case VarDesc::LOD_TENSOR:
       return desc_.lod_tensor().tensor();
+    case VarDesc::LOD_TENSOR_ARRAY:
+      return desc_.tensor_array().tensor();
     default:
       PADDLE_THROW("Unexpected branch.");
   }
@@ -66,6 +84,8 @@ TensorDesc *VarDescBind::mutable_tensor_desc() {
       return desc_.mutable_selected_rows();
     case VarDesc::LOD_TENSOR:
       return desc_.mutable_lod_tensor()->mutable_tensor();
+    case VarDesc::LOD_TENSOR_ARRAY:
+      return desc_.mutable_tensor_array()->mutable_tensor();
     default:
       PADDLE_THROW("Unexpected branch.");
   }

paddle/framework/var_type.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#pragma once
+#include "paddle/framework/framework.pb.h"
+#include "paddle/framework/lod_rank_table.h"
+#include "paddle/framework/lod_tensor.h"
+#include "paddle/framework/lod_tensor_array.h"
+
+namespace paddle {
+namespace framework {
+inline VarDesc::VarType ToVarType(std::type_index type) {
+  if (type.hash_code() == typeid(LoDTensor).hash_code()) {
+    return VarDesc_VarType_LOD_TENSOR;
+  } else if (type.hash_code() == typeid(LoDRankTable).hash_code()) {
+    return VarDesc_VarType_LOD_RANK_TABLE;
+  } else if (type.hash_code() == typeid(LoDTensorArray).hash_code()) {
+    return VarDesc_VarType_LOD_TENSOR_ARRAY;
+  } else if (type.hash_code() == typeid(SelectedRows).hash_code()) {
+    return VarDesc_VarType_SELECTED_ROWS;
+  } else {
+    PADDLE_THROW("ToVarType:Unsupported type %s", type.name());
+  }
+}
+
+template <typename Visitor>
+inline void VisitVarType(const Variable& var, Visitor visitor) {
+  switch (ToVarType(var.Type())) {
+    case VarDesc_VarType_LOD_TENSOR:
+      visitor(var.Get<framework::LoDTensor>());
+      return;
+    case VarDesc_VarType_LOD_RANK_TABLE:
+      visitor(var.Get<LoDRankTable>());
+      return;
+    case VarDesc_VarType_LOD_TENSOR_ARRAY:
+      visitor(var.Get<LoDTensorArray>());
+      return;
+    case VarDesc_VarType_SELECTED_ROWS:
+      visitor(var.Get<SelectedRows>());
+      return;
+    default:
+      PADDLE_THROW("Not supported visit type, %d", ToVarType(var.Type()));
+  }
+}
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/variable.h

@@ -48,6 +48,11 @@ class Variable {
 
   void Clear() { holder_.reset(); }
 
+  std::type_index Type() const {
+    PADDLE_ENFORCE(holder_ != nullptr, "Must hold memory");
+    return holder_->Type();
+  }
+
  private:
   struct Placeholder {
     virtual ~Placeholder() {}

paddle/function/CMakeLists.txt

@@ -45,6 +45,7 @@ if(WITH_GPU)
     add_simple_unittest(BlockExpandOpTest)
     add_simple_unittest(CropOpTest)
     add_simple_unittest(SwitchOpTest)
+    add_simple_unittest(ScaleSubRegionOpTest)
 endif()
 
 add_simple_unittest(Im2ColTest)

paddle/function/FunctionTest.h

@@ -110,6 +110,7 @@ public:
         function2_(FunctionBase::funcRegistrar_.createByType(name2)) {
     function1_->init(config);
     function2_->init(config);
+    initArgsCallback_ = nullptr;
   }
 
   ~Compare2Function() {}
@@ -170,6 +171,10 @@ public:
                                       *seq2_));
   }
 
+  void registerInitCallback(std::function<void(BufferArg&, size_t)> callback) {
+    initArgsCallback_ = callback;
+  }
+
   // output need only contains shape, do not contains data.
   void addOutputs(const BufferArg& output, ArgType argType = ASSIGN_TO) {
     size_t size =
@@ -340,6 +345,10 @@ protected:
         initArg(*func1Inputs_[i]);
       }
 
+      if (initArgsCallback_ != nullptr) {
+        initArgsCallback_(*func1Inputs_[i], i);
+      }
+
       copyArg_(*func1Inputs_[i], *func2Inputs_[i]);
     }
   }
@@ -386,6 +395,7 @@ protected:
   std::shared_ptr<SequenceIdArg> seq1_;
   std::shared_ptr<SequenceIdArg> seq2_;
   test::CopyArgument<DType1, DType2> copyArg_;
+  std::function<void(BufferArg&, size_t)> initArgsCallback_;
 };
 
 class CpuGpuFuncCompare

paddle/function/ScaleSubRegionOp.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 "ScaleSubRegionOp.h"
+#include "paddle/function/TensorShape.h"
+
+namespace paddle {
+
+template <>
+void ScaleSubRegion<DEVICE_TYPE_CPU>(real* outputs,
+                                     const real* inputs,
+                                     const real* indices,
+                                     const TensorShape shape,
+                                     const FuncConfig& conf) {
+  real value = conf.get<real>("value");
+
+  int number = shape[0];
+  int channel = shape[1];
+  int height = shape[2];
+  int width = shape[3];
+
+  memcpy(outputs, inputs, number * channel * height * width * sizeof(real));
+
+  for (int n = 0; n < number; ++n) {
+    // indices start from 1
+    int offset = n * 6;
+    for (int c = indices[offset] - 1; c < indices[offset + 1]; ++c) {
+      for (int h = indices[offset + 2] - 1; h < indices[offset + 3]; ++h) {
+        for (int w = indices[offset + 4] - 1; w < indices[offset + 5]; ++w) {
+          int idx = ((n * channel + c) * height + h) * width + w;
+          outputs[idx] *= value;
+        }
+      }
+    }
+  }
+}
+
+template <>
+void ScaleSubRegionGrad<DEVICE_TYPE_CPU>(const real* inGrad,
+                                         real* outGrad,
+                                         const real* indices,
+                                         const TensorShape shape,
+                                         const FuncConfig& conf) {
+  real value = conf.get<real>("value");
+
+  int number = shape[0];
+  int channel = shape[1];
+  int height = shape[2];
+  int width = shape[3];
+
+  for (int n = 0; n < number; ++n) {
+    for (int c = 0; c < channel; ++c) {
+      for (int h = 0; h < height; ++h) {
+        for (int w = 0; w < width; ++w) {
+          int idx = ((n * channel + c) * height + h) * width + w;
+          int offset = n * 6;
+          if (c >= (indices[offset] - 1) && c <= (indices[offset + 1] - 1) &&
+              h >= (indices[offset + 2] - 1) &&
+              h <= (indices[offset + 3] - 1) &&
+              w >= (indices[offset + 4] - 1) &&
+              w <= (indices[offset + 5] - 1)) {
+            outGrad[idx] += inGrad[idx] * value;
+          } else {
+            outGrad[idx] += inGrad[idx];
+          }
+        }
+      }
+    }
+  }
+}
+
+/**
+ * \brief For each instance, ScaleSubRegion can be used to multiply a value to
+ *        a specified sub continuous region. By providing start index and end
+ *        index for C/H/W, you can specify the location and shape of the region.
+ *
+ * Argument in this Function:
+ * \param inputs    A 4-D tensor with shape [N, C, H, W], only one input.
+ * \param indices   A 2-D tensor with shape [N, 6], indicates the sub region.
+ * \param outputs   A 4-D tensor with same shape as inputs, output value.
+ */
+template <DeviceType Device>
+class ScaleSubRegionFunc : public FunctionBase {
+public:
+  void init(const FuncConfig& config) override { conf_ = config; }
+
+  void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
+    CHECK_EQ(2UL, inputs.size());
+    CHECK_EQ(1UL, outputs.size());
+    CHECK_EQ(outputs[0].getArgType(), ASSIGN_TO);
+
+    TensorShape shape = inputs[0].shape();
+
+    ScaleSubRegion<Device>(outputs[0].data<real>(),
+                           inputs[0].data<real>(),
+                           inputs[1].data<real>(),
+                           shape,
+                           conf_);
+  }
+
+private:
+  FuncConfig conf_;
+};
+
+/**
+ * \brief The backward propagation of ScaleSubRegion Function.
+ *
+ * Argument in this Function:
+ * \param inputs  A 4-D tensor with shape [N, C, H, W], output gradient.
+ * \param indices A 2-D tensor with shape [N, 6], indicates the sub region.
+ * \param outputs A 4-D tensor with shape [N, C, H, W], gradient of input value.
+ */
+
+template <DeviceType Device>
+class ScaleSubRegionGradFunc : public FunctionBase {
+public:
+  void init(const FuncConfig& config) override { conf_ = config; }
+
+  void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
+    CHECK_EQ(2UL, inputs.size());
+    CHECK_EQ(1UL, outputs.size());
+    CHECK_EQ(outputs[0].getArgType(), ADD_TO);
+
+    TensorShape shape = inputs[0].shape();
+
+    ScaleSubRegionGrad<Device>(inputs[0].data<real>(),
+                               outputs[0].data<real>(),
+                               inputs[1].data<real>(),
+                               shape,
+                               conf_);
+  }
+
+private:
+  FuncConfig conf_;
+};
+
+REGISTER_TYPED_FUNC(ScaleSubRegion, CPU, ScaleSubRegionFunc);
+REGISTER_TYPED_FUNC(ScaleSubRegionGrad, CPU, ScaleSubRegionGradFunc);
+#ifdef PADDLE_WITH_CUDA
+REGISTER_TYPED_FUNC(ScaleSubRegion, GPU, ScaleSubRegionFunc);
+REGISTER_TYPED_FUNC(ScaleSubRegionGrad, GPU, ScaleSubRegionGradFunc);
+#endif
+
+}  // namespace paddle

paddle/function/ScaleSubRegionOp.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 "Function.h"
+
+namespace paddle {
+
+/**
+ * \brief Function to multiply a value to values in specified sub continuous
+ *        region. Indices must be provided to indcate the location and shape of
+ *        the region and the multiplied value is passed by configure variable.
+ *
+ *
+ * \param[out] outputs  Output value.
+ * \param[in]  inputs   Input data which contains NCHW information.
+ * \param[in]  indices  Indices data to indcate the sub region.
+ * \param[in]  shape    Tensor shape of input value.
+ * \param[in]  conf     Configure variable which contains the multiplied value.
+ */
+template <DeviceType Device>
+void ScaleSubRegion(real* outputs,
+                    const real* inputs,
+                    const real* indices,
+                    const TensorShape shape,
+                    const FuncConfig& conf);
+
+/**
+ * \brief Backward propagation function of ScaleSubRegion.
+ *
+ * \param[out] inGrad   Gradients of previous layer.
+ * \param[in]  outGrad  Output gradient.
+ * \param[in]  indices  Indices data.
+ * \param[in]  shape    The Shape of input tensor.
+ * \param[in]  conf     Configure variable.
+ */
+template <DeviceType Device>
+void ScaleSubRegionGrad(const real* inGrad,
+                        real* outGrad,
+                        const real* indices,
+                        const TensorShape shape,
+                        const FuncConfig& conf);
+}  // namespace paddle

paddle/function/ScaleSubRegionOpGpu.cu

+/* 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 "ScaleSubRegionOp.h"
+#include "hl_base.h"
+
+namespace paddle {
+
+__global__ void KeScaleSubRegion(real* outputs,
+                                 const real* inputs,
+                                 const real* indices,
+                                 real value,
+                                 int channel,
+                                 int height,
+                                 int width,
+                                 int nthreads) {
+  const int idx = threadIdx.x + blockIdx.x * blockDim.x;
+  if (idx < nthreads) {
+    const int w = idx % width;
+    const int h = (idx / width) % height;
+    const int c = (idx / width / height) % channel;
+    const int n = idx / width / height / channel;
+
+    const int offset = n * 6;
+    if (c >= (indices[offset] - 1) && c <= (indices[offset + 1] - 1) &&
+        h >= (indices[offset + 2] - 1) && h <= (indices[offset + 3] - 1) &&
+        w >= (indices[offset + 4] - 1) && w <= (indices[offset + 5] - 1)) {
+      outputs[idx] = inputs[idx] * value;
+    } else {
+      outputs[idx] = inputs[idx];
+    }
+  }
+}
+
+template <>
+void ScaleSubRegion<DEVICE_TYPE_GPU>(real* outputs,
+                                     const real* inputs,
+                                     const real* indices,
+                                     const TensorShape shape,
+                                     const FuncConfig& conf) {
+  real value = conf.get<real>("value");
+
+  int number = shape[0];
+  int channel = shape[1];
+  int height = shape[2];
+  int width = shape[3];
+
+  size_t nth = number * channel * height * width;
+  int blockSize = 1024;
+  int gridSize = (nth + blockSize - 1) / blockSize;
+
+  KeScaleSubRegion<<<gridSize, blockSize, 0, STREAM_DEFAULT>>>(
+      outputs, inputs, indices, value, channel, height, width, nth);
+  CHECK_SYNC("ScaleSubRegion");
+}
+
+__global__ void KeScaleSubRegionDiff(const real* inGrad,
+                                     real* outGrad,
+                                     const real* indices,
+                                     real value,
+                                     int channel,
+                                     int height,
+                                     int width,
+                                     int nthreads) {
+  const int idx = threadIdx.x + blockIdx.x * blockDim.x;
+  if (idx < nthreads) {
+    const int w = idx % width;
+    const int h = (idx / width) % height;
+    const int c = (idx / width / height) % channel;
+    const int n = idx / width / height / channel;
+
+    const int offset = n * 6;
+    if (c >= (indices[offset] - 1) && c <= (indices[offset + 1] - 1) &&
+        h >= (indices[offset + 2] - 1) && h <= (indices[offset + 3] - 1) &&
+        w >= (indices[offset + 4] - 1) && w <= (indices[offset + 5] - 1)) {
+      outGrad[idx] += inGrad[idx] * value;
+    } else {
+      outGrad[idx] += inGrad[idx];
+    }
+  }
+}
+
+template <>
+void ScaleSubRegionGrad<DEVICE_TYPE_GPU>(const real* inGrad,
+                                         real* outGrad,
+                                         const real* indices,
+                                         const TensorShape shape,
+                                         const FuncConfig& conf) {
+  real value = conf.get<real>("value");
+
+  int number = shape[0];
+  int channel = shape[1];
+  int height = shape[2];
+  int width = shape[3];
+
+  size_t nth = number * channel * height * width;
+  int blockSize = 1024;
+  int gridSize = (nth + blockSize - 1) / blockSize;
+
+  KeScaleSubRegionDiff<<<gridSize, blockSize, 0, STREAM_DEFAULT>>>(
+      inGrad, outGrad, indices, value, channel, height, width, nth);
+  CHECK_SYNC("ScaleSubRegionGrad");
+}
+
+}  // namespace paddle

paddle/function/ScaleSubRegionOpTest.cpp

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include <gtest/gtest.h>
+#include "FunctionTest.h"
+
+namespace paddle {
+
+TEST(ScaleSubRegion, real) {
+  for (size_t numSamples : {5, 32}) {
+    for (size_t channels : {5, 32}) {
+      for (size_t imgSizeH : {5, 33}) {
+        for (size_t imgSizeW : {5, 32}) {
+          for (real value : {-0.5, 0.0, 0.5}) {
+            for (bool firstHalf : {false, true}) {
+              VLOG(3) << " numSamples=" << numSamples
+                      << " channels=" << channels << " imgSizeH=" << imgSizeH
+                      << " imgSizeW=" << imgSizeW;
+
+              for (bool testGrad : {false, true}) {
+                CpuGpuFuncCompare compare(
+                    testGrad ? "ScaleSubRegionGrad" : "ScaleSubRegion",
+                    FuncConfig().set<real>("value", value));
+
+                TensorShape shape{numSamples, channels, imgSizeH, imgSizeW};
+                TensorShape indicesShape{numSamples, 6};
+
+                compare.addInputs(BufferArg(VALUE_TYPE_FLOAT, shape));
+                compare.addInputs(BufferArg(VALUE_TYPE_FLOAT, indicesShape));
+
+                compare.registerInitCallback([=](BufferArg& arg, size_t index) {
+                  if (index == 1) {
+                    real* data = (real*)arg.data();
+
+                    for (size_t i = 0; i < numSamples; ++i) {
+                      size_t offset = i * 6;
+                      data[offset] = firstHalf ? 1 : channels / 2;
+                      data[offset + 1] = firstHalf ? channels / 2 : channels;
+                      data[offset + 2] = firstHalf ? 1 : imgSizeH / 2;
+                      data[offset + 3] = firstHalf ? imgSizeH / 2 : imgSizeH;
+                      data[offset + 4] = firstHalf ? 1 : imgSizeW / 2;
+                      data[offset + 5] = firstHalf ? imgSizeW / 2 : imgSizeW;
+                    }
+                  }
+                });
+
+                compare.addOutputs(
+                    BufferArg(
+                        VALUE_TYPE_FLOAT, shape, testGrad ? ADD_TO : ASSIGN_TO),
+                    testGrad ? ADD_TO : ASSIGN_TO);
+                compare.run();
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+}
+
+}  // namespace paddle

paddle/gserver/layers/ConvBaseProjection.cpp

@@ -17,7 +17,7 @@ limitations under the License. */
 
 namespace paddle {
 
-ThreadLocalD<std::vector<MemoryHandle *>> ConvBaseProjection::convMem_;
+ThreadLocalD<std::vector<MemoryHandlePtr>> ConvBaseProjection::convMem_;
 
 ConvBaseProjection::ConvBaseProjection(const ProjectionConfig &config,
                                        ParameterPtr parameter,
@@ -175,18 +175,18 @@ void ConvBaseProjection::reshape(int batchSize) {
 }
 
 void *ConvBaseProjection::getSpaceBytes(size_t size) {
-  std::vector<MemoryHandle *> &convMem = *convMem_;
+  std::vector<MemoryHandlePtr> &convMem = *convMem_;
   if (convMem.empty()) {
     int numDevices = hl_get_device_count();
     convMem.resize(numDevices);
   }
 
   int devId = hl_get_device();
-  MemoryHandle **localMem = &(convMem[devId]);
-  if (NULL == *localMem || size > (*localMem)->getAllocSize()) {
-    *localMem = new GpuMemoryHandle(size);
+  MemoryHandlePtr localMem = convMem[devId];
+  if (NULL == localMem || size > localMem->getAllocSize()) {
+    localMem = std::make_shared<GpuMemoryHandle>(size);
   }
-  return (*localMem)->getBuf();
+  return localMem->getBuf();
 }
 
 ConvBaseProjection::~ConvBaseProjection() {

paddle/gserver/layers/ConvBaseProjection.h

@@ -105,7 +105,7 @@ protected:
   bool bias_;
 
   std::unique_ptr<Weight> weight_;
-  static ThreadLocalD<std::vector<MemoryHandle*>> convMem_;
+  static ThreadLocalD<std::vector<MemoryHandlePtr>> convMem_;
 };
 
 }  // namespace paddle

paddle/gserver/layers/MKLDNNAddtoLayer.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 "MKLDNNAddtoLayer.h"
+
+using namespace mkldnn;  // NOLINT
+
+namespace paddle {
+
+REGISTER_LAYER(mkldnn_addto, MKLDNNAddtoLayer);
+
+bool MKLDNNAddtoLayer::init(const LayerMap& layerMap,
+                            const ParameterMap& parameterMap) {
+  if (!MKLDNNLayer::init(layerMap, parameterMap)) {
+    return false;
+  }
+
+  layerSize_ = getSize();
+  for (size_t i = 0; i < inputLayers_.size(); i++) {
+    CHECK_EQ(layerSize_, inputLayers_[i]->getSize()) << "input size must equal";
+  }
+  if (biasParameter_.get() != NULL) {
+    biases_ =
+        std::unique_ptr<Weight>(new Weight(1, layerSize_, biasParameter_, 0));
+  }
+  return true;
+}
+
+void MKLDNNAddtoLayer::reshape(
+    int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) {
+  CHECK_EQ(layerSize_, getSize()) << "this layer size can not be changed";
+  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);
+  for (size_t i = 0; i < inputLayers_.size(); i++) {
+    CHECK_EQ(int64_t(bs), inputLayers_[i]->getOutput().getBatchSize());
+    CHECK_EQ(layerSize_, inputLayers_[i]->getSize());
+  }
+
+  oc = ic;
+  oh = ih;
+  ow = iw;
+  reshapeOutput(oh, ow);
+  resizeOutput(bs, oc * oh * ow);
+}
+
+void MKLDNNAddtoLayer::resetFwd(std::vector<primitive>& pipeline,
+                                MKLDNNMatrixPtr& in,
+                                MKLDNNMatrixPtr& wgt,
+                                MKLDNNMatrixPtr& bias,
+                                MKLDNNMatrixPtr& out) {
+  resetFwdBuffers(inVals_, bias, out);
+  in = inVals_[0];
+
+  std::shared_ptr<sum::primitive_desc> fwdPD;
+  std::shared_ptr<sum::primitive_desc> biasPD;
+  resetFwdPD(fwdPD, biasPD, inVals_, bias, out);
+
+  resetFwdPipeline(pipeline, fwdPD, biasPD, inVals_, bias, out);
+}
+
+void MKLDNNAddtoLayer::resetBwd(std::vector<primitive>& pipeline,
+                                MKLDNNMatrixPtr& in,
+                                MKLDNNMatrixPtr& wgt,
+                                MKLDNNMatrixPtr& bias,
+                                MKLDNNMatrixPtr& out) {
+  resetBwdBuffers(inGrads_, bias, out);
+  in = inGrads_[0];
+
+  // backward only need share output grad to input grad
+  for (size_t i = 0; i < inGrads_.size(); i++) {
+    if (inGrads_[i] != nullptr) {
+      inGrads_[i] = out;
+      inputLayers_[i]->getOutputGrad()->setData(inGrads_[i]->getData());
+    }
+  }
+
+  // backward bias
+  bwdBias_ = nullptr;
+  if (bias) {
+    std::vector<float> scales(bs_, 1.0);
+    std::vector<memory::primitive_desc> srcPDs(bs_, bias->getPrimitiveDesc());
+    auto biasPD = sum::primitive_desc(bias->getMemoryDesc(), scales, srcPDs);
+    std::vector<primitive::at> srcs;
+    for (size_t i = 0; i < grads_.size(); ++i) {
+      srcs.push_back(*(grads_[i]));
+    }
+    bwdBias_.reset(new sum(biasPD, srcs, *bias));
+    pipeline.push_back(*bwdBias_);
+  }
+}
+
+void MKLDNNAddtoLayer::updateWeights(const UpdateCallback& callback) {
+  if (biases_ && biases_->getWGrad()) {
+    biases_->getParameterPtr()->incUpdate(callback);
+  }
+}
+
+void MKLDNNAddtoLayer::prepareBias(MKLDNNMatrixPtr& bias,
+                                   const MatrixPtr& biasMat,
+                                   const MKLDNNMatrixPtr& out,
+                                   std::vector<MKLDNNMatrixPtr>& outs) {
+  auto pd = MKLDNNMatrix::createPrimitiveDesc(
+      {(int)layerSize_}, memory::format::x, engine_);
+  bias = MKLDNNMatrix::create(pd, biasMat);
+  outs.clear();
+  real* data = out->getData();
+  CHECK_EQ(bs_ * layerSize_, out->getElementCnt());
+  for (int i = 0; i < bs_; ++i) {
+    MatrixPtr tmp =
+        Matrix::create(data + i * layerSize_, 1, layerSize_, false, false);
+    outs.push_back(MKLDNNMatrix::create(bias->getPrimitiveDesc(), tmp));
+  }
+}
+
+void MKLDNNAddtoLayer::resetFwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
+                                       MKLDNNMatrixPtr& bias,
+                                       MKLDNNMatrixPtr& out) {
+  inputs.resize(inputLayers_.size());
+  for (size_t i = 0; i < inputs.size(); i++) {
+    resetInValue(inputs[i], nullptr, i);
+    CHECK(inputs[i]);
+    inputs[i]->downSpatial();
+  }
+  for (size_t i = 1; i < inputs.size(); i++) {
+    CHECK_PRIMITIVE_DESC_EQ(inputs[i], inputs[0]->getPrimitiveDesc());
+  }
+
+  resetOutValue(out, inputs[0]->getPrimitiveDesc());
+
+  if (biases_ && biases_->getW()) {
+    prepareBias(bias, biases_->getW(), out, vals_);
+  } else {
+    bias = nullptr;
+  }
+}
+
+void MKLDNNAddtoLayer::resetFwdPD(std::shared_ptr<sum::primitive_desc>& pd,
+                                  std::shared_ptr<sum::primitive_desc>& biasPD,
+                                  std::vector<MKLDNNMatrixPtr>& inputs,
+                                  MKLDNNMatrixPtr bias,
+                                  MKLDNNMatrixPtr out) {
+  std::vector<float> scales(inputs.size(), 1.0);
+  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 sum::primitive_desc(out->getMemoryDesc(), scales, srcPDs));
+  CHECK_PRIMITIVE_DESC_EQ(out, pd->dst_primitive_desc());
+
+  biasPD = nullptr;
+  if (bias) {
+    std::vector<float> scales(2, 1.0);
+    std::vector<memory::primitive_desc> srcPDs(2, bias->getPrimitiveDesc());
+    biasPD.reset(
+        new sum::primitive_desc(bias->getMemoryDesc(), scales, srcPDs));
+    CHECK_PRIMITIVE_DESC_EQ(bias, biasPD->dst_primitive_desc());
+  }
+}
+
+void MKLDNNAddtoLayer::resetFwdPipeline(
+    std::vector<primitive>& pipeline,
+    std::shared_ptr<sum::primitive_desc>& pd,
+    std::shared_ptr<sum::primitive_desc>& biasPD,
+    std::vector<MKLDNNMatrixPtr>& inputs,
+    MKLDNNMatrixPtr& bias,
+    MKLDNNMatrixPtr& out) {
+  std::vector<primitive::at> srcs;
+  for (size_t i = 0; i < inputs.size(); i++) {
+    srcs.push_back(*(inputs[i]));
+  }
+  fwd_.reset(new sum(*pd, srcs, *out));
+  pipeline.push_back(*fwd_);
+
+  fwdBias_.clear();
+  if (biasPD == nullptr || bias == nullptr) {
+    return;
+  }
+  fwdBias_.resize(vals_.size());
+  for (size_t i = 0; i < vals_.size(); ++i) {
+    std::vector<primitive::at> srcs;
+    srcs.push_back(*(vals_[i]));
+    srcs.push_back(*bias);
+    fwdBias_[i].reset(new sum(*biasPD, srcs, *vals_[i]));
+    pipeline.push_back(*fwdBias_[i]);
+  }
+}
+
+void MKLDNNAddtoLayer::resetBwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
+                                       MKLDNNMatrixPtr& bias,
+                                       MKLDNNMatrixPtr& out) {
+  CHECK(outVal_);
+  resetOutGrad(out, outVal_->getPrimitiveDesc());
+  CHECK(out);
+
+  inputs.resize(inputLayers_.size());
+  for (size_t i = 0; i < inputs.size(); i++) {
+    resetInGrad(inputs[i], inVal_->getPrimitiveDesc(), i);
+    CHECK_PRIMITIVE_DESC_EQ(inputs[i], out->getPrimitiveDesc());
+  }
+
+  if (biases_ && biases_->getWGrad()) {
+    prepareBias(bias, biases_->getWGrad(), out, grads_);
+  } else {
+    bias = nullptr;
+  }
+}
+
+}  // namespace paddle

paddle/gserver/layers/MKLDNNAddtoLayer.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 Addto layer.
+ *
+ * The config file api is mkldnn_addto
+ */
+class MKLDNNAddtoLayer : public MKLDNNLayer {
+protected:
+  std::vector<MKLDNNMatrixPtr> inVals_;
+  std::vector<MKLDNNMatrixPtr> inGrads_;
+
+  // layer size == ic * ih * iw == oc * oh *ow, and can not be changed
+  size_t layerSize_;
+
+  std::unique_ptr<Weight> biases_;
+
+  // buffers for adding bias
+  std::vector<MKLDNNMatrixPtr> vals_;
+  std::vector<MKLDNNMatrixPtr> grads_;
+  // primitives for adding bias
+  std::vector<std::shared_ptr<mkldnn::primitive>> fwdBias_;
+  std::shared_ptr<mkldnn::primitive> bwdBias_;
+
+public:
+  explicit MKLDNNAddtoLayer(const LayerConfig& config) : MKLDNNLayer(config) {}
+
+  ~MKLDNNAddtoLayer() {}
+
+  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 updateWeights(const UpdateCallback& callback) override;
+
+  void printValueFormat() override {
+    for (size_t i = 0; i < inVals_.size(); ++i) {
+      VLOG(MKLDNN_FMTS) << i << " input: " << 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) << i << " input: " << inGrads_[i]->getFormat() << "<<<";
+    }
+  }
+
+protected:
+  /**
+   * Forward functions: reset buffers(inputs, output, bias),
+   *                    reset primitive descriptor,
+   *                    reset pipeline.
+   */
+  void resetFwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
+                       MKLDNNMatrixPtr& bias,
+                       MKLDNNMatrixPtr& out);
+  void resetFwdPD(std::shared_ptr<mkldnn::sum::primitive_desc>& pd,
+                  std::shared_ptr<mkldnn::sum::primitive_desc>& biasPD,
+                  std::vector<MKLDNNMatrixPtr>& inputs,
+                  MKLDNNMatrixPtr bias,
+                  MKLDNNMatrixPtr out);
+  void resetFwdPipeline(std::vector<mkldnn::primitive>& pipeline,
+                        std::shared_ptr<mkldnn::sum::primitive_desc>& pd,
+                        std::shared_ptr<mkldnn::sum::primitive_desc>& biasPD,
+                        std::vector<MKLDNNMatrixPtr>& inputs,
+                        MKLDNNMatrixPtr& bias,
+                        MKLDNNMatrixPtr& out);
+
+  /**
+   * Backward functions: reset buffers(inputs, output, bias)
+   */
+  void resetBwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
+                       MKLDNNMatrixPtr& bias,
+                       MKLDNNMatrixPtr& out);
+
+  /**
+   * prepare for bias
+   */
+  void prepareBias(MKLDNNMatrixPtr& bias,
+                   const MatrixPtr& biasMat,
+                   const MKLDNNMatrixPtr& out,
+                   std::vector<MKLDNNMatrixPtr>& outs);
+};
+
+}  // namespace paddle

paddle/gserver/layers/MKLDNNBatchNormLayer.cpp

@@ -119,13 +119,12 @@ void MKLDNNBatchNormLayer::reshape(
     int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) {
   reshapeInput(bs, ih, iw);
   oh = ih;
-  ow = ow;
+  ow = iw;
   // ic_ and oc can not be changed
   CHECK_EQ(inputElemenCnt_ / bs / ih / iw, (size_t)ic)
       << "Input channel can not be changed";
   reshapeOutput(oh, ow);
   resizeOutput(bs, oc * oh * ow);
-  printSizeInfo();
 }
 
 void MKLDNNBatchNormLayer::resetFwd(std::vector<primitive>& pipeline,

paddle/gserver/layers/MKLDNNConvLayer.cpp

@@ -102,8 +102,6 @@ void MKLDNNConvLayer::reshape(
 
   reshapeOutput(oh, ow);
   resizeOutput(bs, oc * oh * ow);
-
-  printSizeInfo();
 }
 
 void MKLDNNConvLayer::resetFwd(std::vector<primitive>& pipeline,

paddle/gserver/layers/MKLDNNConvLayer.h

@@ -92,7 +92,7 @@ public:
   void printSizeInfo() override {
     MKLDNNLayer::printSizeInfo();
     VLOG(MKLDNN_SIZES) << getName() << ": fh: " << fh_ << ", fw: " << fw_
-                       << ": ph: " << ph_ << ", pw: " << pw_ << ", sh: " << sh_
+                       << ", ph: " << ph_ << ", pw: " << pw_ << ", sh: " << sh_
                        << ", sw: " << sw_ << ", dh: " << dh_ << ", dw: " << dw_;
   }
 

paddle/gserver/layers/MKLDNNFcLayer.cpp

@@ -60,18 +60,16 @@ void MKLDNNFcLayer::convertWeightsFromPaddle() {
   }
 
   CHECK(wgtVal_) << "should have been initialized";
-  bool hasNoSpatial_ = ih_ == 1 && iw_ == 1;
   auto targetDim = wgtVal_->getDims();
-  auto srcFmt = hasNoSpatial_ ? format::io : format::ihwo;
+  auto srcFmt = targetDim.size() == 2 ? format::io : format::ihwo;
   wgtVal_->reorderDataFrom(wgtVal_, srcFmt, targetDim);
   hasInitedWgt_ = true;
 }
 
 void MKLDNNFcLayer::convertWeightsToPaddle() {
   CHECK(wgtVal_) << "should have been initialized";
-  bool hasNoSpatial_ = ih_ == 1 && iw_ == 1;
   auto targetDim = wgtVal_->getDims();
-  auto dstFmt = hasNoSpatial_ ? format::io : format::ihwo;
+  auto dstFmt = targetDim.size() == 2 ? format::io : format::ihwo;
   wgtVal_->reorderDataTo(wgtVal_, dstFmt, targetDim);
 }
 
@@ -86,8 +84,6 @@ void MKLDNNFcLayer::reshape(
 
   reshapeOutput(oh, ow);
   resizeOutput(bs, oc);
-
-  printSizeInfo();
 }
 
 void MKLDNNFcLayer::resetFwd(std::vector<primitive>& pipeline,

paddle/gserver/layers/MKLDNNLayer.cpp

@@ -77,7 +77,7 @@ void MKLDNNLayer::forward(PassType passType) {
       needResetBwd_ = true;
     }
 
-    if (inputLayers_[0]->getType() == "data") {
+    if (inputLayers_[0]->getType() == "data" && inputLayers_.size() == 1) {
       // Update input value data when input layer is "data" type,
       // since the input value data address might be changed.
       CHECK(extInVal_);
@@ -171,29 +171,27 @@ void MKLDNNLayer::resetWithMatrix(MKLDNNMatrixPtr& dnn,
 }
 
 void MKLDNNLayer::resetInValue(
-    MKLDNNMatrixPtr& in, const std::shared_ptr<memory::primitive_desc>& intPD) {
+    MKLDNNMatrixPtr& in,
+    const std::shared_ptr<memory::primitive_desc>& intPD,
+    size_t inputIdx) {
   cvtInVal_ = nullptr;
   extInVal_ = nullptr;
   in = nullptr;
   CHECK_GT(bs_ * ic_ * ih_ * iw_, 0);
   auto extPD = MKLDNNMatrix::createPrimitiveDesc(
       {bs_, ic_, ih_, iw_}, format::nchw, engine_);
-  const MatrixPtr& inMat = inputLayers_[0]->getOutputValue();
-  in = std::dynamic_pointer_cast<MKLDNNMatrix>(inMat);
-  CHECK_EQ(inputIsOnlyMKLDNN(), in != nullptr);
-  if (in == nullptr || in->getFormat() == format::nc) {
-    in = MKLDNNMatrix::create(extPD, inMat);
-  }
-  extInVal_ = isPaddleFormat(in->getFormat()) ? in : nullptr;
-  if (in->getFormat() == format::nc) {
-    CHECK(ih_ == 1 && iw_ == 1);
+  const MatrixPtr& inMat = inputLayers_[inputIdx]->getOutputValue();
+  extInVal_ = std::dynamic_pointer_cast<MKLDNNMatrix>(inMat);
+  CHECK_EQ(inputIsOnlyMKLDNN(), extInVal_ != nullptr);
+  if (extInVal_ == nullptr || extInVal_->getFormat() == format::nc) {
+    extInVal_ = MKLDNNMatrix::create(extPD, inMat);
   }
+  in = extInVal_;
   if (nullptr == intPD || in->getPrimitiveDesc() == *intPD) {
     return;
   }
   // need create reorder
   in = MKLDNNMatrix::create(*intPD);
-  extInVal_ = extInVal_ ? extInVal_ : MKLDNNMatrix::create(extPD, inMat);
   cvtInVal_ = MKLDNNMatrix::createReorder(extInVal_, in);
   CHECK(cvtInVal_) << "should not be emptry";
 }
@@ -216,11 +214,12 @@ void MKLDNNLayer::resetOutValue(MKLDNNMatrixPtr& out,
 }
 
 void MKLDNNLayer::resetInGrad(MKLDNNMatrixPtr& in,
-                              memory::primitive_desc intPD) {
+                              memory::primitive_desc intPD,
+                              size_t inputIdx) {
   cvtInGrad_ = nullptr;
   extInGrad_ = nullptr;
   in = nullptr;
-  LayerPtr& input = inputLayers_[0];
+  LayerPtr& input = inputLayers_[inputIdx];
   if (input->getOutputGrad() == nullptr) {
     // no need input grad
     return;
@@ -245,7 +244,6 @@ void MKLDNNLayer::resetInGrad(MKLDNNMatrixPtr& in,
     return;
   }
   // need create reorder
-  // TODO(TJ): add macro definition to simplify it
   CHECK(extInVal_ != nullptr && isPaddleFormat(extInVal_->getFormat()))
       << "should have external input value and the format must be nchw(nc)";
   extInGrad_ = MKLDNNMatrix::create(extInVal_->getPrimitiveDesc(), inMat);
@@ -289,7 +287,7 @@ void MKLDNNLayer::resetMergeGrad(MKLDNNMatrixPtr& out) {
     return;
   }
   CHECK(out) << "should have reset internal ouput grad";
-  std::vector<double> scales(outputMap_.size(), 1.0);
+  std::vector<float> scales(outputMap_.size(), 1.0);
   std::vector<memory::primitive_desc> srcPDs;
   std::vector<primitive::at> srcs;
   for (auto it = outputMap_.begin(); it != outputMap_.end(); ++it) {

paddle/gserver/layers/MKLDNNLayer.h

@@ -199,7 +199,8 @@ protected:
    */
   void resetInValue(
       MKLDNNMatrixPtr& in,
-      const std::shared_ptr<mkldnn::memory::primitive_desc>& intPD = nullptr);
+      const std::shared_ptr<mkldnn::memory::primitive_desc>& intPD = nullptr,
+      size_t inputIdx = 0);
 
   /**
    * reset output value from internal primitive desc.
@@ -212,7 +213,9 @@ protected:
    * reset input grad from internal primitive desc.
    * reset both internal and external buffer and create reorder if necessary.
    */
-  void resetInGrad(MKLDNNMatrixPtr& in, mkldnn::memory::primitive_desc intPD);
+  void resetInGrad(MKLDNNMatrixPtr& in,
+                   mkldnn::memory::primitive_desc intPD,
+                   size_t inputIdx = 0);
 
   /**
    * reset output grad from internal primitive desc.

paddle/gserver/layers/MKLDNNPoolLayer.cpp

@@ -71,8 +71,6 @@ void MKLDNNPoolLayer::reshape(
   reshapeOutput(oh, ow);
 
   resizeOutput(bs, oc * oh * ow);
-
-  printSizeInfo();
 }
 
 void MKLDNNPoolLayer::resetFwd(std::vector<primitive>& pipeline,

paddle/gserver/layers/ROIPoolLayer.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 "ROIPoolLayer.h"
+
+namespace paddle {
+
+REGISTER_LAYER(roi_pool, ROIPoolLayer);
+
+bool ROIPoolLayer::init(const LayerMap& layerMap,
+                        const ParameterMap& parameterMap) {
+  Layer::init(layerMap, parameterMap);
+
+  const ROIPoolConfig& layerConf = config_.inputs(0).roi_pool_conf();
+  pooledWidth_ = layerConf.pooled_width();
+  pooledHeight_ = layerConf.pooled_height();
+  spatialScale_ = layerConf.spatial_scale();
+
+  return true;
+}
+
+void ROIPoolLayer::forward(PassType passType) {
+  Layer::forward(passType);
+
+  const ROIPoolConfig& layerConf = config_.inputs(0).roi_pool_conf();
+  height_ = getInput(0).getFrameHeight();
+  if (!height_) height_ = layerConf.height();
+  width_ = getInput(0).getFrameWidth();
+  if (!width_) width_ = layerConf.width();
+  channels_ = getInputValue(0)->getWidth() / width_ / height_;
+
+  size_t batchSize = getInput(0).getBatchSize();
+  size_t numROIs = getInput(1).getBatchSize();
+
+  MatrixPtr dataValue = getInputValue(0);
+  MatrixPtr roiValue = getInputValue(1);
+  resetOutput(numROIs, channels_ * pooledHeight_ * pooledWidth_);
+  MatrixPtr outputValue = getOutputValue();
+
+  if (useGpu_) {  // TODO(guosheng): implement on GPU later
+    MatrixPtr dataCpuBuffer;
+    Matrix::resizeOrCreate(dataCpuBuffer,
+                           dataValue->getHeight(),
+                           dataValue->getWidth(),
+                           false,
+                           false);
+    MatrixPtr roiCpuBuffer;
+    Matrix::resizeOrCreate(roiCpuBuffer,
+                           roiValue->getHeight(),
+                           roiValue->getWidth(),
+                           false,
+                           false);
+    dataCpuBuffer->copyFrom(*dataValue);
+    roiCpuBuffer->copyFrom(*roiValue);
+    dataValue = dataCpuBuffer;
+    roiValue = roiCpuBuffer;
+    MatrixPtr outputCpuBuffer;
+    Matrix::resizeOrCreate(outputCpuBuffer,
+                           outputValue->getHeight(),
+                           outputValue->getWidth(),
+                           false,
+                           false);
+    outputCpuBuffer->copyFrom(*outputValue);
+    outputValue = outputCpuBuffer;
+  }
+
+  real* bottomData = dataValue->getData();
+  size_t batchOffset = dataValue->getWidth();
+  size_t channelOffset = height_ * width_;
+  real* bottomROIs = roiValue->getData();
+  size_t roiOffset = roiValue->getWidth();
+  size_t poolChannelOffset = pooledHeight_ * pooledWidth_;
+
+  real* outputData = outputValue->getData();
+  Matrix::resizeOrCreate(maxIdxs_,
+                         numROIs,
+                         channels_ * pooledHeight_ * pooledWidth_,
+                         false,
+                         false);
+  real* argmaxData = maxIdxs_->getData();
+
+  for (size_t n = 0; n < numROIs; ++n) {
+    // the first five elememts of each RoI should be:
+    // batch_idx, roi_x_start, roi_y_start, roi_x_end, roi_y_end
+    size_t roiBatchIdx = bottomROIs[0];
+    size_t roiStartW = round(bottomROIs[1] * spatialScale_);
+    size_t roiStartH = round(bottomROIs[2] * spatialScale_);
+    size_t roiEndW = round(bottomROIs[3] * spatialScale_);
+    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);
+    real binSizeH =
+        static_cast<real>(roiHeight) / static_cast<real>(pooledHeight_);
+    real binSizeW =
+        static_cast<real>(roiWidth) / static_cast<real>(pooledWidth_);
+    real* batchData = bottomData + batchOffset * roiBatchIdx;
+    for (size_t c = 0; c < channels_; ++c) {
+      for (size_t ph = 0; ph < pooledHeight_; ++ph) {
+        for (size_t pw = 0; pw < pooledWidth_; ++pw) {
+          size_t hstart = static_cast<size_t>(std::floor(ph * binSizeH));
+          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_);
+
+          bool isEmpty = (hend <= hstart) || (wend <= wstart);
+          size_t poolIndex = ph * pooledWidth_ + pw;
+          if (isEmpty) {
+            outputData[poolIndex] = 0;
+            argmaxData[poolIndex] = -1;
+          }
+
+          for (size_t h = hstart; h < hend; ++h) {
+            for (size_t w = wstart; w < wend; ++w) {
+              size_t index = h * width_ + w;
+              if (batchData[index] > outputData[poolIndex]) {
+                outputData[poolIndex] = batchData[index];
+                argmaxData[poolIndex] = index;
+              }
+            }
+          }
+        }
+      }
+      batchData += channelOffset;
+      outputData += poolChannelOffset;
+      argmaxData += poolChannelOffset;
+    }
+    bottomROIs += roiOffset;
+  }
+  if (useGpu_) {
+    getOutputValue()->copyFrom(*outputValue);
+  }
+}
+
+void ROIPoolLayer::backward(const UpdateCallback& callback) {
+  MatrixPtr inGradValue = getInputGrad(0);
+  MatrixPtr outGradValue = getOutputGrad();
+  MatrixPtr roiValue = getInputValue(1);
+
+  if (useGpu_) {
+    MatrixPtr inGradCpuBuffer;
+    Matrix::resizeOrCreate(inGradCpuBuffer,
+                           inGradValue->getHeight(),
+                           inGradValue->getWidth(),
+                           false,
+                           false);
+    MatrixPtr outGradCpuBuffer;
+    Matrix::resizeOrCreate(outGradCpuBuffer,
+                           outGradValue->getHeight(),
+                           outGradValue->getWidth(),
+                           false,
+                           false);
+    MatrixPtr roiCpuBuffer;
+    Matrix::resizeOrCreate(roiCpuBuffer,
+                           roiValue->getHeight(),
+                           roiValue->getWidth(),
+                           false,
+                           false);
+    inGradCpuBuffer->copyFrom(*inGradValue);
+    outGradCpuBuffer->copyFrom(*outGradValue);
+    roiCpuBuffer->copyFrom(*roiValue);
+    inGradValue = inGradCpuBuffer;
+    outGradValue = outGradCpuBuffer;
+    roiValue = roiCpuBuffer;
+  }
+
+  real* bottomROIs = roiValue->getData();
+  size_t numROIs = getInput(1).getBatchSize();
+  size_t roiOffset = getInputValue(1)->getWidth();
+
+  real* inDiffData = inGradValue->getData();
+  size_t batchOffset = getInputValue(0)->getWidth();
+  size_t channelOffset = height_ * width_;
+
+  real* outDiffData = outGradValue->getData();
+  size_t poolChannelOffset = pooledHeight_ * pooledWidth_;
+  real* argmaxData = maxIdxs_->getData();
+
+  for (size_t n = 0; n < numROIs; ++n) {
+    size_t roiBatchIdx = bottomROIs[0];
+    real* batchDiffData = inDiffData + batchOffset * roiBatchIdx;
+    for (size_t c = 0; c < channels_; ++c) {
+      for (size_t ph = 0; ph < pooledHeight_; ++ph) {
+        for (size_t pw = 0; pw < pooledWidth_; ++pw) {
+          size_t poolIndex = ph * pooledWidth_ + pw;
+          if (argmaxData[poolIndex] > 0) {
+            size_t index = static_cast<size_t>(argmaxData[poolIndex]);
+            batchDiffData[index] += outDiffData[poolIndex];
+          }
+        }
+      }
+      batchDiffData += channelOffset;
+      outDiffData += poolChannelOffset;
+      argmaxData += poolChannelOffset;
+    }
+    bottomROIs += roiOffset;
+  }
+
+  if (useGpu_) {
+    getInputGrad(0)->copyFrom(*inGradValue);
+  }
+}
+
+}  // namespace paddle

paddle/gserver/layers/ROIPoolLayer.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 "Layer.h"
+
+namespace paddle {
+
+/**
+ * A layer used by Fast R-CNN to extract feature maps of ROIs from the last
+ * feature map.
+ * - Input: This layer needs two input layers: The first input layer is a
+ *          convolution layer; The second input layer contains the ROI data
+ *          which is the output of ProposalLayer in Faster R-CNN. layers for
+ *          generating bbox location offset and the classification confidence.
+ * - Output: The ROIs' feature map.
+ * Reference:
+ *    Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun.
+ *    Faster R-CNN: Towards Real-Time Object Detection with Region Proposal
+ * Networks
+ */
+
+class ROIPoolLayer : public Layer {
+protected:
+  size_t channels_;
+  size_t width_;
+  size_t height_;
+  size_t pooledWidth_;
+  size_t pooledHeight_;
+  real spatialScale_;
+
+  // Since there is no int matrix, use real maxtrix instead.
+  MatrixPtr maxIdxs_;
+
+public:
+  explicit ROIPoolLayer(const LayerConfig& config) : Layer(config) {}
+
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
+};
+}  // namespace paddle

paddle/gserver/layers/ScaleSubRegionLayer.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 "ScaleSubRegionLayer.h"
+#include "paddle/utils/Stat.h"
+namespace paddle {
+
+REGISTER_LAYER(scale_sub_region, ScaleSubRegionLayer);
+
+bool ScaleSubRegionLayer::init(const LayerMap& layerMap,
+                               const ParameterMap& parameterMap) {
+  Layer::init(layerMap, parameterMap);
+  CHECK_EQ(static_cast<int>(inputLayers_.size()), 2);
+  auto& conf = config_.inputs(0).scale_sub_region_conf();
+  value_ = conf.value();
+
+  createFunction(forward_, "ScaleSubRegion", FuncConfig().set("value", value_));
+  createFunction(
+      backward_, "ScaleSubRegionGrad", FuncConfig().set("value", value_));
+
+  return true;
+}
+
+void ScaleSubRegionLayer::forward(PassType passType) {
+  Layer::forward(passType);
+  auto in0 = getInput(0);
+  imgH_ = in0.getFrameHeight();
+  imgW_ = in0.getFrameWidth();
+  if (imgH_ == 0 || imgW_ == 0) {
+    auto& conf = config_.inputs(0).scale_sub_region_conf();
+    imgH_ = conf.image_conf().img_size_y();
+    imgW_ = conf.image_conf().img_size();
+  }
+  MatrixPtr imgV = in0.value;
+  size_t batchSize = imgV->getHeight();
+  size_t spatialSize = imgH_ * imgW_;
+  channelsNum_ = imgV->getWidth() / spatialSize;
+  shape_ = TensorShape({batchSize, channelsNum_, imgH_, imgW_});
+
+  resetOutput(batchSize, imgV->getWidth());
+  auto& out = getOutput();
+  out.setFrameHeight(imgH_);
+  out.setFrameWidth(imgW_);
+
+  MatrixPtr indicesV = getInputValue(1);
+  indicesShape_ = TensorShape({batchSize, 6});
+
+  REGISTER_TIMER_INFO("ScaleSubRegionForward", getName().c_str());
+  BufferArgs inArgs;
+  BufferArgs outArgs;
+  inArgs.addArg(*imgV, shape_);
+  inArgs.addArg(*indicesV, indicesShape_);
+  outArgs.addArg(*out.value, shape_, ASSIGN_TO);
+  forward_[0]->calc(inArgs, outArgs);
+}
+
+void ScaleSubRegionLayer::backward(const UpdateCallback& callback) {
+  REGISTER_TIMER_INFO("ScaleSubRegionBackward", getName().c_str());
+  BufferArgs inArgs;
+  BufferArgs outArgs;
+  inArgs.addArg(*getOutputGrad(), shape_);
+  inArgs.addArg(*getInputValue(1), indicesShape_);
+  outArgs.addArg(*getInputGrad(0), shape_, ADD_TO);
+  backward_[0]->calc(inArgs, outArgs);
+}
+
+}  // namespace paddle

paddle/gserver/layers/ScaleSubRegionLayer.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 "Layer.h"
+
+namespace paddle {
+
+/**
+ * \brief  For each instance, this layer can be used to multiply a value to a
+ *         specified sub continuous region. By providing start index and end
+ *         index for C/H/W, you can specify the location and shape of the
+ *         region.
+ *
+ *         input_0: Input value.
+ *         input_1: Indices value to specify the location an shape of the
+ *                  region.
+ */
+class ScaleSubRegionLayer : public Layer {
+public:
+  explicit ScaleSubRegionLayer(const LayerConfig& config) : Layer(config) {}
+
+  ~ScaleSubRegionLayer() {}
+
+  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+
+  void forward(PassType passType);
+
+  void backward(const UpdateCallback& callback = nullptr);
+
+protected:
+  TensorShape shape_;
+  TensorShape indicesShape_;
+  size_t imgH_;
+  size_t imgW_;
+  size_t channelsNum_;
+  real value_;
+};
+
+}  // namespace paddle

paddle/gserver/layers/SubSequenceLayer.cpp

@@ -98,8 +98,19 @@ void SubSequenceLayer::forward(PassType passType) {
   CHECK_EQ(numSequences2, numSequences3);
 
   MatrixPtr inputValue = input.value;
-  IVectorPtr offsetValue = offsetSeq.ids;
-  IVectorPtr sizeValue = sizeSeq.ids;
+  IVectorPtr offsetValue;
+  IVectorPtr sizeValue;
+
+  if (useGpu_) {
+    // copy to cpu
+    IVector::resizeOrCreate(offsetValue, offsetSeq.ids->getSize(), false);
+    IVector::resizeOrCreate(sizeValue, sizeSeq.ids->getSize(), false);
+    offsetValue->copyFrom(*offsetSeq.ids);
+    sizeValue->copyFrom(*sizeSeq.ids);
+  } else {
+    offsetValue = offsetSeq.ids;
+    sizeValue = sizeSeq.ids;
+  }
 
   CHECK_EQ(offsetValue->getSize(), numSequences1);
   CHECK_EQ(sizeValue->getSize(), numSequences1);
@@ -176,8 +187,21 @@ void SubSequenceLayer::backward(const UpdateCallback& callback) {
   size_t numSequences1 = startPositions1->getSize() - 1;
   const int* starts1 = startPositions1->getData();
 
-  IVectorPtr offsetValue = getInput(1).ids;
-  IVectorPtr sizeValue = getInput(2).ids;
+  const Argument& offsetSeq = getInput(1);
+  const Argument& sizeSeq = getInput(2);
+  IVectorPtr offsetValue;
+  IVectorPtr sizeValue;
+
+  if (useGpu_) {
+    // copy to cpu
+    IVector::resizeOrCreate(offsetValue, offsetSeq.ids->getSize(), false);
+    IVector::resizeOrCreate(sizeValue, sizeSeq.ids->getSize(), false);
+    offsetValue->copyFrom(*offsetSeq.ids);
+    sizeValue->copyFrom(*sizeSeq.ids);
+  } else {
+    offsetValue = offsetSeq.ids;
+    sizeValue = sizeSeq.ids;
+  }
 
   int* offsets = offsetValue->getData();
   int* sizes = sizeValue->getData();

paddle/gserver/tests/MKLDNNTester.cpp

@@ -132,7 +132,7 @@ void MKLDNNTester::checkForward() {
   VLOG(MKLDNN_TESTS) << "Check Forward";
   printTopDatas();
   double delta =
-      compareMatrix(dnnLayer_->getOutputValue(), refLayer_->getOutputValue());
+      compareMatrix(refLayer_->getOutputValue(), dnnLayer_->getOutputValue());
   EXPECT_LE(fabs(delta), eps_);
 }
 
@@ -147,7 +147,7 @@ void MKLDNNTester::checkBackwardData() {
     VLOG(MKLDNN_ALL) << "Reference Backward Result: InputGrad " << i;
     printMatrix(refDiff);
 
-    double delta = compareMatrix(dnnDiff, refDiff);
+    double delta = compareMatrix(refDiff, dnnDiff);
     EXPECT_LE(fabs(delta), eps_);
     if (isBN) {
       // the other two inputs in batch norm are for moving mean and var
@@ -177,7 +177,7 @@ void MKLDNNTester::checkBackwardWgts() {
                      << parameters_[REF][i]->getName();
     printVector(ref);
 
-    double delta = compareVector(dnn, ref);
+    double delta = compareVector(ref, dnn);
     EXPECT_LE(fabs(delta), eps_);
   }
 

paddle/gserver/tests/test_LayerGrad.cpp

@@ -53,7 +53,7 @@ TEST(Operator, dot_mul) {
 TEST(Projection, context) {
   for (auto contextStart : {-5, -3, -1, 0, 3}) {
     for (auto contextLength : {1, 2, 5, 7}) {
-      for (auto batchSize : {1, 2, 5, 20, 50}) {
+      for (auto batchSize : {1, 2, 5, 20}) {
         for (auto trainablePadding : {false, true}) {
           LOG(INFO) << " contextStart=" << contextStart
                     << " contextLength=" << contextLength
@@ -585,14 +585,14 @@ TEST(Layer, maxoutLayer) {
 }
 void testFcLayer(string format, size_t nnz) {
   TestConfig config;
-  config.biasSize = 4096;
+  config.biasSize = 1024;
   config.layerConfig.set_type("fc");
-  config.layerConfig.set_size(4096);
+  config.layerConfig.set_size(1024);
   config.layerConfig.set_active_type("sigmoid");
   config.layerConfig.set_drop_rate(0.1);
 
   config.inputDefs.push_back(
-      {INPUT_DATA, "layer_0", 8192, nnz, ParaSparse(format)});
+      {INPUT_DATA, "layer_0", 2048, nnz, ParaSparse(format)});
   config.layerConfig.add_inputs();
 
   LOG(INFO) << config.inputDefs[0].sparse.sparse << " "
@@ -609,9 +609,9 @@ void testFcLayer(string format, size_t nnz) {
 }
 
 TEST(Layer, fcLayer) {
-  testFcLayer("", 4096 * 4096 * 2);
-  testFcLayer("csc", 4096 * 40);
-  testFcLayer("csr", 4096 * 40);
+  testFcLayer("", 1024 * 1024 * 2);
+  testFcLayer("csc", 1024 * 10);
+  testFcLayer("csr", 1024 * 10);
 }
 
 TEST(Layer, SelectiveFullyConnectedLayer) {
@@ -1995,7 +1995,7 @@ TEST(Layer, multibox_loss) {
 TEST(Layer, TransLayer) {
   TestConfig config;
   const int height = 128;
-  const int width = 1028;
+  const int width = 256;
   config.layerConfig.set_type("trans");
   config.layerConfig.set_size(width);
 
@@ -2056,6 +2056,43 @@ TEST(Layer, CropLayer) {
   }
 }
 
+TEST(Layer, roi_pool) {
+  TestConfig config;
+  config.layerConfig.set_type("roi_pool");
+  config.biasSize = 0;
+  LayerInputConfig* input = config.layerConfig.add_inputs();
+  ROIPoolConfig* roiPoolConf = input->mutable_roi_pool_conf();
+  roiPoolConf->set_pooled_width(7);
+  roiPoolConf->set_pooled_height(7);
+  roiPoolConf->set_spatial_scale(1. / 16);
+  roiPoolConf->set_width(14);
+  roiPoolConf->set_height(14);
+
+  const size_t roiNum = 10;
+  const size_t roiDim = 10;
+  const size_t batchSize = 5;
+  MatrixPtr roiValue = Matrix::create(roiNum, roiDim, false, false);
+  roiValue->zeroMem();
+  real* roiData = roiValue->getData();
+  for (size_t i = 0; i < roiNum; ++i) {
+    roiData[i * roiDim + 0] = std::rand() % batchSize;
+    roiData[i * roiDim + 1] = std::rand() % 224;  // xMin
+    roiData[i * roiDim + 2] = std::rand() % 224;  // yMin
+    size_t xMin = static_cast<size_t>(roiData[i * roiDim + 1]);
+    size_t yMin = static_cast<size_t>(roiData[i * roiDim + 2]);
+    roiData[i * roiDim + 3] = xMin + std::rand() % (224 - xMin);  // xMax
+    roiData[i * roiDim + 4] = yMin + std::rand() % (224 - yMin);  // yMax
+  }
+
+  config.inputDefs.push_back({INPUT_DATA, "input", 3 * 14 * 14, {}});
+  config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA, "rois", roiValue, {}});
+  config.layerConfig.add_inputs();
+
+  for (auto useGpu : {false, true}) {
+    testLayerGrad(config, "roi_pool", batchSize, false, useGpu, false);
+  }
+}
+
 TEST(Layer, SwitchOrderLayer) {
   TestConfig config;
   // config input_0
@@ -2358,6 +2395,38 @@ TEST(Layer, ScaleShiftLayer) {
   }
 }
 
+TEST(Layer, ScaleSubRegionLayer) {
+  const size_t batchSize = 64;
+  const size_t size = 4096;
+  TestConfig config;
+  config.layerConfig.set_type("scale_sub_region");
+  config.inputDefs.push_back({INPUT_DATA, "input", size, 0});
+  MatrixPtr indicesV = Matrix::create(batchSize, 6, false, false);
+  auto* data = indicesV->getData();
+  for (size_t i = 0; i < batchSize; ++i) {
+    data[i * 2] = 2;
+    data[i * 2 + 1] = 4;
+    data[i * 2 + 2] = 16;
+    data[i * 2 + 3] = 32;
+    data[i * 2 + 4] = 16;
+    data[i * 2 + 5] = 32;
+  }
+  config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA, "indices", indicesV, {}});
+  LayerInputConfig* input = config.layerConfig.add_inputs();
+  ScaleSubRegionConfig* scaleSubRegionConf =
+      input->mutable_scale_sub_region_conf();
+  ImageConfig* imgConf = scaleSubRegionConf->mutable_image_conf();
+  imgConf->set_img_size(32);
+  imgConf->set_img_size_y(32);
+  imgConf->set_channels(4);
+  scaleSubRegionConf->set_value(2.0);
+  config.layerConfig.add_inputs();
+
+  for (auto useGpu : {false, true}) {
+    testLayerGrad(config, "scale_sub_region", batchSize, false, useGpu, false);
+  }
+}
+
 int main(int argc, char** argv) {
   testing::InitGoogleTest(&argc, argv);
   initMain(argc, argv);

paddle/gserver/tests/test_MKLDNN.cpp

@@ -269,22 +269,51 @@ void testBatchNormLayer(const testBatchNormDesc& pm) {
 TEST(MKLDNNLayer, BatchNormLayer) {
   testBatchNormLayer({4, 10, 6, 6});
   testBatchNormLayer({16, 32, 16, 16});
+  testBatchNormLayer({4, 16, 8, 10});
 }
 
-struct testActDesc {
+struct testImageDesc {
   int bs, ic, ih, iw;
 };
 
-static void getAddtoConfig(TestConfig& cfg, const testActDesc& pm) {
+static void getAddtoConfig(TestConfig& cfg,
+                           const testImageDesc& pm,
+                           const size_t nInputs = 1) {
   cfg.biasSize = 0;
   cfg.layerConfig.set_type("addto");
   size_t layerSize = pm.ic * pm.ih * pm.iw;
   cfg.layerConfig.set_size(layerSize);
-  cfg.inputDefs.push_back({INPUT_DATA, "layer_0", layerSize, 0});
-  cfg.layerConfig.add_inputs();
+  cfg.layerConfig.set_active_type("relu");
+  for (size_t i = 0; i < nInputs; ++i) {
+    std::stringstream ss;
+    ss << "layer_" << i;
+    cfg.inputDefs.push_back({INPUT_DATA, ss.str(), layerSize, 0});
+    LayerInputConfig* input = cfg.layerConfig.add_inputs();
+    ImageConfig* img_conf = input->mutable_image_conf();
+    img_conf->set_channels(pm.ic);
+    img_conf->set_img_size_y(pm.ih);
+    img_conf->set_img_size(pm.iw);
+  }
+}
+
+void testAddtoLayer(const testImageDesc& pm, const size_t nInputs) {
+  CHECK_GE(nInputs, 1UL);
+  TestConfig dnnConfig;
+  getAddtoConfig(dnnConfig, pm, nInputs);
+  dnnConfig.layerConfig.set_type("mkldnn_addto");
+  for (auto withBias : {false, true}) {
+    dnnConfig.biasSize = withBias ? pm.ic * pm.ih * pm.iw : 0;
+    RUN_MKLDNN_TEST_LAYER(dnnConfig, "addto", pm)
+  }
+}
+
+TEST(MKLDNNLayer, AddtoLayer) {
+  testAddtoLayer({16, 5, 14, 14}, 1);
+  testAddtoLayer({8, 10, 8, 8}, 2);
+  testAddtoLayer({4, 12, 1, 1}, 3);
 }
 
-void testActivation(std::string actType, const testActDesc& pm) {
+void testActivation(std::string actType, const testImageDesc& pm) {
   // TODO(TJ): remove me when paddle support elu activation
   if (actType == "mkldnn_elu") {
     return;

paddle/math/MKLDNNMatrix.cpp

@@ -152,12 +152,7 @@ void MKLDNNMatrix::downSpatial() {
   }
   memory::desc md = memory::desc(dstDims, getDtype(), dstFmt);
   memory::primitive_desc pd = memory::primitive_desc(md, getEngine());
-  mkldnn_primitive_t result;
-  mkldnn::error::wrap_c_api(
-      mkldnn_primitive_create(&result, pd.get(), nullptr, nullptr),
-      "could not create a memory primitive");
-  reset(result);
-  set_data_handle(data_);
+  resetMKLDNNMemory(pd, data_);
 }
 
 }  // namespace paddle

paddle/math/MKLDNNMatrix.h

@@ -145,6 +145,27 @@ public:
     m_.reset();
   }
 
+  /**
+   * override the CpuMatrix::resize
+   */
+  void resize(size_t newHeight, size_t newWidth) override {
+    m_->resize(newHeight, newWidth);
+    if (data_ == m_->getData() && elementCnt_ == newHeight * newWidth) {
+      return;
+    }
+    CpuMatrix::setData(data_);
+    height_ = newHeight;
+    width_ = newWidth;
+    elementCnt_ = newHeight * newWidth;
+    stride_ = width_;
+    auto pd = mkldnn::memory::primitive_desc(
+        mkldnn::memory::desc({(int)newHeight, (int)newWidth},
+                             getDtype(),
+                             mkldnn::memory::format::nc),
+        getEngine());
+    resetMKLDNNMemory(pd, data_);
+  }
+
   /**
    * override Matrix::getData
    * check data before return
@@ -215,6 +236,17 @@ protected:
                    memory::format srcFmt,
                    memory::format dstFmt,
                    memory::dims dm);
+  /**
+   * reset this MKLDNN Memory from primitve desc
+   */
+  void resetMKLDNNMemory(memory::primitive_desc pd, real* data) {
+    mkldnn_primitive_t result;
+    mkldnn::error::wrap_c_api(
+        mkldnn_primitive_create(&result, pd.get(), nullptr, nullptr),
+        "could not create a memory primitive");
+    reset(result);
+    set_data_handle(data);
+  }
 
 private:
   // save the CpuMatrixPtr in case the buffer released outside

paddle/math/MathFunctions.cpp

@@ -206,7 +206,7 @@ double dotProduct<double>(const int n, const double* x, const double* y) {
 }
 #endif
 
-#if defined(PADDLE_USE_MKL) || defined(PADDLE_USE_MKLML)
+#if defined(PADDLE_USE_MKLML)
 
 template <>
 void vExp<float>(const int n, const float* a, float* r) {
@@ -295,38 +295,6 @@ template void vAdd(const int n, const double* a, const double* b, double* r);
 
 #endif
 
-#ifdef PADDLE_USE_MKL
-template <>
-void vInvSqrt<float>(const int n, const float* a, float* r) {
-  vsInvSqrt(n, a, r);
-}
-
-template <>
-void vInvSqrt<double>(const int n, const double* a, double* r) {
-  vdInvSqrt(n, a, r);
-}
-
-template <>
-void vLog1p<float>(const int n, const float* a, float* r) {
-  vsLog1p(n, a, r);
-}
-
-template <>
-void vLog1p<double>(const int n, const double* a, double* r) {
-  vdLog1p(n, a, r);
-}
-
-template <>
-void vTanh<float>(const int n, const float* a, float* r) {
-  vsTanh(n, a, r);
-}
-
-template <>
-void vTanh<double>(const int n, const double* a, double* r) {
-  vdTanh(n, a, r);
-}
-#else
-
 DEFINE_MATRIX_BINARY_OP(vInvSqrt, b = 1.0f / std::sqrt(a));
 template <class T>
 void vInvSqrt(const int n, const T* a, T* r) {
@@ -357,6 +325,4 @@ template void vLog1p(const int n, const double* a, double* r);
 template void vTanh(const int n, const float* a, float* r);
 template void vTanh(const int n, const double* a, double* r);
 
-#endif
-
 }  // namespace paddle

paddle/math/MathFunctions.h

@@ -21,11 +21,6 @@ limitations under the License. */
 #include <mkl_vml_functions.h>
 #endif
 
-#ifdef PADDLE_USE_MKL
-#include <mkl.h>
-#include <mkl_lapacke.h>
-#endif
-
 #if defined(PADDLE_USE_ATLAS) || defined(PADDLE_USE_VECLIB)
 extern "C" {
 #include <cblas.h>

paddle/math/tests/TensorCheck.h

@@ -169,7 +169,7 @@ void TensorCheck(AssertEq compare,
       count++;
     }
   }
-  EXPECT_EQ(count, 0) << "There are " << count << " different element.";
+  EXPECT_EQ(count, 0) << "There are " << count << " different elements.";
 }
 
 template <typename AssertEq, typename Tensor1, typename Tensor2>